diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index eeda759ff18ccb86ce6a585fe41cb972ea3ae295..e718b32cb6c48d11e73600509a17db107f438708 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -23,7 +23,7 @@ repos: - id: clang-format-with-version-check name: clang-format description: Format files with ClangFormat. - entry: bash ./.clang_format.hook -i + entry: bash ./tools/codestyle/clang_format.hook -i language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$ - repo: local @@ -52,7 +52,7 @@ repos: hooks: - id: copyright_checker name: copyright_checker - entry: python ./.copyright.hook + entry: python ./tools/codestyle/copyright.hook language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto|py)$ exclude: (?!.*third_party)^.*$ | (?!.*book)^.*$ diff --git a/.travis.yml b/.travis.yml index 8c772030925dcad3909f142b08e4d8057a3f89b7..361136ac2c8d899a0d7a4d7945083fcc489551b5 100644 --- a/.travis.yml +++ b/.travis.yml @@ -27,15 +27,6 @@ script: # 43min timeout paddle/scripts/paddle_docker_build.sh ${JOB} if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi; - - | - if [[ "$JOB" != "doc" ]]; then exit 0; fi; - # For document only - if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi; - if [[ "$TRAVIS_BRANCH" != "develop" && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi; - export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh - export DOCS_DIR=`pwd` - cd .. - curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/ notifications: email: on_success: change diff --git a/AUTHORS.md b/AUTHORS.md index 11f227be7148d8d6e055538347a8c31679406c84..41b7193677a0208ba2fa82b72862292572dcb6ef 100644 --- a/AUTHORS.md +++ b/AUTHORS.md @@ -22,6 +22,7 @@ | jczaja | Jacek Czaja | | JiayiFeng | Jia-Yi Feng | | kbinias | Krzysztof Binias | +| kexinzhao | Ke-Xin Zhao | | kuke | Yi-Bing Liu | | lcy-seso | Ying Cao | | lipeng-unisound | Peng Li | @@ -45,6 +46,7 @@ | tianbingsz | Tian-Bing Xu | | tpatejko | Tomasz Patejko | | typhoonzero | Yi Wu | +| velconia | Qi-Yang Min | | wanghaoshuang | Hao-Shuang Wang | | wangyang59 | Yang Wang | | wangzhen-nlp | Zhen Wang | diff --git a/CMakeLists.txt b/CMakeLists.txt index cfaab206e1f321a55119d4a8d65c4a99d3819fff..920c20d6f813c14df8e02593b1c4a5a13cc11ef0 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -55,12 +55,25 @@ option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) -option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF) +option(WITH_DISTRIBUTE "Compile with distributed support" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF) option(WITH_CONTRIB "Compile the third-party contributation" OFF) +option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF) +option(WITH_ANAKIN "Compile with Anakin library" OFF) +option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE}) +option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF) +option(WITH_INFERENCE "Compile fluid inference library" ON) +option(WITH_SYSTEM_BLAS "Use system blas library" OFF) +option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION}) + +# PY_VERSION +if(NOT PY_VERSION) + set(PY_VERSION 2.7) +endif() +set(PYBIND11_PYTHON_VERSION ${PY_VERSION}) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -98,6 +111,11 @@ if(ANDROID OR IOS) add_definitions(-DPADDLE_MOBILE_INFERENCE) endif() +if (APPLE OR WIN32) + set(WITH_MKL OFF CACHE STRING + "Disable MKL for building on mac and windows" FORCE) +endif() + set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING "A path setting third party libraries download & build directories.") @@ -120,6 +138,12 @@ else() set(THIRD_PARTY_BUILD_TYPE Release) endif() +if(WITH_MKL) + option(MKL_SPLIT_GEMM "PaddlePaddle MKL gemm would split to small ones" OFF) + if (MKL_SPLIT_GEMM) + add_definitions(-DPADDLE_MKL_SPLIT_GEMM) + endif() +endif() set(WITH_MKLML ${WITH_MKL}) if (NOT DEFINED WITH_MKLDNN) if (WITH_MKL AND AVX2_FOUND) @@ -129,9 +153,15 @@ if (NOT DEFINED WITH_MKLDNN) set(WITH_MKLDNN OFF) endif() endif() + +if (REPLACE_ENFORCE_GLOG) + add_definitions("-DREPLACE_ENFORCE_GLOG") +endif() ######################################################################################## include(external/mklml) # download mklml package +include(external/xbyak) # download xbyak package +include(external/libxsmm) # download, build, install libxsmm include(external/zlib) # download, build, install zlib include(external/gflags) # download, build, install gflags include(external/glog) # download, build, install glog @@ -147,11 +177,40 @@ include(external/any) # download libn::any include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/cares) -include(external/grpc) +include(external/cub) + +if(WITH_DISTRIBUTE) + if(WITH_GRPC) + include(external/grpc) + message(STATUS "Use grpc framework.") + else() + message(STATUS "Use brpc framework.") + include(external/leveldb) + include(external/brpc) + endif() +endif() + +if(WITH_BRPC_RDMA) + message(STATUS "Use brpc with rdma.") + if(WITH_GRPC) + message(FATAL_ERROR "Can't use grpc with brpc rdma.") + endif() + if(NOT WITH_DISTRIBUTE) + message(FATAL_ERROR "Can't use brpc rdma in no distribute env.") + endif() +endif() + include(external/snappy) # download snappy include(external/snappystream) include(external/threadpool) +set(WITH_ANAKIN OFF CACHE STRING "Disable Anakin first, will add it later." FORCE) +if(WITH_GPU) + include(cuda) + include(tensorrt) + include(external/anakin) +endif() + include(cudnn) # set cudnn libraries, must before configure include(cupti) include(configure) # add paddle env configuration @@ -167,7 +226,7 @@ include(inference_lib) # add paddle fluid inference libraries include_directories("${PADDLE_SOURCE_DIR}") -include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include") +include_directories("${PADDLE_SOURCE_DIR}/paddle/legacy/cuda/include") include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto") include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c") @@ -180,11 +239,6 @@ set(EXTERNAL_LIBS ${PYTHON_LIBRARIES} ) -if(WITH_GPU) - include(cuda) - include(tensorrt) -endif(WITH_GPU) - if(WITH_AMD_GPU) find_package(HIP) include(hip) @@ -194,6 +248,10 @@ if(WITH_MKLML) list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB}) endif() +if(WITH_LIBXSMM) + list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS}) +endif() + if(WITH_MKLDNN) list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB}) endif() @@ -208,7 +266,7 @@ add_subdirectory(proto) if(NOT MOBILE_INFERENCE AND NOT WITH_FLUID_ONLY) # "add_subdirectory(go)" should be placed after the following loine, # because it depends on paddle/optimizer. - add_subdirectory(paddle/optimizer) + add_subdirectory(paddle/legacy/optimizer) endif() # "add_subdirectory(paddle)" and "add_subdirectory(python)" should be @@ -233,7 +291,3 @@ if(WITH_DOC) find_python_module(recommonmark REQUIRED) add_subdirectory(doc) endif() - -if (WITH_CONTRIB) - add_subdirectory(paddle/contrib) -endif() diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index b1b02bcc2f4fd14297715bcf5bfd1617e3d5f0c9..b878f37a5b8e807e5aa346e0074a741f2f8b6cc5 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -159,4 +159,4 @@ This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the - verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework) - verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators) - verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform) -- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/math) +- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math) diff --git a/Dockerfile b/Dockerfile index 4d6165b79a1d94b8f27d7f3ee1b6e2cee5992d31..402adee2ea2822250ebc8f6229fd6a44545d58e5 100644 --- a/Dockerfile +++ b/Dockerfile @@ -23,8 +23,8 @@ ENV HOME /root COPY ./paddle/scripts/docker/root/ /root/ RUN apt-get update && \ - apt-get install -y --allow-downgrades \ - git python-pip python-dev openssh-server bison \ + apt-get install -y --allow-downgrades patchelf \ + git python-pip python-dev python-opencv openssh-server bison \ libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \ wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \ curl sed grep graphviz libjpeg-dev zlib1g-dev \ @@ -80,7 +80,7 @@ RUN pip install pre-commit 'ipython==5.3.0' && \ pip install opencv-python #For docstring checker -RUN pip install pylint pytest astroid isort +RUN pip install pylint pytest astroid isort LinkChecker COPY ./python/requirements.txt /root/ RUN pip install -r /root/requirements.txt diff --git a/README.md b/README.md index 8d89c6b1ec9e4aefbd64328dedb4e8c7cc50c21b..a67cb8ad439f462c361cb6bac2449c3a4b042126 100644 --- a/README.md +++ b/README.md @@ -4,7 +4,6 @@ [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) [![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) [![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) -[![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop) [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) @@ -19,6 +18,22 @@ learning to many products at Baidu. Our vision is to enable deep learning for everyone via PaddlePaddle. Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle. + +### Latest PaddlePaddle Release: [Fluid 0.14.0](https://github.com/PaddlePaddle/Paddle/tree/v0.14.0) +### Install Latest Stable Release: +``` +# Linux CPU +pip install paddlepaddle +# Linux GPU cuda9cudnn7 +pip install paddlepaddle-gpu +# Linux GPU cuda8cudnn7 +pip install paddlepaddle-gpu==0.14.0.post87 +# Linux GPU cuda8cudnn5 +pip install paddlepaddle-gpu==0.14.0.post85 + +# For installation on other platform, refer to http://paddlepaddle.org/ +``` + ## Features - **Flexibility** diff --git a/benchmark/.gitignore b/benchmark/.gitignore index 7b66e8a5b5020fd847982db401665d24ba3a069c..fb4114356d4f37efc8ad672316fd4f99443d9fcd 100644 --- a/benchmark/.gitignore +++ b/benchmark/.gitignore @@ -7,3 +7,6 @@ paddle/rnn/imdb.pkl caffe/image/logs tensorflow/image/logs tensorflow/rnn/logs +fluid/models/*.pyc +fluid/logs +fluid/nohup.out diff --git a/benchmark/fluid/Dockerfile b/benchmark/fluid/Dockerfile new file mode 100644 index 0000000000000000000000000000000000000000..707fadb1fae97cefe8a41715cd57d71754abda41 --- /dev/null +++ b/benchmark/fluid/Dockerfile @@ -0,0 +1,31 @@ +FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04 + +# Use UBUNTU_MIRROR can speed up apt-get speed. +# ARG UBUNTU_MIRROR +# RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi' + +RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop python-opencv +RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/lib/libnccl.so + +# IMPORTANT: +# Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime. +# exmaple: unset http_proxy && unset https_proxy && python fluid_benchmark.py ... + +RUN pip install -U pip +RUN pip install -U kubernetes paddlepaddle + +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python' +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python' +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.imikolov.fetch()" | python' +RUN pip uninstall -y paddlepaddle && mkdir /workspace + +ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin +ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root +RUN chmod +x /usr/bin/paddle_k8s + +ADD *.whl / +RUN pip install /*.whl && rm -f /*.whl + +ENV LD_LIBRARY_PATH=/usr/local/lib +ADD fluid_benchmark.py recordio_converter.py args.py recordio_converter.py run.sh run_fluid_benchmark.sh /workspace/ +ADD models/ /workspace/models/ diff --git a/benchmark/fluid/README.md b/benchmark/fluid/README.md index 7071e9fdcd394a5a4db4d0d599610a72d98c0a3c..28cade4634bb62723bf5120169e202657f548234 100644 --- a/benchmark/fluid/README.md +++ b/benchmark/fluid/README.md @@ -24,14 +24,18 @@ Currently supported `--model` argument include: * Run the following command to start a benchmark job locally: ```bash - python fluid_benchmark.py --model mnist --device GPU + python fluid_benchmark.py --model mnist --device GPU ``` You can choose to use GPU/CPU training. With GPU training, you can specify `--gpus ` to run multi GPU training. + You can set async mode parameter server. With async mode, you can specify + `--async_mode` to train model asynchronous. * Run distributed training with parameter servers: + * see [run_fluid_benchmark.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/fluid/run_fluid_benchmark.sh) as an example. * start parameter servers: ```bash PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver + sleep 15 ``` * start trainers: ```bash @@ -42,13 +46,37 @@ Currently supported `--model` argument include: PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method nccl2 ``` +## Prepare the RecordIO file to Achieve Better Performance + +Run the following command will generate RecordIO files like "mnist.recordio" under the path +and batch_size you choose, you can use batch_size=1 so that later reader can change the batch_size +at any time using `fluid.batch`. + +```bash +python -c 'from recordio_converter import *; prepare_mnist("data", 1)' +``` + ## Run Distributed Benchmark on Kubernetes Cluster +You may need to build a Docker image before submitting a cluster job onto Kubernetes, or you will +have to start all those processes mannually on each node, which is not recommended. + +To build the Docker image, you need to choose a paddle "whl" package to run with, you may either +download it from +http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_en.html or +build it by your own. Once you've got the "whl" package, put it under the current directory and run: + +```bash +docker build -t [your docker image name]:[your docker image tag] . +``` + +Then push the image to a Docker registry that your Kubernetes cluster can reach. + We provide a script `kube_gen_job.py` to generate Kubernetes yaml files to submit distributed benchmark jobs to your cluster. To generate a job yaml, just run: ```bash -python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver " --disttype pserver +python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --gpus 8 --device GPU --update_method pserver " --disttype pserver ``` Then the yaml files are generated under directory `myjob`, you can run: diff --git a/benchmark/fluid/args.py b/benchmark/fluid/args.py new file mode 100644 index 0000000000000000000000000000000000000000..a79f25ccc6ace1594f3f331633130eaace5e175b --- /dev/null +++ b/benchmark/fluid/args.py @@ -0,0 +1,134 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse + +__all__ = ['parse_args', ] + +BENCHMARK_MODELS = [ + "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm" +] + + +def parse_args(): + parser = argparse.ArgumentParser('Fluid model benchmarks.') + parser.add_argument( + '--model', + type=str, + choices=BENCHMARK_MODELS, + default='resnet', + help='The model to run benchmark with.') + parser.add_argument( + '--batch_size', type=int, default=32, help='The minibatch size.') + # args related to learning rate + parser.add_argument( + '--learning_rate', type=float, default=0.001, help='The learning rate.') + # TODO(wuyi): add "--use_fake_data" option back. + parser.add_argument( + '--skip_batch_num', + type=int, + default=5, + help='The first num of minibatch num to skip, for better performance test' + ) + parser.add_argument( + '--iterations', type=int, default=80, help='The number of minibatches.') + parser.add_argument( + '--pass_num', type=int, default=100, help='The number of passes.') + parser.add_argument( + '--data_format', + type=str, + default='NCHW', + choices=['NCHW', 'NHWC'], + help='The data data_format, now only support NCHW.') + parser.add_argument( + '--device', + type=str, + default='GPU', + choices=['CPU', 'GPU'], + help='The device type.') + parser.add_argument( + '--gpus', + type=int, + default=1, + help='If gpus > 1, will use ParallelExecutor to run, else use Executor.') + # this option is available only for vgg and resnet. + parser.add_argument( + '--cpus', + type=int, + default=1, + help='If cpus > 1, will use ParallelDo to run, else use Executor.') + parser.add_argument( + '--data_set', + type=str, + default='flowers', + choices=['cifar10', 'flowers'], + help='Optional dataset for benchmark.') + parser.add_argument( + '--infer_only', action='store_true', help='If set, run forward only.') + parser.add_argument( + '--use_cprof', action='store_true', help='If set, use cProfile.') + parser.add_argument( + '--use_nvprof', + action='store_true', + help='If set, use nvprof for CUDA.') + parser.add_argument( + '--no_test', + action='store_true', + help='If set, do not test the testset during training.') + parser.add_argument( + '--memory_optimize', + action='store_true', + help='If set, optimize runtime memory before start.') + parser.add_argument( + '--use_fake_data', + action='store_true', + help='If set ommit the actual read data operators.') + parser.add_argument( + '--profile', action='store_true', help='If set, profile a few steps.') + parser.add_argument( + '--update_method', + type=str, + default='local', + choices=['local', 'pserver', 'nccl2'], + help='Choose parameter update method, can be local, pserver, nccl2.') + parser.add_argument( + '--no_split_var', + action='store_true', + default=False, + help='Whether split variables into blocks when update_method is pserver') + parser.add_argument( + '--async_mode', + action='store_true', + default=False, + help='Whether start pserver in async mode to support ASGD') + parser.add_argument( + '--use_reader_op', + action='store_true', + help='Whether to use reader op, and must specify the data path if set this to true.' + ) + parser.add_argument( + '--data_path', + type=str, + default="", + help='Directory that contains all the training recordio files.') + parser.add_argument( + '--use_inference_transpiler', + action='store_true', + help='If set, use inference transpiler to optimize the program.') + parser.add_argument( + '--no_random', + action='store_true', + help='If set, keep the random seed and do not shuffle the data.') + args = parser.parse_args() + return args diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py index c1d458970a58bfac2a3369e8964eb100568b28f2..6b22f8f520e3d9c6c89d41a7455a6f9ebbad6d80 100644 --- a/benchmark/fluid/fluid_benchmark.py +++ b/benchmark/fluid/fluid_benchmark.py @@ -24,90 +24,7 @@ import paddle.fluid.core as core import paddle.fluid.profiler as profiler import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler -BENCHMARK_MODELS = [ - "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm" -] - - -def parse_args(): - parser = argparse.ArgumentParser('Fluid model benchmarks.') - parser.add_argument( - '--model', - type=str, - choices=BENCHMARK_MODELS, - default='resnet', - help='The model to run benchmark with.') - parser.add_argument( - '--batch_size', type=int, default=32, help='The minibatch size.') - parser.add_argument( - '--learning_rate', - type=float, - default=0.001, - help='The minibatch size.') - # TODO(wuyi): add "--use_fake_data" option back. - parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test' - ) - parser.add_argument( - '--iterations', type=int, default=80, help='The number of minibatches.') - parser.add_argument( - '--pass_num', type=int, default=100, help='The number of passes.') - parser.add_argument( - '--data_format', - type=str, - default='NCHW', - choices=['NCHW', 'NHWC'], - help='The data data_format, now only support NCHW.') - parser.add_argument( - '--device', - type=str, - default='GPU', - choices=['CPU', 'GPU'], - help='The device type.') - parser.add_argument( - '--gpus', - type=int, - default=1, - help='If gpus > 1, will use ParallelExecutor to run, else use Executor.') - parser.add_argument( - '--data_set', - type=str, - default='flowers', - choices=['cifar10', 'flowers'], - help='Optional dataset for benchmark.') - parser.add_argument( - '--infer_only', action='store_true', help='If set, run forward only.') - parser.add_argument( - '--use_cprof', action='store_true', help='If set, use cProfile.') - parser.add_argument( - '--use_nvprof', - action='store_true', - help='If set, use nvprof for CUDA.') - parser.add_argument( - '--no_test', - action='store_false', - help='If set, test the testset during training.') - parser.add_argument( - '--memory_optimize', - action='store_true', - help='If set, optimize runtime memory before start.') - parser.add_argument( - '--use_fake_data', - action='store_true', - help='If set ommit the actual read data operators.') - parser.add_argument( - '--profile', action='store_true', help='If set, profile a few steps.') - parser.add_argument( - '--update_method', - type=str, - default='local', - choices=['local', 'pserver', 'nccl2'], - help='Choose parameter update method, can be local, pserver, nccl2.') - args = parser.parse_args() - return args +from args import * def append_nccl2_prepare(trainer_id): @@ -142,7 +59,7 @@ def append_nccl2_prepare(trainer_id): "nccl-based dist train.") -def dist_transpile(trainer_id): +def dist_transpile(trainer_id, args): if trainer_id < 0: return None, None @@ -164,7 +81,11 @@ def dist_transpile(trainer_id): training_role = os.getenv("PADDLE_TRAINING_ROLE") t = distribute_transpiler.DistributeTranspiler() - t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) + t.transpile( + trainer_id, + pservers=pserver_endpoints, + trainers=trainers, + sync_mode=not args.async_mode) if training_role == "PSERVER": pserver_program = t.get_pserver_program(current_endpoint) pserver_startup_program = t.get_startup_program(current_endpoint, @@ -175,7 +96,7 @@ def dist_transpile(trainer_id): return train_program, fluid.default_startup_program() else: raise ValueError( - 'TRAINING_ROLE environment variable must be either TRAINER or PSERVER' + 'PADDLE_TRAINING_ROLE environment variable must be either TRAINER or PSERVER' ) @@ -208,36 +129,62 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) exe = fluid.Executor(place) exe.run(startup_prog) - feed_var_list = [ - var for var in train_prog.global_block().vars.itervalues() - if var.is_data - ] - feeder = fluid.DataFeeder(feed_var_list, place) + + # Use inference_transpiler to speedup + if not args.use_reader_op: + feed_var_list = [ + var for var in train_prog.global_block().vars.itervalues() + if var.is_data + ] + feeder = fluid.DataFeeder(feed_var_list, place) iters, num_samples, start_time = 0, 0, time.time() for pass_id in range(args.pass_num): train_losses = [] - for batch_id, data in enumerate(train_reader()): + if not args.use_reader_op: + reader_generator = train_reader() + batch_id = 0 + data = None + while True: + if not args.use_reader_op: + data = next(reader_generator, None) + if data == None: + break + if iters == args.iterations: + break if iters == args.skip_batch_num: start_time = time.time() num_samples = 0 - if iters == args.iterations: - break - loss = exe.run(train_prog, - feed=feeder.feed(data), - fetch_list=[avg_loss]) + + if args.use_reader_op: + try: + loss = exe.run(train_prog, fetch_list=[avg_loss]) + except fluid.core.EnforceNotMet as ex: + break + else: + loss = exe.run(train_prog, + feed=feeder.feed(data), + fetch_list=[avg_loss]) iters += 1 - num_samples += len(data) + batch_id += 1 + # FIXME(wuyi): For use_reader_op, if the current + # pass is not the last, the last batch of this pass + # is also equal to args.batch_size. + if args.use_reader_op: + num_samples += args.batch_size * args.gpus + else: + num_samples += len(data) train_losses.append(loss) print("Pass: %d, Iter: %d, Loss: %f\n" % (pass_id, iters, np.mean(train_losses))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sec\n' % - (num_samples, train_elapsed, examples_per_sec)) - print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))) + print_train_time(start_time, time.time(), num_samples) + print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))), # evaluation - if not args.no_test and batch_acc != None: + if not args.no_test and batch_acc and not args.use_reader_op: + if args.use_inference_transpiler: + t = fluid.InferenceTranspiler() + t.transpile(infer_prog, place) + pass_test_acc = test(exe, infer_prog, test_reader, feeder, batch_acc) print(", Test Accuracy: %f" % pass_test_acc) @@ -251,14 +198,18 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, args, train_prog, startup_prog, nccl_id_var, num_trainers, trainer_id): - feed_var_list = [ - var for var in train_prog.global_block().vars.itervalues() - if var.is_data - ] + place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) + if not args.use_reader_op: + feed_var_list = [ + var for var in train_prog.global_block().vars.itervalues() + if var.is_data + ] + feeder = fluid.DataFeeder(feed_var_list, place) + # generate fake: if args.use_fake_data: for var in feed_var_list: - v = startup_prog.global_block().clone_variable(var) + v = startup_prog.global_block()._clone_variable(var) var.persistable = True v.persistable = True @@ -271,7 +222,6 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, "value": 1.0, "dtype": var.dtype}) - place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) if nccl_id_var and trainer_id == 0: #FIXME(wuyi): wait other trainer to start listening time.sleep(30) @@ -288,12 +238,21 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, num_trainers=num_trainers, trainer_id=trainer_id) - feeder = fluid.DataFeeder(feed_var_list, place) for pass_id in range(args.pass_num): num_samples = 0 iters = 0 start_time = time.time() - for batch_id, data in enumerate(train_reader()): + if not args.use_reader_op: + reader_generator = train_reader() + batch_id = 0 + data = None + while True: + if not args.use_reader_op: + data = next(reader_generator, None) + if data == None: + break + if iters == args.iterations: + break if args.profile and pass_id == 0 and batch_id == 5: profiler.start_profiler("All") elif args.profile and pass_id == 0 and batch_id == 10: @@ -302,47 +261,67 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, if iters == args.skip_batch_num: start_time = time.time() num_samples = 0 - if iters == args.iterations: - break - if args.use_fake_data: - loss, = exe.run([avg_loss.name]) + if args.use_fake_data or args.use_reader_op: + try: + loss, = exe.run([avg_loss.name]) + except fluid.core.EnforceNotMet as ex: + break else: loss, = exe.run([avg_loss.name], feed=feeder.feed(data)) - if args.update_method == "pserver": - exe.bcast_params() - num_samples += len(data) + if args.use_reader_op: + num_samples += args.batch_size * args.gpus + else: + num_samples += len(data) iters += 1 if batch_id % 1 == 0: print("Pass %d, batch %d, loss %s" % (pass_id, batch_id, np.array(loss))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - if not args.no_test and batch_acc != None: + batch_id += 1 + + print_train_time(start_time, time.time(), num_samples) + if not args.no_test and batch_acc and not args.use_reader_op: + # we have not implement record io for test + # skip test when use args.use_reader_op test_acc = test(startup_exe, infer_prog, test_reader, feeder, batch_acc) print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc)) - exit(0) def print_arguments(args): vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and vars(args)['device'] == 'GPU') - print('----------- resnet Configuration Arguments -----------') + print('----------- Configuration Arguments -----------') for arg, value in sorted(vars(args).iteritems()): print('%s: %s' % (arg, value)) print('------------------------------------------------') +def print_train_time(start_time, end_time, num_samples): + train_elapsed = end_time - start_time + examples_per_sec = num_samples / train_elapsed + print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % + (num_samples, train_elapsed, examples_per_sec)) + + +def print_paddle_envs(): + print('----------- Configuration envs -----------') + for k in os.environ: + if "PADDLE_" in k: + print "ENV %s:%s" % (k, os.environ[k]) + print('------------------------------------------------') + + def main(): args = parse_args() print_arguments(args) + print_paddle_envs() + if args.no_random: + fluid.default_startup_program().random_seed = 1 # the unique trainer id, starting from 0, needed by trainer # only nccl_id_var, num_trainers, trainer_id = ( - None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1"))) + None, 1, int(os.getenv("PADDLE_TRAINER_ID", "0"))) if args.use_cprof: pr = cProfile.Profile() @@ -356,7 +335,7 @@ def main(): fluid.memory_optimize(fluid.default_main_program()) if args.update_method == "pserver": - train_prog, startup_prog = dist_transpile(trainer_id) + train_prog, startup_prog = dist_transpile(trainer_id, args) if not train_prog: raise Exception( "Must configure correct environments to run dist train.") diff --git a/benchmark/fluid/kube_gen_job.py b/benchmark/fluid/kube_gen_job.py index 9da8a69af1d7b671b2648b1b3702776c1c0650b0..dfe8b5cdd58456902fa8ec355e9837dface3f7be 100644 --- a/benchmark/fluid/kube_gen_job.py +++ b/benchmark/fluid/kube_gen_job.py @@ -17,6 +17,7 @@ import copy import argparse import random import os +import copy from kube_templates import pserver, trainer, envs @@ -108,10 +109,9 @@ def gen_job(): tn_container["ports"][0]["containerPort"] = spreadport envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname}) - envs.append({"name": "TRAINERS", "value": str(args.trainers)}) - envs.append({"name": "PSERVERS", "value": str(args.pservers)}) + envs.append({"name": "PADDLE_TRAINERS", "value": str(args.trainers)}) + envs.append({"name": "PADDLE_PSERVERS", "value": str(args.pservers)}) envs.append({"name": "ENTRY", "value": args.entry}) - envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)}) envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)}) # NOTE: these directories below are cluster specific, please modify # this settings before you run on your own cluster. @@ -166,17 +166,23 @@ def gen_job(): tn["spec"]["template"]["spec"]["volumes"] = volumes tn_container["volumeMounts"] = volumeMounts - ps_container["env"] = envs - ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"}) + ps_container["env"] = copy.deepcopy(envs) + ps_container["env"].append({ + "name": "PADDLE_TRAINING_ROLE", + "value": "PSERVER" + }) tn_container["env"] = envs if args.disttype == "pserver": tn_container["env"].append({ - "name": "TRAINING_ROLE", + "name": "PADDLE_TRAINING_ROLE", "value": "TRAINER" }) elif args.disttype == "nccl2" or args.disttype == "local": # NCCL2 have no training role, set to plain WORKER - tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"}) + tn_container["env"].append({ + "name": "PADDLE_TRAINING_ROLE", + "value": "WORKER" + }) os.mkdir(args.jobname) if args.disttype == "pserver": diff --git a/benchmark/fluid/models/machine_translation.py b/benchmark/fluid/models/machine_translation.py index 635b3373dd27b21f83afae10b1d24833b81d57eb..17f6b03826ae818a3671ea7f9355a8e8c04b50be 100644 --- a/benchmark/fluid/models/machine_translation.py +++ b/benchmark/fluid/models/machine_translation.py @@ -173,21 +173,6 @@ def seq_to_seq_net(embedding_dim, encoder_size, decoder_size, source_dict_dim, return avg_cost, feeding_list -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - lod_t = core.LoDTensor() - lod_t.set(flattened_data, place) - lod_t.set_lod([lod]) - return lod_t, lod[-1] - - def lodtensor_to_ndarray(lod_tensor): dims = lod_tensor.get_dims() ndarray = np.zeros(shape=dims).astype('float32') @@ -197,6 +182,8 @@ def lodtensor_to_ndarray(lod_tensor): def get_model(args): + if args.use_reader_op: + raise Exception("machine_translation do not support reader op for now.") embedding_dim = 512 encoder_size = 512 decoder_size = 512 @@ -221,7 +208,7 @@ def get_model(args): train_batch_generator = paddle.batch( paddle.reader.shuffle( paddle.dataset.wmt14.train(dict_size), buf_size=1000), - batch_size=args.batch_size) + batch_size=args.batch_size * args.gpus) test_batch_generator = paddle.batch( paddle.reader.shuffle( diff --git a/benchmark/fluid/models/mnist.py b/benchmark/fluid/models/mnist.py index d264bfc12bdb159c06dae81db4949b9ee17268e2..8e740dc6896b7eeeb82170aa13d32987c4df5c48 100644 --- a/benchmark/fluid/models/mnist.py +++ b/benchmark/fluid/models/mnist.py @@ -20,6 +20,7 @@ import numpy as np import argparse import time import cProfile +import os import paddle import paddle.fluid as fluid @@ -65,19 +66,49 @@ def cnn_model(data): def get_model(args): - # Input data - images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - - # Train program - predict = cnn_model(images) - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - # Evaluator - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) + if args.use_reader_op: + filelist = [ + os.path.join(args.data_path, f) for f in os.listdir(args.data_path) + ] + data_file = fluid.layers.open_files( + filenames=filelist, + shapes=[[-1, 1, 28, 28], (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + thread_num=args.gpus, + pass_num=args.pass_num) + data_file = fluid.layers.double_buffer( + fluid.layers.batch( + data_file, batch_size=args.batch_size)) + images, label = fluid.layers.read_file(data_file) + else: + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + if args.device == 'CPU' and args.cpus > 1: + places = fluid.layers.get_places(args.cpus) + pd = fluid.layers.ParallelDo(places) + with pd.do(): + predict = cnn_model(pd.read_input(images)) + label = pd.read_input(label) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + batch_acc = fluid.layers.accuracy(input=predict, label=label) + + pd.write_output(avg_cost) + pd.write_output(batch_acc) + + avg_cost, batch_acc = pd() + avg_cost = fluid.layers.mean(avg_cost) + batch_acc = fluid.layers.mean(batch_acc) + else: + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_acc = fluid.layers.accuracy(input=predict, label=label) # inference program inference_program = fluid.default_main_program().clone() @@ -88,7 +119,7 @@ def get_model(args): # Reader train_reader = paddle.batch( - paddle.dataset.mnist.train(), batch_size=args.batch_size) + paddle.dataset.mnist.train(), batch_size=args.batch_size * args.gpus) test_reader = paddle.batch( paddle.dataset.mnist.test(), batch_size=args.batch_size) return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc diff --git a/benchmark/fluid/models/resnet.py b/benchmark/fluid/models/resnet.py index 9dec8911ed64e09285fb461c4a12adb601535316..d44a9c07d31cfae9d54ad5949b85c77e60eae258 100644 --- a/benchmark/fluid/models/resnet.py +++ b/benchmark/fluid/models/resnet.py @@ -19,6 +19,7 @@ from __future__ import print_function import functools import numpy as np import time +import os import cProfile, pstats, StringIO @@ -26,6 +27,7 @@ import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.profiler as profiler +from recordio_converter import imagenet_train, imagenet_test def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): @@ -122,40 +124,85 @@ def get_model(args): else: dshape = [32, 32, 3] model = resnet_cifar10 - else: + train_reader = paddle.dataset.cifar.train10() + test_reader = paddle.dataset.cifar.test10() + elif args.data_set == "flowers": class_dim = 102 if args.data_format == 'NCHW': dshape = [3, 224, 224] else: dshape = [224, 224, 3] model = resnet_imagenet - - input = fluid.layers.data(name='data', shape=dshape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - predict = model(input, class_dim) - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) + train_reader = paddle.dataset.flowers.train() + test_reader = paddle.dataset.flowers.test() + elif args.data_set == "imagenet": + class_dim = 1000 + if args.data_format == 'NCHW': + dshape = [3, 224, 224] + else: + dshape = [224, 224, 3] + model = resnet_imagenet + if not args.data_path: + raise Exception( + "Must specify --data_path when training with imagenet") + train_reader = imagenet_train(args.data_path) + test_reader = imagenet_test(args.data_path) + + if args.use_reader_op: + filelist = [ + os.path.join(args.data_path, f) for f in os.listdir(args.data_path) + ] + data_file = fluid.layers.open_files( + filenames=filelist, + shapes=[[-1] + dshape, (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + thread_num=args.gpus, + pass_num=args.pass_num) + data_file = fluid.layers.double_buffer( + fluid.layers.batch( + data_file, batch_size=args.batch_size)) + input, label = fluid.layers.read_file(data_file) + else: + input = fluid.layers.data(name='data', shape=dshape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + if args.device == 'CPU' and args.cpus > 1: + places = fluid.layers.get_places(args.cpus) + pd = fluid.layers.ParallelDo(places) + with pd.do(): + predict = model(pd.read_input(input), class_dim) + label = pd.read_input(label) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + batch_acc = fluid.layers.accuracy(input=predict, label=label) + + pd.write_output(avg_cost) + pd.write_output(batch_acc) + + avg_cost, batch_acc = pd() + avg_cost = fluid.layers.mean(avg_cost) + batch_acc = fluid.layers.mean(batch_acc) + else: + predict = model(input, class_dim) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + batch_acc = fluid.layers.accuracy(input=predict, label=label) inference_program = fluid.default_main_program().clone() with fluid.program_guard(inference_program): inference_program = fluid.io.get_inference_program( - target_vars=[batch_acc, batch_size_tensor]) + target_vars=[batch_acc]) optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9) - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.train10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.dataset.cifar.test10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), - batch_size=args.batch_size) - - return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc + batched_train_reader = paddle.batch( + train_reader if args.no_random else paddle.reader.shuffle( + train_reader, buf_size=5120), + batch_size=args.batch_size * args.gpus, + drop_last=True) + batched_test_reader = paddle.batch( + test_reader, batch_size=args.batch_size, drop_last=True) + + return avg_cost, inference_program, optimizer, batched_train_reader,\ + batched_test_reader, batch_acc diff --git a/benchmark/fluid/models/stacked_dynamic_lstm.py b/benchmark/fluid/models/stacked_dynamic_lstm.py index 81a28b5f3aed0c325398b909d700c23df545824a..3231542a17ace99a17c9f9b9bdb3c2527637d9ef 100644 --- a/benchmark/fluid/models/stacked_dynamic_lstm.py +++ b/benchmark/fluid/models/stacked_dynamic_lstm.py @@ -44,6 +44,9 @@ def crop_sentence(reader, crop_size): def get_model(args): + if args.use_reader_op: + raise Exception( + "stacked_dynamic_lstm do not support reader op for now.") lstm_size = 512 emb_dim = 512 crop_size = 1500 @@ -115,25 +118,10 @@ def get_model(args): train_reader = batch( paddle.reader.shuffle( crop_sentence(imdb.train(word_dict), crop_size), buf_size=25000), - batch_size=args.batch_size) + batch_size=args.batch_size * args.gpus) test_reader = batch( paddle.reader.shuffle( crop_sentence(imdb.test(word_dict), crop_size), buf_size=25000), batch_size=args.batch_size) return loss, inference_program, adam, train_reader, test_reader, batch_acc - - -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = numpy.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - res = fluid.LoDTensor() - res.set(flattened_data, place) - res.set_lod([lod]) - return res diff --git a/benchmark/fluid/models/vgg.py b/benchmark/fluid/models/vgg.py index 53856c5f7acd3a4e1476ec57154a880bb6f984c9..932601302d2f5d56b53e3462af886429034d8989 100644 --- a/benchmark/fluid/models/vgg.py +++ b/benchmark/fluid/models/vgg.py @@ -22,6 +22,7 @@ import paddle.fluid as fluid import paddle.fluid.core as core import argparse import functools +import os def vgg16_bn_drop(input): @@ -65,9 +66,25 @@ def get_model(args): else: data_shape = [224, 224, 3] - # Input data - images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') + if args.use_reader_op: + filelist = [ + os.path.join(args.data_path, f) for f in os.listdir(args.data_path) + ] + data_file = fluid.layers.open_files( + filenames=filelist, + shapes=[[-1] + data_shape, (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + thread_num=args.gpus, + pass_num=args.pass_num) + data_file = fluid.layers.double_buffer( + fluid.layers.batch( + data_file, batch_size=args.batch_size)) + images, label = fluid.layers.read_file(data_file) + else: + images = fluid.layers.data( + name='data', shape=data_shape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') # Train program net = vgg16_bn_drop(images) @@ -95,7 +112,7 @@ def get_model(args): paddle.dataset.cifar.train10() if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), buf_size=5120), - batch_size=args.batch_size) + batch_size=args.batch_size * args.gpus) test_reader = paddle.batch( paddle.dataset.cifar.test10() if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), diff --git a/benchmark/fluid/recordio_converter.py b/benchmark/fluid/recordio_converter.py new file mode 100644 index 0000000000000000000000000000000000000000..f2dc39109bf1beaf147b046560c92fbd2416d8e6 --- /dev/null +++ b/benchmark/fluid/recordio_converter.py @@ -0,0 +1,164 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import random +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.dataset import mnist, cifar, flowers, image + + +def convert_2_recordio(py_reader, outfilepath, batch_size, shape_data, + shape_label): + num_batches = 0 + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(py_reader(), batch_size=batch_size) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=shape_data), + fluid.layers.data( + name='label', shape=shape_label, dtype='int64'), + ], + place=fluid.CPUPlace()) + num_batches = fluid.recordio_writer.convert_reader_to_recordio_file( + outfilepath, reader, feeder) + return num_batches + + +def prepare_mnist(outpath, batch_size): + outfilepath = os.path.join(outpath, "mnist.recordio") + convert_2_recordio(mnist.train, outfilepath, batch_size, [784], [1]) + + +def prepare_cifar10(outpath, batch_size): + outfilepath = os.path.join(outpath, "cifar.recordio") + convert_2_recordio(cifar.train10, outfilepath, batch_size, [3, 32, 32], [1]) + + +def prepare_flowers(outpath, batch_size): + outfilepath = os.path.join(outpath, "flowers.recordio") + convert_2_recordio(flowers.train, outfilepath, batch_size, [3, 224, 224], + [1]) + + +def default_mapper(sample): + img, label = sample + img = image.simple_transform( + img, 256, 224, True, mean=[103.94, 116.78, 123.68]) + return img.flatten().astype('float32'), label + + +def imagenet_train(data_dir): + contents = os.listdir(data_dir) + if set(contents) != set( + ["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]): + raise Exception("Imagenet data contents error!") + img2label = dict() + imgfilelist = [] + with open(os.path.join(data_dir, "train.txt")) as fn: + while 1: + l = fn.readline() + if not l: + break + img, lbl = l[:-1].split(" ") + img2label[img] = int(lbl) + imgfilelist.append(img) + # shuffle all, this is slow + random.shuffle(imgfilelist) + + def train_reader(): + for idx, imgfile in enumerate(imgfilelist): + data = image.load_image( + os.path.join(data_dir, "train", imgfile.lower())) + label = [img2label[imgfile], ] + yield [data, label] + + return paddle.reader.map_readers(default_mapper, train_reader) + + +def imagenet_test(data_dir): + contents = os.listdir(data_dir) + if set(contents) != set( + ["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]): + raise Exception("Imagenet data contents error!") + img2label = dict() + imgfilelist = [] + with open(os.path.join(data_dir, "val.txt")) as fn: + while 1: + l = fn.readline() + if not l: + break + img, lbl = l[:-1].split(" ") + img2label[img] = int(lbl) + imgfilelist.append(img) + + def test_reader(): + for idx, imgfile in enumerate(imgfilelist): + base_path = os.path.join(data_dir, "val", imgfile.split(".")[0]) + image_path = ".".join([base_path, "jpeg"]) + data = image.load_image(image_path) + label = [img2label[imgfile], ] + yield [data, label] + + return paddle.reader.map_readers(default_mapper, test_reader) + + +# FIXME(wuyi): delete this when https://github.com/PaddlePaddle/Paddle/pull/11066 is merged +def convert_reader_to_recordio_files( + filename, + batch_per_file, + reader_creator, + feeder, + compressor=core.RecordIOWriter.Compressor.Snappy, + max_num_records=1000, + feed_order=None): + if feed_order is None: + feed_order = feeder.feed_names + f_name, f_ext = os.path.splitext(filename) + assert (f_ext == ".recordio") + + lines = [] + f_idx = 0 + counter = 0 + for idx, batch in enumerate(reader_creator()): + lines.append(batch) + if idx >= batch_per_file and idx % batch_per_file == 0: + filename = "%s-%05d%s" % (f_name, f_idx, f_ext) + with fluid.recordio_writer.create_recordio_writer( + filename, compressor, max_num_records) as writer: + for l in lines: + res = feeder.feed(l) + for each in feed_order: + writer.append_tensor(res[each]) + writer.complete_append_tensor() + counter += 1 + lines = [] + f_idx += 1 + print("written file: ", filename) + return counter + + +def prepare_imagenet(inpath, outpath, batch_size): + r = paddle.batch(imagenet_train(inpath), batch_size=batch_size) + feeder = fluid.DataFeeder( + feed_list=[ + fluid.layers.data( + name="image", shape=[3, 224, 224]), fluid.layers.data( + name="label", shape=[1], dtype='int64') + ], + place=fluid.CPUPlace()) + outpath = os.path.join(outpath, "imagenet.recordio") + convert_reader_to_recordio_files(outpath, 10000, r, feeder) diff --git a/benchmark/fluid/run.sh b/benchmark/fluid/run.sh index f6dfd20bf2ee0b668b6d4238d4511253b2233035..5d9b2db87135e53470b106dcd11a6bcfdc5dbda9 100644 --- a/benchmark/fluid/run.sh +++ b/benchmark/fluid/run.sh @@ -2,6 +2,7 @@ # This script benchmarking the PaddlePaddle Fluid on # single thread single GPU. +mkdir -p logs #export FLAGS_fraction_of_gpu_memory_to_use=0.0 export CUDNN_PATH=/paddle/cudnn_v5 @@ -35,71 +36,74 @@ nohup stdbuf -oL nvidia-smi \ --format=csv \ --filename=mem.log \ -l 1 & + # mnist # mnist gpu mnist 128 -FLAGS_benchmark=true stdbuf -oL python fluid/mnist.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=mnist \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ --iterations=500 \ - 2>&1 | tee -a mnist_gpu_128.log + 2>&1 | tee -a logs/mnist_gpu_128.log # vgg16 # gpu cifar10 128 -FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=vgg16 \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a vgg16_gpu_128.log + 2>&1 | tee -a logs/vgg16_gpu_128.log # flowers gpu 128 -FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=vgg16 \ --device=GPU \ --batch_size=32 \ --data_set=flowers \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a vgg16_gpu_flowers_32.log + 2>&1 | tee -a logs/vgg16_gpu_flowers_32.log # resnet50 # resnet50 gpu cifar10 128 -FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=resnet \ --device=GPU \ --batch_size=128 \ --data_set=cifar10 \ - --model=resnet_cifar10 \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a resnet50_gpu_128.log + 2>&1 | tee -a logs/resnet50_gpu_128.log # resnet50 gpu flowers 64 -FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=resnet \ --device=GPU \ --batch_size=64 \ --data_set=flowers \ - --model=resnet_imagenet \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a resnet50_gpu_flowers_64.log + 2>&1 | tee -a logs/resnet50_gpu_flowers_64.log # lstm # lstm gpu imdb 32 # tensorflow only support batch=32 -FLAGS_benchmark=true stdbuf -oL python fluid/stacked_dynamic_lstm.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=stacked_dynamic_lstm \ --device=GPU \ --batch_size=32 \ --skip_batch_num=5 \ --iterations=30 \ - --hidden_dim=512 \ - --emb_dim=512 \ - --crop_size=1500 \ - 2>&1 | tee -a lstm_gpu_32.log + 2>&1 | tee -a logs/lstm_gpu_32.log # seq2seq # seq2seq gpu wmb 128 -FLAGS_benchmark=true stdbuf -oL python fluid/machine_translation.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=machine_translation \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a lstm_gpu_128.log + 2>&1 | tee -a logs/lstm_gpu_128.log diff --git a/benchmark/fluid/run_fluid_benchmark.sh b/benchmark/fluid/run_fluid_benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..4309a3126c1d72fe1eb2d5ec423075aea4d3ec88 --- /dev/null +++ b/benchmark/fluid/run_fluid_benchmark.sh @@ -0,0 +1,9 @@ +#!/bin/bash + +PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model resnet --device CPU --update_method pserver --iterations=10000 & + +sleep 15 + +CUDA_VISIBLE_DEVICES=0,1 PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model resnet --device GPU --update_method pserver --iterations=10000 --gpus 2 & + +CUDA_VISIBLE_DEVICES=2,3 PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=1 python fluid_benchmark.py --model resnet --device GPU --update_method pserver --iterations=10000 --gpus 2 & diff --git a/benchmark/paddle/image/run.sh b/benchmark/paddle/image/run.sh index 717ed487ba7657db6535efcb1128a355a0f15eaf..5b58a8d773aab795e5439b0f0e5d81bec66b5f56 100755 --- a/benchmark/paddle/image/run.sh +++ b/benchmark/paddle/image/run.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function train() { diff --git a/benchmark/paddle/image/run_mkl_infer.sh b/benchmark/paddle/image/run_mkl_infer.sh index 62c9bf6efd3810f506fd4592b2ba3a21b1b7f0e7..0fad5e04cc992a3ec97591d3833957bb7517a8f3 100755 --- a/benchmark/paddle/image/run_mkl_infer.sh +++ b/benchmark/paddle/image/run_mkl_infer.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function clock_to_seconds() { diff --git a/benchmark/paddle/image/run_mkl_train.sh b/benchmark/paddle/image/run_mkl_train.sh index 03d2d378fb72e36f765d89af788f6ee96fe21d4e..1583bf134a276a08aa2f8e84dc63adbb205a83d6 100755 --- a/benchmark/paddle/image/run_mkl_train.sh +++ b/benchmark/paddle/image/run_mkl_train.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function train() { diff --git a/benchmark/paddle/image/run_openblas_infer.sh b/benchmark/paddle/image/run_openblas_infer.sh index a9a7b8a66717c4be0543c3fe2db293fe199e3dc4..987381cabc2e793886099212660723c122b73bb0 100755 --- a/benchmark/paddle/image/run_openblas_infer.sh +++ b/benchmark/paddle/image/run_openblas_infer.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function clock_to_seconds() { diff --git a/benchmark/paddle/image/run_openblas_train.sh b/benchmark/paddle/image/run_openblas_train.sh index 935cff6f2c97d25d6de556cfee25e27dbe49b5b6..cc64e1d09da02087b1737190a0b75dc7758600a6 100755 --- a/benchmark/paddle/image/run_openblas_train.sh +++ b/benchmark/paddle/image/run_openblas_train.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function train() { diff --git a/benchmark/paddle/rnn/run.sh b/benchmark/paddle/rnn/run.sh index e9dfeb2e525979f47e4ef48f7610dc1007900f2c..f99a562b3f88a98560f4bf7aee98ceee9daefe67 100755 --- a/benchmark/paddle/rnn/run.sh +++ b/benchmark/paddle/rnn/run.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function train() { diff --git a/benchmark/tensorflow/image/run.sh b/benchmark/tensorflow/image/run.sh index eade36beb9df5f8d3978939216e058203e024c1a..cf894fe3f2dca24e3acf863d625b3a7008793b83 100755 --- a/benchmark/tensorflow/image/run.sh +++ b/benchmark/tensorflow/image/run.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function test() { diff --git a/benchmark/tensorflow/image/run_multi.sh b/benchmark/tensorflow/image/run_multi.sh index 69faa4331744f2276e7706185ae10bc507f95764..bf1435bc55b90669e0b8bd893b8ed7bbb99d51e2 100755 --- a/benchmark/tensorflow/image/run_multi.sh +++ b/benchmark/tensorflow/image/run_multi.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function test() { diff --git a/benchmark/tensorflow/rnn/run.sh b/benchmark/tensorflow/rnn/run.sh index bb4c69cb95f965eff35f1c5a60376bf1e84f841b..db10eefdea8676ad34fb84a161f0fc1309147824 100755 --- a/benchmark/tensorflow/rnn/run.sh +++ b/benchmark/tensorflow/rnn/run.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function test() { diff --git a/benchmark/tensorflow/rnn/run_multi.sh b/benchmark/tensorflow/rnn/run_multi.sh index c2d7dd597e6da54cd5c4cda311fbbd18486b4647..ec62fc26b51543f2f8ddfc5e73aa6ff7d611e4dd 100755 --- a/benchmark/tensorflow/rnn/run_multi.sh +++ b/benchmark/tensorflow/rnn/run_multi.sh @@ -1,3 +1,5 @@ +#!/bin/bash + set -e function test() { diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake index e3b9d94215a858c5c9a34e1b7e97540f1876801d..6ed51c648478efb9784d0c43b169c285e740e0f3 100644 --- a/cmake/cblas.cmake +++ b/cmake/cblas.cmake @@ -83,18 +83,20 @@ else() set(REFERENCE_CBLAS_LIB_SEARCH_PATHS ${REFERENCE_CBLAS_ROOT}/lib) endif() -find_path(REFERENCE_CBLAS_INCLUDE_DIR NAMES cblas.h PATHS +if(WITH_SYSTEM_BLAS) + find_path(REFERENCE_CBLAS_INCLUDE_DIR NAMES cblas.h PATHS ${REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS}) -find_library(REFERENCE_CBLAS_LIBRARY NAMES cblas PATHS + find_library(REFERENCE_CBLAS_LIBRARY NAMES cblas PATHS ${REFERENCE_CBLAS_LIB_SEARCH_PATHS}) -if(REFERENCE_CBLAS_INCLUDE_DIR AND REFERENCE_CBLAS_LIBRARY) - set(CBLAS_FOUND ON) - set(CBLAS_PROVIDER REFERENCE) - set(CBLAS_INC_DIR ${REFERENCE_CBLAS_INCLUDE_DIR}) - set(CBLAS_LIBRARIES ${REFERENCE_CBLAS_LIBRARY}) - add_definitions(-DPADDLE_USE_REFERENCE_CBLAS) - message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})") + if(REFERENCE_CBLAS_INCLUDE_DIR AND REFERENCE_CBLAS_LIBRARY) + set(CBLAS_FOUND ON) + set(CBLAS_PROVIDER REFERENCE) + set(CBLAS_INC_DIR ${REFERENCE_CBLAS_INCLUDE_DIR}) + set(CBLAS_LIBRARIES ${REFERENCE_CBLAS_LIBRARY}) + add_definitions(-DPADDLE_USE_REFERENCE_CBLAS) + message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})") + endif() endif() if(IOS_USE_VECLIB_FOR_BLAS AND VECLIB_FOUND) diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 682614742cf1bd3130c638020a2545e16226d4d6..c35096e09b5685ee30f20648e7dd461f71e0b1c4 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -92,8 +92,19 @@ if(WITH_GPU) if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7) message(FATAL_ERROR "TensorRT needs CUDNN >= 7.0 to compile") endif() + if(${TENSORRT_MAJOR_VERSION} VERSION_LESS 4) + message(FATAL_ERROR "Paddle needs TensorRT >= 4.0 to compile") + endif() include_directories(${TENSORRT_INCLUDE_DIR}) endif() + if(WITH_ANAKIN) + if(${CUDA_VERSION_MAJOR} VERSION_LESS 8) + message(FATAL_ERROR "Anakin needs CUDA >= 8.0 to compile") + endif() + if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7) + message(FATAL_ERROR "Anakin needs CUDNN >= 7.0 to compile") + endif() + endif() elseif(WITH_AMD_GPU) add_definitions(-DPADDLE_WITH_HIP) set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__") @@ -115,6 +126,10 @@ endif() set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}") +if(WITH_DISTRIBUTE) + add_definitions(-DPADDLE_WITH_DISTRIBUTE) +endif() + if(WITH_GOLANG) # we need to symlink Paddle directory into GOPATH. If we # don't do it and we have code that depends on Paddle, go @@ -163,3 +178,11 @@ if(WITH_GOLANG) endif() endif(WITH_GOLANG) + +if(WITH_GRPC) + add_definitions(-DPADDLE_WITH_GRPC) +endif(WITH_GRPC) + +if(WITH_BRPC_RDMA) + add_definitions(-DPADDLE_WITH_BRPC_RDMA) +endif(WITH_BRPC_RDMA) diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake index 2c84061ff572de4687b4d496f8ded6deee8d1011..9eebea816cbfc91052c95ecf99ecc4b0bea4e4c2 100644 --- a/cmake/cudnn.cmake +++ b/cmake/cudnn.cmake @@ -21,6 +21,7 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS ${CUDNN_ROOT}/lib64 ${CUDNN_ROOT}/lib ${CUDNN_ROOT}/lib/${TARGET_ARCH}-linux-gnu + ${CUDNN_ROOT}/local/cuda-${CUDA_VERSION}/targets/${TARGET_ARCH}-linux/lib/ $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/lib64 $ENV{CUDNN_ROOT}/lib diff --git a/cmake/external/anakin.cmake b/cmake/external/anakin.cmake new file mode 100644 index 0000000000000000000000000000000000000000..403873a51017b9bb7c888bed77d89ca9b15b68d2 --- /dev/null +++ b/cmake/external/anakin.cmake @@ -0,0 +1,64 @@ +if (NOT WITH_ANAKIN) + return() +endif() + +INCLUDE(ExternalProject) +set(ANAKIN_SOURCE_DIR ${THIRD_PARTY_PATH}/anakin) +# the anakin install dir is only default one now +set(ANAKIN_INSTALL_DIR ${THIRD_PARTY_PATH}/anakin/src/extern_anakin/output) +set(ANAKIN_INCLUDE ${ANAKIN_INSTALL_DIR}) +set(ANAKIN_LIBRARY ${ANAKIN_INSTALL_DIR}) +set(ANAKIN_SHARED_LIB ${ANAKIN_LIBRARY}/libanakin.so) +set(ANAKIN_SABER_LIB ${ANAKIN_LIBRARY}/libanakin_saber_common.so) + +# TODO(luotao): ANAKIN_MODLE_URL will move to demo ci later. +set(ANAKIN_MODLE_URL "http://paddle-inference-dist.bj.bcebos.com/mobilenet_v2.anakin.bin") +execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_SOURCE_DIR}") +execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL}") + +include_directories(${ANAKIN_INCLUDE}) +include_directories(${ANAKIN_INCLUDE}/saber/) + +set(ANAKIN_COMPILE_EXTRA_FLAGS + -Wno-error=unused-but-set-variable -Wno-unused-but-set-variable + -Wno-error=unused-variable -Wno-unused-variable + -Wno-error=format-extra-args -Wno-format-extra-args + -Wno-error=comment -Wno-comment + -Wno-error=format -Wno-format + -Wno-error=switch -Wno-switch + -Wno-error=return-type -Wno-return-type + -Wno-error=non-virtual-dtor -Wno-non-virtual-dtor + -Wno-sign-compare + -Wno-reorder + -Wno-error=cpp) + +ExternalProject_Add( + extern_anakin + ${EXTERNAL_PROJECT_LOG_ARGS} + # TODO(luotao): use PaddlePaddle/Anakin later + GIT_REPOSITORY "https://github.com/luotao1/Anakin" + GIT_TAG "3957ae9263eaa0b1986758dac60a88852afb09be" + PREFIX ${ANAKIN_SOURCE_DIR} + UPDATE_COMMAND "" + CMAKE_ARGS -DUSE_GPU_PLACE=YES + -DUSE_X86_PLACE=YES + -DBUILD_WITH_UNIT_TEST=NO + -DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf + -DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml + -DCUDNN_ROOT=${CUDNN_ROOT} + ${EXTERNAL_OPTIONAL_ARGS} + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR} +) + +message(STATUS "Anakin for inference is enabled") +message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}") + +add_library(anakin_shared SHARED IMPORTED GLOBAL) +set_property(TARGET anakin_shared PROPERTY IMPORTED_LOCATION ${ANAKIN_SHARED_LIB}) +add_dependencies(anakin_shared extern_anakin protobuf mklml) + +add_library(anakin_saber SHARED IMPORTED GLOBAL) +set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB}) +add_dependencies(anakin_saber extern_anakin protobuf mklml) + +list(APPEND external_project_dependencies anakin_shared anakin_saber) diff --git a/cmake/external/brpc.cmake b/cmake/external/brpc.cmake new file mode 100644 index 0000000000000000000000000000000000000000..30b227b6452abf44171a1a4e04569e66b16e67a4 --- /dev/null +++ b/cmake/external/brpc.cmake @@ -0,0 +1,69 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +INCLUDE(ExternalProject) + +find_library(SSL_LIBRARY NAMES ssl) +ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY}) + +find_library(CRYPTO_LIBRARY NAMES crypto) +ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY}) + + +SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc) +SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc) +SET(BRPC_INCLUDE_DIR "${BRPC_INSTALL_DIR}/include" CACHE PATH "brpc include directory." FORCE) +SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc library." FORCE) + +INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR}) + +# Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args +set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib") + +# If minimal .a is need, you can set WITH_DEBUG_SYMBOLS=OFF +ExternalProject_Add( + extern_brpc + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/gongweibao/brpc" + GIT_TAG "7dc04defad1fd4173aae170c3fcbde131b65155a" + PREFIX ${BRPC_SOURCES_DIR} + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_INSTALL_PREFIX=${BRPC_INSTALL_DIR} + -DCMAKE_INSTALL_LIBDIR=${BRPC_INSTALL_DIR}/lib + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE} + -DCMAKE_PREFIX_PATH=${prefix_path} + -DBRPC_WITH_GLOG=ON + -DIOBUF_WITH_HUGE_BLOCK=ON + -DBRPC_WITH_RDMA=${WITH_BRPC_RDMA} + ${EXTERNAL_OPTIONAL_ARGS} + LIST_SEPARATOR | + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${BRPC_INSTALL_DIR} + -DCMAKE_INSTALL_LIBDIR:PATH=${BRPC_INSTALL_DIR}/lib + -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON + -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} +) +ADD_DEPENDENCIES(extern_brpc protobuf ssl crypto leveldb gflags glog gtest snappy) +ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES}) +ADD_DEPENDENCIES(brpc extern_brpc) + + +LIST(APPEND external_project_dependencies brpc) diff --git a/cmake/external/cub.cmake b/cmake/external/cub.cmake new file mode 100644 index 0000000000000000000000000000000000000000..c94849cf4b96746e6c507db2a6310c2f305dacf5 --- /dev/null +++ b/cmake/external/cub.cmake @@ -0,0 +1,35 @@ +if(NOT WITH_GPU) + return() +endif() + +include(ExternalProject) + +set(CUB_SOURCE_DIR ${THIRD_PARTY_PATH}/cub) +set(CUB_INCLUDE_DIR ${CUB_SOURCE_DIR}/src/extern_cub) + +include_directories(${CUB_INCLUDE_DIR}) + +ExternalProject_Add( + extern_cub + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/NVlabs/cub.git" + GIT_TAG "v1.8.0" + PREFIX ${CUB_SOURCE_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + INSTALL_COMMAND "" + TEST_COMMAND "" +) + +if(${CMAKE_VERSION} VERSION_LESS "3.3.0") + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cub_dummy.c) + file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";") + add_library(cub STATIC ${dummyfile}) +else() + add_library(cub INTERFACE) +endif() + +add_dependencies(cub extern_cub) + +LIST(APPEND externl_project_dependencies cub) diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake index 9459f1ddfe85f5607880d3fdd968b494d6af592a..7fb67afbe15a5a019c978092d5ba3a4a0f66d996 100644 --- a/cmake/external/grpc.cmake +++ b/cmake/external/grpc.cmake @@ -33,14 +33,24 @@ ELSE() SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin) ENDIF() +# FIXME(wuyi): do not build zlib cares protobuf twice, find a way to build grpc with them ExternalProject_Add( extern_grpc DEPENDS protobuf zlib - URL "http://paddlepaddledeps.bj.bcebos.com/grpc.tar.xz" + # NOTE(wuyi): + # this package is generated by following steps: + # 1. git clone -b v1.8.x https://github.com/grpc/grpc.git + # 2. git submodule update --init + # 3. keep only zlib, cares, protobuf, boringssl under "third_party", + # checkout and clean other dirs under third_party + # 4. remove .git, and package the directory. + URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz" + URL_MD5 "1f268a2aff6759839dccd256adcc91cf" PREFIX ${GRPC_SOURCES_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" BUILD_IN_SOURCE 1 + PATCH_COMMAND cp ${PADDLE_SOURCE_DIR}/patches/grpc/grpc_library.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/grpc_library.h && cp ${PADDLE_SOURCE_DIR}/patches/grpc/completion_queue.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/completion_queue.h # NOTE(yuyang18): # Disable -Werror, otherwise the compile will fail in MacOS. # It seems that we cannot configure that by make command. @@ -49,7 +59,6 @@ ExternalProject_Add( INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install ) -# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them. ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a") diff --git a/cmake/external/leveldb.cmake b/cmake/external/leveldb.cmake new file mode 100644 index 0000000000000000000000000000000000000000..fb5091731da02b497a14f119e944905eee4979d5 --- /dev/null +++ b/cmake/external/leveldb.cmake @@ -0,0 +1,44 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +INCLUDE(ExternalProject) + +SET(LEVELDB_SOURCES_DIR ${THIRD_PARTY_PATH}/leveldb) +SET(LEVELDB_INSTALL_DIR ${THIRD_PARTY_PATH}/install/leveldb) +SET(LEVELDB_INCLUDE_DIR "${LEVELDB_INSTALL_DIR}/include" CACHE PATH "leveldb include directory." FORCE) +SET(LEVELDB_LIBRARIES "${LEVELDB_INSTALL_DIR}/lib/libleveldb.a" CACHE FILEPATH "leveldb library." FORCE) +INCLUDE_DIRECTORIES(${LEVELDB_INCLUDE_DIR}) + +ExternalProject_Add( + extern_leveldb + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${LEVELDB_SOURCES_DIR} + URL "https://github.com/google/leveldb/archive/v1.18.tar.gz" + URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0" + CONFIGURE_COMMAND "" + BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a + INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/ + && cp ${LEVELDB_SOURCES_DIR}/src/extern_leveldb/libleveldb.a ${LEVELDB_LIBRARIES} + && cp -r ${LEVELDB_SOURCES_DIR}/src/extern_leveldb/include ${LEVELDB_INSTALL_DIR}/ + BUILD_IN_SOURCE 1 +) + +ADD_DEPENDENCIES(extern_leveldb snappy) + +ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES}) +ADD_DEPENDENCIES(leveldb extern_leveldb) + +LIST(APPEND external_project_dependencies leveldb) + diff --git a/cmake/external/libxsmm.cmake b/cmake/external/libxsmm.cmake new file mode 100644 index 0000000000000000000000000000000000000000..530f7ebe2813fb2f00c6b5b4d1f7b2f04fe650b0 --- /dev/null +++ b/cmake/external/libxsmm.cmake @@ -0,0 +1,57 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +OPTION(WITH_LIBXSMM "Compile with libxsmm" OFF) + +IF(NOT WITH_LIBXSMM) + return() +ENDIF() + +IF(WIN32 OR APPLE OR ANDROID OR IOS) + MESSAGE(WARNING "Windows, Mac or Mobile are not supported with libxsmm in Paddle yet.") + SET(WITH_LIBXSMM OFF CACHE STRING "Disable LIBXSMM" FORCE) + return() +ENDIF() + +INCLUDE (ExternalProject) + +SET(LIBXSMM_SOURCES_DIR ${THIRD_PARTY_PATH}/libxsmm) +SET(LIBXSMM_INSTALL_DIR ${THIRD_PARTY_PATH}/install/libxsmm) +SET(LIBXSMM_INCLUDE_DIR "${LIBXSMM_INSTALL_DIR}/include" CACHE PATH "LIBXSMM include directory." FORCE) +SET(LIBXSMM_LIBRARY_DIR "${LIBXSMM_INSTALL_DIR}/lib" CACHE PATH "LIBXSMM library directory." FORCE) +SET(LIBXSMM_LIBS "${LIBXSMM_LIBRARY_DIR}/libxsmm.a" + "${LIBXSMM_LIBRARY_DIR}/libxsmmnoblas.a") + +ExternalProject_Add( + extern_libxsmm + GIT_REPOSITORY "https://github.com/hfp/libxsmm.git" + GIT_TAG "7cc03b5b342fdbc6b6d990b190671c5dbb8489a2" + PREFIX ${LIBXSMM_SOURCES_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_IN_SOURCE 1 + BUILD_COMMAND $(MAKE) --silent PREFIX=${LIBXSMM_INSTALL_DIR} CXX=g++ CC=gcc WARP=0 install + INSTALL_COMMAND "" +) +ADD_LIBRARY(libxsmm STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET libxsmm PROPERTY IMPORTED_LOCATION "${LIBXSMM_LIBRARY_DIR}/libxsmm.a") +SET_PROPERTY(TARGET libxsmm PROPERTY IMPORTED_LOCATION "${LIBXSMM_LIBRARY_DIR}/libxsmmnoblas.a") + +MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}") +include_directories(${LIBXSMM_INCLUDE_DIR}) +ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM) +ADD_DEPENDENCIES(libxsmm extern_libxsmm) +LIST(APPEND external_project_dependencies libxsmm) + diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 25c07850dda7b2f69c2207c37b9d2368632104ec..260985cc8aa4ad0f231798666c048703b64c6d15 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -24,7 +24,7 @@ SET(MKLDNN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/mkldnn) SET(MKLDNN_INC_DIR "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE) IF(WIN32 OR APPLE) - MESSAGE(WARNING + MESSAGE(WARNING "Windows or Mac is not supported with MKLDNN in Paddle yet." "Force WITH_MKLDNN=OFF") SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE) @@ -45,7 +45,8 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML") ELSE() MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN") ENDIF() -SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result") +SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result") +SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") ExternalProject_Add( @@ -53,11 +54,13 @@ ExternalProject_Add( ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "db3424ad44901513c03a1ea31ccaacdf633fbe9f" + GIT_TAG "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} + CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} - CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} + CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} CMAKE_ARGS -DMKLROOT=${MKLML_ROOT} CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG} CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG} diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 8af2765f58717408e3a1ef6b500bb01511bfd8d3..56024edf5be092f81ed893633a8e7cafc8c8d429 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -29,6 +29,8 @@ IF(NOT ${CBLAS_FOUND}) "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE FILEPATH "openblas library." FORCE) + ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS) + SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable") SET(OPENBLAS_COMMIT "v0.2.20") @@ -112,7 +114,17 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR}) SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c) FILE(WRITE ${dummyfile} "const char *dummy_cblas = \"${dummyfile}\";") ADD_LIBRARY(cblas STATIC ${dummyfile}) -TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES}) + +IF("${CBLAS_PROVIDER}" STREQUAL "MKLML") + TARGET_LINK_LIBRARIES(cblas dynload_mklml) +ELSE() + TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES}) +ENDIF("${CBLAS_PROVIDER}" STREQUAL "MKLML") + +IF(WITH_LIBXSMM) + TARGET_LINK_LIBRARIES(cblas ${LIBXSMM_LIBS}) + ADD_DEPENDENCIES(cblas extern_libxsmm) +ENDIF() IF(NOT ${CBLAS_FOUND}) ADD_DEPENDENCIES(cblas extern_openblas) diff --git a/cmake/external/python.cmake b/cmake/external/python.cmake index d7e5571bdbd8ba58d8a08c9426971f1c7b186413..f17b8d46dc2d8ded81ced7de5827d5e7fd5109f0 100644 --- a/cmake/external/python.cmake +++ b/cmake/external/python.cmake @@ -18,8 +18,9 @@ ENDIF() INCLUDE(python_module) -FIND_PACKAGE(PythonInterp 2.7) -FIND_PACKAGE(PythonLibs 2.7) +FIND_PACKAGE(PythonInterp ${PY_VERSION}) +FIND_PACKAGE(PythonLibs ${PY_VERSION}) + # Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE. ADD_LIBRARY(python SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES}) diff --git a/cmake/external/xbyak.cmake b/cmake/external/xbyak.cmake new file mode 100644 index 0000000000000000000000000000000000000000..384c2f9328296ce6a8a6293be6cc47e5063dd3c4 --- /dev/null +++ b/cmake/external/xbyak.cmake @@ -0,0 +1,58 @@ +# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +set(WITH_XBYAK ON) +if(WIN32 OR APPLE) + SET(WITH_XBYAK OFF CACHE STRING "Disable XBYAK in Windows and MacOS" FORCE) + return() +endif() + +include(ExternalProject) + +set(XBYAK_PROJECT extern_xbyak) +set(XBYAK_PREFIX_DIR ${THIRD_PARTY_PATH}/xbyak) +set(XBYAK_INSTALL_ROOT ${THIRD_PARTY_PATH}/install/xbyak) +set(XBYAK_INC_DIR ${XBYAK_INSTALL_ROOT}/include) + +include_directories(${XBYAK_INC_DIR}) +include_directories(${XBYAK_INC_DIR}/xbyak) + +add_definitions(-DPADDLE_WITH_XBYAK) + +# xbyak options +add_definitions(-DXBYAK64) +add_definitions(-DXBYAK_NO_OP_NAMES) + +ExternalProject_Add( + ${XBYAK_PROJECT} + ${EXTERNAL_PROJECT_LOG_ARGS} + DEPENDS "" + GIT_REPOSITORY "https://github.com/herumi/xbyak.git" + GIT_TAG "v5.661" # Jul 26th + PREFIX ${XBYAK_PREFIX_DIR} + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XBYAK_INSTALL_ROOT} + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XBYAK_INSTALL_ROOT} +) + +if (${CMAKE_VERSION} VERSION_LESS "3.3.0") + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/xbyak_dummy.c) + file(WRITE ${dummyfile} "const char *dummy_xbyak = \"${dummyfile}\";") + add_library(xbyak STATIC ${dummyfile}) +else() + add_library(xbyak INTERFACE) +endif() + +add_dependencies(xbyak ${XBYAK_PROJECT}) +list(APPEND external_project_dependencies xbyak) diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 9ddd05b3d9404df29ca1bf634105314b7e6a5b70..82c958073cba92f00a341121e36ba45531b22aec 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -96,6 +96,20 @@ if(NOT APPLE AND NOT ANDROID) set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt") endif(NOT APPLE AND NOT ANDROID) +set_property(GLOBAL PROPERTY FLUID_MODULES "") +# find all fluid modules is used for paddle fluid static library +# for building inference libs +function(find_fluid_modules TARGET_NAME) + get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE) + string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path}) + string(FIND "${__target_path}" "fluid" pos) + if(pos GREATER 1) + get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) + set(fluid_modules ${fluid_modules} ${TARGET_NAME}) + set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}") + endif() +endfunction(find_fluid_modules) + function(merge_static_libs TARGET_NAME) set(libs ${ARGN}) list(REMOVE_DUPLICATES libs) @@ -195,6 +209,15 @@ function(cc_library TARGET_NAME) list(REMOVE_ITEM cc_library_DEPS warpctc) add_dependencies(${TARGET_NAME} warpctc) endif() + # Only deps libmklml.so, not link + if("${cc_library_DEPS};" MATCHES "mklml;") + list(REMOVE_ITEM cc_library_DEPS mklml) + if(NOT "${TARGET_NAME}" MATCHES "dynload_mklml") + list(APPEND cc_library_DEPS dynload_mklml) + endif() + add_dependencies(${TARGET_NAME} mklml) + target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed") + endif() target_link_libraries(${TARGET_NAME} ${cc_library_DEPS}) add_dependencies(${TARGET_NAME} ${cc_library_DEPS}) endif() @@ -234,13 +257,17 @@ function(cc_test TARGET_NAME) set(multiValueArgs SRCS DEPS ARGS) cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_executable(${TARGET_NAME} ${cc_test_SRCS}) - target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog) - add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog) + target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) + add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} ${cc_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) if (${cc_test_SERIAL}) - set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1) + set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1) + + set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true) + set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true) + set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true) endif() endif() endfunction(cc_test) @@ -300,11 +327,15 @@ function(nv_test TARGET_NAME) set(multiValueArgs SRCS DEPS) cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS}) - target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog) - add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog) + target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) + add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) add_test(${TARGET_NAME} ${TARGET_NAME}) if (nv_test_SERIAL) - set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1) + set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1) + + set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true) + set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true) + set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true) endif() endif() endfunction(nv_test) @@ -552,7 +583,9 @@ function(py_test TARGET_NAME) set(multiValueArgs SRCS DEPS ARGS ENVS) cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_test(NAME ${TARGET_NAME} - COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS} + COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true + FLAGS_cpu_deterministic=true + PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS} ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) endif() @@ -610,3 +643,21 @@ function(grpc_library TARGET_NAME) COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}") endfunction() + + +function(brpc_library TARGET_NAME) + set(oneValueArgs PROTO) + set(multiValueArgs SRCS DEPS) + set(options "") + cmake_parse_arguments(brpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + message(STATUS "generating brpc ${brpc_library_PROTO}") + + get_filename_component(ABS_PROTO ${brpc_library_PROTO} ABSOLUTE) + get_filename_component(PROTO_WE ${brpc_library_PROTO} NAME_WE) + get_filename_component(PROTO_PATH ${ABS_PROTO} PATH) + + protobuf_generate_cpp(brpc_proto_srcs brpc_proto_hdrs "${ABS_PROTO}") + cc_library("${TARGET_NAME}_proto" SRCS "${brpc_proto_srcs}") + cc_library("${TARGET_NAME}" SRCS "${brpc_library_SRCS}" DEPS "${TARGET_NAME}_proto" "${brpc_library_DEPS}") +endfunction() diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 236a55d332a91c88d1c5515e7aca4142930a079f..834ab5a9e527355d3664313d38cd4920f6fbf535 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -12,19 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -set_property(GLOBAL PROPERTY FLUID_MODULES "") -# find all fluid modules is used for paddle fluid static library -function(find_fluid_modules TARGET_NAME) - get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE) - string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path}) - string(FIND "${__target_path}" "fluid" pos) - if(pos GREATER 1) - get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) - set(fluid_modules ${fluid_modules} ${TARGET_NAME}) - set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}") - endif() -endfunction(find_fluid_modules) - # make package for paddle fluid shared and static library function(copy TARGET) set(options "") @@ -39,7 +26,7 @@ function(copy TARGET) message(FATAL_ERROR "${TARGET} source numbers are not equal to destination numbers") endif() math(EXPR len "${copy_lib_SRCS_len} - 1") - + add_custom_target(${TARGET} DEPENDS ${copy_lib_DEPS}) foreach(index RANGE ${len}) list(GET copy_lib_SRCS ${index} src) @@ -149,10 +136,23 @@ copy(memory_lib DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail ) +set(inference_deps paddle_fluid_shared paddle_fluid) + +set(module "inference/api") +if (WITH_ANAKIN AND WITH_GPU) + copy(anakin_inference_lib DEPS paddle_inference_api inference_anakin_api + SRCS + ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api + ${ANAKIN_INSTALL_DIR} # anakin release + DSTS ${dst_dir}/inference/anakin ${dst_dir}/inference/anakin) + list(APPEND inference_deps anakin_inference_lib) +endif() + set(module "inference") -copy(inference_lib DEPS paddle_fluid_shared paddle_fluid +copy(inference_lib DEPS ${inference_deps} SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.* - DSTS ${dst_dir}/${module} ${dst_dir}/${module} + ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci + DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ) set(module "platform") diff --git a/cmake/version.cmake b/cmake/version.cmake index cde650128a068faf32f4abfff5cdfdeb656d8577..ac10bdf067be549fe90112aef73fd6e1fbe0ac48 100644 --- a/cmake/version.cmake +++ b/cmake/version.cmake @@ -1,23 +1,46 @@ # Get the latest git tag. set(PADDLE_VERSION $ENV{PADDLE_VERSION}) set(tmp_version "HEAD") +set(TAG_VERSION_REGEX "[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?") +set(COMMIT_VERSION_REGEX "[0-9a-f]+[0-9a-f]+[0-9a-f]+[0-9a-f]+[0-9a-f]+") while ("${PADDLE_VERSION}" STREQUAL "") + # Check current branch name execute_process( - COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 ${tmp_version} + COMMAND ${GIT_EXECUTABLE} rev-parse --abbrev-ref ${tmp_version} WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} - OUTPUT_VARIABLE GIT_TAG_NAME - RESULT_VARIABLE GIT_RESULT + OUTPUT_VARIABLE GIT_BRANCH_NAME + RESULT_VARIABLE GIT_BRANCH_RESULT ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) - if (NOT ${GIT_RESULT}) - # Check the tag is a correct version - if (${GIT_TAG_NAME} MATCHES "v[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?") - string(REPLACE "v" "" PADDLE_VERSION ${GIT_TAG_NAME}) - else() # otherwise, get the previous git tag name. - set(tmp_version "${GIT_TAG_NAME}~1") + if (NOT ${GIT_BRANCH_RESULT}) + execute_process( + COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 --always ${tmp_version} + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} + OUTPUT_VARIABLE GIT_TAG_NAME + RESULT_VARIABLE GIT_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + if (NOT ${GIT_RESULT}) + # Check if current branch is release branch + if (${GIT_BRANCH_NAME} MATCHES "release/${TAG_VERSION_REGEX}") + # Check the tag is a correct version + if (${GIT_TAG_NAME} MATCHES "${COMMIT_VERSION_REGEX}") + # if no tag was found, set PADDLE_VERSION to 0.0.0 to represent latest + set(PADDLE_VERSION "0.0.0") + elseif (${GIT_TAG_NAME} MATCHES "v${TAG_VERSION_REGEX}") + string(REPLACE "v" "" PADDLE_VERSION ${GIT_TAG_NAME}) + else() # otherwise, get the previous git tag name. + set(tmp_version "${GIT_TAG_NAME}~1") + endif() + else() + # otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest + set(PADDLE_VERSION "0.0.0") + endif() + else() + set(PADDLE_VERSION "0.0.0") + message(WARNING "Cannot add paddle version from git tag") endif() else() set(PADDLE_VERSION "0.0.0") - message(WARNING "Cannot add paddle version from git tag") + message(WARNING "Cannot add paddle version for wrong git branch result") endif() endwhile() diff --git a/doc/about/about_us.rst b/doc/about/about_us.rst new file mode 100644 index 0000000000000000000000000000000000000000..f67d8b8130030db8d7e7d10b30271a913bd6272a --- /dev/null +++ b/doc/about/about_us.rst @@ -0,0 +1,53 @@ +========= +关于我们 +========= + +什么是PaddlePaddle +-------------------- + +- PaddlePaddle是百度自主研发并开源的深度学习框架，它能够让开发者和企业安全、快速地实现自己的AI想法 + +- 项目团队汇聚了全球顶级的深度学习科学家，致力于为开发者和企业提供最好的深度学习研发体验 + +- 框架具有易学、易用、安全、高效四大特性，是最适合中国开发者和企业的深度学习工具 + +PaddlePaddle的技术特色 +------------------------- + +- 新一代深度学习框架： PaddlePaddle是基于“深度学习编程语言”的新一代深度学习框架，在保证性能的同时，极大的提升了框架对模型的表达能力，能够描述任意潜在可能出现的模型 + +- 对大规模计算更加友好：经过百度内多种大规模计算业务的打磨，PaddlePaddle在分布式计算上表现优异，基于EDL技术能够节约大量计算资源，同时也能支持大规模稀疏模型的训练 + +- 提供可视化的深度学习：通过Visual DL可以帮助开发者方便的观测训练整体趋势、数据样本质量和中间结果、参数分布和变化趋势、以及模型的结构，帮助开发者更便捷的完成编程过程 + +提供基于PaddlePaddle的教育体系 +-------------------------------- + +- 深度学习课程：百度与中国市场顶级的教育、培训机构共同开发了深度学习精品课程以及学习教材，帮助开发者从零掌握深度学习 + +- 深度学习实训：对于目的是科研和学习的用户，PaddlePaddle提供了无需安装、线上运行的开发环境，并提供算法、算力、数据支持 + +- 线下培训：提供丰富、高质量的线下教育活动，如青年教师培训、线下实战营、沙龙等多种形式的培训和交流 + + +提供基于PaddlePaddle的AI服务 +------------------------------ + +- EadyDL：可以帮助零算法基础的企业快速完成一个深度学习任务，只需少量的数据即可得到优质的模型 + +- AI市场：提供标准化的AI 能力、产品的交易机制，帮助企业快速找到所需，有效开展AI业务 + +- 深度学习竞赛： PaddlePaddle汇聚顶尖深度学习开发者，企业可以发布自己的商业问题，通过竞赛方式快速找到最优的解决方案 + +你对PaddlePaddle有任何的问题都可以通过以下方式联系到我们 +----------------------------------------------------------- + +- 学习/使用问题：可以在 `PaddlePaddle开源社区 `_，以及 `PaddlePaddle中文社区 `_ 向我们反馈 + +- 对PaddlePaddle框架发展的建议：可发送邮件至Paddle-better@baidu.com + +我们期待与你一起打造世界顶级深度学习框架，共同推动AI技术的进步 + + + +PaddlePaddle团队 diff --git a/doc/fluid/api/average.rst b/doc/fluid/api/average.rst new file mode 100644 index 0000000000000000000000000000000000000000..496f5b29875443f0c44f50fcb3ca837f4e7bcd12 --- /dev/null +++ b/doc/fluid/api/average.rst @@ -0,0 +1,16 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +============= +fluid.average +============= + +.. _api_fluid_average_WeightedAverage: + +WeightedAverage +--------------- + +.. autoclass:: paddle.fluid.average.WeightedAverage + :members: + :noindex: + diff --git a/doc/fluid/api/backward.rst b/doc/fluid/api/backward.rst new file mode 100644 index 0000000000000000000000000000000000000000..115e0d24b39928cfc349f72e0a21d6374cd8cd75 --- /dev/null +++ b/doc/fluid/api/backward.rst @@ -0,0 +1,23 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +============== +fluid.backward +============== + +.. _api_fluid_backward_append_backward: + +append_backward +--------------- + +.. autofunction:: paddle.fluid.backward.append_backward + :noindex: + +.. _api_fluid_backward_calc_gradient: + +calc_gradient +------------- + +.. autofunction:: paddle.fluid.backward.calc_gradient + :noindex: + diff --git a/doc/fluid/api/clip.rst b/doc/fluid/api/clip.rst index 3ba096388fc87dda3096a9030fe5749e61112c06..aeefbb95a46e5d5ed46375e388a720fad2711779 100644 --- a/doc/fluid/api/clip.rst +++ b/doc/fluid/api/clip.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -==== -clip -==== +========== +fluid.clip +========== + +.. _api_fluid_clip_ErrorClipByValue: ErrorClipByValue ---------------- @@ -12,6 +14,8 @@ ErrorClipByValue :members: :noindex: +.. _api_fluid_clip_GradientClipByValue: + GradientClipByValue ------------------- @@ -19,6 +23,8 @@ GradientClipByValue :members: :noindex: +.. _api_fluid_clip_GradientClipByNorm: + GradientClipByNorm ------------------ @@ -26,6 +32,8 @@ GradientClipByNorm :members: :noindex: +.. _api_fluid_clip_GradientClipByGlobalNorm: + GradientClipByGlobalNorm ------------------------ @@ -33,15 +41,3 @@ GradientClipByGlobalNorm :members: :noindex: -append_gradient_clip_ops ------------------------- - -.. autofunction:: paddle.fluid.clip.append_gradient_clip_ops - :noindex: - -error_clip_callback -------------------- - -.. autofunction:: paddle.fluid.clip.error_clip_callback - :noindex: - diff --git a/doc/fluid/api/data.rst b/doc/fluid/api/data.rst deleted file mode 100644 index b56c7332cc284649c7e04328e51a7faa78593a39..0000000000000000000000000000000000000000 --- a/doc/fluid/api/data.rst +++ /dev/null @@ -1,10 +0,0 @@ -================================== -Data Reader Interface and DataSets -================================== - -.. toctree:: - :maxdepth: 1 - - data/data_reader.rst - data/image.rst - data/dataset.rst diff --git a/doc/fluid/api/data_feeder.rst b/doc/fluid/api/data_feeder.rst index 3df5c0307ffed9d101da58b385840b115920e906..11d2890f5b3446e37c3ef31e5a17ebebe169dbc8 100644 --- a/doc/fluid/api/data_feeder.rst +++ b/doc/fluid/api/data_feeder.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -=========== -data_feeder -=========== +================= +fluid.data_feeder +================= + +.. _api_fluid_data_feeder_DataFeeder: DataFeeder ---------- diff --git a/doc/fluid/api/evaluator.rst b/doc/fluid/api/evaluator.rst deleted file mode 100644 index c0dc9a0d1d9f2f70948dc3c905dca25d7dd43742..0000000000000000000000000000000000000000 --- a/doc/fluid/api/evaluator.rst +++ /dev/null @@ -1,7 +0,0 @@ -.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` - !DO NOT EDIT THIS FILE MANUALLY! - -========= -evaluator -========= - diff --git a/doc/fluid/api/executor.rst b/doc/fluid/api/executor.rst index f67a14c49f372e67d18ec8e6f87da01109376d22..f23ecc1f80030f20359ce9675130a167722606c9 100644 --- a/doc/fluid/api/executor.rst +++ b/doc/fluid/api/executor.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -======== -executor -======== +============== +fluid.executor +============== + +.. _api_fluid_executor_Executor: Executor -------- @@ -12,27 +14,27 @@ Executor :members: :noindex: +.. _api_fluid_executor_global_scope: + global_scope ------------ .. autofunction:: paddle.fluid.executor.global_scope :noindex: +.. _api_fluid_executor_scope_guard: + scope_guard ----------- .. autofunction:: paddle.fluid.executor.scope_guard :noindex: -switch_scope ------------- - -.. autofunction:: paddle.fluid.executor.switch_scope - :noindex: +.. _api_fluid_executor__switch_scope: -fetch_var ---------- +_switch_scope +------------- -.. autofunction:: paddle.fluid.executor.fetch_var +.. autofunction:: paddle.fluid.executor._switch_scope :noindex: diff --git a/doc/fluid/api/fluid.rst b/doc/fluid/api/fluid.rst new file mode 100644 index 0000000000000000000000000000000000000000..7eab58355c3648d929d3b5d98984adce9034f016 --- /dev/null +++ b/doc/fluid/api/fluid.rst @@ -0,0 +1,362 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +===== +fluid +===== + +.. _api_fluid_Block: + +Block +----- + +.. autoclass:: paddle.fluid.Block + :members: + :noindex: + +.. _api_fluid_Variable: + +Variable +-------- + +.. autoclass:: paddle.fluid.Variable + :members: + :noindex: + +.. _api_fluid_Program: + +Program +------- + +.. autoclass:: paddle.fluid.Program + :members: + :noindex: + +.. _api_fluid_Operator: + +Operator +-------- + +.. autoclass:: paddle.fluid.Operator + :members: + :noindex: + +.. _api_fluid_default_startup_program: + +default_startup_program +----------------------- + +.. autofunction:: paddle.fluid.default_startup_program + :noindex: + +.. _api_fluid_default_main_program: + +default_main_program +-------------------- + +.. autofunction:: paddle.fluid.default_main_program + :noindex: + +.. _api_fluid_program_guard: + +program_guard +------------- + +.. autofunction:: paddle.fluid.program_guard + :noindex: + +.. _api_fluid_get_var: + +get_var +------- + +.. autofunction:: paddle.fluid.get_var + :noindex: + +.. _api_fluid_Executor: + +Executor +-------- + +.. autoclass:: paddle.fluid.Executor + :members: + :noindex: + +.. _api_fluid_global_scope: + +global_scope +------------ + +.. autofunction:: paddle.fluid.global_scope + :noindex: + +.. _api_fluid_scope_guard: + +scope_guard +----------- + +.. autofunction:: paddle.fluid.scope_guard + :noindex: + +.. _api_fluid__switch_scope: + +_switch_scope +------------- + +.. autofunction:: paddle.fluid._switch_scope + :noindex: + + +.. _api_fluid_make_channel: + +make_channel +------------ + +.. autofunction:: paddle.fluid.make_channel + :noindex: + +.. _api_fluid_channel_send: + +channel_send +------------ + +.. autofunction:: paddle.fluid.channel_send + :noindex: + +.. _api_fluid_channel_recv: + +channel_recv +------------ + +.. autofunction:: paddle.fluid.channel_recv + :noindex: + +.. _api_fluid_channel_close: + +channel_close +------------- + +.. autofunction:: paddle.fluid.channel_close + :noindex: + +.. _api_fluid_Select: + +Select +------ + +.. autoclass:: paddle.fluid.Select + :members: + :noindex: + +.. _api_fluid_Trainer: + +Trainer +------- + +.. autoclass:: paddle.fluid.Trainer + :members: + :noindex: + +.. _api_fluid_BeginEpochEvent: + +BeginEpochEvent +--------------- + +.. autoclass:: paddle.fluid.BeginEpochEvent + :members: + :noindex: + +.. _api_fluid_EndEpochEvent: + +EndEpochEvent +------------- + +.. autoclass:: paddle.fluid.EndEpochEvent + :members: + :noindex: + +.. _api_fluid_BeginStepEvent: + +BeginStepEvent +-------------- + +.. autoclass:: paddle.fluid.BeginStepEvent + :members: + :noindex: + +.. _api_fluid_EndStepEvent: + +EndStepEvent +------------ + +.. autoclass:: paddle.fluid.EndStepEvent + :members: + :noindex: + +.. _api_fluid_CheckpointConfig: + +CheckpointConfig +---------------- + +.. autoclass:: paddle.fluid.CheckpointConfig + :members: + :noindex: + +.. _api_fluid_Inferencer: + +Inferencer +---------- + +.. autoclass:: paddle.fluid.Inferencer + :members: + :noindex: + +.. _api_fluid_DistributeTranspiler: + +DistributeTranspiler +-------------------- + +.. autoclass:: paddle.fluid.DistributeTranspiler + :members: + :noindex: + +.. _api_fluid_memory_optimize: + +memory_optimize +--------------- + +.. autofunction:: paddle.fluid.memory_optimize + :noindex: + +.. _api_fluid_release_memory: + +release_memory +-------------- + +.. autofunction:: paddle.fluid.release_memory + :noindex: + +.. _api_fluid_ParallelExecutor: + +ParallelExecutor +---------------- + +.. autoclass:: paddle.fluid.ParallelExecutor + :members: + :noindex: + +.. _api_fluid_ExecutionStrategy: + +ExecutionStrategy +----------------- + +.. autoclass:: paddle.fluid.ExecutionStrategy + :members: + :noindex: + +.. _api_fluid_BuildStrategy: + +BuildStrategy +------------- + +.. autoclass:: paddle.fluid.BuildStrategy + :members: + :noindex: + +.. _api_fluid_create_lod_tensor: + +create_lod_tensor +----------------- + +.. autofunction:: paddle.fluid.create_lod_tensor + :noindex: + +.. _api_fluid_create_random_int_lodtensor: + +create_random_int_lodtensor +--------------------------- + +.. autofunction:: paddle.fluid.create_random_int_lodtensor + :noindex: + +.. _api_fluid_LoDTensor: + +LoDTensor +--------- + +.. autoclass:: paddle.fluid.LoDTensor + :members: + :noindex: + +.. _api_fluid_CPUPlace: + +CPUPlace +-------- + +.. autoclass:: paddle.fluid.CPUPlace + :members: + :noindex: + +.. _api_fluid_CUDAPlace: + +CUDAPlace +--------- + +.. autoclass:: paddle.fluid.CUDAPlace + :members: + :noindex: + +.. _api_fluid_CUDAPinnedPlace: + +CUDAPinnedPlace +--------------- + +.. autoclass:: paddle.fluid.CUDAPinnedPlace + :members: + :noindex: + +.. _api_fluid_Tensor: + +Tensor +------ + +.. autoclass:: paddle.fluid.Tensor + :members: + :noindex: + +.. _api_fluid_ParamAttr: + +ParamAttr +--------- + +.. autoclass:: paddle.fluid.ParamAttr + :members: + :noindex: + +.. _api_fluid_WeightNormParamAttr: + +WeightNormParamAttr +------------------- + +.. autoclass:: paddle.fluid.WeightNormParamAttr + :members: + :noindex: + +.. _api_fluid_DataFeeder: + +DataFeeder +---------- + +.. autoclass:: paddle.fluid.DataFeeder + :members: + :noindex: + +.. _api_fluid_Scope: + +Scope +----- + +.. autoclass:: paddle.fluid.Scope + :members: + :noindex: + diff --git a/doc/fluid/api/gen_doc.py b/doc/fluid/api/gen_doc.py index 89ab880301b6ac687fd61f556f87f03792c37da3..02efce2bf8392c62a7600c272bedcadc6563f927 100644 --- a/doc/fluid/api/gen_doc.py +++ b/doc/fluid/api/gen_doc.py @@ -29,19 +29,27 @@ def parse_arg(): class DocGenerator(object): - def __init__(self, module_name, stream=sys.stdout): + def __init__(self, module_name=None, stream=sys.stdout): + if module_name == "": + module_name = None self.stream = stream - self.module_name = module_name - if not hasattr(fluid, module_name): - raise ValueError("Cannot find fluid.{0}".format(module_name)) + if module_name is None: + self.module_name = "fluid" else: - self.module = getattr(fluid, module_name) + self.module_name = "fluid." + module_name + if module_name is None: + self.module = fluid + else: + if not hasattr(fluid, module_name): + raise ValueError("Cannot find fluid.{0}".format(module_name)) + else: + self.module = getattr(fluid, module_name) self.stream.write('''.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! ''') - self._print_header_(module_name, dot='=', is_title=True) + self._print_header_(self.module_name, dot='=', is_title=True) def print_submodule(self, submodule_name): submodule = getattr(self.module, submodule_name) @@ -60,25 +68,29 @@ class DocGenerator(object): self._print_header_(name, dot='=', is_title=False) def print_item(self, name): - item = getattr(self.module, name) + item = getattr(self.module, name, None) + if item is None: + return if isinstance(item, types.TypeType): self.print_class(name) elif isinstance(item, types.FunctionType): self.print_method(name) else: - raise RuntimeError("Unsupported item {0}".format(name)) + pass def print_class(self, name): + self._print_ref_(name) self._print_header_(name, dot='-', is_title=False) - self.stream.write('''.. autoclass:: paddle.fluid.{0}.{1} + self.stream.write('''.. autoclass:: paddle.{0}.{1} :members: :noindex: '''.format(self.module_name, name)) def print_method(self, name): + self._print_ref_(name) self._print_header_(name, dot='-', is_title=False) - self.stream.write('''.. autofunction:: paddle.fluid.{0}.{1} + self.stream.write('''.. autofunction:: paddle.{0}.{1} :noindex: '''.format(self.module_name, name)) @@ -94,6 +106,10 @@ class DocGenerator(object): self.stream.write('\n') self.stream.write('\n') + def _print_ref_(self, name): + self.stream.write(".. _api_{0}_{1}:\n\n".format("_".join( + self.module_name.split(".")), name)) + def main(): args = parse_arg() diff --git a/doc/fluid/api/gen_doc.sh b/doc/fluid/api/gen_doc.sh index 0f0539355559446fd91f659d61b636db214b5a40..b14ee29873c50fd011f6c48b754767ac8918252a 100755 --- a/doc/fluid/api/gen_doc.sh +++ b/doc/fluid/api/gen_doc.sh @@ -1,7 +1,9 @@ #!/bin/bash -python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst +python gen_doc.py layers --submodules control_flow device io nn ops tensor learning_rate_scheduler detection metric_op tensor > layers.rst -for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer +for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer transpiler recordio_writer backward average profiler do python gen_doc.py ${module} > ${module}.rst done + +python gen_doc.py "" > fluid.rst diff --git a/doc/fluid/api/index_en.rst b/doc/fluid/api/index_en.rst index 29cea9c68221b921939e8e09072d87f9f604e21b..359406819a993e7eaf2155c839373df44d97b103 100644 --- a/doc/fluid/api/index_en.rst +++ b/doc/fluid/api/index_en.rst @@ -1,10 +1,11 @@ -====================== -Fluid -====================== +============= +API Reference +============= .. toctree:: :maxdepth: 1 + fluid.rst layers.rst data_feeder.rst executor.rst @@ -18,3 +19,8 @@ Fluid regularizer.rst io.rst data.rst + transpiler.rst + recordio_writer.rst + backward.rst + average.rst + profiler.rst diff --git a/doc/fluid/api/initializer.rst b/doc/fluid/api/initializer.rst index c49a98c744cdf907630ea8c74791ff2021d996e8..dc0b52b14fd242dfaded1cb9a8e0ab9eb66b0607 100644 --- a/doc/fluid/api/initializer.rst +++ b/doc/fluid/api/initializer.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -=========== -initializer -=========== +================= +fluid.initializer +================= + +.. _api_fluid_initializer_Constant: Constant -------- @@ -12,6 +14,8 @@ Constant :members: :noindex: +.. _api_fluid_initializer_Uniform: + Uniform ------- @@ -19,6 +23,8 @@ Uniform :members: :noindex: +.. _api_fluid_initializer_Normal: + Normal ------ @@ -26,6 +32,8 @@ Normal :members: :noindex: +.. _api_fluid_initializer_Xavier: + Xavier ------ @@ -33,18 +41,42 @@ Xavier :members: :noindex: +.. _api_fluid_initializer_Bilinear: + +Bilinear +-------- + +.. autoclass:: paddle.fluid.initializer.Bilinear + :members: + :noindex: + +.. _api_fluid_initializer_MSRA: + +MSRA +---- + +.. autoclass:: paddle.fluid.initializer.MSRA + :members: + :noindex: + +.. _api_fluid_initializer_force_init_on_cpu: + force_init_on_cpu ----------------- .. autofunction:: paddle.fluid.initializer.force_init_on_cpu :noindex: +.. _api_fluid_initializer_init_on_cpu: + init_on_cpu ----------- .. autofunction:: paddle.fluid.initializer.init_on_cpu :noindex: +.. _api_fluid_initializer_ConstantInitializer: + ConstantInitializer ------------------- @@ -52,6 +84,8 @@ ConstantInitializer :members: :noindex: +.. _api_fluid_initializer_UniformInitializer: + UniformInitializer ------------------ @@ -59,6 +93,8 @@ UniformInitializer :members: :noindex: +.. _api_fluid_initializer_NormalInitializer: + NormalInitializer ----------------- @@ -66,6 +102,8 @@ NormalInitializer :members: :noindex: +.. _api_fluid_initializer_XavierInitializer: + XavierInitializer ----------------- @@ -73,3 +111,21 @@ XavierInitializer :members: :noindex: +.. _api_fluid_initializer_BilinearInitializer: + +BilinearInitializer +------------------- + +.. autoclass:: paddle.fluid.initializer.BilinearInitializer + :members: + :noindex: + +.. _api_fluid_initializer_MSRAInitializer: + +MSRAInitializer +--------------- + +.. autoclass:: paddle.fluid.initializer.MSRAInitializer + :members: + :noindex: + diff --git a/doc/fluid/api/io.rst b/doc/fluid/api/io.rst index dd9d88b669957c22cd0a07fa4b7e219e2d6e5d61..7cee0bc4d9aa2c51517d23a381f14a8f63cc3681 100644 --- a/doc/fluid/api/io.rst +++ b/doc/fluid/api/io.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -== -io -== +======== +fluid.io +======== + +.. _api_fluid_io_save_vars: save_vars --------- @@ -11,51 +13,115 @@ save_vars .. autofunction:: paddle.fluid.io.save_vars :noindex: +.. _api_fluid_io_save_params: + save_params ----------- .. autofunction:: paddle.fluid.io.save_params :noindex: +.. _api_fluid_io_save_persistables: + save_persistables ----------------- .. autofunction:: paddle.fluid.io.save_persistables :noindex: +.. _api_fluid_io_load_vars: + load_vars --------- .. autofunction:: paddle.fluid.io.load_vars :noindex: +.. _api_fluid_io_load_params: + load_params ----------- .. autofunction:: paddle.fluid.io.load_params :noindex: +.. _api_fluid_io_load_persistables: + load_persistables ----------------- .. autofunction:: paddle.fluid.io.load_persistables :noindex: +.. _api_fluid_io_save_inference_model: + save_inference_model -------------------- .. autofunction:: paddle.fluid.io.save_inference_model :noindex: +.. _api_fluid_io_load_inference_model: + load_inference_model -------------------- .. autofunction:: paddle.fluid.io.load_inference_model :noindex: +.. _api_fluid_io_get_inference_program: + get_inference_program --------------------- .. autofunction:: paddle.fluid.io.get_inference_program :noindex: +.. _api_fluid_io_save_checkpoint: + +save_checkpoint +--------------- + +.. autofunction:: paddle.fluid.io.save_checkpoint + :noindex: + +.. _api_fluid_io_load_checkpoint: + +load_checkpoint +--------------- + +.. autofunction:: paddle.fluid.io.load_checkpoint + :noindex: + +.. _api_fluid_io_clean_checkpoint: + +clean_checkpoint +---------------- + +.. autofunction:: paddle.fluid.io.clean_checkpoint + :noindex: + +.. _api_fluid_io_load_persist_vars_without_grad: + +load_persist_vars_without_grad +------------------------------ + +.. autofunction:: paddle.fluid.io.load_persist_vars_without_grad + :noindex: + +.. _api_fluid_io_save_persist_vars_without_grad: + +save_persist_vars_without_grad +------------------------------ + +.. autofunction:: paddle.fluid.io.save_persist_vars_without_grad + :noindex: + +.. _api_fluid_io_get_latest_checkpoint_serial: + +get_latest_checkpoint_serial +---------------------------- + +.. autofunction:: paddle.fluid.io.get_latest_checkpoint_serial + :noindex: + diff --git a/doc/fluid/api/layers.rst b/doc/fluid/api/layers.rst index f53da4d194f8d2428b4121fa1bb31f3fc95a9f64..ecbd8191ccf5aa6046e7875fe8afa2ed0105e4a0 100644 --- a/doc/fluid/api/layers.rst +++ b/doc/fluid/api/layers.rst @@ -1,25 +1,31 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -====== -layers -====== +============ +fluid.layers +============ control_flow ============ +.. _api_fluid_layers_split_lod_tensor: + split_lod_tensor ---------------- .. autofunction:: paddle.fluid.layers.split_lod_tensor :noindex: +.. _api_fluid_layers_merge_lod_tensor: + merge_lod_tensor ---------------- .. autofunction:: paddle.fluid.layers.merge_lod_tensor :noindex: +.. _api_fluid_layers_BlockGuard: + BlockGuard ---------- @@ -27,6 +33,8 @@ BlockGuard :members: :noindex: +.. _api_fluid_layers_BlockGuardWithCompletion: + BlockGuardWithCompletion ------------------------ @@ -34,12 +42,7 @@ BlockGuardWithCompletion :members: :noindex: -StaticRNNMemoryLink -------------------- - -.. autoclass:: paddle.fluid.layers.StaticRNNMemoryLink - :members: - :noindex: +.. _api_fluid_layers_WhileGuard: WhileGuard ---------- @@ -48,6 +51,8 @@ WhileGuard :members: :noindex: +.. _api_fluid_layers_While: + While ----- @@ -55,6 +60,8 @@ While :members: :noindex: +.. _api_fluid_layers_Switch: + Switch ------ @@ -62,78 +69,104 @@ Switch :members: :noindex: +.. _api_fluid_layers_lod_rank_table: + lod_rank_table -------------- .. autofunction:: paddle.fluid.layers.lod_rank_table :noindex: +.. _api_fluid_layers_max_sequence_len: + max_sequence_len ---------------- .. autofunction:: paddle.fluid.layers.max_sequence_len :noindex: +.. _api_fluid_layers_lod_tensor_to_array: + lod_tensor_to_array ------------------- .. autofunction:: paddle.fluid.layers.lod_tensor_to_array :noindex: +.. _api_fluid_layers_array_to_lod_tensor: + array_to_lod_tensor ------------------- .. autofunction:: paddle.fluid.layers.array_to_lod_tensor :noindex: +.. _api_fluid_layers_increment: + increment --------- .. autofunction:: paddle.fluid.layers.increment :noindex: +.. _api_fluid_layers_array_write: + array_write ----------- .. autofunction:: paddle.fluid.layers.array_write :noindex: +.. _api_fluid_layers_create_array: + create_array ------------ .. autofunction:: paddle.fluid.layers.create_array :noindex: +.. _api_fluid_layers_less_than: + less_than --------- .. autofunction:: paddle.fluid.layers.less_than :noindex: +.. _api_fluid_layers_equal: + equal ----- .. autofunction:: paddle.fluid.layers.equal :noindex: +.. _api_fluid_layers_array_read: + array_read ---------- .. autofunction:: paddle.fluid.layers.array_read :noindex: +.. _api_fluid_layers_shrink_memory: + shrink_memory ------------- .. autofunction:: paddle.fluid.layers.shrink_memory :noindex: +.. _api_fluid_layers_array_length: + array_length ------------ .. autofunction:: paddle.fluid.layers.array_length :noindex: +.. _api_fluid_layers_IfElse: + IfElse ------ @@ -141,6 +174,8 @@ IfElse :members: :noindex: +.. _api_fluid_layers_DynamicRNN: + DynamicRNN ---------- @@ -148,6 +183,8 @@ DynamicRNN :members: :noindex: +.. _api_fluid_layers_ConditionalBlock: + ConditionalBlock ---------------- @@ -155,6 +192,8 @@ ConditionalBlock :members: :noindex: +.. _api_fluid_layers_StaticRNN: + StaticRNN --------- @@ -162,12 +201,16 @@ StaticRNN :members: :noindex: +.. _api_fluid_layers_reorder_lod_tensor_by_rank: + reorder_lod_tensor_by_rank -------------------------- .. autofunction:: paddle.fluid.layers.reorder_lod_tensor_by_rank :noindex: +.. _api_fluid_layers_ParallelDo: + ParallelDo ---------- @@ -175,15 +218,27 @@ ParallelDo :members: :noindex: +.. _api_fluid_layers_Print: + Print ----- .. autofunction:: paddle.fluid.layers.Print :noindex: +.. _api_fluid_layers_is_empty: + +is_empty +-------- + +.. autofunction:: paddle.fluid.layers.is_empty + :noindex: + device ====== +.. _api_fluid_layers_get_places: + get_places ---------- @@ -193,12 +248,16 @@ get_places io == +.. _api_fluid_layers_data: + data ---- .. autofunction:: paddle.fluid.layers.data :noindex: +.. _api_fluid_layers_BlockGuardServ: + BlockGuardServ -------------- @@ -206,6 +265,8 @@ BlockGuardServ :members: :noindex: +.. _api_fluid_layers_ListenAndServ: + ListenAndServ ------------- @@ -213,498 +274,790 @@ ListenAndServ :members: :noindex: +.. _api_fluid_layers_Send: + Send ---- .. autofunction:: paddle.fluid.layers.Send :noindex: +.. _api_fluid_layers_Recv: + +Recv +---- + +.. autofunction:: paddle.fluid.layers.Recv + :noindex: + +.. _api_fluid_layers_open_recordio_file: + open_recordio_file ------------------ .. autofunction:: paddle.fluid.layers.open_recordio_file :noindex: +.. _api_fluid_layers_open_files: + open_files ---------- .. autofunction:: paddle.fluid.layers.open_files :noindex: +.. _api_fluid_layers_read_file: + read_file --------- .. autofunction:: paddle.fluid.layers.read_file :noindex: +.. _api_fluid_layers_shuffle: + shuffle ------- .. autofunction:: paddle.fluid.layers.shuffle :noindex: +.. _api_fluid_layers_batch: + batch ----- .. autofunction:: paddle.fluid.layers.batch :noindex: +.. _api_fluid_layers_double_buffer: + double_buffer ------------- .. autofunction:: paddle.fluid.layers.double_buffer :noindex: +.. _api_fluid_layers_random_data_generator: + +random_data_generator +--------------------- + +.. autofunction:: paddle.fluid.layers.random_data_generator + :noindex: + +.. _api_fluid_layers_Preprocessor: + +Preprocessor +------------ + +.. autoclass:: paddle.fluid.layers.Preprocessor + :members: + :noindex: + +.. _api_fluid_layers_load: + +load +---- + +.. autofunction:: paddle.fluid.layers.load + :noindex: + nn == +.. _api_fluid_layers_fc: + fc -- .. autofunction:: paddle.fluid.layers.fc :noindex: +.. _api_fluid_layers_embedding: + embedding --------- .. autofunction:: paddle.fluid.layers.embedding :noindex: +.. _api_fluid_layers_dynamic_lstm: + dynamic_lstm ------------ .. autofunction:: paddle.fluid.layers.dynamic_lstm :noindex: +.. _api_fluid_layers_dynamic_lstmp: + dynamic_lstmp ------------- .. autofunction:: paddle.fluid.layers.dynamic_lstmp :noindex: +.. _api_fluid_layers_dynamic_gru: + dynamic_gru ----------- .. autofunction:: paddle.fluid.layers.dynamic_gru :noindex: +.. _api_fluid_layers_gru_unit: + gru_unit -------- .. autofunction:: paddle.fluid.layers.gru_unit :noindex: +.. _api_fluid_layers_linear_chain_crf: + linear_chain_crf ---------------- .. autofunction:: paddle.fluid.layers.linear_chain_crf :noindex: +.. _api_fluid_layers_crf_decoding: + crf_decoding ------------ .. autofunction:: paddle.fluid.layers.crf_decoding :noindex: +.. _api_fluid_layers_cos_sim: + cos_sim ------- .. autofunction:: paddle.fluid.layers.cos_sim :noindex: +.. _api_fluid_layers_cross_entropy: + cross_entropy ------------- .. autofunction:: paddle.fluid.layers.cross_entropy :noindex: +.. _api_fluid_layers_square_error_cost: + square_error_cost ----------------- .. autofunction:: paddle.fluid.layers.square_error_cost :noindex: +.. _api_fluid_layers_chunk_eval: + chunk_eval ---------- .. autofunction:: paddle.fluid.layers.chunk_eval :noindex: +.. _api_fluid_layers_sequence_conv: + sequence_conv ------------- .. autofunction:: paddle.fluid.layers.sequence_conv :noindex: +.. _api_fluid_layers_conv2d: + conv2d ------ .. autofunction:: paddle.fluid.layers.conv2d :noindex: +.. _api_fluid_layers_conv3d: + +conv3d +------ + +.. autofunction:: paddle.fluid.layers.conv3d + :noindex: + +.. _api_fluid_layers_sequence_pool: + sequence_pool ------------- .. autofunction:: paddle.fluid.layers.sequence_pool :noindex: +.. _api_fluid_layers_sequence_softmax: + sequence_softmax ---------------- .. autofunction:: paddle.fluid.layers.sequence_softmax :noindex: +.. _api_fluid_layers_softmax: + softmax ------- .. autofunction:: paddle.fluid.layers.softmax :noindex: +.. _api_fluid_layers_pool2d: + pool2d ------ .. autofunction:: paddle.fluid.layers.pool2d :noindex: +.. _api_fluid_layers_pool3d: + +pool3d +------ + +.. autofunction:: paddle.fluid.layers.pool3d + :noindex: + +.. _api_fluid_layers_batch_norm: + batch_norm ---------- .. autofunction:: paddle.fluid.layers.batch_norm :noindex: +.. _api_fluid_layers_beam_search_decode: + beam_search_decode ------------------ .. autofunction:: paddle.fluid.layers.beam_search_decode :noindex: +.. _api_fluid_layers_conv2d_transpose: + conv2d_transpose ---------------- .. autofunction:: paddle.fluid.layers.conv2d_transpose :noindex: +.. _api_fluid_layers_conv3d_transpose: + +conv3d_transpose +---------------- + +.. autofunction:: paddle.fluid.layers.conv3d_transpose + :noindex: + +.. _api_fluid_layers_sequence_expand: + sequence_expand --------------- .. autofunction:: paddle.fluid.layers.sequence_expand :noindex: +.. _api_fluid_layers_lstm_unit: + lstm_unit --------- .. autofunction:: paddle.fluid.layers.lstm_unit :noindex: +.. _api_fluid_layers_reduce_sum: + reduce_sum ---------- .. autofunction:: paddle.fluid.layers.reduce_sum :noindex: +.. _api_fluid_layers_reduce_mean: + reduce_mean ----------- .. autofunction:: paddle.fluid.layers.reduce_mean :noindex: +.. _api_fluid_layers_reduce_max: + reduce_max ---------- .. autofunction:: paddle.fluid.layers.reduce_max :noindex: +.. _api_fluid_layers_reduce_min: + reduce_min ---------- .. autofunction:: paddle.fluid.layers.reduce_min :noindex: +.. _api_fluid_layers_reduce_prod: + reduce_prod ----------- .. autofunction:: paddle.fluid.layers.reduce_prod :noindex: +.. _api_fluid_layers_sequence_first_step: + sequence_first_step ------------------- .. autofunction:: paddle.fluid.layers.sequence_first_step :noindex: +.. _api_fluid_layers_sequence_last_step: + sequence_last_step ------------------ .. autofunction:: paddle.fluid.layers.sequence_last_step :noindex: +.. _api_fluid_layers_dropout: + dropout ------- .. autofunction:: paddle.fluid.layers.dropout :noindex: +.. _api_fluid_layers_split: + split ----- .. autofunction:: paddle.fluid.layers.split :noindex: +.. _api_fluid_layers_ctc_greedy_decoder: + ctc_greedy_decoder ------------------ .. autofunction:: paddle.fluid.layers.ctc_greedy_decoder :noindex: +.. _api_fluid_layers_edit_distance: + edit_distance ------------- .. autofunction:: paddle.fluid.layers.edit_distance :noindex: +.. _api_fluid_layers_l2_normalize: + l2_normalize ------------ .. autofunction:: paddle.fluid.layers.l2_normalize :noindex: +.. _api_fluid_layers_matmul: + matmul ------ .. autofunction:: paddle.fluid.layers.matmul :noindex: +.. _api_fluid_layers_topk: + topk ---- .. autofunction:: paddle.fluid.layers.topk :noindex: +.. _api_fluid_layers_warpctc: + warpctc ------- .. autofunction:: paddle.fluid.layers.warpctc :noindex: +.. _api_fluid_layers_sequence_reshape: + sequence_reshape ---------------- .. autofunction:: paddle.fluid.layers.sequence_reshape :noindex: +.. _api_fluid_layers_transpose: + transpose --------- .. autofunction:: paddle.fluid.layers.transpose :noindex: +.. _api_fluid_layers_im2sequence: + im2sequence ----------- .. autofunction:: paddle.fluid.layers.im2sequence :noindex: +.. _api_fluid_layers_nce: + nce --- .. autofunction:: paddle.fluid.layers.nce :noindex: +.. _api_fluid_layers_beam_search: + beam_search ----------- .. autofunction:: paddle.fluid.layers.beam_search :noindex: +.. _api_fluid_layers_row_conv: + row_conv -------- .. autofunction:: paddle.fluid.layers.row_conv :noindex: +.. _api_fluid_layers_multiplex: + multiplex --------- .. autofunction:: paddle.fluid.layers.multiplex :noindex: +.. _api_fluid_layers_layer_norm: + layer_norm ---------- .. autofunction:: paddle.fluid.layers.layer_norm :noindex: +.. _api_fluid_layers_softmax_with_cross_entropy: + softmax_with_cross_entropy -------------------------- .. autofunction:: paddle.fluid.layers.softmax_with_cross_entropy :noindex: +.. _api_fluid_layers_smooth_l1: + smooth_l1 --------- .. autofunction:: paddle.fluid.layers.smooth_l1 :noindex: +.. _api_fluid_layers_one_hot: + one_hot ------- .. autofunction:: paddle.fluid.layers.one_hot :noindex: +.. _api_fluid_layers_autoincreased_step_counter: + autoincreased_step_counter -------------------------- .. autofunction:: paddle.fluid.layers.autoincreased_step_counter :noindex: +.. _api_fluid_layers_reshape: + reshape ------- .. autofunction:: paddle.fluid.layers.reshape :noindex: +.. _api_fluid_layers_lod_reset: + lod_reset --------- .. autofunction:: paddle.fluid.layers.lod_reset :noindex: +.. _api_fluid_layers_lrn: + lrn --- .. autofunction:: paddle.fluid.layers.lrn :noindex: +.. _api_fluid_layers_pad: + pad --- .. autofunction:: paddle.fluid.layers.pad :noindex: +.. _api_fluid_layers_label_smooth: + label_smooth ------------ .. autofunction:: paddle.fluid.layers.label_smooth :noindex: +.. _api_fluid_layers_roi_pool: + roi_pool -------- .. autofunction:: paddle.fluid.layers.roi_pool :noindex: +.. _api_fluid_layers_dice_loss: + +dice_loss +--------- + +.. autofunction:: paddle.fluid.layers.dice_loss + :noindex: + +.. _api_fluid_layers_image_resize: + +image_resize +------------ + +.. autofunction:: paddle.fluid.layers.image_resize + :noindex: + +.. _api_fluid_layers_image_resize_short: + +image_resize_short +------------------ + +.. autofunction:: paddle.fluid.layers.image_resize_short + :noindex: + +.. _api_fluid_layers_resize_bilinear: + +resize_bilinear +--------------- + +.. autofunction:: paddle.fluid.layers.resize_bilinear + :noindex: + +.. _api_fluid_layers_gather: + +gather +------ + +.. autofunction:: paddle.fluid.layers.gather + :noindex: + +.. _api_fluid_layers_random_crop: + +random_crop +----------- + +.. autofunction:: paddle.fluid.layers.random_crop + :noindex: + +.. _api_fluid_layers_mean_iou: + +mean_iou +-------- + +.. autofunction:: paddle.fluid.layers.mean_iou + :noindex: + +.. _api_fluid_layers_relu: + +relu +---- + +.. autofunction:: paddle.fluid.layers.relu + :noindex: + +.. _api_fluid_layers_log: + +log +--- + +.. autofunction:: paddle.fluid.layers.log + :noindex: + +.. _api_fluid_layers_crop: + +crop +---- + +.. autofunction:: paddle.fluid.layers.crop + :noindex: ops === +.. _api_fluid_layers_mean: + mean ---- .. autofunction:: paddle.fluid.layers.mean :noindex: +.. _api_fluid_layers_mul: + mul --- .. autofunction:: paddle.fluid.layers.mul :noindex: +.. _api_fluid_layers_scale: + scale ----- .. autofunction:: paddle.fluid.layers.scale :noindex: +.. _api_fluid_layers_sigmoid_cross_entropy_with_logits: + sigmoid_cross_entropy_with_logits --------------------------------- .. autofunction:: paddle.fluid.layers.sigmoid_cross_entropy_with_logits :noindex: +.. _api_fluid_layers_elementwise_add: + elementwise_add --------------- .. autofunction:: paddle.fluid.layers.elementwise_add :noindex: +.. _api_fluid_layers_elementwise_div: + elementwise_div --------------- .. autofunction:: paddle.fluid.layers.elementwise_div :noindex: +.. _api_fluid_layers_elementwise_sub: + elementwise_sub --------------- .. autofunction:: paddle.fluid.layers.elementwise_sub :noindex: +.. _api_fluid_layers_elementwise_mul: + elementwise_mul --------------- .. autofunction:: paddle.fluid.layers.elementwise_mul :noindex: +.. _api_fluid_layers_elementwise_max: + elementwise_max --------------- .. autofunction:: paddle.fluid.layers.elementwise_max :noindex: +.. _api_fluid_layers_elementwise_min: + elementwise_min --------------- .. autofunction:: paddle.fluid.layers.elementwise_min :noindex: +.. _api_fluid_layers_elementwise_pow: + elementwise_pow --------------- .. autofunction:: paddle.fluid.layers.elementwise_pow :noindex: +.. _api_fluid_layers_clip: + clip ---- .. autofunction:: paddle.fluid.layers.clip :noindex: +.. _api_fluid_layers_clip_by_norm: + clip_by_norm ------------ .. autofunction:: paddle.fluid.layers.clip_by_norm :noindex: +.. _api_fluid_layers_logical_and: + logical_and ----------- .. autofunction:: paddle.fluid.layers.logical_and :noindex: +.. _api_fluid_layers_logical_or: + logical_or ---------- .. autofunction:: paddle.fluid.layers.logical_or :noindex: +.. _api_fluid_layers_logical_xor: + logical_xor ----------- .. autofunction:: paddle.fluid.layers.logical_xor :noindex: +.. _api_fluid_layers_logical_not: + logical_not ----------- .. autofunction:: paddle.fluid.layers.logical_not :noindex: -uniform_random --------------- - -.. autofunction:: paddle.fluid.layers.uniform_random - :noindex: +.. _api_fluid_layers_uniform_random_batch_size_like: uniform_random_batch_size_like ------------------------------ @@ -712,23 +1065,23 @@ uniform_random_batch_size_like .. autofunction:: paddle.fluid.layers.uniform_random_batch_size_like :noindex: +.. _api_fluid_layers_gaussian_random: + gaussian_random --------------- .. autofunction:: paddle.fluid.layers.gaussian_random :noindex: +.. _api_fluid_layers_gaussian_random_batch_size_like: + gaussian_random_batch_size_like ------------------------------- .. autofunction:: paddle.fluid.layers.gaussian_random_batch_size_like :noindex: -cumsum ------- - -.. autofunction:: paddle.fluid.layers.cumsum - :noindex: +.. _api_fluid_layers_scatter: scatter ------- @@ -736,107 +1089,167 @@ scatter .. autofunction:: paddle.fluid.layers.scatter :noindex: +.. _api_fluid_layers_sum: + sum --- .. autofunction:: paddle.fluid.layers.sum :noindex: +.. _api_fluid_layers_slice: + +slice +----- + +.. autofunction:: paddle.fluid.layers.slice + :noindex: + +.. _api_fluid_layers_polygon_box_transform: + +polygon_box_transform +--------------------- + +.. autofunction:: paddle.fluid.layers.polygon_box_transform + :noindex: + +.. _api_fluid_layers_shape: + +shape +----- + +.. autofunction:: paddle.fluid.layers.shape + :noindex: + +.. _api_fluid_layers_iou_similarity: + +iou_similarity +-------------- + +.. autofunction:: paddle.fluid.layers.iou_similarity + :noindex: + +.. _api_fluid_layers_maxout: + +maxout +------ + +.. autofunction:: paddle.fluid.layers.maxout + :noindex: + +.. _api_fluid_layers_sigmoid: + sigmoid ------- .. autofunction:: paddle.fluid.layers.sigmoid :noindex: +.. _api_fluid_layers_logsigmoid: + logsigmoid ---------- .. autofunction:: paddle.fluid.layers.logsigmoid :noindex: +.. _api_fluid_layers_exp: + exp --- .. autofunction:: paddle.fluid.layers.exp :noindex: -relu ----- - -.. autofunction:: paddle.fluid.layers.relu - :noindex: - -tanh +.. _api_fluid_layers_tanh: + +tanh ---- .. autofunction:: paddle.fluid.layers.tanh :noindex: +.. _api_fluid_layers_tanh_shrink: + tanh_shrink ----------- .. autofunction:: paddle.fluid.layers.tanh_shrink :noindex: +.. _api_fluid_layers_softshrink: + softshrink ---------- .. autofunction:: paddle.fluid.layers.softshrink :noindex: +.. _api_fluid_layers_sqrt: + sqrt ---- .. autofunction:: paddle.fluid.layers.sqrt :noindex: +.. _api_fluid_layers_abs: + abs --- .. autofunction:: paddle.fluid.layers.abs :noindex: +.. _api_fluid_layers_ceil: + ceil ---- .. autofunction:: paddle.fluid.layers.ceil :noindex: +.. _api_fluid_layers_floor: + floor ----- .. autofunction:: paddle.fluid.layers.floor :noindex: +.. _api_fluid_layers_cos: + cos --- .. autofunction:: paddle.fluid.layers.cos :noindex: +.. _api_fluid_layers_sin: + sin --- .. autofunction:: paddle.fluid.layers.sin :noindex: +.. _api_fluid_layers_round: + round ----- .. autofunction:: paddle.fluid.layers.round :noindex: +.. _api_fluid_layers_reciprocal: + reciprocal ---------- .. autofunction:: paddle.fluid.layers.reciprocal :noindex: -log ---- - -.. autofunction:: paddle.fluid.layers.log - :noindex: +.. _api_fluid_layers_square: square ------ @@ -844,168 +1257,522 @@ square .. autofunction:: paddle.fluid.layers.square :noindex: +.. _api_fluid_layers_softplus: + softplus -------- .. autofunction:: paddle.fluid.layers.softplus :noindex: +.. _api_fluid_layers_softsign: + softsign -------- .. autofunction:: paddle.fluid.layers.softsign :noindex: +.. _api_fluid_layers_brelu: + brelu ----- .. autofunction:: paddle.fluid.layers.brelu :noindex: +.. _api_fluid_layers_leaky_relu: + leaky_relu ---------- .. autofunction:: paddle.fluid.layers.leaky_relu :noindex: +.. _api_fluid_layers_soft_relu: + soft_relu --------- .. autofunction:: paddle.fluid.layers.soft_relu :noindex: +.. _api_fluid_layers_elu: + elu --- .. autofunction:: paddle.fluid.layers.elu :noindex: +.. _api_fluid_layers_relu6: + relu6 ----- .. autofunction:: paddle.fluid.layers.relu6 :noindex: +.. _api_fluid_layers_pow: + pow --- .. autofunction:: paddle.fluid.layers.pow :noindex: +.. _api_fluid_layers_stanh: + stanh ----- .. autofunction:: paddle.fluid.layers.stanh :noindex: +.. _api_fluid_layers_hard_sigmoid: + +hard_sigmoid +------------ + +.. autofunction:: paddle.fluid.layers.hard_sigmoid + :noindex: + +.. _api_fluid_layers_swish: + +swish +----- + +.. autofunction:: paddle.fluid.layers.swish + :noindex: + +.. _api_fluid_layers_uniform_random: + +uniform_random +-------------- + +.. autofunction:: paddle.fluid.layers.uniform_random + :noindex: + +.. _api_fluid_layers_hard_shrink: + hard_shrink ----------- .. autofunction:: paddle.fluid.layers.hard_shrink :noindex: +.. _api_fluid_layers_cumsum: + +cumsum +------ + +.. autofunction:: paddle.fluid.layers.cumsum + :noindex: + +.. _api_fluid_layers_thresholded_relu: + thresholded_relu ---------------- .. autofunction:: paddle.fluid.layers.thresholded_relu :noindex: -hard_sigmoid ------------- +tensor +====== -.. autofunction:: paddle.fluid.layers.hard_sigmoid +.. _api_fluid_layers_create_tensor: + +create_tensor +------------- + +.. autofunction:: paddle.fluid.layers.create_tensor :noindex: -swish +.. _api_fluid_layers_create_parameter: + +create_parameter +---------------- + +.. autofunction:: paddle.fluid.layers.create_parameter + :noindex: + +.. _api_fluid_layers_create_global_var: + +create_global_var +----------------- + +.. autofunction:: paddle.fluid.layers.create_global_var + :noindex: + +.. _api_fluid_layers_cast: + +cast +---- + +.. autofunction:: paddle.fluid.layers.cast + :noindex: + +.. _api_fluid_layers_concat: + +concat +------ + +.. autofunction:: paddle.fluid.layers.concat + :noindex: + +.. _api_fluid_layers_sums: + +sums +---- + +.. autofunction:: paddle.fluid.layers.sums + :noindex: + +.. _api_fluid_layers_assign: + +assign +------ + +.. autofunction:: paddle.fluid.layers.assign + :noindex: + +.. _api_fluid_layers_fill_constant_batch_size_like: + +fill_constant_batch_size_like +----------------------------- + +.. autofunction:: paddle.fluid.layers.fill_constant_batch_size_like + :noindex: + +.. _api_fluid_layers_fill_constant: + +fill_constant +------------- + +.. autofunction:: paddle.fluid.layers.fill_constant + :noindex: + +.. _api_fluid_layers_argmin: + +argmin +------ + +.. autofunction:: paddle.fluid.layers.argmin + :noindex: + +.. _api_fluid_layers_argmax: + +argmax +------ + +.. autofunction:: paddle.fluid.layers.argmax + :noindex: + +.. _api_fluid_layers_argsort: + +argsort +------- + +.. autofunction:: paddle.fluid.layers.argsort + :noindex: + +.. _api_fluid_layers_ones: + +ones +---- + +.. autofunction:: paddle.fluid.layers.ones + :noindex: + +.. _api_fluid_layers_zeros: + +zeros ----- -.. autofunction:: paddle.fluid.layers.swish +.. autofunction:: paddle.fluid.layers.zeros + :noindex: + +.. _api_fluid_layers_reverse: + +reverse +------- + +.. autofunction:: paddle.fluid.layers.reverse + :noindex: + +learning_rate_scheduler +======================= + +.. _api_fluid_layers_exponential_decay: + +exponential_decay +----------------- + +.. autofunction:: paddle.fluid.layers.exponential_decay + :noindex: + +.. _api_fluid_layers_natural_exp_decay: + +natural_exp_decay +----------------- + +.. autofunction:: paddle.fluid.layers.natural_exp_decay + :noindex: + +.. _api_fluid_layers_inverse_time_decay: + +inverse_time_decay +------------------ + +.. autofunction:: paddle.fluid.layers.inverse_time_decay + :noindex: + +.. _api_fluid_layers_polynomial_decay: + +polynomial_decay +---------------- + +.. autofunction:: paddle.fluid.layers.polynomial_decay + :noindex: + +.. _api_fluid_layers_piecewise_decay: + +piecewise_decay +--------------- + +.. autofunction:: paddle.fluid.layers.piecewise_decay + :noindex: + +.. _api_fluid_layers_noam_decay: + +noam_decay +---------- + +.. autofunction:: paddle.fluid.layers.noam_decay + :noindex: + +.. _api_fluid_layers_append_LARS: + +append_LARS +----------- + +.. autofunction:: paddle.fluid.layers.append_LARS + :noindex: + +detection +========= + +.. _api_fluid_layers_prior_box: + +prior_box +--------- + +.. autofunction:: paddle.fluid.layers.prior_box + :noindex: + +.. _api_fluid_layers_multi_box_head: + +multi_box_head +-------------- + +.. autofunction:: paddle.fluid.layers.multi_box_head + :noindex: + +.. _api_fluid_layers_bipartite_match: + +bipartite_match +--------------- + +.. autofunction:: paddle.fluid.layers.bipartite_match + :noindex: + +.. _api_fluid_layers_target_assign: + +target_assign +------------- + +.. autofunction:: paddle.fluid.layers.target_assign + :noindex: + +.. _api_fluid_layers_detection_output: + +detection_output +---------------- + +.. autofunction:: paddle.fluid.layers.detection_output + :noindex: + +.. _api_fluid_layers_ssd_loss: + +ssd_loss +-------- + +.. autofunction:: paddle.fluid.layers.ssd_loss + :noindex: + +.. _api_fluid_layers_detection_map: + +detection_map +------------- + +.. autofunction:: paddle.fluid.layers.detection_map + :noindex: + +.. _api_fluid_layers_iou_similarity: + +iou_similarity +-------------- + +.. autofunction:: paddle.fluid.layers.iou_similarity + :noindex: + +.. _api_fluid_layers_box_coder: + +box_coder +--------- + +.. autofunction:: paddle.fluid.layers.box_coder + :noindex: + +metric_op +========= + +.. _api_fluid_layers_accuracy: + +accuracy +-------- + +.. autofunction:: paddle.fluid.layers.accuracy + :noindex: + +.. _api_fluid_layers_auc: + +auc +--- + +.. autofunction:: paddle.fluid.layers.auc :noindex: tensor ====== +.. _api_fluid_layers_create_tensor: + create_tensor ------------- .. autofunction:: paddle.fluid.layers.create_tensor :noindex: +.. _api_fluid_layers_create_parameter: + create_parameter ---------------- .. autofunction:: paddle.fluid.layers.create_parameter :noindex: +.. _api_fluid_layers_create_global_var: + create_global_var ----------------- .. autofunction:: paddle.fluid.layers.create_global_var :noindex: +.. _api_fluid_layers_cast: + cast ---- .. autofunction:: paddle.fluid.layers.cast :noindex: +.. _api_fluid_layers_concat: + concat ------ .. autofunction:: paddle.fluid.layers.concat :noindex: +.. _api_fluid_layers_sums: + sums ---- .. autofunction:: paddle.fluid.layers.sums :noindex: +.. _api_fluid_layers_assign: + assign ------ .. autofunction:: paddle.fluid.layers.assign :noindex: +.. _api_fluid_layers_fill_constant_batch_size_like: + fill_constant_batch_size_like ----------------------------- .. autofunction:: paddle.fluid.layers.fill_constant_batch_size_like :noindex: +.. _api_fluid_layers_fill_constant: + fill_constant ------------- .. autofunction:: paddle.fluid.layers.fill_constant :noindex: +.. _api_fluid_layers_argmin: + +argmin +------ + +.. autofunction:: paddle.fluid.layers.argmin + :noindex: + +.. _api_fluid_layers_argmax: + +argmax +------ + +.. autofunction:: paddle.fluid.layers.argmax + :noindex: + +.. _api_fluid_layers_ones: + ones ---- .. autofunction:: paddle.fluid.layers.ones :noindex: +.. _api_fluid_layers_zeros: + zeros ----- .. autofunction:: paddle.fluid.layers.zeros :noindex: -topk ----- +.. _api_fluid_layers_reverse: -.. autofunction:: paddle.fluid.layers.topk - :noindex: - -dice_loss ----- +reverse +------- -.. autofunction:: paddle.fluid.layers.dice_loss +.. autofunction:: paddle.fluid.layers.reverse :noindex: -upsampling_bilinear2d -____ +.. _api_fluid_layers_rank_loss: + +rank_loss +------- -.. autofunction:: paddle.fluid.layers.upsampling_bilinear2d +.. autofunction:: paddle.fluid.layers.rank_loss :noindex: diff --git a/doc/fluid/api/metrics.rst b/doc/fluid/api/metrics.rst index ddf07775d7ea293acd421b8549d03b277ff0611d..0f54b2e2eb7ead353215c5dbd529293794e37123 100644 --- a/doc/fluid/api/metrics.rst +++ b/doc/fluid/api/metrics.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -======= -metrics -======= +============= +fluid.metrics +============= + +.. _api_fluid_metrics_MetricBase: MetricBase ---------- @@ -12,6 +14,8 @@ MetricBase :members: :noindex: +.. _api_fluid_metrics_CompositeMetric: + CompositeMetric --------------- @@ -19,6 +23,26 @@ CompositeMetric :members: :noindex: +.. _api_fluid_metrics_Precision: + +Precision +--------- + +.. autoclass:: paddle.fluid.metrics.Precision + :members: + :noindex: + +.. _api_fluid_metrics_Recall: + +Recall +------ + +.. autoclass:: paddle.fluid.metrics.Recall + :members: + :noindex: + +.. _api_fluid_metrics_Accuracy: + Accuracy -------- @@ -26,6 +50,8 @@ Accuracy :members: :noindex: +.. _api_fluid_metrics_ChunkEvaluator: + ChunkEvaluator -------------- @@ -33,6 +59,8 @@ ChunkEvaluator :members: :noindex: +.. _api_fluid_metrics_EditDistance: + EditDistance ------------ @@ -40,6 +68,8 @@ EditDistance :members: :noindex: +.. _api_fluid_metrics_DetectionMAP: + DetectionMAP ------------ @@ -47,6 +77,8 @@ DetectionMAP :members: :noindex: +.. _api_fluid_metrics_Auc: + Auc --- diff --git a/doc/fluid/api/nets.rst b/doc/fluid/api/nets.rst index 7ae3187304f386a08c5cb8a4ba093423a58a7f36..059733af18517257b6821d95fd628a9e13e6e98e 100644 --- a/doc/fluid/api/nets.rst +++ b/doc/fluid/api/nets.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -==== -nets -==== +========== +fluid.nets +========== + +.. _api_fluid_nets_simple_img_conv_pool: simple_img_conv_pool -------------------- @@ -11,18 +13,24 @@ simple_img_conv_pool .. autofunction:: paddle.fluid.nets.simple_img_conv_pool :noindex: +.. _api_fluid_nets_sequence_conv_pool: + sequence_conv_pool ------------------ .. autofunction:: paddle.fluid.nets.sequence_conv_pool :noindex: +.. _api_fluid_nets_glu: + glu --- .. autofunction:: paddle.fluid.nets.glu :noindex: +.. _api_fluid_nets_scaled_dot_product_attention: + scaled_dot_product_attention ---------------------------- diff --git a/doc/fluid/api/optimizer.rst b/doc/fluid/api/optimizer.rst index df2bd2eace52e78805433bea320f5de95d45bfc7..8d792120f2f16a8c92606b343eb4c3d4368bed14 100644 --- a/doc/fluid/api/optimizer.rst +++ b/doc/fluid/api/optimizer.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -========= -optimizer -========= +=============== +fluid.optimizer +=============== + +.. _api_fluid_optimizer_SGD: SGD --- @@ -12,6 +14,8 @@ SGD :members: :noindex: +.. _api_fluid_optimizer_Momentum: + Momentum -------- @@ -19,6 +23,8 @@ Momentum :members: :noindex: +.. _api_fluid_optimizer_Adagrad: + Adagrad ------- @@ -26,6 +32,8 @@ Adagrad :members: :noindex: +.. _api_fluid_optimizer_Adam: + Adam ---- @@ -33,6 +41,8 @@ Adam :members: :noindex: +.. _api_fluid_optimizer_Adamax: + Adamax ------ @@ -40,6 +50,8 @@ Adamax :members: :noindex: +.. _api_fluid_optimizer_DecayedAdagrad: + DecayedAdagrad -------------- @@ -47,27 +59,16 @@ DecayedAdagrad :members: :noindex: -Adadelta ------------------ - -.. autoclass:: paddle.fluid.optimizer.Adadelta - :members: - :noindex: - -RMSProp ------------------ +.. _api_fluid_optimizer_Ftrl: -.. autoclass:: paddle.fluid.optimizer.RMSProp - :members: - :noindex: - -ModelAverage ------------------ +Ftrl +---- -.. autoclass:: paddle.fluid.optimizer.ModelAverage +.. autoclass:: paddle.fluid.optimizer.Ftrl :members: :noindex: +.. _api_fluid_optimizer_SGDOptimizer: SGDOptimizer ------------ @@ -76,6 +77,8 @@ SGDOptimizer :members: :noindex: +.. _api_fluid_optimizer_MomentumOptimizer: + MomentumOptimizer ----------------- @@ -83,6 +86,8 @@ MomentumOptimizer :members: :noindex: +.. _api_fluid_optimizer_AdagradOptimizer: + AdagradOptimizer ---------------- @@ -90,6 +95,8 @@ AdagradOptimizer :members: :noindex: +.. _api_fluid_optimizer_AdamOptimizer: + AdamOptimizer ------------- @@ -97,6 +104,8 @@ AdamOptimizer :members: :noindex: +.. _api_fluid_optimizer_AdamaxOptimizer: + AdamaxOptimizer --------------- @@ -104,6 +113,8 @@ AdamaxOptimizer :members: :noindex: +.. _api_fluid_optimizer_DecayedAdagradOptimizer: + DecayedAdagradOptimizer ----------------------- @@ -111,25 +122,57 @@ DecayedAdagradOptimizer :members: :noindex: +.. _api_fluid_optimizer_RMSPropOptimizer: -AdadeltaOptimizer ------------------ +RMSPropOptimizer +---------------- -.. autoclass:: paddle.fluid.optimizer.AdadeltaOptimizer +.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer :members: :noindex: +.. _api_fluid_optimizer_FtrlOptimizer: -RMSPropOptimizer ------------------ +FtrlOptimizer +------------- -.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer +.. autoclass:: paddle.fluid.optimizer.FtrlOptimizer + :members: + :noindex: + +.. _api_fluid_optimizer_Adadelta: + +Adadelta +-------- + +.. autoclass:: paddle.fluid.optimizer.Adadelta + :members: + :noindex: + +.. _api_fluid_optimizer_ModelAverage: + +ModelAverage +------------ + +.. autoclass:: paddle.fluid.optimizer.ModelAverage :members: :noindex: - + +.. _api_fluid_optimizer_Optimizer: + Optimizer --------- .. autoclass:: paddle.fluid.optimizer.Optimizer :members: :noindex: + +.. _api_fluid_optimizer_RMSPropOptimizer: + +RMSPropOptimizer +---------------- + +.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer + :members: + :noindex: + diff --git a/doc/fluid/api/param_attr.rst b/doc/fluid/api/param_attr.rst index 8e4ddb2b0492d0fcfcade199fdd6dfe43faa7075..33035bbc7ca5c8d000adeaf1cb79806a3ea64604 100644 --- a/doc/fluid/api/param_attr.rst +++ b/doc/fluid/api/param_attr.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -========== -param_attr -========== +================ +fluid.param_attr +================ + +.. _api_fluid_param_attr_ParamAttr: ParamAttr --------- @@ -12,6 +14,8 @@ ParamAttr :members: :noindex: +.. _api_fluid_param_attr_WeightNormParamAttr: + WeightNormParamAttr ------------------- diff --git a/doc/fluid/api/profiler.rst b/doc/fluid/api/profiler.rst index 74d102dcb0db35766c34e3d14939a8aa5861686b..c750a2d588df56728ac7f73051ab7a9e44dee232 100644 --- a/doc/fluid/api/profiler.rst +++ b/doc/fluid/api/profiler.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -======== -profiler -======== +============== +fluid.profiler +============== + +.. _api_fluid_profiler_cuda_profiler: cuda_profiler ------------- @@ -11,15 +13,35 @@ cuda_profiler .. autofunction:: paddle.fluid.profiler.cuda_profiler :noindex: +.. _api_fluid_profiler_reset_profiler: + reset_profiler -------------- .. autofunction:: paddle.fluid.profiler.reset_profiler :noindex: +.. _api_fluid_profiler_profiler: + profiler -------- .. autofunction:: paddle.fluid.profiler.profiler :noindex: +.. _api_fluid_profiler_start_profiler: + +start_profiler +-------------- + +.. autofunction:: paddle.fluid.profiler.start_profiler + :noindex: + +.. _api_fluid_profiler_stop_profiler: + +stop_profiler +------------- + +.. autofunction:: paddle.fluid.profiler.stop_profiler + :noindex: + diff --git a/doc/fluid/api/recordio_writer.rst b/doc/fluid/api/recordio_writer.rst new file mode 100644 index 0000000000000000000000000000000000000000..f0c12fd115478a29fbd178b533b7490b2f663717 --- /dev/null +++ b/doc/fluid/api/recordio_writer.rst @@ -0,0 +1,23 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +===================== +fluid.recordio_writer +===================== + +.. _api_fluid_recordio_writer_convert_reader_to_recordio_file: + +convert_reader_to_recordio_file +------------------------------- + +.. autofunction:: paddle.fluid.recordio_writer.convert_reader_to_recordio_file + :noindex: + +.. _api_fluid_recordio_writer_convert_reader_to_recordio_files: + +convert_reader_to_recordio_files +-------------------------------- + +.. autofunction:: paddle.fluid.recordio_writer.convert_reader_to_recordio_files + :noindex: + diff --git a/doc/fluid/api/regularizer.rst b/doc/fluid/api/regularizer.rst index 756bc53baa0625aef48dad0c35e7ae57421a70d0..987eaea903520d91c284c8da7a8cb066a1648069 100644 --- a/doc/fluid/api/regularizer.rst +++ b/doc/fluid/api/regularizer.rst @@ -1,9 +1,11 @@ .. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` !DO NOT EDIT THIS FILE MANUALLY! -=========== -regularizer -=========== +================= +fluid.regularizer +================= + +.. _api_fluid_regularizer_append_regularization_ops: append_regularization_ops ------------------------- @@ -11,12 +13,7 @@ append_regularization_ops .. autofunction:: paddle.fluid.regularizer.append_regularization_ops :noindex: -WeightDecayRegularizer ----------------------- - -.. autoclass:: paddle.fluid.regularizer.WeightDecayRegularizer - :members: - :noindex: +.. _api_fluid_regularizer_L1Decay: L1Decay ------- @@ -25,6 +22,8 @@ L1Decay :members: :noindex: +.. _api_fluid_regularizer_L2Decay: + L2Decay ------- @@ -32,6 +31,8 @@ L2Decay :members: :noindex: +.. _api_fluid_regularizer_L1DecayRegularizer: + L1DecayRegularizer ------------------ @@ -39,6 +40,8 @@ L1DecayRegularizer :members: :noindex: +.. _api_fluid_regularizer_L2DecayRegularizer: + L2DecayRegularizer ------------------ diff --git a/doc/fluid/api/transpiler.rst b/doc/fluid/api/transpiler.rst new file mode 100644 index 0000000000000000000000000000000000000000..d2ac04f1449c32cb414cea1b76d7469bbe9ccb85 --- /dev/null +++ b/doc/fluid/api/transpiler.rst @@ -0,0 +1,59 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +================ +fluid.transpiler +================ + +.. _api_fluid_transpiler_DistributeTranspiler: + +DistributeTranspiler +-------------------- + +.. autoclass:: paddle.fluid.transpiler.DistributeTranspiler + :members: + :noindex: + +.. _api_fluid_transpiler_InferenceTranspiler: + +InferenceTranspiler +------------------- + +.. autoclass:: paddle.fluid.transpiler.InferenceTranspiler + :members: + :noindex: + +.. _api_fluid_transpiler_memory_optimize: + +memory_optimize +--------------- + +.. autofunction:: paddle.fluid.transpiler.memory_optimize + :noindex: + +.. _api_fluid_transpiler_release_memory: + +release_memory +-------------- + +.. autofunction:: paddle.fluid.transpiler.release_memory + :noindex: + +.. _api_fluid_transpiler_HashName: + +HashName +-------- + +.. autoclass:: paddle.fluid.transpiler.HashName + :members: + :noindex: + +.. _api_fluid_transpiler_RoundRobin: + +RoundRobin +---------- + +.. autoclass:: paddle.fluid.transpiler.RoundRobin + :members: + :noindex: + diff --git a/doc/fluid/design/concepts/lod_tensor.md b/doc/fluid/design/concepts/lod_tensor.md index d606d7a790b4b0dc18553f2220d39cec8aa619ec..748488f6d5f2f1272e87b89047570632418da8dc 100644 --- a/doc/fluid/design/concepts/lod_tensor.md +++ b/doc/fluid/design/concepts/lod_tensor.md @@ -173,6 +173,7 @@ are transformed into offsets of elements/words as follows: ## Slicing of LoD Tensors + When we use the above 2-level LoD Tensor as the input to a nested-RNN, we need to retrieve certain sequences. Here we define the sequence identified by branch as the **-slice**. For example, the <2>-slice of above example is @@ -189,3 +190,22 @@ and the <2,0>-slice of above slice is 10 12 || ``` + +## Length Representation vs Offset Representation + +The offset representation is an implementation-oriented decision and it makes understanding the idea behind LoDTensor difficult. +Hence, we encapsulate this implementation detail in C++ and expose the original length representation in our Python API. +Specifically, we call this length representation `recursive_sequence_lengths` and users can use the following code to set or get the `recursive_sequence_lengths` of a LoDTensor in Python: +```Python +# length representation of lod called recursive_sequence_lengths +recursive_seq_lens = [[3, 1, 2], [2, 2, 1, 3, 1, 2]] +# Create a LoDTensor that has the above recursive_sequence_lengths info. +# This recursive_sequence_lengths will be converted to an offset representation of LoD in the C++ implementation under the hood. +tensor = fluid.LoDTensor(lod) + +# Set/Change the recursive_sequence_lengths info of LoDTensor +tensor.set_recursive_sequence_lengths([[3, 1, 2]]) +# Get the recursive_sequence_lengths info of a LoDTensor (the offset-based LoD representation stored in C++ will be converted +# back to length-based recursive_sequence_lengths), new_recursive_seq_lens = [[3, 1, 2]] +new_recursive_seq_lens = tensor.recursive_sequence_lengths() +``` diff --git a/doc/fluid/design/concepts/python_data_feeding.md b/doc/fluid/design/concepts/python_data_feeding.md new file mode 100644 index 0000000000000000000000000000000000000000..dffee8e02bacbc99bdfa8c54f1a146de340ad778 --- /dev/null +++ b/doc/fluid/design/concepts/python_data_feeding.md @@ -0,0 +1,130 @@ +# Python Data Feeding + +In the former implementation of Paddle Fluid, there are two ways to feed data: + +- Use `reader_op` in backend C++ side. This method only supports data feeding from recordio files and random data generators, but supports many kinds of `decorated_readers`. For examples, `double_buffer_reader` uses two threads to achieve better performance: one for time-consuming I/O operations, and the other for `Executor::Run()`. See [C++ Data Feeding](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/cpp_data_feeding.md) for details. + +- Feed data directly using `DataFeeder.feed()` in Python codes. It is more flexible than the first way. Many kinds of preprocessing steps can be performed before feeding using Python or any other languages, instead of adding many uncommon `operators` in C++ side. But this method is less efficient: the program cannot read the next mini-batch data before `Executor::Run()` ends. Moreover, `decorated_readers` such as `double_buffer_reader` cannot be used for better performance. + +In this document, we design a Python Data Feeding process combining the efficiency of the first way and the flexibility of the second way. A data queue `LoDTensorBlockingQueue` is designed to be shared by the Python and C++ side, while `LoDTensorArray` is pushed into the queue in Python side and `reader_op` in C++ side reads out the data from the queue. + + +## Design of LoDTensorBlockingQueue +`LoDTensorBlockingQueue` is a blocking queue with a fixed `capacity` and accepts `std::vector` with shapes indicated by `dims`. Since `LoDTensorBlockingQueue` must be constructed using `capacity` and `dims`, it cannot be a `Variable` type. Therefore, a `LoDTensorBlockingQueueHolder` is designed to defer construction of `LoDTensorBlockingQueue`. + +```C++ +class LoDTensorBlockingQueueHolder; + +class LoDTensorBlockingQueue { + friend class LoDTensorBlockingQueueHolder; + private: + // `LoDTensorBlockingQueue` can only be constructed by + // `LoDTensorBlockingQueueHolder::InitOnce()` + LoDTensorBlockingQueue(size_t capacity, const std::vector& dims); + + public: + size_t Size() const { return queue_.Size(); } // Get the current size of the queue + + size_t Cap() const { return queue_.Cap(); }// Get the capacity of the queue + + void Close() { return queue_.Close(); } + + bool IsClosed() const { return queue_.IsClosed(); } + + // Block if Size() == Cap() + // Return false only when queue_.IsClosed() == true + bool Push(const std::vector &lod_tensor_vec); + + // Block if Size() == 0. + // *Success == false when queue_.IsClosed() == true + std::vector Pop(bool *success = nullptr); + + private: + // Use reader::BlockingQueue as the inner data structure + BlockingQueue> queue_; + std::vector dims_; +}; + +class LoDTensorBlockingQueueHolder { + public: + // Call the constructor of `LoDTensorBlockingQueue` to create queue_ + // `InitOnce` can only called once, otherwise an exception would raise + void InitOnce(size_t capacity, const std::vector& dims) { + PADDLE_ENFORCE(queue_ == nullptr); + queue_.reset(new LoDTensorBlockingQueue(capacity, dims)); + } + + const std::shared_ptr& GetQueue() const { return queue_; } + + private: + std::shared_ptr queue_; +}; +``` + +There are some major things that must be concerned: +- `LoDTensorBlockingQueueHolder` should be a `Variable` in global scope, so that `reader_op` can find it when reading data. +- A `Variable` of `LoDTensorBlockingQueueHolder` but not `VarDesc` must be created in Python code before `Executor::Run()` so that `Executor::Run()` can get the feeding data when it is called. +- `Create_reader_op` should accept the name of the `LoDTensorBlockingQueueHolder` variable as an input. + + +## Release of the GIL in pybind +`Pybind11::gil_scoped_release` is used to release GIL (Global Interpreter Lock) when `LoDTensorBlockingQueue::Push()` or `Executor::Run()` method are invoked in Python side, making `LoDTensorBlockingQueue::Push()` and `Executor::Run()` run in parallel. + + +## Design of PyReader +`PyReader` is a reader which holds a `LoDTensorBlockingQueue` object. +```C++ +class PyReader : public ReaderBase { + public: + explicit PyReader(const std::shared_ptr& queue); + + void ReadNext(std::vector* out) override { + bool success; + *out = queue_->Pop(&success); + if (!success) out->clear(); + } + + void ReInit() override { return; } + + private: + std::shared_ptr queue_; +}; +``` + + +## Design of CreatePyReaderOp +`CreatePyReaderOp` is used to create the `PyReader` object. It requires an input `blocking_queue` which indicates the name of the `LoDTensorBlockingQueueHolder` variable. +```C++ +class CreatePyReaderOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + private: + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + auto* out = scope.FindVar(Output("Out")) + ->template GetMutable(); + if (out->Get() != nullptr) return; + + const std::string& queue_name = Input("blocking_queue"); + auto* queue_holder_var = scope.FindVar(queue_name); + PADDLE_ENFORCE(queue_holder_var != nullptr); + auto* queue_holder = queue_holder_var + ->template GetMutable(); + out->Reset(new PyReader(queue_holder->GetQueue())); + } +}; +``` + +## Design of Python codes +The design of Python codes are as follows. First, we construct a variable of `LoDTensorBlockingQueueHolder` and init it with given parameters, returning the `LoDTensorBlockingQueue` object after initialization. After that, a layer of `CreatePyReaderOp` is constructed and accepts the name of the `LoDTensorBlockingQueueHolder` variable. The `LoDTensorBlockingQueue` object and result of the layer are both returned. +```Python +def py_reader(capacity, shapes): + queue_name = unique_name.generate("lod_tensor_blocking_queue") + var = global_scope().var(feeder_name) # create LoDTensorBlockingQueueHolder Variable + feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) # init the queue + out = create_var() + create_py_reader_op_with_queue_name( + inputs={'blocking_queue': queue_name}, + outputs={'Out':[out]}) + return out, feed_queue +``` diff --git a/doc/fluid/design/concepts/var_desc.md b/doc/fluid/design/concepts/var_desc.md index 6750323c0167bf1efbde6ef4fd670e88a5aa502a..8db67f6703d142da71cf06bd4f7e2cb13556f9b0 100644 --- a/doc/fluid/design/concepts/var_desc.md +++ b/doc/fluid/design/concepts/var_desc.md @@ -35,7 +35,7 @@ The computation `Program` consists of nested `Blocks`. Each `Block` will consist ## Definition of VarType -A VarDesc should have a name, type and whether or not it is persistable. The are different kinds of variable types supported in PaddlePaddle, apart from the POD_Types like: `LOD_TENSOR`, `SELECTED_ROWS`, `FEED_MINIBATCH`, `FETCH_LIST`, `STEP_SCOPES`, `LOD_RANK_TABLE`, `LOD_TENSOR_ARRAY`, `PLACE_LIST`, `READER` and `CHANNEL`. These are declared inside `VarType`. A `VarDesc` then looks as the following: +A VarDesc should have a name, type and whether or not it is persistable. There are different kinds of variable types supported in PaddlePaddle, apart from the POD_Types like: `LOD_TENSOR`, `SELECTED_ROWS`, `FEED_MINIBATCH`, `FETCH_LIST`, `STEP_SCOPES`, `LOD_RANK_TABLE`, `LOD_TENSOR_ARRAY`, `PLACE_LIST`, `READER` and `CHANNEL`. These are declared inside `VarType`. A `VarDesc` then looks as the following: ```proto message VarDesc { diff --git a/doc/fluid/design/dist_train/dist_train_nccl2.md b/doc/fluid/design/dist_train/dist_train_nccl2.md new file mode 100644 index 0000000000000000000000000000000000000000..aa7455ec5de0d46d7c2b0cef3b7ebf4754af3cb1 --- /dev/null +++ b/doc/fluid/design/dist_train/dist_train_nccl2.md @@ -0,0 +1,35 @@ +# Distributed Training with NCCL2 + +We design a pattern that can enable training with `ParallelExecutor` and +using [NCCL2](https://developer.nvidia.com/nccl) as it's collective +communication library. + +In `ParallelExecutor` we can use `AllReduce` or `Reduce` and `Broadcast` +to do multi GPU training. And if we initialize NCCL2 communicators as +ranks in a distributed environment, we can simply run the `ParallelExecutor` +as a distributed program! The only thing that may be different than in +the single node version is that we need to broadcast the NCCL unique ID +to all the nodes, and initialize communicators using that ID, so NCCL2 +will know each other as ranks. + +To achieve this feature, we introduce a new operator: `gen_nccl_id` op, +so we are ***not*** "bind to" running NCCL2 with MPI, we can run it in +what ever platform you like. + +It have two running modes: + +1. Generate and broadcast mode, which should be used on trainer 0; +1. Listen and fetch mode, which should be used on trainers other than 0. + +In both two modes, this op can save the NCCL ID into current scope as a +persistable variable, Then we can insert this op at the end of +"startup program" of fluid, so that all workers can get the same ID to +initialize NCCL communicator objects. + + + +The above figure indicates the general process when training with NCCL2 +distributed. Each trainer have the number of communicators equal to the +number of GPUs, but the ranks should match the global ranks number: here +we have total 8 GPUs, so `nranks==8`, for each trainer, the ranks should +be from 0 ~ 3 on trainer 0 and 4 ~ 7 on trainer 1. diff --git a/doc/fluid/design/dist_train/distributed_lookup_table_design.md b/doc/fluid/design/dist_train/distributed_lookup_table_design.md index 988729138926f035750b59eb245dde82502a3ad2..e284e1ec5cdd18d0049ce3c1a8349bbe1248cb48 100644 --- a/doc/fluid/design/dist_train/distributed_lookup_table_design.md +++ b/doc/fluid/design/dist_train/distributed_lookup_table_design.md @@ -1,6 +1,6 @@ # Design Doc: Distributed Lookup Table Operator -A lookup table operator in PaddlePaddle where the table could be out +A distribute lookup table operator in PaddlePaddle where the table could be out of the memory of a computer. ## Background @@ -24,14 +24,14 @@ memory, so we'd need a distributed storage service, which supports the lookup of rows. The following figure illustrates the multiplication of x with two -non-zero elements, or say, two symbols, and a lookup table W: +non-zero elements, or say two symbols, and a lookup table W: ![lookup table](./src/lookup_table.png) ### The Backward Algorithm The backward algorithm computes W'(x) using W(x). W'(x) has the same -scale of size as W(x) and is much smaller than W. +the scale of size as W(x) and is much smaller than W. To optimize W given W', we can do simple SGD update: @@ -44,85 +44,46 @@ $$W = f(W, W')$$ The following figure illustrates the backward pass of the lookup operator: ![lookup table training](./src/lookup_table_training.png) -## Distributed Storage Service - -The forward algorithm requires a distributed storage service for W. -The backward algorithm prefers that the storage system can apply the -optimization algorithm on W. The following two sections describe two -solutions -- the former doesn't require that the storage service can -do optimization, the latter does. - -### Storage Service Doesn't Optimize - -In this design, we use highly-optimized distributed storage, e.g., -memcached, as the storage service, and we run the optimization -algorithm on parameter servers of PaddlePaddle. The following figure -illustrates the training process. - - - - - -Each trainer runs the forward and backward passes using their local -data: - -1. In the forward pass, when a trainer runs the forward algorithm of a - lookup operator, it retrieves W(x) from the storage service. -1. The trainer computes W'(x) in the backward pass using W(x). - -During the global update process: - -1. Each trainer uploads its W'(x) to parameter servers. -1. The parameter server runs the optimization algorithm, e.g., the - Adam optimization algorithm, which requires that - 1. The parameter server retrieves W(x) from memcached, and - 1. The parameter server pushes $\Delta W(x)=f(W(x), lambda \sum_j - W'(x))$ to memcached, where $f$ denotes the optimization - algorithm. - -### Storage Service Does Optimize - -This design is very similar to the above one, except that the -optimization algorithm $f$ runs on the storage service. - -- Pro: parameter servers do not retrieve W(x) from the storage - service, thus saves half network communication. -- Con: the storage service needs to be able to run the optimization - algorithm. - -## Conclusion - -Let us do the "storage service does not optimize" solution first, as a -baseline at least, because it is easier to use a well-optimized -distributed storage service like memcached. We can do the "storage -service does optimize" solution later or at the same time, which, if -implemented carefully, should have better performance than the former. +## Distributed Lookup Table +### Problem 1: The lookup table may be very large. + + In the condition like the search engine and recommendation system, the number of feature Id may be very large, say 100,000,000,000, then for a float value lookup table of size 8, the total size of the table is: + + ``` + 100,000,000,000 * 8 * 4(Bytes) = 2980.23 GB + ``` + +### Solution: Distributed storage + +1. Paddle use [SelectedRows](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/selected_rows.md) as the storage format for the lookup table, the lookup table parameter will be split to multi-machine according to the hash of the feature ID, and data will also be split and send to the same machine to prefetch the parameter. + +1. For common parameters, the trainer will get the whole parameter for training, but for the big lookup table, the trainer can not store the whole parameter. Because the input data feature is very sparse, every time we only need a few parameters for training, so we use `prefetch_op` to only prefetch the parameter needed to trainer. + +### Problem 2. The Id in the lookup table is not sure before training. + + The feature Id is calculated by the hash function because the feature data source is so large, we can not get all the Id before training. So we can not initialize the table before training. + +### Solution: Id auto growth + +At the beginning of training, paddle only malloc the memory for the lookup table at parameter server side, the Id and it's value will not be initialized. During training, when a parameter server received an Id, if it is already in the lookup table, it will return the existing parameter, if the Id does not exist, paddle will add it into the lookup table and initialize the value for it. + +### Problem 3: parameter load and save + +For common parameters, paddle use trainer to save and load them. But for distributed lookup table, trainer cannot do this because it's large size. + +### Solution: Parameter server side save and load + +Paddle support parameter server side save and load for distribute lookup table. Each machine of parameter servers will only save and load part of the whole table. + +## Architecture +The whole architecture of the distribute lookup table is as below: + +### Training steps: +1. Read a batch of data, the data is feature ids. +1. The input ids will be split by `split_ids_op` with the same hash function of the lookup table. +1. The `prefetch_op` use the split result to prefetch parameters back from the lookup table. +1. Run forward-backward to get the gradient of the lookup table. +1. `split_ids_op` split the gradient and then use `send_op` to the parameter server. +1. parameter server update the table with the received gradient. + +![distribute lookup table](./src/distributed_lookup_table.jpeg) diff --git a/doc/fluid/design/dist_train/src/distributed_lookup_table.graffle b/doc/fluid/design/dist_train/src/distributed_lookup_table.graffle new file mode 100644 index 0000000000000000000000000000000000000000..65dfdbbacd219739db6ddfdf243cc16c3c4e8d1e Binary files /dev/null and b/doc/fluid/design/dist_train/src/distributed_lookup_table.graffle differ diff --git a/doc/fluid/design/dist_train/src/distributed_lookup_table.jpeg b/doc/fluid/design/dist_train/src/distributed_lookup_table.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..5353a16fd329f62ff893d32706b9c3c0bcc46a07 Binary files /dev/null and b/doc/fluid/design/dist_train/src/distributed_lookup_table.jpeg differ diff --git a/doc/fluid/design/dist_train/src/fluid_lookup_remote_table.graffle b/doc/fluid/design/dist_train/src/fluid_lookup_remote_table.graffle new file mode 100644 index 0000000000000000000000000000000000000000..96ca6d48f43bd9f49c6861dab006e2037873db87 Binary files /dev/null and b/doc/fluid/design/dist_train/src/fluid_lookup_remote_table.graffle differ diff --git a/doc/fluid/design/dist_train/src/fluid_lookup_remote_table.png b/doc/fluid/design/dist_train/src/fluid_lookup_remote_table.png new file mode 100644 index 0000000000000000000000000000000000000000..afa25ab3b4e427bc595a855b12ab966478e01ed0 Binary files /dev/null and b/doc/fluid/design/dist_train/src/fluid_lookup_remote_table.png differ diff --git a/doc/fluid/design/dist_train/src/ncc2_design.graffle b/doc/fluid/design/dist_train/src/ncc2_design.graffle new file mode 100644 index 0000000000000000000000000000000000000000..7d2753bbb03bc28c7a0054bb0aa424deb072ffbf Binary files /dev/null and b/doc/fluid/design/dist_train/src/ncc2_design.graffle differ diff --git a/doc/fluid/design/dist_train/src/ncc2_design.png b/doc/fluid/design/dist_train/src/ncc2_design.png new file mode 100644 index 0000000000000000000000000000000000000000..da0d5ee81f5dfeb4ca1356601b0bb5870456e3d6 Binary files /dev/null and b/doc/fluid/design/dist_train/src/ncc2_design.png differ diff --git a/doc/fluid/design/ir/overview.md b/doc/fluid/design/ir/overview.md new file mode 100644 index 0000000000000000000000000000000000000000..83ef97c99efeaf27a27f93f0cd3857c0f1bc812e --- /dev/null +++ b/doc/fluid/design/ir/overview.md @@ -0,0 +1,185 @@ +## Motivation + +There is a `gap` between the `Program` defined by +user and the `Executable` that can be scheduled +efficiently on heterogeneous hardware, either locally +or distributedly. + +Usually, the `gap` is bridged by + +* A serious transformations with defined order. + +* These transformations usually involve +`insert, delete, clustering, split, dependency analysis`. + +* Has a simple way to verify and debug each transformation. + +* Flexible to add, remove or customize transformations to fit +the requirements of various algorithms (models) and hardware secenarios. + +Some other events also push us to a better unified pattern. + +* The deep learning framework is built around the concepts of graphs. +To leverage tools such as compilation (e.g. TVM and nGraph) or +cross-framework conversion (e.g. ONNX), we also need a intermediate +representation that can be connected to the rest of the ecosystem. + + +We need a unified pattern to naturally support the requirements +described above. The pattern should fit both training, inference +and other offline serielized model transformations. +Learned from LLVM and other deep learning framework, we draft the +design below. + + +## Design + +### Major Concepts + +#### Node + +`Node` represents an operation that performs some computation or +a variable that is input or output of operation. + +`Node`s are connected to other `Node`s via inputs and outputs. + +Other properties (maybe device placement information) can be added +to `Node` in the future if it's a +common requirement of many other `Pass`es. Otherwise, it should live +in a `Node` wrapper class that is private to some `Pass` or be +a local member of a `Pass`. + +#### Graph + +`Graph` contains a list of `Node`s, which are connected to +each other via inputs and outputs. + +TODO: Better definitions for the graph. + +`Graph` can also contain `Attribute`s. `Attribute`s +can be `any` thing. For example, it can be a list of "wraper" +nodes. The `wrapper` nodes compose `Node`s and provide +helper method for execution or transformation. `Attribute` +can also contain other things that describe some properties of +the `Graph` or `Graph` nodes. `Attribute` can be passed +across `Pass`. However, it should be used with care. + +```cpp +class Graph { + public: + explicit Graph(const ProgramDesc &program); + + bool Has(const std::string &attr_name) const; + + template + AttrType &Get(const std::string &attr_name) const; + + template + void Set(const std::string &attr_name, AttrType *attr); + const std::unordered_set &Nodes() const; + + // Create a normal variable with non-null VarDesc. + ir::Node *CreateVarNode(VarDesc *var_desc); + + // Create a normal runnable operator with OpDesc. + ir::Node *CreateOpNode(OpDesc *op_desc); + + // Create a control dependency var that connects 2 operations. The + // var doesn't hold any data. Other than that, it's no different from + // other var, considering dependency analysis. + ir::Node *CreateControlDepVar(); + + // A more free style way of creating a graph node. Mostly use for test + // or "copy" from another node. Avoid using it if possible. + ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type); + + // Clear all node information of the graph and return the ownership of the + // nodes. + std::vector> ReleaseNodes(); +}; +``` + +#### Pass + +`Pass` represents a transformation of `Graph`. Its input +is a `Graph` and its output is also a `Graph`. For example, +a `Pass` can simply print out the `Graph`. A `Pass` +can also fuse some `Graph`'s `Node`s. + +```cpp +class Pass { + public: + + std::unique_ptr Apply(std::unique_ptr graph) const { + // Some correctness check. + auto new_graph = ApplyImpl(std::move(graph)); + // Some correctness check. + return new_graph; + } + + // Get a reference to the attributed previously set. + template + AttrType &Get(const std::string &attr_name) const; + + // Set a pointer to the attribute. Pass takes ownership of the attribute. + template + void Set(const std::string &attr_name, AttrType *attr) ; + + // Set a pointer to the attribute. Pass doesn't take ownership. Caller + // should delete the attribute. + template + void SetNotOwned(const std::string &attr_name, AttrType *attr); + + protected: + virtual std::unique_ptr ApplyImpl(std::unique_ptr graph) const = 0; +}; + +// In my_pass.cc +class MyPass : public Pass { + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const override { + // do something. + return graph; + } +} +REGISTER_PASS(my_pass, MyPass) +.RequirePassAttr("places") +.RequireGraphAttr("dep_vars"); + + +// To use the pass. +auto my_pass = ir::PassRegistry::Instance().Get("my_pass"); +graph = my_pass->Apply(std::move(graph)); +// Note: to force link my_pass.cc, in the code: +USE_PASS(my_pass); +``` + +#### Optimize + +`Optimize` contains a series of `Pass` with defined order. +`Optimize` transforms a `Graph` that only contains raw +modeling logic to a `Graph` that can be run efficiently while +maintaining the original modeling logic. + + +### Optimize Process + +* Program is first converted to Graph. +* Graph goes through a series of Pass +* Graph is transformed from raw model logic to a +form that is efficient to execute. + +``` +// Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor +auto graph = Graph(program); +graph = PassRegistry::Instance().Get("op_fuse_pass").Apply(std::move(grah)); +// For more complex Pass, Optimize Process can provide Pass attributes. +auto mem_opt_pass = PassRegistry::Instance().Get("memory_optimization_pass"); +mem_opt_pass.SetNotOwned("optimize_level", 1); +mem_opt_pass->Apply(std::move(graph)); +graph = PassRegistry::Instance().Get("multi_devices_pass").Apply(std::move(grah)); +graph = PassRegistry::Instance().Get("multi_devices_check_pass").Apply(std::move(grah)); +Executor exe; +exe.Run(graph); + +``` diff --git a/doc/fluid/design/modules/python_api.md b/doc/fluid/design/modules/python_api.md index 265732a348ea77d21005e335390d99abcdfbd045..83af4e55485c079265d3f2b1e15070825b532c02 100644 --- a/doc/fluid/design/modules/python_api.md +++ b/doc/fluid/design/modules/python_api.md @@ -98,13 +98,13 @@ class Block(objects): def append_operator(self, ...): self.ops.append(Operator(self, ...)) - def prepend_operator(self, ...): # Parameter's ctor prepands initialize operators. + def _prepend_operator(self, ...): # Parameter's ctor prepands initialize operators. self.ops.prepend(Operator(self, ...)) ``` `create_parameter` is necessary because parameters are global variables, defined in the global block, but can be created in some sub-blocks. For example, an FC layer in the step block of an RNN operator. -`prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block. +`_prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block. ### Operator diff --git a/doc/fluid/design/multi_devices/kernel_selection.md b/doc/fluid/design/multi_devices/kernel_selection.md index 967317d5d2eeb818ab14faabca342cc8c4ed717e..4d2aab87b8cf30d03075e96cc4c67070efaf963a 100644 --- a/doc/fluid/design/multi_devices/kernel_selection.md +++ b/doc/fluid/design/multi_devices/kernel_selection.md @@ -74,10 +74,10 @@ void OperatorWithKernel::Run( auto kernel_type_for_var = this->GetKernelTypeForVar(...); if (kernel_type_for_var.place_ != expected_kernel_key.place_) { auto* trans_var = new_scope.Var(var_name); - auto* out = DataTransform(expected_kernel_key, + auto* out = TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in); - CopyVariableWithTensor(...); + SetTensorToVariable(...); } } diff --git a/doc/fluid/design/quantization/fixed_point_quantization.md b/doc/fluid/design/quantization/fixed_point_quantization.md new file mode 100644 index 0000000000000000000000000000000000000000..085352fc5614d693e63a2f7241e868a9649456af --- /dev/null +++ b/doc/fluid/design/quantization/fixed_point_quantization.md @@ -0,0 +1,110 @@ +Fixed-point quantization uses lower bits, for example, 2-bit, 3-bit or 8-bit fixed point to represent weights and activations, which usually are in singe-precision float-point with 32 bits. The fixed-point representation has advantages in reducing memory bandwidth, lowering power consumption and computational resources as well as the model storage requirements. It is especially important for the inference in embedded-device deployment. + +According to some experiments, the apporach to quantize the model trained in float point directly works effectively on the large models, like the VGG model having many parameters. But the accuracy drops a lot for the small model. In order to improve the tradeoff between accuracy and latency, many quantized training apporaches are proposed. + +This document is to design a quantized training framework on Fluid. The first part will introduce how to quantize, The second part will describe the quantized training framework. The last part will illustrate how to calculate the quantization scale. + + +### How to quantize + +There are many ways to quantize the float value to fixed-point value. For example: + +$$ r = min(max(x, a), b)$$ +$$ s = \frac{b - a}{n - 1} $$ +$$ q = \left \lfloor \frac{r - a}{s} \right \rceil $$ + +where, $x$ is the float value to be quantized, $[a, b]$ is the quantization range, $a$ is the minimum value and $b$ is the maximal value. $\left \lfloor \right \rceil$ denotes rounding to the nearest integer. If the quantization level is $k$, $n$ is $2^k$, for example, $k$ is 8 and $n$ is 256. $q$ is the quantized integer. + + +The quantization we applied is parameterized by the number of quantization levels and maximum absolute value: + +$$ M = max(abs(x)) $$ +$$ q = \left \lfloor \frac{x}{M} * (n - 1) \right \rceil $$ + +where, $x$ is the float value to be quantized, $M$ is maximum absolute value. $\left \lfloor \right \rceil$ denotes rounding to the nearest integer. For 8 bit quantization, $n=2^{8}=256$. $q$ is the quantized integer. + + +Wether the *min-max* quantization or *max-abs* quantization, they also can be represent: + +$q = scale * r + b$ + +We call *min-max*, *max-abs* as the quantization arguments, also call them quantization scale or quantization range. + + +How to calculate the quantization scale (or maximum absolute value) for inference will be described in the last part. + + +### Training Framework + +#### Forward pass + +The forward pass is simulated quantization, see Figure 1. + +The training framework is as following figure. + +

+
+Figure 1. Forward in training with simulated quantization. +

+ +- Firstly, both input and weight will be quantized to 8-bit integers. +- Second, do the multiplication (or convolution) operation with integers. +- Third, dequantize the multiplication (or convolution) results to 32-bit float point. +- Finally, do bias-addition in float type of 32 bit. Here, the bias is not quantized. + +For general matrix multiplication (GEMM), quantize for $X$ and $W$: + +$$ X_q = \left \lfloor \frac{X}{X_m} * (n - 1) \right \rceil $$ +$$ W_q = \left \lfloor \frac{W}{W_m} * (n - 1) \right \rceil $$ + +Do GEMM: + +$$ Y = X_q * W_q $$ + + +Dequantize $Y$: + +$$ +\begin{align} +Y_{dq} &=\frac{Y}{(n - 1) * (n - 1)} * X_m * W_m \\\ + &=\frac{X_q * W_q}{(n - 1) * (n - 1)} * X_m * W_m \\\ + &=(\frac{X_q}{n - 1} * X_m) * (\frac{W_q}{n - 1} * W_m) +\end{align} +$$ + +From these formulas, dequantization also can be moved before GEMM, do dequantization for $Xq$ and $Wq$ at first, then do GEMM. The forward workflow in training is equivalent to following framework. + +

+
+Figure 2. Equivalent forward in training with simulated quantization. +

+ +We use this equivalent workflow in the training. In our desigin, there is a quantization transpiler to insert the quantization operator and the de-quantization operator in the Fluid `ProgramDesc`. Since the outputs of quantization and de-quantization operator are still in floating point, they are called faked quantization and de-quantization operator. And the training framework is called simulated quantization. + +#### Backward pass + +See Figure 3. The gradients are calculated by dequantized weights and activations. All inputs and outputs are float point with 32-bit. And in the weight updating process, the gradients will be added to the original weight, not the quantized or dequantized weights. + +

+
+Figure 3. Backward and weight updating in training with simulated quantization. +

+ +So the quantization transipler will change some inputs of the corresponding backward operators. + +### How to calculate quantization scale + +There are two strategies to calculate quantization scale, we call them dynamic and static strategy. The dynamic strategy calculates the quantization scale value each iteration. The static strategy keeps the quantization scale for different inputs. + +For weights, we apply the dynamic strategy in the training, that is to say, the quantization scale will be recalculated during each iteration until the traning is finished. + +For activations, the quantization scales are estimated during training, then used in inference. There are several different ways to estimate them: + + +1. Calculate the mean of maximum absolute during a window. +2. Calculate the max of maximum absolute during a window. +3. Calculate the running mean of maximum absolute during a window, as follows: + + $$ Vt = (1 - k) * V + k * V_{t-1} $$ + + where, $V$ is the maximum absolute value of current batch, $Vt$ is the running mean value. $k$ is a factor, such as 0.9. diff --git a/doc/fluid/design/quantization/quantization_backward_and_optimization.png b/doc/fluid/design/quantization/quantization_backward_and_optimization.png new file mode 100644 index 0000000000000000000000000000000000000000..84f8235ab87cb631992b691f8e05b9c0b6c93da2 Binary files /dev/null and b/doc/fluid/design/quantization/quantization_backward_and_optimization.png differ diff --git a/doc/fluid/design/quantization/quantization_equivalent_forward.png b/doc/fluid/design/quantization/quantization_equivalent_forward.png new file mode 100644 index 0000000000000000000000000000000000000000..df49c864537c047c785da12d24893e54ce0a5341 Binary files /dev/null and b/doc/fluid/design/quantization/quantization_equivalent_forward.png differ diff --git a/doc/fluid/design/quantization/quantization_forward.png b/doc/fluid/design/quantization/quantization_forward.png new file mode 100644 index 0000000000000000000000000000000000000000..0913f61621bb6533bcb10bd1d18120ccaaa96cff Binary files /dev/null and b/doc/fluid/design/quantization/quantization_forward.png differ diff --git a/doc/fluid/dev/api_doc_std_cn.md b/doc/fluid/dev/api_doc_std_cn.md index b50f18f21df0787b9761bf0935ed7f4384ff0f98..7d39b8de1e6dc502ffea5f7882bd6a42b1ed6549 100644 --- a/doc/fluid/dev/api_doc_std_cn.md +++ b/doc/fluid/dev/api_doc_std_cn.md @@ -1,8 +1,9 @@ # API注释撰写标准 -- [API注释模块](#API注释模块) -- [格式及示例](#格式及示例) -- [完整示例](#完整示例) +- [API注释撰写标准](#api) + - [API注释模块](#api) + - [格式及示例](#) + - [完整示例](#) ## API注释模块 @@ -217,4 +218,4 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 ## 完整示例 -fc 的完整注释见[示例](src/fc.py)。 +fc 的完整注释见[示例](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。 diff --git a/doc/fluid/dev/api_doc_std_en.md b/doc/fluid/dev/api_doc_std_en.md index e57072d52fd162e92a3482aef33f99ab9394c532..f175b219750d1c765a6a111c2ec3aa732fa46175 100644 --- a/doc/fluid/dev/api_doc_std_en.md +++ b/doc/fluid/dev/api_doc_std_en.md @@ -1,8 +1,9 @@ # API Doc Standard -- [API Doc Structure](#API Doc Structure) -- [Format and Examples](#Format and Examples) -- [Complete Example](#Complete Example) +- [API Doc Standard](#api-doc-standard) + - [API Doc Structure](#api-doc-structure) + - [Format and Examples](#format-and-examples) + - [Complete Example](#complete-example) ## API Doc Structure @@ -223,4 +224,4 @@ Format and examples of each part of API documantation are as follows: (take fc f ## Complete Example -Complete Example of fc please see [here](src/fc.py)。 +Complete Example of fc please see [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。 diff --git a/doc/fluid/howto/cluster/fluid_cluster_train_cn.md b/doc/fluid/howto/cluster/fluid_cluster_train_cn.md index b99b90056b0a2e51f2668a6d27d94857bdc09c37..55326940ce7c7dbaa5bf19f1950f470527ddf4f0 100644 --- a/doc/fluid/howto/cluster/fluid_cluster_train_cn.md +++ b/doc/fluid/howto/cluster/fluid_cluster_train_cn.md @@ -168,13 +168,13 @@ cd /paddle/python/paddle/fluid/tests/book 第二步，启动Parameter Server： ```bash -PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.2 TRAINERS=2 POD_IP=192.168.1.2 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=PSERVER python test_fit_a_line.py +PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.2 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=192.168.1.2 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=PSERVER python test_fit_a_line.py ``` 执行命令后请等待出现提示： ```Server listening on 192.168.1.2:6174 ```, 表示Paramter Server已经正常启动。 第三步，启动Trainer： ```bash -PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.3 TRAINERS=2 POD_IP=192.168.1.3 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=TRAINER python test_fit_a_line.py +PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.3 PADDLE_TRAINERS=2 PADDLE_CURRENT_IPP=192.168.1.3 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=TRAINER python test_fit_a_line.py ``` 由于我们定义的Trainer的数量是2个，因此需要在另外一个计算节点上再启动一个Trainer。 diff --git a/doc/fluid/howto/cluster/fluid_recordio.md b/doc/fluid/howto/cluster/fluid_recordio.md new file mode 100644 index 0000000000000000000000000000000000000000..92859e8f622d0c155128821c54252113c5016989 --- /dev/null +++ b/doc/fluid/howto/cluster/fluid_recordio.md @@ -0,0 +1,127 @@ +# How to use RecordIO in Fluid + +If you want to use RecordIO as your training data format, you need to convert to your training data +to RecordIO files and reading them in the process of training, PaddlePaddle Fluid provides some +interface to deal with the RecordIO files. + +## Generate RecordIO File + +Before start training with RecordIO files, you need to convert your training data +to RecordIO format by `fluid.recordio_writer.convert_reader_to_recordio_file`, the sample codes +as follows: + +```python + reader = paddle.batch(mnist.train(), batch_size=1) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + fluid.recordio_writer.convert_reader_to_recordio_file('./mnist.recordio', reader, feeder) +``` + +The above code snippet would generate a RecordIO `./mnist.recordio` on your host. + +**NOTE**: we recommend users to set `batch_size=1` when generating the recordio files so that users can +adjust it flexibly while reading it. + +## Use the RecordIO file in a Local Training Job + +PaddlePaddle Fluid provides an interface `fluid.layers.io.open_recordio_file` to load your RecordIO file +and then you can use them as a Layer in your network configuration, the sample codes as follows: + +```python + data_file = fluid.layers.io.open_recordio_file( + filename="./mnist.recordio", + shapes=[(-1, 784),(-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int32"]) + data_file = fluid.layers.io.batch(data_file, batch_size=4) + + img, label = fluid.layers.io.read_file(data_file) + hidden = fluid.layers.fc(input=img, size=100, act='tanh') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + avg_loss = fluid.layers.mean(loss) + + fluid.optimizer.Adam(learning_rate=1e-3).minimize(avg_loss) + + place = fluid.CPUPlace() + + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + avg_loss_np = [] + + # train a pass + batch_id = 0 + while True: + tmp, = exe.run(fetch_list=[avg_loss]) + + avg_loss_np.append(tmp) + print(batch_id) + batch_id += 1 +``` + +## Use the RecordIO files in Distributed Training + +1. generate multiple RecordIO files + +For a distributed training job, you may have multiple trainer nodes, +and one or more RecordIO files for one trainer node, you can use the interface +`fluid.recordio_writer.convert_reader_to_recordio_files` to convert your training data +into multiple RecordIO files, the sample codes as follows: + +```python + reader = paddle.batch(mnist.train(), batch_size=1) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + fluid.recordio_writer.convert_reader_to_recordio_files( + filename_suffix='./mnist.recordio', batch_per_file=100, reader, feeder) +``` + +The above codes would generate multiple RecordIO files on your host like: + +```bash +. + \_mnist-00000.recordio + |-mnist-00001.recordio + |-mnist-00002.recordio + |-mnist-00003.recordio + |-mnist-00004.recordio +``` + +2. open multiple RecordIO files by `fluid.layers.io.open_files` + +For a distributed training job, the distributed operator system will schedule trainer process on multiple nodes, +each trainer process reads parts of the whole training data, we usually take the following approach to make the training +data allocated by each trainer process as uniform as possiable: + +```python +def gen_train_list(file_pattern, trainers, trainer_id): + file_list = glob.glob(file_pattern) + ret_list = [] + for idx, f in enumerate(file_list): + if (idx + trainers) % trainers == trainer_id: + ret_list.append(f) + return ret_list + +trainers = int(os.getenv("PADDLE_TRAINERS")) +trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) +data_file = fluid.layers.io.open_files( + filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0), + thread_num=1, + shapes=[(-1, 784),(-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int32"]) +img, label = fluid.layers.io.read_file(data_files) +... +``` diff --git a/doc/fluid/howto/index_cn.rst b/doc/fluid/howto/index_cn.rst index b7c620179724ebe97a0a47b75a57b376b21ccf90..b57af64f44da82926c4862578f3072960ca5aa92 100644 --- a/doc/fluid/howto/index_cn.rst +++ b/doc/fluid/howto/index_cn.rst @@ -4,5 +4,5 @@ .. toctree:: :maxdepth: 1 + inference/index_cn.rst optimization/index_cn.rst - inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/index_en.rst b/doc/fluid/howto/index_en.rst index f3ca41cdbf1d40ec8afaf045233a38755d8a777a..fd21e167ce3a46da167db1e9d7013804f730e047 100644 --- a/doc/fluid/howto/index_en.rst +++ b/doc/fluid/howto/index_en.rst @@ -5,4 +5,3 @@ HOW TO :maxdepth: 1 optimization/index_en.rst - inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/inference/build_and_install_lib_cn.rst b/doc/fluid/howto/inference/build_and_install_lib_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..91357dd8c8da19f2f33c6f285ed7eb234428b1ab --- /dev/null +++ b/doc/fluid/howto/inference/build_and_install_lib_cn.rst @@ -0,0 +1,97 @@ +安装与编译C++预测库 +=========================== + +直接下载安装 +------------- + +====================== ======================================== +版本说明 C++预测库 +====================== ======================================== +cpu_avx_mkl `fluid.tgz `_ +cpu_avx_openblas `fluid.tgz `_ +cpu_noavx_openblas `fluid.tgz `_ +cuda7.5_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn7_avx_mkl `fluid.tgz `_ +cuda9.0_cudnn7_avx_mkl `fluid.tgz `_ +====================== ======================================== + +从源码编译 +---------- +用户也可以从 PaddlePaddle 核心代码编译C++预测库，只需在编译时配制下面这些编译选项： + +================= ========= +选项 值 +================= ========= +CMAKE_BUILD_TYPE Release +FLUID_INSTALL_DIR 安装路径 +WITH_FLUID_ONLY ON（推荐） +WITH_SWIG_PY OFF（推荐 +WITH_PYTHON OFF（推荐） +WITH_GPU ON/OFF +WITH_MKL ON/OFF +================= ========= + +建议按照推荐值设置，以避免链接不必要的库。其它可选编译选项按需进行设定。 + +下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）： + + .. code-block:: bash + + pip install paddlepaddle-gpu + PADDLE_ROOT=/path/of/capi + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + mkdir build + cd build + cmake -DFLUID_INSTALL_DIR=$PADDLE_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_FLUID_ONLY=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_PYTHON=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=OFF \ + .. + make + make inference_lib_dist + +成功编译后，使用C++预测库所需的依赖（包括：（1）编译出的PaddlePaddle预测库和头文件；（2）第三方链接库和头文件；（3）版本信息与编译选项信息） +均会存放于PADDLE_ROOT目录中。目录结构如下： + + .. code-block:: text + + PaddleRoot/ + ├── CMakeCache.txt + ├── paddle + │   └── fluid + │   ├── framework + │   ├── inference + │   ├── memory + │   ├── platform + │   ├── pybind + │   └── string + ├── third_party + │   ├── boost + │   │   └── boost + │   ├── eigen3 + │   │   ├── Eigen + │   │   └── unsupported + │   └── install + │   ├── gflags + │   ├── glog + │   ├── mklml + │   ├── protobuf + │   ├── snappy + │   ├── snappystream + │   └── zlib + └── version.txt + +version.txt 中记录了该预测库的版本信息，包括Git Commit ID、使用OpenBlas或MKL数学库、CUDA/CUDNN版本号，如： + + .. code-block:: text + + GIT COMMIT ID: c95cd4742f02bb009e651a00b07b21c979637dc8 + WITH_MKL: ON + WITH_GPU: ON + CUDA version: 8.0 + CUDNN version: v5 diff --git a/doc/fluid/howto/inference/index_cn.rst b/doc/fluid/howto/inference/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..a903423548decd0992bf19772fb2cb143f6a12b5 --- /dev/null +++ b/doc/fluid/howto/inference/index_cn.rst @@ -0,0 +1,8 @@ +预测库 +------------ + +.. toctree:: + :maxdepth: 1 + + build_and_install_lib_cn.rst + inference_support_in_fluid_cn.md diff --git a/doc/fluid/howto/inference/inference_support_in_fluid.md b/doc/fluid/howto/inference/inference_support_in_fluid.md deleted file mode 100644 index d272cd3e3bdac49b9ed1a21531de1b0be03d881e..0000000000000000000000000000000000000000 --- a/doc/fluid/howto/inference/inference_support_in_fluid.md +++ /dev/null @@ -1,361 +0,0 @@ -# Fluid Inference使用指南 - -## 目录： - -- Python Inference API -- 编译Fluid Inference库 -- Inference C++ API -- Inference实例 -- Inference计算优化 - -## Python Inference API **[改进中]** -- 保存Inference模型 ([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/io.py#L295)) - - ```python - def save_inference_model(dirname, - feeded_var_names, - target_vars, - executor, - main_program=None, - model_filename=None, - params_filename=None): - ``` - Inference模型和参数将会保存到`dirname`目录下： - - 序列化的模型 - - `model_filename`为`None`，保存到`dirname/__model__` - - `model_filename`非`None`，保存到`dirname/model_filename` - - 参数 - - `params_filename`为`None`，单独保存到各个独立的文件，各文件以参数变量的名字命名 - - `params_filename`非`None`，保存到`dirname/params_filename` - -- 两种存储格式 - - 参数保存到各个独立的文件 - - 如，设置`model_filename`为`None`、`params_filename`为`None` - - ```bash - $ cd recognize_digits_conv.inference.model - $ ls - $ __model__ batch_norm_1.w_0 batch_norm_1.w_2 conv2d_2.w_0 conv2d_3.w_0 fc_1.w_0 batch_norm_1.b_0 batch_norm_1.w_1 conv2d_2.b_0 conv2d_3.b_0 fc_1.b_0 - ``` - - 参数保存到同一个文件 - - 如，设置`model_filename`为`None`、`params_filename`为`__params__` - - ```bash - $ cd recognize_digits_conv.inference.model - $ ls - $ __model__ __params__ - ``` -- 加载Inference模型([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/io.py#L380)) - ```python - def load_inference_model(dirname, - executor, - model_filename=None, - params_filename=None): - ... - return [program, feed_target_names, fetch_targets] - ``` - - -## 编译Fluid Inference库 - - - **不需要额外的CMake选项** - - 1、 配置CMake命令，更多配置请参考[源码编译PaddlePaddle](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/build_from_source_cn.html) - ```bash - $ git clone https://github.com/PaddlePaddle/Paddle.git - $ cd Paddle - $ mkdir build - $ cd build - $ cmake -DCMAKE_INSTALL_PREFIX=your/path/to/paddle_inference_lib \ - -DCMAKE_BUILD_TYPE=Release \ - -DWITH_PYTHON=ON \ - -DWITH_MKL=OFF \ - -DWITH_GPU=OFF \ - .. - ``` - - - 2、 编译PaddlePaddle - ```bash - $ make - ``` - - - 3、 部署。执行如下命令将PaddlePaddle Fluid Inference库部署到`your/path/to/paddle_inference_lib`目录。 - ```bash - $ make inference_lib_dist - ``` - -- 目录结构 - - ```bash - $ cd your/path/to/paddle_inference_lib - $ tree - . - |-- paddle - | `-- fluid - | |-- framework - | |-- inference - | | |-- io.h - | | `-- libpaddle_fluid.so - | |-- memory - | |-- platform - | `-- string - |-- third_party - | |-- eigen3 - | `-- install - | |-- gflags - | |-- glog - | `-- protobuf - `-- ... - ``` - - 假设`PADDLE_ROOT=your/path/to/paddle_inference_lib`。 - - - -## 链接Fluid Inference库 -- 示例项目([链接](https://github.com/luotao1/fluid_inference_example.git)) - - - GCC配置 - ```bash - $ g++ -o a.out -std=c++11 main.cc \ - -I${PADDLE_ROOT}/ \ - -I${PADDLE_ROOT}/third_party/install/gflags/include \ - -I${PADDLE_ROOT}/third_party/install/glog/include \ - -I${PADDLE_ROOT}/third_party/install/protobuf/include \ - -I${PADDLE_ROOT}/third_party/eigen3 \ - -L${PADDLE_ROOT}/paddle/fluid/inference -lpaddle_fluid \ - -lrt -ldl -lpthread - ``` - - - CMake配置 - ```cmake - include_directories(${PADDLE_ROOT}/) - include_directories(${PADDLE_ROOT}/third_party/install/gflags/include) - include_directories(${PADDLE_ROOT}/third_party/install/glog/include) - include_directories(${PADDLE_ROOT}/third_party/install/protobuf/include) - include_directories(${PADDLE_ROOT}/third_party/eigen3) - target_link_libraries(${TARGET_NAME} - ${PADDLE_ROOT}/paddle/fluid/inference/libpaddle_fluid.so - -lrt -ldl -lpthread) - ``` - - - 设置环境变量： - `export LD_LIBRARY_PATH=${PADDLE_ROOT}/paddle/fluid/inference:$LD_LIBRARY_PATH` - - - -## C++ Inference API - -- 推断流程([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/test_helper.h#L91)) - - - 1、 初始化设备 - ```cpp - #include "paddle/fluid/framework/init.h" - paddle::framework::InitDevices(false); - ``` - - - 2、 定义place，executor，scope - ```cpp - auto place = paddle::platform::CPUPlace(); - auto executor = paddle::framework::Executor(place); - auto* scope = new paddle::framework::Scope(); - ``` - - - 3、 加载模型 - ```cpp - #include "paddle/fluid/inference/io.h" - auto inference_program = paddle::inference::Load(executor, *scope, dirname); - // or - auto inference_program = paddle::inference::Load(executor, - *scope, - dirname + "/" + model_filename, - dirname + "/" + params_filename); - ``` - - - 4、 获取`feed_target_names`和`fetch_target_names` - ```cpp - const std::vector& feed_target_names = inference_program->GetFeedTargetNames(); - const std::vector& fetch_target_names = inference_program->GetFetchTargetNames(); - ``` - - - 5、 准备`feed`数据 - ```cpp - #include "paddle/fluid/framework/lod_tensor.h" - std::vector cpu_feeds; - ... - std::map feed_targets; - for (size_t i = 0; i < feed_target_names.size(); ++i) { - // Please make sure that cpu_feeds[i] is right for feed_target_names[i] - feed_targets[feed_target_names[i]] = cpu_feeds[i]; - } - ``` - - - 6、 定义`Tensor`来`fetch`结果 - ```cpp - std::vector cpu_fetchs; - std::map fetch_targets; - for (size_t i = 0; i < fetch_target_names.size(); ++i) { - fetch_targets[fetch_target_names[i]] = cpu_fetchs[i]; - } - ``` - - - 7、 执行`inference_program` - ```cpp - executor.Run(*inference_program, scope, feed_targets, fetch_targets); - ``` - - - 8、 使用`fetch`数据 - ```cpp - for (size_t i = 0; i < cpu_fetchs.size(); ++i) { - std::cout << "lod_i: " << cpu_fetchs[i]->lod(); - std::cout << "dims_i: " << cpu_fetchs[i]->dims(); - std::cout << "result:"; - float* output_ptr = cpu_fetchs[i]->data(); - for (int j = 0; j < cpu_fetchs[i]->numel(); ++j) { - std::cout << " " << output_ptr[j]; - } - std::cout << std::endl; - } - ``` - 针对不同的数据，4. - 8.可执行多次。 - - - 9、 释放内存 - ```cpp - delete scope; - ``` - - -- 接口说明 - - ```cpp - void Run(const ProgramDesc& program, Scope* scope, - std::map& feed_targets, - std::map& fetch_targets, - bool create_vars = true, - const std::string& feed_holder_name = "feed", - const std::string& fetch_holder_name = "fetch"); - ``` - - 使用Python API `save_inference_model`保存的`program`里面包含了`feed_op`和`fetch_op`，用户提供的`feed_targets`、`fetch_targets`必须和`inference_program`中的`feed_op`、`fetch_op`保持一致。 - - 用户提供的`feed_holder_name`和`fetch_holder_name`也必须和`inference_program`中`feed_op`、`fetch_op`保持一致，可使用`SetFeedHolderName`和`SetFetchHolderName`接口重新设置`inferece_program` - - 默认情况下，除了`persistable`属性设置为`True`的`Variable`之外，每次执行`executor.Run`会创建一个局部`Scope`，并且在这个局部`Scope`中创建和销毁所有的`Variable`，以最小化空闲时的内存占用。 - - `persistable`属性为`True`的`Variable`有： - - Operators的参数`w`、`b`等 - - `feed_op`的输入变量 - - `fetch_op`的输出变量 - - -- **不在每次执行时创建和销毁变量 - ([PR](https://github.com/PaddlePaddle/Paddle/pull/9301))** - - 执行`inference_program` - ```cpp - // Call once - executor.CreateVariables(*inference_program, scope, 0); - // Call as many times as you like - executor.Run( - *inference_program, scope, feed_targets, fetch_targets, false); - ``` - - **优点** - - 节省了频繁创建、销毁变量的时间（约占每次`Run`总时间的1% ~ 12%） - - 执行结束后可获取所有Operators的计算结果 - - **缺点** - - 空闲时也会占用大量的内存 - - 在同一个`Scope`中，相同的变量名是公用同一块内存的，容易引起意想不到的错误 - - -- **不在每次执行时创建Op([PR](https://github.com/PaddlePaddle/Paddle/pull/9630))** - - 执行`inference_program` - ```cpp - // Call once - auto ctx = executor.Prepare(*inference_program, 0); - // Call as many times as you like if you have no need to change the inference_program - executor.RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets); - ``` - - **优点** - - 节省了频繁创建、销毁Op的时间 - - **缺点** - - 一旦修改了`inference_program`，则需要重新创建`ctx` - - -- **多线程共享Parameters([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/test_multi_thread_helper.h))** - - 主线程 - - 1、 初始化设备 - - 2、 定义`place`，`executor`，`scope` - - 3、 加载模型，得到`inference_program` - - 从线程 - - **复制`inference_program`得到`copy_program`，修改`copy_program`的`feed_holder_name`和`fetch_holder_name`** - ```cpp - auto copy_program = std::unique_ptr( - new paddle::framework::ProgramDesc(*inference_program)); - std::string feed_holder_name = "feed_" + paddle::string::to_string(thread_id); - std::string fetch_holder_name = "fetch_" + paddle::string::to_string(thread_id); - copy_program->SetFeedHolderName(feed_holder_name); - copy_program->SetFetchHolderName(fetch_holder_name); - ``` - - 4、 获取`copy_program`的`feed_target_names`和`fetch_target_names` - - 5、 准备feed数据，定义Tensor来fetch结果 - - 6、 执行`copy_program` - ```cpp - executor->Run(*copy_program, scope, feed_targets, fetch_targets, true, feed_holder_name, fetch_holder_name); - ``` - - 7、 使用fetch数据 - - 主线程 - - 8、 释放资源 - - -- 基本概念 - - 数据相关： - - [Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/tensor.md)，一个N维数组，数据可以是任意类型（int，float，double等） - - [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md)，带LoD(Level-of-Detail)即序列信息的Tensor - - [Scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md)，记录了变量Variable - - 执行相关： - - [Executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md)，无状态执行器，只跟设备相关 - - Place - - CPUPlace，CPU设备 - - CUDAPlace，CUDA GPU设备 - - 神经网络表示： - - [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). - - 详细介绍请参考[**Paddle Fluid开发者指南**](https://github.com/lcy-seso/learning_notes/blob/master/Fluid/developer's_guid_for_Fluid/Developer's_Guide_to_Paddle_Fluid.md) - - - -## Inference实例 - - 1. fit a line: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_fit_a_line.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_fit_a_line.cc) - 1. image classification: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_image_classification.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_image_classification.cc) - 1. label semantic roles: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_label_semantic_roles.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_label_semantic_roles.cc) - 1. recognize digits: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recognize_digits.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc) - 1. recommender system: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recommender_system.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_recommender_system.cc) - 1. understand sentiment: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_understand_sentiment.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_understand_sentiment.cc) - 1. word2vec: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_word2vec.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_word2vec.cc) - - -## Inference计算优化 -- 使用Python推理优化工具([inference_transpiler](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/inference_transpiler.py)) - ```python - class InferenceTranspiler: - def transpile(self, program, place, scope=None): - ... - if scope is None: - scope = global_scope() - ... - ``` - - 使用`InferenceTranspiler`将会直接修改`program`。 - - 使用`InferenceTranspiler`会修改参数的值，请确保`program`的参数在`scope`内。 -- 支持的优化 - - 融合batch_norm op的计算 -- 使用示例([链接](https://github.com/Xreki/Xreki.github.io/blob/master/fluid/inference/inference_transpiler.py)) - ```python - import paddle.fluid as fluid - # NOTE: Applying the inference transpiler will change the inference_program. - t = fluid.InferenceTranspiler() - t.transpile(inference_program, place, inference_scope) - ``` - - - - -## 内存使用优化 -- 使用Python内存优化工具([memory_optimization_transipiler](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/memory_optimization_transpiler.py)) - ```python - fluid.memory_optimize(inference_program) - ``` diff --git a/doc/fluid/howto/inference/inference_support_in_fluid_cn.md b/doc/fluid/howto/inference/inference_support_in_fluid_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..309b17fccd5c461c9c22beb64eb4c6792b7e4a7a --- /dev/null +++ b/doc/fluid/howto/inference/inference_support_in_fluid_cn.md @@ -0,0 +1,304 @@ +# 使用指南 + +## 目录： + +- Python Inference API +- Inference C++ API +- Inference实例 +- Inference计算优化 + +## Python Inference API **[改进中]** +- 保存Inference模型 ([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/io.py#L295)) + + ```python + def save_inference_model(dirname, + feeded_var_names, + target_vars, + executor, + main_program=None, + model_filename=None, + params_filename=None): + ``` + Inference模型和参数将会保存到`dirname`目录下： + - 序列化的模型 + - `model_filename`为`None`，保存到`dirname/__model__` + - `model_filename`非`None`，保存到`dirname/model_filename` + - 参数 + - `params_filename`为`None`，单独保存到各个独立的文件，各文件以参数变量的名字命名 + - `params_filename`非`None`，保存到`dirname/params_filename` + +- 两种存储格式 + - 参数保存到各个独立的文件 + - 如，设置`model_filename`为`None`、`params_filename`为`None` + + ```bash + $ cd recognize_digits_conv.inference.model + $ ls + $ __model__ batch_norm_1.w_0 batch_norm_1.w_2 conv2d_2.w_0 conv2d_3.w_0 fc_1.w_0 batch_norm_1.b_0 batch_norm_1.w_1 conv2d_2.b_0 conv2d_3.b_0 fc_1.b_0 + ``` + - 参数保存到同一个文件 + - 如，设置`model_filename`为`None`、`params_filename`为`__params__` + + ```bash + $ cd recognize_digits_conv.inference.model + $ ls + $ __model__ __params__ + ``` +- 加载Inference模型([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/io.py#L380)) + ```python + def load_inference_model(dirname, + executor, + model_filename=None, + params_filename=None): + ... + return [program, feed_target_names, fetch_targets] + ``` + +## 链接Fluid Inference库 +- 示例项目([链接](https://github.com/luotao1/fluid_inference_example.git)) + + - GCC配置 + ```bash + $ g++ -o a.out -std=c++11 main.cc \ + -I${PADDLE_ROOT}/ \ + -I${PADDLE_ROOT}/third_party/install/gflags/include \ + -I${PADDLE_ROOT}/third_party/install/glog/include \ + -I${PADDLE_ROOT}/third_party/install/protobuf/include \ + -I${PADDLE_ROOT}/third_party/eigen3 \ + -L${PADDLE_ROOT}/paddle/fluid/inference -lpaddle_fluid \ + -lrt -ldl -lpthread + ``` + + - CMake配置 + ```cmake + include_directories(${PADDLE_ROOT}/) + include_directories(${PADDLE_ROOT}/third_party/install/gflags/include) + include_directories(${PADDLE_ROOT}/third_party/install/glog/include) + include_directories(${PADDLE_ROOT}/third_party/install/protobuf/include) + include_directories(${PADDLE_ROOT}/third_party/eigen3) + target_link_libraries(${TARGET_NAME} + ${PADDLE_ROOT}/paddle/fluid/inference/libpaddle_fluid.so + -lrt -ldl -lpthread) + ``` + + - 设置环境变量： + `export LD_LIBRARY_PATH=${PADDLE_ROOT}/paddle/fluid/inference:$LD_LIBRARY_PATH` + + + +## C++ Inference API + +- 推断流程([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/test_helper.h#L91)) + + - 1、 初始化设备 + ```cpp + #include "paddle/fluid/framework/init.h" + paddle::framework::InitDevices(false); + ``` + + - 2、 定义place，executor，scope + ```cpp + auto place = paddle::platform::CPUPlace(); + auto executor = paddle::framework::Executor(place); + auto* scope = new paddle::framework::Scope(); + ``` + + - 3、 加载模型 + ```cpp + #include "paddle/fluid/inference/io.h" + auto inference_program = paddle::inference::Load(executor, *scope, dirname); + // or + auto inference_program = paddle::inference::Load(executor, + *scope, + dirname + "/" + model_filename, + dirname + "/" + params_filename); + ``` + + - 4、 获取`feed_target_names`和`fetch_target_names` + ```cpp + const std::vector& feed_target_names = inference_program->GetFeedTargetNames(); + const std::vector& fetch_target_names = inference_program->GetFetchTargetNames(); + ``` + + - 5、 准备`feed`数据 + ```cpp + #include "paddle/fluid/framework/lod_tensor.h" + std::vector cpu_feeds; + ... + std::map feed_targets; + for (size_t i = 0; i < feed_target_names.size(); ++i) { + // Please make sure that cpu_feeds[i] is right for feed_target_names[i] + feed_targets[feed_target_names[i]] = cpu_feeds[i]; + } + ``` + + - 6、 定义`Tensor`来`fetch`结果 + ```cpp + std::vector cpu_fetchs; + std::map fetch_targets; + for (size_t i = 0; i < fetch_target_names.size(); ++i) { + fetch_targets[fetch_target_names[i]] = cpu_fetchs[i]; + } + ``` + + - 7、 执行`inference_program` + ```cpp + executor.Run(*inference_program, scope, feed_targets, fetch_targets); + ``` + + - 8、 使用`fetch`数据 + ```cpp + for (size_t i = 0; i < cpu_fetchs.size(); ++i) { + std::cout << "lod_i: " << cpu_fetchs[i]->lod(); + std::cout << "dims_i: " << cpu_fetchs[i]->dims(); + std::cout << "result:"; + float* output_ptr = cpu_fetchs[i]->data(); + for (int j = 0; j < cpu_fetchs[i]->numel(); ++j) { + std::cout << " " << output_ptr[j]; + } + std::cout << std::endl; + } + ``` + 针对不同的数据，4. - 8.可执行多次。 + + - 9、 释放内存 + ```cpp + delete scope; + ``` + + +- 接口说明 + + ```cpp + void Run(const ProgramDesc& program, Scope* scope, + std::map& feed_targets, + std::map& fetch_targets, + bool create_vars = true, + const std::string& feed_holder_name = "feed", + const std::string& fetch_holder_name = "fetch"); + ``` + - 使用Python API `save_inference_model`保存的`program`里面包含了`feed_op`和`fetch_op`，用户提供的`feed_targets`、`fetch_targets`必须和`inference_program`中的`feed_op`、`fetch_op`保持一致。 + - 用户提供的`feed_holder_name`和`fetch_holder_name`也必须和`inference_program`中`feed_op`、`fetch_op`保持一致，可使用`SetFeedHolderName`和`SetFetchHolderName`接口重新设置`inferece_program` + - 默认情况下，除了`persistable`属性设置为`True`的`Variable`之外，每次执行`executor.Run`会创建一个局部`Scope`，并且在这个局部`Scope`中创建和销毁所有的`Variable`，以最小化空闲时的内存占用。 + - `persistable`属性为`True`的`Variable`有： + - Operators的参数`w`、`b`等 + - `feed_op`的输入变量 + - `fetch_op`的输出变量 + + +- **不在每次执行时创建和销毁变量 + ([PR](https://github.com/PaddlePaddle/Paddle/pull/9301))** + - 执行`inference_program` + ```cpp + // Call once + executor.CreateVariables(*inference_program, scope, 0); + // Call as many times as you like + executor.Run( + *inference_program, scope, feed_targets, fetch_targets, false); + ``` + - **优点** + - 节省了频繁创建、销毁变量的时间（约占每次`Run`总时间的1% ~ 12%） + - 执行结束后可获取所有Operators的计算结果 + - **缺点** + - 空闲时也会占用大量的内存 + - 在同一个`Scope`中，相同的变量名是公用同一块内存的，容易引起意想不到的错误 + + +- **不在每次执行时创建Op([PR](https://github.com/PaddlePaddle/Paddle/pull/9630))** + - 执行`inference_program` + ```cpp + // Call once + auto ctx = executor.Prepare(*inference_program, 0); + // Call as many times as you like if you have no need to change the inference_program + executor.RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets); + ``` + - **优点** + - 节省了频繁创建、销毁Op的时间 + - **缺点** + - 一旦修改了`inference_program`，则需要重新创建`ctx` + + +- **多线程共享Parameters([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/test_multi_thread_helper.h))** + - 主线程 + - 1、 初始化设备 + - 2、 定义`place`，`executor`，`scope` + - 3、 加载模型，得到`inference_program` + - 从线程 + - **复制`inference_program`得到`copy_program`，修改`copy_program`的`feed_holder_name`和`fetch_holder_name`** + ```cpp + auto copy_program = std::unique_ptr( + new paddle::framework::ProgramDesc(*inference_program)); + std::string feed_holder_name = "feed_" + paddle::string::to_string(thread_id); + std::string fetch_holder_name = "fetch_" + paddle::string::to_string(thread_id); + copy_program->SetFeedHolderName(feed_holder_name); + copy_program->SetFetchHolderName(fetch_holder_name); + ``` + - 4、 获取`copy_program`的`feed_target_names`和`fetch_target_names` + - 5、 准备feed数据，定义Tensor来fetch结果 + - 6、 执行`copy_program` + ```cpp + executor->Run(*copy_program, scope, feed_targets, fetch_targets, true, feed_holder_name, fetch_holder_name); + ``` + - 7、 使用fetch数据 + - 主线程 + - 8、 释放资源 + + +- 基本概念 + - 数据相关： + - [Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/tensor.md)，一个N维数组，数据可以是任意类型（int，float，double等） + - [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md)，带LoD(Level-of-Detail)即序列信息的Tensor + - [Scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md)，记录了变量Variable + - 执行相关： + - [Executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md)，无状态执行器，只跟设备相关 + - Place + - CPUPlace，CPU设备 + - CUDAPlace，CUDA GPU设备 + - 神经网络表示： + - [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). + + 详细介绍请参考[**Paddle Fluid开发者指南**](https://github.com/lcy-seso/learning_notes/blob/master/Fluid/developer's_guid_for_Fluid/Developer's_Guide_to_Paddle_Fluid.md) + + + +## Inference实例 + + 1. fit a line: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_fit_a_line.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_fit_a_line.cc) + 1. image classification: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_image_classification.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_image_classification.cc) + 1. label semantic roles: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_label_semantic_roles.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_label_semantic_roles.cc) + 1. recognize digits: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recognize_digits.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc) + 1. recommender system: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recommender_system.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_recommender_system.cc) + 1. understand sentiment: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_understand_sentiment.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_understand_sentiment.cc) + 1. word2vec: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_word2vec.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_word2vec.cc) + + +## Inference计算优化 +- 使用Python推理优化工具([inference_transpiler](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/inference_transpiler.py)) + ```python + class InferenceTranspiler: + def transpile(self, program, place, scope=None): + ... + if scope is None: + scope = global_scope() + ... + ``` + - 使用`InferenceTranspiler`将会直接修改`program`。 + - 使用`InferenceTranspiler`会修改参数的值，请确保`program`的参数在`scope`内。 +- 支持的优化 + - 融合batch_norm op的计算 +- 使用示例([链接](https://github.com/Xreki/Xreki.github.io/blob/master/fluid/inference/inference_transpiler.py)) + ```python + import paddle.fluid as fluid + # NOTE: Applying the inference transpiler will change the inference_program. + t = fluid.InferenceTranspiler() + t.transpile(inference_program, place, inference_scope) + ``` + + + + +## 内存使用优化 +- 使用Python内存优化工具([memory_optimization_transipiler](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/memory_optimization_transpiler.py)) + ```python + fluid.memory_optimize(inference_program) + ``` diff --git a/doc/fluid/howto/optimization/benchmark/README.md b/doc/fluid/howto/optimization/benchmark/README.md deleted file mode 120000 index db30af7f53231c687f9ad61ad961a685733cbad0..0000000000000000000000000000000000000000 --- a/doc/fluid/howto/optimization/benchmark/README.md +++ /dev/null @@ -1 +0,0 @@ -../../../../../benchmark/cluster/README.md \ No newline at end of file diff --git a/doc/fluid/howto/optimization/benchmark/vgg16/README.md b/doc/fluid/howto/optimization/benchmark/vgg16/README.md deleted file mode 120000 index ca963ef5f06aa0c2fe507ba7548dca8017358120..0000000000000000000000000000000000000000 --- a/doc/fluid/howto/optimization/benchmark/vgg16/README.md +++ /dev/null @@ -1 +0,0 @@ -../../../../../../benchmark/cluster/vgg16/README.md \ No newline at end of file diff --git a/doc/fluid/howto/optimization/host_memory_profiling_cn.md b/doc/fluid/howto/optimization/host_memory_profiling_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..7fb0883dd937465d15479b29df95078edb50e069 --- /dev/null +++ b/doc/fluid/howto/optimization/host_memory_profiling_cn.md @@ -0,0 +1,89 @@ +# 堆内存分析和优化 + +计算机程序都可能有内存泄漏的风险。**内存泄漏**一般是由于程序在堆(heap)上分配了内存而没有释放，随着程序的运行占用的内存越来越大，一方面会影响程序的稳定性，可能让运行速度越来越慢，或者造成oom，甚至会影响运行程序的机器的稳定性，造成宕机。 + + +目前有很多内存泄漏分析工具，比较经典的有[valgrind](http://valgrind.org/docs/manual/quick-start.html#quick-start.intro), [gperftools](https://gperftools.github.io/gperftools/)。 + +因为Fluid是用Python驱动C++ core来运行，valgrind直接分析非常困难，需要自己编译debug版本的、带valgrind支持的专用Python版本，而且输出的信息中大部分是Python自己的符号和调用信息，分析起来很困难，另外使用valgrind会让程序运行速度变得非常慢，所以不建议使用。 + +本教程主要介绍[gperftools](https://gperftools.github.io/gperftools/)的使用。 + +gperftool主要支持以下四个功能： + +- thread-caching malloc +- heap-checking using tcmalloc +- heap-profiling using tcmalloc +- CPU profiler + +Paddle也提供了基于gperftool的[CPU性能分析教程](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/cpu_profiling_cn.md)。 + +对于堆内存的分析，主要用到thread-caching malloc和heap-profiling using tcmalloc。 + +## 环境 + +本教程基于paddle提供的Docker开发环境paddlepaddle/paddle:latest-dev，基于Ubuntu 16.04.4 LTS环境。 + +## 使用流程 + +- 安装google-perftools + +``` +apt-get install libunwind-dev +apt-get install google-perftools +``` + +- 安装pprof + +``` +go get -u github.com/google/pprof +``` + +- 设置运行环境 + +``` +export PPROF_PATH=/root/gopath/bin/pprof +export PPROF_BINARY_PATH=/root/gopath/bin/pprof +export LD_PRELOAD=/usr/lib/libtcmalloc.so.4 +``` + +- 使用heap profile来运行python程序。本质上是周期性的对堆的分配情况做一次快照。 + +``` +# HEAPPROFILE 设置生成的堆分析文件的目录和文件前缀 +# HEAP_PROFILE_ALLOCATION_INTERVAL 设置每分配多少存储dump一次dump，默认1GB +env HEAPPROFILE="./perf_log/test.log" HEAP_PROFILE_ALLOCATION_INTERVAL=209715200 python trainer.py +``` + +随着程序的运行，会在perf_log这个文件夹下生成很多文件，如下： + +``` +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0001.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0002.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0003.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0004.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0005.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0006.heap +``` + +- 使用pprof对heap文件进行分析。分析有两种模式： + - 完整模式。会对当前heap做一个分析，显示目前分配内存一些调用路径。 + + ``` + pprof --pdf python test.log.0012.heap + ``` + 上述命令会生成一个profile00x.pdf的文件，可以直接打开，例如：[memory_cpu_allocator](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_cpu_allocator.pdf)。从下图可以看出，在CPU版本fluid的运行过程中，分配存储最多的模块式CPUAllocator. 而别的模块相对而言分配内存较少，所以被忽略了，这对于分配内存泄漏是很不方便的，因为泄漏是一个缓慢的过程，在这种图中是无法看到的。 + + ![result](https://user-images.githubusercontent.com/3048612/40964027-a54033e4-68dc-11e8-836a-144910c4bb8c.png) + + - Diff模式。可以对两个时刻的heap做diff，把一些内存分配没有发生变化的模块去掉，而把增量部分显示出来。 + ``` + pprof --pdf --base test.log.0010.heap python test.log.1045.heap + ``` + 生成的结果为：[`memory_leak_protobuf`](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_leak_protobuf.pdf) + + 从图中可以看出：ProgramDesc这个结构，在两个版本之间增长了200MB+，所以这里有很大的内存泄漏的可能性，最终结果也确实证明是这里造成了泄漏。 + + ![result](https://user-images.githubusercontent.com/3048612/40964057-b434d5e4-68dc-11e8-894b-8ab62bcf26c2.png) + ![result](https://user-images.githubusercontent.com/3048612/40964063-b7dbee44-68dc-11e8-9719-da279f86477f.png) + diff --git a/doc/fluid/howto/optimization/timeline.md b/doc/fluid/howto/optimization/timeline.md deleted file mode 100644 index 96481ae2a6e4442d40803f8d5361e5f942502df3..0000000000000000000000000000000000000000 --- a/doc/fluid/howto/optimization/timeline.md +++ /dev/null @@ -1,27 +0,0 @@ -# how to use timeline tool to do profile - -1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number. - - ```python - with profiler.profiler('All', 'total', '/tmp/profile') as prof: - for pass_id in range(pass_num): - for batch_id, data in enumerate(train_reader()): - exe.run(fluid.default_main_program(), - feed=feeder.feed(data), - fetch_list=[], - use_program_cache=True) - ... - ``` - -1. Run `python paddle/tools/timeline.py` to process `/tmp/profile`, it will generate another -file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at -[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details. - -1. Open chrome and visit , use `load` button to load the generated `timeline` file. - - ![chrome tracing](./tracing.jpeg) - -1. The resulting timeline should be like: - - - ![chrome timeline](./timeline.jpeg) diff --git a/doc/fluid/howto/optimization/timeline_cn.md b/doc/fluid/howto/optimization/timeline_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..faf39f276dbddcd4961407ba2d082c9826051cbe --- /dev/null +++ b/doc/fluid/howto/optimization/timeline_cn.md @@ -0,0 +1,32 @@ +# 如何使用timeline工具做性能分析 + +1. 在训练的主循环外加上`profiler.start_profiler(...)`和`profiler.stop_profiler(...)`。运行之后，代码会在`/tmp/profile`目录下生成一个profile的记录文件。 + + **提示：** + 请不要在timeline记录信息时运行太多次迭代，因为timeline中的记录数量和迭代次数是成正比的。 + + ```python + for pass_id in range(pass_num): + for batch_id, data in enumerate(train_reader()): + if pass_id == 0 and batch_id == 5: + profiler.start_profiler("All") + elif pass_id == 0 and batch_id == 10: + profiler.stop_profiler("total", "/tmp/profile") + exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[]) + ... + ``` + +1. 运行`python paddle/tools/timeline.py`来处理`/tmp/profile`，这个程序默认会生成一个`/tmp/timeline`文件，你也可以用命令行参数来修改这个路径，请参考[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py)。 +```python +python Paddle/tools/timeline.py --profile_path=/tmp/profile --timeline_path=timeline +``` + +1. 打开chrome浏览器，访问，用`load`按钮来加载生成的`timeline`文件。 + + ![chrome tracing](./tracing.jpeg) + +1. 结果如下图所示，可以放到来查看timetime的细节信息。 + + ![chrome timeline](./timeline.jpeg) diff --git a/doc/fluid/howto/optimization/timeline_en.md b/doc/fluid/howto/optimization/timeline_en.md new file mode 100644 index 0000000000000000000000000000000000000000..6f963c6b4da6967fb2f493ada917a4b08917fa4c --- /dev/null +++ b/doc/fluid/howto/optimization/timeline_en.md @@ -0,0 +1,33 @@ +# how to use timeline tool to do profile + +1. Add `profiler.start_profiler(...)`和`profiler.stop_profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number. + + ```python + for pass_id in range(pass_num): + for batch_id, data in enumerate(train_reader()): + if pass_id == 0 and batch_id == 5: + profiler.start_profiler("All") + elif pass_id == 0 and batch_id == 10: + profiler.stop_profiler("total", "/tmp/profile") + exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[]) + ... + ``` + +1. Run `python paddle/tools/timeline.py` to process `/tmp/profile`, it will generate another +file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at +[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details. + +```python +python Paddle/tools/timeline.py --profile_path=/tmp/profile --timeline_path=timeline +``` + +1. Open chrome and visit , use `load` button to load the generated `timeline` file. + + ![chrome tracing](./tracing.jpeg) + +1. The resulting timeline should be like: + + + ![chrome timeline](./timeline.jpeg) diff --git a/doc/fluid/howto/performance/error_clip.md b/doc/fluid/howto/performance/error_clip.md index 58aa73b8cd38d01e2426278a3479714e4fb6a3b0..749cf7693c75696feb17f8556224ed03649baa80 100644 --- a/doc/fluid/howto/performance/error_clip.md +++ b/doc/fluid/howto/performance/error_clip.md @@ -78,7 +78,7 @@ def error_clip_callback(block, context): op_desc = block.desc.op(block.desc.op_size() - 1) for grad_n in filter(lambda n: grad_to_var.has_key(n), op_desc.output_arg_names()): - fwd_var = block.var_recursive(grad_to_var[grad_n]) + fwd_var = block.__var_recursive(grad_to_var[grad_n]) error_clip = getattr(fwd_var, "error_clip", None) if not (error_clip is None or isinstance(error_clip, BaseErrorClipAttr)): diff --git a/doc/fluid/index_cn.rst b/doc/fluid/index_cn.rst index d878d192cae7ee9e8b8fdb4f615839c186fdf334..6b1ef3ceed4f7ed5073d42c13ce103e2ab467e58 100644 --- a/doc/fluid/index_cn.rst +++ b/doc/fluid/index_cn.rst @@ -1,12 +1,16 @@ - PaddlePaddle Fluid -========================== +.. PaddlePaddle Fluid documentation master file, created by + sphinx-quickstart on Thu Jun 7 17:04:53 2018. + You can adapt this file completely to your liking, but it should at least + contain the root `toctree` directive. + +############## +欢迎使用 Fluid +############## .. toctree:: - :maxdepth: 1 + :maxdepth: 1 - getstarted/index_cn.rst - build_and_install/index_cn.rst - design/index_cn.rst - howto/index_cn.rst - dev/index_cn.rst - faq/index_cn.rst + new_docs/beginners_guide/index.rst + new_docs/user_guides/index.rst + new_docs/advanced_usage/index.rst + new_docs/faq/index_cn.rst diff --git a/doc/fluid/new_docs/advanced_usage/benchmark.rst b/doc/fluid/new_docs/advanced_usage/benchmark.rst new file mode 100644 index 0000000000000000000000000000000000000000..7854263bf8f64c840492550fb22152582c7d2361 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/benchmark.rst @@ -0,0 +1,120 @@ +################# +如何进行基准测试 +################# + +本文介绍如何给深度学习框架做基准测试。基准测试主要包含验证模型的精度和性能两方面，下文包含搭建测试环境，选择基准测试模型，验证测试结果等几方面内容。 + +验证深度学习框架，可分为训练和测试两个阶段， 验证指标略有不同，本文只介绍训练阶段的指标验证。训练阶段关注的是模型训练集上的精度，训练集是完备的，因此关注大batch\_size下的训练速度,关注吞吐量，例如图像模型常用的batch\_size=128, 多卡情况下会加大；预测阶段关注的是在测试集上的精度，线上服务测试数据不能提前收集，因此关注小batch\_size下的预测速度，关注延迟，例如预测服务常用的batch\_size=1, 4等。 + +`Fluid `__ 是PaddlePaddle从0.11.0版本开始引入的设计，本文的基准测试在该版本上完成。 + + +环境搭建 +"""""""""""" + +基准测试中模型精度和硬件、框架无关，由模型结构和数据共同决定；性能方面由测试硬件和框架性能决定。框架基准测试为了对比框架之间的差异，控制硬件环境，系统库等版本一致。下文中的对比实验都在相同的硬件条件和系统环境条件下进行. + + +不同架构的GPU卡性能差异巨大，在验证模型在GPU上训练性能时，可使用NVIDIA提供的工具:code `nvidia-smi` 检验当前使用的GPU型号，如果测试多卡训练性能，需确认硬件连接是 `nvlink `__ 或 `PCIe `__ 。 同样地，CPU型号会极大影响模型在CPU上的训练性能。可读取`/proc/cpuinfo`中的参数，确认当前正在使用的CPU型号。 + +下载GPU对应的Cuda Tool Kit和 Cudnn，或者使用NVIDIA官方发布的nvidia-docker镜像 `nvidia-docker `__, 镜像内包含了Cuda和Cudnn，本文采用这种方式。 Cuda Tool Kit包含了GPU代码使用到的基础库，影响在此基础上编译出的Fluid二进制运行性能。 + +准备好Cuda环境后，从github上的下载Paddle并源码编译，会生成对应的最适合当前GPU的sm\_arch二进制\ `sm\_arch `__\ 。另外，cudnn对卷积类任务影响巨大，在基准测试中需要小版本一致，例如Cudnn7.0.2与Cudnn7.1.4在Resnet上有5%以上差异。 + + +选择基准模型 +"""""""""""" + +对框架做基准测试，需要覆盖不同训练任务和不同大小的模型，本文中选取了图像和NLP的最为常用的5个模型。 + +============ ============ ================= ============ +任务种类 模型名称 网络结构 数据集 +============ ============ ================= ============ +图像分类 mnist Lenet mnist +图像分类 VGG VGG-16 Flowers102 +图像分类 Resnet Resnet-50 Flowers102 +文本分类 Stacked-LSTM Stacked-LSTM IMDB +机器翻译 seq-seq Stacked-LSTM wmt14 +============ ============ ================= ============ + +其中mnist, VGG, Resnet属于CNN模型, stacked-lstm, seq2seq代表RNN模型。 +`benchmark `__ +基准模型测试脚本中，均跳过了前几个batch的训练过程，原因是加载数据和分配显存受系统当前运行情况影响，会导致统计性能不准确。运行完若干个轮次后，统计对应指标。 + + +基准模型的数据的选择方面，数据量大且验证效果多的公开数据集为首选。图像模型VGG和resnet, 本文选择了 `flowers102 `__ ，图像大小预处理为和Imagenet相同大小，因此性能可直接对比 +NLP模型的公开且影响力大数据集较少，seq2seq模型选择了wmt14数据，stacked-lstm模型中选择了 `imdb `__ 数据。 + + +注意，图像模型每条样本大小相同，图像经过变换后大小一致，因此经过的计算路径基本相同，计算速度和显存占用波动较小，可以从若干个batch的数据中采样得到当前的训练性能数据。而NLP模型由于样本长度不定，计算路径和显存占用也不相同，因此只能完整运行若干个轮次后，统计速度和显存消耗。 +显存分配是特别耗时的操作，因此Fluid默认会占用所有可用显存空间形成显存池，用以加速计算过程中的显存分配。如果需要统计模型真实显存消耗，可设置环境变量`FLAGS_fraction_of_gpu_memory_to_use=0.0`，观察最大显存开销。 + + +测试过程 +"""""""""""" + +- CPU 单机单线程测试 + +测试CPU上单线程的性能，先设置CUDA的环境变量为空，``CUDA_VISIBLE_DEVICES=``，并通过环境变量关闭OpenMP和MKL的多线程 ``OMP_NUM_THREADS=1``， ``MKL_NUM_THREADS=1;``。 +然后代码中设置为使用CPUPlace，如果使用Paddle代码库中的脚本，只需要命令行参数传入 use_gpu=False即可。 + +.. code-block:: python + + >>> import paddle.fluid as fluid + >>> place = fluid.CPUPlace() + +.. code:: bash + + docker run -it --name CASE_NAME --security-opt seccomp=unconfined -v $PWD/benchmark:/benchmark paddlepaddle/paddle:latest-dev /bin/bash + + +- GPU 单机单卡测试 + +本教程使用了Cuda8, Cudnn7.0.1。来源为:code `nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04` + +.. code:: bash + + nvidia-docker run -it --name CASE_NAME --security-opt seccomp=unconfined -v $PWD/benchmark:/benchmark -v /usr/lib/x86_64-linux-gnu:/usr/lib/x86_64-linux-gnu paddlepaddle/paddle:latest-dev /bin/bash +在单卡上测试，设置CUDA的环境变量使用一块GPU，``CUDA_VISIBLE_DEVICES=0`` +然后代码中设置为使用CUDAPlace，如果使用Paddle代码库中的脚本，只需要命令行参数传入 use_gpu=True即可。 + +.. code-block:: python + + >>> import paddle.fluid as fluid + >>> place = fluid.CUDAPlace(0) // 0 指第0块GPU + + +测试结果 +"""""""""""" + +本教程对比相同环境下的Fluid0.12.0和TensorFlow1.4.0的性能表现。 +硬件环境为 CPU: Intel(R) Xeon(R) CPU E5-2660 v4 @ 2.00GHz, GPU: TITAN X(Pascal) 12G x 1, Nvidia-Driver 384.90。 +系统环境为Ubuntu 16.04.3 LTS, 本文中采用了docker环境，系统版本为nvidia-docker17.05.0-ce。 +测试的Fluid版本为\ `v.0.12.0 `__ 。 +TensorFlow版本为\ `v.1.4.0-rc1 `__ 。 +使用的脚本和配置见\ `benchmark `__ 。 +图表中统计单位为samples/秒。 + +- CPU 单机单线程测试结果 + + ================ ==================== =================== + Speed Fluid CPU TensorFlow CPU + ================ ==================== =================== + mnist 1298.75 samples/s 637.57 samples/s + VGG-16 0.4147 images/s 0.1229 images/s + Resnet-50 1.6935 images/s 0.3657 images/s + Stacked-LSTM 472.3225 words/s 48.2293words/s + Seq2Seq 217.1655 words/s 28.6164 words/s + ================ ==================== =================== + +- GPU 单机单卡测试结果 + + =============== ===================== ================= + Speed Fluid GPU TensorFlow GPU + =============== ===================== ================= + mnist 19710.90 samples/s 15576.3 samples/s + VGG-16 59.83327 images/s 40.9967 images/s + Resnet-50 105.84412 97.8923 images/s + Stacked-LSTM 1319.99315 1608.2526 words/s + Seq2Seq 7147.89081 6845.1161 words/s + =============== ===================== ================= diff --git a/doc/fluid/new_docs/advanced_usage/deploy/anakin_arm_benchmark.md b/doc/fluid/new_docs/advanced_usage/deploy/anakin_arm_benchmark.md new file mode 100644 index 0000000000000000000000000000000000000000..08ea379f81d16407ed5f82770b55a34bcf138da8 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/anakin_arm_benchmark.md @@ -0,0 +1,56 @@ +# Anakin ARM 性能测试 + +## 测试环境和参数: ++ 测试模型Mobilenetv1, mobilenetv2, mobilenet-ssd ++ 采用android ndk交叉编译，gcc 4.9，enable neon， ABI： armveabi-v7a with neon -mfloat-abi=softfp ++ 测试平台 + - 荣耀v9(root): 处理器:麒麟960, 4 big cores in 2.36GHz, 4 little cores in 1.8GHz + - nubia z17:处理器:高通835, 4 big cores in 2.36GHz, 4 little cores in 1.9GHz + - 360 N5:处理器:高通653, 4 big cores in 1.8GHz, 4 little cores in 1.4GHz ++ 多线程：openmp ++ 时间：warmup10次，运行10次取均值 ++ ncnn版本：来源于github的master branch中commits ID：307a77f04be29875f40d337cfff6df747df09de6（msg:convert LogisticRegressionOutput)版本 ++ TFlite版本：来源于github的master branch中commits ID：65c05bc2ac19f51f7027e66350bc71652662125c（msg:Removed unneeded file copy that was causing failure in Pi builds)版本 + +在BenchMark中本文将使用**`ncnn`**、**`TFlite`**和**`Anakin`**进行性能对比分析 + +## BenchMark model + +> 注意在性能测试之前，请先将测试model通过[External Converter](#10003)转换为Anakin model +> 对这些model，本文在ARM上进行多线程的单batch size测试。 + +- [Mobilenet v1](#11) *caffe model 可以在[这儿](https://github.com/shicai/MobileNet-Caffe)下载* +- [Mobilenet v2](#22) *caffe model 可以在[这儿](https://github.com/shicai/MobileNet-Caffe)下载* +- [mobilenet-ssd](#33) *caffe model 可以在[这儿](https://github.com/chuanqi305/MobileNet-SSD)下载* + +### mobilenetv1 + + |platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)| + |:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:| + |麒麟960|107.7ms|61.1ms|38.2ms|152.8ms|85.2ms|51.9ms|152.6ms|nan|nan| + |高通835|105.7ms|63.1ms|~~46.8ms~~|152.7ms|87.0ms|~~92.7ms~~|146.9ms|nan|nan| + |高通653|120.3ms|64.2ms|46.6ms|202.5ms|117.6ms|84.8ms|158.6ms|nan|nan| + +### mobilenetv2 + + |platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)| + |:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:| + |麒麟960|93.1ms|53.9ms|34.8ms|144.4ms|84.3ms|55.3ms|100.6ms|nan|nan| + |高通835|93.0ms|55.6ms|41.1ms|139.1ms|88.4ms|58.1ms|95.2ms|nan|nan| + |高通653|106.6ms|64.2ms|48.0ms|199.9ms|125.1ms|98.9ms|108.5ms|nan|nan| + +### mobilenet-ssd + + |platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)| + |:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:| + |麒麟960|213.9ms|120.5ms|74.5ms|307.9ms|166.5ms|104.2ms|nan|nan|nan| + |高通835|213.0ms|125.7ms|~~98.4ms~~|292.9ms|177.9ms|~~167.8ms~~|nan|nan|nan| + |高通653|236.0ms|129.6ms|96.0ms|377.7ms|228.9ms|165.0ms|nan|nan|nan + +## How to run those Benchmark models? + +1. 首先, 使用[External Converter](../docs/Manual/Converter_en.md)对caffe model 进行转换 +2. 然后将转换后的Anakin model和编译好的benchmark_arm 二进制文件通过'adb push'命令上传至测试机 +3. 接着在测试机含有Anakin model的目录中运行'./benchmark_arm ./ anakin_model.anakin.bin 1 10 10 1' 命令 +4. 最后，终端显示器上将会打印该模型的运行时间 +5. 其中运行命令的参数个数和含义可以通过运行'./benchmark_arm'看到 diff --git a/doc/fluid/new_docs/advanced_usage/deploy/anakin_example.md b/doc/fluid/new_docs/advanced_usage/deploy/anakin_example.md new file mode 100644 index 0000000000000000000000000000000000000000..e6b9e18fe2d64b3fda6382bb23a6a818a3e17fbe --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/anakin_example.md @@ -0,0 +1,28 @@ +# Example +Anakin目前只支持NCHW的格式 +示例文件在test/framework/net下 + +## 在NV的GPU上运行CNN模型 +示例文件为打开example_nv_cnn_net.cpp，整体流程如下： +- 将模型的的path设置为anakin模型的路径，初始化NV平台的图对象。 anakin模型可以通过转换器转化caffe或fluid的模型得到 +- 根据模型设置网络图的输入尺寸，进行图优化 +- 根据优化后的网络图初始化网络执行器 +- 取出网络的输入tensor，将数据拷贝到输入tensor +- 运行推导 +- 取出网络的输出tensor + +以NV平台为例演示Anakin框架的使用方法，注意编译时需要打开GPU编译开关 + +## 在X86上运行RNN模型 +示例文件为example_x86_rnn_net.cpp +整体流程与在NV的GPU上运行CNN模型相似，不同之处如下： +- 使用X86标识初始化图对象和网络执行器对象 +- rnn模型的输入尺寸是可变的，初始化图时的输入维度是维度的最大值，输入维度N代表总的词的个数。还需要设置输入tensor的seq_offset来标示这些词是如何划分为句子的,如{0,5,12}表示共有12个词，其中第0到第4个词是第一句话，第5到第11个词是第二句话 + +以X86平台为例演示Anakin框架的使用方法，注意编译时需要打开X86编译开关 + +## 在NV的GPU上使用Anakin的线程池运行CNN模型 +示例文件为example_nv_cnn_net_multi_thread.cpp ，示例使用worker的同步预测接口 +整体流程与在NV的GPU上运行CNN模型相似，不同之处如下： +- 用模型地址和线程池大小初始化worker对象 +- 将输入tensor注入任务队列,获得输出tensor diff --git a/doc/fluid/new_docs/advanced_usage/deploy/anakin_gpu_benchmark.md b/doc/fluid/new_docs/advanced_usage/deploy/anakin_gpu_benchmark.md new file mode 100644 index 0000000000000000000000000000000000000000..667f9396f1169a0d891b9e6b0e912aa5527ab0b8 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/anakin_gpu_benchmark.md @@ -0,0 +1,170 @@ +# Anakin GPU Benchmark + +## Machine: + +> CPU: `12-core Intel(R) Xeon(R) CPU E5-2620 v2 @2.10GHz` +> GPU: `Tesla P4` +> cuDNN: `v7` + + +## Counterpart of anakin : + +The counterpart of **`Anakin`** is the acknowledged high performance inference engine **`NVIDIA TensorRT 3`** , The models which TensorRT 3 doesn't support we use the custom plugins to support. + +## Benchmark Model + +The following convolutional neural networks are tested with both `Anakin` and `TenorRT3`. + You can use pretrained caffe model or the model trained by youself. + +> Please note that you should transform caffe model or others into anakin model with the help of [`external converter ->`](../docs/Manual/Converter_en.md) + + +- [Vgg16](#1) *caffe model can be found [here->](https://gist.github.com/jimmie33/27c1c0a7736ba66c2395)* +- [Yolo](#2) *caffe model can be found [here->](https://github.com/hojel/caffe-yolo-model)* +- [Resnet50](#3) *caffe model can be found [here->](https://github.com/KaimingHe/deep-residual-networks#models)* +- [Resnet101](#4) *caffe model can be found [here->](https://github.com/KaimingHe/deep-residual-networks#models)* +- [Mobilenet v1](#5) *caffe model can be found [here->](https://github.com/shicai/MobileNet-Caffe)* +- [Mobilenet v2](#6) *caffe model can be found [here->](https://github.com/shicai/MobileNet-Caffe)* +- [RNN](#7) *not support yet* + +We tested them on single-GPU with single-thread. + +### VGG16 + +- Latency (`ms`) of different batch + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 8.8690 | 8.2815 | +| 2 | 15.5344 | 13.9116 | +| 4 | 26.6000 | 21.8747 | +| 8 | 49.8279 | 40.4076 | +| 32 | 188.6270 | 163.7660 | + +- GPU Memory Used (`MB`) + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 963 | 997 | +| 2 | 965 | 1039 | +| 4 | 991 | 1115 | +| 8 | 1067 | 1269 | +| 32 | 1715 | 2193 | + + +### Yolo + +- Latency (`ms`) of different batch + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 16.4596| 15.2124 | +| 2 | 26.6347| 25.0442 | +| 4 | 43.3695| 43.5017 | +| 8 | 80.9139 | 80.9880 | +| 32 | 293.8080| 310.8810 | + +- GPU Memory Used (`MB`) + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 1569 | 1775 | +| 2 | 1649 | 1815 | +| 4 | 1709 | 1887 | +| 8 | 1731 | 2031 | +| 32 | 2253 | 2907 | + +### Resnet50 + +- Latency (`ms`) of different batch + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 4.2459 | 4.1061 | +| 2 | 6.2627 | 6.5159 | +| 4 | 10.1277 | 11.3327 | +| 8 | 17.8209 | 20.6680 | +| 32 | 65.8582 | 77.8858 | + +- GPU Memory Used (`MB`) + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 531 | 503 | +| 2 | 543 | 517 | +| 4 | 583 | 541 | +| 8 | 611 | 589 | +| 32 | 809 | 879 | + +### Resnet101 + +- Latency (`ms`) of different batch + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 7.5562 | 7.0837 | +| 2 | 11.6023 | 11.4079 | +| 4 | 18.3650 | 20.0493 | +| 8 | 32.7632 | 36.0648 | +| 32 | 123.2550 | 135.4880 | + +- GPU Memory Used (`MB)` + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 701 | 683 | +| 2 | 713 | 697 | +| 4 | 793 | 721 | +| 8 | 819 | 769 | +| 32 | 1043 | 1059 | + +### MobileNet V1 + +- Latency (`ms`) of different batch + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 45.5156 | 1.3947 | +| 2 | 46.5585 | 2.5483 | +| 4 | 48.4242 | 4.3404 | +| 8 | 52.7957 | 8.1513 | +| 32 | 83.2519 | 31.3178 | + +- GPU Memory Used (`MB`) + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 329 | 283 | +| 2 | 345 | 289 | +| 4 | 371 | 299 | +| 8 | 393 | 319 | +| 32 | 531 | 433 | + +### MobileNet V2 + +- Latency (`ms`) of different batch + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 65.6861 | 2.9842 | +| 2 | 66.6814 | 4.7472 | +| 4 | 69.7114 | 7.4163 | +| 8 | 76.1092 | 12.8779 | +| 32 | 124.9810 | 47.2142 | + +- GPU Memory Used (`MB`) + +| BatchSize | TensorRT | Anakin | +| --- | --- | --- | +| 1 | 341 | 293 | +| 2 | 353 | 301 | +| 4 | 385 | 319 | +| 8 | 421 | 351 | +| 32 | 637 | 551 | + +## How to run those Benchmark models? + +> 1. At first, you should parse the caffe model with [`external converter`](https://github.com/PaddlePaddle/Anakin/blob/b95f31e19993a192e7428b4fcf852b9fe9860e5f/docs/Manual/Converter_en.md). +> 2. Switch to *source_root/benchmark/CNN* directory. Use 'mkdir ./models' to create ./models and put anakin models into this file. +> 3. Use command 'sh run.sh', we will create files in logs to save model log with different batch size. Finally, model latency summary will be displayed on the screen. +> 4. If you want to get more detailed information with op time, you can modify CMakeLists.txt with setting `ENABLE_OP_TIMER` to `YES`, then recompile and run. You will find detailed information in model log file. diff --git a/doc/fluid/new_docs/advanced_usage/deploy/anakin_tutorial.md b/doc/fluid/new_docs/advanced_usage/deploy/anakin_tutorial.md new file mode 100644 index 0000000000000000000000000000000000000000..5efbc89abd469871b318c306e8cb03dd95f0c85b --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/anakin_tutorial.md @@ -0,0 +1,639 @@ +# Anakin 使用教程 ## + +本教程将会简略的介绍Anakin的工作原理，一些基本的Anakin API，以及如何调用这些API。 + +## 内容 ### + +- [Anakin的工作原理](#principle) +- [Anakin APIs](#api) +- [示例代码](#example) + +## Anakin的工作原理 ### + +![Anakin_principle](../pics/anakin_fm_ch.png) + +用Anakin来进行前向计算主要分为三个步骤： + +- 将外部模型通过[Anakin Parser](Converter_ch.md)解析为Anakin模型 + 在使用Anakin之前，用户必须将所有其他模型转换成Anakin模型，我们提供了转换脚本，用户可通过[Anakin Parser](Converter_ch.md)进行模型转换。 +- 生成Anakin计算图 + 加载Anakin模型生成原始计算图，然后需要对原始计算图进行优化。你只需要调用相应的API优化即可。 +- 执行计算图 + Anakin会选择不同硬件平台执行计算图。 + + +## Anakin APIs ### +### Tensor #### + +`Tensor`提供基础的数据操作和管理，为ops提供统一的数据接口。`Tensor`包含以下几个属性： + +- Buffer + 数据存储区 +- Shape + 数据的维度信息 +- Event + 用于异步计算的同步 + + `Tensor` 类包含三个`Shape`对象， 分别是`_shape`, `_valid_shape`和 `offset`。 `_shape`为`tensor`真正空间信息，`_valid_shape`表示当前`tensor`使用的空间信息， `_offset`表示当前`tensor`数据指针相对于真正数据空间的信息。 `Tensor`不同维度与分别与数学中的向量、矩阵等相对应如下表所示。 + + +Dimentions | Math entity | + :----: | :----: +1 | vector +2 | matrix +3 | 3-tensor +n | n-tensor + +#### 声明tensor对象 + +`Tensor`接受三个模板参数: + + +```c++ + template + class Tensor .../* Inherit other class */{ + //some implements + ... + }; +``` + +TargetType是平台类型，如X86，GPU等等，在Anakin内部有相应的标识与之对应；datatype是普通的数据类型，在Anakin内部也有相应的标志与之对应；[LayOutType](#layout)是数据分布类型，如batch x channel x height x width [NxCxHxW], 在Anakin内部用一个struct来标识。 Anakin中数据类型与基本数据类型的对应如下: + +1. TargetType + + Anakin TargetType | platform + :----: | :----:| + NV | NVIDIA GPU + ARM | ARM + AMD | AMD GPU + X86 | X86 + NVHX86 | NVIDIA GPU with Pinned Memory + +2. DataType + +Anakin DataType | C++ | Description +:---: | :---: | :---: | +AK_HALF | short | fp16 +AK_FLOAT | float | fp32 +AK_DOUBLE | double | fp64 +AK_INT8 | char | int8 +AK_INT16 | short | int16 +AK_INT32 | int | int32 +AK_INT64 | long | int64 +AK_UINT8 | unsigned char | uint8 +AK_UINT16 | unsigned short | uint8 +AK_UINT32 | unsigned int | uint32 +AK_STRING | std::string | / +AK_BOOL | bool | / +AK_SHAPE | / | Anakin Shape +AK_TENSOR | / | Anakin Tensor + + +3. LayOutType + +Anakin LayOutType ( Tensor LayOut ) | Tensor Dimention | Tensor Support | Op Support +:---: | :---: | :---: | :---: | +W | 1-D | YES | NO +HW | 2-D | YES | NO +WH | 2-D | YES | NO +NW | 2-D | YES | YES +NHW | 3-D | YES |YES +NCHW ( default ) | 4-D | YES | YES +NHWC | 4-D | YES | NO +NCHW_C4 | 5-D | YES | YES + + +理论上，Anakin支持申明1维以上的tensor，但是对于Anakin中的Op来说，只支持NW、NHW、NCHW、NCHW_C4这四种LayOut，其中NCHW是默认的LayOutType，NCHW_C4是专门针对于int8这种数据类型的。 + + +例子 + +> 下面的代码将展示如何使用tensor， 我们建议先看看这些示例。 + +> 要想获得更多关于tensor的信息， 请参考 *soure_path/core/tensor.h* + +> 1. 使用shape对象初始化tensor +``` c++ + //create a null tensor. A null tensor holds for nothing. + //tensor's buffer is resident at CPU and its datatype is AK_FLOAT. + //tensor's Layout is NCHW(default) + Tensor mytensor; + + //1. using shape object to create a tensor. + Shape shape1(NUM); //1-D shape. NUM is the number of dimention. + Tensor mytensor1(shape1); //1-D tensor. + + // A 4-D shape + Shape shape2(N, C, H, W); // batch x channel x height x width +``` + +>`注意：Shape的维度必须和tensor的`[LayoutType](#layout)`相同，比如Shape(N,C,H,W), 那么Tensor的 LayoutType必须是NCHW，否则会出错。如下列代码所示` + + +```c++ + // A 4-D tensor. + Tensor mytensor2(shape2); //right + + //A 4-D tensor which is resident at GPU and its datatype is AK_INT8 + Tensor mytensor3(shape2); //right + + Tensor mytensor4(shape2); //wrong!! shape's dimetion must be equal to tensor's Layout. + Tensor mytensor5(shape2); //wrong!!!! + +``` + +> 2. 使用现有的数据和shape初始化tensor + +```c++ + + /** + * A construtor of Tensor. + * data_ptr is a pointer to any data type of data + * TargetType is type of a platform [Anakin TargetType] + * id : device id + * shape: a Anakin shape + */ + Tensor(Dtype* data_ptr, TargetType_t target, int id, Shape shape); + + //using existing data feed to a tensor + Tensor mytensor(data_ptr, TargetType, device_id, shape); //shape must has dimention (N, C, H, W). + +``` + +> 3. 使用tensor初始化tensor + +```c++ + Tensor tensor(exist_tensor); +``` + + +> 提示： 你可以用` typedef Tensor Tensor4d_X86 `方便定义tensor + + +#### 填充tensor数据区 + + +填充数据区得看你申明tensor的方式， 下面展示了如何填充tensor的数据区。 + +```c++ +首先来看看tensor的四种声明方式： + +1. Tensor mytensor; +2. Tensor mytensor1(shape1); +3. Tensor mytensor(data_ptr, TargetType, device_id, shape); +4. Tensor tensor(exist_tensor); + + +相关的声明方式的数据填充方法如下： + +1：声明一个空的tensor，此时没有为其分配内存，所以，我们需要手动的为其分配内存。 + + //parama shape + mytensor.re_alloc(Shape shape); + + //Get writable pointer to mytensor. + //parama index (int): where you start to write. + //Dtype is your data type such int, float or double. + Dtype *p = mytensor.mutable_data(index/*=0*/); + //write data to mytensor + for(int i = 0; i < mytensor.size(); i++){ + p[i] = 1.0f; + } + //do something ... + +2: 这种声明方式会自动分配内存 + + //Get writable pointer to mytensor. + //parama index (int): where you start to write. + //Dtype is your data type such int, float or double. + Dtype *p = mytensor1.mutable_data(index/*=0*/); + //write data to mytensor + for(int i = 0; i < mytensor.size(); i++){ + p[i] = 1.0f; + } + //do something ... + + +3：在该种声明方式中，我们仍不需要手动为其分配内存。但在构造函数内部是否为其分配内存，得依情况而定。如果data_ptr和申明的 +tensor都在都一个目标平台上，那么该tensor就会与data_ptr共享内存空间，相反，如果他们不在同一个平台上（如data_ptr在X86上，而 +tensor在GPU上），那么此时tensor就会开辟一个新的内存空间，并将data_ptr所指向的数据拷贝到tensor的buffer中。 + + //Get writable pointer to mytensor. + //parama index (int): where you start to write. + //Dtype is your data type such int, float or double. + Dtype *p = mytensor.mutable_data(index/*=0*/); + //write data to mytensor + for(int i = 0; i < mytensor.size(); i++){ + p[i] = 1.0f; + } + //do something ... + +4：该种方式仍不需要手动分配内存 + + //Get writable pointer to mytensor. + //parama index (int): where you start to write. + //Dtype is your data type such int, float or double. + Dtype *p = mytensor.mutable_data(index/*=0*/); + //write data to mytensor + for(int i = 0; i < mytensor.size(); i++){ + p[i] = 1.0f; + } + //do something ... + + +另外，你还可以获取一个tensor的可读指针，示例如下： + //Get read-only pointer to mytensor. + //parama index (int): where you start to read. + //Dtype is your data type such int, float or double. + Dtype *p = mytensor.data(index/*=0*/); + //do something ... +``` + +如果想更详细的了解tensor，请查阅*soure_path/saber/core/tensor.h* + +#### 获取tensor的shape + +```c++ +//some declarations +// ... +Shape shape = mytensor.shape(); + +//Get a first dimetion size of tesor, if it has. +int d1 = shape[0]; + +//Get a second dimention size of tensor, if it has. +int d2 = shape[1]; + +... + +//Get a n-th dimention size of tensor, if it has. +int dn = shape[n-1]; + + +//Get a tensor's dimention +int dims = mytensor.dims(); + +//Get the size of tensor. +//size = d1 x d2 x ... x dn. +int size = mytensor.size(); + +//Get the size of tensor at interval [Di, Dj) +// form i-th dimention to j-th dimention, but not including the j-th dimention. +// which means di x (di+1) x ... x (dj -1) +int size = mytensor.count(start, end); +``` + +#### 设置tensor的shape + +我们可以用tensor的成员函数set_shape来设置tensor的shape。 下面是set_shape的定义 + + +```c++ +/** + * \brief set a tensor's shape + * \param valid_shape [a Shape object] + * \param shape [a Shape object] + * \param offset [a Shape object] + * \return the status of this operation, that means whether it success * or not. + */ +SaberStatus set_shape(Shape valid_shape, Shape shape = Shape::zero(TensorAPI::layout_dims::value), Shape offset = Shape::minusone(TensorAPI::layout_dims::value)); +``` + +这个成员函数只设置tensor的shape。这些shape对象(valid_shape, shape, offset)的[LayOutType](#layout)必须和当前的tensor的相应三个shape对象的LayOutType相同，如果不同就会出错，返回SaberInvalidValue。 如果相同，那么将成功设置tensor的shape。 + +```c++ + +// some declarations +// ... +//valid_shape, shape , offset are Shape object; +//All these Shape object's LayOutType must be equal to mytensor's. +mytensor.set_shape(valid_shape, shape, offset); + +``` + +#### 重置 tensor的shape + +```c++ +//some declarations +Shape shape, valid_shape, offset; + +//do some initializations +... +mytensor.reshape(valid_shape, shape, offset); +``` + +注意： Reshape操作仍然需要shape的[LayOutType](#layout) 与tensor的相同 + + +### Graph ### + +`Graph`类负责加载Anakin模型生成计算图、对图进行优化、存储模型等操作。 + +#### 图的声明 + +与`Tensor`一样，graph也接受三个模板参数。 + +```c++ + +template +class Graph ... /* inherit other class*/{ + + //some implements + ... + +}; +``` + +前面已经介绍过[TargetType](#target)和[DataType](#datatype)是Anakin内部自定义数据类型。[TargetType](#target)表示平台类型 (如NV、X86), [DataType](#datatype)是Anakin基本数据类型与C++/C中的基本数据类型相对应。 [Precision](#precision)为op所支持的精度类型, 稍后我们在介绍它。 + + +```c++ + +//Create a empty graph object. +Graph graph = Graph tmp(); + +//Create a pointer to a empty graph. +Graph *graph = new Graph(); + +//Create a pointer to a empty graph. +auto graph = new Graph(); + +``` + +#### 加载 Anakin 模型 + +```c++ +//some declarations +... +auto graph = new Graph(); +std::string model_path = "the/path/to/where/your/models/are"; +const char *model_path1 = "the/path/to/where/your/models/are"; + +//Loading Anakin model to generate a compute graph. +auto status = graph->load(model_path); + +//Or this way. +auto status = graph->load(model_path1); +//Check whether load operation success. +if(!status){ + std::cout << "error" << endl; + //do something... +} + +``` + +#### 优化计算图 + +```c++ +//some declarations +... +//Load graph. +... +//According to the ops of loaded graph, optimize compute graph. +graph->Optimize(); + +``` + +> 注意： 第一次加载原始图，必须要优化。 + +#### 保存模型 + +你可以在任何时候保存模型， 特别的， 你可以保存一个优化的模型，这样，下次再加载模型时，就不必进行优化操作。 + + +```c++ +//some declarations +... +//Load graph. +... +// save a model +//save_model_path: the path to where your model is. +auto status = graph->save(save_model_path); + +//Checking +if(!status){ + cout << "error" << endl; + //do somethin... +} +``` + +#### 重新设置计算图里的tensor的shape + +```c++ +//some declarations +... +//Load graph. +... +vector shape{10, 256, 256, 10}; +//input_name : std::string. +//Reshape a tensor named input_name. +graph->Reshape(input_name, shape);//Note: shape is a vector, not a Shape object. +``` + +#### 设置 batch size + +`Graph` 支持重新设置batch size的大小。 + +```c++ +//some declarations +... +//Load graph. +... +//input_name : std::string. +//Reset a tensor named input_name. +int new_batch_size = 4; +graph->ResetBatchSize(input_name, new_batch_size); +``` + +### Net ### + + +`Net` 是计算图的执行器。你可以通过Net对象获得输入和输出 +#### Creating a graph executor + +`Net`接受四个模板参数。 + + +```c++ +template +class Net{ + //some implements + ... + +}; +``` +由于有些Op可能支持多种精度，我们可以通过Precision来指定。OpRunType表示同步或异步类型，异步是默认类型。OpRunType::SYNC表示同步，在GPU上只有单个流；OpRunType::ASYNC表示异步，在GPU上有多个流并以异步方式执行。实际上，Precision和OpRunType都是enum class, 详细设计请参考*source_root/framework/core/types.h*. + + +1. Precision + +Precision | Op support +:---: | :---: +Precision::INT4 | NO +Precision::INT8 | NO +Precision::FP16 | NO +Precision::FP32 | YES +Precision::FP64 | NO + +现在Op的精度只支持FP32， 但在将来我们会支持剩下的Precision. + + + +2. OpRunType + +OpRunType | Sync/Aync |Description +:---: | :---: | :---: +OpRunType::SYNC | Synchronization | single-stream on GPU +OpRunType::ASYNC | Asynchronization | multi-stream on GPU + +用graph对象创建一个执行器。 +```c++ +//some declarations +... +//Create a pointer to a graph. +auto graph = new Graph(); +//do something... +... + +//create a executor +Net executor(*graph); + +``` + +#### 获取输入输出tensor + + +获取输入输出tensor，并填充输入tensor的buffer。如果想要获取输入和输出tensor，那么必须指定输入的名字，如"input_0", "input_1", "input_2", ..., 必须传入如上字符串才能够获得输入tensor。另外，如果想知道input_i对应哪个输入，你需要去dash board查看，如何使用dash board请看[Anakin Parser](Converter_ch.md)。请看如下示例代码 + +```c++ +//some declaratinos +... + +//create a executor +//TargetType is NV [NVIDIA GPU] +Net executor(*graph); + +//Get the first input tensor. +//The following tensors(tensor_in0, tensor_in2 ...) are resident at GPU. +//Note: Member function get_in returns an pointer to tensor. +Tensor* tensor_in0 = executor.get_in("input_0"); + +//If you have multiple input tensors +//You just type this code below. +Tensor* tensor_in1 = executor.get_in("input_1"); +... +auto tensor_inn = executor.get_in("input_n"); +``` + +当得到输入tensor之后，就可以填充它的数据区了。 + +```c++ +//This tensor is resident at GPU. +auto tensor_d_in = executor.get_in("input_0"); + +//If we want to feed above tensor, we must feed the tensor which is resident at host. And then copy the host tensor to the device's one. + +//using Tensor4d = Tensor; +Tensor4d tensor_h_in; //host tensor; +//Tensor tensor_h_in; + +//Allocate memory for host tensor. +tensor_h_in.re_alloc(tensor_d_in->valid_shape()); +//Get a writable pointer to tensor. +float *h_data = tensor_h_in.mutable_data(); + +//Feed your tensor. +/** example +for(int i = 0; i < tensor_h_in.size(); i++){ + h_data[i] = 1.0f; +} +*/ +//Copy host tensor's data to device tensor. +tensor_d_in->copy_from(tensor_h_in); + +// And then +``` + + +类似的，我们可以利用成员函数get_out来获得输出tensor。但与获得输入tensor不同的是， 我们需要指定输入tensor结点的名字，这个可以从dash board中看到，请从[Anakin Parser](Converter_ch.md)中查看dash board的使用方法。假如有个输出结点叫pred_out, 那么我们可以通过如下代码获得相应的输出tensor： +```c++ +//Note: this tensor are resident at GPU. +Tensor* tensor_out_d = executor.get_out("pred_out"); + +``` + + +#### Executing graph + + +当一切准备就绪后，我们就可以执行真正的计算了！ +```c++ +executor.prediction(); +``` + +## 示例代码 ## + +下面的例子展示了如何调用Anakin。 + +在这儿之前， 请确保你已经有了Anakin模型。如果还没有，那么请使用[Anakin Parser](Converter_ch.md)转换你的模型。 + +### Single-thread + +单线程例子在 *source_root/test/framework/net/net_exec_test.cpp`* + +```c++ + +std::string model_path = "your_Anakin_models/xxxxx.anakin.bin"; +// Create an empty graph object. +auto graph = new Graph(); +// Load Anakin model. +auto status = graph->load(model_path); +if(!status ) { + LOG(FATAL) << " [ERROR] " << status.info(); +} +// Reshape +graph->Reshape("input_0", {10, 384, 960, 10}); +// You must optimize graph for the first time. +graph->Optimize(); +// Create a executer. +Net net_executer(*graph); + +//Get your input tensors through some specific string such as "input_0", "input_1", and +//so on. +//And then, feed the input tensor. +//If you don't know Which input do these specific string ("input_0", "input_1") correspond with, you can launch dash board to find out. +auto d_tensor_in_p = net_executer.get_in("input_0"); +Tensor4d h_tensor_in; +auto valid_shape_in = d_tensor_in_p->valid_shape(); +for (int i=0; icopy_from(h_tensor_in); + +//Do inference. +net_executer.prediction(); + +//Get result tensor through the name of output node. +//And also, you need to see the dash board again to find out how many output nodes are and remember their name. + +//For example, you've got a output node named obj_pre_out +//Then, you can get an output tensor. +auto d_tensor_out_0_p = net_executer.get_out("obj_pred_out"); //get_out returns a pointer to output tensor. +auto d_tensor_out_1_p = net_executer.get_out("lc_pred_out"); //get_out returns a pointer to output tensor. +//...... +// do something else ... +//... +//save model. +//You might not optimize the graph when you load the saved model again. +std::string save_model_path = model_path + std::string(".saved"); +auto status = graph->save(save_model_path); +if (!status ) { + LOG(FATAL) << " [ERROR] " << status.info(); +} + +``` diff --git a/doc/fluid/new_docs/advanced_usage/deploy/build_and_install_lib_cn.rst b/doc/fluid/new_docs/advanced_usage/deploy/build_and_install_lib_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..3884284ea020fe94ed9c03ec84c856ee44aa8c3f --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/build_and_install_lib_cn.rst @@ -0,0 +1,99 @@ +.. _install_or_build_cpp_inference_lib: + +安装与编译C++预测库 +=========================== + +直接下载安装 +------------- + +====================== ======================================== +版本说明 C++预测库 +====================== ======================================== +cpu_avx_mkl `fluid.tgz `_ +cpu_avx_openblas `fluid.tgz `_ +cpu_noavx_openblas `fluid.tgz `_ +cuda7.5_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn7_avx_mkl `fluid.tgz `_ +cuda9.0_cudnn7_avx_mkl `fluid.tgz `_ +====================== ======================================== + +从源码编译 +---------- +用户也可以从 PaddlePaddle 核心代码编译C++预测库，只需在编译时配制下面这些编译选项： + +================= ========= +选项 值 +================= ========= +CMAKE_BUILD_TYPE Release +FLUID_INSTALL_DIR 安装路径 +WITH_FLUID_ONLY ON（推荐） +WITH_SWIG_PY OFF（推荐 +WITH_PYTHON OFF（推荐） +WITH_GPU ON/OFF +WITH_MKL ON/OFF +================= ========= + +建议按照推荐值设置，以避免链接不必要的库。其它可选编译选项按需进行设定。 + +下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）： + + .. code-block:: bash + + pip install paddlepaddle-gpu + PADDLE_ROOT=/path/of/capi + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + mkdir build + cd build + cmake -DFLUID_INSTALL_DIR=$PADDLE_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_FLUID_ONLY=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_PYTHON=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=OFF \ + .. + make + make inference_lib_dist + +成功编译后，使用C++预测库所需的依赖（包括：（1）编译出的PaddlePaddle预测库和头文件；（2）第三方链接库和头文件；（3）版本信息与编译选项信息） +均会存放于PADDLE_ROOT目录中。目录结构如下： + + .. code-block:: text + + PaddleRoot/ + ├── CMakeCache.txt + ├── paddle + │   └── fluid + │   ├── framework + │   ├── inference + │   ├── memory + │   ├── platform + │   ├── pybind + │   └── string + ├── third_party + │   ├── boost + │   │   └── boost + │   ├── eigen3 + │   │   ├── Eigen + │   │   └── unsupported + │   └── install + │   ├── gflags + │   ├── glog + │   ├── mklml + │   ├── protobuf + │   ├── snappy + │   ├── snappystream + │   └── zlib + └── version.txt + +version.txt 中记录了该预测库的版本信息，包括Git Commit ID、使用OpenBlas或MKL数学库、CUDA/CUDNN版本号，如： + + .. code-block:: text + + GIT COMMIT ID: c95cd4742f02bb009e651a00b07b21c979637dc8 + WITH_MKL: ON + WITH_GPU: ON + CUDA version: 8.0 + CUDNN version: v5 diff --git a/doc/fluid/new_docs/advanced_usage/deploy/convert_paddle_to_anakin.md b/doc/fluid/new_docs/advanced_usage/deploy/convert_paddle_to_anakin.md new file mode 100644 index 0000000000000000000000000000000000000000..56ca582b2b47f404ede777712830731ea7f4e9b5 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/convert_paddle_to_anakin.md @@ -0,0 +1,73 @@ +# 模型转换指南 + +Anakin 支持不同框架的模型预测。但由于格式的差别，Anakin 需要您预先转换模型。本文档介绍如何转换模型。 + +## 简介 + +Anakin 模型转换器输入支持 Caffe 和 Fluid 两种格式的预测模型，模型包含网络结构（model 或 prototxt）和权重参数（param 或 caffemodel）。 + +模型转换的输出是一个 bin 文件，它作为 Anakin 框架的 graph 参数导入。 + +您还可以使用模型转换器的 launch board 功能生成网络结构的 HTML 预览。 + + +## 系统要求 + +- python 2.7+ +- pyyaml +- flask +- protobuf 3.5+ + + +## 用法 + +### 1、环境 +转换器所需的依赖标注于 *系统要求* 一节。 + +### 2、配置 +您需要对 *config.yaml* 文件进行修改以告知您的需求。工程中给出了 *config.yaml* 示例，下面作进一步说明。 + +#### config.yaml +```bash +OPTIONS: + Framework: CAFFE # 依框架类型填写 CAFFE 或 FLUID + SavePath: ./output # 转换结束后模型的保存位置 + ResultName: googlenet # 输出模型的名字 + Config: + LaunchBoard: ON # 是否生成网络结构预览页面 + Server: + ip: 0.0.0.0 + port: 8888 # 从一个可用端口访问预览页面 + OptimizedGraph: # 当您使用了 Anakin 框架的 Optimized 功能时，才应该打开此项 + enable: OFF + path: /path/to/anakin_optimized_anakin_model/googlenet.anakin.bin.saved + LOGGER: + LogToPath: ./log/ # 生成日志的路径 + WithColor: ON + +TARGET: + CAFFE: + # 当 Framework 为 CAFFE 时需填写 + ProtoPaths: + - /path/to/caffe/src/caffe/proto/caffe.proto + PrototxtPath: /path/to/your/googlenet.prototxt + ModelPath: /path/to/your/googlenet.caffemodel + + FLUID: + # 当 Framework 为 FLUID 时需填写 + Debug: NULL + ProtoPaths: + - / + PrototxtPath: /path/to/fluid/inference_model + ModelPath: /path/to/fluid/inference_model + # ... +``` + +### 3、转换 +在完成配置文件的修改后，您只需执行 ```python converter.py``` 就可以进行模型转换了。 + + +### 4、预览 +最后一步，就是在浏览器中查看令人振奋的转换结果！网址是在 *config.yaml* 中配置的，例如 http://0.0.0.0:8888 。 + +> 注意：若您使用了默认的 IP 地址 0.0.0.0，请在预览时使用真实的服务器地址 real_ip:port 替代它。 diff --git a/doc/fluid/new_docs/advanced_usage/deploy/how_to_add_anakin_op.md b/doc/fluid/new_docs/advanced_usage/deploy/how_to_add_anakin_op.md new file mode 100644 index 0000000000000000000000000000000000000000..f2783eb9f591a31443f2a692ce0eb1bcc9b1063a --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/how_to_add_anakin_op.md @@ -0,0 +1,405 @@ +# 如何增加新的Operator + +## 基本概念 + +简单介绍下几个同Operator相关的基本概念，详情请参考设计文档。 + +```framework```: 上层的逻辑代码，负责从parser中获取参数及weights，添加op时主要修改framework/operator目录下的内容。 + +```saber```: 底层的实现代码，Anakin通过saber封装了不同的backends，不同的实现(impl)分别特化出自己的实现，外层framework通过不同的template进入各自的impl完成调用。各个op的parameter放在saber/saber_funcs_param.h文件中，增加op主要修改saber/funcs下的内容。 + +saber的文件结构： +* saber/funcs下的是各个funcs的外部接口，这一层的op与具体的设备实现无关，只与各op完成的功能有关。由于跟实现(impl)无关，本层文件明均不带impl。 +* saber/funcs/impl下是各个op的impl声明，特定设备需要完成该层声明的特化版本，如saber/funcs/impl/x86实现了上一层impl声明的x86特化版本，saber/funcs/impl/cuda实现了上一层impl声明的NV特化版本。当增加新的backends时需要特化出新的实现。本层代码同实现相关，均带有```impl_```前缀。 +* saber/funcs/impl/cuda/base/cuda_c内有cuda```.cu```扩展名的文件，添加cuda的kernel需要在该文件目录下添加。 +* saber/funcs/impl/cuda/base/sass 内有不同架构的汇编代码编译的静态库。 + +### 涉及到的基类及各个类之前的关系 + +简单介绍相关的基类 + +* ```anakin::Operator```: framework的operator基类，位于framework/core/operator/operator.h + +* ```anakin::saber::BaseFunc```: saber对外的op接口基类，提供统一的对外接口，位于saber/funcs/base.h。BaseFunc的```compute_output_shape```接口只根据input的shape和param的参数计算输出的shape，并通过```tensor```的```set_shape```接口(只设置shape，不分配空间)设置到output中。```operator()```接口为各个op的计算接口。 + +* ```ankain::saber::ImplBase```: saber设备实现的op的接口，所有设备相关实现的基类。位于saber/funcs/impl/impl_base.h。实现版本中这里分为两类，一类以```vender_```为前缀，带有```vender_```代码意为使用第三方库来实现该op，如cudnn的conv，或mkl的conv等等，这类op的性能我们难以调优，因此单独列为一类。另一类是带有源码的saber实现，这些实现都带有```saber_```为前缀，此类实现带有源码，能够通过后续优化不断提升性能，实现起名时需要注意这一点。 + +## 添加operator + +添加一个新的op需要以下几步： + +1. 添加saber的param +2. 定义saber的Operator类 +3. 定义新的impl声明 +3. 完成新的impl实现 +4. 增加framework的实现或特化 + +接下来就针对这几步，以一个简单例子为例介绍实现。 + +例如我们要添加新的Mul op。给出计算公式如下：$$Out = alpha \dot X * Y$$ + +### 为operator增加param + +涉及到的文件：```saber/saber_funcs_param.h```。如果之前已经存在需要添加的op的param，这一步可以跳过。 +这里```XXXParam```是一个```struct```。包含一个无参数的构造函数，含参数的构造函数，复制构造函数，```operator=()```及```operator==()```。 +``` +template // 能够获得target, datatype, layout +struct MulParam{ + MulParam() + : alpha(0) + {} + MulParam(float alpha_in) + : alpha(alpha_in) + {} + MulParam(const MulParam& right) + : alpha(right.alpha) + {} + MulParam &operator=(const MulParam &right) { + alpha = right.alpha; + } + bool operator==(const MulParam &right) { + return alpha == right.alpha; + } + float alpha; +}; +``` + +### 定义Operator类 +涉及到的文件:```saber/funcs/mul.h```。如果之前定义过该op的类，这里需要修改输入的impl定义头文件。 +下面给出一个相对完整的定义结构供参考。 +``` +//不同的设备需要包含对应的operator实现.[详见](#impl) +#ifdef NVIDIA_GPU +#include "saber/funcs/impl/cuda/saber_mul.h" +#include "saber/funcs/impl/cuda/vender_mul.h" +#endif +//如果一个设备现在还没有对应的operator实现，需要包含声明。[详见](#declare) +#ifdef USE_X86_PLACE +#include "saber/funcs/impl/impl_mul.h" +#endif +namespace anakin { +namespace saber { +template +class Mul : public BaseFunc< + Tensor, + Tensor, + Tensor, + ImplBase, MulParam> { +public: + using BaseFunc< + Tensor, + Tensor, + Tensor, + ImplBase, MulParam>::BaseFunc; + Mul() = default; + typedef Tensor InDataTensor; + typedef Tensor OutDataTensor; + typedef Tensor OpTensor; + typedef MulParam Param_t; + typedef std::vector Input_v; + typedef std::vector Output_v; + typedef std::vector Shape_v; + + virtual SaberStatus compute_output_shape(const Input_v &input, + Output_v &output, Param_t ¶m) override { + //计算输出的shape， + Shape output_shape = (input[0]->valid_shape()); + /* code */ + return output[0]->set_shape(output_shape); + } + virtual SaberStatus init_impl(ImplEnum implenum) override { + // 不同设备均使用此init_impl, 此接口创建对应impl的实现。 + switch (implenum) { + case VENDER_IMPL: + this->_impl.push_back(new VenderMul ); + return SaberSuccess; + case SABER_IMPL: + this->_impl.push_back(new SaberMul ); + return SaberSuccess; + default: + return SaberUnImplError; + } + } +private: + virtual void pick_best_static() override { + if (true) // some condition? + this->_best_impl = this->_impl[0]; + } + virtual void pick_best_specify(ImplEnum implenum) override { + this->_best_impl = this->_impl[0]; + } +}; +} // namespace saber +} // namespace anakin +``` + +### 为operator增加新的impl声明 + +涉及的文件:```saber/funcs/impl/impl_mul.h```。不同的设备都特化同一个声明，特化版本放在对应的文件夹下，这里的声明就是给出所有设备的统一声明。下面给出一个参考。 +``` +#include "saber/funcs/impl/impl_macro.h" +namespace anakin{ +namespace saber{ +DEFINE_OP_CLASS(Mul, MulParam); // 第一个参数是op的名字，第二个是对应param的名字 +} +} +``` + +### 完成新的operator特定后端实现 + +涉及的文件:```saber/funcs/impl/xxx/vender_mul.h```或```saber/funcs/impl/xxx/saber_mul.h``` +这里```xxx```指代特定的一种设备。```vender```是指的使用第三方库实现的op，```saber```指的源码实现的op。这里以cuda的vender实现为例，简单介绍一下特化出的函数的几个基本接口。 + +``` +// include 对应的声明 +#include "saber/funcs/impl/impl_mul.h" + +namespace anakin{ +namespace saber{ +template +class VenderMul : + public ImplBase< + Tensor, + Tensor, + Tensor, + MulParam > > +{ +public: + typedef Tensor DataTensor_in; + typedef Tensor DataTensor_out; + typedef Tensor OpTensor; + typedef typename DataTensor_in::Dtype InDataType; + typedef typename DataTensor_out::Dtype OutDataType; + typedef typename OpTensor::Dtype OpDataType; + VenderMul(){} + ~VenderMul() {} + + virtual SaberStatus init(const std::vector& inputs, + std::vector& outputs, + MulParam& param, Context& ctx) { + this->_ctx = ctx; + create(inputs, outputs, param, ctx); + } + + virtual SaberStatus create(const std::vector& inputs, + std::vector& outputs, + MulParam& param, Context& ctx) { + // set内部参数 + } + + virtual SaberStatus dispatch(const std::vector& inputs, + std::vector& outputs, + MulParam& param) { + // dispatch kernel. + } + +private: +}; +} +} +``` +```init```和```create```的区别：```init```接口是第一次初始化op的时候进入的接口，此函数只在第一次初始化op时调用，这个接口一般放一些只需要执行一次的代码，如malloc或者create之类的函数。```create```函数除了第一次init执行外，在输入发生变化或者param发生变化时会再次触发，create一般放置set函数，设置内部变量，当input发生变化时这里执行一些同input或weights直接相关的代码。但create因为触发位置在网络内，如果```create```函数执行了一些严重耗时的操作，这里会拖慢整个op的执行时间，需要慎重选择操作放置的位置。 +### 添加framework的特化 + +涉及的文件:```framework/operators/mul.h```和```framework/operators/mul.cpp```。 +这里简单介绍下如果添加或修改framework内的operator + +``` +#include "framework/core/base.h" +#include "framework/core/data_types.h" +#include "framework/core/operator/operator.h" +#include "utils/logger/logger.h" +#include "saber/funcs/mul.h" // 需要包对应的saber头文件 +namespace anakin { +namespace ops { +template +class MulHelper; + +template +class Mul : public Operator { +public: + Mul() {} + /// forward impl + virtual void operator() (OpContext &ctx, + const std::vector >& ins, + std::vector >& outs) { + LOG(ERROR) << "Not Impl Yet Operator power::type>().type_info()<<">"; + } + friend class MulHelper; +}; +template +class MulHelper : public OperatorHelper { +public: + MulHelper() = default; + ~MulHelper(); + Status InitParam() override; + + Status Init(OpContext &ctx, + const std::vector >& ins, + std::vector >& outs) override; + Status InferShape(const std::vector >& ins, + std::vector >& outs) override; + +public: + saber::MulParam> _param_mul; + saber::Mul _funcs_mul; +}; +} +} /* namespace anakin */ +``` +对应的```.cpp```文件如下： +``` +#include "framework/operators/mul.h" + +namespace anakin { +namespace ops { + +#ifdef USE_CUDA +template<> +void Mul::operator()( + OpContext& ctx, + const std::vector >& ins, + std::vector >& outs) { + auto* impl = + static_cast*>(this->_helper); + auto& param = + static_cast*>(this->_helper)->_param_mul; + impl->_funcs_mul(ins, outs, param, ctx); +} +#endif + +template +Status MulHelper::InitParam() { + auto alpha = GET_PARAMETER(float, alpha); + MulParam> param_mul(alpha); + _param_mul = param_mul; + return Status::OK(); +} + +template +Status MulHelper::Init(OpContext& ctx, + const std::vector >& ins, + std::vector >& outs) { + + SABER_CHECK(_funcs_mul.init(ins, outs, _param_mul, SPECIFY, VENDER_IMPL, ctx)); + return Status::OK(); +} + +template +Status MulHelper::InferShape(const + std::vector >& ins, + std::vector >& outs) { + SABER_CHECK(_funcs_mul.compute_output_shape(ins, outs, _param_mul)); + return Status::OK(); +} + +#ifdef USE_CUDA +template class MulHelper; +#endif +#ifdef USE_ARM_PLACE +template class MulHelper; +#endif +// register helper +#ifdef USE_CUDA +ANAKIN_REGISTER_OP_HELPER(Mul, MulHelper, NV, AK_FLOAT, Precision::FP32); +#endif +#ifdef USE_ARM_PLACE +ANAKIN_REGISTER_OP_HELPER(Mul, MulHelper, ARM, AK_FLOAT, Precision::FP32); +#endif +//! register op +ANAKIN_REGISTER_OP(Mul) +.Doc("Mul operator") +#ifdef USE_CUDA +.__alias__("mul") +#endif +#ifdef USE_ARM_PLACE +.__alias__("mul") +#endif +.num_in(1) +.num_out(1) +.Args("alpha", " alpha of Mul "); //注册 + +} /* namespace ops */ + +} /* namespace anakin */ +``` + +## 实现单元测试 +涉及的文件:```test/saber/xxx/test_saber_funcs_mul_xxx.cpp``` +在对应的test下需要添加新的单元测试 + +``` +TEST(TestSaberFuncNV, test_depthwise_conv) { + + // init tensors and some param. + + // start Reshape & doInfer + Context ctx1(0, 1, 1); + + // create param + MulParam > param(alpha); + + std::vector*> input; + std::vector*> output; + + // create saber op + Mul mul; + + // compute output shape + mul.compute_output_shape(input, output, param); + + // re_alloc output tensors memory based on output shape + output[0]->re_alloc(output[0]->shape()); + + // init saber op(calling init and create) + mul.init(input, output, param, SPECIFY, VENDER_IMPL, ctx1); + + // call operator() + mul(input, output, param, ctx1); + + // cuda specified, record events + cudaStream_t cuda_stream = ctx1.get_compute_stream(); + output[0]->record_event(cuda_stream); + output_dev.sync(); + + // param changed + param.alpha = 2.0; + // auto calling saber op(create and dispatch) + mul(input, output, param, ctx1); + + cudaDeviceSynchronize(); + CUDA_CHECK(cudaPeekAtLastError()); +} + +int main(int argc, const char** argv){ + anakin::saber::Env::env_init(); + + // initial logger + //logger::init(argv[0]); + InitTest(); + RUN_ALL_TESTS(argv[0]); + return 0; +} + +``` +## 调试及注意事项 + +一个op需要有对外的op接口和内部实现，由于存在saber/funcs/impl的非特化版本声明，当有op在某种设备下没有对应实现时，也能够编译，但此时是没有任何实现的空实现， diff --git a/doc/fluid/new_docs/advanced_usage/deploy/how_to_support_new_device_in_anakin.md b/doc/fluid/new_docs/advanced_usage/deploy/how_to_support_new_device_in_anakin.md new file mode 100644 index 0000000000000000000000000000000000000000..a1f75f5e95cfb90f26d3782ba30a6d1887a70424 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/how_to_support_new_device_in_anakin.md @@ -0,0 +1,459 @@ +# 如何支持一个新的设备 + +## 概览 + +添加一个新的设备需要以下3个步骤： + +* [在`CMakeList`中添加设备的支持](#0001) +* [在`saber`中添加设备的实现](#0002) +* [在`framework`中添加设备的具体化或实例化](#0003) + +假设新设备的名称为`TNEW`, 以下将以这个设备名称进行演示。 + +## 在`CMakeList`中添加设备的支持 ## + +* 修改根目录`CMakeList.txt` +```cmake +#select the plantform to build +anakin_option(USE_GPU_PLACE "Select the build mode for GPU place." NO) +anakin_option(USE_X86_PLACE "Select the build mode for X86 place." NO) +anakin_option(USE_ARM_PLACE "Select the build mode for ARM place." NO) +anakin_option(USE_TNEW_PLACE "Select the build mode for ARM place." YES) +``` + +* 修改`saber/CMakeList.txt` + +根据新增设备的目录完善`saber`目录下的`CMakeList.txt`。 +```cmake +if(USE_TNEW_PLACE) + anakin_fetch_files_with_suffix(${ANAKIN_SABER}/core/impl/tnew "cpp" ANAKIN_SABER_BASE_SRC) + anakin_fetch_files_with_suffix(${ANAKIN_SABER}/funcs/impl/tnew "cpp" ANAKIN_SABER_BASE_SRC) +endif() +``` + +* 修改`test/CMakeList.txt` + +新增设备的单测文件放在`test/saber/tnew`目录下，修改`test`目录下的`CMakeList.txt`。 +```cmake +if(USE_TNEW_PLACE) + anakin_fetch_files_with_suffix(${ANAKIN_UNIT_TEST}/saber/tnew "cpp" ANAKIN_TEST_CASE_SRC) +endif() +``` + +* 修改`cmake/anakin_config.h.in` +```c++ +// plantform to use +#cmakedefine USE_GPU_PLACE + +#cmakedefine USE_X86_PLACE + +#cmakedefine USE_ARM_PLACE + +#cmakedefine USE_TNEW_PLACE +``` + +* 其他依赖和编译选项 +修改`cmake`目录下的`compiler_options.cmake`和`find_modules.cmake` + + +## 在`saber`中添加设备的实现 ## +`saber`是`Anakin`的基础计算库，对外提供设备无关的统一的API，设备相关的实现都会封装到`TargetWrapper`中。 + +### 在`saber/saber_types.h`中添加设备 + +```c++ +enum TargetTypeEnum { + eINVALID = -1, + eNV = 1, + eAMD = 2, + eARM = 3, + eX86 = 4, + eNVHX86 = 5, + eTNEW = 6 +}; + +typedef TargetType NV; +typedef TargetType ARM; +typedef TargetType AMD; +typedef TargetType X86; +typedef TargetType TNEW; + +``` + +### 在`saber/core`中添加设备的实现 + +1. 在`target_traits.h`中添加新设备 + +* 增加设备类型 +```c++ +struct __cuda_device{}; +struct __arm_device{}; +struct __amd_device{}; +struct __x86_device{}; +struct __tnew_device{}; +``` + +* `TargetTypeTraits`模板具体化 +```c++ +template <> +struct TargetTypeTraits { + typedef __xxx_target target_category;//根据实际设备是host端还是device端进行选择 + typedef __tnew_device target_type; +}; +``` + +2. 在`data_traits.h`中特化`DataTrait`模板类 + +如果设备需要特殊的数据类型，则特化出设备的`DataTrait`类的实现，例如opencl数据类型的实现如下： +```c++ +#ifdef USE_OPENCL +struct ClMem{ + ClMem(){ + dmem = nullptr; + offset = 0; + } + + ClMem(cl_mem* mem_in, int offset_in = 0) { + dmem = mem_in; + offset = offset_in; + } + + ClMem(ClMem& right) { + dmem = right.dmem; + offset = right.offset; + } + + ClMem& operator=(ClMem& right) { + this->dmem = right.dmem; + this->offset = right.offset; + return *this; + } + + ClMem& operator+(int offset_in) { + this->offset += offset_in; + return *this; + } + + int offset{0}; + cl_mem* dmem; +}; + +template <> +struct DataTrait { + typedef ClMem Dtype; + typedef float dtype; +}; + +template <> +struct DataTrait { + typedef ClMem Dtype; + typedef double dtype; +}; + +template <> +struct DataTrait { + typedef ClMem Dtype; + typedef char dtype; +}; +#endif //use_opencl +``` + +3. 在`target_wrapper.h`中特化`TargetWrapper`模板类 + +特化`TargetWrapper`模板类，在`target_wrapper.h`中声明函数，具体如下： +```c++ +template <> +struct TargetWrapper { //根据TNEW的具体类型修改__xxx_target，__host_target或者__device_target + + typedef xxx_event event_t; //根据设备实现xxx_event + typedef xxx_stream stream_t; //根据设备实现xxx_stream + + static void get_device_count(int& count); + + static void set_device(int id); + + //We should add strategy to avoid malloc directly + static void mem_alloc(void** ptr, size_t n); + + static void mem_free(void* ptr); + + static void mem_set(void* ptr, int value, size_t n); + + static void create_event(event_t& event, bool flag = false); + + static void create_stream(stream_t& stream); + + static void create_stream_with_flag(stream_t& stream, unsigned int flag); + + static void create_stream_with_priority(stream_t& stream, unsigned int flag, int priority); + + static void destroy_stream(stream_t& stream); + + static void destroy_event(event_t& event); + + static void record_event(event_t& event, stream_t stream); + + static void query_event(event_t& event); + + static void sync_event(event_t& event); + + static void sync_stream(event_t& event, stream_t& stream); + + static void sync_memcpy(void* dst, int dst_id, const void* src, int src_id, \ + size_t count, __DtoD); + + static void async_memcpy(void* dst, int dst_id, const void* src, int src_id, \ + size_t count, stream_t& stream, __DtoD); + + static void sync_memcpy(void* dst, int dst_id, const void* src, int src_id, \ + size_t count, __HtoD); + + static void async_memcpy(void* dst, int dst_id, const void* src, int src_id, \ + size_t count, stream_t& stream, __HtoD); + + static void sync_memcpy(void* dst, int dst_id, const void* src, int src_id, \ + size_t count, __DtoH); + + static void async_memcpy(void* dst, int dst_id, const void* src, int src_id, \ + size_t count, stream_t& stream, __DtoH); + + static void sync_memcpy_p2p(void* dst, int dst_dev, const void* src, \ + int src_dev, size_t count); + + static void async_memcpy_p2p(void* dst, int dst_dev, const void* src, \ + int src_dev, size_t count, stream_t& stream); + + static int get_device_id(); +}; + +``` + +4. 在`impl/`目录下添加设备目录和实现 + +在`saber/core/impl`目录下添加设备目录`tnew`。 +* 实现`TargetWrapper`结构体中各函数的定义。 +如果`TargetWrapper`的实现与默认的模板类一致，则不用特化出该类。 + +```c++ +typedef TargetWrapper TNEW_API; +void TNEW_API::get_device_count(int &count) { + // add implementation +} + +void TNEW_API::set_device(int id){ + // add implementation +} + +void TNEW_API::mem_alloc(void** ptr, size_t n){ + // add implementation +} + +void TNEW_API::mem_free(void* ptr){ + if(ptr != nullptr){ + // add implementation + } +} +... + +``` + +* 特化实现`device.h`中的`Device` + +```c++ +template <> +void Device::create_stream() { + // add implementation +} + +template <> +void Device::get_info() { + + // add implementation +} + +``` + +### 在`saber/funcs`中实现设备相关的op + +参考[如何增加新的Operator](addCustomOp.md) + + +## 在`framework`中添加设备的具体化或实例化 ## + +### `framework/core` + +* `net.cpp`中添加实例化 + +```c++ +#ifdef USE_TNEW_PLACE +template class Net; +template class Net; +#endif +``` + +* `operator_func.cpp`中添加实例化 + +```c++ +#ifdef USE_TNEW_PLACE +template class OperatorFunc; +#endif +``` + +* `worker.cpp`中添加实例化 + +```c++ +#ifdef USE_TNEW_PLACE +template class Worker; +template class Worker; +#endif +``` + +* `operator_attr.cpp`中添加实例化 + +```c++ +template +OpAttrWarpper& OpAttrWarpper::__alias__(const std::string& op_name); +template +OpAttrWarpper& OpAttrWarpper::__alias__(const std::string& op_name); +template +OpAttrWarpper& OpAttrWarpper::__alias__(const std::string& op_name); +``` + +* `parameter.h`中添加设备的实现 + +```c++ +#ifdef USE_TNEW_PLACE +template +class PBlock { +public: + typedef Tensor4d::type> type; + + PBlock() { + _inner_tensor = std::make_shared(); + } + ... +} +#endif //TNEW +``` + +* `type_traits_extend.h`中添加设备的实现 + +```c++ +template<> +struct target_host { + typedef saber::X86 type; //根据TNEW选择正确的host type +}; +``` + +### `framework/graph` + +* `graph.cpp`中添加实例化 + +```c++ + #ifdef USE_TNEW_PLACE + template class Graph; + template class Graph; + template class Graph; + #endif +``` + +### `framework/model_parser` + +* `parser.cpp`中添加实例化 + +```c++ + #ifdef USE_TNEW_PLACE + template + Status load(graph::Graph* graph, + const char* model_path); + template + Status load(graph::Graph* graph, + const char* model_path); + template + Status load(graph::Graph* graph, + const char* model_path); + + template + Status save(graph::Graph* graph, + std::string& model_path); + template + Status save(graph::Graph* graph, + std::string& model_path); + template + Status save(graph::Graph* graph, + std::string& model_path); + + template + Status load(graph::Graph* graph, + std::string& model_path); + template + Status load(graph::Graph* graph, + std::string& model_path); + template + Status load(graph::Graph* graph, + std::string& model_path); + + template + Status save(graph::Graph* graph, + const char* model_path); + template + Status save(graph::Graph* graph, + const char* model_path); + template + Status save(graph::Graph* graph, + const char* model_path); + #endif +``` + +* `model_io.cpp`中添加实例化 + +```c++ +#ifdef USE_TNEW_PLACE +template class NodeIO; +template class NodeIO; +template class NodeIO; +#endif +``` + +### `framework/operators` + +为`framework/operators`目录下所有op添加实例化或具体化 +以`activation.cpp`为例，实例化如下： + +```c++ +#ifdef USE_TNEW_PLACE +INSTANCE_ACTIVATION(TNEW, AK_FLOAT, Precision::FP32); +INSTANCE_ACTIVATION(TNEW, AK_FLOAT, Precision::FP16); +INSTANCE_ACTIVATION(TNEW, AK_FLOAT, Precision::INT8); +template class ActivationHelper; +ANAKIN_REGISTER_OP_HELPER(Activation, ActivationHelper, TNEW, AK_FLOAT, Precision::FP32); +#endif +``` + +如果TNEW设备函数的实现与现有模板实现不一致，可以特化实现如下（以init()为例）： +```c++ +#ifdef USE_TNEW_PLACE +INSTANCE_ACTIVATION(TNEW, AK_FLOAT, Precision::FP32); +INSTANCE_ACTIVATION(TNEW, AK_FLOAT, Precision::FP16); +INSTANCE_ACTIVATION(TNEW, AK_FLOAT, Precision::INT8); +template <> +Status ActivationHelper::Init(OpContext &ctx,\ + const std::vector >& ins, \ + std::vector >& outs) { + SABER_CHECK(_funcs_activation.init(ins, outs, _param_activation, SPECIFY, SABER_IMPL, ctx)); //在这里选择实现方式 + return Status::OK(); +} +ANAKIN_REGISTER_OP_HELPER(Activation, ActivationHelper, TNEW, AK_FLOAT, Precision::FP32); +#endif +``` + +在`ANAKIN_REGISTER_OP(Activation)`中添加TNEW的注册 + +```c++ +#ifdef USE_TNEW_PLACE +.__alias__("activation") +#endif +``` + +## 注意事项 +不要修改`Tensor`/`Buffer`/`Env`/`Context`这些类函数的接口和实现 diff --git a/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst b/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst new file mode 100644 index 0000000000000000000000000000000000000000..b782242a6632a5d42a512cf3b830d6e047c064ab --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst @@ -0,0 +1,26 @@ +服务器端部署 - Anakin +##################### + + +使用文档 +~~~~~~~ + +.. toctree:: + :maxdepth: 1 + + install_anakin.md + convert_paddle_to_anakin.md + run_anakin_on_arm.md + anakin_tutorial.md + anakin_example.md + anakin_gpu_benchmark.md + anakin_arm_benchmark.md + +开发文档 +~~~~~~~ + +.. toctree:: + :maxdepth: 1 + + how_to_add_anakin_op.md + how_to_support_new_device_in_anakin.md diff --git a/doc/fluid/new_docs/advanced_usage/deploy/index_mobile.rst b/doc/fluid/new_docs/advanced_usage/deploy/index_mobile.rst new file mode 100644 index 0000000000000000000000000000000000000000..47df6392c123d520c701089db6ee1ae72e4f8ea5 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/index_mobile.rst @@ -0,0 +1,9 @@ +移动端部署 +########## + +.. toctree:: + :maxdepth: 2 + + mobile_build.md + mobile_dev.md + diff --git a/doc/fluid/new_docs/advanced_usage/deploy/index_native.rst b/doc/fluid/new_docs/advanced_usage/deploy/index_native.rst new file mode 100644 index 0000000000000000000000000000000000000000..a5209e8560b31e9f0f776fba9a2b8c5bc150165c --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/index_native.rst @@ -0,0 +1,8 @@ +服务器端部署 - 原生引擎 +####################### + +.. toctree:: + :maxdepth: 2 + + build_and_install_lib_cn.rst + native_infer.rst diff --git a/doc/fluid/new_docs/advanced_usage/deploy/install_anakin.md b/doc/fluid/new_docs/advanced_usage/deploy/install_anakin.md new file mode 100644 index 0000000000000000000000000000000000000000..bb7c1950308622e3de292268a718e6ec688e6ae6 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/install_anakin.md @@ -0,0 +1,69 @@ +## 从源码编译安装Anakin ## + +我们已经在CentOS 7.3上成功的安装和测试了Anakin，对于其他操作系统，我们将很快支持。 + +### 安装概览 ### + +* [在CentOS上安装 Anakin]() +* [在Ubuntu上安装 Anakin]() +* [在ARM上安装 Anakin](run_on_arm_ch.md) +* [验证安装]() + + +### 在CentOS上安装 Anakin ### +#### 1. 系统要求 #### + +* make 3.82+ +* cmake 2.8.12+ +* gcc 4.8.2+ +* g++ 4.8.2+ +* 其他需要补充的。。。 + +#### 2. 编译CPU版Anakin #### + +暂时不支持 + +#### 3. 编译支持NVIDIA GPU的Anakin #### + +- 3.1. 安装依赖 + - 3.1.1 protobuf + >$ git clone https://github.com/google/protobuf + >$ cd protobuf + >$ git submodule update --init --recursive + >$ ./autogen.sh + >$ ./configure --prefix=/path/to/your/insall_dir + >$ make + >$ make check + >$ make install + >$ sudo ldconfig + + + 如安装protobuf遇到任何问题，请访问[这里](https://github.com/google/protobuf/blob/master/src/README.md) + +- 3.2 CUDA Toolkit + - [CUDA 8.0](https://developer.nvidia.com/cuda-zone) or higher. 具体信息参见[NVIDIA's documentation](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/). + - [cuDNN v7](https://developer.nvidia.com/cudnn). 具体信息参见[NVIDIA's documentation](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/). +- 3.3 编译Anakin + >$ git clone https:/xxxxx + >$ cd anakin + >$ mkdir build + >$ camke .. + >$ make + + +#### 4. 编译支持AMD GPU的Anakin #### + +暂时还不支持 + + +### 在Ubuntu上安装 Anakin ### + +暂时还不支持 + + +### 在ARM上安装 Anakin ### + +暂时还不支持 + +### 验证安装 ### +we are coming soon... diff --git a/doc/fluid/new_docs/advanced_usage/deploy/mobile_build.md b/doc/fluid/new_docs/advanced_usage/deploy/mobile_build.md new file mode 100644 index 0000000000000000000000000000000000000000..e51593164987d548e256ddebbc5fa8d960fb5255 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/mobile_build.md @@ -0,0 +1,59 @@ +# 环境搭建 +## 使用 docker +### 1. 安装 docker +安装 docker 的方式，参考官方文档 [https://docs.docker.com/install/](https://docs.docker.com/install/) +### 2. 使用 docker 搭建构建环境 +首先进入 paddle-mobile 的目录下，执行 `docker build` +以 Linux/Mac 为例 (windows 建议在 'Docker Quickstart Terminal' 中执行) +``` +$ docker build -t paddle-mobile:dev - < Dockerfile +``` +使用 `docker images` 可以看到我们新建的 image +``` +$ docker images +REPOSITORY TAG IMAGE ID CREATED SIZE +paddle-mobile dev 33b146787711 45 hours ago 372MB +``` +### 3. 使用 docker 构建 +进入 paddle-mobile 目录，执行 docker run +``` +$ docker run -it --mount type=bind,source=$PWD,target=/paddle-mobile paddle-mobile:dev +root@5affd29d4fc5:/ # cd /paddle-mobile +# 生成构建 android 产出的 Makefile +root@5affd29d4fc5:/ # rm CMakeCache.txt +root@5affd29d4fc5:/ # cmake -DCMAKE_TOOLCHAIN_FILE=tools/toolchains/arm-android-neon.cmake +# 生成构建 linux 产出的 Makefile +root@5affd29d4fc5:/ # rm CMakeCache.txt +root@5affd29d4fc5:/ # cmake -DCMAKE_TOOLCHAIN_FILE=tools/toolchains/arm-linux-gnueabi.cmake +``` +### 4. 设置编译选项 +可以通过 ccmake 设置编译选项 +``` +root@5affd29d4fc5:/ # ccmake . + Page 1 of 1 + CMAKE_ASM_FLAGS + CMAKE_ASM_FLAGS_DEBUG + CMAKE_ASM_FLAGS_RELEASE + CMAKE_BUILD_TYPE + CMAKE_INSTALL_PREFIX /usr/local + CMAKE_TOOLCHAIN_FILE /paddle-mobile/tools/toolchains/arm-android-neon.cmake + CPU ON + DEBUGING ON + FPGA OFF + LOG_PROFILE ON + MALI_GPU OFF + NET googlenet + USE_EXCEPTION ON + USE_OPENMP OFF +``` +修改选项后，按 `c`, `g` 更新 Makefile +### 5. 构建 +使用 make 命令进行构建 +``` +root@5affd29d4fc5:/ # make +``` +### 6. 查看构建产出 +构架产出可以在 host 机器上查看，在 paddle-mobile 的目录下，build 以及 test/build 下，可以使用 adb 指令或者 scp 传输到 device 上执行 + +## 不使用 docker +不使用 docker 的方法，可以直接用 cmake 生成 makefile 后构建。使用 ndk 构建 android 应用需要正确设置 NDK_ROOT。构建 linux 应用需要安装 arm-linux-gnueabi-gcc 或者类似的交叉编译工具，可能需要设置 CC，CXX 环境变量，或者在 tools/toolchains/ 中修改 arm-linux-gnueabi.cmake，或者增加自己需要的 toolchain file。 diff --git a/doc/fluid/new_docs/advanced_usage/deploy/mobile_dev.md b/doc/fluid/new_docs/advanced_usage/deploy/mobile_dev.md new file mode 100644 index 0000000000000000000000000000000000000000..474380f9dbfd2fb8a06630cb1ca3ca5cd14ca9d9 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/mobile_dev.md @@ -0,0 +1,72 @@ +# iOS开发文档 + +## 编译 + +### 一. 使用 build.sh 编译 + +```sh +sh build.sh ios + +# 如果只想编译某个特定模型的 op, 则需执行以下命令 +sh build.sh ios googlenet + +# 在这个文件夹下, 你可以拿到生成的 .a 库 +cd ../build/release/ios/build + +``` + +### 二. 使用 xcode 编译 + +我们提供了 ios 开发更为熟悉的 xcode 编译环境: +在 ios/ 目录下打开 PaddleMobile.xcworkspace 即可编译 PaddleMobile 或者 运行 Demo + +### 三. 集成 + +#### 如使用 c++ 接口 +将 + +``` +libpaddle-mobile.a +io.h +program.h +types.h +lod_tensor.h +tensor.h +``` +拖入工程, io.h 为接口文件, 可在 [github](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/src/io/io.h)上查看接口注释 + +#### 如使用 oc 接口 +将在xcode 编译生成的 +``` +libPaddleMobile.a +PaddleMobile.h +``` +拖入工程, 接口如下: + +``` +/* + 创建单例对象 +*/ ++ (instancetype)sharedInstance; + +/* + load 模型, 开辟内存 +*/ +- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath; + +/* + 进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict +*/ +- (NSArray *)predict:(CGImageRef)image means:(NSArray *)means scale:(float)scale; + +/* + 进行预测 +*/ +- (NSArray *)predict:(CGImageRef)image; + +/* + 清理内存 +*/ +- (void)clear; + +``` diff --git a/doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst b/doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst new file mode 100644 index 0000000000000000000000000000000000000000..3571f81326a9f9ae31a8327c3e288e601f248e4b --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst @@ -0,0 +1,108 @@ +Paddle 预测 API +=============== + +为了更简单方便的预测部署，Fluid 提供了一套高层 API +用来隐藏底层不同的优化实现。 + +`预测库相关代码 `__ +包括 + +- 头文件 ``paddle_inference_api.h`` 定义了所有的接口 +- 库文件\ ``libpaddle_fluid.so`` 或 ``libpaddle_fluid.a`` +- 库文件 ``libpaddle_inference_api.so`` 或 + ``libpaddle_inference_api.a`` + +编译和依赖可以参考 :ref:`install_or_build_cpp_inference_lib` 。 + +下面是一些 API 概念的介绍 + +PaddleTensor +------------ + +PaddleTensor 定义了预测最基本的输入输出的数据格式，其定义是 + +.. code:: cpp + + struct PaddleTensor { + std::string name; // variable name. + std::vector shape; + PaddleBuf data; // blob of data. + PaddleDType dtype; + }; + +- ``name`` 用于指定输入数据对应的 模型中variable 的名字 + （暂时没有用，但会在后续支持任意 target 时启用） +- ``shape`` 表示一个 Tensor 的 shape +- ``data`` 数据以连续内存的方式存储在\ ``PaddleBuf`` + 中，\ ``PaddleBuf`` + 可以接收外面的数据或者独立\ ``malloc``\ 内存，详细可以参考头文件中相关定义。 +- ``dtype`` 表示 Tensor 的数据类型 + +engine +------ + +高层 API 底层有多种优化实现，我们称之为 engine，目前有三种 engine + +- 原生 engine，由 paddle 原生的 forward operator + 组成，可以天然支持所有paddle 训练出的模型， +- Anakin engine，封装了 + `Anakin `__ + ，在某些模型上性能不错，但只能接受自带模型格式，无法支持所有 paddle + 模型， +- TensorRT mixed engine，用子图的方式支持了 + `TensorRT `__ ，支持所有paddle + 模型，并自动切割部分计算子图到 TensorRT 上加速（WIP） + +其实现为 + +.. code:: cpp + + enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. + kAutoMixedTensorRT // Automatically mixing TensorRT with the Fluid ops. + }; + +预测部署过程 +------------ + +总体上分为以下步骤 + +1. 用合适的配置创建 ``PaddlePredictor`` +2. 创建输入用的 ``PaddleTensor``\ ，传入到 ``PaddlePredictor`` 中 +3. 获取输出的 ``PaddleTensor`` ，将结果取出 + +下面完整演示一个简单的模型，部分细节代码隐去 + +.. code:: cpp + + #include "paddle_inference_api.h" + + // 创建一个 config，并修改相关设置 + paddle::NativeConfig config; + config.model_dir = "xxx"; + config.use_gpu = false; + // 创建一个原生的 PaddlePredictor + auto predictor = + paddle::CreatePaddlePredictor(config); + // 创建输入 tensor + int64_t data[4] = {1, 2, 3, 4}; + paddle::PaddleTensor tensor{.name = "", + .shape = std::vector({4, 1}), + .data = PaddleBuf(data, sizeof(data)), + .dtype = PaddleDType::INT64}; + // 创建输出 tensor，输出 tensor 的内存可以复用 + std::vector outputs; + // 执行预测 + CHECK(predictor->Run(slots, &outputs)); + // 获取 outputs ... + +编译时，联编 ``libpaddle_fluid.a/.so`` 和 +``libpaddle_inference_api.a/.so`` 便可。 + +详细代码参考 +------------ + +- `inference + demos `__ +- `复杂单线程/多线程例子 `__ diff --git a/doc/fluid/new_docs/advanced_usage/deploy/run_anakin_on_arm.md b/doc/fluid/new_docs/advanced_usage/deploy/run_anakin_on_arm.md new file mode 100644 index 0000000000000000000000000000000000000000..ebeb38f534ebfc8cb5a41d103abe3bb1de7e379a --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/deploy/run_anakin_on_arm.md @@ -0,0 +1,151 @@ +## 源码编译 Anakin ## + +目前Anakin支持ARM Android平台，采用Android NDK交叉编译工具链，已在mac os和centos上编译和测试通过。 + +### 安装概览 ### + +* [系统需求](#0001) +* [安装第三方依赖](#0002) +* [Anakin源码编译](#0003) +* [验证安装](#0004) + + +### 1. 系统需求 ### + +* 宿主机: linux, mac +* cmake 3.8.2+ +* Android NDK r14, Linux 版本[从这里下载](https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip) + +### 2. 安装第三方依赖 ### + +- 2.1 protobuf3.4.0 + 源码从这里[下载](https://github.com/google/protobuf/releases/tag/v3.4.0) + - 2.1.1 为宿主机编译protobuf + ```bash + $ tar -xzf protobuf-3.4.0.tar.gz + $ cd protobuf-3.4.0 + $ ./autogen.sh + $ ./configure + $ make + $ make check + $ make install + ``` + 上述 $make install 执行后，可在 /usr/local/include/google 找到 libprotobuf 所需的头文件,将整个google文件夹拷贝至Anakin/third-party/arm-android/protobuf/下， + 如有问题，请点[这里](https://github.com/google/protobuf/blob/v3.4.0/src/README.md)。 + 然后将已经生成文件清除。 + ```bash + $ make distclean + ``` + - 2.1.1 交叉编译Android`armeabi-v7a`的protobuf，注意设置ANDROID_NDK的路径，以及ARCH_ABI、HOSTOSN的值， + ```bash + + $ export ANDROID_NDK=your_ndk_path + $ ARCH_ABI="arm-linux-androideabi-4.9" + $ HOSTOSN="darwin-x86_64" + $ export SYSROOT=$ANDROID_NDK/platforms/android-9/arch-arm + $ export PREBUILT=$ANDROID_NDK/toolchains/$ARCH_ABI + $ export LDFLAGS="--sysroot=$SYSROOT" + $ export LD="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/arm-linux-androideabi/bin/ld $LDFLAGS" + $ export LIBS="-llog $ANDROID_NDK/sources/cxx-stl/gnu-libstdc++/4.9/libs/armeabi-v7a/libgnustl_static.a" + $ export CPPFLAGS="" + $ export INCLUDES="-I$ANDROID_NDK/sources/cxx-stl/gnu-libstdc++/4.9/include/ -I$ANDROID_NDK/platforms/android-9/arch-arm/usr/include/ -I$ANDROID_NDK/sources/cxx-stl/gnu-libstdc++/4.9/libs/armeabi-v7a/include/" + $ export CXXFLAGS="-march=armv7-a -mfloat-abi=softfp -DGOOGLE_PROTOBUF_NO_RTTI --sysroot=$SYSROOT" + $ export CCFLAGS="$CXXFLAGS" + $ export CXX="$PREBUILT/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-g++ $CXXFLAGS" + $ export CC="$CXX" + $ export RANLIB="$ANDROID_NDK/toolchains/$ARCH_ABI/prebuilt/$HOSTOSN/bin/arm-linux-androideabi-ranlib" + $ ./autogen.sh + $ ./configure --host=arm-linux-androideabi --with-sysroot=$SYSROOT --enable-cross-compile --with-protoc=protoc --disable-shared CXX="$CXX" CC="$CC" LD="$LD" + $ make + ``` + + 编译生成 *.a 静态库，若希望编译*.so 动态链接库 ，请在./configure参数中改--disable-shared为--disable-static --enable-shared。 + 生成文件在src/.libs/下，将生成的文件拷贝至Anakin/third-party/arm-android/protobuf/lib下。 + 在[cmake](../../cmake/find_modules.cmake)中更新`ARM_RPOTO_ROOT`的路径。 + ```cmake + set(ARM_RPOTO_ROOT "${CMAKE_SOURCE_DIR}/third-party/arm-android/protobuf") + ``` + +- 2.2 opencv 2.4.3+(optional) + Anakin只在examples示例中使用opencv + Android系统的opencv从[这里下载](https://opencv.org/releases.html) + 解压后将 `3rdparty/libs/armeabi-v7a`中的库文件拷贝到`libs/armeabi-v7a` + 在[cmake](../../cmake/find_modules.cmake)中搜索`anakin_find_opencv`, + 并设置 `include_directories` 和 `LINK_DIRECTORIES`为自己安装的库的路径。 + ```cmake + include_directories(${CMAKE_SOURCE_DIR}/third-party/arm-android/opencv/sdk/native/jni/include/) + LINK_DIRECTORIES(${CMAKE_SOURCE_DIR}/third-party/arm-android/opencv/sdk/native/libs/armeabi-v7a/) + ``` +### 3. Anakin源码编译 ### + +#### 编译Android版本 + + 克隆[源码](https://github.com/PaddlePaddle/Anakin/tree/arm) +```bash + cd your_dir + git clone https://github.com/PaddlePaddle/Anakin.git + cd Anakin + git fetch origin arm + git checkout arm + ``` + 修改`android_build.sh` +- 修改NDK路径 + ```bash + #modify "your_ndk_path" to your NDK path + export ANDROID_NDK=your_ndk_path + ``` +- 修改ARM 处理器架构 + 对于32位ARM处理器, 将ANDROID_ABI 设置为 `armeabi-v7a with NEON`， + 对于64位ARM处理器, 可以将ANDROID_ABI 设置为 `armeabi-v7a with NEON`或者`arm64-v8a`。 + 目前我们只支持 `armeabi-v7a with NEON`；`arm64-v8a` 还在开发中。 + ```bash + -DANDROID_ABI="armeabi-v7a with NEON" + ``` +- 设置Android API + 根据Android系统的版本设置API level， 例如API Level 21 -> Android 5.0.1 + ```bash + -DANDROID_NATIVE_API_LEVEL=21 + ``` + +- 选择编译静态库或动态库 + 设置`BUILD_SHARED=NO`编译静态库 + 设置`BUILD_SHARED=YES`编译动态库 + ```bash + -DBUILD_SHARED=NO + ``` +- OpenMP多线程支持 + 设置`USE_OPENMP=YES`开启OpenMP多线程 + ```bash + -DUSE_OPENMP=YES + ``` + +- 编译单测文件 + 设置`BUILD_WITH_UNIT_TEST=YES`将会编译单测文件 + ```bash + -DBUILD_WITH_UNIT_TEST=YES + ``` + +- 编译示例文件 + 设置`BUILD_EXAMPLES=YES`将会编译示例文件 + ```bash + -DBUILD_EXAMPLES=YES + ``` + +- 开启opencv + 如果使用opencv，设置`USE_OPENCV=YES` + ```bash + -DUSE_OPENCV=YES + ``` + +- 开始编译 + 运行脚本 `android_build.sh` 将自动编译Anakin + ```bash + ./android_build.sh + ``` + +### 4. 验证安装 ### + 编译好的库会放在目录`${Anakin_root}/output`下； + 编译好的单测文件会放在`${Anakin_root}/output/unit_test`目录下； + 编译好的示例文件会放在`${Anakin_root}/output/examples`目录下。 + + 对于Android系统，打开设备的调试模式，通过ADB可以访问的目录是`data/local/tmp`，通过ADB push将测试文件、模型和数据发送到设备目录， 运行测试文件。 diff --git a/doc/fluid/new_docs/advanced_usage/development/contribute_to_paddle.md b/doc/fluid/new_docs/advanced_usage/development/contribute_to_paddle.md new file mode 120000 index 0000000000000000000000000000000000000000..1126df7a829ab6d98e58a44e8f9c6459feae9a8b --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/contribute_to_paddle.md @@ -0,0 +1 @@ +../../../dev/contribute_to_paddle_cn.md \ No newline at end of file diff --git a/doc/fluid/new_docs/advanced_usage/development/cpu_profiling_cn.md b/doc/fluid/new_docs/advanced_usage/development/cpu_profiling_cn.md new file mode 120000 index 0000000000000000000000000000000000000000..1381a3b05f6761c60742eb9365708d94ad8a2642 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/cpu_profiling_cn.md @@ -0,0 +1 @@ +../../../howto/optimization/cpu_profiling_cn.md \ No newline at end of file diff --git a/doc/fluid/new_docs/advanced_usage/development/gpu_profiling_cn.rst b/doc/fluid/new_docs/advanced_usage/development/gpu_profiling_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..f2396716bddd4810fa77c738d41f5482aa6d6055 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/gpu_profiling_cn.rst @@ -0,0 +1,242 @@ +============ +GPU性能调优 +============ + +.. contents:: + +此教程将向您分步介绍如何使用内置的定时工具、 **nvprof** 或 **nvvp** 来运行性能分析和调优。 + +- 什么是性能分析？ +- 为什么需要性能分析？ +- 如何进行性能分析？ +- 性能分析工具介绍 +- 详细教程 +- 性能分析小技巧 + +什么是性能分析？ +================ +在软件工程的范畴里，性能分析（Profiling）是一个动态程序分析的术语，它可以指测量一个程序的空间（内存）复杂度或时间复杂度， +也可以说是某些特定指令的使用情况，或者是函数调用的频率和耗时等。通常情况下，分析得到的信息用于协助进行程序的优化。 + +简单来说，性能分析工具是用于给应用程序的性能做定量分析的。如果想很好的理解程序的行为，那程序分析工具是必不可少的利器。简单的性能分析，可以告诉您某个操作到底花了多长时间？而更深入的分析，甚至能解释为什么某个操作花了很长时间？ + +为什么需要性能分析？ +============================ +训练好一个深层神经网络通常要耗费非常长的时间，所以性能也就逐步变成了深度学习领域最重要的指标。 +而优化性能的首要任务，是需要了解哪些步骤拖慢了整体。 +如果某一块根本就不怎么耗时，那也就不需要急着优化性能啦！ + +如何进行性能分析？ +======================== +为了达到性能最优，您可以采用下面五个步骤： + +- 对代码进行性能分析 +- 找到运行慢的部分 +- 找到运行慢的原因 +- 修改成更快的版本 +- 再次对代码进行性能分析 + +Usually, processor has two key performance limits include float point throughput and +memory throughput. For GPU, it also need more parallelism to fulfill its potential. +This is why they can be so fast. + +通常情况下，处理器有两个关键性能限制：一个是浮点计算量，另一个是内存操作量。 +GPU则还需要高并行性，才能发挥其全部能力。这正是它们速度快的原因。 + +性能分析工具介绍 +====================== +就通常的GPU性能分析来说，市面上已经有NVIDIA或第三方提供的众多工具。 + +**nvprof** 是Nvidia性能分析工具， **nvvp** 则是带GUI的Nvidia可视化性能分析工具。 +在这个教程中，我们主要会介绍nvprof和nvvp。 + +:code:`test_GpuProfiler` from :code:`paddle/legacy/math/tests` directory will be used to evaluate +above profilers. + +:code:`paddle/legacy/math/test` 目录中的 :code:`test_GpuProfiler` 就是用于展示上述分析工具的用法。 + +.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp + :language: c++ + :lines: 137-151 + :linenos: + +上述的代码片段包含了两种方法，您可以任意使用一个或两个来对感兴趣的代码段做性能分析。 + +1. :code:`REGISTER_TIMER_INFO` 是一个内置的定时器封装，可以用来计算CPU函数或cuda内核的时间消耗。 + +2. :code:`REGISTER_GPU_PROFILER` is a general purpose wrapper object of :code:`cudaProfilerStart` and :code:`cudaProfilerStop` to avoid +program crashes when CPU version of PaddlePaddle invokes them. + +3. :code:`REGISTER_GPU_PROFILER` 是一个封装对象，封装了 :code:`cudaProfilerStart` 和 :code:`cudaProfileStop` 两个操作；同时其内部实现可以避免纯CPU版本PaddlePaddle在执行本语句时发生崩溃。 + +您会在接下来的部分中获得更多的细节介绍。 + +详细教程 +============ + +内置定时器 +------------ + +如果想要启用PaddlePaddle的内置定时器，您首先需要在相关代码段中加入 :code:`REGISTER_TIMER_INFO`。 +接下来就可以使用 :code:`printStatus` 或者 :code:`printAllStatus` 函数来将信息输出到界面中。 +下面举个简单的例子： + +1. 加入 :code:`REGISTER_TIMER_INFO` 和 :code:`printAllStatus` 函数（如高亮部分）。 + + .. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp + :language: c++ + :lines: 137-151 + :emphasize-lines: 8-12,14 + :linenos: + +2. cmake配置中将 **WITH_TIMER** 打开，重新编译PaddlePaddle。 + + .. code-block:: bash + + cmake .. -DWITH_TIMER=ON + make + +3. 执行您的代码，并观察结果(如高亮部分）。 + + .. code-block:: bash + :emphasize-lines: 1,12-15 + + > ./paddle/legacy/math/tests/test_GpuProfiler + I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/legacy/math/tests/test_GpuProfiler + I1117 11:13:42.845065 2522362816 Util.cpp:130] Calling runInitFunctions + I1117 11:13:42.845208 2522362816 Util.cpp:143] Call runInitFunctions done. + [==========] Running 1 test from 1 test case. + [----------] Global test environment set-up. + [----------] 1 test from Profiler + [ RUN ] Profiler.BilinearFwdBwd + I1117 11:13:42.845310 2522362816 test_GpuProfiler.cpp:114] Enable GPU Profiler Stat: [testBilinearFwdBwd] "numSamples = 10, channels = 16, im + gSizeX = 64, imgSizeY = 64" + I1117 11:13:42.850154 2522362816 ThreadLocal.cpp:37] thread use undeterministic rand seed:20659751 + I1117 11:13:42.981501 2522362816 Stat.cpp:130] ======= StatSet: [GlobalStatInfo] status ====== + I1117 11:13:42.981539 2522362816 Stat.cpp:133] Stat=testBilinearFwdBwd total=136.141 avg=136.141 max=136.141 min=136.141 count=1 + I1117 11:13:42.981572 2522362816 Stat.cpp:141] ======= BarrierStatSet status ====== + I1117 11:13:42.981575 2522362816 Stat.cpp:154] -------------------------------------------------- + [ OK ] Profiler.BilinearFwdBwd (136 ms) + [----------] 1 test from Profiler (136 ms total) + + [----------] Global test environment tear-down + [==========] 1 test from 1 test case ran. (136 ms total) + [ PASSED ] 1 test. + +nvprof 工具 +---------------- + +要使用命令行分析工具 **nvprof**，您按如下步骤操作即可： + +1. 将 :code:`REGISTER_GPU_PROFILER` 函数加到代码中（参考强调部分）。 + + .. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp + :language: c++ + :lines: 137-151 + :emphasize-lines: 6-7 + :linenos: + +2. cmake中将 **WITH_PROFILER** 配置打开，重新编译PaddlePaddle。 + + .. code-block:: bash + + cmake .. -DWITH_PROFILER=ON + make + +3. 使用 **nvprof** 来分析执行文件。 + + .. code-block:: bash + + nvprof ./paddle/legacy/math/tests/test_GpuProfiler + +然后，您就能获得如下的分析结果： + +.. code-block:: bash + + ==78544== Profiling application: ./paddle/legacy/math/tests/test_GpuProfiler + ==78544== Profiling result: + Time(%) Time Calls Avg Min Max Name + 27.60% 9.6305ms 5 1.9261ms 3.4560us 6.4035ms [CUDA memcpy HtoD] + 26.07% 9.0957ms 1 9.0957ms 9.0957ms 9.0957ms KeBilinearInterpBw + 23.78% 8.2977ms 1 8.2977ms 8.2977ms 8.2977ms KeBilinearInterpFw + 22.55% 7.8661ms 2 3.9330ms 1.5798ms 6.2863ms [CUDA memcpy DtoH] + + ==78544== API calls: + Time(%) Time Calls Avg Min Max Name + 46.85% 682.28ms 8 85.285ms 12.639us 682.03ms cudaStreamCreateWithFlags + 39.83% 580.00ms 4 145.00ms 302ns 550.27ms cudaFree + 9.82% 143.03ms 9 15.892ms 8.7090us 142.78ms cudaStreamCreate + 1.23% 17.983ms 7 2.5690ms 23.210us 6.4563ms cudaMemcpy + 1.23% 17.849ms 2 8.9247ms 8.4726ms 9.3768ms cudaStreamSynchronize + 0.66% 9.5969ms 7 1.3710ms 288.43us 2.4279ms cudaHostAlloc + 0.13% 1.9530ms 11 177.54us 7.6810us 591.06us cudaMalloc + 0.07% 1.0424ms 8 130.30us 1.6970us 453.72us cudaGetDevice + 0.04% 527.90us 40 13.197us 525ns 253.99us cudaEventCreateWithFlags + 0.03% 435.73us 348 1.2520us 124ns 42.704us cuDeviceGetAttribute + 0.03% 419.36us 1 419.36us 419.36us 419.36us cudaGetDeviceCount + 0.02% 260.75us 2 130.38us 129.32us 131.43us cudaGetDeviceProperties + 0.02% 222.32us 2 111.16us 106.94us 115.39us cudaLaunch + 0.01% 214.06us 4 53.514us 28.586us 77.655us cuDeviceGetName + 0.01% 115.45us 4 28.861us 9.8250us 44.526us cuDeviceTotalMem + 0.01% 83.988us 4 20.997us 578ns 77.760us cudaSetDevice + 0.00% 38.918us 1 38.918us 38.918us 38.918us cudaEventCreate + 0.00% 34.573us 31 1.1150us 279ns 12.784us cudaDeviceGetAttribute + 0.00% 17.767us 1 17.767us 17.767us 17.767us cudaProfilerStart + 0.00% 15.228us 2 7.6140us 3.5460us 11.682us cudaConfigureCall + 0.00% 14.536us 2 7.2680us 1.1490us 13.387us cudaGetLastError + 0.00% 8.6080us 26 331ns 173ns 783ns cudaSetupArgument + 0.00% 5.5470us 6 924ns 215ns 2.6780us cuDeviceGet + 0.00% 5.4090us 6 901ns 328ns 3.3320us cuDeviceGetCount + 0.00% 4.1770us 3 1.3920us 1.0630us 1.8300us cuDriverGetVersion + 0.00% 3.4650us 3 1.1550us 1.0810us 1.2680us cuInit + 0.00% 830ns 1 830ns 830ns 830ns cudaRuntimeGetVersion + + +nvvp 工具 +-------------- + +如果想使用可视化的分析器 **nvvp**，您可以导入 :code:`nvprof -o ...` 的输出，或者从工具的界面里运行您的应用。 + +**备注: nvvp 也支持CPU的性能分析** (需在nvvp界面中选上才能开启） + +.. image:: nvvp1.png + :align: center + :scale: 33% + +从内核函数的角度， **nvvp** 可以精确说明一个长耗时操作的具体原因。 +同时，如下图所示， **nvvp** 的内核block使用情况、寄存器使用情况和共享内存使用情况能让我们对GPU的整体使用有更好的理解。 + + +.. image:: nvvp2.png + :align: center + :scale: 33% + +而从应用的角度， **nvvp** 可以帮您提供一些定位性能瓶颈的建议。 +例如，下图中就展示了一些关于内存数据迁徙和计算资源利用率的建议，为您做性能调优提供了方向。 + +.. image:: nvvp3.png + :align: center + :scale: 33% + +.. image:: nvvp4.png + :align: center + :scale: 33% + +性能分析小技巧 +================== + +- 开始阶段，从 **nvprof** 和 **nvvp** 的输出信息入手是个不错的选择。 +- 接下来可以考虑下时间线的分析。 +- 如果真想挖掘内核深处的某个秘密，您最好先确认：这一块的耗时比例真的太高，值得深入分析。 +- 可能的情况下，试着让输出的分析数据和理论值对应。 + + 1) 例如，如果我知道内核花了10ms来移动1GB数据，那我会期望分析工具统计到速度是100GB/s。 + 2) 若有不一致之处，很有可能实际应用就是没有按照您的预期情况运行。 +- 了解您的硬件：如果您的GPU理论可以达到6 TFLOPs（6万亿次浮点运算每秒），而当前已经有5.5 TFLOPs了，那估计这里的潜力就没啥好挖的了…… + +性能分析是性能优化的关键一步。有的时候简简单单的改变就能在性能上产生明显的优化效果！ +当然，具体情况因人而异。 + +参考资料 +=========== +Jeremy Appleyard, `GPU Profiling for Deep Learning `_, 2015 diff --git a/doc/fluid/new_docs/advanced_usage/development/host_memory_profiling_cn.md b/doc/fluid/new_docs/advanced_usage/development/host_memory_profiling_cn.md new file mode 120000 index 0000000000000000000000000000000000000000..904968ba4a8d6cc6489c91a0a751e0a33dcc873c --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/host_memory_profiling_cn.md @@ -0,0 +1 @@ +../../../howto/optimization/host_memory_profiling_cn.md \ No newline at end of file diff --git a/doc/fluid/new_docs/advanced_usage/development/new_op.md b/doc/fluid/new_docs/advanced_usage/development/new_op.md new file mode 120000 index 0000000000000000000000000000000000000000..dce0348585b8c484c1418a03a5fde5d78b0afcc9 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/new_op.md @@ -0,0 +1 @@ +../../../dev/new_op_cn.md \ No newline at end of file diff --git a/doc/fluid/new_docs/advanced_usage/development/nvvp1.png b/doc/fluid/new_docs/advanced_usage/development/nvvp1.png new file mode 100644 index 0000000000000000000000000000000000000000..1af23ac3c52929b2b0645d2f9fa4d4c6db1f6e77 Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/nvvp1.png differ diff --git a/doc/fluid/new_docs/advanced_usage/development/nvvp2.png b/doc/fluid/new_docs/advanced_usage/development/nvvp2.png new file mode 100644 index 0000000000000000000000000000000000000000..177c9db708da6863d1075f3e615f5962dbe18b29 Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/nvvp2.png differ diff --git a/doc/fluid/new_docs/advanced_usage/development/nvvp3.png b/doc/fluid/new_docs/advanced_usage/development/nvvp3.png new file mode 100644 index 0000000000000000000000000000000000000000..d8f393667d6569b6f1e61ffccac43fae5888b6db Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/nvvp3.png differ diff --git a/doc/fluid/new_docs/advanced_usage/development/nvvp4.png b/doc/fluid/new_docs/advanced_usage/development/nvvp4.png new file mode 100644 index 0000000000000000000000000000000000000000..51f2f3e183295de6cf8ddaf2b3b8a0862aa35f01 Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/nvvp4.png differ diff --git a/doc/fluid/new_docs/advanced_usage/development/pprof_1.png b/doc/fluid/new_docs/advanced_usage/development/pprof_1.png new file mode 100644 index 0000000000000000000000000000000000000000..8e9edbf377672d0ef40f2fc7bd39e746923550cb Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/pprof_1.png differ diff --git a/doc/fluid/new_docs/advanced_usage/development/pprof_2.png b/doc/fluid/new_docs/advanced_usage/development/pprof_2.png new file mode 100644 index 0000000000000000000000000000000000000000..172ba20399ba974d27f4c072425277b69b02520b Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/pprof_2.png differ diff --git a/doc/fluid/new_docs/advanced_usage/development/timeline.jpeg b/doc/fluid/new_docs/advanced_usage/development/timeline.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..38ec3f80c982857531f30a8bb0fa26ea5bf05385 Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/timeline.jpeg differ diff --git a/doc/fluid/new_docs/advanced_usage/development/timeline_cn.md b/doc/fluid/new_docs/advanced_usage/development/timeline_cn.md new file mode 120000 index 0000000000000000000000000000000000000000..a05540e82a7fa795dcd8e7306261ef9bef57426f --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/timeline_cn.md @@ -0,0 +1 @@ +../../../howto/optimization/timeline_cn.md \ No newline at end of file diff --git a/doc/fluid/new_docs/advanced_usage/development/tracing.jpeg b/doc/fluid/new_docs/advanced_usage/development/tracing.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3a49fc4f8a401a9463b0157e2f38c164ca02dcc5 Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/development/tracing.jpeg differ diff --git a/doc/fluid/new_docs/advanced_usage/development/write_docs.rst b/doc/fluid/new_docs/advanced_usage/development/write_docs.rst new file mode 120000 index 0000000000000000000000000000000000000000..dc536c8bdd4924758d4418bac8e4181ffbb1f780 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/development/write_docs.rst @@ -0,0 +1 @@ +../../../dev/write_docs_cn.rst \ No newline at end of file diff --git a/doc/fluid/new_docs/advanced_usage/index.rst b/doc/fluid/new_docs/advanced_usage/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..dea7c236619a0bdbf402f371571d947d1cdbba65 --- /dev/null +++ b/doc/fluid/new_docs/advanced_usage/index.rst @@ -0,0 +1,23 @@ +######## +进阶使用 +######## + + +.. todo:: + + Complete this guide + +.. toctree:: + :maxdepth: 2 + + deploy/index_native.rst + deploy/index_anakin.rst + deploy/index_mobile.rst + development/contribute_to_paddle.md + development/write_docs.rst + development/new_op.md + development/cpu_profiling_cn.md + development/gpu_profiling_cn.rst + development/host_memory_profiling_cn.md + development/timeline_cn.md + benchmark.rst diff --git a/doc/fluid/new_docs/advanced_usage/pics/anakin_fm_ch.png b/doc/fluid/new_docs/advanced_usage/pics/anakin_fm_ch.png new file mode 100644 index 0000000000000000000000000000000000000000..52d4992a22397119af949aa7c11a9ea6365c167c Binary files /dev/null and b/doc/fluid/new_docs/advanced_usage/pics/anakin_fm_ch.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/.gitignore b/doc/fluid/new_docs/beginners_guide/basics/image_classification/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..dc7c62b06287ad333dd41082e566b0553d3a5341 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/image_classification/.gitignore @@ -0,0 +1,8 @@ +*.pyc +train.log +output +data/cifar-10-batches-py/ +data/cifar-10-python.tar.gz +data/*.txt +data/*.list +data/mean.meta diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog.png new file mode 100644 index 0000000000000000000000000000000000000000..ca8f858a902ea723d886d2b88c2c0a1005301c50 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog_cat.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog_cat.png new file mode 100644 index 0000000000000000000000000000000000000000..38b21f21604b1bb84fc3f6aa96bd5fce45d15a55 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/dog_cat.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/fea_conv0.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/fea_conv0.png new file mode 100644 index 0000000000000000000000000000000000000000..647c822e52cd55d50e5f207978f5e6ada86cf34c Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/fea_conv0.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/flowers.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/flowers.png new file mode 100644 index 0000000000000000000000000000000000000000..04245cef60fe7126ae4c92ba8085273965078bee Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/flowers.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/googlenet.jpeg b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/googlenet.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..249dbf96df61c3352ea5bd80470f6c4a1e03ff10 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/googlenet.jpeg differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/ilsvrc.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/ilsvrc.png new file mode 100644 index 0000000000000000000000000000000000000000..4660ac122e9d533023a21154d35eee29e3b08d27 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/ilsvrc.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/inception.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/inception.png new file mode 100644 index 0000000000000000000000000000000000000000..9591a0c1e8c0165c40ca560be35a7b9a91cd5027 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/inception.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/lenet.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/lenet.png new file mode 100644 index 0000000000000000000000000000000000000000..77f785e03bacd38c4c64a817874a58ff3298d2f3 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/lenet.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/plot.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/plot.png new file mode 100644 index 0000000000000000000000000000000000000000..57e45cc0c27dd99b9918de2ff1228bc6b65f7424 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/plot.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet.png new file mode 100644 index 0000000000000000000000000000000000000000..0aeb4f254639fdbf18e916dc219ca61602596d85 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet_block.jpg b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet_block.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c500eb01a90190ff66150871fe83ec275e2de8d7 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/resnet_block.jpg differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/train_and_test.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/train_and_test.png new file mode 100644 index 0000000000000000000000000000000000000000..c6336a9a69b95dc978719ce68896e3e752e67fed Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/train_and_test.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/vgg16.png b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/vgg16.png new file mode 100644 index 0000000000000000000000000000000000000000..6270eefcfd7071bc1643ee06567e5b81aaf4c177 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/image_classification/image/vgg16.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/image_classification/index.md b/doc/fluid/new_docs/beginners_guide/basics/image_classification/index.md new file mode 100644 index 0000000000000000000000000000000000000000..ce0d2bb1dc0cf73151ee9aceea7e4d7b24af1926 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/image_classification/index.md @@ -0,0 +1,559 @@ + +# 图像分类 + +本教程源代码目录在[book/image_classification](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 + +图像相比文字能够提供更加生动、容易理解及更具艺术感的信息，是人们转递与交换信息的重要来源。在本教程中，我们专注于图像识别领域的一个重要问题，即图像分类。 + +图像分类是根据图像的语义信息将不同类别图像区分开来，是计算机视觉中重要的基本问题，也是图像检测、图像分割、物体跟踪、行为分析等其他高层视觉任务的基础。图像分类在很多领域有广泛应用，包括安防领域的人脸识别和智能视频分析等，交通领域的交通场景识别，互联网领域基于内容的图像检索和相册自动归类，医学领域的图像识别等。 + + +一般来说，图像分类通过手工特征或特征学习方法对整个图像进行全部描述，然后使用分类器判别物体类别，因此如何提取图像的特征至关重要。在深度学习算法之前使用较多的是基于词袋(Bag of Words)模型的物体分类方法。词袋方法从自然语言处理中引入，即一句话可以用一个装了词的袋子表示其特征，袋子中的词为句子中的单词、短语或字。对于图像而言，词袋方法需要构建字典。最简单的词袋模型框架可以设计为**底层特征抽取**、**特征编码**、**分类器设计**三个过程。 + +而基于深度学习的图像分类方法，可以通过有监督或无监督的方式**学习**层次化的特征描述，从而取代了手工设计或选择图像特征的工作。深度学习模型中的卷积神经网络(Convolution Neural Network, CNN)近年来在图像领域取得了惊人的成绩，CNN直接利用图像像素信息作为输入，最大程度上保留了输入图像的所有信息，通过卷积操作进行特征的提取和高层抽象，模型输出直接是图像识别的结果。这种基于"输入-输出"直接端到端的学习方法取得了非常好的效果，得到了广泛的应用。 + +本教程主要介绍图像分类的深度学习模型，以及如何使用PaddlePaddle训练CNN模型。 + +## 效果展示 + +图像分类包括通用图像分类、细粒度图像分类等。图1展示了通用图像分类效果，即模型可以正确识别图像上的主要物体。 + +![dogCatClassification](./image/dog_cat.png) +

+图1. 通用图像分类展示 +

+ + +图2展示了细粒度图像分类-花卉识别的效果，要求模型可以正确识别花的类别。 + +![flowersClassification](./image/flowers.png) +

+图2. 细粒度图像分类展示 +

+ + +一个好的模型既要对不同类别识别正确，同时也应该能够对不同视角、光照、背景、变形或部分遮挡的图像正确识别(这里我们统一称作图像扰动)。图3展示了一些图像的扰动，较好的模型会像聪明的人类一样能够正确识别。 + +![imageVariations](https://raw.githubusercontent.com/PaddlePaddle/book/develop/03.image_classification/image/variations.png) +

+图3. 扰动图片展示[22] +

+ +## 模型概览 + +图像识别领域大量的研究成果都是建立在[PASCAL VOC](http://host.robots.ox.ac.uk/pascal/VOC/)、[ImageNet](http://image-net.org/)等公开的数据集上，很多图像识别算法通常在这些数据集上进行测试和比较。PASCAL VOC是2005年发起的一个视觉挑战赛，ImageNet是2010年发起的大规模视觉识别竞赛(ILSVRC)的数据集，在本章中我们基于这些竞赛的一些论文介绍图像分类模型。 + +在2012年之前的传统图像分类方法可以用背景描述中提到的三步完成，但通常完整建立图像识别模型一般包括底层特征学习、特征编码、空间约束、分类器设计、模型融合等几个阶段。 +1). **底层特征提取**: 通常从图像中按照固定步长、尺度提取大量局部特征描述。常用的局部特征包括SIFT(Scale-Invariant Feature Transform, 尺度不变特征转换) \[[1](#参考文献)\]、HOG(Histogram of Oriented Gradient, 方向梯度直方图) \[[2](#参考文献)\]、LBP(Local Bianray Pattern, 局部二值模式) \[[3](#参考文献)\] 等，一般也采用多种特征描述子，防止丢失过多的有用信息。 +2). **特征编码**: 底层特征中包含了大量冗余与噪声，为了提高特征表达的鲁棒性，需要使用一种特征变换算法对底层特征进行编码，称作特征编码。常用的特征编码包括向量量化编码 \[[4](#参考文献)\]、稀疏编码 \[[5](#参考文献)\]、局部线性约束编码 \[[6](#参考文献)\]、Fisher向量编码 \[[7](#参考文献)\] 等。 +3). **空间特征约束**: 特征编码之后一般会经过空间特征约束，也称作**特征汇聚**。特征汇聚是指在一个空间范围内，对每一维特征取最大值或者平均值，可以获得一定特征不变形的特征表达。金字塔特征匹配是一种常用的特征聚会方法，这种方法提出将图像均匀分块，在分块内做特征汇聚。 +4). **通过分类器分类**: 经过前面步骤之后一张图像可以用一个固定维度的向量进行描述，接下来就是经过分类器对图像进行分类。通常使用的分类器包括SVM(Support Vector Machine, 支持向量机)、随机森林等。而使用核方法的SVM是最为广泛的分类器，在传统图像分类任务上性能很好。 + +这种方法在PASCAL VOC竞赛中的图像分类算法中被广泛使用 \[[18](#参考文献)\]。[NEC实验室](http://www.nec-labs.com/)在ILSVRC2010中采用SIFT和LBP特征，两个非线性编码器以及SVM分类器获得图像分类的冠军 \[[8](#参考文献)\]。 + +Alex Krizhevsky在2012年ILSVRC提出的CNN模型 \[[9](#参考文献)\] 取得了历史性的突破，效果大幅度超越传统方法，获得了ILSVRC2012冠军，该模型被称作AlexNet。这也是首次将深度学习用于大规模图像分类中。从AlexNet之后，涌现了一系列CNN模型，不断地在ImageNet上刷新成绩，如图4展示。随着模型变得越来越深以及精妙的结构设计，Top-5的错误率也越来越低，降到了3.5%附近。而在同样的ImageNet数据集上，人眼的辨识错误率大概在5.1%，也就是目前的深度学习模型的识别能力已经超过了人眼。 + +![ilsvrc](./image/ilsvrc.png) +

+图4. ILSVRC图像分类Top-5错误率 +

+ +### CNN + +传统CNN包含卷积层、全连接层等组件，并采用softmax多类别分类器和多类交叉熵损失函数，一个典型的卷积神经网络如图5所示，我们先介绍用来构造CNN的常见组件。 + +![cnnStructure](./image/lenet.png) +

+图5. CNN网络示例[20] +

+ +- 卷积层(convolution layer): 执行卷积操作提取底层到高层的特征，发掘出图片局部关联性质和空间不变性质。 +- 池化层(pooling layer): 执行降采样操作。通过取卷积输出特征图中局部区块的最大值(max-pooling)或者均值(avg-pooling)。降采样也是图像处理中常见的一种操作，可以过滤掉一些不重要的高频信息。 +- 全连接层(fully-connected layer，或者fc layer): 输入层到隐藏层的神经元是全部连接的。 +- 非线性变化: 卷积层、全连接层后面一般都会接非线性变化层，例如Sigmoid、Tanh、ReLu等来增强网络的表达能力，在CNN里最常使用的为ReLu激活函数。 +- Dropout \[[10](#参考文献)\] : 在模型训练阶段随机让一些隐层节点权重不工作，提高网络的泛化能力，一定程度上防止过拟合。 + +另外，在训练过程中由于每层参数不断更新，会导致下一次输入分布发生变化，这样导致训练过程需要精心设计超参数。如2015年Sergey Ioffe和Christian Szegedy提出了Batch Normalization (BN)算法 \[[14](#参考文献)\] 中，每个batch对网络中的每一层特征都做归一化，使得每层分布相对稳定。BN算法不仅起到一定的正则作用，而且弱化了一些超参数的设计。经过实验证明，BN算法加速了模型收敛过程，在后来较深的模型中被广泛使用。 + +接下来我们主要介绍VGG，GoogleNet和ResNet网络结构。 + +### VGG + +牛津大学VGG(Visual Geometry Group)组在2014年ILSVRC提出的模型被称作VGG模型 \[[11](#参考文献)\] 。该模型相比以往模型进一步加宽和加深了网络结构，它的核心是五组卷积操作，每两组之间做Max-Pooling空间降维。同一组内采用多次连续的3X3卷积，卷积核的数目由较浅组的64增多到最深组的512，同一组内的卷积核数目是一样的。卷积之后接两层全连接层，之后是分类层。由于每组内卷积层的不同，有11、13、16、19层这几种模型，下图展示一个16层的网络结构。VGG模型结构相对简洁，提出之后也有很多文章基于此模型进行研究，如在ImageNet上首次公开超过人眼识别的模型\[[19](#参考文献)\]就是借鉴VGG模型的结构。 + +![vgg16](./image/vgg16.png) +

+图6. 基于ImageNet的VGG16模型 +

+ +### GoogleNet + +GoogleNet \[[12](#参考文献)\] 在2014年ILSVRC的获得了冠军，在介绍该模型之前我们先来了解NIN(Network in Network)模型 \[[13](#参考文献)\] 和Inception模块，因为GoogleNet模型由多组Inception模块组成，模型设计借鉴了NIN的一些思想。 + +NIN模型主要有两个特点：1) 引入了多层感知卷积网络(Multi-Layer Perceptron Convolution, MLPconv)代替一层线性卷积网络。MLPconv是一个微小的多层卷积网络，即在线性卷积后面增加若干层1x1的卷积，这样可以提取出高度非线性特征。2) 传统的CNN最后几层一般都是全连接层，参数较多。而NIN模型设计最后一层卷积层包含类别维度大小的特征图，然后采用全局均值池化(Avg-Pooling)替代全连接层，得到类别维度大小的向量，再进行分类。这种替代全连接层的方式有利于减少参数。 + +Inception模块如下图7所示，图(a)是最简单的设计，输出是3个卷积层和一个池化层的特征拼接。这种设计的缺点是池化层不会改变特征通道数，拼接后会导致特征的通道数较大，经过几层这样的模块堆积后，通道数会越来越大，导致参数和计算量也随之增大。为了改善这个缺点，图(b)引入3个1x1卷积层进行降维，所谓的降维就是减少通道数，同时如NIN模型中提到的1x1卷积也可以修正线性特征。 + +![inception](./image/inception.png) +

+图7. Inception模块 +

+ +GoogleNet由多组Inception模块堆积而成。另外，在网络最后也没有采用传统的多层全连接层，而是像NIN网络一样采用了均值池化层；但与NIN不同的是，池化层后面接了一层到类别数映射的全连接层。除了这两个特点之外，由于网络中间层特征也很有判别性，GoogleNet在中间层添加了两个辅助分类器，在后向传播中增强梯度并且增强正则化，而整个网络的损失函数是这个三个分类器的损失加权求和。 + +GoogleNet整体网络结构如图8所示，总共22层网络：开始由3层普通的卷积组成；接下来由三组子网络组成，第一组子网络包含2个Inception模块，第二组包含5个Inception模块，第三组包含2个Inception模块；然后接均值池化层、全连接层。 + +![googleNet](./image/googlenet.jpeg) +

+图8. GoogleNet[12] +

+ + +上面介绍的是GoogleNet第一版模型(称作GoogleNet-v1)。GoogleNet-v2 \[[14](#参考文献)\] 引入BN层；GoogleNet-v3 \[[16](#参考文献)\] 对一些卷积层做了分解，进一步提高网络非线性能力和加深网络；GoogleNet-v4 \[[17](#参考文献)\] 引入下面要讲的ResNet设计思路。从v1到v4每一版的改进都会带来准确度的提升，介于篇幅，这里不再详细介绍v2到v4的结构。 + + +### ResNet + +ResNet(Residual Network) \[[15](#参考文献)\] 是2015年ImageNet图像分类、图像物体定位和图像物体检测比赛的冠军。针对训练卷积神经网络时加深网络导致准确度下降的问题，ResNet提出了采用残差学习。在已有设计思路(BN, 小卷积核，全卷积网络)的基础上，引入了残差模块。每个残差模块包含两条路径，其中一条路径是输入特征的直连通路，另一条路径对该特征做两到三次卷积操作得到该特征的残差，最后再将两条路径上的特征相加。 + +残差模块如图9所示，左边是基本模块连接方式，由两个输出通道数相同的3x3卷积组成。右边是瓶颈模块(Bottleneck)连接方式，之所以称为瓶颈，是因为上面的1x1卷积用来降维(图示例即256->64)，下面的1x1卷积用来升维(图示例即64->256)，这样中间3x3卷积的输入和输出通道数都较小(图示例即64->64)。 + +![ResNetBlock](./image/resnet_block.jpg) +

+图9. 残差模块 +

+ +图10展示了50、101、152层网络连接示意图，使用的是瓶颈模块。这三个模型的区别在于每组中残差模块的重复次数不同(见图右上角)。ResNet训练收敛较快，成功的训练了上百乃至近千层的卷积神经网络。 + +![ResNet](./image/resnet.png) +

+图10. 基于ImageNet的ResNet模型 +

+ + +## 数据准备 + +通用图像分类公开的标准数据集常用的有[CIFAR](https://www.cs.toronto.edu/~kriz/cifar.html)、[ImageNet](http://image-net.org/)、[COCO](http://mscoco.org/)等，常用的细粒度图像分类数据集包括[CUB-200-2011](http://www.vision.caltech.edu/visipedia/CUB-200-2011.html)、[Stanford Dog](http://vision.stanford.edu/aditya86/ImageNetDogs/)、[Oxford-flowers](http://www.robots.ox.ac.uk/~vgg/data/flowers/)等。其中ImageNet数据集规模相对较大，如[模型概览](#模型概览)一章所讲，大量研究成果基于ImageNet。ImageNet数据从2010年来稍有变化，常用的是ImageNet-2012数据集，该数据集包含1000个类别：训练集包含1,281,167张图片，每个类别数据732至1300张不等，验证集包含50,000张图片，平均每个类别50张图片。 + +由于ImageNet数据集较大，下载和训练较慢，为了方便大家学习，我们使用[CIFAR10]()数据集。CIFAR10数据集包含60,000张32x32的彩色图片，10个类别，每个类包含6,000张。其中50,000张图片作为训练集，10000张作为测试集。图11从每个类别中随机抽取了10张图片，展示了所有的类别。 + +![CIFAR](https://raw.githubusercontent.com/PaddlePaddle/book/develop/03.image_classification/image/cifar.png) +

+图11. CIFAR10数据集[21] +

+ +Paddle API提供了自动加载cifar数据集模块 `paddle.dataset.cifar`。 + +通过输入`python train.py`，就可以开始训练模型了，以下小节将详细介绍`train.py`的相关内容。 + +### 模型结构 + +#### Paddle 初始化 + +让我们从导入 Paddle Fluid API 和辅助模块开始。 + +```python +import paddle +import paddle.fluid as fluid +import numpy +import sys +``` + +本教程中我们提供了VGG和ResNet两个模型的配置。 + +#### VGG + +首先介绍VGG模型结构，由于CIFAR10图片大小和数量相比ImageNet数据小很多，因此这里的模型针对CIFAR10数据做了一定的适配。卷积部分引入了BN和Dropout操作。 +VGG核心模块的输入是数据层，`vgg_bn_drop` 定义了16层VGG结构，每层卷积后面引入BN层和Dropout层，详细的定义如下： + +```python +def vgg_bn_drop(input): +def conv_block(ipt, num_filter, groups, dropouts): +return fluid.nets.img_conv_group( +input=ipt, +pool_size=2, +pool_stride=2, +conv_num_filter=[num_filter] * groups, +conv_filter_size=3, +conv_act='relu', +conv_with_batchnorm=True, +conv_batchnorm_drop_rate=dropouts, +pool_type='max') + +conv1 = conv_block(input, 64, 2, [0.3, 0]) +conv2 = conv_block(conv1, 128, 2, [0.4, 0]) +conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) +conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) +conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) + +drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) +fc1 = fluid.layers.fc(input=drop, size=512, act=None) +bn = fluid.layers.batch_norm(input=fc1, act='relu') +drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) +fc2 = fluid.layers.fc(input=drop2, size=512, act=None) +predict = fluid.layers.fc(input=fc2, size=10, act='softmax') +return predict +``` + +1. 首先定义了一组卷积网络，即conv_block。卷积核大小为3x3，池化窗口大小为2x2，窗口滑动大小为2，groups决定每组VGG模块是几次连续的卷积操作，dropouts指定Dropout操作的概率。所使用的`img_conv_group`是在`paddle.networks`中预定义的模块，由若干组 Conv->BN->ReLu->Dropout 和 一组 Pooling 组成。 + +2. 五组卷积操作，即 5个conv_block。 第一、二组采用两次连续的卷积操作。第三、四、五组采用三次连续的卷积操作。每组最后一个卷积后面Dropout概率为0，即不使用Dropout操作。 + +3. 最后接两层512维的全连接。 + +4. 通过上面VGG网络提取高层特征，然后经过全连接层映射到类别维度大小的向量，再通过Softmax归一化得到每个类别的概率，也可称作分类器。 + +### ResNet + +ResNet模型的第1、3、4步和VGG模型相同，这里不再介绍。主要介绍第2步即CIFAR10数据集上ResNet核心模块。 + +先介绍`resnet_cifar10`中的一些基本函数，再介绍网络连接过程。 + +- `conv_bn_layer` : 带BN的卷积层。 +- `shortcut` : 残差模块的"直连"路径，"直连"实际分两种形式：残差模块输入和输出特征通道数不等时，采用1x1卷积的升维操作；残差模块输入和输出通道相等时，采用直连操作。 +- `basicblock` : 一个基础残差模块，即图9左边所示，由两组3x3卷积组成的路径和一条"直连"路径组成。 +- `bottleneck` : 一个瓶颈残差模块，即图9右边所示，由上下1x1卷积和中间3x3卷积组成的路径和一条"直连"路径组成。 +- `layer_warp` : 一组残差模块，由若干个残差模块堆积而成。每组中第一个残差模块滑动窗口大小与其他可以不同，以用来减少特征图在垂直和水平方向的大小。 + +```python +def conv_bn_layer(input, +ch_out, +filter_size, +stride, +padding, +act='relu', +bias_attr=False): +tmp = fluid.layers.conv2d( +input=input, +filter_size=filter_size, +num_filters=ch_out, +stride=stride, +padding=padding, +act=None, +bias_attr=bias_attr) +return fluid.layers.batch_norm(input=tmp, act=act) + + +def shortcut(input, ch_in, ch_out, stride): +if ch_in != ch_out: +return conv_bn_layer(input, ch_out, 1, stride, 0, None) +else: +return input + + +def basicblock(input, ch_in, ch_out, stride): +tmp = conv_bn_layer(input, ch_out, 3, stride, 1) +tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None, bias_attr=True) +short = shortcut(input, ch_in, ch_out, stride) +return fluid.layers.elementwise_add(x=tmp, y=short, act='relu') + + +def layer_warp(block_func, input, ch_in, ch_out, count, stride): +tmp = block_func(input, ch_in, ch_out, stride) +for i in range(1, count): +tmp = block_func(tmp, ch_out, ch_out, 1) +return tmp +``` + +`resnet_cifar10` 的连接结构主要有以下几个过程。 + +1. 底层输入连接一层 `conv_bn_layer`，即带BN的卷积层。 +2. 然后连接3组残差模块即下面配置3组 `layer_warp` ，每组采用图 10 左边残差模块组成。 +3. 最后对网络做均值池化并返回该层。 + +注意：除过第一层卷积层和最后一层全连接层之外，要求三组 `layer_warp` 总的含参层数能够被6整除，即 `resnet_cifar10` 的 depth 要满足 `$(depth - 2) % 6 == 0$` 。 + +```python +def resnet_cifar10(ipt, depth=32): +# depth should be one of 20, 32, 44, 56, 110, 1202 +assert (depth - 2) % 6 == 0 +n = (depth - 2) / 6 +nStages = {16, 64, 128} +conv1 = conv_bn_layer(ipt, ch_out=16, filter_size=3, stride=1, padding=1) +res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) +res2 = layer_warp(basicblock, res1, 16, 32, n, 2) +res3 = layer_warp(basicblock, res2, 32, 64, n, 2) +pool = fluid.layers.pool2d( +input=res3, pool_size=8, pool_type='avg', pool_stride=1) +predict = fluid.layers.fc(input=pool, size=10, act='softmax') +return predict +``` + +## Infererence Program 配置 + +网络输入定义为 `data_layer` (数据层)，在图像分类中即为图像像素信息。CIFRAR10是RGB 3通道32x32大小的彩色图，因此输入数据大小为3072(3x32x32)。 + +```python +def inference_program(): +# The image is 32 * 32 with RGB representation. +data_shape = [3, 32, 32] +images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + +predict = resnet_cifar10(images, 32) +# predict = vgg_bn_drop(images) # un-comment to use vgg net +return predict +``` + +## Train Program 配置 + +然后我们需要设置训练程序 `train_program`。它首先从推理程序中进行预测。 +在训练期间，它将从预测中计算 `avg_cost`。 +在有监督训练中需要输入图像对应的类别信息，同样通过`fluid.layers.data`来定义。训练中采用多类交叉熵作为损失函数，并作为网络的输出，预测阶段定义网络的输出为分类器得到的概率信息。 + +**注意:** 训练程序应该返回一个数组，第一个返回参数必须是 `avg_cost`。训练器使用它来计算梯度。 + +```python +def train_program(): +predict = inference_program() + +label = fluid.layers.data(name='label', shape=[1], dtype='int64') +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(cost) +accuracy = fluid.layers.accuracy(input=predict, label=label) +return [avg_cost, accuracy] +``` + +## Optimizer Function 配置 + +在下面的 `Adam optimizer`，`learning_rate` 是训练的速度，与网络的训练收敛速度有关系。 + +```python +def optimizer_program(): +return fluid.optimizer.Adam(learning_rate=0.001) +``` + +## 训练模型 + +### Trainer 配置 + +现在，我们需要配置 `Trainer`。`Trainer` 需要接受训练程序 `train_program`, `place` 和优化器 `optimizer_func`。 + +```python +use_cuda = False +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +trainer = fluid.Trainer( +train_func=train_program, +optimizer_func=optimizer_program, +place=place) +``` + +### Data Feeders 配置 + +`cifar.train10()` 每次产生一条样本，在完成shuffle和batch之后，作为训练的输入。 + +```python +# Each batch will yield 128 images +BATCH_SIZE = 128 + +# Reader for training +train_reader = paddle.batch( +paddle.reader.shuffle(paddle.dataset.cifar.train10(), buf_size=50000), +batch_size=BATCH_SIZE) + +# Reader for testing. A separated data set for testing. +test_reader = paddle.batch( +paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) +``` + +### Event Handler + +可以使用`event_handler`回调函数来观察训练过程，或进行测试等, 该回调函数是`trainer.train`函数里设定。 + +`event_handler_plot`可以用来利用回调数据来打点画图: + +![png](./image/train_and_test.png) + +```python +params_dirname = "image_classification_resnet.inference.model" + +from paddle.v2.plot import Ploter + +train_title = "Train cost" +test_title = "Test cost" +cost_ploter = Ploter(train_title, test_title) + +step = 0 +def event_handler_plot(event): +global step +if isinstance(event, fluid.EndStepEvent): +if step % 1 == 0: +cost_ploter.append(train_title, step, event.metrics[0]) +cost_ploter.plot() +step += 1 +if isinstance(event, fluid.EndEpochEvent): +avg_cost, accuracy = trainer.test( +reader=test_reader, +feed_order=['pixel', 'label']) +cost_ploter.append(test_title, step, avg_cost) + +# save parameters +if params_dirname is not None: +trainer.save_params(params_dirname) +``` + +`event_handler` 用来在训练过程中输出文本日志 + +```python +params_dirname = "image_classification_resnet.inference.model" + +# event handler to track training and testing process +def event_handler(event): +if isinstance(event, fluid.EndStepEvent): +if event.step % 100 == 0: +print("\nPass %d, Batch %d, Cost %f, Acc %f" % +(event.step, event.epoch, event.metrics[0], +event.metrics[1])) +else: +sys.stdout.write('.') +sys.stdout.flush() + +if isinstance(event, fluid.EndEpochEvent): +# Test against with the test dataset to get accuracy. +avg_cost, accuracy = trainer.test( +reader=test_reader, feed_order=['pixel', 'label']) + +print('\nTest with Pass {0}, Loss {1:2.2}, Acc {2:2.2}'.format(event.epoch, avg_cost, accuracy)) + +# save parameters +if params_dirname is not None: +trainer.save_params(params_dirname) +``` + +### 训练 + +通过`trainer.train`函数训练: + +**注意:** CPU，每个 Epoch 将花费大约15～20分钟。这部分可能需要一段时间。请随意修改代码，在GPU上运行测试，以提高培训速度。 + +```python +trainer.train( +reader=train_reader, +num_epochs=2, +event_handler=event_handler, +feed_order=['pixel', 'label']) +``` + +一轮训练log示例如下所示，经过1个pass， 训练集上平均 Accuracy 为0.59 ，测试集上平均 Accuracy 为0.6 。 + +```text +Pass 0, Batch 0, Cost 3.869598, Acc 0.164062 +................................................................................................... +Pass 100, Batch 0, Cost 1.481038, Acc 0.460938 +................................................................................................... +Pass 200, Batch 0, Cost 1.340323, Acc 0.523438 +................................................................................................... +Pass 300, Batch 0, Cost 1.223424, Acc 0.593750 +.......................................................................................... +Test with Pass 0, Loss 1.1, Acc 0.6 +``` + +图12是训练的分类错误率曲线图，运行到第200个pass后基本收敛，最终得到测试集上分类错误率为8.54%。 + +![CIFARErrorRate](./image/plot.png) +

+图12. CIFAR10数据集上VGG模型的分类错误率 +

+ +## 应用模型 + +可以使用训练好的模型对图片进行分类，下面程序展示了如何使用 `fluid.Inferencer` 接口进行推断，可以打开注释，更改加载的模型。 + +### 生成预测输入数据 + +`dog.png` is an example image of a dog. Turn it into an numpy array to match the data feeder format. + +```python +# Prepare testing data. +from PIL import Image +import numpy as np +import os + +def load_image(file): +im = Image.open(file) +im = im.resize((32, 32), Image.ANTIALIAS) + +im = np.array(im).astype(np.float32) +# The storage order of the loaded image is W(width), +# H(height), C(channel). PaddlePaddle requires +# the CHW order, so transpose them. +im = im.transpose((2, 0, 1)) # CHW +im = im / 255.0 + +# Add one dimension to mimic the list format. +im = numpy.expand_dims(im, axis=0) +return im + +cur_dir = os.getcwd() +img = load_image(cur_dir + '/image/dog.png') +``` + +### Inferencer 配置和预测 + +`Inferencer` 需要一个 `infer_func` 和 `param_path` 来设置网络和经过训练的参数。 +我们可以简单地插入前面定义的推理程序。 +现在我们准备做预测。 + +```python +inferencer = fluid.Inferencer( +infer_func=inference_program, param_path=params_dirname, place=place) + +# inference +results = inferencer.infer({'pixel': img}) +print("infer results: ", results) +``` + +## 总结 + +传统图像分类方法由多个阶段构成，框架较为复杂，而端到端的CNN模型结构可一步到位，而且大幅度提升了分类准确率。本文我们首先介绍VGG、GoogleNet、ResNet三个经典的模型；然后基于CIFAR10数据集，介绍如何使用PaddlePaddle配置和训练CNN模型，尤其是VGG和ResNet模型；最后介绍如何使用PaddlePaddle的API接口对图片进行预测和特征提取。对于其他数据集比如ImageNet，配置和训练流程是同样的，大家可以自行进行实验。 + + +## 参考文献 + +[1] D. G. Lowe, [Distinctive image features from scale-invariant keypoints](http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf). IJCV, 60(2):91-110, 2004. + +[2] N. Dalal, B. Triggs, [Histograms of Oriented Gradients for Human Detection](http://vision.stanford.edu/teaching/cs231b_spring1213/papers/CVPR05_DalalTriggs.pdf), Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2005. + +[3] Ahonen, T., Hadid, A., and Pietikinen, M. (2006). [Face description with local binary patterns: Application to face recognition](http://ieeexplore.ieee.org/document/1717463/). PAMI, 28. + +[4] J. Sivic, A. Zisserman, [Video Google: A Text Retrieval Approach to Object Matching in Videos](http://www.robots.ox.ac.uk/~vgg/publications/papers/sivic03.pdf), Proc. Ninth Int'l Conf. Computer Vision, pp. 1470-1478, 2003. + +[5] B. Olshausen, D. Field, [Sparse Coding with an Overcomplete Basis Set: A Strategy Employed by V1?](http://redwood.psych.cornell.edu/papers/olshausen_field_1997.pdf), Vision Research, vol. 37, pp. 3311-3325, 1997. + +[6] Wang, J., Yang, J., Yu, K., Lv, F., Huang, T., and Gong, Y. (2010). [Locality-constrained Linear Coding for image classification](http://ieeexplore.ieee.org/abstract/document/5540018/). In CVPR. + +[7] Perronnin, F., Sánchez, J., & Mensink, T. (2010). [Improving the fisher kernel for large-scale image classification](http://dl.acm.org/citation.cfm?id=1888101). In ECCV (4). + +[8] Lin, Y., Lv, F., Cao, L., Zhu, S., Yang, M., Cour, T., Yu, K., and Huang, T. (2011). [Large-scale image clas- sification: Fast feature extraction and SVM training](http://ieeexplore.ieee.org/document/5995477/). In CVPR. + +[9] Krizhevsky, A., Sutskever, I., and Hinton, G. (2012). [ImageNet classification with deep convolutional neu- ral networks](http://www.cs.toronto.edu/~kriz/imagenet_classification_with_deep_convolutional.pdf). In NIPS. + +[10] G.E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R.R. Salakhutdinov. [Improving neural networks by preventing co-adaptation of feature detectors](https://arxiv.org/abs/1207.0580). arXiv preprint arXiv:1207.0580, 2012. + +[11] K. Chatfield, K. Simonyan, A. Vedaldi, A. Zisserman. [Return of the Devil in the Details: Delving Deep into Convolutional Nets](https://arxiv.org/abs/1405.3531). BMVC, 2014。 + +[12] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A., [Going deeper with convolutions](https://arxiv.org/abs/1409.4842). In: CVPR. (2015) + +[13] Lin, M., Chen, Q., and Yan, S. [Network in network](https://arxiv.org/abs/1312.4400). In Proc. ICLR, 2014. + +[14] S. Ioffe and C. Szegedy. [Batch normalization: Accelerating deep network training by reducing internal covariate shift](https://arxiv.org/abs/1502.03167). In ICML, 2015. + +[15] K. He, X. Zhang, S. Ren, J. Sun. [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385). CVPR 2016. + +[16] Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z. [Rethinking the incep-tion architecture for computer vision](https://arxiv.org/abs/1512.00567). In: CVPR. (2016). + +[17] Szegedy, C., Ioffe, S., Vanhoucke, V. [Inception-v4, inception-resnet and the impact of residual connections on learning](https://arxiv.org/abs/1602.07261). arXiv:1602.07261 (2016). + +[18] Everingham, M., Eslami, S. M. A., Van Gool, L., Williams, C. K. I., Winn, J. and Zisserman, A. [The Pascal Visual Object Classes Challenge: A Retrospective]((http://link.springer.com/article/10.1007/s11263-014-0733-5)). International Journal of Computer Vision, 111(1), 98-136, 2015. + +[19] He, K., Zhang, X., Ren, S., and Sun, J. [Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification](https://arxiv.org/abs/1502.01852). ArXiv e-prints, February 2015. + +[20] http://deeplearning.net/tutorial/lenet.html + +[21] https://www.cs.toronto.edu/~kriz/cifar.html + +[22] http://cs231n.github.io/classification/ + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/basics/index.rst b/doc/fluid/new_docs/beginners_guide/basics/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..d16f8b947253a535567ddc8d7b227dd153d9b154 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/index.rst @@ -0,0 +1,18 @@ +################ +深度学习基础知识 +################ + + +.. todo:: + + 概述 + +.. toctree:: + :maxdepth: 2 + + image_classification/index.md + word2vec/index.md + recommender_system/index.md + understand_sentiment/index.md + label_semantic_roles/index.md + machine_translation/index.md diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/.gitignore b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..29b5622a53a1b0847e9f53febf1cc50dcf4f044a --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/.gitignore @@ -0,0 +1,12 @@ +data/train.list +data/test.* +data/conll05st-release.tar.gz +data/conll05st-release +data/predicate_dict +data/label_dict +data/word_dict +data/emb +data/feature +output +predict.res +train.log diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm.png new file mode 100644 index 0000000000000000000000000000000000000000..e63f5ebd6d00f2e4ecf97b9ab2027e74683013f2 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm_en.png new file mode 100755 index 0000000000000000000000000000000000000000..f0a195c24d9ee493f96bb93c28a99e70566be7a4 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bidirectional_stacked_lstm_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example.png new file mode 100755 index 0000000000000000000000000000000000000000..e5f7151c9fcc50a7cf7af485cbbc7e4fccab0c20 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example_en.png new file mode 100755 index 0000000000000000000000000000000000000000..93b44dd4874402ef29ad7bd7d94147609b92e309 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/bio_example_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network.png new file mode 100644 index 0000000000000000000000000000000000000000..592f7ee23bdc88a9a35059612e5ab880bbc9d34b Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network_en.png new file mode 100755 index 0000000000000000000000000000000000000000..c3646312e48db977402fb353dc0c9b4d02269bf4 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/db_lstm_network_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing.png new file mode 100755 index 0000000000000000000000000000000000000000..9265b671735940ed6549e2980064d2ce08baae64 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing_en.png new file mode 100755 index 0000000000000000000000000000000000000000..23f4f45b603e3d60702af2b2464d10fc8deed061 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/dependency_parsing_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/linear_chain_crf.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/linear_chain_crf.png new file mode 100644 index 0000000000000000000000000000000000000000..0778fda74b2ad22ce4b631791a7b028cdef780a5 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/linear_chain_crf.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm.png new file mode 100644 index 0000000000000000000000000000000000000000..3d2914c726b5f4c46e66dfa85d4e88649fede6b3 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm_en.png new file mode 100755 index 0000000000000000000000000000000000000000..0b944ef91e8b5ba4b14d2a35bd8879f261cf8f61 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/image/stacked_lstm_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/index.md b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/index.md new file mode 100644 index 0000000000000000000000000000000000000000..828ca738317992270487647e66b08b6d2f80e209 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/label_semantic_roles/index.md @@ -0,0 +1,568 @@ +# 语义角色标注 + +本教程源代码目录在[book/label_semantic_roles](https://github.com/PaddlePaddle/book/tree/develop/07.label_semantic_roles)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 + +自然语言分析技术大致分为三个层面：词法分析、句法分析和语义分析。语义角色标注是实现浅层语义分析的一种方式。在一个句子中，谓词是对主语的陈述或说明，指出“做什么”、“是什么”或“怎么样，代表了一个事件的核心，跟谓词搭配的名词称为论元。语义角色是指论元在动词所指事件中担任的角色。主要有：施事者（Agent）、受事者（Patient）、客体（Theme）、经验者（Experiencer）、受益者（Beneficiary）、工具（Instrument）、处所（Location）、目标（Goal）和来源（Source）等。 + +请看下面的例子，“遇到” 是谓词（Predicate，通常简写为“Pred”），“小明”是施事者（Agent），“小红”是受事者（Patient），“昨天” 是事件发生的时间（Time），“公园”是事情发生的地点（Location）。 + +$$\mbox{[小明]}_{\mbox{Agent}}\mbox{[昨天]}_{\mbox{Time}}\mbox{[晚上]}_{\mbox{Time}}\mbox{在[公园]}_{\mbox{Location}}\mbox{[遇到]}_{\mbox{Predicate}}\mbox{了[小红]}_{\mbox{Patient}}\mbox{。}$$ + +语义角色标注（Semantic Role Labeling，SRL）以句子的谓词为中心，不对句子所包含的语义信息进行深入分析，只分析句子中各成分与谓词之间的关系，即句子的谓词（Predicate）- 论元（Argument）结构，并用语义角色来描述这些结构关系，是许多自然语言理解任务（如信息抽取，篇章分析，深度问答等）的一个重要中间步骤。在研究中一般都假定谓词是给定的，所要做的就是找出给定谓词的各个论元和它们的语义角色。 + +传统的SRL系统大多建立在句法分析基础之上，通常包括5个流程： + +1. 构建一棵句法分析树，例如，图1是对上面例子进行依存句法分析得到的一棵句法树。 +2. 从句法树上识别出给定谓词的候选论元。 +3. 候选论元剪除；一个句子中的候选论元可能很多，候选论元剪除就是从大量的候选项中剪除那些最不可能成为论元的候选项。 +4. 论元识别：这个过程是从上一步剪除之后的候选中判断哪些是真正的论元，通常当做一个二分类问题来解决。 +5. 对第4步的结果，通过多分类得到论元的语义角色标签。可以看到，句法分析是基础，并且后续步骤常常会构造的一些人工特征，这些特征往往也来自句法分析。 + +![dependencyParsing](./image/dependency_parsing.png) +
+图1. 依存句法分析句法树示例 +
+ +然而，完全句法分析需要确定句子所包含的全部句法信息，并确定句子各成分之间的关系，是一个非常困难的任务，目前技术下的句法分析准确率并不高，句法分析的细微错误都会导致SRL的错误。为了降低问题的复杂度，同时获得一定的句法结构信息，“浅层句法分析”的思想应运而生。浅层句法分析也称为部分句法分析（partial parsing）或语块划分（chunking）。和完全句法分析得到一颗完整的句法树不同，浅层句法分析只需要识别句子中某些结构相对简单的独立成分，例如：动词短语，这些被识别出来的结构称为语块。为了回避 “无法获得准确率较高的句法树” 所带来的困难，一些研究\[[1](#参考文献)\]也提出了基于语块（chunk）的SRL方法。基于语块的SRL方法将SRL作为一个序列标注问题来解决。序列标注任务一般都会采用BIO表示方式来定义序列标注的标签集，我们先来介绍这种表示方法。在BIO表示法中，B代表语块的开始，I代表语块的中间，O代表语块结束。通过B、I、O 三种标记将不同的语块赋予不同的标签，例如：对于一个角色为A的论元，将它所包含的第一个语块赋予标签B-A，将它所包含的其它语块赋予标签I-A，不属于任何论元的语块赋予标签O。 + +我们继续以上面的这句话为例，图1展示了BIO表示方法。 + +![bioExample](./image/bio_example.png) +
+图2. BIO标注方法示例 +
+ +从上面的例子可以看到，根据序列标注结果可以直接得到论元的语义角色标注结果，是一个相对简单的过程。这种简单性体现在：（1）依赖浅层句法分析，降低了句法分析的要求和难度；（2）没有了候选论元剪除这一步骤；（3）论元的识别和论元标注是同时实现的。这种一体化处理论元识别和论元标注的方法，简化了流程，降低了错误累积的风险，往往能够取得更好的结果。 + +与基于语块的SRL方法类似，在本教程中我们也将SRL看作一个序列标注问题，不同的是，我们只依赖输入文本序列，不依赖任何额外的语法解析结果或是复杂的人造特征，利用深度神经网络构建一个端到端学习的SRL系统。我们以[CoNLL-2004 and CoNLL-2005 Shared Tasks](http://www.cs.upc.edu/~srlconll/)任务中SRL任务的公开数据集为例，实践下面的任务：给定一句话和这句话里的一个谓词，通过序列标注的方式，从句子中找到谓词对应的论元，同时标注它们的语义角色。 + +## 模型概览 + +循环神经网络（Recurrent Neural Network）是一种对序列建模的重要模型，在自然语言处理任务中有着广泛地应用。不同于前馈神经网络（Feed-forward Neural Network），RNN能够处理输入之间前后关联的问题。LSTM是RNN的一种重要变种，常用来学习长序列中蕴含的长程依赖关系，我们在[情感分析](https://github.com/PaddlePaddle/book/tree/develop/05.understand_sentiment)一篇中已经介绍过，这一篇中我们依然利用LSTM来解决SRL问题。 + +### 栈式循环神经网络（Stacked Recurrent Neural Network） + +深层网络有助于形成层次化特征，网络上层在下层已经学习到的初级特征基础上，形成更复杂的高级特征。尽管LSTM沿时间轴展开后等价于一个非常“深”的前馈网络，但由于LSTM各个时间步参数共享，`$t-1$`时刻状态到`$t$`时刻的映射，始终只经过了一次非线性映射，也就是说单层LSTM对状态转移的建模是 “浅” 的。堆叠多个LSTM单元，令前一个LSTM`$t$`时刻的输出，成为下一个LSTM单元`$t$`时刻的输入，帮助我们构建起一个深层网络，我们把它称为第一个版本的栈式循环神经网络。深层网络提高了模型拟合复杂模式的能力，能够更好地建模跨不同时间步的模式\[[2](#参考文献)\]。 + +然而，训练一个深层LSTM网络并非易事。纵向堆叠多个LSTM单元可能遇到梯度在纵向深度上传播受阻的问题。通常，堆叠4层LSTM单元可以正常训练，当层数达到4~8层时，会出现性能衰减，这时必须考虑一些新的结构以保证梯度纵向顺畅传播，这是训练深层LSTM网络必须解决的问题。我们可以借鉴LSTM解决 “梯度消失梯度爆炸” 问题的智慧之一：在记忆单元（Memory Cell）这条信息传播的路线上没有非线性映射，当梯度反向传播时既不会衰减、也不会爆炸。因此，深层LSTM模型也可以在纵向上添加一条保证梯度顺畅传播的路径。 + +一个LSTM单元完成的运算可以被分为三部分：（1）输入到隐层的映射（input-to-hidden） ：每个时间步输入信息`$x$`会首先经过一个矩阵映射，再作为遗忘门，输入门，记忆单元，输出门的输入，注意，这一次映射没有引入非线性激活；（2）隐层到隐层的映射（hidden-to-hidden）：这一步是LSTM计算的主体，包括遗忘门，输入门，记忆单元更新，输出门的计算；（3）隐层到输出的映射（hidden-to-output）：通常是简单的对隐层向量进行激活。我们在第一个版本的栈式网络的基础上，加入一条新的路径：除上一层LSTM输出之外，将前层LSTM的输入到隐层的映射作为的一个新的输入，同时加入一个线性映射去学习一个新的变换。 + +图3是最终得到的栈式循环神经网络结构示意图。 + +![lstmStructure](./image/stacked_lstm.png) +

+图3. 基于LSTM的栈式循环神经网络结构示意图 +

+ +### 双向循环神经网络（Bidirectional Recurrent Neural Network） + +在LSTM中，`$t$`时刻的隐藏层向量编码了到`$t$`时刻为止所有输入的信息，但`$t$`时刻的LSTM可以看到历史，却无法看到未来。在绝大多数自然语言处理任务中，我们几乎总是能拿到整个句子。这种情况下，如果能够像获取历史信息一样，得到未来的信息，对序列学习任务会有很大的帮助。 + +为了克服这一缺陷，我们可以设计一种双向循环网络单元，它的思想简单且直接：对上一节的栈式循环神经网络进行一个小小的修改，堆叠多个LSTM单元，让每一层LSTM单元分别以：正向、反向、正向 …… 的顺序学习上一层的输出序列。于是，从第2层开始，`$t$`时刻我们的LSTM单元便总是可以看到历史和未来的信息。图4是基于LSTM的双向循环神经网络结构示意图。 + +![lstmStructure](./image/bidirectional_stacked_lstm.png) +

+图4. 基于LSTM的双向循环神经网络结构示意图 +

+ +需要说明的是，这种双向RNN结构和Bengio等人在机器翻译任务中使用的双向RNN结构\[[3](#参考文献), [4](#参考文献)\] 并不相同，我们会在后续[机器翻译](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.cn.md)任务中，介绍另一种双向循环神经网络。 + +### 条件随机场 (Conditional Random Field) + +使用神经网络模型解决问题的思路通常是：前层网络学习输入的特征表示，网络的最后一层在特征基础上完成最终的任务。在SRL任务中，深层LSTM网络学习输入的特征表示，条件随机场（Conditional Random Filed， CRF）在特征的基础上完成序列标注，处于整个网络的末端。 + +CRF是一种概率化结构模型，可以看作是一个概率无向图模型，结点表示随机变量，边表示随机变量之间的概率依赖关系。简单来讲，CRF学习条件概率`$P(X|Y)$`，其中 `$X = (x_1, x_2, ... , x_n)$` 是输入序列，`$Y = (y_1, y_2, ... , y_n)$` 是标记序列；解码过程是给定 `$X$`序列求解令`$P(Y|X)$`最大的`$Y$`序列，即`$Y^* = \mbox{arg max}_{Y} P(Y | X)$`。 + +序列标注任务只需要考虑输入和输出都是一个线性序列，并且由于我们只是将输入序列作为条件，不做任何条件独立假设，因此输入序列的元素之间并不存在图结构。综上，在序列标注任务中使用的是如图5所示的定义在链式图上的CRF，称之为线性链条件随机场（Linear Chain Conditional Random Field）。 + +![linear_chain_crf](./image/linear_chain_crf.png) +

+图5. 序列标注任务中使用的线性链条件随机场 +

+ +根据线性链条件随机场上的因子分解定理\[[5](#参考文献)\]，在给定观测序列`$X$`时，一个特定标记序列`$Y$`的概率可以定义为： + +$$p(Y | X) = \frac{1}{Z(X)} \text{exp}\left(\sum_{i=1}^{n}\left(\sum_{j}\lambda_{j}t_{j} (y_{i - 1}, y_{i}, X, i) + \sum_{k} \mu_k s_k (y_i, X, i)\right)\right)$$ + +其中`$Z(X)$`是归一化因子，`$t_j$` 是定义在边上的特征函数，依赖于当前和前一个位置，称为转移特征，表示对于输入序列`$X$`及其标注序列在 `$i$`及`$i - 1$`位置上标记的转移概率。`$s_k$`是定义在结点上的特征函数，称为状态特征，依赖于当前位置，表示对于观察序列`$X$`及其`$i$`位置的标记概率。`$\lambda_j$` 和 `$\mu_k$` 分别是转移特征函数和状态特征函数对应的权值。实际上，`$t$`和`$s$`可以用相同的数学形式表示，再对转移特征和状态特在各个位置`$i$`求和有：`$f_{k}(Y, X) = \sum_{i=1}^{n}f_k({y_{i - 1}, y_i, X, i})$`，把`$f$`统称为特征函数，于是`$P(Y|X)$`可表示为： + +$$p(Y|X, W) = \frac{1}{Z(X)}\text{exp}\sum_{k}\omega_{k}f_{k}(Y, X)$$ + +`$\omega$`是特征函数对应的权值，是CRF模型要学习的参数。训练时，对于给定的输入序列和对应的标记序列集合`$D = \left[(X_1, Y_1), (X_2 , Y_2) , ... , (X_N, Y_N)\right]$` ，通过正则化的极大似然估计，求解如下优化目标： + +$$\DeclareMathOperator*{\argmax}{arg\,max} L(\lambda, D) = - \text{log}\left(\prod_{m=1}^{N}p(Y_m|X_m, W)\right) + C \frac{1}{2}\lVert W\rVert^{2}$$ + +这个优化目标可以通过反向传播算法和整个神经网络一起求解。解码时，对于给定的输入序列`$X$`，通过解码算法（通常有：维特比算法、Beam Search）求令出条件概率`$\bar{P}(Y|X)$`最大的输出序列 `$\bar{Y}$`。 + +### 深度双向LSTM（DB-LSTM）SRL模型 + +在SRL任务中，输入是 “谓词” 和 “一句话”，目标是从这句话中找到谓词的论元，并标注论元的语义角色。如果一个句子含有`$n$`个谓词，这个句子会被处理`$n$`次。一个最为直接的模型是下面这样： + +1. 构造输入； +- 输入1是谓词，输入2是句子 +- 将输入1扩展成和输入2一样长的序列，用one-hot方式表示； +2. one-hot方式的谓词序列和句子序列通过词表，转换为实向量表示的词向量序列； +3. 将步骤2中的2个词向量序列作为双向LSTM的输入，学习输入序列的特征表示； +4. CRF以步骤3中模型学习到的特征为输入，以标记序列为监督信号，实现序列标注； + +大家可以尝试上面这种方法。这里，我们提出一些改进，引入两个简单但对提高系统性能非常有效的特征： + +- 谓词上下文：上面的方法中，只用到了谓词的词向量表达谓词相关的所有信息，这种方法始终是非常弱的，特别是如果谓词在句子中出现多次，有可能引起一定的歧义。从经验出发，谓词前后若干个词的一个小片段，能够提供更丰富的信息，帮助消解歧义。于是，我们把这样的经验也添加到模型中，为每个谓词同时抽取一个“谓词上下文” 片段，也就是从这个谓词前后各取`$n$`个词构成的一个窗口片段； +- 谓词上下文区域标记：为句子中的每一个词引入一个0-1二值变量，表示它们是否在“谓词上下文”片段中； + +修改后的模型如下（图6是一个深度为4的模型结构示意图）： + +1. 构造输入 +- 输入1是句子序列，输入2是谓词序列，输入3是谓词上下文，从句子中抽取这个谓词前后各`$n$`个词，构成谓词上下文，用one-hot方式表示，输入4是谓词上下文区域标记，标记了句子中每一个词是否在谓词上下文中； +- 将输入2~3均扩展为和输入1一样长的序列； +2. 输入1~4均通过词表取词向量转换为实向量表示的词向量序列；其中输入1、3共享同一个词表，输入2和4各自独有词表； +3. 第2步的4个词向量序列作为双向LSTM模型的输入；LSTM模型学习输入序列的特征表示，得到新的特性表示序列； +4. CRF以第3步中LSTM学习到的特征为输入，以标记序列为监督信号，完成序列标注； + +![db_lstm_network](./image/db_lstm_network.png) +
+图6. SRL任务上的深层双向LSTM模型 +
+ + +## 数据介绍 + +在此教程中，我们选用[CoNLL 2005](http://www.cs.upc.edu/~srlconll/)SRL任务开放出的数据集作为示例。需要特别说明的是，CoNLL 2005 SRL任务的训练数集和开发集在比赛之后并非免费进行公开，目前，能够获取到的只有测试集，包括Wall Street Journal的23节和Brown语料集中的3节。在本教程中，我们以测试集中的WSJ数据为训练集来讲解模型。但是，由于测试集中样本的数量远远不够，如果希望训练一个可用的神经网络SRL系统，请考虑付费获取全量数据。 + +原始数据中同时包括了词性标注、命名实体识别、语法解析树等多种信息。本教程中，我们使用test.wsj文件夹中的数据进行训练和测试，并只会用到words文件夹（文本序列）和props文件夹（标注结果）下的数据。本教程使用的数据目录如下： + +```text +conll05st-release/ +└── test.wsj +├── props # 标注结果 +└── words # 输入文本序列 +``` + +标注信息源自Penn TreeBank\[[7](#参考文献)\]和PropBank\[[8](#参考文献)\]的标注结果。PropBank标注结果的标签和我们在文章一开始示例中使用的标注结果标签不同，但原理是相同的，关于标注结果标签含义的说明，请参考论文\[[9](#参考文献)\]。 + +原始数据需要进行数据预处理才能被PaddlePaddle处理，预处理包括下面几个步骤: + +1. 将文本序列和标记序列其合并到一条记录中； +2. 一个句子如果含有`$n$`个谓词，这个句子会被处理`$n$`次，变成`$n$`条独立的训练样本，每个样本一个不同的谓词； +3. 抽取谓词上下文和构造谓词上下文区域标记； +4. 构造以BIO法表示的标记； +5. 依据词典获取词对应的整数索引。 + + +```python +# import paddle.v2.dataset.conll05 as conll05 +# conll05.corpus_reader函数完成上面第1步和第2步. +# conll05.reader_creator函数完成上面第3步到第5步. +# conll05.test函数可以获取处理之后的每条样本来供PaddlePaddle训练. +``` + +预处理完成之后一条训练样本包含9个特征，分别是：句子序列、谓词、谓词上下文（占 5 列）、谓词上下区域标志、标注序列。下表是一条训练样本的示例。 + +| 句子序列 | 谓词 | 谓词上下文（窗口 = 5） | 谓词上下文区域标记 | 标注序列 | +|---|---|---|---|---| +| A | set | n't been set . × | 0 | B-A1 | +| record | set | n't been set . × | 0 | I-A1 | +| date | set | n't been set . × | 0 | I-A1 | +| has | set | n't been set . × | 0 | O | +| n't | set | n't been set . × | 1 | B-AM-NEG | +| been | set | n't been set . × | 1 | O | +| set | set | n't been set . × | 1 | B-V | +| . | set | n't been set . × | 1 | O | + + +除数据之外，我们同时提供了以下资源： + +| 文件名称 | 说明 | +|---|---| +| word_dict | 输入句子的词典，共计44068个词 | +| label_dict | 标记的词典，共计106个标记 | +| predicate_dict | 谓词的词典，共计3162个词 | +| emb | 一个训练好的词表，32维 | + +我们在英文维基百科上训练语言模型得到了一份词向量用来初始化SRL模型。在SRL模型训练过程中，词向量不再被更新。关于语言模型和词向量可以参考[词向量](https://github.com/PaddlePaddle/book/blob/develop/04.word2vec/README.cn.md) 这篇教程。我们训练语言模型的语料共有995,000,000个token，词典大小控制为4900,000词。CoNLL 2005训练语料中有5%的词不在这4900,000个词中，我们将它们全部看作未登录词，用``表示。 + +获取词典，打印词典大小： + +```python +import math, os +import numpy as np +import paddle +import paddle.v2.dataset.conll05 as conll05 +import paddle.fluid as fluid +import time + +with_gpu = os.getenv('WITH_GPU', '0') != '0' + +word_dict, verb_dict, label_dict = conll05.get_dict() +word_dict_len = len(word_dict) +label_dict_len = len(label_dict) +pred_dict_len = len(verb_dict) + +print word_dict_len +print label_dict_len +print pred_dict_len +``` + +## 模型配置说明 + +- 定义输入数据维度及模型超参数。 + +```python +mark_dict_len = 2 # 谓上下文区域标志的维度，是一个0-1 2值特征，因此维度为2 +word_dim = 32 # 词向量维度 +mark_dim = 5 # 谓词上下文区域通过词表被映射为一个实向量，这个是相邻的维度 +hidden_dim = 512 # LSTM隐层向量的维度 ： 512 / 4 +depth = 8 # 栈式LSTM的深度 +mix_hidden_lr = 1e-3 + +IS_SPARSE = True +PASS_NUM = 10 +BATCH_SIZE = 10 + +embedding_name = 'emb' +``` + +这里需要特别说明的是hidden_dim = 512指定了LSTM隐层向量的维度为128维，关于这一点请参考PaddlePaddle官方文档中[lstmemory](http://www.paddlepaddle.org/doc/ui/api/trainer_config_helpers/layers.html#lstmemory)的说明。 + +- 如上文提到，我们用基于英文维基百科训练好的词向量来初始化序列输入、谓词上下文总共6个特征的embedding层参数，在训练中不更新。 + +```python +# 这里加载PaddlePaddle上版保存的二进制模型 +def load_parameter(file_name, h, w): +with open(file_name, 'rb') as f: +f.read(16) # skip header. +return np.fromfile(f, dtype=np.float32).reshape(h, w) +``` + +- 8个LSTM单元以“正向/反向”的顺序对所有输入序列进行学习。 + +```python +def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, +**ignored): +# 8 features +predicate_embedding = fluid.layers.embedding( +input=predicate, +size=[pred_dict_len, word_dim], +dtype='float32', +is_sparse=IS_SPARSE, +param_attr='vemb') + +mark_embedding = fluid.layers.embedding( +input=mark, +size=[mark_dict_len, mark_dim], +dtype='float32', +is_sparse=IS_SPARSE) + +word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] +# Since word vector lookup table is pre-trained, we won't update it this time. +# trainable being False prevents updating the lookup table during training. +emb_layers = [ +fluid.layers.embedding( +size=[word_dict_len, word_dim], +input=x, +param_attr=fluid.ParamAttr( +name=embedding_name, trainable=False)) for x in word_input +] +emb_layers.append(predicate_embedding) +emb_layers.append(mark_embedding) + +# 8 LSTM units are trained through alternating left-to-right / right-to-left order +# denoted by the variable `reverse`. +hidden_0_layers = [ +fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') +for emb in emb_layers +] + +hidden_0 = fluid.layers.sums(input=hidden_0_layers) + +lstm_0 = fluid.layers.dynamic_lstm( +input=hidden_0, +size=hidden_dim, +candidate_activation='relu', +gate_activation='sigmoid', +cell_activation='sigmoid') + +# stack L-LSTM and R-LSTM with direct edges +input_tmp = [hidden_0, lstm_0] + +# In PaddlePaddle, state features and transition features of a CRF are implemented +# by a fully connected layer and a CRF layer seperately. The fully connected layer +# with linear activation learns the state features, here we use fluid.layers.sums +# (fluid.layers.fc can be uesed as well), and the CRF layer in PaddlePaddle: +# fluid.layers.linear_chain_crf only +# learns the transition features, which is a cost layer and is the last layer of the network. +# fluid.layers.linear_chain_crf outputs the log probability of true tag sequence +# as the cost by given the input sequence and it requires the true tag sequence +# as target in the learning process. + +for i in range(1, depth): +mix_hidden = fluid.layers.sums(input=[ +fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), +fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') +]) + +lstm = fluid.layers.dynamic_lstm( +input=mix_hidden, +size=hidden_dim, +candidate_activation='relu', +gate_activation='sigmoid', +cell_activation='sigmoid', +is_reverse=((i % 2) == 1)) + +input_tmp = [mix_hidden, lstm] + +# 取最后一个栈式LSTM的输出和这个LSTM单元的输入到隐层映射， +# 经过一个全连接层映射到标记字典的维度，来学习 CRF 的状态特征 +feature_out = fluid.layers.sums(input=[ +fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), +fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') +]) + +return feature_out +``` + +## 训练模型 + +- 我们根据网络拓扑结构和模型参数来构造出trainer用来训练，在构造时还需指定优化方法，这里使用最基本的SGD方法(momentum设置为0)，同时设定了学习率、正则等。 + +- 数据介绍部分提到CoNLL 2005训练集付费，这里我们使用测试集训练供大家学习。conll05.test()每次产生一条样本，包含9个特征，shuffle和组完batch后作为训练的输入。 + +- 通过feeding来指定每一个数据和data_layer的对应关系。 例如 下面feeding表示: conll05.test()产生数据的第0列对应word_data层的特征。 + +- 可以使用event_handler回调函数来观察训练过程，或进行测试等。这里我们打印了训练过程的cost，该回调函数是trainer.train函数里设定。 + +- 通过trainer.train函数训练 + +```python +def train(use_cuda, save_dirname=None, is_local=True): +# define network topology + +# 句子序列 +word = fluid.layers.data( +name='word_data', shape=[1], dtype='int64', lod_level=1) + +# 谓词 +predicate = fluid.layers.data( +name='verb_data', shape=[1], dtype='int64', lod_level=1) + +# 谓词上下文5个特征 +ctx_n2 = fluid.layers.data( +name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) +ctx_n1 = fluid.layers.data( +name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) +ctx_0 = fluid.layers.data( +name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) +ctx_p1 = fluid.layers.data( +name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) +ctx_p2 = fluid.layers.data( +name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) + +# 谓词上下区域标志 +mark = fluid.layers.data( +name='mark_data', shape=[1], dtype='int64', lod_level=1) + +# define network topology +feature_out = db_lstm(**locals()) + +# 标注序列 +target = fluid.layers.data( +name='target', shape=[1], dtype='int64', lod_level=1) + +# 学习 CRF 的转移特征 +crf_cost = fluid.layers.linear_chain_crf( +input=feature_out, +label=target, +param_attr=fluid.ParamAttr( +name='crfw', learning_rate=mix_hidden_lr)) + +avg_cost = fluid.layers.mean(crf_cost) + +sgd_optimizer = fluid.optimizer.SGD( +learning_rate=fluid.layers.exponential_decay( +learning_rate=0.01, +decay_steps=100000, +decay_rate=0.5, +staircase=True)) + +sgd_optimizer.minimize(avg_cost) + +# The CRF decoding layer is used for evaluation and inference. +# It shares weights with CRF layer. The sharing of parameters among multiple layers +# is specified by using the same parameter name in these layers. If true tag sequence +# is provided in training process, `fluid.layers.crf_decoding` calculates labelling error +# for each input token and sums the error over the entire sequence. +# Otherwise, `fluid.layers.crf_decoding` generates the labelling tags. +crf_decode = fluid.layers.crf_decoding( +input=feature_out, param_attr=fluid.ParamAttr(name='crfw')) + +train_data = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.conll05.test(), buf_size=8192), +batch_size=BATCH_SIZE) + +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + +feeder = fluid.DataFeeder( +feed_list=[ +word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, mark, target +], +place=place) +exe = fluid.Executor(place) + +def train_loop(main_program): +exe.run(fluid.default_startup_program()) +embedding_param = fluid.global_scope().find_var( +embedding_name).get_tensor() +embedding_param.set( +load_parameter(conll05.get_embedding(), word_dict_len, word_dim), +place) + +start_time = time.time() +batch_id = 0 +for pass_id in xrange(PASS_NUM): +for data in train_data(): +cost = exe.run(main_program, +feed=feeder.feed(data), +fetch_list=[avg_cost]) +cost = cost[0] + +if batch_id % 10 == 0: +print("avg_cost:" + str(cost)) +if batch_id != 0: +print("second per batch: " + str((time.time( +) - start_time) / batch_id)) +# Set the threshold low to speed up the CI test +if float(cost) < 60.0: +if save_dirname is not None: +fluid.io.save_inference_model(save_dirname, [ +'word_data', 'verb_data', 'ctx_n2_data', +'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data', +'ctx_p2_data', 'mark_data' +], [feature_out], exe) +return + +batch_id = batch_id + 1 + +train_loop(fluid.default_main_program()) +``` + + +## 应用模型 + +训练完成之后，需要依据某个我们关心的性能指标选择最优的模型进行预测，可以简单的选择测试集上标记错误最少的那个模型。以下我们给出一个使用训练后的模型进行预测的示例。 + +```python +def infer(use_cuda, save_dirname=None): +if save_dirname is None: +return + +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +exe = fluid.Executor(place) + +inference_scope = fluid.core.Scope() +with fluid.scope_guard(inference_scope): +# Use fluid.io.load_inference_model to obtain the inference program desc, +# the feed_target_names (the names of variables that will be fed +# data using feed operators), and the fetch_targets (variables that +# we want to obtain data from using fetch operators). +[inference_program, feed_target_names, +fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + +# Setup inputs by creating LoDTensors to represent sequences of words. +# Here each word is the basic element of these LoDTensors and the shape of +# each word (base_shape) should be [1] since it is simply an index to +# look up for the corresponding word vector. +# Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], +# which has only one lod level. Then the created LoDTensors will have only +# one higher level structure (sequence of words, or sentence) than the basic +# element (word). Hence the LoDTensor will hold data for three sentences of +# length 3, 4 and 2, respectively. +# Note that lod info should be a list of lists. +lod = [[3, 4, 2]] +base_shape = [1] +# The range of random integers is [low, high] +word = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=word_dict_len - 1) +pred = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=pred_dict_len - 1) +ctx_n2 = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=word_dict_len - 1) +ctx_n1 = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=word_dict_len - 1) +ctx_0 = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=word_dict_len - 1) +ctx_p1 = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=word_dict_len - 1) +ctx_p2 = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=word_dict_len - 1) +mark = fluid.create_random_int_lodtensor( +lod, base_shape, place, low=0, high=mark_dict_len - 1) + +# Construct feed as a dictionary of {feed_target_name: feed_target_data} +# and results will contain a list of data corresponding to fetch_targets. +assert feed_target_names[0] == 'word_data' +assert feed_target_names[1] == 'verb_data' +assert feed_target_names[2] == 'ctx_n2_data' +assert feed_target_names[3] == 'ctx_n1_data' +assert feed_target_names[4] == 'ctx_0_data' +assert feed_target_names[5] == 'ctx_p1_data' +assert feed_target_names[6] == 'ctx_p2_data' +assert feed_target_names[7] == 'mark_data' + +results = exe.run(inference_program, +feed={ +feed_target_names[0]: word, +feed_target_names[1]: pred, +feed_target_names[2]: ctx_n2, +feed_target_names[3]: ctx_n1, +feed_target_names[4]: ctx_0, +feed_target_names[5]: ctx_p1, +feed_target_names[6]: ctx_p2, +feed_target_names[7]: mark +}, +fetch_list=fetch_targets, +return_numpy=False) +print(results[0].lod()) +np_data = np.array(results[0]) +print("Inference Shape: ", np_data.shape) +``` + +整个程序的入口如下： + +```python +def main(use_cuda, is_local=True): +if use_cuda and not fluid.core.is_compiled_with_cuda(): +return + +# Directory for saving the trained model +save_dirname = "label_semantic_roles.inference.model" + +train(use_cuda, save_dirname, is_local) +infer(use_cuda, save_dirname) + + +main(use_cuda=False) +``` + +## 总结 + +语义角色标注是许多自然语言理解任务的重要中间步骤。这篇教程中我们以语义角色标注任务为例，介绍如何利用PaddlePaddle进行序列标注任务。教程中所介绍的模型来自我们发表的论文\[[10](#参考文献)\]。由于 CoNLL 2005 SRL任务的训练数据目前并非完全开放，教程中只使用测试数据作为示例。在这个过程中，我们希望减少对其它自然语言处理工具的依赖，利用神经网络数据驱动、端到端学习的能力，得到一个和传统方法可比、甚至更好的模型。在论文中我们证实了这种可能性。关于模型更多的信息和讨论可以在论文中找到。 + +## 参考文献 +1. Sun W, Sui Z, Wang M, et al. [Chinese semantic role labeling with shallow parsing](http://www.aclweb.org/anthology/D09-1#page=1513)[C]//Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 3-Volume 3. Association for Computational Linguistics, 2009: 1475-1483. +2. Pascanu R, Gulcehre C, Cho K, et al. [How to construct deep recurrent neural networks](https://arxiv.org/abs/1312.6026)[J]. arXiv preprint arXiv:1312.6026, 2013. +3. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](https://arxiv.org/abs/1406.1078)[J]. arXiv preprint arXiv:1406.1078, 2014. +4. Bahdanau D, Cho K, Bengio Y. [Neural machine translation by jointly learning to align and translate](https://arxiv.org/abs/1409.0473)[J]. arXiv preprint arXiv:1409.0473, 2014. +5. Lafferty J, McCallum A, Pereira F. [Conditional random fields: Probabilistic models for segmenting and labeling sequence data](http://www.jmlr.org/papers/volume15/doppa14a/source/biblio.bib.old)[C]//Proceedings of the eighteenth international conference on machine learning, ICML. 2001, 1: 282-289. +6. 李航. 统计学习方法[J]. 清华大学出版社, 北京, 2012. +7. Marcus M P, Marcinkiewicz M A, Santorini B. [Building a large annotated corpus of English: The Penn Treebank](http://repository.upenn.edu/cgi/viewcontent.cgi?article=1246&context=cis_reports)[J]. Computational linguistics, 1993, 19(2): 313-330. +8. Palmer M, Gildea D, Kingsbury P. [The proposition bank: An annotated corpus of semantic roles](http://www.mitpressjournals.org/doi/pdfplus/10.1162/0891201053630264)[J]. Computational linguistics, 2005, 31(1): 71-106. +9. Carreras X, Màrquez L. [Introduction to the CoNLL-2005 shared task: Semantic role labeling](http://www.cs.upc.edu/~srlconll/st05/papers/intro.pdf)[C]//Proceedings of the Ninth Conference on Computational Natural Language Learning. Association for Computational Linguistics, 2005: 152-164. +10. Zhou J, Xu W. [End-to-end learning of semantic role labeling using recurrent neural networks](http://www.aclweb.org/anthology/P/P15/P15-1109.pdf)[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015. + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/basics/learning_materials.md b/doc/fluid/new_docs/beginners_guide/basics/learning_materials.md new file mode 100644 index 0000000000000000000000000000000000000000..a27499c6ed8d1149c6d519006086febbcae943fa --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/learning_materials.md @@ -0,0 +1,54 @@ +# 学习资料 + +## 要读的第一本书 +基础理论习得的最直接来源就是书本。按机器学习理论、深度学习理论、编程语言三方面划分，这里推荐如下书籍辅助您。 + + +### 机器学习理论 + +在开启深度学习之前，您需要先行掌握机器学习的理论。深度学习是机器学习中的一个分支，两者内在的理论基础存在强关联。 +机器学习理论的书籍教材比较多，这里推荐一本易懂易学的书籍，可以重点关注神经网络部分。 + +书名：《机器学习》（周志华著，清华大学出版社，2016年版） + +### 深度学习理论 + +打好机器学习的理论功底后，您可以开始钻研深度学习的理论。通常深度学习理论会给人留下抽象难懂的印象，且和数学结合紧密。 +为了让您能够顺利入门，这里推荐一份易学易用的教材，无论深度学习理论还是数学理论即可一本搞定。 + +书名：《Deep Learning（深度学习）》（Goodfellow, Bengio, Courville合著，赵申剑、黎彧君、符天凡和李凯合译，人民邮电出版社，2017年版） +此书电子版在Github上已经开源，详情可参考此链接 [《深度学习》](https://github.com/exacity/deeplearningbook-chinese) + +### 编程语言 + +Python方向：这里推荐您学习Python，一方面各大主流深度学习框架的主力支撑编程语言均为Python；另一方面，对比其他语言，Python较为简单易学。 +Python的教材种类较多，这里推荐一本实操和理论性都兼顾的教材，只要完成书中52个习题，跑代码然后发现问题解决，就能逐步上手。 + +书名：《“笨办法”学Python》（Zed Shaw著，王巍巍译，人民邮电出版社，2014年11月版） + + +C++方向：C++语言在底层框架中使用较多，您逐步掌握开源框架的基本操作后，在更高阶的框架应用中会用到这个技能点。 +同前面提到的Python一样，学习C++时需要多上手操作。这里推荐迅速上手C++的书籍，不但能够学习功能和结构，还提供了解决方案的示例。 + +书名：《Essential C++》【美】李普曼（Lippman,S.B.）著，侯捷译，电子工业出版社2013年8月版 + + + +## 要看的视频公开课 + +在学习一门新技术的同时，除了看书，如果有老师面对面教授，可以更快更好的学会知识。相比于线下授课，视频公开课能够在省钱省力的同时，达到易学易掌握的效果。 +目前深度学习的课程多是公开免费的，通过学习您可以更轻松的理解深度学习中的抽象理论，并在实操方面不绕弯路。 +综合课程生动性、可操作性、紧凑性、连续性这些特点，这里推荐如下课程，同步附上网址，便于您查找学习。 + +### 理论知识详解视频课 +[机器学习](http://open.163.com/special/opencourse/machinelearning.html) 斯坦福大学教授吴恩达公开课程，包含相关算法的详细讲解。 + +[AI技术](https://ai.baidu.com/paddlepaddle/player?id=13) 百度推出的“AI核心技术掌握”课程，每节课在20-30分钟左右，从AI技术到深度学习进行全面细致的解读。 + +[深度学习](http://speech.ee.ntu.edu.tw/~tlkagk/courses_ML17_2.html) 台湾李宏毅教授的在线课程，其中是英文课程，会结合国外的科研成果，但也适合新手入门和理解深度学习。 + +[编程语言](https://ai.baidu.com/paddlepaddle/openCourses) Python操作课程，从基础到进阶操作都提供详细说明，每节课时长20分钟左右。 + +### PaddlePaddle实操视频课 +掌握好理论基础，具备编程能力后，您可以开始使用PaddlePaddle Fluid进行实操，从初阶开始学习，向着中高阶努力。 +目前已有PaddlePaddle官方视频公开课在官网呈现,内含PaddlePaddle实战、PaddlePaddle应用场景和机器学习模型讲解课程，帮助开发者从零开始使用PaddlePaddle，从简单场景逐步过渡到工业级应用。[点击这里](http://ai.baidu.com/paddlepaddle/openCourses)您即可开始视频课的学习之旅。 diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/.gitignore b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..6129b9e8645010fcb8372d9dc3dbb568dfa80907 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/.gitignore @@ -0,0 +1,9 @@ +data/wmt14 +data/pre-wmt14 +pretrained/wmt14_model +gen.log +gen_result +train.log +dataprovider_copy_1.py +*.pyc +multi-bleu.perl diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn.png new file mode 100644 index 0000000000000000000000000000000000000000..9d8efd50a49d0305586f550344472ab94c93bed3 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn_en.png new file mode 100755 index 0000000000000000000000000000000000000000..4b35c88fc8ea2c503473c0c15711744e784d6af6 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/bi_rnn_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention.png new file mode 100644 index 0000000000000000000000000000000000000000..1b355e7786d25487a3f564af758c2c52c43b4690 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention_en.png new file mode 100755 index 0000000000000000000000000000000000000000..3728f782ee09d9308d02b42305027b2735467ead Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/decoder_attention_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention.png new file mode 100644 index 0000000000000000000000000000000000000000..28d7a15a3bd65262bde22a3f41b5aa78b46b368a Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention_en.png new file mode 100755 index 0000000000000000000000000000000000000000..ea8585565da1ecaf241654c278c6f9b15e283286 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_attention_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder.png new file mode 100755 index 0000000000000000000000000000000000000000..60aee0017de73f462e35708b1055aff8992c03e1 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder_en.png new file mode 100755 index 0000000000000000000000000000000000000000..6b73798fe632e0873b35c117b86f347c8cf3116a Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/encoder_decoder_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru.png new file mode 100644 index 0000000000000000000000000000000000000000..0cde685b84106650a4df18ce335a23e6338d3d11 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru_en.png new file mode 100755 index 0000000000000000000000000000000000000000..a6af429f23f0f7e82650139bbd8dcbef27a34abe Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/gru_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt.png new file mode 100644 index 0000000000000000000000000000000000000000..bf56d73ebf297fadf522389c7b6836dd379aa097 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt_en.png b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt_en.png new file mode 100755 index 0000000000000000000000000000000000000000..557310e044b2b6687e5ea6895417ed946ac7bc11 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/image/nmt_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/machine_translation/index.md b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/index.md new file mode 100644 index 0000000000000000000000000000000000000000..fc161aaae9c37b0e1a596204e7138025a98adb1d --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/machine_translation/index.md @@ -0,0 +1,448 @@ +# 机器翻译 + +本教程源代码目录在[book/machine_translation](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 + +机器翻译（machine translation, MT）是用计算机来实现不同语言之间翻译的技术。被翻译的语言通常称为源语言（source language），翻译成的结果语言称为目标语言（target language）。机器翻译即实现从源语言到目标语言转换的过程，是自然语言处理的重要研究领域之一。 + +早期机器翻译系统多为基于规则的翻译系统，需要由语言学家编写两种语言之间的转换规则，再将这些规则录入计算机。该方法对语言学家的要求非常高，而且我们几乎无法总结一门语言会用到的所有规则，更何况两种甚至更多的语言。因此，传统机器翻译方法面临的主要挑战是无法得到一个完备的规则集合\[[1](#参考文献)\]。 + +为解决以上问题，统计机器翻译（Statistical Machine Translation, SMT）技术应运而生。在统计机器翻译技术中，转化规则是由机器自动从大规模的语料中学习得到的，而非我们人主动提供规则。因此，它克服了基于规则的翻译系统所面临的知识获取瓶颈的问题，但仍然存在许多挑战：1）人为设计许多特征（feature），但永远无法覆盖所有的语言现象；2）难以利用全局的特征；3）依赖于许多预处理环节，如词语对齐、分词或符号化（tokenization）、规则抽取、句法分析等，而每个环节的错误会逐步累积，对翻译的影响也越来越大。 + +近年来，深度学习技术的发展为解决上述挑战提供了新的思路。将深度学习应用于机器翻译任务的方法大致分为两类：1）仍以统计机器翻译系统为框架，只是利用神经网络来改进其中的关键模块，如语言模型、调序模型等（见图1的左半部分）；2）不再以统计机器翻译系统为框架，而是直接用神经网络将源语言映射到目标语言，即端到端的神经网络机器翻译（End-to-End Neural Machine Translation, End-to-End NMT）（见图1的右半部分），简称为NMT模型。 +![nmt](./image/nmt.png) +

+图1. 基于神经网络的机器翻译系统 +

+ +本教程主要介绍NMT模型，以及如何用PaddlePaddle来训练一个NMT模型。 + +## 效果展示 + +以中英翻译（中文翻译到英文）的模型为例，当模型训练完毕时，如果输入如下已分词的中文句子： +```text +这些 是 希望 的 曙光 和 解脱 的 迹象 . +``` +如果设定显示翻译结果的条数（即[柱搜索算法](#柱搜索算法)的宽度）为3，生成的英语句子如下： +```text +0 -5.36816 These are signs of hope and relief . +1 -6.23177 These are the light of hope and relief . +2 -7.7914 These are the light of hope and the relief of hope . +``` +- 左起第一列是生成句子的序号；左起第二列是该条句子的得分（从大到小），分值越高越好；左起第三列是生成的英语句子。 +- 另外有两个特殊标志：``表示句子的结尾，``表示未登录词（unknown word），即未在训练字典中出现的词。 + +## 模型概览 + +本节依次介绍双向循环神经网络（Bi-directional Recurrent Neural Network），NMT模型中典型的编码器-解码器（Encoder-Decoder）框架以及柱搜索（beam search）算法。 + +### 双向循环神经网络 + +我们已经在[语义角色标注](https://github.com/PaddlePaddle/book/blob/develop/07.label_semantic_roles/README.cn.md)一章中介绍了一种双向循环神经网络，这里介绍Bengio团队在论文\[[2](#参考文献),[4](#参考文献)\]中提出的另一种结构。该结构的目的是输入一个序列，得到其在每个时刻的特征表示，即输出的每个时刻都用定长向量表示到该时刻的上下文语义信息。 + +具体来说，该双向循环神经网络分别在时间维以顺序和逆序——即前向（forward）和后向（backward）——依次处理输入序列，并将每个时间步RNN的输出拼接成为最终的输出层。这样每个时间步的输出节点，都包含了输入序列中当前时刻完整的过去和未来的上下文信息。下图展示的是一个按时间步展开的双向循环神经网络。该网络包含一个前向和一个后向RNN，其中有六个权重矩阵：输入到前向隐层和后向隐层的权重矩阵（`$W_1, W_3$`），隐层到隐层自己的权重矩阵（`$W_2,W_5$`），前向隐层和后向隐层到输出层的权重矩阵（`$W_4, W_6$`）。注意，该网络的前向隐层和后向隐层之间没有连接。 + +![bi_rnn](./image/bi_rnn.png) +

+图3. 按时间步展开的双向循环神经网络 +

+ +### 编码器-解码器框架 + +编码器-解码器（Encoder-Decoder）\[[2](#参考文献)\]框架用于解决由一个任意长度的源序列到另一个任意长度的目标序列的变换问题。即编码阶段将整个源序列编码成一个向量，解码阶段通过最大化预测序列概率，从中解码出整个目标序列。编码和解码的过程通常都使用RNN实现。 +![encoder_decoder](./image/encoder_decoder.png) +

+图4. 编码器-解码器框架 +

+ +#### 编码器 + +编码阶段分为三步： + +1. one-hot vector表示：将源语言句子`$x=\left \{ x_1,x_2,...,x_T \right \}$`的每个词`$x_i$`表示成一个列向量`$w_i\epsilon \left \{ 0,1 \right \}^{\left | V \right |},i=1,2,...,T$`。这个向量`$w_i$`的维度与词汇表大小`$\left | V \right |$` 相同，并且只有一个维度上有值1（该位置对应该词在词汇表中的位置），其余全是0。 + +2. 映射到低维语义空间的词向量：one-hot vector表示存在两个问题，1）生成的向量维度往往很大，容易造成维数灾难；2）难以刻画词与词之间的关系（如语义相似性，也就是无法很好地表达语义）。因此，需再one-hot vector映射到低维的语义空间，由一个固定维度的稠密向量（称为词向量）表示。记映射矩阵为`$C\epsilon R^{K\times \left | V \right |}$`，用`$s_i=Cw_i$`表示第`$i$`个词的词向量，`$K$`为向量维度。 + +3. 用RNN编码源语言词序列：这一过程的计算公式为`$h_i=\varnothing _\theta \left ( h_{i-1}, s_i \right )$`，其中`$h_0$`是一个全零的向量，`$\varnothing _\theta$`是一个非线性激活函数，最后得到的`$\mathbf{h}=\left \{ h_1,..., h_T \right \}$`就是RNN依次读入源语言`$T$`个词的状态编码序列。整句话的向量表示可以采用`$\mathbf{h}$`在最后一个时间步`$T$`的状态编码，或使用时间维上的池化（pooling）结果。 + +第3步也可以使用双向循环神经网络实现更复杂的句编码表示，具体可以用双向GRU实现。前向GRU按照词序列`$(x_1,x_2,...,x_T)$`的顺序依次编码源语言端词，并得到一系列隐层状态`$(\overrightarrow{h_1},\overrightarrow{h_2},...,\overrightarrow{h_T})$`。类似的，后向GRU按照`$(x_T,x_{T-1},...,x_1)$`的顺序依次编码源语言端词，得到`$(\overleftarrow{h_1},\overleftarrow{h_2},...,\overleftarrow{h_T})$`。最后对于词`$x_i$`，通过拼接两个GRU的结果得到它的隐层状态，即`$h_i=\left [ \overrightarrow{h_i^T},\overleftarrow{h_i^T} \right ]^{T}$`。 + +![encoder_attention](./image/encoder_attention.png) +

+图5. 使用双向GRU的编码器 +

+ +#### 解码器 + +机器翻译任务的训练过程中，解码阶段的目标是最大化下一个正确的目标语言词的概率。思路是： + +1. 每一个时刻，根据源语言句子的编码信息（又叫上下文向量，context vector）`$c$`、真实目标语言序列的第`$i$`个词`$u_i$`和`$i$`时刻RNN的隐层状态`$z_i$`，计算出下一个隐层状态`$z_{i+1}$`。计算公式如下： + +$$z_{i+1}=\phi _{\theta '}\left ( c,u_i,z_i \right )$$ + +其中`$\phi _{\theta '}$`是一个非线性激活函数；`$c=q\mathbf{h}$`是源语言句子的上下文向量，在不使用[注意力机制](#注意力机制)时，如果[编码器](#编码器)的输出是源语言句子编码后的最后一个元素，则可以定义`$c=h_T$`；`$u_i$`是目标语言序列的第`$i$`个单词，`$u_0$`是目标语言序列的开始标记``，表示解码开始；`$z_i$`是`$i$`时刻解码RNN的隐层状态，`$z_0$`是一个全零的向量。 + +2. 将`$z_{i+1}$`通过`softmax`归一化，得到目标语言序列的第`$i+1$`个单词的概率分布`$p_{i+1}$`。概率分布公式如下： + +$$p\left ( u_{i+1}|u_{<i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$ + +其中`$W_sz_{i+1}+b_z$`是对每个可能的输出单词进行打分，再用softmax归一化就可以得到第`$i+1$`个词的概率`$p_{i+1}$`。 + +3. 根据`$p_{i+1}$`和`$u_{i+1}$`计算代价。 +4. 重复步骤1~3，直到目标语言序列中的所有词处理完毕。 + +机器翻译任务的生成过程，通俗来讲就是根据预先训练的模型来翻译源语言句子。生成过程中的解码阶段和上述训练过程的有所差异，具体介绍请见[柱搜索算法](#柱搜索算法)。 + +### 柱搜索算法 + +柱搜索（[beam search](http://en.wikipedia.org/wiki/Beam_search)）是一种启发式图搜索算法，用于在图或树中搜索有限集合中的最优扩展节点，通常用在解空间非常大的系统（如机器翻译、语音识别）中，原因是内存无法装下图或树中所有展开的解。如在机器翻译任务中希望翻译“`你好`”，就算目标语言字典中只有3个词（``, ``, `hello`），也可能生成无限句话（`hello`循环出现的次数不定），为了找到其中较好的翻译结果，我们可采用柱搜索算法。 + +柱搜索算法使用广度优先策略建立搜索树，在树的每一层，按照启发代价（heuristic cost）（本教程中，为生成词的log概率之和）对节点进行排序，然后仅留下预先确定的个数（文献中通常称为beam width、beam size、柱宽度等）的节点。只有这些节点会在下一层继续扩展，其他节点就被剪掉了，也就是说保留了质量较高的节点，剪枝了质量较差的节点。因此，搜索所占用的空间和时间大幅减少，但缺点是无法保证一定获得最优解。 + +使用柱搜索算法的解码阶段，目标是最大化生成序列的概率。思路是： + +1. 每一个时刻，根据源语言句子的编码信息`$c$`、生成的第`$i$`个目标语言序列单词`$u_i$`和`$i$`时刻RNN的隐层状态`$z_i$`，计算出下一个隐层状态`$z_{i+1}$`。 +2. 将`$z_{i+1}$`通过`softmax`归一化，得到目标语言序列的第`$i+1$`个单词的概率分布`$p_{i+1}$`。 +3. 根据`$p_{i+1}$`采样出单词`$u_{i+1}$`。 +4. 重复步骤1~3，直到获得句子结束标记``或超过句子的最大生成长度为止。 + +注意：`$z_{i+1}$`和`$p_{i+1}$`的计算公式同[解码器](#解码器)中的一样。且由于生成时的每一步都是通过贪心法实现的，因此并不能保证得到全局最优解。 + +## 数据介绍 + +本教程使用[WMT-14](http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/)数据集中的[bitexts(after selection)](http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/data/bitexts.tgz)作为训练集，[dev+test data](http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/data/dev+test.tgz)作为测试集和生成集。 + +### 数据预处理 + +我们的预处理流程包括两步： +- 将每个源语言到目标语言的平行语料库文件合并为一个文件： +- 合并每个`XXX.src`和`XXX.trg`文件为`XXX`。 +- `XXX`中的第`$i$`行内容为`XXX.src`中的第`$i$`行和`XXX.trg`中的第`$i$`行连接，用'\t'分隔。 +- 创建训练数据的“源字典”和“目标字典”。每个字典都有**DICTSIZE**个单词，包括：语料中词频最高的（DICTSIZE - 3）个单词，和3个特殊符号``（序列的开始）、``（序列的结束）和``（未登录词）。 + +### 示例数据 + +因为完整的数据集数据量较大，为了验证训练流程，PaddlePaddle接口paddle.dataset.wmt14中默认提供了一个经过预处理的[较小规模的数据集](http://paddlepaddle.bj.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz)。 + +该数据集有193319条训练数据，6003条测试数据，词典长度为30000。因为数据规模限制，使用该数据集训练出来的模型效果无法保证。 + +## 模型配置说明 + +下面我们开始根据输入数据的形式配置模型。首先引入所需的库函数以及定义全局变量。 + +```python +import contextlib + +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.framework as framework +import paddle.fluid.layers as pd +from paddle.fluid.executor import Executor +from functools import partial +import os + +dict_size = 30000 +source_dict_dim = target_dict_dim = dict_size +hidden_dim = 32 +word_dim = 16 +batch_size = 2 +max_length = 8 +topk_size = 50 +beam_size = 2 + +decoder_size = hidden_dim +``` + +然后如下实现编码器框架： + +```python +def encoder(is_sparse): +src_word_id = pd.data( +name="src_word_id", shape=[1], dtype='int64', lod_level=1) +src_embedding = pd.embedding( +input=src_word_id, +size=[dict_size, word_dim], +dtype='float32', +is_sparse=is_sparse, +param_attr=fluid.ParamAttr(name='vemb')) + +fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') +lstm_hidden0, lstm_0 = pd.dynamic_lstm(input=fc1, size=hidden_dim * 4) +encoder_out = pd.sequence_last_step(input=lstm_hidden0) +return encoder_out +``` + +再实现训练模式下的解码器： + +```python +def train_decoder(context, is_sparse): +trg_language_word = pd.data( +name="target_language_word", shape=[1], dtype='int64', lod_level=1) +trg_embedding = pd.embedding( +input=trg_language_word, +size=[dict_size, word_dim], +dtype='float32', +is_sparse=is_sparse, +param_attr=fluid.ParamAttr(name='vemb')) + +rnn = pd.DynamicRNN() +with rnn.block(): +current_word = rnn.step_input(trg_embedding) +pre_state = rnn.memory(init=context) +current_state = pd.fc(input=[current_word, pre_state], +size=decoder_size, +act='tanh') + +current_score = pd.fc(input=current_state, +size=target_dict_dim, +act='softmax') +rnn.update_memory(pre_state, current_state) +rnn.output(current_score) + +return rnn() +``` + +实现推测模式下的解码器： + +```python +def decode(context, is_sparse): +init_state = context +array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length) +counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True) + +# fill the first element with init_state +state_array = pd.create_array('float32') +pd.array_write(init_state, array=state_array, i=counter) + +# ids, scores as memory +ids_array = pd.create_array('int64') +scores_array = pd.create_array('float32') + +init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2) +init_scores = pd.data( +name="init_scores", shape=[1], dtype="float32", lod_level=2) + +pd.array_write(init_ids, array=ids_array, i=counter) +pd.array_write(init_scores, array=scores_array, i=counter) + +cond = pd.less_than(x=counter, y=array_len) + +while_op = pd.While(cond=cond) +with while_op.block(): +pre_ids = pd.array_read(array=ids_array, i=counter) +pre_state = pd.array_read(array=state_array, i=counter) +pre_score = pd.array_read(array=scores_array, i=counter) + +# expand the lod of pre_state to be the same with pre_score +pre_state_expanded = pd.sequence_expand(pre_state, pre_score) + +pre_ids_emb = pd.embedding( +input=pre_ids, +size=[dict_size, word_dim], +dtype='float32', +is_sparse=is_sparse) + +# use rnn unit to update rnn +current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb], +size=decoder_size, +act='tanh') +current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score) +# use score to do beam search +current_score = pd.fc(input=current_state_with_lod, +size=target_dict_dim, +act='softmax') +topk_scores, topk_indices = pd.topk(current_score, k=topk_size) +selected_ids, selected_scores = pd.beam_search( +pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) + +pd.increment(x=counter, value=1, in_place=True) + +# update the memories +pd.array_write(current_state, array=state_array, i=counter) +pd.array_write(selected_ids, array=ids_array, i=counter) +pd.array_write(selected_scores, array=scores_array, i=counter) + +pd.less_than(x=counter, y=array_len, cond=cond) + +translation_ids, translation_scores = pd.beam_search_decode( +ids=ids_array, scores=scores_array) + +return translation_ids, translation_scores +``` + +进而，我们定义一个`train_program`来使用`inference_program`计算出的结果，在标记数据的帮助下来计算误差。我们还定义了一个`optimizer_func`来定义优化器。 + +```python +def train_program(is_sparse): +context = encoder(is_sparse) +rnn_out = train_decoder(context, is_sparse) +label = pd.data( +name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) +cost = pd.cross_entropy(input=rnn_out, label=label) +avg_cost = pd.mean(cost) +return avg_cost + + +def optimizer_func(): +return fluid.optimizer.Adagrad( +learning_rate=1e-4, +regularization=fluid.regularizer.L2DecayRegularizer( +regularization_coeff=0.1)) +``` + +## 训练模型 + +### 定义训练环境 +定义您的训练环境，可以指定训练是发生在CPU还是GPU上。 + +```python +use_cuda = False +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +``` + +### 定义数据提供器 +下一步是为训练和测试定义数据提供器。提供器读入一个大小为 `BATCH_SIZE`的数据。`paddle.dataset.wmt.train` 每次会在乱序化后提供一个大小为`BATCH_SIZE`的数据，乱序化的大小为缓存大小`buf_size`。 + +```python +train_reader = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.wmt14.train(dict_size), buf_size=1000), +batch_size=batch_size) +``` + +### 构造训练器(trainer) +训练器需要一个训练程序和一个训练优化函数。 + +```python +is_sparse = False +trainer = fluid.Trainer( +train_func=partial(train_program, is_sparse), +place=place, +optimizer_func=optimizer_func) +``` + +### 提供数据 + +`feed_order`用来定义每条产生的数据和`paddle.layer.data`之间的映射关系。比如，`wmt14.train`产生的第一列的数据对应的是`src_word_id`这个特征。 + +```python +feed_order = [ +'src_word_id', 'target_language_word', 'target_language_next_word' +] +``` + +### 事件处理器 +回调函数`event_handler`在一个之前定义好的事件发生后会被调用。例如，我们可以在每步训练结束后查看误差。 + +```python +def event_handler(event): +if isinstance(event, fluid.EndStepEvent): +if event.step % 10 == 0: +print('pass_id=' + str(event.epoch) + ' batch=' + str(event.step)) + +if event.step == 20: +trainer.stop() +``` + +### 开始训练 +最后，我们传入训练循环数（`num_epoch`）和一些别的参数，调用 `trainer.train` 来开始训练。 + +```python +EPOCH_NUM = 1 + +trainer.train( +reader=train_reader, +num_epochs=EPOCH_NUM, +event_handler=event_handler, +feed_order=feed_order) +``` + +## 应用模型 + +### 定义解码部分 + +使用上面定义的 `encoder` 和 `decoder` 函数来推测翻译后的对应id和分数. + +```python +context = encoder(is_sparse) +translation_ids, translation_scores = decode(context, is_sparse) +``` + +### 定义数据 + +我们先初始化id和分数来生成tensors来作为输入数据。在这个预测例子中，我们用`wmt14.test`数据中的第一个记录来做推测，最后我们用"源字典"和"目标字典"来列印对应的句子结果。 + +```python +init_ids_data = np.array([1 for _ in range(batch_size)], dtype='int64') +init_scores_data = np.array( +[1. for _ in range(batch_size)], dtype='float32') +init_ids_data = init_ids_data.reshape((batch_size, 1)) +init_scores_data = init_scores_data.reshape((batch_size, 1)) +init_lod = [1] * batch_size +init_lod = [init_lod, init_lod] + +init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) +init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) + +test_data = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.wmt14.test(dict_size), buf_size=1000), +batch_size=batch_size) + +feed_order = ['src_word_id'] +feed_list = [ +framework.default_main_program().global_block().var(var_name) +for var_name in feed_order +] +feeder = fluid.DataFeeder(feed_list, place) + +src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size) +``` + +### 测试 +现在我们可以进行预测了。我们要在`feed_order`提供对应参数，放在`executor`上运行以取得id和分数结果 + +```python +exe = Executor(place) +exe.run(framework.default_startup_program()) + +for data in test_data(): +feed_data = map(lambda x: [x[0]], data) +feed_dict = feeder.feed(feed_data) +feed_dict['init_ids'] = init_ids +feed_dict['init_scores'] = init_scores + +results = exe.run( +framework.default_main_program(), +feed=feed_dict, +fetch_list=[translation_ids, translation_scores], +return_numpy=False) + +result_ids = np.array(results[0]) +result_scores = np.array(results[1]) + +print("Original sentence:") +print(" ".join([src_dict[w] for w in feed_data[0][0]])) +print("Translated sentence:") +print(" ".join([trg_dict[w] for w in result_ids])) +print("Corresponding score: ", result_scores) + +break +``` + +## 总结 + +端到端的神经网络机器翻译是近几年兴起的一种全新的机器翻译方法。本章中，我们介绍了NMT中典型的“编码器-解码器”框架。由于NMT是一个典型的Seq2Seq（Sequence to Sequence，序列到序列）学习问题，因此，Seq2Seq中的query改写（query rewriting）、摘要、单轮对话等问题都可以用本教程的模型来解决。 + +## 参考文献 + +1. Koehn P. [Statistical machine translation](https://books.google.com.hk/books?id=4v_Cx1wIMLkC&printsec=frontcover&hl=zh-CN&source=gbs_ge_summary_r&cad=0#v=onepage&q&f=false)[M]. Cambridge University Press, 2009. +2. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](http://www.aclweb.org/anthology/D/D14/D14-1179.pdf)[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014: 1724-1734. +3. Chung J, Gulcehre C, Cho K H, et al. [Empirical evaluation of gated recurrent neural networks on sequence modeling](https://arxiv.org/abs/1412.3555)[J]. arXiv preprint arXiv:1412.3555, 2014. +4. Bahdanau D, Cho K, Bengio Y. [Neural machine translation by jointly learning to align and translate](https://arxiv.org/abs/1409.0473)[C]//Proceedings of ICLR 2015, 2015. +5. Papineni K, Roukos S, Ward T, et al. [BLEU: a method for automatic evaluation of machine translation](http://dl.acm.org/citation.cfm?id=1073135)[C]//Proceedings of the 40th annual meeting on association for computational linguistics. Association for Computational Linguistics, 2002: 311-318. + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/.gitignore b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..f23901aeb3a9e7cd12611fc556742670d04a9bb5 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/.gitignore @@ -0,0 +1,2 @@ +.idea +.ipynb_checkpoints diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/Deep_candidate_generation_model_architecture.en.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/Deep_candidate_generation_model_architecture.en.png new file mode 100644 index 0000000000000000000000000000000000000000..c213608e769f69fb2cfe8597f8e696ee53730e3d Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/Deep_candidate_generation_model_architecture.en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/Deep_candidate_generation_model_architecture.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/Deep_candidate_generation_model_architecture.png new file mode 100644 index 0000000000000000000000000000000000000000..8aedb2204371e7691140ceffa5992f6080bbf097 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/Deep_candidate_generation_model_architecture.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/YouTube_Overview.en.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/YouTube_Overview.en.png new file mode 100644 index 0000000000000000000000000000000000000000..4298567ac5600173343299999965b20612e7affe Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/YouTube_Overview.en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/YouTube_Overview.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/YouTube_Overview.png new file mode 100644 index 0000000000000000000000000000000000000000..a98e7cc67606b31e4c945f7eb907563e46dcef56 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/YouTube_Overview.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/output_32_0.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/output_32_0.png new file mode 100644 index 0000000000000000000000000000000000000000..7fd97b9cc3a0b9105b41591af4e8f8e4646bd681 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/output_32_0.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/rec_regression_network.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/rec_regression_network.png new file mode 100644 index 0000000000000000000000000000000000000000..90c9b09fb78db98391ee199934f2d16efd6d6652 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/rec_regression_network.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/rec_regression_network_en.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/rec_regression_network_en.png new file mode 100755 index 0000000000000000000000000000000000000000..6fc8e11967000ec48c1c0a6fa3c2eaecb80cbb84 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/rec_regression_network_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/text_cnn.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/text_cnn.png new file mode 100644 index 0000000000000000000000000000000000000000..61e63d9147cbc2901706ef80776d706e5368c3c5 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/text_cnn.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/text_cnn_en.png b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/text_cnn_en.png new file mode 100644 index 0000000000000000000000000000000000000000..fbcae2be81141be955076e877b94b0ea5d7e4d4a Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/image/text_cnn_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/recommender_system/index.md b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/index.md new file mode 100644 index 0000000000000000000000000000000000000000..09a07f3dc30abc57ab3731af054dd83491acc9a6 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/recommender_system/index.md @@ -0,0 +1,528 @@ +# 个性化推荐 + +本教程源代码目录在[book/recommender_system](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 + +在网络技术不断发展和电子商务规模不断扩大的背景下，商品数量和种类快速增长，用户需要花费大量时间才能找到自己想买的商品，这就是信息超载问题。为了解决这个难题，推荐系统（Recommender System）应运而生。 + +个性化推荐系统是信息过滤系统（Information Filtering System）的子集，它可以用在很多领域，如电影、音乐、电商和 Feed 流推荐等。推荐系统通过分析、挖掘用户行为，发现用户的个性化需求与兴趣特点，将用户可能感兴趣的信息或商品推荐给用户。与搜索引擎不同，推荐系统不需要用户准确地描述出自己的需求，而是根据分析历史行为建模，主动提供满足用户兴趣和需求的信息。 + +传统的推荐系统方法主要有： + +- 协同过滤推荐（Collaborative Filtering Recommendation）：该方法收集分析用户历史行为、活动、偏好，计算一个用户与其他用户的相似度，利用目标用户的相似用户对商品评价的加权评价值，来预测目标用户对特定商品的喜好程度。优点是可以给用户推荐未浏览过的新产品；缺点是对于没有任何行为的新用户存在冷启动的问题，同时也存在用户与商品之间的交互数据不够多造成的稀疏问题，会导致模型难以找到相近用户。 +- 基于内容过滤推荐[[1](#参考文献)]（Content-based Filtering Recommendation）：该方法利用商品的内容描述，抽象出有意义的特征，通过计算用户的兴趣和商品描述之间的相似度，来给用户做推荐。优点是简单直接，不需要依据其他用户对商品的评价，而是通过商品属性进行商品相似度度量，从而推荐给用户所感兴趣商品的相似商品；缺点是对于没有任何行为的新用户同样存在冷启动的问题。 +- 组合推荐[[2](#参考文献)]（Hybrid Recommendation）：运用不同的输入和技术共同进行推荐，以弥补各自推荐技术的缺点。 + +其中协同过滤是应用最广泛的技术之一，它又可以分为多个子类：基于用户 （User-Based）的推荐[[3](#参考文献)] 、基于物品（Item-Based）的推荐[[4](#参考文献)]、基于社交网络关系（Social-Based）的推荐[[5](#参考文献)]、基于模型（Model-based）的推荐等。1994年明尼苏达大学推出的GroupLens系统[[3](#参考文献)]一般被认为是推荐系统成为一个相对独立的研究方向的标志。该系统首次提出了基于协同过滤来完成推荐任务的思想，此后，基于该模型的协同过滤推荐引领了推荐系统十几年的发展方向。 + +深度学习具有优秀的自动提取特征的能力，能够学习多层次的抽象特征表示，并对异质或跨域的内容信息进行学习，可以一定程度上处理推荐系统冷启动问题[[6](#参考文献)]。本教程主要介绍个性化推荐的深度学习模型，以及如何使用PaddlePaddle实现模型。 + +## 效果展示 + +我们使用包含用户信息、电影信息与电影评分的数据集作为个性化推荐的应用场景。当我们训练好模型后，只需要输入对应的用户ID和电影ID，就可以得出一个匹配的分数（范围[0,5]，分数越高视为兴趣越大），然后根据所有电影的推荐得分排序，推荐给用户可能感兴趣的电影。 + +``` +Input movie_id: 1962 +Input user_id: 1 +Prediction Score is 4.25 +``` + +## 模型概览 + +本章中，我们首先介绍YouTube的视频推荐系统[[7](#参考文献)]，然后介绍我们实现的融合推荐模型。 + +### YouTube的深度神经网络推荐系统 + +YouTube是世界上最大的视频上传、分享和发现网站，YouTube推荐系统为超过10亿用户从不断增长的视频库中推荐个性化的内容。整个系统由两个神经网络组成：候选生成网络和排序网络。候选生成网络从百万量级的视频库中生成上百个候选，排序网络对候选进行打分排序，输出排名最高的数十个结果。系统结构如图1所示： + +![YouTube_Overview](./image/YouTube_Overview.png) +

+图1. YouTube 推荐系统结构 +

+ +#### 候选生成网络（Candidate Generation Network） + +候选生成网络将推荐问题建模为一个类别数极大的多类分类问题：对于一个Youtube用户，使用其观看历史（视频ID）、搜索词记录（search tokens）、人口学信息（如地理位置、用户登录设备）、二值特征（如性别，是否登录）和连续特征（如用户年龄）等，对视频库中所有视频进行多分类，得到每一类别的分类结果（即每一个视频的推荐概率），最终输出概率较高的几百个视频。 + +首先，将观看历史及搜索词记录这类历史信息，映射为向量后取平均值得到定长表示；同时，输入人口学特征以优化新用户的推荐效果，并将二值特征和连续特征归一化处理到[0, 1]范围。接下来，将所有特征表示拼接为一个向量，并输入给非线形多层感知器（MLP，详见[识别数字](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md)教程）处理。最后，训练时将MLP的输出给softmax做分类，预测时计算用户的综合特征（MLP的输出）与所有视频的相似度，取得分最高的`$k$`个作为候选生成网络的筛选结果。图2显示了候选生成网络结构。 + +![Deep_candidate_generation_model_architecture](./image/Deep_candidate_generation_model_architecture.png) +

+图2. 候选生成网络结构 +

+ +对于一个用户`$U$`，预测此刻用户要观看的视频`$\omega$`为视频`$i$`的概率公式为： + +$$P(\omega=i|u)=\frac{e^{v_{i}u}}{\sum_{j \in V}e^{v_{j}u}}$$ + +其中`$u$`为用户`$U$`的特征表示，`$V$`为视频库集合，`$v_i$`为视频库中第`$i$`个视频的特征表示。`$u$`和`$v_i$`为长度相等的向量，两者点积可以通过全连接层实现。 + +考虑到softmax分类的类别数非常多，为了保证一定的计算效率：1）训练阶段，使用负样本类别采样将实际计算的类别数缩小至数千；2）推荐（预测）阶段，忽略softmax的归一化计算（不影响结果），将类别打分问题简化为点积（dot product）空间中的最近邻（nearest neighbor）搜索问题，取与`$u$`最近的`$k$`个视频作为生成的候选。 + +#### 排序网络（Ranking Network） +排序网络的结构类似于候选生成网络，但是它的目标是对候选进行更细致的打分排序。和传统广告排序中的特征抽取方法类似，这里也构造了大量的用于视频排序的相关特征（如视频 ID、上次观看时间等）。这些特征的处理方式和候选生成网络类似，不同之处是排序网络的顶部是一个加权逻辑回归（weighted logistic regression），它对所有候选视频进行打分，从高到底排序后将分数较高的一些视频返回给用户。 + +### 融合推荐模型 +本节会使卷积神经网络（Convolutional Neural Networks）来学习电影名称的表示。下面会依次介绍文本卷积神经网络以及融合推荐模型。 + +#### 文本卷积神经网络（CNN） + +卷积神经网络经常用来处理具有类似网格拓扑结构（grid-like topology）的数据。例如，图像可以视为二维网格的像素点，自然语言可以视为一维的词序列。卷积神经网络可以提取多种局部特征，并对其进行组合抽象得到更高级的特征表示。实验表明，卷积神经网络能高效地对图像及文本问题进行建模处理。 + +卷积神经网络主要由卷积（convolution）和池化（pooling）操作构成，其应用及组合方式灵活多变，种类繁多。本小结我们以如图3所示的网络进行讲解： + +![text_cnn](./image/text_cnn.png) +

+图3. 卷积神经网络文本分类模型 +

+ +假设待处理句子的长度为`$n$`，其中第`$i$`个词的词向量（word embedding）为`$x_i\in\mathbb{R}^k$`，`$k$`为维度大小。 + +首先，进行词向量的拼接操作：将每`$h$`个词拼接起来形成一个大小为`$h$`的词窗口，记为`$x_{i:i+h-1}$`，它表示词序列`$x_{i},x_{i+1},\ldots,x_{i+h-1}$`的拼接，其中，`$i$`表示词窗口中第一个词在整个句子中的位置，取值范围从`$1$`到`$n-h+1$`，`$x_{i:i+h-1}\in\mathbb{R}^{hk}$`。 + +其次，进行卷积操作：把卷积核(kernel)`$w\in\mathbb{R}^{hk}$`应用于包含`$h$`个词的窗口`$x_{i:i+h-1}$`，得到特征`$c_i=f(w\cdot x_{i:i+h-1}+b)$`，其中`$b\in\mathbb{R}$`为偏置项（bias），`$f$`为非线性激活函数，如`$sigmoid$`。将卷积核应用于句子中所有的词窗口`${x_{1:h},x_{2:h+1},\ldots,x_{n-h+1:n}}$`，产生一个特征图（feature map）： + +$$c=[c_1,c_2,\ldots,c_{n-h+1}], c \in \mathbb{R}^{n-h+1}$$ + +接下来，对特征图采用时间维度上的最大池化（max pooling over time）操作得到此卷积核对应的整句话的特征`$\hat c$`，它是特征图中所有元素的最大值： + +$$\hat c=max(c)$$ + +#### 模型概览 + +在融合推荐模型的电影推荐系统中： + +1. 首先，使用用户特征和电影特征作为神经网络的输入，其中： + +- 用户特征融合了四个属性信息，分别是用户ID、性别、职业和年龄。 + +- 电影特征融合了三个属性信息，分别是电影ID、电影类型ID和电影名称。 + +2. 对用户特征，将用户ID映射为维度大小为256的向量表示，输入全连接层，并对其他三个属性也做类似的处理。然后将四个属性的特征表示分别全连接并相加。 + +3. 对电影特征，将电影ID以类似用户ID的方式进行处理，电影类型ID以向量的形式直接输入全连接层，电影名称用文本卷积神经网络得到其定长向量表示。然后将三个属性的特征表示分别全连接并相加。 + +4. 得到用户和电影的向量表示后，计算二者的余弦相似度作为推荐系统的打分。最后，用该相似度打分和用户真实打分的差异的平方作为该回归模型的损失函数。 + +![rec_regression_network](./image/rec_regression_network.png) +

+图4. 融合推荐模型 +

+ +## 数据准备 + +### 数据介绍与下载 + +我们以 [MovieLens 百万数据集（ml-1m）](http://files.grouplens.org/datasets/movielens/ml-1m.zip)为例进行介绍。ml-1m 数据集包含了 6,000 位用户对 4,000 部电影的 1,000,000 条评价（评分范围 1~5 分，均为整数），由 GroupLens Research 实验室搜集整理。 + +Paddle在API中提供了自动加载数据的模块。数据模块为 `paddle.dataset.movielens` + + +```python +import paddle +movie_info = paddle.dataset.movielens.movie_info() +print movie_info.values()[0] +``` + + +```python +# Run this block to show dataset's documentation +# help(paddle.dataset.movielens) +``` + +在原始数据中包含电影的特征数据，用户的特征数据，和用户对电影的评分。 + +例如，其中某一个电影特征为: + + +```python +movie_info = paddle.dataset.movielens.movie_info() +print movie_info.values()[0] +``` + + + + +这表示，电影的id是1，标题是《Toy Story》，该电影被分为到三个类别中。这三个类别是动画，儿童，喜剧。 + + +```python +user_info = paddle.dataset.movielens.user_info() +print user_info.values()[0] +``` + + + + +这表示，该用户ID是1，女性，年龄比18岁还年轻。职业ID是10。 + + +其中，年龄使用下列分布 +* 1: "Under 18" +* 18: "18-24" +* 25: "25-34" +* 35: "35-44" +* 45: "45-49" +* 50: "50-55" +* 56: "56+" + +职业是从下面几种选项里面选则得出: +* 0: "other" or not specified +* 1: "academic/educator" +* 2: "artist" +* 3: "clerical/admin" +* 4: "college/grad student" +* 5: "customer service" +* 6: "doctor/health care" +* 7: "executive/managerial" +* 8: "farmer" +* 9: "homemaker" +* 10: "K-12 student" +* 11: "lawyer" +* 12: "programmer" +* 13: "retired" +* 14: "sales/marketing" +* 15: "scientist" +* 16: "self-employed" +* 17: "technician/engineer" +* 18: "tradesman/craftsman" +* 19: "unemployed" +* 20: "writer" + +而对于每一条训练/测试数据，均为 <用户特征> + <电影特征> + 评分。 + +例如，我们获得第一条训练数据: + + +```python +train_set_creator = paddle.dataset.movielens.train() +train_sample = next(train_set_creator()) +uid = train_sample[0] +mov_id = train_sample[len(user_info[uid].value())] +print "User %s rates Movie %s with Score %s"%(user_info[uid], movie_info[mov_id], train_sample[-1]) +``` + +User rates Movie with Score [5.0] + + +即用户1对电影1193的评价为5分。 + +## 模型配置说明 + +下面我们开始根据输入数据的形式配置模型。首先引入所需的库函数以及定义全局变量。 + + +```python +import math +import sys +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.layers as layers +import paddle.fluid.nets as nets + +IS_SPARSE = True +USE_GPU = False +BATCH_SIZE = 256 +``` + +然后为我们的用户特征综合模型定义模型配置 + +```python +def get_usr_combined_features(): + +USR_DICT_SIZE = paddle.dataset.movielens.max_user_id() + 1 + +uid = layers.data(name='user_id', shape=[1], dtype='int64') + +usr_emb = layers.embedding( +input=uid, +dtype='float32', +size=[USR_DICT_SIZE, 32], +param_attr='user_table', +is_sparse=IS_SPARSE) + +usr_fc = layers.fc(input=usr_emb, size=32) + +USR_GENDER_DICT_SIZE = 2 + +usr_gender_id = layers.data(name='gender_id', shape=[1], dtype='int64') + +usr_gender_emb = layers.embedding( +input=usr_gender_id, +size=[USR_GENDER_DICT_SIZE, 16], +param_attr='gender_table', +is_sparse=IS_SPARSE) + +usr_gender_fc = layers.fc(input=usr_gender_emb, size=16) + +USR_AGE_DICT_SIZE = len(paddle.dataset.movielens.age_table) +usr_age_id = layers.data(name='age_id', shape=[1], dtype="int64") + +usr_age_emb = layers.embedding( +input=usr_age_id, +size=[USR_AGE_DICT_SIZE, 16], +is_sparse=IS_SPARSE, +param_attr='age_table') + +usr_age_fc = layers.fc(input=usr_age_emb, size=16) + +USR_JOB_DICT_SIZE = paddle.dataset.movielens.max_job_id() + 1 +usr_job_id = layers.data(name='job_id', shape=[1], dtype="int64") + +usr_job_emb = layers.embedding( +input=usr_job_id, +size=[USR_JOB_DICT_SIZE, 16], +param_attr='job_table', +is_sparse=IS_SPARSE) + +usr_job_fc = layers.fc(input=usr_job_emb, size=16) + +concat_embed = layers.concat( +input=[usr_fc, usr_gender_fc, usr_age_fc, usr_job_fc], axis=1) + +usr_combined_features = layers.fc(input=concat_embed, size=200, act="tanh") + +return usr_combined_features +``` + +如上述代码所示，对于每个用户，我们输入4维特征。其中包括user_id,gender_id,age_id,job_id。这几维特征均是简单的整数值。为了后续神经网络处理这些特征方便，我们借鉴NLP中的语言模型，将这几维离散的整数值，变换成embedding取出。分别形成usr_emb, usr_gender_emb, usr_age_emb, usr_job_emb。 + +然后，我们对于所有的用户特征，均输入到一个全连接层(fc)中。将所有特征融合为一个200维度的特征。 + +进而，我们对每一个电影特征做类似的变换，网络配置为: + + +```python +def get_mov_combined_features(): + +MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1 + +mov_id = layers.data(name='movie_id', shape=[1], dtype='int64') + +mov_emb = layers.embedding( +input=mov_id, +dtype='float32', +size=[MOV_DICT_SIZE, 32], +param_attr='movie_table', +is_sparse=IS_SPARSE) + +mov_fc = layers.fc(input=mov_emb, size=32) + +CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories()) + +category_id = layers.data( +name='category_id', shape=[1], dtype='int64', lod_level=1) + +mov_categories_emb = layers.embedding( +input=category_id, size=[CATEGORY_DICT_SIZE, 32], is_sparse=IS_SPARSE) + +mov_categories_hidden = layers.sequence_pool( +input=mov_categories_emb, pool_type="sum") + +MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict()) + +mov_title_id = layers.data( +name='movie_title', shape=[1], dtype='int64', lod_level=1) + +mov_title_emb = layers.embedding( +input=mov_title_id, size=[MOV_TITLE_DICT_SIZE, 32], is_sparse=IS_SPARSE) + +mov_title_conv = nets.sequence_conv_pool( +input=mov_title_emb, +num_filters=32, +filter_size=3, +act="tanh", +pool_type="sum") + +concat_embed = layers.concat( +input=[mov_fc, mov_categories_hidden, mov_title_conv], axis=1) + +mov_combined_features = layers.fc(input=concat_embed, size=200, act="tanh") + +return mov_combined_features +``` + +电影标题名称(title)是一个序列的整数，整数代表的是这个词在索引序列中的下标。这个序列会被送入 `sequence_conv_pool` 层，这个层会在时间维度上使用卷积和池化。因为如此，所以输出会是固定长度，尽管输入的序列长度各不相同。 + +最后，我们定义一个`inference_program`来使用余弦相似度计算用户特征与电影特征的相似性。 + +```python +def inference_program(): +usr_combined_features = get_usr_combined_features() +mov_combined_features = get_mov_combined_features() + +inference = layers.cos_sim(X=usr_combined_features, Y=mov_combined_features) +scale_infer = layers.scale(x=inference, scale=5.0) + +return scale_infer +``` + +进而，我们定义一个`train_program`来使用`inference_program`计算出的结果，在标记数据的帮助下来计算误差。我们还定义了一个`optimizer_func`来定义优化器。 + +```python +def train_program(): + +scale_infer = inference_program() + +label = layers.data(name='score', shape=[1], dtype='float32') +square_cost = layers.square_error_cost(input=scale_infer, label=label) +avg_cost = layers.mean(square_cost) + +return [avg_cost, scale_infer] + + +def optimizer_func(): +return fluid.optimizer.SGD(learning_rate=0.2) +``` + + +## 训练模型 + +### 定义训练环境 +定义您的训练环境，可以指定训练是发生在CPU还是GPU上。 + +```python +use_cuda = False +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +``` + +### 定义数据提供器 +下一步是为训练和测试定义数据提供器。提供器读入一个大小为 `BATCH_SIZE`的数据。`paddle.dataset.movielens.train` 每次会在乱序化后提供一个大小为`BATCH_SIZE`的数据，乱序化的大小为缓存大小`buf_size`。 + +```python +train_reader = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.movielens.train(), buf_size=8192), +batch_size=BATCH_SIZE) + +test_reader = paddle.batch( +paddle.dataset.movielens.test(), batch_size=BATCH_SIZE) +``` + +### 构造训练器(trainer) +训练器需要一个训练程序和一个训练优化函数。 + +```python +trainer = fluid.Trainer( +train_func=train_program, place=place, optimizer_func=optimizer_func) +``` + +### 提供数据 + +`feed_order`用来定义每条产生的数据和`paddle.layer.data`之间的映射关系。比如，`movielens.train`产生的第一列的数据对应的是`user_id`这个特征。 + +```python +feed_order = [ +'user_id', 'gender_id', 'age_id', 'job_id', 'movie_id', 'category_id', +'movie_title', 'score' +] +``` + +### 事件处理器 +回调函数`event_handler`在一个之前定义好的事件发生后会被调用。例如，我们可以在每步训练结束后查看误差。 + +```python +# Specify the directory path to save the parameters +params_dirname = "recommender_system.inference.model" + +from paddle.v2.plot import Ploter +test_title = "Test cost" +plot_cost = Ploter(test_title) + + +def event_handler(event): +if isinstance(event, fluid.EndStepEvent): +avg_cost_set = trainer.test( +reader=test_reader, feed_order=feed_order) + +# get avg cost +avg_cost = np.array(avg_cost_set).mean() + +plot_cost.append(test_title, event.step, avg_cost_set[0]) +plot_cost.plot() + +print("avg_cost: %s" % avg_cost) +print('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, +float(avg_cost))) + +if event.step == 20: # Adjust this number for accuracy +trainer.save_params(params_dirname) +trainer.stop() +``` + +### 开始训练 +最后，我们传入训练循环数（`num_epoch`）和一些别的参数，调用 `trainer.train` 来开始训练。 + +```python +trainer.train( +num_epochs=1, +event_handler=event_handler, +reader=train_reader, +feed_order=feed_order) +``` + +## 应用模型 + +### 构建预测器 +传入`inference_program`和`params_dirname`来初始化一个预测器, `params_dirname`用来存放训练过程中的各个参数。 + +```python +inferencer = fluid.Inferencer( +inference_program, param_path=params_dirname, place=place) +``` + +### 生成测试用输入数据 +使用 create_lod_tensor(data, lod, place) 的API来生成细节层次的张量。`data`是一个序列，每个元素是一个索引号的序列。`lod`是细节层次的信息，对应于`data`。比如，data = [[10, 2, 3], [2, 3]] 意味着它包含两个序列，长度分别是3和2。于是相应地 lod = [[3, 2]]，它表明其包含一层细节信息，意味着 `data` 有两个序列，长度分别是3和2。 + +在这个预测例子中，我们试着预测用户ID为1的用户对于电影'Hunchback of Notre Dame'的评分 + +```python +infer_movie_id = 783 +infer_movie_name = paddle.dataset.movielens.movie_info()[infer_movie_id].title +user_id = fluid.create_lod_tensor([[1]], [[1]], place) +gender_id = fluid.create_lod_tensor([[1]], [[1]], place) +age_id = fluid.create_lod_tensor([[0]], [[1]], place) +job_id = fluid.create_lod_tensor([[10]], [[1]], place) +movie_id = fluid.create_lod_tensor([[783]], [[1]], place) # Hunchback of Notre Dame +category_id = fluid.create_lod_tensor([[10, 8, 9]], [[3]], place) # Animation, Children's, Musical +movie_title = fluid.create_lod_tensor([[1069, 4140, 2923, 710, 988]], [[5]], +place) # 'hunchback','of','notre','dame','the' +``` + +### 测试 +现在我们可以进行预测了。我们要提供的`feed_order`应该和训练过程一致。 + + +```python +results = inferencer.infer( +{ +'user_id': user_id, +'gender_id': gender_id, +'age_id': age_id, +'job_id': job_id, +'movie_id': movie_id, +'category_id': category_id, +'movie_title': movie_title +}, +return_numpy=False) +``` + +## 总结 + +本章介绍了传统的推荐系统方法和YouTube的深度神经网络推荐系统，并以电影推荐为例，使用PaddlePaddle训练了一个个性化推荐神经网络模型。推荐系统几乎涵盖了电商系统、社交网络、广告推荐、搜索引擎等领域的方方面面，而在图像处理、自然语言处理等领域已经发挥重要作用的深度学习技术，也将会在推荐系统领域大放异彩。 + +## 参考文献 + +1. [Peter Brusilovsky](https://en.wikipedia.org/wiki/Peter_Brusilovsky) (2007). *The Adaptive Web*. p. 325. +2. Robin Burke , [Hybrid Web Recommender Systems](http://www.dcs.warwick.ac.uk/~acristea/courses/CS411/2010/Book%20-%20The%20Adaptive%20Web/HybridWebRecommenderSystems.pdf), pp. 377-408, The Adaptive Web, Peter Brusilovsky, Alfred Kobsa, Wolfgang Nejdl (Ed.), Lecture Notes in Computer Science, Springer-Verlag, Berlin, Germany, Lecture Notes in Computer Science, Vol. 4321, May 2007, 978-3-540-72078-2. +3. P. Resnick, N. Iacovou, etc. “[GroupLens: An Open Architecture for Collaborative Filtering of Netnews](http://ccs.mit.edu/papers/CCSWP165.html)”, Proceedings of ACM Conference on Computer Supported Cooperative Work, CSCW 1994. pp.175-186. +4. Sarwar, Badrul, et al. "[Item-based collaborative filtering recommendation algorithms.](http://files.grouplens.org/papers/www10_sarwar.pdf)" *Proceedings of the 10th international conference on World Wide Web*. ACM, 2001. +5. Kautz, Henry, Bart Selman, and Mehul Shah. "[Referral Web: combining social networks and collaborative filtering.](http://www.cs.cornell.edu/selman/papers/pdf/97.cacm.refweb.pdf)" Communications of the ACM 40.3 (1997): 63-65. APA +6. Yuan, Jianbo, et al. ["Solving Cold-Start Problem in Large-scale Recommendation Engines: A Deep Learning Approach."](https://arxiv.org/pdf/1611.05480v1.pdf) *arXiv preprint arXiv:1611.05480* (2016). +7. Covington P, Adams J, Sargin E. [Deep neural networks for youtube recommendations](https://static.googleusercontent.com/media/research.google.com/zh-CN//pubs/archive/45530.pdf)[C]//Proceedings of the 10th ACM Conference on Recommender Systems. ACM, 2016: 191-198. + + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/.gitignore b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..667762d327cb160376a4119fa9df9db41b6443b2 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/.gitignore @@ -0,0 +1,10 @@ +data/aclImdb +data/imdb +data/pre-imdb +data/mosesdecoder-master +*.log +model_output +dataprovider_copy_1.py +model.list +*.pyc +.DS_Store diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm.png new file mode 100644 index 0000000000000000000000000000000000000000..98fbea413a98a619004ca669c67f5f867fe974c9 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm_en.png new file mode 100755 index 0000000000000000000000000000000000000000..d73a00bf2c1fca2f9b8c26bccf5ea844fa1db50b Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/lstm_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/rnn.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/rnn.png new file mode 100755 index 0000000000000000000000000000000000000000..26c904102a6e6c4e30f0048b81373ae8c148b355 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/rnn.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm.jpg b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6b2adf70f2b5112a2e82505da5cff9f5fd0c6298 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm.jpg differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm_en.png b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm_en.png new file mode 100755 index 0000000000000000000000000000000000000000..8b5dbd726178b5555c513294e7b10a81acc96ff5 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/image/stacked_lstm_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/index.md b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/index.md new file mode 100644 index 0000000000000000000000000000000000000000..624de7e4d439953c7255481fb0c9d62ce94f3900 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/index.md @@ -0,0 +1,354 @@ +# 情感分析 + +本教程源代码目录在[book/understand_sentiment](https://github.com/PaddlePaddle/book/tree/develop/06.understand_sentiment)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 + +在自然语言处理中，情感分析一般是指判断一段文本所表达的情绪状态。其中，一段文本可以是一个句子，一个段落或一个文档。情绪状态可以是两类，如（正面，负面），（高兴，悲伤）；也可以是三类，如（积极，消极，中性）等等。情感分析的应用场景十分广泛，如把用户在购物网站（亚马逊、天猫、淘宝等）、旅游网站、电影评论网站上发表的评论分成正面评论和负面评论；或为了分析用户对于某一产品的整体使用感受，抓取产品的用户评论并进行情感分析等等。表格1展示了对电影评论进行情感分析的例子： + +| 电影评论 | 类别 | +| -------- | ----- | +| 在冯小刚这几年的电影里，算最好的一部的了| 正面 | +| 很不好看，好像一个地方台的电视剧 | 负面 | +| 圆方镜头全程炫技，色调背景美则美矣，但剧情拖沓，口音不伦不类，一直努力却始终无法入戏| 负面| +|剧情四星。但是圆镜视角加上婺源的风景整个非常有中国写意山水画的感觉，看得实在太舒服了。。|正面| + +

表格 1 电影评论情感分析

+ +在自然语言处理中，情感分析属于典型的**文本分类**问题，即把需要进行情感分析的文本划分为其所属类别。文本分类涉及文本表示和分类方法两个问题。在深度学习的方法出现之前，主流的文本表示方法为词袋模型BOW(bag of words)，话题模型等等；分类方法有SVM(support vector machine), LR(logistic regression)等等。 + +对于一段文本，BOW表示会忽略其词顺序、语法和句法，将这段文本仅仅看做是一个词集合，因此BOW方法并不能充分表示文本的语义信息。例如，句子“这部电影糟糕透了”和“一个乏味，空洞，没有内涵的作品”在情感分析中具有很高的语义相似度，但是它们的BOW表示的相似度为0。又如，句子“一个空洞，没有内涵的作品”和“一个不空洞而且有内涵的作品”的BOW相似度很高，但实际上它们的意思很不一样。 + +本章我们所要介绍的深度学习模型克服了BOW表示的上述缺陷，它在考虑词顺序的基础上把文本映射到低维度的语义空间，并且以端对端（end to end）的方式进行文本表示及分类，其性能相对于传统方法有显著的提升\[[1](#参考文献)\]。 + +## 模型概览 +本章所使用的文本表示模型为卷积神经网络（Convolutional Neural Networks）和循环神经网络(Recurrent Neural Networks)及其扩展。下面依次介绍这几个模型。 + +### 文本卷积神经网络简介（CNN） + +我们在[推荐系统](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system)一节介绍过应用于文本数据的卷积神经网络模型的计算过程，这里进行一个简单的回顾。 + +对卷积神经网络来说，首先使用卷积处理输入的词向量序列，产生一个特征图（feature map），对特征图采用时间维度上的最大池化（max pooling over time）操作得到此卷积核对应的整句话的特征，最后，将所有卷积核得到的特征拼接起来即为文本的定长向量表示，对于文本分类问题，将其连接至softmax即构建出完整的模型。在实际应用中，我们会使用多个卷积核来处理句子，窗口大小相同的卷积核堆叠起来形成一个矩阵，这样可以更高效的完成运算。另外，我们也可使用窗口大小不同的卷积核来处理句子，[推荐系统](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system)一节的图3作为示意画了四个卷积核，不同颜色表示不同大小的卷积核操作。 + +对于一般的短文本分类问题，上文所述的简单的文本卷积网络即可达到很高的正确率\[[1](#参考文献)\]。若想得到更抽象更高级的文本特征表示，可以构建深层文本卷积神经网络\[[2](#参考文献),[3](#参考文献)\]。 + +### 循环神经网络（RNN） + +循环神经网络是一种能对序列数据进行精确建模的有力工具。实际上，循环神经网络的理论计算能力是图灵完备的\[[4](#参考文献)\]。自然语言是一种典型的序列数据（词序列），近年来，循环神经网络及其变体（如long short term memory\[[5](#参考文献)\]等）在自然语言处理的多个领域，如语言模型、句法解析、语义角色标注（或一般的序列标注）、语义表示、图文生成、对话、机器翻译等任务上均表现优异甚至成为目前效果最好的方法。 + +![rnn](./image/rnn.png) +

+图1. 循环神经网络按时间展开的示意图 +

+ +循环神经网络按时间展开后如图1所示：在第`$t$`时刻，网络读入第`$t$`个输入`$x_t$`（向量表示）及前一时刻隐层的状态值`$h_{t-1}$`（向量表示，`$h_0$`一般初始化为`$0$`向量），计算得出本时刻隐层的状态值`$h_t$`，重复这一步骤直至读完所有输入。如果将循环神经网络所表示的函数记为`$f$`，则其公式可表示为： + +$$h_t=f(x_t,h_{t-1})=\sigma(W_{xh}x_t+W_{hh}h_{t-1}+b_h)$$ + +其中`$W_{xh}$`是输入到隐层的矩阵参数，`$W_{hh}$`是隐层到隐层的矩阵参数，`$b_h$`为隐层的偏置向量（bias）参数，`$\sigma$`为`$sigmoid$`函数。 + +在处理自然语言时，一般会先将词（one-hot表示）映射为其词向量（word embedding）表示，然后再作为循环神经网络每一时刻的输入`$x_t$`。此外，可以根据实际需要的不同在循环神经网络的隐层上连接其它层。如，可以把一个循环神经网络的隐层输出连接至下一个循环神经网络的输入构建深层（deep or stacked）循环神经网络，或者提取最后一个时刻的隐层状态作为句子表示进而使用分类模型等等。 + +### 长短期记忆网络（LSTM） + +对于较长的序列数据，循环神经网络的训练过程中容易出现梯度消失或爆炸现象\[[6](#参考文献)\]。为了解决这一问题，Hochreiter S, Schmidhuber J. (1997)提出了LSTM(long short term memory\[[5](#参考文献)\])。 + +相比于简单的循环神经网络，LSTM增加了记忆单元`$c$`、输入门`$i$`、遗忘门`$f$`及输出门`$o$`。这些门及记忆单元组合起来大大提升了循环神经网络处理长序列数据的能力。若将基于LSTM的循环神经网络表示的函数记为`$F$`，则其公式为： + +$$ h_t=F(x_t,h_{t-1})$$ + +`$F$`由下列公式组合而成\[[7](#参考文献)\]： +$$ i_t = \sigma{(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}c_{t-1}+b_i)} $$ +$$ f_t = \sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}c_{t-1}+b_f) $$ +$$ c_t = f_t\odot c_{t-1}+i_t\odot tanh(W_{xc}x_t+W_{hc}h_{t-1}+b_c) $$ +$$ o_t = \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}c_{t}+b_o) $$ +$$ h_t = o_t\odot tanh(c_t) $$ +其中，`$i_t, f_t, c_t, o_t$`分别表示输入门，遗忘门，记忆单元及输出门的向量值，带角标的`$W$`及`$b$`为模型参数，`$tanh$`为双曲正切函数，`$\odot$`表示逐元素（elementwise）的乘法操作。输入门控制着新输入进入记忆单元`$c$`的强度，遗忘门控制着记忆单元维持上一时刻值的强度，输出门控制着输出记忆单元的强度。三种门的计算方式类似，但有着完全不同的参数，它们各自以不同的方式控制着记忆单元`$c$`，如图2所示： + +![lstm](./image/lstm.png) +

+图2. 时刻`$t$`的LSTM [7] +

+ +LSTM通过给简单的循环神经网络增加记忆及控制门的方式，增强了其处理远距离依赖问题的能力。类似原理的改进还有Gated Recurrent Unit (GRU)\[[8](#参考文献)\]，其设计更为简洁一些。**这些改进虽然各有不同，但是它们的宏观描述却与简单的循环神经网络一样（如图2所示），即隐状态依据当前输入及前一时刻的隐状态来改变，不断地循环这一过程直至输入处理完毕：** + +$$ h_t=Recrurent(x_t,h_{t-1})$$ + +其中，`$Recrurent$`可以表示简单的循环神经网络、GRU或LSTM。 + +### 栈式双向LSTM（Stacked Bidirectional LSTM） + +对于正常顺序的循环神经网络，`$h_t$`包含了`$t$`时刻之前的输入信息，也就是上文信息。同样，为了得到下文信息，我们可以使用反方向（将输入逆序处理）的循环神经网络。结合构建深层循环神经网络的方法（深层神经网络往往能得到更抽象和高级的特征表示），我们可以通过构建更加强有力的基于LSTM的栈式双向循环神经网络\[[9](#参考文献)\]，来对时序数据进行建模。 + +如图3所示（以三层为例），奇数层LSTM正向，偶数层LSTM反向，高一层的LSTM使用低一层LSTM及之前所有层的信息作为输入，对最高层LSTM序列使用时间维度上的最大池化即可得到文本的定长向量表示（这一表示充分融合了文本的上下文信息，并且对文本进行了深层次抽象），最后我们将文本表示连接至softmax构建分类模型。 + +![stacked_lstm](./image/stacked_lstm.jpg) +

+图3. 栈式双向LSTM用于文本分类 +

+ + +## 数据集介绍 + +我们以[IMDB情感分析数据集](http://ai.stanford.edu/%7Eamaas/data/sentiment/)为例进行介绍。IMDB数据集的训练集和测试集分别包含25000个已标注过的电影评论。其中，负面评论的得分小于等于4，正面评论的得分大于等于7，满分10分。 +```text +aclImdb +|- test +|-- neg +|-- pos +|- train +|-- neg +|-- pos +``` +Paddle在`dataset/imdb.py`中提实现了imdb数据集的自动下载和读取，并提供了读取字典、训练数据、测试数据等API。 + +## 配置模型 + +在该示例中，我们实现了两种文本分类算法，分别基于[推荐系统](https://github.com/PaddlePaddle/book/tree/develop/05.recommender_system)一节介绍过的文本卷积神经网络，以及[栈式双向LSTM](#栈式双向LSTM（Stacked Bidirectional LSTM）)。我们首先引入要用到的库和定义全局变量： + +```python +import paddle +import paddle.fluid as fluid +from functools import partial +import numpy as np + +CLASS_DIM = 2 +EMB_DIM = 128 +HID_DIM = 512 +BATCH_SIZE = 128 +USE_GPU = False +``` + + +### 文本卷积神经网络 +我们构建神经网络`convolution_net`，示例代码如下。 +需要注意的是：`fluid.nets.sequence_conv_pool` 包含卷积和池化层两个操作。 + +```python +def convolution_net(data, input_dim, class_dim, emb_dim, hid_dim): +emb = fluid.layers.embedding( +input=data, size=[input_dim, emb_dim], is_sparse=True) +conv_3 = fluid.nets.sequence_conv_pool( +input=emb, +num_filters=hid_dim, +filter_size=3, +act="tanh", +pool_type="sqrt") +conv_4 = fluid.nets.sequence_conv_pool( +input=emb, +num_filters=hid_dim, +filter_size=4, +act="tanh", +pool_type="sqrt") +prediction = fluid.layers.fc( +input=[conv_3, conv_4], size=class_dim, act="softmax") +return prediction +``` + +网络的输入`input_dim`表示的是词典的大小，`class_dim`表示类别数。这里，我们使用[`sequence_conv_pool`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/trainer_config_helpers/networks.py) API实现了卷积和池化操作。 + +### 栈式双向LSTM + +栈式双向神经网络`stacked_lstm_net`的代码片段如下： + +```python +def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num): + +emb = fluid.layers.embedding( +input=data, size=[input_dim, emb_dim], is_sparse=True) + +fc1 = fluid.layers.fc(input=emb, size=hid_dim) +lstm1, cell1 = fluid.layers.dynamic_lstm(input=fc1, size=hid_dim) + +inputs = [fc1, lstm1] + +for i in range(2, stacked_num + 1): +fc = fluid.layers.fc(input=inputs, size=hid_dim) +lstm, cell = fluid.layers.dynamic_lstm( +input=fc, size=hid_dim, is_reverse=(i % 2) == 0) +inputs = [fc, lstm] + +fc_last = fluid.layers.sequence_pool(input=inputs[0], pool_type='max') +lstm_last = fluid.layers.sequence_pool(input=inputs[1], pool_type='max') + +prediction = fluid.layers.fc(input=[fc_last, lstm_last], +size=class_dim, +act='softmax') +return prediction +``` +以上的栈式双向LSTM抽象出了高级特征并把其映射到和分类类别数同样大小的向量上。`paddle.activation.Softmax`函数用来计算分类属于某个类别的概率。 + +重申一下，此处我们可以调用`convolution_net`或`stacked_lstm_net`的任何一个。我们以`convolution_net`为例。 + +接下来我们定义预测程序（`inference_program`）。预测程序使用`convolution_net`来对`fluid.layer.data`的输入进行预测。 + +```python +def inference_program(word_dict): +data = fluid.layers.data( +name="words", shape=[1], dtype="int64", lod_level=1) + +dict_dim = len(word_dict) +net = convolution_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM) +return net +``` + +我们这里定义了`training_program`。它使用了从`inference_program`返回的结果来计算误差。我们同时定义了优化函数`optimizer_func`。 + +因为是有监督的学习，训练集的标签也在`paddle.layer.data`中定义了。在训练过程中，交叉熵用来在`paddle.layer.classification_cost`中作为损失函数。 + +在测试过程中，分类器会计算各个输出的概率。第一个返回的数值规定为 损耗(cost)。 + +```python +def train_program(word_dict): +prediction = inference_program(word_dict) +label = fluid.layers.data(name="label", shape=[1], dtype="int64") +cost = fluid.layers.cross_entropy(input=prediction, label=label) +avg_cost = fluid.layers.mean(cost) +accuracy = fluid.layers.accuracy(input=prediction, label=label) +return [avg_cost, accuracy] + + +def optimizer_func(): +return fluid.optimizer.Adagrad(learning_rate=0.002) +``` + +## 训练模型 + +### 定义训练环境 + +定义您的训练是在CPU上还是在GPU上： + + +```python +use_cuda = False +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +``` + +### 定义数据提供器 + +下一步是为训练和测试定义数据提供器。提供器读入一个大小为 BATCH_SIZE的数据。paddle.dataset.imdb.train 每次会在乱序化后提供一个大小为BATCH_SIZE的数据，乱序化的大小为缓存大小buf_size。 + +注意：读取IMDB的数据可能会花费几分钟的时间，请耐心等待。 + +```python +print("Loading IMDB word dict....") +word_dict = paddle.dataset.imdb.word_dict() + +print ("Reading training data....") +train_reader = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.imdb.train(word_dict), buf_size=25000), +batch_size=BATCH_SIZE) +``` + +### 构造训练器(trainer) +训练器需要一个训练程序和一个训练优化函数。 + +```python +trainer = fluid.Trainer( +train_func=partial(train_program, word_dict), +place=place, +optimizer_func=optimizer_func) +``` + +### 提供数据 + +`feed_order`用来定义每条产生的数据和`paddle.layer.data`之间的映射关系。比如，`imdb.train`产生的第一列的数据对应的是`words`这个特征。 + +```python +feed_order = ['words', 'label'] +``` + +### 事件处理器 + +回调函数event_handler在一个之前定义好的事件发生后会被调用。例如，我们可以在每步训练结束后查看误差。 + +```python +# Specify the directory path to save the parameters +params_dirname = "understand_sentiment_conv.inference.model" + +def event_handler(event): +if isinstance(event, fluid.EndStepEvent): +print("Step {0}, Epoch {1} Metrics {2}".format( +event.step, event.epoch, map(np.array, event.metrics))) + +if event.step == 10: +trainer.save_params(params_dirname) +trainer.stop() +``` + +### 开始训练 + +最后，我们传入训练循环数（num_epoch）和一些别的参数，调用 trainer.train 来开始训练。 + +```python +trainer.train( +num_epochs=1, +event_handler=event_handler, +reader=train_reader, +feed_order=feed_order) +``` + +## 应用模型 + +### 构建预测器 + +传入`inference_program`和`params_dirname`来初始化一个预测器, `params_dirname`用来存放训练过程中的各个参数。 + +```python +inferencer = fluid.Inferencer( +inference_program, param_path=params_dirname, place=place) +``` + +### 生成测试用输入数据 + +为了进行预测，我们任意选取3个评论。请随意选取您看好的3个。我们把评论中的每个词对应到`word_dict`中的id。如果词典中没有这个词，则设为`unknown`。 +然后我们用`create_lod_tensor`来创建细节层次的张量。 + +```python +reviews_str = [ +'read the book forget the movie', 'this is a great movie', 'this is very bad' +] +reviews = [c.split() for c in reviews_str] + +UNK = word_dict[''] +lod = [] +for c in reviews: +lod.append([word_dict.get(words, UNK) for words in c]) + +base_shape = [[len(c) for c in lod]] + +tensor_words = fluid.create_lod_tensor(lod, base_shape, place) +``` + +## 应用模型 + +现在我们可以对每一条评论进行正面或者负面的预测啦。 + +```python +results = inferencer.infer({'words': tensor_words}) + +for i, r in enumerate(results[0]): +print("Predict probability of ", r[0], " to be positive and ", r[1], " to be negative for review \'", reviews_str[i], "\'") + +``` + + +## 总结 + +本章我们以情感分析为例，介绍了使用深度学习的方法进行端对端的短文本分类，并且使用PaddlePaddle完成了全部相关实验。同时，我们简要介绍了两种文本处理模型：卷积神经网络和循环神经网络。在后续的章节中我们会看到这两种基本的深度学习模型在其它任务上的应用。 + + +## 参考文献 +1. Kim Y. [Convolutional neural networks for sentence classification](http://arxiv.org/pdf/1408.5882)[J]. arXiv preprint arXiv:1408.5882, 2014. +2. Kalchbrenner N, Grefenstette E, Blunsom P. [A convolutional neural network for modelling sentences](http://arxiv.org/pdf/1404.2188.pdf?utm_medium=App.net&utm_source=PourOver)[J]. arXiv preprint arXiv:1404.2188, 2014. +3. Yann N. Dauphin, et al. [Language Modeling with Gated Convolutional Networks](https://arxiv.org/pdf/1612.08083v1.pdf)[J] arXiv preprint arXiv:1612.08083, 2016. +4. Siegelmann H T, Sontag E D. [On the computational power of neural nets](http://research.cs.queensu.ca/home/akl/cisc879/papers/SELECTED_PAPERS_FROM_VARIOUS_SOURCES/05070215382317071.pdf)[C]//Proceedings of the fifth annual workshop on Computational learning theory. ACM, 1992: 440-449. +5. Hochreiter S, Schmidhuber J. [Long short-term memory](http://web.eecs.utk.edu/~itamar/courses/ECE-692/Bobby_paper1.pdf)[J]. Neural computation, 1997, 9(8): 1735-1780. +6. Bengio Y, Simard P, Frasconi P. [Learning long-term dependencies with gradient descent is difficult](http://www-dsi.ing.unifi.it/~paolo/ps/tnn-94-gradient.pdf)[J]. IEEE transactions on neural networks, 1994, 5(2): 157-166. +7. Graves A. [Generating sequences with recurrent neural networks](http://arxiv.org/pdf/1308.0850)[J]. arXiv preprint arXiv:1308.0850, 2013. +8. Cho K, Van Merriënboer B, Gulcehre C, et al. [Learning phrase representations using RNN encoder-decoder for statistical machine translation](http://arxiv.org/pdf/1406.1078)[J]. arXiv preprint arXiv:1406.1078, 2014. +9. Zhou J, Xu W. [End-to-end learning of semantic role labeling using recurrent neural networks](http://www.aclweb.org/anthology/P/P15/P15-1109.pdf)[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics. 2015. + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/.gitignore b/doc/fluid/new_docs/beginners_guide/basics/word2vec/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a620e0279c310d213d4e6d8e99e666962c11e352 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/word2vec/.gitignore @@ -0,0 +1,3 @@ +data/train.list +data/test.list +data/simple-examples* diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/2d_similarity.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/2d_similarity.png new file mode 100644 index 0000000000000000000000000000000000000000..384f59919a2c8dedb198e97d51434616648932e1 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/2d_similarity.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow.png new file mode 100644 index 0000000000000000000000000000000000000000..76b7d4bc0f99372465bd9aa34721513d39ad0776 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow_en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow_en.png new file mode 100755 index 0000000000000000000000000000000000000000..d985c393e618e9b79df05e4ff0ae57ccc93744d0 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/cbow_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.en.png new file mode 100755 index 0000000000000000000000000000000000000000..2e16ab2f443732b8ef5404a8e7cd2457bc5eee23 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.png new file mode 100644 index 0000000000000000000000000000000000000000..2449dce6a86b43b1b997ff418ed0dba56848463f Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/ngram.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm.png new file mode 100644 index 0000000000000000000000000000000000000000..1e0b40a8f7aefdf46d42761305511f281c08e595 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm_en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm_en.png new file mode 100755 index 0000000000000000000000000000000000000000..158bd64b8f8729dea67834a8d591d21bce8b8564 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/nnlm_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/sentence_emb.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/sentence_emb.png new file mode 100644 index 0000000000000000000000000000000000000000..ce4a8bf4769183cbaff91793753d2350a3ce936c Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/sentence_emb.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram.png new file mode 100644 index 0000000000000000000000000000000000000000..a3ab385845d3dc8b5c670bae91225bc8dd47a8bb Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram_en.png b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram_en.png new file mode 100755 index 0000000000000000000000000000000000000000..3c36c6d1f66eb98ea78c0673965d02a4ee3aa288 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/basics/word2vec/image/skipgram_en.png differ diff --git a/doc/fluid/new_docs/beginners_guide/basics/word2vec/index.md b/doc/fluid/new_docs/beginners_guide/basics/word2vec/index.md new file mode 100644 index 0000000000000000000000000000000000000000..e73a6334ca1acd49379604f24d3d4e463192a902 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/basics/word2vec/index.md @@ -0,0 +1,440 @@ + +# 词向量 + +本教程源代码目录在[book/word2vec](https://github.com/PaddlePaddle/book/tree/develop/04.word2vec)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 + +本章我们介绍词的向量表征，也称为word embedding。词向量是自然语言处理中常见的一个操作，是搜索引擎、广告系统、推荐系统等互联网服务背后常见的基础技术。 + +在这些互联网服务里，我们经常要比较两个词或者两段文本之间的相关性。为了做这样的比较，我们往往先要把词表示成计算机适合处理的方式。最自然的方式恐怕莫过于向量空间模型(vector space model)。 +在这种方式里，每个词被表示成一个实数向量（one-hot vector），其长度为字典大小，每个维度对应一个字典里的每个词，除了这个词对应维度上的值是1，其他元素都是0。 + +One-hot vector虽然自然，但是用处有限。比如，在互联网广告系统里，如果用户输入的query是“母亲节”，而有一个广告的关键词是“康乃馨”。虽然按照常理，我们知道这两个词之间是有联系的——母亲节通常应该送给母亲一束康乃馨；但是这两个词对应的one-hot vectors之间的距离度量，无论是欧氏距离还是余弦相似度(cosine similarity)，由于其向量正交，都认为这两个词毫无相关性。 得出这种与我们相悖的结论的根本原因是：每个词本身的信息量都太小。所以，仅仅给定两个词，不足以让我们准确判别它们是否相关。要想精确计算相关性，我们还需要更多的信息——从大量数据里通过机器学习方法归纳出来的知识。 + +在机器学习领域里，各种“知识”被各种模型表示，词向量模型(word embedding model)就是其中的一类。通过词向量模型可将一个 one-hot vector映射到一个维度更低的实数向量（embedding vector），如`$embedding(Mother's\ Day) = [0.3, 4.2, -1.5, ...], embedding(Carnation) = [0.2, 5.6, -2.3, ...]$`。在这个映射到的实数向量表示中，希望两个语义（或用法）上相似的词对应的词向量“更像”，这样如“母亲节”和“康乃馨”的对应词向量的余弦相似度就不再为零了。 + +词向量模型可以是概率模型、共生矩阵(co-occurrence matrix)模型或神经元网络模型。在用神经网络求词向量之前，传统做法是统计一个词语的共生矩阵`$X$`。`$X$`是一个`$|V| \times |V|$` 大小的矩阵，`$X_{ij}$`表示在所有语料中，词汇表`V`(vocabulary)中第i个词和第j个词同时出现的词数，`$|V|$`为词汇表的大小。对`$X$`做矩阵分解（如奇异值分解，Singular Value Decomposition \[[5](#参考文献)\]），得到的`$U$`即视为所有词的词向量： + +$$X = USV^T$$ + +但这样的传统做法有很多问题：
+1) 由于很多词没有出现，导致矩阵极其稀疏，因此需要对词频做额外处理来达到好的矩阵分解效果；
+2) 矩阵非常大，维度太高(通常达到`$10^6*10^6$`的数量级)；
+3) 需要手动去掉停用词（如although, a,...），不然这些频繁出现的词也会影响矩阵分解的效果。 + + +基于神经网络的模型不需要计算存储一个在全语料上统计的大表，而是通过学习语义信息得到词向量，因此能很好地解决以上问题。在本章里，我们将展示基于神经网络训练词向量的细节，以及如何用PaddlePaddle训练一个词向量模型。 + + +## 效果展示 + +本章中，当词向量训练好后，我们可以用数据可视化算法t-SNE\[[4](#参考文献)\]画出词语特征在二维上的投影（如下图所示）。从图中可以看出，语义相关的词语（如a, the, these; big, huge）在投影上距离很近，语意无关的词（如say, business; decision, japan）在投影上的距离很远。 + +![2d_similarity](./image/2d_similarity.png) +

+图1. 词向量的二维投影 +

+ +另一方面，我们知道两个向量的余弦值在`$[-1,1]$`的区间内：两个完全相同的向量余弦值为1, 两个相互垂直的向量之间余弦值为0，两个方向完全相反的向量余弦值为-1，即相关性和余弦值大小成正比。因此我们还可以计算两个词向量的余弦相似度: + +``` +similarity: 0.899180685161 +please input two words: big huge + +please input two words: from company +similarity: -0.0997506977351 +``` + +以上结果可以通过运行`calculate_dis.py`, 加载字典里的单词和对应训练特征结果得到，我们将在[应用模型](#应用模型)中详细描述用法。 + + +## 模型概览 + +在这里我们介绍三个训练词向量的模型：N-gram模型，CBOW模型和Skip-gram模型，它们的中心思想都是通过上下文得到一个词出现的概率。对于N-gram模型，我们会先介绍语言模型的概念，并在之后的[训练模型](#训练模型)中，带大家用PaddlePaddle实现它。而后两个模型，是近年来最有名的神经元词向量模型，由 Tomas Mikolov 在Google 研发\[[3](#参考文献)\]，虽然它们很浅很简单，但训练效果很好。 + +### 语言模型 + +在介绍词向量模型之前，我们先来引入一个概念：语言模型。 +语言模型旨在为语句的联合概率函数`$P(w_1, ..., w_T)$`建模, 其中`$w_i$`表示句子中的第i个词。语言模型的目标是，希望模型对有意义的句子赋予大概率，对没意义的句子赋予小概率。 +这样的模型可以应用于很多领域，如机器翻译、语音识别、信息检索、词性标注、手写识别等，它们都希望能得到一个连续序列的概率。 以信息检索为例，当你在搜索“how long is a football bame”时（bame是一个医学名词），搜索引擎会提示你是否希望搜索"how long is a football game", 这是因为根据语言模型计算出“how long is a football bame”的概率很低，而与bame近似的，可能引起错误的词中，game会使该句生成的概率最大。 + +对语言模型的目标概率`$P(w_1, ..., w_T)$`，如果假设文本中每个词都是相互独立的，则整句话的联合概率可以表示为其中所有词语条件概率的乘积，即： + +$$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t)$$ + +然而我们知道语句中的每个词出现的概率都与其前面的词紧密相关, 所以实际上通常用条件概率表示语言模型： + +$$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t | w_1, ... , w_{t-1})$$ + + + +### N-gram neural model + +在计算语言学中，n-gram是一种重要的文本表示方法，表示一个文本中连续的n个项。基于具体的应用场景，每一项可以是一个字母、单词或者音节。 n-gram模型也是统计语言模型中的一种重要方法，用n-gram训练语言模型时，一般用每个n-gram的历史n-1个词语组成的内容来预测第n个词。 + +Yoshua Bengio等科学家就于2003年在著名论文 Neural Probabilistic Language Models \[[1](#参考文献)\] 中介绍如何学习一个神经元网络表示的词向量模型。文中的神经概率语言模型（Neural Network Language Model，NNLM）通过一个线性映射和一个非线性隐层连接，同时学习了语言模型和词向量，即通过学习大量语料得到词语的向量表达，通过这些向量得到整个句子的概率。用这种方法学习语言模型可以克服维度灾难（curse of dimensionality）,即训练和测试数据不同导致的模型不准。注意：由于“神经概率语言模型”说法较为泛泛，我们在这里不用其NNLM的本名，考虑到其具体做法，本文中称该模型为N-gram neural model。 + +我们在上文中已经讲到用条件概率建模语言模型，即一句话中第`$t$`个词的概率和该句话的前`$t-1$`个词相关。可实际上越远的词语其实对该词的影响越小，那么如果考虑一个n-gram, 每个词都只受其前面`n-1`个词的影响，则有： + +$$P(w_1, ..., w_T) = \prod_{t=n}^TP(w_t|w_{t-1}, w_{t-2}, ..., w_{t-n+1})$$ + +给定一些真实语料，这些语料中都是有意义的句子，N-gram模型的优化目标则是最大化目标函数: + +$$\frac{1}{T}\sum_t f(w_t, w_{t-1}, ..., w_{t-n+1};\theta) + R(\theta)$$ + +其中`$f(w_t, w_{t-1}, ..., w_{t-n+1})$`表示根据历史n-1个词得到当前词`$w_t$`的条件概率，`$R(\theta)$`表示参数正则项。 + +![nnlm](./image/nnlm.png) +

+图2. N-gram神经网络模型 +

+ +图2展示了N-gram神经网络模型，从下往上看，该模型分为以下几个部分： +- 对于每个样本，模型输入`$w_{t-n+1},...w_{t-1}$`, 输出句子第t个词为字典中`|V|`个词的概率。 + +每个输入词`$w_{t-n+1},...w_{t-1}$`首先通过映射矩阵映射到词向量`$C(w_{t-n+1}),...C(w_{t-1})$`。 + +- 然后所有词语的词向量连接成一个大向量，并经过一个非线性映射得到历史词语的隐层表示： + +$$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$ + +其中，`$x$`为所有词语的词向量连接成的大向量，表示文本历史特征；`$\theta$`、`$U$`、`$b_1$`、`$b_2$`和`$W$`分别为词向量层到隐层连接的参数。`$g$`表示未经归一化的所有输出单词概率，`$g_i$`表示未经归一化的字典中第`$i$`个单词的输出概率。 + +- 根据softmax的定义，通过归一化`$g_i$`, 生成目标词`$w_t$`的概率为： + +$$P(w_t | w_1, ..., w_{t-n+1}) = \frac{e^{g_{w_t}}}{\sum_i^{|V|} e^{g_i}}$$ + +- 整个网络的损失值(cost)为多类分类交叉熵，用公式表示为 + +$$J(\theta) = -\sum_{i=1}^N\sum_{c=1}^{|V|}y_k^{i}log(softmax(g_k^i))$$ + +其中`$y_k^i$`表示第`$i$`个样本第`$k$`类的真实标签(0或1)，`$softmax(g_k^i)$`表示第i个样本第k类softmax输出的概率。 + + + +### Continuous Bag-of-Words model(CBOW) + +CBOW模型通过一个词的上下文（各N个词）预测当前词。当N=2时，模型如下图所示： + +![cbow](./image/cbow.png) +

+图3. CBOW模型 +

+ +具体来说，不考虑上下文的词语输入顺序，CBOW是用上下文词语的词向量的均值来预测当前词。即： + +$$context = \frac{x_{t-1} + x_{t-2} + x_{t+1} + x_{t+2}}{4}$$ + +其中`$x_t$`为第`$t$`个词的词向量，分类分数（score）向量 `$z=U*context$`，最终的分类`$y$`采用softmax，损失函数采用多类分类交叉熵。 + +### Skip-gram model + +CBOW的好处是对上下文词语的分布在词向量上进行了平滑，去掉了噪声，因此在小数据集上很有效。而Skip-gram的方法中，用一个词预测其上下文，得到了当前词上下文的很多样本，因此可用于更大的数据集。 + +![skipgram](./image/skipgram.png) +

+图4. Skip-gram模型 +

+ +如上图所示，Skip-gram模型的具体做法是，将一个词的词向量映射到`$2n$`个词的词向量（`$2n$`表示当前输入词的前后各`$n$`个词），然后分别通过softmax得到这`$2n$`个词的分类损失值之和。 + + +## 数据准备 + +### 数据介绍 + +本教程使用Penn Treebank （PTB）（经Tomas Mikolov预处理过的版本）数据集。PTB数据集较小，训练速度快，应用于Mikolov的公开语言模型训练工具\[[2](#参考文献)\]中。其统计情况如下： + +

+ + + + + + + + + + + + + + + + +
训练数据验证数据测试数据
ptb.train.txtptb.valid.txtptb.test.txt
42068句3370句3761句
+

+ + +### 数据预处理 + +本章训练的是5-gram模型，表示在PaddlePaddle训练时，每条数据的前4个词用来预测第5个词。PaddlePaddle提供了对应PTB数据集的python包`paddle.dataset.imikolov`，自动做数据的下载与预处理，方便大家使用。 + +预处理会把数据集中的每一句话前后加上开始符号``以及结束符号``。然后依据窗口大小（本教程中为5），从头到尾每次向右滑动窗口并生成一条数据。 + +如"I have a dream that one day" 一句提供了5条数据： + +```text + I have a dream +I have a dream that +have a dream that one +a dream that one day +dream that one day +``` + +最后，每个输入会按其单词次在字典里的位置，转化成整数的索引序列，作为PaddlePaddle的输入。 + +## 编程实现 + +本配置的模型结构如下图所示： + +![ngram](./image/ngram.png) +

+图5. 模型配置中的N-gram神经网络模型 +

+ +首先，加载所需要的包： + +```python +import paddle +import paddle.fluid as fluid +import numpy +from functools import partial +import math +import os +import sys +``` + +然后，定义参数： +```python +EMBED_SIZE = 32 # word vector dimension +HIDDEN_SIZE = 256 # hidden layer dimension +N = 5 # train 5-gram +BATCH_SIZE = 32 # batch size + +# can use CPU or GPU +use_cuda = os.getenv('WITH_GPU', '0') != '0' + +word_dict = paddle.dataset.imikolov.build_dict() +dict_size = len(word_dict) +``` + +不同于之前的PaddlePaddle v2版本，在新的Fluid版本里，我们不必再手动计算词向量。PaddlePaddle提供了一个内置的方法`fluid.layers.embedding`，我们就可以直接用它来构造 N-gram 神经网络。 + +- 我们来定义我们的 N-gram 神经网络结构。这个结构在训练和预测中都会使用到。因为词向量比较稀疏，我们传入参数 `is_sparse == True`, 可以加速稀疏矩阵的更新。 + +```python +def inference_program(is_sparse): +first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64') +second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64') +third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64') +fourth_word = fluid.layers.data(name='fourthw', shape=[1], dtype='int64') + +embed_first = fluid.layers.embedding( +input=first_word, +size=[dict_size, EMBED_SIZE], +dtype='float32', +is_sparse=is_sparse, +param_attr='shared_w') +embed_second = fluid.layers.embedding( +input=second_word, +size=[dict_size, EMBED_SIZE], +dtype='float32', +is_sparse=is_sparse, +param_attr='shared_w') +embed_third = fluid.layers.embedding( +input=third_word, +size=[dict_size, EMBED_SIZE], +dtype='float32', +is_sparse=is_sparse, +param_attr='shared_w') +embed_fourth = fluid.layers.embedding( +input=fourth_word, +size=[dict_size, EMBED_SIZE], +dtype='float32', +is_sparse=is_sparse, +param_attr='shared_w') + +concat_embed = fluid.layers.concat( +input=[embed_first, embed_second, embed_third, embed_fourth], axis=1) +hidden1 = fluid.layers.fc(input=concat_embed, +size=HIDDEN_SIZE, +act='sigmoid') +predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax') +return predict_word +``` + +- 基于以上的神经网络结构，我们可以如下定义我们的`训练`方法 + +```python +def train_program(is_sparse): +# The declaration of 'next_word' must be after the invoking of inference_program, +# or the data input order of train program would be [next_word, firstw, secondw, +# thirdw, fourthw], which is not correct. +predict_word = inference_program(is_sparse) +next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') +cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) +avg_cost = fluid.layers.mean(cost) +return avg_cost +``` + +- 现在我们可以开始训练啦。如今的版本较之以前就简单了许多。我们有现成的训练和测试集：`paddle.dataset.imikolov.train()`和`paddle.dataset.imikolov.test()`。两者都会返回一个读取器。在PaddlePaddle中，读取器是一个Python的函数，每次调用，会读取下一条数据。它是一个Python的generator。 + +`paddle.batch` 会读入一个读取器，然后输出一个批次化了的读取器。`event_handler`亦可以一并传入`trainer.train`来时不时的输出每个步骤，批次的训练情况。 + +```python +def optimizer_func(): +# Note here we need to choose more sophisticated optimizers +# such as AdaGrad with a decay rate. The normal SGD converges +# very slowly. +# optimizer=fluid.optimizer.SGD(learning_rate=0.001), +return fluid.optimizer.AdagradOptimizer( +learning_rate=3e-3, +regularization=fluid.regularizer.L2DecayRegularizer(8e-4)) + + +def train(use_cuda, train_program, params_dirname): +train_reader = paddle.batch( +paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) +test_reader = paddle.batch( +paddle.dataset.imikolov.test(word_dict, N), BATCH_SIZE) + +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + +def event_handler(event): +if isinstance(event, fluid.EndStepEvent): +# We output cost every 10 steps. +if event.step % 10 == 0: +outs = trainer.test( +reader=test_reader, +feed_order=['firstw', 'secondw', 'thirdw', 'fourthw', 'nextw']) +avg_cost = outs[0] + +print "Step %d: Average Cost %f" % (event.step, avg_cost) + +# If average cost is lower than 5.8, we consider the model good enough to stop. +# Note 5.8 is a relatively high value. In order to get a better model, one should +# aim for avg_cost lower than 3.5. But the training could take longer time. +if avg_cost < 5.8: +trainer.save_params(params_dirname) +trainer.stop() + +if math.isnan(avg_cost): +sys.exit("got NaN loss, training failed.") + +trainer = fluid.Trainer( +train_func=train_program, +optimizer_func=optimizer_func, +place=place) + +trainer.train( +reader=train_reader, +num_epochs=1, +event_handler=event_handler, +feed_order=['firstw', 'secondw', 'thirdw', 'fourthw', 'nextw']) +``` + +- `trainer.train`将会开始训练。从`event_handler`返回的监控情况如下： + +```python +Step 0: Average Cost 7.337213 +Step 10: Average Cost 6.136128 +Step 20: Average Cost 5.766995 +... +``` + +## 模型应用 +在模型训练后，我们可以用它做一些预测。 + +### 预测下一个词 +我们可以用我们训练过的模型，在得知之前的 N-gram 后，预测下一个词。 + +```python +def infer(use_cuda, inference_program, params_dirname=None): +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +inferencer = fluid.Inferencer( +infer_func=inference_program, param_path=params_dirname, place=place) + +# Setup inputs by creating 4 LoDTensors representing 4 words. Here each word +# is simply an index to look up for the corresponding word vector and hence +# the shape of word (base_shape) should be [1]. The length-based level of +# detail (lod) info of each LoDtensor should be [[1]] meaning there is only +# one lod_level and there is only one sequence of one word on this level. +# Note that lod info should be a list of lists. + +data1 = [[211]] # 'among' +data2 = [[6]] # 'a' +data3 = [[96]] # 'group' +data4 = [[4]] # 'of' +lod = [[1]] + +first_word = fluid.create_lod_tensor(data1, lod, place) +second_word = fluid.create_lod_tensor(data2, lod, place) +third_word = fluid.create_lod_tensor(data3, lod, place) +fourth_word = fluid.create_lod_tensor(data4, lod, place) + +result = inferencer.infer( +{ +'firstw': first_word, +'secondw': second_word, +'thirdw': third_word, +'fourthw': fourth_word +}, +return_numpy=False) + +print(numpy.array(result[0])) +most_possible_word_index = numpy.argmax(result[0]) +print(most_possible_word_index) +print([ +key for key, value in word_dict.iteritems() +if value == most_possible_word_index +][0]) +``` + +在经历3分钟的短暂训练后，我们得到如下的预测。我们的模型预测 `among a group of` 的下一个词是`a`。这比较符合文法规律。如果我们训练时间更长，比如几个小时，那么我们会得到的下一个预测是 `workers`。 + + +```python +[[0.00106646 0.0007907 0.00072041 ... 0.00049024 0.00041355 0.00084464]] +6 +a +``` + +整个程序的入口很简单： + +```python +def main(use_cuda, is_sparse): +if use_cuda and not fluid.core.is_compiled_with_cuda(): +return + +params_dirname = "word2vec.inference.model" + +train( +use_cuda=use_cuda, +train_program=partial(train_program, is_sparse), +params_dirname=params_dirname) + +infer( +use_cuda=use_cuda, +inference_program=partial(inference_program, is_sparse), +params_dirname=params_dirname) + + +main(use_cuda=use_cuda, is_sparse=True) +``` + + +## 总结 +本章中，我们介绍了词向量、语言模型和词向量的关系、以及如何通过训练神经网络模型获得词向量。在信息检索中，我们可以根据向量间的余弦夹角，来判断query和文档关键词这二者间的相关性。在句法分析和语义分析中，训练好的词向量可以用来初始化模型，以得到更好的效果。在文档分类中，有了词向量之后，可以用聚类的方法将文档中同义词进行分组，也可以用 N-gram 来预测下一个词。希望大家在本章后能够自行运用词向量进行相关领域的研究。 + + +## 参考文献 +1. Bengio Y, Ducharme R, Vincent P, et al. [A neural probabilistic language model](http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf)[J]. journal of machine learning research, 2003, 3(Feb): 1137-1155. +2. Mikolov T, Kombrink S, Deoras A, et al. [Rnnlm-recurrent neural network language modeling toolkit](http://www.fit.vutbr.cz/~imikolov/rnnlm/rnnlm-demo.pdf)[C]//Proc. of the 2011 ASRU Workshop. 2011: 196-201. +3. Mikolov T, Chen K, Corrado G, et al. [Efficient estimation of word representations in vector space](https://arxiv.org/pdf/1301.3781.pdf)[J]. arXiv preprint arXiv:1301.3781, 2013. +4. Maaten L, Hinton G. [Visualizing data using t-SNE](https://lvdmaaten.github.io/publications/papers/JMLR_2008.pdf)[J]. Journal of Machine Learning Research, 2008, 9(Nov): 2579-2605. +5. https://en.wikipedia.org/wiki/Singular_value_decomposition + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/index.rst b/doc/fluid/new_docs/beginners_guide/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..e18933dcc0038129077a455892ddd785579f0003 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/index.rst @@ -0,0 +1,15 @@ +######## +新手入门 +######## + +.. todo:: + + 新手入门的导引文字，需要完善。 + +.. toctree:: + :maxdepth: 2 + + install/install_doc.rst + quick_start/index.rst + basics/index.rst + basics/learning_materials.md diff --git a/doc/fluid/new_docs/beginners_guide/install/install_doc.rst b/doc/fluid/new_docs/beginners_guide/install/install_doc.rst new file mode 100644 index 0000000000000000000000000000000000000000..8a66a95f45ea18dbfdc2450694517d5df8c47efd --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/install/install_doc.rst @@ -0,0 +1,543 @@ +.. _how_to_install: + +安装说明 +^^^^^^^^ + +若您的系统为Linux或Windows，您可以使用我们提供的安装包来安装PaddlePaddle。 + +对于MacOS系统，我们暂未提供安装包，您可以使用 **从源码编译** 的方式安装。 + + +.. _install_linux: + +在Linux安装PaddlePaddle +-------- + +推荐您使用 `pip `_ +安装，它是Linux系统下最简单的安装方式。 + +注意事项： + +- PaddlePaddle Python API 依赖Python 2.7版本。 + +执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件。 + + .. code-block:: bash + + pip install paddlepaddle + +您可以通过指定版本号来安装其它版本，例如： + + .. code-block:: bash + + pip install paddlepaddle==0.13.0 + + +如果需要安装支持GPU的版本（cuda9.0_cudnn7_avx_openblas），需要执行： + + .. code-block:: bash + + pip install paddlepaddle-gpu + +PaddlePaddle针对不同需求提供了更多版本的安装包，部分列表如下： + +================================= ======================================== +版本号 版本说明 +================================= ======================================== +paddlepaddle-gpu==0.14.0 使用CUDA 9.0和cuDNN 7编译的0.14.0版本 +paddlepaddle-gpu==0.14.0.post87 使用CUDA 8.0和cuDNN 7编译的0.14.0版本 +paddlepaddle-gpu==0.14.0.post85 使用CUDA 8.0和cuDNN 5编译的0.14.0版本 +paddlepaddle-gpu==0.13.0 使用CUDA 9.0和cuDNN 7编译的0.13.0版本 +paddlepaddle-gpu==0.12.0 使用CUDA 8.0和cuDNN 5编译的0.12.0版本 +paddlepaddle-gpu==0.11.0.post87 使用CUDA 8.0和cuDNN 7编译的0.11.0版本 +paddlepaddle-gpu==0.11.0.post8 使用CUDA 8.0和cuDNN 5编译的0.11.0版本 +paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版本 +================================= ======================================== + +您可以在 `Release History `_ +中找到paddlepaddle-gpu的各个发行版本。 + +如果需要获取并安装最新的PaddlePaddle开发分支，可以从我们的 `CI系统 `_ 中下载最新的whl安装包和c-api开发包并安装。如需登录，请点击“Log in as guest”。 + +.. _FAQ: + +安装常见问题和解决方法 +====================== + +- paddlepaddle*.whl is not a supported wheel on this platform. + +出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。 +请检查Python版本是否为2.7系列。另外最新的pip官方源中的安装包默认是manylinux1标准， +需要使用最新的pip (>9.0.0) 才可以安装。 + +可以使用下面的命令更新您的pip： + + .. code-block:: bash + + pip install --upgrade pip + +如果仍然存在问题，可以执行： + + .. code-block:: bash + + python -c "import pip; print(pip.pep425tags.get_supported())" + +获取当前系统支持的安装包格式，并检查和需安装的包是否匹配。pypi安装包 +可以在 `这里 `_ 找到。 + +如果系统支持的是 linux_x86_64 而安装包是 manylinux1_x86_64 ，需要升级pip版本到最新； +如果系统支持 manylinux1_x86_64 而安装包（本地）是 linux_x86_64， +可以重命名这个whl包为 manylinux1_x86_64 再安装。 + + +.. _install_windows: + +在Windows安装PaddlePaddle +------------------------------ +Windows系统需要通过Docker来使用PaddleaPaddle。Docker是一个虚拟容器，使用Docker可以简化复杂的环境配置工作。 + +我们提供了 `PaddlePaddle_Windows快速安装包 `_， +它能够帮助您安装Docker和PaddlePaddle。 + +* 安装包支持的系统：Windows7，Windows8的所有版本，Windows10的专业版、企业版。 + +* 如果您希望使用GPU提升训练速度，请使用Linux系统安装，Windows系统暂不支持。 + +.. _install_mac: + +在MacOS安装PaddlePaddle +-------- + +对于MacOS系统，我们暂未提供pip安装方式，您可以使用 **源码编译** 的方式安装。 + +.. _others: + +其他安装方式 +------------- + +.. _source: +源码编译（使用Docker镜像） +========== + +.. _requirements: + +需要的软硬件 +""""""""""""" + +为了编译PaddlePaddle，我们需要 + +1. 一台电脑，可以装的是 Linux, Windows 或者 MacOS 操作系统 +2. Docker + +不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要，因为我们会把所有编译工具都安装进一个 Docker 镜像里。 + +.. _build_step: + +编译方法 +""""""""""""" + +PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像可以在 `这里 `_ 找到。 + + +**I. 编译CPU-Only版本的PaddlePaddle，需要执行：** + +.. code-block:: bash + + # 1. 获取源码 + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + # 2. 执行如下命令下载最新版本的docker镜像 + docker run --name paddle-test -v $PWD:/paddle --network=host -it docker.paddlepaddlehub.com/paddle:latest-dev /bin/bash + # 3. 进入docker内执行如下命令编译CPU-Only的二进制安装包 + mkdir -p /paddle/build && cd /paddle/build + cmake .. -DWITH_FLUID_ONLY=ON -DWITH_GPU=OFF -DWITH_TESTING=OFF + make -j$(nproc) + +**II. 编译GPU版本的PaddlePaddle，需要执行：** + +.. code-block:: bash + + # 1. 获取源码 + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + # 2. 安装nvidia-docker + apt-get install nvidia-docker + # 3. 执行如下命令下载支持GPU运行的docker容器 + nvidia-docker run --name paddle-test-gpu -v $PWD:/paddle --network=host -it docker.paddlepaddlehub.com/paddle:latest-dev /bin/bash + # 4. 进入docker内执行如下命令编译GPU版本的PaddlePaddle + mkdir -p /paddle/build && cd /paddle/build + cmake .. -DWITH_FLUID_ONLY=ON -DWITH_GPU=ON -DWITH_TESTING=OFF + make -j$(nproc) + +**注意事项：** + +* 上述有关 :code:`docker` 的命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。 +* 进入 :code:`docker` 后执行 :code:`cmake` 命令，若是出现 :code:`patchelf not found, please install it.` 错误，则执行 :code:`apt-get install -y patchelf` 命令即可解决问题。 +* 若您在使用Docker编译PaddlePaddle遇到问题时， `这个issue `_ 可能会对您有所帮助。 + + +.. _source: +源码编译（不使用Docker镜像） +========== + +如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `附录：编译依赖`_ 之后才能开始编译的步骤。 + +.. _build_step: + +编译方法 +""""""""""""" + +在本机上编译CPU-Only版本的PaddlePaddle，需要执行如下命令： + +.. code-block:: bash + + # 1. 使用virtualenvwrapper创建python虚环境并将工作空间切换到虚环境 [可选] + mkvirtualenv paddle-venv + workon paddle-venv + # 2. 获取源码 + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + # 3. 执行下面的命令编译CPU-Only的二进制 + mkdir build && cd build + cmake .. -DWITH_FLUID_ONLY=ON -DWITH_GPU=OFF -DWITH_TESTING=OFF + make -j4 # 根据机器配备CPU的核心数开启相应的多线程进行编译 + + +**注意事项：** + +* MacOS系统下因为默认安装了cblas库，所以编译时可能会遇到 :code:`use of undeclared identifier 'openblas_set_num_threads'` 错误。因此，在执行cmake命令时需要指定所使用openblas库的头文件路径，具体操作如下： + + .. code-block:: bash + + cd Paddle/build && rm -rf * + cmake .. -DWITH_FLUID_ONLY=ON -DWITH_GPU=OFF -DWITH_TESTING=OFF -DOPENBLAS_INC_DIR=/usr/local/Cellar/openblas/[本机所安装的openblas版本号]/include/ + make -j4 # 根据机器配备CPU的核心数开启相应的多线程进行编译 +* 若您在MacOS系统下从源码编译PaddlePaddle遇到问题时， `这个issue `_ 可能会对您有所帮助。 + +编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装： + +.. code-block:: bash + + pip install build/python/dist/*.whl + +如果机器中已经安装过PaddlePaddle，有两种方法： + +.. code-block:: bash + + 1. 先卸载之前的版本，再重新安装 + pip uninstall paddlepaddle + pip install build/python/dist/*.whl + + 2. 直接升级到更新的版本 + pip install build/python/dist/*.whl -U + +.. _run_test: + +执行单元测试 +""""""""""""" + +如果您期望在编译完成后立即执行所有的单元测试，可以按照下面的方法： + +设置 :code:`RUN_TEST=ON` 和 :code:`WITH_TESTING=ON` 就会在完成编译之后，立即执行单元测试。 +开启 :code:`WITH_GPU=ON` 可以指定同时执行GPU上的单元测试。 + +.. code-block:: bash + + docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" docker.paddlepaddlehub.com/paddle:latest-dev bash -x /paddle/paddle/scripts/paddle_build.sh build + +如果期望执行其中一个单元测试，（比如 :code:`test_sum_op` ）： + +.. code-block:: bash + + docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" docker.paddlepaddlehub.com/paddle:latest-dev bash -x /paddle/paddle/scripts/paddle_build.sh build + cd /paddle/build + ctest -R test_sum_op -V + +.. _faq_docker: + +常见问题 +""""""""""""" + +- 什么是 Docker? + + 如果您没有听说 Docker，可以把它想象为一个类似 virtualenv 的系统，但是虚拟的不仅仅是 Python 的运行环境。 + +- Docker 还是虚拟机？ + + 有人用虚拟机来类比 Docker。需要强调的是：Docker 不会虚拟任何硬件，Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的，性能和把编译工具安装在本机运行一样。 + +- 为什么用 Docker? + + 把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题，其他人可以复现问题以便帮助。 + + 另外，对于习惯使用Windows和MacOS的开发者来说，使用Docker就不用配置交叉编译环境了。 + +- 可以选择不用Docker吗？ + + 当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式，把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程，是因为这个流程比其他方法都更简便。 + +- 学习 Docker 有多难？ + + 理解 Docker 并不难，大概花十分钟看一下 `这篇文章 `_。 + 这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 + +- 可以用 IDE 吗？ + + 当然可以，因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码，我们只需要配置 IDE 来调用 Docker 命令编译源码即可。 + + 很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行 + + .. code-block:: bash + + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + + 就可以按 `Ctrl-C` 和 `c` 键来启动编译了。 + +- 可以并行编译吗？ + + 是的。我们的 Docker image 运行一个 `Bash 脚本 `_。这个脚本调用 :code:`make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 + +- Docker 需要 sudo + + 如果用自己的电脑开发，自然也就有管理员权限（sudo）了。如果用公用的电脑开发，需要请管理员安装和配置好 Docker。此外，PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术，比如 rkt。 + +- 在 Windows/MacOS 上编译很慢 + + Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考 `这个issue `_。 + +- 磁盘不够 + + 本文中的例子里， :code:`docker run` 命令里都用了 :code:`--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 :code:`docker ps -a` 命令看到停止后但是没有删除的 containers。 :code:`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考 `这篇文章 `_ 来清理这些内容。 + + +.. _compile_deps: + +附录：编译依赖 +""""""""""""" + +PaddlePaddle编译需要使用到下面的依赖（包含但不限于），其他的依赖软件，会自动在编译时下载。 + +.. csv-table:: PaddlePaddle编译依赖 + :header: "依赖", "版本", "说明" + :widths: 10, 15, 30 + + "CMake", "3.4", "" + "GCC", "4.8.2", "推荐使用CentOS的devtools2" + "Python", "2.7.x", "依赖libpython2.7.so" + "SWIG", ">=2.0", "" + "wget","","" + "openblas","","" + "pip", ">=9.0", "" + "numpy", "", "" + "protobuf","3.1.0","" + "wheel","","" + "Go", ">=1.8", "可选" + + +.. _build_options: + +附录：编译选项 +""""""""""""" + +PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。 +用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考 +`官方文档 `_ 。 + +在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如： + +.. code-block:: bash + + cmake .. -DWITH_GPU=OFF + +.. csv-table:: 编译选项说明 + :header: "选项", "说明", "默认值" + :widths: 1, 7, 2 + + "WITH_GPU", "是否支持GPU", "ON" + "WITH_C_API", "是否仅编译CAPI", "OFF" + "WITH_DOUBLE", "是否使用双精度浮点数", "OFF" + "WITH_DSO", "是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。", "ON" + "WITH_AVX", "是否编译含有AVX指令集的PaddlePaddle二进制文件", "ON" + "WITH_PYTHON", "是否内嵌PYTHON解释器", "ON" + "WITH_STYLE_CHECK", "是否编译时进行代码风格检查", "ON" + "WITH_TESTING", "是否开启单元测试", "OFF" + "WITH_DOC", "是否编译中英文文档", "OFF" + "WITH_SWIG_PY", "是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练", "Auto" + "WITH_GOLANG", "是否编译go语言的可容错parameter server", "OFF" + "WITH_MKL", "是否使用MKL数学库，如果为否则是用OpenBLAS", "ON" + +BLAS ++++++ + +PaddlePaddle支持 `MKL `_ 和 +`OpenBlAS `_ 两种BLAS库。默认使用MKL。如果使用MKL并且机器含有AVX2指令集， +还会下载MKL-DNN数学库，详细参考 `这里 `_ 。 + +如果关闭MKL，则会使用OpenBLAS作为BLAS库。 + +CUDA/cuDNN ++++++++++++ + +PaddlePaddle在编译时/运行时会自动找到系统中安装的CUDA和cuDNN库进行编译和执行。 +使用参数 :code:`-DCUDA_ARCH_NAME=Auto` 可以指定开启自动检测SM架构，加速编译。 + +PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cuDNN是同一个版本。 +我们推荐使用最新版本的cuDNN。 + +编译选项的设置 +++++++++++++++ + +PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 + +.. code-block:: bash + + cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5 + +注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（ :code:`rm -rf` ）后，再指定。 + +.. _install_docker: + +使用Docker安装运行 +================== + +使用Docker安装和运行PaddlePaddle可以无需考虑依赖环境。 +您可以在 `Docker官网 `_ +获得基本的Docker安装和使用方法。 + +在了解Docker的基本使用方法之后，即可开始下面的步骤： + +.. _docker_pull: + +获取PaddlePaddle的Docker镜像 +"""""""""""""""""""""""""""" + +执行下面的命令获取最新的PaddlePaddle Docker镜像，版本为cpu_avx_mkl： + + .. code-block:: bash + + docker pull paddlepaddle/paddle + +对于国内用户，我们提供了加速访问的镜像源： + + .. code-block:: bash + + docker pull docker.paddlepaddlehub.com/paddle + +下载GPU版本（cuda8.0_cudnn5_avx_mkl）的Docker镜像： + + .. code-block:: bash + + docker pull paddlepaddle/paddle:latest-gpu + docker pull docker.paddlepaddlehub.com/paddle:latest-gpu + +选择下载使用不同的BLAS库的Docker镜像： + + .. code-block:: bash + + # 默认是使用MKL的镜像 + docker pull paddlepaddle/paddle + # 使用OpenBLAS的镜像 + docker pull paddlepaddle/paddle:latest-openblas + +下载指定版本的Docker镜像，可以从 `DockerHub网站 `_ 获取可选的tag，并执行下面的命令： + + .. code-block:: bash + + docker pull paddlepaddle/paddle:[tag] + # 比如： + docker pull docker.paddlepaddlehub.com/paddle:0.11.0-gpu + +.. _docker_run: + +在Docker中执行PaddlePaddle训练程序 +""""""""""""""""""""""""""""""""""" + +假设您已经在当前目录（比如在/home/work）编写了一个PaddlePaddle的程序 :code:`train.py` （可以参考 +`PaddlePaddleBook `_ +编写），就可以使用下面的命令开始执行训练： + + .. code-block:: bash + + cd /home/work + docker run -it -v $PWD:/work paddlepaddle/paddle /work/train.py + +上述命令中， :code:`-it` 参数说明容器已交互式运行； :code:`-v $PWD:/work` +指定将当前路径（Linux中$PWD变量会展开为当前路径的绝对路径）挂载到容器内部的 :code:`/work` +目录； :code:`paddlepaddle/paddle` 指定需要使用的容器； 最后 :code:`/work/train.py` +为容器内执行的命令，即运行训练程序。 + +当然，您也可以进入到Docker容器中，以交互式的方式执行或调试您的代码： + + .. code-block:: bash + docker run -it -v $PWD:/work paddlepaddle/paddle /bin/bash + cd /work + python train.py + +**注：PaddlePaddle Docker镜像为了减小体积，默认没有安装vim，您可以在容器中执行** :code:`apt-get install -y vim` **安装后，在容器中编辑代码。** + +.. _docker_run_book: + +使用Docker启动PaddlePaddle Book教程 +"""""""""""""""""""""""""""""""""""" + +使用Docker可以快速在本地启动一个包含了PaddlePaddle官方Book教程的Jupyter Notebook，可以通过网页浏览。 +PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。 +如果您想要更深入了解deep learning，PaddlePaddle Book一定是您最好的选择。 +大家可以通过它阅读教程，或者制作和分享带有代码、公式、图表、文字的交互式文档。 + +我们提供可以直接运行PaddlePaddle Book的Docker镜像，直接运行： + + .. code-block:: bash + + docker run -p 8888:8888 paddlepaddle/book + +国内用户可以使用下面的镜像源来加速访问： + + .. code-block: bash + + docker run -p 8888:8888 docker.paddlepaddlehub.com/book + +然后在浏览器中输入以下网址： + + .. code-block:: text + + http://localhost:8888/ + +就这么简单，享受您的旅程！ + +.. _docker_run_gpu: + +使用Docker执行GPU训练 +"""""""""""""""""""""""""""" + +为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用 +`nvidia-docker `_ 来运行镜像。 +请不要忘记提前在物理机上安装GPU最新驱动。 + + .. code-block:: bash + + nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash + +**注: 如果没有安装nvidia-docker，可以尝试以下的方法，将CUDA库和Linux设备挂载到Docker容器内：** + + .. code-block:: bash + + export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" + export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') + docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu + +**关于AVX：** + +AVX是一种CPU指令集，可以加速PaddlePaddle的计算。最新的PaddlePaddle Docker镜像默认 +是开启AVX编译的，所以，如果您的电脑不支持AVX，需要单独 +`编译 <./build_from_source_cn.html>`_ PaddlePaddle为no-avx版本。 + +以下指令能检查Linux电脑是否支持AVX： + + .. code-block:: bash + + if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi + +如果输出是No，就需要选择使用no-AVX的镜像 diff --git a/doc/fluid/new_docs/beginners_guide/install/paddleci.png b/doc/fluid/new_docs/beginners_guide/install/paddleci.png new file mode 100644 index 0000000000000000000000000000000000000000..16087ce059aa3c07ce8c927d983eb86351915825 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/install/paddleci.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md new file mode 100644 index 0000000000000000000000000000000000000000..ba43ada5100ed1db7192de9c795b4b8a6596d705 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/README.cn.md @@ -0,0 +1,329 @@ +```eval_rst +.. _quick_start_fit_a_line: +``` +# 线性回归 +让我们从经典的线性回归（Linear Regression \[[1](#参考文献)\]）模型开始这份教程。在这一章里，你将使用真实的数据集建立起一个房价预测模型，并且了解到机器学习中的若干重要概念。 + +本教程源代码目录在[book/fit_a_line](https://github.com/PaddlePaddle/book/tree/develop/01.fit_a_line)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 +给定一个大小为`$n$`的数据集 `${\{y_{i}, x_{i1}, ..., x_{id}\}}_{i=1}^{n}$`，其中`$x_{i1}, \ldots, x_{id}$`是第`$i$`个样本`$d$`个属性上的取值，`$y_i$`是该样本待预测的目标。线性回归模型假设目标`$y_i$`可以被属性间的线性组合描述，即 + +$$y_i = \omega_1x_{i1} + \omega_2x_{i2} + \ldots + \omega_dx_{id} + b, i=1,\ldots,n$$ + +例如，在我们将要建模的房价预测问题里，`$x_{ij}$`是描述房子`$i$`的各种属性（比如房间的个数、周围学校和医院的个数、交通状况等），而 `$y_i$`是房屋的价格。 + +初看起来，这个假设实在过于简单了，变量间的真实关系很难是线性的。但由于线性回归模型有形式简单和易于建模分析的优点，它在实际问题中得到了大量的应用。很多经典的统计学习、机器学习书籍\[[2,3,4](#参考文献)\]也选择对线性模型独立成章重点讲解。 + +## 效果展示 +我们使用从[UCI Housing Data Set](https://archive.ics.uci.edu/ml/datasets/Housing)获得的波士顿房价数据集进行模型的训练和预测。下面的散点图展示了使用模型对部分房屋价格进行的预测。其中，每个点的横坐标表示同一类房屋真实价格的中位数，纵坐标表示线性回归模型根据特征预测的结果，当二者值完全相等的时候就会落在虚线上。所以模型预测得越准确，则点离虚线越近。 + +![BostonHousePricePredictions](./image/predictions.png) +

图1. 预测值 V.S. 真实值

+ +## 模型概览 + +### 模型定义 + +在波士顿房价数据集中，和房屋相关的值共有14个：前13个用来描述房屋相关的各种信息，即模型中的 `$x_i$`；最后一个值为我们要预测的该类房屋价格的中位数，即模型中的 `$y_i$`。因此，我们的模型就可以表示成： + +$$\hat{Y} = \omega_1X_{1} + \omega_2X_{2} + \ldots + \omega_{13}X_{13} + b$$ + +`$\hat{Y}$` 表示模型的预测结果，用来和真实值`$Y$`区分。模型要学习的参数即：`$\omega_1, \ldots, \omega_{13}, b$`。 + +建立模型后，我们需要给模型一个优化目标，使得学到的参数能够让预测值`$\hat{Y}$`尽可能地接近真实值`$Y$`。这里我们引入损失函数（[Loss Function](https://en.wikipedia.org/wiki/Loss_function)，或Cost Function）这个概念。 输入任意一个数据样本的目标值`$y_{i}$`和模型给出的预测值`$\hat{y_{i}}$`，损失函数输出一个非负的实值。这个实值通常用来反映模型误差的大小。 + +对于线性回归模型来讲，最常见的损失函数就是均方误差（Mean Squared Error， [MSE](https://en.wikipedia.org/wiki/Mean_squared_error)）了，它的形式是： + +$$MSE=\frac{1}{n}\sum_{i=1}^{n}{(\hat{Y_i}-Y_i)}^2$$ + +即对于一个大小为`$n$`的测试集，`$MSE$`是`$n$`个数据预测结果误差平方的均值。 + +### 训练过程 + +定义好模型结构之后，我们要通过以下几个步骤进行模型训练 +1. 初始化参数，其中包括权重`$\omega_i$`和偏置`$b$`，对其进行初始化（如0均值，1方差）。 +2. 网络正向传播计算网络输出和损失函数。 +3. 根据损失函数进行反向误差传播 （[backpropagation](https://en.wikipedia.org/wiki/Backpropagation)），将网络误差从输出层依次向前传递, 并更新网络中的参数。 +4. 重复2~3步骤，直至网络训练误差达到规定的程度或训练轮次达到设定值。 + +## 数据集 + +### 数据集介绍 +这份数据集共506行，每行包含了波士顿郊区的一类房屋的相关信息及该类房屋价格的中位数。其各维属性的意义如下： + +

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
属性名解释类型
CRIM该镇的人均犯罪率连续值
ZN占地面积超过25,000平方呎的住宅用地比例连续值
INDUS非零售商业用地比例连续值
CHAS是否邻近 Charles River离散值，1=邻近；0=不邻近
NOX一氧化氮浓度连续值
RM每栋房屋的平均客房数连续值
AGE1940年之前建成的自用单位比例连续值
DIS到波士顿5个就业中心的加权距离连续值
RAD到径向公路的可达性指数连续值
TAX全值财产税率连续值
PTRATIO学生与教师的比例连续值
B1000(BK - 0.63)^2，其中BK为黑人占比连续值
LSTAT低收入人群占比连续值
MEDV同类房屋价格的中位数连续值
+

+ +### 数据预处理 +#### 连续值与离散值 +观察一下数据，我们的第一个发现是：所有的13维属性中，有12维的连续值和1维的离散值（CHAS）。离散值虽然也常使用类似0、1、2这样的数字表示，但是其含义与连续值是不同的，因为这里的差值没有实际意义。例如，我们用0、1、2来分别表示红色、绿色和蓝色的话，我们并不能因此说“蓝色和红色”比“绿色和红色”的距离更远。所以通常对一个有`$d$`个可能取值的离散属性，我们会将它们转为`$d$`个取值为0或1的二值属性或者将每个可能取值映射为一个多维向量。不过就这里而言，因为CHAS本身就是一个二值属性，就省去了这个麻烦。 + +#### 属性的归一化 +另外一个稍加观察即可发现的事实是，各维属性的取值范围差别很大（如图2所示）。例如，属性B的取值范围是[0.32, 396.90]，而属性NOX的取值范围是[0.3850, 0.8170]。这里就要用到一个常见的操作-归一化（normalization）了。归一化的目标是把各位属性的取值范围放缩到差不多的区间，例如[-0.5,0.5]。这里我们使用一种很常见的操作方法：减掉均值，然后除以原取值范围。 + +做归一化（或 [Feature scaling](https://en.wikipedia.org/wiki/Feature_scaling)）至少有以下3个理由： +- 过大或过小的数值范围会导致计算时的浮点上溢或下溢。 +- 不同的数值范围会导致不同属性对模型的重要性不同（至少在训练的初始阶段如此），而这个隐含的假设常常是不合理的。这会对优化的过程造成困难，使训练时间大大的加长。 +- 很多的机器学习技巧/模型（例如L1，L2正则项，向量空间模型-Vector Space Model）都基于这样的假设：所有的属性取值都差不多是以0为均值且取值范围相近的。 + +![featureScale](./image/ranges.png) +

图2. 各维属性的取值范围

+ +#### 整理训练集与测试集 +我们将数据集分割为两份：一份用于调整模型的参数，即进行模型的训练，模型在这份数据集上的误差被称为**训练误差**；另外一份被用来测试，模型在这份数据集上的误差被称为**测试误差**。我们训练模型的目的是为了通过从训练数据中找到规律来预测未知的新数据，所以测试误差是更能反映模型表现的指标。分割数据的比例要考虑到两个因素：更多的训练数据会降低参数估计的方差，从而得到更可信的模型；而更多的测试数据会降低测试误差的方差，从而得到更可信的测试误差。我们这个例子中设置的分割比例为`$8:2$` + + +在更复杂的模型训练过程中，我们往往还会多使用一种数据集：验证集。因为复杂的模型中常常还有一些超参数（[Hyperparameter](https://en.wikipedia.org/wiki/Hyperparameter_optimization)）需要调节，所以我们会尝试多种超参数的组合来分别训练多个模型，然后对比它们在验证集上的表现选择相对最好的一组超参数，最后才使用这组参数下训练的模型在测试集上评估测试误差。由于本章训练的模型比较简单，我们暂且忽略掉这个过程。 + +## 训练 + +`fit_a_line/trainer.py`演示了训练的整体过程。 + +### 配置数据提供器(Datafeeder) +首先我们引入必要的库： +```python +import paddle +import paddle.fluid as fluid +import numpy +``` + +我们通过uci_housing模块引入了数据集合[UCI Housing Data Set](https://archive.ics.uci.edu/ml/datasets/Housing) + +其中，在uci_housing模块中封装了： + +1. 数据下载的过程。下载数据保存在~/.cache/paddle/dataset/uci_housing/housing.data。 +2. [数据预处理](#数据预处理)的过程。 + +接下来我们定义了用于训练和测试的数据提供器。提供器每次读入一个大小为`BATCH_SIZE`的数据批次。如果用户希望加一些随机性，她可以同时定义一个批次大小和一个缓存大小。这样的话，每次数据提供器会从缓存中随机读取批次大小那么多的数据。 + +```python +BATCH_SIZE = 20 + +train_reader = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.uci_housing.train(), buf_size=500), +batch_size=BATCH_SIZE) + +test_reader = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.uci_housing.test(), buf_size=500), +batch_size=BATCH_SIZE) +``` + +### 配置训练程序 +训练程序的目的是定义一个训练模型的网络结构。对于线性回归来讲，它就是一个从输入到输出的简单的全连接层。更加复杂的结果，比如卷积神经网络，递归神经网络等会在随后的章节中介绍。训练程序必须返回`平均损失`作为第一个返回值，因为它会被后面反向传播算法所用到。 + +```python +def train_program(): +y = fluid.layers.data(name='y', shape=[1], dtype='float32') + +# feature vector of length 13 +x = fluid.layers.data(name='x', shape=[13], dtype='float32') +y_predict = fluid.layers.fc(input=x, size=1, act=None) + +loss = fluid.layers.square_error_cost(input=y_predict, label=y) +avg_loss = fluid.layers.mean(loss) + +return avg_loss +``` + +### 定义运算场所 +我们可以定义运算是发生在CPU还是GPU + +```python +use_cuda = False +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() +``` + +### 创建训练器 +训练器会读入一个训练程序和一些必要的其他参数： + +```python +trainer = fluid.Trainer( +train_func=train_program, +place=place, +optimizer_func=fluid.optimizer.SGD(learning_rate=0.001)) +``` + +### 开始提供数据 +PaddlePaddle提供了读取数据者发生器机制来读取训练数据。读取数据者会一次提供多列数据，因此我们需要一个Python的list来定义读取顺序。 + +```python +feed_order=['x', 'y'] +``` + +除此之外，可以定义一个事件相应器来处理类似`打印训练进程`的事件： + +```python +# Specify the directory path to save the parameters +params_dirname = "fit_a_line.inference.model" + +# Plot data +from paddle.v2.plot import Ploter +train_title = "Train cost" +test_title = "Test cost" +plot_cost = Ploter(train_title, test_title) + +step = 0 + +# event_handler to print training and testing info +def event_handler_plot(event): +global step +if isinstance(event, fluid.EndStepEvent): +if event.step % 10 == 0: # every 10 batches, record a test cost +test_metrics = trainer.test( +reader=test_reader, feed_order=feed_order) + +plot_cost.append(test_title, step, test_metrics[0]) +plot_cost.plot() + +if test_metrics[0] < 10.0: +# If the accuracy is good enough, we can stop the training. +print('loss is less than 10.0, stop') +trainer.stop() + +# We can save the trained parameters for the inferences later +if params_dirname is not None: +trainer.save_params(params_dirname) + +step += 1 +``` + +### 开始训练 +我们现在可以通过调用`trainer.train()`来开始训练 + +```python +%matplotlib inline + +# The training could take up to a few minutes. +trainer.train( +reader=train_reader, +num_epochs=100, +event_handler=event_handler_plot, +feed_order=feed_order) +``` + +![trainTestCost](./image/train_and_test.png) + +## 预测 +提供一个`inference_program`和一个`params_dirname`来初始化预测器。`params_dirname`用来存储我们的参数。 + +### 设定预测程序 +类似于`trainer.train`，预测器需要一个预测程序来做预测。我们可以稍加修改我们的训练程序来把预测值包含进来。 + + +```python +def inference_program(): +x = fluid.layers.data(name='x', shape=[13], dtype='float32') +y_predict = fluid.layers.fc(input=x, size=1, act=None) +return y_predict +``` + +### 预测 +预测器会从`params_dirname`中读取已经训练好的模型，来对从未遇见过的数据进行预测。 + +```python +inferencer = fluid.Inferencer( +infer_func=inference_program, param_path=params_dirname, place=place) + +batch_size = 10 +tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32") + +results = inferencer.infer({'x': tensor_x}) +print("infer results: ", results[0]) +``` + +## 总结 +在这章里，我们借助波士顿房价这一数据集，介绍了线性回归模型的基本概念，以及如何使用PaddlePaddle实现训练和测试的过程。很多的模型和技巧都是从简单的线性回归模型演化而来，因此弄清楚线性模型的原理和局限非常重要。 + + +## 参考文献 +1. https://en.wikipedia.org/wiki/Linear_regression +2. Friedman J, Hastie T, Tibshirani R. The elements of statistical learning[M]. Springer, Berlin: Springer series in statistics, 2001. +3. Murphy K P. Machine learning: a probabilistic perspective[M]. MIT press, 2012. +4. Bishop C M. Pattern recognition[J]. Machine Learning, 2006, 128. + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/predictions.png b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/predictions.png new file mode 100644 index 0000000000000000000000000000000000000000..27e4acb1313794f52ad9ad9e874cdadd197ff41f Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/predictions.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/ranges.png b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/ranges.png new file mode 100644 index 0000000000000000000000000000000000000000..5d86b12715f46afbafb7d50e2938e184219b5b95 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/ranges.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/train_and_test.png b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/train_and_test.png new file mode 100644 index 0000000000000000000000000000000000000000..bcd304a6a0baf30ecfbc43e08fc0aca179d05958 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/fit_a_line/image/train_and_test.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/index.rst b/doc/fluid/new_docs/beginners_guide/quick_start/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..f5889ba52b8016596108de48bad59f238c16afc0 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/quick_start/index.rst @@ -0,0 +1,13 @@ +######## +快速入门 +######## + +.. todo:: + + 概述 + +.. toctree:: + :maxdepth: 2 + + fit_a_line/README.cn.md + recognize_digits/README.cn.md diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md new file mode 100644 index 0000000000000000000000000000000000000000..c04a949a3f6550048f2a3447070829aeb640b995 --- /dev/null +++ b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md @@ -0,0 +1,453 @@ +# 识别数字 + +本教程源代码目录在[book/recognize_digits](https://github.com/PaddlePaddle/book/tree/develop/02.recognize_digits)， 初次使用请参考PaddlePaddle[安装教程](https://github.com/PaddlePaddle/book/blob/develop/README.cn.md#运行这本书)。 + +## 背景介绍 +当我们学习编程的时候，编写的第一个程序一般是实现打印"Hello World"。而机器学习（或深度学习）的入门教程，一般都是 [MNIST](http://yann.lecun.com/exdb/mnist/) 数据库上的手写识别问题。原因是手写识别属于典型的图像分类问题，比较简单，同时MNIST数据集也很完备。MNIST数据集作为一个简单的计算机视觉数据集，包含一系列如图1所示的手写数字图片和对应的标签。图片是28x28的像素矩阵，标签则对应着0~9的10个数字。每张图片都经过了大小归一化和居中处理。 + +![MNIST](./image/mnist_example_image.png) +

图1. MNIST图片示例

+ +MNIST数据集是从 [NIST](https://www.nist.gov/srd/nist-special-database-19) 的Special Database 3（SD-3）和Special Database 1（SD-1）构建而来。由于SD-3是由美国人口调查局的员工进行标注，SD-1是由美国高中生进行标注，因此SD-3比SD-1更干净也更容易识别。Yann LeCun等人从SD-1和SD-3中各取一半作为MNIST的训练集（60000条数据）和测试集（10000条数据），其中训练集来自250位不同的标注员，此外还保证了训练集和测试集的标注员是不完全相同的。 + +Yann LeCun早先在手写字符识别上做了很多研究，并在研究过程中提出了卷积神经网络（Convolutional Neural Network），大幅度地提高了手写字符的识别能力，也因此成为了深度学习领域的奠基人之一。如今的深度学习领域，卷积神经网络占据了至关重要的地位，从最早Yann LeCun提出的简单LeNet，到如今ImageNet大赛上的优胜模型VGGNet、GoogLeNet、ResNet等（请参见[图像分类](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification) 教程），人们在图像分类领域，利用卷积神经网络得到了一系列惊人的结果。 + +有很多算法在MNIST上进行实验。1998年，LeCun分别用单层线性分类器、多层感知器（Multilayer Perceptron, MLP）和多层卷积神经网络LeNet进行实验，使得测试集上的误差不断下降（从12%下降到0.7%）\[[1](#参考文献)\]。此后，科学家们又基于K近邻（K-Nearest Neighbors）算法\[[2](#参考文献)\]、支持向量机（SVM）\[[3](#参考文献)\]、神经网络\[[4-7](#参考文献)\]和Boosting方法\[[8](#参考文献)\]等做了大量实验，并采用多种预处理方法（如去除歪曲、去噪、模糊等）来提高识别的准确率。 + +本教程中，我们从简单的模型Softmax回归开始，带大家入门手写字符识别，并逐步进行模型优化。 + + +## 模型概览 + +基于MNIST数据训练一个分类器，在介绍本教程使用的三个基本图像分类网络前，我们先给出一些定义： +- `$X$`是输入：MNIST图片是`$28\times28$` 的二维图像，为了进行计算，我们将其转化为`$784$`维向量，即`$X=\left ( x_0, x_1, \dots, x_{783} \right )$`。 +- `$Y$`是输出：分类器的输出是10类数字（0-9），即`$Y=\left ( y_0, y_1, \dots, y_9 \right )$`，每一维`$y_i$`代表图片分类为第`$i$`类数字的概率。 +- `$L$`是图片的真实标签：`$L=\left ( l_0, l_1, \dots, l_9 \right )$`也是10维，但只有一维为1，其他都为0。 + +### Softmax回归(Softmax Regression) + +最简单的Softmax回归模型是先将输入层经过一个全连接层得到的特征，然后直接通过softmax 函数进行多分类\[[9](#参考文献)\]。 + +输入层的数据`$X$`传到输出层，在激活操作之前，会乘以相应的权重 `$W$` ，并加上偏置变量 `$b$` ，具体如下： + +$$ y_i = \text{softmax}(\sum_j W_{i,j}x_j + b_i) $$ + +其中 `$ \text{softmax}(x_i) = \frac{e^{x_i}}{\sum_j e^{x_j}} $` + +对于有 `$N$` 个类别的多分类问题，指定 `$N$` 个输出节点，`$N$` 维结果向量经过softmax将归一化为 `$N$` 个[0,1]范围内的实数值，分别表示该样本属于这 `$N$` 个类别的概率。此处的 `$y_i$` 即对应该图片为数字 `$i$` 的预测概率。 + +在分类问题中，我们一般采用交叉熵代价损失函数（cross entropy），公式如下： + +$$ \text{crossentropy}(label, y) = -\sum_i label_ilog(y_i) $$ + +图2为softmax回归的网络图，图中权重用蓝线表示、偏置用红线表示、+1代表偏置参数的系数为1。 + +![softmaxRegression](./image/softmax_regression.png) +

图2. softmax回归网络结构图

+ +### 多层感知器(Multilayer Perceptron, MLP) + +Softmax回归模型采用了最简单的两层神经网络，即只有输入层和输出层，因此其拟合能力有限。为了达到更好的识别效果，我们考虑在输入层和输出层中间加上若干个隐藏层\[[10](#参考文献)\]。 + +1. 经过第一个隐藏层，可以得到 `$ H_1 = \phi(W_1X + b_1) $`，其中`$\phi$`代表激活函数，常见的有sigmoid、tanh或ReLU等函数。 +2. 经过第二个隐藏层，可以得到 `$ H_2 = \phi(W_2H_1 + b_2) $`。 +3. 最后，再经过输出层，得到的`$Y=\text{softmax}(W_3H_2 + b_3)$`，即为最后的分类结果向量。 + + +图3为多层感知器的网络结构图，图中权重用蓝线表示、偏置用红线表示、+1代表偏置参数的系数为1。 + +![multilayerPerceptron](./image/mlp.png) +

图3. 多层感知器网络结构图

+ +### 卷积神经网络(Convolutional Neural Network, CNN) + +在多层感知器模型中，将图像展开成一维向量输入到网络中，忽略了图像的位置和结构信息，而卷积神经网络能够更好的利用图像的结构信息。[LeNet-5](http://yann.lecun.com/exdb/lenet/)是一个较简单的卷积神经网络。图4显示了其结构：输入的二维图像，先经过两次卷积层到池化层，再经过全连接层，最后使用softmax分类作为输出层。下面我们主要介绍卷积层和池化层。 + +![cnnStructure](./image/cnn.png) +

图4. LeNet-5卷积神经网络结构

+ +#### 卷积层 + +卷积层是卷积神经网络的核心基石。在图像识别里我们提到的卷积是二维卷积，即离散二维滤波器（也称作卷积核）与二维图像做卷积操作，简单的讲是二维滤波器滑动到二维图像上所有位置，并在每个位置上与该像素点及其领域像素点做内积。卷积操作被广泛应用与图像处理领域，不同卷积核可以提取不同的特征，例如边沿、线性、角等特征。在深层卷积神经网络中，通过卷积操作可以提取出图像低级到复杂的特征。 + +![cnn](https://raw.githubusercontent.com/PaddlePaddle/book/develop/02.recognize_digits/image/conv_layer.png) +

图5. 卷积层图片

+ +图5给出一个卷积计算过程的示例图，输入图像大小为`$H=5,W=5,D=3$`，即`$5 \times 5$`大小的3通道（RGB，也称作深度）彩色图像。这个示例图中包含两（用`$K$`表示）组卷积核，即图中滤波器`$W_0$`和`$W_1$`。在卷积计算中，通常对不同的输入通道采用不同的卷积核，如图示例中每组卷积核包含（`$D=3$`）个`$3 \times 3$`（用`$F \times F$`表示）大小的卷积核。另外，这个示例中卷积核在图像的水平方向（`$W$`方向）和垂直方向（`$H$`方向）的滑动步长为2（用`$S$`表示）；对输入图像周围各填充1（用`$P$`表示）个0，即图中输入层原始数据为蓝色部分，灰色部分是进行了大小为1的扩展，用0来进行扩展。经过卷积操作得到输出为`$3 \times 3 \times 2$`（用`$H_{o} \times W_{o} \times K$`表示）大小的特征图，即`$3 \times 3$`大小的2通道特征图，其中`$H_o$`计算公式为：`$H_o = (H - F + 2 \times P)/S + 1$`，`$W_o$`同理。 而输出特征图中的每个像素，是每组滤波器与输入图像每个特征图的内积再求和，再加上偏置`$b_o$`，偏置通常对于每个输出特征图是共享的。输出特征图`$o[:,:,0]$`中的最后一个`$-2$`计算如图5右下角公式所示。 + +在卷积操作中卷积核是可学习的参数，经过上面示例介绍，每层卷积的参数大小为`$D \times F \times F \times K$`。在多层感知器模型中，神经元通常是全部连接，参数较多。而卷积层的参数较少，这也是由卷积层的主要特性即局部连接和共享权重所决定。 + +- 局部连接：每个神经元仅与输入神经元的一块区域连接，这块局部区域称作感受野（receptive field）。在图像卷积操作中，即神经元在空间维度（spatial dimension，即上图示例H和W所在的平面）是局部连接，但在深度上是全部连接。对于二维图像本身而言，也是局部像素关联较强。这种局部连接保证了学习后的过滤器能够对于局部的输入特征有最强的响应。局部连接的思想，也是受启发于生物学里面的视觉系统结构，视觉皮层的神经元就是局部接受信息的。 + +- 权重共享：计算同一个深度切片的神经元时采用的滤波器是共享的。例如图4中计算`$o[:,:,0]$`的每个每个神经元的滤波器均相同，都为`$W_0$`，这样可以很大程度上减少参数。共享权重在一定程度上讲是有意义的，例如图片的底层边缘特征与特征在图中的具体位置无关。但是在一些场景中是无意的，比如输入的图片是人脸，眼睛和头发位于不同的位置，希望在不同的位置学到不同的特征 (参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/))。请注意权重只是对于同一深度切片的神经元是共享的，在卷积层，通常采用多组卷积核提取不同特征，即对应不同深度切片的特征，不同深度切片的神经元权重是不共享。另外，偏重对同一深度切片的所有神经元都是共享的。 + +通过介绍卷积计算过程及其特性，可以看出卷积是线性操作，并具有平移不变性（shift-invariant），平移不变性即在图像每个位置执行相同的操作。卷积层的局部连接和权重共享使得需要学习的参数大大减小，这样也有利于训练较大卷积神经网络。 + +#### 池化层 + +![pooling](./image/max_pooling.png) +

图6. 池化层图片

+ +池化是非线性下采样的一种形式，主要作用是通过减少网络的参数来减小计算量，并且能够在一定程度上控制过拟合。通常在卷积层的后面会加上一个池化层。池化包括最大池化、平均池化等。其中最大池化是用不重叠的矩形框将输入层分成不同的区域，对于每个矩形框的数取最大值作为输出层，如图6所示。 + +更详细的关于卷积神经网络的具体知识可以参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/ )和[图像分类](https://github.com/PaddlePaddle/book/blob/develop/image_classification/README.md)教程。 + +### 常见激活函数介绍 +- sigmoid激活函数： `$ f(x) = sigmoid(x) = \frac{1}{1+e^{-x}} $` + +- tanh激活函数： `$ f(x) = tanh(x) = \frac{e^x-e^{-x}}{e^x+e^{-x}} $` + +实际上，tanh函数只是规模变化的sigmoid函数，将sigmoid函数值放大2倍之后再向下平移1个单位：tanh(x) = 2sigmoid(2x) - 1 。 + +- ReLU激活函数： `$ f(x) = max(0, x) $` + +更详细的介绍请参考[维基百科激活函数](https://en.wikipedia.org/wiki/Activation_function)。 + +## 数据介绍 + +PaddlePaddle在API中提供了自动加载[MNIST](http://yann.lecun.com/exdb/mnist/)数据的模块`paddle.dataset.mnist`。加载后的数据位于`/home/username/.cache/paddle/dataset/mnist`下： + +

+ + + + + + + + + + + + + + + + + + + + + + + + + + +
文件名称说明
train-images-idx3-ubyte训练数据图片，60,000条数据
train-labels-idx1-ubyte训练数据标签，60,000条数据
t10k-images-idx3-ubyte测试数据图片，10,000条数据
t10k-labels-idx1-ubyte测试数据标签，10,000条数据
+

+ +## Fluid API 概述 + +演示将使用最新的 `Fluid API`。Fluid API是最新的 PaddlePaddle API。它在不牺牲性能的情况下简化了模型配置。 +我们建议使用 Fluid API，因为它更容易学起来。 + +下面是快速的 Fluid API 概述。 +1. `inference_program`：指定如何从数据输入中获得预测的函数。 +这是指定网络流的地方。 + +1. `train_program`：指定如何从 `inference_program` 和`标签值`中获取 `loss` 的函数。 +这是指定损失计算的地方。 + +1. `optimizer_func`: “指定优化器配置的函数。优化器负责减少损失并驱动培训。Paddle 支持多种不同的优化器。 + +1. `Trainer`：PaddlePaddle Trainer 管理由 `train_program` 和 `optimizer` 指定的训练过程。 +通过 `event_handler` 回调函数，用户可以监控培训的进展。 + +1. `Inferencer`：Fluid inferencer 加载 `inference_program` 和由 Trainer 训练的参数。 +然后，它可以推断数据和返回预测。 + +在这个演示中，我们将深入了解它们。 + +## 配置说明 +加载 PaddlePaddle 的 Fluid API 包。 + +```python +import paddle +import paddle.fluid as fluid +``` + +### Program Functions 配置 + +我们需要设置“推理程序”函数。我们想用这个程序来演示三个不同的分类器，每个分类器都定义为 Python 函数。 +我们需要将图像数据馈送到分类器。Paddle 为读取数据提供了一个特殊的层 `layer.data` 层。 +让我们创建一个数据层来读取图像并将其连接到分类网络。 + +- Softmax回归：只通过一层简单的以softmax为激活函数的全连接层，就可以得到分类的结果。 + +```python +def softmax_regression(): +img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') +predict = fluid.layers.fc( +input=img, size=10, act='softmax') +return predict +``` + +- 多层感知器：下面代码实现了一个含有两个隐藏层（即全连接层）的多层感知器。其中两个隐藏层的激活函数均采用ReLU，输出层的激活函数用Softmax。 + +```python +def multilayer_perceptron(): +img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') +# 第一个全连接层，激活函数为ReLU +hidden = fluid.layers.fc(input=img, size=200, act='relu') +# 第二个全连接层，激活函数为ReLU +hidden = fluid.layers.fc(input=hidden, size=200, act='relu') +# 以softmax为激活函数的全连接输出层，输出层的大小必须为数字的个数10 +prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') +return prediction +``` + +- 卷积神经网络LeNet-5: 输入的二维图像，首先经过两次卷积层到池化层，再经过全连接层，最后使用以softmax为激活函数的全连接层作为输出层。 + +```python +def convolutional_neural_network(): +img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') +# 第一个卷积-池化层 +conv_pool_1 = fluid.nets.simple_img_conv_pool( +input=img, +filter_size=5, +num_filters=20, +pool_size=2, +pool_stride=2, +act="relu") +conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) +# 第二个卷积-池化层 +conv_pool_2 = fluid.nets.simple_img_conv_pool( +input=conv_pool_1, +filter_size=5, +num_filters=50, +pool_size=2, +pool_stride=2, +act="relu") +# 以softmax为激活函数的全连接输出层，输出层的大小必须为数字的个数10 +prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') +return prediction +``` + +#### Train Program 配置 +然后我们需要设置训练程序 `train_program`。它首先从分类器中进行预测。 +在训练期间，它将从预测中计算 `avg_cost`。 + +**注意:** 训练程序应该返回一个数组，第一个返回参数必须是 `avg_cost`。训练器使用它来计算梯度。 + +请随意修改代码，测试 Softmax 回归 `softmax_regression`, `MLP` 和 卷积神经网络 `convolutional neural network` 分类器之间的不同结果。 + +```python +def train_program(): +label = fluid.layers.data(name='label', shape=[1], dtype='int64') + +# predict = softmax_regression() # uncomment for Softmax回归 +# predict = multilayer_perceptron() # uncomment for 多层感知器 +predict = convolutional_neural_network() # uncomment for LeNet5卷积神经网络 +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(cost) +acc = fluid.layers.accuracy(input=predict, label=label) +return [avg_cost, acc] + + +# 该模型运行在单个CPU上 +``` + +#### Optimizer Function 配置 + +在下面的 `Adam optimizer`，`learning_rate` 是训练的速度，与网络的训练收敛速度有关系。 + +```python +def optimizer_program(): +return fluid.optimizer.Adam(learning_rate=0.001) +``` + +### 数据集 Feeders 配置 + +下一步，我们开始训练过程。`paddle.dataset.movielens.train()`和`paddle.dataset.movielens.test()`分别做训练和测试数据集。这两个函数各自返回一个reader——PaddlePaddle中的reader是一个Python函数，每次调用的时候返回一个Python yield generator。 + +下面`shuffle`是一个reader decorator，它接受一个reader A，返回另一个reader B —— reader B 每次读入`buffer_size`条训练数据到一个buffer里，然后随机打乱其顺序，并且逐条输出。 + +`batch`是一个特殊的decorator，它的输入是一个reader，输出是一个batched reader —— 在PaddlePaddle里，一个reader每次yield一条训练数据，而一个batched reader每次yield一个minibatch。 + +```python +train_reader = paddle.batch( +paddle.reader.shuffle( +paddle.dataset.mnist.train(), buf_size=500), +batch_size=64) + +test_reader = paddle.batch( +paddle.dataset.mnist.test(), batch_size=64) +``` + +### Trainer 配置 + +现在，我们需要配置 `Trainer`。`Trainer` 需要接受训练程序 `train_program`, `place` 和优化器 `optimizer`。 + +```python +# 该模型运行在单个CPU上 +use_cuda = False # set to True if training with GPU +place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + +trainer = fluid.Trainer( +train_func=train_program, place=place, optimizer_func=optimizer_program) +``` + +#### Event Handler 配置 + +Fluid API 在训练期间为回调函数提供了一个钩子。用户能够通过机制监控培训进度。 +我们将在这里演示两个 `event_handler` 程序。请随意修改 Jupyter 笔记本 ，看看有什么不同。 + +`event_handler` 用来在训练过程中输出训练结果 + +```python +# Save the parameter into a directory. The Inferencer can load the parameters from it to do infer +params_dirname = "recognize_digits_network.inference.model" +lists = [] +def event_handler(event): +if isinstance(event, fluid.EndStepEvent): +if event.step % 100 == 0: +# event.metrics maps with train program return arguments. +# event.metrics[0] will yeild avg_cost and event.metrics[1] will yeild acc in this example. +print "Pass %d, Batch %d, Cost %f" % ( +event.step, event.epoch, event.metrics[0]) + +if isinstance(event, fluid.EndEpochEvent): +avg_cost, acc = trainer.test( +reader=test_reader, feed_order=['img', 'label']) + +print("Test with Epoch %d, avg_cost: %s, acc: %s" % (event.epoch, avg_cost, acc)) + +# save parameters +trainer.save_params(params_dirname) +lists.append((event.epoch, avg_cost, acc)) +``` + +`event_handler_plot` 可以用来在训练过程中画图如下： + +![png](./image/train_and_test.png) + +```python +from paddle.v2.plot import Ploter + +train_title = "Train cost" +test_title = "Test cost" +cost_ploter = Ploter(train_title, test_title) +step = 0 +lists = [] + +# event_handler to plot a figure +def event_handler_plot(event): +global step +if isinstance(event, fluid.EndStepEvent): +if step % 100 == 0: +# event.metrics maps with train program return arguments. +# event.metrics[0] will yeild avg_cost and event.metrics[1] will yeild acc in this example. +cost_ploter.append(train_title, step, event.metrics[0]) +cost_ploter.plot() +step += 1 +if isinstance(event, fluid.EndEpochEvent): +# save parameters +trainer.save_params(params_dirname) + +avg_cost, acc = trainer.test( +reader=test_reader, feed_order=['img', 'label']) +cost_ploter.append(test_title, step, avg_cost) +lists.append((event.epoch, avg_cost, acc)) +``` + +#### 开始训练 + +既然我们设置了 `event_handler` 和 `data reader`，我们就可以开始训练模型了。 + +`feed_order` 用于将数据目录映射到 `train_program` + +```python +trainer.train( +num_epochs=5, +event_handler=event_handler, +reader=train_reader, +feed_order=['img', 'label']) +``` + +训练过程是完全自动的，event_handler里打印的日志类似如下所示： + +``` +Pass 0, Batch 0, Cost 0.125650 +Pass 100, Batch 0, Cost 0.161387 +Pass 200, Batch 0, Cost 0.040036 +Pass 300, Batch 0, Cost 0.023391 +Pass 400, Batch 0, Cost 0.005856 +Pass 500, Batch 0, Cost 0.003315 +Pass 600, Batch 0, Cost 0.009977 +Pass 700, Batch 0, Cost 0.020959 +Pass 800, Batch 0, Cost 0.105560 +Pass 900, Batch 0, Cost 0.239809 +Test with Epoch 0, avg_cost: 0.053097883707459624, acc: 0.9822850318471338 +``` + +训练之后，检查模型的预测准确度。用 MNIST 训练的时候，一般 softmax回归模型的分类准确率为约为 92.34%，多层感知器为97.66%，卷积神经网络可以达到 99.20%。 + + +## 应用模型 + +可以使用训练好的模型对手写体数字图片进行分类，下面程序展示了如何使用 `fluid.Inferencer` 接口进行推断。 + +### Inference 配置 + +`Inference` 需要一个 `infer_func` 和 `param_path` 来设置网络和经过训练的参数。 +我们可以简单地插入在此之前定义的分类器。 + +```python +inferencer = fluid.Inferencer( +# infer_func=softmax_regression, # uncomment for softmax regression +# infer_func=multilayer_perceptron, # uncomment for MLP +infer_func=convolutional_neural_network, # uncomment for LeNet5 +param_path=params_dirname, +place=place) +``` + +### 生成预测输入数据 + +`infer_3.png` 是数字 3 的一个示例图像。把它变成一个 numpy 数组以匹配数据馈送格式。 + +```python +# Prepare the test image +import os +import numpy as np +from PIL import Image +def load_image(file): +im = Image.open(file).convert('L') +im = im.resize((28, 28), Image.ANTIALIAS) +im = np.array(im).reshape(1, 1, 28, 28).astype(np.float32) +im = im / 255.0 * 2.0 - 1.0 +return im + +cur_dir = cur_dir = os.getcwd() +img = load_image(cur_dir + '/image/infer_3.png') +``` + +### 预测 + +现在我们准备做预测。 + +```python +results = inferencer.infer({'img': img}) +lab = np.argsort(results) # probs and lab are the results of one batch data +print "Label of image/infer_3.png is: %d" % lab[0][0][-1] +``` + +## 总结 + +本教程的softmax回归、多层感知器和卷积神经网络是最基础的深度学习模型，后续章节中复杂的神经网络都是从它们衍生出来的，因此这几个模型对之后的学习大有裨益。同时，我们也观察到从最简单的softmax回归变换到稍复杂的卷积神经网络的时候，MNIST数据集上的识别准确率有了大幅度的提升，原因是卷积层具有局部连接和共享权重的特性。在之后学习新模型的时候，希望大家也要深入到新模型相比原模型带来效果提升的关键之处。此外，本教程还介绍了PaddlePaddle模型搭建的基本流程，从dataprovider的编写、网络层的构建，到最后的训练和预测。对这个流程熟悉以后，大家就可以用自己的数据，定义自己的网络模型，并完成自己的训练和预测任务了。 + +## 参考文献 + +1. LeCun, Yann, Léon Bottou, Yoshua Bengio, and Patrick Haffner. ["Gradient-based learning applied to document recognition."](http://ieeexplore.ieee.org/abstract/document/726791/) Proceedings of the IEEE 86, no. 11 (1998): 2278-2324. +2. Wejéus, Samuel. ["A Neural Network Approach to Arbitrary SymbolRecognition on Modern Smartphones."](http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A753279&dswid=-434) (2014). +3. Decoste, Dennis, and Bernhard Schölkopf. ["Training invariant support vector machines."](http://link.springer.com/article/10.1023/A:1012454411458) Machine learning 46, no. 1-3 (2002): 161-190. +4. Simard, Patrice Y., David Steinkraus, and John C. Platt. ["Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis."](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.160.8494&rep=rep1&type=pdf) In ICDAR, vol. 3, pp. 958-962. 2003. +5. Salakhutdinov, Ruslan, and Geoffrey E. Hinton. ["Learning a Nonlinear Embedding by Preserving Class Neighbourhood Structure."](http://www.jmlr.org/proceedings/papers/v2/salakhutdinov07a/salakhutdinov07a.pdf) In AISTATS, vol. 11. 2007. +6. Cireşan, Dan Claudiu, Ueli Meier, Luca Maria Gambardella, and Jürgen Schmidhuber. ["Deep, big, simple neural nets for handwritten digit recognition."](http://www.mitpressjournals.org/doi/abs/10.1162/NECO_a_00052) Neural computation 22, no. 12 (2010): 3207-3220. +7. Deng, Li, Michael L. Seltzer, Dong Yu, Alex Acero, Abdel-rahman Mohamed, and Geoffrey E. Hinton. ["Binary coding of speech spectrograms using a deep auto-encoder."](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.185.1908&rep=rep1&type=pdf) In Interspeech, pp. 1692-1695. 2010. +8. Kégl, Balázs, and Róbert Busa-Fekete. ["Boosting products of base classifiers."](http://dl.acm.org/citation.cfm?id=1553439) In Proceedings of the 26th Annual International Conference on Machine Learning, pp. 497-504. ACM, 2009. +9. Rosenblatt, Frank. ["The perceptron: A probabilistic model for information storage and organization in the brain."](http://psycnet.apa.org/journals/rev/65/6/386/) Psychological review 65, no. 6 (1958): 386. +10. Bishop, Christopher M. ["Pattern recognition."](http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf) Machine Learning 128 (2006): 1-58. + +
+知识共享许可协议
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。 diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn.png new file mode 100644 index 0000000000000000000000000000000000000000..3f5cdaacdc6acce41c5c6c99649be46685cf9903 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn_train_log.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn_train_log.png new file mode 100644 index 0000000000000000000000000000000000000000..65bd17eacd41bbdbdb042bd1ba366eb53663b410 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/cnn_train_log.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/infer_3.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/infer_3.png new file mode 100644 index 0000000000000000000000000000000000000000..030cd60d3b4af9aecd4941204da4ad15f6e1189f Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/infer_3.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/max_pooling.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/max_pooling.png new file mode 100644 index 0000000000000000000000000000000000000000..90b02fa2a735cfcc9efb2de90906325dedcb358c Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/max_pooling.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp.png new file mode 100644 index 0000000000000000000000000000000000000000..9f4d26cd8da32201d0a5e9c72d466301dd2b42a1 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp_train_log.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp_train_log.png new file mode 100644 index 0000000000000000000000000000000000000000..f5a478fdc24f29c17555a2f1451f3f5a079faed9 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mlp_train_log.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mnist_example_image.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mnist_example_image.png new file mode 100644 index 0000000000000000000000000000000000000000..4edd7cabf8a2282f6392ac1421c7ca4afb288589 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/mnist_example_image.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_regression.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_regression.png new file mode 100644 index 0000000000000000000000000000000000000000..40b98298288b9c406fce1cbca9c913753020a94d Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_regression.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_train_log.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_train_log.png new file mode 100644 index 0000000000000000000000000000000000000000..47204941af7f22e68386a70a06ec4f122b83e262 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/softmax_train_log.png differ diff --git a/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/train_and_test.png b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/train_and_test.png new file mode 100644 index 0000000000000000000000000000000000000000..5cb87b450d0398bcfaec0e647c362052069797e7 Binary files /dev/null and b/doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/image/train_and_test.png differ diff --git a/doc/fluid/new_docs/faq/faq.rst b/doc/fluid/new_docs/faq/faq.rst new file mode 100644 index 0000000000000000000000000000000000000000..3b4bd4f895162fa3b0ba12e785e38ad694590b25 --- /dev/null +++ b/doc/fluid/new_docs/faq/faq.rst @@ -0,0 +1,12 @@ +################### +编译安装与单元测试 +################### + +1. 通过pip安装的PaddlePaddle在 :code:`import paddle.fluid` 报找不到 :code:`libmkldnn.so` 或 :code:`libmklml_intel.so` +------------------------------------------------------------------------------------------ +出现这种问题的原因是在导入 :code:`paddle.fluid` 时需要加载 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`， +但是系统没有找到该文件。一般通过pip安装PaddlePaddle时会将 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` +拷贝到 :code:`/usr/local/lib` 路径下，所以解决办法是将该路径加到 :code:`LD_LIBRARY_PATH` 环境变量下， +即： :code:`export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH` 。 + +**注意**：如果是在虚拟环境中安装PaddlePaddle， :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` 可能不在 :code:`/usr/local/lib` 路径下。 diff --git a/doc/fluid/new_docs/faq/index_cn.rst b/doc/fluid/new_docs/faq/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..bb2ed99217609d3a9edd179d4f98ad5b8b649860 --- /dev/null +++ b/doc/fluid/new_docs/faq/index_cn.rst @@ -0,0 +1,9 @@ +FAQ +==== + +本文档对关于PaddlePaddle的一些常见问题提供了解答。如果您的问题未在此处，请您到 `PaddlePaddle社区 `_ 查找答案或直接提 `issue `_ ，我们会及时进行回复。 + +.. toctree:: + :maxdepth: 1 + + faq.rst diff --git a/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_basic_concept.rst b/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_basic_concept.rst new file mode 100644 index 0000000000000000000000000000000000000000..55c3c761f932713ffa2b462b35f9f46a8edae536 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_basic_concept.rst @@ -0,0 +1,392 @@ +================================ +PaddleFluid设计思想和基本使用概念 +================================ + + + +Paddle Fluid 是用来让用户像 PyTorch 和 Tensorflow Eager Execution 一样执行程序。 +在这些系统中，不再有模型这个概念，应用也不再包含一个用于描述 Operator 图或者一系列层的符号描述， +而是像通用程序那样描述训练或者预测的过程。 + + +深度学习平台的演化 +================ + +时至今日，深度学习已成为事实上最流行的机器学习技术。学术界多年研究加上工业界的长期实践提出了若干有效的基本建模单元： +全连接，卷积，循环神经网络等；设计各类训练技巧：初始化方法，跨层连接，各类 norm 技术等； +发明了各种新的优化算法：Adadelta，Adam 等； +各类固定的网络结构：highway, residual, attention 等纷纷涌现，不胜枚举。 +学术界工业界多年的付出共同促成了深度学习方法今日的影响力。 + +学术研究和生产实践中积累了大量的知识，能够很好的解释神经网络中基本模块各自独的学习能力和特性。 +基本模块和训练技术的组合能够搭建出千变万化的神经网络模型。 +基本模块和训练技术是有限的，但他们的组合却是千变万化，这是深度学习方法的魅力所在，也是难度所在。 + +正是这样高度的模块化特性，研究者和工程师们都在努力避免重复造轮子以提高研究和生产的效率， +又进一步催生了深度学习平台技术的发展，深度学习框架已演变成为 AI 基础设施中重要的一部分。 +从 Theano，到 DistBelief，到 TensorFlow；从 Caffe 到 Caffe2； +从 Torch 到 PyTorch；从 PaddlePaddle 到 PaddleFluid， +深度学习平台技术也经历了两代的演化，并向着第三代平台技术迈进。 + +站在历史发展的今天，当我们准备切换尝试使用一个新的深度学习平台作为支持自己学习和研究的工具时， +平台技术都发生了哪些演化，能够为我们的带来什么便利呢？ + +先让我们来看看深度学习框架解决的三大问题： + +- 如何描述计算以支持未来潜在会出现的新模型？ +- 如何高效利用异构设备最大化算力？ +- 如何利用网络中的计算机进行分布式计算来处理千万亿级别的数据？ + +以上三个问题中的第一个和使用者研究者最为密切相关。 +这篇文章我们通过分析 PaddleFluid的设计理念， +来了解一个深度学习框架如何抽象深度学习模型，来看看我们的使用经验如何在不同深度学习平台之间过度和迁移。 + +如何描述计算 +============= + +让我们首先来看看 PaddleFluid 如何描述机器学习模型 + + +PaddleFluid之 :code:`Program` + +如何描述计算很大程度决定了一个神经网络框架计算功能的完备性。 +深度学习模型和方法历经二十多年的发展：“依次执行一组计算的前向， +再以和前向计算相反的顺序执行反向计算，中间无分支无交互”， +这样的模型结构已经无法满足研究者和千千万万框架使用者的想象力。 + +从 `PaddleFluid 的设计目标 `_ 来看， +在如何描述机器学习模型这一核心问题上，PaddleFluid 的目标是： +创造一种新的计算描述方式，不但能够描述至今为止人们已知的主流神经网络模型，并且能够支持未来会出现的任意模型。 + +PaddleFluid 是如何做到支持未来出现的新模型这一目标呢？PaddleFluid 的设计选择是： +对用户来说，用一段 :code:`Program` （在 PaddleFluid 内部会被转化为一种叫作 :code:`ProgramDesc` 的描述语言）， +而不是用计算图来描述机器学习模型。 :code:`Program` 用符合用户使用直觉的方式， +提供一种新的描述语言能够描述任意复杂的机器学习模型。 + +对所有计算机专业同学学习编程语言的第一课一定是建立对“程序语言的三种执行结构：顺序执行，条件选择和循环执行”的认识。 +计算机世界的所有可计算逻辑都是由这三种执行结构表示，用这三种结构描述的逻辑是可计算的。那么同样道理， +对一个神经网络框架来说，如果可以和程序语言一样提供对这三种执行结构的支持，那么将可以描述任意复杂的， +可被计算机计算的机器学习模型。PaddleFluid通过提供对这三种执行结构的支持，来做到对任意复杂模型的描述。 + +具体来说： + +1. Fluid 的核心设计理念都可以类比到程序语言，如果已经有写程序的经验，那么使用 Fluid 构建神经网络模型的体验，将非常接近写程序； + +2. 在 PaddleFluid 中，用户不会显示地感知“计算图”这样的概念，一个机器学习模型被描述为一个 Fluid :code:`Program` （Fluid 内部称之为 :code:`ProgramDesc` ）； + +- 一个 Fluid :code:`Program` 由一组嵌套的 :code:`Block` 构成。 :code:`Block` 的概念可以类比到 C++ 或是 Java 中的一对大括号，或是 Python 语言中的一个缩进快； +- :code:`Block` 中的计算由顺序执行、条件选择或者循环执行三种方式组合，构成复杂的计算逻辑。 + +3. Fluid :code:`Program` 中包含对计算和计算对象的描述。计算的描述称之为 Operator；计算作用的对象（或者说 Operator 的输入和输出）被统一为 Tensor。 + +在描述计算和计算的作用对象这一问题上，各个深度学习框架的选择是相同的，如果有一个平台的使用经验，那么将非常容易在各个平台之间进行迁移。 + +核心使用概念 +============= + +下面，我们将更详细地了解核心使用概念在PaddlePaddle的使用方法。 + +数据表示和计算的对象：Tensor +-------------------------- + +Tensor 是向量矩阵概念的扩展，是神经网络模型计算操作的基本对象。这在是今天所有主流深度学习平台的共同选择。 + +可以简单地将 Tensor 理解为一个 N 维向量，它可以有任意多的维度。一个 Tensor 具有两个基本特征： + +1. 数据类型：每个 Tensor 的所有元素具有同样的、已知的数据类型； + +2. 大小（或者说形状）：即维度的个数（rank，阶）以及各维度的长度。 + +Tensor 某些维度的长度在定义模型阶段可能是未知的，在实际算法执行时才能确定。例如一个 mini-batch 中包含的样本数目（batch size），或者是一个 mini-batch 中序列的最大长度。 + +PaddleFluid中的Tensor +"""""""""""""""""""""" + +PaddleFluid 中也使用 Tensor 作为神经网络中输入输出数据的统一表示。Tensor 的概念在今天主流的深度学习平台中都是完全相同，可以在各个深度学习框架之间直接无缝迁移。 + +在 Fluid 中也同样存在三种特殊的 Tensor： + +1. 模型中的可学习参数 + +模型中的可学习参数生存期和整个训练任务一样长，会接受优化算法的更新。在 PaddleFluid 中同样以 :code:`Variable` 表示； +用户在绝大多数情况下都不需要自己来创建网络中的可学习参数，Fluid 为几乎常见的神经网络基本计算模块都提供了封装。 +以最简单的全连接模型为例，下面的代码片段会直接为全连接层创建连接权值 WW 和偏置（ :code:`bias` ）两个可学习参数， +无需显示地调用 variable 相关接口创建可学习参数。 + + +:: + + import paddle.fluid as fluid + + y = fluid.layers.fc(input=x, size=128, bias_attr=True) + +2. 输入输出Tensor + +整个神经网络的输入数据也是一个特殊的 Tensor，在这个 Tensor 中， +一些维度的大小在定义模型时无法确定（通常包括：batch size； +如果 mini-batch 之间，数据可变，也会包括序列的最大长度，图片的宽度和高度等），在定义模型时需要占位； +PaddleFluid 中使用 :code:`fluid.layers.data` 来接入输入数据， :code:`fluid.layer.data` 需要提供输入 Tensor 的 形状信息， +当遇到无法确定的维度 时， 相应维度指定为 None ，如下面的代码片段所示： + +:: + + import paddle.fluid as fluid + + x = fluid.layers.data(name="x", shape=[2, None, 3], dtype="int64") + +3. 常量 Tensor 在 PaddleFluid 中需要通过组合 Tensor 和 :code:`fluid.layers.assign` 来实现。 + + +计算原语：Operation/Operator +---------------------------- + +Tensor 是今天所有主流深度学习框架的统一数据表示（输入、输出、中间计算结果、模型的可学习参数都是 Tensor）。 +另一方面，对数据的操作，在主流深度学习框架中也高度统一为：Operator/Operation。 +在中文中，通常我们会习惯将其称之为算子。 + +注：在 PaddleFluid 中使用 Operator 称呼对 Tensor 的操作。 + +Operation/Operator 接受多个 Tensor 作为输入，输出若干个 Tensor，表示了从输入到输出的变化。 + +PaddleFluid中的Operator +"""""""""""""""""""""""" + +PaddleFluid 支持的所有算子，可以在 `API 帮助文档 `_ 中查看。 + +为了便于用户使用，在 Python 端，Fluid 中的 Operator 被进一步封装入 :code:`paddle.fluid.layers` ， +:code:`paddle.fluid.networks` 等模块。这是因为：一些常见的对Tensor的操作可能是有更多基础操作构成， +例如：l2 norm 内部由 reduce、elementwise_add，scale 等多个 Operator 组合计算逻辑完成， +为了提高使用的便利性，框架内部对基础 Operator 进行了一些封装，包括创建 Operator 依赖可学习参数， +可学习参数的初始化细节等，减少用户重复开发的成本。 + +对所有深度学习框架都面临同样的封装，在绝大多数情况下，用户很少会直接与框架底层的 Operator 直接打交道，而是使用框架提供的 layers，networks 等模块，降低开发的代码量。不论是什么样的概念，他们在各框架之间的本质和作用都是相同的：对 Tensor 的变换。 + +总结 +>>>>>> + +不论叫作 Operation、Operator 还是 layers，他们在各深度学习平台中的含义和作用都是相同的：对 Tensor 的变换。是一个深度学习平台提供的基础计算能力。可以在每个平台各自的 API 帮助文档中查到。 + +在各个深度学习平台都已加入 ONNX 项目的今天，每个深度学习平台提供给大家的基本算子都已趋同，与此同时，每个平台也各有其特点，会提供一些独特的算子，方便某一类任务的开发。 + +构建模型并执行 +-------------- + +整个训练任务运行方法如下： + +Fluid中的Program和Executor +""""""""""""""""""""""""""" + +1. Fluid 使用 :code:`Program` 描述神经网络模型，对用户来说，并没有计算图的概念。 +用户定义的所有 Tensor 以及对 Tensor 的操作：Operator 都会被加入一段 :code:`Program` 中； + +一段 Program 由嵌套的 :code:`Block` 构成，但用户无需显示地创建 :code:`Block` 或是显示地注意到 :code:`Block` 的存在； +在 Fluid 程序中， :code:`Block` 是在调用 :code:`while_op` ， :code:`if_op` ， :code:`parallel_do` 等特殊 :code:`Operator` 时，由这些 :code:`Operator` 来创建； +对用户使用来说，只需要知道自己正在向一段 Fluid Program 中添加变量（ :code:`Tensor` ）和操作（ :code:`Operator` ）即可。 + +2. Fluid 利用 :code:`Executor` 来执行一段 Fluid :code:`Program` 。 + +为进一步理解 Fluid 中 :code:`Executor` 的作用，需要先解释一下 Fluid 程序的执行流程。 下图展示单机上，Fluid 程序的执行流程： + +.. figure:: fluid_local_train.jpeg + + :scale: 50% + :align: center + + Figure.1 + + Fluid本地训练任务执行流程图 + +1. Fluid 设计思想和灵感非常类似于程序设计语言，和高级编译语言 C++/Java 编写程序的过程非常类似，Fluid 程序执行分为两个重要阶段：编译时和运行时； + +2. 编译期，用户通过调用 Fluid 提供的算子，向一段 :code:`Program` 中添加变量（Tensor）以及对变量的操作（Operators 或者 Layers）。用户只需要描述核心的前向计算，不需要关心反向计算，分布式下，异构设备下如何计算； + +3. 原始的 :code:`Program` 在平台内部转换为中间描述语言： :code:`ProgramDesc` ； + +4. 编译期最重要的一个功能模块是 Transpiler。Transpiler 接受一段 :code:`ProgramDesc` ，输出一段变化后的 :code:`ProgramDesc` ，作为后端 Executor 最终需要执行的 :code:`Fluid Program` ； + +最为常用的 Transipler 包括： + +1. 内存优化 Transipler：通过对变量读写依赖关系分析，插入内存回收 Operator 以维持运行过程中较小的内存开销； + +2. 分布式环境下的 Transpiler：接受用户定义的 local Program ，生成 Parameter Client 和 Parameter Server 执行的两段 :code:`Program` 。 + +3. 后端 Executor 接受 Transpiler 输出的这段 :code:`Program` ，依次执行其中的 Operator（可以类比为程序语言中的指令），在执行过程中会为 Operator 创建所需的输入输出并进行管理。 + +从上面的过程中可以看到，Fluid 程序的执行过程分为：编译器的定义 :code:`Program` ，和创建 :code:`Executor` 运行 :code:`Program` 。 + :code:`Executor` 执行一段 :code:`Program` 的过程是不可交互和不可中断的。 + +在 Fluid 中，可以创建多余一段 :code:`Program` 。默认情况，一个 PaddleFluid 程序中存在 2 段 Program： + +1. :code:`fluid.framework.default_startup_program` ：其中定义了创建模型参数，输入输出，以及模型中可学习参数的初始化等各种操作； + +- :code:`default_startup_program` 可以由框架自动生成，使用时无需显示地创建； +- 如果调用修改了参数的默认初始化方式，框架会自动的将相关的修改加入 :code:`default_startup_program` 。 + +2. :code:`fluid.framework.default_main_program` ：定义了神经网络模型，前向反向计算，以及优化算法对网络中可学习参数的更新； + +- 使用 Fluid 的核心就是构建起 :code:`default_main_program` 。 + +3. PaddleFluid 中的 :code:`Scope` 类似于 TensorFlow 中的 collection 这一概念，但在 Fluid 中 :code:`Scope` 是框架后端概念，用户无法直接操作。因此，在使用框架时无需关心。 + +总结 +""""" + +Fluid 中通过 Executor 来执行一段用户定义的 Fluid :code:`Program` 。 +1. Executor 连接了 Fluid 的前端和后端； + +2. Executor 接受用户定义的原始模型（一段 :code:`Program` ），通过调用系统中不同功能更的 :code:`Transpiler` 完成对原始 :code:`Program` 的变化，进行优化。 + +完整实例：如何完成一个机器学习模型的训练 +=================================== + + + +这一节，我们以 MNIST 手写数字识别问题 —— 机器学习任务的“Hello World”问题和数据，为例，通过一个可以运行的完整实例，来学习上文介绍的概念如何在PaddleFluid 平台使用。 + +步骤1：定义数据 +---------------- + +PaddleFluid 中以 :code:`fluid.layers.data` 来接收输入数据。 + +:: + + import numpy as np + + import paddle.fluid as fluid + import paddle.v2 as paddle + + # define the input layers for the network. + x = fluid.layers.data(name="img", shape=[1, 28, 28], dtype="float32") + y_ = fluid.layers.data(name="label", shape=[1], dtype="int64") + +Fluid 中 Tensor 的第 0 维度固定为 batch size。在上面代码段中，图像输入 :code:`x` 的形状为：[1, 28, 28]。这三个维度的含义分别是：channel 数目，图像的高度和宽度。 + +实际上 Fluid 框架内部,一幅图像输入是一个 4-D Tensor，所有 Tensor 的第 0 维固定为 batch size。框架内部会自动为batch size进行填充占位。无需对batch size指定填充占位。 + +如果除去 batch size（第 0 维度）外，如果 Tensor 某一维度的大小只能在运行时确定，可以在该位置上直接指定 :code:`None` 进行占位。 + +步骤2：定义模型 +-------------- + +通过调用 Fluid 提供的算子定义含有一个隐层的神经网络。Fluid 模型的分为模型结构和优化方法两部分。这一点与 TensorFlow 程序十分相似似，使用概念可以直接对应进行迁移。 + +:: + + # define the network topology. + y = fluid.layers.fc(input=x, size=10, act="softmax") + loss = fluid.layers.cross_entropy(input=y, label=y_) + avg_loss = fluid.layers.mean(loss) + + # define the optimization algorithm. + optimizer = fluid.optimizer.Adam(learning_rate=1e-3) + optimizer.minimize(avg_loss) + +Fluid 使用 Program 而不是计算图描述模型，一般情况下，用户无需关心 Program 的细节，当调用以上 layers 时，会向一个全局的 Program： :code:`fluid.framework.default_main_program` 中插入变量（Tensor）和对变量的操作（上述代码段中的 layers 和 optimzier）。 + +步骤3：参数初始化 +---------------- + +如上文介绍，Fluid 程序中的 Executor 是连接 Fluid 前端和后端的接口。 + +默认一个Fluid模型存在至少两段 Program。用于初始化网络中的可学习参数的那一段 :code:`Program` 叫作 :code:`fluid.default_startup_program()` 。 + +只有执行器 executor 可以执行 Fluid Program，因此，在初始化网络中的可学习参数之前，需要首先创建一个 Fluid executor。 + +:: + + # define the executor. + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + +在以上代码段中， :code:`place` 用于告诉 executor 一段 Fluid Program 在何种设备上执行， +常见的有 :code:`fluid.CPUPlace()` 和 :code:`fluid.CUDAPlace()` 。 + +步骤4：数据输入 + 执行模型训练 +---------------------------- + +我们在步骤 2 中定义的神经网络模型最终被插入一段叫做 :code:`fluid.framework.default_main_program` 的 Fluid Program 中。 + +网络可学习参数初始化之后，可以通过让执行器 Executor 执行这段 :code:`fluid.framework.default_main_program` 来进行训练。 + +:: + + train_reader = paddle.batch( + paddle.reader.shuffle(paddle.dataset.mnist.train(), buf_size=5000), + batch_size=BATCH_SIZE) + feeder = fluid.DataFeeder(place=place, feed_list=[x, y_]) + + for pass_id in range(100): + for batch_id, data in enumerate(train_reader()): + loss = exe.run( + fluid.framework.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_loss]) + print("Cur Cost : %f" % (np.array(loss[0])[0])) + +从上面的代码片段中可以看到，Fluid 程序的训练过程和 TensorFlow 程序的训练过程非常接近， +都放在一个 :code:`for` 循环中，循环读取一个 mini-batch 数据， +调用执行器执行 Fluid :code:`default_main_program` ：接收 mini-batch 输入，在其上进行前向，反向和参数更新计算。 + +`注：上面程序使用了 Fluid 内置的 MNIST 数据，和我们提供给 TensorFlow 示例程序的 MNIST 数据完全一样。` + +步骤5：观察模型效果 +----------------- + +以上步骤已经构成了完整的 Tensorflow 模型训练程序，每个 batch 观察一次 loss，可以直观看到模型的迭代效果： + +.. figure:: fluid_mnist.png + + :scale: 40% + :align: center + + Figure.2 + + Fluid MNIST手写数字识别任务代价下降曲线 + +附：完整代码 +------------ + +:: + + import numpy as np + + import paddle.fluid as fluid + import paddle.v2 as paddle + + + def main(): + BATCH_SIZE = 128 + + # define the input layers for the network. + x = fluid.layers.data(name="img", shape=[1, 28, 28], dtype="float32") + y_ = fluid.layers.data(name="label", shape=[1], dtype="int64") + + # define the network topology. + y = fluid.layers.fc(input=x, size=10, act="softmax") + loss = fluid.layers.cross_entropy(input=y, label=y_) + avg_loss = fluid.layers.mean(loss) + + optimizer = fluid.optimizer.Adam(learning_rate=5e-3) + optimizer.minimize(avg_loss) + + # define the executor. + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + + train_reader = paddle.batch( + paddle.reader.shuffle(paddle.dataset.mnist.train(), buf_size=5000), + batch_size=BATCH_SIZE) + feeder = fluid.DataFeeder(place=place, feed_list=[x, y_]) + + for pass_id in range(100): + for batch_id, data in enumerate(train_reader()): + loss = exe.run( + fluid.framework.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_loss]) + print("Cur Cost : %f" % (np.array(loss[0])[0])) + + if __name__ == "__main__": + main() diff --git a/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_local_train.jpeg b/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_local_train.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..0a495901fafb85987e34acc3c454fb87e8160fca Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_local_train.jpeg differ diff --git a/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_mnist.png b/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_mnist.png new file mode 100644 index 0000000000000000000000000000000000000000..e5ad0ba058c863cf68ef0789e58fcf67b3115fdb Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/basic_concept/fluid_mnist.png differ diff --git a/doc/fluid/new_docs/user_guides/howto/configure_simple_model/index.rst b/doc/fluid/new_docs/user_guides/howto/configure_simple_model/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..5946a2ccb7e43004eae39ec4b3c6112c66c1fd04 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/configure_simple_model/index.rst @@ -0,0 +1,88 @@ +.. _user_guide_configure_simple_model: + +############## +配置简单的网络 +############## + +在解决实际问题时，可以先从逻辑层面对问题进行建模，明确模型所需要的 **输入数据类型**、**计算逻辑**、**求解目标** 以及 **优化算法**。PaddlePaddle提供了丰富的算子来实现模型逻辑。下面以一个简单回归任务举例说明如何使用PaddlePaddle构建模型。该例子完整代码参见 `fit_a_line `_。 + +问题描述及定义 +############## + +问题描述: 给定一组数据 :math:``，求解出函数 :math:`f`，使得 :math:`y=f(x)`，其中 :math:`x\subset X` 表示一条样本的特征，为 :math:`13` 维的实数向量；:math:`y \subset Y` 为一实数表示该样本对应的值。 + +我们可以尝试用回归模型来对问题建模，回归问题的损失函数有很多，这里选择常用的均方误差。为简化问题，这里假定 :math:`f` 为简单的线性变换函数，同时选用随机梯度下降算法来求解模型。 + ++----------------+----------------------------------------------+ +| 输入数据类型 | 样本特征: 13 维 实数 | ++ +----------------------------------------------+ +| | 样本标签: 1 维 实数 | ++----------------+----------------------------------------------+ +| 计算逻辑 | 使用线性模型，产生 1维实数作为模型的预测输出 | ++----------------+----------------------------------------------+ +| 求解目标 | 最小化模型预测输出与样本标签间的均方误差 | ++----------------+----------------------------------------------+ +| 优化算法 | 随机梯度下降 | ++----------------+----------------------------------------------+ + +使用PaddlePadle建模 +################### + +从逻辑层面明确了输入数据格式、模型结构、损失函数以及优化算法后，需要使用PaddlePaddle提供的API及算子来实现模型逻辑。一个典型的模型主要包含4个部分，分别是：输入数据格式定义，模型前向计算逻辑，损失函数以及优化算法。 + +数据层 +------ + +PaddlePaddle提供了 :code:`fluid.layers.data()` 算子来描述输入数据的格式。 + +:code:`fluid.layers.data()` 算子的输出是一个Variable。这个Variable的实际类型是Tensor。Tensor具有强大的表征能力，可以表示多维数据。为了精确描述数据结构，通常需要指定数据shape以及数值类型type。其中shape为一个整数向量，type可以是一个字符串类型。目前支持的数据类型参考 :ref:`user_guide_paddle_support_data_types` 。 模型训练一般会使用batch的方式读取数据，而batch的size在训练过程中可能不固定。data算子会依据实际数据来推断batch size，所以这里提供shape时不用关心batch size，只需关心一条样本的shape即可，更高级用法请参考 :ref:`user_guide_customize_batch_size_rank`。从上知，:math:`x` 为 :math:`13` 维的实数向量，:math:`y` 为实数，可使用下面代码定义数据层： + +.. code-block:: python + + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + +该模型使用的数据比较简单，事实上data算子还可以描述变长的、嵌套的序列数据。也可以使用 :code:`open_files` 打开文件进行训练。更详细的文档可参照 :ref:`user_guide_prepare_data`。 + +前向计算逻辑 +------------ + +实现一个模型最重要的部分是实现计算逻辑，PaddlePaddle提供了丰富的算子。这些算子的封装粒度不同，通常对应一种或一组变换逻辑。算子输出即为对输入数据执行变换后的结果。用户可以灵活使用算子来完成复杂的模型逻辑。比如图像相关任务中会使用较多的卷积算子、序列任务中会使用LSTM/GRU等算子。复杂模型通常会组合多种算子，以完成复杂的变换。PaddlePaddle提供了非常自然的方式来组合算子，一般地可以使用下面的方式： + +.. code-block:: python + + op_1_out = fluid.layers.op_1(input=op_1_in, ...) + op_2_out = fluid.layers.op_2(input=op_1_out, ...) + ... + +其中op_1和op_2表示算子类型，可以是fc来执行线性变换(全连接)，也可以是conv来执行卷积变换等。通过算子的输入输出的连接来定义算子的计算顺序以及数据流方向。上面的例子中，op_1的输出是op_2的输入，那么在执行计算时，会先计算op_1，然后计算op_2。更复杂的模型可能需要使用控制流算子，依据输入数据来动态执行，针对这种情况，PaddlePaddle提供了IfElseOp和WhileOp等。算子的文档可参考 :code:`fluid.layers`。具体到这个任务, 我们使用一个fc算子： + +.. code-block:: python + + y_predict = fluid.layers.fc(input=x, size=1, act=None) + +损失函数 +-------- + +损失函数对应求解目标，我们可以通过最小化损失来求解模型。大多数模型使用的损失函数，输出是一个实数值。但是PaddlePaddle提供的损失算子一般是针对一条样本计算。当输入一个batch的数据时，损失算子的输出有多个值，每个值对应一条样本的损失，所以通常会在损失算子后面使用mean等算子，来对损失做归约。模型在一次前向迭代后会得到一个损失值，PaddlePaddle会自动执行链式求导法则计算模型里面每个参数和变量对应的梯度值。这里使用均方误差损失： + +.. code-block:: python + + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + +优化方法 +-------- + +确定损失函数后，可以通过前向计算得到损失值，然后通过链式求导法则得到参数的梯度值。获取梯度值后需要更新参数，最简单的算法是随机梯度下降法：:math:`w=w - \eta \cdot g`。但是普通的随机梯度下降算法存在一些问题: 比如收敛不稳定等。为了改善模型的训练速度以及效果，学术界先后提出了很多优化算法，包括： :code:`Momentum`、:code:`RMSProp`、:code:`Adam` 等。这些优化算法采用不同的策略来更新模型参数，一般可以针对具体任务和具体模型来选择优化算法。不管使用何种优化算法，学习率一般是一个需要指定的比较重要的超参数，需要通过实验仔细调整。这里采用随机梯度下降算法： + +.. code-block:: python + + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + +更多优化算子可以参考 :code:`fluid.optimizer()` 。 + +下一步做什么？ +############## + +使用PaddlePaddle实现模型时需要关注 **数据层**、**前向计算逻辑**、**损失函数** 和 **优化方法**。不同的任务需要的数据格式不同，涉及的计算逻辑不同，损失函数不同，优化方法也不同。PaddlePaddle提供了丰富的模型示例，可以以这些示例为参考来构建自己的模型结构。用户可以访问 `模型库 `_ 查看官方提供的示例。 diff --git a/doc/fluid/new_docs/user_guides/howto/debug/index.rst b/doc/fluid/new_docs/user_guides/howto/debug/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..0878e17b4069be6b08bc85a35e77ba6421633218 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/debug/index.rst @@ -0,0 +1,10 @@ +############ +Debug 工具 +############ + +PaddlePaddle 提供了如下方式方便 Debug 训练 情况 + +.. toctree:: + :maxdepth: 2 + + visualdl.md diff --git a/doc/fluid/new_docs/user_guides/howto/debug/visualdl.md b/doc/fluid/new_docs/user_guides/howto/debug/visualdl.md new file mode 100644 index 0000000000000000000000000000000000000000..a2f30823a6fcd379f94e6e98d043b0d00681827f --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/debug/visualdl.md @@ -0,0 +1,218 @@ +# VisualDL (Visualize the Deep Learning) +

+ +

+ +## 介绍 +VisualDL是一个面向深度学习任务设计的可视化工具，包含了scalar、参数分布、模型结构、图像可视化等功能，项目正处于高速迭代中，新的组件会不断加入。 + +目前大多数DNN平台均使用Python作为配置语言，VisualDL原生支持python的使用， +通过在模型的Python配置中添加几行，便可以为训练过程提供丰富的可视化支持。 + +除了Python SDK之外，VisualDL底层采用C++编写，其暴露的C++ SDK也可以集成到其他平台中， +实现原生的性能和定制效果。 + +## 组件 +VisualDL 目前支持4种组件： + +- graph +- scalar +- image +- histogram + +### Graph +兼容 ONNX(Open Neural Network Exchange)[https://github.com/onnx/onnx], 通过与 python SDK的结合，VisualDL可以兼容包括 PaddlePaddle, pytorch, mxnet在内的大部分主流DNN平台。 + +

+ +

+ +### Scalar +可以用于展示训练测试的误差趋势 + +

+ +

+ +### Image +可以用于可视化任何tensor，或模型生成的图片 + +

+ +

+ +### Histogram + +用于可视化任何tensor中元素分布的变化趋势 + +

+ +

+ +## 快速尝试 +请使用下面的命令，来快速测试 VisualDL。 + +``` +# 安装，建議是在虚拟环境或anaconda下。 +pip install --upgrade visualdl + +# 运行一个例子，vdl_create_scratch_log 将创建测试日志 +vdl_create_scratch_log +visualDL --logdir=scratch_log --port=8080 + +# 访问 http://127.0.0.1:8080 +``` + +如果以上步骤出现问题，很可能是因为python或pip不同版本或不同位置所致，以下安装方法能解决。 + +## 使用 virtualenv 安装 + +[Virtualenv](https://virtualenv.pypa.io/en/stable/) 能创建独立Python环境，也能确保Python和pip的相对位置正确。 + +在macOS上，安装pip和virtualenv如下： +``` +sudo easy_install pip +pip install --upgrade virtualenv +``` + +在Linux上，安装pip和virtualenv如下: +``` +sudo apt-get install python3-pip python3-dev python-virtualenv +``` + +然后创建一个虚拟环境： +``` +virtualenv ~/vdl # for Python2.7 +virtualenv -p python3 ~/vdl for Python 3.x +``` + +```~/vdl``` 是你的Virtualenv目录, 你也可以选择任一目录。 + +激活虚拟环境如下： +``` +source ~/vdl/bin/activate +``` + +现在再安装 VisualDL 和运行范例： + +``` +pip install --upgrade visualdl + +# 运行一个例子，vdl_create_scratch_log 将创建测试日志 +vdl_create_scratch_log +visualDL --logdir=scratch_log --port=8080 + +# 访问 http://127.0.0.1:8080 +``` + +如果在虚拟环境下仍然遇到安装问题，请尝试以下方法。 + + +## 使用 Anaconda 安装 + +Anaconda是一个用于科学计算的Python发行版，提供了包管理与环境管理的功能，可以很方便地解决多版本python并存、切换以及各种第三方包安装问题。 + +请根据[Anaconda下载网站](https://www.anaconda.com/download) 的指示去下载和安装Anaconda. +下载Python 3.6版本的command-Line installer. + +创建conda环境名字为```vdl```或任何名字: +``` +conda create -n vdl pip python=2.7 # or python=3.3, etc. +``` + +激活conda环境如下: +``` +source activate vdl +``` + +现在再安装 VisualDL 和运行范例： + +``` +pip install --upgrade visualdl + +# 运行一个例子，vdl_create_scratch_log 将创建测试日志 +vdl_create_scratch_log +visualDL --logdir=scratch_log --port=8080 + +# 访问 http://127.0.0.1:8080 +``` + +如果仍然遇到安装问题，请尝试以下用源代码安装方法。 + +### 使用代码安装 +``` +#建議是在虚拟环境或anaconda下。 +git clone https://github.com/PaddlePaddle/VisualDL.git +cd VisualDL + +python setup.py bdist_wheel +pip install --upgrade dist/visualdl-*.whl +``` + +如果打包和安装遇到其他问题，不安装只想运行Visual DL可以看[这里](https://github.com/PaddlePaddle/VisualDL/blob/develop/docs/how_to_dev_frontend_en.md) + + +## SDK +VisualDL 同时提供了python SDK 和 C++ SDK 来实现不同方式的使用。 + +### Python SDK +VisualDL 现在支持 Python 2和 Python 3。 + +以最简单的Scalar组件为例，尝试创建一个scalar组件并插入多个时间步的数据： + +```python +import random +from visualdl import LogWriter + +logdir = "./tmp" +logger = LogWriter(logdir, sync_cycle=10000) + +# mark the components with 'train' label. +with logger.mode("train"): + # create a scalar component called 'scalars/scalar0' + scalar0 = logger.scalar("scalars/scalar0") + +# add some records during DL model running. +for step in range(100): + scalar0.add_record(step, random.random()) +``` + +### C++ SDK +上面 Python SDK 中代码完全一致的C++ SDK用法如下 +```c++ +#include +#include +#include "visualdl/sdk.h" + +namespace vs = visualdl; +namespace cp = visualdl::components; + +int main() { + const std::string dir = "./tmp"; + vs::LogWriter logger(dir, 10000); + + logger.SetMode("train"); + auto tablet = logger.AddTablet("scalars/scalar0"); + + cp::Scalar scalar0(tablet); + + for (int step = 0; step < 1000; step++) { + float v = (float)std::rand() / RAND_MAX; + scalar0.AddRecord(step, v); + } + + return 0; +} +``` +## 启动Board +当训练过程中已经产生了日志数据，就可以启动board进行实时预览可视化信息 + +``` +visualDL --logdir +``` + +board 还支持一下参数来实现远程的访问： + +- `--host` 设定IP +- `--port` 设定端口 +- `--model_pb` 指定 ONNX 格式的模型文件 diff --git a/doc/fluid/new_docs/user_guides/howto/evaluation/index.rst b/doc/fluid/new_docs/user_guides/howto/evaluation/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..6f6698cadcba4d9645fdc4a8a74d899598b96d99 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/evaluation/index.rst @@ -0,0 +1,10 @@ +############ +模型评估和调试 +############ + +PaddlePaddle Fluid提供了常用的模型评估指标，并提供了VisualDL工具可视化模型效果。 + +.. toctree:: + :maxdepth: 2 + + metrics diff --git a/doc/fluid/new_docs/user_guides/howto/evaluation/metrics.rst b/doc/fluid/new_docs/user_guides/howto/evaluation/metrics.rst new file mode 100644 index 0000000000000000000000000000000000000000..f37968a50350a90e698cb1a63bd501635753e7fb --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/evaluation/metrics.rst @@ -0,0 +1,62 @@ +############ +模型评估 +############ + +模型评估是用指标反映模型在预期目标下精度，根据模型任务决定观察指标，作为在训练中调整超参数，评估模型效果的重要依据。 +metric函数的输入为当前模型的预测preds和labels，输出是自定义的。metric函数和loss函数非常相似，但是metric并不是模型训练网络组成部分。 + +用户可以通过训练网络得到当前的预测preds和labels，在Python端定制metric函数；也可以通过定制c++ Operator的方式，在GPU上加速metric计算。 + +paddle.fluid.metrics模块包含该功能 + + +常用指标 +############ + +metric函数根据模型任务不同，指标构建方法因任务而异。 + +回归类型任务labels是实数，因此loss和metric函数构建相同，可参考MSE的方法。 +分类任务常用指标为分类指标，本文提到的一般是二分类指标，多分类和多标签需要查看对应的API文档。例如排序指标auc，多分类可以作为0，1分类任务，auc指标仍然适用。 +Fluid中包含了常用分类指标，例如Precision, Recall, Accuracy等,更多请阅读API文档。以 :ref:`Precision` 为例，具体方法为 + +.. code-block:: python + + >>> import paddle.fluid as fluid + >>> labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + >>> data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + >>> pred = fluid.layers.fc(input=data, size=1000, act="tanh") + >>> acc = fluid.metrics.Precision() + >>> for pass in range(PASSES): + >>> acc.reset() + >>> for data in train_reader(): + >>> loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + >>> acc.update(preds=preds, labels=labels) + >>> numpy_acc = acc.eval() + + +其他任务例如MultiTask Learning，Metric Learning，Learning To Rank各种指标构造方法请参考API文档。 + +自定义指标 +############ +Fluid支持自定义指标，灵活支持各类计算任务。下文通过一个简单的计数器metric函数，实现对模型的评估。 +其中preds是模型预测值，labels是给定的标签。 + +.. code-block:: python + + >>> class MyMetric(MetricBase): + >>> def __init__(self, name=None): + >>> super(MyMetric, self).__init__(name) + >>> self.counter = 0 # simple counter + + >>> def reset(self): + >>> self.counter = 0 + + >>> def update(self, preds, labels): + >>> if not _is_numpy_(preds): + >>> raise ValueError("The 'preds' must be a numpy ndarray.") + >>> if not _is_numpy_(labels): + >>> raise ValueError("The 'labels' must be a numpy ndarray.") + >>> self.counter += sum(preds == labels) + + >>> def eval(self): + >>> return self.counter diff --git a/doc/fluid/new_docs/user_guides/howto/modification/foo.rst b/doc/fluid/new_docs/user_guides/howto/modification/foo.rst new file mode 100644 index 0000000000000000000000000000000000000000..9d43c91a8544c3b281b2e8d556cb8b8e069d7e0a --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/modification/foo.rst @@ -0,0 +1,3 @@ +### +FAQ +### diff --git a/doc/fluid/new_docs/user_guides/howto/prepare_data/feeding_data.rst b/doc/fluid/new_docs/user_guides/howto/prepare_data/feeding_data.rst new file mode 100644 index 0000000000000000000000000000000000000000..c3bf033bb8316eeb4901c0cdc61e0556c8816dac --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/prepare_data/feeding_data.rst @@ -0,0 +1,169 @@ +.. _user_guide_use_numpy_array_as_train_data: + +########################### +使用Numpy Array作为训练数据 +########################### + +PaddlePaddle Fluid支持使用 :code:`fluid.layers.data()` 配置数据层； +再使用 Numpy Array 或者直接使用Python创建C++的 +:code:`fluid.LoDTensor` , 通过 :code:`Executor.run(feed=...)` 传给 +:code:`fluid.Executor` 或 :code:`fluid.ParallelExecutor` 。 + +数据层配置 +########## + +通过 :code:`fluid.layers.data()` 可以配置神经网络中需要的数据层。具体方法为: + +.. code-block:: python + + import paddle.fluid as fluid + + image = fluid.layers.data(name="image", shape=[3, 224, 224]) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + + # use image/label as layer input + prediction = fluid.layers.fc(input=image, size=1000, act="softmax") + loss = fluid.layers.cross_entropy(input=prediction, label=label) + ... + +上段代码中，:code:`image` 和 :code:`label` 是通过 :code:`fluid.layers.data` +创建的两个输入数据层。其中 :code:`image` 是 :code:`[3, 224, 224]` 维度的浮点数据; +:code:`label` 是 :code:`[1]` 维度的整数数据。这里需要注意的是: + +1. Fluid中默认使用 :code:`-1` 表示 batch size 维度，默认情况下会在 :code:`shape` + 的第一个维度添加 :code:`-1` 。 所以 上段代码中， 我们可以接受将一个 + :code:`[32, 3, 224, 224]` 的numpy array传给 :code:`image` 。 如果想自定义batch size + 维度的位置的话，请设置 :code:`fluid.layers.data(append_batch_size=False)` 。 + 请参考进阶使用中的 :ref:`user_guide_customize_batch_size_rank` 。 + + +2. Fluid中用来做类别标签的数据类型是 :code:`int64`，并且标签从0开始。可用数据类型请参考 :ref:`user_guide_paddle_support_data_types`。 + +.. _user_guide_feed_data_to_executor: + +传递训练数据给执行器 +#################### + +:code:`Executor.run` 和 :code:`ParallelExecutor.run` 都接受一个 :code:`feed` 参数。 +这个参数是一个Python的字典。它的键是数据层的名字，例如上文代码中的 :code:`image`。 +它的值是对应的numpy array。 + +例如: + +.. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + exe.run(feed={ + "image": numpy.random.random(size=(32, 3, 224, 224)).astype('float32'), + "label": numpy.random.random(size=(32, 1)).astype('int64') + }) + +进阶使用 +######## + +如何传入序列数据 +---------------- + +序列数据是PaddlePaddle Fluid支持的特殊数据类型，可以使用 :code:`LoDTensor` 作为 +输入数据类型。它需要用户: 1. 传入一个mini-batch需要被训练的所有数据; +2.每个序列的长度信息。 +用户可以使用 :code:`fluid.create_lod_tensor` 来创建 :code:`LoDTensor`。 + +传入序列信息的时候，需要设置序列嵌套深度，:code:`lod_level`。 +例如训练数据是词汇组成的句子，:code:`lod_level=1`；训练数据是 词汇先组成了句子， +句子再组成了段落，那么 :code:`lod_level=2`。 + +例如: + +.. code-block:: python + + sentence = fluid.layers.data(name="sentence", dtype="int64", shape=[1], lod_level=1) + + ... + + exe.run(feed={ + "sentence": create_lod_tensor( + data=numpy.array([1, 3, 4, 5, 3, 6, 8], dtype='int64').reshape(-1, 1), + lod=[4, 1, 2], + place=fluid.CPUPlace() + ) + }) + +训练数据 :code:`sentence` 包含三个样本，他们的长度分别是 :code:`4, 1, 2`。 +他们分别是 :code:`data[0:4]`， :code:`data[4:5]` 和 :code:`data[5:7]`。 + +如何分别设置ParallelExecutor中每个设备的训练数据 +------------------------------------------------ + +用户将数据传递给使用 :code:`ParallelExecutor.run(feed=...)` 时， +可以显示指定每一个训练设备(例如GPU)上的数据。 +用户需要将一个列表传递给 :code:`feed` 参数，列表中的每一个元素都是一个字典。 +这个字典的键是数据层的名字，值是数据层的值。 + +例如: + +.. code-block:: python + + parallel_executor = fluid.ParallelExecutor() + parallel_executor.run( + feed=[ + { + "image": numpy.random.random(size=(32, 3, 224, 224)).astype('float32'), + "label": numpy.random.random(size=(32, 1)).astype('int64') + }, + { + "image": numpy.random.random(size=(16, 3, 224, 224)).astype('float32'), + "label": numpy.random.random(size=(16, 1)).astype('int64') + }, + ] + ) + +上述代码中，GPU0会训练 32 个样本，而 GPU1训练 16 个样本。 + + +.. _user_guide_customize_batch_size_rank: + +自定义BatchSize维度 +------------------- + +PaddlePaddle Fluid默认batch size是数据的第一维度，以 :code:`-1` 表示。但是在高级 +使用中，batch_size 可以固定，也可以是其他维度或者多个维度来表示。这都需要设置 +:code:`fluid.layers.data(append_batch_size=False)` 来完成。 + +1. 固定batch size维度 + + .. code-block:: python + + image = fluid.layers.data(name="image", shape=[32, 784], append_batch_size=False) + + 这里，:code:`image` 永远是一个 :code:`[32, 784]` 大小的矩阵。 + +2. 使用其他维度表示batch size + + .. code-block:: python + + sentence = fluid.layers.data(name="sentence", + shape=[80, -1, 1], + append_batch_size=False, + dtype="int64") + + 这里 :code:`sentence` 的中间维度是batch size。这种数据排布会用在定长的循环神经 + 网络中。 + + +.. _user_guide_paddle_support_data_types: + +Fluid目前支持的数据类型 +----------------------- + +PaddlePaddle Fluid目前支持的数据类型包括: + + * float16： 部分操作支持 + * float32: 主要实数类型 + * float64: 次要实数类型，支持大部分操作 + * int32: 次要标签类型 + * int64: 主要标签类型 + * uint64: 次要标签类型 + * bool: 控制流数据类型 + * int16: 次要标签类型 + * uint8: 输入数据类型，可用于图像像素 \ No newline at end of file diff --git a/doc/fluid/new_docs/user_guides/howto/prepare_data/index.rst b/doc/fluid/new_docs/user_guides/howto/prepare_data/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..56fa928029903f1e3bd3e8064c146797f01b2b85 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/prepare_data/index.rst @@ -0,0 +1,52 @@ +.. _user_guide_prepare_data: + +######## +准备数据 +######## + +PaddlePaddle Fluid支持两种传入数据的方式: + +1. 用户需要使用 :code:`fluid.layers.data` +配置数据输入层，并在 :code:`fluid.Executor` 或 :code:`fluid.ParallelExecutor` +中，使用 :code:`executor.run(feed=...)` 传入训练数据。 + +2. 用户需要先将训练数据 +转换成 Paddle 识别的 :code:`fluid.recordio_writer` ， 再使用 +:code:`fluid.layers.open_files` 以及 :code:`fluid.layers.reader` 配置数据读取。 + +这两种准备数据方法的比较如下: + +.. _user_guide_prepare_data_comparision: + ++------------+----------------------------------+---------------------------------------+ +| | Feed数据 | 使用Reader | ++============+==================================+=======================================+ +| API接口 | :code:`executor.run(feed=...)` | :code:`fluid.layers.reader` | ++------------+----------------------------------+---------------------------------------+ +| 数据格式 | Numpy Array | :code:`fluid.recordio_writer` | ++------------+----------------------------------+---------------------------------------+ +| 数据增强 | Python端使用其他库完成 | 使用Fluid中的Operator 完成 | ++------------+----------------------------------+---------------------------------------+ +| 速度 | 慢 | 快 | ++------------+----------------------------------+---------------------------------------+ +| 推荐用途 | 调试模型 | 工业训练 | ++------------+----------------------------------+---------------------------------------+ + +这些准备数据的详细使用方法，请参考: + +.. toctree:: + :maxdepth: 2 + + feeding_data + use_recordio_reader + +Python Reader +############# + +为了方便用户在Python中定义数据处理流程，PaddlePaddle Fluid支持 Python Reader， +具体请参考: + +.. toctree:: + :maxdepth: 2 + + reader.md diff --git a/doc/fluid/new_docs/user_guides/howto/prepare_data/reader.md b/doc/fluid/new_docs/user_guides/howto/prepare_data/reader.md new file mode 100644 index 0000000000000000000000000000000000000000..aa50e4d26166536eaf8044d527debd8ad46060f6 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/prepare_data/reader.md @@ -0,0 +1,210 @@ +```eval_rst +.. _user_guide_reader: +``` + +# Python Reader + +During the training and testing phases, PaddlePaddle programs need to read data. To help the users write code that performs reading input data, we define the following: + +- A *reader*: A function that reads data (from file, network, random number generator, etc) and yields the data items. +- A *reader creator*: A function that returns a reader function. +- A *reader decorator*: A function, which takes in one or more readers, and returns a reader. +- A *batch reader*: A function that reads data (from *reader*, file, network, random number generator, etc) and yields a batch of data items. + +and also provide a function which can convert a reader to a batch reader, frequently used reader creators and reader decorators. + +## Data Reader Interface + +*Data reader* doesn't have to be a function that reads and yields data items. It can just be any function without any parameters that creates an iterable (anything can be used in `for x in iterable`) as follows: + +``` +iterable = data_reader() +``` + +The item produced from the iterable should be a **single** entry of data and **not** a mini batch. The entry of data could be a single item or a tuple of items. Item should be of one of the [supported types](http://www.paddlepaddle.org/doc/ui/data_provider/pydataprovider2.html?highlight=dense_vector#input-types) (e.g., numpy 1d array of float32, int, list of int etc.) + +An example implementation for single item data reader creator is as follows: + +```python +def reader_creator_random_image(width, height): + def reader(): + while True: + yield numpy.random.uniform(-1, 1, size=width*height) + return reader +``` + +An example implementation for multiple item data reader creator is as follows: +```python +def reader_creator_random_image_and_label(width, height, label): + def reader(): + while True: + yield numpy.random.uniform(-1, 1, size=width*height), label + return reader +``` + +## Batch Reader Interface + +*Batch reader* can be any function without any parameters that creates an iterable (anything can be used in `for x in iterable`). The output of the iterable should be a batch (list) of data items. Each item inside the list should be a tuple. + +Here are some valid outputs: + +```python +# a mini batch of three data items. Each data item consist three columns of data, each of which is 1. +[(1, 1, 1), +(2, 2, 2), +(3, 3, 3)] + +# a mini batch of three data items, each data item is a list (single column). +[([1,1,1],), +([2,2,2],), +([3,3,3],)] +``` + +Please note that each item inside the list must be a tuple, below is an invalid output: +```python + # wrong, [1,1,1] needs to be inside a tuple: ([1,1,1],). + # Otherwise it is ambiguous whether [1,1,1] means a single column of data [1, 1, 1], + # or three columns of data, each of which is 1. +[[1,1,1], +[2,2,2], +[3,3,3]] +``` + +It is easy to convert from a reader to a batch reader: + +```python +mnist_train = paddle.dataset.mnist.train() +mnist_train_batch_reader = paddle.batch(mnist_train, 128) +``` + +It is also straight forward to create a custom batch reader: + +```python +def custom_batch_reader(): + while True: + batch = [] + for i in xrange(128): + batch.append((numpy.random.uniform(-1, 1, 28*28),)) # note that it's a tuple being appended. + yield batch + +mnist_random_image_batch_reader = custom_batch_reader +``` + +## Usage + +Following is how we can use the reader with PaddlePaddle: +The batch reader, a mapping from item(s) to data layer, the batch size and the number of total passes will be passed into `paddle.train` as follows: + +```python +# two data layer is created: +image_layer = paddle.layer.data("image", ...) +label_layer = paddle.layer.data("label", ...) + +# ... +batch_reader = paddle.batch(paddle.dataset.mnist.train(), 128) +paddle.train(batch_reader, {"image":0, "label":1}, 128, 10, ...) +``` + +## Data Reader Decorator + +The *Data reader decorator* takes in a single reader or multiple data readers and returns a new data reader. It is similar to a [python decorator](https://wiki.python.org/moin/PythonDecorators), but it does not use `@` in the syntax. + +Since we have a strict interface for data readers (no parameters and return a single data item), a data reader can be used in a flexible way using data reader decorators. Following are a few examples: + +### Prefetch Data + +Since reading data may take some time and training can not proceed without data, it is generally a good idea to prefetch the data. + +Use `paddle.reader.buffered` to prefetch data: + +```python +buffered_reader = paddle.reader.buffered(paddle.dataset.mnist.train(), 100) +``` + +`buffered_reader` will try to buffer (prefetch) `100` data entries. + +### Compose Multiple Data Readers + +For example, if we want to use a source of real images (say reusing mnist dataset), and a source of random images as input for [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661). + +We can do the following : + +```python +def reader_creator_random_image(width, height): + def reader(): + while True: + yield numpy.random.uniform(-1, 1, size=width*height) + return reader + +def reader_creator_bool(t): + def reader: + while True: + yield t + return reader + +true_reader = reader_creator_bool(True) +false_reader = reader_creator_bool(False) + +reader = paddle.reader.compose(paddle.dataset.mnist.train(), data_reader_creator_random_image(20, 20), true_reader, false_reader) +# Skipped 1 because paddle.dataset.mnist.train() produces two items per data entry. +# And we don't care about the second item at this time. +paddle.train(paddle.batch(reader, 128), {"true_image":0, "fake_image": 2, "true_label": 3, "false_label": 4}, ...) +``` + +### Shuffle + +Given the shuffle buffer size `n`, `paddle.reader.shuffle` returns a data reader that buffers `n` data entries and shuffles them before a data entry is read. + +Example: +```python +reader = paddle.reader.shuffle(paddle.dataset.mnist.train(), 512) +``` + +## Q & A + +### Why does a reader return only a single entry, and not a mini batch? + +Returning a single entry makes reusing existing data readers much easier (for example, if an existing reader returns 3 entries instead if a single entry, the training code will be more complicated because it need to handle cases like a batch size 2). + +We provide a function: `paddle.batch` to turn (a single entry) reader into a batch reader. + +### Why do we need a batch reader, isn't is sufficient to give the reader and batch_size as arguments during training ? + +In most of the cases, it would be sufficient to give the reader and batch_size as arguments to the train method. However sometimes the user wants to customize the order of data entries inside a mini batch, or even change the batch size dynamically. For these cases using a batch reader is very efficient and helpful. + +### Why use a dictionary instead of a list to provide mapping? + +Using a dictionary (`{"image":0, "label":1}`) instead of a list (`["image", "label"]`) gives the advantage that the user can easily reuse the items (e.g., using `{"image_a":0, "image_b":0, "label":1}`) or even skip an item (e.g., using `{"image_a":0, "label":2}`). + +### How to create a custom data reader creator ? + +```python +def image_reader_creator(image_path, label_path, n): + def reader(): + f = open(image_path) + l = open(label_path) + images = numpy.fromfile( + f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32') + images = images / 255.0 * 2.0 - 1.0 + labels = numpy.fromfile(l, 'ubyte', count=n).astype("int") + for i in xrange(n): + yield images[i, :], labels[i] # a single entry of data is created each time + f.close() + l.close() + return reader + +# images_reader_creator creates a reader +reader = image_reader_creator("/path/to/image_file", "/path/to/label_file", 1024) +paddle.train(paddle.batch(reader, 128), {"image":0, "label":1}, ...) +``` + +### How is `paddle.train` implemented + +An example implementation of paddle.train is: + +```python +def train(batch_reader, mapping, batch_size, total_pass): + for pass_idx in range(total_pass): + for mini_batch in batch_reader(): # this loop will never end in online learning. + do_forward_backward(mini_batch, mapping) +``` diff --git a/doc/fluid/new_docs/user_guides/howto/prepare_data/use_recordio_reader.rst b/doc/fluid/new_docs/user_guides/howto/prepare_data/use_recordio_reader.rst new file mode 100644 index 0000000000000000000000000000000000000000..dfda33f1b03516fe2c704f55d095955282b19109 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/prepare_data/use_recordio_reader.rst @@ -0,0 +1,167 @@ +.. _user_guide_use_recordio_as_train_data: + +############################ +使用RecordIO文件作为训练数据 +############################ + +相比于 :ref:`user_guide_use_numpy_array_as_train_data`， +:ref:`user_guide_use_recordio_as_train_data` 的性能更好； +但是用户需要先将训练数据集转换成RecordIO文件格式，再使用 +:code:`fluid.layers.open_files()` 层在神经网络配置中导入 RecordIO 文件。 +用户还可以使用 :code:`fluid.layers.double_buffer()` 加速数据从内存到显存的拷贝， +使用 :code:`fluid.layers.Preprocessor` 工具进行数据增强。 + +将训练数据转换成RecordIO文件格式 +################################ + +:code:`fluid.recordio_writer` 中，每个记录都是一个 +:code:`vector`, 即一个支持序列信息的Tensor数组。这个数组包括训练所需 +的所有特征。例如对于图像分类来说，这个数组可以包含图片和分类标签。 + +用户可以使用 :code:`fluid.recordio_writer.convert_reader_to_recordio_file()` 可以将 +:ref:`user_guide_reader` 转换成一个RecordIO文件。或者可以使用 +:code:`fluid.recordio_writer.convert_reader_to_recordio_files()` 将一个 +:ref:`user_guide_reader` 转换成多个RecordIO文件。 + +具体使用方法为: + +.. code-block:: python + + import paddle.fluid as fluid + import numpy + + def reader_creator(): + def __impl__(): + for i in range(1000): + yield [ + numpy.random.random(size=[3,224,224], dtype="float32"), + numpy.random.random(size=[1], dtype="int64") + ] + return __impl__ + + img = fluid.layers.data(name="image", shape=[3, 224, 224]) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + feeder = fluid.DataFeeder(feed_list=[img, label], place=fluid.CPUPlace()) + + BATCH_SIZE = 32 + reader = paddle.batch(reader_creator(), batch_size=BATCH_SIZE) + fluid.recordio_writer.convert_reader_to_recordio_file( + "train.recordio", feeder=feeder, reader_creator=reader) + +其中 :code:`reader_creator` 创建了一个 :code:`Reader`。 +:ref:`_api_fluid_data_feeder_DataFeeder` +是将 :code:`Reader` 转换成 :code:`LoDTensor` 的工具。详细请参考 +:ref:`user_guide_reader` 。 + +上述程序将 :code:`reader_creator` 的数据转换成了 :code:`train.recordio` 文件， +其中每一个record 含有 32 条样本。如果batch size会在训练过程中调整， +用户可以将每一个Record的样本数设置成1。并参考 +:ref:`user_guide_use_recordio_as_train_data_use_op_create_batch`。 + + +配置神经网络, 打开RecordIO文件 +############################## + +RecordIO文件转换好之后，用户可以使用 :code:`fluid.layers.open_files()` +打开文件，并使用 :code:`fluid.layers.read_file` 读取文件内容。 +简单使用方法如下: + +.. code-block:: python + + import paddle.fluid as fluid + + file_obj = fluid.layers.open_files( + filenames=["train.recordio"], + shape=[[3, 224, 224], [1]], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + pass_num=100 + ) + + image, label = fluid.layers.read_file(file_obj) + +其中如果设置了 :code:`pass_num` ，那么当所有数据读完后，会重新读取数据， +直到读取了 :code:`pass_num` 遍。 + + + +进阶使用 +######## + + +使用 :code:`fluid.layers.double_buffer()` +------------------------------------------ + +:code:`Double buffer` 使用双缓冲技术，将训练数据从内存中复制到显存中。配置双缓冲 +需要使用 :code:`fluid.layers.double_buffer()` 修饰文件对象。 例如: + +.. code-block:: python + + import paddle.fliud as fluid + file_obj = fluid.layers.open_files(...) + file_obj = fluid.layers.double_buffer(file_obj) + + image, label = fluid.layers.read_file(file_obj) + +双缓冲技术可以参考 +`Multiple buffering `_ 。 + +配置数据增强 +------------ + +使用 :code:`fluid.layers.Preprocessor` 可以配置文件的数据增强方法。例如 + +.. code-block:: python + + import paddle.fluid as fluid + file_obj = fluid.layers.open_files(...) + preprocessor = fluid.layers.Preprocessor(reader=data_file) + with preprocessor.block(): + image, label = preprocessor.inputs() + image = image / 2 + label = label + 1 + preprocessor.outputs(image, label) + +如上代码所示，使用 :code:`Preprocessor` 定义了一个数据增强模块，并在 +:code:`with preprocessor.block()` 中定义了数据增强的具体操作。 用户通过配置 +:code:`preprocessor.inputs()` 获得数据文件中的各个字段。 并用 +:code:`preprocessor.outputs()` 标记预处理后的输出。 + +.. _user_guide_use_recordio_as_train_data_use_op_create_batch: + +使用Op组batch +------------- + +使用 :code:`fluid.layers.batch()` 可以在训练的过程中动态的组batch。例如 + +.. code-block:: python + + import paddle.fluid as fluid + file_obj = fluid.layers.open_files(...) + file_obj = fluid.layers.batch(file_obj, batch_size=32) + + img, label = fluid.layers.read_file(file_obj) + +需要注意的是，如果数据集中的最后几个样本不能组成 :code:`batch_size` 大小的批量数据， +那么这几个样本直接组成一个批量数据进行训练。 + +读入数据的shuffle +----------------- + +使用 :code:`fluid.layers.shuffle()` 可以在训练过程中动态重排训练数据。例如 + +.. code-block:: python + + import paddle.fluid as fluid + file_obj = fluid.layers.open_files(...) + file_obj = fliud.layers.shuffle(file_obj, buffer_size=8192) + + img, label = fliud.layers.read_file(file_obj) + +需要注意的是: + +1. :code:`shuffle` 实现方法是: +先读入 :code:`buffer_size` 条样本，再随机的选出样本进行训练。 + +2. :code:`shuffle` 中 :code:`buffer_size` 会占用训练内存，需要确定训练过程中内存 +足够支持缓存 :code:`buffer_size` 条数据。 diff --git a/doc/fluid/new_docs/user_guides/howto/training/checkpoint_doc_cn.md b/doc/fluid/new_docs/user_guides/howto/training/checkpoint_doc_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..c4afd536c67b24a17e4437ecedf779ddcddcbc98 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/checkpoint_doc_cn.md @@ -0,0 +1,60 @@ +# Checkpoint功能使用指南 + +## 背景 +单机/多机在训练过程中会由于软件/硬件的问题出现异常，导致训练中断，进而导致训练无结果或结果不可用，浪费大量时间和机器性能。 + +## 目的 +Checkpoint功能能够在训练中途对训练数据中间数据进行保存，出现异常恢复训练的时候能够加载中途保存的数据继续训练， 实现单机/多机的容错训练的功能。 + +## 说明 +### 目前已实现的参数保存： +1. 基于Trainer 0 实现训练过程中的参数保存 +2. 基于PServer 实现了```Distribute Lookup Table```相关参数保存 +### Fluid Checkpoint 保存数据目录结构： + +``` +checkpoint_dir (用户定义的checkpoint目录) +├── checkpoint_0 (第一次保存) +│ ├── __lockup_table__ (Distribute Lookup Table 目录) +│ │ ├── table_pserver_0 (Pserver 0 号保存的lookup table 数据) +│ │ └── table_pserver_1 +│ ├── __model__ (model 目录) +│ │ └── var.w_1 +│ └── trainer_0 (trainer 自有数据保存) +│ ├── epoch_id +│ └── step_id +└── checkpoint_1 (第二次保存) +``` + +## 使用方法 +### 声明Fluid.CheckpointConfig +用户对checkpoint功能的配置，主要是配置对象```Fluid```中的```CheckpointConfig```. + +```CheckpointConfig``` 包括4个参数： + +| 参数 | 类型 | 说明 | +| - | :-: | - | +| checkpoint_dir | int| checkpoint存储目录 | +| max_num_checkpoints | int | 最大保存的checkpoint副本数 | +| epoch_interval | int | 每隔epoch_interval轮epoch | +| step_interval | int | 每隔step_interval轮step | + +### 在Fluid.Trainer对象的声明中加入Fluid.CheckpointConfig的声明 +Trainer的__init__方法的参数中包含了对```CheckpointConfig```， 需要传入在声明Trainer前声明的```CheckpointConfig```对象。 +如： +```python +config = CheckpointConfig( + checkpoint_dir = "/tmp/ckpt", max_num_checkpoints = 2, + epoch_interval = 2, step_interval = 10) +trainer = Trainer(..., checkpoint_config=config) +``` +定义和声明完成后， 训练在运行过程中就会在指定的step和epoch处进行保存，出现异常时，就会自动从最新的checkpoint目录进行参数恢复啦！ + +## 相关API +[Trainer API 说明](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/trainer.py) + +## 注意 +1. 保证每个训练的```checkpoint_dir``` 与其他训练独立。 +2. 最大副本数量```max_num_checkpoints```需要根据磁盘容量以及模型的大小进行调整， 保证磁盘的可用性。 +3. ```epoch_interval``` 和 ```step_interval``` 不宜过小， 频繁的进行checkpoint会拖慢训练速度。 +4. **分布式训练**的过程中：每个Trainer都会在```checkpoint_dir```目录中保存当前Trainer的参数（只有Trainer 0会保存模型的参数），需要**分布式文件系统(HDFS等)**将同```checkpoint_dir```目录的数据进行合并才能得到完整的数据，恢复训练的时候需要用完整的数据进行恢复。 diff --git a/doc/fluid/new_docs/user_guides/howto/training/checkpoint_doc_en.md b/doc/fluid/new_docs/user_guides/howto/training/checkpoint_doc_en.md new file mode 100644 index 0000000000000000000000000000000000000000..14d37246ca0cab8715e244fda9624d0d59f8ec5f --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/checkpoint_doc_en.md @@ -0,0 +1,62 @@ +# Checkpoint User Guide + +## Background +In many cases, Stand-alone training and Distributed training can be aborted by the software problem or hardware problem. More seriously, we waste so much time and the performance of the machine but get nothing, which makes us frustrating and we have to restart it again. + +## Purpose +The feature of ```Checkpoint``` can save Intermediate model variables, lookup table variable, and other needs data in checkpoint directory. When the exception occurs, we can load these variables from the checkpoint directory immediately. +## Introduce +### Complete Features Currently： +1. The Trainer 0 will save model variables in training. +2. Each of the Trainer will save its own arguments needed. +3. Each of the Parameter Server will save ```Distribute Lookup Table``` variables in training. +### Fluid Checkpoint directory structure： + +``` +checkpoint_dir (the checkpoint directory user define) +├── checkpoint_0 (the first save directory) +│ ├── __lockup_table__ (Distribute Lookup Table directory) +│ │ ├── table_pserver_0 (Lookup table's data about Pserver 0) +│ │ └── table_pserver_1 +│ ├── __model__ (model directory) +│ │ └── var.w_1 +│ └── trainer_0 (each trainer will save its own data) +│ ├── epoch_id +│ └── step_id +└── checkpoint_1 (the second save directory) +``` + +## usage +### Fluid.CheckpointConfig construct +When the user wants to use ```Checkpoint``` feature, the main thing user have to do is declare ```CheckpointConfig``` and construct it. + +```CheckpointConfig``` has 4 member variables need to be initialized： + +| Member Variable | Type | Comment | +| - | :-: | - | +| checkpoint_dir | int| checkpoint directory | +| max_num_checkpoints | int | Maximum number of checkpoint copies | +| epoch_interval | int | epoch interval times | +| step_interval | int | step interval times | + +### Add Fluid.CheckpointConfig's declaration in Fluid.Trainer +Because the initialization of Trainer needs an instance of ```CheckpointConfig```., we should declare ```CheckpointConfig``` in ```Fluid``` first. + +For example： +```python +config = CheckpointConfig( + checkpoint_dir = "/tmp/ckpt", max_num_checkpoints = 2, + epoch_interval = 2, step_interval = 10) +trainer = Trainer(..., checkpoint_config=config) +``` + +After all the things done, the train will save checkpoint at the specified epoch and step, when the train is aborted, the user can restart it, the train will restore from the latest copy. + +## Related API +[Related Trainer API](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/trainer.py) + +## Attention +1. Make the ```checkpoint_dir``` only be used by one train job. +2. The number of ```max_num_checkpoints``` need to be adjusted by the disk size and model size. +3. Too frequently to slow down the train speed, so too ```small epoch_interval``` and ```step_interval``` are not suitable. +4. **In distributed train**, each Trainer will save arguments in its ```checkpoint_dir``` (Only Trainer 0 will save model variables). We need **distributed file system (HDFS, etc)** to merge all the ```checkpoint_dir``` to get the whole data. diff --git a/doc/fluid/new_docs/user_guides/howto/training/cluster_howto.rst b/doc/fluid/new_docs/user_guides/howto/training/cluster_howto.rst new file mode 100644 index 0000000000000000000000000000000000000000..00ec9e819c81fae3263b1f1e6bcedf524f2b3991 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/cluster_howto.rst @@ -0,0 +1,160 @@ +.. _cluster_howto + +Fluid分布式训练使用手册 +==================== + +分布式训练基本思想 +--------------- + +分布式深度学习训练通常分为两种并行化方法：数据并行，模型并行，参考下图： + +.. image:: src/parallelism.png + +在模型并行方式下，模型的层和参数将被分布在多个节点上，模型在一个mini-batch的前向和反向训练中，将经过多次跨\ +节点之间的通信。每个节点只保存整个模型的一部分；在数据并行方式下，每个节点保存有完整的模型的层和参数，每个节点\ +独自完成前向和反向计算，然后完成梯度的聚合并同步的更新所有节点上的参数。Fluid目前版本仅提供数据并行方式，另外\ +诸如模型并行的特例实现（超大稀疏模型训练）功能将在后续的文档中予以说明。 + +在数据并行模式的训练中，Fluid使用了两种通信模式，用于应对不同训练任务对分布式训练的要求，分别为RPC通信和Collective +通信。其中RPC通信方式使用 `gRPC `_ ，Collective通信方式使用 +`NCCL2 `_ 。 + +.. csv-table:: 下面是一个RPC通信和Collective通信的横向对比： + :header: "Feature", "Coolective", "RPC" + + "Ring-Based通信", "Yes", "No" + "异步训练", "Yes", "Yes" + "分布式模型", "No", "Yes" + "容错训练", "No", "Yes" + "性能", "Faster", "Fast" + +- RPC通信方式的结构： + + .. image:: src/dist_train_pserver.png + + 使用RPC通信方式的数据并行分布式训练，会启动多个pserver进程和多个trainer进程，每个pserver进程\ + 会保存一部分模型参数，并负责接收从trainer发送的梯度并更新这些模型参数；每个trainer进程会保存一份\ + 完整的模型，并使用一部分数据进行训练，然后向pserver发送梯度，最后从pserver拉取更新后的参数。 + + pserver进程可以在和trainer完全不同的计算节点上，也可以和trainer公用节点。一个分布式任务所需要的\ + pserver进程个数通常需要根据实际情况调整，已达到最佳的性能，然而通常来说pserver的进程不会比trainer\ + 更多。 + + 在使用GPU训练时，pserver可以选择使用GPU或只使用CPU，如果pserver也使用GPU，则会增加一次从CPU拷贝\ + 接收到的梯度数据到GPU的开销，在某些情况下会导致整体训练性能降低。 + +- NCCL2通信方式的结构： + + .. image:: src/dist_train_nccl2.png + + 使用NCCL2（Collective通信方式）进行分布式训练，是不需要启动pserver进程的，每个trainer进程都保存\ + 一份完整的模型参数，在完成计算梯度之后通过trainer之间的相互通信，Reduce梯度数据到所有节点的所有设备\ + 然后每个节点在各自完成参数更新。 + +使用parameter server方式的训练 +------------------------------ + +使用 :code:`trainer` API，程序可以自动的通过识别环境变量决定是否已分布式方式执行。 + +.. csv-table:: 需要在您的分布式环境中配置的环境变量包括： + :header: "环境变量", "说明" + + "PADDLE_TRAINING_ROLE", "当前进程的角色，可以是PSERVER或TRAINER" + "PADDLE_PSERVER_PORT", "parameter使用的端口" + "PADDLE_PSERVER_IPS", "parameter server的IP地址列表，用逗号分开" + "PADDLE_TRAINERS", "分布式任务中trainer节点的个数" + "PADDLE_CURRENT_IP", "当前节点的IP" + "PADDLE_TRAINER_ID", "trainer节点的id，从0~n-1，不能有重复" + +使用更加底层的 :code:`transpiler` API可以提供自定义的分布式训练的方法，比如可以在同一台机器上， +启动多个pserver和trainer进行训练，使用底层API的方法可以参考下面的样例代码： + +.. code-block:: python + + role = "PSERVER" + trainer_id = 0 + pserver_endpoints = "127.0.0.1:6170,127.0.0.1:6171" + current_endpoint = "127.0.0.1:6170" + trainers = 4 + t = fluid.DistributeTranspiler() + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) + if role == "PSERVER": + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, + pserver_prog) + exe.run(pserver_startup) + exe.run(pserver_prog) + elif role == "TRAINER": + train_loop(t.get_trainer_program()) + + +选择同步或异步训练 +++++++++++++++++++ + +Fluid分布式任务可以支持同步训练或异步训练，在同步训练方式下，所有的trainer节点，会在每个mini-batch +同步地合并所有节点的梯度数据并发送给parameter server完成更新，在异步训练方式下，每个trainer没有相互\ +同步等待的过程，可以独立的parameter server的参数。通常情况下，使用异步训练方式，可以在trainer节点\ +更多的时候比同步训练方式有更高的总体吞吐量。 + +在调用 :code:`transpile` 函数时，默认会生成同步训练的分布式程序，通过指定 :code:`sync_mode=False` +参数即可生成异步训练的程序： + +.. code-block:: python + + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers, sync_mode=False) + + +选择参数分布方法 +++++++++++++++++ + +参数 :code:`split_method` 可以指定参数在parameter server上的分布方式。 + +Fluid默认使用 `RoundRobin `_ +方式将参数分布在多个parameter server上。此方式在默认未关闭参数切分的情况下，参数会较平均的分布在所有的 +parameter server上。如果需要使用其他，可以传入其他的方法，目前可选的方法有： :code:`RoundRobin` 和 +:code:`HashName` 。也可以使用自定义的分布方式，只需要参考 +`这里 `_ +编写自定义的分布函数。 + + +关闭切分参数 +++++++++++++ + +参数 :code:`slice_var_up` 指定是否将较大（大于8192个元素）的参数切分到多个parameter server已均衡计算负载，默认为开启。 + +当模型中的可训练参数体积比较均匀或者使用自定义的参数分布方法是参数均匀分布在多个parameter server上， +可以选择关闭切分参数，这样可以降低切分和重组带来的计算和拷贝开销： + +.. code-block:: python + + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers, slice_var_up=False) + + +使用NCCL2通信方式的训练 +-------------------- + +注NCCL2模式目前仅支持trainer API，NCCL2方式并没有很多可选项，也没有"transpiler"，所以并没有底层API。 +使用NCCL2方式同样需要配置每个节点的环境变量，此处与parameter server模式有所不同，并不需要启动独立的\ +parameter server的进程，只需要启动多个trainer进程即可。 + + +.. csv-table:: NCCL2模式环境变量说明： + :header: "环境变量", "说明" + + "PADDLE_TRAINER_IPS", "所有Trainer节点的IP列表，用逗号分隔" + "PADDLE_TRAINER_ID", "trainer节点的id，从0~n-1，不能有重复" + "PADDLE_PSERVER_PORT", "一个端口，用于在NCCL2初始化时，广播NCCL ID" + "PADDLE_CURRENT_IP", "当前节点的IP" + +目前使用NCCL2进行分布式训练仅支持同步训练方式。使用NCCL2方式的分布式训练，更适合模型体积较大，并需要使用\ +同步训练和GPU训练，如果硬件设备支持RDMA和GPU Direct，可以达到很高的分布式训练性能。 + +注意如果系统中有多个网络设备，需要手动指定NCCL2使用的设备， +假设需要使用 :code:`eth2` 为通信设备，需要设定如下环境变量： + +.. code-block:: bash + + export NCCL_SOCKET_IFNAME=eth2 + +另外NCCL2提供了其他的开关环境变量，比如指定是否开启GPU Direct，是否使用RDMA等，详情可以参考 +`ncclknobs `_ 。 diff --git a/doc/fluid/new_docs/user_guides/howto/training/cluster_quick_start.rst b/doc/fluid/new_docs/user_guides/howto/training/cluster_quick_start.rst new file mode 100644 index 0000000000000000000000000000000000000000..6131c92d6f5386c7e91b2917d25dd7ae830ff182 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/cluster_quick_start.rst @@ -0,0 +1,143 @@ +.. _cluster_quick_start: + +分布式训练快速开始 +================== + +准备工作 +-------- + +在本篇文章中，我们将会在介绍如何快速在一个集群中启动一个 PaddlePaddle +的分布式训练任务，在开始之前，请按如下步骤做些准备工作： + +1. 准备一个至少4个节点的集群，并且保证网络可以联通，在本文中我们使用 + ``*.paddlepaddle.com`` 来表示每个节点的主机名称，您可以根据集群的实际情况来修改它。 + +2. 在开始之前确保已经阅读过 :ref:`how_to_install` + 并且可以在集群的所有节点上可以正常运行 PaddlePaddle。 + +启动集群训练任务 +---------------- + +在启动集群训练脚本时，需要在不同的节点上指定不同的环境变量，具体如下： + ++-----------------+-----------------+-----------------+---------------------+ +| 环境变量 | 数据类型 | 样例 | 描述 | ++=================+=================+=================+=====================+ +| PADDLE_TRAINING | str | PSERVER,TRAINER | 训练节点的角色 | +| _ROLE | | | | ++-----------------+-----------------+-----------------+---------------------+ +| PADDLE_PSERVER_ | str | ps0.paddlepaddl | 所有 pserver | +| IPS | | e.com,ps1.paddl | 节点的 IP | +| | | epaddle.com… | 地址或 | +| | | | hostname, | +| | | | 用“,”分隔 | ++-----------------+-----------------+-----------------+---------------------+ +| PADDLE_PSERVER_ | int | 6174 | pserver | +| PORT | | | 节点监听的端口 | ++-----------------+-----------------+-----------------+---------------------+ +| PADDLE_TRAINERS | int | 2 | 训练任务中 | +| | | | trainer | +| | | | 节点的数量 | ++-----------------+-----------------+-----------------+---------------------+ +| PADDLE_CURRENT_ | str | ps0.paddlepaddl | 当前 pserver | +| IP | | e.com | 节点的 IP | +| | | | 地址或 hostanme | ++-----------------+-----------------+-----------------+---------------------+ +| PADDLE_TRAINER_ | int | 0 | 当前 trainer | +| ID | | | 节点的唯一 ID, | +| | | | 取值范围为从0开始到 | +| | | | PADDLE_TRAINERS-1 | ++-----------------+-----------------+-----------------+---------------------+ + +样例代码 +~~~~~~~~ + +将下面程序代码保存为 ``fluid_dist.py`` + +.. code:: python + + import paddle + import paddle.fluid as fluid + import contextlib + import numpy + import unittest + + # train reader + BATCH_SIZE = 20 + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.train(), buf_size=500), + batch_size=BATCH_SIZE) + + test_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.test(), buf_size=500), + batch_size=BATCH_SIZE) + + + def train_program(): + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + + loss = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_loss = fluid.layers.mean(loss) + + return avg_loss + + def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.001) + + def train(use_cuda, train_program): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + trainer = fluid.Trainer( + train_func=train_program, place=place, optimizer_func=optimizer_func) + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + if event.step == 10: + test_metrics = trainer.test( + reader=test_reader, feed_order=['x', 'y']) + print("step {0}, loss: {1}".format(event.step, test_metrics)) + trainer.stop() + + trainer.train( + reader=train_reader, + num_epochs=100, + event_handler=event_handler, + feed_order=['x', 'y']) + + train(False, train_program) + +启动trainer节点和pserver节点 +~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +.. list-table:: + :header-rows: 1 + + * - 启动节点 + - 启动命令 + - 说明 + * - ps0.paddlepaddle.com + - :code:`PADDLE_TRAINING_ROLE=PSERVER PADDLE_CURRENT_IP=ps0.paddlepaddle.com PADDLE_PSERVER_IPS=ps0.paddlepaddle.com,ps1.paddlepaddle.com PADDLE_TRAINERS=2 PADDLE_PSERVER_PORT=6174 python fluid_dist.py` + - 启动 pserver 节点 + * - ps1.paddlepaddle.com + - :code:`PADDLE_TRAINING_ROLE=PSERVER PADDLE_CURRENT_IP=ps1.paddlepaddle.com PADDLE_PSERVER_IPS=ps0.paddlepaddle.com,ps1.paddlepaddle.com PADDLE_TRAINERS=2 PADDLE_PSERVER_PORT=6174 python fluid_dist.py` + - 启动 pserver 节点 + * - trainer0.paddlepaddle.com + - :code:`PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_IPS=ps0.paddlepaddle.com,ps1.paddlepaddle.com PADDLE_TRAINERS=2 PADDLE_TRAINER_ID=0 PADDLE_PSERVER_PORT=6174 python fluid_dist.py` + - 启动第0号 trainer 节点 + * - trainer1.paddlepaddle.com + - :code:`PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_IPS=ps0.paddlepaddle.com,ps1.paddlepaddle.com PADDLE_TRAINERS=2 PADDLE_TRAINER_ID=1 PADDLE_PSERVER_PORT=6174 python fluid_dist.py` + - 启动第1号 trainer 节点 + +**注意** + +- 需要先启动pserver节点再启动trainer节点 +- 看到trainer节点输出如下日志表示训练任务执行正确 + + .. code:: bash + + step 10, loss: [258.2326202392578] diff --git a/doc/fluid/new_docs/user_guides/howto/training/index.rst b/doc/fluid/new_docs/user_guides/howto/training/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..68475101e26b3f695c8003995cc1c6a95426ff27 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/index.rst @@ -0,0 +1,12 @@ +############ +训练神经网络 +############ + +PaddlePaddle Fluid支持单机训练，和多节点训练。每种训练模式下，都支持多种训练方法。 + +.. toctree:: + :maxdepth: 2 + + single_node + multi_node + save_load_variables diff --git a/doc/fluid/new_docs/user_guides/howto/training/multi_node.rst b/doc/fluid/new_docs/user_guides/howto/training/multi_node.rst new file mode 100644 index 0000000000000000000000000000000000000000..24316f0be0d8f211e680fa15cb432732b5967c79 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/multi_node.rst @@ -0,0 +1,9 @@ +######## +多机训练 +######## + +.. toctree:: + :maxdepth: 2 + + cluster_quick_start.rst + cluster_howto.rst diff --git a/doc/fluid/new_docs/user_guides/howto/training/save_load_variables.rst b/doc/fluid/new_docs/user_guides/howto/training/save_load_variables.rst new file mode 100644 index 0000000000000000000000000000000000000000..a96776f4a17a1d6da170bdff9d81771c38912bb5 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/save_load_variables.rst @@ -0,0 +1,172 @@ +.. _user_guide_save_load_vars: + +################## +保存与载入模型变量 +################## + +模型变量分类 +############ + +在PaddlePaddle Fluid中，所有的模型变量都用 :code:`fluid.Variable()` 作为基类进行表示。 +在该基类之下，模型变量主要可以分为以下几种类别： + +1. 模型参数 + 模型参数是深度学习模型中被训练和学习的变量，在训练过程中，训练框架根据反向传播算法计算出每一个模型参数当前的梯度， + 并用优化器根据梯度对参数进行更新。模型的训练过程本质上可以看做是模型参数不断迭代更新的过程。 + 在PaddlePaddle Fluid中，模型参数用 :code:`fluid.framework.Parameter` 来表示， + 这是一个 :code:`fluid.Variable()` 的派生类，除了 :code:`fluid.Variable()` 具有的各项性质以外， + :code:`fluid.framework.Parameter` 还可以配置自身的初始化方法、更新率等属性。 + +2. 长期变量 + 长期变量指的是在整个训练过程中持续存在、不会因为一个迭代的结束而被销毁的变量，例如动态调节的全局学习率等。 + 在PaddlePaddle Fluid中，长期变量通过将 :code:`fluid.Variable()` 的 :code:`persistable` + 属性设置为 :code:`True` 来表示。所有的模型参数都是长期变量，但并非所有的长期变量都是模型参数。 + +3. 临时变量 + 不属于上面两个类别的所有模型变量都是临时变量，这种类型的变量只在一个训练迭代中存在，在每一个迭代结束后， + 所有的临时变量都会被销毁，然后在下一个迭代开始之前，又会先构造出新的临时变量供本轮迭代使用。 + 一般情况下模型中的大部分变量都属于这一类别，例如输入的训练数据、一个普通的layer的输出等等。 + + + +如何保存模型变量 +################ + +根据用途的不同，我们需要保存的模型变量也是不同的。例如，如果我们只是想保存模型用来进行以后的预测， +那么只保存模型参数就够用了。但如果我们需要保存一个checkpoint以备将来恢复训练， +那么我们应该将各种长期变量都保存下来，甚至还需要记录一下当前的epoch和step的id。 +因为一些模型变量虽然不是参数，但对于模型的训练依然必不可少。 + +因此，根据需求的不同，我们提供了两套API来分别进行模型的参数和checkpoint的保存。 + +保存模型用于对新样本的预测 +========================== + +如果我们保存模型的目的是用于对新样本的预测，那么只保存模型参数就足够了。我们可以使用 +:code:`fluid.io.save_params()` 接口来进行模型参数的保存。 + +例如： + +.. code-block:: python + + import paddle.fluid as fluid + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.save_params(executor=exe, dirname=param_path, main_program=None) + +上面的例子中，通过调用 :code:`fluid.io.save_params` 函数，PaddlePaddle Fluid会对默认 +:code:`fluid.Program` 也就是 :code:`prog` 中的所有模型变量进行扫描， +筛选出其中所有的模型参数，并将这些模型参数保存到指定的 :code:`param_path` 之中。 + + +保存checkpoint用于将来恢复训练 +============================== + +在训练过程中，我们可能希望在一些节点上将当前的训练状态保存下来， +以便在将来需要的时候恢复训练环境继续进行训练。这一般被称作“checkpoint”。 +想要保存checkpoint，可以使用 :code:`fluid.io.save_checkpiont()` 接口。 + +例如： + +.. code-block:: python + + import paddle.fluid as fluid + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./checkpoints" + prog = fluid.default_main_program() + trainer_args = {"epoch_id": 200, + "step_id": 20} # just an example + fluid.io.save_checkpoint(executor=exe, + checkpoint_dir=path, + trainer_id=0, + trainer_args=trainer_args, + main_program=prog, + max_num_checkpoints=3) + +上面的例子中，通过调用 :code:`fluid.io.save_checkpoint` 函数，PaddlePaddle Fluid会对默认 +:code:`fluid.Program` 也就是 :code:`prog` 中的所有模型变量进行扫描， +根据一系列内置的规则自动筛选出其中所有需要保存的变量，并将他们保存到指定的 :code:`path` 目录下。 + +:code:`fluid.io.save_checkpoint` 的各个参数中， :code:`trainer_id` 在单机情况下设置为0即可； :code:`trainer_args` +为一个Python dict，用于给定当前的epoch_id和step_id； +:code:`max_num_checkpoints` 用于表示的最大checkpoint数量， +如果目录中已经存在的checkpoint数量超过这个值，那最早的checkpoint将被删除。 + +如何载入模型变量 +################ + +与模型变量的保存相对应，我们提供了两套API来分别载入模型的参数和载入模型的checkpoint。 + +载入模型用于对新样本的预测 +========================== + +对于通过 :code:`fluid.io.save_params` 保存的模型，可以使用 :code:`fluid.io.load_params` +来进行载入。 + +例如： + +.. code-block:: python + + import paddle.fluid as fluid + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.load_params(executor=exe, dirname=param_path, + main_program=prog) + +上面的例子中，通过调用 :code:`fluid.io.load_params` 函数，PaddlePaddle Fluid会对 +:code:`prog` 中的所有模型变量进行扫描，筛选出其中所有的模型参数， +并尝试从 :code:`param_path` 之中读取加载它们。 + +需要格外注意的是，这里的 :code:`prog` 必须和调用 :code:`fluid.io.save_params` +时所用的 :code:`prog` 中的前向部分完全一致，且不能包含任何参数更新的操作。如果两者存在不一致， +那么可能会导致一些变量未被正确加载；如果错误地包含了参数更新操作，那可能会导致正常预测过程中参数被更改。 +这两个 :code:`fluid.Program` 之间的关系类似于训练 :code:`fluid.Program` +和测试 :code:`fluid.Program` 之间的关系，详见： :ref:`user_guide_test_while_training`。 + +另外，需特别注意运行 :code:`fluid.default_startup_program()` 必须在调用 :code:`fluid.io.load_params` +之前。如果在之后运行，可能会覆盖已加载的模型参数导致错误。 + + +载入checkpoint用于恢复训练 +========================== + +对于通过 :code:`fluid.io.save_checkpoint` 保存的模型，可以使用 :code:`fluid.io.load_checkpoint` +来进行载入。 + +例如： + +.. code-block:: python + + import paddle.fluid as fluid + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./checkpoints" + prog = fluid.default_main_program() + fluid.io.load_checkpoint(executor=exe, checkpoint_dir=path, + serial=9, main_program=prog) + +上面的例子中，通过调用 :code:`fluid.io.save_checkpoint` 函数，PaddlePaddle Fluid会对 +:code:`prog` 中的所有模型变量进行扫描，根据内置规则自动筛选出需要加载的变量， +并尝试从 :code:`path` 之中加载它们。 + +参数 :code:`serial` 用来标记具体要加载的checkpoint的版本号。在保存checkpoint的时候， +一个checkpoint会被保存在一个子目录中，并在目录名上体现出自己的版本号。 +一般越大的版本号表示这个checkpoint越新。 + +这里的 :code:`prog` 必须和调用 :code:`fluid.io.save_checkpoint` 时所用的 :code:`prog` +完全一致，否则会导致变量加载错误或者未加载。另外，与 :code:`fluid.io.save_params` 类似， +运行 :code:`fluid.default_startup_program()` 也必须在 :code:`fluid.io.load_checkpoint` +之前进行。 + +多机checkpoint保存 +################## + +.. toctree:: + :maxdepth: 2 + + checkpoint_doc_cn.md \ No newline at end of file diff --git a/doc/fluid/new_docs/user_guides/howto/training/single_node.rst b/doc/fluid/new_docs/user_guides/howto/training/single_node.rst new file mode 100644 index 0000000000000000000000000000000000000000..23eac0f831f2d6d052b7fc35b536d4ab633df851 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/single_node.rst @@ -0,0 +1,119 @@ +######## +单机训练 +######## + +准备工作 +######## + +要进行PaddlePaddle Fluid单机训练，需要先 :ref:`user_guide_prepare_data` 和 +:ref:`user_guide_configure_simple_model` 。当\ +:ref:`user_guide_configure_simple_model` 完毕后，可以得到两个\ +:code:`fluid.Program`， :code:`startup_program` 和 :code:`main_program`。 +默认情况下，可以使用 :code:`fluid.default_startup_program()` 与\ :code:`fluid.default_main_program()` 获得全局的 :code:`fluid.Program`。 + +例如: + +.. code-block:: python + + import paddle.fluid as fluid + + image = fluid.layers.data(name="image", shape=[784]) + label = fluid.layers.data(name="label", shape=[1]) + hidden = fluid.layers.fc(input=image, size=100, act='relu') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + loss = fluid.layers.mean( + fluid.layers.cross_entropy( + input=prediction, + label=label + ) + ) + + sgd = fluid.optimizer.SGD(learning_rate=0.001) + sgd.minimize(loss) + + # Here the fluid.default_startup_program() and fluid.default_main_program() + # has been constructed. + +在上述模型配置执行完毕后， :code:`fluid.default_startup_program()` 与\ +:code:`fluid.default_main_program()` 配置完毕了。 + +初始化参数 +########## + +参数随机初始化 +============== + +用户配置完模型后，参数初始化操作会被写入到\ +:code:`fluid.default_startup_program()` 中。使用 :code:`fluid.Executor()` 运行 +这一程序，即可在全局 :code:`fluid.global_scope()` 中随机初始化参数。例如: + +.. code-block:: python + + exe = fluid.Executor(fluid.CUDAPlace(0)) + exe.run(program=fluid.default_startup_program()) + +值得注意的是: 如果使用多GPU训练，参数需要先在GPU0上初始化，再经由\ +:code:`fluid.ParallelExecutor` 分发到多张显卡上。 + + +载入预定义参数 +============== + +在神经网络训练过程中，经常会需要载入预定义模型，进而继续进行训练。\ +如何载入预定义参数，请参考 :ref:`user_guide_save_load_vars`。 + + +单卡训练 +######## + +执行单卡训练可以使用 :code:`fluid.Executor()` 中的 :code:`run()` 方法，运行训练\ +:code:`fluid.Program` 即可。在运行的时候，用户可以通过 :code:`run(feed=...)`\ +参数传入数据；用户可以通过 :code:`run(fetch=...)` 获取持久的数据。例如:\ + +.. code-block:: python + + ... + loss = fluid.layers.mean(...) + + exe = fluid.Executor(...) + # the result is an numpy array + result = exe.run(feed={"image": ..., "label": ...}, fetch_list=[loss]) + +这里有几点注意事项: + +1. feed的数据格式，请参考文章 :ref:`user_guide_feed_data_to_executor`。 +2. :code:`Executor.run` 的返回值是 :code:`fetch_list=[...]` 的variable值。被fetch\ + 的Variable必须是persistable的。 :code:`fetch_list` 可以传入Variable的列表，\ + 也可以传入Variable的名字列表。:code:`Executor.run` 返回Fetch结果列表。 +3. 如果需要取回的数据包含序列信息，可以设置 + :code:`exe.run(return_numpy=False, ...)` 直接返回 :code:`fluid.LoDTensor` + 。用户可以直接访问 :code:`fluid.LoDTensor` 中的信息。 + +多卡训练 +######## + +执行多卡训练可以使用 :code:`fluid.ParallelExecutor` 运行训练 +:code:`fluid.Program`。例如: + +.. code-block:: python + + train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name, + main_program=fluid.default_main_program()) + train_exe.run(fetch_list=[loss.name], feed={...}) + +这里有几点注意事项: + +1. :code:`ParallelExecutor` 的构造函数需要指明要执行的 :code:`fluid.Program` , + 并在执行过程中不能修改。默认值是 :code:`fluid.default_main_program()` 。 +2. :code:`ParallelExecutor` 需要明确指定是否使用 CUDA 显卡进行训练。在显卡训练\ + 模式下会占用全部显卡。用户可以配置 `CUDA_VISIBLE_DEVICES `_ 来修改占用\ + 的显卡。 + +进阶使用 +######## + +.. toctree:: + :maxdepth: 2 + + test_while_training + save_load_variables diff --git a/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_nccl2.graffle b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_nccl2.graffle new file mode 100644 index 0000000000000000000000000000000000000000..16f6b8835c4ffb82babca56b62ba44494fd6a947 Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_nccl2.graffle differ diff --git a/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_nccl2.png b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_nccl2.png new file mode 100644 index 0000000000000000000000000000000000000000..587a1a48affdde6809d7f8bf77e1055db7cd8c14 Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_nccl2.png differ diff --git a/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_pserver.graffle b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_pserver.graffle new file mode 100644 index 0000000000000000000000000000000000000000..046c4903231e8ca441884674c08b381766c0bbae Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_pserver.graffle differ diff --git a/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_pserver.png b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_pserver.png new file mode 100644 index 0000000000000000000000000000000000000000..cd2f92ad1a14ac12efc2c257c8aa3d1ae403b2b1 Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/training/src/dist_train_pserver.png differ diff --git a/doc/fluid/new_docs/user_guides/howto/training/src/parallelism.png b/doc/fluid/new_docs/user_guides/howto/training/src/parallelism.png new file mode 100644 index 0000000000000000000000000000000000000000..6c078b5241559a05219447db67b5d8a35aeefd3f Binary files /dev/null and b/doc/fluid/new_docs/user_guides/howto/training/src/parallelism.png differ diff --git a/doc/fluid/new_docs/user_guides/howto/training/test_while_training.rst b/doc/fluid/new_docs/user_guides/howto/training/test_while_training.rst new file mode 100644 index 0000000000000000000000000000000000000000..37d5c0d78179ccead7a81dffb4ae2f0d835a5949 --- /dev/null +++ b/doc/fluid/new_docs/user_guides/howto/training/test_while_training.rst @@ -0,0 +1,120 @@ +.. _user_guide_test_while_training: + +################## +训练过程中评测模型 +################## + +模型的测试评价与训练的 :code:`fluid.Program` 不同。在测试评价中: + +1. 评价测试不进行反向传播，不优化更新参数。 +2. 评价测试执行的操作可以不同。 + + * 例如 BatchNorm 操作，在训练和测试时执行不同的算法。 + + * 评价模型与训练相比可以是完全不同的模型。 + +生成测试 :code:`fluid.Program` +################################# + +通过克隆训练 :code:`fluid.Program` 生成测试 :code:`fluid.Program` +======================================================================= + +:code:`Program.clone()` 方法可以复制出新的 :code:`fluid.Program` 。 通过设置 +:code:`Program.clone(for_test=True)` 复制含有用于测试的操作Program。简单的使用方法如下: + +.. code-block:: python + + import paddle.fluid as fluid + + img = fluid.layers.data(name="image", shape=[784]) + prediction = fluid.layers.fc( + input=fluid.layers.fc(input=img, size=100, act='relu'), + size=10, + act='softmax' + ) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + loss = fluid.layers.mean(fluid.layers.cross_entropy(input=prediction, label=label)) + acc = fluid.layers.accuracy(input=prediction, label=label) + + test_program = fluid.default_main_program().clone(for_test=True) + + adam = fluid.optimizer.Adam(learning_rate=0.001) + adam.minimize(loss) + +在使用 :code:`Optimizer` 之前，将 :code:`fluid.default_main_program()` 复制\ +成一个 :code:`test_program` 。之后使用测试数据运行 :code:`test_program`,\ +就可以做到运行测试程序，而不影响训练结果。 + +分别配置训练 :code:`fluid.Program` 和测试 :code:`fluid.Program` +===================================================================== + +如果训练程序和测试程序相差较大时，用户也可以通过完全定义两个不同的 +:code:`fluid.Program`，分别进行训练和测试。在PaddlePaddle Fluid中，\ +所有的参数都有名字。如果两个不同的操作，甚至两个不同的网络使用了同样名字的参数，\ +那么他们的值和内存空间都是共享的。 + +PaddlePaddle Fluid中使用 :code:`fluid.unique_name` 包来随机初始化用户未定义的\ +参数名称。通过 :code:`fluid.unique_name.guard` 可以确保多次调用某函数\ +参数初始化的名称一致。 + +例如: + +.. code-block:: python + + import paddle.fluid as fluid + + def network(is_test): + file_obj = fluid.layers.open_files(filenames=["test.recordio"] if is_test else ["train.recordio"], ...) + img, label = fluid.layers.read_file(file_obj) + hidden = fluid.layers.fc(input=img, size=100, act="relu") + hidden = fluid.layers.batch_norm(input=hidden, is_test=is_test) + ... + return loss + + with fluid.unique_name.guard(): + train_loss = network(is_test=False) + sgd = fluid.optimizer.SGD(0.001) + sgd.minimize(train_loss) + + test_program = fluid.Program() + with fluid.unique_name.guard(): + with fluid.program_gurad(test_program, fluid.Program()): + test_loss = network(is_test=True) + + # fluid.default_main_program() is the train program + # fluid.test_program is the test program + +执行测试 :code:`fluid.Program` +################################# + +使用 :code:`Executor` 执行测试 :code:`fluid.Program` +======================================================= + +用户可以使用 :code:`Executor.run(program=...)` 来执行测试 +:code:`fluid.Program`。 + +例如 + +.. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + test_acc = exe.run(program=test_program, feed=test_data_batch, fetch_list=[acc]) + print 'Test accuracy is ', test_acc + +使用 :code:`ParallelExecutor` 执行测试 :code:`fluid.Program` +=============================================================== + +用户可以使用训练用的 :code:`ParallelExecutor` 与测试 :code:`fluid.Program` +一起新建一个测试的 :code:`ParallelExecutor` ；再使用测试 +:code:`ParallelExecutor.run` 来执行测试。 + +例如: + +.. code-block:: python + + train_exec = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name) + + test_exec = fluid.ParallelExecutor(use_cuda=True, share_vars_from=train_exec, + main_program=test_program) + test_acc = test_exec.run(fetch_list=[acc], ...) + diff --git a/doc/fluid/new_docs/user_guides/index.rst b/doc/fluid/new_docs/user_guides/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..453cb71cfdf72e031ce0f0517e2db936eca38dfc --- /dev/null +++ b/doc/fluid/new_docs/user_guides/index.rst @@ -0,0 +1,18 @@ +######## +使用指南 +######## + + +.. todo:: + + 完善导引介绍 + +.. toctree:: + :maxdepth: 2 + + howto/prepare_data/index + howto/configure_simple_model/index + howto/training/index + howto/debug/index + howto/evaluation/index + models/index.rst diff --git a/doc/fluid/new_docs/user_guides/models/index.rst b/doc/fluid/new_docs/user_guides/models/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..998e95c4885dc313d9449f5466f80c53d34fe82a --- /dev/null +++ b/doc/fluid/new_docs/user_guides/models/index.rst @@ -0,0 +1,137 @@ +Fluid 模型库 +============ + +图像分类 +-------- + +图像分类是根据图像的语义信息对不同类别图像进行区分，是计算机视觉中重要的基础问题，是物体检测、图像分割、物体跟踪、行为分析、人脸识别等其他高层视觉任务的基础，在许多领域都有着广泛的应用。如：安防领域的人脸识别和智能视频分析等，交通领域的交通场景识别，互联网领域基于内容的图像检索和相册自动归类，医学领域的图像识别等。 + +在深度学习时代，图像分类的准确率大幅度提升，在图像分类任务中，我们向大家介绍了如何在经典的数据集ImageNet上，训练常用的模型，包括AlexNet、VGG、GoogLeNet、ResNet、Inception-v4、MobileNet、DPN(Dual +Path +Network)、SE-ResNeXt模型，也开源了\ `训练的模型 `__\ 方便用户下载使用。同时提供了能够将Caffe模型转换为PaddlePaddle +Fluid模型配置和参数文件的工具。 + +- `AlexNet `__ +- `VGG `__ +- `GoogleNet `__ +- `Residual + Network `__ +- `Inception-v4 `__ +- `MobileNet `__ +- `Dual Path + Network `__ +- `SE-ResNeXt `__ +- `Caffe模型转换为Paddle + Fluid配置和模型文件工具 `__ + +目标检测 +-------- + +目标检测任务的目标是给定一张图像或是一个视频帧，让计算机找出其中所有目标的位置，并给出每个目标的具体类别。对于人类来说，目标检测是一个非常简单的任务。然而，计算机能够“看到”的是图像被编码之后的数字，很难解图像或是视频帧中出现了人或是物体这样的高层语义概念，也就更加难以定位目标出现在图像中哪个区域。与此同时，由于目标会出现在图像或是视频帧中的任何位置，目标的形态千变万化，图像或是视频帧的背景千差万别，诸多因素都使得目标检测对计算机来说是一个具有挑战性的问题。 + +在目标检测任务中，我们介绍了如何基于\ `PASCAL +VOC `__\ 、\ `MS +COCO `__\ 数据训练通用物体检测模型，当前介绍了SSD算法，SSD全称Single Shot MultiBox Detector，是目标检测领域较新且效果较好的检测算法之一，具有检测速度快且检测精度高的特点。 + +开放环境中的检测人脸，尤其是小的、模糊的和部分遮挡的人脸也是一个具有挑战的任务。我们也介绍了如何基于 `WIDER FACE `_ 数据训练百度自研的人脸检测PyramidBox模型，该算法于2018年3月份在WIDER FACE的多项评测中均获得 `第一名 `_。 + +- `Single Shot MultiBox + Detector `__ +- `Face Detector: PyramidBox `_ + +图像语义分割 +------------ + +图像语意分割顾名思义是将图像像素按照表达的语义含义的不同进行分组/分割，图像语义是指对图像内容的理解，例如，能够描绘出什么物体在哪里做了什么事情等，分割是指对图片中的每个像素点进行标注，标注属于哪一类别。近年来用在无人车驾驶技术中分割街景来避让行人和车辆、医疗影像分析中辅助诊断等。 + +在图像语义分割任务中，我们介绍如何基于图像级联网络(Image Cascade +Network,ICNet)进行语义分割，相比其他分割算法，ICNet兼顾了准确率和速度。 + +- `ICNet `__ + +场景文字识别 +------------ + +许多场景图像中包含着丰富的文本信息，对理解图像信息有着重要作用，能够极大地帮助人们认知和理解场景图像的内容。场景文字识别是在图像背景复杂、分辨率低下、字体多样、分布随意等情况下，将图像信息转化为文字序列的过程，可认为是一种特别的翻译过程：将图像输入翻译为自然语言输出。场景图像文字识别技术的发展也促进了一些新型应用的产生，如通过自动识别路牌中的文字帮助街景应用获取更加准确的地址信息等。 + +在场景文字识别任务中，我们介绍如何将基于CNN的图像特征提取和基于RNN的序列翻译技术结合，免除人工定义特征，避免字符分割，使用自动学习到的图像特征，完成端到端地无约束字符定位和识别。当前，介绍了CRNN-CTC模型，后续会引入基于注意力机制的序列到序列模型。 + +- `CRNN-CTC模型 `__ + +语音识别 +-------- + +自动语音识别（Automatic Speech Recognition, +ASR）是将人类声音中的词汇内容转录成计算机可输入的文字的技术。语音识别的相关研究经历了漫长的探索过程，在HMM/GMM模型之后其发展一直较为缓慢，随着深度学习的兴起，其迎来了春天。在多种语言识别任务中，将深度神经网络(DNN)作为声学模型，取得了比GMM更好的性能，使得 +ASR +成为深度学习应用最为成功的领域之一。而由于识别准确率的不断提高，有越来越多的语言技术产品得以落地，例如语言输入法、以智能音箱为代表的智能家居设备等 +—— 基于语言的交互方式正在深刻的改变人类的生活。 + +与 `DeepSpeech `__ +中深度学习模型端到端直接预测字词的分布不同，本实例更接近传统的语言识别流程，以音素为建模单元，关注语言识别中声学模型的训练，利用\ `kaldi `__\ 进行音频数据的特征提取和标签对齐，并集成 +kaldi 的解码器完成解码。 + +- `DeepASR `__ + +机器翻译 +-------- + +机器翻译（Machine +Translation）将一种自然语言(源语言)转换成一种自然语言（目标语音），是自然语言处理中非常基础和重要的研究方向。在全球化的浪潮中，机器翻译在促进跨语言文明的交流中所起的重要作用是不言而喻的。其发展经历了统计机器翻译和基于神经网络的神经机器翻译(Nueural +Machine Translation, NMT)等阶段。在 NMT +成熟后，机器翻译才真正得以大规模应用。而早阶段的 NMT +主要是基于循环神经网络 RNN +的，其训练过程中当前时间步依赖于前一个时间步的计算，时间步之间难以并行化以提高训练速度。因此，非 +RNN 结构的 NMT 得以应运而生，例如基于卷积神经网络 CNN +的结构和基于自注意力机制（Self-Attention）的结构。 + +本实例所实现的 Transformer +就是一个基于自注意力机制的机器翻译模型，其中不再有RNN或CNN结构，而是完全利用 +Attention 学习语言中的上下文依赖。相较于RNN/CNN, +这种结构在单层内计算复杂度更低、易于并行化、对长程依赖更易建模，最终在多种语言之间取得了最好的翻译效果。 + +- `Transformer `__ + +强化学习 +-------- + +强化学习是近年来一个愈发重要的机器学习方向，特别是与深度学习相结合而形成的深度强化学习(Deep +Reinforcement Learning, +DRL)，取得了很多令人惊异的成就。人们所熟知的战胜人类顶级围棋职业选手的 +AlphaGo 就是 DRL +应用的一个典型例子，除游戏领域外，其它的应用还包括机器人、自然语言处理等。 + +深度强化学习的开山之作是在Atari视频游戏中的成功应用， +其可直接接受视频帧这种高维输入并根据图像内容端到端地预测下一步的动作，所用到的模型被称为深度Q网络(Deep +Q-Network, DQN)。本实例就是利用PaddlePaddle Fluid这个灵活的框架，实现了 +DQN 及其变体，并测试了它们在 Atari 游戏中的表现。 + +- `DeepQNetwork `__ + +中文词法分析 +------------ + +中文分词(Word Segmentation)是将连续的自然语言文本，切分出具有语义合理性和完整性的词汇序列的过程。因为在汉语中，词是承担语义的最基本单位，切词是文本分类、情感分析、信息检索等众多自然语言处理任务的基础。 词性标注（Part-of-speech Tagging）是为自然语言文本中的每一个词汇赋予一个词性的过程，这里的词性包括名词、动词、形容词、副词等等。 命名实体识别（Named Entity Recognition，NER）又称作“专名识别”，是指识别自然语言文本中具有特定意义的实体，主要包括人名、地名、机构名、专有名词等。 我们将这三个任务统一成一个联合任务，称为词法分析任务，基于深度神经网络，利用海量标注语料进行训练，提供了一个端到端的解决方案。 + +我们把这个联合的中文词法分析解决方案命名为LAC。LAC既可以认为是Lexical Analysis of Chinese的首字母缩写，也可以认为是LAC Analyzes Chinese的递归缩写。 + +- `LAC `__ + +情感倾向分析 +------------ + +情感倾向分析针对带有主观描述的中文文本，可自动判断该文本的情感极性类别并给出相应的置信度。情感类型分为积极、消极、 中性。情感倾向分析能够帮助企业理解用户消费习惯、分析热点话题和危机舆情监控，为企业提供有力的决策支持。本次我们开放 AI开放平台中情感倾向分析采用的模型(http://ai.baidu.com/tech/nlp/sentiment_classify )， 提供给用户使用。 + +- `Senta `__ + +AnyQ +---- + +`AnyQ `__\ (ANswer Your Questions) +开源项目主要包含面向FAQ集合的问答系统框架、文本语义匹配工具SimNet。 +问答系统框架采用了配置化、插件化的设计，各功能均通过插件形式加入，当前共开放了20+种插件。开发者可以使用AnyQ系统快速构建和定制适用于特定业务场景的FAQ问答系统，并加速迭代和升级。 + +SimNet是百度自然语言处理部于2013年自主研发的语义匹配框架，该框架在百度各产品上广泛应用，主要包括BOW、CNN、RNN、MM-DNN等核心网络结构形式，同时基于该框架也集成了学术界主流的语义匹配模型，如MatchPyramid、MV-LSTM、K-NRM等模型。使用SimNet构建出的模型可以便捷的加入AnyQ系统中，增强AnyQ系统的语义匹配能力。 + +- `SimNet in PaddlePaddle + Fluid `__ diff --git a/doc/mobile/cross_compiling_for_android_cn.md b/doc/mobile/cross_compiling_for_android_cn.md index cdd6917239371a660d0df05bb623f0b94f8f11a3..0607748b751e9f2d606236d9e98868335379b05c 100644 --- a/doc/mobile/cross_compiling_for_android_cn.md +++ b/doc/mobile/cross_compiling_for_android_cn.md @@ -63,16 +63,16 @@ Android的Docker开发镜像向用户提供两个可配置的参数： - 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库 ```bash -$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev +$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev ./paddle/scripts/paddle_build.sh build_android ``` - 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库 ```bash -$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev +$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev ./paddle/scripts/paddle_build.sh build_android ``` -执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。 +执行上述`docker run`命令时，容器执行[paddle/scripts/paddle_build.sh build_android](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/paddle_build.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。 ## 基于Linux交叉编译环境的编译方式 本文档将以Linux x86-64平台为例，介绍交叉编译Android平台上适用的PaddlePaddle库的方法和步骤。 diff --git a/doc/mobile/cross_compiling_for_android_en.md b/doc/mobile/cross_compiling_for_android_en.md index 6af16fc114a2310e364023ec43cc3c64149af8f7..572063e8012efee2d2e142eb57e459e0e8c6382c 100644 --- a/doc/mobile/cross_compiling_for_android_en.md +++ b/doc/mobile/cross_compiling_for_android_en.md @@ -36,7 +36,7 @@ $ docker pull docker.paddlepaddlehub.com/paddle:latest-dev-android We can run the Docker image we just created to build the inference library of PaddlePaddle for Android using the command below: ```bash -$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android +$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android ./paddle/scripts/paddle_build.sh build_android ``` The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`: @@ -70,7 +70,7 @@ The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`: The ARM-64 architecture (`arm64-v8a`) requires at least level 21 of Android API. -The default entry-point of the Docker image, [`paddle/scripts/docker/build_android.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`. For information about other configuration arguments, please continue reading. +The build command, [`paddle/scripts/paddle_build.sh build_android`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/paddle_build.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`. For information about other configuration arguments, please continue reading. The above command generates and outputs the inference library in `$PWD/install_android` and puts third-party libraries in `$PWD/install_android/third_party`. diff --git a/doc/survey/dynamic_graph.md b/doc/survey/dynamic_graph.md new file mode 100644 index 0000000000000000000000000000000000000000..6b80b014b1b1dc50f425e1296f70984c9e9b1cbd --- /dev/null +++ b/doc/survey/dynamic_graph.md @@ -0,0 +1,378 @@ +# Automatic Differentiation with the Tape + +## Automatic Differentiation + +A key challenge in the field of deep learning is to automatically derive the backward pass from the forward pass described algorithmically by researchers. Such a derivation, or a transformation of the forward pass program, has been long studied before the recent prosperity of deep learning in the field known as [automatic differentiation](https://arxiv.org/pdf/1502.05767.pdf). + +## The Tape + +Given the forward pass program (usually in Python in practices), there are two strategies to derive the backward pass: + +1. from the forward pass program itself, or +1. from the execution trace of the forward pass program, which is often known as the *tape*. + +This article surveys systems that follow the latter strategy. + +## Dynamic Network + +When we train a deep learning model, the tape changes every iteration as the input data change, so we have to re-derive the backward pass every iteration. This is known as *dynamic network*. + +Deep learning systems that utilize the idea of dynamic network gained their popularities in recent years. This article surveys two representative systems: [PyTorch](https://pytorch.org/) and [DyNet](https://dynet.readthedocs.io/en/latest/). + +## An Overview + +Both frameworks record a ‘tape’ of the computation and interpreting (or run-time compiling) a transformation of the tape played back in reverse. This tape is a different kind of entity than the original program.[[link]](http://www.bcl.hamilton.ie/~barak/papers/toplas-reverse.pdf) + +Consider the following code feedforward model. + +```python +x = Variable(randn(20, 1))) +label = Variable(randint(1)) +W_1, W_2 = Variable(randn(20, 20)), Variable(randn(10, 20)) +h = matmul(W_1, x) +pred = matmul(W_2, x) +loss = softmax(pred, label) +loss.backward() +``` + +### 1) Dynet uses List to encode the Tape + +During the forward execution, a list of operators, in this case `matmul`, `matmul` and `softmax`, are recorded in the tape, along with the necessary information needed to do the backward such as pointers to the inputs and outputs. Then the tape is played in reverse order at `loss.backward()`. + +
+ +digraph g { + graph [ + rankdir = "LR" + ]; + node [ + fontsize = "16" + shape = "ellipse" + ]; + edge []; + "node0" [ + label = " type: matmul | input: W_1, x | output: h" + shape = "record" + ]; + "node1" [ + label = " type: matmul | input: W_2, h | output: pred" + shape = "record" + ]; + "node2" [ + label = " type: softmax | input: pred, label | output: loss" + shape = "record" + ]; + "node0":f0 -> "node1":f0 []; + "node1":f0 -> "node2":f0 []; +} +
+ +![Alt text](https://g.gravizo.com/svg?digraph%20g%20{%20graph%20[%20rankdir%20=%20%22LR%22%20];%20node%20[%20fontsize%20=%20%2216%22%20shape%20=%20%22ellipse%22%20];%20edge%20[];%20%22node0%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20%3Cf1%3E%20input:%20W_1,%20x%20|%20%3Cf2%3E%20output:%20h%22%20shape%20=%20%22record%22%20];%20%22node1%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20%3Cf1%3E%20input:%20W_2,%20h%20|%20%3Cf2%3E%20output:%20pred%22%20shape%20=%20%22record%22%20];%20%22node2%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20softmax%20|%20%3Cf1%3E%20input:%20pred,%20label%20|%20%3Cf2%3E%20output:%20loss%22%20shape%20=%20%22record%22%20];%20%22node0%22:f0%20-%3E%20%22node1%22:f0%20[%20id%20=%200%20];%20%22node1%22:f0%20-%3E%20%22node2%22:f0%20[%20id%20=%201%20];%20}) + +### 2) Pytorch uses Node Graph to encode the Tape + +The graph is composed of `Variable`s and `Function`s. During the forward execution, a `Variable` records its creator function, e.g. `h.creator = matmul`. And a Function records its inputs' previous/dependent functions `prev_func` through `creator`, e.g. `matmul.prev_func = matmul1`. At `loss.backward()`, a topological sort is performed on all `prev_func`s. Then the grad op is performed by the sorted order. + +
+ +digraph g { + graph [ + rankdir = "LR" + ]; + + subgraph function { + node [ + fontsize = "16" + style = filled + shape = "record" + ]; + "matmul0" [ label = " type: matmul | prev_func: None" ]; + "matmul1" [ label = " type: matmul | prev_func: matmul" ]; + "softmax" [ label = " type: softmax | prev_func: matmul" ]; + } + + subgraph variable { + node [ + fontsize = "16" + shape = "Mrecord" + style = filled + fillcolor = white + ]; + "x" [ label = " x | creator: None" ]; + "label" [ label = " label | creator: None" ]; + "W_1" [ label = " W_1 | creator: None" ]; + "W_2" [ label = " W_2 | creator: None" ]; + "h" [ label = " h | creator: None" ]; + "pred" [ label = " pred | creator: matmul" ]; + "loss" [ label = " loss | creator: softmax" ]; + } + + subgraph data_flow { + "x":f0 -> "matmul0":f0; + "W_1":f0 -> "matmul0":f0; + "matmul0":f0 -> "h":f0; + + "h":f0 -> "matmul1":f0; + "W_2":f0 -> "matmul1":f0; + "matmul1":f0 -> "pred":f0; + + "pred":f0 -> "softmax":f0; + "label":f0 -> "softmax":f0; + "softmax":f0 -> "loss":f0; + } + + subgraph prev_func { + edge [color="red", arrowsize="0.6", penwidth="1", constraint=false]; + "matmul1":f1 -> "matmul0":f0; + "softmax":f1 -> "matmul1":f0; + label = "prev_func"; + } +} +
+ +![Alt text](https://g.gravizo.com/svg?digraph%20g%20{%20graph%20[%20rankdir%20=%20%22LR%22%20];%20subgraph%20function%20{%20node%20[%20fontsize%20=%20%2216%22%20style%20=%20filled%20shape%20=%20%22record%22%20];%20%22matmul0%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20prev_func:%20None%22%20];%20%22matmul1%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20prev_func:%20matmul%22%20];%20%22softmax%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20softmax%20|%20prev_func:%20matmul%22%20];%20}%20subgraph%20variable%20{%20node%20[%20fontsize%20=%20%2216%22%20shape%20=%20%22Mrecord%22%20style%20=%20filled%20fillcolor%20=%20white%20];%20%22x%22%20[%20label%20=%20%22%3Cf0%3E%20x%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22label%22%20[%20label%20=%20%22%3Cf0%3E%20label%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22W_1%22%20[%20label%20=%20%22%3Cf0%3E%20W_1%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22W_2%22%20[%20label%20=%20%22%3Cf0%3E%20W_2%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22h%22%20[%20label%20=%20%22%3Cf0%3E%20h%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22pred%22%20[%20label%20=%20%22%3Cf0%3E%20pred%20|%20%3Cf1%3E%20creator:%20matmul%22%20];%20%22loss%22%20[%20label%20=%20%22%3Cf0%3E%20loss%20|%20%3Cf1%3E%20creator:%20softmax%22%20];%20}%20subgraph%20data_flow%20{%20%22x%22:f0%20-%3E%20%22matmul0%22:f0;%20%22W_1%22:f0%20-%3E%20%22matmul0%22:f0;%20%22matmul0%22:f0%20-%3E%20%22h%22:f0;%20%22h%22:f0%20-%3E%20%22matmul1%22:f0;%20%22W_2%22:f0%20-%3E%20%22matmul1%22:f0;%20%22matmul1%22:f0%20-%3E%20%22pred%22:f0;%20%22pred%22:f0%20-%3E%20%22softmax%22:f0;%20%22label%22:f0%20-%3E%20%22softmax%22:f0;%20%22softmax%22:f0%20-%3E%20%22loss%22:f0;%20}%20subgraph%20prev_func%20{%20edge%20[color=%22red%22,%20arrowsize=%220.6%22,%20penwidth=%221%22,%20constraint=false];%20%22matmul1%22:f1%20-%3E%20%22matmul0%22:f0;%20%22softmax%22:f1%20-%3E%20%22matmul1%22:f0;%20label%20=%20%22prev_func%22;%20}%20}) + +Chainer and Autograd uses the similar techniques to record the forward pass. For details please refer to the appendix. + +## Design choices + +### 1) Dynet's List vs Pytorch's Node Graph + +What's good about List: +1. It avoids a topological sort. One only needs to traverse the list of operators in reverse and calling the corresponding backward operator. +1. It promises effient data parallelism implementations. One could count the time of usage of a certain variable during the construction list. Then in the play back, one knows the calculation of a variable has completed. This enables communication and computation overlapping. + +What's good about Node Graph: +1. More flexibility. PyTorch users can mix and match independent graphs however they like, in whatever threads they like (without explicit synchronization). An added benefit of structuring graphs this way is that when a portion of the graph becomes dead, it is automatically freed. [[2]](https://openreview.net/pdf?id=BJJsrmfCZ) Consider the following example, Pytorch only does backward on SmallNet while Dynet does both BigNet and SmallNet. +```python +result = BigNet(data) +loss = SmallNet(data) +loss.backward() +``` + +### 2) Dynet's Lazy evaluation vs Pytorch's Immediate evaluation + +Dynet builds the list in a symbolic matter. Consider the following example +```python +for epoch in range(num_epochs): + for in_words, out_label in training_data: + dy.renew_cg() + W = dy.parameter(W_p) + b = dy.parameter(b_p) + score_sym = dy.softmax(W*dy.concatenate([E[in_words[0]],E[in_words[1]]])+b) + loss_sym = dy.pickneglogsoftmax(score_sym, out_label) + loss_val = loss_sym.value() + loss_sym.backward() +``` +The computation of `lookup`, `concat`, `matmul` and `softmax` didn't happen until the call of `loss_sym.value()`. This defered execution is useful because it allows some graph-like optimization possible, e.g. kernel fusion. + +Pytorch chooses immediate evaluation. It avoids ever materializing a "forward graph"/"tape" (no need to explicitly call `dy.renew_cg()` to reset the list), recording only what is necessary to differentiate the computation, i.e. `creator` and `prev_func`. + + +## What can fluid learn from them? + +Please refer to `paddle/contrib/dynamic/`. + +# Appendix + +### Overview + +| Framework | Has Tape | Core in C++ | First Release Date | +|-----------|----------|-------------|--------------------| +| Autograd | No | No | Mar 5, 2015 | +| Chainer | No | No | Jun 5, 2015 | +| Pytorch | No | Yes | Aug 31, 2016 | +| Dynet | Yes | Yes | Oct 12, 2016 | + +### Source Code +#### Autograd +[Backward code](https://github.com/HIPS/autograd/blob/442205dfefe407beffb33550846434baa90c4de7/autograd/core.py#L8-L40). In the forward pass, a graph of VJPNode is constructed. +```python +# User API +def make_grad(fun, x): + start_node = VJPNode.new_root() + end_value, end_node = trace(start_node, fun, x) + return backward_pass(g, end_node), end_value + +# trace the forward pass by creating VJPNodes +def trace(start_node, fun, x): + with trace_stack.new_trace() as t: + start_box = new_box(x, t, start_node) + end_box = fun(start_box) + return end_box._value, end_box._node + +def backward_pass(g, end_node): + outgrads = {end_node : (g, False)} + for node in toposort(end_node): + outgrad = outgrads.pop(node) + ingrads = node.vjp(outgrad[0]) + for parent, ingrad in zip(node.parents, ingrads): + outgrads[parent] = add_outgrads(outgrads.get(parent), ingrad) + return outgrad[0] + +# Every VJPNode corresponds to a op_grad +class VJPNode(Node): + __slots__ = ['parents', 'vjp'] + def __init__(self, value, fun, args, kwargs, parent_argnums, parents): + self.parents = parents + vjpmaker = primitive_vjps[fun] + self.vjp = vjpmaker(parent_argnums, value, args, kwargs) +``` +#### Chainer +Example Code +```python +# (1) Function Set definition, creates FunctionNode +model = FunctionSet( + l1=F.Linear(784, 100), + l2=F.Linear(100, 100), + l3=F.Linear(100, 10)).to_gpu() + +# (2) Optimizer Setup +opt = optimizers.SGD() +opt.setup(model) + +# (3) Forward computation +def forward(x, t): + h1 = F.relu(model.l1(x)) + h2 = F.relu(model.l2(h1)) + y = model.l3(h2) + return F.softmax_cross_entropy(y, t) + +# (4) Training loop +for epoch in xrange(n_epoch): + for i in xrange(0, N, b_size): + x = Variable(to_gpu(...)) + t = Variable(to_gpu(...)) + opt.zero_grads() + loss = forward(x, t) + loss.backward() + opt.update() +``` +In `forward(x, t)`, a graph of [`VariableNode`](https://github.com/chainer/chainer/blob/master/chainer/variable.py#L110) and [`FunctionNode`](https://github.com/chainer/chainer/blob/a69103a4aa59d5b318f39b01dbcb858d465b89cf/chainer/function_node.py#L19) is constructed. Every output's `VariableNode.creator` is pointed to the `FunctionNode`. +```python +class FunctionNode(object): + ... + def apply(self, inputs): + outputs = self.forward(inputs) + ret = tuple([variable.Variable(y, requires_grad=requires_grad) + for y in outputs]) + # Topological ordering + self.rank = max([x.rank for x in inputs]) if input_vars else 0 + # Add backward edges + for y in ret: + y.creator_node = self + self.inputs = tuple([x.node for x in input_vars]) + self.outputs = tuple([y.node for y in ret]) + + return ret +``` +`loss.backward()` will calculate the accumulated gradient of all variables. All the backward of `FunctionNode`s will be called based on the topological order. +```python +class VariableNode(object): + ... + def backward(self, retain_grad, loss_scale): + if self.creator_node is None: + return + + cand_funcs = [] + seen_set = set() + grads = {} + + # Initialize error by 1, if this is a loss variable + if self.data.size == 1 and self._grad_var is None: + self.grad = numpy.ones_like(self.data) + grads[self._node] = self._grad_var + + def add_cand(cand): + if cand not in seen_set: + # Negate since heapq is min-heap. This is a global variable + heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand)) + seen_set.add(cand) + + add_cand(self.creator_node) + + while cand_funcs: + _, _, func = heapq.heappop(cand_funcs) + gxs = func.backward_accumulate(func.inputs, func.outputs, func.outputs.grad) + + for x, gx in enumerate(gxs): + if x in grads: + grads[x] += gx + else: + grads[x] = gx + + if x.creator_node is not None: + add_cand(x.creator_node) +``` + +#### PyTorch +Example Code +```python +x = Variable(torch.ones(5, 5)) +y = Variable(torch.ones(5, 5) * 4) +z = x ** 2 + x * 2 + x * y + y +z.backward(torch.ones(5, 5)) +``` +The trace is done by `Variable.creator` and `Function.previous_functions`. +```python +class Variable(object): + def __init__(self, tensor, creator=None, requires_grad=True): + if creator is None: + creator = Leaf(self, requires_grad) + self.data = tensor + self.creator = creator + self._grad = None + + def backward(self, gradient=None): + if gradient is None: + if self.data.numel() != 1: + raise RuntimeError('backward should be called only on a scalar (i.e. 1-element tensor) or with gradient w.r.t. the variable') + gradient = self.data.new(1).fill_(1) + self._execution_engine.run_backward(self, gradient) + +class Function(obejct): + # ... + def _do_forward(self, *input): + unpacked_input = tuple(arg.data for arg in input) + raw_output = self.forward(*unpacked_input) + + # mark output.creator = self for backward trace + output = tuple(Variable(tensor, self) for tensor in raw_output) + + self.previous_functions = [(arg.creator, id(arg)) for arg in input] + self.output_ids = {id(var): i for i, var in enumerate(output)} + return output + + def _do_backward(self, grad_output): + return self.backwaerd(grad_output) +``` +The [backward](https://github.com/pytorch/pytorch/blob/v0.1.1/torch/autograd/engine.py) is similar to Autograd. + +#### DyNet +Example code +```python +model = dy.model() +W_p = model.add_parameters((20, 100)) +b_p = model.add_parameters(20) +E = model.add_lookup_parameters((20000, 50)) +for epoch in range(num_epochs): + for in_words, out_label in training_data: + dy.renew_cg() # init tape + W = dy.parameter(W_p) + b = dy.parameter(b_p) + score_sym = dy.softmax(W*dy.concatenate([E[in_words[0]],E[in_words[1]]])+b) + loss_sym = dy.pickneglogsoftmax(score_sym, out_label) + loss_val = loss_sym.value() + loss_sym.backward() +``` +[forward](https://github.com/clab/dynet/blob/740a9626a13a2732544de142e256ad0d0a166658/dynet/exec.cc#L84-L158), [backward](https://github.com/clab/dynet/blob/740a9626a13a2732544de142e256ad0d0a166658/dynet/exec.cc#L166-L284). The trace is done by creating a tape of expressions in every iteration. Backward is done by traverse the tape in the reverse order. +```c++ +void SimpleExecutionEngine::backward(VariableIndex from_where, bool full) { + ... + for (int i = num_nodes - 1; i >= 0; --i) { + // each node corresponds to an op + node->backward(xs, node_fx, node_dEdfx, ai, node_dEdxai); + } + ... +} +``` diff --git a/doc/survey/op_fusion_design.md b/doc/survey/op_fusion_design.md new file mode 100644 index 0000000000000000000000000000000000000000..d6e48f4f58269b67450cb012f6dcc59e1083abba --- /dev/null +++ b/doc/survey/op_fusion_design.md @@ -0,0 +1,20 @@ +# Operator fusion +Fusing multiple operators together is an important method to optimize the program execution, particularly for GPU or other specialized accelerators. An obvious benefit is to avoid the overhead of saving the intermediate result back into global memory. + +There are generally two ways to fuse operators, fusing directly connected operators and fusing non directly connected operators. The first method is mainly used by [NNVM Compiler](https://github.com/dmlc/tvm/) and [XLA](https://www.tensorflow.org/performance/xla/). The second method is mainly used by Dynet and TensorFlow Fold to do auto-batching. The principle of fusing operator is according to some rules to combine multiple operations into one, for example, `Y = X * W` and `Z = Y + B` can be fused to `Z = X * W + B`, and `Y1 = X1 * W` and `Y2 = X2 * W` can be fused to `[Y1;Y2] = [X1;X2] * W`. In order to get a short-term profit, we decided to try to manually specify these rules. + +## Challenge +The challenge of fusing operators is: + - how to make the rules. + - how to implement these rules efficiently. + +### How to make the rules? + +The problem of determining the best single location for a fusion operator is an NP-hard combinatorial problem. After analysis the operators of the DL model, we found there are two group of operators can be fused explicitly, one is the simple and adjacent operations, for example, `tmp = x + y` and `z = Relu(tmp)`, and the other is the operators that have the same function, for example, a serials of `SGD` or `Momentum`. They usually appear in the model in a large number. So we should think about how to fuse them separately first. + +### How to implement these rules efficiently? +#### How to fuse the adjacent operations efficiently? +Here we use a template function to represent the fused operations. The pros of using a template function are that it is simple and efficient, and the cons are that it is not easy to expand, and it can only be used to express some simple operations. So taking into account our current needs, the template function is more appropriate. + +#### How to fuse the operators that have the same function efficiently? +We take SGD operator as an example, the training model may have hundreds of parameters and correspondingly have the same number of SGD operators. The expression(`w = w - lr*w_g`) of those operators is the same, so during of training, the executor will execute this expression hundreds time in CPU or other specialized accelerators. If we can fuse them and make the address of all `w` and all `w_g` continuous respectively, we only need execute one time. For some accelerators, the time of launching kernel is not neglected, so the time of hundreds of times of launching and executing kernel may be larger than launching and executing only once. There usually are many operators that similar to `SGD` in the DL model, such as `AllReduce` and `FC`. diff --git a/doc/v2/api/config/evaluators.rst b/doc/v2/api/config/evaluators.rst index 9ac972fb193a2fb525edc507f7ba1303d2c8eabe..458d892e825a7a9bbe7843ad5c508bd5a31f5f0f 100644 --- a/doc/v2/api/config/evaluators.rst +++ b/doc/v2/api/config/evaluators.rst @@ -101,7 +101,7 @@ value_printer :noindex: Detection -===== +========== detection_map ------------- diff --git a/doc/v2/api/config/layer.rst b/doc/v2/api/config/layer.rst index 1a6496968cae1fef88142ba9ca3f9e63a81b196d..5a0cfadfce84df41defdf518b7c3a6222d5b30a1 100644 --- a/doc/v2/api/config/layer.rst +++ b/doc/v2/api/config/layer.rst @@ -11,7 +11,7 @@ Data layer data ---- -.. autoclass:: paddle.v2.layer.data +.. autofunction:: paddle.v2.layer.data :noindex: Fully Connected Layers @@ -21,12 +21,12 @@ Fully Connected Layers fc -- -.. autoclass:: paddle.v2.layer.fc +.. autofunction:: paddle.v2.layer.fc :noindex: selective_fc ------------ -.. autoclass:: paddle.v2.layer.selective_fc +.. autofunction:: paddle.v2.layer.selective_fc :noindex: Conv Layers @@ -34,34 +34,34 @@ Conv Layers conv_operator ------------- -.. autoclass:: paddle.v2.layer.conv_operator +.. autofunction:: paddle.v2.layer.conv_operator :noindex: conv_projection --------------- -.. autoclass:: paddle.v2.layer.conv_projection +.. autofunction:: paddle.v2.layer.conv_projection :noindex: conv_shift ---------- -.. autoclass:: paddle.v2.layer.conv_shift +.. autofunction:: paddle.v2.layer.conv_shift :noindex: img_conv -------- -.. autoclass:: paddle.v2.layer.img_conv +.. autofunction:: paddle.v2.layer.img_conv :noindex: .. _api_v2.layer_context_projection: context_projection ------------------ -.. autoclass:: paddle.v2.layer.context_projection +.. autofunction:: paddle.v2.layer.context_projection :noindex: row_conv -------- -.. autoclass:: paddle.v2.layer.row_conv +.. autofunction:: paddle.v2.layer.row_conv :noindex: Image Pooling Layer @@ -69,27 +69,27 @@ Image Pooling Layer img_pool -------- -.. autoclass:: paddle.v2.layer.img_pool +.. autofunction:: paddle.v2.layer.img_pool :noindex: spp --- -.. autoclass:: paddle.v2.layer.spp +.. autofunction:: paddle.v2.layer.spp :noindex: maxout ------ -.. autoclass:: paddle.v2.layer.maxout +.. autofunction:: paddle.v2.layer.maxout :noindex: roi_pool -------- -.. autoclass:: paddle.v2.layer.roi_pool +.. autofunction:: paddle.v2.layer.roi_pool :noindex: pad ---- -.. autoclass:: paddle.v2.layer.pad +.. autofunction:: paddle.v2.layer.pad :noindex: Norm Layer @@ -97,27 +97,27 @@ Norm Layer img_cmrnorm ----------- -.. autoclass:: paddle.v2.layer.img_cmrnorm +.. autofunction:: paddle.v2.layer.img_cmrnorm :noindex: batch_norm ---------- -.. autoclass:: paddle.v2.layer.batch_norm +.. autofunction:: paddle.v2.layer.batch_norm :noindex: sum_to_one_norm --------------- -.. autoclass:: paddle.v2.layer.sum_to_one_norm +.. autofunction:: paddle.v2.layer.sum_to_one_norm :noindex: cross_channel_norm ------------------ -.. autoclass:: paddle.v2.layer.cross_channel_norm +.. autofunction:: paddle.v2.layer.cross_channel_norm :noindex: row_l2_norm ----------- -.. autoclass:: paddle.v2.layer.row_l2_norm +.. autofunction:: paddle.v2.layer.row_l2_norm :noindex: Recurrent Layers @@ -125,22 +125,22 @@ Recurrent Layers recurrent --------- -.. autoclass:: paddle.v2.layer.recurrent +.. autofunction:: paddle.v2.layer.recurrent :noindex: lstmemory --------- -.. autoclass:: paddle.v2.layer.lstmemory +.. autofunction:: paddle.v2.layer.lstmemory :noindex: grumemory --------- -.. autoclass:: paddle.v2.layer.grumemory +.. autofunction:: paddle.v2.layer.grumemory :noindex: gated_unit ----------- -.. autoclass:: paddle.v2.layer.gated_unit +.. autofunction:: paddle.v2.layer.gated_unit :noindex: Recurrent Layer Group @@ -148,32 +148,32 @@ Recurrent Layer Group memory ------ -.. autoclass:: paddle.v2.layer.memory +.. autofunction:: paddle.v2.layer.memory :noindex: recurrent_group --------------- -.. autoclass:: paddle.v2.layer.recurrent_group +.. autofunction:: paddle.v2.layer.recurrent_group :noindex: lstm_step --------- -.. autoclass:: paddle.v2.layer.lstm_step +.. autofunction:: paddle.v2.layer.lstm_step :noindex: gru_step -------- -.. autoclass:: paddle.v2.layer.gru_step +.. autofunction:: paddle.v2.layer.gru_step :noindex: beam_search ------------ -.. autoclass:: paddle.v2.layer.beam_search +.. autofunction:: paddle.v2.layer.beam_search :noindex: get_output ---------- -.. autoclass:: paddle.v2.layer.get_output +.. autofunction:: paddle.v2.layer.get_output :noindex: Mixed Layer @@ -183,54 +183,54 @@ Mixed Layer mixed ----- -.. autoclass:: paddle.v2.layer.mixed +.. autofunction:: paddle.v2.layer.mixed :noindex: .. _api_v2.layer_embedding: embedding --------- -.. autoclass:: paddle.v2.layer.embedding +.. autofunction:: paddle.v2.layer.embedding :noindex: scaling_projection ------------------ -.. autoclass:: paddle.v2.layer.scaling_projection +.. autofunction:: paddle.v2.layer.scaling_projection :noindex: dotmul_projection ----------------- -.. autoclass:: paddle.v2.layer.dotmul_projection +.. autofunction:: paddle.v2.layer.dotmul_projection :noindex: dotmul_operator --------------- -.. autoclass:: paddle.v2.layer.dotmul_operator +.. autofunction:: paddle.v2.layer.dotmul_operator :noindex: full_matrix_projection ---------------------- -.. autoclass:: paddle.v2.layer.full_matrix_projection +.. autofunction:: paddle.v2.layer.full_matrix_projection :noindex: identity_projection ------------------- -.. autoclass:: paddle.v2.layer.identity_projection +.. autofunction:: paddle.v2.layer.identity_projection :noindex: slice_projection ------------------- -.. autoclass:: paddle.v2.layer.slice_projection +.. autofunction:: paddle.v2.layer.slice_projection :noindex: table_projection ---------------- -.. autoclass:: paddle.v2.layer.table_projection +.. autofunction:: paddle.v2.layer.table_projection :noindex: trans_full_matrix_projection ---------------------------- -.. autoclass:: paddle.v2.layer.trans_full_matrix_projection +.. autofunction:: paddle.v2.layer.trans_full_matrix_projection :noindex: Aggregate Layers @@ -245,51 +245,46 @@ AggregateLevel pooling ------- -.. autoclass:: paddle.v2.layer.pooling +.. autofunction:: paddle.v2.layer.pooling :noindex: .. _api_v2.layer_last_seq: last_seq -------- -.. autoclass:: paddle.v2.layer.last_seq +.. autofunction:: paddle.v2.layer.last_seq :noindex: .. _api_v2.layer_first_seq: first_seq --------- -.. autoclass:: paddle.v2.layer.first_seq +.. autofunction:: paddle.v2.layer.first_seq :noindex: sub_seq --------- -.. autoclass:: paddle.v2.layer.sub_seq +.. autofunction:: paddle.v2.layer.sub_seq :noindex: concat ------ -.. autoclass:: paddle.v2.layer.concat +.. autofunction:: paddle.v2.layer.concat :noindex: seq_concat ---------- -.. autoclass:: paddle.v2.layer.seq_concat +.. autofunction:: paddle.v2.layer.seq_concat :noindex: seq_slice --------- -.. autoclass:: paddle.v2.layer.seq_slice - :noindex: - -kmax_sequence_score -------------------- -.. autoclass:: paddle.v2.layer.kmax_sequence_score +.. autofunction:: paddle.v2.layer.seq_slice :noindex: sub_nested_seq -------------- -.. autoclass:: paddle.v2.layer.sub_nested_seq +.. autofunction:: paddle.v2.layer.sub_nested_seq :noindex: Reshaping Layers @@ -297,7 +292,7 @@ Reshaping Layers block_expand ------------ -.. autoclass:: paddle.v2.layer.block_expand +.. autofunction:: paddle.v2.layer.block_expand :noindex: .. _api_v2.layer_expand: @@ -309,22 +304,22 @@ ExpandLevel expand ------ -.. autoclass:: paddle.v2.layer.expand +.. autofunction:: paddle.v2.layer.expand :noindex: repeat ------ -.. autoclass:: paddle.v2.layer.repeat +.. autofunction:: paddle.v2.layer.repeat :noindex: rotate ------ -.. autoclass:: paddle.v2.layer.rotate +.. autofunction:: paddle.v2.layer.rotate :noindex: seq_reshape ----------- -.. autoclass:: paddle.v2.layer.seq_reshape +.. autofunction:: paddle.v2.layer.seq_reshape :noindex: Math Layers @@ -332,94 +327,94 @@ Math Layers addto ----- -.. autoclass:: paddle.v2.layer.addto +.. autofunction:: paddle.v2.layer.addto :noindex: linear_comb ----------- -.. autoclass:: paddle.v2.layer.linear_comb +.. autofunction:: paddle.v2.layer.linear_comb :noindex: interpolation ------------- -.. autoclass:: paddle.v2.layer.interpolation +.. autofunction:: paddle.v2.layer.interpolation :noindex: bilinear_interp --------------- -.. autoclass:: paddle.v2.layer.bilinear_interp +.. autofunction:: paddle.v2.layer.bilinear_interp :noindex: dropout -------- -.. autoclass:: paddle.v2.layer.dropout +.. autofunction:: paddle.v2.layer.dropout :noindex: dot_prod --------- -.. autoclass:: paddle.v2.layer.dot_prod +.. autofunction:: paddle.v2.layer.dot_prod :noindex: out_prod -------- -.. autoclass:: paddle.v2.layer.out_prod +.. autofunction:: paddle.v2.layer.out_prod :noindex: power ----- -.. autoclass:: paddle.v2.layer.power +.. autofunction:: paddle.v2.layer.power :noindex: scaling ------- -.. autoclass:: paddle.v2.layer.scaling +.. autofunction:: paddle.v2.layer.scaling :noindex: clip ---- -.. autoclass:: paddle.v2.layer.clip +.. autofunction:: paddle.v2.layer.clip :noindex: resize ------ -.. autoclass:: paddle.v2.layer.resize +.. autofunction:: paddle.v2.layer.resize :noindex: slope_intercept --------------- -.. autoclass:: paddle.v2.layer.slope_intercept +.. autofunction:: paddle.v2.layer.slope_intercept :noindex: tensor ------ -.. autoclass:: paddle.v2.layer.tensor +.. autofunction:: paddle.v2.layer.tensor :noindex: .. _api_v2.layer_cos_sim: cos_sim ------- -.. autoclass:: paddle.v2.layer.cos_sim +.. autofunction:: paddle.v2.layer.cos_sim :noindex: l2_distance ----------- -.. autoclass:: paddle.v2.layer.l2_distance +.. autofunction:: paddle.v2.layer.l2_distance :noindex: trans ----- -.. autoclass:: paddle.v2.layer.trans +.. autofunction:: paddle.v2.layer.trans :noindex: scale_shift ----------- -.. autoclass:: paddle.v2.layer.scale_shift +.. autofunction:: paddle.v2.layer.scale_shift :noindex: factorization_machine --------------------- -.. autoclass:: paddle.v2.layer.factorization_machine +.. autofunction:: paddle.v2.layer.factorization_machine :noindex: Sampling Layers @@ -427,17 +422,17 @@ Sampling Layers maxid ----- -.. autoclass:: paddle.v2.layer.max_id +.. autofunction:: paddle.v2.layer.max_id :noindex: sampling_id ----------- -.. autoclass:: paddle.v2.layer.sampling_id +.. autofunction:: paddle.v2.layer.sampling_id :noindex: multiplex --------- -.. autoclass:: paddle.v2.layer.multiplex +.. autofunction:: paddle.v2.layer.multiplex :noindex: .. _api_v2.layer_costs: @@ -447,97 +442,97 @@ Cost Layers cross_entropy_cost ------------------ -.. autoclass:: paddle.v2.layer.cross_entropy_cost +.. autofunction:: paddle.v2.layer.cross_entropy_cost :noindex: cross_entropy_with_selfnorm_cost -------------------------------- -.. autoclass:: paddle.v2.layer.cross_entropy_with_selfnorm_cost +.. autofunction:: paddle.v2.layer.cross_entropy_with_selfnorm_cost :noindex: multi_binary_label_cross_entropy_cost ------------------------------------- -.. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost +.. autofunction:: paddle.v2.layer.multi_binary_label_cross_entropy_cost :noindex: classification_cost ------------------- -.. autoclass:: paddle.v2.layer.classification_cost +.. autofunction:: paddle.v2.layer.classification_cost :noindex: huber_regression_cost ------------------------- -.. autoclass:: paddle.v2.layer.huber_regression_cost +.. autofunction:: paddle.v2.layer.huber_regression_cost :noindex: huber_classification_cost ------------------------- -.. autoclass:: paddle.v2.layer.huber_classification_cost +.. autofunction:: paddle.v2.layer.huber_classification_cost :noindex: lambda_cost ----------- -.. autoclass:: paddle.v2.layer.lambda_cost +.. autofunction:: paddle.v2.layer.lambda_cost :noindex: square_error_cost ----------------- -.. autoclass:: paddle.v2.layer.square_error_cost +.. autofunction:: paddle.v2.layer.square_error_cost :noindex: rank_cost --------- -.. autoclass:: paddle.v2.layer.rank_cost +.. autofunction:: paddle.v2.layer.rank_cost :noindex: sum_cost --------- -.. autoclass:: paddle.v2.layer.sum_cost +.. autofunction:: paddle.v2.layer.sum_cost :noindex: crf --- -.. autoclass:: paddle.v2.layer.crf +.. autofunction:: paddle.v2.layer.crf :noindex: crf_decoding ------------ -.. autoclass:: paddle.v2.layer.crf_decoding +.. autofunction:: paddle.v2.layer.crf_decoding :noindex: ctc --- -.. autoclass:: paddle.v2.layer.ctc +.. autofunction:: paddle.v2.layer.ctc :noindex: warp_ctc -------- -.. autoclass:: paddle.v2.layer.warp_ctc +.. autofunction:: paddle.v2.layer.warp_ctc :noindex: nce --- -.. autoclass:: paddle.v2.layer.nce +.. autofunction:: paddle.v2.layer.nce :noindex: hsigmoid --------- -.. autoclass:: paddle.v2.layer.hsigmoid +.. autofunction:: paddle.v2.layer.hsigmoid :noindex: smooth_l1_cost -------------- -.. autoclass:: paddle.v2.layer.smooth_l1_cost +.. autofunction:: paddle.v2.layer.smooth_l1_cost :noindex: multibox_loss -------------- -.. autoclass:: paddle.v2.layer.multibox_loss +.. autofunction:: paddle.v2.layer.multibox_loss :noindex: detection_output ---------------- -.. autoclass:: paddle.v2.layer.detection_output +.. autofunction:: paddle.v2.layer.detection_output :noindex: Check Layer @@ -545,7 +540,7 @@ Check Layer eos --- -.. autoclass:: paddle.v2.layer.eos +.. autofunction:: paddle.v2.layer.eos :noindex: Activation @@ -553,5 +548,5 @@ Activation prelu -------- -.. autoclass:: paddle.v2.layer.prelu +.. autofunction:: paddle.v2.layer.prelu :noindex: diff --git a/doc/v2/api/index_en.rst b/doc/v2/api/index_en.rst index b11cd449affd1dcd9d3f42492961469331350942..5813509dce46677444f0234db8e0eaa4f113e3a0 100644 --- a/doc/v2/api/index_en.rst +++ b/doc/v2/api/index_en.rst @@ -4,8 +4,6 @@ API .. toctree:: :maxdepth: 1 - overview.rst model_configs.rst data.rst run_logic.rst - fluid/index.rst diff --git a/doc/v2/build_and_install/build_from_source_cn.rst b/doc/v2/build_and_install/build_from_source_cn.rst index 741c01ce5428c0046daa5a784da70d4bb492438c..d0dacb104f148c2aeb323365cbd6f014ae00ed5a 100644 --- a/doc/v2/build_and_install/build_from_source_cn.rst +++ b/doc/v2/build_and_install/build_from_source_cn.rst @@ -23,7 +23,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 在 `这里 `__ 找到 paddle_manylinux_devel 镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。 -如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。 +如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 :ref:`编译依赖 <_compile_deps>` 之后才能开始编译的步骤。 编译PaddlePaddle，需要执行： @@ -35,11 +35,16 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 # 2. 可选步骤：源码中构建用于编译PaddlePaddle的Docker镜像 docker build -t paddle:dev . # 3. 执行下面的命令编译CPU-Only的二进制 - docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build + docker run -it -v $PWD:/paddle -w /paddle -e "PYTHON_ABI=cp27-cp27mu" -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build # 4. 或者也可以使用为上述可选步骤构建的镜像（必须先执行第2步） docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev ./paddle/scripts/paddle_build.sh build -注：上述命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。 +注： + +- 上述命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。 + +- 如果您使用的是 manylinux 的镜像进行编译, 那么您需要通过环境变量 :code:`PYTHON_ABI` 来指定一个 `Python ABI `__. +PaddlePaddle目前支持的 Python ABI 有 :code:`cp27-cp27m` 和 :code:`cp27-cp27mu`. 编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装： @@ -106,7 +111,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 - 学习 Docker 有多难？ - 理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 + 理解 Docker 并不难，大概花十分钟看一下 `如何使用Docker `_ 。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 - 我可以用 IDE 吗？ @@ -123,7 +128,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 - 可以并行编译吗？ - 是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 + 是的。我们的 Docker image 运行一个 `Paddle编译Bash脚本 `_ 。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 - Docker 需要 sudo @@ -131,11 +136,11 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 - 在 Windows/MacOS 上编译很慢 - Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。 + Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考 `如何为Windows/Mac计算机上的Docker增加内存和虚拟机 `_ 。 - 磁盘不够 - 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。 + 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考 `如何删除Docker Container `_ 来清理这些内容。 .. _compile_deps: @@ -195,7 +200,7 @@ BLAS PaddlePaddle支持 `MKL `_ 和 `OpenBlAS `_ 两种BLAS库。默认使用MKL。如果使用MKL并且机器含有AVX2指令集， -还会下载MKL-DNN数学库，详细参考 `这里 `_ 。 +还会下载MKL-DNN数学库，详细参考 `mkldnn设计文档 `_ 。 如果关闭MKL，则会使用OpenBLAS作为BLAS库。 @@ -211,7 +216,7 @@ PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行， 编译选项的设置 ++++++++++++++ -PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 +PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 .. code-block:: bash diff --git a/doc/v2/build_and_install/build_from_source_en.rst b/doc/v2/build_and_install/build_from_source_en.rst index b06c43e19dcfc52ad0f074a85517a16744895a3a..664b68da8b7dd3e005ebf3ec34de77729e5ab355 100644 --- a/doc/v2/build_and_install/build_from_source_en.rst +++ b/doc/v2/build_and_install/build_from_source_en.rst @@ -11,7 +11,7 @@ To build PaddlePaddle, you need 1. A computer -- Linux, Windows, MacOS. 2. Docker. -Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. +Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. We run all the tools by running this image. .. _build_step: @@ -26,6 +26,8 @@ you can also find how to build and use paddle_manylinux_devel Docker image from `here `__ Or you can build your own image from source as the optional step below: +If you don't wish to use docker，you need to install several compile dependencies manually as :ref:`Compile Dependencies <_compile_deps>` shows to start compilation. + .. code-block:: bash # 1. clone the source code @@ -34,13 +36,18 @@ Or you can build your own image from source as the optional step below: # 2. Optional: build development docker image from source docker build -t paddle:dev . # 3. Run the following command to build a CPU-Only binaries - docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build + docker run -it -v $PWD:/paddle -w /paddle -e "PYTHON_ABI=cp27-cp27mu" -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build # 4. Or, use your built Docker image to build PaddlePaddle (must run step 2) docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev ./paddle/scripts/paddle_build.sh build -NOTE: The above command try to mount the current working directory (root directory of source code) +NOTE: + +- The above command try to mount the current working directory (root directory of source code) into :code:`/paddle` directory inside docker container. +- You need to pass in the required environment variable :code:`PYTHON_ABI` to specify a `Python ABI `__. +Currently PaddlePaddle supported Python ABIs include :code:`cp27-cp27m` and :code:`cp27-cp27mu` . + When the compile finishes, you can get the output whl package under build/python/dist, then you can choose to install the whl on local machine or copy it to the target machine. @@ -108,7 +115,7 @@ Frequently Asked Questions - How difficult is it to learn Docker? - It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. + It takes you ten minutes to read `an introductory article `_ and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. - Can I use my favorite IDE? @@ -125,7 +132,7 @@ Frequently Asked Questions - Does Docker do parallel building? - Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. + Our building Docker image runs a `Bash script `_ , which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. - Docker requires sudo @@ -133,11 +140,11 @@ Frequently Asked Questions - Docker on Windows/MacOS builds slowly - On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details. + On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to `this issue `_ for details. - Not enough disk space - Examples in this article use option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/). + Examples in this article use option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to `this article `_ . .. _compile_deps: diff --git a/doc/v2/build_and_install/pip_install_cn.rst b/doc/v2/build_and_install/pip_install_cn.rst index 853bdb21bbcf07ae1742d2196dbcfe4668828b7b..095da19cd41d29bfa72ab23abd24bec45f925a86 100644 --- a/doc/v2/build_and_install/pip_install_cn.rst +++ b/doc/v2/build_and_install/pip_install_cn.rst @@ -60,6 +60,7 @@ paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版 "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" .. _pip_dependency: diff --git a/doc/v2/build_and_install/pip_install_en.rst b/doc/v2/build_and_install/pip_install_en.rst index fecf6d3712feac3265100a6121901ba784f7d5cc..8406e4aa1fbb953c3b615b10d1bcb2c45974dde0 100644 --- a/doc/v2/build_and_install/pip_install_en.rst +++ b/doc/v2/build_and_install/pip_install_en.rst @@ -63,6 +63,7 @@ If the links below shows up the login form, just click "Log in as guest" to star "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" .. _pip_dependency: diff --git a/doc/v2/design/cluster_train/large_model_dist_train.md b/doc/v2/design/cluster_train/large_model_dist_train.md index 0c4b5bc24c854b7062d509249bea9c50d42bd5f1..edb0245ea083e791b7f32ac57a330698299fceda 100644 --- a/doc/v2/design/cluster_train/large_model_dist_train.md +++ b/doc/v2/design/cluster_train/large_model_dist_train.md @@ -52,7 +52,7 @@ In `trainer_internal.cpp:L93 trainOneBatch`: When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver. -In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`: +In `legacy/trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`: ```c++ if (fullSize) { diff --git a/doc/v2/design/interface/00.why_plain_c.md b/doc/v2/design/interface/00.why_plain_c.md index a1443093342c5a3ed698fb6b52a751dfc7cb5319..826ff3141bc2512b525cb44ac0f18b376ce57e92 100644 --- a/doc/v2/design/interface/00.why_plain_c.md +++ b/doc/v2/design/interface/00.why_plain_c.md @@ -65,7 +65,7 @@ paddle_error paddle_matrix_get_shape(paddle_matrix matrix, 而在CPP里面实现这个C的接口，文件 `paddle_matrix.cpp` ```cpp -#include "paddle/math/matrix.h" +#include "paddle/legacy/math/matrix.h" extern "C" paddle_error paddle_matrix_shape(paddle_matrix matrix, uint64_t *width, diff --git a/doc/v2/design/mkl/mkldnn.md b/doc/v2/design/mkl/mkldnn.md index bd5bcf6f67168c21cebb046a629b948d1661e75c..4876de0045979be20fa45bdc84d2594516f71c03 100644 --- a/doc/v2/design/mkl/mkldnn.md +++ b/doc/v2/design/mkl/mkldnn.md @@ -18,20 +18,20 @@ Figure 1. PaddlePaddle on IA 具体的完成状态可以参见[这里](https://github.com/PaddlePaddle/Paddle/projects/21)。 ## Contents - -- [Overview](#overview) -- [Actions](#actions) - - [CMake](#cmake) - - [Matrix](#matrix) - - [Layers](#layers) - - [Activations](#activations) - - [Parameters](#parameters) - - [Gradients](#gradients) - - [Unit Tests](#unit-tests) - - [Python API](#python-api) - - [Benchmarking](#benchmarking) - - [Others](#others) -- [Design Concerns](#design-concerns) + +- [Overview](#overview) +- [Actions](#actions) + - [CMake](#cmake) + - [Matrix](#matrix) + - [Layers](#layers) + - [Activations](#activations) + - [Parameters](#parameters) + - [Gradients](#gradients) + - [Unit Tests](#unit-tests) + - [Python API](#python-api) + - [Benchmarking](#benchmarking) + - [Others](#others) +- [Design Concerns](#design-concerns) ## Overview @@ -218,20 +218,20 @@ if use_mkldnn 我们总结出一些特别需要注意的点： 1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数， -我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。 -2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。 +我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。 +2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。 3. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。 -包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。 +包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。 4. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存， 同时数据格式就是`NCHW`，这样下一个cpu device就能拿到正确的数据。 在有普通的CPU layer时， `extOutVal_`和`extOutGrad_`的格式始终是`NCHW`或者`NC`。 ## References 1. [MKL small library](https://github.com/01org/mkl-dnn#linking-your-application)是[Intel MKL](https://software.intel.com/en-us/mkl)的一个子集。 -主要包括了深度学习相关的数学原语与操作，一般由MKL-DNN在发布[新版本](https://github.com/01org/mkl-dnn/releases)时一起更新。 +主要包括了深度学习相关的数学原语与操作，一般由MKL-DNN在发布[新版本](https://github.com/01org/mkl-dnn/releases)时一起更新。 2. [MKL-DNN System Requirements](https://github.com/01org/mkl-dnn#system-requirements)。 目前在PaddlePaddle中，仅会在支持AVX2指令集及以上的机器才使用MKL-DNN。 3. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。 -但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。 +但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。 4. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的cuDNN部分使用的也是`NCHW`，所以不存在这个问题)。 -所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。 +所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。 diff --git a/doc/v2/dev/contribute_to_paddle_cn.md b/doc/v2/dev/contribute_to_paddle_cn.md index add06e42f1bbd221b48eb83e4e84d4a7c89e7483..3244eedf918b93f9351258f1218dfb2d507c1a9c 100644 --- a/doc/v2/dev/contribute_to_paddle_cn.md +++ b/doc/v2/dev/contribute_to_paddle_cn.md @@ -104,7 +104,7 @@ no changes added to commit (use "git add" and/or "git commit -a") ➜ docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest" ``` -关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。 +关于构建和测试的更多信息，请参见[使用Docker安装运行](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/v2/build_and_install/docker_install_cn.rst)。 ## 提交（commit） diff --git a/doc/v2/dev/new_layer_cn.rst b/doc/v2/dev/new_layer_cn.rst index 3115654b2bd87995fa63bb7828fd1b3039aea8cc..e5a14346123d342de0b67757cbbce654bd4180dc 100644 --- a/doc/v2/dev/new_layer_cn.rst +++ b/doc/v2/dev/new_layer_cn.rst @@ -58,7 +58,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。 实现C++类 =================== -一个网络层的C++类需要实现初始化，前向和后向。全连接层的实现位于:code:`paddle/gserver/layers/FullyConnectedLayer.h`及:code:`paddle/gserver/layers/FullyConnectedLayer.cpp`。这里我们展示一份简化过的代码。 +一个网络层的C++类需要实现初始化，前向和后向。全连接层的实现位于:code:`paddle/legacy/gserver/layers/FullyConnectedLayer.h`及:code:`paddle/legacy/gserver/layers/FullyConnectedLayer.cpp`。这里我们展示一份简化过的代码。 这个类需要继承 :code:`paddle::Layer` 这个基类，并且需要重写基类中的以下几个虚函数： @@ -153,7 +153,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。 - 每个层在其 :code:`forward` 函数的开头必须调用 :code:`Layer::forward(passType);` 。 - 之后使用 :code:`reserveOutput(batchSize, size);` 为输出分配内存。由于我们支持训练数据有不同的批次大小，所以这一步是必要的。 :code:`reserveOutput` 会相应地改变输出的尺寸。为了保证效率，如果需要扩大矩阵，我们会重新分配内存；如果需要缩减矩阵，我们会继续使用现有的内存块。 -- 之后使用矩阵运算函数来计算 :math:`\sum_i W_i x + b`。:code:`getInput(i).value` 返回第i个输入矩阵。每个输入都是一个 :math:`batchSize \times dim` 的矩阵，每行表示一个批次中的单个输入。对于我们支持的全部矩阵操作，请参考 :code:`paddle/math/Matrix.h`和:code:`paddle/math/BaseMatrix.h` 。 +- 之后使用矩阵运算函数来计算 :math:`\sum_i W_i x + b`。:code:`getInput(i).value` 返回第i个输入矩阵。每个输入都是一个 :math:`batchSize \times dim` 的矩阵，每行表示一个批次中的单个输入。对于我们支持的全部矩阵操作，请参考 :code:`paddle/legacy/math/Matrix.h`和:code:`paddle/legacy/math/BaseMatrix.h` 。 - 最终，使用 :code:`forwardActivation();` 进行激活操作。这会自动进行网络配置中声明的激活操作。 @@ -262,7 +262,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。 REGISTER_LAYER(fc, FullyConnectedLayer); } -若 :code:`cpp` 被放在 :code:`paddle/gserver/layers` 目录下，其会自动被加入编译列表。 +若 :code:`cpp` 被放在 :code:`paddle/legacy/gserver/layers` 目录下，其会自动被加入编译列表。 写梯度检查单元测试 @@ -270,7 +270,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。 写梯度检查单元测试是一个验证新实现的层是否正确的相对简单的办法。梯度检查单元测试通过有限差分法来验证一个层的梯度。首先对输入做一个小的扰动 :math:`\Delta x` ，然后观察到输出的变化为 :math:`\Delta y` ，那么，梯度就可以通过这个方程计算得到 :math:`\frac{\Delta y}{\Delta x }` 。之后，再用这个梯度去和 :code:`backward` 函数得到的梯度去对比，以保证梯度计算的正确性。需要注意的是梯度检查仅仅验证了梯度的计算，并不保证 :code:`forward` 和 :code:`backward` 函数的实现是正确的。你需要一些更复杂的单元测试来保证你实现的网络层是正确的。 -所有网络层的梯度检查单测都位于 :code:`paddle/gserver/tests/test_LayerGrad.cpp` 。我们建议你在写新网络层时把测试代码放入新的文件中。下面列出了全连接层的梯度检查单元测试。它包含以下几步： +所有网络层的梯度检查单测都位于 :code:`paddle/legacy/gserver/tests/test_LayerGrad.cpp` 。我们建议你在写新网络层时把测试代码放入新的文件中。下面列出了全连接层的梯度检查单元测试。它包含以下几步： + 生成网络层配置。网络层配置包含以下几项： - 偏置参数的大小。（例子中是4096） @@ -322,7 +322,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。 } } -如果你要为了测试而增加新的文件，例如 :code:`paddle/gserver/tests/testFCGrad.cpp` ，你需要把该文件加入 :code:`paddle/gserver/tests/CMakeLists.txt` 中。下面给出了一个例子。当你执行命令 :code:`make tests` 时，所有的单测都会被执行一次。注意，有些层可能需要高精度来保证梯度检查单测正确执行。你需要在配置cmake时将 :code:`WITH_DOUBLE` 设置为 `ON` 。 +如果你要为了测试而增加新的文件，例如 :code:`paddle/legacy/gserver/tests/testFCGrad.cpp` ，你需要把该文件加入 :code:`paddle/legacy/gserver/tests/CMakeLists.txt` 中。下面给出了一个例子。当你执行命令 :code:`make tests` 时，所有的单测都会被执行一次。注意，有些层可能需要高精度来保证梯度检查单测正确执行。你需要在配置cmake时将 :code:`WITH_DOUBLE` 设置为 `ON` 。 .. code-block:: bash diff --git a/doc/v2/dev/new_layer_en.rst b/doc/v2/dev/new_layer_en.rst index b05bb45f11eb253dfb87d6283c29ec6689394d22..ad723738801908a5f48343574c204bdbfc97ee08 100644 --- a/doc/v2/dev/new_layer_en.rst +++ b/doc/v2/dev/new_layer_en.rst @@ -58,7 +58,7 @@ Finally we can use chain rule to calculate :math:`\frac{\partial z}{\partial x}` Implement C++ Class =================== -The C++ class of the layer implements the initialization, forward, and backward part of the layer. The fully connected layer is at :code:`paddle/gserver/layers/FullyConnectedLayer.h` and :code:`paddle/gserver/layers/FullyConnectedLayer.cpp`. We list simplified version of the code below. +The C++ class of the layer implements the initialization, forward, and backward part of the layer. The fully connected layer is at :code:`paddle/legacy/gserver/layers/FullyConnectedLayer.h` and :code:`paddle/legacy/gserver/layers/FullyConnectedLayer.cpp`. We list simplified version of the code below. It needs to derive the base class :code:`paddle::Layer`, and it needs to override the following functions: @@ -154,7 +154,7 @@ The implementation of the forward part has the following steps. - Every layer must call :code:`Layer::forward(passType);` at the beginning of its :code:`forward` function. - Then it allocates memory for the output using :code:`reserveOutput(batchSize, size);`. This step is necessary because we support the batches to have different batch sizes. :code:`reserveOutput` will change the size of the output accordingly. For the sake of efficiency, we will allocate new memory if we want to expand the matrix, but we will reuse the existing memory block if we want to shrink the matrix. -- Then it computes :math:`\sum_i W_i x + b` using Matrix operations. :code:`getInput(i).value` retrieve the matrix of the i-th input. Each input is a :math:`batchSize \times dim` matrix, where each row represents an single input in a batch. For a complete lists of supported matrix operations, please refer to :code:`paddle/math/Matrix.h` and :code:`paddle/math/BaseMatrix.h`. +- Then it computes :math:`\sum_i W_i x + b` using Matrix operations. :code:`getInput(i).value` retrieve the matrix of the i-th input. Each input is a :math:`batchSize \times dim` matrix, where each row represents an single input in a batch. For a complete lists of supported matrix operations, please refer to :code:`paddle/legacy/math/Matrix.h` and :code:`paddle/legacy/math/BaseMatrix.h`. - Finally it applies the activation function using :code:`forwardActivation();`. It will automatically applies the corresponding activation function specifies in the network configuration. @@ -263,7 +263,7 @@ Finally, you can use :code:`REGISTER_LAYER(fc, FullyConnectedLayer);` to registe REGISTER_LAYER(fc, FullyConnectedLayer); } -If the :code:`cpp` file is put into :code:`paddle/gserver/layers`, it will be automatically added to the compilation list. +If the :code:`cpp` file is put into :code:`paddle/legacy/gserver/layers`, it will be automatically added to the compilation list. Write Gradient Check Unit Test @@ -271,7 +271,7 @@ Write Gradient Check Unit Test An easy way to verify the correctness of new layer's implementation is to write a gradient check unit test. Gradient check unit test utilizes finite difference method to verify the gradient of a layer. It modifies the input with a small perturbation :math:`\Delta x` and observes the changes of output :math:`\Delta y`, the gradient can be computed as :math:`\frac{\Delta y}{\Delta x }`. This gradient can be compared with the gradient computed by the :code:`backward` function of the layer to ensure the correctness of the gradient computation. Notice that the gradient check only tests the correctness of the gradient computation, it does not necessarily guarantee the correctness of the implementation of the :code:`forward` and :code:`backward` function. You need to write more sophisticated unit tests to make sure your layer is implemented correctly. -All the gradient check unit tests are located in :code:`paddle/gserver/tests/test_LayerGrad.cpp`. You are recommended to put your test into a new test file if you are planning to write a new layer. The gradient test of the gradient check unit test of the fully connected layer is listed below. It has the following steps. +All the gradient check unit tests are located in :code:`paddle/legacy/gserver/tests/test_LayerGrad.cpp`. You are recommended to put your test into a new test file if you are planning to write a new layer. The gradient test of the gradient check unit test of the fully connected layer is listed below. It has the following steps. + Create layer configuration. A layer configuration can include the following attributes: - size of the bias parameter. (4096 in our example) @@ -323,7 +323,7 @@ All the gradient check unit tests are located in :code:`paddle/gserver/tests/tes } } -If you are creating a new file for the test, such as :code:`paddle/gserver/tests/testFCGrad.cpp`, you need to add the file to :code:`paddle/gserver/tests/CMakeLists.txt`. An example is given below. All the unit tests will run when you execute the command :code:`make tests`. Notice that some layers might need high accuracy for the gradient check unit tests to work well. You need to configure :code:`WITH_DOUBLE` to `ON` when configuring cmake. +If you are creating a new file for the test, such as :code:`paddle/legacy/gserver/tests/testFCGrad.cpp`, you need to add the file to :code:`paddle/legacy/gserver/tests/CMakeLists.txt`. An example is given below. All the unit tests will run when you execute the command :code:`make tests`. Notice that some layers might need high accuracy for the gradient check unit tests to work well. You need to configure :code:`WITH_DOUBLE` to `ON` when configuring cmake. .. code-block:: bash @@ -339,7 +339,7 @@ If you are creating a new file for the test, such as :code:`paddle/gserver/tests Implement Python Wrapper ======================== -Implementing Python wrapper allows us to use the added layer in configuration files. All the Python wrappers are in file :code:`python/paddle/trainer/config_parser.py`. An example of the Python wrapper for fully connected layer is listed below. It has the following steps: +Implementing Python wrapper allows us to use the added layer in configuration files. All the Python wrappers are in file :code:`python/paddle/legacy/trainer/config_parser.py`. An example of the Python wrapper for fully connected layer is listed below. It has the following steps: - Use :code:`@config_layer('fc')` at the decorator for all the Python wrapper class. :code:`fc` is the identifier of the layer. - Implements :code:`__init__` constructor function. diff --git a/doc/v2/faq/build_and_install/index_cn.rst b/doc/v2/faq/build_and_install/index_cn.rst index f292684fb5fe2df06db5239e7f43fdfa1dd2f2bd..0d644777287aea0a572adb6fa40f498f9c147af7 100644 --- a/doc/v2/faq/build_and_install/index_cn.rst +++ b/doc/v2/faq/build_and_install/index_cn.rst @@ -213,3 +213,12 @@ virtualenv本身也是Python的一个包，可以用pip进行安装： 保存并关闭文件。 这样，每次打开终端时就会自动启动名为‘paddle’的Python环境了。 + +10. 通过pip安装的PaddlePaddle在 :code:`import paddle.fluid` 报找不到 :code:`libmkldnn.so` 或 :code:`libmklml_intel.so` +------------------------------------------------------------------------------------------ +出现这种问题的原因是在导入 :code:`paddle.fluid` 时需要加载 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`， +但是系统没有找到该文件。一般通过pip安装PaddlePaddle时会将 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` +拷贝到 :code:`/usr/local/lib` 路径下，所以解决办法是将该路径加到 :code:`LD_LIBRARY_PATH` 环境变量下， +即： :code:`export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH` 。 + +**注意**：如果是在虚拟环境中安装PaddlePaddle， :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` 可能不在 :code:`/usr/local/lib` 路径下。 \ No newline at end of file diff --git a/doc/v2/faq/parameter/index_cn.rst b/doc/v2/faq/parameter/index_cn.rst index 1fa4b3e1311d2007ccba98fde9ff94300ea42c16..987e8cf088be4ee8daa7c28fdc855506cbfd31c7 100644 --- a/doc/v2/faq/parameter/index_cn.rst +++ b/doc/v2/faq/parameter/index_cn.rst @@ -196,6 +196,6 @@ PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数 obj="process", args={"src_dict_path": src_dict_path}) -完整源码可参考 `sequence_recurrent `_ 示例。 +完整源码可参考 `sequence_recurrent `_ 示例。 diff --git a/doc/v2/faq/parameter/index_en.rst b/doc/v2/faq/parameter/index_en.rst index 61c7845af7e531013a06125f7c35b59081dafb42..9edb8dd620f972d019db9c0063cefce616de0ebd 100644 --- a/doc/v2/faq/parameter/index_en.rst +++ b/doc/v2/faq/parameter/index_en.rst @@ -1,5 +1,198 @@ -################# -Parameter Setting -################# +################## +Parameter Settings +################## -TBD +.. contents:: + +1. How to Choose the Learning Rate of SGD Algorithm +-------------------------- + +An important issue when training with :code:`sgd/async_sgd` is to choose the correct value for :code:`learning_rate`. If it is too large, the training may not converge. If too small, the convergence may be slow, resulting in a long training time. + +Usually, we start with a relatively large learning rate. If the training does not converge, then we need to reduce the learning rate continuously by a factor of 10 until the training converges. We examine the convergence of the training by estimating the minimum cost at a constant output of the model. + +If the cost of the training process is significantly higher than the cost of the output, then we judge that the training does not converge. For example, if we have a three-class problem and use multi-class-cross-entropy as the cost, the ratio of 0, 1, and 2 in the data will be :code:`0.2, 0.5, 0.3`. The minimum cost thus will be :code:`-(0.2*log(0.2)+0.5*log(0.5)+0.3*log(0.3))=1.03`. If the cost is greater than this number after training a pass (or even before), then the training may not be converged and the learning rate should be reduced. + +2. How to Implement Learning Rate Annealing +------------------------------------------------ + +We use the Adam algorithm as an example. Set the parameters of :code:`learning_rate_schedule` in the corresponding optimization algorithm as follows: + +.. code-block:: python + +    Optimizer = paddle.optimizer.Adam( +        Learning_rate=1e-3, +        Learning_rate_decay_a=0.5, +        Learning_rate_decay_b=0.75, +        Learning_rate_schedule="poly",) + +PaddlePaddle currently supports 8 learning rate schedules. The 8 learning rate schedules and their corresponding learning rates are calculated as follows: + +* "constant" +   +  Lr = learning_rate + +* "poly" + +  Lr = learning_rate * pow(1 + learning_rate_decay_a * num_samples_processed, -learning_rate_decay_b) + +  Variable :code:`num_samples_processed` is the number of trained samples. + +* "caffe_poly" + +  Lr = learning_rate * pow(1.0 - num_samples_processed / learning_rate_decay_a, learning_rate_decay_b) + +* "exp" + +  Lr = learning_rate * pow(learning_rate_decay_a, num_samples_processed / learning_rate_decay_b) + +* "discexp" + +  Lr = learning_rate * pow(learning_rate_decay_a, floor(num_samples_processed / learning_rate_decay_b)) + +* "linear" + +  Lr = max(learning_rate - learning_rate_decay_a * num_samples_processed, learning_rate_decay_b) + +* "manual" + +  This is a learning rate annealing method that is segmented by the number of trained samples. When using this learning rate schedule, we modify the learning rate attenuation factor piecewise function by changing the parameter :code:`learning_rate_args`. The current learning rate is the product of :code:`learning_rate` and the current attenuation factor. Take the Adam algorithm as an example: + +  .. code-block:: python + +      Optimizer = paddle.optimizer.Adam( +          Learning_rate=1e-3, +          Learning_rate_schedule="manual", +          Learning_rate_args="1000:1.0,2000:0.9,3000:0.8",) + +  In this example, when the number of trained samples is less than or equal to 1000, the learning rate is: code:`1e-3*1.0`; when the number of trained samples is greater than 1000 or less than or equal to 2000, the learning rate is:code:`1e- 3 * 0.9`; when the number of trained samples is greater than 2,000, the learning rate is: code:`1e-3*0.8`. + +* "pass_manual" + +  This is a learning rate annealing method that piecewisely pick values according to the number of trained passes. When using this learning rate schedule, we set the learning rate attenuation factor piecewise function by the parameter :code:`learning_rate_args`. The current learning rate is the product of :code:`learning_rate` and the current attenuation factor. Take the Adam algorithm as an example: + +  .. code-block:: python + +      Optimizer = paddle.optimizer.Adam( +          Learning_rate=1e-3, +          Learning_rate_schedule="pass_manual", +          Learning_rate_args="1:1.0,2:0.9,3:0.8",) + +  In this example, when the number of trained passes is less than or equal to 1, the learning rate is :code:`1e-3*1.0`; when the number of trained passes is greater than 1 or less than 2, the learning rate is :code:`1e- 3 * 0.9`; when the number of trained passes is greater than 2, the learning rate is :code:`1e-3*0.8`. + +3. How to Initialize Parameters +----------------- + +By default, PaddlePaddle initializes parameters with an average of 0 and a standard deviation of :math:`\frac{1}{\sqrt{d}}`, where :math:`d` is the width of the parameter matrix. This initialization method does not produce bad results under normal circumstances. If users want to customize the initialization method, PaddlePaddle provides two ways to initialize the parameters: + +* Gaussian distribution. Set :code:`param_attr` to :code:`param_attr=ParamAttr(initial_mean=0.0, initial_std=1.0)` +* Uniform distribution. Set :code:`param_attr` to :code:`param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0)` + +For example, to set a full connection layer parameter initialization mode and bias initialization mode, you can use the following code: + +.. code-block:: python + +    Hidden = fc_layer(input=ipt, param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0), +                      Bias_attr=ParamAttr(initial_mean=1.0, initial_std=0.0)) + +The above code initializes the bias to 1.0 and initializes the parameters to a uniform distribution of :code:`[1.0, -1.0]`. + +4. How to Share Parameters +--------------- + +PaddlePaddle's parameters use :code:`name` as the ID. Parameters with the same name will share parameters//. We can set the name of the parameters using :code:`ParamAttr(name="YOUR_PARAM_NAME")`. More conveniently, we can make the parameters to be shared use the same :code:`ParamAttr` object. + +A simple fully connected network has its configuration of parameter sharing as follows \: + +.. literalinclude:: ../../python/paddle/trainer_config_helpers/tests/configs/shared_fc.py + +Here :code:`hidden_a` and :code:`hidden_b` have the same parameter and bias. The two input of the softmax layer also use the same parameter :code:`softmax_param`. + +5. How to Load Pre-training Parameters +------------------------ +* For layers that load pre-training parameters, set :code:`is_static = True` so that the parameters of that layer remain unchanged during the training process. Take the embedding layer as an example, the code is as follows: + +.. code-block:: python + +    Emb_para = paddle.attr.Param(name='emb', is_static=True) +    Paddle.layer.embedding(size=word_dim, input=x, param_attr=emb_para) + + +* Load pre-training parameters from the model file into :code:`numpy.array`. After creating the parameters, load the pre-training parameters using :code:`parameters.set()`. The first 16 bytes of the model parameter file saved by PaddlePaddle is the header information. The user must loads : :code:`numpy.array` starting with the 17th byte. Take the embedding layer as an example, the code is as follows: + +.. code-block:: python + +    Def load_parameter(file_name, h, w): +        With open(file_name, 'rb') as f: +            F.read(16) # skip header. +            Return np.fromfile(f, dtype=np.float32).reshape(h, w) + +    Parameters = paddle.parameters.create(my_cost) +    Parameters.set('emb', load_parameter(emb_param_file, 30000, 256)) + +6. Format of the Stored Parameter and How to Convert the File to Plain Text +-------------------------------------------------- + +The model parameter file saved by PaddlePaddle consists of 16 bytes of header information and network parameters. In the header information, the first four bytes show PaddlePaddle's version information. The user should fill in with 0s. The next four bytes represent the number of bytes occupied by each parameter. If the saved network parameter is a float type, the number is four; if it is a double, the number is eight. The third group of four bytes represents the total number of saved parameters. + +When restoring the model parameters saved by PaddlePaddle back to plain text, we use the corresponding data type :code:`numpy.array` to load specific network parameters. At this time, you can skip the header information of the PaddlePaddle model parameter file. If not specified to compile with a precision for double in PaddlePaddle, then the parameter file will be caiculated with a precision for float, and the argument will be stored as a float. In this case, when using :code:`numpy.array`, generally we set :code:`dtype=float32`. An example is as follows: + +.. code-block:: python + +    Def read_parameter(fname, width): +        s = open(fname).read() +        # skip header +        Vec = np.fromstring(s[16:], dtype=np.float32) +        # width is the size of the corresponding layer +        Np.savetxt(fname + ".csv", vec.reshape(width, -1), +                Fmt="%.6f", delimiter=",") + + +When the plaintext parameters are converted into PaddlePaddle loadable model parameters, the header information is constructed first, then the network parameters are written. The following code converts the randomly generated matrix into model parameters that can be loaded by PaddlePaddle: + +.. code-block:: python + +    Def gen_rand_param(param_file, width, height, need_trans): +        Np.random.seed() +        Header = struct.pack("iil", 0, 4, height * width) +        Param = np.float32(np.random.rand(height, width)) +        With open(param_file, "w") as fparam: +            Fparam.write(header + param.tostring()) + +7. A Protocol Message Rejected Because of its Large Size +-------------------------------------------------- ---------- + +If you are training NLP related models, and the following error occurs: + +.. code-block:: bash + +    [libprotobuf ERROR google/protobuf/io/coded_stream.cc:171] A protocol message was rejected because it was too big (more than 67108864 bytes). To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit( ) in google/protobuf/io/coded_stream.h. +    F1205 14:59:50.295174 14703 TrainerConfigHelper.cpp:59] Check failed: m->conf.ParseFromString(configProtoStr) + +The possible reason is that one of the args passed to the dataprovider is too large, which is usually caused by directly passing a large dictionary. A wrongly defineed `_py_data_sources2` is similar to: + +.. code-block:: python + +     Src_dict = dict() +     For line_count, line in enumerate(open(src_dict_path, "r")): +        Src_dict[line.strip()] = line_count + +     Define_py_data_sources2( +        Train_list, +        Test_list, +        Module="dataprovider", +        Obj="process", +        Args={"src_dict": src_dict}) + +The solution is to pass the address of the dictionary as args to the dataprovider, and then load the dictionary according to the address in the dataprovider. Change `_py_data_sources2` to: + +.. code-block:: python + +     Define_py_data_sources2( +        Train_list, +        Test_list, +        Module="dataprovider", +        Obj="process", +        Args={"src_dict_path": src_dict_path}) + +The full source code can be found in the `sequence_recurrent `_ example. diff --git a/doc/v2/howto/capi/compile_paddle_lib_cn.md b/doc/v2/howto/capi/compile_paddle_lib_cn.md index e223fd33a8420abcdfdad53d1cfc5ed160a1b37e..8878ee9d85064ba27708ed92790aa9b83ba316e5 100644 --- a/doc/v2/howto/capi/compile_paddle_lib_cn.md +++ b/doc/v2/howto/capi/compile_paddle_lib_cn.md @@ -18,24 +18,29 @@ cpu_avx_openblas -暂无 +paddle.tgz cpu_noavx_openblas -paddle.tgz +paddle.tgz cuda7.5_cudnn5_avx_mkl -paddle.tgz +paddle.tgz cuda8.0_cudnn5_avx_mkl -paddle.tgz +paddle.tgz cuda8.0_cudnn7_avx_mkl -paddle.tgz - +paddle.tgz + + +cuda9.0_cudnn7_avx_mkl +paddle.tgz + + ### 从源码编译 diff --git a/doc/v2/howto/capi/compile_paddle_lib_en.md b/doc/v2/howto/capi/compile_paddle_lib_en.md index 6212a3081116d988630706e83d2349dd200b73ab..70a6edef27e75af6b38d7d4824c928eba0d29b9a 100644 --- a/doc/v2/howto/capi/compile_paddle_lib_en.md +++ b/doc/v2/howto/capi/compile_paddle_lib_en.md @@ -13,28 +13,33 @@ cpu_avx_mkl -paddle.tgz +paddle.tgz cpu_avx_openblas -- +paddle.tgz cpu_noavx_openblas -paddle.tgz +paddle.tgz cuda7.5_cudnn5_avx_mkl -paddle.tgz +paddle.tgz cuda8.0_cudnn5_avx_mkl -paddle.tgz +paddle.tgz cuda8.0_cudnn7_avx_mkl -paddle.tgz - +paddle.tgz + + +cuda9.0_cudnn7_avx_mkl +paddle.tgz + + ### From source diff --git a/doc/v2/howto/capi/workflow_of_capi_cn.md b/doc/v2/howto/capi/workflow_of_capi_cn.md index 3acdbae28e9b35f8a9104a89c9a5799f8c892334..db1568a2afbea3cca0d4e1fe053ba9536a60ab3d 100644 --- a/doc/v2/howto/capi/workflow_of_capi_cn.md +++ b/doc/v2/howto/capi/workflow_of_capi_cn.md @@ -28,9 +28,9 @@ ### 准备预测模型 -准备预测模型部分，我们以手写数字识别任务为例进行介绍。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression)，网络接受一幅图片作为输入，将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense) 中的相关脚本。 +准备预测模型部分，我们以手写数字识别任务为例进行介绍。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression)，网络接受一幅图片作为输入，将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense) 中的相关脚本。 -调用C-API开发预测程序需要一个训练好的模型，运行[MNIST手写数字识别目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)下的[mnist_v2.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本，在终端执行`python mnist_v2.py`，会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。 +调用C-API开发预测程序需要一个训练好的模型，运行[MNIST手写数字识别目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense)下的[mnist_v2.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/examples/model_inference/dense/mnist_v2.py)脚本，在终端执行`python mnist_v2.py`，会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。 下面，我们将训练结束后存储下来的模型转换成预测模型。 @@ -48,7 +48,7 @@ dump_v2_config(predict, "trainer_config.bin", True) ``` - 对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程，可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化，结果会写入当前运行目录下的`trainer_config.bin`文件中。 + 对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense)这个示例，[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程，可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化，结果会写入当前运行目录下的`trainer_config.bin`文件中。 使用这种方式，需要**在运行时将神经网络的多个可学习参数放在同一个目录中**，C-API可以通过分别指定序列化后的网络结构文件和参数目录来加载训练好的模型。 @@ -68,7 +68,7 @@ merge_v2_model(net, param_file, output_file) ``` - 对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。 + 对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。 #### 注意事项 1. 为使用C-API，在调用`dump_v2_config`序列化神经网络结构时，参数`binary`必须指定为`True`。 @@ -77,10 +77,10 @@ ### 编写预测代码 -预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。 +预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。 #### step 1. 初始化PaddlePaddle运行环境 -第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/main.h#L27) 初始化PaddlePaddle运行环境，该接口接受两个参数：参数的个数和参数列表。 +第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/main.h#L27) 初始化PaddlePaddle运行环境，该接口接受两个参数：参数的个数和参数列表。 #### step2. 加载模型 @@ -88,8 +88,8 @@ 概念上，在 PaddlePaddle 内部，一个GradientMachine类的对象管理着一组计算层（PaddlePaddle Layers）来完成前向和反向计算，并处理与之相关的所有细节。在调用C-API预测时，只需进行前向计算而无需调用反向计算。这篇文档之后部分会使用`gradient machine`来特指调用PaddlePaddle C-API创建的GradientMachine类的对象。每一个 `gradient machine` 都会管理维护一份训练好的模型，下面是C-API提供的，两种常用的模型加载方式： -1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L61)接口，从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型； -1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L88)接口，与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时，通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/multi_thread/main.c)。 +1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/gradient_machine.h#L61)接口，从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型； +1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/gradient_machine.h#L88)接口，与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时，通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/examples/model_inference/multi_thread/main.c)。 - 注意事项 @@ -117,7 +117,7 @@ C-API支持的所有输入数据类型和他们的组织方式，请参考“输 #### step 4. 前向计算 -完成上述准备之后，通过调用 [`paddle_gradient_machine_forward`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L73) 接口完成神经网络的前向计算。 +完成上述准备之后，通过调用 [`paddle_gradient_machine_forward`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/gradient_machine.h#L73) 接口完成神经网络的前向计算。 #### step 5. 清理 diff --git a/doc/v2/howto/optimization/gpu_profiling_cn.rst b/doc/v2/howto/optimization/gpu_profiling_cn.rst index 25bcaccb6975bc21fba2e8c5843da15c69948d72..f2396716bddd4810fa77c738d41f5482aa6d6055 100644 --- a/doc/v2/howto/optimization/gpu_profiling_cn.rst +++ b/doc/v2/howto/optimization/gpu_profiling_cn.rst @@ -50,12 +50,12 @@ GPU则还需要高并行性，才能发挥其全部能力。这正是它们速 **nvprof** 是Nvidia性能分析工具， **nvvp** 则是带GUI的Nvidia可视化性能分析工具。 在这个教程中，我们主要会介绍nvprof和nvvp。 -:code:`test_GpuProfiler` from :code:`paddle/math/tests` directory will be used to evaluate +:code:`test_GpuProfiler` from :code:`paddle/legacy/math/tests` directory will be used to evaluate above profilers. -:code:`paddle/math/test` 目录中的 :code:`test_GpuProfiler` 就是用于展示上述分析工具的用法。 +:code:`paddle/legacy/math/test` 目录中的 :code:`test_GpuProfiler` 就是用于展示上述分析工具的用法。 -.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp +.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp :language: c++ :lines: 137-151 :linenos: @@ -83,7 +83,7 @@ program crashes when CPU version of PaddlePaddle invokes them. 1. 加入 :code:`REGISTER_TIMER_INFO` 和 :code:`printAllStatus` 函数（如高亮部分）。 - .. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp + .. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp :language: c++ :lines: 137-151 :emphasize-lines: 8-12,14 @@ -101,8 +101,8 @@ program crashes when CPU version of PaddlePaddle invokes them. .. code-block:: bash :emphasize-lines: 1,12-15 - > ./paddle/math/tests/test_GpuProfiler - I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/math/tests/test_GpuProfiler + > ./paddle/legacy/math/tests/test_GpuProfiler + I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/legacy/math/tests/test_GpuProfiler I1117 11:13:42.845065 2522362816 Util.cpp:130] Calling runInitFunctions I1117 11:13:42.845208 2522362816 Util.cpp:143] Call runInitFunctions done. [==========] Running 1 test from 1 test case. @@ -130,7 +130,7 @@ nvprof 工具 1. 将 :code:`REGISTER_GPU_PROFILER` 函数加到代码中（参考强调部分）。 - .. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp + .. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp :language: c++ :lines: 137-151 :emphasize-lines: 6-7 @@ -147,13 +147,13 @@ nvprof 工具 .. code-block:: bash - nvprof ./paddle/math/tests/test_GpuProfiler + nvprof ./paddle/legacy/math/tests/test_GpuProfiler 然后，您就能获得如下的分析结果： .. code-block:: bash - ==78544== Profiling application: ./paddle/math/tests/test_GpuProfiler + ==78544== Profiling application: ./paddle/legacy/math/tests/test_GpuProfiler ==78544== Profiling result: Time(%) Time Calls Avg Min Max Name 27.60% 9.6305ms 5 1.9261ms 3.4560us 6.4035ms [CUDA memcpy HtoD] diff --git a/doc/v2/howto/optimization/gpu_profiling_en.rst b/doc/v2/howto/optimization/gpu_profiling_en.rst index 50adb7da24906515cb5977db565e9f8a76599fef..6e439be9bba8935cdd65f1c131cfd3725530ec0e 100644 --- a/doc/v2/howto/optimization/gpu_profiling_en.rst +++ b/doc/v2/howto/optimization/gpu_profiling_en.rst @@ -51,10 +51,10 @@ For general GPU profiling, a bunch of tools are provided from both NVIDIA and th **nvprof** is Nvidia profiler and **nvvp** is (GUI based) Nvidia visual profiler. In this tutorial, we will focus on nvprof and nvvp. -:code:`test_GpuProfiler` from :code:`paddle/math/tests` directory will be used to evaluate +:code:`test_GpuProfiler` from :code:`paddle/legacy/math/tests` directory will be used to evaluate above profilers. -.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp +.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp :language: c++ :lines: 137-151 :linenos: @@ -80,7 +80,7 @@ As a simple example, consider the following: 1. Add :code:`REGISTER_TIMER_INFO` and :code:`printAllStatus` functions (see the emphasize-lines). - .. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp + .. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp :language: c++ :lines: 137-151 :emphasize-lines: 8-12,14 @@ -98,8 +98,8 @@ As a simple example, consider the following: .. code-block:: bash :emphasize-lines: 1,12-15 - > ./paddle/math/tests/test_GpuProfiler - I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/math/tests/test_GpuProfiler + > ./paddle/legacy/math/tests/test_GpuProfiler + I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/legacy/math/tests/test_GpuProfiler I1117 11:13:42.845065 2522362816 Util.cpp:130] Calling runInitFunctions I1117 11:13:42.845208 2522362816 Util.cpp:143] Call runInitFunctions done. [==========] Running 1 test from 1 test case. @@ -127,7 +127,7 @@ To use this command line profiler **nvprof**, you can simply issue the following 1. Add :code:`REGISTER_GPU_PROFILER` function (see the emphasize-lines). - .. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp + .. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp :language: c++ :lines: 137-151 :emphasize-lines: 6-7 @@ -144,13 +144,13 @@ To use this command line profiler **nvprof**, you can simply issue the following .. code-block:: bash - nvprof ./paddle/math/tests/test_GpuProfiler + nvprof ./paddle/legacy/math/tests/test_GpuProfiler Then, you can get the following profiling result: .. code-block:: bash - ==78544== Profiling application: ./paddle/math/tests/test_GpuProfiler + ==78544== Profiling application: ./paddle/legacy/math/tests/test_GpuProfiler ==78544== Profiling result: Time(%) Time Calls Avg Min Max Name 27.60% 9.6305ms 5 1.9261ms 3.4560us 6.4035ms [CUDA memcpy HtoD] diff --git a/doc/v2/howto/rnn/hierarchical_layer_en.rst b/doc/v2/howto/rnn/hierarchical_layer_en.rst index 236f58a160c7f77c28e4b1216b83b3d3cdaaa459..fb668f1babb47f49b2dab6d2411565e99599d8b0 100644 --- a/doc/v2/howto/rnn/hierarchical_layer_en.rst +++ b/doc/v2/howto/rnn/hierarchical_layer_en.rst @@ -1,4 +1,89 @@ -Layers supporting hierarchical sequence as input -================================================ - -TBD +########################### +Layers that Support Hierarchical Sequences as Input +########################### +  +.. contents:: +  +Overview +==== +  +A sequence is a common data type in natural language processing tasks. An independent word can be regarded as a non-sequential input or a 0-level sequence. A sentence made up of words is a single-level sequence; a number of sentences make up a paragraph, which is a double-level sequence. +  +A double-level sequence is a nested sequence where each element is a single-level sequence. This is a very flexible way of organizing data that helps us construct some complex input information. +  +We can define non-sequences, single-level sequences, and double-level sequences at the following levels. +  ++ 0-level sequence: an independent element. Its type can be any input data type supported by PaddlePaddle; ++ Single-level sequence: multiple elements arranged in a row; each element is a 0-level sequence. The order of elements is an important input information; ++ Double-level sequence: multiple elements arranged in a row; each element is a single-layer sequence called a subseq of a double-level sequence, and each element of the subseq is a 0-level sequence. +  +In PaddlePaddle, the following layers accept double-layer sequences as input and perform corresponding calculations. +  +`pooling` +======== +  +The use of pooling is as follows: +  +.. code-block:: bash +  +        Seq_pool = pooling(input=layer, +                           Pooling_type=pooling.Max(), +                           Agg_level=AggregateLevel.TO_SEQUENCE) +         +- `pooling_type` currently supports two types: pooling.Max() and pooling.Avg(). +  +- When ʻagg_level=AggregateLevel.TO_NO_SEQUENCE` (default): +  +  - Effect: a double-level sequence input will be converted into a 0-level sequence, and a single-level sequence will be converted into a 0-level sequence  +  - Input: a double-level sequence or a single-level sequence +  - Output: a 0-level sequence which is the average (or maximum) of the entire input sequence (single or double) +  +- When ʻagg_level=AggregateLevel.TO_SEQUENCE`: +  +  - Effect: a double-level sequence will be transformed into a single-level sequence +  - Input: a double-level sequence +  - Output: a single-level sequence where each element of the sequence is the average (or maximum) value of each subseq element of the original double-level sequence. +  +`last_seq` and `first_seq` +===================== +  +An example of using `last_seq` is as follows (usage of `first_seq` is similar). +  +.. code-block:: bash +  +        Last = last_seq(input=layer, +                        Agg_level=AggregateLevel.TO_SEQUENCE) +         +- When ʻagg_level=AggregateLevel.TO_NO_SEQUENCE` (default): +  +  - Effect: a double-level sequence input will be converted into a 0-level sequence, and a single-level sequence will be converted into a 0-level sequence +  - Input: a double-level sequence or a single-level sequence +  - Output: a 0-level sequence, which is the last or the first element of the input sequence (double or single level). +  +- When ʻagg_level=AggregateLevel.TO_SEQUENCE`: +  - Effect: a double-level sequence will be transformed into a single-level sequence +  - Input: a double-level sequence +  - Output: a single-layer sequence in which each element is the last (or first) element of each subseq in a double-level sequence. +  +`expand` +====== +  +The use of expand is as follows. +  +.. code-block:: bash +  +        Ex = expand(input=layer1, +                    Expand_as=layer2, +                    Expand_level=ExpandLevel.FROM_NO_SEQUENCE) +         +- When `expand_level=ExpandLevel.FROM_NO_SEQUENCE` (default): +  +  - Effect: a 0-level sequence is extended to a single-level sequence or a double-level sequence +  - Input: layer1 must be a 0-level sequence to be extended; layer2 can be a single-level sequence or a double-level sequence that provides the extended length information +  - Output: a single-level sequence or a double-level sequence; the type of the output sequence and the number of elements contained in the sequence are the same as layer2. If the output is a single-level sequence, each element of the single-level sequence will be a copy of the layer1 element. If the output is a double-level sequence, each element in the double-level sequence will be a copy of the layer1 element +  +- When `expand_level=ExpandLevel.FROM_SEQUENCE`: +  +  - Effect: a single-level sequence is extended to a double-level sequence +  - Input: layer1 must be a single-level sequence to be extended; layer2 must be a double-level sequence providing extended length information +  - Output: a double-level sequence with the same number of elements as that of layer2. It is required that the number of elements in the single-level sequence be the same as the number of subseq in the double-level sequences. The i-th element of the single-level sequence (the 0-level sequence) is expanded into a single-level sequence that constitutes the i-th subseq of the output, the double-level sequence. diff --git a/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst b/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst index 67c7b774e9c476a3035037a421c84ebf17a31b09..9d6d417075485dceb1ee71f527b408aa6a6638ea 100644 --- a/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst +++ b/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst @@ -4,7 +4,7 @@ 单双层RNN API对比介绍 ##################### -本文以PaddlePaddle的双层RNN单元测试为示例，用多对效果完全相同的、分别使用单双层RNN作为网络配置的模型，来讲解如何使用双层RNN。本文中所有的例子，都只是介绍双层RNN的API接口，并不是使用双层RNN解决实际的问题。如果想要了解双层RNN在具体问题中的使用，请参考\ :ref:`algo_hrnn_demo`\ 。本文中示例所使用的单元测试文件是\ `test_RecurrentGradientMachine.cpp `_\ 。 +本文以PaddlePaddle的双层RNN单元测试为示例，用多对效果完全相同的、分别使用单双层RNN作为网络配置的模型，来讲解如何使用双层RNN。本文中所有的例子，都只是介绍双层RNN的API接口，并不是使用双层RNN解决实际的问题。如果想要了解双层RNN在具体问题中的使用，请参考\ :ref:`algo_hrnn_demo`\ 。本文中示例所使用的单元测试文件是\ `test_RecurrentGradientMachine.cpp `_\ 。 示例1：双层RNN，子序列间无Memory ================================ @@ -13,8 +13,8 @@ 在本示例中，单层RNN和双层RNN的网络配置，都是将每一句分好词后的句子，使用LSTM作为encoder，压缩成一个向量。区别是RNN使用两层序列模型，将多句话看成一个整体同时使用encoder压缩。二者语意上完全一致。这组语义相同的示例配置如下： -* 单层RNN\: `sequence_layer_group.conf `_ -* 双层RNN\: `sequence_nest_layer_group.conf `_ +* 单层RNN\: `sequence_layer_group.conf `_ +* 双层RNN\: `sequence_nest_layer_group.conf `_ 读取双层序列数据 @@ -24,18 +24,18 @@ - 本例中的原始数据一共有10个样本。每个样本由两部分组成，一个label（此处都为2）和一个已经分词后的句子。这个数据也被单层RNN网络直接使用。 -.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg :language: text - 双层序列数据一共有4个样本。 每个样本间用空行分开，整体数据和原始数据完全一样。但于双层序列的LSTM来说，第一个样本同时encode两条数据成两个向量。这四条数据同时处理的句子数量为\ :code:`[2, 3, 2, 3]`\ 。 -.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg.nest +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg.nest :language: text -其次，对于两种不同的输入数据类型，不同DataProvider对比如下(`sequenceGen.py `_)\： +其次，对于两种不同的输入数据类型，不同DataProvider对比如下(`sequenceGen.py `_)\： -.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py :language: python :lines: 21-39 :linenos: @@ -47,7 +47,7 @@ - words是原始数据中的每一句话，所对应的词表index数组。它是integer_value_sequence类型的，即整数数组。words即为这个数据中的单层时间序列。 - label是原始数据中对于每一句话的分类标签，它是integer_value类型的。 -.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py :language: python :lines: 42-71 :linenos: @@ -64,7 +64,7 @@ 首先，我们看一下单层RNN的配置。代码中9-15行(高亮部分)即为单层RNN序列的使用代码。这里使用了PaddlePaddle预定义好的RNN处理函数。在这个函数中，RNN对于每一个时间步通过了一个LSTM网络。 -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_layer_group.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_layer_group.conf :language: python :lines: 38-63 :linenos: @@ -85,7 +85,7 @@ * 至此，\ :code:`lstm_last`\ 便和单层RNN配置中的\ :code:`lstm_last`\ 具有相同的结果了。 -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_layer_group.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_layer_group.conf :language: python :lines: 38-64 :linenos: @@ -107,7 +107,7 @@ - 单层RNN：过了一个很简单的recurrent_group。每一个时间步，当前的输入y和上一个时间步的输出rnn_state做了一个全链接。 -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn.conf :language: python :lines: 36-48 @@ -116,7 +116,7 @@ - 内层inner_step的recurrent_group和单层序列的几乎一样。除了boot_layer=outer_mem，表示将外层的outer_mem作为内层memory的初始状态。外层outer_step中，outer_mem是一个子句的最后一个向量，即整个双层group是将前一个子句的最后一个向量，作为下一个子句memory的初始状态。 - 从输入数据上看，单双层序列的句子是一样的，只是双层序列将其又做了子序列划分。因此双层序列的配置中，必须将前一个子句的最后一个元素，作为boot_layer传给下一个子句的memory，才能保证和单层序列的配置中“每个时间步都用了上一个时间步的输出结果”一致。 -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn.conf :language: python :lines: 39-66 @@ -134,7 +134,7 @@ **输入不等长** 是指recurrent_group的多个输入序列，在每个时间步的子序列长度可以不相等。但序列输出时，需要指定与某一个输入的序列信息是一致的。使用\ :red:`targetInlink`\ 可以指定哪一个输入和输出序列信息一致，默认指定第一个输入。 -示例3的配置分别为\ `单层不等长RNN `_\ 和\ `双层不等长RNN `_\ 。 +示例3的配置分别为\ `单层不等长RNN `_\ 和\ `双层不等长RNN `_\ 。 示例3对于单层RNN和双层RNN数据完全相同。 @@ -152,14 +152,14 @@ * 单层RNN\: -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py :language: python :lines: 42-59 :linenos: * 双层RNN\ \: -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py :language: python :lines: 41-80 :linenos: diff --git a/doc/v2/howto/rnn/hrnn_rnn_api_compare_en.rst b/doc/v2/howto/rnn/hrnn_rnn_api_compare_en.rst index ae997f0805db5b01a34867c9e8b188c931721920..a4485f7b5edf21871444801230ab1ee191b1137b 100644 --- a/doc/v2/howto/rnn/hrnn_rnn_api_compare_en.rst +++ b/doc/v2/howto/rnn/hrnn_rnn_api_compare_en.rst @@ -4,7 +4,7 @@ API comparision between RNN and hierarchical RNN ##################### -This article takes PaddlePaddle's hierarchical RNN unit test as an example. We will use several examples to illestrate the usage of single-layer and hierarchical RNNs. Each example has two model configurations, one for single-layer, and the other for hierarchical RNN. Although the implementations are different, both the two model configurations' effects are the same. All of the examples in this article only describe the API interface of the hierarchical RNN, while we do not use this hierarchical RNN to solve practical problems. If you want to understand the use of hierarchical RNN in specific issues, please refer to \ :ref:`algo_hrnn_demo`\ 。The unit test file used in this article's example is \ `test_RecurrentGradientMachine.cpp `_\ 。 +This article takes PaddlePaddle's hierarchical RNN unit test as an example. We will use several examples to illestrate the usage of single-layer and hierarchical RNNs. Each example has two model configurations, one for single-layer, and the other for hierarchical RNN. Although the implementations are different, both the two model configurations' effects are the same. All of the examples in this article only describe the API interface of the hierarchical RNN, while we do not use this hierarchical RNN to solve practical problems. If you want to understand the use of hierarchical RNN in specific issues, please refer to \ :ref:`algo_hrnn_demo`\ 。The unit test file used in this article's example is \ `test_RecurrentGradientMachine.cpp `_\ 。 Example 1：Hierarchical RNN without Memory between subsequences ================================ @@ -13,8 +13,8 @@ The classical case in the hierarchical RNN is to perform sequence operations on In this example, the network configuration of single-layer RNNs and hierarchical RNNs are all to use LSTM as en encoder to compress a word-segmented sentence into a vector. The difference is that, RNN uses a hierarchical RNN model, treating multiple sentences as a whole to use encoder to compress simultaneously. They are completely consistent in their semantic meanings. This pair of semantically identical example configurations is as follows： -* RNN\: `sequence_layer_group.conf `_ -* Hierarchical RNN\: `sequence_nest_layer_group.conf `_ +* RNN\: `sequence_layer_group.conf `_ +* Hierarchical RNN\: `sequence_nest_layer_group.conf `_ Reading hierarchical sequence data @@ -24,18 +24,18 @@ Firstly, the original data in this example is as follows \: - The original data in this example has 10 samples. Each of the sample includes two components: a lable(all 2 here), and a word-segmented sentence. This data is used by single RNN as well. -.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg :language: text - The data for hierarchical RNN has 4 samples. Every sample is seperated by a blank line, while the content of the data is the same as the original data. But as for hierarchical LSTM, the first sample will encode two sentences into two vectors simultaneously. The sentence count dealed simultaneously by this 4 samples are \ :code:`[2, 3, 2, 3]`\ . -.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg.nest +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg.nest :language: text -Secondly, as for these two types of different input data formats, the contrast of different DataProviders are as follows (`sequenceGen.py `_)\： +Secondly, as for these two types of different input data formats, the contrast of different DataProviders are as follows (`sequenceGen.py `_)\： -.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py :language: python :lines: 21-39 :linenos: @@ -47,7 +47,7 @@ Secondly, as for these two types of different input data formats, the contrast o - "words" is a list of word table indices corresponding to each word in the sentence in the original data. Its data type is integer_value_sequence, that is integer list. So, "words" is a singler-layer time series in the data. - "label" is the categorical label of each sentence, whose data type is integer_value. -.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py :language: python :lines: 42-71 :linenos: @@ -64,7 +64,7 @@ Model configuration Firstly, let's look at the configuration of single-layer RNN. The hightlighted part of line 9 to line 15 is the usage of single-layer RNN. Here we use the pre-defined RNN process function in PaddlePaddle. In this function, for each time step, RNN passes through an LSTM network. -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_layer_group.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_layer_group.conf :language: python :lines: 38-63 :linenos: @@ -85,7 +85,7 @@ Secondly, let's look at the model configuration of hierarchical RNN which has th * Till now, \ :code:`lstm_last`\ has the same result as \ :code:`lstm_last`\ in single-layer RNN configuration. -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_layer_group.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_layer_group.conf :language: python :lines: 38-64 :linenos: @@ -107,7 +107,7 @@ We select the different parts between single-layer RNN and hierarchical RNN conf - single-layer RNN：passes through a simple recurrent_group. For each time step, the current input y and the last time step's output rnn_state pass through a fully-connected layer. -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn.conf :language: python :lines: 36-48 @@ -116,7 +116,7 @@ We select the different parts between single-layer RNN and hierarchical RNN conf - The recurrent_group of inner layer's inner_step is nearly the same as single-layer sequence, except for the case of boot_layer=outer_mem, which means using the outer layer's outer_mem as the initial state for the inner layer's memory. In the outer layer's out_step, outer_mem is the last vector of a subsequence, that is, the whole hierarchical group uses the last vector of the previous subsequence as the initial state for the next subsequence's memory. - From the aspect of the input data, sentences from single-layer and hierarchical RNN are the same. The only difference is that, hierarchical RNN disassembes the sequence into subsequences. So in the hierarchical RNN configuration, we must use the last element of the previous subsequence as a boot_layer for the memory of the next subsequence, so that it makes no difference with "every time step uses the output of last time step" in the sigle-layer RNN configuration. -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn.conf +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn.conf :language: python :lines: 39-66 @@ -134,7 +134,7 @@ Example 3：hierarchical RNN with unequal length inputs **unequal length inputs** means in the multiple input sequences of recurrent_group, the lengths of subsequences can be unequal. But the output of the sequence, needs to be consistent with one of the input sequences. Using \ :red:`targetInlink`\ can help you specify which of the input sequences and the output sequence can be consistent, by default is the first input. -The configurations of Example 3 are \ `sequence_rnn_multi_unequalength_inputs `_ \ and \ `sequence_nest_rnn_multi_unequalength_inputs `_\ . +The configurations of Example 3 are \ `sequence_rnn_multi_unequalength_inputs `_ \ and \ `sequence_nest_rnn_multi_unequalength_inputs `_\ . The data for the configurations of Example 3's single-layer RNN and hierarchical RNN are exactly the same. @@ -152,14 +152,14 @@ Similar to Example 2's configuration, Example 3's configuration uses single-laye * single-layer RNN\: -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py :language: python :lines: 42-59 :linenos: * hierarchical RNN\ \: -.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py +.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py :language: python :lines: 41-80 :linenos: diff --git a/go/pserver/optimizer.go b/go/pserver/optimizer.go index f17577997bc94b08f3e296c4d6e35682ca3c0e57..eba0c47e195a80fc298f0fdd78c8d6345e963be8 100644 --- a/go/pserver/optimizer.go +++ b/go/pserver/optimizer.go @@ -16,7 +16,7 @@ package pserver // #cgo CFLAGS: -I ../../ // #cgo LDFLAGS: ${SRCDIR}/client/c/libpaddle_go_optimizer.a -lstdc++ -lm -// #include "paddle/optimizer/optimizer.h" +// #include "paddle/legacy/optimizer/optimizer.h" // #include // #include import "C" diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index d722eec1892206ac44c49e7a12d92be0c54df8c0..6653244507742b33d9524a7a0e4a5b2b575d358a 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -1,24 +1,24 @@ if(NOT WITH_FLUID_ONLY) - add_subdirectory(cuda) - add_subdirectory(function) - add_subdirectory(utils) - add_subdirectory(math) - add_subdirectory(gserver) - add_subdirectory(parameter) + add_subdirectory(legacy/cuda) + add_subdirectory(legacy/function) + add_subdirectory(legacy/utils) + add_subdirectory(legacy/math) + add_subdirectory(legacy/gserver) + add_subdirectory(legacy/parameter) if(MOBILE_INFERENCE) - add_subdirectory(capi) + add_subdirectory(legacy/capi) else() - add_subdirectory(pserver) - add_subdirectory(trainer) + add_subdirectory(legacy/pserver) + add_subdirectory(legacy/trainer) add_subdirectory(scripts) if(WITH_C_API) - add_subdirectory(capi) + add_subdirectory(legacy/capi) endif() if(WITH_SWIG_PY) - add_subdirectory(api) + add_subdirectory(legacy/api) endif() endif() endif() diff --git a/paddle/api/Arguments.cpp b/paddle/api/Arguments.cpp deleted file mode 100644 index 62d6a574d55d2748635879a21cbbaa474f070cff..0000000000000000000000000000000000000000 --- a/paddle/api/Arguments.cpp +++ /dev/null @@ -1,174 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" -#include "PaddleAPIPrivate.h" - -#include "paddle/parameter/Argument.h" - -size_t Arguments::getSlotNum() const { return m->outputs.size(); } - -Arguments* Arguments::createArguments(size_t slotNum) { - auto args = new Arguments(); - args->m->outputs.resize(slotNum); - return args; -} - -void Arguments::resize(size_t slotNum) { m->outputs.resize(slotNum); } - -Arguments::Arguments() : m(new ArgumentsPrivate()) {} - -Arguments::~Arguments() { delete m; } - -Arguments* Arguments::createByPaddleArgumentVector(void* ptr) { - auto p = (std::vector*)(ptr); - auto args = new Arguments(); - args->m->outputs = *p; - return args; -} - -Arguments* Arguments::createByPaddleArgument(const void* ptr) { - auto p = (paddle::Argument*)(ptr); - auto args = new Arguments(); - args->m->outputs.push_back(*p); - return args; -} - -Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return Matrix::createByPaddleMatrixPtr(&a.value); -} - -Matrix* Arguments::getSlotGrad(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return Matrix::createByPaddleMatrixPtr(&a.grad); -} - -IVector* Arguments::getSlotIds(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return IVector::createByPaddleVectorPtr(&a.ids); -} - -Matrix* Arguments::getSlotIn(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return Matrix::createByPaddleMatrixPtr(&a.in); -} - -void Arguments::setSlotValue(size_t idx, Matrix* mat) throw(RangeError) { - auto& a = m->getArg(idx); - a.value = m->cast(mat->getSharedPtr()); -} - -void Arguments::setSlotGrad(size_t idx, Matrix* mat) throw(RangeError) { - auto& a = m->getArg(idx); - a.grad = m->cast(mat->getSharedPtr()); -} - -void Arguments::setSlotIn(size_t idx, Matrix* mat) throw(RangeError) { - auto& a = m->getArg(idx); - a.in = m->cast(mat->getSharedPtr()); -} - -void Arguments::setSlotIds(size_t idx, IVector* vec) throw(RangeError) { - auto& a = m->getArg(idx); - auto& v = m->cast(vec->getSharedPtr()); - a.ids = v; -} - -template -static inline void doCopyFromSafely(std::shared_ptr& dest, - std::shared_ptr& src) { - if (src) { - if (dest) { - dest->copyFrom(*src); - } else { - dest = src; - } - } -} - -IVector* Arguments::getSlotSequenceStartPositions(size_t idx) const - throw(RangeError) { - auto& a = m->getArg(idx); - if (a.sequenceStartPositions) { - return IVector::createByPaddleVectorPtr( - &a.sequenceStartPositions->getMutableVector(false)); - } else { - return nullptr; - } -} - -IVector* Arguments::getSlotSubSequenceStartPositions(size_t idx) const - throw(RangeError) { - auto& a = m->getArg(idx); - if (a.subSequenceStartPositions) { - return IVector::createByPaddleVectorPtr( - &a.subSequenceStartPositions->getMutableVector(false)); - } else { - return nullptr; - } -} - -void Arguments::setSlotSequenceStartPositions(size_t idx, - IVector* vec) throw(RangeError) { - auto& a = m->getArg(idx); - auto& v = m->cast(vec->getSharedPtr()); - a.sequenceStartPositions = std::make_shared(v); -} - -void Arguments::setSlotSubSequenceStartPositions( - size_t idx, IVector* vec) throw(RangeError) { - auto& a = m->getArg(idx); - auto& v = m->cast(vec->getSharedPtr()); - a.subSequenceStartPositions = std::make_shared(v); -} - -IVector* Arguments::getSlotSequenceDim(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return IVector::createByPaddleVectorPtr(&a.cpuSequenceDims); -} - -void Arguments::setSlotSequenceDim(size_t idx, IVector* vec) throw(RangeError) { - auto& a = m->getArg(idx); - a.cpuSequenceDims = m->cast(vec->getSharedPtr()); -} - -float Arguments::sum() const { return paddle::Argument::sum(m->outputs); } - -int64_t Arguments::getBatchSize(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return a.getBatchSize(); -} - -void Arguments::setSlotFrameHeight(size_t idx, size_t h) throw(RangeError) { - auto& a = m->getArg(idx); - a.setFrameHeight(h); -} - -void Arguments::setSlotFrameWidth(size_t idx, size_t w) throw(RangeError) { - auto& a = m->getArg(idx); - a.setFrameWidth(w); -} - -size_t Arguments::getSlotFrameHeight(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return a.getFrameHeight(); -} - -size_t Arguments::getSlotFrameWidth(size_t idx) const throw(RangeError) { - auto& a = m->getArg(idx); - return a.getFrameWidth(); -} - -void* Arguments::getInternalArgumentsPtr() const { return &m->outputs; } diff --git a/paddle/api/ConfigParser.cpp b/paddle/api/ConfigParser.cpp deleted file mode 100644 index d362a1e7cf3c8cd05b8c85cfaf8dbbee8b827d4b..0000000000000000000000000000000000000000 --- a/paddle/api/ConfigParser.cpp +++ /dev/null @@ -1,114 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" -#include "PaddleAPIPrivate.h" -#include "paddle/trainer/Trainer.h" - -struct ParameterConfigPrivate { - paddle::ParameterPtr parameter; - paddle::ParameterConfig config; - - inline paddle::ParameterConfig* getConfigPtr() { - if (parameter != nullptr) { - auto& conf = parameter->getConfig(); - return const_cast(&conf); - } else { - return &config; - } - } -}; - -TrainerConfig::TrainerConfig() : m(new TrainerConfigPrivate()) {} - -TrainerConfig::~TrainerConfig() { delete m; } - -TrainerConfig* TrainerConfig::createFromTrainerConfigFile( - const std::string& confPath) { - LOG(INFO) << "load trainer config from " << confPath; - auto conf = std::make_shared(confPath); - auto retv = new TrainerConfig(); - retv->m->conf = conf; - return retv; -} - -TrainerConfig* TrainerConfig::createFromProtoString(const std::string& str) { - auto retv = new TrainerConfig(); - paddle::TrainerConfig trainerConfigProto; - auto conf = std::make_shared(trainerConfigProto); - CHECK(conf->getMutableConfig().ParseFromString(str)); - retv->m->conf = conf; - return retv; -} - -ModelConfig::ModelConfig() : m(new ModelConfigPrivate()) {} - -ModelConfig::~ModelConfig() { delete m; } - -ModelConfig* TrainerConfig::getModelConfig() const { - auto retv = new ModelConfig(); - retv->m->conf = m->conf; - return retv; -} - -ParameterConfig::ParameterConfig() : m(new ParameterConfigPrivate()) {} - -ParameterConfig::~ParameterConfig() { delete m; } - -ParameterConfig* ParameterConfig::createParameterConfigFromParameterSharedPtr( - void* ptr) { - auto& p = *(paddle::ParameterPtr*)(ptr); - if (p != nullptr) { - auto conf = new ParameterConfig(); - conf->m->parameter = p; - return conf; - } else { - return nullptr; - } -} - -ParameterConfig* ParameterConfig::createParameterConfigFromParameterPtr( - void* ptr) { - auto& p = *(paddle::Parameter*)(ptr); - auto conf = new ParameterConfig(); - conf->m->config = p.getConfig(); - return conf; -} - -std::string ParameterConfig::toProtoString() const { - return m->getConfigPtr()->SerializeAsString(); -} - -void* ParameterConfig::getRawPtr() { return m->getConfigPtr(); } - -OptimizationConfig::OptimizationConfig() : m(new OptimizationConfigPrivate()) {} - -OptimizationConfig::~OptimizationConfig() { delete m; } - -std::string OptimizationConfig::toProtoString() { - return m->getConfig().SerializeAsString(); -} - -OptimizationConfig* TrainerConfig::getOptimizationConfig() const { - auto opt_config = new OptimizationConfig(); - opt_config->m->trainer_config = m->conf; - return opt_config; -} - -OptimizationConfig* OptimizationConfig::createFromProtoString( - const std::string& str) { - auto conf = new OptimizationConfig(); - conf->m->config.ParseFromString(str); - return conf; -} diff --git a/paddle/api/GradientMachine.cpp b/paddle/api/GradientMachine.cpp deleted file mode 100644 index 0d9ad30de9c1f3f8f58c856a748abdc050ff8740..0000000000000000000000000000000000000000 --- a/paddle/api/GradientMachine.cpp +++ /dev/null @@ -1,196 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" -#include "PaddleAPIPrivate.h" - -#include "Internal.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" - -std::vector GradientMachine::defaultParamTypes = { - PARAMETER_VALUE, PARAMETER_GRADIENT, PARAMETER_MOMENTUM}; - -GradientMachine::GradientMachine() : m(new GradientMachinePrivate()) {} - -GradientMachine::~GradientMachine() { delete m; } - -GradientMachine* GradientMachine::createFromPaddleModelPtr( - const void* confPtr, - GradientMatchineCreateMode mode, - const std::vector& types) { - auto& conf = *(const paddle::ModelConfig*)(confPtr); - std::vector realTypes; - staticCastVector(&realTypes, types); - auto machineRawPtr = paddle::GradientMachine::create(conf, mode, realTypes); - auto machinePtr = std::shared_ptr(machineRawPtr); - if (machinePtr != nullptr) { - auto machine = new GradientMachine(); - machine->m->machine = machinePtr; - return machine; - } else { - return nullptr; - } -} - -GradientMachine* GradientMachine::createByConfigProtoStr( - const std::string& protoStr, - GradientMatchineCreateMode mode, - const std::vector& types) { - paddle::ModelConfig conf; - conf.ParseFromString(protoStr); - if (conf.IsInitialized()) { - return GradientMachine::createFromPaddleModelPtr(&conf, mode, types); - } else { - return nullptr; - } -} - -GradientMachine* GradientMachine::createByModelConfig( - ModelConfig* conf, - GradientMatchineCreateMode mode, - const std::vector& types) { - auto confPtr = &conf->m->conf->getModelConfig(); - return GradientMachine::createFromPaddleModelPtr(confPtr, mode, types); -} - -void GradientMachine::start() { m->machine->start(); } - -void GradientMachine::finish() { m->machine->finish(); } - -void GradientMachine::onPassEnd() { m->machine->onPassEnd(); } - -void GradientMachine::prefetch(const Arguments& inArgs) { - auto& in = - m->cast>(inArgs.getInternalArgumentsPtr()); - m->machine->prefetch(in); -} - -void GradientMachine::forward(const Arguments& inArgs, - Arguments* outArgs, - PassType passType) { - auto& in = - m->cast>(inArgs.getInternalArgumentsPtr()); - auto& out = m->cast>( - outArgs->getInternalArgumentsPtr()); - paddle::PassType pt = (paddle::PassType)(passType); - m->machine->forward(in, &out, pt); -} - -UpdateCallback::~UpdateCallback() {} - -void UpdateCallback::apply(Parameter* p) { - // UNUSED(p); -} - -class UpdateCallbackWrapper { - public: - explicit UpdateCallbackWrapper(const UpdateCallback& callback) - : callback(const_cast(callback)) {} - - void operator()(paddle::Parameter* param) { - auto p = Parameter::createFromRawPtr(¶m); - // @TODO Use Stack variable instead. - callback.apply(p); - delete p; - } - - private: - UpdateCallback& callback; -}; - -void GradientMachine::backward(const UpdateCallback& callback) { - m->machine->backward(UpdateCallbackWrapper(callback)); -} - -void GradientMachine::forwardBackward(const Arguments& inArgs, - Arguments* outArgs, - PassType passType, - const UpdateCallback& callback) { - auto& in = - m->cast>(inArgs.getInternalArgumentsPtr()); - auto& out = m->cast>( - outArgs->getInternalArgumentsPtr()); - paddle::PassType pt = (paddle::PassType)(passType); - m->machine->forwardBackward(in, &out, pt, UpdateCallbackWrapper(callback)); -} - -void GradientMachine::loadParameters(const std::string& path) { - m->machine->loadParameters(path); -} - -size_t GradientMachine::getParameterSize() const { - return m->machine->getParameters().size(); -} - -Parameter* GradientMachine::getParameter(size_t i) throw(RangeError) { - auto params = m->machine->getParameters(); - if (i < params.size()) { - return Parameter::createFromSharedPtr(&m->machine->getParameters()[i]); - } else { - throw RangeError(); - } -} - -size_t GradientMachine::getNonStaticParameterSize() const { - return m->machine->getNonStaticParameters().size(); -} - -Parameter* GradientMachine::getNonStaticParameter(size_t i) throw(RangeError) { - auto params = m->machine->getNonStaticParameters(); - if (i < params.size()) { - return Parameter::createFromSharedPtr( - &m->machine->getNonStaticParameters()[i]); - } else { - throw RangeError(); - } -} - -void GradientMachine::randParameters() { m->machine->randParameters(); } - -Arguments* GradientMachine::getLayerOutput(const std::string& layerName) const - throw(UnsupportError) { - auto nn = m->machine; - if (nn) { - auto arg = nn->getLayerOutput(layerName); - return Arguments::createByPaddleArgument(&arg); - } else { - throw UnsupportError(); - } -} - -SequenceGenerator* GradientMachine::asSequenceGenerator( - const std::vector& dict, - size_t begin_id, - size_t end_id, - size_t max_length, - size_t beam_size) { - SequenceGenerator* r = - SequenceGenerator::createByGradientMachineSharedPtr(&m->machine); - r->setDict(dict); - r->setBos(begin_id); - r->setEos(end_id); - r->setMaxLength(max_length); - r->setBeamSize(beam_size); - return r; -} - -Evaluator* GradientMachine::makeEvaluator() { - auto ev = new Evaluator(); - ev->m->rawPtr = m->machine->makeEvaluator(); - return ev; -} - -void GradientMachine::eval(Evaluator* evaluator) { - m->machine->eval(evaluator->m->rawPtr); -} diff --git a/paddle/api/Matrix.cpp b/paddle/api/Matrix.cpp deleted file mode 100644 index 8282b4629dc08a7fcd9b52cbc3492ac10d8ed55c..0000000000000000000000000000000000000000 --- a/paddle/api/Matrix.cpp +++ /dev/null @@ -1,317 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/math/Matrix.h" -#include -#include -#include "PaddleAPI.h" -#include "paddle/math/CpuSparseMatrix.h" -#include "paddle/math/SparseMatrix.h" - -struct MatrixPrivate { - std::shared_ptr mat; -}; - -Matrix::Matrix() : m(new MatrixPrivate()) {} - -Matrix* Matrix::createByPaddleMatrixPtr(void* sharedPtr) { - auto* mat = reinterpret_cast(sharedPtr); - if ((*mat) != nullptr) { - auto m = new Matrix(); - m->m->mat = *mat; - return m; - } else { - return nullptr; - } -} - -Matrix* Matrix::createZero(size_t height, size_t width, bool useGpu) { - auto m = new Matrix(); - m->m->mat = paddle::Matrix::create(height, width, useGpu); - m->m->mat->zero(); - return m; -} - -Matrix* Matrix::createDense(const std::vector& data, - size_t height, - size_t width, - bool useGpu) { - auto m = new Matrix(); - m->m->mat = paddle::Matrix::create(height, width, useGpu); - m->m->mat->copyFrom(data.data(), data.size()); - return m; -} - -Matrix* Matrix::createDenseFromNumpy(float* data, - int dim1, - int dim2, - bool copy, - bool useGpu) throw(UnsupportError) { - if (useGpu) { - /// Gpu mode only supports copy=True - if (!copy) { - throw UnsupportError("Gpu mode only supports copy=True"); - } - return Matrix::createGpuDenseFromNumpy(data, dim1, dim2); - } else { - return Matrix::createCpuDenseFromNumpy(data, dim1, dim2, copy); - } -} - -Matrix* Matrix::createCpuDenseFromNumpy(float* data, - int dim1, - int dim2, - bool copy) { - auto m = new Matrix(); - if (copy) { - m->m->mat = paddle::Matrix::create(dim1, dim2); - m->m->mat->copyFrom(data, dim1 * dim2); - } else { - m->m->mat = paddle::Matrix::create(data, dim1, dim2, false); - } - return m; -} - -Matrix* Matrix::createGpuDenseFromNumpy(float* data, int dim1, int dim2) { - auto m = new Matrix(); - m->m->mat = paddle::Matrix::create(dim1, dim2, false, true); - m->m->mat->copyFrom(data, dim1 * dim2); - return m; -} - -Matrix* Matrix::createSparse(size_t height, - size_t width, - size_t nnz, - bool isNonVal, - bool isTrans, - bool useGpu) { - auto m = new Matrix(); - m->m->mat = paddle::Matrix::createSparseMatrix( - height, - width, - nnz, - isNonVal ? paddle::NO_VALUE : paddle::FLOAT_VALUE, - isTrans, - useGpu); - return m; -} - -Matrix::~Matrix() { delete m; } - -size_t Matrix::getHeight() const { return m->mat->getHeight(); } - -size_t Matrix::getWidth() const { return m->mat->getWidth(); } - -float Matrix::get(size_t x, size_t y) const throw(RangeError) { - if (x > this->getWidth() || y > this->getHeight()) { - RangeError e; - throw e; - } - return m->mat->getElement(x, y); -} - -void Matrix::set(size_t x, size_t y, float val) throw(RangeError, - UnsupportError) { - if (x > this->getWidth() || y > this->getHeight()) { - RangeError e; - throw e; - } - auto rawMat = m->mat.get(); - if (auto cDenseMat = dynamic_cast(rawMat)) { - *(cDenseMat->getData() + x + y * cDenseMat->getWidth()) = val; - } else { - UnsupportError e; - throw e; - } -} - -bool Matrix::isSparse() const { - auto raw_mat = m->mat.get(); - return dynamic_cast(raw_mat) != nullptr || - dynamic_cast(raw_mat) != nullptr; -} - -SparseValueType Matrix::getSparseValueType() const throw(UnsupportError) { - auto cpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (cpuSparseMat != nullptr) { - return (SparseValueType)cpuSparseMat->getValueType(); - } else { - auto gpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (gpuSparseMat != nullptr) { - return (SparseValueType)gpuSparseMat->getValueType(); - } else { - UnsupportError e; - throw e; - } - } -} - -SparseFormatType Matrix::getSparseFormat() const throw(UnsupportError) { - auto cpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (cpuSparseMat != nullptr) { - return (SparseFormatType)cpuSparseMat->getFormat(); - } else { - auto gpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (gpuSparseMat != nullptr) { - return SPARSE_CSR; - } else { - UnsupportError e; - throw e; - } - } -} - -IntArray Matrix::getSparseRowCols(size_t i) const - throw(UnsupportError, RangeError) { - auto cpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (cpuSparseMat != nullptr && - cpuSparseMat->getFormat() == paddle::SPARSE_CSR) { - if (i < cpuSparseMat->getHeight()) { - // cpuSparseMat->print(std::cout); - size_t len = cpuSparseMat->getColNum(i); - return IntArray(cpuSparseMat->getRowCols(i), len); - } else { - RangeError e; - throw e; - } - } else { - UnsupportError e; - throw e; - } -} - -IntWithFloatArray Matrix::getSparseRowColsVal(size_t i) const - throw(UnsupportError, RangeError) { - auto cpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (cpuSparseMat != nullptr && - cpuSparseMat->getValueType() == paddle::FLOAT_VALUE) { - if (i < cpuSparseMat->getHeight()) { - return IntWithFloatArray(cpuSparseMat->getRowValues(i), - cpuSparseMat->getRowCols(i), - cpuSparseMat->getColNum(i)); - } else { - RangeError e; - throw e; - } - } else { - UnsupportError e; - throw e; - } -} - -FloatArray Matrix::getData() const { - auto rawMat = m->mat.get(); - if (dynamic_cast(rawMat->getMemoryHandle().get())) { - // is gpu. then copy data - float* data = rawMat->getData(); - size_t len = rawMat->getElementCnt(); - float* cpuData = new float[len]; - hl_memcpy_device2host(cpuData, data, len * sizeof(float)); - FloatArray ret_val(cpuData, len); - ret_val.needFree = true; - return ret_val; - } else { - FloatArray ret_val(rawMat->getData(), rawMat->getElementCnt()); - return ret_val; - } -} - -void Matrix::sparseCopyFrom( - const std::vector& rows, - const std::vector& cols, - const std::vector& vals) throw(UnsupportError) { - auto cpuSparseMat = - std::dynamic_pointer_cast(m->mat); - if (cpuSparseMat != nullptr) { - // LOG(INFO) <<"RowSize = "<isSparse()) { - throw UnsupportError(); - } else { - *dim1 = m->mat->getHeight(); - *dim2 = m->mat->getWidth(); - *view_m_data = new float[(*dim1) * (*dim2)]; - if (auto cpuMat = dynamic_cast(m->mat.get())) { - auto src = cpuMat->getData(); - auto dest = *view_m_data; - std::memcpy(dest, src, sizeof(paddle::real) * (*dim1) * (*dim2)); - } else if (auto gpuMat = dynamic_cast(m->mat.get())) { - auto src = gpuMat->getData(); - auto dest = *view_m_data; - hl_memcpy_device2host( - dest, src, sizeof(paddle::real) * (*dim1) * (*dim2)); - } else { - LOG(WARNING) << "Unexpected Situation"; - throw UnsupportError(); - } - } -} - -void Matrix::copyFromNumpyMat(float* data, - int dim1, - int dim2) throw(UnsupportError, RangeError) { - if (isSparse()) { - throw UnsupportError(); - } else { - if (this->getHeight() == (size_t)dim1 && this->getWidth() == (size_t)dim2) { - if (m->mat->getData() != data) { - m->mat->copyFrom(data, dim1 * dim2); - } - } else { - throw RangeError(); - } - } -} - -bool Matrix::isGpu() const { - auto rawPtr = m->mat.get(); - return dynamic_cast(rawPtr) != nullptr || - dynamic_cast(rawPtr) != nullptr; -} diff --git a/paddle/api/Paddle.i b/paddle/api/Paddle.i deleted file mode 100644 index 3237e73745dca58bed923b20851f0f0039a3487c..0000000000000000000000000000000000000000 --- a/paddle/api/Paddle.i +++ /dev/null @@ -1,202 +0,0 @@ -%module(directors="1") swig_paddle -%include "std_string.i" -%{ -#define SWIG_FILE_WITH_INIT -#include "api/PaddleAPI.h" -%} - -%include "exception.i" -%typemap(throws) UnsupportError %{ - SWIG_exception(SWIG_RuntimeError, $1.what()); - SWIG_fail; -%} - -%include "std_vector.i" -%include "std_pair.i" -#ifdef SWIGPYTHON -%include "numpy.i" -#endif - -%init %{ -#ifdef SWIGPYTHON -import_array(); -#endif -%} - - -namespace std { -%template(vector_int) vector; -%template(vector_uint) vector; -%template(vector_float) vector; -%template(vector_string) vector; -%template(vector_vec_star) vector; -} -#ifdef SWIGPYTHON -%typemap(in) (int argc, char** argv) { - int i = 0; - if (!PyList_Check($input)) { - PyErr_SetString(PyExc_ValueError, "Expecting a list"); - return NULL; - } - $1 = PyList_Size($input); - $2 = (char **) malloc(($1+1)*sizeof(char *)); - for (i = 0; i < $1; i++) { - PyObject *s = PyList_GetItem($input,i); - if (!PyString_Check(s)) { - free($2); - PyErr_SetString(PyExc_ValueError, "List items must be strings"); - return NULL; - } - $2[i] = PyString_AsString(s); - } - $2[i] = 0; -} -%typemap(freearg) (int argc, char** argv) { - if ($2) free($2); -} - -%typemap(out) FloatArray { - $result = PyList_New($1.length); - for (size_t i=0; i<$1.length; ++i) { - PyList_SetItem($result, i, PyFloat_FromDouble($1.buf[i])); - } - if($1.needFree) { - delete [] $1.buf; - } -} - -%typemap(out) IntArray { - $result = PyList_New($1.length); - for (size_t i=0; i<$1.length; ++i) { - PyList_SetItem($result, i, PyInt_FromLong($1.buf[i])); - } - if ($1.needFree) { - delete [] $1.buf; - } -} - -%typemap(out) IntWithFloatArray { - $result = PyList_New($1.length); - for (size_t i=0; i<$1.length; ++i) { - PyList_SetItem($result, i, PyTuple_Pack(2, - PyInt_FromLong($1.idxBuf[i]), - PyFloat_FromDouble($1.valBuf[i]) - )); - } - if ($1.needFree) { - delete [] $1.idxBuf; - delete [] $1.valBuf; - } -} - - -%rename(__getitem__) IVector::get; -%rename(__setitem__) IVector::set; -%rename(__len__) IVector::getSize; -%rename(__getitem__) Vector::get; -%rename(__setitem__) Vector::set; -%rename(__len__) Vector::getSize; -%rename(__len__) Parameter::getSize; -%rename(__call__) ParameterTraverseCallback::apply; -%rename(__repr__) Evaluator::toString; - -%apply (float* INPLACE_ARRAY2, int DIM1, int DIM2) { - (float* data, int dim1, int dim2) -} - -%apply (float** ARGOUTVIEW_ARRAY2, int* DIM1, int* DIM2) { - (float** view_data, int* dim1, int* dim2) -} - -%apply (float** ARGOUTVIEWM_ARRAY2, int* DIM1, int* DIM2) { - (float** view_m_data, int* dim1, int* dim2) -} - -%apply (int** ARGOUTVIEWM_ARRAY1, int* DIM1) { - (int** view_m_data, int* dim1) -} - -%apply (int* INPLACE_ARRAY1, int DIM1) { - (int* data, int dim) -} - -%apply (int** ARGOUTVIEW_ARRAY1, int* DIM1) { - (int** view_data, int* dim1) -} - -%apply (float* INPLACE_ARRAY1, int DIM1) { - (float* data, int dim) -} - -%apply (float** ARGOUTVIEW_ARRAY1, int* DIM1) { - (float** view_data, int* dim1) -} - -%apply (float** ARGOUTVIEWM_ARRAY1, int* DIM1) { - (float** view_m_data, int* dim1) -} - -#endif -// The below functions internally create object by "new", so it should use -// use SWIG to handle gc. There are hints for SWIG to handle GC. -%newobject Matrix::createZero; -%newobject Matrix::createSparse; -%newobject Matrix::createDense; -%newobject Matrix::createDenseFromNumpy; -%newobject Matrix::createCpuDenseFromNumpy; -%newobject Matrix::createGpuDenseFromNumpy; -%newobject Vector::createZero; -%newobject Vector::create; -%newobject Vector::createVectorFromNumpy; -%newobject Vector::createCpuVectorFromNumpy; -%newobject Vector::createGpuVectorFromNumpy; -%newobject IVector::createZero; -%newobject IVector::create; -%newobject IVector::createVectorFromNumpy; -%newobject IVector::createCpuVectorFromNumpy; -%newobject IVector::createGpuVectorFromNumpy; -%newobject Trainer::createByCommandLine; -%newobject Trainer::getForwardOutput; -%newobject Trainer::getLayerOutput; -%newobject Arguments::getSlotValue; -%newobject Arguments::getSlotIds; -%newobject Arguments::getSlotIn; -%newobject Arguments::getSlotSequenceStartPositions; -%newobject Arguments::getSlotSequenceDim; -%newobject Arguments::createArguments; -%newobject GradientMachine::createByConfigProtoStr; -%newobject GradientMachine::createByModelConfig; -%newobject GradientMachine::asSequenceGenerator; -%newobject GradientMachine::getParameter; -%newobject GradientMachine::getLayerOutput; -%newobject GradientMachine::makeEvaluator; -%newobject TrainerConfig::createFromTrainerConfigFile; -%newobject TrainerConfig::getModelConfig; -%newobject TrainerConfig::getOptimizationConfig; -%newobject Parameter::getBuf; -%newobject Parameter::getConfig; -%newobject ParameterOptimizer::create; -%newobject ParameterOptimizer::needSpecialTraversal; -%newobject ParameterUpdater::createLocalUpdater; -%newobject ParameterUpdater::createRemoteUpdater; -%newobject ParameterUpdater::createNewRemoteUpdater; - -%feature("director") UpdateCallback; -%feature("autodoc", 1); // To generate method stub, for code hint in ide - -// Ignore many private class, and method cannot be handled by swig. -%ignore MatrixPrivate; -%ignore TrainerPrivate; -%ignore IVector::operator[]; -%ignore ArgumentsPrivate; -%ignore GradientMachinePrivate; -%ignore TrainerConfigPrivate; -%ignore ModelConfigPrivate; -%ignore ParameterPrivate; -%ignore SequenceGeneratorPrivate; -%ignore VectorPrivate; -%ignore ParameterConfigPrivate; -%ignore OptimizationConfigPrivate; -%ignore ParameterTraverseCallbackPrivate; -%include "utils/GlobalConstants.h" -%include "api/PaddleAPI.h" diff --git a/paddle/api/PaddleAPI.h b/paddle/api/PaddleAPI.h deleted file mode 100644 index 7866122006a996cbe5201c661cab9c81aa82a219..0000000000000000000000000000000000000000 --- a/paddle/api/PaddleAPI.h +++ /dev/null @@ -1,1054 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include -#include -#include -#include "paddle/gserver/gradientmachines/GradientMachine.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/GlobalConstants.h" - -/// Import PaddlePaddle's enumeration into global namespace. -using namespace paddle::enumeration_wrapper; // NOLINT - -/** - * @brief Initialize paddle. - * - * In python, this method should be invoked as - * @code - * import sys - * import paddle - * paddle.initPaddle(sys.argv) - * or you can change arguments as any list of str. - * @endcode - */ -void initPaddle(int argc, char** argv); - -/// Return FLAGS_use_gpu -bool isUsingGpu(); - -/// Set the Flags_use_gpu to the given parameter -void setUseGpu(bool useGpu); - -/// Return true if this py_paddle is compiled in GPU Version -bool isGpuVersion(); - -/// Return FLAGS_trainer_count -int getTrainerCount(); - -/// The Error of IO Operation. Such as file not found, etc. -class IOError {}; - -/// Out of range error -class RangeError {}; - -/// Not support Error, such as access GPU memory directly, etc. -class UnsupportError : public std::runtime_error { - public: - UnsupportError() : std::runtime_error(" ") {} - explicit UnsupportError(const std::string& message) - : std::runtime_error(message) {} -}; - -/// This type will map to python's list of float. -struct FloatArray { - const float* buf; - const size_t length; - bool needFree; // true if the buf is dynamic alloced. - FloatArray(const float* b, const size_t l); -}; - -/// This type will map to python's list of int -struct IntArray { - const int* buf; - const size_t length; - bool needFree; - IntArray(const int* b, const size_t l, bool f = false); -}; - -/// This type will map to python's list of (int, float) -struct IntWithFloatArray { - const float* valBuf; - const int* idxBuf; - const size_t length; - bool needFree; - IntWithFloatArray(const float* v, const int* i, size_t l, bool f = false); -}; - -enum SparseValueType { SPARSE_NON_VALUE = 0, SPARSE_VALUE = 1 }; - -enum SparseFormatType { SPARSE_CSR = 0, SPARSE_CSC = 1 }; - -/** - * In Python, -1UL is hard to write. So define a const value used by python - * side. - */ -const size_t NO_SPARSE_ID = -1UL; - -struct MatrixPrivate; -class Matrix { - Matrix(); // User Cannot Create Matrix. - DISABLE_COPY(Matrix); - static Matrix* createByPaddleMatrixPtr(void* sharedPtr); - - public: - virtual ~Matrix(); - - /** - * Create A Matrix with height,width, which is filled by zero. - */ - static Matrix* createZero(size_t height, - size_t width, - bool useGpu = isUsingGpu()); - - /** - * Create Sparse Matrix. - * - * After create sparse, sparseCopyFrom can be used to fill matrix. - * - * @param nnz Number of non zero values. - * - * @note the default sparse type is SPARSE_CSR. - */ - static Matrix* createSparse(size_t height, - size_t width, - size_t nnz, - bool isNonVal = true, - bool trans = false, - bool useGpu = isUsingGpu()); - - /** - * Create Dense Matrix. - * - * @param data list of float should be passed in python. - * @note the value will be copy into a new matrix. - */ - static Matrix* createDense(const std::vector& data, - size_t height, - size_t width, - bool useGpu = isUsingGpu()); - - static Matrix* createDenseFromNumpy( - float* data, - int dim1, - int dim2, - bool copy = true, - bool useGpu = isUsingGpu()) throw(UnsupportError); - - /** - * Create Cpu Dense Matrix from numpy matrix, dtype=float32 - * - * @param data a numpy matrix. - * @param dim1 dimension of data. - * @param dim2 dimension of data. - * @param copy true if copy into a new matrix, false will create - * matrix inplace. copy = false should be used with extreme - * care because Matrix will share the memory with the given - * numpy array. If the numpy array object is no longer valid, - * the memory space will not be usable. - */ - static Matrix* createCpuDenseFromNumpy(float* data, - int dim1, - int dim2, - bool copy = true); - - /// Create Gpu Dense Matrix from numpy matrix, dtype=float32 - static Matrix* createGpuDenseFromNumpy(float* data, int dim1, int dim2); - - /** - * Cast to numpy matrix. - * - * @note This method take no parameter in python. - * @note This method in python will return a numpy matrix, not void. - * @note Only CpuDenseMatrix is supported. - * - * Example: - * @code - * import paddle - * m = paddle.Matrix.createZero(10,2) - * numpy_mat = m.toNumpyMat() - * @endcode - */ - void toNumpyMatInplace(float** view_data, - int* dim1, - int* dim2) throw(UnsupportError); - - /// Copy To numpy mat. - void copyToNumpyMat(float** view_m_data, - int* dim1, - int* dim2) throw(UnsupportError); - - /// Copy From Numpy Mat - void copyFromNumpyMat(float* data, int dim1, int dim2) throw(UnsupportError, - RangeError); - - /// return true if this matrix is sparse. - bool isSparse() const; - - SparseValueType getSparseValueType() const throw(UnsupportError); - - SparseFormatType getSparseFormat() const throw(UnsupportError); - - IntArray getSparseRowCols(size_t i) const throw(UnsupportError, RangeError); - - IntWithFloatArray getSparseRowColsVal(size_t i) const - throw(UnsupportError, RangeError); - - size_t getHeight() const; - - size_t getWidth() const; - - float get(size_t x, size_t y) const throw(RangeError); - - void set(size_t x, size_t y, float val) throw(RangeError, UnsupportError); - - /// return type is list of float - FloatArray getData() const; - - /** - * Copy from rows, cols, values. - * - * if sparse_nonvalue, the values should be [] - */ - void sparseCopyFrom(const std::vector& rows, - const std::vector& cols, - const std::vector& values = - std::vector()) throw(UnsupportError); - - bool isGpu() const; - - private: - void* getSharedPtr() const; - - MatrixPrivate* m; - friend class Trainer; - friend class GradientMachine; - friend class Arguments; -}; - -struct VectorPrivate; -class Vector { - DISABLE_COPY(Vector); - Vector(); - static Vector* createByPaddleVectorPtr(void* ptr); - - void* getSharedPtr(); - - public: - ~Vector(); - - /// Create Vector filled with zero. - static Vector* createZero(size_t sz, bool useGpu = isUsingGpu()); - - /** - * Create Vector from list of float. - * - * It will create a new vector, and copy data into it. - */ - static Vector* create(const std::vector& data, - bool useGpu = isUsingGpu()); - - static Vector* createVectorFromNumpy( - float* data, - int dim, - bool copy = true, - bool useGpu = isUsingGpu()) throw(UnsupportError); - /** - * Create Cpu Vector from numpy array, which dtype=float32 - * - * If copy is false, it will create vector inplace. - */ - static Vector* createCpuVectorFromNumpy(float* data, - int dim, - bool copy = true); - - /// Create Gpu Vector from numpy array, which dtype=float32 - static Vector* createGpuVectorFromNumpy(float* data, int dim); - - /** - * copy from another vector - * throw(RangeError) if size of src vector is different from size of this - * vector - */ - void copyFrom(Vector* src) throw(RangeError); - - /// Cast to numpy array inplace. - void toNumpyArrayInplace(float** view_data, int* dim1) throw(UnsupportError); - - /// Copy to numpy array. - void copyToNumpyArray(float** view_m_data, int* dim1); - - /// Copy from numpy array. - void copyFromNumpyArray(float* data, int dim); - - /// __getitem__ in python - float get(const size_t idx) const throw(RangeError, UnsupportError); - - /// __setitem__ in python - void set(const size_t idx, float val) throw(RangeError, UnsupportError); - - /// Return is GPU vector or not. - bool isGpu() const; - - /// Return a list of float, the memory is alloced and copied. - FloatArray getData() const; - - /// __len__ in python - size_t getSize() const; - - private: - VectorPrivate* m; - - private: - friend class Parameter; - friend class ParameterOptimizer; - friend struct ParameterTraverseCallbackPrivate; -}; - -struct IVectorPrivate; -class IVector { - IVector(); - DISABLE_COPY(IVector); - static IVector* createByPaddleVectorPtr(void* ptr); - - public: - /// Create IVector filled with zero - static IVector* createZero(size_t sz, bool useGpu = isUsingGpu()); - - /** - * Create IVector from list of int. - * It will create a new vector, and copy data into it. - */ - static IVector* create(const std::vector& data, - bool useGpu = isUsingGpu()); - - static IVector* createVectorFromNumpy( - int* data, - int dim, - bool copy = true, - bool useGpu = isUsingGpu()) throw(UnsupportError); - - /** - * Create Cpu IVector from numpy array, which dtype=int32 - * - * If copy is false, it will create vector inplace - */ - static IVector* createCpuVectorFromNumpy(int* data, - int dim, - bool copy = true); - /** - * Create Gpu IVector from numpy array, which dtype=int32 - */ - static IVector* createGpuVectorFromNumpy(int* data, int dim); - - /// Cast to numpy array inplace. - void toNumpyArrayInplace(int** view_data, int* dim1) throw(UnsupportError); - - /// Copy to numpy array. - void copyToNumpyArray(int** view_m_data, int* dim1); - - /// Copy from numpy array. - void copyFromNumpyArray(int* data, int dim); - - virtual ~IVector(); - - /// Return a list of int, the memory is alloced and copied. - IntArray getData() const; - - /// This method will map to python [] method. - int& operator[](const size_t idx) throw(RangeError, UnsupportError); - - const int& operator[](const size_t idx) const - throw(RangeError, UnsupportError); - - inline int get(const size_t idx) const throw(RangeError, UnsupportError) { - return (*this)[idx]; - } - - inline void set(const size_t idx, int val) throw(RangeError, UnsupportError) { - (*this)[idx] = val; - } - - /// Return true if it is gpu vector. - bool isGpu() const; - - /// This method will map to python __len__(); - size_t getSize() const; - - private: - void* getSharedPtr() const; - - friend class Arguments; - IVectorPrivate* m; -}; - -struct ArgumentsPrivate; - -/// The Arguments is actual a std::vector in paddle. -class Arguments { - private: - Arguments(); // Internal Create. - DISABLE_COPY(Arguments); - - public: - /** - * Create a arguments with size. - * Note that it can be zero. - */ - static Arguments* createArguments(size_t slotNum); - - void resize(size_t slotNum); - - virtual ~Arguments(); - - /** - * Return the slot number that aguments contains. - * - * It is actually the vector's size - */ - size_t getSlotNum() const; - - /** - * The get functions of Arguments - * - * the param idx is the slot id - */ - Matrix* getSlotValue(size_t idx) const throw(RangeError); - Matrix* getSlotGrad(size_t idx) const throw(RangeError); - IVector* getSlotIds(size_t idx) const throw(RangeError); - Matrix* getSlotIn(size_t idx) const throw(RangeError); - IVector* getSlotSequenceStartPositions(size_t idx) const throw(RangeError); - IVector* getSlotSubSequenceStartPositions(size_t idx) const throw(RangeError); - IVector* getSlotSequenceDim(size_t idx) const throw(RangeError); - // End Of get functions of Arguments - - int64_t getBatchSize(size_t idx = 0) const throw(RangeError); - - /** - * The set functions of Arguments. - * - * The param idx is the slot id. - * The other param is the input Matrix or vector. - */ - void setSlotValue(size_t idx, Matrix* mat) throw(RangeError); - void setSlotGrad(size_t idx, Matrix* mat) throw(RangeError); - void setSlotIn(size_t idx, Matrix* mat) throw(RangeError); - void setSlotIds(size_t idx, IVector* vec) throw(RangeError); - void setSlotSequenceStartPositions(size_t idx, - IVector* vec) throw(RangeError); - void setSlotSubSequenceStartPositions(size_t idx, - IVector* vec) throw(RangeError); - void setSlotSequenceDim(size_t idx, IVector* vec) throw(RangeError); - - /** - * Set the frame height of the idx-th Argument. - * - * @param ids The index of which Argument. - * @param h The height value. - */ - void setSlotFrameHeight(size_t idx, size_t h) throw(RangeError); - - /** - * Set the frame height of the idx-th Argument. - * - * @param ids The index of which Argument. - * @param h The height value. - */ - void setSlotFrameWidth(size_t idx, size_t w) throw(RangeError); - - size_t getSlotFrameHeight(size_t idx = 0) const throw(RangeError); - size_t getSlotFrameWidth(size_t idx = 0) const throw(RangeError); - - float sum() const; - - private: - static Arguments* createByPaddleArgumentVector(void* ptr); - static Arguments* createByPaddleArgument(const void* ptr); - void* getInternalArgumentsPtr() const; - - private: - ArgumentsPrivate* m; - friend class Trainer; - friend class GradientMachine; - friend class SequenceGenerator; -}; - -enum GradientMatchineCreateMode { - CREATE_MODE_NORMAL = paddle::GradientMachine::kNormal, - CREATE_MODE_SGD_SPARSE_CPU_TRAINING = - paddle::GradientMachine::kSgdSparseCpuTraining, - CREATE_MODE_TESTING = paddle::GradientMachine::kTesting -}; - -struct ParameterConfigPrivate; -class ParameterConfig { - DISABLE_COPY(ParameterConfig); - ParameterConfig(); - - /** - * Internal methods - */ - static ParameterConfig* createParameterConfigFromParameterSharedPtr( - void* ptr); - static ParameterConfig* createParameterConfigFromParameterPtr(void* ptr); - void* getRawPtr(); - - public: - ~ParameterConfig(); - - /** - * return proto buf string. - */ - std::string toProtoString() const; - - private: - ParameterConfigPrivate* m; - - private: - friend class Parameter; - friend class ParameterOptimizer; - friend struct ParameterTraverseCallbackPrivate; -}; - -struct OptimizationConfigPrivate; -class OptimizationConfig { - DISABLE_COPY(OptimizationConfig); - OptimizationConfig(); - - public: - static OptimizationConfig* createFromProtoString(const std::string& str); - ~OptimizationConfig(); - - /** - * return protobuf string. - */ - std::string toProtoString(); - - private: - OptimizationConfigPrivate* m; - - friend class TrainerConfig; - friend class ParameterOptimizer; - friend class ParameterUpdater; - friend class Trainer; -}; - -struct ParameterPrivate; -class Parameter { - private: - Parameter(); - DISABLE_COPY(Parameter); - - public: - virtual ~Parameter(); - - /** - * get parameter name - */ - std::string getName() const; - - /** - * get buf in Parameter - */ - Vector* getBuf(ParameterType type); - - /** - * get id - */ - size_t getID() const; - - ParameterConfig* getConfig(); - void setValueUpdated(); - - bool save(const std::string& filename) const; - - bool load(const std::string& filename) const; - - size_t getSize() const; - - private: - static Parameter* createFromRawPtr(void* ptr); - static Parameter* createFromSharedPtr(void* ptr); - - private: - ParameterPrivate* m; - friend class UpdateCallbackWrapper; - friend class GradientMachine; - friend class ParameterUpdater; -}; - -struct ModelConfigPrivate; -/** - * You can only get model config from TrainerConfig. - * - * It is used by GradientMachine. - */ -class ModelConfig { - private: - ModelConfig(); - DISABLE_COPY(ModelConfig); - - public: - virtual ~ModelConfig(); - - private: - ModelConfigPrivate* m; - friend class TrainerConfig; - friend struct TrainerConfigPrivate; - friend class GradientMachine; -}; - -struct TrainerConfigPrivate; -/** - * To get TrainerConfig from file. - * - * It is used by GradientMachine. - */ -class TrainerConfig { - private: - TrainerConfig(); - DISABLE_COPY(TrainerConfig); - - public: - virtual ~TrainerConfig(); - - static TrainerConfig* createFromTrainerConfigFile( - const std::string& configPath); - static TrainerConfig* createFromProtoString(const std::string& str); - - ModelConfig* getModelConfig() const; - - OptimizationConfig* getOptimizationConfig() const; - - private: - TrainerConfigPrivate* m; - friend class Trainer; -}; - -/** - * The callback in backword. - * - * You can inherit this class in python. - * - * @code - * class UpdateCallbackInPython(paddle.UpdateCallback): - * def __init__(self): - * paddle.UpdateCallback.__init__(self) - * - * def apply(self, param): - * assert isinstance(param, paddle.Parameter) - * @endcode - */ -class UpdateCallback { - public: - virtual ~UpdateCallback(); - virtual void apply(Parameter* p); -}; - -struct ParameterTraverseCallbackPrivate; -class ParameterTraverseCallback { - DISABLE_COPY(ParameterTraverseCallback); - ParameterTraverseCallback(); - - public: - ~ParameterTraverseCallback(); - - void apply(const std::vector& vecs, - const ParameterConfig& config, - size_t sparseId); - - private: - ParameterTraverseCallbackPrivate* m; - friend class ParameterOptimizer; -}; - -/** - * The ParameterOptimizer Wrapper Class. - * - * Basically same as common/ParameterOptimizer.h - */ -struct ParameterOptimizerPrivate; -class ParameterOptimizer { - DISABLE_COPY(ParameterOptimizer); - ParameterOptimizer(); - - public: - static ParameterOptimizer* create(OptimizationConfig* config); - - ~ParameterOptimizer(); - - void init(size_t numRows, const ParameterConfig* config); - - void startPass(); - - void finishPass(); - - void startBatch(size_t numSamplesProcessed); - - void finishBatch(); - - void update(const std::vector& vecs, - const ParameterConfig& conf, - size_t sparseId = NO_SPARSE_ID); - - std::vector getParameterTypes() const; - - ParameterTraverseCallback* needSpecialTraversal( - const ParameterConfig& config) const; - - private: - ParameterOptimizerPrivate* m; -}; - -class SequenceGenerator; -class Evaluator; -struct GradientMachinePrivate; -class GradientMachine { - private: - GradientMachine(); - DISABLE_COPY(GradientMachine); - - public: - virtual ~GradientMachine(); - - /** - * Create By ProtoStr. - * - * The ProtoStr can be generate by python's protobuf code. - */ - static GradientMachine* createByConfigProtoStr( - const std::string& protoStr, - GradientMatchineCreateMode mode = CREATE_MODE_NORMAL, - const std::vector& parameterTypes = defaultParamTypes); - - /** - * Create by ModelConfig object. - * - * To get ModelConfig, you can get TrainerConfig from config file, then get - * model config by TrainerConfig - */ - static GradientMachine* createByModelConfig( - ModelConfig* conf, - GradientMatchineCreateMode mode = CREATE_MODE_NORMAL, - const std::vector& parameterTypes = defaultParamTypes); - - /** - * @brief finish - */ - void finish(); - - void start(); - - /** - * Prefetch row ids of sparse parameter. - */ - void prefetch(const Arguments& inArgs); - - /** - * Do some thing when train pass ended. - */ - void onPassEnd(); - - /** - * The forward stage of GradientMachine. - * - * @note the outArgs could be zero length arguemnts. - * @note THIS METHOD IS VERY USEFULL FOR PREDICT FROM TRAINED MODEL. - */ - void forward(const Arguments& inArgs, Arguments* outArgs, PassType passType); - - /** - * The backward stage of GradientMachine. - * - * @note Currently the ParameterUpdater is not wrapped in SWIG, so backward - * cannot actually train a network. But you can write a update callback to - * change the parameter or implement a ParameterUpdater in python side. - */ - void backward(const UpdateCallback& callback = UpdateCallback()); - - /** - * Combine forward/backward - */ - void forwardBackward(const Arguments& inArgs, - Arguments* outArgs, - PassType passType, - const UpdateCallback& callback = UpdateCallback()); - - void loadParameters(const std::string& path); - - size_t getParameterSize() const; - Parameter* getParameter(size_t i) throw(RangeError); - - size_t getNonStaticParameterSize() const; - Parameter* getNonStaticParameter(size_t i) throw(RangeError); - - void randParameters(); - - Arguments* getLayerOutput(const std::string& layerName) const - throw(UnsupportError); - - /** - * Create a sequence generator. - * - * @note It just like a paddle_gen_sequence. - */ - SequenceGenerator* asSequenceGenerator( - const std::vector& dict = std::vector(), - size_t begin_id = 0UL, - size_t end_id = 0UL, - size_t max_length = 100UL, - size_t beam_size = -1UL); - - Evaluator* makeEvaluator(); - - void eval(Evaluator* evaluator); - - private: - GradientMachinePrivate* m; - - static GradientMachine* createFromPaddleModelPtr( - const void* confPtr, - GradientMatchineCreateMode mode, - const std::vector& types); - - // Not to use c++ 11 init-list, so we use static var as function default arg. - static std::vector defaultParamTypes; - friend class Trainer; - friend class ParameterUpdater; -}; - -struct ParameterUpdaterPrivate; -class ParameterUpdater { - private: - ParameterUpdater(); - - public: - static ParameterUpdater* createLocalUpdater(OptimizationConfig* config); - static ParameterUpdater* createRemoteUpdater(OptimizationConfig* config, - int passCount, - bool useSparseUpdater); - static ParameterUpdater* createNewRemoteUpdater( - OptimizationConfig* config, - const std::string pserverSpec, - const bool useEtcd) throw(UnsupportError); - ~ParameterUpdater(); - - /** - * @brief initialize Parameter Updater by GradientMachine. - * @param gm - */ - void init(const GradientMachine& gm); - - /** - * @brief begin of a training/testing of one pass. - */ - void startPass(); - - /** - * @brief end of a traning/testing of one pass. - */ - void finishPass(); - - /** - * @brief begin of a training/testing of one batch. - * @param data batch's size - * @return PassType, mostly will be training. - */ - PassType startBatch(size_t batchSize); - - /** - * @brief end of a traning/testing of one batch - * @param cost current batch cost. - */ - void finishBatch(float cost); - - /** - * @brief update a parameter (by local optimizer or by cluster pserver) - * @param param - */ - void update(Parameter* param); - - /** - * @breif only get required sparse rows by default. - * @param fullSize: get full matrix parameter if *fullSize* set - * @param apply: get PARAMETER_APPLY on pserver if *apply* set - */ - void getParametersRemote(bool fullSize = false, bool apply = false); - - /** - * @brief restore the average parameter. - * @note It is only used in AverageOptimizer. Restore will get the current - * PARAMETER_VALUE back. - */ - void restore(); - - /** - * @brief apply. Store the average parameter. - * @note It is only used in AverageOptimizer. Apply will store the current - * PARAMETER_VALUE to buffer, calcaualte current Average Parameter, and save - * it to PARAMETER_VALUE. - */ - void apply(); - - /** - * @brief catchUpWith The Regularization will be delayed in many situations( - * pserver, local sparse). Catch Up means catch the regularization up, apply - * regularization to all params. - */ - void catchUpWith(); - - private: - ParameterUpdaterPrivate* m; -}; - -struct EvaluatorPrivate; -class Evaluator { - private: - Evaluator(); - DISABLE_COPY(Evaluator); - - public: - ~Evaluator(); - - /** - * @brief begin an evaluate stage. - */ - void start(); - - /** - * @brief end an evaluate stage. - */ - void finish(); - - /** - * @brief toString will get a evaluate result. - * - * __repr__ method in python - */ - std::string toString(); - - std::vector getNames() const; - - double getValue(const std::string name) const; - - private: - EvaluatorPrivate* m; - - friend class GradientMachine; -}; - -struct TrainerPrivate; -class Trainer { - private: - TrainerPrivate* m; - Trainer(); - Trainer(TrainerConfig* optConfig, GradientMachine* gm); - DISABLE_COPY(Trainer); - - public: - virtual ~Trainer(); - - /// Create A Trainer By TrainerConfig. using paddle command line. - static Trainer* createByCommandLine() throw(IOError); - - static Trainer* create(TrainerConfig* optConfig, - GradientMachine* gm) throw(IOError); - - /// Start training - void startTrain(); - - /// Finish training - void finishTrain(); - - /// Start a pass. - void startTrainPass(); - - /// Finish a pass - void finishTrainPass(); - - /** - * Train one batch, - * - * @return true if all batch finished. - */ - bool trainOneBatch(size_t batchSize); - - void trainOneDataBatch(size_t batchSize, const Arguments& args); - - void startTestPeriod(); - void testOneDataBatch(size_t batchSize, const Arguments& args); - void finishTestPeriod(); - - void forwardOneBatch(size_t batchSize); - - Arguments* getForwardOutput(); - - Arguments* getLayerOutput(const std::string& layerName) const; -}; - -/// the N-Best results generated from one input sequence. -class ISequenceResults { - public: - virtual ~ISequenceResults(); - - /// Number of result. - virtual size_t getSize() const = 0; - - /** - * Get sentence from dictionary. - * - * @param id the index of result. - * @param split if true, the return sentence will be splited with ' ' by - * each word. Default is false. - */ - virtual std::string getSentence(size_t id, bool split = false) const - throw(RangeError) = 0; - virtual std::vector getSequence(size_t id) const throw(RangeError) = 0; - virtual float getScore(size_t id) const throw(RangeError) = 0; -}; - -struct SequenceGeneratorPrivate; -class SequenceGenerator { - DISABLE_COPY(SequenceGenerator); - SequenceGenerator(); - - public: - virtual ~SequenceGenerator(); - - /** - * Generate Sequence by input. - * - * @note The inArgs is just one sequence of data. - * @note The return will get a N-best generate result by inArgs. - * Sort by score. - */ - ISequenceResults* generateSequence(const Arguments& inArgs) const; - - void setDict(const std::vector& dict); - void setBos(size_t bos); - void setEos(size_t eos); - void setMaxLength(size_t maxlength); - void setBeamSize(size_t beamSize); - - private: - static SequenceGenerator* createByGradientMachineSharedPtr(void* ptr); - friend class GradientMachine; - - private: - SequenceGeneratorPrivate* m; -}; diff --git a/paddle/api/PaddleAPIPrivate.h b/paddle/api/PaddleAPIPrivate.h deleted file mode 100644 index e141fcd761d7db2d3836a6343700ac4a7ca80c16..0000000000000000000000000000000000000000 --- a/paddle/api/PaddleAPIPrivate.h +++ /dev/null @@ -1,97 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#pragma once -#include -#include "PaddleAPI.h" -#include "paddle/gserver/evaluators/Evaluator.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" -#include "paddle/parameter/ParameterUpdaterBase.h" -#include "paddle/trainer/TrainerConfigHelper.h" - -struct GradientMachinePrivate { - std::shared_ptr machine; - - template - inline T& cast(void* ptr) { - return *(T*)(ptr); - } -}; - -struct OptimizationConfigPrivate { - std::shared_ptr trainer_config; - paddle::OptimizationConfig config; - - const paddle::OptimizationConfig& getConfig() { - if (trainer_config != nullptr) { - return trainer_config->getOptConfig(); - } else { - return config; - } - } -}; - -struct TrainerConfigPrivate { - std::shared_ptr conf; - TrainerConfigPrivate() {} -}; - -struct ModelConfigPrivate { - std::shared_ptr conf; -}; - -struct ArgumentsPrivate { - std::vector outputs; - - inline paddle::Argument& getArg(size_t idx) throw(RangeError) { - if (idx < outputs.size()) { - return outputs[idx]; - } else { - RangeError e; - throw e; - } - } - - template - std::shared_ptr& cast(void* rawPtr) const { - return *(std::shared_ptr*)(rawPtr); - } -}; - -struct ParameterUpdaterPrivate { - std::unique_ptr updater; -}; - -struct ParameterPrivate { - std::shared_ptr sharedPtr; - paddle::Parameter* rawPtr; // rawPtr only used in ParameterUpdater, - // in other situation sharedPtr should - // contains value. - - ParameterPrivate() : sharedPtr(nullptr), rawPtr(nullptr) {} - - paddle::Parameter* getPtr() { - if (sharedPtr) { - return sharedPtr.get(); - } else { - return rawPtr; - } - } -}; - -struct EvaluatorPrivate { - paddle::Evaluator* rawPtr; - - EvaluatorPrivate() : rawPtr(nullptr) {} - ~EvaluatorPrivate() { delete rawPtr; } -}; diff --git a/paddle/api/Parameter.cpp b/paddle/api/Parameter.cpp deleted file mode 100644 index 589d22e74e742de2595a9efd17412ddc55159230..0000000000000000000000000000000000000000 --- a/paddle/api/Parameter.cpp +++ /dev/null @@ -1,68 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/parameter/Parameter.h" -#include "PaddleAPI.h" -#include "PaddleAPIPrivate.h" - -Parameter::Parameter() : m(new ParameterPrivate()) {} - -Parameter::~Parameter() { delete m; } - -Parameter* Parameter::createFromRawPtr(void* ptr) { - auto p = new Parameter(); - p->m->rawPtr = *static_cast(ptr); - return p; -} - -Parameter* Parameter::createFromSharedPtr(void* ptr) { - auto& p = *(paddle::ParameterPtr*)(ptr); - if (p == nullptr) { - return nullptr; - } else { - auto retParam = new Parameter(); - retParam->m->sharedPtr = p; - return retParam; - } -} - -std::string Parameter::getName() const { return m->getPtr()->getName(); } - -Vector* Parameter::getBuf(ParameterType type) { - auto buf = m->getPtr()->getBuf(type); - return Vector::createByPaddleVectorPtr(&buf); -} - -ParameterConfig* Parameter::getConfig() { - if (m->sharedPtr) { - return ParameterConfig::createParameterConfigFromParameterSharedPtr( - &m->sharedPtr); - } else { - return ParameterConfig::createParameterConfigFromParameterPtr(m->rawPtr); - } -} - -size_t Parameter::getID() const { return m->getPtr()->getID(); } - -void Parameter::setValueUpdated() { m->getPtr()->setValueUpdated(); } - -bool Parameter::save(const std::string& filename) const { - return m->getPtr()->save(filename); -} - -bool Parameter::load(const std::string& filename) const { - return m->getPtr()->load(filename); -} - -size_t Parameter::getSize() const { return m->getPtr()->getSize(); } diff --git a/paddle/api/ParameterOptimizer.cpp b/paddle/api/ParameterOptimizer.cpp deleted file mode 100644 index d4620be3e6f26cdd4caffffac712e4ef936b222a..0000000000000000000000000000000000000000 --- a/paddle/api/ParameterOptimizer.cpp +++ /dev/null @@ -1,124 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/parameter/ParameterOptimizer.h" -#include -#include "Internal.h" -#include "PaddleAPI.h" -#include "PaddleAPIPrivate.h" - -struct ParameterOptimizerPrivate { - std::unique_ptr optimizer; -}; - -struct ParameterTraverseCallbackPrivate { - paddle::ParameterOptimizer::TraverseCallback callback; - - ParameterTraverseCallbackPrivate() {} - - ParameterTraverseCallbackPrivate( - const paddle::ParameterOptimizer::TraverseCallback& callback) - : callback(callback) {} - - void apply(const std::vector& vecs, - const ParameterConfig& conf, - size_t sparseId) { - std::vector real_vecs; - real_vecs.resize(vecs.size()); - std::transform(vecs.begin(), vecs.end(), real_vecs.begin(), [](Vector* v) { - if (v) { - return *(paddle::VectorPtr*)(v->getSharedPtr()); - } else { - return paddle::VectorPtr(); - } - }); - - paddle::ParameterConfig& real_conf = - *(paddle::ParameterConfig*)(const_cast(conf) - .getRawPtr()); - callback(real_vecs.data(), real_conf, sparseId); - } -}; - -ParameterOptimizer::ParameterOptimizer() : m(new ParameterOptimizerPrivate()) {} - -ParameterOptimizer::~ParameterOptimizer() { delete m; } - -ParameterOptimizer* ParameterOptimizer::create(OptimizationConfig* config) { - CHECK(config != nullptr); - auto retOptimizer = new ParameterOptimizer(); - retOptimizer->m->optimizer.reset( - paddle::ParameterOptimizer::create(config->m->getConfig(), false)); - return retOptimizer; -} - -void ParameterOptimizer::init(size_t numRows, const ParameterConfig* config) { - auto& conf = *(paddle::ParameterConfig*)(const_cast(config) - ->getRawPtr()); - m->optimizer->init(numRows, &conf); -} - -void ParameterOptimizer::startPass() { m->optimizer->startPass(); } - -void ParameterOptimizer::finishPass() { m->optimizer->finishPass(); } - -void ParameterOptimizer::startBatch(size_t numSamplesProcessed) { - constexpr size_t high_1 = 1UL << (sizeof(size_t) * 8 - 1); - CHECK_EQ(numSamplesProcessed & high_1, 0UL); // Safely cast. - m->optimizer->startBatch((int64_t)numSamplesProcessed); -} - -void ParameterOptimizer::finishBatch() { m->optimizer->finishBatch(); } - -void ParameterOptimizer::update(const std::vector& vecs, - const ParameterConfig& conf, - size_t sparseId) { - ParameterTraverseCallbackPrivate invoker( - [&](const paddle::VectorPtr _vecs[], - const paddle::ParameterConfig& config, - size_t sid = -1UL) { m->optimizer->update(_vecs, config, sid); }); - invoker.apply(vecs, conf, sparseId); -} - -std::vector ParameterOptimizer::getParameterTypes() const { - std::vector returnValue; - staticCastVector(&returnValue, m->optimizer->getParameterTypes()); - return returnValue; -} - -ParameterTraverseCallback::ParameterTraverseCallback() - : m(new ParameterTraverseCallbackPrivate()) {} - -ParameterTraverseCallback::~ParameterTraverseCallback() { delete m; } - -void ParameterTraverseCallback::apply(const std::vector& vecs, - const ParameterConfig& conf, - size_t sparseId) { - m->apply(vecs, conf, sparseId); -} - -ParameterTraverseCallback* ParameterOptimizer::needSpecialTraversal( - const ParameterConfig& config) const { - auto& param_config = - *(paddle::ParameterConfig*)const_cast(config) - .getRawPtr(); - auto callback = m->optimizer->needSpecialTraversal(param_config); - if (callback) { - auto retCallback = new ParameterTraverseCallback(); - retCallback->m->callback = callback; - return retCallback; - } else { - return nullptr; - } -} diff --git a/paddle/api/ParameterUpdater.cpp b/paddle/api/ParameterUpdater.cpp deleted file mode 100644 index 63c000c959f67dc682190b73bac24640ca8d0682..0000000000000000000000000000000000000000 --- a/paddle/api/ParameterUpdater.cpp +++ /dev/null @@ -1,99 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" - -#include "PaddleAPIPrivate.h" -#ifndef PADDLE_WITHOUT_GOLANG -#include "paddle/trainer/NewRemoteParameterUpdater.h" -#endif -#include "paddle/trainer/RemoteParameterUpdater.h" -#include "paddle/trainer/ThreadParameterUpdater.h" - -ParameterUpdater::ParameterUpdater() : m(new ParameterUpdaterPrivate()) {} - -ParameterUpdater *ParameterUpdater::createLocalUpdater( - OptimizationConfig *config) { - auto updater = new ParameterUpdater(); - updater->m->updater.reset( - new paddle::SgdThreadUpdater(config->m->getConfig())); - return updater; -} - -ParameterUpdater *ParameterUpdater::createNewRemoteUpdater( - OptimizationConfig *config, - const std::string pserverSpec, - const bool useEtcd) throw(UnsupportError) { -#ifndef PADDLE_WITHOUT_GOLANG - auto updater = new ParameterUpdater(); - updater->m->updater.reset(new paddle::NewRemoteParameterUpdater( - config->m->getConfig(), pserverSpec, useEtcd)); - return updater; -#else - throw UnsupportError("not compiled with WITH_GOLANG"); -#endif -} - -ParameterUpdater *ParameterUpdater::createRemoteUpdater( - OptimizationConfig *config, int passCount, bool useSparseUpdater) { - auto updater = new ParameterUpdater(); - auto remoteUpdater = new paddle::RemoteParameterUpdater( - config->m->getConfig(), passCount, nullptr); - if (useSparseUpdater) { - std::unique_ptr remoteUpdaterPtr(remoteUpdater); - auto sparseRemoteUpdater = - new paddle::SparseRemoteParameterUpdaterComposite( - config->m->getConfig(), - passCount, - false, - std::move(remoteUpdaterPtr)); - updater->m->updater.reset(sparseRemoteUpdater); - } else { - updater->m->updater.reset(remoteUpdater); - } - return updater; -} - -ParameterUpdater::~ParameterUpdater() { delete m; } - -void ParameterUpdater::init(const GradientMachine &gm) { - m->updater->init(gm.m->machine->getNonStaticParameters()); -} - -void ParameterUpdater::startPass() { m->updater->startPass(); } - -void ParameterUpdater::finishPass() { m->updater->finishPass(); } - -PassType ParameterUpdater::startBatch(size_t batchSize) { - return m->updater->startBatch((int64_t)batchSize); -} - -void ParameterUpdater::finishBatch(float cost) { - m->updater->finishBatch(cost); -} - -void ParameterUpdater::update(Parameter *param) { - auto paddleParam = param->m->getPtr(); - m->updater->update(paddleParam); -} - -void ParameterUpdater::getParametersRemote(bool fullSize, bool apply) { - m->updater->getParametersRemote(fullSize, apply); -} - -void ParameterUpdater::restore() { m->updater->restore(); } - -void ParameterUpdater::apply() { m->updater->apply(); } - -void ParameterUpdater::catchUpWith() { m->updater->catchUpWith(); } diff --git a/paddle/api/SequenceGenerator.cpp b/paddle/api/SequenceGenerator.cpp deleted file mode 100644 index 1446c3084238859a759669f3a32c7efde67dcc2b..0000000000000000000000000000000000000000 --- a/paddle/api/SequenceGenerator.cpp +++ /dev/null @@ -1,242 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include "PaddleAPI.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" -#include "paddle/parameter/Argument.h" -#include "paddle/utils/Flags.h" - -// used to represent partial sequence -struct Path { - std::vector ids; - float logProb; - paddle::MachineState machineState; - - Path() { logProb = 0; } - - Path(std::vector& ids, float logProb, paddle::MachineState& machineState) - : ids(ids), logProb(logProb), machineState(machineState) {} - - bool operator<(const Path& other) const { return (logProb > other.logProb); } -}; - -// Return top k (k == beam_size) optimal paths using beam search. The last -// element of inArgs is the Argument of feedback. gradMachine has MaxIdLayer -// as output and outArgs thus stores top k labels and their probabilities per -// position -static void findNBest(paddle::GradientMachine* gradMachine, - std::vector& inArgs, - std::vector& finalPaths, - size_t bos_id, - size_t eos_id, - size_t max_length) { - std::vector paths; - Path emptyPath; - paths.push_back(emptyPath); - finalPaths.clear(); - gradMachine->resetState(); - paddle::Argument feedback = inArgs.back(); - feedback.ids->setElement(0, (int)(bos_id)); - float minFinalPathLogProb = 0; - size_t beam = 0; - int id; - std::vector outArgs; - while (true) { // iterate over each generated word - std::vector newPaths; - paddle::MachineState machineState; - for (size_t j = 0; j < paths.size(); j++) { - Path& path = paths[j]; - if (path.machineState.size() > 0) { - gradMachine->setState(path.machineState); - feedback.ids->setElement(0, path.ids.back()); - } - gradMachine->forward(inArgs, &outArgs, paddle::PASS_TEST); - gradMachine->getState(machineState); - beam = outArgs[0].ids->getSize(); - for (size_t k = 0; k < beam; k++) { - id = outArgs[0].ids->getElement(k); - float prob = outArgs[0].in->getElement(0, k); - std::vector nids(path.ids); - nids.push_back(id); - float newLogProb = path.logProb + log(prob); - Path newPath(nids, newLogProb, machineState); - if (id == (int)eos_id || nids.size() >= max_length) { - finalPaths.push_back(newPath); - if (minFinalPathLogProb > newPath.logProb) { - minFinalPathLogProb = newPath.logProb; - } - } else { - newPaths.push_back(newPath); - } - } - } - - if (newPaths.size() == 0) { - break; - } - std::nth_element(newPaths.begin(), - newPaths.begin() + std::min(beam, newPaths.size()), - newPaths.end()); - if (newPaths.size() > beam) { - newPaths.resize(beam); - } - // pathA < pathB means pathA.logProb > pathB.logProb - float maxPathLogProb = - std::min_element(newPaths.begin(), newPaths.end())->logProb; - if (finalPaths.size() >= beam && minFinalPathLogProb >= maxPathLogProb) { - break; - } - paths = newPaths; - } // end while - - std::partial_sort(finalPaths.begin(), - finalPaths.begin() + std::min(beam, finalPaths.size()), - finalPaths.end()); - if (finalPaths.size() > beam) { - finalPaths.resize(beam); - } -} - -struct SequenceGeneratorPrivate { - std::shared_ptr machine; - std::shared_ptr> dict; - size_t beginPos; - size_t endPos; - size_t maxLength; - - paddle::Argument feedback; - - template - inline T& cast(void* ptr) { - return *(T*)(ptr); - } - - inline void findNBest(std::vector& inArgs, - std::vector& path) { - ::findNBest(machine.get(), inArgs, path, beginPos, endPos, maxLength); - } - - SequenceGeneratorPrivate() - : dict(std::make_shared>()), - beginPos(0UL), - endPos(0UL), - maxLength(0UL), - feedback(__create_feedback__()) {} - - private: - static paddle::Argument __create_feedback__() { - paddle::Argument feedback; - feedback.ids = paddle::IVector::create(/* size= */ 1, FLAGS_use_gpu); - - feedback.sequenceStartPositions = - paddle::ICpuGpuVector::create(/* size= */ 2, /* useGpu= */ false); - feedback.sequenceStartPositions->getMutableData(false)[0] = 0; - feedback.sequenceStartPositions->getMutableData(false)[1] = 1; - return feedback; - } -}; - -SequenceGenerator::SequenceGenerator() : m(new SequenceGeneratorPrivate()) {} - -SequenceGenerator::~SequenceGenerator() { delete m; } - -class PathSequenceResults : public ISequenceResults { - // ISequenceResults interface - public: - PathSequenceResults(const std::shared_ptr>& path, - const std::shared_ptr>& dict) - : path_(path), dict_(dict) {} - - size_t getSize() const { return path_->size(); } - std::string getSentence(size_t id, bool split) const throw(RangeError) { - if (id < getSize()) { - Path& p = (*path_)[id]; - std::ostringstream sout; - std::transform(p.ids.begin(), - p.ids.end(), - std::ostream_iterator(sout, split ? " " : ""), - [&](int id) { return (*dict_)[id]; }); - return sout.str(); - } else { - RangeError e; - throw e; - } - } - std::vector getSequence(size_t id) const throw(RangeError) { - if (id < getSize()) { - Path& p = (*path_)[id]; - return p.ids; - } else { - RangeError e; - throw e; - } - } - float getScore(size_t id) const throw(RangeError) { - if (id < getSize()) { - Path& p = (*path_)[id]; - return p.logProb; - } else { - RangeError e; - throw e; - } - } - - private: - std::shared_ptr> path_; - std::shared_ptr> dict_; -}; - -ISequenceResults* SequenceGenerator::generateSequence( - const Arguments& inArgs) const { - auto& in_args = - m->cast>(inArgs.getInternalArgumentsPtr()); - for (auto& arg : in_args) { - arg.sequenceStartPositions = m->feedback.sequenceStartPositions; - } - in_args.push_back(m->feedback); - auto path = std::make_shared>(); - m->findNBest(in_args, *path); - return new PathSequenceResults(path, m->dict); -} - -SequenceGenerator* SequenceGenerator::createByGradientMachineSharedPtr( - void* ptr) { - SequenceGenerator* r = new SequenceGenerator(); - r->m->machine = r->m->cast>(ptr); - return r; -} - -void SequenceGenerator::setDict(const std::vector& dict) { - *m->dict = dict; -} - -void SequenceGenerator::setBos(size_t bos) { m->beginPos = bos; } - -void SequenceGenerator::setEos(size_t eos) { m->endPos = eos; } - -void SequenceGenerator::setMaxLength(size_t maxLength) { - m->maxLength = maxLength; -} - -void SequenceGenerator::setBeamSize(size_t beamSize) { - if (beamSize != -1UL) { - FLAGS_beam_size = beamSize; - } -} - -ISequenceResults::~ISequenceResults() {} diff --git a/paddle/api/Trainer.cpp b/paddle/api/Trainer.cpp deleted file mode 100644 index 795460b65051b4ec0d9772d2503f123c4a6ea3d0..0000000000000000000000000000000000000000 --- a/paddle/api/Trainer.cpp +++ /dev/null @@ -1,175 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" -#include "PaddleAPIPrivate.h" - -#include -#include -#include - -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/trainer/ParamUtil.h" -#include "paddle/trainer/Trainer.h" -#include "paddle/trainer/TrainerInternal.h" -#include "paddle/utils/Flags.h" - -using paddle::real; - -DECLARE_string(config); -DECLARE_string(init_model_path); -DECLARE_int32(start_pass); - -struct TrainerPrivate : public paddle::Trainer { - bool _trainOneBatch(size_t batchSize); - bool forwardOneBatch(size_t batchSize); - void forwardOneDataBatch(const std::vector& inArgs); - void setBatchSize(size_t batchSize); - std::vector& getForwardOutput(); - - void startTestPeriod(); - void finishTestPeriod(); - void testOneDataBatch(const paddle::DataBatch& dataBatch); - TrainerPrivate() : paddle::Trainer() {} -}; - -Trainer::Trainer() : m(new TrainerPrivate()) { - auto conf = paddle::TrainerConfigHelper::createFromFlags(); - if (conf != nullptr) { - m->init(conf); - } -} - -Trainer::~Trainer() { delete m; } - -Trainer* Trainer::createByCommandLine() throw(IOError) { - auto retv = new Trainer(); - if (retv->m->getConfig().IsInitialized()) { - return retv; - } else { - throw IOError(); - } -} - -Trainer::Trainer(TrainerConfig* config, GradientMachine* gm) - : m(new TrainerPrivate()) { - m->init(config->m->conf, /* testing= */ false, gm ? gm->m->machine : nullptr); -} - -Trainer* Trainer::create(TrainerConfig* config, - GradientMachine* gm) throw(IOError) { - auto retv = new Trainer(config, gm); - if (retv->m->getConfig().IsInitialized()) { - return retv; - } else { - retv->m->getConfig().CheckInitialized(); - throw IOError(); - } -} - -void Trainer::startTrain() { m->startTrain(); } - -void Trainer::finishTrain() { m->finishTrain(); } - -void Trainer::startTrainPass() { m->startTrainPass(); } - -void Trainer::finishTrainPass() { m->finishTrainPass(); } - -void Trainer::trainOneDataBatch(size_t batchSize, const Arguments& inArgs) { - paddle::DataBatch dataBatch; - dataBatch.getStreams() = inArgs.m->outputs; - dataBatch.setSize(batchSize); - m->trainOneDataBatch(dataBatch); -} - -bool Trainer::trainOneBatch(size_t batchSize) { - return m->_trainOneBatch(batchSize); -} - -bool TrainerPrivate::_trainOneBatch(size_t batchSize) { - paddle::DataBatch dataBatch; - CHECK(dataProvider_) << "data_provider is not specified"; - int num = dataProvider_->getNextBatch(batchSize, &dataBatch); - if (num == 0) { - return false; - } - trainOneDataBatch(dataBatch); - return false; -} - -void TrainerPrivate::startTestPeriod() { - if (!tester_) { - createTester(); - } - tester_->startTestPeriod(); -} - -void Trainer::startTestPeriod() { m->startTestPeriod(); } - -void TrainerPrivate::testOneDataBatch(const paddle::DataBatch& dataBatch) { - tester_->testOneDataBatch(dataBatch, &forwardOutput_); -} - -void Trainer::testOneDataBatch(size_t batchSize, const Arguments& args) { - paddle::DataBatch dataBatch; - dataBatch.getStreams() = args.m->outputs; - dataBatch.setSize(batchSize); - m->testOneDataBatch(dataBatch); -} - -void TrainerPrivate::finishTestPeriod() { tester_->finishTestPeriod(); } -void Trainer::finishTestPeriod() { m->finishTestPeriod(); } - -Arguments* Trainer::getLayerOutput(const std::string& layerName) const { - auto nn = this->m->getGradientMachine(); - CHECK(nn) << "trainerInternal_.getGradientMachine() is not NeuralNetwork"; - auto arg = nn->getLayerOutput(layerName); - return Arguments::createByPaddleArgument(&arg); -} - -void Trainer::forwardOneBatch(size_t batchSize) { - m->forwardOneBatch(batchSize); -} - -bool TrainerPrivate::forwardOneBatch(size_t batchSize) { - CHECK(dataProvider_) << "data_provider is not specified"; - paddle::DataBatch dataBatch; - int num = dataProvider_->getNextBatch(batchSize, &dataBatch); - if (num == 0) { - return false; - } - - forwardOneDataBatch(dataBatch.getStreams()); - return true; -} - -void TrainerPrivate::forwardOneDataBatch( - const std::vector& inArgs) { - std::vector& outArgs = forwardOutput_; - - if (config_->getOptConfig().use_sparse_remote_updater()) { - trainerInternal_.getGradientMachine()->prefetch(inArgs); - trainerInternal_.getParameterUpdater()->getParametersRemote(); - } - trainerInternal_.getGradientMachine()->forward( - inArgs, &outArgs, paddle::PASS_TEST); -} - -Arguments* Trainer::getForwardOutput() { - return Arguments::createByPaddleArgumentVector(&m->getForwardOutput()); -} - -std::vector& TrainerPrivate::getForwardOutput() { - return forwardOutput_; -} diff --git a/paddle/api/Util.cpp b/paddle/api/Util.cpp deleted file mode 100644 index 618e87e96459674302d8b468c3ac410e8d3af6a8..0000000000000000000000000000000000000000 --- a/paddle/api/Util.cpp +++ /dev/null @@ -1,60 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" - -#include "paddle/parameter/Parameter.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Util.h" - -#include -#include -#include - -void initPaddle(int argc, char** argv) { - paddle::initMain(argc, argv); - paddle::initPython(argc, argv); - feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); -} - -FloatArray::FloatArray(const float* b, const size_t l) - : buf(b), length(l), needFree(false) {} - -IntArray::IntArray(const int* b, const size_t l, bool f) - : buf(b), length(l), needFree(f) {} - -IntWithFloatArray::IntWithFloatArray(const float* v, - const int* i, - size_t l, - bool f) - : valBuf(v), idxBuf(i), length(l), needFree(f) {} - -bool isUsingGpu() { return FLAGS_use_gpu; } - -void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; } - -bool isGpuVersion() { -#ifndef PADDLE_WITH_CUDA - return false; -#else - return true; -#endif -} - -int getTrainerCount() { return FLAGS_trainer_count; } - -static_assert(NUM_PARAMETER_TYPES == paddle::NUM_PARAMETER_TYPES, - "The Parameter Type should be same in core/api and core/common"); diff --git a/paddle/api/Vector.cpp b/paddle/api/Vector.cpp deleted file mode 100644 index e2a7b974ca78ae3e6e0e66c206a40c8811126b53..0000000000000000000000000000000000000000 --- a/paddle/api/Vector.cpp +++ /dev/null @@ -1,304 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PaddleAPI.h" - -#include "paddle/math/Vector.h" - -#include - -struct IVectorPrivate { - paddle::IVectorPtr vec; -}; - -IVector::IVector() : m(new IVectorPrivate()) {} - -IVector* IVector::createZero(size_t sz, bool useGpu) { - auto v = new IVector(); - v->m->vec = paddle::IVector::create(sz, useGpu); - v->m->vec->zeroMem(); - return v; -} - -IVector* IVector::create(const std::vector& data, bool useGpu) { - auto v = new IVector(); - v->m->vec = paddle::IVector::create(data.size(), useGpu); - v->m->vec->copyFrom(data.data(), data.size()); - return v; -} - -IVector* IVector::createVectorFromNumpy(int* data, - int dim, - bool copy, - bool useGpu) throw(UnsupportError) { - if (useGpu) { - /// if use gpu only copy=true is supported - if (!copy) { - throw UnsupportError("Gpu mode only supports copy=True"); - } - return IVector::createGpuVectorFromNumpy(data, dim); - } else { - return IVector::createCpuVectorFromNumpy(data, dim, copy); - } -} - -IVector* IVector::createCpuVectorFromNumpy(int* data, int dim, bool copy) { - auto v = new IVector(); - if (copy) { - v->m->vec = paddle::IVector::create(dim, false); - v->m->vec->copyFrom(data, dim); - } else { - v->m->vec = paddle::IVector::create(data, dim, false); - } - return v; -} - -IVector* IVector::createGpuVectorFromNumpy(int* data, int dim) { - auto v = new IVector(); - v->m->vec = paddle::IVector::create(dim, true); - v->m->vec->copyFrom(data, dim); - return v; -} - -bool IVector::isGpu() const { - return dynamic_cast(m->vec.get()) != nullptr; -} - -IntArray IVector::getData() const { - if (this->isGpu()) { - int* src = m->vec->getData(); - size_t len = m->vec->getSize(); - int* dest = new int[len]; - hl_memcpy_device2host(dest, src, len * sizeof(int)); - return IntArray(dest, len, true); - } else { - return IntArray(m->vec->getData(), m->vec->getSize()); - } -} - -int& IVector::operator[](const size_t idx) throw(RangeError, UnsupportError) { - if (this->isGpu()) { - UnsupportError e; - throw e; - } else { - if (idx >= m->vec->getSize()) { - RangeError e; - throw e; - } - } - return m->vec->getData()[idx]; -} - -const int& IVector::operator[](const size_t idx) const - throw(RangeError, UnsupportError) { - return (*const_cast(this))[idx]; -} - -IVector* IVector::createByPaddleVectorPtr(void* ptr) { - auto* p = (paddle::IVectorPtr*)ptr; - if ((*p) != nullptr) { - IVector* vec = new IVector(); - vec->m->vec = *p; - return vec; - } else { - return nullptr; - } -} - -IVector::~IVector() { delete m; } - -void* IVector::getSharedPtr() const { return &m->vec; } - -size_t IVector::getSize() const { return m->vec->getSize(); } - -void IVector::toNumpyArrayInplace(int** data, int* dim1) throw(UnsupportError) { - auto v = std::dynamic_pointer_cast(m->vec); - if (v) { - *data = v->getData(); - *dim1 = v->getSize(); - } else { - throw UnsupportError(); - } -} - -void IVector::copyToNumpyArray(int** view_m_data, int* dim1) { - *dim1 = m->vec->getSize(); - *view_m_data = new int[*dim1]; - if (auto cpuVec = dynamic_cast(m->vec.get())) { - std::memcpy(*view_m_data, cpuVec->getData(), sizeof(int) * (*dim1)); - } else if (auto gpuVec = dynamic_cast(m->vec.get())) { - hl_memcpy_device2host( - *view_m_data, gpuVec->getData(), sizeof(int) * (*dim1)); - } else { - LOG(INFO) << "Unexpected situation"; - } -} - -void IVector::copyFromNumpyArray(int* data, int dim) { - m->vec->resize(dim); - m->vec->copyFrom(data, dim); -} - -struct VectorPrivate { - paddle::VectorPtr vec; - - void safeAccessData(const size_t idx, - const std::function& func) const - throw(RangeError, UnsupportError) { - auto cpuVec = std::dynamic_pointer_cast(vec); - if (cpuVec != nullptr) { - if (idx < vec->getSize()) { - func(vec->getData()[idx]); - } else { - throw RangeError(); - } - } else { - throw UnsupportError(); - } - } -}; - -Vector::Vector() : m(new VectorPrivate()) {} - -Vector::~Vector() { delete m; } - -Vector* Vector::createZero(size_t sz, bool useGpu) { - auto retVec = new Vector(); - retVec->m->vec = paddle::Vector::create(sz, useGpu); - retVec->m->vec->zero(); - return retVec; -} - -Vector* Vector::create(const std::vector& data, bool useGpu) { - auto retVec = new Vector(); - retVec->m->vec = paddle::Vector::create(data.size(), useGpu); - retVec->m->vec->copyFrom(data.data(), data.size()); - return retVec; -} - -Vector* Vector::createByPaddleVectorPtr(void* ptr) { - auto& v = *(paddle::VectorPtr*)(ptr); - if (v == nullptr) { - return nullptr; - } else { - auto retVec = new Vector(); - retVec->m->vec = v; - return retVec; - } -} - -Vector* Vector::createVectorFromNumpy(float* data, - int dim, - bool copy, - bool useGpu) throw(UnsupportError) { - if (useGpu) { - /// if use gpu only copy=True is supported - if (!copy) { - throw UnsupportError("Gpu mode only supports copy=True"); - } - return Vector::createGpuVectorFromNumpy(data, dim); - } else { - return Vector::createCpuVectorFromNumpy(data, dim, copy); - } -} - -Vector* Vector::createCpuVectorFromNumpy(float* data, int dim, bool copy) { - CHECK_GT(dim, 0); - auto retVec = new Vector(); - if (copy) { - retVec->m->vec = paddle::Vector::create((size_t)dim, false); - retVec->m->vec->copyFrom(data, dim); - } else { - retVec->m->vec = paddle::Vector::create(data, (size_t)dim, false); - } - return retVec; -} - -Vector* Vector::createGpuVectorFromNumpy(float* data, int dim) { - CHECK_GT(dim, 0); - auto retVec = new Vector(); - retVec->m->vec = paddle::Vector::create((size_t)dim, true); - retVec->m->vec->copyFrom(data, (size_t)dim); - return retVec; -} - -void Vector::toNumpyArrayInplace(float** view_data, - int* dim1) throw(UnsupportError) { - auto v = std::dynamic_pointer_cast(m->vec); - if (v != nullptr) { - *view_data = v->getData(); - *dim1 = (int)v->getSize(); - } else { - throw UnsupportError(); - } -} - -void Vector::copyToNumpyArray(float** view_m_data, int* dim1) { - *dim1 = m->vec->getSize(); - *view_m_data = new float[*dim1]; - if (auto cpuVec = dynamic_cast(m->vec.get())) { - std::memcpy(*view_m_data, cpuVec->getData(), sizeof(float) * (*dim1)); - } else if (auto gpuVec = dynamic_cast(m->vec.get())) { - hl_memcpy_device2host( - *view_m_data, gpuVec->getData(), sizeof(float) * (*dim1)); - } else { - LOG(INFO) << "Unexpected situation"; - } -} - -void Vector::copyFromNumpyArray(float* data, int dim) { - m->vec->resize(dim); - m->vec->copyFrom(data, dim); -} - -FloatArray Vector::getData() const { - if (this->isGpu()) { - float* src = m->vec->getData(); - size_t len = m->vec->getSize(); - float* dest = new float[len]; - hl_memcpy_device2host(dest, src, len * sizeof(float)); - FloatArray ret_val(dest, len); - ret_val.needFree = true; - return ret_val; - } else { - FloatArray ret_val(m->vec->getData(), m->vec->getSize()); - return ret_val; - } -} - -void Vector::copyFrom(Vector* src) throw(RangeError) { - if (src->m->vec->getSize() != m->vec->getSize()) { - throw RangeError(); - } - m->vec->copyFrom(*src->m->vec); -} - -bool Vector::isGpu() const { - return std::dynamic_pointer_cast(m->vec) != nullptr; -} - -float Vector::get(const size_t idx) const throw(RangeError, UnsupportError) { - float r; - m->safeAccessData(idx, [&](float& o) { r = o; }); - return r; -} - -void Vector::set(const size_t idx, float val) throw(RangeError, - UnsupportError) { - m->safeAccessData(idx, [&](float& o) { o = val; }); -} - -size_t Vector::getSize() const { return m->vec->getSize(); } - -void* Vector::getSharedPtr() { return &m->vec; } diff --git a/paddle/capi/Arguments.cpp b/paddle/capi/Arguments.cpp deleted file mode 100644 index 87fac3d6c6abe37b128213d4ffd66f8c1573a910..0000000000000000000000000000000000000000 --- a/paddle/capi/Arguments.cpp +++ /dev/null @@ -1,129 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "arguments.h" -#include "capi_private.h" - -using paddle::capi::cast; - -#define castArg(v) cast(v) -#define castIVec(v) cast(v) - -extern "C" { -paddle_arguments paddle_arguments_create_none() { - return new paddle::capi::CArguments(); -} - -paddle_error paddle_arguments_destroy(paddle_arguments args) { - if (args == nullptr) return kPD_NULLPTR; - delete castArg(args); - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_get_size(paddle_arguments args, uint64_t* size) { - if (args == nullptr || size == nullptr) return kPD_NULLPTR; - *size = castArg(args)->args.size(); - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_resize(paddle_arguments args, uint64_t size) { - if (args == nullptr) return kPD_NULLPTR; - castArg(args)->args.resize(size); - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_set_value(paddle_arguments args, - uint64_t ID, - paddle_matrix mat) { - if (args == nullptr || mat == nullptr) return kPD_NULLPTR; - auto m = paddle::capi::cast(mat); - if (m->mat == nullptr) return kPD_NULLPTR; - auto a = castArg(args); - if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; - a->args[ID].value = m->mat; - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_get_value(paddle_arguments args, - uint64_t ID, - paddle_matrix mat) { - if (args == nullptr || mat == nullptr) return kPD_NULLPTR; - auto m = paddle::capi::cast(mat); - auto a = castArg(args); - if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; - m->mat = a->args[ID].value; - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_get_ids(paddle_arguments args, - uint64_t ID, - paddle_ivector ids) { - if (args == nullptr || ids == nullptr) return kPD_NULLPTR; - auto iv = castIVec(ids); - auto a = castArg(args); - if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; - iv->vec = a->args[ID].ids; - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_set_ids(paddle_arguments args, - uint64_t ID, - paddle_ivector ids) { - //! TODO(lizhao): Complete this method. - if (args == nullptr || ids == nullptr) return kPD_NULLPTR; - auto iv = paddle::capi::cast(ids); - if (iv->vec == nullptr) return kPD_NULLPTR; - auto a = castArg(args); - if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; - a->args[ID].ids = iv->vec; - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_set_frame_shape(paddle_arguments args, - uint64_t ID, - uint64_t frameHeight, - uint64_t frameWidth) { - if (args == nullptr) return kPD_NULLPTR; - auto a = castArg(args); - if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; - a->args[ID].setFrameHeight(frameHeight); - a->args[ID].setFrameWidth(frameWidth); - return kPD_NO_ERROR; -} - -paddle_error paddle_arguments_set_sequence_start_pos(paddle_arguments args, - uint64_t ID, - uint32_t nestedLevel, - paddle_ivector seqPos) { - if (args == nullptr || seqPos == nullptr) return kPD_NULLPTR; - auto iv = paddle::capi::cast(seqPos); - if (iv->vec == nullptr) return kPD_NULLPTR; - auto a = castArg(args); - return a->accessSeqPos(ID, nestedLevel, [&iv](paddle::ICpuGpuVectorPtr& ptr) { - ptr = std::make_shared(iv->vec); - }); -} - -paddle_error paddle_arguments_get_sequence_start_pos(paddle_arguments args, - uint64_t ID, - uint32_t nestedLevel, - paddle_ivector seqPos) { - if (args == nullptr || seqPos == nullptr) return kPD_NULLPTR; - auto iv = paddle::capi::cast(seqPos); - auto a = castArg(args); - return a->accessSeqPos(ID, nestedLevel, [&iv](paddle::ICpuGpuVectorPtr& ptr) { - iv->vec = ptr->getMutableVector(false); - }); -} -} diff --git a/paddle/capi/Main.cpp b/paddle/capi/Main.cpp deleted file mode 100644 index 0a289dede65406facf1f1cba584f4330f2569214..0000000000000000000000000000000000000000 --- a/paddle/capi/Main.cpp +++ /dev/null @@ -1,53 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include "capi_private.h" -#include "main.h" -#include "paddle/trainer/TrainerConfigHelper.h" -#include "paddle/utils/Excepts.h" -#include "paddle/utils/PythonUtil.h" - -static void initPaddle(int argc, char** argv) { - paddle::initMain(argc, argv); - paddle::initPython(argc, argv); -} - -extern "C" { -paddle_error paddle_init(int argc, char** argv) { - static bool isInit = false; - if (isInit) return kPD_NO_ERROR; - - std::vector realArgv; - realArgv.reserve(argc + 1); - realArgv.push_back(strdup("")); - for (int i = 0; i < argc; ++i) { - realArgv.push_back(argv[i]); - } - initPaddle(argc + 1, realArgv.data()); - free(realArgv[0]); - isInit = true; - return kPD_NO_ERROR; -} - -paddle_error paddle_init_thread() { - if (FLAGS_use_gpu) { - hl_init(FLAGS_gpu_id); - } - return kPD_NO_ERROR; -} -} diff --git a/paddle/capi/arguments.h b/paddle/capi/arguments.h deleted file mode 100644 index 69a66bb012c318bc8317c246d690a7f4baffd248..0000000000000000000000000000000000000000 --- a/paddle/capi/arguments.h +++ /dev/null @@ -1,159 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifndef __PADDLE_CAPI_ARGUMENTS_H__ -#define __PADDLE_CAPI_ARGUMENTS_H__ - -#include -#include "config.h" -#include "error.h" -#include "matrix.h" -#include "vector.h" - -/** - * Arguments functions. Each argument means layer output. Arguments means a - * array of arguemnt. - */ -typedef void* paddle_arguments; - -#ifdef __cplusplus -extern "C" { -#endif - -/** - * @brief paddle_arguments_create_none Create a array of arguments, which size - * is zero. - * @return Arguemnts - */ -PD_API paddle_arguments paddle_arguments_create_none(); - -/** - * @brief paddle_arguments_destroy Destroy the arguments - * @param args arguments to destroy - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_destroy(paddle_arguments args); - -/** - * @brief paddle_arguments_get_size Get size of arguments array - * @param [in] args arguments array - * @param [out] size array size - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_get_size(paddle_arguments args, - uint64_t* size); - -/** - * @brief PDArgsResize Resize a arguments array. - * @param args arguments array. - * @param size target size of array - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_resize(paddle_arguments args, - uint64_t size); - -/** - * @brief PDArgsSetValue Set value matrix of one argument in array, which index - * is `ID`. - * @param args arguments array - * @param ID array index - * @param mat matrix pointer - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_set_value(paddle_arguments args, - uint64_t ID, - paddle_matrix mat); - -/** - * @brief PDArgsGetValue Get value matrix of one argument in array, which index - * is `ID`. - * @param [in] args arguments array - * @param [in] ID array index - * @param [out] mat matrix pointer - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_get_value(paddle_arguments args, - uint64_t ID, - paddle_matrix mat); - -/** - * @brief PDArgsGetIds Get the integer vector of one argument in array, which - * index is `ID`. - * @param args arguments array - * @param ID array index - * @param ids integer vector pointer - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_get_ids(paddle_arguments args, - uint64_t ID, - paddle_ivector ids); - -/** - * @brief PDArgsSetIds Set the integer vector of one argument in array, which - * index is `ID`. - * @param [in] args arguments array - * @param [in] ID array index - * @param [out] ids integer vector pointer - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_set_ids(paddle_arguments args, - uint64_t ID, - paddle_ivector ids); - -/** - * @brief paddle_arguments_set_frame_shape Set the fram size of one argument - * in array, which index is `ID`. - * @param [in] args arguments array - * @param [in] ID array index - * @param [in] frameHeight maximum height of input images - * @param [in] frameWidth maximum width of input images - * @return paddle_error - */ -PD_API paddle_error paddle_arguments_set_frame_shape(paddle_arguments args, - uint64_t ID, - uint64_t frameHeight, - uint64_t frameWidth); - -/** - * @brief PDArgsSetSequenceStartPos Set sequence start position vector of one - * argument in array, which index is `ID`. - * @param args arguments array - * @param ID array index - * @param seqPos sequence position array. - * @return paddle_error - */ -PD_API paddle_error -paddle_arguments_set_sequence_start_pos(paddle_arguments args, - uint64_t ID, - uint32_t nestedLevel, - paddle_ivector seqPos); -/** - * @brief PDArgsGetSequenceStartPos Get sequence start position vector of one - * argument in array, which index is `ID`. - * @param [in] args arguments array - * @param [in] ID array index - * @param [out] seqPos sequence position array - * @return paddle_error - */ -PD_API paddle_error -paddle_arguments_get_sequence_start_pos(paddle_arguments args, - uint64_t ID, - uint32_t nestedLevel, - paddle_ivector seqPos); - -#ifdef __cplusplus -} -#endif - -#endif diff --git a/paddle/capi/capi_private.h b/paddle/capi/capi_private.h deleted file mode 100644 index 3332f42a4a6e57fed6ddb20cf7d759d67e7240b5..0000000000000000000000000000000000000000 --- a/paddle/capi/capi_private.h +++ /dev/null @@ -1,82 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "capi.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/parameter/Argument.h" -#pragma once - -namespace paddle { -namespace capi { - -enum CType { kIVECTOR = 0, kMATRIX, kARGUMENTS, kGRADIENT_MACHINE }; - -#define STRUCT_HEADER CType type; - -struct CHeader { - STRUCT_HEADER -}; - -struct CIVector { - STRUCT_HEADER - IVectorPtr vec; - - CIVector() : type(kIVECTOR) {} -}; - -struct CMatrix { - STRUCT_HEADER - MatrixPtr mat; - - CMatrix() : type(kMATRIX) {} -}; - -struct CArguments { - STRUCT_HEADER - std::vector args; - - CArguments() : type(kARGUMENTS) {} - - template - paddle_error accessSeqPos(uint64_t ID, uint32_t nestedLevel, T callback) { - if (ID >= args.size()) return kPD_OUT_OF_RANGE; - switch (nestedLevel) { - case 0: - callback(args[ID].sequenceStartPositions); - break; - case 1: - callback(args[ID].subSequenceStartPositions); - break; - default: - return kPD_OUT_OF_RANGE; - } - return kPD_NO_ERROR; - } -}; - -struct CGradientMachine { - STRUCT_HEADER - paddle::GradientMachinePtr machine; - - CGradientMachine() : type(kGRADIENT_MACHINE) {} -}; - -template -inline T* cast(void* ptr) { - return reinterpret_cast(ptr); -} -} // namespace capi -} // namespace paddle diff --git a/paddle/capi/examples/model_inference/multi_thread/convert_protobin.sh b/paddle/capi/examples/model_inference/multi_thread/convert_protobin.sh deleted file mode 120000 index 3c1b3533523cf1709720d11df7b8e311e0577fe7..0000000000000000000000000000000000000000 --- a/paddle/capi/examples/model_inference/multi_thread/convert_protobin.sh +++ /dev/null @@ -1 +0,0 @@ -../dense/convert_protobin.sh \ No newline at end of file diff --git a/paddle/capi/examples/model_inference/sequence/convert_protobin.sh b/paddle/capi/examples/model_inference/sequence/convert_protobin.sh deleted file mode 120000 index 3c1b3533523cf1709720d11df7b8e311e0577fe7..0000000000000000000000000000000000000000 --- a/paddle/capi/examples/model_inference/sequence/convert_protobin.sh +++ /dev/null @@ -1 +0,0 @@ -../dense/convert_protobin.sh \ No newline at end of file diff --git a/paddle/capi/examples/model_inference/sparse_binary/convert_protobin.sh b/paddle/capi/examples/model_inference/sparse_binary/convert_protobin.sh deleted file mode 120000 index 3c1b3533523cf1709720d11df7b8e311e0577fe7..0000000000000000000000000000000000000000 --- a/paddle/capi/examples/model_inference/sparse_binary/convert_protobin.sh +++ /dev/null @@ -1 +0,0 @@ -../dense/convert_protobin.sh \ No newline at end of file diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp deleted file mode 100644 index 8c3f504e5a2d807c0cc664af486ebab4a82ddec3..0000000000000000000000000000000000000000 --- a/paddle/capi/gradient_machine.cpp +++ /dev/null @@ -1,180 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "gradient_machine.h" -#include "capi_private.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" - -#define cast(v) paddle::capi::cast(v) - -enum GradientMatchineCreateMode { - CREATE_MODE_NORMAL = 0, - CREATE_MODE_TESTING = 4 -}; - -namespace paddle { - -class MyNeuralNetwork : public NeuralNetwork { - public: - MyNeuralNetwork(const std::string& name, NeuralNetwork* network) - : NeuralNetwork(name, network) {} -}; - -NeuralNetwork* newCustomNerualNetwork(const std::string& name, - NeuralNetwork* network) { - return new MyNeuralNetwork(name, network); -} -} // namespace paddle - -extern "C" { -paddle_error paddle_gradient_machine_create_for_inference( - paddle_gradient_machine* machine, void* modelConfigProtobuf, int size) { - if (modelConfigProtobuf == nullptr) return kPD_NULLPTR; - paddle::ModelConfig config; - if (!config.ParseFromArray(modelConfigProtobuf, size) || - !config.IsInitialized()) { - return kPD_PROTOBUF_ERROR; - } - - auto ptr = new paddle::capi::CGradientMachine(); - ptr->machine.reset(paddle::GradientMachine::create( - config, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE})); - *machine = ptr; - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_create_for_inference_with_parameters( - paddle_gradient_machine* machine, void* mergedModel, uint64_t size) { - if (mergedModel == nullptr) return kPD_NULLPTR; - std::istringstream is(std::string(static_cast(mergedModel), size)); - int64_t modelConfigSize = 0; - is.read((char*)(&modelConfigSize), sizeof(modelConfigSize)); - std::string modelConfigProtobuf; - modelConfigProtobuf.resize(modelConfigSize); - is.read(&modelConfigProtobuf[0], modelConfigSize); - paddle::TrainerConfig config; - paddle::ModelConfig modelConfig; - if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) { - if (!modelConfig.ParseFromString(modelConfigProtobuf) || - !modelConfig.IsInitialized()) { - return kPD_PROTOBUF_ERROR; - } - } else { - modelConfig = config.model_config(); - } - auto ptr = new paddle::capi::CGradientMachine(); - ptr->machine.reset(paddle::GradientMachine::create( - modelConfig, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE})); - std::vector& parameters = ptr->machine->getParameters(); - for (auto& para : parameters) { - para->load(is); - } - - *machine = ptr; - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine) { - delete cast(machine); - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_load_parameter_from_disk( - paddle_gradient_machine machine, const char* path) { - auto m = cast(machine); - if (m == nullptr || path == nullptr || m->machine == nullptr) - return kPD_NULLPTR; - m->machine->loadParameters(path); - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_forward(paddle_gradient_machine machine, - paddle_arguments inArgs, - paddle_arguments outArgs, - bool isTrain) { - auto m = cast(machine); - auto in = paddle::capi::cast(inArgs); - auto out = paddle::capi::cast(outArgs); - if (m == nullptr || in == nullptr || out == nullptr || m->machine == nullptr) - return kPD_NULLPTR; - m->machine->forward( - in->args, &out->args, isTrain ? paddle::PASS_TRAIN : paddle::PASS_TEST); - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_create_shared_param( - paddle_gradient_machine origin, - void* modelConfigProtobuf, - int size, - paddle_gradient_machine* slave) { - auto o = cast(origin); - if (origin == nullptr || slave == nullptr || o->machine == nullptr) { - return kPD_NULLPTR; - } - paddle::ModelConfig config; - if (!config.ParseFromArray(modelConfigProtobuf, size) || - !config.IsInitialized()) { - return kPD_PROTOBUF_ERROR; - } - - std::unique_ptr ptr( - new paddle::capi::CGradientMachine()); - auto nn = paddle::NeuralNetwork::create(config); - nn->init(config, - [&o](int paramId, paddle::Parameter* param) { - auto p = o->machine->getParameters()[paramId]; - param->enableSharedType(paddle::PARAMETER_VALUE, - p->getBuf(paddle::PARAMETER_VALUE)); - }, - {paddle::PARAMETER_VALUE}, - false); - ptr->machine.reset(nn); - *slave = ptr.release(); - return kPD_NO_ERROR; -} -} - -paddle_error paddle_gradient_machine_randomize_param( - paddle_gradient_machine machine) { - auto m = cast(machine); - if (m == nullptr || m->machine == nullptr) return kPD_NULLPTR; - m->machine->randParameters(); - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_get_layer_output( - paddle_gradient_machine machine, - const char* layerName, - paddle_arguments args) { - auto m = cast(machine); - auto out = paddle::capi::cast(args); - if (m == nullptr || layerName == nullptr || out == nullptr || - m->machine == nullptr) { - return kPD_NULLPTR; - } - - auto layerOutput = m->machine->getLayerOutput(layerName); - out->args.push_back(layerOutput); - return kPD_NO_ERROR; -} - -paddle_error paddle_gradient_machine_release_layer_output( - paddle_gradient_machine machine) { - auto m = cast(machine); - if (m == nullptr || m->machine == nullptr) { - return kPD_NULLPTR; - } - m->machine->releaseOutput(); - return kPD_NO_ERROR; -} diff --git a/paddle/capi/tests/test_Arguments.cpp b/paddle/capi/tests/test_Arguments.cpp deleted file mode 100644 index bb08adf716bfd6e3c88747616e538e9da89a0e25..0000000000000000000000000000000000000000 --- a/paddle/capi/tests/test_Arguments.cpp +++ /dev/null @@ -1,129 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "capi.h" -#include "gtest/gtest.h" -#include "paddle/utils/ThreadLocal.h" - -static std::vector randomBuffer(size_t bufSize) { - auto& eng = paddle::ThreadLocalRandomEngine::get(); - std::uniform_real_distribution dist(-1.0, 1.0); - std::vector retv; - retv.reserve(bufSize); - for (size_t i = 0; i < bufSize; ++i) { - retv.push_back(dist(eng)); - } - return retv; -} - -TEST(CAPIArguments, create) { - //! TODO(yuyang18): Test GPU Code. - paddle_arguments args = paddle_arguments_create_none(); - uint64_t size; - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_size(args, &size)); - ASSERT_EQ(0UL, size); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); -} - -TEST(CAPIArguments, value) { - paddle_arguments args = paddle_arguments_create_none(); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1)); - - paddle_matrix mat = paddle_matrix_create(128, 64, false); - for (size_t i = 0; i < 128; ++i) { - std::vector sampleBuf = randomBuffer(64); - paddle_matrix_set_row(mat, i, sampleBuf.data()); - } - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(args, 0, mat)); - - paddle_matrix val = paddle_matrix_create_none(); - - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_value(args, 0, val)); - - for (size_t i = 0; i < 128; ++i) { - paddle_real* row1; - paddle_real* row2; - - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, i, &row1)); - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(val, i, &row2)); - ASSERT_EQ(row1, row2); - } - - paddle_ivector ivec = paddle_ivector_create_none(); - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(ivec)); - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(val)); - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat)); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); -} - -TEST(CAPIArguments, ids) { - paddle_arguments args = paddle_arguments_create_none(); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1)); - - paddle_ivector ivec; - int array[3] = {1, 2, 3}; - ivec = paddle_ivector_create(array, 3, true, false); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_ids(args, 0, ivec)); - - paddle_ivector val = paddle_ivector_create_none(); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_ids(args, 0, val)); - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(ivec)); - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(val)); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); -} - -template -void testSequenceHelper(T1 setter, T2 getter) { - paddle_arguments args = paddle_arguments_create_none(); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1)); - - paddle_ivector ivec; - int array[3] = {1, 2, 3}; - ivec = paddle_ivector_create(array, 3, true, false); - ASSERT_EQ(kPD_NO_ERROR, setter(args, 0, ivec)); - - paddle_ivector val = paddle_ivector_create_none(); - ASSERT_EQ(kPD_NO_ERROR, getter(args, 0, val)); - uint64_t size; - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_get_size(val, &size)); - - int* rawBuf; - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_get(val, &rawBuf)); - for (size_t i = 0; i < size; ++i) { - ASSERT_EQ(array[i], rawBuf[i]); - } - - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(ivec)); - ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(val)); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); -} - -TEST(CAPIArguments, Sequence) { - auto testSequence = [](uint32_t nestedLevel) { - testSequenceHelper(std::bind(paddle_arguments_set_sequence_start_pos, - std::placeholders::_1, - std::placeholders::_2, - nestedLevel, - std::placeholders::_3), - std::bind(paddle_arguments_get_sequence_start_pos, - std::placeholders::_1, - std::placeholders::_2, - nestedLevel, - std::placeholders::_3)); - }; - for (uint32_t i = 0; i < 2; ++i) { // test seq and sub-seq. - testSequence(i); - } -} diff --git a/paddle/capi/tests/test_GradientMachine.cpp b/paddle/capi/tests/test_GradientMachine.cpp deleted file mode 100644 index 73b9e477b2a2749250e878cf2174dcf4cc599be1..0000000000000000000000000000000000000000 --- a/paddle/capi/tests/test_GradientMachine.cpp +++ /dev/null @@ -1,117 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include -#include -#include "capi.h" -#include "paddle/utils/ThreadLocal.h" - -static std::vector randomBuffer(size_t bufSize) { - auto& eng = paddle::ThreadLocalRandomEngine::get(); - std::uniform_real_distribution dist(-1.0, 1.0); - std::vector retv; - retv.reserve(bufSize); - for (size_t i = 0; i < bufSize; ++i) { - retv.push_back(dist(eng)); - } - return retv; -} - -TEST(GradientMachine, testPredict) { - //! TODO(yuyang18): Test GPU Code. - paddle::TrainerConfigHelper config("./test_predict_network.py"); - std::string buffer; - ASSERT_TRUE(config.getModelConfig().SerializeToString(&buffer)); - paddle_gradient_machine machine; - - ASSERT_EQ(kPD_NO_ERROR, - paddle_gradient_machine_create_for_inference( - &machine, &buffer[0], (int)buffer.size())); - std::unique_ptr gm( - paddle::GradientMachine::create(config.getModelConfig())); - ASSERT_NE(nullptr, gm); - gm->randParameters(); - gm->saveParameters("./"); - - ASSERT_EQ(kPD_NO_ERROR, - paddle_gradient_machine_load_parameter_from_disk(machine, "./")); - - paddle_gradient_machine machineSlave; - ASSERT_EQ(kPD_NO_ERROR, - paddle_gradient_machine_create_shared_param( - machine, &buffer[0], (int)buffer.size(), &machineSlave)); - std::swap(machineSlave, machine); - paddle_arguments outArgs = paddle_arguments_create_none(); - - paddle_arguments inArgs = paddle_arguments_create_none(); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(inArgs, 1)); - paddle_matrix mat = paddle_matrix_create(1, 100, false); - static_assert(std::is_same::value, ""); - - auto data = randomBuffer(100); - paddle_real* rowPtr; - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr)); - memcpy(rowPtr, data.data(), data.size() * sizeof(paddle_real)); - - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(inArgs, 0, mat)); - ASSERT_EQ(kPD_NO_ERROR, - paddle_gradient_machine_forward(machine, inArgs, outArgs, false)); - - uint64_t sz; - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_size(outArgs, &sz)); - ASSERT_EQ(1UL, sz); - - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_value(outArgs, 0, mat)); - std::vector paddleInArgs; - std::vector paddleOutArgs; - paddleInArgs.resize(1); - paddleInArgs[0].value = - paddle::Matrix::create(data.data(), 1, 100, false, false); - - gm->forward(paddleInArgs, &paddleOutArgs, paddle::PASS_TEST); - - auto matPaddle = paddleOutArgs[0].value; - - uint64_t height, width; - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width)); - ASSERT_EQ(matPaddle->getHeight(), height); - ASSERT_EQ(matPaddle->getWidth(), width); - - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr)); - for (size_t i = 0; i < width; ++i) { - ASSERT_NEAR(matPaddle->getData()[i], rowPtr[i], 1e-5); - } - - ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat)); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(inArgs)); - ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(outArgs)); - std::swap(machineSlave, machine); - ASSERT_EQ(kPD_NO_ERROR, paddle_gradient_machine_destroy(machineSlave)); - ASSERT_EQ(kPD_NO_ERROR, paddle_gradient_machine_destroy(machine)); -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - std::vector argvs; - argvs.push_back(strdup("--use_gpu=false")); - paddle_init((int)argvs.size(), argvs.data()); - for (auto each : argvs) { - free(each); - } - return RUN_ALL_TESTS(); -} diff --git a/paddle/contrib/CMakeLists.txt b/paddle/contrib/CMakeLists.txt deleted file mode 100644 index 4b19256ef4533a09162edf907f6cd51146517e46..0000000000000000000000000000000000000000 --- a/paddle/contrib/CMakeLists.txt +++ /dev/null @@ -1,16 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# - -add_subdirectory(inference) diff --git a/paddle/contrib/float16/float16_transpiler.py b/paddle/contrib/float16/float16_transpiler.py index 91ba101edb65cd45bd5e37a0c6ad25e515593a81..66e0345c299730c113ffbdc8dd3c1fa32f872f3d 100644 --- a/paddle/contrib/float16/float16_transpiler.py +++ b/paddle/contrib/float16/float16_transpiler.py @@ -118,7 +118,7 @@ class Float16Transpiler: for var in self.block.vars.keys(): if var not in args: - self.block.remove_var(var) + self.block._remove_var(var) def _modify_feed_fetch(self): ''' @@ -165,7 +165,7 @@ class Float16Transpiler: dtype=core.VarDesc.VarType.FP16, shape=var.shape, persistable=var.persistable) - self.block.insert_op( + self.block._insert_op( i + 1, type="cast", inputs={"X": var}, @@ -188,7 +188,7 @@ class Float16Transpiler: persistable=var.persistable) find_op(var) var.op.rename_output(var_name, tmp_var_name) - self.block.insert_op( + self.block._insert_op( i, type="cast", inputs={"X": tmp_var}, @@ -253,4 +253,4 @@ class Float16Transpiler: # old var will be replaced by the fp16 var in program desc self.input_map[var.name] = fp16_var_name - self.block.remove_var(var.name) + self.block._remove_var(var.name) diff --git a/paddle/contrib/inference/CMakeLists.txt b/paddle/contrib/inference/CMakeLists.txt deleted file mode 100644 index 6847f7db7fc0f6b41ced1260d409ca6eba9b53eb..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/CMakeLists.txt +++ /dev/null @@ -1,49 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# - -if(APPLE) - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") -endif(APPLE) - -function(inference_api_test TARGET_NAME) - set(options "") - set(oneValueArgs "") - set(multiValueArgs ARGS) - cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - - set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) - cc_test(test_paddle_inference_${TARGET_NAME} - SRCS test_paddle_inference_${TARGET_NAME}.cc - DEPS paddle_fluid_api paddle_inference_api - ARGS --dirname=${PYTHON_TESTS_DIR}/book/) - if(inference_test_ARGS) - set_tests_properties(test_paddle_inference_${TARGET_NAME} - PROPERTIES DEPENDS "${inference_test_ARGS}") - endif() -endfunction(inference_api_test) - - -cc_library(paddle_inference_api - SRCS paddle_inference_api.cc paddle_inference_api_impl.cc - DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) - -if(WITH_TESTING) - cc_test(test_paddle_inference_api - SRCS test_paddle_inference_api.cc - DEPS paddle_inference_api) - - inference_api_test(api_impl - ARGS test_word2vec test_image_classification) -endif() diff --git a/paddle/contrib/inference/paddle_inference_api.cc b/paddle/contrib/inference/paddle_inference_api.cc deleted file mode 100644 index d67e1e7667800d6dd00cb8915b0d6dc7c664970b..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/paddle_inference_api.cc +++ /dev/null @@ -1,15 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - -http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/contrib/inference/paddle_inference_api.h" diff --git a/paddle/contrib/inference/paddle_inference_api.h b/paddle/contrib/inference/paddle_inference_api.h deleted file mode 100644 index b4c7f9bef4d2e83038ff223614a89e1b0493fc6f..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/paddle_inference_api.h +++ /dev/null @@ -1,103 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -/* - * This file contains the definition of a simple Inference API for Paddle. - * - * ATTENTION: It requires some C++ features, for lower version C++ or C, we - * might release another API. - */ - -#pragma once - -#include -#include -#include - -namespace paddle { - -enum PaddleDType { - FLOAT32, - INT64, -}; - -struct PaddleBuf { - void* data; // pointer to the data memory. - size_t length; // number of memory bytes. -}; - -struct PaddleTensor { - std::string name; // variable name. - std::vector shape; - PaddleBuf data; // blob of data. - PaddleDType dtype; -}; - -/* - * A simple Inference API for Paddle. Currently this API can be used by - * non-sequence scenerios. - * TODO(Superjomn) Support another API for NLP-related usages. - */ -class PaddlePredictor { - public: - struct Config; - PaddlePredictor() = default; - PaddlePredictor(const PaddlePredictor&) = delete; - - // Predict an record. - // The caller should be responsible for allocating and releasing the memory of - // `inputs`. `inputs` should be alive until Run returns. caller should be - // responsible for releasing the memory of `output_data`. - virtual bool Run(const std::vector& inputs, - std::vector* output_data) = 0; - - // Clone a predictor that share the model weights, the Cloned predictor should - // be thread-safe. - virtual std::unique_ptr Clone() = 0; - - // Destroy the Predictor. - virtual ~PaddlePredictor() {} - - enum class EngineKind { - kNative = -1, // Use the native Fluid facility. - // TODO(Superjomn) support latter. - // kAnakin, // Use Anakin for inference. - // kTensorRT, // Use TensorRT for inference. - // kAutoMixedAnakin, // Automatically mix Fluid with Anakin. - // kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. - }; - - // The common configs for all the predictors. - struct Config { - std::string model_dir; // path to the model directory. - bool enable_engine{false}; // Enable to execute (part of) the model on - }; -}; - -struct NativeConfig : public PaddlePredictor::Config { - bool use_gpu{false}; - int device; - float fraction_of_gpu_memory; - std::string prog_file; - std::string param_file; - bool share_variables; -}; - -// A factory to help create difference predictor. -template < - typename ConfigT, - PaddlePredictor::EngineKind engine = PaddlePredictor::EngineKind::kNative> -std::unique_ptr CreatePaddlePredictor(const ConfigT& config); - -} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_impl.cc b/paddle/contrib/inference/paddle_inference_api_impl.cc deleted file mode 100644 index 989252f69e42778dfd791cdee02c550f2aa78803..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/paddle_inference_api_impl.cc +++ /dev/null @@ -1,269 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#include -#include -#include -#include -#include -#include -#include -#include - -#include "paddle/contrib/inference/paddle_inference_api_impl.h" - -namespace paddle { -namespace { - -// Timer for timer -class Timer { - public: - double start; - double startu; - void tic() { - struct timeval tp; - gettimeofday(&tp, NULL); - start = tp.tv_sec; - startu = tp.tv_usec; - } - double toc() { - struct timeval tp; - gettimeofday(&tp, NULL); - double used_time_ms = - (tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0; - return used_time_ms; - } -}; - -template -std::string num2str(T a) { - std::stringstream istr; - istr << a; - return istr.str(); -} -} // namespace - -bool NativePaddlePredictor::Init() { - VLOG(3) << "Predictor::init()"; - - // TODO(panyx0718): Should CPU vs GPU device be decided by id? - if (config_.device >= 0) { - place_ = paddle::platform::CUDAPlace(config_.device); - } else { - place_ = paddle::platform::CPUPlace(); - } - paddle::framework::InitDevices(false); - executor_.reset(new paddle::framework::Executor(place_)); - scope_.reset(new paddle::framework::Scope()); - - // Initialize the inference program - if (!config_.model_dir.empty()) { - // Parameters are saved in separate files sited in - // the specified `dirname`. - inference_program_ = paddle::inference::Load( - executor_.get(), scope_.get(), config_.model_dir); - } else if (!config_.prog_file.empty() && !config_.param_file.empty()) { - // All parameters are saved in a single file. - // The file names should be consistent with that used - // in Python API `fluid.io.save_inference_model`. - inference_program_ = paddle::inference::Load( - executor_.get(), scope_.get(), config_.prog_file, config_.param_file); - } else { - LOG(ERROR) << "fail to load inference model."; - return false; - } - ctx_ = executor_->Prepare(*inference_program_, 0); - - // Create variables - // TODO(panyx0718): Why need to test share_variables here? - if (config_.share_variables) { - executor_->CreateVariables(*inference_program_, scope_.get(), 0); - } - // Get the feed_target_names and fetch_target_names - feed_target_names_ = inference_program_->GetFeedTargetNames(); - fetch_target_names_ = inference_program_->GetFetchTargetNames(); - return true; -} - -bool NativePaddlePredictor::Run(const std::vector &inputs, - std::vector *output_data) { - VLOG(3) << "Predictor::predict"; - Timer timer; - timer.tic(); - // set feed variable - std::map feed_targets; - std::vector feeds; - if (!SetFeed(inputs, &feeds)) { - LOG(ERROR) << "fail to set feed"; - return false; - } - for (size_t i = 0; i < feed_target_names_.size(); ++i) { - feed_targets[feed_target_names_[i]] = &feeds[i]; - } - // get fetch variable - std::map fetch_targets; - std::vector fetchs; - fetchs.resize(fetch_target_names_.size()); - for (size_t i = 0; i < fetch_target_names_.size(); ++i) { - fetch_targets[fetch_target_names_[i]] = &fetchs[i]; - } - // Run the inference program - // if share variables, we need not create variables - executor_->RunPreparedContext(ctx_.get(), - scope_.get(), - &feed_targets, - &fetch_targets, - !config_.share_variables); - if (!GetFetch(fetchs, output_data)) { - LOG(ERROR) << "fail to get fetchs"; - return false; - } - VLOG(3) << "predict cost: " << timer.toc() << "ms"; - return true; -} - -std::unique_ptr NativePaddlePredictor::Clone() { - VLOG(3) << "Predictor::clone"; - std::unique_ptr cls(new NativePaddlePredictor(config_)); - - if (!dynamic_cast(cls.get())->Init()) { - LOG(ERROR) << "fail to call Init"; - return nullptr; - } - // fix manylinux compile error. - return std::move(cls); -} - -bool NativePaddlePredictor::SetFeed(const std::vector &inputs, - std::vector *feeds) { - VLOG(3) << "Predictor::set_feed"; - if (inputs.size() != feed_target_names_.size()) { - LOG(ERROR) << "wrong feed input size."; - return false; - } - for (size_t i = 0; i < feed_target_names_.size(); ++i) { - framework::LoDTensor input; - framework::DDim ddim = framework::make_ddim(inputs[i].shape); - void *input_ptr; - if (inputs[i].dtype == PaddleDType::INT64) { - input_ptr = input.mutable_data(ddim, platform::CPUPlace()); - } else if (inputs[i].dtype == PaddleDType::FLOAT32) { - input_ptr = input.mutable_data(ddim, platform::CPUPlace()); - } else { - LOG(ERROR) << "unsupported feed type " << inputs[i].dtype; - return false; - } - - // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. - std::memcpy(static_cast(input_ptr), - inputs[i].data.data, - inputs[i].data.length); - feeds->push_back(input); - } - return true; -} - -bool NativePaddlePredictor::GetFetch( - const std::vector &fetchs, - std::vector *outputs) { - VLOG(3) << "Predictor::get_fetch"; - outputs->resize(fetchs.size()); - for (size_t i = 0; i < fetchs.size(); ++i) { - // TODO(panyx0718): Support fetch of other types. - if (fetchs[i].type() != typeid(float)) { - LOG(ERROR) << "only support fetching float now."; - return false; - } - std::vector shape; - auto dims_i = fetchs[i].dims(); - auto lod = fetchs[i].lod(); - const float *output_ptr = fetchs[i].data(); - // const int64_t* output_ptr = fetchs[i].data(); - auto num = fetchs[i].numel(); - std::vector data; - if (0 == lod.size()) { - std::copy(output_ptr, output_ptr + num, std::back_inserter(data)); - for (int j = 0; j < dims_i.size(); ++j) { - shape.push_back(dims_i[j]); - } - } else { - // for batch detection - // image[0] -> output[0] shape {145, 6} - // image[1] -> output[1] shape {176, 6} - // then, - // the batch output shape {321, 6} - // the lod {{0, 145, 321}} - // so we should append output[0] to {176, 6} - size_t max_dim = 0; - for (size_t j = 1; j < lod[0].size(); j++) { - max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]); - } - size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back(); - if (max_dim > 0) { - data.resize((lod[0].size() - 1) * max_dim * common_dim, 0); - } - for (size_t j = 1; j < lod[0].size(); j++) { - size_t start = lod[0][j - 1] * common_dim; - size_t end = lod[0][j] * common_dim; - if (end > start) { - std::copy(output_ptr + start, - output_ptr + end, - data.begin() + (j - 1) * max_dim * common_dim); - } - } - shape.push_back(lod[0].size() - 1); - shape.push_back(max_dim); - for (int j = 1; j < dims_i.size(); ++j) { - shape.push_back(dims_i[j]); - } - } - - outputs->at(i).shape = shape; - outputs->at(i).data.length = sizeof(float) * data.size(); - outputs->at(i).data.data = malloc(outputs->at(i).data.length); - std::memcpy( - outputs->at(i).data.data, data.data(), outputs->at(i).data.length); - outputs->at(i).dtype = PaddleDType::FLOAT32; - // TODO(panyx0718): support other types? fill tensor name? avoid a copy. - } - return true; -} - -template <> -std::unique_ptr -CreatePaddlePredictor( - const NativeConfig &config) { - VLOG(3) << "create NativePaddlePredictor"; - if (config.use_gpu) { - // 1. GPU memeroy - std::vector flags; - if (config.fraction_of_gpu_memory >= 0.0f || - config.fraction_of_gpu_memory <= 0.95f) { - flags.push_back("dummpy"); - std::string flag = "--fraction_of_gpu_memory_to_use=" + - num2str(config.fraction_of_gpu_memory); - flags.push_back(flag); - VLOG(3) << "set flag: " << flag; - framework::InitGflags(flags); - } - } - - std::unique_ptr predictor(new NativePaddlePredictor(config)); - if (!dynamic_cast(predictor.get())->Init()) { - return nullptr; - } - return std::move(predictor); -} - -} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_impl.h b/paddle/contrib/inference/paddle_inference_api_impl.h deleted file mode 100644 index 84707e223d7aa3d1ebca933923e932b3973613ae..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/paddle_inference_api_impl.h +++ /dev/null @@ -1,62 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#pragma once - -#include -#include -#include -#include - -#include "paddle/contrib/inference/paddle_inference_api.h" - -#include "paddle/fluid/framework/ddim.h" -#include "paddle/fluid/framework/init.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/inference/io.h" -#include "paddle/fluid/platform/profiler.h" - -namespace paddle { - -class NativePaddlePredictor : public PaddlePredictor { - public: - explicit NativePaddlePredictor(const NativeConfig &config) - : config_(config) {} - - bool Init(); - - bool Run(const std::vector &inputs, - std::vector *output_data) override; - - std::unique_ptr Clone() override; - - ~NativePaddlePredictor() override{}; - - private: - bool SetFeed(const std::vector &input_datas, - std::vector *feeds); - bool GetFetch(const std::vector &fetchs, - std::vector *output_data); - - NativeConfig config_; - platform::Place place_; - std::unique_ptr executor_; - std::unique_ptr scope_; - std::unique_ptr ctx_; - std::unique_ptr inference_program_; - std::vector feed_target_names_; - std::vector fetch_target_names_; -}; - -} // namespace paddle diff --git a/paddle/contrib/inference/test_paddle_inference_api.cc b/paddle/contrib/inference/test_paddle_inference_api.cc deleted file mode 100644 index bc7faab6e208a66d7a56e41a56bd743c7644eea2..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/test_paddle_inference_api.cc +++ /dev/null @@ -1,64 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - -http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/contrib/inference/paddle_inference_api.h" - -#include -#include - -namespace paddle { - -/* - * Do not use this, just a demo indicating how to customize a config for a - * specific predictor. - */ -struct DemoConfig : public PaddlePredictor::Config { - float other_config; -}; - -/* - * Do not use this, just a demo indicating how to customize a Predictor. - */ -class DemoPredictor : public PaddlePredictor { - public: - explicit DemoPredictor(const DemoConfig &config) { - LOG(INFO) << "I get other_config " << config.other_config; - } - bool Run(const std::vector &inputs, - std::vector *output_data) override { - LOG(INFO) << "Run"; - return false; - } - - std::unique_ptr Clone() override { return nullptr; } - - ~DemoPredictor() override {} -}; - -template <> -std::unique_ptr CreatePaddlePredictor( - const DemoConfig &config) { - std::unique_ptr x(new DemoPredictor(config)); - return x; -} - -TEST(paddle_inference_api, demo) { - DemoConfig config; - config.other_config = 1.7; - auto predictor = CreatePaddlePredictor(config); - std::vector outputs; - predictor->Run({}, &outputs); -} - -} // namespace paddle diff --git a/paddle/contrib/inference/test_paddle_inference_api_impl.cc b/paddle/contrib/inference/test_paddle_inference_api_impl.cc deleted file mode 100644 index 5240fc2f20211ac5d38c57b71db31d04a6dc536a..0000000000000000000000000000000000000000 --- a/paddle/contrib/inference/test_paddle_inference_api_impl.cc +++ /dev/null @@ -1,153 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - -http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include "gflags/gflags.h" -#include "paddle/contrib/inference/paddle_inference_api_impl.h" -#include "paddle/fluid/inference/tests/test_helper.h" - -DEFINE_string(dirname, "", "Directory of the inference model."); - -namespace paddle { - -PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { - PaddleTensor pt; - pt.data.data = t->data(); - - if (t->type() == typeid(int64_t)) { - pt.data.length = t->numel() * sizeof(int64_t); - pt.dtype = PaddleDType::INT64; - } else if (t->type() == typeid(float)) { - pt.data.length = t->numel() * sizeof(float); - pt.dtype = PaddleDType::FLOAT32; - } else { - LOG(FATAL) << "unsupported type."; - } - pt.shape = framework::vectorize2int(t->dims()); - return pt; -} - -NativeConfig GetConfig() { - NativeConfig config; - config.model_dir = FLAGS_dirname + "word2vec.inference.model"; - LOG(INFO) << "dirname " << config.model_dir; - config.fraction_of_gpu_memory = 0.15; - config.use_gpu = true; - config.device = 0; - config.share_variables = true; - return config; -} - -TEST(paddle_inference_api_impl, word2vec) { - NativeConfig config = GetConfig(); - auto predictor = CreatePaddlePredictor(config); - - framework::LoDTensor first_word, second_word, third_word, fourth_word; - framework::LoD lod{{0, 1}}; - int64_t dict_size = 2073; // The size of dictionary - - SetupLoDTensor(&first_word, lod, static_cast(0), dict_size - 1); - SetupLoDTensor(&second_word, lod, static_cast(0), dict_size - 1); - SetupLoDTensor(&third_word, lod, static_cast(0), dict_size - 1); - SetupLoDTensor(&fourth_word, lod, static_cast(0), dict_size - 1); - - std::vector paddle_tensor_feeds; - paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&first_word)); - paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&second_word)); - paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&third_word)); - paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&fourth_word)); - - std::vector outputs; - ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); - ASSERT_EQ(outputs.size(), 1UL); - size_t len = outputs[0].data.length; - float* data = static_cast(outputs[0].data.data); - for (int j = 0; j < len / sizeof(float); ++j) { - ASSERT_LT(data[j], 1.0); - ASSERT_GT(data[j], -1.0); - } - - std::vector cpu_feeds; - cpu_feeds.push_back(&first_word); - cpu_feeds.push_back(&second_word); - cpu_feeds.push_back(&third_word); - cpu_feeds.push_back(&fourth_word); - - framework::LoDTensor output1; - std::vector cpu_fetchs1; - cpu_fetchs1.push_back(&output1); - - TestInference(config.model_dir, cpu_feeds, cpu_fetchs1); - - float* lod_data = output1.data(); - for (size_t i = 0; i < output1.numel(); ++i) { - EXPECT_LT(lod_data[i] - data[i], 1e-3); - EXPECT_GT(lod_data[i] - data[i], -1e-3); - } - - free(outputs[0].data.data); -} - -TEST(paddle_inference_api_impl, image_classification) { - int batch_size = 2; - bool use_mkldnn = false; - bool repeat = false; - NativeConfig config = GetConfig(); - config.model_dir = - FLAGS_dirname + "image_classification_resnet.inference.model"; - - const bool is_combined = false; - std::vector> feed_target_shapes = - GetFeedTargetShapes(config.model_dir, is_combined); - - framework::LoDTensor input; - // Use normilized image pixels as input data, - // which should be in the range [0.0, 1.0]. - feed_target_shapes[0][0] = batch_size; - framework::DDim input_dims = framework::make_ddim(feed_target_shapes[0]); - SetupTensor( - &input, input_dims, static_cast(0), static_cast(1)); - std::vector cpu_feeds; - cpu_feeds.push_back(&input); - - framework::LoDTensor output1; - std::vector cpu_fetchs1; - cpu_fetchs1.push_back(&output1); - - TestInference(config.model_dir, - cpu_feeds, - cpu_fetchs1, - repeat, - is_combined, - use_mkldnn); - - auto predictor = CreatePaddlePredictor(config); - std::vector paddle_tensor_feeds; - paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&input)); - - std::vector outputs; - ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); - ASSERT_EQ(outputs.size(), 1UL); - size_t len = outputs[0].data.length; - float* data = static_cast(outputs[0].data.data); - float* lod_data = output1.data(); - for (size_t j = 0; j < len / sizeof(float); ++j) { - EXPECT_NEAR(lod_data[j], data[j], 1e-3); - } - free(data); -} - -} // namespace paddle diff --git a/paddle/cuda/include/hl_base.h b/paddle/cuda/include/hl_base.h deleted file mode 100644 index 77f5d82dbe2cad183491033736bac85961b6d320..0000000000000000000000000000000000000000 --- a/paddle/cuda/include/hl_base.h +++ /dev/null @@ -1,250 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include - -#ifdef PADDLE_TYPE_DOUBLE -#define HL_FLOAT_MAX 3.40282347e+38F -#define HL_FLOAT_MIN 1.17549435e-38F -using real = double; -#else -#define HL_FLOAT_MAX 1.7976931348623157e+308 -#define HL_FLOAT_MIN 2.2250738585072014e-308 -using real = float; -#endif - -/** - * The maximum input value for exp, used to avoid overflow problem. - * currently only used for tanh function. - */ -#define EXP_MAX_INPUT 40.0 - -/** - * @brief DIVUP(x, y) is similar to ceil(x / y). - * @note For CUDA, DIVUP will be used to specify - * the size of blockDim. - */ -#ifndef DIVUP -#define DIVUP(x, y) (((x) + (y)-1) / (y)) -#endif - -/** - * HPPL is an internal high performance parallel computing library - * for high-level neural network routines, which can support many - * heterogeneous compute architectures, such as GPU, FPGA, etc. - */ - -/** - * @brief HPPL CUDA Stream. - * - * @note Each thread can use HPPL_STREAM_* after calling hl_init. - * HPPL_STREAM_DEFAULT is HPPL default stream. - */ -typedef enum { - HPPL_STREAM_DEFAULT = 0, /* Thread Default Stream*/ - HPPL_STREAM_1 = 1, - HPPL_STREAM_2 = 2, - HPPL_STREAM_3 = 3, - HPPL_STREAM_4 = 4, - HPPL_THREAD_STREAM_1 = 5, - HPPL_THREAD_STREAM_2 = 6, - HPPL_THREAD_STREAM_3 = 7, - HPPL_THREAD_STREAM_4 = 8, - HPPL_STREAM_END -} hl_stream_t; - -/** - * @brief HPPL activation mode. - */ -typedef enum { - HL_ACTIVATION_SIGMOID = 0, - HL_ACTIVATION_RELU = 1, - HL_ACTIVATION_TANH = 2, - HL_ACTIVATION_LINEAR = 3, - HL_ACTIVATION_END -} hl_activation_mode_t; - -/** - * @brief Transpose type. - */ -typedef enum { - HPPL_OP_N = 0, /* transpose */ - HPPL_OP_T = 1, /* non transpose */ - HPPL_OP_END -} hl_trans_op_t; - -/** - * @brief Lstm value. - * - * @param gateValue input value. - * @param prevStateValue previous state value. - * @param stateValue state value. - * @param stateActiveValue state active value. - * @param outputValue output value. - */ -typedef struct { - real *gateValue; - real *prevStateValue; - real *stateValue; - real *stateActiveValue; - real *outputValue; - real *checkIg; - real *checkFg; - real *checkOg; -} hl_lstm_value; - -/** - * @brief Lstm gradient. - * - * @param gateGrad input gradient. - * @param prevStateGrad previous state gradient. - * @param stateGrad state gradient. - * @param stateActiveGrad state active gradient. - * @param outputGrad output gradient. - */ -typedef struct { - real *gateGrad; - real *prevStateGrad; - real *stateGrad; - real *stateActiveGrad; - real *outputGrad; - real *checkIgGrad; - real *checkFgGrad; - real *checkOgGrad; -} hl_lstm_grad; - -/** - * @brief Gru value. - * - * @param gateWeight gate weight (updateGate + resetGate). - * @param stateWeight frame state weight. - * @param gateValue gate value results. - * @param resetOutputValue resetOutput value. - * @param outputValue output value. - * @param prevOutValue previous output value. - * - */ -typedef struct { - real *gateWeight; - real *stateWeight; - real *gateValue; - real *resetOutputValue; - real *outputValue; - real *prevOutValue; -} hl_gru_value; - -/** - * @brief Gru gradient. - * - * @param gateWeightGrad gate weight gradient. - * @param stateWeightGrad frame state weight gradient. - * @param gateGrad gate gradient results. - * @param resetOutputGrad resetOutput gradient. - * @param outputGrad output gradient. - * @param prevOutGrad previous output gradient. - */ -typedef struct { - real *gateWeightGrad; - real *stateWeightGrad; - real *gateGrad; - real *resetOutputGrad; - real *outputGrad; - real *prevOutGrad; -} hl_gru_grad; - -/** - * @brief Sparse matrix value type. - */ -typedef enum { - HL_NO_VALUE = 0, /* matrix values only 0 or 1 */ - HL_FLOAT_VALUE = 1, - HL_VALUE_END -} hl_matrix_value_t; - -/** - * @brief HPPL matrix format. - */ -typedef enum { - HL_SPARSE_CSR = 0, - HL_SPARSE_CSC = 1, - HL_SPARSE_END -} hl_matrix_format_t; - -typedef struct _hl_matrix_s *hl_matrix_s; - -/** - * @brief HPPL sparse matrix. - * - * @param matrix sparse matrix. - * @param format matrix format. - * @param type the type of matrix values. - * @param rows matrix rows. - * @param cols matrix columns. - * @param nnz nonzero values of sparse matrix. - */ -typedef struct { - hl_matrix_s matrix; - hl_matrix_format_t format; - hl_matrix_value_t type; - int rows; - int cols; - size_t nnz; -} _hl_sparse_matrix_s, *hl_sparse_matrix_s; - -#ifdef __NVCC__ - -#include -#include "paddle/cuda/include/hl_cuda.h" -#include "paddle/utils/Logging.h" - -extern __thread bool g_sync_flag; -extern __thread cudaStream_t default_stream; -#define STREAM_DEFAULT default_stream - -/** - * @brief Check cuda kernel execution. - * @param msg error string - */ -#define CHECK_SYNC(msg) \ - if (true == g_sync_flag) { \ - hl_stream_synchronize(HPPL_STREAM_DEFAULT); \ - cudaError_t err = (cudaError_t)hl_get_device_last_error(); \ - CHECK_EQ(cudaSuccess, err) \ - << "[" << msg << "] " \ - << "CUDA error: " << hl_get_device_error_string((size_t)err); \ - } - -// __shfl has been deprecated as of CUDA 9.0. -#if CUDA_VERSION < 9000 -template -__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { - return __shfl_down(val, delta); -} - -template -__forceinline__ __device__ T -__shfl_sync(unsigned, T val, int src_line, int width) { - return __shfl(val, src_line, width); -} - -#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; -#else -#define FULL_WARP_MASK 0xFFFFFFFF -#define CREATE_SHFL_MASK(mask, predicate) \ - mask = __ballot_sync(FULL_WARP_MASK, (predicate)) -#endif - -#endif // __NVCC__ diff --git a/paddle/cuda/include/hl_gpu_gru.cuh b/paddle/cuda/include/hl_gpu_gru.cuh deleted file mode 100644 index 9fcad2c3bc2fa255e3d7cd3e7940a32fd286751b..0000000000000000000000000000000000000000 --- a/paddle/cuda/include/hl_gpu_gru.cuh +++ /dev/null @@ -1,393 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - - -#ifndef HL_GPU_GRU_CUH_ -#define HL_GPU_GRU_CUH_ - -#ifdef __NVCC__ - -#include "paddle/utils/Logging.h" - -/* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) - */ -template -__global__ void KeGruForwardResetOutput(OpResetOutput opResetOutput, - real *gateValue, - real *resetOutputValue, - real *prevOutputValue, - int frameSize, - int batchSize, - hl_activation_mode_t active_gate) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - resetOutputValue += batchIdx * frameSize; - } - - real rPrevOut = 0; - real rValueResetOutput; - real rValueUpdateGate = gateValue[frameIdx + frameSize * 0]; - real rValueResetGate = gateValue[frameIdx + frameSize * 1]; - - if (prevOutputValue) { - if (isBatch) prevOutputValue += batchIdx * frameSize; - rPrevOut = prevOutputValue[frameIdx]; - } - - opResetOutput(rValueUpdateGate, - rValueResetGate, - rPrevOut, - rValueResetOutput, - hppl::gpu::forward[active_gate]); - - gateValue[frameIdx + frameSize * 0] = rValueUpdateGate; - gateValue[frameIdx + frameSize * 1] = rValueResetGate; - resetOutputValue[frameIdx] = rValueResetOutput; -} - -/* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) - */ -template -__global__ void KeGruForwardFinalOutput(OpFinalOutput opFinalOutput, - real *gateValue, - real *prevOutputValue, - real *outputValue, - int frameSize, - int batchSize, - hl_activation_mode_t active_node) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - outputValue += batchIdx * frameSize; - } - - real rOutput; - real rPrevOut = 0; - real rValueUpdateGate = gateValue[frameIdx + frameSize * 0]; - real rValueFrameState = gateValue[frameIdx + frameSize * 2]; - - if (prevOutputValue) { - if (isBatch) prevOutputValue += batchIdx * frameSize; - rPrevOut = prevOutputValue[frameIdx]; - } - - opFinalOutput(rValueUpdateGate, - rValueFrameState, - rPrevOut, - rOutput, - hppl::gpu::forward[active_node]); - - gateValue[frameIdx + frameSize * 2] = rValueFrameState; - outputValue[frameIdx] = rOutput; -} - -template -void hl_gpu_gru_forward(OpResetOutput opResetOutput, - OpFinalOutput opFinalOutput, - hl_gru_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) { - dim3 threads; - dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); - } else { - threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); - } - - if (value.prevOutValue) { - hl_matrix_mul(value.prevOutValue, HPPL_OP_N, - value.gateWeight, HPPL_OP_N, - value.gateValue, - batchSize, 2*frameSize, frameSize, - /*alpha = */ 1, /*beta = */ 1, - frameSize, 2* frameSize, 3*frameSize); - } - - if (batchSize == 1) { - KeGruForwardResetOutput - <<>>(opResetOutput, - value.gateValue, value.resetOutputValue, value.prevOutValue, - frameSize, batchSize, active_gate); - } else { - KeGruForwardResetOutput - <<>>(opResetOutput, - value.gateValue, value.resetOutputValue, value.prevOutValue, - frameSize, batchSize, active_gate); - } - - if (value.prevOutValue) { - hl_matrix_mul(value.resetOutputValue, HPPL_OP_N, - value.stateWeight, HPPL_OP_N, - value.gateValue + 2*frameSize, - batchSize, frameSize, frameSize, - /*alpha = */ 1, /*beta = */ 1, - frameSize, frameSize, 3*frameSize); - } - - if (batchSize == 1) { - KeGruForwardFinalOutput - <<>>(opFinalOutput, - value.gateValue, value.prevOutValue, value.outputValue, - frameSize, batchSize, active_node); - } else { - KeGruForwardFinalOutput - <<>>(opFinalOutput, - value.gateValue, value.prevOutValue, value.outputValue, - frameSize, batchSize, active_node); - } - - CHECK_SYNC("hl_gpu_gru_forward failed"); -} - -/* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) - */ -template -__global__ void KeGruBackwardStateGrad(OpStateGrad opStateGrad, - real *gateValue, - real *gateGrad, - real *prevOutValue, - real *prevOutGrad, - real *outputGrad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - gateGrad += batchIdx * 3 * frameSize; - outputGrad += batchIdx * frameSize; - } - - real rUpdateGateGrad; - real rFrameStateGrad; - real rPrevOutValue = 0; - real rPrevOutGrad = 0; - real rUpdateGateValue = gateValue[frameIdx + frameSize * 0]; - real rFrameStateValue = gateValue[frameIdx + frameSize * 2]; - real rOutGrad = outputGrad[frameIdx]; - - if (prevOutValue && prevOutGrad) { - if (isBatch) prevOutValue += batchIdx * frameSize; - rPrevOutValue = prevOutValue[frameIdx]; - - if (isBatch) prevOutGrad += batchIdx * frameSize; - rPrevOutGrad = prevOutGrad[frameIdx]; - } - - opStateGrad(rUpdateGateValue, - rUpdateGateGrad, - rFrameStateValue, - rFrameStateGrad, - rPrevOutValue, - rPrevOutGrad, - rOutGrad, - hppl::gpu::backward[active_node]); - - gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad; - gateGrad[frameIdx + frameSize * 2] = rFrameStateGrad; - if (prevOutGrad) { - prevOutGrad[frameIdx] = rPrevOutGrad; - } -} - -/* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) - */ -template -__global__ void KeGruBackwardResetGrad(OpResetGrad opResetGrad, - real *gateValue, - real *gateGrad, - real *prevOutValue, - real *prevOutGrad, - real *resetOutputGrad, - int frameSize, - int batchSize, - hl_activation_mode_t active_gate) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - gateGrad += batchIdx * 3 * frameSize; - resetOutputGrad += batchIdx * frameSize; - } - - real rResetGateGrad; - real rPrevOutValue = 0; - real rPrevOutGrad = 0; - real rResetOutputGrad = 0; - real rUpdateGateValue = gateValue[frameIdx + frameSize * 0]; - real rUpdateGateGrad = gateGrad[frameIdx + frameSize * 0]; - real rResetGateValue = gateValue[frameIdx + frameSize * 1]; - - if (prevOutValue && prevOutGrad) { - if (isBatch) prevOutValue += batchIdx * frameSize; - if (isBatch) prevOutGrad += batchIdx * frameSize; - rPrevOutValue = prevOutValue[frameIdx]; - rPrevOutGrad = prevOutGrad[frameIdx]; - rResetOutputGrad = resetOutputGrad[frameIdx]; - } - - opResetGrad(rUpdateGateValue, - rUpdateGateGrad, - rResetGateValue, - rResetGateGrad, - rPrevOutValue, - rPrevOutGrad, - rResetOutputGrad, - hppl::gpu::backward[active_gate]); - - gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad; - gateGrad[frameIdx + frameSize * 1] = rResetGateGrad; - if (prevOutGrad) { - prevOutGrad[frameIdx] = rPrevOutGrad; - } -} - -template -void hl_gpu_gru_backward(OpStateGrad opStateGrad, - OpResetGrad opResetGrad, - hl_gru_value value, - hl_gru_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) { - dim3 threads; - dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); - } else { - threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); - } - - if (batchSize == 1) { - KeGruBackwardStateGrad - <<>>(opStateGrad, - value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, - grad.outputGrad, frameSize, batchSize, active_node); - } else { - KeGruBackwardStateGrad - <<>>(opStateGrad, - value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, - grad.outputGrad, frameSize, batchSize, active_node); - } - - if (value.prevOutValue && grad.prevOutGrad) { - hl_matrix_mul(grad.gateGrad + 2*frameSize, HPPL_OP_N, - value.stateWeight, HPPL_OP_T, - grad.resetOutputGrad, - batchSize, frameSize, frameSize, - /*alpha = */ 1, /*beta = */ 0, - 3*frameSize, frameSize, frameSize); - if (grad.stateWeightGrad) { - hl_matrix_mul(value.resetOutputValue, HPPL_OP_T, - grad.gateGrad + 2*frameSize, HPPL_OP_N, - grad.stateWeightGrad, - frameSize, frameSize, batchSize, - /*alpha = */ 1, /*beta = */ 1, - frameSize, 3*frameSize, frameSize); - } - } - - if (batchSize == 1) { - KeGruBackwardResetGrad - <<>>(opResetGrad, - value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, - grad.resetOutputGrad, frameSize, batchSize, active_gate); - } else { - KeGruBackwardResetGrad - <<>>(opResetGrad, - value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, - grad.resetOutputGrad, frameSize, batchSize, active_gate); - } - - if (grad.prevOutGrad && value.prevOutValue) { - hl_matrix_mul(grad.gateGrad, HPPL_OP_N, - value.gateWeight, HPPL_OP_T, - grad.prevOutGrad, - batchSize, frameSize, 2*frameSize, - /*alpha = */ 1, /*beta = */ 1, - 3*frameSize, 2*frameSize, frameSize); - if (grad.gateWeightGrad) { - hl_matrix_mul(value.prevOutValue, HPPL_OP_T, - grad.gateGrad, HPPL_OP_N, - grad.gateWeightGrad, - frameSize, 2*frameSize, batchSize, - /*alpha = */ 1, /*beta = */ 1, - frameSize, 3*frameSize, 2*frameSize); - } - } - - CHECK_SYNC("hl_gpu_gru_backward failed"); -} - -#else - -template -void hl_gpu_gru_forward(OpResetOutput opResetOutput, - OpFinalOutput opFinalOutput, - hl_gru_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) {} - -template -void hl_gpu_gru_backward(OpStateGrad opStateGrad, - OpResetGrad opResetGrad, - hl_gru_value value, - hl_gru_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) {} - -#endif - -#endif /* HL_GPU_GRU_CUH_ */ diff --git a/paddle/cuda/include/hl_gpu_lstm.cuh b/paddle/cuda/include/hl_gpu_lstm.cuh deleted file mode 100644 index 92517a44d2353a42d905708fc9aa98727a13a9e9..0000000000000000000000000000000000000000 --- a/paddle/cuda/include/hl_gpu_lstm.cuh +++ /dev/null @@ -1,300 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - - -#ifndef HL_GPU_LSTM_CUH_ -#define HL_GPU_LSTM_CUH_ - -#ifdef __NVCC__ - -#include "paddle/utils/Logging.h" -#include "hl_device_functions.cuh" - -/* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) - */ -template -__global__ void KeLstmForward(Op op, - hl_lstm_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - value.gateValue += batchIdx * frameSize * 4; - value.outputValue += batchIdx * frameSize; - value.stateValue += batchIdx * frameSize; - value.stateActiveValue += batchIdx * frameSize; - } - - real rState; - real rPrevState = 0; - real rStateAtv; - real rOut; - real rValueIn; - real rValueIg; - real rValueFg; - real rValueOg; - real rCheckI = value.checkIg[frameIdx]; - real rCheckF = value.checkFg[frameIdx]; - real rCheckO = value.checkOg[frameIdx]; - - rValueIn = value.gateValue[frameIdx]; - rValueIg = value.gateValue[frameIdx + frameSize]; - rValueFg = value.gateValue[frameIdx + frameSize * 2]; - rValueOg = value.gateValue[frameIdx + frameSize * 3]; - - if (value.prevStateValue) { - if (isBatch) value.prevStateValue += batchIdx * frameSize; - rPrevState = value.prevStateValue[frameIdx]; - } - - op(rValueIn, - rValueIg, - rValueFg, - rValueOg, - rPrevState, - rState, - rStateAtv, - rOut, - rCheckI, - rCheckF, - rCheckO, - hppl::gpu::forward[active_node], - hppl::gpu::forward[active_gate], - hppl::gpu::forward[active_state]); - - value.gateValue[frameIdx] = rValueIn; - value.gateValue[frameIdx + frameSize] = rValueIg; - value.gateValue[frameIdx + frameSize * 2] = rValueFg; - value.gateValue[frameIdx + frameSize * 3] = rValueOg; - - value.stateValue[frameIdx] = rState; - value.stateActiveValue[frameIdx] = rStateAtv; - value.outputValue[frameIdx] = rOut; -} - -/* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) - */ -template -__global__ void KeLstmBackward(Op op, - hl_lstm_value value, - hl_lstm_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - value.gateValue += batchIdx * frameSize * 4; - value.stateValue += batchIdx * frameSize; - value.stateActiveValue += batchIdx * frameSize; - grad.gateGrad += batchIdx * frameSize * 4; - grad.stateGrad += batchIdx * frameSize; - grad.outputGrad += batchIdx * frameSize; - } - - real rValueIn; - real rValueIg; - real rValueFg; - real rValueOg; - real rGradIn; - real rGradIg; - real rGradFg; - real rGradOg; - real rPrevState = 0; - real rPrevStateGrad; - real rState; - real rStateGrad; - real rStateAtv; - real rOutputGrad; - real rCheckI = value.checkIg[frameIdx]; - real rCheckF = value.checkFg[frameIdx]; - real rCheckO = value.checkOg[frameIdx]; - real rCheckIGrad; - real rCheckFGrad; - real rCheckOGrad; - - rValueIn = value.gateValue[frameIdx]; - rValueIg = value.gateValue[frameIdx + frameSize]; - rValueFg = value.gateValue[frameIdx + frameSize * 2]; - rValueOg = value.gateValue[frameIdx + frameSize * 3]; - rState = value.stateValue[frameIdx]; - rStateAtv = value.stateActiveValue[frameIdx]; - rOutputGrad = grad.outputGrad[frameIdx]; - rStateGrad = grad.stateGrad[frameIdx]; - - if (value.prevStateValue) { - if (isBatch) value.prevStateValue += batchIdx * frameSize; - rPrevState = value.prevStateValue[frameIdx]; - } - - op(rValueIn, - rValueIg, - rValueFg, - rValueOg, - rGradIn, - rGradIg, - rGradFg, - rGradOg, - rPrevState, - rPrevStateGrad, - rState, - rStateGrad, - rStateAtv, - rOutputGrad, - rCheckI, - rCheckF, - rCheckO, - rCheckIGrad, - rCheckFGrad, - rCheckOGrad, - hppl::gpu::backward[active_node], - hppl::gpu::backward[active_gate], - hppl::gpu::backward[active_state]); - - grad.gateGrad[frameIdx] = rGradIn; - grad.gateGrad[frameIdx + frameSize ] = rGradIg; - grad.gateGrad[frameIdx + frameSize * 2] = rGradFg; - grad.gateGrad[frameIdx + frameSize * 3] = rGradOg; - grad.stateGrad[frameIdx] = rStateGrad; - if (grad.prevStateGrad) { - if (isBatch) grad.prevStateGrad += batchIdx * frameSize; - grad.prevStateGrad[frameIdx] = rPrevStateGrad; - } - - if (isBatch) { - if (value.prevStateValue) { - if (grad.checkIgGrad) paddle::paddleAtomicAdd(grad.checkIgGrad+frameIdx, rCheckIGrad); - if (grad.checkFgGrad) paddle::paddleAtomicAdd(grad.checkFgGrad+frameIdx, rCheckFGrad); - } - if (grad.checkOgGrad) paddle::paddleAtomicAdd(grad.checkOgGrad+frameIdx, rCheckOGrad); - } else { - if (value.prevStateValue) { - if (grad.checkIgGrad) grad.checkIgGrad[frameIdx] += rCheckIGrad; - if (grad.checkFgGrad) grad.checkFgGrad[frameIdx] += rCheckFGrad; - } - if (grad.checkOgGrad) grad.checkOgGrad[frameIdx] += rCheckOGrad; - } -} - -template -void hl_gpu_lstm_forward(Op op, - hl_lstm_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - dim3 threads; - dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); - } else { - /* framePerBlock = 32 batchPerBlock = 32 */ - threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); - } - - if (batchSize == 1) { - KeLstmForward - <<>>(op, value, - frameSize, batchSize, active_node, active_gate, active_state); - } else { - KeLstmForward - <<>>(op, value, - frameSize, batchSize, active_node, active_gate, active_state); - } - - CHECK_SYNC("hl_gpu_lstm_forward failed"); -} - -template -void hl_gpu_lstm_backward(Op op, - hl_lstm_value value, - hl_lstm_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - dim3 threads; - dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); - } else { - /* framePerBlock = 32 batchPerBlock = 32 */ - threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); - } - - if (batchSize == 1) { - KeLstmBackward - <<>>(op, value, grad, - frameSize, batchSize, active_node, active_gate, active_state); - } else { - KeLstmBackward - <<>>(op, value, grad, - frameSize, batchSize, active_node, active_gate, active_state); - } - - CHECK_SYNC("hl_gpu_lstm_backward failed"); -} - -#else - -template -void hl_gpu_lstm_forward(Op op, - hl_lstm_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) {} - -template -void hl_gpu_lstm_backward(Op op, - hl_lstm_value value, - hl_lstm_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) {} - -#endif - -#endif /* HL_GPU_LSTM_CUH_ */ diff --git a/paddle/cuda/include/hl_gpu_matrix_kernel.cuh b/paddle/cuda/include/hl_gpu_matrix_kernel.cuh deleted file mode 100644 index 0db023ce3745f95ced8b3a33a1d6bcb20066b2ef..0000000000000000000000000000000000000000 --- a/paddle/cuda/include/hl_gpu_matrix_kernel.cuh +++ /dev/null @@ -1,629 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - - - -#ifndef HL_GPU_MATRIX_KERNEL_CUH_ -#define HL_GPU_MATRIX_KERNEL_CUH_ - -#include -#include "paddle/utils/Logging.h" -#include "hl_base.h" - -#ifdef __NVCC__ -/* gpu apply interface */ - -template -__global__ void KeEltWiseUnaryOp(T* A_d, const int border, Op op) { - const int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < border) { - op.gpuOperator(A_d[idx]); - } -} - -template -__global__ void KeEltWiseUnaryOp(T* A_d, - int dimM, - int dimN, - int lda, - Op op) { - const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; - const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; - for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { - for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { - op.gpuOperator(A_d[i * lda + j]); - } - } -} - -template -__global__ void KeEltWiseBinaryOp(T* A_d, T *B_d, const int border, Op op) { - const int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < border) { - op.gpuOperator(A_d[idx], B_d[idx]); - } -} - -template -__global__ void KeEltWiseBinaryOp(T *A_d, - T *B_d, - int dimM, - int dimN, - int lda, - int ldb, - Op op) { - const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; - const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; - for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { - for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { - if (BAsRowVector == 0 && BAsColVector == 0) { - op.gpuOperator(A_d[i * lda + j], B_d[i * ldb + j]); - } else if (BAsRowVector == 1 && BAsColVector == 0) { - op.gpuOperator(A_d[i * lda + j], B_d[j]); - } else if (BAsRowVector == 0 && BAsColVector == 1) { - op.gpuOperator(A_d[i * lda + j], B_d[i * ldb]); - } else { - op.gpuOperator(A_d[i * lda + j], B_d[0]); - } - } - } -} - -template -__global__ void KeEltWiseTernaryOp(T* A_d, - T *B_d, - T *C_d, - const int border, - Op op) { - const int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < border) { - op.gpuOperator(A_d[idx], B_d[idx], C_d[idx]); - } -} - -template -__global__ void KeEltWiseTernaryOp(T* A_d, - T* B_d, - T* C_d, - int dimM, - int dimN, - int lda, - int ldb, - int ldc, - Op op) { - const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; - const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; - for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { - for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { - if (CAsRowVector == 0 && CAsColVector == 0) { - op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[i*ldc + j]); - } else if (CAsRowVector == 1 && CAsColVector == 0) { - op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[j]); - } else if (CAsRowVector == 0 && CAsColVector == 1) { - op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[i*ldc]); - } else { - op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[0]); - } - } - } -} - -template -__global__ void KeEltWiseQuaternaryOp(T* A_d, - T* B_d, - T* C_d, - T* D_d, - const int border, - Op op) { - const int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < border) { - op.gpuOperator(A_d[idx], B_d[idx], C_d[idx], D_d[idx]); - } -} - -template -__global__ void KeEltWiseQuaternaryOp(T* A_d, - T* B_d, - T* C_d, - T* D_d, - int dimM, - int dimN, - int lda, - int ldb, - int ldc, - int ldd, - Op op) { - const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; - const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; - for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { - for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { - op.gpuOperator(A_d[i*lda + j], - B_d[i*ldb + j], C_d[i*ldc + j], D_d[i*ldd + j]); - } - } -} - -/** - * @brief gpu element wise unary operator. - */ -template -void hl_gpu_apply_unary_op(Op op, T* A_d, int dimM, int dimN, int lda) { - CHECK_NOTNULL(A_d); - - if (dimM == 1 || dimN == lda) { - int size = dimM * dimN; - int blockSize = size <= 1024 ? size : 1024; - int gridSize = (size + 1024 - 1) / 1024; - KeEltWiseUnaryOp<<>> - (A_d, size, op); - } else { - int blockSizeY = std::min(32, dimM); - int blockSizeX = (32 / blockSizeY) * 32; - int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); - int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); - dim3 threads(blockSizeX, blockSizeY); - dim3 grid(gridSizeX, gridSizeY); - KeEltWiseUnaryOp<<>> - (A_d, dimM, dimN, lda, op); - } - - CHECK_SYNC("hl_gpu_apply_unary_op failed"); -} - -/** - * @brief gpu element wise binary operator. - */ -template -void hl_gpu_apply_binary_op(Op op, - T* A_d, - T* B_d, - int dimM, - int dimN, - int lda, - int ldb) { - CHECK_NOTNULL(A_d); - - if ((BAsRowVector == 0 && BAsColVector == 0) && - ((dimM == 1) || (dimN == lda && dimN == ldb))) { - int size = dimM * dimN; - int blockSize = size <= 1024 ? size : 1024; - int gridSize = (size + 1024 - 1) / 1024; - KeEltWiseBinaryOp<<>> - (A_d, B_d, size, op); - } else { - int blockSizeY = std::min(32, dimM); - int blockSizeX = (32 / blockSizeY) * 32; - int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); - int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); - dim3 threads(blockSizeX, blockSizeY); - dim3 grid(gridSizeX, gridSizeY); - KeEltWiseBinaryOp - <<>> - (A_d, B_d, dimM, dimN, lda, ldb, op); - } - - CHECK_SYNC("hl_gpu_apply_binary_op failed"); -} - -/** - * @brief gpu element wise ternary operator. - */ -template -void hl_gpu_apply_ternary_op(Op op, - T* A_d, - T* B_d, - T* C_d, - int dimM, - int dimN, - int lda, - int ldb, - int ldc) { - CHECK_NOTNULL(A_d); - - if ((CAsRowVector == 0 && CAsColVector == 0) && - ((dimM == 1) || (dimN == lda && dimN == ldb && dimN == ldc))) { - int size = dimM * dimN; - int blockSize = size <= 1024 ? size : 1024; - int gridSize = (size + 1024 - 1) / 1024; - KeEltWiseTernaryOp<<>> - (A_d, B_d, C_d, size, op); - } else { - int blockSizeY = std::min(32, dimM); - int blockSizeX = (32 / blockSizeY) * 32; - int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); - int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); - dim3 threads(blockSizeX, blockSizeY); - dim3 grid(gridSizeX, gridSizeY); - KeEltWiseTernaryOp - <<>> - (A_d, B_d, C_d, dimM, dimN, lda, ldb, ldc, op); - } - - CHECK_SYNC("hl_gpu_apply_ternary_op failed"); -} - - -/** - * @brief gpu element wise quaternary operator. - */ -template -void hl_gpu_apply_quaternary_op(Op op, - T* A_d, - T* B_d, - T* C_d, - T* D_d, - int dimM, - int dimN, - int lda, - int ldb, - int ldc, - int ldd) { - CHECK_NOTNULL(A_d); - - if ((dimM == 1) || - (dimN == lda && dimN == ldb && dimN == ldc && dimN == ldd)) { - int size = dimM * dimN; - int blockSize = size <= 1024 ? size : 1024; - int gridSize = (size + 1024 - 1) / 1024; - KeEltWiseQuaternaryOp<<>> - (A_d, B_d, C_d, D_d, size, op); - } else { - int blockSizeY = std::min(32, dimM); - int blockSizeX = (32 / blockSizeY) * 32; - int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); - int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); - dim3 threads(blockSizeX, blockSizeY); - dim3 grid(gridSizeX, gridSizeY); - KeEltWiseQuaternaryOp<<>> - (A_d, B_d, C_d, D_d, dimM, dimN, lda, ldb, ldc, ldd, op); - } - - CHECK_SYNC("hl_gpu_apply_quaternary_op failed"); -} - -#else - -template -void hl_gpu_apply_unary_op(Op op, T* A_d, int dimM, int dimN, int lda) {} - -template -void hl_gpu_apply_binary_op(Op op, - T* A_d, - T* B_d, - int dimM, - int dimN, - int lda, - int ldb) {} - -template -void hl_gpu_apply_ternary_op(Op op, - T* A_d, - T* B_d, - T* C_d, - int dimM, - int dimN, - int lda, - int ldb, - int ldc) {} - -template -void hl_gpu_apply_quaternary_op(Op op, - T* A_d, - T* B_d, - T* C_d, - T* D_d, - int dimM, - int dimN, - int lda, - int ldb, - int ldc, - int ldd) {} -#endif - -#ifdef __NVCC__ -/** - * @brief matrix row operator. - */ - -template -__device__ __inline__ real sumRow(Agg agg, Op op, - int idx, int blockSize, - int dimN, real *A) { - real tmp = agg.init(); - int cnt = (dimN + blockSize -1) / blockSize; - for (int i = 0; i < cnt && idx < dimN; i++) { - tmp = agg(tmp, op(A[idx])); - idx += blockSize; - } - return tmp; -} - -template -__device__ __inline__ real sumRow(Agg agg, Op op, - int idx, int blockSize, - int dimN, real *A, real *B) { - real tmp = agg.init(); - int cnt = (dimN + blockSize -1) / blockSize; - for (int i = 0; i < cnt && idx < dimN; i++) { - tmp = agg(tmp, op(A[idx], B[idx])); - idx += blockSize; - } - return tmp; -} - -template -__device__ __inline__ void aggRow(Agg agg, real *row, int size, int tid) { - for (int stride = size/2; stride > 0; stride = stride/2) { - if (tid < stride) { - row[tid] = agg(row[tid], row[tid + stride]); - } - __syncthreads(); - } -} - -template -__global__ void KeMatrixRowOp(Agg agg, Op op, Saver sv, - int dimN, - real *dst, int ld, - real *A, int lda) { - __shared__ real row_s[blockSize]; - int rowId = blockIdx.x + blockIdx.y*gridDim.x; - int tid = threadIdx.x; - - A += rowId*lda; - row_s[tid] = sumRow(agg, op, tid, blockSize, dimN, A); - __syncthreads(); - - aggRow(agg, row_s, blockSize, tid); - __syncthreads(); - - if (tid == 0) { - dst[rowId*ld] = sv(dst[rowId*ld], row_s[0]); - } -} - -template -__global__ void KeMatrixRowOp(Agg agg, Op op, Saver sv, - int dimN, - real *dst, int ld, - real *A, int lda, - real *B, int ldb) { - __shared__ real row_s[blockSize]; - int rowId = blockIdx.x + blockIdx.y*gridDim.x; - int tid = threadIdx.x; - - A += rowId*lda; - B += rowId*ldb; - row_s[tid] = sumRow(agg, op, tid, blockSize, dimN, A, B); - __syncthreads(); - - aggRow(agg, row_s, blockSize, tid); - __syncthreads(); - - if (tid == 0) { - dst[rowId*ld] = sv(dst[rowId*ld], row_s[0]); - } -} - -/** - * @brief matrix column operator. - */ -template -__device__ __inline__ real sumCol(Agg agg, Op op, - int index, int stride, - int dimM, real *A, int lda) { - real tmp = agg.init(); - for (; index < dimM;) { - tmp = agg(tmp, op(A[index*lda])); - index += stride; - } - return tmp; -} - -template -__device__ __inline__ real sumCol(Agg agg, Op op, - int index, int stride, int dimM, - real *A, int lda, real *B, int ldb) { - real tmp = agg.init(); - for (; index < dimM;) { - tmp = agg(tmp, op(A[index*lda], B[index*ldb])); - index += stride; - } - return tmp; -} - -template -__global__ void KeMatrixColumnOp(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, - real *A, int lda) { - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (rowIdx < dimN) { - A += rowIdx; - real tmp = sumCol(agg, op, 0, 1, dimM, A, lda); - dst[rowIdx] = sv(dst[rowIdx], tmp); - } -} - -template -__global__ void KeMatrixColumnOp_S(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, - real *A, int lda) { - __shared__ real col_s[blockDimX*blockDimY]; - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - - if (rowIdx < dimN) { - A += rowIdx; - real tmp = sumCol(agg, op, threadIdx.y, blockDimY, dimM, A, lda); - col_s[threadIdx.x + threadIdx.y*blockDimX] = tmp; - } - __syncthreads(); - - if (rowIdx < dimN) { - if (threadIdx.y ==0) { - real tmp = agg.init(); - for (int i=0; i < blockDimY; i++) { - tmp = agg(tmp, col_s[threadIdx.x + i*blockDimX]); - } - dst[rowIdx] = sv(dst[rowIdx], tmp); - } - } -} - -template -__global__ void KeMatrixColumnOp(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, - real *A, int lda, - real *B, int ldb) { - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (rowIdx < dimN) { - A += rowIdx; - B += rowIdx; - real tmp = sumCol(agg, op, 0, 1, dimM, A, lda, B, ldb); - dst[rowIdx] = sv(dst[rowIdx], tmp); - } -} - -template -__global__ void KeMatrixColumnOp_S(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, - real *A, int lda, - real *B, int ldb) { - __shared__ real col_s[blockDimX*blockDimY]; - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - - if (rowIdx < dimN) { - A += rowIdx; - B += rowIdx; - real tmp = sumCol(agg, op, - threadIdx.y, blockDimY, dimM, A, lda, B, ldb); - col_s[threadIdx.x + threadIdx.y*blockDimX] = tmp; - } - __syncthreads(); - - if (rowIdx < dimN) { - if (threadIdx.y ==0) { - real tmp = agg.init(); - for (int i=0; i < blockDimY; i++) { - tmp = agg(tmp, col_s[threadIdx.x + i*blockDimX]); - } - dst[rowIdx] = sv(dst[rowIdx], tmp); - } - } -} - -#endif - -template -void hl_gpu_matrix_row_op(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, int ld, - real *A, int lda) { -#ifdef __NVCC__ - CHECK_NOTNULL(dst); - CHECK_NOTNULL(A); - - int blocksX = dimM; - int blocksY = 1; - dim3 threads(128, 1); - dim3 grid(blocksX, blocksY); - KeMatrixRowOp<<< grid, threads, 0, STREAM_DEFAULT >>> - (agg, op, sv, dimN, dst, ld, A, lda); - - CHECK_SYNC("hl_matrix_row_op failed"); -#endif -} - -template -void hl_gpu_matrix_row_op(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, int ld, - real *A, int lda, - real *B, int ldb) { -#ifdef __NVCC__ - CHECK_NOTNULL(dst); - CHECK_NOTNULL(A); - - int blocksX = dimM; - int blocksY = 1; - dim3 threads(128, 1); - dim3 grid(blocksX, blocksY); - KeMatrixRowOp<<< grid, threads, 0, STREAM_DEFAULT >>> - (agg, op, sv, dimN, dst, ld, A, lda, B, ldb); - - CHECK_SYNC("hl_matrix_row_op failed"); -#endif -} - -template -void hl_gpu_matrix_column_op(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, - real *A, int lda) { -#ifdef __NVCC__ - if (dimN >= 8192) { - int blocksX = (dimN + 128 -1) / 128; - int blocksY = 1; - dim3 threads(128, 1); - dim3 grid(blocksX, blocksY); - KeMatrixColumnOp - <<< grid, threads, 0, STREAM_DEFAULT >>> - (agg, op, sv, dimM, dimN, dst, A, lda); - } else { - int blocksX = (dimN + 32 -1) / 32; - int blocksY = 1; - dim3 threads(32, 32); - dim3 grid(blocksX, blocksY); - KeMatrixColumnOp_S - <<< grid, threads, 0, STREAM_DEFAULT>>> - (agg, op, sv, dimM, dimN, dst, A, lda); - } - - CHECK_SYNC("hl_matrix_column_op failed"); -#endif -} - -template -void hl_gpu_matrix_column_op(Agg agg, Op op, Saver sv, - int dimM, int dimN, - real *dst, - real *A, int lda, - real *B, int ldb) { -#ifdef __NVCC__ - if (dimN >= 8192) { - int blocksX = (dimN + 128 -1) / 128; - int blocksY = 1; - dim3 threads(128, 1); - dim3 grid(blocksX, blocksY); - KeMatrixColumnOp - <<< grid, threads, 0, STREAM_DEFAULT >>> - (agg, op, sv, dimM, dimN, dst, A, lda, B, ldb); - } else { - int blocksX = (dimN + 32 -1) / 32; - int blocksY = 1; - dim3 threads(32, 32); - dim3 grid(blocksX, blocksY); - KeMatrixColumnOp_S - <<< grid, threads, 0, STREAM_DEFAULT>>> - (agg, op, sv, dimM, dimN, dst, A, lda, B, ldb); - } - - CHECK_SYNC("hl_matrix_column_op failed"); -#endif -} - -#endif /* HL_GPU_MATRIX_KERNEL_CUH_ */ diff --git a/paddle/cuda/src/hl_cuda_aggregate.cu b/paddle/cuda/src/hl_cuda_aggregate.cu deleted file mode 100644 index d30c264127f47da9a48acb71c59cb9e134ced127..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_aggregate.cu +++ /dev/null @@ -1,293 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_aggregate.h" -#include "hl_base.h" -#include "hl_cuda.h" -#include "hl_cuda.ph" -#include "hl_matrix_base.cuh" -#include "hl_thread.ph" -#include "paddle/utils/Logging.h" - -/** - * @brief matrix row operator. - */ -template -__global__ void KeMatrixRowOp(Agg agg, real *E, real *Sum, int dimN) { - __shared__ real sum_s[blockSize]; - int cnt = (dimN + blockSize - 1) / blockSize; - int rowId = blockIdx.x + blockIdx.y * gridDim.x; - int index = rowId * dimN; - int tid = threadIdx.x; - int lmt = tid; - - real tmp = agg.init(); - for (int ii = 0; ii < cnt && lmt < dimN; ii++) { - tmp = agg(tmp, E[index + lmt]); - lmt += blockSize; - } - sum_s[tid] = tmp; - __syncthreads(); - - for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { - if (tid < stride) { - sum_s[tid] = agg(sum_s[tid], sum_s[tid + stride]); - } - __syncthreads(); - } - __syncthreads(); - - if (tid == 0) { - Sum[rowId] = sum_s[0]; - } -} - -template -void hl_matrix_row_op(Agg agg, real *A_d, real *C_d, int dimM, int dimN) { - int blocksX = dimM; - int blocksY = 1; - dim3 threads(128, 1); - dim3 grid(blocksX, blocksY); - - KeMatrixRowOp<<>>( - agg, A_d, C_d, dimN); -} - -void hl_matrix_row_sum(real *A_d, real *C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - hl_matrix_row_op(aggregate::sum(), A_d, C_d, dimM, dimN); - CHECK_SYNC("hl_matrix_row_sum failed"); -} - -void hl_matrix_row_max(real *A_d, real *C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - hl_matrix_row_op(aggregate::max(), A_d, C_d, dimM, dimN); - CHECK_SYNC("hl_matrix_row_max failed"); -} - -void hl_matrix_row_min(real *A_d, real *C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - hl_matrix_row_op(aggregate::min(), A_d, C_d, dimM, dimN); - CHECK_SYNC("hl_matrix_row_min failed"); -} - -/** - * @brief matrix column operator. - */ -template -__global__ void KeMatrixColumnOp( - Agg agg, real *E, real *Sum, int dimM, int dimN) { - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - real tmp = agg.init(); - if (rowIdx < dimN) { - for (int index = 0; index < dimM; index++) { - tmp = agg(tmp, E[dimN * index + rowIdx]); - } - Sum[rowIdx] = tmp; - } -} - -template -__global__ void KeMatrixColumnOp_S( - Agg agg, real *E, real *Sum, int dimM, int dimN) { - __shared__ real _sum[blockDimX * blockDimY]; - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - int index = threadIdx.y; - - real tmp = agg.init(); - if (rowIdx < dimN) { - for (; index < dimM;) { - tmp = agg(tmp, E[dimN * index + rowIdx]); - index += blockDimY; - } - } - _sum[threadIdx.x + threadIdx.y * blockDimX] = tmp; - __syncthreads(); - - if (rowIdx < dimN) { - if (threadIdx.y == 0) { - real tmp = agg.init(); - for (int i = 0; i < blockDimY; i++) { - tmp = agg(tmp, _sum[threadIdx.x + i * blockDimX]); - } - Sum[rowIdx] = tmp; - } - } -} - -template -void hl_matrix_column_op(Agg agg, real *A_d, real *C_d, int dimM, int dimN) { - if (dimN >= 8192) { - int blocksX = (dimN + 128 - 1) / 128; - int blocksY = 1; - dim3 threads(128, 1); - dim3 grid(blocksX, blocksY); - KeMatrixColumnOp<<>>( - agg, A_d, C_d, dimM, dimN); - } else { - int blocksX = (dimN + 32 - 1) / 32; - int blocksY = 1; - dim3 threads(32, 32); - dim3 grid(blocksX, blocksY); - KeMatrixColumnOp_S<<>>( - agg, A_d, C_d, dimM, dimN); - } - - return; -} - -void hl_matrix_column_sum(real *A_d, real *C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - hl_matrix_column_op(aggregate::sum(), A_d, C_d, dimM, dimN); - - CHECK_SYNC("hl_matrix_column_sum failed"); -} - -void hl_matrix_column_max(real *A_d, real *C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - hl_matrix_column_op(aggregate::max(), A_d, C_d, dimM, dimN); - - CHECK_SYNC("hl_matrix_column_max failed"); -} - -void hl_matrix_column_min(real *A_d, real *C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - hl_matrix_column_op(aggregate::min(), A_d, C_d, dimM, dimN); - - CHECK_SYNC("hl_matrix_column_min failed"); -} - -template -__global__ void KeVectorSum(real *E, real *Sum, int dimM) { - __shared__ double sum_s[blockSize]; - int tid = threadIdx.x; - int index = blockIdx.y * blockDim.x + threadIdx.x; - - sum_s[tid] = 0.0f; - while (index < dimM) { - sum_s[tid] += E[index]; - index += blockDim.x * gridDim.y; - } - __syncthreads(); - - for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { - if (tid < stride) { - sum_s[tid] += sum_s[tid + stride]; - } - __syncthreads(); - } - __syncthreads(); - - if (tid == 0) { - Sum[blockIdx.y] = sum_s[0]; - } -} - -void hl_vector_sum(real *A_d, real *C_h, int dimM) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_h); - - int blockSize = 128; - int gridSize = 128; - int blocksX = 1; - int blocksY = gridSize; - dim3 threads(blockSize, 1); - dim3 grid(blocksX, blocksY); - - struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; - hl_event_t hl_event = &hl_event_st; - while (!hl_cuda_event_is_ready(hl_event)) { - } - - KeVectorSum<128><<>>( - A_d, t_resource.gpu_mem, dimM); - KeVectorSum<128><<<1, threads, 0, STREAM_DEFAULT>>>( - t_resource.gpu_mem, t_resource.cpu_mem, 128); - - hl_memcpy_async(C_h, t_resource.cpu_mem, sizeof(real), HPPL_STREAM_DEFAULT); - hl_stream_record_event(HPPL_STREAM_DEFAULT, hl_event); - - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - cudaError_t err = (cudaError_t)hl_get_device_last_error(); - CHECK_EQ(cudaSuccess, err) << "CUDA error: " - << hl_get_device_error_string((size_t)err); -} - -template -__global__ void KeVectorAbsSum(real *E, real *Sum, int dimM) { - __shared__ double sum_s[blockSize]; - int tid = threadIdx.x; - int index = blockIdx.y * blockDim.x + threadIdx.x; - - sum_s[tid] = 0.0f; - while (index < dimM) { - sum_s[tid] += abs(E[index]); - index += blockDim.x * gridDim.y; - } - __syncthreads(); - - for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { - if (tid < stride) { - sum_s[tid] += sum_s[tid + stride]; - } - __syncthreads(); - } - __syncthreads(); - - if (tid == 0) { - Sum[blockIdx.y] = sum_s[0]; - } -} - -void hl_vector_abs_sum(real *A_d, real *C_h, int dimM) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_h); - - int blockSize = 128; - int gridSize = 128; - int blocksX = 1; - int blocksY = gridSize; - dim3 threads(blockSize, 1); - dim3 grid(blocksX, blocksY); - - struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; - hl_event_t hl_event = &hl_event_st; - while (!hl_cuda_event_is_ready(hl_event)) { - } - - KeVectorAbsSum<128><<>>( - A_d, t_resource.gpu_mem, dimM); - KeVectorAbsSum<128><<<1, threads, 0, STREAM_DEFAULT>>>( - t_resource.gpu_mem, t_resource.cpu_mem, 128); - - hl_memcpy_async(C_h, t_resource.cpu_mem, sizeof(real), HPPL_STREAM_DEFAULT); - hl_stream_record_event(HPPL_STREAM_DEFAULT, hl_event); - - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - cudaError_t err = (cudaError_t)hl_get_device_last_error(); - CHECK_EQ(cudaSuccess, err) << "CUDA error: " - << hl_get_device_error_string((size_t)err); -} diff --git a/paddle/cuda/src/hl_cuda_cublas.cc b/paddle/cuda/src/hl_cuda_cublas.cc deleted file mode 100644 index 975df4287894090799c44bc0a4e9e08e4144e68f..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_cublas.cc +++ /dev/null @@ -1,400 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_cuda_cublas.h" -#include -#include "hl_cuda.h" -#include "hl_thread.ph" -#include "paddle/utils/DynamicLoader.h" -#include "paddle/utils/Logging.h" - -namespace dynload { - -std::once_flag cublas_dso_flag; -void *cublas_dso_handle = nullptr; - -/** - * The following macro definition can generate structs - * (for each function) to dynamic load cublas routine - * via operator overloading. - * - * note: default dynamic linked libs - */ -#ifdef PADDLE_USE_DSO -#define DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - cublasStatus_t operator()(Args... args) { \ - typedef cublasStatus_t (*cublasFunc)(Args...); \ - std::call_once(cublas_dso_flag, GetCublasDsoHandle, &cublas_dso_handle); \ - void *p_##__name = dlsym(cublas_dso_handle, #__name); \ - return reinterpret_cast(p_##__name)(args...); \ - } \ - } __name; // struct DynLoad__##__name -#else -#define DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - cublasStatus_t operator()(Args... args) { \ - return __name(args...); \ - } \ - } __name; // struct DynLoad__##__name -#endif - -#define DYNAMIC_LOAD_CUBLAS_V2_WRAP(__name) DYNAMIC_LOAD_CUBLAS_WRAP(__name) - -// include all needed cublas functions in HPPL -// clang-format off -#define CUBLAS_BLAS_ROUTINE_EACH(__macro) \ - __macro(cublasSgemv) \ - __macro(cublasDgemv) \ - __macro(cublasSgemm) \ - __macro(cublasDgemm) \ - __macro(cublasSgeam) \ - __macro(cublasDgeam) \ - -DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasCreate) -DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasDestroy) -DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetStream) -DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetPointerMode) -DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasGetPointerMode) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgemmBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgemmBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasCgemmBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasZgemmBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetrfBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetriBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetrfBatched) -DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetriBatched) -CUBLAS_BLAS_ROUTINE_EACH(DYNAMIC_LOAD_CUBLAS_V2_WRAP) - -#undef DYNAMIC_LOAD_CUBLAS_WRAP -#undef DYNAMIC_LOAD_CUBLAS_V2_WRAP -#undef CUBLAS_BLAS_ROUTINE_EACH - -} /* namespace dynload */ - -// clang-format on -#ifndef PADDLE_TYPE_DOUBLE -#define CUBLAS_GEAM dynload::cublasSgeam -#define CUBLAS_GEMV dynload::cublasSgemv -#define CUBLAS_GEMM dynload::cublasSgemm -#define CUBLAS_GETRF dynload::cublasSgetrfBatched -#define CUBLAS_GETRI dynload::cublasSgetriBatched -#else -#define CUBLAS_GEAM dynload::cublasDgeam -#define CUBLAS_GEMV dynload::cublasDgemv -#define CUBLAS_GEMM dynload::cublasDgemm -#define CUBLAS_GETRF dynload::cublasDgetrfBatched -#define CUBLAS_GETRI dynload::cublasDgetriBatched -#endif - -const char *hl_cublas_get_error_string(cublasStatus_t status) { - switch (status) { - case CUBLAS_STATUS_NOT_INITIALIZED: - return "[cublas status]: not initialized"; - case CUBLAS_STATUS_ALLOC_FAILED: - return "[cublas status]: allocate failed"; - case CUBLAS_STATUS_INVALID_VALUE: - return "[cublas status]: invalid value"; - case CUBLAS_STATUS_ARCH_MISMATCH: - return "[cublas status]: arch mismatch"; - case CUBLAS_STATUS_MAPPING_ERROR: - return "[cublas status]: mapping error"; - case CUBLAS_STATUS_EXECUTION_FAILED: - return "[cublas status]: execution failed"; - case CUBLAS_STATUS_INTERNAL_ERROR: - return "[cublas status]: internal error"; - case CUBLAS_STATUS_SUCCESS: - return "[cublas status]: success"; - default: - return "[cublas status]: unknown error"; - } -} - -/** - * Check build-in cublas function using glog and it also - * support << operator for more details error info. - */ -cublasStatus_t g_cublasStat; -#define CHECK_CUBLAS(cublas_func) \ - g_cublasStat = cublas_func; \ - CHECK_EQ(CUBLAS_STATUS_SUCCESS, g_cublasStat) \ - << "Cublas Error: " << hl_cublas_get_error_string(g_cublasStat) << " " - -void hl_cublas_init(cublasHandle_t *cublas_handle, cudaStream_t stream) { - CHECK_CUBLAS(dynload::cublasCreate(cublas_handle)) - << "[cublas init] Cublas create handle faild!"; - - CHECK_CUBLAS(dynload::cublasSetStream(*cublas_handle, stream)) - << "[cublas init] Cublas set stream faild!"; -} - -void hl_matrix_transpose( - real *A_d, real *C_d, int dimM, int dimN, int lda, int ldc) { - real alpha = 1.0; - real beta = 0.0; - - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - CHECK_CUBLAS(CUBLAS_GEAM(t_resource.handle, - CUBLAS_OP_T, - CUBLAS_OP_N, - dimM, - dimN, - &alpha, - A_d, - lda, - &beta, - nullptr, - dimM, - C_d, - ldc)); - CHECK_SYNC("hl_matrix_transpose failed"); -} - -void hl_matrix_transpose(real *A_d, real *C_d, int dimM, int dimN) { - hl_matrix_transpose(A_d, C_d, dimM, dimN, dimN, dimM); -} - -void hl_matrix_inverse(real *A_d, real *C_d, int dimN, int lda, int ldc) { - /* Solve Ax = I */ - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - /* Step 1: Compute the LU decomposition of matrix A */ - real **inout_h = &A_d; - real **inout_d = (real **)hl_malloc_device(sizeof(real *)); - hl_memcpy(inout_d, inout_h, sizeof(real *)); - - int *pivot_d = (int *)hl_malloc_device(dimN * sizeof(int)); - int *info_d = (int *)t_resource.gpu_mem; - - /* Note: cublasSgetrfBatched is used to calculate a number of - small-sized matrices. There may be a better way to reconstruct - the API for better performance. - */ - CHECK_CUBLAS( - CUBLAS_GETRF(t_resource.handle, dimN, inout_d, lda, pivot_d, info_d, 1)); - - int info_h; - hl_memcpy(&info_h, info_d, sizeof(int)); - if (info_h != 0) { - LOG(FATAL) << "Factorization of matrix failed: matrix may be singular.\n"; - } - - /* Step 2: Compute the inverse of the matrix given its LU decomposition */ - real **out_h = &C_d; - real **out_d = (real **)hl_malloc_device(sizeof(real *)); - hl_memcpy(out_d, out_h, sizeof(real *)); - - CHECK_CUBLAS(CUBLAS_GETRI(t_resource.handle, - dimN, - (const real **)inout_d, - lda, - pivot_d, - out_d, - ldc, - info_d, - 1)); - - hl_memcpy(&info_h, info_d, sizeof(int)); - if (info_h != 0) { - LOG(FATAL) << "Inversion of matrix failed: matrix may be singular.\n"; - } - - hl_free_mem_device(inout_d); - hl_free_mem_device(pivot_d); - hl_free_mem_device(out_d); - - CHECK_SYNC("hl_matrix_inverse failed"); -} - -void hl_matrix_mul(real *A_d, - hl_trans_op_t transa, - real *B_d, - hl_trans_op_t transb, - real *C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta, - int lda, - int ldb, - int ldc) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - - if (dimN == 1 && dimM != 1 && dimK != 1 && transb == HPPL_OP_N) { - int m = (transa == HPPL_OP_N) ? dimM : dimK; - int n = (transa == HPPL_OP_N) ? dimK : dimM; - hl_matrix_mul_vector( - A_d, transa, B_d, C_d, m, n, alpha, beta, lda, ldb, ldc); - return; - } - - if (dimM == 1 && dimN != 1 && dimK != 1 && transa == HPPL_OP_N) { - int m = (transb == HPPL_OP_N) ? dimK : dimN; - int n = (transb == HPPL_OP_N) ? dimN : dimK; - hl_trans_op_t trans = (transb == HPPL_OP_N) ? HPPL_OP_T : HPPL_OP_N; - hl_matrix_mul_vector(B_d, trans, A_d, C_d, m, n, alpha, beta, ldb, 1, 1); - return; - } - - cublasStatus_t stat; - if ((HPPL_OP_N == transa) && (HPPL_OP_N == transb)) { - stat = CUBLAS_GEMM(t_resource.handle, - CUBLAS_OP_N, - CUBLAS_OP_N, - dimN, - dimM, - dimK, - &alpha, - B_d, - ldb, - A_d, - lda, - &beta, - C_d, - ldc); - } else if ((HPPL_OP_T == transa) && (HPPL_OP_N == transb)) { - stat = CUBLAS_GEMM(t_resource.handle, - CUBLAS_OP_N, - CUBLAS_OP_T, - dimN, - dimM, - dimK, - &alpha, - B_d, - ldb, - A_d, - lda, - &beta, - C_d, - ldc); - } else if ((HPPL_OP_N == transa) && (HPPL_OP_T == transb)) { - stat = CUBLAS_GEMM(t_resource.handle, - CUBLAS_OP_T, - CUBLAS_OP_N, - dimN, - dimM, - dimK, - &alpha, - B_d, - ldb, - A_d, - lda, - &beta, - C_d, - ldc); - } else { - LOG(FATAL) << "parameter transa error!"; - } - CHECK_EQ(stat, CUBLAS_STATUS_SUCCESS) << hl_cublas_get_error_string(stat); - CHECK_SYNC("hl_matrix_mul failed"); -} - -void hl_matrix_mul(real *A_d, - hl_trans_op_t transa, - real *B_d, - hl_trans_op_t transb, - real *C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta) { - int lda = (HPPL_OP_N == transa) ? dimK : dimM; - int ldb = (HPPL_OP_N == transb) ? dimN : dimK; - int ldc = dimN; - - hl_matrix_mul(A_d, - transa, - B_d, - transb, - C_d, - dimM, - dimN, - dimK, - alpha, - beta, - lda, - ldb, - ldc); -} - -void hl_matrix_mul_vector(real *A_d, - hl_trans_op_t trans, - real *B_d, - real *C_d, - int dimM, - int dimN, - real alpha, - real beta, - int lda, - int incb, - int incc) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - - cublasStatus_t stat; - if (HPPL_OP_N == trans) { - stat = CUBLAS_GEMV(t_resource.handle, - CUBLAS_OP_T, - dimN, - dimM, - &alpha, - A_d, - lda, - B_d, - incb, - &beta, - C_d, - incc); - } else if (HPPL_OP_T == trans) { - stat = CUBLAS_GEMV(t_resource.handle, - CUBLAS_OP_N, - dimN, - dimM, - &alpha, - A_d, - lda, - B_d, - incb, - &beta, - C_d, - incc); - } else { - LOG(FATAL) << "parameter transa error!"; - } - - CHECK_EQ(stat, CUBLAS_STATUS_SUCCESS) << hl_cublas_get_error_string(stat); - CHECK_SYNC("hl_matrix_mul_vector"); -} - -void hl_matrix_mul_vector(real *A_d, - hl_trans_op_t trans, - real *B_d, - real *C_d, - int dimM, - int dimN, - real alpha, - real beta) { - hl_matrix_mul_vector( - A_d, trans, B_d, C_d, dimM, dimN, alpha, beta, dimN, 1, 1); -} diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc deleted file mode 100644 index dfa935dcff9f7ae9f710d0f01a0217298d8cec04..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_cudnn.cc +++ /dev/null @@ -1,1117 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_cuda_cudnn.h" -#include -#include -#include "hl_cuda_cudnn.ph" -#include "hl_thread.ph" -#include "paddle/utils/DynamicLoader.h" -#include "paddle/utils/Logging.h" - -DEFINE_int32(cudnn_conv_workspace_limit_in_mb, - 4096, - "Specify cuDNN max workspace limit, in units MB, " - "4096MB=4GB by default."); - -namespace dynload { - -std::once_flag cudnn_dso_flag; -void* cudnn_dso_handle = nullptr; - -/** - * The following macro definition can generate structs - * (for each function) to dynamic load cudbnn routine - * via operator overloading: operator () - * - * note: default dynamic linked libs - **/ - -#ifdef PADDLE_USE_DSO - -#define DYNAMIC_LOAD_CUDNN_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - auto operator()(Args... args) -> decltype(__name(args...)) { \ - using cudnn_func = decltype(__name(args...)) (*)(Args...); \ - std::call_once(cudnn_dso_flag, GetCudnnDsoHandle, &cudnn_dso_handle); \ - void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ - return reinterpret_cast(p_##__name)(args...); \ - } \ - } __name; /* struct DynLoad__##__name */ - -#else - -#define DYNAMIC_LOAD_CUDNN_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - auto operator()(Args... args) -> decltype(__name(args...)) { \ - return __name(args...); \ - } \ - } __name; /* struct DynLoad__##__name */ - -#endif - -/** - * include all needed cudnn functions in HPPL - * different cudnn version has different interfaces - **/ -// clang-format off -#define CUDNN_DNN_ROUTINE_EACH(__macro) \ - __macro(cudnnSetTensor4dDescriptor) \ - __macro(cudnnSetTensor4dDescriptorEx) \ - __macro(cudnnGetConvolutionNdForwardOutputDim) \ - __macro(cudnnGetConvolutionForwardAlgorithm) \ - __macro(cudnnCreateTensorDescriptor) \ - __macro(cudnnDestroyTensorDescriptor) \ - __macro(cudnnCreateFilterDescriptor) \ - __macro(cudnnSetFilter4dDescriptor) \ - __macro(cudnnSetPooling2dDescriptor) \ - __macro(cudnnDestroyFilterDescriptor) \ - __macro(cudnnCreateConvolutionDescriptor) \ - __macro(cudnnCreatePoolingDescriptor) \ - __macro(cudnnDestroyPoolingDescriptor) \ - __macro(cudnnSetConvolution2dDescriptor) \ - __macro(cudnnDestroyConvolutionDescriptor) \ - __macro(cudnnCreate) \ - __macro(cudnnDestroy) \ - __macro(cudnnSetStream) \ - __macro(cudnnActivationForward) \ - __macro(cudnnConvolutionForward) \ - __macro(cudnnConvolutionBackwardBias) \ - __macro(cudnnGetConvolutionForwardWorkspaceSize) \ - __macro(cudnnTransformTensor) \ - __macro(cudnnPoolingForward) \ - __macro(cudnnPoolingBackward) \ - __macro(cudnnSoftmaxBackward) \ - __macro(cudnnSoftmaxForward) \ - __macro(cudnnGetVersion) \ - __macro(cudnnGetErrorString) -CUDNN_DNN_ROUTINE_EACH(DYNAMIC_LOAD_CUDNN_WRAP) - -#define CUDNN_DNN_ROUTINE_EACH_R2(__macro) \ - __macro(cudnnAddTensor) \ - __macro(cudnnConvolutionBackwardData) \ - __macro(cudnnConvolutionBackwardFilter) -CUDNN_DNN_ROUTINE_EACH_R2(DYNAMIC_LOAD_CUDNN_WRAP) - -// APIs available after R3: -#if CUDNN_VERSION >= 3000 -#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro) \ - __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize) \ - __macro(cudnnGetConvolutionBackwardDataAlgorithm) \ - __macro(cudnnGetConvolutionBackwardFilterAlgorithm) \ - __macro(cudnnGetConvolutionBackwardDataWorkspaceSize) -CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DYNAMIC_LOAD_CUDNN_WRAP) -#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R3 -#endif - - -// APIs available after R4: -#if CUDNN_VERSION >= 4007 -#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro) \ - __macro(cudnnBatchNormalizationForwardTraining) \ - __macro(cudnnBatchNormalizationForwardInference) \ - __macro(cudnnBatchNormalizationBackward) -CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DYNAMIC_LOAD_CUDNN_WRAP) -#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R4 -#endif - -// APIs in R5 -#if CUDNN_VERSION >= 5000 -#define CUDNN_DNN_ROUTINE_EACH_R5(__macro) \ - __macro(cudnnCreateActivationDescriptor) \ - __macro(cudnnSetActivationDescriptor) \ - __macro(cudnnGetActivationDescriptor) \ - __macro(cudnnDestroyActivationDescriptor) -CUDNN_DNN_ROUTINE_EACH_R5(DYNAMIC_LOAD_CUDNN_WRAP) -#undef CUDNN_DNN_ROUTINE_EACH_R5 -#endif - -#undef CUDNN_DNN_ROUTINE_EACH -// clang-format on -} /* namespace dynload */ - -/** - * Check build-in cudnn function using glog and it **does not** - * support << operator for more details error info. - */ -#define CHECK_CUDNN(cudnnFunc) \ - do { \ - cudnnStatus_t cudnnStat = cudnnFunc; \ - CHECK_EQ(CUDNN_STATUS_SUCCESS, cudnnStat) \ - << "Cudnn Error: " << dynload::cudnnGetErrorString(cudnnStat); \ - } while (0) - -bool g_is_libcudnn_init = false; -int g_cudnn_lib_version = 0; - -void hl_cudnn_desc_init(cudnnTensorDescriptor_t* cudnn_desc) { - CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(cudnn_desc)); -} - -void hl_cudnn_init(cudnnHandle_t* cudnn_handle, cudaStream_t stream) { - size_t cudnn_dso_ver = dynload::cudnnGetVersion(); - size_t cudnn_dso_major = cudnn_dso_ver / 1000; - size_t cudnn_cuh_major = CUDNN_VERSION / 1000; - - // Compare cudnn header version with that of cudnn.so. - CHECK((cudnn_cuh_major < 4 && cudnn_dso_major < 4) || - (cudnn_cuh_major == cudnn_dso_major)) - << "[cudnn init] libcudnn v" << cudnn_dso_major << " with header v" - << cudnn_cuh_major << " unmatched!\n" - << "PaddlePaddle Requirement: " - << "(header v[2-3] with libcudnn v[2-3]) Or " - << "(header v4 with libcudnn v4) Or " - << "(header v5 with libcudnn v5) Or" - << "(header v6 with libcudnn v6)."; - - CHECK(!(CUDNN_VERSION < 6000 && CUDNN_VERSION >= 5000 && CUDA_VERSION < 7050)) - << "cudnn v5 requires cuda version >= 7.5"; - - CHECK(!(CUDNN_VERSION >= 6000 && CUDA_VERSION < 8000)) - << "cudnn v6 requires cuda version >= 8.0"; - - CHECK_CUDNN(dynload::cudnnCreate(cudnn_handle)); - CHECK_CUDNN(dynload::cudnnSetStream(*cudnn_handle, stream)); - - g_is_libcudnn_init = true; - g_cudnn_lib_version = cudnn_dso_ver; -} - -int hl_get_cudnn_lib_version() { return g_cudnn_lib_version; } - -void hl_conv_workspace(hl_tensor_descriptor input, - hl_tensor_descriptor output, - hl_filter_descriptor filter, - hl_convolution_descriptor conv, - int* convFwdAlgo, - size_t* fwdLimitBytes, - int* convBwdDataAlgo, - size_t* bwdDataLimitBytes, - int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes, - bool useDilation) { -#if CUDNN_VERSION >= 4000 - - CHECK_NOTNULL(input); - CHECK_NOTNULL(output); - CHECK_NOTNULL(filter); - CHECK_NOTNULL(conv); - - // Specify workspace limit directly - size_t memoryLimitBytes = - (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb; - - // For dilation - int algo = 0; - - // cudnn convolution forward configuration - cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter); - cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); - // cudnn convolution backward data configuration - cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter); - cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnConvolutionDescriptor_t bwd_data_conv_desc = - GET_CONVOLUTION_DESCRIPTOR(conv); - // cudnn convolution backward filter configuration - cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnConvolutionDescriptor_t bwd_filter_conv_desc = - GET_CONVOLUTION_DESCRIPTOR(conv); - cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter); - - if (useDilation) { - convFwdAlgo = &algo; - convBwdDataAlgo = &algo; - convBwdFilterAlgo = &algo; - } else { - CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm( - t_resource.cudnn_handle, - fwd_src_desc, - fwd_filter_desc, - fwd_conv_desc, - fwd_dest_desc, - CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convFwdAlgo))); - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm( - t_resource.cudnn_handle, - bwd_data_filter_desc, - bwd_data_diff_desc, - bwd_data_conv_desc, - bwd_data_grad_desc, - CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convBwdDataAlgo))); - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm( - t_resource.cudnn_handle, - bwd_filter_src_desc, - bwd_filter_diff_desc, - bwd_filter_conv_desc, - bwd_filter_grad_desc, - CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convBwdFilterAlgo))); - } - - CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize( - t_resource.cudnn_handle, - fwd_src_desc, - fwd_filter_desc, - fwd_conv_desc, - fwd_dest_desc, - static_cast(*convFwdAlgo), - fwdLimitBytes)); - - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( - t_resource.cudnn_handle, - bwd_data_filter_desc, - bwd_data_diff_desc, - bwd_data_conv_desc, - bwd_data_grad_desc, - static_cast(*convBwdDataAlgo), - bwdDataLimitBytes)); - - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( - t_resource.cudnn_handle, - bwd_filter_src_desc, - bwd_filter_diff_desc, - bwd_filter_conv_desc, - bwd_filter_grad_desc, - static_cast(*convBwdFilterAlgo), - bwdFilterLimitBytes)); - -#endif -} - -void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc, - int batch_size, - int feature_maps, - int height, - int width) { - CHECK_NOTNULL(image_desc); - - cudnn_tensor_descriptor hl_desc = - (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor)); - CHECK_NOTNULL(hl_desc); - -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(&hl_desc->desc)); - - CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(hl_desc->desc, - CUDNN_TENSOR_NCHW, - data_type, - batch_size, - feature_maps, - height, - width)); - - hl_desc->format = CUDNN_TENSOR_NCHW; - hl_desc->data_type = data_type; - hl_desc->batch_size = batch_size; - hl_desc->feature_maps = feature_maps; - hl_desc->height = height; - hl_desc->width = width; - - *image_desc = (hl_tensor_descriptor)hl_desc; -} - -void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc) { - CHECK_NOTNULL(image_desc); - - cudnn_tensor_descriptor hl_desc = - (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor)); - CHECK_NOTNULL(hl_desc); - -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(&hl_desc->desc)); - - hl_desc->data_type = data_type; - - *image_desc = (hl_tensor_descriptor)hl_desc; -} - -void hl_tensor_reshape(hl_tensor_descriptor image_desc, - int batch_size, - int feature_maps, - int height, - int width) { - const int stride_w = 1; - const int stride_h = width * stride_w; - const int stride_c = height * stride_h; - const int stride_n = feature_maps * stride_c; - return hl_tensor_reshape(image_desc, - batch_size, - feature_maps, - height, - width, - stride_n, - stride_c, - stride_h, - stride_w); -} - -void hl_tensor_reshape(hl_tensor_descriptor image_desc, - int batch_size, - int feature_maps, - int height, - int width, - int nStride, - int cStride, - int hStride, - int wStride) { - CHECK_NOTNULL(image_desc); - - cudnn_tensor_descriptor hl_desc = (cudnn_tensor_descriptor)image_desc; - CHECK_NOTNULL(hl_desc->desc); - - CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptorEx(hl_desc->desc, - hl_desc->data_type, - batch_size, - feature_maps, - height, - width, - nStride, - cStride, - hStride, - wStride)); - - hl_desc->batch_size = batch_size; - hl_desc->feature_maps = feature_maps; - hl_desc->height = height; - hl_desc->width = width; -} - -void hl_destroy_tensor_descriptor(hl_tensor_descriptor image_desc) { - CHECK_NOTNULL(image_desc); - - cudnn_tensor_descriptor hl_desc = (cudnn_tensor_descriptor)image_desc; - CHECK_NOTNULL(hl_desc->desc); - - CHECK_CUDNN(dynload::cudnnDestroyTensorDescriptor(hl_desc->desc)); - - hl_desc->desc = NULL; - - free(image_desc); -} - -void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc, - hl_pooling_mode_t mode, - int height, - int width, - int height_padding, - int width_padding, - int stride_height, - int stride_width) { - cudnnPoolingMode_t cudnn_mode; - switch (mode) { - case HL_POOLING_MAX: - cudnn_mode = CUDNN_POOLING_MAX; - break; - case HL_POOLING_AVERAGE: - cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; - break; - case HL_POOLING_AVERAGE_INCLUDE_PADDING: - cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; - break; - default: - LOG(FATAL) << "parameter mode error"; - } - - CHECK_NOTNULL(pooling_desc); - - cudnn_pooling_descriptor hl_pooling_desc = - (cudnn_pooling_descriptor)malloc(sizeof(_cudnn_pooling_descriptor)); - CHECK_NOTNULL(hl_pooling_desc); - - CHECK_CUDNN(dynload::cudnnCreatePoolingDescriptor(&hl_pooling_desc->desc)); - - CHECK_CUDNN(dynload::cudnnSetPooling2dDescriptor(hl_pooling_desc->desc, - cudnn_mode, -#if CUDNN_VERSION >= 5000 - CUDNN_PROPAGATE_NAN, -#endif - height, - width, - height_padding, - width_padding, - stride_height, - stride_width)); - - hl_pooling_desc->mode = cudnn_mode; - hl_pooling_desc->window_height = height; - hl_pooling_desc->window_width = width; - hl_pooling_desc->stride_height = stride_height; - hl_pooling_desc->stride_width = stride_width; - - *pooling_desc = (hl_pooling_descriptor)hl_pooling_desc; -} - -void hl_destroy_pooling_descriptor(hl_pooling_descriptor pooling_desc) { - CHECK_NOTNULL(pooling_desc); - - cudnn_pooling_descriptor hl_pooling = (cudnn_pooling_descriptor)pooling_desc; - - CHECK_NOTNULL(hl_pooling->desc); - CHECK_CUDNN(dynload::cudnnDestroyPoolingDescriptor(hl_pooling->desc)); - - hl_pooling->desc = NULL; - - free(pooling_desc); -} - -void hl_pooling_forward(hl_tensor_descriptor input, - real* input_image, - hl_tensor_descriptor output, - real* output_image, - hl_pooling_descriptor pooling) { - cudnnPoolingDescriptor_t pooling_desc; - cudnnTensorDescriptor_t input_desc; - cudnnTensorDescriptor_t output_desc; - - CHECK_NOTNULL(input); - CHECK_NOTNULL(output); - CHECK_NOTNULL(pooling); - CHECK_NOTNULL(input_image); - CHECK_NOTNULL(output_image); - - real alpha = 1.0f; - real beta = 1.0f; - input_desc = ((cudnn_tensor_descriptor)input)->desc; - output_desc = ((cudnn_tensor_descriptor)output)->desc; - pooling_desc = ((cudnn_pooling_descriptor)pooling)->desc; - CHECK_CUDNN(dynload::cudnnPoolingForward(t_resource.cudnn_handle, - pooling_desc, - &alpha, - input_desc, - input_image, - &beta, - output_desc, - output_image)); - CHECK_SYNC("hl_pooling_forward failed"); -} - -void hl_pooling_backward(hl_tensor_descriptor input, - real* input_image, - real* input_image_grad, - hl_tensor_descriptor output, - real* output_image, - real* output_image_grad, - hl_pooling_descriptor pooling) { - cudnnPoolingDescriptor_t pooling_desc; - cudnnTensorDescriptor_t input_desc; - cudnnTensorDescriptor_t output_desc; - - CHECK_NOTNULL(input); - CHECK_NOTNULL(output); - CHECK_NOTNULL(pooling); - CHECK_NOTNULL(input_image); - CHECK_NOTNULL(input_image_grad); - CHECK_NOTNULL(output_image); - CHECK_NOTNULL(output_image_grad); - - real alpha = 1.0f; - real beta = 1.0f; - input_desc = ((cudnn_tensor_descriptor)input)->desc; - output_desc = ((cudnn_tensor_descriptor)output)->desc; - pooling_desc = ((cudnn_pooling_descriptor)pooling)->desc; - CHECK_CUDNN(dynload::cudnnPoolingBackward(t_resource.cudnn_handle, - pooling_desc, - &alpha, - output_desc, - output_image, - output_desc, - output_image_grad, - input_desc, - input_image, - &beta, - input_desc, - input_image_grad)); - CHECK_SYNC("hl_pooling_backward failed"); -} - -void hl_create_filter_descriptor(hl_filter_descriptor* filter, - int input_feature_maps, - int output_feature_maps, - int height, - int width) { - CHECK_NOTNULL(filter); - - cudnn_filter_descriptor hl_filter = - (cudnn_filter_descriptor)malloc(sizeof(_cudnn_filter_descriptor)); - CHECK_NOTNULL(hl_filter); - - CHECK_CUDNN(dynload::cudnnCreateFilterDescriptor(&hl_filter->desc)); - -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnSetFilter4dDescriptor(hl_filter->desc, - data_type, -#if CUDNN_VERSION >= 5000 - CUDNN_TENSOR_NCHW, -#endif - output_feature_maps, - input_feature_maps, - height, - width)); - - hl_filter->data_type = data_type; - hl_filter->output_feature_maps = output_feature_maps; - hl_filter->input_feature_maps = input_feature_maps; - hl_filter->filter_height = height; - hl_filter->filter_width = width; - - *filter = (hl_filter_descriptor)hl_filter; -} - -void hl_destroy_filter_descriptor(hl_filter_descriptor filter) { - CHECK_NOTNULL(filter); - - cudnn_filter_descriptor hl_filter = (cudnn_filter_descriptor)filter; - CHECK_NOTNULL(hl_filter->desc); - - CHECK_CUDNN(dynload::cudnnDestroyFilterDescriptor(hl_filter->desc)); - - hl_filter->desc = NULL; - - free(filter); -} - -void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, - hl_tensor_descriptor image, - hl_filter_descriptor filter, - int padding_height, - int padding_width, - int stride_height, - int stride_width, - int dilation_h, - int dilation_w) { - CHECK_NOTNULL(conv); - - cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc( - sizeof(_cudnn_convolution_descriptor)); - - CHECK_NOTNULL(hl_conv); - CHECK_CUDNN(dynload::cudnnCreateConvolutionDescriptor(&hl_conv->desc)); - - cudnnConvolutionMode_t mode = CUDNN_CROSS_CORRELATION; - -#if CUDNN_VERSION >= 6000 -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc, - padding_height, - padding_width, - stride_height, - stride_width, - dilation_h, - dilation_w, - mode, - data_type)); -#else - if (dilation_h > 1 || dilation_w > 1) { - LOG(FATAL) - << "Current cuDNN version does't support for dilation convolution. " - << "The dilation convolution requires cuDNN >= v6.0."; - } - - CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc, - padding_height, - padding_width, - stride_height, - stride_width, - dilation_h, - dilation_w, - mode)); -#endif - - hl_conv->input_image = image; - hl_conv->filter = filter; - hl_conv->padding_height = padding_height; - hl_conv->padding_width = padding_width; - hl_conv->stride_height = stride_height; - hl_conv->stride_width = stride_width; - hl_conv->upscalex = 1; - hl_conv->upscaley = 1; - hl_conv->mode = mode; - - *conv = (hl_convolution_descriptor)hl_conv; -} - -void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, - hl_tensor_descriptor image, - hl_filter_descriptor filter, - int padding_height, - int padding_width, - int stride_height, - int stride_width, - int dilation_h, - int dilation_w) { - CHECK_NOTNULL(conv); - CHECK_NOTNULL(image); - CHECK_NOTNULL(filter); - - cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); - cudnnConvolutionMode_t mode = CUDNN_CROSS_CORRELATION; - -#if CUDNN_VERSION >= 6000 -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(conv_desc, - padding_height, - padding_width, - stride_height, - stride_width, - dilation_h, - dilation_w, - mode, - data_type)); -#else - CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(conv_desc, - padding_height, - padding_width, - stride_height, - stride_width, - dilation_h, - dilation_w, - mode)); -#endif - - cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)conv; - hl_conv->input_image = image; - hl_conv->filter = filter; - hl_conv->padding_height = padding_height; - hl_conv->padding_width = padding_width; - hl_conv->stride_height = stride_height; - hl_conv->stride_width = stride_width; - hl_conv->upscalex = 1; - hl_conv->upscaley = 1; - hl_conv->mode = mode; -} - -void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) { - CHECK_NOTNULL(conv); - - cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)conv; - CHECK_NOTNULL(hl_conv->desc); - - CHECK_CUDNN(dynload::cudnnDestroyConvolutionDescriptor(hl_conv->desc)); - hl_conv->desc = NULL; - - free(conv); -} - -void hl_convolution_forward(hl_tensor_descriptor input, - real* input_data, - hl_tensor_descriptor output, - real* output_data, - hl_filter_descriptor filter, - real* filter_data, - hl_convolution_descriptor conv, - void* gpuWorkSpace, - size_t sizeInBytes, - int convFwdAlgo) { - CHECK_NOTNULL(input); - CHECK_NOTNULL(output); - CHECK_NOTNULL(filter); - CHECK_NOTNULL(conv); - CHECK_NOTNULL(input_data); - CHECK_NOTNULL(output_data); - CHECK_NOTNULL(filter_data); - cudnnTensorDescriptor_t src_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnTensorDescriptor_t dest_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnFilterDescriptor_t filter_desc = GET_FILTER_DESCRIPTOR(filter); - cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); - real alpha = 1.0f; - real beta = 1.0f; - CHECK_CUDNN(dynload::cudnnConvolutionForward( - t_resource.cudnn_handle, - &alpha, - src_desc, - input_data, - filter_desc, - filter_data, - conv_desc, - static_cast(convFwdAlgo), - gpuWorkSpace, - sizeInBytes, - &beta, - dest_desc, - output_data)); - CHECK_SYNC("hl_convolution_forward failed"); -} - -void hl_convolution_forward_add_bias(hl_tensor_descriptor bias, - real* bias_data, - hl_tensor_descriptor output, - real* output_data) { - CHECK_NOTNULL(bias); - CHECK_NOTNULL(output); - CHECK_NOTNULL(bias_data); - CHECK_NOTNULL(output_data); - - cudnnTensorDescriptor_t output_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnTensorDescriptor_t bias_desc = GET_TENSOR_DESCRIPTOR(bias); - real alpha = 1.0f; - real beta = 1.0f; - - CHECK_CUDNN(dynload::cudnnAddTensor(t_resource.cudnn_handle, -#if CUDNN_VERSION < 4000 - CUDNN_ADD_SAME_C, -#endif - &alpha, - bias_desc, - bias_data, - &beta, - output_desc, - output_data)); - CHECK_SYNC("hl_convolution_forward_add_bias failed"); -} - -void hl_convolution_backward_bias(hl_tensor_descriptor bias, - real* bias_grad_data, - hl_tensor_descriptor output, - real* output_grad_data) { - CHECK_NOTNULL(bias); - CHECK_NOTNULL(output); - CHECK_NOTNULL(bias_grad_data); - CHECK_NOTNULL(output_grad_data); - - real alpha = 1.0f; - real beta = 1.0f; - cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnTensorDescriptor_t bias_desc = GET_TENSOR_DESCRIPTOR(bias); - CHECK_CUDNN(dynload::cudnnConvolutionBackwardBias(t_resource.cudnn_handle, - &alpha, - diff_desc, - output_grad_data, - &beta, - bias_desc, - bias_grad_data)); - CHECK_SYNC("hl_convolution_backward_bias failed"); -} - -void hl_convolution_backward_filter(hl_tensor_descriptor input, - real* input_data, - hl_tensor_descriptor output, - real* output_grad_data, - hl_filter_descriptor filter, - real* filter_grad_data, - hl_convolution_descriptor conv, - void* gpuWorkSpace, - size_t sizeInBytes, - int convBwdFilterAlgo) { - CHECK_NOTNULL(input); - CHECK_NOTNULL(output); - CHECK_NOTNULL(filter); - CHECK_NOTNULL(conv); - CHECK_NOTNULL(input_data); - CHECK_NOTNULL(output_grad_data); - CHECK_NOTNULL(filter_grad_data); - - real alpha = 1.0f; - real beta = 1.0f; - cudnnTensorDescriptor_t src_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); - cudnnFilterDescriptor_t grad_desc = GET_FILTER_DESCRIPTOR(filter); - - CHECK_CUDNN(dynload::cudnnConvolutionBackwardFilter( - t_resource.cudnn_handle, - &alpha, - src_desc, - input_data, - diff_desc, - output_grad_data, - conv_desc, -#if CUDNN_VERSION >= 4000 - static_cast(convBwdFilterAlgo), - gpuWorkSpace, - sizeInBytes, -#endif - &beta, - grad_desc, - filter_grad_data)); - CHECK_SYNC("hl_convolution_backward_filter failed"); -} - -void hl_convolution_backward_data(hl_tensor_descriptor input, - real* input_data_grad, - hl_tensor_descriptor output, - real* output_grad_data, - hl_filter_descriptor filter, - real* filter_data, - hl_convolution_descriptor conv, - void* gpuWorkSpace, - size_t sizeInBytes, - int convBwdDataAlgo) { - real alpha = 1.0f; - real beta = 1.0f; - cudnnFilterDescriptor_t filter_desc = GET_FILTER_DESCRIPTOR(filter); - cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnTensorDescriptor_t grad_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); - - CHECK_CUDNN(dynload::cudnnConvolutionBackwardData( - t_resource.cudnn_handle, - &alpha, - filter_desc, - filter_data, - diff_desc, - output_grad_data, - conv_desc, -#if CUDNN_VERSION >= 4000 - static_cast(convBwdDataAlgo), - gpuWorkSpace, - sizeInBytes, -#endif - &beta, - grad_desc, - input_data_grad)); - CHECK_SYNC("hl_convolution_backward_data failed"); -} - -void hl_softmax_forward(real* input, real* output, int height, int width) { -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(t_resource.cudnn_desc, - CUDNN_TENSOR_NCHW, - data_type, - height, - width, - 1, - 1)); - - real alpha = 1.0f; - real beta = 0.0f; - CHECK_CUDNN(dynload::cudnnSoftmaxForward(t_resource.cudnn_handle, - CUDNN_SOFTMAX_ACCURATE, - CUDNN_SOFTMAX_MODE_CHANNEL, - &alpha, - t_resource.cudnn_desc, - input, - &beta, - t_resource.cudnn_desc, - output)); - CHECK_SYNC("hl_softmax_forward failed"); -} - -void hl_softmax_backward(real* output_value, - real* output_grad, - int height, - int width) { -#ifndef PADDLE_TYPE_DOUBLE - cudnnDataType_t data_type = CUDNN_DATA_FLOAT; -#else - cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; -#endif - CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(t_resource.cudnn_desc, - CUDNN_TENSOR_NCHW, - data_type, - height, - width, - 1, - 1)); - - real alpha = 1.0f; - real beta = 0.0f; - CHECK_CUDNN(dynload::cudnnSoftmaxBackward(t_resource.cudnn_handle, - CUDNN_SOFTMAX_ACCURATE, - CUDNN_SOFTMAX_MODE_CHANNEL, - &alpha, - t_resource.cudnn_desc, - output_value, - t_resource.cudnn_desc, - output_grad, - &beta, - t_resource.cudnn_desc, - output_grad)); - CHECK_SYNC("hl_softmax_backward failed"); -} - -void hl_batch_norm_forward_training(hl_tensor_descriptor inputDesc, - real* input, - hl_tensor_descriptor outputDesc, - real* output, - hl_tensor_descriptor bnParamDesc, - real* scale, - real* bias, - double factor, - real* runningMean, - real* runningInvVar, - double epsilon, - real* savedMean, - real* savedVar) { -#if CUDNN_VERSION >= 4007 - if ((NULL != runningMean && NULL == runningInvVar) || - (NULL == runningMean && NULL != runningInvVar)) { - LOG(FATAL) << "runningMean and runningInvVar can be NULL " - << "but only at the same time."; - } - if ((NULL != savedMean && NULL == savedVar) || - (NULL == savedMean && NULL != savedVar)) { - LOG(FATAL) << "savedMean and savedVar can be NULL " - << "but only at the same time."; - } - - cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc); - cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc); - cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc); - real alpha = 1.0f; - real beta = 1.0f; - cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; - CHECK_CUDNN( - dynload::cudnnBatchNormalizationForwardTraining(t_resource.cudnn_handle, - mode, - &alpha, - &beta, - xDesc, - input, - yDesc, - output, - bnDesc, - scale, - bias, - factor, - runningMean, - runningInvVar, - epsilon, - savedMean, - savedVar)); - - CHECK_SYNC("hl_batch_norm_forward_training failed"); -#else - LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. " - << "But cudnn lib version is " << g_cudnn_lib_version; -#endif -} - -void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc, - real* input, - hl_tensor_descriptor outputDesc, - real* output, - hl_tensor_descriptor bnParamDesc, - real* scale, - real* bias, - real* estimatedMean, - real* estimatedInvVar, - double epsilon) { -#if CUDNN_VERSION >= 4007 - cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc); - cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc); - cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc); - real alpha = 1.0f; - real beta = 1.0f; - cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; - - CHECK_CUDNN( - dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle, - mode, - &alpha, - &beta, - xDesc, - input, - yDesc, - output, - bnDesc, - scale, - bias, - estimatedMean, - estimatedInvVar, - epsilon)); - - CHECK_SYNC("hl_batch_norm_forward_inference failed"); -#else - LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. " - << "But cudnn lib version is " << g_cudnn_lib_version; -#endif -} - -void hl_batch_norm_backward(hl_tensor_descriptor inputDesc, - real* input, - hl_tensor_descriptor outGradDesc, - real* outGrad, - hl_tensor_descriptor inGradDesc, - real* inGrad, - hl_tensor_descriptor dBnParamDesc, - real* scale, - real* scaleGrad, - real* biasGrad, - double epsilon, - real* savedMean, - real* savedInvVar) { -#if CUDNN_VERSION >= 4007 - if ((NULL != savedMean && NULL == savedInvVar) || - (NULL == savedMean && NULL != savedInvVar)) { - LOG(FATAL) << "savedMean and savedVar can be NULL " - << "but only at the same time."; - } - - cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc); - cudnnTensorDescriptor_t dyDesc = GET_TENSOR_DESCRIPTOR(outGradDesc); - cudnnTensorDescriptor_t dxDesc = GET_TENSOR_DESCRIPTOR(inGradDesc); - cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(dBnParamDesc); - real alpha = 1.0f; - real beta = 1.0f; - cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; - CHECK_CUDNN(dynload::cudnnBatchNormalizationBackward(t_resource.cudnn_handle, - mode, - &alpha, - &beta, - &alpha, - &beta, - xDesc, - input, - dyDesc, - outGrad, - dxDesc, - inGrad, - bnDesc, - scale, - scaleGrad, - biasGrad, - epsilon, - savedMean, - savedInvVar)); - - CHECK_SYNC("hl_batch_norm_backward failed"); -#else - LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. " - << "But cudnn lib version is " << g_cudnn_lib_version; -#endif -} diff --git a/paddle/cuda/src/hl_cuda_device.cc b/paddle/cuda/src/hl_cuda_device.cc deleted file mode 100644 index 3025aa48523d67fe3d7ed03f44252d1211d2a46a..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_device.cc +++ /dev/null @@ -1,677 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -// clang-format off -// Because clang-format 4.X and clang-format 3.8+ format -// following lines in different. So disable clang-format. -#include "hl_cuda.h" -#include -#include -#include -#include -#include -#include "hl_cuda.ph" -#include "hl_thread.ph" -#include "paddle/utils/Logging.h" -#include "paddle/utils/DynamicLoader.h" -// clang-format on - -namespace dynload { - -std::once_flag curand_dso_flag; -void *curand_dso_handle = nullptr; - -/** - * The following macro definition can generate structs - * (for each function) to dynamic load curand routine - * via operator overloading. - * - * note: default dynamic linked libs - */ -#ifdef PADDLE_USE_DSO -#define DYNAMIC_LOAD_CURAND_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - curandStatus_t operator()(Args... args) { \ - typedef curandStatus_t (*curandFunc)(Args...); \ - std::call_once(curand_dso_flag, GetCurandDsoHandle, &curand_dso_handle); \ - void *p_##__name = dlsym(curand_dso_handle, #__name); \ - return reinterpret_cast(p_##__name)(args...); \ - } \ - } __name; /* struct DynLoad__##__name */ -#else -#define DYNAMIC_LOAD_CURAND_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - curandStatus_t operator()(Args... args) { \ - return __name(args...); \ - } \ - } __name; /* struct DynLoad__##__name */ -#endif - -/* include all needed curand functions in HPPL */ -// clang-format off -#define CURAND_RAND_ROUTINE_EACH(__macro) \ - __macro(curandCreateGenerator) \ - __macro(curandSetStream) \ - __macro(curandSetPseudoRandomGeneratorSeed)\ - __macro(curandGenerateUniform) \ - __macro(curandGenerateUniformDouble) -// clang-format on - -CURAND_RAND_ROUTINE_EACH(DYNAMIC_LOAD_CURAND_WRAP) - -#undef CURAND_RAND_ROUTINE_EACH -#undef DYNAMIC_LOAD_CURAND_WRAP - -} /* namespace dynload */ - -/** - * @brief global resource. - */ -int g_system_device_num = 0; /* system device number */ -int device_num = 0; /* use device number */ -hl_device_prop *g_device; /* device info table */ -__thread thread_device_resources *t_device; /* device resources table */ -int g_cuda_lib_version = 0; - -/* number of global stream */ -#define NUMBER_OF_GLOBAL_STREAM (HPPL_THREAD_STREAM_1) -/* number of thread stream */ -#define NUMBER_OF_THREAD_STREAM (HPPL_STREAM_END - HPPL_THREAD_STREAM_1) -/* sizeof of device memory */ -#define HPPL_GPU_MEMORY_SIZE (256 * 4) - -/** - * Check build-in cuda function using glog and it **does not** - * support << operator for more details error info. - */ -#define CHECK_CUDA(cudaFunc) \ - do { \ - cudaError_t cudaStat = cudaFunc; \ - CHECK_EQ(cudaSuccess, cudaStat) << "Cuda Error: " \ - << cudaGetErrorString(cudaStat); \ - } while (0) - -/** - * @brief thread resource. - */ -__thread _hl_thread_resource t_resource = {{0}, /* stream */ - 0, /* handle */ - 0, /* gen */ - 0, /* cudnn_handle */ - 0, /* cudnn_desc */ - NULL, /* gen_mutex */ - NULL, /* gpu_mem */ - NULL, /* cpu_mem */ - 0, /* event */ - -1, /* device */ - 0, /* major */ - false}; /* is_init */ - -__thread cudaStream_t default_stream = 0; -__thread bool g_sync_flag = true; -bool hl_start_flag = false; - -inline pid_t gettid() { -#if defined(__APPLE__) || defined(__OSX__) - // syscall is deprecated: first deprecated in macOS 10.12. - // syscall is unsupported; - // syscall pid_t tid = syscall(SYS_thread_selfid); - uint64_t tid; - pthread_threadid_np(NULL, &tid); -#else -#ifndef __NR_gettid -#define __NR_gettid 224 -#endif - pid_t tid = syscall(__NR_gettid); -#endif - CHECK_NE((int)tid, -1); - return tid; -} - -void hl_init(int device) { - CHECK(hl_start_flag) << "[Init failed] hl_start() did not succeed."; - - /* thread has been initialized */ - if (true == t_resource.is_init) { - hl_set_device(device); - return; - } - - /* create thread devcie resources */ - char *tmp; - thread_device_resources device_res; - tmp = (char *)malloc(g_system_device_num * sizeof(thread_device_resources *) + - device_num * sizeof(_thread_device_resources)); - CHECK_NOTNULL(tmp); - t_device = (thread_device_resources *)tmp; - device_res = (thread_device_resources)( - (char *)tmp + g_system_device_num * sizeof(thread_device_resources *)); - memset(t_device, 0, g_system_device_num * sizeof(thread_device_resources *)); - - char *tmp_stream = (char *)malloc(device_num * NUMBER_OF_THREAD_STREAM * - sizeof(cudaStream_t)); - CHECK_NOTNULL(tmp_stream); - - int num = 0; - for (int dev = 0; dev < g_system_device_num; dev++) { - if (!g_device[dev]) { - continue; - } - - t_device[dev] = &device_res[num]; - t_device[dev]->stream = - (cudaStream_t *)(tmp_stream + - num * NUMBER_OF_THREAD_STREAM * sizeof(cudaStream_t)); - - hl_create_thread_resources(dev, t_device[dev]); - num++; - } - - hl_cudnn_desc_init(&t_resource.cudnn_desc); - - /* thread initialization is complete */ - t_resource.is_init = true; - /* set device */ - t_resource.device = -1; - hl_set_device(device); -} - -void hl_fini() { - if (false == t_resource.is_init) { - return; - } - - /* hppl stream fini */ - t_resource.device = -1; - for (int i = NUMBER_OF_GLOBAL_STREAM; i < HPPL_STREAM_END; i++) { - t_resource.stream[i] = 0; - } - - char *tmp = (char *)t_device; - char *tmp_stream = NULL; - for (int dev = 0; dev < g_system_device_num; dev++) { - if (!t_device[dev]) { - continue; - } - if (!tmp_stream) { - tmp_stream = (char *)t_device[dev]->stream; - } - for (int j = 0; j < NUMBER_OF_THREAD_STREAM; j++) { - CHECK_CUDA(cudaStreamDestroy(t_device[dev]->stream[j])); - } - - /* free device memory */ - hl_free_mem_device(t_device[dev]->gpu_mem); - hl_free_mem_host(t_device[dev]->cpu_mem); - CHECK_CUDA(cudaEventDestroy(t_device[dev]->mem_event)); - } - - free(tmp); - free(tmp_stream); - t_resource.is_init = false; -} - -int hl_get_device_count() { return device_num; } - -void hl_set_device(int device) { - if (device == t_resource.device) { - return; - } - - CHECK(device >= 0 && device < g_system_device_num && g_device[device]) - << "Device: " << device << " is not specified in startup."; - - CHECK_CUDA(cudaSetDevice(device)); - - /* switch thread stream */ - for (int i = 0; i < NUMBER_OF_GLOBAL_STREAM; i++) { - t_resource.stream[i] = g_device[device]->device_resources->stream[i]; - } - - if (true == t_resource.is_init) { - for (int i = NUMBER_OF_GLOBAL_STREAM; i < HPPL_STREAM_END; i++) { - t_resource.stream[i] = - t_device[device]->stream[i - NUMBER_OF_GLOBAL_STREAM]; - } - t_resource.gpu_mem = t_device[device]->gpu_mem; - t_resource.cpu_mem = t_device[device]->cpu_mem; - t_resource.event = t_device[device]->mem_event; - } - - t_resource.handle = g_device[device]->device_resources->handle; - t_resource.gen = g_device[device]->device_resources->gen; - t_resource.cudnn_handle = g_device[device]->device_resources->cudnn_handle; - t_resource.gen_mutex = g_device[device]->device_resources->gen_mutex; - t_resource.device = device; - t_resource.major = g_device[device]->major; - default_stream = t_resource.stream[0]; -} - -int hl_get_device() { - int device; - CHECK_CUDA(cudaGetDevice(&device)); - return device; -} - -void *hl_malloc_device(size_t size) { - void *dest_d; - - CHECK(size) << __func__ << ": the size for device memory is 0, please check."; - CHECK_CUDA(cudaMalloc((void **)&dest_d, size)); - - return dest_d; -} - -void hl_free_mem_device(void *dest_d) { - CHECK_NOTNULL(dest_d); - - cudaError_t err = cudaFree(dest_d); - CHECK(cudaSuccess == err || cudaErrorCudartUnloading == err) - << hl_get_device_error_string(); -} - -void *hl_malloc_host(size_t size) { - void *dest_h; - - CHECK(size) << __func__ << ": the size for device memory is 0, please check."; - CHECK_CUDA(cudaHostAlloc((void **)&dest_h, size, cudaHostAllocDefault)); - - return dest_h; -} - -void hl_free_mem_host(void *dest_h) { - CHECK_NOTNULL(dest_h); - - cudaError_t err = cudaFreeHost(dest_h); - CHECK(cudaSuccess == err || cudaErrorCudartUnloading == err) - << hl_get_device_error_string(); -} - -void hl_memcpy(void *dst, void *src, size_t size) { - if (0 == size) { - return; - } - CHECK_NOTNULL(dst); - CHECK_NOTNULL(src); - CHECK_CUDA(cudaMemcpy(dst, src, size, cudaMemcpyDefault)); -} - -void hl_memset_device(void *dest_d, int value, size_t size) { - CHECK_CUDA(cudaMemset(dest_d, value, size)); -} - -void hl_memcpy_host2device(void *dest_d, void *src_h, size_t size) { - if (0 == size) { - return; - } - CHECK_NOTNULL(src_h); - CHECK_NOTNULL(dest_d); - CHECK_CUDA(cudaMemcpy(dest_d, src_h, size, cudaMemcpyHostToDevice)); -} - -void hl_memcpy_device2host(void *dest_h, void *src_d, size_t size) { - if (0 == size) { - return; - } - CHECK_NOTNULL(dest_h); - CHECK_NOTNULL(src_d); - CHECK_CUDA(cudaMemcpy(dest_h, src_d, size, cudaMemcpyDeviceToHost)); -} - -void hl_memcpy_device2device(void *dest_d, void *src_d, size_t size) { - if (0 == size) { - return; - } - CHECK_NOTNULL(dest_d); - CHECK_NOTNULL(src_d); - CHECK_CUDA(cudaMemcpy(dest_d, src_d, size, cudaMemcpyDeviceToDevice)); -} - -void hl_memcpy_async(void *dst, void *src, size_t size, hl_stream_t stream) { - cudaStream_t cu_stream; - - if (0 == size) { - return; - } - CHECK_NOTNULL(dst); - CHECK_NOTNULL(src); - CHECK_LT(stream, HPPL_STREAM_END); - cu_stream = t_resource.stream[stream]; - - CHECK_CUDA(cudaMemcpyAsync(dst, src, size, cudaMemcpyDefault, cu_stream)); -} - -void hl_start() { - hl_specify_devices_start(NULL, 0); - /* set default device */ - hl_set_device(0); -} - -bool hl_device_can_access_peer(int device, int peerDevice) { - int canAccessPeer; - CHECK_CUDA(cudaDeviceCanAccessPeer(&canAccessPeer, device, peerDevice)); - - if (canAccessPeer == 1) { - return true; - } else { - return false; - } -} - -void hl_device_enable_peer_access(int peerDevice) { - cudaError_t err = cudaDeviceEnablePeerAccess(peerDevice, 0); - if (cudaErrorPeerAccessAlreadyEnabled == err) { - cudaGetLastError(); - } else { - CHECK_CUDA(err); - } -} - -void hl_create_global_resources(hl_device_prop device_prop) { - struct cudaDeviceProp cu_prop; - int device = device_prop->device; - global_device_resources device_res = device_prop->device_resources; - - CHECK_CUDA(cudaSetDevice(device)); - /* device properties */ - CHECK_CUDA(cudaGetDeviceProperties(&cu_prop, device)); - - device_prop->major = cu_prop.major; - device_prop->minor = cu_prop.minor; - strncpy(device_prop->device_name, cu_prop.name, 256); - device_prop->device_mem = cu_prop.totalGlobalMem; - - /* create device stream */ - for (int j = 0; j < NUMBER_OF_GLOBAL_STREAM; j++) { - CHECK_CUDA(cudaStreamCreate(&device_res->stream[j])); - } - - /* cublas init */ - hl_cublas_init(&device_res->handle, device_res->stream[0]); - - /* create curand gen */ - CHECK_EQ(dynload::curandCreateGenerator(&device_res->gen, - CURAND_RNG_PSEUDO_DEFAULT), - CURAND_STATUS_SUCCESS) - << "[Start failed] Curand init failed."; - - CHECK_EQ(dynload::curandSetStream(device_res->gen, device_res->stream[0]), - CURAND_STATUS_SUCCESS) - << "[Start failed] Curand set stream failed!"; - - /* create cudnn handle */ - hl_cudnn_init(&device_res->cudnn_handle, device_res->stream[0]); - - int seed = gettid(); - CHECK_EQ(dynload::curandSetPseudoRandomGeneratorSeed(device_res->gen, - seed + device), - CURAND_STATUS_SUCCESS); - - device_res->gen_mutex = (pthread_mutex_t *)(malloc(sizeof(pthread_mutex_t))); - pthread_mutex_init(device_res->gen_mutex, NULL); - - CHECK_CUDA(cudaRuntimeGetVersion(&g_cuda_lib_version)); -} - -int hl_get_cuda_version() { return g_cuda_lib_version; } - -void hl_create_thread_resources(int device, - thread_device_resources device_res) { - CHECK_CUDA(cudaSetDevice(device)); - - /* create thread stream */ - for (int j = 0; j < NUMBER_OF_THREAD_STREAM; j++) { - CHECK_CUDA(cudaStreamCreate(&device_res->stream[j])); - } - - /* allocation device memory */ - device_res->gpu_mem = (real *)hl_malloc_device(HPPL_GPU_MEMORY_SIZE); - - /* allocation host memory */ - device_res->cpu_mem = (real *)hl_malloc_host(HPPL_GPU_MEMORY_SIZE); - - CHECK_CUDA(cudaEventCreate(&device_res->mem_event)); -} - -void hl_specify_devices_start(int *device, int number) { - if (hl_start_flag) return; - - /* 1. get the number of devices */ - CHECK_CUDA(cudaGetDeviceCount(&g_system_device_num)); - CHECK_NE(g_system_device_num, 0) << "[Start failed] there is no GPU device"; - if (device == NULL) { - number = g_system_device_num; - } - - /* 2. check device & create device property table */ - CHECK_LE(number, g_system_device_num) - << "[Start failed] System does not have enough device. " - << "Device number: " << g_system_device_num << "Input number: " << number; - - char *tmp; - hl_device_prop device_prop; - tmp = (char *)malloc(g_system_device_num * sizeof(hl_device_prop *) + - number * sizeof(_hl_device_prop)); - CHECK(tmp) << "[Start failed] System memory is not enough."; - - g_device = (hl_device_prop *)tmp; - device_prop = (hl_device_prop)( - (char *)tmp + g_system_device_num * sizeof(hl_device_prop *)); - memset(g_device, 0, g_system_device_num * sizeof(hl_device_prop *)); - int num = 0; - for (int i = 0; i < number; i++) { - int dev; - if (device == NULL) { - dev = i; - } else { - dev = device[i]; - } - - CHECK_LT(dev, g_system_device_num) - << "[Start failed] The specified device number is " - << "out of range. Max device number: " << g_system_device_num - 1 - << " Specified devcie number: " << dev; - - if (g_device[dev]) { - /* Warning */ - LOG(WARNING) << "[Warning] Repeat specify device: " << dev; - continue; - } - - g_device[dev] = &device_prop[num]; - g_device[dev]->device = dev; - num++; - } - device_num = num; - - /* 3. create global device resources */ - char *tmp_res = (char *)malloc(device_num * sizeof(_global_device_resources)); - CHECK_NOTNULL(tmp_res); - - char *tmp_stream = (char *)malloc(device_num * NUMBER_OF_GLOBAL_STREAM * - sizeof(cudaStream_t)); - CHECK_NOTNULL(tmp_stream); - - num = 0; - for (int i = 0; i < g_system_device_num; i++) { - if (!g_device[i]) { - continue; - } - - g_device[i]->device_resources = (global_device_resources)( - tmp_res + num * sizeof(_global_device_resources)); - g_device[i]->device_resources->stream = - (cudaStream_t *)(tmp_stream + - num * NUMBER_OF_GLOBAL_STREAM * sizeof(cudaStream_t)); - - hl_create_global_resources(g_device[i]); - num++; - } - - /* hl_start() is ok */ - hl_start_flag = true; - /* set default device */ - if (device == NULL) { - hl_set_device(0); - } else { - hl_set_device(device[0]); - } -} - -void hl_rand(real *dest_d, size_t num) { - pthread_mutex_lock(t_resource.gen_mutex); - CHECK_EQ( -#ifndef PADDLE_TYPE_DOUBLE - dynload::curandGenerateUniform(t_resource.gen, dest_d, num), -#else - dynload::curandGenerateUniformDouble(t_resource.gen, dest_d, num), -#endif - CURAND_STATUS_SUCCESS); - pthread_mutex_unlock(t_resource.gen_mutex); - CHECK_SYNC("hl_rand failed"); -} - -void hl_srand(unsigned int seed) { - pthread_mutex_lock(t_resource.gen_mutex); - CHECK_EQ(dynload::curandSetPseudoRandomGeneratorSeed(t_resource.gen, seed), - CURAND_STATUS_SUCCESS); - pthread_mutex_unlock(t_resource.gen_mutex); -} - -void hl_set_sync_flag(bool flag) { g_sync_flag = flag; } - -bool hl_get_sync_flag() { return g_sync_flag; } - -void hl_stream_synchronize(hl_stream_t stream) { - cudaStream_t cu_stream; - - CHECK_LT(stream, HPPL_STREAM_END) << __func__ - << ": the parameter stream is error."; - - cu_stream = t_resource.stream[stream]; - CHECK_CUDA(cudaStreamSynchronize(cu_stream)); -} - -void hl_create_event(hl_event_t *event) { - CHECK_NOTNULL(event); - - struct _hl_event_st *st_event = - (struct _hl_event_st *)malloc(sizeof(struct _hl_event_st)); - - CHECK_CUDA(cudaEventCreate(&st_event->cu_event)); - - *event = st_event; -} - -float hl_event_elapsed_time(hl_event_t start, hl_event_t end) { - float time; - CHECK_NOTNULL(start); - CHECK_NOTNULL(end); - - CHECK_CUDA(cudaEventElapsedTime(&time, start->cu_event, end->cu_event)); - return time; -} - -void hl_stream_record_event(hl_stream_t stream, hl_event_t event) { - cudaStream_t cu_stream; - - CHECK_NOTNULL(event); - CHECK_LT(stream, HPPL_STREAM_END) << __func__ - << ": the parameter stream is error."; - - cu_stream = t_resource.stream[stream]; - CHECK_CUDA(cudaEventRecord(event->cu_event, cu_stream)); -} - -void hl_stream_wait_event(hl_stream_t stream, hl_event_t event) { - cudaStream_t cu_stream; - - CHECK_NOTNULL(event); - CHECK_LT(stream, HPPL_STREAM_END) << __func__ - << ": the parameter stream is error."; - - cu_stream = t_resource.stream[stream]; - CHECK_CUDA(cudaStreamWaitEvent(cu_stream, event->cu_event, 0)); -} - -void hl_destroy_event(hl_event_t event) { - CHECK_NOTNULL(event); - CHECK_CUDA(cudaEventDestroy(event->cu_event)); - - free(event); - event = NULL; -} - -void hl_event_synchronize(hl_event_t event) { - CHECK_NOTNULL(event); - CHECK_CUDA(cudaEventSynchronize(event->cu_event)); -} - -void hl_get_device_name(char *name, int len, int device) { - CHECK_NOTNULL(name); - CHECK(device >= 0 && device < g_system_device_num && g_device[device]) - << "Device(" << device << ") is not specified in startup."; - - strncpy(name, g_device[device]->device_name, len); -} - -void hl_get_device_memory(size_t *mem_size, int device) { - CHECK_NOTNULL(mem_size); - CHECK(device >= 0 && device < g_system_device_num && g_device[device]) - << "Device(" << device << ") is not specified in startup."; - - *mem_size = g_device[device]->device_mem; -} - -void hl_get_device_compute_capability(int *major, int *minor, int device) { - CHECK_NOTNULL(major); - CHECK_NOTNULL(minor); - CHECK(device >= 0 && device < g_system_device_num && g_device[device]) - << "Device(" << device << ") is not specified in startup."; - - *major = g_device[device]->major; - *minor = g_device[device]->minor; -} - -int hl_get_device_last_error() { return (int)cudaGetLastError(); } - -const char *hl_get_device_error_string() { - cudaError_t err = cudaGetLastError(); - return cudaGetErrorString(err); -} - -const char *hl_get_device_error_string(size_t err) { - return cudaGetErrorString((cudaError_t)err); -} - -void hl_device_synchronize() { CHECK_CUDA(cudaDeviceSynchronize()); } -void hl_set_device_flags_block() { - CHECK_CUDA(cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync)); -} - -bool hl_cuda_event_is_ready(hl_event_t event) { - cudaError_t err = cudaEventQuery(event->cu_event); - CHECK(cudaSuccess == err || cudaErrorNotReady == err); - - if (cudaErrorNotReady == err) { - return false; - } - return true; -} - -void hl_profiler_start() { CHECK_CUDA(cudaProfilerStart()); } - -void hl_profiler_end() { CHECK_CUDA(cudaProfilerStop()); } diff --git a/paddle/cuda/src/hl_cuda_lstm.cu b/paddle/cuda/src/hl_cuda_lstm.cu deleted file mode 100644 index b8c4e433a118fb1c5af753751f91c34543b1114c..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_lstm.cu +++ /dev/null @@ -1,876 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_activation_functions.h" -#include "hl_base.h" -#include "hl_cuda_cublas.h" -#include "hl_device_functions.cuh" -#include "paddle/utils/Logging.h" - -typedef hppl::Active::forward t_forward; -typedef hppl::Active::backward t_backward; - -bool hl_lstm_sequence_parallel(int frameSize) { - if (frameSize == 32 || frameSize == 64) { - return true; - } else { - return false; - } -} - -class frameValue { - public: - real *value_; - __device__ frameValue(real *value) : value_(value) {} - template - __device__ inline void init(int start, int length, int idx) { - if (reversed == 0) { - value_ += start * frameSize + idx; - } else { - value_ += (start + length - 1) * frameSize + idx; - } - } - __device__ inline real *getPtr() const { return value_; } - __device__ inline real getValue() { return *value_; } - __device__ inline void setValue(real value) { *value_ = value; } - template - __device__ inline void nextFrame() { - if (reversed == 0) { - value_ += frameSize; - } else { - value_ -= frameSize; - } - } -}; - -__device__ __forceinline__ void ptx_sync(const int id, const int barriers) { - asm volatile("bar.sync %0, %1;" : : "r"(id), "r"(barriers) : "memory"); -} - -__device__ __forceinline__ void ptx_arrive(const int id, const int barriers) { - asm volatile("bar.arrive %0, %1;" : : "r"(id), "r"(barriers) : "memory"); -} - -template -__device__ __forceinline__ real forward_sequence(real value, - real *shValue, - real *state, - real *preOutput, - real *output, - real check, - int index, - t_forward activeNode, - t_forward activeGate, - t_forward activeState) { - real out; - real prevOut; - real state_r; - const int idx = index % frameSize; - const int idy = index / frameSize; - // assert(index < valueSize); - - if (idy == 0) { - value = activeNode(value); - shValue[index] = value; - } - if (idy == 1 || idy == 2) { - state_r = state[idx]; - value += state_r * check; - value = activeGate(value); - shValue[index] = value; - } - ptx_sync(1, valueSize); - if (idy == 3) { - state_r = state[idx]; - state_r = state_r * shValue[idx + frameSize * 2]; - state_r += shValue[idx] * shValue[idx + frameSize]; - state[idx] = state_r; - ptx_arrive(2, frameSize * 2); - value += state_r * check; - value = activeGate(value); - shValue[index] = value; - ptx_sync(3, frameSize * 2); - prevOut = preOutput[idx]; - out = prevOut * value; - output[idx] = out; - } - if (idy == 0) { - ptx_sync(2, frameSize * 2); - prevOut = state[idx]; - prevOut = activeState(prevOut); - preOutput[idx] = prevOut; - ptx_arrive(3, frameSize * 2); - } - return value; -} - -#define OUTPUT_BARRIER_ID 10 -#define OUTPUT_BARRIER_ID2 11 -template -__global__ void KeLstmForward(real *gateValue, - real *state, - real *output, - real *preOutput, - real *checkIg, - real *checkFg, - real *checkOg, - real *weight, - const int *starts, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - __shared__ real shValue[valueSize]; - __shared__ real shState[frameSize]; - __shared__ real shPrevOutput[frameSize]; - __shared__ real shOutput[frameSize]; - - const int index = threadIdx.x; - int start = starts[blockIdx.x]; - int length = starts[blockIdx.x + 1] - start; - - /* init */ - real check; - real value; - frameValue frameGate(gateValue); - frameValue frameState(state); - frameValue frameOutput(output); - frameValue framePreOutput(preOutput); - if (index < valueSize) { - const int idx = index % frameSize; - const int idy = index / frameSize; - frameGate.init(start, length, index); - value = frameGate.getValue(); - if (idy == 0) { - shState[idx] = 0.0; - } else if (idy == 1) { - check = checkIg[idx]; - } else if (idy == 2) { - check = checkFg[idx]; - } else if (idy == 3) { - check = checkOg[idx]; - } - - if (idy == 3) { - frameState.init(start, length, idx); - frameOutput.init(start, length, idx); - framePreOutput.init(start, length, idx); - } - - ptx_sync(1, valueSize); - } - - for (int i = 0; i < length; ++i) { - if (index < valueSize) { - if (valueSize == 128) { - if (i != 0) { - ptx_sync(OUTPUT_BARRIER_ID2, blockSize); - value += shValue[index]; - } - } - value = forward_sequence( - value, - shValue, - shState, - shPrevOutput, - shOutput, - check, - index, - hppl::gpu::forward[active_node], - hppl::gpu::forward[active_gate], - hppl::gpu::forward[active_state]); - const int idx = index % frameSize; - const int idy = index / frameSize; - if (valueSize == 128) { - if (idy == 3) { - ptx_arrive(OUTPUT_BARRIER_ID, frameSize + 128); - } - } - if (valueSize == 256) { - ptx_sync(OUTPUT_BARRIER_ID, valueSize); - } - frameGate.setValue(value); - if (idy == 3) { - frameState.setValue(shState[idx]); - frameOutput.setValue(shOutput[idx]); - framePreOutput.setValue(shPrevOutput[idx]); - frameState.nextFrame(); - frameOutput.nextFrame(); - framePreOutput.nextFrame(); - } - if (i != length - 1) { - frameGate.nextFrame(); - value = frameGate.getValue(); - } - } - if (i != length - 1) { - if (valueSize == 128) { - if (valueSize <= index) { - real B_r[frameSize]; - const int computeIdx = index - valueSize; - if (i == 0) { -#pragma unroll - for (int n = 0; n < frameSize; n++) { - B_r[n] = weight[n * valueSize + computeIdx]; - } - } - ptx_sync(OUTPUT_BARRIER_ID, frameSize + 128); - real A_r[frameSize]; - for (int n = 0; n < frameSize; n++) { - A_r[n] = shOutput[n]; - } - real sum = 0.0f; - for (int n = 0; n < frameSize; n++) { - sum += A_r[n] * B_r[n]; - } - shValue[computeIdx] = sum; - ptx_arrive(OUTPUT_BARRIER_ID2, blockSize); - } - } - if (valueSize == 256) { - real B_r[frameSize]; - if (i == 0) { -#pragma unroll - for (int n = 0; n < frameSize; n++) { - B_r[n] = weight[n * valueSize + index]; - } - } - real sum = 0.0f; - for (int n = 0; n < frameSize; n++) { - sum += shOutput[n] * B_r[n]; - } - value += sum; - } - } - } -} - -void hl_lstm_parallel_forward(real *gateValue, - real *stateValue, - real *preOutputValue, - real *outputValue, - real *checkIg, - real *checkFg, - real *checkOg, - real *weight, - const int *sequence, - int frameSize, - int numSequences, - bool reversed, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - CHECK(frameSize == 32 || frameSize == 64); - dim3 grid(numSequences, 1); - if (!reversed) { - if (frameSize == 32) { - KeLstmForward<128, 32, 0, 128, 256><<>>( - gateValue, - stateValue, - outputValue, - preOutputValue, - checkIg, - checkFg, - checkOg, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 64) { - KeLstmForward<256, 64, 0, 256, 256><<>>( - gateValue, - stateValue, - outputValue, - preOutputValue, - checkIg, - checkFg, - checkOg, - weight, - sequence, - active_node, - active_gate, - active_state); - } - } else { - if (frameSize == 32) { - KeLstmForward<128, 32, 1, 128, 256><<>>( - gateValue, - stateValue, - outputValue, - preOutputValue, - checkIg, - checkFg, - checkOg, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 64) { - KeLstmForward<256, 64, 1, 256, 256><<>>( - gateValue, - stateValue, - outputValue, - preOutputValue, - checkIg, - checkFg, - checkOg, - weight, - sequence, - active_node, - active_gate, - active_state); - } - } - CHECK_SYNC("hl_lstm_parallel_forward failed"); -} - -__device__ __forceinline__ void transpose_32x32(real a[], const int idx) { - const int warp_size = 32; - int addr = idx % warp_size; - unsigned mask = 0u; - CREATE_SHFL_MASK(mask, addr < warp_size); -#pragma unroll - for (int k = 1; k < 32; k++) { - // rSrc[k] = __shfl_sync(rSrc[k], (threadIdx.x + k) % 32, 32); - addr = __shfl_sync(mask, addr, (idx + 1) % 32, 32); - a[k] = __shfl_sync(mask, a[k], addr, 32); - } - -#pragma unroll - for (int tid = 0; tid < 31; tid++) { - real tmp = (idx > tid) ? a[0] : a[1]; -#pragma unroll - for (int k = 31; k > 0; k--) { - a[(k + 1) % 32] = (idx > tid) ? a[k] : a[(k + 1) % 32]; - } - a[1] = tmp; - } - - addr = (32 - idx) % 32; - CREATE_SHFL_MASK(mask, idx % 32 < warp_size); -#pragma unroll - for (int k = 0; k < 32; k++) { - a[k] = __shfl_sync(mask, a[k], addr, 32); - addr = __shfl_sync(mask, addr, (idx + 31) % 32, 32); - } -} - -template -__device__ void backward_sequence(real rGateValue, - real rOutputGrad, - real rPreOutputValue, - real &rGateGrad, - real &rStateGrad, - real *shStateGrad, - real *shStateValue, - real *shGateValue, - real rCheck, - real &rGateValuePrev, - int index, - t_backward activeNode, - t_backward activeGate, - t_backward activeState) { - const int frameIdx = index % frameSize; - const int frameIdy = index / frameSize; - if (frameIdy == 3) { - real rPrevOutputGrad; - rPrevOutputGrad = rOutputGrad * rGateValue; - rStateGrad = activeState(rPrevOutputGrad, rPreOutputValue); - rGateGrad = rOutputGrad * rPreOutputValue; - rGateGrad = activeGate(rGateGrad, rGateValue); - rStateGrad += rGateGrad * rCheck; - shStateGrad[index] = rStateGrad; - ptx_arrive(3, valueSize); - } else if (frameIdy == 1) { - shGateValue[frameIdx + frameSize] = rGateValue; - rStateGrad = rGateGrad * rCheck; - shStateGrad[index] = rStateGrad; - ptx_sync(3, valueSize); - rStateGrad += shStateGrad[frameIdx + frameSize * 2]; - rStateGrad += shStateGrad[frameIdx + frameSize * 3]; - rGateGrad = rStateGrad * shGateValue[frameIdx]; - rGateGrad = activeGate(rGateGrad, rGateValue); - } else if (frameIdy == 2) { - rStateGrad = rStateGrad * rGateValuePrev; - rStateGrad += rGateGrad * rCheck; - shStateGrad[index] = rStateGrad; - ptx_sync(3, valueSize); - rStateGrad += shStateGrad[frameIdx + frameSize]; - rStateGrad += shStateGrad[frameIdx + frameSize * 3]; - rGateValuePrev = rGateValue; - rGateGrad = rStateGrad * shStateValue[frameIdx]; - rGateGrad = activeGate(rGateGrad, rGateValue); - } else if (frameIdy == 0) { - shGateValue[frameIdx] = rGateValue; - ptx_sync(3, valueSize); - rStateGrad = shStateGrad[frameIdx + frameSize]; - rStateGrad += shStateGrad[frameIdx + frameSize * 2]; - rStateGrad += shStateGrad[frameIdx + frameSize * 3]; - rGateGrad = rStateGrad * shGateValue[frameIdx + frameSize]; - rGateGrad = activeNode(rGateGrad, rGateValue); - } -} - -template -__device__ void load_weight(real rWeight[], real *weight, const int index) { - if (valueSize == 128) { - weight += index; -#pragma unroll - for (int n = 0; n < frameSize; n++) { - rWeight[n] = weight[n * valueSize]; - } - transpose_32x32(rWeight, index % 32); - } - if (valueSize == 256) { - int id = (index / 32) % 2; - weight += index - id * 32 + id * 32 * valueSize; -#pragma unroll - for (int n = 0; n < 32; n++) { - rWeight[n] = weight[n * valueSize]; - rWeight[n + 32] = weight[n * valueSize + 32]; - } - transpose_32x32(rWeight, index % 32); - transpose_32x32(&rWeight[32], index % 32); - } -} - -template -__global__ void KeLstmBackward(real *gateValue, - real *gateGrad, - real *stateValue, - real *stateGrad, /* do not need save */ - real *preOutputValue, - real *preOutputGrad, /* do not need save */ - real *checkIg, - real *checkIgGrad, - real *checkFg, - real *checkFgGrad, - real *checkOg, - real *checkOgGrad, - real *outputGrad, - real *weightValue, - const int *starts, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - __shared__ real shGateValue[valueSize]; - __shared__ real shStateGrad[valueSize]; - __shared__ real shStateValue[frameSize]; - __shared__ real shGateGrad[4][frameSize]; - __shared__ real shOutputGrad[4][frameSize]; - const int index = threadIdx.x; - int start = starts[blockIdx.x]; - int length = starts[blockIdx.x + 1] - start; - - const int frameIdx = index % frameSize; - const int frameIdy = index / frameSize; - real rCheck; - real rCheckGrad; - real rGateGrad; - real rStateGrad; - real rGateValuePrev; - real rPreOutputValue; - real rOutputGrad; - real rGateValue; - real rStateValue; - - frameValue frameGateValue(gateValue); - frameValue frameGateGrad(gateGrad); - frameValue framePreOutputValue(preOutputValue); - frameValue frameStateValue(stateValue); - frameValue frameOutputGrad(outputGrad); - if (frameIdy == 0) { - } else if (frameIdy == 1) { - rCheck = checkIg[frameIdx]; - } else if (frameIdy == 2) { - rCheck = checkFg[frameIdx]; - rGateValuePrev = 0.0; - rStateGrad = 0.0; - } else if (frameIdy == 3) { - rCheck = checkOg[frameIdx]; - framePreOutputValue.init(start, length, frameIdx); - frameOutputGrad.init(start, length, frameIdx); - rOutputGrad = frameOutputGrad.getValue(); - rPreOutputValue = framePreOutputValue.getValue(); - frameStateValue.init(start, length, frameIdx); - rStateValue = frameStateValue.getValue(); - } - - frameGateValue.init(start, length, index); - frameGateGrad.init(start, length, index); - rGateValue = frameGateValue.getValue(); - rGateGrad = 0.0; - rCheckGrad = 0.0; - - real B_r[frameSize]; - load_weight(B_r, weightValue, index); - - for (int i = 0; i < length; ++i) { - if (frameIdy == 3) { - if (i != length - 1) { - frameStateValue.nextFrame(); - shStateValue[frameIdx] = frameStateValue.getValue(); - } else { - shStateValue[frameIdx] = 0.0; - } - } - backward_sequence(rGateValue, - rOutputGrad, - rPreOutputValue, - rGateGrad, - rStateGrad, - shStateGrad, - shStateValue, - shGateValue, - rCheck, - rGateValuePrev, - index, - hppl::gpu::backward[active_node], - hppl::gpu::backward[active_gate], - hppl::gpu::backward[active_state]); - if (frameIdy == 3) { - rCheckGrad += rGateGrad * rStateValue; - rStateValue = shStateValue[frameIdx]; - } - - frameGateGrad.setValue(rGateGrad); - frameGateGrad.nextFrame(); - - if (i != length - 1) { - if (frameIdy == 3) { - framePreOutputValue.nextFrame(); - rPreOutputValue = framePreOutputValue.getValue(); - frameOutputGrad.nextFrame(); - rOutputGrad = frameOutputGrad.getValue(); - } else if (frameIdy == 2) { - rCheckGrad += rGateGrad * shStateValue[frameIdx]; - } else if (frameIdy == 1) { - rCheckGrad += rGateGrad * shStateValue[frameIdx]; - } - - frameGateValue.nextFrame(); - rGateValue = frameGateValue.getValue(); - shGateGrad[frameIdy][frameIdx] = rGateGrad; - if (valueSize == 128) { - real sum = 0.0f; -#pragma unroll - for (int n = 0; n < frameSize; n++) { - sum += shGateGrad[frameIdy][n] * B_r[n]; - } - if (frameIdy == 3) { - rOutputGrad += sum; - } else { - shOutputGrad[frameIdy][frameIdx] = sum; - } - } - if (valueSize == 256) { - ptx_sync(5, valueSize); - real A_r[frameSize]; - for (int n = 0; n < frameSize; n++) { - A_r[n] = shGateGrad[frameIdy][n]; - } - real sum = 0.0f; - for (int n = 0; n < frameSize; n++) { - sum += A_r[n] * B_r[n]; - } - if (frameIdy == 3) { - rOutputGrad += sum; - } else { - shOutputGrad[frameIdy][frameIdx] = sum; - } - } - - if (frameIdy == 3) { - ptx_sync(6, valueSize); -#pragma unroll - for (int i = 0; i < 3; i++) { - rOutputGrad += shOutputGrad[i][frameIdx]; - } - } else { - ptx_arrive(6, valueSize); - } - } - } - - /* TODO: Temporary save & merger in another kernel */ - if (frameIdy == 1) { - if (checkIgGrad) - paddle::paddleAtomicAdd(checkIgGrad + frameIdx, rCheckGrad); - } else if (frameIdy == 2) { - if (checkFgGrad) - paddle::paddleAtomicAdd(checkFgGrad + frameIdx, rCheckGrad); - } else if (frameIdy == 3) { - if (checkOgGrad) - paddle::paddleAtomicAdd(checkOgGrad + frameIdx, rCheckGrad); - } -} - -void hl_lstm_parallel_backward_data(real *gateValue, - real *gateGrad, - real *stateValue, - real *stateGrad, - real *preOutputValue, - real *preOutputGrad, - real *outputGrad, - real *checkIg, - real *checkIgGrad, - real *checkFg, - real *checkFgGrad, - real *checkOg, - real *checkOgGrad, - real *weight, - const int *sequence, - int frameSize, - int numSequences, - bool reversed, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate, - hl_activation_mode_t active_state) { - CHECK(frameSize == 32 || frameSize == 64 || frameSize == 128 || - frameSize == 256); - dim3 grid(numSequences, 1); - if (!reversed) { - if (frameSize == 32) { - KeLstmBackward<128, 32, 0><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 64) { - KeLstmBackward<256, 64, 0><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 128) { - KeLstmBackward<512, 128, 0><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 256) { - KeLstmBackward<1024, 256, 0><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } - } else { - if (frameSize == 32) { - KeLstmBackward<128, 32, 1><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 64) { - KeLstmBackward<256, 64, 1><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 128) { - KeLstmBackward<512, 128, 1><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } else if (frameSize == 256) { - KeLstmBackward<1024, 256, 1><<>>( - gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - outputGrad, - weight, - sequence, - active_node, - active_gate, - active_state); - } - } - CHECK_SYNC("hl_lstm_parallel_backward_data"); -} - -template -__global__ void KeSetGradZero(real *gateGrad, - const int *starts, - int valueSize, - int numSequences, - bool reversed) { - // const int tid = threadIdx.x; - - const int frameIdx = blockIdx.x * B_X + threadIdx.x; - const int numSeqId = blockIdx.y * B_Y + threadIdx.y; - - if (numSeqId >= numSequences || frameIdx >= valueSize) return; - - if (!reversed) { - int seqId = starts[numSeqId]; - gateGrad[seqId * valueSize + frameIdx] = 0.0; - } else { - int seqId = starts[numSeqId + 1] - 1; - gateGrad[seqId * valueSize + frameIdx] = 0.0; - } -} - -void hl_lstm_parallel_backward_weight(real *weightGrad, - real *outputValue, - real *gateGrad, - const int *sequence, - int frameSize, - int batchSize, - int numSequences, - bool reversed) { - int valueSize = 4 * frameSize; - dim3 threads(32, 32); - dim3 grid((valueSize + 32 - 1) / 32, (numSequences + 32 - 1) / 32); - KeSetGradZero<32, 32><<>>( - gateGrad, sequence, valueSize, numSequences, reversed); - - if (!reversed) { - hl_matrix_mul(outputValue, - HPPL_OP_T, - gateGrad + valueSize, - HPPL_OP_N, - weightGrad, - frameSize, - valueSize, - batchSize - 1, - 1.0, - 1.0); - } else { - hl_matrix_mul(outputValue + frameSize, - HPPL_OP_T, - gateGrad, - HPPL_OP_N, - weightGrad, - frameSize, - valueSize, - batchSize - 1, - 1.0, - 1.0); - } - CHECK_SYNC("hl_lstm_parallel_backward_weight"); -} diff --git a/paddle/cuda/src/hl_cuda_matrix.cu b/paddle/cuda/src/hl_cuda_matrix.cu deleted file mode 100644 index 3e17c8090c5036037e936af1d6feaa2239251679..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_matrix.cu +++ /dev/null @@ -1,806 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_base.h" -#include "hl_device_functions.cuh" -#include "hl_gpu_matrix_kernel.cuh" -#include "hl_matrix.h" -#include "hl_matrix_apply.cuh" -#include "hl_matrix_ops.cuh" -#include "hl_sequence.h" -#include "hl_sparse.ph" -#include "paddle/utils/Logging.h" - -DEFINE_MATRIX_UNARY_OP(Zero, a = 0); -DEFINE_MATRIX_TERNARY_PARAMETER_OP(_add, TWO_PARAMETER, c = p1 * a + p2 * b); -void hl_matrix_add(real* A_d, - real* B_d, - real* C_d, - int dimM, - int dimN, - real alpha, - real beta) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - - hl_gpu_apply_ternary_op, 0, 0>( - ternary::_add(alpha, beta), - A_d, - B_d, - C_d, - dimM, - dimN, - dimN, - dimN, - dimN); - CHECK_SYNC("hl_matrix_add failed"); -} - -#ifdef PADDLE_TYPE_DOUBLE -#define THRESHOLD 128 -#else -#define THRESHOLD 64 -#endif -__device__ __forceinline__ void findMax(real* I, - real* dfMax_s, - int blockSize, - int base, - int curIdx, - int nextIdx, - int dimN, - real* max) { - dfMax_s[base] = -1.0e20; - while (curIdx < dimN) { - if (dfMax_s[base] < I[nextIdx]) { - dfMax_s[base] = I[nextIdx]; - } - nextIdx += blockSize; - curIdx += blockSize; - } - __syncthreads(); - - for (int stride = blockSize >> 1; stride > 0; stride >>= 1) { - __syncthreads(); - if (base < stride) { - nextIdx = base + stride; - if (dfMax_s[base] < dfMax_s[nextIdx]) { - dfMax_s[base] = dfMax_s[nextIdx]; - } - } - } - - if (0 == base) { - max[0] = dfMax_s[0]; - } - __syncthreads(); -} - -__device__ __forceinline__ void subMaxAndExp(real* I, - real* O, - int curIdx, - int nextIdx, - int blockSize, - int dimN, - real max) { - real val; - while (curIdx < dimN) { - val = I[nextIdx] - max; - if (val < -THRESHOLD) { - val = -THRESHOLD; - } - I[nextIdx] = val; -#ifndef PADDLE_TYPE_DOUBLE - O[nextIdx] = __expf(val); -#else - O[nextIdx] = exp(val); -#endif - nextIdx += blockSize; - curIdx += blockSize; - } - __syncthreads(); -} - -__device__ __forceinline__ void valueSum(real* O, - real* dfMax_s, - int blockSize, - int base, - int curIdx, - int nextIdx, - int dimN) { - dfMax_s[base] = 0; - while (curIdx < dimN) { - dfMax_s[base] += O[nextIdx]; - nextIdx += blockSize; - curIdx += blockSize; - } - __syncthreads(); - - for (int stride = blockSize >> 1; stride > 0; stride >>= 1) { - __syncthreads(); - if (base < stride) { - nextIdx = base + stride; - dfMax_s[base] += dfMax_s[nextIdx]; - } - } - __syncthreads(); -} - -__device__ __forceinline__ void divSum( - real* O, real sum, int curIdx, int nextIdx, int blockSize, int dimN) { - while (curIdx < dimN) { - O[nextIdx] /= sum; - nextIdx += blockSize; - curIdx += blockSize; - } -} - -__device__ __forceinline__ void softmax(real* I, - real* O, - real* dfMax_s, - int blockSize, - int base, - int curIdx, - int nextIdx, - int dimN) { - __shared__ real max; - - // find the max number - findMax(I, dfMax_s, blockSize, base, curIdx, nextIdx, dimN, &max); - - // sub max Value and do Exp operation - subMaxAndExp(I, O, base, nextIdx, blockSize, dimN, max); - - // add dimN values into blockDim.x buffer - // sum is in dfMax_s[0] - valueSum(O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); - - // divided by sum - divSum(O, dfMax_s[0], curIdx, nextIdx, blockSize, dimN); -} - -template -__global__ void KeMatrixSoftMax(real* O, real* I, int dimN) { - int base = threadIdx.x; - __shared__ real dfMax_s[blockSize]; - int nextIdx = blockIdx.x * dimN + base; - int curIdx = base; - - softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); -} - -void hl_matrix_softmax(real* A_d, real* C_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - dim3 block(512, 1); - dim3 grid(dimM, 1); - KeMatrixSoftMax<512><<>>(C_d, A_d, dimN); - CHECK_SYNC("hl_matrix_softmax failed"); -} - -template -__global__ void KeSequenceSoftMax(real* O, real* I, const int* index) { - int base = threadIdx.x; - int bid = blockIdx.x; - __shared__ real dfMax_s[blockSize]; - - int start = index[bid]; - int dimN = index[bid + 1] - start; - - int nextIdx = start + base; - int curIdx = base; - - softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); -} - -void hl_sequence_softmax_forward(real* A_d, - real* C_d, - const int* index, - int numSequence) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - dim3 block(512, 1); - dim3 grid(numSequence, 1); - KeSequenceSoftMax<512><<>>(C_d, A_d, index); - CHECK_SYNC("hl_sequence_softmax_forward failed"); -} - -__global__ void KeMatrixDerivative( - real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) { - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - int colIdx = blockIdx.y * blockDim.y + threadIdx.y; - int index; - - if (rowIdx < dimM && colIdx < dimN) { - index = rowIdx * dimN + colIdx; - grad_d[index] = output_d[index] * (grad_d[index] - sftmaxSum_d[rowIdx]); - } -} - -void hl_matrix_softmax_derivative( - real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) { - CHECK_NOTNULL(grad_d); - CHECK_NOTNULL(output_d); - CHECK_NOTNULL(sftmaxSum_d); - - int blocksX = (dimM + 0) / 1; - int blocksY = (dimN + 1024 - 1) / 1024; - dim3 threads(1, 1024); - dim3 grid(blocksX, blocksY); - - KeMatrixDerivative<<>>( - grad_d, output_d, sftmaxSum_d, dimM, dimN); - CHECK_SYNC("hl_matrix_softmax_derivative failed"); -} - -__global__ void KeMatrixMultiBinaryCrossEntropy( - real* output, real* entropy, int* row, int* col, int dimM, int dimN) { - int index = blockIdx.x * blockDim.x + threadIdx.x; - if (index < dimM) { - for (int i = 0; i < dimN; i++) { - entropy[index] -= log(1 - output[index * dimN + i]); - } - int* row_col = col + row[index]; - int col_num = row[index + 1] - row[index]; - for (int i = 0; i < col_num; i++) { - real o = output[index * dimN + row_col[i]]; - entropy[index] -= log(o / (1 - o)); - } - } -} - -void hl_matrix_multi_binary_cross_entropy(real* output, - real* entropy, - hl_sparse_matrix_s csr_mat, - int dimM, - int dimN) { - CHECK_NOTNULL(output); - CHECK_NOTNULL(entropy); - CHECK_NOTNULL(csr_mat); - CHECK_EQ(csr_mat->format, HL_SPARSE_CSR); - int n_threads = 1024; - int blocks = (dimM + n_threads - 1) / n_threads; - dim3 threads(n_threads); - dim3 grid(blocks); - hl_csr_matrix mat = (hl_csr_matrix)(csr_mat->matrix); - KeMatrixMultiBinaryCrossEntropy<<>>( - output, entropy, mat->csr_row, mat->csr_col, dimM, dimN); - CHECK_SYNC("hl_matrix_multi_binary_cross_entropy failed"); -} - -__global__ void KeMatrixMultiBinaryCrossEntropyBp( - real* output, real* grad, int* row, int* col, int dimM, int dimN) { - int row_idx = blockIdx.x * blockDim.x + threadIdx.x; - if (row_idx < dimM) { - for (int i = 0; i < dimN; i++) { - int index = row_idx * dimN + i; - grad[index] += 1.0 / (1 - output[index]); - } - int col_num = row[row_idx + 1] - row[row_idx]; - int* row_col = col + row[row_idx]; - for (int i = 0; i < col_num; i++) { - int index = row_idx * dimN + row_col[i]; - grad[index] -= 1.0 / (output[index] * (1 - output[index])); - } - } -} - -void hl_matrix_multi_binary_cross_entropy_bp( - real* output, real* grad, hl_sparse_matrix_s csr_mat, int dimM, int dimN) { - CHECK_NOTNULL(output); - CHECK_NOTNULL(grad); - CHECK_NOTNULL(csr_mat); - CHECK_EQ(csr_mat->format, HL_SPARSE_CSR); - int n_threads = 1024; - int blocks = (dimM + n_threads - 1) / n_threads; - dim3 threads(n_threads); - dim3 grid(blocks); - hl_csr_matrix mat = (hl_csr_matrix)(csr_mat->matrix); - KeMatrixMultiBinaryCrossEntropyBp<<>>( - output, grad, mat->csr_row, mat->csr_col, dimM, dimN); - CHECK_SYNC("hl_matrix_multi_binary_cross_entropy_bp failed"); -} - -__global__ void KeMatrixCrossEntropy( - real* O, real* E, int* label, int dimM, int dimN) { - int index = blockIdx.x * blockDim.x + threadIdx.x; - int newBase; - if (index < dimM) { - newBase = label[index]; - newBase = newBase % dimN; - E[index] = -log(O[index * dimN + newBase]); - } -} - -void hl_matrix_cross_entropy( - real* A_d, real* C_d, int* label_d, int dimM, int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - - int blocks = (dimM + 1024 - 1) / 1024; - dim3 threads(1024, 1); - dim3 grid(blocks, 1); - KeMatrixCrossEntropy<<>>( - A_d, C_d, label_d, dimM, dimN); - CHECK_SYNC("hl_matrix_cross_entropy failed"); -} - -__global__ void KeMatrixCrossEntropyBp( - real* grad_d, real* output_d, int* label_d, int dimM, int dimN) { - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - int colIdx = blockIdx.y * blockDim.y + threadIdx.y; - int index; - if (rowIdx < dimM && colIdx < dimN) { - index = rowIdx * dimN + colIdx; - if (label_d[rowIdx] == colIdx) { - grad_d[index] -= 1.0f / output_d[index]; - } - } -} - -void hl_matrix_cross_entropy_bp( - real* grad_d, real* output_d, int* label_d, int dimM, int dimN) { - CHECK_NOTNULL(grad_d); - CHECK_NOTNULL(output_d); - CHECK_NOTNULL(label_d); - - int blocksX = (dimM + 0) / 1; - int blocksY = (dimN + 1024 - 1) / 1024; - dim3 threads(1, 1024); - dim3 grid(blocksX, blocksY); - KeMatrixCrossEntropyBp<<>>( - grad_d, output_d, label_d, dimM, dimN); - CHECK_SYNC("hl_matrix_cross_entropy_bp failed"); -} - -void hl_matrix_zero_mem(real* data, int num) { - hl_gpu_apply_unary_op(unary::Zero(), data, 1, num, num); -} - -__global__ void KeParamReluForward(real* output, - real* input, - real* w, - int width, - int height, - int partial_sum) { - int tx = blockIdx.x * blockDim.x + threadIdx.x; - int ty = blockIdx.y * blockDim.y + threadIdx.y; - if (tx < width && ty < height) { - int index = ty * width + tx; - output[index] = - input[index] > 0 ? input[index] : input[index] * w[tx / partial_sum]; - } -} - -void hl_param_relu_forward(real* output, - real* input, - real* w, - int width, - int height, - int partial_sum) { - CHECK_NOTNULL(output); - CHECK_NOTNULL(input); - CHECK_NOTNULL(w); - dim3 threads(16, 16); - int blockX = (width + 16 - 1) / 16; - int blockY = (height + 16 - 1) / 16; - dim3 grid(blockX, blockY); - KeParamReluForward<<>>( - output, input, w, width, height, partial_sum); - CHECK_SYNC("hl_param_relu_forward failed"); -} - -template -__global__ void KeParamReluBackWardW(real* grad_w, - real* grad_o, - real* input, - int width, - int height, - int partial_sum) { - const int tid = threadIdx.x; - __shared__ real temp[blockSize]; - grad_o += partial_sum * blockIdx.x; - input += partial_sum * blockIdx.x; - real tmp = 0.0; - for (int index = tid; index < partial_sum * height; index += blockSize) { - int row = index / partial_sum; - int offset = row * width + (index - row * partial_sum); - if (input[offset] < 0) { - tmp += grad_o[offset] * input[offset]; - } - } - temp[tid] = tmp; - __syncthreads(); - for (int s = blockSize / 2; s > 0; s >>= 1) { - if (tid < s) { - temp[tid] += temp[tid + s]; - } - __syncthreads(); - } - if (tid == 0) { - grad_w[blockIdx.x] += temp[0]; - } -} - -void hl_param_relu_backward_w(real* grad_w, - real* grad_o, - real* input, - int width, - int height, - int partial_sum) { - CHECK_NOTNULL(grad_w); - CHECK_NOTNULL(grad_o); - CHECK_NOTNULL(input); - const int blockSize = 1024; - int grid_num = width / partial_sum; - dim3 threads(blockSize, 1); - dim3 grid(grid_num, 1); - KeParamReluBackWardW<<>>( - grad_w, grad_o, input, width, height, partial_sum); - CHECK_SYNC("hl_param_relu_backward_w failed"); -} - -__global__ void KeParamReluBackwardDiff(real* grad_o, - real* input, - real* w, - real* diff, - int width, - int height, - int partial_sum) { - int tx = blockIdx.x * blockDim.x + threadIdx.x; - int ty = blockIdx.y * blockDim.y + threadIdx.y; - if (tx < width && ty < height) { - int index = ty * width + tx; - diff[index] += grad_o[index] * (input[index] > 0 ? 1 : w[tx / partial_sum]); - } -} - -void hl_param_relu_backward_diff(real* grad_o, - real* data, - real* w, - real* diff, - int width, - int height, - int partial_sum) { - CHECK_NOTNULL(grad_o); - CHECK_NOTNULL(data); - CHECK_NOTNULL(w); - CHECK_NOTNULL(diff); - dim3 threads(16, 16); - int blockX = (width + 16 - 1) / 16; - int blockY = (height + 16 - 1) / 16; - dim3 grid(blockX, blockY); - KeParamReluBackwardDiff<<>>( - grad_o, data, w, diff, width, height, partial_sum); - CHECK_SYNC("hl_param_relu_backward_diff failed"); -} - -__global__ void KeMatrixAddSharedBias( - real* A, real* B, const int channel, const int M, const int N, real scale) { - int index = blockIdx.x * blockDim.x + threadIdx.x; - int dim = N / channel; - if (index < M * N) { - int i = index % N; - i = i / dim; - A[index] += scale * B[i]; - } -} - -void hl_matrix_add_shared_bias(real* A_d, - real* B_d, - const int channel, - const int dimM, - const int dimN, - real scale) { - const int blocks = 512; - const int grids = DIVUP(dimM * dimN, blocks); - KeMatrixAddSharedBias<<>>( - A_d, B_d, channel, dimM, dimN, scale); - CHECK_SYNC("hl_matrix_add_shared_bias failed"); -} - -template -__global__ void KeMatrixCollectSharedBias(real* B, - real* A, - const int channel, - const int M, - const int N, - const int dim, - const int limit, - real scale) { - if (dim < limit) { - int index = blockIdx.x * blockDim.x + threadIdx.x; - if (index < channel) { - real sum = 0.0; - for (int i = 0; i < M; ++i) { - for (int j = 0; j < dim; ++j) { - sum += A[i * N + index * dim + j]; - } - } - B[index] += scale * sum; - } - } else { - const int tid = threadIdx.x; - const int bid = blockIdx.x; - __shared__ real smem[blockSize]; - real sum = 0.0; - for (int j = 0; j < ((dim * M + blockSize - 1) / blockSize); ++j) { - int n = j * blockSize + tid; - int m = n / dim; - int w = n % dim; - smem[tid] = (m < M && w < dim) ? A[m * N + bid * dim + w] : 0.0; - __syncthreads(); - simpleReduce(smem, tid, blockSize); - sum += smem[0]; - } - if (tid == 0) { - B[bid] += scale * sum; - } - } -} - -void hl_matrix_collect_shared_bias(real* B_d, - real* A_d, - const int channel, - const int dimM, - const int dimN, - real scale) { - const int dim = dimN / channel; - const int blocks = 256; - const int limit = 64; - int grids = (dimM * dim) < limit ? DIVUP(channel, blocks) : channel; - - KeMatrixCollectSharedBias<<>>( - B_d, A_d, channel, dimM, dimN, dim, limit, scale); - CHECK_SYNC("hl_matrix_collect_shared_bias failed"); -} - -__global__ void keMatrixRotate( - real* mat, real* matRot, int dimM, int dimN, bool clockWise) { - int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < dimM * dimN) { - int i = idx / dimN; - int j = idx % dimN; - if (clockWise) { - matRot[j * dimM + i] = mat[(dimM - i - 1) * dimN + j]; - } else { - matRot[j * dimM + i] = mat[i * dimN + (dimN - j - 1)]; - } - } -} - -void hl_matrix_rotate( - real* mat, real* matRot, int dimM, int dimN, bool clockWise) { - CHECK_NOTNULL(mat); - CHECK_NOTNULL(matRot); - const int threads = 512; - const int blocks = DIVUP(dimM * dimN, threads); - keMatrixRotate<<>>( - mat, matRot, dimM, dimN, clockWise); - CHECK_SYNC("hl_matrix_rotate failed"); -} - -__global__ void keMatrixVol2Col(int num_kernels, - const real* dataSrc, - real* dataDst, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - int depth_col, - int height_col, - int width_col) { - for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; - index += blockDim.x * gridDim.x) { - int w_out = index % width_col; - int h_out = (index / width_col) % height_col; - int d_out = (index / width_col / height_col) % depth_col; - int channel_in = index / width_col / height_col / depth_col; - int channel_out = channel_in * filterD * filterH * filterW; - int w_in = w_out * strideW - paddingW; - int h_in = h_out * strideH - paddingH; - int d_in = d_out * strideD - paddingD; - - dataDst += - ((channel_out * depth_col + d_out) * height_col + h_out) * width_col + - w_out; - dataSrc += ((channel_in * depth + d_in) * height + h_in) * width + w_in; - for (int k = 0; k < filterD; ++k) { - for (int i = 0; i < filterH; ++i) { - for (int j = 0; j < filterW; ++j) { - int d = d_in + k; - int h = h_in + i; - int w = w_in + j; - *dataDst = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 && - w < width) - ? dataSrc[(k * height + i) * width + j] - : 0; - dataDst += depth_col * height_col * width_col; - } - } - } - } -} - -void hl_matrix_vol2Col(const real* dataSrc, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - real* dataDst) { - int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1; - int height_col = (height + 2 * paddingH - filterH) / strideH + 1; - int width_col = (width + 2 * paddingW - filterW) / strideW + 1; - int num_kernels = channels * depth_col * height_col * width_col; - - const int threads = 512; - const int blocks = DIVUP(num_kernels, threads); - - keMatrixVol2Col<<>>(num_kernels, - dataSrc, - dataDst, - depth, - height, - width, - filterD, - filterH, - filterW, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - depth_col, - height_col, - width_col); - CHECK_SYNC("hl_matrix_vol2Col failed"); -} - -__global__ void keMatrixCol2Vol(int num_kernels, - real* dataDst, - const real* dataSrc, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - int depth_col, - int height_col, - int width_col, - real alpha, - real beta) { - for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; - index += blockDim.x * gridDim.x) { - real srcVal = 0; - real dstVal = dataDst[index]; - int w = index % width + paddingW; - int h = (index / width) % height + paddingH; - int d = (index / width / height) % depth + paddingD; - int c = index / width / height / depth; - // compute the start and end of the output - int w_col_start = (w < filterW) ? 0 : (w - filterW) / strideW + 1; - int w_col_end = min(w / strideW + 1, width_col); - int h_col_start = (h < filterH) ? 0 : (h - filterH) / strideH + 1; - int h_col_end = min(h / strideH + 1, height_col); - int d_col_start = (d < filterD) ? 0 : (d - filterD) / strideD + 1; - int d_col_end = min(d / strideD + 1, depth_col); - - int offset = (c * filterD * filterW * filterH + d * filterW * filterH + - h * filterW + w) * - depth_col * height_col * width_col; - - int coeff_d_col = - (1 - strideD * filterW * filterH * depth_col) * height_col * width_col; - int coeff_h_col = - (1 - strideH * filterW * depth_col * height_col) * width_col; - int coeff_w_col = (1 - strideW * depth_col * height_col * width_col); - - for (int d_col = d_col_start; d_col < d_col_end; ++d_col) { - for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { - for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { - srcVal += dataSrc[offset + d_col * coeff_d_col + h_col * coeff_h_col + - w_col * coeff_w_col]; - } - } - } - dataDst[index] = alpha * srcVal + beta * dstVal; - } -} - -void hl_matrix_col2Vol(real* dataDst, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - const real* dataSrc, - real alpha, - real beta) { - int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1; - int height_col = (height + 2 * paddingH - filterH) / strideH + 1; - int width_col = (width + 2 * paddingW - filterW) / strideW + 1; - int num_kernels = channels * depth * height * width; - - const int threads = 512; - const int blocks = DIVUP(num_kernels, threads); - - keMatrixCol2Vol<<>>(num_kernels, - dataDst, - dataSrc, - depth, - height, - width, - filterD, - filterH, - filterW, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - depth_col, - height_col, - width_col, - alpha, - beta); - - CHECK_SYNC("hl_matrix_col2Vol failed"); -} - -__global__ void keVectorCast2Int(int* out, real* vec, int size) { - for (int i = threadIdx.x; i < (size); i += blockDim.x) { - out[i] = int(vec[i]); - } -} - -void hl_vector_cast2int(int* out, real* vec, int size) { - keVectorCast2Int<<<1, 512, 0, STREAM_DEFAULT>>>(out, vec, size); - CHECK_SYNC("hl_vector_cast2int failed"); -} diff --git a/paddle/cuda/src/hl_cuda_sequence.cu b/paddle/cuda/src/hl_cuda_sequence.cu deleted file mode 100644 index a3a5f038de7c0a68ee2e387d83b2272907164e90..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_sequence.cu +++ /dev/null @@ -1,408 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_base.h" -#include "hl_device_functions.cuh" -#include "paddle/utils/Logging.h" - -__global__ void KeMaxSequenceForward(real* input, - const int* sequence, - real* output, - int* index, - int numSequences, - int dim) { - int dimIdx = threadIdx.x; - int sequenceId = blockIdx.x; - if (sequenceId >= numSequences) return; - int start = sequence[sequenceId]; - int end = sequence[sequenceId + 1]; - - for (int i = dimIdx; i < dim; i += blockDim.x) { - real tmp = -HL_FLOAT_MAX; - int tmpId = -1; - for (int insId = start; insId < end; insId++) { - if (tmp < input[insId * dim + i]) { - tmp = input[insId * dim + i]; - tmpId = insId; - } - } - output[sequenceId * dim + i] = tmp; - index[sequenceId * dim + i] = tmpId; - } -} - -void hl_max_sequence_forward(real* input, - const int* sequence, - real* output, - int* index, - int numSequences, - int dim) { - CHECK_NOTNULL(input); - CHECK_NOTNULL(sequence); - CHECK_NOTNULL(output); - CHECK_NOTNULL(index); - - dim3 threads(256, 1); - dim3 grid(numSequences, 1); - KeMaxSequenceForward<<>>( - input, sequence, output, index, numSequences, dim); - CHECK_SYNC("hl_max_sequence_forward failed"); -} - -__global__ void KeMaxSequenceBackward( - real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) { - int idx = threadIdx.x + blockIdx.x * blockDim.x; - int colIdx = idx % dim; - if (idx < numSequences * dim) { - int insId = index[idx]; - inputGrad[insId * dim + colIdx] += outputGrad[idx]; - } -} - -void hl_max_sequence_backward( - real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) { - CHECK_NOTNULL(outputGrad); - CHECK_NOTNULL(index); - CHECK_NOTNULL(inputGrad); - - unsigned int blocks = (numSequences * dim + 128 - 1) / 128; - dim3 threads(128, 1); - dim3 grid(blocks, 1); - KeMaxSequenceBackward<<>>( - outputGrad, index, inputGrad, numSequences, dim); - CHECK_SYNC("hl_max_sequence_backward failed"); -} - -template -__global__ void KeMatrixAddRows(real* output, - real* table, - int* ids, - int numSamples, - int tableSize, - int dim) { - int idx = threadIdx.x; - int idy = threadIdx.y; - int sampleId = blockIdx.x + idy * gridDimX; - - while (sampleId < numSamples) { - int tableId = ids[sampleId]; - if ((0 <= tableId) && (tableId < tableSize)) { - real* outputData = output + sampleId * dim; - real* tableData = table + tableId * dim; - for (int i = idx; i < dim; i += blockDimX) { - if (AddRow == 0) { - outputData[i] += tableData[i]; - } else { - paddle::paddleAtomicAdd(&tableData[i], outputData[i]); - } - } - } - sampleId += blockDimY * gridDimX; - } -} - -template -__global__ void KeSequence2Batch(real* batch, - real* sequence, - const int* batchIndex, - int seqWidth, - int batchCount) { - int idx = threadIdx.x; - int idy = threadIdx.y; - int id = blockIdx.x + idy * gridDimX; - while (id < batchCount) { - int seqId = batchIndex[id]; - real* batchData = batch + id * seqWidth; - real* seqData = sequence + seqId * seqWidth; - for (int i = idx; i < seqWidth; i += blockDimX) { - if (seq2batch) { - if (isAdd) { - batchData[i] += seqData[i]; - } else { - batchData[i] = seqData[i]; - } - } else { - if (isAdd) { - seqData[i] += batchData[i]; - } else { - seqData[i] = batchData[i]; - } - } - } - id += blockDimY * gridDimX; - } -} - -void hl_sequence2batch_copy(real* batch, - real* sequence, - const int* batchIndex, - int seqWidth, - int batchCount, - bool seq2batch) { - CHECK_NOTNULL(sequence); - CHECK_NOTNULL(batch); - CHECK_NOTNULL(batchIndex); - - dim3 threads(128, 8); - dim3 grid(8, 1); - if (seq2batch) { - KeSequence2Batch<128, 8, 8, 1, 0><<>>( - batch, sequence, batchIndex, seqWidth, batchCount); - } else { - KeSequence2Batch<128, 8, 8, 0, 0><<>>( - batch, sequence, batchIndex, seqWidth, batchCount); - } - CHECK_SYNC("hl_sequence2batch_copy failed"); -} - -void hl_sequence2batch_add(real* batch, - real* sequence, - int* batchIndex, - int seqWidth, - int batchCount, - bool seq2batch) { - CHECK_NOTNULL(sequence); - CHECK_NOTNULL(batch); - CHECK_NOTNULL(batchIndex); - - dim3 threads(128, 8); - dim3 grid(8, 1); - if (seq2batch) { - KeSequence2Batch<128, 8, 8, 1, 1><<>>( - batch, sequence, batchIndex, seqWidth, batchCount); - } else { - KeSequence2Batch<128, 8, 8, 0, 1><<>>( - batch, sequence, batchIndex, seqWidth, batchCount); - } - CHECK_SYNC("hl_sequence2batch_add failed"); -} - -template -__global__ void KeSequence2BatchPadding(real* batch, - real* sequence, - const int* sequenceStartPositions, - const size_t sequenceWidth, - const size_t maxSequenceLength, - const size_t numSequences) { - int batchIdx = blockIdx.y; - int sequenceStart = sequenceStartPositions[batchIdx]; - int sequenceLength = sequenceStartPositions[batchIdx + 1] - sequenceStart; - - int sequenceIdx = blockIdx.x * blockDim.y + threadIdx.y; - int batchBaseIdx = (sequenceIdx * numSequences + batchIdx) * sequenceWidth; - int sequenceBaseIdx = (sequenceStart + sequenceIdx) * sequenceWidth; - - real scale = normByTimes ? (1.0f / (real)sequenceLength) : 1.0f; - - if (sequenceIdx < sequenceLength) { - if (seq2batch) { - /* sequence -> batch */ - for (int i = threadIdx.x; i < sequenceWidth; i += blockDim.x) { - batch[batchBaseIdx + i] = scale * sequence[sequenceBaseIdx + i]; - } - } else { - /* batch -> sequence */ - for (int i = threadIdx.x; i < sequenceWidth; i += blockDim.x) { - sequence[sequenceBaseIdx + i] = scale * batch[batchBaseIdx + i]; - } - } - } else if (sequenceIdx < maxSequenceLength) { - if (seq2batch) { - /* sequence -> batch */ - for (int i = threadIdx.x; i < sequenceWidth; i += blockDim.x) { - batch[batchBaseIdx + i] = 0; - } - } - } -} - -void hl_sequence2batch_copy_padding(real* batch, - real* sequence, - const int* sequenceStartPositions, - const size_t sequenceWidth, - const size_t maxSequenceLength, - const size_t numSequences, - bool normByTimes, - bool seq2batch) { - CHECK_NOTNULL(batch); - CHECK_NOTNULL(sequence); - CHECK_NOTNULL(sequenceStartPositions); - - if (!normByTimes && numSequences == 1) { - size_t elementCount = maxSequenceLength * sequenceWidth; - if (seq2batch) { - /* sequence -> batch */ - hl_memcpy_device2device(batch, sequence, sizeof(real) * elementCount); - } else { - /* batch -> sequence */ - hl_memcpy_device2device(sequence, batch, sizeof(real) * elementCount); - } - return; - } - - const int CUDA_BLOCK_SIZE = 512; - - /* At least use 32 threads to copy sequenceWidth elements, - and at least 8 elements for each thread. */ - int blockDimX = ((((sequenceWidth + 7) >> 3) + 31) >> 5) << 5; - blockDimX = (blockDimX < CUDA_BLOCK_SIZE) ? blockDimX : CUDA_BLOCK_SIZE; - - int blockDimY = CUDA_BLOCK_SIZE / blockDimX; - dim3 threads(blockDimX, blockDimY); - - int gridDimX = (maxSequenceLength + blockDimY - 1) / blockDimY; - int gridDimY = numSequences; - dim3 grid(gridDimX, gridDimY); - - if (seq2batch) { - /* sequence -> batch */ - if (normByTimes) { - KeSequence2BatchPadding<1, 1><<>>( - batch, - sequence, - sequenceStartPositions, - sequenceWidth, - maxSequenceLength, - numSequences); - } else { - KeSequence2BatchPadding<0, 1><<>>( - batch, - sequence, - sequenceStartPositions, - sequenceWidth, - maxSequenceLength, - numSequences); - } - } else { - /* batch -> sequence */ - if (normByTimes) { - KeSequence2BatchPadding<1, 0><<>>( - batch, - sequence, - sequenceStartPositions, - sequenceWidth, - maxSequenceLength, - numSequences); - } else { - KeSequence2BatchPadding<0, 0><<>>( - batch, - sequence, - sequenceStartPositions, - sequenceWidth, - maxSequenceLength, - numSequences); - } - } - - CHECK_SYNC("hl_sequence2batch_copy_padding failed"); -} - -__device__ inline float my_rsqrt(float x) { return rsqrtf(x); } - -__device__ inline double my_rsqrt(double x) { return rsqrt(x); } - -__global__ void KeSequenceAvgForward(real* dst, - real* src, - const int* starts, - int height, - int width, - const int mode) { - int gid = blockIdx.x * blockDim.x + threadIdx.x; - int row = gid / width; - int col = gid % width; - - if (gid < height * width) { - int start = starts[row]; - int end = starts[row + 1]; - int seqLength = end - start; - if (seqLength == 0) return; - real sum = 0.0; - for (int i = start; i < end; i++) { - sum += src[i * width + col]; - } - sum = mode == 1 ? sum : (mode == 0 ? sum / seqLength - : sum * my_rsqrt((real)seqLength)); - dst[gid] += sum; - } -} - -void hl_sequence_avg_forward(real* dst, - real* src, - const int* starts, - int height, - int width, - const int mode) { - CHECK_NOTNULL(dst); - CHECK_NOTNULL(src); - CHECK_NOTNULL(starts); - - int block = 512; - int grid = DIVUP(width * height, 512); - - CHECK(mode == 0 || mode == 1 || mode == 2) - << "mode error in hl_sequence_avg_forward!"; - - KeSequenceAvgForward<<>>( - dst, src, starts, height, width, mode); - CHECK_SYNC("hl_sequence_avg_forward failed"); -} - -__global__ void KeSequenceAvgBackward(real* dst, - real* src, - const int* starts, - int height, - int width, - const int mode) { - int gid = blockIdx.x * blockDim.x + threadIdx.x; - int row = gid / width; - int col = gid % width; - - if (gid < height * width) { - int start = starts[row]; - int end = starts[row + 1]; - int seqLength = end - start; - if (seqLength == 0) return; - real grad = src[gid]; - grad = mode == 1 ? grad : (mode == 0 ? grad / seqLength - : grad * my_rsqrt((real)seqLength)); - for (int i = start; i < end; i++) { - dst[i * width + col] += grad; - } - } -} - -void hl_sequence_avg_backward(real* dst, - real* src, - const int* starts, - int height, - int width, - const int mode) { - CHECK_NOTNULL(dst); - CHECK_NOTNULL(src); - CHECK_NOTNULL(starts); - - int block = 512; - int grid = DIVUP(width * height, 512); - - CHECK(mode == 0 || mode == 1 || mode == 2) - << "mode error in hl_sequence_avg_backward!"; - - KeSequenceAvgBackward<<>>( - dst, src, starts, height, width, mode); - CHECK_SYNC("hl_sequence_avg_backward failed"); -} diff --git a/paddle/cuda/src/hl_cuda_sparse.cu b/paddle/cuda/src/hl_cuda_sparse.cu deleted file mode 100644 index 432041fed5ab1ffc02dabcd4644fa70a6473fba1..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_cuda_sparse.cu +++ /dev/null @@ -1,1262 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_cuda.h" -#include "hl_cuda_sparse.cuh" -#include "hl_matrix_apply.cuh" -#include "hl_matrix_ops.cuh" -#include "hl_sparse.h" -#include "hl_sparse.ph" -#include "paddle/utils/Logging.h" - -DEFINE_MATRIX_UNARY_PARAMETER_OP(mul_scalar, ONE_PARAMETER, a = a * p); -DEFINE_MATRIX_UNARY_OP(Zero, a = 0); - -void hl_matrix_csr2dense(hl_sparse_matrix_s A_d, - real *C_d, - int dimM, - int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - CHECK(dimM > 0 && dimN > 0 && A_d->rows == dimM && A_d->cols == dimN); - CHECK(A_d->format == HL_SPARSE_CSR) << "matrix format error!"; - - if (A_d->nnz == 0) { - hl_gpu_apply_unary_op(unary::Zero(), C_d, dimM, dimN, dimN); - return; - } - - /* nnz != 0 */ - hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); - CHECK((A_d2->csr_val || A_d->type == HL_NO_VALUE) && A_d2->csr_row && - A_d2->csr_col) - << "parameter transa error!"; - - int blocksX = (dimN + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; - int blocksY = (dimM + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; - dim3 threads(CU_CSR2DENSE_THREAD_X, CU_CSR2DENSE_THREAD_X); - dim3 grid(blocksX, blocksY); - - if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsr2Dense<0><<>>( - A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, C_d, dimM, dimN); - } else if (A_d->type == HL_FLOAT_VALUE) { - KeSMatrixCsr2Dense<1><<>>( - A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, C_d, dimM, dimN); - } else { - } - CHECK_SYNC("hl_matrix_csr2dense failed"); -} - -void hl_matrix_csc2dense(hl_sparse_matrix_s A_d, - real *C_d, - int dimM, - int dimN) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(C_d); - CHECK(dimM > 0 && dimN > 0 && A_d->rows == dimM && A_d->cols == dimN); - CHECK(A_d->format == HL_SPARSE_CSC) << "matrix format error!"; - - if (A_d->nnz == 0) { - hl_gpu_apply_unary_op(unary::Zero(), C_d, dimM, dimN, dimN); - return; - } - - /* nnz != 0 */ - hl_csc_matrix A_d2 = (hl_csc_matrix)(A_d->matrix); - CHECK((A_d2->csc_val || A_d->type == HL_NO_VALUE) && A_d2->csc_row && - A_d2->csc_col) - << "parameter transa error!"; - - int blocksX = (dimN + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; - int blocksY = (dimM + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; - dim3 threads(CU_CSR2DENSE_THREAD_X, CU_CSR2DENSE_THREAD_X); - dim3 grid(blocksX, blocksY); - - if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsc2Dense<0><<>>( - A_d2->csc_val, A_d2->csc_row, A_d2->csc_col, C_d, dimM, dimN); - } else if (A_d->type == HL_FLOAT_VALUE) { - KeSMatrixCsc2Dense<1><<>>( - A_d2->csc_val, A_d2->csc_row, A_d2->csc_col, C_d, dimM, dimN); - } else { - } - CHECK_SYNC("hl_matrix_csc2dense failed"); -} - -void hl_malloc_sparse_matrix(hl_sparse_matrix_s *A_d, - hl_matrix_format_t format, - hl_matrix_value_t value_type, - int dimM, - int dimN, - int nnz) { - CHECK_NOTNULL(A_d); - CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) - << "sparse matrix format error!"; - CHECK(value_type == HL_FLOAT_VALUE || value_type == HL_NO_VALUE) - << "sparse matrix value type error!"; - /* avoid malloc 0 bytes */ - int nnz_s = (nnz == 0 ? 1 : nnz); - - if (format == HL_SPARSE_CSR) { - CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - - char *tmp = - (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); - CHECK_NOTNULL(tmp); - - hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); - csr->sparsity = -1.0; - - if (value_type == HL_NO_VALUE) { - csr->csr_val = NULL; - csr->nnz_s = nnz_s; - csr->row_s = dimM + 1; - csr->csr_row = (int *)hl_malloc_device((dimM + 1) * sizeof(int)); - csr->csr_col = (int *)hl_malloc_device((nnz_s) * sizeof(int)); - - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csr; - } else if (value_type == HL_FLOAT_VALUE) { - csr->nnz_s = nnz_s; - csr->row_s = dimM + 1; - csr->csr_val = (real *)hl_malloc_device((nnz_s) * sizeof(real)); - csr->csr_row = (int *)hl_malloc_device((dimM + 1) * sizeof(int)); - csr->csr_col = (int *)hl_malloc_device((nnz_s) * sizeof(int)); - - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csr; - } - } else if (format == HL_SPARSE_CSC) { - CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - - char *tmp = - (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); - CHECK_NOTNULL(tmp); - - hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); - csc->sparsity = -1.0f; - - if (value_type == HL_NO_VALUE) { - csc->csc_val = NULL; - csc->nnz_s = nnz_s; - csc->col_s = dimN + 1; - csc->csc_row = (int *)hl_malloc_device((nnz_s) * sizeof(int)); - csc->csc_col = (int *)hl_malloc_device((dimN + 1) * sizeof(int)); - - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csc; - } else if (value_type == HL_FLOAT_VALUE) { - csc->nnz_s = nnz_s; - csc->col_s = dimN + 1; - csc->csc_val = (real *)hl_malloc_device((nnz_s) * sizeof(real)); - csc->csc_row = (int *)hl_malloc_device((nnz_s) * sizeof(int)); - csc->csc_col = (int *)hl_malloc_device((dimN + 1) * sizeof(int)); - - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csc; - } - } - - (*A_d)->format = format; - (*A_d)->type = value_type; - (*A_d)->rows = dimM; - (*A_d)->cols = dimN; - (*A_d)->nnz = nnz; -} - -void hl_free_sparse_matrix(hl_sparse_matrix_s A_d) { - CHECK_NOTNULL(A_d); - CHECK(A_d->format == HL_SPARSE_CSR || A_d->format == HL_SPARSE_CSC) - << "sparse matrix format error!"; - - if (A_d->matrix == NULL) { - free(A_d); - return; - } - - if (A_d->format == HL_SPARSE_CSR) { - hl_csr_matrix csr = (hl_csr_matrix)A_d->matrix; - if (csr->csr_val != NULL) { - hl_free_mem_device(csr->csr_val); - csr->csr_val = NULL; - } - - if (csr->csr_row != NULL) { - hl_free_mem_device(csr->csr_row); - csr->csr_row = NULL; - } - - if (csr->csr_col != NULL) { - hl_free_mem_device(csr->csr_col); - csr->csr_col = NULL; - } - - A_d->matrix = NULL; - free(A_d); - } else if (A_d->format == HL_SPARSE_CSC) { - hl_csc_matrix csc = (hl_csc_matrix)A_d->matrix; - if (csc->csc_val != NULL) { - hl_free_mem_device(csc->csc_val); - csc->csc_val = NULL; - } - - if (csc->csc_row != NULL) { - hl_free_mem_device(csc->csc_row); - csc->csc_row = NULL; - } - - if (csc->csc_col != NULL) { - hl_free_mem_device(csc->csc_col); - csc->csc_col = NULL; - } - - A_d->matrix = NULL; - free(A_d); - } -} - -void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, - void *dest_d, - size_t size, - hl_matrix_format_t format, - hl_matrix_value_t value_type, - int dimM, - int dimN, - int nnz) { - CHECK_NOTNULL(A_d); - CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) - << "sparse matrix format error!"; - - if (format == HL_SPARSE_CSR) { - CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - - size_t size_ = (dimM + 1) * sizeof(int) + nnz * sizeof(int); - if (value_type != HL_NO_VALUE) { - size_ += nnz * sizeof(real); - } - CHECK_LE(size_, size) << "dest_d size(" << size - << ") too small, should bigger than(" << size_ - << ")!"; - - char *tmp = - (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); - CHECK_NOTNULL(tmp); - - hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); - - if (value_type == HL_NO_VALUE) { - csr->csr_val = NULL; - csr->csr_row = (int *)dest_d; - csr->csr_col = (int *)((char *)dest_d + (dimM + 1) * sizeof(int)); - } else { - csr->csr_val = (real *)dest_d; - csr->csr_row = (int *)((char *)dest_d + nnz * sizeof(real)); - csr->csr_col = (int *)((char *)dest_d + nnz * sizeof(real) + - (dimM + 1) * sizeof(int)); - } - csr->nnz_s = nnz; - csr->row_s = dimM + 1; - csr->sparsity = -1.0; - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csr; - } else if (format == HL_SPARSE_CSC) { - CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - - size_t size_ = (dimN + 1) * sizeof(int) + nnz * sizeof(int); - if (value_type != HL_NO_VALUE) { - size_ += nnz * sizeof(real); - } - CHECK_LE(size_, size) << "dest_d size(" << size - << ") too small, should bigger than(" << size_ - << ")!"; - - char *tmp = - (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); - CHECK_NOTNULL(tmp); - - hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); - if (value_type == HL_NO_VALUE) { - csc->csc_val = NULL; - csc->csc_col = (int *)dest_d; - csc->csc_row = (int *)((char *)dest_d + (dimN + 1) * sizeof(int)); - } else { - csc->csc_val = (real *)dest_d; - csc->csc_col = (int *)((char *)dest_d + nnz * sizeof(real)); - csc->csc_row = (int *)((char *)dest_d + nnz * sizeof(real) + - (dimN + 1) * sizeof(int)); - } - csc->nnz_s = nnz; - csc->col_s = dimN + 1; - csc->sparsity = -1.0f; - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csc; - } - - (*A_d)->format = format; - (*A_d)->type = value_type; - (*A_d)->rows = dimM; - (*A_d)->cols = dimN; - (*A_d)->nnz = nnz; -} - -void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, - real *value_d, - int *rows_d, - int *cols_d, - hl_matrix_format_t format, - hl_matrix_value_t value_type, - int dimM, - int dimN, - int nnz) { - CHECK_NOTNULL(A_d); - CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - - CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) - << "sparse matrix format error!"; - - if (format == HL_SPARSE_CSR) { - char *tmp = - (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); - CHECK_NOTNULL(tmp); - - hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); - csr->csr_row = rows_d; - csr->csr_col = cols_d; - csr->csr_val = value_d; - csr->nnz_s = nnz; - csr->row_s = dimM + 1; - csr->sparsity = -1.0; - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csr; - } else if (format == HL_SPARSE_CSC) { - char *tmp = - (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); - CHECK_NOTNULL(tmp); - - hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); - csc->csc_row = rows_d; - csc->csc_col = cols_d; - csc->csc_val = value_d; - csc->nnz_s = nnz; - csc->col_s = dimN + 1; - csc->sparsity = -1.0f; - *A_d = (hl_sparse_matrix_s)tmp; - (*A_d)->matrix = (hl_matrix_s)csc; - } - - (*A_d)->format = format; - (*A_d)->type = value_type; - (*A_d)->rows = dimM; - (*A_d)->cols = dimN; - (*A_d)->nnz = nnz; -} - -void hl_destruct_sparse_matrix(hl_sparse_matrix_s A_d) { - CHECK_NOTNULL(A_d); - free(A_d); -} - -void hl_memcpy_csr_matrix(hl_sparse_matrix_s csr_matrix, - real *csr_val, - int *csr_row, - int *csr_col, - hl_stream_t stream) { - CHECK_NOTNULL(csr_matrix); - CHECK_EQ(csr_matrix->format, HL_SPARSE_CSR) - << "csr_matrix is not csr format!"; - CHECK_NOTNULL(csr_matrix->matrix); - - hl_csr_matrix csr = (hl_csr_matrix)(csr_matrix->matrix); - CHECK_LE(csr_matrix->nnz, csr->nnz_s) << "copy size " << csr_matrix->nnz - << " is big than alloc size " - << csr->nnz_s; - - CHECK_LE((csr_matrix->rows + 1), csr->row_s) - << "copy size " << (csr_matrix->rows + 1) << " is big than alloc size " - << csr->row_s; - - CHECK(csr_matrix->type == HL_FLOAT_VALUE || csr_matrix->type == HL_NO_VALUE) - << "sparse matrix value type error!"; - - if (csr_matrix->type == HL_NO_VALUE) { - if (csr_row == NULL && csr_col == NULL) { - return; - } else if (csr_row != NULL && csr_col != NULL) { - hl_memcpy_async( - csr->csr_row, csr_row, (csr_matrix->rows + 1) * sizeof(int), stream); - - hl_memcpy_async( - csr->csr_col, csr_col, (csr_matrix->nnz) * sizeof(int), stream); - } else { - LOG(FATAL) << "parameter csr_row or csr_col is null pointer!"; - } - } else if (csr_matrix->type == HL_FLOAT_VALUE) { - if (csr_val == NULL && csr_row == NULL && csr_col == NULL) { - return; - } else if (csr_val != NULL && csr_row == NULL && csr_col == NULL) { - hl_memcpy_async( - csr->csr_val, csr_val, (csr_matrix->nnz) * sizeof(real), stream); - } else if (csr_val != NULL && csr_row != NULL && csr_col != NULL) { - hl_memcpy_async( - csr->csr_val, csr_val, (csr_matrix->nnz) * sizeof(real), stream); - hl_memcpy_async( - csr->csr_row, csr_row, (csr_matrix->rows + 1) * sizeof(int), stream); - hl_memcpy_async( - csr->csr_col, csr_col, (csr_matrix->nnz) * sizeof(int), stream); - } else { - LOG(FATAL) << "parameter csr_row or csr_col is null pointer!"; - } - } - - csr->sparsity = ((float)csr_matrix->nnz) / ((float)csr_matrix->rows) / - ((float)csr_matrix->cols); -} - -void hl_memcpy_csc_matrix(hl_sparse_matrix_s csc_matrix, - real *csc_val, - int *csc_row, - int *csc_col, - hl_stream_t stream) { - CHECK_NOTNULL(csc_matrix); - CHECK_EQ(csc_matrix->format, HL_SPARSE_CSC) - << "csc_matrix is not csc format error!"; - - hl_csc_matrix csc = (hl_csc_matrix)(csc_matrix->matrix); - CHECK_LE(csc_matrix->nnz, csc->nnz_s) << "copy size " << csc_matrix->nnz - << " is big than alloc size " - << csc->nnz_s; - - CHECK_LE((csc_matrix->cols + 1), csc->col_s) - << "copy size " << (csc_matrix->cols + 1) << " is big than alloc size " - << csc->col_s; - - CHECK(csc_matrix->type == HL_FLOAT_VALUE || csc_matrix->type == HL_NO_VALUE) - << "sparse matrix value type error!"; - - if (csc_matrix->type == HL_NO_VALUE) { - if (csc_row == NULL && csc_col == NULL) { - return; - } else if (csc_row != NULL && csc_col != NULL) { - hl_memcpy_async( - csc->csc_row, csc_row, (csc_matrix->nnz) * sizeof(int), stream); - hl_memcpy_async( - csc->csc_col, csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); - } else { - LOG(FATAL) << "parameter csc_row or csc_col is null pointer!"; - } - } else if (csc_matrix->type == HL_FLOAT_VALUE) { - if (csc_val == NULL && csc_row == NULL && csc_col == NULL) { - return; - } else if (csc_val != NULL && csc_row == NULL && csc_col == NULL) { - hl_memcpy_async( - csc->csc_val, csc_val, (csc_matrix->nnz) * sizeof(real), stream); - } else if (csc_val != NULL && csc_row != NULL && csc_col != NULL) { - hl_memcpy_async( - csc->csc_val, csc_val, (csc_matrix->nnz) * sizeof(real), stream); - hl_memcpy_async( - csc->csc_row, csc_row, (csc_matrix->nnz) * sizeof(int), stream); - hl_memcpy_async( - csc->csc_col, csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); - } else { - LOG(FATAL) << "parameter csc_row or csc_col is null pointer!"; - } - } - - csc->sparsity = ((float)csc_matrix->nnz) / ((float)csc_matrix->rows) / - ((float)csc_matrix->cols); -} - -void hl_memcpy_sparse_matrix(hl_sparse_matrix_s dst, - hl_sparse_matrix_s src, - hl_stream_t stream) { - CHECK(dst && src && dst->matrix && src->matrix) - << "parameter dst or src is null pointer!"; - CHECK_EQ(dst->format, src->format) << "sparse matrix format does not match!"; - CHECK(dst->type != HL_FLOAT_VALUE || src->type != HL_NO_VALUE) - << "src sparse matrix is no value, dst sparse matrix has value!"; - - if (dst->format == HL_SPARSE_CSR) { - dst->rows = src->rows; - dst->cols = src->cols; - dst->nnz = src->nnz; - hl_csr_matrix csr = (hl_csr_matrix)src->matrix; - hl_memcpy_csr_matrix(dst, csr->csr_val, csr->csr_row, csr->csr_col, stream); - } else if (dst->format == HL_SPARSE_CSC) { - dst->rows = src->rows; - dst->cols = src->cols; - dst->nnz = src->nnz; - hl_csc_matrix csc = (hl_csc_matrix)src->matrix; - hl_memcpy_csc_matrix(dst, csc->csc_val, csc->csc_row, csc->csc_col, stream); - } else { - LOG(FATAL) << "sparse matrix format error!"; - } -} - -/** - * Calculate beta * C, if beta is zero, C does not have to be a valid input. - */ -static void _beta_mul_c(real *c, int dimM, int dimN, real beta) { - if (beta == 0.0) { - hl_gpu_apply_unary_op(unary::Zero(), c, dimM, dimN, dimN); - } else { - if (beta != 1.0) { - hl_gpu_apply_unary_op(unary::mul_scalar(beta), c, dimM, dimN, dimN); - } - } - - return; -} - -void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, - hl_trans_op_t transa, - real *B_d, - hl_trans_op_t transb, - real *C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta) { - CHECK_EQ(transb, HPPL_OP_N); - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - CHECK(dimM > 0 && dimN > 0 && dimK > 0); - CHECK_EQ(A_d->format, HL_SPARSE_CSR) << "matrix format error!"; - - if ((HPPL_OP_N == transa && (A_d->rows != dimM || A_d->cols != dimK)) || - (HPPL_OP_T == transa && (A_d->rows != dimK || A_d->cols != dimM))) { - LOG(FATAL) << "parameter error!"; - } - - if (A_d->nnz == 0) { - _beta_mul_c(C_d, dimM, dimN, beta); - return; - } - - /* nnz != 0 */ - hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); - if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || - A_d2->csr_row == NULL || A_d2->csr_col == NULL) { - LOG(FATAL) << "parameter error!"; - } - - if (HPPL_OP_N == transa) { - int blocksX = (dimN + CU_CSRMM_BLOCK_N - 1) / CU_CSRMM_BLOCK_N; - int blocksY = (dimM + CU_CSRMM_THREAD_Y - 1) / CU_CSRMM_THREAD_Y; - dim3 threads(CU_CSRMM_THREAD_X, CU_CSRMM_THREAD_Y); - dim3 grid(blocksX, blocksY); - - /* sparsity pattern */ - // A_d->sparsity; - if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsrMulDense<0><<>>( - C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixCsrMulDense<1><<>>( - C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else if (HPPL_OP_T == transa) { - _beta_mul_c(C_d, dimM, dimN, beta); - - int blocksX = - (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / CU_CSC_MUL_DENSE_BLOCK_N; - int blocksY = - (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / CU_CSC_MUL_DENSE_BLOCK_K; - dim3 threads(CU_CSC_MUL_DENSE_THREAD_X, CU_CSC_MUL_DENSE_THREAD_Y); - dim3 grid(blocksX, blocksY); - if (A_d->type == HL_NO_VALUE) { - KeSMatrixCscMulDense<0><<>>( - C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixCscMulDense<1><<>>( - C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else { - LOG(FATAL) << "parameter transa error!"; - } - - CHECK_SYNC("hl_matrix_csr_mul_dense failed"); -} - -void hl_matrix_dense_mul_csc(real *A_d, - hl_trans_op_t transa, - hl_sparse_matrix_s B_d, - hl_trans_op_t transb, - real *C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta) { - CHECK_EQ(transa, HPPL_OP_N); - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - - if (dimM <= 0 || dimN <= 0 || dimK <= 0 || - ((transb == HPPL_OP_N) && (B_d->rows != dimK || B_d->cols != dimN)) || - ((transb == HPPL_OP_T) && (B_d->rows != dimN || B_d->cols != dimK))) { - LOG(FATAL) << "parameter dims error!"; - } - - CHECK_EQ(B_d->format, HL_SPARSE_CSC) << "matrix format error!"; - - if (B_d->nnz == 0) { - _beta_mul_c(C_d, dimM, dimN, beta); - return; - } - - /* nnz != 0 */ - hl_csc_matrix B_d2 = (hl_csc_matrix)(B_d->matrix); - if ((B_d2->csc_val == NULL && B_d->type != HL_NO_VALUE) || - B_d2->csc_row == NULL || B_d2->csc_col == NULL) { - LOG(FATAL) << "parameter B is null!"; - } - - if (transb == HPPL_OP_N) { - int blocksX = (dimM + CU_CSCMM_BLOCK_M_BEST - 1) / CU_CSCMM_BLOCK_M_BEST; - int blocksY = (dimN + CU_CSCMM_BLOCK_N_BEST - 1) / CU_CSCMM_BLOCK_N_BEST; - dim3 threads(CU_CSCMM_THREAD_X_BEST, CU_CSCMM_THREAD_Y_BEST); - dim3 grid(blocksX, blocksY); - - if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsc<0><<>>( - C_d, - A_d, - B_d2->csc_val, - B_d2->csc_row, - B_d2->csc_col, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixDenseMulCsc<1><<>>( - C_d, - A_d, - B_d2->csc_val, - B_d2->csc_row, - B_d2->csc_col, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else if (transb == HPPL_OP_T) { - _beta_mul_c(C_d, dimM, dimN, beta); - int blocksX = 1 + (dimK - 1) / CU_DM_CSR_THREAD_X; - int blocksY = 1 + (dimM - 1) / CU_DM_CSR_BLOCK_M; - dim3 threads(CU_DM_CSR_THREAD_X, CU_DM_CSR_THREAD_Y); - dim3 grid(blocksX, blocksY); - if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsr<0><<>>( - C_d, - A_d, - B_d2->csc_val, - B_d2->csc_col, - B_d2->csc_row, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixDenseMulCsr<1><<>>( - C_d, - A_d, - B_d2->csc_val, - B_d2->csc_col, - B_d2->csc_row, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else { - LOG(FATAL) << "parameter transb error!"; - } - - CHECK_SYNC("hl_matrix_dense_mul_csc failed"); -} - -void hl_matrix_dense_mul_csr(real *A_d, - hl_trans_op_t transa, - hl_sparse_matrix_s B_d, - hl_trans_op_t transb, - real *C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta) { - CHECK_EQ(transa, HPPL_OP_N); - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - - if (dimM <= 0 || dimN <= 0 || dimK <= 0 || - (transb == HPPL_OP_N && (B_d->rows != dimK || B_d->cols != dimN)) || - (transb == HPPL_OP_T && (B_d->rows != dimN || B_d->cols != dimK))) { - LOG(FATAL) << "parameter dims error!"; - } - - CHECK_EQ(B_d->format, HL_SPARSE_CSR) << "matrix format error!"; - - if (B_d->nnz == 0) { - _beta_mul_c(C_d, dimM, dimN, beta); - return; - } - - /* nnz != 0 */ - hl_csr_matrix B_d2 = (hl_csr_matrix)(B_d->matrix); - if ((B_d2->csr_val == NULL && B_d->type != HL_NO_VALUE) || - B_d2->csr_row == NULL || B_d2->csr_col == NULL) { - LOG(FATAL) << "parameter transa error!"; - } - - if (transb == HPPL_OP_N) { - _beta_mul_c(C_d, dimM, dimN, beta); - int blocksX = 1 + (dimK - 1) / CU_DM_CSR_THREAD_X; - int blocksY = 1 + (dimM - 1) / CU_DM_CSR_BLOCK_M; - dim3 threads(CU_DM_CSR_THREAD_X, CU_DM_CSR_THREAD_Y); - dim3 grid(blocksX, blocksY); - if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsr<0><<>>( - C_d, - A_d, - B_d2->csr_val, - B_d2->csr_row, - B_d2->csr_col, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixDenseMulCsr<1><<>>( - C_d, - A_d, - B_d2->csr_val, - B_d2->csr_row, - B_d2->csr_col, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else if (transb == HPPL_OP_T) { - int blocksX = (dimM + CU_CSCMM_BLOCK_M_BEST - 1) / CU_CSCMM_BLOCK_M_BEST; - int blocksY = (dimN + CU_CSCMM_BLOCK_N_BEST - 1) / CU_CSCMM_BLOCK_N_BEST; - dim3 threads(CU_CSCMM_THREAD_X_BEST, CU_CSCMM_THREAD_Y_BEST); - dim3 grid(blocksX, blocksY); - if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsc<0><<>>( - C_d, - A_d, - B_d2->csr_val, - B_d2->csr_col, - B_d2->csr_row, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixDenseMulCsc<1><<>>( - C_d, - A_d, - B_d2->csr_val, - B_d2->csr_col, - B_d2->csr_row, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else { - LOG(FATAL) << "parameter transb error!"; - } - - CHECK_SYNC("hl_matrix_dense_mul_csr failed"); -} - -void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, - hl_trans_op_t transa, - real *B_d, - hl_trans_op_t transb, - real *C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta) { - CHECK_EQ(transb, HPPL_OP_N); - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - CHECK(dimM > 0 && dimN > 0 && dimK > 0) << "parameter error!"; - CHECK_EQ(A_d->format, HL_SPARSE_CSC) << "matrix format error!"; - - if ((HPPL_OP_N == transa && (A_d->rows != dimM || A_d->cols != dimK)) || - (HPPL_OP_T == transa && (A_d->rows != dimK || A_d->cols != dimM))) { - LOG(FATAL) << "parameter error!"; - } - - if (A_d->nnz == 0) { - _beta_mul_c(C_d, dimM, dimN, beta); - return; - } - - /* nnz != 0 */ - hl_csc_matrix A_d2 = (hl_csc_matrix)(A_d->matrix); - if ((A_d2->csc_val == NULL && A_d->type != HL_NO_VALUE) || - A_d2->csc_row == NULL || A_d2->csc_col == NULL) { - LOG(FATAL) << "parameter error!"; - } - - if (HPPL_OP_N == transa) { - _beta_mul_c(C_d, dimM, dimN, beta); - - int blocksX = - (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / CU_CSC_MUL_DENSE_BLOCK_N; - int blocksY = - (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / CU_CSC_MUL_DENSE_BLOCK_K; - dim3 threads(CU_CSC_MUL_DENSE_THREAD_X, CU_CSC_MUL_DENSE_THREAD_Y); - dim3 grid(blocksX, blocksY); - if (A_d->type == HL_NO_VALUE) { - KeSMatrixCscMulDense<0><<>>( - C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixCscMulDense<1><<>>( - C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else if (HPPL_OP_T == transa) { - int blocksX = (dimN + CU_CSRMM_BLOCK_N - 1) / CU_CSRMM_BLOCK_N; - int blocksY = (dimM + CU_CSRMM_THREAD_Y - 1) / CU_CSRMM_THREAD_Y; - dim3 threads(CU_CSRMM_THREAD_X, CU_CSRMM_THREAD_Y); - dim3 grid(blocksX, blocksY); - - /* sparsity pattern */ - // A_d->sparsity; - if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsrMulDense<0><<>>( - C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } else { - KeSMatrixCsrMulDense<1><<>>( - C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); - } - } else { - LOG(FATAL) << "parameter transa error!"; - } - - CHECK_SYNC("hl_matrix_csc_mul_dense failed"); -} - -void hl_sparse_matrix_mul(real *A_d, - hl_trans_op_t transa, - real *B_d, - hl_trans_op_t transb, - hl_sparse_matrix_s C_d, - int dimM, - int dimN, - int dimK, - real alpha, - real beta) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - CHECK_NOTNULL(C_d); - CHECK(dimM > 0 && dimN > 0 && dimK > 0) << "parameter error!"; - CHECK_NE(C_d->type, HL_NO_VALUE) << "C value type error!"; - - if (C_d->nnz == 0) return; - - if (C_d->format == HL_SPARSE_CSC) { - hl_csc_matrix C_d2 = (hl_csc_matrix)(C_d->matrix); - if (C_d2->csc_val == NULL || C_d2->csc_row == NULL || - C_d2->csc_col == NULL) { - LOG(FATAL) << "parameter error!"; - } - - if (beta != 1.0) { - hl_gpu_apply_unary_op( - unary::mul_scalar(beta), C_d2->csc_val, 1, C_d->nnz, C_d->nnz); - } - - int blocksX = dimN; - int blocksY = 1; - dim3 threads(CU_CSCMM_DMD2CSC_THREAD_X, 1); - dim3 grid(blocksX, blocksY); - bool transA = transa == HPPL_OP_T ? 1 : 0; - bool transB = transb == HPPL_OP_T ? 1 : 0; - KeSMatrixDenseMulDense2CSC<<>>( - C_d2->csc_val, - C_d2->csc_row, - C_d2->csc_col, - A_d, - B_d, - transA, - transB, - dimM, - dimN, - dimK, - alpha, - beta); - CHECK_SYNC("hl_sparse_matrix_mul failed"); - } else { - hl_csr_matrix C_d2 = (hl_csr_matrix)(C_d->matrix); - if ((C_d2->csr_val == NULL && C_d->type != HL_NO_VALUE) || - C_d2->csr_row == NULL || C_d2->csr_col == NULL) { - LOG(FATAL) << "parameter error!"; - } - - if (beta != 1.0) { - hl_gpu_apply_unary_op( - unary::mul_scalar(beta), C_d2->csr_val, 1, C_d->nnz, C_d->nnz); - } - - bool transA = transa == HPPL_OP_T ? 1 : 0; - bool transB = transb == HPPL_OP_T ? 1 : 0; - if (!transB) { - int blocksX = dimM; - int blocksY = 1; - dim3 threads(CU_CSCMM_DMD2CSR_THREAD_X, 1); - dim3 grid(blocksX, blocksY); - - KeSMatrixDenseMulDense2CSR<<>>( - C_d2->csr_val, - C_d2->csr_row, - C_d2->csr_col, - A_d, - B_d, - transA, - transB, - dimM, - dimN, - dimK, - alpha, - beta); - CHECK_SYNC("hl_sparse_matrix_mul failed"); - } else { - CHECK(!transA) << "Not supported A is trans and B is not trans!"; - - dim3 block(CU_BLOCK_SIZE, 1); - int avgNnzPerRow = C_d->nnz / dimM; - avgNnzPerRow = avgNnzPerRow > 0 ? avgNnzPerRow : 1; - int gridx = DIVUP(avgNnzPerRow, CU_BLOCK_SIZE); - dim3 grid(gridx, dimM); - KeSMatrixDenseMulDenseTrans2CSR<<>>( - C_d2->csr_val, - C_d2->csr_row, - C_d2->csr_col, - A_d, - B_d, - transA, - transB, - dimM, - dimN, - dimK, - alpha, - beta); - CHECK_SYNC("hl_sparse_matrix_mul failed"); - } - } -} - -void hl_memcpy_from_csc_matrix(real *csc_val, - size_t val_size, - int *csc_row, - size_t row_size, - int *csc_col, - size_t col_size, - hl_sparse_matrix_s csc_matrix, - hl_stream_t stream) { - CHECK_NOTNULL(csc_matrix); - CHECK_NOTNULL(csc_row); - CHECK_NOTNULL(csc_col); - - CHECK_EQ(csc_matrix->format, HL_SPARSE_CSC) - << "csc_matrix is not csc format error!"; - - if (csc_matrix->nnz > row_size || - csc_matrix->cols + 1 > static_cast(col_size)) { - LOG(FATAL) << "size not match!"; - } - - hl_csc_matrix csc = (hl_csc_matrix)(csc_matrix->matrix); - hl_memcpy_async((void *)csc_row, - (void *)csc->csc_row, - (csc_matrix->nnz) * sizeof(int), - stream); - hl_memcpy_async((void *)csc_col, - (void *)csc->csc_col, - (csc_matrix->cols + 1) * sizeof(int), - stream); - if (csc_matrix->type == HL_FLOAT_VALUE) { - if (csc_val != NULL) { - CHECK_LE(csc_matrix->nnz, val_size) << "size not match!"; - hl_memcpy_async((void *)csc_val, - (void *)csc->csc_val, - (csc_matrix->nnz) * sizeof(real), - stream); - } else { - LOG(FATAL) << "parameter csr_val is null pointer!"; - } - } -} - -void hl_memcpy_from_csr_matrix(real *csr_val, - size_t val_size, - int *csr_row, - size_t row_size, - int *csr_col, - size_t col_size, - hl_sparse_matrix_s csr_matrix, - hl_stream_t stream) { - CHECK_NOTNULL(csr_matrix); - CHECK_NOTNULL(csr_row); - CHECK_NOTNULL(csr_col); - CHECK_EQ(csr_matrix->format, HL_SPARSE_CSR) - << "csr_matrix is not csr format error!"; - - if (csr_matrix->nnz > col_size || - csr_matrix->rows + 1 > static_cast(row_size)) { - LOG(FATAL) << "size not match!"; - } - - hl_csr_matrix csr = (hl_csr_matrix)(csr_matrix->matrix); - hl_memcpy_async((void *)csr_row, - (void *)csr->csr_row, - (csr_matrix->rows + 1) * sizeof(int), - stream); - hl_memcpy_async((void *)csr_col, - (void *)csr->csr_col, - (csr_matrix->nnz) * sizeof(int), - stream); - if (csr_matrix->type == HL_FLOAT_VALUE) { - if (csr_val != NULL) { - CHECK_LE(csr_matrix->nnz, val_size) << "size not match!"; - hl_memcpy_async((void *)csr_val, - (void *)csr->csr_val, - (csr_matrix->nnz) * sizeof(real), - stream); - } else { - LOG(FATAL) << "parameter csr_val is null pointer!"; - } - } -} - -void hl_sparse_matrix_column_sum( - real *A_d, hl_sparse_matrix_s B_d, int dimM, int dimN, real scale) { - if (B_d->format == HL_SPARSE_CSR) { - hl_matrix_csr_column_sum(A_d, B_d, dimM, dimN, scale); - } else { - LOG(FATAL) << "Not support CSC format error!"; - } -} - -void hl_matrix_csr_column_sum( - real *A_d, hl_sparse_matrix_s B_d, int dimM, int dimN, real scale) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - - if (dimM <= 0 || dimN <= 0 || (B_d->rows != dimM || B_d->cols != dimN)) { - LOG(FATAL) << "parameter dims error!"; - } - - hl_csr_matrix B_d2 = (hl_csr_matrix)(B_d->matrix); - if ((B_d2->csr_val == NULL && B_d->type != HL_NO_VALUE) || - B_d2->csr_row == NULL || B_d2->csr_col == NULL) { - LOG(FATAL) << "parameter B is null!"; - } - - if (B_d->nnz == 0) return; - - int nnz = B_d->nnz; - int block = 512; - int grid = DIVUP(nnz, 512); - KeSMatrixCsrColumnSum<<>>( - A_d, B_d2->csr_val, B_d2->csr_col, nnz); - - CHECK_SYNC("hl_matrix_csr_column_sum failed"); -} - -void hl_sparse_matrix_add_bias(hl_sparse_matrix_s A_d, real *B_d, real scale) { - if (A_d->format == HL_SPARSE_CSR) { - hl_matrix_csr_add_bias(A_d, B_d, scale); - } else { - LOG(FATAL) << "Not support CSC format error!"; - } -} - -void hl_matrix_csr_add_bias(hl_sparse_matrix_s A_d, real *B_d, real scale) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - - hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); - if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || - A_d2->csr_row == NULL || A_d2->csr_col == NULL) { - LOG(FATAL) << "parameter A_d is null!"; - } - - if (A_d->nnz == 0) return; - - int nnz = A_d->nnz; - int block = 512; - int grid = DIVUP(nnz, 512); - KeSMatrixCsrAddBias<<>>( - A_d2->csr_val, A_d2->csr_col, B_d, scale, nnz); - - CHECK_SYNC("hl_sparse_matrix_add_bias failed"); -} - -void hl_sparse_matrix_add_dense(hl_sparse_matrix_s A_d, - real *B_d, - int dimM, - int dimN, - real alpha, - real beta) { - if (A_d->format == HL_SPARSE_CSR) { - hl_matrix_csr_add_dense(A_d, B_d, dimM, dimN, alpha, beta); - } else { - LOG(FATAL) << "Not support CSC format error!"; - } -} - -void hl_matrix_csr_add_dense(hl_sparse_matrix_s A_d, - real *B_d, - int dimM, - int dimN, - real alpha, - real beta) { - CHECK_NOTNULL(A_d); - CHECK_NOTNULL(B_d); - - if (dimM <= 0 || dimN <= 0 || A_d->rows != dimM || A_d->cols != dimN) { - LOG(FATAL) << "parameter dim error!"; - } - - hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); - if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || - A_d2->csr_row == NULL || A_d2->csr_col == NULL) { - LOG(FATAL) << "parameter A_d is null!"; - } - - if (A_d->nnz == 0) return; - - int gridX = DIVUP((A_d->nnz / dimM), 512); - gridX = gridX > 0 ? gridX : 1; - dim3 block(512, 1); - dim3 grid(gridX, dimM); - KeSMatrixCsrAddDense<<>>(A_d2->csr_val, - A_d2->csr_row, - A_d2->csr_col, - B_d, - alpha, - beta, - dimM, - dimN); - - CHECK_SYNC("hl_sparse_matrix_add_dense failed"); -} - -int *hl_sparse_matrix_get_rows(hl_sparse_matrix_s sMat) { - __sparse_get_return__(sMat, row); -} - -int *hl_sparse_matrix_get_cols(hl_sparse_matrix_s sMat) { - __sparse_get_return__(sMat, col); -} - -real *hl_sparse_matrix_get_value(hl_sparse_matrix_s sMat) { - __sparse_get_return__(sMat, val); -} diff --git a/paddle/cuda/src/hl_table_apply.cu b/paddle/cuda/src/hl_table_apply.cu deleted file mode 100644 index efa4bef02ba5f5fe9ae449b44bbdc844e5745307..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_table_apply.cu +++ /dev/null @@ -1,124 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_base.h" -#include "hl_cuda.h" -#include "hl_device_functions.cuh" -#include "paddle/utils/Logging.h" - -template -__global__ void KeMatrixAddRows(real* output, - int ldo, - real* table, - int ldt, - int* ids, - int numSamples, - int tableSize, - int dim) { - int idx = threadIdx.x; - int idy = blockIdx.x + threadIdx.y * gridDimX; - - while (idy < numSamples) { - int tableId = ids[idy]; - if ((0 <= tableId) && (tableId < tableSize)) { - real* out = output + idy * ldo; - real* tab = table + tableId * ldt; - for (int i = idx; i < dim; i += blockDimX) { - if (AddRow) { - paddle::paddleAtomicAdd(&tab[i], out[i]); - } else { - out[i] += tab[i]; - } - } - } - idy += blockDimY * gridDimX; - } -} - -void hl_matrix_select_rows(real* output, - int ldo, - real* table, - int ldt, - int* ids, - int numSamples, - int tableSize, - int dim) { - CHECK_NOTNULL(output); - CHECK_NOTNULL(table); - CHECK_NOTNULL(ids); - - dim3 threads(128, 8); - dim3 grid(8, 1); - KeMatrixAddRows<128, 8, 8, 0><<>>( - output, ldo, table, ldt, ids, numSamples, tableSize, dim); - - CHECK_SYNC("hl_matrix_select_rows failed"); -} - -void hl_matrix_add_to_rows(real* table, - int ldt, - real* input, - int ldi, - int* ids, - int numSamples, - int tableSize, - int dim) { - CHECK_NOTNULL(input); - CHECK_NOTNULL(table); - CHECK_NOTNULL(ids); - - dim3 threads(128, 8); - dim3 grid(8, 1); - KeMatrixAddRows<128, 8, 8, 1><<>>( - input, ldi, table, ldt, ids, numSamples, tableSize, dim); - - CHECK_SYNC("hl_matrix_add_to_rows failed"); -} - -template -__global__ void KeVectorSelect( - T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) { - int idx = threadIdx.x + blockDimX * blockIdx.x; - while (idx < sizei) { - int index = ids[idx]; - // check(index < sizes); - dst[idx] = src[index]; - idx += blockDimX * gridDimX; - } -} - -template -void hl_vector_select_from( - T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) { - CHECK_NOTNULL(dst); - CHECK_NOTNULL(src); - CHECK_NOTNULL(ids); - CHECK_EQ(sized, sizei); - - dim3 threads(512, 1); - dim3 grid(8, 1); - KeVectorSelect<<>>( - dst, sized, src, sizes, ids, sizei); - - CHECK_SYNC("hl_vector_select_from failed"); -} - -template void hl_vector_select_from(real* dst, - int sized, - const real* src, - int sizes, - const int* ids, - int sizei); -template void hl_vector_select_from( - int* dst, int sized, const int* src, int sizes, const int* ids, int sizei); diff --git a/paddle/cuda/src/hl_top_k.cu b/paddle/cuda/src/hl_top_k.cu deleted file mode 100644 index b17290557c4f635a963d88525409b9373b057a4b..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_top_k.cu +++ /dev/null @@ -1,481 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/cuda/include/hl_base.h" -#include "paddle/cuda/include/hl_sparse.ph" -#include "paddle/cuda/include/hl_top_k.h" -#include "paddle/utils/Logging.h" - -// using namespace hppl; - -struct Pair { - __device__ __forceinline__ Pair() {} - - __device__ __forceinline__ Pair(real value, int id) : v_(value), id_(id) {} - - __device__ __forceinline__ void set(real value, int id) { - v_ = value; - id_ = id; - } - - __device__ __forceinline__ void operator=(const Pair& in) { - v_ = in.v_; - id_ = in.id_; - } - - __device__ __forceinline__ bool operator<(const real value) const { - return (v_ < value); - } - - __device__ __forceinline__ bool operator<(const Pair& in) const { - return (v_ < in.v_) || ((v_ == in.v_) && (id_ > in.id_)); - } - - __device__ __forceinline__ bool operator>(const Pair& in) const { - return (v_ > in.v_) || ((v_ == in.v_) && (id_ < in.id_)); - } - - real v_; - int id_; -}; - -__device__ __forceinline__ void addTo(Pair topK[], - const Pair& p, - int beamSize) { - for (int k = beamSize - 2; k >= 0; k--) { - if (topK[k] < p) { - topK[k + 1] = topK[k]; - } else { - topK[k + 1] = p; - return; - } - } - topK[0] = p; -} - -template -__device__ __forceinline__ void addTo(Pair topK[], const Pair& p) { - for (int k = beamSize - 2; k >= 0; k--) { - if (topK[k] < p) { - topK[k + 1] = topK[k]; - } else { - topK[k + 1] = p; - return; - } - } - topK[0] = p; -} - -template -__device__ __forceinline__ void getTopK( - Pair topK[], real* src, int idx, int dim, int beamSize) { - while (idx < dim) { - if (topK[beamSize - 1] < src[idx]) { - Pair tmp(src[idx], idx); - addTo(topK, tmp, beamSize); - } - idx += blockSize; - } -} - -template -__device__ __forceinline__ void getTopK( - Pair topK[], real* src, int idx, int dim, const Pair& max, int beamSize) { - while (idx < dim) { - if (topK[beamSize - 1] < src[idx]) { - Pair tmp(src[idx], idx); - if (tmp < max) { - addTo(topK, tmp, beamSize); - } - } - idx += blockSize; - } -} - -template -__device__ __forceinline__ void getTopK( - Pair topK[], real* val, int* col, int idx, int dim, int beamSize) { - while (idx < dim) { - if (topK[beamSize - 1] < val[idx]) { - Pair tmp(val[idx], col[idx]); - addTo(topK, tmp, beamSize); - } - idx += blockSize; - } -} - -template -__device__ __forceinline__ void getTopK(Pair topK[], - real* val, - int* col, - int idx, - int dim, - const Pair& max, - int beamSize) { - while (idx < dim) { - if (topK[beamSize - 1] < val[idx]) { - Pair tmp(val[idx], col[idx]); - if (tmp < max) { - addTo(topK, tmp, beamSize); - } - } - idx += blockSize; - } -} - -template -__device__ __forceinline__ void threadGetTopK(Pair topK[], - int& beam, - int beamSize, - real* src, - bool& firstStep, - bool& isEmpty, - Pair& max, - int dim, - const int tid) { - if (beam > 0) { - int length = beam < beamSize ? beam : beamSize; - if (firstStep) { - firstStep = false; - getTopK(topK, src, tid, dim, length); - } else { - for (int k = 0; k < maxLength; k++) { - if (k < maxLength - beam) { - topK[k] = topK[k + beam]; - } else { - topK[k].set(-HL_FLOAT_MAX, -1); - } - } - if (!isEmpty) { - getTopK(topK + maxLength - beam, src, tid, dim, max, length); - } - } - - max = topK[maxLength - 1]; - if (max.id_ == -1) isEmpty = true; - beam = 0; - } -} - -template -__device__ __forceinline__ void threadGetTopK(Pair topK[], - int& beam, - int beamSize, - real* val, - int* col, - bool& firstStep, - bool& isEmpty, - Pair& max, - int dim, - const int tid) { - if (beam > 0) { - int length = beam < beamSize ? beam : beamSize; - if (firstStep) { - firstStep = false; - getTopK(topK, val, col, tid, dim, length); - } else { - for (int k = 0; k < maxLength; k++) { - if (k < maxLength - beam) { - topK[k] = topK[k + beam]; - } else { - topK[k].set(-HL_FLOAT_MAX, -1); - } - } - if (!isEmpty) { - getTopK( - topK + maxLength - beam, val, col, tid, dim, max, length); - } - } - - max = topK[maxLength - 1]; - if (max.id_ == -1) isEmpty = true; - beam = 0; - } -} - -template -__device__ __forceinline__ void blockReduce(Pair* shTopK, - int* maxId, - Pair topK[], - real** topVal, - int** topIds, - int& beam, - int& beamSize, - const int tid, - const int warp) { - while (true) { - __syncthreads(); - if (tid < blockSize / 2) { - if (shTopK[tid] < shTopK[tid + blockSize / 2]) { - maxId[tid] = tid + blockSize / 2; - } else { - maxId[tid] = tid; - } - } - __syncthreads(); - for (int stride = blockSize / 4; stride > 0; stride = stride / 2) { - if (tid < stride) { - if (shTopK[maxId[tid]] < shTopK[maxId[tid + stride]]) { - maxId[tid] = maxId[tid + stride]; - } - } - __syncthreads(); - } - __syncthreads(); - - if (tid == 0) { - **topVal = shTopK[maxId[0]].v_; - **topIds = shTopK[maxId[0]].id_; - (*topVal)++; - (*topIds)++; - } - if (tid == maxId[0]) beam++; - if (--beamSize == 0) break; - __syncthreads(); - - // NOTE(zcd): temporary solution - unsigned mask = 0u; - CREATE_SHFL_MASK(mask, true); - - if (tid == maxId[0]) { - if (beam < maxLength) { - shTopK[tid] = topK[beam]; - } - } - if (maxId[0] / 32 == warp) { - if (__shfl_sync(mask, beam, (maxId[0]) % 32, 32) == maxLength) break; - } - } -} - -/** - * Each block compute one sample. - * In a block: - * 1. every thread get top maxLength value; - * 2. merge to shTopK, block reduce and get max value; - * 3. go to the second setp, until one thread's topK value is null; - * 4. go to the first setp, until get the topK value. - */ -template -__global__ void KeMatrixTopK(real* topVal, - int ldv, - int* topIds, - real* src, - int lds, - int dim, - int beamSize) { - __shared__ Pair shTopK[blockSize]; - __shared__ int maxId[blockSize / 2]; - const int tid = threadIdx.x; - const int warp = threadIdx.x / 32; - src += blockIdx.x * lds; - topVal += blockIdx.x * ldv; - topIds += blockIdx.x * beamSize; - - Pair topK[maxLength]; // NOLINT - int beam = maxLength; - Pair max; - bool isEmpty = false; - bool firstStep = true; - - for (int k = 0; k < maxLength; k++) { - topK[k].set(-HL_FLOAT_MAX, -1); - } - while (beamSize) { - threadGetTopK( - topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); - - shTopK[tid] = topK[0]; - blockReduce( - shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); - } -} - -template -__global__ void KeSMatrixTopK(real* topVal, - int ldv, - int* topIds, - real* val, - int* row, - int* col, - int beamSize) { - __shared__ Pair shTopK[blockSize]; - __shared__ int maxId[blockSize / 2]; - const int tid = threadIdx.x; - const int warp = threadIdx.x / 32; - topVal += blockIdx.x * ldv; - topIds += blockIdx.x * beamSize; - - Pair topK[maxLength]; // NOLINT - int beam = maxLength; - Pair max; - bool isEmpty = false; - bool firstStep = true; - - int start = row[blockIdx.x]; - int end = row[blockIdx.x + 1]; - int dim = end - start; - val += start; - col += start; - - if (beamSize > dim) { - // if the number of values to sort are less than the output size, - // use -1 to indicate the end of valid sorted values. - if (tid == 0) { - topIds[dim] = -1; - } - - beamSize = dim; - } - - for (int k = 0; k < maxLength; k++) { - topK[k].set(-HL_FLOAT_MAX, -1); - } - while (beamSize) { - threadGetTopK( - topK, beam, beamSize, val, col, firstStep, isEmpty, max, dim, tid); - - shTopK[tid] = topK[0]; - blockReduce( - shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); - } -} - -void hl_matrix_top_k(real* topVal, - int ldv, - int* topIds, - real* src, - int lds, - int dim, - int beamSize, - int numSamples) { - CHECK_NOTNULL(topVal); - CHECK_NOTNULL(topIds); - CHECK_NOTNULL(src); - - if (beamSize > dim) beamSize = dim; - - dim3 threads(256, 1); - dim3 grid(numSamples, 1); - KeMatrixTopK<5, 256><<>>( - topVal, ldv, topIds, src, lds, dim, beamSize); - - CHECK_SYNC("hl_matrix_top_k failed"); -} - -void hl_sparse_matrix_top_k(real* topVal, - int ldv, - int* topIds, - hl_sparse_matrix_s src, - int beamSize, - int numSamples) { - CHECK_NOTNULL(topVal); - CHECK_NOTNULL(topIds); - CHECK_NOTNULL(src); - CHECK_EQ(src->format, HL_SPARSE_CSR) << "sparse matrix format error!"; - - hl_csr_matrix csr = (hl_csr_matrix)src->matrix; - if (csr->csr_val == NULL || csr->csr_row == NULL || csr->csr_col == NULL) { - LOG(FATAL) << "parameter src is null!"; - } - - dim3 threads(256, 1); - dim3 grid(numSamples, 1); - KeSMatrixTopK<5, 256><<>>( - topVal, ldv, topIds, csr->csr_val, csr->csr_row, csr->csr_col, beamSize); - - CHECK_SYNC("hl_sparse_matrix_top_k failed"); -} - -/** - * Each block compute one sample. - * In a block: - * 1. every thread get top maxLength value; - * 2. merge to shTopK, block reduce and get max value; - * 3. go to the second setp, until one thread's topK value is null; - * 4. go to the first setp, until get the topK value. - */ -template -__global__ void KeMatrixTopKClassificationError(real* topVal, - int ldv, - int* topIds, - real* src, - int lds, - int dim, - int beamSize, - int* label, - real* recResult) { - __shared__ Pair shTopK[blockSize]; - __shared__ int maxId[blockSize / 2]; - const int tid = threadIdx.x; - const int warp = threadIdx.x / 32; - src += blockIdx.x * lds; - topVal += blockIdx.x * ldv; - topIds += blockIdx.x * beamSize; - - Pair topK[maxLength]; // NOLINT - int beam = maxLength; - Pair max; - bool isEmpty = false; - bool firstStep = true; - int topkSize = beamSize; - - for (int k = 0; k < maxLength; k++) { - topK[k].set(-HL_FLOAT_MAX, -1); - } - - while (beamSize) { - threadGetTopK( - topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); - - shTopK[tid] = topK[0]; - blockReduce( - shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); - } - - __syncthreads(); - if (tid == 0) { - for (int i = 0; i < topkSize; i++) { - if (*--topIds == label[blockIdx.x]) { - recResult[blockIdx.x] = 0; - break; - } - recResult[blockIdx.x] = 1.0f; - } - } -} - -void hl_matrix_classification_error(real* topVal, - int ldv, - int* topIds, - real* src, - int lds, - int dim, - int topkSize, - int numSamples, - int* label, - real* recResult) { - CHECK_NOTNULL(topVal); - CHECK_NOTNULL(topIds); - CHECK_NOTNULL(src); - - if (topkSize > dim) topkSize = dim; - - dim3 threads(256, 1); - dim3 grid(numSamples, 1); - KeMatrixTopKClassificationError<5, 256><<>>( - topVal, ldv, topIds, src, lds, dim, topkSize, label, recResult); - - CHECK_SYNC("hl_matrix_top_k classification error failed"); -} diff --git a/paddle/cuda/src/hl_warpctc_wrap.cc b/paddle/cuda/src/hl_warpctc_wrap.cc deleted file mode 100644 index 5111bceaff224f2467fe1b6c92daed03414dd12e..0000000000000000000000000000000000000000 --- a/paddle/cuda/src/hl_warpctc_wrap.cc +++ /dev/null @@ -1,151 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "hl_warpctc_wrap.h" -#include -#include "paddle/utils/DynamicLoader.h" -#include "paddle/utils/Logging.h" - -namespace dynload { - -std::once_flag warpctc_dso_flag; -void* warpctc_dso_handle = nullptr; - -/** - * The following macro definition can generate structs - * (for each function) to dynamic load warpctc routine - * via operator overloading. When PADDLE_USE_DSO is - * false, you need to add the path of libwarp-ctc.so to - * the linked-libs of paddle or to LD_PRELOAD. - */ -#define DYNAMIC_LOAD_WARPCTC_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - auto operator()(Args... args) -> decltype(__name(args...)) { \ - using warpctcFunc = decltype(__name(args...)) (*)(Args...); \ - std::call_once( \ - warpctc_dso_flag, GetWarpCTCDsoHandle, &warpctc_dso_handle); \ - void* p_##_name = dlsym(warpctc_dso_handle, #__name); \ - return reinterpret_cast(p_##_name)(args...); \ - } \ - } __name; // struct DynLoad__##__name - -// include all needed warp-ctc functions -DYNAMIC_LOAD_WARPCTC_WRAP(get_warpctc_version) -DYNAMIC_LOAD_WARPCTC_WRAP(ctcGetStatusString) -DYNAMIC_LOAD_WARPCTC_WRAP(compute_ctc_loss) -DYNAMIC_LOAD_WARPCTC_WRAP(get_workspace_size) - -#undef DYNAMIC_LOAD_WARPCTC_WRAP - -} /* namespace dynload */ - -#define WARPCTC_GET_VERSION dynload::get_warpctc_version -#define WARPCTC_GET_STATUS_STRING dynload::ctcGetStatusString - -static int g_warpctcVersion = -1; -#ifndef PADDLE_TYPE_DOUBLE -#define WARPCTC_COMPUTE_LOSS dynload::compute_ctc_loss -#define WARPCTC_GET_WORKSPACE_SIZE dynload::get_workspace_size -#else -hl_warpctc_status_t fatal(...) { - LOG(FATAL) << "warp-ctc [version " << g_warpctcVersion - << "] Error: not support double precision."; - // both of get_warpctc_version() and get_workspace_size() return an ctcStatus - // type value - return CTC_STATUS_EXECUTION_FAILED; -} -#define WARPCTC_COMPUTE_LOSS fatal -#define WARPCTC_GET_WORKSPACE_SIZE fatal -#endif - -/** - * Check build-in warp-ctc function using glog and it also - * support << operator for more details error info. - */ -#define CHECK_WARPCTC(warpctcStat) \ - CHECK_EQ(CTC_STATUS_SUCCESS, warpctcStat) \ - << "warp-ctc [version " << g_warpctcVersion \ - << "] Error: " << WARPCTC_GET_STATUS_STRING(warpctcStat) << " " - -void hl_warpctc_init(const size_t blank, - bool useGpu, - hl_warpctc_options_t* options) { - CHECK_NOTNULL(options); - - g_warpctcVersion = WARPCTC_GET_VERSION(); - - if (useGpu) { -#ifdef __NVCC__ - options->loc = CTC_GPU; - options->stream = STREAM_DEFAULT; -#else - LOG(FATAL) << "[warpctc init] GPU is not enabled."; -#endif - } else { - options->loc = CTC_CPU; - options->num_threads = 1; - } - - options->blank_label = blank; -} - -void hl_warpctc_compute_loss(const real* batchInput, - real* batchGrad, - const int* cpuLabels, - const int* cpuLabelLengths, - const int* cpuInputLengths, - const size_t numClasses, - const size_t numSequences, - real* cpuCosts, - void* workspace, - hl_warpctc_options_t* options) { - CHECK_NOTNULL(batchInput); - CHECK_NOTNULL(cpuLabels); - CHECK_NOTNULL(cpuLabelLengths); - CHECK_NOTNULL(cpuInputLengths); - CHECK_NOTNULL(cpuCosts); - CHECK_NOTNULL(workspace); - CHECK_NOTNULL(options); - - CHECK_WARPCTC(WARPCTC_COMPUTE_LOSS(batchInput, - batchGrad, - cpuLabels, - cpuLabelLengths, - cpuInputLengths, - numClasses, - numSequences, - cpuCosts, - workspace, - *options)); -} - -void hl_warpctc_get_workspace_size(const int* cpuLabelLengths, - const int* cpuInputLengths, - const size_t numClasses, - const size_t numSequences, - hl_warpctc_options_t* options, - size_t* bytes) { - CHECK_NOTNULL(cpuLabelLengths); - CHECK_NOTNULL(cpuInputLengths); - CHECK_NOTNULL(options); - CHECK_NOTNULL(bytes); - - CHECK_WARPCTC(WARPCTC_GET_WORKSPACE_SIZE(cpuLabelLengths, - cpuInputLengths, - numClasses, - numSequences, - *options, - bytes)); -} diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec new file mode 100644 index 0000000000000000000000000000000000000000..c020ff45ad3f3a72bf8a88622df333c1765a3d21 --- /dev/null +++ b/paddle/fluid/API.spec @@ -0,0 +1,411 @@ +paddle.fluid.Program.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.block ArgSpec(args=['self', 'index'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.clone ArgSpec(args=['self', 'for_test'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.Program.copy_data_info_from ArgSpec(args=['self', 'other'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.create_block ArgSpec(args=['self', 'parent_idx'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.Program.current_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.get_desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.global_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.inference_optimize ArgSpec(args=['self', 'export_for_deployment'], varargs=None, keywords=None, defaults=(True,)) +paddle.fluid.Program.list_vars ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.optimized_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.prune ArgSpec(args=['self', 'targets'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.rollback ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Program.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.Operator.__init__ ArgSpec(args=['self', 'block', 'desc', 'type', 'inputs', 'outputs', 'attrs'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.Operator.all_attrs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.attr_type ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.block_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.block_attr_id ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.blocks_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.blocks_attr_ids ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.has_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.has_kernel ArgSpec(args=['self', 'op_type'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.input ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.output ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.rename_input ArgSpec(args=['self', 'old_name', 'new_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.rename_output ArgSpec(args=['self', 'old_name', 'new_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.set_attr ArgSpec(args=['self', 'name', 'val'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Operator.to_string ArgSpec(args=['self', 'throw_on_error'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Parameter.__init__ ArgSpec(args=['self', 'block', 'shape', 'dtype'], varargs=None, keywords='kwargs', defaults=None) +paddle.fluid.Parameter.astype ArgSpec(args=['self', 'dtype'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Parameter.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None) +paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None) +paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.get_var ArgSpec(args=['name', 'program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False)) +paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None) +paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.Trainer.__init__ ArgSpec(args=['self', 'train_func', 'optimizer_func', 'param_path', 'place', 'parallel', 'checkpoint_config'], varargs=None, keywords=None, defaults=(None, None, False, None)) +paddle.fluid.Trainer.save_params ArgSpec(args=['self', 'param_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Trainer.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Trainer.test ArgSpec(args=['self', 'reader', 'feed_order'], varargs=None, keywords=None, defaults=None) +paddle.fluid.Trainer.train ArgSpec(args=['self', 'num_epochs', 'event_handler', 'reader', 'feed_order'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.BeginEpochEvent.__init__ ArgSpec(args=['self', 'epoch_id'], varargs=None, keywords=None, defaults=None) +paddle.fluid.EndEpochEvent.__init__ ArgSpec(args=['self', 'epoch_id'], varargs=None, keywords=None, defaults=None) +paddle.fluid.BeginStepEvent.__init__ ArgSpec(args=['self', 'epoch_id', 'step_id'], varargs=None, keywords=None, defaults=None) +paddle.fluid.EndStepEvent.__init__ ArgSpec(args=['self', 'epoch_id', 'step_id', 'metrics'], varargs=None, keywords=None, defaults=None) +paddle.fluid.CheckpointConfig.__init__ ArgSpec(args=['self', 'checkpoint_dir', 'max_num_checkpoints', 'epoch_interval', 'step_interval'], varargs=None, keywords=None, defaults=(None, 3, 1, 10)) +paddle.fluid.Inferencer.__init__ ArgSpec(args=['self', 'infer_func', 'param_path', 'place', 'parallel'], varargs=None, keywords=None, defaults=(None, False)) +paddle.fluid.Inferencer.infer ArgSpec(args=['self', 'inputs', 'return_numpy'], varargs=None, keywords=None, defaults=(True,)) +paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True)) +paddle.fluid.InferenceTranspiler.__init__ +paddle.fluid.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0)) +paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.DistributeTranspilerConfig.__init__ +paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id'], varargs=None, keywords='kwargs', defaults=(None, None, None, None, None, 1, 0)) +paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True)) +paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None +paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None +paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ReduceStrategy, arg0: int) -> None +paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.BuildStrategy) -> None +paddle.fluid.create_lod_tensor ArgSpec(args=['data', 'recursive_seq_lens', 'place'], varargs=None, keywords=None, defaults=None) +paddle.fluid.create_random_int_lodtensor ArgSpec(args=['recursive_seq_lens', 'base_shape', 'place', 'low', 'high'], varargs=None, keywords=None, defaults=None) +paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.io.save_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.io.load_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.io.load_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.io.load_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True)) +paddle.fluid.io.load_inference_model ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.io.get_inference_program ArgSpec(args=['target_vars', 'main_program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.initializer.ConstantInitializer.__init__ ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False)) +paddle.fluid.initializer.UniformInitializer.__init__ ArgSpec(args=['self', 'low', 'high', 'seed'], varargs=None, keywords=None, defaults=(-1.0, 1.0, 0)) +paddle.fluid.initializer.NormalInitializer.__init__ ArgSpec(args=['self', 'loc', 'scale', 'seed'], varargs=None, keywords=None, defaults=(0.0, 1.0, 0)) +paddle.fluid.initializer.XavierInitializer.__init__ ArgSpec(args=['self', 'uniform', 'fan_in', 'fan_out', 'seed'], varargs=None, keywords=None, defaults=(True, None, None, 0)) +paddle.fluid.initializer.BilinearInitializer.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.initializer.MSRAInitializer.__init__ ArgSpec(args=['self', 'uniform', 'fan_in', 'seed'], varargs=None, keywords=None, defaults=(True, None, 0)) +paddle.fluid.initializer.force_init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None) +paddle.fluid.initializer.init_on_cpu ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.layers.fc ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'use_mkldnn', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, False, None, False, None)) +paddle.fluid.layers.embedding ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32')) +paddle.fluid.layers.dynamic_lstm ArgSpec(args=['input', 'size', 'h_0', 'c_0', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'float32', None)) +paddle.fluid.layers.dynamic_lstmp ArgSpec(args=['input', 'size', 'proj_size', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'proj_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'tanh', 'float32', None)) +paddle.fluid.layers.dynamic_gru ArgSpec(args=['input', 'size', 'param_attr', 'bias_attr', 'is_reverse', 'gate_activation', 'candidate_activation', 'h_0'], varargs=None, keywords=None, defaults=(None, None, False, 'sigmoid', 'tanh', None)) +paddle.fluid.layers.gru_unit ArgSpec(args=['input', 'hidden', 'size', 'param_attr', 'bias_attr', 'activation', 'gate_activation'], varargs=None, keywords=None, defaults=(None, None, 'tanh', 'sigmoid')) +paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.crf_decoding ArgSpec(args=['input', 'param_attr', 'label'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None)) +paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, False, None, None)) +paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, False, None, None)) +paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn'], varargs=None, keywords=None, defaults=(None, None, True)) +paddle.fluid.layers.softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None)) +paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'use_mkldnn', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, False, None)) +paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'use_mkldnn', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, False, None)) +paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'use_mkldnn', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, False, None, None, None, False, False)) +paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) +paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) +paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None)) +paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None)) +paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) +paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) +paddle.fluid.layers.reduce_max ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) +paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) +paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) +paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None)) +paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)) +paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None)) +paddle.fluid.layers.l2_normalize ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)) +paddle.fluid.layers.matmul ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'name'], varargs=None, keywords=None, defaults=(False, False, None)) +paddle.fluid.layers.topk ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_times'], varargs=None, keywords=None, defaults=(0, False)) +paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None)) +paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None)) +paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None)) +paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1)) +paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None)) +paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.lrn ArgSpec(args=['input', 'n', 'k', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(5, 1.0, 0.0001, 0.75, None)) +paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None)) +paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None)) +paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)) +paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)) +paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR')) +paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)) +paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.random_crop ArgSpec(args=['x', 'shape', 'seed'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.relu ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.log ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None)) +paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)) +paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True)) +paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)) +paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.shuffle ArgSpec(args=['reader', 'buffer_size'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.batch ArgSpec(args=['reader', 'batch_size'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.double_buffer ArgSpec(args=['reader', 'place', 'name'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes', 'lod_levels', 'for_parallel'], varargs=None, keywords=None, defaults=(True,)) +paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)) +paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.Preprocessor.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.Preprocessor.outputs ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None) +paddle.fluid.layers.load ArgSpec(args=['out', 'file_path', 'load_as_fp16'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.create_tensor ArgSpec(args=['dtype', 'name', 'persistable'], varargs=None, keywords=None, defaults=(None, False)) +paddle.fluid.layers.create_parameter ArgSpec(args=['shape', 'dtype', 'name', 'attr', 'is_bias', 'default_initializer'], varargs=None, keywords=None, defaults=(None, None, False, None)) +paddle.fluid.layers.create_global_var ArgSpec(args=['shape', 'value', 'dtype', 'persistable', 'force_cpu', 'name'], varargs=None, keywords=None, defaults=(False, False, None)) +paddle.fluid.layers.cast ArgSpec(args=['x', 'dtype'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.concat ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(0, None)) +paddle.fluid.layers.sums ArgSpec(args=['input', 'out'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.assign ArgSpec(args=['input', 'output'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.fill_constant_batch_size_like ArgSpec(args=['input', 'shape', 'dtype', 'value', 'input_dim_idx', 'output_dim_idx'], varargs=None, keywords=None, defaults=(0, 0)) +paddle.fluid.layers.fill_constant ArgSpec(args=['shape', 'dtype', 'value', 'force_cpu', 'out'], varargs=None, keywords=None, defaults=(False, None)) +paddle.fluid.layers.argmin ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,)) +paddle.fluid.layers.argmax ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,)) +paddle.fluid.layers.argsort ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(-1, None)) +paddle.fluid.layers.ones ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.Switch.case ArgSpec(args=['self', 'condition'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.Switch.default ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.increment ArgSpec(args=['x', 'value', 'in_place'], varargs=None, keywords=None, defaults=(1.0, True)) +paddle.fluid.layers.array_write ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.create_array ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.less_than ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None)) +paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords='ignored', defaults=(None,)) +paddle.fluid.layers.array_read ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.array_length ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.IfElse.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.IfElse.false_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.IfElse.input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.IfElse.output ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None) +paddle.fluid.layers.IfElse.true_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.DynamicRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.DynamicRNN.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.layers.DynamicRNN.memory ArgSpec(args=['self', 'init', 'shape', 'value', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, False, 'float32')) +paddle.fluid.layers.DynamicRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) +paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.DynamicRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.DynamicRNN.update_memory ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.StaticRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.StaticRNN.memory ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1)) +paddle.fluid.layers.StaticRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) +paddle.fluid.layers.StaticRNN.step ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.StaticRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.StaticRNN.step_output ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.StaticRNN.update_memory ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.reorder_lod_tensor_by_rank ArgSpec(args=['x', 'rank_table'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.ParallelDo.__init__ ArgSpec(args=['self', 'places', 'use_nccl', 'name'], varargs=None, keywords=None, defaults=(False, None)) +paddle.fluid.layers.ParallelDo.do ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.ParallelDo.get_parameters ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.ParallelDo.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.ParallelDo.read_input ArgSpec(args=['self', 'var'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.ParallelDo.write_output ArgSpec(args=['self', 'var'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.Print ArgSpec(args=['input', 'first_n', 'message', 'summarize', 'print_tensor_name', 'print_tensor_type', 'print_tensor_shape', 'print_tensor_lod', 'print_phase'], varargs=None, keywords=None, defaults=(-1, None, -1, True, True, True, True, 'both')) +paddle.fluid.layers.is_empty ArgSpec(args=['x', 'cond'], varargs=None, keywords='ignored', defaults=(None,)) +paddle.fluid.layers.mean ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.mul ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.scale ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_add ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_div ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_sub ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_mul ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_max ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_min ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elementwise_pow ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.clip ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.clip_by_norm ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.logical_and ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.logical_or ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.logical_xor ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.logical_not ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.uniform_random_batch_size_like ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.gaussian_random ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.gaussian_random_batch_size_like ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.scatter ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.sum ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.slice ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.shape ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.maxout ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.sigmoid ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.logsigmoid ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.exp ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.tanh ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.tanh_shrink ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.softshrink ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.sqrt ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.abs ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.ceil ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.floor ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.cos ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.sin ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.round ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.reciprocal ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.square ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.softplus ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.softsign ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.brelu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.leaky_relu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.soft_relu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.elu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.relu6 ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.pow ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.stanh ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.hard_sigmoid ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.swish ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.uniform_random ArgSpec(args=['shape', 'dtype', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False)) +paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)) +paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.detection_output ArgSpec(args=['loc', 'scores', 'prior_box', 'prior_box_var', 'background_label', 'nms_threshold', 'nms_top_k', 'keep_top_k', 'score_threshold', 'nms_eta'], varargs=None, keywords=None, defaults=(0, 0.3, 400, 200, 0.01, 1.0)) +paddle.fluid.layers.ssd_loss ArgSpec(args=['location', 'confidence', 'gt_box', 'gt_label', 'prior_box', 'prior_box_var', 'background_label', 'overlap_threshold', 'neg_pos_ratio', 'neg_overlap', 'loc_loss_weight', 'conf_loss_weight', 'match_type', 'mining_type', 'normalize', 'sample_size'], varargs=None, keywords=None, defaults=(None, 0, 0.5, 3.0, 0.5, 1.0, 1.0, 'per_prediction', 'max_negative', True, None)) +paddle.fluid.layers.detection_map ArgSpec(args=['detect_res', 'label', 'class_num', 'background_label', 'overlap_threshold', 'evaluate_difficult', 'has_state', 'input_states', 'out_states', 'ap_version'], varargs=None, keywords=None, defaults=(0, 0.3, True, None, None, None, 'integral')) +paddle.fluid.layers.rpn_target_assign ArgSpec(args=['loc', 'scores', 'anchor_box', 'gt_box', 'rpn_batch_size_per_im', 'fg_fraction', 'rpn_positive_overlap', 'rpn_negative_overlap'], varargs=None, keywords=None, defaults=(256, 0.25, 0.7, 0.3)) +paddle.fluid.layers.anchor_generator ArgSpec(args=['input', 'anchor_sizes', 'aspect_ratios', 'variance', 'stride', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, [0.1, 0.1, 0.2, 0.2], None, 0.5, None)) +paddle.fluid.layers.iou_similarity ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.box_coder ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.polygon_box_transform ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None)) +paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk'], varargs=None, keywords=None, defaults=('ROC', 200, 1)) +paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.natural_exp_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.inverse_time_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.polynomial_decay ArgSpec(args=['learning_rate', 'decay_steps', 'end_learning_rate', 'power', 'cycle'], varargs=None, keywords=None, defaults=(0.0001, 1.0, False)) +paddle.fluid.layers.piecewise_decay ArgSpec(args=['boundaries', 'values'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.noam_decay ArgSpec(args=['d_model', 'warmup_steps'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.append_LARS ArgSpec(args=['params_grads', 'learning_rate', 'weight_decay'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.InitState.__init__ ArgSpec(args=['self', 'init', 'shape', 'value', 'init_boot', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, None, False, 'float32')) +paddle.fluid.contrib.StateCell.__init__ ArgSpec(args=['self', 'inputs', 'states', 'out_state', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.StateCell.compute_state ArgSpec(args=['self', 'inputs'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.StateCell.get_input ArgSpec(args=['self', 'input_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.StateCell.get_state ArgSpec(args=['self', 'state_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.StateCell.out_state ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.StateCell.set_state ArgSpec(args=['self', 'state_name', 'state_value'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.StateCell.state_updater ArgSpec(args=['self', 'updater'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.StateCell.update_states ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.TrainingDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.TrainingDecoder.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.contrib.TrainingDecoder.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) +paddle.fluid.contrib.TrainingDecoder.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.TrainingDecoder.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.BeamSearchDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'init_ids', 'init_scores', 'target_dict_dim', 'word_dim', 'input_var_dict', 'topk_size', 'sparse_emb', 'max_len', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=({}, 50, True, 100, 1, 1, None)) +paddle.fluid.contrib.BeamSearchDecoder.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False)) +paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array', 'value'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.transpiler.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True)) +paddle.fluid.transpiler.InferenceTranspiler.__init__ +paddle.fluid.transpiler.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.transpiler.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0)) +paddle.fluid.transpiler.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.transpiler.HashName.__init__ ArgSpec(args=['self', 'pserver_endpoints'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.HashName.dispatch ArgSpec(args=['self', 'varlist'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.HashName.reset ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.RoundRobin.__init__ ArgSpec(args=['self', 'pserver_endpoints'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.RoundRobin.dispatch ArgSpec(args=['self', 'varlist'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.RoundRobin.reset ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.transpiler.DistributeTranspilerConfig.__init__ +paddle.fluid.nets.simple_img_conv_pool ArgSpec(args=['input', 'num_filters', 'filter_size', 'pool_size', 'pool_stride', 'pool_padding', 'pool_type', 'global_pooling', 'conv_stride', 'conv_padding', 'conv_dilation', 'conv_groups', 'param_attr', 'bias_attr', 'act', 'use_cudnn', 'use_mkldnn'], varargs=None, keywords=None, defaults=(0, 'max', False, 1, 0, 1, 1, None, None, None, True, False)) +paddle.fluid.nets.sequence_conv_pool ArgSpec(args=['input', 'num_filters', 'filter_size', 'param_attr', 'act', 'pool_type'], varargs=None, keywords=None, defaults=(None, 'sigmoid', 'max')) +paddle.fluid.nets.glu ArgSpec(args=['input', 'dim'], varargs=None, keywords=None, defaults=(-1,)) +paddle.fluid.nets.scaled_dot_product_attention ArgSpec(args=['queries', 'keys', 'values', 'num_heads', 'dropout_rate'], varargs=None, keywords=None, defaults=(1, 0.0)) +paddle.fluid.optimizer.SGDOptimizer.__init__ ArgSpec(args=['self', 'learning_rate'], varargs=None, keywords='kwargs', defaults=None) +paddle.fluid.optimizer.SGDOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'use_nesterov'], varargs=None, keywords='kwargs', defaults=(False,)) +paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon'], varargs=None, keywords='kwargs', defaults=(1e-06,)) +paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon'], varargs=None, keywords='kwargs', defaults=(0.001, 0.9, 0.999, 1e-08)) +paddle.fluid.optimizer.AdamOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.AdamaxOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon'], varargs=None, keywords='kwargs', defaults=(0.001, 0.9, 0.999, 1e-08)) +paddle.fluid.optimizer.AdamaxOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'decay', 'epsilon'], varargs=None, keywords='kwargs', defaults=(0.95, 1e-06)) +paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.FtrlOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power'], varargs=None, keywords='kwargs', defaults=(0.0, 0.0, -0.5)) +paddle.fluid.optimizer.FtrlOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.RMSPropOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum'], varargs=None, keywords='kwargs', defaults=(0.95, 1e-06, 0.0)) +paddle.fluid.optimizer.RMSPropOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.AdadeltaOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho'], varargs=None, keywords='kwargs', defaults=(1e-06, 0.95)) +paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window'], varargs=None, keywords='kwargs', defaults=(10000, 10000)) +paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.optimizer.ModelAverage.restore ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None) +paddle.fluid.backward.append_backward ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.regularizer.L1DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)) +paddle.fluid.regularizer.L2DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)) +paddle.fluid.LoDTensor.__init__ 1. __init__(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None 2. __init__(self: paddle.fluid.core.LoDTensor) -> None +paddle.fluid.LoDTensor.has_valid_recursive_sequence_lengths has_valid_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> bool +paddle.fluid.LoDTensor.lod lod(self: paddle.fluid.core.LoDTensor) -> List[List[int]] +paddle.fluid.LoDTensor.recursive_sequence_lengths recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> List[List[int]] +paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None 2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None 3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None 4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None 5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None 6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None 7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None 8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None 9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None 10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None 11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None 12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None 13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None 14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None 15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None 16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None 17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None 18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None 19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None 20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None 21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None +paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None +paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None +paddle.fluid.LoDTensor.shape shape(self: paddle.fluid.core.Tensor) -> List[int] +paddle.fluid.LoDTensorArray.__init__ __init__(self: paddle.fluid.core.LoDTensorArray) -> None +paddle.fluid.LoDTensorArray.append append(self: paddle.fluid.core.LoDTensorArray, arg0: paddle.fluid.core.LoDTensor) -> None +paddle.fluid.CPUPlace.__init__ __init__(self: paddle.fluid.core.CPUPlace) -> None +paddle.fluid.CUDAPlace.__init__ __init__(self: paddle.fluid.core.CUDAPlace, arg0: int) -> None +paddle.fluid.CUDAPinnedPlace.__init__ __init__(self: paddle.fluid.core.CUDAPinnedPlace) -> None +paddle.fluid.ParamAttr.__init__ ArgSpec(args=['self', 'name', 'initializer', 'learning_rate', 'regularizer', 'trainable', 'gradient_clip', 'do_model_average'], varargs=None, keywords=None, defaults=(None, None, 1.0, None, True, None, False)) +paddle.fluid.WeightNormParamAttr.__init__ ArgSpec(args=['self', 'dim'], varargs=None, keywords='kwargs', defaults=(None,)) +paddle.fluid.DataFeeder.__init__ ArgSpec(args=['self', 'feed_list', 'place', 'program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.DataFeeder.decorate_reader ArgSpec(args=['self', 'reader', 'multi_devices', 'num_places', 'drop_last'], varargs=None, keywords=None, defaults=(None, True)) +paddle.fluid.DataFeeder.feed ArgSpec(args=['self', 'iterable'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DataFeeder.feed_parallel ArgSpec(args=['self', 'iterable', 'num_places'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.clip.ErrorClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.clip.GradientClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.clip.GradientClipByNorm.__init__ ArgSpec(args=['self', 'clip_norm'], varargs=None, keywords=None, defaults=None) +paddle.fluid.clip.GradientClipByGlobalNorm.__init__ ArgSpec(args=['self', 'clip_norm', 'group_name'], varargs=None, keywords=None, defaults=('default_group',)) +paddle.fluid.profiler.cuda_profiler ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.profiler.reset_profiler ArgSpec(args=[], varargs=None, keywords=None, defaults=None) +paddle.fluid.profiler.profiler ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.profiler.start_profiler ArgSpec(args=['state'], varargs=None, keywords=None, defaults=None) +paddle.fluid.profiler.stop_profiler ArgSpec(args=['sorted_key', 'profile_path'], varargs=None, keywords=None, defaults=(None, '/tmp/profile')) +paddle.fluid.unique_name.generate ArgSpec(args=['key'], varargs=None, keywords=None, defaults=None) +paddle.fluid.unique_name.switch ArgSpec(args=['new_generator'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.unique_name.guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.recordio_writer.convert_reader_to_recordio_file ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)) +paddle.fluid.recordio_writer.convert_reader_to_recordio_files ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)) +paddle.fluid.Scope.__init__ __init__(self: paddle.fluid.core.Scope) -> None +paddle.fluid.Scope.drop_kids drop_kids(self: paddle.fluid.core.Scope) -> None +paddle.fluid.Scope.find_var find_var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable +paddle.fluid.Scope.new_scope new_scope(self: paddle.fluid.core.Scope) -> paddle.fluid.core.Scope +paddle.fluid.Scope.var var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable diff --git a/paddle/fluid/CMakeLists.txt b/paddle/fluid/CMakeLists.txt index d274d96c29bdbf5973d568d783369c3975bdc436..2577e59d9cf24c26b7c04aa00cdde6cde17f7206 100644 --- a/paddle/fluid/CMakeLists.txt +++ b/paddle/fluid/CMakeLists.txt @@ -5,5 +5,7 @@ add_subdirectory(operators) add_subdirectory(pybind) add_subdirectory(string) add_subdirectory(recordio) -# NOTE: please add subdirectory inference at last. -add_subdirectory(inference) +if(WITH_INFERENCE) + # NOTE: please add subdirectory inference at last. + add_subdirectory(inference) +endif() diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index ed1e70c6460b513c1d2e1add18ac037f71d36944..1d62792b80dd002b894da28be9162fc7d3ce054e 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -1,4 +1,5 @@ add_subdirectory(details) +add_subdirectory(ir) # ddim lib proto_library(framework_proto SRCS framework.proto) @@ -6,10 +7,11 @@ cc_library(ddim SRCS ddim.cc DEPS eigen3 boost) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) nv_test(dim_test SRCS dim_test.cu DEPS ddim) cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context) +cc_test(data_type_test SRCS data_type_test.cc DEPS data_type place tensor) if(WITH_GPU) - nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type) + nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context) else() - cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type) + cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context) endif() cc_test(tensor_test SRCS tensor_test.cc DEPS tensor) @@ -21,12 +23,18 @@ endif() cc_test(eigen_test SRCS eigen_test.cc DEPS tensor) -nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context init) +if(WITH_GPU) + nv_test(mixed_vector_test SRCS mixed_vector_test.cc mixed_vector_test.cu DEPS place memory device_context tensor) +else() + cc_test(mixed_vector_test SRCS mixed_vector_test.cc DEPS place memory device_context tensor) +endif() + cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio) cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory) -nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init) +nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor) cc_library(reader SRCS reader.cc DEPS lod_tensor ddim) +cc_test(reader_test SRCS reader_test.cc DEPS reader) cc_test(variable_test SRCS variable_test.cc) @@ -38,7 +46,7 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope) cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor) nv_test(data_device_transform_test SRCS data_device_transform_test.cu - DEPS operator op_registry init math_function) + DEPS operator op_registry device_context math_function) if(WITH_GPU) nv_library(data_type_transform SRCS data_type_transform.cu DEPS tensor) @@ -63,7 +71,7 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context) cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog shape_inference data_transform lod_tensor profiler) -cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init) +cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context) cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog) cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc) @@ -83,11 +91,16 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor) cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog) -cc_library(executor SRCS executor.cc DEPS op_registry device_context scope -framework_proto glog lod_rank_table feed_fetch_method) +if(WITH_DISTRIBUTE) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr) + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) +else() + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method) +endif() -cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor) +cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph graph_viz_pass multi_devices_graph_pass multi_devices_graph_print_pass multi_devices_graph_check_pass) cc_library(prune SRCS prune.cc DEPS framework_proto) cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context) @@ -96,14 +109,14 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry cc_library(selected_rows SRCS selected_rows.cc DEPS tensor) cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows) -cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece operator) -cc_test(init_test SRCS init_test.cc DEPS init) - cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto) cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc) - + # cc_test(channel_test SRCS channel_test.cc) cc_test(tuple_test SRCS tuple_test.cc ) -cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op - channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op - conditional_block_op while_op assign_op print_op executor proto_desc) + +# disable test temporarily. +# TODO https://github.com/PaddlePaddle/Paddle/issues/11971 +# cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op +# channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op +# conditional_block_op while_op assign_op print_op executor proto_desc) diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index fd409ed4c0f7a504686765909e9c71692aab8824..f537e4b9e569dd4c513ac0efde7240833bcf04b6 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -169,17 +169,13 @@ void BlockDesc::Flush() { } if (need_update_) { - auto &op_field = *this->desc_->mutable_ops(); - this->ClearPBOps(); - op_field.Reserve(static_cast(ops_.size())); + this->desc_->mutable_ops()->Clear(); for (auto &op_desc : ops_) { - op_field.AddAllocated(op_desc->Proto()); + this->desc_->mutable_ops()->Add()->CopyFrom(*op_desc->Proto()); } - auto &var_field = *this->desc_->mutable_vars(); - this->ClearPBVars(); - var_field.Reserve(static_cast(vars_.size())); + this->desc_->mutable_vars()->Clear(); for (auto &var_desc : vars_) { - var_field.AddAllocated(var_desc.second->Proto()); + this->desc_->mutable_vars()->Add()->CopyFrom(*var_desc.second->Proto()); } need_update_ = false; } @@ -200,7 +196,7 @@ BlockDesc::BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc) vars_[var_desc.name()].reset(new VarDesc(var_desc)); } for (const proto::OpDesc &op_desc : desc_->ops()) { - ops_.emplace_back(new OpDesc(op_desc, prog, this)); + ops_.emplace_back(new OpDesc(op_desc, this)); } } @@ -209,7 +205,7 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, : prog_(prog), desc_(desc) { need_update_ = true; for (auto &op : other.ops_) { - ops_.emplace_back(new OpDesc(*op->Proto(), prog, this)); + ops_.emplace_back(new OpDesc(*op, this)); } for (auto &it : other.vars_) { auto *var = new VarDesc(*it.second); @@ -217,22 +213,6 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, } } -void BlockDesc::ClearPBOps() { - auto ops = this->desc_->mutable_ops(); - while (!ops->empty()) { - // we do not own the OpDesc, so release the ownership. - ops->ReleaseLast(); - } -} - -void BlockDesc::ClearPBVars() { - auto vars = this->desc_->mutable_vars(); - while (!vars->empty()) { - // we do not own the VarDesc, so release the ownership. - vars->ReleaseLast(); - } -} - void BlockDesc::SetForwardBlockID(int32_t forward_block_id) { PADDLE_ENFORCE(!desc_->has_forward_block_idx(), "Parent block ID has been set to %d. Cannot set to %d", diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index 600601669c5d56a3ffc2fb9c804ffad5fde58f0b..960ca39e1eadd3c064beb0e2c1342a406c4f0b6a 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -41,11 +41,6 @@ class BlockDesc { BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, ProgramDesc *prog); - ~BlockDesc() { - this->ClearPBVars(); - this->ClearPBOps(); - } - int32_t ID() const { return desc_->idx(); } int32_t Parent() const { return desc_->parent_idx(); } @@ -93,9 +88,8 @@ class BlockDesc { OpDesc *InsertOp(size_t index); /* - * Remove Op and its input/output variables. - * Note that for either input or output variable, if it is also an input or - * output variable of other ops, we should remain it. + * Only remove op itself, + * do nothing to its input and output variables */ void RemoveOp(size_t s, size_t e); @@ -105,7 +99,7 @@ class BlockDesc { size_t OpSize() const { return ops_.size(); } - OpDesc *Op(int idx) { return ops_.at(idx).get(); } + OpDesc *Op(int idx) const { return ops_.at(idx).get(); } void Flush(); @@ -113,10 +107,6 @@ class BlockDesc { ProgramDesc *Program() const { return this->prog_; } - private: - void ClearPBOps(); - void ClearPBVars(); - private: ProgramDesc *prog_; // not_own proto::BlockDesc *desc_; // not_own diff --git a/paddle/fluid/framework/data_device_transform_test.cu b/paddle/fluid/framework/data_device_transform_test.cu index a91fe5c99d397ef1bf04f6d22e988b6d3f33e500..f2c55e533a2747325b1b16fdada37945a8ed3c42 100644 --- a/paddle/fluid/framework/data_device_transform_test.cu +++ b/paddle/fluid/framework/data_device_transform_test.cu @@ -14,13 +14,13 @@ limitations under the License. */ #include "gtest/gtest.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/elementwise_op_function.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/init.h" namespace paddle { namespace framework { diff --git a/paddle/fluid/framework/data_layout.h b/paddle/fluid/framework/data_layout.h index 9c5e2cf7ccdcea2822da42210ff1fdb915a9a4ec..b611bb77b4e1ec05b8bd029ac37cefba346c6eb0 100644 --- a/paddle/fluid/framework/data_layout.h +++ b/paddle/fluid/framework/data_layout.h @@ -27,6 +27,7 @@ enum class DataLayout { kNHWC = 0, kNCHW = 1, kAnyLayout = 2, + kMKLDNN = 3, // all layouts supported by MKLDNN internally }; inline DataLayout StringToDataLayout(const std::string& str) { @@ -41,6 +42,8 @@ inline DataLayout StringToDataLayout(const std::string& str) { return DataLayout::kNCHW; } else if (s == "ANYLAYOUT") { return DataLayout::kAnyLayout; + } else if (s == "MKLDNNLAYOUT") { + return DataLayout::kMKLDNN; } else { PADDLE_THROW("Unknown storage order string: %s", s); } @@ -54,8 +57,10 @@ inline std::string DataLayoutToString(const DataLayout& data_layout) { return "NCHW"; case DataLayout::kAnyLayout: return "ANY_LAYOUT"; + case DataLayout::kMKLDNN: + return "MKLDNNLAYOUT"; default: - PADDLE_THROW("unknown DataLayou %d", data_layout); + PADDLE_THROW("unknown DataLayout %d", data_layout); } } diff --git a/paddle/fluid/framework/data_layout_transform.cc b/paddle/fluid/framework/data_layout_transform.cc index 60ec60a427ba9046ce690eb75c27cd322fdd726d..cd00b7de7338982308acfa1f1e8c38e010c6a43b 100644 --- a/paddle/fluid/framework/data_layout_transform.cc +++ b/paddle/fluid/framework/data_layout_transform.cc @@ -16,6 +16,9 @@ #include #include "paddle/fluid/operators/math/math_function.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace framework { @@ -88,5 +91,84 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var, out->set_layout(expected_kernel_type.data_layout_); } +#ifdef PADDLE_WITH_MKLDNN +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; + +void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) { + switch (type) { + case mkldnn::memory::data_type::f32: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::s8: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::u8: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::s16: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::s32: + return platform::to_void_cast(tensor.data()); + default: + PADDLE_THROW("wrong mkldnn type provided"); + } +} +#endif + +void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, + const OpKernelType& expected_kernel_type, + const Tensor& in, Tensor* out) { + auto in_layout = kernel_type_for_var.data_layout_; + auto out_layout = expected_kernel_type.data_layout_; + + PADDLE_ENFORCE( + in_layout == DataLayout::kMKLDNN && out_layout != DataLayout::kMKLDNN, + "TransDataLayoutFromMKLDNN only supports transform from MKLDNN to " + "non-MKLDNN"); + +#ifdef PADDLE_WITH_MKLDNN + PADDLE_ENFORCE(in.format() != memory::format::format_undef && + in.format() != memory::format::any, + "Input tensor should have specified memory format"); + + // Set default as NCHW in case not specified + out_layout = + out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout; + + auto& pool = platform::DeviceContextPool::Instance(); + auto* dev_ctx = dynamic_cast( + pool.Get(expected_kernel_type.place_)); + auto& cpu_engine = dev_ctx->GetEngine(); + + std::vector in_tz = paddle::framework::vectorize2int(in.dims()); + std::vector out_tz = in_tz; + + memory::data_type in_type = ToMKLDNNDataType(in.type()); + PADDLE_ENFORCE(in_type != memory::data_type::data_undef, + "Input tensor type is not supported: ", in.type().name()); + memory::data_type out_type = in_type; + + auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format()); + auto out_format = + platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout)); + + void* in_data = GetDataFromTensor(in, in_type); + + // output tensor has the same dims as input. Reorder don't change dims + out->Resize(in.dims()); + + auto out_data = out->mutable_data(expected_kernel_type.place_, in.type()); + + auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data); + auto out_memory = + memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data); + + platform::Reorder(in_memory, out_memory); + + out->set_layout(out_layout); + // reset format since the out tensor will be feed to non-MKLDNN OPkernel + out->set_format(memory::format::format_undef); +#endif +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_layout_transform.h b/paddle/fluid/framework/data_layout_transform.h index 06b638663dd334837a3bcb7737e507fcbc871c7a..90bb206ec6b698bc23ad1a5c9609a25186ec6de8 100644 --- a/paddle/fluid/framework/data_layout_transform.h +++ b/paddle/fluid/framework/data_layout_transform.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include "paddle/fluid/framework/op_kernel_type.h" #include "paddle/fluid/framework/tensor.h" @@ -22,6 +23,51 @@ namespace paddle { namespace framework { +#ifdef PADDLE_WITH_MKLDNN +using MKLDNNFormat = mkldnn::memory::format; +using MKLDNNDataType = mkldnn::memory::data_type; + +inline MKLDNNFormat ToMKLDNNFormat(const DataLayout& layout) { + switch (layout) { + case DataLayout::kNHWC: + return MKLDNNFormat::nhwc; + case DataLayout::kNCHW: + return MKLDNNFormat::nchw; + default: + PADDLE_THROW("Fail to convert layout %s to MKLDNN format", + DataLayoutToString(layout)); + } +} + +inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) { + switch (format) { + case MKLDNNFormat::nhwc: + return DataLayout::kNHWC; + case MKLDNNFormat::nchw: + return DataLayout::kNCHW; + default: + PADDLE_THROW("Fail to convert MKLDNN format to paddle layout"); + } +} + +inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) { + static const std::map dict{ + {std::type_index(typeid(float)), MKLDNNDataType::f32}, // NOLINT + {std::type_index(typeid(char)), MKLDNNDataType::s8}, // NOLINT + {std::type_index(typeid(unsigned char)), MKLDNNDataType::u8}, + {std::type_index(typeid(int16_t)), MKLDNNDataType::s16}, + {std::type_index(typeid(int32_t)), MKLDNNDataType::s32}}; + auto iter = dict.find(type); + if (iter != dict.end()) return iter->second; + return MKLDNNDataType::data_undef; +} + +#endif + +void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, + const OpKernelType& expected_kernel_type, + const Tensor& in, Tensor* out); + std::vector GetAxis(const DataLayout& from, const DataLayout& to); void TransDataLayout(const OpKernelType& kernel_type_for_var, diff --git a/paddle/fluid/framework/data_transform.cc b/paddle/fluid/framework/data_transform.cc index 9c277a27da5af34fc9fb18ca073e369c05ecdf22..82872224501709080ff02a13464d58543a0abda8 100644 --- a/paddle/fluid/framework/data_transform.cc +++ b/paddle/fluid/framework/data_transform.cc @@ -18,26 +18,57 @@ limitations under the License. */ #include "paddle/fluid/framework/data_layout_transform.h" #include "paddle/fluid/framework/data_type_transform.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace framework { -static void PassTensorData(Tensor* from, Tensor* to) { +static void PassTensorData(Tensor *from, Tensor *to) { to->ShareDataWith(*from); *from = Tensor(); } -void DataTransform(const OpKernelType& expected_kernel_type, - const OpKernelType& kernel_type_for_var, - const Tensor& input_tensor, Tensor* output_tensor) { +void TransformData(const OpKernelType &expected_kernel_type, + const OpKernelType &kernel_type_for_var, + const Tensor &input_tensor, Tensor *output_tensor) { bool transformed = false; Tensor in; in.ShareDataWith(input_tensor); Tensor out; + DataLayout lin = kernel_type_for_var.data_layout_; + DataLayout lout = expected_kernel_type.data_layout_; // do layout transform - if (NeedTransformLayout(expected_kernel_type.data_layout_, - kernel_type_for_var.data_layout_)) { - TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out); + if (NeedTransformLayout(lout, lin)) { + if (lin == DataLayout::kMKLDNN || lout == DataLayout::kMKLDNN) { + PADDLE_ENFORCE( + !(lin == DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN), + "No layout transform needed between two MKLDNN OPKernels"); + + if (lin != DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN) { +#ifdef PADDLE_WITH_MKLDNN + // Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel + // Just set layout/format. No real transform occur + + auto out_format = platform::MKLDNNFormatForSize(in.dims().size(), + ToMKLDNNFormat(lin)); + + out.ShareDataWith(input_tensor); + out.set_layout(DataLayout::kMKLDNN); + out.set_format(out_format); +#endif + } else { + // Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel + // Do transform via MKLDNN lib + TransDataLayoutFromMKLDNN(kernel_type_for_var, expected_kernel_type, in, + &out); + } + } else { + // Case3 - transfrom between Non-MKLDNN OPKernels + TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out); + } transformed = true; PassTensorData(&out, &in); } @@ -62,17 +93,17 @@ void DataTransform(const OpKernelType& expected_kernel_type, output_tensor->ShareDataWith(in); } -void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor, - Variable* out_var) { +void SetTensorToVariable(const Variable &in_var, const Tensor &tensor, + Variable *out_var) { if (in_var.IsType()) { - auto& in_lod_tensor = in_var.Get(); - auto* tran_lod_tensor = out_var->GetMutable(); + auto &in_lod_tensor = in_var.Get(); + auto *tran_lod_tensor = out_var->GetMutable(); tran_lod_tensor->set_lod(in_lod_tensor.lod()); tran_lod_tensor->set_layout(in_lod_tensor.layout()); tran_lod_tensor->ShareDataWith(tensor); } else if (in_var.IsType()) { - auto& in_selected_rows = in_var.Get(); - auto* trans_selected_rows = out_var->GetMutable(); + auto &in_selected_rows = in_var.Get(); + auto *trans_selected_rows = out_var->GetMutable(); trans_selected_rows->set_height(in_selected_rows.height()); trans_selected_rows->set_rows(in_selected_rows.rows()); trans_selected_rows->mutable_value()->ShareDataWith(tensor); diff --git a/paddle/fluid/framework/data_transform.h b/paddle/fluid/framework/data_transform.h index dee5d8c7c1126013742460df1d94bb364220ad09..ae3ab051bda2e698801cc6fe6e3ddddf039f5385 100644 --- a/paddle/fluid/framework/data_transform.h +++ b/paddle/fluid/framework/data_transform.h @@ -30,12 +30,15 @@ limitations under the License. */ namespace paddle { namespace framework { -void DataTransform(const OpKernelType& expected_kernel_type, - const OpKernelType& kernel_type_for_var, - const Tensor& input_tensor, Tensor* out); - -void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor, - Variable* out_var); +void TransformData(const OpKernelType &expected_kernel_type, + const OpKernelType &kernel_type_for_var, + const Tensor &input_tensor, Tensor *out); + +/** + * Set OutVar from InVar, except the tensor is shared with `tensor` + */ +void SetTensorToVariable(const Variable &in_var, const Tensor &tensor, + Variable *out_var); } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc index b6b93cf422a60c1d8e9cb8b477efd562f9fe4758..1a9ce746ea840bc088d222cc4e9bc05159d64734 100644 --- a/paddle/fluid/framework/data_type.cc +++ b/paddle/fluid/framework/data_type.cc @@ -17,6 +17,8 @@ #include #include +using float16 = paddle::platform::float16; + namespace paddle { namespace framework { @@ -28,6 +30,9 @@ struct DataTypeMap { }; static DataTypeMap* InitDataTypeMap(); +// C++11 removes the need for manual locking. Concurrent execution shall wait if +// a static local variable is already being initialized. +// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex static DataTypeMap& gDataTypeMap() { static DataTypeMap* g_data_type_map_ = InitDataTypeMap(); return *g_data_type_map_; @@ -50,7 +55,7 @@ static DataTypeMap* InitDataTypeMap() { RegisterType(retv, proto_type, #cc_type) // NOTE: Add your customize type here. - RegType(platform::float16, proto::VarType::FP16); + RegType(float16, proto::VarType::FP16); RegType(float, proto::VarType::FP32); RegType(double, proto::VarType::FP64); RegType(int, proto::VarType::INT32); diff --git a/paddle/fluid/framework/data_type_test.cc b/paddle/fluid/framework/data_type_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..54c41c55ba63c0b2001cfcb6a9e94fbb0036d437 --- /dev/null +++ b/paddle/fluid/framework/data_type_test.cc @@ -0,0 +1,40 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#include "paddle/fluid/framework/data_type.h" + +#include +#include "gtest/gtest.h" +#include "paddle/fluid/framework/tensor.h" + +TEST(DataType, float16) { + using paddle::framework::Tensor; + using paddle::platform::CPUPlace; + using paddle::platform::float16; + namespace f = paddle::framework; + f::proto::VarType::Type dtype = f::proto::VarType::FP16; + + Tensor tensor; + CPUPlace cpu; + tensor.mutable_data(cpu, f::ToTypeIndex(dtype)); + + // test fp16 tensor + EXPECT_EQ(tensor.type(), std::type_index(typeid(float16))); + + // test fp16 size + EXPECT_EQ(f::SizeOfType(f::ToTypeIndex(dtype)), 2u); + + // test debug info + std::string type = "float16"; + EXPECT_STREQ(f::DataTypeToString(dtype).c_str(), type.c_str()); +} diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index 1bcd8412eb2d618b923bcd0557d118af62271f4a..8f6c4163d6ee11fbe83f603f6148c2ac6175324d 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -1,34 +1,37 @@ -cc_library(var_handle SRCS var_handle.cc DEPS place) +cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node) cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor) cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry) cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry) -cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base) -cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph) +cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper) +cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper) +cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc DEPS multi_devices_helper) cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows) if(WITH_GPU) - nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory - dynload_cuda) - set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle) + nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory + dynload_cuda variable_visitor) nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda) nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda) else() - set(multi_devices_graph_builder_deps) + cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory + variable_visitor) cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim) cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) endif() +cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor) cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) +cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) -cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle - scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle) +cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle + scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle) -cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto) +cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto) cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context) @@ -36,5 +39,6 @@ cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_ha device_context broadcast_op_handle) cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory device_context gather_op_handle) +cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_executor.cc DEPS ssa_graph_executor) #cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory # device_context reduce_op_handle ) diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..bf493a3fa44e48deec734250d04b2a413c3ed9da --- /dev/null +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -0,0 +1,144 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#include + +#include "paddle/fluid/framework/details/all_reduce_op_handle.h" +#include "paddle/fluid/framework/details/container_cast.h" +#include "paddle/fluid/framework/details/reduce_and_gather.h" +#include "paddle/fluid/framework/details/variable_visitor.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace framework { +namespace details { + +#ifdef PADDLE_WITH_CUDA +AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, + const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *ctxs) + : OpHandleBase(node), + local_scopes_(local_scopes), + places_(places), + nccl_ctxs_(ctxs) { + if (nccl_ctxs_) { + for (auto &p : places_) { + this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p); + } + } +} +#else +AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, + const std::vector &local_scopes, + const std::vector &places) + : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {} +#endif + +void AllReduceOpHandle::RunImpl() { + platform::RecordEvent r("all_reduce", nullptr); + if (NoDummyInputSize() == 1) { + return; // No need to all reduce when GPU count = 1; + } else { + // Wait input done + WaitInputVarGenerated(); + auto in_var_handles = DynamicCast(this->Inputs()); + auto out_var_handles = DynamicCast(this->Outputs()); + PADDLE_ENFORCE_EQ( + in_var_handles.size(), places_.size(), + "The NoDummyInputSize should be equal to the number of places."); + PADDLE_ENFORCE_EQ( + in_var_handles.size(), out_var_handles.size(), + "The NoDummyInputSize and NoDummyOutputSize should be equal."); + + std::vector lod_tensors; + for (size_t i = 0; i < local_scopes_.size(); ++i) { + auto *s = local_scopes_[i]; + auto &local_scope = *s->FindVar(kLocalExecScopeName)->Get(); + auto &lod_tensor = + local_scope.FindVar(in_var_handles[i]->name_)->Get(); + lod_tensors.emplace_back(&lod_tensor); + PADDLE_ENFORCE_EQ(in_var_handles[i]->name_, out_var_handles[i]->name_, + "The name of input and output should be equal."); + } + + if (platform::is_gpu_place(lod_tensors[0]->place())) { +#ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr."); + int dtype = -1; + size_t numel = 0; + std::vector> all_reduce_calls; + for (size_t i = 0; i < local_scopes_.size(); ++i) { + auto &p = places_[i]; + auto &lod_tensor = *lod_tensors[i]; + void *buffer = const_cast(lod_tensor.data()); + + if (dtype == -1) { + dtype = platform::ToNCCLDataType(lod_tensor.type()); + } + + if (numel == 0) { + numel = static_cast(lod_tensor.numel()); + } + + int dev_id = boost::get(p).device; + auto &nccl_ctx = nccl_ctxs_->at(dev_id); + auto stream = nccl_ctx.stream(); + auto comm = nccl_ctx.comm_; + all_reduce_calls.emplace_back([=] { + PADDLE_ENFORCE(platform::dynload::ncclAllReduce( + buffer, buffer, numel, static_cast(dtype), + ncclSum, comm, stream)); + }); + } + this->RunAndRecordEvent([&] { + platform::NCCLGroupGuard guard; + for (auto &call : all_reduce_calls) { + call(); + } + }); +#else + PADDLE_THROW("Not compiled with CUDA"); +#endif + } else { // Special handle CPU only Operator's gradient. Like CRF + auto &trg = *this->local_scopes_[0] + ->FindVar(kLocalExecScopeName) + ->Get() + ->FindVar(out_var_handles[0]->name_) + ->GetMutable(); + + // Reduce All Tensor to trg in CPU + ReduceLoDTensor func(lod_tensors, &trg); + VisitDataType(ToDataType(lod_tensors[0]->type()), func); + + for (size_t i = 1; i < local_scopes_.size(); ++i) { + auto &scope = + *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get(); + auto &p = places_[i]; + auto *var = scope.FindVar(out_var_handles[i]->name_); + auto *dev_ctx = dev_ctxes_[p]; + + RunAndRecordEvent(p, [&trg, var, dev_ctx, p] { + auto &tensor_gpu = *var->GetMutable(); + auto &tensor_cpu = trg; + TensorCopy(tensor_cpu, p, *dev_ctx, &tensor_gpu); + }); + } + } + } +} + +std::string AllReduceOpHandle::Name() const { return "all_reduce"; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.h b/paddle/fluid/framework/details/all_reduce_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..f6ef3a1367b91b6abf8ce74a91f73056efd0f84e --- /dev/null +++ b/paddle/fluid/framework/details/all_reduce_op_handle.h @@ -0,0 +1,59 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include + +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/nccl_helper.h" +#endif + +namespace paddle { +namespace framework { +namespace details { + +struct AllReduceOpHandle : public OpHandleBase { +#ifdef PADDLE_WITH_CUDA + AllReduceOpHandle(ir::Node *node, const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *ctxs); +#else + AllReduceOpHandle(ir::Node *node, const std::vector &local_scopes, + const std::vector &places); +#endif + std::string Name() const override; + + // Delay and buffer nccl_all_reduce together can significantly increase + // performance. Disable this feature by returning false. + bool IsMultiDeviceTransfer() override { return true; }; + + protected: + void RunImpl() override; + + private: + std::vector local_scopes_; + std::vector places_; +#ifdef PADDLE_WITH_CUDA + const platform::NCCLContextMap *nccl_ctxs_; +#endif +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index d5ca061944f33939cea59a5275e691b1966194fa..1d9f1bd6e417e30f0799f0bbed1739cedb4e8fbf 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -73,6 +73,9 @@ void BroadcastOpHandle::RunImpl() { int root_id = boost::get(in_tensor.place()).device; std::vector> broadcast_calls; + int type = platform::ToNCCLDataType(in_tensor.type()); + size_t numel = static_cast(in_tensor.numel()); + for (auto out_var_handle : out_var_handles) { Variable *out_var = var_scopes.at(out_var_handle->scope_idx_) ->FindVar(out_var_handle->name_); @@ -87,13 +90,11 @@ void BroadcastOpHandle::RunImpl() { send_recv_buffer = const_cast(in_tensor.data()); out_handle = out_var_handle; } else { - send_recv_buffer = - VariableVisitor::GetMutableTensor(out_var).mutable_data( - out_var_handle->place_); + send_recv_buffer = VariableVisitor::GetMutableTensor(out_var) + .Resize(in_tensor.dims()) + .mutable_data(out_var_handle->place_); } - int type = platform::ToNCCLDataType(in_tensor.type()); - size_t numel = static_cast(in_tensor.numel()); broadcast_calls.emplace_back( [send_recv_buffer, numel, type, root_id, &nccl_ctx] { PADDLE_ENFORCE(platform::dynload::ncclBcast( diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index 629aa00cb817c4b1446e7b750ca62a7c6b1db670..fe4e733e43417977df324fde808f52b228a27d19 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -35,10 +35,13 @@ namespace details { struct BroadcastOpHandle : public OpHandleBase { public: #ifdef PADDLE_WITH_CUDA - BroadcastOpHandle(const std::vector &local_scopes, + BroadcastOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, const platform::NCCLContextMap *nccl_ctxs) - : local_scopes_(local_scopes), places_(places), nccl_ctxs_(nccl_ctxs) { + : OpHandleBase(node), + local_scopes_(local_scopes), + places_(places), + nccl_ctxs_(nccl_ctxs) { if (nccl_ctxs_) { for (auto &p_ctx : nccl_ctxs_->contexts_) { dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); @@ -46,9 +49,9 @@ struct BroadcastOpHandle : public OpHandleBase { } } #else - BroadcastOpHandle(const std::vector &local_scopes, + BroadcastOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places) - : local_scopes_(local_scopes), places_(places) {} + : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {} #endif std::string Name() const override; @@ -59,8 +62,8 @@ struct BroadcastOpHandle : public OpHandleBase { void RunImpl() override; private: - const std::vector &local_scopes_; - const std::vector &places_; + std::vector local_scopes_; + std::vector places_; #ifdef PADDLE_WITH_CUDA const platform::NCCLContextMap *nccl_ctxs_; #endif diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.cc b/paddle/fluid/framework/details/broadcast_op_handle_test.cc index c6e923ef77ff03413eefe4f26457a5322747618e..1413f7bd9ac515ae7dceee62de8f3bc74e3a2efc 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.cc @@ -96,48 +96,61 @@ struct TestBroadcastOpHandle { } param_scopes_[input_scope_idx]->Var("input"); + std::unique_ptr n( + new ir::Node("node0", ir::Node::Type::kOperation)); if (use_gpu_) { #ifdef PADDLE_WITH_CUDA - op_handle_.reset( - new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get())); + op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_, + nccl_ctxs_.get())); #else PADDLE_THROW("CUDA is not support."); #endif } else { #ifdef PADDLE_WITH_CUDA - op_handle_.reset( - new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get())); + op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_, + nccl_ctxs_.get())); #else - op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_)); + op_handle_.reset( + new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_)); #endif } - auto* in_var_handle = - new VarHandle(1, input_scope_idx, "input", gpu_list_[input_scope_idx]); + std::unique_ptr v( + new ir::Node("node1", ir::Node::Type::kVariable)); + auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input", + gpu_list_[input_scope_idx]); vars_.emplace_back(in_var_handle); op_handle_->AddInput(in_var_handle); // add dummy var - vars_.emplace_back(new DummyVarHandle()); + + std::unique_ptr v2( + new ir::Node("node2", ir::Node::Type::kVariable)); + vars_.emplace_back(new DummyVarHandle(v2.get())); DummyVarHandle* dummy_var_handle = static_cast(vars_.back().get()); - dummy_var_handle->generated_op_ = nullptr; + dummy_var_handle->ClearGeneratedOp(); op_handle_->AddInput(dummy_var_handle); for (size_t j = 0; j < gpu_list_.size(); ++j) { if (!use_gpu_) { op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); } - VarHandle* out_var_handle = new VarHandle(2, j, "out", gpu_list_[j]); + std::unique_ptr v3( + new ir::Node("node3", ir::Node::Type::kVariable)); + VarHandle* out_var_handle = + new VarHandle(v3.get(), 2, j, "out", gpu_list_[j]); vars_.emplace_back(out_var_handle); op_handle_->AddOutput(out_var_handle); } // add dummy var - vars_.emplace_back(new DummyVarHandle()); + std::unique_ptr v4( + new ir::Node("node4", ir::Node::Type::kVariable)); + vars_.emplace_back(new DummyVarHandle(v4.get())); DummyVarHandle* out_dummy_var_handle = static_cast(vars_.back().get()); - out_dummy_var_handle->generated_op_ = nullptr; + out_dummy_var_handle->ClearGeneratedOp(); op_handle_->AddOutput(out_dummy_var_handle); } diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index 91bdfe6134ffbd1404336c9d6d1222a505084b2b..8714a42162bda3d5ad12e7925fe8cc4e693f51b1 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -14,11 +14,33 @@ #pragma once +#include + namespace paddle { namespace framework { namespace details { struct BuildStrategy { + // ParallelExecutor supports two modes of ReduceStrategy, kAllReduce and + // kReduce, for CPU and GPU. If you use kAllReduce, different threads + // optimize their parameters separately. If you use kReduce, the optimizations + // of parameters are distributed to different threads. + // For example, a model has 100 parameters and is running with four threads, + // if you choose kAllReduce, every thread is to optimize 100 parameters + // separately, if you choose kReduce, every thread is to optimize 25 + // parameters. + // Of particular note is, if you use kReduce when using CPU training, + // all the parameters are shared between different threads. This feature will + // save memory. + // FIXME(zcd): The result of the two modes(kAllReduce and kReduce) maybe not + // equal for GPU. Because, the result of the different order of summing maybe + // different, for example, the result of `a+b+c+d` may be different with the + // result of `c+a+b+d`. + // For GPU, the implementation of kAllReduce and kReduce is adopted NCCL, + // so the result of kAllReduce and kReduce maybe not equal. + // For CPU, if you want to fix the order of summing to make the result + // of kAllReduce and kReduce no diff, you can add + // `FLAGS_cpu_deterministic=true` to env. enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 }; enum class GradientScaleStrategy { @@ -29,6 +51,10 @@ struct BuildStrategy { ReduceStrategy reduce_{ReduceStrategy::kAllReduce}; GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice}; + + std::string debug_graphviz_path_{""}; + + bool enable_data_balance_{false}; }; } // namespace details diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc index df05bb06333d6b964f2f5434c3d43214e5d2cb7a..b6282debdb4eb6b1f29c39e54ac4f3e2296838da 100644 --- a/paddle/fluid/framework/details/computation_op_handle.cc +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -19,9 +19,10 @@ namespace paddle { namespace framework { namespace details { -ComputationOpHandle::ComputationOpHandle(const OpDesc &op_desc, Scope *scope, +ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place) - : op_(framework::OpRegistry::CreateOp(op_desc)), + : OpHandleBase(node), + op_(framework::OpRegistry::CreateOp(*node->Op())), scope_(scope), place_(place) {} @@ -35,8 +36,8 @@ void ComputationOpHandle::RunImpl() { bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) { bool need_wait = - in_var && in_var->generated_op_ && - in_var->generated_op_->DeviceContext(place_) != dev_ctxes_[place_]; + in_var && in_var->GeneratedOp() && + in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_]; return need_wait; } diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index f048f973fdeb6cf7d1485cda8cea7d530d9ba465..d9fcd92427ef38b131b4ce782c0ada37765682db 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -28,8 +28,7 @@ namespace framework { namespace details { struct ComputationOpHandle : public OpHandleBase { public: - ComputationOpHandle(const OpDesc &op_desc, Scope *scope, - platform::Place place); + ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place); std::string Name() const override; diff --git a/paddle/fluid/framework/details/data_balance_op_handle.cc b/paddle/fluid/framework/details/data_balance_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..525d24322442ef4dd6e8c24212af61c908959b87 --- /dev/null +++ b/paddle/fluid/framework/details/data_balance_op_handle.cc @@ -0,0 +1,154 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/data_balance_op_handle.h" +#include +#include "paddle/fluid/framework/details/container_cast.h" + +namespace paddle { +namespace framework { +namespace details { + +#ifdef PADDLE_WITH_CUDA +DataBalanceOpHandle::DataBalanceOpHandle( + ir::Node *node, const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *ctxs) + : OpHandleBase(node), local_scopes_(local_scopes), places_(places) { + if (ctxs) { + for (auto &p : places_) { + this->dev_ctxes_[p] = ctxs->DevCtx(p); + } + } +} +#else +DataBalanceOpHandle::DataBalanceOpHandle( + ir::Node *node, const std::vector &local_scopes, + const std::vector &places) + : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {} +#endif + +std::string DataBalanceOpHandle::Name() const { return "data balance"; } + +std::vector> DataBalanceOpHandle::GetBalancePlan( + const std::vector &device_sizes) { + int device_num = device_sizes.size(); + int total_size = 0; + int empty_num = 0; + std::vector> size_device_vec; + size_device_vec.reserve(device_num); + for (int i = 0; i < device_num; ++i) { + if (device_sizes[i] == 0) { + ++empty_num; + } + total_size += device_sizes[i]; + size_device_vec.push_back({{device_sizes[i], i}}); + } + std::vector> res; + if (empty_num == 0) { + // No need to do data balance. + return res; + } + if (total_size < device_num) { + // No enough data. + PADDLE_THROW_EOF(); + } + std::sort(size_device_vec.begin(), size_device_vec.end(), + [](const std::array &a, const std::array &b) { + return a[0] > b[0]; + }); + int expected_device_size = total_size / device_num; + int src_idx = 0; + for (int dst_idx = device_num - empty_num; dst_idx < device_num; ++dst_idx) { + if (size_device_vec[src_idx][0] <= expected_device_size) { + ++src_idx; + PADDLE_ENFORCE_LT( + src_idx, device_num - empty_num, + "In current srategy an empty tensor should not be copy source."); + } + size_device_vec[src_idx][0] -= expected_device_size; + size_device_vec[dst_idx][0] += expected_device_size; + res.push_back({{size_device_vec[src_idx][1], size_device_vec[dst_idx][1], + expected_device_size}}); + } + return res; +} + +void DataBalanceOpHandle::RunImpl() { + PADDLE_ENFORCE_GT(places_.size(), 1, + "Data balance can only be enabled when the number of " + "places to run larger than 1."); + auto in_var_handles = DynamicCast(inputs_); + auto out_var_handles = DynamicCast(outputs_); + PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0); + PADDLE_ENFORCE_EQ( + in_var_handles.size(), out_var_handles.size(), + "The NoDummyInputSize and NoDummyOutputSize should be equal."); + int data_num = in_var_handles.size() / places_.size(); + WaitInputVarGenerated(); + std::vector> lod_tensors(data_num); + std::vector device_sizes; + for (int i = 0; i < static_cast(in_var_handles.size()); ++i) { + PADDLE_ENFORCE_EQ(in_var_handles[i]->name_, out_var_handles[i]->name_, + "The name of input and output should be equal."); + int place_idx = i / data_num; + int data_idx = i % data_num; + auto *local_scope = + local_scopes_[place_idx]->FindVar(kLocalExecScopeName)->Get(); + auto *tensor_var = local_scope->FindVar(in_var_handles[i]->name_); + PADDLE_ENFORCE(tensor_var->IsType()); + auto *tensor = tensor_var->GetMutable(); + lod_tensors[data_idx].push_back(tensor); + int ins_size = + tensor->lod().empty() ? tensor->dims()[0] : tensor->NumElements(); + if (data_idx == 0) { + device_sizes.emplace_back(ins_size); + } else { + PADDLE_ENFORCE_EQ( + ins_size, device_sizes.at(place_idx), + "All data on the same device shall have the same batch size."); + } + } + const auto &balance_plan = GetBalancePlan(device_sizes); + + for (const auto &trans : balance_plan) { + for (int data_idx = 0; data_idx < data_num; ++data_idx) { + LoDTensor *src_tensor = lod_tensors[data_idx][trans[0]]; + LoDTensor *dst_tensor = lod_tensors[data_idx][trans[1]]; + int trans_ins_size = trans[2]; + LoD src_lod = src_tensor->lod(); + int src_ins_size = + src_lod.empty() ? src_tensor->dims()[0] : src_tensor->NumElements(); + int cut_point = src_ins_size - trans_ins_size; + if (!src_lod.empty()) { + for (auto &level : src_lod) { + cut_point = level[cut_point]; + } + } + TensorCopySync(src_tensor->Slice(cut_point, src_tensor->dims()[0]), + dst_tensor->place(), dst_tensor); + src_tensor->ShareDataWith(src_tensor->Slice(0, cut_point)); + if (!src_lod.empty()) { + dst_tensor->set_lod(SliceInLevel( + src_lod, 0, src_ins_size - trans_ins_size, src_ins_size)); + src_tensor->set_lod( + SliceInLevel(src_lod, 0, 0, src_ins_size - trans_ins_size)); + } + } + } +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/data_balance_op_handle.h b/paddle/fluid/framework/details/data_balance_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..0462fb6ec713eb977f420a9cb485c0273e782496 --- /dev/null +++ b/paddle/fluid/framework/details/data_balance_op_handle.h @@ -0,0 +1,59 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/nccl_helper.h" +#endif + +namespace paddle { +namespace framework { +namespace details { + +struct DataBalanceOpHandle : public OpHandleBase { + public: +#ifdef PADDLE_WITH_CUDA + DataBalanceOpHandle(ir::Node *node, const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *ctxs); +#else + DataBalanceOpHandle(ir::Node *node, const std::vector &local_scopes, + const std::vector &places); +#endif + + std::string Name() const override; + + bool IsMultiDeviceTransfer() override { return false; }; + + protected: + void RunImpl() override; + + private: + // std::vector<(src_dev_id, dst_dev_id, trans_size)> + std::vector> GetBalancePlan( + const std::vector &batch_size_per_device); + + const std::vector local_scopes_; + const std::vector places_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/exception_holder.h b/paddle/fluid/framework/details/exception_holder.h new file mode 100644 index 0000000000000000000000000000000000000000..c97b364de1ecae21e97351196389615187932b5e --- /dev/null +++ b/paddle/fluid/framework/details/exception_holder.h @@ -0,0 +1,95 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "glog/logging.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace details { + +class ExceptionHolder { + public: + void Catch(std::exception_ptr eptr) { + try { + std::rethrow_exception(eptr); + } catch (platform::EOFException exp) { + Catch(exp); + } catch (platform::EnforceNotMet exp) { + Catch(exp); + } catch (...) { + LOG(FATAL) << "Unknown exception caught"; + } + } + + bool IsCaught() const { + std::lock_guard lock(mu_); + return exception_.get() != nullptr; + } + + void ReThrow() { + std::lock_guard lock(mu_); + switch (type_) { + case kNone: + break; + case kEnforceNotMet: { + auto e = *static_cast(exception_.get()); + throw e; + } + case kEOF: { + auto e = *static_cast(exception_.get()); + throw e; + } + } + ClearImpl(); + } + + void Clear() { + std::lock_guard lock(mu_); + ClearImpl(); + } + + private: + void ClearImpl() { + exception_.reset(); + type_ = kNone; + } + + void Catch(const platform::EnforceNotMet& exp) { + std::lock_guard lock(mu_); + exception_.reset(new platform::EnforceNotMet(exp)); + type_ = kEnforceNotMet; + } + + void Catch(const platform::EOFException& exp) { + std::lock_guard lock(mu_); + // EOFException will not cover up existing EnforceNotMet. + if (exception_.get() == nullptr) { + exception_.reset(new platform::EOFException(exp)); + type_ = kEOF; + } + } + + enum ExceptionType { kNone, kEnforceNotMet, kEOF }; + ExceptionType type_{kNone}; + + std::unique_ptr exception_; + mutable std::mutex mu_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h index e8d510ec955602b5a3f73ca06caa121886eb150b..716d674fa29bad9321fc20979775c06f26bf4679 100644 --- a/paddle/fluid/framework/details/execution_strategy.h +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -20,8 +20,9 @@ namespace details { struct ExecutionStrategy { size_t num_threads_{0}; - bool use_event_{true}; + bool use_cuda_{true}; bool allow_op_delay_{false}; + size_t num_iteration_per_drop_scope_{100}; }; } // namespace details diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc index 224e8e1f6efd7a894591ac51c929517cae7539ce..fe18b2060c5cd7e157374da53c5a985f70545ab7 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.cc +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -21,13 +21,16 @@ namespace paddle { namespace framework { namespace details { -FetchOpHandle::FetchOpHandle(FeedFetchList *data, size_t offset, +FetchOpHandle::FetchOpHandle(ir::Node *node, FeedFetchList *data, size_t offset, std::vector *local_scopes) - : data_(data), offset_(offset), local_scopes_(local_scopes) {} + : OpHandleBase(node), + data_(data), + offset_(offset), + local_scopes_(local_scopes) {} FetchOpHandle::~FetchOpHandle() { for (auto *input_var : inputs_) { - input_var->pending_ops_.erase(this); + input_var->RemoveOutput(this, this->Node()); } } @@ -67,8 +70,8 @@ void FetchOpHandle::RunImpl() { #endif } else { tensors_[i].ShareDataWith(t); - tensors_[i].set_lod(t.lod()); } + tensors_[i].set_lod(t.lod()); } this->WaitAndMergeCPUTensors(); @@ -77,8 +80,8 @@ void FetchOpHandle::RunImpl() { void FetchOpHandle::WaitInputVarGenerated(const platform::Place &place) { auto cpu_ctx = platform::DeviceContextPool::Instance().Get(place); for (auto *input : inputs_) { - if (input->generated_op_) { - input->generated_op_->RecordWaitEventOnCtx(cpu_ctx); + if (input->GeneratedOp()) { + input->GeneratedOp()->RecordWaitEventOnCtx(cpu_ctx); } } } diff --git a/paddle/fluid/framework/details/fetch_op_handle.h b/paddle/fluid/framework/details/fetch_op_handle.h index e09bdd1d3338bb175c1ddae35b53f98197b68e9a..6ce42f92d7f1e81eeafd1eb5c28ce3564a5ffebc 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.h +++ b/paddle/fluid/framework/details/fetch_op_handle.h @@ -28,7 +28,7 @@ namespace details { struct FetchOpHandle : public OpHandleBase { public: - FetchOpHandle(FeedFetchList *data, size_t offset, + FetchOpHandle(ir::Node *node, FeedFetchList *data, size_t offset, std::vector *local_scopes); ~FetchOpHandle(); diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.cc b/paddle/fluid/framework/details/fuse_vars_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..018c9bff71e553d8a3641f06f10b350453676b24 --- /dev/null +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.cc @@ -0,0 +1,51 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/fuse_vars_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { + +void FuseVarsOpHandle::RunImpl() { + WaitInputVarGenerated(place_); + + auto in_var_handles = DynamicCast(this->Inputs()); + auto out_var_handles = DynamicCast(this->Outputs()); + PADDLE_ENFORCE_EQ(in_var_handles.size(), 0); + PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), ""); + + auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); + + auto out_var_handle = out_var_handles[0]; + auto out_var = scope->Var(out_var_handle->name_); + + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({total_numel_}).mutable_data(this->place_, type_); + + int64_t s = 0; + for (size_t i = 1; i < out_var_handles.size(); ++i) { + auto out_name = out_var_handles[i]->name_; + auto out_t = scope->Var(out_name)->GetMutable(); + auto numel = this->inputs_numel_.at(out_name); + out_t->ShareDataWith(out_tensor->Slice(s, s + numel)); + s += numel; + } + this->RunAndRecordEvent([] {}); +} + +std::string FuseVarsOpHandle::Name() const { return "fuse vars"; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.h b/paddle/fluid/framework/details/fuse_vars_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..3f360c510a4fdc0caaeb15d862b217ef41b8ea6e --- /dev/null +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.h @@ -0,0 +1,65 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/framework/details/container_cast.h" +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { +namespace details { + +struct FuseVarsOpHandle : public OpHandleBase { + public: + FuseVarsOpHandle(ir::Node *node, Scope *local_scope, + const platform::Place &place, + const std::unordered_map &inputs_numel, + const std::type_index &var_type) + : OpHandleBase(node), + local_scope_(local_scope), + place_(place), + inputs_numel_(inputs_numel), + type_(var_type) { + total_numel_ = 0; + for (auto in_numel : inputs_numel) { + PADDLE_ENFORCE_GT(in_numel.second, 0); + total_numel_ += in_numel.second; + } + } + + std::string Name() const override; + + bool IsMultiDeviceTransfer() override { return false; }; + + protected: + void RunImpl() override; + + private: + Scope *local_scope_; + const platform::Place place_; + const std::unordered_map inputs_numel_; + const std::type_index type_; + int64_t total_numel_; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/gather_op_handle.cc b/paddle/fluid/framework/details/gather_op_handle.cc index 2be02304566cf5dbe348fa01fc4171990eafd158..9aae19fc73de4387186da47c55710c94d53f1b88 100644 --- a/paddle/fluid/framework/details/gather_op_handle.cc +++ b/paddle/fluid/framework/details/gather_op_handle.cc @@ -20,9 +20,10 @@ namespace paddle { namespace framework { namespace details { -GatherOpHandle::GatherOpHandle(const std::vector &local_scopes, +GatherOpHandle::GatherOpHandle(ir::Node *node, + const std::vector &local_scopes, const std::vector &places) - : local_scopes_(local_scopes), places_(places) {} + : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {} void GatherOpHandle::RunImpl() { if (places_.size() == 1) return; diff --git a/paddle/fluid/framework/details/gather_op_handle.h b/paddle/fluid/framework/details/gather_op_handle.h index d11ef8556aa8840949ca8dc7aa176413f70b9f22..d9afbc6547e18e8886c414ff150e332cfaf9b0c3 100644 --- a/paddle/fluid/framework/details/gather_op_handle.h +++ b/paddle/fluid/framework/details/gather_op_handle.h @@ -30,7 +30,7 @@ namespace details { struct GatherOpHandle : public OpHandleBase { public: - GatherOpHandle(const std::vector &local_scopes, + GatherOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places); std::string Name() const override; diff --git a/paddle/fluid/framework/details/gather_op_handle_test.cc b/paddle/fluid/framework/details/gather_op_handle_test.cc index 3cce2cc1640b3866130126424ff8fef18b8befc6..c9b94d1e1039df6ff27f9ffe225b2a50c35a5c50 100644 --- a/paddle/fluid/framework/details/gather_op_handle_test.cc +++ b/paddle/fluid/framework/details/gather_op_handle_test.cc @@ -70,6 +70,7 @@ struct TestGatherOpHandle { } void InitGatherOp(size_t input_scope_idx) { + std::vector> nodes; for (size_t j = 0; j < gpu_list_.size(); ++j) { local_scopes_.push_back(&(g_scope_.NewScope())); Scope& local_scope = local_scopes_.back()->NewScope(); @@ -81,30 +82,37 @@ struct TestGatherOpHandle { } param_scopes_[input_scope_idx]->Var("out"); - op_handle_.reset(new GatherOpHandle(local_scopes_, gpu_list_)); + nodes.emplace_back(new ir::Node("node", ir::Node::Type::kOperation)); + op_handle_.reset( + new GatherOpHandle(nodes.back().get(), local_scopes_, gpu_list_)); // add input for (size_t j = 0; j < gpu_list_.size(); ++j) { op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); - auto* in_var_handle = new VarHandle(1, j, "input", gpu_list_[j]); + nodes.emplace_back(new ir::Node("node1", ir::Node::Type::kVariable)); + auto* in_var_handle = + new VarHandle(nodes.back().get(), 1, j, "input", gpu_list_[j]); vars_.emplace_back(in_var_handle); op_handle_->AddInput(in_var_handle); } // add dummy var - vars_.emplace_back(new DummyVarHandle()); + nodes.emplace_back(new ir::Node("node2", ir::Node::Type::kVariable)); + vars_.emplace_back(new DummyVarHandle(nodes.back().get())); DummyVarHandle* in_dummy_var_handle = static_cast(vars_.back().get()); - in_dummy_var_handle->generated_op_ = nullptr; + in_dummy_var_handle->ClearGeneratedOp(); op_handle_->AddInput(in_dummy_var_handle); // add output - auto* out_var_handle = - new VarHandle(2, input_scope_idx, "out", gpu_list_[input_scope_idx]); + nodes.emplace_back(new ir::Node("node3", ir::Node::Type::kVariable)); + auto* out_var_handle = new VarHandle(nodes.back().get(), 2, input_scope_idx, + "out", gpu_list_[input_scope_idx]); vars_.emplace_back(out_var_handle); op_handle_->AddOutput(out_var_handle); // add dummy var - vars_.emplace_back(new DummyVarHandle()); + nodes.emplace_back(new ir::Node("node4", ir::Node::Type::kVariable)); + vars_.emplace_back(new DummyVarHandle(nodes.back().get())); DummyVarHandle* dummy_var_handle = static_cast(vars_.back().get()); op_handle_->AddOutput(dummy_var_handle); diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc deleted file mode 100644 index d8e711994c5dba15ce0a1c237558b121888902e3..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ /dev/null @@ -1,491 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. -#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" -#include -#include -#include "paddle/fluid/framework/details/broadcast_op_handle.h" -#include "paddle/fluid/framework/details/computation_op_handle.h" -#include "paddle/fluid/framework/details/reduce_op_handle.h" -#include "paddle/fluid/framework/details/rpc_op_handle.h" -#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" -#include "paddle/fluid/framework/op_info.h" -#include "paddle/fluid/framework/scope.h" - -#ifdef PADDLE_WITH_CUDA -#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" -#endif - -#include -#include - -DEFINE_string(ssa_graph_path, "/tmp/ssa_graph.dot", - "the ssa graph path only print with GLOG_v=10," - "default /tmp/graph.dot"); - -namespace paddle { -namespace framework { -namespace details { - -#ifdef PADDLE_WITH_CUDA -MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( - const std::vector &places, - const std::string &loss_var_name, - const std::unordered_set ¶ms, - const std::vector &local_scopes, - platform::NCCLContextMap *nccl_ctxs, const BuildStrategy &strategy) - : loss_var_name_(loss_var_name), - places_(places), - local_scopes_(local_scopes), - nccl_ctxs_(nccl_ctxs), - strategy_(strategy) { -#else -MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( - const std::vector &places, - const std::string &loss_var_name, - const std::unordered_set ¶ms, - const std::vector &local_scopes, const BuildStrategy &strategy) - : loss_var_name_(loss_var_name), - places_(places), - local_scopes_(local_scopes), - strategy_(strategy) { -#endif - for (auto &p : params) { - grad_names_.insert(GradVarName(p)); - } -} - -void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, - const OpDesc &op, - size_t place_id) const { - auto p = places_[place_id]; - auto *op_handle = result->ops_.back().get(); - op_handle->SetDeviceContext(p, - platform::DeviceContextPool::Instance().Get(p)); - - for (auto &each_var_name : op.InputArgumentNames()) { - VarHandle *var = - CreateOrGetLatestVarHandle(result, each_var_name, p, place_id); - op_handle->AddInput(var); - } - - for (auto &each_var_name : op.OutputArgumentNames()) { - CreateOpOutput(result, op_handle, each_var_name, p, place_id); - } -} - -std::vector MultiDevSSAGraphBuilder::FindDistTrainSendVars( - const ProgramDesc &program) const { - std::vector send_vars; - // since parameters are all in block 0, - // it's enough to only scan send ops in block 0 - for (auto *op : program.Block(0).AllOps()) { - // TODO(Yancey1989): use a graceful method to find send op, - // instead of the the hard code string - if (op->Type() == "send_vars") { - auto op_vars = op->InputArgumentNames(); - send_vars.reserve(send_vars.size() + - std::distance(op_vars.begin(), op_vars.end())); - send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end()); - } - } - return send_vars; -} - -std::vector MultiDevSSAGraphBuilder::FindDistTrainRecvVars( - const ProgramDesc &program) const { - std::vector recv_vars; - for (auto *op : program.Block(0).AllOps()) { - // TODO(Yancey1989): use a graceful method to find recv op, - // instead of the hard code string - if (op->Type() == "recv") { - auto op_vars = op->OutputArgumentNames(); - recv_vars.reserve(recv_vars.size() + - std::distance(op_vars.begin(), op_vars.end())); - recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end()); - } - } - return recv_vars; -} - -bool MultiDevSSAGraphBuilder::IsDistTrainOp( - const OpDesc &op, const std::vector &send_vars, - const std::vector &recv_vars) const { - if (send_vars.size() == 0 || recv_vars.size() == 0) { - return false; - } - - /** - * Check any of opvars contains `.block` and in sendvars - */ - auto checker = [](const std::vector &opvars, - const std::vector &rpc_vars) -> bool { - for (auto &var : opvars) { - // a variable name with the suffix `.block` means it's a splited - // variable by (DistributeTranspiler) - // [python/paddle/fluid/transpiler/distribute_transpiler.py] - if (var.find(".block") != std::string::npos && - std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) { - return true; - } - } - return false; - }; - - return checker(op.OutputArgumentNames(), send_vars) || - checker(op.InputArgumentNames(), recv_vars); -} - -std::unique_ptr MultiDevSSAGraphBuilder::Build( - const ProgramDesc &program) const { - std::unordered_map var_types; - for (auto *var : program.Block(0).AllVars()) { - var_types[var->Name()] = var->GetType(); - } - - auto graph = new SSAGraph(); - SSAGraph &result = *graph; - std::unordered_set og_has_been_broadcast; - - // We cannot invoke resize. It is a bug of GCC 4.8 - result.vars_ = std::vector< - std::unordered_map>>>( - places_.size()); - - // find send/recv vars so that we can place the distributed training - // realted op in the place 0 - auto send_vars = FindDistTrainSendVars(program); - auto recv_vars = FindDistTrainRecvVars(program); - - size_t cur_device_id = 0; - std::vector> var_name_on_devices; - std::vector> bcast_var_name_set; - var_name_on_devices.resize(places_.size()); - bcast_var_name_set.resize(places_.size()); - - bool is_forwarding = true; - for (auto *op : program.Block(0).AllOps()) { - if (boost::get( - op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == - static_cast(OpRole::kRPC)) { - // append rpc op if program is distributed trainer main program. - // always use the first device - CreateRPCOp(&result, *op); - } else if (IsDistTrainOp(*op, send_vars, recv_vars)) { - CreateDistTrainOp(&result, *op); - } else if (IsScaleLossOp(*op)) { - // user can customize loss@grad if not use_default_grad_scale_ - if (strategy_.gradient_scale_ != - BuildStrategy::GradientScaleStrategy::kCustomized) { - CreateScaleLossGradOp(&result); - } - is_forwarding = false; - } else { - int op_dev_id = GetOpDeviceID(var_name_on_devices, *op); - if (op_dev_id == -1) { // var on all device - CreateComputationalOps(&result, *op, places_.size()); - } else { - CreateComputationalOp(&result, *op, op_dev_id); - for (auto &var_name : op->OutputArgumentNames()) { - var_name_on_devices[op_dev_id].emplace(var_name); - } - } - if (!is_forwarding && places_.size() > 1) { - // Currently, we assume that once gradient is generated, it can be - // broadcast, and each gradient is only broadcast once. - if (static_cast(boost::get(op->GetAttr( - OpProtoAndCheckerMaker::OpRoleAttrName())) & - static_cast(OpRole::kBackward))) { - try { - auto backward_vars = - boost::get>(op->GetNullableAttr( - OpProtoAndCheckerMaker::OpRoleVarAttrName())); - - PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0); - - for (size_t i = 0; i < backward_vars.size(); i += 2) { - auto &p_name = backward_vars[i]; - auto &g_name = backward_vars[i + 1]; - VLOG(10) << "Bcast " << g_name << " for parameter " << p_name; - - switch (strategy_.reduce_) { - case BuildStrategy::ReduceStrategy::kReduce: - CreateReduceOp(&result, g_name, cur_device_id); - var_name_on_devices[cur_device_id].emplace(g_name); - bcast_var_name_set[cur_device_id].emplace(p_name); - cur_device_id = (cur_device_id + 1) % places_.size(); - break; - case BuildStrategy::ReduceStrategy::kAllReduce: - if (IsSparseGradient(var_types, g_name)) { - CreateReduceOp(&result, g_name, 0); - CreateBroadcastOp(&result, g_name, 0); - } else { - InsertNCCLAllReduceOp(&result, g_name); - } - break; - } - } - } catch (boost::bad_get e) { - } - } - } - } - } - - // Insert BCast Ops - for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) { - auto &to_bcast_set = bcast_var_name_set[dev_id]; - for (auto &bcast_name : to_bcast_set) { - CreateBroadcastOp(&result, bcast_name, dev_id); - } - } - /* - Dependency graph has been constructed. However, there are still data - harzaeds need to be handled. - */ - PolishGraphToSupportDataHazards(&result); - - /* - * Only variables should be the leaves of graph. - */ - AddOutputToLeafOps(&result); - - if (VLOG_IS_ON(10)) { - std::ofstream fout(FLAGS_ssa_graph_path); - PrintGraphviz(*graph, fout); - } - - return std::unique_ptr(graph); -} - -bool MultiDevSSAGraphBuilder::IsSparseGradient( - const std::unordered_map &var_types, - const std::string &og) const { - PADDLE_ENFORCE(var_types.count(og) != 0); - if (var_types.at(og) == proto::VarType::SELECTED_ROWS) { - return true; - } - return false; -} - -void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result, - const std::string &p_name, - size_t src_dev_id) const { -#ifdef PADDLE_WITH_CUDA - auto *op_handle = new BroadcastOpHandle(local_scopes_, places_, nccl_ctxs_); -#else - auto *op_handle = new BroadcastOpHandle(local_scopes_, places_); -#endif - - result->ops_.emplace_back(op_handle); - auto *in = result->vars_.at(src_dev_id).at(p_name).back().get(); - op_handle->AddInput(in); - - for (size_t i = 0; i < places_.size(); ++i) { - auto &vars = result->vars_.at(i).at(p_name); - auto &p = places_[i]; - auto *out_var = new VarHandle(vars.size(), i, p_name, p); - vars.emplace_back(out_var); - op_handle->AddOutput(out_var); -#ifndef ADDLE_WITH_CUDA - op_handle->SetDeviceContext(p, - platform::DeviceContextPool::Instance().Get(p)); -#endif - } -} - -void MultiDevSSAGraphBuilder::CreateComputationalOp(SSAGraph *result, - const OpDesc &op, - int dev_id) const { - result->ops_.emplace_back( - new ComputationOpHandle(op, local_scopes_[dev_id], places_[dev_id])); - CreateOpHandleIOs(result, op, dev_id); -} - -void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp( - SSAGraph *result, const std::string &og) const { -#ifdef PADDLE_WITH_CUDA - result->ops_.emplace_back( - new NCCLAllReduceOpHandle(local_scopes_, places_, *nccl_ctxs_)); - auto *op_handle = result->ops_.back().get(); - - for (size_t i = 0; i < places_.size(); ++i) { - auto &p = places_[i]; - auto &vars = result->vars_[i][og]; - PADDLE_ENFORCE(!vars.empty()); - auto &prev_grad = vars.back(); - op_handle->AddInput(prev_grad.get()); - - auto var = new VarHandle(vars.size() - 1, i, og, p); - vars.emplace_back(var); - op_handle->AddOutput(var); - } -#else - PADDLE_ENFORCE("Not implemented"); -#endif -} - -bool MultiDevSSAGraphBuilder::IsParameterGradientOnce( - const std::string &og, - std::unordered_set *og_has_been_broadcast) const { - bool is_pg_once = - grad_names_.count(og) != 0 && og_has_been_broadcast->count(og) == 0; - if (is_pg_once) { - // Insert NCCL AllReduce Op - og_has_been_broadcast->insert(og); - } - return is_pg_once; -} - -int MultiDevSSAGraphBuilder::GetOpDeviceID( - const std::vector> &var_name_on_devices, - const OpDesc &op) const { - if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) { - return -1; - } - - int var_dev_id = -1; - for (auto &var_name : op.InputArgumentNames()) { - if (var_dev_id != -1) break; - for (size_t i = 0; i < var_name_on_devices.size(); ++i) { - if (var_name_on_devices[i].count(var_name)) { - var_dev_id = static_cast(i); - break; - } - } - } - return var_dev_id; -} - -void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const { - for (size_t i = 0; i < places_.size(); ++i) { -// Insert ScaleCost OpHandle -#ifdef PADDLE_WITH_CUDA - auto *communication_dev_ctx = nccl_ctxs_->DevCtx(places_[i]); -#else - auto *communication_dev_ctx = - platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); -#endif - - auto *op_handle = - new ScaleLossGradOpHandle(local_scopes_.size(), local_scopes_[i], - places_[i], communication_dev_ctx); - result->ops_.emplace_back(op_handle); - - // FIXME: Currently ScaleLossGradOp only use device_count as scale - // factor. So it does not depend on any other operators. - // VarHandle *loss = GetVarHandle(loss_var_name, place); - // loss->pending_ops_.emplace_back(op_handle); - // op_handle->inputs_.emplace_back(loss); - - CreateOpOutput(result, op_handle, GradVarName(loss_var_name_), places_[i], - i); - } -} - -void MultiDevSSAGraphBuilder::CreateComputationalOps(SSAGraph *result, - const OpDesc &op, - size_t num_places) const { - for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) { - auto p = places_[scope_idx]; - auto s = local_scopes_[scope_idx]; - result->ops_.emplace_back(new ComputationOpHandle(op, s, p)); - CreateOpHandleIOs(result, op, scope_idx); - } -} - -VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result, - const std::string &og, - int dst_dev_id) const { -#ifdef PADDLE_WITH_CUDA - result->ops_.emplace_back( - new ReduceOpHandle(local_scopes_, places_, nccl_ctxs_)); -#else - result->ops_.emplace_back(new ReduceOpHandle(local_scopes_, places_)); -#endif - auto *op_handle = result->ops_.back().get(); - - for (size_t i = 0; i < places_.size(); ++i) { - auto &vars = result->vars_[i][og]; -#ifndef PADDLE_WITH_CUDA - auto &p = places_[i]; - op_handle->SetDeviceContext(p, - platform::DeviceContextPool::Instance().Get(p)); -#endif - PADDLE_ENFORCE(!vars.empty()); - auto &prev_grad = vars.back(); - op_handle->AddInput(prev_grad.get()); - } - auto &vars = result->vars_[dst_dev_id][og]; - auto var = - new VarHandle(vars.size() - 1, dst_dev_id, og, places_[dst_dev_id]); - vars.emplace_back(var); - op_handle->AddOutput(var); - return var; -} - -void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op, - const std::string &prev_op_name) const { - for (auto &prev_op : result->ops_) { - if (prev_op->Name() == prev_op_name) { - auto *dep_var = new DummyVarHandle(); - prev_op->AddOutput(dep_var); - result->dep_vars_.emplace(dep_var); - op->AddInput(dep_var); - } - } -} - -void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result, - const OpDesc &op) const { - CreateComputationalOp(result, op, 0); - if (op.Type() == "concat") { - ConnectOp(result, result->ops_.back().get(), "fetch_barrier"); - } -} - -void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result, - const OpDesc &op) const { - auto &p = places_[0]; - auto *s = local_scopes_[0]; - result->ops_.emplace_back(new RPCOpHandle(op, s, p, op.Type())); - - if (op.Type() == "send_barrier") { - ConnectOp(result, result->ops_.back().get(), "send_vars"); - } else if (op.Type() == "recv") { - ConnectOp(result, result->ops_.back().get(), "send_barrier"); - } else if (op.Type() == "fetch_barrier") { - ConnectOp(result, result->ops_.back().get(), "recv"); - } else if (op.Type() == "send_vars") { - // do nothing - } else { - PADDLE_THROW( - "rpc op should be in [" - "send_vars, send_barrier. recv, fetch_barrier]"); - } - - // TODO(Yancey1989): schedule rpc op on different place may - // increate throughput - CreateOpHandleIOs(result, op, 0); -} - -bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const { - return boost::get( - op.GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == - (static_cast(OpRole::kBackward) | - static_cast(OpRole::kLoss)) && - !loss_var_name_.empty(); // If loss_var is empty. This is test mode -} -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h deleted file mode 100644 index e07597dbd80889c366babe79455beb12c9eb80d9..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ /dev/null @@ -1,117 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once -#include -#include -#include - -#include "paddle/fluid/framework/details/build_strategy.h" -#include "paddle/fluid/framework/details/ssa_graph_builder.h" - -namespace paddle { -namespace platform { -class NCCLContextMap; -} - -namespace framework { -class Scope; -namespace details { - -class MultiDevSSAGraphBuilder : public SSAGraphBuilder { - public: -#ifdef PADDLE_WITH_CUDA - MultiDevSSAGraphBuilder(const std::vector &places, - const std::string &loss_var_name, - const std::unordered_set ¶ms, - const std::vector &local_scopes, - platform::NCCLContextMap *nccl_ctxs, - const BuildStrategy &strategy); -#else - MultiDevSSAGraphBuilder(const std::vector &places, - const std::string &loss_var_name, - const std::unordered_set ¶ms, - const std::vector &local_scopes, - const BuildStrategy &strategy); -#endif - - std::unique_ptr Build(const ProgramDesc &program) const override; - - private: - void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op, - size_t place_id) const; - - private: - std::string loss_var_name_; - const std::vector &places_; - const std::vector &local_scopes_; - std::unordered_set grad_names_; - -#ifdef PADDLE_WITH_CUDA - platform::NCCLContextMap *nccl_ctxs_; -#endif - - bool IsScaleLossOp(const OpDesc &op) const; - - void CreateRPCOp(SSAGraph *result, const OpDesc &op) const; - void CreateDistTrainOp(SSAGraph *result, const OpDesc &op) const; - - /** - * Is this operator as the end-point operator before/after send operator. - */ - bool IsDistTrainOp(const OpDesc &op, - const std::vector &send_vars, - const std::vector &recv_vars) const; - - std::vector FindDistTrainSendVars( - const ProgramDesc &program) const; - - std::vector FindDistTrainRecvVars( - const ProgramDesc &program) const; - - void ConnectOp(SSAGraph *result, OpHandleBase *op, - const std::string &prev_op_name) const; - - void CreateComputationalOps(SSAGraph *result, const OpDesc &op, - size_t num_places) const; - - void CreateScaleLossGradOp(SSAGraph *result) const; - VarHandle *CreateReduceOp(SSAGraph *result, const std::string &og, - int dst_dev_id) const; - void CreateComputationalOp(SSAGraph *result, const OpDesc &op, - int dev_id) const; - - bool IsParameterGradientOnce( - const std::string &og, - std::unordered_set *og_has_been_broadcast) const; - - int GetOpDeviceID( - const std::vector> &var_name_on_devices, - const OpDesc &op) const; - - void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const; - - void CreateBroadcastOp(SSAGraph *result, const std::string &p_name, - size_t src_dev_id) const; - - bool IsSparseGradient( - const std::unordered_map &var_types, - const std::string &og) const; - - private: - BuildStrategy strategy_; -}; -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..c9c255864a2477ed29873f8521acce37fa928c06 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc @@ -0,0 +1,94 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" +#include +#include "paddle/fluid/framework/ir/graph.h" + +namespace paddle { +namespace framework { +namespace details { + +bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const { + std::unordered_map pending_ops; + std::unordered_set pending_vars; + std::unordered_set ready_vars; + std::unordered_set ready_ops; + + auto insert_pending_var = [&](VarHandleBase *var) { + pending_vars.insert(var); + if (var->GeneratedOp() == nullptr) { + ready_vars.emplace(var); + } + }; + + for (auto &var_map : graph->Get(kGraphVars)) { + for (auto &name_pair : var_map) { + for (auto &version_pair : name_pair.second) { + insert_pending_var(version_pair.get()); + } + } + } + + for (auto &var : graph->Get(kGraphDepVars)) { + insert_pending_var(var.get()); + } + + for (auto &op : graph->Get(kGraphOps)) { + if (op->Inputs().empty()) { + ready_ops.insert(op.get()); + } else { + pending_ops.insert({op.get(), op.get()->NoDupInputSize()}); + } + } + + auto run_all_ops = [&](std::unordered_set &set) { + for (auto *op : set) { + for (auto out : op->Outputs()) { + ready_vars.emplace(out); + } + } + set.clear(); + }; + + while (!pending_vars.empty()) { + run_all_ops(ready_ops); + + if (ready_vars.empty()) { + return false; + } + + for (auto ready_var : ready_vars) { + pending_vars.erase(ready_var); + for (auto *op : ready_var->PendingOps()) { + auto &deps = --pending_ops[op]; + if (deps == 0) { + ready_ops.insert(op); + } + } + } + ready_vars.clear(); + } + return true; +} +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(multi_devices_check_pass, + paddle::framework::details::SSAGraghBuilderWithChecker) + .RequireGraphAttr(paddle::framework::details::kGraphVars) + .RequireGraphAttr(paddle::framework::details::kGraphDepVars) + .RequireGraphAttr(paddle::framework::details::kGraphOps) + .RequireGraphAttr(paddle::framework::details::kShardedVarDevice); diff --git a/paddle/fluid/framework/details/multi_devices_graph_check_pass.h b/paddle/fluid/framework/details/multi_devices_graph_check_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..1e2b1867c376956d7d2dac465c13e2f3f64ba7eb --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_check_pass.h @@ -0,0 +1,38 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/framework/details/multi_devices_helper.h" + +#include + +namespace paddle { +namespace framework { +namespace details { + +class SSAGraghBuilderWithChecker : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override { + PADDLE_ENFORCE(IsValidGraph(graph.get())); + return graph; + } + + bool IsValidGraph(const ir::Graph* graph) const; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..c5a13e7e1f45e1eb9b4271880630c52d30022f4b --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -0,0 +1,846 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/details/all_reduce_op_handle.h" +#include "paddle/fluid/framework/details/broadcast_op_handle.h" +#include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/data_balance_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_graph_pass.h" +#include "paddle/fluid/framework/details/reduce_op_handle.h" +#include "paddle/fluid/framework/details/rpc_op_handle.h" +#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/ir/node.h" +#include "paddle/fluid/framework/op_info.h" +#include "paddle/fluid/framework/scope.h" + +namespace paddle { +namespace framework { +namespace details { +namespace { +void PolishGraphToSupportDataHazards(ir::Graph *graph) { + for (auto &var_map : graph->Get(kGraphVars)) { + for (auto &name_pair : var_map) { + if (name_pair.second.size() <= 1) { + continue; + } + auto it_new = name_pair.second.rbegin(); + auto it_old = name_pair.second.rbegin(); + ++it_old; + for (; it_old != name_pair.second.rend(); it_new = it_old, ++it_old) { + OpHandleBase *write_op = (*it_new)->GeneratedOp(); + const auto &read_ops = (*it_old)->PendingOps(); + + for (auto *read_op : read_ops) { + // Manually add a dependency var from read_op to write_op; + if (read_op == write_op) { + // Read Write is the same op. + continue; + } + bool has_dep = false; + for (auto *r_out : read_op->Outputs()) { + for (auto *w_in : write_op->Inputs()) { + if (r_out->Node() == w_in->Node()) { + has_dep = true; + break; + } + } + } + if (has_dep) continue; + + auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); + read_op->AddOutput(dep_var); + write_op->AddInput(dep_var); + graph->Get(kGraphDepVars).emplace(dep_var); + } + } + } + } +} + +VarHandle *CreateOrGetLatestVarHandle(ir::Graph *graph, ir::Node *node, + const platform::Place &place, + size_t place_offset) { + auto &var_holders = graph->Get(kGraphVars)[place_offset]; + auto &var_holder = var_holders[node->Name()]; + VarHandle *var = nullptr; + if (var_holder.empty()) { + if (node->Var()) { + var = new VarHandle(graph->CreateVarNode(node->Var()), 0, place_offset, + node->Name(), place); + } else { + var = new VarHandle( + graph->CreateEmptyNode(node->Name(), ir::Node::Type::kVariable), 0, + place_offset, node->Name(), place); + } + var_holder.emplace_back(var); + } else { + var = var_holder.rbegin()->get(); + } + return var; +} + +void CreateOpOutput(ir::Graph *graph, OpHandleBase *op_handle, + ir::Node *new_node, const platform::Place &place, + size_t place_offset) { + auto &vars = + graph->Get(kGraphVars)[place_offset][new_node->Name()]; + size_t version = vars.size(); + auto var = + new VarHandle(new_node, version, place_offset, new_node->Name(), place); + vars.emplace_back(var); + op_handle->AddOutput(var); +} + +void AddOutputToLeafOps(ir::Graph *graph) { + for (auto &op : graph->Get(kGraphOps)) { + if (!op->Outputs().empty()) { + continue; + } + auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar()); + graph->Get(kGraphDepVars).emplace(dummy_leaf); + op->AddOutput(dummy_leaf); + } +} +} // namespace + +static const char kLossVarName[] = "loss_var_name"; +static const char kPlaces[] = "places"; +static const char kParams[] = "params"; +static const char kLocalScopes[] = "local_scopes"; +static const char kStrategy[] = "strategy"; + +void MultiDevSSAGraphBuilder::Init() const { + loss_var_name_ = Get(kLossVarName); + places_ = Get>(kPlaces); + local_scopes_ = Get>(kLocalScopes); + strategy_ = Get(kStrategy); +#ifdef PADDLE_WITH_CUDA + nccl_ctxs_ = &Get("nccl_ctxs"); +#endif + + for (auto &p : Get>(kParams)) { + grad_names_.insert(GradVarName(p)); + } + balance_vars_.resize(places_.size(), 0); + if (strategy_.enable_data_balance_ && places_.size() == 1) { + LOG(WARNING) << "It is no need to enable data balance when there is only " + "one place. enable_data_balance is set to False."; + strategy_.enable_data_balance_ = false; + } +} + +void MultiDevSSAGraphBuilder::CreateOpHandleIOs(ir::Graph *result, + ir::Node *node, + size_t place_id) const { + auto p = places_[place_id]; + auto *op_handle = result->Get(kGraphOps).back().get(); + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); + + for (ir::Node *input : node->inputs) { + VarHandle *var = CreateOrGetLatestVarHandle(result, input, p, place_id); + op_handle->AddInput(var); + } + + for (ir::Node *output : node->outputs) { + ir::Node *new_node = nullptr; + if (output->Var()) { + new_node = result->CreateVarNode(output->Var()); + } else { + new_node = + result->CreateEmptyNode(output->Name(), ir::Node::Type::kVariable); + } + CreateOpOutput(result, op_handle, new_node, p, place_id); + } +} + +std::vector MultiDevSSAGraphBuilder::FindDistTrainSendVars( + const std::vector &nodes) const { + std::vector send_vars; + // since parameters are all in block 0, + // it's enough to only scan send ops in block 0 + for (auto &node : nodes) { + OpDesc *op = node->Op(); + // TODO(Yancey1989): use a graceful method to find send op, + // instead of the the hard code string + if (op->Type() == "send") { + auto op_vars = op->InputArgumentNames(); + send_vars.reserve(send_vars.size() + + std::distance(op_vars.begin(), op_vars.end())); + send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end()); + } + } + return send_vars; +} + +std::vector MultiDevSSAGraphBuilder::FindDistTrainRecvVars( + const std::vector &nodes) const { + std::vector recv_vars; + for (auto &node : nodes) { + OpDesc *op = node->Op(); + // TODO(Yancey1989): use a graceful method to find recv op, + // instead of the hard code string + if (op->Type() == "recv") { + auto op_vars = op->OutputArgumentNames(); + recv_vars.reserve(recv_vars.size() + + std::distance(op_vars.begin(), op_vars.end())); + recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end()); + } + } + return recv_vars; +} + +bool MultiDevSSAGraphBuilder::IsDistTrainOp( + ir::Node *node, const std::vector &send_vars, + const std::vector &recv_vars) const { + if (send_vars.size() == 0 || recv_vars.size() == 0) { + return false; + } + + /** + * Check any of opvars contains `.block` and in sendvars + */ + auto checker = [](const std::vector &opvars, + const std::vector &rpc_vars) -> bool { + for (auto &var : opvars) { + // a variable name with the suffix `.block` means it's a splited + // variable by (DistributeTranspiler) + // [python/paddle/fluid/transpiler/distribute_transpiler.py] + if (var.find(".block") != std::string::npos && + std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) { + return true; + } + } + return false; + }; + + std::vector input_var_names; + std::vector output_var_names; + for (ir::Node *input : node->inputs) { + input_var_names.push_back(input->Name()); + } + for (ir::Node *output : node->outputs) { + output_var_names.push_back(output->Name()); + } + + return checker(output_var_names, send_vars) || + checker(input_var_names, recv_vars); +} + +size_t MultiDevSSAGraphBuilder::GetAppropriateDeviceID( + const std::vector &var_names) const { + int64_t numel_sum = 0; + for (auto var_name : var_names) { + if (all_vars_.find(var_name) == all_vars_.end()) continue; + auto var_desc = all_vars_.at(var_name); + PADDLE_ENFORCE_NOT_NULL(var_desc); + auto dim = framework::make_ddim(var_desc->GetShape()); + int64_t numel = framework::product(dim); + PADDLE_ENFORCE_GT(numel, 0); + numel_sum += numel; + } + + auto smallest = + std::min_element(std::begin(balance_vars_), std::end(balance_vars_)); + size_t dev_id = + static_cast(std::distance(std::begin(balance_vars_), smallest)); + balance_vars_[dev_id] += numel_sum; + return dev_id; +} + +// Topology sort the graph nodes from inputs to outputs. +// Since SSAGraphBuilder depends on forward/backward nodes to assign devices +// to parameter/gradients before optimizer ops, topo sort is insufficient. ( +// some optimizer ops might not depend on any nodes), we manually move all +// optimizer nodes after last backward nodes. +// However, the assumption by SSAGraphBuilder should be relaxed in the future. +std::vector SortOpsAndDelayOptimizeOp(const ir::Graph &graph) { + std::vector ret = ir::TopologySortOperations(graph); + size_t last_backward = 0; + for (size_t i = 0; i < ret.size(); ++i) { + if (boost::get( + ret[i]->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == + static_cast(OpRole::kBackward)) { + last_backward = i; + } + } + + std::vector optimize_ops; + std::vector sorted_ret; + for (size_t i = 0; i < ret.size(); ++i) { + if (i < last_backward) { + if (boost::get(ret[i]->Op()->GetAttr( + OpProtoAndCheckerMaker::OpRoleAttrName())) == + static_cast(OpRole::kOptimize)) { + optimize_ops.push_back(ret[i]); + } else { + sorted_ret.push_back(ret[i]); + } + } else if (i == last_backward) { + sorted_ret.push_back(ret[i]); + // Verify that no operations before optimize ops depends on optimize ops. + std::unordered_set optimize_set(optimize_ops.begin(), + optimize_ops.end()); + for (ir::Node *n : sorted_ret) { + for (ir::Node *in : n->inputs) { + for (ir::Node *pre_n : in->inputs) { + PADDLE_ENFORCE(optimize_set.find(pre_n) == optimize_set.end(), + "optimize operations cannot be depended by forward " + "or backward node %s -> %s", + pre_n->Name(), n->Name()); + } + } + } + sorted_ret.insert(sorted_ret.end(), optimize_ops.begin(), + optimize_ops.end()); + } else { + sorted_ret.push_back(ret[i]); + } + } + return sorted_ret; +} + +std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( + std::unique_ptr graph) const { + Init(); + // Give the topology sort order and rebuild the graph structure. + std::vector sorted_ops = SortOpsAndDelayOptimizeOp(*graph); + auto nodes = graph->ReleaseNodes(); + ir::Graph &result = *graph; + + for (auto &node : nodes) { + if (node->NodeType() == ir::Node::Type::kVariable && node->Var()) { + all_vars_.emplace(node->Name(), node->Var()); + } + } + std::unordered_set og_has_been_broadcast; + + // We cannot invoke resize. It is a bug of GCC 4.8 + result.Set(kGraphVars, new GraphVars(places_.size())); + result.Set(kGraphDepVars, new GraphDepVars); + result.Set(kGraphOps, new GraphOps); + result.Set(kShardedVarDevice, new ShardedVarDevice); + + // find send/recv vars so that we can place the distributed training + // related op in the place 0 + auto send_vars = FindDistTrainSendVars(sorted_ops); + auto recv_vars = FindDistTrainRecvVars(sorted_ops); + + std::vector> bcast_var_name_set; + bcast_var_name_set.resize(places_.size()); + + size_t cur_device_id = 0; + bool is_forwarding = true; + + for (ir::Node *node : sorted_ops) { + if (boost::get( + node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == + static_cast(OpRole::kRPC)) { + CreateRPCOp(&result, node); + } else if (IsDistTrainOp(node, send_vars, recv_vars)) { + CreateDistTrainOp(&result, node); + } else if (IsScaleLossOp(node)) { + // user can customize loss@grad if not use_default_grad_scale_ + if (strategy_.gradient_scale_ != + BuildStrategy::GradientScaleStrategy::kCustomized) { + // TODO(paddle-dev): Why is there no input for this op_handle? + auto loss_grad_name = node->Op()->OutputArgumentNames()[0]; + CreateScaleLossGradOp(&result, loss_grad_name); + } + // This assumes the backward generating code will ensure IsScaleLossOp + // is true only for the op that scale the final scalar loss. + // It also assumes backward op will always follow the forward op in + // the block. + is_forwarding = false; + } else { + int op_dev_id = GetOpDeviceID(result, node); + if (op_dev_id != -1) { // This op only runs on one specific device. + CreateComputationalOp(&result, node, op_dev_id); + for (ir::Node *n : node->outputs) { + graph->Get(kShardedVarDevice) + .emplace(n->Name(), op_dev_id); + } + } else { + // This op runs on all devices, and its output may have parameter's + // gradients. + // TODO(paddle-dev): Why is so special about "read" op? + if (node->Op()->Type() == "read" && strategy_.enable_data_balance_) { + node->Op()->SetAttr("throw_eof_exp", false); + CreateComputationalOps(&result, node, places_.size()); + const auto &data_var_names = node->Op()->Output("Out"); + InsertDataBalanceOp(&result, data_var_names); + } else { + CreateComputationalOps(&result, node, places_.size()); + } + + if (!is_forwarding && places_.size() > 1) { + // Currently, we assume that once gradient is generated, it can be + // broadcast, and each gradient is only broadcast once. + if (static_cast(boost::get(node->Op()->GetAttr( + OpProtoAndCheckerMaker::OpRoleAttrName())) & + static_cast(OpRole::kBackward))) { + try { + auto backward_vars = boost::get>( + node->Op()->GetNullableAttr( + OpProtoAndCheckerMaker::OpRoleVarAttrName())); + + PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0); + + for (size_t i = 0; i < backward_vars.size(); i += 2) { + auto &p_name = backward_vars[i]; + auto &g_name = backward_vars[i + 1]; + VLOG(10) << "Bcast " << g_name << " for parameter " << p_name; + + switch (strategy_.reduce_) { + case BuildStrategy::ReduceStrategy::kReduce: + cur_device_id = GetAppropriateDeviceID({g_name}); + CreateReduceOp(&result, g_name, cur_device_id); + graph->Get(kShardedVarDevice) + .emplace(g_name, cur_device_id); + bcast_var_name_set[cur_device_id].emplace(p_name); + break; + case BuildStrategy::ReduceStrategy::kAllReduce: + if (IsSparseGradient(g_name)) { + CreateReduceOp(&result, g_name, 0); + CreateBroadcastOp(&result, g_name, 0); + } else { + InsertAllReduceOp(&result, g_name); + } + break; + default: + LOG(FATAL) << "Unknown reduce strategy "; + break; + } + } + } catch (boost::bad_get e) { + } + } + } + } + } + } + + bool use_gpu = false; +#ifdef PADDLE_WITH_CUDA + use_gpu = nccl_ctxs_ != nullptr; +#endif + + if (use_gpu || + strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) { + // Insert BCast Ops + for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) { + auto &to_bcast_set = bcast_var_name_set[dev_id]; + for (auto &bcast_name : to_bcast_set) { + CreateBroadcastOp(&result, bcast_name, dev_id); + } + } + } + /* + Dependency graph has been constructed. However, there are still data + hazards need to be handled. + */ + PolishGraphToSupportDataHazards(&result); + + /* + * Only variables should be the leaves of graph. + */ + AddOutputToLeafOps(&result); + PADDLE_ENFORCE(!ir::HasCircle(result)); + return graph; +} + +bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const { + PADDLE_ENFORCE(all_vars_.count(og) != 0); + if (all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS) { + return true; + } + return false; +} + +void MultiDevSSAGraphBuilder::SetCommunicationContext( + OpHandleBase *op_handle, const platform::Place &p) const { +#ifdef PADDLE_WITH_CUDA + if (nccl_ctxs_ == nullptr) { + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); + } +#else + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); +#endif +} + +void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result, + const std::string &p_name, + size_t src_dev_id) const { +#ifdef PADDLE_WITH_CUDA + auto *op_handle = new BroadcastOpHandle( + result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation), + local_scopes_, places_, nccl_ctxs_); +#else + auto *op_handle = new BroadcastOpHandle( + result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation), + local_scopes_, places_); +#endif + result->Get(kGraphOps).emplace_back(op_handle); + + auto *in = + result->Get(kGraphVars).at(src_dev_id).at(p_name).back().get(); + op_handle->AddInput(in); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + SetCommunicationContext(op_handle, p); + auto &vars = result->Get(kGraphVars).at(i).at(p_name); + auto *out_var = new VarHandle( + result->CreateEmptyNode(p_name, ir::Node::Type::kVariable), vars.size(), + i, p_name, p); + vars.emplace_back(out_var); + op_handle->AddOutput(out_var); + } +} + +void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result, + ir::Node *node, + int dev_id) const { + result->Get(kGraphOps).emplace_back( + new ComputationOpHandle(result->CreateOpNode(node->Op()), + local_scopes_[dev_id], places_[dev_id])); + CreateOpHandleIOs(result, node, dev_id); +} + +void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, + const std::string &og) const { +#ifdef PADDLE_WITH_CUDA + result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( + result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), + local_scopes_, places_, nccl_ctxs_)); +#else + result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( + result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), + local_scopes_, places_)); +#endif + auto *op_handle = result->Get(kGraphOps).back().get(); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + SetCommunicationContext(op_handle, p); + auto &vars = result->Get(kGraphVars)[i][og]; + PADDLE_ENFORCE(!vars.empty()); + auto &prev_grad = vars.back(); + op_handle->AddInput(prev_grad.get()); + + auto var = + new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable), + vars.size(), i, og, p); + vars.emplace_back(var); + op_handle->AddOutput(var); + } +} + +void MultiDevSSAGraphBuilder::InsertDataBalanceOp( + ir::Graph *result, const std::vector &datas) const { +#ifdef PADDLE_WITH_CUDA + result->Get(kGraphOps).emplace_back(new DataBalanceOpHandle( + result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation), + local_scopes_, places_, nccl_ctxs_)); +#else + result->Get(kGraphOps).emplace_back(new DataBalanceOpHandle( + result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation), + local_scopes_, places_)); +#endif + auto *op_handle = result->Get(kGraphOps).back().get(); + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + SetCommunicationContext(op_handle, p); + for (const std::string &d_name : datas) { + auto &vars = result->Get(kGraphVars)[i][d_name]; + PADDLE_ENFORCE(!vars.empty()); + op_handle->AddInput(vars.back().get()); + auto var = new VarHandle( + result->CreateEmptyNode(d_name, ir::Node::Type::kVariable), + vars.size(), i, d_name, p); + vars.emplace_back(var); + op_handle->AddOutput(var); + } + } +} + +bool MultiDevSSAGraphBuilder::IsParameterGradientOnce( + const std::string &og, + std::unordered_set *og_has_been_broadcast) const { + bool is_pg_once = + grad_names_.count(og) != 0 && og_has_been_broadcast->count(og) == 0; + if (is_pg_once) { + // Insert NCCL AllReduce Op + og_has_been_broadcast->insert(og); + } + return is_pg_once; +} + +int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph, + ir::Node *node) const { + if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) { + return -1; + } + int op_role = boost::get( + node->Op()->GetAttr(framework::OpProtoAndCheckerMaker::OpRoleAttrName())); + if (op_role != static_cast(framework::OpRole::kOptimize)) { + return -1; + } + auto param_grad = boost::get>( + node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName())); + + PADDLE_ENFORCE_EQ(param_grad.size(), 2U); + int dev_id = GetVarDeviceID(graph, param_grad[1]); + PADDLE_ENFORCE_NE(dev_id, -1, "dev_id should not be -1.[%s, %s, %s]", + node->Op()->Type(), param_grad[0], param_grad[1]); + return dev_id; +} + +int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph, + const std::string &varname) const { + auto &sharded_var_device = graph.Get(kShardedVarDevice); + auto got = sharded_var_device.find(varname); + return got == sharded_var_device.end() ? -1 : got->second; +} + +void MultiDevSSAGraphBuilder::CreateScaleLossGradOp( + ir::Graph *result, const std::string &loss_grad_name) const { + for (size_t i = 0; i < places_.size(); ++i) { +// Insert ScaleCost OpHandle +#ifdef PADDLE_WITH_CUDA + auto *communication_dev_ctx = + nccl_ctxs_ ? nccl_ctxs_->DevCtx(places_[i]) + : platform::DeviceContextPool::Instance().Get(places_[i]); +#else + auto *communication_dev_ctx = + platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); +#endif + auto *op_handle = new ScaleLossGradOpHandle( + result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation), + local_scopes_.size(), local_scopes_[i], places_[i], + communication_dev_ctx); + result->Get(kGraphOps).emplace_back(op_handle); + + // FIXME: Currently ScaleLossGradOp only use device_count as scale + // factor. So it does not depend on any other operators. + // VarHandle *loss = GetVarHandle(loss_var_name, place); + // loss->pending_ops_.emplace_back(op_handle); + // op_handle->inputs_.emplace_back(loss); + + CreateOpOutput( + result, op_handle, + result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable), + places_[i], i); + } +} + +void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result, + ir::Node *node, + size_t num_places) const { + for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) { + auto p = places_[scope_idx]; + auto s = local_scopes_[scope_idx]; + result->Get(kGraphOps).emplace_back( + new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p)); + CreateOpHandleIOs(result, node, scope_idx); + } +} + +VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, + const std::string &og, + int dst_dev_id) const { +#ifdef PADDLE_WITH_CUDA + result->Get(kGraphOps).emplace_back(new ReduceOpHandle( + result->CreateEmptyNode("reduce", ir::Node::Type::kOperation), + local_scopes_, places_, nccl_ctxs_)); +#else + result->Get(kGraphOps).emplace_back(new ReduceOpHandle( + result->CreateEmptyNode("reduce", ir::Node::Type::kOperation), + local_scopes_, places_)); +#endif + auto *op_handle = result->Get(kGraphOps).back().get(); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + SetCommunicationContext(op_handle, p); + auto &vars = result->Get(kGraphVars)[i][og]; + PADDLE_ENFORCE(!vars.empty()); + auto &prev_grad = vars.back(); + op_handle->AddInput(prev_grad.get()); + } + auto &vars = result->Get(kGraphVars)[dst_dev_id][og]; + auto var = + new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable), + vars.size(), dst_dev_id, og, places_[dst_dev_id]); + vars.emplace_back(var); + op_handle->AddOutput(var); + return var; +} + +// Find the first occurence of `prev_op_name` and make current `op` depend +// on it. +void MultiDevSSAGraphBuilder::ConnectOp(ir::Graph *result, OpHandleBase *op, + const std::string &prev_op_name) const { + for (auto &prev_op : result->Get(kGraphOps)) { + if (prev_op->Name() == prev_op_name) { + auto *dep_var = new DummyVarHandle(result->CreateControlDepVar()); + prev_op->AddOutput(dep_var); + result->Get(kGraphDepVars).emplace(dep_var); + op->AddInput(dep_var); + } + } +} + +void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, + ir::Node *node) const { + int op_dev_id = -1; + std::vector input_var_names; + std::vector output_var_names; + for (ir::Node *input : node->inputs) { + input_var_names.push_back(input->Name()); + } + for (ir::Node *output : node->outputs) { + output_var_names.push_back(output->Name()); + } + + if (node->Op()->Type() == "split_byref" || + node->Op()->Type() == "split_selected_rows") { + // TODO(paddle-dev): getting the first var is not safe. + op_dev_id = GetVarDeviceID(*result, input_var_names[0]); + if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) { + op_dev_id = GetAppropriateDeviceID(input_var_names); + for (auto &varname : input_var_names) { + result->Get(kShardedVarDevice) + .emplace(varname, op_dev_id); + } + } + for (auto &varname : output_var_names) { + result->Get(kShardedVarDevice) + .emplace(varname, op_dev_id); + } + } else if (node->Op()->Type() == "concat") { + op_dev_id = GetVarDeviceID(*result, input_var_names[0]); + for (auto &varname : output_var_names) { + result->Get(kShardedVarDevice) + .emplace(varname, op_dev_id); + } + } else { + PADDLE_ENFORCE( + "the distribute training related op should be in [split_byref, " + "concat]."); + } + + PADDLE_ENFORCE(op_dev_id != -1, + "can not find right place for distributed op: %s", + node->Op()->Type()); + + CreateComputationalOp(result, node, op_dev_id); + if (node->Op()->Type() == "concat") { + ConnectOp(result, result->Get(kGraphOps).back().get(), + "fetch_barrier"); + } +} + +// Create RPC related op handles that connects its in ops and out ops. +void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, + ir::Node *node) const { + int op_dev_id = -1; + if (node->Op()->Type() == "send") { + // TODO(paddle-dev): getting the first var is not safe. + op_dev_id = GetVarDeviceID(*result, node->inputs[0]->Name()); + PADDLE_ENFORCE(!ir::IsControlDepVar(*node->inputs[0]), + "This hack no longer holds, please fix."); + // the variable name which contains .block means it was splited by + // split_byref op + // so that we can balance the variable blocks to all the pserver + // instances. + if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce && + node->inputs[0]->Name().find(".block") == std::string::npos) { + std::vector input_var_names; + for (ir::Node *n : node->inputs) { + input_var_names.push_back(n->Name()); + } + op_dev_id = GetAppropriateDeviceID(input_var_names); + for (auto &varname : input_var_names) { + result->Get(kShardedVarDevice) + .emplace(varname, op_dev_id); + } + } + } else if (node->Op()->Type() == "recv") { + std::vector output_var_names; + for (ir::Node *n : node->outputs) { + output_var_names.push_back(n->Name()); + } + op_dev_id = GetAppropriateDeviceID(output_var_names); + for (auto &varname : output_var_names) { + result->Get(kShardedVarDevice) + .emplace(varname, op_dev_id); + } + } else { + // send_barrier and fetch_barrier op can be scheduled on device 0 + op_dev_id = 0; + } + + PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s", + node->Op()->Type()); + + result->Get(kGraphOps).emplace_back(new RPCOpHandle( + result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id], + node->Op()->Type(), places_[op_dev_id])); + + // TODO(panyx0718): This might not be needed anymore. + if (node->Op()->Type() == "send_barrier") { + ConnectOp(result, result->Get(kGraphOps).back().get(), "send"); + } else if (node->Op()->Type() == "recv") { + ConnectOp(result, result->Get(kGraphOps).back().get(), + "send_barrier"); + } else if (node->Op()->Type() == "fetch_barrier") { + ConnectOp(result, result->Get(kGraphOps).back().get(), "recv"); + } else if (node->Op()->Type() == "send") { + // do nothing + } else { + PADDLE_THROW( + "rpc op should be in [" + "send, send_barrier. recv, fetch_barrier]"); + } + + CreateOpHandleIOs(result, node, op_dev_id); +} + +bool MultiDevSSAGraphBuilder::IsScaleLossOp(ir::Node *node) const { + return boost::get( + node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == + (static_cast(OpRole::kBackward) | + static_cast(OpRole::kLoss)) && + !loss_var_name_.empty(); // If loss_var is empty. This is test mode +} +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(multi_devices_pass, + paddle::framework::details::MultiDevSSAGraphBuilder) + .RequirePassAttr(paddle::framework::details::kLossVarName) + .RequirePassAttr(paddle::framework::details::kPlaces) + .RequirePassAttr(paddle::framework::details::kParams) + .RequirePassAttr(paddle::framework::details::kLocalScopes) + .RequirePassAttr(paddle::framework::details::kStrategy); diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.h b/paddle/fluid/framework/details/multi_devices_graph_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..7a6f238f9cf7af18cb10ea271e453fec1902c833 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.h @@ -0,0 +1,115 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include +#include + +#include "paddle/fluid/framework/details/build_strategy.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/ir/graph.h" + +namespace paddle { +namespace platform { +class NCCLContextMap; +} + +namespace framework { +class Scope; +namespace details { + +class MultiDevSSAGraphBuilder : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; + + private: + void CreateOpHandleIOs(ir::Graph *result, ir::Node *node, + size_t device_id) const; + void Init() const; + + private: + mutable std::string loss_var_name_; + mutable std::vector places_; + mutable std::vector local_scopes_; + mutable std::unordered_set grad_names_; + +#ifdef PADDLE_WITH_CUDA + mutable platform::NCCLContextMap *nccl_ctxs_; +#endif + + int GetVarDeviceID(const ir::Graph &graph, const std::string &varname) const; + + bool IsScaleLossOp(ir::Node *node) const; + + void CreateRPCOp(ir::Graph *result, ir::Node *node) const; + void CreateDistTrainOp(ir::Graph *result, ir::Node *node) const; + + /** + * Is this operator as the end-point operator before/after send operator. + */ + bool IsDistTrainOp(ir::Node *node, const std::vector &send_vars, + const std::vector &recv_vars) const; + + std::vector FindDistTrainSendVars( + const std::vector &nodes) const; + + std::vector FindDistTrainRecvVars( + const std::vector &nodes) const; + + void ConnectOp(ir::Graph *result, OpHandleBase *op, + const std::string &prev_op_name) const; + + void CreateComputationalOps(ir::Graph *result, ir::Node *node, + size_t num_places) const; + + void CreateScaleLossGradOp(ir::Graph *result, + const std::string &loss_grad_name) const; + + VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og, + int dst_dev_id) const; + void CreateComputationalOp(ir::Graph *result, ir::Node *node, + int dev_id) const; + + bool IsParameterGradientOnce( + const std::string &og, + std::unordered_set *og_has_been_broadcast) const; + + int GetOpDeviceID(const ir::Graph &graph, ir::Node *node) const; + + void InsertAllReduceOp(ir::Graph *result, const std::string &og) const; + + void InsertDataBalanceOp(ir::Graph *result, + const std::vector &datas) const; + + void CreateBroadcastOp(ir::Graph *result, const std::string &p_name, + size_t src_dev_id) const; + + bool IsSparseGradient(const std::string &og) const; + + size_t GetAppropriateDeviceID( + const std::vector &var_names) const; + + private: + mutable BuildStrategy strategy_; + mutable std::unordered_map all_vars_; + mutable std::vector balance_vars_; + + void SetCommunicationContext(OpHandleBase *op_handle, + const platform::Place &p) const; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..69944a42b688a9ea5ff29f75f18dd4b156848a27 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc @@ -0,0 +1,86 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" +#include +#include "paddle/fluid/framework/ir/graph.h" + +namespace paddle { +namespace framework { +namespace details { + +template +static inline void IterAllVar(const ir::Graph &graph, Callback callback) { + for (auto &each : graph.Get(kGraphVars)) { + for (auto &pair1 : each) { + for (auto &pair2 : pair1.second) { + callback(*pair2); + } + } + } + + for (auto &var : graph.Get(kGraphDepVars)) { + callback(*var); + } +} + +void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph, + std::ostream &sout) const { + size_t var_id = 0; + std::unordered_map vars; + + sout << "digraph G {\n"; + + IterAllVar(graph, [&](const VarHandleBase &var) { + auto *var_ptr = &var; + auto *var_handle_ptr = dynamic_cast(var_ptr); + auto *dummy_ptr = dynamic_cast(var_ptr); + + size_t cur_var_id = var_id++; + vars[var_ptr] = cur_var_id; + + if (var_handle_ptr) { + sout << "var_" << cur_var_id << " [label=\"" << var_handle_ptr->name_ + << "\\n" + << var_handle_ptr->place_ << "\\n" + << var_handle_ptr->version_ << "\"]" << std::endl; + } else if (dummy_ptr) { + sout << "var_" << cur_var_id << " [label=\"dummy\"]" << std::endl; + } + }); + + size_t op_id = 0; + for (auto &op : graph.Get(kGraphOps)) { + std::string op_name = "op_" + std::to_string(op_id++); + sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]" + << std::endl; + for (auto in : op->Inputs()) { + std::string var_name = "var_" + std::to_string(vars[in]); + sout << var_name << " -> " << op_name << std::endl; + } + + for (auto out : op->Outputs()) { + std::string var_name = "var_" + std::to_string(vars[out]); + sout << op_name << " -> " << var_name << std::endl; + } + } + + sout << "}\n"; +} +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(multi_devices_print_pass, + paddle::framework::details::SSAGraghBuilderWithPrinter); diff --git a/paddle/fluid/framework/details/multi_devices_graph_print_pass.h b/paddle/fluid/framework/details/multi_devices_graph_print_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..c00685fa1629c0722c315c726053c2cba8bf17e7 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_print_pass.h @@ -0,0 +1,52 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include +#include "paddle/fluid/framework/details/multi_devices_helper.h" + +namespace paddle { +namespace framework { +namespace details { + +class SSAGraphPrinter { + public: + virtual ~SSAGraphPrinter() {} + virtual void Print(const ir::Graph& graph, std::ostream& sout) const = 0; +}; + +class GraphvizSSAGraphPrinter : public SSAGraphPrinter { + public: + void Print(const ir::Graph& graph, std::ostream& sout) const override; +}; + +class SSAGraghBuilderWithPrinter : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override { + std::unique_ptr fout( + new std::ofstream(Get("debug_graphviz_path"))); + PADDLE_ENFORCE(fout->good()); + Get("graph_printer").Print(*graph, *fout); + return graph; + } +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_helper.cc b/paddle/fluid/framework/details/multi_devices_helper.cc new file mode 100644 index 0000000000000000000000000000000000000000..0242274a16c50508f2c0294264c175515c7293ef --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_helper.cc @@ -0,0 +1,20 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#include "paddle/fluid/framework/details/multi_devices_helper.h" + +namespace paddle { +namespace framework { +namespace details {} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_helper.h b/paddle/fluid/framework/details/multi_devices_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..175c5a9950be69d7bf6ae9e386af762007a18a51 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_helper.h @@ -0,0 +1,57 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/details/var_handle.h" + +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/platform/place.h" + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +// all variable in each devices. +// The outside vector is the device vector. Each element of this vector is a +// map from variable name to variables. The variables, who have the same name, +// will have a differsent version. The offset in the +// `std::vector>` is the version of varaibles. +typedef std::vector< + std::unordered_map>>> + GraphVars; +const char kGraphVars[] = "vars"; + +// aux variables to represent dependency. Useful to resolve data hazard. +typedef std::unordered_set> GraphDepVars; +const char kGraphDepVars[] = "dep_vars"; + +// all operators. NOTE that even we use a vector here, the operators is +// unordered. +typedef std::vector> GraphOps; +const char kGraphOps[] = "ops"; + +typedef std::unordered_map ShardedVarDevice; +const char kShardedVarDevice[] = "sharded_var_device"; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc deleted file mode 100644 index 95aa599cd3e403e9cc66b2b5ad35d0d214d1ab5b..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" -#include -#include "paddle/fluid/framework/details/reduce_and_gather.h" - -namespace paddle { -namespace framework { -namespace details { -NCCLAllReduceOpHandle::NCCLAllReduceOpHandle( - const std::vector &local_scopes, - const std::vector &places, - const platform::NCCLContextMap &ctxs) - : local_scopes_(local_scopes), places_(places), nccl_ctxs_(ctxs) { - for (auto &p : places_) { - this->dev_ctxes_[p] = nccl_ctxs_.DevCtx(p); - } -} - -void NCCLAllReduceOpHandle::RunImpl() { - if (inputs_.size() == 1) { - return; // No need to all reduce when GPU count = 1; - } else { - // Wait input done - WaitInputVarGenerated(); - - auto &var_name = static_cast(this->inputs_[0])->name_; - int dtype = -1; - size_t numel = 0; - - std::vector lod_tensors; - - for (size_t i = 0; i < local_scopes_.size(); ++i) { - auto *s = local_scopes_[i]; - auto &local_scope = *s->FindVar(kLocalExecScopeName)->Get(); - - auto &lod_tensor = local_scope.FindVar(var_name)->Get(); - lod_tensors.emplace_back(&lod_tensor); - } - - if (platform::is_gpu_place(lod_tensors[0]->place())) { - std::vector> all_reduce_calls; - for (size_t i = 0; i < local_scopes_.size(); ++i) { - auto &p = places_[i]; - auto &lod_tensor = *lod_tensors[i]; - void *buffer = const_cast(lod_tensor.data()); - - if (dtype == -1) { - dtype = platform::ToNCCLDataType(lod_tensor.type()); - } - - if (numel == 0) { - numel = static_cast(lod_tensor.numel()); - } - - int dev_id = boost::get(p).device; - auto &nccl_ctx = nccl_ctxs_.at(dev_id); - auto stream = nccl_ctx.stream(); - auto comm = nccl_ctx.comm_; - all_reduce_calls.emplace_back([=] { - PADDLE_ENFORCE(platform::dynload::ncclAllReduce( - buffer, buffer, numel, static_cast(dtype), - ncclSum, comm, stream)); - }); - } - this->RunAndRecordEvent([&] { - platform::NCCLGroupGuard guard; - for (auto &call : all_reduce_calls) { - call(); - } - }); - } else { // Special handle CPU only Operator's gradient. Like CRF - auto &trg = *this->local_scopes_[0] - ->FindVar(kLocalExecScopeName) - ->Get() - ->Var() - ->GetMutable(); - - // Reduce All Tensor to trg in CPU - ReduceLoDTensor func(lod_tensors, &trg); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); - - for (size_t i = 0; i < local_scopes_.size(); ++i) { - auto &scope = - *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get(); - auto &p = places_[i]; - auto *var = scope.FindVar(var_name); - auto *dev_ctx = dev_ctxes_[p]; - - RunAndRecordEvent(p, [&trg, var, dev_ctx, p] { - auto &tensor_gpu = *var->GetMutable(); - auto &tensor_cpu = trg; - TensorCopy(tensor_cpu, p, *dev_ctx, &tensor_gpu); - }); - } - } - } -} - -std::string NCCLAllReduceOpHandle::Name() const { return "nccl_all_reduce"; } -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h deleted file mode 100644 index a0c321843e3fc5abcbd1ef2ce2e153250269aa7d..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h +++ /dev/null @@ -1,51 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include - -#include "paddle/fluid/framework/details/op_handle_base.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/platform/nccl_helper.h" - -namespace paddle { -namespace framework { -namespace details { - -struct NCCLAllReduceOpHandle : public OpHandleBase { - NCCLAllReduceOpHandle(const std::vector &local_scopes, - const std::vector &places, - const platform::NCCLContextMap &ctxs); - - std::string Name() const override; - - // Delay and buffer nccl_all_reduce together can significantly increase - // performance. Disable this feature by returning false. - bool IsMultiDeviceTransfer() override { return true; }; - - protected: - void RunImpl() override; - - private: - const std::vector &local_scopes_; - const std::vector &places_; - const platform::NCCLContextMap &nccl_ctxs_; -}; - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index 6b064650b4f09737836bda4a43fa421720077929..ee9f9184da65467b82794c99fe3e95b108373753 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -11,8 +11,8 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. - #include "paddle/fluid/framework/details/op_handle_base.h" +#include namespace paddle { namespace framework { @@ -39,9 +39,9 @@ OpHandleBase::~OpHandleBase() { #endif } -void OpHandleBase::Run(bool use_event) { +void OpHandleBase::Run(bool use_cuda) { #ifdef PADDLE_WITH_CUDA - if (events_.empty() && use_event) { + if (events_.empty() && use_cuda) { for (auto &p : dev_ctxes_) { int dev_id = boost::get(p.first).device; PADDLE_ENFORCE(cudaSetDevice(dev_id)); @@ -50,7 +50,7 @@ void OpHandleBase::Run(bool use_event) { } } #else - PADDLE_ENFORCE(!use_event); + PADDLE_ENFORCE(!use_cuda); #endif RunImpl(); @@ -58,8 +58,10 @@ void OpHandleBase::Run(bool use_event) { void OpHandleBase::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { #ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE_NOT_NULL(waited_ctx); if (platform::is_cpu_place(waited_ctx->GetPlace()) || events_.empty()) { for (auto &dev_ctx : dev_ctxes_) { + PADDLE_ENFORCE_NOT_NULL(dev_ctx.second); dev_ctx.second->Wait(); } } else { @@ -78,19 +80,21 @@ void OpHandleBase::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { void OpHandleBase::AddInput(VarHandleBase *in) { this->inputs_.emplace_back(in); - in->pending_ops_.insert(this); + node_->inputs.push_back(in->Node()); + in->AddOutput(this, this->Node()); } void OpHandleBase::AddOutput(VarHandleBase *out) { outputs_.emplace_back(out); - out->generated_op_ = this; + node_->outputs.push_back(out->Node()); + out->AddInput(this, this->Node()); } void OpHandleBase::WaitInputVarGenerated() { for (auto in_var : inputs_) { if (NeedWait(in_var)) { for (auto &pair : dev_ctxes_) { - in_var->generated_op_->RecordWaitEventOnCtx(pair.second); + in_var->GeneratedOp()->RecordWaitEventOnCtx(pair.second); } } } @@ -99,20 +103,29 @@ void OpHandleBase::WaitInputVarGenerated() { void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) { for (auto *in : inputs_) { if (NeedWait(in)) { - in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[place]); + in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[place]); } } } +size_t OpHandleBase::NoDummyInputSize() const { + size_t cnt = 0; + for (auto *in : inputs_) { + if (dynamic_cast(in) == nullptr) { + ++cnt; + } + } + return cnt; +} + bool OpHandleBase::NeedWait(VarHandleBase *in_var) { - return in_var && in_var->generated_op_; + return in_var && in_var->GeneratedOp(); } void OpHandleBase::RunAndRecordEvent(const std::function &callback) { #ifdef PADDLE_WITH_CUDA if (!events_.empty()) { // Use event std::function method = callback; - for (auto &p : dev_ctxes_) { method = [method, p, this]() { static_cast(p.second)->RecordEvent( diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 8f94206a87dbae8a81727ca48718886bbabbe25c..2d7f18942890245249dd0619a40bb43833c9a2ee 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -13,10 +13,11 @@ // limitations under the License. #pragma once +#include #include #include - #include "paddle/fluid/framework/details/var_handle.h" +#include "paddle/fluid/framework/ir/node.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/macros.h" @@ -26,9 +27,11 @@ namespace details { constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@"; +// Wraps ir::Node and provide helper utilities. +// It's responsible for populating necessary fields of ir::Node. class OpHandleBase { public: - OpHandleBase() {} + explicit OpHandleBase(ir::Node *node) : node_(node) {} virtual ~OpHandleBase(); @@ -36,7 +39,7 @@ class OpHandleBase { virtual std::string Name() const = 0; - void Run(bool use_event); + void Run(bool use_cuda); virtual void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx); @@ -80,6 +83,10 @@ class OpHandleBase { const std::vector &Outputs() const { return outputs_; } + size_t NoDummyInputSize() const; + + ir::Node *Node() { return node_; } + protected: void RunAndRecordEvent(const std::function &callback); @@ -88,11 +95,10 @@ class OpHandleBase { virtual void RunImpl() = 0; + ir::Node *node_; std::vector inputs_; std::vector outputs_; - std::unordered_map - dev_ctxes_; + std::map dev_ctxes_; #ifdef PADDLE_WITH_CUDA std::unordered_map events_; diff --git a/paddle/fluid/framework/details/reduce_and_gather.h b/paddle/fluid/framework/details/reduce_and_gather.h index 2b95a284990da8f9b7c16d6e4221eb1ed061f74b..e28264eb32756f77ef5baed3dff77ba9f0943160 100644 --- a/paddle/fluid/framework/details/reduce_and_gather.h +++ b/paddle/fluid/framework/details/reduce_and_gather.h @@ -35,12 +35,16 @@ struct ReduceLoDTensor { PADDLE_ENFORCE(!src_tensors_.empty()); auto &t0 = *src_tensors_[0]; PADDLE_ENFORCE_NE(t0.numel(), 0); + dst_tensor_.Resize(t0.dims()); T *dst = dst_tensor_.mutable_data(platform::CPUPlace()); - std::copy(t0.data(), t0.data() + t0.numel(), dst); - for (size_t i = 1; i < src_tensors_.size(); ++i) { + for (size_t i = 0; i < src_tensors_.size(); ++i) { auto &t = *src_tensors_[i]; + if (dst == t.data()) { + continue; + } + PADDLE_ENFORCE_EQ(t.dims(), t0.dims()); PADDLE_ENFORCE_EQ(t.type(), t0.type()); std::transform(t.data(), t.data() + t.numel(), dst, dst, @@ -52,8 +56,7 @@ struct ReduceLoDTensor { inline void GatherSelectedRows( const std::vector &src_selecte_rows_, const std::vector &in_places, - const std::unordered_map &dev_ctxes, + const std::map &dev_ctxes, const platform::Place &out_place, SelectedRows *dst_selecte_rows) { PADDLE_ENFORCE(!src_selecte_rows_.empty()); diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index 7160e346dad0615e2fd32b70c096880af0359e1a..6c7e5c1fb06620b1c071b00fcfcc1b4a29bf8d62 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -16,12 +16,18 @@ #include "paddle/fluid/framework/details/container_cast.h" #include "paddle/fluid/framework/details/reduce_and_gather.h" #include "paddle/fluid/framework/details/variable_visitor.h" +#include "paddle/fluid/platform/profiler.h" + +DEFINE_bool( + cpu_deterministic, false, + "Whether to make the result of computation deterministic in CPU side."); namespace paddle { namespace framework { namespace details { void ReduceOpHandle::RunImpl() { + platform::RecordEvent r("reduce", nullptr); if (places_.size() == 1) return; // the input and output may have dummy var. auto in_var_handles = DynamicCast(inputs_); @@ -89,11 +95,33 @@ void ReduceOpHandle::RunImpl() { } else { std::vector lod_tensors = GetInputValues(in_var_handles, var_scopes); + if (paddle::platform::is_cpu_place(lod_tensors[0]->place())) { this->RunAndRecordEvent([&] { - ReduceLoDTensor func(lod_tensors, - out_var->GetMutable()); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); + // FIXME(zcd): The order of summing is important, + // especially when the type of data is float or double. + // For example, the result of `a+b+c+d` may be different + // with the result of `c+a+b+d`, so the summing order should be fixed. + if (!FLAGS_cpu_deterministic) { + ReduceLoDTensor func(lod_tensors, + out_var->GetMutable()); + VisitDataType(ToDataType(lod_tensors[0]->type()), func); + } else { + // We sum lod_tensors to reduce_sum_trg which is in local_scopes_0 + // here, but it doesn't mean reduce_sum_trg must be in local_scopes_0. + auto &reduce_sum_trg = *this->local_scopes_[0] + ->FindVar(kLocalExecScopeName) + ->Get() + ->FindVar(out_var_handle->name_) + ->GetMutable(); + ReduceLoDTensor func(lod_tensors, &reduce_sum_trg); + VisitDataType(ToDataType(lod_tensors[0]->type()), func); + + auto trg = out_var->GetMutable(); + if (reduce_sum_trg.data() != trg->data()) { + TensorCopy(reduce_sum_trg, platform::CPUPlace(), trg); + } + } }); } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) { #ifdef PADDLE_WITH_CUDA diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index c652a2f4eb0f9b73cb19ebbd9d0809210b280ad3..a6289b055f97b7b0e57928358d84117b33cf2df8 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -32,15 +32,18 @@ namespace framework { namespace details { struct ReduceOpHandle : public OpHandleBase { - const std::vector &local_scopes_; - const std::vector &places_; + std::vector local_scopes_; + std::vector places_; #ifdef PADDLE_WITH_CUDA const platform::NCCLContextMap *nccl_ctxs_; - ReduceOpHandle(const std::vector &local_scopes, + ReduceOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, const platform::NCCLContextMap *nccl_ctxs) - : local_scopes_(local_scopes), places_(places), nccl_ctxs_(nccl_ctxs) { + : OpHandleBase(node), + local_scopes_(local_scopes), + places_(places), + nccl_ctxs_(nccl_ctxs) { if (nccl_ctxs_) { for (auto &p_ctx : nccl_ctxs_->contexts_) { dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); @@ -48,9 +51,9 @@ struct ReduceOpHandle : public OpHandleBase { } } #else - ReduceOpHandle(const std::vector &local_scopes, + ReduceOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places) - : local_scopes_(local_scopes), places_(places) {} + : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {} #endif std::string Name() const override; diff --git a/paddle/fluid/framework/details/reduce_op_handle_test.cc b/paddle/fluid/framework/details/reduce_op_handle_test.cc index ffdd7c14eb5097cc8285da090e4a72e1e3f43d86..3a9a58412391b188c5e804b41fa47b3607a36bd1 100644 --- a/paddle/fluid/framework/details/reduce_op_handle_test.cc +++ b/paddle/fluid/framework/details/reduce_op_handle_test.cc @@ -84,6 +84,7 @@ struct TestReduceOpHandle { } void InitReduceOp(size_t out_scope_idx) { + std::vector> nodes; // init scope for (size_t j = 0; j < gpu_list_.size(); ++j) { local_scopes_.push_back(&(g_scope_.NewScope())); @@ -96,19 +97,21 @@ struct TestReduceOpHandle { } param_scopes_[out_scope_idx]->Var("out"); + nodes.emplace_back(new ir::Node("node")); if (use_gpu_) { #ifdef PADDLE_WITH_CUDA - op_handle_.reset( - new ReduceOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get())); + op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_, + gpu_list_, nccl_ctxs_.get())); #else PADDLE_THROW("CUDA is not support."); #endif } else { #ifdef PADDLE_WITH_CUDA - op_handle_.reset( - new ReduceOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get())); + op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_, + gpu_list_, nccl_ctxs_.get())); #else - op_handle_.reset(new ReduceOpHandle(local_scopes_, gpu_list_)); + op_handle_.reset( + new ReduceOpHandle(nodes.back().get(), local_scopes_, gpu_list_)); #endif } @@ -118,8 +121,10 @@ struct TestReduceOpHandle { if (!use_gpu_) { op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); } - auto *in_var_handle = new VarHandle(1, j, "input", gpu_list_[j]); - in_var_handle->generated_op_ = nullptr; + nodes.emplace_back(new ir::Node("node1")); + auto *in_var_handle = + new VarHandle(nodes.back().get(), 1, j, "input", gpu_list_[j]); + in_var_handle->ClearGeneratedOp(); vars_.emplace_back(in_var_handle); op_handle_->AddInput(in_var_handle); } @@ -128,12 +133,13 @@ struct TestReduceOpHandle { vars_.emplace_back(new DummyVarHandle()); DummyVarHandle *in_dummy_var_handle = static_cast(vars_.back().get()); - in_dummy_var_handle->generated_op_ = nullptr; + in_dummy_var_handle->ClearGeneratedOp(); op_handle_->AddInput(in_dummy_var_handle); // add output - auto *out_var_handle = - new VarHandle(2, out_scope_idx, "out", gpu_list_[out_scope_idx]); + nodes.emplace_back(new ir::Node("node2")); + auto *out_var_handle = new VarHandle(nodes.back().get(), 2, out_scope_idx, + "out", gpu_list_[out_scope_idx]); vars_.emplace_back(out_var_handle); op_handle_->AddOutput(out_var_handle); diff --git a/paddle/fluid/framework/details/rpc_op_handle.cc b/paddle/fluid/framework/details/rpc_op_handle.cc index 7f4da4c01de1010467d839ee5490c5e0d02d8c24..f44b374edb29228dff5a8bf003d945291f166d49 100644 --- a/paddle/fluid/framework/details/rpc_op_handle.cc +++ b/paddle/fluid/framework/details/rpc_op_handle.cc @@ -13,18 +13,20 @@ // limitations under the License. #include "paddle/fluid/framework/details/rpc_op_handle.h" +#include "paddle/fluid/framework/ir/graph.h" namespace paddle { namespace framework { namespace details { -RPCOpHandle::RPCOpHandle(const framework::OpDesc &op_desc, - const Scope *local_scope, const platform::Place &place, - const std::string &name) - : op_(framework::OpRegistry::CreateOp(op_desc)), +RPCOpHandle::RPCOpHandle(ir::Node *node, const framework::OpDesc &op_desc, + const Scope *local_scope, const std::string &name, + const platform::Place &place) + : OpHandleBase(node), + op_(framework::OpRegistry::CreateOp(op_desc)), local_scope_(local_scope), - place_(place), - name_(name) {} + name_(name), + place_(place) {} void RPCOpHandle::RunImpl() { // TODO(wuyi): need further analysis whether wait VarDummyHandle. @@ -32,11 +34,11 @@ void RPCOpHandle::RunImpl() { for (auto *in : inputs_) { auto &p = static_cast(in)->place_; // FIXME(Yancey1989): need a better solution instead of use DebugString() - if (in->DebugString() == "dummy") { // HACK + if (ir::IsControlDepVar(*in->Node())) { // HACK continue; } - if (in->generated_op_) { - in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[p]); + if (in->GeneratedOp()) { + in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[p]); } } auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); diff --git a/paddle/fluid/framework/details/rpc_op_handle.h b/paddle/fluid/framework/details/rpc_op_handle.h index d28b7721720d808a8d81701c3811eae16121fb41..7f99cdeacf618a9496eaef98520685d6d1621ae1 100644 --- a/paddle/fluid/framework/details/rpc_op_handle.h +++ b/paddle/fluid/framework/details/rpc_op_handle.h @@ -28,8 +28,9 @@ namespace framework { namespace details { struct RPCOpHandle : public OpHandleBase { - RPCOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope, - const platform::Place& place, const std::string& name); + RPCOpHandle(ir::Node* node, const framework::OpDesc& op_desc, + const Scope* local_scope, const std::string& name, + const platform::Place& place); std::string Name() const override; @@ -43,8 +44,8 @@ struct RPCOpHandle : public OpHandleBase { private: std::unique_ptr op_; const Scope* local_scope_; - const platform::Place& place_; const std::string name_; + platform::Place place_; }; } // namespace details diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc index d9c387e79dc71288e7330597fed57171d447f31b..609e18581957f62b040e04e937873b7a8fa5785a 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc @@ -19,10 +19,14 @@ namespace paddle { namespace framework { namespace details { -ScaleLossGradOpHandle::ScaleLossGradOpHandle(size_t num_dev, Scope *scope, +ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, + Scope *scope, platform::Place place, platform::DeviceContext *dev_ctx) - : coeff_(static_cast(1.0 / num_dev)), scope_(scope), place_(place) { + : OpHandleBase(node), + coeff_(static_cast(1.0 / num_dev)), + scope_(scope), + place_(place) { dev_ctxes_[place_] = dev_ctx; } diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.h b/paddle/fluid/framework/details/scale_loss_grad_op_handle.h index d93d599d46f130cf98f39f15697ce994a31e20c3..523b55724c82d4e2bef0520c10e5708c952a3ecc 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.h +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.h @@ -25,7 +25,8 @@ namespace framework { namespace details { struct ScaleLossGradOpHandle : public OpHandleBase { - ScaleLossGradOpHandle(size_t num_dev, Scope *scope, platform::Place place, + ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, Scope *scope, + platform::Place place, platform::DeviceContext *context); ~ScaleLossGradOpHandle() final; diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc new file mode 100644 index 0000000000000000000000000000000000000000..5bd974d6b789a2f085c0a69de5e133187342f587 --- /dev/null +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc @@ -0,0 +1,89 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h" +#include +#include +#include +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace framework { +namespace details { +ScopeBufferedSSAGraphExecutor::ScopeBufferedSSAGraphExecutor( + ExecutionStrategy strategy, std::vector local_scopes, + std::vector var_infos, std::vector places, + std::unique_ptr &&underlying_executor) + : strategy_(std::move(strategy)), + underlying_executor_(std::move(underlying_executor)), + local_scopes_(std::move(local_scopes)), + var_infos_(std::move(var_infos)), + places_(std::move(places)) {} + +FeedFetchList ScopeBufferedSSAGraphExecutor::Run( + const std::vector &fetch_tensors) { + if (drop_scope_counter_ == 0) { + // Create local scopes. + for (auto it = local_scopes_.rbegin(); it != local_scopes_.rend(); ++it) { + auto &scope = *it; + Scope &local_scope = scope->NewScope(); + *scope->Var(details::kLocalExecScopeName)->GetMutable() = + &local_scope; + + for (auto &info : var_infos_) { + if (scope->FindVar(info.name_) != nullptr) { + continue; + } + + if (info.persistable_) { // Persistable + InitializeVariable(scope->Var(info.name_), info.type_); + } else { + InitializeVariable(local_scope.Var(info.name_), info.type_); + } + } + } + } + std::vector fetch_data; + std::exception_ptr eptr; + try { + fetch_data = underlying_executor_->Run(fetch_tensors); + } catch (...) { + eptr = std::current_exception(); + } + + platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr); + drop_scope_counter_ += 1; + if (!fetch_tensors.empty() || + drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) { + drop_scope_counter_ = 0; + // Wait All computational streams + for (auto p : places_) { + platform::DeviceContextPool::Instance().Get(p)->Wait(); + } + for (auto &scope : local_scopes_) { + auto &local_scope = + *scope->Var(details::kLocalExecScopeName)->GetMutable(); + scope->DeleteScope(local_scope); + } + } + if (eptr) { + std::rethrow_exception(eptr); + } else { + return fetch_data; + } +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h new file mode 100644 index 0000000000000000000000000000000000000000..5e87e0bf50b51d2b630aba06a5907dd721754d1f --- /dev/null +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h @@ -0,0 +1,61 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/details/var_handle.h" + +#include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/ssa_graph_executor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/place.h" +namespace paddle { +namespace framework { +namespace details { + +struct VariableInfo { + std::string name_; + proto::VarType::Type type_; + bool persistable_; +}; + +class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor { + public: + ScopeBufferedSSAGraphExecutor( + ExecutionStrategy strategy, std::vector local_scopes, + std::vector var_infos, std::vector places, + std::unique_ptr&& underlying_executor); + + const ir::Graph& Graph() const override { + return underlying_executor_->Graph(); + } + + FeedFetchList Run(const std::vector& fetch_tensors) override; + + private: + size_t drop_scope_counter_{0}; + + ExecutionStrategy strategy_; + std::unique_ptr underlying_executor_; + std::vector local_scopes_; + std::vector var_infos_; + std::vector places_; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph.cc b/paddle/fluid/framework/details/ssa_graph.cc deleted file mode 100644 index 1b8c889449059c563ea39f86250075ac2537cdbe..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/ssa_graph.cc +++ /dev/null @@ -1,15 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/ssa_graph.h" diff --git a/paddle/fluid/framework/details/ssa_graph.h b/paddle/fluid/framework/details/ssa_graph.h deleted file mode 100644 index e996a00c162186e47e77d007503ac67caa9f8024..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/ssa_graph.h +++ /dev/null @@ -1,49 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include -#include - -#include "paddle/fluid/framework/details/op_handle_base.h" -#include "paddle/fluid/framework/details/var_handle.h" - -namespace paddle { -namespace framework { -namespace details { - -// A SSA graph used by parallel executor. -struct SSAGraph { - // all variable in each devices. - // The outside vector is the device vector. Each element of this vector is a - // map from variable name to variables. The variables, who have the same name, - // will have a different version. The offset in the - // `std::vector>` is the version of varaibles. - std::vector< - std::unordered_map>>> - vars_; - - // aux variables to represent dependency. Useful to resolve data hazard. - std::unordered_set> dep_vars_; - - // all operators. NOTE that even we use a vector here, the operators is - // unordered. - std::vector> ops_; -}; - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_builder.cc b/paddle/fluid/framework/details/ssa_graph_builder.cc deleted file mode 100644 index 6a567527550883add08031e50aa8de2b204cf13d..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/ssa_graph_builder.cc +++ /dev/null @@ -1,146 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/ssa_graph_builder.h" - -namespace paddle { -namespace framework { -namespace details { -void SSAGraphBuilder::PolishGraphToSupportDataHazards(SSAGraph *graph) { - for (auto &var_map : graph->vars_) { - for (auto &name_pair : var_map) { - if (name_pair.second.size() <= 1) { - continue; - } - auto it_new = name_pair.second.rbegin(); - auto it_old = name_pair.second.rbegin(); - ++it_old; - for (; it_old != name_pair.second.rend(); it_new = it_old, ++it_old) { - auto *write_op = (*it_new)->generated_op_; - auto &read_ops = (*it_old)->pending_ops_; - - for (auto *read_op : read_ops) { - // Manually add a dependency var from read_op to write_op; - if (read_op == write_op) { - // Read Write is the same op. - continue; - } - - auto *dep_var = new DummyVarHandle(); - read_op->AddOutput(dep_var); - write_op->AddInput(dep_var); - graph->dep_vars_.emplace(dep_var); - } - } - } - } -} - -VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle( - SSAGraph *graph, const std::string &each_var_name, - const platform::Place &place, size_t place_offset) { - auto &var_holders = graph->vars_[place_offset]; - auto &var_holder = var_holders[each_var_name]; - VarHandle *var = nullptr; - if (var_holder.empty()) { - var = new VarHandle(0, place_offset, each_var_name, place); - var_holder.emplace_back(var); - } else { - var = var_holder.rbegin()->get(); - } - return var; -} - -void SSAGraphBuilder::CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, - const std::string &each_var_name, - const platform::Place &place, - size_t place_offset) { - auto &vars = graph->vars_[place_offset][each_var_name]; - size_t version = vars.size(); - auto var = new VarHandle(version, place_offset, each_var_name, place); - vars.emplace_back(var); - op_handle->AddOutput(var); -} - -template -void IterAllVar(const SSAGraph &graph, Callback callback) { - for (auto &each : graph.vars_) { - for (auto &pair1 : each) { - for (auto &pair2 : pair1.second) { - callback(*pair2); - } - } - } - - for (auto &var : graph.dep_vars_) { - callback(*var); - } -} - -void SSAGraphBuilder::PrintGraphviz(const SSAGraph &graph, std::ostream &sout) { - size_t var_id = 0; - std::unordered_map vars; - - sout << "digraph G {\n"; - - IterAllVar(graph, [&](const VarHandleBase &var) { - auto *var_ptr = &var; - auto *var_handle_ptr = dynamic_cast(var_ptr); - auto *dummy_ptr = dynamic_cast(var_ptr); - - size_t cur_var_id = var_id++; - vars[var_ptr] = cur_var_id; - - if (var_handle_ptr) { - sout << "var_" << cur_var_id << " [label=\"" << var_handle_ptr->name_ - << "\\n" - << var_handle_ptr->place_ << "\\n" - << var_handle_ptr->version_ << "\"]" << std::endl; - } else if (dummy_ptr) { - sout << "var_" << cur_var_id << " [label=\"dummy\"]" << std::endl; - } - }); - - size_t op_id = 0; - for (auto &op : graph.ops_) { - std::string op_name = "op_" + std::to_string(op_id++); - sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]" - << std::endl; - for (auto in : op->Inputs()) { - std::string var_name = "var_" + std::to_string(vars[in]); - sout << var_name << " -> " << op_name << std::endl; - } - - for (auto out : op->Outputs()) { - std::string var_name = "var_" + std::to_string(vars[out]); - sout << op_name << " -> " << var_name << std::endl; - } - } - - sout << "}\n"; -} - -void SSAGraphBuilder::AddOutputToLeafOps(SSAGraph *graph) { - for (auto &op : graph->ops_) { - if (!op->Outputs().empty()) { - continue; - } - auto *dummy_leaf = new DummyVarHandle(); - graph->dep_vars_.emplace(dummy_leaf); - op->AddOutput(dummy_leaf); - } -} -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_builder.h b/paddle/fluid/framework/details/ssa_graph_builder.h deleted file mode 100644 index 64e5d93081eb76c56898bbeb530e37364619fdbb..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/ssa_graph_builder.h +++ /dev/null @@ -1,63 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include - -#include "paddle/fluid/framework/details/ssa_graph.h" -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/platform/place.h" - -namespace paddle { -namespace framework { -namespace details { - -class SSAGraphBuilder { - public: - SSAGraphBuilder() {} - virtual ~SSAGraphBuilder() {} - virtual std::unique_ptr Build(const ProgramDesc &program) const = 0; - - DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder); - - protected: - /** - * We only handle write after read(WAR), since it should not have a write - * after write in program. If there are write after write operators, we need - * prune them. - * - * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR) - */ - static void PolishGraphToSupportDataHazards(SSAGraph *graph); - - static VarHandle *CreateOrGetLatestVarHandle(SSAGraph *graph, - const std::string &each_var_name, - const platform::Place &place, - size_t place_offset); - - // Add an output variable (each_var_name, place, place_offset) to op_handle, - // which belongs to graph - static void CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, - const std::string &each_var_name, - const platform::Place &place, size_t place_offset); - - static void AddOutputToLeafOps(SSAGraph *graph); - - static void PrintGraphviz(const SSAGraph &graph, std::ostream &sout); -}; -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_executor.cc b/paddle/fluid/framework/details/ssa_graph_executor.cc index 8da6ca889b89999e0f6f974503cea476c9de97f3..09b97bd0d98dc4ad1124dcbc495cff921bf03efc 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/ssa_graph_executor.cc @@ -17,10 +17,6 @@ namespace paddle { namespace framework { namespace details { - -SSAGraphExecutor::SSAGraphExecutor(std::unique_ptr &&graph) - : graph_(std::move(graph)) {} - SSAGraphExecutor::~SSAGraphExecutor() {} } // namespace details diff --git a/paddle/fluid/framework/details/ssa_graph_executor.h b/paddle/fluid/framework/details/ssa_graph_executor.h index a8833b7388ab907020a260d356f1484ffd227658..96fffb7d9430cd00b3823ada9fbe9a65a6bd718c 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.h +++ b/paddle/fluid/framework/details/ssa_graph_executor.h @@ -18,8 +18,8 @@ #include #include -#include "paddle/fluid/framework/details/ssa_graph.h" #include "paddle/fluid/framework/feed_fetch_type.h" +#include "paddle/fluid/framework/ir/graph.h" namespace paddle { namespace framework { @@ -28,15 +28,13 @@ class SSAGraphExecutor { DISABLE_COPY_AND_ASSIGN(SSAGraphExecutor); public: - // Steal graph inside - explicit SSAGraphExecutor(std::unique_ptr &&graph); + SSAGraphExecutor() {} virtual ~SSAGraphExecutor(); - virtual FeedFetchList Run(const std::vector &fetch_tensors) = 0; + virtual const ir::Graph& Graph() const = 0; - protected: - std::unique_ptr graph_; + virtual FeedFetchList Run(const std::vector& fetch_tensors) = 0; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 815f739371e77d953a28be99b38ec1b8ff26506c..c9e331ef359f853263f8dad38dd0a2be4d9618ad 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -14,14 +14,17 @@ #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/platform/profiler.h" + namespace paddle { namespace framework { namespace details { ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::unique_ptr &&graph) - : SSAGraphExecutor(std::move(graph)), + std::unique_ptr &&graph) + : graph_(std::move(graph)), pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_) : nullptr), local_scopes_(local_scopes), @@ -32,6 +35,8 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( FeedFetchList ThreadedSSAGraphExecutor::Run( const std::vector &fetch_tensors) { + std::unique_ptr event( + new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr)); std::unordered_map pending_ops; std::unordered_set pending_vars; BlockingQueue ready_vars; @@ -43,18 +48,18 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( std::unordered_set delayed_ops; // Transform SSAGraph to pending_ops & pending_vars - for (auto &var_map : graph_->vars_) { + for (auto &var_map : graph_->Get(details::kGraphVars)) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { InsertPendingVar(&pending_vars, &ready_vars, version_pair.get()); } } } - for (auto &var : graph_->dep_vars_) { + for (auto &var : graph_->Get(details::kGraphDepVars)) { InsertPendingVar(&pending_vars, &ready_vars, var.get()); } - for (auto &op : graph_->ops_) { + for (auto &op : graph_->Get(details::kGraphOps)) { if (op->Inputs().empty()) { // Special case, Op has no input. ready_ops.insert(op.get()); } else { @@ -64,11 +69,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // Step 2. Insert FetchOps std::vector> fetch_ops; + std::vector> tmp_nodes; std::unordered_set> fetch_dependencies; FeedFetchList fetch_data(fetch_tensors.size()); - InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, - &pending_vars, &ready_vars, &fetch_data); + InsertFetchOps(fetch_tensors, &fetch_ops, &tmp_nodes, &fetch_dependencies, + &pending_ops, &pending_vars, &ready_vars, &fetch_data); auto run_all_ops = [&](std::unordered_set &set) { for (auto *op : set) { @@ -78,6 +84,11 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( set.clear(); }; + // Clean run context + run_op_futures_.clear(); + exception_holder_.Clear(); + event.reset(nullptr); + // Step 3. Execution while (!pending_vars.empty()) { // 1. Run All Ready ops @@ -96,10 +107,11 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( auto cur_ready_vars = ready_vars.PopAll(1, &timeout); if (timeout) { - if (exception_) { - auto exp = *exception_; - exception_.reset(); - throw exp; + if (exception_holder_.IsCaught()) { + for (auto &run_op_future : run_op_futures_) { + run_op_future.wait(); + } + exception_holder_.ReThrow(); } else { continue; } @@ -108,7 +120,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // Find the ready_ops after the ready_var. for (auto ready_var : cur_ready_vars) { pending_vars.erase(ready_var); - for (auto *op : ready_var->pending_ops_) { + for (auto *op : ready_var->PendingOps()) { auto &deps = pending_ops[op]; --deps; if (deps == 0) { @@ -134,6 +146,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( void ThreadedSSAGraphExecutor::InsertFetchOps( const std::vector &fetch_tensors, std::vector> *fetch_ops, + std::vector> *temp_nodes, std::unordered_set> *fetch_dependencies, std::unordered_map *pending_ops, std::unordered_set *pending_vars, @@ -141,7 +154,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps( std::unordered_map> fetched_vars; for (auto &fetch_var_name : fetch_tensors) { - for (auto &var_map : graph_->vars_) { + for (auto &var_map : graph_->Get(details::kGraphVars)) { auto it = var_map.find(fetch_var_name); if (it != var_map.end()) { fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); @@ -151,8 +164,16 @@ void ThreadedSSAGraphExecutor::InsertFetchOps( for (size_t i = 0; i < fetch_tensors.size(); ++i) { auto &var_name = fetch_tensors[i]; - auto &vars = fetched_vars.at(var_name); - auto *op = new FetchOpHandle(fetch_data, i, &local_scopes_); + auto fetched_var_it = fetched_vars.find(var_name); + PADDLE_ENFORCE(fetched_var_it != fetched_vars.end(), + "Cannot find fetched variable.(Perhaps the main_program " + "is not set to ParallelExecutor)"); + + auto &vars = fetched_var_it->second; + + temp_nodes->emplace_back(new ir::Node("fetch", ir::Node::Type::kOperation)); + auto *op = new FetchOpHandle(temp_nodes->back().get(), fetch_data, i, + &local_scopes_); fetch_ops->emplace_back(op); for (auto &p : places_) { @@ -163,7 +184,8 @@ void ThreadedSSAGraphExecutor::InsertFetchOps( op->AddInput(var); } - auto *fetch_dummy = new DummyVarHandle(); + temp_nodes->emplace_back(new ir::Node("fetch", ir::Node::Type::kOperation)); + auto *fetch_dummy = new DummyVarHandle(temp_nodes->back().get()); op->AddOutput(fetch_dummy); fetch_dependencies->emplace(fetch_dummy); this->InsertPendingVar(pending_vars, ready_vars, fetch_dummy); @@ -181,28 +203,29 @@ void ThreadedSSAGraphExecutor::InsertPendingVar( std::unordered_set *pending_vars, BlockingQueue *ready_vars, VarHandleBase *var) const { pending_vars->insert(var); - if (var->generated_op_ == nullptr) { + if (var->GeneratedOp() == nullptr) { ready_vars->Push(var); } } + void ThreadedSSAGraphExecutor::RunOp( BlockingQueue *ready_var_q, details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { - VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); - op->Run(strategy_.use_event_); + if (VLOG_IS_ON(10)) { + VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); + } + op->Run(strategy_.use_cuda_); VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); VLOG(10) << op << " " << op->Name() << "Signal posted"; - } catch (platform::EnforceNotMet ex) { - exception_.reset(new platform::EnforceNotMet(ex)); } catch (...) { - LOG(FATAL) << "Unknown exception catched"; + exception_holder_.Catch(std::current_exception()); } }; if (pool_) { - pool_->enqueue(op_run); + run_op_futures_.emplace_back(pool_->enqueue(op_run)); } else { op_run(); } diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index 1f7f88d75218e757e4555ad093f3cd6558f624dd..9135c1f5d435d5e2c60eb90c80803361aa31a3c4 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -15,6 +15,7 @@ #pragma once #include +#include #include #include #include @@ -23,9 +24,11 @@ #include #include "ThreadPool.h" // ThreadPool in thrird party #include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/details/exception_holder.h" #include "paddle/fluid/framework/details/execution_strategy.h" #include "paddle/fluid/framework/details/fetch_op_handle.h" #include "paddle/fluid/framework/details/ssa_graph_executor.h" +#include "paddle/fluid/framework/ir/graph.h" namespace paddle { namespace framework { @@ -38,8 +41,9 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::unique_ptr &&graph); + std::unique_ptr &&graph); + const ir::Graph &Graph() const override { return *graph_; } // Run a SSAGraph by a thread pool // Use topological sort algorithm FeedFetchList Run(const std::vector &fetch_tensors) override; @@ -51,11 +55,12 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { details::OpHandleBase *op); private: + std::unique_ptr graph_; std::unique_ptr<::ThreadPool> pool_; std::vector local_scopes_; std::vector places_; platform::DeviceContextPool fetch_ctxs_; - std::unique_ptr exception_; + ExceptionHolder exception_holder_; std::atomic running_ops_; void InsertPendingOp(std::unordered_map *pending_ops, @@ -68,6 +73,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { void InsertFetchOps( const std::vector &fetch_tensors, std::vector> *fetch_ops, + std::vector> *temp_nodes, std::unordered_set> *fetch_dependencies, std::unordered_map *pending_ops, std::unordered_set *pending_vars, @@ -75,6 +81,8 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { private: ExecutionStrategy strategy_; + // use std::list because clear(), push_back, and for_each are O(1) + std::list> run_op_futures_; }; } // namespace details diff --git a/paddle/fluid/framework/details/var_handle.cc b/paddle/fluid/framework/details/var_handle.cc index 6f00abd9473a84a77ed1a39015e2ae079e00be79..5457870e9ff5d7cf67c9c7076b9aae94eeada779 100644 --- a/paddle/fluid/framework/details/var_handle.cc +++ b/paddle/fluid/framework/details/var_handle.cc @@ -26,7 +26,7 @@ std::string VarHandle::DebugString() const { return ss.str(); } -std::string DummyVarHandle::DebugString() const { return "dummy"; } +std::string DummyVarHandle::DebugString() const { return node_->Name(); } } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/var_handle.h b/paddle/fluid/framework/details/var_handle.h index cae9af7217660fb7e4b8535ee8e022fb3a127668..d8c2bc40b9458a1d5a7dd8a32277d04f69295f09 100644 --- a/paddle/fluid/framework/details/var_handle.h +++ b/paddle/fluid/framework/details/var_handle.h @@ -13,11 +13,14 @@ // limitations under the License. #pragma once + +#include #include #include #include #include +#include "paddle/fluid/framework/ir/node.h" #include "paddle/fluid/platform/place.h" namespace paddle { @@ -25,19 +28,60 @@ namespace framework { namespace details { class OpHandleBase; +// Wraps ir::Node and provide helper utilities. +// It's responsible for populating necessary fields of ir::Node. +// // VarHandleBase is the var node in the dependency graph. // A variable can only be generated by a single operator. i.e. // This is a single assignment graph. struct VarHandleBase { + explicit VarHandleBase(ir::Node* node) : node_(node) {} + virtual ~VarHandleBase(); + virtual std::string DebugString() const = 0; + void AddInput(OpHandleBase* in, ir::Node* node) { + node_->inputs.clear(); + node_->inputs.push_back(node); + generated_op_ = in; + } + + void AddOutput(OpHandleBase* out, ir::Node* node) { + if (pending_ops_.find(out) == pending_ops_.end()) { + pending_ops_.insert(out); + node_->outputs.push_back(node); + } + } + + void RemoveOutput(OpHandleBase* out, ir::Node* node) { + pending_ops_.erase(out); + node_->outputs.erase( + std::remove(node_->outputs.begin(), node_->outputs.end(), node), + node_->outputs.end()); + } + + void ClearGeneratedOp() { + generated_op_ = nullptr; + node_->inputs.clear(); + } + + OpHandleBase* GeneratedOp() { return generated_op_; } + + const std::unordered_set& PendingOps() const { + return pending_ops_; + } + + ir::Node* Node() { return node_; } + + protected: // The operator who generate this variable. nullptr if the variable // is a root node. OpHandleBase* generated_op_{nullptr}; // Operators which depend on this variable ready. std::unordered_set pending_ops_; + ir::Node* node_; }; // VarHandle is actually a single version of Runtime Variable. @@ -46,11 +90,14 @@ struct VarHandleBase { // // NOTE: runtime variables have place. struct VarHandle : public VarHandleBase { + explicit VarHandle(ir::Node* node) : VarHandleBase(node) {} + std::string DebugString() const override; - VarHandle(size_t version, size_t scope_index, std::string name, - platform::Place place) - : version_(version), + VarHandle(ir::Node* node, size_t version, size_t scope_index, + std::string name, platform::Place place) + : VarHandleBase(node), + version_(version), scope_idx_(scope_index), name_(std::move(name)), place_(std::move(place)) {} @@ -70,6 +117,8 @@ struct VarHandle : public VarHandleBase { // Dummy Variable. It is used to represent dependencies between operators struct DummyVarHandle : public VarHandleBase { + explicit DummyVarHandle(ir::Node* node) : VarHandleBase(node) {} + std::string DebugString() const override; }; diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 863053c32b190f4e8497b16f3edd76cb2f76168b..dad170ed78c64202b5c812bd8682887fe3b736d6 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -20,10 +20,12 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" DECLARE_bool(benchmark); +DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run"); namespace paddle { namespace framework { @@ -43,6 +45,14 @@ ExecutorPrepareContext::~ExecutorPrepareContext() { Executor::Executor(const platform::Place& place) : place_(place) {} +void Executor::Close() { +#ifdef PADDLE_WITH_DISTRIBUTE + ::paddle::operators::distributed::RPCClient::GetInstance< + ::paddle::operators::distributed::GRPCClient>() + ->SendComplete(); +#endif +} + void InitializeVariable(Variable* var, proto::VarType::Type var_type) { if (var_type == proto::VarType::LOD_TENSOR) { var->GetMutable(); @@ -115,6 +125,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, bool create_local_scope, bool create_vars) { platform::RecordBlock b(block_id); + if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc); auto ctx = Prepare(pdesc, block_id); RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars); } @@ -214,16 +225,18 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, const std::string& feed_holder_name, const std::string& fetch_holder_name) { platform::RecordBlock b(kProgramId); + if (FLAGS_use_mkldnn) EnableMKLDNN(program); bool has_feed_ops = has_feed_operators(program.Block(0), *feed_targets, feed_holder_name); bool has_fetch_ops = has_fetch_operators(program.Block(0), *fetch_targets, fetch_holder_name); ProgramDesc* copy_program = const_cast(&program); + std::unique_ptr unique_ptr_of_copy_program; if (!has_feed_ops || !has_fetch_ops) { - copy_program = std::unique_ptr(new ProgramDesc(program)).get(); + unique_ptr_of_copy_program.reset(new ProgramDesc(program)); + copy_program = unique_ptr_of_copy_program.get(); } - auto* global_block = copy_program->MutableBlock(0); if (!has_feed_ops) { @@ -280,13 +293,14 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, std::unique_ptr Executor::Prepare( const ProgramDesc& program, int block_id) { - auto* ctx = new ExecutorPrepareContext(program, block_id); + std::unique_ptr ctx( + new ExecutorPrepareContext(program, block_id)); PADDLE_ENFORCE_LT(static_cast(block_id), program.Size()); auto& block = program.Block(block_id); for (auto& op_desc : block.AllOps()) { ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc)); } - return std::unique_ptr(ctx); + return ctx; } std::vector> Executor::Prepare( @@ -305,7 +319,8 @@ std::vector> Executor::Prepare( } void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, - bool create_local_scope, bool create_vars) { + bool create_local_scope, bool create_vars, + bool keep_kids) { Scope* local_scope = scope; if (create_vars) { if (create_local_scope) { @@ -315,7 +330,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } for (auto& op : ctx->ops_) { - VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); op->Run(*local_scope, place_); if (FLAGS_benchmark) { @@ -324,12 +338,20 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } } platform::DeviceContextPool::Instance().Get(place_)->Wait(); - if (create_vars && create_local_scope) { + if (local_scope != scope) { scope->DeleteScope(local_scope); } else { - // Delete the local scopes created in operators. - scope->DropKids(); + if (!keep_kids) { + // By default, we should delete all kid scopes after run executor because + // some operators may create local scope when running, such as while_op. + // But when while_op also create a local executor to run it's sub block, + // the sub scopes it created should not be dropped immediately, because + // while_grad_op will use some variables created during while_op run, so + // we need to keep the kids and wait for the outer executor to drop them. + scope->DropKids(); + } } + if (FLAGS_benchmark) { VLOG(2) << "-------------------------------------------------------"; VLOG(2) << "Memory used after deleting local scope: " @@ -376,5 +398,22 @@ void Executor::RunPreparedContext( } } +void Executor::EnableMKLDNN(const ProgramDesc& program) { +#ifdef PADDLE_WITH_MKLDNN + VLOG(3) << "use_mkldnn=True"; + for (size_t bid = 0; bid < program.Size(); ++bid) { + auto* block = const_cast(program).MutableBlock(bid); + for (auto* op : block->AllOps()) { + if (op->HasAttr("use_mkldnn")) { + op->SetAttr("use_mkldnn", true); + } + } + } +#else + LOG(WARNING) + << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option"; +#endif +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 0c3c23611d95e0da67cabfb8fb2755a4a52c991b..214ca3dc492c31d4c683790a6ae051be467401c9 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -44,6 +44,12 @@ class Executor { explicit Executor(const platform::Place& place); + /* + * Close this Executor. + * Calling this method will send complete messages to all pserver instances. + */ + void Close(); + /* @Brief * Runtime evaluation of the given ProgramDesc under certain Scope * @@ -71,7 +77,7 @@ class Executor { void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, bool create_local_scope = true, - bool create_vars = true); + bool create_vars = true, bool keep_kids = false); void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, std::map* feed_targets, @@ -81,6 +87,8 @@ class Executor { const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); + void EnableMKLDNN(const ProgramDesc& program); + private: const platform::Place place_; }; diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index d35125fe8c3c8018c38650dc87b2b1474ded6058..2cf14bd371831ab682166f4256d6966b5ab278c8 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -27,6 +27,7 @@ enum AttrType { BOOLEANS = 7; BLOCK = 8; LONG = 9; + BLOCKS = 10; } // OpDesc describes an instance of a C++ framework::OperatorBase @@ -46,6 +47,7 @@ message OpDesc { repeated bool bools = 11; optional int32 block_idx = 12; optional int64 l = 13; + repeated int32 blocks_idx = 14; }; message Var { @@ -71,6 +73,7 @@ message OpProto { optional bool duplicable = 3 [ default = false ]; optional bool intermediate = 4 [ default = false ]; optional bool dispensable = 5 [ default = false ]; + optional string reuse = 6; } // AttrProto describes the C++ type Attribute. diff --git a/paddle/fluid/framework/init.cc b/paddle/fluid/framework/init.cc deleted file mode 100644 index 85beae775b96c3b7e08a2795bcd0ec79b24faeb4..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/init.cc +++ /dev/null @@ -1,126 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#include // for strdup -#include -#include -#include - -#include "paddle/fluid/framework/init.h" -#include "paddle/fluid/framework/operator.h" -#include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/place.h" -#include "paddle/fluid/string/piece.h" - -namespace paddle { -namespace framework { - -std::once_flag gflags_init_flag; -std::once_flag p2p_init_flag; - -void InitGflags(std::vector argv) { - std::call_once(gflags_init_flag, [&]() { - argv.insert(argv.begin(), "dummy"); - int argc = argv.size(); - char **arr = new char *[argv.size()]; - std::string line; - for (size_t i = 0; i < argv.size(); i++) { - arr[i] = &argv[i][0]; - line += argv[i]; - line += ' '; - } - google::ParseCommandLineFlags(&argc, &arr, true); - VLOG(1) << "Init commandline: " << line; - }); -} - -void InitP2P(std::vector devices) { -#ifdef PADDLE_WITH_CUDA - std::call_once(p2p_init_flag, [&]() { - int count = devices.size(); - for (int i = 0; i < count; ++i) { - for (int j = 0; j < count; ++j) { - if (devices[i] == devices[j]) continue; - int can_acess = -1; - PADDLE_ENFORCE( - cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]), - "Failed to test P2P access."); - if (can_acess != 1) { - LOG(WARNING) << "Cannot enable P2P access from " << devices[i] - << " to " << devices[j]; - } else { - cudaSetDevice(devices[i]); - cudaDeviceEnablePeerAccess(devices[j], 0); - } - } - } - }); -#endif -} - -void InitDevices(bool init_p2p) { - /*Init all available devices by default */ - std::vector devices; -#ifdef PADDLE_WITH_CUDA - try { - int count = platform::GetCUDADeviceCount(); - for (int i = 0; i < count; ++i) { - devices.push_back(i); - } - } catch (const std::exception &exp) { - LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime."; - } -#else - LOG(WARNING) - << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; -#endif - InitDevices(init_p2p, devices); -} - -void InitDevices(bool init_p2p, const std::vector devices) { - std::vector places; - int count = 0; -#ifdef PADDLE_WITH_CUDA - try { - count = platform::GetCUDADeviceCount(); - } catch (const std::exception &exp) { - LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime."; - } -#else - LOG(WARNING) - << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; -#endif - - for (size_t i = 0; i < devices.size(); ++i) { - if (devices[i] >= count || devices[i] < 0) { - LOG(WARNING) << "Invalid devices id."; - continue; - } - places.emplace_back(platform::CUDAPlace(devices[i])); - } - if (init_p2p) { - InitP2P(devices); - } - places.emplace_back(platform::CPUPlace()); - platform::DeviceContextPool::Init(places); -} - -void InitGLOG(const std::string &prog_name) { - // glog will not hold the ARGV[0] inside. - // Use strdup to alloc a new string. - google::InitGoogleLogging(strdup(prog_name.c_str())); - google::InstallFailureSignalHandler(); -} - -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/init_test.cc b/paddle/fluid/framework/init_test.cc deleted file mode 100644 index 928e2d14abea604cf483f4bc1e1c58fbae04dd21..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/init_test.cc +++ /dev/null @@ -1,40 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#include "gtest/gtest.h" - -#include "paddle/fluid/framework/init.h" -#include "paddle/fluid/platform/device_context.h" - -TEST(InitDevices, CPU) { - using paddle::framework::InitDevices; - using paddle::platform::DeviceContextPool; - -#ifndef PADDLE_WITH_CUDA - InitDevices(true); - DeviceContextPool& pool = DeviceContextPool::Instance(); - ASSERT_EQ(pool.size(), 1U); -#endif -} - -TEST(InitDevices, CUDA) { - using paddle::framework::InitDevices; - using paddle::platform::DeviceContextPool; - -#ifdef PADDLE_WITH_CUDA - int count = paddle::platform::GetCUDADeviceCount(); - InitDevices(true); - DeviceContextPool& pool = DeviceContextPool::Instance(); - ASSERT_EQ(pool.size(), 1U + static_cast(count)); -#endif -} diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..923a7083d4f30b646bbab03d79992b275aa2b403 --- /dev/null +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -0,0 +1,12 @@ +cc_library(node SRCS node.cc DEPS proto_desc) +cc_library(graph SRCS graph.cc DEPS node) +cc_library(graph_helper SRCS graph_helper.cc DEPS graph) +cc_library(pass SRCS pass.cc DEPS graph node graph_helper) +cc_library(graph_viz_pass SRCS graph_viz_pass.cc DEPS graph pass graph_helper) +cc_library(graph_traits SRCS graph_traits.cc DEPS graph) +cc_library(graph_pattern_detecter SRCS graph_pattern_detecter.cc DEPS graph graph_helper graph_traits) + +cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper) +cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry) +cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry) +cc_test(test_graph_pattern_detecter SRCS graph_pattern_detecter_tester.cc DEPS graph_pattern_detecter) diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc new file mode 100644 index 0000000000000000000000000000000000000000..f87d5212c0cd87a5a63cf2d54ca677516ab45816 --- /dev/null +++ b/paddle/fluid/framework/ir/graph.cc @@ -0,0 +1,252 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/op_proto_maker.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/var_desc.h" + +namespace paddle { +namespace framework { +namespace ir { + +std::vector FindDistTrainSendVars( + const std::vector &nodes) { + std::vector send_vars; + // since parameters are all in block 0, + // it's enough to only scan send ops in block 0 + for (auto &node : nodes) { + auto op_vars = node->Op()->InputArgumentNames(); + send_vars.reserve(send_vars.size() + + std::distance(op_vars.begin(), op_vars.end())); + send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end()); + } + return send_vars; +} + +std::vector FindDistTrainRecvVars( + const std::vector &nodes) { + std::vector recv_vars; + for (auto &node : nodes) { + auto op_vars = node->Op()->OutputArgumentNames(); + recv_vars.reserve(recv_vars.size() + + std::distance(op_vars.begin(), op_vars.end())); + recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end()); + } + return recv_vars; +} + +bool IsDistTrainOp(ir::Node *node, const std::vector &send_vars, + const std::vector &recv_vars) { + if (send_vars.size() == 0 || recv_vars.size() == 0) { + return false; + } + + /** + * Check any of opvars contains `.block` and in sendvars + */ + auto checker = [](const std::vector &opvars, + const std::vector &rpc_vars) -> bool { + for (auto &var : opvars) { + // a variable name with the suffix `.block` means it's a splited + // variable by (DistributeTranspiler) + // [python/paddle/fluid/transpiler/distribute_transpiler.py] + if (var.find(".block") != std::string::npos && + std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) { + return true; + } + } + return false; + }; + + std::vector input_var_names; + std::vector output_var_names; + for (ir::Node *input : node->inputs) { + input_var_names.push_back(input->Name()); + } + for (ir::Node *output : node->outputs) { + output_var_names.push_back(output->Name()); + } + + return checker(output_var_names, send_vars) || + checker(input_var_names, recv_vars); +} + +Graph::Graph(const ProgramDesc &program) : program_(program) { + VLOG(3) << "block in program:" << program_.Size(); + std::unordered_map all_vars; + for (auto *var : program.Block(0).AllVars()) { + all_vars.emplace(var->Name(), var); + } + + std::map> var_nodes; + for (auto *op : program.Block(0).AllOps()) { + ir::Node *node = CreateOpNode(op); + // For input args, reuse the same var name if it was created before. + // Otherwise, create a new one. + for (auto &each_var_name : op->InputArgumentNames()) { + ir::Node *var = nullptr; + if (var_nodes.find(each_var_name) != var_nodes.end()) { + var = var_nodes.at(each_var_name).back(); + } else if (all_vars.count(each_var_name) != 0) { + var = CreateVarNode(all_vars.at(each_var_name)); + var_nodes[each_var_name].push_back(var); + } else { + // Operation input var can be optional (dispensable). Which means + // the operation doesn't really need the var at runtime. In this + // case, the no-existed var is ready at the beginning. + var = CreateEmptyNode(each_var_name, ir::Node::Type::kVariable); + var_nodes[each_var_name].push_back(var); + } + node->inputs.push_back(var); + var->outputs.push_back(node); + } + // For output args, always create a new var. + for (auto &each_var_name : op->OutputArgumentNames()) { + ir::Node *var = CreateVarNode(all_vars.at(each_var_name)); + var_nodes[each_var_name].push_back(var); + node->outputs.push_back(var); + var->inputs.push_back(node); + } + } + + std::vector send_ops; + ir::Node *send_bar = nullptr; + std::vector recv_ops; + ir::Node *fetch_bar = nullptr; + for (ir::Node *node : Nodes()) { + if (node->Name() == "send") { + send_ops.push_back(node); + } else if (node->Name() == "send_barrier") { + PADDLE_ENFORCE(!send_bar, "only has one send barrier"); + send_bar = node; + } else if (node->Name() == "recv") { + recv_ops.push_back(node); + } else if (node->Name() == "fetch_barrier") { + PADDLE_ENFORCE(!fetch_bar, "only has one fetch barrier"); + fetch_bar = node; + } + } + if (send_bar) { + for (ir::Node *send : send_ops) { + ir::Node *dep_var = CreateControlDepVar(); + send->outputs.push_back(dep_var); + dep_var->inputs.push_back(send); + send_bar->inputs.push_back(dep_var); + dep_var->outputs.push_back(send_bar); + } + for (ir::Node *recv : recv_ops) { + ir::Node *dep_var = CreateControlDepVar(); + recv->inputs.push_back(dep_var); + dep_var->outputs.push_back(recv); + send_bar->outputs.push_back(dep_var); + dep_var->inputs.push_back(send_bar); + } + } + if (fetch_bar) { + for (ir::Node *recv : recv_ops) { + ir::Node *dep_var = CreateControlDepVar(); + recv->outputs.push_back(dep_var); + dep_var->inputs.push_back(recv); + fetch_bar->inputs.push_back(dep_var); + dep_var->outputs.push_back(fetch_bar); + } + } + + std::vector send_vars = FindDistTrainSendVars(send_ops); + std::vector recv_vars = FindDistTrainRecvVars(recv_ops); + for (ir::Node *node : Nodes()) { + if (IsDistTrainOp(node, send_vars, recv_vars)) { + if (fetch_bar && node->Name() == "concat") { + ir::Node *dep_var = CreateControlDepVar(); + fetch_bar->outputs.push_back(dep_var); + dep_var->inputs.push_back(fetch_bar); + node->inputs.push_back(dep_var); + dep_var->outputs.push_back(node); + } + } + } + + /** + * We should handle write after read(WAR) and write after write(WAW) here. + * Because some of the operators of the program can be executed parallelly. + * So, to make the program running in the right order, we should add the + * dependence of WAR and WAW. + * + * + * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR) + */ + + for (auto &var : var_nodes) { + auto &versions = var.second; + if (versions.size() <= 1) continue; + + auto it_new = versions.rbegin(); + auto it_old = versions.rbegin(); + ++it_old; + for (; it_old != versions.rend(); it_new = it_old, ++it_old) { + ir::Node *write_op = + (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0]; + const auto &read_ops = (*it_old)->outputs; + + PADDLE_ENFORCE(write_op, "The write_op should not be empty."); + + // Add write after write dependence + ir::Node *upstream_op = + (*it_old)->inputs.empty() ? nullptr : (*it_old)->inputs[0]; + if (upstream_op) { + ir::Node *dep_var = CreateControlDepVar(); + write_op->inputs.push_back(dep_var); + upstream_op->outputs.push_back(dep_var); + dep_var->outputs.push_back(write_op); + dep_var->inputs.push_back(upstream_op); + } + + for (auto *read_op : read_ops) { + // Manually add a dependency var from read_op to write_op; + if (read_op == write_op) { + // Read Write is the same op. + continue; + } + // 2 ops might have been connected via other vars. + bool has_dep = false; + for (ir::Node *r_out : read_op->outputs) { + for (ir::Node *w_in : write_op->inputs) { + if (r_out == w_in) { + has_dep = true; + break; + } + } + } + if (has_dep) continue; + + ir::Node *dep_var = CreateControlDepVar(); + read_op->outputs.push_back(dep_var); + dep_var->inputs.push_back(read_op); + write_op->inputs.push_back(dep_var); + dep_var->outputs.push_back(write_op); + } + } + } +} + +bool IsControlDepVar(const ir::Node &var) { + return var.Name().find(ir::Node::kControlDepVarName) != std::string::npos; +} +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h new file mode 100644 index 0000000000000000000000000000000000000000..5736a5c4e232698085936303d1f23760649f8245 --- /dev/null +++ b/paddle/fluid/framework/ir/graph.h @@ -0,0 +1,163 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include + +#include "paddle/fluid/framework/ir/node.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/variant.h" + +namespace paddle { +namespace framework { +namespace ir { + +/* + * The graph is a Directed Acyclic Single Static Assignment Graph. + * + * In more detail, the following properties must hold: + * + * The graph shouldn't contain cycle. Each node is a black-box to the graph + * so the node itself could be a loop operator. + * + * Each Variable-type node has only one input (thus single static assignment). + * + * The output/input of operator is variable and the output/input of variable + * is operator. + * + * The following data harzards in Program are addressed in the Graph: + * + * Write-After-Read + * a = op1(x) + * x = op2(b) + * A control-dependency connection is created bettwen op1 and op2 such that + * op1->op2, so as to ensure correct order. + * + * Write-After-Write + * x = op1(a) + * x = op2(b) + * A control-dependency connection is created between op1 and op2 such that + * op1->op2, so as to ensure correct order. + * + * Other properties currently hold, but is not enforced yet: + * + * Variable-type node (not control dep) with the same variable name share + * the same underlying VarDesc. + */ +class Graph { + public: + explicit Graph(const ProgramDesc &program); + + virtual ~Graph() { + for (auto &attr : attrs_) { + attr_dels_[attr.first](); + } + attrs_.clear(); + attr_dels_.clear(); + } + + bool Has(const std::string &attr_name) const { + return attrs_.find(attr_name) != attrs_.end(); + } + + template + AttrType &Get(const std::string &attr_name) const { + PADDLE_ENFORCE(Has(attr_name), "%s attr not registered for graph.", + attr_name); + return *boost::any_cast(attrs_.at(attr_name)); + } + + template + void Set(const std::string &attr_name, AttrType *attr) { + PADDLE_ENFORCE(attrs_.count(attr_name) == 0, "%s already set in the graph", + attr_name); + attrs_[attr_name] = attr; + attr_dels_[attr_name] = [attr, attr_name]() { + VLOG(3) << "deleting " << attr_name; + delete attr; + }; + } + + const std::unordered_set &Nodes() const { return node_set_; } + + // Create a normal variable with non-null VarDesc. + ir::Node *CreateVarNode(VarDesc *var_desc) { + return AddNode(new ir::Node(var_desc)); + } + + // Create a normal runnable operator with OpDesc. + ir::Node *CreateOpNode(OpDesc *op_desc) { + return AddNode(new ir::Node(op_desc)); + } + + // Create a control dependency var that connects 2 operations. The + // var doesn't hold any data. Other than that, it's no different from + // other var, considering dependency analysis. + ir::Node *CreateControlDepVar() { + // TODO(panyx0718): control var name should be really unique. + const std::string name = string::Sprintf( + "%s@%llu", ir::Node::kControlDepVarName, node_set_.size()); + return AddNode(new ir::Node(name, ir::Node::Type::kVariable)); + } + + // A more free style way of creating a graph node. Mostly use for test + // or "copy" from another node. Avoid using it if possible. + ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type) { + return AddNode(new ir::Node(name, type)); + } + + // Clear all node information of the graph and return the ownership of the + // nodes. + std::vector> ReleaseNodes() { + std::vector> ret; + for (auto &n : nodes_) { + ret.emplace_back(n.second.release()); + } + nodes_.clear(); + node_set_.clear(); + return ret; + } + + private: + // This method takes ownership of `node`. + ir::Node *AddNode(ir::Node *node) { + PADDLE_ENFORCE(node_set_.find(node) == node_set_.end()); + nodes_[node].reset(node); + node_set_.insert(node); + return node; + } + + void RemoveNode(ir::Node *node) { + PADDLE_ENFORCE(node_set_.find(node) != node_set_.end()); + node_set_.erase(node); + nodes_.erase(node); + } + + // NOTE: program_ shouldn't be exposed to user. + const ProgramDesc &program_; + std::map attrs_; + std::map> attr_dels_; + std::map> nodes_; + std::unordered_set node_set_; +}; + +bool IsControlDepVar(const ir::Node &var); +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc new file mode 100644 index 0000000000000000000000000000000000000000..b1c19e6535150130822e9f48685241e62de5b064 --- /dev/null +++ b/paddle/fluid/framework/ir/graph_helper.cc @@ -0,0 +1,118 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "paddle/fluid/framework/ir/graph_helper.h" + +namespace paddle { +namespace framework { +namespace ir { +namespace { +void SortHelper( + const std::map> &adj_list, + ir::Node *node, std::unordered_set *visited, + std::vector *ret) { + visited->insert(node); + + for (auto adj : adj_list.at(node)) { + if (visited->find(adj) == visited->end()) { + SortHelper(adj_list, adj, visited, ret); + } + } + + VLOG(3) << "topology sort insert: " << node->Name() + << reinterpret_cast(node) << " input " << node->inputs.size(); + ret->push_back(node); +} + +bool HasCircleHelper( + ir::Node *node, + const std::map> &adj_list, + std::unordered_set *visited, + std::unordered_set *in_trace) { + if (visited->find(node) == visited->end()) { + visited->insert(node); + in_trace->insert(node); + + for (ir::Node *in : adj_list.at(node)) { + if (visited->find(in) == visited->end() && + HasCircleHelper(in, adj_list, visited, in_trace)) { + return true; + } else if (in_trace->find(in) != in_trace->end()) { + return true; + } + } + } + in_trace->erase(node); + return false; +} + +bool HasCircleInternal( + const std::map> &adj_list) { + std::unordered_set visited; + std::unordered_set in_trace; + for (auto &adj : adj_list) { + if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace)) { + return true; + } + } + return false; +} +} // namespace + +bool HasCircle(const Graph &graph) { + return HasCircleInternal(BuildOperationAdjList(graph)); +} + +std::vector TopologySortOperations(const Graph &graph) { + std::map> adj_list = + BuildOperationAdjList(graph); + PADDLE_ENFORCE(!HasCircleInternal(adj_list)); + std::unordered_set visited; + std::vector ret; + for (auto adj : adj_list) { + if (visited.find(adj.first) == visited.end()) { + SortHelper(adj_list, adj.first, &visited, &ret); + } + } + return ret; +} + +std::map> BuildOperationAdjList( + const Graph &graph) { + std::map> adj_list; + + for (auto &n : graph.Nodes()) { + if (n->NodeType() != ir::Node::Type::kOperation) continue; + if (adj_list.find(n) == adj_list.end()) { + adj_list[n] = std::unordered_set(); + } + for (auto &var : n->inputs) { + for (auto &adj_n : var->inputs) { + PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation); + adj_list[n].insert(adj_n); + VLOG(3) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) + << " -> " << n->Name() << reinterpret_cast(n) + << " via " << var->Name() << reinterpret_cast(var); + } + } + } + return adj_list; +} + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_helper.h b/paddle/fluid/framework/ir/graph_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..cd6c53a07f8f56781989739d995226bd02b3d3d0 --- /dev/null +++ b/paddle/fluid/framework/ir/graph_helper.h @@ -0,0 +1,40 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/node.h" + +namespace paddle { +namespace framework { +namespace ir { +// Test if the graph contains circle. +bool HasCircle(const Graph &graph); + +// Topology Sort the operations in the graph from inputs to outputs. +// `graph` cannot contain circle. +std::vector TopologySortOperations(const Graph &graph); + +// Build an adjacency list of operations for the `graph`. +std::map> BuildOperationAdjList( + const Graph &graph); + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_helper_test.cc b/paddle/fluid/framework/ir/graph_helper_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..a260dd3da2a7863c06e51aa4feafd824ea254139 --- /dev/null +++ b/paddle/fluid/framework/ir/graph_helper_test.cc @@ -0,0 +1,125 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/ir/graph.h" +#include +#include "gtest/gtest.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/program_desc.h" + +namespace paddle { +namespace framework { +namespace ir { + +void BuildCircleGraph(Graph* g) { + ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation); + ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable); + + o1->outputs.push_back(v1); + o1->inputs.push_back(v1); + v1->inputs.push_back(o1); + v1->outputs.push_back(o1); +} + +void BuildCircleGraph2(Graph* g) { + ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation); + ir::Node* o2 = g->CreateEmptyNode("op2", Node::Type::kOperation); + ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable); + ir::Node* v2 = g->CreateEmptyNode("var2", Node::Type::kVariable); + + o1->outputs.push_back(v1); + o2->inputs.push_back(v1); + v1->inputs.push_back(o1); + v1->outputs.push_back(o2); + + o2->outputs.push_back(v2); + o1->inputs.push_back(v2); + v2->inputs.push_back(o2); + v2->outputs.push_back(o1); +} + +void BuildNoCircleGraph(Graph* g) { + ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation); + ir::Node* o2 = g->CreateEmptyNode("op2", Node::Type::kOperation); + ir::Node* o3 = g->CreateEmptyNode("op3", Node::Type::kOperation); + ir::Node* o4 = g->CreateEmptyNode("op4", Node::Type::kOperation); + ir::Node* o5 = g->CreateEmptyNode("op5", Node::Type::kOperation); + ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable); + ir::Node* v2 = g->CreateEmptyNode("var2", Node::Type::kVariable); + ir::Node* v3 = g->CreateEmptyNode("var3", Node::Type::kVariable); + ir::Node* v4 = g->CreateEmptyNode("var4", Node::Type::kVariable); + + // o1->v1->o2 + o1->outputs.push_back(v1); + o2->inputs.push_back(v1); + v1->inputs.push_back(o1); + v1->outputs.push_back(o2); + // o2->v2->o3 + // o2->v2->o4 + o2->outputs.push_back(v2); + o3->inputs.push_back(v2); + o4->inputs.push_back(v2); + v2->inputs.push_back(o2); + v2->outputs.push_back(o3); + v2->outputs.push_back(o4); + // o2->v3->o5 + o2->outputs.push_back(v3); + o5->inputs.push_back(v3); + v3->inputs.push_back(o2); + v3->outputs.push_back(o5); + // o3-v4->o5 + o3->outputs.push_back(v4); + o5->inputs.push_back(v4); + v4->inputs.push_back(o3); + v4->outputs.push_back(o5); +} + +TEST(GraphHelperTest, Basic) { + ProgramDesc prog; + + Graph g(prog); + BuildCircleGraph(&g); + ASSERT_TRUE(HasCircle(g)); + + Graph g2(prog); + BuildCircleGraph2(&g2); + ASSERT_TRUE(HasCircle(g2)); + + auto adj_list = BuildOperationAdjList(g2); + for (auto& adj : adj_list) { + auto& adj_set = adj.second; + if (adj.first->Name() == "op1") { + ASSERT_EQ((*adj_set.begin())->Name(), "op2"); + } else if (adj.first->Name() == "op2") { + ASSERT_EQ((*adj_set.begin())->Name(), "op1"); + } else { + ASSERT_TRUE(false); + } + } + + Graph g3(prog); + BuildNoCircleGraph(&g3); + ASSERT_FALSE(HasCircle(g3)); + auto sorted = TopologySortOperations(g3); + std::map node_map; + for (size_t i = 0; i < sorted.size(); ++i) { + node_map[sorted[i]->Name()] = i; + } + ASSERT_EQ(node_map.at("op1"), 0UL); + ASSERT_EQ(node_map.at("op2"), 1UL); + ASSERT_TRUE(node_map.at("op3") < node_map.at("op5")); +} +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_pattern_detecter.cc b/paddle/fluid/framework/ir/graph_pattern_detecter.cc new file mode 100644 index 0000000000000000000000000000000000000000..f27d9b0509aa4561cfd1e5da3b46a3a085cc888c --- /dev/null +++ b/paddle/fluid/framework/ir/graph_pattern_detecter.cc @@ -0,0 +1,186 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include + +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/ir/graph_pattern_detecter.h" +#include "paddle/fluid/framework/ir/graph_traits.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace ir { + +PDNode* PDPattern::NewNode(PDNode::teller_t&& teller, const std::string& name) { + nodes_.emplace_back(new PDNode(std::move(teller), name)); + auto* cur = nodes_.back().get(); + return cur; +} + +void PDPattern::AddEdge(PDNode* a, PDNode* b) { + PADDLE_ENFORCE(a); + PADDLE_ENFORCE(b); + PADDLE_ENFORCE(a != b, "can't connect to the same nodes."); + edges_.emplace_back(a, b); +} + +void GraphPatternDetecter::operator()(Graph* graph, + GraphPatternDetecter::handle_t handler) { + if (!MarkPDNodesInGraph(*graph)) return; + auto subgraphs = DetectPatterns(); + UniquePatterns(&subgraphs); + RemoveOverlappedMatch(&subgraphs); + + for (auto& g : subgraphs) { + handler(g, graph); + } +} + +bool GraphPatternDetecter::MarkPDNodesInGraph(const ir::Graph& graph) { + if (graph.Nodes().empty()) return false; + + for (auto& node : GraphTraits::DFS(graph)) { + for (const auto& pdnode : pattern_.nodes()) { + if (pdnode->Tell(&node)) { + pdnodes2nodes_[pdnode.get()].insert(&node); + } + } + } + return !pdnodes2nodes_.empty(); +} + +struct HitGroup { + std::unordered_map roles; + + bool Match(Node* node, PDNode* pat) { + return !roles.count(pat) || roles.at(pat) == node; + } + + void Register(Node* node, PDNode* pat) { roles[pat] = node; } +}; + +// Tell whether Node a links to b. +bool IsNodesLink(Node* a, Node* b) { + for (auto* node : a->outputs) { + if (b == node) { + return true; + } + } + return false; +} + +std::vector +GraphPatternDetecter::DetectPatterns() { + // Init empty subgraphs. + std::vector result; + std::vector init_groups; + PADDLE_ENFORCE(!pattern_.edges().empty(), "At least one edge is needed"); + auto* first_pnode = pattern_.edges().front().first; + if (!pdnodes2nodes_.count(first_pnode)) return result; + for (auto* node : pdnodes2nodes_[first_pnode]) { + HitGroup group; + group.roles[first_pnode] = node; + init_groups.emplace_back(group); + } + + int step = 0; + std::array, 2> bi_records; + bi_records[0] = std::move(init_groups); + + // Extend a PDNode to subgraphs by deducing the connection relations defined + // in edges of PDNodes. + for (const auto& edge : pattern_.edges()) { + // Each role has two PDNodes, which indicates two roles. + // Detect two Nodes that can match these two roles and they are connected. + auto& pre_groups = bi_records[step % 2]; + auto& cur_groups = bi_records[1 - (step++ % 2)]; + cur_groups.clear(); + // source -> target + for (Node* source : pdnodes2nodes_[edge.first]) { + for (Node* target : pdnodes2nodes_[edge.second]) { + // TODO(Superjomn) add some prune strategies. + for (const auto& group : pre_groups) { + HitGroup new_group = group; + if (IsNodesLink(source, target) && + new_group.Match(source, edge.first)) { + new_group.Register(source, edge.first); + if (new_group.Match(target, edge.second)) { + new_group.Register(target, edge.second); + cur_groups.push_back(new_group); + // TODO(Superjomn) need to unique + } + } + } + } + } + } + + for (auto& group : bi_records[step % 2]) { + GraphPatternDetecter::subgraph_t subgraph; + for (auto& role : group.roles) { + subgraph.emplace(role.first, role.second); + } + result.emplace_back(subgraph); + } + return result; +} + +void GraphPatternDetecter::UniquePatterns( + std::vector* subgraphs) { + if (subgraphs->empty()) return; + std::vector result; + + std::unordered_set set; + for (auto& g : *subgraphs) { + size_t key = 0; + for (auto& item : g) { + key ^= std::hash{}(item.first); + key ^= std::hash{}(item.second); + } + if (!set.count(key)) { + result.emplace_back(g); + set.insert(key); + } + } + *subgraphs = result; +} + +void GraphPatternDetecter::RemoveOverlappedMatch( + std::vector* subgraphs) { + std::vector result; + std::unordered_set node_set; + + for (const auto& subgraph : *subgraphs) { + bool valid = true; + for (auto& item : subgraph) { + if (node_set.count(item.second)) { + valid = false; + break; + } + } + if (valid) { + for (auto& item : subgraph) { + node_set.insert(item.second); + } + result.push_back(subgraph); + } + } + *subgraphs = result; +} + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_pattern_detecter.h b/paddle/fluid/framework/ir/graph_pattern_detecter.h new file mode 100644 index 0000000000000000000000000000000000000000..1778bf00000f60e5cf8b2a585bf7e5dae0a582eb --- /dev/null +++ b/paddle/fluid/framework/ir/graph_pattern_detecter.h @@ -0,0 +1,181 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#ifdef PADDLE_WITH_TESTING +#include +#endif + +#include +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/node.h" + +namespace paddle { +namespace framework { +namespace ir { + +// Some basic torminolygies: +// - PDPattern: a pattern defined as a data flow graph. +// - PDNode: the node in the pattern, each PDNode represents an `ir::Node` +// that meets some conditions defined in `PDNode.teller`. +// - A pattern is defined with PDNodes with edges. + +// Pattern detector node. This node helps to build a pattern. +struct PDNode { + // tell whether an ir::Node* is a candidation for a PDNode. + using teller_t = std::function; + + PDNode(teller_t&& teller, const std::string& name = "") + : teller_(teller), name_(name) { + PADDLE_ENFORCE(teller_ != nullptr, "invalid teller functer is set."); + } + + PDNode(PDNode&& other) = default; + + std::vector inlinks; + std::vector outlinks; + + bool Tell(Node* node) const { + PADDLE_ENFORCE(teller_ != nullptr, "teller should be set for a PDNode"); + return teller_(node); + } + + const std::string& name() const { return name_; } + + PDNode(const PDNode&) = delete; + PDNode& operator=(const PDNode&) = delete; + + private: + teller_t teller_; + std::string name_; +}; + +/* + * A pattern in a graph, which defined with PDNode and edges. Most graph + * patterns can be divided into PDNodes and link relations between them. + * + * For example, the FC fusion need to filter the MUL and ELEMENTWISE_ADD + * operators from the computation graph, the MUL's output should have only one + * consumer which is the ELEMENTWISE_ADD. + * This pattern can be defined as with the following pseudo codes + * + * // Create two operator PDNodes. + * MUL = PDPattern.NewNode() + * ELE = PDPattern.NewNode() + * // Create the variable PDNodes. + * MUL_out = PDPattern.NewNode() + * // Add teller to define some rules that help to filter the target Nodes. + * MUL.teller = lambda(node): node->IsOp() && node->Op()->Type == "mul"; + * ELE.teller = lambda(node): \ + * node->IsOp() && node->Op()->Type == "elementwise_add"; + * MUL_out.teller = lambda(node): node->IsVar() && (MUL in node->inputs) + * && (ELE in node->outputs) + * + * One can add more specific tellers for PDNodes or edges, both the Operator + * and Variable Nodes can be ruled in PDNode.teller. + * + * PDPattern can record the general patterns, such as the pattern represents + * - Op in CPU -> Op in GPU -> Op in CPU, to findout the IO abnormal place. + * - Ops whose inputs and outputs share the same variables + */ +class PDPattern { + public: + using edge_t = std::pair; + + void AddEdge(PDNode* a, PDNode* b); + + PDNode* NewNode(PDNode::teller_t&& teller, const std::string& name = ""); + + const std::vector>& nodes() const { return nodes_; } + const std::vector& edges() const { return edges_; } + + private: +#ifdef PADDLE_WITH_TESTING + FRIEND_TEST(PDPattern, AddEdge); + FRIEND_TEST(PDPattern, NewNode); +#endif + + std::vector> nodes_; + std::vector edges_; +}; + +/* + * GraphPatternDetecter helps to detect the specific patterns in the graph. + * Input a pattern, output a list of the matched subgraphs/nodes. + * This helper can be used to support fuse(conv+batchnorm => batchnorm e.g.). + * + * The algorithm has three phases: + * 1. Mark the nodes that match the defined PDNodes in a PDPattern, + * 2. Extend a PDNode to subgraphs by deducing the connection relation defined + * in PAPattern(the edges), + * 3. Get the filtered subgraphs and treat them with a pre-defined handler. + * + * Usage: + * // Create a detector + * GraphPatternDetecter detector; + * // Define the detector's pattern, by adding PDNode and define the edges. + * auto* node0 = detector.mutable_pattern().AddNode(...) + * auto* node1 = detector.mutable_pattern().AddNode(...) + * node0->teller = some lambda. + * node1->teller = some lambda. + * detector.mutable_pattern().AddEdge(node0, node1); + * // Create an handler, to define the behavior of treating the filtered + * // subgraphs that comply with the patterns. + * GraphPatternDetecter::handle_t handler = some labmda + * // Execute the detector. + * detector(&graph, handler); + */ +class GraphPatternDetecter { + public: + using subgraph_t = std::unordered_map; + + // Operate on the detected pattern. + using handle_t = + std::function; + + void operator()(Graph* graph, handle_t handler); + + const PDPattern& pattern() const { return pattern_; } + PDPattern* mutable_pattern() { return &pattern_; } + + private: + // Mark the nodes that fits the pattern. + bool MarkPDNodesInGraph(const ir::Graph& graph); + + // Detect all the pattern and output the hit records. + std::vector DetectPatterns(); + + // Remove duplicate patterns. + void UniquePatterns(std::vector* subgraphs); + + // Remove overlapped match subgraphs, when overlapped, keep the previous one. + void RemoveOverlappedMatch(std::vector* subgraphs); + +#ifdef PADDLE_WITH_TESTING + FRIEND_TEST(GraphPatternDetecter, MarkPDNodesInGraph); + FRIEND_TEST(GraphPatternDetecter, DetectPatterns); +#endif + + private: + using hit_rcd_t = + std::pair; + PDPattern pattern_; + std::vector marked_records_; + std::unordered_map> pdnodes2nodes_; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_pattern_detecter_tester.cc b/paddle/fluid/framework/ir/graph_pattern_detecter_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..993c885a810fe80a170ed190b892b148d85e8b5f --- /dev/null +++ b/paddle/fluid/framework/ir/graph_pattern_detecter_tester.cc @@ -0,0 +1,172 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/graph_pattern_detecter.h" + +#include + +namespace paddle { +namespace framework { +namespace ir { + +void BuildGraph(Graph* g) { + ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation); + ir::Node* o2 = g->CreateEmptyNode("op2", Node::Type::kOperation); + ir::Node* o3 = g->CreateEmptyNode("op3", Node::Type::kOperation); + ir::Node* o4 = g->CreateEmptyNode("op4", Node::Type::kOperation); + ir::Node* o5 = g->CreateEmptyNode("op5", Node::Type::kOperation); + ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable); + ir::Node* v2 = g->CreateEmptyNode("var2", Node::Type::kVariable); + ir::Node* v3 = g->CreateEmptyNode("var3", Node::Type::kVariable); + ir::Node* v4 = g->CreateEmptyNode("var4", Node::Type::kVariable); + + // o1->v1->o2 + o1->outputs.push_back(v1); + o2->inputs.push_back(v1); + v1->inputs.push_back(o1); + v1->outputs.push_back(o2); + // o2->v2->o3 + // o2->v2->o4 + o2->outputs.push_back(v2); + o3->inputs.push_back(v2); + o4->inputs.push_back(v2); + v2->inputs.push_back(o2); + v2->outputs.push_back(o3); + v2->outputs.push_back(o4); + // o2->v3->o5 + o2->outputs.push_back(v3); + o5->inputs.push_back(v3); + v3->inputs.push_back(o2); + v3->outputs.push_back(o5); + // o3-v4->o5 + o3->outputs.push_back(v4); + o5->inputs.push_back(v4); + v4->inputs.push_back(o3); + v4->outputs.push_back(o5); +} + +TEST(PDPattern, NewNode) { + PDPattern x; + auto* n = x.NewNode([](Node* x) { return true; }); + ASSERT_TRUE(n); + ASSERT_EQ(x.nodes_.size(), 1UL); +} + +TEST(PDPattern, AddEdge) { + PDPattern x; + auto* a = x.NewNode([](Node* x) { return true; }); + auto* b = x.NewNode([](Node* x) { return true; }); + ASSERT_TRUE(a); + ASSERT_TRUE(b); + x.AddEdge(a, b); + ASSERT_EQ(x.nodes_.size(), 2UL); + ASSERT_EQ(x.edges_.size(), 1UL); + ASSERT_EQ(x.edges_.front().first, a); + ASSERT_EQ(x.edges_.front().second, b); + + ASSERT_EQ(x.nodes().size(), 2UL); + ASSERT_EQ(x.edges().size(), 1UL); + ASSERT_EQ(x.edges().front().first, a); + ASSERT_EQ(x.edges().front().second, b); +} + +TEST(GraphPatternDetecter, MarkPDNodesInGraph) { + GraphPatternDetecter x; + // mark o2, o3, v2 + + // The pattern is a graph: + // o2(a node named o2) -> v2(a node named v2) + // v2 -> o3(a node named o3) + auto* o2 = x.pattern_.NewNode([](Node* node) { + // The teller can be any condition, such as op type, or variable's shape. + return node && node->Name() == "op2" && node->IsOp(); + }); + auto* o3 = x.pattern_.NewNode([](Node* node) { + // The teller can be any condition, such as op type, or variable's shape. + return node && node->Name() == "op3" && node->IsOp(); + }); + auto* v2 = x.pattern_.NewNode([](Node* node) { + // The teller can be any condition, such as op type, or variable's shape. + return node && node->Name() == "var2" && node->IsVar(); + }); + + ASSERT_FALSE(o2->Tell(nullptr)); + ASSERT_FALSE(o3->Tell(nullptr)); + ASSERT_FALSE(v2->Tell(nullptr)); + + x.pattern_.AddEdge(o2, v2); + x.pattern_.AddEdge(v2, o3); + + ASSERT_EQ(x.pattern_.edges().size(), 2UL); + ASSERT_EQ(x.pattern_.edges()[0].first, o2); + ASSERT_EQ(x.pattern_.edges()[0].second, v2); + ASSERT_EQ(x.pattern_.edges()[1].first, v2); + ASSERT_EQ(x.pattern_.edges()[1].second, o3); + + ProgramDesc program; + Graph graph(program); + BuildGraph(&graph); + + x.MarkPDNodesInGraph(graph); + + ASSERT_EQ(x.pdnodes2nodes_.size(), 3UL); + + auto subgraphs = x.DetectPatterns(); + ASSERT_EQ(subgraphs.size(), 1UL); +} + +TEST(GraphPatternDetecter, MultiSubgraph) { + ProgramDesc program; + Graph graph(program); + BuildGraph(&graph); + + GraphPatternDetecter x; + + // The pattern is a graph: + // op -> var + auto* any_op = x.mutable_pattern()->NewNode( + [](Node* node) { + return node->IsOp() && (node->Name() == "op2" || node->Name() == "op3"); + }, + "OP0"); + auto* any_var = x.mutable_pattern()->NewNode( + [](Node* node) { return node->IsVar(); }, "VAR"); + auto* any_op1 = x.mutable_pattern()->NewNode( + [](Node* node) { return node->IsOp(); }, "OP1"); + + x.mutable_pattern()->AddEdge(any_op, any_var); + x.mutable_pattern()->AddEdge(any_var, any_op1); + + int count = 0; + GraphPatternDetecter::handle_t handle = [&]( + const GraphPatternDetecter::subgraph_t& s, Graph* g) { + LOG(INFO) << "Detect " << s.at(any_op)->Name() << " -> " + << s.at(any_var)->Name() << " -> " << s.at(any_op1)->Name(); + count++; + }; + + x(&graph, handle); + + // 1. Detect op3 -> var4 -> op5 + // 2. Detect op2 -> var2 -> op3 + // 3. Detect op2 -> var2 -> op4 + // 4. Detect op2 -> var3 -> op5 + // But 2 and 3 and 4 overlapped, so keep 2, so the final choices are 1 and 2 + ASSERT_GE(count, 1UL); + ASSERT_LE(count, 2UL); +} + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_test.cc b/paddle/fluid/framework/ir/graph_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..b1b8d1c586c98a327a8e5b4890ced00022155e6b --- /dev/null +++ b/paddle/fluid/framework/ir/graph_test.cc @@ -0,0 +1,208 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/ir/graph.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" + +namespace paddle { +namespace framework { + +class NOP : public OperatorBase { + public: + NOP(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + private: + void RunImpl(const Scope &scope, + const platform::Place &place) const override {} +}; + +class SumOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", "").AsDuplicable(); + AddComment(""); + } +}; + +class SumOpVarTypeInference : public VarTypeInference { + public: + void operator()(const OpDesc &op_desc, BlockDesc *block) const override { + auto &inputs = op_desc.Input("X"); + auto default_var_type = proto::VarType::SELECTED_ROWS; + + bool any_input_is_lod_tensor = std::any_of( + inputs.begin(), inputs.end(), [block](const std::string &name) { + return block->Var(name)->GetType() == proto::VarType::LOD_TENSOR; + }); + if (any_input_is_lod_tensor) { + default_var_type = proto::VarType::LOD_TENSOR; + } + + auto out_var_name = op_desc.Output("Out").front(); + block->Var(out_var_name)->SetType(default_var_type); + } +}; + +class DummyOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", "").AsDuplicable(); + AddComment(""); + } +}; + +class DummyOpVarTypeInference : public VarTypeInference { + public: + void operator()(const OpDesc &op_desc, BlockDesc *block) const override {} +}; +} // namespace framework +} // namespace paddle + +REGISTER_OPERATOR(sum, paddle::framework::NOP, paddle::framework::SumOpMaker, + paddle::framework::SumOpVarTypeInference); +REGISTER_OPERATOR(dummy, paddle::framework::NOP, paddle::framework::SumOpMaker, + paddle::framework::SumOpVarTypeInference); +REGISTER_OPERATOR(sum_without_infer_var_type, paddle::framework::NOP, + paddle::framework::SumOpMaker); + +namespace paddle { +namespace framework { + +TEST(GraphTest, Basic) { + ProgramDesc prog; + auto *op = prog.MutableBlock(0)->AppendOp(); + op->SetType("sum"); + op->SetInput("X", {"test_a", "test_b", "test_c"}); + op->SetOutput("Out", {"test_out"}); + op->SetAttr("op_role", 1); + + prog.MutableBlock(0)->Var("test_a")->SetType(proto::VarType::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarType::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test_c")->SetType(proto::VarType::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test_out"); + + op->InferVarType(prog.MutableBlock(0)); + + ASSERT_EQ(proto::VarType::SELECTED_ROWS, + prog.MutableBlock(0)->Var("test_out")->GetType()); + + prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarType::LOD_TENSOR); + op->InferVarType(prog.MutableBlock(0)); + ASSERT_EQ(proto::VarType::LOD_TENSOR, + prog.MutableBlock(0)->Var("test_out")->GetType()); + + std::unique_ptr g(new ir::Graph(prog)); + std::vector nodes(g->Nodes().begin(), g->Nodes().end()); + for (ir::Node *n : nodes) { + if (n->Name() == "sum") { + ASSERT_EQ(n->inputs.size(), 3UL); + ASSERT_EQ(n->outputs.size(), 1UL); + } else if (n->Name() == "test_a" || n->Name() == "test_b" || + n->Name() == "test_c") { + ASSERT_EQ(n->inputs.size(), 0UL); + ASSERT_EQ(n->outputs.size(), 1UL); + } else if (n->Name() == "test_out") { + ASSERT_EQ(n->inputs.size(), 1UL); + ASSERT_EQ(n->outputs.size(), 0UL); + } + } + ASSERT_EQ(nodes.size(), 5); +} + +TEST(GraphTest, WriteAfterRead) { + // void Test() { + ProgramDesc prog; + auto *op = prog.MutableBlock(0)->AppendOp(); + op->SetType("sum"); + op->SetInput("X", {"a"}); + op->SetOutput("Out", {"b"}); + op->SetAttr("op_role", 1); + + op = prog.MutableBlock(0)->AppendOp(); + op->SetType("dummy"); + op->SetInput("X", {"c"}); + op->SetOutput("Out", {"a"}); + op->SetAttr("op_role", 1); + + prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR); + + std::unique_ptr g(new ir::Graph(prog)); + ir::Node *control_dep1 = nullptr; + ir::Node *control_dep2 = nullptr; + for (ir::Node *n : g->Nodes()) { + if (n->Name() == "sum") { + ASSERT_EQ(n->outputs[0]->Name(), "b"); + ASSERT_TRUE(ir::IsControlDepVar(*n->outputs[1])); + control_dep1 = n->outputs[1]; + ASSERT_EQ(n->outputs.size(), 2); + } + if (n->Name() == "dummy") { + ASSERT_EQ(n->inputs[0]->Name(), "c"); + ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1])); + control_dep2 = n->inputs[1]; + ASSERT_EQ(n->inputs.size(), 2); + } + } + ASSERT_EQ(control_dep1, control_dep2); +} + +TEST(GraphTest, WriteAfterWrite) { + // void Test() { + ProgramDesc prog; + auto *op = prog.MutableBlock(0)->AppendOp(); + op->SetType("sum"); + op->SetInput("X", {"a"}); + op->SetOutput("Out", {"b"}); + op->SetAttr("op_role", 1); + + op = prog.MutableBlock(0)->AppendOp(); + op->SetType("dummy"); + op->SetInput("X", {"c"}); + op->SetOutput("Out", {"b"}); + op->SetAttr("op_role", 1); + + prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR); + + std::unique_ptr g(new ir::Graph(prog)); + ir::Node *control_dep1 = nullptr; + ir::Node *control_dep2 = nullptr; + for (ir::Node *n : g->Nodes()) { + if (n->Name() == "sum") { + ASSERT_EQ(n->outputs[0]->Name(), "b"); + ASSERT_TRUE(ir::IsControlDepVar(*n->outputs[1])); + ASSERT_EQ(n->outputs.size(), 2); + control_dep1 = n->outputs[1]; + } + if (n->Name() == "dummy") { + ASSERT_EQ(n->inputs[0]->Name(), "c"); + ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1])); + control_dep2 = n->inputs[1]; + ASSERT_EQ(n->inputs.size(), 2); + ASSERT_EQ(control_dep1, control_dep2); + } + } +} +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_traits.cc b/paddle/fluid/framework/ir/graph_traits.cc new file mode 100644 index 0000000000000000000000000000000000000000..8f548913e4e1d9d5bc5bdace8b92db9065cf3b5e --- /dev/null +++ b/paddle/fluid/framework/ir/graph_traits.cc @@ -0,0 +1,69 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/graph_traits.h" + +namespace paddle { +namespace framework { +namespace ir { + +// +// NodesDFSIterator +// +NodesDFSIterator::NodesDFSIterator(const std::vector &source) { + for (auto *x : source) stack_.push(x); +} + +NodesDFSIterator::NodesDFSIterator(NodesDFSIterator &&other) noexcept + : stack_(std::move(other.stack_)), + visited_(std::move(other.visited_)) {} + +NodesDFSIterator::NodesDFSIterator(const NodesDFSIterator &other) + : stack_(other.stack_), visited_(other.visited_) {} + +Node &NodesDFSIterator::operator*() { + PADDLE_ENFORCE(!stack_.empty()); + return *stack_.top(); +} + +NodesDFSIterator &NodesDFSIterator::operator++() { + PADDLE_ENFORCE(!stack_.empty(), "the iterator exceeds range"); + visited_.insert(stack_.top()); + auto *cur = stack_.top(); + stack_.pop(); + for (auto *x : cur->outputs) { + if (!visited_.count(x)) { + stack_.push(x); + } + } + return *this; +} +bool NodesDFSIterator::operator==(const NodesDFSIterator &other) { + if (stack_.empty()) return other.stack_.empty(); + if ((!stack_.empty()) && (!other.stack_.empty())) { + return stack_.top() == other.stack_.top(); + } + return false; +} + +NodesDFSIterator &NodesDFSIterator::operator=(const NodesDFSIterator &other) { + stack_ = other.stack_; + visited_ = other.visited_; + return *this; +} +Node *NodesDFSIterator::operator->() { return stack_.top(); } + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_traits.h b/paddle/fluid/framework/ir/graph_traits.h new file mode 100644 index 0000000000000000000000000000000000000000..edbe45acb98326ee3bf1d86495832ec8469b634e --- /dev/null +++ b/paddle/fluid/framework/ir/graph_traits.h @@ -0,0 +1,90 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/node.h" + +namespace paddle { +namespace framework { +namespace ir { + +template +class iterator_range { + IteratorT begin_, end_; + + public: + template + explicit iterator_range(Container &&c) : begin_(c.begin()), end_(c.end()) {} + + iterator_range(const IteratorT &begin, const IteratorT &end) + : begin_(begin), end_(end) {} + + const IteratorT &begin() const { return begin_; } + const IteratorT &end() const { return end_; } +}; + +// DFS iterator on nodes. +struct NodesDFSIterator + : public std::iterator { + NodesDFSIterator() = default; + explicit NodesDFSIterator(const std::vector &source); + NodesDFSIterator(NodesDFSIterator &&other) noexcept; + NodesDFSIterator(const NodesDFSIterator &other); + + Node &operator*(); + NodesDFSIterator &operator++(); + // TODO(Superjomn) current implementation just compare the first + // element, need to compare the graph and all the elements in the queue and + // set. + NodesDFSIterator &operator=(const NodesDFSIterator &other); + bool operator==(const NodesDFSIterator &other); + bool operator!=(const NodesDFSIterator &other) { return !(*this == other); } + Node *operator->(); + + private: + std::stack stack_; + std::unordered_set visited_; +}; + +/* + * GraphTraits contains some graph traversal algorithms. + * + * Usage: + * + */ +struct GraphTraits { + static iterator_range DFS(const Graph &g) { + auto start_points = ExtractStartPoints(g); + NodesDFSIterator x(start_points); + return iterator_range(NodesDFSIterator(start_points), + NodesDFSIterator()); + } + + private: + // The nodes those have no input will be treated as start points. + static std::vector ExtractStartPoints(const Graph &g) { + std::vector result; + for (auto *node : g.Nodes()) { + if (node->inputs.empty()) { + result.push_back(node); + } + } + return result; + } +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_viz_pass.cc b/paddle/fluid/framework/ir/graph_viz_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..8cb812d1388bf74d173a4dc7a99561e730f8e95a --- /dev/null +++ b/paddle/fluid/framework/ir/graph_viz_pass.cc @@ -0,0 +1,72 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "paddle/fluid/framework/ir/graph_viz_pass.h" + +namespace paddle { +namespace framework { +namespace ir { +static const char kGraphVizPath[] = "graph_viz_path"; + +std::unique_ptr GraphVizPass::ApplyImpl( + std::unique_ptr graph) const { + const std::string graph_viz_path = Get(kGraphVizPath); + std::unique_ptr fout(new std::ofstream(graph_viz_path)); + PADDLE_ENFORCE(fout->good()); + std::ostream& sout = *fout; + + size_t var_id = 0; + std::unordered_map vars; + + sout << "digraph G {\n"; + + for (const ir::Node* n : graph->Nodes()) { + if (n->NodeType() != ir::Node::Type::kVariable) continue; + size_t cur_var_id = var_id++; + vars[n] = cur_var_id; + + sout << "var_" << cur_var_id << " [label=\"" << n->Name() << "\"]" + << std::endl; + } + + size_t op_id = 0; + for (const ir::Node* n : graph->Nodes()) { + if (n->NodeType() != ir::Node::Type::kOperation) continue; + std::string op_name = "op_" + std::to_string(op_id++); + sout << op_name << " [label=\"" << n->Name() << "\", shape=rect]" + << std::endl; + for (auto in : n->inputs) { + std::string var_name = "var_" + std::to_string(vars[in]); + sout << var_name << " -> " << op_name << std::endl; + } + + for (auto out : n->outputs) { + std::string var_name = "var_" + std::to_string(vars[out]); + sout << op_name << " -> " << var_name << std::endl; + } + } + + sout << "}\n"; + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(graph_viz_pass, paddle::framework::ir::GraphVizPass) + .RequirePassAttr(paddle::framework::ir::kGraphVizPath); diff --git a/paddle/fluid/framework/ir/graph_viz_pass.h b/paddle/fluid/framework/ir/graph_viz_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..1fd8c8a26e9581ccf605d4271a49ec2e90d8b997 --- /dev/null +++ b/paddle/fluid/framework/ir/graph_viz_pass.h @@ -0,0 +1,38 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +class GraphVizPass : public Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/node.cc b/paddle/fluid/framework/ir/node.cc new file mode 100644 index 0000000000000000000000000000000000000000..aca77da8d674f29b89c023717cdcd061232d023a --- /dev/null +++ b/paddle/fluid/framework/ir/node.cc @@ -0,0 +1,23 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/ir/node.h" + +namespace paddle { +namespace framework { +namespace ir { +const char Node::kControlDepVarName[] = "__control_var"; +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h new file mode 100644 index 0000000000000000000000000000000000000000..9c0765ab8ce16733ac021aefc8c7b2bb779319f3 --- /dev/null +++ b/paddle/fluid/framework/ir/node.h @@ -0,0 +1,79 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/var_desc.h" +#include "paddle/fluid/platform/macros.h" + +namespace paddle { +namespace framework { +namespace ir { + +class Node { + public: + enum class Type { kOperation, kVariable }; + static const char kControlDepVarName[]; + + explicit Node(const std::string& name, Type type) + : name_(name), var_desc_(nullptr), op_desc_(nullptr), type_(type) {} + + explicit Node(VarDesc* var_desc) + : name_(var_desc->Name()), + var_desc_(var_desc), + op_desc_(nullptr), + type_(Type::kVariable) {} + + explicit Node(OpDesc* op_desc) + : name_(op_desc->Type()), + var_desc_(nullptr), + op_desc_(op_desc), + type_(Type::kOperation) {} + + Type NodeType() const { return type_; } + + std::string Name() const { return name_; } + + VarDesc* Var() { + PADDLE_ENFORCE(type_ == Type::kVariable); + return var_desc_; + } + + OpDesc* Op() { + PADDLE_ENFORCE(type_ == Type::kOperation); + return op_desc_; + } + + bool IsOp() const { return type_ == Type::kOperation; } + bool IsVar() const { return type_ == Type::kVariable; } + + std::vector inputs; + std::vector outputs; + + protected: + const std::string name_; + VarDesc* var_desc_; + OpDesc* op_desc_; + Type type_; + + private: + DISABLE_COPY_AND_ASSIGN(Node); +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/pass.cc b/paddle/fluid/framework/ir/pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..d7158eba62686be57499df697466797e4034ea8f --- /dev/null +++ b/paddle/fluid/framework/ir/pass.cc @@ -0,0 +1,46 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/ir/pass.h" +#include "paddle/fluid/framework/ir/graph_helper.h" + +namespace paddle { +namespace framework { +namespace ir { +std::unique_ptr Pass::Apply(std::unique_ptr graph) const { + PADDLE_ENFORCE(!applied_, "Pass can only Apply() once."); + PADDLE_ENFORCE(graph.get(), "graph passed to Pass::Apply() cannot be empty."); + for (const std::string& attr : required_pass_attrs_) { + PADDLE_ENFORCE(attrs_.find(attr) != attrs_.end(), + "Required pass atrribute %s not set.", attr); + } + for (const std::string& attr : required_graph_attrs_) { + PADDLE_ENFORCE(graph->Has(attr), "Required graph atrribute %s not set.", + attr); + } + auto applied_graph = ApplyImpl(std::move(graph)); + // TODO(panyx0718): Add more verifications. + PADDLE_ENFORCE(!HasCircle(*applied_graph), + "Illegal Pass. Generated graph shouldn't has cycle."); + applied_ = true; + return applied_graph; +} + +PassRegistry& PassRegistry::Instance() { + static PassRegistry g_pass_info_map; + return g_pass_info_map; +} +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/pass.h b/paddle/fluid/framework/ir/pass.h new file mode 100644 index 0000000000000000000000000000000000000000..0f14083d259172f5b5f1ed80c7d38312d711beb5 --- /dev/null +++ b/paddle/fluid/framework/ir/pass.h @@ -0,0 +1,200 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/node.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/platform/variant.h" + +namespace paddle { +namespace framework { +namespace ir { +template +struct PassRegistrar; + +class Pass { + public: + Pass() = default; + virtual ~Pass() { + for (auto &attr : attrs_) { + if (attr_dels_.find(attr.first) != attr_dels_.end()) { + attr_dels_[attr.first](); + } + } + attrs_.clear(); + attr_dels_.clear(); + } + + std::unique_ptr Apply(std::unique_ptr graph) const; + + // Get a reference to the attributed previously set. + template + AttrType &Get(const std::string &attr_name) const { + PADDLE_ENFORCE(attrs_.find(attr_name) != attrs_.end(), + "%s attr not registered for pass.", attr_name); + return *boost::any_cast(attrs_.at(attr_name)); + } + + // Set a pointer to the attribute. Pass takes ownership of the attribute. + template + void Set(const std::string &attr_name, AttrType *attr) { + PADDLE_ENFORCE(attrs_.count(attr_name) == 0, "%s already set in the pass", + attr_name); + attrs_[attr_name] = attr; + attr_dels_[attr_name] = [attr, attr_name]() { + VLOG(3) << "deleting " << attr_name; + delete attr; + }; + } + + // Set a pointer to the attribute. Pass doesn't take ownership. Caller + // should delete the attribute. + template + void SetNotOwned(const std::string &attr_name, AttrType *attr) { + PADDLE_ENFORCE(attrs_.count(attr_name) == 0); + attrs_[attr_name] = attr; + } + + protected: + virtual std::unique_ptr ApplyImpl( + std::unique_ptr graph) const = 0; + + private: + template + friend struct PassRegistrar; + + void RegisterRequiredPassAttrs(const std::unordered_set &attrs) { + required_pass_attrs_.insert(attrs.begin(), attrs.end()); + } + + void RegisterRequiredGraphAttrs( + const std::unordered_set &attrs) { + required_graph_attrs_.insert(attrs.begin(), attrs.end()); + } + + mutable bool applied_{false}; + std::unordered_set required_pass_attrs_; + std::unordered_set required_graph_attrs_; + std::map attrs_; + std::map> attr_dels_; +}; + +using PassCreator = std::function()>; + +class Registrar { + public: + // In our design, various kinds of passes, + // have their corresponding registry and registrar. The action of + // registration is in the constructor of a global registrar variable, which + // are not used in the code that calls package framework, and would + // be removed from the generated binary file by the linker. To avoid such + // removal, we add Touch to all registrar classes and make USE_PASS macros to + // call this method. So, as long as the callee code calls USE_PASS, the global + // registrar variable won't be removed by the linker. + void Touch() {} +}; + +class PassRegistry { + public: + static PassRegistry &Instance(); + + bool Has(const std::string &pass_type) const { + return map_.find(pass_type) != map_.end(); + } + + void Insert(const std::string &pass_type, const PassCreator &pass_creator) { + PADDLE_ENFORCE(!Has(pass_type), "Pass %s has been registered", pass_type); + map_.insert({pass_type, pass_creator}); + } + + std::unique_ptr Get(const std::string &pass_type) const { + PADDLE_ENFORCE(Has(pass_type), "Pass %s has not been registered", + pass_type); + return map_.at(pass_type)(); + } + + private: + PassRegistry() = default; + std::unordered_map map_; + + DISABLE_COPY_AND_ASSIGN(PassRegistry); +}; + +template +struct PassRegistrar : public Registrar { + explicit PassRegistrar(const char *pass_type) { + PADDLE_ENFORCE(!PassRegistry::Instance().Has(pass_type), + "'%s' is registered more than once.", pass_type); + PassRegistry::Instance().Insert( + pass_type, [this]() -> std::unique_ptr { + std::unique_ptr pass(new PassType()); + pass->RegisterRequiredPassAttrs(this->required_pass_attrs_); + pass->RegisterRequiredGraphAttrs(this->required_graph_attrs_); + return pass; + }); + } + + PassRegistrar &RequirePassAttr(const std::string &attr) { + required_pass_attrs_.insert(attr); + return *this; + } + + PassRegistrar &RequireGraphAttr(const std::string &attr) { + required_graph_attrs_.insert(attr); + return *this; + } + + private: + std::unordered_set required_pass_attrs_; + std::unordered_set required_graph_attrs_; +}; + +#define STATIC_ASSERT_PASS_GLOBAL_NAMESPACE(uniq_name, msg) \ + struct __test_global_namespace_##uniq_name##__ {}; \ + static_assert(std::is_same<::__test_global_namespace_##uniq_name##__, \ + __test_global_namespace_##uniq_name##__>::value, \ + msg) + +// Register a new pass that can be applied on the IR. +#define REGISTER_PASS(pass_type, pass_class) \ + STATIC_ASSERT_PASS_GLOBAL_NAMESPACE( \ + __reg_pass__##pass_type, \ + "REGISTER_PASS must be called in global namespace"); \ + static ::paddle::framework::ir::PassRegistrar \ + __pass_registrar_##pass_type##__(#pass_type); \ + int TouchPassRegistrar_##pass_type() { \ + __pass_registrar_##pass_type##__.Touch(); \ + return 0; \ + } \ + static ::paddle::framework::ir::PassRegistrar \ + &__pass_tmp_registrar_##pass_type##__ __attribute__((unused)) = \ + __pass_registrar_##pass_type##__ + +#define USE_PASS(pass_type) \ + STATIC_ASSERT_PASS_GLOBAL_NAMESPACE( \ + __use_pass_itself_##pass_type, \ + "USE_PASS must be called in global namespace"); \ + extern int TouchPassRegistrar_##pass_type(); \ + static int use_pass_itself_##pass_type##_ __attribute__((unused)) = \ + TouchPassRegistrar_##pass_type() + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/pass_test.cc b/paddle/fluid/framework/ir/pass_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..5b5011412ed39e033a7a65921e9c64ce2d54c638 --- /dev/null +++ b/paddle/fluid/framework/ir/pass_test.cc @@ -0,0 +1,112 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/ir/pass.h" +#include +#include "gtest/gtest.h" +#include "paddle/fluid/framework/ir/graph.h" + +namespace paddle { +namespace framework { +namespace ir { +void BuildCircleGraph(Graph* g) { + ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation); + ir::Node* o2 = g->CreateEmptyNode("op2", Node::Type::kOperation); + ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable); + ir::Node* v2 = g->CreateEmptyNode("var2", Node::Type::kVariable); + + o1->outputs.push_back(v1); + o2->inputs.push_back(v1); + v1->inputs.push_back(o1); + v1->outputs.push_back(o2); + + o2->outputs.push_back(v2); + o1->inputs.push_back(v2); + v2->inputs.push_back(o2); + v2->outputs.push_back(o1); +} + +class TestPass : public Pass { + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const { + graph->Set("copy_test_pass_attr", new int); + graph->Set("copy_test_graph_attr", new int); + + int test_pass_attr = this->Get("test_pass_attr"); + graph->Get("copy_test_pass_attr") = test_pass_attr + 1; + + int test_graph_attr = graph->Get("test_graph_attr"); + graph->Get("copy_test_graph_attr") = test_graph_attr + 1; + return graph; + } +}; + +TEST(PassTest, TestPassAttrCheck) { + ProgramDesc prog; + auto pass = PassRegistry::Instance().Get("test_pass"); + std::unique_ptr graph(new Graph(prog)); + std::string exception; + try { + graph = pass->Apply(std::move(graph)); + } catch (paddle::platform::EnforceNotMet e) { + exception = std::string(e.what()); + } + ASSERT_TRUE(exception.find("test_pass_attr not set") != exception.npos); + + int val = 1; + graph.reset(new Graph(prog)); + pass->SetNotOwned("test_pass_attr", &val); + + try { + graph = pass->Apply(std::move(graph)); + } catch (paddle::platform::EnforceNotMet e) { + exception = std::string(e.what()); + } + ASSERT_TRUE(exception.find("test_graph_attr not set") != exception.npos); + + graph.reset(new Graph(prog)); + graph->Set("test_graph_attr", new int); + graph->Get("test_graph_attr") = 1; + graph = pass->Apply(std::move(graph)); + ASSERT_EQ(graph->Get("copy_test_pass_attr"), 2); + ASSERT_EQ(graph->Get("copy_test_graph_attr"), 2); + + try { + graph = pass->Apply(std::move(graph)); + } catch (paddle::platform::EnforceNotMet e) { + exception = std::string(e.what()); + } + ASSERT_TRUE(exception.find("Pass can only Apply() once") != exception.npos); + + pass = PassRegistry::Instance().Get("test_pass"); + pass->SetNotOwned("test_pass_attr", &val); + graph.reset(new Graph(prog)); + BuildCircleGraph(graph.get()); + graph->Set("test_graph_attr", new int); + graph->Get("test_graph_attr") = 2; + try { + auto tmp = pass->Apply(std::move(graph)); + } catch (paddle::platform::EnforceNotMet e) { + exception = std::string(e.what()); + } + ASSERT_TRUE(exception.find("shouldn't has cycle") != exception.npos); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(test_pass, paddle::framework::ir::TestPass) + .RequirePassAttr("test_pass_attr") + .RequireGraphAttr("test_graph_attr"); diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc index a56674cbe216e312c4394ef537140122352dc785..919029c38f2f26a6f5e02da645c4f7718044cdae 100644 --- a/paddle/fluid/framework/lod_tensor.cc +++ b/paddle/fluid/framework/lod_tensor.cc @@ -20,6 +20,7 @@ limitations under the License. */ #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/memory/memory.h" @@ -51,8 +52,6 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) { } std::ostream &operator<<(std::ostream &os, const LoDTensor &t) { - PADDLE_ENFORCE(t.type().hash_code() == typeid(float).hash_code()); - if (!platform::is_cpu_place(t.place())) { LoDTensor tt; framework::TensorCopy(t, platform::CPUPlace(), &tt); @@ -70,7 +69,13 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) { // only print first ten elements int64_t size = t.numel() < 10 ? t.numel() : 10; for (int64_t i = 0; i < size; ++i) { - os << t.data()[i] << " "; + if (IsType(t.type())) { + os << t.data()[i] << " "; + } else if (IsType(t.type())) { + os << t.data()[i] << " "; + } else { + PADDLE_THROW("LoDTensor data type not in [float, int64_t]"); + } } return os; @@ -85,6 +90,7 @@ std::string LoDToString(const LoD &lod) { LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin, size_t elem_end) { PADDLE_ENFORCE_LT(level, in.size()); + PADDLE_ENFORCE_LT(elem_begin, elem_end); PADDLE_ENFORCE_LT(elem_end, in[level].size()); LoD res; @@ -306,19 +312,22 @@ void WriteToRecordIO(recordio::Writer *writer, writer->Write(buffer.str()); } -std::vector ReadFromRecordIO( - recordio::Scanner *scanner, const platform::DeviceContext &dev_ctx) { - std::vector result; - if (scanner->HasNext()) { - std::istringstream sin(scanner->Next()); - uint32_t sz; - sin.read(reinterpret_cast(&sz), sizeof(uint32_t)); - result.resize(sz); - for (uint32_t i = 0; i < sz; ++i) { - DeserializeFromStream(sin, &result[i], dev_ctx); - } +bool ReadFromRecordIO(recordio::Scanner *scanner, + const platform::DeviceContext &dev_ctx, + std::vector *result_ptr) { + if (!scanner->HasNext()) { + return false; } - return result; + std::istringstream sin(scanner->Next()); + uint32_t sz; + sin.read(reinterpret_cast(&sz), sizeof(uint32_t)); + auto &result = *result_ptr; + result.resize(sz); + for (uint32_t i = 0; i < sz; ++i) { + DeserializeFromStream(sin, &result[i], dev_ctx); + } + + return true; } std::vector LoDTensor::SplitLoDTensor( @@ -380,7 +389,7 @@ void LoDTensor::MergeLoDTensor( LoD new_lod = lod_tensors[0]->lod(); for (size_t i = 1; i < lod_tensors.size(); ++i) { auto *t = lod_tensors[i]; - PADDLE_ENFORCE_EQ(new_type.hash_code(), t->type().hash_code()); + PADDLE_ENFORCE_EQ(new_type, t->type()); PADDLE_ENFORCE_EQ(new_layout, t->layout()); PADDLE_ENFORCE_EQ(framework::product(new_dim) / new_dim[0], @@ -388,6 +397,7 @@ void LoDTensor::MergeLoDTensor( new_dim[0] += t->dims()[0]; auto &lod = t->lod(); + PADDLE_ENFORCE_EQ(new_lod.size(), lod.size()); for (size_t j = 0; j < lod.size(); ++j) { auto &sub_lod = new_lod[j]; auto &offset = sub_lod.back(); @@ -410,5 +420,38 @@ void LoDTensor::MergeLoDTensor( } } +LoD ConvertToLengthBasedLoD(const LoD &offset_lod) { + LoD length_lod; + length_lod.reserve(offset_lod.size()); + for (size_t lvl = 0; lvl < offset_lod.size(); ++lvl) { + std::vector level; + if (offset_lod[lvl].size() > 0) { + level.reserve(offset_lod[lvl].size() - 1); + } + for (size_t idx = 0; idx < offset_lod[lvl].size() - 1; ++idx) { + level.push_back(offset_lod[lvl][idx + 1] - offset_lod[lvl][idx]); + } + length_lod.push_back(level); + } + return length_lod; +} + +LoD ConvertToOffsetBasedLoD(const LoD &length_lod) { + LoD offset_lod; + offset_lod.reserve(length_lod.size()); + for (size_t lvl = 0; lvl < length_lod.size(); ++lvl) { + std::vector level; + level.reserve(length_lod[lvl].size() + 1); + size_t tmp = 0; + level.push_back(tmp); + for (size_t idx = 0; idx < length_lod[lvl].size(); ++idx) { + tmp += length_lod[lvl][idx]; + level.push_back(tmp); + } + offset_lod.push_back(level); + } + return offset_lod; +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/lod_tensor.h b/paddle/fluid/framework/lod_tensor.h index 1159fee39b0737402c60448dcbe69e7535c9d6e1..e9b473d547252e80ed26ec61e1a33fbe1742dbe0 100644 --- a/paddle/fluid/framework/lod_tensor.h +++ b/paddle/fluid/framework/lod_tensor.h @@ -223,8 +223,23 @@ extern void WriteToRecordIO(recordio::Writer* writer, const std::vector& tensor, const platform::DeviceContext& dev_ctx); -extern std::vector ReadFromRecordIO( - recordio::Scanner* scanner, const platform::DeviceContext& dev_ctx); +extern bool ReadFromRecordIO(recordio::Scanner* scanner, + const platform::DeviceContext& dev_ctx, + std::vector* result_ptr); + +/* + * Convert between length-based LoD and offset-based LoD. + * The implementation of LoDTensor class use offset-based LoD. + * However, we want to expose the more user-friendly length-based + * LoD to the Python side instead. + * + * Example: + * If offset_lod = [[0, 2, 3],[0, 3, 5, 9]] + * then length_lod = [[2, 1], [3, 2, 4]] + */ +LoD ConvertToLengthBasedLoD(const LoD& offset_lod); + +LoD ConvertToOffsetBasedLoD(const LoD& length_lod); } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc index 2ceffc93319359683e87e7fec2d18784c9bf02f3..cd50aaa26054b78f1b1e8f0d470b397892155a2b 100644 --- a/paddle/fluid/framework/lod_tensor_test.cc +++ b/paddle/fluid/framework/lod_tensor_test.cc @@ -26,6 +26,20 @@ namespace paddle { namespace framework { +TEST(LoD, PrintLoDTensor) { + LoDTensor tensor1; + tensor1.mutable_data(platform::CPUPlace()); + tensor1.data()[0] = 0.2; + tensor1.data()[1] = 0.5; + LOG(INFO) << tensor1; + + LoDTensor tensor2; + tensor2.mutable_data(platform::CPUPlace()); + tensor2.data()[0] = 1; + tensor2.data()[1] = 2; + LOG(INFO) << tensor2; +} + TEST(LoD, data) { LoD lod{{0, 1, 2}}; lod.push_back({0, 2, 4, 5}); @@ -37,7 +51,7 @@ TEST(LoD, data) { } } -TEST(LodExpand, test) { +TEST(LoD, ExpandLoD) { LoD lod{{0, 2}}; LoDTensor tensor; tensor.set_lod(lod); @@ -228,6 +242,38 @@ TEST(LoD, CheckAbsLoD) { ASSERT_FALSE(CheckAbsLoD(abs_lod0)); } +TEST(LoD, ConvertToLengthBasedLoD) { + LoD offset_lod; + offset_lod.push_back(std::vector({0, 2})); + offset_lod.push_back(std::vector({0, 1, 3})); + offset_lod.push_back(std::vector({0, 2, 4, 5})); + + LoD length_lod = ConvertToLengthBasedLoD(offset_lod); + + LoD expected; + expected.push_back(std::vector({2})); + expected.push_back(std::vector({1, 2})); + expected.push_back(std::vector({2, 2, 1})); + + EXPECT_EQ(length_lod, expected); +} + +TEST(LoD, ConvertToOffsetBasedLoD) { + LoD length_lod; + length_lod.push_back(std::vector({2})); + length_lod.push_back(std::vector({1, 2})); + length_lod.push_back(std::vector({2, 2, 1})); + + LoD offset_lod = ConvertToOffsetBasedLoD(length_lod); + + LoD expected; + expected.push_back(std::vector({0, 2})); + expected.push_back(std::vector({0, 1, 3})); + expected.push_back(std::vector({0, 2, 4, 5})); + + EXPECT_EQ(offset_lod, expected); +} + template static void TestRecordIO() { LoDTensor tensor; @@ -255,11 +301,12 @@ static void TestRecordIO() { { std::unique_ptr stream_ptr(stream); recordio::Scanner scanner(std::move(stream_ptr)); - auto tensors = ReadFromRecordIO(&scanner, ctx); + std::vector tensors; + ASSERT_TRUE(ReadFromRecordIO(&scanner, ctx, &tensors)); ASSERT_EQ(tensors.size(), static_cast(2)); assert_tensor_ok(tensors[0]); assert_tensor_ok(tensors[1]); - tensors = ReadFromRecordIO(&scanner, ctx); + ASSERT_TRUE(ReadFromRecordIO(&scanner, ctx, &tensors)); ASSERT_EQ(tensors.size(), static_cast(2)); assert_tensor_ok(tensors[0]); assert_tensor_ok(tensors[1]); diff --git a/paddle/fluid/framework/lod_tensor_test.cu b/paddle/fluid/framework/lod_tensor_test.cu index e3efbe4c464493af87e33510647d8c67d457a76d..b9950627ca378cb9607681799bd7fe5bfce2bf50 100644 --- a/paddle/fluid/framework/lod_tensor_test.cu +++ b/paddle/fluid/framework/lod_tensor_test.cu @@ -17,9 +17,9 @@ #include #include "gtest/gtest.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/platform/assert.h" +#include "paddle/fluid/platform/init.h" #include "paddle/fluid/platform/place.h" __global__ void test(size_t* a, int size) { diff --git a/paddle/fluid/framework/mixed_vector.h b/paddle/fluid/framework/mixed_vector.h index 29b3396bc9854cd3d3ac8d4283f48019c9a9c55f..7836ecb1272a07a79a70c9cb040335f9a42e5684 100644 --- a/paddle/fluid/framework/mixed_vector.h +++ b/paddle/fluid/framework/mixed_vector.h @@ -16,6 +16,7 @@ #include #include +#include #include #include "paddle/fluid/framework/tensor.h" @@ -26,6 +27,7 @@ namespace paddle { namespace framework { +#if defined(PADDLE_WITH_CUDA) // Vector implements the std::vector interface, and can get Data or // MutableData from any place. The data will be synced implicitly inside. template @@ -37,11 +39,11 @@ class Vector { Vector() { InitEmpty(); } // Fill vector with value. The vector size is `count`. - explicit Vector(size_t count, const T& value = T()) { + explicit Vector(size_t count, const T &value = T()) { InitEmpty(); if (count != 0) { resize(count); - T* ptr = begin(); + T *ptr = begin(); for (size_t i = 0; i < count; ++i) { ptr[i] = value; } @@ -59,7 +61,7 @@ class Vector { // implicit cast from std::vector. template - Vector(const std::vector& dat) { // NOLINT + Vector(const std::vector &dat) { // NOLINT if (dat.size() == 0) { InitEmpty(); } else { @@ -68,10 +70,10 @@ class Vector { } // Copy ctor - Vector(const Vector& other) { this->operator=(other); } + Vector(const Vector &other) { this->operator=(other); } // Copy operator - Vector& operator=(const Vector& other) { + Vector &operator=(const Vector &other) { if (other.size() != 0) { this->InitByIter(other.size(), other.begin(), other.end()); } else { @@ -81,7 +83,7 @@ class Vector { } // Move ctor - Vector(Vector&& other) { + Vector(Vector &&other) { this->size_ = other.size_; this->flag_ = other.flag_; if (other.cuda_vec_.memory_size()) { @@ -93,13 +95,13 @@ class Vector { } // CPU data access method. Mutable. - T& operator[](size_t i) { + T &operator[](size_t i) { MutableCPU(); - return const_cast(cpu_vec_.data())[i]; + return const_cast(cpu_vec_.data())[i]; } // CPU data access method. Immutable. - const T& operator[](size_t i) const { + const T &operator[](size_t i) const { ImmutableCPU(); return cpu_vec_.data()[i]; } @@ -107,43 +109,43 @@ class Vector { // std::vector iterator methods. Based on CPU data access method size_t size() const { return size_; } - T* begin() { return capacity() == 0 ? &EmptyDummy() : &this->operator[](0); } + T *begin() { return capacity() == 0 ? &EmptyDummy() : &this->operator[](0); } - T* end() { + T *end() { return capacity() == 0 ? &EmptyDummy() : &this->operator[](size()); } - T& front() { return *begin(); } + T &front() { return *begin(); } - T& back() { + T &back() { auto it = end(); --it; return *it; } - const T* begin() const { + const T *begin() const { return capacity() == 0 ? &EmptyDummy() : &this->operator[](0); } - const T* end() const { + const T *end() const { return capacity() == 0 ? &EmptyDummy() : &this->operator[](size()); } - const T* cbegin() const { return begin(); } + const T *cbegin() const { return begin(); } - const T* cend() const { return end(); } + const T *cend() const { return end(); } - const T& back() const { + const T &back() const { auto it = end(); --it; return *it; } - T* data() { return begin(); } + T *data() { return begin(); } - const T* data() const { return begin(); } + const T *data() const { return begin(); } - const T& front() const { return *begin(); } + const T &front() const { return *begin(); } // end of std::vector iterator methods // assign this from iterator. @@ -169,7 +171,7 @@ class Vector { void Extend(It begin, It end) { size_t pre_size = size_; resize(pre_size + (end - begin)); - T* ptr = this->begin() + pre_size; + T *ptr = this->begin() + pre_size; for (; begin < end; ++begin, ++ptr) { *ptr = *begin; } @@ -183,9 +185,9 @@ class Vector { MutableCPU(); Tensor cpu_tensor; platform::Place cpu = platform::CPUPlace(); - T* ptr = cpu_tensor.mutable_data( + T *ptr = cpu_tensor.mutable_data( framework::make_ddim({static_cast(size)}), cpu); - const T* old_ptr = + const T *old_ptr = cpu_vec_.memory_size() == 0 ? nullptr : cpu_vec_.data(); if (old_ptr != nullptr) { std::copy(old_ptr, old_ptr + size_, ptr); @@ -196,7 +198,7 @@ class Vector { } // get cuda ptr. immutable - const T* CUDAData(platform::Place place) const { + const T *CUDAData(platform::Place place) const { PADDLE_ENFORCE(platform::is_gpu_place(place), "CUDA Data must on CUDA place"); ImmutableCUDA(place); @@ -204,10 +206,10 @@ class Vector { } // get cuda ptr. mutable - T* CUDAMutableData(platform::Place place) { - const T* ptr = CUDAData(place); + T *CUDAMutableData(platform::Place place) { + const T *ptr = CUDAData(place); flag_ = kDirty | kDataInCUDA; - return const_cast(ptr); + return const_cast(ptr); } // clear @@ -228,7 +230,7 @@ class Vector { } // the unify method to access CPU or CUDA data. immutable. - const T* Data(platform::Place place) const { + const T *Data(platform::Place place) const { if (platform::is_gpu_place(place)) { return CUDAData(place); } else { @@ -237,7 +239,7 @@ class Vector { } // the unify method to access CPU or CUDA data. mutable. - T* MutableData(platform::Place place) { + T *MutableData(platform::Place place) { if (platform::is_gpu_place(place)) { return CUDAMutableData(place); } else { @@ -253,7 +255,7 @@ class Vector { return result; } - bool operator==(const Vector& other) const { + bool operator==(const Vector &other) const { if (size() != other.size()) return false; auto it1 = cbegin(); auto it2 = other.cbegin(); @@ -274,7 +276,7 @@ class Vector { template void InitByIter(size_t size, Iter begin, Iter end) { platform::Place cpu = platform::CPUPlace(); - T* ptr = this->cpu_vec_.template mutable_data( + T *ptr = this->cpu_vec_.template mutable_data( framework::make_ddim({static_cast(size)}), cpu); for (size_t i = 0; i < size; ++i) { *ptr++ = *begin++; @@ -368,7 +370,7 @@ class Vector { } } - static T& EmptyDummy() { + static T &EmptyDummy() { static T dummy = T(); return dummy; } @@ -379,5 +381,52 @@ class Vector { size_t size_; }; -} // namespace framework +#else // PADDLE_WITH_CUDA + +template +class CPUVector : public std::vector> { + public: + CPUVector() : std::vector() {} + CPUVector(size_t count, const T &value = T()) // NOLINT + : std::vector(count, value) {} + CPUVector(std::initializer_list init) : std::vector(init) {} + CPUVector(const std::vector &other) : std::vector(other) {} // NOLINT + CPUVector(const CPUVector &other) : std::vector(other) {} + CPUVector(CPUVector &&other) : std::vector(std::move(other)) {} + CPUVector(std::vector &&other) // NOLINT + : std::vector(std::move(other)) {} + CPUVector &operator=(const CPUVector &other) { + this->assign(other.begin(), other.end()); + return *this; + } + CPUVector &operator=(const std::vector &other) { + this->assign(other.begin(), other.end()); + return *this; + } + + friend std::ostream &operator<<(std::ostream &os, const CPUVector &other) { + std::stringstream ss; + for (auto v : other) { + os << v << " "; + } + return os; + } + + T &operator[](size_t id) { return this->at(id); } + + const T &operator[](size_t id) const { return this->at(id); } + + template + void Extend(const D &begin, const D &end) { + this->reserve(this->size() + size_t(end - begin)); + this->insert(this->end(), begin, end); + } +}; + +template +using Vector = CPUVector; + +#endif // PADDLE_WITH_CUDA + +}; // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/mixed_vector_test.cc b/paddle/fluid/framework/mixed_vector_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..0599c8d384641606b0a5ebb5ba1781b56f539e63 --- /dev/null +++ b/paddle/fluid/framework/mixed_vector_test.cc @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include + +#include "glog/logging.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/mixed_vector.h" + +template +using vec = paddle::framework::Vector; + +TEST(mixed_vector, CPU_VECTOR) { + vec tmp; + for (int i = 0; i < 10; ++i) { + tmp.push_back(i); + } + ASSERT_EQ(tmp.size(), 10UL); + vec tmp2; + tmp2 = tmp; + ASSERT_EQ(tmp2.size(), 10UL); + for (int i = 0; i < 10; ++i) { + ASSERT_EQ(tmp2[i], i); + ASSERT_EQ(tmp2[i], tmp[i]); + } + int cnt = 0; + for (auto& t : tmp2) { + ASSERT_EQ(t, cnt); + ++cnt; + } +} + +TEST(mixed_vector, InitWithCount) { + paddle::framework::Vector vec(10, 10); + for (int i = 0; i < 10; ++i) { + ASSERT_EQ(vec[i], 10); + } +} + +TEST(mixed_vector, ForEach) { + vec tmp; + for (auto& v : tmp) { + VLOG(3) << v; + } +} + +TEST(mixed_vector, Reserve) { + paddle::framework::Vector vec; + vec.reserve(1); + vec.push_back(0); + vec.push_back(0); + vec.push_back(0); +} + +TEST(mixed_vector, Resize) { + paddle::framework::Vector vec; + vec.resize(1); + vec.push_back(0); + vec.push_back(0); + vec.push_back(0); +} diff --git a/paddle/fluid/framework/mixed_vector_test.cu b/paddle/fluid/framework/mixed_vector_test.cu index d57f82510833d6a0cea7009cf1f0b49543812f8d..4b0caa8d350dde0462e5fdcca743df919358a364 100644 --- a/paddle/fluid/framework/mixed_vector_test.cu +++ b/paddle/fluid/framework/mixed_vector_test.cu @@ -11,7 +11,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ + #include +#include #include "glog/logging.h" #include "gtest/gtest.h" @@ -21,26 +23,6 @@ template using vec = paddle::framework::Vector; -TEST(mixed_vector, CPU_VECTOR) { - vec tmp; - for (int i = 0; i < 10; ++i) { - tmp.push_back(i); - } - ASSERT_EQ(tmp.size(), 10UL); - vec tmp2; - tmp2 = tmp; - ASSERT_EQ(tmp2.size(), 10UL); - for (int i = 0; i < 10; ++i) { - ASSERT_EQ(tmp2[i], i); - ASSERT_EQ(tmp2[i], tmp[i]); - } - int cnt = 0; - for (auto& t : tmp2) { - ASSERT_EQ(t, cnt); - ++cnt; - } -} - static __global__ void multiply_10(int* ptr) { for (int i = 0; i < 10; ++i) { ptr[i] *= 10; @@ -91,24 +73,3 @@ TEST(mixed_vector, MultiGPU) { ASSERT_EQ(tmp[i], i * 100); } } - -TEST(mixed_vector, InitWithCount) { - paddle::framework::Vector vec(10, 10); - for (int i = 0; i < 10; ++i) { - ASSERT_EQ(vec[i], 10); - } -} - -TEST(mixed_vector, ForEach) { - vec tmp; - for (auto& v : tmp) { - } -} - -TEST(mixed_vector, Reserve) { - paddle::framework::Vector vec; - vec.reserve(1); - vec.push_back(0); - vec.push_back(0); - vec.push_back(0); -} diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 09b67e5a1741c68c5f5487340e8fc86ff31e00a4..03f7e71c03b8dd75d2a47cb4c6d1ef1a71792cf3 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -103,7 +103,7 @@ void OpDesc::CopyFrom(const OpDesc &op_desc) { need_update_ = true; } -OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block) +OpDesc::OpDesc(const proto::OpDesc &desc, BlockDesc *block) : desc_(desc), need_update_(false) { // restore inputs_ int input_size = desc_.inputs_size(); @@ -211,6 +211,12 @@ void OpDesc::SetBlockAttr(const std::string &name, BlockDesc *block) { need_update_ = true; } +void OpDesc::SetBlocksAttr(const std::string &name, + std::vector blocks) { + this->attrs_[name] = blocks; + need_update_ = true; +} + void OpDesc::SetAttrMap( const std::unordered_map &attr_map) { attrs_ = attr_map; @@ -232,7 +238,20 @@ Attribute OpDesc::GetNullableAttr(const std::string &name) const { } } -int OpDesc::GetBlockAttr(const std::string &name) const { +std::vector OpDesc::GetBlocksAttrIds(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + auto blocks = boost::get>(it->second); + + std::vector ids; + for (auto n : blocks) { + ids.push_back(n->ID()); + } + + return ids; +} + +int OpDesc::GetBlockAttrId(const std::string &name) const { auto it = attrs_.find(name); PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); return boost::get(it->second)->ID(); @@ -305,6 +324,13 @@ struct SetAttrDescVisitor : public boost::static_visitor { void operator()(const std::vector &v) const { VectorToRepeated(v, attr_->mutable_bools()); } + void operator()(const std::vector &v) const { + std::vector blocks_idx; + for (auto blk : v) { + blocks_idx.push_back(blk->ID()); + } + VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx()); + } void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); } void operator()(int64_t v) const { attr_->set_l(v); } void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index 1a330db7cc5555a939950043ac90a321573b292d..b77d84125a23b81c3de4123bea6f0e09cd6d1e90 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -33,13 +33,14 @@ class OpDesc { OpDesc(const std::string &type, const VariableNameMap &inputs, const VariableNameMap &outputs, const AttributeMap &attrs); - OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block); + OpDesc(const proto::OpDesc &desc, BlockDesc *block); explicit OpDesc(BlockDesc *block) : block_(block) {} OpDesc(const OpDesc &other, BlockDesc *block) { *this = other; block_ = block; + need_update_ = true; } void CopyFrom(const OpDesc &op_desc); @@ -76,11 +77,15 @@ class OpDesc { void SetBlockAttr(const std::string &name, BlockDesc *block); + void SetBlocksAttr(const std::string &name, std::vector blocks); + Attribute GetAttr(const std::string &name) const; Attribute GetNullableAttr(const std::string &name) const; - int GetBlockAttr(const std::string &name) const; + int GetBlockAttrId(const std::string &name) const; + + std::vector GetBlocksAttrIds(const std::string &name) const; void Rename(const std::string &old_name, const std::string &new_name); diff --git a/paddle/fluid/framework/op_info.cc b/paddle/fluid/framework/op_info.cc index b99e82f8c4358b60a014c6fc7c61c9bbb8683834..af75baa5c4b98f7d092834c05eb57e9c7e131b29 100644 --- a/paddle/fluid/framework/op_info.cc +++ b/paddle/fluid/framework/op_info.cc @@ -17,13 +17,12 @@ limitations under the License. */ namespace paddle { namespace framework { -static OpInfoMap* g_op_info_map = nullptr; - +// C++11 removes the need for manual locking. Concurrent execution shall wait if +// a static local variable is already being initialized. +// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex OpInfoMap& OpInfoMap::Instance() { - if (g_op_info_map == nullptr) { - g_op_info_map = new OpInfoMap(); - } - return *g_op_info_map; + static OpInfoMap g_op_info_map; + return g_op_info_map; } } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_kernel_type.h b/paddle/fluid/framework/op_kernel_type.h index fab20d75f5a45257f243333c1998d7b2549a25f9..c59b232191c49ccb47bb9f51dcaf2fd9280fae19 100644 --- a/paddle/fluid/framework/op_kernel_type.h +++ b/paddle/fluid/framework/op_kernel_type.h @@ -87,10 +87,17 @@ inline std::string KernelTypeToString(const OpKernelType& kernel_key) { } inline bool NeedTransformLayout(const DataLayout& l, const DataLayout& r) { - return l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r; + bool ret = + (l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r); +#ifdef PADDLE_WITH_MKLDNN + // Layout transform needed for either non-MKLDNN to MKLDNN or vice versa + ret |= (l != DataLayout::kMKLDNN && r == DataLayout::kMKLDNN); + ret |= (l == DataLayout::kMKLDNN && r != DataLayout::kMKLDNN); +#endif + return ret; } -inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) { +inline bool NeedTransform(const OpKernelType& l, const OpKernelType& r) { return (!platform::places_are_same_class(l.place_, r.place_)) || (l.data_type_ != r.data_type_) || NeedTransformLayout(l.data_layout_, r.data_layout_); diff --git a/paddle/fluid/framework/op_kernel_type_test.cc b/paddle/fluid/framework/op_kernel_type_test.cc index db95861c510b52a5b52229541434e6437d3fb9f4..3e17a512ce154de88ac890f3b29f03385595d95c 100644 --- a/paddle/fluid/framework/op_kernel_type_test.cc +++ b/paddle/fluid/framework/op_kernel_type_test.cc @@ -29,6 +29,13 @@ TEST(OpKernelType, ToString) { ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type), "data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type[" "CUDNN]"); + + using CUDAPlace = paddle::platform::CUDAPlace; + OpKernelType op_kernel_type2(DataType::FP16, CUDAPlace(0), DataLayout::kNCHW, + LibraryType::kCUDNN); + ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type2), + "data_type[float16]:data_layout[NCHW]:place[CUDAPlace(0)]:library_" + "type[CUDNN]"); } TEST(OpKernelType, Hash) { diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc index ae9f4efd44acdcdff2806deea6826e4089459a78..2288c7fe6609a765612b468d69ad35101b92b384 100644 --- a/paddle/fluid/framework/op_proto_maker.cc +++ b/paddle/fluid/framework/op_proto_maker.cc @@ -21,6 +21,7 @@ namespace framework { void OpProtoAndCheckerMaker::Validate() { validated_ = true; CheckNoDuplicatedInOutAttrs(); + CheckReuseVars(); } OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput( @@ -39,6 +40,40 @@ OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput( return OpProtoAndCheckerMaker::VariableBuilder{output}; } +void OpProtoAndCheckerMaker::Reuse(const std::string& name, + const std::string& reused_name) { + bool found = false; + proto::OpProto::Var* var; + + for (auto& var : proto_->inputs()) { + if (var.name() == reused_name) { + found = true; + break; + } + } + PADDLE_ENFORCE(found == true, + "Input/Output name: %s reused_name: %s, one of them is not " + "exists or not matched.", + name, reused_name); + + found = false; + for (int i = 0; i < proto_->outputs().size(); ++i) { + var = proto_->mutable_outputs()->Mutable(i); + if (var->name() == name) { + PADDLE_ENFORCE(!var->has_reuse(), + "Output(%s) has been set reused var of %s", name, + var->reuse()); + found = true; + var->set_reuse(reused_name); + break; + } + } + PADDLE_ENFORCE(found == true, + "Input/Output name: %s reused_name: %s, one of them is not " + "exists or not matched.", + name, reused_name); +} + void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { std::unordered_set names; auto checker = [&](const std::string& name) { @@ -56,6 +91,24 @@ void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { } } +void OpProtoAndCheckerMaker::CheckReuseVars() { + std::unordered_set names; + for (auto& input : proto_->inputs()) { + names.insert(input.name()); + } + auto checker = [&](const std::string& name, const std::string& reused) { + PADDLE_ENFORCE( + names.count(reused), + "Output [%s] reuse Input [%s], but the input is not registered.", name, + reused); + }; + for (auto& output : proto_->outputs()) { + if (output.has_reuse()) { + checker(output.name(), output.reuse()); + } + } +} + void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto, OpAttrChecker* attr_checker) { proto_ = proto; diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 8493b9d8b326c71a33b95bf95e5fc1743c686eb7..80970291c9c234f1306162f4ffa3c2528f88c35f 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once #include +#include + #include "glog/logging.h" #include "paddle/fluid/framework/attribute.h" #include "paddle/fluid/framework/framework.pb.h" @@ -64,6 +66,11 @@ class OpProtoAndCheckerMaker { var_->set_dispensable(true); return *this; } + + VariableBuilder &Reuse(const std::string &name) { + var_->set_reuse(name); + return *this; + } }; VariableBuilder AddInput(const std::string &name, const std::string &comment); @@ -71,6 +78,8 @@ class OpProtoAndCheckerMaker { VariableBuilder AddOutput(const std::string &name, const std::string &comment); + void Reuse(const std::string &name, const std::string &reused_name); + template TypedAttrChecker &AddAttr(const std::string &name, const std::string &comment, @@ -89,6 +98,8 @@ class OpProtoAndCheckerMaker { void CheckNoDuplicatedInOutAttrs(); void Validate(); + void CheckReuseVars(); + proto::OpProto *proto_; OpAttrChecker *op_checker_; bool validated_{false}; diff --git a/paddle/fluid/framework/op_proto_maker_test.cc b/paddle/fluid/framework/op_proto_maker_test.cc index a8030d377fdb4d4aef74b315e21792dad10fac96..b71c7b646857e11f291748c4c7c2af92b6d53231 100644 --- a/paddle/fluid/framework/op_proto_maker_test.cc +++ b/paddle/fluid/framework/op_proto_maker_test.cc @@ -47,3 +47,120 @@ TEST(ProtoMaker, DuplicatedInOut) { ASSERT_THROW(proto_maker(&op_proto, &op_checker), paddle::platform::EnforceNotMet); } + +class TestInplaceProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "input of test op"); + AddOutput("XOut", "output of test op").Reuse("X"); + } +}; + +class TestInplaceProtoMaker2 + : public paddle::framework::OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "input of test op"); + AddOutput("XOut", "output of test op").Reuse("X"); + AddOutput("NoOut", "output of test op").Reuse("NotExists"); + } +}; + +TEST(ProtoMaker, InplaceOutput) { + paddle::framework::proto::OpProto op_proto, op_proto2; + paddle::framework::OpAttrChecker op_checker; + TestInplaceProtoMaker proto_maker; + TestInplaceProtoMaker2 proto_maker2; + + proto_maker(&op_proto, &op_checker); + + ASSERT_THROW(proto_maker2(&op_proto2, &op_checker), + paddle::platform::EnforceNotMet); +} + +// normal reuse +class TestReuseProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "input of test op"); + AddInput("Y", "input of test op"); + AddOutput("Out", "output of test op"); + AddOutput("XOut", "output of test op"); + // avoid destructor exception. + // Validate(); + TestReuse(); + } + + virtual void TestReuse() {} +}; + +// test duplicate reuse error +class TestReuseProtoMaker2 : public TestReuseProtoMaker { + public: + void TestReuse() { + Reuse("Out", "X"); + Reuse("Out", "Y"); + } +}; + +// NotExists Input +class TestReuseProtoMaker3 : public TestReuseProtoMaker { + public: + void TestReuse() { + Reuse("Out", "NotExists"); + Reuse("XOut", "X"); + } +}; + +// NotExists Output +class TestReuseProtoMaker4 : public TestReuseProtoMaker { + public: + void TestReuse() { Reuse("NotExists", "X"); } +}; + +TEST(ProtoMaker, Reuse) { + paddle::framework::proto::OpProto op_proto; + paddle::framework::OpAttrChecker op_checker; + TestReuseProtoMaker proto_maker; + proto_maker(&op_proto, &op_checker); +} + +// NOTE(dzhwinter): +// There is a Fatal CHECK on base class destructor, which will call abort inside +// instead of +// throw an exception. If we throw an exception in Make(), we will trigger the +// CHECK and terminate the tests. +// +// I had tried to replace the default CHECK with a exception, however, it's +// still not supported by glog. +// the details: +// https://github.com/google/glog/issues/249 +// https://github.com/facebookresearch/TensorComprehensions/issues/351 +/* +TEST(ProtoMaker, ReuseWithException) { + paddle::framework::proto::OpProto op_proto2, op_proto3, op_proto4; + paddle::framework::OpAttrChecker op_checker; + TestReuseProtoMaker2 proto_maker2; + TestReuseProtoMaker3 proto_maker3; + TestReuseProtoMaker4 proto_maker4; + EXPECT_THROW(proto_maker2(&op_proto2, &op_checker), + paddle::platform::EnforceNotMet); + + EXPECT_THROW(proto_maker3(&op_proto3, &op_checker), + paddle::platform::EnforceNotMet); + + EXPECT_THROW(proto_maker4(&op_proto4, &op_checker), + paddle::platform::EnforceNotMet); +} + +void FailureFunction() { + throw std::runtime_error("Check failed in destructor."); + // return 0; +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + google::InstallFailureFunction(&FailureFunction); + return RUN_ALL_TESTS(); +} +*/ diff --git a/paddle/fluid/framework/op_registry.h b/paddle/fluid/framework/op_registry.h index 748317438b44bc4af84f13b25f8e4f88386388fb..e7dfa608b48f89a2155e43c7e63e31154675cd38 100644 --- a/paddle/fluid/framework/op_registry.h +++ b/paddle/fluid/framework/op_registry.h @@ -76,6 +76,20 @@ class OpRegistry { template struct OpKernelRegistrarFunctor; +template +inline void RegisterKernelClass(const char* op_type, const char* library_type, + Func func) { + std::string library(library_type); + std::string data_layout = "ANYLAYOUT"; + if (library == "MKLDNN") { + data_layout = "MKLDNNLAYOUT"; + } + OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(), + StringToDataLayout(data_layout), + StringToLibraryType(library_type)); + OperatorWithKernel::AllOpKernels()[op_type][key] = func; +} + template struct OpKernelRegistrarFunctor { using KERNEL_TYPE = @@ -83,10 +97,10 @@ struct OpKernelRegistrarFunctor { void operator()(const char* op_type, const char* library_type) const { using T = typename KERNEL_TYPE::ELEMENT_TYPE; - OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(), - DataLayout::kAnyLayout, StringToLibraryType(library_type)); - OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE); - + RegisterKernelClass( + op_type, library_type, [](const framework::ExecutionContext& ctx) { + KERNEL_TYPE().Compute(ctx); + }); constexpr auto size = std::tuple_size>::value; OpKernelRegistrarFunctor func; @@ -99,7 +113,8 @@ struct OpKernelRegistrarFunctor { void operator()(const char* op_type, const char* library_type) const {} }; -// User can register many kernel in one place. The data type could be different. +// User can register many kernel in one place. The data type could be +// different. template class OpKernelRegistrar : public Registrar { public: @@ -109,6 +124,47 @@ class OpKernelRegistrar : public Registrar { } }; +template +struct OpKernelRegistrarFunctorEx; + +template +class OpKernelRegistrarEx : public Registrar { + public: + explicit OpKernelRegistrarEx(const char* op_type, const char* library_type) { + OpKernelRegistrarFunctorEx + func; + func(op_type, library_type); + } +}; + +template +struct OpKernelRegistrarFunctorEx { + void operator()(const char* op_type, const char* library_type) const {} +}; + +template +struct OpKernelRegistrarFunctorEx { + using Functor = + typename std::tuple_element>::type; + using T = + typename std::tuple_element>::type; + + void operator()(const char* op_type, const char* library_type) const { + RegisterKernelClass(op_type, library_type, Functor()); + + constexpr auto size = + std::tuple_size>::value; + OpKernelRegistrarFunctorEx= size, I + 2, + DataTypeAndKernelType...> + func; + func(op_type, library_type); + } +}; + /** * check if MACRO is used in GLOBAL NAMESPACE. */ @@ -126,21 +182,15 @@ class OpKernelRegistrar : public Registrar { VarTypeInference InferShapeBase */ -#define REGISTER_OPERATOR(op_type, op_class, ...) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op__##op_type, \ - "REGISTER_OPERATOR must be called in global namespace"); \ - class _OpClass_##op_type##_ : public op_class { \ - public: \ - DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \ - DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \ - }; \ - static ::paddle::framework::OperatorRegistrar<_OpClass_##op_type##_, \ - ##__VA_ARGS__> \ - __op_registrar_##op_type##__(#op_type); \ - int TouchOpRegistrar_##op_type() { \ - __op_registrar_##op_type##__.Touch(); \ - return 0; \ +#define REGISTER_OPERATOR(op_type, op_class, ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op__##op_type, \ + "REGISTER_OPERATOR must be called in global namespace"); \ + static ::paddle::framework::OperatorRegistrar \ + __op_registrar_##op_type##__(#op_type); \ + int TouchOpRegistrar_##op_type() { \ + __op_registrar_##op_type##__.Touch(); \ + return 0; \ } #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \ @@ -149,15 +199,15 @@ class OpKernelRegistrar : public Registrar { /** * Macro to register OperatorKernel. */ -#define REGISTER_OP_KERNEL(op_type, LIBRARY_TYPE, place_class, ...) \ +#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...) \ STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op_kernel_##op_type##_##LIBRARY_TYPE##__, \ + __reg_op_kernel_##op_type##_##library_type##__, \ "REGISTER_OP_KERNEL must be called in global namespace"); \ static ::paddle::framework::OpKernelRegistrar \ - __op_kernel_registrar_##op_type##_##LIBRARY_TYPE##__(#op_type, \ - #LIBRARY_TYPE); \ - int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE() { \ - __op_kernel_registrar_##op_type##_##LIBRARY_TYPE##__.Touch(); \ + __op_kernel_registrar_##op_type##_##library_type##__(#op_type, \ + #library_type); \ + int TouchOpKernelRegistrar_##op_type##_##library_type() { \ + __op_kernel_registrar_##op_type##_##library_type##__.Touch(); \ return 0; \ } @@ -167,6 +217,25 @@ class OpKernelRegistrar : public Registrar { #define REGISTER_OP_CPU_KERNEL(op_type, ...) \ REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) +#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class, ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op_kernel_##op_type##_##library_type##__, \ + "REGISTER_OP_KERNEL_EX must be called in global namespace"); \ + static ::paddle::framework::OpKernelRegistrarEx \ + __op_kernel_registrar_##op_type##_##library_type##__(#op_type, \ + #library_type); \ + int TouchOpKernelRegistrar_##op_type##_##library_type() { \ + __op_kernel_registrar_##op_type##_##library_type##__.Touch(); \ + return 0; \ + } + +#define REGISTER_OP_CUDA_KERNEL_FUNCTOR(op_type, ...) \ + REGISTER_OP_KERNEL_EX(op_type, CUDA, ::paddle::platform::CUDAPlace, \ + __VA_ARGS__) + +#define REGISTER_OP_CPU_KERNEL_FUNCTOR(op_type, ...) \ + REGISTER_OP_KERNEL_EX(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) + /** * Macro to mark what Operator and Kernel * we will use and tell the compiler to diff --git a/paddle/fluid/framework/op_registry_test.cc b/paddle/fluid/framework/op_registry_test.cc index 18b1649cc71d5edd5b07740bbad1fe8f81128898..04996d7b09cecc3c330a47153c9b10310f1792f4 100644 --- a/paddle/fluid/framework/op_registry_test.cc +++ b/paddle/fluid/framework/op_registry_test.cc @@ -193,15 +193,10 @@ TEST(OpRegistry, CustomChecker) { ASSERT_EQ(test_attr, 4); } -class CosineOpComplete : public paddle::framework::CosineOp { - public: - DEFINE_OP_CONSTRUCTOR(CosineOpComplete, paddle::framework::CosineOp); - DEFINE_OP_CLONE_METHOD(CosineOpComplete); -}; - TEST(OperatorRegistrar, Test) { paddle::framework::OperatorRegistrar< - CosineOpComplete, paddle::framework::CosineOpProtoAndCheckerMaker> + paddle::framework::CosineOp, + paddle::framework::CosineOpProtoAndCheckerMaker> reg("cos"); } diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index f87d5521492418d2daf5b7fba1500c4bb31e10f5..d04f7744961b2561977f4d36d0073a97557043da 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/framework/data_transform.h" #include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/shape_inference.h" #include "paddle/fluid/framework/var_type.h" @@ -57,7 +58,11 @@ static DDim GetDims(const Scope& scope, const std::string& name, } if (var->IsType()) { - return var->Get().dims(); + const LoDTensor& tensor = var->Get(); + if (UNLIKELY(!tensor.IsInitialized())) { + return DDim({-1}); + } + return tensor.dims(); } else if (var->IsType()) { if (get_actual_dim) { return var->Get().value().dims(); @@ -69,6 +74,39 @@ static DDim GetDims(const Scope& scope, const std::string& name, } } +static std::string GetDtype(const Scope& scope, const std::string& name) { + Variable* var = scope.FindVar(name); + if (var == nullptr) { + return ""; + } + + if (var->IsType()) { + const LoDTensor& tensor = var->Get(); + if (UNLIKELY(!tensor.IsInitialized())) { + return ""; + } + return DataTypeToString(ToDataType(tensor.type())); + } else if (var->IsType()) { + return DataTypeToString( + ToDataType(var->Get().value().type())); + } else { + return ""; + } +} + +static int GetRowSize(const Scope& scope, const std::string& name) { + Variable* var = scope.FindVar(name); + if (var == nullptr) { + return -1; + } + + if (var->IsType()) { + return var->Get().rows().size(); + } + + return -1; +} + static LoD GetLoD(const Scope& scope, const std::string& name) { Variable* var = scope.FindVar(name); auto default_lod = LoD({{}}); @@ -78,13 +116,18 @@ static LoD GetLoD(const Scope& scope, const std::string& name) { } if (var->IsType()) { - return var->Get().lod(); + const LoDTensor& tensor = var->Get(); + if (UNLIKELY(!tensor.IsInitialized())) { + return default_lod; + } + return tensor.lod(); } else { return default_lod; } } void OperatorBase::Run(const Scope& scope, const platform::Place& place) { + VLOG(4) << place << " " << DebugStringEx(&scope); if (platform::is_gpu_place(place)) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("Cannot run operator on place %s", place); @@ -93,7 +136,10 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { platform::SetDeviceId(dev_id); #endif } + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + platform::RecordEvent record_event(Type(), pool.Get(place)); RunImpl(scope, place); + VLOG(3) << place << " " << DebugStringEx(&scope); } bool OperatorBase::HasInputs(const std::string& name) const { @@ -153,6 +199,12 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { for (size_t i = 0; i < input.second.size(); ++i) { ss << input.second[i]; if (scope) { + int row_size = GetRowSize(*scope, input.second[i]); + if (row_size >= 0) { + ss << "[row_size=" << row_size << "]"; + } + std::string dtype = GetDtype(*scope, input.second[i]); + ss << ":" << dtype; ss << "[" << GetDims(*scope, input.second[i], true) << "]"; ss << "(" << GetLoD(*scope, input.second[i]) << ")"; } @@ -173,6 +225,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { for (size_t i = 0; i < output.second.size(); ++i) { ss << output.second[i]; if (scope) { + int row_size = GetRowSize(*scope, output.second[i]); + if (row_size >= 0) { + ss << "[row_size=" << row_size << "]"; + } ss << "[" << GetDims(*scope, output.second[i], true) << "]"; ss << "(" << GetLoD(*scope, output.second[i]) << ")"; } @@ -293,6 +349,38 @@ static Tensor* GetMutableTensorFromVar(Variable* var) { } } +bool ExecutionContext::HasInput(const std::string& name) const { + if (!op_.HasInputs(name)) { + return false; + } + auto& ins = Inputs(name); + size_t length = ins.size(); + if (length == 0) { + return false; + } + PADDLE_ENFORCE_EQ(length, 1UL, + "Input %s should not have more than one inputs", name); + auto arg = ins[0]; + auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg); + return var != nullptr; +} + +bool ExecutionContext::HasOutput(const std::string& name) const { + if (!op_.HasOutputs(name)) { + return false; + } + auto& outs = Outputs(name); + size_t length = outs.size(); + if (length == 0) { + return false; + } + PADDLE_ENFORCE_EQ(length, 1UL, + "Output %s should not have more than one inputs", name); + auto arg = outs[0]; + auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg); + return var != nullptr; +} + template <> const Tensor* ExecutionContext::Input(const std::string& name) const { auto* var = InputVar(name); @@ -444,10 +532,25 @@ class RuntimeInferShapeContext : public InferShapeContext { auto* out_tensor = out_var->GetMutable(); out_tensor->set_lod(in_tensor.lod()); - // TODO(dzhwinter) : reuse ShareLoD in most operators. - // Need to call ShareLayout explicitly in sequence related ops. - // Shall we have a better method to shared info between in/out Tensor? - out_tensor->set_layout(in_tensor.layout()); +// TODO(dzhwinter) : reuse ShareLoD in most operators. +// Need to call ShareLayout explicitly in sequence related ops. +// Shall we have a better method to shared info between in/out Tensor? +#ifdef PADDLE_WITH_MKLDNN + // Fix me: ugly workaround below + // Correct solution: + // set_layout() should NOT be called here (i.e. ShareLoD). Instead, + // layout of output tensor should be set "manually" in Compute() + // of each OPKernel. The reason layout should NOT be shared between + // input and output "automatically" (now by InferShape()->ShareLoD()) + // is that layout transform may occur after InferShape(). + // Workaround: + // Skip set_layout() when input layout is kMKLDNN + // This is to avoid kMKLDNN is populated wrongly into a non-MKLDNN + // OPKernel. In all MKLDNN OPkernel, set_layout(kMKLDNN) should be called + // in Compute() + if (in_tensor.layout() != DataLayout::kMKLDNN) +#endif + out_tensor->set_layout(in_tensor.layout()); } void ShareLayout(const std::string& in, const std::string& out, size_t i = 0, @@ -522,8 +625,7 @@ static void CheckTensorNANOrInf(const std::string& name, if (tensor.memory_size() == 0) { return; } - if (tensor.type().hash_code() != typeid(float).hash_code() && // NOLINT - tensor.type().hash_code() != typeid(double).hash_code()) { // NOLINT + if (!IsType(tensor.type()) && !IsType(tensor.type())) { return; } PADDLE_ENFORCE(!framework::TensorContainsInf(tensor), @@ -539,9 +641,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope, platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto* dev_ctx = pool.Get(place); - // For profiling, don't move out of this function because that will result - // in the failure of multi-GPU profiling. - platform::RecordEvent record_event(Type(), dev_ctx); // check if op[type] has kernel registered. auto& all_op_kernels = AllOpKernels(); auto kernels_iter = all_op_kernels.find(type_); @@ -550,8 +649,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope, "There are no kernels which are registered in the %s operator.", type_); } - ExecutionContext ctx(*this, scope, *dev_ctx); - OpKernelMap& kernels = kernels_iter->second; // TODO(dzhwinter) : kernel fallback mechanism will be added when all the @@ -561,77 +658,127 @@ void OperatorWithKernel::RunImpl(const Scope& scope, // Do selection // } - auto expected_kernel_key = this->GetExpectedKernelType(ctx); + auto expected_kernel_key = + this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx)); VLOG(3) << "expected_kernel_key:" << expected_kernel_key; auto kernel_iter = kernels.find(expected_kernel_key); +#ifdef PADDLE_WITH_MKLDNN + // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set + if (kernel_iter == kernels.end() && + expected_kernel_key.library_type_ == LibraryType::kMKLDNN) { + VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one"; + expected_kernel_key.library_type_ = LibraryType::kPlain; + expected_kernel_key.data_layout_ = DataLayout::kAnyLayout; + kernel_iter = kernels.find(expected_kernel_key); + } +#endif if (kernel_iter == kernels.end()) { PADDLE_THROW("op %s does not have kernel for %s", type_, KernelTypeToString(expected_kernel_key)); } - // do data transform - Scope& new_scope = scope.NewScope(); + // do data transformScope &transfer_scope; + std::vector transfered_inplace_vars; + auto* transfer_scope = + TryTransferData(scope, expected_kernel_key, &transfered_inplace_vars); - std::vector inplace_vars; - for (auto& var_name_item : this->Inputs()) { - for (auto& var_name : var_name_item.second) { - auto* var = scope.FindVar(var_name); - if (var && VarIsTensor(var)) { - auto* tensor_in = GetTensorFromVar(var); - if (tensor_in->IsInitialized()) { - auto kernel_type_for_var = this->GetKernelTypeForVar( - var_name_item.first, *tensor_in, expected_kernel_key); - if (TransFromNeeded(kernel_type_for_var, expected_kernel_key)) { - auto out_var_names = OutputVars(true); - if (std::find(out_var_names.begin(), out_var_names.end(), - var_name) != out_var_names.end()) { - inplace_vars.push_back(var_name); - } - VLOG(3) << "Transform Variable " << var_name << " from " - << kernel_type_for_var << " to " << expected_kernel_key; - auto* trans_var = new_scope.Var(var_name); - std::shared_ptr out(new Tensor); - DataTransform(expected_kernel_key, kernel_type_for_var, *tensor_in, - out.get()); - CopyVariableWithTensor(*var, *(out.get()), trans_var); - } - } - } - } + // exec scope is the scope that kernel actually executed on. + const Scope& exec_scope = + (transfer_scope == nullptr ? scope : *transfer_scope); + + if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) { + dev_ctx = pool.Get(expected_kernel_key.place_); } - auto* new_dev_ctx = pool.Get(expected_kernel_key.place_); - kernel_iter->second->Compute( - ExecutionContext(*this, new_scope, *new_dev_ctx)); + kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx)); - for (auto& var_name : inplace_vars) { - VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; - auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name)); - auto* transformed_tensor = GetTensorFromVar(new_scope.FindVar(var_name)); - original_tensor->ShareDataWith(*transformed_tensor); + if (!transfered_inplace_vars.empty()) { + // there is inplace variable has been transfered. + TransferInplaceVarsBack(scope, transfered_inplace_vars, *transfer_scope); } /*For profiling/benchmark only*/ if (FLAGS_benchmark) { - new_dev_ctx->Wait(); + dev_ctx->Wait(); } if (FLAGS_check_nan_inf) { for (auto& vname : OutputVars(true)) { - auto* var = new_scope.FindVar(vname); + auto* var = exec_scope.FindVar(vname); if (var == nullptr) continue; if (var->IsType()) { CheckTensorNANOrInf(vname, var->Get()); + } else if (var->IsType()) { + CheckTensorNANOrInf(vname, var->Get().value()); } } } } +void OperatorWithKernel::TransferInplaceVarsBack( + const Scope& scope, const std::vector& inplace_vars, + const Scope& transfer_scope) const { + for (auto& var_name : inplace_vars) { + VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; + auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name)); + auto* transformed_tensor = + GetTensorFromVar(transfer_scope.FindVar(var_name)); + original_tensor->ShareDataWith(*transformed_tensor); + } +} + +Scope* OperatorWithKernel::TryTransferData( + const Scope& scope, const OpKernelType& expected_kernel_key, + std::vector* transfered_inplace_vars) const { + Scope* new_scope = nullptr; + for (auto& var_name_item : Inputs()) { + for (auto& var_name : var_name_item.second) { + auto* var = scope.FindVar(var_name); + // Only tensor can be tranfer to another device. + if (var == nullptr || !VarIsTensor(var)) { + continue; + } + + auto* tensor_in = GetTensorFromVar(var); + if (!tensor_in->IsInitialized()) { + continue; + } + + auto kernel_type_for_var = GetKernelTypeForVar( + var_name_item.first, *tensor_in, expected_kernel_key); + + if (!NeedTransform(kernel_type_for_var, expected_kernel_key)) { + continue; + } + + auto out_var_names = OutputVars(true); + if (std::find(out_var_names.begin(), out_var_names.end(), var_name) != + out_var_names.end()) { + transfered_inplace_vars->emplace_back(var_name); + } + + VLOG(3) << "Transform Variable " << var_name << " from " + << kernel_type_for_var << " to " << expected_kernel_key; + + if (new_scope == nullptr) { + new_scope = &scope.NewScope(); + } + + auto* trans_var = new_scope->Var(var_name); + Tensor out; + TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out); + SetTensorToVariable(*var, out, trans_var); + } + } + + return new_scope; +} proto::VarType::Type OperatorWithKernel::IndicateDataType( const ExecutionContext& ctx) const { auto& scope = ctx.scope(); int data_type = -1; + std::string last_input_name; for (auto& input : this->inputs_) { for (auto& ipt_name : input.second) { auto* var = scope.FindVar(ipt_name); @@ -646,9 +793,12 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType( } if (t != nullptr) { int tmp = static_cast(ToDataType(t->type())); - PADDLE_ENFORCE(tmp == data_type || data_type == -1, - "DataType of Paddle Op %s must be the same.", Type()); + PADDLE_ENFORCE( + tmp == data_type || data_type == -1, + "DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)", + Type(), last_input_name, data_type, ipt_name, tmp); data_type = tmp; + last_input_name = ipt_name; } } } @@ -665,7 +815,8 @@ OpKernelType OperatorWithKernel::GetExpectedKernelType( OpKernelType OperatorWithKernel::GetKernelTypeForVar( const std::string& var_name, const Tensor& tensor, const OpKernelType& expected_kernel_type) const { - return OpKernelType(expected_kernel_type.data_type_, tensor.place()); + return OpKernelType(expected_kernel_type.data_type_, tensor.place(), + tensor.layout()); } } // namespace framework diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 2f480e00c100d579e100de17d3feb957f5ef6167..1040eb882baea624e972faf4af3094119df72308 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -121,10 +121,6 @@ class OperatorBase { //! Get all outputs variable names virtual std::vector OutputVars(bool has_intermediate) const; - // Return a new operator instance, which is as same as this. - // Use unique_ptr to prevent caller forget to delete this pointer. - virtual std::unique_ptr Clone() const = 0; - protected: std::string type_; // NOTE: in case of OpGrad, inputs_ contains: @@ -145,37 +141,6 @@ class OperatorBase { const platform::Place& place) const = 0; }; -// Macro for define a clone method. -// If you are writing an kernel operator, `Clone` will be defined when you -// register it. i.e. `Clone` method is not needed to define by yourself. -#define DEFINE_OP_CLONE_METHOD(cls) \ - std::unique_ptr<::paddle::framework::OperatorBase> Clone() const final { \ - return std::unique_ptr<::paddle::framework::OperatorBase>(new cls(*this)); \ - } - -// Macro for define a default constructor for Operator. -// You can also use -// using PARENT_CLASS::PARENT_CLASS; -// to use parent's constructor. -#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls) \ - cls(const std::string& type, \ - const ::paddle::framework::VariableNameMap& inputs, \ - const ::paddle::framework::VariableNameMap& outputs, \ - const paddle::framework::AttributeMap& attrs) \ - : parent_cls(type, inputs, outputs, attrs) {} - -class NOP : public OperatorBase { - public: - using OperatorBase::OperatorBase; - std::unique_ptr Clone() const override { - return std::unique_ptr(new NOP(*this)); - } - - private: - void RunImpl(const Scope& scope, - const platform::Place& place) const override {} -}; - class ExecutionContext { public: ExecutionContext(const OperatorBase& op, const Scope& scope, @@ -191,9 +156,9 @@ class ExecutionContext { return op_.Attr(name); } - bool HasInput(const std::string& name) const { return op_.HasInputs(name); } + bool HasInput(const std::string& name) const; - bool HasOutput(const std::string& name) const { return op_.HasOutputs(name); } + bool HasOutput(const std::string& name) const; size_t InputSize(const std::string& name) const { return op_.Inputs(name).size(); @@ -347,9 +312,9 @@ class OpKernel : public OpKernelBase { class OperatorWithKernel : public OperatorBase { public: + using OpKernelFunc = std::function; using OpKernelMap = - std::unordered_map, - OpKernelType::Hash>; + std::unordered_map; OperatorWithKernel(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) @@ -384,6 +349,20 @@ class OperatorWithKernel : public OperatorBase { // same. proto::VarType::Type IndicateDataType(const ExecutionContext& ctx) const; void RunImpl(const Scope& scope, const platform::Place& place) const final; + + /** + * Transfer data from scope to a transfered scope. If there is no data need to + * be tranfered, it returns nullptr. + * + * * transfered_inplace_vars is a output vector. + */ + Scope* TryTransferData( + const Scope& scope, const OpKernelType& expected_kernel_key, + std::vector* transfered_inplace_vars) const; + + void TransferInplaceVarsBack(const Scope& scope, + const std::vector& inplace_vars, + const Scope& exec_scope) const; }; extern bool OpSupportGPU(const std::string& op_type); diff --git a/paddle/fluid/framework/operator_test.cc b/paddle/fluid/framework/operator_test.cc index 74043b5d7990178976baf2fad991ae03f9c8dd25..ac9dd8245ad4e0e8842f219b23d3866b03fdaedb 100644 --- a/paddle/fluid/framework/operator_test.cc +++ b/paddle/fluid/framework/operator_test.cc @@ -13,10 +13,10 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "gtest/gtest.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/init.h" namespace paddle { namespace framework { @@ -247,26 +247,3 @@ TEST(OpKernel, multi_inputs) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); op->Run(scope, cpu_place); } - -class OperatorClone : public paddle::framework::OperatorBase { - public: - DEFINE_OP_CLONE_METHOD(OperatorClone); - OperatorClone(const std::string& type, - const paddle::framework::VariableNameMap& inputs, - const paddle::framework::VariableNameMap& outputs, - const paddle::framework::AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - private: - void RunImpl(const paddle::framework::Scope& scope, - const paddle::platform::Place& place) const override {} -}; - -TEST(Operator, Clone) { - paddle::framework::InitDevices(true); - OperatorClone a("ABC", paddle::framework::VariableNameMap{}, - paddle::framework::VariableNameMap{}, - paddle::framework::AttributeMap{}); - auto b = a.Clone(); - ASSERT_EQ(a.Type(), b->Type()); -} diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 50c3468d556bfe05d6c41906cf35cb671f711b1e..275cb8c592c3c0b153d31149570cd6596b9e1a7f 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -18,17 +18,81 @@ limitations under the License. */ #include #include +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_viz_pass.h" + #ifdef PADDLE_WITH_CUDA #include "paddle/fluid/platform/nccl_helper.h" #endif -#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" +#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" +#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h" #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace framework { +std::unique_ptr ApplyParallelExecutorPass( + const ProgramDesc &main_program, const std::vector &places, + const std::string &loss_var_name, + const std::unordered_set ¶m_names, + const std::vector &local_scopes, const bool use_cuda, +#ifdef PADDLE_WITH_CUDA + const BuildStrategy &strategy, platform::NCCLContextMap *nccl_ctxs) { +#else + const BuildStrategy &strategy) { +#endif + // Convert the program to graph. + std::unique_ptr graph(new ir::Graph(main_program)); + + // Apply a graph viz pass to record a graph. + if (!strategy.debug_graphviz_path_.empty()) { + auto viz_pass = ir::PassRegistry::Instance().Get("graph_viz_pass"); + const std::string graph_path = string::Sprintf( + "%s%s", strategy.debug_graphviz_path_.c_str(), "_original_graph"); + viz_pass->Set("graph_viz_path", new std::string(graph_path)); + graph = viz_pass->Apply(std::move(graph)); + } + + // Convert graph to run on multi-devices. + auto multi_devices_pass = + ir::PassRegistry::Instance().Get("multi_devices_pass"); + multi_devices_pass->SetNotOwned>("places", + &places); + multi_devices_pass->SetNotOwned("loss_var_name", + &loss_var_name); + multi_devices_pass->SetNotOwned>( + "params", ¶m_names); + multi_devices_pass->SetNotOwned>("local_scopes", + &local_scopes); + multi_devices_pass->SetNotOwned("strategy", &strategy); + +#ifdef PADDLE_WITH_CUDA + platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr; + multi_devices_pass->SetNotOwned("nccl_ctxs", nctx); +#endif + graph = multi_devices_pass->Apply(std::move(graph)); + + // Apply a graph print pass to record a graph with device info. + if (!strategy.debug_graphviz_path_.empty()) { + auto multi_devices_print_pass = + ir::PassRegistry::Instance().Get("multi_devices_print_pass"); + multi_devices_print_pass->SetNotOwned( + "debug_graphviz_path", &strategy.debug_graphviz_path_); + multi_devices_print_pass->Set( + "graph_printer", new details::GraphvizSSAGraphPrinter); + graph = multi_devices_print_pass->Apply(std::move(graph)); + } + + // Verify that the graph is correct for multi-device executor. + auto multi_devices_check_pass = + ir::PassRegistry::Instance().Get("multi_devices_check_pass"); + graph = multi_devices_check_pass->Apply(std::move(graph)); + return graph; +} + class ParallelExecutorPrivate { public: explicit ParallelExecutorPrivate(const std::vector &places) @@ -42,9 +106,9 @@ class ParallelExecutorPrivate { #ifdef PADDLE_WITH_CUDA std::unique_ptr nccl_ctxs_; #endif - - std::vector> var_types_; - bool own_local_scope; + bool own_local_scope_; + bool use_cuda_; + bool use_all_reduce_; }; std::vector &ParallelExecutor::GetLocalScopes() { @@ -61,69 +125,107 @@ ParallelExecutor::ParallelExecutor( size_t num_trainers, size_t trainer_id) : member_(new ParallelExecutorPrivate(places)) { member_->global_scope_ = scope; + member_->use_cuda_ = exec_strategy.use_cuda_; + member_->use_all_reduce_ = + build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce; + + if (!member_->use_all_reduce_) { + PADDLE_ENFORCE(places.size() > 1, + "If you set build_strategy.reduce with 'Reduce'," + "the number of places must be greater than 1."); + } // Step 1. Bcast the params to devs. // Create local scopes if (local_scopes.empty()) { - member_->own_local_scope = true; + member_->own_local_scope_ = true; member_->local_scopes_.emplace_back(member_->global_scope_); for (size_t i = 1; i < member_->places_.size(); ++i) { member_->local_scopes_.emplace_back(&scope->NewScope()); } } else { - member_->own_local_scope = false; + member_->own_local_scope_ = false; PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size()); for (size_t i = 0; i < member_->places_.size(); ++i) { member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope()); } } + if (member_->use_cuda_) { // Bcast Parameters to all GPUs #ifdef PADDLE_WITH_CUDA - auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); - ncclUniqueId *nccl_id = nullptr; - if (nccl_id_var != nullptr) { - nccl_id = nccl_id_var->GetMutable(); - } - member_->nccl_ctxs_.reset(new platform::NCCLContextMap( - member_->places_, nccl_id, num_trainers, trainer_id)); + auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); + ncclUniqueId *nccl_id = nullptr; + if (nccl_id_var != nullptr) { + nccl_id = nccl_id_var->GetMutable(); + } + member_->nccl_ctxs_.reset(new platform::NCCLContextMap( + member_->places_, nccl_id, num_trainers, trainer_id)); +#else + PADDLE_THROW("Not compiled with CUDA"); #endif - if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 && - local_scopes.empty()) { // Is CUDA - BCastParamsToGPUs(bcast_vars); } -// Startup Program has been run. All local scopes has correct parameters. -// Step 2. Convert main_program to SSA form and dependency graph. Also, insert + if (member_->local_scopes_.size() != 1 && local_scopes.empty()) { + BCastParamsToDevices(bcast_vars); + } + // Startup Program has been run. All local scopes has correct parameters. + + // Step 2. Create vars in each scope; + std::vector var_infos; + for (auto *var : main_program.Block(0).AllVars()) { + var_infos.emplace_back(); + var_infos.back().name_ = var->Name(); + var_infos.back().type_ = var->GetType(); + var_infos.back().persistable_ = var->Persistable(); + } + +// Step 3. Convert main_program to SSA form and dependency graph. Also, insert // ncclOp #ifdef PADDLE_WITH_CUDA - details::MultiDevSSAGraphBuilder builder( - member_->places_, loss_var_name, params, member_->local_scopes_, - member_->nccl_ctxs_.get(), build_strategy); + std::unique_ptr graph = ApplyParallelExecutorPass( + main_program, member_->places_, loss_var_name, params, + member_->local_scopes_, member_->use_cuda_, build_strategy, + member_->nccl_ctxs_.get()); #else - details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name, - params, member_->local_scopes_, - build_strategy); + std::unique_ptr graph = ApplyParallelExecutorPass( + main_program, member_->places_, loss_var_name, params, + member_->local_scopes_, member_->use_cuda_, build_strategy); #endif - auto graph = builder.Build(main_program); member_->executor_.reset(new details::ThreadedSSAGraphExecutor( exec_strategy, member_->local_scopes_, places, std::move(graph))); - - // Step 3. Create vars in each scope; - for (auto *var : main_program.Block(0).AllVars()) { - member_->var_types_.emplace_back(var->Name(), var->GetType(), - var->Persistable()); - } + member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor( + exec_strategy, member_->local_scopes_, std::move(var_infos), + member_->places_, std::move(member_->executor_))); } -void ParallelExecutor::BCastParamsToGPUs( +void ParallelExecutor::BCastParamsToDevices( const std::unordered_set &vars) const { -#ifdef PADDLE_WITH_CUDA - auto *main_scope = member_->local_scopes_[0]; - + // the initializing bcast, all vars would be bcast from device(0), + // otherwise + // bcast from the specified device. + bool initializing = member_->executor_ ? false : true; for (auto &var : vars) { - auto *main_var = main_scope->FindVar(var); + int var_dev_id = -1; + if (member_->executor_) { + auto &sharded_var_device = + member_->executor_->Graph().Get( + details::kShardedVarDevice); + if (sharded_var_device.find(var) != sharded_var_device.end()) { + var_dev_id = sharded_var_device.at(var); + } + } + + if (!initializing && var_dev_id == -1) continue; + + framework::Variable *main_var = nullptr; + if (initializing) { + main_var = member_->local_scopes_[0]->FindVar(var); + } else { + main_var = member_->local_scopes_[var_dev_id]->FindVar(var); + } + if (main_var == nullptr || !main_var->IsType()) { continue; } @@ -131,13 +233,16 @@ void ParallelExecutor::BCastParamsToGPUs( auto &main_tensor = main_var->Get(); auto &dims = main_tensor.dims(); if (paddle::platform::is_gpu_place(main_tensor.place())) { +#ifdef PADDLE_WITH_CUDA + std::vector buffers; size_t numel = main_tensor.numel(); ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type()); - platform::NCCLGroupGuard guard; for (size_t i = 0; i < member_->places_.size(); ++i) { auto place = member_->places_[i]; void *buffer; - if (i == 0) { + + if ((initializing && i == 0) || + (!initializing && static_cast(i) == var_dev_id)) { buffer = const_cast(main_tensor.data()); } else { auto local_scope = member_->local_scopes_[i]; @@ -145,66 +250,62 @@ void ParallelExecutor::BCastParamsToGPUs( t->Resize(dims); buffer = t->mutable_data(place, main_tensor.type()); } - auto &nccl_ctx = member_->nccl_ctxs_->at(place); - platform::dynload::ncclBcast(buffer, numel, data_type, 0, - nccl_ctx.comm_, nccl_ctx.stream()); + buffers.push_back(buffer); } + + PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(), + "variables' buffer size to bcast NOT equal to places"); + { + platform::NCCLGroupGuard guard; + for (size_t i = 0; i < member_->places_.size(); ++i) { + auto &nccl_ctx = member_->nccl_ctxs_->at(member_->places_[i]); + if (initializing) { + platform::dynload::ncclBcast(buffers[i], numel, data_type, 0, + nccl_ctx.comm_, nccl_ctx.stream()); + } else { + if (var_dev_id >= 0) { + platform::dynload::ncclBcast(buffers[i], numel, data_type, + var_dev_id, nccl_ctx.comm_, + nccl_ctx.stream()); + } + } + } + member_->nccl_ctxs_->WaitAll(); + } + +#else + PADDLE_THROW("Not compiled with CUDA"); +#endif } else { platform::CPUPlace cpu; - for (size_t i = 1; i < member_->places_.size(); ++i) { + for (size_t i = 0; i < member_->places_.size(); ++i) { + if ((initializing && i == 0) || + (!initializing && static_cast(i) == var_dev_id)) + continue; + auto local_scope = member_->local_scopes_[i]; auto *t = local_scope->Var(var)->GetMutable(); - t->Resize(dims); - t->mutable_data(cpu, main_tensor.type()); - paddle::framework::TensorCopy(main_tensor, cpu, t); + + // FIXME(zcd): LR_DECAY_COUNTER should not be shared. This is a hot fix. + if (member_->use_all_reduce_ || member_->use_cuda_ || + var == "@LR_DECAY_COUNTER@") { + t->Resize(dims); + t->mutable_data(cpu, main_tensor.type()); + paddle::framework::TensorCopy(main_tensor, cpu, t); + } else { + t->ShareDataWith(main_tensor); + } } } - member_->nccl_ctxs_->WaitAll(); } -#else - PADDLE_THROW("Not compiled with CUDA"); -#endif } void ParallelExecutor::Run(const std::vector &fetch_tensors, const std::string &fetched_var_name) { platform::RecordBlock b(0); - // Create local scopes. - for (auto it = member_->local_scopes_.rbegin(); - it != member_->local_scopes_.rend(); ++it) { - auto &scope = *it; - Scope &local_scope = scope->NewScope(); - *scope->Var(details::kLocalExecScopeName)->GetMutable() = - &local_scope; - - for (auto &name_type_pair : member_->var_types_) { - if (scope->FindVar(std::get<0>(name_type_pair)) != nullptr) { - continue; - } - - if (std::get<2>(name_type_pair)) { // Persistable - InitializeVariable(scope->Var(std::get<0>(name_type_pair)), - std::get<1>(name_type_pair)); - } else { - InitializeVariable(local_scope.Var(std::get<0>(name_type_pair)), - std::get<1>(name_type_pair)); - } - } - } - auto fetch_data = member_->executor_->Run(fetch_tensors); *member_->global_scope_->Var(fetched_var_name)->GetMutable() = fetch_data; - - // Wait All computational streams - for (auto p : member_->places_) { - platform::DeviceContextPool::Instance().Get(p)->Wait(); - } - for (auto &scope : member_->local_scopes_) { - auto &local_scope = - *scope->Var(details::kLocalExecScopeName)->GetMutable(); - scope->DeleteScope(local_scope); - } } void ParallelExecutor::FeedTensorsIntoLocalScopes( @@ -242,7 +343,7 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes( } ParallelExecutor::~ParallelExecutor() { - if (member_->own_local_scope) { + if (member_->own_local_scope_) { for (size_t i = 1; i < member_->local_scopes_.size(); ++i) { member_->global_scope_->DeleteScope(member_->local_scopes_[i]); } @@ -251,3 +352,8 @@ ParallelExecutor::~ParallelExecutor() { } // namespace framework } // namespace paddle + +USE_PASS(graph_viz_pass); +USE_PASS(multi_devices_pass); +USE_PASS(multi_devices_check_pass); +USE_PASS(multi_devices_print_pass); diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index 5247e790649e76567f4527d54499d6e95dac5c27..5fb748fa205d5e9dbd2943b615c69aedd0e7a26f 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -19,12 +19,14 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/multi_devices_graph_pass.h" #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" + namespace paddle { namespace framework { @@ -64,7 +66,7 @@ class ParallelExecutor { void Run(const std::vector &fetch_tensors, const std::string &fetched_var_name); - void BCastParamsToGPUs(const std::unordered_set &vars) const; + void BCastParamsToDevices(const std::unordered_set &vars) const; private: ParallelExecutorPrivate *member_; diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 64fb028f83a539d17885186d5d8ee6ef26f095e9..20bdc7830f32564448a69e9cd76c02585b7a1aca 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -51,12 +51,15 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) { auto *block = desc_.mutable_blocks(i); blocks_.emplace_back(new BlockDesc(*o.blocks_[i], block, this)); } - for (auto &block : blocks_) { - for (auto *op : block->AllOps()) { - for (const auto &attr : op->Proto()->attrs()) { - if (attr.type() == proto::AttrType::BLOCK) { - size_t blk_idx = attr.block_idx(); - op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); + for (size_t block_id = 0; block_id < blocks_.size(); ++block_id) { + auto all_ops = blocks_[block_id]->AllOps(); + for (size_t op_id = 0; op_id < all_ops.size(); ++op_id) { + auto &op = all_ops[op_id]; + for (const std::string &attr_name : op->AttrNames()) { + if (op->GetAttrType(attr_name) == proto::AttrType::BLOCK) { + int sub_block_id = + o.Block(block_id).Op(op_id)->GetBlockAttrId(attr_name); + op->SetBlockAttr(attr_name, MutableBlock(sub_block_id)); } } } @@ -86,6 +89,16 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) { for (auto &block_desc : *desc_.mutable_blocks()) { blocks_.emplace_back(new BlockDesc(this, &block_desc)); } + for (auto &block : blocks_) { + for (auto *op : block->AllOps()) { + for (const auto &attr : op->Proto()->attrs()) { + if (attr.type() == proto::AttrType::BLOCK) { + size_t blk_idx = attr.block_idx(); + op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); + } + } + } + } } const std::vector ProgramDesc::GetFeedTargetNames() { diff --git a/paddle/fluid/framework/reader.cc b/paddle/fluid/framework/reader.cc index 0b36f1116d15004b355e854e101abb9ad3297836..40eafda9bf294f7e8ddd067e9014447f4de1cc0e 100644 --- a/paddle/fluid/framework/reader.cc +++ b/paddle/fluid/framework/reader.cc @@ -13,29 +13,62 @@ // limitations under the License. #include "paddle/fluid/framework/reader.h" +#include namespace paddle { namespace framework { -ReaderBase::~ReaderBase() {} - -FileReader::FileReader(const std::vector &dims) : dims_(dims) {} -void FileReader::ReadNext(std::vector *out) { +void ReaderBase::ReadNext(std::vector *out) { + std::lock_guard lock(mu_); + PADDLE_ENFORCE_EQ(status_, ReaderStatus::kRunning); ReadNextImpl(out); - if (out->empty()) { - return; - } +} - PADDLE_ENFORCE_EQ(out->size(), dims_.size()); - for (size_t i = 0; i < dims_.size(); ++i) { - auto &actual = (*out)[i].dims(); - auto &expect = dims_[i]; +void ReaderBase::InsertDecoratedReader( + const std::shared_ptr &decorated_reader) { + std::lock_guard guard(mu_); + decorated_readers_.emplace_back(decorated_reader); +} - PADDLE_ENFORCE_EQ(actual.size(), expect.size()); - for (int j = 0; j < actual.size(); ++j) { - // PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1); +std::unordered_set ReaderBase::GetEndPoints() { + std::unordered_set result; + std::deque queue; + queue.emplace_back(this); + while (!queue.empty()) { // BFS search + auto *front = queue.front(); + queue.pop_front(); + if (front->decorated_readers_.empty()) { + result.emplace(front); + } else { + for (auto &reader : front->decorated_readers_) { + if (auto *reader_ptr = reader.lock().get()) { + queue.emplace_back(reader_ptr); + } + } } } + + return result; } + +void ReaderBase::Shutdown() { + std::lock_guard lock(mu_); + if (status_ != ReaderStatus::kStopped) { + ShutdownImpl(); + status_ = ReaderStatus::kStopped; + } +} + +void ReaderBase::Start() { + std::lock_guard lock(mu_); + if (status_ != ReaderStatus::kRunning) { + StartImpl(); + status_ = ReaderStatus::kRunning; + } +} + +ReaderBase::~ReaderBase() {} + +DecoratedReader::~DecoratedReader() { reader_->Shutdown(); } } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/reader.h b/paddle/fluid/framework/reader.h index 3a413941df964c8d9454fafc6030c377c10f9fb1..82562bf883d88787858912f7039cf8fef003eccf 100644 --- a/paddle/fluid/framework/reader.h +++ b/paddle/fluid/framework/reader.h @@ -15,6 +15,7 @@ #pragma once #include +#include #include #include "paddle/fluid/framework/ddim.h" @@ -26,58 +27,116 @@ namespace framework { class ReaderBase { public: - virtual void ReadNext(std::vector* out) = 0; + virtual void ReadNext(std::vector* out); - virtual void ReInit() = 0; + virtual void Shutdown(); - virtual ~ReaderBase(); -}; + virtual void Start(); -class DecoratedReader : public ReaderBase { - public: - explicit DecoratedReader(ReaderBase* reader) : ReaderBase(), reader_(reader) { - PADDLE_ENFORCE_NOT_NULL(reader_); - } + // Return the readers which are the end of decorating chain. Basically + // they are readers just before read op. + std::unordered_set GetEndPoints(); - void ReInit() override { reader_->ReInit(); } + virtual ~ReaderBase(); protected: - ReaderBase* reader_; + virtual void ReadNextImpl(std::vector* out) {} + + virtual void ShutdownImpl() {} + + virtual void StartImpl() {} + + enum ReaderStatus { kRunning, kStopped }; + + ReaderStatus status_{kRunning}; + + mutable std::mutex mu_; + + private: + friend class DecoratedReader; + // These methods can be only invoked inside DecoratedReader to record the + // decorating chain. + void InsertDecoratedReader( + const std::shared_ptr& decorated_reader); + // A set of which readers that decorated this reader. + std::vector> decorated_readers_; }; -class FileReader : public ReaderBase { +class DecoratedReader : public ReaderBase, + public std::enable_shared_from_this { public: - explicit FileReader(const std::vector& dims); + explicit DecoratedReader(const std::shared_ptr& reader) + : ReaderBase(), reader_(reader) { + PADDLE_ENFORCE_NOT_NULL(reader_); + } - void ReadNext(std::vector* out) override; + void RegisterDecorateChain() { + reader_->InsertDecoratedReader(shared_from_this()); + } + + ~DecoratedReader(); protected: - virtual void ReadNextImpl(std::vector* out) = 0; + void ShutdownImpl() override { reader_->Shutdown(); } - private: - std::vector dims_; + void StartImpl() override { reader_->Start(); } + + std::shared_ptr reader_; }; +// FileReader is just a conceptual class. +class FileReader : public ReaderBase {}; + // The ReaderHolder is used as reader' unified wrapper, // making it easier to access different type reader in Variables. class ReaderHolder { public: - void Reset(ReaderBase* reader) { reader_.reset(reader); } + template + void Reset(const std::shared_ptr& reader) { + auto reader_base = std::dynamic_pointer_cast(reader); + PADDLE_ENFORCE_NOT_NULL(reader_base); + reader_ = reader_base; + } - ReaderBase* Get() const { return reader_.get(); } + const std::shared_ptr& Get() const { return reader_; } void ReadNext(std::vector* out) { PADDLE_ENFORCE_NOT_NULL(reader_); reader_->ReadNext(out); } - void ReInit() { + + void ResetAll() { + auto end_readers = reader_->GetEndPoints(); + for (auto* reader : end_readers) { + reader->Shutdown(); + } + for (auto* reader : end_readers) { + reader->Start(); + } + } + + void Shutdown() { PADDLE_ENFORCE_NOT_NULL(reader_); - reader_->ReInit(); + reader_->Shutdown(); } + void Start() { + PADDLE_ENFORCE_NOT_NULL(reader_); + reader_->Start(); + } + + operator const std::shared_ptr&() const { return this->reader_; } + private: - std::unique_ptr reader_; + std::shared_ptr reader_; }; +template +inline std::shared_ptr MakeDecoratedReader(ARGS&&... args) { + std::shared_ptr reader(new T(std::forward(args)...)); + reader->RegisterDecorateChain(); + return reader; +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/reader_test.cc b/paddle/fluid/framework/reader_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..f0d07cb7c1367576084b9494e7758103bb45d1e5 --- /dev/null +++ b/paddle/fluid/framework/reader_test.cc @@ -0,0 +1,52 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/reader.h" +#include +#include "gtest/gtest.h" + +class StubDecoratedReader : public paddle::framework::DecoratedReader { + public: + explicit StubDecoratedReader(const std::shared_ptr &reader) + : DecoratedReader(reader) {} + + void ReadNextImpl(std::vector *out) override {} +}; + +class StubRootReader : public paddle::framework::ReaderBase { + public: + void ReadNextImpl(std::vector *out) override {} +}; + +TEST(READER, decorate_chain) { + auto root = std::make_shared(); + auto end_point1 = + paddle::framework::MakeDecoratedReader(root); + auto end_point2 = + paddle::framework::MakeDecoratedReader(root); + + { + auto endpoints = root->GetEndPoints(); + ASSERT_EQ(endpoints.size(), 2U); + ASSERT_NE(endpoints.count(end_point1.get()), 0); + ASSERT_NE(endpoints.count(end_point2.get()), 0); + } + + { + auto end_point3 = + paddle::framework::MakeDecoratedReader(root); + ASSERT_EQ(root->GetEndPoints().size(), 3U); + } + { ASSERT_EQ(root->GetEndPoints().size(), 2U); } +} diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 9091713158c8071d5386f14250e3c546284e7fd0..50f374e3703a97f6c1fdb4b14fdeb0b603f9ac86 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -34,13 +34,7 @@ DEFINE_bool( namespace paddle { namespace framework { -Scope::~Scope() { - DropKids(); - for (auto& kv : vars_) { - VLOG(3) << "Destroy variable " << kv.first; - delete kv.second; - } -} +Scope::~Scope() { DropKids(); } Scope& Scope::NewScope() const { std::unique_lock lock(mutex_); @@ -49,45 +43,37 @@ Scope& Scope::NewScope() const { } Variable* Scope::Var(const std::string& name) { - auto* v = FindVarLocally(name); - if (v != nullptr) return v; - v = new Variable(); - vars_[name] = v; - VLOG(3) << "Create variable " << name; - v->name_ = &(vars_.find(name)->first); - return v; + std::unique_lock lock(mutex_); + return VarInternal(name); } Variable* Scope::Var(std::string* name) { - auto var_name = string::Sprintf("%p.%d", this, vars_.size()); + std::unique_lock lock(mutex_); + auto new_name = string::Sprintf("%p.%d", this, vars_.size()); if (name != nullptr) { - *name = var_name; + *name = new_name; } - return Var(var_name); + return VarInternal(new_name); } Variable* Scope::FindVar(const std::string& name) const { - auto var = FindVarLocally(name); - if (var != nullptr) { - return var; - } - return (parent_ == nullptr) ? nullptr : parent_->FindVar(name); + std::unique_lock lock(mutex_); + return FindVarInternal(name); } const Scope* Scope::FindScope(const Variable* var) const { - for (auto& kv : vars_) { - if (kv.second == var) { - return this; - } - } - return (parent_ == nullptr) ? nullptr : parent_->FindScope(var); + std::unique_lock lock(mutex_); + return FindScopeInternal(var); } + void Scope::DropKids() { + std::unique_lock lock(mutex_); for (Scope* s : kids_) delete s; kids_.clear(); } std::vector Scope::LocalVarNames() const { + std::unique_lock lock(mutex_); std::vector known_vars; known_vars.reserve(this->vars_.size()); for (auto& p : vars_) { @@ -110,10 +96,10 @@ void Scope::DeleteScope(Scope* scope) const { } void Scope::EraseVars(const std::vector& var_names) { + std::unique_lock lock(mutex_); std::set var_set(var_names.begin(), var_names.end()); for (auto it = vars_.begin(); it != vars_.end();) { if (var_set.find(it->first) != var_set.end()) { - delete it->second; it = vars_.erase(it); } else { ++it; @@ -123,25 +109,60 @@ void Scope::EraseVars(const std::vector& var_names) { void Scope::Rename(const std::string& origin_name, const std::string& new_name) const { + std::unique_lock lock(mutex_); + RenameInternal(origin_name, new_name); +} + +std::string Scope::Rename(const std::string& origin_name) const { + std::unique_lock lock(mutex_); + auto new_name = string::Sprintf("%p.%d", this, vars_.size()); + RenameInternal(origin_name, new_name); + return new_name; +} + +Variable* Scope::VarInternal(const std::string& name) { + auto* v = FindVarLocally(name); + if (v != nullptr) return v; + + v = new Variable(); + vars_[name].reset(v); + VLOG(3) << "Create variable " << name; + v->name_ = &(vars_.find(name)->first); + return v; +} + +const Scope* Scope::FindScopeInternal(const Variable* var) const { + for (auto& kv : vars_) { + if (kv.second.get() == var) { + return this; + } + } + return (parent_ == nullptr) ? nullptr : parent_->FindScope(var); +} + +void Scope::RenameInternal(const std::string& origin_name, + const std::string& new_name) const { auto origin_it = vars_.find(origin_name); PADDLE_ENFORCE(origin_it != vars_.end(), "Cannot find original variable with name %s", origin_name); auto new_it = vars_.find(new_name); PADDLE_ENFORCE(new_it == vars_.end(), "The variable with name %s is already in the scope", new_name); - vars_[new_name] = origin_it->second; + vars_[new_name].reset(origin_it->second.release()); vars_.erase(origin_it); } -std::string Scope::Rename(const std::string& origin_name) const { - auto var_name = string::Sprintf("%p.%d", this, vars_.size()); - Rename(origin_name, var_name); - return var_name; +Variable* Scope::FindVarInternal(const std::string& name) const { + auto var = FindVarLocally(name); + if (var != nullptr) { + return var; + } + return (parent_ == nullptr) ? nullptr : parent_->FindVar(name); } Variable* Scope::FindVarLocally(const std::string& name) const { auto it = vars_.find(name); - if (it != vars_.end()) return it->second; + if (it != vars_.end()) return it->second.get(); return nullptr; } diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index abc82e452d732638a2f7315022074850f299a7ea..e246241c0abfbc7bdcaf38d073cc58fc36a4f737 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -47,15 +47,18 @@ class Scope { Scope& NewScope() const; /// Create a variable with given name if it doesn't exist. + /// Caller doesn't own the returned Variable. Variable* Var(const std::string& name); /// Create a variable with a scope-unique name. + /// Caller doesn't own the returned Variable. Variable* Var(std::string* name = nullptr); void EraseVars(const std::vector& var_names); /// Find a variable in the scope or any of its ancestors. Returns /// nullptr if cannot find. + /// Caller doesn't own the returned Variable. Variable* FindVar(const std::string& name) const; const Scope* parent() const { return parent_; } @@ -78,13 +81,30 @@ class Scope { // Rename variable to a new name and return the new name std::string Rename(const std::string& origin_name) const; - Variable* FindVarLocally(const std::string& name) const; + protected: + mutable std::unordered_map> vars_; private: // Call Scope::NewScope for a sub-scope. explicit Scope(Scope const* parent) : parent_(parent) {} - mutable std::unordered_map vars_; + // Called by Var. + Variable* VarInternal(const std::string& name); + + // Called by FindScope. + const Scope* FindScopeInternal(const Variable* var) const; + + // Called by Rename. + void RenameInternal(const std::string& origin_name, + const std::string& new_name) const; + + // Called by FindVar recursively. + Variable* FindVarInternal(const std::string& name) const; + + // Called by FindVarInternal and Var. + Variable* FindVarLocally(const std::string& name) const; + + // Scope in `kids_` are owned by this class. mutable std::list kids_; Scope const* parent_{nullptr}; diff --git a/paddle/fluid/framework/tensor.cc b/paddle/fluid/framework/tensor.cc index e97ada06f06d0538f17160220e3aa3f4ffc55520..56bb9142dabe0d5546e321e675a5acba7bf4d306 100644 --- a/paddle/fluid/framework/tensor.cc +++ b/paddle/fluid/framework/tensor.cc @@ -15,5 +15,103 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor.h" namespace paddle { -namespace framework {} +namespace framework { +extern size_t SizeOfType(std::type_index type); +void Tensor::check_memory_size() const { + PADDLE_ENFORCE_NOT_NULL( + holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); + PADDLE_ENFORCE_LE( + numel() * SizeOfType(type()), memory_size(), + "Tensor's dims_ is out of bound. Call Tensor::mutable_data " + "first to re-allocate memory.\n" + "or maybe the required data-type mismatches the data already stored."); +} + +size_t Tensor::memory_size() const { + return holder_ == nullptr ? 0UL : holder_->size() - offset_; +} + +void* Tensor::mutable_data(platform::Place place, std::type_index type) { + if (holder_ != nullptr) { + holder_->set_type(type); + } + PADDLE_ENFORCE_GE(numel(), 0, + "When calling this method, the Tensor's numel must be " + "equal or larger than zero. " + "Please check Tensor::Resize has been called first."); + int64_t size = numel() * SizeOfType(type); + /* some versions of boost::variant don't have operator!= */ + if (holder_ == nullptr || !(holder_->place() == place) || + holder_->size() < size + offset_) { + if (platform::is_cpu_place(place)) { + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); + } else if (platform::is_gpu_place(place) || + platform::is_cuda_pinned_place(place)) { +#ifndef PADDLE_WITH_CUDA + PADDLE_THROW( + "CUDAPlace or CUDAPinnedPlace is not supported in CPU-only mode."); + } +#else + if (platform::is_gpu_place(place)) { + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); + } else if (platform::is_cuda_pinned_place(place)) { + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); + } + } +#endif + offset_ = 0; + } + return reinterpret_cast(reinterpret_cast(holder_->ptr()) + + offset_); +} + +void* Tensor::mutable_data(platform::Place place) { + PADDLE_ENFORCE(this->holder_ != nullptr, + "Cannot invoke mutable data if current hold nothing."); + return mutable_data(place, holder_->type()); +} + +Tensor& Tensor::ShareDataWith(const Tensor& src) { + src.check_memory_size(); + *this = src; + return *this; +} + +Tensor Tensor::Slice(int begin_idx, int end_idx) const { + check_memory_size(); + PADDLE_ENFORCE_GE(begin_idx, 0, + "The start row index must be greater than 0."); + PADDLE_ENFORCE_LE(end_idx, dims_[0], "The end row index is out of bound."); + PADDLE_ENFORCE_LT( + begin_idx, end_idx, + "The start row index must be lesser than the end row index."); + + if (dims_[0] == 1) { + return *this; + } else { + size_t base = numel() / dims_[0]; + Tensor dst; + dst.holder_ = holder_; + dst.set_layout(layout_); + DDim dst_dims = dims_; + dst_dims[0] = end_idx - begin_idx; + dst.Resize(dst_dims); + dst.offset_ = offset_ + begin_idx * base * SizeOfType(type()); + return dst; + } +} + +Tensor& Tensor::Resize(const DDim& dims) { + dims_ = dims; + return *this; +} + +const DDim& Tensor::dims() const { return dims_; } + +int64_t Tensor::numel() const { return product(dims_); } + +} // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/tensor.h b/paddle/fluid/framework/tensor.h index 6f878541e6de1deec1829145b1b325ecd176a034..0bbfd66148e9bc9080654bf1b0b34477115a0e6b 100644 --- a/paddle/fluid/framework/tensor.h +++ b/paddle/fluid/framework/tensor.h @@ -34,6 +34,28 @@ namespace framework { class LoDTensor; class Tensor { +#ifdef PADDLE_WITH_MKLDNN + + public: + inline mkldnn::memory::format format() const { return format_; } + + inline void set_format(const mkldnn::memory::format format) { + format_ = format; + } + + protected: + /** + * @brief the detail format of memory block which have layout as kMKLDNN + * + * @note MKLDNN lib support various memory format like nchw, nhwc, nChw8C, + * nChw16c, etc. For a MKLDNN memory block, layout will be set as + * DataLayout::kMKLDNN meanwhile detail memory format will be kept in + * this field. + */ + + mkldnn::memory::format format_ = mkldnn::memory::format::format_undef; +#endif + public: template friend struct EigenTensor; @@ -54,26 +76,24 @@ class Tensor { /*! Return a pointer to mutable memory block. */ template - inline T* data(); + T* data(); /*! Return a pointer to constant memory block. */ template - inline const T* data() const; + const T* data() const; inline bool IsInitialized() const; - inline void switch_place(platform::Place new_place); - /** * @brief Return a pointer to mutable memory block. * @note If not exist, then allocation. */ template - inline T* mutable_data(platform::Place place); + T* mutable_data(platform::Place place); - inline void* mutable_data(platform::Place place, std::type_index type); + void* mutable_data(platform::Place place, std::type_index type); - inline void* mutable_data(platform::Place place); + void* mutable_data(platform::Place place); /** * @brief Return a pointer to mutable memory block. @@ -84,19 +104,19 @@ class Tensor { * @note If not exist, then allocation. */ template - inline T* mutable_data(DDim dims, platform::Place place); + T* mutable_data(DDim dims, platform::Place place); /*! Return the dimensions of the memory block. */ - inline const DDim& dims() const; + const DDim& dims() const; /*! Return the numel of the memory block. */ - inline int64_t numel() const; + int64_t numel() const; /*! Resize the dimensions of the memory block. */ - inline Tensor& Resize(const DDim& dims); + Tensor& Resize(const DDim& dims); /*! The internal of two tensors share the same memory block. */ - inline Tensor& ShareDataWith(const Tensor& src); + Tensor& ShareDataWith(const Tensor& src); /** * @brief Return a sub-tensor of the given tensor. @@ -106,7 +126,7 @@ class Tensor { * @param[in] end_idx The index of the end row(exclusive) to slice. * The index number begins from 0. */ - inline Tensor Slice(int begin_idx, int end_idx) const; + Tensor Slice(int begin_idx, int end_idx) const; platform::Place place() const { PADDLE_ENFORCE_NOT_NULL( @@ -123,11 +143,11 @@ class Tensor { // memory size returns the holding memory size in byte. size_t memory_size() const; - inline void check_memory_size() const; + void check_memory_size() const; - inline DataLayout layout() const { return layout_; } + DataLayout layout() const { return layout_; } - inline void set_layout(const DataLayout layout) { layout_ = layout; } + void set_layout(const DataLayout layout) { layout_ = layout; } private: /** @@ -197,8 +217,10 @@ class Tensor { * N,C,H,W for respectively the batch size, the number of * feature maps, the height. */ - - DataLayout layout_ = DataLayout::kNHWC; + // Fix me: here just change the default layout to kNCHW + // it doesn't fix the real issue, i.e. feeder should set up tensor layout + // according to actual input data + DataLayout layout_ = DataLayout::kNCHW; /** * @brief A PlaceHolder may be shared by more than one tensor. @@ -210,15 +232,6 @@ class Tensor { size_t offset_; }; -inline void Tensor::switch_place(platform::Place new_place) { - if (holder_->place() == new_place) { - return; - } - - // TODO(tonyyang-svail): do memcpy here. - PADDLE_THROW("Not Implemented"); -} - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index 2f19ec0f0a9338e2b96d1f64eac45387bae4d1eb..b7b62eef23ec351686378c913d18fc72308fd7b2 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -20,27 +20,12 @@ limitations under the License. */ namespace paddle { namespace framework { -extern size_t SizeOfType(std::type_index type); -inline void Tensor::check_memory_size() const { - PADDLE_ENFORCE_NOT_NULL( - holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); - PADDLE_ENFORCE_LE( - numel() * SizeOfType(type()), memory_size(), - "Tensor's dims_ is out of bound. Call Tensor::mutable_data " - "first to re-allocate memory.\n" - "or maybe the required data-type mismatches the data already stored."); -} - -inline size_t Tensor::memory_size() const { - return holder_ == nullptr ? 0UL : holder_->size() - offset_; -} - template inline const T* Tensor::data() const { check_memory_size(); - PADDLE_ENFORCE(std::is_same::value || - holder_->type() == std::type_index(typeid(T)), - "Tensor holds the wrong type, it holds %s", + bool valid = std::is_same::value || + holder_->type() == std::type_index(typeid(T)); + PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); return reinterpret_cast( @@ -52,9 +37,9 @@ inline bool Tensor::IsInitialized() const { return holder_ != nullptr; } template inline T* Tensor::data() { check_memory_size(); - PADDLE_ENFORCE(std::is_same::value || - holder_->type() == std::type_index(typeid(T)), - "Tensor holds the wrong type, it holds %s", + bool valid = std::is_same::value || + holder_->type() == std::type_index(typeid(T)); + PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); return reinterpret_cast(reinterpret_cast(holder_->ptr()) + offset_); @@ -73,89 +58,15 @@ inline T* Tensor::mutable_data(platform::Place place) { return reinterpret_cast(mutable_data(place, typeid(T))); } -inline void* Tensor::mutable_data(platform::Place place, std::type_index type) { - if (holder_ != nullptr) { - holder_->set_type(type); - } - PADDLE_ENFORCE_GE(numel(), 0, - "When calling this method, the Tensor's numel must be " - "equal or larger than zero. " - "Please check Tensor::Resize has been called first."); - int64_t size = numel() * SizeOfType(type); - /* some versions of boost::variant don't have operator!= */ - if (holder_ == nullptr || !(holder_->place() == place) || - holder_->size() < size + offset_) { - if (platform::is_cpu_place(place)) { - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); - } else if (platform::is_gpu_place(place) || - platform::is_cuda_pinned_place(place)) { -#ifndef PADDLE_WITH_CUDA - PADDLE_THROW( - "CUDAPlace or CUDAPinnedPlace is not supported in CPU-only mode."); - } -#else - if (platform::is_gpu_place(place)) { - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); - } else if (platform::is_cuda_pinned_place(place)) { - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); - } - } -#endif - offset_ = 0; - } - return reinterpret_cast(reinterpret_cast(holder_->ptr()) + - offset_); -} - -inline void* Tensor::mutable_data(platform::Place place) { - PADDLE_ENFORCE(this->holder_ != nullptr, - "Cannot invoke mutable data if current hold nothing."); - return mutable_data(place, holder_->type()); -} - -inline Tensor& Tensor::ShareDataWith(const Tensor& src) { - src.check_memory_size(); - *this = src; - return *this; -} - -inline Tensor Tensor::Slice(int begin_idx, int end_idx) const { - check_memory_size(); - PADDLE_ENFORCE_GE(begin_idx, 0, - "The start row index must be greater than 0."); - PADDLE_ENFORCE_LE(end_idx, dims_[0], "The end row index is out of bound."); - PADDLE_ENFORCE_LT( - begin_idx, end_idx, - "The start row index must be lesser than the end row index."); - - if (dims_[0] == 1) { - return *this; - } else { - size_t base = numel() / dims_[0]; - Tensor dst; - dst.holder_ = holder_; - dst.set_layout(layout_); - DDim dst_dims = dims_; - dst_dims[0] = end_idx - begin_idx; - dst.Resize(dst_dims); - dst.offset_ = offset_ + begin_idx * base * SizeOfType(type()); - return dst; - } -} - -inline Tensor& Tensor::Resize(const DDim& dims) { - dims_ = dims; - return *this; -} - -inline const DDim& Tensor::dims() const { return dims_; } - -inline int64_t Tensor::numel() const { return product(dims_); } - inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { + int rank = src.dims().size(); + PADDLE_ENFORCE_GE( + rank, 2, + "'ReshapeToMatrix()' is only used for flatten high rank " + "tensors to matrixs. Can not be used in reshaping vectors."); + if (rank == 2) { + return src; + } Tensor res; res.ShareDataWith(src); res.Resize(flatten_to_2d(src.dims(), num_col_dims)); diff --git a/paddle/fluid/framework/tensor_test.cc b/paddle/fluid/framework/tensor_test.cc index e1012de2ec36eb4a858202d56a678b6a204c2f0a..cb2061c06a429d8e8116001a4aa4e8c46ea13428 100644 --- a/paddle/fluid/framework/tensor_test.cc +++ b/paddle/fluid/framework/tensor_test.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/tensor.h" #include #include +#include "paddle/fluid/platform/float16.h" namespace framework = paddle::framework; namespace platform = paddle::platform; @@ -209,7 +210,21 @@ TEST(Tensor, ReshapeToMatrix) { TEST(Tensor, Layout) { framework::Tensor src; - ASSERT_EQ(src.layout(), framework::DataLayout::kNHWC); + ASSERT_EQ(src.layout(), framework::DataLayout::kNCHW); src.set_layout(framework::DataLayout::kAnyLayout); ASSERT_EQ(src.layout(), framework::DataLayout::kAnyLayout); } + +TEST(Tensor, FP16) { + using platform::float16; + framework::Tensor src; + float16* src_ptr = src.mutable_data({2, 3}, platform::CPUPlace()); + for (int i = 0; i < 2 * 3; ++i) { + src_ptr[i] = static_cast(i); + } + EXPECT_EQ(src.memory_size(), 2 * 3 * sizeof(float16)); + // EXPECT a human readable error message + // src.data(); + // Tensor holds the wrong type, it holds N6paddle8platform7float16E at + // [/paddle/Paddle/paddle/fluid/framework/tensor_impl.h:43] +} diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index e5bc74755f46449296a153e8b330968e6d9f1e1d..ab693004cfb038fd92afd9c60e0fcb4e16b9f8a9 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -15,6 +15,7 @@ #include #include #include +#include "paddle/fluid/framework/data_type.h" namespace paddle { namespace framework { @@ -69,7 +70,22 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); auto stream = reinterpret_cast(ctx).stream(); - memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); + if (platform::is_same_place(src_place, dst_place)) { + memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, + stream); + } else { + if (platform::is_same_place(ctx_place, src_place)) { + memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, + stream); + platform::DeviceContextPool::Instance().Get(src.place())->Wait(); + } else if (platform::is_same_place(ctx_place, dst_place)) { + platform::DeviceContextPool::Instance().Get(src.place())->Wait(); + memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, + stream); + } else { + PADDLE_THROW("ctx is not belong to dst_gpu_place or src_gpu_place."); + } + } } #endif } @@ -78,10 +94,10 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, Tensor* dst) { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); const platform::DeviceContext* dev_ctx; - if (platform::is_gpu_place(src.place())) { - dev_ctx = pool.Get(src.place()); - } else { + if (platform::is_gpu_place(dst_place)) { dev_ctx = pool.Get(dst_place); + } else { + dev_ctx = pool.Get(src.place()); } TensorCopy(src, dst_place, *dev_ctx, dst); } @@ -246,7 +262,8 @@ void TensorToStream(std::ostream& os, const Tensor& tensor, os.write(out.data(), size); } { // the 3rd field, tensor data - uint64_t size = tensor.memory_size(); + uint64_t size = tensor.numel() * framework::SizeOfType(tensor.type()); + auto* data_ptr = tensor.data(); PADDLE_ENFORCE(size < std::numeric_limits::max(), "Index overflow when writing tensor"); @@ -316,6 +333,9 @@ void TensorFromStream(std::istream& is, Tensor* tensor, tensor->Resize(framework::make_ddim(dims)); void* buf; auto ctx = platform::CPUDeviceContext(); + size_t size = + tensor->numel() * + framework::SizeOfType(framework::ToTypeIndex(desc.data_type())); if (platform::is_gpu_place(dev_ctx.GetPlace())) { #ifdef PADDLE_WITH_CUDA Tensor cpu_tensor; @@ -323,7 +343,7 @@ void TensorFromStream(std::istream& is, Tensor* tensor, framework::VisitDataType( desc.data_type(), DeserializedDataFunctor(&buf, &cpu_tensor, ctx.GetPlace())); - is.read(static_cast(buf), cpu_tensor.memory_size()); + is.read(static_cast(buf), size); auto dst_place = dev_ctx.GetPlace(); framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor); #else @@ -333,7 +353,7 @@ void TensorFromStream(std::istream& is, Tensor* tensor, framework::VisitDataType( desc.data_type(), DeserializedDataFunctor(&buf, tensor, ctx.GetPlace())); - is.read(static_cast(buf), tensor->memory_size()); + is.read(static_cast(buf), size); } } } diff --git a/paddle/fluid/framework/tensor_util.h b/paddle/fluid/framework/tensor_util.h index dca279b69382b80e055f661cefe84b81326704b5..4457382ade37a12f5f3613fc4113fbf1f6f91124 100644 --- a/paddle/fluid/framework/tensor_util.h +++ b/paddle/fluid/framework/tensor_util.h @@ -23,10 +23,25 @@ limitations under the License. */ namespace paddle { namespace framework { +// NOTE(zcd): Because TensorCopy is an async operation, when the src_place +// and dst_place are two different GPU, to ensure that the operation can +// be carried out correctly, there is a src_ctx wait operation in TensorCopy. +// If ctx_place and src_place are the same, src_ctx.Wait() is added +// after memory::Copy; if ctx_place and dst_place are the same, +// src_ctx.Wait() is added before memory::Copy. void TensorCopy(const Tensor& src, const platform::Place& dst_place, const platform::DeviceContext& ctx, Tensor* dst); + +// NOTE(zcd): If the src.place() and dst_place are two different GPU, +// the copy operation is carried out on the dst_place's stream. This is +// very important, because TensorCopy is an async operator, and in most +// case, once this copy operator returns, dst is to be used in dst_place's +// stream, if this copy operation is carried out on the src_place's stream, +// when dst is used in dst_place's stream the copy operation may be +// not completed. void TensorCopy(const Tensor& src, const platform::Place& dst_place, Tensor* dst); + void TensorCopySync(const Tensor& src, const platform::Place& dst_place, Tensor* dst); diff --git a/paddle/fluid/framework/threadpool.cc b/paddle/fluid/framework/threadpool.cc index f26f212d4d5793b88fd1e6d782cdf983bf341879..18cdca3a658a6a89e6ab637a7f38825756acfea8 100644 --- a/paddle/fluid/framework/threadpool.cc +++ b/paddle/fluid/framework/threadpool.cc @@ -20,6 +20,9 @@ DEFINE_int32(io_threadpool_size, 100, "number of threads used for doing IO, default 100"); +DEFINE_int32(dist_threadpool_size, 0, + "number of threads used for distributed executed."); + namespace paddle { namespace framework { @@ -35,6 +38,10 @@ void ThreadPool::Init() { if (threadpool_.get() == nullptr) { // TODO(Yancey1989): specify the max threads number int num_threads = std::thread::hardware_concurrency(); + if (FLAGS_dist_threadpool_size > 0) { + num_threads = FLAGS_dist_threadpool_size; + VLOG(1) << "set dist_threadpool_size to " << num_threads; + } PADDLE_ENFORCE_GT(num_threads, 0); threadpool_.reset(new ThreadPool(num_threads)); } diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h index 4879209ece9fdfea91e484a4118c00a2a2a2b4f7..e099e40f121ff13657e563eb608feecbca0551be 100644 --- a/paddle/fluid/framework/type_defs.h +++ b/paddle/fluid/framework/type_defs.h @@ -35,7 +35,8 @@ using VariableNameMap = std::map>; using Attribute = boost::variant, std::vector, std::vector, bool, - std::vector, BlockDesc*, int64_t>; + std::vector, BlockDesc*, int64_t, + std::vector>; using AttributeMap = std::unordered_map; diff --git a/paddle/fluid/framework/var_type.h b/paddle/fluid/framework/var_type.h index 2b646d78f0b23ec3e065c891826856c2341d4ac1..429997c8b89fef7aa164e878095ab3b5c9998e5b 100644 --- a/paddle/fluid/framework/var_type.h +++ b/paddle/fluid/framework/var_type.h @@ -24,18 +24,24 @@ limitations under the License. */ namespace paddle { namespace framework { + +template +bool IsType(const std::type_index& type_index) { + return type_index == std::type_index(typeid(T)); +} + inline proto::VarType::Type ToVarType(std::type_index type) { - if (type.hash_code() == typeid(LoDTensor).hash_code()) { + if (IsType(type)) { return proto::VarType_Type_LOD_TENSOR; - } else if (type.hash_code() == typeid(LoDRankTable).hash_code()) { + } else if (IsType(type)) { return proto::VarType_Type_LOD_RANK_TABLE; - } else if (type.hash_code() == typeid(LoDTensorArray).hash_code()) { + } else if (IsType(type)) { return proto::VarType_Type_LOD_TENSOR_ARRAY; - } else if (type.hash_code() == typeid(SelectedRows).hash_code()) { + } else if (IsType(type)) { return proto::VarType_Type_SELECTED_ROWS; - } else if (type.hash_code() == typeid(ReaderHolder).hash_code()) { + } else if (IsType(type)) { return proto::VarType_Type_READER; - } else if (type.hash_code() == typeid(ChannelHolder).hash_code()) { + } else if (IsType(type)) { return proto::VarType_Type_CHANNEL; } else { PADDLE_THROW("ToVarType:Unsupported type %s", type.name()); diff --git a/paddle/fluid/framework/var_type_inference_test.cc b/paddle/fluid/framework/var_type_inference_test.cc index 14b81ddfecb8c996ae8709910c022a074e91eb3c..7842168f603885ce7dc87d2a01dfa4f544389faa 100644 --- a/paddle/fluid/framework/var_type_inference_test.cc +++ b/paddle/fluid/framework/var_type_inference_test.cc @@ -22,6 +22,17 @@ limitations under the License. */ namespace paddle { namespace framework { +class NOP : public OperatorBase { + public: + NOP(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + private: + void RunImpl(const Scope &scope, + const platform::Place &place) const override {} +}; + class SumOpMaker : public OpProtoAndCheckerMaker { public: void Make() { diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index ec16a1c600a3bafc1c4cbbd920360253c106e3a1..ba7645aa02413f28a648f35e381da7824604a455 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -1,4 +1,11 @@ -set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init) +# analysis and tensorrt must be added before creating static library, +# otherwise, there would be undefined reference to them in static library. +add_subdirectory(analysis) +if (TENSORRT_FOUND) + add_subdirectory(tensorrt) +endif() + +set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor) # TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal? cc_library(paddle_fluid_api @@ -7,15 +14,24 @@ cc_library(paddle_fluid_api get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) -if(WITH_CONTRIB) - set(fluid_modules "${fluid_modules}" paddle_inference_api) +# paddle_fluid_origin exclude inference api interface +cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api) + +if(NOT APPLE) + add_subdirectory(api) endif() # Create static library -cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api) +cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api) +if(NOT APPLE) + # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac. + set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym") + set_target_properties(paddle_fluid PROPERTIES LINK_FLAGS "${LINK_FLAGS}") +endif() + # Create shared library cc_library(paddle_fluid_shared SHARED - SRCS io.cc + SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc DEPS ${fluid_modules} paddle_fluid_api) set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) @@ -23,14 +39,21 @@ if(NOT APPLE) # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac. set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map") set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}") + # check symbol hidden + FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake + "execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh" + " ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_fluid.so\" RESULT_VARIABLE symbol_res)\n" + "if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n" + " message(FATAL_ERROR \"Check symbol failed.\")\n" + "endif()\n") + add_custom_command( + OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol" + COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake" + DEPENDS paddle_fluid_shared) + add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol") endif() if(WITH_TESTING) # both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book add_subdirectory(tests/book) - add_subdirectory(analysis) -endif() - -if (TENSORRT_FOUND) - add_subdirectory(tensorrt) endif() diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt index 9faf5bb3036775a2ba0c08d3d6ea17ffa73753c6..27fe575cb6167a726ff92a8f3d2e47b6f536ba39 100644 --- a/paddle/fluid/inference/analysis/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -1,17 +1,45 @@ -set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init) -cc_library(analysis SRCS dot.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc fluid_to_data_flow_graph_pass.cc - DEPS paddle_fluid) +cc_library(analysis SRCS pass_manager.cc dot.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc + fluid_to_data_flow_graph_pass.cc + data_flow_graph_to_fluid_pass.cc + dfg_graphviz_draw_pass.cc + tensorrt_subgraph_pass.cc + tensorrt_subgraph_node_mark_pass.cc + analyzer.cc + helper.cc + model_store_pass.cc + DEPS framework_proto proto_desc) cc_test(test_node SRCS node_tester.cc DEPS analysis) cc_test(test_dot SRCS dot_tester.cc DEPS analysis) +cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis) set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) -cc_test(test_data_flow_graph SRCS data_flow_graph_tester.cc DEPS analysis ${FLUID_CORE_MODULES} paddle_fluid - ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model) -set_tests_properties(test_data_flow_graph PROPERTIES DEPENDS test_word2vec) +function (inference_analysis_test TARGET) + if(WITH_TESTING) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS) + cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) -cc_test(test_subgraph_splitter - SRCS subgraph_splitter_tester.cc - DEPS analysis paddle_fluid tensor - ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model) -set_tests_properties(test_subgraph_splitter PROPERTIES DEPENDS test_word2vec) + set(mem_opt "") + if(WITH_GPU) + set(mem_opt "--fraction_of_gpu_memory_to_use=0.5") + endif() + cc_test(${TARGET} + SRCS "${analysis_test_SRCS}" + DEPS analysis + ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt}) + set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec) + endif(WITH_TESTING) +endfunction(inference_analysis_test) + +inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc) +inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc) +inference_analysis_test(test_fluid_to_data_flow_graph_pass SRCS fluid_to_data_flow_graph_pass_tester.cc) +inference_analysis_test(test_subgraph_splitter SRCS subgraph_splitter_tester.cc) +inference_analysis_test(test_dfg_graphviz_draw_pass SRCS dfg_graphviz_draw_pass_tester.cc) +inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_tester.cc) +inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc) +inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc) +inference_analysis_test(test_analyzer SRCS analyzer_tester.cc) +inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc) diff --git a/paddle/fluid/inference/analysis/README.md b/paddle/fluid/inference/analysis/README.md new file mode 100644 index 0000000000000000000000000000000000000000..70adb4a974cc5f9911cb302840bbef7ec2591505 --- /dev/null +++ b/paddle/fluid/inference/analysis/README.md @@ -0,0 +1,58 @@ +# Inference Analysis + +The `inference/analysis` module is used to analyze and optimize the inference program, +it references some philosophy from `LLVM/analysis`, +and make the various optimization features be pluggable and co-exist in a pipeline. + +We borrowed some concepts from LLVM, such as + +- [Pass](./pass.h)es to implement optimization that traverse the inference program, +- [DataFlowGraph](./data_flow_graph.h) to represent the data flow graph built from a program, +- [PassManager](./pass_manager.h) to manage a sequence of `Pass`es over a graph. + +There are some other basic concepts here + +- [Node](./node.h), the node in a `DataFlowGraph`, + - `Function`, the Operator in Fluid, + - `Value`, the Variable in Fluid; +- [Argument](./argument.h), the argument that treat as the input and output of all `Pass`es in the pipeline, + +## How it works + +The `inference/analysis` module make all the passes in a pipeline, and works in such way: + +1. Build a `DataFlowGraph` from a Fluid inference ProgramDesc, +2. Call the middle passes one by one, the same `DataFlowGraph` is passed across all the passes, +3. Transform a new ProgramDesc from the modified `DataFlowGraph`. + +The new optimization features can be added as an independent `Pass` and controlled by gflags, +each pass will generate unified debug information or visualization for better debugging. + +## Supported Passes + +### `FluidToDataFlowGraphPass` +Transform the fluid `ProgramDesc` to a `DataFlowGraph` to give an abstract representation for all the middle passes, +this should be the first pass of the pipeline. + +### `DataFlowGraphToFluidPass` +Generate a final `ProgramDesc` from a data flow graph, this should be the last pass of the pipeline. + +### `TensorRTSubgraphNodeMarkPass` +Mark the `Node` that are supported by TensorRT, +this pass will generate a visualization file which can be used for debugging. + +### `TensorRTSubGraphPass` +Split the sub-graph that are can be accelerated by TensorRT. + +### `DFG_GraphvizDrawPass` +This pass is just for debug, it will visualize the `DataFlowGraph` using the [graphviz](http://www.graphviz.org) tool. + +It can be used as a helper class that draws the modified graph after each pass. + +## Utilities + +There is some helper legacy/function/class for analysis. + +- [dot.h](./dot.h) give a easy to use interface for generating `DOT` codes, +- [graph_traits.h](./graph_traits.h) contains the interfaces of the graph traversal algorithms, it uses `iterator`to make the algorithms easy to share across different passes, +there are some implementations in [data_flow_graph.cc](./data_flow_graph.cc) , such as BFS and DFS.. diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc new file mode 100644 index 0000000000000000000000000000000000000000..9318f1089781b30468cf4d3c7151d0dd26e50a9c --- /dev/null +++ b/paddle/fluid/inference/analysis/analyzer.cc @@ -0,0 +1,100 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/analyzer.h" +#include +#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h" +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include "paddle/fluid/inference/analysis/model_store_pass.h" +#include "paddle/fluid/inference/analysis/pass_manager.h" +#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h" +#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h" + +namespace paddle { + +DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, true, + "Enable subgraph to TensorRT engine for acceleration"); + +DEFINE_string(inference_analysis_graphviz_log_root, "./", + "Graphviz debuger for data flow graphs."); + +DEFINE_string(inference_analysis_output_storage_path, "", + "optimized model output path"); + +namespace inference { +namespace analysis { + +class DfgPassManagerImpl final : public DfgPassManager { + public: + DfgPassManagerImpl() { + // TODO(Superjomn) set the key with pass reprs. + AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass); + if (FLAGS_inference_analysis_enable_tensorrt_subgraph_engine) { + auto trt_teller = [&](const Node* node) { + std::unordered_set teller_set( + {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax"}); + if (!node->IsFunction()) return false; + + const auto* func = static_cast(node); + if (teller_set.count(func->func_type())) { + return true; + } else { + return false; + } + }; + + AddPass("tensorrt-subgraph-marker", + new TensorRTSubgraphNodeMarkPass(trt_teller)); + AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller)); + } + AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass); + if (!FLAGS_inference_analysis_output_storage_path.empty()) { + AddPass("model-store-pass", new ModelStorePass); + } + } + + std::string repr() const override { return "dfg-pass-manager"; } + std::string description() const override { return "DFG pass manager."; } + + private: + void AddPass(const std::string& name, Pass* pass) { + LOG(INFO) << "Adding pass " << name; + Register(name, pass); + AddGraphvizDebugerPass(pass); + } + + // Add the graphviz debuger pass if the parent pass has one. + void AddGraphvizDebugerPass(Pass* pass) { + auto* debuger_pass = pass->CreateGraphvizDebugerPass(); + if (debuger_pass) { + LOG(INFO) << " - register debug pass [" << debuger_pass->repr() << "]"; + Register(debuger_pass->repr(), debuger_pass); + } + } +}; + +Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); } + +void Analyzer::Run(Argument* argument) { + for (auto& x : data_) { + PADDLE_ENFORCE(x->Initialize(argument)); + x->RunAll(); + PADDLE_ENFORCE(x->Finalize()); + } +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h new file mode 100644 index 0000000000000000000000000000000000000000..c82fdfff86c91b4e07e3c1b80987d3d8d796ad23 --- /dev/null +++ b/paddle/fluid/inference/analysis/analyzer.h @@ -0,0 +1,65 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +/* + * This file contains Analyzer, an class that exposed as a library that analyze + * and optimize Fluid ProgramDesc for inference. Similar to LLVM, it has + * multiple flags to + * control whether an process is applied on the program. + * + * The processes are called Passes in analysis, the Passes are placed in a + * pipeline, the first Pass is the FluidToDataFlowGraphPass which transforms a + * Fluid ProgramDesc to + * a data flow graph, the last Pass is DataFlowGraphToFluidPass which transforms + * a data flow graph to a Fluid ProgramDesc. The passes in the middle of the + * pipeline can be any Passes + * which take a node or data flow graph as input. + * + * The Analyzer can be used in two methods, the first is a executable file which + * can be used to pre-process the inference model and can be controlled by + * passing difference command flags; + * the other way is to compose inside the inference API as a runtime pre-process + * phase in the inference service. + */ + +#include +#include "paddle/fluid/inference/analysis/pass.h" +#include "paddle/fluid/inference/analysis/pass_manager.h" + +namespace paddle { + +// TODO(Superjomn) add a definition flag like PADDLE_WITH_TENSORRT and hide this +// flag if not available. +DECLARE_bool(inference_analysis_enable_tensorrt_subgraph_engine); +DECLARE_string(inference_analysis_graphviz_log_root); +DECLARE_string(inference_analysis_output_storage_path); + +namespace inference { +namespace analysis { + +class Analyzer : public OrderedRegistry { + public: + // Register all the pass-managers. + Analyzer(); + + void Run(Argument* argument); + + DISABLE_COPY_AND_ASSIGN(Analyzer); +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/analyzer_main.cc b/paddle/fluid/inference/analysis/analyzer_main.cc new file mode 100644 index 0000000000000000000000000000000000000000..5e1fe3eb797cdced56a61aa2db0c3d18601824f8 --- /dev/null +++ b/paddle/fluid/inference/analysis/analyzer_main.cc @@ -0,0 +1,33 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file implements analysizer -- an executation help to analyze and + * optimize trained model. + */ +#include "paddle/fluid/inference/analysis/analyzer.h" +#include +#include + +int main(int argc, char** argv) { + google::ParseCommandLineFlags(&argc, &argv, true); + using paddle::inference::analysis::Analyzer; + using paddle::inference::analysis::Argument; + + Argument argument; + Analyzer analyzer; + analyzer.Run(&argument); + + return 0; +} diff --git a/paddle/fluid/inference/analysis/analyzer_tester.cc b/paddle/fluid/inference/analysis/analyzer_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..24bfb3993cf569561980006b6627b56327dd0f67 --- /dev/null +++ b/paddle/fluid/inference/analysis/analyzer_tester.cc @@ -0,0 +1,41 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/analyzer.h" +#include +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(Analyzer, analysis_without_tensorrt) { + FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = false; + Argument argument; + argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir)); + Analyzer analyser; + analyser.Run(&argument); +} + +TEST(Analyzer, analysis_with_tensorrt) { + FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = true; + Argument argument; + argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir)); + Analyzer analyser; + analyser.Run(&argument); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/argument.cc b/paddle/fluid/inference/analysis/argument.cc new file mode 100644 index 0000000000000000000000000000000000000000..cb0263d5d98e86b612696ebde66d17fb2543809b --- /dev/null +++ b/paddle/fluid/inference/analysis/argument.cc @@ -0,0 +1,15 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/argument.h" diff --git a/paddle/fluid/inference/analysis/argument.h b/paddle/fluid/inference/analysis/argument.h new file mode 100644 index 0000000000000000000000000000000000000000..a17d6281a2976f0600c7ce94c2d43e65d30de265 --- /dev/null +++ b/paddle/fluid/inference/analysis/argument.h @@ -0,0 +1,72 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file defines the class Argument, which is the input and output of the + * analysis module. All the fields that needed either by Passes or PassManagers + * are contained in Argument. + * + * TODO(Superjomn) Find some way better to contain the fields when it grow too + * big. + */ + +#pragma once + +#include +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/analysis/data_flow_graph.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * The argument definition of both Pass and PassManagers. + * + * All the fields should be registered here for clearness. + */ +struct Argument { + Argument() = default; + explicit Argument(const std::string& fluid_model_dir) + : fluid_model_dir(new std::string(fluid_model_dir)) {} + // The directory of the trained model. + std::unique_ptr fluid_model_dir; + // The path of `__model__` and `param`, this is used when the file name of + // model and param is changed. + std::unique_ptr fluid_model_program_path; + std::unique_ptr fluid_model_param_path; + + // The graph that process by the Passes or PassManagers. + std::unique_ptr main_dfg; + + // The original program desc. + std::unique_ptr origin_program_desc; + + // The processed program desc. + std::unique_ptr transformed_program_desc; + + // The output storage path of ModelStorePass. + std::unique_ptr model_output_store_path; +}; + +#define UNLIKELY(condition) __builtin_expect(static_cast(condition), 0) +#define ANALYSIS_ARGUMENT_CHECK_FIELD(field__) \ + if (UNLIKELY(!(field__))) { \ + LOG(ERROR) << "field " << #field__ << " should be set."; \ + return false; \ + } + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph.cc b/paddle/fluid/inference/analysis/data_flow_graph.cc index 4220451e3caee62caa51af5bc33d6dd3fd891018..7f64bc75ae8ad40a268739cdc36051e76af9f49a 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph.cc @@ -14,12 +14,13 @@ limitations under the License. */ #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/dot.h" +#include "paddle/fluid/inference/analysis/node.h" namespace paddle { namespace inference { namespace analysis { -// It is a better idea that the inputs and outputs of this graph is set manully +// It is a better idea that the inputs and outputs of this graph is set manually // before, but there must be a Pass that helps to prune the unnecessary ops that // do not contribute to the given targets, so in this pass, analysis and get the // inputs and outputs is OK. @@ -49,6 +50,25 @@ void DataFlowGraph::Build() { outputs.push_back(out); } } + + Clean(); +} + +void DataFlowGraph::Clean() { + for (auto &node : nodes.nodes()) { + std::unordered_set inlinks_set(node->inlinks.begin(), + node->inlinks.end()); + std::unordered_set outlinks_set(node->outlinks.begin(), + node->outlinks.end()); + if (inlinks_set.size() < node->inlinks.size()) { + LOG(INFO) << "Clean: node " << node->repr() << " prune duplicate inputs"; + node->inlinks.assign(inlinks_set.begin(), inlinks_set.end()); + } + if (outlinks_set.size() < node->outlinks.size()) { + LOG(INFO) << "Clean: node " << node->repr() << " prune duplicate inputs"; + node->outlinks.assign(outlinks_set.begin(), outlinks_set.end()); + } + } } std::string DataFlowGraph::DotString() const { @@ -57,19 +77,7 @@ std::string DataFlowGraph::DotString() const { // Add nodes for (size_t i = 0; i < nodes.size(); i++) { const Node &node = nodes.Get(i); - switch (node.type()) { - case Node::Type::kValue: - dot.AddNode(node.repr(), node.dot_attrs()); - break; - case Node::Type::kFunction: - dot.AddNode(node.repr(), node.dot_attrs()); - break; - case Node::Type::kFunctionBlock: - dot.AddNode(node.repr(), node.dot_attrs()); - break; - default: - PADDLE_THROW("unsupported Node type %d", static_cast(node.type())); - } + dot.AddNode(node.repr(), node.dot_attrs()); } // Add edges @@ -82,6 +90,20 @@ std::string DataFlowGraph::DotString() const { return dot.Build(); } +std::string DataFlowGraph::HumanReadableInfo(bool show_values, + bool show_functions) const { + std::stringstream values, functions; + for (auto &n : nodes.nodes()) { + if (show_values && n->IsValue()) { + values << n->repr() << "\n"; + } + if (show_functions && n->IsFunction()) { + functions << n->repr() << "\n"; + } + } + return "Values:\n" + values.str() + "\n\n" + "Functions:\n" + functions.str(); +} + // // NodesBFSIterator // @@ -138,7 +160,7 @@ bool GraphTraits::NodesBFSIterator::operator==( if ((!queue_.empty()) && (!other.queue_.empty())) { return queue_.front() == other.queue_.front() && visited_.size() == other.visited_.size(); // here need to check the - // equality of queue and + // equality of queue and // visited. Just a light but week implementation. } return false; @@ -200,6 +222,149 @@ Node *GraphTraits::NodesDFSIterator::operator->() { return stack_.top(); } +inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) { + return node.inlinks.size() == n; +} + +GraphTraits::NodesTSIterator::NodesTSIterator( + const std::vector &source) { + PADDLE_ENFORCE(!source.empty(), + "Start points of topological sorting should not be empty!"); + // CHECK all the inputs' in-degree is 0 + for (auto *node : source) { + PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0)); + } + + std::unordered_set visited; + std::unordered_set to_visit{source.begin(), source.end()}; + + std::vector inlink_visited; + while (!to_visit.empty()) { + std::vector queue(to_visit.begin(), to_visit.end()); + for (auto *p : queue) { + if (p->deleted()) { + visited.insert(p); + to_visit.erase(p); + continue; + } + inlink_visited.clear(); + + std::copy_if(p->inlinks.begin(), p->inlinks.end(), + std::back_inserter(inlink_visited), + [&](Node *x) { return visited.count(x); }); + + if (inlink_visited.size() == p->inlinks.size()) { + sorted_.push_back(p); + for (auto *_ : p->outlinks) { + if (!visited.count(_)) { + to_visit.insert(_); + } + } + + to_visit.erase(p); + visited.insert(p); + } + } + } +} + +GraphTraits::NodesTSIterator::NodesTSIterator( + const paddle::inference::analysis::GraphTraits< + DataFlowGraph>::NodesTSIterator &other) + : sorted_(other.sorted_), cursor_(other.cursor_) {} + +Node &GraphTraits::NodesTSIterator::operator*() { + PADDLE_ENFORCE_LT(cursor_, sorted_.size()); + return *sorted_[cursor_]; +} + +paddle::inference::analysis::GraphTraits::NodesTSIterator + &GraphTraits::NodesTSIterator::operator++() { + if (++cursor_ >= sorted_.size()) { + sorted_.clear(); + cursor_ = 0; + } + return *this; +} +paddle::inference::analysis::GraphTraits::NodesTSIterator & +GraphTraits::NodesTSIterator::operator=( + const paddle::inference::analysis::GraphTraits< + DataFlowGraph>::NodesTSIterator &other) { + cursor_ = other.cursor_; + sorted_ = other.sorted_; + return *this; +} + +bool GraphTraits::NodesTSIterator::operator==( + const paddle::inference::analysis::GraphTraits< + DataFlowGraph>::NodesTSIterator &other) { + return sorted_ == other.sorted_ && cursor_ == other.cursor_; +} + +Node *GraphTraits::NodesTSIterator::operator->() { + PADDLE_ENFORCE_LT(cursor_, sorted_.size()); + return sorted_[cursor_]; +} + +std::pair, std::vector> +ExtractInputAndOutputOfSubGraph(std::vector &graph) { // NOLINT + std::unordered_set nodes(graph.begin(), graph.end()); + std::unordered_set inputs; + std::unordered_set outputs; + // Input a Value, check whether its inlink is in the subgraph. + auto inlink_in_subgraph = [&](Node *n) { + for (auto *in : n->inlinks) { + if (nodes.count(in)) return true; + } + return false; + }; + for (auto &node : graph) { + for (auto *in : node->inlinks) { + // The Value that is written by nodes inside a sub-graph shouldn't be the + // input of the sub-graph. + if (!nodes.count(in) && in->type() == Node::Type::kValue && + !inlink_in_subgraph(in)) { + inputs.insert(in); + } + } + for (auto *out : node->outlinks) { + if (!nodes.count(out) && out->type() == Node::Type::kValue) { + outputs.insert(out); + } + } + } + return std::make_pair(std::vector(inputs.begin(), inputs.end()), + std::vector(outputs.begin(), outputs.end())); +} + +void FilterRedundantOutputOfSubGraph(DataFlowGraph *graph) { + std::vector op_nodes; + for (auto &node : GraphTraits(graph).nodes_in_TS()) { + if (node.type() == Node::Type::kValue || node.deleted()) { + continue; + } + op_nodes.push_back(&node); + } + size_t op_num = op_nodes.size(); + for (size_t i = 0; i < op_num; i++) { + if (op_nodes[i]->type() == Node::Type::kFunction) continue; + std::unordered_set follow_up_input_names; + for (size_t j = i + 1; j < op_num; j++) { + for (auto *in : op_nodes[j]->inlinks) { + follow_up_input_names.insert(in->name()); + } + } + std::vector filtered_subgraph_outlinks; + for (auto *out : op_nodes[i]->outlinks) { + if (follow_up_input_names.count(out->name())) { + filtered_subgraph_outlinks.push_back(out); + } + } + PADDLE_ENFORCE_GE(filtered_subgraph_outlinks.size(), 1UL); + op_nodes[i]->outlinks = filtered_subgraph_outlinks; + } +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph.h b/paddle/fluid/inference/analysis/data_flow_graph.h index 913e344d371ddf3ea05a53c216e5b3bea8f11c7b..bb3ec6bbc1d9555386aba8837b019d2511653258 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.h +++ b/paddle/fluid/inference/analysis/data_flow_graph.h @@ -36,6 +36,8 @@ namespace analysis { /* * DataFlowGraph - A container of Value and Function Nodes. + * + * This is the base graph for any other type of graphs, such as SSA or CFG. */ struct DataFlowGraph { NodeMap nodes; @@ -47,6 +49,13 @@ struct DataFlowGraph { // Output a DOT graph file for debug. std::string DotString() const; + + std::string HumanReadableInfo(bool show_values = true, + bool show_functions = true) const; + + private: + // Remove duplicate edges and so on. + void Clean(); }; /* @@ -103,6 +112,32 @@ struct GraphTraits { std::unordered_set visited_; }; + // Topological sorting iterator on nodes. + struct NodesTSIterator + : public std::iterator { + NodesTSIterator() = default; + explicit NodesTSIterator(const std::vector &source); + NodesTSIterator(NodesTSIterator &&other) + : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) { + other.cursor_ = 0; + } + NodesTSIterator(const NodesTSIterator &other); + + Node &operator*(); + NodesTSIterator &operator++(); + // TODO(Superjomn) current implementation just compare the first + // element, need to compare the graph and all the elements in the queue and + // set. + NodesTSIterator &operator=(const NodesTSIterator &other); + bool operator==(const NodesTSIterator &other); + bool operator!=(const NodesTSIterator &other) { return !(*this == other); } + Node *operator->(); + + private: + std::vector sorted_; + size_t cursor_{0}; + }; + explicit GraphTraits(DataFlowGraph *graph) : graph_(graph) {} // default use BFS to visit the nodes. @@ -115,17 +150,24 @@ struct GraphTraits { iterator_range nodes_in_DFS() { return iterator_range(nodes_dfs_begin(), nodes_dfs_end()); } + iterator_range nodes_in_TS() { + return iterator_range(nodes_ts_begin(), nodes_ts_end()); + } private: NodesBFSIterator nodes_bfs_begin() { return NodesBFSIterator(graph_->inputs); } NodesBFSIterator nodes_bfs_end() { return NodesBFSIterator(); } + NodesDFSIterator nodes_dfs_begin() { return NodesDFSIterator(graph_->inputs); } NodesDFSIterator nodes_dfs_end() { return NodesDFSIterator(); } + NodesTSIterator nodes_ts_begin() { return NodesTSIterator(graph_->inputs); } + NodesTSIterator nodes_ts_end() { return NodesTSIterator(); } + private: DataFlowGraph *graph_; }; @@ -133,30 +175,10 @@ struct GraphTraits { // Extract the inputs and outputs of a graph. The inputs and outputs of a // sub-graph is the inputs nodes and output nodes that doesn't inside the // sub-graph. -std::pair< - std::vector, - std::vector< - Node *>> static ExtractInputAndOutputOfSubGraph(std::vector - &graph) { - std::unordered_set nodes(graph.begin(), graph.end()); - std::unordered_set inputs; - std::unordered_set outputs; - for (auto &node : graph) { - for (auto *in : node->inlinks) { - if (!nodes.count(in) && in->type() == Node::Type::kValue) { - inputs.insert(in); - } - } - for (auto *out : node->outlinks) { - if (!nodes.count(out) && out->type() == Node::Type::kValue) { - outputs.insert(out); - } - } - } - return std::make_pair(std::vector(inputs.begin(), inputs.end()), - std::vector(outputs.begin(), outputs.end())); -} +std::pair, std::vector> +ExtractInputAndOutputOfSubGraph(std::vector &graph); // NOLINT +void FilterRedundantOutputOfSubGraph(DataFlowGraph *graph); } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc index 9d7cceeb65888b8ba3fdf39e88fc2877abd82d11..a881262665f156812da9e1576aa29b05fc398499 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc @@ -20,15 +20,15 @@ namespace inference { namespace analysis { TEST(DataFlowGraph, BFS) { - auto desc = LoadProgramDesc(); + auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); auto dfg = ProgramDescToDFG(desc); dfg.Build(); - for (auto* in : dfg.inputs) { + for (auto *in : dfg.inputs) { LOG(INFO) << "inputs: " << in->name() << " " << static_cast(in->type()); } - for (auto* out : dfg.outputs) { + for (auto *out : dfg.outputs) { LOG(INFO) << "outputs: " << out->name() << " " << static_cast(out->type()); } @@ -44,7 +44,7 @@ TEST(DataFlowGraph, BFS) { } TEST(DataFlowGraph, DFS) { - auto desc = LoadProgramDesc(); + auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); auto dfg = ProgramDescToDFG(desc); dfg.Build(); GraphTraits trait(&dfg); @@ -57,6 +57,71 @@ TEST(DataFlowGraph, DFS) { ASSERT_EQ(count, dfg.nodes.size()); } +// Topological sorting. +/* + * Graph topology + * inputs: 0, 1, 2 + * 0 -> 4 + * 0 -> 5 + * 1 -> 6 + * 2 -> 7 + * 4 -> 5 + * 4 -> 7 + * 4 -> 3 + * 7 -> 3 + */ +TEST(DataFlowGraph, TS) { + DataFlowGraph graph; + + for (int i = 0; i < 8; i++) { + auto *node = graph.nodes.Create(Node::Type::kValue); + node->SetName("node-" + std::to_string(i)); + } + + auto add_link = [&](int i, int j) { + Node *source = graph.nodes.GetMutable(i); + Node *target = graph.nodes.GetMutable(j); + target->inlinks.push_back(source); + source->outlinks.push_back(target); + }; + + graph.inputs.push_back(graph.nodes.GetMutable(0)); + graph.inputs.push_back(graph.nodes.GetMutable(1)); + graph.inputs.push_back(graph.nodes.GetMutable(2)); + + add_link(0, 4); + add_link(0, 5); + add_link(1, 6); + add_link(2, 7); + add_link(4, 5); + add_link(4, 7); + add_link(4, 3); + add_link(7, 3); + + auto its = GraphTraits(&graph).nodes_in_TS(); + std::vector sorted_ids; + for (auto it = its.begin(); it != its.end(); ++it) { + LOG(INFO) << it->name(); + sorted_ids.push_back(it->id()); + } + + // Assert a occurs prior to b in the sorted_ids. + auto assert_positive_sequence_pair = [&](int a, int b) { + auto a_offset = std::find(sorted_ids.begin(), sorted_ids.end(), a); + auto b_offset = std::find(sorted_ids.begin(), sorted_ids.end(), b); + ASSERT_LT(a_offset, b_offset); + }; + + assert_positive_sequence_pair(2, 7); + assert_positive_sequence_pair(7, 3); + assert_positive_sequence_pair(4, 3); + assert_positive_sequence_pair(0, 4); + assert_positive_sequence_pair(0, 5); + assert_positive_sequence_pair(1, 6); + assert_positive_sequence_pair(4, 5); + assert_positive_sequence_pair(4, 7); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..18c32fa09199003f17183207828cdfe4e627ae1a --- /dev/null +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc @@ -0,0 +1,267 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h" +#include +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/proto_desc.h" +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" + +namespace paddle { +namespace inference { + +DEFINE_int32(tensorrt_max_batchsize, 3, "TensorRT maximum batch size"); +DEFINE_int32(tensorrt_workspace_size, 2048, "TensorRT workspace size"); + +namespace analysis { + +using framework::proto::ProgramDesc; + +std::vector ExtractParameters( + const std::vector> &nodes); + +bool DataFlowGraphToFluidPass::Initialize(Argument *argument) { + ANALYSIS_ARGUMENT_CHECK_FIELD(argument) + ANALYSIS_ARGUMENT_CHECK_FIELD(argument->origin_program_desc) + PADDLE_ENFORCE(!argument->transformed_program_desc); + // The transformed_program_desc should inherit all the VarDesc and BlockDesc + // from the original program desc. The operators of the main block(the first + // block) should rewritten by data flow graph. + argument->transformed_program_desc.reset( + new ProgramDesc(*argument->origin_program_desc)); + argument->transformed_program_desc->mutable_blocks(framework::kRootBlockIndex) + ->clear_ops(); + desc_ = argument->transformed_program_desc.get(); + argument_ = argument; + return true; +} + +bool DataFlowGraphToFluidPass::Finalize() { return true; } + +void DataFlowGraphToFluidPass::Run(DataFlowGraph *graph) { + FilterRedundantOutputOfSubGraph(graph); + LOG(INFO) << "graph.inputs " << graph->inputs.size(); + for (auto &node : GraphTraits(graph).nodes_in_TS()) { + if (node.deleted()) continue; + + switch (node.type()) { + case Node::Type::kFunction: { + LOG(INFO) << "add function " << node.repr(); + AddFluidOp(&node); + } break; + case Node::Type::kFunctionBlock: { + LOG(INFO) << "add engine op " << node.repr() << " , " + << static_cast(&node)->subgraph.size(); + AddEngineOp(&node); + } break; + default: + continue; + } + } + + PADDLE_ENFORCE(argument_->transformed_program_desc.get()); +} + +void DataFlowGraphToFluidPass::AddFluidOp(Node *node) { + auto *ori_op = static_cast(node->pb_desc()); + // currently only the main block is analyzed. + auto *main_block = desc_->mutable_blocks(framework::kRootBlockIndex); + auto *op = main_block->add_ops(); + *op = *ori_op; // copy the attributes, by default, these will not be changed + // by analysis phrase. + // The inputs and outputs of the existing ops are not changed by tensorrt + // subgraph pass. + // NOTE It might be changed by other passes in the long run. +} + +void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph, + framework::proto::BlockDesc *block) { + static int counter{0}; + PADDLE_ENFORCE(node->IsFunctionBlock()); + framework::OpDesc desc; + auto *func = static_cast(node); + + // collect inputs + std::unordered_set input_names; + for (auto *x : func->inlinks) { + input_names.insert(x->name()); + } + desc.SetInput( + "Xs", std::vector(input_names.begin(), input_names.end())); + + std::unordered_set output_names; + for (auto *x : func->outlinks) { + output_names.insert(x->name()); + } + + std::vector output_temp(output_names.begin(), + output_names.end()); + desc.SetOutput("Ys", output_temp); + desc.SetType("tensorrt_engine"); + + std::unordered_map output_name_map; + + // The following procedure is used to rename all the intermediate + // variables and the output variables of the subgraph. + // Why we do this? + // During the transition from fluid OP to tensorrt OP, we map + // the input and output Tensor(fluid data structure) of fluid OP + // to the correspondin ITensor (trt data structure) through the + // Tensor name. When we set up ITensor for an variable, we must + // ensure that it has not been set before. + // If there is variable in the fluid graph, which is not only the + // input of a OP, but also the output of a Op, there will be problems. + // So we have to rename the variable in the subgraph to make sure + // it is either an OP's input or an OP's output. + + auto subgraph_nodes = func->subgraph; + for (int index = 0; index < block->ops_size(); index++) { + framework::proto::OpDesc *op = block->mutable_ops(index); + auto correspond_node = subgraph_nodes[index]; + PADDLE_ENFORCE_EQ(correspond_node->name(), op->type()); + + std::unordered_map var2id; + for (auto *in_var : correspond_node->inlinks) { + var2id[in_var->name()] = in_var->id(); + } + // rename for the input variables of op inside subgraph + for (int i = 0; i < op->inputs_size(); i++) { + framework::proto::OpDesc_Var *in_var = op->mutable_inputs(i); + std::vector replaced_names; + for (int k = 0; k < in_var->arguments_size(); k++) { + std::string arg_value = in_var->arguments(k); + if (input_names.count(arg_value)) { + replaced_names.push_back(arg_value); + } else { + replaced_names.push_back(arg_value + + std::to_string(var2id[arg_value])); + } + } + in_var->clear_arguments(); + for (size_t k = 0; k < replaced_names.size(); k++) { + in_var->add_arguments(replaced_names[k]); + } + } + var2id.clear(); + for (auto out_var : correspond_node->outlinks) { + var2id[out_var->name()] = out_var->id(); + } + + // rename for the output variables of op inside subgraph + for (int i = 0; i < op->outputs_size(); i++) { + framework::proto::OpDesc_Var *out_var = op->mutable_outputs(i); + std::vector replaced_names; + for (int k = 0; k < out_var->arguments_size(); k++) { + std::string arg_value = out_var->arguments(k); + if (output_names.count(arg_value)) { + output_name_map[arg_value] = + arg_value + std::to_string(var2id[arg_value]); + } + replaced_names.push_back(arg_value + std::to_string(var2id[arg_value])); + } + out_var->clear_arguments(); + for (size_t k = 0; k < replaced_names.size(); k++) { + out_var->add_arguments(replaced_names[k]); + } + } + } + // When tensorrt engine runs at the end of the operation, + // output_mapping help us copy the data from the renamed ITensor + // to Tensor. + std::vector output_mapping; + for (auto name : output_names) { + PADDLE_ENFORCE(output_name_map.count(name) != 0); + output_mapping.push_back(output_name_map[name]); + } + + PADDLE_ENFORCE(!block->vars().empty(), "the block has no var-desc"); + // Set attrs + SetAttr(desc.Proto(), "subgraph", block->SerializeAsString()); + SetAttr(desc.Proto(), "engine_uniq_key", "trt-" + std::to_string(counter++)); + SetAttr(desc.Proto(), "max_batch", FLAGS_tensorrt_max_batchsize); + SetAttr(desc.Proto(), "max_workspace", FLAGS_tensorrt_workspace_size); + SetAttr(desc.Proto(), "parameters", ExtractParameters(graph.nodes.nodes())); + SetAttr(desc.Proto(), "output_name_mapping", output_mapping); + node->SetPbMsg(desc.Proto()->SerializeAsString()); +} + +std::vector ExtractParameters( + const std::vector> &nodes) { + std::vector parameters; + for (const auto &node : nodes) { + if (!node->IsValue()) continue; + PADDLE_ENFORCE(!node->pb_msg().empty(), "pb_msg should be set first"); + framework::proto::VarDesc var; + var.ParseFromString(node->pb_msg()); + if (var.persistable()) { + parameters.push_back(var.name()); + } + } + return parameters; +} + +void DataFlowGraphToFluidPass::AddEngineOp(Node *node) { + // TODO(Superjomn) Here need to expose some arguments for default setting. + PADDLE_ENFORCE(node->IsFunctionBlock()); + auto *block_node = static_cast(node); + framework::proto::BlockDesc proto; + framework::BlockDesc block_desc(nullptr, &proto); + block_desc.Proto()->set_parent_idx(-1); + block_desc.Proto()->set_idx(0); + LOG(INFO) << "origin variable size: " + << argument_->origin_program_desc->blocks(0).vars().size(); + LOG(INFO) << "transformed variable size: " + << block_desc.Proto()->vars().size(); + // copy ops. + + for (auto *node : block_node->subgraph) { + auto *op = block_desc.AppendOp(); + PADDLE_ENFORCE(!node->pb_msg().empty()); + op->Proto()->ParseFromString(node->pb_msg()); + } + + *block_desc.Proto()->mutable_vars() = + argument_->origin_program_desc->blocks(0).vars(); + PADDLE_ENFORCE(!block_desc.Proto()->vars().empty()); + CreateTrtEngineOp(node, *argument_->main_dfg, block_desc.Proto()); + auto *main_block = desc_->mutable_blocks(framework::kRootBlockIndex); + auto *op = main_block->add_ops(); + PADDLE_ENFORCE(!node->pb_msg().empty(), "failed to set desc for block"); + op->ParseFromString(node->pb_msg()); +} + +namespace { +class DFG_DebuggerPass : public DFG_GraphvizDrawPass { + public: + using Config = DFG_GraphvizDrawPass::Config; + explicit DFG_DebuggerPass(const Config &config) + : DFG_GraphvizDrawPass(config) {} + + std::string repr() const override { return "dfg-to-fluid-debuger-pass"; } + + bool Finalize() override { return true; } +}; +} // namespace + +Pass *DataFlowGraphToFluidPass::CreateGraphvizDebugerPass() const { + return new DFG_DebuggerPass(DFG_GraphvizDrawPass::Config( + FLAGS_inference_analysis_graphviz_log_root, + "data_flow_graph_to_fluid_graphviz_debugger")); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..59c47365aa6c8ad5886c4515850d264f69cc4670 --- /dev/null +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h @@ -0,0 +1,62 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +/* + * This file implements the transformation from fluid ProgramDesc to data flow + * graph. + */ + +#pragma once + +#include +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { + +DECLARE_int32(tensorrt_max_batchsize); +DECLARE_int32(tensorrt_workspace_size); + +namespace analysis { +class DataFlowGraphToFluidPass final : public DataFlowGraphPass { + public: + DataFlowGraphToFluidPass() = default; + + bool Initialize(Argument *argument) override; + bool Finalize() override; + + void Run(DataFlowGraph *graph) override; + + std::string repr() const override { return "DFG to fluid"; } + std::string description() const override { + return "Transform a DFG to a Fluid ProgramDesc"; + } + + Pass *CreateGraphvizDebugerPass() const override; + + protected: + // Add a Fluid Op into the ProgramDesc. + void AddFluidOp(Node *node); + // Add a EngineOp into the ProgramDesc. + void AddEngineOp(Node *node); + + private: + framework::proto::ProgramDesc *desc_; + Argument *argument_; +}; +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc index dcee75cee50ede1d2b660e88e06544440bd5ef77..4ef381db295b986b91173a728b6d98640f6f4f51 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc @@ -26,22 +26,21 @@ namespace paddle { namespace inference { namespace analysis { -TEST_F(DFG_Tester, Test) { - framework::proto::ProgramDesc new_desc; - DataFlowGraph graph; +TEST(DataFlowGraph, Test) { + Argument argument(FLAGS_inference_model_dir); FluidToDataFlowGraphPass pass0; DataFlowGraphToFluidPass pass1; - pass0.Initialize(desc); - pass1.Initialize(&new_desc); + ASSERT_TRUE(pass0.Initialize(&argument)); + ASSERT_TRUE(pass1.Initialize(&argument)); - pass0.Run(&graph); - pass1.Run(&graph); + pass0.Run(argument.main_dfg.get()); + pass1.Run(argument.main_dfg.get()); pass0.Finalize(); pass1.Finalize(); - LOG(INFO) << graph.nodes.size(); + LOG(INFO) << argument.main_dfg->nodes.size(); } }; // namespace analysis diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..c05b0e5d4690d0a447edf63a149903704bc2c9be --- /dev/null +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc @@ -0,0 +1,59 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +int DFG_GraphvizDrawPass::counter_{0}; + +void DFG_GraphvizDrawPass::Run(DataFlowGraph *graph) { + auto content = Draw(graph); + auto dot_path = GenDotPath(); + std::ofstream file(dot_path); + file.write(content.c_str(), content.size()); + file.close(); + + auto png_path = dot_path.substr(0, dot_path.size() - 4) + ".png"; + std::string message; + LOG(INFO) << "draw to " << png_path; + ExecShellCommand("dot -Tpng " + dot_path + " -o " + png_path, &message); +} + +std::string DFG_GraphvizDrawPass::Draw(DataFlowGraph *graph) { + Dot dot; + // Add nodes + for (size_t i = 0; i < graph->nodes.size(); i++) { + const Node &node = graph->nodes.Get(i); + if (config_.display_deleted_node || !node.deleted()) { + dot.AddNode(node.repr(), node.dot_attrs()); + } + } + // Add edges + for (size_t i = 0; i < graph->nodes.size(); i++) { + const Node &node = graph->nodes.Get(i); + if (!config_.display_deleted_node && node.deleted()) continue; + for (auto &out : node.outlinks) { + if (!config_.display_deleted_node && out->deleted()) continue; + dot.AddEdge(node.repr(), out->repr(), {}); + } + } + return dot.Build(); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..17445ab4407a159ca11345bc9a9226b3ad0044f0 --- /dev/null +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h @@ -0,0 +1,78 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file create an DFG_GraphvizDrawPass which helps to draw a data flow + * graph's structure using graphviz. + */ + +#pragma once + +#include +#include +#include "paddle/fluid/inference/analysis/dot.h" +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Output a dot file and write to some place. + */ +class DFG_GraphvizDrawPass : public DataFlowGraphPass { + public: + struct Config { + Config(const std::string &dir, const std::string &id, + bool display_deleted_node = false) + : dir(dir), id(id), display_deleted_node(display_deleted_node) {} + + // The directory to store the .dot or .png files. + const std::string dir; + // The identifier for this dot file. + const std::string id; + // Whether to display deleted nodes, default false. + const bool display_deleted_node; + }; + + explicit DFG_GraphvizDrawPass(const Config &config) : config_(config) {} + + bool Initialize(Argument *argument) override { return true; } + void Run(DataFlowGraph *graph) override; + bool Finalize() override { return true; } + + std::string repr() const override { return "DFG graphviz drawer"; } + std::string description() const override { + return "Debug a DFG by draw with graphviz"; + } + + protected: + // A counter to add a number prefix to the debugger image output so that they + // will sort in the triggered order. + static int counter_; + + // Path of the dot file to output. + std::string GenDotPath() const { + return config_.dir + "/" + std::to_string(counter_++) + "-graph_" + + config_.id + ".dot"; + } + + virtual std::string Draw(DataFlowGraph *graph); + + Config config_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..928be7917047382d9b86294f6039b26b0ebf6f49 --- /dev/null +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc @@ -0,0 +1,54 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" + +#include +#include +#include +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(DFG_GraphvizDrawPass, dfg_graphviz_draw_pass_tester) { + Argument argument(FLAGS_inference_model_dir); + FluidToDataFlowGraphPass pass0; + ASSERT_TRUE(pass0.Initialize(&argument)); + pass0.Run(argument.main_dfg.get()); + + // auto dfg = ProgramDescToDFG(*argument.origin_program_desc); + + DFG_GraphvizDrawPass::Config config("./", "test"); + DFG_GraphvizDrawPass pass(config); + pass.Initialize(&argument); + pass.Run(argument.main_dfg.get()); + + // test content + std::ifstream file("./0-graph_test.dot"); + ASSERT_TRUE(file.is_open()); + + std::string line; + int no{0}; + while (std::getline(file, line)) { + no++; + } + // DFG is sensitive to ProgramDesc, be careful to change the existing models. + ASSERT_EQ(no, 83); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc index 9f67c989cca4a936cd320b73efaae277263fb3e2..511631d3e067f14bc1230d9e4b4d92dbe604e1d4 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc @@ -12,38 +12,69 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include #include +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" namespace paddle { namespace inference { namespace analysis { -FluidToDataFlowGraphPass::FluidToDataFlowGraphPass() {} - -bool FluidToDataFlowGraphPass::Initialize() { return Pass::Initialize(); } +bool FluidToDataFlowGraphPass::Initialize(Argument *argument) { + ANALYSIS_ARGUMENT_CHECK_FIELD(argument); + if (argument->origin_program_desc) { + LOG(WARNING) << "argument's origin_program_desc is already set, might " + "duplicate called"; + } + if (!argument->fluid_model_program_path) { + ANALYSIS_ARGUMENT_CHECK_FIELD(argument->fluid_model_dir); + argument->fluid_model_program_path.reset( + new std::string(*argument->fluid_model_dir + "/__model__")); + } + ANALYSIS_ARGUMENT_CHECK_FIELD(argument->fluid_model_program_path); + auto program = LoadProgramDesc(*argument->fluid_model_program_path); + argument->origin_program_desc.reset( + new framework::proto::ProgramDesc(program)); -bool FluidToDataFlowGraphPass::Initialize( - const framework::proto::ProgramDesc &desc) { - desc_ = &desc; + if (!argument->main_dfg) { + argument->main_dfg.reset(new DataFlowGraph); + } + desc_ = argument->origin_program_desc.get(); return true; } -bool FluidToDataFlowGraphPass::Finalize() { return Pass::Finalize(); } +bool FluidToDataFlowGraphPass::Finalize() { return true; } void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { + PADDLE_ENFORCE(graph); + PADDLE_ENFORCE(desc_); // insert vars + // The `var2id` keeps a map from a variable's name to its Node-id, the Node-id + // will keep updating to its latest alias during the graph-building. std::unordered_map var2id; auto &main_block = desc_->blocks(framework::kRootBlockIndex); for (int i = 0; i < main_block.vars_size(); i++) { const auto &var = main_block.vars(i); auto *v = graph->nodes.Create(Node::Type::kValue); v->SetName(var.name()); - v->SetExtraInfo(const_cast(static_cast(&var))); + v->SetPbDesc(const_cast(static_cast(&var))); + v->SetPbMsg(var.SerializeAsString()); var2id[var.name()] = v->id(); } + + // The variables in a SSA can only write once, so if a variable is written + // multiple times(quite common in our ProgramDesc design), multiple alias + // Nodes of this variable will be created, and each will just write once. + + // An set that keep all the names of the variables(the original, not alias) + // that have been written(as outputs). Once an Op's output variable hit the + // set, it should create a new alias and update the global alias for this + // variable. And that make a Data Flow Graph a SSA. + std::unordered_set unique_written_vars; for (int i = 0; i < main_block.ops_size(); i++) { const auto &op = main_block.ops(i); auto *o = graph->nodes.Create(Node::Type::kFunction); @@ -51,9 +82,10 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { static_cast(o)->SetFuncType(op.type()); // Link to the original protobuf message's memory, make it easier to // generate from a data flow graph to fluid ProgramDesc. - o->SetExtraInfo(const_cast(static_cast(&op))); + o->SetPbDesc(const_cast(static_cast(&op))); + o->SetPbMsg(op.SerializeAsString()); + // set inputs and outputs - // TODO(Superjomn) make sure the InputNames is the real variable name. for (int j = 0; j < op.inputs_size(); j++) { auto &in_var = op.inputs(j); for (int k = 0; k < in_var.arguments_size(); k++) { @@ -66,8 +98,21 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { auto &out_var = op.outputs(j); for (int k = 0; k < out_var.arguments_size(); k++) { auto *out = graph->nodes.GetMutable(var2id[out_var.arguments(k)]); + if (unique_written_vars.count(out)) { + // Loop found, for example, a = op(a), use SSA, change to a1 = op(a). + auto *out_alias = graph->nodes.Create(Node::Type::kValue); + out_alias->SetName(out->name()); + out_alias->SetPbDesc(out->pb_desc()); + out_alias->SetPbMsg(out->pb_msg()); + var2id[out_alias->name()] = + out_alias->id(); // update variable's alias Node + LOG(INFO) << "loop found in graph, create SSA alias node [" + << out_alias->repr() << "] for [" << out->repr() << "]"; + out = out_alias; + } out->inlinks.push_back(o); o->outlinks.push_back(out); + unique_written_vars.insert(out); } } } @@ -75,9 +120,20 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { graph->Build(); } -Pass *FluidToDataFlowGraphPass::CreatePrinterPass( - std::ostream &os, const std::string &banner) const { - return nullptr; +namespace { +class DFG_DebuggerPass : public DFG_GraphvizDrawPass { + public: + using Config = DFG_GraphvizDrawPass::Config; + explicit DFG_DebuggerPass(const Config &config) + : DFG_GraphvizDrawPass(config) {} + std::string repr() const override { return "fluid-to-dfg-debuger-pass"; } + bool Finalize() override { return true; } +}; +} + +Pass *FluidToDataFlowGraphPass::CreateGraphvizDebugerPass() const { + return new DFG_DebuggerPass(DFG_GraphvizDrawPass::Config( + FLAGS_inference_analysis_graphviz_log_root, "fluid-to-dfg-debuger")); } } // namespace analysis diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h index 33517e57becdffc0416f204247eac5feadb7ed82..fb948bf2242abcbc1e841fd3b8457e63358782c5 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h @@ -30,19 +30,23 @@ namespace inference { namespace analysis { /* - * Transform a FluidDesc to a data flow graph. + * Transform a FluidDesc to a SSA. */ class FluidToDataFlowGraphPass final : public DataFlowGraphPass { public: - FluidToDataFlowGraphPass(); - bool Initialize() override; - bool Initialize(const framework::proto::ProgramDesc &desc) override; + FluidToDataFlowGraphPass() = default; + + bool Initialize(Argument *argument) override; bool Finalize() override; void Run(DataFlowGraph *graph) override; - Pass *CreatePrinterPass(std::ostream &os, - const std::string &banner) const override; + std::string repr() const override { return "fluid-to-data-flow-graph"; } + std::string description() const override { + return "transform a fluid ProgramDesc to a data flow graph."; + } + + Pass *CreateGraphvizDebugerPass() const override; private: framework::proto::ProgramDesc const *desc_; diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc index 817d32c92cdbdc234eef9ed5156891c2b11ced4c..d218dcd05015aa4636c16569de4addf4936c8cd5 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc @@ -21,15 +21,16 @@ namespace paddle { namespace inference { namespace analysis { -TEST_F(DFG_Tester, Init) { +TEST(FluidToDataFlowGraphPass, Test) { FluidToDataFlowGraphPass pass; - pass.Initialize(); - pass.Initialize(desc); - DataFlowGraph graph; - pass.Run(&graph); - ASSERT_GT(graph.nodes.size(), 0); + Argument argument(FLAGS_inference_model_dir); + pass.Initialize(&argument); + pass.Run(argument.main_dfg.get()); + // Analysis is sensitive to ProgramDesc, careful to change the original model. + ASSERT_EQ(argument.main_dfg->nodes.size(), 38UL); pass.Finalize(); - LOG(INFO) << '\n' << graph.DotString(); + ASSERT_FALSE(argument.main_dfg->DotString().empty()); + EXPECT_FALSE(argument.main_dfg->inputs.empty()); } } // namespace analysis diff --git a/paddle/fluid/inference/analysis/helper.cc b/paddle/fluid/inference/analysis/helper.cc new file mode 100644 index 0000000000000000000000000000000000000000..ca40c01fc57dbcc2ca16770a1b7d798de8b5625b --- /dev/null +++ b/paddle/fluid/inference/analysis/helper.cc @@ -0,0 +1,60 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/framework/framework.pb.h" + +namespace paddle { +namespace inference { +namespace analysis { + +template <> +void SetAttr(framework::proto::OpDesc *op, const std::string &name, + const std::string &data) { + auto *attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::STRING); + attr->set_s(data); +} +template <> +void SetAttr(framework::proto::OpDesc *op, const std::string &name, + const int &data) { + auto *attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::INT); + attr->set_i(data); +} +template <> +void SetAttr(framework::proto::OpDesc *op, const std::string &name, + const int64_t &data) { + auto *attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::LONG); + attr->set_l(data); +} +template <> +void SetAttr>(framework::proto::OpDesc *op, + const std::string &name, + const std::vector &data) { + auto *attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::STRINGS); + for (const auto &s : data) { + attr->add_strings(s.c_str()); + } +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h index 153dca576bd6734d62f00c4a7cb9b503506b33e2..a0f912b251d5ea29594a7f601d5b2bce91201790 100644 --- a/paddle/fluid/inference/analysis/helper.h +++ b/paddle/fluid/inference/analysis/helper.h @@ -14,16 +14,26 @@ limitations under the License. */ #pragma once +#include +#include #include +#include #include #include +#include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/variable.h" #include "paddle/fluid/platform/enforce.h" namespace paddle { namespace inference { namespace analysis { +template +void SetAttr(framework::proto::OpDesc *op, const std::string &name, + const T &data); + template int AccuDims(Vec &&vec, int size) { int res = 1; @@ -33,7 +43,7 @@ int AccuDims(Vec &&vec, int size) { return res; } -#define SET_TYPE(type__) dic_[typeid(type__).hash_code()] = #type__; +#define SET_TYPE(type__) dic_[std::type_index(typeid(type__))] = #type__; /* * Map typeid to representation. */ @@ -45,14 +55,14 @@ struct DataTypeNamer { template const std::string &repr() const { - auto x = typeid(T).hash_code(); + auto x = std::type_index(typeid(T)); PADDLE_ENFORCE(dic_.count(x), "unknown type for representation"); return dic_.at(x); } - const std::string &repr(size_t &hash) const { // NOLINT - PADDLE_ENFORCE(dic_.count(hash), "unknown type for representation"); - return dic_.at(hash); + const std::string &repr(const std::type_index &type) const { // NOLINT + PADDLE_ENFORCE(dic_.count(type), "unknown type for representation"); + return dic_.at(type); } private: @@ -60,11 +70,10 @@ struct DataTypeNamer { SET_TYPE(int); SET_TYPE(bool); SET_TYPE(float); + SET_TYPE(void *); } - std::unordered_map - dic_; + std::unordered_map dic_; }; #undef SET_TYPE @@ -90,7 +99,7 @@ template class OrderedRegistry { public: T *Register(const std::string &name, T *x) { - PADDLE_ENFORCE(!dic_.count(name)); + PADDLE_ENFORCE(!dic_.count(name), "duplicate key [%s]", name); dic_[name] = data_.size(); data_.emplace_back(std::unique_ptr(x)); return data_.back().get(); @@ -107,6 +116,41 @@ class OrderedRegistry { std::vector> data_; }; +template +T &GetFromScope(const framework::Scope &scope, const std::string &name) { + framework::Variable *var = scope.FindVar(name); + PADDLE_ENFORCE(var != nullptr); + return *var->GetMutable(); +} + +static void ExecShellCommand(const std::string &cmd, std::string *message) { + char buffer[128]; + std::shared_ptr pipe(popen(cmd.c_str(), "r"), pclose); + if (!pipe) { + LOG(ERROR) << "error running command: " << cmd; + return; + } + while (!feof(pipe.get())) { + if (fgets(buffer, 128, pipe.get()) != nullptr) { + *message += buffer; + } + } +} + +static framework::proto::ProgramDesc LoadProgramDesc( + const std::string &model_path) { + std::ifstream fin(model_path, std::ios::in | std::ios::binary); + PADDLE_ENFORCE(fin.is_open(), "Cannot open file %s", model_path); + fin.seekg(0, std::ios::end); + std::string buffer(fin.tellg(), ' '); + fin.seekg(0, std::ios::beg); + fin.read(&buffer[0], buffer.size()); + fin.close(); + framework::proto::ProgramDesc program_desc; + program_desc.ParseFromString(buffer); + return program_desc; +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/model_store_pass.cc b/paddle/fluid/inference/analysis/model_store_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..1c429176424bd5c1d8fa5e015c19d698f966880e --- /dev/null +++ b/paddle/fluid/inference/analysis/model_store_pass.cc @@ -0,0 +1,63 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include + +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/analysis/argument.h" +#include "paddle/fluid/inference/analysis/model_store_pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +void ModelStorePass::Run(DataFlowGraph *x) { + if (!argument_->fluid_model_param_path) { + PADDLE_ENFORCE_NOT_NULL(argument_->fluid_model_dir); + argument_->fluid_model_param_path.reset( + new std::string(*argument_->fluid_model_dir + "param")); + } + PADDLE_ENFORCE_NOT_NULL(argument_->model_output_store_path); + // Directly copy param file to destination. + std::stringstream ss; + // NOTE these commands only works on linux. + ss << "mkdir -p " << *argument_->model_output_store_path; + LOG(INFO) << "run command: " << ss.str(); + PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0); + ss.str(""); + + ss << "cp " << *argument_->fluid_model_dir << "/*" + << " " << *argument_->model_output_store_path; + LOG(INFO) << "run command: " << ss.str(); + PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0); + + // Store program + PADDLE_ENFORCE_NOT_NULL(argument_->transformed_program_desc, + "program desc is not transformed, should call " + "DataFlowGraphToFluidPass first."); + const std::string program_output_path = + *argument_->model_output_store_path + "/__model__"; + std::ofstream file(program_output_path, std::ios::binary); + PADDLE_ENFORCE(file.is_open(), "failed to open %s to write.", + program_output_path); + const std::string serialized_message = + argument_->transformed_program_desc->SerializeAsString(); + file.write(serialized_message.c_str(), serialized_message.size()); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/model_store_pass.h b/paddle/fluid/inference/analysis/model_store_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..fac7083925776b6209d49255c9e67b930cb1250b --- /dev/null +++ b/paddle/fluid/inference/analysis/model_store_pass.h @@ -0,0 +1,53 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file defines ModelStorePass, which store the runtime DFG to a Paddle + * model in the disk, and that model can be reloaded for prediction. + */ + +#pragma once +#include +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class ModelStorePass : public DataFlowGraphPass { + public: + bool Initialize(Argument* argument) override { + if (!argument) { + LOG(ERROR) << "invalid argument"; + return false; + } + argument_ = argument; + return true; + } + + void Run(DataFlowGraph* x) override; + + std::string repr() const override { return "DFG-store-pass"; } + std::string description() const override { + return R"DD(This file defines ModelStorePass, which store the runtime DFG to a Paddle + model in the disk, and that model can be reloaded for prediction again.)DD"; + } + + private: + Argument* argument_{nullptr}; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/model_store_pass_tester.cc b/paddle/fluid/inference/analysis/model_store_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..5f3526dd504e77e58d79b4f675db86a22fd0f26b --- /dev/null +++ b/paddle/fluid/inference/analysis/model_store_pass_tester.cc @@ -0,0 +1,43 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/model_store_pass.h" + +#include +#include +#include "paddle/fluid/inference/analysis/analyzer.h" + +namespace paddle { +namespace inference { +namespace analysis { + +DEFINE_string(inference_model_dir, "", "Model path"); + +TEST(DFG_StorePass, test) { + Analyzer analyzer; + Argument argument(FLAGS_inference_model_dir); + argument.model_output_store_path.reset( + new std::string("./_dfg_store_pass_tmp")); + // disable storage in alalyzer + FLAGS_inference_analysis_output_storage_path = ""; + analyzer.Run(&argument); + + ModelStorePass pass; + pass.Initialize(&argument); + pass.Run(argument.main_dfg.get()); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/node.cc b/paddle/fluid/inference/analysis/node.cc index fe060526080b1ee01aa98f2ff06fb2191eddf9da..f2e918f3ff41d9db0c3ec38561015967bed26f4e 100644 --- a/paddle/fluid/inference/analysis/node.cc +++ b/paddle/fluid/inference/analysis/node.cc @@ -20,6 +20,17 @@ namespace paddle { namespace inference { namespace analysis { +template <> +std::string &NodeAttr::As() { + if (data_.empty()) { + type_index_ = std::type_index(typeid(std::string)); + } + PADDLE_ENFORCE_EQ(type_index_, std::type_index(typeid(std::string))); + return data_; +} + +std::string &NodeAttr::String() { return As(); } + std::vector Value::dot_attrs() const { return std::vector({Dot::Attr("style", "filled,rounded"), Dot::Attr("shape", "box"), @@ -40,6 +51,9 @@ Node *NodeMap::Create(Node::Type type) { case Node::Type::kValue: nodes_.emplace_back(new Value); break; + case Node::Type::kFunctionBlock: + nodes_.emplace_back(new FunctionBlock); + break; default: PADDLE_THROW("Not supported node type."); } diff --git a/paddle/fluid/inference/analysis/node.h b/paddle/fluid/inference/analysis/node.h index 7972ca25c92186a8c55a76de645f4fdbb089e8d3..47e524bc5c4a6b1324d5f182053129311487522d 100644 --- a/paddle/fluid/inference/analysis/node.h +++ b/paddle/fluid/inference/analysis/node.h @@ -25,6 +25,7 @@ limitations under the License. */ #include #include +#include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/inference/analysis/device.h" #include "paddle/fluid/inference/analysis/dot.h" #include "paddle/fluid/inference/analysis/helper.h" @@ -35,6 +36,44 @@ namespace analysis { class NodeMap; +// A helper class to maintain the status from Pass. +struct NodeAttr { + // NOTE T should be a primary type or a struct combined by several primary + // types. + // NOTE the STL containers should not use here. + // Some usages + // Attr attr; + // attr.Bool() = true; + + bool &Bool() { return As(); } + float &Float() { return As(); } + int32_t &Int32() { return As(); } + int64_t &Int64() { return As(); } + void *&Pointer() { return As(); } + std::string &String(); + + private: + template + T &As() { + // init storage in the first usage. + if (data_.empty()) { + VLOG(4) << "resize data to " << sizeof(T); + type_index_ = std::type_index(typeid(T)); + data_.resize(sizeof(T)); + } + PADDLE_ENFORCE(framework::IsType(type_index_), + "type not matched, origin is %s, want %s", + DataTypeNamer::Global().repr(type_index_), + DataTypeNamer::Global().repr()); + PADDLE_ENFORCE_EQ(data_.size(), sizeof(T), "Node attr type recast error"); + return *reinterpret_cast(&data_[0]); + } + + private: + std::string data_; + std::type_index type_index_{typeid(NodeAttr)}; +}; + /* * Node Representation. * @@ -50,8 +89,6 @@ class Node { Node() = default; - struct Attr; - // Cast to a subclass type, Function for example. template Subclass &As() { @@ -71,12 +108,20 @@ class Node { // Get an additional attribute and convert it to T data type. NOTE this will // silently create a new attribute if not exists. - Attr &attr(const std::string &name) { return attrs_[name]; } + NodeAttr &attr(const std::string &name) const { return attrs_[name]; } int id() const { return id_; } - bool deleted() const { return deleted_; } + // The Protobuf description is set/get with a void* to decouple Node interface + // from a specific kind of Protobuf message. + void SetPbDesc(void *pb) { attr("pb_desc").Pointer() = pb; } + void *pb_desc() const { return attr("pb_desc").Pointer(); } + + void SetPbMsg(const std::string &s) { attr("pb_msg").String() = s; } + const std::string &pb_msg() const { return attr("pb_msg").String(); } + void SetDeleted() { deleted_ = true; } + bool deleted() const { return deleted_; } void SetName(const std::string &name) { name_ = name; } const std::string &name() const { return name_; } @@ -84,52 +129,12 @@ class Node { void SetType(Type type) { type_ = type; } Type type() const { return type_; } - void *extra_info() const { return extra_info_; } - void SetExtraInfo(void *extra_info) { extra_info_ = extra_info; } - // Input links. std::vector inlinks; // Output links. std::vector outlinks; - // A helper class to maintain the status from Pass. - // TODO(superjomn) add a checker here to ensure the T is primary. - struct Attr { - // NOTE T should be a primary type or a struct combined by several primary - // types. - // NOTE the STL containers should not use here. - // Some usages - // Attr attr; - // T data; - // attr.data.assign((char*)data, sizeof(data)); - - bool &Bool() { return As(); } - float &Float() { return As(); } - int32_t &Int32() { return As(); } - int64_t &Int64() { return As(); } - - private: - template - T &As() { - // init storage in the first usage. - if (data_.empty()) { - VLOG(4) << "resize data to " << sizeof(T); - type_hash_ = typeid(T).hash_code(); - data_.resize(sizeof(T)); - } - PADDLE_ENFORCE(type_hash_ == typeid(T).hash_code(), - "type not matched, origin is %s, want %s", - DataTypeNamer::Global().repr(type_hash_), - DataTypeNamer::Global().repr()); - PADDLE_ENFORCE_EQ(data_.size(), sizeof(T), "Node attr type recast error"); - return *reinterpret_cast(&data_[0]); - } - - private: - std::string data_; - size_t type_hash_{std::numeric_limits::max()}; - }; - + // Type checks. bool IsFunction() const { return type_ == Node::Type::kFunction; } bool IsValue() const { return type_ == Node::Type::kValue; } bool IsFunctionBlock() const { return type_ == Node::Type::kFunctionBlock; } @@ -148,10 +153,7 @@ class Node { Type type_{Type::kNone}; // Mark this node is deleted by some pass. bool deleted_{false}; - - void *extra_info_; - - mutable std::unordered_map attrs_; + mutable std::unordered_map attrs_; }; class Function; @@ -214,6 +216,10 @@ class Function : public Node { struct FunctionBlock : public Node { std::string repr() const override { return "block-" + std::to_string(id()); } std::vector subgraph; + + protected: + FunctionBlock() { SetType(Node::Type::kFunctionBlock); } + friend class NodeMap; }; class NodeMap { @@ -228,7 +234,7 @@ class NodeMap { void Delete(size_t id); - const std::vector> &nodes() { return nodes_; } + const std::vector> &nodes() const { return nodes_; } size_t size() const { return nodes_.size(); } diff --git a/paddle/fluid/inference/analysis/node_attr_flags.h b/paddle/fluid/inference/analysis/node_attr_flags.h new file mode 100644 index 0000000000000000000000000000000000000000..a3f70e5419a66969e8fb20152a8a8ace39316f57 --- /dev/null +++ b/paddle/fluid/inference/analysis/node_attr_flags.h @@ -0,0 +1,32 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file contains all the flags that declared in Node::Attr. + * + * The Node::Attr is designed to share information between different passes, one + * can get other's attributes in a Node by the flags in this file. + */ +#pragma once +namespace paddle { +namespace inference { +namespace analysis { + +#define DECLARE_NODE_ATTR(flag__) const char ATTR_##flag__[] = #flag__; + +DECLARE_NODE_ATTR(supported_by_tensorrt) // bool + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/pass.h b/paddle/fluid/inference/analysis/pass.h index aa0e8667b5e4a9e6156c25fcad03bb8eee3287f6..6806f9ff7dada2c1e2328e1ffbfd225afefcf474 100644 --- a/paddle/fluid/inference/analysis/pass.h +++ b/paddle/fluid/inference/analysis/pass.h @@ -19,6 +19,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/inference/analysis/argument.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/analysis/node.h" @@ -30,19 +31,11 @@ namespace analysis { class Pass { public: Pass() = default; - virtual ~Pass() {} - // Virtual method overridden by subclasses to do only necessary initialization - // before any pass is run. - virtual bool Initialize() { return false; } - // There is some passes such as FlowToDataFlowGraphPass that needs a - // ProgramDesc. Here use the native ProgramDesc ProtoBuf message, so that it - // only couple with the proto file. - virtual bool Initialize(const framework::proto::ProgramDesc &desc) { - return false; - } - // There are some Passes such as DataFlowGraphToFluidPass that will output a - // ProgramDesc. - virtual bool Initialize(framework::proto::ProgramDesc *desc) { return false; } + virtual ~Pass() = default; + // Mutable Pass. + virtual bool Initialize(Argument *argument) { return false; } + // Readonly Pass. + virtual bool Initialize(const Argument &argument) { return false; } // Virtual method overriden by subclasses to do any necessary clean up after // all passes have run. @@ -50,8 +43,14 @@ class Pass { // Get a Pass appropriate to print the Node this pass operates on. virtual Pass *CreatePrinterPass(std::ostream &os, - const std::string &banner) const = 0; + const std::string &banner) const { + return nullptr; + } + // Create a debugger Pass that draw the DFG by graphviz toolkit. + virtual Pass *CreateGraphvizDebugerPass() const { return nullptr; } + + virtual void Run() { LOG(FATAL) << "not valid"; } // Run on a single Node. virtual void Run(Node *x) { LOG(FATAL) << "not valid"; } // Run on a single Function. @@ -60,6 +59,11 @@ class Pass { virtual void Run(FunctionBlock *x) { LOG(FATAL) << "not valid"; } // Run on a single DataFlowGraph. virtual void Run(DataFlowGraph *x) { LOG(FATAL) << "not valid"; } + + // Human-readable short representation. + virtual std::string repr() const = 0; + // Human-readable long description. + virtual std::string description() const = 0; }; // NodePass process on any Node types. diff --git a/paddle/fluid/inference/analysis/pass_manager.cc b/paddle/fluid/inference/analysis/pass_manager.cc new file mode 100644 index 0000000000000000000000000000000000000000..b428bb22b1f0c5c1a47fc4c46c9070c1ace4a228 --- /dev/null +++ b/paddle/fluid/inference/analysis/pass_manager.cc @@ -0,0 +1,56 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/pass_manager.h" +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +bool PassManager::Initialize(Argument* argument) { + argument_ = argument; + for (auto& pass : data_) { + LOG(INFO) << "Initializing pass " << pass->repr(); + if (!pass->Initialize(argument)) { + LOG(ERROR) << "Failed to initialize pass [" << pass->repr() << "]"; + return false; + } + } + return true; +} + +void DfgPassManager::RunAll() { + PADDLE_ENFORCE(argument_); + for (auto& pass : data_) { + VLOG(4) << "Running pass [" << pass->repr() << "]"; + pass->Run(argument_->main_dfg.get()); + } +} + +void NodePassManager::RunAll() { + PADDLE_ENFORCE(argument_); + PADDLE_ENFORCE(argument_->main_dfg.get()); + auto trait = + GraphTraits(argument_->main_dfg.get()).nodes_in_DFS(); + for (auto& node : trait) { + for (auto& pass : data_) { + pass->Run(&node); + } + } +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/pass_manager.h b/paddle/fluid/inference/analysis/pass_manager.h new file mode 100644 index 0000000000000000000000000000000000000000..81a17e0287a5aef8a328e43380ee3691f5a32379 --- /dev/null +++ b/paddle/fluid/inference/analysis/pass_manager.h @@ -0,0 +1,106 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file defines the logic of pass management. The analysis for inference is + * a pipeline of Passes, a PassManager is a agency that helps to manage the + * executation of the Passes. + * + * There are two modes of Passes, the first one is called NodePass and takes + * an Node as input and output; the second one is called DFGPass and takes a + * DFG(Data Flow Graph) as input and output. It is hard to put all the passes in + * the same pipeline, there are two kinds of PassManagers, both takes a DFG as + * input and output a DFG, but the Passes inside are different: + * + * 1. NodePassManager: the passes inside are all NodePasses, it can have + * different graph trivial algorithm, for example, DFS_NodePassManager will + * trigger the passes in depth first order; + * 2. DfgPassManager: the passes inside are all DfgPasses. + */ + +#pragma once + +#include +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * PassManager is the base class for all pass managers, a pass manager has + * several Pass-es registered, and execute them in the linear order. + */ +class PassManager : public OrderedRegistry { + public: + PassManager() = default; + // Call all the passes' Initialize methods. The desc and data_flow_graph are + // globally shared, so pass them as the arguemnts for all the pass managers. + virtual bool Initialize(const Argument& argument) { return false; } + + virtual bool Initialize(Argument* argument); + + // Call all the passes' Finalize methods. + virtual bool Finalize() { + for (auto& pass : data_) { + if (!pass->Finalize()) { + LOG(ERROR) << "Failed to finalize pass [" << pass->repr() << "]"; + return false; + } + } + return true; + } + + // Run all the passes. + virtual void RunAll() = 0; + + // Short identifier. + virtual std::string repr() const = 0; + // Long description. + virtual std::string description() const = 0; + + virtual ~PassManager() = default; + + protected: + Argument* argument_{nullptr}; +}; + +/* + * A pass manager that process a DFG. + */ +class DfgPassManager : public PassManager { + public: + DfgPassManager() = default; + + void RunAll() override; + + virtual ~DfgPassManager() = default; +}; + +/* + * A pass manager that process a Node each time. + */ +class NodePassManager : public PassManager { + public: + NodePassManager() = default; + + void RunAll() override; + + virtual ~NodePassManager() = default; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/pass_manager_tester.cc b/paddle/fluid/inference/analysis/pass_manager_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..13423e4837e12a96e7a5dfc9ca3f59bf8b14746a --- /dev/null +++ b/paddle/fluid/inference/analysis/pass_manager_tester.cc @@ -0,0 +1,89 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h" +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include "paddle/fluid/inference/analysis/pass_manager.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class TestDfgPassManager final : public DfgPassManager { + public: + TestDfgPassManager() = default; + virtual ~TestDfgPassManager() = default; + // Short identifier. + std::string repr() const override { return "test-pass-manager"; } + // Long description. + std::string description() const override { return "test doc"; } +}; + +class TestNodePassManager final : public NodePassManager { + public: + virtual ~TestNodePassManager() = default; + + std::string repr() const override { return "test-node-pass-manager"; } + std::string description() const override { return "test doc"; } +}; + +class TestNodePass final : public NodePass { + public: + virtual ~TestNodePass() = default; + + bool Initialize(Argument* argument) override { return true; } + + void Run(Node* node) override { + LOG(INFO) << "- Processing node " << node->repr(); + } + + std::string repr() const override { return "test-node"; } + std::string description() const override { return "some doc"; } +}; + +TEST(PassManager, DFG_pass_manager) { + TestDfgPassManager manager; + DFG_GraphvizDrawPass::Config config("./", "dfg.dot"); + + manager.Register("fluid-to-flow-graph", new FluidToDataFlowGraphPass); + manager.Register("graphviz", new DFG_GraphvizDrawPass(config)); + manager.Register("dfg-to-fluid", new DataFlowGraphToFluidPass); + + Argument argument(FLAGS_inference_model_dir); + + ASSERT_TRUE(&argument); + ASSERT_TRUE(manager.Initialize(&argument)); + manager.RunAll(); +} + +TEST(PassManager, Node_pass_manager) { + Argument argument(FLAGS_inference_model_dir); + // Pre-process: initialize the DFG with the ProgramDesc first. + FluidToDataFlowGraphPass pass0; + pass0.Initialize(&argument); + pass0.Run(argument.main_dfg.get()); + + TestNodePassManager manager; + manager.Register("test-node-pass", new TestNodePass); + ASSERT_TRUE(manager.Initialize(&argument)); + manager.RunAll(); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc index 43ccac96c84e987ad1f494af3e314c810fc1ffe3..80809d4c43ca08298bad25cf614dcb4117d3f99a 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.cc +++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc @@ -76,7 +76,7 @@ void UnionFindCombine(const node_map_t &node_map, size_t a, size_t b) { std::vector> SubGraphSplitter::ExtractSubGraphs() { std::vector marked_nodes; - for (auto &node : GraphTraits(graph_).nodes()) { + for (auto &node : GraphTraits(graph_).nodes_in_TS()) { if (node.attr(kMarkerAttrName).Bool()) { marked_nodes.push_back(&node); } @@ -119,10 +119,12 @@ void SubGraphFuse::operator()() { ReplaceNodesWithSubGraphs(); } void SubGraphFuse::ReplaceNodesWithSubGraphs() { auto subgraphs = SubGraphSplitter(graph_, node_inside_subgraph_teller_)(); for (auto &subgraph : subgraphs) { + std::unordered_set subgraph_uniq(subgraph.begin(), subgraph.end()); // replace this sub-graph with the first node. Two steps: 1. Create a Block // Node that contains this subgraph 2. Mark the nodes inside the sub-graph // as deleted. 3. Replace the deleted node with the new Block Node. - auto *block_node = graph_->nodes.Create(Node::Type::kFunctionBlock); + auto *block_node = static_cast( + graph_->nodes.Create(Node::Type::kFunctionBlock)); auto io = ExtractInputAndOutputOfSubGraph(subgraph); block_node->inlinks = std::move(io.first); block_node->outlinks = std::move(io.second); @@ -130,21 +132,25 @@ void SubGraphFuse::ReplaceNodesWithSubGraphs() { // TODO(Superjomn) need a unified mechanism to treat deleted node in each // pass. node->SetDeleted(); + block_node->subgraph.push_back(node); } - std::unordered_map - delelte_node_map; // deleted node to BlockNode - for (auto *n : block_node->inlinks) { - n->inlinks.clear(); - } - for (auto *n : block_node->outlinks) { - n->outlinks.clear(); - } - for (auto *n : block_node->inlinks) { - n->outlinks.push_back(block_node); + // Change all the sub-graph's inputs and outputs corresponding inlink and + // outlink to this sub-graph node. + auto inlink_or_outlink_cleaner = [&](std::vector &nodes) { + for (auto *&n : nodes) { + if (subgraph_uniq.count(n)) { + n = block_node; + } + } + std::unordered_set uniq(nodes.begin(), nodes.end()); + nodes.assign(uniq.begin(), uniq.end()); + }; + for (auto *i : block_node->inlinks) { + inlink_or_outlink_cleaner(i->outlinks); } - for (auto *n : block_node->outlinks) { - n->inlinks.push_back(n); + for (auto *&o : block_node->outlinks) { + inlink_or_outlink_cleaner(o->inlinks); } } } diff --git a/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc b/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc index 0644c0db12e3daabba76dbaad33847f5624b157a..39cc433b40fad17f4f12359d4e907a250a88bd63 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc +++ b/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc @@ -19,22 +19,23 @@ namespace paddle { namespace inference { namespace analysis { -TEST_F(DFG_Tester, Split) { - auto desc = LoadProgramDesc(); +SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) { + if (node->type() != Node::Type::kFunction) return false; + const auto* func = static_cast(node); + if (func->func_type() == "elementwise_add" || func->func_type() == "relu" || + func->func_type() == "conv2d" || func->func_type() == "mul" || + func->func_type() == "sigmoid" || func->func_type() == "softmax") { + LOG(INFO) << "sub-graph marked " << node->repr(); + return true; + } + return false; +}; + +TEST(SubGraphSplitter, Split) { + auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); auto dfg = ProgramDescToDFG(desc); LOG(INFO) << "spliter\n" << dfg.DotString(); - SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) { - if (node->type() != Node::Type::kFunction) return false; - const auto* func = static_cast(node); - if (func->func_type() == "elementwise_add" || func->func_type() == "relu" || - func->func_type() == "conv2d" || func->func_type() == "mul" || - func->func_type() == "sigmoid" || func->func_type() == "softmax") { - LOG(INFO) << "sub-graph marked " << node->repr(); - return true; - } - return false; - }; ASSERT_GT(dfg.nodes.size(), 5UL); auto subgraphs = SubGraphSplitter(&dfg, teller)(); @@ -62,6 +63,28 @@ TEST_F(DFG_Tester, Split) { ASSERT_EQ(subgraphs.back().size(), 6UL); } +TEST(SubGraphSplitter, Fuse) { + auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); + auto dfg = ProgramDescToDFG(desc); + + size_t count0 = dfg.nodes.size(); + + SubGraphFuse fuse(&dfg, teller); + fuse(); + + int count1 = 0; + for (auto& node : dfg.nodes.nodes()) { + if (node->deleted()) { + LOG(INFO) << "deleted " << node->repr(); + } + count1 += node->deleted(); + } + + // At least one nodes should be deleted. + ASSERT_EQ(dfg.nodes.size(), count0 + 1); // added a new FunctionBlock + ASSERT_EQ(6, count1); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..f736e385c11add152dc9ab9485bf1de40f80b2f3 --- /dev/null +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc @@ -0,0 +1,80 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include + +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" +#include "paddle/fluid/inference/analysis/node_attr_flags.h" +#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +void TensorRTSubgraphNodeMarkPass::Run(DataFlowGraph *graph) { + for (auto &node : graph->nodes.nodes()) { + node->attr(ATTR_supported_by_tensorrt).Bool() = teller_(node.get()); + } +} + +class DfgDebuggerPass : public DFG_GraphvizDrawPass { + public: + explicit DfgDebuggerPass(const DFG_GraphvizDrawPass::Config &config) + : DFG_GraphvizDrawPass(config) {} + + std::string repr() const override { + return "tensorrt-subgraph-node-mark-debugger"; + } + + bool Finalize() override { return true; } + + protected: + std::string Draw(DataFlowGraph *graph) override { + Dot dot; + // Add nodes + for (size_t i = 0; i < graph->nodes.size(); i++) { + const Node &node = graph->nodes.Get(i); + if (config_.display_deleted_node || !node.deleted()) { + auto dot_attr = node.dot_attrs(); + if (node.attr(ATTR_supported_by_tensorrt).Bool()) { + dot_attr.assign( + {Dot::Attr{"color", "green"}, Dot::Attr{"style", "filled"}}); + } + dot.AddNode(node.repr(), dot_attr); + } + } + // Add edges + for (size_t i = 0; i < graph->nodes.size(); i++) { + const Node &node = graph->nodes.Get(i); + if (!config_.display_deleted_node && node.deleted()) continue; + for (auto &in : node.inlinks) { + if (!config_.display_deleted_node && in->deleted()) continue; + dot.AddEdge(in->repr(), node.repr(), {}); + } + } + return dot.Build(); + } +}; + +Pass *TensorRTSubgraphNodeMarkPass::CreateGraphvizDebugerPass() const { + DFG_GraphvizDrawPass::Config config( + FLAGS_inference_analysis_graphviz_log_root, "tensorrt_marked_node"); + return new DfgDebuggerPass(config); +} +bool TensorRTSubgraphNodeMarkPass::Finalize() { return true; } + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..c558a6ebbde371071c7330a14cc986bf764d1773 --- /dev/null +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h @@ -0,0 +1,60 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file defines TensorRTSubgraphNodeMarkPass which helps to mark the ops + * that supported by TensorRT engine. + */ + +#pragma once + +#include +#include "paddle/fluid/inference/analysis/pass.h" +#include "paddle/fluid/inference/analysis/subgraph_splitter.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Mark the operators that TensorRT engine supports. + */ +class TensorRTSubgraphNodeMarkPass : public DataFlowGraphPass { + public: + using teller_t = SubGraphSplitter::NodeInsideSubgraphTeller; + + explicit TensorRTSubgraphNodeMarkPass(const teller_t& teller) + : teller_(teller) {} + + bool Initialize(Argument* argument) override { return true; } + + // This class get a sub-graph as input and determine whether to transform this + // sub-graph into TensorRT. + void Run(DataFlowGraph* graph) override; + + std::string repr() const override { return "tensorrt-sub-subgraph-mark"; } + std::string description() const override { + return "tensorrt sub-graph mark pass"; + } + + Pass* CreateGraphvizDebugerPass() const override; + bool Finalize() override; + + private: + teller_t teller_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass_tester.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..c1d932878e559180af987594535959afdf475587 --- /dev/null +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass_tester.cc @@ -0,0 +1,50 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h" + +#include +#include "paddle/fluid/inference/analysis/node_attr_flags.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(TensorRTSubgraphNodeMarkPass, test) { + // init + FluidToDataFlowGraphPass pass; + Argument argument(FLAGS_inference_model_dir); + ASSERT_TRUE(pass.Initialize(&argument)); + pass.Run(argument.main_dfg.get()); + + TensorRTSubgraphNodeMarkPass::teller_t teller = [](const Node* node) { + return node->IsFunction() && + static_cast(node)->func_type() == "mul"; + }; + TensorRTSubgraphNodeMarkPass pass1(teller); + ASSERT_TRUE(pass1.Initialize(&argument)); + pass1.Run(argument.main_dfg.get()); + + int counter{0}; + for (auto& node : argument.main_dfg->nodes.nodes()) { + counter += node->attr(ATTR_supported_by_tensorrt).Bool(); + } + ASSERT_EQ(counter, 2); + LOG(INFO) << counter << " nodes marked"; +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..faf876de6d65d20cf7a084cd97392cfc8d791a42 --- /dev/null +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc @@ -0,0 +1,36 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h" +#include "paddle/fluid/inference/analysis/subgraph_splitter.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TensorRTSubGraphPass::TensorRTSubGraphPass( + const TensorRTSubGraphPass::NodeInsideSubgraphTeller &teller) + : node_inside_subgraph_teller_(teller) {} + +void TensorRTSubGraphPass::Run(DataFlowGraph *graph) { + SubGraphFuse(graph, node_inside_subgraph_teller_)(); + VLOG(4) << "debug info " + << graph->HumanReadableInfo(false /*show_values*/, + true /*show_functions*/); +} + +} // namespace analysis +} // namespace inference + +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..c6741a92095d33d261a4e1667c87a8ca02e51a9f --- /dev/null +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/inference/analysis/node.h" +#include "paddle/fluid/inference/analysis/pass.h" +#include "paddle/fluid/inference/analysis/subgraph_splitter.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Parse the graph and replace TensorRT supported nodes with SubGraphNode + */ +class TensorRTSubGraphPass : public DataFlowGraphPass { + public: + // Tell whether to transform a sub-graph into TensorRT. + using NodeInsideSubgraphTeller = SubGraphFuse::NodeInsideSubgraphTeller; + + explicit TensorRTSubGraphPass(const NodeInsideSubgraphTeller& teller); + + bool Initialize(Argument* argument) override { return true; } + + // This class get a sub-graph as input and determine whether to transform this + // sub-graph into TensorRT. + void Run(DataFlowGraph* graph) override; + + bool Finalize() override { return true; } + + std::string repr() const override { return "tensorrt-sub-graph"; } + std::string description() const override { return "tensorrt sub graph pass"; } + + private: + NodeInsideSubgraphTeller node_inside_subgraph_teller_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass_tester.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..67a5af83d89b771536ea11be51b35244ff5c09d6 --- /dev/null +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass_tester.cc @@ -0,0 +1,69 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h" + +#include +#include +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +DEFINE_string(dot_dir, "./", ""); + +TEST(TensorRTSubGraphPass, main) { + std::unordered_set teller_set( + {"elementwise_add", "mul", "sigmoid"}); + SubGraphSplitter::NodeInsideSubgraphTeller teller = [&](const Node* node) { + if (node->type() != Node::Type::kFunction) return false; + const auto* func = static_cast(node); + if (teller_set.count(func->func_type())) return true; + return false; + }; + + Argument argument(FLAGS_inference_model_dir); + + DFG_GraphvizDrawPass::Config config{FLAGS_dot_dir, "origin"}; + DFG_GraphvizDrawPass::Config config1{FLAGS_dot_dir, "fusion"}; + + DFG_GraphvizDrawPass dfg_pass(config); + DFG_GraphvizDrawPass dfg_pass1(config1); + FluidToDataFlowGraphPass pass0; + TensorRTSubGraphPass trt_pass(std::move(teller)); + + dfg_pass.Initialize(&argument); + dfg_pass1.Initialize(&argument); + pass0.Initialize(&argument); + trt_pass.Initialize(&argument); + + argument.main_dfg.reset(new DataFlowGraph); + pass0.Run(argument.main_dfg.get()); + dfg_pass.Run(argument.main_dfg.get()); + trt_pass.Run(argument.main_dfg.get()); + dfg_pass1.Run(argument.main_dfg.get()); + + // Check the TRT op's block desc + for (auto& node : argument.main_dfg->nodes.nodes()) { + if (node->IsFunctionBlock()) { + LOG(INFO) << "get function block"; + } + } +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/ut_helper.h b/paddle/fluid/inference/analysis/ut_helper.h index 722fa99a48a5f2b0e778904de0c35977d0ee3cc0..1073a6f686eaeeaaae2d93ab044149b7df518085 100644 --- a/paddle/fluid/inference/analysis/ut_helper.h +++ b/paddle/fluid/inference/analysis/ut_helper.h @@ -15,33 +15,31 @@ limitations under the License. */ #pragma once #include #include +#include #include #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" -#include "paddle/fluid/inference/analysis/ut_helper.h" -#include "paddle/fluid/inference/io.h" +#include "paddle/fluid/inference/analysis/helper.h" namespace paddle { namespace inference { + +// Read ProgramDesc from a __model__ file, defined in io.cc +extern void ReadBinaryFile(const std::string& filename, std::string* contents); + namespace analysis { DEFINE_string(inference_model_dir, "", "inference test model dir"); -static framework::proto::ProgramDesc LoadProgramDesc( - const std::string& model_dir = FLAGS_inference_model_dir) { - paddle::platform::CPUPlace place; - paddle::framework::Executor executor(place); - paddle::framework::Scope scope; - auto program = Load(&executor, &scope, model_dir); - return *program->Proto(); -} - static DataFlowGraph ProgramDescToDFG( const framework::proto::ProgramDesc& desc) { DataFlowGraph graph; FluidToDataFlowGraphPass pass; - pass.Initialize(desc); + Argument argument; + argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir)); + argument.origin_program_desc.reset(new framework::proto::ProgramDesc(desc)); + pass.Initialize(&argument); pass.Run(&graph); pass.Finalize(); return graph; @@ -49,9 +47,12 @@ static DataFlowGraph ProgramDescToDFG( class DFG_Tester : public ::testing::Test { protected: - void SetUp() override { desc = LoadProgramDesc(FLAGS_inference_model_dir); } + void SetUp() override { + auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); + argument.origin_program_desc.reset(new framework::proto::ProgramDesc(desc)); + } - framework::proto::ProgramDesc desc; + Argument argument; }; } // namespace analysis diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..83867e0a2c198f265cb36be3a71546796dfbec67 --- /dev/null +++ b/paddle/fluid/inference/api/CMakeLists.txt @@ -0,0 +1,78 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +if(APPLE) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") +endif(APPLE) + + +set(inference_deps paddle_inference_api paddle_fluid_api) + +if(WITH_GPU AND TENSORRT_FOUND) + set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine) +endif() + +function(inference_api_test TARGET_NAME) + if (WITH_TESTING) + set(options "") + set(oneValueArgs SRC) + set(multiValueArgs ARGS) + cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) + cc_test(${TARGET_NAME} + SRCS ${inference_test_SRC} + DEPS "${inference_deps}" + ARGS --dirname=${PYTHON_TESTS_DIR}/book/) + if(inference_test_ARGS) + set_tests_properties(${TARGET_NAME} + PROPERTIES DEPENDS "${inference_test_ARGS}") + endif() + endif(WITH_TESTING) +endfunction(inference_api_test) + +cc_library(paddle_inference_api SRCS api.cc api_impl.cc DEPS lod_tensor) + +cc_test(test_paddle_inference_api + SRCS api_tester.cc + DEPS paddle_inference_api) + +inference_api_test(test_api_impl SRC api_impl_tester.cc + ARGS test_word2vec test_image_classification) + +if(WITH_GPU AND TENSORRT_FOUND) +cc_library(paddle_inference_tensorrt_subgraph_engine + SRCS api_tensorrt_subgraph_engine.cc + DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter) + +inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec) +endif() + +if (WITH_ANAKIN) # only needed in CI + # compile the libinference_anakin_api.a and anakin.so. + nv_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber) + #nv_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin) + function(anakin_target target_name) + target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) + endfunction() + anakin_target(inference_anakin_api) + #anakin_target(inference_anakin_api_shared) + if (WITH_TESTING) + cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc + ARGS --model=${ANAKIN_SOURCE_DIR}/mobilenet_v2.anakin.bin + DEPS inference_anakin_api dynload_cuda SERIAL) + target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) + endif(WITH_TESTING) +endif() diff --git a/paddle/contrib/inference/README.md b/paddle/fluid/inference/api/README.md similarity index 100% rename from paddle/contrib/inference/README.md rename to paddle/fluid/inference/api/README.md diff --git a/paddle/fluid/inference/api/api.cc b/paddle/fluid/inference/api/api.cc new file mode 100644 index 0000000000000000000000000000000000000000..63c3f0d7b3f5c2b9246e2b041796caf5eb562826 --- /dev/null +++ b/paddle/fluid/inference/api/api.cc @@ -0,0 +1,95 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +namespace paddle { + +int PaddleDtypeSize(PaddleDType dtype) { + switch (dtype) { + case PaddleDType::FLOAT32: + return sizeof(float); + case PaddleDType::INT64: + return sizeof(int64_t); + default: + assert(false); + return -1; + } +} + +PaddleBuf::PaddleBuf(PaddleBuf&& other) + : data_(other.data_), + length_(other.length_), + memory_owned_(other.memory_owned_) { + other.memory_owned_ = false; + other.data_ = nullptr; + other.length_ = 0; +} + +PaddleBuf::PaddleBuf(const PaddleBuf& other) { *this = other; } + +PaddleBuf& PaddleBuf::operator=(const PaddleBuf& other) { + if (!other.memory_owned_) { + data_ = other.data_; + length_ = other.length_; + memory_owned_ = other.memory_owned_; + } else { + Resize(other.length()); + memcpy(data_, other.data(), other.length()); + length_ = other.length(); + memory_owned_ = true; + } + return *this; +} + +PaddleBuf& PaddleBuf::operator=(PaddleBuf&& other) { + // only the buffer with external memory can be copied + data_ = other.data_; + length_ = other.length_; + memory_owned_ = other.memory_owned_; + other.data_ = nullptr; + other.length_ = 0; + other.memory_owned_ = false; + return *this; +} + +void PaddleBuf::Resize(size_t length) { + // Only the owned memory can be reset, the external memory can't be changed. + if (length_ == length) return; + if (memory_owned_) { + Free(); + } + data_ = new char[length]; + length_ = length; + memory_owned_ = true; +} + +void PaddleBuf::Reset(void* data, size_t length) { + Free(); + memory_owned_ = false; + data_ = data; + length_ = length; +} + +void PaddleBuf::Free() { + if (memory_owned_ && data_) { + assert(length_ > 0); + delete[] static_cast(data_); + data_ = nullptr; + length_ = 0; + } +} + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_anakin_engine.cc b/paddle/fluid/inference/api/api_anakin_engine.cc new file mode 100644 index 0000000000000000000000000000000000000000..6b374ceefbc180a5c22abe591f12e1c3d89bc64a --- /dev/null +++ b/paddle/fluid/inference/api/api_anakin_engine.cc @@ -0,0 +1,167 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/api/api_anakin_engine.h" +#include +#include + +namespace paddle { + +template +PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor( + const AnakinConfig &config) { + CHECK(Init(config)); +} + +template +bool PaddleInferenceAnakinPredictor::Init(const AnakinConfig &config) { + if (!(graph_.load(config.model_file))) { + LOG(FATAL) << "fail to load graph from " << config.model_file; + return false; + } + auto inputs = graph_.get_ins(); + for (auto &input_str : inputs) { + graph_.ResetBatchSize(input_str, config.max_batch_size); + } + // optimization for graph + if (!(graph_.Optimize())) { + return false; + } + // construct executer + if (executor_p_ == nullptr) { + executor_p_ = new anakin::Net(graph_, true); + } + return true; +} + +template +bool PaddleInferenceAnakinPredictor::Run( + const std::vector &inputs, + std::vector *output_data, int batch_size) { + for (const auto &input : inputs) { + if (input.dtype != PaddleDType::FLOAT32) { + LOG(ERROR) << "Only support float type inputs. " << input.name + << "'s type is not float"; + return false; + } + auto d_tensor_in_p = executor_p_->get_in(input.name); + auto net_shape = d_tensor_in_p->valid_shape(); + if (net_shape.size() != input.shape.size()) { + LOG(ERROR) << " input " << input.name + << "'s shape size should be equal to that of net"; + return false; + } + int sum = 1; + for_each(input.shape.begin(), input.shape.end(), [&](int n) { sum *= n; }); + if (sum > net_shape.count()) { + graph_.Reshape(input.name, input.shape); + delete executor_p_; + executor_p_ = new anakin::Net(graph_, true); + d_tensor_in_p = executor_p_->get_in(input.name); + } + + anakin::saber::Shape tmp_shape; + for (auto s : input.shape) { + tmp_shape.push_back(s); + } + d_tensor_in_p->reshape(tmp_shape); + + float *d_data_p = d_tensor_in_p->mutable_data(); + if (cudaMemcpy(d_data_p, static_cast(input.data.data()), + d_tensor_in_p->valid_size() * sizeof(float), + cudaMemcpyHostToDevice) != 0) { + LOG(ERROR) << "copy data from CPU to GPU error"; + return false; + } + cudaStreamSynchronize(NULL); + } + cudaDeviceSynchronize(); + executor_p_->prediction(); + cudaDeviceSynchronize(); + + if (output_data->empty()) { + LOG(ERROR) << "At least one output should be set with tensors' names."; + return false; + } + for (auto &output : *output_data) { + auto *tensor = executor_p_->get_out(output.name); + output.shape = tensor->valid_shape(); + if (output.data.length() < tensor->valid_size() * sizeof(float)) { + output.data.Resize(tensor->valid_size() * sizeof(float)); + } + // Copy data from GPU -> CPU + if (cudaMemcpy(output.data.data(), tensor->mutable_data(), + tensor->valid_size() * sizeof(float), + cudaMemcpyDeviceToHost) != 0) { + LOG(ERROR) << "copy data from GPU to CPU error"; + return false; + } + cudaStreamSynchronize(NULL); + } + return true; +} + +template +anakin::Net + &PaddleInferenceAnakinPredictor::get_executer() { + return *executor_p_; +} + +// the cloned new Predictor of anakin share the same net weights from original +// Predictor +template +std::unique_ptr +PaddleInferenceAnakinPredictor::Clone() { + VLOG(3) << "Anakin Predictor::clone"; + std::unique_ptr cls( + new PaddleInferenceAnakinPredictor()); + // construct executer from other graph + auto anakin_predictor_p = + dynamic_cast *>(cls.get()); + if (!anakin_predictor_p) { + LOG(ERROR) << "fail to call Init"; + return nullptr; + } + anakin_predictor_p->get_executer().init(graph_); + + return std::move(cls); +} + +template class PaddleInferenceAnakinPredictor; +template class PaddleInferenceAnakinPredictor; + +// A factory to help create difference predictor. +template <> +std::unique_ptr CreatePaddlePredictor< + AnakinConfig, PaddleEngineKind::kAnakin>(const AnakinConfig &config) { + VLOG(3) << "Anakin Predictor create."; + if (config.target_type == AnakinConfig::NVGPU) { + VLOG(3) << "Anakin Predictor create on [ NVIDIA GPU ]."; + std::unique_ptr x( + new PaddleInferenceAnakinPredictor(config)); + return x; + } else if (config.target_type == AnakinConfig::X86) { + VLOG(3) << "Anakin Predictor create on [ Intel X86 ]."; + std::unique_ptr x( + new PaddleInferenceAnakinPredictor(config)); + return x; + } else { + VLOG(3) << "Anakin Predictor create on unknown platform."; + return nullptr; + } +}; + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_anakin_engine.h b/paddle/fluid/inference/api/api_anakin_engine.h new file mode 100644 index 0000000000000000000000000000000000000000..836badd9799228c6c294dcad5df73d039d36a1ff --- /dev/null +++ b/paddle/fluid/inference/api/api_anakin_engine.h @@ -0,0 +1,65 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains the implementation of inference API with Anakin engine + * embeded, this API can only support Anakin models. + */ + +#pragma once + +#include + +#include "framework/core/net/net.h" +#include "framework/graph/graph.h" +#include "paddle/fluid/inference/api/paddle_inference_api.h" +#include "saber/core/shape.h" +#include "saber/saber_types.h" + +namespace paddle { + +template +class PaddleInferenceAnakinPredictor : public PaddlePredictor { + public: + PaddleInferenceAnakinPredictor() {} + + explicit PaddleInferenceAnakinPredictor(const AnakinConfig& config); + + // NOTE Unlike the native engine, the buffers of anakin engine's output_data + // should be allocated first. + bool Run(const std::vector& inputs, + std::vector* output_data, + int batch_size = -1) override; + + std::unique_ptr Clone() override; + + anakin::Net& + get_executer(); + + ~PaddleInferenceAnakinPredictor() override { + delete executor_p_; + executor_p_ = nullptr; + }; + + private: + bool Init(const AnakinConfig& config); + + anakin::graph::Graph + graph_; + anakin::Net* + executor_p_{nullptr}; + AnakinConfig config_; +}; + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_anakin_engine_tester.cc b/paddle/fluid/inference/api/api_anakin_engine_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..62e820b68c79a47d963bb174663bfc8c4ac22de3 --- /dev/null +++ b/paddle/fluid/inference/api/api_anakin_engine_tester.cc @@ -0,0 +1,66 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "gflags/gflags.h" +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +DEFINE_string(model, "", "Directory of the inference model(mobile_v2)."); + +namespace paddle { + +AnakinConfig GetConfig() { + AnakinConfig config; + // using AnakinConfig::X86 if you need to use cpu to do inference + config.target_type = AnakinConfig::NVGPU; + config.model_file = FLAGS_model; + config.device = 0; + config.max_batch_size = 1; + return config; +} + +TEST(inference, anakin) { + AnakinConfig config = GetConfig(); + auto predictor = + CreatePaddlePredictor(config); + + float data[1 * 3 * 224 * 224] = {1.0f}; + PaddleTensor tensor; + tensor.name = "input_0"; + tensor.shape = std::vector({1, 3, 224, 224}); + tensor.data = PaddleBuf(data, sizeof(data)); + tensor.dtype = PaddleDType::FLOAT32; + + // For simplicity, we set all the slots with the same data. + std::vector paddle_tensor_feeds(1, tensor); + + PaddleTensor tensor_out; + tensor_out.name = "prob_out"; + tensor_out.shape = std::vector({}); + tensor_out.data = PaddleBuf(); + tensor_out.dtype = PaddleDType::FLOAT32; + + std::vector outputs(1, tensor_out); + + ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); + + float* data_o = static_cast(outputs[0].data.data()); + for (size_t j = 0; j < outputs[0].data.length(); ++j) { + LOG(INFO) << "output[" << j << "]: " << data_o[j]; + } +} + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..e31c637e969f7a86f4f185abb4f0f01d3303db75 --- /dev/null +++ b/paddle/fluid/inference/api/api_impl.cc @@ -0,0 +1,313 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include +#include + +#include "paddle/fluid/inference/api/api_impl.h" +#include "paddle/fluid/platform/profiler.h" + +DEFINE_bool(profile, false, "Turn on profiler for fluid"); + +namespace paddle { +namespace { + +// Timer for timer +class Timer { + public: + double start; + double startu; + void tic() { + struct timeval tp; + gettimeofday(&tp, NULL); + start = tp.tv_sec; + startu = tp.tv_usec; + } + double toc() { + struct timeval tp; + gettimeofday(&tp, NULL); + double used_time_ms = + (tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0; + return used_time_ms; + } +}; + +template +std::string num2str(T a) { + std::stringstream istr; + istr << a; + return istr.str(); +} +} // namespace + +bool NativePaddlePredictor::Init( + std::shared_ptr parent_scope) { + VLOG(3) << "Predictor::init()"; + + if (FLAGS_profile) { + LOG(WARNING) << "Profiler is actived, might affect the performance"; + LOG(INFO) << "You can turn off by set gflags '-profile false'"; + + auto tracking_device = config_.use_gpu ? platform::ProfilerState::kAll + : platform::ProfilerState::kCPU; + platform::EnableProfiler(tracking_device); + } + + if (config_.use_gpu) { + place_ = paddle::platform::CUDAPlace(config_.device); + } else { + place_ = paddle::platform::CPUPlace(); + } + if (parent_scope) { + scope_ = parent_scope; + sub_scope_ = &(parent_scope->NewScope()); + PADDLE_ENFORCE_NOT_NULL(sub_scope_, "create sub scope fail"); + } else { + paddle::framework::InitDevices(false); + scope_.reset(new paddle::framework::Scope()); + } + + executor_.reset(new paddle::framework::Executor(place_)); + + // Initialize the inference program + if (!config_.model_dir.empty()) { + // Parameters are saved in separate files sited in + // the specified `dirname`. + inference_program_ = paddle::inference::Load(executor_.get(), scope_.get(), + config_.model_dir); + } else if (!config_.prog_file.empty() && !config_.param_file.empty()) { + // All parameters are saved in a single file. + // The file names should be consistent with that used + // in Python API `fluid.io.save_inference_model`. + inference_program_ = paddle::inference::Load( + executor_.get(), scope_.get(), config_.prog_file, config_.param_file); + } else { + LOG(ERROR) << "fail to load inference model."; + return false; + } + + ctx_ = executor_->Prepare(*inference_program_, 0); + executor_->CreateVariables(*inference_program_, + sub_scope_ ? sub_scope_ : scope_.get(), 0); + + // Get the feed_target_names and fetch_target_names + feed_target_names_ = inference_program_->GetFeedTargetNames(); + fetch_target_names_ = inference_program_->GetFetchTargetNames(); + return true; +} + +NativePaddlePredictor::~NativePaddlePredictor() { + if (FLAGS_profile) { + platform::DisableProfiler(platform::EventSortingKey::kTotal, + "./profile.log"); + } + if (sub_scope_) { + scope_->DeleteScope(sub_scope_); + } +} + +bool NativePaddlePredictor::Run(const std::vector &inputs, + std::vector *output_data, + int batch_size) { + VLOG(3) << "Predictor::predict"; + Timer timer; + timer.tic(); + // set feed variable + std::map feed_targets; + std::vector feeds; + if (!SetFeed(inputs, &feeds)) { + LOG(ERROR) << "fail to set feed"; + return false; + } + for (size_t i = 0; i < feed_target_names_.size(); ++i) { + VLOG(4) << "setting " << i << "-th target"; + feed_targets[feed_target_names_[i]] = &feeds[i]; + } + // get fetch variable + std::map fetch_targets; + std::vector fetchs; + fetchs.resize(fetch_target_names_.size()); + for (size_t i = 0; i < fetch_target_names_.size(); ++i) { + fetch_targets[fetch_target_names_[i]] = &fetchs[i]; + } + // Run the inference program + // if share variables, we need not create variables + VLOG(4) << "Run prepared context"; + executor_->RunPreparedContext( + ctx_.get(), sub_scope_ != nullptr ? sub_scope_ : scope_.get(), + &feed_targets, &fetch_targets, + false, /* don't create local scope each time*/ + false /* don't create variable eatch time */); + VLOG(4) << "Finish prepared context"; + if (!GetFetch(fetchs, output_data)) { + LOG(ERROR) << "fail to get fetches"; + return false; + } + VLOG(3) << "predict cost: " << timer.toc() << "ms"; + return true; +} + +std::unique_ptr NativePaddlePredictor::Clone() { + VLOG(3) << "Predictor::clone"; + std::unique_ptr cls(new NativePaddlePredictor(config_)); + + if (!dynamic_cast(cls.get())->Init(scope_)) { + LOG(ERROR) << "fail to call Init"; + return nullptr; + } + // fix manylinux compile error. + return std::move(cls); +} + +bool NativePaddlePredictor::SetFeed(const std::vector &inputs, + std::vector *feeds) { + VLOG(3) << "Predictor::set_feed"; + if (inputs.size() != feed_target_names_.size()) { + LOG(ERROR) << "wrong feed input size."; + return false; + } + for (size_t i = 0; i < feed_target_names_.size(); ++i) { + framework::LoDTensor input; + framework::DDim ddim = framework::make_ddim(inputs[i].shape); + void *input_ptr; + if (inputs[i].dtype == PaddleDType::INT64) { + input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + } else if (inputs[i].dtype == PaddleDType::FLOAT32) { + input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + } else { + LOG(ERROR) << "unsupported feed type " << inputs[i].dtype; + return false; + } + + // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. + std::memcpy(static_cast(input_ptr), inputs[i].data.data(), + inputs[i].data.length()); + // TODO(Superjomn) Low performance, need optimization for heavy LoD copy. + framework::LoD lod; + for (auto &level : inputs[i].lod) { + lod.emplace_back(level); + } + input.set_lod(lod); + + feeds->push_back(input); + } + return true; +} + +bool NativePaddlePredictor::GetFetch( + const std::vector &fetchs, + std::vector *outputs) { + VLOG(3) << "Predictor::get_fetch"; + outputs->resize(fetchs.size()); + for (size_t i = 0; i < fetchs.size(); ++i) { + // TODO(panyx0718): Support fetch of other types. + if (fetchs[i].type() != typeid(float)) { + LOG(ERROR) << "only support fetching float now."; + return false; + } + std::vector shape; + auto dims_i = fetchs[i].dims(); + auto lod = fetchs[i].lod(); + const float *output_ptr = fetchs[i].data(); + // const int64_t* output_ptr = fetchs[i].data(); + auto num = fetchs[i].numel(); + std::vector data; + if (0 == lod.size()) { + std::copy(output_ptr, output_ptr + num, std::back_inserter(data)); + for (int j = 0; j < dims_i.size(); ++j) { + shape.push_back(dims_i[j]); + } + } else { + // for batch detection + // image[0] -> output[0] shape {145, 6} + // image[1] -> output[1] shape {176, 6} + // then, + // the batch output shape {321, 6} + // the lod {{0, 145, 321}} + // so we should append output[0] to {176, 6} + size_t max_dim = 0; + for (size_t j = 1; j < lod[0].size(); j++) { + max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]); + } + size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back(); + if (max_dim > 0) { + data.resize((lod[0].size() - 1) * max_dim * common_dim, 0); + } + for (size_t j = 1; j < lod[0].size(); j++) { + size_t start = lod[0][j - 1] * common_dim; + size_t end = lod[0][j] * common_dim; + if (end > start) { + std::copy(output_ptr + start, output_ptr + end, + data.begin() + (j - 1) * max_dim * common_dim); + } + } + shape.push_back(lod[0].size() - 1); + shape.push_back(max_dim); + for (int j = 1; j < dims_i.size(); ++j) { + shape.push_back(dims_i[j]); + } + } + + outputs->at(i).shape = shape; + auto &buffer = outputs->at(i).data; + if (buffer.empty() || buffer.length() < sizeof(float) * data.size()) { + buffer.Resize(sizeof(float) * data.size()); + } + std::memcpy(buffer.data(), data.data(), buffer.length()); + // copy LoD + for (const auto &level : fetchs[i].lod()) { + outputs->at(i).lod.emplace_back(level); + } + outputs->at(i).dtype = PaddleDType::FLOAT32; + // TODO(panyx0718): support other types? fill tensor name? avoid a copy. + } + return true; +} + +template <> +std::unique_ptr CreatePaddlePredictor< + NativeConfig, PaddleEngineKind::kNative>(const NativeConfig &config) { + VLOG(3) << "create NativePaddlePredictor"; + if (config.use_gpu) { + // 1. GPU memeroy + PADDLE_ENFORCE_GT( + config.fraction_of_gpu_memory, 0.f, + "fraction_of_gpu_memory in the config should be set to range (0., 1.]"); + PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device); + std::vector flags; + if (config.fraction_of_gpu_memory >= 0.0f || + config.fraction_of_gpu_memory <= 0.95f) { + flags.push_back("dummpy"); + std::string flag = "--fraction_of_gpu_memory_to_use=" + + num2str(config.fraction_of_gpu_memory); + flags.push_back(flag); + VLOG(3) << "set flag: " << flag; + framework::InitGflags(flags); + } + } + + std::unique_ptr predictor(new NativePaddlePredictor(config)); + if (!dynamic_cast(predictor.get())->Init(nullptr)) { + return nullptr; + } + return std::move(predictor); +} + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_impl.h b/paddle/fluid/inference/api/api_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..4f28c3cd34bade4189871210e6168c6c1c610c2c --- /dev/null +++ b/paddle/fluid/inference/api/api_impl.h @@ -0,0 +1,66 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include +#include +#include +#include + +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/inference/io.h" +#include "paddle/fluid/platform/init.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { + +class NativePaddlePredictor : public PaddlePredictor { + public: + explicit NativePaddlePredictor(const NativeConfig &config) + : config_(config) {} + + // will only create sub scope if have global scope + bool Init(std::shared_ptr parent_scope); + + bool Run(const std::vector &inputs, + std::vector *output_data, + int batch_size = -1) override; + + std::unique_ptr Clone() override; + + ~NativePaddlePredictor() override; + + protected: + bool SetFeed(const std::vector &input_datas, + std::vector *feeds); + bool GetFetch(const std::vector &fetchs, + std::vector *output_data); + + NativeConfig config_; + platform::Place place_; + std::unique_ptr executor_; + std::shared_ptr scope_; + std::unique_ptr ctx_; + std::unique_ptr inference_program_; + std::vector feed_target_names_; + std::vector fetch_target_names_; + // Do not use unique_ptr, use parent scope to delete + framework::Scope *sub_scope_{nullptr}; +}; + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_impl_tester.cc b/paddle/fluid/inference/api/api_impl_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..fc1364b80ac1ee2d304eb2fe429eae5f56967516 --- /dev/null +++ b/paddle/fluid/inference/api/api_impl_tester.cc @@ -0,0 +1,288 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include // NOLINT + +#include "gflags/gflags.h" +#include "paddle/fluid/inference/api/api_impl.h" +#include "paddle/fluid/inference/tests/test_helper.h" + +DEFINE_string(dirname, "", "Directory of the inference model."); + +namespace paddle { + +PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { + PaddleTensor pt; + + if (t->type() == typeid(int64_t)) { + pt.data.Reset(t->data(), t->numel() * sizeof(int64_t)); + pt.dtype = PaddleDType::INT64; + } else if (t->type() == typeid(float)) { + pt.data.Reset(t->data(), t->numel() * sizeof(float)); + pt.dtype = PaddleDType::FLOAT32; + } else { + LOG(FATAL) << "unsupported type."; + } + pt.shape = framework::vectorize2int(t->dims()); + return pt; +} + +NativeConfig GetConfig() { + NativeConfig config; + config.model_dir = FLAGS_dirname + "word2vec.inference.model"; + LOG(INFO) << "dirname " << config.model_dir; + config.fraction_of_gpu_memory = 0.15; +#ifdef PADDLE_WITH_CUDA + config.use_gpu = true; +#else + config.use_gpu = false; +#endif + config.device = 0; + return config; +} + +void MainWord2Vec(bool use_gpu) { + NativeConfig config = GetConfig(); + auto predictor = CreatePaddlePredictor(config); + config.use_gpu = use_gpu; + + framework::LoDTensor first_word, second_word, third_word, fourth_word; + framework::LoD lod{{0, 1}}; + int64_t dict_size = 2073; // The size of dictionary + + SetupLoDTensor(&first_word, lod, static_cast(0), dict_size - 1); + SetupLoDTensor(&second_word, lod, static_cast(0), dict_size - 1); + SetupLoDTensor(&third_word, lod, static_cast(0), dict_size - 1); + SetupLoDTensor(&fourth_word, lod, static_cast(0), dict_size - 1); + + std::vector paddle_tensor_feeds; + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&first_word)); + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&second_word)); + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&third_word)); + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&fourth_word)); + + std::vector outputs; + ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); + ASSERT_EQ(outputs.size(), 1UL); + size_t len = outputs[0].data.length(); + float* data = static_cast(outputs[0].data.data()); + for (size_t j = 0; j < len / sizeof(float); ++j) { + ASSERT_LT(data[j], 1.0); + ASSERT_GT(data[j], -1.0); + } + + std::vector cpu_feeds; + cpu_feeds.push_back(&first_word); + cpu_feeds.push_back(&second_word); + cpu_feeds.push_back(&third_word); + cpu_feeds.push_back(&fourth_word); + + framework::LoDTensor output1; + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); + + TestInference(config.model_dir, cpu_feeds, cpu_fetchs1); + + float* lod_data = output1.data(); + for (int i = 0; i < output1.numel(); ++i) { + EXPECT_LT(lod_data[i] - data[i], 1e-3); + EXPECT_GT(lod_data[i] - data[i], -1e-3); + } +} + +void MainImageClassification(bool use_gpu) { + int batch_size = 2; + bool repeat = false; + NativeConfig config = GetConfig(); + config.use_gpu = use_gpu; + config.model_dir = + FLAGS_dirname + "image_classification_resnet.inference.model"; + + const bool is_combined = false; + std::vector> feed_target_shapes = + GetFeedTargetShapes(config.model_dir, is_combined); + + framework::LoDTensor input; + // Use normilized image pixels as input data, + // which should be in the range [0.0, 1.0]. + feed_target_shapes[0][0] = batch_size; + framework::DDim input_dims = framework::make_ddim(feed_target_shapes[0]); + SetupTensor(&input, input_dims, static_cast(0), + static_cast(1)); + std::vector cpu_feeds; + cpu_feeds.push_back(&input); + + framework::LoDTensor output1; + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); + + TestInference( + config.model_dir, cpu_feeds, cpu_fetchs1, repeat, is_combined); + + auto predictor = CreatePaddlePredictor(config); + std::vector paddle_tensor_feeds; + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&input)); + + std::vector outputs; + ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); + ASSERT_EQ(outputs.size(), 1UL); + size_t len = outputs[0].data.length(); + float* data = static_cast(outputs[0].data.data()); + float* lod_data = output1.data(); + for (size_t j = 0; j < len / sizeof(float); ++j) { + EXPECT_NEAR(lod_data[j], data[j], 1e-3); + } +} + +void MainThreadsWord2Vec(bool use_gpu) { + NativeConfig config = GetConfig(); + config.use_gpu = use_gpu; + auto main_predictor = CreatePaddlePredictor(config); + + // prepare inputs data and reference results + constexpr int num_jobs = 3; + std::vector> jobs(num_jobs); + std::vector> paddle_tensor_feeds(num_jobs); + std::vector refs(num_jobs); + for (size_t i = 0; i < jobs.size(); ++i) { + // each job has 4 words + jobs[i].resize(4); + for (size_t j = 0; j < 4; ++j) { + framework::LoD lod{{0, 1}}; + int64_t dict_size = 2073; // The size of dictionary + SetupLoDTensor(&jobs[i][j], lod, static_cast(0), dict_size - 1); + paddle_tensor_feeds[i].push_back(LodTensorToPaddleTensor(&jobs[i][j])); + } + + // get reference result of each job + std::vector ref_feeds; + std::vector ref_fetches(1, &refs[i]); + for (auto& word : jobs[i]) { + ref_feeds.push_back(&word); + } + TestInference(config.model_dir, ref_feeds, ref_fetches); + } + + // create threads and each thread run 1 job + std::vector threads; + for (int tid = 0; tid < num_jobs; ++tid) { + threads.emplace_back([&, tid]() { + auto predictor = main_predictor->Clone(); + auto& local_inputs = paddle_tensor_feeds[tid]; + std::vector local_outputs; + ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); + + // check outputs range + ASSERT_EQ(local_outputs.size(), 1UL); + const size_t len = local_outputs[0].data.length(); + float* data = static_cast(local_outputs[0].data.data()); + for (size_t j = 0; j < len / sizeof(float); ++j) { + ASSERT_LT(data[j], 1.0); + ASSERT_GT(data[j], -1.0); + } + + // check outputs correctness + float* ref_data = refs[tid].data(); + EXPECT_EQ(refs[tid].numel(), static_cast(len / sizeof(float))); + for (int i = 0; i < refs[tid].numel(); ++i) { + EXPECT_NEAR(ref_data[i], data[i], 1e-3); + } + }); + } + for (int i = 0; i < num_jobs; ++i) { + threads[i].join(); + } +} + +void MainThreadsImageClassification(bool use_gpu) { + constexpr int num_jobs = 4; // each job run 1 batch + constexpr int batch_size = 1; + NativeConfig config = GetConfig(); + config.use_gpu = use_gpu; + config.model_dir = + FLAGS_dirname + "image_classification_resnet.inference.model"; + + auto main_predictor = CreatePaddlePredictor(config); + std::vector jobs(num_jobs); + std::vector> paddle_tensor_feeds(num_jobs); + std::vector refs(num_jobs); + for (size_t i = 0; i < jobs.size(); ++i) { + // prepare inputs + std::vector> feed_target_shapes = + GetFeedTargetShapes(config.model_dir, /*is_combined*/ false); + feed_target_shapes[0][0] = batch_size; + framework::DDim input_dims = framework::make_ddim(feed_target_shapes[0]); + SetupTensor(&jobs[i], input_dims, 0.f, 1.f); + paddle_tensor_feeds[i].push_back(LodTensorToPaddleTensor(&jobs[i])); + + // get reference result of each job + std::vector ref_feeds(1, &jobs[i]); + std::vector ref_fetches(1, &refs[i]); + TestInference(config.model_dir, ref_feeds, ref_fetches); + } + + // create threads and each thread run 1 job + std::vector threads; + for (int tid = 0; tid < num_jobs; ++tid) { + threads.emplace_back([&, tid]() { + auto predictor = main_predictor->Clone(); + auto& local_inputs = paddle_tensor_feeds[tid]; + std::vector local_outputs; + ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); + + // check outputs correctness + ASSERT_EQ(local_outputs.size(), 1UL); + const size_t len = local_outputs[0].data.length(); + float* data = static_cast(local_outputs[0].data.data()); + float* ref_data = refs[tid].data(); + EXPECT_EQ((size_t)refs[tid].numel(), len / sizeof(float)); + for (int i = 0; i < refs[tid].numel(); ++i) { + EXPECT_NEAR(ref_data[i], data[i], 1e-3); + } + }); + } + for (int i = 0; i < num_jobs; ++i) { + threads[i].join(); + } +} + +TEST(inference_api_native, word2vec_cpu) { MainWord2Vec(false /*use_gpu*/); } +TEST(inference_api_native, word2vec_cpu_threads) { + MainThreadsWord2Vec(false /*use_gpu*/); +} +TEST(inference_api_native, image_classification_cpu) { + MainThreadsImageClassification(false /*use_gpu*/); +} +TEST(inference_api_native, image_classification_cpu_threads) { + MainThreadsImageClassification(false /*use_gpu*/); +} + +#ifdef PADDLE_WITH_CUDA +TEST(inference_api_native, word2vec_gpu) { MainWord2Vec(true /*use_gpu*/); } +TEST(inference_api_native, word2vec_gpu_threads) { + MainThreadsWord2Vec(true /*use_gpu*/); +} +TEST(inference_api_native, image_classification_gpu) { + MainThreadsImageClassification(true /*use_gpu*/); +} +TEST(inference_api_native, image_classification_gpu_threads) { + MainThreadsImageClassification(true /*use_gpu*/); +} + +#endif + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc new file mode 100644 index 0000000000000000000000000000000000000000..45b5a7638b7dc6a54bbd905766fd5c284cb6aea1 --- /dev/null +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc @@ -0,0 +1,152 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/api/api_impl.h" +#include "paddle/fluid/inference/api/paddle_inference_api.h" +#include "paddle/fluid/inference/utils/singleton.h" +#include "paddle/fluid/operators/tensorrt_engine_op.h" + +namespace paddle { + +using inference::analysis::Argument; +using inference::Singleton; +using inference::analysis::Analyzer; +using framework::proto::ProgramDesc; + +class TensorRTSubgraphPredictor : public NativePaddlePredictor { + public: + explicit TensorRTSubgraphPredictor(const TensorRTConfig& config) + : NativePaddlePredictor(config), config_(config) {} + + bool Init(const std::shared_ptr& parent_scope) { + VLOG(3) << "Predictor::init()"; + + if (config_.use_gpu) { + place_ = paddle::platform::CUDAPlace(config_.device); + } else { + place_ = paddle::platform::CPUPlace(); + } + if (parent_scope) { + scope_ = parent_scope; + sub_scope_ = &(parent_scope->NewScope()); + } else { + paddle::framework::InitDevices(false); + scope_.reset(new paddle::framework::Scope()); + } + + executor_.reset(new paddle::framework::Executor(place_)); + + // Initialize the inference program + if (!config_.model_dir.empty()) { + // Parameters are saved in separate files sited in + // the specified `dirname`. + inference_program_ = paddle::inference::Load( + executor_.get(), scope_.get(), config_.model_dir); + } else if (!config_.prog_file.empty() && !config_.param_file.empty()) { + // All parameters are saved in a single file. + // The file names should be consistent with that used + // in Python API `fluid.io.save_inference_model`. + inference_program_ = paddle::inference::Load( + executor_.get(), scope_.get(), config_.prog_file, config_.param_file); + } else { + LOG(ERROR) << "fail to load inference model."; + return false; + } + + OptimizeInferenceProgram(); + ctx_ = executor_->Prepare(*inference_program_, 0); + + VLOG(5) << "to create variables"; + executor_->CreateVariables(*inference_program_, + sub_scope_ ? sub_scope_ : scope_.get(), 0); + + // Get the feed_target_names and fetch_target_names + feed_target_names_ = inference_program_->GetFeedTargetNames(); + fetch_target_names_ = inference_program_->GetFetchTargetNames(); + return true; + } + + bool Run(const std::vector& inputs, + std::vector* output_data, + int batch_size = -1) override { + PADDLE_ENFORCE_GT(batch_size, 0, + "TensorRT engine needs the argument batch_size set"); + FLAGS_tensorrt_engine_batch_size = batch_size; + return NativePaddlePredictor::Run(inputs, output_data, batch_size); + } + + void OptimizeInferenceProgram() { + // Analyze inference_program + Argument argument; + if (!config_.model_dir.empty()) { + argument.fluid_model_dir.reset(new std::string(config_.model_dir)); + } else { + PADDLE_ENFORCE( + !config_.param_file.empty(), + "Either model_dir or (param_file, prog_file) should be set."); + PADDLE_ENFORCE(!config_.prog_file.empty()); + argument.fluid_model_program_path.reset( + new std::string(config_.prog_file)); + argument.fluid_model_param_path.reset( + new std::string(config_.param_file)); + } + argument.origin_program_desc.reset( + new ProgramDesc(*inference_program_->Proto())); + Singleton::Global().Run(&argument); + CHECK(argument.transformed_program_desc); + VLOG(5) << "transformed program:\n" + << argument.transformed_program_desc->SerializeAsString(); + VLOG(5) << "to prepare executor"; + inference_program_.reset( + new framework::ProgramDesc(*argument.transformed_program_desc)); + } + + private: + TensorRTConfig config_; +}; + +template <> +std::unique_ptr +CreatePaddlePredictor( + const TensorRTConfig& config) { + VLOG(3) << "create TensorRTSubgraphPredictor"; + if (config.use_gpu) { + // 1. GPU memeroy + PADDLE_ENFORCE_GT( + config.fraction_of_gpu_memory, 0.f, + "fraction_of_gpu_memory in the config should be set to range (0., 1.]"); + PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device); + std::vector flags; + if (config.fraction_of_gpu_memory >= 0.0f || + config.fraction_of_gpu_memory <= 0.95f) { + flags.push_back("dummpy"); + std::string flag = "--fraction_of_gpu_memory_to_use=" + + std::to_string(config.fraction_of_gpu_memory); + flags.push_back(flag); + VLOG(3) << "set flag: " << flag; + framework::InitGflags(flags); + } + } + + std::unique_ptr predictor( + new TensorRTSubgraphPredictor(config)); + if (!dynamic_cast(predictor.get()) + ->Init(nullptr)) { + return nullptr; + } + return std::move(predictor); +} + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..fcbf9b89d608e7961e3ef81ac1c70e083dae1cc0 --- /dev/null +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc @@ -0,0 +1,92 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +namespace paddle { + +DEFINE_string(dirname, "", "Directory of the inference model."); + +void CompareTensorRTWithFluid(bool enable_tensorrt) { + FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = enable_tensorrt; + + //# 1. Create PaddlePredictor with a config. + NativeConfig config0; + config0.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config0.use_gpu = true; + config0.fraction_of_gpu_memory = 0.3; + config0.device = 0; + + TensorRTConfig config1; + config1.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config1.use_gpu = true; + config1.fraction_of_gpu_memory = 0.3; + config1.device = 0; + + auto predictor0 = + CreatePaddlePredictor(config0); + auto predictor1 = + CreatePaddlePredictor(config1); + + for (int batch_id = 0; batch_id < 1; batch_id++) { + //# 2. Prepare input. + std::vector data(20); + for (int i = 0; i < 20; i++) data[i] = i; + + PaddleTensor tensor; + tensor.shape = std::vector({10, 1}); + tensor.data = PaddleBuf(data.data(), data.size() * sizeof(int64_t)); + tensor.dtype = PaddleDType::INT64; + + // For simplicity, we set all the slots with the same data. + std::vector slots(4, tensor); + + //# 3. Run + std::vector outputs0; + std::vector outputs1; + CHECK(predictor0->Run(slots, &outputs0)); + CHECK(predictor1->Run(slots, &outputs1, 10)); + + //# 4. Get output. + ASSERT_EQ(outputs0.size(), 1UL); + ASSERT_EQ(outputs1.size(), 1UL); + + const size_t num_elements = outputs0.front().data.length() / sizeof(float); + const size_t num_elements1 = outputs1.front().data.length() / sizeof(float); + EXPECT_EQ(num_elements, num_elements1); + + auto *data0 = static_cast(outputs0.front().data.data()); + auto *data1 = static_cast(outputs1.front().data.data()); + + ASSERT_GT(num_elements, 0UL); + for (size_t i = 0; i < std::min(num_elements, num_elements1); i++) { + EXPECT_NEAR(data0[i], data1[i], 1e-3); + } + } +} + +TEST(paddle_inference_api_tensorrt_subgraph_engine, without_tensorrt) { + CompareTensorRTWithFluid(false); +} + +TEST(paddle_inference_api_tensorrt_subgraph_engine, with_tensorrt) { + CompareTensorRTWithFluid(true); +} + +} // namespace paddle diff --git a/paddle/fluid/inference/api/api_tester.cc b/paddle/fluid/inference/api/api_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..7a579610eefda24c911edd28b5f3a178aa10ab1e --- /dev/null +++ b/paddle/fluid/inference/api/api_tester.cc @@ -0,0 +1,64 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +namespace paddle { + +/* + * Do not use this, just a demo indicating how to customize a config for a + * specific predictor. + */ +struct DemoConfig : public PaddlePredictor::Config { + float other_config; +}; + +/* + * Do not use this, just a demo indicating how to customize a Predictor. + */ +class DemoPredictor : public PaddlePredictor { + public: + explicit DemoPredictor(const DemoConfig &config) { + LOG(INFO) << "I get other_config " << config.other_config; + } + bool Run(const std::vector &inputs, + std::vector *output_data, + int batch_size = 0) override { + LOG(INFO) << "Run"; + return false; + } + + std::unique_ptr Clone() override { return nullptr; } + + ~DemoPredictor() override {} +}; + +template <> +std::unique_ptr CreatePaddlePredictor( + const DemoConfig &config) { + std::unique_ptr x(new DemoPredictor(config)); + return x; +} + +TEST(paddle_inference_api, demo) { + DemoConfig config; + config.other_config = 1.7; + auto predictor = CreatePaddlePredictor(config); + std::vector outputs; + predictor->Run({}, &outputs); +} + +} // namespace paddle diff --git a/paddle/fluid/inference/api/demo_ci/.gitignore b/paddle/fluid/inference/api/demo_ci/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..1269488f7fb1f4b56a8c0e5eb48cecbfadfa9219 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/.gitignore @@ -0,0 +1 @@ +data diff --git a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..ba73a6eaa6fc885b6b56c2d6330394e2f9c384bf --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt @@ -0,0 +1,73 @@ +cmake_minimum_required(VERSION 3.0) + +project(cpp_inference_demo CXX C) + +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11") + +if(NOT DEFINED PADDLE_LIB) + message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib") +endif() +if(NOT DEFINED DEMO_NAME) + message(FATAL_ERROR "please set DEMO_NAME with -DDEMO_NAME=demo_name") +endif() + +option(WITH_MKL "Compile demo with MKL/OpenBlas support, default use MKL." ON) +option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." OFF) +option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON) + +if(WITH_GPU) + set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library") +endif() + +include_directories("${PADDLE_LIB}") +include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") +include_directories("${PADDLE_LIB}/third_party/install/glog/include") +include_directories("${PADDLE_LIB}/third_party/install/gflags/include") +include_directories("${PADDLE_LIB}/third_party/install/snappy/include") +include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") +include_directories("${PADDLE_LIB}/third_party/install/zlib/include") + +include_directories("${PADDLE_LIB}/third_party/boost") +include_directories("${PADDLE_LIB}/third_party/eigen3") + +link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") +link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") +link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") +link_directories("${PADDLE_LIB}/third_party/install/glog/lib") +link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") +link_directories("${PADDLE_LIB}/third_party/install/zlib/lib") + +add_executable(${DEMO_NAME} ${DEMO_NAME}.cc) + +if(WITH_MKL) + include_directories("${PADDLE_LIB}/third_party/install/mklml/include") + set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so + ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5.so) + set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn") + if(EXISTS ${MKLDNN_PATH}) + include_directories("${MKLDNN_PATH}/include") + set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0) + endif() +else() + set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a) +endif() + +# Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a +if(WITH_STATIC_LIB) + set(DEPS + ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a) +else() + set(DEPS + ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.so) +endif() +set(EXTERNAL_LIB "-lrt -ldl -lpthread") + +set(DEPS ${DEPS} + ${MATH_LIB} ${MKLDNN_LIB} + glog gflags protobuf snappystream snappy z + ${EXTERNAL_LIB}) +if(WITH_GPU) + set(DEPS ${DEPS} ${CUDA_LIB}/libcudart.so) +endif() + +target_link_libraries(${DEMO_NAME} ${DEPS}) diff --git a/paddle/fluid/inference/api/demo_ci/README.md b/paddle/fluid/inference/api/demo_ci/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7f013da7f30acd84ec484773f4ea716a08efa0ff --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/README.md @@ -0,0 +1,26 @@ +# Inference Demos + +There are several demos: + +- simple_on_word2vec: + - Follow the C++ codes is in `simple_on_word2vec.cc`. + - It is suitable for word2vec model. +- vis_demo: + - Follow the C++ codes is in `vis_demo.cc`. + - It is suitable for mobilenet, se_resnext50 and ocr three models. + - Input data format: + - Each line contains a single record + - Each record's format is + ``` + \t + ``` + +To build and execute the demos, simply run +``` +./run.sh $PADDLE_ROOT $TURN_ON_MKL $TEST_GPU_CPU +``` +- It will build and execute the demos in both static and shared library. +- `$PADDLE_ROOT`: paddle library path +- `$TURN_ON_MKL`: use MKL or Openblas +- `$TEST_GPU_CPU`: test both GPU/CPU mode or only CPU mode +- NOTE: for simple_on_word2vec, must run `ctest -R test_word2vec -R` to obtain word2vec model at first. diff --git a/paddle/fluid/inference/api/demo_ci/clean.sh b/paddle/fluid/inference/api/demo_ci/clean.sh new file mode 100755 index 0000000000000000000000000000000000000000..0d9f3d2aa237acaf3bd7adb031b1f2a73c555352 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/clean.sh @@ -0,0 +1,4 @@ +set -x +cd `dirname $0` +rm -rf build/ data/ +set +x diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh new file mode 100755 index 0000000000000000000000000000000000000000..3e829dd726b132844a45427b7b0b39eedf197496 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/run.sh @@ -0,0 +1,81 @@ +set -x +PADDLE_ROOT=$1 +TURN_ON_MKL=$2 # use MKL or Openblas +TEST_GPU_CPU=$3 # test both GPU/CPU mode or only CPU mode +if [ $2 == ON ]; then + # You can export yourself if move the install path + MKL_LIB=${PADDLE_ROOT}/build/fluid_install_dir/third_party/install/mklml/lib + export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${MKL_LIB} +fi +if [ $3 == ON ]; then + use_gpu_list='true false' +else + use_gpu_list='false' +fi + +# download vis_demo data +function download() { + dir_name=$1 + mkdir -p $dir_name + cd $dir_name + wget -q ${URL_ROOT}$dir_name.tar.gz + tar xzf *.tar.gz + cd .. +} +URL_ROOT=http://paddlemodels.bj.bcebos.com/inference-vis-demos%2F +mkdir -p data +cd data +vis_demo_list='se_resnext50 ocr mobilenet' +for vis_demo_name in $vis_demo_list; do + download $vis_demo_name +done +cd .. + +# compile and test the demo +mkdir -p build +cd build + +for WITH_STATIC_LIB in ON OFF; do + # -----simple_on_word2vec----- + rm -rf * + cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \ + -DWITH_MKL=$TURN_ON_MKL \ + -DDEMO_NAME=simple_on_word2vec \ + -DWITH_GPU=$TEST_GPU_CPU \ + -DWITH_STATIC_LIB=$WITH_STATIC_LIB + make -j + word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model' + if [ -d $word2vec_model ]; then + for use_gpu in $use_gpu_list; do + ./simple_on_word2vec \ + --dirname=$word2vec_model \ + --use_gpu=$use_gpu + if [ $? -ne 0 ]; then + echo "simple_on_word2vec demo runs fail." + exit 1 + fi + done + fi + # ---------vis_demo--------- + rm -rf * + cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \ + -DWITH_MKL=$TURN_ON_MKL \ + -DDEMO_NAME=vis_demo \ + -DWITH_GPU=$TEST_GPU_CPU \ + -DWITH_STATIC_LIB=$WITH_STATIC_LIB + make -j + for use_gpu in $use_gpu_list; do + for vis_demo_name in $vis_demo_list; do + ./vis_demo \ + --modeldir=../data/$vis_demo_name/model \ + --data=../data/$vis_demo_name/data.txt \ + --refer=../data/$vis_demo_name/result.txt \ + --use_gpu=$use_gpu + if [ $? -ne 0 ]; then + echo "vis demo $vis_demo_name runs fail." + exit 1 + fi + done + done +done +set +x diff --git a/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc new file mode 100644 index 0000000000000000000000000000000000000000..03ac79e9edf0d7ce6e167c3d34af5ba84bbc0e72 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc @@ -0,0 +1,142 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains a simple demo for how to take a model for inference. + */ + +#include +#include +#include +#include //NOLINT +#include "paddle/fluid/inference/paddle_inference_api.h" +#include "paddle/fluid/platform/enforce.h" + +DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_bool(use_gpu, false, "Whether use gpu."); + +namespace paddle { +namespace demo { + +void Main(bool use_gpu) { + //# 1. Create PaddlePredictor with a config. + NativeConfig config; + if (FLAGS_dirname.empty()) { + LOG(INFO) << "Usage: ./simple_on_word2vec --dirname=path/to/your/model"; + exit(1); + } + config.model_dir = FLAGS_dirname; + config.use_gpu = use_gpu; + config.fraction_of_gpu_memory = 0.15; + config.device = 0; + auto predictor = + CreatePaddlePredictor(config); + + for (int batch_id = 0; batch_id < 3; batch_id++) { + //# 2. Prepare input. + int64_t data[4] = {1, 2, 3, 4}; + + PaddleTensor tensor; + tensor.shape = std::vector({4, 1}); + tensor.data = PaddleBuf(data, sizeof(data)); + tensor.dtype = PaddleDType::INT64; + + // For simplicity, we set all the slots with the same data. + std::vector slots(4, tensor); + + //# 3. Run + std::vector outputs; + CHECK(predictor->Run(slots, &outputs)); + + //# 4. Get output. + PADDLE_ENFORCE(outputs.size(), 1UL); + // Check the output buffer size and result of each tid. + PADDLE_ENFORCE(outputs.front().data.length(), 33168UL); + float result[5] = {0.00129761, 0.00151112, 0.000423564, 0.00108815, + 0.000932706}; + const size_t num_elements = outputs.front().data.length() / sizeof(float); + // The outputs' buffers are in CPU memory. + for (size_t i = 0; i < std::min(5UL, num_elements); i++) { + PADDLE_ENFORCE(static_cast(outputs.front().data.data())[i], + result[i]); + } + } +} + +void MainThreads(int num_threads, bool use_gpu) { + // Multi-threads only support on CPU + // 0. Create PaddlePredictor with a config. + NativeConfig config; + config.model_dir = FLAGS_dirname; + config.use_gpu = use_gpu; + config.fraction_of_gpu_memory = 0.15; + config.device = 0; + auto main_predictor = + CreatePaddlePredictor(config); + + std::vector threads; + for (int tid = 0; tid < num_threads; ++tid) { + threads.emplace_back([&, tid]() { + // 1. clone a predictor which shares the same parameters + auto predictor = main_predictor->Clone(); + constexpr int num_batches = 3; + for (int batch_id = 0; batch_id < num_batches; ++batch_id) { + // 2. Dummy Input Data + int64_t data[4] = {1, 2, 3, 4}; + PaddleTensor tensor; + tensor.shape = std::vector({4, 1}); + tensor.data = PaddleBuf(data, sizeof(data)); + tensor.dtype = PaddleDType::INT64; + + std::vector inputs(4, tensor); + std::vector outputs; + // 3. Run + CHECK(predictor->Run(inputs, &outputs)); + + // 4. Get output. + PADDLE_ENFORCE(outputs.size(), 1UL); + // Check the output buffer size and result of each tid. + PADDLE_ENFORCE(outputs.front().data.length(), 33168UL); + float result[5] = {0.00129761, 0.00151112, 0.000423564, 0.00108815, + 0.000932706}; + const size_t num_elements = + outputs.front().data.length() / sizeof(float); + // The outputs' buffers are in CPU memory. + for (size_t i = 0; i < std::min(5UL, num_elements); i++) { + PADDLE_ENFORCE(static_cast(outputs.front().data.data())[i], + result[i]); + } + } + }); + } + for (int i = 0; i < num_threads; ++i) { + threads[i].join(); + } +} + +} // namespace demo +} // namespace paddle + +int main(int argc, char** argv) { + google::ParseCommandLineFlags(&argc, &argv, true); + paddle::demo::Main(false /* use_gpu*/); + paddle::demo::MainThreads(1, false /* use_gpu*/); + paddle::demo::MainThreads(4, false /* use_gpu*/); + if (FLAGS_use_gpu) { + paddle::demo::Main(true /*use_gpu*/); + paddle::demo::MainThreads(1, true /*use_gpu*/); + paddle::demo::MainThreads(4, true /*use_gpu*/); + } + return 0; +} diff --git a/paddle/fluid/inference/api/demo_ci/utils.h b/paddle/fluid/inference/api/demo_ci/utils.h new file mode 100644 index 0000000000000000000000000000000000000000..cb8990671162dff47228736e69617229528cc093 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/utils.h @@ -0,0 +1,67 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include +#include +#include "paddle/fluid/inference/paddle_inference_api.h" + +namespace paddle { +namespace demo { + +static void split(const std::string& str, char sep, + std::vector* pieces) { + pieces->clear(); + if (str.empty()) { + return; + } + size_t pos = 0; + size_t next = str.find(sep, pos); + while (next != std::string::npos) { + pieces->push_back(str.substr(pos, next - pos)); + pos = next + 1; + next = str.find(sep, pos); + } + if (!str.substr(pos).empty()) { + pieces->push_back(str.substr(pos)); + } +} + +/* + * Get a summary of a PaddleTensor content. + */ +static std::string SummaryTensor(const PaddleTensor& tensor) { + std::stringstream ss; + int num_elems = tensor.data.length() / PaddleDtypeSize(tensor.dtype); + + ss << "data[:10]\t"; + switch (tensor.dtype) { + case PaddleDType::INT64: { + for (int i = 0; i < std::min(num_elems, 10); i++) { + ss << static_cast(tensor.data.data())[i] << " "; + } + break; + } + case PaddleDType::FLOAT32: + for (int i = 0; i < std::min(num_elems, 10); i++) { + ss << static_cast(tensor.data.data())[i] << " "; + } + break; + } + return ss.str(); +} + +} // namespace demo +} // namespace paddle diff --git a/paddle/fluid/inference/api/demo_ci/vis_demo.cc b/paddle/fluid/inference/api/demo_ci/vis_demo.cc new file mode 100644 index 0000000000000000000000000000000000000000..3800d49b34738d5a272033d75cb415ae9ad1fb8f --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/vis_demo.cc @@ -0,0 +1,154 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains demo for mobilenet, se-resnext50 and ocr. + */ + +#include +#include // use glog instead of PADDLE_ENFORCE to avoid importing other paddle header files. +#include +#include +#include "paddle/fluid/inference/demo_ci/utils.h" +#include "paddle/fluid/platform/enforce.h" + +#ifdef PADDLE_WITH_CUDA +DECLARE_double(fraction_of_gpu_memory_to_use); +#endif +DEFINE_string(modeldir, "", "Directory of the inference model."); +DEFINE_string(refer, "", "path to reference result for comparison."); +DEFINE_string( + data, "", + "path of data; each line is a record, format is " + "'\t data; + std::vector shape; +}; + +void split(const std::string& str, char sep, std::vector* pieces); + +Record ProcessALine(const std::string& line) { + VLOG(3) << "process a line"; + std::vector columns; + split(line, '\t', &columns); + CHECK_EQ(columns.size(), 2UL) + << "data format error, should be \t"; + + Record record; + std::vector data_strs; + split(columns[0], ' ', &data_strs); + for (auto& d : data_strs) { + record.data.push_back(std::stof(d)); + } + + std::vector shape_strs; + split(columns[1], ' ', &shape_strs); + for (auto& s : shape_strs) { + record.shape.push_back(std::stoi(s)); + } + VLOG(3) << "data size " << record.data.size(); + VLOG(3) << "data shape size " << record.shape.size(); + return record; +} + +void CheckOutput(const std::string& referfile, const PaddleTensor& output) { + std::string line; + std::ifstream file(referfile); + std::getline(file, line); + auto refer = ProcessALine(line); + file.close(); + + size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); + VLOG(3) << "predictor output numel " << numel; + VLOG(3) << "reference output numel " << refer.data.size(); + PADDLE_ENFORCE_EQ(numel, refer.data.size()); + switch (output.dtype) { + case PaddleDType::INT64: { + for (size_t i = 0; i < numel; ++i) { + PADDLE_ENFORCE_EQ(static_cast(output.data.data())[i], + refer.data[i]); + } + break; + } + case PaddleDType::FLOAT32: + for (size_t i = 0; i < numel; ++i) { + PADDLE_ENFORCE_LT( + fabs(static_cast(output.data.data())[i] - refer.data[i]), + 1e-5); + } + break; + } +} + +/* + * Use the native fluid engine to inference the demo. + */ +void Main(bool use_gpu) { + NativeConfig config; + config.param_file = FLAGS_modeldir + "/__params__"; + config.prog_file = FLAGS_modeldir + "/__model__"; + config.use_gpu = use_gpu; + config.device = 0; + if (FLAGS_use_gpu) { + config.fraction_of_gpu_memory = 0.1; // set by yourself + } + + VLOG(3) << "init predictor"; + auto predictor = + CreatePaddlePredictor(config); + + VLOG(3) << "begin to process data"; + // Just a single batch of data. + std::string line; + std::ifstream file(FLAGS_data); + std::getline(file, line); + auto record = ProcessALine(line); + file.close(); + + // Inference. + PaddleTensor input; + input.shape = record.shape; + input.data = + PaddleBuf(record.data.data(), record.data.size() * sizeof(float)); + input.dtype = PaddleDType::FLOAT32; + + VLOG(3) << "run executor"; + std::vector output; + predictor->Run({input}, &output); + + VLOG(3) << "output.size " << output.size(); + auto& tensor = output.front(); + VLOG(3) << "output: " << SummaryTensor(tensor); + + // compare with reference result + CheckOutput(FLAGS_refer, tensor); +} + +} // namespace demo +} // namespace paddle + +int main(int argc, char** argv) { + google::ParseCommandLineFlags(&argc, &argv, true); + paddle::demo::Main(false /* use_gpu*/); + if (FLAGS_use_gpu) { + paddle::demo::Main(true /*use_gpu*/); + } + return 0; +} diff --git a/paddle/fluid/inference/api/high_level_api.md b/paddle/fluid/inference/api/high_level_api.md new file mode 100644 index 0000000000000000000000000000000000000000..8b8b6916d7e2b1a2f9fd09e9dfd2fe5a332461f5 --- /dev/null +++ b/paddle/fluid/inference/api/high_level_api.md @@ -0,0 +1,60 @@ +# Inference High-level APIs +This document describes the high-level inference APIs, one can use them to deploy a Paddle model for an application quickly. + +The APIs are described in `paddle_inference_api.h`, just one header file, and two libaries `libpaddle_fluid.so` and `libpaddle_fluid_api.so` are needed for a deployment. + +## PaddleTensor +We provide the `PaddleTensor` data structure to give a general tensor interface. + +The definition is + +```c++ +struct PaddleTensor { + std::string name; // variable name. + std::vector shape; + PaddleBuf data; // blob of data. + PaddleDType dtype; +}; +``` + +The data is stored in a continuous memory `PaddleBuf,` and a `PaddleDType` specifies tensor's data type. +The `name` field is used to specify the name of an input variable, +that is important when there are multiple inputs and need to distinguish which variable to set. + +## engine +The inference APIs has two different underlying engines + +- the native engine, which is consists of the native operators and framework, +- the Anakin engine, which has an Anakin library embedded. + +The native engine takes a native Paddle model as input, and supports any model that trained by Paddle, +the Anakin engine is faster for some model, +but it can only take the Anakin model as input(user need to transform the format first manually) and currently not all Paddle models are supported. + +```c++ +enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. +}; +``` + +## PaddlePredictor and how to create one +The main interface is `PaddlePredictor,` there are following methods + +- `bool Run(const std::vector& inputs, std::vector* output_data)` + - take inputs and output `output_data.` +- `Clone` to clone a predictor from an existing one, with model parameter shared. + +There is a factory method to help create a predictor, and the user takes the ownership of this object. + +```c++ +template +std::unique_ptr CreatePaddlePredictor(const ConfigT& config); +``` + +By specifying the engine kind and config, one can get a specific implementation. + +## Reference + +- [paddle_inference_api.h](./paddle_inference_api.h) +- [some demos](./demo_ci) diff --git a/paddle/fluid/inference/api/high_level_api_cn.md b/paddle/fluid/inference/api/high_level_api_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..2fb914592cbcb1b0c3f2ef33ff9cf4c295e427b6 --- /dev/null +++ b/paddle/fluid/inference/api/high_level_api_cn.md @@ -0,0 +1,87 @@ +# Paddle 预测 API + +为了更简单方便的预测部署，Fluid 提供了一套高层 API 用来隐藏底层不同的优化实现。 + +预测库包含: + +- 头文件 `paddle_inference_api.h` 定义了所有的接口 +- 库文件`libpaddle_fluid.so` 或 `libpaddle_fluid.a` +- 库文件 `libpaddle_inference_api.so` 或 `libpaddle_inference_api.a` + +下面是详细的一些 API 概念介绍 + +## PaddleTensor + +PaddleTensor 定义了预测最基本的输入输出的数据格式，其定义是 + +```c++ +struct PaddleTensor { + std::string name; // variable name. + std::vector shape; + PaddleBuf data; // blob of data. + PaddleDType dtype; +}; +``` + +- `name` 用于指定输入数据对应的 模型中variable 的名字 （暂时没有用，但会在后续支持任意 target 时启用） +- `shape` 表示一个 Tensor 的 shape +- `data` 数据以连续内存的方式存储在`PaddleBuf` 中，`PaddleBuf` 可以接收外面的数据或者独立`malloc`内存，详细可以参考头文件中相关定义。 +- `dtype` 表示 Tensor 的数据类型 + +## engine + +高层 API 底层有多种优化实现，我们称之为 engine，目前有三种 engine + +- 原生 engine，由 paddle 原生的 forward operator 组成，可以天然支持所有paddle 训练出的模型， +- Anakin engine，封装了 [Anakin](https://github.com/PaddlePaddle/Anakin) ，在某些模型上性能不错，但只能接受自带模型格式，无法支持所有 paddle 模型， +- TensorRT mixed engine，用子图的方式支持了 [TensorRT](https://developer.nvidia.com/tensorrt) ，支持所有paddle 模型，并自动切割部分计算子图到 TensorRT 上加速（WIP） + +其实现为 + +```c++ +enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. + kAutoMixedTensorRT // Automatically mixing TensorRT with the Fluid ops. +}; +``` + +## 预测部署过程 + +总体上分为以下步骤 + +1. 用合适的配置创建 `PaddlePredictor` +2. 创建输入用的 `PaddleTensor`，传入到 `PaddlePredictor` 中 +3. 获取输出的 `PaddleTensor` ，将结果取出 + +下面完整演示一个简单的模型，部分细节代码隐去 + +```c++ +#include "paddle_inference_api.h" + +// 创建一个 config，并修改相关设置 +paddle::NativeConfig config; +config.model_dir = "xxx"; +config.use_gpu = false; +// 创建一个原生的 PaddlePredictor +auto predictor = + paddle::CreatePaddlePredictor(config); +// 创建输入 tensor +int64_t data[4] = {1, 2, 3, 4}; +paddle::PaddleTensor tensor{.name = "", + .shape = std::vector({4, 1}), + .data = PaddleBuf(data, sizeof(data)), + .dtype = PaddleDType::INT64}; +// 创建输出 tensor，输出 tensor 的内存可以复用 +std::vector outputs; +// 执行预测 +CHECK(predictor->Run(slots, &outputs)); +// 获取 outputs ... +``` + +编译时，联编 `libpaddle_fluid.a/.so` 和 `libpaddle_inference_api.a/.so` 便可。 + +## 详细代码参考 + +- [inference demos](./demo_ci) +- [复杂单线程/多线程例子](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/api/test_api_impl.cc) diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h new file mode 100644 index 0000000000000000000000000000000000000000..794534467be066e91db2b4c204913ab2cf12dbfd --- /dev/null +++ b/paddle/fluid/inference/api/paddle_inference_api.h @@ -0,0 +1,155 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains the definition of a simple Inference API for Paddle. + * + * ATTENTION: It requires some C++11 features, for lower version C++ or C, we + * might release another API. + */ + +#pragma once + +#include +#include +#include +#include + +namespace paddle { + +enum PaddleDType { + FLOAT32, + INT64, +}; + +class PaddleBuf { + public: + PaddleBuf() = default; + PaddleBuf(PaddleBuf&& other); + // Copy only available when memory is managed externally. + explicit PaddleBuf(const PaddleBuf&); + PaddleBuf& operator=(const PaddleBuf&); + PaddleBuf& operator=(PaddleBuf&&); + // Do not own the memory. + PaddleBuf(void* data, size_t length) + : data_(data), length_(length), memory_owned_{false} {} + // Own memory. + explicit PaddleBuf(size_t length) + : data_(new char[length]), length_(length), memory_owned_(true) {} + // Resize to `length` bytes. + void Resize(size_t length); + // Reset to external memory. + void Reset(void* data, size_t length); + bool empty() const { return length_ == 0; } + void* data() const { return data_; } + size_t length() const { return length_; } + + ~PaddleBuf() { Free(); } + + private: + void Free(); + void* data_{nullptr}; // pointer to the data memory. + size_t length_{0}; // number of memory bytes. + bool memory_owned_{true}; +}; + +struct PaddleTensor { + PaddleTensor() = default; + std::string name; // variable name. + std::vector shape; + PaddleBuf data; // blob of data. + PaddleDType dtype; + std::vector> lod; // lod data +}; + +enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. + kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. + // TODO(Superjomn) support following engines latter. + // kTensorRT, // Use TensorRT for inference. + // kAutoMixedAnakin, // Automatically mix Fluid with Anakin. +}; + +/* + * A simple Inference API for Paddle. Currently this API can be used by + * non-sequence scenerios. + */ +class PaddlePredictor { + public: + struct Config; + PaddlePredictor() = default; + PaddlePredictor(const PaddlePredictor&) = delete; + PaddlePredictor& operator=(const PaddlePredictor&) = delete; + + // Predict an record. + // The caller should be responsible for allocating and releasing the memory of + // `inputs`. `inputs` should be available until Run returns. Caller should be + // responsible for the output tensor's buffer, either allocated or passed from + // outside. + virtual bool Run(const std::vector& inputs, + std::vector* output_data, + int batch_size = -1) = 0; + + // Clone a predictor that share the model weights, the Cloned predictor should + // be thread-safe. + virtual std::unique_ptr Clone() = 0; + + // Destroy the Predictor. + virtual ~PaddlePredictor() = default; + + // The common configs for all the predictors. + struct Config { + std::string model_dir; // path to the model directory. + }; +}; + +struct NativeConfig : public PaddlePredictor::Config { + // GPU related fields. + bool use_gpu{false}; + int device{0}; + float fraction_of_gpu_memory{-1.f}; // Negative to notify initialization. + + std::string prog_file; + std::string param_file; +}; + +// Configurations for Anakin engine. +struct AnakinConfig : public PaddlePredictor::Config { + enum TargetType { NVGPU = 0, X86 }; + int device; + std::string model_file; + int max_batch_size{-1}; + TargetType target_type; +}; + +struct TensorRTConfig : public NativeConfig { + // Determine whether a subgraph will be executed by TRT. + int min_subgraph_size{1}; +}; + +// A factory to help create different predictors. +// +// FOR EXTENSION DEVELOPER: +// Different predictors are designated by config type and engine kind. Similar +// configs can be merged, but there shouldn't be a huge config containing +// different fields for more than one kind of predictors. +// +// Similarly, each engine kind should map to a unique predictor implementation. +template +std::unique_ptr CreatePaddlePredictor(const ConfigT& config); + +int PaddleDtypeSize(PaddleDType dtype); + +} // namespace paddle diff --git a/paddle/fluid/inference/check_symbol.sh b/paddle/fluid/inference/check_symbol.sh new file mode 100755 index 0000000000000000000000000000000000000000..12b7b3e7e5982f193e48596b867953fc93841b61 --- /dev/null +++ b/paddle/fluid/inference/check_symbol.sh @@ -0,0 +1,12 @@ +#!/bin/bash + +lib=$1 +if [ $# -ne 1 ]; then echo "No input library"; exit -1 ; fi + +num_paddle_syms=$(nm -D ${lib} | grep paddle | wc -l) +num_google_syms=$(nm -D ${lib} | grep google | grep -v paddle | grep T | wc -l) + +if [ $num_paddle_syms -le 0 ]; then echo "Have no paddle symbols"; exit -1 ; fi +if [ $num_google_syms -ge 1 ]; then echo "Have some google symbols"; exit -1 ; fi + +exit 0 diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index 65db7c7b5008dcb301e741ec17c3623715e10bb8..181868977dd8f2568486ed0c4e1f260a69795896 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -20,16 +20,20 @@ limitations under the License. */ #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/feed_fetch_type.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/pybind/pybind.h" DEFINE_string(devices, "", "The devices to be used which is joined by comma."); DEFINE_bool(init_p2p, false, "Whether to init p2p."); +DEFINE_int32(math_num_threads, 1, + "Number of threads used to run math functions."); namespace paddle { namespace inference { void Init(const std::vector argv) { framework::InitGflags(argv); + platform::SetNumThreads(FLAGS_math_num_threads); // init devices std::vector devices; std::string token; diff --git a/paddle/fluid/inference/io.h b/paddle/fluid/inference/io.h index caf599b1a68783f155cd134c2a29e9ffa49a0895..01b50b3670cb9da2e0be232a61ea6129dd83aa20 100644 --- a/paddle/fluid/inference/io.h +++ b/paddle/fluid/inference/io.h @@ -18,9 +18,9 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/init.h" namespace paddle { namespace inference { diff --git a/paddle/fluid/inference/paddle_fluid.sym b/paddle/fluid/inference/paddle_fluid.sym new file mode 100644 index 0000000000000000000000000000000000000000..ef2a04d788aa86b7f6a61c4af479d70d1137f374 --- /dev/null +++ b/paddle/fluid/inference/paddle_fluid.sym @@ -0,0 +1 @@ +*paddle* diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 5ada1d631269209e912e2d4817382ea2c6c67353..6863b035d8cd9dfb21aed3947226a796778912a4 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,10 +1,26 @@ # Add TRT tests -nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine) -# This test is not stable -# See https://paddleci.ngrok.io/viewLog.html?tab=buildLog&buildTypeId=Paddle_PrCi2&buildId=36834&_focus=8828 -#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc -# DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine -# SERIAL) +nv_library(tensorrt_converter + SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc +activation_op.cc softmax_op.cc + DEPS tensorrt_engine operator scope framework_proto op_registry) + +nv_test(test_op_converter SRCS test_op_converter.cc DEPS + ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter) + nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) +nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) +nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine activation_op SERIAL) +nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine conv_op SERIAL) +nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine pool_op SERIAL) + +nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine elementwise_add_op SERIAL) + +nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index 6297051e5a30f1daa512d25d5aa3ab3b2f79f1d1..e1cace9cc1b06f036f52e82b7b86c99a02d50f50 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { @@ -21,10 +22,11 @@ namespace tensorrt { class ReluOpConverter : public OpConverter { public: ReluOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { // Here the two nullptr looks strange, that's because the // framework::OpDesc's constructor is strange. - framework::OpDesc op_desc(op, nullptr, nullptr); + framework::OpDesc op_desc(op, nullptr); LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " "type is Relu"; const nvinfer1::ITensor* input_tensor = @@ -32,12 +34,17 @@ class ReluOpConverter : public OpConverter { nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( engine_, Activation, *const_cast(input_tensor), nvinfer1::ActivationType::kRELU); - engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0)); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } } }; -REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); - } // namespace tensorrt } // namespace inference } // namespace paddle + +REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 209936c3bafb0d31546856dc36c1b48053a0634b..dba1d50b2d1c487ced8e6ca51f2d257641ad5fc7 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -20,15 +20,65 @@ namespace tensorrt { class Conv2dOpConverter : public OpConverter { public: - Conv2dOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { LOG(INFO) << "convert a fluid conv2d op to tensorrt conv layer without bias"; + + framework::OpDesc op_desc(op, nullptr); + PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Filter").size(), 1); // Y is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1); + + auto* X = engine_->GetITensor(op_desc.Input("Input").front()); + // Declare weights + auto* Y_v = scope.FindVar(op_desc.Input("Filter").front()); + PADDLE_ENFORCE_NOT_NULL(Y_v); + auto* Y_t = Y_v->GetMutable(); + auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + + PADDLE_ENFORCE_EQ(Y_t->dims().size(), 4UL); + const int n_output = Y_t->dims()[0]; + const int filter_h = Y_t->dims()[2]; + const int filter_w = Y_t->dims()[3]; + + const int groups = boost::get(op_desc.GetAttr("groups")); + const std::vector dilations = + boost::get>(op_desc.GetAttr("dilations")); + const std::vector strides = + boost::get>(op_desc.GetAttr("strides")); + const std::vector paddings = + boost::get>(op_desc.GetAttr("paddings")); + + nvinfer1::DimsHW nv_ksize(filter_h, filter_w); + nvinfer1::DimsHW nv_dilations(dilations[0], dilations[1]); + nvinfer1::DimsHW nv_strides(strides[0], strides[1]); + nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); + + TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, + static_cast(weight_data), + Y_t->memory_size() / sizeof(float)}; + + TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0}; + auto* layer = TRT_ENGINE_ADD_LAYER( + engine_, Convolution, *const_cast(X), n_output, + nv_ksize, weight.get(), bias.get()); + PADDLE_ENFORCE(layer != nullptr); + layer->setStride(nv_strides); + layer->setPadding(nv_paddings); + layer->setDilation(nv_dilations); + layer->setNbGroups(groups); + + auto output_name = op_desc.Output("Output").front(); + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { + engine_->DeclareOutput(output_name); + } } }; -REGISTER_TRT_OP_CONVERTER(conv2d, Conv2dOpConverter); - } // namespace tensorrt } // namespace inference } // namespace paddle + +REGISTER_TRT_OP_CONVERTER(conv2d, Conv2dOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..3744550f60a1696aedd8a3ecd24f1b21d22325b9 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc @@ -0,0 +1,210 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class ElementwiseWeightOpConverter : public OpConverter { + public: + ElementwiseWeightOpConverter() {} + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + // Here the two nullptr looks strange, that's because the + // framework::OpDesc's constructor is strange. + framework::OpDesc op_desc(op, nullptr); + LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer"; + + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); + + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + nvinfer1::Dims dims_x = X->getDimensions(); + PADDLE_ENFORCE(dims_x.nbDims >= 3); + + auto* Y_v = scope.FindVar(op_desc.Input("Y").front()); + PADDLE_ENFORCE_NOT_NULL(Y_v); + auto* Y_t = Y_v->GetMutable(); + auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + auto scale_mode = nvinfer1::ScaleMode::kELEMENTWISE; + + std::vector dims_y = framework::vectorize2int(Y_t->dims()); + if (static_cast(dims_y.size()) == dims_x.nbDims + 1) { + if (dims_y[0] == 1) dims_y.erase(dims_y.begin()); + } + + if (static_cast(dims_y.size()) == 1 && dims_y[0] == dims_x.d[0]) { + scale_mode = nvinfer1::ScaleMode::kCHANNEL; + } else if (static_cast(dims_y.size()) == dims_x.nbDims && + dims_y[0] == dims_x.d[0]) { + scale_mode = nvinfer1::ScaleMode::kELEMENTWISE; + for (int i = 1; i < dims_x.nbDims; i++) { + if (dims_y[i] != dims_x.d[i]) { + scale_mode = nvinfer1::ScaleMode::kCHANNEL; + break; + } + } + if (scale_mode == nvinfer1::ScaleMode::kCHANNEL) { + for (int i = 1; i < dims_x.nbDims; i++) { + if (dims_y[i] != 1) + PADDLE_THROW( + "TensorRT unsupported weight shape for Elementwise op!"); + } + } + } else { + PADDLE_THROW("TensorRT unsupported weight Shape for Elementwise op!"); + } + + TensorRTEngine::Weight shift_weights{nvinfer1::DataType::kFLOAT, + static_cast(weight_data), + Y_t->memory_size() / sizeof(float)}; + TensorRTEngine::Weight scale_weights{nvinfer1::DataType::kFLOAT, nullptr, + 0}; + TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr, + 0}; + + nvinfer1::IScaleLayer* layer = TRT_ENGINE_ADD_LAYER( + engine_, Scale, *const_cast(X), scale_mode, + shift_weights.get(), scale_weights.get(), power_weights.get()); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } + } +}; + +class ElementwiseTensorOpConverter : public OpConverter { + public: + ElementwiseTensorOpConverter() {} + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + // Here the two nullptr looks strange, that's because the + // framework::OpDesc's constructor is strange. + framework::OpDesc op_desc(op, nullptr); + LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer"; + + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); + + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + auto* Y = engine_->GetITensor(op_desc.Input("Y").front()); + nvinfer1::Dims dims_x = X->getDimensions(); + nvinfer1::Dims dims_y = Y->getDimensions(); + + // The two input tensor should have the same dims + PADDLE_ENFORCE(dims_x.nbDims >= 3); + if (dims_x.nbDims == dims_y.nbDims) { + for (int i = 0; i < dims_x.nbDims; i++) { + if (dims_x.d[i] != dims_y.d[i]) + PADDLE_THROW("TensorRT unsupported tensor shape for Elementwise op!"); + } + } else { + PADDLE_THROW("TensorRT unsupported tensor shape for Elementwise op!"); + } + + auto op_pair = ops.find(op_type_); + if (op_pair == ops.end()) { + PADDLE_THROW("Wrong elementwise op type!"); + } + nvinfer1::IElementWiseLayer* layer = TRT_ENGINE_ADD_LAYER( + engine_, ElementWise, *const_cast(X), + *const_cast(Y), op_pair->second); + + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } + } + + protected: + static const std::unordered_map + ops; + std::string op_type_; +}; + +const std::unordered_map + ElementwiseTensorOpConverter::ops = { + {"add", nvinfer1::ElementWiseOperation::kSUM}, + {"mul", nvinfer1::ElementWiseOperation::kPROD}, + {"sub", nvinfer1::ElementWiseOperation::kSUB}, + {"div", nvinfer1::ElementWiseOperation::kDIV}, + {"min", nvinfer1::ElementWiseOperation::kMIN}, + {"pow", nvinfer1::ElementWiseOperation::kPOW}, + {"max", nvinfer1::ElementWiseOperation::kMAX}, +}; + +class ElementwiseTensorAddOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorAddOpConverter() { op_type_ = "add"; } +}; + +class ElementwiseTensorMulOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorMulOpConverter() { op_type_ = "mul"; } +}; + +class ElementwiseTensorSubOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorSubOpConverter() { op_type_ = "sub"; } +}; + +class ElementwiseTensorDivOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorDivOpConverter() { op_type_ = "div"; } +}; + +class ElementwiseTensorMinOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorMinOpConverter() { op_type_ = "min"; } +}; + +class ElementwiseTensorMaxOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorMaxOpConverter() { op_type_ = "max"; } +}; + +class ElementwiseTensorPowOpConverter : public ElementwiseTensorOpConverter { + public: + ElementwiseTensorPowOpConverter() { op_type_ = "pow"; } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(elementwise_add_weight, ElementwiseWeightOpConverter); + +REGISTER_TRT_OP_CONVERTER(elementwise_add_tensor, + ElementwiseTensorAddOpConverter); +REGISTER_TRT_OP_CONVERTER(elementwise_sub_tensor, + ElementwiseTensorSubOpConverter); +REGISTER_TRT_OP_CONVERTER(elementwise_div_tensor, + ElementwiseTensorDivOpConverter); +REGISTER_TRT_OP_CONVERTER(elementwise_mul_tensor, + ElementwiseTensorMulOpConverter); +REGISTER_TRT_OP_CONVERTER(elementwise_max_tensor, + ElementwiseTensorMaxOpConverter); +REGISTER_TRT_OP_CONVERTER(elementwise_min_tensor, + ElementwiseTensorMinOpConverter); +REGISTER_TRT_OP_CONVERTER(elementwise_pow_tensor, + ElementwiseTensorPowOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..39fe1f609d7b94638506877fc301f19ef33ec8ac --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -0,0 +1,119 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +// Reorder the elements from istrides to ostrides, borrowed from TRT convert in +// tensorflow. +// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/tensorrt/convert/convert_nodes.cc#L318 +template +void Reorder2(nvinfer1::DimsHW shape, const T* idata, nvinfer1::DimsHW istrides, + T* odata, nvinfer1::DimsHW ostrides) { + for (int h = 0; h < shape.h(); ++h) { + for (int w = 0; w < shape.w(); ++w) { + odata[h * ostrides.h() + w * ostrides.w()] = + idata[h * istrides.h() + w * istrides.w()]; + } + } +} +// indata c * k +// Reorder the data layout from CK to KC. +void ReorderCKtoKC(TensorRTEngine::Weight& iweights, // NOLINT + TensorRTEngine::Weight* oweights) { + int c = iweights.dims[0]; + int k = iweights.dims[1]; + oweights->dims.assign({k, c}); + nvinfer1::DimsHW istrides = {1, k}; + nvinfer1::DimsHW ostrides = {c, 1}; + Reorder2({k, c}, static_cast(iweights.get().values), istrides, + static_cast(const_cast(oweights->get().values)), + ostrides); +} + +/* + * FC converter convert a MUL op in Fluid to a FC layer in TRT. + */ +class FcOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) << "convert a fluid fc op to tensorrt fc layer without bias"; + + framework::OpDesc op_desc(op, nullptr); + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); + + // Declare inputs + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + + // Declare weights + auto* Y_v = scope.FindVar(op_desc.Input("Y").front()); + PADDLE_ENFORCE_NOT_NULL(Y_v); + auto* Y_t = Y_v->GetMutable(); + // This may trigger a GPU->CPU copy, because TRT's weight can only be + // assigned from CPU memory, that can't be avoided. + auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + PADDLE_ENFORCE_EQ(Y_t->dims().size(), 2UL); // a matrix + size_t n_output = Y_t->dims()[1]; + + framework::LoDTensor tmp; + tmp.Resize(Y_t->dims()); + memcpy(tmp.mutable_data(platform::CPUPlace()), weight_data, + Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float)); + TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, + static_cast(weight_data), + Y_t->memory_size() / sizeof(float)}; + TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT, + static_cast(tmp.data()), + Y_t->memory_size() / sizeof(float)); + weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]}); + tmp_weight.dims = weight.dims; + + // The data layout of TRT FC layer's weight is different from fluid's FC, + // need to reorder the elements. + ReorderCKtoKC(weight, &tmp_weight); + + // Currently, the framework can only handle one fluid op -> one TRT layer, + // but fc fuses `mul` and `bias` (2 fluid ops), so here is a trick, just + // handle `mul`, leave `add` as another layer. + // DEBUG + TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0}; + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, + *const_cast(X), + n_output, tmp_weight.get(), bias.get()); + + auto output_name = op_desc.Output("Out").front(); + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(fc, FcOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc index ed09f54bde00d12aaec829ba90cc08ebfef57e92..514eb659a8da73b6e56b5d17148ec0cb2aeaa135 100644 --- a/paddle/fluid/inference/tensorrt/convert/mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -23,11 +23,11 @@ namespace tensorrt { */ class MulOpConverter : public OpConverter { public: - MulOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { - VLOG(4) << "convert a fluid mul op to tensorrt fc layer without bias"; + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; - framework::OpDesc op_desc(op, nullptr, nullptr); + framework::OpDesc op_desc(op, nullptr); // Declare inputs auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); auto* input2 = engine_->GetITensor(op_desc.Input("Y")[0]); @@ -36,12 +36,17 @@ class MulOpConverter : public OpConverter { engine_, MatrixMultiply, *const_cast(input1), false, *const_cast(input2), false); - engine_->DeclareOutput(layer, 0, op_desc.Output("Out")[0]); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } } }; -REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); - } // namespace tensorrt } // namespace inference } // namespace paddle + +REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index 1cd3ed9a00acead2599420f88499bd0d74c2974b..41faaf7212accaaec238062b1340e8da8fa6be33 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/inference/utils/singleton.h" @@ -31,27 +32,72 @@ namespace tensorrt { class OpConverter { public: OpConverter() {} - virtual void operator()(const framework::proto::OpDesc& op) {} - void Run(const framework::proto::OpDesc& op, TensorRTEngine* engine) { - std::string type = op.type(); - auto* it = Registry::Lookup(type); - PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); - it->SetEngine(engine); - (*it)(op); - } + // Converter logic for an op. + virtual void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, + bool test_mode = false) {} + + // Convert a single fluid operator and add the corresponding layer to TRT. + // test_mode: whether the instance executes in an unit test. + void ConvertOp(const framework::proto::OpDesc& op, + const std::unordered_set& parameters, + const framework::Scope& scope, TensorRTEngine* engine, + bool test_mode = false) { + framework::OpDesc op_desc(op, nullptr); + + OpConverter* it{nullptr}; + + if (op_desc.Type() == "mul") { + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1UL); + std::string Y = op_desc.Input("Y")[0]; + if (parameters.count(Y)) { + it = Registry::Lookup("fc"); + } + } + if (op_desc.Type().find("elementwise") != std::string::npos) { + static std::unordered_set add_tensor_op_set{ + "add", "mul", "sub", "div", "max", "min", "pow"}; + // TODO(xingzhaolong): all mul, sub, div + // static std::unordered_set add_weight_op_set {"add", "mul", + // "sub", "div"}; + static std::unordered_set add_weight_op_set{"add"}; + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1UL); + int op_type_len = op_desc.Type().size(); + std::string op_type = op_desc.Type().substr(op_type_len - 3, op_type_len); + std::string Y = op_desc.Input("Y")[0]; + if (parameters.count(Y)) { + PADDLE_ENFORCE(add_weight_op_set.count(op_type) > 0, + "Unsupported elementwise type" + op_type); + it = + Registry::Lookup("elementwise_" + op_type + "_weight"); + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", + op_desc.Type()); + } else { + PADDLE_ENFORCE(add_tensor_op_set.count(op_type) > 0, + "Unsupported elementwise type" + op_type); + it = + Registry::Lookup("elementwise_" + op_type + "_tensor"); + } + } - // convert fluid op to tensorrt layer - void ConvertOp(const framework::proto::OpDesc& op, TensorRTEngine* engine) { - OpConverter::Run(op, engine); + if (!it) { + it = Registry::Lookup(op_desc.Type()); + } + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", + op_desc.Type()); + it->SetEngine(engine); + (*it)(op, scope, test_mode); } - // convert fluid block to tensorrt network + // Convert a fluid block to tensorrt network, NOTE it just convert operators, + // the INetwork's inputs and outputs should specified in some other modules. void ConvertBlock(const framework::proto::BlockDesc& block, - TensorRTEngine* engine) { + const std::unordered_set& parameters, + const framework::Scope& scope, TensorRTEngine* engine) { for (int i = 0; i < block.ops_size(); i++) { const auto& op = block.ops(i); - OpConverter::Run(op, engine); + ConvertOp(op, parameters, scope, engine); } } @@ -62,6 +108,9 @@ class OpConverter { // TensorRT engine TensorRTEngine* engine_{nullptr}; + protected: + bool test_mode_; + private: // registered op converter map, whose key is the fluid op type, and value is // the pointer position of corresponding OpConverter class. @@ -70,14 +119,25 @@ class OpConverter { framework::Scope* scope_{nullptr}; }; -#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ - struct trt_##op_type__##_converter { \ - trt_##op_type__##_converter() { \ - Registry::Register(#op_type__); \ - } \ - }; \ - trt_##op_type__##_converter trt_##op_type__##_converter__; - } // namespace tensorrt } // namespace inference } // namespace paddle + +#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ + struct trt_##op_type__##_converter : public ::paddle::framework::Registrar { \ + trt_##op_type__##_converter() { \ + ::paddle::inference:: \ + Registry::Register< \ + ::paddle::inference::tensorrt::Converter__>(#op_type__); \ + } \ + }; \ + trt_##op_type__##_converter trt_##op_type__##_converter__; \ + int TouchConverterRegister_##op_type__() { \ + trt_##op_type__##_converter__.Touch(); \ + return 0; \ + } + +#define USE_TRT_CONVERTER(op_type__) \ + extern int TouchConverterRegister_##op_type__(); \ + static int use_op_converter_trt_##op_type__ __attribute__((unused)) = \ + TouchConverterRegister_##op_type__(); diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..11cad95361867476c6f775af778015da37f1cfb1 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc @@ -0,0 +1,80 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * Pool2dOp, IPoolingLayer in TRT. This Layer doesn't has weights. + */ +class Pool2dOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) + << "convert a fluid pool2d op to tensorrt pool2d layer without bias"; + framework::OpDesc op_desc(op, nullptr); + // Declare inputs + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); + auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); + + std::string pool_type = + boost::get(op_desc.GetAttr("pooling_type")); + std::vector ksize = + boost::get>(op_desc.GetAttr("ksize")); + std::vector strides = + boost::get>(op_desc.GetAttr("strides")); + std::vector paddings = + boost::get>(op_desc.GetAttr("paddings")); + + const nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); + const nvinfer1::DimsHW nv_strides(strides[0], strides[1]); + const nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); + + PADDLE_ENFORCE_EQ(input1->getDimensions().nbDims, 3UL); + + nvinfer1::PoolingType nv_pool_type = nvinfer1::PoolingType::kMAX; + if (pool_type == "max") { + nv_pool_type = nvinfer1::PoolingType::kMAX; + } else if (pool_type == "avg") { + nv_pool_type = nvinfer1::PoolingType::kAVERAGE; + } else { + PADDLE_THROW("TensorRT unsupported pooling type!"); + } + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, + *const_cast(input1), + nv_pool_type, nv_ksize); + PADDLE_ENFORCE_NOT_NULL(layer, "pool layer could not be created."); + layer->setStride(nv_strides); + layer->setPadding(nv_paddings); + + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(pool2d); +REGISTER_TRT_OP_CONVERTER(pool2d, Pool2dOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/softmax_op.cc b/paddle/fluid/inference/tensorrt/convert/softmax_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0064f90fd7944403c14d4d47616ea82f681ceb74 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/softmax_op.cc @@ -0,0 +1,49 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * SoftMaxOp, ISoftMaxLayer in TRT. This Layer doesn't has weights. + */ +class SoftMaxOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) + << "convert a fluid softmax op to tensorrt softmax layer without bias"; + framework::OpDesc op_desc(op, nullptr); + // Declare inputs + auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, SoftMax, + *const_cast(input1)); + + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(softmax); +REGISTER_TRT_OP_CONVERTER(softmax, SoftMaxOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc index 86ca2ca08eb14265e1bfe7abd5eb6af5c83b8a5c..e82762ea03ecd00bce7cfb83b130a3436ccbfed3 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -1,106 +1,47 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ #include -#include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" -#include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/place.h" - -USE_OP(relu); +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" namespace paddle { namespace inference { namespace tensorrt { -void Compare(const std::string op_type, float input, float expect) { +TEST(ReluOpConverter, main) { framework::Scope scope; - platform::CUDAPlace place; - platform::CUDADeviceContext ctx(place); - - // init fluid op and variable - auto x_var = scope.Var("X"); - auto x_tensor = x_var->GetMutable(); - x_tensor->Resize({1, 1}); - x_tensor->mutable_data(place); - std::vector init; - init.push_back(input); - framework::TensorFromVector(init, ctx, x_tensor); - - auto out_var = scope.Var("Out"); - auto out_tensor = out_var->GetMutable(); - out_tensor->Resize({1, 1}); - out_tensor->mutable_data(place); - - framework::OpDesc op_desc; - op_desc.SetType(op_type); - op_desc.SetInput("X", {"X"}); - op_desc.SetOutput("Out", {"Out"}); - - auto op = framework::OpRegistry::CreateOp(*op_desc.Proto()); - - // run fluid op - op->Run(scope, place); - // get fluid output - std::vector out1; - framework::TensorToVector(*out_tensor, ctx, &out1); - - // init tensorrt op - cudaStream_t stream; - ASSERT_EQ(0, cudaStreamCreate(&stream)); - TensorRTEngine* engine = new TensorRTEngine(1, 1 << 10, &stream); - engine->InitNetwork(); - engine->DeclareInput("X", nvinfer1::DataType::kFLOAT, - nvinfer1::DimsCHW{1, 1, 1}); - // convert op - OpConverter op_converter; - op_converter.ConvertOp(*op_desc.Proto(), engine); - - engine->DeclareOutput("Out"); - engine->FreezeNetwork(); - - // convert LoDTensor to ITensor - size_t size = x_tensor->memory_size(); - EngineIOConverter::ConvertInput(op_type, *x_tensor, - engine->buffer("X").buffer, size, &stream); - // run tensorrt Outp - engine->Execute(1); - // convert ITensor to LoDTensor - EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer, - out_tensor, size, &stream); - // get tensorrt output - std::vector out2; - framework::TensorToVector(*out_tensor, ctx, &out2); - - // compare - ASSERT_EQ(out1[0], out2[0]); - ASSERT_EQ(out1[0], expect); - - delete engine; - cudaStreamDestroy(stream); -} - -TEST(OpConverter, ConvertRelu) { - Compare("relu", 1, 1); // relu(1) = 1 - Compare("relu", -5, 0); // relu(-5) = 0 + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000); + validator.DeclInputVar("relu-X", nvinfer1::Dims2(10, 6)); + validator.DeclOutputVar("relu-Out", nvinfer1::Dims2(10, 6)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("relu"); + desc.SetInput("X", {"relu-X"}); + desc.SetOutput("Out", {"relu-Out"}); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(5); } } // namespace tensorrt } // namespace inference } // namespace paddle -USE_OP(activation); +USE_OP(relu); diff --git a/paddle/fluid/inference/tensorrt/convert/test_conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/test_conv2d_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f8711c6b60d74639529624c25429bc245de46479 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_conv2d_op.cc @@ -0,0 +1,57 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(conv2d_op, test) { + std::unordered_set parameters({"conv2d-Y"}); + framework::Scope scope; + TRTConvertValidation validator(5, parameters, scope, 1 << 15); + + validator.DeclInputVar("conv2d-X", nvinfer1::Dims3(2, 5, 5)); + validator.DeclParamVar("conv2d-Y", nvinfer1::Dims4(3, 2, 3, 3)); + validator.DeclOutputVar("conv2d-Out", nvinfer1::Dims3(3, 5, 5)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("conv2d"); + desc.SetInput("Input", {"conv2d-X"}); + desc.SetInput("Filter", {"conv2d-Y"}); + desc.SetOutput("Output", {"conv2d-Out"}); + + const std::vector strides({1, 1}); + const std::vector paddings({1, 1}); + const std::vector dilations({1, 1}); + const int groups = 1; + + desc.SetAttr("strides", strides); + desc.SetAttr("paddings", paddings); + desc.SetAttr("dilations", dilations); + desc.SetAttr("groups", groups); + + validator.SetOp(*desc.Proto()); + + validator.Execute(3); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle +USE_OP(conv2d); diff --git a/paddle/fluid/inference/tensorrt/convert/test_elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/test_elementwise_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..7537d02a35b66a41c158cd8eb1b1e5d4107e7d84 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_elementwise_op.cc @@ -0,0 +1,73 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(elementwise_op, add_weight_test) { + std::unordered_set parameters({"elementwise_add-Y"}); + framework::Scope scope; + TRTConvertValidation validator(10, parameters, scope, 1 << 15); + validator.DeclInputVar("elementwise_add-X", nvinfer1::DimsCHW(10, 3, 3)); + validator.DeclParamVar("elementwise_add-Y", nvinfer1::Dims3(10, 1, 1)); + // validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2)); + validator.DeclOutputVar("elementwise_add-Out", nvinfer1::DimsCHW(10, 3, 3)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("elementwise_add"); + desc.SetInput("X", {"elementwise_add-X"}); + desc.SetInput("Y", {"elementwise_add-Y"}); + desc.SetOutput("Out", {"elementwise_add-Out"}); + + int axis = 1; + desc.SetAttr("axis", axis); + + validator.SetOp(*desc.Proto()); + + validator.Execute(8); +} + +TEST(elementwise_op, add_tensor_test) { + std::unordered_set parameters; + framework::Scope scope; + TRTConvertValidation validator(8, parameters, scope, 1 << 15); + validator.DeclInputVar("elementwise_add-X", nvinfer1::DimsCHW(10, 3, 3)); + validator.DeclInputVar("elementwise_add-Y", nvinfer1::Dims3(10, 3, 3)); + // validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2)); + validator.DeclOutputVar("elementwise_add-Out", nvinfer1::DimsCHW(10, 3, 3)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("elementwise_add"); + desc.SetInput("X", {"elementwise_add-X"}); + desc.SetInput("Y", {"elementwise_add-Y"}); + desc.SetOutput("Out", {"elementwise_add-Out"}); + + // the defalut axis of elementwise op is -1 + + validator.SetOp(*desc.Proto()); + + validator.Execute(8); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle +USE_OP(elementwise_add); diff --git a/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..1ae2668e733aad23241c63b9985e708396d0b1bc --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc @@ -0,0 +1,46 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(fc_op, test) { + std::unordered_set parameters({"mul-Y"}); + framework::Scope scope; + TRTConvertValidation validator(10, parameters, scope, 1000); + validator.DeclInputVar("mul-X", nvinfer1::Dims3(10, 1, 1)); + validator.DeclParamVar("mul-Y", nvinfer1::Dims2(10, 2)); + validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(1, 2)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("mul"); + desc.SetInput("X", {"mul-X"}); + desc.SetInput("Y", {"mul-Y"}); + desc.SetOutput("Out", {"mul-Out"}); + + validator.SetOp(*desc.Proto()); + + validator.Execute(10); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle +USE_OP(mul); diff --git a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc index d8b61d5f08ffd071c112b4677fcb6f6f50784bbc..3d34cd7d5d0deca4d83a3f5b5ed0fb396c6acd56 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc @@ -21,7 +21,9 @@ namespace inference { namespace tensorrt { TEST(MulOpConverter, main) { - TRTConvertValidation validator(10, 1000); + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000, false); validator.DeclInputVar("mul-X", nvinfer1::Dims2(10, 6)); validator.DeclInputVar("mul-Y", nvinfer1::Dims2(6, 10)); validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(10, 10)); @@ -37,7 +39,7 @@ TEST(MulOpConverter, main) { validator.SetOp(*desc.Proto()); LOG(INFO) << "execute"; - validator.Execute(10); + validator.Execute(2); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index 9ae7de9cbfa656fbcbb48557bd4b548115897c6d..d6651a5b244ba31a01220e6299cb2016ae61fe64 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -12,9 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + #include #include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { namespace inference { @@ -24,12 +25,46 @@ TEST(OpConverter, ConvertBlock) { framework::ProgramDesc prog; auto* block = prog.MutableBlock(0); auto* conv2d_op = block->AppendOp(); + + // init trt engine + cudaStream_t stream_; + std::unique_ptr engine_; + engine_.reset(new TensorRTEngine(5, 1 << 15, &stream_)); + engine_->InitNetwork(); + PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0); + + engine_->DeclareInput("conv2d-X", nvinfer1::DataType::kFLOAT, + nvinfer1::Dims3(2, 5, 5)); + conv2d_op->SetType("conv2d"); + conv2d_op->SetInput("Input", {"conv2d-X"}); + conv2d_op->SetInput("Filter", {"conv2d-Y"}); + conv2d_op->SetOutput("Output", {"conv2d-Out"}); + + const std::vector strides({1, 1}); + const std::vector paddings({1, 1}); + const std::vector dilations({1, 1}); + const int groups = 1; + + conv2d_op->SetAttr("strides", strides); + conv2d_op->SetAttr("paddings", paddings); + conv2d_op->SetAttr("dilations", dilations); + conv2d_op->SetAttr("groups", groups); + + // init scope + framework::Scope scope; + std::vector dim_vec = {3, 2, 3, 3}; + auto* x = scope.Var("conv2d-Y"); + auto* x_tensor = x->GetMutable(); + x_tensor->Resize(framework::make_ddim(dim_vec)); OpConverter converter; - converter.ConvertBlock(*block->Proto(), nullptr /*TensorRTEngine*/); + converter.ConvertBlock(*block->Proto(), {"conv2d-Y"}, scope, + engine_.get() /*TensorRTEngine*/); } } // namespace tensorrt } // namespace inference } // namespace paddle + +USE_TRT_CONVERTER(conv2d) diff --git a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c5dddbc8cd37b9fb1ba39382af2da5ad045f3af2 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc @@ -0,0 +1,60 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ +#include +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(Pool2dOpConverter, main) { + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(5, parameters, scope, 1 << 15); + + // The ITensor's Dims should not contain the batch size. + // So, the ITensor's Dims of input and output should be C * H * W. + validator.DeclInputVar("pool2d-X", nvinfer1::Dims3(3, 4, 4)); + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("pool2d"); + desc.SetInput("X", {"pool2d-X"}); + desc.SetOutput("Out", {"pool2d-Out"}); + + std::vector ksize({2, 2}); + std::vector strides({2, 2}); + std::vector paddings({0, 0}); + std::string pooling_t = "max"; + + desc.SetAttr("pooling_type", pooling_t); + desc.SetAttr("ksize", ksize); + desc.SetAttr("strides", strides); + desc.SetAttr("paddings", paddings); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(3); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(pool2d); diff --git a/paddle/fluid/inference/tensorrt/convert/test_softmax_op.cc b/paddle/fluid/inference/tensorrt/convert/test_softmax_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..503ce71f7fb4377bb4304569b7484fb25abdb284 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_softmax_op.cc @@ -0,0 +1,49 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(SoftMaxOpConverter, main) { + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(8, parameters, scope, 1000); + + std::vector tensor_shape{8, 10}; + validator.DeclInputVar("softmax-X", tensor_shape, + nvinfer1::DimsCHW(10, 1, 1)); + validator.DeclOutputVar("softmax-Out", nvinfer1::DimsCHW(10, 1, 1)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("softmax"); + desc.SetInput("X", {"softmax-X"}); + desc.SetOutput("Out", {"softmax-Out"}); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(3); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(softmax); diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h index e46c577cdae145c0d4ceb6bfa307f03d313514ce..4265f33f28fe36b1745baf4761c3c85e3a281d6b 100644 --- a/paddle/fluid/inference/tensorrt/convert/ut_helper.h +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -27,6 +27,7 @@ limitations under the License. */ #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace inference { @@ -38,7 +39,7 @@ namespace tensorrt { float random(float low, float high) { static std::random_device rd; static std::mt19937 mt(rd()); - std::uniform_real_distribution dist(1.0, 10.0); + std::uniform_real_distribution dist(low, high); return dist(mt); } @@ -48,6 +49,7 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place, size_t num_elements = analysis::AccuDims(dims, dims.size()); PADDLE_ENFORCE_GT(num_elements, 0); auto* data = tensor->mutable_data(place); + for (size_t i = 0; i < num_elements; i++) { *(data + i) = random(0., 1.); } @@ -61,58 +63,86 @@ class TRTConvertValidation { public: TRTConvertValidation() = delete; - explicit TRTConvertValidation(int batch_size, int workspace_size = 1024) { + TRTConvertValidation(int max_batch_size, + const std::unordered_set& parameters, + framework::Scope& scope, // NOLINT + int workspace_size = 1 << 10, bool if_add_batch = true) + : parameters_(parameters), + scope_(scope), + if_add_batch_(if_add_batch), + max_batch_size_(max_batch_size) { // create engine. - engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_)); + engine_.reset(new TensorRTEngine(max_batch_size, workspace_size, &stream_)); engine_->InitNetwork(); PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0); } // Declare a Variable as input with random initialization. + void DeclInputVar(const std::string& name, const std::vector tensor_dims, + const nvinfer1::Dims& trt_dims) { + DeclVar(name, tensor_dims); + engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, trt_dims); + } + void DeclInputVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims); // Declare TRT inputs. engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, dims); } + // Declare a parameter varaible in the scope. + void DeclParamVar(const std::string& name, const nvinfer1::Dims& dims) { + DeclVar(name, dims, true); + } + void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims); } - void DeclVar(const std::string& name, const nvinfer1::Dims& dims) { + void DeclVar(const std::string& name, const std::vector dim_vec) { platform::CPUPlace place; platform::CPUDeviceContext ctx(place); - // Init Fluid tensor. - std::vector dim_vec(dims.nbDims); - for (int i = 0; i < dims.nbDims; i++) { - dim_vec[i] = dims.d[i]; - } auto* x = scope_.Var(name); auto* x_tensor = x->GetMutable(); x_tensor->Resize(framework::make_ddim(dim_vec)); RandomizeTensor(x_tensor, place, ctx); } + // Declare a variable in a fluid Scope. + void DeclVar(const std::string& name, const nvinfer1::Dims& dims, + bool is_param = false) { + // Init Fluid tensor. + std::vector dim_vec(dims.d, dims.d + dims.nbDims); + // There is no batchsize in ITensor's shape, but We should add it to + // tensor's shape of fluid. If the variable is not parameter and the + // if_add_batch_ flag is true, add the max batchsize to dim_vec. + if (is_param != true && if_add_batch_ == true) + dim_vec.insert(dim_vec.begin(), max_batch_size_); + + DeclVar(name, dim_vec); + } void SetOp(const framework::proto::OpDesc& desc) { op_ = framework::OpRegistry::CreateOp(desc); - OpConverter op_converter; - op_converter.ConvertOp(desc, engine_.get()); + Singleton::Global().ConvertOp( + desc, parameters_, scope_, engine_.get(), true /*test_mode*/); engine_->FreezeNetwork(); // Declare outputs. - op_desc_.reset(new framework::OpDesc(desc, nullptr, nullptr)); + op_desc_.reset(new framework::OpDesc(desc, nullptr)); // Set Inputs. for (const auto& input : op_desc_->InputArgumentNames()) { + if (parameters_.count(input)) continue; auto* var = scope_.FindVar(input); PADDLE_ENFORCE(var); auto tensor = var->GetMutable(); + engine_->SetInputFromCPU( - input, static_cast(tensor->data()), + input, static_cast(tensor->data()), sizeof(float) * analysis::AccuDims(tensor->dims(), tensor->dims().size())); } @@ -120,26 +150,36 @@ class TRTConvertValidation { void Execute(int batch_size) { // Execute Fluid Op - // Execute TRT + PADDLE_ENFORCE_LE(batch_size, max_batch_size_); platform::CPUPlace place; platform::CPUDeviceContext ctx(place); - engine_->Execute(batch_size); - op_->Run(scope_, place); + // Execute TRT. + engine_->Execute(batch_size); + cudaStreamSynchronize(*engine_->stream()); ASSERT_FALSE(op_desc_->OutputArgumentNames().empty()); + const size_t output_space_size = 3000; for (const auto& output : op_desc_->OutputArgumentNames()) { std::vector fluid_out; - std::vector trt_out(200); - engine_->GetOutputInCPU(output, &trt_out[0], 200 * sizeof(float)); + std::vector trt_out(output_space_size); + engine_->GetOutputInCPU(output, &trt_out[0], output_space_size); + cudaStreamSynchronize(*engine_->stream()); auto* var = scope_.FindVar(output); auto tensor = var->GetMutable(); framework::TensorToVector(*tensor, ctx, &fluid_out); + + size_t fluid_out_size = fluid_out.size(); + if (if_add_batch_ == true) { + fluid_out_size = + batch_size * (framework::product(tensor->dims()) / max_batch_size_); + } // Compare two output ASSERT_FALSE(fluid_out.empty()); - for (size_t i = 0; i < fluid_out.size(); i++) { - EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 0.001); + for (size_t i = 0; i < fluid_out_size; i++) { + // Loose the threshold for CI in different machine model. + EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 2e-5); } } } @@ -149,9 +189,16 @@ class TRTConvertValidation { private: std::unique_ptr engine_; cudaStream_t stream_; - framework::Scope scope_; std::unique_ptr op_; std::unique_ptr op_desc_; + const std::unordered_set& parameters_; + framework::Scope& scope_; + // The ITensor of trt does not cotain the batch size, + // bug, in most cases, we need to set batch size for + // fluid's tensor shape. This variable indicates + // whether to add batch size to tensor shape of fluid. + bool if_add_batch_; + int max_batch_size_; }; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index a88236ae98e1816fc43796ead596c432b798d7de..b821c3d0bf425c46fae634fbf53f7ee63100ca5c 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -1,7 +1,7 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. +Licensed under the Apache License, Version 2.0 (the "License"); you may not use +this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 @@ -26,26 +26,32 @@ namespace paddle { namespace inference { namespace tensorrt { -void TensorRTEngine::Build(const DescType& paddle_model) { +int TensorRTEngine::runtime_batch_ = 1; + +void TensorRTEngine::Build(const DescType &paddle_model) { PADDLE_ENFORCE(false, "not implemented"); } void TensorRTEngine::Execute(int batch_size) { - std::vector buffers; - for (auto& buf : buffers_) { + batch_size_ = batch_size; + std::vector buffers; + for (auto &buf : buffers_) { PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated"); PADDLE_ENFORCE_GT(buf.max_size, 0); PADDLE_ENFORCE(buf.device == DeviceType::GPU); buffers.push_back(buf.buffer); } + PADDLE_ENFORCE_NOT_NULL(stream_); infer_context_->enqueue(batch_size, buffers.data(), *stream_, nullptr); cudaStreamSynchronize(*stream_); + SetRuntimeBatch(batch_size); } TensorRTEngine::~TensorRTEngine() { + cudaStreamSynchronize(*stream_); // clean buffer - for (auto& buf : buffers_) { - if (buf.buffer != nullptr) { + for (auto &buf : buffers_) { + if (buf.device == DeviceType::GPU && buf.buffer != nullptr) { PADDLE_ENFORCE_EQ(0, cudaFree(buf.buffer)); buf.buffer = nullptr; buf.max_size = 0; @@ -69,96 +75,122 @@ void TensorRTEngine::FreezeNetwork() { // allocate GPU buffers. buffers_.resize(buffer_sizes_.size()); - for (auto& item : buffer_sizes_) { + for (auto &item : buffer_sizes_) { + // The output buffers are not set in the network building phrase, need to + // infer from the TesorRT network. if (item.second == 0) { auto slot_offset = infer_engine_->getBindingIndex(item.first.c_str()); auto dims = infer_engine_->getBindingDimensions(slot_offset); item.second = kDataTypeSize[static_cast( infer_engine_->getBindingDataType(slot_offset))] * - analysis::AccuDims(dims.d, dims.nbDims); + analysis::AccuDims(dims.d, dims.nbDims) * max_batch_; + PADDLE_ENFORCE_GT(item.second, 0); } - auto& buf = buffer(item.first); + + auto &buf = buffer(item.first); + buf.max_size = item.second * max_batch_; CHECK(buf.buffer == nullptr); // buffer should be allocated only once. - PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second)); - buf.size = buf.max_size = item.second; + + PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second * max_batch_)); + buf.size = 0; + PADDLE_ENFORCE_LE(buf.max_size, 1 << 30); // 10G buf.device = DeviceType::GPU; } } -nvinfer1::ITensor* TensorRTEngine::DeclareInput(const std::string& name, +nvinfer1::ITensor *TensorRTEngine::DeclareInput(const std::string &name, nvinfer1::DataType dtype, - const nvinfer1::Dims& dims) { + const nvinfer1::Dims &dims) { PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate input name %s", name); PADDLE_ENFORCE(infer_network_ != nullptr, "should initnetwork first"); - auto* input = infer_network_->addInput(name.c_str(), dtype, dims); + auto *input = infer_network_->addInput(name.c_str(), dtype, dims); PADDLE_ENFORCE(input, "infer network add input %s failed", name); buffer_sizes_[name] = kDataTypeSize[static_cast(dtype)] * - analysis::AccuDims(dims.d, dims.nbDims); + analysis::AccuDims(dims.d, dims.nbDims) * max_batch_; + PADDLE_ENFORCE(input->isNetworkInput()); TensorRTEngine::SetITensor(name, input); return input; } -void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer* layer, int offset, - const std::string& name) { +void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer *layer, int offset, + const std::string &name) { PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", name); - auto* output = layer->getOutput(offset); + auto *output = layer->getOutput(offset); + SetITensor(name, output); PADDLE_ENFORCE(output != nullptr); output->setName(name.c_str()); + PADDLE_ENFORCE(!output->isNetworkInput()); infer_network_->markOutput(*output); + PADDLE_ENFORCE(output->isNetworkOutput()); // output buffers' size can only be decided latter, set zero here to mark this // and will reset latter. buffer_sizes_[name] = 0; } -void TensorRTEngine::DeclareOutput(const std::string& name) { +void TensorRTEngine::DeclareOutput(const std::string &name) { PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", name); - auto* output = TensorRTEngine::GetITensor(name); + auto *output = TensorRTEngine::GetITensor(name); PADDLE_ENFORCE(output != nullptr); output->setName(name.c_str()); + PADDLE_ENFORCE(!output->isNetworkInput()); infer_network_->markOutput(*output); // output buffers' size can only be decided latter, set zero here to mark this // and will reset latter. buffer_sizes_[name] = 0; } -void* TensorRTEngine::GetOutputInGPU(const std::string& name) { +void *TensorRTEngine::GetOutputInGPU(const std::string &name) { return buffer(name).buffer; } -void TensorRTEngine::GetOutputInGPU(const std::string& name, void* dst, +void TensorRTEngine::GetOutputInGPU(const std::string &name, void *dst, size_t max_size) { // determine data size + auto *output = TensorRTEngine::GetITensor(name); + nvinfer1::Dims dims = output->getDimensions(); + auto dim_size = analysis::AccuDims(dims.d, dims.nbDims); + size_t dst_size = dim_size * runtime_batch_ * + kDataTypeSize[static_cast(output->getType())]; + auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); PADDLE_ENFORCE_GT(it->second, 0); - PADDLE_ENFORCE_GE(max_size, it->second); - auto& buf = buffer(name); + PADDLE_ENFORCE_LE(dst_size, it->second); + PADDLE_ENFORCE_GE(max_size, dst_size); + auto &buf = buffer(name); PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before"); - PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, it->second, + PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, dst_size, cudaMemcpyDeviceToDevice, *stream_), 0); } -void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, +void TensorRTEngine::GetOutputInCPU(const std::string &name, void *dst, size_t max_size) { // determine data size + + auto *output = TensorRTEngine::GetITensor(name); + nvinfer1::Dims dims = output->getDimensions(); + auto dim_size = analysis::AccuDims(dims.d, dims.nbDims); + size_t dst_size = dim_size * runtime_batch_ * + kDataTypeSize[static_cast(output->getType())]; auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); PADDLE_ENFORCE_GT(it->second, 0); - PADDLE_ENFORCE_GE(max_size, it->second); - auto& buf = buffer(name); + PADDLE_ENFORCE_LE(dst_size, it->second); + PADDLE_ENFORCE_GE(max_size, dst_size); + auto &buf = buffer(name); PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before"); - PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, it->second, + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, dst_size, cudaMemcpyDeviceToHost, *stream_)); } -Buffer& TensorRTEngine::buffer(const std::string& name) { +Buffer &TensorRTEngine::buffer(const std::string &name) { PADDLE_ENFORCE(infer_engine_ != nullptr, "call FreezeNetwork first."); auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); @@ -166,19 +198,23 @@ Buffer& TensorRTEngine::buffer(const std::string& name) { return buffers_[slot_offset]; } -void TensorRTEngine::SetInputFromCPU(const std::string& name, const void* data, +void TensorRTEngine::SetInputFromCPU(const std::string &name, const void *data, size_t size) { - auto& buf = buffer(name); + auto &buf = buffer(name); PADDLE_ENFORCE_NOT_NULL(buf.buffer); + PADDLE_ENFORCE_NOT_NULL(data); + PADDLE_ENFORCE_NOT_NULL(stream_); PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small"); PADDLE_ENFORCE(buf.device == DeviceType::GPU); + buf.size = size; PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size, cudaMemcpyHostToDevice, *stream_)); } -void TensorRTEngine::SetInputFromGPU(const std::string& name, const void* data, +void TensorRTEngine::SetInputFromGPU(const std::string &name, const void *data, size_t size) { - auto& buf = buffer(name); + auto &buf = buffer(name); + buf.size = size; PADDLE_ENFORCE_NOT_NULL(buf.buffer); PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small"); PADDLE_ENFORCE(buf.device == DeviceType::GPU); @@ -186,19 +222,25 @@ void TensorRTEngine::SetInputFromGPU(const std::string& name, const void* data, cudaMemcpyDeviceToDevice, *stream_)); } -void TensorRTEngine::SetITensor(const std::string& name, - nvinfer1::ITensor* tensor) { +void TensorRTEngine::SetITensor(const std::string &name, + nvinfer1::ITensor *tensor) { PADDLE_ENFORCE(tensor != nullptr); PADDLE_ENFORCE_EQ(0, itensor_map_.count(name), "duplicate ITensor name %s", name); itensor_map_[name] = tensor; } -nvinfer1::ITensor* TensorRTEngine::GetITensor(const std::string& name) { +nvinfer1::ITensor *TensorRTEngine::GetITensor(const std::string &name) { PADDLE_ENFORCE(itensor_map_.count(name), "no ITensor %s", name); return itensor_map_[name]; } +void TensorRTEngine::SetRuntimeBatch(size_t batch_size) { + runtime_batch_ = batch_size; +} + +int TensorRTEngine::GetRuntimeBatch() { return runtime_batch_; } + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index d9d3163b66d4c4c302d12edcc42f00e1cdfa5a30..694468c419c20089de1cdecff1a903ad0cc6e99f 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -21,6 +21,7 @@ limitations under the License. */ #include #include "paddle/fluid/inference/engine.h" #include "paddle/fluid/inference/tensorrt/helper.h" +#include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace inference { @@ -37,23 +38,28 @@ class TensorRTEngine : public EngineBase { // Weight is model parameter. class Weight { public: - Weight(nvinfer1::DataType dtype, void* value, int num_elem) { + Weight(nvinfer1::DataType dtype, void* value, size_t num_elem) { w_.type = dtype; w_.values = value; w_.count = num_elem; } const nvinfer1::Weights& get() { return w_; } + std::vector dims; + private: nvinfer1::Weights w_; }; - TensorRTEngine(int max_batch, int max_workspace, cudaStream_t* stream, + TensorRTEngine(int max_batch, int max_workspace, + cudaStream_t* stream = nullptr, nvinfer1::ILogger& logger = NaiveLogger::Global()) : max_batch_(max_batch), max_workspace_(max_workspace), - stream_(stream), - logger_(logger) {} + stream_(stream ? stream : &default_stream_), + logger_(logger) { + cudaStreamCreate(&default_stream_); + } virtual ~TensorRTEngine(); @@ -111,13 +117,22 @@ class TensorRTEngine : public EngineBase { nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); } nvinfer1::INetworkDefinition* network() { return infer_network_.get(); } + void SetRuntimeBatch(size_t batch_size); + int GetRuntimeBatch(); private: // the max batch size int max_batch_; + // the runtime batch size + static int runtime_batch_; // the max memory size the engine uses int max_workspace_; + + // batch size of the current data, will be updated each Executation. + int batch_size_{-1}; cudaStream_t* stream_; + // If stream_ is not set from outside, hold its own stream. + cudaStream_t default_stream_; nvinfer1::ILogger& logger_; std::vector buffers_; @@ -129,7 +144,11 @@ class TensorRTEngine : public EngineBase { // TensorRT related internal members template struct Destroyer { - void operator()(T* x) { x->destroy(); } + void operator()(T* x) { + if (x) { + x->destroy(); + } + } }; template using infer_ptr = std::unique_ptr>; @@ -153,6 +172,38 @@ class TensorRTEngine : public EngineBase { #define TRT_ENGINE_ADD_LAYER(engine__, layer__, ARGS...) \ engine__->network()->add##layer__(ARGS); +/* + * Helper to control the TensorRT engine's creation and deletion. + */ +class TRT_EngineManager { + public: + bool HasEngine(const std::string& name) const { + return engines_.count(name) != 0; + } + + // Get an engine called `name`. + TensorRTEngine* Get(const std::string& name) const { + return engines_.at(name).get(); + } + + // Create or get an engine called `name` + TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream, + const std::string& name) { + auto* p = new TensorRTEngine(max_batch, max_workspace, stream); + engines_[name].reset(p); + return p; + } + + void DeleteALl() { + for (auto& item : engines_) { + item.second.reset(nullptr); + } + } + + private: + std::unordered_map> engines_; +}; + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc index e635f0f87d577a1f1ac74687ee60f762be525418..dc03702990587bf5e65d28da662d10df4d882110 100644 --- a/paddle/fluid/inference/tensorrt/test_engine.cc +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -28,7 +28,7 @@ class TensorRTEngineTest : public ::testing::Test { protected: void SetUp() override { ASSERT_EQ(0, cudaStreamCreate(&stream_)); - engine_ = new TensorRTEngine(1, 1 << 10, &stream_); + engine_ = new TensorRTEngine(10, 1 << 10, &stream_); engine_->InitNetwork(); } @@ -71,7 +71,7 @@ TEST_F(TensorRTEngineTest, add_layer) { LOG(INFO) << "to get output"; float y_cpu; - engine_->GetOutputInCPU("y", &y_cpu, sizeof(float)); + engine_->GetOutputInCPU("y", &y_cpu, 1 * sizeof(float)); LOG(INFO) << "to checkout output"; ASSERT_EQ(y_cpu, x_v * 2 + 3); @@ -103,11 +103,80 @@ TEST_F(TensorRTEngineTest, add_layer_multi_dim) { LOG(INFO) << "to get output"; float y_cpu[2] = {-1., -1.}; - engine_->GetOutputInCPU("y", &y_cpu[0], sizeof(float) * 2); + + auto dims = engine_->GetITensor("y")->getDimensions(); + ASSERT_EQ(dims.nbDims, 3); + ASSERT_EQ(dims.d[0], 2); + ASSERT_EQ(dims.d[1], 1); + engine_->GetOutputInCPU("y", &y_cpu[0], 2 * sizeof(float)); ASSERT_EQ(y_cpu[0], 4.5); ASSERT_EQ(y_cpu[1], 14.5); } +TEST_F(TensorRTEngineTest, test_conv2d) { + // Weight in CPU memory. + float raw_weight[9] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0}; + float raw_bias[1] = {0}; + + TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, 9); + TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, 1); + auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT, + nvinfer1::Dims3{1, 3, 3}); + auto* conv_layer = + TRT_ENGINE_ADD_LAYER(engine_, Convolution, *x, 1, nvinfer1::DimsHW{3, 3}, + weight.get(), bias.get()); + PADDLE_ENFORCE(conv_layer != nullptr); + conv_layer->setStride(nvinfer1::DimsHW{1, 1}); + conv_layer->setPadding(nvinfer1::DimsHW{1, 1}); + + engine_->DeclareOutput(conv_layer, 0, "y"); + engine_->FreezeNetwork(); + ASSERT_EQ(engine_->engine()->getNbBindings(), 2); + + float x_v[18] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, + 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0}; + engine_->SetInputFromCPU("x", reinterpret_cast(&x_v), + 18 * sizeof(float)); + engine_->Execute(2); + + LOG(INFO) << "to get output"; + float* y_cpu = new float[18]; + engine_->GetOutputInCPU("y", &y_cpu[0], 18 * sizeof(float)); + ASSERT_EQ(y_cpu[0], 4.0); + ASSERT_EQ(y_cpu[1], 6.0); +} + +TEST_F(TensorRTEngineTest, test_pool2d) { + // Weight in CPU memory. + auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT, + nvinfer1::Dims3{1, 2, 2}); + + nvinfer1::PoolingType pool_t = nvinfer1::PoolingType::kAVERAGE; + auto* pool_layer = + TRT_ENGINE_ADD_LAYER(engine_, Pooling, *const_cast(x), + pool_t, nvinfer1::DimsHW{2, 2}); + + PADDLE_ENFORCE(pool_layer != nullptr); + pool_layer->setStride(nvinfer1::DimsHW{1, 1}); + pool_layer->setPadding(nvinfer1::DimsHW{0, 0}); + + engine_->DeclareOutput(pool_layer, 0, "y"); + engine_->FreezeNetwork(); + ASSERT_EQ(engine_->engine()->getNbBindings(), 2); + + float x_v[8] = {1.0, 2.0, 5.0, 0.0, 2.0, 3.0, 5.0, 10.0}; + engine_->SetInputFromCPU("x", reinterpret_cast(&x_v), + 8 * sizeof(float)); + engine_->Execute(2); + + LOG(INFO) << "to get output"; + float* y_cpu = new float[2]; + engine_->GetOutputInCPU("y", &y_cpu[0], 2 * sizeof(float)); + + ASSERT_EQ(y_cpu[0], 2.0); + ASSERT_EQ(y_cpu[1], 5.0); +} + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/book/CMakeLists.txt b/paddle/fluid/inference/tests/book/CMakeLists.txt index dbb81462b8273bd701e9c9f530eaf69817abd6a1..017fc4cd7b11c150cb941fffca2606a4d707330f 100644 --- a/paddle/fluid/inference/tests/book/CMakeLists.txt +++ b/paddle/fluid/inference/tests/book/CMakeLists.txt @@ -17,7 +17,7 @@ function(inference_test TARGET_NAME) string(REGEX REPLACE "^_$" "" arg "${arg}") cc_test(test_inference_${TARGET_NAME}${arg} SRCS test_inference_${TARGET_NAME}.cc - DEPS paddle_fluid + DEPS paddle_fluid_origin ARGS --dirname=${PYTHON_TESTS_DIR}/book/${TARGET_NAME}${arg}.inference.model) set_tests_properties(test_inference_${TARGET_NAME}${arg} PROPERTIES DEPENDS test_${TARGET_NAME}) @@ -38,3 +38,11 @@ inference_test(recommender_system) #inference_test(rnn_encoder_decoder) #inference_test(understand_sentiment ARGS conv) inference_test(word2vec) + +# This is an unly work around to make this test run +# TODO(TJ): clean me up +cc_test(test_inference_nlp + SRCS test_inference_nlp.cc + DEPS paddle_fluid_origin + ARGS + --model_path=${PADDLE_BINARY_DIR}/python/paddle/fluid/tests/book/recognize_digits_mlp.inference.model) diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index 987da18116cc6f4902bd66ae317f2470a8bc5057..60c761c5281e2f535aab0200c93fb738addcdb87 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -21,7 +21,6 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); DEFINE_int32(batch_size, 1, "Batch size of input data"); DEFINE_int32(repeat, 1, "Running the inference program repeat times"); DEFINE_bool(skip_cpu, false, "Skip the cpu test"); -DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference"); TEST(inference, image_classification) { if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { @@ -59,10 +58,8 @@ TEST(inference, image_classification) { // Run inference on CPU LOG(INFO) << "--- CPU Runs: ---"; LOG(INFO) << "Batch size is " << FLAGS_batch_size; - LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn; TestInference( - dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined, - FLAGS_use_mkldnn); + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined); LOG(INFO) << output1.dims(); } diff --git a/paddle/fluid/inference/tests/book/test_inference_nlp.cc b/paddle/fluid/inference/tests/book/test_inference_nlp.cc new file mode 100644 index 0000000000000000000000000000000000000000..e2a3e9d46ef9f303d191d59253ffbe9f4826184b --- /dev/null +++ b/paddle/fluid/inference/tests/book/test_inference_nlp.cc @@ -0,0 +1,221 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include // NOLINT +#include "gflags/gflags.h" +#include "gtest/gtest.h" +#include "paddle/fluid/inference/tests/test_helper.h" +#include "paddle/fluid/platform/cpu_helper.h" + +DEFINE_string(model_path, "", "Directory of the inference model."); +DEFINE_string(data_file, "", "File of input index data."); +DEFINE_int32(repeat, 100, "Running the inference program repeat times"); +DEFINE_bool(prepare_vars, true, "Prepare variables before executor"); +DEFINE_int32(num_threads, 1, "Number of threads should be used"); +DECLARE_bool(use_mkldnn); +DECLARE_int32(paddle_num_threads); + +inline double GetCurrentMs() { + struct timeval time; + gettimeofday(&time, NULL); + return 1e+3 * time.tv_sec + 1e-3 * time.tv_usec; +} + +// This function just give dummy data for recognize_digits model. +size_t DummyData(std::vector* out) { + paddle::framework::LoDTensor input; + SetupTensor(&input, {1, 1, 28, 28}, -1.f, 1.f); + out->emplace_back(input); + return 1; +} + +// Load the input word index data from file and save into LodTensor. +// Return the size of words. +size_t LoadData(std::vector* out, + const std::string& filename) { + if (filename.empty()) { + return DummyData(out); + } + + size_t sz = 0; + std::fstream fin(filename); + std::string line; + out->clear(); + while (getline(fin, line)) { + std::istringstream iss(line); + std::vector ids; + std::string field; + while (getline(iss, field, ' ')) { + ids.push_back(stoi(field)); + } + if (ids.size() >= 1024) { + // Synced with NLP guys, they will ignore input larger then 1024 + continue; + } + + paddle::framework::LoDTensor words; + paddle::framework::LoD lod{{0, ids.size()}}; + words.set_lod(lod); + int64_t* pdata = words.mutable_data( + {static_cast(ids.size()), 1}, paddle::platform::CPUPlace()); + memcpy(pdata, ids.data(), words.numel() * sizeof(int64_t)); + out->emplace_back(words); + sz += ids.size(); + } + return sz; +} + +// Split input data samples into small pieces jobs as balanced as possible, +// according to the number of threads. +void SplitData( + const std::vector& datasets, + std::vector>* jobs, + const int num_threads) { + size_t s = 0; + jobs->resize(num_threads); + while (s < datasets.size()) { + for (auto it = jobs->begin(); it != jobs->end(); it++) { + it->emplace_back(&datasets[s]); + s++; + if (s >= datasets.size()) { + break; + } + } + } +} + +void ThreadRunInfer( + const int tid, paddle::framework::Scope* scope, + const std::vector>& jobs) { + // maybe framework:ProgramDesc is not thread-safe + paddle::platform::CPUPlace place; + paddle::framework::Executor executor(place); + auto& sub_scope = scope->NewScope(); + auto inference_program = + paddle::inference::Load(&executor, scope, FLAGS_model_path); + + auto ctx = executor.Prepare(*inference_program, /*block_id*/ 0); + executor.CreateVariables(*inference_program, &sub_scope, /*block_id*/ 0); + + const std::vector& feed_target_names = + inference_program->GetFeedTargetNames(); + const std::vector& fetch_target_names = + inference_program->GetFetchTargetNames(); + + PADDLE_ENFORCE_EQ(fetch_target_names.size(), 1UL); + std::map fetch_targets; + paddle::framework::LoDTensor outtensor; + fetch_targets[fetch_target_names[0]] = &outtensor; + + std::map feed_targets; + PADDLE_ENFORCE_EQ(feed_target_names.size(), 1UL); + + auto& inputs = jobs[tid]; + auto start_ms = GetCurrentMs(); + for (size_t i = 0; i < inputs.size(); ++i) { + feed_targets[feed_target_names[0]] = inputs[i]; + executor.RunPreparedContext(ctx.get(), &sub_scope, &feed_targets, + &fetch_targets, false /*create_local_scope*/); + } + auto stop_ms = GetCurrentMs(); + scope->DeleteScope(&sub_scope); + LOG(INFO) << "Tid: " << tid << ", process " << inputs.size() + << " samples, avg time per sample: " + << (stop_ms - start_ms) / inputs.size() << " ms"; +} + +TEST(inference, nlp) { + if (FLAGS_model_path.empty()) { + LOG(FATAL) << "Usage: ./example --model_path=path/to/your/model"; + } + if (FLAGS_data_file.empty()) { + LOG(WARNING) << "No data file provided, will use dummy data!" + << "Note: if you use nlp model, please provide data file."; + } + LOG(INFO) << "Model Path: " << FLAGS_model_path; + LOG(INFO) << "Data File: " << FLAGS_data_file; + + std::vector datasets; + size_t num_total_words = LoadData(&datasets, FLAGS_data_file); + LOG(INFO) << "Number of samples (seq_len<1024): " << datasets.size(); + LOG(INFO) << "Total number of words: " << num_total_words; + + // 0. Call `paddle::framework::InitDevices()` initialize all the devices + std::unique_ptr scope( + new paddle::framework::Scope()); + + paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + + double start_ms = 0, stop_ms = 0; + if (FLAGS_num_threads > 1) { + std::vector> jobs; + SplitData(datasets, &jobs, FLAGS_num_threads); + std::vector> threads; + start_ms = GetCurrentMs(); + for (int i = 0; i < FLAGS_num_threads; ++i) { + threads.emplace_back( + new std::thread(ThreadRunInfer, i, scope.get(), std::ref(jobs))); + } + for (int i = 0; i < FLAGS_num_threads; ++i) { + threads[i]->join(); + } + stop_ms = GetCurrentMs(); + } else { + // 1. Define place, executor, scope + paddle::platform::CPUPlace place; + paddle::framework::Executor executor(place); + + // 2. Initialize the inference_program and load parameters + std::unique_ptr inference_program; + inference_program = InitProgram(&executor, scope.get(), FLAGS_model_path, + /*model combined*/ false); + // always prepare context + std::unique_ptr ctx; + ctx = executor.Prepare(*inference_program, 0); + if (FLAGS_prepare_vars) { + executor.CreateVariables(*inference_program, scope.get(), 0); + } + // preapre fetch + const std::vector& fetch_target_names = + inference_program->GetFetchTargetNames(); + PADDLE_ENFORCE_EQ(fetch_target_names.size(), 1UL); + std::map fetch_targets; + paddle::framework::LoDTensor outtensor; + fetch_targets[fetch_target_names[0]] = &outtensor; + + // prepare feed + const std::vector& feed_target_names = + inference_program->GetFeedTargetNames(); + PADDLE_ENFORCE_EQ(feed_target_names.size(), 1UL); + std::map feed_targets; + + // feed data and run + start_ms = GetCurrentMs(); + for (size_t i = 0; i < datasets.size(); ++i) { + feed_targets[feed_target_names[0]] = &(datasets[i]); + executor.RunPreparedContext(ctx.get(), scope.get(), &feed_targets, + &fetch_targets, !FLAGS_prepare_vars); + } + stop_ms = GetCurrentMs(); + LOG(INFO) << "Tid: 0, process " << datasets.size() + << " samples, avg time per sample: " + << (stop_ms - start_ms) / datasets.size() << " ms"; + } + LOG(INFO) << "Total inference time with " << FLAGS_num_threads + << " threads : " << (stop_ms - start_ms) / 1000.0 + << " sec, QPS: " << datasets.size() / ((stop_ms - start_ms) / 1000); +} diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 01b8dc0be662da22fe15a79cd9abfe5fa92c9577..695790a37dce889e838462b401ca4e89f09271d5 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -22,6 +22,8 @@ limitations under the License. */ #include "paddle/fluid/inference/io.h" #include "paddle/fluid/platform/profiler.h" +DECLARE_bool(use_mkldnn); + template void SetupTensor(paddle::framework::LoDTensor* input, paddle::framework::DDim dims, T lower, T upper) { @@ -133,24 +135,11 @@ std::vector> GetFeedTargetShapes( return feed_target_shapes; } -void EnableMKLDNN( - const std::unique_ptr& program) { - for (size_t bid = 0; bid < program->Size(); ++bid) { - auto* block = program->MutableBlock(bid); - for (auto* op : block->AllOps()) { - if (op->HasAttr("use_mkldnn")) { - op->SetAttr("use_mkldnn", true); - } - } - } -} - template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, const std::vector& cpu_fetchs, - const int repeat = 1, const bool is_combined = false, - const bool use_mkldnn = false) { + const int repeat = 1, const bool is_combined = false) { // 1. Define place, executor, scope auto place = Place(); auto executor = paddle::framework::Executor(place); @@ -182,9 +171,6 @@ void TestInference(const std::string& dirname, "init_program", paddle::platform::DeviceContextPool::Instance().Get(place)); inference_program = InitProgram(&executor, scope, dirname, is_combined); - if (use_mkldnn) { - EnableMKLDNN(inference_program); - } } // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, @@ -210,7 +196,10 @@ void TestInference(const std::string& dirname, fetch_targets[fetch_target_names[i]] = cpu_fetchs[i]; } - // 6. Run the inference program + // 6. If export Flags_use_mkldnn=True, use mkldnn related ops. + if (FLAGS_use_mkldnn) executor.EnableMKLDNN(*inference_program); + + // 7. Run the inference program { if (!CreateVars) { // If users don't want to create and destroy variables every time they @@ -221,13 +210,14 @@ void TestInference(const std::string& dirname, // Ignore the profiling results of the first run std::unique_ptr ctx; + bool CreateLocalScope = CreateVars; if (PrepareContext) { ctx = executor.Prepare(*inference_program, 0); executor.RunPreparedContext(ctx.get(), scope, &feed_targets, - &fetch_targets, true, CreateVars); + &fetch_targets, CreateLocalScope, CreateVars); } else { executor.Run(*inference_program, scope, &feed_targets, &fetch_targets, - true, CreateVars); + CreateLocalScope, CreateVars); } // Enable the profiler @@ -243,10 +233,11 @@ void TestInference(const std::string& dirname, // Note: if you change the inference_program, you need to call // executor.Prepare() again to get a new ExecutorPrepareContext. executor.RunPreparedContext(ctx.get(), scope, &feed_targets, - &fetch_targets, CreateVars); + &fetch_targets, CreateLocalScope, + CreateVars); } else { executor.Run(*inference_program, scope, &feed_targets, &fetch_targets, - CreateVars); + CreateLocalScope, CreateVars); } } diff --git a/paddle/fluid/memory/detail/buddy_allocator.cc b/paddle/fluid/memory/detail/buddy_allocator.cc index 4194ba197948b47003863196efdac1c08a7ae4f6..c2f45fdc99b87bc12c2aadf1985de6e98a24fce7 100644 --- a/paddle/fluid/memory/detail/buddy_allocator.cc +++ b/paddle/fluid/memory/detail/buddy_allocator.cc @@ -15,12 +15,17 @@ limitations under the License. */ #include "paddle/fluid/memory/detail/buddy_allocator.h" #include "glog/logging.h" +DEFINE_bool(free_idle_memory, false, + "If it is true, Paddle will try to free idle memory trunks during " + "running time."); + namespace paddle { namespace memory { namespace detail { -BuddyAllocator::BuddyAllocator(SystemAllocator* system_allocator, - size_t min_chunk_size, size_t max_chunk_size) +BuddyAllocator::BuddyAllocator( + std::unique_ptr system_allocator, size_t min_chunk_size, + size_t max_chunk_size) : min_chunk_size_(min_chunk_size), max_chunk_size_(max_chunk_size), cache_(system_allocator->UseGpu()), @@ -151,13 +156,14 @@ void BuddyAllocator::Free(void* p) { pool_.insert( IndexSizeAddress(block->index(cache_), block->total_size(cache_), block)); - // Clean up if existing too much free memory - - // Prefer freeing fallback allocation first - CleanIdleFallBackAlloc(); + if (FLAGS_free_idle_memory) { + // Clean up if existing too much free memory + // Prefer freeing fallback allocation first + CleanIdleFallBackAlloc(); - // Free normal allocation - CleanIdleNormalAlloc(); + // Free normal allocation + CleanIdleNormalAlloc(); + } } size_t BuddyAllocator::Used() { return total_used_; } diff --git a/paddle/fluid/memory/detail/buddy_allocator.h b/paddle/fluid/memory/detail/buddy_allocator.h index 2f39d774d6fb6a2bc37877eb2f8b90bebd3cda28..f0c83efc23ce39c4fc89296d672e1e55751851bf 100644 --- a/paddle/fluid/memory/detail/buddy_allocator.h +++ b/paddle/fluid/memory/detail/buddy_allocator.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include // NOLINT #include #include @@ -32,8 +33,8 @@ namespace detail { class BuddyAllocator { public: - BuddyAllocator(SystemAllocator* system_allocator, size_t min_chunk_size, - size_t max_chunk_size); + BuddyAllocator(std::unique_ptr system_allocator, + size_t min_chunk_size, size_t max_chunk_size); ~BuddyAllocator(); @@ -103,7 +104,7 @@ class BuddyAllocator { private: /*! Allocate CPU/GPU memory from system */ - SystemAllocator* system_allocator_; + std::unique_ptr system_allocator_; std::mutex mutex_; }; diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc index d5390529163491c2711e50ffad236534e88b73ee..9b1ab1e228dd758b52975abc4c4aa0bdeadbe2de 100644 --- a/paddle/fluid/memory/detail/system_allocator.cc +++ b/paddle/fluid/memory/detail/system_allocator.cc @@ -43,14 +43,16 @@ void* CPUAllocator::Alloc(size_t* index, size_t size) { *index = 0; // unlock memory - void* p; + void* p = nullptr; #ifdef PADDLE_WITH_MKLDNN // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp // memory alignment - PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0); + PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0, "Alloc %ld error!", + size); #else - PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0); + PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0, "Alloc %ld error!", + size); #endif PADDLE_ENFORCE(p, "Fail to allocate CPU memory: size = %d .", size); diff --git a/paddle/fluid/memory/malloc.cc b/paddle/fluid/memory/malloc.cc index 0c74f62de5c6f5d432ee928945db6dcf385ca209..7c800b3c164049244770ceb2070b177d8307e85e 100644 --- a/paddle/fluid/memory/malloc.cc +++ b/paddle/fluid/memory/malloc.cc @@ -12,6 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include + #include "paddle/fluid/memory/malloc.h" #include "glog/logging.h" @@ -20,6 +22,12 @@ limitations under the License. */ #include "paddle/fluid/memory/detail/system_allocator.h" #include "paddle/fluid/platform/gpu_info.h" +DEFINE_bool(init_allocated_mem, false, + "It is a mistake that the values of the memory allocated by " + "BuddyAllocator are always zeroed in some op's implementation. " + "To find this error in time, we use init_allocated_mem to indicate " + "that initializing the allocated memory with a small value " + "during unit testing."); DECLARE_double(fraction_of_gpu_memory_to_use); namespace paddle { @@ -28,12 +36,15 @@ namespace memory { using BuddyAllocator = detail::BuddyAllocator; BuddyAllocator* GetCPUBuddyAllocator() { + static std::once_flag init_flag; static detail::BuddyAllocator* a = nullptr; - if (a == nullptr) { - a = new detail::BuddyAllocator(new detail::CPUAllocator, - platform::CpuMinChunkSize(), - platform::CpuMaxChunkSize()); - } + + std::call_once(init_flag, []() { + a = new detail::BuddyAllocator( + std::unique_ptr(new detail::CPUAllocator), + platform::CpuMinChunkSize(), platform::CpuMaxChunkSize()); + }); + return a; } @@ -41,6 +52,9 @@ template <> void* Alloc(platform::CPUPlace place, size_t size) { VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place); void* p = GetCPUBuddyAllocator()->Alloc(size); + if (FLAGS_init_allocated_mem) { + memset(p, 0xEF, size); + } VLOG(10) << " pointer=" << p; return p; } @@ -59,27 +73,33 @@ size_t Used(platform::CPUPlace place) { #ifdef PADDLE_WITH_CUDA BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { - static BuddyAllocator** as = NULL; - if (as == NULL) { + static std::once_flag init_flag; + static detail::BuddyAllocator** a_arr = nullptr; + + std::call_once(init_flag, [gpu_id]() { int gpu_num = platform::GetCUDADeviceCount(); - as = new BuddyAllocator*[gpu_num]; - for (int gpu = 0; gpu < gpu_num; gpu++) { - as[gpu] = nullptr; + PADDLE_ENFORCE(gpu_id < gpu_num, "gpu_id:%d should < gpu_num:%d", gpu_id, + gpu_num); + + a_arr = new BuddyAllocator*[gpu_num]; + for (int i = 0; i < gpu_num; i++) { + a_arr[i] = nullptr; + platform::SetDeviceId(i); + a_arr[i] = new BuddyAllocator( + std::unique_ptr(new detail::GPUAllocator(i)), + platform::GpuMinChunkSize(), platform::GpuMaxChunkSize()); + + VLOG(10) << "\n\nNOTE: each GPU device use " + << FLAGS_fraction_of_gpu_memory_to_use * 100 + << "% of GPU memory.\n" + << "You can set GFlags environment variable '" + << "FLAGS_fraction_of_gpu_memory_to_use" + << "' to change the fraction of GPU usage.\n\n"; } - } + }); + platform::SetDeviceId(gpu_id); - if (!as[gpu_id]) { - as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator(gpu_id), - platform::GpuMinChunkSize(), - platform::GpuMaxChunkSize()); - VLOG(10) << "\n\nNOTE: each GPU device use " - << FLAGS_fraction_of_gpu_memory_to_use * 100 - << "% of GPU memory.\n" - << "You can set GFlags environment variable '" - << "FLAGS_fraction_of_gpu_memory_to_use" - << "' to change the fraction of GPU usage.\n\n"; - } - return as[gpu_id]; + return a_arr[gpu_id]; } template <> @@ -104,6 +124,9 @@ void* Alloc(platform::CUDAPlace place, size_t size) { LOG(WARNING) << "GPU memory used: " << Used(place); platform::SetDeviceId(cur_dev); } + if (FLAGS_init_allocated_mem) { + cudaMemset(ptr, 0xEF, size); + } return ptr; } @@ -113,12 +136,16 @@ void Free(platform::CUDAPlace place, void* p) { } BuddyAllocator* GetCUDAPinnedBuddyAllocator() { - static BuddyAllocator* ba = NULL; - if (ba == NULL) { - ba = new BuddyAllocator(new detail::CUDAPinnedAllocator, + static std::once_flag init_flag; + static BuddyAllocator* ba = nullptr; + + std::call_once(init_flag, []() { + ba = new BuddyAllocator(std::unique_ptr( + new detail::CUDAPinnedAllocator), platform::CUDAPinnedMinChunkSize(), platform::CUDAPinnedMaxChunkSize()); - } + }); + return ba; } @@ -137,6 +164,9 @@ void* Alloc(platform::CUDAPinnedPlace place, LOG(WARNING) << "cudaMallocHost Cannot allocate " << size << " bytes in CUDAPinnedPlace"; } + if (FLAGS_init_allocated_mem) { + memset(ptr, 0xEF, size); + } return ptr; } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index de6ff29c6f8edbcf930546ff157a1c226e1311db..e8b5dec9d49f5613cec92441d19ab7dc1a1ad90c 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -166,10 +166,13 @@ function(op_library TARGET) # NOTE(*): activation use macro to regist the kernels, set use_op manually. if(${TARGET} STREQUAL "activation") file(APPEND ${pybind_file} "USE_OP(relu);\n") - elseif(${TARGET} STREQUAL "reduce") - file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n") elseif(${TARGET} STREQUAL "fake_dequantize") file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n") + elseif(${TARGET} STREQUAL "tensorrt_engine_op") + message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference") + elseif(${TARGET} STREQUAL "fc") + # HACK: fc only have mkldnn and cpu, which would mismatch the cpu only condition + file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n") else() file(APPEND ${pybind_file} "USE_OP(${TARGET});\n") endif() @@ -186,47 +189,68 @@ else() set(DEPS_OPS ${DEPS_OPS} nccl_op) endif() -add_subdirectory(detail) +set(DISTRIBUTE_DEPS "") if(WITH_DISTRIBUTE) + add_subdirectory(distributed) - set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) + set(DISTRIBUTE_DEPS "") + if(WITH_GRPC) + set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node) + else() + set(DISTRIBUTE_DEPS sendrecvop_brpc brpc leveldb snappystream snappy protobuf ssl crypto zlib node) + if(WITH_BRPC_RDMA) + find_library(IBVERBS_LIBRARY NAMES ibverbs) + ADD_LIBRARY(ibverbs SHARED IMPORTED GLOBAL) + SET_PROPERTY(TARGET ibverbs PROPERTY IMPORTED_LOCATION ${IBVERBS_LIBRARY}) + + + find_library(RDMACM_LIBRARY NAMES rdmacm) + ADD_LIBRARY(rdmacm SHARED IMPORTED GLOBAL) + SET_PROPERTY(TARGET rdmacm PROPERTY IMPORTED_LOCATION ${RDMACM_LIBRARY}) + + set(DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} ibverbs rdmacm) + endif() + endif() + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - op_library(send_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(prefetch_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(prefetch_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(recv_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(listen_and_serv_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(listen_and_serv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(send_vars_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_vars_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS}) - op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - set_source_files_properties(fetch_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + foreach(dist_op "prefetch_op" "checkpoint_notify_op" "listen_and_serv_op" "send_op" "recv_op" "send_barrier_op" "fetch_barrier_op") + op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + endforeach() + #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op # listen_and_serv_op sum_op executor SERIAL) if(WITH_GPU) set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op - listen_and_serv_op executor SERIAL) - op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc) + cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op ${DISTRIBUTE_DEPS} executor SERIAL) + if(WITH_GRPC) + op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc) + else() + op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_brpc) + endif() set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) else() set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op) endif() else() - set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op fetch_barrier_op gen_nccl_id_op) + set(DEPS_OPS ${DEPS_OPS} checkpoint_notify_op prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op) endif() op_library(cross_entropy_op DEPS cross_entropy) -op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax) +if(WITH_GPU) + op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax cub) +else() + op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax) +endif() + op_library(softmax_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) if (WITH_GPU AND TENSORRT_FOUND) - op_library(tensorrt_engine_op DEPS tensorrt_engine) + op_library(tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter) + nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc + DEPS tensorrt_engine_op + analysis) else() set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op) endif() @@ -245,15 +269,21 @@ op_library(max_sequence_len_op DEPS lod_rank_table) op_library(sequence_conv_op DEPS context_project) op_library(sequence_pool_op DEPS sequence_pooling) op_library(lstm_op DEPS sequence2batch lstm_compute) +op_library(hierarchical_sigmoid_op DEPS matrix_bit_code) op_library(lstmp_op DEPS sequence2batch lstm_compute) op_library(gru_op DEPS sequence2batch gru_compute) op_library(recurrent_op DEPS executor) op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale) op_library(cos_sim_op DEPS cos_sim_functor) op_library(parallel_do_op DEPS executor) +op_library(unsqueeze_op DEPS reshape_op) +op_library(squeeze_op DEPS reshape_op) +op_library(extract_rows_op DEPS memory) +op_library(flatten_op DEPS reshape_op) if (WITH_GPU) op_library(conv_op DEPS vol2col depthwise_conv im2col) + op_library(layer_norm_op DEPS cub) else() op_library(conv_op DEPS vol2col im2col) endif() @@ -273,12 +303,6 @@ op_library(channel_recv_op DEPS concurrency) list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) -# The fully connected layer is deleted when the WITH_MKLDNN flag is OFF -# Because the fully connected layer has only one MKLDNN's operator -if(NOT WITH_MKLDNN) - list(REMOVE_ITEM GENERAL_OPS fc_op) -endif(NOT WITH_MKLDNN) - foreach(src ${GENERAL_OPS}) op_library(${src}) endforeach() @@ -296,6 +320,7 @@ foreach(src ${DETECTION_LIBRARY}) endforeach() set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") +set(GLOB_DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} CACHE INTERNAL "distributed dependency") cc_test(gather_test SRCS gather_test.cc DEPS tensor) cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc index b892ac77d9ed60210ddadaecb1a4f214e5a25180..137bca5e2b8e2754aed274970e08b03ee816a7f2 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -12,16 +12,20 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "mkldnn.hpp" #include "paddle/fluid/operators/activation_op.h" -#include "paddle/fluid/operators/mkldnn_activation_op.h" #include "paddle/fluid/platform/mkldnn_helper.h" namespace paddle { namespace operators { -using paddle::framework::Tensor; -using paddle::platform::MKLDNNDeviceContext; +using framework::DataLayout; +using framework::Tensor; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::stream; +using platform::GetMKLDNNFormat; +using platform::MKLDNNDeviceContext; +using platform::to_void_cast; namespace { std::string gethash(const mkldnn::memory::dims &operand_dims, @@ -35,222 +39,294 @@ std::string gethash(const mkldnn::memory::dims &operand_dims, }; return dim2str(operand_dims) + std::to_string(algorithm); } +} // namespace + +template +class MKLDNNActivationKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *x = ctx.Input("X"); + PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN && + x->format() != memory::format::format_undef, + "Wrong layout/format set for Input x tensor"); + + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto &attr : attrs) { + *attr.second = ctx.Attr(attr.first); + } + functor(ctx); + } +}; -template -void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, - const T alpha = 0, const T beta = 0) { +template +class MKLDNNActivationGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *diff_y = ctx.Input(framework::GradVarName("Out")); + PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN && + diff_y->format() != memory::format::format_undef, + "Wrong layout/format set for Input OutGrad tensor"); + + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto &attr : attrs) { + *attr.second = ctx.Attr(attr.first); + } + functor(ctx); + } +}; + +template +void eltwise_forward(const framework::ExecutionContext &ctx, + mkldnn::algorithm algorithm, const T alpha = 0, + const T beta = 0) { PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); - auto &dev_ctx = ctx.template device_context(); const auto &mkldnn_engine = dev_ctx.GetEngine(); - // get buffers - const auto *src = ctx.template Input("X"); - const auto *src_data = src->template data(); + const auto *x = ctx.Input("X"); + auto *y = ctx.Output("Out"); - auto *dst = ctx.template Output("Out"); - T *dst_data = dst->template mutable_data(ctx.GetPlace()); + const T *x_data = x->data(); + T *y_data = y->mutable_data(ctx.GetPlace()); - // get memory dim - PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4, + PADDLE_ENFORCE(x->dims().size() == 2 || x->dims().size() == 4, "Input dim must be with 2 or 4"); - std::vector src_tz = framework::vectorize2int(src->dims()); + + std::vector src_tz = framework::vectorize2int(x->dims()); + + auto src_format = + src_tz.size() == 2 ? mkldnn::memory::format::nc : x->format(); const std::string key = gethash(src_tz, algorithm); const std::string key_src_data = key + ctx.op().Output("Out") + "@eltwise_fwd_src_data"; - const std::string key_src_mem = key + "@eltwise_fwd_src_mem"; - const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem"; - const std::string key_fwd = key + "@eltwise_fwd"; + const std::string key_src_layout = + key + ctx.op().Output("Out") + "@eltwise_fwd_src_layout"; + const std::string key_with_layout = key + std::to_string(src_format); + const std::string key_src_mem = key_with_layout + "@eltwise_fwd_src_mem"; + const std::string key_dst_mem = key_with_layout + "@eltwise_fwd_dst_mem"; + const std::string key_fwd = key_with_layout + "@eltwise_fwd"; + const std::string key_fwd_pd = key_with_layout + "@eltwise_fwd_pd"; + + // save input data and layout to be referred in backward path + auto p_src_data = std::make_shared(x_data); + dev_ctx.SetBlob(key_src_data, p_src_data); + auto p_src_layout = std::make_shared(src_format); + dev_ctx.SetBlob(key_src_layout, p_src_layout); auto p_fwd = std::static_pointer_cast( dev_ctx.GetBlob(key_fwd)); - // save input data to be referred in backward path - auto p_src_data = std::make_shared(src_data); - dev_ctx.SetBlob(key_src_data, p_src_data); + std::shared_ptr dst_memory; if (p_fwd == nullptr) { - // create memory description - auto data_md = src_tz.size() == 2 - ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nc) - : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // create memory primitives - auto p_src_mem = std::make_shared(mkldnn::memory( - {data_md, mkldnn_engine}, platform::to_void_cast(src_data))); - dev_ctx.SetBlob(key_src_mem, p_src_mem); - - auto p_dst_mem = std::make_shared(mkldnn::memory( - {data_md, mkldnn_engine}, platform::to_void_cast(dst_data))); - dev_ctx.SetBlob(key_dst_mem, p_dst_mem); - - auto fwd_desc = mkldnn::eltwise_forward::desc( - mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); - auto p_fwd_pd = std::make_shared( - fwd_desc, mkldnn_engine); - const std::string key_fwd_pd = key + "eltwise_fwd_pd"; - dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd); - p_fwd = std::make_shared( - *p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get())); + // create mkldnn memory for input X + auto src_md = platform::MKLDNNMemDesc( + src_tz, platform::MKLDNNGetDataType(), src_format); + auto src_memory = std::shared_ptr( + new memory({src_md, mkldnn_engine}, to_void_cast(x_data))); + // save src_memory to be referred in backward path + dev_ctx.SetBlob(key_src_mem, src_memory); + + // create primitive descriptor for activation forward and save it + auto forward_desc = mkldnn::eltwise_forward::desc( + mkldnn::prop_kind::forward_training, algorithm, + src_memory->get_primitive_desc().desc(), alpha, beta); + auto forward_pd = std::make_shared( + forward_desc, mkldnn_engine); + + // save prim desc into global device context to be referred in backward path + dev_ctx.SetBlob(key_fwd_pd, forward_pd); + + // create mkldnn memory for output y + dst_memory = + std::make_shared(forward_pd->dst_primitive_desc(), y_data); + + dev_ctx.SetBlob(key_dst_mem, dst_memory); + + // create activation primitive + p_fwd = std::make_shared(*forward_pd, *src_memory, + *dst_memory); dev_ctx.SetBlob(key_fwd, p_fwd); } else { // primitives already exist - auto p_src_mem = + auto src_memory = std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); - PADDLE_ENFORCE(p_src_mem != nullptr, - "Fail to find eltwise p_src_mem in device context."); - auto p_dst_mem = + PADDLE_ENFORCE(src_memory != nullptr, + "Fail to find eltwise src_memory in device context."); + dst_memory = std::static_pointer_cast(dev_ctx.GetBlob(key_dst_mem)); - PADDLE_ENFORCE(p_dst_mem != nullptr, - "Fail to find eltwise p_src_mem in device context."); + PADDLE_ENFORCE(dst_memory != nullptr, + "Fail to find eltwise dst_memory in device context."); - p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data)); - p_dst_mem->set_data_handle(dst_data); + src_memory->set_data_handle(platform::to_void_cast(x_data)); + dst_memory->set_data_handle(y_data); } // push primitive to stream and wait until it's executed - std::vector pipeline = {*(p_fwd.get())}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline; + pipeline.push_back(*p_fwd); + stream(stream::kind::eager).submit(pipeline).wait(); + + y->set_layout(DataLayout::kMKLDNN); + y->set_format(GetMKLDNNFormat(*dst_memory)); } -template -void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, - const T alpha = 0, const T beta = 0) { +template +void eltwise_grad(const framework::ExecutionContext &ctx, + mkldnn::algorithm algorithm, const T alpha = 0, + const T beta = 0) { auto &dev_ctx = ctx.template device_context(); const auto &mkldnn_engine = dev_ctx.GetEngine(); - // get buffers - const auto *out = ctx.template Input("Out"); - - auto *dout = ctx.template Input(framework::GradVarName("Out")); - const auto *diff_dst = dout->template data(); + const auto *diff_y = ctx.Input(framework::GradVarName("Out")); + auto *diff_x = ctx.Output(framework::GradVarName("X")); - auto *dx = - ctx.template Output(framework::GradVarName("X")); - const T *diff_src = dx->template mutable_data(ctx.GetPlace()); + const T *diff_y_data = diff_y->data(); + T *diff_x_data = diff_x->mutable_data(ctx.GetPlace()); - // get memory dim - std::vector src_tz = framework::vectorize2int(out->dims()); + std::vector diff_dst_tz = framework::vectorize2int(diff_y->dims()); - const std::string key = gethash(src_tz, algorithm); - const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem"; - const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem"; - const std::string key_grad = key + "@eltwise_grad"; + auto diff_y_format = + diff_dst_tz.size() == 2 ? mkldnn::memory::format::nc : diff_y->format(); + const std::string key = gethash(diff_dst_tz, algorithm); const std::string key_src_data = key + ctx.op().Input("Out") + "@eltwise_fwd_src_data"; + const std::string key_src_layout = + key + ctx.op().Input("Out") + "@eltwise_fwd_src_layout"; + const auto p_src_layout = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_layout)); + const std::string key_src_mem = + key + std::to_string(*p_src_layout) + "@eltwise_fwd_src_mem"; + const std::string key_fwd_pd = + key + std::to_string(*p_src_layout) + "@eltwise_fwd_pd"; + const std::string key_with_layouts = + key + std::to_string(*p_src_layout) + "-" + std::to_string(diff_y_format); + const std::string key_diff_src_mem = + key_with_layouts + "@eltwise_diff_src_mem"; + const std::string key_diff_dst_mem = + key_with_layouts + "@eltwise_diff_dst_mem"; + const std::string key_grad = key_with_layouts + "@eltwise_grad"; + const auto p_src_data = std::static_pointer_cast(dev_ctx.GetBlob(key_src_data)); - const std::string key_src_mem = key + "@eltwise_fwd_src_mem"; - auto p_src_mem = + auto src_memory = std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); - p_src_mem->set_data_handle(*p_src_data.get()); + PADDLE_ENFORCE(src_memory != nullptr, + "Fail to find src_memory in device context"); + src_memory->set_data_handle(*p_src_data.get()); + + std::shared_ptr diff_src_memory; - auto p_grad = std::static_pointer_cast( + auto p_grad = std::static_pointer_cast( dev_ctx.GetBlob(key_grad)); if (p_grad == nullptr) { - // create memory description - auto data_md = src_tz.size() == 2 - ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nc) - : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // create memory primitives - std::shared_ptr p_diff_src_mem = - std::make_shared(mkldnn::memory( - {data_md, mkldnn_engine}, platform::to_void_cast(diff_src))); - dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem); - std::shared_ptr p_diff_dst_mem = - std::make_shared(mkldnn::memory( - {data_md, mkldnn_engine}, platform::to_void_cast(diff_dst))); - dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem); - - auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, - alpha, beta); - - const std::string key_fwd_pd = key + "eltwise_fwd_pd"; - auto *p_fwd_pd = static_cast( - dev_ctx.GetBlob(key_fwd_pd).get()); - - auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( - bwd_desc, mkldnn_engine, *p_fwd_pd); - + // create mkldnn memory for input diff_y + auto diff_dst_md = platform::MKLDNNMemDesc( + diff_dst_tz, platform::MKLDNNGetDataType(), diff_y_format); + auto diff_dst_memory = std::shared_ptr( + new memory({diff_dst_md, mkldnn_engine}, to_void_cast(diff_y_data))); + dev_ctx.SetBlob(key_diff_dst_mem, diff_dst_memory); + + // retrieve eltwise primitive desc from device context + auto forward_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_fwd_pd)); + PADDLE_ENFORCE(forward_pd != nullptr, + "Fail to find eltwise_fwd_pd in device context"); + + // ceate primitive descriptor for activation backward + auto backward_desc = mkldnn::eltwise_backward::desc( + algorithm, diff_dst_memory->get_primitive_desc().desc(), + src_memory->get_primitive_desc().desc(), alpha, beta); + auto backward_pd = mkldnn::eltwise_backward::primitive_desc( + backward_desc, mkldnn_engine, *forward_pd); + + // create mkldnn memory for output diff_src + diff_src_memory = std::make_shared( + backward_pd.diff_src_primitive_desc(), diff_x_data); + dev_ctx.SetBlob(key_diff_src_mem, diff_src_memory); + + // create activation backward primitive p_grad = std::make_shared( - eltwise_bwd_prim_desc, *static_cast(p_src_mem.get()), - *(static_cast(p_diff_dst_mem.get())), - *(static_cast(p_diff_src_mem.get()))); + backward_pd, *src_memory, *diff_dst_memory, *diff_src_memory); + dev_ctx.SetBlob(key_grad, p_grad); } else { // primitives already exist - auto p_diff_src_mem = std::static_pointer_cast( + diff_src_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_diff_src_mem)); - auto p_diff_dst_mem = std::static_pointer_cast( + auto diff_dst_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_diff_dst_mem)); - p_diff_src_mem->set_data_handle( - platform::to_void_reinterpret_cast(diff_src)); - p_diff_dst_mem->set_data_handle( - platform::to_void_reinterpret_cast(diff_dst)); + diff_src_memory->set_data_handle( + platform::to_void_reinterpret_cast(diff_x_data)); + diff_dst_memory->set_data_handle( + platform::to_void_reinterpret_cast(diff_y_data)); } // push primitive to stream and wait until it's executed - std::vector pipeline = {*(p_grad.get())}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline; + pipeline.push_back(*p_grad); + stream(stream::kind::eager).submit(pipeline).wait(); + + diff_x->set_layout(DataLayout::kMKLDNN); + diff_x->set_format(GetMKLDNNFormat(*diff_src_memory)); } -} // anonymous namespace template struct MKLDNNActivationFunc : public BaseActivationFunctor { - template - void operator()(const ExecContext &ctx) const { + void operator()(const framework::ExecutionContext &ctx) const { eltwise_forward(ctx, algorithm); } }; template struct MKLDNNActivationGradFunc : public BaseActivationFunctor { - template - void operator()(const ExecContext &ctx) const { + void operator()(const framework::ExecutionContext &ctx) const { eltwise_grad(ctx, algorithm); } }; template -using ReluMkldnnFunctor = +using ReluMKLDNNFunctor = MKLDNNActivationFunc; template -using TanhMkldnnFunctor = +using TanhMKLDNNFunctor = MKLDNNActivationFunc; template -using SqrtMkldnnFunctor = +using SqrtMKLDNNFunctor = MKLDNNActivationFunc; template -using AbsMkldnnFunctor = +using AbsMKLDNNFunctor = MKLDNNActivationFunc; template -using ReluMkldnnGradFunctor = +using ReluMKLDNNGradFunctor = MKLDNNActivationGradFunc; template -using TanhMkldnnGradFunctor = +using TanhMKLDNNGradFunctor = MKLDNNActivationGradFunc; template -using SqrtMkldnnGradFunctor = +using SqrtMKLDNNGradFunctor = MKLDNNActivationGradFunc; template -using AbsMkldnnGradFunctor = +using AbsMKLDNNGradFunctor = MKLDNNActivationGradFunc; } // namespace operators } // namespace paddle @@ -265,9 +341,9 @@ namespace ops = paddle::operators; ops::MKLDNNActivationGradKernel>); #define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro) \ - __macro(relu, ReluMkldnnFunctor, ReluMkldnnGradFunctor); \ - __macro(tanh, TanhMkldnnFunctor, TanhMkldnnGradFunctor); \ - __macro(sqrt, SqrtMkldnnFunctor, SqrtMkldnnGradFunctor); \ - __macro(abs, AbsMkldnnFunctor, AbsMkldnnGradFunctor); + __macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor); \ + __macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor); \ + __macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor); \ + __macro(abs, AbsMKLDNNFunctor, AbsMKLDNNGradFunctor); FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL); diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index dd71c66a75a039429f6e4b1771bb31508bb6b56d..286b03d7b7d11a50f33f0190c1a5b9097ed0f4a2 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -19,13 +19,15 @@ limitations under the License. */ namespace paddle { namespace operators { +using paddle::framework::Tensor; + #define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \ class OP_NAME##OpMaker \ : public ::paddle::framework::OpProtoAndCheckerMaker { \ public: \ void Make() override { \ - AddInput("X", "Input of " #OP_NAME "operator"); \ - AddOutput("Out", "Output of" #OP_NAME "operator"); \ + AddInput("X", "Input of " #OP_NAME " operator"); \ + AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \ AddAttr("use_mkldnn", \ "(bool, default false) Only used in mkldnn kernel") \ .SetDefault(false); \ @@ -58,14 +60,15 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx, const framework::OperatorWithKernel& oper, const std::string& name) { framework::LibraryType library{framework::LibraryType::kPlain}; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN auto it = oper.Attrs().find("use_mkldnn"); if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() && platform::CanMKLDNNBeUsed(ctx)) { library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif - framework::DataLayout layout = framework::DataLayout::kAnyLayout; return framework::OpKernelType( framework::ToDataType(ctx.Input(name)->type()), ctx.GetPlace(), layout, library); @@ -80,6 +83,7 @@ class ActivationOp : public framework::OperatorWithKernel { ctx->ShareLoD("X", /*->*/ "Out"); } + protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return GetKernelType(ctx, *this, "X"); @@ -94,6 +98,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); } + protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return GetKernelType(ctx, *this, "Out"); @@ -110,7 +115,7 @@ $$out = \frac{1}{1 + e^{-x}}$$ __attribute__((unused)) constexpr char LogSigmoidDoc[] = R"DOC( Logsigmoid Activation Operator -$$out = \log \frac{1}{1 + e^{-x}}$$ +$$out = \\log \\frac{1}{1 + e^{-x}}$$ )DOC"; @@ -131,14 +136,14 @@ $out = \max(x, 0)$ __attribute__((unused)) constexpr char TanhDoc[] = R"DOC( Tanh Activation Operator. -$$out = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ +$$out = \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ )DOC"; __attribute__((unused)) constexpr char TanhShrinkDoc[] = R"DOC( TanhShrink Activation Operator. -$$out = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ +$$out = x - \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ )DOC"; @@ -194,7 +199,7 @@ $out = [x]$ __attribute__((unused)) constexpr char ReciprocalDoc[] = R"DOC( Reciprocal Activation Operator. -$$out = \frac{1}{x}$$ +$$out = \\frac{1}{x}$$ )DOC"; @@ -250,15 +255,14 @@ class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "Output of Softshrink operator"); AddAttr("lambda", "non-negative offset").SetDefault(0.5f); AddComment(R"DOC( -Softshrink Activation Operator. +:strong:`Softshrink Activation Operator` -$$ -out = \begin{cases} - x - \lambda, \text{if } x > \lambda \\ - x + \lambda, \text{if } x < -\lambda \\ - 0, \text{otherwise} - \end{cases} -$$ +.. math:: + out = \begin{cases} + x - \lambda, \text{if } x > \lambda \\ + x + \lambda, \text{if } x < -\lambda \\ + 0, \text{otherwise} + \end{cases} )DOC"); } @@ -269,18 +273,18 @@ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "Input of HardShrink operator"); AddOutput("Out", "Output of HardShrink operator"); - AddAttr("threshold", "The value of threshold for HardShrink") + AddAttr("threshold", + "The value of threshold for HardShrink. [default: 0.5]") .SetDefault(0.5f); AddComment(R"DOC( -HardShrink Activation Operator. +:strong:`HardShrink activation operator` -$$ -out = \begin{cases} - x, \text{if } x > \lambda \\ - x, \text{if } x < -\lambda \\ - 0, \text{otherwise} - \end{cases} -$$ +.. math:: + out = \begin{cases} + x, \text{if } x > \lambda \\ + x, \text{if } x < -\lambda \\ + 0, \text{otherwise} + \end{cases} )DOC"); } @@ -381,7 +385,7 @@ class STanhOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( STanh Activation Operator. -$$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ +$$out = b * \\frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ )DOC"); } @@ -392,18 +396,18 @@ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "Input of ThresholdedRelu operator"); AddOutput("Out", "Output of ThresholdedRelu operator"); - AddAttr("threshold", "The threshold location of activation") + AddAttr("threshold", + "The threshold location of activation. [default 1.0].") .SetDefault(1.0f); AddComment(R"DOC( -ThresholdedRelu Activation Operator. +:strong:`ThresholdedRelu activation operator` -$$ -out = \begin{cases} - x, \text{if } x > threshold \\ - 0, \text{otherwise} - \end{cases} -$$ +.. math:: + out = \begin{cases} + x, \text{if } x > threshold \\ + 0, \text{otherwise} + \end{cases} )DOC"); } }; @@ -442,7 +446,7 @@ class SwishOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Swish Activation Operator. -$$out = \frac{x}{1 + e^{- \beta x}}$$ +$$out = \\frac{x}{1 + e^{- \beta x}}$$ )DOC"); } diff --git a/paddle/fluid/operators/adam_op.cc b/paddle/fluid/operators/adam_op.cc index 99b0239855d6241b064a5883c2be3d58078b3b61..5d670fe3b9d99a31a628ff707ff860564eca952e 100644 --- a/paddle/fluid/operators/adam_op.cc +++ b/paddle/fluid/operators/adam_op.cc @@ -56,9 +56,12 @@ class AdamOp : public framework::OperatorWithKernel { "Beta2 power accumulator should have 1 dimension"); auto param_dims = ctx->GetInputDim("Param"); - PADDLE_ENFORCE_EQ( - param_dims, ctx->GetInputDim("Grad"), - "Param and Grad input of AdamOp should have same dimension"); + if (ctx->GetInputsVarType("Grad")[0] == + framework::proto::VarType::LOD_TENSOR) { + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Grad"), + "Param and Grad input of AdamOp should have same dimension"); + } PADDLE_ENFORCE_EQ( param_dims, ctx->GetInputDim("Moment1"), "Param and Moment1 input of AdamOp should have same dimension"); @@ -89,9 +92,9 @@ class AdamOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator"); AddInput("Beta2Pow", "(Tensor) Input beta2 power accumulator"); - AddOutput("ParamOut", "(Tensor) Output parameter"); - AddOutput("Moment1Out", "(Tensor) Output first moment"); - AddOutput("Moment2Out", "(Tensor) Output second moment"); + AddOutput("ParamOut", "(Tensor) Output parameter").Reuse("Param"); + AddOutput("Moment1Out", "(Tensor) Output first moment").Reuse("Moment1"); + AddOutput("Moment2Out", "(Tensor) Output second moment").Reuse("Moment2"); AddAttr("beta1", "(float, default 0.9) " diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h index f82ff47b52490c354f383515d430d14e24cbf6af..84a584f424823a450effd4c36e9da600f5851da2 100644 --- a/paddle/fluid/operators/adam_op.h +++ b/paddle/fluid/operators/adam_op.h @@ -282,6 +282,10 @@ class AdamOpKernel : public framework::OpKernel { } else if (grad_var->IsType()) { auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); + if (grad.rows().size() == 0) { + VLOG(3) << "grad row size is 0!!"; + return; + } // merge duplicated rows if any. scatter::MergeAdd merge_func; auto grad_merge = @@ -289,11 +293,18 @@ class AdamOpKernel : public framework::OpKernel { auto& grad_tensor = grad_merge.value(); const T* grad_data = grad_tensor.template data(); int64_t* rows = nullptr; +// When compiled without CUDA, the CUDAMutableData() interface should not be +// provided. +#if defined(PADDLE_WITH_CUDA) if (platform::is_gpu_place(ctx.GetPlace())) { rows = grad_merge.mutable_rows()->CUDAMutableData(ctx.GetPlace()); } else { +#endif rows = grad_merge.mutable_rows()->data(); + +#if defined(PADDLE_WITH_CUDA) } +#endif auto row_numel = grad_tensor.numel() / grad_merge.rows().size(); SparseAdamFunctor functor( diff --git a/paddle/fluid/operators/arg_max_op.cc b/paddle/fluid/operators/arg_max_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8174d3735859b1fac40cd4c07545f34874d31ab7 --- /dev/null +++ b/paddle/fluid/operators/arg_max_op.cc @@ -0,0 +1,33 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OPERATOR(arg_max, paddle::operators::ArgMinMaxOp, + paddle::operators::ArgMaxOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL( + arg_max, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel); diff --git a/paddle/fluid/operators/arg_max_op.cu b/paddle/fluid/operators/arg_max_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..a147d77a9e9c577984028e1a6ed9582dda622069 --- /dev/null +++ b/paddle/fluid/operators/arg_max_op.cu @@ -0,0 +1,31 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OP_CUDA_KERNEL( + arg_max, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel); diff --git a/paddle/fluid/operators/arg_min_max_op_base.h b/paddle/fluid/operators/arg_min_max_op_base.h new file mode 100644 index 0000000000000000000000000000000000000000..6cbdaefeda099c36a864289ef8195c20d09c55e6 --- /dev/null +++ b/paddle/fluid/operators/arg_min_max_op_base.h @@ -0,0 +1,160 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/string/printf.h" + +namespace paddle { +namespace operators { + +enum ArgMinMaxType { kArgMin, kArgMax }; + +template +struct ArgMinMaxFunctor {}; + +#define DECLARE_ARG_MIN_MAX_FUNCTOR(eigen_op_type, enum_argminmax_value) \ + template \ + struct ArgMinMaxFunctor { \ + void operator()(const DeviceContext& ctx, const framework::LoDTensor& in, \ + framework::LoDTensor* out, int64_t axis) { \ + auto in_eigen = framework::EigenTensor::From(in); \ + auto out_eigen = framework::EigenTensor::From(*out); \ + out_eigen.device(*(ctx.eigen_device())) = \ + in_eigen.eigen_op_type(axis).template cast(); \ + } \ + } + +DECLARE_ARG_MIN_MAX_FUNCTOR(argmin, ArgMinMaxType::kArgMin); +DECLARE_ARG_MIN_MAX_FUNCTOR(argmax, ArgMinMaxType::kArgMax); + +template +class ArgMinMaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& x = *(ctx.Input("X")); + auto& out = *(ctx.Output("Out")); + out.mutable_data(ctx.GetPlace()); + auto axis = ctx.Attr("axis"); + auto& dev_ctx = ctx.template device_context(); + +#define CALL_ARG_MINMAX_FUNCTOR(rank) \ + ArgMinMaxFunctor \ + functor##rank; \ + functor##rank(dev_ctx, x, &out, axis) + + switch (x.dims().size()) { + case 1: + CALL_ARG_MINMAX_FUNCTOR(1); + break; + case 2: + CALL_ARG_MINMAX_FUNCTOR(2); + break; + case 3: + CALL_ARG_MINMAX_FUNCTOR(3); + break; + case 4: + CALL_ARG_MINMAX_FUNCTOR(4); + break; + case 5: + CALL_ARG_MINMAX_FUNCTOR(5); + break; + case 6: + CALL_ARG_MINMAX_FUNCTOR(6); + break; + default: + PADDLE_THROW( + "%s operator doesn't supports tensors whose ranks are greater " + "than 6.", + (EnumArgMinMaxValue == kArgMin ? "argmin" : "argmax")); + break; +#undef CALL_ARG_MINMAX_FUNCTOR + } + } +}; + +template +using ArgMinKernel = + ArgMinMaxKernel; + +template +using ArgMaxKernel = + ArgMinMaxKernel; + +class ArgMinMaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + const auto& x_dims = ctx->GetInputDim("X"); + int64_t axis = ctx->Attrs().Get("axis"); + PADDLE_ENFORCE(axis >= -x_dims.size() && axis < x_dims.size(), + "'axis' must be inside [-Rank(X), Rank(X))"); + + auto x_rank = x_dims.size(); + if (axis < 0) axis += x_rank; + + std::vector vec; + for (int64_t i = 0; i < axis; i++) vec.push_back(x_dims[i]); + for (int64_t i = axis + 1; i < x_rank; i++) vec.push_back(x_dims[i]); + ctx->SetOutputDim("Out", framework::make_ddim(vec)); + } +}; + +class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker { + protected: + virtual const char* OpName() const = 0; + virtual const char* Name() const = 0; + + public: + void Make() override { + AddInput("X", "Input tensor."); + AddOutput("Out", "Output tensor."); + AddAttr("axis", "The axis in which to compute the arg indics."); + AddComment(string::Sprintf(R"DOC( + %s Operator. + + Computes the indices of the %s elements of the input tensor's element + along the provided axis. +)DOC", + OpName(), Name())); + } +}; + +class ArgMinOpMaker : public BaseArgMinMaxOpMaker { + protected: + const char* OpName() const override { return "ArgMin"; } + const char* Name() const override { return "min"; } +}; + +class ArgMaxOpMaker : public BaseArgMinMaxOpMaker { + protected: + const char* OpName() const override { return "ArgMax"; } + const char* Name() const override { return "max"; } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/arg_min_op.cc b/paddle/fluid/operators/arg_min_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..41f188029f17dbe8717afc0ca0760a39edc24b54 --- /dev/null +++ b/paddle/fluid/operators/arg_min_op.cc @@ -0,0 +1,33 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OPERATOR(arg_min, paddle::operators::ArgMinMaxOp, + paddle::operators::ArgMinOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL( + arg_min, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel); diff --git a/paddle/fluid/operators/arg_min_op.cu b/paddle/fluid/operators/arg_min_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..4d020508505a6ebac8be41ce1e4f99d436b67ab5 --- /dev/null +++ b/paddle/fluid/operators/arg_min_op.cu @@ -0,0 +1,31 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OP_CUDA_KERNEL( + arg_min, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel); diff --git a/paddle/fluid/operators/argsort_op.cc b/paddle/fluid/operators/argsort_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a2f5a2545701991263c1ef842e9275b1edbfd2ca --- /dev/null +++ b/paddle/fluid/operators/argsort_op.cc @@ -0,0 +1,87 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/argsort_op.h" + +namespace paddle { +namespace operators { + +class ArgsortOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ArgsortOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ArgsortOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Indices"), + "Output(Indices) of ArgsortOp should not be null."); + + auto in_dims = ctx->GetInputDim("X"); + int axis = ctx->Attrs().Get("axis"); + + auto num_dims = in_dims.size(); + PADDLE_ENFORCE(axis < num_dims, + "Attr(axis) %d of ArgsortOp is out of bounds for Input(X)'s " + "rank %d.", + axis, num_dims); + PADDLE_ENFORCE(axis >= -num_dims, + "Attr(axis) %d of ArgsortOp must be not less than " + "-rank(Input(X)) (%d).", + axis, num_dims); + + ctx->SetOutputDim("Out", in_dims); + ctx->SetOutputDim("Indices", in_dims); + ctx->ShareLoD("X", "Out"); + ctx->ShareLoD("X", "Indices"); + } +}; + +class ArgsortOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) The input of Argsort op."); + AddOutput("Out", + "(Tensor) The sorted tensor of Argsort op, with the same " + "shape as Input(X)."); + AddOutput("Indices", + "(Tensor) The indices of a tensor giving the sorted order, with " + "the same shape as Input(X)."); + AddComment(R"DOC( +Argsort operator + +Performs sorting on the input tensor along the given axis and outputs two +tensors, Output(Out) and Output(Indices). They reserve the same shape +with Input(X), and Output(Out) represents the sorted tensor while +Output(Indices) gives the sorted order along the given axis Attr(axis). + + )DOC"); + AddAttr("axis", + "(int, default -1) The axis along which to sort the tensor. " + "When axis < 0, the actual axis will be the |axis|'th " + "counting backwards. Default -1, the last dimension.") + .SetDefault(-1); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(argsort, ops::ArgsortOp, ops::ArgsortOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(argsort, + ops::ArgsortKernel, + ops::ArgsortKernel); diff --git a/paddle/fluid/operators/argsort_op.cu b/paddle/fluid/operators/argsort_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..7d5199aae7da4eed5afa6b8bd64c04a540b915d4 --- /dev/null +++ b/paddle/fluid/operators/argsort_op.cu @@ -0,0 +1,151 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/argsort_op.h" +#include "paddle/fluid/platform/assert.h" +#include "paddle/fluid/platform/cuda_device_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using platform::PADDLE_CUDA_NUM_THREADS; + +const int kMaxRank = 9; // The max rank of a tensor allowed in Fluid + +__global__ void ComputeTargetIdx(const int64_t* in_dims, int dims_size, + int axis, int64_t n, int64_t* trg_idx, + int64_t* med_ids) { + int64_t index = threadIdx.x + blockDim.x * blockIdx.x; + if (index < n) { + int64_t shape_out_axis[kMaxRank - 1] = {0}; + int64_t dims_out_axis[kMaxRank - 1] = {0}; + int64_t tmp = index; + int64_t pos_in_axis = 0; + int64_t i = dims_size - 2; + int64_t dim_axis = 0; + for (int64_t j = dims_size - 1; j >= 0; --j) { + int64_t dim = in_dims[j]; + if (j != axis) { + shape_out_axis[i] = tmp % dim; + dims_out_axis[i] = dim; + i--; + } else { + dim_axis = dim; + pos_in_axis = tmp % dim_axis; + } + tmp /= dim; + } + int64_t group = (dims_size > 1) ? shape_out_axis[0] : 0; + for (int64_t j = 0; j < dims_size - 2; ++j) { + group = group * dims_out_axis[j + 1] + shape_out_axis[j + 1]; + } + + int64_t traget_idx = group * dim_axis + pos_in_axis; + trg_idx[index] = traget_idx; + med_ids[traget_idx] = pos_in_axis; + } +} + +template +__global__ void PermuteInData(const T* in, const int64_t* trg_idx, int64_t n, + T* med_out) { + int index = threadIdx.x + blockDim.x * blockIdx.x; + if (index < n) { + med_out[trg_idx[index]] = in[index]; + } +} + +template +__global__ void Sort(int64_t axis_dim, int64_t groups, T* med_out, + int64_t* med_ids) { + int index = threadIdx.x + blockDim.x * blockIdx.x; + if (index < groups) { + thrust::sort_by_key(thrust::device, med_out + index * axis_dim, + med_out + axis_dim * (1 + index), + med_ids + index * axis_dim); + } +} + +template +__global__ void PermuteMediateData(const T* med_out, const int64_t* med_ids, + const int64_t* trg_idx, int64_t n, T* out, + int64_t* indices) { + int index = threadIdx.x + blockDim.x * blockIdx.x; + if (index < n) { + out[index] = med_out[trg_idx[index]]; + indices[index] = med_ids[trg_idx[index]]; + } +} + +template +class ArgsortOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* indices = ctx.Output("Indices"); + int axis = ctx.Attr("axis"); + + auto in_dims = input->dims(); + axis = (axis < 0) ? (in_dims.size() + axis) : axis; + + const T* in_data = input->data(); + T* out_data = output->mutable_data(ctx.GetPlace()); + int64_t* ids_data = indices->mutable_data(ctx.GetPlace()); + + int64_t numel = input->numel(); + int64_t groups = numel / in_dims[axis]; + + std::vector in_dims_vec = vectorize(in_dims); + thrust::device_vector in_dims_dev(in_dims_vec.begin(), + in_dims_vec.end()); + int64_t* in_dims_data = thrust::raw_pointer_cast(in_dims_dev.data()); + // Mediate tensor for sorting data and indices + Tensor mediate_output, mediate_indices; + T* med_out_data = + mediate_output.mutable_data(input->dims(), ctx.GetPlace()); + int64_t* med_ids_data = + mediate_indices.mutable_data(in_dims, ctx.GetPlace()); + // Target index of each element along the given axis in the mediate tensors + Tensor trg_idx_t; + int64_t* trg_idx = trg_idx_t.mutable_data(in_dims, ctx.GetPlace()); + + auto stream = ctx.cuda_device_context().stream(); + const int num_threads = PADDLE_CUDA_NUM_THREADS; + + ComputeTargetIdx<<<(numel - 1) / num_threads + 1, num_threads, 0, stream>>>( + in_dims_data, in_dims.size(), axis, numel, trg_idx, med_ids_data); + + PermuteInData<<<(numel - 1) / num_threads + 1, num_threads, 0, stream>>>( + in_data, trg_idx, numel, med_out_data); + + Sort<<<(groups - 1) / num_threads + 1, num_threads, 0, stream>>>( + in_dims[axis], groups, med_out_data, med_ids_data); + + PermuteMediateData<<<(numel - 1) / num_threads + 1, num_threads, 0, + stream>>>(med_out_data, med_ids_data, trg_idx, numel, + out_data, ids_data); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_CUDA_KERNEL(argsort, paddle::operators::ArgsortOpCUDAKernel, + paddle::operators::ArgsortOpCUDAKernel); diff --git a/paddle/fluid/operators/argsort_op.h b/paddle/fluid/operators/argsort_op.h new file mode 100644 index 0000000000000000000000000000000000000000..7e9112cfb7cbe5f783b04729fb4dff3676c922bc --- /dev/null +++ b/paddle/fluid/operators/argsort_op.h @@ -0,0 +1,81 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class ArgsortKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* indices = ctx.Output("Indices"); + int axis = ctx.Attr("axis"); + + auto in_dims = input->dims(); + axis = (axis < 0) ? (in_dims.size() + axis) : axis; + + const T* in_data = input->data(); + T* out_data = output->mutable_data(ctx.GetPlace()); + int64_t* ids_data = indices->mutable_data(ctx.GetPlace()); + + int64_t groups = input->numel() / in_dims[axis]; + int64_t stride = (axis == in_dims.size() - 1) + ? 1 + : framework::product(framework::slice_ddim( + in_dims, axis + 1, in_dims.size())); + + for (int64_t i = 0; i < groups; ++i) { + int64_t idx = i; + std::vector shape_vec(in_dims.size(), 0); + for (int64_t dim = in_dims.size() - 1; dim >= 0; --dim) { + if (dim != axis) { + shape_vec[dim] = idx % in_dims[dim]; + idx /= in_dims[dim]; + } + } + + int64_t start_index = shape_vec[0]; + for (int64_t dim = 0; dim < in_dims.size() - 1; ++dim) { + start_index = start_index * in_dims[dim + 1] + shape_vec[dim + 1]; + } + + std::vector org_index_vec(in_dims[axis], start_index); + for (int64_t j = 1; j < in_dims[axis]; ++j) { + org_index_vec[j] += j * stride; + } + + std::sort(org_index_vec.begin(), org_index_vec.end(), + [in_data](const int64_t v1, const int64_t v2) { + return in_data[v1] < in_data[v2]; + }); + + for (size_t j = 0; j < org_index_vec.size(); ++j) { + int64_t index = start_index + j * stride; + out_data[index] = in_data[org_index_vec[j]]; + ids_data[index] = (org_index_vec[j] - start_index) / stride; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/assign_value_op.cc b/paddle/fluid/operators/assign_value_op.cc index 4ad6f3443db33fd14b67091d14fd877b951730ff..a757916be7f6ece9b783d51d1051aac6a276795b 100644 --- a/paddle/fluid/operators/assign_value_op.cc +++ b/paddle/fluid/operators/assign_value_op.cc @@ -70,6 +70,7 @@ $$Out = values$$ namespace ops = paddle::operators; -REGISTER_OPERATOR(assign_value, ops::AssignValueOp, ops::AssignValueOpMaker); +REGISTER_OPERATOR(assign_value, ops::AssignValueOp, ops::AssignValueOpMaker, + paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL(assign_value, ops::AssignValueKernel, ops::AssignValueKernel); diff --git a/paddle/fluid/operators/auc_op.cc b/paddle/fluid/operators/auc_op.cc index c9871a9fe6b3b0d0cf671c2d155715f92c94fd8f..5edecd18e673da326ec119cf9a383f24f8045089 100644 --- a/paddle/fluid/operators/auc_op.cc +++ b/paddle/fluid/operators/auc_op.cc @@ -24,26 +24,34 @@ class AucOp : public framework::OperatorWithKernel { protected: void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Out"), "Input of Out should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Indices"), - "Input of Indices should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Predict"), + "Input of Out should not be null."); PADDLE_ENFORCE(ctx->HasInput("Label"), "Input of Label should not be null."); - auto inference_height = ctx->GetInputDim("Out")[0]; + auto predict_width = ctx->GetInputDim("Predict")[1]; + PADDLE_ENFORCE_EQ(predict_width, 2, "Only support binary classification"); + auto predict_height = ctx->GetInputDim("Predict")[0]; auto label_height = ctx->GetInputDim("Label")[0]; - PADDLE_ENFORCE_EQ(inference_height, label_height, + PADDLE_ENFORCE_EQ(predict_height, label_height, "Out and Label should have same height."); + int num_thres = ctx->Attrs().Get("num_thresholds"); + ctx->SetOutputDim("AUC", {1}); - ctx->ShareLoD("Out", /*->*/ "AUC"); + ctx->SetOutputDim("TPOut", {num_thres}); + ctx->SetOutputDim("TNOut", {num_thres}); + ctx->SetOutputDim("FPOut", {num_thres}); + ctx->SetOutputDim("FNOut", {num_thres}); + + ctx->ShareLoD("Predict", /*->*/ "AUC"); } protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("Out")->type()), + framework::ToDataType(ctx.Input("Predict")->type()), ctx.device_context()); } }; @@ -51,22 +59,25 @@ class AucOp : public framework::OperatorWithKernel { class AucOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Out", - "A floating point 2D tensor, values are in the range [0, 1]." - "Each row is sorted in descending order. This input should be the" - "output of topk." + AddInput("Predict", + "A floating point 2D tensor with shape [batch_size, 2], values " + "are in the range [0, 1]." "Typically, this tensor indicates the probability of each label"); - AddInput("Indices", - "An int 2D tensor, indicating the indices of original" - "tensor before sorting. Typically, this tensor indicates which " - "label the probability stands for."); AddInput("Label", - "A 2D int tensor indicating the label of the training data." - "The height is batch size and width is always 1."); + "A 2D int tensor indicating the label of the training data. " + "shape: [batch_size, 1]"); + AddInput("TP", "True-Positive value."); + AddInput("FP", "False-Positive value."); + AddInput("TN", "True-Negative value."); + AddInput("FN", "False-Negative value."); // TODO(typhoonzero): support weight input AddOutput("AUC", "A scalar representing the " "current area-under-the-curve."); + AddOutput("TPOut", "True-Positive value."); + AddOutput("FPOut", "False-Positive value."); + AddOutput("TNOut", "True-Negative value."); + AddOutput("FNOut", "False-Negative value."); AddAttr("curve", "Curve type, can be 'ROC' or 'PR'.") .SetDefault("ROC"); diff --git a/paddle/fluid/operators/auc_op.h b/paddle/fluid/operators/auc_op.h index 8b016c3d31ad83e66baeb298c61840cc529efa1e..0a18585edb54a76aff5ae72ecc71e0eebb9f9361 100644 --- a/paddle/fluid/operators/auc_op.h +++ b/paddle/fluid/operators/auc_op.h @@ -31,58 +31,54 @@ template class AucKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - auto* inference = ctx.Input("Out"); + auto* predict = ctx.Input("Predict"); auto* label = ctx.Input("Label"); auto* auc = ctx.Output("AUC"); + // Only use output var for now, make sure it's persistable and + // not cleaned up for each batch. + auto* true_positive = ctx.Output("TPOut"); + auto* false_positive = ctx.Output("FPOut"); + auto* true_negative = ctx.Output("TNOut"); + auto* false_negative = ctx.Output("FNOut"); - float* auc_data = auc->mutable_data(ctx.GetPlace()); + auto* auc_data = auc->mutable_data(ctx.GetPlace()); std::string curve = ctx.Attr("curve"); int num_thresholds = ctx.Attr("num_thresholds"); - std::vector thresholds_list; + std::vector thresholds_list; thresholds_list.reserve(num_thresholds); for (int i = 1; i < num_thresholds - 1; i++) { - thresholds_list[i] = static_cast(i) / (num_thresholds - 1); + thresholds_list[i] = static_cast(i) / (num_thresholds - 1); } - const float kEpsilon = 1e-7; + const double kEpsilon = 1e-7; thresholds_list[0] = 0.0f - kEpsilon; thresholds_list[num_thresholds - 1] = 1.0f + kEpsilon; - size_t batch_size = inference->dims()[0]; - size_t inference_width = inference->dims()[1]; + size_t batch_size = predict->dims()[0]; + size_t inference_width = predict->dims()[1]; - const T* inference_data = inference->data(); - const int64_t* label_data = label->data(); + const T* inference_data = predict->data(); + const auto* label_data = label->data(); - // Create local tensor for storing the curve: TP, FN, TN, FP - // TODO(typhoonzero): use eigen op to caculate these values. - Tensor true_positive, false_positive, true_negative, false_negative; - - true_positive.Resize({num_thresholds}); - false_negative.Resize({num_thresholds}); - true_negative.Resize({num_thresholds}); - false_positive.Resize({num_thresholds}); - - int64_t* tp_data = true_positive.mutable_data(ctx.GetPlace()); - int64_t* fn_data = false_negative.mutable_data(ctx.GetPlace()); - int64_t* tn_data = true_negative.mutable_data(ctx.GetPlace()); - int64_t* fp_data = false_positive.mutable_data(ctx.GetPlace()); + auto* tp_data = true_positive->mutable_data(ctx.GetPlace()); + auto* fn_data = false_negative->mutable_data(ctx.GetPlace()); + auto* tn_data = true_negative->mutable_data(ctx.GetPlace()); + auto* fp_data = false_positive->mutable_data(ctx.GetPlace()); for (int idx_thresh = 0; idx_thresh < num_thresholds; idx_thresh++) { - // caculate TP, FN, TN, FP for current thresh + // calculate TP, FN, TN, FP for current thresh int64_t tp = 0, fn = 0, tn = 0, fp = 0; for (size_t i = 0; i < batch_size; i++) { - // NOTE: label_data used as bool, labels >0 will be treated as true. + // NOTE: label_data used as bool, labels > 0 will be treated as true. if (label_data[i]) { - // use first(max) data in each row - if (inference_data[i * inference_width] >= + if (inference_data[i * inference_width + 1] >= (thresholds_list[idx_thresh])) { tp++; } else { fn++; } } else { - if (inference_data[i * inference_width] >= + if (inference_data[i * inference_width + 1] >= (thresholds_list[idx_thresh])) { fp++; } else { @@ -91,27 +87,27 @@ class AucKernel : public framework::OpKernel { } } // store rates - tp_data[idx_thresh] = tp; - fn_data[idx_thresh] = fn; - tn_data[idx_thresh] = tn; - fp_data[idx_thresh] = fp; + tp_data[idx_thresh] += tp; + fn_data[idx_thresh] += fn; + tn_data[idx_thresh] += tn; + fp_data[idx_thresh] += fp; } // epsilon to avoid divide by zero. - float epsilon = 1e-6; + double epsilon = 1e-6; // Riemann sum to caculate auc. Tensor tp_rate, fp_rate, rec_rate; tp_rate.Resize({num_thresholds}); fp_rate.Resize({num_thresholds}); rec_rate.Resize({num_thresholds}); - float* tp_rate_data = tp_rate.mutable_data(ctx.GetPlace()); - float* fp_rate_data = fp_rate.mutable_data(ctx.GetPlace()); - float* rec_rate_data = rec_rate.mutable_data(ctx.GetPlace()); + auto* tp_rate_data = tp_rate.mutable_data(ctx.GetPlace()); + auto* fp_rate_data = fp_rate.mutable_data(ctx.GetPlace()); + auto* rec_rate_data = rec_rate.mutable_data(ctx.GetPlace()); for (int i = 0; i < num_thresholds; i++) { - tp_rate_data[i] = (static_cast(tp_data[i]) + epsilon) / + tp_rate_data[i] = (static_cast(tp_data[i]) + epsilon) / (tp_data[i] + fn_data[i] + epsilon); fp_rate_data[i] = - static_cast(fp_data[i]) / (fp_data[i] + tn_data[i] + epsilon); - rec_rate_data[i] = (static_cast(tp_data[i]) + epsilon) / + static_cast(fp_data[i]) / (fp_data[i] + tn_data[i] + epsilon); + rec_rate_data[i] = (static_cast(tp_data[i]) + epsilon) / (tp_data[i] + fp_data[i] + epsilon); } *auc_data = 0.0f; diff --git a/paddle/fluid/operators/average_accumulates_op.cc b/paddle/fluid/operators/average_accumulates_op.cc index 25864e95d7e290c7f684501893e99c828c511979..f389eab605e087c535b9918264e6502217062505 100644 --- a/paddle/fluid/operators/average_accumulates_op.cc +++ b/paddle/fluid/operators/average_accumulates_op.cc @@ -19,28 +19,28 @@ namespace operators { template <> void GetAccumulators( - const framework::ExecutionContext& ctx, int64_t* num_updates_, - int64_t* num_accumulates_, int64_t* old_num_accumulates_) { + const framework::ExecutionContext& ctx, int64_t* num_updates, + int64_t* num_accumulates, int64_t* old_num_accumulates) { auto* in_old_num_accumulates = ctx.Input("in_old_num_accumulates"); auto* in_num_accumulates = ctx.Input("in_num_accumulates"); auto* in_num_updates = ctx.Input("in_num_updates"); - *old_num_accumulates_ = in_old_num_accumulates->data()[0]; - *num_accumulates_ = in_num_accumulates->data()[0]; - *num_updates_ = in_num_updates->data()[0]; + *old_num_accumulates = in_old_num_accumulates->data()[0]; + *num_accumulates = in_num_accumulates->data()[0]; + *num_updates = in_num_updates->data()[0]; } template <> void SetAccumulators( - const framework::ExecutionContext& ctx, int64_t num_updates_, - int64_t num_accumulates_, int64_t old_num_accumulates_) { + const framework::ExecutionContext& ctx, int64_t num_updates, + int64_t num_accumulates, int64_t old_num_accumulates) { auto* out_old_num_accumulates = ctx.Output("out_old_num_accumulates"); auto* out_num_accumulates = ctx.Output("out_num_accumulates"); auto* out_num_updates = ctx.Output("out_num_updates"); - out_old_num_accumulates->data()[0] = old_num_accumulates_; - out_num_accumulates->data()[0] = num_accumulates_; - out_num_updates->data()[0] = num_updates_; + out_old_num_accumulates->data()[0] = old_num_accumulates; + out_num_accumulates->data()[0] = num_accumulates; + out_num_updates->data()[0] = num_updates; } class AverageAccumulatesOp : public framework::OperatorWithKernel { @@ -177,7 +177,7 @@ class AverageAccumulatesOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( AverageAccumulates Operator. -Accumulate the sum of parameter whtin sliding window. The size of sliding window is +Accumulate the sum of parameter within sliding window. The size of sliding window is determined by 'average_window', 'max_average_window' and 'min_average_window'. Memory was shared by Input(in_sum_1) and Output(out_sum_1) which acts as an accumulator 'sum_1'. 'sum_2', 'sum_3', 'num_accumulates', 'old_num_accumulates' and 'num_updates' were the same as 'sum_1'. diff --git a/paddle/fluid/operators/average_accumulates_op.h b/paddle/fluid/operators/average_accumulates_op.h index 07ac5ced11605f6d0d5164d1c0f69acbd7bbed60..3958d3f685470f2505abf0e8bfd269d3834970ae 100644 --- a/paddle/fluid/operators/average_accumulates_op.h +++ b/paddle/fluid/operators/average_accumulates_op.h @@ -54,8 +54,9 @@ class AverageAccumulatesKernel : public framework::OpKernel { float average_window = ctx.Attr("average_window"); int64_t max_average_window = ctx.Attr("max_average_window"); int64_t min_average_window = ctx.Attr("min_average_window"); - min_average_window = - std::min(min_average_window, max_average_window); + PADDLE_ENFORCE_LE(min_average_window, max_average_window, + "min_average_window shouldn't be larger than " + "max_average_window"); // Get inputs auto* param = ctx.Input("param"); diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc index 0e4a56d4a45a732cfcf43b09228bc0c44df5924c..9ab2179b5fe689762704039c5f67dd080e530aa5 100644 --- a/paddle/fluid/operators/batch_norm_mkldnn_op.cc +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -19,22 +19,15 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; +using batch_norm_bwd = mkldnn::batch_normalization_backward; +using batch_norm_fwd = mkldnn::batch_normalization_forward; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; using paddle::platform::MKLDNNDeviceContext; using paddle::platform::MKLDNNMemDesc; -using mkldnn::memory; - -template -using EigenArrayMap = - Eigen::Map>; -template -using ConstEigenArrayMap = - Eigen::Map>; -template -using EigenVectorArrayMap = Eigen::Map>; -template -using ConstEigenVectorArrayMap = - Eigen::Map>; +using platform::to_void_cast; namespace { template @@ -64,24 +57,16 @@ void run_batch_norm_op(Args &&... args) { mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } -template -inline void *cast_const_to_void(const T *t) { - return static_cast(const_cast(t)); -} } // namespace template class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { - auto data_layout_str = ctx.Attr("data_layout"); - auto data_layout = framework::StringToDataLayout(data_layout_str); - PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, - "MKLDNN batch normalization handles only NCHW data layout"); - const float epsilon = ctx.Attr("epsilon"); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); + const bool fuse_with_relu = ctx.Attr("fuse_with_relu"); const auto *x = ctx.Input("X"); const auto *mean = ctx.Input("Mean"); @@ -99,41 +84,57 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { const auto *scale = ctx.Input("Scale"); const auto *shift = ctx.Input("Bias"); - y->mutable_data(ctx.GetPlace()); - mean_out->mutable_data(ctx.GetPlace()); - variance_out->mutable_data(ctx.GetPlace()); + PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN && + x->format() != memory::format::format_undef, + "Wrong layout/format set for Input x tensor"); + + const T *x_data = x->data(); + const T *mean_data = mean->data(); + const T *variance_data = variance->data(); + T *y_data = y->mutable_data(ctx.GetPlace()); + T *mean_out_data = mean_out->mutable_data(ctx.GetPlace()); + T *variance_out_data = variance_out->mutable_data(ctx.GetPlace()); + T *batch_mean_data = nullptr; + T *batch_variance_data = nullptr; if (!is_test) { - batch_mean->mutable_data(ctx.GetPlace()); - batch_variance->mutable_data(ctx.GetPlace()); + batch_mean_data = batch_mean->mutable_data(ctx.GetPlace()); + batch_variance_data = batch_variance->mutable_data(ctx.GetPlace()); } auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring : mkldnn::prop_kind::forward_training; - auto dims = paddle::framework::vectorize2int(x->dims()); - - auto src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - - auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; - auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine}; - - auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data())}; - auto dst = mkldnn::memory{dst_pd, y->data()}; + auto src_tz = paddle::framework::vectorize2int(x->dims()); + auto scale_tz = paddle::framework::vectorize2int(scale->dims()); + PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1"); + const unsigned int ic = scale_tz[0]; unsigned flags = mkldnn::use_scale_shift; if (is_test) flags |= mkldnn::use_global_stats; + if (fuse_with_relu) flags |= mkldnn::fuse_bn_relu; + + // create mkldnn memory from input x tensor + mkldnn::memory::format input_format = + platform::MKLDNNFormatForSize(src_tz.size(), x->format()); + auto src_memory = memory( + {{{src_tz}, memory::data_type::f32, input_format}, mkldnn_engine}, + to_void_cast(x_data)); + + // create primitive descriptor for batch norm forward using bn_fwd_types = bn_type_traits; - auto batch_norm_fwd_desc = - bn_fwd_types::op_desc{propagation, src_md, epsilon, flags}; - auto batch_norm_fwd_pd = - bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ + propagation, src_memory.get_primitive_desc().desc(), epsilon, flags}; + std::shared_ptr batch_norm_fwd_pd = + std::shared_ptr( + new batch_norm_fwd::primitive_desc(batch_norm_fwd_desc, + mkldnn_engine)); - const unsigned int ic = dims[1]; + // Save the pd to be used in backward pass + const std::string key = ctx.op().Output("SavedMean"); + const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; + dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd); // MKLDNN requires a single piece of memory for scale and shift/bias data const size_t scaleshift_size = 2 * ic; @@ -143,73 +144,58 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { copy_to_weights(scale->data(), scale->data() + ic, shift->data(), shift->data() + ic, &scaleshift_data); - auto scaleshift_memory = mkldnn::memory{ - batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + // crate mkldnn memory for weights(scale/shift) + auto scaleshift_memory = memory(batch_norm_fwd_pd->weights_primitive_desc(), + scaleshift_data.data()); - if (is_test) { - auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), - cast_const_to_void(mean->data())}; + // create mkldnn memory for output y tensor + auto dst_memory = memory(batch_norm_fwd_pd->dst_primitive_desc(), y_data); + if (is_test) { + // create mkldnn memory for stats (as input) + auto mean_memory = memory(batch_norm_fwd_pd->mean_primitive_desc(), + to_void_cast(mean_data)); auto variance_memory = - mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), - cast_const_to_void(variance->data())}; + memory(batch_norm_fwd_pd->variance_primitive_desc(), + to_void_cast(variance_data)); run_batch_norm_op( - batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory, + *batch_norm_fwd_pd, src_memory, + (const mkldnn::primitive::at &)mean_memory, (const mkldnn::primitive::at &)variance_memory, scaleshift_memory, - dst); + dst_memory); } else { + // create mkldnn memory for stats (as output) auto mean_memory = - mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), - cast_const_to_void(batch_mean->data())}; + memory(batch_norm_fwd_pd->mean_primitive_desc(), batch_mean_data); + auto variance_memory = memory( + batch_norm_fwd_pd->variance_primitive_desc(), batch_variance_data); - auto variance_memory = - mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), - cast_const_to_void(batch_variance->data())}; - - run_batch_norm_op(batch_norm_fwd_pd, src, - scaleshift_memory, dst, + run_batch_norm_op(*batch_norm_fwd_pd, src_memory, + scaleshift_memory, dst_memory, mean_memory, variance_memory); } if (!is_test) { - const unsigned int in = dims[0]; - const unsigned int sample_size = x->numel() / in / ic; - - // saved_xx is use just in this batch of data - EigenVectorArrayMap saved_mean_e( - batch_mean->mutable_data(ctx.GetPlace()), ic); - EigenVectorArrayMap saved_variance_e( - batch_variance->mutable_data(ctx.GetPlace()), ic); - saved_mean_e.setZero(); - saved_variance_e.setZero(); - - const unsigned int x_arr_size = in * ic; - ConstEigenArrayMap x_arr(x->data(), sample_size, x_arr_size); - for (unsigned int nc = 0; nc < x_arr_size; ++nc) { - saved_mean_e(nc % ic) += x_arr.col(nc).sum(); - } - saved_mean_e /= in * sample_size; - for (unsigned int nc = 0; nc < x_arr_size; ++nc) { - saved_variance_e(nc % ic) += - (x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm(); - } - saved_variance_e /= in * sample_size; - - ConstEigenVectorArrayMap mean_arr{mean->data(), ic}; - ConstEigenVectorArrayMap variance_arr{variance->data(), ic}; - - EigenVectorArrayMap running_mean_arr( - mean_out->mutable_data(ctx.GetPlace()), ic); - EigenVectorArrayMap running_var_arr( - variance_out->mutable_data(ctx.GetPlace()), ic); + // mkldnn only compute stats for current batch + // so we need compute momentum stats via Eigen lib + EigenVectorArrayMap batch_mean_e(batch_mean_data, ic); + EigenVectorArrayMap batch_variance_e(batch_variance_data, ic); + ConstEigenVectorArrayMap mean_e(mean_data, ic); + ConstEigenVectorArrayMap variance_e{variance_data, ic}; + + EigenVectorArrayMap running_mean_e(mean_out_data, ic); + EigenVectorArrayMap running_variance_e(variance_out_data, ic); auto one_minus_momentum = 1. - momentum; - running_mean_arr = - mean_arr * momentum + saved_mean_e * one_minus_momentum; - running_var_arr = - variance_arr * momentum + saved_variance_e * one_minus_momentum; + running_mean_e = mean_e * momentum + batch_mean_e * one_minus_momentum; + running_variance_e = + variance_e * momentum + batch_variance_e * one_minus_momentum; } + + y->set_layout(DataLayout::kMKLDNN); + y->set_format( + (memory::format)dst_memory.get_primitive_desc().desc().data.format); } }; @@ -217,11 +203,6 @@ template class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { public: void Compute(const paddle::framework::ExecutionContext &ctx) const override { - auto data_layout_str = ctx.Attr("data_layout"); - auto data_layout = framework::StringToDataLayout(data_layout_str); - PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, - "MKLDNN batch normalization handles only NCHW data layout"); - auto &dev_ctx = ctx.template device_context(); auto mkldnn_engine = dev_ctx.GetEngine(); @@ -238,88 +219,138 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto *diff_scale = ctx.Output(framework::GradVarName("Scale")); auto *diff_shift = ctx.Output(framework::GradVarName("Bias")); - diff_x->mutable_data(ctx.GetPlace()); - diff_scale->mutable_data(ctx.GetPlace()); - diff_shift->mutable_data(ctx.GetPlace()); - - auto dims = paddle::framework::vectorize2int(x->dims()); - unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats; - - auto src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto diff_src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto diff_dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN && + diff_y->format() != memory::format::format_undef, + "Wrong layout/format set for Input diff_y tensor"); + + const T *x_data = x->data(); + const T *diff_y_data = diff_y->data(); + const T *batch_mean_data = batch_mean->data(); + const T *batch_variance_data = batch_variance->data(); + const T *scale_data = scale->data(); + const T *shift_data = shift->data(); + T *diff_x_data = diff_x->mutable_data(ctx.GetPlace()); + T *diff_scale_data = diff_scale->mutable_data(ctx.GetPlace()); + T *diff_shift_data = diff_shift->mutable_data(ctx.GetPlace()); + + auto src_tz = paddle::framework::vectorize2int(x->dims()); + auto diff_src_tz = src_tz; + auto dst_tz = src_tz; + auto diff_dst_tz = dst_tz; + auto scale_tz = paddle::framework::vectorize2int(scale->dims()); + PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1"); + + const unsigned int ic = scale_tz[0]; + + // Retrieve bn_fwd_pd from device context + const std::string key = ctx.op().Input("SavedMean"); + const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; + auto batch_norm_fwd_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_batch_norm_fwd_pd)); + PADDLE_ENFORCE(batch_norm_fwd_pd != nullptr, + "Fail to find batch_norm_fwd_pd in device context"); using bn_bwd_types = bn_type_traits; - using bn_fwd_types = bn_type_traits; - auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ - mkldnn::prop_kind::forward_training, src_md, epsilon, flags}; - auto batch_norm_fwd_pd = - bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + // create mkldnn memory from input diff_y tensor - auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ - mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags}; - auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ - batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd}; + mkldnn::memory::format dst_format = + platform::MKLDNNFormatForSize(src_tz.size(), diff_y->format()); - auto src = mkldnn::memory{{src_md, mkldnn_engine}, - cast_const_to_void(x->data())}; + auto user_diff_dst_memory = memory( + {{{diff_dst_tz}, memory::data_type::f32, dst_format}, mkldnn_engine}, + to_void_cast(diff_y_data)); - auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(), - cast_const_to_void(batch_mean->data())}; + // create mkldnn memory from input x tensor + mkldnn::memory::format input_format = + platform::MKLDNNFormatForSize(src_tz.size(), x->format()); - auto variance = - mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(), - cast_const_to_void(batch_variance->data())}; + auto src_memory = memory( + {{{src_tz}, memory::data_type::f32, input_format}, mkldnn_engine}, + to_void_cast(x_data)); - auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine}, - cast_const_to_void(diff_y->data())}; + // for diff_dst, try to use same format as dst in forward pass + auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc(); + auto diff_dst_md = diff_dst_pd.desc(); - const unsigned int ic = dims[1]; + // create primitive descriptor for batch norm backward + unsigned flags = mkldnn::use_scale_shift; + auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ + mkldnn::prop_kind::backward, diff_dst_md, + src_memory.get_primitive_desc().desc(), epsilon, flags}; + auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ + batch_norm_bwd_desc, mkldnn_engine, *batch_norm_fwd_pd}; + + // reorder user_diff_dst if it's not in preferred format + auto diff_dst_memory = user_diff_dst_memory; + primitive reorder_diff_dst; + bool is_diff_dst_reordered = false; + if (diff_dst_pd != user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = memory(diff_dst_pd); + reorder_diff_dst = reorder(user_diff_dst_memory, diff_dst_memory); + is_diff_dst_reordered = true; + } + + // create mkldnn memory for input tensors (src/mean/variance) + auto mean_memory = memory(batch_norm_bwd_pd.mean_primitive_desc(), + to_void_cast(batch_mean_data)); + auto variance_memory = memory(batch_norm_bwd_pd.variance_primitive_desc(), + to_void_cast(batch_variance_data)); + // MKLDNN requires a single piece of memory for scale and shift/bias data const size_t scaleshift_size = 2 * ic; std::vector scaleshift_data; scaleshift_data.reserve(scaleshift_size); - copy_to_weights(scale->data(), scale->data() + ic, shift->data(), - shift->data() + ic, &scaleshift_data); + copy_to_weights(scale_data, scale_data + ic, shift_data, shift_data + ic, + &scaleshift_data); - auto scaleshift_memory = mkldnn::memory{ - batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + // create mkldnn memory for input tensors (scale/shift) + auto scaleshift_memory = memory(batch_norm_bwd_pd.weights_primitive_desc(), + scaleshift_data.data()); + // create mkldnn memory for output diff weights (combined scale/shift) std::vector diff_scaleshift_data; diff_scaleshift_data.reserve(scaleshift_size); - copy_to_weights(diff_scale->data(), diff_scale->data() + ic, - diff_shift->data(), diff_shift->data() + ic, - &diff_scaleshift_data); - auto diff_scaleshift_memory = - mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(), - diff_scaleshift_data.data()}; - - auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine}, - static_cast(diff_x->data())}; - - run_batch_norm_op( - batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory, - diff_src, diff_scaleshift_memory); - + memory(batch_norm_bwd_pd.diff_weights_primitive_desc(), + diff_scaleshift_data.data()); + + // here assume diff_src is in the same format of src + auto diff_src_memory = memory(src_memory.get_primitive_desc(), diff_x_data); + + // finally create batch_norm backward primitive + auto batch_norm_bwd_prim = + batch_norm_bwd(batch_norm_bwd_pd, src_memory, mean_memory, + variance_memory, diff_dst_memory, scaleshift_memory, + diff_src_memory, diff_scaleshift_memory); + + // execute optional reorder and batch_norm backward primitive + std::vector pipeline; + if (is_diff_dst_reordered) pipeline.push_back(reorder_diff_dst); + pipeline.push_back(batch_norm_bwd_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + // copy back diff sacle/shift to output tensors (diff scale/shift) + diff_scaleshift_data.resize(scaleshift_size); auto it = std::begin(diff_scaleshift_data); - std::copy(it, std::next(it, ic), diff_scale->data()); + std::copy(it, std::next(it, ic), diff_scale_data); std::copy(std::next(it, ic), std::end(diff_scaleshift_data), - diff_shift->data()); + diff_shift_data); + + // set layout/format of output tensors + diff_x->set_layout(DataLayout::kMKLDNN); + diff_x->set_format((memory::format)diff_src_memory.get_primitive_desc() + .desc() + .data.format); } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace, +REGISTER_OP_KERNEL(batch_norm, MKLDNN, ::paddle::platform::CPUPlace, ops::BatchNormMKLDNNOpKernel); -REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace, +REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, ::paddle::platform::CPUPlace, ops::BatchNormMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 6ec8c9d18b466142acdb46b0f46826a2aca7a47e..5912a1a17cbd29c3ebd83f37133c044f0905c8bd 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -22,22 +22,6 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; -using LoDTensor = framework::LoDTensor; -using DataLayout = framework::DataLayout; - -template -using EigenArrayMap = - Eigen::Map>; -template -using ConstEigenArrayMap = - Eigen::Map>; -template -using EigenVectorArrayMap = Eigen::Map>; -template -using ConstEigenVectorArrayMap = - Eigen::Map>; - class BatchNormOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -110,17 +94,19 @@ class BatchNormOp : public framework::OperatorWithKernel { ctx.Input("Variance")->type()), "Variance input should be of float type"); - framework::LibraryType library_{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, - library_); + library); } }; @@ -149,13 +135,15 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Variance", "The global variance (for training) " "or estimated Variance (for testing)"); - AddOutput("Y", "result after normalization"); + AddOutput("Y", "result after normalization").Reuse("X"); AddOutput("MeanOut", "Share memory with Mean. " - "Store the global mean when training"); + "Store the global mean when training") + .Reuse("Mean"); AddOutput("VarianceOut", "Share memory with Variance. " - "Store the global Variance when training"); + "Store the global Variance when training") + .Reuse("Variance"); AddOutput("SavedMean", "Mean of the current mini batch, " "will apply to output when training") @@ -167,6 +155,9 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { AddAttr("use_mkldnn", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); + AddAttr("fuse_with_relu", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Batch Normalization. @@ -225,6 +216,18 @@ class BatchNormKernel saved_mean_e.setZero(); saved_variance_e.setZero(); + EigenVectorArrayMap running_mean_arr( + mean_out->mutable_data(ctx.GetPlace()), C); + EigenVectorArrayMap running_var_arr( + variance_out->mutable_data(ctx.GetPlace()), C); + + if ((N * sample_size) == 1) { + LOG(WARNING) << "Only 1 element in normalization dimension, " + << "we skip the batch norm calculation, let y = x."; + framework::TensorCopySync(*x, ctx.GetPlace(), y); + return; + } + switch (data_layout) { case DataLayout::kNCHW: { ConstEigenArrayMap x_arr(x->data(), sample_size, N * C); @@ -256,10 +259,6 @@ class BatchNormKernel PADDLE_THROW("Unknown storage order: %s", data_layout_str); } - EigenVectorArrayMap running_mean_arr( - mean_out->mutable_data(ctx.GetPlace()), C); - EigenVectorArrayMap running_var_arr( - variance_out->mutable_data(ctx.GetPlace()), C); running_mean_arr = running_mean_arr * momentum + saved_mean_e * (1. - momentum); running_var_arr = @@ -366,18 +365,21 @@ class BatchNormGradOp : public framework::OperatorWithKernel { PADDLE_THROW("can't find Y@GRAD"); } - framework::LibraryType library_{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout, library_); + layout, library); } }; @@ -433,6 +435,11 @@ class BatchNormGradKernel d_bias_arr.setZero(); d_scale_arr.setZero(); + if ((N * sample_size) == 1) { + framework::TensorCopySync(*d_y, ctx.GetPlace(), d_x); + return; + } + const auto scale_inv_var_nhw = scale_arr * inv_var_arr / (N * sample_size); switch (data_layout) { diff --git a/paddle/fluid/operators/batch_norm_op.cu.cc b/paddle/fluid/operators/batch_norm_op.cu.cc index 550dd32d36767f90e880415bfffaf01aeb623609..ca6cd8669352fd5814f25a04433ca97fe4abe9ff 100644 --- a/paddle/fluid/operators/batch_norm_op.cu.cc +++ b/paddle/fluid/operators/batch_norm_op.cu.cc @@ -72,6 +72,9 @@ class BatchNormKernel int N, C, H, W, D; ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D); + auto *y = ctx.Output("Y"); + y->mutable_data(ctx.GetPlace()); + // ------------------- cudnn descriptors --------------------- cudnnTensorDescriptor_t data_desc_; cudnnTensorDescriptor_t bn_param_desc_; @@ -93,7 +96,7 @@ class BatchNormKernel mode_ = CUDNN_BATCHNORM_SPATIAL; #endif - VLOG(1) << "Setting descriptors."; + VLOG(3) << "Setting descriptors."; std::vector dims; std::vector strides; if (data_layout == DataLayout::kNCHW) { @@ -113,11 +116,6 @@ class BatchNormKernel const auto *scale = ctx.Input("Scale"); const auto *bias = ctx.Input("Bias"); - auto *y = ctx.Output("Y"); - - // alloc memory - y->mutable_data(ctx.GetPlace()); - auto &dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); @@ -162,22 +160,28 @@ class BatchNormKernel functor(dev_ctx, saved_mean, static_cast>(0)); functor(dev_ctx, saved_variance, static_cast>(0)); - double this_factor = 1. - momentum; - - CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationForwardTraining( - handle, mode_, CudnnDataType::kOne(), CudnnDataType::kZero(), - data_desc_, x->template data(), data_desc_, - y->template mutable_data(ctx.GetPlace()), bn_param_desc_, - scale->template data>(), - bias->template data>(), this_factor, - mean_out->template mutable_data>( - ctx.GetPlace()), - variance_out->template mutable_data>( - ctx.GetPlace()), - epsilon, saved_mean->template mutable_data>( - ctx.GetPlace()), - saved_variance->template mutable_data>( - ctx.GetPlace()))); + if ((N * H * W * D) == 1) { + LOG(WARNING) << "Only 1 element in normalization dimension, " + << "we skip the batch norm calculation, let y = x."; + framework::TensorCopySync(*x, ctx.GetPlace(), y); + } else { + double this_factor = 1. - momentum; + + CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationForwardTraining( + handle, mode_, CudnnDataType::kOne(), CudnnDataType::kZero(), + data_desc_, x->template data(), data_desc_, + y->template mutable_data(ctx.GetPlace()), bn_param_desc_, + scale->template data>(), + bias->template data>(), this_factor, + mean_out->template mutable_data>( + ctx.GetPlace()), + variance_out->template mutable_data>( + ctx.GetPlace()), + epsilon, saved_mean->template mutable_data>( + ctx.GetPlace()), + saved_variance->template mutable_data>( + ctx.GetPlace()))); + } } // clean when exit. @@ -209,6 +213,25 @@ class BatchNormGradKernel int N, C, H, W, D; ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D); + // init output + auto *d_x = ctx.Output(framework::GradVarName("X")); + auto *d_scale = ctx.Output(framework::GradVarName("Scale")); + auto *d_bias = ctx.Output(framework::GradVarName("Bias")); + + d_x->mutable_data(ctx.GetPlace()); + d_scale->mutable_data(ctx.GetPlace()); + d_bias->mutable_data(ctx.GetPlace()); + + auto &dev_ctx = ctx.template device_context(); + if ((N * H * W * D) == 1) { + framework::TensorCopySync(*d_y, ctx.GetPlace(), d_x); + math::SetConstant> + functor; + functor(dev_ctx, d_scale, static_cast>(0)); + functor(dev_ctx, d_bias, static_cast>(0)); + return; + } + PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL); PADDLE_ENFORCE_EQ(scale->dims()[0], C); @@ -247,21 +270,11 @@ class BatchNormGradKernel CUDNN_ENFORCE(platform::dynload::cudnnDeriveBNTensorDescriptor( bn_param_desc_, data_desc_, mode_)); - // init output - auto *d_x = ctx.Output(framework::GradVarName("X")); - auto *d_scale = ctx.Output(framework::GradVarName("Scale")); - auto *d_bias = ctx.Output(framework::GradVarName("Bias")); - - d_x->mutable_data(ctx.GetPlace()); - d_scale->mutable_data(ctx.GetPlace()); - d_bias->mutable_data(ctx.GetPlace()); - const auto *saved_mean = ctx.Input("SavedMean"); const auto *saved_var = ctx.Input("SavedVariance"); const void *saved_mean_data = saved_mean->template data(); const void *saved_var_data = saved_var->template data(); - auto &dev_ctx = ctx.template device_context(); CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward( dev_ctx.cudnn_handle(), mode_, CudnnDataType::kOne(), CudnnDataType::kZero(), CudnnDataType::kOne(), diff --git a/paddle/fluid/operators/batch_norm_op.h b/paddle/fluid/operators/batch_norm_op.h index 9e5fc41598f29336074335f3624a2300ad018d09..5e3d630d6889e445c5e84fa836d2d81bb7266779 100644 --- a/paddle/fluid/operators/batch_norm_op.h +++ b/paddle/fluid/operators/batch_norm_op.h @@ -19,6 +19,22 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; +using DataLayout = framework::DataLayout; + +template +using EigenArrayMap = + Eigen::Map>; +template +using ConstEigenArrayMap = + Eigen::Map>; +template +using EigenVectorArrayMap = Eigen::Map>; +template +using ConstEigenVectorArrayMap = + Eigen::Map>; + template class BatchNormKernel : public framework::OpKernel { public: diff --git a/paddle/fluid/operators/batch_size_like.h b/paddle/fluid/operators/batch_size_like.h index 483c9f8c2191fa4eb98b91112f9d6753e2fbddc3..fc15d56891cf7af10a91ca22a09c84fa2e52d465 100644 --- a/paddle/fluid/operators/batch_size_like.h +++ b/paddle/fluid/operators/batch_size_like.h @@ -54,18 +54,18 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel { class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() final { - AddInput("Input", - "(Tensor) Tensor " - "whose input_dim_idx'th dimension specifies the batch_size"); + AddInput( + "Input", + "Tensor whose input_dim_idx'th dimension specifies the batch_size"); AddOutput("Out", - "(Tensor) Tensor of specified shape will be filled " + "Tensor of specified shape will be filled " "with the specified value"); - AddAttr>("shape", "(vector) The shape of the output"); + AddAttr>("shape", "The shape of the output"); AddAttr("input_dim_idx", - "(int, default 0) The index of input's batch size dimension") + "default 0. The index of input's batch size dimension") .SetDefault(0); AddAttr("output_dim_idx", - "(int, default 0) The index of output's batch size dimension") + "default 0. The index of output's batch size dimension") .SetDefault(0); Apply(); } diff --git a/paddle/fluid/operators/beam_search_decode_op.cc b/paddle/fluid/operators/beam_search_decode_op.cc index c3dd22119ddab8ecf9213ee274e4cbd4f05e78fd..10d678111f5325e495b24286e6ecf651230393fe 100644 --- a/paddle/fluid/operators/beam_search_decode_op.cc +++ b/paddle/fluid/operators/beam_search_decode_op.cc @@ -12,8 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/beam_search_decode_op.h" +#include #include + +#include "paddle/fluid/operators/beam_search_decode_op.h" #include "paddle/fluid/platform/device_context.h" namespace paddle { @@ -22,8 +24,11 @@ namespace operators { struct BeamSearchDecodeFunctor { BeamSearchDecodeFunctor(const LoDTensorArray& step_ids, const LoDTensorArray& step_scores, - LoDTensor* id_tensor, LoDTensor* score_tensor) - : step_ids_origin_(step_ids), + LoDTensor* id_tensor, LoDTensor* score_tensor, + size_t beam_size, int end_id) + : beam_size_(beam_size), + end_id_(end_id), + step_ids_origin_(step_ids), step_scores_origin_(step_scores), id_tensor_(id_tensor), score_tensor_(score_tensor) { @@ -37,9 +42,11 @@ struct BeamSearchDecodeFunctor { // Copy all tensors in the input tensor array for (auto& step_id : step_ids_origin_) { framework::LoDTensor out; - dev_ctx->Wait(); - framework::TensorCopy(step_id, platform::CPUPlace(), *dev_ctx, &out); - dev_ctx->Wait(); + if (step_id.numel() > 0) { + dev_ctx->Wait(); + framework::TensorCopy(step_id, platform::CPUPlace(), *dev_ctx, &out); + dev_ctx->Wait(); + } out.set_lod(step_id.lod()); step_ids_.push_back(out); @@ -53,9 +60,12 @@ struct BeamSearchDecodeFunctor { // Copy all tensors in the input tensor array for (auto& step_score : step_scores_origin_) { framework::LoDTensor out; - dev_ctx->Wait(); - framework::TensorCopy(step_score, platform::CPUPlace(), *dev_ctx, &out); - dev_ctx->Wait(); + if (step_score.numel() > 0) { + dev_ctx->Wait(); + framework::TensorCopy(step_score, platform::CPUPlace(), *dev_ctx, + &out); + dev_ctx->Wait(); + } out.set_lod(step_score.lod()); step_scores_.push_back(out); @@ -67,6 +77,8 @@ struct BeamSearchDecodeFunctor { void operator()() const; bool tensor_on_gpu_; + size_t beam_size_; + int end_id_; const LoDTensorArray& step_ids_origin_; const LoDTensorArray& step_scores_origin_; LoDTensorArray step_ids_ = LoDTensorArray(); @@ -77,14 +89,14 @@ struct BeamSearchDecodeFunctor { template void BeamSearchDecodeFunctor::operator()() const { - BeamSearchDecoder beam_search_decoder; + BeamSearchDecoder beam_search_decoder(beam_size_, end_id_); // Check if the tensor is on GPU. If so, use the CPU copy instead if (tensor_on_gpu_) { - beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_, - score_tensor_); + beam_search_decoder.Backtrace(step_ids_, step_scores_, id_tensor_, + score_tensor_); } else { - beam_search_decoder.PackAllSteps(step_ids_origin_, step_scores_origin_, - id_tensor_, score_tensor_); + beam_search_decoder.Backtrace(step_ids_origin_, step_scores_origin_, + id_tensor_, score_tensor_); } } @@ -122,13 +134,17 @@ class BeamSearchDecodeOp : public framework::OperatorBase { "Level of LodTensor should be 2"); } + size_t beam_size = ctx.Attr("beam_size"); + int end_id = ctx.Attr("end_id"); + // prepare output LoDTensor* sentenceIds = ctx.Output("SentenceIds"); LoDTensor* sentenceScores = ctx.Output("SentenceScores"); framework::VisitDataType( framework::ToDataType(scores->at(0).type()), - BeamSearchDecodeFunctor(*ids, *scores, sentenceIds, sentenceScores)); + BeamSearchDecodeFunctor(*ids, *scores, sentenceIds, sentenceScores, + beam_size, end_id)); } }; @@ -137,18 +153,32 @@ class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("Ids", "(LodTensorArray)" - "score of the candidate words in each step"); + "The LodTensorArray containing the selected ids of all steps"); AddInput("Scores", "(LodTensorArray)" - "score of the candidate words in each step"); - AddOutput("SentenceIds", - "(LodTensor)" - "All possible result sentences of word ids"); - AddOutput("SentenceScores", - "(LodTensor)" - "All possible result sentences of word scores"); + "The LodTensorArray containing the selected scores of all steps"); + AddOutput( + "SentenceIds", + "(LodTensor)" + "An LodTensor containing all generated id sequences for all source " + "sentences"); + AddOutput( + "SentenceScores", + "(LodTensor)" + "An LodTensor containing scores corresponding to Output(SentenceIds)"); + AddAttr("beam_size", "beam size for beam search"); + AddAttr("end_id", + "the token id which indicates the end of a sequence"); AddComment(R"DOC( -Pack the result of Beam search op into SentenceIds and SentenceScores. +Beam Search Decode Operator. This Operator constructs the full hypotheses for +each source sentence by walking back along the LoDTensorArray Input(ids) +whose lods can be used to restore the path in the beam search tree. + +The Output(SentenceIds) and Output(SentenceScores) separately contain the +generated id sequences and the corresponding scores. The shapes and lods of the +two LodTensor are same. The lod level is 2 and the two levels separately +indicate how many hypotheses each source sentence has and how many ids each +hypothesis has. )DOC"); } }; @@ -172,10 +202,12 @@ class BeamSearchDecodeInferVarType : public framework::VarTypeInference { void operator()(const framework::OpDesc& op_desc, framework::BlockDesc* block) const override { for (auto& o : op_desc.Output("SentenceIds")) { - block->Var(o)->SetType(framework::proto::VarType::LOD_TENSOR); + auto& sentence_ids = block->FindRecursiveOrCreateVar(o); + sentence_ids.SetType(framework::proto::VarType::LOD_TENSOR); } for (auto& o : op_desc.Output("SentenceScores")) { - block->Var(o)->SetType(framework::proto::VarType::LOD_TENSOR); + auto& sentence_scores = block->FindRecursiveOrCreateVar(o); + sentence_scores.SetType(framework::proto::VarType::LOD_TENSOR); } } }; diff --git a/paddle/fluid/operators/beam_search_decode_op.h b/paddle/fluid/operators/beam_search_decode_op.h index 3c01f81c83555b985bb6b7a9e3330ab594a62863..6aefc5446f167eebb0da673b3fbdf7ed128daa98 100644 --- a/paddle/fluid/operators/beam_search_decode_op.h +++ b/paddle/fluid/operators/beam_search_decode_op.h @@ -14,7 +14,9 @@ limitations under the License. */ #pragma once +#include #include + #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/op_registry.h" @@ -25,42 +27,12 @@ using LoDTensor = framework::LoDTensor; using LoDTensorArray = framework::LoDTensorArray; // all the lod have 2 levels. -// The First is source level, the second is sentence level. -// source level describe how many candidate words for this source. -// sentence level describe these candidates belong to which prefix +// The first is source level, the second is sentence level. +// source level describe how many prefixes (branchs) for each source sentece +// (beam). sentence level describe how these candidates belong to the prefixes. const size_t kSourceLevel = 0; const size_t kSentenceLevel = 1; -template -struct BeamNode { - BeamNode(int64_t word_id, T score) : word_id_(word_id), score_(score) {} - - ~BeamNode() { - if (parent_) { - parent_->DropKid(this); - if (parent_->kids_.size() == 0UL) { - delete parent_; - } - } - VLOG(3) << "Delete BeamNode root with word_id:" << this->word_id_; - } - - void AppendTo(BeamNode* parent) { - parent_ = parent; - parent->kids_.insert(this); - } - - void DropKid(BeamNode* kid) { kids_.erase(kid); } - - BeamNode* parent_ = nullptr; - std::unordered_set kids_; - int64_t word_id_; - T score_; -}; - -template -using BeamNodeVector = std::vector>>; - template struct Sentence { std::vector word_ids; @@ -72,24 +44,8 @@ using SentenceVector = std::vector>; template struct BeamSearchDecoder { - /** - * make a BeamNode and all it's related prefix BeanNode into a Sentence. - */ - Sentence MakeSentence(const BeamNode* node) const; - - /** - * Param: - * cur_ids: LoDTensor of One step for word ID - * cur_scores: LoDTensor of One Step for word score - * prefixes_list: prefixes for each source sentence. - * sentence_vector_list: result sentence_vector for each source sentence. - * Return: - * a new prefixes list for each source of current step - */ - std::vector> PackTwoSteps( - const LoDTensor& cur_ids, const LoDTensor& cur_scores, - std::vector>* prefixes_list, - std::vector>* sentence_vector_list) const; + BeamSearchDecoder(size_t beam_size, int end_id) + : beam_size_(beam_size), end_id_(end_id) {} /** * convert the result sentence_vector for each source sentence into two @@ -100,107 +56,30 @@ struct BeamSearchDecoder { * sentence_vector_list: sentence_vector for each source sentence. * id_tensor: result LoDTensor for sentences of id. * score_tensor: result LoDTensor for sentences of score. + * reverse: whether ids of sentence in sentence_vector_list is reversed + * sort_by_score: whether to sort hypotheses of each sentence by scores. */ void ConvertSentenceVectorToLodTensor( std::vector> sentence_vector_list, LoDTensor* id_tensor, - LoDTensor* score_tensor) const; + LoDTensor* score_tensor, bool reverse = true, + bool sort_by_score = true) const; /** - * Pack all steps of id/score LodTensor into sentence LoDTensor - * it's main logic is: - * ```python - * prefix - * result_sentence - * result_lod_tensor - * - * for (step in steps): - * prefix = PackTwoSteps(prefix, step, &result_sentence) - * ConvertSentenceVectorToLodTensor(result_sentence, &result_lod_tensor) - * ``` + * Gather the hypotheses for each source sentence by backtrace though the + * LoDTensorArray step_ids whose lods reserve the path in the tree. */ - void PackAllSteps(const LoDTensorArray& step_ids, - const LoDTensorArray& step_scores, LoDTensor* id_tensor, - LoDTensor* score_tensor) const; -}; - -template -Sentence BeamSearchDecoder::MakeSentence(const BeamNode* node) const { - Sentence sentence; - while (node != nullptr) { - sentence.word_ids.emplace_back(node->word_id_); - sentence.scores.emplace_back(node->score_); - node = node->parent_; - } - - std::reverse(std::begin(sentence.word_ids), std::end(sentence.word_ids)); - std::reverse(std::begin(sentence.scores), std::end(sentence.scores)); - - return sentence; -} - -template -std::vector> BeamSearchDecoder::PackTwoSteps( - const LoDTensor& cur_ids, const LoDTensor& cur_scores, - std::vector>* prefixes_list, - std::vector>* sentence_vector_list) const { - std::vector> result; + void Backtrace(const LoDTensorArray& step_ids, + const LoDTensorArray& step_scores, LoDTensor* id_tensor, + LoDTensor* score_tensor) const; - for (size_t src_idx = 0; src_idx < cur_ids.lod()[kSourceLevel].size() - 1; - ++src_idx) { - size_t src_start = cur_ids.lod().at(kSourceLevel)[src_idx]; - size_t src_end = cur_ids.lod().at(kSourceLevel)[src_idx + 1]; - - BeamNodeVector beam_nodes; - - // if prefixes size is 0, it means this is the first step. In this step, - // all candidate id is the start of candidate sentences. - if (prefixes_list->empty()) { - PADDLE_ENFORCE_EQ(cur_ids.lod().at(kSourceLevel).back(), - cur_ids.lod().at(kSentenceLevel).back(), - "in the first step"); - for (size_t id_idx = src_start; id_idx < src_end; ++id_idx) { - beam_nodes.push_back(std::unique_ptr>(new BeamNode( - cur_ids.data()[id_idx], cur_scores.data()[id_idx]))); - } - } else { - BeamNodeVector& prefixes = prefixes_list->at(src_idx); - SentenceVector& sentence_vector = (*sentence_vector_list)[src_idx]; - - PADDLE_ENFORCE_EQ(src_end - src_start, prefixes.size(), - "prefix and candidate set number should be the same"); - - auto candidate_offset = cur_ids.lod()[kSentenceLevel]; - for (size_t prefix_idx = 0; prefix_idx < prefixes.size(); ++prefix_idx) { - std::unique_ptr>& prefix = prefixes[prefix_idx]; - size_t candidate_start = candidate_offset[src_start + prefix_idx]; - size_t candidate_end = candidate_offset[src_start + prefix_idx + 1]; - if (candidate_start == candidate_end) { - VLOG(3) << "this sentence has no more candidate, " - "add to result sentence and rm it from beam tree"; - sentence_vector.push_back(MakeSentence(prefix.get())); - prefix.reset(); - } else { - for (size_t candidate_idx = candidate_start; - candidate_idx < candidate_end; ++candidate_idx) { - auto* candidate = - new BeamNode(cur_ids.data()[candidate_idx], - cur_scores.data()[candidate_idx]); - candidate->AppendTo(prefix.get()); - beam_nodes.push_back(std::unique_ptr>(candidate)); - } - prefix.release(); - } - } - } - result.push_back(std::move(beam_nodes)); - } - return result; -} + size_t beam_size_; + int end_id_; +}; template void BeamSearchDecoder::ConvertSentenceVectorToLodTensor( std::vector> sentence_vector_list, LoDTensor* id_tensor, - LoDTensor* score_tensor) const { + LoDTensor* score_tensor, bool reverse, bool sort_by_score) const { size_t src_num = sentence_vector_list.size(); PADDLE_ENFORCE_NE(src_num, 0, "src_num should not be 0"); @@ -211,11 +90,29 @@ void BeamSearchDecoder::ConvertSentenceVectorToLodTensor( std::vector score_data; for (size_t src_idx = 0; src_idx < src_num; ++src_idx) { + if (sort_by_score) { + sort(sentence_vector_list[src_idx].begin(), + sentence_vector_list[src_idx].end(), + [reverse](const Sentence& a, const Sentence& b) { + if (reverse) + return a.scores.front() > b.scores.front(); + else + return a.scores.back() > b.scores.back(); + }); + } for (Sentence& sentence : sentence_vector_list[src_idx]) { - id_data.insert(id_data.end(), sentence.word_ids.begin(), - sentence.word_ids.end()); - score_data.insert(score_data.end(), sentence.scores.begin(), - sentence.scores.end()); + if (reverse) { + id_data.insert(id_data.end(), sentence.word_ids.rbegin(), + sentence.word_ids.rend()); + score_data.insert(score_data.end(), sentence.scores.rbegin(), + sentence.scores.rend()); + } else { + id_data.insert(id_data.end(), sentence.word_ids.begin(), + sentence.word_ids.end()); + score_data.insert(score_data.end(), sentence.scores.begin(), + sentence.scores.end()); + } + sentence_level_lod.push_back(sentence_level_lod.back() + sentence.word_ids.size()); } @@ -243,39 +140,75 @@ void BeamSearchDecoder::ConvertSentenceVectorToLodTensor( } template -void BeamSearchDecoder::PackAllSteps(const LoDTensorArray& step_ids, - const LoDTensorArray& step_scores, - LoDTensor* id_tensor, - LoDTensor* score_tensor) const { +void BeamSearchDecoder::Backtrace(const LoDTensorArray& step_ids, + const LoDTensorArray& step_scores, + LoDTensor* id_tensor, + LoDTensor* score_tensor) const { PADDLE_ENFORCE(!step_ids.empty(), "step num should be larger than 0"); PADDLE_ENFORCE_EQ(step_ids.size(), step_scores.size(), "step_ids and step_scores should be the same"); const size_t step_num = step_ids.size(); const size_t src_num = step_ids.at(0).lod().at(kSourceLevel).size() - 1; + std::vector> sentence_vector_list( + src_num, SentenceVector(beam_size_)); + std::vector> prefix_idx_vector_list(src_num); + for (int step_id = step_num - 1; step_id >= 0; --step_id) { + auto& cur_ids = step_ids.at(step_id); + auto& cur_scores = step_scores.at(step_id); + for (size_t src_idx = 0; src_idx < src_num; ++src_idx) { + // for each source sentence + auto& sentence_vector = sentence_vector_list.at(src_idx); + auto& prefix_idx_vector = prefix_idx_vector_list.at(src_idx); + size_t src_prefix_start = cur_ids.lod().at(kSourceLevel)[src_idx]; + size_t src_prefix_end = cur_ids.lod().at(kSourceLevel)[src_idx + 1]; + if (prefix_idx_vector.empty()) { // be finished and pruned at this step + // or the last time step + for (size_t prefix_idx = src_prefix_start; prefix_idx < src_prefix_end; + ++prefix_idx) { + size_t candidate_start = cur_ids.lod().at(kSentenceLevel)[prefix_idx]; + size_t candidate_end = + cur_ids.lod().at(kSentenceLevel)[prefix_idx + 1]; + for (size_t candidate_idx = candidate_start; + candidate_idx < candidate_end; ++candidate_idx) { + prefix_idx_vector.push_back(prefix_idx); + size_t idx = prefix_idx_vector.size() - 1; + auto cur_id = cur_ids.data()[candidate_idx]; + auto cur_score = cur_scores.data()[candidate_idx]; + sentence_vector.at(idx).word_ids.push_back(cur_id); + sentence_vector.at(idx).scores.push_back(cur_score); + } + } + } else { // use prefix_idx_vector to backtrace + size_t src_candidate_start = + cur_ids.lod().at(kSentenceLevel)[src_prefix_start]; + size_t prefix_idx = src_prefix_start; + size_t candidate_num = + cur_ids.lod().at(kSentenceLevel)[prefix_idx + 1] - + cur_ids.lod().at(kSentenceLevel)[prefix_idx]; + for (size_t idx = 0; idx < prefix_idx_vector.size(); ++idx) { + auto candidate_idx = prefix_idx_vector.at(idx); + auto cur_id = cur_ids.data()[candidate_idx]; + auto cur_score = cur_scores.data()[candidate_idx]; + if (cur_id != end_id_ || sentence_vector.at(idx).word_ids.empty()) { + // to skip redundant end tokens + sentence_vector.at(idx).word_ids.push_back(cur_id); + sentence_vector.at(idx).scores.push_back(cur_score); + } - PADDLE_ENFORCE_GT(src_num, 0UL, "source num should be larger than 0"); - - // previous prefixes for each step, - // the init length is 0, means this is the first step. - std::vector> beamnode_vector_list(0); - std::vector> sentence_vector_list(src_num); - - // pack all steps for one batch first, then another batch - for (size_t step_id = 0; step_id < step_num; ++step_id) { - beamnode_vector_list = - PackTwoSteps(step_ids.at(step_id), step_scores.at(step_id), - &beamnode_vector_list, &sentence_vector_list); - } - // append last beam_node to result - for (size_t src_idx = 0; src_idx < src_num; ++src_idx) { - for (auto& beam_node : beamnode_vector_list.at(src_idx)) { - sentence_vector_list[src_idx].push_back(MakeSentence(beam_node.get())); - beam_node.reset(); + while (src_candidate_start + candidate_num <= + candidate_idx) { // search the corresponding prefix + prefix_idx++; + candidate_num += cur_ids.lod().at(kSentenceLevel)[prefix_idx + 1] - + cur_ids.lod().at(kSentenceLevel)[prefix_idx]; + } + prefix_idx_vector.at(idx) = prefix_idx; + } + } } } ConvertSentenceVectorToLodTensor(sentence_vector_list, id_tensor, - score_tensor); + score_tensor, true, true); } } // namespace operators diff --git a/paddle/fluid/operators/beam_search_decode_op_test.cc b/paddle/fluid/operators/beam_search_decode_op_test.cc index 36f9594969c416c694928811012baf94332bbd91..88339e38d89db3f79abf232d6b0d035b759739a6 100644 --- a/paddle/fluid/operators/beam_search_decode_op_test.cc +++ b/paddle/fluid/operators/beam_search_decode_op_test.cc @@ -20,15 +20,11 @@ using LoD = paddle::framework::LoD; using LoDTensor = paddle::framework::LoDTensor; using LoDTensorArray = paddle::framework::LoDTensorArray; -template -using BeamNode = paddle::operators::BeamNode; template using BeamSearchDecoder = paddle::operators::BeamSearchDecoder; template using Sentence = paddle::operators::Sentence; template -using BeamNodeVector = paddle::operators::BeamNodeVector; -template using SentenceVector = paddle::operators::SentenceVector; namespace paddle { @@ -77,138 +73,50 @@ void GenerateExample(const std::vector& level_0, } // namespace test } // namespace paddle -TEST(BeamSearchDecodeOp, DeleteBeamNode) { - auto* root = new BeamNode(0, 0); - auto* b1 = new BeamNode(1, 1); - auto* b2 = new BeamNode(2, 2); - auto* b3 = new BeamNode(3, 3); - - b1->AppendTo(root); - b2->AppendTo(root); - b3->AppendTo(b1); - - delete b3; - delete b2; -} - -TEST(BeamSearchDecodeOp, MakeSentence) { - auto* root = new BeamNode(0, 0); - auto* b1 = new BeamNode(1, 1); - auto* end = new BeamNode(2, 2); - b1->AppendTo(root); - end->AppendTo(b1); - - BeamSearchDecoder helper; - Sentence sentence = helper.MakeSentence(end); - delete end; - - std::vector expect_ids = {0, 1, 2}; - ASSERT_EQ(sentence.word_ids, expect_ids); - - std::vector expect_scores = {0, 1, 2}; - ASSERT_EQ(sentence.scores, expect_scores); -} - -TEST(BeamSearchDecodeOp, PackTwoStepsFistStep) { - CPUPlace place; - - LoDTensorArray ids; - LoDTensorArray scores; - - paddle::test::GenerateExample( - std::vector{0, 2, 6}, std::vector{0, 1, 2, 3, 4, 5, 6}, - std::vector{1, 2, 3, 4, 5, 6}, &ids, &scores); - - std::vector> beamnode_vector_list; - std::vector> sentence_vector_list( - 2, SentenceVector()); - - BeamSearchDecoder helper; - beamnode_vector_list = helper.PackTwoSteps( - ids[0], scores[0], &beamnode_vector_list, &sentence_vector_list); - ASSERT_EQ(beamnode_vector_list.size(), 2UL); - ASSERT_EQ(beamnode_vector_list[0].size(), 2UL); - ASSERT_EQ(beamnode_vector_list[1].size(), 4UL); -} - -TEST(BeamSearchDecodeOp, PackTwoSteps) { - CPUPlace place; - - // first source has three prefix - BeamNodeVector source0_prefixes; - source0_prefixes.push_back( - std::unique_ptr>(new BeamNode(1, 1))); - source0_prefixes.push_back( - std::unique_ptr>(new BeamNode(0, 0))); - source0_prefixes.push_back( - std::unique_ptr>(new BeamNode(3, 3))); - - // second source has two prefix - BeamNodeVector source1_prefixes; - source1_prefixes.push_back( - std::unique_ptr>(new BeamNode(4, 4))); - source1_prefixes.push_back( - std::unique_ptr>(new BeamNode(5, 5))); - - std::vector> beamnode_vector_list; - std::vector> sentence_vector_list( - 2, SentenceVector()); - - beamnode_vector_list.push_back(std::move(source0_prefixes)); - beamnode_vector_list.push_back(std::move(source1_prefixes)); - - // generate data for one step - LoDTensorArray ids; - LoDTensorArray scores; - - paddle::test::GenerateExample(std::vector{0, 3, 5}, - std::vector{0, 1, 1, 3, 4, 5}, - std::vector{0, 1, 2, 3, 4}, &ids, &scores); - - BeamSearchDecoder helper1; - beamnode_vector_list = helper1.PackTwoSteps( - ids[0], scores[0], &beamnode_vector_list, &sentence_vector_list); - - ASSERT_EQ(sentence_vector_list[0].size(), 1UL); - ASSERT_EQ(sentence_vector_list[1].size(), 0UL); - ASSERT_EQ(beamnode_vector_list[0].size(), 3UL); - ASSERT_EQ(beamnode_vector_list[1].size(), 2UL); -} - -TEST(BeamSearchDecodeOp, PackAllSteps) { +TEST(BeamSearchDecodeOp, Backtrace) { CPUPlace place; - // we will constuct a sample data with 3 steps and 2 source sentences + // Construct sample data with 5 steps and 2 source sentences + // beam_size = 2, start_id = 0, end_id = 1 LoDTensorArray ids; LoDTensorArray scores; paddle::test::GenerateExample( - std::vector{0, 3, 6}, std::vector{0, 1, 2, 3, 4, 5, 6}, - std::vector{1, 2, 3, 4, 5, 6}, &ids, &scores); + std::vector{0, 1, 2}, std::vector{0, 1, 2}, + std::vector{0, 0}, &ids, &scores); // start with start_id + paddle::test::GenerateExample(std::vector{0, 1, 2}, + std::vector{0, 2, 4}, + std::vector{2, 3, 4, 5}, &ids, &scores); + paddle::test::GenerateExample(std::vector{0, 2, 4}, + std::vector{0, 2, 2, 4, 4}, + std::vector{3, 1, 5, 4}, &ids, &scores); + paddle::test::GenerateExample(std::vector{0, 2, 4}, + std::vector{0, 1, 2, 3, 4}, + std::vector{1, 1, 3, 5}, &ids, &scores); paddle::test::GenerateExample( - std::vector{0, 3, 6}, std::vector{0, 1, 1, 3, 5, 5, 6}, - std::vector{0, 1, 2, 3, 4, 5}, &ids, &scores); - paddle::test::GenerateExample(std::vector{0, 3, 6}, - std::vector{0, 0, 1, 2, 3, 4, 5}, - std::vector{0, 1, 2, 3, 4}, &ids, &scores); + std::vector{0, 2, 4}, + std::vector{0, 0, 0, 2, + 2}, // the branchs of the first source sentence + // are pruned since finished + std::vector{5, 1}, + &ids, &scores); - ASSERT_EQ(ids.size(), 3UL); - ASSERT_EQ(scores.size(), 3UL); + ASSERT_EQ(ids.size(), 5UL); + ASSERT_EQ(scores.size(), 5UL); - BeamSearchDecoder helper; + BeamSearchDecoder helper(2, 1); // beam_size = 2, end_id = 1 LoDTensor id_tensor; LoDTensor score_tensor; - helper.PackAllSteps(ids, scores, &id_tensor, &score_tensor); + helper.Backtrace(ids, scores, &id_tensor, &score_tensor); LoD lod = id_tensor.lod(); - std::vector expect_source_lod = {0, 4, 8}; + std::vector expect_source_lod = {0, 2, 4}; EXPECT_EQ(lod[0], expect_source_lod); - std::vector expect_sentence_lod = {0, 1, 3, 6, 9, 10, 13, 16, 19}; + std::vector expect_sentence_lod = {0, 4, 7, 12, 17}; EXPECT_EQ(lod[1], expect_sentence_lod); - // 2| 1, 0| 3, 1, 0| 3, 2, 1| 5| 4, 3, 2| 4, 4, 3| 6, 5, 4 - std::vector expect_data = {2, 1, 0, 3, 1, 0, 3, 2, 1, 5, - 4, 3, 2, 4, 4, 3, 6, 5, 4}; + std::vector expect_data = {0, 2, 3, 1, 0, 2, 1, 0, 4, + 5, 3, 5, 0, 4, 5, 3, 1}; ASSERT_EQ(id_tensor.dims()[0], static_cast(expect_data.size())); for (size_t i = 0; i < expect_data.size(); ++i) { ASSERT_EQ(id_tensor.data()[i], diff --git a/paddle/fluid/operators/beam_search_op.cc b/paddle/fluid/operators/beam_search_op.cc index df0b50881f4e3ec6f57bdb2b63033931059c486e..62771d09f112785ca1ba741a0ba239b1f0234633 100644 --- a/paddle/fluid/operators/beam_search_op.cc +++ b/paddle/fluid/operators/beam_search_op.cc @@ -12,25 +12,26 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/beam_search_op.h" - #include #include #include #include + #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/beam_search_op.h" namespace paddle { namespace operators { void BeamSearch::operator()(const framework::LoDTensor &pre_ids, + const framework::LoDTensor &pre_scores, framework::LoDTensor *selected_ids, framework::LoDTensor *selected_scores) { auto abs_lod = framework::ToAbsOffset(ids_->lod()); auto &high_level = abs_lod[lod_level_]; - auto items = SelectTopBeamSizeItems(); + auto items = SelectTopBeamSizeItems(pre_ids, pre_scores); auto selected_items = ToMap(items, high_level.back()); VLOG(3) << "selected_items:"; for (size_t i = 0; i < selected_items.size(); ++i) { @@ -39,7 +40,8 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, VLOG(3) << ItemToString(item); } } - PruneEndidCandidates(pre_ids, &selected_items); + + PruneEndBeams(pre_ids, &selected_items); // calculate the output tensor's height size_t num_instances = std::accumulate( std::begin(selected_items), std::end(selected_items), 0, @@ -61,12 +63,6 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, size_t low_offset = 0; for (auto &items : selected_items) { low_level.push_back(low_offset); - sort(items.begin(), items.end(), [](const Item &a, const Item &b) { - if (a.offset < b.offset) { - return true; - } - return a.id < b.id; - }); for (auto &item : items) { ids_data[low_offset] = item.id; scores_data[low_offset] = item.score; @@ -86,21 +82,31 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, selected_scores->set_lod(lod); } -int BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids, - std::vector> *items) { +void BeamSearch::PruneEndBeams(const framework::LoDTensor &pre_ids, + std::vector> *items) { auto *pre_ids_data = pre_ids.data(); - - int res = 0; - for (size_t offset = 0; offset < items->size(); offset++) { - auto prefix_id = pre_ids_data[offset]; - if (prefix_id == end_id_) { - items->at(offset).clear(); - } else { - res++; + auto abs_lod = framework::ToAbsOffset(ids_->lod()); + auto &high_level = abs_lod[lod_level_]; + for (size_t src_idx = 0; src_idx < high_level.size() - 1; ++src_idx) { + size_t src_prefix_start = high_level[src_idx]; + size_t src_prefix_end = high_level[src_idx + 1]; + bool finish_flag = true; + for (size_t offset = src_prefix_start; offset < src_prefix_end; offset++) { + for (auto &item : items->at(offset)) { + if (item.id != static_cast(end_id_) || + pre_ids_data[offset] != end_id_) { + finish_flag = false; + break; + } + } + if (!finish_flag) break; + } + if (finish_flag) { // all branchs of the beam (source sentence) end and + // prune this beam + for (size_t offset = src_prefix_start; offset < src_prefix_end; offset++) + items->at(offset).clear(); } } - - return res; } std::vector> BeamSearch::ToMap( @@ -115,19 +121,17 @@ std::vector> BeamSearch::ToMap( return result; } -std::vector> -BeamSearch::SelectTopBeamSizeItems() { +std::vector> BeamSearch::SelectTopBeamSizeItems( + const framework::LoDTensor &pre_ids, + const framework::LoDTensor &pre_scores) { std::vector> result; std::vector items; // for each source sentence, select the top beam_size items across all // candidate sets. - while (NextItemSet(&items)) { - std::nth_element(std::begin(items), std::begin(items) + beam_size_, - std::end(items), [](const Item &a, const Item &b) { - // TODO(superjom) make score's comparation customizable. - // partial sort in descending order - return a.score > b.score; - }); + while (NextItemSet(pre_ids, pre_scores, &items)) { + std::nth_element( + std::begin(items), std::begin(items) + beam_size_, std::end(items), + [](const Item &a, const Item &b) { return a.score > b.score; }); // prune the top beam_size items. if (items.size() > beam_size_) { items.resize(beam_size_); @@ -146,7 +150,9 @@ BeamSearch::SelectTopBeamSizeItems() { } // the candidates of a source -bool BeamSearch::NextItemSet(std::vector *items) { +bool BeamSearch::NextItemSet(const framework::LoDTensor &pre_ids, + const framework::LoDTensor &pre_scores, + std::vector *items) { if (sent_offset_ >= ids_->NumElements(lod_level_)) { return false; } @@ -164,14 +170,24 @@ bool BeamSearch::NextItemSet(std::vector *items) { instance_dim *= ids.dims()[i]; } + auto *pre_ids_data = pre_ids.data(); + auto *pre_scores_data = pre_scores.data(); items->clear(); items->reserve(framework::product(ids.dims())); for (size_t offset = abs_lod[lod_level_][sent_offset_]; offset < abs_lod[lod_level_][sent_offset_ + 1]; offset++) { - for (size_t d = 0; d < instance_dim; d++) { - const size_t dim_offset = offset * instance_dim + d; - items->emplace_back(offset, ids_data[dim_offset], - scores_data[dim_offset]); + auto pre_id = pre_ids_data[offset]; + auto pre_score = pre_scores_data[offset]; + if (pre_id == end_id_) { + // Allocate all probability mass to eos_id for finished branchs and the + // other candidate ids can be ignored. + items->emplace_back(offset, end_id_, pre_score); + } else { + for (size_t d = 0; d < instance_dim; d++) { + const size_t dim_offset = offset * instance_dim + d; + items->emplace_back(offset, ids_data[dim_offset], + scores_data[dim_offset]); + } } } @@ -199,15 +215,27 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { // inputs and outputs stored in proto - AddInput("pre_ids", "ids in previous step"); - AddInput("ids", "a LoDTensor of shape of [None,k]"); + AddInput("pre_ids", + "(LoDTensor) The LoDTensor containing the selected ids at the " + "previous step. It should be a tensor with shape (batch_size, 1) " + "and lod `[[0, 1, ... , batch_size], [0, 1, ..., batch_size]]` at " + "thefirst step."); + AddInput("pre_scores", + "(LoDTensor) The LoDTensor containing the accumulated " + "scores corresponding to the selected ids at the previous step."); + AddInput("ids", + "(LoDTensor) The LoDTensor containing the candidates ids. Its " + "shape should be (batch_size * beam_size, K), where K supposed to " + "be beam_size."); AddInput("scores", - "a LoDTensor that has the same shape and LoD with `ids`"); + "(LoDTensor) The LodTensor containing the accumulated scores " + "corresponding to Input(ids) and its shape is the same as the " + "shape of Input(ids)."); AddOutput("selected_ids", - "a LoDTensor that stores the IDs selected by beam search"); - AddOutput( - "selected_scores", - "a LoDTensor that has the same shape and LoD with `selected_ids`"); + "A LodTensor that stores the IDs selected by beam search."); + AddOutput("selected_scores", + "A LoDTensor containing the accumulated scores corresponding to " + "Output(selected_ids)."); // Attributes stored in AttributeMap AddAttr("level", "the level of LoDTensor"); @@ -215,8 +243,21 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker { AddAttr("end_id", "the token id which indicates the end of a sequence"); - AddComment( - "This is a beam search operator that help to generate sequences."); + AddComment(R"DOC( +This operator does the search in beams for one time step. +Specifically, it selects the top-K candidate word ids of current step from +Input(ids) according to their Input(scores) for all source sentences, +where K is Attr(beam_size) and Input(ids), Input(scores) are predicted results +from the computation cell. Additionally, Input(pre_ids) and Input(pre_scores) +are the output of beam_search at previous step, they are needed for special use +to handle ended candidate translations. The paths linking prefixes and selected +candidates are organized and reserved in lod. + +Note that the Input(scores) passed in should be accumulated scores, and +length penalty should be done with extra operators before calculating the +accumulated scores if needed, also suggest finding top-K before it and +using the top-K candidates following. +)DOC"); } }; @@ -253,10 +294,12 @@ class BeamSearchInferVarType : public framework::VarTypeInference { void operator()(const framework::OpDesc &op_desc, framework::BlockDesc *block) const override { for (auto &o : op_desc.Output("selected_ids")) { - block->Var(o)->SetType(framework::proto::VarType::LOD_TENSOR); + auto &selected_ids = block->FindRecursiveOrCreateVar(o); + selected_ids.SetType(framework::proto::VarType::LOD_TENSOR); } for (auto &o : op_desc.Output("selected_scores")) { - block->Var(o)->SetType(framework::proto::VarType::LOD_TENSOR); + auto &selected_scores = block->FindRecursiveOrCreateVar(o); + selected_scores.SetType(framework::proto::VarType::LOD_TENSOR); } } }; diff --git a/paddle/fluid/operators/beam_search_op.h b/paddle/fluid/operators/beam_search_op.h index 46bc4f6f936929050276e8b3b93f1eddd62ac638..b5e2ed05924cc8b7bc06058b9b1103ba10be486e 100644 --- a/paddle/fluid/operators/beam_search_op.h +++ b/paddle/fluid/operators/beam_search_op.h @@ -132,6 +132,7 @@ class BeamSearch { * that means no candidates is provided, and the task will stop running. */ void operator()(const framework::LoDTensor& pre_ids, + const framework::LoDTensor& pre_scores, framework::LoDTensor* selected_ids, framework::LoDTensor* selected_scores); /* @@ -153,14 +154,16 @@ class BeamSearch { protected: /* - * Delete all the records that follows the end token. + * Prune the source sentences all branchs finished, and it is optional. + * Pruning must one step later than finishing (thus pre_ids is needed here), + * since the end tokens must be writed out. */ - int PruneEndidCandidates(const framework::LoDTensor& pre_ids, - std::vector>* items); + void PruneEndBeams(const framework::LoDTensor& pre_ids, + std::vector>* items); /* * Transform the items into a map whose key is offset, value is the items. - * NOTE low performance + * NOTE low performance. */ std::vector> ToMap( const std::vector>& inputs, size_t element_num); @@ -168,12 +171,16 @@ class BeamSearch { /* * For each source, select top beam_size records. */ - std::vector> SelectTopBeamSizeItems(); + std::vector> SelectTopBeamSizeItems( + const framework::LoDTensor& pre_ids, + const framework::LoDTensor& pre_scores); /* * Get the items of next source sequence, return false if no remaining items. */ - bool NextItemSet(std::vector* items); + bool NextItemSet(const framework::LoDTensor& pre_ids, + const framework::LoDTensor& pre_scores, + std::vector* items); private: size_t beam_size_; @@ -192,24 +199,25 @@ template class BeamSearchOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* ids_var = context.Input("ids"); - auto* scores_var = context.Input("scores"); - auto* pre_ids_var = context.Input("pre_ids"); - PADDLE_ENFORCE_NOT_NULL(ids_var); - PADDLE_ENFORCE_NOT_NULL(scores_var); - PADDLE_ENFORCE_NOT_NULL(pre_ids_var); + auto* ids = context.Input("ids"); + auto* scores = context.Input("scores"); + auto* pre_ids = context.Input("pre_ids"); + auto* pre_scores = context.Input("pre_scores"); + PADDLE_ENFORCE_NOT_NULL(ids); + PADDLE_ENFORCE_NOT_NULL(scores); + PADDLE_ENFORCE_NOT_NULL(pre_ids); + PADDLE_ENFORCE_NOT_NULL(pre_scores); size_t level = context.Attr("level"); size_t beam_size = context.Attr("beam_size"); int end_id = context.Attr("end_id"); - BeamSearch alg(*ids_var, *scores_var, level, beam_size, end_id); - auto selected_ids_var = - context.Output("selected_ids"); - auto selected_scores_var = + BeamSearch alg(*ids, *scores, level, beam_size, end_id); + auto selected_ids = context.Output("selected_ids"); + auto selected_scores = context.Output("selected_scores"); - PADDLE_ENFORCE_NOT_NULL(selected_ids_var); - PADDLE_ENFORCE_NOT_NULL(selected_scores_var); - alg(*pre_ids_var, selected_ids_var, selected_scores_var); + PADDLE_ENFORCE_NOT_NULL(selected_ids); + PADDLE_ENFORCE_NOT_NULL(selected_scores); + alg(*pre_ids, *pre_scores, selected_ids, selected_scores); } }; } // namespace operators diff --git a/paddle/fluid/operators/beam_search_op_test.cc b/paddle/fluid/operators/beam_search_op_test.cc index ec666359aa2bd81f1323b54f9a03235740c3a696..c4f4b478fbfc87e4178155132781214575c1e6b0 100644 --- a/paddle/fluid/operators/beam_search_op_test.cc +++ b/paddle/fluid/operators/beam_search_op_test.cc @@ -30,7 +30,7 @@ using std::endl; void CreateInput(LoDTensor* ids, LoDTensor* scores) { LoD lod; - vector level0({0, 1, 4}); + vector level0({0, 2, 4}); vector level1({0, 1, 2, 3, 4}); lod.push_back(level0); lod.push_back(level1); @@ -64,17 +64,22 @@ TEST(beam_search_op, run) { for (int i = 0; i < 4; i++) { pre_ids.mutable_data(place)[i] = i + 1; } + LoDTensor pre_scores; + pre_scores.Resize(framework::make_ddim(vector(4, 1))); + for (int i = 0; i < 4; i++) { + pre_scores.mutable_data(place)[i] = 0.1 * (i + 1); + } - BeamSearch beamsearch(ids, scores, (int64_t)0, (int64_t)2, 0); + BeamSearch beamsearch(ids, scores, (size_t)0, (size_t)2, 0); LoDTensor sids, sscores; - beamsearch(pre_ids, &sids, &sscores); + beamsearch(pre_ids, pre_scores, &sids, &sscores); LOG(INFO) << "score: " << sscores << endl; ASSERT_EQ(sids.lod(), sscores.lod()); - vector tids({2, 4, 3, 8}); - vector tscores({0.3, 0.5, 0.9, 0.7}); + vector tids({4, 2, 3, 8}); + vector tscores({0.5, 0.6, 0.9, 0.7}); for (int i = 0; i < 4; i++) { ASSERT_EQ(tids[i], sids.data()[i]); diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc index d46fda54e7a9d5bc737a7ec2116daca33ffa015f..2dc3399da183fbcf7664066f6f7ce12db3dc6d5e 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -34,22 +34,38 @@ class BilinearInterpOp : public framework::OperatorWithKernel { int out_w = ctx->Attrs().Get("out_w"); PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4"); + if (ctx->HasInput("OutSize")) { + auto out_size_dim = ctx->GetInputDim("OutSize"); + PADDLE_ENFORCE_EQ(out_size_dim.size(), 1, + "OutSize's dimension size must be 1"); + PADDLE_ENFORCE_EQ(out_size_dim[0], 2, "OutSize's dim[0] must be 2"); + } std::vector dim_out({dim_x[0], dim_x[1], out_h, out_w}); ctx->SetOutputDim("Out", framework::make_ddim(dim_out)); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace()); + } }; class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "(Tensor) The input tensor of bilinear interpolation, " + "The input tensor of bilinear interpolation, " "This is a 4-D tensor with shape of (N x C x h x w)"); - AddOutput("Out", - "(Tensor) The dimension of output is (N x C x out_h x out_w]"); + AddInput("OutSize", + "This is a 1-D tensor with two number. " + "The first number is height and the second number is width.") + .AsDispensable(); + AddOutput("Out", "The dimension of output is (N x C x out_h x out_w)"); - AddAttr("out_h", "(int) output height of bilinear interpolation op."); - AddAttr("out_w", "(int) output width of bilinear interpolation op."); + AddAttr("out_h", "output height of bilinear interpolation op."); + AddAttr("out_w", "output width of bilinear interpolation op."); AddComment(R"DOC( Bilinear interpolation is an extension of linear interpolation for interpolating functions of two variables (e.g. H-direction and @@ -78,6 +94,12 @@ class BilinearInterpOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), dim_x); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace()); + } }; } // namespace operators @@ -88,6 +110,7 @@ REGISTER_OPERATOR(bilinear_interp, ops::BilinearInterpOp, ops::BilinearInterpOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(bilinear_interp_grad, ops::BilinearInterpOpGrad); -REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel); +REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel, + ops::BilinearInterpKernel); REGISTER_OP_CPU_KERNEL(bilinear_interp_grad, ops::BilinearInterpGradKernel); diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu index 510190f1aaf02960284216a1bedd409011088499..4c1971538495c6f111e9db18f4014786f6f0dd58 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cu +++ b/paddle/fluid/operators/bilinear_interp_op.cu @@ -102,10 +102,21 @@ class BilinearInterpOpCUDAKernel : public framework::OpKernel { auto* input_t = ctx.Input("X"); // float tensor auto* output_t = ctx.Output("Out"); // float tensor auto* input = input_t->data(); - auto* output = output_t->mutable_data(ctx.GetPlace()); int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + auto out_dims = output_t->dims(); + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + auto* output = output_t->mutable_data( + {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); + int batch_size = input_t->dims()[0]; int channels = input_t->dims()[1]; int in_h = input_t->dims()[2]; @@ -139,8 +150,8 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* d_input_t = ctx.Output(framework::GradVarName("X")); auto* d_output_t = ctx.Input(framework::GradVarName("Out")); - auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto* d_output = d_output_t->data(); + auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto& device_ctx = ctx.template device_context(); @@ -149,6 +160,16 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + int batch_size = d_input_t->dims()[0]; int channels = d_input_t->dims()[1]; int in_h = d_input_t->dims()[2]; diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h index f6cd77e4d49b53ecde6a84908cdffc7e1e02ac6a..70847cb8c1abe2e94bc844ab8117d1f23fea533b 100644 --- a/paddle/fluid/operators/bilinear_interp_op.h +++ b/paddle/fluid/operators/bilinear_interp_op.h @@ -24,11 +24,18 @@ class BilinearInterpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* input_t = ctx.Input("X"); // float tensor auto* output_t = ctx.Output("Out"); // float tensor + auto out_dims = output_t->dims(); auto* input = input_t->data(); - auto* output = output_t->mutable_data(ctx.GetPlace()); - int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + auto out_size_data = out_size_t->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + auto* output = output_t->mutable_data( + {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); int batch_size = input_t->dims()[0]; int channels = input_t->dims()[1]; int in_h = input_t->dims()[2]; @@ -39,8 +46,10 @@ class BilinearInterpKernel : public framework::OpKernel { int in_chw = channels * in_hw; int out_chw = channels * out_hw; - T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; if (in_h == out_h && in_w == out_w) { memcpy(output, input, input_t->numel() * sizeof(T)); @@ -49,24 +58,24 @@ class BilinearInterpKernel : public framework::OpKernel { for (int i = 0; i < out_h; ++i) { // loop for images int h = ratio_h * i; int hid = (h < in_h - 1) ? 1 : 0; - T h1lambda = ratio_h * i - h; - T h2lambda = 1 - h1lambda; + float h1lambda = ratio_h * i - h; + float h2lambda = 1.f - h1lambda; for (int j = 0; j < out_w; ++j) { int w = ratio_w * j; int wid = (w < in_w - 1) ? 1 : 0; - T w1lambda = ratio_w * j - w; - T w2lambda = 1 - w1lambda; + float w1lambda = ratio_w * j - w; + float w2lambda = 1.f - w1lambda; // calculate four position for bilinear interpolation const T* in_pos = &input[k * in_chw + h * in_w + w]; T* out_pos = &output[k * out_chw + i * out_w + j]; for (int c = 0; c < channels; ++c) { // loop for channels // bilinear interpolation - out_pos[0] = + out_pos[0] = static_cast( h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[wid]) + h1lambda * (w2lambda * in_pos[hid * in_w] + - w1lambda * in_pos[hid * in_w + wid]); + w1lambda * in_pos[hid * in_w + wid])); in_pos += in_hw; out_pos += out_hw; } @@ -83,9 +92,8 @@ class BilinearInterpGradKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* d_input_t = ctx.Output(framework::GradVarName("X")); auto* d_output_t = ctx.Input(framework::GradVarName("Out")); - auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto* d_output = d_output_t->data(); - + auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto& device_ctx = ctx.template device_context(); math::SetConstant zero; @@ -93,6 +101,14 @@ class BilinearInterpGradKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + auto out_size_data = out_size_t->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + int batch_size = d_input_t->dims()[0]; int channels = d_input_t->dims()[1]; int in_h = d_input_t->dims()[2]; @@ -103,8 +119,10 @@ class BilinearInterpGradKernel : public framework::OpKernel { int in_chw = channels * in_hw; int out_chw = channels * out_hw; - T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; if (in_h == out_h && in_w == out_w) { memcpy(d_input, d_output, d_input_t->numel() * sizeof(T)); @@ -113,22 +131,24 @@ class BilinearInterpGradKernel : public framework::OpKernel { for (int i = 0; i < out_h; ++i) { // loop for images int h = ratio_h * i; int hid = (h < in_h - 1) ? 1 : 0; - T h1lambda = ratio_h * i - h; - T h2lambda = 1 - h1lambda; + float h1lambda = ratio_h * i - h; + float h2lambda = 1 - h1lambda; for (int j = 0; j < out_w; ++j) { int w = ratio_w * j; int wid = (w < in_w - 1) ? 1 : 0; - T w1lambda = ratio_w * j - w; - T w2lambda = 1 - w1lambda; + float w1lambda = ratio_w * j - w; + float w2lambda = 1 - w1lambda; T* in_pos = &d_input[k * in_chw + h * in_w + w]; const T* out_pos = &d_output[k * out_chw + i * out_w + j]; for (int c = 0; c < channels; ++c) { // loop for channels - in_pos[0] += h2lambda * w2lambda * out_pos[0]; - in_pos[wid] += h2lambda * w1lambda * out_pos[0]; - in_pos[hid * in_w] += h1lambda * w2lambda * out_pos[0]; - in_pos[hid * in_w + wid] += h1lambda * w1lambda * out_pos[0]; + in_pos[0] += static_cast(h2lambda * w2lambda * out_pos[0]); + in_pos[wid] += static_cast(h2lambda * w1lambda * out_pos[0]); + in_pos[hid * in_w] += + static_cast(h1lambda * w2lambda * out_pos[0]); + in_pos[hid * in_w + wid] += + static_cast(h1lambda * w1lambda * out_pos[0]); in_pos += in_hw; out_pos += out_hw; } diff --git a/paddle/fluid/operators/checkpoint_notify_op.cc b/paddle/fluid/operators/checkpoint_notify_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..3a2527e407bb179c4873fa3ffe2e8f22fb47faf7 --- /dev/null +++ b/paddle/fluid/operators/checkpoint_notify_op.cc @@ -0,0 +1,88 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include // NOLINT +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/send_recv_util.h" +#include "paddle/fluid/string/printf.h" + +namespace paddle { +namespace operators { + +class CheckpointNotifyOp : public framework::OperatorBase { + public: + CheckpointNotifyOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& place) const override { + std::vector epmap = Attr>("epmap"); + std::string dir = Attr("dir"); + std::string lookup_table_name = Attr("lookup_table"); + + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); + for (size_t i = 0; i < epmap.size(); i++) { + auto lookup_table_save_dir = + string::Sprintf("%s/%s_%d", dir, lookup_table_name, i); + rpc_client->AsyncCheckpointNotify(epmap[i], lookup_table_save_dir); + VLOG(3) << "checkpoint notify sending lookup table: " << lookup_table_name + << " and dir:" << dir << " to " << epmap[i]; + } + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); + } +}; + +class CheckpointNotifyOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() { + AddAttr>("epmap", + "(string vector, default 127.0.0.1:6164)" + "Parameter Server endpoints in the order") + .SetDefault({"127.0.0.1:6164"}); + AddAttr( + "dir", "(string, default '') indicate the folder checkpoint will use"); + AddAttr("lookup_table", + "(string, default '') the lookup table name"); + AddComment(R"DOC( +CheckpointNotify operator + +This operator will send lookup table and it's checkpoint direcoty to listen_and_serve op at +the parameter server. +)DOC"); + } +}; + +class CheckpointNotifyOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(checkpoint_notify, ops::CheckpointNotifyOp, + paddle::framework::EmptyGradOpMaker, + ops::CheckpointNotifyOpMaker, + ops::CheckpointNotifyOpShapeInference); diff --git a/paddle/fluid/operators/chunk_eval_op.cc b/paddle/fluid/operators/chunk_eval_op.cc index 62636bb2f9078768180ab1e0016e3565617d24d2..dc43c69be0bcea2b82e1d61a9a5b2e03129d4f8e 100644 --- a/paddle/fluid/operators/chunk_eval_op.cc +++ b/paddle/fluid/operators/chunk_eval_op.cc @@ -91,32 +91,31 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker { "(int64_t). The number of chunks both in Inference and Label on the " "given mini-batch."); AddAttr("num_chunk_types", - "(int). The number of chunk type. See below for details."); - AddAttr( - "chunk_scheme", - "(string, default IOB). The labeling scheme indicating " - "how to encode the chunks. Must be IOB, IOE, IOBES or plain. See below " - "for details.") + "The number of chunk type. See the description for details."); + AddAttr("chunk_scheme", + "The labeling scheme indicating " + "how to encode the chunks. Must be IOB, IOE, IOBES or " + "plain. See the description" + "for details.") .SetDefault("IOB"); AddAttr>("excluded_chunk_types", - "(list) A list including chunk type ids " + "A list including chunk type ids " "indicating chunk types that are not counted. " - "See below for details.") + "See the description for details.") .SetDefault(std::vector{}); AddComment(R"DOC( For some basics of chunking, please refer to -‘Chunking with Support Vector Machines ’. +'Chunking with Support Vector Machines '. - -CheckEvalOp computes the precision, recall, and F1-score of chunk detection, +ChunkEvalOp computes the precision, recall, and F1-score of chunk detection, and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes. Here is a NER example of labeling for these tagging schemes: - - Li Ming works at Agricultural Bank of China in Beijing. - IO: I-PER I-PER O O I-ORG I-ORG I-ORG I-ORG O I-LOC - IOB: B-PER I-PER O O B-ORG I-ORG I-ORG I-ORG O B-LOC - IOE: I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC - IOBES: B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC + + Li Ming works at Agricultural Bank of China in Beijing. + IO I-PER I-PER O O I-ORG I-ORG I-ORG I-ORG O I-LOC + IOB B-PER I-PER O O B-ORG I-ORG I-ORG I-ORG O B-LOC + IOE I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC + IOBES B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC There are three chunk types(named entity types) including PER(person), ORG(organization) and LOC(LOCATION), and we can see that the labels have the form -. @@ -124,31 +123,31 @@ and LOC(LOCATION), and we can see that the labels have the form -("force_cpu", - "(bool, default false) Force fill output variable to cpu " + "Force fill output variable to cpu " "memory. Otherwise, fill output variable to the running " - "device") - .SetDefault(false); - AddOutput("Out", string::Sprintf( - "(LoDTensor) n-dim bool tensor. Each element is %s", - comment.equation)); - AddComment(string::Sprintf(R"DOC(%s Operator - + "device [default true].") + .SetDefault(true); + AddOutput("Out", string::Sprintf("n-dim bool tensor. Each element is %s", + comment.equation)); + AddComment(string::Sprintf(R"DOC( It operates element-wise on X and Y, and returns the Out. Each of them is a N-dim tensor. X and Y could be any type. The each element of the Out tensor is -calculated by %s +calculated by $%s$ )DOC", - comment.type, comment.equation)); - AddAttr("axis", - "(int, default -1). The start dimension index " - "for broadcasting Y onto X.") + comment.equation)); + AddAttr( + "axis", + "The start dimension index for broadcasting Y onto X. [default -1]") .SetDefault(-1) .EqualGreaterThan(-1); } diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc index 38337f9aa52435c445420047957500d21069506a..c72405593788493e10a1293b0c722e2d11c6e312 100644 --- a/paddle/fluid/operators/concat_op.cc +++ b/paddle/fluid/operators/concat_op.cc @@ -107,7 +107,13 @@ REGISTER_OPERATOR(concat, ops::ConcatOp, ops::ConcatOpMaker, false> /* set false to disable empty grad */); REGISTER_OPERATOR(concat_grad, ops::ConcatOpGrad); REGISTER_OP_CPU_KERNEL( - concat, ops::ConcatKernel); + concat, ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel); REGISTER_OP_CPU_KERNEL( concat_grad, - ops::ConcatGradKernel); + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel); diff --git a/paddle/fluid/operators/concat_op.cu.cc b/paddle/fluid/operators/concat_op.cu.cc index 590eca9d066ff7549939e62ddbfedc8ab76bb5e7..8e38e5231fbf6955ff8a9680a241a4a4ba1b924d 100644 --- a/paddle/fluid/operators/concat_op.cu.cc +++ b/paddle/fluid/operators/concat_op.cu.cc @@ -15,7 +15,13 @@ limitations under the License. */ #include "paddle/fluid/operators/concat_op.h" namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( - concat, ops::ConcatKernel); + concat, ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel); REGISTER_OP_CUDA_KERNEL( concat_grad, - ops::ConcatGradKernel); + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel); diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h index 1b1b8bf5ed959dd9c2ce8c9f5c905a75b81865fd..a496301526f58875ff51aeaa5b2094c3c656531c 100644 --- a/paddle/fluid/operators/concat_op.h +++ b/paddle/fluid/operators/concat_op.h @@ -60,34 +60,45 @@ template class ConcatGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { - auto* in = ctx.Input(framework::GradVarName("Out")); + auto* out_grad = + ctx.Input(framework::GradVarName("Out")); + auto ins = ctx.MultiInput("X"); + auto out_var_names = ctx.Outputs(framework::GradVarName("X")); auto outs = ctx.MultiOutput(framework::GradVarName("X")); int64_t axis = static_cast(ctx.Attr("axis")); + // get output tensor that the name is not kEmptyVarName + std::vector outputs; + for (size_t j = 0; j < outs.size(); ++j) { + if (out_var_names[j] != framework::kEmptyVarName) { + outs[j]->mutable_data(ctx.GetPlace()); + outputs.push_back(outs[j]); + } else { + outputs.push_back(nullptr); + } + } + // Sometimes direct copies will be faster, this maybe need deeply analysis. if (axis == 0 && outs.size() < 10) { size_t input_offset = 0; - auto in_stride = framework::stride_numel(in->dims()); + const auto in_stride = framework::stride_numel(out_grad->dims()); - for (auto& out : outs) { - out->mutable_data(ctx.GetPlace()); - auto out_stride = framework::stride_numel(out->dims()); - StridedNumelCopyWithAxis(ctx.device_context(), axis, out->data(), - out_stride, in->data() + input_offset, - in_stride, out_stride[axis]); + for (size_t i = 0; i < outs.size(); ++i) { + auto out_stride = framework::stride_numel(ins[i]->dims()); + auto* out = outputs[i]; + if (out != nullptr) { + StridedNumelCopyWithAxis( + ctx.device_context(), axis, out->data(), out_stride, + out_grad->data() + input_offset, in_stride, out_stride[axis]); + } input_offset += out_stride[axis]; } } else { - std::vector outputs(outs.size()); - for (size_t j = 0; j < outs.size(); ++j) { - outs[j]->mutable_data(ctx.GetPlace()); - outputs[j] = *outs[j]; - } - auto& dev_ctx = ctx.template device_context(); paddle::operators::math::ConcatGradFunctor concat_grad_functor; - concat_grad_functor(dev_ctx, *in, static_cast(axis), &outputs); + concat_grad_functor(dev_ctx, *out_grad, ins, static_cast(axis), + &outputs); } } }; diff --git a/paddle/fluid/operators/conditional_block_op.cc b/paddle/fluid/operators/conditional_block_op.cc index 5984f80d04bdeb232f8e24264ae979725af24ef4..580fde753816c30b188b8a99cc63fcbafde64e25 100644 --- a/paddle/fluid/operators/conditional_block_op.cc +++ b/paddle/fluid/operators/conditional_block_op.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/var_type.h" namespace paddle { namespace operators { @@ -47,7 +48,7 @@ class ConditionalOp : public framework::OperatorBase { if (!(ips.size() == 1UL && ips[0]->IsInitialized())) { PADDLE_THROW("should have one initialized input as condition"); } - if (!(ips[0]->type().hash_code() == typeid(bool).hash_code() && // NOLINT + if (!(framework::IsType(ips[0]->type()) && // NOLINT ips[0]->numel() == 1)) { PADDLE_THROW( "condition input's data type should be bool, " @@ -204,9 +205,10 @@ class ConditionalBlockGradInferShape : public framework::InferShapeBase { context->SetOutputsDim(framework::GradVarName("Params"), context->GetInputsDim("Params")); } - PADDLE_ENFORCE(context->HasOutputs(framework::GradVarName("X"))); - context->SetOutputsDim(framework::GradVarName("X"), - context->GetInputsDim("X")); + if (context->HasOutputs(framework::GradVarName("X"))) { + context->SetOutputsDim(framework::GradVarName("X"), + context->GetInputsDim("X")); + } } }; diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index 7a7b8b76e43b1f91a3ba2767c217993cc39f26b6..22cbf680c0670552fb014043c69fcadc56863529 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -20,10 +20,10 @@ limitations under the License. */ #include "paddle/fluid/platform/cudnn_helper.h" #include "paddle/fluid/platform/float16.h" -DEFINE_bool(cudnn_algo_use_autotune, true, +DEFINE_bool(cudnn_deterministic, false, "Whether allow using an autotuning algorithm for convolution " "operator. The autotuning algorithm may be non-deterministic. If " - "false, the algorithm is deterministic."); + "true, the algorithm is deterministic."); namespace paddle { namespace operators { @@ -77,7 +77,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { // cudnn 7 can support groups, no need to do it mannually // FIXME(typhoonzero): find a better way to disable groups // rather than setting it to 1. - PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount( + CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount( cudnn_conv_desc, groups)); groups = 1; #endif @@ -129,7 +129,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); - PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, workspace_size_limit, &algo)); @@ -140,18 +140,18 @@ class CUDNNConvOpKernel : public framework::OpKernel { if (dev_ctx.GetComputeCapability() >= 70 && std::type_index(typeid(T)) == std::type_index(typeid(platform::float16))) { - PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( + CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( cudnn_conv_desc, CUDNN_TENSOR_OP_MATH)); // Currently tensor core is only enabled using this algo algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; } else { - PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( + CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( cudnn_conv_desc, CUDNN_DEFAULT_MATH)); } #endif // get workspace size able to allocate - PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize( + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize( handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, cudnn_output_desc, algo, &workspace_size_in_bytes)); // It is possible for float16 on Volta GPU to allocate more memory than @@ -165,7 +165,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { // ------------------- cudnn conv forward --------------------- ScalingParamType alpha = 1.0f, beta = 0.0f; for (int i = 0; i < groups; i++) { - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward( + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward( handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in, cudnn_filter_desc, filter_data + i * group_offset_filter, cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes, @@ -218,7 +218,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { // cudnn 7 can support groups, no need to do it mannually // FIXME(typhoonzero): find a better way to disable groups // rather than setting it to 1. - PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount( + CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount( cudnn_conv_desc, groups)); groups = 1; #endif @@ -272,8 +272,8 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); if (input_grad) { - if (FLAGS_cudnn_algo_use_autotune) { - PADDLE_ENFORCE( + if (!FLAGS_cudnn_deterministic) { + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( handle, cudnn_filter_desc, // dyDesc: Handle to the previously initialized input @@ -289,7 +289,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1; } - PADDLE_ENFORCE( + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( handle, cudnn_filter_desc, cudnn_output_grad_desc, cudnn_conv_desc, cudnn_input_desc, data_algo, &tmp_size)); @@ -297,8 +297,8 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } if (filter_grad) { - if (FLAGS_cudnn_algo_use_autotune) { - PADDLE_ENFORCE( + if (!FLAGS_cudnn_deterministic) { + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc, cudnn_filter_desc, @@ -308,7 +308,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1; } - PADDLE_ENFORCE( + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc, cudnn_filter_desc, filter_algo, &tmp_size)); @@ -326,7 +326,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { // Because beta is zero, it is unnecessary to reset input_grad. for (int i = 0; i < groups; i++) { - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( handle, &alpha, cudnn_filter_desc, filter_data + i * group_offset_filter, cudnn_output_grad_desc, output_grad_data + i * group_offset_out, cudnn_conv_desc, data_algo, @@ -339,7 +339,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { T* filter_grad_data = filter_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset filter_grad. for (int i = 0; i < groups; i++) { - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in, cudnn_output_grad_desc, output_grad_data + i * group_offset_out, cudnn_conv_desc, filter_algo, cudnn_workspace, diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index 63d371310d2a26a1460e527fc51923dfd6e0b8bc..f07ab5a33b87d7945e5fcdf8f3644f0711ce643b 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -18,6 +18,204 @@ namespace paddle { namespace operators { +using framework::DataLayout; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; +using platform::to_void_cast; +using platform::GetMKLDNNFormat; + +class ConvMKLDNNHandler : public platform::MKLDNNHandler { + public: + ConvMKLDNNHandler( + std::shared_ptr conv_pd, + const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : platform::MKLDNNHandler(dev_ctx, engine, base_key) { + conv_pd_ = conv_pd; + } + + ConvMKLDNNHandler( + std::shared_ptr conv_pd, + std::shared_ptr + conv_bwd_data_pd, + std::shared_ptr + conv_bwd_weights_pd, + const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : platform::MKLDNNHandler(dev_ctx, engine, base_key), + conv_pd_(conv_pd), + conv_bwd_weights_pd_(conv_bwd_weights_pd), + conv_bwd_data_pd_(conv_bwd_data_pd) { + // If we are in Grad operatgor then update a key with BWD suffix to + // distinguish from FWD memory primitives + key_ += "-BWD"; + } + + std::shared_ptr AcquireSrcMemoryFromWeightsPrimitive( + const std::shared_ptr user_memory_p, + std::vector& pipeline) { // NOLINT + auto src_pd = conv_bwd_weights_pd_->src_primitive_desc(); + auto user_pd = user_memory_p->get_primitive_desc(); + return this->AcquireMemory(src_pd, user_pd, user_memory_p, + "@weights-src_mem_p", pipeline); + } + + std::shared_ptr AcquireDiffDstMemoryFromWeightsPrimitive( + const std::shared_ptr user_memory_p, + std::vector& pipeline) { // NOLINT + auto diff_dst_pd = conv_bwd_weights_pd_->diff_dst_primitive_desc(); + auto user_pd = user_memory_p->get_primitive_desc(); + return this->AcquireMemory(diff_dst_pd, user_pd, user_memory_p, + "@weights-diff_dst_mem_p", pipeline); + } + + std::shared_ptr AcquireDiffWeightsMemoryFromWeightsPrimitive( + void* ptr) { + return this->AcquireMemoryFromPrimitive( + conv_bwd_weights_pd_->diff_weights_primitive_desc(), ptr, + "@diff_weights_mem_p"); + } + + std::shared_ptr AcquireDiffDstMemoryFromDataPrimitive( + const std::shared_ptr user_memory_p, + std::vector& pipeline) { // NOLINT + auto diff_dst_pd = conv_bwd_data_pd_->diff_dst_primitive_desc(); + auto user_pd = user_memory_p->get_primitive_desc(); + return this->AcquireMemory(diff_dst_pd, user_pd, user_memory_p, + "@data-diff_dst_mem_p", pipeline); + } + + std::shared_ptr AcquireWeightsMemoryFromDataPrimitive( + const std::shared_ptr user_weights_memory_p, + std::vector& pipeline) { // NOLINT + auto weights_pd = conv_bwd_data_pd_->weights_primitive_desc(); + auto user_pd = user_weights_memory_p->get_primitive_desc(); + return this->AcquireMemory(weights_pd, user_pd, user_weights_memory_p, + "@data-weights_mem_p", pipeline); + } + + std::shared_ptr AcquireDiffSrcMemoryFromDataPrimitive( + void* ptr) { + return this->AcquireMemoryFromPrimitive( + conv_bwd_data_pd_->diff_src_primitive_desc(), ptr, "@diff_src_mem_p"); + } + + std::shared_ptr AcquireDstMemoryFromPrimitive(void* ptr) { + return this->AcquireMemoryFromPrimitive(conv_pd_->dst_primitive_desc(), ptr, + "@dst_mem_p"); + } + + std::shared_ptr AcquireSrcMemoryFromPrimitive( + const std::shared_ptr user_memory_p, + std::vector& pipeline) { // NOLINT + auto src_pd = conv_pd_->src_primitive_desc(); + auto user_pd = user_memory_p->get_primitive_desc(); + return this->AcquireMemory(src_pd, user_pd, user_memory_p, "@src_mem_p", + pipeline); + } + + std::shared_ptr AcquireWeightsMemoryFromPrimitive( + const std::shared_ptr user_weights_memory_p, + std::vector& pipeline) { // NOLINT + auto user_weights_pd = user_weights_memory_p->get_primitive_desc(); + auto weights_pd = conv_pd_->weights_primitive_desc(); + return this->AcquireMemory(weights_pd, user_weights_pd, + user_weights_memory_p, "@weights_mem_p", + pipeline); + } + + std::shared_ptr AcquireConvolution( + std::shared_ptr src_memory_p, + std::shared_ptr weights_memory_p, + std::shared_ptr dst_memory_p) { + auto prim_key = key_ + "@conv_p"; + auto conv_p = std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false), + "Fail to find convolution primitive in device context"); + if (conv_p == nullptr) { + conv_p = std::make_shared( + *conv_pd_, *(src_memory_p), *(weights_memory_p.get()), + *(dst_memory_p.get())); + + dev_ctx_.SetBlob(prim_key, conv_p); + } else { + is_reusing_ = true; + } + return conv_p; + } + + std::shared_ptr + AcquireConvolutionBackwardWeights( + std::shared_ptr src_memory_p, + std::shared_ptr diff_dst_memory_p, + std::shared_ptr diff_weights_memory_p) { + auto prim_key = key_ + "@conv_bwd_weights_p"; + auto conv_bwd_weights_p = + std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE( + (conv_bwd_weights_p != nullptr) || (is_reusing_ == false), + "Fail to find convolution bwd weights primitive in device context"); + if (conv_bwd_weights_p == nullptr) { + // create backward conv primitive for weights + conv_bwd_weights_p = + std::make_shared( + *conv_bwd_weights_pd_, *src_memory_p, *diff_dst_memory_p, + *diff_weights_memory_p); + dev_ctx_.SetBlob(prim_key, conv_bwd_weights_p); + } else { + is_reusing_ = true; + } + return conv_bwd_weights_p; + } + + std::shared_ptr + AcquireConvolutionBackwardData( + std::shared_ptr diff_dst_memory_p, + std::shared_ptr weights_memory_p, + std::shared_ptr diff_src_memory_p) { + auto prim_key = key_ + "@conv_bwd_data_p"; + auto conv_bwd_data_p = + std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE( + (conv_bwd_data_p != nullptr) || (is_reusing_ == false), + "Fail to find convolution bwd data primitive in device context"); + if (conv_bwd_data_p == nullptr) { + conv_bwd_data_p = std::make_shared( + *conv_bwd_data_pd_, *diff_dst_memory_p, *weights_memory_p, + *diff_src_memory_p); + dev_ctx_.SetBlob(prim_key, conv_bwd_data_p); + } else { + is_reusing_ = true; + } + return conv_bwd_data_p; + } + + // Generate keys for storing/retriving primitives for this operator + // TODO(jczaja): Make hashing function more optimial + static std::string GetHash(memory::dims& input_dims, // NOLINT + memory::dims& weights_dims, // NOLINT + std::vector& strides, // NOLINT + std::vector& paddings, // NOLINT + std::vector& dilations, // NOLINT + int groups, const std::string& suffix) { + return dims2str(input_dims) + dims2str(weights_dims) + dims2str(strides) + + dims2str(paddings) + dims2str(dilations) + std::to_string(groups) + + suffix; + } + + private: + std::shared_ptr conv_pd_; + std::shared_ptr + conv_bwd_weights_pd_; + std::shared_ptr + conv_bwd_data_pd_; +}; + template class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { public: @@ -33,10 +231,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto* filter = ctx.Input("Filter"); auto* output = ctx.Output("Output"); - // Get an unique name from "argument" name of "Output" variable - // This name will be used as key when saving info into device context - const std::string key = ctx.op().Output("Output"); - const std::string key_conv_pd = key + "@conv_pd"; + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != memory::format::format_undef, + "Wrong layout/format set for Input tensor"); + PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && + filter->format() != memory::format::format_undef, + "Wrong layout/format set for Filter tensor"); std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -63,49 +263,77 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { paddle::framework::vectorize2int(filter->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // TODO(pzelazko-intel): support more formats + // Get unique name for storing MKLDNN primitives + const std::string key = ConvMKLDNNHandler::GetHash( + src_tz, weights_tz, strides, paddings, dilations, groups, + ctx.op().Output("Output")); + const std::string key_conv_pd = key + "@conv_pd"; + + std::vector pipeline; + + auto user_src_md = platform::MKLDNNMemDesc( + {src_tz}, platform::MKLDNNGetDataType(), input->format()); + auto user_weights_md = platform::MKLDNNMemDesc( + {weights_tz}, platform::MKLDNNGetDataType(), filter->format()); + + /* create memory descriptor for convolution without specified format + * ('any') which lets a primitive (convolution in this case) choose + * the memory format preferred for best performance + */ + std::string data_format = ctx.Attr("data_format"); + auto chosen_memory_format = + platform::data_format_to_memory_format(data_format); + auto src_md = platform::MKLDNNMemDesc( - src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - auto weights_md = - platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, - mkldnn::memory::format::oihw); + src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + auto weights_md = platform::MKLDNNMemDesc( + weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto dst_md = platform::MKLDNNMemDesc( - dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - - auto src_memory = - mkldnn::memory({src_md, mkldnn_engine}, - reinterpret_cast(const_cast(input_data))); - auto weights_memory = - mkldnn::memory({weights_md, mkldnn_engine}, - reinterpret_cast(const_cast(filter_data))); - auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data); + dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + // create a conv primitive descriptor and save it for usage in backward std::shared_ptr conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, mkldnn_engine); - - // save conv_pd into global device context to be referred in backward path + // Save conv_pd/src_memory/weights_memory for backward pass dev_ctx.SetBlob(key_conv_pd, conv_pd); + ConvMKLDNNHandler handler(conv_pd, dev_ctx, mkldnn_engine, key); + + // create mkldnn memory from input tensors (data/weights) + auto user_src_memory_p = + handler.AcquireSrcMemory(user_src_md, to_void_cast(input_data)); + auto user_weights_memory_p = handler.AcquireWeightsMemory( + user_weights_md, to_void_cast(filter_data)); + + // create reorder primitive if the input format is not the preferred one + auto src_memory_p = + handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); + auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive( + user_weights_memory_p, pipeline); + auto dst_memory_p = + handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); + // create convolution op primitive - auto conv_prim = mkldnn::convolution_forward(*conv_pd, src_memory, - weights_memory, dst_memory); + auto conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p, + dst_memory_p); // push primitive to stream and wait until it's executed - std::vector pipeline{conv_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + pipeline.push_back(*conv_p); + stream(stream::kind::eager).submit(pipeline).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(GetMKLDNNFormat(*dst_memory_p)); } private: std::unique_ptr - ConvFwdPrimitiveDesc(const mkldnn::memory::desc& src, - const mkldnn::memory::desc& weights, - const mkldnn::memory::desc& dst, - const std::vector& strides, + ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights, + const memory::desc& dst, const std::vector& strides, const std::vector& paddings, const mkldnn::engine& engine) const { - mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; - mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; + memory::dims stride_dims = {strides[0], strides[1]}; + memory::dims padding_dims = {paddings[0], paddings[1]}; auto conv_desc = mkldnn::convolution_forward::desc( mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights, @@ -139,15 +367,25 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { Tensor* input_grad = ctx.Output(framework::GradVarName("Input")); Tensor* filter_grad = ctx.Output(framework::GradVarName("Filter")); - if (!input_grad && !filter_grad) return; + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != memory::format::format_undef, + "Wrong layout/format set for Input tensor"); + PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && + filter->format() != memory::format::format_undef, + "Wrong layout/format set for Filter tensor"); + PADDLE_ENFORCE(output->layout() == DataLayout::kMKLDNN && + output->format() != memory::format::format_undef, + "Wrong layout/format set for Output tensor"); + PADDLE_ENFORCE(output_grad->layout() == DataLayout::kMKLDNN && + output_grad->format() != memory::format::format_undef, + "Wrong layout/format set for output_grad tensor"); - // Get an unique name from "argument" name of "Output" variable - // This name will be used as key when saving info into device context - const std::string key = ctx.op().Input("Output"); - const std::string key_conv_pd = key + "@conv_pd"; + if (!input_grad && !filter_grad) return; std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); + std::vector dilations = ctx.Attr>("dilations"); + int groups = ctx.Attr("groups"); const T* input_data = input->data(); const T* filter_data = filter->data(); @@ -167,24 +405,43 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { paddle::framework::vectorize2int(filter->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // TODO(pzelazko-intel): support more formats + // Get an unique name from "argument" name of "Output" variable + // as well as attributes of primitive to be created + // This name will be used as key when saving info into device context + const std::string key = + ConvMKLDNNHandler::GetHash(src_tz, weights_tz, strides, paddings, + dilations, groups, ctx.op().Input("Output")); + + const std::string key_conv_pd = key + "@conv_pd"; + std::vector pipeline; + + // Create user memory descriptors + auto user_src_md = platform::MKLDNNMemDesc( + {src_tz}, platform::MKLDNNGetDataType(), input->format()); + auto user_weights_md = platform::MKLDNNMemDesc( + {weights_tz}, platform::MKLDNNGetDataType(), filter->format()); + auto user_diff_dst_md = platform::MKLDNNMemDesc( + {dst_tz}, platform::MKLDNNGetDataType(), output_grad->format()); + + /* create memory descriptor for conv backward without specified format + * ('any') which lets a primitive (conv backward in this case) choose + * the memory format preferred for best performance + */ + std::string data_format = ctx.Attr("data_format"); + auto chosen_memory_format = + platform::data_format_to_memory_format(data_format); + auto src_md = platform::MKLDNNMemDesc( - src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto diff_src_md = platform::MKLDNNMemDesc( - src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - auto weights_md = - platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, - mkldnn::memory::format::oihw); - auto diff_weights_md = - platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, - mkldnn::memory::format::oihw); + src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + auto weights_md = platform::MKLDNNMemDesc( + weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + auto diff_weights_md = platform::MKLDNNMemDesc( + weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto diff_dst_md = platform::MKLDNNMemDesc( - dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); - // create memory - auto diff_dst_memory = mkldnn::memory( - {diff_weights_md, mkldnn_engine}, - reinterpret_cast(const_cast(output_grad_data))); // Retrieve conv_pd from device context auto conv_pd = std::static_pointer_cast( @@ -192,83 +449,77 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { PADDLE_ENFORCE(conv_pd != nullptr, "Fail to find conv_pd in device context"); + // create backward convolution weights primitive descriptor + auto conv_bwd_weights_desc = mkldnn::convolution_backward_weights::desc( + mkldnn::convolution_direct, src_md, diff_weights_md, diff_dst_md, + strides, paddings, paddings, mkldnn::padding_kind::zero); + auto conv_bwd_weights_pd = + std::make_shared( + conv_bwd_weights_desc, mkldnn_engine, *conv_pd); + + // create backward convolution data primitive descriptor + auto conv_bwd_data_desc = mkldnn::convolution_backward_data::desc( + mkldnn::convolution_direct, diff_src_md, weights_md, diff_dst_md, + strides, paddings, paddings, mkldnn::padding_kind::zero); + auto conv_bwd_data_pd = + std::make_shared( + conv_bwd_data_desc, mkldnn_engine, *conv_pd); + + ConvMKLDNNHandler handler(conv_pd, conv_bwd_data_pd, conv_bwd_weights_pd, + dev_ctx, mkldnn_engine, key); + + // create mkldnn memory from input tensors (data/weights) + auto user_src_memory_p = + handler.AcquireSrcMemory(user_src_md, to_void_cast(input_data)); + auto user_weights_memory_p = handler.AcquireWeightsMemory( + user_weights_md, to_void_cast(filter_data)); + auto user_diff_dst_memory_p = handler.AcquireDiffDstMemory( + user_diff_dst_md, to_void_cast(output_grad_data)); + // create backward conv primitive for weights if (filter_grad) { - // create primitive descriptor - mkldnn::convolution_backward_weights::primitive_desc conv_bwd_weights_pd = - ConvBwdWeightsPrimitiveDesc(src_md, diff_weights_md, diff_dst_md, - strides, paddings, *conv_pd, - mkldnn_engine); - - // create memory - auto diff_weights_memory = - mkldnn::memory({diff_weights_md, mkldnn_engine}, - reinterpret_cast(filter_grad_data)); - auto src_memory = - mkldnn::memory({src_md, mkldnn_engine}, - reinterpret_cast(const_cast(input_data))); + auto src_memory_p = handler.AcquireSrcMemoryFromWeightsPrimitive( + user_src_memory_p, pipeline); - // create backward conv primitive for weights - auto conv_bwd_weights_prim = mkldnn::convolution_backward_weights( - conv_bwd_weights_pd, src_memory, diff_dst_memory, - diff_weights_memory); + auto diff_dst_memory_4filter_p = + handler.AcquireDiffDstMemoryFromWeightsPrimitive( + user_diff_dst_memory_p, pipeline); - // push primitive and execute it - std::vector pipeline{conv_bwd_weights_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); - } + auto diff_weights_memory_p = + handler.AcquireDiffWeightsMemoryFromWeightsPrimitive( + reinterpret_cast(filter_grad_data)); - if (input_grad) { - // create primitive descriptor - mkldnn::convolution_backward_data::primitive_desc conv_bwd_data_pd = - ConvBwdDataPrimitiveDesc(diff_src_md, weights_md, diff_dst_md, - strides, paddings, *conv_pd, mkldnn_engine); - - // create memory - auto diff_src_memory = mkldnn::memory( - {diff_src_md, mkldnn_engine}, - reinterpret_cast(const_cast(input_grad_data))); - auto weights_memory = - mkldnn::memory({weights_md, mkldnn_engine}, - reinterpret_cast(const_cast(filter_data))); - - // create backward conv primitive for data - auto conv_bwd_data_prim = mkldnn::convolution_backward_data( - conv_bwd_data_pd, diff_dst_memory, weights_memory, diff_src_memory); + auto conv_bwd_weights_p = handler.AcquireConvolutionBackwardWeights( + src_memory_p, diff_dst_memory_4filter_p, diff_weights_memory_p); // push primitive to stream and wait until it's executed - std::vector pipeline{conv_bwd_data_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + pipeline.push_back(*conv_bwd_weights_p); + + filter_grad->set_layout(DataLayout::kMKLDNN); + filter_grad->set_format(GetMKLDNNFormat(*diff_weights_memory_p)); } - } // Compute() - private: - mkldnn::convolution_backward_weights::primitive_desc - ConvBwdWeightsPrimitiveDesc( - const mkldnn::memory::desc& src, const mkldnn::memory::desc& diff_weights, - const mkldnn::memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const mkldnn::convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine) const { - auto conv_bwd_weights_desc = mkldnn::convolution_backward_weights::desc( - mkldnn::convolution_direct, src, diff_weights, diff_dst, strides, - paddings, paddings, mkldnn::padding_kind::zero); - return mkldnn::convolution_backward_weights::primitive_desc( - conv_bwd_weights_desc, engine, conv_pd); - } + if (input_grad) { + auto weights_memory_p = handler.AcquireWeightsMemoryFromDataPrimitive( + user_weights_memory_p, pipeline); - mkldnn::convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc( - const mkldnn::memory::desc& diff_src, const mkldnn::memory::desc& weights, - const mkldnn::memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const mkldnn::convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine) const { - auto conv_bwd_data_desc = mkldnn::convolution_backward_data::desc( - mkldnn::convolution_direct, diff_src, weights, diff_dst, strides, - paddings, paddings, mkldnn::padding_kind::zero); - return mkldnn::convolution_backward_data::primitive_desc(conv_bwd_data_desc, - engine, conv_pd); - } + auto diff_dst_memory_4data_p = + handler.AcquireDiffDstMemoryFromDataPrimitive(user_diff_dst_memory_p, + pipeline); + + auto diff_src_memory_p = handler.AcquireDiffSrcMemoryFromDataPrimitive( + reinterpret_cast(input_grad_data)); + + auto conv_bwd_data_p = handler.AcquireConvolutionBackwardData( + diff_dst_memory_4data_p, weights_memory_p, diff_src_memory_p); + + pipeline.push_back(*conv_bwd_data_p); + + input_grad->set_layout(DataLayout::kMKLDNN); + input_grad->set_format(GetMKLDNNFormat(*diff_src_memory_p)); + } + stream(stream::kind::eager).submit(pipeline).wait(); + } // Compute() }; } // namespace operators diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 697d91484257984b104a13b0572cf19b16f8d37e..37153d58439a90190eb2ad82d5dcc145e22dfa48 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -75,6 +75,10 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { framework::LibraryType library{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library = framework::LibraryType::kCUDNN; @@ -84,6 +88,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif @@ -99,9 +104,6 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( "float16 can only be used when CUDNN is used"); } - std::string data_format = ctx.Attr("data_format"); - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout = framework::StringToDataLayout(data_format); return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, library); } @@ -122,7 +124,8 @@ void Conv2DOpMaker::Make() { "input image channels divided by the groups."); AddOutput("Output", "(Tensor) The output tensor of convolution operator. " - "The format of output tensor is also NCHW."); + "The format of output tensor is also NCHW.") + .Reuse("Input"); AddAttr>("strides", "(vector default:{1, 1}), the " "strides(h_stride, w_stride) of " @@ -217,7 +220,8 @@ void Conv3DOpMaker::Make() { "input image channels divided by the groups."); AddOutput("Output", "(Tensor) The output tensor of convolution operator." - "The format of output tensor is also NCDHW."); + "The format of output tensor is also NCDHW.") + .Reuse("Input"); AddAttr>("strides", "(vector, default:{1, 1, 1}), the " "strides(d_stride, h_stride, w_stride) of " @@ -309,6 +313,10 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOpGrad::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -318,12 +326,10 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType( if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), layout_, library_); diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc index 038ea8999072f562104c5386ed18b6b275816345..82fff68e7557b3f0b44e6faf2a50e5a0ecbba589 100644 --- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc @@ -87,7 +87,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); // Get the algorithm - PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc, // dxDesc: Handle to the previously initialized output tensor // descriptor. @@ -95,7 +95,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { workspace_size_limit, &algo)); // get workspace size able to allocate - PADDLE_ENFORCE( + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc, cudnn_output_desc, algo, &workspace_size_in_bytes)); @@ -110,7 +110,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; for (int g = 0; g < groups; g++) { - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( handle, &alpha, cudnn_filter_desc, filter_data + filter_offset * g, cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes, &beta, @@ -178,11 +178,11 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { auto handle = dev_ctx.cudnn_handle(); if (input_grad) { // choose backward algorithm for data - PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( handle, cudnn_output_desc, cudnn_filter_desc, cudnn_conv_desc, cudnn_input_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, workspace_size_limit, &data_algo)); - PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize( + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize( handle, cudnn_output_desc, cudnn_filter_desc, cudnn_conv_desc, cudnn_input_desc, data_algo, &fwd_ws_size)); workspace_size_in_bytes = std::max(workspace_size_in_bytes, fwd_ws_size); @@ -190,7 +190,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { if (filter_grad) { // choose backward algorithm for filter - PADDLE_ENFORCE( + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( handle, cudnn_output_desc, cudnn_input_desc, cudnn_conv_desc, cudnn_filter_desc, @@ -198,7 +198,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { workspace_size_limit, &filter_algo)); // get workspace for backwards filter algorithm - PADDLE_ENFORCE( + CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( handle, cudnn_output_desc, cudnn_input_desc, cudnn_conv_desc, cudnn_filter_desc, filter_algo, &bwd_filter_ws_size)); @@ -222,7 +222,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. for (int g = 0; g < groups; g++) { - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward( + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward( handle, &alpha, cudnn_output_desc, output_grad_data + output_grad_offset * g, cudnn_filter_desc, filter_data + filter_offset * g, cudnn_conv_desc, data_algo, @@ -237,7 +237,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { // Because beta is zero, it is unnecessary to reset filter_grad. // Gradient with respect to the filter for (int g = 0; g < groups; g++) { - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( handle, &alpha, cudnn_output_desc, output_grad_data + output_grad_offset * g, cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc, filter_algo, diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc index 0b363f5c43f9fc191790e5cca629ffc46eb9388c..eeb98ee44f206dbfbe1f61689aa9843122ae3f92 100644 --- a/paddle/fluid/operators/conv_transpose_op.cc +++ b/paddle/fluid/operators/conv_transpose_op.cc @@ -156,7 +156,7 @@ Parameters(strides, paddings) are two elements. These two elements represent hei and width, respectively. The input(X) size and output(Out) size may be different. -Example: +For an example: Input: Input shape: $(N, C_{in}, H_{in}, W_{in})$ Filter shape: $(C_{in}, C_{out}, H_f, W_f)$ @@ -302,6 +302,7 @@ framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType( namespace ops = paddle::operators; +// conv2d_transpose REGISTER_OPERATOR(conv2d_transpose, ops::ConvTransposeOp, ops::Conv2DTransposeOpMaker, paddle::framework::DefaultGradOpDescMaker); @@ -317,6 +318,7 @@ REGISTER_OP_CPU_KERNEL( ops::GemmConvTransposeGradKernel); +// conv3d_transpose REGISTER_OPERATOR(conv3d_transpose, ops::ConvTransposeOp, ops::Conv3DTransposeOpMaker, paddle::framework::DefaultGradOpDescMaker); @@ -331,3 +333,19 @@ REGISTER_OP_CPU_KERNEL( ops::GemmConvTransposeGradKernel, ops::GemmConvTransposeGradKernel); + +// depthwise conv2d_transpose +REGISTER_OPERATOR(depthwise_conv2d_transpose, ops::ConvTransposeOp, + ops::Conv2DTransposeOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(depthwise_conv2d_transpose_grad, ops::ConvTransposeOpGrad); + +REGISTER_OP_CPU_KERNEL( + depthwise_conv2d_transpose, + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); +REGISTER_OP_CPU_KERNEL( + depthwise_conv2d_transpose_grad, + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); diff --git a/paddle/fluid/operators/conv_transpose_op.cu.cc b/paddle/fluid/operators/conv_transpose_op.cu.cc index 640fa7d14a079debeceb54d8775c4ede7da1b536..a6d5665df83ae5c89d42840e91a6abd853fedd12 100644 --- a/paddle/fluid/operators/conv_transpose_op.cu.cc +++ b/paddle/fluid/operators/conv_transpose_op.cu.cc @@ -15,25 +15,28 @@ limitations under the License. */ #include "paddle/fluid/operators/conv_transpose_op.h" namespace ops = paddle::operators; +using CUDA = paddle::platform::CUDADeviceContext; -REGISTER_OP_CUDA_KERNEL( - conv2d_transpose, - ops::GemmConvTransposeKernel, - ops::GemmConvTransposeKernel); -REGISTER_OP_CUDA_KERNEL( - conv2d_transpose_grad, - ops::GemmConvTransposeGradKernel, - ops::GemmConvTransposeGradKernel); - -REGISTER_OP_CUDA_KERNEL( - conv3d_transpose, - ops::GemmConvTransposeKernel, - ops::GemmConvTransposeKernel); -REGISTER_OP_CUDA_KERNEL( - conv3d_transpose_grad, - ops::GemmConvTransposeGradKernel, - ops::GemmConvTransposeGradKernel); +// conv2d +REGISTER_OP_CUDA_KERNEL(conv2d_transpose, + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); +REGISTER_OP_CUDA_KERNEL(conv2d_transpose_grad, + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); + +// conv3d +REGISTER_OP_CUDA_KERNEL(conv3d_transpose, + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); +REGISTER_OP_CUDA_KERNEL(conv3d_transpose_grad, + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); + +// depthwise conv2d +REGISTER_OP_CUDA_KERNEL(depthwise_conv2d_transpose, + ops::DepthwiseConvTransposeKernel, + ops::DepthwiseConvTransposeKernel); +REGISTER_OP_CUDA_KERNEL(depthwise_conv2d_transpose_grad, + ops::DepthwiseConvTransposeGradKernel, + ops::DepthwiseConvTransposeGradKernel); diff --git a/paddle/fluid/operators/conv_transpose_op.h b/paddle/fluid/operators/conv_transpose_op.h index 1dcfc651fdd79aed50736d05d38ec8576b183d41..0d9c6a62fec1ea24bee5c24b4a7b792781f14d9e 100644 --- a/paddle/fluid/operators/conv_transpose_op.h +++ b/paddle/fluid/operators/conv_transpose_op.h @@ -17,6 +17,7 @@ limitations under the License. */ #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/depthwise_conv.h" #include "paddle/fluid/operators/math/im2col.h" #include "paddle/fluid/operators/math/vol2col.h" @@ -316,5 +317,74 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { } } }; + +template +class DepthwiseConvTransposeKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* input = context.Input("Input"); + Tensor filter = *context.Input("Filter"); + Tensor* output = context.Output("Output"); + output->mutable_data(context.GetPlace()); + + int groups = context.Attr("groups"); + PADDLE_ENFORCE_EQ(groups, filter.dims()[0]); + + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + std::vector dilations = context.Attr>("dilations"); + for (auto v : dilations) { + PADDLE_ENFORCE_EQ(v, 1); + } + + output->mutable_data(context.GetPlace()); + auto& dev_ctx = context.template device_context(); + math::SetConstant set_zero; + set_zero(dev_ctx, output, static_cast(0)); + + math::DepthwiseConvInputGradFunctor + depthwiseConvInputGrad; + depthwiseConvInputGrad(dev_ctx, *output, filter, *input, strides, paddings, + output); + } +}; + +template +class DepthwiseConvTransposeGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* input = context.Input("Input"); + const Tensor* output_grad = + context.Input(framework::GradVarName("Output")); + Tensor* input_grad = + context.Output(framework::GradVarName("Input")); + Tensor* filter_grad = + context.Output(framework::GradVarName("Filter")); + Tensor filter = *context.Input("Filter"); + + if (!input_grad && !filter_grad) return; + + auto& dev_ctx = context.template device_context(); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + + if (input_grad) { + math::DepthwiseConvFunctor depthwiseConv; + depthwiseConv(dev_ctx, *output_grad, filter, strides, paddings, + input_grad); + } + + if (filter_grad) { + math::SetConstant set_zero; + filter_grad->mutable_data(context.GetPlace()); + set_zero(dev_ctx, filter_grad, static_cast(0)); + + math::DepthwiseConvFilterGradFunctor + depthwiseConvFilterGrad; + depthwiseConvFilterGrad(dev_ctx, *output_grad, *input, strides, paddings, + filter_grad); + } + } +}; } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/cos_sim_op.cc b/paddle/fluid/operators/cos_sim_op.cc index 046dd11910bb0ff46b567c3b89883582782205d3..8f3644039f9950a8a70e2fd66c20837a5f52bd7f 100644 --- a/paddle/fluid/operators/cos_sim_op.cc +++ b/paddle/fluid/operators/cos_sim_op.cc @@ -76,9 +76,9 @@ class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { .AsIntermediate(); AddComment(R"DOC( -Cosine Similarity Operator. +**Cosine Similarity Operator** -$Out = X^T * Y / (\sqrt{X^T * X} * \sqrt{Y^T * Y})$ +$Out = \frac{X^T * Y}{(\sqrt{X^T * X} * \sqrt{Y^T * Y})}$ The input X and Y must have the same shape, except that the 1st dimension of input Y could be just 1 (different from input X), which will be diff --git a/paddle/fluid/operators/crf_decoding_op.cc b/paddle/fluid/operators/crf_decoding_op.cc index 40f43936db662f2b18ffa540da4794755b5d6fc7..c27befe1143baa68add4b56f3572eab75272c3a5 100644 --- a/paddle/fluid/operators/crf_decoding_op.cc +++ b/paddle/fluid/operators/crf_decoding_op.cc @@ -53,21 +53,18 @@ sequence of observed tags. The output of this operator changes according to whether Input(Label) is given: 1. Input(Label) is given: - -This happens in training. This operator is used to co-work with the chunk_eval -operator. - -When Input(Label) is given, the crf_decoding operator returns a row vector -with shape [N x 1] whose values are fixed to be 0, indicating an incorrect -prediction, or 1 indicating a tag is correctly predicted. Such an output is the -input to chunk_eval operator. + This happens in training. This operator is used to co-work with the chunk_eval + operator. + When Input(Label) is given, the crf_decoding operator returns a row vector + with shape [N x 1] whose values are fixed to be 0, indicating an incorrect + prediction, or 1 indicating a tag is correctly predicted. Such an output is the + input to chunk_eval operator. 2. Input(Label) is not given: - -This is the standard decoding process. + This is the standard decoding process. The crf_decoding operator returns a row vector with shape [N x 1] whose values -range from 0 to maximum tag number - 1. Each element indicates an index of a +range from 0 to maximum tag number - 1, Each element indicates an index of a predicted tag. )DOC"); } diff --git a/paddle/fluid/operators/crop_op.cc b/paddle/fluid/operators/crop_op.cc index 669b3bbe9df4cae1aa381184092dfa51157ab6a3..a2a871efa850df5101be7c27ebd81456acace7e1 100644 --- a/paddle/fluid/operators/crop_op.cc +++ b/paddle/fluid/operators/crop_op.cc @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -48,6 +48,13 @@ class CropOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", y_dim); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } }; class CropOpMaker : public framework::OpProtoAndCheckerMaker { @@ -60,13 +67,19 @@ class CropOpMaker : public framework::OpProtoAndCheckerMaker { "The input used as reference for cropping, " "which is of the same dimensions as X.") .AsDispensable(); + AddInput("Offsets", + "The input used to describe offsets in runtime, which is a " + "1-D vector whose size equals to the rank of input 'X'. The " + "elements data type must be int.") + .AsDispensable(); AddOutput("Out", "The output of crop op, " "which is of the same dimensions as X."); AddAttr>("offsets", "A list describing offsets to be cropped. " "The size of offsets list should be the same as " - "the dimension size of input X."); + "the dimension size of input X.") + .SetDefault(std::vector()); AddAttr>("shape", "A list describing the shape of output. " "The size of shape list should be the same as " @@ -77,6 +90,17 @@ Crop Operator. Crop input into output, as specified by offsets and shape. +There are two ways to set the offsets: +1. In runtime: Using the input 'Offsets', which is a Vairbale and can be + output of other operators. This way is suitable for + dynamic offsets. +2. In network configuration: Using the attribute 'offsets', which will be + set in Python configure script. This way is + suitable for fixed offsets. +You CANNOT use these two ways at the same time. An exception will be raised +if input 'Offset' is configured and meanwhile the attribute 'offsets' is +not empty. + There are two ways to set shape: 1. reference input: crop input X into the same shape as reference input. The dimension of reference input should @@ -146,6 +170,15 @@ class CropOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(x_grad_name, x_dims); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.Input(framework::GradVarName("Out")) + ->type()), + ctx.device_context()); + } }; } // namespace operators @@ -155,6 +188,7 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(crop, ops::CropOp, ops::CropOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(crop_grad, ops::CropOpGrad); -REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel); +REGISTER_OP_CPU_KERNEL( + crop, ops::CropKernel); REGISTER_OP_CPU_KERNEL( crop_grad, ops::CropGradKernel); diff --git a/paddle/fluid/operators/crop_op.cu b/paddle/fluid/operators/crop_op.cu index 1a391860463dba14ad0de755ceb659bc9f64adc9..b75678217e36aa2297c68a7f8e2a9dfafadaca72 100644 --- a/paddle/fluid/operators/crop_op.cu +++ b/paddle/fluid/operators/crop_op.cu @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -16,6 +16,7 @@ limitations under the License. */ #include "paddle/fluid/operators/crop_op.h" namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel); +REGISTER_OP_CUDA_KERNEL( + crop, ops::CropKernel); REGISTER_OP_CUDA_KERNEL( crop_grad, ops::CropGradKernel); diff --git a/paddle/fluid/operators/crop_op.h b/paddle/fluid/operators/crop_op.h index f05c2e23284e3a24cf48442996f671ec6084c391..2d7d33bd4f9b42b644444912570375bad92ba6c2 100644 --- a/paddle/fluid/operators/crop_op.h +++ b/paddle/fluid/operators/crop_op.h @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -27,36 +27,106 @@ template ; using framework::Tensor; -template +static std::vector GetOffsets(const framework::ExecutionContext& ctx) { + std::vector res; + int rank = ctx.Input("X")->dims().size(); + if (ctx.HasInput("Offsets")) { + PADDLE_ENFORCE(ctx.Attr>("offsets").empty(), + "Input 'Offsets' and attribute 'offsets' should not be used " + "at the same time."); + const auto* offsets_tensor = ctx.Input("Offsets"); + PADDLE_ENFORCE_EQ(offsets_tensor->dims().size(), 1); + PADDLE_ENFORCE_EQ( + rank, offsets_tensor->dims()[0], + "Offsets size should be equal to dimension size of input tensor."); + const int* offsets_data; + framework::Tensor cpu_tmp_tensor; + if (platform::is_cpu_place(offsets_tensor->place())) { + offsets_data = offsets_tensor->data(); + } else { + framework::TensorCopySync(*offsets_tensor, platform::CPUPlace(), + &cpu_tmp_tensor); + offsets_data = cpu_tmp_tensor.data(); + } + res = std::vector(offsets_data, offsets_data + rank); + } else { + res = ctx.Attr>("offsets"); + PADDLE_ENFORCE_EQ( + rank, static_cast(res.size()), + "Offsets size should be equal to dimension size of input tensor."); + } + return res; +} + +template +void CropFunction(const framework::ExecutionContext& context) { + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + auto out_dims = out->dims(); + if (out_dims[0] == -1) { + out_dims[0] = x->dims()[0]; + } + out->mutable_data(out_dims, context.GetPlace()); + auto x_stride = framework::stride(x->dims()); + auto out_stride = framework::stride(out->dims()); + auto offsets = GetOffsets(context); + int64_t offset = 0; + for (size_t i = 0; i < offsets.size(); ++i) { + offset += (x_stride[i] * offsets[i]); + } + + auto x_tensor = EigenTensor::From(*x); + auto out_tensor = EigenTensor::From(*out); + Eigen::array e_offsets; + Eigen::array e_shape; + for (size_t i = 0; i < D; ++i) { + e_offsets[i] = offsets[i]; + e_shape[i] = out->dims()[i]; + } + auto& place = + *context.template device_context().eigen_device(); + out_tensor.device(place) = x_tensor.slice(e_offsets, e_shape); +} + +template class CropKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* x = context.Input("X"); - auto* out = context.Output("Out"); - const T* x_data = x->data(); - T* out_data = out->mutable_data(context.GetPlace()); - auto x_stride = framework::stride(x->dims()); - auto out_stride = framework::stride(out->dims()); - auto offsets = context.Attr>("offsets"); - PADDLE_ENFORCE_EQ( - x->dims().size(), static_cast(offsets.size()), - "Offsets size should be equal to dimension size of input tensor."); - int64_t offset = 0; - for (size_t i = 0; i < offsets.size(); ++i) { - offset += (x_stride[i] * offsets[i]); + int rank = context.Input("X")->dims().size(); + switch (rank) { + case 1: + CropFunction(context); + break; + case 2: + CropFunction(context); + break; + case 3: + CropFunction(context); + break; + case 4: + CropFunction(context); + break; + case 5: + CropFunction(context); + break; + case 6: + CropFunction(context); + break; + default: + PADDLE_THROW( + "CropOp only support tensors with no more than 6 dimensions."); } - StridedMemcpy(context.device_context(), x_data + offset, x_stride, - out->dims(), out_stride, out_data); } }; template void CropGradFunction(const framework::ExecutionContext& context) { auto* d_x = context.Output(framework::GradVarName("X")); + auto* x = context.Input("X"); if (d_x != nullptr) { auto* d_out = context.Input(framework::GradVarName("Out")); - d_x->mutable_data(context.GetPlace()); - auto offsets = context.Attr>("offsets"); + d_x->mutable_data(x->dims(), context.GetPlace()); + auto offsets = GetOffsets(context); Eigen::array, D> paddings; for (size_t i = 0; i < D; ++i) { paddings[i].first = offsets[i]; diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index a3bec3da45136bca5cb2763e7ffd6b67703a1813..578ab63bc380ee62d76e34b7cf3cbd590bfa2eda 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -28,23 +28,26 @@ class CrossEntropyOp : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("X"); auto label_dims = ctx->GetInputDim("Label"); - PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input(X)'s rank should be 2."); - PADDLE_ENFORCE_EQ(label_dims.size(), 2UL, - "Input(Label)'s rank should be 2."); - PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0], - "The 1st dimension of Input(X) and Input(Label) should " - "be equal."); + int rank = x_dims.size(); + PADDLE_ENFORCE_EQ(rank, label_dims.size(), + "Input(X) and Input(Label) shall have the same rank."); + PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1), + framework::slice_ddim(label_dims, 0, rank - 1), + "Input(X) and Input(Label) shall have the same shape " + "except the last dimension."); if (ctx->Attrs().Get("soft_label")) { - PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1], - "If Attr(soft_label) == true, the 2nd dimension of " + PADDLE_ENFORCE_EQ(x_dims[rank - 1], label_dims[rank - 1], + "If Attr(soft_label) == true, the last dimension of " "Input(X) and Input(Label) should be equal."); } else { - PADDLE_ENFORCE_EQ(label_dims[1], 1UL, - "If Attr(softLabel) == false, the 2nd dimension of " + PADDLE_ENFORCE_EQ(label_dims[rank - 1], 1UL, + "If Attr(softLabel) == false, the last dimension of " "Input(Label) should be 1."); } - ctx->SetOutputDim("Y", {x_dims[0], 1}); + auto y_dims = x_dims; + y_dims[rank - 1] = 1; + ctx->SetOutputDim("Y", y_dims); ctx->ShareLoD("X", /*->*/ "Y"); } @@ -74,24 +77,28 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("X"); auto label_dims = ctx->GetInputDim("Label"); auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y")); - PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); - PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2."); - PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2."); - PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0], - "The 1st dimension of Input(X) and Input(Label) should " - "be equal."); - PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0], - "The 1st dimension of Input(X) and Input(Y@Grad) should " - "be equal."); - PADDLE_ENFORCE_EQ(dy_dims[1], 1, - "The 2nd dimension of Input(Y@Grad) should be 1."); + int rank = x_dims.size(); + PADDLE_ENFORCE_EQ(dy_dims.size(), rank, + "Input(Y@Grad) and Input(X) should have the same rank."); + PADDLE_ENFORCE_EQ(label_dims.size(), rank, + "Input(Label) and Input(X) should have the same rank."); + PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1), + framework::slice_ddim(label_dims, 0, rank - 1), + "The Input(X) and Input(Label) should have the same " + "shape except the last dimension."); + PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1), + framework::slice_ddim(dy_dims, 0, rank - 1), + "The Input(X) and Input(Y@Grad) should have the same " + "shape except the last dimension."); + PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1, + "The last dimension of Input(Y@Grad) should be 1."); if (ctx->Attrs().Get("soft_label")) { - PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1], - "When Attr(soft_label) == true, the 2nd dimension of " + PADDLE_ENFORCE_EQ(x_dims[rank - 1], label_dims[rank - 1], + "When Attr(soft_label) == true, the last dimension of " "Input(X) and Input(Label) should be equal."); } else { - PADDLE_ENFORCE_EQ(label_dims[1], 1, - "When Attr(soft_label) == false, the 2nd dimension of " + PADDLE_ENFORCE_EQ(label_dims[rank - 1], 1, + "When Attr(soft_label) == false, the last dimension of " "Input(Label) should be 1."); } ctx->SetOutputDim(framework::GradVarName("X"), x_dims); @@ -113,18 +120,20 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "(Tensor, default Tensor), a 2-D tensor with shape [N x D]," - " where N is the batch size and D is the number of classes. " - "This input is a probability computed by the previous operator, " - "which is almost always the result of a softmax operator."); - AddInput("Label", - "(Tensor), the ground truth which is a 2-D tensor. When " - "soft_label is set to false, Label is a Tensor with shape " - "[N x 1]. When soft_label is set to true, Label is a " - "Tensor with shape [N x D]."); + "(Tensor, default Tensor), a tensor whose last dimension " + "size is equal to the number of classes. This input is a " + "probability computed by the previous operator, which is almost " + "always the result of a softmax operator."); + AddInput( + "Label", + "(Tensor), the tensor which represents the ground truth. It has the " + "same shape with 'X' except the last dimension. When soft_label is set " + "to false, the last dimension size is 1; when soft_label is set to " + "true, the last dimension size is equal to the number of classes."); AddOutput("Y", - "(Tensor, default Tensor), a 2-D tensor with shape " - "[N x 1]. The cross entropy loss."); + "(Tensor, default Tensor), a tensor whose shape is same " + "with 'X' except that the last dimension size is 1. It " + "represents the cross entropy loss."); AddAttr("soft_label", "(bool, default false), a flag indicating whether to " "interpretate the given labels as soft labels.") @@ -132,6 +141,12 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( CrossEntropy Operator. +The input 'X' and 'Label' will first be logically flattened to 2-D matrixs. +The matrix's second dimension(row length) is as same as the original last +dimension, and the first dimension(column length) is the product of all other +original dimensions. Then the softmax computation will take palce on each raw +of flattened matrixs. + It supports both standard cross-entropy and soft-label cross-entropy loss computation. 1) One-hot cross-entropy: diff --git a/paddle/fluid/operators/cross_entropy_op.h b/paddle/fluid/operators/cross_entropy_op.h index 19a2aec92b267ece94685ce34604b7d1cfa5d209..36b58d80144d242277f6fc970a3a61a6721d4b50 100644 --- a/paddle/fluid/operators/cross_entropy_op.h +++ b/paddle/fluid/operators/cross_entropy_op.h @@ -33,8 +33,13 @@ class CrossEntropyOpKernel : public framework::OpKernel { auto* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); + int rank = x->dims().size(); + Tensor x_2d = framework::ReshapeToMatrix(*x, rank - 1); + Tensor labels_2d = framework::ReshapeToMatrix(*labels, rank - 1); + Tensor y_2d = framework::ReshapeToMatrix(*y, rank - 1); + math::CrossEntropyFunctor()( - ctx.template device_context(), y, x, labels, + ctx.template device_context(), &y_2d, &x_2d, &labels_2d, ctx.Attr("soft_label")); } }; @@ -98,9 +103,12 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel { auto* dy = ctx.Input(framework::GradVarName("Y")); auto* label = ctx.Input("Label"); auto* dx = ctx.Output(framework::GradVarName("X")); - auto* dx_data = dx->mutable_data(ctx.GetPlace()); + T* dx_data = dx->mutable_data(ctx.GetPlace()); - int64_t class_num = x->dims()[1]; + // Following computation only depends on the last dimension size. So it's + // unnecessary to convert tensors to 2-D views. + int rank = x->dims().size(); + int64_t class_num = x->dims()[rank - 1]; if (ctx.Attr("soft_label")) { XeSoftlabelGradFunctor functor(dx_data, dy->data(), x->data(), label->data(), diff --git a/paddle/fluid/operators/cumsum_op.cc b/paddle/fluid/operators/cumsum_op.cc index 92bb835e8f18e17ae1355fdec29f43b8ffb70460..5302b822d6b9f232e9ccd0d03cc549d7d5044ebf 100644 --- a/paddle/fluid/operators/cumsum_op.cc +++ b/paddle/fluid/operators/cumsum_op.cc @@ -30,19 +30,19 @@ class CumOp : public framework::OperatorWithKernel { class CumsumOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "Input of Cumsum operator"); - AddOutput("Out", "Output of Cumsum operator"); + AddInput("X", "Input of cumsum operator"); + AddOutput("Out", "Output of cumsum operator"); AddAttr("axis", - "(int, default -1). The dimenstion to accumulate along. " - "-1 means the last dimenstion") + "The dimenstion to accumulate along. -1 means the last " + "dimenstion [default -1].") .SetDefault(-1) .EqualGreaterThan(-1); AddAttr("exclusive", - "bool, default false). Whether to perform exclusive cumsum") + "Whether to perform exclusive cumsum. [default false].") .SetDefault(false); AddAttr("reverse", - "bool, default false). If true, the cumsum is performed in " - "the reversed direction") + "If true, the cumsum is performed in the reversed direction. " + "[default false].") .SetDefault(false); AddComment(R"DOC( The cumulative sum of the elements along a given axis. diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/detail/CMakeLists.txt deleted file mode 100644 index cf20530513cf6cd420e56b2f6378225f73c2bc8b..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ /dev/null @@ -1,12 +0,0 @@ -if(WITH_DISTRIBUTE) - grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc - request_handler_impl.cc rpc_server.cc grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor - selected_rows memory) - set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - set_source_files_properties(serde_test.cc grpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(serde_test SRCS serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr - cares zlib protobuf sendrecvop_grpc SERIAL) - cc_test(grpc_server_test SRCS grpc_server_test.cc DEPS sendrecvop_grpc - grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor - proto_desc lookup_table_op SERIAL) -endif() diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.cc b/paddle/fluid/operators/detail/bytebuffer_stream.cc deleted file mode 100644 index a14171563edb0ac9a22b7ae493c965de3efb7823..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/bytebuffer_stream.cc +++ /dev/null @@ -1,88 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -// NOTE: This file was originally created by tensorflow -// (https://github.com/tensorflow/tensorflow/) we borrow this -// file and did some modifications so that we can send gRPC -// requests without too much copying of the tensor data. - -#include "paddle/fluid/operators/detail/bytebuffer_stream.h" - -namespace paddle { -namespace operators { -namespace detail { - -GrpcByteBufferSource::GrpcByteBufferSource() {} - -bool GrpcByteBufferSource::Init(const grpc::ByteBuffer& src) { - cur_ = -1; - left_ = 0; - ptr_ = nullptr; - byte_count_ = 0; - bool ok = src.Dump(&slices_).ok(); - if (!ok) { - slices_.clear(); - } - return ok; -} - -bool GrpcByteBufferSource::Next(const void** data, int* size) { - // Use loop instead of if in case buffer contained empty slices. - while (left_ == 0) { - // Advance to next slice. - cur_++; - if (cur_ >= slices_.size()) { - return false; - } - const ::grpc::Slice& s = slices_[cur_]; - left_ = s.size(); - ptr_ = reinterpret_cast(s.begin()); - } - - *data = ptr_; - *size = left_; - byte_count_ += left_; - ptr_ += left_; - left_ = 0; - return true; -} - -void GrpcByteBufferSource::BackUp(int count) { - ptr_ -= count; - left_ += count; - byte_count_ -= count; -} - -bool GrpcByteBufferSource::Skip(int count) { - const void* data; - int size; - while (Next(&data, &size)) { - if (size >= count) { - BackUp(size - count); - return true; - } - // size < count; - count -= size; - } - // error or we have too large count; - return false; -} - -google::protobuf::int64 GrpcByteBufferSource::ByteCount() const { - return byte_count_; -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.h b/paddle/fluid/operators/detail/bytebuffer_stream.h deleted file mode 100644 index 054dd4ff294414cca55d7e033f2c5403bbb85526..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/bytebuffer_stream.h +++ /dev/null @@ -1,188 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -// NOTE: This file was originally created by tensorflow -// (https://github.com/tensorflow/tensorflow/) we borrow this -// file and did some modifications so that we can send gRPC -// requests without too much copying of the tensor data. - -#pragma once - -#include - -#include "google/protobuf/io/coded_stream.h" -#include "google/protobuf/io/zero_copy_stream.h" -#include "grpc++/grpc++.h" - -namespace grpc { -// A ZeroCopyInputStream that reads from grpc_byte_buffer -class GrpcBufferReader final - : public ::google::protobuf::io::ZeroCopyInputStream { - typedef void (CoreCodegenInterface::*OldReaderInitAPI)( - grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer); - typedef int (CoreCodegenInterface::*NewReaderInitAPI)( - grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer); - void ReaderInit(OldReaderInitAPI ptr, grpc_byte_buffer_reader* reader, - grpc_byte_buffer* buffer) { - (g_core_codegen_interface->*ptr)(reader, buffer); - } - void ReaderInit(NewReaderInitAPI ptr, grpc_byte_buffer_reader* reader, - grpc_byte_buffer* buffer) { - int result = (g_core_codegen_interface->*ptr)(reader, buffer); - (void)result; - } - - public: - explicit GrpcBufferReader(grpc_byte_buffer* buffer) - : byte_count_(0), backup_count_(0) { - ReaderInit(&CoreCodegenInterface::grpc_byte_buffer_reader_init, &reader_, - buffer); - } - ~GrpcBufferReader() override { - g_core_codegen_interface->grpc_byte_buffer_reader_destroy(&reader_); - } - - bool Next(const void** data, int* size) override { - if (backup_count_ > 0) { - *data = GRPC_SLICE_START_PTR(slice_) + GRPC_SLICE_LENGTH(slice_) - - backup_count_; - GPR_CODEGEN_ASSERT(backup_count_ <= INT_MAX); - *size = static_cast(backup_count_); - backup_count_ = 0; - return true; - } - if (!g_core_codegen_interface->grpc_byte_buffer_reader_next(&reader_, - &slice_)) { - return false; - } - g_core_codegen_interface->grpc_slice_unref(slice_); - *data = GRPC_SLICE_START_PTR(slice_); - // On win x64, int is only 32bit - GPR_CODEGEN_ASSERT(GRPC_SLICE_LENGTH(slice_) <= INT_MAX); - byte_count_ += * size = static_cast(GRPC_SLICE_LENGTH(slice_)); - return true; - } - - void BackUp(int count) override { backup_count_ = count; } - - bool Skip(int count) override { - const void* data; - int size; - while (Next(&data, &size)) { - if (size >= count) { - BackUp(size - count); - return true; - } - // size < count; - count -= size; - } - // error or we have too large count; - return false; - } - - ::google::protobuf::int64 ByteCount() const override { - return byte_count_ - backup_count_; - } - - private: - int64_t byte_count_; - int64_t backup_count_; - grpc_byte_buffer_reader reader_; - grpc_slice slice_; -}; - -}; // namespace grpc - -namespace paddle { -namespace operators { -namespace detail { -// Source provides a way for a particular RPC implementation to provide -// received data to ParseFrom. -class Source { - public: - virtual ~Source() {} - - // Return the stream that contains the data to be parsed. - // Note that this method might be invoked more than once if - // ParseFrom needs to fall back to a more expensive parsing method. - // Every call must return a stream pointing at the beginning of - // the serialized RecvTensorResponse. - // - // Note that a subsequent call to contents() invalidates previous - // results of contents(). - // - // Ownership of the returned stream is retained by the Source and - // should not be deleted by the caller. - virtual ::google::protobuf::io::ZeroCopyInputStream* contents() = 0; -}; - -// A ZeroCopyInputStream that reads from a grpc::ByteBuffer. -class GrpcByteBufferSource - : public ::google::protobuf::io::ZeroCopyInputStream { - public: - GrpcByteBufferSource(); - bool Init(const ::grpc::ByteBuffer& src); // Can be called multiple times. - bool Next(const void** data, int* size) override; - void BackUp(int count) override; - bool Skip(int count) override; - ::google::protobuf::int64 ByteCount() const override; - - private: - std::vector<::grpc::Slice> slices_; - size_t cur_; // Current slice index. - int left_; // Number of bytes in slices_[cur_] left to yield. - const char* ptr_; // Address of next byte in slices_[cur_] to yield. - ::google::protobuf::int64 byte_count_; -}; - -class GrpcByteBufferSourceWrapper : public Source { - public: - explicit GrpcByteBufferSourceWrapper(GrpcByteBufferSource* source) - : source_(source) {} - ::google::protobuf::io::ZeroCopyInputStream* contents() override { - return source_; - } - - private: - GrpcByteBufferSource* source_; -}; - -class GrpcByteSource : public Source { - public: - explicit GrpcByteSource(grpc_byte_buffer* buffer) : buffer_(buffer) {} - ~GrpcByteSource() override { DeleteStream(); } - - typedef ::grpc::GrpcBufferReader Reader; - - ::google::protobuf::io::ZeroCopyInputStream* contents() override { - DeleteStream(); - stream_ = new (&space_) Reader(buffer_); - return stream_; - } - - private: - void DeleteStream() { - if (stream_) { - stream_->~Reader(); - } - } - - grpc_byte_buffer* buffer_; // Not owned - Reader* stream_ = nullptr; // Points into space_ if non-nullptr - char space_[sizeof(Reader)]; -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc deleted file mode 100644 index da9ca1a0c1d55018141f0e4285fe35d7c437fd55..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ /dev/null @@ -1,289 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/detail/grpc_client.h" - -#include - -#include - -#include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/platform/profiler.h" - -namespace paddle { -namespace operators { -namespace detail { - -std::once_flag RPCClient::init_flag_; - -std::unique_ptr RPCClient::rpc_client_(nullptr); - -RPCClient* RPCClient::GetInstance() { - std::call_once(init_flag_, &RPCClient::Init); - return rpc_client_.get(); -} - -void RPCClient::Init() { - if (rpc_client_.get() == nullptr) { - rpc_client_.reset(new RPCClient()); - } -} - -bool RPCClient::AsyncSendVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out) { - const platform::DeviceContext* p_ctx = &ctx; - const std::string ep_val = ep; - const std::string var_name_val = var_name; - const framework::Scope* p_scope = &scope; - const auto ch = GetChannel(ep_val); - - framework::AsyncIO([var_name_val, p_ctx, ep_val, p_scope, time_out, ch, - this] { - auto* var = p_scope->FindVar(var_name_val); - - ::grpc::ByteBuffer req; - SerializeToByteBuffer(var_name_val, var, *p_ctx, &req); - - // varhandle - VarHandle var_h; - var_h.ep = ep_val; - var_h.scope = p_scope; - var_h.name = var_name_val; - var_h.ctx = p_ctx; - - // stub context - SendProcessor* s = new SendProcessor(ch); - s->Prepare(var_h, time_out); - s->response_call_back_ = nullptr; - - auto call = s->stub_g_.PrepareUnaryCall( - s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); - call->StartCall(); - call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); - }); - req_count_++; - - return true; -} - -void ProcGetResponse(const VarHandle& var_h, - const ::grpc::ByteBuffer& ret_msg) { - framework::Variable* outvar = nullptr; - DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar); -} - -template -void RequestToByteBuffer(const T& proto, ::grpc::ByteBuffer* result) { - ::grpc::Slice slice(proto.ByteSizeLong()); - proto.SerializeWithCachedSizesToArray(const_cast(slice.begin())); - ::grpc::ByteBuffer tmp(&slice, 1); - result->Swap(&tmp); -} - -bool RPCClient::AsyncGetVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out) { - const platform::DeviceContext* p_ctx = &ctx; - const std::string ep_val = ep; - const std::string var_name_val = var_name; - const framework::Scope* p_scope = &scope; - const auto ch = GetChannel(ep_val); - - framework::AsyncIO([var_name_val, ep_val, p_scope, p_ctx, time_out, ch, - this] { - // prepare input - sendrecv::VariableMessage req; - req.set_varname(var_name_val); - ::grpc::ByteBuffer buf; - RequestToByteBuffer(req, &buf); - - // var handle - VarHandle var_h; - var_h.ep = ep_val; - var_h.scope = p_scope; - var_h.name = var_name_val; - var_h.ctx = p_ctx; - - // stub context - GetProcessor* s = new GetProcessor(ch); - s->Prepare(var_h, time_out); - s->response_call_back_ = ProcGetResponse; - - auto call = s->stub_g_.PrepareUnaryCall( - s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); - call->StartCall(); - call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); - }); - - req_count_++; - - return true; -} - -bool RPCClient::AsyncPrefetchVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& in_var_name, - const std::string& out_var_name, - int64_t time_out) { - const platform::DeviceContext* p_ctx = &ctx; - const std::string ep_val = ep; - const std::string in_var_name_val = in_var_name; - const std::string out_var_name_val = out_var_name; - const framework::Scope* p_scope = &scope; - const auto ch = GetChannel(ep_val); - - framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, - time_out, ch, this] { - auto* var = p_scope->FindVar(in_var_name_val); - - ::grpc::ByteBuffer req; - SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req, out_var_name_val); - - // var handle - VarHandle var_h; - var_h.ep = ep_val; - var_h.scope = p_scope; - var_h.name = out_var_name_val; - var_h.ctx = p_ctx; - - // stub context - GetProcessor* s = new GetProcessor(ch); - s->Prepare(var_h, time_out); - s->response_call_back_ = ProcGetResponse; - - auto call = s->stub_g_.PrepareUnaryCall( - s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, - &cq_); - call->StartCall(); - call->Finish(&s->reply_, &s->status_, static_cast(s)); - }); - - req_count_++; - return true; -} - -void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { - const auto ch = GetChannel(ep); - - BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); - s->Prepare(time_out); - - sendrecv::VariableMessage req; - req.set_varname(BATCH_BARRIER_MESSAGE); - auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); - rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); - req_count_++; -} - -void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { - const auto ch = GetChannel(ep); - FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); - s->Prepare(time_out); - - sendrecv::VariableMessage req; - req.set_varname(FETCH_BARRIER_MESSAGE); - auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); - rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); - req_count_++; -} - -bool RPCClient::Wait() { - VLOG(3) << "RPCClient begin Wait()" - << " req_count_:" << req_count_; - if (req_count_ <= 0) { - return true; - } - const size_t kReqCnt = req_count_; - bool a[kReqCnt]; - std::vector> waits(req_count_); - std::mutex mu; - - for (int i = 0; i < req_count_; i++) { - waits[i] = framework::AsyncIO([i, &a, &mu, this] { - bool ret = Proceed(); - std::lock_guard l(mu); - a[i] = ret; - }); - } - - for (int i = 0; i < req_count_; i++) { - waits[i].wait(); - } - - int last_req_count = req_count_; - req_count_ = 0; - - for (int i = 0; i < last_req_count; i++) { - if (!a[i]) { - return false; - } - } - - return true; -} - -bool RPCClient::Proceed() { - void* tag = NULL; - bool ok = false; - - // request counts. - if (!cq_.Next(&tag, &ok)) { - LOG(ERROR) << "Get meets CompletionQueue error"; - return false; - } - - GPR_ASSERT(ok); - PADDLE_ENFORCE(tag); - - // TODO(gongwb): add more retries. - BaseProcessor* c = static_cast(tag); - if (!c->status_.ok()) { - LOG(ERROR) << "proc param error:" << c->var_h_.String() - << " grpc error:" << c->status_.error_message(); - delete c; - return false; - } - - c->Process(); - delete c; - return true; -} -std::shared_ptr RPCClient::GetChannel(const std::string& ep) { - // TODO(Yancey1989): make grpc client completely thread-safe - std::unique_lock lock(mutex_); - auto it = channels_.find(ep); - if (it != channels_.end()) { - return it->second; - } - - grpc::ChannelArguments args; - args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE); - args.SetMaxSendMessageSize(std::numeric_limits::max()); - args.SetMaxReceiveMessageSize(std::numeric_limits::max()); - - auto ch = - grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args); - channels_[ep] = ch; - return ch; -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h deleted file mode 100644 index 449d5105afb8c02294a0ef57610e7de1b1631b35..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/grpc_client.h +++ /dev/null @@ -1,215 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include - -#include // NOLINT -#include -#include -#include -#include -#include // NOLINT -#include -#include - -#include "grpc++/generic/generic_stub.h" -#include "grpc++/grpc++.h" -#include "grpc++/support/byte_buffer.h" -#include "grpc++/support/slice.h" -#include "grpc/support/log.h" -#include "paddle/fluid/framework/blocking_queue.h" -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN - -namespace paddle { -namespace operators { -namespace detail { - -struct VarHandle { - std::string ep; - const platform::DeviceContext* ctx; - const framework::Scope* scope; - std::string name; - - std::string String() const { - std::ostringstream s; - s << "name:[" << name << "] ep:[" << ep << "]"; - return s.str(); - } -}; - -void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg); - -class BaseProcessor { - public: - explicit BaseProcessor(std::shared_ptr ch) { - context_ = nullptr; - } - - virtual ~BaseProcessor() {} - - virtual void Prepare(const VarHandle& var_info, int64_t time_out) { - context_.reset(new grpc::ClientContext()); - var_h_ = var_info; - - std::chrono::system_clock::time_point deadline = - std::chrono::system_clock::now() + std::chrono::milliseconds(time_out); - - context_->set_deadline(deadline); - } - - virtual void Prepare(int64_t time_out) { - context_.reset(new grpc::ClientContext()); - - std::chrono::system_clock::time_point deadline = - std::chrono::system_clock::now() + std::chrono::milliseconds(time_out); - - context_->set_deadline(deadline); - } - - virtual void Process() = 0; - - std::unique_ptr context_; - grpc::Status status_; - VarHandle var_h_; -}; - -typedef std::function - RequestSendCallBack; - -class SendProcessor : public BaseProcessor { - public: - explicit SendProcessor(std::shared_ptr ch) - : BaseProcessor(ch), stub_g_(ch) {} - - virtual ~SendProcessor() {} - - virtual void Process() { - if (response_call_back_) { - response_call_back_(var_h_, reply_); - } - } - - ::grpc::GenericStub stub_g_; - ::grpc::ByteBuffer reply_; - RequestSendCallBack response_call_back_ = nullptr; -}; - -typedef std::function - RequestGetCallBack; - -class GetProcessor : public BaseProcessor { - public: - explicit GetProcessor(std::shared_ptr ch) - : BaseProcessor(ch), stub_g_(ch) {} - - virtual ~GetProcessor() {} - - virtual void Process() { - if (response_call_back_) { - response_call_back_(var_h_, reply_); - } - } - - ::grpc::ByteBuffer reply_; - ::grpc::GenericStub stub_g_; - RequestGetCallBack response_call_back_ = ProcGetResponse; -}; - -class BatchBarrierProcessor : public BaseProcessor { - public: - explicit BatchBarrierProcessor(std::shared_ptr ch) - : BaseProcessor(ch) { - stub_ = sendrecv::SendRecvService::NewStub(ch); - } - - virtual ~BatchBarrierProcessor() {} - - virtual void Process() {} - sendrecv::VoidMessage reply_; - std::unique_ptr stub_; -}; - -class FetchBarrierProcessor : public BaseProcessor { - public: - explicit FetchBarrierProcessor(std::shared_ptr ch) - : BaseProcessor(ch) { - stub_ = sendrecv::SendRecvService::NewStub(ch); - } - - virtual ~FetchBarrierProcessor() {} - - virtual void Process() {} - sendrecv::VariableMessage reply_; - std::unique_ptr stub_; -}; - -class RPCClient { - public: - RPCClient() {} - - static RPCClient* GetInstance(); - - bool AsyncSendVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out = 600 * 1000); - - bool AsyncGetVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out = 600 * 1000); - - bool AsyncPrefetchVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& in_var_name, - const std::string& out_var_name, - int64_t time_out = 600 * 1000); - - void AsyncSendBatchBarrier(const std::string& ep, - int64_t time_out = 600 * 1000); - - void AsyncSendFetchBarrier(const std::string& ep, - int64_t time_out = 600 * 1000); - - bool Wait(); - - private: - bool Proceed(); - std::shared_ptr GetChannel(const std::string& ep); - // Init is called by GetInstance. - static void Init(); - - private: - grpc::CompletionQueue cq_; - std::map> channels_; - std::atomic req_count_{0}; - std::mutex mutex_; - static std::unique_ptr rpc_client_; - static std::once_flag init_flag_; - DISABLE_COPY_AND_ASSIGN(RPCClient); -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc deleted file mode 100644 index e73756d89004bc48339c0aa31dd0857c2ca6722d..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ /dev/null @@ -1,356 +0,0 @@ -/*Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include "paddle/fluid/operators/detail/grpc_server.h" - -using ::grpc::ServerAsyncResponseWriter; - -namespace paddle { -namespace operators { -namespace detail { -enum CallStatus { PROCESS = 0, FINISH }; - -// reference: -// https://stackoverflow.com/questions/41732884/grpc-multiple-services-in-cpp-async-server -class RequestBase { - public: - explicit RequestBase(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, - RequestHandler* request_handler, int req_id) - : service_(service), - cq_(cq), - status_(PROCESS), - request_handler_(request_handler), - req_id_(req_id) { - PADDLE_ENFORCE(cq_); - } - virtual ~RequestBase() {} - virtual void Process() = 0; - - CallStatus Status() { return status_; } - void SetStatus(CallStatus status) { status_ = status; } - virtual std::string GetReqName() = 0; - - protected: - ::grpc::ServerContext ctx_; - GrpcService::AsyncService* service_; - ::grpc::ServerCompletionQueue* cq_; - CallStatus status_; - RequestHandler* request_handler_; - int req_id_; -}; - -class RequestSend final : public RequestBase { - public: - explicit RequestSend(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, - RequestHandler* request_handler, int req_id) - : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { - request_.reset(new VariableResponse(request_handler->scope(), - request_handler->dev_ctx(), - !request_handler->sync_mode())); - int method_id = static_cast(detail::GrpcMethod::kSendVariable); - service_->RequestAsyncUnary( - method_id, &ctx_, request_.get(), &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id))); - } - - virtual ~RequestSend() {} - - std::string GetReqName() override { return request_->Varname(); } - - void Process() override { - std::string varname = GetReqName(); - VLOG(3) << "RequestSend var_name:" << varname; - - auto scope = request_->GetMutableLocalScope(); - auto invar = request_->GetVar(); - framework::Variable* outvar = nullptr; - - request_handler_->Handle(varname, scope, invar, &outvar); - - status_ = FINISH; - responder_.Finish(reply_, ::grpc::Status::OK, - reinterpret_cast(static_cast(req_id_))); - } - - protected: - sendrecv::VoidMessage reply_; - std::shared_ptr request_; - ServerAsyncResponseWriter responder_; -}; - -class RequestGet final : public RequestBase { - public: - explicit RequestGet(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, - RequestHandler* request_handler, int req_id) - : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { - auto method_id = static_cast(detail::GrpcMethod::kGetVariable); - service_->RequestAsyncUnary( - method_id, &ctx_, &request_, &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id))); - } - - virtual ~RequestGet() {} - - std::string GetReqName() override { return request_.varname(); } - - void Process() override { - // proc request. - std::string varname = request_.varname(); - VLOG(3) << "RequestGet " << varname; - - auto scope = request_handler_->scope(); - auto invar = scope->FindVar(varname); - framework::Variable* outvar = nullptr; - - request_handler_->Handle(varname, scope, invar, &outvar); - - if (outvar) { - SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), - &reply_); - } - - status_ = FINISH; - responder_.Finish(reply_, ::grpc::Status::OK, - reinterpret_cast(static_cast(req_id_))); - } - - protected: - sendrecv::VariableMessage request_; - ::grpc::ByteBuffer reply_; - ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; -}; - -class RequestPrefetch final : public RequestBase { - public: - explicit RequestPrefetch(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, - RequestHandler* request_handler, int req_id) - : RequestBase(service, cq, request_handler, req_id), - responder_(&ctx_), - local_scope_(nullptr) { - request_.reset(new VariableResponse(request_handler->scope(), - request_handler->dev_ctx(), true)); - int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); - service_->RequestAsyncUnary( - method_id, &ctx_, request_.get(), &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id))); - } - - virtual ~RequestPrefetch() {} - - std::string GetReqName() override { return request_->Varname(); } - - void Process() override { - // prefetch process... - std::string varname = request_->OutVarname(); - VLOG(3) << "RequestPrefetch " << varname; - - auto scope = request_->GetMutableLocalScope(); - auto invar = scope->FindVar(varname); - framework::Variable* outvar = nullptr; - - request_handler_->Handle(varname, scope, invar, &outvar); - - SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), - &reply_); - - status_ = FINISH; - responder_.Finish(reply_, ::grpc::Status::OK, - reinterpret_cast(static_cast(req_id_))); - } - - protected: - std::shared_ptr request_; - ::grpc::ByteBuffer reply_; - ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; - framework::Scope* local_scope_; -}; - -void AsyncGRPCServer::WaitServerReady() { - VLOG(3) << "AsyncGRPCServer is wait server ready"; - std::unique_lock lock(this->mutex_ready_); - condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); - VLOG(3) << "AsyncGRPCServer WaitSeverReady"; -} - -void AsyncGRPCServer::StartServer() { - ::grpc::ServerBuilder builder; - builder.AddListeningPort(bind_address_, ::grpc::InsecureServerCredentials(), - &selected_port_); - - builder.SetMaxSendMessageSize(std::numeric_limits::max()); - builder.SetMaxReceiveMessageSize(std::numeric_limits::max()); - builder.RegisterService(&service_); - - for (auto t : rpc_call_map_) { - rpc_cq_[t.first].reset(builder.AddCompletionQueue().release()); - } - - server_ = builder.BuildAndStart(); - LOG(INFO) << "Server listening on " << bind_address_ - << " selected port: " << selected_port_; - - std::function f = - std::bind(&AsyncGRPCServer::TryToRegisterNewOne, this, - std::placeholders::_1, std::placeholders::_2); - - for (auto& t : rpc_call_map_) { - auto& rpc_name = t.first; - auto& cq = rpc_cq_[rpc_name]; - auto threadnum = rpc_thread_num_[rpc_name]; - auto& reqs = rpc_reqs_[rpc_name]; - - reqs.reserve(kRequestBufSize); - - for (int i = 0; i < kRequestBufSize; i++) { - TryToRegisterNewOne(rpc_name, i); - } - - for (int i = 0; i < threadnum; i++) { - rpc_threads_[rpc_name].emplace_back(new std::thread(std::bind( - &AsyncGRPCServer::HandleRequest, this, cq.get(), rpc_name, f))); - VLOG(3) << t.first << " creates threads!"; - } - } - - { - std::lock_guard lock(this->mutex_ready_); - ready_ = 1; - } - condition_ready_.notify_all(); - - // wait server - server_->Wait(); - - for (auto& t : rpc_threads_) { - auto& threads = t.second; - for (size_t i = 0; i < threads.size(); ++i) { - threads[i]->join(); - VLOG(3) << t.first << " threads ends!"; - } - } -} - -void AsyncGRPCServer::ShutdownQueue() { - for (auto& t : rpc_cq_) { - t.second->Shutdown(); - VLOG(3) << t.first << " shutdown!"; - } -} - -void AsyncGRPCServer::ShutDownImpl() { - std::unique_lock lock(cq_mutex_); - is_shut_down_ = true; - ShutdownQueue(); - - VLOG(3) << "server_ shutdown!"; - server_->Shutdown(); -} - -void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, - int req_id) { - std::unique_lock lock(cq_mutex_); - if (is_shut_down_) { - VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; - return; - } - - VLOG(4) << "register send rpc_name:" << rpc_name - << ", handler:" << rpc_call_map_[kRequestSend]; - - auto& reqs = rpc_reqs_[rpc_name]; - auto& handler = rpc_call_map_[rpc_name]; - auto& cq = rpc_cq_[rpc_name]; - - RequestBase* b = nullptr; - if (rpc_name == kRequestSend) { - b = new RequestSend(&service_, cq.get(), handler, req_id); - } else if (rpc_name == kRequestGet) { - b = new RequestGet(&service_, cq.get(), handler, req_id); - } else if (rpc_name == kRequestPrefetch) { - b = new RequestPrefetch(&service_, cq.get(), handler, req_id); - } else { - PADDLE_ENFORCE(false, "not surpported rpc"); - } - - reqs[req_id] = b; - - VLOG(4) << "Create RequestSend status:" << b->Status(); -} - -void AsyncGRPCServer::HandleRequest( - ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, - std::function TryToRegisterNewOne) { - void* tag = NULL; - bool ok = false; - - while (true) { - VLOG(3) << "HandleRequest " << rpc_name << " wait next"; - if (!cq->Next(&tag, &ok)) { - LOG(INFO) << "CompletionQueue " << rpc_name << " shutdown!"; - break; - } - - int req_id = static_cast(reinterpret_cast(tag)); - VLOG(3) << "HandleRequest " << rpc_name << ", req_id:" << req_id - << " get next"; - - auto& reqs = rpc_reqs_[rpc_name]; - RequestBase* base = nullptr; - { - PADDLE_ENFORCE(req_id >= 0 && req_id < kRequestBufSize); - std::unique_lock lock(cq_mutex_); - base = reqs[req_id]; - } - - // reference: - // https://github.com/tensorflow/tensorflow/issues/5596 - // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM - // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I - if (!ok) { - LOG(WARNING) << "completion queue:" << rpc_name - << " recv no regular event:argument name[" - << base->GetReqName() << "]"; - TryToRegisterNewOne(rpc_name, req_id); - delete base; - continue; - } - - VLOG(3) << "queue id:" << rpc_name << ", req_id:" << req_id - << ", status:" << base->Status(); - - switch (base->Status()) { - case PROCESS: { - base->Process(); - break; - } - case FINISH: { - TryToRegisterNewOne(rpc_name, req_id); - delete base; - break; - } - default: { assert(false); } - } - } -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h deleted file mode 100644 index d1fcbc414f123c5c4810d9cecf807a406aa2c405..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/grpc_server.h +++ /dev/null @@ -1,90 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include -#include // NOLINT -#include -#include - -#include "grpc++/grpc++.h" -#include "paddle/fluid/framework/blocking_queue.h" -#include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/grpc_service.h" -#include "paddle/fluid/operators/detail/request_handler.h" -#include "paddle/fluid/operators/detail/rpc_server.h" -#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/platform/profiler.h" - -namespace paddle { -namespace operators { -namespace detail { - -class RequestBase; - -class AsyncGRPCServer final : public RPCServer { - public: - explicit AsyncGRPCServer(const std::string& address, int client_num) - : RPCServer(address, client_num), ready_(0) {} - - virtual ~AsyncGRPCServer() {} - void WaitServerReady() override; - void StartServer() override; - - private: - void HandleRequest( - ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, - std::function TryToRegisterNewOne); - - void TryToRegisterNewOne(const std::string& rpc_name, int req_id); - void ShutdownQueue(); - void ShutDownImpl() override; - - private: - static const int kRequestBufSize = 100; - - std::mutex cq_mutex_; - volatile bool is_shut_down_ = false; - - GrpcService::AsyncService service_; - std::unique_ptr<::grpc::Server> server_; - - // condition of the sub program - std::mutex barrier_mutex_; - mutable int barrier_cond_step_; - std::condition_variable barrier_condition_; - - std::mutex mutex_ready_; - std::condition_variable condition_ready_; - - int ready_; - - std::map> rpc_cq_; - std::map>> rpc_threads_; - std::map> rpc_reqs_; -}; - -}; // namespace detail -}; // namespace operators -}; // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/grpc_server_test.cc deleted file mode 100644 index f97f638701cfb263f28dddbdc3bc80fb16468744..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/grpc_server_test.cc +++ /dev/null @@ -1,159 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include // NOLINT - -#include "gtest/gtest.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/detail/grpc_server.h" - -#include "paddle/fluid/framework/block_desc.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/operator.h" - -#include "paddle/fluid/operators/detail/request_handler_impl.h" - -namespace framework = paddle::framework; -namespace platform = paddle::platform; -namespace detail = paddle::operators::detail; - -USE_OP(lookup_table); - -std::unique_ptr g_rpc_service; -std::unique_ptr g_req_handler; - -framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) { - auto root_block = program->MutableBlock(0); - auto* block = program->AppendBlock(*root_block); - - framework::VariableNameMap input({{"W", {"w"}}, {"Ids", {"ids"}}}); - framework::VariableNameMap output({{"Output", {"out"}}}); - auto op = block->AppendOp(); - op->SetType("lookup_table"); - op->SetInput("W", {"w"}); - op->SetInput("Ids", {"ids"}); - op->SetOutput("Out", {"out"}); - - auto& out = *root_block->Var("out"); - out.SetType(framework::proto::VarType::SELECTED_ROWS); - out.SetShape({10, 10}); - - return block; -} - -void CreateVarsOnScope(framework::Scope* scope, platform::CPUPlace* place) { - auto w_var = scope->Var("w"); - w_var->GetMutable(); - - auto out_var = scope->Var("out"); - out_var->GetMutable(); - - auto ids_var = scope->Var("ids"); - ids_var->GetMutable(); -} - -void InitTensorsOnClient(framework::Scope* scope, platform::CPUPlace* place, - int64_t rows_numel) { - CreateVarsOnScope(scope, place); - auto ids_var = scope->Var("ids")->GetMutable(); - auto rows = ids_var->mutable_rows(); - for (int64_t i = 0; i < rows_numel; ++i) rows->push_back(i * 2); - ids_var->mutable_value()->Resize({rows_numel, 1}); - ids_var->mutable_value()->mutable_data(*place); -} - -void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place, - int64_t rows_numel) { - CreateVarsOnScope(scope, place); - auto w = scope->Var("w")->GetMutable(); - auto rows = w->mutable_rows(); - for (int64_t i = 0; i < rows_numel; ++i) rows->push_back(i); - auto w_value = w->mutable_value(); - w_value->Resize({rows_numel, 10}); - - auto ptr = w_value->mutable_data(*place); - - for (int64_t i = 0; i < w_value->numel(); ++i) { - ptr[i] = static_cast(i / 10); - } -} - -void StartServer() { - framework::ProgramDesc program; - framework::Scope scope; - platform::CPUPlace place; - framework::Executor exe(place); - platform::CPUDeviceContext ctx(place); - auto* block = AppendPrefetchBlcok(&program); - auto prepared = exe.Prepare(program, block->ID()); - InitTensorsOnServer(&scope, &place, 10); - - g_req_handler->SetProgram(&program); - g_req_handler->SetPrefetchPreparedCtx(std::move(prepared)); - g_req_handler->SetDevCtx(&ctx); - g_req_handler->SetScope(&scope); - g_req_handler->SetExecutor(&exe); - - g_rpc_service->RegisterRPC(detail::kRequestPrefetch, g_req_handler.get()); - g_req_handler->SetRPCServer(g_rpc_service.get()); - - std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get())); - - // FIXME(gongwb): don't use hard time. - sleep(10); - LOG(INFO) << "got nccl id and stop server..."; - g_rpc_service->ShutDown(); - server_thread.join(); -} - -TEST(PREFETCH, CPU) { - g_req_handler.reset(new detail::RequestPrefetchHandler(true)); - g_rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", 1)); - - std::thread server_thread(StartServer); - g_rpc_service->WaitServerReady(); - - detail::RPCClient client; - int port = g_rpc_service->GetSelectedPort(); - std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port); - - framework::Scope scope; - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); - { - // create var on local scope - int64_t rows_numel = 5; - InitTensorsOnClient(&scope, &place, rows_numel); - std::string in_var_name("ids"); - std::string out_var_name("out"); - - client.AsyncPrefetchVariable(ep, ctx, scope, in_var_name, out_var_name); - client.Wait(); - auto var = scope.Var(out_var_name); - auto value = var->GetMutable()->value(); - auto ptr = value.mutable_data(place); - - for (int64_t i = 0; i < rows_numel; ++i) { - EXPECT_EQ(ptr[0 + i * value.dims()[1]], static_cast(i * 2)); - } - } - - server_thread.join(); - LOG(INFO) << "begin reset"; - g_rpc_service.reset(nullptr); - g_req_handler.reset(nullptr); -} diff --git a/paddle/fluid/operators/detail/grpc_service.h b/paddle/fluid/operators/detail/grpc_service.h deleted file mode 100644 index e0505c2b9d0903837713d7e0032b01ab091c2e04..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/grpc_service.h +++ /dev/null @@ -1,123 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include "paddle/fluid/operators/detail/variable_response.h" - -#include "paddle/fluid/platform/profiler.h" - -// NOTE: This method was originally created by tensorflow -// (https://github.com/tensorflow/tensorflow/) we borrow this -// method and did some modifications so that we can parse gRPC -// requests without too much copying of the tensor data. - -namespace grpc { -class CompletionQueue; -class Channel; -class RpcService; -class ServerCompletionQueue; -class ServerContext; - -// Support parsing/unparsing of tensorflow::VariableResponse. -// Wire-format is identical to RecvVariableResponse. -template <> -class SerializationTraits { - public: - static Status Serialize( - const paddle::operators::detail::VariableResponse& msg, - grpc_byte_buffer** bp, bool* own_buffer) { - PADDLE_ENFORCE(false, "SerializationTraits::Serialize not implemented!"); - return Status(); - } - static Status Deserialize(grpc_byte_buffer* buffer, - paddle::operators::detail::VariableResponse* msg, - int max_message_size = INT_MAX) { - if (buffer == nullptr) { - return Status(StatusCode::INTERNAL, "No payload"); - } - - Status result = g_core_codegen_interface->ok(); - if (result.ok()) { - paddle::operators::detail::GrpcByteSource source(buffer); - int ret = msg->Parse(&source); - if (ret != 0) { - result = Status(StatusCode::INTERNAL, "VariableResponse parse error"); - } - } - g_core_codegen_interface->grpc_byte_buffer_destroy(buffer); - return result; - } -}; -} // namespace grpc - -namespace paddle { -namespace operators { -namespace detail { - -enum class GrpcMethod { - kSendVariable, - kGetVariable, - kPrefetchVariable, -}; - -static const int kGrpcNumMethods = - static_cast(GrpcMethod::kPrefetchVariable) + 1; - -inline const char* GrpcMethodName(GrpcMethod id) { - switch (id) { - case GrpcMethod::kSendVariable: - return "/sendrecv.SendRecvService/SendVariable"; - case GrpcMethod::kGetVariable: - return "/sendrecv.SendRecvService/GetVariable"; - case GrpcMethod::kPrefetchVariable: - return "/sendrecv.SendRecvService/PrefetchVariable"; - } - - // Shouldn't be reached. - PADDLE_ENFORCE(false, "Invalid id: not found valid method name"); - return nullptr; -} - -class GrpcService final { - public: - class AsyncService : public ::grpc::Service { - public: - AsyncService() { - for (int i = 0; i < kGrpcNumMethods; ++i) { - AddMethod(new ::grpc::internal::RpcServiceMethod( - GrpcMethodName(static_cast(i)), - ::grpc::internal::RpcMethod::NORMAL_RPC, nullptr)); - ::grpc::Service::MarkMethodAsync(i); - } - } - virtual ~AsyncService() {} - - // Make RequestAsyncUnary public for grpc_call.h - using ::grpc::Service::RequestAsyncUnary; - }; -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/macros.h b/paddle/fluid/operators/detail/macros.h new file mode 100644 index 0000000000000000000000000000000000000000..6f4a15caa5542a45cd8e26a72b055ca8948069d0 --- /dev/null +++ b/paddle/fluid/operators/detail/macros.h @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#ifdef PADDLE_WITH_DISTRIBUTE + +#ifdef PADDLE_WITH_GRPC + +#include "paddle/fluid/operators/distributed/grpc_client.h" +#include "paddle/fluid/operators/distributed/grpc_server.h" +#define RPCSERVER_T paddle::operators::distributed::AsyncGRPCServer +#define RPCCLIENT_T paddle::operators::distributed::GRPCClient + +#else // PADDLE_WITH_GRPC + +#include "paddle/fluid/operators/distributed/brpc_client.h" +#include "paddle/fluid/operators/distributed/brpc_server.h" +#define RPCSERVER_T paddle::operators::distributed::AsyncBRPCServer +#define RPCCLIENT_T paddle::operators::distributed::BRPCClient + +#endif // PADDLE_WITH_GRPC + +#endif // PADDLE_WITH_DISTRIBUTE diff --git a/paddle/fluid/operators/detail/proto_encoder_helper.h b/paddle/fluid/operators/detail/proto_encoder_helper.h deleted file mode 100644 index d91d054b2507f32d1e948dde33da06a70cabe775..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/proto_encoder_helper.h +++ /dev/null @@ -1,149 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -// NOTE: This file was originally created by tensorflow -// (https://github.com/tensorflow/tensorflow/) we borrow this -// file and did some modifications so that we can send gRPC -// requests without too much copying of the tensor data. - -#pragma once - -#include - -#include "grpc++/grpc++.h" -#include "paddle/fluid/platform/enforce.h" - -namespace paddle { -namespace operators { -namespace detail { - -char* EncodeVarint32(char* dst, uint32_t v) { - // Operate on characters as unsigneds - unsigned char* ptr = reinterpret_cast(dst); - static const int B = 128; - if (v < (1 << 7)) { - *(ptr++) = v; - } else if (v < (1 << 14)) { - *(ptr++) = v | B; - *(ptr++) = v >> 7; - } else if (v < (1 << 21)) { - *(ptr++) = v | B; - *(ptr++) = (v >> 7) | B; - *(ptr++) = v >> 14; - } else if (v < (1 << 28)) { - *(ptr++) = v | B; - *(ptr++) = (v >> 7) | B; - *(ptr++) = (v >> 14) | B; - *(ptr++) = v >> 21; - } else { - *(ptr++) = v | B; - *(ptr++) = (v >> 7) | B; - *(ptr++) = (v >> 14) | B; - *(ptr++) = (v >> 21) | B; - *(ptr++) = v >> 28; - } - return reinterpret_cast(ptr); -} - -char* EncodeVarint64(char* dst, uint64_t v) { - static const int B = 128; - unsigned char* ptr = reinterpret_cast(dst); - while (v >= B) { - *(ptr++) = (v & (B - 1)) | B; - v >>= 7; - } - *(ptr++) = static_cast(v); - return reinterpret_cast(ptr); -} - -int VarintLength(uint64_t v) { - int len = 1; - while (v >= 128) { - v >>= 7; - len++; - } - return len; -} - -class ProtoEncodeHelper { - public: - ProtoEncodeHelper(char* buf, int max_size) - : base_(buf), p_(buf), limit_(base_ + max_size) {} - - ~ProtoEncodeHelper() { - // Make sure callers didn't do operations that went over max_size promised - PADDLE_ENFORCE_LE(p_, limit_); - } - - const char* data() const { return base_; } - size_t size() const { return p_ - base_; } - - void WriteUint64(int tag, uint64_t v) { - Encode32(combine(tag, WIRETYPE_VARINT)); - Encode64(v); - } - void WriteBool(int tag, bool v) { - Encode32(combine(tag, WIRETYPE_VARINT)); - EncodeBool(v); - } - void WriteString(int tag, const std::string& v) { - Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED)); - Encode32(v.size()); - EncodeBytes(v.data(), v.size()); - } - void WriteVarlengthBeginning(int tag, uint32_t len) { - Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED)); - Encode32(len); - } - void WriteRawBytes(const std::string& v) { EncodeBytes(v.data(), v.size()); } - - private: - // Note: this module's behavior must match the protocol buffer wire encoding - // format. - enum { - WIRETYPE_VARINT = 0, - WIRETYPE_LENGTH_DELIMITED = 2, - }; - static uint32_t combine(uint32_t tag, uint32_t type) { - return ((tag << 3) | type); - } - inline void Encode32(uint32_t v) { - if (v < 128) { - // Fast path for single-byte values. Many of the calls will use a - // constant value for v, so the comparison will get optimized away - // when Encode32 is inlined into the caller. - *p_ = v; - p_++; - } else { - p_ = EncodeVarint32(p_, v); - } - } - void Encode64(uint64_t v) { p_ = EncodeVarint64(p_, v); } - void EncodeBool(bool v) { - *p_ = (v ? 1 : 0); // Equal to varint32 encoding of 0 or 1 - p_++; - } - void EncodeBytes(const char* bytes, int N) { - memcpy(p_, bytes, N); - p_ += N; - } - - char* base_; - char* p_; - char* limit_; // Just for CHECKs -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler.h b/paddle/fluid/operators/detail/request_handler.h deleted file mode 100644 index 4bc5e7f10ee2a8939d230fe96517bd9f56c13933..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/request_handler.h +++ /dev/null @@ -1,127 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include - -#include -#include -#include -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" - -namespace paddle { -namespace operators { -namespace detail { - -constexpr char kRequestSend[] = "RequestSend"; -constexpr char kRequestGet[] = "RequestGet"; -constexpr char kRequestPrefetch[] = "RequestPrefetch"; - -class RPCServer; - -class RequestHandler { - public: - explicit RequestHandler(bool sync_mode) - : sync_mode_(sync_mode), - dev_ctx_(nullptr), - executor_(nullptr), - scope_(nullptr), - program_(nullptr), - rpc_server_(nullptr) {} - - virtual ~RequestHandler() {} - - // Set attributes. - void SetScope(framework::Scope* scope) { scope_ = scope; } - void SetDevCtx(const platform::DeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; } - void SetProgram(framework::ProgramDesc* program) { program_ = program; } - void SetExecutor(framework::Executor* executor) { executor_ = executor; } - void SetPrefetchPreparedCtx( - std::unique_ptr prepared) { - prefetch_ctx_.reset(prepared.release()); - } - - // Used for async. - void SetGradToPreparedCtx( - std::unordered_map< - std::string, std::shared_ptr>* g) { - grad_to_prepared_ctx_ = g; - } - - void SetRPCServer(RPCServer* rpc_server) { rpc_server_ = rpc_server; } - - // Get attributes. - bool sync_mode() { return sync_mode_; } - framework::Scope* scope() { return scope_; } - const platform::DeviceContext* dev_ctx() { return dev_ctx_; } - framework::ExecutorPrepareContext* prefetch_ctx() { - return prefetch_ctx_.get(); - } - framework::ProgramDesc* program() { return program_; } - framework::Executor* executor() { return executor_; } - std::vector& sparse_vars() { return sparse_vars_; } - - // This function processes user's rpc request. - // The implemention is in request_handler_impl. - // example: - // std::string varname = request_.varname(); - // - // auto scope = request_handler_->scope(); - // auto invar = scope->FindVar(varname); - // framework::Variable* outvar = nullptr; - // - // request_handler_->Handle(varname, scope, invar, &outvar); - // if (outvar) { - // SerializeToByteBuffer(varname, outvar, - // *request_handler_->dev_ctx(), &reply_); - // } - virtual bool Handle(const std::string& varname, framework::Scope* scope, - framework::Variable* var, - framework::Variable** outvar) = 0; - - protected: - const bool sync_mode_; - - const platform::DeviceContext* dev_ctx_; - framework::Executor* executor_; - framework::Scope* scope_; - framework::ProgramDesc* program_; - std::unique_ptr prefetch_ctx_; - - // Used for async. - std::unordered_map>* - grad_to_prepared_ctx_; - - // Record received sparse variables, so that - // we could reset those after execute optimize program - std::vector sparse_vars_; - RPCServer* rpc_server_; - - std::mutex sparse_var_mutex_; -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.cc b/paddle/fluid/operators/detail/request_handler_impl.cc deleted file mode 100644 index f16c06d52f4fb86d51083a8b3b98d05a64c1af74..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/request_handler_impl.cc +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include -#include -#include - -#include "paddle/fluid/framework/blocking_queue.h" -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" -#include "paddle/fluid/operators/detail/rpc_server.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/operators/detail/variable_response.h" - -namespace paddle { -namespace operators { -namespace detail { - -bool RequestSendHandler::Handle(const std::string& varname, - framework::Scope* scope, - framework::Variable* invar, - framework::Variable** outvar) { - VLOG(4) << "RequestSendHandler:" << varname; - - // Async - if (!sync_mode_) { - try { - executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), - scope); - } catch (std::exception& e) { - LOG(ERROR) << "async: run sub program error " << e.what(); - return false; - } - return true; - } - - // Sync - if (varname == BATCH_BARRIER_MESSAGE) { - VLOG(3) << "sync: recv batch barrier message"; - rpc_server_->IncreaseBatchBarrier(kRequestSend); - } else { - VLOG(3) << "sync: received var_name: " << varname; - if (sync_mode_) { - rpc_server_->WaitCond(kRequestSend); - } - - if (invar == nullptr) { - LOG(ERROR) << "sync: Can not find server side var: " << varname; - PADDLE_THROW("sync: Can not find server side var"); - return false; - } - - if (invar->IsType()) { - std::unique_lock lock(sparse_var_mutex_); - sparse_vars_.push_back(invar); - } - } - - return true; -} - -bool RequestGetHandler::Handle(const std::string& varname, - framework::Scope* scope, - framework::Variable* invar, - framework::Variable** outvar) { - VLOG(4) << "RequestGetHandler:" << varname; - - if (varname != FETCH_BARRIER_MESSAGE) { - if (sync_mode_) { - rpc_server_->WaitCond(kRequestGet); - } - *outvar = scope_->FindVar(varname); - return true; - } - - // FETCH_BARRIER_MESSAGE - if (sync_mode_) { - VLOG(3) << "sync: recv fetch barrier message"; - rpc_server_->IncreaseBatchBarrier(kRequestGet); - } - - return true; -} - -bool RequestPrefetchHandler::Handle(const std::string& varname, - framework::Scope* scope, - framework::Variable* invar, - framework::Variable** outvar) { - VLOG(4) << "RequestPrefetchHandler " << varname; - - auto var_desc = program_->Block(0).FindVar(varname); - *outvar = scope->FindVar(varname); - InitializeVariable(*outvar, var_desc->GetType()); - executor_->RunPreparedContext(prefetch_ctx_.get(), scope); - - return true; -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.h b/paddle/fluid/operators/detail/request_handler_impl.h deleted file mode 100644 index 8d0c62232b68ad6c05e751c25103802ee12db57e..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/request_handler_impl.h +++ /dev/null @@ -1,64 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include - -#include -#include -#include -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/request_handler.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" - -namespace paddle { -namespace operators { -namespace detail { - -class RequestSendHandler final : public RequestHandler { - public: - explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {} - virtual ~RequestSendHandler() {} - bool Handle(const std::string& varname, framework::Scope* scope, - framework::Variable* var, framework::Variable** outvar) override; -}; - -class RequestGetHandler final : public RequestHandler { - public: - explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {} - virtual ~RequestGetHandler() {} - bool Handle(const std::string& varname, framework::Scope* scope, - framework::Variable* var, framework::Variable** outvar) override; -}; - -class RequestPrefetchHandler final : public RequestHandler { - public: - explicit RequestPrefetchHandler(bool sync_mode) : RequestHandler(sync_mode) {} - virtual ~RequestPrefetchHandler() {} - bool Handle(const std::string& varname, framework::Scope* scope, - framework::Variable* var, framework::Variable** outvar) override; -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.cc b/paddle/fluid/operators/detail/rpc_server.cc deleted file mode 100644 index 448763372a8c224cc68319a4a444915896b68234..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/rpc_server.cc +++ /dev/null @@ -1,113 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include -#include -#include -#include - -#include "paddle/fluid/operators/detail/rpc_server.h" - -namespace paddle { -namespace operators { -namespace detail { - -void RPCServer::ShutDown() { - LOG(INFO) << "RPCServer ShutDown "; - ShutDownImpl(); - - exit_flag_ = true; - barrier_cond_.notify_all(); - rpc_cond_.notify_all(); -} - -void RPCServer::SavePort() const { - auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); - std::ofstream port_file; - port_file.open(file_path); - port_file << selected_port_; - port_file.close(); - VLOG(4) << "selected port written to " << file_path; -} - -void RPCServer::WaitBarrier(const std::string& rpc_name) { - std::unique_lock lock(this->mutex_); - barrier_cond_.wait(lock, [=] { - return (barrier_counter_[rpc_name] >= client_num_ || exit_flag_.load()); - }); - - VLOG(3) << "batch_barrier_:" << barrier_counter_[rpc_name]; -} - -void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) { - VLOG(3) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; - int b = 0; - { - std::unique_lock lock(mutex_); - b = ++barrier_counter_[rpc_name]; - } - - VLOG(3) << "RPCServer IncreaseBatchBarrier " << rpc_name - << ", barrier_count:" << b << ", fan_in" << client_num_; - - if (b >= client_num_) { - barrier_cond_.notify_all(); - } -} - -void RPCServer::ResetBarrierCounter() { - VLOG(3) << "RPCServer ResetBarrierCounter "; - std::unique_lock lock(mutex_); - for (auto& t : barrier_counter_) { - t.second = 0; - } -} - -void RPCServer::RegisterRPC(const std::string& rpc_name, - RequestHandler* handler, int thread_num) { - rpc_call_map_[rpc_name] = handler; - rpc_thread_num_[rpc_name] = thread_num; - - static int cond = -1; - rpc_cond_map_[rpc_name] = ++cond; - VLOG(4) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler - << ", cond:" << rpc_cond_map_[rpc_name]; -} - -void RPCServer::SetCond(const std::string& rpc_name) { - VLOG(3) << "RPCServer SetCond " << rpc_name; - { - std::unique_lock lock(mutex_); - cur_cond_ = rpc_cond_map_[rpc_name]; - } - - rpc_cond_.notify_all(); -} - -void RPCServer::WaitCond(const std::string& rpc_name) { - VLOG(3) << "RPCServer WaitCond " << rpc_name; - int cond = 0; - { - std::unique_lock lock(mutex_); - cond = rpc_cond_map_[rpc_name]; - } - - std::unique_lock lock(mutex_); - rpc_cond_.wait( - lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); }); -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.h b/paddle/fluid/operators/detail/rpc_server.h deleted file mode 100644 index c2e7ae706c9dc6776e09b25e424b30f110c3855d..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/rpc_server.h +++ /dev/null @@ -1,91 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include -#include // NOLINT -#include -#include -#include "paddle/fluid/operators/detail/request_handler.h" - -namespace paddle { -namespace operators { -namespace detail { - -class RPCServer { - public: - explicit RPCServer(const std::string& address, int client_num) - : cur_cond_(0), - bind_address_(address), - exit_flag_(false), - selected_port_(0), - client_num_(client_num) {} - - virtual ~RPCServer() {} - virtual void StartServer() = 0; - virtual void WaitServerReady() = 0; - - void ShutDown(); - - bool IsExit() { return exit_flag_.load(); } - - int GetSelectedPort() const { return selected_port_; } - void SavePort() const; - - // RegisterRPC, register the rpc method name to a handler - // class, and auto generate a condition id for this call - // to be used for the barrier. - void RegisterRPC(const std::string& rpc_name, RequestHandler* handler, - int thread_num = 5); - - // Wait util all the clients have reached the barrier for one - // rpc method. This function should be called in the - // RequestHandler if you want to run the server/client in a - // synchronous mode. - void WaitBarrier(const std::string& rpc_name); - - void SetCond(const std::string& rpc_name); - void WaitCond(const std::string& rpc_name); - void IncreaseBatchBarrier(const std::string rpc_name); - void ResetBarrierCounter(); - - protected: - virtual void ShutDownImpl() = 0; - - private: - std::mutex mutex_; - std::unordered_map barrier_counter_; - std::condition_variable barrier_cond_; - - std::unordered_map rpc_cond_map_; - std::atomic cur_cond_; - std::condition_variable rpc_cond_; - - protected: - std::string bind_address_; - std::atomic exit_flag_; - int selected_port_; - - const int client_num_; - - std::unordered_map rpc_call_map_; - std::unordered_map rpc_thread_num_; - friend class RequestHandler; -}; - -}; // namespace detail -}; // namespace operators -}; // namespace paddle diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto deleted file mode 100644 index a244afc46f3247c7e6e8481b09b5c729a2a569f7..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/send_recv.proto +++ /dev/null @@ -1,79 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. Licensed under -the Apache License, Version 2.0 (the "License"); you may not use this file -except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -syntax = "proto3"; -package sendrecv; - -service SendRecvService { - // For parameter server round-robin like hashing, do not split tensors. - // Send and recv only one tensor - // TODO(typhoonzero): add streaming API - rpc SendVariable(VariableMessage) returns (VoidMessage) {} - // Argument VariableMessage for GetVariable should only contain varname. - rpc GetVariable(VariableMessage) returns (VariableMessage) {} - // pre-fetch variable by given variable name and Ids - rpc PrefetchVariable(VariableMessage) returns (VariableMessage) {} -} - -// VariableMessage is serialized paddle variable message. -// It can be: -// LoDTensor -// SelectedRows -enum VarType { - LOD_TENSOR = 0; - SELECTED_ROWS = 1; - NCCL_ID = 2; -} - -// NOTICE(gongwb):don't modify this proto if you are not -// not familar with how we serialize in sendrecvop_utils.h -// and deserilize it in variable_response.h. -message VariableMessage { - enum Type { - // Pod Types - BOOL = 0; - INT16 = 1; - INT32 = 2; - INT64 = 3; - FP16 = 4; - FP32 = 5; - FP64 = 6; - } - - message LodData { repeated int64 lod_data = 1; } - string varname = 1; - // TODO(Yancey1989): reference framework::proto::VarDesc::VarType - VarType type = 2; - // bool persistable is not needed for sending. - // tensor info: - Type data_type = 3; - repeated int64 dims = 4; - - // lod details: - int64 lod_level = 5; - repeated LodData lod = 6; - // selected_rows height, aka. original dim0 - int64 slr_height = 7; - // tensor data - bytes serialized = 8; - // selected_rows data - bytes rows = 9; - // Look up table block execution output variable name. - string out_varname = 10; - // If 1, the ps server will start profiling, the ps - // server stops profiling and generates a profile to /tmp/profile_ps_* - // when profile switches from 1 to 2. - int64 profile = 11; -} - -message VoidMessage {} diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc deleted file mode 100644 index 507b465435609a91ebca97dd70b176c3b79bee02..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/sendrecvop_utils.cc +++ /dev/null @@ -1,232 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" - -#ifdef PADDLE_WITH_CUDA -#include -#endif -#include -#include // NOLINT - -#include "google/protobuf/io/coded_stream.h" -#include "google/protobuf/io/zero_copy_stream.h" -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/operators/detail/bytebuffer_stream.h" -#include "paddle/fluid/operators/detail/proto_encoder_helper.h" -#include "paddle/fluid/operators/detail/variable_response.h" -#include "paddle/fluid/platform/profiler.h" - -namespace paddle { -namespace operators { -namespace detail { - -using VarMsg = sendrecv::VariableMessage; - -void GetTensorPayload(framework::Variable* var, - const platform::DeviceContext& ctx, VarMsg* request, - void** payload, size_t* payload_size) { - auto tensor = var->Get(); - // FIXME(wuyi): data types in send_recv.proto is copied from - // framework.proto - request->set_data_type( - static_cast(framework::ToDataType(tensor.type()))); - for (auto& dim : framework::vectorize(tensor.dims())) { - request->add_dims(dim); - } - const framework::LoD lod = tensor.lod(); - if (lod.size() > 0) { - request->set_lod_level(lod.size()); - for (auto& each : lod) { - VarMsg::LodData* lod_inner = request->add_lod(); - for (auto& d : each) { - lod_inner->add_lod_data(d); - } - } - } - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef PADDLE_WITH_CUDA - PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); - platform::CUDAPinnedPlace cuda_pinned; - auto& gpu_dev_ctx = static_cast(ctx); - auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type()); - *payload = memory::Alloc(cuda_pinned, copy_size); - - memory::Copy(cuda_pinned, *payload, - boost::get(tensor.place()), - reinterpret_cast(tensor.data()), copy_size, - gpu_dev_ctx.stream()); - ctx.Wait(); -#endif - } else { - *payload = tensor.data(); - } - *payload_size = tensor.numel() * framework::SizeOfType(tensor.type()); -} - -void GetSelectedRowsPayload(framework::Variable* var, - const platform::DeviceContext& ctx, VarMsg* request, - void** payload, size_t* payload_size) { - auto* slr = var->GetMutable(); - request->set_data_type( - static_cast(framework::ToDataType(slr->value().type()))); - request->set_lod_level(0); - request->set_slr_height(slr->height()); - - for (auto& dim : framework::vectorize(slr->value().dims())) { - request->add_dims(dim); - } - - auto* tensor = slr->mutable_value(); - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef PADDLE_WITH_CUDA - platform::CUDAPinnedPlace cuda_pinned; - auto& gpu_dev_ctx = static_cast(ctx); - auto copy_size = tensor->numel() * framework::SizeOfType(tensor->type()); - *payload = memory::Alloc(cuda_pinned, copy_size); - memory::Copy(cuda_pinned, *payload, - boost::get(tensor->place()), - reinterpret_cast(tensor->data()), copy_size, - gpu_dev_ctx.stream()); - ctx.Wait(); -#endif - } else { - *payload = slr->mutable_value()->data(); - } - *payload_size = tensor->numel() * framework::SizeOfType(tensor->type()); -} - -void SerializeToByteBuffer(const std::string& name, framework::Variable* var, - const platform::DeviceContext& ctx, - ::grpc::ByteBuffer* msg, - const std::string& out_name) { - // Default DestroyCallback does nothing, When using GPU - // the CPU buffer need to be freed. - DestroyCallback destroy_callback = [](void* backing) {}; - VarMsg request; - void* payload = nullptr; - size_t payload_size; - - request.set_varname(name); - // Note: normally the profiler is enabled in 1 trainer, hence only - // 1 trainer returns true for ShouldSendProfileState(). It tells PS - // servers the trainer's profiling state so that PS can follow the - // trainer. - if (platform::ShouldSendProfileState()) { - if (platform::IsProfileEnabled()) { - request.set_profile(platform::kEnableProfiler); - } else { - request.set_profile(platform::kDisableProfiler); - } - } - if (!out_name.empty()) { - request.set_out_varname(out_name); - } - if (var->IsType()) { - request.set_type(::sendrecv::LOD_TENSOR); - GetTensorPayload(var, ctx, &request, &payload, &payload_size); - } else if (var->IsType()) { - request.set_type(::sendrecv::SELECTED_ROWS); - GetSelectedRowsPayload(var, ctx, &request, &payload, &payload_size); -#ifdef PADDLE_WITH_CUDA - } else if (var->IsType()) { - request.set_type(::sendrecv::NCCL_ID); -#endif - } else { - PADDLE_THROW("Serialize does not support type: %s", - typeid(var->Type()).name()); - } - - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef PADDLE_WITH_CUDA - // GPU data is copied to CPU buffer when sending, - // free the buffer when possible. - destroy_callback = [](void* backing) { - platform::CUDAPinnedPlace cuda_pinned; - memory::Free(cuda_pinned, backing); - }; -#endif - } - - std::string header; - request.AppendToString(&header); - auto buffer = std::unique_ptr(new char[1024]); - void* buf = buffer.get(); - ProtoEncodeHelper e(static_cast(buf), 1024); - e.WriteRawBytes(std::string(header.data(), header.size())); -// NCCLID is copied directly to the message, return bytebuffer -// with only one slice if serializing NCCLID. -#ifdef PADDLE_WITH_CUDA - if (var->IsType()) { - e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, - NCCL_UNIQUE_ID_BYTES); - const ncclUniqueId& uid = var->Get(); - e.WriteRawBytes(std::string(uid.internal, NCCL_UNIQUE_ID_BYTES)); - - // for serialize NCCL_ID - ::grpc::Slice slices(e.size()); - memcpy(const_cast(slices.begin()), e.data(), e.size()); - ::grpc::ByteBuffer tmp(&slices, 1); - msg->Swap(&tmp); - return; - } -#endif - - e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); - // steal reference of tensor data - ::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows - int num_slices = 2; // only SelectedRows have rows buffer - slices[0] = ::grpc::Slice(e.size()); - memcpy(const_cast(slices[0].begin()), e.data(), e.size()); - slices[1] = ::grpc::Slice( - grpc_slice_new_with_user_data(payload, payload_size, destroy_callback, - static_cast(payload)), - ::grpc::Slice::STEAL_REF); - - if (var->IsType()) { - auto* slr = var->GetMutable(); - ProtoEncodeHelper e2(static_cast(buf), 128); - size_t rows_memory_size = - slr->rows().size() * framework::SizeOfType(typeid(int64_t)); - e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); - slices[2] = ::grpc::Slice(e2.size()); - memcpy(const_cast(slices[2].begin()), e2.data(), e2.size()); - - slices[3] = ::grpc::Slice( - grpc_slice_new_with_user_data( - const_cast( - reinterpret_cast(slr->rows().data())), - rows_memory_size, [](void* backing) {}, - const_cast( - reinterpret_cast(slr->rows().data()))), - ::grpc::Slice::STEAL_REF); - num_slices = 4; - } - - ::grpc::ByteBuffer tmp(&slices[0], num_slices); - msg->Swap(&tmp); -} - -void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, - const platform::DeviceContext& ctx, - const framework::Scope* scope, - framework::Variable** var) { - operators::detail::VariableResponse resp(scope, &ctx); - PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); - *var = resp.GetVar(); -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.h b/paddle/fluid/operators/detail/sendrecvop_utils.h deleted file mode 100644 index c72e1bd076f670458f3915072154847db6205092..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/sendrecvop_utils.h +++ /dev/null @@ -1,76 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/framework/tensor_util.h" -#include "paddle/fluid/framework/var_type.h" - -#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" - -namespace paddle { -namespace operators { -namespace detail { - -#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" -#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV" -#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV" - -static int64_t GetTimestamp() { - struct timeval tp; - gettimeofday(&tp, NULL); - return tp.tv_sec * 1000 + tp.tv_usec / 1000; -} - -typedef void (*DestroyCallback)(void*); - -void SerializeToByteBuffer(const std::string& name, framework::Variable* var, - const platform::DeviceContext& ctx, - ::grpc::ByteBuffer* msg, - const std::string& out_varname = std::string()); - -void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, - const platform::DeviceContext& ctx, - const framework::Scope* scope, - framework::Variable** var); - -inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) { - switch (type) { - case sendrecv::VariableMessage::FP32: - return typeid(float); // NOLINT - case sendrecv::VariableMessage::FP64: - return typeid(double); // NOLINT - case sendrecv::VariableMessage::INT32: - return typeid(int); // NOLINT - case sendrecv::VariableMessage::INT64: - return typeid(int64_t); // NOLINT - case sendrecv::VariableMessage::BOOL: - return typeid(bool); // NOLINT - default: - PADDLE_THROW("Not support type %d", type); - } -} - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/serde_test.cc b/paddle/fluid/operators/detail/serde_test.cc deleted file mode 100644 index 15892295e6901fe649788c9e34604008fc8cbdfa..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/serde_test.cc +++ /dev/null @@ -1,221 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include // NOLINT - -#include "google/protobuf/text_format.h" -#include "gtest/gtest.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/tensor_util.h" -#include "paddle/fluid/framework/variable.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/operators/detail/variable_response.h" -#include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/platform/place.h" -#include "paddle/fluid/string/printf.h" - -namespace framework = paddle::framework; -namespace platform = paddle::platform; -namespace operators = paddle::operators; -namespace math = paddle::operators::math; -namespace memory = paddle::memory; - -void RunSerdeTestSelectedRows(platform::Place place) { - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - - // serialize var to ByteBuffer - framework::Variable var; - auto* slr = var.GetMutable(); - slr->set_height(1000); - auto* tensor = slr->mutable_value(); - auto* rows = slr->mutable_rows(); - tensor->Resize(framework::make_ddim({564, 128})); - tensor->mutable_data(place); - int tensor_numel = 564 * 128; - math::set_constant(ctx, tensor, 32.7); - for (int i = 0; i < 564; ++i) rows->push_back(i); - - ::grpc::ByteBuffer msg; - operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); - EXPECT_GT(msg.Length(), static_cast(0)); - - // deserialize - std::vector<::grpc::Slice> slices; - (void)msg.Dump(&slices); - std::string tmp; - for (const auto& s : slices) { - tmp.append(reinterpret_cast(s.begin()), s.size()); - } - - sendrecv::VariableMessage varmsg; - EXPECT_TRUE(varmsg.ParseFromString(tmp)); - - // deserialize bytebuffer - EXPECT_EQ(varmsg.varname(), "myvar"); - EXPECT_EQ(varmsg.type(), 1); - - const float* tensor_data = - reinterpret_cast(varmsg.serialized().data()); - const int64_t* rows_data = - reinterpret_cast(varmsg.rows().data()); - for (int i = 0; i < tensor_numel; ++i) { - EXPECT_FLOAT_EQ(tensor_data[i], 32.7); - } - for (int i = 0; i < 564; ++i) { - EXPECT_EQ(rows_data[i], i); - } - - // deserialize zero-copy - // framework::Variable var2; - // operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); - framework::Scope scope; - scope.Var("myvar"); - operators::detail::VariableResponse resp(&scope, &ctx); - EXPECT_EQ(resp.Parse(msg), 0); - - framework::Variable* var2 = resp.GetVar(); - - auto* slr2 = var2->GetMutable(); - auto* tensor2 = slr2->mutable_value(); - auto* rows2 = slr2->mutable_rows(); - float* tensor_data2 = nullptr; - framework::Tensor tmp_tensor; - - if (platform::is_gpu_place(ctx.GetPlace())) { - platform::CPUPlace cpu; - framework::TensorCopy(*tensor2, cpu, &tmp_tensor); - tensor_data2 = tmp_tensor.data(); - } else { - tensor_data2 = const_cast(tensor2->data()); - } - const int64_t* rows_data2 = rows2->data(); - - for (int i = 0; i < tensor_numel; ++i) { - EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); - } - for (size_t i = 0; i < rows2->size(); ++i) { - EXPECT_EQ(rows_data2[i], static_cast(i)); - } - EXPECT_EQ(slr2->height(), 1000); -} - -void RunTestLodTensor(platform::Place place, int from_type = 0) { - // serialize var to ByteBuffer - framework::Variable var; - auto* tensor = var.GetMutable(); - tensor->Resize(framework::make_ddim({512, 8, 4, 2})); - framework::LoD lod; - lod.push_back(framework::Vector({1, 3, 8})); - tensor->set_lod(lod); - int tensor_numel = 512 * 8 * 4 * 2; - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - tensor->mutable_data(place); - math::set_constant(ctx, tensor, 31.9); - - ::grpc::ByteBuffer msg; - operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); - EXPECT_GT(msg.Length(), static_cast(0)); - - // deserialize - std::vector<::grpc::Slice> slices; - (void)msg.Dump(&slices); - std::string tmp; - for (const auto& s : slices) { - tmp.append(reinterpret_cast(s.begin()), s.size()); - } - sendrecv::VariableMessage varmsg; - EXPECT_TRUE(varmsg.ParseFromString(tmp)); - EXPECT_EQ(varmsg.varname(), "myvar"); - EXPECT_EQ(varmsg.type(), 0); - EXPECT_EQ(varmsg.dims()[0], 512); - EXPECT_EQ(varmsg.dims()[1], 8); - EXPECT_EQ(varmsg.dims()[2], 4); - EXPECT_EQ(varmsg.dims()[3], 2); - EXPECT_EQ(varmsg.lod_level(), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); - EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); - - const float* tensor_data = - reinterpret_cast(varmsg.serialized().data()); - for (int i = 0; i < tensor_numel; ++i) { - EXPECT_FLOAT_EQ(tensor_data[i], 31.9); - } - - // message binary - std::string str; - varmsg.SerializeToString(&str); - - // message bytebuffer - ::grpc::Slice slices_2[1]; - int num_slices = 1; - slices_2[0] = ::grpc::Slice(str.length()); - memcpy(const_cast(slices_2[0].begin()), str.c_str(), str.length()); - ::grpc::ByteBuffer bytebuffer2(&slices_2[0], num_slices); - - // deserialize zero-copy - framework::Scope scope; - scope.Var("myvar"); - operators::detail::VariableResponse resp(&scope, &ctx); - if (from_type == 0) { - EXPECT_EQ(resp.Parse(msg), 0); - } else { - EXPECT_EQ(resp.Parse(bytebuffer2), 0); - } - - framework::Variable* var2 = resp.GetVar(); - - auto tensor2 = var2->Get(); - float* tensor_data2 = nullptr; - framework::Tensor tmp_tensor; - - if (platform::is_gpu_place(ctx.GetPlace())) { - platform::CPUPlace cpu; - framework::TensorCopy(tensor2, cpu, &tmp_tensor); - tensor_data2 = tmp_tensor.data(); - } else { - tensor_data2 = const_cast(tensor2.data()); - } - - EXPECT_EQ(varmsg.lod_level(), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); - EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); - for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9); -} - -TEST(LodTensor, Run) { - platform::CPUPlace place; - RunTestLodTensor(place); - RunTestLodTensor(place, 1); -#ifdef PADDLE_WITH_CUDA - platform::CUDAPlace gpu(0); - RunTestLodTensor(gpu); - RunTestLodTensor(gpu, 1); -#endif -} - -TEST(SelectedRows, Run) { - platform::CPUPlace place; - RunSerdeTestSelectedRows(place); - -#ifdef PADDLE_WITH_CUDA - platform::CUDAPlace gpu; - RunSerdeTestSelectedRows(gpu); -#endif -} diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/detail/variable_response.cc deleted file mode 100644 index 24cb91a3bb820a0e5d51aaa49154434919080f69..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/variable_response.cc +++ /dev/null @@ -1,485 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/operators/detail/variable_response.h" - -#include -#include -#include -#ifdef PADDLE_WITH_CUDA -#include -#endif -#include "paddle/fluid/platform/profiler.h" - -#include "paddle/fluid/operators/detail/send_recv.pb.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" - -namespace paddle { -namespace operators { -namespace detail { - -enum WireType { - WIRETYPE_VARINT = 0, - WIRETYPE_LENGTH_DELIMITED = 2, -}; - -inline int GetTagFieldNumber(uint32_t tag) { return tag >> 3; } - -inline WireType GetTagWireType(uint32_t tag) { - return static_cast(tag & 0x7); -} - -bool ReadVarintSizeAsInt(::google::protobuf::io::CodedInputStream* input, - int* result) { - uint64_t v; - if (input->ReadVarint64(&v) && v <= static_cast(INT_MAX)) { - *result = static_cast(v); - return true; - } else { - return false; - } -} - -bool ReadRaw(::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& dev_ctx, platform::Place place, - void* dest, int size) { - const void* data = NULL; - int size_to_write = 0; - int length = size; - int total_written = 0; - - if (platform::is_gpu_place(place)) { -#ifdef PADDLE_WITH_CUDA - auto& gpu_dev_ctx = - static_cast(dev_ctx); - platform::CPUPlace cpu; - - char* p = reinterpret_cast(dest); - while (total_written < length) { - if (!input->GetDirectBufferPointer(&data, &size_to_write)) { - return false; - } - // NOTE: if raw buffer is large and have two neighbor fields of raw - // buffers GetDirectBufferPointer can get all of them, use length to - // truncate it. - if (total_written + size_to_write > length) { - size_to_write = length - total_written; - } - memory::Copy(boost::get(place), - reinterpret_cast(p), cpu, data, size_to_write, - gpu_dev_ctx.stream()); - p += size_to_write; - total_written += size_to_write; - - input->Skip(size_to_write); - } - gpu_dev_ctx.Wait(); -#else - PADDLE_THROW("Unexpected branch"); -#endif - return true; - } - - char* p = reinterpret_cast(dest); - while (total_written < length) { - if (!input->GetDirectBufferPointer(&data, &size_to_write)) { - return false; - } - // NOTE: if raw buffer is large and have two neighbor fields of raw buffers - // GetDirectBufferPointer can get all of them, use length to truncate it. - if (total_written + size_to_write > length) { - size_to_write = length - total_written; - } - // TODO(gongwb): can we avoid copy? - platform::CPUPlace cpu; - memory::Copy(cpu, reinterpret_cast(p), cpu, data, size_to_write); - - p += size_to_write; - total_written += size_to_write; - - input->Skip(size_to_write); - } - - return true; -} - -bool VariableResponse::CopyLodTensorData( - ::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& ctx, const framework::DDim& dims, - int length) { - auto* tensor = GetVar()->GetMutable(); - tensor->Resize(dims); - - framework::LoD lod; - for (int i = 0; i < meta_.lod_level(); ++i) { - framework::Vector v; - for (int j = 0; j < meta_.lod(i).lod_data_size(); ++j) { - v.push_back(meta_.lod(i).lod_data(j)); - } - lod.push_back(v); - } - tensor->set_lod(lod); - - void* tensor_data = - tensor->mutable_data(ctx.GetPlace(), ToTypeIndex(meta_.data_type())); - - if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { - return false; - } - - return true; -} - -inline framework::DDim GetDims( - const ::google::protobuf::RepeatedField<::google::protobuf::int64>& dims) { - std::vector vecdims; - for (auto& d : dims) { - vecdims.push_back(d); - } - return framework::make_ddim(vecdims); -} - -bool VariableResponse::CopySelectRowsTensorData( - ::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& ctx, const framework::DDim& dims, - int length) { - auto* slr = GetVar()->GetMutable(); - slr->set_height(meta_.slr_height()); - auto* tensor = slr->mutable_value(); - tensor->Resize(dims); - PADDLE_ENFORCE_EQ( - static_cast(tensor->numel()), - length / framework::SizeOfType( - paddle::operators::detail::ToTypeIndex(meta_.data_type()))); - void* tensor_data = tensor->mutable_data( - ctx.GetPlace(), - paddle::operators::detail::ToTypeIndex(meta_.data_type())); - - if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { - return false; - } - - return true; -} - -bool VariableResponse::CopySelectRowsData( - ::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& ctx, int length) { - auto* slr = GetVar()->GetMutable(); - slr->mutable_rows()->resize(length / - framework::SizeOfType(typeid(int64_t))); // int64 - int64_t* rows_data = slr->mutable_rows()->data(); - - // copy rows CPU data, GPU data will be copied lazily. - platform::CPUPlace cpu; - if (!ReadRaw(input, ctx, cpu, rows_data, length)) { - return false; - } - - return true; -} - -bool ParseLodData(::google::protobuf::io::CodedInputStream* input, - std::vector* lod) { - while (true) { - auto p = input->ReadTagWithCutoff(127); - int tag = GetTagFieldNumber(p.first); - WireType wt = GetTagWireType(p.first); - - if (!p.second) { - return (tag == 0); - } - - switch (tag) { - case sendrecv::VariableMessage_LodData::kLodDataFieldNumber: { - uint64_t v; - if (wt == WIRETYPE_VARINT) { - if (!input->ReadVarint64(&v)) { - return false; - } - lod->push_back(v); - break; - } - - if (wt == WIRETYPE_LENGTH_DELIMITED) { - int num_bytes = 0; - if (!input->ReadVarintSizeAsInt(&num_bytes)) { - return tag; - } - int start_pos = input->CurrentPosition(); - while (input->CurrentPosition() - start_pos < num_bytes) { - uint64_t v; - if (!input->ReadVarint64(&v)) { - return tag; - } - lod->push_back(v); - } - break; - } - - return false; - } - default: { return false; } - } - } - - return true; -} - -int VariableResponse::Parse(const ::grpc::ByteBuffer& byte_buffer) { - GrpcByteBufferSource source; - source.Init(byte_buffer); - GrpcByteBufferSourceWrapper r(&source); - - return Parse(&r); -} - -int VariableResponse::Parse(Source* source) { - ::google::protobuf::io::ZeroCopyInputStream* input_stream = - source->contents(); - ::google::protobuf::io::CodedInputStream input(input_stream); - input.SetTotalBytesLimit(INT_MAX, INT_MAX); - - while (true) { - auto p = input.ReadTagWithCutoff(127); - int tag = GetTagFieldNumber(p.first); - WireType wt = GetTagWireType(p.first); - if (!p.second) { - if (tag != 0) { - return -1; - } - return 0; - } - - switch (tag) { - case sendrecv::VariableMessage::kVarnameFieldNumber: { - uint32_t length; - if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) { - return tag; - } - - std::string temp; - if (!input.ReadString(&temp, length)) { - return tag; - } - - meta_.set_varname(temp); - break; - } - case sendrecv::VariableMessage::kTypeFieldNumber: { - uint32_t v; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { - return tag; - } - - meta_.set_type(static_cast<::sendrecv::VarType>(v)); - break; - } - case sendrecv::VariableMessage::kDataTypeFieldNumber: { - uint32_t v = 0; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { - return tag; - } - - meta_.set_data_type(static_cast<::sendrecv::VariableMessage_Type>(v)); - break; - } - case sendrecv::VariableMessage::kDimsFieldNumber: { - // not packed - if (wt == WIRETYPE_VARINT) { - uint64_t v; - if (!input.ReadVarint64(&v)) { - return tag; - } - meta_.add_dims(v); - break; - } - - // packed - if (wt == WIRETYPE_LENGTH_DELIMITED) { - int num_bytes = 0; - if (!input.ReadVarintSizeAsInt(&num_bytes)) { - return tag; - } - int start_pos = input.CurrentPosition(); - while (input.CurrentPosition() - start_pos < num_bytes) { - uint64_t v; - if (!input.ReadVarint64(&v)) { - return tag; - } - meta_.add_dims(v); - } - break; - } - return tag; - } - case sendrecv::VariableMessage::kLodLevelFieldNumber: { - uint64_t v = 0; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { - return tag; - } - meta_.set_lod_level(static_cast(v)); - break; - } - case sendrecv::VariableMessage::kLodFieldNumber: { - int length = 0; - if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &length)) { - return tag; - } - - std::pair<::google::protobuf::io::CodedInputStream::Limit, int> p = - input.IncrementRecursionDepthAndPushLimit(length); - - std::vector lod_data; - if (p.second < 0 || !ParseLodData(&input, &lod_data)) { - return tag; - } - - if (!input.DecrementRecursionDepthAndPopLimit(p.first)) { - return false; - } - - if (lod_data.size() == 0) { - break; - } - - auto lod = meta_.add_lod(); - for (uint32_t i = 0; i < lod_data.size(); i++) { - lod->add_lod_data(lod_data[i]); - } - break; - } - case sendrecv::VariableMessage::kSlrHeightFieldNumber: { - uint64_t v = 0; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { - return tag; - } - meta_.set_slr_height(static_cast(v)); - break; - } - case sendrecv::VariableMessage::kSerializedFieldNumber: { - PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || - meta_.type() == sendrecv::LOD_TENSOR || - meta_.type() == sendrecv::NCCL_ID) && - meta_.varname() != "", - "meta info should be got first!"); - - int num_bytes = 0; - if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &num_bytes)) { - return tag; - } - - if (meta_.type() == sendrecv::NCCL_ID) { -#ifdef PADDLE_WITH_CUDA - auto* var = scope_->FindVar(meta_.varname()); - if (var != nullptr) { - ncclUniqueId* id = var->GetMutable(); - if (!ReadRaw(&input, *dev_ctx_, platform::CPUPlace(), id->internal, - num_bytes)) { - return tag; - } - } - break; -#else - PADDLE_THROW("Not compiled with CUDA!"); -#endif - } - - framework::DDim dims = GetDims(meta_.dims()); - if (meta_.type() == sendrecv::LOD_TENSOR) { - PADDLE_ENFORCE(meta_.lod_size() >= 0, - "lod info should be got first!"); - if (!CopyLodTensorData(&input, *dev_ctx_, dims, num_bytes)) { - return tag; - } - break; - } - - if (meta_.type() == sendrecv::SELECTED_ROWS) { - if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, num_bytes)) { - return tag; - } - break; - } - - return tag; - } - case sendrecv::VariableMessage::kRowsFieldNumber: { - PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || - meta_.type() == sendrecv::LOD_TENSOR) && - meta_.varname() != "", - "meta info should be got first!"); - - int num_bytes = 0; - if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &num_bytes)) { - return tag; - } - - if (!CopySelectRowsData(&input, *dev_ctx_, num_bytes)) { - return tag; - } - break; - } - case sendrecv::VariableMessage::kOutVarnameFieldNumber: { - uint32_t length; - if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) { - return tag; - } - - std::string temp; - if (!input.ReadString(&temp, length)) { - return tag; - } - - meta_.set_out_varname(temp); - break; - } - case sendrecv::VariableMessage::kProfileFieldNumber: { - uint64_t profiling = 0; - if (!input.ReadVarint64(&profiling)) { - return tag; - } - meta_.set_profile(profiling); - int64_t listener_id = platform::ListenerId(); - if (listener_id <= 0) { - break; - } - if (profiling == platform::kEnableProfiler && - !platform::IsProfileEnabled()) { - platform::EnableProfiler(platform::ProfilerState::kCPU); - } else if (profiling == platform::kDisableProfiler && - platform::IsProfileEnabled()) { - // TODO(panyx0718): Should we allow to customize file dir. - platform::DisableProfiler( - platform::EventSortingKey::kDefault, - string::Sprintf("/tmp/profile_ps_%lld", listener_id)); - } - break; - } - default: { - // Unknown tag, return unknown error. - return -1; - } - } - } - - return 0; -} - -}; // namespace detail -}; // namespace operators -}; // namespace paddle diff --git a/paddle/fluid/operators/detail/variable_response.h b/paddle/fluid/operators/detail/variable_response.h deleted file mode 100644 index 69cfd784f8dd4f129f50c6882061e53e8535b949..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detail/variable_response.h +++ /dev/null @@ -1,104 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/framework/var_type.h" - -#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" - -#include "google/protobuf/io/coded_stream.h" -#include "google/protobuf/io/zero_copy_stream.h" -#include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/operators/detail/bytebuffer_stream.h" - -namespace paddle { -namespace operators { -namespace detail { - -class VariableResponse { - public: - VariableResponse(const framework::Scope* scope, - const platform::DeviceContext* dev_ctx, - bool create_scope = false) - : scope_(scope), dev_ctx_(dev_ctx), create_scope_(create_scope) { - if (create_scope) { - local_scope_ = &scope->NewScope(); - } - } - - virtual ~VariableResponse() { - if (create_scope_) { - scope_->DeleteScope(local_scope_); - } - } - - // return: - // 0:ok. - // -1: unkown error. - // other: number of error field. - int Parse(Source* source); - - // return: - // 0:ok. - // -1: unkown error. - // other: number of error field. - int Parse(const ::grpc::ByteBuffer& byte_buffer); - - const framework::Scope& GetLocalScope() const { return *local_scope_; } - - framework::Scope* GetMutableLocalScope() const { return local_scope_; } - - inline std::string Varname() const { return meta_.varname(); } - inline std::string OutVarname() const { return meta_.out_varname(); } - - // should call parse first. - framework::Variable* GetVar() { - if (create_scope_) { - return local_scope_->Var(meta_.varname()); - } - return scope_->FindVar(meta_.varname()); - } - - private: - bool CopySelectRowsTensorData(::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& ctx, - const framework::DDim& dims, int length); - - bool CopySelectRowsData(::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& ctx, int length); - - bool CopyLodTensorData(::google::protobuf::io::CodedInputStream* input, - const platform::DeviceContext& ctx, - const framework::DDim& dims, int length); - - private: - const framework::Scope* scope_; - const platform::DeviceContext* dev_ctx_; - bool create_scope_ = false; - framework::Scope* local_scope_ = nullptr; - // only Skeleton - sendrecv::VariableMessage meta_; -}; - -}; // namespace detail -}; // namespace operators -}; // namespace paddle diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt index 20d960f9fee1eae42b2241fb96c163e15db5e24d..a44d84cd7b99107fef09a6b4dfa60172fabd718b 100644 --- a/paddle/fluid/operators/detection/CMakeLists.txt +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -22,10 +22,13 @@ iou_similarity_op.cu) detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc) detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc) detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu) +detection_library(anchor_generator_op SRCS anchor_generator_op.cc +anchor_generator_op.cu) detection_library(target_assign_op SRCS target_assign_op.cc target_assign_op.cu) detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc - polygon_box_transform_op.cu) +polygon_box_transform_op.cu) +detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc) # Export local libraries to parent set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE) diff --git a/paddle/fluid/operators/detection/anchor_generator_op.cc b/paddle/fluid/operators/detection/anchor_generator_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0c0155a0a977846b1300d93b4c3fef0e71fc1d26 --- /dev/null +++ b/paddle/fluid/operators/detection/anchor_generator_op.cc @@ -0,0 +1,154 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/detection/anchor_generator_op.h" + +namespace paddle { +namespace operators { + +class AnchorGeneratorOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of AnchorGeneratorOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Anchors"), + "Output(Anchors) of AnchorGeneratorOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Variances"), + "Output(Variances) of AnchorGeneratorOp should not be null."); + + auto input_dims = ctx->GetInputDim("Input"); + PADDLE_ENFORCE(input_dims.size() == 4, "The layout of input is NCHW."); + + auto anchor_sizes = ctx->Attrs().Get>("anchor_sizes"); + auto aspect_ratios = ctx->Attrs().Get>("aspect_ratios"); + auto stride = ctx->Attrs().Get>("stride"); + auto variances = ctx->Attrs().Get>("variances"); + + size_t num_anchors = aspect_ratios.size() * anchor_sizes.size(); + + std::vector dim_vec(4); + dim_vec[0] = input_dims[2]; + dim_vec[1] = input_dims[3]; + dim_vec[2] = num_anchors; + dim_vec[3] = 4; + ctx->SetOutputDim("Anchors", framework::make_ddim(dim_vec)); + ctx->SetOutputDim("Variances", framework::make_ddim(dim_vec)); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Input")->type()), + ctx.device_context()); + } +}; + +class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", + "(Tensor, default Tensor), " + "the input feature is a tensor with a rank of 4. " + "The layout is NCHW."); + AddOutput("Anchors", + "(Tensor, default Tensor), the output is a " + "tensor with a rank of 4. The layout is [H, W, num_anchors, 4]. " + "H is the height of input, W is the width of input, num_anchors " + "is the box count of each position. " + "Each anchor is in (xmin, ymin, xmax, ymax) format"); + AddOutput("Variances", + "(Tensor, default Tensor), the expanded variances for " + "normalizing bbox regression targets. The layout is [H, W, " + "num_anchors, 4]. " + "H is the height of input, W is the width of input, num_anchors " + "is the box count of each position. " + "Each variance is in (xcenter, ycenter, w, h) format"); + + AddAttr>( + "anchor_sizes", + "(vector) List of Region Proposal Network(RPN) anchor sizes " + " given in absolute pixels e.g. (64, 128, 256, 512)." + " For instance, the anchor size of 64 means the area of this anchor " + "equals to 64**2.") + .AddCustomChecker([](const std::vector& anchor_sizes) { + PADDLE_ENFORCE_GT(anchor_sizes.size(), 0, + "Size of anchor_sizes must be at least 1."); + for (size_t i = 0; i < anchor_sizes.size(); ++i) { + PADDLE_ENFORCE_GT(anchor_sizes[i], 0.0, + "anchor_sizes[%d] must be positive.", i); + } + }); + AddAttr>( + "aspect_ratios", + "(vector) List of Region Proposal Network(RPN) anchor aspect " + "ratios, e.g. (0.5, 1, 2)." + "For instacne, the aspect ratio of 0.5 means the height / width of " + "this anchor equals 0.5."); + + AddAttr>("variances", + "(vector) List of variances to be used " + "in box regression deltas") + .AddCustomChecker([](const std::vector& variances) { + PADDLE_ENFORCE_EQ(variances.size(), 4, + "Must and only provide 4 variance."); + for (size_t i = 0; i < variances.size(); ++i) { + PADDLE_ENFORCE_GT(variances[i], 0.0, + "variance[%d] must be greater than 0.", i); + } + }); + + AddAttr>("stride", + "Anchors stride across width and height, " + "with a default of (16, 16)") + .SetDefault(std::vector(2, 16.0)) + .AddCustomChecker([](const std::vector& stride) { + PADDLE_ENFORCE_EQ( + stride.size(), 2, + "Must and only provide 2 stride for width and height."); + for (size_t i = 0; i < stride.size(); ++i) { + PADDLE_ENFORCE_GT(stride[i], 0.0, + "stride[%d] should be larger than 0.", i); + } + }); + + AddAttr("offset", + "(float) " + "Anchor center offset, with a default of 0.5") + .SetDefault(0.5); + AddComment(R"DOC( +AnchorGenerator operator +Generates anchors for Faster RCNN, FPN etc. algorithm. +Each position of the input produce N anchors, N = + size(anchor_sizes) * size(aspect_ratios). + +Please get more information from the following papers: +https://arxiv.org/abs/1506.01497. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(anchor_generator, ops::AnchorGeneratorOp, + ops::AnchorGeneratorOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL(anchor_generator, ops::AnchorGeneratorOpKernel, + ops::AnchorGeneratorOpKernel); diff --git a/paddle/fluid/operators/detection/anchor_generator_op.cu b/paddle/fluid/operators/detection/anchor_generator_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..3cc9bbeee1eeed17142a6b1bd23b45aff9cf745f --- /dev/null +++ b/paddle/fluid/operators/detection/anchor_generator_op.cu @@ -0,0 +1,132 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/detection/anchor_generator_op.h" + +namespace paddle { +namespace operators { + +template +__global__ void GenAnchors(T* out, const T* aspect_ratios, const int ar_num, + const T* anchor_sizes, const int as_num, + const T* stride, const int sd_num, const int height, + const int width, const T offset) { + int num_anchors = as_num * ar_num; + int box_num = height * width * num_anchors; + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < box_num; + i += blockDim.x * gridDim.x) { + int h_idx = i / (num_anchors * width); + int w_idx = (i / num_anchors) % width; + T stride_width = stride[0]; + T stride_height = stride[1]; + T x_ctr = (w_idx * stride_width) + offset * (stride_width - 1); + T y_ctr = (h_idx * stride_height) + offset * (stride_height - 1); + T area, area_ratios; + T base_w, base_h; + T scale_w, scale_h; + T anchor_width, anchor_height; + int anch_idx = i % num_anchors; + int ar_idx = anch_idx / as_num; + int as_idx = anch_idx % as_num; + T aspect_ratio = aspect_ratios[ar_idx]; + T anchor_size = anchor_sizes[as_idx]; + area = stride_width * stride_height; + area_ratios = area / aspect_ratio; + base_w = round(sqrt(area_ratios)); + base_h = round(base_w * aspect_ratio); + scale_w = anchor_size / stride_width; + scale_h = anchor_size / stride_height; + anchor_width = scale_w * base_w; + anchor_height = scale_h * base_h; + + T xmin = (x_ctr - 0.5 * (anchor_width - 1)); + T ymin = (y_ctr - 0.5 * (anchor_height - 1)); + T xmax = (x_ctr + 0.5 * (anchor_width - 1)); + T ymax = (y_ctr + 0.5 * (anchor_height - 1)); + out[i * 4] = xmin; + out[i * 4 + 1] = ymin; + out[i * 4 + 2] = xmax; + out[i * 4 + 3] = ymax; + } +} + +template +__global__ void SetVariance(T* out, const T* var, const int vnum, + const int num) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + out[i] = var[i % vnum]; + } +} + +template +class AnchorGeneratorOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("Input"); + auto* anchors = ctx.Output("Anchors"); + auto* vars = ctx.Output("Variances"); + + auto anchor_sizes = ctx.Attr>("anchor_sizes"); + auto aspect_ratios = ctx.Attr>("aspect_ratios"); + auto stride = ctx.Attr>("stride"); + auto variances = ctx.Attr>("variances"); + + T offset = static_cast(ctx.Attr("offset")); + + auto width = input->dims()[3]; + auto height = input->dims()[2]; + + int num_anchors = aspect_ratios.size() * anchor_sizes.size(); + + int box_num = width * height * num_anchors; + + int block = 512; + int grid = (box_num + block - 1) / block; + + auto stream = + ctx.template device_context().stream(); + + anchors->mutable_data(ctx.GetPlace()); + vars->mutable_data(ctx.GetPlace()); + + framework::Tensor ar; + framework::TensorFromVector(aspect_ratios, ctx.device_context(), &ar); + + framework::Tensor as; + framework::TensorFromVector(anchor_sizes, ctx.device_context(), &as); + + framework::Tensor sd; + framework::TensorFromVector(stride, ctx.device_context(), &sd); + + GenAnchors<<>>( + anchors->data(), ar.data(), aspect_ratios.size(), as.data(), + anchor_sizes.size(), sd.data(), stride.size(), height, width, + offset); + + framework::Tensor v; + framework::TensorFromVector(variances, ctx.device_context(), &v); + grid = (box_num * 4 + block - 1) / block; + SetVariance<<>>(vars->data(), v.data(), + variances.size(), box_num * 4); + } +}; // namespace operators + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(anchor_generator, + ops::AnchorGeneratorOpCUDAKernel, + ops::AnchorGeneratorOpCUDAKernel); diff --git a/paddle/fluid/operators/detection/anchor_generator_op.h b/paddle/fluid/operators/detection/anchor_generator_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e0e499d76a19ba5f6b91ba4c8797684fb53c7caa --- /dev/null +++ b/paddle/fluid/operators/detection/anchor_generator_op.h @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/transform.h" + +namespace paddle { +namespace operators { + +template +class AnchorGeneratorOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("Input"); + auto* anchors = ctx.Output("Anchors"); + auto* vars = ctx.Output("Variances"); + + auto anchor_sizes = ctx.Attr>("anchor_sizes"); + auto aspect_ratios = ctx.Attr>("aspect_ratios"); + auto stride = ctx.Attr>("stride"); + auto variances = ctx.Attr>("variances"); + + T offset = static_cast(ctx.Attr("offset")); + + auto feature_width = input->dims()[3]; + auto feature_height = input->dims()[2]; + + T stride_width, stride_height; + stride_width = stride[0]; + stride_height = stride[1]; + + int num_anchors = aspect_ratios.size() * anchor_sizes.size(); + + anchors->mutable_data(ctx.GetPlace()); + vars->mutable_data(ctx.GetPlace()); + + auto e_anchors = framework::EigenTensor::From(*anchors); + for (int h_idx = 0; h_idx < feature_height; ++h_idx) { + for (int w_idx = 0; w_idx < feature_width; ++w_idx) { + T x_ctr = (w_idx * stride_width) + offset * (stride_width - 1); + T y_ctr = (h_idx * stride_height) + offset * (stride_height - 1); + T area, area_ratios; + T base_w, base_h; + T scale_w, scale_h; + T anchor_width, anchor_height; + int idx = 0; + for (size_t r = 0; r < aspect_ratios.size(); ++r) { + auto ar = aspect_ratios[r]; + for (size_t s = 0; s < anchor_sizes.size(); ++s) { + auto anchor_size = anchor_sizes[s]; + area = stride_width * stride_height; + area_ratios = area / ar; + base_w = round(sqrt(area_ratios)); + base_h = round(base_w * ar); + scale_w = anchor_size / stride_width; + scale_h = anchor_size / stride_height; + anchor_width = scale_w * base_w; + anchor_height = scale_h * base_h; + e_anchors(h_idx, w_idx, idx, 0) = + (x_ctr - 0.5 * (anchor_width - 1)); + e_anchors(h_idx, w_idx, idx, 1) = + (y_ctr - 0.5 * (anchor_height - 1)); + e_anchors(h_idx, w_idx, idx, 2) = + (x_ctr + 0.5 * (anchor_width - 1)); + e_anchors(h_idx, w_idx, idx, 3) = + (y_ctr + 0.5 * (anchor_height - 1)); + idx++; + } + } + } + } + + framework::Tensor var_t; + var_t.mutable_data( + framework::make_ddim({1, static_cast(variances.size())}), + ctx.GetPlace()); + auto var_et = framework::EigenTensor::From(var_t); + for (size_t i = 0; i < variances.size(); ++i) { + var_et(0, i) = variances[i]; + } + + int anchor_num = feature_height * feature_width * num_anchors; + auto var_dim = vars->dims(); + vars->Resize({anchor_num, static_cast(variances.size())}); + + auto e_vars = framework::EigenMatrix::From(*vars); + e_vars = var_et.broadcast(Eigen::DSizes(anchor_num, 1)); + + vars->Resize(var_dim); + } +}; // namespace operators + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detection/bipartite_match_op.cc b/paddle/fluid/operators/detection/bipartite_match_op.cc index d437ad5c19828331c749244404ba80d0f3acda2a..c23b65fe4dead3ca01a447d03877e3359b19e656 100644 --- a/paddle/fluid/operators/detection/bipartite_match_op.cc +++ b/paddle/fluid/operators/detection/bipartite_match_op.cc @@ -51,6 +51,12 @@ class BipartiteMatchOp : public framework::OperatorWithKernel { } }; +template +bool DistPairDescend(std::tuple pair1, + std::tuple pair2) { + return std::get<2>(pair1) > std::get<2>(pair2); +} + template class BipartiteMatchKernel : public framework::OpKernel { public: @@ -58,46 +64,76 @@ class BipartiteMatchKernel : public framework::OpKernel { // The match_dist must be initialized to 0 at first. void BipartiteMatch(const Tensor& dist, int* match_indices, T* match_dist) const { - constexpr T kEPS = static_cast(1e-6); PADDLE_ENFORCE_EQ(dist.dims().size(), 2, "The rank of dist must be 2."); int64_t row = dist.dims()[0]; int64_t col = dist.dims()[1]; auto* dist_data = dist.data(); - std::vector row_pool; - for (int i = 0; i < row; ++i) { - row_pool.push_back(i); - } - while (row_pool.size() > 0) { - int max_idx = -1; - int max_row_idx = -1; - T max_dist = -1; - for (int64_t j = 0; j < col; ++j) { - if (match_indices[j] != -1) { - continue; + // Test result: When row==130 the speed of these two methods almost the same + if (row >= 130) { + std::vector> match_pair; + + for (int64_t i = 0; i < row; ++i) { + for (int64_t j = 0; j < col; ++j) { + match_pair.push_back(std::make_tuple(i, j, dist_data[i * col + j])); } - for (size_t k = 0; k < row_pool.size(); ++k) { - int m = row_pool[k]; - // distance is 0 between m-th row and j-th column - if (dist_data[m * col + j] < kEPS) { + } + std::sort(match_pair.begin(), match_pair.end(), DistPairDescend); + std::vector row_indices(row, -1); + + int64_t idx = 0; + for (int64_t k = 0; k < row * col; ++k) { + int64_t i = std::get<0>(match_pair[k]); + int64_t j = std::get<1>(match_pair[k]); + T dist = std::get<2>(match_pair[k]); + + if (idx >= row) { + break; + } + if (match_indices[j] == -1 && row_indices[i] == -1 && dist > 0) { + match_indices[j] = i; + row_indices[i] = j; + match_dist[j] = dist; + idx += 1; + } + } + } else { + constexpr T kEPS = static_cast(1e-6); + std::vector row_pool; + for (int i = 0; i < row; ++i) { + row_pool.push_back(i); + } + while (row_pool.size() > 0) { + int max_idx = -1; + int max_row_idx = -1; + T max_dist = -1; + for (int64_t j = 0; j < col; ++j) { + if (match_indices[j] != -1) { continue; } - if (dist_data[m * col + j] > max_dist) { - max_idx = j; - max_row_idx = m; - max_dist = dist_data[m * col + j]; + for (size_t k = 0; k < row_pool.size(); ++k) { + int m = row_pool[k]; + // distance is 0 between m-th row and j-th column + if (dist_data[m * col + j] < kEPS) { + continue; + } + if (dist_data[m * col + j] > max_dist) { + max_idx = j; + max_row_idx = m; + max_dist = dist_data[m * col + j]; + } } } - } - if (max_idx == -1) { - // Cannot find good match. - break; - } else { - PADDLE_ENFORCE_EQ(match_indices[max_idx], -1); - match_indices[max_idx] = max_row_idx; - match_dist[max_idx] = max_dist; - // Erase the row index. - row_pool.erase( - std::find(row_pool.begin(), row_pool.end(), max_row_idx)); + if (max_idx == -1) { + // Cannot find good match. + break; + } else { + PADDLE_ENFORCE_EQ(match_indices[max_idx], -1); + match_indices[max_idx] = max_row_idx; + match_dist[max_idx] = max_dist; + // Erase the row index. + row_pool.erase( + std::find(row_pool.begin(), row_pool.end(), max_row_idx)); + } } } } diff --git a/paddle/fluid/operators/detection/box_coder_op.cc b/paddle/fluid/operators/detection/box_coder_op.cc index 74848005d0bea6e5c818ff999727aa2b8ad51d84..d0f95f727fdbc82777147e3e8ada6ad4f7a35e60 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cc +++ b/paddle/fluid/operators/detection/box_coder_op.cc @@ -22,21 +22,21 @@ class BoxCoderOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("PriorBox"), "Input(PriorBox) of BoxCoderOp should not be null."); - PADDLE_ENFORCE(ctx->HasInput("PriorBoxVar"), - "Input(PriorBoxVar) of BoxCoderOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("TargetBox"), "Input(TargetBox) of BoxCoderOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("OutputBox"), "Output(OutputBox) of BoxCoderOp should not be null."); auto prior_box_dims = ctx->GetInputDim("PriorBox"); - auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar"); auto target_box_dims = ctx->GetInputDim("TargetBox"); PADDLE_ENFORCE_EQ(prior_box_dims.size(), 2, "The rank of Input of PriorBoxVar must be 2"); PADDLE_ENFORCE_EQ(prior_box_dims[1], 4, "The shape of PriorBox is [N, 4]"); - PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims); + if (ctx->HasInput("PriorBoxVar")) { + auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar"); + PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims); + } auto code_type = GetBoxCodeType(ctx->Attrs().Get("code_type")); if (code_type == BoxCodeType::kEncodeCenterSize) { @@ -71,9 +71,11 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "of the coordinate system. [xmax, ymax] is the right bottom " "coordinate of the anchor box."); AddInput("PriorBoxVar", - "(Tensor, default Tensor) " + "(Tensor, default Tensor, optional) " "PriorBoxVar is a 2-D Tensor with shape [M, 4] holds M group " - "of variance."); + "of variance. PriorBoxVar will set all elements to 1 by " + "default.") + .AsDispensable(); AddInput( "TargetBox", "(LoDTensor or Tensor) This input can be a 2-D LoDTensor with shape " @@ -91,6 +93,10 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "the code type used with the target box") .SetDefault("encode_center_size") .InEnum({"encode_center_size", "decode_center_size"}); + AddAttr("box_normalized", + "(bool, default true) " + "whether treat the priorbox as a noramlized box") + .SetDefault(true); AddOutput("OutputBox", "(LoDTensor or Tensor) " "When code_type is 'encode_center_size', the output tensor of " @@ -100,23 +106,36 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "and M represents the number of deocded boxes."); AddComment(R"DOC( -Bounding Box Coder Operator. + +Bounding Box Coder. + Encode/Decode the target bounding box with the priorbox information. + The Encoding schema described below: -ox = (tx - px) / pw / pxv -oy = (ty - py) / ph / pyv -ow = log(abs(tw / pw)) / pwv -oh = log(abs(th / ph)) / phv + + ox = (tx - px) / pw / pxv + + oy = (ty - py) / ph / pyv + + ow = log(abs(tw / pw)) / pwv + + oh = log(abs(th / ph)) / phv + The Decoding schema described below: -ox = (pw * pxv * tx * + px) - tw / 2 -oy = (ph * pyv * ty * + py) - th / 2 -ow = exp(pwv * tw) * pw + tw / 2 -oh = exp(phv * th) * ph + th / 2 -where tx, ty, tw, th denote the target box's center coordinates, width and -height respectively. Similarly, px, py, pw, ph denote the priorbox's(anchor) -center coordinates, width and height. pxv, pyv, pwv, phv denote the variance -of the priorbox and ox, oy, ow, oh denote the encoded/decoded coordinates, -width and height. + + ox = (pw * pxv * tx * + px) - tw / 2 + + oy = (ph * pyv * ty * + py) - th / 2 + + ow = exp(pwv * tw) * pw + tw / 2 + + oh = exp(phv * th) * ph + th / 2 + +where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, width +and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote the +priorbox's (anchor) center coordinates, width and height. `pxv`, `pyv`, `pwv`, +`phv` denote the variance of the priorbox and `ox`, `oy`, `ow`, `oh` denote the +encoded/decoded coordinates, width and height. )DOC"); } }; @@ -127,5 +146,6 @@ width and height. namespace ops = paddle::operators; REGISTER_OPERATOR(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker, paddle::framework::EmptyGradOpMaker); -REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel, - ops::BoxCoderKernel); +REGISTER_OP_CPU_KERNEL( + box_coder, ops::BoxCoderKernel, + ops::BoxCoderKernel); diff --git a/paddle/fluid/operators/detection/box_coder_op.cu b/paddle/fluid/operators/detection/box_coder_op.cu index 8cef8e03439df4ca5b0fa94176a21a36f9eb9f70..a7af111f63d654319dd1d90d2032956951dfe49e 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cu +++ b/paddle/fluid/operators/detection/box_coder_op.cu @@ -20,15 +20,16 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data, const T* prior_box_var_data, const T* target_box_data, const int row, const int col, const int len, - T* output) { + const bool normalized, T* output) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < row * col) { const int row_idx = idx / col; const int col_idx = idx % col; - T prior_box_width = - prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len]; - T prior_box_height = - prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1]; + T prior_box_width = prior_box_data[col_idx * len + 2] - + prior_box_data[col_idx * len] + (normalized == false); + T prior_box_height = prior_box_data[col_idx * len + 3] - + prior_box_data[col_idx * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2; T prior_box_center_y = (prior_box_data[col_idx * len + 3] + @@ -41,20 +42,24 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data, T target_box_center_y = (target_box_data[row_idx * len + 3] + target_box_data[row_idx * len + 1]) / 2; - T target_box_width = - target_box_data[row_idx * len + 2] - target_box_data[row_idx * len]; - T target_box_height = - target_box_data[row_idx * len + 3] - target_box_data[row_idx * len + 1]; + T target_box_width = target_box_data[row_idx * len + 2] - + target_box_data[row_idx * len] + (normalized == false); + T target_box_height = target_box_data[row_idx * len + 3] - + target_box_data[row_idx * len + 1] + + (normalized == false); - output[idx * len] = (target_box_center_x - prior_box_center_x) / - prior_box_width / prior_box_var_data[col_idx * len]; - output[idx * len + 1] = (target_box_center_y - prior_box_center_y) / - prior_box_height / - prior_box_var_data[col_idx * len + 1]; - output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)) / - prior_box_var_data[col_idx * len + 2]; - output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)) / - prior_box_var_data[col_idx * len + 3]; + output[idx * len] = + (target_box_center_x - prior_box_center_x) / prior_box_width; + output[idx * len + 1] = + (target_box_center_y - prior_box_center_y) / prior_box_height; + output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)); + output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)); + if (prior_box_var_data) { + output[idx * len] /= prior_box_var_data[col_idx * len]; + output[idx * len + 1] /= prior_box_var_data[col_idx * len + 1]; + output[idx * len + 2] /= prior_box_var_data[col_idx * len + 2]; + output[idx * len + 3] /= prior_box_var_data[col_idx * len + 3]; + } } } @@ -63,42 +68,56 @@ __global__ void DecodeCenterSizeKernel(const T* prior_box_data, const T* prior_box_var_data, const T* target_box_data, const int row, const int col, const int len, - T* output) { + const bool normalized, T* output) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < row * col) { const int col_idx = idx % col; - T prior_box_width = - prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len]; - T prior_box_height = - prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1]; + T prior_box_width = prior_box_data[col_idx * len + 2] - + prior_box_data[col_idx * len] + (normalized == false); + T prior_box_height = prior_box_data[col_idx * len + 3] - + prior_box_data[col_idx * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2; T prior_box_center_y = (prior_box_data[col_idx * len + 3] + prior_box_data[col_idx * len + 1]) / 2; - - T target_box_width = exp(prior_box_var_data[col_idx * len + 2] * + T target_box_width, target_box_height; + T target_box_center_x, target_box_center_y; + if (prior_box_var_data) { + target_box_width = exp(prior_box_var_data[col_idx * len + 2] * target_box_data[idx * len + 2]) * prior_box_width; - T target_box_height = exp(prior_box_var_data[col_idx * len + 3] * + target_box_height = exp(prior_box_var_data[col_idx * len + 3] * target_box_data[idx * len + 3]) * prior_box_height; - T target_box_center_x = prior_box_var_data[col_idx * len] * + target_box_center_x = prior_box_var_data[col_idx * len] * target_box_data[idx * len] * prior_box_width + prior_box_center_x; - T target_box_center_y = prior_box_var_data[col_idx * len + 1] * + target_box_center_y = prior_box_var_data[col_idx * len + 1] * target_box_data[idx * len + 1] * prior_box_height + prior_box_center_y; + } else { + target_box_width = exp(target_box_data[idx * len + 2]) * prior_box_width; + target_box_height = + exp(target_box_data[idx * len + 3]) * prior_box_height; + target_box_center_x = + target_box_data[idx * len] * prior_box_width + prior_box_center_x; + target_box_center_y = target_box_data[idx * len + 1] * prior_box_height + + prior_box_center_y; + } output[idx * len] = target_box_center_x - target_box_width / 2; output[idx * len + 1] = target_box_center_y - target_box_height / 2; - output[idx * len + 2] = target_box_center_x + target_box_width / 2; - output[idx * len + 3] = target_box_center_y + target_box_height / 2; + output[idx * len + 2] = + target_box_center_x + target_box_width / 2 - (normalized == false); + output[idx * len + 3] = + target_box_center_y + target_box_height / 2 - (normalized == false); } } -template +template class BoxCoderCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -109,6 +128,11 @@ class BoxCoderCUDAKernel : public framework::OpKernel { auto* target_box = context.Input("TargetBox"); auto* output_box = context.Output("OutputBox"); + const T* prior_box_data = prior_box->data(); + const T* target_box_data = target_box->data(); + const T* prior_box_var_data = nullptr; + if (prior_box_var) prior_box_var_data = prior_box_var->data(); + if (target_box->lod().size()) { PADDLE_ENFORCE_EQ(target_box->lod().size(), 1, "Only support 1 level of LoD."); @@ -120,22 +144,19 @@ class BoxCoderCUDAKernel : public framework::OpKernel { int grid = (row * col + block - 1) / block; auto& device_ctx = context.cuda_device_context(); - const T* prior_box_data = prior_box->data(); - const T* prior_box_var_data = prior_box_var->data(); - const T* target_box_data = target_box->data(); - output_box->mutable_data({row, col, len}, context.GetPlace()); T* output = output_box->data(); auto code_type = GetBoxCodeType(context.Attr("code_type")); + bool normalized = context.Attr("box_normalized"); if (code_type == BoxCodeType::kEncodeCenterSize) { EncodeCenterSizeKernel<<>>( prior_box_data, prior_box_var_data, target_box_data, row, col, len, - output); + normalized, output); } else if (code_type == BoxCodeType::kDecodeCenterSize) { DecodeCenterSizeKernel<<>>( prior_box_data, prior_box_var_data, target_box_data, row, col, len, - output); + normalized, output); } } }; @@ -144,5 +165,7 @@ class BoxCoderCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(box_coder, ops::BoxCoderCUDAKernel, - ops::BoxCoderCUDAKernel); +REGISTER_OP_CUDA_KERNEL( + box_coder, + ops::BoxCoderCUDAKernel, + ops::BoxCoderCUDAKernel); diff --git a/paddle/fluid/operators/detection/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h index 77fc6c2b62af42e6526b889aeef2d9bab795baec..5ed8520acddfa8fe2105a7c1615bcb3243cb130f 100644 --- a/paddle/fluid/operators/detection/box_coder_op.h +++ b/paddle/fluid/operators/detection/box_coder_op.h @@ -28,26 +28,28 @@ inline BoxCodeType GetBoxCodeType(const std::string& type) { PADDLE_THROW("Not support type %s.", type); } -template +template class BoxCoderKernel : public framework::OpKernel { public: - void EncodeCenterSize(const framework::Tensor& target_box, - const framework::Tensor& prior_box, - const framework::Tensor& prior_box_var, - T* output) const { - int64_t row = target_box.dims()[0]; - int64_t col = prior_box.dims()[0]; - int64_t len = prior_box.dims()[1]; - auto* target_box_data = target_box.data(); - auto* prior_box_data = prior_box.data(); - auto* prior_box_var_data = prior_box_var.data(); + void EncodeCenterSize(const framework::Tensor* target_box, + const framework::Tensor* prior_box, + const framework::Tensor* prior_box_var, + const bool normalized, T* output) const { + int64_t row = target_box->dims()[0]; + int64_t col = prior_box->dims()[0]; + int64_t len = prior_box->dims()[1]; + auto* target_box_data = target_box->data(); + auto* prior_box_data = prior_box->data(); + const T* prior_box_var_data = nullptr; + if (prior_box_var) prior_box_var_data = prior_box_var->data(); for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { - T prior_box_width = - prior_box_data[j * len + 2] - prior_box_data[j * len]; - T prior_box_height = - prior_box_data[j * len + 3] - prior_box_data[j * len + 1]; + T prior_box_width = prior_box_data[j * len + 2] - + prior_box_data[j * len] + (normalized == false); + T prior_box_height = prior_box_data[j * len + 3] - + prior_box_data[j * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2; T prior_box_center_y = @@ -57,67 +59,89 @@ class BoxCoderKernel : public framework::OpKernel { (target_box_data[i * len + 2] + target_box_data[i * len]) / 2; T target_box_center_y = (target_box_data[i * len + 3] + target_box_data[i * len + 1]) / 2; - T target_box_width = - target_box_data[i * len + 2] - target_box_data[i * len]; - T target_box_height = - target_box_data[i * len + 3] - target_box_data[i * len + 1]; + T target_box_width = target_box_data[i * len + 2] - + target_box_data[i * len] + (normalized == false); + T target_box_height = target_box_data[i * len + 3] - + target_box_data[i * len + 1] + + (normalized == false); size_t offset = i * col * len + j * len; - output[offset] = (target_box_center_x - prior_box_center_x) / - prior_box_width / prior_box_var_data[j * len]; - output[offset + 1] = (target_box_center_y - prior_box_center_y) / - prior_box_height / prior_box_var_data[j * len + 1]; + output[offset] = + (target_box_center_x - prior_box_center_x) / prior_box_width; + output[offset + 1] = + (target_box_center_y - prior_box_center_y) / prior_box_height; output[offset + 2] = - std::log(std::fabs(target_box_width / prior_box_width)) / - prior_box_var_data[j * len + 2]; + std::log(std::fabs(target_box_width / prior_box_width)); output[offset + 3] = - std::log(std::fabs(target_box_height / prior_box_height)) / - prior_box_var_data[j * len + 3]; + std::log(std::fabs(target_box_height / prior_box_height)); + if (prior_box_var) { + output[offset] /= prior_box_var_data[j * len]; + output[offset + 1] /= prior_box_var_data[j * len + 1]; + output[offset + 2] /= prior_box_var_data[j * len + 2]; + output[offset + 3] /= prior_box_var_data[j * len + 3]; + } } } } - void DecodeCenterSize(const framework::Tensor& target_box, - const framework::Tensor& prior_box, - const framework::Tensor& prior_box_var, - T* output) const { - int64_t row = target_box.dims()[0]; - int64_t col = prior_box.dims()[0]; - int64_t len = prior_box.dims()[1]; - - auto* target_box_data = target_box.data(); - auto* prior_box_data = prior_box.data(); - auto* prior_box_var_data = prior_box_var.data(); + void DecodeCenterSize(const framework::Tensor* target_box, + const framework::Tensor* prior_box, + const framework::Tensor* prior_box_var, + const bool normalized, T* output) const { + int64_t row = target_box->dims()[0]; + int64_t col = prior_box->dims()[0]; + int64_t len = prior_box->dims()[1]; + + auto* target_box_data = target_box->data(); + auto* prior_box_data = prior_box->data(); + const T* prior_box_var_data = nullptr; + if (prior_box_var) prior_box_var_data = prior_box_var->data(); for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { size_t offset = i * col * len + j * len; - T prior_box_width = - prior_box_data[j * len + 2] - prior_box_data[j * len]; - T prior_box_height = - prior_box_data[j * len + 3] - prior_box_data[j * len + 1]; + T prior_box_width = prior_box_data[j * len + 2] - + prior_box_data[j * len] + (normalized == false); + T prior_box_height = prior_box_data[j * len + 3] - + prior_box_data[j * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2; T prior_box_center_y = (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2; - T target_box_center_x = prior_box_var_data[j * len] * + T target_box_center_x = 0, target_box_center_y = 0; + T target_box_width = 0, target_box_height = 0; + if (prior_box_var) { + target_box_center_x = prior_box_var_data[j * len] * target_box_data[offset] * prior_box_width + prior_box_center_x; - T target_box_center_y = prior_box_var_data[j * len + 1] * + target_box_center_y = prior_box_var_data[j * len + 1] * target_box_data[offset + 1] * prior_box_height + prior_box_center_y; - T target_box_width = std::exp(prior_box_var_data[j * len + 2] * + target_box_width = std::exp(prior_box_var_data[j * len + 2] * target_box_data[offset + 2]) * prior_box_width; - T target_box_height = std::exp(prior_box_var_data[j * len + 3] * + target_box_height = std::exp(prior_box_var_data[j * len + 3] * target_box_data[offset + 3]) * prior_box_height; + } else { + target_box_center_x = + target_box_data[offset] * prior_box_width + prior_box_center_x; + target_box_center_y = target_box_data[offset + 1] * prior_box_height + + prior_box_center_y; + target_box_width = + std::exp(target_box_data[offset + 2]) * prior_box_width; + target_box_height = + std::exp(target_box_data[offset + 3]) * prior_box_height; + } output[offset] = target_box_center_x - target_box_width / 2; output[offset + 1] = target_box_center_y - target_box_height / 2; - output[offset + 2] = target_box_center_x + target_box_width / 2; - output[offset + 3] = target_box_center_y + target_box_height / 2; + output[offset + 2] = + target_box_center_x + target_box_width / 2 - (normalized == false); + output[offset + 3] = + target_box_center_y + target_box_height / 2 - (normalized == false); } } } @@ -139,11 +163,14 @@ class BoxCoderKernel : public framework::OpKernel { output_box->mutable_data({row, col, len}, context.GetPlace()); auto code_type = GetBoxCodeType(context.Attr("code_type")); + bool normalized = context.Attr("box_normalized"); T* output = output_box->data(); if (code_type == BoxCodeType::kEncodeCenterSize) { - EncodeCenterSize(*target_box, *prior_box, *prior_box_var, output); + EncodeCenterSize(target_box, prior_box, prior_box_var, normalized, + output); } else if (code_type == BoxCodeType::kDecodeCenterSize) { - DecodeCenterSize(*target_box, *prior_box, *prior_box_var, output); + DecodeCenterSize(target_box, prior_box, prior_box_var, normalized, + output); } } }; diff --git a/paddle/fluid/operators/detection/iou_similarity_op.cc b/paddle/fluid/operators/detection/iou_similarity_op.cc index 8e58605fcea04f9ffa97ce8cca53c073e7068aaf..9c89b7ca9af1b235659554afc805600d31ef8ea6 100644 --- a/paddle/fluid/operators/detection/iou_similarity_op.cc +++ b/paddle/fluid/operators/detection/iou_similarity_op.cc @@ -68,15 +68,16 @@ class IOUSimilarityOpMaker : public framework::OpProtoAndCheckerMaker { "representing pairwise iou scores."); AddComment(R"DOC( -IOU Similarity Operator. +**IOU Similarity Operator** + Computes intersection-over-union (IOU) between two box lists. - Box list 'X' should be a LoDTensor and 'Y' is a common Tensor, - boxes in 'Y' are shared by all instance of the batched inputs of X. - Given two boxes A and B, the calculation of IOU is as follows: +Box list 'X' should be a LoDTensor and 'Y' is a common Tensor, +boxes in 'Y' are shared by all instance of the batched inputs of X. +Given two boxes A and B, the calculation of IOU is as follows: $$ IOU(A, B) = -\frac{area(A\cap B)}{area(A)+area(B)-area(A\cap B)} +\\frac{area(A\\cap B)}{area(A)+area(B)-area(A\\cap B)} $$ )DOC"); diff --git a/paddle/fluid/operators/detection/mine_hard_examples_op.cc b/paddle/fluid/operators/detection/mine_hard_examples_op.cc index d4a09bae3a98e4518f9885c1e9182f7033a0d262..54a4b87ec8f13c4d474aad4cc0b8159cd5f59d1c 100644 --- a/paddle/fluid/operators/detection/mine_hard_examples_op.cc +++ b/paddle/fluid/operators/detection/mine_hard_examples_op.cc @@ -227,6 +227,9 @@ class MineHardExamplesOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_GT( neg_pos_ratio, 0.0f, "neg_pos_ratio must greater than zero in max_negative mode"); + PADDLE_ENFORCE_LT( + neg_dist_threshold, 1.0f, + "neg_dist_threshold must less than one in max_negative mode"); PADDLE_ENFORCE_GT( neg_dist_threshold, 0.0f, "neg_dist_threshold must greater than zero in max_negative mode"); diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cc b/paddle/fluid/operators/detection/polygon_box_transform_op.cc index 335e8dd470f851d8c5f6bdbc94cfc343da269034..568d50d457d838d5f11605710c0d3b987af01d10 100644 --- a/paddle/fluid/operators/detection/polygon_box_transform_op.cc +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cc @@ -83,11 +83,13 @@ class PolygonBoxTransformOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( PolygonBoxTransform Operator. + +PolygonBoxTransform Operator is used to transform the coordinate shift to the real coordinate. + The input is the final geometry output in detection network. We use 2*n numbers to denote the coordinate shift from n corner vertices of the polygon_box to the pixel location. As each distance offset contains two numbers (xi, yi), the geometry output contains 2*n channels. -PolygonBoxTransform Operator is used to transform the coordinate shift to the real coordinate. )DOC"); } }; diff --git a/paddle/fluid/operators/detection/prior_box_op.cc b/paddle/fluid/operators/detection/prior_box_op.cc index 4e35c38e4e03d4d0f00601812fdc4803519b89ae..b5cb6a724c095eb849f3a184f13843e1a0cca92f 100644 --- a/paddle/fluid/operators/detection/prior_box_op.cc +++ b/paddle/fluid/operators/detection/prior_box_op.cc @@ -149,6 +149,13 @@ class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker { "(float) " "Prior boxes center offset.") .SetDefault(0.5); + AddAttr( + "min_max_aspect_ratios_order", + "(bool) If set True, the output prior box is in order of" + "[min, max, aspect_ratios], which is consistent with Caffe." + "Please note, this order affects the weights order of convolution layer" + "followed by and does not affect the final detection results.") + .SetDefault(false); AddComment(R"DOC( Prior box operator Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm. diff --git a/paddle/fluid/operators/detection/prior_box_op.cu b/paddle/fluid/operators/detection/prior_box_op.cu index f67e6ca91c0852b5a3be35d23246884d1157caa4..1ea8cfc1d2af8cc6c332768a467cdcd4c0166319 100644 --- a/paddle/fluid/operators/detection/prior_box_op.cu +++ b/paddle/fluid/operators/detection/prior_box_op.cu @@ -28,8 +28,8 @@ __global__ void GenPriorBox(T* out, const T* aspect_ratios, const int height, const int im_width, const int as_num, const T offset, const T step_width, const T step_height, const T* min_sizes, - const T* max_sizes, const int min_num, - bool is_clip) { + const T* max_sizes, const int min_num, bool is_clip, + bool min_max_aspect_ratios_order) { int num_priors = max_sizes ? as_num * min_num + min_num : as_num * min_num; int box_num = height * width * num_priors; for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < box_num; @@ -44,14 +44,28 @@ __global__ void GenPriorBox(T* out, const T* aspect_ratios, const int height, T min_size = min_sizes[m]; if (max_sizes) { int s = p % (as_num + 1); - if (s < as_num) { - T ar = aspect_ratios[s]; - bw = min_size * sqrt(ar) / 2.; - bh = min_size / sqrt(ar) / 2.; + if (!min_max_aspect_ratios_order) { + if (s < as_num) { + T ar = aspect_ratios[s]; + bw = min_size * sqrt(ar) / 2.; + bh = min_size / sqrt(ar) / 2.; + } else { + T max_size = max_sizes[m]; + bw = sqrt(min_size * max_size) / 2.; + bh = bw; + } } else { - T max_size = max_sizes[m]; - bw = sqrt(min_size * max_size) / 2.; - bh = bw; + if (s == 0) { + bw = bh = min_size / 2.; + } else if (s == 1) { + T max_size = max_sizes[m]; + bw = sqrt(min_size * max_size) / 2.; + bh = bw; + } else { + T ar = aspect_ratios[s - 1]; + bw = min_size * sqrt(ar) / 2.; + bh = min_size / sqrt(ar) / 2.; + } } } else { int s = p % as_num; @@ -94,6 +108,8 @@ class PriorBoxOpCUDAKernel : public framework::OpKernel { auto variances = ctx.Attr>("variances"); auto flip = ctx.Attr("flip"); auto clip = ctx.Attr("clip"); + auto min_max_aspect_ratios_order = + ctx.Attr("min_max_aspect_ratios_order"); std::vector aspect_ratios; ExpandAspectRatios(input_aspect_ratio, flip, &aspect_ratios); @@ -149,7 +165,7 @@ class PriorBoxOpCUDAKernel : public framework::OpKernel { GenPriorBox<<>>( boxes->data(), r.data(), height, width, im_height, im_width, aspect_ratios.size(), offset, step_width, step_height, min.data(), - max_data, min_num, clip); + max_data, min_num, clip, min_max_aspect_ratios_order); framework::Tensor v; framework::TensorFromVector(variances, ctx.device_context(), &v); diff --git a/paddle/fluid/operators/detection/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h index 1c62fd8d2c4d4e4deba4ca6442efbaff83e36c35..4e226abbb51c271502f0ca5419d488643b5a1a82 100644 --- a/paddle/fluid/operators/detection/prior_box_op.h +++ b/paddle/fluid/operators/detection/prior_box_op.h @@ -68,6 +68,8 @@ class PriorBoxOpKernel : public framework::OpKernel { auto variances = ctx.Attr>("variances"); auto flip = ctx.Attr("flip"); auto clip = ctx.Attr("clip"); + auto min_max_aspect_ratios_order = + ctx.Attr("min_max_aspect_ratios_order"); std::vector aspect_ratios; ExpandAspectRatios(input_aspect_ratio, flip, &aspect_ratios); @@ -108,26 +110,59 @@ class PriorBoxOpKernel : public framework::OpKernel { int idx = 0; for (size_t s = 0; s < min_sizes.size(); ++s) { auto min_size = min_sizes[s]; - // priors with different aspect ratios - for (size_t r = 0; r < aspect_ratios.size(); ++r) { - float ar = aspect_ratios[r]; - box_width = min_size * sqrt(ar) / 2.; - box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; - } - if (max_sizes.size() > 0) { - auto max_size = max_sizes[s]; - // square prior with size sqrt(minSize * maxSize) - box_width = box_height = sqrt(min_size * max_size) / 2.; + if (min_max_aspect_ratios_order) { + box_width = box_height = min_size / 2.; e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; idx++; + if (max_sizes.size() > 0) { + auto max_size = max_sizes[s]; + // square prior with size sqrt(minSize * maxSize) + box_width = box_height = sqrt(min_size * max_size) / 2.; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; + idx++; + } + // priors with different aspect ratios + for (size_t r = 0; r < aspect_ratios.size(); ++r) { + float ar = aspect_ratios[r]; + if (fabs(ar - 1.) < 1e-6) { + continue; + } + box_width = min_size * sqrt(ar) / 2.; + box_height = min_size / sqrt(ar) / 2.; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; + idx++; + } + } else { + // priors with different aspect ratios + for (size_t r = 0; r < aspect_ratios.size(); ++r) { + float ar = aspect_ratios[r]; + box_width = min_size * sqrt(ar) / 2.; + box_height = min_size / sqrt(ar) / 2.; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; + idx++; + } + if (max_sizes.size() > 0) { + auto max_size = max_sizes[s]; + // square prior with size sqrt(minSize * maxSize) + box_width = box_height = sqrt(min_size * max_size) / 2.; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; + idx++; + } } } } diff --git a/paddle/fluid/operators/detection/rpn_target_assign_op.cc b/paddle/fluid/operators/detection/rpn_target_assign_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..9a1643d5b35c067ba9064286bab32019fb34fbe8 --- /dev/null +++ b/paddle/fluid/operators/detection/rpn_target_assign_op.cc @@ -0,0 +1,283 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; +template +using EigenMatrix = framework::EigenMatrix; + +class RpnTargetAssignOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("DistMat"), + "Input(DistMat) of RpnTargetAssignOp should not be null"); + + PADDLE_ENFORCE( + ctx->HasOutput("LocationIndex"), + "Output(LocationIndex) of RpnTargetAssignOp should not be null"); + PADDLE_ENFORCE( + ctx->HasOutput("ScoreIndex"), + "Output(ScoreIndex) of RpnTargetAssignOp should not be null"); + PADDLE_ENFORCE( + ctx->HasOutput("TargetLabel"), + "Output(TargetLabel) of RpnTargetAssignOp should not be null"); + + auto in_dims = ctx->GetInputDim("DistMat"); + PADDLE_ENFORCE_EQ(in_dims.size(), 2, + "The rank of Input(DistMat) must be 2."); + } +}; + +template +class RpnTargetAssignKernel : public framework::OpKernel { + public: + void ScoreAssign(const T* dist_data, const Tensor& anchor_to_gt_max, + const int row, const int col, const float pos_threshold, + const float neg_threshold, int64_t* target_label_data, + std::vector* fg_inds, std::vector* bg_inds) const { + int fg_offset = fg_inds->size(); + int bg_offset = bg_inds->size(); + for (int64_t i = 0; i < row; ++i) { + const T* v = dist_data + i * col; + T max_dist = *std::max_element(v, v + col); + for (int64_t j = 0; j < col; ++j) { + T val = dist_data[i * col + j]; + if (val == max_dist) target_label_data[j] = 1; + } + } + + // Pick the fg/bg and count the number + for (int64_t j = 0; j < col; ++j) { + if (anchor_to_gt_max.data()[j] > pos_threshold) { + target_label_data[j] = 1; + } else if (anchor_to_gt_max.data()[j] < neg_threshold) { + target_label_data[j] = 0; + } + if (target_label_data[j] == 1) { + fg_inds->push_back(fg_offset + j); + } else if (target_label_data[j] == 0) { + bg_inds->push_back(bg_offset + j); + } + } + } + + void ReservoirSampling(const int num, const int offset, + std::minstd_rand engine, + std::vector* inds) const { + std::uniform_real_distribution uniform(0, 1); + const int64_t size = static_cast(inds->size()); + if (size > num) { + for (int64_t i = num; i < size; ++i) { + int rng_ind = std::floor(uniform(engine) * i); + if (rng_ind < num) + std::iter_swap(inds->begin() + rng_ind + offset, + inds->begin() + i + offset); + } + } + } + + void RpnTargetAssign(const framework::ExecutionContext& ctx, + const Tensor& dist, const float pos_threshold, + const float neg_threshold, const int rpn_batch_size, + const int fg_num, std::minstd_rand engine, + std::vector* fg_inds, std::vector* bg_inds, + int64_t* target_label_data) const { + auto* dist_data = dist.data(); + int64_t row = dist.dims()[0]; + int64_t col = dist.dims()[1]; + int fg_offset = fg_inds->size(); + int bg_offset = bg_inds->size(); + + // Calculate the max IoU between anchors and gt boxes + Tensor anchor_to_gt_max; + anchor_to_gt_max.mutable_data( + framework::make_ddim({static_cast(col), 1}), + platform::CPUPlace()); + auto& place = *ctx.template device_context() + .eigen_device(); + auto x = EigenMatrix::From(dist); + auto x_col_max = EigenMatrix::From(anchor_to_gt_max); + x_col_max.device(place) = + x.maximum(Eigen::DSizes(0)) + .reshape(Eigen::DSizes(static_cast(col), 1)); + // Follow the Faster RCNN's implementation + ScoreAssign(dist_data, anchor_to_gt_max, row, col, pos_threshold, + neg_threshold, target_label_data, fg_inds, bg_inds); + // Reservoir Sampling + ReservoirSampling(fg_num, fg_offset, engine, fg_inds); + int bg_num = rpn_batch_size - fg_inds->size(); + ReservoirSampling(bg_num, bg_offset, engine, bg_inds); + } + + void Compute(const framework::ExecutionContext& context) const override { + auto* dist = context.Input("DistMat"); + auto* loc_index = context.Output("LocationIndex"); + auto* score_index = context.Output("ScoreIndex"); + auto* tgt_lbl = context.Output("TargetLabel"); + + auto col = dist->dims()[1]; + int64_t n = dist->lod().size() == 0UL + ? 1 + : static_cast(dist->lod().back().size() - 1); + if (dist->lod().size()) { + PADDLE_ENFORCE_EQ(dist->lod().size(), 1UL, + "Only support 1 level of LoD."); + } + int rpn_batch_size = context.Attr("rpn_batch_size_per_im"); + float pos_threshold = context.Attr("rpn_positive_overlap"); + float neg_threshold = context.Attr("rpn_negative_overlap"); + float fg_fraction = context.Attr("fg_fraction"); + + int fg_num = static_cast(rpn_batch_size * fg_fraction); + + int64_t* target_label_data = + tgt_lbl->mutable_data({n * col, 1}, context.GetPlace()); + + auto& dev_ctx = context.device_context(); + math::SetConstant iset; + iset(dev_ctx, tgt_lbl, static_cast(-1)); + + std::vector fg_inds; + std::vector bg_inds; + std::random_device rnd; + std::minstd_rand engine; + int seed = + context.Attr("fix_seed") ? context.Attr("seed") : rnd(); + engine.seed(seed); + + if (n == 1) { + RpnTargetAssign(context, *dist, pos_threshold, neg_threshold, + rpn_batch_size, fg_num, engine, &fg_inds, &bg_inds, + target_label_data); + } else { + auto lod = dist->lod().back(); + for (size_t i = 0; i < lod.size() - 1; ++i) { + Tensor one_ins = dist->Slice(lod[i], lod[i + 1]); + RpnTargetAssign(context, one_ins, pos_threshold, neg_threshold, + rpn_batch_size, fg_num, engine, &fg_inds, &bg_inds, + target_label_data + i * col); + } + } + int* loc_index_data = loc_index->mutable_data( + {static_cast(fg_inds.size())}, context.GetPlace()); + int* score_index_data = score_index->mutable_data( + {static_cast(fg_inds.size() + bg_inds.size())}, + context.GetPlace()); + memcpy(loc_index_data, reinterpret_cast(&fg_inds[0]), + fg_inds.size() * sizeof(int)); + memcpy(score_index_data, reinterpret_cast(&fg_inds[0]), + fg_inds.size() * sizeof(int)); + memcpy(score_index_data + fg_inds.size(), + reinterpret_cast(&bg_inds[0]), bg_inds.size() * sizeof(int)); + } +}; + +class RpnTargetAssignOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput( + "DistMat", + "(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape " + "[K, M]. It is pair-wise distance matrix between the entities " + "represented by each row and each column. For example, assumed one " + "entity is A with shape [K], another entity is B with shape [M]. The " + "DistMat[i][j] is the distance between A[i] and B[j]. The bigger " + "the distance is, the better macthing the pairs are. Please note, " + "This tensor can contain LoD information to represent a batch of " + "inputs. One instance of this batch can contain different numbers of " + "entities."); + AddAttr( + "rpn_positive_overlap", + "Minimum overlap required between an anchor and ground-truth " + "box for the (anchor, gt box) pair to be a positive example.") + .SetDefault(0.7); + AddAttr( + "rpn_negative_overlap", + "Maximum overlap allowed between an anchor and ground-truth " + "box for the (anchor, gt box) pair to be a negative examples.") + .SetDefault(0.3); + AddAttr( + "fg_fraction", + "Target fraction of RoI minibatch that " + "is labeled foreground (i.e. class > 0), 0-th class is background.") + .SetDefault(0.25); + AddAttr("rpn_batch_size_per_im", + "Total number of RPN examples per image.") + .SetDefault(256); + AddAttr("fix_seed", + "A flag indicating whether to use a fixed seed to generate " + "random mask. NOTE: DO NOT set this flag to true in " + "training. Setting this flag to true is only useful in " + "unittest.") + .SetDefault(false); + AddAttr("seed", "RpnTargetAssign random seed.").SetDefault(0); + AddOutput( + "LocationIndex", + "(Tensor), The indexes of foreground anchors in all RPN anchors, the " + "shape of the LocationIndex is [F], F depends on the value of input " + "tensor and attributes."); + AddOutput( + "ScoreIndex", + "(Tensor), The indexes of foreground and background anchors in all " + "RPN anchors(The rest anchors are ignored). The shape of the " + "ScoreIndex is [F + B], F and B depend on the value of input " + "tensor and attributes."); + AddOutput("TargetLabel", + "(Tensor), The target labels of each anchor with shape " + "[K * M, 1], " + "K and M is the same as they are in DistMat."); + AddComment(R"DOC( +This operator can be, for given the IoU between the ground truth bboxes and the +anchors, to assign classification and regression targets to each prediction. +The Score index and LocationIndex will be generated according to the DistMat. +The rest anchors would not contibute to the RPN training loss + +ScoreIndex is composed of foreground anchor indexes(positive labels) and +background anchor indexes(negative labels). LocationIndex is exactly same +as the foreground anchor indexes since we can not assign regression target to +the background anchors. + +The classification targets(TargetLabel) is a binary class label (of being +an object or not). Following the paper of Faster-RCNN, the positive labels +are two kinds of anchors: (i) the anchor/anchors with the highest IoU +overlap with a ground-truth box, or (ii) an anchor that has an IoU overlap +higher than rpn_positive_overlap(0.7) with any ground-truth box. Note that +a single ground-truth box may assign positive labels to multiple anchors. +A non-positive anchor is when its IoU ratio is lower than rpn_negative_overlap +(0.3) for all ground-truth boxes. Anchors that are neither positive nor +negative do not contribute to the training objective. + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(rpn_target_assign, ops::RpnTargetAssignOp, + ops::RpnTargetAssignOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(rpn_target_assign, ops::RpnTargetAssignKernel, + ops::RpnTargetAssignKernel); diff --git a/paddle/fluid/operators/detection/target_assign_op.h b/paddle/fluid/operators/detection/target_assign_op.h index 3d529737414d54f05e8c82ede1d6068e5d261110..7f989dfca699d498432f8df3f86c44723faeb980 100644 --- a/paddle/fluid/operators/detection/target_assign_op.h +++ b/paddle/fluid/operators/detection/target_assign_op.h @@ -106,7 +106,11 @@ class TargetAssignKernel : public framework::OpKernel { int64_t k = x->dims()[2]; auto x_lod = x->lod().back(); +#if defined(PADDLE_WITH_CUDA) size_t* x_lod_data = x_lod.MutableData(ctx.GetPlace()); +#else + size_t* x_lod_data = x_lod.data(); +#endif TargetAssignFunctor functor(x_data, match_idx_data, x_lod_data, mismatch_value, n, m, p, k, out_data, @@ -121,7 +125,11 @@ class TargetAssignKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(neg_indices->lod().size(), 1UL); const int* neg_idx_data = neg_indices->data(); auto neg_lod = neg_indices->lod().back(); +#if defined(PADDLE_WITH_CUDA) size_t* neg_lod_data = neg_lod.MutableData(ctx.GetPlace()); +#else + size_t* neg_lod_data = neg_lod.data(); +#endif NegTargetAssignFunctor neg_trg_functor; neg_trg_functor(device_ctx, neg_idx_data, neg_lod_data, n, m, k, mismatch_value, out_data, out_wt_data); diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index 716c8625d35308f98582e6802e90d99d643e188b..d7f49a9590e4ef4ca4d2ad5a92572c70e6bfb6ac 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -175,12 +175,12 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Detection mAP evaluate operator. The general steps are as follows. First, calculate the true positive and - false positive according to the input of detection and labels, then - calculate the mAP evaluate value. - Supporting '11 point' and 'integral' mAP algorithm. Please get more information - from the following articles: - https://sanchom.wordpress.com/tag/average-precision/ - https://arxiv.org/abs/1512.02325 +false positive according to the input of detection and labels, then +calculate the mAP evaluate value. +Supporting '11 point' and 'integral' mAP algorithm. Please get more information +from the following articles: +https://sanchom.wordpress.com/tag/average-precision/ +https://arxiv.org/abs/1512.02325 )DOC"); } diff --git a/paddle/fluid/operators/distributed/CMakeLists.txt b/paddle/fluid/operators/distributed/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..da5d20505e9b06c0717af8d79d5456a9ade1e89c --- /dev/null +++ b/paddle/fluid/operators/distributed/CMakeLists.txt @@ -0,0 +1,43 @@ +if(NOT WITH_DISTRIBUTE) + return() +endif() + +if(WITH_GRPC) + set(cc_generic_services "false") +else() + set(cc_generic_services "true") +endif() +configure_file(send_recv.proto.in ${CMAKE_CURRENT_SOURCE_DIR}/send_recv.proto @ONLY) + +if(WITH_GRPC) + grpc_library(sendrecvop_grpc SRCS grpc_bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc + request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc grpc_variable_response.cc grpc_serde.cc + PROTO send_recv.proto + DEPS lod_tensor selected_rows memory) + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(grpc_serde_test SRCS grpc_serde_test.cc + DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL) + cc_test(rpc_server_test SRCS rpc_server_test.cc + DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_sparse_table_op SERIAL) + return() +endif() + + +set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + +set_source_files_properties(brpc_server.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc + brpc_variable_response.cc brpc_sendrecvop_utils.cc brpc_rdma_pool.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + +brpc_library(sendrecvop_brpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc rpc_client.cc request_handler_impl.cc brpc_sendrecvop_utils.cc + brpc_variable_response.cc variable_response.cc sendrecvop_utils.cc brpc_rdma_pool.cc + PROTO send_recv.proto + DEPS lod_tensor selected_rows memory) + +set(brpc_test_depends sendrecvop_brpc brpc ssl crypto protobuf leveldb gflags glog executor proto_desc lookup_table_op snappystream snappy) + +cc_test(brpc_server_test SRCS rpc_server_test.cc + DEPS ${brpc_test_depends} SERIAL) + +cc_test(brpc_serde_test SRCS brpc_serde_test.cc + DEPS ${brpc_test_depends} SERIAL) diff --git a/paddle/fluid/operators/distributed/brpc_client.cc b/paddle/fluid/operators/distributed/brpc_client.cc new file mode 100644 index 0000000000000000000000000000000000000000..b394c678fb6503eb73a1e11e6feb814251e9e940 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_client.cc @@ -0,0 +1,180 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/distributed/brpc_client.h" +#include "paddle/fluid/framework/threadpool.h" + +namespace paddle { +namespace operators { +namespace distributed { + +DEFINE_int32(brpc_channel_num, 24, + "Number of channels to send requests connected to one server"); +DEFINE_int32(timeout_ms, 30000, "RPC timeout in milliseconds"); +DEFINE_int32(max_retry, 3, "Max retries(not including the first RPC)"); + +BRPCClient::~BRPCClient() { Wait(); } + +void HandleSendResponse(brpc::Controller* cntl, + sendrecv::VoidMessage* response) { + // std::unique_ptr makes sure cntl/response will be deleted before returning. + std::unique_ptr cntl_guard(cntl); + std::unique_ptr response_guard(response); + + if (cntl->Failed()) { + LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText(); + return; + } + LOG(INFO) << "Received response from " << cntl->remote_side() + << " latency=" << cntl->latency_us() << "us"; +} + +bool BRPCClient::AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string var_name_val = var_name; + const framework::Scope* p_scope = &scope; + const auto ch_ptr = GetChannel(ep_val); + + framework::AsyncIO( + [var_name_val, p_ctx, ep_val, p_scope, time_out, ch_ptr, this] { + auto ch_ctx = ch_ptr->Pop(); + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); + cntl->set_timeout_ms(time_out); + + google::protobuf::Closure* done = + brpc::NewCallback(&HandleSendResponse, cntl, response); + + sendrecv::VariableMessage request; + ch_ctx->stub->SendVariable(cntl, &request, response, done); + }); + req_count_++; + + return true; +} + +void HandleGetResponse(brpc::Controller* cntl, + sendrecv::VariableMessage* response) { + // std::unique_ptr makes sure cntl/response will be deleted before returning. + std::unique_ptr cntl_guard(cntl); + std::unique_ptr response_guard(response); + + if (cntl->Failed()) { + LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText(); + return; + } + LOG(INFO) << "Received response from " << cntl->remote_side() + << " latency=" << cntl->latency_us() << "us"; + + // framework::Variable* outvar = nullptr; + // DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar); +} + +bool BRPCClient::AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string var_name_val = var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO( + [var_name_val, ep_val, p_scope, p_ctx, time_out, ch, this] {}); + + req_count_++; + + return true; +} + +bool BRPCClient::AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string in_var_name_val = in_var_name; + const std::string out_var_name_val = out_var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, + time_out, ch, this] {}); + + req_count_++; + return true; +} + +void BRPCClient::AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out) { + req_count_++; +} + +void BRPCClient::AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out) { + req_count_++; +} + +void BRPCClient::Wait() { + std::unique_lock lk(sync_mutex_); + sync_cond_.wait(lk, [this] { return req_count_ == 0; }); +} + +ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) { + { + std::lock_guard guard(chan_mutex_); + auto it = channels_.find(ep); + if (it != channels_.end()) { + return it->second; + } + } + + ChannelQueuePtr q(new framework::BlockingQueue()); + + brpc::ChannelOptions options; + options.protocol = "baidu_std"; + options.connection_type = "pooled"; + options.connect_timeout_ms = 100; + options.timeout_ms = FLAGS_timeout_ms /*milliseconds*/; + options.max_retry = FLAGS_max_retry; + for (int i = 0; i < FLAGS_brpc_channel_num; ++i) { + std::shared_ptr c(new ChannelContext()); + if (c->channel.Init(ep.c_str(), &options) != 0) { + LOG(ERROR) << "Fail to initialize channel"; + return nullptr; + } + + c->stub.reset(new sendrecv::SendRecvService_Stub( + static_cast(&c->channel))); + q->Push(c); + } + + { + std::lock_guard guard(chan_mutex_); + channels_[ep] = q; + } + + return q; +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_client.h b/paddle/fluid/operators/distributed/brpc_client.h new file mode 100644 index 0000000000000000000000000000000000000000..8ff1f0a6076b3574c42065edcbac50eb75b3b483 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_client.h @@ -0,0 +1,98 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include // NOLINT +#include +#include +#include +#include +#include // NOLINT +#include +#include + +#include "brpc/channel.h" +#include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN + +namespace paddle { +namespace operators { +namespace distributed { + +struct ChannelContext { + brpc::Channel channel; + std::shared_ptr stub; +}; + +typedef std::shared_ptr ChannelContextPtr; +typedef std::shared_ptr> + ChannelQueuePtr; + +class BRPCClient : public RPCClient { + public: + BRPCClient() {} + virtual ~BRPCClient(); + + bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + + bool AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + + bool AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + + void AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) override; + + void AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) override; + + void Wait() override; + + private: + void Proceed(); + ChannelQueuePtr GetChannel(const std::string& ep); + + private: + std::unordered_map channels_; + + // mutex for Wait client sync + std::mutex sync_mutex_; + std::condition_variable sync_cond_; + std::atomic req_count_{0}; + + // mutex for GetChannel thread safety + std::mutex chan_mutex_; + DISABLE_COPY_AND_ASSIGN(BRPCClient); +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_server.cc b/paddle/fluid/operators/distributed/brpc_server.cc new file mode 100644 index 0000000000000000000000000000000000000000..862167f02084cfe81db1c0936bbfb0415fa85721 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_server.cc @@ -0,0 +1,144 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/distributed/brpc_server.h" +#include "paddle/fluid/operators/distributed/request_handler.h" + +namespace sendrecv { + +typedef std::unordered_map + HandlerMap; + +class BRPCServiceImpl : public SendRecvService { + public: + explicit BRPCServiceImpl(const HandlerMap& rpc_call_map) + : request_send_h_(nullptr), + request_get_h_(nullptr), + request_prefetch_h_(nullptr) { + auto it = rpc_call_map.find(paddle::operators::distributed::kRequestSend); + if (it != rpc_call_map.end()) { + request_send_h_ = it->second; + } + + it = rpc_call_map.find(paddle::operators::distributed::kRequestSend); + if (it != rpc_call_map.end()) { + request_get_h_ = it->second; + } + + it = rpc_call_map.find(paddle::operators::distributed::kRequestPrefetch); + if (it != rpc_call_map.end()) { + request_prefetch_h_ = it->second; + } + } + + virtual ~BRPCServiceImpl() {} + + void SendVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VoidMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE(request_send_h_ != nullptr, + "RequestSend handler should be registed first!"); + brpc::ClosureGuard done_guard(done); + + paddle::framework::Scope* local_scope = request_send_h_->scope(); + paddle::framework::Variable* outvar = nullptr; + paddle::framework::Variable* invar = nullptr; + + std::string varname = request->varname(); + + if (!request_send_h_->sync_mode()) { + local_scope = &request_send_h_->scope()->NewScope(); + invar = local_scope->Var(varname); + } else { + invar = local_scope->FindVar(varname); + } + + request_send_h_->Handle(varname, local_scope, invar, &outvar); + + if (!request_send_h_->sync_mode()) { + request_send_h_->scope()->DeleteScope(local_scope); + } + } + + void GetVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VariableMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE(request_get_h_ != nullptr, + "RequestGet handler should be registed first!"); + } + + void PrefetchVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, + VariableMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE(request_prefetch_h_ != nullptr, + "kRequestPrefetch handler should be registed first!"); + } + + private: + paddle::operators::distributed::RequestHandler* request_send_h_; + paddle::operators::distributed::RequestHandler* request_get_h_; + paddle::operators::distributed::RequestHandler* request_prefetch_h_; +}; +} // namespace sendrecv + +namespace paddle { +namespace operators { +namespace distributed { + +void AsyncBRPCServer::StartServer() { + // Instance of your service. + sendrecv::BRPCServiceImpl service_impl(rpc_call_map_); + + // Add the service into server. Notice the second parameter, because the + // service is put on stack, we don't want server to delete it, otherwise + // use brpc::SERVER_OWNS_SERVICE. + if (server_.AddService(&service_impl, brpc::SERVER_DOESNT_OWN_SERVICE) != 0) { + LOG(FATAL) << "Fail to add service"; + return; + } + + brpc::ServerOptions options; + options.idle_timeout_sec = idle_timeout_s_; + options.max_concurrency = max_concurrency_; + if (server_.Start(bind_address_.c_str(), &options) != 0) { + LOG(FATAL) << "Fail to start EchoServer" << bind_address_; + return; + } + + butil::EndPoint ep = server_.listen_address(); + selected_port_ = ep.port; + + { + std::lock_guard lock(this->mutex_ready_); + ready_ = 1; + } + condition_ready_.notify_all(); + + server_.Join(); +} + +void AsyncBRPCServer::ShutDownImpl() { server_.Stop(1000); } + +void AsyncBRPCServer::WaitServerReady() { + VLOG(3) << "AsyncGRPCServer is wait server ready"; + std::unique_lock lock(this->mutex_ready_); + condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); + VLOG(3) << "AsyncGRPCServer WaitSeverReady"; +} + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_server.h b/paddle/fluid/operators/distributed/brpc_server.h new file mode 100644 index 0000000000000000000000000000000000000000..85a7ad0dfe843dad483d43631b69a79d75211ce9 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_server.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include // NOLINT +#include // NOLINT +#include + +#include "brpc/server.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class AsyncBRPCServer final : public RPCServer { + public: + explicit AsyncBRPCServer(const std::string& address, int client_num) + : RPCServer(address, client_num), ready_(0) {} + + virtual ~AsyncBRPCServer() {} + void StartServer() override; + void WaitServerReady() override; + + private: + void ShutDownImpl() override; + + brpc::Server server_; + + static constexpr int idle_timeout_s_ = -1; + static constexpr int max_concurrency_ = 0; + + std::mutex mutex_ready_; + std::condition_variable condition_ready_; + int ready_; +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_bytebuffer_stream.cc b/paddle/fluid/operators/distributed/grpc_bytebuffer_stream.cc new file mode 100644 index 0000000000000000000000000000000000000000..d192f54ee0c924b772045d9b6a01701f640e07c7 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_bytebuffer_stream.cc @@ -0,0 +1,88 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +// NOTE: This file was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// file and did some modifications so that we can send gRPC +// requests without too much copying of the tensor data. + +#include "paddle/fluid/operators/distributed/grpc_bytebuffer_stream.h" + +namespace paddle { +namespace operators { +namespace distributed { + +GrpcByteBufferSource::GrpcByteBufferSource() {} + +bool GrpcByteBufferSource::Init(const grpc::ByteBuffer& src) { + cur_ = -1; + left_ = 0; + ptr_ = nullptr; + byte_count_ = 0; + bool ok = src.Dump(&slices_).ok(); + if (!ok) { + slices_.clear(); + } + return ok; +} + +bool GrpcByteBufferSource::Next(const void** data, int* size) { + // Use loop instead of if in case buffer contained empty slices. + while (left_ == 0) { + // Advance to next slice. + cur_++; + if (cur_ >= slices_.size()) { + return false; + } + const ::grpc::Slice& s = slices_[cur_]; + left_ = s.size(); + ptr_ = reinterpret_cast(s.begin()); + } + + *data = ptr_; + *size = left_; + byte_count_ += left_; + ptr_ += left_; + left_ = 0; + return true; +} + +void GrpcByteBufferSource::BackUp(int count) { + ptr_ -= count; + left_ += count; + byte_count_ -= count; +} + +bool GrpcByteBufferSource::Skip(int count) { + const void* data; + int size; + while (Next(&data, &size)) { + if (size >= count) { + BackUp(size - count); + return true; + } + // size < count; + count -= size; + } + // error or we have too large count; + return false; +} + +google::protobuf::int64 GrpcByteBufferSource::ByteCount() const { + return byte_count_; +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_bytebuffer_stream.h b/paddle/fluid/operators/distributed/grpc_bytebuffer_stream.h new file mode 100644 index 0000000000000000000000000000000000000000..e9074574cdd163bbf7e62939df9283352706f840 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_bytebuffer_stream.h @@ -0,0 +1,170 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +// NOTE: This file was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// file and did some modifications so that we can send gRPC +// requests without too much copying of the tensor data. + +#pragma once + +#include + +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "grpc++/grpc++.h" +#include "paddle/fluid/operators/distributed/variable_response.h" + +namespace grpc { +// A ZeroCopyInputStream that reads from grpc_byte_buffer +class GrpcBufferReader final + : public ::google::protobuf::io::ZeroCopyInputStream { + typedef void (CoreCodegenInterface::*OldReaderInitAPI)( + grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer); + typedef int (CoreCodegenInterface::*NewReaderInitAPI)( + grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer); + void ReaderInit(OldReaderInitAPI ptr, grpc_byte_buffer_reader* reader, + grpc_byte_buffer* buffer) { + (g_core_codegen_interface->*ptr)(reader, buffer); + } + void ReaderInit(NewReaderInitAPI ptr, grpc_byte_buffer_reader* reader, + grpc_byte_buffer* buffer) { + int result = (g_core_codegen_interface->*ptr)(reader, buffer); + (void)result; + } + + public: + explicit GrpcBufferReader(grpc_byte_buffer* buffer) + : byte_count_(0), backup_count_(0) { + ReaderInit(&CoreCodegenInterface::grpc_byte_buffer_reader_init, &reader_, + buffer); + } + ~GrpcBufferReader() override { + g_core_codegen_interface->grpc_byte_buffer_reader_destroy(&reader_); + } + + bool Next(const void** data, int* size) override { + if (backup_count_ > 0) { + *data = GRPC_SLICE_START_PTR(slice_) + GRPC_SLICE_LENGTH(slice_) - + backup_count_; + GPR_CODEGEN_ASSERT(backup_count_ <= INT_MAX); + *size = static_cast(backup_count_); + backup_count_ = 0; + return true; + } + if (!g_core_codegen_interface->grpc_byte_buffer_reader_next(&reader_, + &slice_)) { + return false; + } + g_core_codegen_interface->grpc_slice_unref(slice_); + *data = GRPC_SLICE_START_PTR(slice_); + // On win x64, int is only 32bit + GPR_CODEGEN_ASSERT(GRPC_SLICE_LENGTH(slice_) <= INT_MAX); + byte_count_ += * size = static_cast(GRPC_SLICE_LENGTH(slice_)); + return true; + } + + void BackUp(int count) override { backup_count_ = count; } + + bool Skip(int count) override { + const void* data; + int size; + while (Next(&data, &size)) { + if (size >= count) { + BackUp(size - count); + return true; + } + // size < count; + count -= size; + } + // error or we have too large count; + return false; + } + + ::google::protobuf::int64 ByteCount() const override { + return byte_count_ - backup_count_; + } + + private: + int64_t byte_count_; + int64_t backup_count_; + grpc_byte_buffer_reader reader_; + grpc_slice slice_; +}; + +}; // namespace grpc + +namespace paddle { +namespace operators { +namespace distributed { + +// A ZeroCopyInputStream that reads from a grpc::ByteBuffer. +class GrpcByteBufferSource + : public ::google::protobuf::io::ZeroCopyInputStream { + public: + GrpcByteBufferSource(); + bool Init(const ::grpc::ByteBuffer& src); // Can be called multiple times. + bool Next(const void** data, int* size) override; + void BackUp(int count) override; + bool Skip(int count) override; + ::google::protobuf::int64 ByteCount() const override; + + private: + std::vector<::grpc::Slice> slices_; + size_t cur_; // Current slice index. + int left_; // Number of bytes in slices_[cur_] left to yield. + const char* ptr_; // Address of next byte in slices_[cur_] to yield. + ::google::protobuf::int64 byte_count_; +}; + +class GrpcByteBufferSourceWrapper : public Source { + public: + explicit GrpcByteBufferSourceWrapper(GrpcByteBufferSource* source) + : source_(source) {} + ::google::protobuf::io::ZeroCopyInputStream* contents() override { + return source_; + } + + private: + GrpcByteBufferSource* source_; +}; + +class GrpcByteSource : public Source { + public: + explicit GrpcByteSource(grpc_byte_buffer* buffer) : buffer_(buffer) {} + ~GrpcByteSource() override { DeleteStream(); } + + typedef ::grpc::GrpcBufferReader Reader; + + ::google::protobuf::io::ZeroCopyInputStream* contents() override { + DeleteStream(); + stream_ = new (&space_) Reader(buffer_); + return stream_; + } + + private: + void DeleteStream() { + if (stream_) { + stream_->~Reader(); + } + } + + grpc_byte_buffer* buffer_; // Not owned + Reader* stream_ = nullptr; // Points into space_ if non-nullptr + char space_[sizeof(Reader)]; +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_client.cc b/paddle/fluid/operators/distributed/grpc_client.cc new file mode 100644 index 0000000000000000000000000000000000000000..b4f60c9ff9a41d5cb7dbe4e7a7694a84bab8e940 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_client.cc @@ -0,0 +1,329 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/distributed/grpc_client.h" + +#include + +#include + +#include "glog/logging.h" // For VLOG +#include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/distributed/grpc_serde.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace operators { +namespace distributed { + +void GRPCClient::InitImpl() { InitEventLoop(); } + +void GRPCClient::InitEventLoop() { + // start the client process thread + // TODO(wuyi): can make this in a threadpool + client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this))); +} + +void GRPCClient::SendComplete() { + std::unique_lock lk(completed_mutex_); + if (!completed_) { + for (auto& it : channels_) { + VLOG(3) << "send complete message to " << it.first; + this->AsyncSendComplete(it.first); + } + PADDLE_ENFORCE(this->Wait(), "internal grpc error"); + completed_ = true; + } +} + +GRPCClient::~GRPCClient() { + stopped_ = true; + Wait(); + cq_.Shutdown(); + { + std::lock_guard guard(chan_mutex_); + for (auto& it : channels_) { + it.second.reset(); + } + channels_.clear(); + } + + client_thread_->join(); +} + +bool GRPCClient::AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string var_name_val = var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO([var_name_val, p_ctx, ep_val, p_scope, time_out, ch, + this] { + auto* var = p_scope->FindVar(var_name_val); + + ::grpc::ByteBuffer req; + SerializeToByteBuffer(var_name_val, var, *p_ctx, &req); + + // varhandle + VarHandle var_h; + var_h.ep = ep_val; + var_h.scope = p_scope; + var_h.name = var_name_val; + var_h.ctx = p_ctx; + var_h.method = "Send"; + + VLOG(3) << var_h.String() << " begin"; + + // stub context + SendProcessor* s = new SendProcessor(ch); + s->Prepare(var_h, time_out); + s->response_call_back_ = nullptr; + + auto call = s->stub_g_.PrepareUnaryCall( + s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); + call->StartCall(); + call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + }); + req_count_++; + + return true; +} + +void ProcGetResponse(const VarHandle& var_h, + const ::grpc::ByteBuffer& ret_msg) { + framework::Variable* outvar = nullptr; + DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar); +} + +template +void RequestToByteBuffer(const T& proto, ::grpc::ByteBuffer* result) { + ::grpc::Slice slice(proto.ByteSizeLong()); + proto.SerializeWithCachedSizesToArray(const_cast(slice.begin())); + ::grpc::ByteBuffer tmp(&slice, 1); + result->Swap(&tmp); +} + +bool GRPCClient::AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string var_name_val = var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO([var_name_val, ep_val, p_scope, p_ctx, time_out, ch, + this] { + // prepare input + sendrecv::VariableMessage req; + req.set_varname(var_name_val); + ::grpc::ByteBuffer buf; + RequestToByteBuffer(req, &buf); + + // var handle + VarHandle var_h; + var_h.ep = ep_val; + var_h.scope = p_scope; + var_h.name = var_name_val; + var_h.ctx = p_ctx; + var_h.method = "Get"; + + VLOG(3) << var_h.String() << " begin"; + + // stub context + GetProcessor* s = new GetProcessor(ch); + s->Prepare(var_h, time_out); + s->response_call_back_ = ProcGetResponse; + + auto call = s->stub_g_.PrepareUnaryCall( + s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); + call->StartCall(); + call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + }); + + req_count_++; + + return true; +} + +bool GRPCClient::AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string in_var_name_val = in_var_name; + const std::string out_var_name_val = out_var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, + time_out, ch, this] { + auto* var = p_scope->FindVar(in_var_name_val); + + ::grpc::ByteBuffer req; + SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req, out_var_name_val); + + // var handle + VarHandle var_h; + var_h.ep = ep_val; + var_h.scope = p_scope; + var_h.name = out_var_name_val; + var_h.ctx = p_ctx; + var_h.method = "Prefetch"; + + VLOG(3) << var_h.String() << " begin"; + + // stub context + GetProcessor* s = new GetProcessor(ch); + s->Prepare(var_h, time_out); + s->response_call_back_ = ProcGetResponse; + + auto call = s->stub_g_.PrepareUnaryCall( + s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, + &cq_); + call->StartCall(); + call->Finish(&s->reply_, &s->status_, static_cast(s)); + }); + + req_count_++; + return true; +} + +void GRPCClient::AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out) { + const auto ch = GetChannel(ep); + + BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); + s->Prepare(time_out); + + sendrecv::VariableMessage req; + req.set_varname(BATCH_BARRIER_MESSAGE); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + req_count_++; +} + +void GRPCClient::AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out) { + const auto ch = GetChannel(ep); + FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); + s->Prepare(time_out); + + sendrecv::VariableMessage req; + req.set_varname(FETCH_BARRIER_MESSAGE); + auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + req_count_++; +} + +void GRPCClient::AsyncSendComplete(const std::string& ep, int64_t time_out) { + const auto ch = GetChannel(ep); + + BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); + s->Prepare(time_out); + + sendrecv::VariableMessage req; + req.set_varname(COMPLETE_MESSAGE); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + req_count_++; +} + +void GRPCClient::AsyncCheckpointNotify(const std::string& ep, + const std::string& dir, + int64_t time_out) { + const auto ch = GetChannel(ep); + + CheckpointNotifyProcessor* s = new CheckpointNotifyProcessor(ch); + s->Prepare(time_out); + + sendrecv::VariableMessage req; + req.set_varname(CHECKPOINT_SAVE_MESSAGE); + req.set_out_varname(dir); + + auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + req_count_++; +} + +bool GRPCClient::Wait() { + std::unique_lock lk(sync_mutex_); + sync_cond_.wait(lk, [this] { return (req_count_ == 0 || ok_ == false); }); + return ok_; +} + +void GRPCClient::Proceed() { + void* tag = nullptr; + bool ok = false; + + while (!stopped_ && cq_.Next(&tag, &ok)) { + BaseProcessor* c = static_cast(tag); + GPR_ASSERT(ok); + PADDLE_ENFORCE(c); + if (c->status_.ok()) { + VLOG(3) << c->var_h_.String() << " process"; + c->Process(); + } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) { + LOG(ERROR) << c->var_h_.String() + << " meets grpc error:" << c->status_.error_message(); + { + std::lock_guard lk(sync_mutex_); + ok_ = false; + } + sync_cond_.notify_all(); + } else { + LOG(FATAL) << c->var_h_.String() + << " meets grpc error:" << c->status_.error_message(); + } + delete c; + { + std::lock_guard lk(sync_mutex_); + req_count_--; + } + sync_cond_.notify_all(); + } +} + +std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { + std::lock_guard guard(chan_mutex_); + auto it = channels_.find(ep); + if (it != channels_.end()) { + return it->second; + } + + // Channel configurations: + grpc::ChannelArguments args; + args.SetInt(GRPC_ARG_MAX_RECONNECT_BACKOFF_MS, 2000); + args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE); + args.SetMaxSendMessageSize(std::numeric_limits::max()); + args.SetMaxReceiveMessageSize(std::numeric_limits::max()); + + auto ch = + grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args); + channels_[ep] = ch; + return ch; +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_client.h b/paddle/fluid/operators/distributed/grpc_client.h new file mode 100644 index 0000000000000000000000000000000000000000..0c95ffeb5ce7e1586c5968fb122acd12c0c0196e --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_client.h @@ -0,0 +1,246 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include // NOLINT +#include // NOLINT +#include +#include +#include +#include +#include // NOLINT +#include +#include // NOLINT +#include + +#include "grpc++/channel.h" +#include "grpc++/generic/generic_stub.h" +#include "grpc++/grpc++.h" +#include "grpc++/support/byte_buffer.h" +#include "grpc++/support/slice.h" +#include "grpc/support/log.h" +#include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN + +namespace paddle { +namespace operators { +namespace distributed { + +void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg); + +class BaseProcessor { + public: + explicit BaseProcessor(std::shared_ptr ch) { + context_ = nullptr; + } + + virtual ~BaseProcessor() {} + + virtual void Prepare(const VarHandle& var_info, int64_t time_out) { + context_.reset(new grpc::ClientContext()); + var_h_ = var_info; + context_->set_wait_for_ready(true); + if (time_out) { + std::chrono::system_clock::time_point deadline = + std::chrono::system_clock::now() + + std::chrono::milliseconds(time_out); + context_->set_deadline(deadline); + } + } + + virtual void Prepare(int64_t time_out) { + context_.reset(new grpc::ClientContext()); + context_->set_wait_for_ready(true); + + std::chrono::system_clock::time_point deadline = + std::chrono::system_clock::now() + std::chrono::milliseconds(time_out); + + context_->set_deadline(deadline); + } + + virtual void Process() = 0; + + std::unique_ptr context_; + grpc::Status status_; + VarHandle var_h_; +}; + +typedef std::function + RequestSendCallBack; + +class SendProcessor : public BaseProcessor { + public: + explicit SendProcessor(std::shared_ptr ch) + : BaseProcessor(ch), stub_g_(ch) {} + + virtual ~SendProcessor() {} + + virtual void Process() { + if (response_call_back_) { + response_call_back_(var_h_, reply_); + } + } + + ::grpc::GenericStub stub_g_; + ::grpc::ByteBuffer reply_; + RequestSendCallBack response_call_back_ = nullptr; +}; + +typedef std::function + RequestGetCallBack; + +class GetProcessor : public BaseProcessor { + public: + explicit GetProcessor(std::shared_ptr ch) + : BaseProcessor(ch), stub_g_(ch) {} + + virtual ~GetProcessor() {} + + virtual void Process() { + if (response_call_back_) { + response_call_back_(var_h_, reply_); + } + } + + ::grpc::ByteBuffer reply_; + ::grpc::GenericStub stub_g_; + RequestGetCallBack response_call_back_ = ProcGetResponse; +}; + +class BatchBarrierProcessor : public BaseProcessor { + public: + explicit BatchBarrierProcessor(std::shared_ptr ch) + : BaseProcessor(ch) { + stub_ = sendrecv::SendRecvService::NewStub(ch); + } + + virtual ~BatchBarrierProcessor() {} + + virtual void Process() {} + sendrecv::VoidMessage reply_; + std::unique_ptr stub_; +}; + +class FetchBarrierProcessor : public BaseProcessor { + public: + explicit FetchBarrierProcessor(std::shared_ptr ch) + : BaseProcessor(ch) { + stub_ = sendrecv::SendRecvService::NewStub(ch); + } + + virtual ~FetchBarrierProcessor() {} + + virtual void Process() {} + sendrecv::VariableMessage reply_; + std::unique_ptr stub_; +}; + +class CheckpointNotifyProcessor : public BaseProcessor { + public: + explicit CheckpointNotifyProcessor(std::shared_ptr ch) + : BaseProcessor(ch) { + stub_ = sendrecv::SendRecvService::NewStub(ch); + } + + virtual ~CheckpointNotifyProcessor() {} + + virtual void Process() {} + sendrecv::VoidMessage reply_; + std::unique_ptr stub_; +}; + +class GRPCClient : public RPCClient { + public: + GRPCClient() : ok_(true), completed_(false), stopped_(false) {} + virtual ~GRPCClient(); + + bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + + bool AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + + bool AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + + void AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) override; + + void AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) override; + + void AsyncCheckpointNotify(const std::string& ep, const std::string& dir, + int64_t time_out = FLAGS_rpc_deadline) override; + + void AsyncSendComplete(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) override; + + bool Wait() override; + + void SendComplete() override; + + protected: + void InitImpl() override; + + private: + // InitEventLoop should only be called by Init() + void InitEventLoop(); + + void Proceed(); + + std::shared_ptr GetChannel(const std::string& ep); + + private: + grpc::CompletionQueue cq_; + std::unordered_map> channels_; + std::unique_ptr client_thread_; + + // mutex for Wait client sync + std::mutex sync_mutex_; + std::condition_variable sync_cond_; + std::atomic req_count_{0}; + bool ok_; + + // mutex for GetChannel thread safety + std::mutex chan_mutex_; + DISABLE_COPY_AND_ASSIGN(GRPCClient); + + // mutex for sending complete message only once + std::mutex completed_mutex_; + bool completed_; + + volatile bool stopped_; +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc_serde.cc new file mode 100644 index 0000000000000000000000000000000000000000..3f8796713a6b89a308113981614673e07e8d367f --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_serde.cc @@ -0,0 +1,157 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +#include +#endif +#include +#include // NOLINT + +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/operators/distributed/grpc_bytebuffer_stream.h" +#include "paddle/fluid/operators/distributed/grpc_serde.h" +#include "paddle/fluid/operators/distributed/grpc_variable_response.h" +#include "paddle/fluid/operators/distributed/proto_encoder_helper.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace operators { +namespace distributed { + +void SerializeToByteBuffer(const std::string& name, framework::Variable* var, + const platform::DeviceContext& ctx, + ::grpc::ByteBuffer* msg, + const std::string& out_name) { + // Default DestroyCallback does nothing, When using GPU + // the CPU buffer need to be freed. + DestroyCallback destroy_callback = [](void* backing) {}; + VarMsg request; + void* payload = nullptr; + size_t payload_size; + + request.set_varname(name); + // Note: normally the profiler is enabled in 1 trainer, hence only + // 1 trainer returns true for ShouldSendProfileState(). It tells PS + // servers the trainer's profiling state so that PS can follow the + // trainer. + if (platform::ShouldSendProfileState()) { + if (platform::IsProfileEnabled()) { + request.set_profile(platform::kEnableProfiler); + } else { + request.set_profile(platform::kDisableProfiler); + } + } + if (!out_name.empty()) { + request.set_out_varname(out_name); + } + if (var->IsType()) { + request.set_type(::sendrecv::LOD_TENSOR); + GetTensorPayload(var, ctx, &request, &payload, &payload_size); + } else if (var->IsType()) { + request.set_type(::sendrecv::SELECTED_ROWS); + GetSelectedRowsPayload(var, ctx, &request, &payload, &payload_size); +#ifdef PADDLE_WITH_CUDA + } else if (var->IsType()) { + request.set_type(::sendrecv::NCCL_ID); +#endif + } else { + PADDLE_THROW("Serialize does not support type: %s", + typeid(var->Type()).name()); + } + + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + // GPU data is copied to CPU buffer when sending, + // free the buffer when possible. + destroy_callback = [](void* backing) { + platform::CUDAPinnedPlace cuda_pinned; + memory::Free(cuda_pinned, backing); + }; +#endif + } + + std::string header; + request.AppendToString(&header); + auto buffer = std::unique_ptr(new char[1024]); + void* buf = buffer.get(); + ProtoEncodeHelper e(static_cast(buf), 1024); + e.WriteRawBytes(std::string(header.data(), header.size())); +// NCCLID is copied directly to the message, return bytebuffer +// with only one slice if serializing NCCLID. +#ifdef PADDLE_WITH_CUDA + if (var->IsType()) { + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, + NCCL_UNIQUE_ID_BYTES); + const ncclUniqueId& uid = var->Get(); + e.WriteRawBytes(std::string(uid.internal, NCCL_UNIQUE_ID_BYTES)); + + // for serialize NCCL_ID + ::grpc::Slice slices(e.size()); + memcpy(const_cast(slices.begin()), e.data(), e.size()); + ::grpc::ByteBuffer tmp(&slices, 1); + msg->Swap(&tmp); + return; + } +#endif + + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); + // steal reference of tensor data + ::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows + int num_slices = 2; // only SelectedRows have rows buffer + slices[0] = ::grpc::Slice(e.size()); + memcpy(const_cast(slices[0].begin()), e.data(), e.size()); + slices[1] = ::grpc::Slice( + grpc_slice_new_with_user_data(payload, payload_size, destroy_callback, + static_cast(payload)), + ::grpc::Slice::STEAL_REF); + + if (var->IsType()) { + auto* slr = var->GetMutable(); + ProtoEncodeHelper e2(static_cast(buf), 128); + size_t rows_memory_size = + slr->rows().size() * framework::SizeOfType(typeid(int64_t)); + e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); + slices[2] = ::grpc::Slice(e2.size()); + memcpy(const_cast(slices[2].begin()), e2.data(), e2.size()); + + slices[3] = ::grpc::Slice( + grpc_slice_new_with_user_data( + const_cast( + reinterpret_cast(slr->rows().data())), + rows_memory_size, [](void* backing) {}, + const_cast( + reinterpret_cast(slr->rows().data()))), + ::grpc::Slice::STEAL_REF); + num_slices = 4; + } + + ::grpc::ByteBuffer tmp(&slices[0], num_slices); + msg->Swap(&tmp); +} + +void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, + const platform::DeviceContext& ctx, + const framework::Scope* scope, + framework::Variable** var) { + operators::distributed::GRPCVariableResponse resp(scope, &ctx); + PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); + *var = resp.GetVar(); +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_serde.h b/paddle/fluid/operators/distributed/grpc_serde.h new file mode 100644 index 0000000000000000000000000000000000000000..450c41dcd6b1bf9a33d3bbef3a1c94a2f83ff322 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_serde.h @@ -0,0 +1,50 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" + +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +namespace paddle { +namespace operators { +namespace distributed { + +typedef void (*DestroyCallback)(void*); + +void SerializeToByteBuffer(const std::string& name, framework::Variable* var, + const platform::DeviceContext& ctx, + ::grpc::ByteBuffer* msg, + const std::string& out_varname = std::string()); + +void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, + const platform::DeviceContext& ctx, + const framework::Scope* scope, + framework::Variable** var); + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_serde_test.cc b/paddle/fluid/operators/distributed/grpc_serde_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..96ea05e74ed76768248a27ab435dc801b7d1b995 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_serde_test.cc @@ -0,0 +1,223 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include // NOLINT + +#include "google/protobuf/text_format.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/framework/variable.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/grpc_serde.h" +#include "paddle/fluid/operators/distributed/grpc_variable_response.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/string/printf.h" + +namespace framework = paddle::framework; +namespace platform = paddle::platform; +namespace operators = paddle::operators; +namespace math = paddle::operators::math; +namespace memory = paddle::memory; + +void RunSerdeTestSelectedRows(platform::Place place) { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + // serialize var to ByteBuffer + framework::Variable var; + auto* slr = var.GetMutable(); + slr->set_height(1000); + auto* tensor = slr->mutable_value(); + auto* rows = slr->mutable_rows(); + tensor->Resize(framework::make_ddim({564, 128})); + tensor->mutable_data(place); + int tensor_numel = 564 * 128; + math::set_constant(ctx, tensor, 32.7); + for (int i = 0; i < 564; ++i) rows->push_back(i); + + ::grpc::ByteBuffer msg; + operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg); + EXPECT_GT(msg.Length(), static_cast(0)); + + // deserialize + std::vector<::grpc::Slice> slices; + (void)msg.Dump(&slices); + std::string tmp; + for (const auto& s : slices) { + tmp.append(reinterpret_cast(s.begin()), s.size()); + } + + sendrecv::VariableMessage varmsg; + EXPECT_TRUE(varmsg.ParseFromString(tmp)); + + // deserialize bytebuffer + EXPECT_EQ(varmsg.varname(), "myvar"); + EXPECT_EQ(varmsg.type(), 1); + + const float* tensor_data = + reinterpret_cast(varmsg.serialized().data()); + const int64_t* rows_data = + reinterpret_cast(varmsg.rows().data()); + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data[i], 32.7); + } + for (int i = 0; i < 564; ++i) { + EXPECT_EQ(rows_data[i], i); + } + + // deserialize zero-copy + // framework::Variable var2; + // operators::distributed::DeserializeFromByteBuffer(msg, ctx, &var2); + framework::Scope scope; + scope.Var("myvar"); + operators::distributed::GRPCVariableResponse resp(&scope, &ctx); + EXPECT_EQ(resp.Parse(msg), 0); + + framework::Variable* var2 = resp.GetVar(); + + auto* slr2 = var2->GetMutable(); + auto* tensor2 = slr2->mutable_value(); + auto* rows2 = slr2->mutable_rows(); + float* tensor_data2 = nullptr; + framework::Tensor tmp_tensor; + + if (platform::is_gpu_place(ctx.GetPlace())) { + platform::CPUPlace cpu; + framework::TensorCopy(*tensor2, cpu, &tmp_tensor); + tensor_data2 = tmp_tensor.data(); + } else { + tensor_data2 = const_cast(tensor2->data()); + } + const int64_t* rows_data2 = rows2->data(); + + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); + } + for (size_t i = 0; i < rows2->size(); ++i) { + EXPECT_EQ(rows_data2[i], static_cast(i)); + } + EXPECT_EQ(slr2->height(), 1000); +} + +void RunTestLodTensor(platform::Place place, int from_type = 0) { + // serialize var to ByteBuffer + framework::Variable var; + auto* tensor = var.GetMutable(); + tensor->Resize(framework::make_ddim({512, 8, 4, 2})); + framework::LoD lod; + lod.push_back(framework::Vector({1, 3, 8})); + tensor->set_lod(lod); + int tensor_numel = 512 * 8 * 4 * 2; + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + tensor->mutable_data(place); + math::set_constant(ctx, tensor, 31.9); + + ::grpc::ByteBuffer msg; + operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg); + EXPECT_GT(msg.Length(), static_cast(0)); + + // deserialize + std::vector<::grpc::Slice> slices; + (void)msg.Dump(&slices); + std::string tmp; + for (const auto& s : slices) { + tmp.append(reinterpret_cast(s.begin()), s.size()); + } + sendrecv::VariableMessage varmsg; + EXPECT_TRUE(varmsg.ParseFromString(tmp)); + EXPECT_EQ(varmsg.varname(), "myvar"); + EXPECT_EQ(varmsg.type(), 0); + EXPECT_EQ(varmsg.dims()[0], 512); + EXPECT_EQ(varmsg.dims()[1], 8); + EXPECT_EQ(varmsg.dims()[2], 4); + EXPECT_EQ(varmsg.dims()[3], 2); + EXPECT_EQ(varmsg.lod_level(), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); + EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); + + const float* tensor_data = + reinterpret_cast(varmsg.serialized().data()); + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data[i], 31.9); + } + + // message binary + std::string str; + varmsg.SerializeToString(&str); + + // message bytebuffer + ::grpc::Slice slices_2[1]; + int num_slices = 1; + slices_2[0] = ::grpc::Slice(str.length()); + memcpy(const_cast(slices_2[0].begin()), str.c_str(), str.length()); + ::grpc::ByteBuffer bytebuffer2(&slices_2[0], num_slices); + + // deserialize zero-copy + framework::Scope scope; + scope.Var("myvar"); + operators::distributed::GRPCVariableResponse resp(&scope, &ctx); + if (from_type == 0) { + EXPECT_EQ(resp.Parse(msg), 0); + } else { + EXPECT_EQ(resp.Parse(bytebuffer2), 0); + } + + framework::Variable* var2 = resp.GetVar(); + + auto tensor2 = var2->Get(); + float* tensor_data2 = nullptr; + framework::Tensor tmp_tensor; + + if (platform::is_gpu_place(ctx.GetPlace())) { + platform::CPUPlace cpu; + framework::TensorCopy(tensor2, cpu, &tmp_tensor); + tensor_data2 = tmp_tensor.data(); + } else { + tensor_data2 = const_cast(tensor2.data()); + } + + EXPECT_EQ(varmsg.lod_level(), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); + EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); + for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9); +} + +TEST(LodTensor, Run) { + platform::CPUPlace place; + RunTestLodTensor(place); + RunTestLodTensor(place, 1); +#ifdef PADDLE_WITH_CUDA + platform::CUDAPlace gpu(0); + RunTestLodTensor(gpu); + RunTestLodTensor(gpu, 1); +#endif +} + +TEST(SelectedRows, Run) { + platform::CPUPlace place; + RunSerdeTestSelectedRows(place); + +#ifdef PADDLE_WITH_CUDA + platform::CUDAPlace gpu; + RunSerdeTestSelectedRows(gpu); +#endif +} diff --git a/paddle/fluid/operators/distributed/grpc_server.cc b/paddle/fluid/operators/distributed/grpc_server.cc new file mode 100644 index 0000000000000000000000000000000000000000..8edb00276df3ade1b320fbf2873e8b54ff3e1464 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_server.cc @@ -0,0 +1,415 @@ +/*Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "paddle/fluid/operators/distributed/grpc_serde.h" +#include "paddle/fluid/operators/distributed/grpc_server.h" + +using ::grpc::ServerAsyncResponseWriter; + +namespace paddle { +namespace operators { +namespace distributed { +enum CallStatus { PROCESS = 0, FINISH }; + +// reference: +// https://stackoverflow.com/questions/41732884/grpc-multiple-services-in-cpp-async-server +class RequestBase { + public: + explicit RequestBase(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : service_(service), + cq_(cq), + status_(PROCESS), + request_handler_(request_handler), + req_id_(req_id) { + PADDLE_ENFORCE(cq_); + } + virtual ~RequestBase() {} + virtual void Process() = 0; + + std::string Status2String(const std::string& method) { + std::string status = "Process"; + if (status_ == FINISH) { + status = "Finish"; + } + + std::ostringstream s; + s << method << " name:[" << GetReqName() << "]" + << ", ep:[" << ctx_.peer() << "]" + << " " << status << " using req_id:" << req_id_; + return s.str(); + } + + CallStatus Status() const { + std::lock_guard l(status_mu_); + return status_; + } + + template + void Finish(const T& reply, ServerAsyncResponseWriter* responder) { + std::lock_guard l(status_mu_); + status_ = FINISH; + responder->Finish(reply, ::grpc::Status::OK, + reinterpret_cast(static_cast(req_id_))); + } + virtual std::string GetReqName() = 0; + + protected: + mutable std::mutex status_mu_; + ::grpc::ServerContext ctx_; + GrpcService::AsyncService* service_; + ::grpc::ServerCompletionQueue* cq_; + CallStatus status_; + RequestHandler* request_handler_; + int req_id_; +}; + +class RequestSend final : public RequestBase { + public: + explicit RequestSend(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { + request_.reset(new GRPCVariableResponse(request_handler->scope(), + request_handler->dev_ctx(), + !request_handler->sync_mode())); + int method_id = static_cast(distributed::GrpcMethod::kSendVariable); + service_->RequestAsyncUnary( + method_id, &ctx_, request_.get(), &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); + } + virtual ~RequestSend() {} + std::string GetReqName() override { return request_->Varname(); } + + void Process() override { + std::string varname = GetReqName(); + VLOG(4) << "RequestSend var_name:" << varname; + + auto scope = request_->GetMutableLocalScope(); + auto invar = request_->GetVar(); + framework::Variable* outvar = nullptr; + + request_handler_->Handle(varname, scope, invar, &outvar); + Finish(reply_, &responder_); + } + + protected: + sendrecv::VoidMessage reply_; + std::shared_ptr request_; + ServerAsyncResponseWriter responder_; +}; + +class RequestGet final : public RequestBase { + public: + explicit RequestGet(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { + auto method_id = static_cast(distributed::GrpcMethod::kGetVariable); + service_->RequestAsyncUnary( + method_id, &ctx_, &request_, &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); + } + + virtual ~RequestGet() {} + + std::string GetReqName() override { return request_.varname(); } + + void Process() override { + // proc request. + std::string varname = request_.varname(); + VLOG(4) << "RequestGet " << varname; + + auto scope = request_handler_->scope(); + auto invar = scope->FindVar(varname); + framework::Variable* outvar = nullptr; + + request_handler_->Handle(varname, scope, invar, &outvar); + + if (outvar) { + SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), + &reply_); + } + Finish(reply_, &responder_); + } + + protected: + sendrecv::VariableMessage request_; + ::grpc::ByteBuffer reply_; + ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; +}; + +class RequestPrefetch final : public RequestBase { + public: + explicit RequestPrefetch(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), + responder_(&ctx_), + local_scope_(nullptr) { + request_.reset(new GRPCVariableResponse(request_handler->scope(), + request_handler->dev_ctx(), true)); + int method_id = + static_cast(distributed::GrpcMethod::kPrefetchVariable); + service_->RequestAsyncUnary( + method_id, &ctx_, request_.get(), &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); + } + + virtual ~RequestPrefetch() {} + + std::string GetReqName() override { return request_->Varname(); } + + void Process() override { + // prefetch process... + std::string in_var_name = request_->Varname(); + std::string out_var_name = request_->OutVarname(); + VLOG(4) << "RequestPrefetch, in_var_name: " << in_var_name + << " out_var_name: " << out_var_name; + + auto scope = request_->GetMutableLocalScope(); + auto invar = scope->FindVar(in_var_name); + // out var must be created in local scope! + framework::Variable* outvar = scope->Var(out_var_name); + + request_handler_->Handle(in_var_name, scope, invar, &outvar, out_var_name); + + SerializeToByteBuffer(out_var_name, outvar, *request_handler_->dev_ctx(), + &reply_); + Finish(reply_, &responder_); + } + + protected: + std::shared_ptr request_; + ::grpc::ByteBuffer reply_; + ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; + framework::Scope* local_scope_; +}; + +class RequestCheckpointNotify final : public RequestBase { + public: + explicit RequestCheckpointNotify(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { + request_.reset(new GRPCVariableResponse(request_handler->scope(), + request_handler->dev_ctx())); + int method_id = + static_cast(distributed::GrpcMethod::kCheckpointNotify); + service_->RequestAsyncUnary( + method_id, &ctx_, request_.get(), &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); + } + + virtual ~RequestCheckpointNotify() {} + + std::string GetReqName() override { return request_->Varname(); } + + void Process() override { + auto scope = request_->GetMutableLocalScope(); + + std::string checkpoint_notify = request_->Varname(); + std::string checkpoint_dir = request_->OutVarname(); + + VLOG(4) << "RequestCheckpointNotify notify: " << checkpoint_notify + << ", dir: " << checkpoint_dir; + + request_handler_->Handle(checkpoint_notify, scope, nullptr, nullptr, + checkpoint_dir); + Finish(reply_, &responder_); + } + + protected: + std::shared_ptr request_; + sendrecv::VoidMessage reply_; + ServerAsyncResponseWriter responder_; +}; + +void AsyncGRPCServer::WaitServerReady() { + VLOG(4) << "AsyncGRPCServer is wait server ready"; + std::unique_lock lock(this->mutex_ready_); + condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); + VLOG(4) << "AsyncGRPCServer WaitSeverReady"; +} + +void AsyncGRPCServer::StartServer() { + ::grpc::ServerBuilder builder; + builder.AddListeningPort(bind_address_, ::grpc::InsecureServerCredentials(), + &selected_port_); + + builder.SetMaxSendMessageSize(std::numeric_limits::max()); + builder.SetMaxReceiveMessageSize(std::numeric_limits::max()); + builder.RegisterService(&service_); + + for (auto t : rpc_call_map_) { + rpc_cq_[t.first].reset(builder.AddCompletionQueue().release()); + } + + server_ = builder.BuildAndStart(); + LOG(INFO) << "Server listening on " << bind_address_ + << " selected port: " << selected_port_; + + std::function f = + std::bind(&AsyncGRPCServer::TryToRegisterNewOne, this, + std::placeholders::_1, std::placeholders::_2); + + for (auto& t : rpc_call_map_) { + auto& rpc_name = t.first; + auto& cq = rpc_cq_[rpc_name]; + auto threadnum = rpc_thread_num_[rpc_name]; + auto& reqs = rpc_reqs_[rpc_name]; + + reqs.reserve(kRequestBufSize); + + for (int i = 0; i < kRequestBufSize; i++) { + VLOG(6) << "TryToRegisterNewOne on RPC NAME: " << rpc_name << " I: " << i; + TryToRegisterNewOne(rpc_name, i); + } + + for (int i = 0; i < threadnum; i++) { + rpc_threads_[rpc_name].emplace_back(new std::thread(std::bind( + &AsyncGRPCServer::HandleRequest, this, cq.get(), rpc_name, f))); + VLOG(4) << t.first << " creates threads!"; + } + } + + { + std::lock_guard lock(this->mutex_ready_); + ready_ = 1; + } + condition_ready_.notify_all(); + + // wait server + server_->Wait(); + + for (auto& t : rpc_threads_) { + auto& threads = t.second; + for (size_t i = 0; i < threads.size(); ++i) { + threads[i]->join(); + VLOG(4) << t.first << " threads ends!"; + } + } +} + +void AsyncGRPCServer::ShutdownQueue() { + for (auto& t : rpc_cq_) { + t.second->Shutdown(); + VLOG(4) << t.first << " queue shutdown!"; + } +} + +void AsyncGRPCServer::ShutDownImpl() { + std::unique_lock lock(cq_mutex_); + is_shut_down_ = true; + ShutdownQueue(); + + VLOG(4) << "server_ shutdown!"; + server_->Shutdown(); +} + +void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, + int req_id) { + std::unique_lock lock(cq_mutex_); + if (is_shut_down_) { + VLOG(4) << "shutdown, do not TryToRegisterNewSendOne"; + return; + } + + VLOG(4) << "TryToRegisterNewOne on RPC NAME: " << rpc_name + << " REQ ID: " << req_id; + + auto& reqs = rpc_reqs_[rpc_name]; + auto& handler = rpc_call_map_[rpc_name]; + auto& cq = rpc_cq_[rpc_name]; + + RequestBase* b = nullptr; + if (rpc_name == kRequestSend) { + b = new RequestSend(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestGet) { + b = new RequestGet(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestPrefetch) { + b = new RequestPrefetch(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestCheckpoint) { + b = new RequestCheckpointNotify(&service_, cq.get(), handler, req_id); + } else { + PADDLE_ENFORCE(false, "not supported rpc"); + } + + reqs[req_id] = b; + + VLOG(4) << "Create RequestSend status:" << b->Status(); +} + +void AsyncGRPCServer::HandleRequest( + ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, + std::function TryToRegisterNewOne) { + void* tag = NULL; + bool ok = false; + + while (true) { + VLOG(4) << "HandleRequest " << rpc_name << " wait next"; + if (!cq->Next(&tag, &ok)) { + VLOG(3) << "CompletionQueue " << rpc_name << " shutdown!"; + break; + } + + int req_id = static_cast(reinterpret_cast(tag)); + VLOG(4) << "HandleRequest " << rpc_name << ", req_id:" << req_id + << " get next"; + + auto& reqs = rpc_reqs_[rpc_name]; + RequestBase* base = nullptr; + { + PADDLE_ENFORCE(req_id >= 0 && req_id < kRequestBufSize); + std::unique_lock lock(cq_mutex_); + base = reqs[req_id]; + } + + VLOG(3) << base->Status2String(rpc_name); + + // reference: + // https://github.com/tensorflow/tensorflow/issues/5596 + // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM + // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I + if (!ok) { + LOG(WARNING) << "completion queue:" << rpc_name + << " recv no regular event" + << " context:" << base->Status2String(rpc_name); + TryToRegisterNewOne(rpc_name, req_id); + delete base; + continue; + } + + switch (base->Status()) { + case PROCESS: { + base->Process(); + break; + } + case FINISH: { + TryToRegisterNewOne(rpc_name, req_id); + delete base; + break; + } + default: { assert(false); } + } + } +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_server.h b/paddle/fluid/operators/distributed/grpc_server.h new file mode 100644 index 0000000000000000000000000000000000000000..d2524f5e65db6dedab78f45e17380359b58a3d11 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_server.h @@ -0,0 +1,89 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include // NOLINT +#include +#include + +#include "grpc++/grpc++.h" +#include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/distributed/grpc_service.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class RequestBase; + +class AsyncGRPCServer final : public RPCServer { + public: + explicit AsyncGRPCServer(const std::string& address, int client_num) + : RPCServer(address, client_num), ready_(0) {} + + virtual ~AsyncGRPCServer() {} + void WaitServerReady() override; + void StartServer() override; + + private: + // HandleRequest needs to be thread-safe. + void HandleRequest( + ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, + std::function TryToRegisterNewOne); + + void TryToRegisterNewOne(const std::string& rpc_name, int req_id); + void ShutdownQueue(); + void ShutDownImpl() override; + + private: + static const int kRequestBufSize = 100; + + std::mutex cq_mutex_; + volatile bool is_shut_down_ = false; + + GrpcService::AsyncService service_; + std::unique_ptr<::grpc::Server> server_; + + // condition of the sub program + std::condition_variable barrier_condition_; + + std::mutex mutex_ready_; + std::condition_variable condition_ready_; + + int ready_; + + std::map> rpc_cq_; + std::map>> rpc_threads_; + std::map> rpc_reqs_; +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_service.h b/paddle/fluid/operators/distributed/grpc_service.h new file mode 100644 index 0000000000000000000000000000000000000000..9ae9a31a003cbb1f808fd1127a5dd78511aa3e99 --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_service.h @@ -0,0 +1,127 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "paddle/fluid/operators/distributed/grpc_variable_response.h" +#include "paddle/fluid/platform/profiler.h" + +// NOTE: This method was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// method and did some modifications so that we can parse gRPC +// requests without too much copying of the tensor data. + +namespace grpc { +class CompletionQueue; +class Channel; +class RpcService; +class ServerCompletionQueue; +class ServerContext; + +// Support parsing/unparsing of tensorflow::VariableResponse. +// Wire-format is identical to RecvVariableResponse. +template <> +class SerializationTraits< + paddle::operators::distributed::GRPCVariableResponse> { + public: + static Status Serialize( + const paddle::operators::distributed::GRPCVariableResponse& msg, + grpc_byte_buffer** bp, bool* own_buffer) { + PADDLE_ENFORCE(false, "SerializationTraits::Serialize not implemented!"); + return Status(); + } + static Status Deserialize( + grpc_byte_buffer* buffer, + paddle::operators::distributed::GRPCVariableResponse* msg, + int max_message_size = INT_MAX) { + if (buffer == nullptr) { + return Status(StatusCode::INTERNAL, "No payload"); + } + + Status result = g_core_codegen_interface->ok(); + if (result.ok()) { + paddle::operators::distributed::GrpcByteSource source(buffer); + int ret = msg->Parse(&source); + if (ret != 0) { + result = Status(StatusCode::INTERNAL, "VariableResponse parse error"); + } + } + g_core_codegen_interface->grpc_byte_buffer_destroy(buffer); + return result; + } +}; +} // namespace grpc + +namespace paddle { +namespace operators { +namespace distributed { + +enum class GrpcMethod { + kSendVariable, + kGetVariable, + kPrefetchVariable, + kCheckpointNotify, +}; + +static const int kGrpcNumMethods = + static_cast(GrpcMethod::kCheckpointNotify) + 1; + +inline const char* GrpcMethodName(GrpcMethod id) { + switch (id) { + case GrpcMethod::kSendVariable: + return "/sendrecv.SendRecvService/SendVariable"; + case GrpcMethod::kGetVariable: + return "/sendrecv.SendRecvService/GetVariable"; + case GrpcMethod::kPrefetchVariable: + return "/sendrecv.SendRecvService/PrefetchVariable"; + case GrpcMethod::kCheckpointNotify: + return "/sendrecv.SendRecvService/CheckpointNotify"; + } + + // Shouldn't be reached. + PADDLE_ENFORCE(false, "Invalid id: not found valid method name"); + return nullptr; +} + +class GrpcService final { + public: + class AsyncService : public ::grpc::Service { + public: + AsyncService() { + for (int i = 0; i < kGrpcNumMethods; ++i) { + AddMethod(new ::grpc::internal::RpcServiceMethod( + GrpcMethodName(static_cast(i)), + ::grpc::internal::RpcMethod::NORMAL_RPC, nullptr)); + ::grpc::Service::MarkMethodAsync(i); + } + } + virtual ~AsyncService() {} + + // Make RequestAsyncUnary public for grpc_call.h + using ::grpc::Service::RequestAsyncUnary; + }; +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_variable_response.cc b/paddle/fluid/operators/distributed/grpc_variable_response.cc new file mode 100644 index 0000000000000000000000000000000000000000..34d47f3ec0f3025109447b66078b724607d2953a --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_variable_response.cc @@ -0,0 +1,308 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include +#ifdef PADDLE_WITH_CUDA +#include +#endif + +#include "paddle/fluid/operators/distributed/grpc_variable_response.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace operators { +namespace distributed { + +enum WireType { + WIRETYPE_VARINT = 0, + WIRETYPE_LENGTH_DELIMITED = 2, +}; + +inline int GetTagFieldNumber(uint32_t tag) { return tag >> 3; } + +inline WireType GetTagWireType(uint32_t tag) { + return static_cast(tag & 0x7); +} + +bool ReadVarintSizeAsInt(::google::protobuf::io::CodedInputStream* input, + int* result) { + uint64_t v; + if (input->ReadVarint64(&v) && v <= static_cast(INT_MAX)) { + *result = static_cast(v); + return true; + } else { + return false; + } +} + +int GRPCVariableResponse::Parse(const ::grpc::ByteBuffer& byte_buffer) { + GrpcByteBufferSource source; + source.Init(byte_buffer); + GrpcByteBufferSourceWrapper r(&source); + + return Parse(&r); +} + +bool ParseLodData(::google::protobuf::io::CodedInputStream* input, + std::vector* lod) { + while (true) { + auto p = input->ReadTagWithCutoff(127); + int tag = GetTagFieldNumber(p.first); + WireType wt = GetTagWireType(p.first); + + if (!p.second) { + return (tag == 0); + } + + switch (tag) { + case sendrecv::VariableMessage_LodData::kLodDataFieldNumber: { + uint64_t v; + if (wt == WIRETYPE_VARINT) { + if (!input->ReadVarint64(&v)) { + return false; + } + lod->push_back(v); + break; + } + + if (wt == WIRETYPE_LENGTH_DELIMITED) { + int num_bytes = 0; + if (!input->ReadVarintSizeAsInt(&num_bytes)) { + return tag; + } + int start_pos = input->CurrentPosition(); + while (input->CurrentPosition() - start_pos < num_bytes) { + uint64_t v; + if (!input->ReadVarint64(&v)) { + return tag; + } + lod->push_back(v); + } + break; + } + + return false; + } + default: { return false; } + } + } + + return true; +} + +int GRPCVariableResponse::Parse(Source* source) { + ::google::protobuf::io::ZeroCopyInputStream* input_stream = + source->contents(); + ::google::protobuf::io::CodedInputStream input(input_stream); + input.SetTotalBytesLimit(INT_MAX, INT_MAX); + + while (true) { + auto p = input.ReadTagWithCutoff(127); + int tag = GetTagFieldNumber(p.first); + WireType wt = GetTagWireType(p.first); + if (!p.second) { + if (tag != 0) { + return -1; + } + return 0; + } + + switch (tag) { + case sendrecv::VariableMessage::kVarnameFieldNumber: { + uint32_t length; + if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) { + return tag; + } + + std::string temp; + if (!input.ReadString(&temp, length)) { + return tag; + } + + meta_.set_varname(temp); + break; + } + case sendrecv::VariableMessage::kTypeFieldNumber: { + uint32_t v; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { + return tag; + } + + meta_.set_type(static_cast<::sendrecv::VarType>(v)); + break; + } + case sendrecv::VariableMessage::kDataTypeFieldNumber: { + uint32_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { + return tag; + } + + meta_.set_data_type(static_cast<::sendrecv::VariableMessage_Type>(v)); + break; + } + case sendrecv::VariableMessage::kDimsFieldNumber: { + // not packed + if (wt == WIRETYPE_VARINT) { + uint64_t v; + if (!input.ReadVarint64(&v)) { + return tag; + } + meta_.add_dims(v); + break; + } + + // packed + if (wt == WIRETYPE_LENGTH_DELIMITED) { + int num_bytes = 0; + if (!input.ReadVarintSizeAsInt(&num_bytes)) { + return tag; + } + int start_pos = input.CurrentPosition(); + while (input.CurrentPosition() - start_pos < num_bytes) { + uint64_t v; + if (!input.ReadVarint64(&v)) { + return tag; + } + meta_.add_dims(v); + } + break; + } + return tag; + } + case sendrecv::VariableMessage::kLodLevelFieldNumber: { + uint64_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + return tag; + } + meta_.set_lod_level(static_cast(v)); + break; + } + case sendrecv::VariableMessage::kLodFieldNumber: { + int length = 0; + if (wt != WIRETYPE_LENGTH_DELIMITED || + !ReadVarintSizeAsInt(&input, &length)) { + return tag; + } + + std::pair<::google::protobuf::io::CodedInputStream::Limit, int> p = + input.IncrementRecursionDepthAndPushLimit(length); + + std::vector lod_data; + if (p.second < 0 || !ParseLodData(&input, &lod_data)) { + return tag; + } + + if (!input.DecrementRecursionDepthAndPopLimit(p.first)) { + return tag; + } + + if (lod_data.size() == 0) { + break; + } + + auto lod = meta_.add_lod(); + for (uint32_t i = 0; i < lod_data.size(); i++) { + lod->add_lod_data(lod_data[i]); + } + break; + } + case sendrecv::VariableMessage::kSlrHeightFieldNumber: { + uint64_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + return tag; + } + meta_.set_slr_height(static_cast(v)); + break; + } + case sendrecv::VariableMessage::kSerializedFieldNumber: { + int num_bytes = 0; + if (wt != WIRETYPE_LENGTH_DELIMITED || + !ReadVarintSizeAsInt(&input, &num_bytes)) { + return tag; + } + + if (!ProcSerializedField(tag, &input, num_bytes)) { + return tag; + } + + break; + } + case sendrecv::VariableMessage::kRowsFieldNumber: { + PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || + meta_.type() == sendrecv::LOD_TENSOR) && + meta_.varname() != "", + "meta info should be got first!"); + + int num_bytes = 0; + if (wt != WIRETYPE_LENGTH_DELIMITED || + !ReadVarintSizeAsInt(&input, &num_bytes)) { + return tag; + } + + if (!CopySelectRowsData(&input, *dev_ctx_, num_bytes)) { + return tag; + } + break; + } + case sendrecv::VariableMessage::kOutVarnameFieldNumber: { + uint32_t length; + if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) { + return tag; + } + + std::string temp; + if (!input.ReadString(&temp, length)) { + return tag; + } + + meta_.set_out_varname(temp); + break; + } + case sendrecv::VariableMessage::kProfileFieldNumber: { + uint64_t profiling = 0; + if (!input.ReadVarint64(&profiling)) { + return tag; + } + meta_.set_profile(profiling); + int64_t listener_id = platform::ListenerId(); + if (listener_id <= 0) { + break; + } + if (profiling == platform::kEnableProfiler && + !platform::IsProfileEnabled()) { + platform::EnableProfiler(platform::ProfilerState::kCPU); + } else if (profiling == platform::kDisableProfiler && + platform::IsProfileEnabled()) { + // TODO(panyx0718): Should we allow to customize file dir. + platform::DisableProfiler( + platform::EventSortingKey::kDefault, + string::Sprintf("/tmp/profile_ps_%lld", listener_id)); + } + break; + } + default: { + // Unknown tag, return unknown error. + return -1; + } + } + } + + return 0; +} + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/grpc_variable_response.h b/paddle/fluid/operators/distributed/grpc_variable_response.h new file mode 100644 index 0000000000000000000000000000000000000000..89df07c92cd33bcb76c8539b5566d74fa21bba5e --- /dev/null +++ b/paddle/fluid/operators/distributed/grpc_variable_response.h @@ -0,0 +1,58 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" + +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/distributed/grpc_bytebuffer_stream.h" +#include "paddle/fluid/operators/distributed/variable_response.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class GRPCVariableResponse : public VariableResponse { + public: + GRPCVariableResponse(const framework::Scope* scope, + const platform::DeviceContext* dev_ctx, + bool create_scope = false) + : VariableResponse(scope, dev_ctx, create_scope) {} + + virtual ~GRPCVariableResponse() {} + + int Parse(Source* source) override; + + // return: + // 0:ok. + // -1: unkown error. + // other: number of error field. + int Parse(const ::grpc::ByteBuffer& byte_buffer); +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/proto_encoder_helper.h b/paddle/fluid/operators/distributed/proto_encoder_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..2fab02e32fe18ee04f86a69bb5bae1cbe7c6762c --- /dev/null +++ b/paddle/fluid/operators/distributed/proto_encoder_helper.h @@ -0,0 +1,149 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +// NOTE: This file was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// file and did some modifications so that we can send gRPC +// requests without too much copying of the tensor data. + +#pragma once + +#include + +#include "grpc++/grpc++.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace distributed { + +char* EncodeVarint32(char* dst, uint32_t v) { + // Operate on characters as unsigneds + unsigned char* ptr = reinterpret_cast(dst); + static const int B = 128; + if (v < (1 << 7)) { + *(ptr++) = v; + } else if (v < (1 << 14)) { + *(ptr++) = v | B; + *(ptr++) = v >> 7; + } else if (v < (1 << 21)) { + *(ptr++) = v | B; + *(ptr++) = (v >> 7) | B; + *(ptr++) = v >> 14; + } else if (v < (1 << 28)) { + *(ptr++) = v | B; + *(ptr++) = (v >> 7) | B; + *(ptr++) = (v >> 14) | B; + *(ptr++) = v >> 21; + } else { + *(ptr++) = v | B; + *(ptr++) = (v >> 7) | B; + *(ptr++) = (v >> 14) | B; + *(ptr++) = (v >> 21) | B; + *(ptr++) = v >> 28; + } + return reinterpret_cast(ptr); +} + +char* EncodeVarint64(char* dst, uint64_t v) { + static const int B = 128; + unsigned char* ptr = reinterpret_cast(dst); + while (v >= B) { + *(ptr++) = (v & (B - 1)) | B; + v >>= 7; + } + *(ptr++) = static_cast(v); + return reinterpret_cast(ptr); +} + +int VarintLength(uint64_t v) { + int len = 1; + while (v >= 128) { + v >>= 7; + len++; + } + return len; +} + +class ProtoEncodeHelper { + public: + ProtoEncodeHelper(char* buf, int max_size) + : base_(buf), p_(buf), limit_(base_ + max_size) {} + + ~ProtoEncodeHelper() { + // Make sure callers didn't do operations that went over max_size promised + PADDLE_ENFORCE_LE(p_, limit_); + } + + const char* data() const { return base_; } + size_t size() const { return p_ - base_; } + + void WriteUint64(int tag, uint64_t v) { + Encode32(combine(tag, WIRETYPE_VARINT)); + Encode64(v); + } + void WriteBool(int tag, bool v) { + Encode32(combine(tag, WIRETYPE_VARINT)); + EncodeBool(v); + } + void WriteString(int tag, const std::string& v) { + Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED)); + Encode32(v.size()); + EncodeBytes(v.data(), v.size()); + } + void WriteVarlengthBeginning(int tag, uint32_t len) { + Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED)); + Encode32(len); + } + void WriteRawBytes(const std::string& v) { EncodeBytes(v.data(), v.size()); } + + private: + // Note: this module's behavior must match the protocol buffer wire encoding + // format. + enum { + WIRETYPE_VARINT = 0, + WIRETYPE_LENGTH_DELIMITED = 2, + }; + static uint32_t combine(uint32_t tag, uint32_t type) { + return ((tag << 3) | type); + } + inline void Encode32(uint32_t v) { + if (v < 128) { + // Fast path for single-byte values. Many of the calls will use a + // constant value for v, so the comparison will get optimized away + // when Encode32 is inlined into the caller. + *p_ = v; + p_++; + } else { + p_ = EncodeVarint32(p_, v); + } + } + void Encode64(uint64_t v) { p_ = EncodeVarint64(p_, v); } + void EncodeBool(bool v) { + *p_ = (v ? 1 : 0); // Equal to varint32 encoding of 0 or 1 + p_++; + } + void EncodeBytes(const char* bytes, int N) { + memcpy(p_, bytes, N); + p_ += N; + } + + char* base_; + char* p_; + char* limit_; // Just for CHECKs +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/request_handler.h b/paddle/fluid/operators/distributed/request_handler.h new file mode 100644 index 0000000000000000000000000000000000000000..64ac7281848f91302bc0aa3cb81dd198e56fb653 --- /dev/null +++ b/paddle/fluid/operators/distributed/request_handler.h @@ -0,0 +1,158 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" + +namespace paddle { +namespace operators { +namespace distributed { + +constexpr char kRequestSend[] = "RequestSend"; +constexpr char kRequestGet[] = "RequestGet"; +constexpr char kRequestPrefetch[] = "RequestPrefetch"; +constexpr char kRequestCheckpoint[] = "RequestCheckpoint"; +constexpr char kRequestPassBarrier[] = "RequestPassBarrier"; + +#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" +#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV" +#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV" +#define COMPLETE_MESSAGE "COMPLETE@RECV" + +#define CHECKPOINT_SAVE_MESSAGE "SAVE@CHECKPOINTNOTIFY" +#define CHECKPOINT_LOAD_MESSAGE "LOAD@CHECKPOINTNOTIFY" + +class RPCServer; + +struct VarHandle { + // RPC endpoint. + std::string ep; + const platform::DeviceContext* ctx; + const framework::Scope* scope; + // Variable name. + std::string name; + // RPC method name. + std::string method; + + std::string String() const { + std::ostringstream s; + s << method << " name:[" << name << "], ep:[" << ep << "]"; + return s.str(); + } +}; + +class RequestHandler { + public: + explicit RequestHandler(bool sync_mode) + : sync_mode_(sync_mode), + dev_ctx_(nullptr), + executor_(nullptr), + scope_(nullptr), + program_(nullptr), + rpc_server_(nullptr) {} + + virtual ~RequestHandler() {} + + // Set attributes. + void SetScope(framework::Scope* scope) { scope_ = scope; } + void SetDevCtx(const platform::DeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; } + void SetProgram(framework::ProgramDesc* program) { program_ = program; } + void SetExecutor(framework::Executor* executor) { executor_ = executor; } + + // Used for dist lookup table prefetch + void SetPrefetchPreparedCtx( + std::unordered_map< + std::string, std::shared_ptr>* g) { + prefetch_var_name_to_prepared_ctx_ = g; + } + + void SetCheckpointNotifyPreparedCtx( + std::shared_ptr g) { + checkpoint_prepared_ctx_ = g; + } + + // Used for async. + void SetGradToPreparedCtx( + std::unordered_map< + std::string, std::shared_ptr>* g) { + grad_to_prepared_ctx_ = g; + } + + void SetRPCServer(RPCServer* rpc_server) { rpc_server_ = rpc_server; } + + // Get attributes. + bool sync_mode() { return sync_mode_; } + framework::Scope* scope() { return scope_; } + const platform::DeviceContext* dev_ctx() { return dev_ctx_; } + framework::ProgramDesc* program() { return program_; } + framework::Executor* executor() { return executor_; } + + // This function processes user's rpc request. + // The implemention is in request_handler_impl. + // example: + // std::string varname = request_.varname(); + // + // auto scope = request_handler_->scope(); + // auto invar = scope->FindVar(varname); + // framework::Variable* outvar = nullptr; + // + // request_handler_->Handle(varname, scope, invar, &outvar); + // if (outvar) { + // SerializeToByteBuffer(varname, outvar, + // *request_handler_->dev_ctx(), &reply_); + // } + virtual bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") = 0; + + protected: + const bool sync_mode_; + + const platform::DeviceContext* dev_ctx_; + framework::Executor* executor_; + framework::Scope* scope_; + framework::ProgramDesc* program_; + + // used for distribute lookup table prefetch + std::unordered_map>* + prefetch_var_name_to_prepared_ctx_; + // used for checkpoint notify + std::shared_ptr checkpoint_prepared_ctx_; + + // Used for async. + std::unordered_map>* + grad_to_prepared_ctx_; + + RPCServer* rpc_server_; +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..de1a503154deb967eb4389a9f43b86c05626d966 --- /dev/null +++ b/paddle/fluid/operators/distributed/request_handler_impl.cc @@ -0,0 +1,144 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" +#include "paddle/fluid/string/printf.h" + +namespace paddle { +namespace operators { +namespace distributed { + +// define LOOKUP_TABLE_PATH for checkpoint notify to save lookup table variables +// to directory specified. +constexpr char LOOKUP_TABLE_PATH[] = "kLookupTablePath"; + +bool RequestSendHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + VLOG(4) << "RequestSendHandler:" << varname; + + // Async + if (!sync_mode_) { + rpc_server_->Profiler().OneStep(); + try { + executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), + scope); + } catch (std::exception& e) { + LOG(ERROR) << "async: run sub program error " << e.what(); + return false; + } + return true; + } + + // Sync + if (varname == BATCH_BARRIER_MESSAGE) { + VLOG(3) << "sync: recv BATCH_BARRIER_MESSAGE"; + rpc_server_->IncreaseBatchBarrier(kRequestSend); + } else if (varname == COMPLETE_MESSAGE) { + VLOG(3) << "sync: recv complete message"; + rpc_server_->Complete(); + } else { + VLOG(3) << "sync: received var_name: " << varname; + rpc_server_->WaitCond(kRequestSend); + VLOG(3) << "sync: processing received var: " << varname; + + if (invar == nullptr) { + LOG(FATAL) << "sync: Can not find server side var: " << varname; + return false; + } + if (invar->IsType()) { + std::unique_lock lock(mutex_sparse_vars_); + sparse_vars_.push_back(invar); + } + } + return true; +} + +void RequestSendHandler::ResetSparseVarRecorder() { + std::unique_lock lock(mutex_sparse_vars_); + for (auto* var : sparse_vars_) { + var->GetMutable()->mutable_rows()->clear(); + } + sparse_vars_.clear(); +} + +bool RequestGetHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + VLOG(4) << "RequestGetHandler:" << varname; + if (sync_mode_) { + if (varname == FETCH_BARRIER_MESSAGE) { + VLOG(3) << "sync: recv fetch barrier message"; + rpc_server_->IncreaseBatchBarrier(kRequestGet); + } else { + rpc_server_->WaitCond(kRequestGet); + *outvar = scope_->FindVar(varname); + } + } else { + if (varname != FETCH_BARRIER_MESSAGE && varname != COMPLETE_MESSAGE) { + *outvar = scope_->FindVar(varname); + } + } + return true; +} + +bool RequestPrefetchHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + VLOG(4) << "RequestPrefetchHandler " << varname; + + auto var_desc = program_->Block(0).FindVar(out_var_name); + InitializeVariable(*outvar, var_desc->GetType()); + executor_->RunPreparedContext( + (*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope); + + return true; +} + +bool RequestCheckpointHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + PADDLE_ENFORCE( + checkpoint_notify_id != -1, + "when checkpoint_notify_id = -1, there should be no RPC invoke."); + + auto* lt_var = scope->FindVar(LOOKUP_TABLE_PATH)->GetMutable(); + lt_var->clear(); + lt_var->append(out_var_name); + VLOG(4) << "RequestCheckpointHandler update var kLookupTablePath to: " + << out_var_name; + executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope); + return true; +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/request_handler_impl.h b/paddle/fluid/operators/distributed/request_handler_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..87185500f2ffc3a8578eea339cc7a1e2b0e46631 --- /dev/null +++ b/paddle/fluid/operators/distributed/request_handler_impl.h @@ -0,0 +1,86 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/distributed/request_handler.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class RequestSendHandler final : public RequestHandler { + public: + explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestSendHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; + void ResetSparseVarRecorder(); + + private: + std::mutex mutex_sparse_vars_; + std::vector sparse_vars_; +}; + +class RequestGetHandler final : public RequestHandler { + public: + explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestGetHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; +}; + +class RequestPrefetchHandler final : public RequestHandler { + public: + explicit RequestPrefetchHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestPrefetchHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; +}; + +class RequestCheckpointHandler final : public RequestHandler { + public: + explicit RequestCheckpointHandler(bool sync_mode, int checkpoint_notify_id) + : RequestHandler(sync_mode) { + this->checkpoint_notify_id = checkpoint_notify_id; + } + virtual ~RequestCheckpointHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; + + private: + int checkpoint_notify_id; +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/rpc_client.cc b/paddle/fluid/operators/distributed/rpc_client.cc new file mode 100644 index 0000000000000000000000000000000000000000..b5ec9fe5367beb97b3cc7298102deff1e8ca4ec9 --- /dev/null +++ b/paddle/fluid/operators/distributed/rpc_client.cc @@ -0,0 +1,30 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "gflags/gflags.h" + +// default to 3min to avoid temprary network failures. +DEFINE_int32(rpc_deadline, 180000, "deadline timeouts for rpc"); + +namespace paddle { +namespace operators { +namespace distributed { + +std::once_flag RPCClient::init_flag_; +std::unique_ptr RPCClient::rpc_client_(nullptr); + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/rpc_client.h b/paddle/fluid/operators/distributed/rpc_client.h new file mode 100644 index 0000000000000000000000000000000000000000..22a022a5d25e5c6628b80294494b87ca105a04c7 --- /dev/null +++ b/paddle/fluid/operators/distributed/rpc_client.h @@ -0,0 +1,97 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include "gflags/gflags.h" + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" + +DECLARE_int32(rpc_deadline); + +namespace paddle { +namespace operators { +namespace distributed { + +class RPCClient { + public: + RPCClient() {} + virtual ~RPCClient() {} + virtual bool AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + virtual bool AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + virtual bool AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + virtual void AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + virtual void AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + virtual void AsyncCheckpointNotify(const std::string& ep, + const std::string& dir, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + virtual void AsyncSendComplete(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline) = 0; + + // Complete tells all the pserver instances that finishe the training, + // the pserver can reduce it's barrier count, and continue to train + // with other trainers. + virtual void SendComplete() = 0; + + virtual bool Wait() = 0; + + template + static RPCClient* GetInstance() { + std::call_once(init_flag_, &RPCClient::Init); + return rpc_client_.get(); + } + + // Init is called by GetInstance. + template + static void Init() { + if (rpc_client_.get() == nullptr) { + rpc_client_.reset(new T()); + rpc_client_->InitImpl(); + } + } + + protected: + virtual void InitImpl() {} + + private: + static std::once_flag init_flag_; + static std::unique_ptr rpc_client_; +}; +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/rpc_server.cc b/paddle/fluid/operators/distributed/rpc_server.cc new file mode 100644 index 0000000000000000000000000000000000000000..406e7294c190172347d432fb155c2a81c43dda25 --- /dev/null +++ b/paddle/fluid/operators/distributed/rpc_server.cc @@ -0,0 +1,161 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include +#include + +#include "paddle/fluid/operators/distributed/rpc_server.h" +#include "paddle/fluid/platform/profiler.h" + +DEFINE_int32(rpc_server_profile_period, 0, + "the period of listen_and_serv to do profile"); +DEFINE_string(rpc_server_profile_path, "/dev/null", + "the profile log file path"); + +namespace paddle { +namespace operators { +namespace distributed { + +RPCServerProfiler::RPCServerProfiler(int profile_period, + const std::string& profile_log_path) + : profile_period_(profile_period), profile_log_path_(profile_log_path) { + step_ = 0; +} + +void RPCServerProfiler::OneStep() { + PADDLE_ENFORCE_LE(step_, profile_period_, + "step_ should not be larger then " + "profile_period_"); + if (profile_period_ <= 0) { + return; + } + + if (step_ == 0) { + auto pf_state = paddle::platform::ProfilerState::kCPU; + paddle::platform::EnableProfiler(pf_state); + } + if (step_ == profile_period_) { + paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kTotal, + profile_log_path_); + step_ = 0; + } else { + step_++; + } +} + +void RPCServer::ShutDown() { + LOG(INFO) << "RPCServer ShutDown "; + ShutDownImpl(); + + exit_flag_ = true; + barrier_cond_.notify_all(); + rpc_cond_.notify_all(); +} + +void RPCServer::SavePort() const { + auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); + std::ofstream port_file; + port_file.open(file_path); + port_file << selected_port_; + port_file.close(); + VLOG(4) << "selected port written to " << file_path; +} + +void RPCServer::WaitBarrier(const std::string& rpc_name) { + std::unique_lock lock(this->mutex_); + barrier_cond_.wait(lock, [this, &rpc_name] { + return ((barrier_counter_[rpc_name] == client_num_ && client_num_ != 0) || + exit_flag_.load()); + }); + + VLOG(3) << "batch_barrier_: " << rpc_name << " " + << barrier_counter_[rpc_name]; +} + +void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) { + VLOG(4) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; + int b = 0; + std::unique_lock lock(mutex_); + b = ++barrier_counter_[rpc_name]; + if (b >= client_num_) { + lock.unlock(); + barrier_cond_.notify_all(); + lock.lock(); + } +} + +void RPCServer::Complete() { + { + std::unique_lock lock(mutex_); + client_num_--; + VLOG(4) << "decrease client_num to: " << client_num_; + if (cur_cond_.load() == rpc_cond_map_[kRequestGet]) { + barrier_counter_[kRequestGet]--; + } + } + barrier_cond_.notify_all(); +} + +int RPCServer::GetClientNum() { + std::unique_lock lock(mutex_); + return client_num_; +} + +void RPCServer::ResetBarrierCounter() { + VLOG(3) << "RPCServer ResetBarrierCounter "; + std::unique_lock lock(mutex_); + for (auto& t : barrier_counter_) { + t.second = 0; + } +} + +void RPCServer::RegisterRPC(const std::string& rpc_name, + RequestHandler* handler, int thread_num) { + rpc_call_map_[rpc_name] = handler; + rpc_thread_num_[rpc_name] = thread_num; + + static int cond = -1; + rpc_cond_map_[rpc_name] = ++cond; + VLOG(4) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler + << ", cond:" << rpc_cond_map_[rpc_name]; +} + +void RPCServer::SetCond(const std::string& rpc_name) { + VLOG(3) << "RPCServer SetCond " << rpc_name; + { + std::unique_lock lock(mutex_); + cur_cond_ = rpc_cond_map_[rpc_name]; + } + + rpc_cond_.notify_all(); +} + +void RPCServer::WaitCond(const std::string& rpc_name) { + VLOG(4) << "RPCServer WaitCond " << rpc_name; + int cond = 0; + { + std::unique_lock lock(mutex_); + cond = rpc_cond_map_[rpc_name]; + } + + std::unique_lock lock(mutex_); + rpc_cond_.wait( + lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); }); +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/rpc_server.h b/paddle/fluid/operators/distributed/rpc_server.h new file mode 100644 index 0000000000000000000000000000000000000000..d813ba03e2fbec6e808f59f814a9b2f4bfbcd77b --- /dev/null +++ b/paddle/fluid/operators/distributed/rpc_server.h @@ -0,0 +1,115 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include // NOLINT +#include +#include + +#include "paddle/fluid/operators/distributed/request_handler.h" + +DECLARE_int32(rpc_server_profile_period); +DECLARE_string(rpc_server_profile_path); + +namespace paddle { +namespace operators { +namespace distributed { + +class RPCServerProfiler { + public: + RPCServerProfiler(int profile_period, const std::string& profile_log_path); + void OneStep(); + + private: + const int profile_period_; + std::string profile_log_path_; + int step_; +}; + +class RPCServer { + public: + explicit RPCServer(const std::string& address, int client_num) + : cur_cond_(0), + profiler_(FLAGS_rpc_server_profile_period, + FLAGS_rpc_server_profile_path), + bind_address_(address), + exit_flag_(false), + selected_port_(0), + client_num_(client_num) {} + + virtual ~RPCServer() {} + virtual void StartServer() = 0; + virtual void WaitServerReady() = 0; + + void ShutDown(); + + bool IsExit() { return exit_flag_.load(); } + + int GetSelectedPort() const { return selected_port_; } + + int GetClientNum(); + + void SavePort() const; + + // RegisterRPC, register the rpc method name to a handler + // class, and auto generate a condition id for this call + // to be used for the barrier. + void RegisterRPC(const std::string& rpc_name, RequestHandler* handler, + int thread_num = 5); + + // Wait util all the clients have reached the barrier for one + // rpc method. This function should be called in the + // RequestHandler if you want to run the server/client in a + // synchronous mode. + void WaitBarrier(const std::string& rpc_name); + + void SetCond(const std::string& rpc_name); + void WaitCond(const std::string& rpc_name); + void IncreaseBatchBarrier(const std::string rpc_name); + + void Complete(); + + void ResetBarrierCounter(); + RPCServerProfiler& Profiler() { return profiler_; } + + protected: + virtual void ShutDownImpl() = 0; + + private: + std::mutex mutex_; + std::unordered_map barrier_counter_; + std::condition_variable barrier_cond_; + + std::unordered_map rpc_cond_map_; + std::atomic cur_cond_; + std::condition_variable rpc_cond_; + RPCServerProfiler profiler_; + + protected: + std::string bind_address_; + std::atomic exit_flag_; + int selected_port_; + int client_num_; + + std::unordered_map rpc_call_map_; + std::unordered_map rpc_thread_num_; + friend class RequestHandler; +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/rpc_server_test.cc b/paddle/fluid/operators/distributed/rpc_server_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..b50830c362d3f6ecf38affbfa6a1ffe2ed77e125 --- /dev/null +++ b/paddle/fluid/operators/distributed/rpc_server_test.cc @@ -0,0 +1,184 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include // NOLINT + +#include "gtest/gtest.h" +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" + +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" + +namespace framework = paddle::framework; +namespace platform = paddle::platform; +namespace distributed = paddle::operators::distributed; + +USE_NO_KERNEL_OP(lookup_sparse_table); + +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; + +framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) { + auto root_block = program->MutableBlock(0); + auto* block = program->AppendBlock(*root_block); + + framework::VariableNameMap input({{"W", {"w"}}, {"Ids", {"ids"}}}); + framework::VariableNameMap output({{"Output", {"out"}}}); + auto op = block->AppendOp(); + op->SetType("lookup_sparse_table"); + op->SetInput("W", {"w"}); + op->SetInput("Ids", {"ids"}); + op->SetOutput("Out", {"out"}); + + auto& out = *root_block->Var("out"); + out.SetType(framework::proto::VarType::LOD_TENSOR); + out.SetShape({10, 10}); + + return block; +} + +void CreateVarsOnScope(framework::Scope* scope, platform::CPUPlace* place) { + auto w_var = scope->Var("w"); + w_var->GetMutable(); + + auto out_var = scope->Var("out"); + out_var->GetMutable(); + + auto ids_var = scope->Var("ids"); + ids_var->GetMutable(); +} + +void InitTensorsOnClient(framework::Scope* scope, platform::CPUPlace* place, + int64_t rows_numel) { + CreateVarsOnScope(scope, place); + auto ids_var = scope->Var("ids")->GetMutable(); + int64_t* ids_ptr = + ids_var->mutable_data(framework::DDim({rows_numel, 1}), *place); + for (int64_t i = 0; i < rows_numel; ++i) ids_ptr[i] = i * 2; +} + +void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place, + int64_t rows_numel) { + CreateVarsOnScope(scope, place); + auto w = scope->Var("w")->GetMutable(); + auto rows = w->mutable_rows(); + for (int64_t i = 0; i < rows_numel; ++i) rows->push_back(i); + auto w_value = w->mutable_value(); + w_value->Resize({rows_numel, 10}); + + auto ptr = w_value->mutable_data(*place); + + for (int64_t i = 0; i < w_value->numel(); ++i) { + ptr[i] = static_cast(i / 10); + } +} + +void StartServer(const std::string& rpc_name) { + framework::ProgramDesc program; + framework::Scope scope; + platform::CPUPlace place; + framework::Executor exe(place); + platform::CPUDeviceContext ctx(place); + auto* block = AppendPrefetchBlcok(&program); + std::string in_var_name("ids"); + std::vector prefetch_block_ids{block->ID()}; + auto prepared = exe.Prepare(program, prefetch_block_ids); + InitTensorsOnServer(&scope, &place, 10); + + std::unordered_map> + prefetch_var_name_to_prepared; + prefetch_var_name_to_prepared[in_var_name] = prepared[0]; + + g_req_handler->SetProgram(&program); + g_req_handler->SetPrefetchPreparedCtx(&prefetch_var_name_to_prepared); + g_req_handler->SetDevCtx(&ctx); + g_req_handler->SetScope(&scope); + g_req_handler->SetExecutor(&exe); + + g_rpc_service->RegisterRPC(rpc_name, g_req_handler.get()); + g_req_handler->SetRPCServer(g_rpc_service.get()); + + std::thread server_thread( + std::bind(&distributed::RPCServer::StartServer, g_rpc_service.get())); + + server_thread.join(); +} + +TEST(PREFETCH, CPU) { + g_req_handler.reset(new distributed::RequestPrefetchHandler(true)); + g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); + + std::thread server_thread(StartServer, distributed::kRequestPrefetch); + g_rpc_service->WaitServerReady(); + + int port = g_rpc_service->GetSelectedPort(); + std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port); + + framework::Scope scope; + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + { + // create var on local scope + int64_t rows_numel = 5; + InitTensorsOnClient(&scope, &place, rows_numel); + std::string in_var_name("ids"); + std::string out_var_name("out"); + + client->AsyncPrefetchVar(ep, ctx, scope, in_var_name, out_var_name); + client->Wait(); + auto var = scope.Var(out_var_name); + auto value = var->GetMutable(); + auto ptr = value->mutable_data(place); + + for (int64_t i = 0; i < rows_numel; ++i) { + EXPECT_EQ(ptr[0 + i * value->dims()[1]], static_cast(i * 2)); + } + } + + g_rpc_service->ShutDown(); + server_thread.join(); + LOG(INFO) << "begin reset"; + g_rpc_service.reset(nullptr); + g_req_handler.reset(nullptr); +} + +TEST(COMPLETE, CPU) { + g_req_handler.reset(new distributed::RequestSendHandler(true)); + g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 2)); + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); + PADDLE_ENFORCE(client != nullptr); + std::thread server_thread(StartServer, distributed::kRequestSend); + g_rpc_service->WaitServerReady(); + int port = g_rpc_service->GetSelectedPort(); + std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port); + client->AsyncSendComplete(ep); + client->Wait(); + + EXPECT_EQ(g_rpc_service->GetClientNum(), 1); + + g_rpc_service->ShutDown(); + server_thread.join(); + g_rpc_service.reset(nullptr); + g_req_handler.reset(nullptr); +} diff --git a/paddle/fluid/operators/distributed/send_recv.proto.in b/paddle/fluid/operators/distributed/send_recv.proto.in new file mode 100644 index 0000000000000000000000000000000000000000..8b0a09abe1d05dda10eda0030eb91cb9ca40683e --- /dev/null +++ b/paddle/fluid/operators/distributed/send_recv.proto.in @@ -0,0 +1,84 @@ + +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. Licensed under +the Apache License, Version 2.0 (the "License"); you may not use this file +except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +syntax = "proto3"; +package sendrecv; + +option cc_generic_services = @cc_generic_services@; + +service SendRecvService { + // For parameter server round-robin like hashing, do not split tensors. + // Send and recv only one tensor + // TODO(typhoonzero): add streaming API + rpc SendVariable(VariableMessage) returns (VoidMessage) {} + // Argument VariableMessage for GetVariable should only contain varname. + rpc GetVariable(VariableMessage) returns (VariableMessage) {} + // pre-fetch variable by given variable name and Ids + rpc PrefetchVariable(VariableMessage) returns (VariableMessage) {} + + rpc CheckpointNotify(VariableMessage) returns (VoidMessage) {} +} + +// VariableMessage is serialized paddle variable message. +// It can be: +// LoDTensor +// SelectedRows +enum VarType { + LOD_TENSOR = 0; + SELECTED_ROWS = 1; + NCCL_ID = 2; +} + +// NOTICE(gongwb):don't modify this proto if you are not +// not familar with how we serialize in sendrecvop_utils.h +// and deserilize it in variable_response.h. +message VariableMessage { + enum Type { + // Pod Types + BOOL = 0; + INT16 = 1; + INT32 = 2; + INT64 = 3; + FP16 = 4; + FP32 = 5; + FP64 = 6; + } + + message LodData { repeated int64 lod_data = 1; } + string varname = 1; + // TODO(Yancey1989): reference framework::proto::VarDesc::VarType + VarType type = 2; + // bool persistable is not needed for sending. + // tensor info: + Type data_type = 3; + repeated int64 dims = 4; + + // lod details: + int64 lod_level = 5; + repeated LodData lod = 6; + // selected_rows height, aka. original dim0 + int64 slr_height = 7; + // tensor data + bytes serialized = 8; + // selected_rows data + bytes rows = 9; + // Look up table block execution output variable name. + string out_varname = 10; + // If 1, the ps server will start profiling, the ps + // server stops profiling and generates a profile to /tmp/profile_ps_* + // when profile switches from 1 to 2. + int64 profile = 11; +} + +message VoidMessage {} diff --git a/paddle/fluid/operators/distributed/sendrecvop_utils.cc b/paddle/fluid/operators/distributed/sendrecvop_utils.cc new file mode 100644 index 0000000000000000000000000000000000000000..6a3f8fd544bc5d669b725765a863b42ec069a7b6 --- /dev/null +++ b/paddle/fluid/operators/distributed/sendrecvop_utils.cc @@ -0,0 +1,116 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +#include +#endif +#include +#include // NOLINT + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/variable_response.h" + +namespace paddle { +namespace operators { +namespace distributed { + +using VarMsg = sendrecv::VariableMessage; + +#ifdef PADDLE_WITH_CUDA +void* GetVarPayLoad(const std::string varname, int64_t size) { + platform::CUDAPinnedPlace cuda_pinned; + return memory::Alloc(cuda_pinned, size); +} +#endif + +void GetTensorPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size) { + auto tensor = var->Get(); + // FIXME(wuyi): data types in send_recv.proto is copied from + // framework.proto + request->set_data_type( + static_cast(framework::ToDataType(tensor.type()))); + for (auto& dim : framework::vectorize(tensor.dims())) { + request->add_dims(dim); + } + const framework::LoD lod = tensor.lod(); + if (lod.size() > 0) { + request->set_lod_level(lod.size()); + for (auto& each : lod) { + VarMsg::LodData* lod_inner = request->add_lod(); + for (auto& d : each) { + lod_inner->add_lod_data(d); + } + } + } + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); + // platform::CUDAPinnedPlace cuda_pinned; + auto& gpu_dev_ctx = static_cast(ctx); + auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type()); + *payload = GetVarPayLoad(request->varname(), copy_size); + + platform::CUDAPinnedPlace cuda_pinned; + memory::Copy(cuda_pinned, *payload, + boost::get(tensor.place()), + reinterpret_cast(tensor.data()), copy_size, + gpu_dev_ctx.stream()); + + ctx.Wait(); +#endif + } else { + *payload = tensor.data(); + } + *payload_size = tensor.numel() * framework::SizeOfType(tensor.type()); +} + +void GetSelectedRowsPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size) { + auto* slr = var->GetMutable(); + request->set_data_type( + static_cast(framework::ToDataType(slr->value().type()))); + request->set_lod_level(0); + request->set_slr_height(slr->height()); + + for (auto& dim : framework::vectorize(slr->value().dims())) { + request->add_dims(dim); + } + + auto* tensor = slr->mutable_value(); + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + auto& gpu_dev_ctx = static_cast(ctx); + auto copy_size = tensor->numel() * framework::SizeOfType(tensor->type()); + *payload = GetVarPayLoad(request->varname(), copy_size); + + platform::CUDAPinnedPlace cuda_pinned; + memory::Copy(cuda_pinned, *payload, + boost::get(tensor->place()), + reinterpret_cast(tensor->data()), copy_size, + gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + *payload = slr->mutable_value()->data(); + } + *payload_size = tensor->numel() * framework::SizeOfType(tensor->type()); +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/sendrecvop_utils.h b/paddle/fluid/operators/distributed/sendrecvop_utils.h new file mode 100644 index 0000000000000000000000000000000000000000..4d08d3c77afa3c1f2b4d7602f7199558bb5a79c0 --- /dev/null +++ b/paddle/fluid/operators/distributed/sendrecvop_utils.h @@ -0,0 +1,63 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/framework/var_type.h" + +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +namespace paddle { +namespace operators { +namespace distributed { + +using VarMsg = sendrecv::VariableMessage; + +void GetTensorPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size); + +void GetSelectedRowsPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size); + +inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) { + switch (type) { + case sendrecv::VariableMessage::FP32: + return typeid(float); // NOLINT + case sendrecv::VariableMessage::FP64: + return typeid(double); // NOLINT + case sendrecv::VariableMessage::INT32: + return typeid(int); // NOLINT + case sendrecv::VariableMessage::INT64: + return typeid(int64_t); // NOLINT + case sendrecv::VariableMessage::BOOL: + return typeid(bool); // NOLINT + default: + PADDLE_THROW("Not support type %d", type); + } +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/variable_response.cc b/paddle/fluid/operators/distributed/variable_response.cc new file mode 100644 index 0000000000000000000000000000000000000000..8e38b3713f28b045e9214db68aec50f0ba6c06f6 --- /dev/null +++ b/paddle/fluid/operators/distributed/variable_response.cc @@ -0,0 +1,219 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/distributed/variable_response.h" +#include +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" + +namespace paddle { +namespace operators { +namespace distributed { + +bool VariableResponse::ReadRaw(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& dev_ctx, + platform::Place place, void* dest, + int64_t size) { + const void* data = NULL; + int size_to_write = 0; + int64_t length = size; + int total_written = 0; + + if (platform::is_gpu_place(place)) { +#ifdef PADDLE_WITH_CUDA + auto& gpu_dev_ctx = + static_cast(dev_ctx); + platform::CPUPlace cpu; + + char* p = reinterpret_cast(dest); + while (total_written < length) { + if (!input->GetDirectBufferPointer(&data, &size_to_write)) { + return false; + } + // NOTE: if raw buffer is large and have two neighbor fields of raw + // buffers GetDirectBufferPointer can get all of them, use length to + // truncate it. + if (total_written + size_to_write > length) { + size_to_write = length - total_written; + } + // This log is useful to see how long a internal block size is of rpc. + VLOG(7) << "copy " << size_to_write << " data to CUDAPlace"; + memory::Copy(boost::get(place), + reinterpret_cast(p), cpu, data, size_to_write, + gpu_dev_ctx.stream()); + p += size_to_write; + total_written += size_to_write; + + input->Skip(size_to_write); + } + gpu_dev_ctx.Wait(); +#else + PADDLE_THROW("Unexpected branch"); +#endif + return true; + } + + char* p = reinterpret_cast(dest); + while (total_written < length) { + if (!input->GetDirectBufferPointer(&data, &size_to_write)) { + return false; + } + // NOTE: if raw buffer is large and have two neighbor fields of raw buffers + // GetDirectBufferPointer can get all of them, use length to truncate it. + if (total_written + size_to_write > length) { + size_to_write = length - total_written; + } + // TODO(gongwb): can we avoid copy? + platform::CPUPlace cpu; + // This log is useful to see how long a internal block size is of rpc. + VLOG(7) << "copy " << size_to_write << " data to CPUPlace"; + memory::Copy(cpu, reinterpret_cast(p), cpu, data, size_to_write); + + p += size_to_write; + total_written += size_to_write; + + input->Skip(size_to_write); + } + + return true; +} + +bool VariableResponse::CopyLodTensorData( + ::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, const framework::DDim& dims, + int length) { + auto* tensor = GetVar()->GetMutable(); + tensor->Resize(dims); + + framework::LoD lod; + for (int i = 0; i < meta_.lod_level(); ++i) { + framework::Vector v; + for (int j = 0; j < meta_.lod(i).lod_data_size(); ++j) { + v.push_back(meta_.lod(i).lod_data(j)); + } + lod.push_back(v); + } + tensor->set_lod(lod); + + void* tensor_data = + tensor->mutable_data(ctx.GetPlace(), ToTypeIndex(meta_.data_type())); + + if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { + return false; + } + + return true; +} + +inline framework::DDim GetDims( + const ::google::protobuf::RepeatedField<::google::protobuf::int64>& dims) { + std::vector vecdims; + for (auto& d : dims) { + vecdims.push_back(d); + } + return framework::make_ddim(vecdims); +} + +bool VariableResponse::CopySelectRowsTensorData( + ::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, const framework::DDim& dims, + int length) { + auto* slr = GetVar()->GetMutable(); + slr->set_height(meta_.slr_height()); + auto* tensor = slr->mutable_value(); + tensor->Resize(dims); + PADDLE_ENFORCE_EQ(static_cast(tensor->numel()), + length / framework::SizeOfType( + paddle::operators::distributed::ToTypeIndex( + meta_.data_type()))); + void* tensor_data = tensor->mutable_data( + ctx.GetPlace(), + paddle::operators::distributed::ToTypeIndex(meta_.data_type())); + + if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { + return false; + } + + return true; +} + +bool VariableResponse::CopySelectRowsData( + ::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, int length) { + auto* slr = GetVar()->GetMutable(); + slr->mutable_rows()->resize(length / + framework::SizeOfType(typeid(int64_t))); // int64 + int64_t* rows_data = slr->mutable_rows()->data(); + + // copy rows CPU data, GPU data will be copied lazily. + platform::CPUPlace cpu; + if (!ReadRaw(input, ctx, cpu, rows_data, length)) { + return false; + } + + return true; +} + +bool VariableResponse::ProcSerializedField( + int tag, ::google::protobuf::io::CodedInputStream* input, + int64_t num_bytes) { + PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || + meta_.type() == sendrecv::LOD_TENSOR || + meta_.type() == sendrecv::NCCL_ID) && + meta_.varname() != "", + "meta info should be got first!"); + + if (meta_.type() == sendrecv::NCCL_ID) { +#ifdef PADDLE_WITH_CUDA + auto* var = scope_->FindVar(meta_.varname()); + if (var != nullptr) { + ncclUniqueId* id = var->GetMutable(); + if (!ReadRaw(input, *dev_ctx_, platform::CPUPlace(), id->internal, + num_bytes)) { + return false; + } + } + return true; +#else + PADDLE_THROW("Not compiled with CUDA!"); + return false; +#endif + } + + VLOG(7) << "ProcSerializedField:" << meta_.varname() + << ", type:" << meta_.type() << std::endl; + framework::DDim dims = GetDims(meta_.dims()); + if (meta_.type() == sendrecv::LOD_TENSOR) { + PADDLE_ENFORCE(meta_.lod_size() >= 0, "lod info should be got first!"); + if (!CopyLodTensorData(input, *dev_ctx_, dims, num_bytes)) { + return false; + } + + return true; + } + + if (meta_.type() == sendrecv::SELECTED_ROWS) { + if (!CopySelectRowsTensorData(input, *dev_ctx_, dims, num_bytes)) { + return false; + } + return true; + } + + PADDLE_ENFORCE("not supported var types:", meta_.varname(), meta_.type()); + + return false; +} + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/variable_response.h b/paddle/fluid/operators/distributed/variable_response.h new file mode 100644 index 0000000000000000000000000000000000000000..6aec52ca00f59a42ecca01da8df1680ce4eda432 --- /dev/null +++ b/paddle/fluid/operators/distributed/variable_response.h @@ -0,0 +1,126 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" + +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +namespace paddle { +namespace operators { +namespace distributed { + +// Source provides a way for a particular RPC implementation to provide +// received data to ParseFrom. +class Source { + public: + virtual ~Source() {} + + // Return the stream that contains the data to be parsed. + // Note that this method might be invoked more than once if + // ParseFrom needs to fall back to a more expensive parsing method. + // Every call must return a stream pointing at the beginning of + // the serialized RecvTensorResponse. + // + // Note that a subsequent call to contents() invalidates previous + // results of contents(). + // + // Ownership of the returned stream is retained by the Source and + // should not be deleted by the caller. + virtual ::google::protobuf::io::ZeroCopyInputStream* contents() = 0; +}; + +class VariableResponse { + public: + VariableResponse(const framework::Scope* scope, + const platform::DeviceContext* dev_ctx, + bool create_scope = false) + : scope_(scope), dev_ctx_(dev_ctx), create_scope_(create_scope) { + if (create_scope) { + local_scope_ = &scope->NewScope(); + } + } + + virtual ~VariableResponse() { + if (create_scope_) { + scope_->DeleteScope(local_scope_); + } + } + + int Parse(Source* source, const sendrecv::VariableMessage& meta) { + meta_ = meta; + return Parse(source); + } + + // return: + // 0:ok. + // -1: unkown error. + // other: number of error field. + virtual int Parse(Source* source) = 0; + + inline const framework::Scope& GetLocalScope() const { return *local_scope_; } + inline framework::Scope* GetMutableLocalScope() const { return local_scope_; } + inline std::string Varname() const { return meta_.varname(); } + inline std::string OutVarname() const { return meta_.out_varname(); } + + // should call parse first. + framework::Variable* GetVar() { + if (create_scope_) { + return local_scope_->Var(meta_.varname()); + } + return scope_->FindVar(meta_.varname()); + } + + protected: + bool ReadRaw(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& dev_ctx, platform::Place place, + void* dest, int64_t size); + + bool CopySelectRowsTensorData(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, + const framework::DDim& dims, int length); + + bool CopySelectRowsData(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, int length); + + bool CopyLodTensorData(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, + const framework::DDim& dims, int length); + + bool ProcSerializedField(int tag, + ::google::protobuf::io::CodedInputStream* input, + int64_t num_bytes); + + protected: + const framework::Scope* scope_; + const platform::DeviceContext* dev_ctx_; + bool create_scope_ = false; + framework::Scope* local_scope_ = nullptr; + + sendrecv::VariableMessage meta_; +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/elementwise_add_mkldnn_op.cc b/paddle/fluid/operators/elementwise_add_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c86cd57316078778e5930c9b524b931d523028d7 --- /dev/null +++ b/paddle/fluid/operators/elementwise_add_mkldnn_op.cc @@ -0,0 +1,193 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/memory/memcpy.h" +#include "paddle/fluid/operators/elementwise_add_op.h" +#include "paddle/fluid/operators/elementwise_op_function.h" + +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using framework::DataLayout; +using framework::Tensor; +using mkldnn::memory; +using mkldnn::reorder; +using mkldnn::primitive; +using mkldnn::stream; +using mkldnn::sum; + +template +class EltwiseAddMKLDNNKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& dev_ctx = + ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); + const T* x_data = x->data(); + const T* y_data = y->data(); + T* z_data = z->mutable_data(ctx.GetPlace()); + + int axis = ctx.Attr("axis"); + + auto x_dims = x->dims(); + auto y_dims_untrimed = y->dims(); + auto z_dims = z->dims(); + + // Execute default elementwise_add operator when + // broadcast operations need to performed. + if (x_dims != y_dims_untrimed) { + auto sum_func = [](T a, T b) -> T { return a + b; }; + + TransformFunctor + functor( + x, y, z, + ctx.template device_context(), + sum_func); + + axis = (axis == -1 ? x_dims.size() - y_dims_untrimed.size() : axis); + PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), + "Axis should be in range [0, x_dims)"); + + auto y_dims = trim_trailing_singular_dims(y_dims_untrimed); + axis = (y_dims.size() == 0) ? x_dims.size() : axis; + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post); + + if (post == 1) { + functor.RunRowWise(n, pre); + } else { + functor.RunMidWise(n, pre, post); + } + z->set_layout(DataLayout::kMKLDNN); + z->set_format(x->format()); + } else { + PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN && + x->format() != memory::format::format_undef, + "Wrong layout/format set for X tensor"); + PADDLE_ENFORCE(y->layout() == DataLayout::kMKLDNN && + y->format() != memory::format::format_undef, + "Wrong layout/format set for Y tensor"); + + std::vector src_x_tz = framework::vectorize2int(x_dims); + std::vector src_y_tz = framework::vectorize2int(y_dims_untrimed); + std::vector dst_tz = framework::vectorize2int(z_dims); + + std::vector srcs_pd; + std::vector srcs; + std::vector scales = {1.0f, 1.0f}; + + auto src_x_pd = memory::primitive_desc( + {{src_x_tz}, memory::data_type::f32, x->format()}, mkldnn_engine); + auto src_y_pd = memory::primitive_desc( + {{src_y_tz}, memory::data_type::f32, y->format()}, mkldnn_engine); + auto src_x_memory = + memory(src_x_pd, paddle::platform::to_void_cast(x_data)); + auto src_y_memory = + memory(src_y_pd, paddle::platform::to_void_cast(y_data)); + + srcs_pd.push_back(src_x_pd); + srcs_pd.push_back(src_y_pd); + srcs.push_back(src_x_memory); + srcs.push_back(src_y_memory); + + auto dst_md = + memory::desc({dst_tz}, memory::data_type::f32, memory::format::any); + + // create primitive descriptor for sum + auto sum_pd = sum::primitive_desc(dst_md, scales, srcs_pd); + + // create mkldnn memory for dst + memory dst_memory = memory(sum_pd.dst_primitive_desc(), z_data); + + std::vector inputs; + inputs.push_back(srcs[0]); + inputs.push_back(srcs[1]); + + // create sum primitive + auto sum_prim = sum(sum_pd, inputs, dst_memory); + + std::vector pipeline; + pipeline.push_back(sum_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + z->set_layout(DataLayout::kMKLDNN); + z->set_format( + (memory::format)dst_memory.get_primitive_desc().desc().data.format); + } + } +}; + +template +class EltwiseAddMKLDNNGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + using Tensor = framework::Tensor; + + auto* dout = ctx.Input(framework::GradVarName("Out")); + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Output(framework::GradVarName("Y")); + int axis = ctx.Attr("axis"); + // skip out, x, y, + // dout length is larger or equal than dx, dy. + auto* out = dout; + auto *x = dout, *y = dout; + + auto set_mkldnn_format = [](Tensor* in, const Tensor* out) { + in->set_layout(DataLayout::kMKLDNN); + in->set_format(out->format()); + }; + + if (dx != nullptr && dy != nullptr && dx->dims() == dy->dims()) { + if (dx->dims() == dy->dims()) { + auto blas = math::GetBlas(ctx); + if (dx) { + blas.VCOPY(dout->numel(), dout->data(), + dx->mutable_data(ctx.GetPlace())); + set_mkldnn_format(dx, dout); + } + + if (dy) { + blas.VCOPY(dout->numel(), dout->data(), + dy->mutable_data(ctx.GetPlace())); + set_mkldnn_format(dy, dout); + } + } + } else { + // Execute default kernel when broadcast is needed + ElemwiseExplicitGradCompute, IdentityGrad>( + ctx, *x, *y, *out, *dout, axis, dx, dy, IdentityGrad(), + IdentityGrad()); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(elementwise_add, MKLDNN, ::paddle::platform::CPUPlace, + ops::EltwiseAddMKLDNNKernel) + +REGISTER_OP_KERNEL(elementwise_add_grad, MKLDNN, ::paddle::platform::CPUPlace, + ops::EltwiseAddMKLDNNGradKernel) diff --git a/paddle/fluid/operators/elementwise_add_op.cc b/paddle/fluid/operators/elementwise_add_op.cc index d2c20537136fc3ac9d1bece24a2238f26215c922..3c97ac995c649ecd0d196a584240e1e7ac04f08e 100644 --- a/paddle/fluid/operators/elementwise_add_op.cc +++ b/paddle/fluid/operators/elementwise_add_op.cc @@ -15,7 +15,9 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_add_op.h" #include "paddle/fluid/operators/elementwise_op.h" namespace ops = paddle::operators; -REGISTER_ELEMWISE_OP(elementwise_add, "Add", "Out = X + Y"); +REGISTER_ELEMWISE_GRAD_MAKER(elementwise_add, Add); +REGISTER_ELEMWISE_EXPLICIT_OP(elementwise_add, "Add", "Out = X + Y", "Out", + "X"); REGISTER_OP_CPU_KERNEL( elementwise_add, ops::ElementwiseAddKernel, diff --git a/paddle/fluid/operators/elementwise_add_op.h b/paddle/fluid/operators/elementwise_add_op.h index baf04c30b17cb333fc8a6544afd6c479442f835b..5356105e2e551c0528694091608fc7585dce66d2 100644 --- a/paddle/fluid/operators/elementwise_add_op.h +++ b/paddle/fluid/operators/elementwise_add_op.h @@ -95,9 +95,10 @@ void default_elementwise_add_grad(const framework::ExecutionContext& ctx, framework::Tensor* dy) { int axis = ctx.Attr("axis"); - ElemwiseGradCompute, IdentityGrad>( - ctx, *x, *y, *out, *dout, axis, dx, dy, IdentityGrad(), - IdentityGrad()); + ElemwiseExplicitGradCompute, + IdentityGrad>(ctx, *x, *y, *out, *dout, axis, + dx, dy, IdentityGrad(), + IdentityGrad()); } template @@ -140,14 +141,15 @@ class ElementwiseAddGradKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { using Tensor = framework::Tensor; - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* out = ctx.Input("Out"); auto* dout = ctx.Input(framework::GradVarName("Out")); auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); + // skip out, x, y + auto* out = dout; + auto *x = dout, *y = dout; - if (platform::is_cpu_place(ctx.GetPlace()) && (x->dims() == y->dims())) { + if (platform::is_cpu_place(ctx.GetPlace()) && dx != nullptr && + dy != nullptr && (dx->dims() == dy->dims())) { elementwise_add_grad(ctx, x, y, out, dout, dx, dy); } else { default_elementwise_add_grad(ctx, x, y, out, dout, dx, diff --git a/paddle/fluid/operators/elementwise_div_op.cc b/paddle/fluid/operators/elementwise_div_op.cc index 824b1221e5a77c8799dc34820b7f0db180c2439e..84c8a65e5f859d276ae6d5f1a3f25c9d713a7a61 100644 --- a/paddle/fluid/operators/elementwise_div_op.cc +++ b/paddle/fluid/operators/elementwise_div_op.cc @@ -15,7 +15,9 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_div_op.h" #include "paddle/fluid/operators/elementwise_op.h" namespace ops = paddle::operators; + REGISTER_ELEMWISE_OP(elementwise_div, "Div", "Out = X / Y"); + REGISTER_OP_CPU_KERNEL( elementwise_div, ops::ElementwiseDivKernel, diff --git a/paddle/fluid/operators/elementwise_mul_op.cc b/paddle/fluid/operators/elementwise_mul_op.cc index ba343909bb87b4f2efa56c0a4ff664b278e90c60..7cd67e74de6b9c4fbc718f60b4f671ccab2f9956 100644 --- a/paddle/fluid/operators/elementwise_mul_op.cc +++ b/paddle/fluid/operators/elementwise_mul_op.cc @@ -15,7 +15,7 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_mul_op.h" #include "paddle/fluid/operators/elementwise_op.h" namespace ops = paddle::operators; -REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\odot\\ Y"); +REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\\\odot Y"); REGISTER_OP_CPU_KERNEL( elementwise_mul, ops::ElementwiseMulKernel, diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index f4cec8ad971abebe8d6dff1a384c8414269148a5..d8a12e800ad733800c1ec333f15d31d4dcd1a3a5 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -14,8 +14,12 @@ limitations under the License. */ #pragma once #include +#include "paddle/fluid/framework/data_layout.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace operators { @@ -40,6 +44,21 @@ class ElementwiseOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", x_dim); ctx->ShareLoD("X", /*->*/ "Out"); } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("X")->type()); + +#ifdef PADDLE_WITH_MKLDNN + if (platform::CanMKLDNNBeUsed(ctx)) { + return framework::OpKernelType(input_data_type, ctx.GetPlace(), + framework::DataLayout::kMKLDNN, + framework::LibraryType::kMKLDNN); + } +#endif + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class ElementwiseOpInferVarType : public framework::VarTypeInference { @@ -59,55 +78,62 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { void Make() final { AddInput("X", "(Tensor), The first input tensor of elementwise op."); AddInput("Y", "(Tensor), The second input tensor of elementwise op."); + // AddOutput("SavedShape", "(Tensor), save X, Y shape for grad to save + // memory.").AsIntermediate(); AddOutput("Out", "The output of elementwise op."); AddAttr("axis", "(int, default -1). The start dimension index " "for broadcasting Y onto X.") .SetDefault(-1) .EqualGreaterThan(-1); + AddAttr("use_mkldnn", "(bool, default false). Used by MKLDNN.") + .SetDefault(false); AddComment(string::Sprintf(R"DOC( -Limited Elementwise %s Operator. +Limited Elementwise %s Operator The equation is: $$%s$$ -$X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be -smaller than or equal to the dimensions of $X$. +- $X$: a tensor of any dimension. +- $Y$: a tensor whose dimensions must be less than or equal to the dimensions of $X$. There are two cases for this operator: -1. The shape of $Y$ is same with $X$; -2. The shape of $Y$ is a congiguous subsequencet of $X$. The trailing dimensions - of size 1 for $Y$ will be ignored for the consideration of subsequence. +1. The shape of $Y$ is the same with $X$. +2. The shape of $Y$ is a continuous subsequence of $X$. For case 2: -$Y$ will be broadcasted to match the shape of $X$ and axis should be -set to index of the start dimension to broadcast $Y$ onto $X$. +1. Broadcast $Y$ to match the shape of $X$, where $axis$ is the start dimension index + for broadcasting $Y$ onto $X$. +2. If $axis$ is -1 (default), $axis = rank(X) - rank(Y)$. +3. The trailing dimensions of size 1 for $Y$ will be ignored for the consideration of + subsequence, such as shape(Y) = (2, 1) => (2). -If axis is -1, it is treated as axis=rank(X)-rank(Y). +For example: -For example .. code-block:: python shape(X) = (2, 3, 4, 5), shape(Y) = (,) shape(X) = (2, 3, 4, 5), shape(Y) = (5,) - shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5) + shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5), with axis=-1(default) or axis=2 shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0 shape(X) = (2, 3, 4, 5), shape(Y) = (2, 1), with axis=0 -Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details) -information. However, the output only shares the LoD information with input $X$. +The inputs $X$ and $Y$ can carry the different LoD information. +But the output only shares the LoD information with the input $X$. )DOC", GetName(), GetEquation())); + SetReuse(); } protected: virtual std::string GetName() const = 0; virtual std::string GetEquation() const = 0; + virtual void SetReuse() {} }; class ElementwiseOpGrad : public framework::OperatorWithKernel { @@ -137,10 +163,74 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(y_grad_name, y_dims); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto input_data_type = framework::ToDataType( + ctx.Input(framework::GradVarName("Out"))->type()); + +#ifdef PADDLE_WITH_MKLDNN + if (platform::CanMKLDNNBeUsed(ctx)) { + return framework::OpKernelType(input_data_type, ctx.GetPlace(), + framework::DataLayout::kMKLDNN, + framework::LibraryType::kMKLDNN); + } +#endif + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } +}; + +// For Add, Sub op, the X, Out is not needed. +class ElementwiseOpExplicitGrad : public ElementwiseOpGrad { + public: + using operators::ElementwiseOpGrad::ElementwiseOpGrad; + using operators::ElementwiseOpGrad::GetExpectedKernelType; + using Tensor = framework::Tensor; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); + ctx->SetOutputDim(x_grad_name, out_dims); + } + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(y_grad_name)) { + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); + auto y_dims = ctx->GetInputDim("Y"); + ctx->SetOutputDim(y_grad_name, y_dims); + } + } }; + } // namespace operators } // namespace paddle +/* +*/ + +#define REGISTER_ELEMWISE_GRAD_MAKER(kernel_type, op_name) \ + class kernel_type##GradMaker \ + : public paddle::framework::SingleGradOpDescMaker { \ + public: \ + using ::paddle::framework::SingleGradOpDescMaker::SingleGradOpDescMaker; \ + \ + protected: \ + std::unique_ptr Apply() const override { \ + auto* op = new paddle::framework::OpDesc(); \ + op->SetType(#kernel_type "_grad"); \ + op->SetInput("Y", Input("Y")); \ + op->SetInput(::paddle::framework::GradVarName("Out"), \ + OutputGrad("Out")); \ + op->SetAttrMap(Attrs()); \ + op->SetOutput(::paddle::framework::GradVarName("X"), InputGrad("X")); \ + op->SetOutput(::paddle::framework::GradVarName("Y"), InputGrad("Y")); \ + return std::unique_ptr<::paddle::framework::OpDesc>(op); \ + } \ + } + #define REGISTER_ELEMWISE_OP(op_type, op_name, equation) \ class __ElemwiseOp##op_type##Maker__ \ : public ::paddle::operators::ElementwiseOpMaker { \ @@ -153,3 +243,18 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { ::paddle::operators::ElementwiseOpInferVarType, \ ::paddle::framework::DefaultGradOpDescMaker); \ REGISTER_OPERATOR(op_type##_grad, ::paddle::operators::ElementwiseOpGrad) + +#define REGISTER_ELEMWISE_EXPLICIT_OP(op_type, op_name, equation, ...) \ + class __ElemwiseOp##op_type##Maker__ \ + : public ::paddle::operators::ElementwiseOpMaker { \ + protected: \ + virtual std::string GetName() const { return op_name; } \ + virtual std::string GetEquation() const { return equation; } \ + virtual void SetReuse() { Reuse(__VA_ARGS__); } \ + }; \ + REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp, \ + __ElemwiseOp##op_type##Maker__, \ + ::paddle::operators::ElementwiseOpInferVarType, \ + op_type##GradMaker); \ + REGISTER_OPERATOR(op_type##_grad, \ + ::paddle::operators::ElementwiseOpExplicitGrad) diff --git a/paddle/fluid/operators/elementwise_op_function.h b/paddle/fluid/operators/elementwise_op_function.h index 8b052611f80ddf874ca48c1c58e13346528a834e..bc3e95e904f8b6c2cdd2ae6685bf67580178e6b6 100644 --- a/paddle/fluid/operators/elementwise_op_function.h +++ b/paddle/fluid/operators/elementwise_op_function.h @@ -13,7 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" @@ -65,17 +67,21 @@ inline void get_mid_dims(const framework::DDim& x_dims, } } -inline void trim_trailing_singular_dims(framework::DDim* dims) { +inline framework::DDim trim_trailing_singular_dims( + const framework::DDim& dims) { // Remove trailing dimensions of size 1 for y - auto actual_dims_size = dims->size(); + auto actual_dims_size = dims.size(); for (; actual_dims_size != 0; --actual_dims_size) { - if ((*dims)[actual_dims_size - 1] != 1) break; + if (dims[actual_dims_size - 1] != 1) break; } - if (actual_dims_size != dims->size()) { - auto actual_dims = framework::vectorize(*dims); - actual_dims.resize(actual_dims_size); - *dims = framework::make_ddim(actual_dims); + + std::vector trim_dims; + trim_dims.resize(actual_dims_size); + for (int i = 0; i < actual_dims_size; ++i) { + trim_dims[i] = dims[i]; } + framework::DDim actual_dims = framework::make_ddim(trim_dims); + return actual_dims; } template @@ -456,6 +462,71 @@ static void ElemwiseGradBroadcast2CUDA(cudaStream_t stream, const T* x, #endif +template +void ElemwiseGradComputeNoBroadcast( + const framework::ExecutionContext& ctx, const framework::DDim& x_dim, + const framework::DDim& y_dim, const framework::Tensor& x, + const framework::Tensor& y, const framework::Tensor& out, + const framework::Tensor& dout, int axis, framework::Tensor* dx, + framework::Tensor* dy, DX_OP dx_op, DY_OP dy_op) { + size_t N = static_cast(framework::product(x_dim)); + platform::ForRange for_range( + ctx.template device_context(), N); + for_range(ElemwiseGradNoBroadcast{ + x.data(), y.data(), out.data(), dout.data(), dx_op, dy_op, + dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), + dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())}); +} + +template +void ElemwiseGradComputeWithBroadcast( + const framework::ExecutionContext& ctx, const framework::DDim& x_dim, + const framework::DDim& y_dim_untrimed, const framework::Tensor& x, + const framework::Tensor& y, const framework::Tensor& out, + const framework::Tensor& dout, int axis, framework::Tensor* dx, + framework::Tensor* dy, DX_OP dx_op, DY_OP dy_op) { + axis = (axis == -1 ? x_dim.size() - y_dim_untrimed.size() : axis); + auto y_dim = trim_trailing_singular_dims(y_dim_untrimed); + axis = (y_dim.size() == 0) ? x_dim.size() : axis; + + int pre, n, post; + get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post); + if (post == 1) { + int h = pre; + int w = n; + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef __NVCC__ + ElemwiseGradBroadcast1CUDA( + ctx.template device_context().stream(), x.data(), + y.data(), out.data(), dout.data(), h, w, dx_op, dy_op, + dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), + dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); +#endif + } else { + ElemwiseGradBroadcast1CPU( + x.data(), y.data(), out.data(), dout.data(), h, w, dx_op, + dy_op, dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), + dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); + } + } else { + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef __NVCC__ + ElemwiseGradBroadcast2CUDA( + ctx.template device_context().stream(), x.data(), + y.data(), out.data(), dout.data(), pre, n, post, dx_op, + dy_op, dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), + dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); +#endif + } else { + ElemwiseGradBroadcast2CPU( + x.data(), y.data(), out.data(), dout.data(), pre, n, post, + dx_op, dy_op, + dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), + dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); + } + } +} + template void ElemwiseGradCompute(const framework::ExecutionContext& ctx, const framework::Tensor& x, const framework::Tensor& y, @@ -463,63 +534,50 @@ void ElemwiseGradCompute(const framework::ExecutionContext& ctx, const framework::Tensor& dout, int axis, framework::Tensor* dx, framework::Tensor* dy, DX_OP dx_op, DY_OP dy_op) { + const framework::DDim& x_dim = x.dims(); + const framework::DDim& y_dim = y.dims(); if (x.dims() == y.dims()) { - size_t N = static_cast(framework::product(x.dims())); - platform::ForRange for_range( - ctx.template device_context(), N); - for_range(ElemwiseGradNoBroadcast{ - x.data(), y.data(), out.data(), dout.data(), dx_op, dy_op, - dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), - dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())}); + ElemwiseGradComputeNoBroadcast( + ctx, x_dim, y_dim, x, y, out, dout, axis, dx, dy, dx_op, dy_op); } else { // Y is a scalar - auto x_dim = x.dims(); - auto y_dim = y.dims(); - - axis = (axis == -1 ? x_dim.size() - y_dim.size() : axis); - trim_trailing_singular_dims(&y_dim); - axis = (y_dim.size() == 0) ? x_dim.size() : axis; - - int pre, n, post; - get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post); - if (post == 1) { - int h = pre; - int w = n; - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef __NVCC__ - ElemwiseGradBroadcast1CUDA( - ctx.template device_context().stream(), x.data(), - y.data(), out.data(), dout.data(), h, w, dx_op, dy_op, - dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), - dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); -#endif - } else { - ElemwiseGradBroadcast1CPU( - x.data(), y.data(), out.data(), dout.data(), h, w, - dx_op, dy_op, - dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), - dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); - } - } else { - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef __NVCC__ - ElemwiseGradBroadcast2CUDA( - ctx.template device_context().stream(), x.data(), - y.data(), out.data(), dout.data(), pre, n, post, dx_op, - dy_op, - dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), - dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); -#endif - } else { - ElemwiseGradBroadcast2CPU( - x.data(), y.data(), out.data(), dout.data(), pre, n, - post, dx_op, dy_op, - dx == nullptr ? nullptr : dx->mutable_data(ctx.GetPlace()), - dy == nullptr ? nullptr : dy->mutable_data(ctx.GetPlace())); - } + ElemwiseGradComputeWithBroadcast( + ctx, x_dim, y_dim, x, y, out, dout, axis, dx, dy, dx_op, dy_op); + } +} + +// NOTE(dzhwinter): Only used in elementwise_add, elementwise_sub. +// explicit gradient can cut off X, Y, Out from gradient op +// In elementwise_add, elementwise_sub, we use dout as fake X, Y, Out to reuse +// elementwise code. +template +void ElemwiseExplicitGradCompute(const framework::ExecutionContext& ctx, + const framework::Tensor& x, + const framework::Tensor& y, + const framework::Tensor& out, + const framework::Tensor& dout, int axis, + framework::Tensor* dx, framework::Tensor* dy, + DX_OP dx_op, DY_OP dy_op) { + if (dy == nullptr) { + const framework::DDim& dx_dims = dout.dims(); + auto dy_dims = dx_dims; + ElemwiseGradComputeNoBroadcast( + ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op); + } else { + if (dout.dims() == dy->dims()) { + const framework::DDim& dx_dims = dout.dims(); + const framework::DDim& dy_dims = dy->dims(); + ElemwiseGradComputeNoBroadcast( + ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op); + } else { // Y is a scalar + auto dx_dims = dout.dims(); + const framework::DDim& dy_dims = dy->dims(); + ElemwiseGradComputeWithBroadcast( + ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op); } } } +// Deprecated template void ElementwiseGradCompute(const framework::ExecutionContext& ctx, @@ -547,7 +605,7 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx, } axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); - trim_trailing_singular_dims(&y_dims); + trim_trailing_singular_dims(y_dims); axis = (y_dims.size() == 0) ? x_dims.size() : axis; int pre, n, post; @@ -574,19 +632,19 @@ void ElementwiseComputeEx(const framework::ExecutionContext& ctx, x, y, z, ctx.template device_context(), func); auto x_dims = x->dims(); - auto y_dims = y->dims(); - PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + auto y_dims_untrimed = y->dims(); + PADDLE_ENFORCE_GE(x_dims.size(), y_dims_untrimed.size(), "Rank of first input must >= rank of second input."); - if (x_dims == y_dims) { + if (x_dims == y_dims_untrimed) { functor.Run(); return; } - axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + axis = (axis == -1 ? x_dims.size() - y_dims_untrimed.size() : axis); PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), "Axis should be in range [0, x_dims)"); - trim_trailing_singular_dims(&y_dims); + auto y_dims = trim_trailing_singular_dims(y_dims_untrimed); axis = (y_dims.size() == 0) ? x_dims.size() : axis; int pre, n, post; diff --git a/paddle/fluid/operators/elementwise_sub_op.cc b/paddle/fluid/operators/elementwise_sub_op.cc index a7562b166b373ee2a8c9b6f379431d88d3e45fcb..b7224261e6a7ca82dff92a25f5fe8818c08e676d 100644 --- a/paddle/fluid/operators/elementwise_sub_op.cc +++ b/paddle/fluid/operators/elementwise_sub_op.cc @@ -15,7 +15,10 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_sub_op.h" #include "paddle/fluid/operators/elementwise_op.h" namespace ops = paddle::operators; -REGISTER_ELEMWISE_OP(elementwise_sub, "Sub", "Out = X - Y"); +REGISTER_ELEMWISE_GRAD_MAKER(elementwise_sub, Sub); +REGISTER_ELEMWISE_EXPLICIT_OP(elementwise_sub, "Sub", "Out = X - Y", "Out", + "X"); + REGISTER_OP_CPU_KERNEL( elementwise_sub, ops::ElementwiseSubKernel, diff --git a/paddle/fluid/operators/elementwise_sub_op.h b/paddle/fluid/operators/elementwise_sub_op.h index fe088b8203722a43b9aba7be3878b8f4ca68ba12..11c7e3fe628001f095836a788f2bcc7c4ee7ad4b 100644 --- a/paddle/fluid/operators/elementwise_sub_op.h +++ b/paddle/fluid/operators/elementwise_sub_op.h @@ -4,7 +4,7 @@ Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, @@ -55,14 +55,15 @@ class ElementwiseSubGradKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { using Tensor = framework::Tensor; - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* out = ctx.Input("Out"); auto* dout = ctx.Input(framework::GradVarName("Out")); auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); int axis = ctx.Attr("axis"); - ElemwiseGradCompute, SubGradDY>( + // skip out, x, y + auto* out = dout; + auto *x = dout, *y = dout; + + ElemwiseExplicitGradCompute, SubGradDY>( ctx, *x, *y, *out, *dout, axis, dx, dy, SubGradDX(), SubGradDY()); } }; diff --git a/paddle/fluid/operators/extract_rows_op.cc b/paddle/fluid/operators/extract_rows_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..9a297d03cfb041e584159a5fc5ba214f8ac404b4 --- /dev/null +++ b/paddle/fluid/operators/extract_rows_op.cc @@ -0,0 +1,103 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class ExtractRowsOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ExtractRowsOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ExtractRowsOp should not be null."); + PADDLE_ENFORCE_EQ(ctx->GetInputsVarType("X")[0], + framework::proto::VarType::SELECTED_ROWS, + "The type of input(X) must be SelectedRows."); + auto in_dims = ctx->GetInputDim("X"); + + ctx->SetOutputDim( + "Out", framework::make_ddim(std::vector{in_dims[0], 1})); + } +}; + +class ExtractRowsOp : public framework::OperatorBase { + public: + ExtractRowsOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : framework::OperatorBase(type, inputs, outputs, attrs) {} + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto &in = scope.FindVar(Input("X"))->Get(); + auto out = scope.FindVar(Output("Out"))->GetMutable(); + + auto in_rows = in.rows(); + auto out_dim = framework::make_ddim( + std::vector{static_cast(in_rows.size()), 1}); + auto dst_ptr = out->mutable_data(out_dim, in.place()); + + if (paddle::platform::is_gpu_place(in.place())) { +#ifdef PADDLE_WITH_CUDA + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto *dev_ctx = pool.Get(in.place()); + auto src_ptr = in_rows.Data(in.place()); + auto stream = + reinterpret_cast(*dev_ctx) + .stream(); + memory::Copy(boost::get(out->place()), dst_ptr, + boost::get(in.place()), src_ptr, + in_rows.size() * sizeof(int64_t), stream); +#else + PADDLE_THROW("Not compiled with CUDA."); +#endif + } else { + memory::Copy(platform::CPUPlace(), dst_ptr, platform::CPUPlace(), + in_rows.data(), in_rows.size() * sizeof(int64_t)); + } + } +}; + +class ExtractRowsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(SelectedRows). The input tensor of extract_rows operator," + " and its type is SelectedRows."); + AddOutput("Out", "(Tensor). The the rows of input(X)."); + + AddComment(R"DOC( + ExtractRows Operator. + +The function of extract_rows_op is extracting the rows from the input(X) +whose type is SelectedRows. + + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(extract_rows, ops::ExtractRowsOp, ops::ExtractRowsOpMaker, + ops::ExtractRowsOpInferShape); diff --git a/paddle/fluid/operators/fake_quantize_op.cc b/paddle/fluid/operators/fake_quantize_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a91e0f520e93c01bc5af09b691af2d5a6deda9f2 --- /dev/null +++ b/paddle/fluid/operators/fake_quantize_op.cc @@ -0,0 +1,112 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/fake_quantize_op.h" +#include + +namespace paddle { +namespace operators { + +class FakeQuantizeOp : public framework::OperatorWithKernel { + public: + FakeQuantizeOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FakeQuantizeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of FakeQuantizeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutMovingScale"), + "OutMovingScale(Out) of FakeQuantizeOp should not be null"); + // if (ctx->HasInput("InMovingScale")) { + ctx->SetOutputDim("OutMovingScale", ctx->GetInputDim("InMovingScale")); + //} + // if (ctx->HasInput("InScales")) { + PADDLE_ENFORCE(ctx->HasOutput("OutScales"), + "OutScales(Out) of FakeQuantizeOp should not be null"); + ctx->SetOutputDim("OutScales", ctx->GetInputDim("InScales")); + // PADDLE_ENFORCE_EQ(ctx->Inputs("InScales")[0], + // ctx->Outputs("OutScales")[0], + // "Mean and MeanOut should share the same memory"); + //} + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class FakeQuantizeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) Input tensor of scale operator."); + AddInput("InScales", "(Tensor) scale buffer, used in static quantization.") + .AsDispensable(); + AddInput("InMovingScale", "Last scale, used in static quantization.") + .AsDispensable(); + AddInput("InCurrentIter", + "Last iteration number, used in static quantization.") + .AsDispensable(); + AddOutput("Out", "(Tensor) Output of quantized low level tensor."); + AddOutput("OutScales", + "(Tensor) scale buffer, used in static quantization.") + .AsDispensable(); + AddOutput("OutMovingScale", " Current scale"); + AddOutput("OutCurrentIter", "Current iteration number.").AsDispensable(); + AddAttr("quantize_type", + "(string, default abs_max)" + "The scaling tpe of the quantize operator.") + .SetDefault("abs_max"); + AddAttr("window_size", "(int, default 10000)").SetDefault(10000); + AddAttr("bit_length", "(int, default 8)") + .SetDefault(8) + .AddCustomChecker([](const int &bit_length) { + PADDLE_ENFORCE(bit_length >= 1 && bit_length <= 16, + "'bit_length' should be between 1 and 16."); + }); + AddAttr("is_test", "").SetDefault(false); + AddComment(R"DOC( +FakeQuantize operator + +quantize_type = abs_max: + + $$scale = max(abs(x))$$ + +quantize_type = range_abs_max: + + $$scale = max(max(abs(x)), history_abs_max)$$ + +quantize_type = moving_average_abs_max: + + $$scale = 0.1*scale+0.9*new_abs_max)$$ + +$$Out = scale*X$$ + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(fake_quantize, ops::FakeQuantizeOp, ops::FakeQuantizeOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL( + fake_quantize, + ops::FakeQuantizeKernel, + ops::FakeQuantizeKernel); diff --git a/paddle/fluid/operators/fake_quantize_op.cu b/paddle/fluid/operators/fake_quantize_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..be0c6730a5119090600a27c66510b2a095c54583 --- /dev/null +++ b/paddle/fluid/operators/fake_quantize_op.cu @@ -0,0 +1,272 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/operators/fake_quantize_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +template +__global__ void FindAbsMaxKernel(const int n, const T* in, T* out) { + int bid = threadIdx.x + blockIdx.x * blockDim.x; + int tid = threadIdx.x; + + extern __shared__ T shared_max_data[]; + if (gridDim.x > 1) { + shared_max_data[tid] = T(0); + for (int i = bid; i < n; i += blockDim.x * gridDim.x) { + T tmp = fabs(in[i]); + if (tmp > shared_max_data[tid]) { + shared_max_data[tid] = tmp; + } + } + } else { + if (bid < n) { + shared_max_data[tid] = fabs(in[bid]); + } else { + shared_max_data[tid] = T(0); + } + } + __syncthreads(); + + for (int i = blockDim.x / 2; i > 0; i >>= 1) { + if (tid < i && shared_max_data[tid] < shared_max_data[tid + i]) { + shared_max_data[tid] = shared_max_data[tid + i]; + } + __syncthreads(); + } + if (tid == 0) { + out[blockIdx.x] = shared_max_data[0]; + } +} + +float FindAbsMaxGpu(const platform::CUDADeviceContext& ctx, const float* array, + int length) { + float host_max; + int kNumTheads = 1024; + int gridDimx = (kNumTheads - 1 + length) / kNumTheads; + gridDimx = (gridDimx > kNumTheads) ? kNumTheads : gridDimx; + framework::Tensor t; + float* device_max = t.mutable_data(framework::make_ddim({gridDimx}), + platform::CUDAPlace()); + FindAbsMaxKernel<<>>(length, array, device_max); + FindAbsMaxKernel< + float><<<1, kNumTheads, kNumTheads * sizeof(float), ctx.stream()>>>( + gridDimx, device_max, device_max); + PADDLE_ENFORCE_EQ( + cudaMemcpy(&host_max, device_max, sizeof(float), cudaMemcpyDeviceToHost), + cudaSuccess, "cudaMemcpy failed"); + return host_max; +} + +template +__global__ void ApplySaturateKernel(const int n, const T* in, T* out, + int* num_saturate, const T min, + const T max) { + int bid = threadIdx.x + blockIdx.x * blockDim.x; + int tid = threadIdx.x; + + extern __shared__ int shared_count[]; + shared_count[tid] = 0; + for (int i = bid; i < n; i += blockDim.x * gridDim.x) { + if (in[i] > max) { + out[i] = max; + shared_count[tid] += 1; + } else if (in[i] < min) { + out[i] = min; + shared_count[tid] += 1; + } else { + out[i] = in[i]; + } + } + __syncthreads(); + + for (int i = blockDim.x / 2; i > 0; i >>= 1) { + if (tid < i) { + shared_count[tid] += shared_count[tid + i]; + } + __syncthreads(); + } + if (tid == 0) { + num_saturate[blockIdx.x] = shared_count[0]; + } +} + +template +__global__ void ReduceKernel(const int n, const T* in, T* out) { + int tid = threadIdx.x; + extern __shared__ T shared_sum[]; + if (tid < n) { + shared_sum[tid] = in[tid]; + } else { + shared_sum[tid] = T(0); + } + __syncthreads(); + // blockDim.x must >= n + for (int i = (n + 1) / 2; i > 0; i >>= 1) { + if (tid < i) { + shared_sum[tid] += shared_sum[tid + i]; + } + __syncthreads(); + } + if (tid == 0) { + out[0] = shared_sum[0]; + } +} + +template +int ApplySaturateGpu(const platform::CUDADeviceContext& ctx, const int n, + const T* in, T* out, const T min, const T max) { + int host_num_saturate; + int kNumTheads = 1024; + int gridDimx = (n + kNumTheads - 1) / kNumTheads; + gridDimx = (gridDimx > kNumTheads) ? kNumTheads : gridDimx; + framework::Tensor t; + int* device_num_saturate = t.mutable_data( + framework::make_ddim({gridDimx}), platform::CUDAPlace()); + ApplySaturateKernel< + T><<>>( + n, in, out, device_num_saturate, min, max); + ReduceKernel<<<1, kNumTheads, kNumTheads * sizeof(T), ctx.stream()>>>( + gridDimx, device_num_saturate, device_num_saturate); + PADDLE_ENFORCE_EQ(cudaSuccess, + cudaMemcpy(&host_num_saturate, device_num_saturate, + sizeof(int), cudaMemcpyDeviceToHost), + "cudaMemcpy failed"); + return host_num_saturate; +} + +template +class FakeQuantizeCUDAKernel : public framework::OpKernel { + public: + T FindRangeAbsMax(const platform::CUDADeviceContext& ctx, + framework::Tensor* scale_list, framework::Tensor* out_scale, + const T& cur_scale, int window_size, + int current_iter) const { + T* sl = scale_list->mutable_data(platform::CPUPlace()); + T remove_tmp = sl[current_iter]; + sl[current_iter] = cur_scale; + T& max_scale = out_scale->mutable_data(platform::CPUPlace())[0]; + if (max_scale < cur_scale) { + max_scale = cur_scale; + } else if (fabs(remove_tmp - max_scale) < 1e-6) { + int size = (current_iter > window_size) ? window_size : current_iter; + max_scale = T(FindAbsMaxGpu(ctx, scale_list->data(), size)); + } + return max_scale; + } + + T FindMovingAverageAbsMmax(framework::Tensor* in_scale, + framework::Tensor* out_scale, + const T& cur_scale) const { + T* ins = in_scale->mutable_data(platform::CPUPlace()); + T* outs = out_scale->mutable_data(platform::CPUPlace()); + outs[0] = 0.9 * cur_scale + 0.1 * ins[0]; + return T(outs[0]); + } + + virtual void Compute(const framework::ExecutionContext& context) const { + PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()), + "This kernel only runs on GPU device."); + auto& device_ctx = context.cuda_device_context(); + auto* tensor = context.Output("Out"); + auto* in = context.Input("X"); + const bool is_test = context.Attr("is_test"); + tensor->mutable_data(in->place()); + context.Output("OutMovingScale") + ->mutable_data( + context.Input("InMovingScale")->place()); + auto quantize_type = + static_cast(context.Attr("quantize_type")); + if (quantize_type == std::string("range_abs_max")) { + context.Output("OutScales") + ->mutable_data( + context.Input("InScales")->place()); + context.Output("OutCurrentIter") + ->mutable_data( + context.Input("InCurrentIter")->place()); + } + + T scale = T(1); + int window_size = context.Attr("window_size"); + T bin_cnt = (T)((1 << (context.Attr("bit_length") - 1)) - 1); + if (quantize_type == std::string("abs_max")) { + auto* saving_scale = context.Output("OutMovingScale"); + scale = (T)FindAbsMaxGpu(device_ctx, in->data(), in->numel()); + saving_scale->mutable_data(platform::CPUPlace())[0] = scale; + + auto& device_ctx = context.template device_context(); + auto* scale_list = context.Output("OutScales"); + math::SetConstant scalar; + scale_list->mutable_data(context.GetPlace()); + scalar(device_ctx, scale_list, static_cast(0)); + auto* iter = context.Output("OutCurrentIter"); + iter->mutable_data(context.GetPlace()); + scalar(device_ctx, iter, static_cast(0)); + } else if (quantize_type == std::string("range_abs_max")) { + auto* moving_scale = const_cast( + context.Input("InMovingScale")); + if (is_test) { + scale = moving_scale->mutable_data(platform::CPUPlace())[0]; + } else { + auto* it = const_cast( + context.Input("InCurrentIter")); + auto* iter = context.Output("OutCurrentIter"); + int* last_iter = it->mutable_data(platform::CPUPlace()); + int* current_iter = iter->mutable_data(platform::CPUPlace()); + auto* scale_list = context.Output("OutScales"); + auto* saving_scale = + context.Output("OutMovingScale"); + scale = (T)FindAbsMaxGpu(device_ctx, in->data(), in->numel()); + scale = FindRangeAbsMax(device_ctx, scale_list, saving_scale, scale, + window_size, current_iter[0]); + (*current_iter) = (*last_iter) + 1; + } + } else if (quantize_type == std::string("moving_average_abs_max")) { + auto* moving_scale = const_cast( + context.Input("InMovingScale")); + if (is_test) { + scale = moving_scale->mutable_data(platform::CPUPlace())[0]; + } else { + scale = (T)FindAbsMaxGpu(device_ctx, in->data(), in->numel()); + auto* saving_scale = + context.Output("OutMovingScale"); + scale = FindMovingAverageAbsMmax( + const_cast(moving_scale), saving_scale, scale); + } + } + + ApplySaturateGpu(device_ctx, in->numel(), in->data(), + tensor->mutable_data(in->place()), -scale, scale); + scale = bin_cnt / scale; + + auto& dev = + *context.template device_context().eigen_device(); + auto eigen_out = framework::EigenVector::Flatten(*tensor); + auto eigen_in = framework::EigenVector::Flatten(*tensor); + eigen_out.device(dev) = (scale * eigen_in).round(); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_CUDA_KERNEL(fake_quantize, + paddle::operators::FakeQuantizeCUDAKernel< + paddle::platform::CUDADeviceContext, float>, + paddle::operators::FakeQuantizeCUDAKernel< + paddle::platform::CUDADeviceContext, double>); diff --git a/paddle/fluid/operators/fake_quantize_op.h b/paddle/fluid/operators/fake_quantize_op.h new file mode 100644 index 0000000000000000000000000000000000000000..80f71d85dde39f773cc489fb79effcc775c5010a --- /dev/null +++ b/paddle/fluid/operators/fake_quantize_op.h @@ -0,0 +1,155 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/clip_op.h" +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/platform/transform.h" + +namespace paddle { +namespace operators { + +using platform::Transform; + +template +class FakeQuantizeKernel : public framework::OpKernel { + public: + T FindAbsMax(framework::Tensor* in, int n) const { + T* p = in->mutable_data(platform::CPUPlace()); + T abs_max = (T)0.00000001; + for (int i = 0; i < n; i++) { + T tmp = fabs(p[i]); + if (tmp > abs_max) abs_max = tmp; + } + return T(abs_max); + } + T FindRangeAbsMax(framework::Tensor* scale_list, framework::Tensor* out_scale, + const T& cur_scale, int window_size, + int current_iter) const { + T* sl = scale_list->mutable_data(platform::CPUPlace()); + T remove_tmp = sl[current_iter]; + sl[current_iter] = cur_scale; + T& max_scale = out_scale->mutable_data(platform::CPUPlace())[0]; + if (max_scale < cur_scale) { + max_scale = cur_scale; + } else if (fabs(remove_tmp - max_scale) < 1e-6) { + int size = (current_iter > window_size) ? window_size : current_iter; + max_scale = T(FindAbsMax(scale_list, size)); + } + return max_scale; + } + + T FindMovingAverageAbsMmax(framework::Tensor* in_scale, + framework::Tensor* out_scale, + const T& cur_scale) const { + T* ins = in_scale->mutable_data(platform::CPUPlace()); + T* outs = out_scale->mutable_data(platform::CPUPlace()); + outs[0] = 0.9 * cur_scale + 0.1 * ins[0]; + return T(outs[0]); + } + + virtual void Compute(const framework::ExecutionContext& context) const { + auto* tensor = context.Output("Out"); + auto* in = context.Input("X"); + const bool is_test = context.Attr("is_test"); + tensor->mutable_data(in->place()); + + auto* oms_tensor = context.Output("OutMovingScale"); + oms_tensor->mutable_data(in->place()); + + auto quantize_type = + static_cast(context.Attr("quantize_type")); + if (quantize_type == std::string("range_abs_max")) { + auto* oss_tensor = context.Output("OutScales"); + oss_tensor->mutable_data( + context.Input("InScales")->place()); + auto* oci_tensor = context.Output("OutCurrentIter"); + oci_tensor->mutable_data( + context.Input("InCurrentIter")->place()); + } + + T scale = static_cast(1); + int window_size = context.Attr("window_size"); + int bit_length = context.Attr("bit_length"); + int bin_cnt = std::pow(2, bit_length - 1) - 1; + + auto& dev = + *context.template device_context().eigen_device(); + auto raw_in = framework::EigenVector::Flatten(*in); + if (quantize_type == std::string("abs_max")) { + auto* saving_scale = context.Output("OutMovingScale"); + auto scale_out = framework::EigenVector::Flatten(*saving_scale); + scale_out.device(dev) = raw_in.abs().maximum(); + scale = scale_out(0); + + auto& device_ctx = context.template device_context(); + auto* scale_list = context.Output("OutScales"); + math::SetConstant scalar; + scale_list->mutable_data(context.GetPlace()); + scalar(device_ctx, scale_list, static_cast(0)); + auto* iter = context.Output("OutCurrentIter"); + iter->mutable_data(context.GetPlace()); + scalar(device_ctx, iter, static_cast(0)); + } else if (quantize_type == std::string("range_abs_max")) { + auto* moving_scale = context.Input("InMovingScale"); + if (is_test) { + scale = moving_scale->data()[0]; + } else { + auto* it = context.Input("InCurrentIter"); + auto* iter = context.Output("OutCurrentIter"); + const int* last_iter = it->data(); + int* current_iter = iter->mutable_data(platform::CPUPlace()); + auto* scale_list = context.Output("OutScales"); + auto* saving_scale = + context.Output("OutMovingScale"); + auto scale_out = framework::EigenVector::Flatten(*saving_scale); + scale_out.device(dev) = raw_in.abs().maximum(); + scale = saving_scale->mutable_data(platform::CPUPlace())[0]; + scale = FindRangeAbsMax(scale_list, saving_scale, scale, window_size, + current_iter[0]); + saving_scale->mutable_data(platform::CPUPlace())[0] = scale; + (*current_iter) = (*last_iter) + 1; + } + } else if (quantize_type == std::string("moving_average_abs_max")) { + auto* moving_scale = context.Input("InMovingScale"); + if (is_test) { + scale = moving_scale->data()[0]; + } else { + auto* saving_scale = + context.Output("OutMovingScale"); + auto scale_out = framework::EigenVector::Flatten(*saving_scale); + scale_out.device(dev) = raw_in.abs().maximum(); + scale = saving_scale->mutable_data(platform::CPUPlace())[0]; + scale = FindMovingAverageAbsMmax( + const_cast(moving_scale), saving_scale, scale); + saving_scale->mutable_data(platform::CPUPlace())[0] = scale; + } + } + + Transform trans; + trans(context.template device_context(), in->data(), + in->data() + in->numel(), tensor->mutable_data(in->place()), + ClipFunctor(-scale, scale)); + auto eigen_out = framework::EigenVector::Flatten(*tensor); + auto eigen_in = framework::EigenVector::Flatten(*tensor); + eigen_out.device(dev) = (bin_cnt / scale * eigen_in).round(); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/fc_mkldnn_op.cc b/paddle/fluid/operators/fc_mkldnn_op.cc index 847b7b0c12e1679501dbe83d578b23ca2aef3e9e..e595f1a627cfefbb91b070b898046cf135dc4988 100644 --- a/paddle/fluid/operators/fc_mkldnn_op.cc +++ b/paddle/fluid/operators/fc_mkldnn_op.cc @@ -115,6 +115,7 @@ class MKLDNNMemory { template class FCMKLDNNOpKernel : public paddle::framework::OpKernel { + public: void Compute(const paddle::framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); @@ -124,13 +125,16 @@ class FCMKLDNNOpKernel : public paddle::framework::OpKernel { auto input = ctx.Input("Input"); auto w = ctx.Input("W"); + auto bias = ctx.Input("Bias"); PADDLE_ENFORCE(input->dims().size() == 2 || input->dims().size() == 4, "Input must be with 2 or 4 dimensions, i.e. NCHW"); + // TODO(intel friends): the native weight format is io, + // but the mkldnn weight format is oihw, which may need be transposed. PADDLE_ENFORCE(w->dims().size() == 2 || w->dims().size() == 4, "Weights must be with 2 or 4 dimensions, i.e. OI or OIHW"); - bool with_bias = ctx.Attr("bias_attr"); + bool with_bias = bias != nullptr; MKLDNNMD md(input, w, with_bias); std::shared_ptr pd = @@ -153,6 +157,7 @@ class FCMKLDNNOpKernel : public paddle::framework::OpKernel { auto dst_memory = mem.dst(output_data); auto src_memory = mem.src(input_data); auto weights_memory = mem.weights(w_data); + // TODO(intel friends): bias memory should also be obtain from bias->data() auto bias_memory = mem.bias(); auto forward = with_bias ? mkldnn::inner_product_forward( @@ -215,7 +220,8 @@ class FCMKLDNNGradOpKernel : public paddle::framework::OpKernel { const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); const T* out_grad_data = out_grad->data(); - bool with_bias = ctx.Attr("bias_attr"); + auto bias = ctx.Input("Bias"); + bool with_bias = bias != nullptr; MKLDNNMD md(input, w, with_bias); MKLDNNMemory mem(&md, mkldnn_engine); diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 8843a1c44b7004ba5d7935f75d3c99d9c30fc6c0..099ca52c8e945a0e93c2f13adb612158c67397cf 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -14,6 +14,9 @@ limitations under the License. */ #include "paddle/fluid/operators/fc_op.h" #include +#include "paddle/fluid/operators/math/blas.h" + +DECLARE_int32(paddle_num_threads); namespace paddle { namespace operators { @@ -25,16 +28,24 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const { "Out(Output) of Fully Connected should not be null."); PADDLE_ENFORCE(ctx->HasInput("W"), "W(Input) of Fully Connected should not be null."); - + // NCHW auto in_dims = ctx->GetInputDim("Input"); + // IO, I=C*H*W auto w_dims = ctx->GetInputDim("W"); std::vector output_shape({in_dims[0], w_dims[1]}); + if (ctx->HasInput("Bias")) { + auto bias_dims = ctx->GetInputDim("Bias"); + PADDLE_ENFORCE_EQ(bias_dims[0], 1, "The shape of Bias must be [1, dim]."); + PADDLE_ENFORCE_EQ(bias_dims[1], w_dims[1], + "The shape of Bias must be [1, dim]."); + } PADDLE_ENFORCE(in_dims.size() == 2 || in_dims.size() == 4, "Fully Connected input should be 2-D or 4-D tensor."); - - PADDLE_ENFORCE(w_dims.size() == 2 || w_dims.size() == 4, - "Fully Connected input should be 2-D or 4-D tensor."); + PADDLE_ENFORCE_EQ(w_dims.size(), 2UL, + "Fully Connected input should be 2-D tensor."); + PADDLE_ENFORCE_EQ(framework::product(in_dims) / in_dims[0], w_dims[0], + "Fully Connected input and weigth size do not match."); ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); ctx->ShareLoD("Input", "Out"); @@ -42,9 +53,12 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType FCOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { - framework::LibraryType library{framework::LibraryType::kMKLDNN}; - framework::DataLayout layout{framework::DataLayout::kAnyLayout}; - + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + if (ctx.Attr("use_mkldnn")) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), layout, library); @@ -60,27 +74,39 @@ void FCOpGrad::InferShape(framework::InferShapeContext* ctx) const { if (ctx->HasOutput(framework::GradVarName("W"))) { ctx->SetOutputDim(framework::GradVarName("W"), w_dims); } + + if (ctx->HasInput("Bias")) { + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), + "Should have bias grad"); + auto bias_dims = ctx->GetInputDim("Bias"); + ctx->SetOutputDim(framework::GradVarName("Bias"), bias_dims); + } } framework::OpKernelType FCOpGrad::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { - framework::LibraryType library{framework::LibraryType::kMKLDNN}; - framework::DataLayout layout{framework::DataLayout::kAnyLayout}; - + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + if (ctx.Attr("use_mkldnn")) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), layout, library); } void FCOpMaker::Make() { - AddInput("Input", "(Tensor) The input tensor of fully connected operator. "); - AddInput("W", "(Tensor), The second input tensor of fc op."); + AddInput("Input", + "(Tensor), The input tensor of fully connected operator with format " + "(NCHW). "); + AddInput("W", "(Tensor), The weight fc op with shape (I, O)."); + AddInput("Bias", "(Tensor, optional) Bias vector with shape (1 x O") + .AsDispensable(); AddOutput("Out", "(Tensor) The output tensor of fully connected operator. "); AddAttr("use_mkldnn", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); - AddAttr("bias_attr", "(bool, default false) Only used in mkldnn kernel") - .SetDefault(false); AddComment(R"DOC( Fully Connected Operator. @@ -94,9 +120,47 @@ void FCOpMaker::Make() { )DOC"); } +template +class FCOpKernel : public framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + auto input = ctx.Input("Input"); + auto w = ctx.Input("W"); + auto bias = ctx.Input("Bias"); + auto output = ctx.Output("Out"); + auto in_dims = input->dims(); + auto w_dims = w->dims(); + + auto& dev_ctx = ctx.template device_context(); + auto blas = math::GetBlas(dev_ctx); + const T* input_data = input->data(); + const T* w_data = w->data(); + T* output_data = output->mutable_data(ctx.GetPlace()); + + blas.GEMM(CblasNoTrans, CblasNoTrans, in_dims[0], w_dims[1], w_dims[0], + static_cast(1), input_data, w_data, static_cast(0), + output_data); + + if (bias) { + const T* bias_data = bias->data(); +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for if (FLAGS_paddle_num_threads > 1) +#endif + for (int bs = 0; bs < in_dims[0]; bs++) { + blas.AXPY(w_dims[1], static_cast(1), bias_data, + output_data + bs * w_dims[1]); + } + } + } +}; + } // namespace operators } // namespace paddle -REGISTER_OPERATOR(fc, paddle::operators::FCOp, paddle::operators::FCOpMaker, +namespace ops = paddle::operators; +REGISTER_OPERATOR(fc, ops::FCOp, ops::FCOpMaker, paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(fc_grad, paddle::operators::FCOpGrad); +REGISTER_OPERATOR(fc_grad, ops::FCOpGrad); +REGISTER_OP_CPU_KERNEL(fc, ops::FCOpKernel, ops::FCOpKernel); diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc index bcb3e63ed7dbc775c1de6c4522f0548ea48a6cf0..dc7ef664958238ddbd48745bd59cc7db28e49f5b 100644 --- a/paddle/fluid/operators/feed_op.cc +++ b/paddle/fluid/operators/feed_op.cc @@ -31,7 +31,6 @@ class FeedOp : public framework::OperatorBase { const platform::Place &place) const override { // get device context from pool auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place); - platform::RecordEvent record_event(Type(), dev_ctx); auto feed_var_name = Input("X"); auto *feed_var = scope.FindVar(feed_var_name); diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc index 79ec02f52094121d01c6bda2a5d99d2211893e89..d9cd956dfdff3d009d38ee5088f5396080580483 100644 --- a/paddle/fluid/operators/fetch_barrier_op.cc +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -19,8 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" - -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -37,21 +36,16 @@ class FetchBarrierOp : public framework::OperatorBase { void RunImpl(const framework::Scope& scope, const platform::Place& place) const override { std::vector eps = Attr>("endpoints"); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - // For profiling - platform::RecordEvent record_event(Type(), &ctx); - - auto rpc_client = detail::RPCClient::GetInstance(); - - PADDLE_ENFORCE(rpc_client->Wait()); + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); for (auto& ep : eps) { VLOG(3) << "fetch barrier, ep: " << ep; rpc_client->AsyncSendFetchBarrier(ep); } - PADDLE_ENFORCE(rpc_client->Wait()); + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } }; diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc index 1640a2a22c69a0e3ab81a2889d6105b2cf4162b7..c197b45e8196a47def6465128e8ca39d8daefed6 100644 --- a/paddle/fluid/operators/fetch_op.cc +++ b/paddle/fluid/operators/fetch_op.cc @@ -30,9 +30,6 @@ class FetchOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - platform::RecordEvent record_event(Type(), pool.Get(place)); - auto fetch_var_name = Input("X"); auto *fetch_var = scope.FindVar(fetch_var_name); PADDLE_ENFORCE(fetch_var != nullptr, diff --git a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc index 1ae78675a0cac8a72aeaef1227b631a41e4a10b2..453a1b32a0171a2ca88879ab3287e89c4d3c7759 100644 --- a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc +++ b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc @@ -32,16 +32,16 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp { class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { protected: void Apply() override { - AddAttr("dtype", - "(int, default 5 (FP32)) " - "Output data type") + AddAttr( + "dtype", + "It could be numpy.dtype. Output data type. Default is float32") .SetDefault(framework::proto::VarType::FP32); - AddAttr("value", "(float, default 0) The value to be filled") + AddAttr("value", "default 0. The value to be filled") .SetDefault(0.0f); AddComment(R"DOC( -FillConstantBatchSizeLike Operator. - -Fill up a variable with specified constant value. +This function creates a tensor of specified *shape*, *dtype* and batch size, +and initializes this with a constant supplied in *value*. The batch size is +obtained from the `input` tensor. )DOC"); } diff --git a/paddle/fluid/operators/flatten_op.cc b/paddle/fluid/operators/flatten_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..fdda01381e117cecffb2a05f8399f3ad82a46339 --- /dev/null +++ b/paddle/fluid/operators/flatten_op.cc @@ -0,0 +1,169 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class FlattenOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input (X) of Flatten op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output (Output) of Flatten op should not be null."); + const auto &axis = ctx->Attrs().Get("axis"); + const auto &in_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE(axis >= 0, "The axis should be greater than or equal to 0."); + PADDLE_ENFORCE( + axis <= in_dims.size(), + "The axis should be less than or equal to input tensor's rank."); + + const auto &out_dims = GetOutputShape(axis, in_dims); + ctx->SetOutputDim("Out", framework::make_ddim(out_dims)); + if (in_dims[0] == out_dims[0]) { + // Only pass LoD when the first dimension of output and Input(X) + // are the same. + ctx->ShareLoD("X", "Out"); + } + } + + static std::vector GetOutputShape(const int axis, + const framework::DDim &in_dims) { + int64_t outer = 1, inner = 1; + for (int i = 0; i < in_dims.size(); ++i) { + if (i < axis) { + outer *= in_dims[i]; + } else { + inner *= in_dims[i]; + } + } + std::vector out_shape(2); + out_shape[0] = outer; + out_shape[1] = inner; + return out_shape; + } +}; + +class FlattenOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto &axis = Attr("axis"); + auto in_dims = + scope.FindVar(Input("X"))->Get().dims(); + const auto &out_dims = FlattenOpInferShape::GetOutputShape(axis, in_dims); + + framework::AttributeMap attrs; + attrs["shape"] = out_dims; + attrs["inplace"] = false; + // Invoke Reshape Op + auto reshape_op = framework::OpRegistry::CreateOp( + "reshape", {{"X", {Input("X")}}, {"Shape", {}}}, + {{"Out", {Output("Out")}}}, attrs); + reshape_op->Run(scope, place); + } +}; + +class FlattenOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) A tensor of rank >= axis."); + AddOutput("Out", + "A 2D tensor is reshaped input tensor. The input dimensions" + "up to axis are flattened to the outer dimension of the output" + "and the remaining input dimensions are flattened into the inner" + "dimension of the output."); + AddAttr("axis", + "(int)" + "Indicate up to which input dimensions (exclusive) should be" + "flattened to the outer dimension of the output. The value" + "for axis must be in the range [0, R], where R is the rank of" + "the input tensor. When axis = 0, the shape of the output" + "tensor is (1, (d_0 X d_1 ... d_n), where the shape of the" + "input tensor is (d_0, d_1, ... d_n).") + .SetDefault(1); + AddComment(R"DOC( +Flatten Operator + +Flattens the input tensor into a 2D matrix. + +Examples: +Case 1: + Given + X.shape = (3, 100, 100, 4) + and + axis = 2 + We get: + Out.shape = (3 * 100, 4 * 100) + +Case 2: + Given + X.shape = (3, 100, 100, 4) + and + axis = 0 + We get: + Out.shape = (1, 3 * 100 * 100 * 4) +)DOC"); + } +}; + +class FlattenGradInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *context) const override { + context->SetOutputDim(framework::GradVarName("X"), + context->GetInputDim("X")); + context->ShareLoD("X", framework::GradVarName("X")); + } +}; + +class FlattenGradOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto dx_name = Output(framework::GradVarName("X")); + auto dout_name = Input(framework::GradVarName("Out")); + auto in_dims = + scope.FindVar(Input("X"))->Get().dims(); + framework::AttributeMap attrs; + attrs["shape"] = framework::vectorize2int(in_dims); + attrs["inplace"] = false; + + auto reshape_op = framework::OpRegistry::CreateOp( + "reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}}, + attrs); + reshape_op->Run(scope, place); + } +}; + +} // namespace operators +} // namespace paddle + +USE_OP(reshape); + +namespace ops = paddle::operators; +REGISTER_OPERATOR(flatten, ops::FlattenOp, ops::FlattenOpMaker, + ops::FlattenOpInferShape, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(flatten_grad, ops::FlattenGradOp, ops::FlattenGradInferShape); diff --git a/paddle/fluid/operators/fused_elemwise_activation_op.cc b/paddle/fluid/operators/fused_elemwise_activation_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a6fd0aeb021dce40339c32251af130d5984dccd2 --- /dev/null +++ b/paddle/fluid/operators/fused_elemwise_activation_op.cc @@ -0,0 +1,221 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "paddle/fluid/operators/fused_elemwise_activation_op.h" + +namespace paddle { +namespace operators { + +class FusedElemwiseActivationOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE( + ctx->HasInput("X"), + "Input(X) of FusedElemwiseActivationOp op should not be null."); + PADDLE_ENFORCE( + ctx->HasInput("Y"), + "Input(Y) of FusedElemwiseActivationOp op should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Output(Out) of FusedElemwiseActivationOp op should not be null."); + + auto x_dim = ctx->GetInputDim("X"); + auto y_dim = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(), + "Rank of first input must >= rank of second input."); + + ctx->SetOutputDim("Out", x_dim); + ctx->ShareLoD("X", /*->*/ "Out"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("X")->type(), + ctx.Input("Y")->type(), + "The element's type of input should be the same."); + auto input_data_type = + framework::ToDataType(ctx.Input("X")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } +}; + +class FusedElemwiseActivationMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(vector)"); + AddInput("Y", "(vector)"); + AddOutput("Out", "vector"); + AddAttr("axis", + "axis is used by elementwise_op, the default value is -1.") + .SetDefault(-1); + AddAttr("scale", + "scale is used by scale_op, the default value is 0.0.") + .SetDefault(0.0); + AddAttr("recomputation", + "Whether to recompute the Out." + "fused_elemwise_activation_grad has two methods to get the " + "dx and dy, one " + "is to use the 'Out', and the other is not to use it. " + "The former method will save the time of recomputing the " + "'Out', but it must occupy the memory to store the 'out'. " + "While, the later method can avoid occupying the memory, " + "but it must recompute the 'Out'. The default value is true.") + .SetDefault(true); + AddAttr>("functor_list", + "The functors that should be fused.") + .AddCustomChecker([&](const std::vector &functor_list) { + PADDLE_ENFORCE(ValidCheck(functor_list)); + }); + + AddComment(R"DOC( +FusedElemwiseActivation Operator. + +At present, FusedElemwiseActivation only supports Two kinds of compound +operators (elementwise_op and activation_op): + + Z = Binary(X, Unary(Y)) + Z = Unary(Binary(X, Y)) + +The attributions of activation_op can be get from fused_elemwise_activation_op's +attributions. functor_list records the functors to be fused, for example +"scale,elementwise_add". + +)DOC"); + } + + private: + bool ValidCheck(const std::vector &functors) { + std::unordered_set unary_fun = {"scale", "relu"}; + std::unordered_set binary_fun = {"elementwise_add"}; + + std::string unary_fun_str; + if (binary_fun.count(functors[0])) { + unary_fun_str = functors[1]; + } else if (binary_fun.count(functors[1])) { + unary_fun_str = functors[0]; + } else { + PADDLE_THROW("%s and %s are not included in fused_list.", functors[0], + functors[1]); + } + PADDLE_ENFORCE_EQ(unary_fun.count(unary_fun_str), 1, + "%s is not included in fused_list.", unary_fun_str); + return true; + } +}; + +class FusedElemwiseActivationGradMaker + : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *op_desc_ptr = new framework::OpDesc(); + op_desc_ptr->SetType(this->ForwardOpType() + "_grad"); + + for (auto &input_param : this->InputNames()) { + op_desc_ptr->SetInput(input_param, this->Input(input_param)); + op_desc_ptr->SetOutput(framework::GradVarName(input_param), + this->InputGrad(input_param, true)); + } + + for (auto &output_param : this->OutputNames()) { + op_desc_ptr->SetInput(output_param, this->Output(output_param)); + op_desc_ptr->SetInput(framework::GradVarName(output_param), + this->OutputGrad(output_param)); + } + op_desc_ptr->SetAttrMap(this->Attrs()); + + std::vector functor_names = + boost::get>( + op_desc_ptr->GetAttr("functor_list")); + functor_names[0] += "_grad"; + functor_names[1] += "_grad"; + op_desc_ptr->SetAttr("functor_list", functor_names); + return std::unique_ptr(op_desc_ptr); + } +}; + +class FusedElemwiseActivationOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); + + PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + "Rank of first input must >= rank of second input."); + + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type_index = ctx.Input("X")->type(); + PADDLE_ENFORCE_EQ(input_data_type_index, + ctx.Input("Y")->type(), + "The element's type of input should be the same."); + PADDLE_ENFORCE_EQ( + input_data_type_index, + ctx.Input(framework::GradVarName("Out"))->type(), + "The element's type of input should be the same."); + + auto input_data_type = framework::ToDataType(input_data_type_index); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(fused_elemwise_activation, ops::FusedElemwiseActivationOp, + ops::FusedElemwiseActivationMaker, + ops::FusedElemwiseActivationGradMaker); +REGISTER_OPERATOR(fused_elemwise_activation_grad, + ops::FusedElemwiseActivationOpGrad); + +REGISTER_OP_CPU_KERNEL( + fused_elemwise_activation, + ops::FusedElemwiseActivationKernel, + ops::FusedElemwiseActivationKernel); + +REGISTER_OP_CPU_KERNEL( + fused_elemwise_activation_grad, + ops::FusedElemwiseActivationGradKernel, + ops::FusedElemwiseActivationGradKernel); diff --git a/paddle/fluid/operators/fused_elemwise_activation_op.cu b/paddle/fluid/operators/fused_elemwise_activation_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..e1d2b16b4b5e3a480777f834c2cbeb6d00a755e4 --- /dev/null +++ b/paddle/fluid/operators/fused_elemwise_activation_op.cu @@ -0,0 +1,30 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/fused_elemwise_activation_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + fused_elemwise_activation, + ops::FusedElemwiseActivationKernel, + ops::FusedElemwiseActivationKernel); + +REGISTER_OP_CUDA_KERNEL( + fused_elemwise_activation_grad, + ops::FusedElemwiseActivationGradKernel, + ops::FusedElemwiseActivationGradKernel); diff --git a/paddle/fluid/operators/fused_elemwise_activation_op.h b/paddle/fluid/operators/fused_elemwise_activation_op.h new file mode 100644 index 0000000000000000000000000000000000000000..fe0017b824532b1210d0ae3e51983d63d081f12a --- /dev/null +++ b/paddle/fluid/operators/fused_elemwise_activation_op.h @@ -0,0 +1,425 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/elementwise_op_function.h" +#include "paddle/fluid/operators/math/functors.h" + +namespace math = paddle::operators::math; + +namespace paddle { +namespace operators { + +// CompoundFunctors +// For example: Z = Binary(X, Unary(Y)) +template +struct BinaryCompoundFunctor { + BinaryCompoundFunctor(const BinaryFun &binary_fun, const UnaryFun &unary_fun) + : binary_fun_(binary_fun), unary_fun_(unary_fun) {} + + inline HOSTDEVICE T operator()(T x, T y) { + return binary_fun_(x, unary_fun_(y)); + } + + private: + BinaryFun binary_fun_; + UnaryFun unary_fun_; +}; + +// For example: Z = Unary(Binary(X, Y)) +template +struct UnaryCompoundFunctor { + UnaryCompoundFunctor(const UnaryFun &unary_fun, const BinaryFun &binary_fun) + : unary_fun_(unary_fun), binary_fun_(binary_fun) {} + + inline HOSTDEVICE T operator()(T x, T y) { + return unary_fun_(binary_fun_(x, y)); + } + + private: + UnaryFun unary_fun_; + BinaryFun binary_fun_; +}; + +// FIXME(zcd): DBinaryFun and DUnaryFun have to method to get +// the dx, one is to use the 'out', and the other is not to use it. +// the former method will save the time of recomputing the +// 'out', but it must occupy the memory to store the 'out'. +// While the later method can avoid occupying this memory, +// but it must recompute the 'out'. + +template +struct BinaryCompoundGradDxFunctor { + BinaryCompoundGradDxFunctor(const DBinaryFun &d_binary_fun, + const UnaryFun &unary_fun) + : d_binary_fun_(d_binary_fun), unary_fun_(unary_fun) {} + + inline HOSTDEVICE T operator()(T x, T y, T out, T dout) { + if (Recomputation) { + return dout * d_binary_fun_(x, unary_fun_(y)); + } else { + return dout * d_binary_fun_(x, unary_fun_(y), out); + } + } + + private: + DBinaryFun d_binary_fun_; + UnaryFun unary_fun_; +}; + +template +struct BinaryCompoundGradDyFunctor { + BinaryCompoundGradDyFunctor(const DBinaryFun &d_binary_fun, + const UnaryFun &unary_fun, + const DUnaryFun &d_unary_fun) + : d_binary_fun_(d_binary_fun), + unary_fun_(unary_fun), + d_unary_fun_(d_unary_fun) {} + + inline HOSTDEVICE T operator()(T x, T y, T out, T dout) { + if (Recomputation) { + return dout * d_binary_fun_(unary_fun_(y), x) * d_unary_fun_(y); + } else { + return dout * d_binary_fun_(unary_fun_(y), x, out) * d_unary_fun_(y); + } + } + + private: + DBinaryFun d_binary_fun_; + UnaryFun unary_fun_; + DUnaryFun d_unary_fun_; +}; + +template +struct UnaryCompoundGradDxFunctor { + UnaryCompoundGradDxFunctor(const DUnaryFun &d_unary_fun, + const BinaryFun &binary_fun, + const DBinaryFun &d_binary_fun) + : d_unary_fun_(d_unary_fun), + binary_fun_(binary_fun), + d_binary_fun_(d_binary_fun) {} + + inline HOSTDEVICE T operator()(T x, T y, T out, T dout) { + T base; + if (Recomputation) { + base = dout * d_unary_fun_(binary_fun_(x, y)); + } else { + base = dout * d_unary_fun_(binary_fun_(x, y), out); + } + return base * d_binary_fun_(x, y); + } + + private: + DUnaryFun d_unary_fun_; + BinaryFun binary_fun_; + DBinaryFun d_binary_fun_; +}; + +template +struct UnaryCompoundGradDyFunctor { + UnaryCompoundGradDyFunctor(const DUnaryFun &d_unary_fun, + const BinaryFun &binary_fun, + const DBinaryFun &d_binary_fun) + : d_unary_fun_(d_unary_fun), + binary_fun_(binary_fun), + d_binary_fun_(d_binary_fun) {} + + inline HOSTDEVICE T operator()(T x, T y, T out, T dout) { + T base; + if (Recomputation) { + base = dout * d_unary_fun_(binary_fun_(x, y)); + } else { + base = dout * d_unary_fun_(binary_fun_(x, y), out); + } + return base * d_binary_fun_(y, x); + } + + private: + DUnaryFun d_unary_fun_; + BinaryFun binary_fun_; + DBinaryFun d_binary_fun_; +}; + +template +static void RunBinaryCompoundFunctor(const framework::ExecutionContext &ctx, + const BinaryFunctor &binary_functor, + const UnaryFunctor &unary_functor, + const framework::Tensor *in_x, + const framework::Tensor *in_y, + framework::Tensor *output) { + int axis = ctx.Attr("axis"); + using BinaryCompoundFunctor = + BinaryCompoundFunctor; + + ElementwiseComputeEx( + ctx, in_x, in_y, axis, + BinaryCompoundFunctor(binary_functor, unary_functor), output); +} + +template +static void RunUnaryCompoundFunctors(const framework::ExecutionContext &ctx, + const UnaryFunctor &unary_functor, + const BinaryFunctor &binary_functor, + const framework::Tensor *in_x, + const framework::Tensor *in_y, + framework::Tensor *output) { + int axis = ctx.Attr("axis"); + + using UnaryCompoundFunctor = + UnaryCompoundFunctor; + + ElementwiseComputeEx( + ctx, in_x, in_y, axis, + UnaryCompoundFunctor(unary_functor, binary_functor), output); +} + +template +static void RunBinaryCompoundGradFunctors( + const framework::ExecutionContext &ctx, + const BinaryGradFunctor &binary_grad_functor, + const UnaryFunctor &unary_functor, + const UnaryGradFunctor &unary_grad_functor, const framework::Tensor *in_x, + const framework::Tensor *in_y, const framework::Tensor *in_out, + const framework::Tensor *in_out_grad, framework::Tensor *x_grad, + framework::Tensor *y_grad) { + int axis = ctx.Attr("axis"); + + using BinaryCompoundDxFunctor = + BinaryCompoundGradDxFunctor; + using BinaryCompoundDyFunctor = + BinaryCompoundGradDyFunctor; + + ElemwiseGradCompute( + ctx, *in_x, *in_y, *in_out, *in_out_grad, axis, x_grad, y_grad, + BinaryCompoundDxFunctor(binary_grad_functor, unary_functor), + BinaryCompoundDyFunctor(binary_grad_functor, unary_functor, + unary_grad_functor)); +} + +template +static void RunUnaryCompoundGradFunctors( + const framework::ExecutionContext &ctx, + const UnaryGradFunctor &unary_grad_functor, + const BinaryFunctor &binary_functor, + const BinaryGradFunctor &binary_grad_functor, const framework::Tensor *in_x, + const framework::Tensor *in_y, const framework::Tensor *in_out, + const framework::Tensor *in_out_grad, framework::Tensor *x_grad, + framework::Tensor *y_grad) { + int axis = ctx.Attr("axis"); + + using UnaryCompoundDxFunctor = + UnaryCompoundGradDxFunctor; + using UnaryCompoundDyFunctor = + UnaryCompoundGradDyFunctor; + + ElemwiseGradCompute( + ctx, *in_x, *in_y, *in_out, *in_out_grad, axis, x_grad, y_grad, + UnaryCompoundDxFunctor(unary_grad_functor, binary_functor, + binary_grad_functor), + UnaryCompoundDyFunctor(unary_grad_functor, binary_functor, + binary_grad_functor)); +} + +template +static void RunFunctors(const framework::ExecutionContext &ctx, + const framework::Tensor *in_x, + const framework::Tensor *in_y, + framework::Tensor *output) { + auto &functors = ctx.Attr>("functor_list"); + auto funcs_str = functors[0] + "," + functors[1]; + // TODO(zcd): The following code can be refined. + if (funcs_str == "elementwise_add,scale") { + // Z = Binary(X, Unary(Y)) + T scale = static_cast(ctx.Attr("scale")); + RunBinaryCompoundFunctor, + math::ScaleFunctor>( + ctx, math::AddFunctor(), math::ScaleFunctor(scale), in_x, in_y, + output); + } else if (funcs_str == "scale,elementwise_add") { + // Z = Unary(Binary(X, Y)) + T scale = static_cast(ctx.Attr("scale")); + RunUnaryCompoundFunctors, + math::AddFunctor>( + ctx, math::ScaleFunctor(scale), math::AddFunctor(), in_x, in_y, + output); + } else if (funcs_str == "elementwise_add,relu") { + RunBinaryCompoundFunctor, + math::ReluFunctor>( + ctx, math::AddFunctor(), math::ReluFunctor(), in_x, in_y, output); + } else if (funcs_str == "relu,elementwise_add") { + RunUnaryCompoundFunctors, + math::AddFunctor>( + ctx, math::ReluFunctor(), math::AddFunctor(), in_x, in_y, output); + } else { + PADDLE_THROW("%s has not been implemented.", funcs_str); + } +} + +template +static void RunGradFunctors(const framework::ExecutionContext &ctx, + const framework::Tensor *in_x, + const framework::Tensor *in_y, + const framework::Tensor *in_out, + const framework::Tensor *in_out_grad, + framework::Tensor *x_grad, + framework::Tensor *y_grad) { + auto &functors = ctx.Attr>("functor_list"); + auto funcs_str = functors[0] + "," + functors[1]; + + bool recomputation = ctx.Attr("recomputation"); + + // TODO(zcd): The following code can be refined. for example, use registion + if (funcs_str == "elementwise_add_grad,scale_grad") { + // The backward of Z = Binary(X, Unary(Y)) + T scale = static_cast(ctx.Attr("scale")); + if (recomputation) { + RunBinaryCompoundGradFunctors, + math::ScaleFunctor, + math::ScaleGradFunctor, true>( + ctx, math::AddGradFunctor(), math::ScaleFunctor(scale), + math::ScaleGradFunctor(scale), in_x, in_y, in_out, in_out_grad, + x_grad, y_grad); + } else { + RunBinaryCompoundGradFunctors, + math::ScaleFunctor, + math::ScaleGradFunctor, false>( + ctx, math::AddGradFunctor(), math::ScaleFunctor(scale), + math::ScaleGradFunctor(scale), in_x, in_y, in_out, in_out_grad, + x_grad, y_grad); + } + } else if (funcs_str == "scale_grad,elementwise_add_grad") { + // The backward of Z = Unary(Binary(X, Y)) + T scale = static_cast(ctx.Attr("scale")); + if (recomputation) { + RunUnaryCompoundGradFunctors, + math::AddFunctor, math::AddGradFunctor, + true>(ctx, math::ScaleGradFunctor(scale), + math::AddFunctor(), + math::AddGradFunctor(), in_x, in_y, + in_out, in_out_grad, x_grad, y_grad); + } else { + RunUnaryCompoundGradFunctors, + math::AddFunctor, math::AddGradFunctor, + false>(ctx, math::ScaleGradFunctor(scale), + math::AddFunctor(), + math::AddGradFunctor(), in_x, in_y, + in_out, in_out_grad, x_grad, y_grad); + } + } else if (funcs_str == "elementwise_add_grad,relu_grad") { + if (recomputation) { + RunBinaryCompoundGradFunctors, + math::ReluFunctor, + math::ReluGradFunctor, true>( + ctx, math::AddGradFunctor(), math::ReluFunctor(), + math::ReluGradFunctor(), in_x, in_y, in_out, in_out_grad, x_grad, + y_grad); + } else { + RunBinaryCompoundGradFunctors, + math::ReluFunctor, + math::ReluGradFunctor, false>( + ctx, math::AddGradFunctor(), math::ReluFunctor(), + math::ReluGradFunctor(), in_x, in_y, in_out, in_out_grad, x_grad, + y_grad); + } + } else if (funcs_str == "relu_grad,elementwise_add_grad") { + if (recomputation) { + RunUnaryCompoundGradFunctors, + math::AddFunctor, math::AddGradFunctor, + true>(ctx, math::ReluGradFunctor(), + math::AddFunctor(), + math::AddGradFunctor(), in_x, in_y, + in_out, in_out_grad, x_grad, y_grad); + } else { + RunUnaryCompoundGradFunctors, + math::AddFunctor, math::AddGradFunctor, + false>(ctx, math::ReluGradFunctor(), + math::AddFunctor(), + math::AddGradFunctor(), in_x, in_y, + in_out, in_out_grad, x_grad, y_grad); + } + } else { + PADDLE_THROW("%s has not been implemented.", funcs_str); + } +} + +template +class FusedElemwiseActivationKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto &in_x = detail::Ref(ctx.Input("X"), + "Cannot get input tensor %s, variable name = %s", + "X", ctx.op().Input("X")); + auto &in_y = detail::Ref(ctx.Input("Y"), + "Cannot get input tensor %s, variable name = %s", + "Y", ctx.op().Input("Y")); + auto &output = detail::Ref(ctx.Output("Out"), + "Cannot get input tensor %s, variable name = %s", + "Out", ctx.op().Output("Out")); + + RunFunctors(ctx, &in_x, &in_y, &output); + } +}; + +template +class FusedElemwiseActivationGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto &in_x = detail::Ref(ctx.Input("X"), + "Cannot get input tensor %s, variable name = %s", + "X", ctx.op().Input("X")); + auto &in_y = detail::Ref(ctx.Input("Y"), + "Cannot get input tensor %s, variable name = %s", + "Y", ctx.op().Input("Y")); + auto &in_out = detail::Ref(ctx.Input("Out"), + "Cannot get input tensor %s, variable name = %s", + "Out", ctx.op().Input("Out")); + auto &in_out_grad = + detail::Ref(ctx.Input(framework::GradVarName("Out")), + "Cannot get input tensor %s, variable name = %s", + framework::GradVarName("Out"), + ctx.op().Input(framework::GradVarName("Out"))); + + framework::Tensor *x_grad = + ctx.Output(framework::GradVarName("X")); + framework::Tensor *y_grad = + ctx.Output(framework::GradVarName("Y")); + + RunGradFunctors(ctx, &in_x, &in_y, &in_out, &in_out_grad, + x_grad, y_grad); + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index e21b57258928856a10d6e86c3e2c6e81fb241ee3..aa3e05b83b23569a4dd9c83294916e289f993abc 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -33,7 +33,6 @@ class GatherOp : public framework::OperatorWithKernel { auto index_dims = ctx->GetInputDim("Index"); PADDLE_ENFORCE(index_dims.size() == 1); int batch_size = ctx->GetInputDim("Index")[0]; - PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); framework::DDim output_dims(ctx->GetInputDim("X")); output_dims[0] = batch_size; ctx->SetOutputDim("Out", output_dims); diff --git a/paddle/fluid/operators/gather_test.cc b/paddle/fluid/operators/gather_test.cc index 9c0561b016fdbfa8e48535eaa673a3f85bc936e5..f6b156eb30dae154395b34dcfc26319cd89edbca 100644 --- a/paddle/fluid/operators/gather_test.cc +++ b/paddle/fluid/operators/gather_test.cc @@ -43,7 +43,8 @@ TEST(Gather, GatherData) { auto* cpu_place = new paddle::platform::CPUPlace(); paddle::platform::CPUDeviceContext ctx(*cpu_place); paddle::operators::CPUGather(ctx, *src, *index, output); - + delete cpu_place; + cpu_place = NULL; for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); diff --git a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc index 8050f61d4546f3351645f23ddcc63b2c49f17929..4a974281481c8bc02589b428098475d73b8a0ba5 100644 --- a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc @@ -36,11 +36,12 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { void Apply() override { AddAttr("mean", "(float, default 0.0) " - "mean of random tensor.") + "The mean (or center) of the gaussian distribution.") .SetDefault(.0f); AddAttr("std", "(float, default 1.0) " - "std of random tensor.") + "The standard deviation (std, or spread) of the " + "gaussian distribution.") .SetDefault(1.0f); AddAttr("seed", "(int, default 0) " @@ -55,9 +56,11 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { .SetDefault(framework::proto::VarType::FP32); AddComment(R"DOC( -GaussianRandom Operator. Used to initialize tensors with gaussian random generator. +The defalut mean of the distribution is 0. and defalut standard +deviation (std) of the distribution is 1.. Uers can set mean and std +by input arguments. )DOC"); } }; diff --git a/paddle/fluid/operators/gaussian_random_mkldnn_op.cc b/paddle/fluid/operators/gaussian_random_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..76b00b396c1349eff5db1059268e7cf280a8fc64 --- /dev/null +++ b/paddle/fluid/operators/gaussian_random_mkldnn_op.cc @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/operators/mean_op.h" + +namespace paddle { +namespace operators { + +using framework::DataLayout; +template +class GaussianMKLDNNKernel : public paddle::framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + float mean = context.Attr("mean"); + float std = context.Attr("std"); + auto* tensor = context.Output("Out"); + T* data = tensor->mutable_data(context.GetPlace()); + + unsigned int seed = static_cast(context.Attr("seed")); + std::minstd_rand engine; + if (seed == 0) { + seed = std::random_device()(); + } + engine.seed(seed); + std::normal_distribution dist(mean, std); + int64_t size = tensor->numel(); + for (int64_t i = 0; i < size; ++i) { + data[i] = dist(engine); + } + + // The format of output is set as the mkldnn's format + // TODO(@mozga-intel) The format of matrix sets inside the another layers. + tensor->set_layout(DataLayout::kMKLDNN); + tensor->set_format(mkldnn::memory::format::oihw); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(gaussian_random, MKLDNN, ::paddle::platform::CPUPlace, + ops::GaussianMKLDNNKernel); diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc index 815c1bb50988be49ca9996e368a59344c6583d58..1488aab1926b5b4ba7bceed582700f5a11fc6c93 100644 --- a/paddle/fluid/operators/gaussian_random_op.cc +++ b/paddle/fluid/operators/gaussian_random_op.cc @@ -15,6 +15,10 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace operators { @@ -62,9 +66,20 @@ class GaussianRandomOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { + framework::LibraryType library{framework::LibraryType::kPlain}; + framework::DataLayout layout{framework::DataLayout::kAnyLayout}; + +#ifdef PADDLE_WITH_MKLDNN + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } +#endif + return framework::OpKernelType( static_cast(ctx.Attr("dtype")), - ctx.device_context()); + ctx.device_context(), layout, library); } }; @@ -95,7 +110,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { "(int, default 5(FP32)) " "Output data type.") .SetDefault(framework::proto::VarType::FP32); - + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( GaussianRandom Operator. diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc index 4bce2d322d825110a446c9bc5eccdacf0ba3c943..697c239e59d158428ae9ba9f7feded19637dff28 100644 --- a/paddle/fluid/operators/gen_nccl_id_op.cc +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -21,9 +21,8 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/detail/grpc_server.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/platform/nccl_helper.h" namespace paddle { @@ -61,12 +60,18 @@ class GenNCCLIdOp : public framework::OperatorBase { std::vector endpoint_list = Attr>("endpoint_list"); - detail::RPCClient client; + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); + for (auto& ep : endpoint_list) { VLOG(3) << "sending nccl id to " << ep; - client.AsyncSendVariable(ep, dev_ctx, *scope, NCCL_ID_VARNAME); + client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME); + } + client->Wait(); + for (auto& ep : endpoint_list) { + client->AsyncSendBatchBarrier(ep); } - client.Wait(); + client->Wait(); VLOG(3) << "sending completed..."; } @@ -76,10 +81,12 @@ class GenNCCLIdOp : public framework::OperatorBase { // NOTE: Can not use unique_ptr here because the default // deleter will call GRPC Server's base class's dtor and // that will cause a wired crash. - detail::RequestSendHandler rpc_h(true); - detail::AsyncGRPCServer rpc_service(endpoint, 1); - rpc_service.RegisterRPC(detail::kRequestSend, &rpc_h); - rpc_h.SetRPCServer(&rpc_service); + distributed::RequestSendHandler rpc_h(true); + std::unique_ptr rpc_service( + new RPCSERVER_T(endpoint, 1)); + + rpc_service->RegisterRPC(distributed::kRequestSend, &rpc_h); + rpc_h.SetRPCServer(rpc_service.get()); framework::ProgramDesc empty_program; framework::Executor executor(dev_ctx.GetPlace()); @@ -89,12 +96,13 @@ class GenNCCLIdOp : public framework::OperatorBase { rpc_h.SetExecutor(&executor); std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::StartServer, &rpc_service)); - rpc_service.SetCond(detail::kRequestSend); + std::bind(&distributed::RPCServer::StartServer, rpc_service.get())); + + rpc_service->SetCond(distributed::kRequestSend); VLOG(3) << "start getting nccl id from trainer 0..."; - rpc_service.WaitBarrier(detail::kRequestSend); + rpc_service->WaitBarrier(distributed::kRequestSend); VLOG(3) << "got nccl id and stop server..."; - rpc_service.ShutDown(); + rpc_service->ShutDown(); VLOG(3) << "rpc server stopped"; server_thread.join(); } diff --git a/paddle/fluid/operators/get_places_op.cc b/paddle/fluid/operators/get_places_op.cc index eafc364a15fa17cc5107bba737b0b44e712b0bef..db6ff7825690176ded0ab957764ed8411d3cd804 100644 --- a/paddle/fluid/operators/get_places_op.cc +++ b/paddle/fluid/operators/get_places_op.cc @@ -85,7 +85,7 @@ class GetPlacesOpProtoMaker : public framework::OpProtoAndCheckerMaker { .InEnum({"CUDA", "CPU", "AUTO"}) .SetDefault("AUTO"); AddComment(R"DOC( -Returns a list of places based on flags. The list will be used for parallel +Returns a list of places based on arguments. The list will be used for parallel execution. )DOC"); } diff --git a/paddle/fluid/operators/hierarchical_sigmoid_op.cc b/paddle/fluid/operators/hierarchical_sigmoid_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..dadd054b9a6f8d44f4e5832888052bffde34c827 --- /dev/null +++ b/paddle/fluid/operators/hierarchical_sigmoid_op.cc @@ -0,0 +1,167 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/hierarchical_sigmoid_op.h" +#include + +namespace paddle { +namespace operators { + +/** + * Organize the classes into a binary tree. At each node, a sigmoid function + * is used to calculate the probability of belonging to the right branch. + * This idea is from "F. Morin, Y. Bengio (AISTATS 05): + * Hierarchical Probabilistic Neural Network Language Model." + * + * Here we uses a simple way of making the binary tree. + * Assuming the number of classes C = 6, + * The classes are organized as a binary tree in the following way: + * + * @code{.py} + * *-*-*- 2 + * | | |- 3 + * | | + * | |-*- 4 + * | |- 5 + * | + * |-*- 0 + * |- 1 + * @endcode + * + * where * indicates an internal node, and each leaf node represents a class. + * - Node 0 ... C-2 are internal nodes. + * - Node C-1 ... 2C-2 are leaf nodes. + * - Class c is represented by leaf node \f$c+C-1\f$. + * + * We assign an id for each node: + * - the id of root be 0. + * - the left child of a node i is 2*i+1. + * - the right child of a node i is 2*i+2. + * + * It's easy to see that: + * - the parent of node i is \f$\left\lfloor(i-1)/2\right\rfloor\f$. + * - the j-th level ancestor of node i is + * \f$\left\lfloor(i+1)/2^{j+1}\right\rfloor - 1\f$. + * - A node i is a left child of its parent if \f$(i-1)\%2==0\f$. + * + */ + +class HierarchicalSigmoidOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("PreOut"), + "Output(PreOut) should not be null."); + const int64_t batch_size = ctx->GetInputDim("X")[0]; + std::vector output_shape({batch_size, 1}); + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.GetPlace()); + } +}; + +template +class HierarchicalSigmoidOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor, required) The input tensor with shape [N, D], " + "where N is the size of mini-batch, and D is the feature size."); + AddInput("W", + "(Tensor, required), The parameters of hierarchical " + "sigmoid operator, each of them is a 2-D tensor, the shape is" + "[num_classes - 1, D]."); + AddInput("Label", + "(Tensor, required), The labels of training data. It's a" + "tensor with shape [N, 1]."); + AddInput("Bias", + "(Tensor, optional), The bias is a tensor with shape" + "[1, num_classes - 1]."); + AddOutput("Out", + "(Tensor, required) The output of hierarchical sigmoid operator." + "The shape is [N, 1]."); + AddOutput("PreOut", + "(Tensor, required) A intermedia 2-D tensor with shape " + "[batch_size, code_length], where code_length represents the " + "maximum path length from root to leaf nodes.") + .AsIntermediate(); + AddAttr("num_classes", "(int, required), The number of classes") + .SetDefault(2); + AddComment(R"DOC( +The hierarchical sigmoid operator organize the classes into a binary tree. +At each node, a sigmoid function is used to calculate the probability of +belonging to the right branch. This idea is from +"F. Morin, Y. Bengio (AISTATS 05): +Hierarchical Probabilistic Neural Network Language Model." + )DOC"); + } +}; + +class HierarchicalSigmoidGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("PreOut"), + "Input(Preout) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("W")), + "Output(W@Grad should not be null.)"); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X"))); + if (ctx->HasOutput(framework::GradVarName("Bias"))) { + ctx->SetOutputDim(framework::GradVarName("Bias"), + ctx->GetInputDim("Bias")); + } + ctx->SetOutputDim(framework::GradVarName("W"), ctx->GetInputDim("W")); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.GetPlace()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(hierarchical_sigmoid, ops::HierarchicalSigmoidOp, + ops::HierarchicalSigmoidOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(hierarchical_sigmoid_grad, ops::HierarchicalSigmoidGradOp); +REGISTER_OP_CPU_KERNEL( + hierarchical_sigmoid, + ops::HierarchicalSigmoidOpKernel, + ops::HierarchicalSigmoidOpKernel); +REGISTER_OP_CPU_KERNEL( + hierarchical_sigmoid_grad, + ops::HierarchicalSigmoidGradOpKernel, + ops::HierarchicalSigmoidGradOpKernel); diff --git a/paddle/fluid/operators/hierarchical_sigmoid_op.h b/paddle/fluid/operators/hierarchical_sigmoid_op.h new file mode 100644 index 0000000000000000000000000000000000000000..64096a717b12ed231344649f5eb76b7e4b9af4a6 --- /dev/null +++ b/paddle/fluid/operators/hierarchical_sigmoid_op.h @@ -0,0 +1,135 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/clip_op.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/matrix_bit_code.h" +#include "paddle/fluid/platform/transform.h" +namespace paddle { +namespace operators { + +template +using EigenMatrix = framework::EigenMatrix; +using platform::Transform; + +template +class HierarchicalSigmoidOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* w = ctx.Input("W"); + auto* label = ctx.Input("Label"); + auto* bias = ctx.Input("Bias"); + auto* out = ctx.Output("Out"); + auto* pre_out = ctx.Output("PreOut"); + size_t num_classes = static_cast(ctx.Attr("num_classes")); + int64_t code_length = math::FindLastSet(num_classes - 1); + int64_t batch_size = in->dims()[0]; + framework::Tensor sum; + auto& dev_ctx = ctx.template device_context(); + auto* pre_out_data = pre_out->mutable_data( + framework::make_ddim({batch_size, code_length}), ctx.GetPlace()); + auto pre_out_mat = EigenMatrix::From(*pre_out); + // Not all class(leaf) nodes' path lengths equal code_length, thus init as + // 0s can avoid out of path's loss. + math::SetConstant zero; + zero(dev_ctx, pre_out, static_cast(0.0)); + auto& place = *ctx.template device_context().eigen_device(); + math::RowwiseSum row_sum; + math::MatrixBitCodeFunctor bit_code(num_classes, label->data()); + + std::vector sum_dims({batch_size, 1UL}); + sum.mutable_data(framework::make_ddim(sum_dims), ctx.GetPlace()); + auto sum_mat = EigenMatrix::From(sum); + out->mutable_data(ctx.GetPlace()); + auto out_mat = framework::EigenVector::Flatten(*out); + if (bias) { + bit_code.Add(pre_out, *bias); + } + bit_code.Mul(pre_out, *w, *in); + // clip to [-40, 40] + Transform trans; + trans(ctx.template device_context(), pre_out_data, + pre_out_data + pre_out->numel(), pre_out_data, + ClipFunctor(static_cast(-40.0), static_cast(40.0))); + bit_code.Sum(*pre_out, out, static_cast(-1)); + // use softrelu to calculate cross entropy + pre_out_mat.device(place) = (static_cast(1.0) + pre_out_mat.exp()).log(); + row_sum(dev_ctx, *pre_out, &sum); + // TODO(guosheng): Subtract the out of path's loss, since not all + // class(leaf) nodes' path lengths equal code_length. But it won't break the + // gradient check since both have the out of path's loss and will cancel out + // each other. + out_mat.device(place) = sum_mat + out_mat; + } +}; + +template +class HierarchicalSigmoidGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* w = ctx.Input("W"); + auto* in_grad = ctx.Output(framework::GradVarName("X")); + auto* w_grad = ctx.Output(framework::GradVarName("W")); + auto* bias_grad = + ctx.Output(framework::GradVarName("Bias")); + auto* label = ctx.Input("Label"); + auto* pre_out = ctx.Input("PreOut"); + auto* out_grad = + ctx.Input(framework::GradVarName("Out")); + framework::Tensor pre_out_grad; + + pre_out_grad.mutable_data(pre_out->dims(), ctx.GetPlace()); + in_grad->mutable_data(ctx.GetPlace()); + w_grad->mutable_data(ctx.GetPlace()); + auto& dev_ctx = ctx.template device_context(); + math::SetConstant zero; + zero(dev_ctx, in_grad, static_cast(0.0)); + zero(dev_ctx, w_grad, static_cast(0.0)); + + size_t num_classes = static_cast(ctx.Attr("num_classes")); + math::MatrixBitCodeFunctor bit_code(num_classes, label->data()); + + auto& place = *ctx.template device_context().eigen_device(); + auto pre_out_mat = EigenMatrix::From(*pre_out); + auto pre_out_grad_mat = EigenMatrix::From(pre_out_grad); + auto out_grad_mat = EigenMatrix::From(*out_grad); + Eigen::array bcast({{1, static_cast(pre_out_grad.dims()[1])}}); + + // softrelu derivative + pre_out_grad_mat.device(place) = + static_cast(1.0) - static_cast(1.0) / pre_out_mat.exp(); + bit_code.Sub(&pre_out_grad); // the gradient of clip(w * x + b) + pre_out_grad_mat.device(place) = + pre_out_grad_mat * out_grad_mat.broadcast(bcast); + // TODO(guosheng): multiply pre_out_grad with subgradient of clipping to + // be consistent with the clipping in forward. + if (bias_grad) { + bias_grad->mutable_data(ctx.GetPlace()); + zero(dev_ctx, bias_grad, static_cast(0.0)); + bit_code.AddGrad(pre_out_grad, bias_grad); + } + bit_code.MulGradWeight(pre_out_grad, w_grad, *in); + bit_code.MulGradError(pre_out_grad, *w, in_grad); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/im2sequence_op.cc b/paddle/fluid/operators/im2sequence_op.cc index 0669661d225c664010fce97f0a526b62988b92c5..8efd43928aac994c7630a213f6724e8f50abc7e0 100644 --- a/paddle/fluid/operators/im2sequence_op.cc +++ b/paddle/fluid/operators/im2sequence_op.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/im2sequence_op.h" +#include #include namespace paddle { @@ -28,27 +29,18 @@ class Im2SequenceOp : public framework::OperatorWithKernel { "Input(X) of Im2SequenceOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of Im2SequenceOp op should not be null."); - auto in_dim = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(in_dim.size(), 4, "Input(X) format must be 4D tensor, eg., NCHW."); + int img_channels = in_dim[1]; auto kernels = ctx->Attrs().Get>("kernels"); auto strides = ctx->Attrs().Get>("strides"); auto paddings = ctx->Attrs().Get>("paddings"); - int batch_size = in_dim[0]; - int img_channels = in_dim[1]; - int img_height = in_dim[2]; - int img_width = in_dim[3]; - - int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0], - paddings[2], strides[0]); - int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1], - paddings[3], strides[1]); - - ctx->SetOutputDim("Out", {batch_size * output_height * output_width, - img_channels * kernels[0] * kernels[1]}); + ctx->SetOutputDim("Out", + {in_dim[0], img_channels * kernels[0] * kernels[1]}); } }; @@ -61,6 +53,10 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker { "C: channels" "H: height" "W: width"); + AddInput("Y", + "(Tensor) The input tensor of image real size(H, W)." + "2-D with shape [batchsize, 2]") + .AsDispensable(); AddOutput("Out", "(LodTensor) The output data of im2sequence op,"); AddAttr>("kernels", "(vector), the " @@ -73,6 +69,13 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker { "(vector default:{0, 0, 0, 0}), the " "paddings(up_pad, left_pad, down_pad, right_pad)") .SetDefault({0, 0, 0, 0}); + AddAttr>("out_stride", + "the attribute is valid only when input(Y)" + "is not NULL.this attribute represents the" + "scaling of the pic through the CNN" + "(vector dedault:{1,1}),the out_stride" + " (out_stride_height, out_stride_width)") + .SetDefault({1, 1}); AddComment(R"DOC( This op uses kernels to scan images and converts these images to sequences. After expanding, The number of time steps are output_height * output_width @@ -123,7 +126,7 @@ output.data = [[ 6. 2. 8. 3. 2. 4. 6. 3.] [ 7. 1. 7. 9. 2. 1. 3. 5.] [ 5. 7. 2. 4. 1. 3. 9. 0.] [ 7. 9. 4. 8. 3. 5. 0. 8.]] -output.dims = {8, 9} +output.dims = {8, 8} output.lod = [[0, 4, 8]] )DOC"); diff --git a/paddle/fluid/operators/im2sequence_op.h b/paddle/fluid/operators/im2sequence_op.h index d792c68f784d8ffec0eb303a6ab9b59c9f121fa7..4a9942819414d552eb69bd0b30b66aab76a2dbf4 100644 --- a/paddle/fluid/operators/im2sequence_op.h +++ b/paddle/fluid/operators/im2sequence_op.h @@ -13,6 +13,7 @@ limitations under the License. */ #pragma once +#include #include #include "paddle/fluid/framework/data_layout.h" #include "paddle/fluid/framework/eigen.h" @@ -39,50 +40,107 @@ class Im2SequenceKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { const Tensor* in = ctx.Input("X"); LoDTensor* out = ctx.Output("Out"); - out->mutable_data(ctx.GetPlace()); - // TODO(wanghaoshuang): Add layout checker after 'set_layout' - // being available for python API - // PADDLE_ENFORCE_EQ(in->layout(), framework::DataLayout::kNCHW, - // "Input(X) layout must be NCHW"); auto in_dim = in->dims(); int batch_size = in_dim[0]; int img_channels = in_dim[1]; int img_height = in_dim[2]; int img_width = in_dim[3]; - auto kernels = ctx.Attr>("kernels"); auto strides = ctx.Attr>("strides"); auto paddings = ctx.Attr>("paddings"); - int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0], - paddings[2], strides[0]); - int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1], - paddings[3], strides[1]); - - const std::vector dilations({1, 1}); - - auto out_dims = out->dims(); - out->Resize({batch_size, out->numel() / batch_size}); - for (int i = 0; i < batch_size; i++) { - const Tensor src = - in->Slice(i, i + 1).Resize({img_channels, img_height, img_width}); - Tensor dst = out->Slice(i, i + 1).Resize( - {output_height, output_width, img_channels, kernels[0], kernels[1]}); - - math::Im2ColFunctor f; - auto& dev_ctx = ctx.template device_context(); - f(dev_ctx, src, dilations, strides, paddings, &dst); - } - out->Resize(out_dims); - - // set lod information - // TODO(wanghaoshuang): Move this to InferShape - framework::LoD lod(1); - lod[0].reserve(batch_size + 1); - for (int i = 0, offset = 0; i < batch_size + 1; ++i) { + if (ctx.HasInput("Y") && batch_size > 1) { + const Tensor* imgrealsize = ctx.Input("Y"); + auto out_stride = ctx.Attr>("out_stride"); + Tensor cpu_shape_tensor; + TensorCopySync(*imgrealsize, platform::CPUPlace(), &cpu_shape_tensor); + std::vector imgreal_h; + std::vector imgreal_w; + std::vector output_height; + std::vector output_width; + int result = 0; + for (int i = 0; i < batch_size; i++) { + int tmp_real_h = static_cast((cpu_shape_tensor.data())[2 * i]); + int tmp_real_w = + static_cast((cpu_shape_tensor.data())[2 * i + 1]); + if (tmp_real_h % out_stride[0] == 0) { + tmp_real_h = tmp_real_h / out_stride[0]; + } else { + tmp_real_h = tmp_real_h / out_stride[0] + 1; + } + if (tmp_real_w % out_stride[1] == 0) { + tmp_real_w = tmp_real_w / out_stride[1]; + } else { + tmp_real_w = tmp_real_w / out_stride[1] + 1; + } + imgreal_h.push_back(tmp_real_h); + imgreal_w.push_back(tmp_real_w); + output_height.push_back(Im2SeqOutputSize( + imgreal_h[i], kernels[0], paddings[0], paddings[2], strides[0])); + output_width.push_back(Im2SeqOutputSize( + imgreal_w[i], kernels[1], paddings[1], paddings[3], strides[1])); + result += output_height[i] * output_width[i]; + } + + out->mutable_data({result, img_channels * kernels[0] * kernels[1]}, + ctx.GetPlace()); + + const std::vector dilations({1, 1}); + int offset_out = 0; + for (int i = 0; i < batch_size; i++) { + const Tensor src = + in->Slice(i, i + 1).Resize({img_channels, img_height, img_width}); + Tensor dst = out->Slice(offset_out, + offset_out + output_height[i] * output_width[i]) + .Resize({output_height[i], output_width[i], + img_channels, kernels[0], kernels[1]}); + offset_out += output_height[i] * output_width[i]; + + math::Im2ColFunctor f; + auto& dev_ctx = ctx.template device_context(); + f(dev_ctx, src, dilations, strides, paddings, &dst); + } + framework::LoD lod(1); + lod[0].reserve(batch_size + 1); + int offset = 0; + lod[0].push_back(offset); + for (int i = 0; i < batch_size; ++i) { + offset += output_height[i] * output_width[i]; + lod[0].push_back(offset); + } + out->set_lod(lod); + } else { + int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0], + paddings[2], strides[0]); + int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1], + paddings[3], strides[1]); + out->mutable_data({batch_size * output_height * output_width, + img_channels * kernels[0] * kernels[1]}, + ctx.GetPlace()); + const std::vector dilations({1, 1}); + auto out_dims = out->dims(); + out->Resize({batch_size, out->numel() / batch_size}); + for (int i = 0; i < batch_size; i++) { + const Tensor src = + in->Slice(i, i + 1).Resize({img_channels, img_height, img_width}); + Tensor dst = + out->Slice(i, i + 1).Resize({output_height, output_width, + img_channels, kernels[0], kernels[1]}); + + math::Im2ColFunctor f; + auto& dev_ctx = ctx.template device_context(); + f(dev_ctx, src, dilations, strides, paddings, &dst); + } + out->Resize(out_dims); + framework::LoD lod(1); + lod[0].reserve(batch_size + 1); + int offset = 0; lod[0].push_back(offset); - offset += output_height * output_width; + for (int i = 0; i < batch_size; ++i) { + offset += output_height * output_width; + lod[0].push_back(offset); + } + out->set_lod(lod); } - out->set_lod(lod); } }; diff --git a/paddle/fluid/operators/layer_norm_op.cc b/paddle/fluid/operators/layer_norm_op.cc index ab097d31e9ab5eafa788539170e7e405df697625..14ce1da2e97186a50ed8bd52223a500c4c57b328 100644 --- a/paddle/fluid/operators/layer_norm_op.cc +++ b/paddle/fluid/operators/layer_norm_op.cc @@ -62,36 +62,33 @@ class LayerNormOp : public framework::OperatorWithKernel { class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "(LoDTensor) The input tensor."); + AddInput("X", "The input tensor."); AddInput("Scale", - "(Tensor, optional) Scale is a 1-dimensional tensor of size " + "(optional) Scale is a 1-dimensional tensor of size " "H(`begin_norm_axis` splits the tensor(`X`) to a matrix [N,H])." "It is applied to the output.") .AsDispensable(); AddInput("Bias", - "(Tensor, optional) Bias is a 1-dimensional tensor of size " + "(optional) Bias is a 1-dimensional tensor of size " "H(`begin_norm_axis` splits the tensor(`X`) to a matrix [N,H])." "It is applied to the output.") .AsDispensable(); - AddOutput("Y", "(LoDTensor) Result after normalization."); - AddOutput("Mean", "(Tensor) Mean of the current mini batch.") - .AsIntermediate(); - AddOutput("Variance", "(Tensor) Variance of the current mini batch.") + AddOutput("Y", "Result after normalization."); + AddOutput("Mean", "Mean of the current mini batch.").AsIntermediate(); + AddOutput("Variance", "Variance of the current mini batch.") .AsIntermediate(); AddAttr("epsilon", - "(float, default 1e-5) Constant for " - "numerical stability") + "Constant for numerical stability [default 1e-5].") .SetDefault(1e-5) .AddCustomChecker([](const float &epsilon) { PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f, "'epsilon' should be between 0.0 and 0.001."); }); AddAttr("begin_norm_axis", - "(int default:1), the " - "axis of `begin_norm_axis ... Rank(X) - 1` will be " + "the axis of `begin_norm_axis ... Rank(X) - 1` will be " "normalized. `begin_norm_axis` splits the tensor(`X`) to a " - "matrix [N,H].") + "matrix [N,H]. [default 1].") .SetDefault(1) .AddCustomChecker([](const int &begin_norm_axis) { PADDLE_ENFORCE_GT(begin_norm_axis, 0, @@ -99,10 +96,14 @@ class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker { }); AddComment(R"DOC( -Layer Normalization. -Layer Norm has been implemented as discussed in the paper: -https://arxiv.org/abs/1607.06450 -... +Assume feature vectors exist on dimensions +:attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics +along these dimensions for each feature vector :math:`a` with size +:math:`H`, then normalize each feature vector using the corresponding +statistics. After that, apply learnable gain and bias on the normalized +tensor to scale and shift if :attr:`scale` and :attr:`shift` are set. + +Refer to `Layer Normalization `_ )DOC"); } }; diff --git a/paddle/fluid/operators/layer_norm_op.cu b/paddle/fluid/operators/layer_norm_op.cu index 6840e1e08f3d5bc84a05f15e30982c7cfb59680b..0886c41a1b582881faf24f5531d414db4e4db71c 100644 --- a/paddle/fluid/operators/layer_norm_op.cu +++ b/paddle/fluid/operators/layer_norm_op.cu @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -12,8 +12,512 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include "paddle/fluid/operators/layer_norm_op.h" +namespace paddle { +namespace operators { + +inline static int GetDesiredBlockDim(int block_dim) { + const int kMaxBlockDim = 512; + return block_dim >= kMaxBlockDim + ? kMaxBlockDim + : (1 << (static_cast(std::log2f(block_dim)))); +} + +#define FIXED_BLOCK_DIM_CASE_BASE(log2_block_dim, ...) \ + case (1 << (log2_block_dim)): { \ + constexpr auto kBlockDim = (1 << (log2_block_dim)); \ + __VA_ARGS__; \ + } break + +#define FIXED_BLOCK_DIM_CASE(...) \ + FIXED_BLOCK_DIM_CASE_BASE(9, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(8, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(7, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(6, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(5, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(4, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(3, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(2, ##__VA_ARGS__); \ + FIXED_BLOCK_DIM_CASE_BASE(1, ##__VA_ARGS__) + +static __device__ __forceinline__ float real_sqrt(float x) { return sqrtf(x); } +static __device__ __forceinline__ double real_sqrt(double x) { return sqrt(x); } + +template +struct PairForLayerNorm { + __device__ __forceinline__ PairForLayerNorm() {} + __device__ __forceinline__ PairForLayerNorm(const T &first, const T &second) + : first_(first), second_(second) {} + + T first_; + T second_; +}; + +template +struct PairForLayerNormAddFunctor { + __device__ __forceinline__ PairForLayerNorm operator()( + const PairForLayerNorm &p1, const PairForLayerNorm &p2) { + return PairForLayerNorm(p1.first_ + p2.first_, p1.second_ + p2.second_); + } +}; + +template +__global__ void LayerNormForward(const T *x, const T *scale, const T *bias, + T *y, T *mean, T *var, float epsilon, + int feature_size) { + using BlockReduce = cub::BlockReduce, BlockDim>; + __shared__ typename BlockReduce::TempStorage temp_storage; + + int beg_idx = blockIdx.x * feature_size + threadIdx.x; + int end_idx = (blockIdx.x + 1) * feature_size; + + // Step 1: Reduce to calculate mean and var + T mean_val = static_cast(0); + T var_val = static_cast(0); + for (int i = beg_idx; i < end_idx; i += BlockDim) { + T tmp = x[i]; + mean_val += tmp; + var_val += (tmp * tmp); + } + auto pair = BlockReduce(temp_storage) + .Reduce(PairForLayerNorm(mean_val, var_val), + PairForLayerNormAddFunctor()); + if (threadIdx.x == 0) { + auto tmp = pair.first_ / feature_size; + mean[blockIdx.x] = tmp; + var[blockIdx.x] = pair.second_ / feature_size - tmp * tmp; + } + __syncthreads(); + mean_val = mean[blockIdx.x]; + var_val = static_cast(real_sqrt(var[blockIdx.x] + epsilon)); + + // Step 2: Calculate y + if (scale != nullptr) { + if (bias != nullptr) { + for (int i = beg_idx, j = threadIdx.x; i < end_idx; + i += BlockDim, j += BlockDim) { + y[i] = scale[j] * (x[i] - mean_val) / var_val + bias[j]; + } + } else { + for (int i = beg_idx, j = threadIdx.x; i < end_idx; + i += BlockDim, j += BlockDim) { + y[i] = scale[j] * (x[i] - mean_val) / var_val; + } + } + } else { // scale == nullptr + if (bias != nullptr) { + for (int i = beg_idx, j = threadIdx.x; i < end_idx; + i += BlockDim, j += BlockDim) { + y[i] = (x[i] - mean_val) / var_val + bias[j]; + } + } else { + for (int i = beg_idx, j = threadIdx.x; i < end_idx; + i += BlockDim, j += BlockDim) { + y[i] = (x[i] - mean_val) / var_val; + } + } + } +} + +// Make sure that d_scale != nullptr && d_bias != nullptr +// Since d_scale != nullptr, scale would not be nullptr +template +__global__ void LayerNormBackwardGradientAll(const T *x, const T *d_y, + T *d_scale, T *d_bias, T *d_x, + const T *mean, const T *var, + const T *scale, float epsilon, + int batch_size, int feature_size) { + using BlockReduce = cub::BlockReduce, BlockDim>; + __shared__ typename BlockReduce::TempStorage temp_storage; + + int beg_idx = threadIdx.x * feature_size + blockIdx.x; + int end_idx = batch_size * feature_size + blockIdx.x; + int stride = BlockDim * feature_size; + + T d_scale_partial = 0, d_bias_partial = 0; + + for (int i = beg_idx; i < end_idx; i += stride) { + int row_idx = i / feature_size; + auto var_val = static_cast(real_sqrt(var[row_idx] + epsilon)); + d_scale_partial += d_y[i] * (x[i] - mean[row_idx]) / var_val; + d_bias_partial += d_y[i]; + if (HasDx) { + d_x[i] = d_y[i] * scale[blockIdx.x] / var_val; + } + } + + auto pair = BlockReduce(temp_storage) + .Reduce(PairForLayerNorm(d_scale_partial, d_bias_partial), + PairForLayerNormAddFunctor()); + + if (threadIdx.x == 0) { + d_scale[blockIdx.x] = pair.first_; + d_bias[blockIdx.x] = pair.second_; + } +} + +// Make sure that there is only one true expression: d_scale != nullptr +// or d_bias != nullptr +// Notice: scale may be nullptr +template +__global__ void LayerNormBackwardGradientScaleOrBias( + const T *x, const T *d_y, T *d_scale, T *d_bias, T *d_x, const T *mean, + const T *var, const T *scale, float epsilon, int batch_size, + int feature_size) { + using BlockReduce = cub::BlockReduce; + __shared__ typename BlockReduce::TempStorage temp_storage; + int beg_idx = threadIdx.x * feature_size + blockIdx.x; + int end_idx = batch_size * feature_size + blockIdx.x; + int stride = BlockDim * feature_size; + T d_scale_or_d_bias_partial = 0; + + for (int i = beg_idx; i < end_idx; i += stride) { + int row_idx = i / feature_size; + auto var_val = static_cast(real_sqrt(var[row_idx] + epsilon)); + if (HasDScale) { + d_scale_or_d_bias_partial += d_y[i] * (x[i] - mean[row_idx]) / var_val; + } else { // d_bias != nullptr + d_scale_or_d_bias_partial += d_y[i]; + } + + if (HasDx) { + if (scale != nullptr) { + d_x[i] = d_y[i] * scale[blockIdx.x] / var_val; + } else { + d_x[i] = d_y[i] / var_val; + } + } + } + + d_scale_or_d_bias_partial = + BlockReduce(temp_storage).Reduce(d_scale_or_d_bias_partial, cub::Sum()); + + if (threadIdx.x == 0) { + if (HasDScale) { + d_scale[blockIdx.x] = d_scale_or_d_bias_partial; + } else { + d_bias[blockIdx.x] = d_scale_or_d_bias_partial; + } + } +} + +template +__global__ void LayerNormBackwardPostProcessToCalculateDX(const T *x, T *d_x, + const T *mean, + const T *var, + float epsilon, + int feature_size) { + using BlockReduce = cub::BlockReduce, BlockDim>; + __shared__ typename BlockReduce::TempStorage temp_storage; + __shared__ T d_x_reduce_tmp[2]; + + int beg_idx = blockIdx.x * feature_size + threadIdx.x; + int end_idx = (blockIdx.x + 1) * feature_size; + + T block_mean = mean[blockIdx.x]; + T block_var = var[blockIdx.x]; + T d_x_mean_partial = 0, d_x_var_partial = 0; + for (int i = beg_idx; i < end_idx; i += BlockDim) { + d_x_mean_partial += d_x[i]; + d_x_var_partial += d_x[i] * (x[i] - block_mean); + } + + auto pair = + BlockReduce(temp_storage) + .Reduce(PairForLayerNorm(d_x_mean_partial, d_x_var_partial), + PairForLayerNormAddFunctor()); + + if (threadIdx.x == 0) { + d_x_reduce_tmp[0] = pair.first_ / feature_size; + d_x_reduce_tmp[1] = pair.second_ / (feature_size * (block_var + epsilon)); + } + __syncthreads(); + + d_x_mean_partial = d_x_reduce_tmp[0]; + d_x_var_partial = d_x_reduce_tmp[1]; + for (int i = beg_idx; i < end_idx; i += BlockDim) { + d_x[i] -= d_x_mean_partial; + d_x[i] -= (x[i] - block_mean) * d_x_var_partial; + } +} + +// Here, we only calculate d_x +template +__global__ void LayerNormBackwardGradientOnlyDX(const T *x, const T *d_y, + T *d_x, const T *mean, + const T *var, const T *scale, + float epsilon, + int feature_size) { + using BlockReduce = cub::BlockReduce, BlockDim>; + __shared__ typename BlockReduce::TempStorage temp_storage; + __shared__ T d_x_reduce_tmp[2]; + + int beg_idx = blockIdx.x * feature_size + threadIdx.x; + int end_idx = (blockIdx.x + 1) * feature_size; + + T block_mean = mean[blockIdx.x], block_var = var[blockIdx.x]; + T d_x_mean_partial = 0, d_x_var_partial = 0; + for (int i = beg_idx; i < end_idx; i += BlockDim) { + auto var_val = static_cast(real_sqrt(block_var + epsilon)); + if (scale != nullptr) { + int col_idx = i % feature_size; + d_x[i] = d_y[i] * scale[col_idx] / var_val; + } else { + d_x[i] = d_y[i] / var_val; + } + d_x_mean_partial += d_x[i]; + d_x_var_partial += d_x[i] * (x[i] - block_mean); + } + + auto pair = + BlockReduce(temp_storage) + .Reduce(PairForLayerNorm(d_x_mean_partial, d_x_var_partial), + PairForLayerNormAddFunctor()); + + if (threadIdx.x == 0) { + d_x_reduce_tmp[0] = pair.first_ / feature_size; + d_x_reduce_tmp[1] = pair.second_ / (feature_size * (block_var + epsilon)); + } + __syncthreads(); + + d_x_mean_partial = d_x_reduce_tmp[0]; + d_x_var_partial = d_x_reduce_tmp[1]; + for (int i = beg_idx; i < end_idx; i += BlockDim) { + d_x[i] -= d_x_mean_partial; + d_x[i] -= (x[i] - block_mean) * d_x_var_partial; + } +} + +template +__global__ void LayerNormBackwardWhenBatchSizeIsOne( + const T *x, const T *d_y, T *d_x, T *d_scale, T *d_bias, const T *mean, + const T *var, const T *scale, float epsilon, int feature_size) { + int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < feature_size) { + auto var_val = static_cast(real_sqrt(var[idx] + epsilon)); + if (d_x != nullptr) { + if (d_scale == nullptr) { + d_x[idx] = d_y[idx] / var_val; + } else { + d_x[idx] = d_y[idx] * scale[idx] / var_val; + } + } + + if (d_scale != nullptr) { + d_scale[idx] = d_y[idx] * (x[idx] - mean[idx]) / var_val; + } + + if (d_bias != nullptr) d_bias[idx] = d_y[idx]; + } +} + +template +static void LayerNormBackward(const T *x, const T *d_y, const T *scale, + const T *mean, const T *var, T *d_x, T *d_scale, + T *d_bias, float epsilon, int batch_size, + int feature_size, cudaStream_t stream) { + const int kMaxBlockDim = 512; + int gradient_flag = ((d_x != nullptr ? 1 : 0) << 2) | + ((d_scale != nullptr ? 1 : 0) << 1) | + ((d_bias != nullptr ? 1 : 0)); + if (gradient_flag == 0) return; + + if (batch_size == 1) { + LayerNormBackwardWhenBatchSizeIsOne< + T><<<(feature_size + kMaxBlockDim - 1) / kMaxBlockDim, kMaxBlockDim, 0, + stream>>>(x, d_y, d_x, d_scale, d_bias, mean, var, scale, epsilon, + feature_size); + + if (d_x != nullptr) { + switch (GetDesiredBlockDim(feature_size)) { + FIXED_BLOCK_DIM_CASE(LayerNormBackwardPostProcessToCalculateDX< + T, kBlockDim><<<1, kBlockDim, 0, stream>>>( + x, d_x, mean, var, epsilon, feature_size)); + } + } + return; + } + + auto block_dim = GetDesiredBlockDim(batch_size); + switch (gradient_flag) { + case 1: // d_x == nulptr, d_scale == nullptr, d_bias != nullptr + switch (block_dim) { + FIXED_BLOCK_DIM_CASE(LayerNormBackwardGradientScaleOrBias< + T, kBlockDim, false, + false><<>>( + x, d_y, d_scale, d_bias, d_x, mean, var, scale, epsilon, batch_size, + feature_size)); + } + break; + case 2: // d_x == nullptr, d_scale != nullptr, d_bias == nullptr + switch (block_dim) { + FIXED_BLOCK_DIM_CASE(LayerNormBackwardGradientScaleOrBias< + T, kBlockDim, false, + true><<>>( + x, d_y, d_scale, d_bias, d_x, mean, var, scale, epsilon, batch_size, + feature_size)); + } + break; + case 3: // d_x == nullptr, d_scale != nulptr, d_bias != nullptr + switch (block_dim) { + FIXED_BLOCK_DIM_CASE( + LayerNormBackwardGradientAll< + T, kBlockDim, false><<>>( + x, d_y, d_scale, d_bias, d_x, mean, var, scale, epsilon, + batch_size, feature_size)); + } + break; + case 4: // d_x != nullptr, d_scale == nullptr, d_bias == nullptr + switch (GetDesiredBlockDim(feature_size)) { + FIXED_BLOCK_DIM_CASE( + LayerNormBackwardGradientOnlyDX< + T, kBlockDim><<>>( + x, d_y, d_x, mean, var, scale, epsilon, feature_size)); + } + break; + case 5: // d_x != nulptr, d_scale == nullptr, d_bias != nullptr + switch (block_dim) { + FIXED_BLOCK_DIM_CASE(LayerNormBackwardGradientScaleOrBias< + T, kBlockDim, true, + false><<>>( + x, d_y, d_scale, d_bias, d_x, mean, var, scale, epsilon, batch_size, + feature_size)); + } + switch (GetDesiredBlockDim(feature_size)) { + FIXED_BLOCK_DIM_CASE( + LayerNormBackwardPostProcessToCalculateDX< + T, kBlockDim><<>>( + x, d_x, mean, var, epsilon, feature_size)); + } + break; + case 6: // d_x != nullptr, d_scale != nullptr, d_bias == nullptr + switch (block_dim) { + FIXED_BLOCK_DIM_CASE(LayerNormBackwardGradientScaleOrBias< + T, kBlockDim, true, + true><<>>( + x, d_y, d_scale, d_bias, d_x, mean, var, scale, epsilon, batch_size, + feature_size)); + } + switch (GetDesiredBlockDim(feature_size)) { + FIXED_BLOCK_DIM_CASE( + LayerNormBackwardPostProcessToCalculateDX< + T, kBlockDim><<>>( + x, d_x, mean, var, epsilon, feature_size)); + } + break; + case 7: // d_x != nullptr, d_scale != nullptr, d_bias != nullptr + switch (block_dim) { + FIXED_BLOCK_DIM_CASE( + LayerNormBackwardGradientAll< + T, kBlockDim, true><<>>( + x, d_y, d_scale, d_bias, d_x, mean, var, scale, epsilon, + batch_size, feature_size)); + } + switch (GetDesiredBlockDim(feature_size)) { + FIXED_BLOCK_DIM_CASE( + LayerNormBackwardPostProcessToCalculateDX< + T, kBlockDim><<>>( + x, d_x, mean, var, epsilon, feature_size)); + } + break; + default: + break; + } +} + +template +class LayerNormKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const float epsilon = ctx.Attr("epsilon"); + auto *scale = ctx.Input("Scale"); + auto *bias = ctx.Input("Bias"); + auto *x = ctx.Input("X"); + + auto *y = ctx.Output("Y"); + auto *mean = ctx.Output("Mean"); + auto *var = ctx.Output("Variance"); + const auto begin_norm_axis = ctx.Attr("begin_norm_axis"); + + const auto x_dims = x->dims(); + auto *x_data = x->data(); + auto *y_data = y->mutable_data(ctx.GetPlace()); + auto *mean_data = mean->mutable_data(ctx.GetPlace()); + auto *var_data = var->mutable_data(ctx.GetPlace()); + auto *scale_data = (scale == nullptr ? nullptr : scale->data()); + auto *bias_data = (bias == nullptr ? nullptr : bias->data()); + + auto matrix_dim = framework::flatten_to_2d(x_dims, begin_norm_axis); + int batch_size = static_cast(matrix_dim[0]); + int feature_size = static_cast(matrix_dim[1]); + + auto stream = ctx.cuda_device_context().stream(); + + switch (GetDesiredBlockDim(feature_size)) { + FIXED_BLOCK_DIM_CASE( + LayerNormForward<<>>( + x_data, scale_data, bias_data, y_data, mean_data, var_data, + epsilon, feature_size)); + default: + PADDLE_THROW( + "Product from begin_norm_axis to end must be larger than 1"); + break; + } + } +}; + +template +class LayerNormGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const float epsilon = ctx.Attr("epsilon"); + // d_x, d_scale, d_bias may be nullptr + auto *d_x = ctx.Output(framework::GradVarName("X")); + auto *d_scale = ctx.Output(framework::GradVarName("Scale")); + auto *d_bias = ctx.Output(framework::GradVarName("Bias")); + + auto *x = ctx.Input("X"); + auto *mean = ctx.Input("Mean"); + auto *var = ctx.Input("Variance"); + auto *scale = ctx.Input("Scale"); + auto *d_y = ctx.Input(framework::GradVarName("Y")); + + auto *x_data = x->data(); + auto *d_y_data = d_y->data(); + auto *mean_data = mean->data(); + auto *var_data = var->data(); + auto *scale_data = (scale == nullptr ? nullptr : scale->data()); + auto *d_scale_data = + (d_scale == nullptr ? nullptr + : d_scale->mutable_data(ctx.GetPlace())); + auto *d_bias_data = + (d_bias == nullptr ? nullptr : d_bias->mutable_data(ctx.GetPlace())); + auto *d_x_data = + (d_x == nullptr ? nullptr : d_x->mutable_data(ctx.GetPlace())); + + const auto &x_dims = x->dims(); + const auto begin_norm_axis = ctx.Attr("begin_norm_axis"); + auto matrix_dim = framework::flatten_to_2d(x_dims, begin_norm_axis); + int batch_size = static_cast(matrix_dim[0]); + int feature_size = static_cast(matrix_dim[1]); + + auto stream = ctx.cuda_device_context().stream(); + + LayerNormBackward(x_data, d_y_data, scale_data, mean_data, var_data, + d_x_data, d_scale_data, d_bias_data, epsilon, + batch_size, feature_size, stream); + } +}; + +#undef FIXED_BLOCK_DIM_CASE_BASE +#undef FIXED_BLOCK_DIM_CASE +} // namespace operators +} // namespace paddle + namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( layer_norm, diff --git a/paddle/fluid/operators/linear_chain_crf_op.cc b/paddle/fluid/operators/linear_chain_crf_op.cc index e38525cd7f44de020f364ffd16e71a439048347f..ea1ca7f59db22bee973a8827a88e2fb80265fa51 100644 --- a/paddle/fluid/operators/linear_chain_crf_op.cc +++ b/paddle/fluid/operators/linear_chain_crf_op.cc @@ -67,8 +67,6 @@ class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker { "mini-batch. Note: S is equal to the sequence number in a mini-batch. " "The output is no longer a LoDTensor."); AddComment(R"DOC( -LinearChainCRF Operator. - Conditional Random Field defines an undirected probabilistic graph with nodes denoting random variables and edges denoting dependencies between these variables. CRF learns the conditional probability $P(Y|X)$, where @@ -86,6 +84,7 @@ CRF. Please refer to http://www.cs.columbia.edu/~mcollins/fb.pdf and http://cseweb.ucsd.edu/~elkan/250Bwinter2012/loglinearCRFs.pdf for details. Equation: + 1. Denote Input(Emission) to this operator as $x$ here. 2. The first D values of Input(Transition) to this operator are for starting weights, denoted as $a$ here. @@ -108,6 +107,7 @@ Finally, the linear chain CRF operator outputs the logarithm of the conditional likelihood of each training sample in a mini-batch. NOTE: + 1. The feature function for a CRF is made up of the emission features and the transition features. The emission feature weights are NOT computed in this operator. They MUST be computed first before this operator is called. diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 71e75c25321812c849e205460217b174d80654be..f196e18fe122af9536230752096a2d90de8ab527 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -19,15 +19,17 @@ limitations under the License. */ #include // NOLINT #include -#include "paddle/fluid/operators/detail/grpc_server.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "gflags/gflags.h" + +#include "paddle/fluid/operators/detail/macros.h" + +#include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" -#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -void RunServer(std::shared_ptr service) { +void RunServer(std::shared_ptr service) { service->StartServer(); VLOG(4) << "RunServer thread end"; } @@ -60,6 +62,8 @@ static void ParallelExecuteBlocks( framework::Async([&executor, &prepared, &program, &scope, idx]() { int run_block = idx; // thread local try { + VLOG(3) << "running server block: " << run_block + << "pointer: " << prepared[run_block].get(); executor->RunPreparedContext(prepared[run_block].get(), scope); } catch (const std::exception &e) { LOG(ERROR) << "run sub program error " << e.what(); @@ -89,37 +93,51 @@ void ListenAndServOp::SavePort() const { rpc_service_->SavePort(); } -void ListenAndServOp::RunSyncLoop(framework::Executor *executor, - framework::ProgramDesc *program, - framework::Scope *recv_scope, - framework::BlockDesc *prefetch_block) const { +static int64_t GetTimestamp() { + struct timeval tp; + gettimeofday(&tp, NULL); + return tp.tv_sec * 1000 + tp.tv_usec / 1000; +} + +void ListenAndServOp::RunSyncLoop( + framework::Executor *executor, framework::ProgramDesc *program, + framework::Scope *recv_scope, + const std::vector &prefetch_block_id_list, + const int checkpoint_point_block_id) const { + VLOG(2) << "RunSyncLoop"; size_t num_blocks = program->Size(); + auto optimize_blocks = + Attr>(kOptimizeBlocks); PADDLE_ENFORCE_GE(num_blocks, 2, "server program should have at least 2 blocks"); - std::vector block_list; - for (size_t blkid = 1; blkid < num_blocks; ++blkid) { - block_list.push_back(blkid); + // Prepare all the server block + std::vector optimize_blocks_list; + for (size_t i = 1; i < program->Size(); ++i) { + optimize_blocks_list.push_back(i); } - auto optimize_prepared = executor->Prepare(*program, block_list); - // Insert placeholder for block0 which holds current op itself. + auto optimize_prepared = executor->Prepare(*program, optimize_blocks_list); + // Insert placeholder for block0 which holds current op itself, + // NOTE the first block in `optimize_prepared` should never be ran. optimize_prepared.insert( optimize_prepared.begin(), std::shared_ptr(nullptr)); + // Trainers will get all parameters from pserver in the + // startup program, so we will wait RequestGet first + rpc_service_->SetCond(distributed::kRequestGet); + rpc_service_->WaitBarrier(distributed::kRequestGet); rpc_service_->ResetBarrierCounter(); - // Record received sparse variables, so that - // we could reset those after execute optimize program - std::vector sparse_vars; while (true) { + rpc_service_->Profiler().OneStep(); // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. - rpc_service_->SetCond(detail::kRequestSend); - rpc_service_->WaitBarrier(detail::kRequestSend); + rpc_service_->SetCond(distributed::kRequestSend); + rpc_service_->WaitBarrier(distributed::kRequestSend); if (rpc_service_->IsExit()) { LOG(WARNING) << "get exit!rpc_processor break!"; - rpc_service_->SetCond(detail::kRequestGet); + rpc_service_->SetCond(distributed::kRequestGet); break; } @@ -127,43 +145,39 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, // and this will still work. // The optimize blocks which have the same parent ID would run parallel // TODO(Yancey1989): need to use ParallelExecutor for future - int32_t last_parent_blkid = program->Block(1).Parent(); + int32_t last_parent_blkid = optimize_blocks[0]->Parent(); std::vector parallel_blkids; - parallel_blkids.push_back(1); - double ts = detail::GetTimestamp(); - for (size_t blkid = 2; blkid < num_blocks; ++blkid) { - if (blkid != static_cast(prefetch_block->ID())) { - if (program->Block(blkid).Parent() != last_parent_blkid) { - ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, - program, recv_scope); - parallel_blkids.clear(); - last_parent_blkid = program->Block(blkid).Parent(); - } - parallel_blkids.push_back(blkid); + parallel_blkids.push_back(optimize_blocks[0]->ID()); + double ts = GetTimestamp(); + for (size_t i = 1; i < optimize_blocks.size(); ++i) { + // skip the first optimize block because it is already in the + // parallel_blkids. + int blkid = optimize_blocks[i]->ID(); + if (program->Block(blkid).Parent() != last_parent_blkid) { + ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, + program, recv_scope); + parallel_blkids.clear(); + last_parent_blkid = program->Block(blkid).Parent(); } + parallel_blkids.push_back(blkid); } ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, program, recv_scope); - VLOG(2) << "run all blocks spent " << detail::GetTimestamp() - ts << "(ms)"; - - // Reset the received sparse variables, the sum operator would not - // sum the input sparse variables which rows is empty at the next - // mini-batch. - // TODO(Yancey1989): move the reset action into an operator, we couldn't - // have any hide logic in the operator. - for (framework::Variable *var : sparse_vars) { - var->GetMutable()->mutable_rows()->clear(); - } + VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)"; - rpc_service_->SetCond(detail::kRequestGet); - rpc_service_->WaitBarrier(detail::kRequestGet); + rpc_service_->SetCond(distributed::kRequestGet); + rpc_service_->WaitBarrier(distributed::kRequestGet); rpc_service_->ResetBarrierCounter(); + // reset received sparse vars to avoid reuse it in the next mini-batch + dynamic_cast(request_send_handler_.get()) + ->ResetSparseVarRecorder(); } // while(true) } void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, - framework::ProgramDesc *program) const { - VLOG(3) << "RunAsyncLoop in"; + framework::ProgramDesc *program, + framework::Scope *recv_scope) const { + VLOG(2) << "RunAsyncLoop"; // grad name to block id std::unordered_map grad_to_block_id; std::unordered_map id_to_grad; @@ -190,6 +204,10 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, block_list.push_back(blkid); } auto optimize_prepared = executor->Prepare(*program, block_list); + // execute global block if needed + if (block_list[0] == 1 && id_to_grad.count(1) == 0) { + executor->RunPreparedContext(optimize_prepared[0].get(), recv_scope); + } std::unordered_map> grad_to_prepared_ctx; @@ -201,10 +219,9 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, request_get_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); request_prefetch_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); - VLOG(3) << "RunAsyncLoop into while"; while (true) { if (rpc_service_->IsExit()) { - LOG(INFO) << "get exit!rpc_processor break!"; + VLOG(4) << "get exit!rpc_processor break!"; break; } @@ -212,19 +229,22 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, } // while(true) } -static void FillRequestCtx(detail::RequestHandler *h, framework::Scope *scope, - platform::DeviceContext *dev_ctx, - framework::Executor *executor, - framework::ProgramDesc *program, - framework::ExecutorPrepareContext *prefetch_ctx, - detail::RPCServer *rpc_server) { +static void FillRequestCtx( + distributed::RequestHandler *h, framework::Scope *scope, + platform::DeviceContext *dev_ctx, framework::Executor *executor, + framework::ProgramDesc *program, + std::unordered_map> + *prefetch_ctx, + std::shared_ptr checkpoint_ctx, + distributed::RPCServer *rpc_server) { h->SetScope(scope); h->SetDevCtx(dev_ctx); h->SetExecutor(executor); h->SetProgram(program); - h->SetPrefetchPreparedCtx(std::move( - std::unique_ptr(prefetch_ctx))); + h->SetPrefetchPreparedCtx(prefetch_ctx); h->SetRPCServer(rpc_server); + h->SetCheckpointNotifyPreparedCtx(checkpoint_ctx); } void ListenAndServOp::RunImpl(const framework::Scope &scope, @@ -240,38 +260,83 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, PADDLE_ENFORCE(!rpc_service_); std::string endpoint = Attr("endpoint"); + int checkpoint_block_id = Attr(kCheckpointBlockId); - LOG(INFO) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in - << ", end_point:" << endpoint; + VLOG(4) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in + << ", end_point:" << endpoint + << ", checkpoint_block_id: " << checkpoint_block_id; - // request_handler_.reset(new detail::GRPCRequestSendHandler(sync_mode)); - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, fan_in)); - request_send_handler_.reset(new detail::RequestSendHandler(sync_mode)); - request_get_handler_.reset(new detail::RequestGetHandler(sync_mode)); - request_prefetch_handler_.reset( - new detail::RequestPrefetchHandler(sync_mode)); + rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in)); - rpc_service_->RegisterRPC(detail::kRequestSend, request_send_handler_.get()); - rpc_service_->RegisterRPC(detail::kRequestGet, request_get_handler_.get()); - rpc_service_->RegisterRPC(detail::kRequestPrefetch, + request_send_handler_.reset(new distributed::RequestSendHandler(sync_mode)); + request_get_handler_.reset(new distributed::RequestGetHandler(sync_mode)); + request_prefetch_handler_.reset( + new distributed::RequestPrefetchHandler(sync_mode)); + request_checkpoint_handler_.reset(new distributed::RequestCheckpointHandler( + sync_mode, checkpoint_block_id)); + + rpc_service_->RegisterRPC(distributed::kRequestSend, + request_send_handler_.get()); + rpc_service_->RegisterRPC(distributed::kRequestGet, + request_get_handler_.get()); + rpc_service_->RegisterRPC(distributed::kRequestPrefetch, request_prefetch_handler_.get()); - - auto *optimize_block = Attr(kOptimizeBlock); - auto *prefetch_block = Attr(kPrefetchBlock); - auto *program = optimize_block->Program(); + rpc_service_->RegisterRPC(distributed::kRequestCheckpoint, + request_checkpoint_handler_.get()); + + auto optimize_blocks = + Attr>(kOptimizeBlocks); + PADDLE_ENFORCE(optimize_blocks.size() >= 1, + "optimize blocks should be 1 at least on the pserver side."); + auto *program = optimize_blocks[0]->Program(); framework::Executor executor(dev_place); + std::shared_ptr ckpt_pre_context = nullptr; + if (checkpoint_block_id != -1) { + auto ctx = executor.Prepare(*program, checkpoint_block_id); + // see: https://stackoverflow.com/a/14856553 + ckpt_pre_context = std::move(ctx); + } + // prepare for prefetch - VLOG(3) << "prefetch block id is " << prefetch_block->ID(); - auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID()); + std::vector prefetch_block_id_list; + std::unordered_map block_id_to_prefetch_var_name; + + auto prefetch_var_name_to_block_id_str = + Attr>(kPrefetchVarNameToBlockId); + for (const auto &prefetch_var_name_and_id : + prefetch_var_name_to_block_id_str) { + std::vector pieces; + split(prefetch_var_name_and_id, ':', &pieces); + VLOG(3) << "after split, prefetch_var = " << pieces[0] + << ", id=" << pieces[1]; + PADDLE_ENFORCE_EQ(pieces.size(), 2); + + int block_id = std::stoi(pieces[1]); + prefetch_block_id_list.push_back(block_id); + block_id_to_prefetch_var_name[block_id] = pieces[0]; + } - auto f = std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope, - &dev_ctx, &executor, program, prefetch_prepared.release(), - rpc_service_.get()); + auto prefetch_prepared = executor.Prepare(*program, prefetch_block_id_list); + + std::unordered_map> + prefetch_var_name_to_prepared_ctx; + for (size_t i = 0; i < prefetch_block_id_list.size(); ++i) { + auto block_id = prefetch_block_id_list[i]; + auto prefetch_var_name = block_id_to_prefetch_var_name[block_id]; + prefetch_var_name_to_prepared_ctx[prefetch_var_name] = prefetch_prepared[i]; + } + + auto f = + std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope, &dev_ctx, + &executor, program, &prefetch_var_name_to_prepared_ctx, + ckpt_pre_context, rpc_service_.get()); f(request_send_handler_.get()); f(request_get_handler_.get()); f(request_prefetch_handler_.get()); + f(request_checkpoint_handler_.get()); // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); @@ -285,9 +350,10 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, // Write to a file of server selected port for python use. SavePort(); if (sync_mode) { - RunSyncLoop(&executor, program, &recv_scope, prefetch_block); + RunSyncLoop(&executor, program, &recv_scope, prefetch_block_id_list, + checkpoint_block_id); } else { - RunAsyncLoop(&executor, program); + RunAsyncLoop(&executor, program, &recv_scope); } } @@ -309,17 +375,23 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { "a map from grad name to it's optimize block id") .SetDefault({}); AddAttr("sync_mode", "if works at sync_mode or not").SetDefault(true); - AddAttr(kOptimizeBlock, - "BlockID to run on server side."); - AddAttr(kPrefetchBlock, - "prefetch block to run on server side."); + AddAttr>( + kOptimizeBlocks, "Optimize blocks to run on server side.") + .SetDefault({}); + AddAttr>(kPrefetchVarNameToBlockId, + "prefetch blocks to run on server side.") + .SetDefault({}); AddAttr("Fanin", "How many clients send to this server.") .SetDefault(1); + AddAttr(kCheckpointBlockId, + "BolckID to run save checkpoint on pserer.") + .SetDefault(-1); } }; void SignalHandler::StopAndExit(int signal_num) { - VLOG(3) << "Catch interrupt signal: " << signal_num << ", program will exit"; + // Do not use VLOG here for the device for printing maybe already released. + // exit will release interal allocated resoureces. exit(0); } diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 87952cb0e683596b2b0395890b6e25b15f74d7e2..978969cc515c7954b59f2bf7a4f2c0e1b13f9bc0 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -18,21 +18,23 @@ limitations under the License. */ #include #include #include +#include #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/request_handler.h" -#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" namespace paddle { namespace operators { -constexpr char kOptimizeBlock[] = "OptimizeBlock"; -constexpr char kPrefetchBlock[] = "PrefetchBlock"; +constexpr char kOptimizeBlocks[] = "optimize_blocks"; +constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id"; +constexpr char kCheckpointBlockId[] = "checkpint_block_id"; -void RunServer(std::shared_ptr service); +void RunServer(std::shared_ptr service); class ListenAndServOp : public framework::OperatorBase { public: @@ -46,10 +48,12 @@ class ListenAndServOp : public framework::OperatorBase { void RunSyncLoop(framework::Executor* executor, framework::ProgramDesc* program, framework::Scope* recv_scope, - framework::BlockDesc* prefetch_block) const; + const std::vector& prefetch_block_id_list, + const int checkpoint_point_block_id) const; void RunAsyncLoop(framework::Executor* executor, - framework::ProgramDesc* program) const; + framework::ProgramDesc* program, + framework::Scope* recv_scope) const; void SavePort() const; @@ -61,10 +65,13 @@ class ListenAndServOp : public framework::OperatorBase { const platform::Place& dev_place) const override; protected: - mutable std::shared_ptr rpc_service_; - mutable std::shared_ptr request_send_handler_; - mutable std::shared_ptr request_get_handler_; - mutable std::shared_ptr request_prefetch_handler_; + mutable std::shared_ptr rpc_service_; + mutable std::shared_ptr request_send_handler_; + mutable std::shared_ptr request_get_handler_; + mutable std::shared_ptr + request_prefetch_handler_; + mutable std::shared_ptr + request_checkpoint_handler_; mutable std::shared_ptr server_thread_; }; diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index 93f45cff8a26201b1fbb1c44141e125a67c44037..27e26cb1b5c1e831f05dac299489628b92eaa58c 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -31,9 +31,8 @@ class LoadOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { - auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place); - platform::RecordEvent record_event(Type(), dev_ctx); - + // FIXME(yuyang18): We save variable to local file now, but we should change + // it to save an output stream. auto filename = Attr("file_path"); std::ifstream fin(filename); PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s for load op", @@ -44,9 +43,25 @@ class LoadOp : public framework::OperatorBase { PADDLE_ENFORCE(out_var != nullptr, "Output variable %s cannot be found", out_var_name); - auto *tensor = out_var->GetMutable(); + if (out_var->IsType()) { + LoadLodTensor(fin, place, out_var); + } else if (out_var->IsType()) { + LoadSelectedRows(fin, place, out_var); + } else { + PADDLE_ENFORCE( + false, + "Load only support LoDTensor and SelectedRows, %s has wrong type", + out_var_name); + } + } - DeserializeFromStream(fin, tensor, *dev_ctx); + void LoadLodTensor(std::istream &fin, const platform::Place &place, + framework::Variable *var) const { + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + auto *tensor = var->GetMutable(); + DeserializeFromStream(fin, tensor, dev_ctx); auto load_as_fp16 = Attr("load_as_fp16"); auto in_dtype = framework::ToDataType(tensor->type()); @@ -63,36 +78,40 @@ class LoadOp : public framework::OperatorBase { &fp16_tensor); // reset output tensor - out_var->Clear(); - tensor = out_var->GetMutable(); + var->Clear(); + tensor = var->GetMutable(); tensor->set_lod(fp16_tensor.lod()); tensor->ShareDataWith(fp16_tensor); } } + + void LoadSelectedRows(std::istream &fin, const platform::Place &place, + framework::Variable *var) const { + auto *selectedRows = var->GetMutable(); + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + framework::DeserializeFromStream(fin, selectedRows, dev_ctx); + } }; class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddOutput("Out", "(Tensor) The tensor need to be loaded"); + AddOutput("Out", "The LoDTensor / SelectedRows need to be loaded"); AddAttr( "load_as_fp16", - "(boolean, default false)" "If true, the tensor will be first loaded and then " "converted to float16 data type. Otherwise, the tensor will be " - "directly loaded without data type conversion.") + "directly loaded without data type conversion. Default is false.") .SetDefault(false); AddAttr("file_path", - "(string) " - "Variable will be loaded from \"file_path\".") + R"(Variable will be loaded from "file_path")") .AddCustomChecker( [](const std::string &path) { return !path.empty(); }); - AddComment(R"DOC( -Load Operator. - -Load operator will load a tensor variable from disk file. - -)DOC"); + AddComment( + "Load operator will load a LoDTensor / SelectedRows variable from disk " + "file."); } }; } // namespace operators diff --git a/paddle/fluid/operators/logical_op.cc b/paddle/fluid/operators/logical_op.cc index db109f5cd053d84718ac85bd4693ecece12ce172..26970db8d2af62bb06fce4eb1a1f21fd41617bd1 100644 --- a/paddle/fluid/operators/logical_op.cc +++ b/paddle/fluid/operators/logical_op.cc @@ -146,6 +146,6 @@ REGISTER_UNARY_LOGICAL_OP(logical_not, "$$Out = !X$$"); REGISTER_UNARY_LOGICAL_KERNEL(logical_not, CPU, paddle::operators::LogicalNotFunctor); REGISTER_BINARY_LOGICAL_OP(logical_xor, - "$$Out = (X || Y) \\, \\&\\& \\, !(X \\&\\& Y)$$"); + "$$Out = (X || Y) \\&\\& !(X \\&\\& Y)$$"); REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, CPU, paddle::operators::LogicalXorFunctor); diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index bda499432214b8841c8dfc406ee45ca0367920e7..d77b095c5d783a2a9fab87eb8b458117a6a3d225 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -32,20 +32,21 @@ class LookupTableOp : public framework::OperatorWithKernel { auto table_dims = ctx->GetInputDim("W"); auto ids_dims = ctx->GetInputDim("Ids"); + int ids_rank = ids_dims.size(); - auto ids_var_type = ctx->GetInputsVarType("Ids").front(); - // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type - // is LoDTensor, this tensor contains the ids to be looked up in W - // and it must be a column vector with rank = 2 while the 2nd dimension - // size must be 1, when Ids's type is SelectedRows, the rows of Ids - // contains the ids to be looked up in W; - if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { - PADDLE_ENFORCE_EQ(ids_dims.size(), 2); - PADDLE_ENFORCE_EQ(ids_dims[1], 1); - } + PADDLE_ENFORCE_EQ(table_dims.size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[ids_rank - 1], 1, + "The last dimension of the 'Ids' tensor must be 1."); + + auto output_dims = + framework::vectorize(framework::slice_ddim(ids_dims, 0, ids_rank - 1)); + output_dims.push_back(table_dims[1]); + ctx->SetOutputDim("Out", framework::make_ddim(output_dims)); - ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]}); - ctx->ShareLoD("Ids", /*->*/ "Out"); + if (ctx->GetOutputsVarType("Out")[0] == + framework::proto::VarType::LOD_TENSOR) { + ctx->ShareLoD("Ids", /*->*/ "Out"); + } } protected: @@ -62,17 +63,11 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("W", "(Tensor) The input represents embedding tensors, " "which is a learnable parameter."); - AddInput( - "Ids", - "(Tensor or SelectedRows) Ids's type can be Tensor or " - "SelectedRows, when Ids's type is Tensor, this tensor contains " - "the ids to be looked up in W and it must be a column vector with " - "rank = 2 while the 2nd dimension size must be 1; when Ids's type is " - "SelectedRows, the rows of Ids contains the ids to be looked up " - "in W."); - AddOutput("Out", - "(Tensor or SelectedRows) The lookup results, which have the " - "same type as W."); + AddInput("Ids", + "An input with type int32 or int64 " + "contains the ids to be looked up in W. " + "The last dimension size must be 1."); + AddOutput("Out", "The lookup results, which have the same type as W."); AddAttr("is_sparse", "(boolean, default false) " "Sparse update.") @@ -90,15 +85,10 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { Lookup Table Operator. This operator is used to perform lookups on the parameter W, -then concatenated into a dense or sparse tensor. - -The type of Ids(Input) is SelectedRows, Tensor or LoDTensor, when Ids's -type is SelectedRows, the rows of Ids contains the ids to be looked up in W; -when Ids's type is Tensor, this tensor contains the ids to be looked up in W -and it must be a column vector with rank = 2 while the 2nd dimension size must be 1, -at this time, Ids can carry the LoD (Level of Details) information, or not, and -the output only shares the LoD information with input Ids. +then concatenated into a dense tensor. +The input Ids can carry the LoD (Level of Details) information, +or not. And the output only shares the LoD information with input Ids. )DOC"); } diff --git a/paddle/fluid/operators/lookup_table_op.cu b/paddle/fluid/operators/lookup_table_op.cu index 77722c50d39003d9342afb04a61ae3aaf6b21100..74823dab09cac358f647c074ac2f2ee2fed17e55 100644 --- a/paddle/fluid/operators/lookup_table_op.cu +++ b/paddle/fluid/operators/lookup_table_op.cu @@ -23,7 +23,7 @@ namespace operators { template -__global__ void LookupTable(T* output, const T* table, const int64_t* ids, +__global__ void LookupTable(T *output, const T *table, const int64_t *ids, const int64_t N, const int64_t K, const int64_t D, const int64_t padding_idx) { int idx = threadIdx.x; @@ -33,8 +33,8 @@ __global__ void LookupTable(T* output, const T* table, const int64_t* ids, int64_t id = ids[idy]; PADDLE_ASSERT(id >= 0); PADDLE_ASSERT(id < N); - T* out = output + idy * D; - const T* tab = table + id * D; + T *out = output + idy * D; + const T *tab = table + id * D; for (int i = idx; i < D; i += BlockDimX) { if (PaddingFlag) { if (id == padding_idx) @@ -50,7 +50,7 @@ __global__ void LookupTable(T* output, const T* table, const int64_t* ids, } template -__global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids, +__global__ void LookupTableGrad(T *table, const T *output, const int64_t *ids, const int64_t N, const int64_t K, const int64_t D) { int idx = threadIdx.x; @@ -60,8 +60,8 @@ __global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids, int id = ids[idy]; PADDLE_ASSERT(id >= 0); PADDLE_ASSERT(id < N); - const T* out = output + idy * D; - T* tab = table + id * D; + const T *out = output + idy * D; + T *tab = table + id * D; for (int i = idx; i < D; i += BlockDimX) { paddle::platform::CudaAtomicAdd(&tab[i], out[i]); } @@ -72,36 +72,19 @@ __global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids, template class LookupTableCUDAKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - auto* table_t = context.Input("W"); + void Compute(const framework::ExecutionContext &context) const override { + auto *table_t = context.Input("W"); + auto *ids_t = context.Input("Ids"); + auto *output_t = context.Output("Out"); int64_t padding_idx = context.Attr("padding_idx"); - auto* ids_var = context.InputVar("Ids"); - Tensor* output_t = context.Output("Out"); - - int64_t* ids; - int64_t K; - - // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type - // is LoDTensor, this tensor contains the ids to be looked up in W; - // when Ids's type is SelectedRows, the rows of Ids contains the - // ids to be looked up in W. - if (ids_var->IsType()) { - auto* ids_t = context.Input("Ids"); - ids = const_cast(ids_t->data()); - K = ids_t->numel(); - } else if (ids_var->IsType()) { - auto* ids_t = context.Input("Ids"); - ids = const_cast(ids_t->rows().CUDAData(context.GetPlace())); - K = ids_t->rows().size(); - output_t->Resize({K, table_t->dims()[1]}); - } else { - PADDLE_THROW("Unsupported Variable Type of Ids"); - } size_t N = table_t->dims()[0]; size_t D = table_t->dims()[1]; - auto* table = table_t->data(); - auto* output = output_t->mutable_data(context.GetPlace()); + size_t K = ids_t->numel(); + + auto *ids = ids_t->data(); + auto *table = table_t->data(); + auto *output = output_t->mutable_data(context.GetPlace()); dim3 threads(128, 8); dim3 grids(8, 1); @@ -122,41 +105,44 @@ class LookupTableCUDAKernel : public framework::OpKernel { template class LookupTableGradCUDAKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - auto& dev_ctx = + void Compute(const framework::ExecutionContext &context) const override { + auto &dev_ctx = context.template device_context(); bool is_sparse = context.Attr("is_sparse"); // Since paddings are not trainable and fixed in forward, the gradient of // paddings makes no sense and we don't deal with it in backward. if (is_sparse) { - auto* ids = context.Input("Ids"); - auto* table = context.Input("W"); - auto* d_output = context.Input(framework::GradVarName("Out")); - auto* d_table = context.Output(framework::GradVarName("W")); + auto *ids = context.Input("Ids"); + auto *table = context.Input("W"); + auto *d_output = context.Input(framework::GradVarName("Out")); + auto *d_table = context.Output(framework::GradVarName("W")); - auto* ids_data = ids->data(); - auto ids_dim = ids->dims(); + auto *ids_data = ids->data(); + int64_t ids_num = ids->numel(); auto stream = dev_ctx.stream(); // copy GPU memory to CPU pinned memory framework::Vector new_rows; - new_rows.resize(ids_dim[0]); + new_rows.resize(ids_num); auto gpu_place = boost::get(context.GetPlace()); // TODO(yuyang18): Strange code here. memory::Copy(platform::CPUPlace(), new_rows.CUDAMutableData(context.GetPlace()), gpu_place, - ids_data, ids_dim[0] * sizeof(int64_t), stream); + ids_data, ids_num * sizeof(int64_t), stream); d_table->set_rows(new_rows); - auto* d_table_value = d_table->mutable_value(); - d_table_value->Resize({ids_dim[0], table->dims()[1]}); + auto *d_table_value = d_table->mutable_value(); + d_table_value->Resize({ids_num, table->dims()[1]}); d_table_value->mutable_data(context.GetPlace()); - auto* d_table_data = d_table_value->data(); - auto* d_output_data = d_output->data(); - PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output->dims()); + auto *d_table_data = d_table_value->data(); + auto *d_output_data = d_output->data(); + auto d_output_dims = d_output->dims(); + PADDLE_ENFORCE_EQ( + d_table_value->dims(), + framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1)); memory::Copy(gpu_place, d_table_data, gpu_place, d_output_data, d_output->numel() * sizeof(T), stream); @@ -168,9 +154,9 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { int N = d_table_t->dims()[0]; int D = d_table_t->dims()[1]; int K = ids_t->numel(); - const int64_t* ids = ids_t->data(); - const T* d_output = d_output_t->data(); - T* d_table = d_table_t->mutable_data(context.GetPlace()); + const int64_t *ids = ids_t->data(); + const T *d_output = d_output_t->data(); + T *d_table = d_table_t->mutable_data(context.GetPlace()); auto t = framework::EigenVector::Flatten(*d_table_t); t.device(*dev_ctx.eigen_device()) = t.constant(static_cast(0)); diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h index d482506bf0361c11a019e32efbf348a64aaf5164..f5c10ced8305b64c6386c5051804f8c9a8f71802 100644 --- a/paddle/fluid/operators/lookup_table_op.h +++ b/paddle/fluid/operators/lookup_table_op.h @@ -36,43 +36,13 @@ template class LookupTableKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { + auto *ids_t = context.Input("Ids"); // int tensor + auto *output_t = context.Output("Out"); // float tensor auto *table_var = context.InputVar("W"); - auto *ids_var = context.InputVar("Ids"); - Tensor *output_t = context.Output("Out"); - int64_t padding_idx = context.Attr("padding_idx"); - - DDim table_dim; - if (table_var->IsType()) { - table_dim = context.Input("W")->dims(); - } else if (table_var->IsType()) { - auto *table_t = context.Input("W"); - table_dim = table_t->value().dims(); - } else { - PADDLE_THROW( - "The parameter W of a LookupTable " - "must be either LoDTensor or SelectedRows"); - } - - int64_t *ids; - int64_t ids_numel; - - // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type - // is LoDTensor, this tensor contains the ids to be looked up in W; - // when Ids's type is SelectedRows, the rows of Ids contains the - // ids to be looked up in W. - if (ids_var->IsType()) { - auto *ids_t = context.Input("Ids"); - ids = const_cast(ids_t->data()); - ids_numel = ids_t->numel(); - } else if (ids_var->IsType()) { - auto *ids_t = context.Input("Ids"); - ids = const_cast(ids_t->rows().data()); - ids_numel = ids_t->rows().size(); - output_t->Resize({ids_numel, table_dim[1]}); - } else { - PADDLE_THROW("Unsupported Variable Type of Ids"); - } + int64_t padding_idx = context.Attr("padding_idx"); + int64_t *ids = const_cast(ids_t->data()); + int64_t ids_numel = ids_t->numel(); if (table_var->IsType()) { auto *table_t = context.Input("W"); @@ -139,17 +109,17 @@ class LookupTableGradKernel : public framework::OpKernel { auto *d_table = context.Output(framework::GradVarName("W")); auto *ids_data = ids->data(); - auto ids_dim = ids->dims(); + int64_t ids_num = ids->numel(); framework::Vector new_rows; - new_rows.reserve(ids_dim[0]); - for (int64_t i = 0; i < ids_dim[0]; i++) { + new_rows.reserve(ids_num); + for (int64_t i = 0; i < ids_num; i++) { new_rows.push_back(ids_data[i]); } d_table->set_rows(new_rows); auto *d_table_value = d_table->mutable_value(); - d_table_value->Resize({ids_dim[0], table_dim[1]}); + d_table_value->Resize({ids_num, table_dim[1]}); d_table_value->mutable_data(context.GetPlace()); d_table->set_height(table_dim[0]); @@ -157,7 +127,10 @@ class LookupTableGradKernel : public framework::OpKernel { auto *d_output_data = d_output->data(); auto *d_table_data = d_table_value->data(); - PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output->dims()); + auto d_output_dims = d_output->dims(); + PADDLE_ENFORCE_EQ( + d_table_value->dims(), + framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1)); memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel()); } else { auto *ids = context.Input("Ids"); @@ -165,10 +138,9 @@ class LookupTableGradKernel : public framework::OpKernel { auto *d_table = context.Output(framework::GradVarName("W")); auto *ids_data = ids->data(); - auto ids_dim = ids->dims(); int N = table_dim[0]; - int D = d_output->dims()[1]; + int D = table_dim[1]; auto *d_output_data = d_output->data(); auto *d_table_data = d_table->mutable_data(context.GetPlace()); diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc index 52b9cd7fb7019b738098a8649f23277afd40e938..52b459a6a2e56b7c256efdb535b4652c64bae23c 100644 --- a/paddle/fluid/operators/lrn_op.cc +++ b/paddle/fluid/operators/lrn_op.cc @@ -124,16 +124,17 @@ namespace { framework::OpKernelType GetExpectedLRNKernel( const framework::ExecutionContext& ctx) { framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); #ifdef PADDLE_WITH_MKLDNN if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), layout_, library_); diff --git a/paddle/fluid/operators/lstm_op.cc b/paddle/fluid/operators/lstm_op.cc index 4751e3e8025e51a687f8fcfd25e603b61e762f6d..3225bf9bb63d57969ce9ae0e4a74e8f466c8c2d0 100644 --- a/paddle/fluid/operators/lstm_op.cc +++ b/paddle/fluid/operators/lstm_op.cc @@ -184,34 +184,32 @@ Long-Short Term Memory (LSTM) Operator. The defalut implementation is diagonal/peephole connection (https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows: -$$ -i_t = \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) \\ +$$ i_t = \\sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) $$ -f_t = \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) \\ +$$ f_t = \\sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) $$ -\tilde{c_t} = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) \\ +$$ \\tilde{c_t} = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) $$ -o_t = \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) \\ +$$ o_t = \\sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) $$ -c_t = f_t \odot c_{t-1} + i_t \odot \tilde{c_t} \\ +$$ c_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} $$ -h_t = o_t \odot act_h(c_t) -$$ +$$ h_t = o_t \\odot act_h(c_t) $$ -where the W terms denote weight matrices (e.g. $W_{xi}$ is the matrix -of weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$ -are diagonal weight matrices for peephole connections. In our implementation, -we use vectors to reprenset these diagonal weight matrices. The b terms -denote bias vectors ($b_i$ is the input gate bias vector), $\sigma$ -is the non-line activations, such as logistic sigmoid function, and -$i, f, o$ and $c$ are the input gate, forget gate, output gate, -and cell activation vectors, respectively, all of which have the same size as -the cell output activation vector $h$. - -The $\odot$ is the element-wise product of the vectors. $act_g$ and $act_h$ -are the cell input and cell output activation functions and `tanh` is usually -used for them. $\tilde{c_t}$ is also called candidate hidden state, -which is computed based on the current input and the previous hidden state. +- W terms denote weight matrices (e.g. $W_{xi}$ is the matrix + of weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$ + are diagonal weight matrices for peephole connections. In our implementation, + we use vectors to reprenset these diagonal weight matrices. +- The b terms denote bias vectors ($b_i$ is the input gate bias vector). +- $\sigma$ is the non-line activations, such as logistic sigmoid function. +- $i, f, o$ and $c$ are the input gate, forget gate, output gate, + and cell activation vectors, respectively, all of which have the same size as + the cell output activation vector $h$. +- The $\odot$ is the element-wise product of the vectors. +- $act_g$ and $act_h$ are the cell input and cell output activation functions + and `tanh` is usually used for them. +- $\tilde{c_t}$ is also called candidate hidden state, + which is computed based on the current input and the previous hidden state. Set `use_peepholes` False to disable peephole connection. The formula is omitted here, please refer to the paper diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index 53a478c1ac0bdf8c0a3f3721161779ef10cb14f8..d2b772d11379c218be77277b89f3ded7b59ab9f3 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -51,16 +51,17 @@ math_library(sequence_padding) math_library(sequence_pooling DEPS math_function) math_library(sequence_scale) math_library(softmax DEPS math_function) +math_library(matrix_bit_code) math_library(unpooling) math_library(vol2col) -cc_test(math_function_test SRCS math_function_test.cc) +cc_test(math_function_test SRCS math_function_test.cc DEPS math_function) cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selected_rows_functor) cc_test(im2col_test SRCS im2col_test.cc DEPS im2col) cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col) cc_test(sequence_padding_test SRCS sequence_padding_test.cc DEPS sequence_padding) if(WITH_GPU) - nv_test(math_function_gpu_test SRCS math_function_test.cu) - nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor) + nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function) + nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function) endif() cc_test(concat_test SRCS concat_test.cc DEPS concat) diff --git a/paddle/fluid/operators/math/blas.h b/paddle/fluid/operators/math/blas.h index 1a37cb39d56066b8380338b9710a441e41518c39..70f88f24f682e05972ca73ef7b50f96be50d1ef4 100644 --- a/paddle/fluid/operators/math/blas.h +++ b/paddle/fluid/operators/math/blas.h @@ -18,28 +18,15 @@ #include "paddle/fluid/framework/tensor.h" #ifdef PADDLE_WITH_MKLML -#include -#include -#include +#include "paddle/fluid/platform/dynload/mklml.h" #endif -#ifdef PADDLE_USE_OPENBLAS -#include -#include +#ifdef PADDLE_WITH_LIBXSMM +#include #endif -#ifndef LAPACK_FOUND -extern "C" { -#include // NOLINT -int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda, - int* ipiv); -int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda, - int* ipiv); -int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda, - const int* ipiv); -int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, - const int* ipiv); -} +#ifdef PADDLE_USE_OPENBLAS +#include #endif namespace paddle { diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h index ae20406bc21d5e08359be8295cd98495dda7813b..a0802ef90ca7e30a2b22d187cb9092163518d8e9 100644 --- a/paddle/fluid/operators/math/blas_impl.h +++ b/paddle/fluid/operators/math/blas_impl.h @@ -12,6 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. #pragma once +#include #include #include "paddle/fluid/operators/math/math_function.h" @@ -22,83 +23,142 @@ namespace math { template struct CBlas; +#ifdef PADDLE_WITH_MKLML template <> struct CBlas { template static void GEMM(ARGS... args) { - cblas_sgemm(args...); + platform::dynload::cblas_sgemm(args...); } +#ifdef PADDLE_WITH_LIBXSMM template - static void AXPY(ARGS... args) { - cblas_saxpy(args...); + static void SMM_GEMM(ARGS... args) { + libxsmm_sgemm(args...); } +#endif -#ifdef PADDLE_WITH_MKLML template - static void VADD(ARGS... args) { - vsAdd(args...); + static void AXPY(ARGS... args) { + platform::dynload::cblas_saxpy(args...); } -#endif template static void VCOPY(ARGS... args) { - cblas_scopy(args...); + platform::dynload::cblas_scopy(args...); } template static void GEMV(ARGS... args) { - cblas_sgemv(args...); + platform::dynload::cblas_sgemv(args...); } -#ifdef PADDLE_WITH_MKLML template static void GEMM_BATCH(ARGS... args) { - cblas_sgemm_batch(args...); + platform::dynload::cblas_sgemm_batch(args...); + } + + template + static void VADD(ARGS... args) { + platform::dynload::vsAdd(args...); } -#endif }; template <> struct CBlas { template static void GEMM(ARGS... args) { - cblas_dgemm(args...); + platform::dynload::cblas_dgemm(args...); } +#ifdef PADDLE_WITH_LIBXSMM + template + static void SMM_GEMM(ARGS... args) { + libxsmm_dgemm(args...); + } +#endif + template static void AXPY(ARGS... args) { - cblas_daxpy(args...); + platform::dynload::cblas_daxpy(args...); + } + + template + static void VCOPY(ARGS... args) { + platform::dynload::cblas_dcopy(args...); + } + + template + static void GEMV(ARGS... args) { + platform::dynload::cblas_dgemv(args...); + } + + template + static void GEMM_BATCH(ARGS... args) { + platform::dynload::cblas_dgemm_batch(args...); } -#ifdef PADDLE_WITH_MKLML template static void VADD(ARGS... args) { - vdAdd(args...); + platform::dynload::vdAdd(args...); + } +}; + +#else + +template <> +struct CBlas { + template + static void GEMM(ARGS... args) { + cblas_sgemm(args...); + } + + template + static void AXPY(ARGS... args) { + cblas_saxpy(args...); } -#endif template static void VCOPY(ARGS... args) { - cblas_dcopy(args...); + cblas_scopy(args...); } template static void GEMV(ARGS... args) { - cblas_dgemv(args...); + cblas_sgemv(args...); } +}; -#ifdef PADDLE_WITH_MKLML +template <> +struct CBlas { template - static void GEMM_BATCH(ARGS... args) { - cblas_dgemm_batch(args...); + static void GEMM(ARGS... args) { + cblas_dgemm(args...); + } + + template + static void AXPY(ARGS... args) { + cblas_daxpy(args...); + } + + template + static void VCOPY(ARGS... args) { + cblas_dcopy(args...); + } + + template + static void GEMV(ARGS... args) { + cblas_dgemv(args...); } -#endif }; +#endif template <> struct CBlas { static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); } + static void SMM_GEMM(...) { + PADDLE_THROW("float16 SMM_GEMM not supported on CPU"); + } #ifdef PADDLE_WITH_MKLML static void GEMM_BATCH(...) { PADDLE_THROW("float16 GEMM_BATCH not supported on CPU"); @@ -106,6 +166,64 @@ struct CBlas { #endif }; +template +inline bool UseXSMM(const int &m, const int &n, const int &k, bool transa, + bool transb, const T &alpha, const T &beta) { +#ifdef PADDLE_WITH_LIBXSMM + // Refer to https://github.com/hfp/libxsmm/blob/master/README.md + // But the threshold is custom + constexpr int LIBXSMM_THRESHOLD = 20 * 20 * 20; + if (m * n * k > LIBXSMM_THRESHOLD || transa || transb || + std::abs(alpha - static_cast(1) > + std::numeric_limits::epsilon()) || + std::abs(beta) > std::numeric_limits::epsilon()) { + return false; + } else { + return true; + } +#endif + return false; +} + +template <> +inline bool UseXSMM(const int &m, const int &n, const int &k, + bool transa, bool transb, + const platform::float16 &alpha, + const platform::float16 &beta) { + return false; +} + +template +inline void GEMM_WARP(CBLAS_ORDER order, CBLAS_TRANSPOSE transA, + CBLAS_TRANSPOSE transB, int M, int N, int K, T alpha, + const T *A, int lda, const T *B, int ldb, T beta, T *C, + int ldc) { +#ifdef PADDLE_WITH_LIBXSMM + if (UseXSMM(M, N, K, transA != CblasNoTrans, transB != CblasNoTrans, alpha, + beta)) { + // Note: SMM use ColMajor + const char transa = 'N'; + const char transb = 'N'; + CBlas::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &ldb, A, &lda, + &beta, C, &ldc); + return; + } +#endif + +#ifdef PADDLE_MKL_SPLIT_GEMM + constexpr int bs = 2; + if (M % bs == 0 && transA == CblasNoTrans && transB == CblasNoTrans) { + for (int off = 0; off < M; off += bs) { + CBlas::GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, bs, N, K, alpha, + A + off * lda, lda, B, ldb, beta, C + off * ldb, ldc); + } + return; + } +#endif + CBlas::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + template <> template void Blas::GEMM(CBLAS_TRANSPOSE transA, @@ -115,8 +233,8 @@ void Blas::GEMM(CBLAS_TRANSPOSE transA, int lda = (transA == CblasNoTrans) ? K : M; int ldb = (transB == CblasNoTrans) ? N : K; int ldc = N; - CBlas::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, - beta, C, ldc); + GEMM_WARP(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); } template <> @@ -125,9 +243,9 @@ void Blas::GEMM(bool transA, bool transB, int M, int N, int K, T alpha, const T *A, int lda, const T *B, int ldb, T beta, T *C, int ldc) const { - CBlas::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, - transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, - lda, B, ldb, beta, C, ldc); + GEMM_WARP(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, + lda, B, ldb, beta, C, ldc); } template diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat.cc index cc69212466b72f3fa82e8f5f58b4f3229dab28ec..55c8a472aca7fe700ef6a3f96bed1496d7b12b80 100644 --- a/paddle/fluid/operators/math/concat.cc +++ b/paddle/fluid/operators/math/concat.cc @@ -70,21 +70,23 @@ template class ConcatGradFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, const int axis, - std::vector* outputs) { + const framework::Tensor& input, + const std::vector& ref_inputs, + const int axis, std::vector* outputs) { // TODO(zcd): Add input data validity checking - int num = outputs->size(); + size_t num = outputs->size(); int input_rows = 1; - auto dim_0 = outputs->at(0).dims(); + auto dim_0 = ref_inputs[0]->dims(); for (int i = 0; i < axis; ++i) { input_rows *= dim_0[i]; } + int input_cols = 0; std::vector output_cols(outputs->size()); - for (int i = 0; i < num; ++i) { - int t_cols = outputs->at(i).numel() / input_rows; + for (size_t i = 0; i < num; ++i) { + int t_cols = ref_inputs[i]->numel() / input_rows; input_cols += t_cols; output_cols[i] = t_cols; } @@ -94,11 +96,14 @@ class ConcatGradFunctor { for (int k = 0; k < input_rows; ++k) { const T* src_ptr = input.data() + k * input_cols; int col_idx = 0; - for (int j = 0; j < num; ++j) { + for (size_t j = 0; j < num; ++j) { int col_len = output_cols[j]; - T* dst_ptr = outputs->at(j).data() + k * col_len; - memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx, - sizeof(T) * col_len); + auto* out_tensor = outputs->at(j); + if (out_tensor != nullptr) { + T* dst_ptr = out_tensor->data() + k * col_len; + memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx, + sizeof(T) * col_len); + } col_idx += col_len; } } diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat.cu index 4285d38dcd6a4124543cdd2246c82a8203f5a281..5863d74fca21de8b77bc208fb95d8fd52562f7a7 100644 --- a/paddle/fluid/operators/math/concat.cu +++ b/paddle/fluid/operators/math/concat.cu @@ -22,43 +22,24 @@ namespace paddle { namespace operators { namespace math { -template -__device__ T upper_bound(const T* first, T count, T val) { - const T* orig = first; - const T* it = nullptr; - T step = 0; - while (count > 0) { - it = first; - step = count / 2; - it += step; - if (!(val < *it)) { - first = ++it; - count -= step + 1; - } else { - count = step; - } - } - return first - orig; -} - template __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size, const int output_rows, const int output_cols, T* output) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; - int segment = upper_bound(input_cols, col_size, tid_x) - 1; - - int curr_offset = input_cols[segment]; - int curr_segment = segment; + int curr_segment = 0; + int curr_offset = input_cols[0]; for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) { - T curr_col_offset; - while ((curr_col_offset = input_cols[curr_segment + 1]) <= tid_x) { + int curr_col_offset = input_cols[curr_segment + 1]; + while (curr_col_offset <= tid_x) { curr_offset = curr_col_offset; ++curr_segment; + curr_col_offset = input_cols[curr_segment + 1]; } int local_col = tid_x - curr_offset; int segment_width = curr_col_offset - curr_offset; + T* input_ptr = inputs[curr_segment]; int tid_y = blockIdx.y * blockDim.y + threadIdx.y; for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y) @@ -89,23 +70,25 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row, const int in_col, const int* out_cols, int out_cols_size, T** outputs_data) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; - int segment = upper_bound(out_cols, out_cols_size, tid_x) - 1; - int curr_offset = out_cols[segment]; - int curr_segment = segment; + int curr_segment = 0; + int curr_offset = out_cols[0]; for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { - T curr_col_offset; - while ((curr_col_offset = out_cols[curr_segment + 1]) <= tid_x) { + int curr_col_offset = out_cols[curr_segment + 1]; + while (curr_col_offset <= tid_x) { curr_offset = curr_col_offset; ++curr_segment; + curr_col_offset = out_cols[curr_segment + 1]; } int local_col = tid_x - curr_offset; int segment_width = curr_col_offset - curr_offset; T* output_ptr = outputs_data[curr_segment]; - int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) - output_ptr[tid_y * segment_width + local_col] = - input_data[tid_y * in_col + tid_x]; + if (output_ptr != nullptr) { + int tid_y = blockIdx.y * blockDim.y + threadIdx.y; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * segment_width + local_col] = + input_data[tid_y * in_col + tid_x]; + } } } @@ -118,10 +101,12 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row, int split = tid_x / fixed_out_col; int in_offset = tid_x - split * fixed_out_col; T* output_ptr = outputs_data[split]; - int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) - output_ptr[tid_y * fixed_out_col + in_offset] = - input_data[tid_y * in_col + tid_x]; + if (output_ptr != nullptr) { + int tid_y = blockIdx.y * blockDim.y + threadIdx.y; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * fixed_out_col + in_offset] = + input_data[tid_y * in_col + tid_x]; + } } } @@ -203,17 +188,18 @@ template class ConcatGradFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& input, const int axis, - std::vector* outputs) { + const framework::Tensor& input, + const std::vector& ref_inputs, + const int axis, std::vector* outputs) { // TODO(zcd): Add input data validity checking int o_num = outputs->size(); int out_row = 1; - auto dim_0 = outputs->at(0).dims(); + auto dim_0 = ref_inputs[0]->dims(); for (int i = 0; i < axis; ++i) { out_row *= dim_0[i]; } - int out_col = outputs->at(0).numel() / out_row; + int out0_col = ref_inputs[0]->numel() / out_row; int in_col = 0, in_row = out_row; bool sameShape = true; @@ -223,13 +209,17 @@ class ConcatGradFunctor { outputs_cols[0] = 0; for (int i = 0; i < o_num; ++i) { - int t_col = outputs->at(i).numel() / out_row; + int t_col = ref_inputs.at(i)->numel() / out_row; if (sameShape) { - if (t_col != out_col) sameShape = false; + if (t_col != out0_col) sameShape = false; } in_col += t_col; outputs_cols[i + 1] = in_col; - outputs_ptr[i] = outputs->at(i).data(); + if (outputs->at(i) != nullptr) { + outputs_ptr[i] = outputs->at(i)->data(); + } else { + outputs_ptr[i] = nullptr; + } } T** dev_out_gpu_data = @@ -255,7 +245,7 @@ class ConcatGradFunctor { if (sameShape) { KernelConcatGrad<<>>( - input.data(), in_row, in_col, out_col, dev_out_gpu_data); + input.data(), in_row, in_col, out0_col, dev_out_gpu_data); } else { const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace()); KernelConcatGrad<<>>( diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat.h index 041ce8bf8a2e9528a004c076ead4471a3837c1a6..9e080f2e8be23768dcea47b577043beef37b2eaf 100644 --- a/paddle/fluid/operators/math/concat.h +++ b/paddle/fluid/operators/math/concat.h @@ -57,7 +57,8 @@ template class ConcatGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - const int axis, std::vector* outputs); + const std::vector& ref_inputs, + const int axis, std::vector* outputs); }; } // namespace math diff --git a/paddle/fluid/operators/math/detail/avx_functions.cc b/paddle/fluid/operators/math/detail/avx_functions.cc index b95109d3f73505fa6b5438326804a2b348fb3668..5641f914523771f47bd7f814bfd39964a53deefc 100644 --- a/paddle/fluid/operators/math/detail/avx_functions.cc +++ b/paddle/fluid/operators/math/detail/avx_functions.cc @@ -17,7 +17,7 @@ limitations under the License. */ #include #include "paddle/fluid/operators/math/detail/activation_functions.h" // TODO(qingqing) refine this dependence -#include "paddle/cuda/src/avx_mathfun.h" +#include "paddle/legacy/cuda/src/avx_mathfun.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/functors.h b/paddle/fluid/operators/math/functors.h new file mode 100644 index 0000000000000000000000000000000000000000..ad2f49ccbf5ff37d33cc9e71c1a683571f4f8137 --- /dev/null +++ b/paddle/fluid/operators/math/functors.h @@ -0,0 +1,71 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +namespace paddle { +namespace operators { +namespace math { + +// AddFunctor +template +struct AddFunctor { + // out = x + y; + inline HOSTDEVICE T operator()(T x, T y) { return x + y; } +}; + +template +struct AddGradFunctor { + inline HOSTDEVICE T operator()(T x, T y) { return 1; } + + inline HOSTDEVICE T operator()(T x, T y, T out) const { return 1; } +}; + +template +struct ScaleFunctor { + explicit ScaleFunctor(const T coeff) : coeff_(coeff) {} + + inline HOSTDEVICE T operator()(T ele) { return ele * coeff_; } + + private: + T coeff_; +}; + +template +struct ScaleGradFunctor { + explicit ScaleGradFunctor(T coeff) : coeff_(coeff) {} + + inline HOSTDEVICE T operator()(T x) { return coeff_; } + + inline HOSTDEVICE T operator()(T x, T out) { return coeff_; } + + private: + T coeff_; +}; + +template +struct ReluFunctor { + inline HOSTDEVICE T operator()(T x) { return x * (x > 0); } +}; + +template +struct ReluGradFunctor { + inline HOSTDEVICE T operator()(T x) { return x > 0 ? 1 : 0; } + + inline HOSTDEVICE T operator()(T x, T out) { return x > 0 ? 1 : 0; } +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/im2col.cc b/paddle/fluid/operators/math/im2col.cc index 336d6febc2ce3a55e82ed613bbc1081101f822f0..1472edbbf47e3e4d6b22c65349713904b13647d2 100644 --- a/paddle/fluid/operators/math/im2col.cc +++ b/paddle/fluid/operators/math/im2col.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/im2col.h" #include +#include "paddle/fluid/operators/math/im2col_cfo_cpu.h" namespace paddle { namespace operators { @@ -35,51 +36,18 @@ class Im2ColFunctordims().size() == 5); - int im_channels = im.dims()[0]; - int im_height = im.dims()[1]; - int im_width = im.dims()[2]; - int filter_height = col->dims()[1]; - int filter_width = col->dims()[2]; - int col_height = col->dims()[3]; - int col_width = col->dims()[4]; - - PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - - ((dilation[0] * (filter_height - 1) + 1))) / - stride[0] + - 1, - col_height, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - - ((dilation[1] * (filter_width - 1) + 1))) / - stride[1] + - 1, - col_width, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - - int channels_col = im_channels * filter_height * filter_width; - - const T* im_data = im.data(); - T* col_data = col->data(); - for (int c = 0; c < channels_col; ++c) { - int w_offset = c % filter_width; - int h_offset = (c / filter_width) % filter_height; - int c_im = c / (filter_width * filter_height); - for (int h = 0; h < col_height; ++h) { - int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0]; - for (int w = 0; w < col_width; ++w) { - int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1]; - int col_idx = (c * col_height + h) * col_width + w; - int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx; - - col_data[col_idx] = (im_row_idx < 0 || im_row_idx >= im_height || - im_col_idx < 0 || im_col_idx >= im_width) - ? static_cast(0) - : im_data[im_idx]; - } + if (stride[0] == 1 && stride[1] == 1 && dilation[0] == 1 && + dilation[1] == 1) { + if (padding[0] == 0 && padding[1] == 0) { + im2col_sh1sw1dh1dw1ph0pw0(im, col); + return; + } else if (padding[0] == 1 && padding[1] == 1) { + im2col_sh1sw1dh1dw1ph1pw1(im, col); + return; } + // TODO(TJ): complete padding >=2 } + im2col_common(im, dilation, stride, padding, col); } }; @@ -178,17 +146,6 @@ class Im2ColFunctordims()[0]; int col_width = col->dims()[1]; - PADDLE_ENFORCE_EQ( - (im_height + padding[0] + padding[2] - filter_height) / stride[0] + 1, - col_height, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - PADDLE_ENFORCE_EQ( - (im_width + padding[1] + padding[3] - filter_width) / stride[1] + 1, - col_width, - "col_width and padding(padding_left, padding_right) are " - "inconsistent."); - const T* im_data = im.data(); T* col_data = col->data(); diff --git a/paddle/fluid/operators/math/im2col.cu b/paddle/fluid/operators/math/im2col.cu index eecb233d22cea06da016b2671fd606b70eddf5a5..4897767f4d88d9e079f05c921153923c4eb354b0 100644 --- a/paddle/fluid/operators/math/im2col.cu +++ b/paddle/fluid/operators/math/im2col.cu @@ -77,21 +77,6 @@ class Im2ColFunctordims()[3]; int col_width = col->dims()[4]; - PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - - (dilation[0] * (filter_height - 1) + 1)) / - stride[0] + - 1, - col_height, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - - (dilation[1] * (filter_width - 1) + 1)) / - stride[1] + - 1, - col_width, - "col_width and padding(padding_left, padding_right) are " - "inconsistent."); - int num_outputs = im_channels * col_height * col_width; int blocks = (num_outputs + 1024 - 1) / 1024; int block_x = 512; @@ -274,21 +259,6 @@ class Im2ColFunctordims()[0]; int col_width = col->dims()[1]; - PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] - - (dilation[0] * (filter_height - 1) + 1)) / - stride[0] + - 1, - col_height, - "Output_height and padding(padding_up, padding_down) are " - "inconsistent."); - PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] - - (dilation[1] * (filter_width - 1) + 1)) / - stride[1] + - 1, - col_width, - "col_width and padding(padding_left, padding_right) are " - "inconsistent."); - int block_dim_x = 0; int block_dim_y = 0; if (filter_height <= 4 && filter_width <= 4) { diff --git a/paddle/fluid/operators/math/im2col_cfo_cpu.h b/paddle/fluid/operators/math/im2col_cfo_cpu.h new file mode 100644 index 0000000000000000000000000000000000000000..0d32bc5bd0d7f25479370959cabeb9b9c9e7e2d6 --- /dev/null +++ b/paddle/fluid/operators/math/im2col_cfo_cpu.h @@ -0,0 +1,252 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/tensor.h" + +namespace paddle { +namespace operators { +namespace math { + +/** + * The most common im2col algorithm. + * Support dilation, stride and padding. + */ +template +inline void im2col_common(const framework::Tensor& im, + const std::vector& dilation, + const std::vector& stride, + const std::vector& padding, + framework::Tensor* col) { + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[1]; + int filter_width = col->dims()[2]; + int output_height = col->dims()[3]; + int output_width = col->dims()[4]; + int channels_col = im_channels * filter_height * filter_width; + + const T* im_data = im.data(); + T* col_data = col->data(); + for (int c = 0; c < channels_col; ++c) { + int w_offset = c % filter_width; + int h_offset = (c / filter_width) % filter_height; + int c_im = c / (filter_width * filter_height); + for (int h = 0; h < output_height; ++h) { + int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0]; + for (int w = 0; w < output_width; ++w) { + int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1]; + int col_idx = (c * output_height + h) * output_width + w; + int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx; + col_data[col_idx] = (im_row_idx < 0 || im_row_idx >= im_height || + im_col_idx < 0 || im_col_idx >= im_width) + ? static_cast(0) + : im_data[im_idx]; + } + } + } +} + +/** + * im2col algorithm with strides == 1, dilations == 1, paddings == 0 + */ +template +inline void im2col_sh1sw1dh1dw1ph0pw0(const framework::Tensor& im, + framework::Tensor* col) { + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[1]; + int filter_width = col->dims()[2]; + int output_height = col->dims()[3]; + int output_width = col->dims()[4]; + + const T* im_data = im.data(); + T* col_data = col->data(); + int col_matrix_width = output_width * output_height; + int im_size = im_height * im_width; + size_t copy_size = sizeof(T) * output_width; + const T* im_data_oh = im_data; + T* dst_data_oh = col_data; + for (int oh = 0; oh < output_height; ++oh) { + const T* src_data_ic = im_data_oh; + T* dst_data = dst_data_oh; + for (int ic = 0; ic < im_channels; ++ic) { + const T* src_data = src_data_ic; + for (int kh = 0; kh < filter_height; ++kh) { + for (int kw = 0; kw < filter_width; ++kw) { + std::memcpy(dst_data, src_data + kw, copy_size); + dst_data = dst_data + col_matrix_width; + } + src_data = src_data + im_width; + } + src_data_ic = src_data_ic + im_size; + } + im_data_oh = im_data_oh + im_width; + dst_data_oh = dst_data_oh + output_width; + } +} + +/** + * im2col algorithm with strides == 1, dilations == 1, paddings == 1 + * and filter_width == 1 have a special implementation + */ +template +inline void im2col_sh1sw1dh1dw1ph1pw1(const framework::Tensor& im, + framework::Tensor* col) { + int im_channels = im.dims()[0]; + int im_height = im.dims()[1]; + int im_width = im.dims()[2]; + int filter_height = col->dims()[1]; + int filter_width = col->dims()[2]; + int output_height = col->dims()[3]; + int output_width = col->dims()[4]; + + constexpr int plh = 1; + constexpr int prh = 1; + constexpr int plw = 1; + constexpr int prw = 1; + + const T* im_data = im.data(); + T* col_data = col->data(); + int im_size = im_height * im_width; + int col_matrix_width = output_width * output_height; + int col_block_fh = filter_width * col_matrix_width; // fw*oh*ow + int col_block_ic = filter_height * col_block_fh; // fh*fw*oh*ow + + // fill height padding + { + size_t copy_size = sizeof(T) * output_width; + T* col_start_l = col_data; + T* col_start_r = col_data + (filter_height - 1) * col_block_fh + + col_matrix_width - output_width; + for (int ic = 0; ic < im_channels; ++ic) { + T* dst_data_l = col_start_l; + T* dst_data_r = col_start_r; + for (int kw = 0; kw < filter_width; ++kw) { + std::memset(dst_data_l, 0, copy_size); + std::memset(dst_data_r, 0, copy_size); + dst_data_l = dst_data_l + col_matrix_width; + dst_data_r = dst_data_r + col_matrix_width; + } + col_start_l = col_start_l + col_block_ic; + col_start_r = col_start_r + col_block_ic; + } + } + + auto pad = static_cast(0); + if (filter_width == 1) { + // fill width padding + T* dst_data_ic = col_data; + for (int ic = 0; ic < im_channels; ++ic) { + T* dst_data_kh = dst_data_ic; + for (int kh = 0; kh < filter_height; ++kh) { + T* dst_data = dst_data_kh; + for (int oh = 0; oh < output_height; ++oh) { + *dst_data = pad; + dst_data = dst_data + output_width - 1; + *dst_data = pad; + ++dst_data; + } + dst_data_kh = dst_data_kh + col_block_fh; + } + dst_data_ic = dst_data_ic + col_block_ic; + } + // fill core + size_t copy_size = sizeof(T) * (output_width - plw - prw); + for (int oh = 0; oh < output_height; ++oh) { + const T* im_data_start = + im_data + (oh - plh > 0 ? oh - plh : 0) * im_width; + T* dst_data = col_data + oh * output_width; + for (int ic = 0; ic < im_channels; ++ic) { + const T* src_data = im_data_start + ic * im_size; + for (int kh = 0; kh < filter_height; ++kh) { + if ((oh < plh && kh < plh) || (oh > (output_height - prh - 1) && + kh > (filter_height - prh - 1))) { + dst_data = dst_data + col_matrix_width; + continue; + } + std::memcpy(dst_data + plw, src_data, copy_size); + dst_data = dst_data + col_matrix_width; + src_data = src_data + im_width; + } + } + } + return; + } + + // filter_width != 1 + // fill width padding + T* dst_data_ic = col_data; + for (int ic = 0; ic < im_channels; ++ic) { + T* dst_data_kh = dst_data_ic; + for (int kh = 0; kh < filter_height; ++kh) { + for (T* dst_data : + {dst_data_kh, dst_data_kh + (filter_width - prw) * col_matrix_width + + output_width - 1}) { + // TODO(TJ): from plh, saving repeated assignment + for (int oh = 0; oh < output_height; ++oh) { + *dst_data = pad; + dst_data = dst_data + output_width; + } + } + dst_data_kh = dst_data_kh + col_block_fh; + } + dst_data_ic = dst_data_ic + col_block_ic; + } + + // TODO(TJ): use array like: size_t copy_size[kw]={sizeof(T) * + // (output_width-1)} + // length of copy_size is equal kw. + for (int oh = 0; oh < output_height; ++oh) { + const T* im_data_start = im_data + (oh - plh > 0 ? oh - plh : 0) * im_width; + T* dst_data = col_data + oh * output_width; + for (int ic = 0; ic < im_channels; ++ic) { + const T* src_data = im_data_start + ic * im_size; + for (int kh = 0; kh < filter_height; ++kh) { + if ((oh < plh && kh < plh) || (oh > (output_height - prh - 1) && + kh > (filter_height - prh - 1))) { + dst_data = dst_data + filter_width * col_matrix_width; + continue; + } + // TODO(TJ): reuse plw-kw outside this for + // try to unify + for (int kw = 0; kw < plw; ++kw) { + std::memcpy(dst_data + (plw - kw), src_data, + sizeof(T) * (output_width - (plw - kw))); + dst_data = dst_data + col_matrix_width; + } + for (int kw = plw; kw < filter_width - prw; ++kw) { + std::memcpy(dst_data, src_data + (kw - plw), + sizeof(T) * output_width); + dst_data = dst_data + col_matrix_width; + } + int i = 1; + for (int kw = filter_width - prw; kw < filter_width; ++kw, ++i) { + std::memcpy(dst_data, src_data + (kw - plw), + sizeof(T) * (output_width - i)); + dst_data = dst_data + col_matrix_width; + } + src_data = src_data + im_width; + } + } + } +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/im2col_test.cc b/paddle/fluid/operators/math/im2col_test.cc index 8e3f0f286823c383bb0c44d0e7887040ec9b20a0..ae2c90b33a4298ada4fd01aa2a44ebdf10d036d4 100644 --- a/paddle/fluid/operators/math/im2col_test.cc +++ b/paddle/fluid/operators/math/im2col_test.cc @@ -14,7 +14,9 @@ limitations under the License. */ #include "paddle/fluid/operators/math/im2col.h" #include +#include #include +#include "paddle/fluid/operators/math/im2col_cfo_cpu.h" template void testIm2col() { @@ -167,3 +169,104 @@ TEST(math, im2col) { paddle::platform::CUDAPlace>(); #endif } + +#define PREPARE_IM2COL_CPU \ + paddle::platform::CPUPlace place; \ + paddle::platform::CPUDeviceContext context(place); \ + paddle::framework::Tensor input; \ + paddle::framework::Tensor out; \ + paddle::framework::Tensor ref; \ + std::vector padding({ph, pw}); \ + std::vector stride({1, 1}); \ + std::vector dilation({1, 1}); \ + float* input_ptr = input.mutable_data({ic, ih, iw}, place); \ + for (int i = 0; i < input.numel(); ++i) { \ + input_ptr[i] = static_cast(i + 1); \ + } \ + int output_height = (ih - fh + padding[0] * 2) / stride[0] + 1; \ + int output_width = (iw - fw + padding[1] * 2) / stride[1] + 1; \ + out.mutable_data({ic, fh, fw, output_height, output_width}, place); \ + ref.mutable_data({ic, fh, fw, output_height, output_width}, place); \ + paddle::operators::math::Im2ColFunctor< \ + paddle::operators::math::ColFormat::kCFO, \ + paddle::platform::CPUDeviceContext, float> \ + im2col + +void testIm2colCPU(int ic, int ih, int iw, int fh, int fw, int ph, int pw) { + PREPARE_IM2COL_CPU; + + im2col(context, input, dilation, stride, padding, &out); + paddle::operators::math::im2col_common(input, dilation, stride, + padding, &ref); + + float* ref_data = ref.data(); + float* out_data = out.data(); + for (int i = 0; i < out.numel(); ++i) { + EXPECT_EQ(out_data[i], ref_data[i]); + } +} + +void benchIm2col(int ic, int ih, int iw, int fh, int fw, int ph, int pw) { + PREPARE_IM2COL_CPU; + constexpr int repeat = 100; + auto GetCurrentMs = []() -> double { + struct timeval time; + gettimeofday(&time, NULL); + return 1e+3 * time.tv_sec + 1e-3 * time.tv_usec; + }; + auto t1 = GetCurrentMs(); + for (int i = 0; i < repeat; ++i) { + im2col(context, input, dilation, stride, padding, &out); + } + auto t2 = GetCurrentMs(); + + for (int i = 0; i < repeat; ++i) { + paddle::operators::math::im2col_common(input, dilation, stride, + padding, &ref); + } + auto t3 = GetCurrentMs(); + + LOG(INFO) << "before: " << (t3 - t2) / repeat + << ",after: " << (t2 - t1) / repeat + << ",boost: " << ((t3 - t2) / (t2 - t1) - 1) * 100 << "%"; +} + +TEST(math, im2col_cputest) { + // padding_h == padding_w + for (int p = 0; p < 4; ++p) { + // width == height + testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 5, /*fh*/ 4, /*fw*/ 4, /*ph*/ p, + /*pw*/ p); + testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 4, /*fh*/ 3, /*fw*/ 3, /*ph*/ p, + /*pw*/ p); + testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 4, /*fh*/ 2, /*fw*/ 2, /*ph*/ p, + /*pw*/ p); + + // height != width + testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 4, /*fh*/ 2, /*fw*/ 3, /*ph*/ p, + /*pw*/ p); + testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 4, /*fh*/ 1, /*fw*/ 3, /*ph*/ p, + /*pw*/ p); + testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 5, /*fh*/ 3, /*fw*/ 1, /*ph*/ p, + /*pw*/ p); + + // filter == 1 + testIm2colCPU(/*ic*/ 3, /*ih*/ 4, /*iw*/ 4, /*fh*/ 1, /*fw*/ 1, /*ph*/ p, + /*pw*/ p); + testIm2colCPU(/*ic*/ 3, /*ih*/ 3, /*iw*/ 4, /*fh*/ 1, /*fw*/ 1, /*ph*/ p, + /*pw*/ p); + } + + // padding_h != padding_w + testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 4, /*fh*/ 2, /*fw*/ 3, /*ph*/ 1, + /*pw*/ 2); + + // benchmark + for (int p : {0, 1}) { + for (int k : {1, 3, 5}) { + LOG(INFO) << "padding == " << p << ", filter == " << k; + benchIm2col(/*ic*/ 3, /*ih*/ 224, /*iw*/ 224, /*fh*/ k, /*fw*/ k, + /*ph*/ p, /*pw*/ p); + } + } +} diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index d39154c6f88d6d17c1719eb9a5b048211f4bb52b..c3387be6daa3bd34a6e3410ced23fce5d65f2cf7 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -30,6 +30,7 @@ template struct SetConstant; template struct SetConstant; template struct SetConstant; template struct SetConstant; +template struct SetConstant; #define DEFINE_CPU_TRANS(RANK) \ template struct Transpose -#include -#include +#include "paddle/fluid/platform/dynload/mklml.h" #endif #ifdef PADDLE_USE_OPENBLAS #include -#include -#endif - -#ifndef LAPACK_FOUND -extern "C" { -#include // NOLINT -int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda, - int* ipiv); -int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda, - int* ipiv); -int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda, - const int* ipiv); -int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, - const int* ipiv); -} #endif #include diff --git a/paddle/fluid/operators/math/math_function_impl.h b/paddle/fluid/operators/math/math_function_impl.h index b9bd49d77d935e985705f78402ffe1ea90f24cb3..895a7019aa10e5d9bb8f0c17e433a4344eac3bf4 100644 --- a/paddle/fluid/operators/math/math_function_impl.h +++ b/paddle/fluid/operators/math/math_function_impl.h @@ -155,7 +155,7 @@ class RowwiseSum { PADDLE_ENFORCE_EQ(in_dims.size(), 2U); auto height = in_dims[0]; auto size = in_dims[1]; - PADDLE_ENFORCE_EQ(out->numel(), size); + PADDLE_ENFORCE_EQ(out->numel(), height); T* out_buf = out->mutable_data(out->place()); const T* in_buf = input.data(); diff --git a/paddle/fluid/operators/math/math_function_test.cc b/paddle/fluid/operators/math/math_function_test.cc index 3719a264e90ea7d1a99eb9589ce4fd0d8e074781..2343e0ee965303c9fdb2ad3faf9ddf6e5bb7782f 100644 --- a/paddle/fluid/operators/math/math_function_test.cc +++ b/paddle/fluid/operators/math/math_function_test.cc @@ -54,8 +54,64 @@ TEST(math_function, gemm_notrans_cblas) { EXPECT_EQ(input3_ptr[6], 86); EXPECT_EQ(input3_ptr[7], 99); } +#ifdef PADDLE_WITH_LIBXSMM +template +void MklSmmCompare(int m, int n, int k) { + paddle::framework::Tensor mat_a; + paddle::framework::Tensor mat_b; + paddle::framework::Tensor mat_c_smm; + paddle::framework::Tensor mat_c_mkl; + auto* cpu_place = new paddle::platform::CPUPlace(); + + T* A = mat_a.mutable_data({m, k}, *cpu_place); + T* B = mat_b.mutable_data({k, n}, *cpu_place); + T* CSMM = mat_c_smm.mutable_data({m, n}, *cpu_place); + T* CMKL = mat_c_mkl.mutable_data({m, n}, *cpu_place); + T alpha = static_cast(1); + T beta = static_cast(0); + for (int i = 0; i < mat_a.numel(); ++i) { + A[i] = static_cast(i); + } + for (int i = 0; i < mat_b.numel(); ++i) { + B[i] = static_cast(i); + } + // lda,ldb,ldc follow RowMajor + int lda = k; + int ldb = n; + int ldc = n; + + auto smm = [&, m, n, k, lda, ldb, ldc, alpha, beta]() { + const char transa = 'N'; + const char transb = 'N'; + paddle::operators::math::CBlas::SMM_GEMM(&transa, &transb, &n, &m, &k, + &alpha, B, &ldb, A, &lda, &beta, + CSMM, &ldc); + }; -TEST(math_function, gemm_trans_clbas) { + auto mkl = [&, m, n, k, lda, ldb, ldc, alpha, beta]() { + paddle::operators::math::CBlas::GEMM(CblasRowMajor, CblasNoTrans, + CblasNoTrans, m, n, k, alpha, A, + lda, B, ldb, beta, CMKL, ldc); + }; + + smm(); + mkl(); + ASSERT_EQ(mat_c_mkl.numel(), mat_c_smm.numel()); + for (int i = 0; i < mat_c_mkl.numel(); ++i) { + EXPECT_FLOAT_EQ(CSMM[i], CMKL[i]); + } +} +TEST(math_function, gemm_mkl_vs_smm) { + MklSmmCompare(1, 2, 3); + MklSmmCompare(1, 2, 3); + MklSmmCompare(3, 2, 1); + MklSmmCompare(3, 2, 1); + MklSmmCompare(3, 8, 5); + MklSmmCompare(3, 8, 5); +} +#endif + +TEST(math_function, gemm_trans_cblas) { paddle::framework::Tensor input1; paddle::framework::Tensor input2; paddle::framework::Tensor input3; @@ -77,6 +133,8 @@ TEST(math_function, gemm_trans_clbas) { paddle::platform::CPUDeviceContext context(*cpu_place); GetBlas(context).GEMM(false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1, input3_ptr + 1, 4); + delete cpu_place; + cpu_place = NULL; EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); @@ -170,3 +228,57 @@ TEST(math_funciton, set_constant) { } delete ctx; } + +template +void GemmWarpTest(int m, int n, int k, T alpha, T beta) { + paddle::framework::Tensor mat_a; + paddle::framework::Tensor mat_b; + paddle::framework::Tensor mat_c_ref; + paddle::framework::Tensor mat_c_mkl; + auto* cpu_place = new paddle::platform::CPUPlace(); + + T* A = mat_a.mutable_data({m, k}, *cpu_place); + T* B = mat_b.mutable_data({k, n}, *cpu_place); + T* CREF = mat_c_ref.mutable_data({m, n}, *cpu_place); + T* CMKL = mat_c_mkl.mutable_data({m, n}, *cpu_place); + + ASSERT_EQ(mat_c_mkl.numel(), mat_c_ref.numel()); + for (int i = 0; i < mat_a.numel(); ++i) { + A[i] = static_cast(i); + } + for (int i = 0; i < mat_b.numel(); ++i) { + B[i] = static_cast(i + 1); + } + for (int i = 0; i < mat_c_ref.numel(); ++i) { + CREF[i] = static_cast(i + 2); + CMKL[i] = CREF[i]; + } + + // this would call gemm_warp + paddle::platform::CPUDeviceContext context(*cpu_place); + GetBlas(context).GEMM(CblasNoTrans, CblasNoTrans, m, n, k, alpha, A, B, + beta, CREF); + + // lda,ldb,ldc follow RowMajor + int lda = k; + int ldb = n; + int ldc = n; + paddle::operators::math::CBlas::GEMM(CblasRowMajor, CblasNoTrans, + CblasNoTrans, m, n, k, alpha, A, lda, + B, ldb, beta, CMKL, ldc); + + for (int i = 0; i < mat_c_mkl.numel(); ++i) { + EXPECT_FLOAT_EQ(CREF[i], CMKL[i]); + } +} + +TEST(math_function, gemm_warp) { + GemmWarpTest(3, 2, 5, 1.f, 0.f); + GemmWarpTest(3, 2, 5, 2.f, 1.f); + GemmWarpTest(8, 5, 6, 1.f, 0.f); + GemmWarpTest(8, 5, 6, 2.f, 1.f); + GemmWarpTest(3, 2, 5, 1.0, 0.0); + GemmWarpTest(3, 2, 5, 2.0, 1.0); + GemmWarpTest(8, 5, 6, 1.0, 0.0); + GemmWarpTest(8, 5, 6, 2.0, 1.0); +} diff --git a/paddle/fluid/operators/math/matrix_bit_code.cc b/paddle/fluid/operators/math/matrix_bit_code.cc new file mode 100644 index 0000000000000000000000000000000000000000..1e56e297396c6e37867a53f039478191f0caf08e --- /dev/null +++ b/paddle/fluid/operators/math/matrix_bit_code.cc @@ -0,0 +1,176 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/matrix_bit_code.h" +#include +namespace paddle { +namespace operators { +namespace math { + +template +void MatrixBitCodeFunctor::Add(framework::Tensor* tmat, + const framework::Tensor& vec) { + SimpleCodeTable code_table(num_classes_); + size_t batch_size = tmat->dims()[0]; + size_t width = tmat->dims()[1]; + for (size_t i = 0; i < batch_size; ++i) { + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + size_t index = code.calc_index(j); + tmat->data()[i * width + j] += vec.data()[index]; + } + } +} + +template +void MatrixBitCodeFunctor::AddGrad(const framework::Tensor& tmat, + framework::Tensor* vec) { + SimpleCodeTable code_table(num_classes_); + size_t batch_size = tmat.dims()[0]; + size_t width = tmat.dims()[1]; + for (size_t i = 0; i < batch_size; ++i) { + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + size_t index = code.calc_index(j); + vec->data()[index] += tmat.data()[i * width + j]; + } + } +} + +template +void MatrixBitCodeFunctor::Sum(const framework::Tensor& tmat, + framework::Tensor* sum, T scale_sum) { + SimpleCodeTable code_table(num_classes_); + size_t num_samples = tmat.dims()[0]; + size_t o_width = tmat.dims()[1]; + for (size_t i = 0; i < num_samples; ++i) { + T sm = static_cast(0.0); + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + if (code.calc_bit(j)) { + // calc_bit starts from right most bit, while data in tmat[i] is in the + // reverse order. + sm += tmat.data()[i * o_width + j]; + } + } + sum->data()[i] = scale_sum * sm; + } +} + +template +void MatrixBitCodeFunctor::Mul(framework::Tensor* tmat, + const framework::Tensor& weight, + const framework::Tensor& input) { + SimpleCodeTable code_table(num_classes_); + size_t num_samples = tmat->dims()[0]; + size_t tmat_width = tmat->dims()[1]; + size_t input_width = input.dims()[1]; + size_t weight_width = weight.dims()[1]; + auto tmat_value = tmat->data(); + auto weight_value = weight.data(); + auto input_value = input.data(); + for (size_t i = 0; i < num_samples; ++i) { + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + size_t index = code.calc_index(j); + T sum = static_cast(0.0); + for (size_t k = 0; k < input_width; ++k) { + sum += weight_value[weight_width * index + k] * + input_value[input_width * i + k]; + } + tmat_value[i * tmat_width + j] += sum; + } + } +} + +template +void MatrixBitCodeFunctor::MulGradWeight(const framework::Tensor& tmat, + framework::Tensor* weight, + const framework::Tensor& input) { + SimpleCodeTable code_table(num_classes_); + size_t num_samples = tmat.dims()[0]; + size_t input_width = input.dims()[1]; + size_t tmat_width = tmat.dims()[1]; + size_t weight_width = weight->dims()[1]; + auto tmat_value = tmat.data(); + auto weight_value = weight->data(); + auto input_value = input.data(); + for (size_t i = 0; i < num_samples; ++i) { + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + size_t index = code.calc_index(j); + + for (size_t k = 0; k < input_width; ++k) { + weight_value[weight_width * index + k] += + tmat_value[i * tmat_width + j] * input_value[input_width * i + k]; + } + } + } +} + +template +void MatrixBitCodeFunctor::MulGradError(const framework::Tensor& tmat, + const framework::Tensor& weight, + framework::Tensor* input) { + SimpleCodeTable code_table(num_classes_); + size_t num_samples = tmat.dims()[0]; + size_t tmat_width = tmat.dims()[1]; + size_t input_width = input->dims()[1]; + size_t weight_width = weight.dims()[1]; + auto tmat_value = tmat.data(); + auto weight_value = weight.data(); + auto input_value = input->data(); + + for (size_t i = 0; i < num_samples; ++i) { + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + size_t index = code.calc_index(j); + + for (size_t k = 0; k < input_width; ++k) { + input_value[input_width * i + k] += + tmat_value[i * tmat_width + j] * + weight_value[weight_width * index + k]; + } + } + } +} + +template +void MatrixBitCodeFunctor::Sub(framework::Tensor* tmat) { + SimpleCodeTable code_table(num_classes_); + size_t num_samples = tmat->dims()[0]; + size_t o_width = tmat->dims()[1]; + for (size_t i = 0; i < num_samples; ++i) { + auto code = code_table(static_cast(ids_[i])); + int code_length = code.get_length(); + for (int j = 0; j < code_length; ++j) { + if (code.calc_bit(j)) { + tmat->data()[i * o_width + j] -= 1; + } + } + } +} + +template class MatrixBitCodeFunctor; +template class MatrixBitCodeFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/matrix_bit_code.h b/paddle/fluid/operators/math/matrix_bit_code.h new file mode 100644 index 0000000000000000000000000000000000000000..5454d58f371afb5f5d6a1c3208318f80d4e0aa36 --- /dev/null +++ b/paddle/fluid/operators/math/matrix_bit_code.h @@ -0,0 +1,143 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace operators { +namespace math { +/** + * SimpleCodeTable class should support 3 functions: + * + * size_t size() + * return the number of ids + * + * int get_max_code_length() + * return the maximal code length + * + * SimpleCode operator()(size_t i) + * return the i-th code. Code class is descriebed below. + * + * SimpleCode class should support 3 functions: + * + * int get_length() + * return the length of the code + * + * size_t cal_index(int bit) + * bit ranges from 0 to get_length() - 1 + * return the index for the (1+bit) level parent + * + * bool calc_bit(int bit) + * return true if the bit level parent is the right child of (1+bit) level + * parent + * + */ + +/** + * return the 1-based index of the highest bit set + * + * for x > 0: + * \f[ + * FindLastSet(x) = 1 + \floor*{\log_{2}x} + * \f] + */ +inline constexpr size_t FindLastSet(size_t x) { + return std::is_same::value + ? (x ? 8 * sizeof(x) - __builtin_clz(x) : 0) + : (std::is_same::value // NOLINT + ? (x ? 8 * sizeof(x) - __builtin_clzl(x) : 0) + : (x ? 8 * sizeof(x) - __builtin_clzll(x) : 0)); +} + +struct SimpleCode { + SimpleCode(size_t code, size_t num_classes) : c_(code + num_classes) {} + /** + * Here the id of root shoud be 1 rather than 0, thus the encoding of class c + * is `c + num_classes` and all siblings can get the same weight indice using + * prefixes. + * Weight index is the prefixes of encoding, thus leave out the right most + * bit in calc_index. + * Binary classification path is the suffixes of encoding, thus leave out the + * left most bit in calc_bit. + */ + inline size_t calc_index(int bit) const { return (c_ >> (bit + 1)) - 1; } + inline bool calc_bit(int bit) const { return c_ & (1 << bit); } + inline int get_length() const { return FindLastSet(c_) - 1; } + + private: + size_t c_; +}; + +struct SimpleCodeTable { + explicit SimpleCodeTable(size_t num_classes) : num_classes_(num_classes) {} + SimpleCode operator()(size_t code) const { + return SimpleCode(code, num_classes_); + } + size_t size() const { return num_classes_; } + int get_max_code_length() const { return FindLastSet(num_classes_ - 1); } + + private: + size_t num_classes_; +}; + +template +class MatrixBitCodeFunctor { + public: + explicit MatrixBitCodeFunctor(size_t num_classes, const int64_t* ids) + : num_classes_(num_classes), ids_(ids) {} + /* For j < code_length + tmat(i, j) += vec(0, index(i, j)) + */ + void Add(framework::Tensor* tmat, const framework::Tensor& vec); + + /* For j < code_length + vec(0, index(i, j)) += tmat(i, j) + */ + void AddGrad(const framework::Tensor& tmat, framework::Tensor* vec); + + /* For j < code_length + sum(i, 0) = \sum_j bit(i, j) * tmat(i, j) + */ + void Sum(const framework::Tensor& tmat, framework::Tensor* sum, T scale_sum); + + /* For j < code_length + tmat(i, j) -= bit(i, j) + */ + void Sub(framework::Tensor* tmat); + /* For j < code_length + input.row(i) += tmat(i, j) * weight.row(index(i, j)) + */ + void Mul(framework::Tensor* tmat, const framework::Tensor& weight, + const framework::Tensor& input); + + /* For index(i, j) >= 0: + weight.row(index(i, j)) += tmat(i, j) * input.row(i) + */ + void MulGradWeight(const framework::Tensor& tmat, framework::Tensor* weight, + const framework::Tensor& input); + /* For j < code_length + input.row(i) += tmat(i, j) * weight.row(index(i, j)) + */ + void MulGradError(const framework::Tensor& tmat, + const framework::Tensor& weight, framework::Tensor* input); + + size_t num_classes_; + const int64_t* ids_; +}; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/sequence2batch.h b/paddle/fluid/operators/math/sequence2batch.h index 62e6307ae9f4236a38c49daaf09fc05c54268159..07372235a7c23832e528c3e852a4747f4244b833 100644 --- a/paddle/fluid/operators/math/sequence2batch.h +++ b/paddle/fluid/operators/math/sequence2batch.h @@ -78,7 +78,7 @@ class LoDTensor2BatchFunctor { auto lods = lod_tensor.lod(); PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now."); - auto lod = lods[0]; + const auto& lod = lods[0]; std::vector seq_info; for (size_t seq_id = 0; seq_id < lod.size() - 1; ++seq_id) { diff --git a/paddle/fluid/operators/math/softmax.cu b/paddle/fluid/operators/math/softmax.cu index a579182ec1bd5d10d95bbf8c6f5a0e70ceaaaf4b..3effe776258cb541dbba32f63eda457d917011f4 100644 --- a/paddle/fluid/operators/math/softmax.cu +++ b/paddle/fluid/operators/math/softmax.cu @@ -52,7 +52,7 @@ void SoftmaxCUDNNFunctor::operator()( xDesc.descriptor(layout, cudnn_tensor_dims); cudnnTensorDescriptor_t cudnn_y_desc = xDesc.descriptor(layout, cudnn_tensor_dims); - PADDLE_ENFORCE(platform::dynload::cudnnSoftmaxForward( + CUDNN_ENFORCE(platform::dynload::cudnnSoftmaxForward( context.cudnn_handle(), CUDNN_SOFTMAX_ACCURATE, CUDNN_SOFTMAX_MODE_INSTANCE, CudnnDataType::kOne(), cudnn_x_desc, X->data(), CudnnDataType::kZero(), cudnn_y_desc, @@ -83,7 +83,7 @@ void SoftmaxGradCUDNNFunctor::operator()( dxDesc.descriptor(layout, cudnn_tensor_dims); cudnnTensorDescriptor_t cudnn_ygrad_desc = dyDesc.descriptor(layout, cudnn_tensor_dims); - PADDLE_ENFORCE(platform::dynload::cudnnSoftmaxBackward( + CUDNN_ENFORCE(platform::dynload::cudnnSoftmaxBackward( context.cudnn_handle(), CUDNN_SOFTMAX_ACCURATE, CUDNN_SOFTMAX_MODE_INSTANCE, CudnnDataType::kOne(), cudnn_y_desc, Y->data(), cudnn_ygrad_desc, YGrad->data(), diff --git a/paddle/fluid/operators/max_sequence_len_op.cc b/paddle/fluid/operators/max_sequence_len_op.cc index 8e508b68eeab69a4595904dcc3ea0a541d9ab6e6..b1e69f375d3274aade3184af02f7f914dba5db71 100644 --- a/paddle/fluid/operators/max_sequence_len_op.cc +++ b/paddle/fluid/operators/max_sequence_len_op.cc @@ -42,10 +42,15 @@ class MaxSeqenceLenOp : public framework::OperatorBase { class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("RankTable", "The lod_rank_table."); - AddOutput("Out", "The max sequence length."); - AddComment( - R"DOC(Calculate the max sequence length through lod_rank_table.)DOC"); + AddInput("RankTable", "Input variable which is a LoDRankTable object"); + AddOutput("Out", "The max sequence length"); + AddComment(R"DOC( + Given a LoDRankTable object, this layer returns the max length of + a batch of sequences. In fact, a LoDRankTable object contains a list of + tuples() and the list is already sorted by + sequence length in descending order, so the operator just returns the + sequence length of the first tuple element +)DOC"); } }; diff --git a/paddle/fluid/operators/mean_iou_op.cc b/paddle/fluid/operators/mean_iou_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a60f245f53e342fd9c1382fdda33a011a7fb06d6 --- /dev/null +++ b/paddle/fluid/operators/mean_iou_op.cc @@ -0,0 +1,110 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/mean_iou_op.h" + +namespace paddle { +namespace operators { + +class MeanIoUOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Predictions"), + "Input (Predictions) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input (labels) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutMeanIou"), + "Output (OutMeanIou) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutWrong"), + "Output (OutWrong) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutCorrect"), + "Output (OutWrong) of MeanIoU op should not be null."); + + int64_t num_classes = + static_cast(ctx->Attrs().Get("num_classes")); + + ctx->SetOutputDim("OutMeanIou", {1}); + ctx->SetOutputDim("OutWrong", {num_classes}); + ctx->SetOutputDim("OutCorrect", {num_classes}); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Predictions")->type()), + ctx.GetPlace()); + } +}; + +class MeanIoUOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Predictions", + "(Tensor), A Tensor of prediction results for semantic labels" + " with type int32 or int64. The rank should be greater than 1."); + AddInput( + "Labels", + "(Tensor), A Tensor of ground truth labels with type int32 or int64." + "Its shape should be the same as Input(Predictions)."); + AddInput("InWrongs", + "(vector), A list of Tensor with shape " + "[num_classes]. They are used to collect wrong number among " + "batches. Empty list is also valid here.") + .AsDuplicable() + .AsDispensable(); + AddInput( + "InCorrects", + "(vector), A list of Tensor with shape " + "[num_classes]. They are used to collect correct number among batches. " + "Empty list is also valid here.") + .AsDuplicable() + .AsDispensable(); + AddInput("InMeanIou", + "(vector), A list of Tensor that Output(mean_iou) should " + "be added to. Empty list is also valid here.") + .AsDuplicable() + .AsDispensable(); + AddOutput("OutMeanIou", + "(vector), A Tensor representing the" + " mean intersection-over-union with shape [1]."); + AddOutput("OutWrong", "(Tensor), A Tensor with shape [num_classes]. "); + AddOutput("OutCorrect", "(Tensor), A Tensor with shape [num_classes]. "); + AddAttr("num_classes", "(int), The possible number of labels."); + + AddComment(R"DOC( +mean-IOU Operator. +Mean Intersection-Over-Union is a common evaluation metric for +semantic image segmentation, which first computes the IOU for each +semantic class and then computes the average over classes. +IOU is defined as follows: + IOU = true_positive / (true_positive + false_positive + false_negative). +It is based on pixel level area while "IOU Similarity Operator" +is based on area of rectangle. + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(mean_iou, ops::MeanIoUOp, ops::MeanIoUOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(mean_iou, ops::MeanIoUKernel, + ops::MeanIoUKernel, + ops::MeanIoUKernel); diff --git a/paddle/fluid/operators/mean_iou_op.cu b/paddle/fluid/operators/mean_iou_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..83bb4dde46fa241affad3788e3381b6ecd8aa098 --- /dev/null +++ b/paddle/fluid/operators/mean_iou_op.cu @@ -0,0 +1,164 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/mean_iou_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/gpu_info.h" + +namespace paddle { +namespace operators { + +using platform::PADDLE_CUDA_NUM_THREADS; + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__global__ void CountCUDAKernel(const int num_classes, const int count, + const T* predictions, const T* labels, + int* wrong, int* correct) { + extern __shared__ int blcok_cache[]; + int* wrong_c = blcok_cache; + int* correct_c = blcok_cache + num_classes; + // init cache + for (int i = threadIdx.x; i < num_classes * 2; i += blockDim.x) { + blcok_cache[i] = 0; + } + __syncthreads(); + + T pred; + T label; + CUDA_1D_KERNEL_LOOP(i, count) { + pred = predictions[i]; + label = labels[i]; + if (pred == label) { + atomicAdd(correct_c + pred, 1); + } else { + atomicAdd(wrong_c + pred, 1); + atomicAdd(wrong_c + label, 1); + } + } + + __syncthreads(); + + for (int i = threadIdx.x; i < num_classes; i += blockDim.x) { + atomicAdd(wrong + i, wrong_c[i]); + atomicAdd(correct + i, correct_c[i]); + } +} + +__global__ void ComputeIoUCUDAKernel(const int num_classes, int* wrong, + int* correct, float* ious, float* iou) { + __shared__ int valid_count_c; + if (threadIdx.x == 0) { + valid_count_c = 0; + } + __syncthreads(); + CUDA_1D_KERNEL_LOOP(i, num_classes) { + int wrong_n = wrong[i]; + int correct_n = correct[i]; + int denominator = wrong_n + correct_n; + if (denominator > 0) { + atomicAdd(&valid_count_c, 1); + ious[i] = static_cast(correct_n) / denominator; + } else { + ious[i] = 0; + } + } + __syncthreads(); + if (threadIdx.x == 0) { + float iou_sum = 0; + for (int i = 0; i < num_classes; ++i) { + iou_sum += ious[i]; + } + iou[0] += iou_sum / valid_count_c; + } +} + +template +class MeanIoUCUDAOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& place = *ctx.template device_context() + .eigen_device(); + // get input and output tensor + auto* predictions = ctx.Input("Predictions"); + auto* labels = ctx.Input("Labels"); + auto* out_mean_iou = ctx.Output("OutMeanIou"); + auto* out_wrong = ctx.Output("OutWrong"); + auto* out_correct = ctx.Output("OutCorrect"); + int num_classes = static_cast(ctx.Attr("num_classes")); + + // Get data ptr + const T* predictions_data = predictions->data(); + const T* labels_data = labels->data(); + int* out_wrong_data = out_wrong->mutable_data(ctx.GetPlace()); + int* out_correct_data = out_correct->mutable_data(ctx.GetPlace()); + float* out_mean_iou_data = + out_mean_iou->mutable_data(ctx.GetPlace()); + + // Get Eigen tensor + auto out_mean_iou_t = EigenTensor::From(*out_mean_iou); + auto out_wrong_t = EigenTensor::From(*out_wrong); + auto out_correct_t = EigenTensor::From(*out_correct); + + // Temporary tensor + Tensor ious; + float* ious_data = ious.mutable_data( + {static_cast(num_classes)}, ctx.GetPlace()); + auto ious_t = EigenTensor::From(ious); + + // Init out_wrong, out_correct and out_mean_iou + out_wrong_t.device(place) = out_wrong_t.constant(0); + out_correct_t.device(place) = out_correct_t.constant(0); + out_mean_iou_t.device(place) = out_mean_iou_t.constant(0.0f); + + // collect pre wrong, correct and mean_iou + auto in_mean_ious = ctx.MultiInput("InMeanIou"); + for (int i = 0; i < in_mean_ious.size(); ++i) { + out_mean_iou_t.device(place) += + EigenTensor::From(*in_mean_ious[i]); + } + auto in_wrongs = ctx.MultiInput("InWrongs"); + for (int i = 0; i < in_wrongs.size(); ++i) { + out_wrong_t.device(place) += EigenTensor::From(*in_wrongs[i]); + } + auto in_corrects = ctx.MultiInput("InCorrects"); + for (int i = 0; i < in_corrects.size(); ++i) { + out_correct_t.device(place) += EigenTensor::From(*in_corrects[i]); + } + // compute + auto stream = ctx.cuda_device_context().stream(); + int block = PADDLE_CUDA_NUM_THREADS; + int grid = (predictions->numel() + block - 1) / block; + int cache_size = (num_classes * 2 + 1) * sizeof(int); + CountCUDAKernel<<>>( + num_classes, predictions->numel(), predictions_data, labels_data, + out_wrong_data, out_correct_data); + ctx.device_context().Wait(); + ComputeIoUCUDAKernel<<<1, block, 0, stream>>>(num_classes, out_wrong_data, + out_correct_data, ious_data, + out_mean_iou_data); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(mean_iou, ops::MeanIoUCUDAOpKernel, + ops::MeanIoUCUDAOpKernel, + ops::MeanIoUCUDAOpKernel); diff --git a/paddle/fluid/operators/mean_iou_op.h b/paddle/fluid/operators/mean_iou_op.h new file mode 100644 index 0000000000000000000000000000000000000000..9fa00e60e05504e0bb8658c6908e4d4ac46b2ca4 --- /dev/null +++ b/paddle/fluid/operators/mean_iou_op.h @@ -0,0 +1,117 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +using Tensor = framework::Tensor; + +template +using EigenTensor = framework::EigenTensor; + +template +class MeanIoUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& place = *ctx.template device_context() + .eigen_device(); + // get input and output tensor + auto* predictions = ctx.Input("Predictions"); + auto* labels = ctx.Input("Labels"); + auto* out_mean_iou = ctx.Output("OutMeanIou"); + auto* out_wrong = ctx.Output("OutWrong"); + auto* out_correct = ctx.Output("OutCorrect"); + int num_classes = static_cast(ctx.Attr("num_classes")); + + // get data ptr + const T* predictions_data = predictions->data(); + const T* labels_data = labels->data(); + float* out_mean_iou_data = + out_mean_iou->mutable_data(ctx.GetPlace()); + int* out_wrong_data = out_wrong->mutable_data(ctx.GetPlace()); + int* out_correct_data = out_correct->mutable_data(ctx.GetPlace()); + + // get eigen tensor + auto out_mean_iou_t = EigenTensor::From(*out_mean_iou); + auto out_wrong_t = EigenTensor::From(*out_wrong); + auto out_correct_t = EigenTensor::From(*out_correct); + + // Tmp tensor + Tensor denominator; + Tensor valid_count; + Tensor iou_sum; + + // get data ptr of tmp tensor + int* denominator_data = denominator.mutable_data( + {static_cast(num_classes)}, ctx.GetPlace()); + int* valid_count_data = valid_count.mutable_data({1}, ctx.GetPlace()); + float* iou_sum_data = iou_sum.mutable_data({1}, ctx.GetPlace()); + + // get eigen tensor of tmp tensor + auto denominator_t = EigenTensor::From(denominator); + auto valid_count_t = EigenTensor::From(valid_count); + auto iou_sum_t = EigenTensor::From(iou_sum); + + // init out_wrong, out_correct and out_mean_iou + out_wrong_t = out_wrong_t.constant(0); + out_correct_t = out_correct_t.constant(0); + out_mean_iou_t = out_mean_iou_t.constant(0); + + // collect pre wrong, correct and mean_iou + auto in_mean_ious = ctx.MultiInput("InMeanIou"); + for (size_t i = 0; i < in_mean_ious.size(); ++i) { + out_mean_iou_t.device(place) += + EigenTensor::From(*in_mean_ious[i]); + } + auto in_wrongs = ctx.MultiInput("InWrongs"); + for (size_t i = 0; i < in_wrongs.size(); ++i) { + out_wrong_t.device(place) += EigenTensor::From(*in_wrongs[i]); + } + auto in_corrects = ctx.MultiInput("InCorrects"); + for (size_t i = 0; i < in_corrects.size(); ++i) { + out_correct_t.device(place) += EigenTensor::From(*in_corrects[i]); + } + + // compute + for (int64_t i = 0; i < predictions->numel(); ++i) { + if (predictions_data[i] == labels_data[i]) { + out_correct_data[predictions_data[i]] += 1; + } else { + out_wrong_data[labels_data[i]] += 1; + out_wrong_data[predictions_data[i]] += 1; + } + } + + denominator_t = out_wrong_t + out_correct_t; + valid_count_t = + (denominator_t > denominator_t.constant(0.0f)).cast().sum(); + + for (int i = 0; i < num_classes; ++i) { + if (denominator_data[i] == 0) { + denominator_data[i] = 1; + } + } + + iou_sum_t = + (out_correct_t.cast() / denominator_t.cast()).sum(); + out_mean_iou_data[0] += (iou_sum_data[0] / valid_count_data[0]); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc index 74477eb439dc202c3f5f17fdf3e1647bc5c23512..9e0bebd17c02a3ce010b77142757b8789cfbcdd9 100644 --- a/paddle/fluid/operators/mean_op.cc +++ b/paddle/fluid/operators/mean_op.cc @@ -33,12 +33,10 @@ class MeanOp : public framework::OperatorWithKernel { class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "The input of mean op"); - AddOutput("Out", "The output of mean op"); + AddInput("X", "(Tensor) The input of mean op"); + AddOutput("Out", "(Tensor) The output of mean op").Reuse("X"); AddComment(R"DOC( -Mean Operator. - -Out is a scalar which is the mean of all elements in X. +Mean Operator calculates the mean of all elements in X. )DOC"); } diff --git a/paddle/fluid/operators/merge_ids_op.cc b/paddle/fluid/operators/merge_ids_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c6ec4ab047d5e91625e646fd26108d2e477cdce5 --- /dev/null +++ b/paddle/fluid/operators/merge_ids_op.cc @@ -0,0 +1,128 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/merge_ids_op.h" + +namespace paddle { +namespace operators { + +class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); + AddInput( + "X", + "(LoDTensors) multi input tensor with shape{batch_num, N}, N is the " + "size of embedding table") + .AsDuplicable(); + AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors."); + + AddComment(R"DOC( +Merge multi LoDTensor's into one according to Ids's shard num. + + +split_ids_op -> prefetch_op -> merge_ids_op + + +merge_ids_op should be used after split_ids_op and prefetch_op, split_ids_op + will split input Ids into multiple tensors according to Id's shard number. +prefetch_op will send them to parameter server to prefetch embedding value +back. During split, the order of ids is disordered. In merge_ids_op we use +the original Ids to restore the order of the fetched embedding value and + also pass the lod information to the merged output. + + +Example: + + Ids = [1,2,3,4,5,6] # 3 shared + +split_ids_op -> + + Id0 = [3, 6] # id % 3 == 0 + Id1 = [1, 4] # id % 3 == 1 + Id2 = [2, 5] # id % 3 == 2 + +prefetch_op -> + + X0 = [[0.3 0.3] # 3 + [0.6 0.6]] # 6 + X1 = [[0.1 0.1] # 1 + [0.4 0.4]] # 4 + X2 = [[0.2 0.2] # 2 + [0.5 0.5]] # 5 + +merge_ids_op -> + + Out = [[0.1 0.1] # 1 + [0.2 0.2] # 2 + [0.3 0.3] # 3 + [0.4 0.4] # 4 + [0.5 0.5] # 5 + [0.6 0.6]] # 6 +)DOC"); + } +}; + +class MergeIdsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids."); + PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out."); + + auto ids_var_type = ctx->GetInputsVarType("Ids").front(); + auto ids_dims = ctx->GetInputDim("Ids"); + if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { + PADDLE_ENFORCE_EQ(ids_dims.size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[1], 1); + } + auto x_var_type = ctx->GetInputsVarType("X"); + for (auto &var_type : x_var_type) { + PADDLE_ENFORCE_EQ(var_type, framework::proto::VarType::LOD_TENSOR, + "input X only support lod tensors"); + } + ctx->ShareLoD("Ids", "Out"); + } + + private: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.MultiInput("X").front()->type()), + ctx.GetPlace()); + } +}; + +class MergeIdsOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto *input_var = block->Var(op_desc.Input("Ids")[0]); + for (auto &out_var : op_desc.Output("Out")) { + block->Var(out_var)->SetType(input_var->GetType()); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(merge_ids, ops::MergeIdsOp, ops::MergeIdsOpMaker, + ops::MergeIdsOpInferVarType); +REGISTER_OP_CPU_KERNEL( + merge_ids, ops::MergeIdsOpKernel); diff --git a/paddle/fluid/operators/merge_ids_op.h b/paddle/fluid/operators/merge_ids_op.h new file mode 100644 index 0000000000000000000000000000000000000000..83712a8519c6817151e1922c606c0fdd4682a2db --- /dev/null +++ b/paddle/fluid/operators/merge_ids_op.h @@ -0,0 +1,92 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" + +namespace paddle { +namespace operators { + +template +class MergeIdsOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto place = ctx.GetPlace(); + if (!platform::is_cpu_place(place)) { + PADDLE_THROW("MergeIds do not support GPU kernel"); + } + VLOG(3) << "run in MergeIdsOpKernel"; + + const auto *ids_var = ctx.InputVar("Ids"); + PADDLE_ENFORCE(ids_var->IsType(), + "only support to merge Ids of LoDTensor"); + + const auto &ids_tensor = ids_var->Get(); + const auto &ids_dims = ids_tensor.dims(); + const int64_t *ids = ids_tensor.data(); + + auto x_tensors = ctx.MultiInput("X"); + + auto *out = ctx.Output("Out"); + + int batch_size = 0; + int embedding_size = 0; + for (auto &input : x_tensors) { + if (framework::product(input->dims()) != 0) { + if (embedding_size == 0) { + embedding_size = input->dims()[1]; + } + PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1], + "embedding size of all input should be the same"); + batch_size += input->dims()[0]; + } + } + PADDLE_ENFORCE_EQ( + batch_size, ids_dims[0], + "the batch size of ids and merged embedding value should be the same"); + + const size_t shard_num = x_tensors.size(); + + if (shard_num == 1) { + VLOG(3) << "only one shard, we can copy the data directly"; + TensorCopy(*x_tensors[0], place, out); + } else { + std::vector in_indexs(shard_num, 0); + auto *out_data = out->mutable_data( + framework::make_ddim({batch_size, embedding_size}), place); + // copy data from ins[shard_num] to out. + for (int i = 0; i < ids_dims[0]; ++i) { + int64_t id = ids[i]; + size_t shard_id = static_cast(id) % shard_num; + int index = in_indexs[shard_id]; + memcpy(out_data + embedding_size * i, + x_tensors[shard_id]->data() + index * embedding_size, + sizeof(T) * embedding_size); + in_indexs[shard_id] += 1; + } + + for (size_t i = 0; i < shard_num; ++i) { + PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0], + "after merge, all data in x_tensor should be used"); + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/merge_lod_tensor_op.cc b/paddle/fluid/operators/merge_lod_tensor_op.cc index a16861b3b77fc980ab932b9d88859b38ec36108b..2dc1467b0d4816d5cc0535eb62e936cf342a241c 100644 --- a/paddle/fluid/operators/merge_lod_tensor_op.cc +++ b/paddle/fluid/operators/merge_lod_tensor_op.cc @@ -44,8 +44,10 @@ class MergeLoDTensorOp : public framework::OperatorBase { scope.FindVar(Output("Out"))->GetMutable(); auto level = static_cast(Attr("level")); - auto &mask_dim = mask.dims(); + PADDLE_ENFORCE(in_true.numel() || in_false.numel(), + "Input(InTrue) or Input(InFalse) should be initialized."); + auto &mask_dim = mask.dims(); std::unique_ptr cpu_mask{new framework::LoDTensor()}; if (platform::is_cpu_place(mask.place())) { cpu_mask->ShareDataWith(mask); @@ -59,19 +61,27 @@ class MergeLoDTensorOp : public framework::OperatorBase { } auto *mask_data = cpu_mask->data(); - int rank = in_true.dims().size(); - platform::Place place = in_true.place(); - std::type_index data_type = in_true.type(); - framework::DDim in_true_dims = - framework::slice_ddim(in_true.dims(), 1, rank); - + platform::Place place = dev_place; int64_t batch_size = in_true.dims()[0] + in_false.dims()[0]; - auto in_true_dim_vec = framework::vectorize(in_true_dims); - in_true_dim_vec.insert(in_true_dim_vec.begin(), batch_size); + std::type_index data_type = + in_true.IsInitialized() ? in_true.type() : in_false.type(); + int rank; + framework::DDim in_dims; + if (in_true.IsInitialized()) { + rank = in_true.dims().size(); + in_dims = framework::slice_ddim(in_true.dims(), 1, rank); + } else { + rank = in_false.dims().size(); + in_dims = framework::slice_ddim(in_false.dims(), 1, rank); + } + + auto in_dim_vec = framework::vectorize(in_dims); + in_dim_vec.insert(in_dim_vec.begin(), batch_size); - framework::DDim out_dims = framework::make_ddim(in_true_dim_vec); + framework::DDim out_dims = framework::make_ddim(in_dim_vec); out->Resize(out_dims); + out->mutable_data(place, data_type); auto *out_lod = out->mutable_lod(); diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index dcd73e3c3e40f80e07b73944d1f0cc57fea010d3..5f43c5810812260c4384349bdb709716c9a182f5 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -98,7 +98,7 @@ The update equations are as follows: $$ velocity = mu * velocity + gradient \\ if (use\_nesterov): \\ - param = param - gradient * learning\_rate + mu * velocity * learning\_rate \\ + param = param - (gradient + mu * velocity) * learning\_rate \\ else: \\ param = param - learning\_rate * velocity. \\ $$ diff --git a/paddle/fluid/operators/momentum_op.cu b/paddle/fluid/operators/momentum_op.cu index 5eb9d9950248bb50bb823f071c7fff0ddcc47234..a3932db1f3a50305d585cd3d5e86fa1b527df78b 100644 --- a/paddle/fluid/operators/momentum_op.cu +++ b/paddle/fluid/operators/momentum_op.cu @@ -30,7 +30,7 @@ __global__ void MomentumKernel(const T* p, const T* g, const T* v, T g_val = g[i]; T v_new = v[i] * mu + g_val; v_out[i] = v_new; - p_out[i] = p[i] - (g_val - v_new * mu) * lr; + p_out[i] = p[i] - (g_val + v_new * mu) * lr; } } else { for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; diff --git a/paddle/fluid/operators/momentum_op.h b/paddle/fluid/operators/momentum_op.h index 04a1929b84a93af6465bacfe7974a1530296946d..264726040fb566a52b8c0cdee0a1524197d2a675 100644 --- a/paddle/fluid/operators/momentum_op.h +++ b/paddle/fluid/operators/momentum_op.h @@ -46,7 +46,7 @@ class MomentumOpKernel : public framework::OpKernel { v_out = v * mu + g; if (use_nesterov) { - p_out = p - (g - v_out * mu) * lr[0]; + p_out = p - (g + v_out * mu) * lr[0]; } else { p_out = p - lr[0] * v_out; } diff --git a/paddle/fluid/operators/multiplex_op.cc b/paddle/fluid/operators/multiplex_op.cc index a4363fd25d57edb5c2509904a1f55634832613be..18ad46cb5eeeab2169136e40cebdaa53c0bfd587 100644 --- a/paddle/fluid/operators/multiplex_op.cc +++ b/paddle/fluid/operators/multiplex_op.cc @@ -62,26 +62,46 @@ class MultiplexOp : public framework::OperatorWithKernel { class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Ids", "The index tensor of multiplex operator."); - AddInput("X", "The candidate tensors of multiplex operator.") + AddInput("Ids", + "Tensor, index variable which is a 2-D tensor with shape " + "[M, 1] where M is the batch size."); + AddInput("X", + "A list of variables to gather from. All variables have the same " + "shape and the rank is at least 2.") .AsDuplicable(); AddOutput("Out", "The output tensor of multiplex operator."); AddComment(R"DOC( -Multiplex Operator. - -Multiplex multiple tensors according to the index provided by the index tensor. - -Ids: the index tensor. -X[0 : N - 1]: the candidate tensors for output (N >= 2). -For each index i from 0 to batchSize - 1, the output is the i-th row of the +Referring to the given index variable, this layer selects rows from the +input variables to construct a multiplex variable. Assuming that there are +:math:`m` input variables and :math:`I_i` represents the i-th input +variable and :math:`i` is in [0, :math:`m`). All input variables are +tensors with same shape [:math:`d_0`, :math:`d_1`, ..., :math:`d_R`]. +Please note that rank of the input tensor should be at least 2. Each input +variable will be treated as a 2-D matrix with shape [:math:`M`, :math:`N`] +where :math:`M` for :math:`d_0` and :math:`N` for :math:`d_1` * :math:`d_2` +* ... * :math:`d_R`. Let :math:`I_i[j]` be the j-th row of the i-th input +variable. The given index variable should be a 2-D tensor with shape +[:math:`M`, 1]. Let `ID[i]` be the i-th index value of the index variable. +Then the output variable will be a tensor with shape [:math:`d_0`, +:math:`d_1`, ..., :math:`d_R`]. If we treat the output tensor as a 2-D +matrix with shape [:math:`M`, :math:`N`] and let :math:`O[i]` be the i-th +row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`. + +* Ids: the index tensor. + +* X[0 : N - 1]: the candidate tensors for output (N >= 2). + +* For each index i from 0 to batchSize - 1, the output is the i-th row of the the (Ids[i])-th tensor. For i-th row of the output tensor: -$$y[i] = x_{k}[i]$$ +$$ +y[i] = x_{k}[i] +$$ -where `y` is the output tensor, `x_{k}` is the k-th input tensor, -and `k = Ids[i]`. +where $y$ is the output tensor, $x_{k}$ is the k-th input tensor, +and $k = Ids[i]$. )DOC"); } diff --git a/paddle/fluid/operators/nccl_op_test.cu.cc b/paddle/fluid/operators/nccl_op_test.cu.cc index ef54d79fdf2becde98c68044d14bd4347773b975..d5fb7a12e5d9757f3e639f6de7f0129bd531e2a1 100644 --- a/paddle/fluid/operators/nccl_op_test.cu.cc +++ b/paddle/fluid/operators/nccl_op_test.cu.cc @@ -19,7 +19,6 @@ limitations under the License. */ #include // NOLINT #include -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/op_desc.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/program_desc.h" @@ -27,6 +26,7 @@ limitations under the License. */ #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/gpu_info.h" +#include "paddle/fluid/platform/init.h" #include "paddle/fluid/platform/place.h" USE_NO_KERNEL_OP(ncclInit); diff --git a/paddle/fluid/operators/nce_op.cc b/paddle/fluid/operators/nce_op.cc index 06092e680a1efbef379ccf40fdf476769f820429..e471f04662a1fa3e8e77a2db37f0da4521682018 100644 --- a/paddle/fluid/operators/nce_op.cc +++ b/paddle/fluid/operators/nce_op.cc @@ -128,8 +128,10 @@ class NCEOpMaker : public framework::OpProtoAndCheckerMaker { "user should avoid setting this attribute.") .SetDefault({}); AddComment(R"DOC( -Compute and return the noise-contrastive estimation training loss. -See [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf). +Compute and return the noise-contrastive estimation training loss. See +`Noise-contrastive estimation: A new estimation principle for unnormalized +statistical models + `_. By default this operator uses a uniform distribution for sampling. )DOC"); } diff --git a/paddle/fluid/operators/norm_op.cc b/paddle/fluid/operators/norm_op.cc index cdbc975c02214721ceae3a338741101ef32d7ee9..aa19c62c83648814e86b1e7062424be3693e4b98 100644 --- a/paddle/fluid/operators/norm_op.cc +++ b/paddle/fluid/operators/norm_op.cc @@ -16,40 +16,34 @@ limitations under the License. */ namespace paddle { namespace operators { -template class NormOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput( - "X", - "(Tensor) The input tensor of norm operator. " - "The format of input tensor is NCHW. Where N is batch size, C is the " - "number of channels, H and W is the height and width of feature."); - AddInput("Scale", - "(Tensor) The input tensor of norm operator. " - "The format of input tensor is C * 1."); - AddAttr("epsilon", - "(float, default 1e-10) Constant " - "for numerical stability.") + AddInput("X", "(Tensor) A tensor of rank >= axis."); + AddAttr("axis", + "The axis on which to apply normalization. If axis < 0, " + "the dimension to normalization is rank(X) + axis. -1 is " + "the last dimension."); + AddAttr("epsilon", + "(float, default 1e-10) The epsilon value is used " + "to avoid division by zero.") .SetDefault(1.0e-10f); - AddOutput("Out", - "(Tensor) The output tensor of norm operator." - "N * M." - "M = C * H * W"); + AddOutput("Norm", + "(Tensor) A tensor saved the `sqrt(sum(x) + epsion)` will " + "be used in backward kernel.") + .AsIntermediate(); + AddOutput("Out", "(Tensor) A tensor of the same shape as X."); AddComment(R"DOC( - "Input shape: $(N, C, H, W)$ - Scale shape: $(C, 1)$ - Output shape: $(N, C, H, W)$ - Where - forward - $$ - [\frac {x_{1}}{\sqrt{\sum{x_{i}^{2}}}} \frac {x_{2}}{\sqrt{\sum{x_{i}^{2}}}} \frac {x_{3}}{\sqrt{\sum{x_{i}^{2}}}} \cdot \cdot \cdot \frac {x_{n}}{\sqrt{\sum{x_{i}^{2}}}}] - $$ - backward - $$ - \frac{\frac{\mathrm{d}L }{\mathrm{d}y_{1}} - \frac {x_{1}\sum {\frac{\mathrm{d} L}{\mathrm{d} y_{j}}}x_{j}}{\sum x_{j}^{2}} }{\sqrt{\sum{x_{j}^{2}}}} - $$ - )DOC"); + +Given a tensor, apply 2-normalization along the provided axis. + +$$ +y = \frac{x}{ \sqrt{\sum {x^2} + epsion }} +$$ + +where, $\sum {x^2}$ is calculated along the `axis` dimension. + +)DOC"); } }; @@ -58,15 +52,15 @@ class NormOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of NormOp" - "should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Scale"), - "Input(Scale) of NormOp" - "should not be null."); + "Input(X) of NormOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of NormOp should not be null."); - auto in_x_dims = ctx->GetInputDim("X"); - ctx->SetOutputDim("Out", in_x_dims); + auto xdim = ctx->GetInputDim("X"); + ctx->SetOutputDim("Out", xdim); + int axis = ctx->Attrs().Get("axis"); + if (axis < 0) axis = xdim.size() + axis; + xdim[axis] = 1; + ctx->SetOutputDim("Norm", xdim); } }; @@ -84,12 +78,12 @@ class NormOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(norm, ops::NormOp, ops::NormOpMaker, +using CPU = paddle::platform::CPUDeviceContext; + +REGISTER_OPERATOR(norm, ops::NormOp, ops::NormOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(norm_grad, ops::NormOpGrad); -REGISTER_OP_CPU_KERNEL( - norm, ops::NormKernel, - ops::NormKernel); -REGISTER_OP_CPU_KERNEL( - norm_grad, ops::NormGradKernel, - ops::NormGradKernel); +REGISTER_OP_CPU_KERNEL(norm, ops::NormKernel, + ops::NormKernel); +REGISTER_OP_CPU_KERNEL(norm_grad, ops::NormGradKernel, + ops::NormGradKernel); diff --git a/paddle/fluid/operators/norm_op.cu b/paddle/fluid/operators/norm_op.cu index d1d9be50742b54a3b6f068fd43ec4b16696183bf..1d0021d33ff9ee65c3366183466b94266e6c2999 100644 --- a/paddle/fluid/operators/norm_op.cu +++ b/paddle/fluid/operators/norm_op.cu @@ -16,9 +16,9 @@ limitations under the License. */ #include "paddle/fluid/operators/norm_op.h" namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - norm, ops::NormKernel, - ops::NormKernel); -REGISTER_OP_CUDA_KERNEL( - norm_grad, ops::NormGradKernel, - ops::NormGradKernel); +using CUDA = paddle::platform::CUDADeviceContext; + +REGISTER_OP_CUDA_KERNEL(norm, ops::NormKernel, + ops::NormKernel); +REGISTER_OP_CUDA_KERNEL(norm_grad, ops::NormGradKernel, + ops::NormGradKernel); diff --git a/paddle/fluid/operators/norm_op.h b/paddle/fluid/operators/norm_op.h index 0ad29e8a0385c46a07842930378ed7a040564437..3167bdc8ac718b23435690577e4163826d14a332 100644 --- a/paddle/fluid/operators/norm_op.h +++ b/paddle/fluid/operators/norm_op.h @@ -19,156 +19,110 @@ limitations under the License. */ namespace paddle { namespace operators { -template +inline void GetDims(const framework::DDim& dim, int axis, int* pre, int* n, + int* post) { + *pre = 1; + *post = 1; + *n = dim[axis]; + for (int i = 0; i < axis; ++i) { + (*pre) *= dim[i]; + } + for (int i = axis + 1; i < dim.size(); ++i) { + (*post) *= dim[i]; + } +} + +template class NormKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - const framework::Tensor* in_x = context.Input("X"); - const framework::Tensor* scale = context.Input("Scale"); - auto* out = context.Output("Out"); - auto epsilon = static_cast(context.Attr("epsilon")); - out->mutable_data(context.GetPlace()); - int batch_size = in_x->dims()[0]; - int channels = in_x->dims()[1]; - int height = in_x->dims()[2]; - int width = in_x->dims()[3]; - int fea_len = height * width; - auto* place = - context.template device_context().eigen_device(); - auto x = - framework::EigenMatrix::From( - *in_x, framework::make_ddim({batch_size, fea_len * channels})); - // get square - framework::Tensor x_square; - x_square.mutable_data(in_x->dims(), context.GetPlace()); - auto x_square_eigen = - framework::EigenMatrix::From( - x_square, framework::make_ddim({batch_size, fea_len * channels})); - x_square_eigen.device(*place) = x.square(); - auto scale_eigen = - framework::EigenVector::Flatten( - *scale); - for (int n = 0; n < batch_size; ++n) { - framework::Tensor in_x_batch = in_x->Slice(n, n + 1); - auto in_x_batch_eigen = - framework::EigenMatrix::From( - in_x_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor x_square_batch = x_square.Slice(n, n + 1); - auto x_square_batch_eigen = - framework::EigenMatrix::From( - x_square_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor out_batch = out->Slice(n, n + 1); - auto out_batch_eigen = - framework::EigenMatrix::From( - out_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor tmp_tensor; - tmp_tensor.mutable_data(framework::make_ddim({1, fea_len}), - context.GetPlace()); - auto tmp = framework::EigenVector::Flatten(tmp_tensor); - // get colsum and sqrt , inverse - auto dim = Eigen::array({{0}}); - tmp.device(*place) = x_square_batch_eigen.sum(dim); - tmp.device(*place) = (tmp + epsilon).sqrt().inverse(); - Eigen::array broadcast_dim_col; - broadcast_dim_col[1] = 1; - broadcast_dim_col[0] = channels; - out_batch_eigen.device(*place) = - in_x_batch_eigen * (tmp.broadcast(broadcast_dim_col)); - Eigen::array broadcast_dim_row; - broadcast_dim_row[1] = fea_len; - broadcast_dim_row[0] = 1; - out_batch_eigen.device(*place) = - out_batch_eigen * (scale_eigen.broadcast(broadcast_dim_row)); - } + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_x = ctx.Input("X"); + auto* out_y = ctx.Output("Out"); + auto* out_norm = ctx.Output("Norm"); + out_y->mutable_data(ctx.GetPlace()); + out_norm->mutable_data(ctx.GetPlace()); + + auto xdim = in_x->dims(); + auto ndim = out_norm->dims(); + T eps = static_cast(ctx.Attr("epsilon")); + int axis = ctx.Attr("axis"); + if (axis < 0) axis = xdim.size() + axis; + int pre, n, post; + GetDims(xdim, axis, &pre, &n, &post); + + auto* place = ctx.template device_context().eigen_device(); + + Eigen::DSizes shape(pre, n, post); + Eigen::DSizes norm_shape(pre, post); + + auto x_e = framework::EigenVector::Flatten(*in_x); + auto y_e = framework::EigenVector::Flatten(*out_y); + auto norm_e = framework::EigenVector::Flatten(*out_norm); + auto x = x_e.reshape(shape); + auto y = y_e.reshape(shape); + auto norm = norm_e.reshape(norm_shape); + + Eigen::DSizes rdim(1); + // y = x / sqrt((sum(x * x) + epsilon)) + // norm = sqrt(sum(x * x) + epsilon) + auto sum = x.pow(2).sum(rdim) + eps; + norm.device(*place) = sum.sqrt(); + // y = x / norm + Eigen::DSizes rshape(pre, 1, post); + Eigen::DSizes bcast(1, n, 1); + y.device(*place) = x / norm.reshape(rshape).broadcast(bcast); } }; template class NormGradKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - const framework::Tensor* in_x = context.Input("X"); - const framework::Tensor* scale = context.Input("Scale"); - const framework::Tensor* out_grad = - context.Input(framework::GradVarName("Out")); - auto epsilon = static_cast(context.Attr("epsilon")); - framework::Tensor* in_x_grad = - context.Output(framework::GradVarName("X")); - in_x_grad->mutable_data(context.GetPlace()); - int batch_size = in_x->dims()[0]; - int channels = in_x->dims()[1]; - int height = in_x->dims()[2]; - int width = in_x->dims()[3]; - int fea_len = height * width; - auto* place = - context.template device_context().eigen_device(); - - auto scale_eigen = - framework::EigenVector::Flatten( - *scale); - auto x = - framework::EigenMatrix::From( - *in_x, framework::make_ddim({batch_size, fea_len * channels})); - // get square - framework::Tensor x_square; - x_square.mutable_data(in_x->dims(), context.GetPlace()); - auto x_square_eigen = - framework::EigenMatrix::From( - x_square, framework::make_ddim({batch_size, fea_len * channels})); - x_square_eigen.device(*place) = x.square(); - - for (int n = 0; n < batch_size; ++n) { - framework::Tensor in_x_batch = in_x->Slice(n, n + 1); - auto in_x_batch_eigen = - framework::EigenMatrix::From( - in_x_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor in_g_batch = in_x_grad->Slice(n, n + 1); - auto in_g_batch_eigen = - framework::EigenMatrix::From( - in_g_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor x_square_batch = x_square.Slice(n, n + 1); - auto x_square_batch_eigen = - framework::EigenMatrix::From( - x_square_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor outg_batch = out_grad->Slice(n, n + 1); - auto outg_batch_eigen = - framework::EigenMatrix::From( - outg_batch, framework::make_ddim({channels, fea_len})); - - framework::Tensor tmp_tensor; - tmp_tensor.mutable_data(framework::make_ddim({1, fea_len}), - context.GetPlace()); - auto tmp_eigen = - framework::EigenVector::Flatten(tmp_tensor); - auto dim = Eigen::array({{0}}); - tmp_eigen.device(*place) = (in_x_batch_eigen * outg_batch_eigen).sum(dim); - framework::Tensor norm_tmp_tensor; - norm_tmp_tensor.mutable_data(framework::make_ddim({1, fea_len}), - context.GetPlace()); - auto norm_tmp_eigen = - framework::EigenVector::Flatten(norm_tmp_tensor); - norm_tmp_eigen.device(*place) = - (x_square_batch_eigen.sum(dim) + epsilon).sqrt(); - Eigen::array broadcast_dim_col; - broadcast_dim_col[1] = 1; - broadcast_dim_col[0] = channels; - in_g_batch_eigen.device(*place) = - in_x_batch_eigen * tmp_eigen.broadcast(broadcast_dim_col); - in_g_batch_eigen.device(*place) = - in_g_batch_eigen / - (norm_tmp_eigen * norm_tmp_eigen).broadcast(broadcast_dim_col); - in_g_batch_eigen.device(*place) = outg_batch_eigen - in_g_batch_eigen; - // outg_batch_eigen + (in_g_batch_eigen * -1); - in_g_batch_eigen.device(*place) = - in_g_batch_eigen / norm_tmp_eigen.broadcast(broadcast_dim_col); - Eigen::array broadcast_dim_row; - broadcast_dim_row[1] = fea_len; - broadcast_dim_row[0] = 1; - in_g_batch_eigen.device(*place) = - in_g_batch_eigen * (scale_eigen.broadcast(broadcast_dim_row)); - } + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_x = ctx.Input("X"); + auto* in_norm = ctx.Input("Norm"); + auto* in_dy = ctx.Input(framework::GradVarName("Out")); + auto* out_dx = ctx.Output(framework::GradVarName("X")); + out_dx->mutable_data(ctx.GetPlace()); + + auto xdim = in_x->dims(); + int axis = ctx.Attr("axis"); + if (axis < 0) axis = xdim.size() + axis; + int pre, n, post; + GetDims(xdim, axis, &pre, &n, &post); + + auto* place = ctx.template device_context().eigen_device(); + + auto x_e = framework::EigenVector::Flatten(*in_x); + auto dy_e = framework::EigenVector::Flatten(*in_dy); + auto norm_e = framework::EigenVector::Flatten(*in_norm); + auto dx_e = framework::EigenVector::Flatten(*out_dx); + + Eigen::DSizes shape(pre, n, post); + Eigen::DSizes norm_shape(pre, post); + auto x = x_e.reshape(shape); + auto dy = dy_e.reshape(shape); + auto norm = norm_e.reshape(norm_shape); + auto dx = dx_e.reshape(shape); + + framework::Tensor rsum; + rsum.mutable_data({pre, post}, ctx.GetPlace()); + auto sum = framework::EigenTensor::From(rsum); + + Eigen::DSizes rdim(1); + Eigen::DSizes bcast(1, n, 1); + Eigen::DSizes rshape(pre, 1, post); + + // dx = ( dy/sqrt(sum(x*x)) ) * [1 - x*sum(x) / (sum(x*x) + e)] + // = [dy - dy * x * sum(x) / (sum(x*x) + e)] / sqrt(sum(x*x)) + // = [dy - x * sum(x*dy) / (sum(x*x) + e)] / sqrt(sum(x*x)) + // 1. sum = sum(x*dy) + sum.device(*place) = (x * dy).sum(rdim); + // 2. dx = x * sum + dx.device(*place) = sum.reshape(rshape).broadcast(bcast) * x; + // 3. dx / (sum(x*x) + e) + // where, norm.pow(2) = sum(x*x) + e, which is calculated in forward. + dx.device(*place) = dx / norm.pow(2).broadcast(bcast); + // 4. [dy - dx] / sqrt(sum(x*x)) + dx.device(*place) = (dy - dx) / norm.broadcast(bcast); } }; } // namespace operators diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index 1012640d5e2052e4f347ad458cea9072a004f334..eb09470f37eabb5524f774bc289fc68f5884c540 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -18,7 +18,6 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/operators/detail/safe_ref.h" -#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -164,14 +163,11 @@ class ParallelDoOp : public framework::OperatorBase { auto &place = places[place_idx]; auto *cur_scope = sub_scopes[place_idx]; - workers.emplace_back( - framework::Async([program, cur_scope, place, block, place_idx] { - // Give the thread an id to distinguish parallel block with same id. - platform::RecordThread rt(static_cast(place_idx) + 1); - framework::Executor executor(place); - executor.Run(*program, cur_scope, block->ID(), - false /*create_local_scope*/); - })); + workers.emplace_back(framework::Async([program, cur_scope, place, block] { + framework::Executor executor(place); + executor.Run(*program, cur_scope, block->ID(), + false /*create_local_scope*/); + })); } for (auto &worker : workers) { worker.wait(); @@ -242,14 +238,11 @@ class ParallelDoGradOp : public framework::OperatorBase { auto *cur_scope = sub_scopes[i]; // execute - workers.emplace_back( - framework::Async([program, cur_scope, place, block, i] { - // Give the thread an id to distinguish parallel block with same id. - platform::RecordThread rt(static_cast(i) + 1); - framework::Executor executor(place); - executor.Run(*program, cur_scope, block->ID(), - false /*create_local_scope*/); - })); + workers.emplace_back(framework::Async([program, cur_scope, place, block] { + framework::Executor executor(place); + executor.Run(*program, cur_scope, block->ID(), + false /*create_local_scope*/); + })); } for (auto &worker : workers) { worker.wait(); @@ -295,7 +288,7 @@ class ParallelDoGradOp : public framework::OperatorBase { auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}}, - framework::AttributeMap{}); + framework::AttributeMap{{"use_mkldnn", {false}}}); VLOG(10) << sum_op->DebugStringEx(sub_scopes[0]); sum_op->Run(*sub_scopes[0], places[0]); WaitOnPlace(places[0]); diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index d60a99994edc926456706ff6a3ba998a3e5e7dd5..31f083565fddee66aea1485ed71f41b6199f4502 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -81,7 +81,7 @@ class PoolCUDNNOpKernel : public framework::OpKernel { // ------------------- cudnn pool algorithm --------------------- auto handle = ctx.cuda_device_context().cudnn_handle(); ScalingParamType alpha = 1.0f, beta = 0.0f; - PADDLE_ENFORCE(platform::dynload::cudnnPoolingForward( + CUDNN_ENFORCE(platform::dynload::cudnnPoolingForward( handle, cudnn_pool_desc, &alpha, cudnn_input_desc, input_data, &beta, cudnn_output_desc, output_data)); } @@ -135,7 +135,11 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { PoolingMode pooling_mode; if (pooling_type == "max") { - pooling_mode = PoolingMode::kMaximum; + if (FLAGS_cudnn_deterministic) { + pooling_mode = PoolingMode::kMaximumDeterministic; + } else { + pooling_mode = PoolingMode::kMaximum; + } } else { pooling_mode = PoolingMode::kAverage; } @@ -150,7 +154,7 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { T *input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. - PADDLE_ENFORCE(platform::dynload::cudnnPoolingBackward( + CUDNN_ENFORCE(platform::dynload::cudnnPoolingBackward( handle, cudnn_pool_desc, &alpha, cudnn_output_desc, output_data, cudnn_output_desc, output_grad_data, cudnn_input_desc, input_data, &beta, cudnn_input_desc, input_grad_data)); diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/pool_mkldnn_op.cc index 60e936298defe7c6ce8a33bdc7de05b52eb950e7..5341187d1ce9400ac34750ab691608e76158ae0d 100644 --- a/paddle/fluid/operators/pool_mkldnn_op.cc +++ b/paddle/fluid/operators/pool_mkldnn_op.cc @@ -18,16 +18,24 @@ limitations under the License. */ namespace paddle { namespace operators { -using mkldnn::memory; // Note: paddle has also "memory" namespace -using mkldnn::pooling_forward; +using framework::DataLayout; +using mkldnn::memory; using mkldnn::pooling_backward; +using mkldnn::pooling_forward; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; +using platform::to_void_cast; // Generate keys for storing/retriving primitives for this operator // TODO(jczaja): Make hashing function more optimial -static std::string gethash(memory::dims& input_dims, std::string& pooling_type, - std::vector& ksize, std::vector& strides, - std::vector& paddings, std::string suffix) { - auto dims2str = [](memory::dims& operand_dims) { +static std::string gethash(const memory::dims& input_dims, + const std::string& pooling_type, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, + const std::string& suffix) { + auto dims2str = [](const memory::dims& operand_dims) { std::string dstr = ""; for (size_t i = 0; i < operand_dims.size(); ++i) { dstr += std::to_string(operand_dims[i]) + "-"; @@ -52,8 +60,9 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { const Tensor* input = ctx.Input("X"); Tensor* output = ctx.Output("Out"); - // Get an unique name from "argument" name of "Out" variable - // This name will be used as key when saving info into device context + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != memory::format::format_undef, + "Wrong layout/format set for Input tensor"); std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); @@ -79,6 +88,9 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector src_tz = paddle::framework::vectorize2int(input->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + auto input_format = input->format(); + memory::format output_format{memory::format::format_undef}; + const std::string key = gethash(src_tz, pooling_type, ksize, strides, paddings, ctx.op().Output("Out")); const std::string key_pool_p = key + "@pool_p"; @@ -91,16 +103,17 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { auto pool_p = std::static_pointer_cast(dev_ctx.GetBlob(key_pool_p)); if (pool_p == nullptr) { - // TODO(pzelazko-intel): support more formats + auto src_md = platform::MKLDNNMemDesc( + src_tz, platform::MKLDNNGetDataType(), input_format); - auto src_md = - platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); - auto dst_md = - platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); + /* create memory descriptor for pooling without specified format + * ('any') which lets a primitive (pooling in this case) choose + * the memory format preferred for best performance + */ + auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32, + mkldnn::memory::format::any); - std::shared_ptr pool_pd = + std::shared_ptr pool_pd = CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, pooling_type, mkldnn_engine); @@ -113,20 +126,22 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { // save pool_workspace_memory to be referred in backward path dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); - auto pool_src_memory_p = std::make_shared( - memory::primitive_desc{src_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p); + auto src_memory = std::make_shared(pool_pd->src_primitive_desc(), + to_void_cast(input_data)); + auto dst_memory = + std::make_shared(pool_pd->dst_primitive_desc(), output_data); - auto pool_dst_memory_p = std::make_shared( - memory::primitive_desc{dst_md, mkldnn_engine}, - static_cast(output_data)); - dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p); + dev_ctx.SetBlob(key_pool_src_mem_p, src_memory); + dev_ctx.SetBlob(key_pool_dst_mem_p, dst_memory); + + pool_p = std::make_shared(*pool_pd, *(src_memory.get()), + *(dst_memory.get()), + *workspace_memory); - pool_p = std::make_shared( - *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()), - *workspace_memory); dev_ctx.SetBlob(key_pool_p, pool_p); + + output_format = + (memory::format)dst_memory->get_primitive_desc().desc().data.format; } else { // Primitives already exist auto pool_src_memory_p = @@ -137,14 +152,20 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { std::static_pointer_cast(dev_ctx.GetBlob(key_pool_dst_mem_p)); PADDLE_ENFORCE(pool_dst_memory_p != nullptr, "Fail to find pooling dst mem_p in device context"); - pool_src_memory_p->set_data_handle( - reinterpret_cast(const_cast(input_data))); + pool_src_memory_p->set_data_handle(to_void_cast(input_data)); pool_dst_memory_p->set_data_handle(output_data); + + output_format = (memory::format)pool_dst_memory_p->get_primitive_desc() + .desc() + .data.format; } // push primitive to stream and wait until it's executed std::vector pipeline{*(pool_p.get())}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + stream(stream::kind::eager).submit(pipeline).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(output_format); } private: @@ -191,6 +212,13 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); Tensor* in_x_grad = ctx.Output(framework::GradVarName("X")); + PADDLE_ENFORCE(in_x->layout() == DataLayout::kMKLDNN && + in_x->format() != memory::format::format_undef, + "Wrong layout/format set for Input X tensor"); + PADDLE_ENFORCE(out_grad->layout() == DataLayout::kMKLDNN && + out_grad->format() != memory::format::format_undef, + "Wrong layout/format set for Input output_grad tensor"); + std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); @@ -209,6 +237,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const T* out_grad_data = out_grad->data(); T* in_x_grad_data = in_x_grad->mutable_data(ctx.GetPlace()); + memory::format in_x_grad_format{memory::format::format_undef}; std::vector diff_src_tz = paddle::framework::vectorize2int(in_x_grad->dims()); @@ -222,39 +251,48 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const std::string key_pool_bwd_p = key + "@pool_bwd_p"; const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p"; const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p"; + const std::string key_pool_src_mem_p = key + "@pool_src_mem_p"; + const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p"; const std::string key_pool_pd = key + "@pool_pd"; const std::string key_pool_workspace_memory = key + "@pool_workspace_memory"; + auto user_diff_dst_memory = + memory({{{diff_dst_tz}, memory::data_type::f32, out_grad->format()}, + mkldnn_engine}, + to_void_cast(out_grad_data)); + + std::shared_ptr diff_src_memory; + std::shared_ptr diff_dst_memory; + auto dst_memory = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_dst_mem_p)); + PADDLE_ENFORCE(dst_memory != nullptr, + "Fail to find dst_memory in device context"); + + primitive reorder_diff_dst; + bool is_diff_dst_reordered = false; auto pool_bwd_p = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_bwd_p)); if (pool_bwd_p == nullptr) { - auto diff_src_md = - platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); - auto diff_dst_md = - platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); + // Retrieve src_memory/dst_memory saved in forward pass + auto src_memory = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_src_mem_p)); + PADDLE_ENFORCE(src_memory != nullptr, + "Fail to find src_memory in device context"); // Retrieve pool_pd/pool_workspace_memory from device context auto pool_pd = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_pd)); PADDLE_ENFORCE(pool_pd != nullptr, "Fail to find pool_pd in device context"); - - auto workspace_memory = std::static_pointer_cast( + auto workspace_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_workspace_memory)); PADDLE_ENFORCE(workspace_memory != nullptr, "Fail to find workspace_memory in device context"); - auto pool_diff_src_memory_p = std::make_shared(memory( - {diff_src_md, mkldnn_engine}, static_cast(in_x_grad_data))); - dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p); - - auto pool_diff_dst_memory_p = std::make_shared( - memory({diff_dst_md, mkldnn_engine}, - static_cast(const_cast(out_grad_data)))); - dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p); + // create memory descriptors for pooling + auto diff_src_md = src_memory.get()->get_primitive_desc().desc(); + auto diff_dst_md = dst_memory.get()->get_primitive_desc().desc(); auto pool_bwd_desc = mkldnn::pooling_backward::desc( pooling_type == "max" ? mkldnn::algorithm::pooling_max @@ -264,35 +302,74 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( pool_bwd_desc, mkldnn_engine, *pool_pd); + // reorder between user_diff_dst and pool diff_dst if needed + diff_dst_memory = std::make_shared(user_diff_dst_memory); + if (memory::primitive_desc(dst_memory->get_primitive_desc()) != + user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = + std::make_shared(dst_memory.get()->get_primitive_desc()); + reorder_diff_dst = reorder(user_diff_dst_memory, *diff_dst_memory); + is_diff_dst_reordered = true; + } + + diff_src_memory = std::make_shared( + pool_bwd_pd.diff_src_primitive_desc(), in_x_grad_data); + + dev_ctx.SetBlob(key_pool_diff_src_mem_p, diff_src_memory); + dev_ctx.SetBlob(key_pool_diff_dst_mem_p, diff_dst_memory); + pool_bwd_p = std::make_shared( - pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory, - *(pool_diff_src_memory_p)); + pool_bwd_pd, *(diff_dst_memory.get()), *workspace_memory, + *(diff_src_memory)); dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p); + } else { // Primitives already exist - auto pool_diff_src_memory_p = std::static_pointer_cast( + diff_src_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_diff_src_mem_p)); - PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr, + PADDLE_ENFORCE(diff_src_memory != nullptr, "Fail to find pooling src mem_p in device context"); - auto pool_diff_dst_memory_p = std::static_pointer_cast( + diff_dst_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_diff_dst_mem_p)); - PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr, + PADDLE_ENFORCE(diff_dst_memory != nullptr, "Fail to find pooling dst mem_p in device context"); - pool_diff_src_memory_p->set_data_handle( - reinterpret_cast(in_x_grad_data)); - pool_diff_dst_memory_p->set_data_handle(const_cast(out_grad_data)); + + diff_src_memory->set_data_handle(reinterpret_cast(in_x_grad_data)); + diff_dst_memory->set_data_handle(const_cast(out_grad_data)); + + // reorder between user_diff_dst and pool diff_dst if needed + if (memory::primitive_desc(dst_memory->get_primitive_desc()) != + user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = + std::make_shared(dst_memory.get()->get_primitive_desc()); + reorder_diff_dst = reorder(user_diff_dst_memory, *diff_dst_memory); + is_diff_dst_reordered = true; + } } + in_x_grad_format = (memory::format)diff_src_memory->get_primitive_desc() + .desc() + .data.format; + // push primitive to stream and wait until it's executed - std::vector pipeline{*(pool_bwd_p.get())}; + std::vector pipeline; + if (is_diff_dst_reordered) { + pipeline.push_back(reorder_diff_dst); + } + pipeline.push_back(*(pool_bwd_p.get())); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + + in_x_grad->set_layout(DataLayout::kMKLDNN); + in_x_grad->set_format(in_x_grad_format); } // Compute() }; } // namespace operators } // namespace paddle +namespace ops = paddle::operators; + REGISTER_OP_KERNEL(pool2d, MKLDNN, ::paddle::platform::CPUPlace, - paddle::operators::PoolMKLDNNOpKernel); + ops::PoolMKLDNNOpKernel); REGISTER_OP_KERNEL(pool2d_grad, MKLDNN, ::paddle::platform::CPUPlace, - paddle::operators::PoolMKLDNNGradOpKernel); + ops::PoolMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index f4fb2b132fe8d59cb50f5a1f7359240ac50445fe..f8ad63690e84339da0390d4ddd2db45f25db385a 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -83,6 +83,9 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const { framework::OpKernelType PoolOp::GetExpectedKernelType( const framework::ExecutionContext &ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -92,11 +95,10 @@ framework::OpKernelType PoolOp::GetExpectedKernelType( if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), layout_, library_); @@ -112,6 +114,9 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { framework::OpKernelType PoolOpGrad::GetExpectedKernelType( const framework::ExecutionContext &ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -121,6 +126,7 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType( if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif @@ -129,8 +135,6 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType( PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN, "float16 can only be used when CUDNN is used"); } - std::string data_format = ctx.Attr("data_format"); - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, library_); } @@ -147,7 +151,8 @@ void Pool2dOpMaker::Make() { "The format of output tensor is also NCHW, " "where N is batch size, C is the number of channels, " "H is the height of the feature, " - "and W is the width of the feature."); + "and W is the width of the feature.") + .Reuse("X"); AddAttr("pooling_type", "(string), pooling type, can be \"max\" for max-pooling " @@ -199,8 +204,6 @@ void Pool2dOpMaker::Make() { // TODO(dzhwinter): need to registered layout transform function AddComment(R"DOC( -Pool2d Operator. - The pooling2d operation calculates the output based on the input, pooling_type and ksize, strides, paddings parameters. Input(X) and output(Out) are in NCHW format, where N is batch size, C is the @@ -210,19 +213,28 @@ These two elements represent height and width, respectively. The input(X) size and output(Out) size may be different. Example: + Input: + X shape: $(N, C, H_{in}, W_{in})$ + Output: + Out shape: $(N, C, H_{out}, W_{out})$ + For ceil_mode = false: $$ - H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\ - W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 + H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 + $$ + $$ + W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 $$ For ceil_mode = true: $$ - H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0] + strides[0] - 1)}{strides[0]} + 1 \\ - W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1 + H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0] + strides[0] - 1)}{strides[0]} + 1 + $$ + $$ + W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1 $$ )DOC"); @@ -240,7 +252,8 @@ void Pool3dOpMaker::Make() { "The format of output tensor is also NCDHW, " "where N is batch size, C is " "the number of channels, and D, H and W is the depth, height and " - "width of the feature, respectively."); + "width of the feature, respectively.") + .Reuse("X"); AddAttr("pooling_type", "(string) Pooling type, can be \"max\" for max-pooling " diff --git a/paddle/fluid/operators/positive_negative_pair_op.h b/paddle/fluid/operators/positive_negative_pair_op.h index f20f33bbeb19766d6974ea17b155cac363c01fb2..db0a1002f47944c5d926fb5a51b84536dcf446b8 100644 --- a/paddle/fluid/operators/positive_negative_pair_op.h +++ b/paddle/fluid/operators/positive_negative_pair_op.h @@ -14,7 +14,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/utils/Logging.h" +#include "paddle/legacy/utils/Logging.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index e0a9b24ac8978418a1a4ece62286e022bec8b834..4b804740a06f9e29704f2b3f58a90191e3559347 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -18,7 +18,7 @@ limitations under the License. */ #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/operators/send_recv_util.h" namespace paddle { @@ -41,19 +41,19 @@ class PrefetchOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - auto rpc_client = detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { VLOG(3) << "sending " << ins[i] << " to " << epmap[i] << " to get " << outs[i] << " back"; - rpc_client->AsyncPrefetchVariable(epmap[i], ctx, scope, ins[i], - outs[i]); + rpc_client->AsyncPrefetchVar(epmap[i], ctx, scope, ins[i], outs[i]); } else { VLOG(3) << "don't send no-initialied variable: " << ins[i]; } } - PADDLE_ENFORCE(rpc_client->Wait()); + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } }; diff --git a/paddle/fluid/operators/print_op.cc b/paddle/fluid/operators/print_op.cc index db7634918a5179a61304315ecd08350d23fb4642..cceac402951ae6bf3fe0b4c96af5b7ce9ca1ba0e 100644 --- a/paddle/fluid/operators/print_op.cc +++ b/paddle/fluid/operators/print_op.cc @@ -16,6 +16,7 @@ #include #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/framework/variable.h" namespace paddle { @@ -62,7 +63,7 @@ struct Formater { } } void PrintDtype() { - if (dtype.hash_code() != typeid(const char).hash_code()) { + if (!framework::IsType(dtype)) { CLOG << "\tdtype: " << dtype.name() << std::endl; } } @@ -83,15 +84,15 @@ struct Formater { void PrintData(size_t size) { PADDLE_ENFORCE_NOT_NULL(data); // print float - if (dtype.hash_code() == typeid(const float).hash_code()) { + if (framework::IsType(dtype)) { Display(size); - } else if (dtype.hash_code() == typeid(const double).hash_code()) { + } else if (framework::IsType(dtype)) { Display(size); - } else if (dtype.hash_code() == typeid(const int).hash_code()) { + } else if (framework::IsType(dtype)) { Display(size); - } else if (dtype.hash_code() == typeid(const int64_t).hash_code()) { + } else if (framework::IsType(dtype)) { Display(size); - } else if (dtype.hash_code() == typeid(const bool).hash_code()) { + } else if (framework::IsType(dtype)) { Display(size); } else { CLOG << "\tdata: unprintable type: " << dtype.name() << std::endl; diff --git a/paddle/fluid/operators/random_crop_op.cc b/paddle/fluid/operators/random_crop_op.cc index b14b559e31dd422f8ebe4002988a9746dfdf28a2..123fa44fa3ddbc9343b9629be63fdefdf12b4646 100644 --- a/paddle/fluid/operators/random_crop_op.cc +++ b/paddle/fluid/operators/random_crop_op.cc @@ -20,7 +20,6 @@ class RandomCropOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( @@ -36,11 +35,16 @@ class RandomCropOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Seed", "The random seed."); AddOutput("Out", "The cropped instance batch."); AddOutput("SeedOut", "The random seed after random cropping.") - .AsDispensable(); + .AsIntermediate(); AddAttr>("shape", "The shape of a cropped instance."); + AddAttr("startup_seed", + "If the input 'Seed' is not initialized, the 'startup_seed' " + "will be used to replace it. Even so, the seed after random " + "crop will also be outputed to the 'SeedOut'.") + .SetDefault(0); AddComment(R"DOC( - This operator takes a batch of instance, and do random cropping on each instance. - It means that cropping positions differs on each instance, which is determined + This operator takes a batch of instance, and do random cropping on each instance. + It means that cropping positions differs on each instance, which is determined by an uniform random generator. All cropped instances have the same shape, which is determined by the operator's attribute 'shape'. )DOC"); @@ -50,8 +54,6 @@ class RandomCropOpMaker : public framework::OpProtoAndCheckerMaker { class RandomCropOpInferShape : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext* ctx) const override { - auto seed_dim = ctx->GetInputDim("Seed"); - PADDLE_ENFORCE(seed_dim.size() == 1 && seed_dim[0] == 1); auto shape = ctx->Attrs().Get>("shape"); auto x_dim = ctx->GetInputDim("X"); PADDLE_ENFORCE_GT(x_dim.size(), static_cast(shape.size())); @@ -63,7 +65,6 @@ class RandomCropOpInferShape : public framework::InferShapeBase { out_dim[x_i] = shape[shape_i]; } ctx->SetOutputDim("Out", framework::make_ddim(out_dim)); - ctx->SetOutputDim("SeedOut", framework::make_ddim({1})); } }; diff --git a/paddle/fluid/operators/random_crop_op.h b/paddle/fluid/operators/random_crop_op.h index f3261cbdc986b0cc724315c1eb92b8b84e18c742..d68ba9d661698bb0d33b139f5748daec2ead6595 100644 --- a/paddle/fluid/operators/random_crop_op.h +++ b/paddle/fluid/operators/random_crop_op.h @@ -142,16 +142,22 @@ template class RandomCropKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& ctx) const { - auto& seed_tensor = detail::Ref(ctx.Input("Seed")); int64_t seed = 0; - if (platform::is_cpu_place(seed_tensor.place())) { - seed = *seed_tensor.data(); + auto& seed_tensor = detail::Ref(ctx.Input("Seed")); + if (seed_tensor.IsInitialized()) { + if (platform::is_cpu_place(seed_tensor.place())) { + seed = *seed_tensor.data(); + } else { + LOG(WARNING) << "It is slow to place seed in GPU memory. Please verify " + "your program"; + framework::LoDTensor cpu_seed; + framework::TensorCopySync(seed_tensor, platform::CPUPlace(), &cpu_seed); + seed = *cpu_seed.data(); + } } else { - LOG(WARNING) << "It is slow to place seed in GPU memory. Please verify " - "your program"; - framework::LoDTensor cpu_seed; - framework::TensorCopySync(seed_tensor, platform::CPUPlace(), &cpu_seed); - seed = *cpu_seed.data(); + VLOG(5) << "WARNING: The input 'Seed' is not initialized, use attribute " + "'startup_seed' instead."; + seed = ctx.Attr("startup_seed"); } auto shape = ctx.Attr>("shape"); auto& x = detail::Ref(ctx.Input("X")); @@ -171,7 +177,7 @@ class RandomCropKernel : public framework::OpKernel { engine.discard(functor.prod_batchsize_dims_ * (functor.rank_ - functor.num_batchsize_dims_)); *ctx.Output("SeedOut")->mutable_data( - platform::CPUPlace()) = engine(); + framework::make_ddim({1}), platform::CPUPlace()) = engine(); } }; diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index 72a27d43584d55cd0859c63577ae85ff0f5fdfa8..a0d640b2020958af53a4405ae886eadb2a1e117e 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -65,10 +66,26 @@ class ReadOp : public framework::OperatorBase { .GetMutable(); std::vector out_arg_names = Outputs("Out"); std::vector ins; + + // For profiling + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(dev_place); + platform::RecordEvent record_event(Type(), &ctx); + reader->ReadNext(&ins); - PADDLE_ENFORCE(!ins.empty(), "There is no next data."); + if (ins.empty()) { + if (Attr("throw_eof_exp")) { + PADDLE_THROW_EOF(); + } else { + ins.resize(out_arg_names.size()); + for (auto& tensor : ins) { + // data type is not important for subsequent DataBalanceOpHandle + tensor.mutable_data(framework::make_ddim({0}), dev_place); + } + } + } PADDLE_ENFORCE_EQ(ins.size(), out_arg_names.size()); - for (size_t i = 0; i < ins.size(); ++i) { + for (size_t i = 0; i < out_arg_names.size(); ++i) { auto* out = scope.FindVar(out_arg_names[i])->GetMutable(); out->ShareDataWith(ins[i]); @@ -82,6 +99,14 @@ class ReadOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("Reader", "(ReaderHolder) The executed reader."); AddOutput("Out", "(LoDTensor) The output data.").AsDuplicable(); + AddAttr( + "throw_eof_exp", + "If set true, an exception will be thrown when the Reader " + "yields empty (which means there is no next data).\n" + "NOTES: This flag must be true always. It will be set to false" + " only when the data-balance is enabled in ParallelExecutor" + " and it is set by ParallelExecutor instance, not users.") + .SetDefault(true); AddComment(R"DOC( Read Operator diff --git a/paddle/fluid/operators/reader/CMakeLists.txt b/paddle/fluid/operators/reader/CMakeLists.txt index 62532036f86bfb82465ccd9e0ec526299489932a..728197377df04df8c993a48bc282431473fe9959 100644 --- a/paddle/fluid/operators/reader/CMakeLists.txt +++ b/paddle/fluid/operators/reader/CMakeLists.txt @@ -15,15 +15,16 @@ function(reader_library TARGET_NAME) PARENT_SCOPE) endfunction() -reader_library(open_files_op SRCS open_files_op.cc) +cc_library(buffered_reader SRCS buffered_reader.cc DEPS reader simple_threadpool) +reader_library(open_files_op SRCS open_files_op.cc DEPS buffered_reader) reader_library(create_random_data_generator_op SRCS create_random_data_generator_op.cc) reader_library(create_shuffle_reader_op SRCS create_shuffle_reader_op.cc) reader_library(create_batch_reader_op SRCS create_batch_reader_op.cc) reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_op.cc) -reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc) +reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc DEPS buffered_reader) reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc) -reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc) reader_library(create_custom_reader_op SRCS create_custom_reader_op.cc) +reader_library(create_py_reader_op SRCS create_py_reader_op.cc) cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc) # Export local libraries to parent diff --git a/paddle/fluid/operators/reader/blocking_queue.h b/paddle/fluid/operators/reader/blocking_queue.h index 71684b14176edc8f71efbefa9a7decffc8f3011e..28cc91a5ed5d74994e5b960a0a4dd3c6a5e6cdcc 100644 --- a/paddle/fluid/operators/reader/blocking_queue.h +++ b/paddle/fluid/operators/reader/blocking_queue.h @@ -81,6 +81,15 @@ class BlockingQueue { } } + void ReOpen() { + std::lock_guard lock(mutex_); + closed_ = false; + std::deque new_deque; + queue_.swap(new_deque); + send_cv_.notify_all(); + receive_cv_.notify_all(); + } + void Close() { std::lock_guard lock(mutex_); closed_ = true; @@ -88,24 +97,29 @@ class BlockingQueue { receive_cv_.notify_all(); } - bool IsClosed() { + bool IsClosed() const { std::lock_guard lock(mutex_); return closed_; } - size_t Cap() { + size_t Cap() const { std::lock_guard lock(mutex_); return capacity_; } + size_t Size() const { + std::lock_guard lock(mutex_); + return queue_.size(); + } + private: size_t capacity_; bool closed_; std::deque queue_; - std::mutex mutex_; - std::condition_variable receive_cv_; - std::condition_variable send_cv_; + mutable std::mutex mutex_; + mutable std::condition_variable receive_cv_; + mutable std::condition_variable send_cv_; }; } // namespace reader } // namespace operators diff --git a/paddle/fluid/operators/reader/buffered_reader.cc b/paddle/fluid/operators/reader/buffered_reader.cc new file mode 100644 index 0000000000000000000000000000000000000000..26ff221dfa0768bd2bcc9e6485a32485f0212ac6 --- /dev/null +++ b/paddle/fluid/operators/reader/buffered_reader.cc @@ -0,0 +1,108 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reader/buffered_reader.h" +#include + +namespace paddle { +namespace operators { +namespace reader { +BufferedReader::~BufferedReader() { + reader_->Shutdown(); + while (!position_.empty()) { + position_.front().wait(); + position_.pop(); + } +} + +BufferedReader::BufferedReader( + const std::shared_ptr &reader, + const platform::Place &place, size_t buffer_size) + : framework::DecoratedReader(reader), + thread_pool_(1), + place_(place), + buffer_size_(buffer_size) { + cpu_buffer_.resize(buffer_size); + gpu_buffer_.resize(buffer_size); + ReadTillBufferFullAsync(); +} + +void BufferedReader::ReadTillBufferFullAsync() { + PADDLE_ENFORCE_EQ(position_.size(), 0U); + for (size_t i = 0; i < buffer_size_; ++i) { + ReadAsync(i); + } +} + +void BufferedReader::ReadAsync(size_t i) { + position_.emplace(thread_pool_.enqueue([this, i]() -> size_t { + TensorVec &cpu = cpu_buffer_[i]; + reader_->ReadNext(&cpu); + + if (cpu.empty()) { + return -1UL; + } + + if (platform::is_gpu_place(place_)) { + TensorVec &gpu = gpu_buffer_[i]; + gpu.resize(cpu.size()); + for (size_t i = 0; i < cpu.size(); ++i) { + framework::TensorCopySync(cpu[i], place_, &gpu[i]); + gpu[i].set_lod(cpu[i].lod()); + } + } + return i; + })); +} + +void BufferedReader::ShutdownImpl() { + reader_->Shutdown(); + while (!position_.empty()) { + position_.pop(); + } + prev_pos_ = -1UL; +} + +void BufferedReader::StartImpl() { + reader_->Start(); + ReadTillBufferFullAsync(); +} + +void BufferedReader::ReadNextImpl(std::vector *out) { + if (position_.empty()) { + out->clear(); + return; + } + size_t i = position_.front().get(); + position_.pop(); + + if (i == -1UL) { + ReadNextImpl(out); + return; + } + + *out = platform::is_gpu_place(place_) ? gpu_buffer_[i] : cpu_buffer_[i]; + + // Do not push current position into ReadAsync. Push the previous position + // Since all computation in fluid are async, change the data of + // current position may cause data error. + if (prev_pos_ != -1Ul) { + ReadAsync(prev_pos_); + } + prev_pos_ = i; +} + +} // namespace reader +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reader/buffered_reader.h b/paddle/fluid/operators/reader/buffered_reader.h new file mode 100644 index 0000000000000000000000000000000000000000..cbe2bc1b5fdd69d1a843b768e3289acd621369a6 --- /dev/null +++ b/paddle/fluid/operators/reader/buffered_reader.h @@ -0,0 +1,66 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include "ThreadPool.h" +#include "paddle/fluid/framework/reader.h" + +namespace paddle { +namespace operators { +namespace reader { + +class BufferedReader : public framework::DecoratedReader { + using TensorVec = std::vector; + using VecFuture = std::future; + + public: + BufferedReader(const std::shared_ptr& reader, + const platform::Place& place, size_t buffer_size); + + ~BufferedReader() override; + + private: + void ReadTillBufferFullAsync(); + + void ReadAsync(size_t i); + + protected: + void ShutdownImpl() override; + void StartImpl() override; + void ReadNextImpl(std::vector* out) override; + + private: + ThreadPool thread_pool_; + platform::Place place_; + const size_t buffer_size_; + + std::queue> position_; + + // The buffer for reading data. + // NOTE: the simplest way to implement buffered reader is do not use any + // buffer, just read async and create futures as buffer size. However, to + // malloc tensors every time is extremely slow. Here we store all data in + // buffers and prevent alloc every time. + std::vector cpu_buffer_; + std::vector gpu_buffer_; + size_t prev_pos_{-1UL}; +}; + +} // namespace reader +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reader/create_batch_reader_op.cc b/paddle/fluid/operators/reader/create_batch_reader_op.cc index 4cc7cbc6e89b0712faf9ad9c51480bce00da15f5..e17c2ffd39eea31fe85933eda144ab97cf8c3dd8 100644 --- a/paddle/fluid/operators/reader/create_batch_reader_op.cc +++ b/paddle/fluid/operators/reader/create_batch_reader_op.cc @@ -20,15 +20,19 @@ namespace reader { class BatchReader : public framework::DecoratedReader { public: - BatchReader(ReaderBase* reader, int batch_size) - : DecoratedReader(reader), batch_size_(batch_size) { + BatchReader(const std::shared_ptr& reader, int batch_size, + bool discard_leftover) + : DecoratedReader(reader), + batch_size_(static_cast(batch_size)), + discard_leftover_(discard_leftover) { buffer_.reserve(batch_size_); } - void ReadNext(std::vector* out) override; + void ReadNextImpl(std::vector* out) override; private: - int batch_size_; + size_t batch_size_; + bool discard_leftover_; std::vector> buffer_; }; @@ -46,8 +50,9 @@ class CreateBatchReaderOp : public framework::OperatorBase { } const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) ->Get(); - out->Reset( - new BatchReader(underlying_reader.Get(), Attr("batch_size"))); + out->Reset(framework::MakeDecoratedReader( + underlying_reader, Attr("batch_size"), + Attr("discard_leftover"))); } }; @@ -57,6 +62,10 @@ class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase { AddAttr("batch_size", "How many instances the batch reader yields each time.") .GreaterThan(0); + AddAttr("discard_leftover", + "If true, the leftover instances that are not enough for a " + "new batch will be discarded.") + .SetDefault(true); AddComment(R"DOC( CreateBatchReader Operator @@ -66,10 +75,10 @@ class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase { } }; -void BatchReader::ReadNext(std::vector* out) { +void BatchReader::ReadNextImpl(std::vector* out) { buffer_.clear(); buffer_.reserve(batch_size_); - for (int i = 0; i < batch_size_; ++i) { + for (size_t i = 0; i < batch_size_; ++i) { buffer_.push_back(std::vector()); reader_->ReadNext(&buffer_.back()); if (buffer_.back().empty()) { @@ -77,15 +86,18 @@ void BatchReader::ReadNext(std::vector* out) { break; } } + if (discard_leftover_ && buffer_.size() < batch_size_) { + buffer_.clear(); + } // Concat instances out->clear(); if (buffer_.empty()) { // if buffer_ is empty, the 'out' will return as an empty vector. return; } - int out_num = buffer_[0].size(); + size_t out_num = buffer_[0].size(); out->reserve(out_num); - for (int j = 0; j < out_num; ++j) { + for (size_t j = 0; j < out_num; ++j) { // Merge shape and check date type std::type_index batch_type = buffer_[0][j].type(); framework::DDim batch_shape = buffer_[0][j].dims(); diff --git a/paddle/fluid/operators/reader/create_custom_reader_op.cc b/paddle/fluid/operators/reader/create_custom_reader_op.cc index 331224a59899b4a7d517ca4f7141fb5b8f4f5168..85394b336fc967fc6973131fbedda4c796825185 100644 --- a/paddle/fluid/operators/reader/create_custom_reader_op.cc +++ b/paddle/fluid/operators/reader/create_custom_reader_op.cc @@ -22,7 +22,8 @@ namespace reader { class CustomReader : public framework::DecoratedReader { public: - CustomReader(ReaderBase* reader, const framework::BlockDesc& sub_block, + CustomReader(const std::shared_ptr& reader, + const framework::BlockDesc& sub_block, const std::vector& source_var_names, const std::vector& sink_var_names) : DecoratedReader(reader), @@ -32,12 +33,13 @@ class CustomReader : public framework::DecoratedReader { source_var_names_(source_var_names), sink_var_names_(sink_var_names) {} - void ReadNext(std::vector* out) override; + void ReadNextImpl(std::vector* out) override; private: const framework::ProgramDesc program_; int sub_block_id_; framework::Executor exe_; + framework::Scope scope_; std::vector source_var_names_; std::vector sink_var_names_; @@ -58,10 +60,10 @@ class CreateCustomReaderOp : public framework::OperatorBase { } const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) ->Get(); - out->Reset( - new CustomReader(underlying_reader.Get(), *sub_block, - Attr>("source_var_names"), - Attr>("sink_var_names"))); + out->Reset(framework::MakeDecoratedReader( + underlying_reader, *sub_block, + Attr>("source_var_names"), + Attr>("sink_var_names"))); } }; @@ -141,7 +143,7 @@ class CustomReaderInferVarType : public framework::VarTypeInference { } }; -void CustomReader::ReadNext(std::vector* out) { +void CustomReader::ReadNextImpl(std::vector* out) { out->clear(); std::vector underlying_outs; reader_->ReadNext(&underlying_outs); @@ -157,23 +159,24 @@ void CustomReader::ReadNext(std::vector* out) { // The scope for CustomReader's sub-block should be independent and shouldn't // be any other computation scope's child. Otherwise, data preprocessing and // compution cannot be concurrent. - framework::Scope scope; + framework::Scope* exe_scope = &scope_.NewScope(); // 1. Copy LoDTensors from underlying reader's output to source variables. for (size_t i = 0; i < source_var_names_.size(); ++i) { - framework::Variable* var = scope.Var(source_var_names_[i]); + framework::Variable* var = exe_scope->Var(source_var_names_[i]); framework::LoDTensor* tensor = var->GetMutable(); tensor->ShareDataWith(underlying_outs[i]); tensor->set_lod(underlying_outs[i].lod()); } // 2. Run the sub-block. - exe_.Run(program_, &scope, sub_block_id_, false, true); + exe_.Run(program_, exe_scope, sub_block_id_, false, true); // 3. Copy LoDTensors from sink variables to out. out->resize(sink_var_names_.size()); for (size_t i = 0; i < sink_var_names_.size(); ++i) { - const auto& tensor = detail::Ref(scope.FindVar(sink_var_names_[i])) + const auto& tensor = detail::Ref(exe_scope->FindVar(sink_var_names_[i])) .Get(); framework::TensorCopySync(tensor, platform::CPUPlace(), &(*out)[i]); } + scope_.DeleteScope(exe_scope); } } // namespace reader diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc index bc830a2b72e657f79f4c94e24428d38ff2b7c42e..ed719f91d0980480aa62a5cd3c1f819e6c0e7475 100644 --- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -12,73 +12,12 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include // NOLINT - -#include "paddle/fluid/operators/reader/blocking_queue.h" +#include "paddle/fluid/operators/reader/buffered_reader.h" #include "paddle/fluid/operators/reader/reader_op_registry.h" namespace paddle { namespace operators { namespace reader { - -// 'Double buffer' means we shall maintain two batches of input data at the same -// time. So the kCacheSize shoul be at least 2. -static constexpr size_t kCacheSize = 3; -// There will be two bacthes out of the channel during training: -// 1. the one waiting to be sent to the channel -// 2. the one just be received from the channel, which is also being used by -// subsequent operators. -// So the channel size should be kChacheSize - 2 -static constexpr size_t kChannelSize = 1; // kCacheSize - 2 - -class DoubleBufferReader : public framework::DecoratedReader { - public: - explicit DoubleBufferReader( - ReaderBase* reader, platform::Place target_place = platform::CPUPlace()) - : DecoratedReader(reader), place_(target_place) { - cpu_tensor_cache_.resize(kCacheSize); - gpu_tensor_cache_.resize(kCacheSize); -#ifdef PADDLE_WITH_CUDA - if (platform::is_gpu_place(place_)) { - for (size_t i = 0; i < kCacheSize; ++i) { - ctxs_.emplace_back(new platform::CUDADeviceContext( - boost::get(place_))); - } - } -#endif - StartPrefetcher(); - } - - void ReadNext(std::vector* out) override; - void ReInit() override; - - ~DoubleBufferReader() { EndPrefetcher(); } - - private: - void StartPrefetcher() { - channel_ = new reader::BlockingQueue(kChannelSize); - prefetcher_ = std::thread([this] { PrefetchThreadFunc(); }); - } - - void EndPrefetcher() { - channel_->Close(); - if (prefetcher_.joinable()) { - prefetcher_.join(); - } - delete channel_; - channel_ = nullptr; - } - - void PrefetchThreadFunc(); - - std::thread prefetcher_; - reader::BlockingQueue* channel_; - platform::Place place_; - std::vector> cpu_tensor_cache_; - std::vector> gpu_tensor_cache_; - std::vector> ctxs_; -}; - class CreateDoubleBufferReaderOp : public framework::OperatorBase { public: using framework::OperatorBase::OperatorBase; @@ -108,7 +47,8 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase { place = platform::CUDAPlace(static_cast(num)); } - out->Reset(new DoubleBufferReader(underlying_reader.Get(), place)); + out->Reset(framework::MakeDecoratedReader(underlying_reader, + place, 2)); } }; @@ -135,57 +75,6 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase { } }; -void DoubleBufferReader::ReadNext(std::vector* out) { - size_t cached_tensor_id; - if (channel_->Receive(&cached_tensor_id)) { - if (platform::is_gpu_place(place_)) { - *out = gpu_tensor_cache_[cached_tensor_id]; - } else { - // CPU place - *out = cpu_tensor_cache_[cached_tensor_id]; - } - } else { - out->clear(); - } -} - -void DoubleBufferReader::ReInit() { - reader_->ReInit(); - EndPrefetcher(); - StartPrefetcher(); -} - -void DoubleBufferReader::PrefetchThreadFunc() { - VLOG(5) << "A new prefetch thread starts."; - size_t cached_tensor_id = 0; - while (true) { - auto& cpu_batch = cpu_tensor_cache_[cached_tensor_id]; - reader_->ReadNext(&cpu_batch); - if (cpu_batch.empty()) { - // The underlying reader have no next data. - break; - } - if (platform::is_gpu_place(place_)) { - auto& gpu_batch = gpu_tensor_cache_[cached_tensor_id]; - gpu_batch.resize(cpu_batch.size()); - for (size_t i = 0; i < cpu_batch.size(); ++i) { - // TODO(fengjiayi): Use asynchronous TensorCopy instead - framework::TensorCopySync(cpu_batch[i], place_, &gpu_batch[i]); - gpu_batch[i].set_lod(cpu_batch[i].lod()); - } - } - if (!channel_->Send(cached_tensor_id)) { - VLOG(5) << "WARNING: The double buffer channel has been closed. The " - "prefetch thread will terminate."; - break; - } - ++cached_tensor_id; - cached_tensor_id %= kCacheSize; - } - channel_->Close(); - VLOG(5) << "Prefetch thread terminates."; -} - } // namespace reader } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc index 249b0b7c6dbc8b8104bce95562e6e9b2a28c77f8..0a225597d34f43c7fb82aeae2552cdf16c8ba566 100644 --- a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc +++ b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc @@ -21,26 +21,25 @@ namespace reader { class MultiPassReader : public framework::DecoratedReader { public: - MultiPassReader(ReaderBase* reader, int pass_num) + MultiPassReader(const std::shared_ptr& reader, int pass_num) : DecoratedReader(reader), pass_num_(pass_num), pass_count_(0) {} - void ReadNext(std::vector* out) override { + void ReadNextImpl(std::vector* out) override { reader_->ReadNext(out); - if (out->empty()) { + if (out->empty() && pass_count_ < pass_num_ - 1) { + reader_->Shutdown(); + reader_->Start(); + reader_->ReadNext(out); ++pass_count_; - if (pass_count_ < pass_num_) { - reader_->ReInit(); - reader_->ReadNext(out); - } } } - void ReInit() override { + private: + void StartImpl() override { pass_count_ = 0; - reader_->ReInit(); + reader_->Start(); } - private: int pass_num_; mutable int pass_count_; }; @@ -60,7 +59,8 @@ class CreateMultiPassReaderOp : public framework::OperatorBase { const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) ->Get(); int pass_num = Attr("pass_num"); - out->Reset(new MultiPassReader(underlying_reader.Get(), pass_num)); + out->Reset(framework::MakeDecoratedReader( + underlying_reader, pass_num)); } }; diff --git a/paddle/fluid/operators/reader/create_py_reader_op.cc b/paddle/fluid/operators/reader/create_py_reader_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0f31ca1a94326956ae5e6dffd582daedeb55a9e3 --- /dev/null +++ b/paddle/fluid/operators/reader/create_py_reader_op.cc @@ -0,0 +1,89 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h" +#include "paddle/fluid/operators/reader/reader_op_registry.h" + +namespace paddle { +namespace operators { +namespace reader { + +class PyReader : public framework::FileReader { + public: + explicit PyReader(const std::shared_ptr& queue) + : framework::FileReader() { + PADDLE_ENFORCE(queue != nullptr, "LoDTensorBlockingQueue must not be null"); + queue_ = queue; + } + + void ReadNext(std::vector* out) override { + bool success; + *out = queue_->Pop(&success); + if (!success) out->clear(); + } + + ~PyReader() { queue_->Close(); } + + void Shutdown() override { queue_->Close(); } + + void Start() override { queue_->ReOpen(); } + + private: + std::shared_ptr queue_; +}; + +class CreatePyReaderOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + auto* out = scope.FindVar(Output("Out")) + ->template GetMutable(); + if (out->Get() != nullptr) return; + + const std::string& queue_name = Input("blocking_queue"); + auto* queue_holder_var = scope.FindVar(queue_name); + PADDLE_ENFORCE_NOT_NULL( + queue_holder_var, + "No LoDTensorBlockingQueueHolder variable with name %s found", + queue_name); + auto* queue_holder = + queue_holder_var->template GetMutable(); + + out->Reset(std::make_shared(queue_holder->GetQueue())); + } +}; + +class CreatePyReaderOpMaker : public FileReaderMakerBase { + protected: + void Apply() override { + AddInput("blocking_queue", + "Name of the `LoDTensorBlockingQueueHolder` variable"); + + AddComment(R"DOC( + Create PyReader to support LoDTensor data feeding in Python side. + )DOC"); + } +}; + +} // namespace reader +} // namespace operators +} // namespace paddle + +namespace reader = ::paddle::operators::reader; + +REGISTER_FILE_READER_OPERATOR(create_py_reader, reader::CreatePyReaderOp, + reader::CreatePyReaderOpMaker); diff --git a/paddle/fluid/operators/reader/create_random_data_generator_op.cc b/paddle/fluid/operators/reader/create_random_data_generator_op.cc index 5b7e8a063a034f0be056065826fca0fe807bc9a7..e5c116dfcd71ef40597ca19d1da0b51038baaad1 100644 --- a/paddle/fluid/operators/reader/create_random_data_generator_op.cc +++ b/paddle/fluid/operators/reader/create_random_data_generator_op.cc @@ -19,11 +19,11 @@ namespace operators { namespace reader { template -class RandomDataGenerator : public framework::ReaderBase { +class RandomDataGenerator : public framework::FileReader { public: RandomDataGenerator(const std::vector& shapes, float low, float high) - : framework::ReaderBase(), low_(low), high_(high), shapes_(shapes) { + : framework::FileReader(), low_(low), high_(high), shapes_(shapes) { PADDLE_ENFORCE_LE(low, high, "'low' shouldn't be greater than 'high'.(%f vs %f)", low, high); @@ -32,7 +32,7 @@ class RandomDataGenerator : public framework::ReaderBase { dist_ = std::uniform_real_distribution(low_, high_); } - void ReadNext(std::vector* out) override { + void ReadNextImpl(std::vector* out) override { out->clear(); out->reserve(shapes_.size()); for (const framework::DDim& shape : shapes_) { @@ -51,8 +51,6 @@ class RandomDataGenerator : public framework::ReaderBase { } } - void ReInit() override { return; } - private: float low_; float high_; @@ -79,8 +77,8 @@ class CreateRandomDataGeneratorOp : public framework::OperatorBase { std::vector shapes = RestoreShapes(shape_concat, ranks); auto* out = scope.FindVar(Output("Out")) ->template GetMutable(); - out->Reset(new RandomDataGenerator(shapes, Attr("low"), - Attr("high"))); + out->Reset(std::make_shared>( + shapes, Attr("low"), Attr("high"))); } }; diff --git a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc index 282ec3f36b98e7aa62d71fb04f72721a5464e21c..a08a9dbd0da46e73082cdd24c019e8d210d8bcc4 100644 --- a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc +++ b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc @@ -21,10 +21,8 @@ namespace reader { template class RecordIOFileReader : public framework::FileReader { public: - explicit RecordIOFileReader(const std::string& filename, - const std::vector& dims) - : FileReader(dims), - scanner_(filename), + explicit RecordIOFileReader(const std::string& filename) + : scanner_(filename), dev_ctx_(*platform::DeviceContextPool::Instance().Get( platform::CPUPlace())) { if (ThreadSafe) { @@ -33,18 +31,21 @@ class RecordIOFileReader : public framework::FileReader { LOG(INFO) << "Creating file reader" << filename; } - void ReInit() override { scanner_.Reset(); } - protected: void ReadNextImpl(std::vector* out) override { + std::unique_ptr> guard; if (ThreadSafe) { - std::lock_guard guard(*mutex_); - *out = framework::ReadFromRecordIO(&scanner_, dev_ctx_); - } else { - *out = framework::ReadFromRecordIO(&scanner_, dev_ctx_); + guard.reset(new std::lock_guard(*mutex_)); + } + + bool ok = framework::ReadFromRecordIO(&scanner_, dev_ctx_, out); + if (!ok) { + out->clear(); } } + void StartImpl() override { scanner_.Reset(); } + private: std::unique_ptr mutex_; recordio::Scanner scanner_; @@ -58,31 +59,26 @@ class CreateRecordIOReaderOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope& scope, const platform::Place& dev_place) const override { - const auto& shape_concat = Attr>("shape_concat"); - const auto& ranks = Attr>("ranks"); - PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty()); - PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0), - static_cast(shape_concat.size()), - "The accumulate of all ranks should be equal to the " - "shape concat's length."); std::string filename = Attr("filename"); - auto* out = scope.FindVar(Output("Out")) ->template GetMutable(); - out->Reset(new RecordIOFileReader( - filename, RestoreShapes(shape_concat, ranks))); + out->Reset(std::make_shared>(filename)); } }; class CreateRecordIOReaderOpMaker : public FileReaderMakerBase { protected: void Apply() override { - AddAttr("filename", "The filename of record io reader"); + AddAttr( + "filename", + "The filename of record file. This file will given to reader."); AddComment(R"DOC( - CreateRecordIOReader Operator +Open a recordio file and return the reader object. The returned reader object +is thread-safe. - Create a reader from a record io file +NOTE: This is a very low-level API. It is used for debugging data file or +training. Please use `open_files` instead of this API for production usage. )DOC"); } }; diff --git a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc index fd233be945932eee9f9a3c0c578a43d5b7cc83aa..3f72890a7cee1453585d50afa04fa62a9b059dc3 100644 --- a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc +++ b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc @@ -23,7 +23,8 @@ namespace reader { class ShuffleReader : public framework::DecoratedReader { public: - ShuffleReader(ReaderBase* reader, size_t buffer_size, size_t seed = 0) + ShuffleReader(const std::shared_ptr& reader, size_t buffer_size, + size_t seed = 0) : DecoratedReader(reader), buffer_size_(buffer_size), seed_(seed) { VLOG(10) << "Create shuffle reader of " << reader_; if (seed_ == 0) { @@ -33,7 +34,7 @@ class ShuffleReader : public framework::DecoratedReader { ReloadBuffer(); } - void ReadNext(std::vector* out) override { + void ReadNextImpl(std::vector* out) override { out->clear(); if (iteration_pos_ >= buffer_.size()) { VLOG(10) << "Resetting shuffle buffer"; @@ -46,6 +47,17 @@ class ShuffleReader : public framework::DecoratedReader { } private: + void ShutdownImpl() override { + reader_->Shutdown(); + buffer_.clear(); + iteration_pos_ = 0; + } + + void StartImpl() override { + reader_->Start(); + ReloadBuffer(); + } + void ReloadBuffer() { buffer_.clear(); buffer_.reserve(buffer_size_); @@ -85,9 +97,8 @@ class CreateShuffleReaderOp : public framework::OperatorBase { } const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) ->Get(); - out->Reset( - new ShuffleReader(underlying_reader.Get(), - static_cast(Attr("buffer_size")))); + out->Reset(framework::MakeDecoratedReader( + underlying_reader, static_cast(Attr("buffer_size")))); } }; diff --git a/paddle/fluid/operators/reader/create_threaded_reader_op.cc b/paddle/fluid/operators/reader/create_threaded_reader_op.cc deleted file mode 100644 index 1db70f3e9699dba604569c36dc35025dfe2c94fe..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/reader/create_threaded_reader_op.cc +++ /dev/null @@ -1,78 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/operators/detail/safe_ref.h" -#include "paddle/fluid/operators/reader/reader_op_registry.h" - -namespace paddle { -namespace operators { -namespace reader { - -class ThreadedReader : public framework::DecoratedReader { - public: - explicit ThreadedReader(ReaderBase* reader) : DecoratedReader(reader) {} - - void ReadNext(std::vector* out) override { - std::lock_guard lock(mutex_); - reader_->ReadNext(out); - } - - void ReInit() override { reader_->ReInit(); } - - private: - std::mutex mutex_; -}; - -class CreateThreadedReaderOp : public framework::OperatorBase { - public: - using framework::OperatorBase::OperatorBase; - - private: - void RunImpl(const framework::Scope& scope, - const platform::Place& dev_place) const override { - auto* out = detail::Ref(scope.FindVar(Output("Out"))) - .GetMutable(); - if (out->Get() != nullptr) { - return; - } - const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) - ->Get(); - out->Reset(new ThreadedReader(underlying_reader.Get())); - } -}; - -class CreateThreadedReaderOpMaker : public DecoratedReaderMakerBase { - protected: - void Apply() override { - AddComment(R"DOC( - CreateThreadedReader Operator - - This operator creates a threaded reader. A threaded reader's - 'ReadNext()' can be invoked by several threads at the same - time. - When the attribute 'safe_mode' is true, the threaded reader's - 'ReInit()' is disabled to avoid unexpected bugs in multi-thread - environment. - )DOC"); - } -}; - -} // namespace reader -} // namespace operators -} // namespace paddle - -namespace reader = paddle::operators::reader; -REGISTER_DECORATED_READER_OPERATOR(create_threaded_reader, - reader::CreateThreadedReaderOp, - reader::CreateThreadedReaderOpMaker); diff --git a/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h b/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h new file mode 100644 index 0000000000000000000000000000000000000000..4f7cfc24ec035349f3c85e84d876ad9b5b5493a6 --- /dev/null +++ b/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h @@ -0,0 +1,89 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include + +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/operators/reader/blocking_queue.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace operators { +namespace reader { + +class LoDTensorBlockingQueueHolder; + +class LoDTensorBlockingQueue { + friend class LoDTensorBlockingQueueHolder; + + private: + LoDTensorBlockingQueue(size_t capacity, + const std::vector& dims) + : queue_(capacity), dims_(dims) {} + + public: + bool Push(const std::vector& lod_tensor_vec) { + return queue_.Send(lod_tensor_vec); + } + + bool Push(std::vector&& lod_tensor_vec) { + return queue_.Send(std::move(lod_tensor_vec)); + } + + std::vector Pop(bool* ok = nullptr) { + std::vector lod_tensor_vec; + bool success = queue_.Receive(&lod_tensor_vec); + if (ok != nullptr) *ok = success; + return lod_tensor_vec; + } + + inline size_t Cap() const { return queue_.Cap(); } + + inline size_t Size() const { return queue_.Size(); } + + inline void ReOpen() { queue_.ReOpen(); } + + inline void Close() { queue_.Close(); } + + inline bool IsClosed() const { return queue_.IsClosed(); } + + private: + BlockingQueue> queue_; + std::vector dims_; +}; + +class LoDTensorBlockingQueueHolder { + public: + void InitOnce(size_t capacity, const std::vector& dims) { + PADDLE_ENFORCE( + queue_ == nullptr, + "LoDTensorBlockingQueueHolder::InitOnce() can only be called once"); + queue_.reset(new LoDTensorBlockingQueue(capacity, dims)); + } + + inline const std::shared_ptr& GetQueue() const { + return queue_; + } + + private: + std::shared_ptr queue_; +}; + +} // namespace reader +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc index 8c0dac65dd691954b112bfa61622d399b2b9c3e5..38223e069975a08791d58d6ae10e2112b79a61fe 100644 --- a/paddle/fluid/operators/reader/open_files_op.cc +++ b/paddle/fluid/operators/reader/open_files_op.cc @@ -12,150 +12,200 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include #include // NOLINT - +#include "ThreadPool.h" +#include "paddle/fluid/framework/blocking_queue.h" #include "paddle/fluid/operators/reader/blocking_queue.h" +#include "paddle/fluid/operators/reader/buffered_reader.h" #include "paddle/fluid/operators/reader/reader_op_registry.h" namespace paddle { namespace operators { namespace reader { -class MultiFileReader : public framework::ReaderBase { +class IReaderContainer { public: - MultiFileReader(const std::vector& file_names, - const std::vector& dims, size_t thread_num, - size_t buffer_size) - : file_names_(file_names), dims_(dims), buffer_size_(buffer_size) { - prefetchers_.resize(thread_num); - StartNewScheduler(); + virtual ~IReaderContainer() {} + virtual void AppendReader( + std::unique_ptr&& readers) = 0; + virtual void Stop() = 0; + virtual void Start() = 0; + virtual void ReadNext(std::vector* out) = 0; +}; + +class OrderedReaderContainer : public IReaderContainer { + public: + void AppendReader(std::unique_ptr&& reader) override { + pending_.emplace(std::move(reader)); } - void ReadNext(std::vector* out) override; - void ReInit() override; + void Stop() override { + while (!pending_.empty()) { + MoveFrontPendingToDone(); + } + } - ~MultiFileReader() { EndScheduler(); } + void Start() override { std::swap(done_, pending_); } + + void ReadNext(std::vector* out) override { + if (!pending_.empty()) { + pending_.front()->ReadNext(out); + if (out->empty()) { + MoveFrontPendingToDone(); + ReadNext(out); + } + } else { + out->clear(); + } + } private: - void StartNewScheduler(); - void EndScheduler(); - void ScheduleThreadFunc(); - void PrefetchThreadFunc(std::string file_name, size_t thread_idx); - - std::vector file_names_; - std::vector dims_; - std::thread scheduler_; - std::vector prefetchers_; - size_t buffer_size_; - reader::BlockingQueue* waiting_file_idx_; - reader::BlockingQueue* available_thread_idx_; - reader::BlockingQueue>* buffer_; + void MoveFrontPendingToDone() { + pending_.front()->Shutdown(); + pending_.front()->Start(); + done_.emplace(move(pending_.front())); + pending_.pop(); + } + + std::queue> pending_; + std::queue> done_; }; -void MultiFileReader::ReadNext(std::vector* out) { - if (!buffer_->Receive(out)) { - out->clear(); - } -} - -void MultiFileReader::ReInit() { - EndScheduler(); - StartNewScheduler(); -} - -void MultiFileReader::StartNewScheduler() { - size_t thread_num = prefetchers_.size(); - waiting_file_idx_ = new reader::BlockingQueue(file_names_.size()); - available_thread_idx_ = new reader::BlockingQueue(thread_num); - buffer_ = new reader::BlockingQueue>( - buffer_size_); - - for (size_t i = 0; i < file_names_.size(); ++i) { - waiting_file_idx_->Send(i); - } - waiting_file_idx_->Close(); - for (size_t i = 0; i < thread_num; ++i) { - available_thread_idx_->Send(i); - } +class PreemptiveReaderContainer : public IReaderContainer { + using ReaderList = std::list>; - scheduler_ = std::thread([this] { ScheduleThreadFunc(); }); -} + struct FutureItem { + std::vector data_; + ReaderList::iterator reader_it_; + std::exception_ptr exception_; + }; -void MultiFileReader::EndScheduler() { - available_thread_idx_->Close(); - buffer_->Close(); - waiting_file_idx_->Close(); - if (scheduler_.joinable()) { - scheduler_.join(); - } - delete buffer_; - delete available_thread_idx_; - delete waiting_file_idx_; -} - -void MultiFileReader::ScheduleThreadFunc() { - VLOG(5) << "MultiFileReader schedule thread starts."; - size_t completed_thread_num = 0; - size_t thread_idx; - while (available_thread_idx_->Receive(&thread_idx)) { - std::thread& prefetcher = prefetchers_[thread_idx]; - if (prefetcher.joinable()) { - prefetcher.join(); - } - size_t file_idx; - if (waiting_file_idx_->Receive(&file_idx)) { - // Still have files to read. Start a new prefetch thread. - std::string file_name = file_names_[file_idx]; - prefetcher = std::thread([this, file_name, thread_idx] { - PrefetchThreadFunc(file_name, thread_idx); - }); - } else { - // No more file to read. - ++completed_thread_num; - if (completed_thread_num == prefetchers_.size()) { - buffer_->Close(); - break; + using FutureList = std::list>; + + public: + explicit PreemptiveReaderContainer(size_t thread_num) : pool_(thread_num) {} + + void Stop() override { + if (!pending_.empty()) { + for (auto& reader : pending_) { + reader->Shutdown(); + } + for (auto& fu : futures_) { + fu.wait(); + } + futures_.clear(); + for (auto& reader : pending_) { + reader->Start(); + done_.emplace_back(std::move(reader)); } + pending_.clear(); + bool timeout; + complete_queue_.PopAll(1000, &timeout); + PADDLE_ENFORCE(!timeout); } } - // If users invoke ReInit() when scheduler is running, it will close the - // 'avaiable_thread_idx_' and prefecther threads have no way to tell scheduler - // to release their resource. So a check is needed before scheduler ends. - for (auto& p : prefetchers_) { - if (p.joinable()) { - p.join(); + + void Start() override { + for (auto& reader : done_) { + AppendReader(std::move(reader)); } + done_.clear(); } - VLOG(5) << "MultiFileReader schedule thread terminates."; -} - -void MultiFileReader::PrefetchThreadFunc(std::string file_name, - size_t thread_idx) { - VLOG(5) << "The prefetch thread of file '" << file_name << "' starts."; - std::unique_ptr reader = - CreateReaderByFileName(file_name, dims_); - while (true) { - std::vector ins; - reader->ReadNext(&ins); - if (ins.empty()) { - break; + + void ReadNext(std::vector* out) override { + if (!pending_.empty()) { + auto future_it = complete_queue_.Pop(); + FutureItem item = future_it->get(); + if (item.exception_) { + for (auto it = futures_.begin(); it != futures_.end(); ++it) { + if (it != future_it) { + it->wait(); // Wait all other threads complete. + } + } + std::rethrow_exception(item.exception_); + + } else if (item.data_.empty()) { // reader done. + done_.emplace_back(std::move(*item.reader_it_)); + pending_.erase(item.reader_it_); + futures_.erase(future_it); + ReadNext(out); + } else { + *out = item.data_; + // continue read async + ReadAsync(item.reader_it_, &future_it); + } + } else { + out->clear(); } - try { - buffer_->Send(std::move(ins)); - } catch (paddle::platform::EnforceNotMet e) { - VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch " - "thread of file '" - << file_name << "' will terminate."; - break; + } + + private: + void AppendReader(std::unique_ptr&& reader) override { + pending_.emplace_back(std::move(reader)); + auto reader_it = pending_.end(); + --reader_it; + + futures_.emplace_back(); + auto future_it = futures_.end(); + --future_it; + + ReadAsync(reader_it, &future_it); + } + + void ReadAsync(const ReaderList::iterator& reader_it, + FutureList::iterator* future_it_ptr) { + auto& future_it = *future_it_ptr; + *future_it = pool_.enqueue([reader_it, future_it, this] { + try { + FutureItem item; + item.reader_it_ = reader_it; + (*reader_it)->ReadNext(&item.data_); + if (item.data_.empty()) { + (*reader_it)->Shutdown(); + (*reader_it)->Start(); + } + complete_queue_.Push(future_it); + return item; + } catch (...) { + FutureItem item; + item.exception_ = std::current_exception(); + complete_queue_.Push(future_it); + return item; + } + }); + } + + FutureList futures_; + ThreadPool pool_; + framework::BlockingQueue complete_queue_; + std::list> pending_; + std::list> done_; +}; + +class MultiFileReader : public framework::ReaderBase { + public: + MultiFileReader(const std::vector& file_names, + std::unique_ptr&& container) + : container_(std::move(container)) { + for (auto& fn : file_names) { + container_->AppendReader(CreateReaderByFileName(fn)); } } - if (!available_thread_idx_->Send(thread_idx)) { - VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. " - "Fail to send thread_idx."; + ~MultiFileReader() { container_->Stop(); } + + protected: + void ReadNextImpl(std::vector* out) override { + container_->ReadNext(out); } - VLOG(5) << "The prefetch thread of file '" << file_name << "' terminates."; -} + void ShutdownImpl() override { container_->Stop(); } + void StartImpl() override { container_->Start(); } + + private: + std::unique_ptr container_; +}; class OpenFilesOp : public framework::OperatorBase { public: @@ -173,14 +223,27 @@ class OpenFilesOp : public framework::OperatorBase { "shape concat's length."); const auto& file_names = Attr>("file_names"); PADDLE_ENFORCE(!file_names.empty(), "No file to be read!"); - const size_t thread_num = Attr("thread_num"); - const size_t buffer_size = Attr("buffer_size"); + bool is_test = Attr("is_test"); auto* out = scope.FindVar(Output("Out")) ->template GetMutable(); - out->Reset(new MultiFileReader(file_names, - RestoreShapes(shape_concat, ranks), - thread_num, buffer_size)); + std::unique_ptr container; + + if (is_test) { + container.reset(new OrderedReaderContainer()); + } else { + container.reset(new PreemptiveReaderContainer( + static_cast(Attr("thread_num")))); + } + + std::shared_ptr reader( + new MultiFileReader(file_names, std::move(container))); + auto buffer_size = Attr("buffer_size"); + if (buffer_size > 1) { + reader = framework::MakeDecoratedReader( + reader, platform::CPUPlace(), buffer_size); + } + out->Reset(reader); } }; @@ -188,9 +251,7 @@ class OpenFilesOpMaker : public FileReaderMakerBase { protected: void Apply() override { AddAttr>("file_names", "Files to be read."); - AddAttr("thread_num", "The maximal concurrent prefetch thread number.") - .GreaterThan(0); - AddAttr("buffer_size", "The size of prefetch buffer.").GreaterThan(0); + AddAttr("is_test", "Used for testing data.").SetDefault(false); AddComment(R"DOC( OpenFiles Operator @@ -198,6 +259,11 @@ class OpenFilesOpMaker : public FileReaderMakerBase { An OpenFilesOp creates a MultiFileReader, which is able to read data multi-threaded from multiple files. )DOC"); + AddAttr("thread_num", + "The maximal concurrent prefetch thread number. Used only " + "when is_test = False"); + AddAttr("buffer_size", "The reading buffer of these files.") + .GreaterThan(0); } }; diff --git a/paddle/fluid/operators/reader/reader_op_registry.cc b/paddle/fluid/operators/reader/reader_op_registry.cc index 612e1f5eca3a4836db1fd167fc6bb63400d20177..b82aab1214992be73d876a42424234e3cea46455 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.cc +++ b/paddle/fluid/operators/reader/reader_op_registry.cc @@ -39,7 +39,7 @@ std::unordered_map& FileReaderRegistry() { } std::unique_ptr CreateReaderByFileName( - const std::string& file_name, const std::vector& dims) { + const std::string& file_name) { size_t separator_pos = file_name.find_last_of(kFileFormatSeparator); PADDLE_ENFORCE_NE(separator_pos, std::string::npos, "File name illegal! A legal file name should be like: " @@ -49,12 +49,12 @@ std::unique_ptr CreateReaderByFileName( auto itor = FileReaderRegistry().find(filetype); PADDLE_ENFORCE(itor != FileReaderRegistry().end(), "No file reader registered for '%s' format.", filetype); - framework::ReaderBase* reader = (itor->second)(file_name, dims); + framework::ReaderBase* reader = (itor->second)(file_name); return std::unique_ptr(reader); } void FileReaderMakerBase::Make() { - AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable(); + AddOutput("Out", "(ReaderHolder): The created random reader.").AsDuplicable(); AddAttr>("shape_concat", "The concat of all data's shapes."); AddAttr>( "ranks", diff --git a/paddle/fluid/operators/reader/reader_op_registry.h b/paddle/fluid/operators/reader/reader_op_registry.h index 244bf15f068a47efc29ee54492cdbdeb10025020..25c3e7d77b788d38daf6dee1fc79e5c1c97e8842 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.h +++ b/paddle/fluid/operators/reader/reader_op_registry.h @@ -25,22 +25,21 @@ namespace reader { static constexpr char kFileFormatSeparator[] = "."; -using FileReaderCreator = std::function&)>; +using FileReaderCreator = + std::function; std::unordered_map& FileReaderRegistry(); template int RegisterFileReader(const std::string& filetype) { - FileReaderRegistry()[filetype] = []( - const std::string& fn, const std::vector& dims) { - return new Reader(fn, dims); + FileReaderRegistry()[filetype] = [](const std::string& fn) { + return new Reader(fn); }; return 0; } std::unique_ptr CreateReaderByFileName( - const std::string& file_name, const std::vector& dims); + const std::string& file_name); extern std::vector RestoreShapes( const std::vector& shape_concat, const std::vector& ranks); diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 9c1cee7022a9b9a98f026f7602f0f7badc44a49b..162bfcbb0844d29385d0f8ad5d25a3f8de6bd41b 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase { auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {pg_names[param_id], new_inside_name}}}, - {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{}); + {{"Out", {pg_names[param_id]}}}, + framework::AttributeMap{{"use_mkldnn", {false}}}); sum_op->Run(cur_scope, place); cur_scope.Rename(new_inside_name, inside_grad_name); diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index d8ddb7b448910b5e0e6e71742eb2fdc6a225c919..4a6ce938a5f337d035b21f562d46daf606236db0 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -19,8 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" - -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -41,17 +40,16 @@ class RecvOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - // For profiling - platform::RecordEvent record_event(Type(), &ctx); - auto rpc_client = detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); for (size_t i = 0; i < outs.size(); i++) { VLOG(3) << "getting " << outs[i] << " from " << epmap[i]; - rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); + rpc_client->AsyncGetVar(epmap[i], ctx, scope, outs[i]); } if (sync_mode) { - PADDLE_ENFORCE(rpc_client->Wait()); + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } } }; @@ -77,9 +75,15 @@ This operator can get variables from server side. } }; +class RecvOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker); +REGISTER_OPERATOR(recv, ops::RecvOp, paddle::framework::EmptyGradOpMaker, + ops::RecvOpMaker, ops::RecvOpShapeInference); diff --git a/paddle/fluid/operators/reduce_max_op.cc b/paddle/fluid/operators/reduce_max_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..95d3768e1fdf6947659c7b3a1c9d57fad741472a --- /dev/null +++ b/paddle/fluid/operators/reduce_max_op.cc @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_REDUCE_OP(reduce_max); +REGISTER_OP_CPU_KERNEL( + reduce_max, ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL( + reduce_max_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_max_op.cu b/paddle/fluid/operators/reduce_max_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..0d86b3127e42f7ee14ba57b1c762e8128a0f2d54 --- /dev/null +++ b/paddle/fluid/operators/reduce_max_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_max, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_max_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.cc b/paddle/fluid/operators/reduce_mean_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..fc258c2496340b47d24dc89f16f7419dbb4b0d95 --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.cc @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_mean_op.h" + +REGISTER_REDUCE_OP(reduce_mean); +REGISTER_OP_CPU_KERNEL(reduce_mean, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_mean_grad, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.cu b/paddle/fluid/operators/reduce_mean_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..960cb3235be7f4cc98b97d3b088ceaeb3d4a4209 --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_mean_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_mean, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_mean_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.h b/paddle/fluid/operators/reduce_mean_op.h new file mode 100644 index 0000000000000000000000000000000000000000..1359679c4767d2032bf3e3a90849ad2a2ef3e829 --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.h @@ -0,0 +1,39 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct MeanFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->mean(dim); + } +}; + +struct MeanGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + dx->device(place) = dy->broadcast(dim) / dx->constant(size); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_min_max_op.h b/paddle/fluid/operators/reduce_min_max_op.h new file mode 100644 index 0000000000000000000000000000000000000000..ec59f3e71c1c702655a3feed10935b2f5a29d8a8 --- /dev/null +++ b/paddle/fluid/operators/reduce_min_max_op.h @@ -0,0 +1,50 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct MaxFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->maximum(dim); + } +}; + +struct MinFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->minimum(dim); + } +}; + +struct MaxOrMinGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + auto equals = (*x) == y->broadcast(dim); + auto ones = dx->constant(1); + auto zeros = dx->constant(0); + // If there are multiple minimum or maximum elements, the subgradient of + // each is the set [0, 1], and we pass gradient to all of them here. + dx->device(place) = dy->broadcast(dim) * equals.select(ones, zeros); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_min_op.cc b/paddle/fluid/operators/reduce_min_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..330a86d2e4237a10d8cf6fd40025540edf08d897 --- /dev/null +++ b/paddle/fluid/operators/reduce_min_op.cc @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_REDUCE_OP(reduce_min); +REGISTER_OP_CPU_KERNEL( + reduce_min, ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL( + reduce_min_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_min_op.cu b/paddle/fluid/operators/reduce_min_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..da466f805eff4709dc23471baef03e94052ee6c1 --- /dev/null +++ b/paddle/fluid/operators/reduce_min_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_min, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_min_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_op.cc b/paddle/fluid/operators/reduce_op.cc deleted file mode 100644 index e293fd5e410b2a34b3c71ea674607ba9d7654535..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/reduce_op.cc +++ /dev/null @@ -1,186 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/reduce_op.h" - -#include -#include -#include - -namespace paddle { -namespace operators { - -using framework::Tensor; - -class ReduceOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of ReduceOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of ReduceOp should not be null."); - auto x_dims = ctx->GetInputDim("X"); - auto x_rank = x_dims.size(); - PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); - auto dims = ctx->Attrs().Get>("dim"); - for (size_t i = 0; i < dims.size(); ++i) { - if (dims[i] < 0) dims[i] = x_rank + dims[i]; - PADDLE_ENFORCE_LT( - dims[i], x_rank, - "The dim should be in the range [-rank(input), rank(input))."); - } - sort(dims.begin(), dims.end()); - bool reduce_all = ctx->Attrs().Get("reduce_all"); - bool keep_dim = ctx->Attrs().Get("keep_dim"); - if (reduce_all) { - if (keep_dim) - ctx->SetOutputDim( - "Out", framework::make_ddim(std::vector(x_rank, 1))); - else - ctx->SetOutputDim("Out", {1}); - } else { - auto dims_vector = vectorize(x_dims); - if (keep_dim) { - for (size_t i = 0; i < dims.size(); ++i) { - dims_vector[dims[i]] = 1; - } - } else { - const int kDelFlag = -2; - for (size_t i = 0; i < dims.size(); ++i) { - dims_vector[dims[i]] = kDelFlag; - } - dims_vector.erase( - remove(dims_vector.begin(), dims_vector.end(), kDelFlag), - dims_vector.end()); - } - auto out_dims = framework::make_ddim(dims_vector); - ctx->SetOutputDim("Out", out_dims); - if (dims[0] != 0) { - // Only pass LoD when not reducing on the first dim. - ctx->ShareLoD("X", /*->*/ "Out"); - } - } - } -}; - -class ReduceGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null."); - auto x_dims = ctx->GetInputDim("X"); - auto x_rank = x_dims.size(); - PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); - auto dims = ctx->Attrs().Get>("dim"); - for (size_t i = 0; i < dims.size(); ++i) { - if (dims[i] < 0) dims[i] = x_rank + dims[i]; - PADDLE_ENFORCE_LT( - dims[i], x_rank, - "The dim should be in the range [-rank(input), rank(input))."); - } - sort(dims.begin(), dims.end()); - auto x_grad_name = framework::GradVarName("X"); - if (ctx->HasOutput(x_grad_name)) { - ctx->SetOutputDim(x_grad_name, x_dims); - ctx->ShareLoD("X", /*->*/ x_grad_name); - } - } -}; - -class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() final { - AddInput("X", - "(Tensor) The input tensor. Tensors with rank at most 6 are " - "supported."); - AddOutput("Out", "(Tensor) The result tensor."); - AddAttr>( - "dim", - "(list, default {0}) The dimensions to reduce. " - "Must be in the range [-rank(input), rank(input)). " - "If `dim[i] < 0`, the dims[i] to reduce is `rank + dims[i]`. " - "Note that reducing on the first dim will make the LoD info lost.") - .SetDefault({0}); - AddAttr("keep_dim", - "(bool, default false) " - "If true, retain the reduced dimension with length 1.") - .SetDefault(false); - AddAttr("reduce_all", - "(bool, default false) " - "If true, output a scalar reduced along all dimensions.") - .SetDefault(false); - AddComment(string::Sprintf(R"DOC( -%s Operator. - -This operator computes the %s of input tensor along the given dimension. -The result tensor has 1 fewer dimension than the input unless keep_dim is true. -If reduce_all is true, just reduce along all dimensions and output a scalar. - -)DOC", - GetOpType(), GetName())); - } - - protected: - virtual std::string GetName() const = 0; - virtual std::string GetOpType() const = 0; -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; - -#define REGISTER_REDUCE_OP(op_name) \ - class __##op_name##Maker__ : public ops::ReduceOpMaker { \ - protected: \ - virtual std::string GetName() const { return #op_name; } \ - virtual std::string GetOpType() const { return "Reduce " #op_name; } \ - }; \ - REGISTER_OPERATOR(reduce_##op_name, ops::ReduceOp, __##op_name##Maker__, \ - paddle::framework::DefaultGradOpDescMaker); \ - REGISTER_OPERATOR(reduce_##op_name##_grad, ops::ReduceGradOp) - -REGISTER_REDUCE_OP(sum); -REGISTER_REDUCE_OP(mean); -REGISTER_REDUCE_OP(max); -REGISTER_REDUCE_OP(min); -REGISTER_REDUCE_OP(prod); - -#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ - REGISTER_OP_CPU_KERNEL(reduce_type, \ - ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel); \ - REGISTER_OP_CPU_KERNEL( \ - reduce_type##_grad, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel); - -FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL); diff --git a/paddle/fluid/operators/reduce_op.cu b/paddle/fluid/operators/reduce_op.cu deleted file mode 100644 index ae29587f55847315b1d84f1344677e753fe01a9b..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/reduce_op.cu +++ /dev/null @@ -1,41 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#define EIGEN_USE_GPU -#include "paddle/fluid/operators/reduce_op.h" - -namespace ops = paddle::operators; - -#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor) \ - REGISTER_OP_CUDA_KERNEL( \ - reduce_type, ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel); \ - REGISTER_OP_CUDA_KERNEL( \ - reduce_type##_grad, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel); - -FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL); diff --git a/paddle/fluid/operators/reduce_op.h b/paddle/fluid/operators/reduce_op.h index cd19cc1460a6b4d4201f21f6f27f988c1547b88a..72b6cf1773d5bcc42e40e72111179d454d2bb4a9 100644 --- a/paddle/fluid/operators/reduce_op.h +++ b/paddle/fluid/operators/reduce_op.h @@ -14,105 +14,20 @@ limitations under the License. */ #pragma once +#include +#include #include -#include "glog/logging.h" -#include "paddle/fluid/framework/eigen.h" -#include "paddle/fluid/framework/op_registry.h" + +#include "paddle/fluid/operators/reduce_op_function.h" namespace paddle { namespace operators { -using Tensor = framework::Tensor; -using DDim = framework::DDim; -template -using EigenTensor = framework::EigenTensor; -template -using EigenScalar = framework::EigenScalar; -template -using EigenVector = framework::EigenVector; - -struct SumFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->sum(dim); - } -}; - -struct SumGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - dx->device(place) = dy->broadcast(dim); - } -}; - -struct MeanFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->mean(dim); - } -}; - -struct MeanGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - dx->device(place) = dy->broadcast(dim) / dx->constant(size); - } -}; - -struct MaxFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->maximum(dim); - } -}; - -struct MinFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->minimum(dim); - } -}; - -struct MaxOrMinGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - auto equals = (*x) == y->broadcast(dim); - auto ones = dx->constant(1); - auto zeros = dx->constant(0); - // If there are multiple minimum or maximum elements, the subgradient of - // each is the set [0, 1], and we pass gradient to all of them here. - dx->device(place) = dy->broadcast(dim) * equals.select(ones, zeros); - } -}; - -struct ProdFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->prod(dim); - } -}; - -struct ProdGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - dx->device(place) = dy->broadcast(dim) * y->broadcast(dim) * x->inverse(); - } -}; - -#define HANDLE_DIM(NDIM, RDIM) \ - if (ndim == NDIM && rdim == RDIM) { \ - ReduceCompute(context); \ +#define HANDLE_DIM(NDIM, RDIM) \ + if (ndim == NDIM && rdim == RDIM) { \ + ReduceFunctor( \ + context.template device_context(), *input, output, \ + dims, keep_dim); \ } template @@ -120,11 +35,15 @@ class ReduceKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { bool reduce_all = context.Attr("reduce_all"); + auto* input = context.Input("X"); + auto* output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + + auto dims = context.Attr>("dim"); + bool keep_dim = context.Attr("keep_dim"); + if (reduce_all) { // Flatten and reduce 1-D tensor - auto* input = context.Input("X"); - auto* output = context.Output("Out"); - output->mutable_data(context.GetPlace()); auto x = EigenVector::Flatten(*input); auto out = EigenScalar::From(*output); auto& place = @@ -133,17 +52,18 @@ class ReduceKernel : public framework::OpKernel { Functor functor; functor(place, &x, &out, reduce_dim); } else { - int ndim = context.Input("X")->dims().size(); - int rdim = context.Attr>("dim").size(); - HANDLE_DIM(6, 5); - HANDLE_DIM(6, 4); - HANDLE_DIM(6, 3); - HANDLE_DIM(6, 2); - HANDLE_DIM(6, 1); - HANDLE_DIM(5, 4); - HANDLE_DIM(5, 3); - HANDLE_DIM(5, 2); - HANDLE_DIM(5, 1); + int ndim = input->dims().size(); + int rdim = dims.size(); + // comments for accelerating compiling temporarily. + // HANDLE_DIM(6, 5); + // HANDLE_DIM(6, 4); + // HANDLE_DIM(6, 3); + // HANDLE_DIM(6, 2); + // HANDLE_DIM(6, 1); + // HANDLE_DIM(5, 4); + // HANDLE_DIM(5, 3); + // HANDLE_DIM(5, 2); + // HANDLE_DIM(5, 1); HANDLE_DIM(4, 3); HANDLE_DIM(4, 2); HANDLE_DIM(4, 1); @@ -153,48 +73,6 @@ class ReduceKernel : public framework::OpKernel { HANDLE_DIM(1, 1); } } - - private: - template - void ReduceCompute(const framework::ExecutionContext& context) const { - auto* input = context.Input("X"); - auto* output = context.Output("Out"); - output->mutable_data(context.GetPlace()); - - auto x = EigenTensor::From(*input); - auto x_rank = static_cast(x.dimensions().size()); - auto dims = context.Attr>("dim"); - auto reduce_dim = Eigen::array(); - for (size_t i = 0; i < dims.size(); ++i) { - if (dims[i] < 0) dims[i] = x_rank + dims[i]; - reduce_dim[i] = dims[i]; - } - // construct the squeezed output tensor - bool keep_dim = context.Attr("keep_dim"); - DDim out_dims = output->dims(); - if (keep_dim && x_rank > 1) { - const int kDelFlag = -2; - auto dims_vector = vectorize(out_dims); - for (size_t i = 0; i < dims.size(); ++i) { - dims_vector[dims[i]] = kDelFlag; - } - dims_vector.erase( - remove(dims_vector.begin(), dims_vector.end(), kDelFlag), - dims_vector.end()); - out_dims = framework::make_ddim(dims_vector); - } - auto& place = - *context.template device_context().eigen_device(); - Functor functor; - - if (D == 1) { - auto out = EigenScalar::From(*output); - functor(place, &x, &out, reduce_dim); - } else { - auto out = EigenTensor::From(*output, out_dims); - functor(place, &x, &out, reduce_dim); - } - } }; template @@ -202,12 +80,15 @@ class ReduceGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { bool reduce_all = context.Attr("reduce_all"); + auto dims = context.Attr>("dim"); + + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Out"); + auto* input2 = context.Input(framework::GradVarName("Out")); + auto* output = context.Output(framework::GradVarName("X")); + output->mutable_data(context.GetPlace()); + if (reduce_all) { - auto* input0 = context.Input("X"); - auto* input1 = context.Input("Out"); - auto* input2 = context.Input(framework::GradVarName("Out")); - auto* output = context.Output(framework::GradVarName("X")); - output->mutable_data(context.GetPlace()); auto x = EigenVector::Flatten(*input0); auto x_reduce = EigenVector::From(*input1); auto x_reduce_grad = EigenVector::From(*input2); @@ -220,74 +101,172 @@ class ReduceGradKernel : public framework::OpKernel { functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, broadcast_dim[0]); } else { - int rank = context.Input("X")->dims().size(); + int rank = input0->dims().size(); switch (rank) { case 1: - ReduceGradCompute<1>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 2: - ReduceGradCompute<2>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 3: - ReduceGradCompute<3>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 4: - ReduceGradCompute<4>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 5: - ReduceGradCompute<5>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 6: - ReduceGradCompute<6>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; } } } +}; - private: - template - void ReduceGradCompute(const framework::ExecutionContext& context) const { - auto* input0 = context.Input("X"); - auto* input1 = context.Input("Out"); - auto* input2 = context.Input(framework::GradVarName("Out")); - auto* output = context.Output(framework::GradVarName("X")); +class ReduceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; - output->mutable_data(context.GetPlace()); - auto x = EigenTensor::From(*input0); - auto x_grad = EigenTensor::From(*output); - auto x_rank = static_cast(x.dimensions().size()); - auto dims = context.Attr>("dim"); - auto x_dims = input0->dims(); - auto reduced_dims_v = vectorize(x_dims); - Eigen::array broadcast_dim; - for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReduceOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReduceOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + auto dims = ctx->Attrs().Get>("dim"); + for (size_t i = 0; i < dims.size(); ++i) { + if (dims[i] < 0) dims[i] = x_rank + dims[i]; + PADDLE_ENFORCE_LT( + dims[i], x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + } + sort(dims.begin(), dims.end()); + bool reduce_all = ctx->Attrs().Get("reduce_all"); + bool keep_dim = ctx->Attrs().Get("keep_dim"); + if (reduce_all) { + if (keep_dim) + ctx->SetOutputDim( + "Out", framework::make_ddim(std::vector(x_rank, 1))); + else + ctx->SetOutputDim("Out", {1}); + } else { + auto dims_vector = vectorize(x_dims); + if (keep_dim) { + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = 1; + } + } else { + const int kDelFlag = -2; + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = kDelFlag; + } + dims_vector.erase( + remove(dims_vector.begin(), dims_vector.end(), kDelFlag), + dims_vector.end()); + } + auto out_dims = framework::make_ddim(dims_vector); + ctx->SetOutputDim("Out", out_dims); + if (dims[0] != 0) { + // Only pass LoD when not reducing on the first dim. + ctx->ShareLoD("X", /*->*/ "Out"); + } + } + } +}; + +class ReduceGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; - int broad_cats_times = 1; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + auto dims = ctx->Attrs().Get>("dim"); for (size_t i = 0; i < dims.size(); ++i) { if (dims[i] < 0) dims[i] = x_rank + dims[i]; - reduced_dims_v[dims[i]] = 1; - broadcast_dim[dims[i]] = x_dims[dims[i]]; - broad_cats_times *= x_dims[dims[i]]; + PADDLE_ENFORCE_LT( + dims[i], x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + } + sort(dims.begin(), dims.end()); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + ctx->ShareLoD("X", /*->*/ x_grad_name); } - auto reduced_dims = framework::make_ddim(reduced_dims_v); - auto x_reduce = EigenTensor::From(*input1, reduced_dims); - auto x_reduce_grad = EigenTensor::From(*input2, reduced_dims); + } +}; + +class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() final { + AddInput("X", + "(Tensor) The input tensor. Tensors with rank at most 6 are " + "supported."); + AddOutput("Out", "(Tensor) The result tensor."); + AddAttr>( + "dim", + "(list, default {0}) The dimensions to reduce. " + "Must be in the range [-rank(input), rank(input)). " + "If `dim[i] < 0`, the dims[i] to reduce is `rank + dims[i]`. " + "Note that reducing on the first dim will make the LoD info lost.") + .SetDefault({0}); + AddAttr("keep_dim", + "(bool, default false) " + "If true, retain the reduced dimension with length 1.") + .SetDefault(false); + AddAttr("reduce_all", + "(bool, default false) " + "If true, output a scalar reduced along all dimensions.") + .SetDefault(false); + AddComment(string::Sprintf(R"DOC( +%s Operator. - auto& place = - *context.template device_context().eigen_device(); +This operator computes the %s of input tensor along the given dimension. +The result tensor has 1 fewer dimension than the input unless keep_dim is true. +If reduce_all is true, just reduce along all dimensions and output a scalar. - Functor functor; - functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, - broad_cats_times); +)DOC", + GetOpType(), GetName())); } + + protected: + virtual std::string GetName() const = 0; + virtual std::string GetOpType() const = 0; }; } // namespace operators } // namespace paddle -#define FOR_EACH_KERNEL_FUNCTOR(__macro) \ - __macro(reduce_sum, SumFunctor, SumGradFunctor); \ - __macro(reduce_mean, MeanFunctor, MeanGradFunctor); \ - __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \ - __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor); \ - __macro(reduce_prod, ProdFunctor, ProdGradFunctor); +namespace ops = paddle::operators; + +#define REGISTER_REDUCE_OP(op_name) \ + class __##op_name##Maker__ : public ops::ReduceOpMaker { \ + protected: \ + virtual std::string GetName() const { return #op_name; } \ + virtual std::string GetOpType() const { return "Reduce " #op_name; } \ + }; \ + REGISTER_OPERATOR(op_name, ops::ReduceOp, __##op_name##Maker__, \ + paddle::framework::DefaultGradOpDescMaker); \ + REGISTER_OPERATOR(op_name##_grad, ops::ReduceGradOp) diff --git a/paddle/fluid/operators/reduce_op_function.h b/paddle/fluid/operators/reduce_op_function.h new file mode 100644 index 0000000000000000000000000000000000000000..3da27bc8ac8d448471b9ff3779ac6aca59fac523 --- /dev/null +++ b/paddle/fluid/operators/reduce_op_function.h @@ -0,0 +1,109 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using DDim = framework::DDim; +template +using EigenTensor = framework::EigenTensor; +template +using EigenScalar = framework::EigenScalar; +template +using EigenVector = framework::EigenVector; + +template +void ReduceFunctor(const DeviceContext& context, const framework::Tensor& input, + framework::Tensor* output, const std::vector& dims, + bool keep_dim) { + auto x = EigenTensor::From(input); + auto x_rank = static_cast(x.dimensions().size()); + auto reduce_dim = Eigen::array(); + std::vector dims_ref = dims; + for (size_t i = 0; i < dims_ref.size(); ++i) { + if (dims_ref[i] < 0) dims_ref[i] = x_rank + dims_ref[i]; + reduce_dim[i] = dims_ref[i]; + } + // construct the squeezed output tensor + DDim out_dims = output->dims(); + if (keep_dim && x_rank > 1) { + const int kDelFlag = -2; + auto dims_vector = framework::vectorize(out_dims); + for (size_t i = 0; i < dims_ref.size(); ++i) { + dims_vector[dims_ref[i]] = kDelFlag; + } + dims_vector.erase(remove(dims_vector.begin(), dims_vector.end(), kDelFlag), + dims_vector.end()); + out_dims = framework::make_ddim(dims_vector); + } + auto& place = *context.eigen_device(); + Functor functor; + + if (D == 1) { + auto out = EigenScalar::From(*output); + functor(place, &x, &out, reduce_dim); + } else { + auto out = EigenTensor::From(*output, out_dims); + functor(place, &x, &out, reduce_dim); + } +} + +template +void ReduceGradFunctor(const DeviceContext& context, + const framework::Tensor& input0, + const framework::Tensor& input1, + const framework::Tensor& input2, + framework::Tensor* output, + const std::vector& dims) { + auto x = EigenTensor::From(input0); + auto x_grad = EigenTensor::From(*output); + auto x_rank = static_cast(x.dimensions().size()); + auto x_dims = input0.dims(); + auto reduced_dims_v = framework::vectorize(x_dims); + std::vector dims_ref = dims; + Eigen::array broadcast_dim; + for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1; + + int broad_cats_times = 1; + for (size_t i = 0; i < dims_ref.size(); ++i) { + if (dims_ref[i] < 0) { + dims_ref[i] = x_rank + dims_ref[i]; + } + reduced_dims_v[dims_ref[i]] = 1; + broadcast_dim[dims_ref[i]] = x_dims[dims_ref[i]]; + broad_cats_times *= x_dims[dims_ref[i]]; + } + auto reduced_dims = framework::make_ddim(reduced_dims_v); + auto x_reduce = EigenTensor::From(input1, reduced_dims); + auto x_reduce_grad = EigenTensor::From(input2, reduced_dims); + + auto& place = *context.eigen_device(); + + Functor functor; + functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, + broad_cats_times); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_prod_op.cc b/paddle/fluid/operators/reduce_prod_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..713728b99757a6f3bb128f665d5576ac64eef8ec --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.cc @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_prod_op.h" + +REGISTER_REDUCE_OP(reduce_prod); +REGISTER_OP_CPU_KERNEL(reduce_prod, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_prod_grad, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_prod_op.cu b/paddle/fluid/operators/reduce_prod_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..d62e677d92cffecf629d1684026b0c7bcfec29e3 --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_prod_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_prod, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_prod_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_prod_op.h b/paddle/fluid/operators/reduce_prod_op.h new file mode 100644 index 0000000000000000000000000000000000000000..97748113e092719aceed9d806ca6242077111532 --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.h @@ -0,0 +1,39 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct ProdFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->prod(dim); + } +}; + +struct ProdGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + dx->device(place) = dy->broadcast(dim) * y->broadcast(dim) * x->inverse(); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_sum_op.cc b/paddle/fluid/operators/reduce_sum_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f0e5f6580fbc9e70562cb2fdd7e0c5d8729bc9a7 --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.cc @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_sum_op.h" + +REGISTER_REDUCE_OP(reduce_sum); +REGISTER_OP_CPU_KERNEL( + reduce_sum, ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL( + reduce_sum_grad, + ops::ReduceSumGradKernel, + ops::ReduceSumGradKernel, + ops::ReduceSumGradKernel, + ops::ReduceSumGradKernel); diff --git a/paddle/fluid/operators/reduce_sum_op.cu b/paddle/fluid/operators/reduce_sum_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..f2e16955a50dc6a7feda9fbaf968c929ef3d8a4f --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_sum_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_sum, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_sum_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_sum_op.h b/paddle/fluid/operators/reduce_sum_op.h new file mode 100644 index 0000000000000000000000000000000000000000..3e8d1bbdba504669bc06e0637094e3bee840adf2 --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.h @@ -0,0 +1,97 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +// use for loop to speed up Eigen broadcast. 4 timer faster then broadcast +template +class ReduceSumGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto dims = context.Attr>("dim"); + if (context.GetPlace().type() == typeid(platform::CPUPlace) && + dims.size() == 1) { + auto* input0 = context.Input("X"); + auto* input2 = context.Input(framework::GradVarName("Out")); + auto* output = context.Output(framework::GradVarName("X")); + output->mutable_data(context.GetPlace()); + const auto* input2_d = input2->data(); + auto* output_d = output->data(); + + // handle reduce_all + if (input2->dims().size() == 1 && input2->dims()[0] == 1) { + for (int64_t i = 0; i < framework::product(input0->dims()); ++i) { + output_d[i] = input2_d[0]; + } + return; + } + + // handle reduce by one dimension + int reduce_dim_index = dims[0]; + if (reduce_dim_index < 0) { + reduce_dim_index += input0->dims().size(); + } + + auto& input_dim = input0->dims(); + int64_t before_dim = 1; + for (int i = 0; i < reduce_dim_index; ++i) { + before_dim *= input_dim[i]; + } + int64_t reduce_dim = input_dim[reduce_dim_index]; + int64_t after_dim = 1; + for (int i = reduce_dim_index + 1; i < input_dim.size(); ++i) { + after_dim *= input_dim[i]; + } + for (int64_t i = 0; i < before_dim; ++i) { + for (int64_t j = 0; j < reduce_dim; ++j) { + for (int64_t k = 0; k < after_dim; ++k) { + output_d[i * reduce_dim * after_dim + j * after_dim + k] = + input2_d[i * after_dim + k]; + } + } + } + return; + } + + // default use Eigen broadcast + ReduceGradKernel kernel; + kernel.Compute(context); + } +}; + +struct SumFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->sum(dim); + } +}; + +struct SumGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + dx->device(place) = dy->eval().broadcast(dim); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index 7f743f577fbcdaf6f62e01031e25ef09a842c2e9..a1dfe39c3a4f84f5e4aaa2306813a7decf0e49ea 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -12,14 +12,108 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/reshape_op.h" - #include #include +#include "paddle/fluid/framework/op_registry.h" namespace paddle { namespace operators { +class ReshapeOp : public framework::OperatorWithKernel { + public: + ReshapeOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReshapeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReshapeOp should not be null."); + + const std::vector &shape = ctx->Attrs().Get>("shape"); + PADDLE_ENFORCE(!shape.empty(), + "The shape information must be set by Attr(shape)."); + + if (ctx->HasInput("Shape") && ctx->IsRuntime()) { + // If true, set the shape of Output(Out) according to Input(Shape) in + // ReshapeKernel with ExecutionContext. Also check LoD in ReshapeKernel. + ctx->ShareLoD("X", /*->*/ "Out"); + return; + } + + auto x_dims = ctx->GetInputDim("X"); + auto out_dims = ValidateShape(shape, x_dims); + ctx->SetOutputDim("Out", out_dims); + if (x_dims[0] == out_dims[0]) { + // Only pass LoD when the first dimension of output and Input(X) + // are the same. + ctx->ShareLoD("X", /*->*/ "Out"); + } + } + + static framework::DDim ValidateShape(const std::vector shape, + const framework::DDim &in_dims) { + const int64_t in_size = framework::product(in_dims); + // only one dimension can be set to -1, whose size will be automatically + // infered. + const int64_t unk_dim_val = -1; + const int64_t copy_dim_val = 0; + + std::vector output_shape(shape.size(), 0); + int64_t capacity = 1; + int unk_dim_idx = -1; + for (size_t i = 0; i < shape.size(); ++i) { + if (shape[i] == unk_dim_val) { + PADDLE_ENFORCE( + unk_dim_idx == -1, + "Only one input dimension of Attr(shape) can be unknown."); + unk_dim_idx = i; + } else if (shape[i] == copy_dim_val) { + PADDLE_ENFORCE( + static_cast(i) < in_dims.size(), + "The index of dimension to copy from input shape must be less " + "than the size of input shape."); + } else { + PADDLE_ENFORCE( + shape[i] > 0, + "Each input dimension of Attr(shape) must not be negtive except " + "one unknown dimension."); + } + + capacity *= (shape[i] ? shape[i] : in_dims[i]); + output_shape[i] = + (shape[i] ? static_cast(shape[i]) : in_dims[i]); + } + + if (unk_dim_idx != -1) { + if (in_size > 0) { + // in_size < 0 and is un-determinate in compile time, skip the check, + // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], + // capacity = -24, in_size = -8, output_shape[0] = 0 + // the following check will fail. + output_shape[unk_dim_idx] = -in_size / capacity; + PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, + "Invalid shape is given."); + } else { + output_shape[unk_dim_idx] = -1; + } + } else { + PADDLE_ENFORCE_EQ(capacity, in_size, "Invalid shape is given."); + } + return framework::make_ddim(output_shape); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -33,12 +127,6 @@ class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "(Tensor). The output tensor of reshape operator."); AddAttr>( "shape", "(std::vector) Target shape of reshape operator."); - AddAttr("inplace", - "(default: false) Change the source tensor's shape without " - "memory copy. When Attr(inplace) is set true, the output " - "tensor shares memory with Input(X), otherwise, a new output " - "tensor is created, and its data are copied from Input(x).") - .SetDefault(false); AddComment(R"DOC( Reshape Operator. @@ -107,19 +195,78 @@ class ReshapeGradOp : public framework::OperatorWithKernel { } }; +class ReshapeKernel { + public: + void operator()(const framework::ExecutionContext &ctx) const { + auto *out = ctx.Output("Out"); + auto *in = ctx.Input("X"); + + auto *shape_tensor = ctx.HasInput("Shape") + ? ctx.Input("Shape") + : nullptr; + + framework::DDim out_dims = out->dims(); + + if (shape_tensor) { + auto *shape_data = shape_tensor->data(); + framework::Tensor cpu_shape_tensor; + if (platform::is_gpu_place(shape_tensor->place())) { + TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor); + shape_data = cpu_shape_tensor.data(); + } + auto shape = + std::vector(shape_data, shape_data + shape_tensor->numel()); + out_dims = ReshapeOp::ValidateShape(shape, in->dims()); + } + if (!in->lod().empty()) { + PADDLE_ENFORCE_EQ( + out_dims[0], in->dims()[0], + "Reshape operator cannot reshape an input sequence batch " + "into an output sequence batch that has a different " + "number of time steps. Please consider using " + "sequence_reshape op."); + } + + out->mutable_data(ctx.GetPlace(), in->type()); + framework::TensorCopySync(*in, ctx.GetPlace(), out); + out->Resize(out_dims); + } +}; + +class ReshapeGradKernel { + public: + void operator()(const framework::ExecutionContext &ctx) const { + auto *d_out = ctx.Input(framework::GradVarName("Out")); + auto *d_x = ctx.Output(framework::GradVarName("X")); + auto in_dims = d_x->dims(); + + d_x->mutable_data(ctx.GetPlace(), d_out->type()); + framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x); + d_x->Resize(in_dims); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -using CPU = paddle::platform::CPUDeviceContext; REGISTER_OPERATOR(reshape, ops::ReshapeOp, ops::ReshapeOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(reshape_grad, ops::ReshapeGradOp); -REGISTER_OP_CPU_KERNEL(reshape, ops::ReshapeKernel, - ops::ReshapeKernel, - ops::ReshapeKernel, - ops::ReshapeKernel); -REGISTER_OP_CPU_KERNEL(reshape_grad, ops::ReshapeGradKernel, - ops::ReshapeGradKernel, - ops::ReshapeGradKernel, - ops::ReshapeGradKernel); +REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape, float, ops::ReshapeKernel, double, + ops::ReshapeKernel, int, ops::ReshapeKernel, + int64_t, ops::ReshapeKernel); +REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape_grad, float, ops::ReshapeGradKernel, + double, ops::ReshapeGradKernel, int, + ops::ReshapeGradKernel, int64_t, + ops::ReshapeGradKernel); + +#ifdef PADDLE_WITH_CUDA +REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape, float, ops::ReshapeKernel, double, + ops::ReshapeKernel, int, ops::ReshapeKernel, + int64_t, ops::ReshapeKernel); +REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape_grad, float, ops::ReshapeGradKernel, + double, ops::ReshapeGradKernel, int, + ops::ReshapeGradKernel, int64_t, + ops::ReshapeGradKernel); +#endif diff --git a/paddle/fluid/operators/reshape_op.cu b/paddle/fluid/operators/reshape_op.cu deleted file mode 100644 index c628c634e2bc9ae260948a6e7ccf786cbd6c5c3c..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/reshape_op.cu +++ /dev/null @@ -1,26 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/reshape_op.h" -using CUDA = paddle::platform::CUDADeviceContext; - -REGISTER_OP_CUDA_KERNEL(reshape, paddle::operators::ReshapeKernel, - paddle::operators::ReshapeKernel, - paddle::operators::ReshapeKernel, - paddle::operators::ReshapeKernel); -REGISTER_OP_CUDA_KERNEL(reshape_grad, - paddle::operators::ReshapeGradKernel, - paddle::operators::ReshapeGradKernel, - paddle::operators::ReshapeGradKernel, - paddle::operators::ReshapeGradKernel); diff --git a/paddle/fluid/operators/reshape_op.h b/paddle/fluid/operators/reshape_op.h deleted file mode 100644 index 3dd8c7c11eca241e747bfa129962032d882ce44c..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/reshape_op.h +++ /dev/null @@ -1,189 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include - -#include "paddle/fluid/framework/eigen.h" -#include "paddle/fluid/framework/op_registry.h" - -namespace paddle { -namespace operators { - -class ReshapeOp : public framework::OperatorWithKernel { - public: - ReshapeOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : OperatorWithKernel(type, inputs, outputs, attrs) {} - - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of ReshapeOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of ReshapeOp should not be null."); - - const std::vector &shape = ctx->Attrs().Get>("shape"); - PADDLE_ENFORCE(!shape.empty(), - "The shape information must be set by Attr(shape)."); - - if (ctx->HasInput("Shape") && ctx->IsRuntime()) { - // If true, set the shape of Output(Out) according to Input(Shape) in - // ReshapeKernel with ExecutionContext. Also check LoD in ReshapeKernel. - ctx->ShareLoD("X", /*->*/ "Out"); - return; - } - - auto x_dims = ctx->GetInputDim("X"); - auto out_dims = ValidateShape(shape, x_dims); - ctx->SetOutputDim("Out", out_dims); - if (x_dims[0] == out_dims[0]) { - // Only pass LoD when the first dimension of output and Input(X) - // are the same. - ctx->ShareLoD("X", /*->*/ "Out"); - } - } - - static framework::DDim ValidateShape(const std::vector shape, - const framework::DDim &in_dims) { - const int64_t in_size = framework::product(in_dims); - // only one dimension can be set to -1, whose size will be automatically - // infered. - const int64_t unk_dim_val = -1; - const int64_t copy_dim_val = 0; - - std::vector output_shape(shape.size(), 0); - int64_t capacity = 1; - int unk_dim_idx = -1; - for (size_t i = 0; i < shape.size(); ++i) { - if (shape[i] == unk_dim_val) { - PADDLE_ENFORCE( - unk_dim_idx == -1, - "Only one input dimension of Attr(shape) can be unknown."); - unk_dim_idx = i; - } else if (shape[i] == copy_dim_val) { - PADDLE_ENFORCE( - static_cast(i) < in_dims.size(), - "The index of dimension to copy from input shape must be less " - "than the size of input shape."); - } else { - PADDLE_ENFORCE( - shape[i] > 0, - "Each input dimension of Attr(shape) must not be negtive except " - "one unknown dimension."); - } - - capacity *= (shape[i] ? shape[i] : in_dims[i]); - output_shape[i] = - (shape[i] ? static_cast(shape[i]) : in_dims[i]); - } - - if (unk_dim_idx != -1) { - if (in_size > 0) { - // in_size < 0 and is un-determinate in compile time, skip the check, - // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], - // capacity = -24, in_size = -8, output_shape[0] = 0 - // the following check will fail. - output_shape[unk_dim_idx] = -in_size / capacity; - PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, - "Invalid shape is given."); - } else { - output_shape[unk_dim_idx] = -1; - } - } else { - PADDLE_ENFORCE_EQ(capacity, in_size, "Invalid shape is given."); - } - return framework::make_ddim(output_shape); - } - - protected: - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext &ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); - } -}; - -template -class ReshapeKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext &ctx) const { - auto *out = ctx.Output("Out"); - auto *in = ctx.Input("X"); - - auto *shape_tensor = ctx.HasInput("Shape") - ? ctx.Input("Shape") - : nullptr; - - framework::DDim out_dims = out->dims(); - - if (shape_tensor) { - auto *shape_data = shape_tensor->data(); - framework::Tensor cpu_shape_tensor; - if (platform::is_gpu_place(ctx.GetPlace())) { - TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor); - shape_data = cpu_shape_tensor.data(); - } - auto shape = - std::vector(shape_data, shape_data + shape_tensor->numel()); - out_dims = ReshapeOp::ValidateShape(shape, in->dims()); - } - if (!in->lod().empty()) { - PADDLE_ENFORCE_EQ( - out_dims[0], in->dims()[0], - "Reshape operator cannot reshape an input sequence batch " - "into an output sequence batch that has a different " - "number of time steps. Please consider using " - "sequence_reshape op."); - } - - bool inplace = ctx.Attr("inplace"); - out->Resize(out_dims); - if (!inplace) { - out->mutable_data(ctx.GetPlace()); - framework::TensorCopySync(*in, ctx.GetPlace(), out); - out->Resize(out_dims); - } else { - out->ShareDataWith(*in); - out->Resize(out_dims); - } - } -}; - -template -class ReshapeGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext &ctx) const { - auto *d_out = ctx.Input(framework::GradVarName("Out")); - auto *d_x = ctx.Output(framework::GradVarName("X")); - - d_x->mutable_data(ctx.GetPlace()); - bool inplace = ctx.Attr("inplace"); - - auto in_dims = d_x->dims(); - if (!inplace) { - framework::TensorCopy(*d_out, ctx.GetPlace(), ctx.device_context(), d_x); - ctx.device_context().Wait(); - d_x->Resize(in_dims); - } else { - d_x->ShareDataWith(*d_out); - d_x->Resize(in_dims); - } - } -}; -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/reverse_op.cc b/paddle/fluid/operators/reverse_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a20f7d231fa9ea313581ac0629a87fa5f4a88ce5 --- /dev/null +++ b/paddle/fluid/operators/reverse_op.cc @@ -0,0 +1,107 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reverse_op.h" +#include + +namespace paddle { +namespace operators { + +class ReverseOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + const auto& x_dims = ctx->GetInputDim("X"); + const auto& axis = ctx->Attrs().Get>("axis"); + PADDLE_ENFORCE(!axis.empty(), "'axis' can not be empty."); + for (int a : axis) { + PADDLE_ENFORCE_LT(a, x_dims.size(), + "The axis must be less than input tensor's rank."); + } + ctx->SetOutputDim("Out", x_dims); + } +}; + +class ReverseOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "The LoDTensor to be flipped."); + AddOutput("Out", "The LoDTensor after flipping."); + AddAttr>( + "axis", "The axises that along which order of elements is reversed."); + AddComment(R"DOC( + Reverse Operator. + + Reverse the order of elements in the input LoDTensor along given axises. + + Case 1: + Given + X = [[1, 2, 3, 4, 5] + [6, 7, 8, 9, 10] + [11, 12, 13, 14, 15]], + and + axis = [0], + we get: + Out = [[11, 12, 13, 14, 15] + [6, 7, 8, 9, 10] + [1, 2, 3, 4, 5]]. + + Case 2: + Given + X = [[[1, 2, 3, 4] + [5, 6, 7, 8]] + [[9, 10, 11, 12] + [13, 14, 15, 16]]], + and + axis = [0, 2], + we get: + Out = [[[12, 11, 10, 9] + [16, 15, 14, 13]] + [[4, 3, 2, 1] + [8, 7, 6, 5]]], + )DOC"); + } +}; + +class ReverseGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDesc(); + grad_op->SetType("reverse"); + grad_op->SetInput("X", OutputGrad("Out")); + grad_op->SetOutput("Out", InputGrad("X")); + grad_op->SetAttr("axis", GetAttr("axis")); + return std::unique_ptr(grad_op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(reverse, ops::ReverseOp, ops::ReverseOpMaker, + ops::ReverseGradMaker); +REGISTER_OPERATOR(reverse_grad, ops::ReverseOp); +REGISTER_OP_CPU_KERNEL( + reverse, ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel) diff --git a/paddle/fluid/operators/reverse_op.cu b/paddle/fluid/operators/reverse_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..635c41529b38f2dd287b00ed2e5659e11f619e78 --- /dev/null +++ b/paddle/fluid/operators/reverse_op.cu @@ -0,0 +1,24 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reverse_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + reverse, ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel) diff --git a/paddle/fluid/operators/reverse_op.h b/paddle/fluid/operators/reverse_op.h new file mode 100644 index 0000000000000000000000000000000000000000..9063cd59bba5c6307b55a500455908a5fd278390 --- /dev/null +++ b/paddle/fluid/operators/reverse_op.h @@ -0,0 +1,87 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +struct ReverseFunctor { + void operator()(const DeviceContext& context, const framework::LoDTensor& in, + framework::LoDTensor* out, const std::vector& axis) { + Eigen::array reverse_axis; + for (int i = 0; i < Rank; ++i) { + reverse_axis[i] = false; + } + for (int a : axis) { + reverse_axis[a] = true; + } + + auto in_eigen = framework::EigenTensor::From(in); + auto out_eigen = framework::EigenTensor::From(*out); + auto* dev = context.eigen_device(); + + out_eigen.device(*dev) = in_eigen.reverse(reverse_axis); + } +}; + +template +class ReverseKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + const auto& axis = context.Attr>("axis"); + int rank = x->dims().size(); + auto& dev_ctx = context.template device_context(); + + switch (rank) { + case 1: + ReverseFunctor functor1; + functor1(dev_ctx, *x, out, axis); + break; + case 2: + ReverseFunctor functor2; + functor2(dev_ctx, *x, out, axis); + break; + case 3: + ReverseFunctor functor3; + functor3(dev_ctx, *x, out, axis); + break; + case 4: + ReverseFunctor functor4; + functor4(dev_ctx, *x, out, axis); + break; + case 5: + ReverseFunctor functor5; + functor5(dev_ctx, *x, out, axis); + break; + case 6: + ReverseFunctor functor6; + functor6(dev_ctx, *x, out, axis); + break; + default: + PADDLE_THROW( + "Reserve operator doesn't supports tensors whose ranks are greater " + "than 6."); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/roi_pool_op.cc b/paddle/fluid/operators/roi_pool_op.cc index 293abb0ea4f1ac03c3889ce2937ef8fa0845db73..d6d209d5de041500a9b4893d70800a58e8ee1e1d 100644 --- a/paddle/fluid/operators/roi_pool_op.cc +++ b/paddle/fluid/operators/roi_pool_op.cc @@ -139,7 +139,20 @@ class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker { "The pooled output width.") .SetDefault(1); AddComment(R"DOC( -ROIPool operator +**ROIPool Operator** + +Region of interest pooling (also known as RoI pooling) is to perform +is to perform max pooling on inputs of nonuniform sizes to obtain +fixed-size feature maps (e.g. 7*7). + +The operator has three steps: + +1. Dividing each region proposal into equal-sized sections with + the pooled_width and pooled_height + +2. Finding the largest value in each section + +3. Copying these max values to the output buffer ROI Pooling for Faster-RCNN. The link below is a further introduction: https://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 20f140f962c3aac364a1239a663d5f340bbeb6b2..10b1b0c899d833d70fa6afe51998fe210899e3c3 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -78,23 +78,23 @@ class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "(LoDTensor), the input(X) is a LodTensor, which supports " + "the input(X) is a LodTensor, which supports " "variable time-length input sequences. The underlying tensor " "in this LoDTensor is a matrix with shape (T x N), where T " "is the total time steps in this mini-batch and N is the input " "data dimension."); AddInput("Filter", - "(Tensor), the input(Filter) is a learnable parameter. It " + "the input(Filter) is a learnable parameter. It " "is a 2-D tensor with shape (future_context x N), where, " "future_context is the future context length and N is the data " "dimension."); AddOutput("Out", - "(LoDTensor), the output(Out) is a LodTensor, which supports " + "the output(Out) is a LodTensor, which supports " "variable time-length input sequences. The underlying tensor " "in this LodTensor is a matrix with shape T x N, i.e., the " "same shape as X."); AddComment(R"DOC( -Row-convolution Operator. +:strong:`Row-convolution operator` The row convolution is called lookahead convolution. This operator was introduced in the following paper for DeepSpeech2: @@ -114,9 +114,23 @@ and a filter ($W$) of size $context \times d$, the output sequence is convolved as: $$ -out_{i, :} = \sum_{j=i}^{i + context} in_{j,:} \dot W_{i-j, :} +out_{i, :} = \\sum_{j=i}^{i + context} in_{j,:} \\cdot W_{i-j, :} $$ +In the above equation: + +* $Out_{i}$: The i-th row of output variable with shape [1, D]. + +* $\\tau$: Future context size. + +* $X_{j}$: The j-th row of input variable with shape [1, D]. + +* $W_{i-j}$: The (i-j)-th row of parameters with shape [1, D]. + +More details about row_conv please refer to +the design document +https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645 . + )DOC"); } }; diff --git a/paddle/fluid/operators/save_load_op_test.cc b/paddle/fluid/operators/save_load_op_test.cc index c4fcc61af4b75e6dc7d5c31e20c5fff358637af5..ccaea0eef2906953d922e097348b6c0a86dad6f1 100644 --- a/paddle/fluid/operators/save_load_op_test.cc +++ b/paddle/fluid/operators/save_load_op_test.cc @@ -139,6 +139,7 @@ TEST(LoadFP16Op, CPU) { save_op->Run(scope, place); auto load_var = scope.Var("out_var"); + load_var->GetMutable(); auto load_op = paddle::framework::OpRegistry::CreateOp( "load", {}, {{"Out", {"out_var"}}}, attrs); load_op->Run(scope, place); diff --git a/paddle/fluid/operators/save_op.cc b/paddle/fluid/operators/save_op.cc index e6d27e2dedd7668b93bd8ddc330a897d1c6fa732..201a51130d6b6f94104e2dabf9e7facffa672ae0 100644 --- a/paddle/fluid/operators/save_op.cc +++ b/paddle/fluid/operators/save_op.cc @@ -22,11 +22,17 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/variable.h" #include "paddle/fluid/platform/device_context.h" namespace paddle { namespace operators { +// define LOOKUP_TABLE_PATH for checkpoint notify to save lookup table variables +// to directory specified. +constexpr char LOOKUP_TABLE_PATH[] = "kLookupTablePath"; + // TODO(yuyang18): If the functions below are needed by other files, move them // to paddle::filesystem namespace. constexpr char kSEP = '/'; @@ -67,9 +73,27 @@ class SaveOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { + auto iname = Input("X"); + auto *var = scope.FindVar(iname); + PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s for save_op", + iname); + + if (var->IsType()) { + SaveLodTensor(place, var); + } else if (var->IsType()) { + SaveSelectedRows(scope, place, var); + } else { + PADDLE_ENFORCE( + false, + "SaveOp only support LoDTensor and SelectedRows, %s has wrong type", + iname); + } + } + + void SaveLodTensor(const platform::Place &place, + framework::Variable *var) const { auto filename = Attr("file_path"); auto overwrite = Attr("overwrite"); - auto save_as_fp16 = Attr("save_as_fp16"); if (FileExists(filename) && !overwrite) { PADDLE_THROW("%s is existed, cannot save to it when overwrite=false", @@ -78,26 +102,19 @@ class SaveOp : public framework::OperatorBase { MkDirRecursively(DirName(filename).c_str()); - // FIXME(yuyang18): We save variable to local file now, but we should change - // it to save an output stream. - std::ofstream fout(filename); - PADDLE_ENFORCE(static_cast(fout), "Cannot open %s to write", - filename); - - auto iname = Input("X"); - auto *var = scope.FindVar(iname); - PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s for save_op", - iname); - - PADDLE_ENFORCE(var->IsType(), - "SaveOp only support LoDTensor, %s has wrong type", iname); - auto &tensor = var->Get(); // get device context from pool platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(place); + // FIXME(yuyang18): We save variable to local file now, but we should change + // it to save an output stream. + std::ofstream fout(filename); + PADDLE_ENFORCE(static_cast(fout), "Cannot open %s to write", + filename); + + auto save_as_fp16 = Attr("save_as_fp16"); auto in_dtype = framework::ToDataType(tensor.type()); auto out_dtype = save_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; @@ -112,17 +129,43 @@ class SaveOp : public framework::OperatorBase { } else { framework::SerializeToStream(fout, tensor, dev_ctx); } + fout.close(); + } + + void SaveSelectedRows(const framework::Scope &scope, + const platform::Place &place, + framework::Variable *var) const { + auto *lt_var = scope.FindVar(LOOKUP_TABLE_PATH)->GetMutable(); + PADDLE_ENFORCE( + lt_var != nullptr, + "Can not find variable kLookupTablePath for SaveSelectedRows"); + std::string filename = lt_var->data(); + VLOG(4) << "SaveSelectedRows get File name: " << filename; + + auto &selectedRows = var->Get(); + + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + // FIXME(yuyang18): We save variable to local file now, but we should change + // it to save an output stream. + std::ofstream fout(filename); + PADDLE_ENFORCE(static_cast(fout), "Cannot open %s to write", + filename); + framework::SerializeToStream(fout, selectedRows, dev_ctx); + fout.close(); } }; class SaveOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", "(Tensor ) Input tensor to be saved"); + AddInput("X", "(Tensor ) Input LoDTensor and SelectedRows to be saved"); AddComment(R"DOC( Save operator -This operator will serialize and write a tensor variable to file on disk. +This operator will serialize and write LoDTensor / SelectedRows variable to file on disk. )DOC"); AddAttr("overwrite", "(boolean, default true)" @@ -142,9 +185,26 @@ This operator will serialize and write a tensor variable to file on disk. } }; +class SaveOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto out_var_name = op_desc.Output(LOOKUP_TABLE_PATH).front(); + auto &out_var = block->FindRecursiveOrCreateVar(out_var_name); + auto var_type = framework::proto::VarType::RAW; + out_var.SetType(var_type); + } +}; + +class SaveOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override {} +}; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(save, ops::SaveOp, ops::SaveOpProtoMaker); +REGISTER_OPERATOR(save, ops::SaveOp, paddle::framework::EmptyGradOpMaker, + ops::SaveOpProtoMaker, ops::SaveOpVarTypeInference, + ops::SaveOpShapeInference); diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc index 4687e21e7155fc7309fb28c881c0d47152df9ad5..7f8822e40053b5bcd394f446138a2292d80b69bf 100644 --- a/paddle/fluid/operators/scale_op.cc +++ b/paddle/fluid/operators/scale_op.cc @@ -41,13 +41,13 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("X", "(Tensor) Input tensor of scale operator."); AddOutput("Out", "(Tensor) Output tensor of scale operator."); AddComment(R"DOC( -Scale operator +**Scale operator** + +Multiply the input tensor with a float scalar to scale the input tensor. $$Out = scale*X$$ )DOC"); - AddAttr("scale", - "(float, default 1.0)" - "The scaling factor of the scale operator.") + AddAttr("scale", "The scaling factor of the scale operator.") .SetDefault(1.0); } }; diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index bcd8e81609a37cc544f5a5cc4188400c1632a668..1866a86048acbefadcb4d82cd6309cd16f0352d6 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -19,8 +19,8 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/macros.h" -#include "paddle/fluid/operators/detail/grpc_client.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -39,23 +39,19 @@ class SendBarrierOp : public framework::OperatorBase { std::vector eps = Attr>("endpoints"); bool sync_mode = Attr("sync_mode"); - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - // For profiling - platform::RecordEvent record_event(Type(), &ctx); - - auto rpc_client = detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode; // need to wait before sending send_barrier message - PADDLE_ENFORCE(rpc_client->Wait()); + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); if (sync_mode) { for (auto& ep : eps) { VLOG(3) << "send barrier, ep: " << ep; rpc_client->AsyncSendBatchBarrier(ep); } - PADDLE_ENFORCE(rpc_client->Wait()); + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } } }; diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index a5150f242ca3b0befafa2443f0bc466e2aea85e4..3cd42f2d059532b7090e66ce21de8e5cb014adf1 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -16,10 +16,9 @@ limitations under the License. */ #include #include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/operators/send_recv_util.h" #include "paddle/fluid/platform/profiler.h" @@ -36,51 +35,28 @@ class SendOp : public framework::OperatorBase { void RunImpl(const framework::Scope& scope, const platform::Place& place) const override { auto ins = Inputs("X"); - auto outs = Outputs("Out"); - std::vector epmap = Attr>("epmap"); - std::vector endpoints = - Attr>("endpoints"); - bool sync_mode = Attr("sync_mode"); + std::vector epmap = Attr>("epmap"); + int sync_send = Attr("sync_mode"); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - // For profiling - platform::RecordEvent record_event(Type(), &ctx); - - auto rpc_client = detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; - rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]); + // TODO(Yancey1989): we need to use an IO threadpool which has + // a larger number of threads than the computing threadpool. + rpc_client->AsyncSendVar(epmap[i], ctx, scope, ins[i]); } else { VLOG(3) << "don't send no-initialied variable: " << ins[i]; } } - PADDLE_ENFORCE(rpc_client->Wait()); - - if (sync_mode) { - for (auto& ep : endpoints) { - VLOG(3) << "batch barrier, ep: " << ep; - rpc_client->AsyncSendBatchBarrier(ep); - } - PADDLE_ENFORCE(rpc_client->Wait()); - } - - if (outs.size() > 0) { - for (size_t i = 0; i < outs.size(); i++) { - VLOG(2) << "getting " << outs[i] << " from " << epmap[i]; - rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); - } - PADDLE_ENFORCE(rpc_client->Wait()); - // tell pservers that current trainer have called fetch - for (auto& ep : endpoints) { - VLOG(2) << "send fetch barrier, ep: " << ep; - rpc_client->AsyncSendFetchBarrier(ep); - } - PADDLE_ENFORCE(rpc_client->Wait()); + if (sync_send) { + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } } }; @@ -88,26 +64,22 @@ class SendOp : public framework::OperatorBase { class SendOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() { - AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable(); - AddOutput("Out", "(Tensor) Output tensor to be received from server") + AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") .AsDuplicable(); AddComment(R"DOC( Send operator -This operator will send tensor to recv_op at the parameter server. +This operator will send variables to listen_and_serve op at the parameter server. )DOC"); - // TODO(typhoonzero): remove this attr generate de-duplicated vector from - // epmap when initializing. - AddAttr>("endpoints", - "(string vector, default 127.0.0.1:6164)" - "Server endpoints to send variables to.") - .SetDefault({}); + AddAttr("sync_mode", + "(int, default 0)" + "sync send or async send.") + .SetDefault(0); AddAttr>("epmap", "(string vector, default 127.0.0.1:6164)" "Server endpoints in the order of input " "variables for mapping") - .SetDefault({}); - AddAttr("sync_mode", "work in sync_mode or not").SetDefault(true); + .SetDefault({"127.0.0.1:6164"}); } }; diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index e550552b195b768d68ec64e9c3b5889b56ca719f..aee6180add5708d31f7ce927b37c4524a291fe3c 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -129,7 +129,10 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { // sub program run in listen_and_serv_op, for simple test we use sum f::ProgramDesc program; const auto &root_block = program.Block(0); + std::vector optimize_blocks; auto *optimize_block = program.AppendBlock(root_block); + optimize_blocks.push_back(optimize_block); + auto *prefetch_block = program.AppendBlock(root_block); // X for server side tensors, RX for received tensors, must be of same shape. AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block, @@ -139,7 +142,7 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { attrs.insert({"Fanin", 1}); attrs.insert({"ParamList", std::vector({"Out"})}); attrs.insert({"GradList", std::vector({"x1"})}); - attrs.insert({"OptimizeBlock", optimize_block}); + attrs.insert({"optimize_blocks", optimize_blocks}); attrs.insert({"PrefetchBlock", prefetch_block}); attrs.insert({"grad_to_block_id", std::vector({""})}); attrs.insert({"sync_mode", true}); diff --git a/paddle/fluid/operators/send_recv_util.h b/paddle/fluid/operators/send_recv_util.h index deab005149027caffa962783df944fad7110382f..dc26c53c64f06ce21856fb5af8f2a5eb3fc75bb7 100644 --- a/paddle/fluid/operators/send_recv_util.h +++ b/paddle/fluid/operators/send_recv_util.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include +#include "paddle/fluid/framework/ir/node.h" namespace paddle { namespace operators { @@ -22,7 +23,10 @@ inline bool NeedSend(const framework::Scope& scope, const std::string& varname) { // dummy variable is only used in parallel executor to represent // some dependency relationship, we don't need to send/recv it. - if (varname == "dummy") return false; + // TODO(paddle-dev): Why would parallel executor logic leaked into here? + if (varname.find(framework::ir::Node::kControlDepVarName) != + std::string::npos) + return false; auto* var = scope.FindVar(varname); PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.", varname); diff --git a/paddle/fluid/operators/send_vars_op.cc b/paddle/fluid/operators/send_vars_op.cc deleted file mode 100644 index fe839dab6924618c8a4c39868d9bf86056a0be40..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/send_vars_op.cc +++ /dev/null @@ -1,100 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include // NOLINT -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/send_recv_util.h" -#include "paddle/fluid/platform/profiler.h" - -namespace paddle { -namespace operators { - -class SendVarsOp : public framework::OperatorBase { - public: - SendVarsOp(const std::string& type, const framework::VariableNameMap& inputs, - const framework::VariableNameMap& outputs, - const framework::AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - void RunImpl(const framework::Scope& scope, - const platform::Place& place) const override { - auto ins = Inputs("X"); - - std::vector epmap = Attr>("epmap"); - int sync_send = Attr("sync_send"); - - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - - // For profiling - platform::RecordEvent record_event(Type(), &ctx); - - auto rpc_client = detail::RPCClient::GetInstance(); - - for (size_t i = 0; i < ins.size(); i++) { - if (NeedSend(scope, ins[i])) { - VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; - // TODO(Yancey1989): we need to use an IO threadpool which has - // a larger number of threads than the computing threadpool. - rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]); - } else { - VLOG(3) << "don't send no-initialied variable: " << ins[i]; - } - } - if (sync_send) { - rpc_client->Wait(); - } - } -}; - -class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") - .AsDuplicable(); - AddComment(R"DOC( -Send operator - -This operator will send variables to listen_and_serve op at the parameter server. -)DOC"); - AddAttr("sync_send", - "(int, default 0)" - "sync send or async send.") - .SetDefault(0); - AddAttr>("epmap", - "(string vector, default 127.0.0.1:6164)" - "Server endpoints in the order of input " - "variables for mapping") - .SetDefault({"127.0.0.1:6164"}); - } -}; - -class SendVarsOpShapeInference : public framework::InferShapeBase { - public: - void operator()(framework::InferShapeContext* ctx) const override {} -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; - -REGISTER_OPERATOR(send_vars, ops::SendVarsOp, - paddle::framework::EmptyGradOpMaker, ops::SendVarsOpMaker, - ops::SendVarsOpShapeInference); diff --git a/paddle/fluid/operators/sequence_expand_op.h b/paddle/fluid/operators/sequence_expand_op.h index d62c387c3eebf9df0ab532f4e891da006f239468..39301e1ac0971dfe0ca7854257f10ddeb60f1000 100644 --- a/paddle/fluid/operators/sequence_expand_op.h +++ b/paddle/fluid/operators/sequence_expand_op.h @@ -151,9 +151,6 @@ struct SequenceExpandGradFunctor { const framework::Vector& x_lod, /*expand source lod*/ const framework::Vector& ref_lod, /*expand referenced lod*/ LoDTensor* dx) { - math::SetConstant set_zero; - set_zero(context, dx, static_cast(0)); - int dout_offset = 0; for (size_t i = 1; i < ref_lod.size(); ++i) { int repeat_num = ref_lod[i] - ref_lod[i - 1]; @@ -187,6 +184,10 @@ class SequenceExpandGradKernel : public framework::OpKernel { g_x->mutable_data(context.GetPlace()); g_x->set_lod(x->lod()); + auto& dev_ctx = context.template device_context(); + math::SetConstant set_zero; + set_zero(dev_ctx, g_x, static_cast(0)); + auto& y_lod = y->lod(); if (ref_level == -1) ref_level = y_lod.size() - 1; // just copy the gradient diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index 7a2bdeac09d61603f437ff10d58d0542bb3c3689..fef230e42d07a5ed73b7a7a6ab682694675bb9d2 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -74,7 +74,8 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Grad", "(Tensor or SelectedRows) Input gradient"); AddOutput("ParamOut", "(Tensor or SelectedRows, same with Param) " - "Output parameter, should share the same memory with Param"); + "Output parameter, should share the same memory with Param") + .Reuse("Param"); AddComment(R"DOC( SGD operator diff --git a/paddle/fluid/operators/sgd_op.h b/paddle/fluid/operators/sgd_op.h index f9e0596191d0b86686e0fa36265806111c774b38..2685ce217ee0f0d3e89f3751e96218dcd19bead4 100644 --- a/paddle/fluid/operators/sgd_op.h +++ b/paddle/fluid/operators/sgd_op.h @@ -114,7 +114,7 @@ class SGDOpKernel : public framework::OpKernel { int64_t id_index = param.Index(grad.rows()[i]); PADDLE_ENFORCE_GE(id_index, static_cast(0), "id should be in the table"); - for (size_t j = 0; j < grad_row_width; j++) { + for (int64_t j = 0; j < grad_row_width; j++) { out_data[id_index * grad_row_width + j] -= lr[0] * grad_data[i * grad_row_width + j]; } diff --git a/paddle/fluid/operators/shape_op.cc b/paddle/fluid/operators/shape_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..1be9fe47af71d31ce2e0eba807ea4a43601f8aca --- /dev/null +++ b/paddle/fluid/operators/shape_op.cc @@ -0,0 +1,57 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/shape_op.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class ShapeOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input (Input) of get_shape op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output (Out) of get_shape op should not be null."); + auto in_dim = ctx->GetInputDim("Input"); + ctx->SetOutputDim("Out", {in_dim.size()}); + } +}; + +class ShapeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", "(Tensor), The input tensor."); + AddOutput("Out", + "(Tensor), The shape of input tensor, the data type of the shape" + " is int32_t, will be on the same device with the input Tensor."); + AddComment(R"DOC( +Shape Operator + +Get the shape of input tensor. Only support CPU input Tensor now. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(shape, ops::ShapeOp, ops::ShapeOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(shape, ops::ShapeKernel, ops::ShapeKernel, + ops::ShapeKernel, ops::ShapeKernel); diff --git a/paddle/fluid/operators/shape_op.cu b/paddle/fluid/operators/shape_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..d8fa9515abf807ab4ae3c47e8e1b1cf7f30440a8 --- /dev/null +++ b/paddle/fluid/operators/shape_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/shape_op.h" + +REGISTER_OP_CUDA_KERNEL(shape, paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel); diff --git a/paddle/fluid/operators/shape_op.h b/paddle/fluid/operators/shape_op.h new file mode 100644 index 0000000000000000000000000000000000000000..0d510a505583c55e26a26bfc6e5d6192899b3d9e --- /dev/null +++ b/paddle/fluid/operators/shape_op.h @@ -0,0 +1,38 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ShapeKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_t = ctx.Input("Input"); + auto* out_t = ctx.Output("Out"); + auto out_data = out_t->mutable_data(platform::CPUPlace()); + auto in_dims = in_t->dims(); + for (int i = 0; i < in_dims.size(); ++i) { + out_data[i] = in_dims[i]; + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc index 135e2a6f7f877c9ef159a4542b834d5627649e81..c3b0fe32098cb4b41ccc155db58809ef9f1bf46b 100644 --- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc +++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -113,14 +113,14 @@ The logistic loss is given as follows: $$loss = -Labels * \log(\sigma(X)) - (1 - Labels) * \log(1 - \sigma(X))$$ -We know that $$\sigma(X) = (1 / (1 + \exp(-X)))$$. By substituting this we get: +We know that $$\sigma(X) = \\frac{1}{1 + \exp(-X)}$$. By substituting this we get: $$loss = X - X * Labels + \log(1 + \exp(-X))$$ For stability and to prevent overflow of $$\exp(-X)$$ when X < 0, we reformulate the loss as follows: - $$loss = \max(X, 0) - X * Labels + \log(1 + \exp(-|X|))$$ + $$loss = \max(X, 0) - X * Labels + \log(1 + \exp(-\|X\|))$$ Both the input `X` and `Labels` can carry the LoD (Level of Details) information. However the output only shares the LoD with input `X`. diff --git a/paddle/fluid/operators/slice_op.cc b/paddle/fluid/operators/slice_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..4bd23d594134f227e86b01fd75b7e202dd76c11b --- /dev/null +++ b/paddle/fluid/operators/slice_op.cc @@ -0,0 +1,133 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/slice_op.h" +#include +#include + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class SliceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input (Input) of slice op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output (Out) of slice op should not be null."); + + auto in_dims = ctx->GetInputDim("Input"); + PADDLE_ENFORCE(in_dims.size() < 7, + "The rank of input should be less than 7."); + framework::DDim out_dims(in_dims); + auto axes = ctx->Attrs().Get>("axes"); + auto starts = ctx->Attrs().Get>("starts"); + auto ends = ctx->Attrs().Get>("ends"); + + PADDLE_ENFORCE_EQ(starts.size(), ends.size()); + PADDLE_ENFORCE_EQ(starts.size(), axes.size()); + int dim_value, start, end; + for (size_t i = 0; i < axes.size(); ++i) { + dim_value = out_dims[axes[i]]; + start = starts[i] < 0 ? (starts[i] + dim_value) : starts[i]; + end = ends[i] < 0 ? (ends[i] + dim_value) : ends[i]; + start = std::max(start, 0); + end = std::max(end, 0); + start = std::min(start, dim_value); + end = std::min(end, dim_value); + start = std::min(start, end); + out_dims[axes[i]] = end - start; + } + ctx->SetOutputDim("Out", out_dims); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Input")->type()), + ctx.GetPlace()); + } +}; + +class SliceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", "Tensor of data to extract slices from."); + AddOutput("Out", "Sliced data tensor."); + + AddAttr>( + "axes", + "(list) Axes that `starts` and `ends` apply to. It's optional." + "If not present, will be treated as [0, 1, ..., len(`starts`) - 1]."); + AddAttr>( + "starts", + "(list) Starting indices of corresponding axis in `axes`"); + AddAttr>( + "ends", + "(list) Starting indices of corresponding axis in `axes`."); + + AddComment(R"DOC( +Slice Operator. + +Produces a slice of the input tensor along multiple axes. Similar to numpy: +https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html +Slice uses `axes`, `starts` and `ends` attributes to specify the start and +end dimension for each axis in the list of axes, it uses this information +to slice the input data tensor. If a negative value is passed for any of +the start or end indices, it represents number of elements before the end +of that dimension. If the value passed to start or end is larger than +the n (the number of elements in this dimension), it represents n. +For slicing to the end of a dimension with unknown size, it is recommended +to pass in INT_MAX. If axes are omitted, they are set to [0, ..., ndim-1]. +Following examples will explain how slice works: + + .. code-block:: text + + Cast1: + Given: + data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] + axes = [0, 1] + starts = [1, 0] + ends = [2, 3] + Then: + result = [ [5, 6, 7], ] + + Cast2: + Given: + data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] + starts = [0, 1] + ends = [-1, 1000] + Then: + result = [ [2, 3, 4], ] +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(slice, ops::SliceOp, ops::SliceOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL( + slice, ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel); diff --git a/paddle/fluid/operators/slice_op.cu b/paddle/fluid/operators/slice_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..8c1767c70b19d1386af9610ef3405eb487a39878 --- /dev/null +++ b/paddle/fluid/operators/slice_op.cu @@ -0,0 +1,22 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/slice_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + slice, ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel); diff --git a/paddle/fluid/operators/slice_op.h b/paddle/fluid/operators/slice_op.h new file mode 100644 index 0000000000000000000000000000000000000000..ba231aee176564b91a642912ce0b32bcdef8cfc1 --- /dev/null +++ b/paddle/fluid/operators/slice_op.h @@ -0,0 +1,88 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class SliceKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + int rank = ctx.Input("Input")->dims().size(); + switch (rank) { + case 1: + SliceCompute<1>(ctx); + break; + case 2: + SliceCompute<2>(ctx); + break; + case 3: + SliceCompute<3>(ctx); + break; + case 4: + SliceCompute<4>(ctx); + break; + case 5: + SliceCompute<5>(ctx); + break; + case 6: + SliceCompute<6>(ctx); + break; + } + } + + private: + template + void SliceCompute(const framework::ExecutionContext& context) const { + auto& place = + *context.template device_context().eigen_device(); + auto in = context.Input("Input"); + auto out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + auto out_dims = out->dims(); + auto in_dims = in->dims(); + auto axes = context.Attr>("axes"); + auto starts = context.Attr>("starts"); + + auto offsets = Eigen::array(); + auto extents = Eigen::array(); + for (size_t i = 0; i < D; ++i) { + offsets[i] = 0; + extents[i] = out_dims[i]; + } + int start; + for (size_t i = 0; i < axes.size(); ++i) { + start = starts[i]; + if (start < 0) { + start = (start + in_dims[axes[i]]); + } + start = std::max(start, 0); + offsets[axes[i]] = start; + } + auto in_t = + framework::EigenTensor::From( + *in); + auto out_t = + framework::EigenTensor::From( + *out); + out_t.device(place) = in_t.slice(offsets, extents); + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc index 5596fa0648ccc151bc0d11de9c556599428a8d71..2bdb23e999621b10799b5163f326bc4b66a437e6 100644 --- a/paddle/fluid/operators/softmax_cudnn_op.cu.cc +++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc @@ -30,8 +30,16 @@ class SoftmaxCUDNNKernel : public framework::OpKernel { // allocate memory on device. Out->mutable_data(context.GetPlace()); + auto dims = X->dims(); + auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1); + framework::LoDTensor flattened_x; + framework::LoDTensor flattened_out; + flattened_x.ShareDataWith(*X).Resize(flattened_dims); + flattened_out.ShareDataWith(*Out).Resize(flattened_dims); + math::SoftmaxCUDNNFunctor()( - context.template device_context(), X, Out); + context.template device_context(), + &flattened_x, &flattened_out); } }; @@ -46,9 +54,18 @@ class SoftmaxGradCUDNNKernel : public framework::OpKernel { // allocate memory on device. dX->mutable_data(context.GetPlace()); + auto dims = Out->dims(); + auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1); + framework::LoDTensor flattened_out; + framework::LoDTensor flattened_d_out; + framework::LoDTensor flattened_d_x; + flattened_out.ShareDataWith(*Out).Resize(flattened_dims); + flattened_d_out.ShareDataWith(*dOut).Resize(flattened_dims); + flattened_d_x.ShareDataWith(*dX).Resize(flattened_dims); + math::SoftmaxGradCUDNNFunctor()( - context.template device_context(), Out, - dOut, dX); + context.template device_context(), + &flattened_out, &flattened_d_out, &flattened_d_x); } }; diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/softmax_mkldnn_op.cc index 14b57b11fefb2b726531cb164dbf479f8df26b24..01819f53e3ab0973f6140c5a81f18f954b6a0376 100644 --- a/paddle/fluid/operators/softmax_mkldnn_op.cc +++ b/paddle/fluid/operators/softmax_mkldnn_op.cc @@ -26,9 +26,82 @@ using paddle::platform::MKLDNNMemDesc; using mkldnn::memory; // Note: paddle has also "memory" namespace using mkldnn::primitive; -using mkldnn::softmax_forward; using mkldnn::prop_kind; +using mkldnn::softmax_backward; +using mkldnn::softmax_forward; using mkldnn::stream; +using platform::to_void_cast; + +class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler { + public: + SoftmaxMKLDNNHandler( + std::shared_ptr softmax_pd, + const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : platform::MKLDNNHandler(dev_ctx, engine, base_key), + softmax_pd_(softmax_pd) {} + + SoftmaxMKLDNNHandler( + std::shared_ptr softmax_pd, + std::shared_ptr softmax_bwd_pd, + const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : platform::MKLDNNHandler(dev_ctx, engine, base_key), + softmax_pd_(softmax_pd), + softmax_bwd_pd_(softmax_bwd_pd) { + // If we are in Grad operatgor then update a key with BWD suffix to + // distinguish from FWD memory primitives + key_ += "-BWD"; + } + + std::shared_ptr AcquireSoftmax( + std::shared_ptr dst_memory_p, + std::shared_ptr src_memory_p) { + /*Generate key*/ + auto prim_key = key_ + "@softmax_p"; + + auto softmax_p = std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE((softmax_p != nullptr) || (is_reusing_ == false), + "Fail to find softmax primitive in device context"); + if (softmax_p == nullptr) { + softmax_p = std::make_shared( + *(softmax_pd_.get()), + *(static_cast(src_memory_p.get())), + *(static_cast(dst_memory_p.get()))); + dev_ctx_.SetBlob(prim_key, softmax_p); + } else { + is_reusing_ = true; + } + + return softmax_p; + } + + std::shared_ptr AcquireSoftmaxBackward( + std::shared_ptr dst_memory_p, + std::shared_ptr diff_dst_memory_p, + std::shared_ptr diff_src_memory_p) { + auto prim_key = key_ + "@softmax_bwd_p"; + auto softmax_bwd_p = std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE((softmax_bwd_p != nullptr) || (is_reusing_ == false), + "Fail to find softmax backward primitive in device context"); + if (softmax_bwd_p == nullptr) { + softmax_bwd_p = std::make_shared( + *softmax_bwd_pd_, *(dst_memory_p.get()), *(diff_dst_memory_p.get()), + *(diff_src_memory_p.get())); + dev_ctx_.SetBlob(prim_key, softmax_bwd_p); + } else { + is_reusing_ = true; + } + + return softmax_bwd_p; + } + + private: + std::shared_ptr softmax_pd_; + std::shared_ptr softmax_bwd_pd_; +}; template class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { @@ -40,70 +113,51 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { auto mkldnn_engine = dev_ctx.GetEngine(); const Tensor* input = ctx.Input("X"); Tensor* output = ctx.Output("Out"); - PADDLE_ENFORCE(input->dims().size() == 2UL, - "The input of softmax op must be a 2D matrix."); - const T* input_data = input->data(); - // allocate memory for output - T* output_data = output->mutable_data(ctx.GetPlace()); - std::vector src_tz = paddle::framework::vectorize2int(input->dims()); - std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // MKL-DNN does support softmax over selected axis. Having 2D Tensor, - // we will make normalization after final eg. axis: 1 - PADDLE_ENFORCE(((src_tz[0] == dst_tz[0]) && (src_tz[1] == dst_tz[1])), - "Softmax input and output dimensions should match"); + PADDLE_ENFORCE_EQ( + input->dims(), output->dims(), + "The shape of softmax's input and output must be identical."); + + // make sure 'output' holds memory, which will be shared by + // 'flattened_output' later. + output->mutable_data(ctx.GetPlace()); + + // flatten input and output to 2-D matrixs + auto dims = input->dims(); // input and output share the same shape + auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1); + framework::Tensor flattened_input; + framework::Tensor flattened_output; + flattened_input.ShareDataWith(*input).Resize(flattened_dims); + flattened_output.ShareDataWith(*output).Resize(flattened_dims); + + const T* input_data = flattened_input.data(); + T* output_data = flattened_output.mutable_data(ctx.GetPlace()); + + std::vector src_tz = paddle::framework::vectorize2int(flattened_dims); + std::vector dst_tz = src_tz; // Same memory descriptor to be used for input and output memory::dims softmax_tz = {src_tz[0], src_tz[1]}; // Generate keys for storing/retriving primitives for this operator - // TODO(jczaja): Each MKLDNN operator may have diffrent hashing function - auto gethash = [](memory::dims& operand_dims) { - return std::string(std::to_string(operand_dims[0]) + "-" + - std::to_string(operand_dims[1])); - }; - const std::string key = gethash(softmax_tz); - const std::string key_softmax_p = key + "@softmax_p"; - const std::string key_softmax_src_mem_p = key + "@softmax_src_mem_p"; - const std::string key_softmax_dst_mem_p = key + "@softmax_dst_mem_p"; - - std::shared_ptr softmax_p = dev_ctx.GetBlob(key_softmax_p); - if (softmax_p == nullptr) { - // Currently only NC data format is supported - auto softmax_md = - MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc); - // Normalization is made after innermost dimension eg. C out of NC - auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring, - softmax_md, 1 /*dim: C*/); - // create memory primitives - auto softmax_src_memory_p = std::make_shared( - memory::primitive_desc{softmax_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - dev_ctx.SetBlob(key_softmax_src_mem_p, softmax_src_memory_p); - auto softmax_dst_memory_p = std::make_shared( - memory::primitive_desc{softmax_md, mkldnn_engine}, - static_cast(output_data)); - dev_ctx.SetBlob(key_softmax_dst_mem_p, softmax_dst_memory_p); - - auto softmax_forward_pd = - std::make_shared(softmax_desc, - mkldnn_engine); - softmax_p = std::make_shared( - *(softmax_forward_pd.get()), - *(static_cast(softmax_src_memory_p.get())), - *(static_cast(softmax_dst_memory_p.get()))); - dev_ctx.SetBlob(key_softmax_p, softmax_p); - } else { - // Primitives already exist - auto src_memory_p = std::static_pointer_cast( - dev_ctx.GetBlob(key_softmax_src_mem_p)); - PADDLE_ENFORCE(src_memory_p != nullptr, - "Fail to find softmax src mem_p in device context"); - auto dst_memory_p = std::static_pointer_cast( - dev_ctx.GetBlob(key_softmax_dst_mem_p)); - PADDLE_ENFORCE(dst_memory_p != nullptr, - "Fail to find softmax dst mem_p in device context"); - src_memory_p->set_data_handle( - reinterpret_cast(const_cast(input_data))); - dst_memory_p->set_data_handle(output_data); - } + const std::string key = + platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Output("Out")); + const std::string key_softmax_pd = key + "@softmax_pd"; + + // Currently only NC data format is supported + auto softmax_md = MKLDNNMemDesc( + {softmax_tz}, platform::MKLDNNGetDataType(), memory::format::nc); + // Normalization is made after innermost dimension eg. C out of NC + auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring, + softmax_md, 1 /*dim: C*/); + auto softmax_pd = std::make_shared( + softmax_desc, mkldnn_engine); + dev_ctx.SetBlob(key_softmax_pd, softmax_pd); + + SoftmaxMKLDNNHandler handler(softmax_pd, dev_ctx, mkldnn_engine, key); + auto softmax_src_memory_p = + handler.AcquireSrcMemory(softmax_md, to_void_cast(input_data)); + auto softmax_dst_memory_p = + handler.AcquireDstMemory(softmax_md, to_void_cast(output_data)); + auto softmax_p = + handler.AcquireSoftmax(softmax_dst_memory_p, softmax_src_memory_p); std::vector pipeline{ *(static_cast(softmax_p.get()))}; @@ -120,6 +174,88 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { } }; +template +class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + auto& dev_ctx = ctx.template device_context(); + auto mkldnn_engine = dev_ctx.GetEngine(); + const Tensor* output = ctx.Input("Out"); + auto* dout = ctx.template Input(framework::GradVarName("Out")); + auto* dx = + ctx.template Output(framework::GradVarName("X")); + + PADDLE_ENFORCE_EQ( + dout->dims(), dx->dims(), + "The shape of softmax_grad's input and output must be identical."); + + // make sure 'dx' holds memory, which will be shared by 'flattened_dx' + // later. + dx->template mutable_data(ctx.GetPlace()); + + auto dims = dout->dims(); // input and output share the same shape + auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1); + framework::Tensor flattened_output; + framework::Tensor flattened_dout; + framework::Tensor flattened_dx; + flattened_output.ShareDataWith(*output).Resize(flattened_dims); + flattened_dout.ShareDataWith(*dout).Resize(flattened_dims); + flattened_dx.ShareDataWith(*dx).Resize(flattened_dims); + + const T* dst_data = flattened_output.data(); + const T* diff_dst_ptr = flattened_dout.template data(); + T* diff_src_ptr = flattened_dx.template mutable_data(ctx.GetPlace()); + + std::vector dst_tz = paddle::framework::vectorize2int(flattened_dims); + std::vector src_tz(dst_tz); + + // Same memory descriptor to be used for input and output + memory::dims softmax_tz = {src_tz[0], src_tz[1]}; + // Currently only supports NC data format + // retrieve eltwise primitive desc from device context + const std::string key = + platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Input("Out")); + const std::string key_softmax_pd = key + "@softmax_pd"; + + auto softmax_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_softmax_pd)); + PADDLE_ENFORCE(softmax_pd != nullptr, + "Fail to find softmax_pd in device context"); + + // TODO(jczaja): Add layouts support when there is a need to do so + // Two dimensional softmax does support NC format + auto data_softmax_md = MKLDNNMemDesc( + {softmax_tz}, platform::MKLDNNGetDataType(), memory::format::nc); + auto diff_softmax_md = MKLDNNMemDesc( + {softmax_tz}, platform::MKLDNNGetDataType(), memory::format::nc); + // Normalization is made after innermost dimension eg. C out of NC + auto softmax_bwd_desc = + softmax_backward::desc(diff_softmax_md, data_softmax_md, 1 /* dim: C*/); + auto softmax_bwd_pd = + std::make_shared( + softmax_bwd_desc, mkldnn_engine, *softmax_pd); + + SoftmaxMKLDNNHandler handler(softmax_pd, softmax_bwd_pd, dev_ctx, + mkldnn_engine, key); + auto dst_memory_p = + handler.AcquireDstMemory(data_softmax_md, to_void_cast(dst_data)); + auto diff_dst_memory_p = handler.AcquireDiffDstMemory( + diff_softmax_md, to_void_cast(diff_dst_ptr)); + auto diff_src_memory_p = handler.AcquireDiffSrcMemory( + diff_softmax_md, to_void_cast(diff_src_ptr)); + + // Get primitve from device context + auto softmax_bwd_p = handler.AcquireSoftmaxBackward( + dst_memory_p, diff_dst_memory_p, diff_src_memory_p); + + std::vector pipeline{*softmax_bwd_p}; + stream(stream::kind::eager).submit(pipeline).wait(); + } +}; } // namespace operators } // namespace paddle @@ -127,3 +263,5 @@ namespace ops = paddle::operators; REGISTER_OP_KERNEL(softmax, MKLDNN, ::paddle::platform::CPUPlace, ops::SoftmaxMKLDNNKernel); +REGISTER_OP_KERNEL(softmax_grad, MKLDNN, ::paddle::platform::CPUPlace, + ops::SoftmaxMKLDNNGradKernel); diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index cc256aa627bdda0609f496cab93a2dec7d95f348..bb081238820b9ee3ae095442d21cfce11f7b41e5 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -37,10 +37,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of SoftmaxOp should not be null."); - auto x_dims = ctx->GetInputDim("X"); - PADDLE_ENFORCE(x_dims.size() == 2UL, - "The input of softmax op must be a matrix."); - ctx->SetOutputDim("Out", x_dims); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); ctx->ShareLoD("X", /*->*/ "Out"); } @@ -49,6 +46,9 @@ class SoftmaxOp : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { // choose cudnn kernel if the runtime supported. framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -58,6 +58,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif @@ -68,9 +69,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { "float16 can only be used on GPU place"); } - std::string data_format = ctx.Attr("data_format"); - return framework::OpKernelType(input_data_type, ctx.GetPlace(), - framework::StringToDataLayout(data_format), + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, library_); } }; @@ -79,9 +78,10 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "The input tensor of softmax. " - "2-D with shape [batch_size, input_feature_dimensions]."); - AddOutput("Out", "The normalized values with the same shape as X."); + "The input tensor of softmax, " + "whose last dimension is the input_feature_dimensions."); + AddOutput("Out", "The normalized values with the same shape as X.") + .Reuse("X"); AddAttr( "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") @@ -102,20 +102,23 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Softmax Operator. -The input of the softmax operator is a 2-D tensor with shape N x K (N is the -batch_size, K is the dimension of input feature). The output tensor has the -same shape as the input tensor. +The input of the softmax operator is a tensor of any rank. The output tensor +has the same shape as the input. -For each row of the input tensor, the softmax operator squashes the -K-dimensional vector of arbitrary real values to a K-dimensional vector of real -values in the range [0, 1] that add up to 1. +The input tensor will first be logically flattened to a 2-D matrix. The matrix's +second dimension(row length) is as same as the last dimension of the input +tensor, and the first dimension(column length) is the product of all other +dimensions of the input tensor. For each row of the matrix, the softmax operator +squashes the K-dimensional(K is the width of the matrix, which is also the size +of the input tensor's last dimension) vector of arbitrary real values to a +K-dimensional vector of real values in the range [0, 1] that add up to 1. It computes the exponential of the given dimension and the sum of exponential values of all the other dimensions in the K-dimensional vector input. Then the ratio of the exponential of the given dimension and the sum of exponential values of all the other dimensions is the output of the softmax operator. -For each row $i$ and each column $j$ in Input(X), we have: +For each row $i$ and each column $j$ in the matrix, we have: $$Out[i, j] = \frac{\exp(X[i, j])}{\sum_j(exp(X[i, j])}$$ )DOC"); @@ -134,7 +137,8 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { ctx->GetInputDim(framework::GradVarName("Out")), "Input(Out) and its gradients should have a same shape."); - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->SetOutputDim(framework::GradVarName("X"), + ctx->GetInputDim(framework::GradVarName("Out"))); } protected: @@ -142,25 +146,58 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { // choose cudnn kernel if the runtime supported. framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - framework::StringToDataLayout(data_format), library_); +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; + } +#endif + auto input_data_type = framework::ToDataType( + ctx.Input(framework::GradVarName("Out"))->type()); + if (input_data_type == framework::proto::VarType::FP16) { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "float16 can only be used on GPU place"); + } + + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, + library_); } }; +class SoftmaxOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("softmax_grad"); + + op->SetInput("Out", Output("Out")); + op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + return std::unique_ptr(op); + } +}; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, - paddle::framework::DefaultGradOpDescMaker); + ops::SoftmaxOpGradMaker); REGISTER_OPERATOR(softmax_grad, ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL( softmax, ops::SoftmaxKernel, diff --git a/paddle/fluid/operators/softmax_op.h b/paddle/fluid/operators/softmax_op.h index 600da45a0bbb69b76d59c981e195fc03a49b0504..cf1eeb017d666f605a431aa54637d8cbc99c7c46 100644 --- a/paddle/fluid/operators/softmax_op.h +++ b/paddle/fluid/operators/softmax_op.h @@ -31,8 +31,12 @@ class SoftmaxKernel : public framework::OpKernel { // allocate memory on device. Out->mutable_data(context.GetPlace()); + int rank = X->dims().size(); + Tensor X_2d = framework::ReshapeToMatrix(*X, rank - 1); + Tensor Out_2d = framework::ReshapeToMatrix(*Out, rank - 1); + math::SoftmaxFunctor()( - context.template device_context(), X, Out); + context.template device_context(), &X_2d, &Out_2d); } }; @@ -47,8 +51,14 @@ class SoftmaxGradKernel : public framework::OpKernel { // allocate memory on device. dX->mutable_data(context.GetPlace()); + int rank = Out->dims().size(); + Tensor Out_2d = framework::ReshapeToMatrix(*Out, rank - 1); + Tensor dOut_2d = framework::ReshapeToMatrix(*dOut, rank - 1); + Tensor dX_2d = framework::ReshapeToMatrix(*dX, rank - 1); + math::SoftmaxGradFunctor()( - context.template device_context(), Out, dOut, dX); + context.template device_context(), &Out_2d, &dOut_2d, + &dX_2d); } }; diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu index 8f7840cee1dd95a828fd4ac8815e335a5db47e3d..a559b01ed32a48e3befb37c2ae8935b4f3a4acb0 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -14,6 +14,8 @@ limitations under the License. */ #define EIGEN_USE_GPU +#include +#include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/operators/softmax_with_cross_entropy_op.h" namespace paddle { @@ -53,8 +55,196 @@ __global__ void SoftCrossEntropyGradientKernel(T* logit_grad, logit_grad[ids] = loss_grad[row_ids] * (logit_grad[ids] - labels[ids]); } } + } // namespace +static __device__ __forceinline__ float real_exp(float x) { return expf(x); } +static __device__ __forceinline__ double real_exp(double x) { return exp(x); } +static __device__ __forceinline__ float real_log(float x) { + return math::TolerableValue()(logf(x)); +} +static __device__ __forceinline__ double real_log(double x) { + return math::TolerableValue()(log(x)); +} + +/** In the following codes, 3 CUDA kernels are implemented to calculate softmax + * and loss **/ +/* + Supposing the x is `logits` and y is `labels`, the equations are as +followings: + + cross\_entropy_i = \sum_{j}[- y_i_j * log({e^{x_i_j}/\sum_{j}e^{x_i_j}})] + = \sum_{j}[- y_i_j * log({e^{x_i_j - max_i}/\sum_{j}e^{x_i_j-max_i}})] + = \sum_{j}[-y_i_j * (x_i_j - max_i - log\sum_{j}e^{x_i_j - max_i})] + = \sum_{j}[-y_i_j * (x_i_j - max_i - logDiffMaxSum_i)] + = \sum_{j}(-y_i_j * tmp_i_j) + + softmax_i_j = e^{tmp_i_j} + +where: + max_i = \max_{j}{x_i_j} + logDiffMaxSum_i = log\sum_{j}e^{x_i_j - max_i} + tmp_i_j = x_i_j - max_i - logDiffMaxSum_i + +Therefore, the calculation can be separated into 3 steps: +Step 1: row-wise operation to calculate max_i +Step 2: row-wise operation to calculate logDiffMaxSum_i +Step 3: caculate tmp_i_j, and finally get softmax_i_j and cross\_entropy_i + +To save memory, we can share memory among max_i, logDiffMaxSum_i and +cross\_entropy_i. +In this way, the 3 steps should be changed to: +Step 1 (RowReductionForMax): row-wise operation to calculate max_i +Step 2 (RowReductionForDiffMaxSum): calculate immediate result of softmax'_i_j = +x_i_j - max_i, and row-wise operation to calculate logDiffMaxSum_i +Step 3 (RowReductionForSoftmaxAndCrossEntropy): calculate tmp_i_j = softmax'_i_j +- logDiffMaxSum_i, and finally get softmax_i_j and cross\_entropy_i +*/ + +// There are 3 kinds of reduce algorithms in cub: +// BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY +// BLOCK_REDUCE_RAKING +// BLOCK_REDUCE_WARP_REDUCTIONS (default) +template +using BlockReduce = + cub::BlockReduce; + +template +using BlockReduceTempStorage = typename BlockReduce::TempStorage; + +// Make sure that BlockDim <= feature_size +// This kernel is used to calculate the max element of each row +template +__global__ void RowReductionForMax(const T* logits_data, T* max_data, + int feature_size) { + __shared__ BlockReduceTempStorage temp_storage; + + auto beg_idx = feature_size * blockIdx.x + threadIdx.x; + auto end_idx = feature_size * (blockIdx.x + 1); + + T cur_max = logits_data[beg_idx]; + beg_idx += BlockDim; + while (beg_idx < end_idx) { + if (cur_max < logits_data[beg_idx]) { + cur_max = logits_data[beg_idx]; + } + beg_idx += BlockDim; + } + + cur_max = BlockReduce(temp_storage).Reduce(cur_max, cub::Max()); + + if (threadIdx.x == 0) { + max_data[blockIdx.x] = cur_max < -64 ? -64 : cur_max; + } +} + +// Make sure that BlockDim <= feature_size +template +__global__ void RowReductionForDiffMaxSum(const T* logits_data, T* max_data, + T* softmax, int feature_size) { + __shared__ BlockReduceTempStorage temp_storage; + + auto beg_idx = feature_size * blockIdx.x + threadIdx.x; + auto end_idx = feature_size * (blockIdx.x + 1); + + auto block_max = max_data[blockIdx.x]; + + softmax[beg_idx] = logits_data[beg_idx] - block_max; + T diff_max_sum = real_exp(softmax[beg_idx]); + beg_idx += BlockDim; + while (beg_idx < end_idx) { + softmax[beg_idx] = logits_data[beg_idx] - block_max; + diff_max_sum += real_exp(softmax[beg_idx]); + beg_idx += BlockDim; + } + + diff_max_sum = + BlockReduce(temp_storage).Reduce(diff_max_sum, cub::Sum()); + if (threadIdx.x == 0) max_data[blockIdx.x] = real_log(diff_max_sum); +} + +// Make sure that BlockDim <= feature_size +template +__global__ void RowReductionForSoftmaxAndCrossEntropy(const T* logits_data, + const T* labels_data, + T* loss_data, T* softmax, + int feature_size) { + __shared__ BlockReduceTempStorage temp_storage; + + auto beg_idx = feature_size * blockIdx.x + threadIdx.x; + auto end_idx = feature_size * (blockIdx.x + 1); + + // log_diff_max_sum shares memory with loss + auto block_log_diff_max_sum = loss_data[blockIdx.x]; + auto tmp = softmax[beg_idx] - block_log_diff_max_sum; + softmax[beg_idx] = real_exp(tmp); + auto loss = -labels_data[beg_idx] * tmp; + beg_idx += BlockDim; + while (beg_idx < end_idx) { + tmp = softmax[beg_idx] - block_log_diff_max_sum; + softmax[beg_idx] = real_exp(tmp); + loss -= (labels_data[beg_idx] * tmp); + beg_idx += BlockDim; + } + + loss = BlockReduce(temp_storage).Reduce(loss, cub::Sum()); + if (threadIdx.x == 0) loss_data[blockIdx.x] = loss; +} + +template +__global__ void SetSoftmaxToOneWhenFeatureSizeIsOne(T* out, int batch_size) { + auto idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < batch_size) out[idx] = static_cast(1); +} + +template +static void SoftmaxWithCrossEntropyFusedKernel(const T* logits_data, + const T* labels_data, + T* softmax_data, T* loss_data, + int batch_size, int feature_size, + cudaStream_t stream) { + constexpr int kMaxBlockDim = 512; + int block_dim = feature_size >= kMaxBlockDim + ? kMaxBlockDim + : (1 << static_cast(std::log2(feature_size))); + +#define CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(BlockDim) \ + case BlockDim: \ + RowReductionForMax<<>>( \ + logits_data, loss_data, feature_size); \ + RowReductionForDiffMaxSum<<>>( \ + logits_data, loss_data, softmax_data, feature_size); \ + RowReductionForSoftmaxAndCrossEntropy< \ + T, BlockDim><<>>( \ + logits_data, labels_data, loss_data, softmax_data, feature_size); \ + break + + switch (block_dim) { + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(512); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(256); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(128); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(64); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(32); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(16); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(8); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(4); + CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(2); + case 1: + SetSoftmaxToOneWhenFeatureSizeIsOne<<<(batch_size + kMaxBlockDim - 1) / + kMaxBlockDim, + kMaxBlockDim, 0, stream>>>( + softmax_data, batch_size); + cudaMemsetAsync(loss_data, 0, batch_size, stream); + break; + default: + PADDLE_THROW("BlockDim must be 2^n in softmax_with_cross_entropy_op"); + break; + } + +#undef CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL +} + template class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel { public: @@ -66,14 +256,24 @@ class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel { Tensor* softmax = context.Output("Softmax"); Tensor* loss = context.Output("Loss"); - softmax->mutable_data(context.GetPlace()); - loss->mutable_data(context.GetPlace()); - - math::SoftmaxFunctor()( - context.cuda_device_context(), logits, softmax); - math::CrossEntropyFunctor()( - context.cuda_device_context(), loss, softmax, labels, - context.Attr("soft_label")); + auto* softmax_data = softmax->mutable_data(context.GetPlace()); + auto* loss_data = loss->mutable_data(context.GetPlace()); + + auto soft_label = context.Attr("soft_label"); + if (soft_label) { + int batch_size = logits->dims()[0]; + int feature_size = logits->dims()[1]; + auto* logits_data = logits->data(); + auto* labels_data = labels->data(); + SoftmaxWithCrossEntropyFusedKernel( + logits_data, labels_data, softmax_data, loss_data, batch_size, + feature_size, context.cuda_device_context().stream()); + } else { + math::SoftmaxCUDNNFunctor()(context.cuda_device_context(), logits, + softmax); + math::CrossEntropyFunctor()( + context.cuda_device_context(), loss, softmax, labels, false); + } } }; diff --git a/paddle/fluid/operators/split_ids_op.h b/paddle/fluid/operators/split_ids_op.h index d263426e073d95ad6d584c7370baf596587a993d..c4af5a65fc5f81c1af7c1fdcca637ca37c940637 100644 --- a/paddle/fluid/operators/split_ids_op.h +++ b/paddle/fluid/operators/split_ids_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" @@ -67,10 +68,15 @@ class SplitIdsOpKernel : public framework::OpKernel { const auto &ids_rows = ids_selected_rows->rows(); auto outs = ctx.MultiOutput("Out"); const size_t shard_num = outs.size(); + for (auto &out : outs) { + out->mutable_rows()->clear(); + } // get rows for outputs - for (auto &id : ids_rows) { - size_t shard_id = static_cast(id) % shard_num; - outs[shard_id]->mutable_rows()->push_back(id); + std::unordered_map id_to_index; + for (size_t i = 0; i < ids_rows.size(); ++i) { + id_to_index[ids_rows[i]] = i; + size_t shard_id = static_cast(ids_rows[i]) % shard_num; + outs[shard_id]->mutable_rows()->push_back(ids_rows[i]); } int64_t row_width = ids_dims[1]; @@ -80,7 +86,8 @@ class SplitIdsOpKernel : public framework::OpKernel { {static_cast(out->rows().size()), row_width}); T *output = out->mutable_value()->mutable_data(ddim, place); for (int64_t i = 0; i < ddim[0]; ++i) { - memcpy(output + i * row_width, ids + out->rows()[i] * row_width, + memcpy(output + i * row_width, + ids + id_to_index[out->rows()[i]] * row_width, row_width * sizeof(T)); } } diff --git a/paddle/fluid/operators/split_op.cc b/paddle/fluid/operators/split_op.cc index 5e2b2a994534c2fb1e053c067b36651d358b9da8..d661b276bc31bf0c3ab181d706ffdccec89f0632 100644 --- a/paddle/fluid/operators/split_op.cc +++ b/paddle/fluid/operators/split_op.cc @@ -115,4 +115,7 @@ USE_CPU_ONLY_OP(concat); REGISTER_OPERATOR(split, ops::SplitOp, ops::SplitOpMaker, ops::SplitGradMaker); REGISTER_OP_CPU_KERNEL(split, - ops::SplitOpKernel); + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel); diff --git a/paddle/fluid/operators/split_op.cu.cc b/paddle/fluid/operators/split_op.cu.cc index efa378af857a8881f25c76379ba7cf81e64c80bb..18e0904681753aff7f3deac96efb6d62f389a031 100644 --- a/paddle/fluid/operators/split_op.cu.cc +++ b/paddle/fluid/operators/split_op.cu.cc @@ -15,4 +15,7 @@ limitations under the License. */ #include "paddle/fluid/operators/split_op.h" namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( - split, ops::SplitOpKernel); + split, ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel); diff --git a/paddle/fluid/operators/squeeze_op.cc b/paddle/fluid/operators/squeeze_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6c507baf3a0ab0a557d29a53700685753616193b --- /dev/null +++ b/paddle/fluid/operators/squeeze_op.cc @@ -0,0 +1,202 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class SqueezeOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SqueezeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SqueezeOp should not be null."); + + const auto &x_dims = ctx->GetInputDim("X"); + // Check input tensor dims (<6) Eigen limit. + PADDLE_ENFORCE(x_dims.size() <= 6, + "Invalid dimnesions, the rank of Input(X) " + "should be in the range of [1, 6] (Eigen limit)."); + + const auto &axes = ctx->Attrs().Get>("axes"); + for (int a : axes) { + PADDLE_ENFORCE_LT(a, x_dims.size(), + "The squeeze axis should be less than input " + "tensor's rank."); + } + + auto out_dims = GetOutputShape(axes, x_dims); + ctx->SetOutputDim("Out", out_dims); + if (x_dims[0] == out_dims[0]) { + // Only pass LoD when the first dimension of output and Input(X) + // are the same. + ctx->ShareLoD("X", "Out"); + } + } + + static framework::DDim GetOutputShape(const std::vector squeeze_dims, + const framework::DDim &in_dims) { + size_t num_squeeze_dims = squeeze_dims.size(); + int cnt_squeezed_dims = 0; + bool should_squeeze[9] = {false}; + + // Determines number of dimensions of output tensor after squeeze. + // Mark and count the dimensions need to be squeezed + if (num_squeeze_dims == 0) { + for (int idx = 0; idx < in_dims.size(); ++idx) { + if (in_dims[idx] == 1) { + should_squeeze[idx] = true; + ++cnt_squeezed_dims; + } + } + } else { + for (size_t idx = 0; idx < num_squeeze_dims; ++idx) { + int current = squeeze_dims[idx] < 0 ? squeeze_dims[idx] + in_dims.size() + : squeeze_dims[idx]; + // Check current index, the upper limit has beed checked in line 36. + PADDLE_ENFORCE(current >= 0, + "Invalid axis, the negative axis is out of range."); + PADDLE_ENFORCE(in_dims[current] == 1, + "Invalid axis index, the axis that will be squeezed " + "should be equal to 1."); + + if (!(should_squeeze[current])) { + ++cnt_squeezed_dims; + } + should_squeeze[current] = true; + } + } + + // Make output dimensions + std::vector output_shape(in_dims.size() - cnt_squeezed_dims, 0); + for (int in_idx = 0, out_idx = 0; in_idx < in_dims.size(); ++in_idx) { + if (!should_squeeze[in_idx]) { + output_shape[out_idx++] = in_dims[in_idx]; + } + } + + return framework::make_ddim(output_shape); + } +}; + +class SqueezeOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto &axes = Attr>("axes"); + auto x_dims = scope.FindVar(Input("X"))->Get().dims(); + auto out_dims = SqueezeOpInferShape::GetOutputShape(axes, x_dims); + + framework::AttributeMap attrs; + attrs["shape"] = framework::vectorize2int(out_dims); + attrs["inplace"] = Attr("inplace"); + // Invoke Reshape Op + auto reshape_op = framework::OpRegistry::CreateOp( + "reshape", {{"X", {Input("X")}}, {"Shape", {}}}, + {{"Out", {Output("Out")}}}, attrs); + reshape_op->Run(scope, place); + } +}; + +class SqueezeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor). The input tensor of squeeze operator."); + AddOutput("Out", "(Tensor). The output tensor of squeeze operator."); + AddAttr>("axes", + "(std::vector). List of integers," + " indicating the dimensions to squeeze.") + .SetDefault({}); + AddAttr("inplace", + "(default: false) Squeeze the source tensor's shape without " + "memory copy. When Attr(inplace) is set true, the output " + "tensor shares memory with Input(X), otherwise, a new output " + "tensor is created, and its data are copied from Input(x).") + .SetDefault(false); + AddComment(R"DOC( + Squeeze Operator. + + Remove single-dimensional entries from the shape of a tensor. + Takes a parameter axes with a list of axes to squeeze. + If axes is not provided, all the single dimensions will be removed from the shape. + If an axis is selected with shape entry not equal to one, an error is raised. + + Examples: + Case 1: + Given + X.shape = (1, 3, 1, 5) + and + axes = [0] + we get: + Out.shape = (3, 1, 5) + + Case 2: + Given + X.shape = (1, 3, 1, 5) + and + axes = [] + we get: + Out.shape = (3, 5) + )DOC"); + } +}; + +class SqueezeGradInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *context) const override { + context->SetOutputDim(framework::GradVarName("X"), + context->GetInputDim("X")); + context->ShareLoD("X", framework::GradVarName("X")); + } +}; + +class SqueezeGradOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto dx_name = Output(framework::GradVarName("X")); + auto dout_name = Input(framework::GradVarName("Out")); + auto x_dims = scope.FindVar(Input("X"))->Get().dims(); + framework::AttributeMap attrs; + attrs["shape"] = framework::vectorize2int(x_dims); + attrs["inplace"] = Attr("inplace"); + + auto reshape_op = framework::OpRegistry::CreateOp( + "reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}}, + attrs); + reshape_op->Run(scope, place); + } +}; + +} // namespace operators +} // namespace paddle + +// Tell linker to use reshape op +USE_OP(reshape); + +namespace ops = paddle::operators; +REGISTER_OPERATOR(squeeze, ops::SqueezeOp, ops::SqueezeOpMaker, + ops::SqueezeOpInferShape, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(squeeze_grad, ops::SqueezeGradOp, ops::SqueezeGradInferShape); diff --git a/paddle/fluid/operators/sum_mkldnn_op.cc b/paddle/fluid/operators/sum_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d2035777ee2289291a02594ee289156504df09d9 --- /dev/null +++ b/paddle/fluid/operators/sum_mkldnn_op.cc @@ -0,0 +1,240 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/*Licensed under the Apache License, Version 2.0(the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "mkldnn.hpp" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/operators/sum_op.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using paddle::framework::Tensor; +using paddle::platform::MKLDNNDeviceContext; +using paddle::platform::CPUDeviceContext; +using framework::DataLayout; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::stream; +using mkldnn::sum; +using mkldnn::reorder; +using platform::to_void_cast; + +template +class SumMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + auto& dev_ctx = ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + auto in_vars = ctx.MultiInputVar("X"); + + const int N = in_vars.size(); + auto out_var = ctx.OutputVar("Out"); + bool in_place = out_var == in_vars[0]; + + if (out_var->IsType()) { + LoDTensor* output = ctx.Output("Out"); + T* output_data = output->mutable_data(ctx.GetPlace()); + + std::vector dst_tz = framework::vectorize2int(output->dims()); + auto src_tz = dst_tz; + memory::format output_format{memory::format::format_undef}; + std::vector scales; + std::vector srcs_mpd; + std::vector srcs_mem; + + PADDLE_ENFORCE(in_vars[0]->IsType(), + "Input[0] must be LoDTensors"); + auto& input0 = in_vars[0]->Get(); + PADDLE_ENFORCE(input0.layout() == DataLayout::kMKLDNN && + input0.format() != memory::format::format_undef, + "Wrong layout/format for inputs[0]"); + + memory::format input_format = input0.format(); + + if (src_tz.size() == 1 && (input_format == memory::format::nchw || + input_format == memory::format::nhwc)) { + input_format = memory::format::x; + } + if (src_tz.size() == 2 && (input_format == memory::format::nchw || + input_format == memory::format::nhwc)) { + input_format = memory::format::nc; + } + + for (int i = 0; i < N; i++) { + PADDLE_ENFORCE(in_vars[i]->IsType(), + "all inputs must be all LoDTensors"); + auto& input = in_vars[i]->Get(); + PADDLE_ENFORCE(input.layout() == DataLayout::kMKLDNN && + input.format() != memory::format::format_undef, + "Wrong layout/format for inputs"); + + if (input.numel() == 0) { + continue; + } + + const T* input_data = input.data(); + + auto src_md = + memory::desc(src_tz, memory::data_type::f32, input_format); + auto src_mpd = memory::primitive_desc(src_md, mkldnn_engine); + auto src_mem = memory(src_mpd, to_void_cast(input_data)); + srcs_mpd.push_back(src_mpd); + srcs_mem.push_back(src_mem); + scales.push_back(1.0); + } + + auto dst_md = + memory::desc(dst_tz, memory::data_type::f32, memory::format::any); + + auto sum_pd = sum::primitive_desc(dst_md, scales, srcs_mpd); + + std::shared_ptr dst_mem; + if (in_place) { + dst_mem.reset(new memory(sum_pd.dst_primitive_desc())); + } else { + dst_mem.reset(new memory(sum_pd.dst_primitive_desc(), output_data)); + } + std::vector inputs; + for (size_t i = 0; i < srcs_mem.size(); ++i) { + inputs.push_back(srcs_mem[i]); + } + + auto sum_prim = mkldnn::sum(sum_pd, inputs, *dst_mem); + output_format = (memory::format)platform::GetMKLDNNFormat(sum_pd); + + primitive reorder_prim; + std::shared_ptr target_mem; + if (in_place) { + output_format = input_format; + target_mem.reset(new memory( + {{{src_tz}, memory::data_type::f32, output_format}, mkldnn_engine}, + output_data)); + reorder_prim = reorder(*dst_mem, *target_mem); + } + + std::vector pipeline; + pipeline.push_back(sum_prim); + if (in_place) pipeline.push_back(reorder_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(output_format); + } else if (out_var->IsType()) { + // TODO(@mozga-intel) Add MKLDNN SelectedRows support + std::unique_ptr in0; + if (in_place) { + // If is in_place, we store the input[0] to in0 + auto& in_sel0 = in_vars[0]->Get(); + auto& rows = in_sel0.rows(); + in0.reset(new framework::SelectedRows(rows, in_sel0.height())); + in0->mutable_value()->ShareDataWith(in_sel0.value()); + } + + auto get_selected_row = [&](size_t i) -> const SelectedRows& { + if (i == 0 && in0) { + return *in0.get(); + } else { + return in_vars[i]->Get(); + } + }; + auto* out = ctx.Output("Out"); + out->mutable_rows()->clear(); + auto* out_value = out->mutable_value(); + + // Runtime InferShape + size_t first_dim = 0; + for (int i = 0; i < N; i++) { + auto& sel_row = get_selected_row(i); + first_dim += sel_row.rows().size(); + } + auto in_dim = + framework::vectorize(get_selected_row(N - 1).value().dims()); + in_dim[0] = static_cast(first_dim); + + out_value->Resize(framework::make_ddim(in_dim)); + + // if all the input sparse vars are empty, no need to + // merge these vars. + if (first_dim == 0UL) { + return; + } + out_value->mutable_data(ctx.GetPlace()); + math::SelectedRowsAddTo functor; + int64_t offset = 0; + for (int i = 0; i < N; i++) { + auto& sel_row = get_selected_row(i); + if (sel_row.rows().size() == 0) { + continue; + } + PADDLE_ENFORCE_EQ(out->height(), sel_row.height()); + functor(ctx.template device_context(), sel_row, + offset, out); + offset += sel_row.value().numel(); + } + } else if (out_var->IsType()) { + // TODO(@mozga-intel) Add MKLDNN LoDTensorArray support + auto& out_array = *out_var->GetMutable(); + for (size_t i = in_place ? 1 : 0; i < in_vars.size(); ++i) { + PADDLE_ENFORCE(in_vars[i]->IsType(), + "Only support all inputs are TensorArray"); + auto& in_array = in_vars[i]->Get(); + + for (size_t i = 0; i < in_array.size(); ++i) { + if (in_array[i].numel() != 0) { + if (i >= out_array.size()) { + out_array.resize(i + 1); + } + if (out_array[i].numel() == 0) { + framework::TensorCopy(in_array[i], in_array[i].place(), + ctx.device_context(), &out_array[i]); + out_array[i].set_lod(in_array[i].lod()); + } else { + PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod()); + auto in = EigenVector::Flatten(in_array[i]); + auto result = EigenVector::Flatten(out_array[i]); + result.device(*ctx.template device_context() + .eigen_device()) = result + in; + } + } + } + } + } else { + PADDLE_THROW("Unexpected branch, output variable type is %s", + out_var->Type().name()); + } + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_KERNEL(sum, MKLDNN, ::paddle::platform::CPUPlace, + paddle::operators::SumMKLDNNOpKernel); diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index bcc5e22d4a77349e7cde9a43b83f23d4c867d994..fe7c7039c7dec714e265ede1b7167fd800ddc2f7 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -18,6 +18,10 @@ limitations under the License. */ #include "paddle/fluid/framework/var_type_inference.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace operators { using framework::Tensor; @@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { auto x_vars = ctx.MultiInputVar("X"); + + framework::LibraryType library{framework::LibraryType::kPlain}; + framework::DataLayout layout{framework::DataLayout::kAnyLayout}; + +#ifdef PADDLE_WITH_MKLDNN + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } +#endif + if (x_vars[0]->IsType()) { int dtype = -1; for (auto& x_var : x_vars) { @@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel { "Sum operator should have at least one tensor"); return framework::OpKernelType( - static_cast(dtype), - ctx.device_context()); + static_cast(dtype), ctx.GetPlace(), + layout, library); } else if (x_vars[0]->IsType()) { for (auto& var : x_vars) { auto& value = var->Get().value(); if (value.IsInitialized()) { return framework::OpKernelType(framework::ToDataType(value.type()), - ctx.device_context()); + ctx.device_context(), layout, library); } } // if input sparse vars are not initialized, use an default kernel type. return framework::OpKernelType(framework::proto::VarType::FP32, - ctx.device_context()); + ctx.device_context(), layout, library); } else if (x_vars[0]->IsType()) { for (auto& x_var : x_vars) { auto& array = x_var->Get(); for (auto& each : array) { if (each.numel() != 0) { return framework::OpKernelType(framework::ToDataType(each.type()), - ctx.device_context()); + ctx.device_context(), layout, + library); } } } @@ -115,7 +132,10 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); - AddOutput("Out", "(Tensor) The output tensor of sum operator."); + AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X"); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Sum operator. @@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference { framework::BlockDesc* block) const override { auto& inputs = op_desc.Input("X"); auto var_type = framework::proto::VarType::SELECTED_ROWS; - for (auto& name : op_desc.Input("X")) { VLOG(10) << name << " " << block->FindRecursiveOrCreateVar(name).GetType(); @@ -206,6 +225,7 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker, ops::SumOpVarTypeInference); + REGISTER_OP_CPU_KERNEL( sum, ops::SumKernel, ops::SumKernel, diff --git a/paddle/fluid/operators/tensor_array_read_write_op.cc b/paddle/fluid/operators/tensor_array_read_write_op.cc index c703d11eeccf8418250f00c801f47418ee9c85ae..a2d44284e9de1ace42cabbce82e0b45929432d7b 100644 --- a/paddle/fluid/operators/tensor_array_read_write_op.cc +++ b/paddle/fluid/operators/tensor_array_read_write_op.cc @@ -38,15 +38,14 @@ class WriteToArrayOp : public ArrayOp { << " to " << offset + 1; out->resize(offset + 1); } + auto *out_tensor = &out->at(offset); + out_tensor->set_lod(x_tensor.lod()); if (x_tensor.memory_size() > 0) { - auto *out_tensor = &out->at(offset); - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(place); TensorCopy(x_tensor, place, dev_ctx, out_tensor); - out_tensor->set_lod(x_tensor.lod()); } else { VLOG(10) << "WARNING: The input tensor 'x_tensor' holds no memory, so " "nothing has been written to output array[" diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index 83e768b4dc9c607b0f73d7183462d772ae7ab994..ee3078876c15b06a887064f08dc0c05d450b5f77 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -14,26 +14,113 @@ #ifdef PADDLE_WITH_CUDA -#include "paddle/fluid/operators/tensorrt_engine_op.h" +#include +#include + #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/inference/utils/singleton.h" +#include "paddle/fluid/operators/tensorrt_engine_op.h" namespace paddle { + +DEFINE_int32(tensorrt_engine_batch_size, 1, "the batch_size of TensorRT"); + namespace operators { +using inference::Singleton; +using inference::tensorrt::TRT_EngineManager; + +using FluidDT = framework::proto::VarType_Type; +using TRT_DT = nvinfer1::DataType; + +namespace { + +TRT_DT FluidDataType2TRT(FluidDT type) { + switch (type) { + case FluidDT::VarType_Type_FP32: + return TRT_DT::kFLOAT; + case FluidDT::VarType_Type_INT32: + return TRT_DT::kINT32; + default: + return TRT_DT::kINT32; + } + PADDLE_THROW("unkown type"); + return TRT_DT::kINT32; +} + +nvinfer1::Dims Vec2TRT_Dims(const std::vector &shape) { + PADDLE_ENFORCE_GT(shape.size(), 1UL, + "TensorRT' tensor input requires at least 2 dimensions"); + PADDLE_ENFORCE_LE(shape.size(), 4UL, + "TensorRT' tensor input requires at most 4 dimensions"); + PADDLE_ENFORCE_EQ(shape.size(), 4UL); + return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]); +} + +} // namespace + template -void paddle::operators::TensorRTEngineKernel::Prepare( +void TensorRTEngineKernel::Prepare( const framework::ExecutionContext &context) const { + VLOG(4) << "Prepare engine"; // Get the ProgramDesc and pass to convert. - const auto &block = context.Attr("subgraph"); - max_batch_ = context.Attr("max_batch"); + framework::proto::BlockDesc block_desc; + block_desc.ParseFromString(context.Attr("subgraph")); + int max_batch = context.Attr("max_batch"); auto max_workspace = context.Attr("max_workspace"); - engine_.reset(new inference::tensorrt::TensorRTEngine( - max_batch_, max_workspace, nullptr)); + auto params = context.Attr>("parameters"); + std::unordered_set parameters; + for (const auto ¶m : params) { + parameters.insert(param); + } + + std::vector output_maps = + context.Attr>("output_name_mapping"); + + // TODO(Superjomn) replace this with a different stream + auto *engine = Singleton::Global().Create( + max_batch, max_workspace, nullptr /*engine hold its own stream*/, + context.Attr("engine_uniq_key")); + engine->InitNetwork(); + + framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); + VLOG(4) << "parsed var size " << block.AllVars().size(); + // Add inputs + VLOG(4) << "declare inputs"; + for (auto &input : context.Inputs("Xs")) { + if (parameters.count(input)) continue; + VLOG(4) << "declare input " << input; + auto *var = block.FindVar(input); + // TensorRT engine need to create parameters. The parameter's description + // should be set in + PADDLE_ENFORCE(var, "no variable called %s", input); + PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, + "TensorRT engine only takes LoDTensor as input"); + auto shape = var->GetShape(); + // For the special batch_size placeholder -1, drop it and pass the real + // shape of data. + // TODO(Superjomn) fix this with batch broadcast, or it can't handle + // variational batch size. + if (shape[0] == -1) { + shape[0] = FLAGS_tensorrt_engine_batch_size; + } + engine->DeclareInput( + input, FluidDataType2TRT( + var->Proto()->type().lod_tensor().tensor().data_type()), + Vec2TRT_Dims(shape)); + } + inference::Singleton::Global().ConvertBlock( - block, engine_.get()); - engine_->FreezeNetwork(); + block_desc, parameters, context.scope(), engine); + + // Add outputs + for (auto &output : output_maps) { + engine->DeclareOutput(output); + } + + engine->FreezeNetwork(); } class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { @@ -41,7 +128,10 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("Xs", "A list of inputs.").AsDuplicable(); AddOutput("Ys", "A list of outputs").AsDuplicable(); - AddAttr("subgraph", "the subgraph"); + AddAttr("subgraph", "the subgraph."); + AddAttr("engine_uniq_key", "unique key for the TRT engine."); + AddAttr("max_batch", "the maximum batch size."); + AddAttr("max_workspace", "the maximum batch size."); AddComment("TensorRT engine operator."); } }; diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index fe273d386c529be3df05a955f492e2c39d4d8812..2cbe1213a2f428a3ce56b06f97636baeb4b66c26 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -16,13 +16,22 @@ #ifdef PADDLE_WITH_CUDA +#include +#include + #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/engine.h" namespace paddle { + +DECLARE_int32(tensorrt_engine_batch_size); + namespace operators { +using inference::Singleton; +using inference::tensorrt::TRT_EngineManager; + class TensorRTEngineOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -32,9 +41,12 @@ class TensorRTEngineOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { + auto input0 = ctx.Inputs("Xs").front(); framework::OpKernelType kt = framework::OpKernelType( - framework::ToDataType( - ctx.Input("pre_ids")->type()), + framework::ToDataType(ctx.scope() + .FindVar(input0) + ->GetMutable() + ->type()), platform::CPUPlace()); return kt; } @@ -44,38 +56,47 @@ template class TensorRTEngineKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - if (!engine_) { + auto engine_name = context.Attr("engine_uniq_key"); + if (!Singleton::Global().HasEngine(engine_name)) { Prepare(context); } + auto* engine = Singleton::Global().Get(engine_name); auto input_names = context.op().Inputs("Xs"); PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs"); - // Try to determine a batch_size - auto* tensor0 = context.Input(input_names.front()); - PADDLE_ENFORCE_NOT_NULL(tensor0); - int batch_size = tensor0->dims()[0]; - PADDLE_ENFORCE_LE(batch_size, max_batch_); + PADDLE_ENFORCE_LE(FLAGS_tensorrt_engine_batch_size, + context.Attr("max_batch")); + std::vector output_maps = + context.Attr>("output_name_mapping"); + + auto params = context.Attr>("parameters"); + std::unordered_set parameters; + for (const auto& param : params) { + parameters.insert(param); + } // Convert input tensor from fluid to engine. for (const auto& x : context.Inputs("Xs")) { + if (parameters.count(x)) continue; // convert input and copy to TRT engine's buffer - auto* v = context.scope().FindVar(x); - PADDLE_ENFORCE_NOT_NULL(v, "no variable called %s", x); - auto& t = v->Get(); + auto& t = inference::analysis::GetFromScope( + context.scope(), x); if (platform::is_cpu_place(t.place())) { - engine_->SetInputFromCPU(x, static_cast(t.data()), - t.memory_size()); + engine->SetInputFromCPU(x, static_cast(t.data()), + t.memory_size()); } else { - engine_->SetInputFromGPU(x, static_cast(t.data()), - t.memory_size()); + engine->SetInputFromGPU(x, static_cast(t.data()), + t.memory_size()); } } // Execute the engine. - PADDLE_ENFORCE_GT(batch_size, 0); - engine_->Execute(batch_size); + PADDLE_ENFORCE_GT(FLAGS_tensorrt_engine_batch_size, 0); + engine->Execute(FLAGS_tensorrt_engine_batch_size); + // Convert output tensor from engine to fluid + int output_index = 0; for (const auto& y : context.Outputs("Ys")) { // convert output and copy to fluid. - nvinfer1::ITensor* trt_t = engine_->GetITensor(y); + nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]); auto dims = trt_t->getDimensions(); // Use the output ITensor's dims to reshape the Fluid Tensor. std::vector ddim(dims.d, dims.d + dims.nbDims); @@ -83,25 +104,32 @@ class TensorRTEngineKernel : public framework::OpKernel { auto* fluid_v = context.scope().FindVar(y); PADDLE_ENFORCE_NOT_NULL(fluid_v, "no output variable called %s", y); auto* fluid_t = fluid_v->GetMutable(); + fluid_t->Resize(framework::make_ddim(ddim)); - auto size = inference::analysis::AccuDims(dims.d, dims.nbDims); - if (platform::is_cpu_place(fluid_t->place())) { - engine_->GetOutputInCPU( - y, fluid_t->mutable_data(platform::CPUPlace()), size); - } else { - engine_->GetOutputInGPU( - y, fluid_t->mutable_data(platform::CUDAPlace()), size); - } + + // TODO(Superjomn) find some way to determine which device to output the + // tensor. + // if (platform::is_cpu_place(fluid_t->place())) { + // TODO(Superjomn) change this float to dtype size. + auto size = inference::analysis::AccuDims(dims.d, dims.nbDims) * + FLAGS_tensorrt_engine_batch_size; + engine->GetOutputInCPU(output_maps[output_index], + fluid_t->mutable_data(platform::CPUPlace()), + size * sizeof(float)); + //} else { + // engine->GetOutputInGPU( + // y, fluid_t->mutable_data(platform::CUDAPlace()), + // size * sizeof(float)); + //} + output_index += 1; } + + cudaStreamSynchronize(*engine->stream()); } protected: // Build the engine. void Prepare(const framework::ExecutionContext& context) const; - - private: - mutable std::unique_ptr engine_; - mutable int max_batch_{0}; }; } // namespace operators diff --git a/paddle/fluid/operators/tensorrt_engine_op_test.cc b/paddle/fluid/operators/tensorrt_engine_op_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..37657fa0b0498986fe67027415279af1775e58b9 --- /dev/null +++ b/paddle/fluid/operators/tensorrt_engine_op_test.cc @@ -0,0 +1,218 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +USE_CPU_ONLY_OP(tensorrt_engine); + +namespace paddle { +namespace operators { + +namespace { +void CreateCPUTensor(framework::Scope* scope, const std::string& name, + const std::vector& shape) { + auto* var = scope->Var(name); + auto* tensor = var->GetMutable(); + auto dims = framework::make_ddim(shape); + tensor->Resize(dims); + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + inference::tensorrt::RandomizeTensor(tensor, place, ctx); +} + +void AddTensorToBlockDesc(framework::proto::BlockDesc* block, + const std::string& name, + const std::vector& shape) { + using framework::proto::VarType; + auto* var = block->add_vars(); + framework::VarDesc desc(name); + desc.SetType(VarType::LOD_TENSOR); + desc.SetDataType(VarType::FP32); + desc.SetShape(shape); + *var = *desc.Proto(); +} + +} // namespace + +using inference::analysis::SetAttr; + +TEST(TensorRTEngineOp, manual) { + framework::ProgramDesc program; + auto* block_ = program.Proto()->add_blocks(); + block_->set_idx(0); + block_->set_parent_idx(-1); + + LOG(INFO) << "create block desc"; + framework::BlockDesc block_desc(&program, block_); + LOG(INFO) << "create fc op"; + auto* fc0 = block_desc.AppendOp(); + fc0->SetType("fc"); + fc0->SetInput("X", std::vector({"x"})); // 4 x 1 x 1 + fc0->SetInput("Y", std::vector({"y"})); // 4 x 6 + fc0->SetOutput("Out", std::vector({"z"})); // 6 x 1 x 1 + + LOG(INFO) << "create fc op"; + auto* fc1 = block_desc.AppendOp(); + fc1->SetType("fc"); + fc1->SetInput("X", std::vector({"z"})); + fc1->SetInput("Y", std::vector({"y0"})); // 6 x 8 + fc1->SetOutput("Out", std::vector({"z0"})); // 8 x 1 x 1 + + // Set inputs' variable shape in BlockDesc + // the batch size is 2, so the dims of 'x' is {2, 4, 1, 1} + AddTensorToBlockDesc(block_, "x", std::vector({2, 4, 1, 1})); + AddTensorToBlockDesc(block_, "y", std::vector({4, 6})); + AddTensorToBlockDesc(block_, "y0", std::vector({6, 8})); + AddTensorToBlockDesc(block_, "z", std::vector({2, 6})); + + // It is wired, need to copy manually. + *block_->add_ops() = *fc0->Proto(); + *block_->add_ops() = *fc1->Proto(); + + ASSERT_EQ(block_->ops_size(), 2); + + LOG(INFO) << "create tensorrt desc"; + framework::OpDesc engine_op_desc(nullptr); + engine_op_desc.SetType("tensorrt_engine"); + engine_op_desc.SetInput("Xs", std::vector({"x"})); + engine_op_desc.SetOutput("Ys", std::vector({"z0"})); + SetAttr(engine_op_desc.Proto(), "subgraph", + block_->SerializeAsString()); + SetAttr(engine_op_desc.Proto(), "max_batch", 100); + SetAttr(engine_op_desc.Proto(), "max_workspace", 1 << 10); + SetAttr(engine_op_desc.Proto(), "engine_uniq_key", "a_engine"); + SetAttr>(engine_op_desc.Proto(), "parameters", + std::vector({})); + SetAttr>(engine_op_desc.Proto(), + "output_name_mapping", + std::vector({"z0"})); + + LOG(INFO) << "create engine op"; + auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto()); + LOG(INFO) << "engine_op " << engine_op.get(); + + framework::Scope scope; + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + // Prepare variables. + CreateCPUTensor(&scope, "x", std::vector({2, 4})); + CreateCPUTensor(&scope, "y", std::vector({4, 6})); + CreateCPUTensor(&scope, "z", std::vector({2, 6})); + + CreateCPUTensor(&scope, "y0", std::vector({6, 8})); + CreateCPUTensor(&scope, "z0", std::vector({2, 8})); + + // Execute them. + LOG(INFO) << "engine_op run"; + engine_op->Run(scope, place); +} + +void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { + framework::ProgramDesc program; + framework::Scope scope; + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + + auto* block_ = program.Proto()->add_blocks(); + block_->set_idx(0); + block_->set_parent_idx(-1); + + using shape_t = std::vector; + + LOG(INFO) << "create block desc"; + framework::BlockDesc block_desc(&program, block_); + + auto AddFCLayer = [&](const std::string& x_name, const std::string& y_name, + const std::string& z_name, bool x_created, + const shape_t& x_shape, const shape_t& y_shape, + const shape_t& z_shape) { + LOG(INFO) << "create fc op"; + auto* fc = block_desc.AppendOp(); + fc->SetType("mul"); + fc->SetInput("X", std::vector({x_name})); + fc->SetInput("Y", std::vector({y_name})); + fc->SetOutput("Out", std::vector({z_name})); + + // Set inputs' variable shape in BlockDesc + if (!x_created) { + AddTensorToBlockDesc(block_, x_name, + std::vector({batch_size, input_dim, 1, 1})); + } + AddTensorToBlockDesc(block_, y_name, + std::vector({input_dim, output_dim})); + AddTensorToBlockDesc(block_, z_name, + std::vector({batch_size, output_dim})); + + // Prepare variables. + if (!x_created) { + CreateCPUTensor(&scope, x_name, std::vector(x_shape)); + } + CreateCPUTensor(&scope, y_name, std::vector(y_shape)); + CreateCPUTensor(&scope, z_name, std::vector(z_shape)); + + // It is wired, need to copy manually. + *block_->add_ops() = *fc->Proto(); + }; + + // Test with 4 layer FC + AddFCLayer("x0", "y0", "z0", false, {batch_size, input_dim}, + {input_dim, output_dim}, {batch_size, output_dim}); + AddFCLayer("z0", "y1", "z1", true, {}, {output_dim, output_dim}, + {batch_size, output_dim}); + AddFCLayer("z1", "y2", "z2", true, {}, {output_dim, output_dim}, + {batch_size, output_dim}); + AddFCLayer("z2", "y3", "z3", true, {}, {output_dim, output_dim}, + {batch_size, output_dim}); + + LOG(INFO) << "create tensorrt desc"; + framework::OpDesc engine_op_desc(nullptr); + engine_op_desc.SetType("tensorrt_engine"); + engine_op_desc.SetInput("Xs", std::vector({"x0"})); + engine_op_desc.SetOutput("Ys", std::vector({"z3"})); + + SetAttr(engine_op_desc.Proto(), "subgraph", + block_->SerializeAsString()); + SetAttr(engine_op_desc.Proto(), "max_batch", batch_size); + SetAttr(engine_op_desc.Proto(), "max_workspace", 2 << 10); + SetAttr>( + engine_op_desc.Proto(), "parameters", + std::vector({"y0", "y1", "y2", "y3"})); + SetAttr(engine_op_desc.Proto(), "engine_uniq_key", "b_engine"); + + SetAttr>(engine_op_desc.Proto(), + "output_name_mapping", + std::vector({"z3"})); + + auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto()); + + // Execute them. + engine_op->Run(scope, place); +} + +// Test with a larger FC layer. +TEST(TensorRTEngineOp, fc) { Execute(40, 28, 28); } + +} // namespace operators +} // namespace paddle + +USE_TRT_CONVERTER(fc) diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc index a845ba2eb038fa6a8e70dfbac06c31c19dbb9e3e..e2b7b6b8e447381229e4ad594b7974bc0aa159d5 100644 --- a/paddle/fluid/operators/test_send_nccl_id.cc +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -20,25 +20,28 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/detail/grpc_server.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/nccl_helper.h" #include "paddle/fluid/string/printf.h" +#ifdef PADDLE_WITH_GRPC +#include "paddle/fluid/operators/send_recv_util.h" +#endif + USE_NO_KERNEL_OP(listen_and_serv); namespace f = paddle::framework; namespace p = paddle::platform; namespace m = paddle::operators::math; -namespace detail = paddle::operators::detail; +namespace distributed = paddle::operators::distributed; namespace string = paddle::string; -std::unique_ptr g_rpc_service; -std::unique_ptr g_req_handler; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; void StartServer() { f::Scope scope; @@ -54,24 +57,23 @@ void StartServer() { g_req_handler->SetProgram(&empty_program); g_req_handler->SetExecutor(&executor); - g_rpc_service->RegisterRPC(detail::kRequestSend, g_req_handler.get()); + g_rpc_service->RegisterRPC(distributed::kRequestSend, g_req_handler.get()); g_req_handler->SetRPCServer(g_rpc_service.get()); std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get())); + std::bind(&distributed::RPCServer::StartServer, g_rpc_service.get())); - g_rpc_service->SetCond(detail::kRequestSend); - std::cout << "before WaitFanInOfSend" << std::endl; - g_rpc_service->WaitBarrier(detail::kRequestSend); + g_rpc_service->SetCond(distributed::kRequestSend); + g_rpc_service->WaitBarrier(distributed::kRequestSend); LOG(INFO) << "got nccl id and stop server..."; g_rpc_service->ShutDown(); server_thread.join(); } -TEST(SendNcclId, GrpcServer) { - g_req_handler.reset(new detail::RequestSendHandler(true)); - g_rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", 1)); +TEST(SendNcclId, RPCServer) { + g_req_handler.reset(new distributed::RequestSendHandler(true)); + g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); std::thread server_thread(StartServer); g_rpc_service->WaitServerReady(); @@ -88,12 +90,15 @@ TEST(SendNcclId, GrpcServer) { int port = g_rpc_service->GetSelectedPort(); std::string ep = string::Sprintf("127.0.0.1:%d", port); - detail::RPCClient client; + + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); + LOG(INFO) << "connect to server" << ep; - client.AsyncSendVariable(ep, dev_ctx, scope, NCCL_ID_VARNAME); - client.Wait(); - client.AsyncSendBatchBarrier(ep); - client.Wait(); + client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME); + client->Wait(); + client->AsyncSendBatchBarrier(ep); + client->Wait(); server_thread.join(); g_rpc_service.reset(nullptr); diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc index c17d1afc309c65035063348d4934ea1783b018ed..4a8ac441cfaf642fde58ee30865a22e83c065498 100644 --- a/paddle/fluid/operators/top_k_op.cc +++ b/paddle/fluid/operators/top_k_op.cc @@ -50,7 +50,7 @@ class TopkOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor) The input of Topk op"); - AddOutput("Out", "(Tensor) The output tensor of Topk op"); + AddOutput("Out", "(Tensor) The output tensor of Topk op").Reuse("X"); AddOutput("Indices", "(Tensor) The indices of Topk elements of input"); AddComment(R"DOC( Top K operator diff --git a/paddle/fluid/operators/top_k_op.h b/paddle/fluid/operators/top_k_op.h index 7ddb82ef6ff063868a4b9b603b8ab89700b9dd13..054dd481994d03f71b0ed5dc73e103085f6c91aa 100644 --- a/paddle/fluid/operators/top_k_op.h +++ b/paddle/fluid/operators/top_k_op.h @@ -60,6 +60,7 @@ class TopkKernel : public framework::OpKernel { #endif for (size_t i = 0; i < row; i++) { std::vector> vec; + vec.reserve(col); for (size_t j = 0; j < col; j++) { vec.push_back(std::pair(eg_input(i, j), j)); } diff --git a/paddle/fluid/operators/uniform_random_batch_size_like_op.cc b/paddle/fluid/operators/uniform_random_batch_size_like_op.cc index 78fee77df8151221459b0afa0d6789bfe82cfda5..75d6181749e4e9bd81a3c02de69caf0acd81eef9 100644 --- a/paddle/fluid/operators/uniform_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/uniform_random_batch_size_like_op.cc @@ -35,10 +35,10 @@ class UniformRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { protected: void Apply() override { AddComment(R"DOC( -Uniform random operator +UniformRandomBatchSizeLike operator. This operator initializes a tensor with the same batch_size as the Input tensor - with random values sampled from a uniform distribution. +with random values sampled from a uniform distribution. )DOC"); AddAttr("min", diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index 137ea91caedabc3167146d91b063dbe9e2e2b931..edd1baa4ace4e246190afcd12b0716f1dd38e243 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -86,32 +86,24 @@ class UniformRandomOp : public framework::OperatorWithKernel { class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddOutput("Out", "(Tensor) The output tensor of uniform random op"); + AddOutput("Out", "The output tensor of uniform random op"); AddComment(R"DOC( -Uniform random operator. - This operator initializes a tensor with random values sampled from a -uniform distribution. +uniform distribution. The random result is in set [min, max]. )DOC"); - AddAttr>("shape", - "(vector) The shape of the output tensor"); - AddAttr("min", - "(float, default -1.0) " - "Minimum value of uniform random") + AddAttr>("shape", "The shape of the output tensor"); + AddAttr("min", "Minimum value of uniform random. [default -1.0].") .SetDefault(-1.0f); - AddAttr("max", - "(float, default 1.0) " - "Maximun value of uniform random") + AddAttr("max", "Maximun value of uniform random. [default 1.0].") .SetDefault(1.0f); AddAttr("seed", - "(int, default 0) " "Random seed used for generating samples. " "0 means use a seed generated by the system." "Note that if seed is not 0, this operator will always " - "generate the same random numbers every time.") + "generate the same random numbers every time. [default 0].") .SetDefault(0); - AddAttr("dtype", "(int, default 5(FP32)) Output tensor data type") + AddAttr("dtype", "Output tensor data type. [default 5(FP32)].") .SetDefault(framework::proto::VarType::FP32); } }; diff --git a/paddle/fluid/operators/unsqueeze_op.cc b/paddle/fluid/operators/unsqueeze_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f2a15fdf572e0de30f9949dda5020e130b0c5585 --- /dev/null +++ b/paddle/fluid/operators/unsqueeze_op.cc @@ -0,0 +1,191 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class UnsqueezeOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of UnsqueezeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of UnsqueezeOp should not be null."); + + const auto &axes = ctx->Attrs().Get>("axes"); + const auto &x_dims = ctx->GetInputDim("X"); + // Validity Check: input tensor dims (<6). + PADDLE_ENFORCE(x_dims.size() <= 6, + "Invalid dimensions, the rank of Input(X) " + "should be in the range of [1, 6] (Eigen limit)"); + auto out_dims = GetOutputShape(axes, x_dims); + ctx->SetOutputDim("Out", out_dims); + if (x_dims[0] == out_dims[0]) { + // Only pass LoD when the first dimension of output and Input(X) + // are the same. + ctx->ShareLoD("X", "Out"); + } + } + + static framework::DDim GetOutputShape(const std::vector unsqz_dims, + const framework::DDim &in_dims) { + int output_size = in_dims.size() + static_cast(unsqz_dims.size()); + int cur_output_size = in_dims.size(); + std::vector output_shape(output_size, 0); + + // Validity Check: rank range. + PADDLE_ENFORCE(output_size <= 6, + "The output tensor's rank should be less than 6."); + + for (int axis : unsqz_dims) { + int cur = axis < 0 ? axis + cur_output_size + 1 : axis; + // Vaildity Check: the axis bound + PADDLE_ENFORCE( + cur >= 0 && cur <= cur_output_size, + "The unsqueeze dims must be within range of current rank."); + // Move old axis, and insert new axis + for (int i = cur_output_size; i >= cur; --i) { + if (output_shape[i] == 1) { + // Move axis + output_shape[i + 1] = 1; + output_shape[i] = 0; + } + } + output_shape[cur] = 1; + // Add the output size. + cur_output_size++; + } + + // Make output shape + for (int in_idx = 0, out_idx = 0; out_idx < output_size; ++out_idx) { + if (output_shape[out_idx] == 0) { + output_shape[out_idx] = in_dims[in_idx++]; + } + } + + return framework::make_ddim(output_shape); + } +}; + +class UnsqueezeOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto &axes = Attr>("axes"); + auto x_dims = scope.FindVar(Input("X"))->Get().dims(); + auto out_dims = UnsqueezeOpInferShape::GetOutputShape(axes, x_dims); + + framework::AttributeMap attrs; + attrs["shape"] = framework::vectorize2int(out_dims); + attrs["inplace"] = Attr("inplace"); + // Invoke Reshape op. + auto reshape_op = framework::OpRegistry::CreateOp( + "reshape", {{"X", {Input("X")}}, {"Shape", {}}}, + {{"Out", {Output("Out")}}}, attrs); + reshape_op->Run(scope, place); + } +}; + +class UnsqueezeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor). The input tensor of unsqueeze operator."); + AddOutput("Out", "(Tensor). The output tensor of unsqueeze operator."); + AddAttr>("axes", + "(std::vector). List of integers," + " indicating the dimensions to be inserted") + .AddCustomChecker([](const std::vector &axes) { + PADDLE_ENFORCE(!axes.empty(), + "Invalid axes, The unsqueeze axes is empty."); + // Validity Check: axes dims (<6). + PADDLE_ENFORCE(static_cast(axes.size()) < 6, + "Invalid dimensions, dynamic dimensions should be " + "within [1, 6] dimensions (Eigen limit)."); + // Validity Check: the range of unsqueeze aixs. + for (int axis : axes) { + PADDLE_ENFORCE(axis < 6, + "Invalid dimensions, input axis should be" + " within [1, 6] dimensions (Eigen limit)."); + } + }); + AddAttr( + "inplace", + "(default: false) Unsqueeze the source tensor's shape without " + "memory copy. When Attr(inplace) is set true, the output " + "tensor shares memory with Input(X), otherwise, a new output " + "tensor is created, and its data are copied from Input(x).") + .SetDefault(false); + AddComment(R"DOC( + Unsqueeze Operator. + + Insert single-dimensional entries to the shape of a tensor. + Takes one required argument axes, a list of dimensions that will be inserted. + Dimension indices in axes are as seen in the output tensor. + + For example: + Given a tensor such that tensor with shape [3, 4, 5], + then Unsqueeze(tensor, axes=[0, 4]) has shape [1, 3, 4, 5, 1] + )DOC"); + } +}; + +class UnsqueezeGradInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->ShareLoD("X", framework::GradVarName("X")); + } +}; + +class UnsqueezeGradOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto dx_name = Output(framework::GradVarName("X")); + auto dout_name = Input(framework::GradVarName("Out")); + auto x_dims = scope.FindVar(Input("X"))->Get().dims(); + + framework::AttributeMap attrs; + attrs["shape"] = framework::vectorize2int(x_dims); + attrs["inplace"] = Attr("inplace"); + + auto reshape_op = framework::OpRegistry::CreateOp( + "reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}}, + attrs); + reshape_op->Run(scope, place); + } +}; + +} // namespace operators +} // namespace paddle + +// Tell linker to use reshape op. +USE_OP(reshape); + +namespace ops = paddle::operators; +REGISTER_OPERATOR(unsqueeze, ops::UnsqueezeOp, ops::UnsqueezeOpMaker, + ops::UnsqueezeOpInferShape, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(unsqueeze_grad, ops::UnsqueezeGradOp, + ops::UnsqueezeGradInferShape); diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 175c3ac5d79f24e47d21417df8e3eaeb4d5b2335..733157ea05ed39434b9a750e3a94ea548f512ce6 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/operators/detail/safe_ref.h" namespace paddle { @@ -135,15 +136,14 @@ class WhileGradOp : public framework::OperatorBase { auto &og_inside = detail::Ref(cur_scope.Var(inside_og_name), "Cannot find inside gradient %s", inside_og_name); - if (og_outside.Type().hash_code() == - typeid(framework::LoDTensor).hash_code()) { + if (framework::IsType(og_outside.Type())) { auto &outside_tensor = og_outside.Get(); auto &inside_tensor = detail::Ref(og_inside.GetMutable()); inside_tensor.set_lod(outside_tensor.lod()); inside_tensor.ShareDataWith(outside_tensor); - } else if (og_outside.Type().hash_code() == - typeid(framework::LoDTensorArray).hash_code()) { + } else if (framework::IsType( + og_outside.Type())) { auto &outside_array = og_outside.Get(); auto &inside_array = detail::Ref(og_inside.GetMutable()); @@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase { ->set_lod(inside_tensor.lod()); } } - auto new_inside_name = cur_scope.Rename(inside_grad_name); auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {pg_names[param_id], new_inside_name}}}, - {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{}); + {{"Out", {pg_names[param_id]}}}, + framework::AttributeMap{{"use_mkldnn", {false}}}); sum_op->Run(cur_scope, dev_place); cur_scope.Rename(new_inside_name, inside_grad_name); } diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index b29035bafd34fa81dc6b59691142fe74439202b8..f08c0e8e345179bf198ca9d50278b7f65e03ca2c 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -18,7 +18,11 @@ else() endif() cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece enforce) -cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog enforce) +set(CPU_INFO_DEPS gflags glog enforce) +IF(WITH_XBYAK) + list(APPEND CPU_INFO_DEPS xbyak) +ENDIF() +cc_library(cpu_info SRCS cpu_info.cc DEPS ${CPU_INFO_DEPS}) cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info) nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce) @@ -28,6 +32,9 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags) add_subdirectory(dynload) +cc_library(cpu_helper SRCS cpu_helper.cc DEPS cblas enforce) +cc_test(cpu_helper_test SRCS cpu_helper_test.cc DEPS cpu_helper) + IF(WITH_GPU) set(GPU_CTX_DEPS dynload_cuda dynamic_loader) ELSE() @@ -42,10 +49,12 @@ ENDIF() # memcpy depends on device_context, here add deps individually for # avoiding cycle dependencies -cc_library(device_context SRCS device_context.cc DEPS malloc - place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS}) +cc_library(device_context SRCS device_context.cc init.cc DEPS malloc + place eigen3 stringpiece cpu_helper framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS}) nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info) +cc_test(init_test SRCS init_test.cc DEPS device_context) + nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context) @@ -53,5 +62,9 @@ cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framewo cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) cc_test(profiler_test SRCS profiler_test.cc DEPS profiler) -nv_test(float16_gpu_test SRCS float16_test.cu) -cc_test(float16_test SRCS float16_test.cc) +nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor) +cc_test(float16_test SRCS float16_test.cc DEPS lod_tensor) + +IF(WITH_GPU) + nv_test(cuda_helper_test SRCS cuda_helper_test.cu) +ENDIF() diff --git a/paddle/fluid/platform/assert.h b/paddle/fluid/platform/assert.h index 123d3598f4f4753f70889e415aff0f41b7d212f7..2ce9b31bb81de867ff4ed6ee14afddecd95317b9 100644 --- a/paddle/fluid/platform/assert.h +++ b/paddle/fluid/platform/assert.h @@ -17,7 +17,7 @@ limitations under the License. */ #define STRINGIFY(x) #x #define TOSTRING(x) STRINGIFY(x) -#if defined(__APPLE__) && defined(__CUDA_ARCH__) && !defined(NDEBUG) +#if defined(__CUDA_ARCH__) #include #define PADDLE_ASSERT(e) \ do { \ @@ -38,6 +38,9 @@ limitations under the License. */ } while (0) #else #include -#define PADDLE_ASSERT(e) assert(e) +// For cuda, the assertions can affect performance and it is therefore +// recommended to disable them in production code +// https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#assertion +#define PADDLE_ASSERT(e) assert((e)) #define PADDLE_ASSERT_MSG(e, m) assert((e) && (m)) #endif diff --git a/paddle/fluid/platform/cpu_helper.cc b/paddle/fluid/platform/cpu_helper.cc new file mode 100644 index 0000000000000000000000000000000000000000..234a04b5c2eb5ee643e8a4e723b28331cd8e6ee0 --- /dev/null +++ b/paddle/fluid/platform/cpu_helper.cc @@ -0,0 +1,44 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/platform/cpu_helper.h" +#include "paddle/fluid/platform/enforce.h" + +#ifdef PADDLE_WITH_MKLML +#include +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef PADDLE_USE_OPENBLAS +#include +#endif + +namespace paddle { +namespace platform { + +void SetNumThreads(int num_threads) { +#ifdef PADDLE_USE_OPENBLAS + int real_num_threads = num_threads > 1 ? num_threads : 1; + openblas_set_num_threads(real_num_threads); +#elif defined(PADDLE_WITH_MKLML) + int real_num_threads = num_threads > 1 ? num_threads : 1; + platform::dynload::MKL_Set_Num_Threads(real_num_threads); + omp_set_num_threads(num_threads); +#else + PADDLE_ENFORCE(false, "To be implemented."); +#endif +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/cpu_helper.h b/paddle/fluid/platform/cpu_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..78fc392b632ef92d4ae08de2051041fc0bf6778b --- /dev/null +++ b/paddle/fluid/platform/cpu_helper.h @@ -0,0 +1,26 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +namespace paddle { +namespace platform { + +//! Set the number of threads in use. +void SetNumThreads(int num_threads); + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/cpu_helper_test.cc b/paddle/fluid/platform/cpu_helper_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..dc1b2b56cd98ca6259c46a76231dbc99482970c1 --- /dev/null +++ b/paddle/fluid/platform/cpu_helper_test.cc @@ -0,0 +1,22 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/platform/cpu_helper.h" + +#include "gtest/gtest.h" + +TEST(CpuHelper, SetNumThread) { + paddle::platform::SetNumThreads(1); + paddle::platform::SetNumThreads(4); +} diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc index 4fc9aae8e36e9b43d65fab0b92ec3a2549057128..7d53a684d6068c79659719159696ef5aebfeaa2b 100644 --- a/paddle/fluid/platform/cpu_info.cc +++ b/paddle/fluid/platform/cpu_info.cc @@ -14,6 +14,11 @@ limitations under the License. */ #include "paddle/fluid/platform/cpu_info.h" +#ifdef PADDLE_WITH_XBYAK +#include "xbyak/xbyak.h" +#include "xbyak/xbyak_util.h" +#endif + #ifdef __APPLE__ #include #include @@ -21,12 +26,23 @@ limitations under the License. */ #include #endif +#include #include "gflags/gflags.h" DEFINE_double(fraction_of_cpu_memory_to_use, 1, "Default use 100% of CPU memory for PaddlePaddle," "reserve the rest for page tables, etc"); +DEFINE_uint64(initial_cpu_memory_in_mb, +#ifdef PADDLE_WITH_MKLDNN + /* Aligned with mozga-intel, MKLDNN need at least 5000 MB + * to obtain the best performance*/ + 5000, +#else + 500, +#endif + "Initial CPU memory for PaddlePaddle, in MD unit."); + DEFINE_double( fraction_of_cuda_pinned_memory_to_use, 0.5, "Default use 50% of CPU memory as the pinned_memory for PaddlePaddle," @@ -63,8 +79,11 @@ size_t CpuMinChunkSize() { } size_t CpuMaxChunkSize() { - // Allow to allocate the maximum chunk size is roughly 3% of CPU memory. - return CpuMaxAllocSize() / 32; + // Allow to allocate the maximum chunk size is roughly 3% of CPU memory, + // or the initial_cpu_memory_in_mb. + return std::min( + static_cast(CpuMaxAllocSize() / 32), + static_cast(FLAGS_initial_cpu_memory_in_mb * 1 << 20)); } size_t CUDAPinnedMaxAllocSize() { @@ -84,5 +103,39 @@ size_t CUDAPinnedMaxChunkSize() { return CUDAPinnedMaxAllocSize() / 256; } +#ifdef PADDLE_WITH_XBYAK +namespace jit { + +static Xbyak::util::Cpu cpu; +bool MayIUse(const cpu_isa_t cpu_isa) { + using namespace Xbyak::util; // NOLINT + switch (cpu_isa) { + case sse42: + return cpu.has(Cpu::tSSE42); + case avx2: + return cpu.has(Cpu::tAVX2); + case avx512_common: + return cpu.has(Cpu::tAVX512F); + case avx512_core: + return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) && + cpu.has(Cpu::tAVX512VL) && cpu.has(Cpu::tAVX512DQ); + case avx512_core_vnni: + return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) && + cpu.has(Cpu::tAVX512VL) && cpu.has(Cpu::tAVX512DQ) && + cpu.has(Cpu::tAVX512_VNNI); + case avx512_mic: + return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512CD) && + cpu.has(Cpu::tAVX512ER) && cpu.has(Cpu::tAVX512PF); + case avx512_mic_4ops: + return true && MayIUse(avx512_mic) && cpu.has(Cpu::tAVX512_4FMAPS) && + cpu.has(Cpu::tAVX512_4VNNIW); + case isa_any: + return true; + } + return false; +} + +} // namespace jit +#endif } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/cpu_info.h b/paddle/fluid/platform/cpu_info.h index f06c2b67fe4385f427322e9bb2f3080fdd3acc94..f5f67667594f1ab80058533e4c5d5b04c2592b60 100644 --- a/paddle/fluid/platform/cpu_info.h +++ b/paddle/fluid/platform/cpu_info.h @@ -37,5 +37,25 @@ size_t CUDAPinnedMinChunkSize(); //! Get the maximum chunk size for buddy allocator. size_t CUDAPinnedMaxChunkSize(); +#ifdef PADDLE_WITH_XBYAK +namespace jit { + +typedef enum { + isa_any, + sse42, + avx2, + avx512_common, + avx512_core, + avx512_core_vnni, + avx512_mic, + avx512_mic_4ops, +} cpu_isa_t; // Instruction set architecture + +// May I use some instruction +inline bool MayIUse(const cpu_isa_t cpu_isa); + +} // namespace jit +#endif + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h index ecec4178f2d9937920e52eb74bf9068b84e741a0..23457ff5fe1ec27094113ba0dde26adc64c716b5 100644 --- a/paddle/fluid/platform/cuda_device_function.h +++ b/paddle/fluid/platform/cuda_device_function.h @@ -14,6 +14,10 @@ limitations under the License. */ #pragma once #include +// NOTE(): support float16 to half in header file. +#define PADDLE_CUDA_FP16 +#include +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace platform { @@ -36,6 +40,18 @@ __forceinline__ __device__ T CudaShuffleDownSync(unsigned mask, T val, #endif } +// CUDA 9.0 have native compatible float16 shfl_down +#if CUDA_VERSION < 9000 +template <> +__forceinline__ __device__ float16 CudaShuffleDownSync(unsigned mask, + float16 val, int delta, + int width) { + half tmp = static_cast(val); + __shfl_down(tmp, static_cast(delta), width); + return float16(tmp); +} +#endif + template __forceinline__ __device__ T CudaShuffleSync(unsigned mask, T val, int src_line, int width = 32) { @@ -46,6 +62,11 @@ __forceinline__ __device__ T CudaShuffleSync(unsigned mask, T val, int src_line, #endif } +template +HOSTDEVICE T Infinity() { + return INFINITY; +} + template __device__ T reduceSum(T val, int tid, int len) { // NOTE(zcd): The warp size should be taken from the diff --git a/paddle/fluid/platform/cuda_helper_test.cu b/paddle/fluid/platform/cuda_helper_test.cu new file mode 100644 index 0000000000000000000000000000000000000000..ca5ca1caeb23f01c047feeccf9c276b2dcd1cb68 --- /dev/null +++ b/paddle/fluid/platform/cuda_helper_test.cu @@ -0,0 +1,153 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include + +#define PADDLE_CUDA_FP16 +#include "paddle/fluid/platform/cuda_device_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/float16.h" + +using paddle::platform::PADDLE_CUDA_NUM_THREADS; +using paddle::platform::float16; + +template +__global__ void AddKernel(const T* data_a, T* data_b, size_t num) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + paddle::platform::CudaAtomicAdd(&data_b[i], data_a[i]); + } +} + +template +struct AddFunctor { + T operator()(const T& a, const T& b) { return a + b; } +}; + +template +void TestCase(size_t num) { + T *in1, *in2, *out; + T *d_in1, *d_in2; + size_t size = sizeof(T) * num; + cudaMalloc(reinterpret_cast(&d_in1), size); + cudaMalloc(reinterpret_cast(&d_in2), size); + in1 = reinterpret_cast(malloc(size)); + in2 = reinterpret_cast(malloc(size)); + out = reinterpret_cast(malloc(size)); + std::minstd_rand engine; + std::uniform_real_distribution dist(0.0, 1.0); + for (size_t i = 0; i < num; ++i) { + in1[i] = static_cast(dist(engine)); + in2[i] = static_cast(dist(engine)); + } + cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); + AddKernel<<<1, PADDLE_CUDA_NUM_THREADS>>>(d_in1, d_in2, num); + cudaDeviceSynchronize(); + cudaMemcpy(out, d_in2, size, cudaMemcpyDeviceToHost); + cudaDeviceSynchronize(); + for (size_t i = 0; i < num; ++i) { + // NOTE(dzhwinter): the float16 add has small underflow/overflow + // so we use EXPECT_NEAR to check the result. + EXPECT_NEAR(static_cast(out[i]), + static_cast(AddFunctor()(in1[i], in2[i])), 0.001); + } + free(in1); + free(in2); + free(out); + cudaFree(d_in1); + cudaFree(d_in2); +} + +// cuda primitives +TEST(CudaAtomic, Add) { + TestCase(static_cast(10)); + TestCase(static_cast(1024 * 1024)); + + TestCase(static_cast(10)); + TestCase(static_cast(1024 * 1024)); +} + +TEST(CudaAtomic, float16) { + TestCase(static_cast(1)); + TestCase(static_cast(2)); + TestCase(static_cast(3)); + + TestCase(static_cast(10)); + TestCase(static_cast(1024 * 1024)); +} + +// unalignment of uint8 +void TestUnalign(size_t num, const int shift_bit) { + PADDLE_ENFORCE(num % 2 == 0, "must be a multiple of 2"); + float16 *in1, *in2, *out; + float16 *d_in1, *d_in2; + size_t size = sizeof(uint8_t) * (num + shift_bit); + size_t array_size = sizeof(float16) * (num / 2); + + cudaMalloc(reinterpret_cast(&d_in1), size); + cudaMalloc(reinterpret_cast(&d_in2), size); + in1 = reinterpret_cast(malloc(size)); + in2 = reinterpret_cast(malloc(size)); + out = reinterpret_cast(malloc(size)); + + // right shift 1, mimic the unalignment of address + float16* r_in1 = + reinterpret_cast(reinterpret_cast(in1) + shift_bit); + float16* r_in2 = + reinterpret_cast(reinterpret_cast(in2) + shift_bit); + + std::minstd_rand engine; + std::uniform_real_distribution dist(0.0, 1.0); + for (size_t i = 0; i < num / 2; ++i) { + r_in1[i] = static_cast(dist(engine)); + r_in2[i] = static_cast(dist(engine)); + } + cudaMemcpy(d_in1, r_in1, array_size, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, r_in2, array_size, cudaMemcpyHostToDevice); + AddKernel<<<1, PADDLE_CUDA_NUM_THREADS>>>(d_in1, d_in2, num / 2); + cudaDeviceSynchronize(); + cudaMemcpy(out, d_in2, array_size, cudaMemcpyDeviceToHost); + cudaDeviceSynchronize(); + for (size_t i = 0; i < num / 2; ++i) { + // NOTE(dzhwinter): the float16 add has small underflow/overflow + // so we use EXPECT_NEAR to check the result. + EXPECT_NEAR(static_cast(out[i]), + static_cast(AddFunctor()(r_in1[i], r_in2[i])), + 0.001); + } + free(in1); + free(in2); + free(out); + cudaFree(d_in1); + cudaFree(d_in2); +} + +TEST(CudaAtomic, float16Unalign) { + // same with float16 testcase + TestUnalign(static_cast(2), /*shift_bit*/ 2); + TestUnalign(static_cast(1024), /*shift_bit*/ 2); + TestUnalign(static_cast(1024 * 1024), /*shift_bit*/ 2); + + // shift the address. + TestUnalign(static_cast(2), /*shift_bit*/ 1); + TestUnalign(static_cast(1024), /*shift_bit*/ 1); + TestUnalign(static_cast(1024 * 1024), /*shift_bit*/ 1); + + TestUnalign(static_cast(2), /*shift_bit*/ 3); + TestUnalign(static_cast(1024), /*shift_bit*/ 3); + TestUnalign(static_cast(1024 * 1024), /*shift_bit*/ 3); +} diff --git a/paddle/fluid/platform/cuda_primitives.h b/paddle/fluid/platform/cuda_primitives.h index d535ed2f89df6a0b311ec068ecd92c8e3183cee7..67ea64833d3b844d88a2e5996f860ef165bd8ffd 100644 --- a/paddle/fluid/platform/cuda_primitives.h +++ b/paddle/fluid/platform/cuda_primitives.h @@ -14,12 +14,14 @@ limitations under the License. */ #pragma once #include +#include +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace platform { #define CUDA_ATOMIC_WRAPPER(op, T) \ - __device__ __forceinline__ T CudaAtomic##op(T* address, const T val) + __device__ __forceinline__ T CudaAtomic##op(T *address, const T val) #define USE_CUDA_ATOMIC(op, T) \ CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); } @@ -42,17 +44,17 @@ CUDA_ATOMIC_WRAPPER(Add, int64_t) { static_assert(sizeof(int64_t) == sizeof(long long int), // NOLINT "long long should be int64"); return CudaAtomicAdd( - reinterpret_cast(address), // NOLINT - static_cast(val)); // NOLINT + reinterpret_cast(address), // NOLINT + static_cast(val)); // NOLINT } #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600 USE_CUDA_ATOMIC(Add, double); #else CUDA_ATOMIC_WRAPPER(Add, double) { - unsigned long long int* address_as_ull = // NOLINT - reinterpret_cast(address); // NOLINT - unsigned long long int old = *address_as_ull, assumed; // NOLINT + unsigned long long int *address_as_ull = // NOLINT + reinterpret_cast(address); // NOLINT + unsigned long long int old = *address_as_ull, assumed; // NOLINT do { assumed = old; @@ -64,6 +66,67 @@ CUDA_ATOMIC_WRAPPER(Add, double) { return __longlong_as_double(old); } +#endif + +#ifdef PADDLE_CUDA_FP16 +// NOTE(dzhwinter): cuda do not have atomicCAS for half. +// Just use the half address as a unsigned value address and +// do the atomicCAS. According to the value store at high 16 bits +// or low 16 bits, then do a different sum and CAS. +// Given most warp-threads will failed on the atomicCAS, so this +// implemented should be avoided in high concurrency. It's will be +// slower than the way convert value into 32bits and do a full atomicCAS. + +// convert the value into float and do the add arithmetic. +// then store the result into a uint32. +inline static __device__ uint32_t add_to_low_half(uint32_t val, float x) { + float16 low_half; + // the float16 in lower 16bits + low_half.x = static_cast(val & 0xFFFFu); + low_half = static_cast(static_cast(low_half) + x); + return (val & 0xFFFF0000u) | low_half.x; +} + +inline static __device__ uint32_t add_to_high_half(uint32_t val, float x) { + float16 high_half; + // the float16 in higher 16bits + high_half.x = static_cast(val >> 16); + high_half = static_cast(static_cast(high_half) + x); + return (val & 0xFFFFu) | (static_cast(high_half.x) << 16); +} + +CUDA_ATOMIC_WRAPPER(Add, float16) { + // concrete packed float16 value may exsits in lower or higher 16bits + // of the 32bits address. + uint32_t *address_as_ui = reinterpret_cast( + reinterpret_cast(address) - + (reinterpret_cast(address) & 0x02)); + float val_f = static_cast(val); + uint32_t old = *address_as_ui; + uint32_t sum; + uint32_t newval; + uint32_t assumed; + if (((uintptr_t)address & 0x02) == 0) { + // the float16 value stay at lower 16 bits of the address. + do { + assumed = old; + old = atomicCAS(address_as_ui, assumed, add_to_low_half(assumed, val_f)); + } while (old != assumed); + float16 ret; + ret.x = old & 0xFFFFu; + return ret; + } else { + // the float16 value stay at higher 16 bits of the address. + do { + assumed = old; + old = atomicCAS(address_as_ui, assumed, add_to_high_half(assumed, val_f)); + } while (old != assumed); + float16 ret; + ret.x = old >> 16; + return ret; + } +} + #endif } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h index c0d399d078f73743836fc2a0c1d4b1e6b31ecd83..bb8b14bb9fa41942c3aa653ca224c0842fbf9a00 100644 --- a/paddle/fluid/platform/cudnn_helper.h +++ b/paddle/fluid/platform/cudnn_helper.h @@ -22,6 +22,8 @@ limitations under the License. */ #include "paddle/fluid/platform/float16.h" #include "paddle/fluid/platform/macros.h" +DECLARE_bool(cudnn_deterministic); + namespace paddle { namespace platform { @@ -57,13 +59,12 @@ inline const char* cudnnGetErrorString(cudnnStatus_t status) { #define CUDNN_VERSION_MIN(major, minor, patch) \ (CUDNN_VERSION >= ((major)*1000 + (minor)*100 + (patch))) -#define CUDNN_ENFORCE(condition) \ - do { \ - cudnnStatus_t status = condition; \ - if (status != CUDNN_STATUS_SUCCESS) { \ - VLOG(1) << ::paddle::platform::cudnnGetErrorString(status); \ - PADDLE_THROW("cuDNN call failed"); \ - } \ +#define CUDNN_ENFORCE(condition) \ + do { \ + cudnnStatus_t status = condition; \ + if (UNLIKELY(status != CUDNN_STATUS_SUCCESS)) { \ + PADDLE_THROW(::paddle::platform::cudnnGetErrorString(status)); \ + } \ } while (false) enum class DataLayout { // Not use @@ -76,8 +77,44 @@ enum class DataLayout { // Not use enum class PoolingMode { kMaximum, kAverage, + kMaximumDeterministic, }; +#if CUDNN_VERSION < 6000 +#pragma message "CUDNN version under 6.0 is supported at best effort." +#pragma message "We strongly encourage you to move to 6.0 and above." +#pragma message "This message is intended to annoy you enough to update." +#pragma message \ + "please see https://docs.nvidia.com/deeplearning/sdk/cudnn-release-notes/" + +inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { + switch (mode) { + case PoolingMode::kMaximumDeterministic: + return CUDNN_POOLING_MAX; + case PoolingMode::kAverage: + return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kMaximum: + return CUDNN_POOLING_MAX; + default: + PADDLE_THROW("Unexpected pooling mode."); + } +} +#else + +inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { + switch (mode) { + case PoolingMode::kMaximumDeterministic: + return CUDNN_POOLING_MAX_DETERMINISTIC; + case PoolingMode::kAverage: + return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kMaximum: + return CUDNN_POOLING_MAX; + default: + PADDLE_THROW("Unexpected pooling mode."); + } +} +#endif // CUDNN_VERSION < 6000 + template class CudnnDataType; @@ -293,9 +330,7 @@ class ScopedPoolingDescriptor { PADDLE_ENFORCE_EQ(kernel.size(), pads.size()); PADDLE_ENFORCE_EQ(kernel.size(), strides.size()); PADDLE_ENFORCE(dynload::cudnnSetPoolingNdDescriptor( - desc_, (mode == PoolingMode::kMaximum - ? CUDNN_POOLING_MAX - : CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING), + desc_, (GetPoolingMode(mode)), CUDNN_PROPAGATE_NAN, // Always propagate nans. kernel.size(), kernel.data(), pads.data(), strides.data())); return desc_; diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index 1f733d71bdfb777d4a2f316a5fefc3c874879862..2cc26da013f59f5b7ee1747d57baca9c1c0efe2c 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -10,6 +10,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/platform/device_context.h" +#include #include #include #include @@ -35,7 +36,7 @@ DeviceContextPool::DeviceContextPool( const std::vector& places) { PADDLE_ENFORCE_GT(places.size(), 0); using PtrType = std::unique_ptr; - std::unordered_set set; + std::set set; for (auto& p : places) { set.insert(p); } @@ -175,7 +176,6 @@ CUDADeviceContext::~CUDADeviceContext() { Place CUDADeviceContext::GetPlace() const { return place_; } void CUDADeviceContext::Wait() const { - std::lock_guard guard(mutex_); PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); PADDLE_ENFORCE(cudaGetLastError()); } diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index a9c1984616bc731e0557f2cb89282423aa9c3bac..88e0383146c1adf2752a362091996bad9cfcce5e 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -11,6 +11,7 @@ limitations under the License. */ #pragma once #include +#include // NOLINT #include #include #include @@ -26,12 +27,12 @@ limitations under the License. */ #include #endif +#include +#include "glog/logging.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/place.h" #include "unsupported/Eigen/CXX11/Tensor" -#include "glog/logging.h" - namespace paddle { namespace platform { @@ -100,7 +101,7 @@ class CUDADeviceContext : public DeviceContext { template void RecordEvent(cudaEvent_t ev, Callback callback) { - std::lock_guard guard(mutex_); + std::lock_guard guard(mtx_); callback(); PADDLE_ENFORCE(cudaEventRecord(ev, stream_)); } @@ -110,8 +111,6 @@ class CUDADeviceContext : public DeviceContext { std::unique_ptr eigen_device_; std::unique_ptr eigen_stream_; - - mutable std::recursive_mutex mutex_; cudaStream_t stream_; cudnnHandle_t cudnn_handle_; cublasHandle_t cublas_handle_; @@ -119,6 +118,8 @@ class CUDADeviceContext : public DeviceContext { int compute_capability; int multi_process; int max_threads_per_mp; + + std::mutex mtx_; }; template <> @@ -200,9 +201,7 @@ class DeviceContextPool { private: static DeviceContextPool* pool; - std::unordered_map, PlaceHash> - device_contexts_; + std::map> device_contexts_; DISABLE_COPY_AND_ASSIGN(DeviceContextPool); }; diff --git a/paddle/fluid/platform/device_context_test.cu b/paddle/fluid/platform/device_context_test.cu index fa806aba6d8747beebc3eed2c661b326dd62fd76..171d2979a0218ad5e22112190a59866b3e0b617f 100644 --- a/paddle/fluid/platform/device_context_test.cu +++ b/paddle/fluid/platform/device_context_test.cu @@ -69,19 +69,3 @@ TEST(Device, DeviceContextPool) { ASSERT_NE(dev_ctx, nullptr); } } - -int main(int argc, char** argv) { - std::vector places; - - places.emplace_back(paddle::platform::CPUPlace()); - int count = paddle::platform::GetCUDADeviceCount(); - for (int i = 0; i < count; ++i) { - places.emplace_back(paddle::platform::CUDAPlace(i)); - } - - VLOG(0) << " DeviceCount " << count; - paddle::platform::DeviceContextPool::Init(places); - - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index 1a9be044e024e4b1dda5ef7d515c65f3a7513710..dc1d751141187edb7738e42c41514614d4d399b0 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -30,9 +30,6 @@ limitations under the License. */ namespace paddle { namespace platform { namespace { -// Current thread's id. Note, we don't distinguish nested threads -// for now. -thread_local int cur_thread_id = 0; // Tracking the nested block stacks of each thread. thread_local std::deque block_id_stack; // Tracking the nested event stacks. @@ -192,6 +189,8 @@ void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, } } // namespace +#endif // PADDLE_WITH_CUPTI + class DeviceTracerImpl : public DeviceTracer { public: DeviceTracerImpl() : enabled_(false) {} @@ -247,6 +246,8 @@ class DeviceTracerImpl : public DeviceTracer { if (enabled_) { return; } + +#ifdef PADDLE_WITH_CUPTI EnableActivity(); // Register callbacks for buffer requests and completed by CUPTI. @@ -265,6 +266,7 @@ class DeviceTracerImpl : public DeviceTracer { dynload::cuptiEnableCallback(1, subscriber_, CUPTI_CB_DOMAIN_DRIVER_API, CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)); CUPTI_CALL(dynload::cuptiGetTimestamp(&start_ns_)); +#endif // PADDLE_WITH_CUPTI enabled_ = true; } @@ -316,17 +318,21 @@ class DeviceTracerImpl : public DeviceTracer { } void Disable() { +#ifdef PADDLE_WITH_CUPTI // flush might cause additional calls to DeviceTracker. dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED); +#endif // PADDLE_WITH_CUPTI std::lock_guard l(trace_mu_); +#ifdef PADDLE_WITH_CUPTI DisableActivity(); dynload::cuptiUnsubscribe(subscriber_); CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_)); - PADDLE_ENFORCE(dynload::cuptiFinalize()); +#endif // PADDLE_WITH_CUPTI enabled_ = false; } private: +#ifdef PADDLE_WITH_CUPTI static void CUPTIAPI ApiCallback(void *userdata, CUpti_CallbackDomain domain, CUpti_CallbackId cbid, const void *cbdata) { auto *cbInfo = reinterpret_cast(cbdata); @@ -344,7 +350,8 @@ class DeviceTracerImpl : public DeviceTracer { VLOG(1) << "Unhandled API Callback for " << domain << " " << cbid; } } - + CUpti_SubscriberHandle subscriber_; +#endif // PADDLE_WITH_CUPTI std::mutex trace_mu_; bool enabled_; uint64_t start_ns_; @@ -353,45 +360,9 @@ class DeviceTracerImpl : public DeviceTracer { std::vector mem_records_; std::vector cpu_records_; std::unordered_map correlations_; - CUpti_SubscriberHandle subscriber_; -}; - -#endif // PADDLE_WITH_CUPTI - -class DeviceTracerDummy : public DeviceTracer { - public: - DeviceTracerDummy() {} - - void AddAnnotation(uint64_t id, const std::string &anno) {} - - void AddCPURecords(const std::string &anno, uint64_t start_ns, - uint64_t end_ns, int64_t device_id, int64_t thread_id) {} - - void AddMemRecords(const std::string &name, uint64_t start_ns, - uint64_t end_ns, int64_t device_id, int64_t stream_id, - uint32_t correlation_id, uint64_t bytes) {} - - void AddKernelRecords(uint64_t start, uint64_t end, int64_t device_id, - int64_t stream_id, uint32_t correlation_id) {} - - bool IsEnabled() { return false; } - - void Enable() {} - - proto::Profile GenProfile(const std::string &profile_path) { - return proto::Profile(); - } - - void Disable() {} }; -void CreateTracer(DeviceTracer **t) { -#ifdef PADDLE_WITH_CUPTI - *t = new DeviceTracerImpl(); -#else - *t = new DeviceTracerDummy(); -#endif // PADDLE_WITH_CUPTI -} +void CreateTracer(DeviceTracer **t) { *t = new DeviceTracerImpl(); } DeviceTracer *GetDeviceTracer() { std::call_once(tracer_once_flag, CreateTracer, &tracer); @@ -414,12 +385,5 @@ void SetCurBlock(int block_id) { block_id_stack.push_back(block_id); } void ClearCurBlock() { block_id_stack.pop_back(); } int BlockDepth() { return block_id_stack.size(); } - -void SetCurThread(int thread_id) { cur_thread_id = thread_id; } - -void ClearCurThread() { cur_thread_id = 0; } - -int CurThread() { return cur_thread_id; } - } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/device_tracer.h b/paddle/fluid/platform/device_tracer.h index 0375c7439c29d4122e8ff6b58734dad4f504b7a2..322996fb4f54d34ebbb034a6e1de420e9c532545 100644 --- a/paddle/fluid/platform/device_tracer.h +++ b/paddle/fluid/platform/device_tracer.h @@ -13,6 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include +#include // NOLINT #include #include "paddle/fluid/platform/dynload/cupti.h" @@ -25,6 +28,12 @@ namespace platform { // WARN: Under Development. Don't depend on it yet. ////////////////////// +inline uint64_t PosixInNsec() { + struct timeval tv; + gettimeofday(&tv, nullptr); + return 1000 * (static_cast(tv.tv_sec) * 1000000 + tv.tv_usec); +} + // DeviceTracer performs the following tasks: // 1. Register cuda callbacks for various events: kernel, memcpy, etc. // 2. Collect cuda statistics: start/end ts, memory, etc. @@ -99,9 +108,5 @@ std::string CurAnnotation(); void SetCurBlock(int block_id); void ClearCurBlock(); int BlockDepth(); - -void SetCurThread(int thread_id); -void ClearCurThread(); -int CurThread(); } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/dynload/CMakeLists.txt b/paddle/fluid/platform/dynload/CMakeLists.txt index 364c4901b297dbd647faae85b01f682a1daace9c..9da787a4073fa002f75154f7c4fba54e9ed8efa6 100644 --- a/paddle/fluid/platform/dynload/CMakeLists.txt +++ b/paddle/fluid/platform/dynload/CMakeLists.txt @@ -1,14 +1,23 @@ cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce) -list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc nccl.cc) +list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc) + +# There is no macOS version of NCCL. +if (NOT APPLE) + list(APPEND CUDA_SRCS nccl.cc) +endif() + if (TENSORRT_FOUND) list(APPEND CUDA_SRCS tensorrt.cc) endif() - configure_file(cupti_lib_path.h.in ${CMAKE_CURRENT_BINARY_DIR}/cupti_lib_path.h) if (CUPTI_FOUND) list(APPEND CUDA_SRCS cupti.cc) endif(CUPTI_FOUND) nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader) cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc) +if (WITH_MKLML) + cc_library(dynload_mklml SRCS mklml.cc DEPS dynamic_loader mklml) +endif() +# TODO(TJ): add iomp, mkldnn? diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h index 81acaff87d3c2025cf0d6185a1590b018bfbd83c..25bcda7eedc1ef42f75fb8fd1439f0c8f55015c3 100644 --- a/paddle/fluid/platform/dynload/cublas.h +++ b/paddle/fluid/platform/dynload/cublas.h @@ -45,7 +45,7 @@ extern void *cublas_dso_handle; std::call_once(cublas_dso_flag, []() { \ cublas_dso_handle = paddle::platform::dynload::GetCublasDsoHandle(); \ }); \ - void *p_##__name = dlsym(cublas_dso_handle, #__name); \ + static void *p_##__name = dlsym(cublas_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h index 34d83e395694f55eafca74d63ebf363169ab30e8..77e46fa768b62c277d7b4027de7173e39a5672b4 100644 --- a/paddle/fluid/platform/dynload/cudnn.h +++ b/paddle/fluid/platform/dynload/cudnn.h @@ -39,7 +39,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name); cudnn_dso_handle = paddle::platform::dynload::GetCUDNNDsoHandle(); \ }); \ EnforceCUDNNLoaded(#__name); \ - void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ + static void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/cupti.h b/paddle/fluid/platform/dynload/cupti.h index e64de7c20fc9d145e51cfc4528e321b3c4ec86c8..e8f4a82ef132be9e4ec3fb76f11766046a2ff638 100644 --- a/paddle/fluid/platform/dynload/cupti.h +++ b/paddle/fluid/platform/dynload/cupti.h @@ -45,7 +45,7 @@ extern void *cupti_dso_handle; std::call_once(cupti_dso_flag, []() { \ cupti_dso_handle = paddle::platform::dynload::GetCUPTIDsoHandle(); \ }); \ - void *p_##__name = dlsym(cupti_dso_handle, #__name); \ + static void *p_##__name = dlsym(cupti_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ @@ -72,7 +72,6 @@ extern void *cupti_dso_handle; __macro(cuptiGetResultString); \ __macro(cuptiActivityGetNumDroppedRecords); \ __macro(cuptiActivityFlushAll); \ - __macro(cuptiFinalize); \ __macro(cuptiSubscribe); \ __macro(cuptiUnsubscribe); \ __macro(cuptiEnableCallback); \ diff --git a/paddle/fluid/platform/dynload/curand.h b/paddle/fluid/platform/dynload/curand.h index 46ad4379d5f9572d415ef1d747077217ae29391e..5b9e0820e0b319fe7a636a57a0029caf038b4db3 100644 --- a/paddle/fluid/platform/dynload/curand.h +++ b/paddle/fluid/platform/dynload/curand.h @@ -34,7 +34,7 @@ extern void *curand_dso_handle; std::call_once(curand_dso_flag, []() { \ curand_dso_handle = paddle::platform::dynload::GetCurandDsoHandle(); \ }); \ - void *p_##__name = dlsym(curand_dso_handle, #__name); \ + static void *p_##__name = dlsym(curand_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/dynamic_loader.cc b/paddle/fluid/platform/dynload/dynamic_loader.cc index 19c01dc5a968c7e1d2b0f15cf9a0e8427004e58b..93bf7c13516ffa4baca6a30f1daf946939726d85 100644 --- a/paddle/fluid/platform/dynload/dynamic_loader.cc +++ b/paddle/fluid/platform/dynload/dynamic_loader.cc @@ -36,8 +36,6 @@ DEFINE_string(cuda_dir, "", DEFINE_string(warpctc_dir, "", "Specify path for loading libwarpctc.so."); -DEFINE_string(lapack_dir, "", "Specify path for loading liblapack.so."); - DEFINE_string(nccl_dir, "", "Specify path for loading nccl library, such as libcublas, " "libcurand. For instance, /usr/local/cuda/lib64. If default, " @@ -49,6 +47,8 @@ DEFINE_string( tensorrt_dir, "", "Specify path for loading tensorrt library, such as libnvinfer.so."); +DEFINE_string(mklml_dir, "", "Specify path for loading libmklml_intel.so."); + namespace paddle { namespace platform { namespace dynload { @@ -76,6 +76,7 @@ static inline void* GetDsoHandleFromDefaultPath(const std::string& dso_path, VLOG(3) << "Try to find library: " << dso_path << " from default system path."; // default search from LD_LIBRARY_PATH/DYLD_LIBRARY_PATH + // and /usr/local/lib path void* dso_handle = dlopen(dso_path.c_str(), dynload_flags); // DYLD_LIBRARY_PATH is disabled after Mac OS 10.11 to @@ -97,6 +98,10 @@ static inline void* GetDsoHandleFromDefaultPath(const std::string& dso_path, } #endif + if (nullptr == dso_handle) { + LOG(WARNING) << "Can not find library: " << dso_path + << ". Please try to add the lib path to LD_LIBRARY_PATH."; + } return dso_handle; } @@ -182,14 +187,6 @@ void* GetWarpCTCDsoHandle() { #endif } -void* GetLapackDsoHandle() { -#if defined(__APPLE__) || defined(__OSX__) - return GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapacke.dylib"); -#else - return GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapacke.so"); -#endif -} - void* GetNCCLDsoHandle() { #if defined(__APPLE__) || defined(__OSX__) return GetDsoHandleFromSearchPath(FLAGS_nccl_dir, "libnccl.dylib"); @@ -206,6 +203,14 @@ void* GetTensorRtDsoHandle() { #endif } +void* GetMKLMLDsoHandle() { +#if defined(__APPLE__) || defined(__OSX__) + return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.dylib"); +#else + return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.so"); +#endif +} + } // namespace dynload } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/dynload/dynamic_loader.h b/paddle/fluid/platform/dynload/dynamic_loader.h index 0de3559b6088086cb52c254535b6ec42da7dd724..84fd2ce9987628a5ed29e4125a03dedb96e416c1 100644 --- a/paddle/fluid/platform/dynload/dynamic_loader.h +++ b/paddle/fluid/platform/dynload/dynamic_loader.h @@ -23,9 +23,9 @@ void* GetCUDNNDsoHandle(); void* GetCUPTIDsoHandle(); void* GetCurandDsoHandle(); void* GetWarpCTCDsoHandle(); -void* GetLapackDsoHandle(); void* GetNCCLDsoHandle(); void* GetTensorRtDsoHandle(); +void* GetMKLMLDsoHandle(); } // namespace dynload } // namespace platform diff --git a/paddle/fluid/platform/dynload/mklml.cc b/paddle/fluid/platform/dynload/mklml.cc new file mode 100644 index 0000000000000000000000000000000000000000..0f61a5e09b3243cbdf570ba7c28a260f181d8848 --- /dev/null +++ b/paddle/fluid/platform/dynload/mklml.cc @@ -0,0 +1,30 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/platform/dynload/mklml.h" + +namespace paddle { +namespace platform { +namespace dynload { + +std::once_flag mklml_dso_flag; +void* mklml_dso_handle = nullptr; + +#define DEFINE_WRAP(__name) DynLoad__##__name __name + +MKLML_ROUTINE_EACH(DEFINE_WRAP); + +} // namespace dynload +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/dynload/mklml.h b/paddle/fluid/platform/dynload/mklml.h new file mode 100644 index 0000000000000000000000000000000000000000..17acefe8cde01809572e4c86cbdccfed9a477a51 --- /dev/null +++ b/paddle/fluid/platform/dynload/mklml.h @@ -0,0 +1,71 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include // NOLINT +#include "paddle/fluid/platform/dynload/dynamic_loader.h" + +namespace paddle { +namespace platform { +namespace dynload { + +extern std::once_flag mklml_dso_flag; +extern void* mklml_dso_handle; + +/** + * The following macro definition can generate structs + * (for each function) to dynamic load mklml routine + * via operator overloading. + */ +#define DYNAMIC_LOAD_MKLML_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + using mklmlFunc = decltype(&::__name); \ + std::call_once(mklml_dso_flag, []() { \ + mklml_dso_handle = paddle::platform::dynload::GetMKLMLDsoHandle(); \ + }); \ + static void* p_##_name = dlsym(mklml_dso_handle, #__name); \ + return reinterpret_cast(p_##_name)(args...); \ + } \ + }; \ + extern DynLoad__##__name __name + +#define DECLARE_DYNAMIC_LOAD_MKLML_WRAP(__name) DYNAMIC_LOAD_MKLML_WRAP(__name) + +#define MKLML_ROUTINE_EACH(__macro) \ + __macro(cblas_sgemm); \ + __macro(cblas_saxpy); \ + __macro(cblas_scopy); \ + __macro(cblas_sgemv); \ + __macro(cblas_sgemm_batch); \ + __macro(cblas_dgemm); \ + __macro(cblas_daxpy); \ + __macro(cblas_dcopy); \ + __macro(cblas_dgemv); \ + __macro(cblas_dgemm_batch); \ + __macro(vsAdd); \ + __macro(vdAdd); \ + __macro(MKL_Set_Num_Threads) + +MKLML_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_MKLML_WRAP); + +#undef DYNAMIC_LOAD_MKLML_WRAP + +} // namespace dynload +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/dynload/nccl.h b/paddle/fluid/platform/dynload/nccl.h index 37902ae20c5d9d64486232bbd468375c4a50a615..575516f81870fc9f7b92919ffc20a201cb5cbce8 100644 --- a/paddle/fluid/platform/dynload/nccl.h +++ b/paddle/fluid/platform/dynload/nccl.h @@ -37,7 +37,7 @@ extern void* nccl_dso_handle; std::call_once(nccl_dso_flag, []() { \ nccl_dso_handle = paddle::platform::dynload::GetNCCLDsoHandle(); \ }); \ - void* p_##__name = dlsym(nccl_dso_handle, #__name); \ + static void* p_##__name = dlsym(nccl_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/tensorrt.h b/paddle/fluid/platform/dynload/tensorrt.h index f584a49da0fefe0b064b5fb55b01ec132225ce5e..5d67658b94af75680a100e13eed7b6b052162e00 100644 --- a/paddle/fluid/platform/dynload/tensorrt.h +++ b/paddle/fluid/platform/dynload/tensorrt.h @@ -40,7 +40,7 @@ extern void* tensorrt_dso_handle; paddle::platform::dynload::GetTensorRtDsoHandle(); \ PADDLE_ENFORCE(tensorrt_dso_handle, "load tensorrt so failed"); \ }); \ - void* p_##__name = dlsym(tensorrt_dso_handle, #__name); \ + static void* p_##__name = dlsym(tensorrt_dso_handle, #__name); \ PADDLE_ENFORCE(p_##__name, "load %s failed", #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ diff --git a/paddle/fluid/platform/dynload/warpctc.h b/paddle/fluid/platform/dynload/warpctc.h index 7c70649d21c547beb824576d4a8ecf6219a9bddf..d157c1fda789b98f06ad069d2a9c4f421ff82dcd 100644 --- a/paddle/fluid/platform/dynload/warpctc.h +++ b/paddle/fluid/platform/dynload/warpctc.h @@ -40,7 +40,7 @@ extern void* warpctc_dso_handle; std::call_once(warpctc_dso_flag, []() { \ warpctc_dso_handle = paddle::platform::dynload::GetWarpCTCDsoHandle(); \ }); \ - void* p_##_name = dlsym(warpctc_dso_handle, #__name); \ + static void* p_##_name = dlsym(warpctc_dso_handle, #__name); \ return reinterpret_cast(p_##_name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index 7b8c29e1e642ec6bb4023afd8c083311b8b31812..566485cd3c383640047d97f40b452735e8c8c171 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -44,8 +44,10 @@ limitations under the License. */ #include "paddle/fluid/platform/dynload/cublas.h" #include "paddle/fluid/platform/dynload/cudnn.h" #include "paddle/fluid/platform/dynload/curand.h" +#ifndef __APPLE__ #include "paddle/fluid/platform/dynload/nccl.h" -#endif +#endif // __APPLE__ +#endif // PADDLE_WITH_CUDA namespace paddle { namespace platform { @@ -100,6 +102,15 @@ struct EnforceNotMet : public std::exception { const char* what() const noexcept { return err_str_.c_str(); } }; +struct EOFException : public std::exception { + std::string err_str_; + EOFException(const char* err_msg, const char* f, int l) { + err_str_ = string::Sprintf("%s at [%s:%d]", err_msg, f, l); + } + + const char* what() const noexcept { return err_str_.c_str(); } +}; + // Because most enforce conditions would evaluate to true, we can use // __builtin_expect to instruct the C++ compiler to generate code that // always forces branch prediction of true. @@ -111,7 +122,11 @@ template inline typename std::enable_if::type throw_on_error( bool stat, const Args&... args) { if (UNLIKELY(!(stat))) { +#ifndef REPLACE_ENFORCE_GLOG throw std::runtime_error(string::Sprintf(args...)); +#else + LOG(FATAL) << string::Sprintf(args...); +#endif } } @@ -121,8 +136,12 @@ template inline typename std::enable_if::type throw_on_error( cudaError_t e, const Args&... args) { if (UNLIKELY(e)) { +#ifndef REPLACE_ENFORCE_GLOG throw thrust::system_error(e, thrust::cuda_category(), string::Sprintf(args...)); +#else + LOG(FATAL) << string::Sprintf(args...); +#endif } } @@ -130,8 +149,12 @@ template inline typename std::enable_if::type throw_on_error( curandStatus_t stat, const Args&... args) { if (stat != CURAND_STATUS_SUCCESS) { +#ifndef REPLACE_ENFORCE_GLOG throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(), string::Sprintf(args...)); +#else + LOG(FATAL) << string::Sprintf(args...); +#endif } } @@ -141,8 +164,12 @@ inline typename std::enable_if::type throw_on_error( if (stat == CUDNN_STATUS_SUCCESS) { return; } else { +#ifndef REPLACE_ENFORCE_GLOG throw std::runtime_error(platform::dynload::cudnnGetErrorString(stat) + string::Sprintf(args...)); +#else + LOG(FATAL) << string::Sprintf(args...); +#endif } } @@ -171,20 +198,30 @@ inline typename std::enable_if::type throw_on_error( } else if (stat == CUBLAS_STATUS_LICENSE_ERROR) { err = "CUBLAS: license error, "; } +#ifndef REPLACE_ENFORCE_GLOG throw std::runtime_error(err + string::Sprintf(args...)); +#else + LOG(FATAL) << err << string::Sprintf(args...); +#endif } +#ifndef __APPLE__ template inline typename std::enable_if::type throw_on_error( ncclResult_t stat, const Args&... args) { if (stat == ncclSuccess) { return; } else { +#ifndef REPLACE_ENFORCE_GLOG throw std::runtime_error(platform::dynload::ncclGetErrorString(stat) + string::Sprintf(args...)); +#else + LOG(FATAL) << platform::dynload::ncclGetErrorString(stat) + << string::Sprintf(args...); +#endif } } - +#endif // __APPLE__ #endif // PADDLE_WITH_CUDA template @@ -200,6 +237,7 @@ inline void throw_on_error(T e) { __FILE__, __LINE__); \ } while (false) +#ifndef REPLACE_ENFORCE_GLOG #define PADDLE_ENFORCE(...) \ do { \ try { \ @@ -209,7 +247,15 @@ inline void throw_on_error(T e) { __FILE__, __LINE__); \ } \ } while (false) +#else +#define PADDLE_ENFORCE(...) ::paddle::platform::throw_on_error(__VA_ARGS__); +#endif +#define PADDLE_THROW_EOF() \ + do { \ + throw ::paddle::platform::EOFException("There is no next data.", __FILE__, \ + __LINE__); \ + } while (false) /* * Some enforce helpers here, usage: * int a = 1; diff --git a/paddle/fluid/platform/enforce_test.cc b/paddle/fluid/platform/enforce_test.cc index 57d751cc00b5f11f1ba1a3b0c9a6b7ce9e79f586..0e8684581a93f076b1a077cc52e966d3c88cf078 100644 --- a/paddle/fluid/platform/enforce_test.cc +++ b/paddle/fluid/platform/enforce_test.cc @@ -210,3 +210,14 @@ TEST(ENFORCE_USER_DEFINED_CLASS, NE) { Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}}; ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet); } + +TEST(EOF_EXCEPTION, THROW_EOF) { + bool caught_eof = false; + try { + PADDLE_THROW_EOF(); + } catch (paddle::platform::EOFException error) { + caught_eof = true; + EXPECT_TRUE(HasPrefix(StringPiece(error.what()), "There is no next data.")); + } + EXPECT_TRUE(caught_eof); +} diff --git a/paddle/fluid/platform/float16.h b/paddle/fluid/platform/float16.h index ffd183af68514dbb1a8b3de39000c9ca3f56ddc3..efb021c838e3680ab2cdd1c4b298cf7ec2186478 100644 --- a/paddle/fluid/platform/float16.h +++ b/paddle/fluid/platform/float16.h @@ -67,8 +67,11 @@ struct float16; } // namespace platform } // namespace paddle +// NOTE(): +// Do not move the eigen.h header, otherwise the eigen_vector will failed. #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/platform/hostdevice.h" +#include "unsupported/Eigen/CXX11/Tensor" namespace paddle { namespace platform { @@ -898,6 +901,30 @@ struct is_pod { is_standard_layout::value; }; +template <> +struct is_floating_point + : std::integral_constant< + bool, std::is_same::type>::value> {}; +template <> +struct is_signed { + static const bool value = true; +}; + +template <> +struct is_unsigned { + static const bool value = false; +}; + +inline bool isnan(const paddle::platform::float16& a) { + return paddle::platform::isnan(a); +} + +inline bool isinf(const paddle::platform::float16& a) { + return paddle::platform::isinf(a); +} + template <> struct numeric_limits { static const bool is_specialized = true; diff --git a/paddle/fluid/platform/float16_test.cc b/paddle/fluid/platform/float16_test.cc index a589e32b61a9b6a44bdc4529eee715d987d6922c..27e930e6e0a76982b3f27619f38a4a08d82cafa1 100644 --- a/paddle/fluid/platform/float16_test.cc +++ b/paddle/fluid/platform/float16_test.cc @@ -13,8 +13,8 @@ limitations under the License. */ #include #include "gtest/gtest.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/platform/init.h" namespace paddle { namespace platform { @@ -141,10 +141,36 @@ TEST(float16, lod_tensor_cpu) { } } +TEST(float16, floating) { + // compile time assert. + PADDLE_ASSERT(std::is_floating_point::value); +} + TEST(float16, print) { float16 a = float16(1.0f); std::cout << a << std::endl; } +// CPU test +TEST(float16, isinf) { + float16 a; + a.x = 0x7c00; + float16 b = float16(INFINITY); + float16 c = static_cast(INFINITY); + EXPECT_EQ(std::isinf(a), true); + EXPECT_EQ(std::isinf(b), true); + EXPECT_EQ(std::isinf(c), true); +} + +TEST(float16, isnan) { + float16 a; + a.x = 0x7fff; + float16 b = float16(NAN); + float16 c = static_cast(NAN); + EXPECT_EQ(std::isnan(a), true); + EXPECT_EQ(std::isnan(b), true); + EXPECT_EQ(std::isnan(c), true); +} + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/float16_test.cu b/paddle/fluid/platform/float16_test.cu index 577fc24ceb1d3c83cc0546dc5db9c8c7c1f01f86..e2b7ca9b03809113c31af8ff4d3ad3713748f330 100644 --- a/paddle/fluid/platform/float16_test.cu +++ b/paddle/fluid/platform/float16_test.cu @@ -11,11 +11,13 @@ limitations under the License. */ #include "paddle/fluid/platform/float16.h" +#include #include +#include +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor_util.h" -#include "paddle/utils/Logging.h" #define ARITHMETIC_KERNEL(op_type, sign) \ __global__ void op_type(const half* in1, const half* in2, half* out) { \ @@ -241,6 +243,72 @@ TEST(float16, lod_tensor_on_gpu) { } } +template +struct Functor { + bool operator()(const T& val) { + return std::type_index(typeid(T)) == + std::type_index(typeid(platform::float16)); + } +}; + +TEST(float16, typeid) { + // the framework heavily used typeid hash + Functor functor; + float16 a = float16(.0f); + Functor functor2; + int b(0); + + // compile time assert + PADDLE_ASSERT(functor(a) == true); + PADDLE_ASSERT(functor2(b) == false); +} + +// GPU test +TEST(float16, isinf) { + float16 a; + a.x = 0x7c00; + float16 b = float16(INFINITY); + // underflow to 0 + float16 native_a(5e-40f); + // overflow to inf + float16 native_b(5e40f); + EXPECT_EQ(std::isinf(a), true); + EXPECT_EQ(std::isinf(b), true); + EXPECT_EQ(std::isinf(native_b), true); + EXPECT_EQ(native_a, float16(0)); +} + +TEST(float16, isnan) { + float16 a; + a.x = 0x7fff; + float16 b = float16(NAN); + float16 c = float16(5e40); + // inf * +-0 will get a nan + float16 d = c * float16(0); + EXPECT_EQ(std::isnan(a), true); + EXPECT_EQ(std::isnan(b), true); + EXPECT_EQ(std::isnan(d), true); +} + +TEST(float16, cast) { + float16 a; + a.x = 0x0070; + auto b = a; + { + // change semantic, keep the same value + float16 c = reinterpret_cast(reinterpret_cast(b)); + EXPECT_EQ(b, c); + } + + { + // use uint32 low 16 bit store float16 + uint32_t c = reinterpret_cast(b); + float16 d; + d.x = c; + EXPECT_EQ(b, d); + } +} + } // namespace platform } // namespace paddle #endif // PADDLE_CUDA_FP16 diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc new file mode 100644 index 0000000000000000000000000000000000000000..6f1f0c4796f3bae2fb419bf103cb6c0c5489bf65 --- /dev/null +++ b/paddle/fluid/platform/init.cc @@ -0,0 +1,133 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#include // for strdup +#include +#include +#include + +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/cpu_helper.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/init.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/string/piece.h" + +DEFINE_int32(paddle_num_threads, 1, + "Number of threads for each paddle instance."); + +namespace paddle { +namespace framework { + +std::once_flag gflags_init_flag; +std::once_flag p2p_init_flag; + +void InitGflags(std::vector argv) { + std::call_once(gflags_init_flag, [&]() { + argv.insert(argv.begin(), "dummy"); + int argc = argv.size(); + char **arr = new char *[argv.size()]; + std::string line; + for (size_t i = 0; i < argv.size(); i++) { + arr[i] = &argv[i][0]; + line += argv[i]; + line += ' '; + } + google::ParseCommandLineFlags(&argc, &arr, true); + VLOG(1) << "Init commandline: " << line; + }); +} + +void InitP2P(std::vector devices) { +#ifdef PADDLE_WITH_CUDA + std::call_once(p2p_init_flag, [&]() { + int count = devices.size(); + for (int i = 0; i < count; ++i) { + for (int j = 0; j < count; ++j) { + if (devices[i] == devices[j]) continue; + int can_acess = -1; + PADDLE_ENFORCE( + cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]), + "Failed to test P2P access."); + if (can_acess != 1) { + LOG(WARNING) << "Cannot enable P2P access from " << devices[i] + << " to " << devices[j]; + } else { + cudaSetDevice(devices[i]); + cudaDeviceEnablePeerAccess(devices[j], 0); + } + } + } + }); +#endif +} + +void InitDevices(bool init_p2p) { + /*Init all available devices by default */ + std::vector devices; +#ifdef PADDLE_WITH_CUDA + try { + int count = platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + devices.push_back(i); + } + } catch (const std::exception &exp) { + LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime."; + } +#else + LOG(WARNING) + << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; +#endif + InitDevices(init_p2p, devices); +} + +void InitDevices(bool init_p2p, const std::vector devices) { + std::vector places; + int count = 0; +#ifdef PADDLE_WITH_CUDA + try { + count = platform::GetCUDADeviceCount(); + } catch (const std::exception &exp) { + LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime."; + } +#else + LOG(WARNING) + << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; +#endif + + for (size_t i = 0; i < devices.size(); ++i) { + if (devices[i] >= count || devices[i] < 0) { + LOG(WARNING) << "Invalid devices id."; + continue; + } + places.emplace_back(platform::CUDAPlace(devices[i])); + } + if (init_p2p) { + InitP2P(devices); + } + places.emplace_back(platform::CPUPlace()); + platform::DeviceContextPool::Init(places); +#ifndef PADDLE_WITH_MKLDNN + platform::SetNumThreads(FLAGS_paddle_num_threads); +#endif +} + +void InitGLOG(const std::string &prog_name) { + // glog will not hold the ARGV[0] inside. + // Use strdup to alloc a new string. + google::InitGoogleLogging(strdup(prog_name.c_str())); + google::InstallFailureSignalHandler(); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/init.h b/paddle/fluid/platform/init.h similarity index 100% rename from paddle/fluid/framework/init.h rename to paddle/fluid/platform/init.h diff --git a/paddle/fluid/platform/init_test.cc b/paddle/fluid/platform/init_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..eef1470a90c7da15efff965fc8f66dfa616ba25f --- /dev/null +++ b/paddle/fluid/platform/init_test.cc @@ -0,0 +1,40 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#include "gtest/gtest.h" + +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/init.h" + +TEST(InitDevices, CPU) { + using paddle::framework::InitDevices; + using paddle::platform::DeviceContextPool; + +#ifndef PADDLE_WITH_CUDA + InitDevices(true); + DeviceContextPool& pool = DeviceContextPool::Instance(); + ASSERT_EQ(pool.size(), 1U); +#endif +} + +TEST(InitDevices, CUDA) { + using paddle::framework::InitDevices; + using paddle::platform::DeviceContextPool; + +#ifdef PADDLE_WITH_CUDA + int count = paddle::platform::GetCUDADeviceCount(); + InitDevices(true); + DeviceContextPool& pool = DeviceContextPool::Instance(); + ASSERT_EQ(pool.size(), 1U + static_cast(count)); +#endif +} diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index f1187620d81ff3bc1deef2106edb54d6199fa927..10a3ad256b17ba41380cdc0377905d03188cbaa3 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -14,8 +14,10 @@ limitations under the License. */ #pragma once #include +#include #include #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/place.h" namespace paddle { namespace platform { @@ -86,5 +88,180 @@ inline mkldnn::memory::data_type MKLDNNGetDataType() { return mkldnn::memory::f32; } +inline void Reorder(const mkldnn::memory& src, const mkldnn::memory& dst) { + auto reorder_prim = mkldnn::reorder(src, dst); + std::vector pipeline; + pipeline.push_back(reorder_prim); + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} + +inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) { + return static_cast( + memory.get_primitive_desc().desc().data.format); +} + +inline mkldnn::memory::format GetMKLDNNFormat( + const mkldnn::sum::primitive_desc& memory) { + return static_cast( + memory.dst_primitive_desc().desc().data.format); +} + +class MKLDNNHandler { + public: + MKLDNNHandler(const MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : dev_ctx_(dev_ctx), + engine_(engine), + key_(base_key), + is_reusing_(false) {} + + std::shared_ptr AcquireSrcMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_src_mem_p"); + } + + std::shared_ptr AcquireWeightsMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_weights_mem_p"); + } + + std::shared_ptr AcquireDstMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_dst_mem_p"); + } + + std::shared_ptr AcquireDiffDstMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_diff_dst_mem_p"); + } + + std::shared_ptr AcquireDiffSrcMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_diff_src_mem_p"); + } + + std::shared_ptr AcquireMemoryFromPrimitive( + mkldnn::memory::primitive_desc mdp, void* ptr, + const std::string& suffix) { + auto local_key = key_ + suffix; + auto mem_p = + std::static_pointer_cast(dev_ctx_.GetBlob(local_key)); + PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false), + "Fail to find mem primitive in device context"); + if (mem_p == nullptr) { + mem_p = std::make_shared(mdp, ptr); + dev_ctx_.SetBlob(local_key, mem_p); + } else { + mem_p->set_data_handle(ptr); + // Mark that reusing happenned. All primitives from operator instance + // should be reused or none of them. So we check consistency + is_reusing_ = true; + } + return mem_p; + } + + std::shared_ptr AcquireMemory(const mkldnn::memory::desc& md, + void* ptr, + const std::string& suffix) { + /*Generate key*/ + auto local_key = key_ + suffix; + auto mem_p = + std::static_pointer_cast(dev_ctx_.GetBlob(local_key)); + PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false), + "Fail to find mem primitive in device context"); + if (mem_p == nullptr) { + mem_p = std::make_shared( + mkldnn::memory::primitive_desc{md, engine_}, ptr); + dev_ctx_.SetBlob(local_key, mem_p); + } else { + mem_p->set_data_handle(ptr); + // Mark that reusing happenned. All primitives from operator instance + // should be reused or none of them. So we check consistency + is_reusing_ = true; + } + return mem_p; + } + + std::shared_ptr AcquireMemory( + mkldnn::memory::primitive_desc& mpd, // NOLINT + mkldnn::memory::primitive_desc& user_mpd, // NOLINT + const std::shared_ptr user_memory_p, + const std::string& suffix, + std::vector& pipeline) { // NOLINT + // create reorder primitive if the input format is not the preferred one + auto local_key = key_ + suffix; + auto key_reorder_p = key_ + suffix + "reorder_p"; + + auto target_memory_p = + std::static_pointer_cast(dev_ctx_.GetBlob(local_key)); + PADDLE_ENFORCE((target_memory_p != nullptr) || (is_reusing_ == false), + "Fail to find mem primitive in device context"); + if (target_memory_p == nullptr) { + target_memory_p = user_memory_p; + std::shared_ptr reorder_p; + if (mpd != user_mpd) { + target_memory_p = std::make_shared(mpd); + + auto reorder_p = + std::make_shared(*user_memory_p, *target_memory_p); + dev_ctx_.SetBlob(key_reorder_p, reorder_p); + pipeline.push_back(*reorder_p); + } + dev_ctx_.SetBlob(local_key, target_memory_p); + } else { + // Make reorder if needed + auto reorder_p = std::static_pointer_cast( + dev_ctx_.GetBlob(key_reorder_p)); + if (reorder_p != nullptr) { + pipeline.push_back(*reorder_p); + } + is_reusing_ = true; + } + return target_memory_p; + } + + static std::string GetHash(mkldnn::memory::dims& operand_dims, // NOLINT + const std::string& suffix) { + return dims2str(operand_dims) + suffix; + } + + protected: + static std::string dims2str(const mkldnn::memory::dims& operand_dims) { + std::string dstr = ""; + for (size_t i = 0; i < operand_dims.size(); ++i) { + dstr += std::to_string(operand_dims[i]) + "-"; + } + return dstr; + } + + protected: + const MKLDNNDeviceContext& dev_ctx_; + mkldnn::engine engine_; + std::string key_; + bool is_reusing_; +}; + +inline mkldnn::memory::format MKLDNNFormatForSize( + size_t dims_size, mkldnn::memory::format data_format) { + if (dims_size == 1) { + return mkldnn::memory::format::x; + } else if (dims_size == 2) { + return mkldnn::memory::format::nc; + } + return data_format; +} + +inline mkldnn::memory::format data_format_to_memory_format( + const std::string& data_format) { + switch (framework::StringToDataLayout(data_format)) { + case framework::DataLayout::kNHWC: + return mkldnn::memory::format::nhwc; + case framework::DataLayout::kNCHW: + return mkldnn::memory::format::nchw; + default: + return mkldnn::memory::format::any; + } +} + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 6f8e3f22db54d166cf97cfdd3d009058207a7ca5..cc46c88fd1f9a5d1bacad26beed6fd0af6405310 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -41,6 +41,11 @@ inline ncclDataType_t ToNCCLDataType(std::type_index type) { } } +// NOTE(minqiyang): according to the ncclGroupEnd documentations: +// https://docs.nvidia.com/deeplearning/sdk/nccl-api/ncclapidoc.html, +// ncclGroupEnd will wait for all communicators to be initialized, which will +// cause blocking problem when a runtime_error was thrown, so try only guard +// NCCL actions when use it. class NCCLGroupGuard { public: static std::mutex &NCCLMutex() { diff --git a/paddle/fluid/platform/place.h b/paddle/fluid/platform/place.h index ad54a878996bd36f2d714f6554b44c89dae3fd0c..e3ee504f3d042d6a99036e34507c4c8bee306750 100644 --- a/paddle/fluid/platform/place.h +++ b/paddle/fluid/platform/place.h @@ -30,6 +30,7 @@ struct CPUPlace { // needed for variant equality comparison inline bool operator==(const CPUPlace &) const { return true; } inline bool operator!=(const CPUPlace &) const { return false; } + inline bool operator<(const CPUPlace &) const { return false; } }; struct CUDAPlace { @@ -42,6 +43,7 @@ struct CUDAPlace { return device == o.device; } inline bool operator!=(const CUDAPlace &o) const { return !(*this == o); } + inline bool operator<(const CUDAPlace &o) const { return device < o.device; } int device; }; @@ -52,6 +54,7 @@ struct CUDAPinnedPlace { // needed for variant equality comparison inline bool operator==(const CUDAPinnedPlace &) const { return true; } inline bool operator!=(const CUDAPinnedPlace &) const { return false; } + inline bool operator<(const CUDAPinnedPlace &) const { return false; } }; struct IsCUDAPlace : public boost::static_visitor { @@ -89,18 +92,6 @@ bool is_cuda_pinned_place(const Place &); bool places_are_same_class(const Place &, const Place &); bool is_same_place(const Place &, const Place &); -struct PlaceHash { - std::size_t operator()(const Place &p) const { - constexpr size_t num_dev_bits = 4; - std::hash ihash; - size_t dev_id = 0; - if (is_gpu_place(p)) { - dev_id = boost::get(p).device; - } - return ihash(dev_id << num_dev_bits | p.which()); - } -}; - std::ostream &operator<<(std::ostream &, const Place &); template diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 3d8d64e4c2758675067834810ebb9aee1e88fdb9..652a6ec7a4e2e823b28f39b449570cd375e88e18 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -15,7 +15,6 @@ limitations under the License. */ #include "paddle/fluid/platform/profiler.h" #include -#include #include #include #include @@ -97,12 +96,6 @@ inline uint64_t GetTimeInNsec() { .count(); } -inline uint64_t PosixInNsec() { - struct timeval tv; - gettimeofday(&tv, nullptr); - return 1000 * (static_cast(tv.tv_sec) * 1000000 + tv.tv_usec); -} - Event::Event(EventType type, std::string name, uint32_t thread_id, const DeviceContext* dev_ctx) : type_(type), name_(name), thread_id_(thread_id), has_cuda_(false) { @@ -110,6 +103,8 @@ Event::Event(EventType type, std::string name, uint32_t thread_id, has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false; if (has_cuda_) { auto* cuda_dev_ctx = static_cast(dev_ctx); + PADDLE_ENFORCE(cudaSetDevice( + boost::get(cuda_dev_ctx->GetPlace()).device)); PADDLE_ENFORCE(cudaGetDevice(&device_)); PADDLE_ENFORCE(cudaEventCreate(&event_)); auto stream = cuda_dev_ctx->stream(); @@ -127,6 +122,7 @@ double Event::CpuElapsedMs(const Event& e) const { double Event::CudaElapsedMs(const Event& e) const { #ifdef PADDLE_WITH_CUDA + if (!has_cuda_) return 0.0; PADDLE_ENFORCE(e.has_cuda() && has_cuda()); PADDLE_ENFORCE(e.device() == device()); PADDLE_ENFORCE(cudaEventSynchronize(event_)); @@ -175,6 +171,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) { RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) : is_enabled_(false), start_ns_(PosixInNsec()) { + std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled) return; is_enabled_ = true; dev_ctx_ = dev_ctx; @@ -185,11 +182,12 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) } RecordEvent::~RecordEvent() { + std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(), - BlockDepth(), CurThread()); + BlockDepth(), g_thread_id); } ClearCurAnnotation(); PopEvent(name_, dev_ctx_); @@ -197,6 +195,7 @@ RecordEvent::~RecordEvent() { RecordBlock::RecordBlock(int block_id) : is_enabled_(false), start_ns_(PosixInNsec()) { + std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled) return; is_enabled_ = true; SetCurBlock(block_id); @@ -204,27 +203,18 @@ RecordBlock::RecordBlock(int block_id) } RecordBlock::~RecordBlock() { + std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { // We try to put all blocks at the same nested depth in the // same timeline lane. and distinguish the using thread_id. tracer->AddCPURecords(name_, start_ns_, PosixInNsec(), BlockDepth(), - CurThread()); + g_thread_id); } ClearCurBlock(); } -RecordThread::RecordThread(int thread_id) { - if (g_state == ProfilerState::kDisabled) return; - SetCurThread(thread_id); -} - -RecordThread::~RecordThread() { - if (g_state == ProfilerState::kDisabled) return; - ClearCurThread(); -} - void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, "Can't enbale profling, since the input state is ", @@ -280,12 +270,13 @@ struct EventItem { double min_time; double max_time; double ave_time; + float ratio; }; // Print results void PrintProfiler(const std::vector>& events_table, const std::string& sorted_domain, const size_t name_width, - const size_t data_width) { + const size_t data_width, double total) { // Output header information std::cout << "\n------------------------->" << " Profiling Report " @@ -310,7 +301,8 @@ void PrintProfiler(const std::vector>& events_table, std::cout << std::setw(name_width) << "Event" << std::setw(data_width) << "Calls" << std::setw(data_width) << "Total" << std::setw(data_width) << "Min." << std::setw(data_width) - << "Max." << std::setw(data_width) << "Ave." << std::endl; + << "Max." << std::setw(data_width) << "Ave." + << std::setw(data_width) << "Ratio." << std::endl; for (size_t i = 0; i < events_table.size(); ++i) { for (size_t j = 0; j < events_table[i].size(); ++j) { const EventItem& event_item = events_table[i][j]; @@ -319,7 +311,9 @@ void PrintProfiler(const std::vector>& events_table, << std::setw(data_width) << event_item.total_time << std::setw(data_width) << event_item.min_time << std::setw(data_width) << event_item.max_time - << std::setw(data_width) << event_item.ave_time << std::endl; + << std::setw(data_width) << event_item.ave_time + << std::setw(data_width) << event_item.total_time / total + << std::endl; } } std::cout << std::endl; @@ -369,6 +363,7 @@ void ParseEvents(const std::vector>& events, std::vector> events_table; size_t max_name_width = 0; + double total = 0.; // the total time for (size_t i = 0; i < events.size(); i++) { std::list pushed_events; std::vector event_items; @@ -389,6 +384,7 @@ void ParseEvents(const std::vector>& events, g_state == ProfilerState::kAll) ? rit->CudaElapsedMs(events[i][j]) : rit->CpuElapsedMs(events[i][j]); + total += event_time; std::string event_name = "thread" + std::to_string(rit->thread_id()) + "::" + rit->name(); @@ -397,7 +393,8 @@ void ParseEvents(const std::vector>& events, if (event_idx.find(event_name) == event_idx.end()) { event_idx[event_name] = event_items.size(); EventItem event_item = {event_name, 1, event_time, - event_time, event_time, event_time}; + event_time, event_time, event_time, + 0.}; event_items.push_back(event_item); } else { int index = event_idx[event_name]; @@ -441,7 +438,7 @@ void ParseEvents(const std::vector>& events, } // Print report - PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12); + PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, total); } void DisableProfiler(EventSortingKey sorted_key, diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index bf43925373a12cd9ff2155d68c42d0266ba4df60..c99d9c807d1bfb45d1ce0725b84b9fff09049511 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -95,11 +95,6 @@ struct RecordBlock { uint64_t start_ns_; }; -struct RecordThread { - explicit RecordThread(int thread_id); - ~RecordThread(); -}; - // Return the event list of all threads. Assumed the returned value calls // event_lists, event_lists[i][j] represents the j-th Event of i-th thread. std::vector> GetAllEvents(); diff --git a/paddle/fluid/platform/variant.h b/paddle/fluid/platform/variant.h index 45f60fc9d76560b133fa06198a24c7eaccc24088..dc9fad29f281a1c6ac300b48f9e600ff802a5752 100644 --- a/paddle/fluid/platform/variant.h +++ b/paddle/fluid/platform/variant.h @@ -38,6 +38,7 @@ limitations under the License. */ #endif #endif +#include #include #include #include diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt index 4fef351c2118e43697606c90a616cd870e78cd77..89ca4f781273e99bbb83216c238dfc5c88c0a22b 100644 --- a/paddle/fluid/pybind/CMakeLists.txt +++ b/paddle/fluid/pybind/CMakeLists.txt @@ -2,13 +2,13 @@ if(WITH_PYTHON) if(WITH_AMD_GPU) hip_library(paddle_pybind SHARED SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc - DEPS pybind python proto_desc memory executor prune init profiler feed_fetch_method + DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method parallel_executor ${GLOB_OP_LIB}) else() cc_library(paddle_pybind SHARED SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc - DEPS pybind python proto_desc memory executor prune init profiler feed_fetch_method + DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method parallel_executor ${GLOB_OP_LIB}) if(NOT APPLE AND NOT ANDROID) diff --git a/paddle/fluid/pybind/exception.cc b/paddle/fluid/pybind/exception.cc index 08a2f185e117718d07ba984f76dfe5bf8229c33c..831f30e35fd3e01ce0f0524f6f85dd59494f5353 100644 --- a/paddle/fluid/pybind/exception.cc +++ b/paddle/fluid/pybind/exception.cc @@ -18,10 +18,13 @@ namespace paddle { namespace pybind { void BindException(pybind11::module* m) { + static pybind11::exception eof(*m, "EOFException"); static pybind11::exception exc(*m, "EnforceNotMet"); pybind11::register_exception_translator([](std::exception_ptr p) { try { if (p) std::rethrow_exception(p); + } catch (const platform::EOFException& e) { + eof(e.what()); } catch (const platform::EnforceNotMet& e) { exc(e.what()); } diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index bcf6d4dd3087060c016e53722cde80704ef2e834..be623703c2480774bb04a6bc0c5b00b699d7bb16 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -145,14 +145,14 @@ void BindBlockDesc(pybind11::module *m) { .def_property_readonly("id", &pd::BlockDesc::ID) .def_property_readonly("parent", &pd::BlockDesc::Parent) .def("get_forward_block_idx", &pd::BlockDesc::ForwardBlockID) - .def("set_forward_block_idx", &pd::BlockDesc::SetForwardBlockID) + .def("_set_forward_block_idx", &pd::BlockDesc::SetForwardBlockID) .def("append_op", &pd::BlockDesc::AppendOp, pybind11::return_value_policy::reference) - .def("prepend_op", &pd::BlockDesc::PrependOp, + .def("_prepend_op", &pd::BlockDesc::PrependOp, pybind11::return_value_policy::reference) - .def("insert_op", &pd::BlockDesc::InsertOp, + .def("_insert_op", &pd::BlockDesc::InsertOp, pybind11::return_value_policy::reference) - .def("remove_op", &pd::BlockDesc::RemoveOp) + .def("_remove_op", &pd::BlockDesc::RemoveOp) .def("var", [](pd::BlockDesc &self, pybind11::bytes byte_name) { std::string name = byte_name; @@ -165,7 +165,7 @@ void BindBlockDesc(pybind11::module *m) { return self.HasVar(name); }, pybind11::return_value_policy::reference) - .def("rename_var", + .def("_rename_var", [](pd::BlockDesc &self, const pybind11::bytes &byte_name, const pybind11::bytes &byte_name_new) { std::string name = byte_name; @@ -189,7 +189,7 @@ void BindBlockDesc(pybind11::module *m) { return self.FindVarRecursive(name); }, pybind11::return_value_policy::reference) - .def("remove_var", + .def("_remove_var", [](pd::BlockDesc &self, pybind11::bytes byte_name) { std::string name = byte_name; return self.RemoveVar(name); @@ -268,7 +268,8 @@ void BindOpDesc(pybind11::module *m) { .value("STRINGS", pd::proto::AttrType::STRINGS) .value("BOOL", pd::proto::AttrType::BOOLEAN) .value("BOOLS", pd::proto::AttrType::BOOLEANS) - .value("BLOCK", pd::proto::AttrType::BLOCK); + .value("BLOCK", pd::proto::AttrType::BLOCK) + .value("BLOCKS", pd::proto::AttrType::BLOCKS); pybind11::class_ op_desc(*m, "OpDesc", ""); op_desc @@ -293,13 +294,15 @@ void BindOpDesc(pybind11::module *m) { .def("set_attr", &pd::OpDesc::SetAttr) .def("attr", &pd::OpDesc::GetAttr) .def("set_block_attr", &pd::OpDesc::SetBlockAttr) + .def("set_blocks_attr", &pd::OpDesc::SetBlocksAttr) .def("set_serialized_attr", [](pd::OpDesc &self, const std::string &name, const pybind11::bytes &seriralized) { std::string ser(seriralized); self.SetAttr(name, ser); }) - .def("block_attr", &pd::OpDesc::GetBlockAttr) + .def("block_attr_id", &pd::OpDesc::GetBlockAttrId) + .def("blocks_attr_ids", &pd::OpDesc::GetBlocksAttrIds) .def("check_attrs", &pd::OpDesc::CheckAttrs) .def("infer_shape", &pd::OpDesc::InferShape) .def("infer_var_type", &pd::OpDesc::InferVarType) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 3af8941be69fe507bc105e26b608ec768e4b5998..7127bb38f6ddf8a55c1741d1f0ef18c8d9067fba 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include #include #include +#include #include // NOLINT // for call_once #include #include @@ -24,7 +25,6 @@ limitations under the License. */ #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/feed_fetch_method.h" #include "paddle/fluid/framework/framework.pb.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/lod_rank_table.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/lod_tensor_array.h" @@ -34,7 +34,9 @@ limitations under the License. */ #include "paddle/fluid/framework/reader.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/operators/activation_op.h" +#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h" #include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/init.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/pybind/const_value.h" @@ -65,6 +67,14 @@ bool IsCompiledWithCUDA() { #endif } +bool IsCompiledWithDIST() { +#ifdef PADDLE_WITH_DISTRIBUTE + return true; +#else + return false; +#endif +} + PYBIND11_PLUGIN(core) { py::module m("core", "C++ core of PaddlePaddle"); @@ -77,37 +87,37 @@ PYBIND11_PLUGIN(core) { py::class_(m, "Tensor", py::buffer_protocol()) .def_buffer( [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); }) - .def("get_dims", + .def("_get_dims", [](const Tensor &self) { return vectorize(self.dims()); }) - .def("set_dims", + .def("_set_dims", [](Tensor &self, const std::vector &dim) { self.Resize(make_ddim(dim)); }) - .def("set_layout", + .def("_set_layout", [](Tensor &self, const std::string &layout) { self.set_layout(StringToDataLayout(layout)); }) - .def("alloc_float", + .def("_alloc_float", [](Tensor &self, paddle::platform::CUDAPlace &place) { self.mutable_data(place); }) - .def("alloc_float", + .def("_alloc_float", [](Tensor &self, paddle::platform::CPUPlace &place) { self.mutable_data(place); }) - .def("alloc_int", + .def("_alloc_int", [](Tensor &self, paddle::platform::CPUPlace &place) { self.mutable_data(place); }) - .def("alloc_int", + .def("_alloc_int", [](Tensor &self, paddle::platform::CUDAPlace &place) { self.mutable_data(place); }) - .def("alloc_int", + .def("_alloc_int", [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) { self.mutable_data(place); }) - .def("alloc_float", + .def("_alloc_float", [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) { self.mutable_data(place); }) @@ -135,37 +145,84 @@ PYBIND11_PLUGIN(core) { .def("set", PyCUDAPinnedTensorSetFromArray) #endif .def("shape", [](Tensor &self) { return vectorize(self.dims()); }) - .def("set_float_element", TensorSetElement) - .def("get_float_element", TensorGetElement) - .def("set_double_element", TensorSetElement) - .def("get_double_element", TensorGetElement) - .def("dtype", [](Tensor &self) { return ToDataType(self.type()); }); + .def("_set_float_element", TensorSetElement) + .def("_get_float_element", TensorGetElement) + .def("_set_double_element", TensorSetElement) + .def("_get_double_element", TensorGetElement) + .def("_dtype", [](Tensor &self) { return ToDataType(self.type()); }); py::class_(m, "LoDTensor") .def_buffer( [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); }) - .def( - "__init__", - [](LoDTensor &instance, const std::vector> &lod) { - LoD new_lod; - new_lod.reserve(lod.size()); - std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); - new (&instance) LoDTensor(new_lod); - }) + .def("__init__", + [](LoDTensor &instance, const std::vector> + &recursive_sequence_lengths) { + LoD new_lod; + new_lod.reserve(recursive_sequence_lengths.size()); + std::copy(recursive_sequence_lengths.begin(), + recursive_sequence_lengths.end(), + std::back_inserter(new_lod)); + LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod); + PADDLE_ENFORCE( + CheckLoD(new_offset_lod, -1), + "the provided recursive_sequence_lengths info is invalid"); + new (&instance) LoDTensor(new_offset_lod); + }) .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); }) + // We implement offset based LOD in C++ while we use length based with + // Python API. So we changed set_lod to set_recursive_sequence_lengths to + // avoid misuse. + // The discussion is here: + // https://github.com/PaddlePaddle/Paddle/issues/10855 .def("set_lod", [](LoDTensor &self, const std::vector> &lod) { + // the input lod is offset-based level-of-detail info LoD new_lod; new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); + PADDLE_ENFORCE(CheckLoD(new_lod, vectorize(self.dims()).front()), + "the provided lod info is invalid"); self.set_lod(new_lod); }) - .def("lod", [](LoDTensor &self) -> std::vector> { - auto lod = self.lod(); - std::vector> new_lod; - new_lod.reserve(lod.size()); - std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); - return new_lod; + .def("set_recursive_sequence_lengths", + [](LoDTensor &self, const std::vector> + &recursive_sequence_lengths) { + // the input recursive_sequence_lengths is length-based + // level-of-detail info + LoD new_lod; + new_lod.reserve(recursive_sequence_lengths.size()); + std::copy(recursive_sequence_lengths.begin(), + recursive_sequence_lengths.end(), + std::back_inserter(new_lod)); + LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod); + PADDLE_ENFORCE( + CheckLoD(new_offset_lod, vectorize(self.dims()).front()), + "the provided recursive_sequence_lengths info is invalid"); + self.set_lod(new_offset_lod); + }) + .def("lod", + [](LoDTensor &self) -> std::vector> { + // output the offset-based lod info + LoD lod = self.lod(); + std::vector> new_lod; + new_lod.reserve(lod.size()); + std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); + return new_lod; + }) + // Set above comments of set_lod. + .def("recursive_sequence_lengths", + [](LoDTensor &self) -> std::vector> { + // output the length-based lod info + LoD lod = ConvertToLengthBasedLoD(self.lod()); + std::vector> new_lod; + new_lod.reserve(lod.size()); + std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); + return new_lod; + }) + .def("has_valid_recursive_sequence_lengths", [](LoDTensor &self) -> bool { + // Check that the lod info is valid and match the outermost + // dimension of the LoDTensor data + return CheckLoD(self.lod(), vectorize(self.dims()).front()); }); py::class_(m, "SelectedRows") @@ -191,15 +248,11 @@ PYBIND11_PLUGIN(core) { #endif }) .def("rows", [](SelectedRows &self) { -#ifndef PADDLE_WITH_CUDA - return self.rows(); -#else - auto rows = self.rows(); - std::vector new_rows; - new_rows.reserve(rows.size()); - std::copy(rows.begin(), rows.end(), std::back_inserter(new_rows)); - return new_rows; -#endif + auto rows = self.rows(); + std::vector new_rows; + new_rows.reserve(rows.size()); + std::copy(rows.begin(), rows.end(), std::back_inserter(new_rows)); + return new_rows; }); py::class_(m, "Variable", R"DOC(Variable Class. @@ -248,7 +301,39 @@ All parameter, weight, gradient are variables in Paddle. py::return_value_policy::reference); py::class_(m, "Reader", "") - .def("reset", &framework::ReaderHolder::ReInit); + .def("reset", &framework::ReaderHolder::ResetAll); + + using LoDTensorBlockingQueue = + ::paddle::operators::reader::LoDTensorBlockingQueue; + using LoDTensorBlockingQueueHolder = + ::paddle::operators::reader::LoDTensorBlockingQueueHolder; + py::class_>( + m, "LoDTensorBlockingQueue", "") + .def("push", + [](LoDTensorBlockingQueue &self, + const std::vector &lod_tensor_vec) { + pybind11::gil_scoped_release release; + return self.Push(lod_tensor_vec); + }) + .def("size", &LoDTensorBlockingQueue::Size) + .def("capacity", &LoDTensorBlockingQueue::Cap) + .def("close", &LoDTensorBlockingQueue::Close) + .def("is_closed", &LoDTensorBlockingQueue::IsClosed); + + m.def("init_lod_tensor_blocking_queue", + [](Variable &var, size_t capacity, + const std::vector> &shapes) + -> std::shared_ptr { + std::vector dims(shapes.size()); + std::transform(shapes.begin(), shapes.end(), dims.begin(), + [](const std::vector &shape) { + return make_ddim(shape); + }); + auto *holder = var.GetMutable(); + holder->InitOnce(capacity, dims); + return holder->GetQueue(); + }, + py::return_value_policy::copy); py::class_(m, "Scope", "") .def("var", @@ -309,8 +394,10 @@ All parameter, weight, gradient are variables in Paddle. InferenceOptimize(*(origin.Proto()), &pruned_desc); return new ProgramDesc(pruned_desc); }); - m.def("empty_var_name", []() { return framework::kEmptyVarName; }); - m.def("grad_var_suffix", []() { return framework::kGradVarSuffix; }); + m.def("empty_var_name", + []() { return std::string(framework::kEmptyVarName); }); + m.def("grad_var_suffix", + []() { return std::string(framework::kGradVarSuffix); }); m.def_submodule( "var_names", "The module will return special predefined variable name in Paddle") @@ -413,9 +500,12 @@ All parameter, weight, gradient are variables in Paddle. py::class_(m, "Executor") .def(py::init()) - .def("run", - (void (Executor::*)(const ProgramDesc &, Scope *, int, bool, bool)) & - Executor::Run); + .def("close", &Executor::Close) + .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope, + int block_id, bool create_local_scope, bool create_vars) { + pybind11::gil_scoped_release release; + self.Run(prog, scope, block_id, create_local_scope, create_vars); + }); m.def("init_gflags", framework::InitGflags); m.def("init_glog", framework::InitGLOG); @@ -423,6 +513,7 @@ All parameter, weight, gradient are variables in Paddle. [](bool init_p2p) { framework::InitDevices(init_p2p); }); m.def("is_compiled_with_cuda", IsCompiledWithCUDA); + m.def("is_compiled_with_dist", IsCompiledWithDIST); #ifdef PADDLE_WITH_CUDA m.def("is_float16_supported", [](const platform::CUDAPlace &place) -> bool { // Only GPUs with Compute Capability >= 53 support float16 @@ -449,6 +540,8 @@ All parameter, weight, gradient are variables in Paddle. }); py::class_(m, "LoDTensorArray") + .def("__init__", + [](LoDTensorArray &instance) { new (&instance) LoDTensorArray(); }) .def("__getitem__", [](LoDTensorArray &self, size_t i) { return &self.at(i); }, py::return_value_policy::reference) @@ -509,16 +602,24 @@ All parameter, weight, gradient are variables in Paddle. self.num_threads_ = num_threads; }) .def_property( - "use_event", - [](const ExecutionStrategy &self) { return self.use_event_; }, - [](ExecutionStrategy &self, bool use_event) { - self.use_event_ = use_event; + "use_cuda", + [](const ExecutionStrategy &self) { return self.use_cuda_; }, + [](ExecutionStrategy &self, bool use_cuda) { + self.use_cuda_ = use_cuda; }) .def_property( "allow_op_delay", [](const ExecutionStrategy &self) { return self.allow_op_delay_; }, [](ExecutionStrategy &self, bool allow_op_delay) { self.allow_op_delay_ = allow_op_delay; + }) + .def_property( + "num_iteration_per_drop_scope", + [](const ExecutionStrategy &self) { + return self.num_iteration_per_drop_scope_; + }, + [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) { + self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope; }); py::class_ build_strategy(pe, "BuildStrategy"); @@ -545,7 +646,17 @@ All parameter, weight, gradient are variables in Paddle. [](BuildStrategy &self, BuildStrategy::GradientScaleStrategy strategy) { self.gradient_scale_ = strategy; - }); + }) + .def_property( + "debug_graphviz_path", + [](const BuildStrategy &self) { return self.debug_graphviz_path_; }, + [](BuildStrategy &self, const std::string &path) { + self.debug_graphviz_path_ = path; + }) + .def_property( + "enable_data_balance", + [](const BuildStrategy &self) { return self.enable_data_balance_; }, + [](BuildStrategy &self, bool b) { self.enable_data_balance_ = b; }); pe.def(py::init &, const std::unordered_set &, @@ -553,7 +664,7 @@ All parameter, weight, gradient are variables in Paddle. const std::string &, Scope *, std::vector &, const ExecutionStrategy &, const BuildStrategy &, size_t, size_t>()) - .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs) + .def("_bcast_params", &ParallelExecutor::BCastParamsToDevices) // NOTE: even we return a vec* to Python use reference policy. // We still cannot get local_scope from this vector, since the element // of vec will be freed by Python GC. We can only return Scope* @@ -567,7 +678,12 @@ All parameter, weight, gradient are variables in Paddle. &ParallelExecutor::FeedTensorsIntoLocalScopes) .def("feed_and_split_tensor_into_local_scopes", &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes) - .def("run", &ParallelExecutor::Run); + .def("run", [](ParallelExecutor &self, + const std::vector &fetch_tensors, + const std::string &fetched_var_name) { + pybind11::gil_scoped_release release; + self.Run(fetch_tensors, fetched_var_name); + }); BindRecordIOWriter(&m); return m.ptr(); diff --git a/paddle/fluid/pybind/recordio.cc b/paddle/fluid/pybind/recordio.cc index 330d104e0a774d905e463566f85bd2e64a080190..f83b026d4d50772b969c4316964b70a68b27442b 100644 --- a/paddle/fluid/pybind/recordio.cc +++ b/paddle/fluid/pybind/recordio.cc @@ -30,7 +30,9 @@ class RecordIOWriter { public: RecordIOWriter(const std::string& filename, recordio::Compressor compressor, size_t max_num_record) - : stream_(filename), writer_(&stream_, compressor, max_num_record) {} + : closed_(false), + stream_(filename), + writer_(&stream_, compressor, max_num_record) {} void AppendTensor(const framework::LoDTensor& tensor) { tensors_.push_back(tensor); @@ -47,9 +49,17 @@ class RecordIOWriter { PADDLE_ENFORCE(tensors_.empty()); writer_.Flush(); stream_.close(); + closed_ = true; + } + + ~RecordIOWriter() { + if (!closed_) { + Close(); + } } private: + bool closed_; std::vector tensors_; std::ofstream stream_; recordio::Writer writer_; diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h index 93b09ed6922b32a5531224acc470daf0d97f95bd..3e2ea1ef88b03f5b2576c1cee2b5d26a439943da 100644 --- a/paddle/fluid/pybind/tensor_py.h +++ b/paddle/fluid/pybind/tensor_py.h @@ -97,7 +97,7 @@ struct CastToPyBufferImpl { inline pybind11::buffer_info CastToPyBuffer(const framework::Tensor &tensor) { auto buffer_info = details::CastToPyBufferImpl()(tensor); + uint8_t, platform::float16>()(tensor); return buffer_info; } @@ -146,7 +146,7 @@ void PyCPUTensorSetFromArray( template <> // This following specialization maps uint16_t in the parameter type to // platform::float16. -void PyCPUTensorSetFromArray( +inline void PyCPUTensorSetFromArray( framework::Tensor *self, pybind11::array_t @@ -185,7 +185,7 @@ void PyCUDATensorSetFromArray( template <> // This following specialization maps uint16_t in the parameter type to // platform::float16. -void PyCUDATensorSetFromArray( +inline void PyCUDATensorSetFromArray( framework::Tensor *self, pybind11::array_t @@ -224,7 +224,7 @@ void PyCUDAPinnedTensorSetFromArray( template <> // This following specialization maps uint16_t in the parameter type to // platform::float16. -void PyCUDAPinnedTensorSetFromArray( +inline void PyCUDAPinnedTensorSetFromArray( framework::Tensor *self, pybind11::array_t diff --git a/paddle/fluid/recordio/chunk.cc b/paddle/fluid/recordio/chunk.cc index 82d9aa601cf450b8f90573d6c582bb12ced7a48a..6c65d9160c059ac143ee258b2bdaed5915a1dca1 100644 --- a/paddle/fluid/recordio/chunk.cc +++ b/paddle/fluid/recordio/chunk.cc @@ -119,40 +119,56 @@ bool Chunk::Write(std::ostream& os, Compressor ct) const { } bool Chunk::Parse(std::istream& sin) { - Header hdr; - bool ok = hdr.Parse(sin); + ChunkParser parser(sin); + if (!parser.Init()) { + return false; + } + Clear(); + while (parser.HasNext()) { + Add(parser.Next()); + } + return true; +} + +ChunkParser::ChunkParser(std::istream& sin) : in_(sin) {} +bool ChunkParser::Init() { + pos_ = 0; + bool ok = header_.Parse(in_); if (!ok) { return ok; } - auto beg_pos = sin.tellg(); - uint32_t crc = Crc32Stream(sin, hdr.CompressSize()); - PADDLE_ENFORCE_EQ(hdr.Checksum(), crc); - Clear(); - sin.seekg(beg_pos, sin.beg); - std::unique_ptr compressed_stream; - switch (hdr.CompressType()) { + auto beg_pos = in_.tellg(); + uint32_t crc = Crc32Stream(in_, header_.CompressSize()); + PADDLE_ENFORCE_EQ(header_.Checksum(), crc); + in_.seekg(beg_pos, in_.beg); + + switch (header_.CompressType()) { case Compressor::kNoCompress: break; case Compressor::kSnappy: - compressed_stream.reset(new snappy::iSnappyStream(sin)); + compressed_stream_.reset(new snappy::iSnappyStream(in_)); break; default: PADDLE_THROW("Not implemented"); } + return true; +} - std::istream& stream = compressed_stream ? *compressed_stream : sin; +bool ChunkParser::HasNext() const { return pos_ < header_.NumRecords(); } - for (uint32_t i = 0; i < hdr.NumRecords(); ++i) { - uint32_t rec_len; - stream.read(reinterpret_cast(&rec_len), sizeof(uint32_t)); - std::string buf; - buf.resize(rec_len); - stream.read(&buf[0], rec_len); - PADDLE_ENFORCE_EQ(rec_len, stream.gcount()); - Add(buf); +std::string ChunkParser::Next() { + if (!HasNext()) { + return ""; } - return true; + ++pos_; + std::istream& stream = compressed_stream_ ? *compressed_stream_ : in_; + uint32_t rec_len; + stream.read(reinterpret_cast(&rec_len), sizeof(uint32_t)); + std::string buf; + buf.resize(rec_len); + stream.read(&buf[0], rec_len); + PADDLE_ENFORCE_EQ(rec_len, stream.gcount()); + return buf; } - } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/recordio/chunk.h b/paddle/fluid/recordio/chunk.h index 71a1556a33bfa5c937d6a799d2818cd5a5ef2094..cfb954a591679c2d2c4f42ecd99ca0c8bd1084cf 100644 --- a/paddle/fluid/recordio/chunk.h +++ b/paddle/fluid/recordio/chunk.h @@ -13,6 +13,7 @@ // limitations under the License. #pragma once +#include #include #include @@ -53,9 +54,20 @@ class Chunk { DISABLE_COPY_AND_ASSIGN(Chunk); }; -size_t CompressData(const char* in, size_t in_length, Compressor ct, char* out); +class ChunkParser { + public: + explicit ChunkParser(std::istream& sin); + + bool Init(); + std::string Next(); + bool HasNext() const; -void DeflateData(const char* in, size_t in_length, Compressor ct, char* out); + private: + Header header_; + uint32_t pos_{0}; + std::istream& in_; + std::unique_ptr compressed_stream_; +}; } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/recordio/scanner.cc b/paddle/fluid/recordio/scanner.cc index 88b4d4001bc1b6dc935a9aabc2db5edfb55a60e4..a0a2f984228db0e7a015630655a3176aa4d1a5a4 100644 --- a/paddle/fluid/recordio/scanner.cc +++ b/paddle/fluid/recordio/scanner.cc @@ -22,35 +22,34 @@ namespace paddle { namespace recordio { Scanner::Scanner(std::unique_ptr &&stream) - : stream_(std::move(stream)) { + : stream_(std::move(stream)), parser_(*stream_) { Reset(); } -Scanner::Scanner(const std::string &filename) { - stream_.reset(new std::ifstream(filename)); +Scanner::Scanner(const std::string &filename) + : stream_(new std::ifstream(filename)), parser_(*stream_) { + PADDLE_ENFORCE(static_cast(*stream_), "Cannot open file %s", filename); Reset(); } void Scanner::Reset() { stream_->clear(); stream_->seekg(0, std::ios::beg); - ParseNextChunk(); + parser_.Init(); } std::string Scanner::Next() { - PADDLE_ENFORCE(!eof_, "StopIteration"); - auto rec = cur_chunk_.Record(offset_++); - if (offset_ == cur_chunk_.NumRecords()) { - ParseNextChunk(); + if (stream_->eof()) { + return ""; } - return rec; -} -void Scanner::ParseNextChunk() { - eof_ = !cur_chunk_.Parse(*stream_); - offset_ = 0; + auto res = parser_.Next(); + if (!parser_.HasNext() && HasNext()) { + parser_.Init(); + } + return res; } -bool Scanner::HasNext() const { return !eof_; } +bool Scanner::HasNext() const { return !stream_->eof(); } } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/recordio/scanner.h b/paddle/fluid/recordio/scanner.h index 34f1b0c78d6b5af6072a993579e1866d38c6d009..0d885dd87a2f819ba1d9f76259196f6cfff0b2a0 100644 --- a/paddle/fluid/recordio/scanner.h +++ b/paddle/fluid/recordio/scanner.h @@ -37,11 +37,7 @@ class Scanner { private: std::unique_ptr stream_; - Chunk cur_chunk_; - size_t offset_; - bool eof_; - - void ParseNextChunk(); + ChunkParser parser_; }; } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/string/printf.h b/paddle/fluid/string/printf.h index 062095a1c3e977c0bcc89346ead765acb023bcf7..47de23377398423dabf3b0ed5b670e564f57cdfb 100644 --- a/paddle/fluid/string/printf.h +++ b/paddle/fluid/string/printf.h @@ -83,6 +83,13 @@ void Fprintf(std::ostream& out, const char* fmt, const Args&... args) { tinyformat::vformat(out, fmt, tinyformat::makeFormatList(args...)); } +template +std::string Sprintf(const Args&... args) { + std::ostringstream oss; + Fprintf(oss, ""); + return oss.str(); +} + template std::string Sprintf(const char* fmt, const Args&... args) { std::ostringstream oss; diff --git a/paddle/fluid/string/printf_test.cc b/paddle/fluid/string/printf_test.cc index 678029f93534ab374bd29083f8991d632ccdd5a1..544b12ef3a877a6e84c136433799301edaa4abdf 100644 --- a/paddle/fluid/string/printf_test.cc +++ b/paddle/fluid/string/printf_test.cc @@ -27,4 +27,5 @@ TEST(StringPrintf, StringPrintf) { EXPECT_EQ(std::string("Wednesday, July 27, 14:44"), paddle::string::Sprintf("%s, %s %d, %.2d:%.2d", weekday, month, day, hour, min)); + EXPECT_EQ(std::string(""), paddle::string::Sprintf()); } diff --git a/paddle/fluid/train/demo/demo_trainer.cc b/paddle/fluid/train/demo/demo_trainer.cc index 813d8386868558bd62a9d5670d540ddeddb2b77d..a0757b53f37b29de0b3802c345b1ad9db69f16e9 100644 --- a/paddle/fluid/train/demo/demo_trainer.cc +++ b/paddle/fluid/train/demo/demo_trainer.cc @@ -12,15 +12,17 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include #include #include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/init.h" #include "paddle/fluid/platform/place.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace train { @@ -93,11 +95,21 @@ int main() { auto loss_var = scope.Var(loss_name); + paddle::platform::ProfilerState pf_state; + pf_state = paddle::platform::ProfilerState::kCPU; + paddle::platform::EnableProfiler(pf_state); + clock_t t1 = clock(); + for (int i = 0; i < 10; ++i) { executor.Run(*train_program.get(), &scope, 0, false, true); std::cout << "step: " << i << " loss: " << loss_var->Get().data()[0] << std::endl; } + + clock_t t2 = clock(); + paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kTotal, + "run_paddle_op_profiler"); + std::cout << "run_time = " << t2 - t1 << std::endl; return 0; } diff --git a/paddle/function/BufferArg.cpp b/paddle/function/BufferArg.cpp deleted file mode 100644 index 2dc931c5d7e727679d435470544e60f9b5ce2bde..0000000000000000000000000000000000000000 --- a/paddle/function/BufferArg.cpp +++ /dev/null @@ -1,52 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "BufferArg.h" -#include "paddle/math/SparseMatrix.h" - -namespace paddle { - -const SequenceArg& BufferArg::sequence() const { - CHECK_EQ(bufferType_, TENSOR_SEQUENCE_DATA); - return dynamic_cast(*this); -} - -const SparseMatrixArg& BufferArg::sparse() const { - CHECK_EQ(bufferType_, TENSOR_SPARSE); - return dynamic_cast(*this); -} - -SparseMatrixArg::SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType) - : BufferArg(sparse, argType), - row_(reinterpret_cast(sparse.getRows()), VALUE_TYPE_INT32), - col_(reinterpret_cast(sparse.getCols()), VALUE_TYPE_INT32), - nnz_(sparse.getElementCnt()), - format_(static_cast(sparse.getFormat())), - type_(static_cast(sparse.getValueType())) { - bufferType_ = TENSOR_SPARSE; -} - -SparseMatrixArg::SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType) - : BufferArg(sparse, argType), - row_(reinterpret_cast(sparse.getRows()), VALUE_TYPE_INT32), - col_(reinterpret_cast(sparse.getCols()), VALUE_TYPE_INT32), - nnz_(sparse.getElementCnt()), - format_(static_cast(sparse.getFormat())), - type_(static_cast(sparse.getValueType())) { - bufferType_ = TENSOR_SPARSE; -} - -} // namespace paddle diff --git a/paddle/function/BufferArg.h b/paddle/function/BufferArg.h deleted file mode 100644 index 6de8c94e778c8d1439b2a2aa3c581a5a3cf70261..0000000000000000000000000000000000000000 --- a/paddle/function/BufferArg.h +++ /dev/null @@ -1,364 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include - -#include "TensorShape.h" -#include "TensorType.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -enum BufferType { - TENSOR_UNKNOWN = 0, - TENSOR_NORMAL = 1, - TENSOR_SEQUENCE_ID = 2, - TENSOR_SEQUENCE_DATA = 3, - TENSOR_SPARSE = 4 -}; - -class BufferArg; -class SequenceArg; -class SparseMatrixArg; - -/** - * \brief BufferArg used as the argument type of Function. - * - * The arguments of the Paddle Function have four Buffer types. - * 1. BufferArg for a dense Buffer of any dimension. - * 2. SequenceIdArg for a Buffer of sequence start positions. - * 3. SequenceArg for a Buffer of sequence data. - * 4. SparseMatrixArg for a Buffer of sparse matrix. - * - * Buffer shape - * For most buffers, the first dimension `shape()[0]` represents - * the size of the mini-batch. - * - * Buffer argType - * There is an ArgType property for the BufferArg used as Function Output. - * Whether the result of the Function calculation is assigned to the - * output Buffer or added to the output Buffer is determined by the - * argType_ property of the output BufferArg. - */ - -// ArgType is only used by output BufferArg. -// For input argument, argType_ is ignored. -// For output argument, need to set the argType_ of the BufferArg. -enum ArgType { - UNSPECIFIED = 0, - ASSIGN_TO = 1, - ADD_TO = 2, -}; -class BufferArg { - public: - void setArgType(ArgType argType) { argType_ = argType; } - - ArgType getArgType() const { return argType_; } - - public: - BufferArg(ValueType valueType, - const TensorShape& shape, - ArgType argType = UNSPECIFIED) - : buf_(nullptr), valueType_(valueType), shape_(shape), argType_(argType) { - bufferType_ = TENSOR_NORMAL; - } - - BufferArg(void* buf, - ValueType valueType, - const TensorShape& shape, - ArgType argType = UNSPECIFIED) - : buf_(buf), valueType_(valueType), shape_(shape), argType_(argType) { - bufferType_ = TENSOR_NORMAL; - } - - BufferArg(void* buf, ValueType valueType) : buf_(buf), valueType_(valueType) { - bufferType_ = TENSOR_NORMAL; - } - - BufferArg(const Matrix& matrix, ArgType argType = UNSPECIFIED) - : buf_( - const_cast(reinterpret_cast(matrix.getData()))), - valueType_(DataType::value), - shape_(2), - argType_(argType) { - bufferType_ = TENSOR_NORMAL; - shape_.setDim(0, matrix.getHeight()); - shape_.setDim(1, matrix.getWidth()); - } - - BufferArg(const Matrix& matrix, - const TensorShape& shape, - ArgType argType = UNSPECIFIED) - : buf_( - const_cast(reinterpret_cast(matrix.getData()))), - valueType_(DataType::value), - shape_(shape), - argType_(argType) { - bufferType_ = TENSOR_NORMAL; - CHECK_EQ(matrix.getElementCnt(), shape.getElements()); - } - - BufferArg(const Vector& vector, ArgType argType = UNSPECIFIED) - : buf_( - const_cast(reinterpret_cast(vector.getData()))), - valueType_(DataType::value), - shape_(1), - argType_(argType) { - bufferType_ = TENSOR_NORMAL; - shape_.setDim(0, vector.getSize()); - } - - BufferArg(const IVector& vector, ArgType argType = UNSPECIFIED) - : buf_( - const_cast(reinterpret_cast(vector.getData()))), - valueType_(VALUE_TYPE_INT32), - shape_(1), - argType_(argType) { - bufferType_ = TENSOR_NORMAL; - shape_.setDim(0, vector.getSize()); - } - - template - typename Tensor::Matrix matrix() const { - CHECK(buf_); - CHECK(valueType_ == DataType::value); - // CHECK(deviceType_ == DType); - CHECK_EQ((size_t)2, shape_.ndims()); - return typename Tensor::Matrix( - reinterpret_cast(buf_), shape_[0], shape_[1]); - } - - template - typename Tensor::Vector vector() const { - CHECK(buf_); - CHECK(valueType_ == DataType::value); - // CHECK(deviceType_ == DType); - CHECK_EQ((size_t)1, shape_.ndims()); - return typename Tensor::Vector( - shape_[0], reinterpret_cast(buf_)); - } - - virtual ~BufferArg() {} - - template - T* data() const { - return reinterpret_cast(buf_); - } - - void* data() const { return buf_; } - ValueType valueType() const { return valueType_; } - BufferType bufferType() const { return bufferType_; } - const TensorShape& shape() const { return shape_; } - bool isSparseArg() const { return TENSOR_SPARSE == bufferType_; } - bool isSequenceArg() const { return TENSOR_SEQUENCE_DATA == bufferType_; } - virtual size_t numElements() const { return shape_.getElements(); } - - const SequenceArg& sequence() const; - const SparseMatrixArg& sparse() const; - - protected: - void* buf_; - ValueType valueType_; - TensorShape shape_; - BufferType bufferType_{TENSOR_UNKNOWN}; - ArgType argType_{UNSPECIFIED}; - // TODO(tianbing), add deviceType_ - // leading dimensions. The size is dims_.size() - // Dims lds_; -}; - -// sequence start positions in a mini-batch of sequences -// shape_.ndims() == 1 -// valueType_ = int32 -// if a < b then value_.buf_[a] < value_.buf_[b] -class SequenceIdArg : public BufferArg { - public: - SequenceIdArg(const TensorShape& shape, ArgType argType = UNSPECIFIED) - : BufferArg(VALUE_TYPE_INT32, shape, argType) { - bufferType_ = TENSOR_SEQUENCE_ID; - CHECK_EQ(shape_.ndims(), 1UL); - CHECK_GE(shape_[0], 1UL); - numSeqs_ = shape_[0] - 1; - } - - SequenceIdArg(void* buf, - const TensorShape& shape, - ArgType argType = UNSPECIFIED) - : BufferArg(buf, VALUE_TYPE_INT32, shape, argType) { - bufferType_ = TENSOR_SEQUENCE_ID; - CHECK_EQ(shape_.ndims(), 1UL); - numSeqs_ = shape_[0] - 1; - } - - SequenceIdArg(const IVector& vector) : BufferArg(vector) { - bufferType_ = TENSOR_SEQUENCE_ID; - numSeqs_ = shape_[0] - 1; - } - - ~SequenceIdArg() {} - - size_t numSeqs() const { return numSeqs_; } - - private: - size_t numSeqs_; -}; - -// sequences data -// For mini-batch calculate, -// one batch can contain more than one sequence of data. -// SequenceArg can be used to represent sequences that contain multiple -// unequal lengths. -class SequenceArg : public BufferArg { - public: - SequenceArg(ValueType valueType, - const TensorShape& shape, - ArgType argType = UNSPECIFIED) - : BufferArg(valueType, shape, argType), - startPositions_(TensorShape({shape[0]})) { - bufferType_ = TENSOR_SEQUENCE_DATA; - } - - SequenceArg(void* buf, - ValueType valueType, - const TensorShape& shape, - const SequenceIdArg& startPositions, - ArgType argType = UNSPECIFIED) - : BufferArg(buf, valueType, shape, argType), - startPositions_(startPositions) { - bufferType_ = TENSOR_SEQUENCE_DATA; - } - - SequenceArg(const Matrix& matrix, - const IVector& vector, - ArgType argType = UNSPECIFIED) - : BufferArg(matrix, argType), startPositions_(vector) { - bufferType_ = TENSOR_SEQUENCE_DATA; - } - - ~SequenceArg() {} - - void* getIdBuf() const { return startPositions_.data(); } - size_t numSeqs() const { return startPositions_.numSeqs(); } - SequenceIdArg& getSequenceId() { return startPositions_; } - const SequenceIdArg& getSequenceId() const { return startPositions_; } - - private: - SequenceIdArg startPositions_; -}; - -// sparse matrix -// valueType_ == float or double -// shape_.ndims() == 2 -class SparseMatrixArg : public BufferArg { - public: - SparseMatrixArg(void* buf, - ValueType valueType, - const TensorShape& shape, - const BufferArg& row, - const BufferArg& col, - size_t nnz, - SparseFormat format, - SparseValueType type, - ArgType argType = UNSPECIFIED) - : BufferArg(buf, valueType, shape, argType), - row_(row), - col_(col), - nnz_(nnz), - format_(static_cast(format)), - type_(static_cast(type)) { - bufferType_ = TENSOR_SPARSE; - CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE)); - CHECK_EQ(shape_.ndims(), 2UL); - CHECK_EQ(row_.shape().ndims(), 1UL); - CHECK_EQ(col_.shape().ndims(), 1UL); - if (format_ == T_SPARSE_CSR) { - CHECK_EQ(nnz, col.shape()[0]); - } else if (format_ == T_SPARSE_CSC) { - CHECK_EQ(nnz, row.shape()[0]); - } - } - - SparseMatrixArg(ValueType valueType, - const TensorShape& shape, - size_t nnz, - SparseFormat format, - SparseValueType type, - ArgType argType = UNSPECIFIED) - : BufferArg(valueType, shape, argType), - row_(BufferArg(nullptr, VALUE_TYPE_INT32)), - col_(BufferArg(nullptr, VALUE_TYPE_INT32)), - nnz_(nnz), - format_(static_cast(format)), - type_(static_cast(type)) { - bufferType_ = TENSOR_SPARSE; - CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE)); - CHECK_EQ(shape_.ndims(), 2UL); - - /// len of row_ : height + 1 (CSR) or nnz (CSC), buf_ == nullptr - row_ = (format_ == T_SPARSE_CSR - ? BufferArg(VALUE_TYPE_INT32, TensorShape{shape_[0] + 1}) - : BufferArg(VALUE_TYPE_INT32, TensorShape{nnz})); - /// len of col_ : width + 1 (CSC) or nnz (CSR), buf_ == nullptr - col_ = (format_ == T_SPARSE_CSR - ? BufferArg(VALUE_TYPE_INT32, TensorShape{nnz}) - : BufferArg(VALUE_TYPE_INT32, TensorShape{shape_[1] + 1})); - } - - SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType = UNSPECIFIED); - - SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType = UNSPECIFIED); - - template - typename Tensor::SparseMatrix SparseMatrix() const { - CHECK(buf_); - CHECK(valueType_ == DataType::value); - // CHECK(deviceType_ == DType); - CHECK_EQ(2UL, shape_.ndims()); - return typename Tensor::SparseMatrix( - reinterpret_cast(buf_), - reinterpret_cast(row_.data()), - reinterpret_cast(col_.data()), - shape_[0], - shape_[1], - nnz_, - static_cast(type_), - static_cast(format_), - false); - } - - ~SparseMatrixArg() {} - - void* getRowBuf() const { return row_.data(); } - - void* getColBuf() const { return col_.data(); } - - size_t nnz() const { return nnz_; } - - size_t numElements() const override { return nnz_; } - - SparseDataFormat dataFormat() const { return format_; } - - SparseDataType dataType() const { return type_; } - - private: - BufferArg row_; - BufferArg col_; - size_t nnz_; - SparseDataFormat format_; - SparseDataType type_; -}; - -} // namespace paddle diff --git a/paddle/function/BufferArgTest.cpp b/paddle/function/BufferArgTest.cpp deleted file mode 100644 index 1a6e0110afb64c8b4f164d71e31e5f9bfcdee4a8..0000000000000000000000000000000000000000 --- a/paddle/function/BufferArgTest.cpp +++ /dev/null @@ -1,38 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "BufferArg.h" -#include -#include "paddle/math/MemoryHandle.h" - -namespace paddle { - -TEST(BufferTest, BufferArg) { - TensorShape shape({8, 10}); - CpuMemoryHandle memory(shape.getElements() * - sizeOfValuType(VALUE_TYPE_FLOAT)); - BufferArg buffer(memory.getBuf(), VALUE_TYPE_FLOAT, shape); - EXPECT_EQ(buffer.data(), memory.getBuf()); -} - -TEST(BufferTest, SequenceIdArg) { - TensorShape shape({10}); - CpuMemoryHandle memory(shape.getElements() * - sizeOfValuType(VALUE_TYPE_INT32)); - SequenceIdArg buffer(memory.getBuf(), shape); - EXPECT_EQ(buffer.data(), memory.getBuf()); - EXPECT_EQ(buffer.numSeqs(), 9U); -} - -} // namespace paddle diff --git a/paddle/function/ContextProjectionOp.cpp b/paddle/function/ContextProjectionOp.cpp deleted file mode 100644 index 1187842452460ac3fd71f48150fab6467f93dc6c..0000000000000000000000000000000000000000 --- a/paddle/function/ContextProjectionOp.cpp +++ /dev/null @@ -1,412 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ContextProjectionOp.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" - -namespace paddle { -/** - * Context Projection Forward with CPU Matrix Device. - * - */ -template <> -void ContextProjectionForward(CpuMatrix& out_mat, - const CpuMatrix& input_mat, - const CpuMatrix& weight_mat, - const CpuIVector& seq_vec, - size_t context_length, - int context_start, - size_t begin_pad) { - const int* starts = seq_vec.getData(); - const size_t num_sequences = seq_vec.getSize() - 1; - for (size_t i = 0; i < num_sequences; ++i) { - for (size_t j = 0; j < context_length; ++j) { - int begin = starts[i] + context_start + j; - int end = starts[i + 1] + context_start + j; - int dst_begin = starts[i]; - int dst_end = starts[i + 1]; - if (begin < starts[i]) { - int64_t pad_size = - std::min(starts[i] - begin, starts[i + 1] - starts[i]); - MatrixPtr mat = out_mat.subMatrix(starts[i], pad_size); - if (weight_mat) { - MatrixPtr sub = - const_cast(weight_mat).subMatrix(j, pad_size); - mat->addAtOffset(*sub, j * input_mat.getWidth()); - } - dst_begin = starts[i] + pad_size; - begin = starts[i]; - } - if (end > starts[i + 1]) { - int64_t pad_size = - std::min(end - starts[i + 1], starts[i + 1] - starts[i]); - MatrixPtr mat = out_mat.subMatrix(starts[i + 1] - pad_size, pad_size); - if (weight_mat) { - MatrixPtr sub = - const_cast(weight_mat) - .subMatrix(begin_pad + context_start + j - pad_size, - pad_size); - mat->addAtOffset(*sub, j * input_mat.getWidth()); - } - dst_end = starts[i + 1] - pad_size; - end = starts[i + 1]; - } - if (end <= begin) continue; - MatrixPtr src = - const_cast(input_mat).subMatrix(begin, end - begin); - MatrixPtr dst = out_mat.subMatrix(dst_begin, dst_end - dst_begin); - dst->addAtOffset(*src, j * input_mat.getWidth()); - } - } -} - -/** - * Paddle Function for Context Projection Forward. - * Calculate the output layer value sequence after context projection. - * - * What is Context Projection for a sequence? - * For example, assumed input (x) has 4 words and the dimension of each word - * representation is 2. If we use zero to pad instead of learned weight to pad, - * and the context_lenth is 3, the output (y) is: - * - * @code - * x = [a1, a2; - * b1, b2; - * c1, c2; - * d1, d2] - * y = [0, 0, a1, a2, b1, b2; - * a1, a2, b1, b2, c1, c2; - * b1, b2, c1, c2, d1, d2; - * c1, c2, d1, d2, 0, 0] - * @endcode - * - * \param outputs[0].matrix output layer value, n * (d * l) - * \param outputs[0].vector start position sequence, n * 1 - * \param inputs[0].matrix input layer value, n * d - * \param inputs[0].vector start position sequence, n * 1 - * \param inputs[1].matrix input layer weight, pad * d - */ -template -class ContextProjectionForwardFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - context_length_ = config.get("context_length"); - context_start_ = config.get("context_start"); - begin_pad_ = config.get("begin_pad"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK(1UL == inputs.size() || 2UL == inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) - << "SequenceArg required here"; - const auto val_seqs = dynamic_cast(inputs[0]); - auto out_seq = dynamic_cast(outputs[0]); - - CHECK(out_seq.data() && val_seqs.data() && val_seqs.getSequenceId().data()); - CHECK_EQ(out_seq.shape().ndims(), 2UL); - CHECK_EQ(val_seqs.shape().ndims(), 2UL); - /// dim of output = dim of input * context_length - CHECK_EQ(out_seq.shape()[1], val_seqs.shape()[1] * context_length_); - /// input and output has the same batch_size - CHECK_EQ(val_seqs.shape()[0], out_seq.shape()[0]); - if (2UL == inputs.size()) { - CHECK_EQ(inputs[1].shape().ndims(), 2UL); - /// dim of input == dim of weight - CHECK_EQ(val_seqs.shape()[1], inputs[1].shape()[1]); - } - - CHECK_EQ(out_seq.getArgType(), ADD_TO); - auto out_mat = out_seq.matrix(); - const auto in_mat = val_seqs.matrix(); - const auto w_mat = - (2UL == inputs.size() && inputs[1].data()) - ? inputs[1].matrix() - : typename Tensor::Matrix(nullptr, 0, 0); - const auto seq_vec = val_seqs.getSequenceId().vector(); - - ContextProjectionForward(out_mat, - in_mat, - w_mat, - seq_vec, - context_length_, - context_start_, - begin_pad_); - } - - private: - size_t context_length_; - int context_start_; - size_t begin_pad_; -}; - -/** - * Context Projection Backward with CPU Matrix Device. - * - */ -template <> -void ContextProjectionBackward(const CpuMatrix& out_grad_mat, - CpuMatrix& in_grad_mat, - CpuMatrix& w_grad_mat, - const CpuIVector& seq_vec, - size_t context_length, - int context_start, - size_t begin_pad, - bool is_padding, - size_t total_pad) { - size_t input_dim = in_grad_mat ? in_grad_mat.getWidth() - : w_grad_mat ? w_grad_mat.getWidth() : 0; - const int* starts = seq_vec.getData(); - size_t num_sequences = seq_vec.getSize() - 1; - for (size_t i = 0; i < num_sequences; ++i) { - for (size_t j = 0; j < context_length; ++j) { - int begin = starts[i] + context_start + j; - int end = starts[i + 1] + context_start + j; - int dst_begin = starts[i]; - int dst_end = starts[i + 1]; - if (begin < starts[i]) { - int64_t pad_size = - std::min(starts[i] - begin, starts[i + 1] - starts[i]); - if (is_padding && w_grad_mat) { - MatrixPtr mat = const_cast(out_grad_mat) - .subMatrix(starts[i], pad_size); - MatrixPtr sub = w_grad_mat.subMatrix(j, pad_size); - sub->addAtOffset(*mat, j * input_dim); - } - dst_begin = starts[i] + pad_size; - begin = starts[i]; - } - if (end > starts[i + 1]) { - int64_t pad_size = - std::min(end - starts[i + 1], starts[i + 1] - starts[i]); - if (is_padding && w_grad_mat) { - MatrixPtr mat = const_cast(out_grad_mat) - .subMatrix(starts[i + 1] - pad_size, pad_size); - MatrixPtr sub = w_grad_mat.subMatrix( - begin_pad + context_start + j - pad_size, pad_size); - sub->addAtOffset(*mat, j * input_dim); - } - dst_end = starts[i + 1] - pad_size; - end = starts[i + 1]; - } - if (end <= begin) continue; - if (!in_grad_mat) continue; - MatrixPtr src = in_grad_mat.subMatrix(begin, end - begin); - MatrixPtr dst = const_cast(out_grad_mat) - .subMatrix(dst_begin, dst_end - dst_begin); - src->addAtOffset(*dst, j * input_dim); - } - } -} - -/** - * Context Projection Backward Function. - * Update the weight gradient and input layer gradient with backprop - * - * \param inputs[0].matrix output layer grad, n * (d * l) - * \param inputs[0].vector start position sequence, n * 1 - * \param outputs[0].matrix input layer grad, n * d - * \param outputs[0].vector start position sequence, n * 1 - * \param outputs[1] weight grad, pad * d - */ -template -class ContextProjectionBackwardFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - context_length_ = config.get("context_length"); - context_start_ = config.get("context_start"); - begin_pad_ = config.get("begin_pad"); - is_padding_ = config.get("is_padding"); - total_pad_ = config.get("total_pad"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK(1UL == outputs.size() || 2UL == outputs.size()); - CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) - << "SequenceArg required here"; - const auto in_seq = dynamic_cast(inputs[0]); - auto out_seq = dynamic_cast(outputs[0]); - CHECK(in_seq.data() && in_seq.getSequenceId().data()); - CHECK_EQ(in_seq.shape().ndims(), 2UL); - CHECK_EQ(out_seq.shape().ndims(), 2UL); - CHECK_EQ(out_seq.getSequenceId().shape().ndims(), 1UL); - - /// input and output grad has the same batch_size - CHECK_EQ(out_seq.shape()[0], in_seq.shape()[0]); - /// dim of output grad = dim of input grad * context_length - CHECK_EQ(in_seq.shape()[1], out_seq.shape()[1] * context_length_); - CHECK_EQ(out_seq.getArgType(), ADD_TO); - - if (2UL == outputs.size()) { - CHECK_EQ(outputs[1].shape().ndims(), 2UL); - /// dim of input grad == dim of weight - CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]); - CHECK_EQ(outputs[1].getArgType(), ADD_TO); - } - - const auto seq_vec = in_seq.getSequenceId().vector(); - const auto out_grad_mat = in_seq.matrix(); - auto in_grad_mat = - !out_seq.data() ? typename Tensor::Matrix(nullptr, 0, 0) - : out_seq.matrix(); - auto w_grad_mat = - (2UL == outputs.size() && outputs[1].data()) - ? outputs[1].matrix() - : typename Tensor::Matrix(nullptr, 0, 0); - - ContextProjectionBackward(out_grad_mat, - in_grad_mat, - w_grad_mat, - seq_vec, - context_length_, - context_start_, - begin_pad_, - is_padding_, - total_pad_); - } - - private: - size_t context_length_; - int context_start_; - size_t begin_pad_; - bool is_padding_; - size_t total_pad_; -}; - -/** - * Context Projection Backward Data Function - * Update input layer grad - * input: sequence of output layer grad - * output: sequence of input layer grad - * - * \param outputs[0].matrix input layer grad, n * d - * \param outputs[0].vector start position sequence, n * 1 - * \param inputs[0].matrix output layer grad, n * (d * l) - * \param inputs[0].vector start positon sequence, n * 1 - */ -template -class ContextProjectionBackwardDataFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - context_length_ = config.get("context_length"); - context_start_ = config.get("context_start"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) - << "SequenceArg required here"; - const auto in_seq = dynamic_cast(inputs[0]); - const auto out_seq = dynamic_cast(outputs[0]); - - CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceId().data()); - CHECK_EQ(out_seq.shape().ndims(), 2UL); - CHECK_EQ(in_seq.shape().ndims(), 2UL); - CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL); - /// output layer grad dim == input layer grad dim * context_length_ - CHECK_EQ(in_seq.shape().ndims(), out_seq.shape().ndims() * context_length_); - /// input and output has the same batch_size - CHECK_EQ(in_seq.shape()[0], out_seq.shape()[0]); - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - - const auto out_grad_mat = in_seq.matrix(); - const auto seq_vec = in_seq.getSequenceId().vector(); - auto in_grad_mat = out_seq.matrix(); - - ContextProjectionBackwardData( - out_grad_mat, in_grad_mat, seq_vec, context_length_, context_start_); - } - - private: - size_t context_length_; - int context_start_; -}; - -/** - * Context Projection Backward Weight Function - * Update weight grad by backprop - * input: sequence of output layer grad - * output: weight grad - * - * \param outputs[0] weight grad, pad * d - * \param inputs[0].matrix output layer grad, n * (d * l) - * \param inputs[0].vecotr start positon sequence, n * 1 - */ -template -class ContextProjectionBackwardWeightFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - context_length_ = config.get("context_length"); - context_start_ = config.get("context_start"); - begin_pad_ = config.get("begin_pad"); - total_pad_ = config.get("total_pad"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK(inputs[0].isSequenceArg()) << "SequenceArg required here"; - const auto in_seq = dynamic_cast(inputs[0]); - CHECK(in_seq.data() && in_seq.getSequenceId().data() && outputs[0].data()); - CHECK_EQ(outputs[0].shape().ndims(), 2UL); - CHECK_EQ(in_seq.shape().ndims(), 2UL); - CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL); - CHECK_EQ(in_seq.shape()[0], outputs[0].shape()[0]); - /// output layer grad dim == weight dim * context_length_ - CHECK_EQ(in_seq.shape()[1], outputs[0].shape()[1] * context_length_); - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - - const auto seq_vec = in_seq.getSequenceId().vector(); - const auto out_grad_mat = in_seq.matrix(); - auto w_grad_mat = outputs[0].matrix(); - ContextProjectionBackwardWeight(out_grad_mat, - w_grad_mat, - seq_vec, - context_length_, - context_start_, - total_pad_, - begin_pad_); - } - - private: - size_t context_length_; - int context_start_; - size_t begin_pad_; - size_t total_pad_; -}; - -REGISTER_TYPED_FUNC(ContextProjectionForward, - CPU, - ContextProjectionForwardFunc); -REGISTER_TYPED_FUNC(ContextProjectionBackward, - CPU, - ContextProjectionBackwardFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(ContextProjectionForward, - GPU, - ContextProjectionForwardFunc); -REGISTER_TYPED_FUNC(ContextProjectionBackward, - GPU, - ContextProjectionBackwardFunc); -REGISTER_TYPED_FUNC(ContextProjectionBackwardData, - GPU, - ContextProjectionBackwardDataFunc); -REGISTER_TYPED_FUNC(ContextProjectionBackwardWeight, - GPU, - ContextProjectionBackwardWeightFunc); -#endif -} // namespace paddle diff --git a/paddle/function/ContextProjectionOpTest.cpp b/paddle/function/ContextProjectionOpTest.cpp deleted file mode 100644 index d805c3ae927321fc74946e202b98401b6b3cd0f7..0000000000000000000000000000000000000000 --- a/paddle/function/ContextProjectionOpTest.cpp +++ /dev/null @@ -1,114 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "FunctionTest.h" -#include "paddle/math/Matrix.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT - -void testMatrixProjectionForward(int context_start, - size_t context_length, - bool is_padding, - size_t batch_size, - size_t input_dim) { - size_t pad = std::max(0, -context_start) + - std::max(0, (int)(context_start + context_length - 1)); - if (pad == 0) is_padding = false; - - CpuGpuFuncCompare test( - "ContextProjectionForward", - FuncConfig() - .set("context_length", context_length) - .set("context_start", context_start) - .set("begin_pad", (size_t)std::max(0, -context_start))); - - // prepare input arguments - test.addSequence(SequenceIdArg(TensorShape{batch_size})); - test.addInputs( - SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim})); - if (is_padding) { // weight - test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim})); - } - test.addOutputs( - SequenceArg(VALUE_TYPE_FLOAT, - TensorShape{batch_size, input_dim * context_length}), - ADD_TO); - - // run Function - test.run(); -} - -void testMatrixProjectionBackward(int context_start, - size_t context_length, - bool is_padding, - size_t batch_size, - size_t input_dim) { - size_t pad = std::max(0, -context_start) + - std::max(0, (int)(context_start + context_length - 1)); - if (pad == 0) is_padding = false; - - CpuGpuFuncCompare test( - "ContextProjectionBackward", - FuncConfig() - .set("context_length", context_length) - .set("context_start", context_start) - .set("begin_pad", (size_t)std::max(0, -context_start)) - .set("is_padding", is_padding) - .set("total_pad", pad)); - - // prepare input arguments - test.addSequence(SequenceIdArg(TensorShape{batch_size})); - test.addInputs(SequenceArg( - VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim * context_length})); - test.addOutputs( - SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim}), - ADD_TO); - if (is_padding) { // weight - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim}), - ADD_TO); - } - - // run Function - test.run(); -} - -TEST(ContextProjection, Projection) { - for (auto context_start : {-5, -3, -1, 0, 3}) { - for (auto context_length : {1, 2, 5, 7}) { - for (auto trainable_padding : {false, true}) { - for (auto batch_size : {1, 2, 5, 20, 100}) { - for (auto input_dim : {15, 32, 63, 128, 200}) { - VLOG(3) << " context_start=" << context_start - << " context_length=" << context_length - << " trainable_padding=" << trainable_padding - << " batch_size=" << batch_size - << " input_dim=" << input_dim; - testMatrixProjectionForward(context_start, - context_length, - trainable_padding, - batch_size, - input_dim); - testMatrixProjectionBackward(context_start, - context_length, - trainable_padding, - batch_size, - input_dim); - } - } - } - } - } -} diff --git a/paddle/function/CosSimOp.cpp b/paddle/function/CosSimOp.cpp deleted file mode 100644 index 2c25e1af44965d30591faeccc9a181e36c7e0a0f..0000000000000000000000000000000000000000 --- a/paddle/function/CosSimOp.cpp +++ /dev/null @@ -1,240 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CosSimOp.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" - -namespace paddle { -/** - * Cosine Similarity for CpuMatrix - * - * \param out_mat, output value, size: nSamples * 1. - * \param in1_mat, input value 1, size: nSamples * dim. - * \param in2_mat, input value 2, size: n2 * dim (n2 == 1 or n2 == nSamples). - * \param scale, default 1.0 - * - */ -template <> -void CosSimForward(CpuMatrix& out_mat, - const CpuMatrix& in1_mat, - const CpuMatrix& in2_mat, - real scale) { - CHECK(out_mat.getData() && in1_mat.getData() && in2_mat.getData()); - size_t num_samples = out_mat.getHeight(); - size_t dim = in1_mat.getWidth(); - /// column vector [nSamples, 1] - real* out = out_mat.getData(); - const real* x = in1_mat.getData(); - const real* y = in2_mat.getData(); - - /// in2 might only have one row or full rows - CHECK(in2_mat.getHeight() == 1LU || in2_mat.getHeight() == num_samples); - size_t inc = (in2_mat.getHeight() == 1LU) ? 0 : dim; - for (size_t i = 0; i < num_samples; ++i, x += dim, y += inc) { - real square_sum_x = 0; - real square_sum_y = 0; - real xy = 0; - for (size_t j = 0; j < dim; ++j) { - square_sum_x += x[j] * x[j]; - square_sum_y += y[j] * y[j]; - xy += x[j] * y[j]; - } - CHECK(square_sum_x > 0 && square_sum_y > 0); - out[i] = scale * xy / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y)); - } -} - -/** - * Cosine Similarity - * for each row i, - * out[i] = scale * cos(input1[i], input2[i]) - * = scale * /sqrt(|input1[i]|^2 * |input2[i]|^2) - * when input2 only has one row, then for each row i, - * out[i] = cos(input1[i], input2[0]) - * - * \param inputs[0] input matrix 1, size: nSamples * dim. - * \param inputs[1] input matrix 2, size: n2 * dim (n2 == 1 or n2 == nSamples). - * \param outputs[0] output matrix, size : nSamples * 1. - */ - -template -class CosSimForwardFunc : public FunctionBase { - void init(const FuncConfig& config) override { - scale_ = config.get("scale"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(inputs.size(), 2UL); - CHECK_EQ(outputs.size(), 1UL); - - CHECK_EQ(inputs[0].shape().ndims(), 2UL); - CHECK_EQ(inputs[1].shape().ndims(), 2UL); - CHECK_EQ(outputs[0].shape().ndims(), 2UL); - - CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]); - CHECK_EQ(inputs[0].shape()[1], inputs[1].shape()[1]); - CHECK_EQ(outputs[0].shape()[1], 1UL); - - CHECK(outputs[0].data() && inputs[0].data() && inputs[1].data()); - - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - auto out_mat = outputs[0].matrix(); - const auto in1_mat = inputs[0].matrix(); - const auto in2_mat = inputs[1].matrix(); - - CosSimForward(out_mat, in1_mat, in2_mat, scale_); - } - - private: - real scale_; -}; - -/** - * Cosine Similarity Derivative for CpuMatrix - * - * \param in1_grad forward input grad 1, size: nSamples * dim. - * \param in2_grad forward input grad 2, - * size: n2 * dim (n2 == 1 or n2 == nSamples). - * - * \param out_grad backward loss output grad, size : nSamples * 1. - * \param out_val forward output value, size: nSamples * 1. - * \param in1_val forward input value 1, size: nSamples * dim. - * \param in2_val forward input value 2, - * size: n2 * dim (n2 == 1 or n2 == nSamples). - * \param scale, default 1.0 - */ -template <> -void CosSimBackward(const CpuMatrix& out_grad, - const CpuMatrix& out_val, - const CpuMatrix& in1_val, - const CpuMatrix& in2_val, - CpuMatrix& in1_grad, - CpuMatrix& in2_grad, - real scale) { - CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() && - in2_val.getData() && in1_grad.getData() && in2_grad.getData()); - CHECK_EQ(out_val.useGpu_, false) << "Matrix type are GPU, CPU required"; - - const real* grad = out_grad.getData(); - const real* out = out_val.getData(); - const real* prev_out_x = in1_val.getData(); - const real* prev_out_y = in2_val.getData(); - real* prev_grad_x = in1_grad.getData(); - real* prev_grad_y = in2_grad.getData(); - - size_t num_samples = out_grad.getHeight(); - size_t dim = in1_val.getWidth(); - CHECK_EQ(in2_val.getHeight(), in2_grad.getHeight()); - CHECK(in2_val.getHeight() == 1LU || in2_val.getHeight() == num_samples); - size_t inc = (in2_val.getHeight() == 1LU) ? 0 : dim; - for (size_t i = 0; i < num_samples; ++i, - prev_out_x += dim, - prev_out_y += inc, - prev_grad_x += dim, - prev_grad_y += inc) { - real square_sum_x = 0; - real square_sum_y = 0; - real xy = 0; - for (size_t j = 0; j < dim; ++j) { - square_sum_x += prev_out_x[j] * prev_out_x[j]; - square_sum_y += prev_out_y[j] * prev_out_y[j]; - xy += prev_out_x[j] * prev_out_y[j]; - } - CHECK(square_sum_x > 0 && square_sum_y > 0); - if (xy == 0) { - real reciprocal = - 1.0f / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y)); - for (size_t j = 0; j < dim; ++j) { - prev_grad_x[j] += scale * grad[i] * prev_out_y[j] * reciprocal; - prev_grad_y[j] += scale * grad[i] * prev_out_x[j] * reciprocal; - } - } else { - real reciprocal_xy = 1.0f / xy; - real reciprocal_square_sum_x = 1.0f / square_sum_x; - real reciprocal_square_sum_y = 1.0f / square_sum_y; - for (size_t j = 0; j < dim; ++j) { - prev_grad_x[j] += - out[i] * grad[i] * (prev_out_y[j] * reciprocal_xy - - prev_out_x[j] * reciprocal_square_sum_x); - prev_grad_y[j] += - out[i] * grad[i] * (prev_out_x[j] * reciprocal_xy - - prev_out_y[j] * reciprocal_square_sum_y); - } - } - } -} - -/** - * Cosine Similarity backward Derivative - * - * \param outputs[0] forward input grad 1, size: nSamples * dim. - * \param outputs[1] forward input grad 2, - * size: n2 * dim (n2 == 1 or n2 == nSamples). - * - * \param inputs[0] backward loss output grad, size : nSamples * 1. - * \param inputs[1] forward output value, size: nSamples * 1. - * \param inputs[2] forward input value 1, size: nSamples * dim. - * \param inputs[3] forward input value 2, - * size: n2 * dim (n2 == 1 or n2 == nSamples). - */ -template -class CosSimBackwardFunc : public FunctionBase { - void init(const FuncConfig& config) override { - scale_ = config.get("scale"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(inputs.size(), 4UL); - CHECK_EQ(outputs.size(), 2UL); - /// dim of out_grad and out_val == 1, column vector - CHECK_EQ(inputs[0].shape()[1], 1UL); - CHECK_EQ(inputs[1].shape()[1], 1UL); - /// nSamples of out_grad == out_val == in_val1 == in_grad1 - CHECK_EQ(inputs[1].shape()[0], inputs[0].shape()[0]); - CHECK_EQ(inputs[0].shape()[0], inputs[0].shape()[0]); - CHECK_EQ(outputs[0].shape()[0], inputs[0].shape()[0]); - /// dim of in1_val1 == in_val2 == in_grad1 == in_grad2 - CHECK_EQ(inputs[3].shape()[1], inputs[2].shape()[1]); - CHECK_EQ(outputs[0].shape()[1], inputs[2].shape()[1]); - CHECK_EQ(outputs[1].shape()[1], inputs[2].shape()[1]); - - CHECK(inputs[0].data() && inputs[1].data() && inputs[2].data() && - inputs[3].data() && outputs[0].data() && outputs[1].data()); - - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - CHECK_EQ(outputs[1].getArgType(), ADD_TO); - - const auto out_grad = inputs[0].matrix(); - const auto out_val = inputs[1].matrix(); - const auto in1_val = inputs[2].matrix(); - const auto in2_val = inputs[3].matrix(); - auto in1_grad = outputs[0].matrix(); - auto in2_grad = outputs[1].matrix(); - - CosSimBackward( - out_grad, out_val, in1_val, in2_val, in1_grad, in2_grad, scale_); - } - - private: - real scale_; -}; - -REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc); -REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc); -REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc); -#endif -} // namespace paddle diff --git a/paddle/function/CosSimOpTest.cpp b/paddle/function/CosSimOpTest.cpp deleted file mode 100644 index 42b02da0cb07a57e030a3edb08bea23203efd688..0000000000000000000000000000000000000000 --- a/paddle/function/CosSimOpTest.cpp +++ /dev/null @@ -1,64 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "FunctionTest.h" -#include "paddle/math/Matrix.h" - -using namespace paddle; // NOLINT - -void testCosSimForward(size_t height_x, - size_t height_y, - size_t width, - real scale) { - CpuGpuFuncCompare test("CosSimForward", FuncConfig().set("scale", scale)); - // prepare input arguments - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width})); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width})); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}), - ASSIGN_TO); - // run Function - test.run(); -} - -void testCosSimBackward(size_t height_x, - size_t height_y, - size_t width, - real scale) { - CpuGpuFuncCompare test("CosSimBackward", FuncConfig().set("scale", scale)); - // prepare input arguments - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1})); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1})); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width})); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width})); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}), - ADD_TO); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}), - ADD_TO); - // run Function - test.run(); -} - -TEST(Matrix, cosSim) { - for (auto height_x : {10, 100, 1000}) { - for (auto height_y : {1, height_x}) { - for (auto width : {10, 100, 1000}) { - for (auto scale : {1.0, 2.0}) { - testCosSimForward(height_x, height_y, width, scale); - testCosSimBackward(height_x, height_y, width, scale); - } - } - } - } -} diff --git a/paddle/function/CropOp.cpp b/paddle/function/CropOp.cpp deleted file mode 100644 index 5bd98910fe838751935f8ef2387ce96e755c6df1..0000000000000000000000000000000000000000 --- a/paddle/function/CropOp.cpp +++ /dev/null @@ -1,177 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CropOp.h" -#include "paddle/function/TensorShape.h" -#include "paddle/math/Vector.h" - -namespace paddle { - -template <> -void Crop(real* outputs, - const real* inputs, - const TensorShape inShape, - const TensorShape outShape, - const FuncConfig& conf) { - std::vector crop_corner = - conf.get>("crop_corner"); - int cCrop = crop_corner[1]; - int hCrop = crop_corner[2]; - int wCrop = crop_corner[3]; - - int num = inShape[0]; - int inC = inShape[1]; - int inH = inShape[2]; - int inW = inShape[3]; - - int outC = outShape[1]; - int outH = outShape[2]; - int outW = outShape[3]; - - for (int n = 0; n < num; n++) { - for (int c = 0; c < outC; c++) { - for (int h = 0; h < outH; h++) { - int outoff = ((n * outC + c) * outH + h) * outW; - int inoff = ((n * inC + c + cCrop) * inH + h + hCrop) * inW + wCrop; - memcpy(outputs + outoff, inputs + inoff, outW * sizeof(real)); - } - } - } -} - -template <> -void CropGrad(const real* inGrad, - real* outGrad, - const TensorShape inShape, - const TensorShape outShape, - const FuncConfig& conf) { - std::vector crop_corner = - conf.get>("crop_corner"); - int cCrop = crop_corner[1]; - int hCrop = crop_corner[2]; - int wCrop = crop_corner[3]; - - int num = outShape[0]; - int outC = outShape[1]; - int outH = outShape[2]; - int outW = outShape[3]; - - int inC = inShape[1]; - int inH = inShape[2]; - int inW = inShape[3]; - - for (int n = 0; n < num; n++) { - for (int c = 0; c < inC; c++) { - for (int h = 0; h < inH; h++) { - int outoff = ((n * outC + c + cCrop) * outH + h + hCrop) * outW + wCrop; - int inoff = ((n * inC + c) * inH + h) * inW; - CpuVector inG = CpuVector(inW, const_cast(inGrad + inoff)); - CpuVector outG = CpuVector(inW, outGrad + outoff); - outG += inG; - } - } - } -} - -/** - * \brief Crop input according to the specify corner and shape. - * The input and output is a 4D tensor. In CropFunc, we only - * crop the 2nd to 4th dimension. - * - * Argument in this Function: - * \param pad_ A struct object contains the cropping corner and shape. - * \param inputs A 4D tensor, only one input. - * \param outputs A 4D tensor, the output value after cropping. - * - * For example, - * Input(2,2,2,3) = [ - * [ [[1,2,3], [3,4,5]], - * [[2,3,5], [1,6,7]] ], - * [ [[4,3,1], [1,8,7]], - * [[3,8,9], [2,3,5]] ] - * ] # the input shape is (2,2,2,3) - * - * pad_: if corner = (0,1,1) and crop_shape = (2,1,2) - * Output(2,2,1,2) = [ - * [ [[4,5]], - * [[6,7]] ], - * [ [[8,7]], - * [[3,5]] ] - * ] # the input shape is (2,2,2,3) - */ -template -class CropFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { conf_ = config; } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - - TensorShape inShape = inputs[0].shape(); - TensorShape outShape = outputs[0].shape(); - - Crop(outputs[0].data(), - inputs[0].data(), - inShape, - outShape, - conf_); - } - - private: - FuncConfig conf_; -}; - -/** - * \brief The backward propagation of cropping Function. - * - * Argument in this Function: - * \param crop_ The same meaning as it in CropFunc. - * \param inputs The gradient with respect to the output value of CropFunc. - * \param outputs The gradient with respect to the input value of CropFunc. - */ - -template -class CropGradFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { conf_ = config; } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - - TensorShape outShape = outputs[0].shape(); - TensorShape inShape = inputs[0].shape(); - - CropGrad(inputs[0].data(), - outputs[0].data(), - inShape, - outShape, - conf_); - } - - private: - FuncConfig conf_; -}; - -REGISTER_TYPED_FUNC(Crop, CPU, CropFunc); -REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(Crop, GPU, CropFunc); -REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc); -#endif - -} // namespace paddle diff --git a/paddle/function/CrossMapNormalOp.cpp b/paddle/function/CrossMapNormalOp.cpp deleted file mode 100644 index 7ff9227e5c2702d9d5334db501730b57ec10bfe3..0000000000000000000000000000000000000000 --- a/paddle/function/CrossMapNormalOp.cpp +++ /dev/null @@ -1,344 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CrossMapNormalOp.h" -#include "paddle/math/Vector.h" - -namespace paddle { - -template <> -void CrossMapNormal(real* outputs, - real* denoms, - const real* inputs, - size_t numSamples, - size_t channels, - size_t height, - size_t width, - size_t size, - real scale, - real pow) { - size_t oneImage = height * width; - size_t oneSample = channels * oneImage; - - CpuVector outputsV(numSamples * oneSample, outputs); - CpuVector inputsV(numSamples * oneSample, const_cast(inputs)); - CpuVector denomsV(numSamples * oneSample, denoms); - - // f(x) = x * ( 1 + scale * SUM((x)^2) )^(-pow) - // x represents inputs - // f(x) represents outputs - // denoms save the intermediate result for backward - denomsV = denomsV.constant(1.0); - const int start = -((int)size - 1) / 2; - const int end = (int)size + start; - for (size_t i = 0; i < numSamples; i++) { - real* oneDenom = denoms + i * oneSample; - real* oneInput = const_cast(inputs) + i * oneSample; - for (int c = 0; c < (int)channels; c++) { - CpuVector denom(oneImage, oneDenom + c * oneImage); - for (int s = start; s < end; s++) { - if (c + s >= 0 && c + s < (int)channels) { - CpuVector input(oneImage, oneInput + (c + s) * oneImage); - denom += input.square() * scale; - } - } - } - } - - outputsV = inputsV * denomsV.pow(-pow); -} - -template <> -void CrossMapNormalGrad(real* inputsGrad, - const real* inputsValue, - const real* outputsValue, - const real* outputsGrad, - const real* denoms, - size_t numSamples, - size_t channels, - size_t height, - size_t width, - size_t size, - real scale, - real pow) { - size_t oneSample = channels * height * width; - std::function oneImage = [=](real* data, - size_t offset) { - return CpuVector(height * width, data + offset); - }; - - const int start = -((int)size) / 2; - const int end = (int)size + start; - const real ratio = -(real)2 * scale * pow; - for (size_t i = 0; i < numSamples; i++) { - size_t sOffset = i * oneSample; - real* oneInputGrad = inputsGrad + sOffset; - real* oneInputValue = const_cast(inputsValue) + sOffset; - real* oneDenom = const_cast(denoms) + sOffset; - real* oneOutputGrad = const_cast(outputsGrad) + sOffset; - real* oneOutputValue = const_cast(outputsValue) + sOffset; - - for (int c = 0; c < (int)channels; c++) { - size_t cOffset = c * height * width; - CpuVector inputGrad = oneImage(oneInputGrad, cOffset); - CpuVector inputValue = oneImage(oneInputValue, cOffset); - CpuVector denom = oneImage(oneDenom, cOffset); - CpuVector outputGrad = oneImage(oneOutputGrad, cOffset); - - inputGrad = inputGrad + denom.pow(-pow) * outputGrad; - for (int s = start; s < end; s++) { - if (c + s >= 0 && c + s < (int)channels) { - size_t offset = (c + s) * height * width; - CpuVector output = oneImage(oneOutputValue, offset); - CpuVector outputGrad = oneImage(oneOutputGrad, offset); - CpuVector denom = oneImage(oneDenom, offset); - - inputGrad += ((outputGrad * output * ratio) / denom) * inputValue; - } - } - } - } -} - -/** - * \brief Normalization with across maps. - * - * This Function comes from the paper - * "ImageNet Classification with Deep Convolutional Neural Networks". - * - * The original formula is: - * - * Input(i, x, y) - * Output(i, x, y) = ---------------------------------------------- - * -- upper - * (k + alpha * > (Input(j, x, y))^2) ^ (beta) - * -- j = lower - * - * upper is `min(C, c + N/2)` - * lower if `max(0, c - N/2)` - * - * Function implementation: - * - * inputs and outpus is NCHW format, while input.shape.ndims() is equal 4. - * And the meaning of each dimension(0-3) is respectively batch size, - * feature maps, rows and columns. - * - * Input and Output in the above formula is for each map(i) of one image, and - * Input(i, x, y), Output(i, x, y) represents an element in an image. - * - * C is the number of feature maps of one image, and N is a hyper-parameters - * is configured when Function is initialized. The sum in the denominator - * is the sum of the same position in the neighboring maps. - * - * In the implementation of Function, k is equal to 1, - * so Function has no argument for k. - * - * Function Arguments: - * - * \param size_ represent N - * \param scale_ represent alpha - * \param pow_ represent beta - * \param inputs[0] represent Input - * \param outputs[0] represent Output - * \param outputs[1] represent The denominator in the formula(except beta) - * - * Note: - * Save output[1] is to simplify the backward calculation. - * TODO, if only consider the forward calculation, we can optimize to - * remove the output[1]. - */ -template -class CrossMapNormalFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - // function arguments - size_ = config.get("size"); - scale_ = config.get("scale"); - pow_ = config.get("pow"); - - // number of inputs and outputs - numInputs_ = 1; - numOutputs_ = 2; - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - check(inputs, outputs); - // ArgType check still on here, - // not sure whether it is better to put inside the check. - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - CHECK_EQ(outputs[1].getArgType(), ASSIGN_TO); - size_t batchSize = inputs[0].shape()[0]; - size_t maps = inputs[0].shape()[1]; - size_t rows = inputs[0].shape()[2]; - size_t columns = inputs[0].shape()[3]; - - CrossMapNormal(outputs[0].data(), - outputs[1].data(), - inputs[0].data(), - batchSize, - maps, - rows, - columns, - size_, - scale_, - pow_); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - - CHECK_EQ(inputs[0].shape().ndims(), (size_t)4); - CHECK(inputs[0].shape() == outputs[0].shape()); - CHECK(inputs[0].shape() == outputs[1].shape()); - } - - // Only need the shape of the input, can calculate the - // floating-point operation. - size_t ops(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ((size_t)numInputs_, inputs.size()); - size_t batchSize = inputs[0].shape()[0]; - size_t maps = inputs[0].shape()[1]; - size_t rows = inputs[0].shape()[2]; - size_t columns = inputs[0].shape()[3]; - - // number of floating-point operations - // an approximate value - size_t ops = batchSize * maps * rows * columns * (size_ * 2 + 3); - - return ops; - } - - private: - size_t size_; - real scale_; - real pow_; -}; - -/** - * \brief Backward calculation for normalization with across maps. - * - * Function implementation: - * - * The implementation of this Function is derived from the - * CrossMapNormalFunc implementation. - * - * InputGrad = OutputGrad * denoms ^ (-beta) - * -- upper - * + > (OutputGrad * OutputValue * (-2 * alpha * beta) / denoms) * InputValue - * -- lower - * - * The data of inputs/outputs format is the same as the forward interface - * and is NCHW. - * - * The upper and lower is the same as forward. The logic of the sum - * is also the same as forward. - * - * Function Arguments: - * - * \param size_ represent N - * \param scale_ represent alpha - * \param pow_ represent beta - * \param inputs[0] represent InputValue, inputs[0] of CrossMapNormalFunc - * \param inputs[1] represent OutputValue, outputs[0] of CrossMapNormalFunc - * \param inputs[2] represent OutputGrad - * \param inputs[3] represent denoms, outputs[1] of CrossMapNormalFunc - * This is the intermediate result that is - * preserved in the forward calculation. - * \param outputs[0] represent InputGrad - */ -template -class CrossMapNormalGradFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - // function arguments - size_ = config.get("size"); - scale_ = config.get("scale"); - pow_ = config.get("pow"); - - // number of inputs and outputs - numInputs_ = 4; - numOutputs_ = 1; - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - check(inputs, outputs); - if (outputs[0].getArgType() != ADD_TO) { - // Currently, some algorithm implementations are ASSIGN_TO mode, - // if need to support the ADD_TO calculation, need to clear the output. - typename Tensor::Vector tmp( - outputs[0].shape().getElements(), outputs[0].data()); - tmp.zero(); - } - - size_t batchSize = inputs[0].shape()[0]; - size_t maps = inputs[0].shape()[1]; - size_t rows = inputs[0].shape()[2]; - size_t columns = inputs[0].shape()[3]; - - CrossMapNormalGrad(outputs[0].data(), - inputs[0].data(), - inputs[1].data(), - inputs[2].data(), - inputs[3].data(), - batchSize, - maps, - rows, - columns, - size_, - scale_, - pow_); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - - CHECK_EQ(inputs[0].shape().ndims(), (size_t)4); - CHECK(inputs[0].shape() == inputs[1].shape()); - CHECK(inputs[0].shape() == inputs[2].shape()); - CHECK(inputs[0].shape() == inputs[3].shape()); - CHECK(inputs[0].shape() == outputs[0].shape()); - } - - // Only need the shape of one input, can calculate the - // floating-point operation. - size_t ops(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_LT((size_t)1, inputs.size()); - size_t batchSize = inputs[0].shape()[0]; - size_t maps = inputs[0].shape()[1]; - size_t rows = inputs[0].shape()[2]; - size_t columns = inputs[0].shape()[3]; - - // number of floating-point operations - // an approximate value - size_t ops = batchSize * maps * rows * columns * (size_ * 4 + 2); - - return ops; - } - - private: - size_t size_; - real scale_; - real pow_; -}; - -REGISTER_TYPED_FUNC(CrossMapNormal, CPU, CrossMapNormalFunc); -REGISTER_TYPED_FUNC(CrossMapNormalGrad, CPU, CrossMapNormalGradFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(CrossMapNormal, GPU, CrossMapNormalFunc); -REGISTER_TYPED_FUNC(CrossMapNormalGrad, GPU, CrossMapNormalGradFunc); -#endif - -} // namespace paddle diff --git a/paddle/function/DepthwiseConvOpGpu.cu b/paddle/function/DepthwiseConvOpGpu.cu deleted file mode 100644 index 2c0e71b19b22abac25d273d8bbeddc330e67f8b0..0000000000000000000000000000000000000000 --- a/paddle/function/DepthwiseConvOpGpu.cu +++ /dev/null @@ -1,376 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "DepthwiseConvOp.h" -#include "paddle/math/BaseMatrix.h" - -namespace paddle { - -// CUDA kernel to compute the depthwise convolution forward pass -template -__global__ void ConvolutionDepthwiseForward(const int nthreads, - const T* const inputData, - const T* const filterData, - const int batchSize, - const int outputChannels, - const int outputHeight, - const int outputWidth, - const int inputChannels, - const int inputHeight, - const int inputWidth, - const int filterMultiplier, - const int filterHeight, - const int filterWidth, - const int strideH, - const int strideW, - const int paddingH, - const int paddingW, - T* const outputData) { - int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; - - if (index < nthreads) { - const int batch = index / outputChannels / outputHeight / outputWidth; - const int c_out = (index / outputHeight / outputWidth) % outputChannels; - const int h_out = (index / outputWidth) % outputHeight; - const int w_out = index % outputWidth; - - const int c_in = c_out / filterMultiplier; - const T* weight = filterData + c_out * filterHeight * filterWidth; - T value = 0; - const int h_in_start = -paddingH + h_out * strideH; - const int w_in_start = -paddingW + w_out * strideW; - const int h_in_end = -paddingH + h_out * strideH + filterHeight - 1; - const int w_in_end = -paddingW + w_out * strideW + filterWidth - 1; - if ((h_in_start >= 0) && (h_in_end < inputHeight) && (w_in_start >= 0) && - (w_in_end < inputWidth)) { - for (int kh = 0; kh < filterHeight; ++kh) { - for (int kw = 0; kw < filterWidth; ++kw) { - const int h_in = -paddingH + h_out * strideH + kh; - const int w_in = -paddingW + w_out * strideW + kw; - const int offset = - ((batch * inputChannels + c_in) * inputHeight + h_in) * - inputWidth + - w_in; - value += (*weight) * inputData[offset]; - ++weight; - } - } - } else { - for (int kh = 0; kh < filterHeight; ++kh) { - for (int kw = 0; kw < filterWidth; ++kw) { - const int h_in = -paddingH + h_out * strideH + kh; - const int w_in = -paddingW + w_out * strideW + kw; - if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) && - (w_in < inputWidth)) { - const int offset = - ((batch * inputChannels + c_in) * inputHeight + h_in) * - inputWidth + - w_in; - value += (*weight) * inputData[offset]; - } - ++weight; - } - } - } - outputData[index] = value; - } -} - -// CUDA kernel to compute the depthwise convolution backprop w.r.t input. -template -__global__ void ConvolutionDepthwiseInputBackward(const int nthreads, - const T* const top_diff, - const T* const weight_data, - const int num, - const int outputChannels, - const int outputHeight, - const int outputWidth, - const int inputChannels, - const int inputHeight, - const int inputWidth, - const int filterMultiplier, - const int filterHeight, - const int filterWidth, - const int strideH, - const int strideW, - const int paddingH, - const int paddingW, - T* const bottom_diff) { - int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; - if (index < nthreads) { - const int batch = index / inputChannels / inputHeight / inputWidth; - const int c_in = (index / inputHeight / inputWidth) % inputChannels; - const int h_in = (index / inputWidth) % inputHeight; - const int w_in = index % inputWidth; - - const int c_out_start = c_in * filterMultiplier; - - int h_out_start = (h_in - filterHeight + paddingH + strideH) / strideH; - h_out_start = 0 > h_out_start ? 0 : h_out_start; - int h_out_end = (h_in + paddingH) / strideH; - h_out_end = outputHeight - 1 < h_out_end ? outputHeight - 1 : h_out_end; - int w_out_start = (w_in - filterWidth + paddingW + strideW) / strideW; - w_out_start = 0 > w_out_start ? 0 : w_out_start; - int w_out_end = (w_in + paddingW) / strideW; - w_out_end = outputWidth - 1 < w_out_end ? outputWidth - 1 : w_out_end; - - T value = 0; - - for (int c_out = c_out_start; c_out < c_out_start + filterMultiplier; - c_out++) { - for (int h_out = h_out_start; h_out <= h_out_end; ++h_out) { - const int filter_h = h_in + paddingH - h_out * strideH; - for (int w_out = w_out_start; w_out <= w_out_end; ++w_out) { - const int filter_w = w_in + paddingW - w_out * strideW; - const int filter_offset = c_out * filterHeight * filterWidth + - filter_h * filterWidth + filter_w; - const int top_diff_offset = - ((batch * outputChannels + c_out) * outputHeight + h_out) * - outputWidth + - w_out; - value += top_diff[top_diff_offset] * weight_data[filter_offset]; - } - } - } - bottom_diff[index] += value; - } -} - -// CUDA kernel to compute the depthwise convolution backprop w.r.t filter. -template -__global__ void ConvolutionDepthwiseFilterBackward(const int num_i, - const int nthreads, - const T* const top_diff, - const T* const inputData, - const int num, - const int outputChannels, - const int outputHeight, - const int outputWidth, - const int inputChannels, - const int inputHeight, - const int inputWidth, - const int filterMultiplier, - const int filterHeight, - const int filterWidth, - const int strideH, - const int strideW, - const int paddingH, - const int paddingW, - T* const buffer_data) { - int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; - if (index < nthreads) { - const int h_out = (index / outputWidth) % outputHeight; - const int w_out = index % outputWidth; - const int kh = - (index / filterWidth / outputHeight / outputWidth) % filterHeight; - const int kw = (index / outputHeight / outputWidth) % filterWidth; - const int h_in = -paddingH + h_out * strideH + kh; - const int w_in = -paddingW + w_out * strideW + kw; - if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) && - (w_in < inputWidth)) { - const int c_out = - index / (filterHeight * filterWidth * outputHeight * outputWidth); - const int c_in = c_out / filterMultiplier; - const int batch = num_i; - const int top_offset = - ((batch * outputChannels + c_out) * outputHeight + h_out) * - outputWidth + - w_out; - const int bottom_offset = - ((batch * inputChannels + c_in) * inputHeight + h_in) * inputWidth + - w_in; - buffer_data[index] = top_diff[top_offset] * inputData[bottom_offset]; - } else { - buffer_data[index] = 0; - } - } -} - -template -class DepthwiseConvFunctor { - public: - void operator()(const T* inputData, - const T* filterData, - int batchSize, - int outputChannels, - int outputHeight, - int outputWidth, - int inputChannels, - int inputHeight, - int inputWidth, - int filterMultiplier, - int filterHeight, - int filterWidth, - int strideH, - int strideW, - int paddingH, - int paddingW, - T* outputData) { - int outputSize = batchSize * outputChannels * outputHeight * outputWidth; - - size_t blocks = (outputSize + 1024 - 1) / 1024; - size_t blockX = 512; - size_t blockY = (blocks + 512 - 1) / 512; - dim3 threads(1024, 1); - dim3 grid(blockX, blockY); - - ConvolutionDepthwiseForward<<>>( - outputSize, - inputData, - filterData, - batchSize, - outputChannels, - outputHeight, - outputWidth, - inputChannels, - inputHeight, - inputWidth, - filterMultiplier, - filterHeight, - filterWidth, - strideH, - strideW, - paddingH, - paddingW, - outputData); - } -}; - -template -class DepthwiseConvGradInputFunctor { - public: - void operator()(const T* outputGrad, - const T* filterData, - int batchSize, - int outputChannels, - int outputHeight, - int outputWidth, - int inputChannels, - int inputHeight, - int inputWidth, - int filterMultiplier, - int filterHeight, - int filterWidth, - int strideH, - int strideW, - int paddingH, - int paddingW, - T* inputGrad) { - int inputSize = batchSize * inputChannels * inputHeight * inputWidth; - - size_t blocks = (inputSize + 1024 - 1) / 1024; - size_t blockX = 512; - size_t blockY = (blocks + 512 - 1) / 512; - dim3 threads(1024, 1); - dim3 grid(blockX, blockY); - - ConvolutionDepthwiseInputBackward - // NOLINT_NEXT_LINE(whitespace/operators) - <<>>(inputSize, - outputGrad, - filterData, - batchSize, - outputChannels, - outputHeight, - outputWidth, - inputChannels, - inputHeight, - inputWidth, - filterMultiplier, - filterHeight, - filterWidth, - strideH, - strideW, - paddingH, - paddingW, - inputGrad); - } -}; - -template -class DepthwiseConvGradFilterFunctor { - public: - void operator()(const T* outputGrad, - const T* inputData, - int batchSize, - int outputChannels, - int outputHeight, - int outputWidth, - int inputChannels, - int inputHeight, - int inputWidth, - int filterMultiplier, - int filterHeight, - int filterWidth, - int strideH, - int strideW, - int paddingH, - int paddingW, - T* colData, - T* filterGrad) { - int colDataSize = outputChannels * filterHeight * filterWidth * - outputHeight * outputWidth; - - size_t blocks = (colDataSize + 1024 - 1) / 1024; - size_t blockX = 512; - size_t blockY = (blocks + 512 - 1) / 512; - dim3 threads(1024, 1); - dim3 grid(blockX, blockY); - BaseMatrix filterGradMatrix(outputChannels * filterHeight * filterWidth, - 1, - filterGrad, - false, - true); - - for (int i = 0; i < batchSize; i++) { - ConvolutionDepthwiseFilterBackward< - T><<>>(i, - colDataSize, - outputGrad, - inputData, - batchSize, - outputChannels, - outputHeight, - outputWidth, - inputChannels, - inputHeight, - inputWidth, - filterMultiplier, - filterHeight, - filterWidth, - strideH, - strideW, - paddingH, - paddingW, - colData); - int K = outputHeight * outputWidth; - int M = colDataSize / K; - - BaseMatrix colMatrix(M, K, colData, false, true); - filterGradMatrix.sumRows(colMatrix, (T)1.0, (T)1.0); - } - } -}; - -#ifdef PADDLE_TYPE_DOUBLE -template class DepthwiseConvGradInputFunctor; -template class DepthwiseConvFunctor; -template class DepthwiseConvGradFilterFunctor; -#else -template class DepthwiseConvGradInputFunctor; -template class DepthwiseConvFunctor; -template class DepthwiseConvGradFilterFunctor; -#endif - -} // namespace paddle diff --git a/paddle/function/EigenGemm.cpp b/paddle/function/EigenGemm.cpp deleted file mode 100644 index 8e9dbbd7a154095a7298bb2f59a82d13a60f9bd3..0000000000000000000000000000000000000000 --- a/paddle/function/EigenGemm.cpp +++ /dev/null @@ -1,102 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/function/EigenThreadDevice.h" - -namespace paddle { - -template -struct EigenBlasGemm { - typedef Eigen::TensorMap, - Eigen::Aligned> - EigenMatrix; - - static void compute(const bool transA, - const bool transB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc) { - Eigen::array sizeA; - if (transA) { - sizeA[0] = K; - sizeA[1] = M; - CHECK_EQ(M, lda); - } else { - sizeA[0] = M; - sizeA[1] = K; - CHECK_EQ(K, lda); - } - Eigen::array sizeB; - if (transB) { - sizeB[0] = N; - sizeB[1] = K; - CHECK_EQ(K, ldb); - } else { - sizeB[0] = K; - sizeB[1] = N; - CHECK_EQ(N, ldb); - } - Eigen::array sizeC = {{M, ldc}}; - Eigen::array offsetC = {{0, 0}}; - Eigen::array extentC = {{M, N}}; - - const EigenMatrix a(const_cast(A), sizeA); - const EigenMatrix b(const_cast(B), sizeB); - EigenMatrix c(C, sizeC); - - typedef typename Eigen::Tensor::DimensionPair DimPair; - Eigen::array dims; - dims[0] = DimPair(1, 0); - dims[0].first = transA ? 0 : 1; - dims[0].second = transB ? 1 : 0; - - auto* device = EigenDeviceWarpper::device(); - if (N == ldc) { - if (alpha == T(1) && beta == T(0)) { - c.device(*device) = a.contract(b, dims); - } else if (alpha == T(1) && beta == T(1)) { - c.device(*device) += a.contract(b, dims); - } else { - c.device(*device) = alpha * a.contract(b, dims) + beta * c; - } - } else { - if (alpha == T(1) && beta == T(0)) { - c.slice(offsetC, extentC).device(*device) = a.contract(b, dims); - } else if (alpha == T(1) && beta == T(1)) { - c.slice(offsetC, extentC).device(*device) += a.contract(b, dims); - } else { - c.slice(offsetC, extentC).device(*device) = - alpha * a.contract(b, dims) + beta * c.slice(offsetC, extentC); - } - } - EigenDeviceWarpper::free_device(device); - } -}; - -#ifdef PADDLE_TYPE_DOUBLE -template struct EigenBlasGemm; -#else -template struct EigenBlasGemm; -#endif - -} // namespace paddle diff --git a/paddle/function/Function.h b/paddle/function/Function.h deleted file mode 100644 index a6c14ef29b760faa393c37bd2357824a061c7b38..0000000000000000000000000000000000000000 --- a/paddle/function/Function.h +++ /dev/null @@ -1,214 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "BufferArg.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Any.h" -#include "paddle/utils/ClassRegistrar.h" -#include "paddle/utils/Error.h" - -namespace paddle { - -/** - * Function Configuration. - * The argument type of Function::init. - */ -class FuncConfig { - public: - template - T get(const std::string& key, Error* err = nullptr) const { - try { - return any_cast(valueMap_.at(key)); - } catch (std::exception& e) { // could be cast or out of range exception. - if (err) { - *err = Error(e.what()); - } else { - LOG(FATAL) << "Cannot get key " << key << " with error " << e.what(); - } - return T(); - } - } - - template - FuncConfig& set(const std::string& key, T v, Error* err = nullptr) { - auto it = valueMap_.find(key); - if (it != valueMap_.end()) { // already contains key. - if (err) { - *err = Error("Key %s is already set in FuncConfig", key.c_str()); - } else { - LOG(FATAL) << "Key " << key << " is already set in FuncConfig."; - } - return *this; - } - valueMap_[key] = any(v); - return *this; - } - - protected: - mutable std::unordered_map valueMap_; -}; - -/** - * Argument type for Function::calc(). - * A BufferArgs contains a set of BufferArg, - * because Function can have multiple inputs and outputs. - * - * addArg() with Matix object used to adapt Layer Argument. - * Will create a BufferArg object in addArg(), - * and free in destructor of BufferArgs. - * - * addArg() with BufferArg object, just save BufferArg object address, - * and the caller needs to guarantee the validity of the BufferArg object - * in the BufferArgs life time. - */ -class BufferArgs { - public: - BufferArgs() {} - - ~BufferArgs() { - for (auto arg : _args_) { - delete arg; - } - } - - size_t size() const { return args_.size(); } - - // add argument into BufferArgs - // Tensor can be Matrix, Vector, IVector. - // For inputs, do not need argType. - // For outputs, the argType needs to be specified as ASSIGN_TO or ADD_TO. - void addArg(const Matrix& arg, ArgType argType = UNSPECIFIED) { - _args_.push_back(new BufferArg(arg, argType)); - addArg(*_args_.back()); - } - - void addArg(const Vector& arg, ArgType argType = UNSPECIFIED) { - _args_.push_back(new BufferArg(arg, argType)); - addArg(*_args_.back()); - } - - void addArg(const IVector& arg, ArgType argType = UNSPECIFIED) { - _args_.push_back(new BufferArg(arg, argType)); - addArg(*_args_.back()); - } - - // Add arg into BufferArgs and reshape the arg. - // - // For example, arg represents an image buffer, - // but Matrix can only represent a two-dimensional Tensor. - // So need an extra argument to describe the shape of the image buffer. - void addArg(const Matrix& arg, - const TensorShape& shape, - ArgType argType = UNSPECIFIED); - - void addArg(const CpuSparseMatrix& arg, ArgType argType = UNSPECIFIED); - void addArg(const GpuSparseMatrix& arg, ArgType argType = UNSPECIFIED); - - void addArg(const Matrix& matrix, - const IVector& vector, - ArgType argType = UNSPECIFIED); - - // get argument - const BufferArg& operator[](size_t num) const { - CHECK_LT(num, args_.size()); - return *args_[num]; - } - - void addArg(BufferArg& arg) { args_.push_back(&arg); } - - void addArg(SequenceIdArg& arg) { args_.push_back(&arg); } - - void addArg(SequenceArg& arg) { args_.push_back(&arg); } - - void addArg(SparseMatrixArg& arg) { args_.push_back(&arg); } - - private: - std::vector args_; - // The BufferArg object is constructed and freed by BufferArgs. - std::vector _args_; -}; - -/** - * \brief Base class for Function. - * The basic Function implementation requires override init and calc interfaces. - * - * The caller needs to ensure the validity of the arguments - * during Function execution. - * - * Function inputs are readonly, Function outputs have two modes: ASSIGN_TO - * and ADD_TO. - * If output.getArgType() == ASSIGN_TO, this is assign mode, and the calculation - * result of Function assigned to the output BufferArg. - * If output.getArgType() == ADD_TO, this is add mode, and the calculation - * result of Function need added to the output BufferArg. - * - * For example: - * ASSIGN_TO: output = Function(inputs) - * ADD_TO: output += Function(inputs) - * If Function has more than one output, each output can have different modes. - */ -class FunctionBase { - public: - virtual ~FunctionBase() {} - - virtual void init(const FuncConfig& config) {} - - virtual void calc(const BufferArgs& inputs, const BufferArgs& outputs) {} - - // This member function is used to check whether the BufferType and shape of - // the inputs and outputs arguments of the Function are correct. - // General calc function which will call this check to do arguments check. - // And before the calc called, the caller can also check their own arguments. - virtual void check(const BufferArgs& inputs, const BufferArgs& outputs) {} - - // Calculate the number of floating-point operations of this Function. - // The inputs and outputs arguments do not need to contain the actual data, - // only the shape. - // And some Functions have the same input and output shapes, - // so you may not need to enter the complete number of arguments. - // But entering the full arguments is always correct for this interface. - virtual size_t ops(const BufferArgs& inputs, const BufferArgs& outputs) { - return 0; - } - - int getNumInputs() const { return numInputs_; } - - int getNumOutputs() const { return numOutputs_; } - - static ClassRegistrar funcRegistrar_; - - protected: - // numInputs_ and numOutputs_ represents the maximum - // input and output supported by Function. - // Some functions are optimized for input and output, - // so when comparing the number of arguments, for these functions - // inputs.size() <= numInputs_ or outputs.size() <= numOutputs_ - size_t numInputs_; - size_t numOutputs_; -}; - -#define FUNC_NAME(typeName, deviceName) #typeName "-" #deviceName - -#define REGISTER_TYPED_FUNC(typeName, deviceName, className) \ - static InitFunction __reg_type_##typeName##deviceName([]() { \ - FunctionBase::funcRegistrar_ \ - .registerClass>( \ - FUNC_NAME(typeName, deviceName)); \ - }) - -} // namespace paddle diff --git a/paddle/function/FunctionTest.cpp b/paddle/function/FunctionTest.cpp deleted file mode 100644 index f5e6ca3f515a7fcd1498979703a0a59ddca40742..0000000000000000000000000000000000000000 --- a/paddle/function/FunctionTest.cpp +++ /dev/null @@ -1,166 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Function.h" -#include -#include "paddle/math/SparseMatrix.h" - -namespace paddle { - -template -void FunctionApi(typename Tensor::Matrix& output, - const typename Tensor::Matrix& input); - -template <> -void FunctionApi(CpuMatrix& output, const CpuMatrix& input) { - EXPECT_EQ(output.getHeight(), 100U); - EXPECT_EQ(output.getWidth(), 200U); -} - -template <> -void FunctionApi(GpuMatrix& output, const GpuMatrix& input) { - EXPECT_EQ(output.getHeight(), 10U); - EXPECT_EQ(output.getWidth(), 20U); -} - -template -void Function(const BufferArgs& arguments) { - const auto input = arguments[0].matrix(); - auto output = arguments[1].matrix(); - FunctionApi(output, input); -} - -TEST(Function, BufferArgs) { - CpuMatrix cpuInput = CpuMatrix(100, 200); - CpuMatrix cpuOutput = CpuMatrix(100, 200); - BufferArgs cpuArgments; - cpuArgments.addArg(cpuInput); - cpuArgments.addArg(cpuOutput); - Function(cpuArgments); - - GpuMatrix gpuInput = GpuMatrix(10, 20); - GpuMatrix gpuOutput = GpuMatrix(10, 20); - BufferArgs gpuArgments; - gpuArgments.addArg(gpuInput); - gpuArgments.addArg(gpuOutput); - Function(gpuArgments); -} - -/** - * Some tests case are used to check the consistency between the BufferArg type - * argument received by Function and the original type argument. - * - * Use Case: - * TEST() { - * Matrix matrix(...); - * CheckBufferArg lambda = [=](const BufferArg& arg) { - * // check matrix and arg are equivalent - * EXPECT_EQ(matrix, arg); - * } - * - * BufferArgs argments{matrix...}; - * std::vector checkFunc{lambda...}; - * testBufferArgs(argments, checkFunc); - * } - */ -typedef std::function CheckBufferArg; - -void testBufferArgs(const BufferArgs& inputs, - const std::vector& check) { - EXPECT_EQ(inputs.size(), check.size()); - for (size_t i = 0; i < inputs.size(); i++) { - check[i](inputs[i]); - } -} - -void testBufferArgs(const BufferArgs& inputs, const CheckBufferArg& check) { - EXPECT_EQ(inputs.size(), 1U); - check(inputs[0]); -} - -TEST(Arguments, Matrix) { - MatrixPtr matrix = Matrix::create(100, 200); - CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 2U); - EXPECT_EQ(arg.shape()[0], 100U); - EXPECT_EQ(arg.shape()[1], 200U); - EXPECT_EQ(arg.data(), matrix->getData()); - - EXPECT_EQ(arg.matrix().getHeight(), matrix->getHeight()); - EXPECT_EQ(arg.matrix().getWidth(), matrix->getWidth()); - EXPECT_EQ(arg.matrix().getData(), matrix->getData()); - }; - - BufferArgs argments; - argments.addArg(*matrix); - std::vector checkFunc; - checkFunc.push_back(check); - testBufferArgs(argments, checkFunc); -} - -TEST(Arguments, Vector) { - VectorPtr vector = Vector::create(100, false); - CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 1U); - EXPECT_EQ(arg.shape()[0], 100U); - EXPECT_EQ(arg.data(), vector->getData()); - - CpuVector inVector = arg.vector(); - EXPECT_EQ(inVector.getSize(), vector->getSize()); - EXPECT_EQ(inVector.getData(), vector->getData()); - }; - - BufferArgs argments; - argments.addArg(*vector); - std::vector checkFunc; - checkFunc.push_back(check); - testBufferArgs(argments, checkFunc); -} - -TEST(Arguments, CpuSparseMatrix) { - CpuSparseMatrix sparse(200, 300, 50); - CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 2U); - EXPECT_EQ(arg.shape()[0], 200U); - EXPECT_EQ(arg.shape()[1], 300U); - EXPECT_EQ(arg.data(), sparse.getData()); - // CHECK_EQ(arg.sparse().nnz(), 50); - // CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT); - // CHECK_EQ(arg.sparse().dataType(), SPARSE_FLOAT_VALUE); - EXPECT_EQ(arg.sparse().getRowBuf(), sparse.getRows()); - EXPECT_EQ(arg.sparse().getColBuf(), sparse.getCols()); - }; - - BufferArgs argments; - argments.addArg(sparse); - std::vector checkFunc; - checkFunc.push_back(check); - testBufferArgs(argments, checkFunc); -} - -TEST(Arguments, BufferArg) { - BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3}); - CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 3U); - EXPECT_EQ(arg.shape()[0], 1U); - EXPECT_EQ(arg.shape()[1], 2U); - EXPECT_EQ(arg.shape()[2], 3U); - }; - - BufferArgs argments; - argments.addArg(arg); - testBufferArgs(argments, check); -} - -} // namespace paddle diff --git a/paddle/function/FunctionTest.h b/paddle/function/FunctionTest.h deleted file mode 100644 index 14003d2c885c8f846f9445ad8844869c9112816e..0000000000000000000000000000000000000000 --- a/paddle/function/FunctionTest.h +++ /dev/null @@ -1,410 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Function.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" -#include "paddle/math/tests/TensorCheck.h" -#include "paddle/testing/TestUtil.h" - -namespace paddle { - -typedef std::shared_ptr BufferArgPtr; - -namespace test { -template -struct Allocator; - -template <> -struct Allocator { - using type = CpuMemoryHandle; -}; - -template <> -struct Allocator { - using type = GpuMemoryHandle; -}; - -// Copy argument1 to argument2 -template -class CopyArgument { - public: - void operator()(const BufferArg& arg1, BufferArg& arg2) { - CHECK_EQ(arg1.valueType(), arg2.valueType()); - CHECK_LE(arg1.shape().getElements(), arg2.shape().getElements()); - - if (arg1.valueType() == VALUE_TYPE_INT32) { - IVectorPtr vector1 = - IVector::create((int*)arg1.data(), - arg1.shape().getElements(), - DType1 == DEVICE_TYPE_CPU ? false : true); - IVectorPtr vector2 = - IVector::create((int*)arg2.data(), - arg2.shape().getElements(), - DType2 == DEVICE_TYPE_CPU ? false : true); - vector2->copyFrom(*vector1); - } else { - VectorPtr vector1 = - Vector::create((real*)arg1.data(), - arg1.shape().getElements(), - DType1 == DEVICE_TYPE_CPU ? false : true); - VectorPtr vector2 = - Vector::create((real*)arg2.data(), - arg2.shape().getElements(), - DType2 == DEVICE_TYPE_CPU ? false : true); - vector2->copyFrom(*vector1); - } - } -}; -} // namespace test - -/** - * \brief A class for comparing two Functions of different implementations. - * For example, can be used to compare the CPU and GPU implementation - * of the function is consistent. - * - * Use case: - * // Initializes a test object, the corresponding cpu and gpu Function - * // are constructed according to FunctionName and FuncConfig. - * CpuGpuFuncCompare test(FunctionName, FuncConfig); - * // Prepare inputs and outputs arguments. - * // Here the input and output can not contain real data, - * // only contains the argument type and shape. - * test.addInputs(input1); - * test.addInputs(input2); - * test.addOutputs(output1); - * test.addOutputs(output2); - * // Run. - * // Will according to the type and shape of arguments(inputs_/outputs_), - * // automatic initialization cpu and gpu function required arguments - * // (cpuInputs_/cpuOutputs_/gpuInputs_/gpuOutputs_). - * // Call the CPU and GPU Function calculation results. - * // Compares CPU and GPU calculation results for consistency. - * test.run(); - */ -template -class Compare2Function { - public: - typedef typename test::Allocator::type Allocator1; - typedef typename test::Allocator::type Allocator2; - typedef typename Tensor::Vector Vector1; - typedef typename Tensor::Vector Vector2; - typedef typename Tensor::SparseMatrix SparseMatrix1; - typedef typename Tensor::SparseMatrix SparseMatrix2; - - Compare2Function(const std::string& name1, - const std::string& name2, - const FuncConfig& config) - : function1_(FunctionBase::funcRegistrar_.createByType(name1)), - function2_(FunctionBase::funcRegistrar_.createByType(name2)) { - function1_->init(config); - function2_->init(config); - initArgsCallback_ = nullptr; - } - - ~Compare2Function() {} - - // input need only contains shape, do not contains data. - void addInputs(const BufferArg& input) { - size_t size = - input.shape().getElements() * sizeOfValuType(input.valueType()); - func1Memory_.emplace_back(std::make_shared(size)); - func2Memory_.emplace_back(std::make_shared(size)); - - func1Inputs_.emplace_back(std::make_shared( - func1Memory_.back()->getBuf(), input.valueType(), input.shape())); - func2Inputs_.emplace_back(std::make_shared( - func2Memory_.back()->getBuf(), input.valueType(), input.shape())); - } - - // assume one copy of sequence is shared by different SequenceArgs - void addSequence(const SequenceIdArg& input) { - CHECK_EQ(input.shape().ndims(), 1UL); - size_t batchSize = input.shape()[0]; - size_t numSeqs = batchSize / 10 + 1; - size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32); - func1Memory_.emplace_back(std::make_shared(sizeId)); - func2Memory_.emplace_back(std::make_shared(sizeId)); - seq1_ = std::make_shared(func1Memory_.back()->getBuf(), - TensorShape{numSeqs + 1}); - seq2_ = std::make_shared(func2Memory_.back()->getBuf(), - TensorShape{numSeqs + 1}); - /// init sequence Id - initArg(*seq1_, batchSize); - - copyArg_(*seq1_, *seq2_); - } - - void addInputs(const SequenceArg& input) { - CHECK_EQ(input.shape().ndims(), 2UL); - size_t batchSize = input.shape()[0]; - if (!seq1_ || !seq2_) { // sequence not exist - addSequence(SequenceIdArg(TensorShape{batchSize})); - } - - size_t size = - input.shape().getElements() * sizeOfValuType(input.valueType()); - func1Memory_.emplace_back(std::make_shared(size)); - func2Memory_.emplace_back(std::make_shared(size)); - - /// SequenceArg - func1Inputs_.emplace_back( - std::make_shared(func1Memory_.back()->getBuf(), - input.valueType(), - input.shape(), - *seq1_)); - func2Inputs_.emplace_back( - std::make_shared(func2Memory_.back()->getBuf(), - input.valueType(), - input.shape(), - *seq2_)); - } - - void registerInitCallback(std::function callback) { - initArgsCallback_ = callback; - } - - // output need only contains shape, do not contains data. - void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) { - size_t size = - output.shape().getElements() * sizeOfValuType(output.valueType()); - func1Memory_.emplace_back(std::make_shared(size)); - func2Memory_.emplace_back(std::make_shared(size)); - - func1Outputs_.emplace_back( - std::make_shared(func1Memory_.back()->getBuf(), - output.valueType(), - output.shape(), - argType)); - func2Outputs_.emplace_back( - std::make_shared(func2Memory_.back()->getBuf(), - output.valueType(), - output.shape(), - argType)); - } - - /// add and init output sparse matrix - void addOutputs(const SparseMatrixArg& output, ArgType argType = ASSIGN_TO) { - sparse1_ = std::make_shared( - output.shape()[0], - output.shape()[1], - output.nnz(), - static_cast(output.dataType()), - static_cast(output.dataFormat())); - - sparse2_ = std::make_shared( - output.shape()[0], - output.shape()[1], - output.nnz(), - static_cast(output.dataType()), - static_cast(output.dataFormat())); - - /// init sparse matrix - hl_stream_t stream(HPPL_STREAM_1); - sparse1_->randomizeUniform(); - sparse2_->copyFrom(*sparse1_, stream); - hl_stream_synchronize(stream); - - func1Outputs_.emplace_back( - std::make_shared(*sparse1_, argType)); - func2Outputs_.emplace_back( - std::make_shared(*sparse2_, argType)); - } - - void addOutputs(const SequenceArg& output, ArgType argType = ASSIGN_TO) { - CHECK_EQ(output.shape().ndims(), 2UL); - size_t batchSize = output.shape()[0]; - - if (!seq1_ || !seq2_) { // sequence not exist - addSequence(SequenceIdArg(TensorShape{batchSize})); - } - size_t size = - output.shape().getElements() * sizeOfValuType(output.valueType()); - func1Memory_.emplace_back(std::make_shared(size)); - func2Memory_.emplace_back(std::make_shared(size)); - - /// SequenceArg - func1Outputs_.emplace_back( - std::make_shared(func1Memory_.back()->getBuf(), - output.valueType(), - output.shape(), - *seq1_, - argType)); - func2Outputs_.emplace_back( - std::make_shared(func2Memory_.back()->getBuf(), - output.valueType(), - output.shape(), - *seq2_, - argType)); - } - - void addInputs(const SparseMatrixArg& input) { - sparse1_ = std::make_shared( - input.shape()[0], - input.shape()[1], - input.nnz(), - static_cast(input.dataType()), - static_cast(input.dataFormat())); - - sparse2_ = std::make_shared( - input.shape()[0], - input.shape()[1], - input.nnz(), - static_cast(input.dataType()), - static_cast(input.dataFormat())); - - /// init sparse matrix - hl_stream_t stream(HPPL_STREAM_1); - sparse1_->randomizeUniform(); - sparse2_->copyFrom(*sparse1_, stream); - hl_stream_synchronize(stream); - - func1Inputs_.emplace_back(std::make_shared(*sparse1_)); - func2Inputs_.emplace_back(std::make_shared(*sparse2_)); - } - - void run() { - // prepare cpu/gpu arguments - initInputs(); - - initOutputs(); - // function calculate - auto callFunction = [](FunctionBase* function, - std::vector& inputs, - std::vector& outputs) { - BufferArgs inArgs; - BufferArgs outArgs; - for (auto arg : inputs) { - inArgs.addArg(*arg); - } - for (auto arg : outputs) { - outArgs.addArg(*arg); - } - function->calc(inArgs, outArgs); - }; - - callFunction(function1_.get(), func1Inputs_, func1Outputs_); - callFunction(function2_.get(), func2Inputs_, func2Outputs_); - - // check outputs - compareOutputs(); - } - - std::shared_ptr getFunction1() const { return function1_; } - - std::shared_ptr getFunction2() const { return function2_; } - - protected: - // only init cpu argument, gpu argument copy from cpu argument. - void initArg(BufferArg& arg) { - Vector1 vector(arg.shape().getElements(), (real*)arg.data()); - vector.uniform(0.001, 1); - } - - void initArg(SequenceArg& arg) { - /// init only matrix - Vector1 vector(arg.shape().getElements(), (real*)arg.data()); - vector.uniform(0.001, 1); - } - - void initArg(SequenceIdArg& arg, size_t batchSize) { - size_t numSeqs = arg.numSeqs(); - int* buf = reinterpret_cast(arg.data()); - int pos = 0; - size_t maxLen = 2 * batchSize / numSeqs; - for (int i = 0; i < (int)numSeqs; ++i) { - int len = 1 + uniformRandom(std::min( - maxLen, batchSize - pos - numSeqs + i)); - buf[i] = pos; - pos += len; - VLOG(1) << " len=" << len; - } - buf[numSeqs] = batchSize; - } - - void initInputs() { - for (size_t i = 0; i < func1Inputs_.size(); i++) { - if (func1Inputs_[i]->isSparseArg()) { - continue; /// sparse matrix already init - } - - if (func1Inputs_[i]->isSequenceArg()) { - initArg(dynamic_cast(*func1Inputs_[i])); - } else { - initArg(*func1Inputs_[i]); - } - - if (initArgsCallback_ != nullptr) { - initArgsCallback_(*func1Inputs_[i], i); - } - - copyArg_(*func1Inputs_[i], *func2Inputs_[i]); - } - } - - void initOutputs() { - for (size_t i = 0; i < func1Outputs_.size(); i++) { - if (func1Outputs_[i]->isSparseArg()) { - continue; /// sparse matrix already init - } - - if (func1Outputs_[i]->isSequenceArg()) { - initArg(dynamic_cast(*func1Outputs_[i])); - } else { - initArg(*func1Outputs_[i]); - } - - copyArg_(*func1Outputs_[i], *func2Outputs_[i]); - } - } - - void compareOutputs() { - for (size_t i = 0; i < func1Outputs_.size(); i++) { - // TODO, Need a BufferCheck used to compare the two buffers. - const auto cpu = func1Outputs_[i]; - const auto gpu = func2Outputs_[i]; - CHECK_EQ(cpu->numElements(), gpu->numElements()); - Vector1 cpuVector(cpu->numElements(), (real*)cpu->data()); - Vector2 gpuVector(gpu->numElements(), (real*)gpu->data()); - autotest::TensorCheckErr(cpuVector, gpuVector); - } - } - - protected: - std::shared_ptr function1_; - std::shared_ptr function2_; - std::vector> func1Memory_; - std::vector> func2Memory_; - std::vector func1Inputs_; - std::vector func1Outputs_; - std::vector func2Inputs_; - std::vector func2Outputs_; - std::shared_ptr sparse1_; - std::shared_ptr sparse2_; - std::shared_ptr seq1_; - std::shared_ptr seq2_; - test::CopyArgument copyArg_; - std::function initArgsCallback_; -}; - -class CpuGpuFuncCompare - : public Compare2Function { - public: - CpuGpuFuncCompare(const std::string& name, const FuncConfig& config) - : Compare2Function(name + "-CPU", name + "-GPU", config) {} - - ~CpuGpuFuncCompare() {} -}; - -} // namespace paddle diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp deleted file mode 100644 index 5b023e2c10e5040a28660d555efceb0e26b40d49..0000000000000000000000000000000000000000 --- a/paddle/function/GemmConvOp.cpp +++ /dev/null @@ -1,522 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvOp.h" -#include "GemmFunctor.h" -#include "Im2Col.h" -#include "paddle/math/MemoryHandle.h" - -namespace paddle { - -/* - * \brief Forward calculation of convolution. - */ -template -class GemmConvFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - check(inputs, outputs); - // TODO(hedaoyuan): Need to define some index macros, - // to avoid useing 0 and 1. - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - - real beta; - if (outputs[0].getArgType() == ADD_TO) { - beta = 1.0; - } else { - beta = 0.0; - } - - size_t batchSize = input[0]; - size_t inputChannels = input[1]; - size_t inputHeight = input[2]; - size_t inputWidth = input[3]; - size_t filterHeight = getFilterHeight(filter); - size_t filterWidth = getFilterWidth(filter); - size_t outputChannels = output[1]; - size_t outputHeight = output[2]; - size_t outputWidth = output[3]; - - real* inputData = inputs[0].data(); - real* filterData = inputs[1].data(); - real* outputData = outputs[0].data(); - bool needIm2col = isNeedIm2col(filter); - - TensorShape imShape = - TensorShape({inputChannels / groups_, inputHeight, inputWidth}); - - TensorShape colShape; - real* colData = NULL; - - if (needIm2col) { - colShape = TensorShape({inputChannels / groups_, - filterHeight, - filterWidth, - outputHeight, - outputWidth}); - resizeBuffer(colShape.getElements()); - colData = reinterpret_cast(memory_->getBuf()); - } - - Im2ColFunctor im2col; - size_t inputOffset = imShape.getElements(); - size_t outputOffset = - (outputChannels / groups_) * outputHeight * outputWidth; - size_t filterOffset = filter.getElements() / groups_; - - for (size_t i = 0; i < batchSize; i++) { - for (size_t g = 0; g < groups_; g++) { - if (needIm2col) { - im2col(inputData + g * inputOffset, - imShape, - colData, - colShape, - strideH(), - strideW(), - paddingH(), - paddingW(), - dilationH(), - dilationW()); - } else { - colData = inputData + g * inputOffset; - } - int M = outputChannels / groups_; - int N = outputHeight * outputWidth; - int K = inputChannels / groups_ * filterHeight * filterWidth; - BlasGemm::compute(false, - false, - M, - N, - K, - 1.0f, - filterData + g * filterOffset, - K, - colData, - N, - beta, - outputData + g * outputOffset, - N); - } - inputData += inputChannels * inputHeight * inputWidth; - outputData += outputChannels * outputHeight * outputWidth; - } - } -}; - -#ifdef PADDLE_MOBILE_INFERENCE - -/* - * \brief Forward calculation of convolution, optimized for mobile. - */ -template -class GemmConvMobileFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - check(inputs, outputs); - // TODO(hedaoyuan): Need to define some index macros, - // to avoid useing 0 and 1. - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - - real beta; - if (outputs[0].getArgType() == ADD_TO) { - beta = 1.0; - } else { - beta = 0.0; - } - - size_t batchSize = input[0]; - size_t inputChannels = input[1]; - size_t inputHeight = input[2]; - size_t inputWidth = input[3]; - size_t filterHeight = getFilterHeight(filter); - size_t filterWidth = getFilterWidth(filter); - size_t outputChannels = output[1]; - size_t outputHeight = output[2]; - size_t outputWidth = output[3]; - - real* inputData = inputs[0].data(); - real* filterData = inputs[1].data(); - real* outputData = outputs[0].data(); - real* colData = NULL; - bool needIm2col = isNeedIm2col(filter); - - TensorShape imShape = - TensorShape({inputChannels / groups_, inputHeight, inputWidth}); - TensorShape colShape; - - // Max col matrix width 4096, Max col matrix size 4M. - size_t outputHeightSteps = - std::min(std::max(4096 / outputWidth, (size_t)1), outputHeight); - size_t maxColWidth = outputHeightSteps * outputWidth; - size_t channelSteps = - std::min(std::max((1048576 / maxColWidth) / filterHeight * filterWidth, - (size_t)1), - inputChannels / groups_); - size_t maxColHeight = channelSteps * filterHeight * filterWidth; - - if (needIm2col) { - colShape = TensorShape({inputChannels / groups_, - filterHeight, - filterWidth, - outputHeight, - outputWidth}); - - resizeBuffer(maxColHeight * maxColWidth * sizeof(real)); - colData = reinterpret_cast(memory_->getBuf()); - } - - Im2ColMobileFunctor im2col; - size_t inputOffset = imShape.getElements(); - size_t outputOffset = - (outputChannels / groups_) * outputHeight * outputWidth; - size_t filterOffset = filter.getElements() / groups_; - - int nStride = outputHeight * outputWidth; - int kStride = inputChannels / groups_ * filterHeight * filterWidth; - for (size_t i = 0; i < batchSize; i++) { - filterData = inputs[1].data(); - for (size_t g = 0; g < groups_; g++) { - if (needIm2col) { - real beta_ = beta; - for (size_t ic = 0; ic < inputChannels / groups_; - ic += channelSteps) { - int channels = std::min(inputChannels / groups_ - ic, channelSteps); - for (size_t oh = 0; oh < outputHeight; oh += outputHeightSteps) { - int height = std::min(outputHeight - oh, outputHeightSteps); - - int M = outputChannels / groups_; - int N = height * outputWidth; - int K = channels * filterHeight * filterWidth; - // im2col - im2col(inputData, - imShape, - colData, - colShape, - strideH(), - strideW(), - paddingH(), - paddingW(), - dilationH(), - dilationW(), - channels, - oh, - height, - N); - - // gemm - BlasGemm::compute( - false, - false, - M, - N, - K, - 1.0f, - filterData + ic * filterHeight * filterWidth, - kStride, - colData, - N, - beta_, - outputData + oh * outputWidth, - nStride); - } - beta_ = 1.0; - } - } else { - int M = outputChannels / groups_; - int N = outputHeight * outputWidth; - int K = inputChannels / groups_ * filterHeight * filterWidth; - BlasGemm::compute(false, - false, - M, - N, - K, - 1.0f, - filterData, - K, - inputData, - N, - beta, - outputData, - N); - } - inputData += inputOffset; - outputData += outputOffset; - filterData += filterOffset; - } - } - - memory_.reset(); - } -}; - -#endif - -/* - * \brief Backward input calculation of convolution. - */ -template -class GemmConvGradInputFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& output = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& input = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - check(inputs, outputs); - // Since the implementation of Col2ImFunctor is ADD_TO, - // this function only supports ADD_TO mode. - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - const TensorShape& output = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& input = outputs[0].shape(); - - size_t batchSize = input[0]; - size_t inputChannels = input[1]; - size_t inputHeight = input[2]; - size_t inputWidth = input[3]; - size_t filterHeight = getFilterHeight(filter); - size_t filterWidth = getFilterWidth(filter); - size_t outputChannels = output[1]; - size_t outputHeight = output[2]; - size_t outputWidth = output[3]; - - real* outputGrad = inputs[0].data(); - real* filterData = inputs[1].data(); - real* inputGrad = outputs[0].data(); - bool needIm2col = isNeedIm2col(filter); - - TensorShape imShape = - TensorShape({inputChannels / groups_, inputHeight, inputWidth}); - - TensorShape colShape; - real* colData = NULL; - - if (needIm2col) { - colShape = TensorShape({inputChannels / groups_, - filterHeight, - filterWidth, - outputHeight, - outputWidth}); - resizeBuffer(colShape.getElements()); - colData = reinterpret_cast(memory_->getBuf()); - } - - Col2ImFunctor col2im; - size_t inputOffset = imShape.getElements(); - size_t outputOffset = - (outputChannels / groups_) * outputHeight * outputWidth; - size_t filterOffset = filter.getElements() / groups_; - - for (size_t i = 0; i < batchSize; i++) { - for (size_t g = 0; g < groups_; g++) { - int K = outputChannels / groups_; - int N = outputHeight * outputWidth; - int M = inputChannels / groups_ * filterHeight * filterWidth; - real scale = 0.0f; - if (!needIm2col) { - colData = inputGrad + g * inputOffset; - scale = 1.0f; - } - BlasGemm::compute(true, - false, - M, - N, - K, - 1.0f, - filterData + g * filterOffset, - M, - outputGrad + g * outputOffset, - N, - scale, - colData, - N); - if (needIm2col) { - col2im(inputGrad + g * inputOffset, - imShape, - colData, - colShape, - strideH(), - strideW(), - paddingH(), - paddingW(), - dilationH(), - dilationW()); - } - } - inputGrad += inputChannels * inputHeight * inputWidth; - outputGrad += outputChannels * outputHeight * outputWidth; - } - } -}; - -/* - * \brief Backward filter calculation of convolution. - */ -template -class GemmConvGradFilterFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& output = inputs[0].shape(); - const TensorShape& input = inputs[1].shape(); - const TensorShape& filter = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - check(inputs, outputs); - const TensorShape& output = inputs[0].shape(); - const TensorShape& input = inputs[1].shape(); - const TensorShape& filter = outputs[0].shape(); - - real beta; - if (outputs[0].getArgType() == ADD_TO) { - beta = 1.0; - } else { - beta = 0.0; - } - - size_t batchSize = input[0]; - size_t inputChannels = input[1]; - size_t inputHeight = input[2]; - size_t inputWidth = input[3]; - size_t filterHeight = getFilterHeight(filter); - size_t filterWidth = getFilterWidth(filter); - size_t outputChannels = output[1]; - size_t outputHeight = output[2]; - size_t outputWidth = output[3]; - - real* outputGrad = inputs[0].data(); - real* inputData = inputs[1].data(); - real* filterGrad = outputs[0].data(); - bool needIm2col = isNeedIm2col(filter); - - TensorShape imShape = - TensorShape({inputChannels / groups_, inputHeight, inputWidth}); - - TensorShape colShape; - real* colData = NULL; - - if (needIm2col) { - colShape = TensorShape({inputChannels / groups_, - filterHeight, - filterWidth, - outputHeight, - outputWidth}); - resizeBuffer(colShape.getElements()); - colData = reinterpret_cast(memory_->getBuf()); - } - - Im2ColFunctor im2col; - size_t inputOffset = imShape.getElements(); - size_t outputOffset = - (outputChannels / groups_) * outputHeight * outputWidth; - size_t filterOffset = filter.getElements() / groups_; - for (size_t i = 0; i < batchSize; i++) { - for (size_t g = 0; g < groups_; g++) { - if (needIm2col) { - im2col(inputData + g * inputOffset, - imShape, - colData, - colShape, - strideH(), - strideW(), - paddingH(), - paddingW(), - dilationH(), - dilationW()); - } else { - colData = inputData + g * inputOffset; - } - int M = outputChannels / groups_; - int K = outputHeight * outputWidth; - int N = inputChannels / groups_ * filterHeight * filterWidth; - BlasGemm::compute(false, - true, - M, - N, - K, - 1.0f, - outputGrad + g * outputOffset, - K, - colData, - K, - i == 0 ? beta : 1.0f, - filterGrad + g * filterOffset, - N); - } - inputData += inputChannels * inputHeight * inputWidth; - outputGrad += outputChannels * outputHeight * outputWidth; - } - } -}; - -#ifdef PADDLE_MOBILE_INFERENCE -REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvMobileFunction); -#else -REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction); -#endif -REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction); -REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction); -REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction); -REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction); -#endif - -} // namespace paddle diff --git a/paddle/function/GemmFunctor.cpp b/paddle/function/GemmFunctor.cpp deleted file mode 100644 index 0b1fe1b67d8fd6caf86a08bc05e250b1936e9f85..0000000000000000000000000000000000000000 --- a/paddle/function/GemmFunctor.cpp +++ /dev/null @@ -1,90 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "GemmFunctor.h" -#include "paddle/math/MathFunctions.h" - -namespace paddle { - -template -struct BlasGemm { - static void compute(const bool transA, - const bool transB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc) { -#ifdef PADDLE_USE_EIGEN_FOR_BLAS - EigenBlasGemm::compute( - transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc); -#else - gemm(transA == false ? CblasNoTrans : CblasTrans, - transB == false ? CblasNoTrans : CblasTrans, - M, - N, - K, - alpha, - A, - lda, - B, - ldb, - beta, - C, - ldc); -#endif - } -}; - -template -struct BlasGemm { - static void compute(const bool transA, - const bool transB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc) { - hl_matrix_mul((T*)A, - transA == false ? HPPL_OP_N : HPPL_OP_T, - (T*)B, - transB == false ? HPPL_OP_N : HPPL_OP_T, - C, - M, - N, - K, - alpha, - beta, - lda, - ldb, - ldc); - } -}; - -template struct BlasGemm; -template struct BlasGemm; - -} // namespace paddle diff --git a/paddle/function/Im2ColTest.cpp b/paddle/function/Im2ColTest.cpp deleted file mode 100644 index 967c5b91536608364b4181707b843799b1764c3f..0000000000000000000000000000000000000000 --- a/paddle/function/Im2ColTest.cpp +++ /dev/null @@ -1,223 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Im2Col.h" -#include -#include "Function.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/tests/TensorCheck.h" - -namespace paddle { - -template -void TestIm2ColFunctor() { - for (size_t channels : {1, 5, 32}) { - for (size_t inputHeight : {5, 33, 100}) { - for (size_t inputWidth : {5, 32, 96}) { - for (size_t filterHeight : {1, 5}) { - for (size_t filterWidth : {3, 7}) { - for (size_t stride : {1, 2}) { - for (size_t padding : {0, 1}) { - for (size_t dilation : {1, 3}) { - size_t filterSizeH = (filterHeight - 1) * dilation + 1; - size_t filterSizeW = (filterWidth - 1) * dilation + 1; - if (inputHeight + 2 * padding < filterSizeH || - inputWidth + 2 * padding < filterSizeW) - break; - if (padding >= filterSizeH || padding >= filterSizeW) break; - size_t outputHeight = - (inputHeight - filterSizeH + 2 * padding) / stride + 1; - size_t outputWidth = - (inputWidth - filterSizeW + 2 * padding) / stride + 1; - - TensorShape imShape = - TensorShape({channels, inputHeight, inputWidth}); - TensorShape colShape1 = TensorShape({channels, - filterHeight, - filterWidth, - outputHeight, - outputWidth}); - TensorShape colShape2 = TensorShape({outputHeight, - outputWidth, - channels, - filterHeight, - filterWidth}); - - size_t height = channels * filterHeight * filterWidth; - size_t width = outputHeight * outputWidth; - VectorPtr input1 = - Vector::create(imShape.getElements(), false); - VectorPtr input2 = - Vector::create(imShape.getElements(), false); - MatrixPtr output1 = - Matrix::create(height, width, false, false); - MatrixPtr output2 = - Matrix::create(width, height, false, false); - input1->uniform(0.001, 1); - input2->copyFrom(*input1); - - Im2ColFunctor im2Col1; - Im2ColFunctor im2Col2; - im2Col1(input1->getData(), - imShape, - output1->getData(), - colShape1, - stride, - stride, - padding, - padding, - dilation, - dilation); - im2Col2(input2->getData(), - imShape, - output2->getData(), - colShape2, - stride, - stride, - padding, - padding, - dilation, - dilation); - - // The transposition of the result of ColFormat == kCFO - // is equal to the result of ColFormat == kOCF. - MatrixPtr test; - output2->transpose(test, true); - autotest::TensorCheckErr(*output1, *test); - - Col2ImFunctor col2Im1; - Col2ImFunctor col2Im2; - - col2Im1(input1->getData(), - imShape, - output1->getData(), - colShape1, - stride, - stride, - padding, - padding, - dilation, - dilation); - col2Im2(input2->getData(), - imShape, - output2->getData(), - colShape2, - stride, - stride, - padding, - padding, - dilation, - dilation); - autotest::TensorCheckErr(*input1, *input2); - } - } - } - } - } - } - } - } -} - -TEST(Im2ColFunctor, CPU) { TestIm2ColFunctor(); } - -#ifdef PADDLE_WITH_CUDA - -TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor(); } - -#endif - -template -void TestIm2ColMobileFunctor() { - for (size_t channels : {32}) { - for (size_t inputHeight : {33, 100}) { - for (size_t inputWidth : {32, 96}) { - for (size_t filterHeight : {5}) { - for (size_t filterWidth : {7}) { - for (size_t stride : {2}) { - for (size_t padding : {1}) { - for (size_t dilation : {1, 3}) { - size_t filterSizeH = (filterHeight - 1) * dilation + 1; - size_t filterSizeW = (filterWidth - 1) * dilation + 1; - if (inputHeight + 2 * padding < filterSizeH || - inputWidth + 2 * padding < filterSizeW) - break; - if (padding >= filterSizeH || padding >= filterSizeW) break; - size_t outputHeight = - (inputHeight - filterSizeH + 2 * padding) / stride + 1; - size_t outputWidth = - (inputWidth - filterSizeW + 2 * padding) / stride + 1; - - TensorShape imShape = - TensorShape({channels, inputHeight, inputWidth}); - TensorShape colShape1 = TensorShape({channels, - filterHeight, - filterWidth, - outputHeight, - outputWidth}); - - size_t height = channels * filterHeight * filterWidth; - size_t width = outputHeight * outputWidth; - VectorPtr input1 = - Vector::create(imShape.getElements(), false); - VectorPtr input2 = - Vector::create(imShape.getElements(), false); - MatrixPtr output1 = - Matrix::create(height, width, false, false); - MatrixPtr output2 = - Matrix::create(height, width, false, false); - input1->uniform(0.001, 1); - input2->copyFrom(*input1); - - Im2ColFunctor im2Col1; - Im2ColMobileFunctor im2Col2; - im2Col1(input1->getData(), - imShape, - output1->getData(), - colShape1, - stride, - stride, - padding, - padding, - dilation, - dilation); - im2Col2(input2->getData(), - imShape, - output2->getData(), - colShape1, - stride, - stride, - padding, - padding, - dilation, - dilation, - channels, - 0, - outputHeight, - outputHeight * outputWidth); - - autotest::TensorCheckEqual(*output1, *output2); - } - } - } - } - } - } - } - } -} - -TEST(Im2ColFunctor, Mobile) { TestIm2ColMobileFunctor(); } - -} // namespace paddle diff --git a/paddle/function/MulOp.cpp b/paddle/function/MulOp.cpp deleted file mode 100644 index 7bf36c8050a8c33d836ce98dc7f3cf6d3de38d55..0000000000000000000000000000000000000000 --- a/paddle/function/MulOp.cpp +++ /dev/null @@ -1,347 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MulOp.h" -#include "GemmFunctor.h" -#include "paddle/math/SIMDFunctions.h" -#include "paddle/utils/ThreadLocal.h" - -namespace { -inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { - for (unsigned int i = 0; i < len; ++i) { - a[i] += (1.0 == scaleB) ? b[i] : scaleB * b[i]; - } -} - -inline void colVecAddTo( - real* a, real* b, real c, size_t len, size_t aWidth, size_t bWidth) { - for (unsigned int i = 0; i < len; ++i) { - a[i * aWidth] += (1.0 == c) ? b[i * bWidth] : b[i * bWidth] * c; - } -} -} // namespace - -namespace paddle { -/// sparse matrix (+)= dense matrix * dense matrix -template <> -void MulOp(CpuSparseMatrix& out, - const CpuMatrix& a, - const CpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - CHECK_EQ(out.getValueType(), FLOAT_VALUE); - if (scaleT == 0) { - out.zeroMem(); - } - const real* A = a.getData(); - const real* B = b.getData(); - real* C = out.getValue(); - int* rows = out.getRows(); - int* cols = out.getCols(); - size_t width = out.getWidth(); - size_t height = out.getHeight(); - - /// SPARSE_CSC, {a any, b not trans} - if (out.getFormat() == SPARSE_CSC) { - /// b not trans and a any - CHECK(!bTrans); - size_t m = !aTrans ? a.getWidth() : a.getHeight(); - for (size_t i = 0; i < width; i++) { - size_t start = out.getColStartIdx(i); - size_t end = out.getColStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - real sum = 0; - size_t rowIdx = rows[j]; - for (size_t k = 0; k < m; k++) { - sum += (!aTrans ? A[rowIdx * m + k] : A[k * height + rowIdx]) * - B[k * width + i]; - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - return; - } - - /// SPARSE_CSR, {a any, b not trans} or {a not trans, b trans} - if (out.getFormat() == SPARSE_CSR) { - /// a and b can not both transpose - CHECK(!(aTrans && bTrans)); - size_t m = a.getWidth(); - for (size_t i = 0; i < height; i++) { - size_t start = out.getRowStartIdx(i); - size_t end = out.getRowStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - real sum = 0; - size_t colIdx = cols[j]; - for (size_t k = 0; k < m; k++) { - sum += (!aTrans ? A[i * m + k] : A[k * height + i]) * - (!bTrans ? B[k * width + colIdx] : B[colIdx * m + k]); - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - return; - } -} - -/// dense matrix (+)= dense matrix * dense matrix -template <> -void MulOp(CpuMatrix& out, - const CpuMatrix& a, - const CpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - BlasGemm::compute( - aTrans, - bTrans, - out.getHeight(), - out.getWidth(), - !aTrans ? a.getWidth() : a.getHeight(), - scaleAB, - a.getData(), - a.getStride(), - b.getData(), - b.getStride(), - scaleT, - out.getData(), - out.getStride()); -} - -/// dense matrix (+)= sparse matrix * dense matrix -template <> -void MulOp(CpuMatrix& out, - const CpuSparseMatrix& a, - const CpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - if (scaleT == 0) { - out.zeroMem(); - } - const real* B = b.getData(); - real* C = out.getData(); - if (out.getWidth() % 32 == 0) { - CHECK_EQ((size_t)B % 32, 0UL); - CHECK_EQ((size_t)C % 32, 0UL); - } - - int* cols = a.getCols(); - real* values = a.getValue(); - for (size_t i = 0; i < a.getHeight(); ++i) { - const int start = a.getRowStartIdx(i); - const int end = a.getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - vecAddTo(!aTrans ? out.getRow(i) : out.getRow(cols[j]), - !aTrans ? const_cast(b).getRow(cols[j]) - : const_cast(b).getRow(i), - (a.getValueType() == FLOAT_VALUE) ? values[j] : (real)1.0, - out.getWidth()); - } - } -} - -/// dense matrix (+)= dense matrix * sparse matrix -template <> -void MulOp(CpuMatrix& out, - const CpuMatrix& a, - const CpuSparseMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - if (scaleT == 0) { - out.zeroMem(); - } - real* A = const_cast(a.getData()); - real* B = const_cast(b.getValue()); - real* C = out.getData(); - int* rows = b.getRows(); - int* cols = b.getCols(); - - /// SPARSE_CSC format - if (b.getFormat() == SPARSE_CSC) { - for (size_t j = 0; j < b.getWidth(); ++j) { - int start = b.getColStartIdx(j); - int end = b.getColStartIdx(j + 1); - for (int i = start; i < end; ++i) { - colVecAddTo(!bTrans ? C + j : C + rows[i], - !bTrans ? A + rows[i] : A + j, - (b.getValueType() == NO_VALUE) ? (real)1.0 : B[i], - out.getHeight(), - out.getWidth(), - a.getWidth()); - } - } - return; - } - - /// SPARSE_CSR format - if (b.getFormat() == SPARSE_CSR) { - for (size_t j = 0; j < b.getHeight(); ++j) { - int start = b.getRowStartIdx(j); - int end = b.getRowStartIdx(j + 1); - for (int i = start; i < end; ++i) { - colVecAddTo(!bTrans ? C + cols[i] : C + j, - !bTrans ? A + j : A + cols[i], - (b.getValueType() == NO_VALUE) ? (real)1.0 : B[i], - out.getHeight(), - out.getWidth(), - a.getWidth()); - } - } - return; - } -} - -/** - * mul operator - * out = scaleT * out + scaleAB * (A * B) - * here, scaleT in {0, 1}, scaleAB == 1, - * out = A * B, ASSIGN_TO - * out += A * B, ADD_TO - * - * - * \param outputs[0] output matrix (out), M * N, - * could be either Sparse or Dense Matrix - * M is num of rows, N is num of columns - * \param inputs[0] first input matrix (A), M * K (if non-trans) - * could be either Sparse or Dense Matrix - * M is num of rows, K is num of columns - * \param inputs[1] second input matrix (B), K * N (if non-trans) - * could be either Sparse or Dense Matrix - * K is num of rows, N is num of columns - * - * Support eight Mul operators, with both GPU and CPU devices - * For each device, four Mul operators are supported: - * 1. dense (out) = dense (A) * dense (B) - * 2. dense (out) = sparse (A) * dense (B) - * sparse matrix only support SPARSE_CSR format - * 3. dense (out) = dense (A) * sparse (B) - * sparse matrix support SPARSE_CSC and SPARSE_CSR formats - * 4. sparse (out) = dense (A) * dense (B) - * sparse matrix support SPARSE_CSC and SPARSE_CSR formats - * - */ -template -class MulFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { - aTrans_ = config.get("aTrans"); - bTrans_ = config.get("bTrans"); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK(!aTrans_ || !bTrans_) - << "Not support both a and b are transpose matrices"; - - CHECK_EQ((size_t)2, inputs.size()); - CHECK_EQ((size_t)1, outputs.size()); - CHECK(inputs[0].data() && inputs[1].data() && outputs[0].data()); - CHECK_EQ(inputs[0].shape().ndims(), (size_t)2); - CHECK_EQ(inputs[1].shape().ndims(), (size_t)2); - CHECK_EQ(outputs[0].shape().ndims(), (size_t)2); - - size_t aRow = !aTrans_ ? inputs[0].shape()[0] : inputs[0].shape()[1]; - size_t aCol = !aTrans_ ? inputs[0].shape()[1] : inputs[0].shape()[0]; - size_t bRow = !bTrans_ ? inputs[1].shape()[0] : inputs[1].shape()[1]; - size_t bCol = !bTrans_ ? inputs[1].shape()[1] : inputs[1].shape()[0]; - /// C = A * B, or C += A * B, for matrix format - CHECK_EQ(aCol, bRow); - CHECK_EQ(aRow, outputs[0].shape()[0]); - CHECK_EQ(bCol, outputs[0].shape()[1]); - - /// only support C = A * B (ASSIGN_TO) or C += A * B (ADD_TO) - real scaleT = (outputs[0].getArgType() == ADD_TO) ? 1.0 : 0.0; - - /// support dense = not both sparse * sparse - /// or sparse = dense * dense - CHECK((!outputs[0].isSparseArg() && - !(inputs[0].isSparseArg() && inputs[1].isSparseArg())) || - (outputs[0].isSparseArg() && !inputs[0].isSparseArg() && - !inputs[1].isSparseArg())); - - auto outMat = outputs[0].matrix(); - /// dense matrix = dense matrix * dense matrix - if (!inputs[0].isSparseArg() && !inputs[1].isSparseArg() && - !outputs[0].isSparseArg()) { - MulOp(outMat, - inputs[0].matrix(), - inputs[1].matrix(), - 1.0, // scaleAB - scaleT, - aTrans_, - bTrans_); - return; - } - - /// dense matrix = dense matrix * sparse matrix - if (!inputs[0].isSparseArg() && inputs[1].isSparseArg() && - !outputs[0].isSparseArg()) { - CHECK(!aTrans_) << "Not supported a transpose"; - MulOp(outMat, - inputs[0].matrix(), - inputs[1].sparse().SparseMatrix(), - 1.0, // scaleAB - scaleT, - aTrans_, - bTrans_); - return; - } - - /// dense matrix = sparse matrix * dense matrix - if (inputs[0].isSparseArg() && !inputs[1].isSparseArg() && - !outputs[0].isSparseArg()) { - CHECK(!bTrans_) << "Not supported b transpose"; - CHECK_EQ(inputs[0].sparse().dataFormat(), T_SPARSE_CSR) - << "Only supported SPARSE_CSR format for sparse matrix a"; - MulOp(outMat, - inputs[0].sparse().SparseMatrix(), - inputs[1].matrix(), - 1.0, // scaleAB - scaleT, - aTrans_, - bTrans_); - return; - } - - /// sparse matrix = dense matrix * dense matrix - auto outSparseMat = outputs[0].sparse().SparseMatrix(); - if (!inputs[0].isSparseArg() && !inputs[1].isSparseArg() && - outputs[0].isSparseArg()) { - MulOp(outSparseMat, - inputs[0].matrix(), - inputs[1].matrix(), - 1.0, // scaleAB - scaleT, - aTrans_, - bTrans_); - return; - } - } - - private: - bool aTrans_; - bool bTrans_; -}; - -REGISTER_TYPED_FUNC(MulOp, CPU, MulFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(MulOp, GPU, MulFunc); -#endif -} // namespace paddle diff --git a/paddle/function/MulOp.h b/paddle/function/MulOp.h deleted file mode 100644 index e6057be4e54b3cc2b3502b9a93825d4b53037c91..0000000000000000000000000000000000000000 --- a/paddle/function/MulOp.h +++ /dev/null @@ -1,102 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Function.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" - -namespace paddle { -/// CPU, dense matrix (+)= dense matrix * dense matrix -template -void MulOp(CpuMatrix& out, - const CpuMatrix& a, - const CpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// CPU, dense matrix (+)= sparse matrix * dense matrix -template -void MulOp(CpuMatrix& out, - const CpuSparseMatrix& a, - const CpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// CPU, dense matrix (+)= dense matrix * sparse matrix -template -void MulOp(CpuMatrix& out, - const CpuMatrix& a, - const CpuSparseMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// CPU, sparse matrix (+)= dense matrix * dense matrix -template -void MulOp(CpuSparseMatrix& out, - const CpuMatrix& a, - const CpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// GPU, dense matrix (+)= dense matrix * dense matrix -template -void MulOp(GpuMatrix& out, - const GpuMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// GPU, dense matrix (+)= sparse matrix * dense matrix -template -void MulOp(GpuMatrix& out, - const GpuSparseMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// GPU, dense matrix (+)= dense matrix * sparse matrix -template -void MulOp(GpuMatrix& out, - const GpuMatrix& a, - const GpuSparseMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -/// GPU, sparse matrix (+)= dense matrix * dense matrix -template -void MulOp(GpuSparseMatrix& out, - const GpuMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans); - -} // namespace paddle diff --git a/paddle/function/MulOpGpu.cu b/paddle/function/MulOpGpu.cu deleted file mode 100644 index d63416a8e45346089bac23100742b8afc99b8e77..0000000000000000000000000000000000000000 --- a/paddle/function/MulOpGpu.cu +++ /dev/null @@ -1,130 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MulOp.h" -#include "hl_base.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" - -namespace paddle { -/// dense matrix (+)= dense matrix * dense matrix -template <> -void MulOp(GpuMatrix& out, - const GpuMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; - hl_matrix_mul(const_cast(a.getData()), - !aTrans ? HPPL_OP_N : HPPL_OP_T, - const_cast(b.getData()), - !bTrans ? HPPL_OP_N : HPPL_OP_T, - const_cast(out.getData()), - out.getHeight(), - out.getWidth(), - !aTrans ? a.getWidth() : a.getHeight(), - scaleAB, - scaleT, - a.getStride(), - b.getStride(), - out.getStride()); -} - -/// dense matrix (+)= sparse matrix * dense matrix -template <> -void MulOp(GpuMatrix& out, - const GpuSparseMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - CHECK(out.isContiguous()); - CHECK(b.isContiguous()); - CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; - hl_matrix_csr_mul_dense(a.sMatrix_.get(), - aTrans ? HPPL_OP_T : HPPL_OP_N, - const_cast(b.getData()), - HPPL_OP_N, - const_cast(out.getData()), - out.getHeight(), - out.getWidth(), - b.getHeight(), - scaleAB, - scaleT); -} - -/// dense matrix (+)= dense matrix * sparse matrix -template <> -void MulOp(GpuMatrix& out, - const GpuMatrix& a, - const GpuSparseMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - CHECK(out.isContiguous()); - CHECK(a.isContiguous()); - CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; - - if (b.format_ == SPARSE_CSC) { - hl_matrix_dense_mul_csc(const_cast(a.getData()), - HPPL_OP_N, - b.sMatrix_.get(), - bTrans ? HPPL_OP_T : HPPL_OP_N, - const_cast(out.getData()), - out.getHeight(), - out.getWidth(), - a.getWidth(), - scaleAB, - scaleT); - } else { - hl_matrix_dense_mul_csr(const_cast(a.getData()), - HPPL_OP_N, - b.sMatrix_.get(), - bTrans ? HPPL_OP_T : HPPL_OP_N, - const_cast(out.getData()), - out.getHeight(), - out.getWidth(), - a.getWidth(), - scaleAB, - scaleT); - } -} - -/// sparse matrix (+)= dense matrix * dense matrix -template <> -void MulOp(GpuSparseMatrix& out, - const GpuMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT, - bool aTrans, - bool bTrans) { - CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; - hl_sparse_matrix_mul(const_cast(a.getData()), - aTrans ? HPPL_OP_T : HPPL_OP_N, - const_cast(b.getData()), - bTrans ? HPPL_OP_T : HPPL_OP_N, - out.sMatrix_.get(), - out.getHeight(), - out.getWidth(), - !bTrans ? b.getHeight() : b.getWidth(), - scaleAB, - scaleT); -} - -} // namespace paddle diff --git a/paddle/function/MulOpTest.cpp b/paddle/function/MulOpTest.cpp deleted file mode 100644 index 4e1ebd749c0cd083c025e43a321d6992a11786ff..0000000000000000000000000000000000000000 --- a/paddle/function/MulOpTest.cpp +++ /dev/null @@ -1,212 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "FunctionTest.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" -#include "paddle/math/tests/test_matrixUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT - -/** - * C += A * B, A, B, C dense matrix - * dense = dense * dense - */ -void testFuncDDDMatrix( - bool transa, bool transb, size_t dimM, size_t dimN, size_t dimK) { - real scaleT = 1.0; - size_t heightA = (transa == false) ? dimM : dimK; - size_t widthA = (transa == false) ? dimK : dimM; - size_t heightB = (transb == false) ? dimK : dimN; - size_t widthB = (transb == false) ? dimN : dimK; - size_t heightC = dimM; - size_t widthC = dimN; - // init Test object - CpuGpuFuncCompare test( - "MulOp", FuncConfig().set("aTrans", transa).set("bTrans", transb)); - // prepare input arguments - /// matrix A : HA * WA - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{heightA, widthA})); - /// matrix B: HB * WB - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{heightB, widthB})); - - /// output matrix C: HC * WC - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{heightC, widthC}), - scaleT == 1.0 ? ADD_TO : ASSIGN_TO); - // run Function - test.run(); -} - -TEST(MulOp, DDDMatrixMul) { - LOG(INFO) << "function test for dense = dense * dense matrix"; - for (const auto transa : {false, true}) { - for (const auto transb : {false, true}) { - for (const auto dimM : {1, 10, 100}) { - for (const auto dimN : {1, 10}) { - for (const auto dimK : {8}) { - if (transa && transb) { - continue; - } - VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') - << " transa=" << transa << " transb=" << transb - << " dimM=" << std::setw(5) << dimM - << " dimN=" << std::setw(5) << dimN - << " dimK=" << std::setw(5) << dimK; - testFuncDDDMatrix(transa, transb, dimM, dimN, dimK); - } - } - } - } - } -} - -/** - * C += A * B, B, C dense, A sparse - * dense = sparse * dense - */ -void testFuncDSparseDMatrix( - size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) { - real scaleT = 1.0; - // init Test object - CpuGpuFuncCompare test( - "MulOp", FuncConfig().set("aTrans", false).set("bTrans", false)); - // prepare input arguments - /// sparse matrix A : M * K - test.addInputs(SparseMatrixArg( - VALUE_TYPE_FLOAT, TensorShape{dimM, dimK}, nnz, FORMAT, FLOAT_VALUE)); - /// matrix B: K * N - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimK, dimN})); - - /// output matrix C: M * N - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimN}), - scaleT == 1.0 ? ADD_TO : ASSIGN_TO); - // run Function - test.run(); -} - -TEST(MuLOp, DSparseDMul) { - LOG(INFO) << "function test for dense = sparse * dense matrix"; - for (const auto dimM : {10, 100, 1000}) { - for (const auto dimN : {10, 100}) { - for (const auto dimK : {3, 10}) { - for (const auto nnz : {3, 10}) { - for (const auto FORMAT : {SPARSE_CSR}) { - VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') - << " dimM=" << std::setw(5) << dimM - << " dimN=" << std::setw(5) << dimN - << " dimK=" << std::setw(5) << dimK - << " nnz=" << std::setw(5) << nnz - << " format=" << std::setw(5) << FORMAT; - testFuncDSparseDMatrix(dimM, dimN, dimK, nnz, FORMAT); - } - } - } - } - } -} - -/** - * C += A * B, A, C dense, B sparse - * dense = dense * sparse - */ -void testFuncDDSparseMatrix( - size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) { - real scaleT = 1.0; - // init Test object - CpuGpuFuncCompare test( - "MulOp", FuncConfig().set("aTrans", false).set("bTrans", false)); - // prepare input arguments - /// matrix A : M * K - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimK})); - - /// matrix B: K * N - test.addInputs(SparseMatrixArg( - VALUE_TYPE_FLOAT, TensorShape{dimK, dimN}, nnz, FORMAT, FLOAT_VALUE)); - - /// output matrix C: M * N - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimN}), - scaleT == 1.0 ? ADD_TO : ASSIGN_TO); - // run Function - test.run(); -} - -TEST(MulOp, DDSparseMul) { - LOG(INFO) << "function test for dense = dense * sparse matrix"; - for (const auto dimM : {10, 100, 1000}) { - for (const auto dimN : {10, 100}) { - for (const auto dimK : {3, 10}) { - for (const auto nnz : {3, 10}) { - for (const auto FORMAT : {SPARSE_CSR, SPARSE_CSC}) { - VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') - << " dimM=" << std::setw(5) << dimM - << " dimN=" << std::setw(5) << dimN - << " dimK=" << std::setw(5) << dimK - << " nnz=" << std::setw(5) << nnz - << " format=" << std::setw(5) << FORMAT; - testFuncDDSparseMatrix(dimM, dimN, dimK, nnz, FORMAT); - } - } - } - } - } -} - -/** - * C += A * B, A sparse, B, C dense - * sparse = dense * dense - */ -void testFuncSparseDDMatrix( - size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) { - real scaleT = 1.0; - // init Test object - CpuGpuFuncCompare test( - "MulOp", FuncConfig().set("aTrans", false).set("bTrans", false)); - // prepare input arguments - /// matrix A : M * K - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimK})); - - /// matrix B: K * N - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimK, dimN})); - - /// output sparse matrix C: M * N - test.addOutputs( - SparseMatrixArg( - VALUE_TYPE_FLOAT, TensorShape{dimM, dimN}, nnz, FORMAT, FLOAT_VALUE), - scaleT == 1.0 ? ADD_TO : ASSIGN_TO); - // run Function - test.run(); -} - -TEST(MulOp, SparseDDMul) { - LOG(INFO) << "function test for sparse = dense * dense matrix"; - for (const auto dimM : {10, 100, 1000}) { - for (const auto dimN : {10, 100}) { - for (const auto dimK : {3, 10}) { - for (const auto nnz : {3, 10}) { - for (const auto FORMAT : {SPARSE_CSC, SPARSE_CSR}) { - VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') - << " dimM=" << std::setw(5) << dimM - << " dimN=" << std::setw(5) << dimN - << " dimK=" << std::setw(5) << dimK - << " nnz=" << std::setw(5) << nnz - << " format=" << std::setw(5) << FORMAT; - testFuncSparseDDMatrix(dimM, dimN, dimK, nnz, FORMAT); - } - } - } - } - } -} diff --git a/paddle/function/PadOp.cpp b/paddle/function/PadOp.cpp deleted file mode 100644 index 5d7515e8c053439b95fb18de3c8ffe70705600a3..0000000000000000000000000000000000000000 --- a/paddle/function/PadOp.cpp +++ /dev/null @@ -1,215 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PadOp.h" -#include "paddle/math/Vector.h" - -namespace paddle { - -template <> -void Pad(real* outputs, - const real* inputs, - const int num, - const int inC, - const int inH, - const int inW, - const PadConf& pad) { - int cstart = pad.channel[0], cend = pad.channel[1]; - int hstart = pad.height[0], hend = pad.height[1]; - int wstart = pad.width[0], wend = pad.width[1]; - int outC = inC + cstart + cend; - int outH = inH + hstart + hend; - int outW = inW + wstart + wend; - for (int i = 0; i < num; i++) { - for (int c = 0; c < inC; c++) { - for (int h = 0; h < inH; h++) { - int inoff = ((i * inC + c) * inH + h) * inW; - int outoff = - ((i * outC + c + cstart) * outH + h + hstart) * outW + wstart; - memcpy(outputs + outoff, inputs + inoff, inW * sizeof(real)); - } - } - } -} - -template <> -void PadGrad(real* inGrad, - const real* outGrad, - const int num, - const int inC, - const int inH, - const int inW, - const PadConf& pad) { - int cstart = pad.channel[0], cend = pad.channel[1]; - int hstart = pad.height[0], hend = pad.height[1]; - int wstart = pad.width[0], wend = pad.width[1]; - int outC = inC + cstart + cend; - int outH = inH + hstart + hend; - int outW = inW + wstart + wend; - for (int i = 0; i < num; i++) { - for (int c = 0; c < inC; c++) { - for (int h = 0; h < inH; h++) { - int inoff = ((i * inC + c) * inH + h) * inW; - int outoff = - ((i * outC + c + cstart) * outH + h + hstart) * outW + wstart; - CpuVector inG = CpuVector(inW, inGrad + inoff); - CpuVector outG = CpuVector(inW, const_cast(outGrad + outoff)); - inG += outG; - } - } - } -} - -static inline PadConf castToPadConf(const FuncConfig& conf) { - return {conf.get>("channel"), - conf.get>("height"), - conf.get>("width")}; -} - -/** - * \brief Padding zeros to input according to the specify dimension. - * The struct pad_ contains the padding size in each dimension. - * The input and output is a 4D tensor. In PadFunc, we only - * pad zeros to the 2nd to 4th dimension. - * - * Argument in this Function: - * \param pad_ A struct object contains the padding size in each dimension. - * It has six integers. The channelStart and channelEnd indicate - * how many zeros to add before and after the input in channel - * dimension. And the heightStart and heightEnd indicate padding - * in height dimension. The widthStart and widthEnd indicate the - * padding in width dimension. - * \param inputs A 4D tensor, only one input. - * \param outputs A 4D tensor, the output value after padding. - * - * For example, - * Input(2,2,2,3) = [ - * [ [[1,2,3], [3,4,5]], - * [[2,3,5], [1,6,7]] ], - * [ [[4,3,1], [1,8,7]], - * [[3,8,9], [2,3,5]] ] - * ] # the shape is (1,2,2,3) - * - * pad_: if channelStart = channelEnd = 1, others are 0. - * Output(2,4,2,3) = [ - * [ [[0,0,0], [0,0,0]], - * [[1,2,3], [3,4,5]], - * [[2,3,5], [1,6,7]], - * [[0,0,0], [0,0,0]] ], - * [ [[0,0,0], [0,0,0]], - * [[4,3,1], [1,8,7]], - * [[3,8,9], [2,3,5]], - * [[0,0,0], [0,0,0]] ] - * ] # the shape is (2,4,2,3) - * - * pad_: if widthStart = 1, widthEnd = 2, others are 0. - * Output(2,2,2,6) = [ - * [ [[0,1,2,3,0,0], [0,3,4,5,0,0]], - * [[0,2,3,5,0,0], [0,1,6,7,0,0]] ], - * [ [[0,4,3,1,0,0], [0,1,8,7,0,0]], - * [[0,3,8,9,0,0], [0,2,3,5,0,0]] ], - * ] # the shape is (2,2,2,6) - * - * pad_: if heightStart = 1, heightEnd = 1, others are 0. - * Output(2,2,4,3) = [ - * [ [[0,0,0], [1,2,3], [3,4,5], [0,0,0]], - * [[0,0,0], [2,3,5], [1,6,7], [0,0,0]] ], - * [ [[0,0,0], [4,3,1], [1,8,7], [0,0,0]], - * [[0,0,0], [3,8,9], [2,3,5], [0,0,0]] ], - * ] # the shape is (2,2,4,3) - */ - -template -class PadFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { pad_ = castToPadConf(config); } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - - size_t num = inputs[0].shape()[0]; - size_t inC = inputs[0].shape()[1]; - size_t inH = inputs[0].shape()[2]; - size_t inW = inputs[0].shape()[3]; - typename Tensor::Vector vec(outputs[0].shape().getElements(), - outputs[0].data()); - vec.zero(); - - Pad(outputs[0].data(), - inputs[0].data(), - num, - inC, - inH, - inW, - pad_); - } - - private: - PadConf pad_; -}; - -/** - * \brief The backward propagation of padding Function. Remove the elements - * in the padding positions of forward. - * - * Argument in this Function: - * \param pad_ The same meaning as it in PadFunc. - * \param inputs The gradient with respect to the output value of PadFunc. - * \param outputs The gradient with respect to the input value of PadFunc. - */ - -template -class PadGradFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { pad_ = castToPadConf(config); } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - - size_t num = outputs[0].shape()[0]; - size_t inC = outputs[0].shape()[1]; - size_t inH = outputs[0].shape()[2]; - size_t inW = outputs[0].shape()[3]; - - if (outputs[0].getArgType() != ADD_TO) { - // for unit test - typename Tensor::Vector tmp( - outputs[0].shape().getElements(), outputs[0].data()); - tmp.zero(); - } - - PadGrad(outputs[0].data(), - inputs[0].data(), - num, - inC, - inH, - inW, - pad_); - } - - private: - PadConf pad_; -}; - -REGISTER_TYPED_FUNC(Pad, CPU, PadFunc); -REGISTER_TYPED_FUNC(PadGrad, CPU, PadGradFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(Pad, GPU, PadFunc); -REGISTER_TYPED_FUNC(PadGrad, GPU, PadGradFunc); -#endif - -} // namespace paddle diff --git a/paddle/function/RowConvOp.cpp b/paddle/function/RowConvOp.cpp deleted file mode 100644 index 129e9334582fad011c259e8ab8268b00a7fab7b6..0000000000000000000000000000000000000000 --- a/paddle/function/RowConvOp.cpp +++ /dev/null @@ -1,225 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "RowConvOp.h" -#include -#include "paddle/math/Vector.h" - -namespace paddle { - -template <> -void RowConv(CpuMatrix& out, - const CpuMatrix& in, - const CpuMatrix& filter, - const CpuIVector& seq) { - const int* starts = seq.getData(); - const size_t numSeq = seq.getSize() - 1; - const size_t contextLength = filter.getHeight(); - for (size_t i = 0; i < numSeq; ++i) { - size_t begin = starts[i]; - size_t end = starts[i + 1]; - for (size_t j = begin; j < end; ++j) { - MatrixPtr x; - MatrixPtr w; - if ((j + contextLength) < end) { - x = (const_cast(in)).subMatrix(j, contextLength); - w = (const_cast(filter)).subMatrix(0, contextLength); - } else { - x = (const_cast(in)).subMatrix(j, end - j); - w = (const_cast(filter)).subMatrix(0, end - j); - } - MatrixPtr y = out.subMatrix(j, 1); - y->addDotMulVMM(*x, *w); - } - } -} - -template <> -void RowConvGrad(const CpuMatrix& outG, - const CpuMatrix& in, - const CpuMatrix& filter, - CpuMatrix& inG, - CpuMatrix& filterG, - const CpuIVector& seq) { - // gradient w.r.t filter - const int* starts = seq.getData(); - const size_t numSeq = seq.getSize() - 1; - const size_t contextLength = filter.getHeight(); - if (filterG) { - for (size_t i = 0; i < numSeq; ++i) { - size_t begin = starts[i]; - size_t end = starts[i + 1]; - size_t steps = end - begin; - for (size_t j = 0; j < contextLength && (begin + j) < end; ++j) { - MatrixPtr x = - (const_cast(in)).subMatrix(begin + j, steps - j); - MatrixPtr dy = - (const_cast(outG)).subMatrix(begin, steps - j); - MatrixPtr dw = filterG.subMatrix(j, 1); - dw->addDotMulVMM(*dy, *x); - } - } - } - - // gradient w.r.t input feature - if (inG) { - for (size_t i = 0; i < numSeq; ++i) { - size_t begin = starts[i]; - size_t end = starts[i + 1]; - size_t steps = end - begin; - for (size_t j = 0; j < steps; ++j) { - MatrixPtr dx = inG.subMatrix(begin + j, 1); - for (size_t t = 0; t < contextLength; ++t) { - if (int(j - t) >= 0) { - MatrixPtr dy = - (const_cast(outG)).subMatrix(begin + j - t, 1); - MatrixPtr w = (const_cast(filter)).subMatrix(t, 1); - dx->addDotMul(*dy, *w, 1.0, 1.0); - } - } - } - } - } -} - -/** - * \brief The row convolution is called lookahead convolution. It is firstly - * introduced in deep-speech2 system. The bidirectional RNN that learns - * representation for a sequence by performing a forward and a backward pass - * through the entire sequence. However, unlike unidirectional RNNs, - * bidirectional RNNs are challenging to deploy in an online and low-latency - * setting. The lookahead convolution incorporates information from future - * subsequences in a computationally efficient manner to improve unidirectional - * recurrent neural networks. - * - * The connection of row convolution is different form the 1D sequence - * convolution. Assumed that, the future context-length is k, that is to say, - * it can get the output at timestep t by using the the input feature from t-th - * timestep to (t+k)-th timestep. Assumed that the hidden dim of input - * activations are d, the activations r_t for the new layer at time-step t are: - * - * - * -- k + 1 - * r(t,i) = > W(i,j) * h(t+j-1, i), for (1 <= i <= d) - * -- j = 1 - * - * - * The weight shape is: (k + 1) x d - * Function Arguments: - * - * \param inputs[0] The input activations. - * \param inputs[0] The filter (or weight) and shape is (k+1) x d. - * \param outputs[1] The output activations. - * - * [1] Dario Amodei, etc. Deep Speech 2 : End-to-End Speech Recognition in - * English - * and Mandarin. https://arxiv.org/abs/1512.02595 - */ - -template -class RowConvFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override {} - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - // check - CHECK_EQ(2UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - // TODO(qingqing): support ASSIGN_TO. - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) - << "SequenceArg required here."; - const auto in = dynamic_cast(inputs[0]); - auto out = dynamic_cast(outputs[0]); - auto w = inputs[1]; - CHECK(in.data() && out.data() && in.getSequenceId().data()); - CHECK_EQ(in.shape().ndims(), 2UL); - CHECK(in.shape() == out.shape()); - CHECK_EQ(w.shape()[1], in.shape()[1]); - - auto outMat = out.matrix(); - const auto inMat = in.matrix(); - const auto wMat = w.matrix(); - const auto seqId = in.getSequenceId().vector(); - - RowConv(outMat, inMat, wMat, seqId); - } -}; - -/** - * \brief The backward of row convolution function. This function calculated - * the gradient w.r.t filter and the gradient w.r.t input activations(or data). - * - * Argument in this Function: - * - * \param inputs[0] The gradient w.r.t output activations. - * \param inputs[1] The input activations. - * \param inputs[2] The filter (or weight) and shape is (k+1) x d. - * \param outputs[0] The gradient w.r.t input activations. - * \param outputs[1] The gradient w.r.r filter. - * - * Abbreviation: - * w.r.t: with respect to. - */ - -template -class RowConvGradFunc : public FunctionBase { - // TODO(qingqing): split into RowConvDataFunc and RowConvWeightFunc - public: - void init(const FuncConfig& config) override {} - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - // check - CHECK_EQ(3UL, inputs.size()); - CHECK_EQ(2UL, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - CHECK_EQ(outputs[1].getArgType(), ADD_TO); - CHECK(inputs[0].isSequenceArg() && inputs[1].isSequenceArg() && - outputs[0].isSequenceArg()) - << "SequenceArg required here."; - - const auto outGrad = dynamic_cast(inputs[0]); - const auto in = dynamic_cast(inputs[1]); - const auto w = inputs[2]; - auto inGrad = dynamic_cast(outputs[0]); - auto wGrad = outputs[1]; - - CHECK_EQ(in.shape().ndims(), 2UL); - CHECK(in.shape() == inGrad.shape()); - CHECK(in.shape() == outGrad.shape()); - CHECK_EQ(wGrad.shape()[1], in.shape()[1]); - - const auto outGMat = outGrad.matrix(); - const auto inMat = in.matrix(); - const auto wMat = w.matrix(); - auto inGMat = inGrad.data() - ? inGrad.matrix() - : typename Tensor::Matrix(nullptr, 0, 0); - auto wGMat = wGrad.data() - ? wGrad.matrix() - : typename Tensor::Matrix(nullptr, 0, 0); - const auto seqId = in.getSequenceId().vector(); - - RowConvGrad(outGMat, inMat, wMat, inGMat, wGMat, seqId); - } -}; - -REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc); -REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc); -REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc); -#endif - -} // namespace paddle diff --git a/paddle/function/RowConvOpGpu.cu b/paddle/function/RowConvOpGpu.cu deleted file mode 100644 index f820ee9a9713ce17547aa03945dc3c291ef50a59..0000000000000000000000000000000000000000 --- a/paddle/function/RowConvOpGpu.cu +++ /dev/null @@ -1,373 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/cuda/include/hl_base.h" -#include "paddle/function/RowConvOp.h" - -namespace paddle { - -template -__global__ void KeRowConv(real* y, - const real* x, - const real* w, - const int* starts, - const int height, - const int width, - const int numSeq, - const int context) { - const int tidx = threadIdx.x; - const int tidy = threadIdx.y; - const int blky = blockDim.y; - const int gidx = blockIdx.x * blockDim.x; - - __shared__ real sw[BLOCK_H][BLOCK_W]; - - for (int i = tidy; i < context; i += blky) { - sw[i][tidx] = gidx + tidx < width ? w[i * width + gidx + tidx] : 0.0; - } - - __syncthreads(); - - for (int i = 0; i < numSeq; ++i) { - const int start = starts[i]; - const int end = starts[i + 1]; - const int steps = end - start; - for (int j = tidy; j < steps; j += blky) { - real sum = 0; - int off = (start + j) * width; - for (int t = 0; t < context; ++t) { - if ((start + j + t) < end) { - int xoff = off + t * width; - real xVal = gidx + tidx < width ? x[xoff + gidx + tidx] : 0.0; - sum += sw[t][tidx] * xVal; - } - } - if (gidx + tidx < width) { - y[off + gidx + tidx] += sum; - } - } - } -} - -__global__ void KeRowConv2(real* y, - const real* x, - const real* w, - const int* starts, - const int height, - const int width, - const int numSeq, - const int context) { - const int tidx = threadIdx.x; - const int tidy = threadIdx.y; - const int blky = blockDim.y; - const int gidx = blockIdx.x * blockDim.x; - - for (int i = 0; i < numSeq; ++i) { - const int start = starts[i]; - const int end = starts[i + 1]; - const int steps = end - start; - for (int j = tidy; j < steps; j += blky) { - int off = (start + j) * width; - real sum = 0; - for (int t = 0; t < context && (start + j + t) < end; ++t) { - int xoff = off + t * width; - real xd = gidx + tidx < width ? x[xoff + gidx + tidx] : 0.0; - real wd = gidx + tidx < width ? w[t * width + gidx + tidx] : 0.0; - sum += wd * xd; - } - if (gidx + tidx < width) { - y[off + gidx + tidx] += sum; - } - } - } -} - -template <> -void RowConv(GpuMatrix& out, // NOLINT - const GpuMatrix& in, - const GpuMatrix& filter, - const GpuIVector& seq) { - const size_t numSeq = seq.getSize() - 1; - const size_t contextLength = filter.getHeight(); - const size_t height = in.getHeight(); - const size_t width = in.getWidth(); - - real* y = out.getData(); - const real* x = in.getData(); - const real* w = filter.getData(); - const int* starts = seq.getData(); - - dim3 dimBlock(32, 32); - dim3 dimGrid(DIVUP(width, dimBlock.x), 1); - - if (contextLength <= 32) { - KeRowConv<32, 32><<>>( - y, x, w, starts, height, width, numSeq, contextLength); - } else { - KeRowConv2<<>>( - y, x, w, starts, height, width, numSeq, contextLength); - } - CHECK_SYNC("RowConv"); -} - -template -__global__ void KeRowConvBwWeight(real* dw, - const real* x, - const real* dy, - const int* starts, - const int height, - const int width, - const int numSeq, - const int context) { - const int tidx = threadIdx.x; - const int tidy = threadIdx.y; - const int blky = blockDim.y; - const int gidx = blockIdx.x * blockDim.x; - - __shared__ real sh_x[BLOCK_W][BLOCK_H]; - __shared__ real sh_dy[BLOCK_W][BLOCK_H + CONTEXT - 1]; - __shared__ real sh_dw[CONTEXT][BLOCK_W]; - - if (tidy < context) { - sh_dw[tidy][tidx] = 0.0; - } - __syncthreads(); - - // NOTE(zcd): temporary solution - unsigned mask = 0u; - CREATE_SHFL_MASK(mask, true); - - for (int i = 0; i < numSeq; ++i) { - const int start = starts[i]; - const int end = starts[i + 1]; - const int steps = end - start; - const int size = ((steps + BLOCK_H - 1) / BLOCK_H) * BLOCK_H; - for (int j = tidy; j < size; j += BLOCK_H) { - int xoff = gidx + tidx; - int yoff = start + j; - - // transpose - sh_x[tidx][tidy] = - (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0; - sh_dy[tidx][tidy + context - 1] = - (xoff < width && yoff < end) ? dy[yoff * width + xoff] : 0.0; - __syncthreads(); - if (tidy < (context - 1)) { - yoff = yoff - context + 1; - sh_dy[tidx][tidy] = - (xoff < width && yoff >= start) ? dy[yoff * width + xoff] : 0.0; - } - __syncthreads(); - - for (int t = 0; t < context; t++) { - real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx + context - 1 - t]; - __syncthreads(); - // warp size and blockDim.x is 32. - - for (int offset = 16; offset > 0; offset /= 2) - val += __shfl_down_sync(mask, val, offset); - - __syncthreads(); - if (tidx == 0) { - sh_dw[t][tidy] += val; - } - __syncthreads(); - } - } - } - - for (int t = tidy; (t < context) && ((gidx + tidx) < width); t += blky) { - dw[t * width + gidx + tidx] += sh_dw[t][tidx]; - } -} - -template -__global__ void KeRowConvBwWeight2(real* dw, - const real* x, - const real* dy, - const int* starts, - const int height, - const int width, - const int numSeq, - const int context) { - const int tidx = threadIdx.x; - const int tidy = threadIdx.y; - const int gidx = blockIdx.x * blockDim.x; - - __shared__ real sh_x[BLOCK_H][BLOCK_W]; - __shared__ real sh_dy[BLOCK_H][BLOCK_W]; - - // NOTE(zcd): temporary solution - unsigned mask = 0u; - CREATE_SHFL_MASK(mask, true); - - for (int i = 0; i < numSeq; ++i) { - const int start = starts[i]; - const int end = starts[i + 1]; - const int steps = end - start; - - const int size = ((steps + BLOCK_H - 1) / BLOCK_H) * BLOCK_H; - for (int j = tidy; j < size; j += BLOCK_H) { - int xoff = gidx + tidx; - int yoff = start + j; - - // transpose - sh_x[tidx][tidy] = - (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0; - __syncthreads(); - - for (int t = 0; t < context; t++) { - sh_dy[tidx][tidy] = - (xoff < width && (yoff - t) >= start && yoff - t < end) - ? dy[(yoff - t) * width + xoff] - : 0.0; - __syncthreads(); - - real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx]; - __syncthreads(); - // warp size and blockDim.x is 32. - for (int offset = 16; offset > 0; offset /= 2) - val += __shfl_down_sync(mask, val, offset); - - __syncthreads(); - - if (tidx == 0 && (gidx + tidy) < width) { - dw[t * width + gidx + tidy] += val; - } - } - } - } -} - -template -__global__ void KeRowConvBwData(real* dx, - const real* w, - const real* dy, - const int* starts, - const int height, - const int width, - const int numSeq, - const int context) { - const int tidx = threadIdx.x; - const int tidy = threadIdx.y; - const int blky = blockDim.y; - const int gidx = blockIdx.x * blockDim.x; - - __shared__ real sw[BLOCK_H][BLOCK_W]; - - for (int i = tidy; i < context; i += blky) { - sw[i][tidx] = gidx + tidx < width ? w[i * width + gidx + tidx] : 0.0; - } - - __syncthreads(); - - for (int i = 0; i < numSeq; ++i) { - const int start = starts[i]; - const int end = starts[i + 1]; - const int steps = end - start; - for (int j = tidy; j < steps; j += blky) { - real sum = 0; - int off = (start + j) * width; - for (int t = 0; t < context && (j - t) >= 0; ++t) { - int dyOff = off - t * width; - real dyVal = gidx + tidx < width ? dy[dyOff + gidx + tidx] : 0.0; - sum += sw[t][tidx] * dyVal; - } - if (gidx + tidx < width) { - dx[off + gidx + tidx] += sum; - } - } - } -} - -__global__ void KeRowConvBwData2(real* dx, - const real* w, - const real* dy, - const int* starts, - const int height, - const int width, - const int numSeq, - const int context) { - const int tidx = threadIdx.x; - const int tidy = threadIdx.y; - const int blky = blockDim.y; - const int gidx = blockIdx.x * blockDim.x; - - for (int i = 0; i < numSeq; ++i) { - const int start = starts[i]; - const int end = starts[i + 1]; - const int steps = end - start; - for (int j = tidy; j < steps; j += blky) { - real sum = 0; - int off = (start + j) * width; - for (int t = 0; t < context && (j - t) >= 0; ++t) { - int dyOff = off - t * width; - real dyVal = gidx + tidx < width ? dy[dyOff + gidx + tidx] : 0.0; - real wVal = gidx + tidx < width ? w[t * width + gidx + tidx] : 0.0; - sum += wVal * dyVal; - } - if (gidx + tidx < width) { - dx[off + gidx + tidx] += sum; - } - } - } -} - -template <> -void RowConvGrad(const GpuMatrix& outG, - const GpuMatrix& in, - const GpuMatrix& filter, - GpuMatrix& inG, // NOLINT - GpuMatrix& filterG, // NOLINT - const GpuIVector& seq) { - const size_t numSeq = seq.getSize() - 1; - const size_t contextLength = filter.getHeight(); - const size_t height = in.getHeight(); - const size_t width = in.getWidth(); - - const real* dy = outG.getData(); - const real* x = in.getData(); - const real* w = filter.getData(); - const int* starts = seq.getData(); - - if (filterG) { - dim3 dimBlock(32, 32); - dim3 dimGrid(DIVUP(width, dimBlock.x), 1); - real* dw = filterG.getData(); - if (contextLength <= 32) { - KeRowConvBwWeight<32, 32, 32><<>>( - dw, x, dy, starts, height, width, numSeq, contextLength); - } else { - KeRowConvBwWeight2<32, 32><<>>( - dw, x, dy, starts, height, width, numSeq, contextLength); - } - } - - if (inG) { - real* dx = inG.getData(); - dim3 dimBlock2(32, 32); - dim3 dimGrid2(DIVUP(width, dimBlock2.x), 1); - if (contextLength <= 64) { - KeRowConvBwData<32, 64><<>>( - dx, w, dy, starts, height, width, numSeq, contextLength); - } else { - KeRowConvBwData2<<>>( - dx, w, dy, starts, height, width, numSeq, contextLength); - } - } - - CHECK_SYNC("RowConvGrad"); -} - -} // namespace paddle diff --git a/paddle/function/ScaleSubRegionOp.cpp b/paddle/function/ScaleSubRegionOp.cpp deleted file mode 100644 index 9a06ef2a96f25b5b7326049df2a708637f319561..0000000000000000000000000000000000000000 --- a/paddle/function/ScaleSubRegionOp.cpp +++ /dev/null @@ -1,155 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ScaleSubRegionOp.h" -#include "paddle/function/TensorShape.h" - -namespace paddle { - -template <> -void ScaleSubRegion(real* outputs, - const real* inputs, - const real* indices, - const TensorShape shape, - const FuncConfig& conf) { - real value = conf.get("value"); - - int number = shape[0]; - int channel = shape[1]; - int height = shape[2]; - int width = shape[3]; - - memcpy(outputs, inputs, number * channel * height * width * sizeof(real)); - - for (int n = 0; n < number; ++n) { - // indices start from 1 - int offset = n * 6; - for (int c = indices[offset] - 1; c < indices[offset + 1]; ++c) { - for (int h = indices[offset + 2] - 1; h < indices[offset + 3]; ++h) { - for (int w = indices[offset + 4] - 1; w < indices[offset + 5]; ++w) { - int idx = ((n * channel + c) * height + h) * width + w; - outputs[idx] *= value; - } - } - } - } -} - -template <> -void ScaleSubRegionGrad(const real* inGrad, - real* outGrad, - const real* indices, - const TensorShape shape, - const FuncConfig& conf) { - real value = conf.get("value"); - - int number = shape[0]; - int channel = shape[1]; - int height = shape[2]; - int width = shape[3]; - - for (int n = 0; n < number; ++n) { - for (int c = 0; c < channel; ++c) { - for (int h = 0; h < height; ++h) { - for (int w = 0; w < width; ++w) { - int idx = ((n * channel + c) * height + h) * width + w; - int offset = n * 6; - if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) && - h >= (indices[offset + 2] - 1) && - h <= (indices[offset + 3] - 1) && - w >= (indices[offset + 4] - 1) && - w <= (indices[offset + 5] - 1)) { - outGrad[idx] += inGrad[idx] * value; - } else { - outGrad[idx] += inGrad[idx]; - } - } - } - } - } -} - -/** - * \brief For each instance, ScaleSubRegion can be used to multiply a value to - * a specified sub continuous region. By providing start index and end - * index for C/H/W, you can specify the location and shape of the region. - * - * Argument in this Function: - * \param inputs A 4-D tensor with shape [N, C, H, W], only one input. - * \param indices A 2-D tensor with shape [N, 6], indicates the sub region. - * \param outputs A 4-D tensor with same shape as inputs, output value. - */ -template -class ScaleSubRegionFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { conf_ = config; } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(2UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - - TensorShape shape = inputs[0].shape(); - - ScaleSubRegion(outputs[0].data(), - inputs[0].data(), - inputs[1].data(), - shape, - conf_); - } - - private: - FuncConfig conf_; -}; - -/** - * \brief The backward propagation of ScaleSubRegion Function. - * - * Argument in this Function: - * \param inputs A 4-D tensor with shape [N, C, H, W], output gradient. - * \param indices A 2-D tensor with shape [N, 6], indicates the sub region. - * \param outputs A 4-D tensor with shape [N, C, H, W], gradient of input value. - */ - -template -class ScaleSubRegionGradFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override { conf_ = config; } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(2UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ADD_TO); - - TensorShape shape = inputs[0].shape(); - - ScaleSubRegionGrad(inputs[0].data(), - outputs[0].data(), - inputs[1].data(), - shape, - conf_); - } - - private: - FuncConfig conf_; -}; - -REGISTER_TYPED_FUNC(ScaleSubRegion, CPU, ScaleSubRegionFunc); -REGISTER_TYPED_FUNC(ScaleSubRegionGrad, CPU, ScaleSubRegionGradFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(ScaleSubRegion, GPU, ScaleSubRegionFunc); -REGISTER_TYPED_FUNC(ScaleSubRegionGrad, GPU, ScaleSubRegionGradFunc); -#endif - -} // namespace paddle diff --git a/paddle/function/SwitchOp.cpp b/paddle/function/SwitchOp.cpp deleted file mode 100644 index 750fb6bf28baf050b1f9f965a1a9b315363e5645..0000000000000000000000000000000000000000 --- a/paddle/function/SwitchOp.cpp +++ /dev/null @@ -1,140 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SwitchOp.h" -#include "paddle/math/Vector.h" - -namespace paddle { - -template <> -void NCHW2NHWC(real* outputs, - const real* inputs, - const int num, - const int inC, - const int inH, - const int inW, - const int argType) { - for (int n = 0; n < num; ++n) { - for (int c = 0; c < inC; ++c) { - for (int h = 0; h < inH; ++h) { - for (int w = 0; w < inW; ++w) { - if (argType == ADD_TO) { - outputs[((n * inH + h) * inW + w) * inC + c] += *(inputs++); - } else { - outputs[((n * inH + h) * inW + w) * inC + c] = *(inputs++); - } - } - } - } - } -} - -template <> -void NHWC2NCHW(real* outputs, - const real* inputs, - const int num, - const int inH, - const int inW, - const int inC, - const int argType) { - for (int n = 0; n < num; ++n) { - for (int h = 0; h < inH; ++h) { - for (int w = 0; w < inW; ++w) { - for (int c = 0; c < inC; ++c) { - if (argType == ADD_TO) { - outputs[((n * inC + c) * inH + h) * inW + w] += *(inputs++); - } else { - outputs[((n * inC + c) * inH + h) * inW + w] = *(inputs++); - } - } - } - } - } -} - -/** - * \brief Switch dimension order of image input. - * The input and output is a 4D tensor. Switch order - * 'batch_size,channels, height, width' to - * order 'batch_size, height, width, channels'. - * - * Argument in this Function: - * \param inputs input data with order 'batch_size,channels, height, width'. - * \param outputs output data with order 'batch_size, height, width, channels'. - */ -template -class NCHW2NHWCFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override {} - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - - size_t num = inputs[0].shape()[0]; - size_t inC = inputs[0].shape()[1]; - size_t inH = inputs[0].shape()[2]; - size_t inW = inputs[0].shape()[3]; - NCHW2NHWC(outputs[0].data(), - inputs[0].data(), - num, - inC, - inH, - inW, - outputs[0].getArgType()); - } -}; - -/** - * \brief Switch dimension order of image input. - * The input and output is a 4D tensor. Switch order - * 'batch_size, height, width, channels' to - * order 'batch_size, channels, height, width'. - * - * Argument in this Function: - * \param inputs input data with order 'batch_size, height, width, channels'. - * \param outputs output data with order 'batch_size, channels, height, width'. - */ -template -class NHWC2NCHWFunc : public FunctionBase { - public: - void init(const FuncConfig& config) override {} - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(1UL, inputs.size()); - CHECK_EQ(1UL, outputs.size()); - - size_t num = inputs[0].shape()[0]; - size_t inH = inputs[0].shape()[1]; - size_t inW = inputs[0].shape()[2]; - size_t inC = inputs[0].shape()[3]; - - NHWC2NCHW(outputs[0].data(), - inputs[0].data(), - num, - inH, - inW, - inC, - outputs[0].getArgType()); - } -}; - -REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc); -REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc); -#ifdef PADDLE_WITH_CUDA -REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc); -REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc); -#endif - -} // namespace paddle diff --git a/paddle/function/TensorType.h b/paddle/function/TensorType.h deleted file mode 100644 index b384591bd8852bbdc61bf9aa678ce613732c369a..0000000000000000000000000000000000000000 --- a/paddle/function/TensorType.h +++ /dev/null @@ -1,149 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/math/Matrix.h" - -namespace paddle { - -enum ValueType { - VALUE_TYPE_INT32 = 0, - VALUE_TYPE_FLOAT = 1, - VALUE_TYPE_DOUBLE = 2, - VALUE_TYPE_BYTE = 3 -}; - -enum DeviceType { - DEVICE_TYPE_UNSPECIFIED = 0, - DEVICE_TYPE_CPU = 1, - DEVICE_TYPE_GPU = 2 -}; - -enum SparseDataType { T_NO_VALUE = 0, T_FLOAT_VALUE = 1 }; - -enum SparseDataFormat { T_SPARSE_CSR = 0, T_SPARSE_CSC = 1 }; - -inline int sizeOfValuType(ValueType valueType) { - if (valueType == VALUE_TYPE_INT32) { - return 4; - } else if (valueType == VALUE_TYPE_FLOAT) { - return 4; - } else if (valueType == VALUE_TYPE_DOUBLE) { - return 8; - } else { - LOG(FATAL) << "Unknown type: " << valueType; - return 0; - } -} - -template -struct DataType; - -template <> -struct DataType { - static const ValueType value = VALUE_TYPE_FLOAT; -}; - -template <> -struct DataType { - static const ValueType value = VALUE_TYPE_DOUBLE; -}; - -template <> -struct DataType { - static const ValueType value = VALUE_TYPE_INT32; -}; - -namespace detail { - -template -struct MatrixT; - -template <> -struct MatrixT { - using type = CpuMatrix; -}; - -template <> -struct MatrixT { - using type = GpuMatrix; -}; - -template <> -struct MatrixT { - using type = void; // Not implemented -}; - -template <> -struct MatrixT { - using type = void; // Not implemented -}; - -template -struct SparseMatrixT; - -template <> -struct SparseMatrixT { - using type = CpuSparseMatrix; -}; - -template <> -struct SparseMatrixT { - using type = GpuSparseMatrix; -}; - -template <> -struct SparseMatrixT { - using type = void; // Not implemented -}; - -template <> -struct SparseMatrixT { - using type = void; // Not implemented -}; - -template -struct VectorT; - -template <> -struct VectorT { - using type = CpuVector; -}; - -template <> -struct VectorT { - using type = GpuVector; -}; - -template <> -struct VectorT { - using type = CpuIVector; -}; - -template <> -struct VectorT { - using type = GpuIVector; -}; - -} // namespace detail - -template -struct Tensor { - typedef typename detail::VectorT::type Vector; - typedef typename detail::MatrixT::type Matrix; - typedef typename detail::SparseMatrixT::type SparseMatrix; -}; - -} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp deleted file mode 100644 index d7ac83da41aaba5cd38b042d0381dea527f9c42d..0000000000000000000000000000000000000000 --- a/paddle/function/neon/NeonDepthwiseConv.cpp +++ /dev/null @@ -1,120 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "NeonDepthwiseConv.h" -#include "paddle/function/ConvOp.h" - -namespace paddle { - -#if defined(__ARM_NEON__) || defined(__ARM_NEON) - -template -class NeonDepthwiseConvFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - check(inputs, outputs); - - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - - int batchSize = input[0]; - int inputChannels = input[1]; - int inputHeight = input[2]; - int inputWidth = input[3]; - int filterHeight = getFilterHeight(filter); - int filterWidth = getFilterWidth(filter); - int outputChannels = output[1]; - int outputHeight = output[2]; - int outputWidth = output[3]; - int filterMultiplier = outputChannels / groups_; - CHECK_EQ(static_cast(inputChannels), groups_); - - // only support strideH() == strideW() and filterHeight == filterWidth. - CHECK_EQ(strideH(), strideW()); - CHECK_EQ(filterHeight, filterWidth); - - float* inputData = inputs[0].data(); - float* filterData = inputs[1].data(); - float* outputData = outputs[0].data(); - - // padding the input - float* inputPadding = inputData; - int padInputHeight = inputHeight + 2 * paddingH(); - int padInputWidth = inputWidth + 2 * paddingW(); - int newSize = - batchSize * (inputChannels + 1) * padInputHeight * padInputWidth; - - resizeBuffer(newSize); - inputPadding = reinterpret_cast(memory_->getBuf()); - neon::Padding::run(inputData, - inputPadding, - batchSize * inputChannels, - inputHeight, - inputWidth, - padInputHeight, - padInputWidth); - - std::function - DepthWiseConv; - - if (filterWidth == 3 && strideW() == 1) { - DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; - } else if (filterWidth == 3 && strideW() == 2) { - DepthWiseConv = neon::DepthwiseConvKernel<3, 2>::run; - } else if (filterWidth == 4 && strideW() == 1) { - DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; - } else if (filterWidth == 4 && strideW() == 2) { - DepthWiseConv = neon::DepthwiseConvKernel<4, 2>::run; - } else { - LOG(FATAL) << "Not supported"; - } - - for (int i = 0; i < batchSize; i++) { - DepthWiseConv(inputPadding, - filterData, - padInputHeight, - padInputWidth, - outputChannels, - outputHeight, - outputWidth, - filterMultiplier, - outputData); - inputPadding += inputChannels * padInputHeight * padInputWidth; - outputData += outputChannels * outputHeight * outputWidth; - } - } -}; - -#ifndef PADDLE_TYPE_DOUBLE -REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction); -#endif - -#endif - -} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConvTranspose.cpp b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp deleted file mode 100644 index 1fc5daf6078bbd5b4506ff2e0832e2cc3ec48fe3..0000000000000000000000000000000000000000 --- a/paddle/function/neon/NeonDepthwiseConvTranspose.cpp +++ /dev/null @@ -1,136 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "NeonDepthwiseConv.h" -#include "paddle/function/ConvOp.h" - -namespace paddle { - -#if defined(__ARM_NEON__) || defined(__ARM_NEON) - -template -class NeonDepthwiseConvTransposeFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - check(inputs, outputs); - - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - - int batchSize = input[0]; - int inputChannels = input[1]; - int inputHeight = input[2]; - int inputWidth = input[3]; - int filterHeight = getFilterHeight(filter); - int filterWidth = getFilterWidth(filter); - int outputChannels = output[1]; - int outputHeight = output[2]; - int outputWidth = output[3]; - int filterMultiplier = outputChannels / groups_; - CHECK_EQ(inputChannels, groups_); - - // only support strideH() == strideW() and filterHeight == filterWidth. - CHECK_EQ(strideH(), strideW()); - CHECK_EQ(paddingH(), paddingW()); - CHECK_EQ(filterHeight, filterWidth); - - float* inputData = inputs[0].data(); - float* filterData = inputs[1].data(); - float* outputData = outputs[0].data(); - - // padding the input, input -> inputPadding - float* inputPadding = inputData; - int padInputHeight = - (inputHeight - 1) * strideH() + 2 * filterHeight - 1 - 2 * paddingH(); - int padInputWidth = - (inputWidth - 1) * strideW() + 2 * filterWidth - 1 - 2 * paddingW(); - - if (padInputHeight > inputHeight || padInputWidth > inputWidth) { - int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; - resizeBuffer(newSize); - inputPadding = reinterpret_cast(memory_->getBuf()); - if (strideH() == 1) { - neon::Padding::run(inputData, - inputPadding, - batchSize * inputChannels, - inputHeight, - inputWidth, - padInputHeight, - padInputWidth); - } else if (strideH() == 2) { - neon::StridePadding::run(inputData, - inputPadding, - batchSize * inputChannels, - inputHeight, - inputWidth, - padInputHeight, - padInputWidth); - } else { - LOG(FATAL) << "Not supported"; - } - } - - std::function - DepthWiseConv; - - if (filterWidth == 3) { - DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; - } else if (filterWidth == 4) { - DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; - } else { - LOG(FATAL) << "Not supported"; - } - - for (int i = 0; i < batchSize; i++) { - DepthWiseConv(inputPadding, - filterData, - padInputHeight, - padInputWidth, - outputChannels, - outputHeight, - outputWidth, - filterMultiplier, - outputData); - inputPadding += inputChannels * padInputHeight * padInputWidth; - outputData += outputChannels * outputHeight * outputWidth; - } - } -}; - -#ifndef PADDLE_TYPE_DOUBLE - -REGISTER_TYPED_FUNC(NeonDepthwiseConvTranspose, - CPU, - NeonDepthwiseConvTransposeFunction); - -#endif - -#endif - -} // namespace paddle diff --git a/paddle/function/nnpack/NNPACKConvOp.cpp b/paddle/function/nnpack/NNPACKConvOp.cpp deleted file mode 100644 index 48c997b50d8c73b25c58801c30e597c9d1f3232a..0000000000000000000000000000000000000000 --- a/paddle/function/nnpack/NNPACKConvOp.cpp +++ /dev/null @@ -1,247 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "nnpack.h" -#include "paddle/function/ConvOp.h" - -DEFINE_bool(nnpack_allocate_outside, - true, - "Allocate and free workspace memory outside the NNPACK interface."); -DEFINE_int32(nnpack_num_threads, - 0, - "The number of nnpack threads" - "default: 0; 0 to disable threadpool."); - -namespace paddle { - -nnp_convolution_algorithm get_nnp_convolution_algorithm( - const std::string& algorithm) { - if (algorithm == "auto") { - return nnp_convolution_algorithm_auto; - } else if (algorithm == "ft8x8") { - return nnp_convolution_algorithm_ft8x8; - } else if (algorithm == "ft16x16") { - return nnp_convolution_algorithm_ft16x16; - } else if (algorithm == "wt8x8") { - return nnp_convolution_algorithm_wt8x8; - } else if (algorithm == "implicit-gemm") { - return nnp_convolution_algorithm_implicit_gemm; - } else if (algorithm == "direct") { - return nnp_convolution_algorithm_direct; - } else { - return nnp_convolution_algorithm_auto; - } -} - -template -class NNPACKConvFunction : public ConvFunctionBase { - public: - void init(const FuncConfig& config) override { - ConvFunctionBase::init(config); - algorithm_ = get_nnp_convolution_algorithm(config.get("algo")); - transform_strategy_ = nnp_convolution_transform_strategy_compute; - nnp_status status = nnp_initialize(); - CHECK_EQ(status, nnp_status_success); - workspaceBuffer_ = nullptr; - workspaceSize_ = 0; - - create_nnpack_threadpool(); - } - - ~NNPACKConvFunction() { - if (workspaceBuffer_) { - free(workspaceBuffer_); - } - } - - void check(const BufferArgs& inputs, const BufferArgs& outputs) override { - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - checkShape(input, filter, output); - } - - void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { - CHECK_EQ(numInputs_, inputs.size()); - CHECK_EQ(numOutputs_, outputs.size()); - CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); - check(inputs, outputs); - const TensorShape& input = inputs[0].shape(); - const TensorShape& filter = inputs[1].shape(); - const TensorShape& output = outputs[0].shape(); - - size_t batchSize = input[0]; - size_t inputChannels = input[1]; - size_t inputHeight = input[2]; - size_t inputWidth = input[3]; - size_t filterHeight = getFilterHeight(filter); - size_t filterWidth = getFilterWidth(filter); - size_t outputChannels = output[1]; - size_t outputHeight = output[2]; - size_t outputWidth = output[3]; - - nnp_size inputSize = {.width = inputWidth, .height = inputHeight}; - nnp_padding padding = {.top = (size_t)paddingH(), - .right = (size_t)paddingW(), - .bottom = (size_t)paddingH(), - .left = (size_t)paddingW()}; - nnp_size kernelSize = {.width = filterWidth, .height = filterHeight}; - nnp_size outputSubsampling = {.width = (size_t)strideW(), - .height = (size_t)strideH()}; - - float* inputData = inputs[0].data(); - float* filterData = inputs[1].data(); - float* outputData = outputs[0].data(); - - void* bufferPtr = nullptr; - size_t* sizePtr = nullptr; - size_t needSize; - if (FLAGS_nnpack_allocate_outside) { - if (batchSize == 1) { - nnp_status status = nnp_convolution_inference(algorithm_, - transform_strategy_, - inputChannels, - outputChannels, - inputSize, - padding, - kernelSize, - outputSubsampling, - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, - &needSize, - nnp_activation_identity, - nullptr, - nullptr, - nullptr); - CHECK_EQ(status, nnp_status_success); - } else { - // only supports stride = 1 - CHECK_EQ(strideH(), 1); - CHECK_EQ(strideW(), 1); - nnp_status status = nnp_convolution_output(algorithm_, - batchSize, - inputChannels, - outputChannels, - inputSize, - padding, - kernelSize, - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, - &needSize, - nnp_activation_identity, - nullptr, - nullptr, - nullptr); - CHECK_EQ(status, nnp_status_success); - } - - VLOG(3) << "workspace size is " << needSize; - if (needSize > workspaceSize_) { - workspaceSize_ = needSize; - if (workspaceBuffer_) { - free(workspaceBuffer_); - } else { - posix_memalign(&workspaceBuffer_, 64, needSize); - } - } - - if (needSize) { - bufferPtr = workspaceBuffer_; - sizePtr = &needSize; - } - } - - size_t inputOffset = inputChannels / groups_ * inputHeight * inputWidth; - size_t outputOffset = outputChannels / groups_ * outputHeight * outputWidth; - size_t filterOffset = filter.getElements() / groups_; - - if (batchSize == 1) { - for (size_t g = 0; g < groups_; g++) { - nnp_status status = - nnp_convolution_inference(algorithm_, - transform_strategy_, - inputChannels / groups_, - outputChannels / groups_, - inputSize, - padding, - kernelSize, - outputSubsampling, - inputData + inputOffset * g, - filterData + filterOffset * g, - nullptr, /* bias */ - outputData + outputOffset * g, - bufferPtr, - sizePtr, - nnp_activation_identity, - nullptr, - threadpool_, /* threadpool */ - nullptr); - CHECK_EQ(status, nnp_status_success); - } - } else { - // only supports stride = 1 - CHECK_EQ(strideH(), 1); - CHECK_EQ(strideW(), 1); - - // TODO(hedaoyuan): There has some bug when batchSize > 1 and groups_ > 1. - CHECK_EQ(groups_, static_cast(1)); - nnp_status status = nnp_convolution_output(algorithm_, - batchSize, - inputChannels, - outputChannels, - inputSize, - padding, - kernelSize, - inputData, - filterData, - nullptr, /* bias */ - outputData, - bufferPtr, - sizePtr, - nnp_activation_identity, - nullptr, - threadpool_, /* threadpool */ - nullptr); - CHECK_EQ(status, nnp_status_success); - } - } - - static void create_nnpack_threadpool() { - if (FLAGS_nnpack_num_threads && threadpool_ == nullptr) { - threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads); - VLOG(3) << "Number of threads " - << pthreadpool_get_threads_count(threadpool_); - } - } - - private: - nnp_convolution_algorithm algorithm_; - nnp_convolution_transform_strategy transform_strategy_; - void* workspaceBuffer_; - size_t workspaceSize_; - static pthreadpool_t threadpool_; -}; - -template -pthreadpool_t NNPACKConvFunction::threadpool_ = nullptr; - -REGISTER_TYPED_FUNC(NNPACKConv, CPU, NNPACKConvFunction); - -} // namespace paddle diff --git a/paddle/function/nnpack/NNPACKConvOpTest.cpp b/paddle/function/nnpack/NNPACKConvOpTest.cpp deleted file mode 100644 index c80ffb5d5d255465e9a2fa251fb9a6c61f96e7ec..0000000000000000000000000000000000000000 --- a/paddle/function/nnpack/NNPACKConvOpTest.cpp +++ /dev/null @@ -1,30 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/function/ConvOpTest.h" - -namespace paddle { - -TEST(NNPACK, Forward) { - Convolution( - "GemmConv-CPU", "NNPACKConv-CPU", forward); -} - -TEST(NNPACK, Depthwise) { - DepthwiseConvolution( - "GemmConv-CPU", "NNPACKConv-CPU", forward); -} - -} // namespace paddle diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp deleted file mode 100644 index 71c238fbfe9f32f3764601ebb441336931f8ef5f..0000000000000000000000000000000000000000 --- a/paddle/gserver/activations/ActivationFunction.cpp +++ /dev/null @@ -1,509 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ActivationFunction.h" - -#include -#include -#include -#include -#include -#include -#include "paddle/parameter/Argument.h" -#include "paddle/utils/ClassRegistrar.h" -#include "paddle/utils/Logging.h" - -#ifdef PADDLE_WITH_MKLDNN -#include "MKLDNNActivation.h" -#endif - -namespace paddle { - -static ClassRegistrar gActivationRegistrar; -/** - * @def ACTIVATION_CLASS_NAME - * @brief Macro for getting derived activation class name - * @note ACTIVATION_CLASS_NAME(softmax) softmax_; - * means softmaxActivation softmax_; - */ -#define ACTIVATION_CLASS_NAME(ACTIVATION_NAME) ACTIVATION_NAME##Activation -/** - * @def BEGIN_DEFINE_ACTIVATION - * @brief Macro for defining a devried activation class - */ -#define BEGIN_DEFINE_ACTIVATION(ACTIVATION_NAME) \ - class ACTIVATION_CLASS_NAME(ACTIVATION_NAME) : public ActivationFunction { \ - private: \ - static const std::string name; \ - \ - public: \ - const std::string& getName() const { return name; } -/** - * @def END_DEFINE_ACTIVATION - * @brief Macro for registering a derived activation class - */ -#define END_DEFINE_ACTIVATION(ACTIVATION_NAME) \ - } \ - ; \ - const std::string ACTIVATION_CLASS_NAME(ACTIVATION_NAME)::name = \ - #ACTIVATION_NAME; \ - static InitFunction __reg_activation__##ACTIVATION_NAME([] { \ - gActivationRegistrar \ - .registerClass( \ - #ACTIVATION_NAME); \ - }); - -/** - * @brief The IdentityActivation class - * - * Do nothing when forward/backward. - */ -class IdentityActivation : public ActivationFunction { - public: - static const std::string name; - Error __must_check forward(Argument& act) { - (void)act; - return Error(); - } - Error __must_check backward(Argument& act) { - (void)act; - return Error(); - } - const std::string& getName() const { return name; } -}; -const std::string IdentityActivation::name = ""; -static InitFunction __reg_activation__identity([] { - gActivationRegistrar.registerClass(""); - gActivationRegistrar.registerClass("linear"); -}); - -/** - * @brief Sigmoid Activation - * \f[ - * f(z) = \frac{1}{1+exp(-z)} - * \f] - */ -BEGIN_DEFINE_ACTIVATION(sigmoid) -Error __must_check forward(Argument& act) { - act.value->sigmoid(*act.value); - return Error(); -} -Error __must_check backward(Argument& act) { - act.grad->sigmoidDerivative(*act.value); - return Error(); -} -END_DEFINE_ACTIVATION(sigmoid) - -/** - * @brief Softmax Activation - * \f[ - * P(y=j|x) = \frac{e^{x^Tw_j}}{\sum^K_{k=1}e^{x^Tw_k}} - * \f] - */ -BEGIN_DEFINE_ACTIVATION(softmax) -private: -MatrixPtr sftMaxSum_; -MatrixPtr sftMaxDot_; - -public: -Error __must_check forward(Argument& act) { - act.value->softmax(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - MatrixPtr outputV = act.value; - MatrixPtr outputG = act.grad; - - if (outputG->useGpu()) { - outputG->softmaxBackward(*outputV); - } else { - SetDevice device(act.deviceId); - Matrix::resizeOrCreate(sftMaxDot_, - outputG->getHeight(), - outputG->getWidth(), - /* trans */ false, - useGpu(act.deviceId)); - Matrix::resizeOrCreate(sftMaxSum_, - outputG->getHeight(), - 1, - /* trans */ false, - useGpu(act.deviceId)); - - sftMaxDot_->dotMul(*outputG, *outputV); - sftMaxSum_->colMerge(*sftMaxDot_); - - act.grad->softmaxDerivative(*act.value, *sftMaxSum_); - } - return Error(); -} -END_DEFINE_ACTIVATION(softmax) - -/** - * @brief Sequence_softmax Activation - * @note Softmax on all frames of one sequence. - * Width of frame must be one. - */ -BEGIN_DEFINE_ACTIVATION(sequence_softmax) -private: -ACTIVATION_CLASS_NAME(softmax) softmax_; -Argument argument_; - -public: -Error __must_check forward(Argument& act) { - if (act.value->getWidth() != 1UL) { - return Error( - "Input width for each timestep of sequence softmax should be 1"); - } - - if (!argument_.value) { - argument_.value = Matrix::create(nullptr, - /* height= */ 1, - 1, - /* trans= */ false, - useGpu(act.deviceId)); - argument_.grad = Matrix::create(nullptr, - /* height= */ 1, - 1, - /* trans= */ false, - useGpu(act.deviceId)); - } - - auto starts = - act.hasSubseq() - ? act.subSequenceStartPositions->getVector(useGpu(act.deviceId)) - : act.sequenceStartPositions->getVector(useGpu(act.deviceId)); - act.value->sequenceSoftmax(*act.value, *starts); - return Error(); -} - -Error __must_check backward(Argument& act) { - if (act.value->getWidth() != 1UL) { - return Error( - "Input width for each timestep of sequence softmax should be 1"); - } - - size_t numSequences = - act.hasSubseq() ? act.getNumSubSequences() : act.getNumSequences(); - const int* starts = act.getCpuStartPositions(); - - for (size_t i = 0; i < numSequences; ++i) { - // TODO(Dangqingqing) optimization for GPU - size_t offset = starts[i]; - size_t size = starts[i + 1] - starts[i]; - argument_.value->setData(act.value->getData() + offset, 1UL, size); - argument_.grad->setData(act.grad->getData() + offset, 1UL, size); - - Error err = softmax_.backward(argument_); - if (!err.isOK()) return err; - } - return Error(); -} -END_DEFINE_ACTIVATION(sequence_softmax) - -/* - * @brief SoftSign Activation. - * \f[ - * f(z) = \frac{z}{1 + |z|} - * \f] - */ -BEGIN_DEFINE_ACTIVATION(softsign) -private: -MatrixPtr denominator_; - -Error __must_check forward(Argument& act) { - size_t height = act.value->getHeight(); - size_t width = act.value->getWidth(); - Matrix::resizeOrCreate( - denominator_, height, width, false, useGpu(act.deviceId)); - denominator_->assign(*act.value); - denominator_->abs2(); - denominator_->add(1.); - - act.value->dotDiv(*act.value, *denominator_); - return Error(); -} - -Error __must_check backward(Argument& act) { - denominator_->square2(); - denominator_->scalarDiv(*denominator_, 1.); - act.grad->dotMul(*act.grad, *denominator_); - return Error(); -} -END_DEFINE_ACTIVATION(softsign) - -/** - * @brief Relu Activation. - * forward. y = max(0, z) - * - * derivative of relu is: - * - * 1 if z > 0 - * - * 0 otherwise. - */ -BEGIN_DEFINE_ACTIVATION(relu) -Error __must_check forward(Argument& act) { - act.value->relu(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->reluDerivative(*act.value); - return Error(); -} -END_DEFINE_ACTIVATION(relu) - -/** - * @brief BRelu Activation. - * - * forward. y = min(24, max(0, z)) - * - * derivative of brelu is: - * - * 1 if 0 < z < 24 - * - * 0 otherwise. - * - * TODO(yuyang18): Remove magic number 24 or make it configuable. - */ -BEGIN_DEFINE_ACTIVATION(brelu) -Error __must_check forward(Argument& act) { - act.value->brelu(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->breluDerivative(*act.value); - return Error(); -} -END_DEFINE_ACTIVATION(brelu) - -/** - * @brief Tanh Activation. - * \f[ - * f(z) = tanh(z)=\frac{e^z-e^{-z}}{e^z+e^{-z}} - * \f] - */ -BEGIN_DEFINE_ACTIVATION(tanh) -Error __must_check forward(Argument& act) { - act.value->tanh(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->tanhDerivative(*act.value); - return Error(); -} -END_DEFINE_ACTIVATION(tanh) - -/** - * @brief Scaled Tanh Activation - * \f[ - * f(z) = 1.7159 * tanh(2/3*z) - * \f] - */ -BEGIN_DEFINE_ACTIVATION(stanh) -private: -real a, b; - -public: -ACTIVATION_CLASS_NAME(stanh)() : a(1.7159), b(2. / 3.) {} -Error __must_check forward(Argument& act) { - act.value->scaledTanh(*act.value, a, b); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->scaledTanhDerivative(*act.value, a, b); - return Error(); -} -END_DEFINE_ACTIVATION(stanh) - -/** - * @brief Soft Relu Activation. - * \f[ - * f(z) = ln(1+e^z) - * \f] - */ -BEGIN_DEFINE_ACTIVATION(softrelu) -Error __must_check forward(Argument& act) { - act.value->softrelu(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->softreluDerivative(*act.value); - return Error(); -} -END_DEFINE_ACTIVATION(softrelu) - -/** - * @brief Abs Activation. - * Forward: f(z) = abs(z) - * - * Derivative: - * - * 1 if z>0 - * - * -1 if z<0 - * - * 0 if z=0 - */ -BEGIN_DEFINE_ACTIVATION(abs) -Error __must_check forward(Argument& act) { - SetDevice device(act.deviceId); - Matrix::resizeOrCreate(act.in, - act.value->getHeight(), - act.value->getWidth(), - /* trans */ false, - useGpu(act.deviceId)); - - act.in->copyFrom(*act.value); - act.value->abs2(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->absDerivative(*act.in); - return Error(); -} -END_DEFINE_ACTIVATION(abs) - -/** - * @brief Square Activation. - * \f[ - * f(z) = z^2. - * \f] - */ -BEGIN_DEFINE_ACTIVATION(square) -Error __must_check forward(Argument& act) { - SetDevice device(act.deviceId); - Matrix::resizeOrCreate(act.in, - act.value->getHeight(), - act.value->getWidth(), - /* trans */ false, - useGpu(act.deviceId)); - - act.in->copyFrom(*act.value); - act.value->square2(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->squareDerivative(*act.in); - return Error(); -} -END_DEFINE_ACTIVATION(square) - -/** - * @brief Exponential Activation. - * \f[ - * f(z) = e^z - * \f] - */ -BEGIN_DEFINE_ACTIVATION(exponential) -Error __must_check forward(Argument& act) { - act.value->exp2(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->expDerivative(*act.value); - return Error(); -} -END_DEFINE_ACTIVATION(exponential) - -/** - * @brief Reciprocal Activation. - * \f[ - * f(z) = 1/z - * \f] - */ -BEGIN_DEFINE_ACTIVATION(reciprocal) -Error __must_check forward(Argument& act) { - act.value->reciprocal2(); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->dotMulSquare(*act.value); - act.grad->neg(); - return Error(); -} -END_DEFINE_ACTIVATION(reciprocal) - -/** - * @brief Square Root Activation. - * \f[ - * f(z) = sqrt(z) - * \f] - */ -BEGIN_DEFINE_ACTIVATION(sqrt) -Error __must_check forward(Argument& act) { - act.value->sqrt2(); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->dotDiv(*act.grad, *act.value); - act.grad->mulScalar(0.5); - return Error(); -} -END_DEFINE_ACTIVATION(sqrt) - -/** - * @brief Logarithm Activation. - * \f[ - * f(z) = log(z) - * \f] - */ -BEGIN_DEFINE_ACTIVATION(log) -Error __must_check forward(Argument& act) { - SetDevice device(act.deviceId); - Matrix::resizeOrCreate(act.in, - act.value->getHeight(), - act.value->getWidth(), - /* trans */ false, - useGpu(act.deviceId)); - - act.in->copyFrom(*act.value); - act.value->log2(*act.value); - return Error(); -} - -Error __must_check backward(Argument& act) { - act.grad->dotDiv(*act.grad, *act.in); - return Error(); -} -END_DEFINE_ACTIVATION(log) - -ActivationFunction* ActivationFunction::create(const std::string& type) { -#ifdef PADDLE_WITH_MKLDNN - if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) { - return MKLDNNActivation::create(type); - } -#endif - - return gActivationRegistrar.createByType(type); -} - -std::vector ActivationFunction::getAllRegisteredTypes() { - std::vector types; - gActivationRegistrar.forEachType( - [&](const std::string& type) { types.push_back(type); }); - return types; -} - -} // namespace paddle diff --git a/paddle/gserver/activations/ActivationFunction.h b/paddle/gserver/activations/ActivationFunction.h deleted file mode 100644 index 8e2e144769f2e668a9a8f02890d29c4a7fe128a3..0000000000000000000000000000000000000000 --- a/paddle/gserver/activations/ActivationFunction.h +++ /dev/null @@ -1,66 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include "paddle/utils/Error.h" - -namespace paddle { - -struct Argument; -/** - * @brief Activation function is a function that transforms a set of input - * signals into an output signals. The purpose of the activation function - * is to introduce non-liearilty into the network. - * - * @note Common activation function are provieded, including linear, - * sigmoid, softmax, sequence_max, relu, brelu, tanh, stanh, - * softrelu, abs, square, exponential. - * - */ -class ActivationFunction { - public: - static ActivationFunction* create(const std::string& type); - static std::vector getAllRegisteredTypes(); - - ActivationFunction() {} - - virtual ~ActivationFunction() {} - - /** - * @brief Foward propagation - * - * act.value <- f(act.value), - * where f is the activation function. - * Suppose that before calling forward(), act.value is x and - * after forward() is called, act.value is y, then y = f(x). - * - * Usually, act is Layer::output_ - */ - virtual Error __must_check forward(Argument& act) = 0; - - /** - * @brief Backward propagaion - * - * x and y are defined in the above comment for forward(). - * - Before calling backward(), act.grad = dE / dy, where E is the error/cost - * - After backward() returns, act.grad = dE / dx = (dE/dy) * (dy/dx) - */ - virtual Error __must_check backward(Argument& act) = 0; - - virtual const std::string& getName() const = 0; -}; - -} // namespace paddle diff --git a/paddle/gserver/activations/MKLDNNActivation.cpp b/paddle/gserver/activations/MKLDNNActivation.cpp deleted file mode 100644 index 672444c6561adbeb78c3c453f12ab6aaedeed646..0000000000000000000000000000000000000000 --- a/paddle/gserver/activations/MKLDNNActivation.cpp +++ /dev/null @@ -1,249 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MKLDNNActivation.h" -#include "mkldnn.hpp" -#include "paddle/utils/ClassRegistrar.h" - -namespace paddle { - -static ClassRegistrar gMKLDNNActivationRegistrar; -/** - * @def MKLDNN_ACTIVATION_CLASS_NAME - * @note MKLDNN_ACTIVATION_CLASS_NAME(relu) relu_; - * means mkldnn_reluActivation relu_; - */ -#define MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) mkldnn_##ACT_TYPE##Activation - -/** - * @def BEGIN_MKLDNN_ACTIVATION - */ -#define BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ - class MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) : public BASE_CLASS { -/** - * @def END_MKLDNN_ACTIVATION - */ -#define END_MKLDNN_ACTIVATION(ACT_TYPE) \ - private: \ - static const std::string name; \ - \ - public: \ - const std::string& getName() const { return name; } \ - } \ - ; \ - const std::string MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::name = \ - "mkldnn_" #ACT_TYPE; \ - static InitFunction __reg_activation__mkldnn_##ACT_TYPE([] { \ - gMKLDNNActivationRegistrar \ - .registerClass( \ - "mkldnn_" #ACT_TYPE); \ - }); - -/** - * @def DEFINE_MKLDNN_ACTIVATION - */ -#define DEFINE_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ - BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ - END_MKLDNN_ACTIVATION(ACT_TYPE) - -/** - * @def DEFINE_MKLDNN_ELTWISE_ACTIVATION - */ -#define DEFINE_MKLDNN_ELTWISE_ACTIVATION( \ - ACT_TYPE, BASE_CLASS, ALPHA, BWD_ALPHA) \ - BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ - private: \ - static const float alpha; \ - static const float bwdAlpha; \ - \ - public: \ - float getAlpha() const { return alpha; } \ - float getBwdAlpha() const { return bwdAlpha; } \ - END_MKLDNN_ACTIVATION(ACT_TYPE) \ - const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::alpha = ALPHA; \ - const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::bwdAlpha = BWD_ALPHA; - -/** - * @brief MKLDNN Relu Activation. - * Actually mkldnn_relu is Leaky Relu. - * f(x) = x (x >= 0) - * f(x) = negative_slope * x (x < 0) - * @note the negative_slope should be -0.f in forward - */ -DEFINE_MKLDNN_ELTWISE_ACTIVATION(relu, MKLDNNEltwiseActivation, -0.f, 0.f) - -/** - * @brief MKLDNN Tanh Activation. - */ -DEFINE_MKLDNN_ELTWISE_ACTIVATION(tanh, MKLDNNEltwiseActivation, 0.f, 0.f) - -/** - * @brief MKLDNN ELU(Exponential Linear Unit) Activation. - * f(x) = x (x >= 0) - * f(x) = negative_slope * (exp(x) - 1) (x < 0) - */ -DEFINE_MKLDNN_ELTWISE_ACTIVATION(elu, MKLDNNEltwiseActivation, 0.f, 0.f) - -mkldnn::algorithm MKLDNNEltwiseActivation::getAlgo(std::string type) const { - const std::map algoMap = { - {"relu", algorithm::eltwise_relu}, - {"tanh", algorithm::eltwise_tanh}, - {"elu", algorithm::eltwise_elu}}; - type.erase(0, 7); // remove mkldnn_ - algorithm algo = (algorithm)0; - mapGet(type, algoMap, &algo); - return algo; -} - -void MKLDNNEltwiseActivation::resetFwd(Argument& act) { - if (cnt_ == act.value->getElementCnt()) { - return; - } - MKLDNNActivation::resetFwd(act); - // note: alpha represents the NegativeSlope when used in relu. - float alpha = getAlpha(); - float beta = getBeta(); - algorithm algo = getAlgo(this->getName()); - auto fwdDesc = eltwise_fwd::desc(mkldnn::prop_kind::forward_training, - algo, - val_->getMemoryDesc(), - alpha, - beta); - fwdPD_.reset(new eltwise_fwd::primitive_desc(fwdDesc, *engine_)); - // use inplace for forward but save input value before submit - inVal_ = val_; - copyInVal_ = nullptr; - if (act.grad && algo == algorithm::eltwise_tanh) { - // tanh need save src input for backward - inVal_ = MKLDNNMatrix::create(val_->getPrimitiveDesc()); - copyInVal_ = std::make_shared(*val_, *inVal_); - CHECK(copyInVal_) << "should not be emptry"; - pipelineFwd_.push_back(*copyInVal_); - } - fwd_.reset(new eltwise_fwd(*fwdPD_, *val_, *val_)); - pipelineFwd_.push_back(*fwd_); - needResetBwd_ = true; -} - -void MKLDNNEltwiseActivation::resetBwd(Argument& act) { - if (!needResetBwd_) { - return; - } - VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; - needResetBwd_ = false; - algorithm algo = getAlgo(this->getName()); - float alpha = getBwdAlpha(); - float beta = getBeta(); - grad_ = MKLDNNMatrix::create(val_->getPrimitiveDesc(), act.grad); - auto eng = CPUEngine::Instance().getEngine(); - auto bwdDesc = eltwise_bwd::desc( - algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta); - auto bwdPD = eltwise_bwd::primitive_desc(bwdDesc, eng, *fwdPD_); - CHECK(inVal_); - bwd_.reset(new eltwise_bwd(bwdPD, *inVal_, *grad_, *grad_)); - pipelineBwd_.clear(); - pipelineBwd_.push_back(*bwd_); -} - -/** - * @brief MKLDNN Softmax Activation - */ -DEFINE_MKLDNN_ACTIVATION(softmax, MKLDNNSoftmaxActivation) - -void MKLDNNSoftmaxActivation::resetFwd(Argument& act) { - if (cnt_ == act.value->getElementCnt()) { - return; - } - MKLDNNActivation::resetFwd(act); - int axis = 1; - auto fwdDesc = softmax_fwd::desc( - mkldnn::prop_kind::forward_scoring, val_->getMemoryDesc(), axis); - auto fwdPD = softmax_fwd::primitive_desc(fwdDesc, *engine_); - fwd_.reset(new softmax_fwd(fwdPD, *val_, *val_)); - pipelineFwd_.push_back(*fwd_); -} - -Error __must_check MKLDNNSoftmaxActivation::forward(Argument& act) { - resetFwd(act); - stream_->submit(pipelineFwd_); - real* v = act.value->getData(); - real threshold = exp(-64); -#pragma omp parallel for - for (size_t i = 0; i < act.value->getElementCnt(); ++i) { - v[i] = v[i] < threshold ? threshold : v[i]; - } - return Error(); -} - -Error __must_check MKLDNNSoftmaxActivation::backward(Argument& act) { - MatrixPtr outputV = act.value; - MatrixPtr outputG = act.grad; - Matrix::resizeOrCreate(sftMaxDot_, - outputG->getHeight(), - outputG->getWidth(), - /* trans */ false, - /* useGpu */ false); - Matrix::resizeOrCreate(sftMaxSum_, - outputG->getHeight(), - 1, - /* trans */ false, - /* useGpu */ false); - sftMaxDot_->dotMul(*outputG, *outputV); - sftMaxSum_->colMerge(*sftMaxDot_); - act.grad->softmaxDerivative(*act.value, *sftMaxSum_); - return Error(); -} - -ActivationFunction* MKLDNNActivation::create(const std::string& type) { - return gMKLDNNActivationRegistrar.createByType(type); -} - -std::vector MKLDNNActivation::getAllRegisteredTypes() { - std::vector types; - gMKLDNNActivationRegistrar.forEachType( - [&](const std::string& type) { types.push_back(type); }); - return types; -} - -void MKLDNNActivation::resetFwd(Argument& act) { - VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; - cnt_ = act.value->getElementCnt(); - pipelineFwd_.clear(); - stream_.reset(new MKLDNNStream()); - engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0)); - val_ = std::dynamic_pointer_cast(act.value); - if (val_ == nullptr) { - int bs = act.getBatchSize(); - int ih = act.getFrameHeight() > 0 ? act.getFrameHeight() : 1; - int iw = act.getFrameWidth() > 0 ? act.getFrameWidth() : 1; - int ic = cnt_ / bs / ih / iw; - CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw); - val_ = MKLDNNMatrix::create( - {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_, act.value); - CHECK(val_); - val_->downSpatial(); - } -} - -Error __must_check MKLDNNActivation::forward(Argument& act) { - resetFwd(act); - stream_->submit(pipelineFwd_); - return Error(); -} -Error __must_check MKLDNNActivation::backward(Argument& act) { - resetBwd(act); - stream_->submit(pipelineBwd_); - return Error(); -} -} // namespace paddle diff --git a/paddle/gserver/activations/MKLDNNActivation.h b/paddle/gserver/activations/MKLDNNActivation.h deleted file mode 100644 index eece1b9c37e72624dffd119804c65f7bd36e20fb..0000000000000000000000000000000000000000 --- a/paddle/gserver/activations/MKLDNNActivation.h +++ /dev/null @@ -1,119 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "ActivationFunction.h" -#include "mkldnn.hpp" -#include "paddle/gserver/layers/MKLDNNBase.h" -#include "paddle/math/MKLDNNMatrix.h" -#include "paddle/parameter/Argument.h" - -namespace paddle { - -/** - * @brief Base class of MKLDNN Activation. - * Common activation function are provieded, - * including mkldnn_relu, mkldnn_elu, mkldnn_tanh, mkldnn_softmax - */ -class MKLDNNActivation : public ActivationFunction { - protected: - // input value element count - size_t cnt_; - // should not merge the resetBwd into resetFwd, - // because the grad data would be changing before backward. - bool needResetBwd_; - // mkldnn matrix, primitive, stream and pipeline - MKLDNNMatrixPtr val_; - MKLDNNMatrixPtr grad_; - std::shared_ptr engine_; - std::shared_ptr stream_; - std::shared_ptr fwd_; - std::shared_ptr bwd_; - std::vector pipelineFwd_; - std::vector pipelineBwd_; - - public: - MKLDNNActivation() : cnt_(0), needResetBwd_(true) {} - ~MKLDNNActivation() {} - static ActivationFunction* create(const std::string& type); - static std::vector getAllRegisteredTypes(); - virtual const std::string& getName() const = 0; - /** - * reset the forward primitives - */ - virtual void resetFwd(Argument& act); - /** - * reset the backward primitives, - * can not merge this functions into resetFwd as the grad data - * would be changing before backward. - */ - virtual void resetBwd(Argument& act) {} - virtual Error __must_check forward(Argument& act); - virtual Error __must_check backward(Argument& act); -}; - -/** - * @brief Base class of MKLDNN Eltwise Activation, - * includes mkldnn_relu, mkldnn_elu and mkldnn_tanh. - */ -class MKLDNNEltwiseActivation : public MKLDNNActivation { - typedef mkldnn::eltwise_forward eltwise_fwd; - typedef mkldnn::eltwise_backward eltwise_bwd; - typedef mkldnn::algorithm algorithm; - - protected: - // save the forward primitive desc, which can be used backward - std::shared_ptr fwdPD_; - // eltwise_bwd need src input value - MKLDNNMatrixPtr inVal_; - // use for copy data - std::shared_ptr copyInVal_; - - public: - MKLDNNEltwiseActivation() {} - ~MKLDNNEltwiseActivation() {} - virtual const std::string& getName() const = 0; - - // in common, the alpha of forward and backward should be equal. - // but for relu, to avoid negative value, they should be opposite - virtual float getAlpha() const = 0; - virtual float getBwdAlpha() const = 0; - virtual float getBeta() const { return 0.f; } - virtual algorithm getAlgo(std::string type) const; - void resetFwd(Argument& act) override; - void resetBwd(Argument& act) override; -}; - -/** - * @brief Base class of MKLDNN softmax Activation, - * only have mkldnn forward, use cpu implement for backward. - */ -class MKLDNNSoftmaxActivation : public MKLDNNActivation { - typedef mkldnn::softmax_forward softmax_fwd; - - private: - // for backward - MatrixPtr sftMaxSum_; - MatrixPtr sftMaxDot_; - - public: - MKLDNNSoftmaxActivation() {} - ~MKLDNNSoftmaxActivation() {} - virtual const std::string& getName() const = 0; - void resetFwd(Argument& act) override; - Error __must_check forward(Argument& act) override; - Error __must_check backward(Argument& act) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/dataproviders/DataProvider.cpp b/paddle/gserver/dataproviders/DataProvider.cpp deleted file mode 100644 index 580cf821c685b3daf7f015bc137c6d5ea31ef100..0000000000000000000000000000000000000000 --- a/paddle/gserver/dataproviders/DataProvider.cpp +++ /dev/null @@ -1,410 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "DataProvider.h" - -#include -#include -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/StringUtil.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -void BufferBatch::swap(BufferBatch* bufBatch) { - DataBatch* batchData = bufBatch->getDataBatch(); - hl_event_t hlEvent = bufBatch->getCuEvent(); - hl_stream_t hlStream = bufBatch->getCuStream(); - bufBatch->setDataBatch(batchData_); - bufBatch->setCuStream(hlStream_); - bufBatch->setCuEvent(hlEvent_); - - batchData_ = batchData; - hlEvent_ = hlEvent; - hlStream_ = hlStream; -} - -void BufferBatch::clone(DataBatch* srcBatch, bool useGpu) { - if (batchData_ == NULL) { - batchData_ = new DataBatch(); - } - std::vector& destData = batchData_->getStreams(); - int numStreams = srcBatch->getNumStreams(); - destData.resize(numStreams); - batchData_->setSize(srcBatch->getSize()); - if (useGpu) { - createCuEvent(); - } - - for (int i = 0; i < numStreams; i++) { - destData[i].resizeAndCopyFrom(srcBatch->getStream(i), useGpu, hlStream_); - } - if (useGpu) { - hl_stream_record_event(hlStream_, hlEvent_); - } -} - -DoubleBuffer::DoubleBuffer(DataProvider* dataPool, - bool useGpu, - int64_t batchSize) { - batchSize_ = batchSize; - dataPool_ = dataPool; - useGpu_ = useGpu; - dataQueue_ = new BufferBatchQueue(); - bufferQueue_ = new BufferBatchQueue(); - - // insert a empty buffer - bufferQueue_->enqueue(new BufferBatch()); - stopping_ = false; - pending_ = true; -} - -DoubleBuffer::~DoubleBuffer() { - finishAsyncLoad(); - while (dataQueue_->size()) { - BufferBatch* dataBtch = dataQueue_->dequeue(); - delete dataBtch; - dataBtch = NULL; - } - while (bufferQueue_->size()) { - BufferBatch* bufBtch = bufferQueue_->dequeue(); - delete bufBtch; - bufBtch = NULL; - } - delete dataQueue_; - dataQueue_ = NULL; - delete bufferQueue_; - bufferQueue_ = NULL; -} - -void DoubleBuffer::removeOneBatch(DataBatch* dataBatch) { - // get data - BufferBatch* batch = dataQueue_->dequeue(); - batch->syncEvent(); // when use GPU, need synchronized with the cuEvent - *dataBatch = *(batch->getDataBatch()); - - // push anothor buffer - if (*usingBatch_ == nullptr) { - *usingBatch_ = std::make_shared(); - } - - // Mark the using-batch - batch->swap((*usingBatch_).get()); - bufferQueue_->enqueue(batch); - - if (0 == dataBatch->getSize()) { - setPending(true); - } -} - -void DoubleBuffer::insertOneBatch(DataBatch* batch) { - while (!bufferQueue_->waitNotEmptyFor(2 /* seconds */)) { // time out - if (stopping_) return; - } - BufferBatch* bufBatch = bufferQueue_->dequeue(); - // clone and copy the data from an Threadlocal Variable - bufBatch->clone(batch, useGpu_); - dataQueue_->enqueue(bufBatch); -} - -void DoubleBuffer::asyncLoadBatch() { - int64_t actualSize = 0; - if (useGpu_) { - hl_set_device(FLAGS_gpu_id); - } - setPending(false); - - while (true) { - taskReadySem_.wait(); - if (stopping_) break; - - while (batchSize_ == 0 && !stopping_) { - usleep(5); - } - if (stopping_) break; - - do { - DataBatch newBatch; - { - REGISTER_TIMER("getNextBatchInternal"); - actualSize = dataPool_->getNextBatchInternal(batchSize_, &newBatch); - } - insertOneBatch(&newBatch); - } while (actualSize > 0 && !stopping_); - } -} - -void DoubleBuffer::startAsyncLoad() { - if (asyncLoader_ == nullptr) { - asyncLoader_.reset(new std::thread([this]() { this->asyncLoadBatch(); })); - } - taskReadySem_.post(); -} - -ClassRegistrar - DataProvider::registrar_; - -DataProvider* DataProvider::create(const DataConfig& config, - const ModelConfig& modelConfig, - bool useGpu) { - return registrar_.createByType(config.type(), config, modelConfig, useGpu); -} - -REGISTER_DATA_PROVIDER(simple, SimpleDataProvider); -REGISTER_DATA_PROVIDER(dummy, DummyDataProvider); - -int64_t DataProvider::getNextBatch(int64_t size, DataBatch* batch) { - int64_t batchSize = doubleBuffer_ ? getNextBatchFromBuffer(size, batch) - : getNextBatchInternal(size, batch); - - if (!batchSize) return 0; - - if (!config_.constant_slots_size()) return batchSize; - - auto& constantSlots = *constantSlots_; - constantSlots.resize(config_.constant_slots_size()); - - for (int i = 0; i < config_.constant_slots_size(); ++i) { - MemoryHandlePtr handle = - constantSlots[i] ? constantSlots[i]->getMemoryHandle() : nullptr; - Matrix::resizeOrCreate(constantSlots[i], - batchSize, - 1, // = width - false, // = trans - useGpu_); // = useGpu - if (handle != constantSlots[i]->getMemoryHandle()) { - // memory buf was reallocated. We need to initialize the value - constantSlots[i]->assign(config_.constant_slots(i)); - } - batch->appendData(constantSlots[i], - batch->getStream(0).sequenceStartPositions); - } - - return batchSize; -} - -int64_t DataProvider::getNextBatchFromBuffer(int64_t size, DataBatch* batch) { - CHECK(doubleBuffer_ != nullptr); - - if (doubleBuffer_->getBatchSize() != size) { - doubleBuffer_->setBatchSize(size); - } - - doubleBuffer_->removeOneBatch(batch); - return batch->getSize(); -} - -void DataProvider::initAsyncLoader() { - if (doubleBuffer_ == nullptr) { - doubleBuffer_.reset(new DoubleBuffer(this, useGpu_)); - } - useGpu_ = false; // Avoid D2D copy, it will delay the computing performance -} - -SimpleDataProviderBase::SimpleDataProviderBase(const DataConfig& config, - bool useGpu, - bool withInfo) - : DataProvider(config, useGpu) { - /* initialize the size of a sample, and the buffer */ - sampleDim_ = config_.feat_dim() * (2 * config_.context_len() + 1); - bufferCapacity_ = config_.buffer_capacity(); - withInfo_ = withInfo; - sampleNumInBuf_ = 0; - nextItemIndex_ = 0; - - /* malloc buffer in cpu */ - hInputDataBuf_ = std::make_shared(bufferCapacity_, sampleDim_); - hInputLabelBuf_ = std::make_shared(bufferCapacity_); - hInputInfoBuf_ = std::make_shared(bufferCapacity_); -} - -void SimpleDataProviderBase::shuffle() { - int i, t; - int len = sampleNumInBuf_; - std::vector temp(sampleDim_); - real* data = hInputDataBuf_->getData(); - int* label = hInputLabelBuf_->getData(); - int* info = hInputInfoBuf_->getData(); - int sampleSz = sizeof(real) * sampleDim_; - for (i = 0; i < len; i++) { - int randNum = rand(); // NOLINT TODO(yuyang18): Use rand_r instead? - t = randNum % (len - i) + i; - // swap - if (i != t) { - // swap data - memcpy(&temp[0], &data[i * sampleDim_], sampleSz); - memcpy(&data[i * sampleDim_], &data[t * sampleDim_], sampleSz); - memcpy(&data[t * sampleDim_], &temp[0], sampleSz); - std::swap(label[i], label[t]); - if (withInfo_) { - std::swap(info[i], info[t]); - } - } - } -} - -int64_t SimpleDataProviderBase::getNextBatchInternal(int64_t size, - DataBatch* batch) { - CHECK(batch != NULL); - batch->clear(); - - int64_t startIndex; - int64_t cpySize; - - std::lock_guard guard(lock_); - if (sampleNumInBuf_ - nextItemIndex_ < size) { - int64_t n = fillBuffer(); - VLOG(1) << "fillBuffer return " << n << " samples.\n"; - } - - startIndex = nextItemIndex_; - cpySize = std::min(size, sampleNumInBuf_ - nextItemIndex_); - nextItemIndex_ += cpySize; - - if (cpySize > 0) { - real* data = hInputDataBuf_->getData() + startIndex * sampleDim_; - int* label = hInputLabelBuf_->getData() + startIndex; - int* info = hInputInfoBuf_->getData() + startIndex; - - MatrixPtr& dataBatch = *dataBatch_; // get the thread local object - IVectorPtr& labelBatch = *labelBatch_; // get the thread local object - IVectorPtr& infoBatch = *infoBatch_; // get the thread local object - if (!dataBatch) { - dataBatch = Matrix::create(cpySize, sampleDim_, false, useGpu_); - labelBatch = IVector::create(cpySize, useGpu_); - if (withInfo_) { - infoBatch = IVector::create(cpySize, 0); - } - } else { - dataBatch->resize(cpySize, sampleDim_); - labelBatch->resize(cpySize); - if (withInfo_) { - infoBatch->resize(cpySize); - } - } - dataBatch->copyFrom(data, cpySize * sampleDim_); - labelBatch->copyFrom(label, cpySize); - batch->appendData(dataBatch); - batch->appendLabel(labelBatch); - if (withInfo_) { - infoBatch->copyFrom(info, cpySize); - batch->appendLabel(infoBatch); - } - } - - batch->setSize(cpySize); - return cpySize; -} - -void SimpleDataProviderBase::reset() { - sampleNumInBuf_ = 0; - nextItemIndex_ = 0; - DataProvider::reset(); -} - -int64_t SimpleDataProviderBase::getSize() { - LOG(FATAL) << "Currently, not implemented"; - return 0; -} - -int64_t SimpleDataProviderBase::fillBuffer() { - int64_t n = sampleNumInBuf_ - nextItemIndex_; - - /* flash the remaining data to the beginning of the buffer */ - if (n > 0) { - hInputDataBuf_->copyFrom( - hInputDataBuf_->getData() + nextItemIndex_ * sampleDim_, - n * sampleDim_); - hInputLabelBuf_->copyFrom(hInputLabelBuf_->getData() + nextItemIndex_, n); - if (withInfo_) { - hInputInfoBuf_->copyFrom(hInputInfoBuf_->getData() + nextItemIndex_, n); - } - } - - sampleNumInBuf_ = - n + fillBufferImp(hInputDataBuf_->getData() + n * sampleDim_, - hInputLabelBuf_->getData() + n, - hInputInfoBuf_->getData() + n, - bufferCapacity_ - n); - - /* for stachastic gradient training */ - if (!skipShuffle_) { - shuffle(); - } - - nextItemIndex_ = 0; - - return sampleNumInBuf_; -} - -SimpleDataProvider::SimpleDataProvider(const DataConfig& config, bool useGpu) - : SimpleDataProviderBase(config, useGpu, /* withInfo= */ false), - currentSampleIndex_(0) { - loadData(config_.files()); -} - -SimpleDataProvider::~SimpleDataProvider() {} - -int64_t SimpleDataProvider::fillBufferImp(real* data, - int* label, - int* info, - int64_t size) { - (void)info; - int64_t n = std::min(labels_.size() - currentSampleIndex_, size); - memcpy(data, - &data_[currentSampleIndex_ * sampleDim_], - n * sampleDim_ * sizeof(real)); - memcpy(label, &labels_[currentSampleIndex_], sizeof(int) * n); - currentSampleIndex_ += n; - - return n; -} - -void SimpleDataProvider::reset() { - currentSampleIndex_ = 0; - SimpleDataProviderBase::reset(); -} - -void SimpleDataProvider::loadData(const std::string& fileName) { - std::ifstream is(fileName); - CHECK(is) << "Fail to open " << fileName; - std::string line; - while (is) { - if (!getline(is, line)) break; - LOG(INFO) << "load data file " << line; - loadDataFile(line); - } - LOG(INFO) << "read done, num of instance=" << labels_.size() - << " data size=" << data_.size(); -} - -void SimpleDataProvider::loadDataFile(const std::string& fileName) { - std::ifstream is(fileName); - std::string line; - std::vector pieces; - while (is) { - if (!getline(is, line)) break; - str::split(line, ' ', &pieces); - CHECK_EQ((uint64_t)(sampleDim_ + 1), pieces.size()) - << " Dimension mismatch, " << pieces.size() - 1 << " in " << fileName - << " " << sampleDim_ << " from config"; - labels_.push_back(atoi(pieces[0].c_str())); - for (int i = 0; i < sampleDim_; ++i) { - data_.push_back(atof(pieces[i + 1].c_str())); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/dataproviders/DataProvider.h b/paddle/gserver/dataproviders/DataProvider.h deleted file mode 100644 index 21822b10c2ebf1d353195794cf8f49e02b64c177..0000000000000000000000000000000000000000 --- a/paddle/gserver/dataproviders/DataProvider.h +++ /dev/null @@ -1,480 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include -#include -#include -#include -#include -#include -#include - -#include "DataConfig.pb.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" -#include "paddle/math/Vector.h" -#include "paddle/parameter/Argument.h" -#include "paddle/utils/ClassRegistrar.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Queue.h" -#include "paddle/utils/ThreadLocal.h" -#include "paddle/utils/Util.h" - -namespace paddle { -/** - * @def REGISTER_DATA_PROVIDER - * @brief Macro for registering a data provider. The class type should contain - * a consturctor with parameter (DataConfig, bool). - */ -#define REGISTER_DATA_PROVIDER(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name([]() { \ - DataProvider::registrar_.registerClass( \ - #__type_name, \ - [](DataConfig conf, ModelConfig, bool useGpu) -> DataProvider* { \ - DataProvider* dp = new __class_name(conf, useGpu); \ - return dp; \ - }); \ - }) - -/** - * @def REGISTER_DATA_PROVIDER_EX - * @brief Macro for registering a data provider, which contains a constructor - * with parameter (DataConfig, ModelConfig, bool). - */ -#define REGISTER_DATA_PROVIDER_EX(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name([] { \ - DataProvider::registrar_.registerClass<__class_name>(#__type_name); \ - }) - -class DataBatch; -class BufferBatch; -typedef std::shared_ptr DataBatchPtr; -typedef std::shared_ptr BufferBatchPtr; -/** - * @brief Data for batch training a neural network - */ -class DataBatch { - public: - DataBatch() : size_(0) { data_.clear(); } - /** - * @brief Get batch size - * @return batch size - */ - int64_t getSize() const { return size_; } - /** - * @brief Get num of sequences of sequence data - * @return num of sequences - */ - int64_t getNumSequences() const { - if (data_.empty()) return size_; - return data_[0].sequenceStartPositions - ? data_[0].sequenceStartPositions->getSize() - 1 - : size_; - } - /** - * @brief Set batch size - * @param[in] size size - */ - void setSize(int64_t size) { size_ = size; } - /** - * @brief Get size of argument vector - * @return size of argument vector - * @note For usual supervised learning, input data and label is needed, - * then there will be two argument. - */ - int64_t getNumStreams() const { return data_.size(); } - - /** - * @brief Get a argument with index i - * @param[in] i index in argument vector - * @return a argument with index i - */ - const Argument& getStream(int i) const { return data_[i]; } - /** - * @brief Get all argument - * @return an argument vector - */ - std::vector& getStreams() { return data_; } - /** - * @brief Get all argument const - * @return an argument vector - */ - std::vector getStreams() const { return data_; } - /** - * @brief Clear DataBatch - */ - void clear() { - data_.clear(); - size_ = 0; - } - - /** - * @brief Append data to DataBatch - * @param[in] data matrix data - * @note The order in which each data stream is appended must match the order - * specified in stream_names of DataConfig. The stream_names can be obtained - * using DataProvider::getStreamNames(). - */ - void appendData(MatrixPtr data) { - Argument argu; - argu.value = data; - data_.push_back(argu); - } - - /** - * @brief Append sequence data to DataBatch - * @param[in] data matrix data - * @param[in] sequenceStartPositions sequence data - * @note The order in which each data stream is appended must match the order - * specified in stream_names of DataConfig. The stream_names can be obtained - * using DataProvider::getStreamNames(). - */ - void appendData(const MatrixPtr& data, - const ICpuGpuVectorPtr& sequenceStartPositions) { - Argument argu; - argu.value = data; - argu.sequenceStartPositions = sequenceStartPositions; - data_.push_back(argu); - } - /** - * @brief Append label data - * @param[in] label label data - * @param[in] value matrix data, default null - */ - void appendLabel(IVectorPtr label, MatrixPtr value = nullptr) { - Argument argu; - argu.ids = label; - argu.value = value; - data_.push_back(argu); - } - - /* - * @brief Append argument - * @param[in] argus DataBatch.getStreams() - * @param[in] size DataBatch.getSize() - * @param[in] dataId sub dataprovider id (in MultiDataProvider) - */ - void appendArguments(const std::vector& argus, - int size, - int dataId) { - size_ += size; - for (const auto& argu : argus) { - data_.push_back(argu); - data_.back().dataId = dataId; - } - } - - protected: - /** - * @brief batch size - */ - int64_t size_; - /** - * @brief A batch data consist of a Argument vector, - * An argument corresponds to a type of input data. - */ - std::vector data_; -}; - -class BufferBatch { - public: - BufferBatch() { - hlStream_ = HPPL_STREAM_DEFAULT; - hlEvent_ = NULL; - batchData_ = NULL; - } - ~BufferBatch() { - if (hlEvent_) { - hl_destroy_event(hlEvent_); - hlEvent_ = NULL; - } - delete batchData_; - batchData_ = NULL; - } - - void setDataBatch(DataBatch* batchData) { batchData_ = batchData; } - DataBatch* getDataBatch() { return batchData_; } - - void setCuStream(hl_stream_t stream) { hlStream_ = stream; } - hl_stream_t getCuStream() const { return hlStream_; } - - void setCuEvent(hl_event_t event) { hlEvent_ = event; } - - hl_event_t getCuEvent() const { return hlEvent_; } - - void createCuEvent() { - if (!hlEvent_) { - hlStream_ = HPPL_STREAM_1; - hl_create_event(&hlEvent_); - } - } - - void syncEvent() { - if (hlEvent_) { - hl_stream_wait_event(hlStream_, hlEvent_); - } - } - - void swap(BufferBatch* bufBatch); - void clone(DataBatch* srcBatch, bool useGpu); - - protected: - DataBatch* batchData_; - hl_stream_t hlStream_; - hl_event_t hlEvent_; -}; - -class DataProvider; -typedef std::shared_ptr DataProviderPtr; - -typedef Queue BufferBatchQueue; - -class DoubleBuffer { - public: - DoubleBuffer(DataProvider* dataPool, bool useGpu, int64_t batchSize = 0); - virtual ~DoubleBuffer(); - void removeOneBatch(DataBatch* dataBatch); - - void setBatchSize(int64_t newBatchSize) { batchSize_ = newBatchSize; } - - int64_t getBatchSize() { return batchSize_; } - - void startAsyncLoad(); - void finishAsyncLoad() { - stopping_ = true; - taskReadySem_.post(); - if (asyncLoader_) { - asyncLoader_->join(); - } - } - - void setPending(bool pending) { pending_ = pending; } - - protected: - virtual void asyncLoadBatch(); - void insertOneBatch(DataBatch* batch); - - DataProvider* dataPool_; - bool useGpu_; - int32_t batchSize_; - ThreadLocal usingBatch_; - BufferBatchQueue* dataQueue_; - BufferBatchQueue* bufferQueue_; - std::unique_ptr asyncLoader_; - Semaphore taskReadySem_; - bool stopping_; - bool pending_; -}; - -/** - * @brief Base class for DataProvider, which supplies data for training - * @note It can supplies multiple streams of data. - * For typical supervised training, there are two streams: - * one is for input, one is for label. - */ -class DataProvider { - public: - static ClassRegistrar registrar_; - static DataProvider* create(const DataConfig& config, - const ModelConfig& modelConfig, - bool useGpu = FLAGS_use_gpu); - - /** - * @brief create only used for unittest. - */ - inline static DataProvider* create(const DataConfig& config, - bool useGpu = FLAGS_use_gpu) { - return create(config, ModelConfig(), useGpu); - } - - DataProvider(const DataConfig& config, bool useGpu) - : config_(config), - skipShuffle_(false), - usageRatio_(config.usage_ratio()), - useGpu_(useGpu) { - if (config_.async_load_data()) { - initAsyncLoader(); - } - } - virtual ~DataProvider() {} - - const DataConfig& getConfig() const { return config_; } - - void setSkipShuffle() { skipShuffle_ = true; } - - /** - * @brief Get next batch of training samples - * @param[in] size size of training samples to get - * @param[out] batch a batch of training samples - * @return actual size of obtained training samples - */ - int64_t getNextBatch(int64_t size, DataBatch* batch); - - /** - * @brief Shuffle the data set - */ - virtual void shuffle() = 0; - - /** - * @brief reset all the value of index - * @note reset() must be called before any calls to getNextBatch() - * IMPORTANT: subclass reset() should always call the base class reset() - * at the end of the function - */ - virtual void reset() { - if (doubleBuffer_ != nullptr) { - doubleBuffer_->startAsyncLoad(); - } - } - - /** - * @brief Get the size of training samples - * @return the number of training samples in the data set. - * @note return -1 to indicate unlimited number of samples. - */ - virtual int64_t getSize() = 0; - - /** - * @brief Get next batch training samples internally - * @param[in] size size of training samples to get - * @param[out] batch a batch of training samples - * @return actual size of obtained training samples - */ - virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch) = 0; - - protected: - DataConfig config_; - bool skipShuffle_; - float usageRatio_; - bool useGpu_; - std::unique_ptr doubleBuffer_; - ThreadLocal> constantSlots_; - /** - * @@brief Get next batch training samples from buffer - * @param[in] size size of training samples to get - * @param[out] batch a batch of training samples - * @return actual size of obtained training samples - */ - int64_t getNextBatchFromBuffer(int64_t size, DataBatch* batch); - - void initAsyncLoader(); -}; - -/** - * A data provider which does nothing. It only serves as providing - * necessary configurations such as stream_names - */ -class DummyDataProvider : public DataProvider { - public: - DummyDataProvider(const DataConfig& config, bool useGpu) - : DataProvider(config, useGpu) {} - virtual void shuffle() {} - virtual void reset() { DataProvider::reset(); } - virtual int64_t getSize() { return 0; } - virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch) { - (void)size; - (void)batch; - return 0; - } -}; - -/** - * Data provider for one input and one integer label. - */ -class SimpleDataProviderBase : public DataProvider { - protected: - /// sample feature dimension - int64_t sampleDim_; - /// the number of samples - int64_t bufferCapacity_; - int64_t sampleNumInBuf_; - /// next item to read in buffer - int64_t nextItemIndex_; - /// some user defined info for validation - bool withInfo_; - - /// data buffer: bufferCapacity_ * nDataDim_ - CpuMatrixPtr hInputDataBuf_; - - /// label buffer:bufferCapacity_ * 1 - CpuIVectorPtr hInputLabelBuf_; - - /// info buffer:bufferCapacity_ * 1 - CpuIVectorPtr hInputInfoBuf_; - - ThreadLocal dataBatch_; - ThreadLocal labelBatch_; - ThreadLocal infoBatch_; - - RWLock lock_; - - public: - SimpleDataProviderBase(const DataConfig& config, bool useGpu, bool withInfo); - ~SimpleDataProviderBase() {} - - void shuffle(); - - virtual void reset(); - - virtual int64_t getSize(); - - virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); - - /// return the number of samples in the buffer - int64_t fillBuffer(); - - protected: - /** - * @brief Fill at most size samples into data and label. - * - * Each input is stored in contiguous memory locations in data. - * - * data[n * sampleDim_] .. data[n * sampleDim_ + sampleDim_ - 1] is for - * the input of the n-th sample. - * - * label[n] is the label for the n-th sample. - */ - virtual int64_t fillBufferImp(real* data, - int* label, - int* info, - int64_t size) = 0; -}; - -class SimpleDataProvider : public SimpleDataProviderBase { - public: - SimpleDataProvider(const DataConfig& config, bool useGpu); - ~SimpleDataProvider(); - virtual void reset(); - - protected: - void loadData(const std::string& fileName); - void loadDataFile(const std::string& fileName); - virtual int64_t fillBufferImp(real* data, - int* label, - int* info, - int64_t size); - - protected: - size_t currentSampleIndex_; - std::vector labels_; - std::vector data_; -}; - -} // namespace paddle diff --git a/paddle/gserver/dataproviders/MultiDataProvider.cpp b/paddle/gserver/dataproviders/MultiDataProvider.cpp deleted file mode 100644 index f71947ef3946284b7ecfb50851100fe43bd78857..0000000000000000000000000000000000000000 --- a/paddle/gserver/dataproviders/MultiDataProvider.cpp +++ /dev/null @@ -1,122 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MultiDataProvider.h" -#include -#include "paddle/utils/Logging.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -using namespace std; - -MultiDataProvider::MultiDataProvider(const DataConfig& config, - const ModelConfig& modelConfig, - bool useGpu) - : DataProvider(config, useGpu) { - bool atLeastOneMainDataFlag = false; - totalDataRatio_ = 0; - LOG(INFO) << "MultiDataProvider: sub data provider size: " - << config.sub_data_configs_size(); - LOG(INFO) << "MultiDataProvider: for_test: " << config.for_test(); - isTestMode_ = config.for_test(); - for (int i = 0; i < config.sub_data_configs_size(); i++) { - LOG(INFO) << "dataRatio of sub(" << i - << ") is: " << config.sub_data_configs(i).data_ratio(); - totalDataRatio_ += config.sub_data_configs(i).data_ratio(); - if (config.sub_data_configs(i).is_main_data()) { - LOG(INFO) << "main data is [" << i << "]"; - atLeastOneMainDataFlag = true; - } - } - CHECK(atLeastOneMainDataFlag) << "all sub dataproviders in MultiData do not" - << " have is_main_data flag"; - LOG(INFO) << "totalDataRatio_=" << totalDataRatio_; - DataConfig subConfig; - int subDataProviderCount = config.sub_data_configs_size(); - if (isTestMode()) { - LOG(INFO) << "construct MultiDataProvider in test mode"; - } else { - LOG(INFO) << "construct MultiDataProvider in train mode"; - } - subDataProviders_.resize(subDataProviderCount); - for (int i = 0; i < subDataProviderCount; i++) { - subConfig = config.sub_data_configs(i); - if (subConfig.async_load_data()) { - LOG(INFO) << "can not use async_load_data in sub dataprovider of " - "MultiDataProvider"; - subConfig.set_async_load_data(false); - } - subDataProviders_[i] = std::unique_ptr( - DataProvider::create(subConfig, modelConfig, useGpu_)); - } -} - -void MultiDataProvider::reset() { - for (auto& elem : subDataProviders_) { - elem->reset(); - } - DataProvider::reset(); -} - -void MultiDataProvider::shuffle() { - for (auto& elem : subDataProviders_) { - elem->shuffle(); - } -} - -int64_t MultiDataProvider::getNextBatchInternal(int64_t size, - DataBatch* batch) { - batch->clear(); - for (size_t i = 0; i < subDataProviders_.size(); ++i) { - // calc size according to data ratio - int64_t subSize = - (int64_t)(1.0 * size * config_.sub_data_configs(i).data_ratio() / - totalDataRatio_); - DataBatch subBatch; - int64_t realSize = - subDataProviders_[i]->getNextBatchInternal(subSize, &subBatch); - if (realSize == 0) { - // current subDataProvider has no data - if (!isTestMode()) { - // in train mode - if (config_.sub_data_configs(i).is_main_data()) { - // is main data provider. then return 0 - batch->clear(); - return 0; - } else { - // not main data provider, reset current subDataProvider and try again - subDataProviders_[i]->reset(); - subBatch.clear(); - realSize = - subDataProviders_[i]->getNextBatchInternal(subSize, &subBatch); - CHECK_GT(realSize, 0); - } - } else { - // in test mode, make an empty argument - Argument emptyArgu; - std::vector argus; - argus.push_back(emptyArgu); - batch->appendArguments(argus, 0, -1); - continue; - } - } - batch->appendArguments(subBatch.getStreams(), subBatch.getSize(), i); - } - return batch->getSize(); -} - -REGISTER_DATA_PROVIDER_EX(multi, MultiDataProvider); - -} // namespace paddle diff --git a/paddle/gserver/dataproviders/PyDataProvider.cpp b/paddle/gserver/dataproviders/PyDataProvider.cpp deleted file mode 100644 index dadf1b4cf27f248c7353aaad50dc22d4f6431cca..0000000000000000000000000000000000000000 --- a/paddle/gserver/dataproviders/PyDataProvider.cpp +++ /dev/null @@ -1,498 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PyDataProvider.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -#ifndef PADDLE_NO_PYTHON -REGISTER_DATA_PROVIDER(py, PyDataProvider); -#endif - -PyDataProvider::PyDataProvider(const DataConfig& config, - bool useGpu, - bool loadDataAll) - : DataProvider(config, useGpu), batchSize_(0) { - PyGuard guard; - pyModuleName_ = config_.load_data_module(); - pyClassName_ = config_.load_data_object(); - if (config_.load_data_args() != "") { - pyUserArgs_["load_data_args"] = config_.load_data_args(); - } - - if (loadDataAll) { - std::vector fileList; - if (!config_.files().empty()) { - loadFileList(config_.files(), fileList); - } - loadData(fileList); - } -} - -void PyDataProvider::loadData(const std::vector& fileList) { - VLOG(1) << "module:" << pyModuleName_ << " class:" << pyClassName_; - classInstance_ = - createPythonClass(pyModuleName_, pyClassName_, fileList, pyUserArgs_); - CHECK(classInstance_) << "Create class instance failed."; - PyObjectPtr obj(PyObject_CallMethod( - classInstance_.get(), const_cast("getHeader"), NULL)); - CHECK_PY(obj) << "Call function getHeader failed."; - std::string headerInfo = - std::string(PyString_AsString(obj.get()), PyString_Size(obj.get())); - parseHeaderData(headerInfo); - feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); -} - -void PyDataProvider::parseHeaderData(const std::string& headerData) { - char* pHeader = const_cast(headerData.c_str()); - char* pHeaderEnd = pHeader + headerData.size(); - slotNum_ = readT(pHeader, pHeaderEnd); - unsigned int useSequenceFlag = readT(pHeader, pHeaderEnd); - isIID_ = useSequenceFlag != 1; - slots_.clear(); - slots_.reserve(slotNum_); - for (size_t i = 0; i < slotNum_; ++i) { - unsigned int slotType = readT(pHeader, pHeaderEnd); - unsigned int slotDim = readT(pHeader, pHeaderEnd); - slots_.emplace_back(); - slots_.back().dim = slotDim; - slots_.back().type = static_cast(slotType); - } -} - -void PyDataProvider::resetSlots() { - for (auto& slot : slots_) { - slot.indexData.clear(); - slot.denseData.clear(); - slot.sparseNonValueData.clear(); - slot.sparseFloatValueData.clear(); - slot.indices.clear(); - slot.sequenceStartPositions.clear(); - slot.sampleSequenceIdVec.clear(); - slot.subSequenceStartPositions.clear(); - slot.strData.clear(); - } -} - -void PyDataProvider::fillDenseSlot(ProtoSlot& slot, - char*& data, - const char* dataEnd) { - unsigned int dim = slot.dim; - slot.sampleNum = readT(data, dataEnd); - slot.denseData.resize(slot.sampleNum * dim); -#ifdef PADDLE_TYPE_DOUBLE - CHECK_LE(data + sizeof(real) * dim * slot.sampleNum, dataEnd) - << "std::copy data is out of range"; - // PyDataProvider always provide data in float - float* dat = reinterpret_cast(data); - std::copy(dat, dat + slot.sampleNum * dim, slot.denseData.begin()); -#else - memcpyWithCheck(slot.denseData.data(), - data, - sizeof(real) * dim * slot.sampleNum, - dataEnd); -#endif - // PyDataProvider always provide data in float - data += sizeof(float) * dim * slot.sampleNum; -} - -void PyDataProvider::fillSparseNonValueSlot(ProtoSlot& slot, - char*& data, - const char* dataEnd) { - slot.sampleNum = readT(data, dataEnd); - unsigned int* indexPtr = (unsigned int*)data; - CHECK_LE(data + sizeof(unsigned int) * slot.sampleNum, dataEnd) - << "Vector assign value is out of range"; - slot.indices.assign(indexPtr, indexPtr + slot.sampleNum); - data += sizeof(unsigned int) * slot.sampleNum; - unsigned int length = 0; - length = readT(data, dataEnd); - slot.indices.push_back(length); - slot.sparseNonValueData.resize(length); - memcpyWithCheck(slot.sparseNonValueData.data(), - data, - sizeof(unsigned int) * length, - dataEnd); - data += sizeof(unsigned int) * length; -} - -void PyDataProvider::fillSparseValueSlot(ProtoSlot& slot, - char*& data, - const char* dataEnd) { - slot.sampleNum = readT(data, dataEnd); - unsigned int* indexPtr = (unsigned int*)data; - CHECK_LE(data + sizeof(unsigned int) * slot.sampleNum, dataEnd) - << "Vector assign value is out of range"; - slot.indices.assign(indexPtr, indexPtr + slot.sampleNum); - data += sizeof(unsigned int) * slot.sampleNum; - unsigned int length = 0; - length = readT(data, dataEnd); - unsigned int* colPtr = reinterpret_cast(data); - CHECK_LE(data + sizeof(unsigned int) * length, dataEnd) - << "Data is out of range"; - data += sizeof(unsigned int) * length; - size_t colLen = readT(data, dataEnd); - CHECK_EQ(colLen, length); - float* valuePtr = reinterpret_cast(data); - CHECK_LE(data + sizeof(real) * length, dataEnd) << "Data is out of range"; - data += sizeof(real) * length; - slot.indices.push_back(length); - slot.sparseFloatValueData.resize(length); - for (unsigned int ii = 0; ii < length; ++ii) { - slot.sparseFloatValueData[ii].col = colPtr[ii]; - slot.sparseFloatValueData[ii].value = valuePtr[ii]; - } -} - -void PyDataProvider::fillIndexSlot(ProtoSlot& slot, - char*& data, - const char* dataEnd) { - slot.sampleNum = readT(data, dataEnd); - CHECK_LE(data + sizeof(unsigned int) * slot.sampleNum, dataEnd) - << "Vector assign is out of range"; - slot.indexData.assign(reinterpret_cast(data), - reinterpret_cast(data) + slot.sampleNum); - data += sizeof(unsigned int) * slot.sampleNum; -} - -void PyDataProvider::fillStringSlot(ProtoSlot& slot, - char*& data, - const char* dataEnd) { - slot.sampleNum = readT(data, dataEnd); - for (unsigned int i = 0; i < slot.sampleNum; ++i) { - size_t len = readT(data, dataEnd); - auto str_begin = data; - data += len; - CHECK_LE(data, dataEnd) << "Data is out of range"; - slot.strData.emplace_back(str_begin, len); - } -} - -void PyDataProvider::fillSlotsByStr(const std::string& samples) { - char* data = const_cast(samples.c_str()); - char* dataEnd = data + samples.size(); - batchSize_ = readT(data, dataEnd); - if (0 == batchSize_) { - return; - } - - for (size_t j = 0; j < slotNum_; ++j) { - auto& slot = slots_[j]; - CHECK(SlotDef::INDEX >= slot.type || SlotDef::STRING == slot.type) - << " Slot type:" << slot.type << " is out of range."; - CHECK_GE(slot.type, SlotDef::VECTOR_DENSE) << " Slot type:" << slot.type - << " is out of range."; - switch (slot.type) { - case SlotDef::VECTOR_DENSE: - fillDenseSlot(slot, data, dataEnd); - break; - case SlotDef::VECTOR_SPARSE_NON_VALUE: - fillSparseNonValueSlot(slot, data, dataEnd); - break; - case SlotDef::VECTOR_SPARSE_VALUE: - fillSparseValueSlot(slot, data, dataEnd); - break; - case SlotDef::INDEX: - fillIndexSlot(slot, data, dataEnd); - break; - case SlotDef::VAR_MDIM_DENSE: - LOG(FATAL) << "Not implemented"; - break; - case SlotDef::VAR_MDIM_INDEX: - LOG(FATAL) << "Not implemented"; - break; - case SlotDef::STRING: - fillStringSlot(slot, data, dataEnd); - break; - } - } - // read sequenceStartPositions - for (size_t j = 0; j < slotNum_; ++j) { - auto& slot = slots_[j]; - if (!iidData()) { - unsigned int sequenceNum = readT(data, dataEnd); - slot.sequenceNum = sequenceNum; - for (size_t i = 0; i < sequenceNum; ++i) { - slot.sequenceStartPositions.push_back( - readT(data, dataEnd)); - } - for (size_t i = 0; i < sequenceNum; ++i) { - size_t begin = slot.sequenceStartPositions[i]; - size_t end = (i < sequenceNum - 1) ? slot.sequenceStartPositions[i + 1] - : slot.sampleNum; - for (size_t ii = begin; ii < end; ++ii) { - slot.sampleSequenceIdVec.push_back(ii); - } - } - } else { - for (size_t i = 0; i < slot.sampleNum; ++i) { - slot.sampleSequenceIdVec.push_back(i); - } - } - } - // read subSequenceStartPositions, not all slots have this infomation. - for (size_t j = 0; j < slotNum_; ++j) { - auto& slot = slots_[j]; - if (!iidData() && data != dataEnd) { - unsigned int subSequenceNum = readT(data, dataEnd); - slot.subSequenceNum = subSequenceNum; - for (size_t i = 0; i < subSequenceNum; ++i) { - slot.subSequenceStartPositions.push_back( - readT(data, dataEnd)); - } - } - } -} - -void PyDataProvider::reset() { - { // Invoke PyDataProvider Reset - PyGuard guard; - PyObjectPtr obj(PyObject_CallMethod( - classInstance_.get(), const_cast("reset"), NULL)); - CHECK_PY(obj) << "Call function reset failed."; - } - - if (!skipShuffle_) { - // Invoke PyDataProvider Shuffle - shuffle(); - } - DataProvider::reset(); -} - -void PyDataProvider::shuffle() { - // py shuffle - PyGuard guard; - PyObjectPtr obj(PyObject_CallMethod( - classInstance_.get(), const_cast("shuffle"), NULL)); - CHECK_PY(obj) << "Call function shuffle failed."; -} - -void PyDataProvider::handleDenseSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments) { - unsigned int dim = slot.dim; - Matrix::resizeOrCreate(cpuArguments[slotIndex].value, - slot.sampleNum, - dim, - false, // trans = false - false); // useGpu = false - real* buf = cpuArguments[slotIndex].value->getData(); - for (size_t i = 0; i < slot.sampleNum; ++i) { - memcpyWithCheck(buf + i * dim, - slot.denseData.data() + slot.sampleSequenceIdVec[i] * dim, - sizeof(real) * dim, - slot.denseData.data() + slot.denseData.size()); - } -} - -void PyDataProvider::handleSparseNonValueSlot( - ProtoSlot& slot, size_t slotIndex, std::vector& cpuArguments) { - unsigned int dim = slot.dim; - if (!(cpuArguments[slotIndex].value)) { - cpuArguments[slotIndex].value = - Matrix::createSparseMatrix(slot.sampleNum, - dim, - slot.sampleNum /*DEFAULT_AVG_WIDTH = 1*/, - NO_VALUE, - SPARSE_CSR, - false, - useGpu_); - } - auto mat = cpuArguments[slotIndex].value; - mat->resize(slot.sampleNum, dim, slot.sampleNum, NO_VALUE, SPARSE_CSR); - if (std::dynamic_pointer_cast(mat)) { - std::dynamic_pointer_cast(mat)->copyFrom( - slot.sampleSequenceIdVec.data(), - slot.indices.data(), - slot.sparseNonValueData.data(), - HPPL_STREAM_1); - } else if (std::dynamic_pointer_cast(mat)) { - std::dynamic_pointer_cast(mat)->copyFrom( - slot.sampleSequenceIdVec.data(), - slot.indices.data(), - slot.sparseNonValueData.data()); - } else { - LOG(FATAL) << "Not Supported"; - } -} - -void PyDataProvider::handleSparseValueSlot( - ProtoSlot& slot, size_t slotIndex, std::vector& cpuArguments) { - unsigned int dim = slot.dim; - if (!(cpuArguments[slotIndex].value)) { - cpuArguments[slotIndex].value = - Matrix::createSparseMatrix(slot.sampleNum, - dim, - slot.sampleNum /*DEFAULT_AVG_WIDTH = 1*/, - FLOAT_VALUE, - SPARSE_CSR, - false, - useGpu_); - } - auto mat = cpuArguments[slotIndex].value; - mat->resize(slot.sampleNum, dim, slot.sampleNum, FLOAT_VALUE, SPARSE_CSR); - if (std::dynamic_pointer_cast(mat)) { - std::dynamic_pointer_cast(mat)->copyFrom( - slot.sampleSequenceIdVec.data(), - slot.indices.data(), - slot.sparseFloatValueData.data(), - HPPL_STREAM_DEFAULT); - } else if (std::dynamic_pointer_cast(mat)) { - std::dynamic_pointer_cast(mat)->copyFrom( - slot.sampleSequenceIdVec.data(), - slot.indices.data(), - slot.sparseFloatValueData.data()); - } else { - LOG(FATAL) << "Not Supported"; - } -} - -void PyDataProvider::handleIndexSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments) { - IVector::resizeOrCreate(cpuArguments[slotIndex].ids, - slot.sampleNum, - /*useGpu_*/ false); - int* buf = cpuArguments[slotIndex].ids->getData(); - for (size_t i = 0; i < slot.sampleNum; ++i) { - buf[i] = slot.indexData[slot.sampleSequenceIdVec[i]]; - } -} - -void PyDataProvider::handleStringSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments) { - if (cpuArguments[slotIndex].strs) { - cpuArguments[slotIndex].strs->resize(slot.sampleNum); - } else { - cpuArguments[slotIndex].strs = - std::make_shared>(slot.sampleNum); - } - for (size_t i = 0; i < slot.sampleNum; ++i) { - (*cpuArguments[slotIndex].strs)[i] = - slot.strData[slot.sampleSequenceIdVec[i]]; - } -} - -int64_t PyDataProvider::getNextBatchInternal(int64_t size, DataBatch* batch) { - PyGuard guard; - PyObjectPtr obj(PyObject_CallMethod(classInstance_.get(), - const_cast("getNextBatch"), - const_cast("i"), - size)); - CHECK_PY(obj) << "Call function getNextBatch failed."; - const std::string& samples = - std::string(PyString_AsString(obj.get()), PyString_Size(obj.get())); - resetSlots(); - fillSlotsByStr(samples); - size = batchSize_; - if (size <= 0) return 0; - - DataBatch& cpuBatch = *cpuBatch_; - std::vector& cpuArguments = cpuBatch.getStreams(); - cpuBatch.setSize(size); - cpuArguments.resize(slotNum_); - - if (!iidData()) { - for (size_t j = 0; j < slotNum_; ++j) { - auto& slot = slots_[j]; - ICpuGpuVector::resizeOrCreate(cpuArguments[j].sequenceStartPositions, - slot.sequenceNum + 1, - /* useGpu= */ false); - int* buf = cpuArguments[j].sequenceStartPositions->getMutableData(false); - std::copy(slot.sequenceStartPositions.begin(), - slot.sequenceStartPositions.end(), - buf); - buf[slot.sequenceStartPositions.size()] = slot.sampleNum; - - if (slot.subSequenceStartPositions.size()) { - ICpuGpuVector::resizeOrCreate(cpuArguments[j].subSequenceStartPositions, - slot.subSequenceNum + 1, - /* useGpu= */ false); - int* buf = - cpuArguments[j].subSequenceStartPositions->getMutableData(false); - std::copy(slot.subSequenceStartPositions.begin(), - slot.subSequenceStartPositions.end(), - buf); - buf[slot.subSequenceNum] = slot.sampleNum; - // check subSequenceStartPositions and sequenceStartPositions - cpuArguments[j].checkSubset(); - } - } - } - - for (size_t slotIndex = 0; slotIndex < slotNum_; ++slotIndex) { - auto& slot = slots_[slotIndex]; - SlotDef::SlotType slotType = slot.type; - switch (slotType) { - case SlotDef::VECTOR_DENSE: - handleDenseSlot(slot, slotIndex, cpuArguments); - break; - case SlotDef::VECTOR_SPARSE_NON_VALUE: - handleSparseNonValueSlot(slot, slotIndex, cpuArguments); - break; - case SlotDef::VECTOR_SPARSE_VALUE: - handleSparseValueSlot(slot, slotIndex, cpuArguments); - break; - case SlotDef::INDEX: - handleIndexSlot(slot, slotIndex, cpuArguments); - break; - case SlotDef::VAR_MDIM_DENSE: - LOG(FATAL) << "Not implemented"; - break; - case SlotDef::VAR_MDIM_INDEX: - LOG(FATAL) << "Not implemented"; - break; - case SlotDef::STRING: - handleStringSlot(slot, slotIndex, cpuArguments); - break; - } - } - - if (useGpu_) { - std::vector& cpuArguments = cpuBatch.getStreams(); - DataBatch& gpuBatch = *gpuBatch_; - std::vector& gpuArguments = gpuBatch.getStreams(); - gpuArguments.resize(cpuArguments.size()); - gpuBatch.setSize(size); - for (size_t i = 0; i < slotNum_; ++i) { - SlotDef::SlotType slotType = slots_[i].type; - if (SlotDef::VECTOR_SPARSE_VALUE == slotType || - SlotDef::VECTOR_SPARSE_NON_VALUE == slotType) { - gpuArguments[i] = cpuArguments[i]; - gpuArguments[i].sequenceStartPositions = - cpuArguments[i].sequenceStartPositions; - - if (slots_[i].subSequenceStartPositions.size()) { - gpuArguments[i].subSequenceStartPositions = - cpuArguments[i].subSequenceStartPositions; - } - } else { - gpuArguments[i].resizeAndCopyFrom( - cpuArguments[i], useGpu_, HPPL_STREAM_1); - } - } - hl_stream_synchronize(HPPL_STREAM_1); - *batch = gpuBatch; - } else { - *batch = cpuBatch; - } - - return batch->getSize(); -} - -} // namespace paddle diff --git a/paddle/gserver/dataproviders/PyDataProvider.h b/paddle/gserver/dataproviders/PyDataProvider.h deleted file mode 100644 index da50dd4e2ebb743ef45af319bc713ed7ac3d3e10..0000000000000000000000000000000000000000 --- a/paddle/gserver/dataproviders/PyDataProvider.h +++ /dev/null @@ -1,124 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "DataFormat.pb.h" -#include "DataProvider.h" - -#include - -namespace paddle { - -class PyDataProvider : public DataProvider { - public: - PyDataProvider(const DataConfig& config, - bool useGpu, - bool loadDataAll = true); - - virtual void reset(); - - // Note this size includes the sequences which are skipped because they - // are longer than the batch size - virtual int64_t getSize() { - LOG(FATAL) << "Not implement yet"; - return -1; - } - virtual void shuffle(); - - virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); - - protected: - struct ProtoSlot; - // return false if each each sample is one sequence, i.e., independent - // of other samples. - inline bool iidData() const { return isIID_; } - - void parseHeaderData(const std::string& headerData); - void fillDenseSlot(ProtoSlot& slot, char*& data, const char* dataEnd); - void fillSparseNonValueSlot(ProtoSlot& slot, - char*& data, - const char* dataEnd); - void fillSparseValueSlot(ProtoSlot& slot, char*& data, const char* dataEnd); - void fillIndexSlot(ProtoSlot& slot, char*& data, const char* dataEnd); - void fillStringSlot(ProtoSlot& slot, char*& data, const char* dataEnd); - void fillSlotsByStr(const std::string& samples); - void handleDenseSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments); - void handleSparseNonValueSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments); - void handleSparseValueSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments); - void handleIndexSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments); - void handleStringSlot(ProtoSlot& slot, - size_t slotIndex, - std::vector& cpuArguments); - void resetSlots(); - void loadData(const std::vector& fileList); - - protected: - struct ProtoSlot { - SlotDef::SlotType type; - int dim; - unsigned int sampleNum; - unsigned int sequenceNum; - unsigned int subSequenceNum; - // Store the data of index type slot - std::vector indexData; - // Store the data of dense type slot - std::vector denseData; - // Store the data of sparseNonValue type slot - std::vector sparseNonValueData; - // Store the data of sparseValue type slot - std::vector sparseFloatValueData; - // Used to store the index of each sample in slot values - std::vector indices; - // The starting position of each sequence in samples - // The last element should be the number of samples - // If empty, each sample is one sequence. - std::vector sequenceStartPositions; - // The index id of sequences in slot - std::vector sampleSequenceIdVec; - // The starting position of each subsequence in samples - // The last element should be the number of subsequence - // If empty, each sequence of sample has no subsequence. - std::vector subSequenceStartPositions; - // Store the data of string type slot - std::vector strData; - }; - std::vector slots_; - - PyObjectPtr classInstance_; - unsigned int batchSize_; - unsigned int slotNum_; - // if use sequence, isIID_ equals false, otherwise it is true. - bool isIID_; - // The name of python module name - std::string pyModuleName_; - // The name of python class name - std::string pyClassName_; - // User args set in config - std::map pyUserArgs_; - - ThreadLocalD cpuBatch_; - ThreadLocalD gpuBatch_; -}; - -} // namespace paddle diff --git a/paddle/gserver/dataproviders/PyDataProvider2.cpp b/paddle/gserver/dataproviders/PyDataProvider2.cpp deleted file mode 100644 index 54ee091e8f257f76b113d4ca6f8a7c3989c0c1df..0000000000000000000000000000000000000000 --- a/paddle/gserver/dataproviders/PyDataProvider2.cpp +++ /dev/null @@ -1,1031 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifndef PADDLE_NO_PYTHON - -#include -#include -#include -#include -#include -#include -#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION -#include - -#include "DataProvider.h" - -#include "paddle/utils/Locks.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -namespace unittest { - -static std::unique_ptr> - OnPoolFilled; - -namespace pydp2 { - -void setOnPoolFilledHook(const std::function& callback) { - OnPoolFilled.reset(new std::function()); - *OnPoolFilled = callback; -} - -void clearOnPoolFilledHook() { OnPoolFilled.reset(); } - -} // namespace pydp2 -} // namespace unittest - -/** - * Slot type - */ -enum SlotType { - ST_DENSE = 0, - ST_NON_SPARSE_VALUE = 1, - ST_SPARSE_VALUE = 2, - ST_INDEX = 3 -}; - -/** - * Sequence type - */ -enum SeqType { SQT_NONE = 0, SQT_SEQ, SQT_SUBSEQ }; - -/** - * Cache Type. - */ -enum CacheType { - NO_CACHE = 0, // Each pass will load data from PyDataProvider2. - CACHE_PASS_IN_MEM = 1, // First pass will load data from PyDataProvider2, - // then cache all data in memory. Load data from - // memory in rest passes. -}; - -struct SlotHeader { // Slot Header will parse from python object's slots field. - size_t dim; - SlotType slotType; - SeqType seqType; -}; - -inline std::ostream& operator<<(std::ostream& os, const SlotHeader& header) { - os << "Dim = " << header.dim << " Type = " << header.slotType - << " SeqType = " << header.seqType; - return os; -} - -/** - * FieldScanner Interface. - * - * It will read python object, and fill to argument's each slot. - * There are two steps, prepare and fill. Scanner will alloc memory during - * prepare step, fill data into argument during fill step. - */ -class IFieldScanner { - public: - DISABLE_COPY(IFieldScanner); - /** - * Ctor. - * @param headerPtr slot header that scanner belong to. - */ - explicit IFieldScanner(SlotHeader* headerPtr) : headerPtr_(headerPtr) {} - virtual ~IFieldScanner() {} - - /** - * Start prepare step. - */ - virtual void startPrepare(Argument& argument) {} - - /** - * Prepare step. - * - * @note the obj could be a timestep of sample or whole sample. It depends - * what scanner it is. - */ - virtual void prepare(Argument& argument, PyObject* obj) {} - - /** - * Finish Prepare step. - */ - virtual void finishPrepare(Argument& argument) {} - - /** - * Start fill step. - */ - virtual void startFill(Argument& argument) {} - - /** - * Fill step. - * - * @note the obj could be a timestep of sample or whole sample. It depends - * what scanner it is. - */ - virtual void fill(Argument& argument, PyObject* obj) {} - - /** - * Finish fill step. - */ - virtual void finishFill(Argument& argument) {} - - /** - * Factory method. Create a scanner by header. The final scanner may be - * combine many scanners. - * - * @note Fatal if header is not support. - */ - static IFieldScanner* create(SlotHeader* header); - - protected: - SlotHeader* headerPtr_; -}; - -/** - * Py Data Provider Cache Interface. - */ -class IPyDataProviderCache { - public: - virtual ~IPyDataProviderCache() {} - - /** - * invoke when DataProvider::reset() - * @return true if read data from python. - */ - virtual bool reset() = 0; - - /** - * invoke when these data are used by DataProvider, and need to clear. - * @param [inout] data used data. - * - * @note The implemented class must clear these data array. Or if you want to - * delete the PyObjectPtr later, you should make sure the paddle process only - * have one active thread calling python code (use PyGuard otherwise). - */ - virtual void drop(std::deque* data) = 0; - - /** - * Return whole data in cache. - */ - virtual std::deque* load() = 0; - - /** - * Factory method. Convert CacheType to IPyDataProviderCache* - */ - static IPyDataProviderCache* create(CacheType ct); -}; - -/** - * PyDataProvider2. - * - * For usage, please refer python module 'paddle.trainer.PyDataProvider2' - * - * Here, we start a thread to read data. It is totally asynchronous for reading - * data. And it support cache strategies. - */ -class PyDataProvider2 : public DataProvider { - public: - /** - * Ctor - */ - PyDataProvider2(const DataConfig& config, - const ModelConfig& modelConfig, - bool useGpu) - : DataProvider(config, useGpu), callingContextCreated_(2) { - if (PyArray_API == NULL) import_array(); - auto& args = config.load_data_args(); - PyObjectPtr kwargs = PyObjectPtr(PyDict_New()); - if (!args.empty()) { - kwargs = callPythonFuncRetPyObj( - "paddle.trainer.PyDataProvider2", "deserialize_args", {args}); - } - - py::DictHelper kwargsDict(kwargs); - kwargsDict.setBool("is_train", !config.for_test()); - std::vector inputs; - inputs.reserve(modelConfig.input_layer_names().size()); - std::copy(modelConfig.input_layer_names().begin(), - modelConfig.input_layer_names().end(), - std::back_inserter(inputs)); - kwargsDict.setStringList("input_order", inputs); - - // kwargs is keyword arguemts to create object. - this->createPyDataObj(config.load_data_module(), - config.load_data_object(), - config.files(), - std::move(kwargs)); - DBG << "Instance " << instance_.get() << " loaded."; - this->readPyFields(config.for_test()); - DBG << "Py Field Done"; - } - - /** - * Dtor - * @note will stop loading thread when destructing - */ - virtual ~PyDataProvider2() { resetImpl(false); } - - private: - void createPyDataObj(const std::string& model, - const std::string& className, - const std::string& fileListName, - PyObjectPtr&& kwargs // NOLINT - ) { - LOG(INFO) << "loading dataprovider " << model << "::" << className; - - PyObjectPtr module = py::import(model); - PyObjectPtr moduleDict(PyModule_GetDict(module.get())); - CHECK_PY(moduleDict) << "Invoke module.__dict__ error"; - PyObjectPtr cls(PyDict_GetItemString(moduleDict.get(), className.c_str())); - CHECK_PY(cls) << "load class " << className.c_str() << "error"; - - // If there are multiple python instance share same module, the PyObjectPtr - // only for instance will make python reference-count error. - // - // So here, we increase reference count manually. - Py_XINCREF(module.get()); - Py_XINCREF(moduleDict.get()); - Py_XINCREF(cls.get()); - - PyObjectPtr fileListInPy = loadPyFileLists(fileListName); - PyDict_SetItemString(kwargs.get(), "file_list", fileListInPy.get()); - { - PyGuard guard; - instance_.reset(PyObject_Call(cls.get(), zeroTuple_.get(), kwargs.get())); - } - CHECK_PY(instance_) << "Cannot Create instance"; - } - - void readPyFields(bool testing) { - py::ObjectHelper self(this->instance_); - bool ok; - - this->skipShuffle_ = - !self.getBoolAttr("should_shuffle", &ok /*isBoolType*/); - if (!ok) { - this->skipShuffle_ = testing; // shuffle when is training, skip shuffle - // when is testing. - } - DBG << "Provider Skip Shuffle " << this->skipShuffle_; - - this->poolSize_ = self.getIntAttr("pool_size", &ok); - if (!ok) { - this->poolSize_ = -1UL; - } - this->minPoolSize_ = self.getIntAttr("min_pool_size", &ok); - if (!ok) { - this->minPoolSize_ = -1UL; - } - this->minPoolSize_ = std::min(this->poolSize_, this->minPoolSize_); - - this->canOverBatchSize_ = self.getBoolAttr("can_over_batch_size"); - - calcBatchSize_.reset(self.getAttr("calc_batch_size")); - if (this->calcBatchSize_ && !py::isCallable(this->calcBatchSize_)) { - this->calcBatchSize_.reset(); - } - - generator_.reset(self.getAttr("generator")); - CHECK(py::isCallable(generator_)); - - // Reading slots. - PyObjectPtr slotsPtr(self.getAttr("slots")); - py::SequenceHelper slots(slotsPtr); - headers_.reserve(slots.size()); - for (size_t i = 0; i < slots.size(); ++i) { - headers_.emplace_back(); - auto& header = headers_.back(); - PyObject* hdPtr = slots[i]; - CHECK(hdPtr != nullptr); - Py_XINCREF(hdPtr); - PyObjectPtr headerPtrWrap(hdPtr); - py::ObjectHelper hd(headerPtrWrap); - header.dim = hd.getIntAttrWithError("dim"); - header.seqType = (SeqType)hd.getIntAttrWithError("seq_type"); - header.slotType = (SlotType)hd.getIntAttrWithError("type"); - } - - DBG << "Data header size " << headers_.size(); - for (auto& header : headers_) { - DBG << header; - } - cache_.reset(IPyDataProviderCache::create( - (CacheType)self.getIntAttrWithError("cache"))); - } - - PyObjectPtr loadPyFileLists(const std::string& fileListName) { - loadFileList(fileListName, fileLists_); - PyObject* lst = PyList_New(fileLists_.size()); - for (size_t i = 0; i < fileLists_.size(); ++i) { - PyList_SET_ITEM(lst, i, PyString_FromString(fileLists_[i].c_str())); - } - return PyObjectPtr(lst); - } - - void loadThread() { - DBG << "Creating context"; - for (auto& filename : fileLists_) { - PyGuard g; - py::CallableHelper generator(this->generator_); - generator.setArgsSize(2); - generator.getArgs().set(0, instance_); - generator.getArgs().set(1, PyString_FromString(filename.c_str()), true); - callingContexts_.emplace_back(generator()); - CHECK_PY(callingContexts_.back()) << "Generator error."; - CHECK(PyIter_Check(callingContexts_.back())); - } - DBG << "Create context done"; - callingContextCreated_.wait(); - - PositionRandom p(skipShuffle_); - - while (!exit_ && !callingContexts_.empty()) { - PyObject* data = nullptr; - - { // Read data. - size_t cid = p(callingContexts_.size()); - bool atEnd; - data = py::iterNext(callingContexts_[cid], &atEnd); - if (atEnd || data == nullptr) { - if (cid != 0) { - std::swap(callingContexts_[cid], callingContexts_[0]); - cid = 0; - } - - PyObjectPtr front; - { - std::unique_lock l(mtx_); - front = pop_get_front(callingContexts_); - } - { - PyGuard g; - front.reset(); - } - this->pullCV_.notify_all(); - continue; - } - } - - size_t additionalBatchSize = 1; - if (calcBatchSize_) { - PyGuard guard; - py::CallableHelper calcBatchSize(this->calcBatchSize_); - calcBatchSize.setArgsSize(1); - calcBatchSize.getArgs().set(0, data); - PyObjectPtr bs(calcBatchSize()); - CHECK_PY(bs); - bool ok; - additionalBatchSize = py::castInt(bs.get(), &ok); - CHECK(ok) << "CalcBatchSize must return int or long"; - } - - if (this->loadThread_) { // wait poolActualSize < poolSize; - std::unique_lock l(mtx_); - pushCV_.wait(l, [this] { return this->poolActualSize_ < poolSize_; }); - } - - { - std::lock_guard guard(mtx_); - poolActualSize_ += additionalBatchSize; - dataPool_.emplace_back(data); - } - pullCV_.notify_all(); - } - DBG << "load thread end"; - } - - inline void resetImpl(bool startNewThread) { - DBG << "Reseting " << startNewThread; - exit_.store(true); - if (loadThread_) { // is loading. - loadThread_->join(); - loadThread_.reset(); - } - { - PyGuard g; - callingContexts_.clear(); - this->pullCV_.notify_one(); - } - - std::lock_guard guard(mutexForReset_); - { - PyGuard g; - dataPool_.clear(); - } - poolActualSize_ = 0; - - if (startNewThread && cache_->reset()) { - DBG << "Start new thread."; - loadThread_.reset(new std::thread([this] { - exit_ = false; - loadThread(); - })); - callingContextCreated_.wait(); - } - DBG << "Reset done"; - exit_ = false; - } - - private: - std::unique_ptr loadThread_; - std::atomic exit_; - std::deque callingContexts_; - std::deque dataPool_; - size_t poolActualSize_; - std::condition_variable pushCV_; - std::condition_variable pullCV_; - std::mutex mtx_; - - std::mutex mutexForReset_; - - ThreadBarrier callingContextCreated_; - std::unique_ptr cache_; - - PyObjectPtr instance_; - size_t poolSize_; - size_t minPoolSize_; - bool canOverBatchSize_; - PyObjectPtr calcBatchSize_; - PyObjectPtr generator_; - std::vector fileLists_; - std::vector headers_; - static PyObjectPtr zeroTuple_; - - class PositionRandom { - public: - inline explicit PositionRandom(bool skipRand) - : eng_(ThreadLocalRandomEngine::get()), skipRand_(skipRand) {} - - inline size_t operator()(size_t len) { - if (!skipRand_) { - if (!dist_ || dist_->b() != len - 1) { - dist_.reset(new std::uniform_int_distribution(0, len - 1)); - } - return (*dist_)(eng_); - } else { - return 0; - } - } - - private: - std::default_random_engine& eng_; - std::unique_ptr> dist_; - bool skipRand_; - }; - - // DataProvider interface - public: - /** - * Resetting the PyDataProvider. May start reading thread here. - */ - virtual void reset() { - resetImpl(true); - DataProvider::reset(); - } - - /** - * Shuffle. Do nothing because PyDataProvider do shuffle implicitly by random - * select data from datapool. - */ - void shuffle() {} - - /** - * Not limited size. - */ - int64_t getSize() { return -1; } - - /** - * Loading a batch of data. - */ - int64_t getNextBatchInternal(int64_t size_, DataBatch* batch) { - std::lock_guard guard(mutexForReset_); - REGISTER_TIMER("PyDP2.getNextBatchInternal") - CHECK_GE(size_, 0); - size_t size = (size_t)size_; - if (loadThread_) { // loading from thread should wait for data pool ready. - // but, loading from cache, cache object should ensure - // data pool ready. - std::unique_lock l(mtx_); - pullCV_.wait(l, [this, &size] { - return this->poolActualSize_ >= std::max(size, this->minPoolSize_) || - callingContexts_.empty(); - }); - - if (unittest::OnPoolFilled) { - (*unittest::OnPoolFilled)(this->poolActualSize_); - } - } - std::deque data; - size_t bsize = 0; - std::deque* poolPtr = nullptr; - - if (this->loadThread_) { // loading from thread. - poolPtr = &this->dataPool_; - } else { // loading from cache. - poolPtr = this->cache_->load(); - } - if (exit_) { - // PyDataProvider is destructing. - return 0; - } - CHECK(poolPtr != nullptr); - - std::deque& pool = *poolPtr; - - while (bsize < size && !pool.empty()) { - { - // move data from pool to data - std::lock_guard guard(mtx_); - if (skipShuffle_) { - size_t i = 0; - CHECK(pool[i] != nullptr); - data.emplace_back(std::move(pool[i])); - pool.pop_front(); - } else { // when shuffle, use swap to drop only last pool element. - size_t i = ThreadLocalRand::rand() % pool.size(); - CHECK(pool[i] != nullptr); - if (i != 0) { - std::swap(pool[i], pool.front()); - } - data.emplace_back(std::move(pool.front())); - pool.pop_front(); - } - - if (calcBatchSize_) { // custom calc batch size. - PyGuard guard; - Py_INCREF(data.back().get()); - py::CallableHelper calcBatchSize(calcBatchSize_); - calcBatchSize.setArgsSize(1); - calcBatchSize.getArgs().set(0, data.back()); - PyObjectPtr customBatchSize(calcBatchSize()); - bool ok; - size_t tmp = py::castInt(customBatchSize.get(), &ok); - CHECK(ok) << "calc_batch_size must return int"; - - if (bsize + tmp > size && !canOverBatchSize_) { - // Put data back. - pool.push_front(std::move(data.back())); - data.pop_back(); - break; - } else { - bsize += tmp; - } - } else { - bsize += 1; - } - } - } - - if (this->loadThread_) { - { - std::lock_guard g(mtx_); - poolActualSize_ -= bsize; - } - this->pushCV_.notify_all(); - } - - if (bsize == 0) { // end of pass. In data pool, cannot get any data. - return 0; - } - - DataBatch cpuBatch; - cpuBatch.setSize(bsize); - auto& inArgs = cpuBatch.getStreams(); - inArgs.resize(headers_.size()); - std::vector> scanners; - scanners.reserve(headers_.size()); - for (auto& header : headers_) { - scanners.emplace_back(IFieldScanner::create(&header)); - } - DBG << "Scanner created."; - for (size_t i = 0; i < headers_.size(); ++i) { - scanners[i]->startPrepare(inArgs[i]); - } - for (auto& d : data) { - py::SequenceHelper s(d); - for (size_t i = 0; i < headers_.size(); ++i) { - scanners[i]->prepare(inArgs[i], s[i]); - } - } - for (size_t i = 0; i < headers_.size(); ++i) { - scanners[i]->finishPrepare(inArgs[i]); - } - for (size_t i = 0; i < headers_.size(); ++i) { - scanners[i]->startFill(inArgs[i]); - } - for (auto& d : data) { - py::SequenceHelper s(d); - for (size_t i = 0; i < headers_.size(); ++i) { - scanners[i]->fill(inArgs[i], s[i]); - } - } - - for (size_t i = 0; i < headers_.size(); ++i) { - scanners[i]->finishFill(inArgs[i]); - } - - { - PyGuard g; - cache_->drop(&data); - } - - DBG << "Reading CPU Batch Done."; - - if (useGpu_) { - std::vector& cpuArguments = cpuBatch.getStreams(); - DataBatch& gpuBatch = *batch; - std::vector& gpuArguments = gpuBatch.getStreams(); - gpuArguments.resize(cpuArguments.size()); - gpuBatch.setSize(bsize); - for (size_t i = 0; i < headers_.size(); ++i) { - gpuArguments[i].resizeAndCopyFrom( - cpuArguments[i], useGpu_, HPPL_STREAM_1); - } - hl_stream_synchronize(HPPL_STREAM_1); - } else { - *batch = cpuBatch; - } - return bsize; - } -}; - -PyObjectPtr PyDataProvider2::zeroTuple_(PyTuple_New(0)); - -REGISTER_DATA_PROVIDER_EX(py2, PyDataProvider2); - -/** - * Scanner for dense slot. - */ -class DenseScanner : public IFieldScanner { - public: - explicit DenseScanner(SlotHeader* ptr) : IFieldScanner(ptr), height_(0) {} - - /** - * Prepare. - * @param argument target argument - * @param obj each timestep of a sample. - */ - virtual void prepare(Argument& argument, PyObject* obj) { ++height_; } - - virtual void finishPrepare(Argument& argument) { - Matrix::resizeOrCreate( - argument.value, height_, headerPtr_->dim, false, false); - height_ = 0; - } - - /** - * Fill argument from obj. - * @param argument - * @param obj - */ - virtual void fill(Argument& argument, PyObject* obj) { - real* dat = argument.value->getData() + height_ * headerPtr_->dim; - if (PyArray_Check(obj)) { - auto dtype = PyArray_DTYPE((PyArrayObject*)obj); - if (dtype->type == 'f' && dtype->elsize == sizeof(real)) { - real* data = (real*)PyArray_DATA((PyArrayObject*)obj); - auto sz = PyArray_SIZE((PyArrayObject*)obj); - std::copy(data, data + sz, dat); - } else { - LOG(FATAL) << "You should yield float" << sizeof(real) * 8 << " array"; - } - } else { - py::SequenceHelper s(obj); - // TODO(yuyang18): Here we can use AVX or SSE to accelerate memory copy. - for (size_t i = 0; i < headerPtr_->dim; ++i) { - dat[i] = (real)s.getDouble(i); - } - } - ++height_; - } - - private: - size_t height_; -}; - -/** - * Scanner for index slot - */ -class IndexScanner : public IFieldScanner { - public: - explicit IndexScanner(SlotHeader* ptr) : IFieldScanner(ptr), cnt_(0) {} - - /** - * Prepare memory space. - * - * @note obj is a single timestep of sample - */ - virtual void prepare(Argument& argument, PyObject* obj) { ++cnt_; } - - virtual void finishPrepare(Argument& argument) { - IVector::resizeOrCreate(argument.ids, cnt_, false); - cnt_ = 0; - } - - /** - * Fill one index to argument. - */ - virtual void fill(Argument& argument, PyObject* obj) { - bool ok; - argument.ids->getData()[cnt_++] = py::castInt(obj, &ok); - CHECK(ok) << "Cannot cast int " << py::repr(obj); - } - - private: - size_t cnt_; -}; - -class SparseNonValueScanner : public IFieldScanner { - public: - explicit SparseNonValueScanner(SlotHeader* ptr) - : IFieldScanner(ptr), nnz_(0), height_(0) {} - - /** - * Prepare memory space - * @note obj is a timestep of one sample. - */ - virtual void prepare(Argument& argument, PyObject* obj) { - ++height_; - nnz_ += py::SequenceHelper(obj).size(); - } - - virtual void finishPrepare(Argument& argument) { - Matrix::resizeOrCreateSparseMatrix( - argument.value, height_, headerPtr_->dim, nnz_, NO_VALUE); - } - - virtual void startFill(Argument& argument) { - auto smat = (CpuSparseMatrix*)(argument.value.get()); - smat->getRows()[0] = 0; - nnz_ = 0; - height_ = 1; - } - - /** - * Fill one sparse vector to argument. - * @note obj is a timestep of one sample. - */ - virtual void fill(Argument& argument, PyObject* obj) { - py::SequenceHelper s(obj); - auto sz = s.size(); - auto smat = (CpuSparseMatrix*)(argument.value.get()); - int* row = smat->getRows(); - int* col = smat->getCols(); - real* dat = smat->getData(); - row[height_] = row[height_ - 1] + (int)sz; - - for (decltype(sz) i = 0; i < sz; ++i) { - setData(col + nnz_, dat + nnz_, s[i]); - ++nnz_; - } - ++height_; - } - - protected: - /** - * Set a single sparse index and value. - * @param [out] col sparse index - * @param [out] dat sparse value - * @param [in] obj Python Object. For sparse_non_value is a PyInt or PyLong. - * For sparse_value is a Tuple (int, float). - */ - virtual void setData(int* col, real* dat, PyObject* obj) { - bool ok; - *col = py::castInt(obj, &ok); - CHECK(ok); - } - - size_t nnz_; - size_t height_; -}; - -class SparseValueScanner : public SparseNonValueScanner { - public: - explicit SparseValueScanner(SlotHeader* ptr) : SparseNonValueScanner(ptr) {} - - virtual void finishPrepare(Argument& argument) { - Matrix::resizeOrCreateSparseMatrix( - argument.value, height_, headerPtr_->dim, nnz_, FLOAT_VALUE); - } - - protected: - virtual void setData(int* col, real* dat, PyObject* obj) { - py::SequenceHelper s(obj); - SparseNonValueScanner::setData(col, dat, s[0]); - *dat = (real)s.getDouble(1); - } -}; - -/** - * Sequence Scanner. Scanner for sequence or sub-sequence. - */ -class SequenceScanner : public IFieldScanner { - public: - /** - * Ctor - * @param innerScanner inner scanner for each timestep or sub-sequence. - * @param getSeqStartPos A callback, (Argument) => ICpuGpuVectorPtr. - * return a sequence start position or a sub-sequence - * start position. - */ - SequenceScanner( - std::unique_ptr&& innerScanner, - const std::function& getSeqStartPos) - : IFieldScanner(nullptr), - inner_(std::move(innerScanner)), - cnt_(0), - getSeqStartPos_(getSeqStartPos) {} - - /** - * Start prepare. Invoke inner->startPrepare too. - */ - virtual void startPrepare(Argument& argument) { - inner_->startPrepare(argument); - } - - /** - * Prepare. obj is a list or tuple. it will invoke inner_->prepare for each - * element of sequence obj. - */ - virtual void prepare(Argument& argument, PyObject* obj) { - py::SequenceHelper s(obj); - ++cnt_; - for (size_t i = 0; i < s.size(); ++i) { - inner_->prepare(argument, s[i]); - } - } - - /** - * Finish prepare. invoke inner_->finishPrepare too. - */ - virtual void finishPrepare(Argument& argument) { - ICpuGpuVector::resizeOrCreate(getSeqStartPos_(argument), cnt_ + 1, false); - inner_->finishPrepare(argument); - } - - /** - * Start fill. invoke inner->startFill too. - */ - virtual void startFill(Argument& argument) { - getSeqStartPos_(argument)->getMutableData(false)[0] = 0; - cnt_ = 1; - inner_->startFill(argument); - } - - /** - * Fill. Obj is a tuple or list. invoke inner->fill for each element of - * sequence obj. And set seqStartPos at same time. The seqStartPos will be - * calculated by getSeqStartPos callback passed in ctor. - */ - virtual void fill(Argument& argument, PyObject* obj) { - getSeqStartPos_(argument)->getMutableData(false)[cnt_] = - getSeqStartPos_(argument)->getMutableData(false)[cnt_ - 1] + - (int)getSize(obj); - py::SequenceHelper s(obj); - ++cnt_; - for (size_t i = 0; i < s.size(); ++i) { - inner_->fill(argument, s[i]); - } - } - - /** - * Finish fill. will invoke inner->finishFill too. - */ - virtual void finishFill(Argument& argument) { inner_->finishFill(argument); } - - protected: - size_t getSize(PyObject* obj) { - py::SequenceHelper s(obj); - auto sc = dynamic_cast(inner_.get()); - if (sc) { - size_t sum = 0; - for (size_t i = 0; i < s.size(); ++i) { - sum += sc->getSize(s[i]); - } - return sum; - } else { - return s.size(); - } - } - - private: - std::unique_ptr inner_; - size_t cnt_; - std::function getSeqStartPos_; -}; - -IFieldScanner* IFieldScanner::create(SlotHeader* header) { - IFieldScanner* retv = nullptr; - switch (header->slotType) { - case ST_DENSE: - retv = new DenseScanner(header); - break; - case ST_INDEX: - retv = new IndexScanner(header); - break; - case ST_NON_SPARSE_VALUE: - retv = new SparseNonValueScanner(header); - break; - case ST_SPARSE_VALUE: - retv = new SparseValueScanner(header); - break; - default: - LOG(FATAL) << "Not implemented " << header->slotType; - } - - switch (header->seqType) { - case SQT_NONE: - break; - case SQT_SUBSEQ: - retv = new SequenceScanner(std::unique_ptr(retv), - [](Argument& arg) -> ICpuGpuVectorPtr& { - return arg.subSequenceStartPositions; - }); - // fall through, not break; - case SQT_SEQ: - retv = new SequenceScanner(std::unique_ptr(retv), - [](Argument& arg) -> ICpuGpuVectorPtr& { - return arg.sequenceStartPositions; - }); - break; - default: - LOG(FATAL) << "Not implemented"; - } - - return retv; -} - -/** - * No Cache Strategy. Will destruct old data immediately and load data from - * python every pass. - */ -class NoCacheStrategy : public IPyDataProviderCache { - public: - virtual bool reset() { return true; } - - virtual void drop(std::deque* data) { data->clear(); } - - virtual std::deque* load() { return nullptr; } -}; - -/** - * Cache One Pass In Memory strategy. - * - * In first pass, will load data from python and store them in memory. - * The rest passes, will load data from memory. - */ -class CacheOnePassInMemory : public IPyDataProviderCache { - public: - CacheOnePassInMemory() - : objPool_(new std::deque()), - droppedPool_(new std::deque()) {} - - virtual bool reset() { - if (objPool_->empty() && droppedPool_->empty()) { - return true; - } else if (objPool_->empty()) { - std::swap(objPool_, droppedPool_); - return false; - } else { - LOG(FATAL) << "Unexpected branch"; - } - } - - virtual void drop(std::deque* data) { - size_t orgSize = droppedPool_->size(); - droppedPool_->resize(orgSize + data->size()); - for (size_t i = 0; i < data->size(); ++i) { - std::swap((*droppedPool_)[orgSize + i], (*data)[i]); - } - data->clear(); - } - - virtual std::deque* load() { return objPool_.get(); } - - private: - std::unique_ptr> objPool_; - std::unique_ptr> droppedPool_; -}; - -IPyDataProviderCache* IPyDataProviderCache::create(CacheType ct) { - switch (ct) { - case NO_CACHE: - return new NoCacheStrategy(); - case CACHE_PASS_IN_MEM: - return new CacheOnePassInMemory(); - default: - LOG(FATAL) << "Not implemented"; - } -} -} // namespace paddle - -#endif diff --git a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp deleted file mode 100644 index c6cd41de9a1a22470d8659eb90d1ac2b075b2df9..0000000000000000000000000000000000000000 --- a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp +++ /dev/null @@ -1,320 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Evaluator.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/utils/StringUtil.h" - -namespace paddle { - -/** - * calculate sequence-to-sequence edit distance - */ -class CTCErrorEvaluator : public Evaluator { - private: - MatrixPtr outActivations_; - int numTimes_, numClasses_, numSequences_, blank_; - real deletions_, insertions_, substitutions_; - int seqClassficationError_; - mutable std::unordered_map evalResults_; - - std::vector path2String(const std::vector& path) { - std::vector str; - str.clear(); - int prevLabel = -1; - for (std::vector::const_iterator label = path.begin(); - label != path.end(); - label++) { - if (*label != blank_ && - (str.empty() || *label != str.back() || prevLabel == blank_)) { - str.push_back(*label); - } - prevLabel = *label; - } - return str; - } - - std::vector bestLabelSeq() { - std::vector path; - path.clear(); - real* acts = outActivations_->getData(); - for (int i = 0; i < numTimes_; ++i) { - path.push_back(std::max_element(acts + i * numClasses_, - acts + (i + 1) * numClasses_) - - (acts + i * numClasses_)); - } - return path2String(path); - } - - /* "sp, dp, ip" is the weighting parameter of "substitution, deletion, - * insertion" - * in edit-distance error */ - real stringAlignment(std::vector& gtStr, - std::vector& recogStr, - bool backtrace = true, - real sp = 1.0, - real dp = 1.0, - real ip = 1.0) { - std::vector> matrix; - int substitutions, deletions, insertions; - real distance; - int n = gtStr.size(); - int m = recogStr.size(); - - if (n == 0) { - substitutions = 0; - deletions = 0; - insertions = m; - distance = m; - } else if (m == 0) { - substitutions = 0; - deletions = n; - insertions = 0; - distance = n; - } else { - substitutions = 0; - deletions = 0; - insertions = 0; - distance = 0; - // initialize the matrix - matrix.resize(n + 1); - for (int i = 0; i < n + 1; ++i) { - matrix[i].resize(m + 1); - for (int j = 0; j < m + 1; ++j) { - matrix[i][j] = 0; - } - } - for (int i = 0; i < n + 1; ++i) { - matrix[i][0] = i; - } - for (int j = 0; j < m + 1; ++j) { - matrix[0][j] = j; - } - - // calculate the insertions, substitutions and deletions - for (int i = 1; i < n + 1; ++i) { - int s_i = gtStr[i - 1]; - for (int j = 1; j < m + 1; ++j) { - int t_j = recogStr[j - 1]; - int cost = (s_i == t_j) ? 0 : 1; - const int above = matrix[i - 1][j]; - const int left = matrix[i][j - 1]; - const int diag = matrix[i - 1][j - 1]; - const int cell = std::min(above + 1, std::min(left + 1, diag + cost)); - matrix[i][j] = cell; - } - } - - if (backtrace) { - size_t i = n; - size_t j = m; - substitutions = 0; - deletions = 0; - insertions = 0; - - while (i != 0 && j != 0) { - if (matrix[i][j] == matrix[i - 1][j - 1]) { - --i; - --j; - } else if (matrix[i][j] == matrix[i - 1][j - 1] + 1) { - ++substitutions; - --i; - --j; - } else if (matrix[i][j] == matrix[i - 1][j] + 1) { - ++deletions; - --i; - } else { - ++insertions; - --j; - } - } - while (i != 0) { - ++deletions; - --i; - } - while (j != 0) { - ++insertions; - --j; - } - int diff = substitutions + deletions + insertions; - if (diff != matrix[n][m]) { - LOG(ERROR) << "Found path with distance " << diff - << " but Levenshtein distance is " << matrix[n][m]; - } - - distance = (sp * substitutions) + (dp * deletions) + (ip * insertions); - } else { - distance = (real)matrix[n][m]; - } - } - real maxLen = std::max(m, n); - deletions_ += deletions / maxLen; - insertions_ += insertions / maxLen; - substitutions_ += substitutions / maxLen; - - if (distance != 0) { - seqClassficationError_ += 1; - } - - return distance / maxLen; - } - - real editDistance( - real* output, int numTimes, int numClasses, int* labels, int labelsLen) { - numTimes_ = numTimes; - numClasses_ = numClasses; - blank_ = numClasses_ - 1; - outActivations_ = Matrix::create(output, numTimes, numClasses); - std::vector recogStr, gtStr; - recogStr = bestLabelSeq(); - for (int i = 0; i < labelsLen; ++i) { - gtStr.push_back(labels[i]); - } - - return stringAlignment(gtStr, recogStr); - } - - void storeLocalValues() const { - evalResults_["error"] = numSequences_ ? totalScore_ / numSequences_ : 0; - evalResults_["deletion_error"] = - numSequences_ ? deletions_ / numSequences_ : 0; - evalResults_["insertion_error"] = - numSequences_ ? insertions_ / numSequences_ : 0; - evalResults_["substitution_error"] = - numSequences_ ? substitutions_ / numSequences_ : 0; - evalResults_["sequence_error"] = - (real)seqClassficationError_ / numSequences_; - } - - public: - CTCErrorEvaluator() - : numTimes_(0), - numClasses_(0), - numSequences_(0), - blank_(0), - deletions_(0), - insertions_(0), - substitutions_(0), - seqClassficationError_(0) {} - - virtual real evalImp(std::vector& arguments) { - CHECK_EQ(arguments.size(), (size_t)2); - Argument output, label; - output.resizeAndCopyFrom(arguments[0], false, HPPL_STREAM_DEFAULT); - label.resizeAndCopyFrom(arguments[1], false, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - CHECK(label.sequenceStartPositions); - CHECK(label.ids); - size_t numSequences = label.sequenceStartPositions->getSize() - 1; - const int* labelStarts = label.sequenceStartPositions->getData(false); - const int* outputStarts = output.sequenceStartPositions->getData(false); - real totalErr = 0; - for (size_t i = 0; i < numSequences; ++i) { - real err = 0; - err = editDistance( - output.value->getData() + output.value->getWidth() * outputStarts[i], - outputStarts[i + 1] - outputStarts[i], - output.value->getWidth(), - label.ids->getData() + labelStarts[i], - labelStarts[i + 1] - labelStarts[i]); - - totalErr += err; - } - - return totalErr; - } - - virtual void eval(const NeuralNetwork& nn) { - Evaluator::eval(nn); - std::vector arguments; - arguments.reserve(config_.input_layers_size()); - for (const std::string& name : config_.input_layers()) { - arguments.push_back(nn.getLayer(name)->getOutput()); - } - } - - virtual void updateSamplesNum(const std::vector& arguments) { - numSequences_ += arguments[1].getNumSequences(); - } - - virtual void start() { - Evaluator::start(); - numSequences_ = 0; - blank_ = 0; - deletions_ = 0; - insertions_ = 0; - substitutions_ = 0; - seqClassficationError_ = 0; - } - - virtual void printStats(std::ostream& os) const { - storeLocalValues(); - os << config_.name() << " error = " << evalResults_["error"]; - os << " deletions error = " << evalResults_["deletion_error"]; - os << " insertions error = " << evalResults_["insertion_error"]; - os << " substitution error = " << evalResults_["substitution_error"]; - os << " sequence error = " << evalResults_["sequence_error"]; - } - - virtual void distributeEval(ParameterClient2* client) { - double buf[6] = {totalScore_, - (double)deletions_, - (double)insertions_, - (double)substitutions_, - (double)seqClassficationError_, - (double)numSequences_}; - client->reduce(buf, buf, 6, FLAGS_trainer_id, 0); - totalScore_ = buf[0]; - deletions_ = (real)buf[1]; - insertions_ = (real)buf[2]; - substitutions_ = (real)buf[3]; - seqClassficationError_ = (int)buf[4]; - numSequences_ = (int)buf[5]; - } - - void getNames(std::vector* names) { - storeLocalValues(); - names->reserve(names->size() + evalResults_.size()); - for (auto it = evalResults_.begin(); it != evalResults_.end(); ++it) { - names->push_back(config_.name() + "." + it->first); - } - } - - real getValue(const std::string& name, Error* err) const { - storeLocalValues(); - - std::vector buffers; - paddle::str::split(name, '.', &buffers); - auto it = evalResults_.find(buffers[buffers.size() - 1]); - - if (it == evalResults_.end()) { - *err = Error("Evaluator does not have the key %s", name.c_str()); - return 0.0f; - } - - return it->second; - } - - std::string getType(const std::string& name, Error* err) const { - this->getValue(name, err); - if (!err->isOK()) { - return ""; - } - return "ctc_edit_distance"; - } -}; - -REGISTER_EVALUATOR(ctc_edit_distance, CTCErrorEvaluator); - -} // namespace paddle diff --git a/paddle/gserver/evaluators/ChunkEvaluator.cpp b/paddle/gserver/evaluators/ChunkEvaluator.cpp deleted file mode 100644 index a2216293b1ab3a32e9cc903b805ca0aca10d58c1..0000000000000000000000000000000000000000 --- a/paddle/gserver/evaluators/ChunkEvaluator.cpp +++ /dev/null @@ -1,296 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include "paddle/math/Vector.h" -#include "paddle/utils/StringUtil.h" - -#include "Evaluator.h" - -namespace paddle { - -/** - * Chunk evaluator is used to evaluate segment labelling accuracy for a - * sequence. It calculates the chunk detection F1 score. - * - * A chunk is correctly detected if its beginning, end and type are correct. - * Other chunk type is ignored. - * For each label in the label sequence, we have - * - * @code - * tagType = label % numTagType - * chunkType = label / numTagType - * otherChunkType = numChunkTypes - * @endcode - * - * The total number of different labels is numTagType*numChunkTypes+1 - * We support 4 labelling scheme - * The tag type for each of the scheme is shown as follows: - * - * @code - * Scheme Begin Inside End Single - * plain 0 - - - - * IOB 0 1 - - - * IOE - 0 1 - - * IOBES 0 1 2 3 - * @endcode - * - * 'plain' means the whole chunk must contain exactly the same chunk label. - */ -class ChunkEvaluator : public Evaluator { - int otherChunkType_; - int numChunkTypes_; // number of chunk types besides other chunk type - int numTagTypes_; - int tagBegin_; - int tagInside_; - int tagEnd_; - int tagSingle_; - - int64_t numLabelSegments_; - int64_t numOutputSegments_; - int64_t numCorrect_; - - struct Segment { - int begin; - int end; - int type; - bool operator==(const Segment& y) const { - return begin == y.begin && end == y.end && type == y.type; - } - }; - - std::vector labelSegments_; - std::vector outputSegments_; - std::set excludedChunkTypes_; - mutable std::unordered_map values_; - - public: - virtual void init(const EvaluatorConfig& config) { - Evaluator::init(config); - if (config.chunk_scheme() == "IOB") { - numTagTypes_ = 2; - tagBegin_ = 0; - tagInside_ = 1; - tagEnd_ = -1; - tagSingle_ = -1; - } else if (config.chunk_scheme() == "IOE") { - numTagTypes_ = 2; - tagBegin_ = -1; - tagInside_ = 0; - tagEnd_ = 1; - tagSingle_ = -1; - } else if (config.chunk_scheme() == "IOBES") { - numTagTypes_ = 4; - tagBegin_ = 0; - tagInside_ = 1; - tagEnd_ = 2; - tagSingle_ = 3; - } else if (config.chunk_scheme() == "plain") { - numTagTypes_ = 1; - tagBegin_ = -1; - tagInside_ = -1; - tagEnd_ = -1; - tagSingle_ = -1; - } else { - LOG(FATAL) << "Unknown chunk scheme: " << config.chunk_scheme(); - } - CHECK(config.has_num_chunk_types()) << "Missing num_chunk_types in config"; - otherChunkType_ = numChunkTypes_ = config.num_chunk_types(); - - // the chunks of types in excludedChunkTypes_ will not be counted - auto& tmp = config.excluded_chunk_types(); - excludedChunkTypes_.insert(tmp.begin(), tmp.end()); - } - - virtual void start() { - Evaluator::start(); - numLabelSegments_ = 0; - numOutputSegments_ = 0; - numCorrect_ = 0; - } - - virtual void printStats(std::ostream& os) const { - storeLocalValues(); - os << config_.name() << "=" << values_["F1-score"] - << " true_chunks=" << numLabelSegments_ - << " result_chunks=" << numOutputSegments_ - << " correct_chunks=" << numCorrect_; - } - - virtual void distributeEval(ParameterClient2* client) { - int64_t buf[3] = {numLabelSegments_, numOutputSegments_, numCorrect_}; - client->reduce(buf, buf, 3, FLAGS_trainer_id, 0); - numLabelSegments_ = buf[0]; - numOutputSegments_ = buf[1]; - numCorrect_ = buf[2]; - } - - virtual real evalImp(std::vector& arguments) { - CHECK_EQ(arguments.size(), (size_t)2); - IVectorPtr& output = arguments[0].ids; - IVectorPtr& label = arguments[1].ids; - CHECK(!output->useGpu() && !label->useGpu()) << "Not supported"; - auto sequenceStartPositions = - arguments[1].sequenceStartPositions->getVector(false); - CHECK_EQ(output->getSize(), label->getSize()); - CHECK(sequenceStartPositions); - size_t numSequences = sequenceStartPositions->getSize() - 1; - const int* starts = sequenceStartPositions->getData(); - for (size_t i = 0; i < numSequences; ++i) { - eval1(output->getData() + starts[i], - label->getData() + starts[i], - starts[i + 1] - starts[i]); - } - return 0; - } - - void eval1(int* output, int* label, int length) { - getSegments(output, length, outputSegments_); - getSegments(label, length, labelSegments_); - size_t i = 0, j = 0; - while (i < outputSegments_.size() && j < labelSegments_.size()) { - if (outputSegments_[i] == labelSegments_[j] && - excludedChunkTypes_.count(outputSegments_[i].type) != 1) { - ++numCorrect_; - } - if (outputSegments_[i].end < labelSegments_[j].end) { - ++i; - } else if (outputSegments_[i].end > labelSegments_[j].end) { - ++j; - } else { - ++i; - ++j; - } - } - for (auto& segment : labelSegments_) { - if (excludedChunkTypes_.count(segment.type) != 1) ++numLabelSegments_; - } - for (auto& segment : outputSegments_) { - if (excludedChunkTypes_.count(segment.type) != 1) ++numOutputSegments_; - } - } - - void getSegments(int* label, int length, std::vector& segments) { - segments.clear(); - segments.reserve(length); - int chunkStart = 0; - bool inChunk = false; - int tag = -1; - int type = otherChunkType_; - for (int i = 0; i < length; ++i) { - int prevTag = tag; - int prevType = type; - CHECK_LE(label[i], numChunkTypes_ * numTagTypes_); - tag = label[i] % numTagTypes_; - type = label[i] / numTagTypes_; - if (inChunk && isChunkEnd(prevTag, prevType, tag, type)) { - Segment segment{ - chunkStart, // begin - i - 1, // end - prevType, - }; - segments.push_back(segment); - inChunk = false; - } - if (isChunkBegin(prevTag, prevType, tag, type)) { - chunkStart = i; - inChunk = true; - } - } - if (inChunk) { - Segment segment{ - chunkStart, // begin - length - 1, // end - type, - }; - segments.push_back(segment); - } - } - - // whether (prevTag, prevType) is the end of a chunk - bool isChunkEnd(int prevTag, int prevType, int tag, int type) { - if (prevType == otherChunkType_) return false; - if (type == otherChunkType_) return true; - if (type != prevType) return true; - if (prevTag == tagBegin_) return tag == tagBegin_ || tag == tagSingle_; - if (prevTag == tagInside_) return tag == tagBegin_ || tag == tagSingle_; - if (prevTag == tagEnd_) return true; - if (prevTag == tagSingle_) return true; - return false; - } - - // whether (tag, type) is the beginning of a chunk - bool isChunkBegin(int prevTag, int prevType, int tag, int type) { - if (prevType == otherChunkType_) return type != otherChunkType_; - if (type == otherChunkType_) return false; - if (type != prevType) return true; - if (tag == tagBegin_) return true; - if (tag == tagInside_) return prevTag == tagEnd_ || prevTag == tagSingle_; - if (tag == tagEnd_) return prevTag == tagEnd_ || prevTag == tagSingle_; - if (tag == tagSingle_) return true; - return false; - } - - // three metrics: precision, recall and F1-score - void getNames(std::vector* names) { - storeLocalValues(); - names->reserve(names->size() + values_.size()); - for (auto it = values_.begin(); it != values_.end(); ++it) { - names->push_back(config_.name() + "." + it->first); - } - } - - // get value by field name - real getValue(const std::string& name, Error* err) const { - storeLocalValues(); - std::vector buffers; - paddle::str::split(name, '.', &buffers); - auto it = values_.find(buffers.back()); - if (it == values_.end()) { // not found - *err = Error("No such key %s", name.c_str()); - return 0.0f; - } - - return it->second; - } - - // get type of evaluator - std::string getType(const std::string& name, Error* err) const { - this->getValue(name, err); - if (!err->isOK()) { - return ""; - } - return "chunk"; - } - - private: - void storeLocalValues() const { - CHECK_GE(numOutputSegments_, 0); - CHECK_GE(numLabelSegments_, 0); - double precision = - !numOutputSegments_ ? 0 : (double)numCorrect_ / numOutputSegments_; - double recall = - !numLabelSegments_ ? 0 : (double)numCorrect_ / numLabelSegments_; - values_["precision"] = precision; - values_["recall"] = recall; - values_["F1-score"] = - !numCorrect_ ? 0 : 2 * precision * recall / (precision + recall); - } -}; - -REGISTER_EVALUATOR(chunk, ChunkEvaluator); - -} // namespace paddle diff --git a/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp b/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp deleted file mode 100644 index ddb8ebca784db4a83c328ff75f5c50c7aecd7352..0000000000000000000000000000000000000000 --- a/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp +++ /dev/null @@ -1,308 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Evaluator.h" -#include "paddle/gserver/layers/DetectionUtil.h" - -using std::map; -using std::vector; -using std::pair; -using std::make_pair; - -namespace paddle { - -/** - * @brief detection map Evaluator - * - * The config file api is detection_map_evaluator. - */ -class DetectionMAPEvaluator : public Evaluator { - public: - DetectionMAPEvaluator() - : evaluateDifficult_(false), cpuOutput_(nullptr), cpuLabel_(nullptr) {} - - virtual void start() { - Evaluator::start(); - allTruePos_.clear(); - allFalsePos_.clear(); - numPos_.clear(); - } - - virtual real evalImp(std::vector& arguments) { - overlapThreshold_ = config_.overlap_threshold(); - backgroundId_ = config_.background_id(); - evaluateDifficult_ = config_.evaluate_difficult(); - apType_ = config_.ap_type(); - - MatrixPtr detectTmpValue = arguments[0].value; - Matrix::resizeOrCreate(cpuOutput_, - detectTmpValue->getHeight(), - detectTmpValue->getWidth(), - false, - false); - - MatrixPtr labelTmpValue = arguments[1].value; - Matrix::resizeOrCreate(cpuLabel_, - labelTmpValue->getHeight(), - labelTmpValue->getWidth(), - false, - false); - - cpuOutput_->copyFrom(*detectTmpValue); - cpuLabel_->copyFrom(*labelTmpValue); - - Argument label = arguments[1]; - const int* labelIndex = label.sequenceStartPositions->getData(false); - size_t batchSize = label.getNumSequences(); - - vector>> allGTBBoxes; - vector>>> allDetectBBoxes; - - for (size_t n = 0; n < batchSize; ++n) { - map> bboxes; - for (int i = labelIndex[n]; i < labelIndex[n + 1]; ++i) { - vector bbox; - getBBoxFromLabelData(cpuLabel_->getData() + i * 6, 1, bbox); - int c = cpuLabel_->getData()[i * 6]; - bboxes[c].push_back(bbox[0]); - } - allGTBBoxes.push_back(bboxes); - } - - size_t n = 0; - const real* cpuOutputData = cpuOutput_->getData(); - for (size_t imgId = 0; imgId < batchSize; ++imgId) { - map>> bboxes; - size_t curImgId = static_cast((cpuOutputData + n * 7)[0]); - while (curImgId == imgId && n < cpuOutput_->getHeight()) { - vector label; - vector score; - vector bbox; - getBBoxFromDetectData(cpuOutputData + n * 7, 1, label, score, bbox); - bboxes[label[0]].push_back(make_pair(score[0], bbox[0])); - ++n; - curImgId = static_cast((cpuOutputData + n * 7)[0]); - } - allDetectBBoxes.push_back(bboxes); - } - - for (size_t n = 0; n < batchSize; ++n) { - for (map>::iterator it = - allGTBBoxes[n].begin(); - it != allGTBBoxes[n].end(); - ++it) { - size_t count = 0; - if (evaluateDifficult_) { - count = it->second.size(); - } else { - for (size_t i = 0; i < it->second.size(); ++i) - if (!(it->second[i].isDifficult)) ++count; - } - if (numPos_.find(it->first) == numPos_.end() && count != 0) { - numPos_[it->first] = count; - } else { - numPos_[it->first] += count; - } - } - } - - // calcTFPos - calcTFPos(batchSize, allGTBBoxes, allDetectBBoxes); - - return 0; - } - - virtual void printStats(std::ostream& os) const { - real mAP = calcMAP(); - os << "Detection mAP=" << mAP; - } - - virtual void distributeEval(ParameterClient2* client) { - LOG(FATAL) << "Distribute detection evaluation not implemented."; - } - - protected: - void calcTFPos(const size_t batchSize, - const vector>>& allGTBBoxes, - const vector>>>& - allDetectBBoxes) { - for (size_t n = 0; n < allDetectBBoxes.size(); ++n) { - if (allGTBBoxes[n].size() == 0) { - for (map>>::const_iterator - it = allDetectBBoxes[n].begin(); - it != allDetectBBoxes[n].end(); - ++it) { - size_t label = it->first; - for (size_t i = 0; i < it->second.size(); ++i) { - allTruePos_[label].push_back(make_pair(it->second[i].first, 0)); - allFalsePos_[label].push_back(make_pair(it->second[i].first, 1)); - } - } - } else { - for (map>>::const_iterator - it = allDetectBBoxes[n].begin(); - it != allDetectBBoxes[n].end(); - ++it) { - size_t label = it->first; - vector> predBBoxes = it->second; - if (allGTBBoxes[n].find(label) == allGTBBoxes[n].end()) { - for (size_t i = 0; i < predBBoxes.size(); ++i) { - allTruePos_[label].push_back(make_pair(predBBoxes[i].first, 0)); - allFalsePos_[label].push_back(make_pair(predBBoxes[i].first, 1)); - } - } else { - vector gtBBoxes = - allGTBBoxes[n].find(label)->second; - vector visited(gtBBoxes.size(), false); - // Sort detections in descend order based on scores - std::sort(predBBoxes.begin(), - predBBoxes.end(), - sortScorePairDescend); - for (size_t i = 0; i < predBBoxes.size(); ++i) { - real maxOverlap = -1.0; - size_t maxIdx = 0; - for (size_t j = 0; j < gtBBoxes.size(); ++j) { - real overlap = - jaccardOverlap(predBBoxes[i].second, gtBBoxes[j]); - if (overlap > maxOverlap) { - maxOverlap = overlap; - maxIdx = j; - } - } - if (maxOverlap > overlapThreshold_) { - if (evaluateDifficult_ || - (!evaluateDifficult_ && !gtBBoxes[maxIdx].isDifficult)) { - if (!visited[maxIdx]) { - allTruePos_[label].push_back( - make_pair(predBBoxes[i].first, 1)); - allFalsePos_[label].push_back( - make_pair(predBBoxes[i].first, 0)); - visited[maxIdx] = true; - } else { - allTruePos_[label].push_back( - make_pair(predBBoxes[i].first, 0)); - allFalsePos_[label].push_back( - make_pair(predBBoxes[i].first, 1)); - } - } - } else { - allTruePos_[label].push_back(make_pair(predBBoxes[i].first, 0)); - allFalsePos_[label].push_back( - make_pair(predBBoxes[i].first, 1)); - } - } - } - } - } - } - } - - real calcMAP() const { - real mAP = 0.0; - size_t count = 0; - for (map::const_iterator it = numPos_.begin(); - it != numPos_.end(); - ++it) { - size_t label = it->first; - size_t labelNumPos = it->second; - if (labelNumPos == 0 || allTruePos_.find(label) == allTruePos_.end()) - continue; - vector> labelTruePos = allTruePos_.find(label)->second; - vector> labelFalsePos = - allFalsePos_.find(label)->second; - // Compute average precision. - vector tpCumSum; - getAccumulation(labelTruePos, &tpCumSum); - vector fpCumSum; - getAccumulation(labelFalsePos, &fpCumSum); - std::vector precision, recall; - size_t num = tpCumSum.size(); - // Compute Precision. - for (size_t i = 0; i < num; ++i) { - CHECK_LE(tpCumSum[i], labelNumPos); - precision.push_back(static_cast(tpCumSum[i]) / - static_cast(tpCumSum[i] + fpCumSum[i])); - recall.push_back(static_cast(tpCumSum[i]) / labelNumPos); - } - // VOC2007 style - if (apType_ == "11point") { - vector maxPrecisions(11, 0.0); - int startIdx = num - 1; - for (int j = 10; j >= 0; --j) - for (int i = startIdx; i >= 0; --i) { - if (recall[i] < j / 10.) { - startIdx = i; - if (j > 0) maxPrecisions[j - 1] = maxPrecisions[j]; - break; - } else { - if (maxPrecisions[j] < precision[i]) - maxPrecisions[j] = precision[i]; - } - } - for (int j = 10; j >= 0; --j) mAP += maxPrecisions[j] / 11; - ++count; - } else if (apType_ == "Integral") { - // Nature integral - real averagePrecisions = 0.; - real prevRecall = 0.; - for (size_t i = 0; i < num; ++i) { - if (fabs(recall[i] - prevRecall) > 1e-6) - averagePrecisions += precision[i] * fabs(recall[i] - prevRecall); - prevRecall = recall[i]; - } - mAP += averagePrecisions; - ++count; - } else { - LOG(FATAL) << "Unkown ap version: " << apType_; - } - } - if (count != 0) mAP /= count; - return mAP * 100; - } - - void getAccumulation(vector> inPairs, - vector* accuVec) const { - std::stable_sort( - inPairs.begin(), inPairs.end(), sortScorePairDescend); - accuVec->clear(); - size_t sum = 0; - for (size_t i = 0; i < inPairs.size(); ++i) { - sum += inPairs[i].second; - accuVec->push_back(sum); - } - } - - std::string getTypeImpl() const { return "detection_map"; } - - real getValueImpl() const { return calcMAP(); } - - private: - real overlapThreshold_; // overlap threshold when determining whether matched - bool evaluateDifficult_; // whether evaluate difficult ground truth - size_t backgroundId_; // class index of background - std::string apType_; // how to calculate mAP (Integral or 11point) - - MatrixPtr cpuOutput_; - MatrixPtr cpuLabel_; - - map numPos_; // counts of true objects each classification - map>> - allTruePos_; // true positive prediction - map>> - allFalsePos_; // false positive prediction -}; - -REGISTER_EVALUATOR(detection_map, DetectionMAPEvaluator); - -} // namespace paddle diff --git a/paddle/gserver/evaluators/Evaluator.cpp b/paddle/gserver/evaluators/Evaluator.cpp deleted file mode 100644 index 941fb8fb539d58cca22ecf563d2effa816243c3b..0000000000000000000000000000000000000000 --- a/paddle/gserver/evaluators/Evaluator.cpp +++ /dev/null @@ -1,1361 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/gserver/evaluators/Evaluator.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/StringUtil.h" - -DECLARE_int32(trainer_id); - -namespace paddle { - -void Evaluator::eval(const NeuralNetwork& nn) { - std::vector arguments; - arguments.reserve(config_.input_layers_size()); - for (const std::string& name : config_.input_layers()) { - arguments.push_back(nn.getLayer(name)->getOutput()); - } - SetDevice device(arguments[0].deviceId); - real score = evalImp(arguments); - totalScore_ += score; - updateSamplesNum(arguments); -} -/** - * @brief classification error Evaluator - * - * The config file api is classification_error_evaluator. - */ -class ClassificationErrorEvaluator : public Evaluator { - public: - /* - ClassificationErrorEvaluator() : totalScore2_(0) {} - - virtual void start() { - Evaluator::start(); - totalScore2_ = 0; - } */ - - virtual void updateSamplesNum(const std::vector& arguments) { - if (3 == arguments.size()) { - numSamples_ += arguments[2].value->getSum(); - } else { - numSamples_ += arguments[0].getBatchSize(); - } - } - - MatrixPtr calcError(std::vector& arguments) { - CHECK_GE(arguments.size(), (size_t)2); - CHECK_LE(arguments.size(), (size_t)3); - MatrixPtr& output = arguments[0].value; - IVectorPtr& label = arguments[1].ids; - MatrixPtr& multiBinaryLabel = arguments[1].value; // For multi binary label - bool supportWeight = (3 == arguments.size()) ? true : false; - MatrixPtr weight = supportWeight ? arguments[2].value : nullptr; - if (nullptr == output || - (nullptr == label && nullptr == multiBinaryLabel) || - (supportWeight && nullptr == weight)) { - return 0; - } - - if (label != nullptr) { - CHECK_EQ(label->getSize(), output->getHeight()); - } else { - CHECK_EQ(multiBinaryLabel->getHeight(), output->getHeight()); - CHECK_EQ(multiBinaryLabel->getWidth(), output->getWidth()); - } - if (supportWeight) { - CHECK_EQ(output->getHeight(), weight->getHeight()); - CHECK_EQ((size_t)1, weight->getWidth()); - } - - const MatrixPtr errorMat = Matrix::create(output->getHeight(), - 1, - /* trans= */ false, - useGpu(arguments[0].deviceId)); - - errorMat->zeroMem(); - - if (label != nullptr) { - errorMat->classificationError(*output, *label, config_.top_k()); - } else if (dynamic_cast(multiBinaryLabel.get()) || - dynamic_cast(multiBinaryLabel.get())) { - errorMat->classificationErrorMulti( - *output, *multiBinaryLabel, config_.classification_threshold()); - } else { - errorMat->binaryClassificationError( - 0, *output, *multiBinaryLabel, config_.classification_threshold()); - } - - if (supportWeight) { - errorMat->dotMul(*errorMat, *weight); - } - return errorMat; - } - - void printStats(std::ostream& os) const { - if (config_.top_k() == 1) { - os << config_.name() << "=" - << (numSamples_ ? totalScore_ / numSamples_ : 0); - } else { - os << " top_" << config_.top_k() - << "_error=" << (numSamples_ ? totalScore_ / numSamples_ : 0); - } - } - - virtual real evalImp(std::vector& arguments) { - MatrixPtr errorMat = calcError(arguments); - return errorMat->getSum(); - } - - virtual void distributeEval(ParameterClient2* client) { - mergeResultsOfAllClients(client); - } - - // Evaluator interface - protected: - std::string getTypeImpl() const { return "classification_error"; } -}; - -/** - * @brief sequence classification error Evaluator - * @note sequence level classification error stats, - * if any frame in one sequence has error, the sequence is error - */ -class SequenceClassificationErrorEvaluator - : public ClassificationErrorEvaluator { - public: - virtual void updateSamplesNum(const std::vector& arguments) { - numSamples_ += arguments[0].getNumSequences(); - } - - virtual real evalImp(std::vector& arguments) { - auto sequenceStartPositions = - arguments[0].sequenceStartPositions->getVector(false); - CHECK(sequenceStartPositions != nullptr); - const int* starts = sequenceStartPositions->getData(); - - MatrixPtr errorMat = calcError(arguments); - - int errCounter = 0; - CpuVector errorVec(0, nullptr); - for (size_t i = 0; i < sequenceStartPositions->getSize() - 1; ++i) { - errorVec.subVecFrom( - errorMat->getData(), starts[i], starts[i + 1] - starts[i]); - if (errorVec.getSum() > 0) { - errCounter += 1; - } - } - - return static_cast(errCounter); - } - - virtual void distributeEval(ParameterClient2* client) { - mergeResultsOfAllClients(client); - } - - // Evaluator interface - protected: - std::string getTypeImpl() const { return "seq_classification_error"; } -}; -REGISTER_EVALUATOR(seq_classification_error, - SequenceClassificationErrorEvaluator); -/** - * @brief sum Evaluator - * Calculate the sum of output or label - * - * The config file api is sum_evaluator. - */ -class SumEvaluator : public Evaluator { - public: - SumEvaluator() : cpuLabel_(nullptr), cpuWeight_(nullptr) {} - - virtual void updateSamplesNum(const std::vector& arguments) { - if (2 == arguments.size()) { - numSamples_ += arguments[1].value->getSum(); - } else { - numSamples_ += arguments[0].getBatchSize(); - } - } - - virtual real evalImp(std::vector& arguments) { - REGISTER_TIMER("SumEvaluator"); - CHECK_GE(arguments.size(), (size_t)1); - CHECK_LE(arguments.size(), (size_t)2); - bool supportWeight = (2 == arguments.size()) ? true : false; - if (supportWeight) { - if (nullptr == arguments[1].value) { - return 0; - } - CHECK_EQ(arguments[1].value->getWidth(), (size_t)1); - } - - // The sum of output - if (arguments[0].value) { - if (supportWeight) { - CHECK_EQ(arguments[0].value->getHeight(), - arguments[1].value->getHeight()); - MatrixPtr tmpMat = Matrix::create(arguments[0].value->getHeight(), - arguments[0].value->getWidth(), - /* trans= */ false, - arguments[0].value->useGpu()); - tmpMat->copyFrom(*arguments[0].value); - tmpMat->rowScale(0, *tmpMat, *arguments[1].value); - return tmpMat->getSum(); - } else { - return arguments[0].value->getSum(); - } - // The sum of label - } else if (arguments[0].ids) { - size_t insNum = arguments[0].ids->getSize(); - IVectorPtr label = arguments[0].ids; - MatrixPtr weight = supportWeight ? arguments[1].value : nullptr; - if (dynamic_cast(label.get())) { - IVector::resizeOrCreate(cpuLabel_, insNum, false); - cpuLabel_->copyFrom(*arguments[0].ids); - - if (supportWeight) { - CHECK_EQ(insNum, arguments[1].value->getHeight()); - Matrix::resizeOrCreate(cpuWeight_, insNum, (size_t)1, false, false); - cpuWeight_->copyFrom(*arguments[1].value); - } - - label = cpuLabel_; - weight = cpuWeight_; - } - - if (supportWeight) { - real score = 0.0; - int* labelD = label->getData(); - real* weightD = weight->getData(); - for (size_t i = 0; i < insNum; ++i) { - score += (labelD[i] * weightD[i]); - } - return score; - } else { - return label->getSum(); - } - } else { - return 0; - } - } - - virtual void distributeEval(ParameterClient2* client) { - mergeResultsOfAllClients(client); - } - - private: - IVectorPtr cpuLabel_; - MatrixPtr cpuWeight_; - - // Evaluator interface - protected: - std::string getTypeImpl() const { return "sum"; } -}; -/** - * @brief column sum Evaluator - * @note column sum for the colIdx-th column * - * - colIdx = 0: the 0-th column. - * - colIdx > 0: the colIdx-th column. - * - colIdx < 0: the last colIdx-th column. - * - * The config file api is column_sum_evaluator. - * - */ -class ColumnSumEvaluator : public Evaluator { - public: - explicit ColumnSumEvaluator(int32_t colIdx) - : colIdx_(colIdx), colNum_(0), sum_(nullptr) {} - - virtual void start() { - Evaluator::start(); - if (nullptr != sum_) { - sum_->zeroMem(); - } - } - - virtual void updateSamplesNum(const std::vector& arguments) { - if (2 == arguments.size()) { - numSamples_ += arguments[1].value->getSum(); - } else { - numSamples_ += arguments[0].getBatchSize(); - } - } - - virtual real evalImp(std::vector& arguments) { - REGISTER_TIMER("ColumnSumEvaluator"); - CHECK_GE(arguments.size(), (size_t)1); - CHECK_LE(arguments.size(), (size_t)2); - bool supportWeight = (2 == arguments.size()) ? true : false; - if (nullptr == arguments[0].value || - (supportWeight && nullptr == arguments[1].value)) { - return 0; - } - - size_t insNum = arguments[0].value->getHeight(); - size_t colNum = arguments[0].value->getWidth(); - if (nullptr == sum_) { - sum_ = Matrix::create((size_t)1, colNum, false, /* useGpu */ false); - colNum_ = colNum; - sum_->zeroMem(); - } else { - CHECK_EQ(colNum, sum_->getWidth()); - } - - if (supportWeight) { - CHECK_EQ(insNum, arguments[1].value->getHeight()); - CHECK_EQ((size_t)1, arguments[1].value->getWidth()); - MatrixPtr tmpMat = Matrix::create(insNum, colNum); - if (arguments[0].value->useGpu()) { - tmpMat->copyFrom(*arguments[0].value); - } - if (!arguments[1].value->useGpu()) { - if (!arguments[0].value->useGpu()) { - tmpMat->rowScale(0, *arguments[0].value, *arguments[1].value); - } else { - tmpMat->rowScale(0, *tmpMat, *arguments[1].value); - } - } else { - MatrixPtr tmp2 = Matrix::create(insNum, 1); - tmp2->copyFrom(*arguments[1].value); - if (!arguments[0].value->useGpu()) { - tmpMat->rowScale(0, *arguments[0].value, *tmp2); - } else { - tmpMat->rowScale(0, *tmpMat, *tmp2); - } - } - sum_->accumulateColSum(*tmpMat); - } else { - if (!arguments[0].value->useGpu()) { - sum_->accumulateColSum(*arguments[0].value); - } else { - MatrixPtr tmpMat = Matrix::create(insNum, colNum); - tmpMat->copyFrom(*arguments[0].value); - sum_->accumulateColSum(*tmpMat); - } - } - return 0; - } - - virtual void printStats(std::ostream& os) const { - CHECK(colIdx_ + (int32_t)colNum_ >= 0 && colIdx_ - (int32_t)colNum_ < 0) - << "column index [" << colIdx_ << "] out of range [-" << colNum_ << ", " - << colNum_ << ")"; - size_t colIdx = 0; - if (colIdx_ >= 0) { - colIdx = colIdx_; - } else { - colIdx = colNum_ + colIdx_; - } - os << config_.name() << "=" - << (numSamples_ ? sum_->getElement(0, colIdx) / numSamples_ : 0); - } - - void distributeEval(ParameterClient2* client) { - client->reduce( - sum_->getData(), sum_->getData(), colNum_, FLAGS_trainer_id, 0); - client->reduce(&numSamples_, &numSamples_, 1, FLAGS_trainer_id, 0); - } - - private: - int32_t colIdx_; - size_t colNum_; - MatrixPtr sum_; /* cpu matrix */ - - // Evaluator interface - protected: - std::string getTypeImpl() const { - if (colIdx_ == -1) - return "last-column-sum"; - else - return "column-sum"; - } -}; - -void AucEvaluator::start() { - Evaluator::start(); - memset(statPos_, 0, sizeof(statPos_)); - memset(statNeg_, 0, sizeof(statNeg_)); -} - -real AucEvaluator::evalImp(std::vector& arguments) { - REGISTER_TIMER("AucEvaluator"); - CHECK_GE(arguments.size(), (size_t)2); - CHECK_LE(arguments.size(), (size_t)3); - MatrixPtr output = arguments[0].value; - IVectorPtr label = arguments[1].ids; - MatrixPtr labelval = arguments[1].value; - bool supportWeight = (3 == arguments.size()) ? true : false; - MatrixPtr weight = supportWeight ? arguments[2].value : nullptr; - - if (nullptr == output || (supportWeight && nullptr == weight)) { - return 0; - } - size_t insNum = output->getHeight(); - size_t outputDim = output->getWidth(); - // Copy label from value to a vector. - if (nullptr == label && nullptr != labelval) { - // label width is 1 - CHECK_EQ(1U, labelval->getWidth()); - VectorPtr vec = - Vector::create(labelval->getData(), insNum, output->useGpu()); - label = vec->castToInt(); - } - - CHECK_EQ(insNum, label->getSize()); - if (supportWeight) { - CHECK_EQ(insNum, weight->getHeight()); - CHECK_EQ((size_t)1, weight->getWidth()); - } - - CHECK(colIdx_ + (int32_t)outputDim >= 0 && colIdx_ - (int32_t)outputDim < 0) - << "column index [" << colIdx_ << "] out of range [-" << outputDim << ", " - << outputDim << ")"; - realColumnIdx_ = 0; - if (colIdx_ >= 0) { - realColumnIdx_ = colIdx_; - } else { - realColumnIdx_ = outputDim + colIdx_; - } - - if (dynamic_cast(output.get())) { - Matrix::resizeOrCreate(cpuOutput_, - insNum, - outputDim, - /* trans=*/false, - /* useGpu=*/false); - cpuOutput_->copyFrom(*output); - IVector::resizeOrCreate(cpuLabel_, insNum, false); - cpuLabel_->copyFrom(*label); - - if (supportWeight) { - Matrix::resizeOrCreate(cpuWeight_, insNum, (size_t)1, false, false); - cpuWeight_->copyFrom(*weight); - } - - output = cpuOutput_; - label = cpuLabel_; - weight = cpuWeight_; - } - - real* outputD = output->getData(); - int* labelD = label->getData(); - real* weightD = supportWeight ? weight->getData() : nullptr; - size_t pos = realColumnIdx_; - - for (size_t i = 0; i < insNum; ++i) { - real value = outputD[pos]; - uint32_t binIdx = static_cast(value * kBinNum_); - CHECK(binIdx <= kBinNum_) << "bin index [" << binIdx - << "] out of range, predict value[" << value - << "]"; - real w = supportWeight ? weightD[i] : 1.0; - if (labelD[i] == kNegativeLabel_) { - statNeg_[binIdx] += w; - } else { - statPos_[binIdx] += w; - } - pos += outputDim; - } - return 0; -} - -void AucEvaluator::distributeEval(ParameterClient2* client) { - client->reduce(statPos_, statPos_, kBinNum_ + 1, FLAGS_trainer_id, 0); - client->reduce(statNeg_, statNeg_, kBinNum_ + 1, FLAGS_trainer_id, 0); -} - -double AucEvaluator::calcAuc() const { - double totPos = 0.0; - double totNeg = 0.0; - double totPosPrev = 0.0; - double totNegPrev = 0.0; - double auc = 0.0; - - int64_t idx = kBinNum_; - while (idx >= 0) { - totPosPrev = totPos; - totNegPrev = totNeg; - totPos += statPos_[idx]; - totNeg += statNeg_[idx]; - auc += trapezoidArea(totNeg, totNegPrev, totPos, totPosPrev); - --idx; - } - - if (totPos > 0.0 && totNeg > 0.0) { - return auc / totPos / totNeg; - } else { - return 0.0; - } -} - -real AucEvaluator::getValueImpl() const { return calcAuc(); } - -std::string AucEvaluator::getTypeImpl() const { - if (colIdx_ == -1) { - return "last-column-auc"; - } else { - return "auc"; - } -} - -// class RankAucEvaluator -REGISTER_EVALUATOR(rankauc, RankAucEvaluator); - -void RankAucEvaluator::start() { Evaluator::start(); } -void RankAucEvaluator::updateSamplesNum( - const std::vector& arguments) { - numSamples_ += arguments[0].getNumSequences(); -} -real RankAucEvaluator::evalImp(std::vector& arguments) { - CHECK_GE(arguments.size(), 2U); - CHECK_LE(arguments.size(), 3U); - double batchAuc = 0.0; - output_ = arguments[0].value; - click_ = arguments[1].value; - size_t batchSize = output_->getHeight(); - CHECK(!output_->useGpu()) << "RankAUC evaluator does not support GPU!"; - - if (arguments.size() == 3U) { - pv_ = arguments[2].value; - } else { - Matrix::resizeOrCreate(pv_, batchSize, 1, false, false); - std::fill(pv_->getData(), pv_->getData() + batchSize, 1.0); - } - - real* outputData = output_->getData(); - real* clickData = click_->getData(); - real* pvData = pv_->getData(); - - auto startPos = arguments[0].sequenceStartPositions->getVector(false); - const int* startPosData = startPos->getData(); - size_t batchNum = startPos->getSize() - 1; - for (size_t i = 0; i < batchNum; ++i) { - int beginPos = startPosData[i]; - int endPos = startPosData[i + 1]; - batchAuc += calcRankAuc(outputData + beginPos, - clickData + beginPos, - pvData + beginPos, - endPos - beginPos); - } - return batchAuc; -} - -double RankAucEvaluator::calcRankAuc(real* outputData, - real* clickData, - real* pvData, - size_t size) { - outputPair_.clear(); - for (size_t i = 0; i < size; ++i) { - outputPair_.push_back(std::make_pair(outputData[i], i)); - } - std::sort(outputPair_.begin(), - outputPair_.end(), - [](const std::pair& a, const std::pair& b) { - return a.first > b.first; - }); - double aucTmp = 0.0; - double clickSum = 0.0; - double oldClickSum = 0.0; - double noClick = 0.0; - double noClickSum = 0.0; - - double lastScore = outputPair_[0].first + 1.0; - for (size_t i = 0; i < size; ++i) { - if (lastScore != outputPair_[i].first) { - aucTmp += (clickSum + oldClickSum) * noClick / 2.0; - oldClickSum = clickSum; - noClick = 0.0; - lastScore = outputPair_[i].first; - } - size_t id = outputPair_[i].second; - noClick += pvData[id] - clickData[id]; - noClickSum += noClick; - clickSum += clickData[id]; - } - aucTmp += (clickSum + oldClickSum) * noClick / 2.0; - return (clickSum * noClickSum) == 0.0 ? 0.0 - : aucTmp / (clickSum * noClickSum); -} - -std::string RankAucEvaluator::getTypeImpl() const { return "rankauc"; } - -// class PrecisionRecallEvaluator -REGISTER_EVALUATOR(precision_recall, PrecisionRecallEvaluator); - -void PrecisionRecallEvaluator::start() { - Evaluator::start(); - statsInfo_.clear(); - values_.clear(); -} - -real PrecisionRecallEvaluator::evalImp(std::vector& arguments) { - REGISTER_TIMER("PrecisionRecallEvaluator"); - CHECK_GE(arguments.size(), (size_t)2); - CHECK_LE(arguments.size(), (size_t)3); - MatrixPtr output = arguments[0].value; - IVectorPtr label = arguments[1].ids; - MatrixPtr multiBinaryLabel = arguments[1].value; - bool supportWeight = (3 == arguments.size()) ? true : false; - MatrixPtr weight = supportWeight ? arguments[2].value : nullptr; - if (nullptr == output || (nullptr == label && nullptr == multiBinaryLabel) || - (supportWeight && nullptr == weight)) { - return 0; - } - - size_t insNum = output->getHeight(); - size_t outputDim = output->getWidth(); - if (label != nullptr) { - CHECK_EQ(insNum, label->getSize()); - } else { - CHECK_EQ(insNum, multiBinaryLabel->getHeight()); - CHECK_EQ(outputDim, multiBinaryLabel->getWidth()); - } - if (supportWeight) { - CHECK_EQ(insNum, weight->getHeight()); - CHECK_EQ((size_t)1, weight->getWidth()); - } - - if (statsInfo_.size() != outputDim) { - statsInfo_.clear(); - statsInfo_.resize(outputDim); - } - - isMultiBinaryLabel_ = (nullptr == label) ? true : false; - if (label != nullptr) { - if (dynamic_cast(output.get())) { - Matrix::resizeOrCreate(cpuOutput_, insNum, outputDim, false, false); - cpuOutput_->copyFrom(*output); - IVector::resizeOrCreate(cpuLabel_, insNum, false); - cpuLabel_->copyFrom(*label); - if (supportWeight) { - Matrix::resizeOrCreate(cpuWeight_, insNum, (size_t)1, false, false); - cpuWeight_->copyFrom(*weight); - } - - output = cpuOutput_; - label = cpuLabel_; - weight = cpuWeight_; - } - calcStatsInfo(output, label, weight); - } else { - // Not support GPU for multi binary labels - CHECK(dynamic_cast(multiBinaryLabel.get())); - calcStatsInfoMulti(output, multiBinaryLabel, weight); - } - return 0; -} - -void PrecisionRecallEvaluator::printStats(std::ostream& os) const { - PrintStatsInfo info; - bool containMacroMicroInfo = getStatsInfo(&info); - os << "positive_label=" << config_.positive_label() - << " precision=" << info.precision << " recall=" << info.recall - << " F1-score=" << info.f1; - if (containMacroMicroInfo) { - os << "macro-average-precision=" << info.macroAvgPrecision - << " macro-average-recall=" << info.macroAvgRecall - << " macro-average-F1-score=" << info.macroAvgF1Score; - if (!isMultiBinaryLabel_) { - // precision and recall are equal in this case - os << " micro-average-precision=" << info.microAvgPrecision; - } else { - os << " micro-average-precision=" << info.microAvgPrecision - << " micro-average-recall=" << info.microAvgRecall - << " micro-average-F1-score=" << info.microAvgF1Score; - } - } -} - -void PrecisionRecallEvaluator::calcStatsInfo(const MatrixPtr& output, - const IVectorPtr& label, - const MatrixPtr& weight) { - size_t insNum = output->getHeight(); - size_t dim = output->getWidth(); - real* outputD = output->getData(); - int* labelD = label->getData(); - real* weightD = (weight != nullptr) ? weight->getData() : nullptr; - for (size_t i = 0; i < insNum; ++i) { - CHECK_GE(labelD[i], 0); - CHECK_LT((size_t)labelD[i], dim); - size_t maxIdx = 0; - real maxValue = outputD[i * dim]; - for (size_t j = 1; j < dim; ++j) { - size_t idx = i * dim + j; - if (maxValue < outputD[idx]) { - maxIdx = j; - maxValue = outputD[idx]; - } - } - - real w = (weightD != nullptr) ? weightD[i] : 1.0; - if (maxIdx == (size_t)labelD[i]) { - statsInfo_[maxIdx].TP += w; // true positive for labelD[i] - // true negative for all labels except for labelD[i] - for (size_t j = 0; j < dim; ++j) { - statsInfo_[j].TN += w; - } - statsInfo_[maxIdx].TN -= w; - } else { - statsInfo_[labelD[i]].FN += w; // false negative for labelD[i] - statsInfo_[maxIdx].FP += w; // false positive for maxIdx - // true negatives for all labels except for maxIdx and labelD[i] - for (size_t j = 0; j < dim; ++j) { - statsInfo_[j].TN += w; - } - statsInfo_[maxIdx].TN -= w; - statsInfo_[labelD[i]].TN -= w; - } - } -} - -void PrecisionRecallEvaluator::calcStatsInfoMulti(const MatrixPtr& output, - const MatrixPtr& label, - const MatrixPtr& weight) { - size_t insNum = output->getHeight(); - size_t dim = output->getWidth(); - real* outputD = output->getData(); - auto labelD = dynamic_cast(label.get()); - real* weightD = (weight != nullptr) ? weight->getData() : nullptr; - real threshold = config_.classification_threshold(); - for (size_t i = 0; i < insNum; ++i) { - for (size_t j = 0; j < dim; ++j) { - real w = (weightD != nullptr) ? weightD[i] : 1.0; - size_t idx = i * dim + j; - if (outputD[idx] < threshold) { - statsInfo_[j].TN += w; // true negative - } else { - statsInfo_[j].FP += w; // false positive - } - } - - const int* cols = labelD->getRowCols(i); - for (size_t j = 0; j < labelD->getColNum(i); ++j) { - CHECK_LT(size_t(cols[j]), dim); - real w = (weightD != nullptr) ? weightD[i] : 1.0; - size_t idx = i * dim + cols[j]; - if (outputD[idx] < threshold) { - statsInfo_[cols[j]].FN += w; // false negative - statsInfo_[cols[j]].TN -= w; // true negative - } else { - statsInfo_[cols[j]].TP += w; // true positive - statsInfo_[cols[j]].FP -= w; // false positive - } - } - } -} - -void PrecisionRecallEvaluator::storeLocalValues() const { - if (this->values_.size() == 0) { - PrintStatsInfo info; - bool containMacroMicroInfo = getStatsInfo(&info); - values_["precision"] = info.precision; - values_["recal"] = info.recall; - values_["F1-score"] = info.f1; - if (containMacroMicroInfo) { - values_["macro-average-precision"] = info.macroAvgPrecision; - values_["macro-average-recall"] = info.macroAvgRecall; - values_["macro-average-F1-score"] = info.macroAvgF1Score; - if (!isMultiBinaryLabel_) { - // precision and recall are equal in this case - values_["micro-average-precision"] = info.microAvgPrecision; - } else { - values_["micro-average-precision"] = info.microAvgPrecision; - values_["micro-average-recall"] = info.microAvgRecall; - values_["micro-average-F1-score"] = info.microAvgF1Score; - } - } - } -} - -void PrecisionRecallEvaluator::getNames(std::vector* names) { - this->storeLocalValues(); - names->reserve(this->values_.size()); - for (auto it = this->values_.begin(); it != this->values_.end(); ++it) { - names->push_back(this->config_.name() + "." + it->first); - } -} - -real PrecisionRecallEvaluator::getValue(const std::string& name, - Error* err) const { - this->storeLocalValues(); - std::vector buffers; - paddle::str::split(name, '.', &buffers); - auto it = this->values_.find(buffers[buffers.size() - 1]); - if (it == this->values_.end()) { // not found - *err = Error("No such key %s", name.c_str()); - return .0f; - } - - return it->second; -} - -std::string PrecisionRecallEvaluator::getType(const std::string& name, - Error* err) const { - this->getValue(name, err); - if (!err->isOK()) { - return ""; - } - return "precision_recall"; -} - -void PrecisionRecallEvaluator::distributeEval(ParameterClient2* client) { - size_t size = 4 * statsInfo_.size(); - double* buf = new double[size]; - for (size_t i = 0; i < statsInfo_.size(); ++i) { - buf[4 * i + 0] = statsInfo_[i].TP; - buf[4 * i + 1] = statsInfo_[i].TN; - buf[4 * i + 2] = statsInfo_[i].FP; - buf[4 * i + 3] = statsInfo_[i].FN; - } - client->reduce(buf, buf, size, FLAGS_trainer_id, 0); - for (size_t i = 0; i < statsInfo_.size(); ++i) { - statsInfo_[i].TP = buf[4 * i + 0]; - statsInfo_[i].TN = buf[4 * i + 1]; - statsInfo_[i].FP = buf[4 * i + 2]; - statsInfo_[i].FN = buf[4 * i + 3]; - } - delete[] buf; -} - -bool PrecisionRecallEvaluator::getStatsInfo( - PrecisionRecallEvaluator::PrintStatsInfo* info) const { - int label = config_.positive_label(); - if (label != -1) { - CHECK(label >= 0 && label < (int)statsInfo_.size()) - << "positive_label [" << label << "] should be in range [0, " - << statsInfo_.size() << ")"; - info->precision = calcPrecision(statsInfo_[label].TP, statsInfo_[label].FP); - info->recall = calcRecall(statsInfo_[label].TP, statsInfo_[label].FN); - info->f1 = calcF1Score(info->precision, info->recall); - return false; - } - - // micro average method: precision = (TP1+TP2)/(TP1+FP1+TP2+FP2) - // macro average method: precision = (precision1+precision2)/2 - double microTotalTP = 0; - double microTotalFP = 0; - double microTotalFN = 0; - info->macroAvgPrecision = 0; - info->macroAvgRecall = 0; - size_t numLabels = statsInfo_.size(); - for (size_t i = 0; i < numLabels; ++i) { - microTotalTP += statsInfo_[i].TP; - microTotalFP += statsInfo_[i].FP; - microTotalFN += statsInfo_[i].FN; - info->macroAvgPrecision += - calcPrecision(statsInfo_[i].TP, statsInfo_[i].FP); - info->macroAvgRecall += calcRecall(statsInfo_[i].TP, statsInfo_[i].FN); - } - info->macroAvgPrecision /= numLabels; - info->macroAvgRecall /= numLabels; - info->macroAvgF1Score = - calcF1Score(info->macroAvgPrecision, info->macroAvgRecall); - - info->microAvgPrecision = calcPrecision(microTotalTP, microTotalFP); - info->microAvgRecall = calcPrecision(microTotalTP, microTotalFN); - info->microAvgF1Score = - calcF1Score(info->microAvgPrecision, info->microAvgRecall); - return true; -} - -REGISTER_EVALUATOR(pnpair, PnpairEvaluator); -void PnpairEvaluator::start() { - Evaluator::start(); - memset(pairArray_, 0, sizeof(pairArray_)); - predictArray_.clear(); -} - -real PnpairEvaluator::evalImp(std::vector& arguments) { - CHECK_GE(arguments.size(), 3UL); - CHECK_LE(arguments.size(), 4UL); - MatrixPtr output = arguments[0].value; - IVectorPtr label = arguments[1].ids; - IVectorPtr info = arguments[2].ids; - bool supportWeight = (4 == arguments.size()) ? true : false; - MatrixPtr weight = supportWeight ? arguments[3].value : nullptr; - if (nullptr == output || nullptr == label || - (supportWeight && nullptr == weight)) { - return 0; - } - size_t height = output->getHeight(); - size_t width = output->getWidth(); - CHECK_EQ(height, label->getSize()); - CHECK_EQ(height, info->getSize()); - if (supportWeight) { - CHECK_EQ(height, weight->getHeight()); - CHECK_EQ((size_t)1, weight->getWidth()); - } - - if (dynamic_cast(output.get())) { - Matrix::resizeOrCreate(cpuOutput_, height, width, false, false); - IVector::resizeOrCreate(cpuLabel_, height, false); - IVector::resizeOrCreate(cpuInfo_, height, false); - cpuOutput_->copyFrom(*output); - cpuLabel_->copyFrom(*label); - cpuInfo_->copyFrom(*info); - - output = cpuOutput_; - label = cpuLabel_; - info = cpuInfo_; - - if (supportWeight) { - Matrix::resizeOrCreate(cpuWeight_, height, (size_t)1, false, false); - cpuWeight_->copyFrom(*weight); - weight = cpuWeight_; - } - } - - real* outputs = output->getData(); - int* labels = label->getData(); - int* infos = info->getData(); - real* weights = supportWeight ? weight->getData() : nullptr; - for (size_t i = 0; i < output->getHeight(); i++) { - real y1 = outputs[i * width + (width - 1)]; - real w = supportWeight ? weights[i] : 1.0; - predictArray_.push_back(PredictionResult(y1, labels[i], infos[i], w)); - } - return 0; -} - -void PnpairEvaluator::stat(size_t start, - size_t end, - PredictionResult* answers, - double& pos, - double& neg, - double& spe) { - for (size_t i = start; i < end; i++) { - for (size_t j = i + 1; j < end; j++) { - CHECK_EQ(answers[i].queryid, answers[j].queryid); - // The pair weight is the mean of the two samples' weight - double weight = (answers[i].weight + answers[j].weight) / 2.0; - if (answers[i].label != answers[j].label) { - if ((answers[i].out > answers[j].out && - answers[i].label > answers[j].label) || - (answers[i].out < answers[j].out && - answers[i].label < answers[j].label)) { - pos += weight; - } else if ((answers[i].out > answers[j].out && - answers[i].label < answers[j].label) || - (answers[i].out < answers[j].out && - answers[i].label > answers[j].label)) { - neg += weight; - } else { - spe += weight; - } - } - } - } -} - -void PnpairEvaluator::calc(std::vector& predictArray) { - std::sort(predictArray.begin(), - predictArray.end(), - [](const PredictionResult& x, const PredictionResult& y) { - return x.queryid < y.queryid; - }); - - double pos = 0; - double neg = 0; - double special = 0; - auto start = predictArray.begin(); - while (start != predictArray.end()) { - auto end = std::find_if( - start + 1, predictArray.end(), [=](const PredictionResult& x) { - return x.queryid != start->queryid; - }); - CHECK(end != start); - stat(start - predictArray.begin(), - end - predictArray.begin(), - predictArray.data(), - pos, - neg, - special); - - start = end; - } - - pairArray_[0] += pos; - pairArray_[1] += neg; - - LOG(INFO) << " calc total pos pair: " << pos - << " calc total neg pair: " << neg - << " calc total special pair: " << special; -} - -std::string PnpairEvaluator::getTypeImpl() const { return "pnpair"; } - -ClassRegistrar Evaluator::registrar_; -Evaluator* Evaluator::create(const EvaluatorConfig& config) { - Evaluator* evaluator = registrar_.createByType(config.type()); - evaluator->init(config); - return evaluator; -} - -REGISTER_EVALUATOR(classification_error, ClassificationErrorEvaluator); -REGISTER_EVALUATOR(sum, SumEvaluator); -static InitFunction __reg_type_auc_sum__([]() { - Evaluator::registrar_.registerClass( - "last-column-sum", [] { return new ColumnSumEvaluator(-1); }); - Evaluator::registrar_.registerClass("last-column-auc", - [] { return new AucEvaluator(-1); }); -}); - -/** - * @brief print value of each layer. - * - * The config file api is value_printer_evaluator. - */ -class ValuePrinter : public NotGetableEvaluator { - public: - virtual void eval(const NeuralNetwork& nn) { - for (const std::string& name : config_.input_layers()) { - nn.getLayer(name)->getOutput().printValueString(LOG(INFO), - "layer=" + name + " "); - } - } - - virtual void updateSamplesNum(const std::vector& arguments) {} - - virtual real evalImp(std::vector& arguments) { return 0; } -}; -REGISTER_EVALUATOR(value_printer, ValuePrinter); - -/** - * @brief print gradient of each layer. - * - * The config file api is gradient_printer_evaluator. - */ -class GradientPrinter : public NotGetableEvaluator { - public: - virtual void eval(const NeuralNetwork& nn) { - for (const std::string& name : config_.input_layers()) { - const Argument& argu = nn.getLayer(name)->getOutput(); - if (argu.grad) { - std::ostringstream os; - argu.grad->print(os); - LOG(INFO) << "layer=" << name << " grad matrix:\n" << os.str(); - } - } - } - - virtual void updateSamplesNum(const std::vector& arguments) {} - - virtual real evalImp(std::vector& arguments) { return 0; } -}; -REGISTER_EVALUATOR(gradient_printer, GradientPrinter); -/** - * @brief print row max id vctor of each layer - * - * The config file api is maxid_printer_evaluator. - */ -class MaxIdPrinter : public NotGetableEvaluator { - private: - IVectorPtr maxIds_; - MatrixPtr maxValues_; - - public: - MaxIdPrinter() {} - - virtual void eval(const NeuralNetwork& nn) { - for (const std::string& name : config_.input_layers()) { - const Argument& argu = nn.getLayer(name)->getOutput(); - if (argu.value) { - size_t height = argu.value->getHeight(); - size_t width = config_.num_results(); - IVector::resizeOrCreate(maxIds_, height * width, false); - Matrix::resizeOrCreate(maxValues_, height, width, false); - argu.value->rowMax(*maxIds_, *maxValues_); - std::ostringstream os; - int* ids = maxIds_->getData(); - real* values = maxValues_->getData(); - for (size_t i = 0; i < height; ++i) { - for (size_t j = 0; j < width; ++j) { - size_t pos = i * width + j; - os << ids[pos] << " : " << values[pos] << ", "; - } - os << std::endl; - } - LOG(INFO) << "layer=" << name << " row max id vector:\n" << os.str(); - } - } - } - - virtual void updateSamplesNum(const std::vector& arguments) {} - - virtual real evalImp(std::vector& arguments) { return 0; } -}; -REGISTER_EVALUATOR(max_id_printer, MaxIdPrinter); -/** - * @brief print sequence max frames of each layer - * - * The config file api is maxframe_printer_evaluator. - */ -class MaxFramePrinter : public NotGetableEvaluator { - private: - IVectorPtr maxIds_; - MatrixPtr maxValues_; - MatrixPtr value_; - - public: - MaxFramePrinter() { - value_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, false); - } - - virtual void eval(const NeuralNetwork& nn) { - for (const std::string& name : config_.input_layers()) { - const Argument& argu = nn.getLayer(name)->getOutput(); - - CHECK_EQ(argu.value->getWidth(), 1LU); - size_t numSequences = argu.getNumSequences(); - const int* starts = argu.sequenceStartPositions->getData(false); - - std::ostringstream os; - for (size_t i = 0; i < numSequences; ++i) { - size_t offset = starts[i]; - size_t size = starts[i + 1] - starts[i]; - value_->setData(argu.value->getData() + offset, 1LU, size); - - size_t height = 1LU; - size_t width = std::min((size_t)config_.num_results(), size); - IVector::resizeOrCreate(maxIds_, height * width, false); - Matrix::resizeOrCreate(maxValues_, height, width, false); - - value_->rowMax(*maxIds_, *maxValues_); - - int* ids = maxIds_->getData(); - real* values = maxValues_->getData(); - for (size_t j = 0; j < width; ++j) { - os << ids[j] << " : " << values[j] << ", "; - } - os << "total " << size << " frames" << std::endl; - } - LOG(INFO) << "layer=" << name << " sequence max frames:\n" << os.str(); - } - } - - virtual void updateSamplesNum(const std::vector& arguments) {} - - virtual real evalImp(std::vector& arguments) { return 0; } -}; -REGISTER_EVALUATOR(max_frame_printer, MaxFramePrinter); - -/** - * @brief print text according to index matrix and a dictionary. - * - * There can be multiple input to this layer: - * - If there is only one input, the input must be a matrix containing - * the sequence of indices; - * - If there are more than one input, the first input should be ids, - * and are interpreted as sample ids. - * - * The output format will be: - * - * - sequence without sub-sequence, and there is probability. - * - * @code - * id \t prob space_seperated_tokens_from_dictionary_according_to_seq - * @endcode - * - * - sequence without sub-sequence, and there is not probability. - * - * @code - * id \t space_seperated_tokens_from_dictionary_according_to_seq - * @endcode - * - * - sequence with sub-sequence, and there is not probability. - * - * @code - * id \t space_seperated_tokens_from_dictionary_according_to_sub_seq - * \t \t space_seperated_tokens_from_dictionary_according_to_sub_seq - * ... - * @endcode - * - * Typically SequenceTextPrinter layer takes output of maxid or RecurrentGroup - * with maxid (when generating) as an input. - * - * The config file api is seqtext_printer_evaluator. - * - */ -class SequenceTextPrinter : public NotGetableEvaluator { - private: - /// dict_file, which contains a list of tokens - std::vector dict_; - /// result_file, which is the output file - std::ofstream os_; - /// True/False, to indicate whether to use space to separate output tokens. - /// Default is True. No space is added if set to False. - bool delimited_; - /// store the cpu version of argument.ids - std::vector cpuIds_; - /// store the probability associated with each sequence - std::vector cpuIn_; - - public: - SequenceTextPrinter() {} - - virtual void init(const EvaluatorConfig& config) { - Evaluator::init(config); - if (!config.dict_file().empty()) { - loadFileList(config.dict_file(), dict_); - } - - os_.open(config.result_file(), std::ofstream::trunc); - CHECK(os_.is_open()) << "Failed to open file " << config.result_file(); - delimited_ = config.delimited(); - } - - virtual void updateSamplesNum(const std::vector& arguments) {} - - virtual real evalImp(std::vector& arguments) { - CHECK_GE(arguments.size(), 1LU); - bool hasId = arguments.size() > 1; - size_t numSequences = arguments[0].getNumSequences(); - if (hasId) { - CHECK_EQ(arguments[0].ids->getSize(), numSequences) - << "first input must be sample id."; - } - for (size_t i = hasId ? 1 : 0; i < arguments.size(); ++i) { - CHECK_EQ((size_t)arguments[i].getNumSequences(), numSequences); - } - - auto resizeVector = [](IVectorPtr& dest, const IVectorPtr& src) { - if (src && src->useGpu()) { - IVector::resizeOrCreate(dest, src->getSize(), false); - dest->copyFrom(*src); - } else { - dest = src; - } - }; - - auto resizeMatrix = [](MatrixPtr& dest, const MatrixPtr& src) { - if (src && src->useGpu()) { - Matrix::resizeOrCreate( - dest, src->getHeight(), src->getWidth(), false, false); - dest->copyFrom(*src); - } else { - dest = src; - } - }; - - cpuIds_.resize(arguments.size()); - cpuIn_.resize(arguments.size()); - for (size_t i = 0; i < arguments.size(); ++i) { - resizeVector(cpuIds_[i], arguments[i].ids); - resizeMatrix(cpuIn_[i], arguments[i].in); - } - - int* sampleIds = nullptr; - if (hasId) { - sampleIds = cpuIds_[0]->getData(); - } - - for (size_t i = 0; i < numSequences; ++i) { - os_ << (hasId ? sampleIds[i] : i); - for (size_t j = hasId ? 1 : 0; j < arguments.size(); ++j) { - int* output = cpuIds_[j]->getData(); - const int* starts = arguments[j].sequenceStartPositions->getData(false); - - auto seqPrint = [&](int start, int end) { - os_ << "\t"; - for (int k = start; k < end; k++) { - int id = output[k]; - os_ << (delimited_ ? " " : ""); - if (!dict_.empty()) { - CHECK_LT((size_t)id, dict_.size()); - os_ << dict_[id]; - } else { - os_ << id; - } - } - }; - - if (arguments[j].hasSubseq()) { - // print sequence with sub-sequence - const int* subStarts = - arguments[j].subSequenceStartPositions->getData(false); - int subSeqId_start = 0; - int subSeqId_end = 0; - for (size_t k = 0; k < (size_t)arguments[j].getNumSubSequences() + 1; - ++k) { - if (starts[i] == subStarts[k]) subSeqId_start = k; - if (starts[i + 1] == subStarts[k]) subSeqId_end = k; - } - for (int k = subSeqId_start; k < subSeqId_end; k++) { - seqPrint(subStarts[k], subStarts[k + 1]); - os_ << std::endl; - } - - } else { - // print sequence without sub-sequence - if (arguments[j].in) { // beam print - real* probs = cpuIn_[j]->rowBuf(i); - os_ << std::endl; - int start = starts[i]; - int seqEnd = starts[i + 1]; - for (size_t k = 0; k < arguments[j].in->getWidth(); ++k) { - if (start == seqEnd) { - break; - } - int end = start + output[start] + 2; - CHECK_LE(end, seqEnd); - CHECK_EQ(output[end - 1], -1); - os_ << k << "\t" << probs[k]; - seqPrint(start + 1, end - 1); - os_ << std::endl; - start = end; - } - } else { - seqPrint(starts[i], starts[i + 1]); - } - } - } - os_ << std::endl; - } - return 0; - } -}; -REGISTER_EVALUATOR(seq_text_printer, SequenceTextPrinter); -/** - * @brief print classification error. - * - * The config file api is classification_error_printer_evaluator. - */ -class ClassificationErrorPrinter : public ClassificationErrorEvaluator { - public: - virtual void updateSamplesNum(const std::vector& arguments) {} - - virtual real evalImp(std::vector& arguments) { - MatrixPtr errorMat = calcError(arguments); - - std::ostringstream os; - errorMat->print(os); - LOG(INFO) << "Printer=" << config_.name() << " Classification Error:\n" - << os.str(); - - if (auto startPos = arguments[0].sequenceStartPositions) { - std::ostringstream os; - startPos->getVector(false)->print(os, startPos->getSize()); - LOG(INFO) << "Printer=" << config_.name() << " sequence pos vector:\n" - << os.str(); - } - return 0; - } -}; -REGISTER_EVALUATOR(classification_error_printer, ClassificationErrorPrinter); - -std::string DummyEvaluator::getTypeImpl() const { return "dummy"; } - -} // namespace paddle diff --git a/paddle/gserver/evaluators/Evaluator.h b/paddle/gserver/evaluators/Evaluator.h deleted file mode 100644 index 42948f1097d9a12600f4b11646a47e45b9bf4e96..0000000000000000000000000000000000000000 --- a/paddle/gserver/evaluators/Evaluator.h +++ /dev/null @@ -1,510 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "ModelConfig.pb.h" -#include "paddle/parameter/Argument.h" -#include "paddle/pserver/ParameterClient2.h" -#include "paddle/utils/ClassRegistrar.h" -#include "paddle/utils/Error.h" - -namespace paddle { - -class NeuralNetwork; -/** - * @def REGISTER_EVALUATOR - * @brief Macro for registering evaluator class - */ - -#define REGISTER_EVALUATOR(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name([]() { \ - Evaluator::registrar_.registerClass<__class_name>(#__type_name); \ - }) -/** - * @brief Base class for Evaluator - * Evaluating the performance of a model is very important. - * It indicates how successful the scores(predictions) of a datasets - * has been by a trained model. - */ -class Evaluator { - public: - static Evaluator* create(const EvaluatorConfig& config); - - Evaluator() : numSamples_(0), totalScore_(0) {} - - virtual ~Evaluator() {} - - virtual void init(const EvaluatorConfig& config) { config_ = config; } - - /** - * @brief start to evaluate some data - */ - virtual void start() { - numSamples_ = 0; - totalScore_ = 0; - } - - /** - * @brief Process a batch of data. - */ - virtual void eval(const NeuralNetwork& nn); - - /** - * @brief Process a batch of data. - * @return the score for the batch if it make sense to sum the score across - * batches. - * @note Otherwise evaluator should return 0 and override finish() and - * printStats() to do the right calculation. - */ - virtual real evalImp(std::vector& arguments) = 0; - - /** - * @brief Update the number of processed samples - */ - virtual void updateSamplesNum(const std::vector& arguments) { - numSamples_ += arguments[0].getBatchSize(); - } - - /// finish() should be called before distributeEval - virtual void distributeEval(ParameterClient2* client) { - LOG(FATAL) << "Not implemeted"; - } - - void mergeResultsOfAllClients(ParameterClient2* client) { - double data[2] = {totalScore_, numSamples_}; - client->reduce(data, data, 2, FLAGS_trainer_id, 0); - totalScore_ = data[0]; - numSamples_ = data[1]; - } - - /** - * @brief finish the evaluation. - */ - virtual void finish() {} - - /** - * @brief print the statistics of evaluate result - * @note finish() should be called before printStats - */ - virtual void printStats(std::ostream& os) const { - os << config_.name() << "=" - << (numSamples_ ? totalScore_ / numSamples_ : 0); - } - - friend std::ostream& operator<<(std::ostream& os, - const Evaluator& evaluator) { - evaluator.printStats(os); - return os; - } - - friend std::ostream&& operator<<(std::ostream&& os, // NOLINT - const Evaluator& evaluator) { - evaluator.printStats(os); - return std::move(os); - } - - static ClassRegistrar registrar_; - - /** - * @brief getNames will return all field names of current evaluator. - * - * The format of name is `evaluator_name.evaluator_fields`. If the evaluator - * has multiple field, the name could be `evaluator_name.field1`. For example - * the PrecisionRecallEvaluator contains `precision`, `recall` fields. The get - * names will return `precision_recall_evaluator.precision`, - * `precision_recall_evaluator.recal`, etc. - * - * Also, if current Evaluator is a combined evaluator. getNames will return - * all names of all evaluators inside the combined evaluator. - * - * @param names [out]: the field names of current evaluator. - * @note Never clear the names parameter inside getNames. - */ - virtual void getNames(std::vector* names) { - names->push_back(config_.name()); - } - - /** - * @brief getValue will return the current evaluate value of one field. - * - * @param name: The field name of current evaluator. - * @param err [out]: The error state. - * - * @return The evaluate value(metric). - */ - virtual real getValue(const std::string& name, Error* err) const { - if (name != config_.name()) { - *err = Error("no such name of evaluator %s", name.c_str()); - return .0f; - } - return this->getValueImpl(); - } - - /** - * @brief getType will return the evaluator type by field name. - * - * Evaluate Type is the current type of evaluator in string. Such as 'auc', - * 'precision_recall'. In combined evaluator, different name may get different - * evaluate type because it could be evaluated by different evaluator inside. - * - * @param name: The field name of current Evaluator. - * @param err: The error state. nullptr means don't care. - * @return the evaluator type string. - */ - virtual std::string getType(const std::string& name, Error* err) const { - if (name != config_.name()) { - *err = Error("no such name of evaluator %s", name.c_str()); - return std::string(); - } - return this->getTypeImpl(); - } - - protected: - /** - * @brief getValueImpl The simplest way to define getValue result. If this - * evaluator doesn't contain multiple fields, and do not throw any error, just - * implemented this method to get the evaluate result(metric). - * @return Evaluate result(metric). - */ - virtual real getValueImpl() const { - return numSamples_ != .0 ? totalScore_ / numSamples_ : .0; - } - - /** - * @brief getTypeImpl The simplest way to define getType result. If this - * evaluator doesn't combine many evaluators, the get type should only return - * itself type. - * @return Evaluator type. - */ - virtual std::string getTypeImpl() const { return "base"; } - - protected: - EvaluatorConfig config_; - double numSamples_; - double totalScore_; -}; - -/** - * @brief The NotGetableEvaluator class is the base class of evaluator that - * cannot get value in runtime. The most NotGetableEvaluator is Printer - * Evaluator, which is only used to debug network configuration. - */ -class NotGetableEvaluator : public Evaluator { - // Evaluator interface - public: - void getNames(std::vector* names) {} - - real getValue(const std::string& name, Error* err) const { - *err = Error("Not implemented"); - return .0f; - } - - std::string getType(const std::string& name, Error* err) const { - *err = Error("Not implemented"); - return ""; - } -}; - -class DummyEvaluator : public Evaluator { - public: - DummyEvaluator() {} - virtual void init(const EvaluatorConfig&) {} - virtual void start() {} - virtual void eval(const NeuralNetwork&) {} - virtual real evalImp(std::vector& arguments) { - (void)arguments; - return -1; - } - virtual void finish() {} - virtual void printStats(std::ostream&) const {} - - // Evaluator interface - protected: - std::string getTypeImpl() const; -}; -/** - * @brief evaluate AUC using colIdx-th column as prediction. - * The AUC(Area Under the Curve) is a common evaluation metric - * for binary classification problems. It computes the area under - * the receiver operating characteristic(ROC) curve. - * - * @note colIdx-th column - * - * - colIdx = 0: the 0-th column. - * - colIdx > 0: the colIdx-th column. - * - colIdx < 0: the last colIdx-th column. - * - * The config file api is auc_evaluator. - * - */ -class AucEvaluator : public Evaluator { - public: - AucEvaluator(int32_t colIdx) - : colIdx_(colIdx), - realColumnIdx_(0), - cpuOutput_(nullptr), - cpuLabel_(nullptr), - cpuWeight_(nullptr) {} - - virtual void start(); - - virtual real evalImp(std::vector& arguments); - - virtual void printStats(std::ostream& os) const { - os << config_.name() << "=" << calcAuc(); - } - - virtual void distributeEval(ParameterClient2* client); - - private: - static const uint32_t kBinNum_ = (1 << 24) - 1; - static const int kNegativeLabel_ = 0; - double statPos_[kBinNum_ + 1]; - double statNeg_[kBinNum_ + 1]; - int32_t colIdx_; - uint32_t realColumnIdx_; - MatrixPtr cpuOutput_; - IVectorPtr cpuLabel_; - MatrixPtr cpuWeight_; - - AucEvaluator() {} - - inline static double trapezoidArea(double X1, - double X2, - double Y1, - double Y2) { - return (X1 > X2 ? (X1 - X2) : (X2 - X1)) * (Y1 + Y2) / 2.0; - } - - double calcAuc() const; - - // Evaluator interface - protected: - real getValueImpl() const; - std::string getTypeImpl() const; -}; - -/** - * @brief RankAucEvaluator calculates the AUC of each list (i.e., titles - * under the same query), and averages them. Each list should be organized - * as a sequence. The inputs of this evaluator is [output, click, pv]. If pv - * is not provided, it will be set to 1. The types of click and pv are - * dense value. - */ -class RankAucEvaluator : public Evaluator { - public: - // evaluate ranking AUC - virtual void start(); - - virtual void updateSamplesNum(const std::vector& arguments); - - virtual real evalImp(std::vector& arguments); - - virtual void distributeEval(ParameterClient2* client) { - mergeResultsOfAllClients(client); - } - - private: - MatrixPtr output_; - MatrixPtr click_; - MatrixPtr pv_; - std::vector> outputPair_; - - double calcRankAuc(real* outputData, - real* clickData, - real* pvData, - size_t size); - - // Evaluator interface - protected: - std::string getTypeImpl() const; -}; - -/** - * @brief precision, recall and f1 score Evaluator - * \f[ - * precision = \frac{tp}{tp+tn} \\ - * recall=\frac{tp}{tp+fn} \\ - * f1=2*\frac{precsion*recall}{precision+recall} - * \f] - * - * The config file api is precision_recall_evaluator. - */ -class PrecisionRecallEvaluator : public Evaluator { - public: - // Evaluate precision, recall and F1 score - PrecisionRecallEvaluator() - : isMultiBinaryLabel_(false), - cpuOutput_(nullptr), - cpuLabel_(nullptr), - cpuWeight_(nullptr) {} - - virtual void start(); - - virtual real evalImp(std::vector& arguments); - - virtual void printStats(std::ostream& os) const; - - virtual void distributeEval(ParameterClient2* client); - - void getNames(std::vector* names); - - real getValue(const std::string& name, Error* err) const; - - std::string getType(const std::string& name, Error* err) const; - - struct StatsInfo { - /// numbers of true positives - double TP; - /// numbers of true negatives - double TN; - /// numbers of false positives - double FP; - /// numbers of false negatives - double FN; - - StatsInfo() : TP(0.0), TN(0.0), FP(0.0), FN(0.0) {} - }; - - private: - bool isMultiBinaryLabel_; - std::vector statsInfo_; - - MatrixPtr cpuOutput_; - IVectorPtr cpuLabel_; - MatrixPtr cpuWeight_; - - struct PrintStatsInfo { - double precision; - double recall; - double f1; - double macroAvgPrecision; - double macroAvgRecall; - double macroAvgF1Score; - double microAvgPrecision; - double microAvgRecall; - double microAvgF1Score; - }; - - bool getStatsInfo(PrintStatsInfo* info) const; - - void calcStatsInfo(const MatrixPtr& output, - const IVectorPtr& label, - const MatrixPtr& weight); - - void calcStatsInfoMulti(const MatrixPtr& output, - const MatrixPtr& label, - const MatrixPtr& weight); - - inline static double calcPrecision(double TP, double FP) { - if (TP > 0.0 || FP > 0.0) { - return TP / (TP + FP); - } else { - return 1.0; - } - } - - inline static double calcRecall(double TP, double FN) { - if (TP > 0.0 || FN > 0.0) { - return TP / (TP + FN); - } else { - return 1.0; - } - } - - inline static double calcF1Score(double precision, double recall) { - if (precision > 0.0 || recall > 0.0) { - return 2 * precision * recall / (precision + recall); - } else { - return 0; - } - } - - mutable std::unordered_map values_; - - void storeLocalValues() const; -}; - -/* - * @brief positive-negative pair rate Evaluator - * - * The config file api is pnpair_evaluator. - */ -class PnpairEvaluator : public Evaluator { - public: - PnpairEvaluator() - : cpuOutput_(nullptr), - cpuLabel_(nullptr), - cpuInfo_(nullptr), - cpuWeight_(nullptr) {} - - virtual void start(); - virtual real evalImp(std::vector& arguments); - - struct PredictionResult { - PredictionResult(real __out, int __label, int __queryid, real __weight) - : out(__out), label(__label), queryid(__queryid), weight(__weight) {} - real out; - int label; - int queryid; - real weight; - }; - std::vector predictArray_; - void printPredictResults() { - std::ofstream fs(FLAGS_predict_file); - CHECK(fs) << "Fail to open " << FLAGS_predict_file; - for (auto& res : predictArray_) { - fs << res.out << " " << res.label << " " << res.queryid << std::endl; - } - } - - void stat(size_t start, - size_t end, - PredictionResult* answers, - double& pos, - double& neg, - double& spe); - void calc(std::vector& predictArray); - - virtual void finish() { calc(predictArray_); } - - virtual void printStats(std::ostream& os) const { - os << " pos/neg=" << this->getValueImpl(); - } - - virtual void distributeEval(ParameterClient2* client) { - client->reduce(pairArray_, pairArray_, kPairArrayNum_, FLAGS_trainer_id, 0); - LOG(INFO) << " distribute eval calc total pos pair: " << pairArray_[0] - << " calc total neg pair: " << pairArray_[1]; - } - - private: - static const uint32_t kPairArrayNum_ = 2; - double pairArray_[kPairArrayNum_]; - MatrixPtr cpuOutput_; - IVectorPtr cpuLabel_; - IVectorPtr cpuInfo_; - MatrixPtr cpuWeight_; - - // Evaluator interface - protected: - real getValueImpl() const { - return pairArray_[0] / ((pairArray_[1] <= 0) ? 1.0 : pairArray_[1]); - } - std::string getTypeImpl() const; -}; - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/GradientMachine.cpp b/paddle/gserver/gradientmachines/GradientMachine.cpp deleted file mode 100644 index 654024e8a47c1e538f25823da78dce6a7a093975..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/GradientMachine.cpp +++ /dev/null @@ -1,104 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "GradientMachine.h" - -#include -#include "paddle/utils/Logging.h" - -#include "NeuralNetwork.h" -#include "hl_gpu.h" - -#ifndef PADDLE_MOBILE_INFERENCE -#include "GradientMachineMode.h" -#include "MultiGradientMachine.h" -#include "MultiNetwork.h" -#include "ParallelNeuralNetwork.h" -#endif - -namespace paddle { - -GradientMachine* GradientMachine::create( - const ModelConfig& config, - int mode, - const std::vector& parameterTypes) { -#ifndef PADDLE_MOBILE_INFERENCE - if (auto gm = IGradientMachineMode::tryCreateGradientMachine(mode, config)) { - return gm; - } - if (FLAGS_trainer_count > 1) { - return new MultiGradientMachine(config, FLAGS_use_gpu); - } -#endif - if (FLAGS_trainer_count == 1) { // single -#ifndef PADDLE_MOBILE_INFERENCE - NeuralNetwork* nn; - if (config.type() == "multi_nn") { - /* multi submodel calculate, thread(s) will be initialized inside */ - nn = new MultiNetwork("root"); - } else if (FLAGS_parallel_nn) { - /* multi threads calculate */ - nn = new ParallelNeuralNetwork(); - } else { - /* single thread calculate */ - nn = NeuralNetwork::create(config); - } -#else - NeuralNetwork* nn = NeuralNetwork::create(config); -#endif - ParamInitCallback testParamInitCb = [](int paramId, Parameter* para) { - para->enableType(PARAMETER_VALUE); - }; - nn->init( - config, mode == kTesting ? testParamInitCb : nullptr, parameterTypes); - return nn; - } - LOG(FATAL) << "Unknown model type: " << config.type(); - return nullptr; -} - -void GradientMachine::saveParameters(const std::string& dir) const { - LOG(INFO) << "Saving parameters to " << dir; - - for (auto& para : parameters_) { - std::string filename = dir + "/" + para->getName(); - if (para->isFullSize()) { - para->save(filename); - } - } -} - -void GradientMachine::loadParameters(const std::string& dir) { - LOG(INFO) << "Loading parameters from " << dir; - - for (auto& para : parameters_) { - std::string filename = dir + "/" + para->getName(); - if (para->isFullSize()) { - para->load(filename); - } - } -} - -void GradientMachine::randParameters() { - LOG(INFO) << "Initing parameters.."; - - for (auto& para : parameters_) { - if (para->isFullSize()) { - para->randomize(); - } - } - LOG(INFO) << "Init parameters done."; -} - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/GradientMachine.h b/paddle/gserver/gradientmachines/GradientMachine.h deleted file mode 100644 index 22cf5d265f429ecbcea1808a54c85d7e89f8bc99..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/GradientMachine.h +++ /dev/null @@ -1,250 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include - -#include "ModelConfig.pb.h" -#include "TrainerConfig.pb.h" -#include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/gserver/layers/Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/parameter/ParameterUpdaterBase.h" -#include "paddle/utils/Thread.h" - -#ifndef PADDLE_MOBILE_INFERENCE -#include "paddle/gserver/evaluators/Evaluator.h" -#endif - -namespace paddle { -/** - * @brief A gradient machine is capable of calculating some outputs given - * some inputs and performing gradient calculation based on the - * derivative from the outputs. - * - * A gradient machine can be either a full neural network or part of a neural - * network. - * - * Usage for training: - * - * 1. Prepare inArgs. Put your input data into inArgs[i].value. - * - * 2. Call forward(inArgs, &outArgs) - * - * 3. Calculate gradient with respect to outArgs[i]->value - * and fill them into outArgs[i]->grad. - * This step can be skipped if your the outputs are from cost layers. - * - * 4. Call backward(). After backward, gradient of each parameter is - * accumulated to getParameters()[i]->getBuf(PARAMETER_GRADIENT) - * - * 5. Update parameter value getParameters()[i]->getBuf(PARAMETER_VALUE) using - * gradients. - * - * 6. Clear gradients to zero. - * - * Usage for prediction: - * - * 1. Prepare inArgs. Put your input data into inArgs[i].value. - * - * 2. Call forward(inArgs, &outArgs) - * - * 3. Obtain the prediction result from outArgs[i] - */ - -typedef std::vector MachineState; - -class GradientMachine; - -typedef std::shared_ptr GradientMachinePtr; - -class GradientMachine { - public: - enum CreateMode { - kNormal = 0, - kSgdSparseCpuTraining = 3, - kTesting = 4, - kCustom = 10 - }; - - /** - * Create a gradient machine from ModelConfig - * Parameter will have parameterTypes - */ - static GradientMachine* create( - const ModelConfig& config, - int mode = kNormal, - const std::vector& parameterTypes = - std::vector{ - PARAMETER_VALUE, PARAMETER_GRADIENT, PARAMETER_MOMENTUM}); - - virtual ~GradientMachine() {} - - /** - * Prefetch row ids of sparse parameter. - */ - virtual void prefetch(const std::vector& inArgs) { (void)inArgs; } - - /** - * @brief Forward propagation. - * - * Calculate outputs (outArgs) based the inputs (inArgs) - * - * @note: if passType==PASS_TEST, then backward() should not be called - */ - virtual void forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType) = 0; - - /** - * @brief Backward propagation. - * - * Calculate the gradient of inArgs and parameter. - * - * This function should only be called after a corresponding forward() call. - * The caller is responsible for filling the correct grad for the outArgs - * obtained using forward(). - * - * It may also change the grad field for the inArgs supplied at forward() - */ - virtual void backward(const UpdateCallback& callback = nullptr) = 0; - - /** - * Combine forward() and backward(). For multithread training, this - * may be faster. - * - * @note: passType PASS_TEST is not allowed for forwardBackward(). - */ - virtual void forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback = nullptr) { - forward(inArgs, outArgs, passType); - backward(callback); - } - - virtual Argument getLayerOutput(const std::string& layerName) = 0; - - // see comment in Layer.h for the function with the same name - virtual void resetState() {} - - // set machine state - virtual void setState(const MachineState& machineState) {} - - // save machine state - virtual void getState(MachineState& machineState) {} - - virtual void onPassEnd() = 0; - -#ifndef PADDLE_MOBILE_INFERENCE - /** - * Create an evaluator which can be used for eval() - */ - virtual Evaluator* makeEvaluator() const = 0; - - /** - * evaluate using the given evaluator - */ - virtual void eval(Evaluator* evaluator) const = 0; -#endif - - std::vector& getParameters() { return parameters_; } - - std::vector& getNonStaticParameters() { - if (nonStaticParameters_.empty()) { - for (auto para : parameters_) { - if (!para->isStatic()) { - nonStaticParameters_.push_back(para); - } - } - } - return nonStaticParameters_; - } - - inline bool hasStaticParameters() { - return parameters_.size() != getNonStaticParameters().size(); - } - - /** - * @brief Used before formal training, start work-threads and set - * trainer Parameters; - * - * @note This function will only been implemented and used in a - * multithreaded environment. - */ - virtual void start() {} - - /** - * @brief check each work-thread whether is failed/error/finish, - * if not, return ture, and yes return false. - * - * @note This function will only been implemented and used in a - * multithreaded environment. - */ - virtual void finish() {} - - /** - * @brief set the training status a "finished" value, the sub_work_threads - * will option the change, and then exit. - * - * @note This function will only been implemented and used in a - * multithreaded environment. - */ - virtual bool trainIsOn() { return true; } - - /** - * @brief when all or some of the sub-workThreads are suspended to waiting - * controller's instructions, and after some processing done in the - * controller, it will call this function to wake up all the pending - * thread. - * - * @note This function will only been implemented and used in a - * multithreaded environment. - */ - virtual void restart() {} - - /// Set the gradient of the output from outside. - virtual void setOutputGrad(const std::vector& args) { - LOG(FATAL) << "Not implemented!"; - } - - void saveParameters(const std::string& dir) const; - - void loadParameters(const std::string& dir); - - void randParameters(); - - virtual void getStats(real& cost, int64_t& numProcessed) { - (void)cost; - (void)numProcessed; - } - - /** - * @brief Release the middle layer's output memory. - * - * @note This function is used for memory optimization in inference. - */ - virtual void releaseOutput() {} - - protected: - virtual void onLoadParameter() {} - - std::vector parameters_; - std::vector nonStaticParameters_; -}; - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp deleted file mode 100644 index b8d4d28f0f309a5f7348605e8d35e160e7fd5552..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp +++ /dev/null @@ -1,894 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MultiGradientMachine.h" - -#include "paddle/utils/Logging.h" - -#include "paddle/utils/Stat.h" - -#include "NeuralNetwork.h" -#include "ParallelNeuralNetwork.h" - -DEFINE_bool(allow_only_one_model_on_one_gpu, - true, - "If true, do not allow multiple models on one GPU device"); - -namespace paddle { - -// get types of the parameters which need to be merged after backward() -static void fillMergeTypes(PassType passType, - std::vector* mergeTypes) { - mergeTypes->clear(); - if (passType != PASS_TEST) { - mergeTypes->push_back(PARAMETER_GRADIENT); - } -} - -MultiGradientMachine::MultiGradientMachine(const ModelConfig& config, - bool useGpu) - : useGpu_(useGpu), - trainerBarrier_(FLAGS_trainer_count), - allBarrier_(FLAGS_trainer_count + 1), - inArgsCopied_(false) { - isPassGrad_ = false; - numThreads_ = FLAGS_trainer_count; - if (useGpu) { - //! TODO(yuyang18): When useGpu=false && paddle is not compiled with gpu, - //! the hl_get_device_count will get an error result. It seems should return - //! 0 when hppl is not compiled as gpu version. - numDevices_ = hl_get_device_count(); - } else { - numDevices_ = 0; - } - ParamInitCallback mainParamInitCb = [](int paramId, Parameter* para) { - // only create buf for CPU parameters - // GPU parameters will be created in each thread - if (para->useGpu()) return; - - if (para->isSparseRemoteUpdate()) { - para->enableType(PARAMETER_VALUE, - FLAGS_loadsave_parameters_in_pserver - ? Parameter::MAT_SPARSE_ROW_PREFETCH - : Parameter::MAT_SPARSE_ROW_PREFETCH_FULL_SIZE); - para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW); - } else if (para->isGradSparseUpdate()) { - para->enableType(PARAMETER_VALUE); - para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW_IDS); - SparseRowIdsCpuMatrix* mat = dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get()); - mat->setNumOfThreads(FLAGS_trainer_count); - } else if (para->isValueShared()) { - para->enableType(PARAMETER_VALUE, Parameter::MAT_VALUE_SHARED); - if (!para->isStatic()) { - para->enableType(PARAMETER_GRADIENT); - } - } else { - para->enableType(PARAMETER_VALUE); - if (!para->isStatic()) { - para->enableType(PARAMETER_GRADIENT); - } - } - }; - - NeuralNetwork* nn = NeuralNetwork::create(config); - nn->init(config, mainParamInitCb); - gradientMachine_.reset(nn); - parameters_ = gradientMachine_->getParameters(); - - numLogicalDevices_ = 0; - if (useGpu_) { - numLogicalDevices_ = 1; - - for (size_t pid = 0; pid < parameters_.size(); pid++) { - if (parameters_[pid]->getConfig().device() + 1 > numLogicalDevices_) { - numLogicalDevices_ = parameters_[pid]->getConfig().device() + 1; - } - } - LOG(INFO) << "numLogicalDevices=" << numLogicalDevices_ - << " numThreads=" << numThreads_ << " numDevices=" << numDevices_; - - if (numLogicalDevices_ * numThreads_ > numDevices_ && - FLAGS_allow_only_one_model_on_one_gpu) { - LOG(FATAL) << "trainer_count * num_devices_in_model " - << "(" << numThreads_ << "*" << numLogicalDevices_ << ")" - << "=" << numThreads_ * numLogicalDevices_ - << " exceeds number of GPU devices(" << numDevices_ << ")"; - } - numLogicalDevices_ = std::min(numLogicalDevices_, numDevices_); - - /* Enables direct access to memory allocations on a peer device */ - for (int i = 0; i < numThreads_; i++) { - for (int d = 0; d < numLogicalDevices_; ++d) { - enablePeerAccess(logicalDeviceId2RealDeviceId(d, i), - logicalDeviceId2RealDeviceId(d, i + 1)); - enablePeerAccess(logicalDeviceId2RealDeviceId(d, i), - logicalDeviceId2RealDeviceId(d, i - 1)); - } - } - } - - for (int i = 0; i < numThreads_; ++i) { - threads_.emplace_back(new TrainerThread(config, i, this)); - } - - bufferSizes_.resize(numLogicalDevices_, 0); - paraMainThread_.reserve(parameters_.size()); - int pid = 0; - for (auto& para : parameters_) { - if (para->isStatic() || !para->useGpu()) { - paraMainThread_.push_back(0); - } else { - int end = pid++ % numThreads_; - paraMainThread_.push_back(end); - int paraDeviceId = para->getDeviceId(); - if (paraDeviceId == -1) paraDeviceId = 0; - paraDeviceId = paraDeviceId % numLogicalDevices_; - if (para->getSize() > bufferSizes_[paraDeviceId]) { - bufferSizes_[paraDeviceId] = para->getSize(); - VLOG(1) << "bufferSize[" << paraDeviceId << "]" << para->getSize(); - } - } - } - - // TODO(xuwei06) Instead of using maximal buffer size, we may use a smaller - // fixed buffer size and use pipeline to dispatch parameter value and merge - // parameter gradient, which may be faster. - - // combination of all trainers mainPara into GradientMachine parameters - hasNonstaticCpuParamters_ = false; - for (size_t pid = 0; pid < parameters_.size(); pid++) { - if (parameters_[pid]->useGpu()) { - parameters_[pid] = threads_[paraMainThread_[pid]]->getParameters()[pid]; - } else if (!parameters_[pid]->isStatic()) { - hasNonstaticCpuParamters_ = true; - } - } - - gradBufs_.resize(numThreads_); - for (int i = 0; i < numThreads_; ++i) { - gradBufs_[i].resize(numLogicalDevices_); - for (int d = 0; d < numLogicalDevices_; ++d) { - gradBufs_[i][d].sem.post(); - } - } - - outArgStream_ = HPPL_STREAM_1; - - start(); -} - -void MultiGradientMachine::start() { - for (auto& thread : threads_) { - thread->start(); - } -} - -void MultiGradientMachine::finish() { - for (auto& thread : threads_) { - thread->stop(); - } -} - -std::vector*> -MultiGradientMachine::getSlaveParameters() { - std::vector*> vec; - vec.reserve(threads_.size()); - for (auto& thread : threads_) { - vec.push_back(&thread->getParameters()); - } - return vec; -} - -void MultiGradientMachine::notifyGradientTransfer(int paramId) { - gradQueue_.enqueue(paramId); -} - -void MultiGradientMachine::allocGradBufs() { - if (numLogicalDevices_ == 0) return; - if (gradBufs_[0][0].bufs.size() >= mergeTypes_.size()) return; - - for (int i = 0; i < numThreads_; i++) { - for (int d = 0; d < numLogicalDevices_; ++d) { - if (bufferSizes_[d] == 0) continue; - SetDevice device(logicalDeviceId2RealDeviceId(d, i)); - for (size_t j = 0; j < mergeTypes_.size(); j++) { - gradBufs_[i][d].bufs.push_back( - Vector::create(bufferSizes_[d], /* useGpu= */ true)); - } - } - } -} - -void MultiGradientMachine::prefetch(const std::vector& inArgs) { - // Each gradient machine in threads needs to do prefetch on its own - // part of inArgs. So we need to first divide inArgs to each thread - inArgs_ = inArgs; - startTask(TASK_COPY_IN_ARGS); - - for (auto& para : parameters_) { - if (para->isSparseRemoteUpdate()) { - auto mat = dynamic_cast( - para->getMat(PARAMETER_VALUE).get()); - mat->clearIndices(); - } - } - - waitForCopyInArgs(); - - // Because SparsePrefetchRowCpuMatrix can only be changed by ONE thread - // at one time, we need to do prefetch sequentially - for (auto& thread : threads_) { - thread->prefetch(); - } - - for (auto& para : parameters_) { - if (para->isSparseRemoteUpdate()) { - auto mat = dynamic_cast( - para->getMat(PARAMETER_VALUE).get()); - mat->setupIndices(); - auto matGrad = dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get()); - matGrad->reserveStore(); - } - } -} - -void MultiGradientMachine::forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType) { - forwardImp(inArgs, outArgs, passType, TASK_FORWARD); -} - -void MultiGradientMachine::forwardImp(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - TaskType taskType) { - updateThreadParameters(); - passType_ = passType; - - if (!inArgsCopied_) { - inArgs_ = inArgs; - inArgsCopied_ = false; - } - - fillMergeTypes(passType, &mergeTypes_); - allocGradBufs(); - startTask(taskType); - - getOutArgs(outArgs, passType); -} - -void MultiGradientMachine::backward(const UpdateCallback& callback) { - backwardCallback_ = callback; - startTask(TASK_BACKWARD); - backwardImp(callback); -} - -void MultiGradientMachine::forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback) { - backwardCallback_ = callback; - forwardImp(inArgs, outArgs, passType, TASK_FORWARD_BACKWARD); - backwardImp(callback); -} - -Argument MultiGradientMachine::getLayerOutput(const std::string& layerName) { - std::vector args; - args.reserve(threads_.size()); - - for (auto& thread : threads_) { - args.push_back(thread->getGradientMachine()->getLayerOutput(layerName)); - } - outLayerArgs_.concat(args, false /* use_gpu */, outArgStream_, passType_); - - return outLayerArgs_; -} - -void MultiGradientMachine::backwardImp(const UpdateCallback& callback) { - for (size_t i = 0; i < parameters_.size(); i++) { - if (!parameters_[i]->useGpu() || parameters_[i]->isStatic()) continue; - REGISTER_TIMER("controller_dequeue"); - gradQueue_.dequeue(); - } - if (hasNonstaticCpuParamters()) { - waitAfterMerge(); - if (backwardCallback_) { - for (auto& para : parameters_) { - if (!para->useGpu() && !para->isStatic()) { - backwardCallback_(para.get()); - } - } - } - } -} - -void MultiGradientMachine::updateThreadParameters() { - for (size_t pid = 0; pid < parameters_.size(); ++pid) { - if (!parameters_[pid]->useGpu()) continue; - if (!parameters_[pid]->isValueUpdated()) continue; - parameters_[pid]->clearValueUpdated(); - for (int i = 0; i < (int)threads_.size(); i++) { - threads_[i]->incUpdateCounter(); - } - // NotifyValueReady should happen after that all threads' incUpdateCounter() - // are called so that the counters are correct when notifyValueReady() - // is called. - threads_[paraMainThread_[pid]]->notifyValueReady(pid); - } -} - -void MultiGradientMachine::onPassEnd() { - for (auto& thread : threads_) { - thread->onPassEnd(); - } -} - -Evaluator* MultiGradientMachine::makeEvaluator() const { - return threads_[0]->getGradientMachine()->makeEvaluator(); -} - -void MultiGradientMachine::eval(Evaluator* evaluator) const { - for (auto& thread : threads_) { - SetDevice device(thread->getDeviceId()); - if (thread->hasInputData()) { - thread->getGradientMachine()->eval(evaluator); - } - } -} - -void MultiGradientMachine::getOutArgs(std::vector* outArgs, - PassType passType) { - for (auto& thread : threads_) { - REGISTER_TIMER("waitOutArgs"); - thread->waitOutArgsReady(); - } - - outArgs_.resize(threads_[threads_.size() - 1]->getOutArgs().size()); - - REGISTER_TIMER("copyOutArgs"); - for (size_t i = 0; i < outArgs_.size(); ++i) { - std::vector args; - args.reserve(threads_.size()); - for (auto& thread : threads_) { - // If the thread input is empty, then the output is empty. - auto tmp = thread->getOutArgs(); - if (tmp.size() > 0) { - args.push_back(tmp[i]); - } - } - outArgs_[i].concat(args, useGpu_, outArgStream_, passType); - } - - if (useGpu_) { - hl_stream_synchronize(outArgStream_); - } - - *outArgs = outArgs_; -} - -void MultiGradientMachine::setOutputGrad(const std::vector& args) { - CHECK_EQ(args.size(), outArgs_.size()); - for (size_t i = 0; i < args.size(); i++) { - outArgs_[i].grad = args[i].grad; - } -} - -void MultiGradientMachine::startTask(TaskType taskType) { - taskType_ = taskType; - for (auto& thread : threads_) { - thread->notifyTaskReady(); - } -} - -TrainerThread::TrainerThread(const ModelConfig& config, - int threadId, - MultiGradientMachine* multiMachine) - : multiMachine_(multiMachine), - config_(config), - threadId_(threadId), - inArgsCopied_(false) { - int numThreads = multiMachine->getNumThreads(); - - auto& mainParas = multiMachine->getParameters(); - - using std::placeholders::_1; - using std::placeholders::_2; - - partnerId_ = mod(threadId_ - 1, numThreads); - - deviceId_ = !multiMachine_->useGpu() - ? -1 - : multiMachine_->logicalDeviceId2RealDeviceId(0, threadId_); - SetDevice gpuDevice(deviceId_); - - NeuralNetwork* nn = nullptr; - if (!multiMachine->useGpu() || !FLAGS_parallel_nn) { - nn = NeuralNetwork::create(config); - } else { - nn = new ParallelNeuralNetwork(); - for (auto& paraConfig : *config_.mutable_parameters()) { - if (paraConfig.device() != -1) { - paraConfig.set_device(multiMachine_->logicalDeviceId2RealDeviceId( - paraConfig.device(), threadId_)); - } - } - for (auto& layerConfig : *config_.mutable_layers()) { - if (layerConfig.device() != -1) { - layerConfig.set_device(multiMachine_->logicalDeviceId2RealDeviceId( - layerConfig.device(), threadId_)); - } - } - } - // Only GPU do not share parameter values with main paramters. - ParamInitCallback slaveParamInitCb = - std::bind(parameterInitNN, _1, _2, &mainParas); - nn->init(config_, slaveParamInitCb); - gradientMachine_.reset(nn); - parameters_ = gradientMachine_->getParameters(); - if (!FLAGS_parallel_nn) { - for (auto& para : parameters_) { - para->setDevice(deviceId_); - } - } - - backwardCallback_ = - std::bind(&TrainerThread::backwardCallback, this, std::placeholders::_1); - - gradStream_ = HPPL_STREAM_2; - valueStream_ = HPPL_STREAM_3; - stopping_ = true; - updateCounter_ = 0; - parameterUpdated_ = false; -} - -TrainerThread::~TrainerThread() { stop(); } - -void TrainerThread::start() { - if (!stopping_) return; - - stopping_ = false; - - gradientMachine_->start(); - - computeThread_.reset(new std::thread([this]() { computeThread(); })); - - if (multiMachine_->useGpu()) { - gradCollectThread_.reset( - new std::thread([this]() { gradCollectThread(); })); - - valueDispatchThread_.reset( - new std::thread([this]() { valueDispatchThread(); })); - - copyThread_.reset(new std::thread([this]() { copyGradToBufferThread(); })); - } -} - -void TrainerThread::stop() { - if (stopping_) return; - - stopping_ = true; - - if (computeThread_) { - taskReadySem_.post(); - computeThread_->join(); - } - if (gradCollectThread_) { - gradQueue_.enqueue(0); - gradCollectThread_->join(); - } - if (copyThread_) { - gradBufQueue_.enqueue(0); - copyThread_->join(); - } - if (valueDispatchThread_) { - valueReadyQueue_.enqueue(0); - valueDispatchThread_->join(); - } -} - -void TrainerThread::computeThread() { - VLOG(1) << "gradComputeThread " << threadId_; - - if (deviceId_ >= 0) { - hl_init(deviceId_); - } - - while (true) { - { - REGISTER_TIMER("taskSem_wait"); - taskReadySem_.wait(); - } - - if (stopping_) break; - - switch (multiMachine_->getTaskType()) { - case MultiGradientMachine::TASK_FORWARD_BACKWARD: - forward(); - backward(); - break; - case MultiGradientMachine::TASK_FORWARD: - forward(); - break; - case MultiGradientMachine::TASK_BACKWARD: - backward(); - break; - case MultiGradientMachine::TASK_COPY_IN_ARGS: - batchSize_ = copyInArgs(); - inArgsCopied_ = true; - multiMachine_->waitForCopyInArgs(); - break; - } - } -} - -void TrainerThread::prefetch() { - SetDevice setDevice(deviceId_); - gradientMachine_->prefetch(inArgs_); -} - -void TrainerThread::forward() { - if (!inArgsCopied_) { - REGISTER_TIMER("copyInArgs"); - batchSize_ = copyInArgs(); - } else { - inArgsCopied_ = false; - } - - if (multiMachine_->getPassType() != PASS_TEST) { - REGISTER_TIMER("clearGradient"); - // For main parameter, the user of MultiGpuSyncMachine is responsible - // for setting the gradient to zero - for (size_t i = 0; i < parameters_.size(); i++) { - if (parameters_[i]->useGpu()) { - if (multiMachine_->paraMainThread(i) != threadId_) { - SetDevice device(parameters_[i]->getDeviceId()); - parameters_[i]->clearGradient(); - } - } else { - parameters_[i]->clearGradient(); - } - } - } - - { - REGISTER_TIMER("wait_value"); - valueReadyCond_.wait([this]() { return !parameterUpdated_; }); - } - - { fillMergeTypes(multiMachine_->getPassType(), &mergeTypes_); } - - { - REGISTER_TIMER("thread_forward"); - if (batchSize_ > 0) { - gradientMachine_->forward( - inArgs_, &outArgs_, multiMachine_->getPassType()); - } else { - outArgs_.clear(); - } - } - outArgsReadySem_.post(); -} - -void TrainerThread::backward() { - REGISTER_TIMER("thread_backward"); - if (multiMachine_->isPassGrad()) { - copyOutputGrad(); - } - if (batchSize_ > 0) { - gradientMachine_->backward(backwardCallback_); - } else { - for (size_t i = parameters_.size(); i > 0; i--) { - backwardCallback(parameters_[i - 1].get()); - } - } - if (multiMachine_->hasNonstaticCpuParamters()) { - mergeCpuGradients(); - } -} - -void TrainerThread::backwardCallback(Parameter* para) { - // CPU parameters are merged in the end - if (!para->useGpu() || para->isStatic()) return; - - int paramId = para->getID(); - if (multiMachine_->getNumThreads() == 1) { - // no need to do merge if there is only one thread - doCallback(paramId); - } else if (threadId_ == mod(multiMachine_->paraMainThread(paramId) - 1, - multiMachine_->getNumThreads())) { - notifyCopyGradToBuffer(paramId); - } else { - notifyGradientCollect(paramId); - } -} - -void TrainerThread::copyGradToBufferThread() { - VLOG(1) << "copyGradToBufferThread " << threadId_; - - if (deviceId_ >= 0) { - hl_init(deviceId_); - } - auto& partnerThread = multiMachine_->getThread(partnerId_); - auto& gradBufs = multiMachine_->getGradBuf(partnerId_); - - while (true) { - int pid = gradBufQueue_.dequeue(); - if (stopping_) break; - - int pdeviceId = multiMachine_->realDeviceId2LogicalDeviceId( - parameters_[pid]->getDeviceId(), threadId_); - - auto& gradBuf = gradBufs[pdeviceId]; - - { - REGISTER_TIMER("waitBufferReady"); - gradBuf.sem.wait(); - } - - { - REGISTER_TIMER("copyGradToBuffer"); - SetDevice setDevice(parameters_[pid]->getDeviceId()); - for (size_t i = 0; i < mergeTypes_.size(); ++i) { - gradBuf.bufs[i]->resize( - parameters_[pid]->getBuf(mergeTypes_[i])->getSize()); - gradBuf.bufs[i]->copyFrom(*parameters_[pid]->getBuf(mergeTypes_[i]), - gradStream_); - } - hl_stream_synchronize(gradStream_); - } - partnerThread->notifyGradientCollect(pid); - } -} - -void TrainerThread::gradCollectThread() { - VLOG(1) << "gradCollectThread " << threadId_; - - if (deviceId_ >= 0) { - hl_init(deviceId_); - } - - std::vector gradReadyCount(parameters_.size(), 0); - - auto& gradBufs = multiMachine_->getGradBuf(threadId_); - - while (true) { - int pid = gradQueue_.dequeue(); - if (stopping_) break; - - if (++gradReadyCount[pid] < 2) continue; - gradReadyCount[pid] = 0; - int pdeviceId = multiMachine_->realDeviceId2LogicalDeviceId( - parameters_[pid]->getDeviceId(), threadId_); - - auto& gradBuf = gradBufs[pdeviceId]; - - { - REGISTER_TIMER("mergeGrad"); - for (size_t i = 0; i < mergeTypes_.size(); ++i) { - ParameterType type = mergeTypes_[i]; - const VectorPtr& localGrad = parameters_[pid]->getBuf(type); - SetDevice setDevice(parameters_[pid]->getDeviceId()); - localGrad->add(*gradBuf.bufs[i]); - } - } - - gradBuf.sem.post(); - - if (multiMachine_->paraMainThread(pid) == threadId_) { - doCallback(pid); - } else { - notifyCopyGradToBuffer(pid); - } - } -} - -void TrainerThread::doCallback(int pid) { - REGISTER_TIMER("callback"); - auto& gpuThreads = multiMachine_->getAllThreads(); - if (multiMachine_->getBackwardCallback()) { - // The callback supplied by the user of MultiGradientMachine may handle - // the parameter update using the gradient. - multiMachine_->getBackwardCallback()(parameters_[pid].get()); - if (parameters_[pid]->isValueUpdated()) { - parameters_[pid]->clearValueUpdated(); - for (auto& thread : gpuThreads) { - thread->incUpdateCounter(); - } - notifyValueReady(pid); - } - } - multiMachine_->notifyGradientTransfer(pid); -} - -void TrainerThread::valueDispatchThread() { - VLOG(1) << "valueDispatchThread " << threadId_; - - if (deviceId_ >= 0) { - hl_init(deviceId_); - } - - auto& thread = multiMachine_->getThread(partnerId_); - - while (true) { - int pid; - { - REGISTER_TIMER("value_dequeue"); - pid = valueReadyQueue_.dequeue(); - } - if (stopping_) break; - - if (multiMachine_->paraMainThread(pid) == partnerId_) continue; - - { - REGISTER_TIMER("copyValue"); - SetDevice setDevice(parameters_[pid]->getDeviceId()); - thread->getValueBuf(pid)->copyFrom(*getValueBuf(pid), valueStream_); - hl_stream_synchronize(valueStream_); - } - - thread->notifyValueReady(pid); - } -} - -void TrainerThread::notifyValueReady(int paramId) { - if (--updateCounter_ == 0) { - valueReadyCond_.notify_all([this] { parameterUpdated_ = false; }); - } - - notifyValueDispatch(paramId); -} - -int TrainerThread::copyInArgs() { - const std::vector& fullInArgs = multiMachine_->getInArgs(); - int numThreads = multiMachine_->getAllThreads().size(); - int32_t numSequences = fullInArgs[0].getNumSequences(); - int32_t startSeq = numSequences * threadId_ / numThreads; - int32_t endSeq = numSequences * (threadId_ + 1) / numThreads; - int32_t copySize = endSeq - startSeq; - - /** - * For the first copy, need to allocate space here - */ - if (inArgs_.size() == 0) { - inArgs_.resize(fullInArgs.size()); - } - - if (copySize == 0) { - return 0; - } - - for (size_t i = 0; i < fullInArgs.size(); i++) { - inArgs_[i].resizeAndCopyFrom( - fullInArgs[i], - startSeq, - copySize, - FLAGS_parallel_nn ? false : multiMachine_->useGpu()); - } - return copySize; -} - -void TrainerThread::mergeCpuGradients() { - CHECK_EQ(mergeTypes_.size(), 1UL); - CHECK_EQ(mergeTypes_[0], PARAMETER_GRADIENT); - - { - REGISTER_TIMER("waitbeforeMerge"); - multiMachine_->waitBeforeMerge(); - } - std::vector*> slaveParameters = - multiMachine_->getSlaveParameters(); - - CHECK(slaveParameters.size()); - for (auto& para : multiMachine_->getNonStaticParameters()) { - if (para->useGpu()) continue; - if (para->isSparseRemoteUpdate()) { - REGISTER_TIMER("mergeRemoteGradSparse"); - mergeGradSparseRemote(para.get(), slaveParameters); - } else if (para->isGradSparseUpdate()) { - REGISTER_TIMER("mergeGradSparse"); - mergeGradSparse(para.get(), slaveParameters); - } else { - REGISTER_TIMER("mergeGradDense"); - mergeGradDense(para.get(), slaveParameters); - } - } - { - REGISTER_TIMER("waitbeforeMerge"); - multiMachine_->waitAfterMerge(); - } -} - -void TrainerThread::mergeGradSparse( - Parameter* para, - std::vector*>& slaveParameters) { - size_t pid = para->getID(); - SparseRowIdsCpuMatrix* mainMat = dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get()); - std::vector& ids = mainMat->getIds(threadId_); - - for (auto slaveParams : slaveParameters) { - SparseRowCpuMatrix* mat = dynamic_cast( - (*slaveParams)[pid]->getMat(PARAMETER_GRADIENT).get()); - mat->addTo(*mainMat, ids, threadId_, multiMachine_->getNumThreads()); - // we use a sample hash method(%) instead of range partition, - // because range partition has balance issue sometimes, - // when feature ids are not generated from hashcode. - } - uniqueIds(ids); -} - -void TrainerThread::mergeGradSparseRemote( - Parameter* para, - std::vector*>& slaveParameters) { - size_t pid = para->getID(); - SparseRowCpuMatrix* mainMat = - dynamic_cast(para->getMat(PARAMETER_GRADIENT).get()); - - mainMat->checkIndices(); - mainMat->zeroMemThread(threadId_, multiMachine_->getNumThreads()); - - for (auto slaveParams : slaveParameters) { - SparseRowCpuMatrix* mat = dynamic_cast( - (*slaveParams)[pid]->getMat(PARAMETER_GRADIENT).get()); - mat->addTo(*mainMat, threadId_, multiMachine_->getNumThreads()); - } -} - -void TrainerThread::mergeGradDense( - Parameter* para, - std::vector*>& slaveParameters) { - size_t pid = para->getID(); - auto interval = calcSplitArrayInterval(para->getSize(), - (size_t)threadId_, - multiMachine_->getNumThreads(), - 8LU /*for avx*/); - size_t startSeq = interval.first; - size_t copySize = interval.second - interval.first; - - // setup sub bufs - CpuVector destGrad(0, nullptr); - destGrad.subVecFrom(*para->getBuf(PARAMETER_GRADIENT), startSeq, copySize); - - // merge - CpuVector slaveGradSub(0, nullptr); - for (auto slaveParams : slaveParameters) { - slaveGradSub.subVecFrom( - *(*slaveParams)[pid]->getBuf(PARAMETER_GRADIENT), startSeq, copySize); - destGrad.add(slaveGradSub); - } -} - -void TrainerThread::copyOutputGrad() { - const std::vector& outputGradArgs = multiMachine_->outArgs_; - int numThreads = multiMachine_->getAllThreads().size(); - int32_t numSequences = outputGradArgs[0].getNumSequences(); - int32_t startSeq = numSequences * threadId_ / numThreads; - int32_t endSeq = numSequences * (threadId_ + 1) / numThreads; - int32_t copySize = endSeq - startSeq; - outArgs_.resize(outputGradArgs.size()); - for (size_t i = 0; i < outputGradArgs.size(); i++) { - outArgs_[i].resizeAndCopyFrom(outputGradArgs[i], - startSeq, - copySize, - multiMachine_->useGpu(), - HPPL_STREAM_DEFAULT); - } - if (multiMachine_->useGpu()) { - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - } - gradientMachine_->setOutputGrad(outArgs_); -} -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.h b/paddle/gserver/gradientmachines/MultiGradientMachine.h deleted file mode 100644 index eff7d5284c6dd4898344203b50acc94ae61b4d59..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/MultiGradientMachine.h +++ /dev/null @@ -1,478 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include - -#include "GradientMachine.h" - -#include "hl_gpu.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/Queue.h" - -namespace paddle { - -class TrainerThread; - -typedef Queue PidQueue; -typedef std::unique_ptr TrainerThreadPtr; - -struct GradBuffer { - /// GradBuffer is used for gathering gradient for GPU parameters - int paramId; - - /// sem is used to notify that the local gradient merge of the current thread - /// finished for the current thread. - Semaphore sem; - - // bufs[mergeIndex] - std::vector bufs; -}; - -/** - * A MultiGradientMachine is a synchronous GradientMachine which devides - * one data batch into several smaller batches and assign each one small batch - * to one computint thread for computation. After each thread finishes - * computation, it merges result (including output Argument and gradient during - * backward()). It basically is the same as single thread gradient machine, - * except that it uses multi-thread to do the computation. - * - * It handles GPU and Cpu parameters differently. In GPU, one computing thread - * generally corresponds to one GPU device. Thus, each thread keeps a separate - * copy of the parameter in its own device's memory. In CPU, we only need to - keep - * one copy of the parameters in the main memory. After, each computing thread - * computes its own parameter gradient, the update process needs to accumulate - * the parameter gradients from all the computing threads, and update the - * accumulated parameter gradient to the corresponding parameter value. - * - * Each GPU parameter is assigned to a thread called its main thread. For each - * parameter, the accumulation of its gradients and the update of its value - * happens in its main thread. The main thread first gather the parameter - * gradients from all the computing thread. Then, it performs parameter update. - * After a gradient is updated by the main thread, it is scattered to all the - * computing thread so that the parameters in all the computing threads are - * synchronized. The scatter and gather process are implemented by ring-style - * communication. Assume we have N computing threads, its thread ids will be - * 0, 1, ..., N-1. For each parameter, the id of the main thread is specified - in - * paraMainThread_[pid], where pid is the id of the parameter. Each thread i - only - * sends data to its partner thread (i - 1) % N. For example, for a parameter - * gradient that is computed in thread 4, and its main thread is 2. Its - * traveling process would be 4, 5,..., N-1, 0, 1, 2. In each step, the - gradient - * buffer is added to the local gradient, and the local gradient is then copied - * to the gradient buffer of the next thread. At last, its main thread 2 will - * get the accumulated parameter gradient. For the same parameter, after its - * value is updated, the value's traveling process would be 2, 1, 0, N-1, ... - 3. - * At the end, all the computing threads would have the updated parameter - value. - * - * A computing thread (TrainerThread) uses 4 threads to do different jobs: - * - * 1. computeThread(): performing forward(), backward(), prefetch(). - * - * 2. valueDispatchThread(): copying parameter values to partner thread. - * - * 3. copyGradToBufferThread(): copying parameter gradient to partner thread. - * - * 4. gradCollectThread(): merging the gradient from step 3 with local gradient - * and call the callback supplied by the user to update parameter value. - * - * CPU parameter value has only one copy. And their gradients are merged at the - * end of backward(). - * - * * Handling of sparse update - * Currently, sparse update is only supported for CPU parameters. - - * Sparse updates refers to gradient caculation where the gradient is sparse. - For - * example, if the input argument to a 'fc' layer is sparse, the gradient of - the - * weight matrix of this layer will be sparse. It is usually more efficient to - * treat the gradient explicitly as sparse vector during the parameter update. - - * There are two types of sparse updates called local sparse update and remote - * sparse update. - - * For both types of sparse updates, there is one copy of parameter value and - * gradient called main parameter value and gradient, and there is a copy of - * parameter value and gradient for each computing thread called slave - parameter - * value and gradient. The slave parameter values are always shared with the - * corresponding main parameter value. The slave parameter grad is a sparse row - * matrix. The sparse pattern for slave parameter grads are different, because - * the small batches for each computing thread might have different sparsity - * pattern. - - * 1. Local sparse update - * - * Main parameter value type is MAT_NORMAL. It is a dense matrix. - * - * Main parameter grad type is MAT_SPARSE_ROW_IDS (SparseRowIdsCpuMatrix) - * It is also a dense matrix, but the updated values are specified by IDS. - * - * Slave parameter value shares with main parameter value. - * - * Slave parameter grad type is MAT_SPARSE_ROW_AUTO_GROW - * (SparseAutoGrowRowCpuMatrix). It is a sparse row matrix. - * - * During backward() of each TrainerThread, SparseAutoGrowRowCpuMatrix will - * gather all the non-zero gradient. And After backward(), they will be - merged - * into main parameter grad (SparseRowIdsCpuMatrix), with indices indicating - * which rows have nonzero gradient. - * - * 2. Remote sparse update - * - * Main parameter value type is MAT_SPARSE_ROW_PREFETCH(_FULL_SIZE) - * (SparsePrefetchRowCpuMatrix). MAT_SPARSE_ROW_PREFETCH is a sparse matrix. - * MAT_SPARSE_ROW_PREFETCH_FULL_SIZE is a dense matrix. However, only the - * parameter values that are prefetched is up-to-date. - * - * Main parameter grad type is MAT_SPARSE_ROW (SparseRowCpuMatrix). - * And it shares sparse pattern with value by sharing indexDictHandle_, - which - * is an internal data structure used by SparseRowCpuMatrixto specify the - * sparsity pattern of Slave parameter value shares with main parameter - value. - * - * Slave parameter grad type is MAT_SPARSE_ROW_AUTO_GROW - * (SparsePrefetchRowCpuMatrix). It is a sparse row matrix - * - * During prefetch(), all the layers will indicates which rows of each - * parameter are needed. Then the framework will retrieve those rows from - * parameter server. - * - * During backward() of each TrainerThread, SparseAutoGrowRowCpuMatrix will - * gather all the non-zero gradient. And After backward(), they will be - merged - * into main parameter grad (SparseRowCpuMatrix). And the framework will - send - * the merged gradient to parameter server. - */ -class MultiGradientMachine : public GradientMachine { - public: - enum TaskType { - TASK_FORWARD_BACKWARD = 0, - TASK_FORWARD = 1, - TASK_BACKWARD = 2, - TASK_COPY_IN_ARGS = 3, - }; - - explicit MultiGradientMachine(const ModelConfig& config, bool useGpu); - - virtual void start(); - - virtual void finish(); - - virtual void prefetch(const std::vector& inArgs); - - virtual void forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType); - - virtual void backward(const UpdateCallback& callback = nullptr); - - void forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback); - - virtual Argument getLayerOutput(const std::string& layerName); - - virtual void onPassEnd(); - - virtual Evaluator* makeEvaluator() const; - - virtual void eval(Evaluator* evaluator) const; - - bool useGpu() const { return useGpu_; } - - /// @return whether to pass the gradients in outArgs_ to each threads. - bool isPassGrad() { return isPassGrad_; } - - /// @brief set whether to pass the gradient in outArgs_ to each threads. - void setPassGrad(bool isPass) { isPassGrad_ = isPass; } - - /// Set the gradients of the outputs. - /// The gradietns will be copied to each thread in the computing threads. - virtual void setOutputGrad(const std::vector& args); - - protected: - friend class TrainerThread; - - std::vector& getAllThreads() { return threads_; } - /// Calculate the real device id based on the logical device id and the - /// thread id. - int logicalDeviceId2RealDeviceId(int logicalId, int threadId = 0) const { - if (logicalId == -1) { - logicalId = 0; - } - return mod(logicalId + FLAGS_gpu_id + threadId * numLogicalDevices_, - numDevices_); - } - - /// Calculate the logical device id based on the real device id and the - /// thread id. - int realDeviceId2LogicalDeviceId(int realId, int threadId = 0) const { - if (realId == -1) { - return 0; - } else { - return mod(realId - FLAGS_gpu_id - threadId * numLogicalDevices_, - numDevices_); - } - } - - std::vector*> getSlaveParameters(); - - bool hasNonstaticCpuParamters() const { return hasNonstaticCpuParamters_; } - - /// Called TrainerThread to wait before merging CPU parameter gradients. - void waitBeforeMerge() { trainerBarrier_.wait(); } - - /// called by MultiGradientMachine and TrainerThread to wait after merging - /// CPU parameter graidents. - void waitAfterMerge() { allBarrier_.wait(); } - - /// called by MultiGradientMachine and TrainerThread to wait for copyInArgs() - /// finishing - void waitForCopyInArgs() { allBarrier_.wait(); } - - TrainerThreadPtr& getThread(int threadId) { return threads_[threadId]; } - - std::vector& getGradBuf(int threadId) { - return gradBufs_[threadId]; - } - - PassType getPassType() const { return passType_; } - - /// Called by TrainerThread to notify MultiGradientMachine that the gradient - /// for paramId is ready - void notifyGradientTransfer(int paramId); - - const std::vector& getInArgs() { return inArgs_; } - - TaskType getTaskType() const { return taskType_; } - - const UpdateCallback& getBackwardCallback() const { - return backwardCallback_; - } - - int getNumDevices() const { return numDevices_; } - - int getNumLogicalDevices() const { return numLogicalDevices_; } - - int getNumThreads() const { return numThreads_; } - - int paraMainThread(int pid) const { return paraMainThread_[pid]; } - - protected: - virtual void forwardImp(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - TaskType taskType); - - virtual void backwardImp(const UpdateCallback& callback = NULL); - - /// update all parameters - void updateThreadParameters(); - - void startTask(TaskType taskType); - - void getOutArgs(std::vector* outArgs, PassType passType); - - void allocGradBufs(); - - protected: - bool useGpu_; - - bool hasNonstaticCpuParamters_; - - /// store main parameter only - std::unique_ptr gradientMachine_; - - std::vector threads_; - std::vector paraMainThread_; - std::vector> gradBufs_; // [threadId][deviceId] - std::vector bufferSizes_; - - PassType passType_; - TaskType taskType_; - PidQueue gradQueue_; - std::vector inArgs_; - std::vector outArgs_; - hl_stream_t outArgStream_; - - Argument outLayerArgs_; - - /// ParameterType which needs to be merged from each GPU - std::vector mergeTypes_; - int numDevices_; /* number of gpu devices */ - int numLogicalDevices_; // number of GPU used by one NN - int numThreads_; /* number of train threads */ - - UpdateCallback backwardCallback_; - - /// barrrier for threads_ - ThreadBarrier trainerBarrier_; - - /// barrier for both MultiGradientMachine and threds_ - ThreadBarrier allBarrier_; - - /// indicate whether inArgs is copied before forward() - bool inArgsCopied_; - - /// Whether to copy the gradient back from an external input. - bool isPassGrad_; -}; - -class TrainerThread { - public: - TrainerThread(const ModelConfig& config, - int threadId, - MultiGradientMachine* multiMachine); - - ~TrainerThread(); - - void start(); - - void onPassEnd() { gradientMachine_->onPassEnd(); } - - void waitOutArgsReady() { outArgsReadySem_.wait(); } - - void notifyTaskReady() { taskReadySem_.post(); } - - int getDeviceId() const { return deviceId_; } - - GradientMachine* getGradientMachine() { return gradientMachine_.get(); } - - const std::vector& getParameters() { return parameters_; } - - void stop(); - - void notifyValueReady(int paramId); - - const VectorPtr& getValueBuf(int paramId) { - return parameters_[paramId]->getBuf(PARAMETER_VALUE); - } - - const std::vector& getOutArgs() { return outArgs_; } - - void incUpdateCounter(int n = 1) { - updateCounter_ += n; - parameterUpdated_ = true; - } - - void notifyGradientCollect(int paramId) { gradQueue_.enqueue(paramId); } - - void notifyCopyGradToBuffer(int paramId) { gradBufQueue_.enqueue(paramId); } - - void notifyValueDispatch(int paramId) { valueReadyQueue_.enqueue(paramId); } - - void prefetch(); - - /// copy the output gradient from the main GradientMachine. - void copyOutputGrad(); - - /// Whether the thread has input data. - bool hasInputData() { return batchSize_ != 0; } - - protected: - void mergeCpuGradients(); - - void mergeGradSparse( - Parameter* para, - std::vector*>& slaveParameters); - - void mergeGradSparseRemote( - Parameter* para, - std::vector*>& slaveParameters); - - void mergeGradDense( - Parameter* para, - std::vector*>& slaveParameters); - - void computeThread(); - void valueDispatchThread(); - void copyGradToBufferThread(); - void gradCollectThread(); - - int copyInArgs(); - void forward(); - void backward(); - void backwardCallback(Parameter* para); - - /// call the actuall callback supplied by the caller of - /// GradientMachine::backward - void doCallback(int pid); - - protected: - MultiGradientMachine* multiMachine_; - ModelConfig config_; - /// whether the thread should stop - bool stopping_; - /// the threads form which to collect gradient - int partnerId_; - /// from 0 to threads-1 - int threadId_; - int deviceId_; - std::unique_ptr gradientMachine_; - std::vector parameters_; - - /// ParameterType which needs to be merged from each GPU - std::vector mergeTypes_; - - /// compute thread - std::unique_ptr computeThread_; - std::vector inArgs_; - std::vector outArgs_; - Semaphore taskReadySem_; - Semaphore outArgsReadySem_; - - /// copy thread - std::unique_ptr copyThread_; - /// queue of gradient needs to be copied to partner - PidQueue gradBufQueue_; - hl_stream_t gradStream_; - - /// grad merge thread - std::unique_ptr gradCollectThread_; - /// queue of gradient needs to be merged with gradient coopied by - /// copyGradToBufferThread - PidQueue gradQueue_; - UpdateCallback backwardCallback_; - - /// value dispatch thread - std::unique_ptr valueDispatchThread_; - /// queue of the parameter whose the vale are ready for copy - PidQueue valueReadyQueue_; - - /// used to notify all the parameter values are ready - LockedCondition valueReadyCond_; - - hl_stream_t valueStream_; - /// how many parameters are updated - std::atomic updateCounter_; - bool parameterUpdated_; - - /// indicate whether inArgs is copied before forward() - bool inArgsCopied_; - int batchSize_; -}; - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/MultiNetwork.cpp b/paddle/gserver/gradientmachines/MultiNetwork.cpp deleted file mode 100644 index 5f3d09dda26772850828e6d44e8cc65635b314dc..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/MultiNetwork.cpp +++ /dev/null @@ -1,185 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -#include "MultiNetwork.h" - -#include "NeuralNetwork.h" -#include "ParallelNeuralNetwork.h" - -namespace paddle { - -void MultiNetwork::init(const ModelConfig& config, - ParamInitCallback callback, - const std::vector& parameterTypes, - bool useGpu) { - CHECK_GT(config.sub_models_size(), 1) << "sub_models_size should GT 1"; - // check submodel[0] is root - CHECK_EQ("root", config.sub_models(0).name()) - << "sub_models(0) should be root"; - // ignore root - subNetworks_.resize(config.sub_models_size() - 1); - // base class - NeuralNetwork::init(config, callback, parameterTypes, useGpu); - // sub networks - for (int i = 1; i < config.sub_models_size(); ++i) { - std::string subModelName = config.sub_models(i).name(); - if (FLAGS_parallel_nn) { - subNetworks_[i - 1] = std::unique_ptr( - new ParallelNeuralNetwork(subModelName, this)); - } else { - subNetworks_[i - 1] = std::unique_ptr( - NeuralNetwork::newNeuralNetwork(subModelName, this)); - } - subNetworks_[i - 1]->init(config); - } -} - -void MultiNetwork::prefetch(const std::vector& inArgs) { - std::vector> argumentGroups; - Argument::splitByDataId(inArgs, &argumentGroups); - // check group size is equal to sub network size - CHECK_EQ(argumentGroups.size(), subNetworks_.size()); - for (size_t i = 0; i < subNetworks_.size(); i++) { - if (argumentGroups[i].size() == 1 && argumentGroups[i][0].dataId == -1) { - // check input args: if dataId is -1, then skip this sub network - continue; - } - subNetworks_[i]->prefetch(argumentGroups[i]); - } -} - -void MultiNetwork::forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType) { - // split inArgs to several vectors - std::vector> argumentGroups; - Argument::splitByDataId(inArgs, &argumentGroups); - - // check group size is equal to sub network size - CHECK_EQ(argumentGroups.size(), subNetworks_.size()); - std::vector tempOutArgs; - outArgs->clear(); - - for (size_t i = 0; i < subNetworks_.size(); i++) { - tempOutArgs.clear(); - if (argumentGroups[i].size() == 1 && argumentGroups[i][0].dataId == -1) { - // check input args: if dataId is -1, then skip this sub network - continue; - } - subNetworks_[i]->forward(argumentGroups[i], &tempOutArgs, passType); - for (const auto& elem : tempOutArgs) { - outArgs->push_back(elem); - outArgs->back().dataId = i; - } - } -} - -void MultiNetwork::backward(const UpdateCallback& callback) { - for (size_t i = 0; i < subNetworks_.size(); i++) { - subNetworks_[i]->backward(callback); - } -} - -void MultiNetwork::forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback) { - forward(inArgs, outArgs, passType); - backward(callback); -} - -void MultiNetwork::onPassEnd() { - for (size_t i = 0; i < subNetworks_.size(); i++) { - subNetworks_[i]->onPassEnd(); - } -} - -void MultiNetwork::start() { - for (auto& subNetwork : subNetworks_) { - subNetwork->start(); - } -} - -void MultiNetwork::finish() { - for (size_t i = 0; i < subNetworks_.size(); i++) { - subNetworks_[i]->finish(); - } -} - -class MultiCombinedEvaluator : public Evaluator { - public: - MultiCombinedEvaluator() {} - void addEvaluator(std::unique_ptr&& evaluator) { - evaluators_.emplace_back(std::move(evaluator)); - } - virtual void start() { - for (auto& evaluator : evaluators_) { - evaluator->start(); - } - } - - virtual void finish() { - for (auto& evaluator : evaluators_) { - evaluator->finish(); - } - } - - virtual void eval(const NeuralNetwork& nn) { - const MultiNetwork& multiNetwork = dynamic_cast(nn); - CHECK_EQ(evaluators_.size(), multiNetwork.getSubNetworks().size()); - int size = evaluators_.size(); - for (int i = 0; i < size; i++) { - // one evaluator for one subNetwork - evaluators_[i]->eval(*multiNetwork.getSubNetworks()[i]); - } - } - - virtual real evalImp(std::vector& arguments) { - (void)arguments; - return -1; - } - - virtual void printStats(std::ostream& os) const { - for (auto& evaluator : evaluators_) { - evaluator->printStats(os); - os << ' '; - } - } - - virtual void distributeEval(ParameterClient2* client) { - for (auto& evaluator : evaluators_) { - evaluator->distributeEval(client); - } - } - - protected: - std::vector> evaluators_; -}; - -Evaluator* MultiNetwork::makeEvaluator() const { - MultiCombinedEvaluator* multiCombinedEvaluator = new MultiCombinedEvaluator(); - for (size_t i = 0; i < subNetworks_.size(); i++) { - std::unique_ptr evaluator(subNetworks_[i]->makeEvaluator()); - multiCombinedEvaluator->addEvaluator(std::move(evaluator)); - } - return multiCombinedEvaluator; -} - -void MultiNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); } - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/MultiNetwork.h b/paddle/gserver/gradientmachines/MultiNetwork.h deleted file mode 100644 index 495d5592017b5fb937fb8243bf12a5f2f30d67e7..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/MultiNetwork.h +++ /dev/null @@ -1,64 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "GradientMachine.h" -#include "NeuralNetwork.h" - -#include "paddle/utils/Locks.h" - -namespace paddle { - -class MultiNetwork : public NeuralNetwork { - public: - explicit MultiNetwork(std::string subModelName = "") - : NeuralNetwork(subModelName) {} - - virtual void init(const ModelConfig& config, - ParamInitCallback callback, - const std::vector& parameterTypes, - bool useGpu); - - virtual void prefetch(const std::vector& inArgs); - - virtual void forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType); - - virtual void backward(const UpdateCallback& callback = nullptr); - - void forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback); - - virtual void onPassEnd(); - - virtual Evaluator* makeEvaluator() const; - - virtual void eval(Evaluator* evaluator) const; - - const std::vector>& getSubNetworks() const { - return subNetworks_; - } - - virtual void start(); - - virtual void finish(); - - protected: - std::vector> subNetworks_; -}; -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp deleted file mode 100644 index ac60a3a3408d37b66cb712d893c6b93a1750f448..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp +++ /dev/null @@ -1,548 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Util.h" - -#include "NeuralNetwork.h" -#include "hl_gpu.h" -#include "paddle/utils/CustomStackTrace.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -#ifdef PADDLE_WITH_MKLDNN -#include "paddle/gserver/layers/MKLDNNLayer.h" -#endif - -#ifndef PADDLE_MOBILE_INFERENCE -#include "MultiNetwork.h" -#include "RecurrentGradientMachine.h" -#include "paddle/gserver/layers/AgentLayer.h" -#endif - -namespace paddle { -void parameterInitNN(int paramId, - Parameter* para, - std::vector* sharedParams) { - // Create parameters values. - if (!para->useGpu() && sharedParams) { - para->enableSharedType(PARAMETER_VALUE, - (*sharedParams)[paramId]->getBuf(PARAMETER_VALUE), - (*sharedParams)[paramId]->getMat(PARAMETER_VALUE)); - } else { - if (para->isSparseRemoteUpdate()) { - para->enableType(PARAMETER_VALUE, - FLAGS_loadsave_parameters_in_pserver - ? Parameter::MAT_SPARSE_ROW_PREFETCH - : Parameter::MAT_SPARSE_ROW_PREFETCH_FULL_SIZE); - } else { - para->enableType(PARAMETER_VALUE); - } - } - // Create parameter gradients. - if (para->isSparseRemoteUpdate() && !sharedParams) { - para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW); - } else if (para->isGradSparseUpdate()) { - para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW_AUTO_GROW); - } else if (!para->isStatic()) { - para->enableType(PARAMETER_GRADIENT); - } -} - -NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) { -#ifndef PADDLE_MOBILE_INFERENCE - if (config.type() == "recurrent_nn") { - return newNeuralNetwork("root"); - } else if (config.type() == "multi_nn") { - return new MultiNetwork("root"); - } else { - return newNeuralNetwork(); - } -#else - return new NeuralNetwork(); -#endif -} - -std::map NeuralNetwork::dllInitMap; - -void NeuralNetwork::init(const ModelConfig& config, - ParamInitCallback callback, - const std::vector& parameterTypes, - bool useGpu) { - using std::placeholders::_1; - using std::placeholders::_2; - ParamInitCallback paramCallback = nullptr; - if (callback != nullptr) { - paramSelfInited_ = false; - paramCallback = callback; - } else { - paramSelfInited_ = true; - paramCallback = std::bind(parameterInitNN, _1, _2, nullptr); - } - config_ = config; - - if (rootNetwork_ != nullptr) { - // direct use parameters_ and parameterMap_ from base network - CHECK_EQ((size_t)config.parameters_size(), - rootNetwork_->getParameters().size()); - parameters_ = rootNetwork_->getParameters(); - parameterMap_ = *(rootNetwork_->getParameterMap()); - } else { - parameters_.reserve(config.parameters_size()); - for (const auto& para_config : config.parameters()) { - auto parameter = std::make_shared(para_config, - useGpu, - /*initialize=*/false); - paramCallback(parameters_.size(), parameter.get()); - if (!callback) { - for (ParameterType type : - (parameter->isStatic() - ? std::vector{PARAMETER_VALUE} - : parameterTypes)) { - if (type != PARAMETER_VALUE && type != PARAMETER_GRADIENT) { - parameter->enableType(type); - } - } - } - parameter->setID(parameters_.size()); - parameters_.push_back(parameter); - CHECK(!parameterMap_.count(parameter->getName())); - parameterMap_[parameter->getName()] = parameter; - } - } - - auto layerCreate = [&](const LayerConfig& layer_config) { - auto layer = Layer::create(layer_config); - CHECK(layer) << "Create layer failed. Layer name:" << layer->getName(); - layers_.push_back(layer); - CHECK(!layerMap_.count(layer->getName())); - layerMap_[layer->getName()] = layer; - }; - - auto subModelConfig = std::find_if(config.sub_models().begin(), - config.sub_models().end(), - [=](const SubModelConfig& sub_model) { - return sub_model.name() == subModelName_; - }); - bool useSubModel = (subModelConfig != config.sub_models().end()); - CHECK_EQ(useSubModel, !subModelName_.empty()); - if (useSubModel) { - layers_.reserve(subModelConfig->layer_names_size()); - for (const auto& layer_name : subModelConfig->layer_names()) { - auto layer_config = - std::find_if(config.layers().begin(), - config.layers().end(), - [=](const LayerConfig& layer_config) { - return layer_config.name() == layer_name; - }); - CHECK(layer_config != config.layers().end()); - layerCreate(*layer_config); - } - } else { - layers_.reserve(config.layers_size()); - for (const auto& layer_config : config.layers()) { - bool useLayer = true; - if (config.has_external_config()) { - useLayer = true; - for (const auto& name : config.external_config().layer_names()) { - if (layer_config.name() == name) { - useLayer = false; - break; - } - } - } - if (useLayer) { - layerCreate(layer_config); - } - } - } - - for (const auto& layer : layers_) { - layer->init(layerMap_, parameterMap_); - layer->initSubNetwork(this /*root*/, config_, parameterTypes, useGpu); - } - - for (const auto& layer_name : - (useSubModel ? subModelConfig->input_layer_names() - : config.input_layer_names())) { - auto it = layerMap_.find(layer_name); - CHECK(it != layerMap_.end()); - dataLayers_.push_back(std::dynamic_pointer_cast(it->second)); - } - - for (const auto& layer_name : - (useSubModel ? subModelConfig->output_layer_names() - : config.output_layer_names())) { - auto it = layerMap_.find(layer_name); - CHECK(it != layerMap_.end()); - outputLayers_.push_back(it->second); - } - - for (const auto& layer : layers_) { - const auto& name = layer->getName(); - bool isMiddleLayer = true; - - // if data layer - for (const auto& dataLayer : dataLayers_) { - if (name == dataLayer->getName()) { - isMiddleLayer = false; - break; - } - } - - // if output layer - for (const auto& dataLayer : outputLayers_) { - if (name == dataLayer->getName()) { - isMiddleLayer = false; - break; - } - } - - if (isMiddleLayer) { - middleLayers_.push_back(layer); - } - } -} - -void NeuralNetwork::connect(LayerPtr agentLayer, - LayerPtr realLayer, - int height) { -#ifndef PADDLE_MOBILE_INFERENCE - AgentLayer* agent = dynamic_cast(agentLayer.get()); - CHECK_NOTNULL(agent); - agent->setRealLayer(realLayer, height); -#endif -} - -void NeuralNetwork::connect(std::string agentLayerName, - NeuralNetwork* srcNN, - std::string realLayerName) { - connect(this->getLayer(agentLayerName), srcNN->getLayer(realLayerName)); -} - -void NeuralNetwork::prefetch(const std::vector& inArgs) { - CHECK_EQ(inArgs.size(), dataLayers_.size()); - - if (paramSelfInited_) { - for (auto& para : parameters_) { - if (para->isSparseRemoteUpdate()) { - auto mat = dynamic_cast( - para->getMat(PARAMETER_VALUE).get()); - para->clearGradient(); - if (mat) mat->clearIndices(); - } - } - } - - for (size_t i = 0; i != dataLayers_.size(); ++i) { - if (FLAGS_parallel_nn) { - const_cast(inArgs[i]).deviceId = -1; - } - dataLayers_[i]->setData(inArgs[i]); - } - - for (auto& layer : layers_) { - layer->prefetch(); - } - - if (paramSelfInited_) { - for (auto& para : parameters_) { - if (para->isSparseRemoteUpdate()) { - auto mat = dynamic_cast( - para->getMat(PARAMETER_VALUE).get()); - mat->setupIndices(); - auto matGrad = dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get()); - matGrad->reserveStore(); - } - } - } -} - -void NeuralNetwork::forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType) { - CHECK_EQ(inArgs.size(), dataLayers_.size()); - outArgs->resize(outputLayers_.size()); - for (size_t i = 0; i != dataLayers_.size(); ++i) { - dataLayers_[i]->setData(inArgs[i]); - } - - gLayerStackTrace.set_stage(true); - - { - for (auto& layer : layers_) { - REGISTER_TIMER_INFO("ForwardTimer", layer->getName().c_str()); - gLayerStackTrace.push(layer->getName()); - layer->forward(passType); - gLayerStackTrace.pop(layer->getName()); - } - } - - outArgs->clear(); - outArgs->reserve(outputLayers_.size()); - for (auto& layer : outputLayers_) { - outArgs->push_back(layer->getOutput()); - } -} - -void NeuralNetwork::resetState() { - for (auto& layer : layers_) { - layer->resetState(); - } -} - -void NeuralNetwork::setState(const MachineState& machineState) { - for (size_t i = 0; i < layers_.size(); i++) { - if (machineState[i] != nullptr) { - layers_[i]->setState(machineState[i]); - } - } -} - -void NeuralNetwork::getState(MachineState& machineState) { - machineState.clear(); - machineState.reserve(layers_.size()); - for (auto& layer : layers_) { - LayerStatePtr p = layer->getState(); - machineState.push_back(p); - } -} - -void NeuralNetwork::backward(const UpdateCallback& callback) { - gLayerStackTrace.set_stage(false); - FOR_EACH_R(layer, layers_) { - REGISTER_TIMER_INFO("BackwardTimer", (*layer)->getName().c_str()); - gLayerStackTrace.push((*layer)->getName()); - if ((*layer)->needGradient()) { - (*layer)->backward(callback); - } - gLayerStackTrace.pop((*layer)->getName()); - } -} - -void NeuralNetwork::finish() { -#ifdef PADDLE_WITH_MKLDNN - FOR_EACH_R(layer, layers_) { - MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(*layer); - if (dnnLayer) { - dnnLayer->convertWeightsToPaddle(); - } - } -#endif -} - -Argument NeuralNetwork::getLayerOutput(const std::string& layerName) { - return getLayer(layerName)->getOutput(); -} - -void NeuralNetwork::onPassEnd() { - for (auto& layer : layers_) { - layer->onPassEnd(); - } -} - -void NeuralNetwork::releaseOutput() { - for (auto& layer : middleLayers_) { - Argument& arg = layer->getOutput(); - arg.value.reset(); - } -} - -#ifndef PADDLE_MOBILE_INFERENCE - -class CombinedEvaluator : public Evaluator { - public: - void addEvaluator(std::unique_ptr&& evaluator) { - evaluators_.emplace_back(std::move(evaluator)); - } - void start() override { - for (auto& evaluator : evaluators_) { - evaluator->start(); - } - } - - void finish() override { - for (auto& evaluator : evaluators_) { - evaluator->finish(); - } - } - - void eval(const NeuralNetwork& nn) override { - for (auto& evaluator : evaluators_) { - evaluator->eval(nn); - } - } - real evalImp(std::vector& arguments) override { - (void)arguments; - return -1; - } - void printStats(std::ostream& os) const override { - for (auto& evaluator : evaluators_) { - evaluator->printStats(os); - os << ' '; - } - } - - void distributeEval(ParameterClient2* client) override { - for (auto& evaluator : evaluators_) { - evaluator->distributeEval(client); - } - } - - protected: - std::vector> evaluators_; - - // Evaluator interface - public: - /** - * @brief getNames will return all inside evaluators' names. - * @param names [out]: return names. - */ - void getNames(std::vector* names) override { - for (auto& eval : evaluators_) { - eval->getNames(names); - } - } - - /** - * @brief getValue could get all inside evaluators' value. - */ - real getValue(const std::string& name, Error* err) const override { - return this->getMethodHelper( - name, err, [&name, err](const std::unique_ptr& eval) { - return eval->getValue(name, err); - }); - } - - /** - * @brief getType could get all inside evaluators' type. - */ - std::string getType(const std::string& name, Error* err) const override { - return this->getMethodHelper( - name, err, [&name, err](const std::unique_ptr& eval) { - return eval->getType(name, err); - }); - } - - private: - template - T getMethodHelper(const std::string& name, - Error* err, - const std::function&)>& - callback) const { - for (auto& eval : evaluators_) { - std::vector names; - eval->getNames(&names); - if (std::find(names.begin(), names.end(), name) != names.end()) { - return callback(eval); - } - } - *err = Error("No such key %s", name.c_str()); - return T(); - } -}; - -class SubnetEvaluator : public CombinedEvaluator { - public: - SubnetEvaluator(const std::string& layerName, - std::unique_ptr&& evaluator) - : layerName_(layerName) { - addEvaluator(std::move(evaluator)); - } - void eval(const NeuralNetwork& nn) override { - const LayerPtr& layer = nn.getLayer(layerName_); - CHECK(layer) << "Nonexisted layer: " << layerName_ << " in submodel " - << nn.getName(); - bool accessed = false; - layer->accessSubNetwork([this, &accessed](NeuralNetwork& subnet) { - subnet.eval(evaluators_[0].get()); - accessed = true; - }); - CHECK(accessed) << "There is no subnetwork for layer " << layerName_ - << " in submodel " << nn.getName(); - } - - protected: - std::string layerName_; -}; - -Evaluator* NeuralNetwork::makeEvaluator() const { - CombinedEvaluator* combinedEvaluator = new CombinedEvaluator(); - auto subModelConfig = std::find_if(config_.sub_models().begin(), - config_.sub_models().end(), - [=](const SubModelConfig& sub_model) { - return sub_model.name() == subModelName_; - }); - bool useSubModel = (subModelConfig != config_.sub_models().end()); - CHECK_EQ(useSubModel, !subModelName_.empty()); - if (useSubModel) { - // create the evaluators that belong to CURRENT submodel - for (int i = 0; i < subModelConfig->evaluator_names_size(); ++i) { - // find evaluator by name - auto thisEvalConfig = std::find_if( - config_.evaluators().begin(), - config_.evaluators().end(), - [=](const EvaluatorConfig& ecfg) { - return ecfg.name() == subModelConfig->evaluator_names(i); - }); - bool validConfig = (thisEvalConfig != config_.evaluators().end()); - if (validConfig) { - std::unique_ptr evaluator( - Evaluator::create(*thisEvalConfig)); - combinedEvaluator->addEvaluator(std::move(evaluator)); - } - } - for (auto& layer : layers_) { - layer->accessSubNetwork( - [layer, combinedEvaluator](NeuralNetwork& subnet) { - std::unique_ptr subEvaluator(new SubnetEvaluator( - layer->getName(), - std::unique_ptr(subnet.makeEvaluator()))); - combinedEvaluator->addEvaluator(std::move(subEvaluator)); - }); - } - } else { - for (const EvaluatorConfig& evalConfig : config_.evaluators()) { - std::unique_ptr evaluator(Evaluator::create(evalConfig)); - combinedEvaluator->addEvaluator(std::move(evaluator)); - } - } - return combinedEvaluator; -} - -void NeuralNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); } - -#endif - -void NeuralNetwork::setOutputGrad(const std::vector& args) { - CHECK_GE(outputLayers_.size(), args.size()); - for (size_t i = 0; i < args.size(); ++i) { - outputLayers_[i]->getOutput().grad = args[i].grad; - } -} - -extern NeuralNetwork* newCustomNerualNetwork(const std::string& name, - NeuralNetwork* network) - __attribute__((weak)); - -NeuralNetwork* NeuralNetwork::newNeuralNetwork(const std::string& name, - NeuralNetwork* rootNetwork) { - if (newCustomNerualNetwork) { - return newCustomNerualNetwork(name, rootNetwork); - } else { - return new NeuralNetwork(name, rootNetwork); - } -} - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.h b/paddle/gserver/gradientmachines/NeuralNetwork.h deleted file mode 100644 index 3e5615c8f0b30ab1283d41e025496051869289dc..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/NeuralNetwork.h +++ /dev/null @@ -1,179 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include - -#include "ModelConfig.pb.h" -#include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" -#include "paddle/gserver/layers/CostLayer.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/gserver/layers/Layer.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/utils/ClassRegistrar.h" - -namespace paddle { -/* - * @brief Init function for the parameters. - * @param paramId: the id of the parameter to init. - * @param para: the pointer to the parameter to init. - * @param sharedParams: the pointer to an array of the parameter to be shared. - * If it is null, no parameter sharing is used. - * Only CPU paramters can be shared. - * It handles CPU, CPU sparse, CPU sparse remote, - * and GPU parameters differently. If the type - * of a parameter is NORMAL. Basically nothing need to be done. - * CPU value: NORMAL. - * CPU param: NORMAL. - * - * CPU sparse value: NORMAL. - * CPU sparse gradient: MAT_SPARSE_ROW_AUTO_GROW. - * - * CPU sparse remote value: MAT_SPARSE_ROW_PREFETCH(_FULL_SIZE). - * CPU sparse remote gradient: MAT_SPARSE_ROW (!sharedParams) - * MAT_SPARSE_ROW_AUTO_GROW (sharedParams) - * - * GPU value: NORMAL - * GPU param: NORMAL - */ -void parameterInitNN(int paramId, - Parameter* para, - std::vector* sharedParams); - -class NeuralNetwork : public GradientMachine { - public: - virtual void init(const ModelConfig& config, - ParamInitCallback callback = nullptr, - const std::vector& parameterTypes = - std::vector{PARAMETER_VALUE, - PARAMETER_GRADIENT, - PARAMETER_MOMENTUM}, - bool useGpu = FLAGS_use_gpu); - - /** - * Connect two submodels and - * down-submodel's output become up-submodel's input. - * By default, connection is one by one, - * If the agent height is smaller than real layer, *height* has to be filled. - * - * @param realLayer The down-submodel's output layer. - * @param agentLayer The up-submodel's input agent layer. - */ - static void connect(LayerPtr agentLayer, LayerPtr realLayer, int height = 0); - void connect(std::string agentLayerName, - NeuralNetwork* srcNN, - std::string realLayerName); - - virtual void prefetch(const std::vector& inArgs); - - virtual void forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType); - - virtual void backward(const UpdateCallback& callback = nullptr); - - virtual Argument getLayerOutput(const std::string& layerName); - - const LayerPtr& getLayer(const std::string& layerName) const { - auto it = layerMap_.find(layerName); - CHECK(it != layerMap_.end()) << "Unknown layer " << layerName; - return it->second; - } - - virtual void onPassEnd(); - -#ifndef PADDLE_MOBILE_INFERENCE - virtual Evaluator* makeEvaluator() const; - - virtual void eval(Evaluator* evaluator) const; -#endif - - virtual void resetState(); - virtual void setOutputGrad(const std::vector& args); - - /// set machine state - virtual void setState(const MachineState& machineState); - - /// get machine state - virtual void getState(MachineState& machineState); - - static NeuralNetwork* create(const ModelConfig& config); - - ParameterMap* getParameterMap() { return ¶meterMap_; } - - /** - * @brief Access each layer as a for each loop. - * @param callback invoke with each layer. - */ - template - void forEachLayer(T callback) { - for (auto& l : layers_) { - if (callback(l)) { - break; - } - } - } - - static NeuralNetwork* newNeuralNetwork(const std::string& name = "", - NeuralNetwork* rootNetwork = nullptr); - - const std::string& getName() const { return subModelName_; } - - /// some finish work, like convert the weight format of MKLDNNLayers - void finish(); - - /** - * @brief Release the middle layer's output memory. - * - * @note This function is used for memory optimization in inference. - */ - void releaseOutput(); - - protected: - /** - * The constructor of NeuralNetwork. - * The sub networks can get parameters_ and parameterMap_ - * from base NeuralNetwork. - * - * @param subModelName The name of sub-model. - * @param rootNetwork It used in MultiNetwork. - */ - NeuralNetwork(std::string subModelName = "", - NeuralNetwork* rootNetwork = nullptr) - : subModelName_(subModelName), rootNetwork_(rootNetwork) {} - - std::string subModelName_; - ModelConfig config_; - std::vector layers_; - ParameterMap parameterMap_; - LayerMap layerMap_; - - std::vector dataLayers_; - std::vector outputLayers_; - std::vector middleLayers_; - - static std::map dllInitMap; - - NeuralNetwork* rootNetwork_; - - /// Whether parameter of this NN is initialized by its own - /// (i.e., not by callback supplied with the caller) - bool paramSelfInited_; -}; - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp deleted file mode 100644 index 85cfc59fbef7017f8dea7fdfecd18aa3e75a871c..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.cpp +++ /dev/null @@ -1,213 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -#include "ParallelNeuralNetwork.h" - -#include -#include - -namespace paddle { - -void ParallelNeuralNetwork::init( - const ModelConfig& config, - ParamInitCallback callback, - const std::vector& parameterTypes, - bool useGpu) { - NeuralNetwork::init(config, callback, parameterTypes, useGpu); - - if (config.type() == "recurrent_nn") { - LOG(FATAL) - << "You can not add `--parallel_nn=true` on the command line, " - << "parallel_nn training mode does not support the recurrent_nn model."; - } - - useGpu_ = useGpu; - numDevices_ = 0; - if (useGpu_) { - numDevices_ = hl_get_device_count(); - } - - for (auto& layer : layers_) { - int deviceId = layer->getDeviceId(); - CHECK_LT(deviceId, numDevices_); - addComputeThread(deviceId); - } -} - -void ParallelNeuralNetwork::addComputeThread(int deviceId) { - for (auto& thread : threads_) { - if (thread->getDeviceId() == deviceId) { - return; - } - } - - threads_.emplace_back(new ParallelThread( - threads_.size(), deviceId, deviceId >= 0 ? useGpu_ : false)); -} - -void ParallelNeuralNetwork::waitAllThread() { - for (auto& thread : threads_) { - thread->jobEnqueue(NULL, TASK_END_LAYER); - } - - for (size_t i = 0; i < threads_.size(); i++) { - threads_[i]->queue_.waitEmpty(); - } -} - -void ParallelNeuralNetwork::dispatchByDeviceId(int deviceId, - LayerPtr layer, - TaskType task) { - for (auto& thread : threads_) { - if (thread->getDeviceId() == deviceId) { - thread->jobEnqueue(layer, task); - return; - } - } - LOG(FATAL) << "No specific device thread "; -} - -void ParallelNeuralNetwork::forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType) { - for (auto& thread : threads_) { - thread->setForwardPassType(passType); - } - CHECK_EQ(inArgs.size(), dataLayers_.size()); - outArgs->resize(outputLayers_.size()); - for (size_t i = 0; i != dataLayers_.size(); ++i) { - const_cast(inArgs[i]).deviceId = -1; - dataLayers_[i]->setData(inArgs[i]); - } - - for (auto& layer : layers_) { - dispatchByDeviceId(layer->getDeviceId(), layer, TASK_FORWARD); - } - - { - REGISTER_TIMER("forwardTime"); - waitAllThread(); - } - outArgs->clear(); - outArgs->reserve(outputLayers_.size()); - for (auto& layer : outputLayers_) { - outArgs->push_back(layer->getOutput()); - } -} - -void ParallelNeuralNetwork::backward(const UpdateCallback& callback) { - for (auto& thread : threads_) { - thread->setBackwardCallback(callback); - } - - FOR_EACH_R(layer, layers_) { - dispatchByDeviceId((*layer)->getDeviceId(), *layer, TASK_BACKWARD); - } - { - REGISTER_TIMER("backwardTime"); - waitAllThread(); - } -} - -void ParallelNeuralNetwork::forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback) { - forward(inArgs, outArgs, passType); - backward(callback); -} - -void ParallelNeuralNetwork::start() { - for (auto& thread : threads_) { - thread->start(); - } -} - -ParallelThread::ParallelThread(int threadId, int deviceId, bool useGpu) - : threadId_(threadId), deviceId_(deviceId), useGpu_(useGpu) {} - -ParallelThread::~ParallelThread() { stop(); } - -void ParallelThread::stop() { - if (computeThread_) { - jobEnqueue(NULL, TASK_THREAD_FINISH); - computeThread_->join(); - computeThread_.reset(nullptr); - } -} - -void ParallelThread::computeThread() { - LOG(INFO) << "gradComputeThread " << threadId_; - - if (useGpu_) { - hl_init(deviceId_); - } - - while (true) { - struct Job job_work = queue_.dequeue(); - - if (job_work.task_ == TASK_END_LAYER) { - continue; - } else if (job_work.task_ == TASK_THREAD_FINISH) { - break; - } - - if (TASK_FORWARD == job_work.task_) { - { - REGISTER_TIMER_INFO("waitInputValue", - job_work.layer_->getName().c_str()); - job_work.layer_->waitInputValue(); - } - { - REGISTER_TIMER_INFO("threadForwardTimer", - job_work.layer_->getName().c_str()); - job_work.layer_->forward(passType_); - } - { - REGISTER_TIMER_INFO("copyOutputToOtherDevice", - job_work.layer_->getName().c_str()); - job_work.layer_->copyOutputToOtherDevice(); - } - } else { - { - REGISTER_TIMER_INFO("waitAndMergeOutputGrad", - job_work.layer_->getName().c_str()); - job_work.layer_->waitAndMergeOutputGrad(); - } - { - REGISTER_TIMER_INFO("threadBackwardTimer", - job_work.layer_->getName().c_str()); - job_work.layer_->backward(backwardCallback_); - } - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - job_work.layer_->markAllInputGrad(); - } - } -} - -void ParallelThread::start() { - computeThread_.reset(new std::thread([this]() { computeThread(); })); -} - -void ParallelThread::jobEnqueue(LayerPtr layer, TaskType task) { - struct Job job_work; - job_work.layer_ = layer; - job_work.task_ = task; - queue_.enqueue(job_work); -} - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp deleted file mode 100644 index 73ac8cda721f200c1a02cd9c1d9456df70d7b7d2..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp +++ /dev/null @@ -1,1501 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "RecurrentGradientMachine.h" -#include -#include -#include -#include -#include -#include "NeuralNetwork.h" -#include "paddle/gserver/layers/AgentLayer.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -DEFINE_string(diy_beam_search_prob_so, "", "the diy beam search cost so"); - -static const char* DIY_CALC_PROB_SYMBOL_NAME = "calc_prob"; -static const char* DIY_START_CALC_PROB_SYMBOL_NAME = "start_calc_prob"; -static const char* DIY_FINISH_CALC_PROB_SYMBOL_NAME = "finish_calc_prob"; - -namespace paddle { - -/** - * Start Custom Calculate Probability callback type. - * - * @param nNode, nodes: the path will be explored. nNodes is array size. - * nodes is array elements. - * - * @return: A custom handler id that will passed to another callback. - */ -typedef int (*DiyStartCalcProbCallback)(size_t nNodes, int* nodes); - -/** - * Doing Custom Calculation of Probability callback type. - * - * @param handler: User custom handler. The return value from start calc prob. - * @param nNode, nodes: Array. The current path. - * @param curProb: The current log probability that neural network returns. - * - * @return: Log probability which user calculated, it will be updated to this - * path. - * @NOTE: Return -INFINITY will DROP this path IMMEDIATELY!! - */ -typedef real (*DiyCalcProbCallback)( - int handler, size_t nNodes, int* nodes, real curProb, bool atEos); - -/** - * Finish Custom Calculation of Probability callback type. - * - * @param handler: User custom handler. The return value from start calc prob. - */ -typedef void (*DiyStopCalcProbCallback)(int handler); - -static DiyCalcProbCallback gDiyProbMethod = nullptr; -static DiyStartCalcProbCallback gDiyProbStart = nullptr; -static DiyStopCalcProbCallback gDiyProbStop = nullptr; -static void* gDiyProbHandle = nullptr; - -static void exit_diy_prob() { dlclose(gDiyProbHandle); } - -template -static inline SymbolType loadDiySymbol(const char* symbolName) { - void* sym = dlsym(gDiyProbHandle, symbolName); - CHECK(sym) << "Cannot load symbol " << symbolName << " from " - << FLAGS_diy_beam_search_prob_so; - return reinterpret_cast(sym); -} - -static InitFunction __init__diy_prob_method( - [] { - std::string soName = FLAGS_diy_beam_search_prob_so; - if (!soName.empty()) { - gDiyProbHandle = dlopen(soName.c_str(), RTLD_LAZY); - CHECK(gDiyProbHandle) << "Cannot Open DIY Prob So " << soName; - atexit(exit_diy_prob); - gDiyProbMethod = - loadDiySymbol(DIY_CALC_PROB_SYMBOL_NAME); - gDiyProbStart = loadDiySymbol( - DIY_START_CALC_PROB_SYMBOL_NAME); - gDiyProbStop = loadDiySymbol( - DIY_FINISH_CALC_PROB_SYMBOL_NAME); - } - }, - std::numeric_limits::max()); - -class BeamSearchControlCallbacks { - public: - RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback - beamSearchCandidateAdjust; - RecurrentGradientMachine::NormOrDropNodeCallback normOrDropNode; - RecurrentGradientMachine::DropCallback stopDetermineCandidates; - - //! for gcc46 aggregate initialization is not very well, so we need to - //! explicit - BeamSearchControlCallbacks( - const RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback& - candidateAdjust, - const RecurrentGradientMachine::NormOrDropNodeCallback& norm, - const RecurrentGradientMachine::DropCallback& stop) - : beamSearchCandidateAdjust(candidateAdjust), - normOrDropNode(norm), - stopDetermineCandidates(stop) {} -}; - -class BeamSearchStatisticsCallbacks { - public: - RecurrentGradientMachine::EachStepCallback onEachStepStarted; - RecurrentGradientMachine::EachStepCallback onEachStepStoped; - - BeamSearchStatisticsCallbacks( - const RecurrentGradientMachine::EachStepCallback& start, - const RecurrentGradientMachine::EachStepCallback& stop) - : onEachStepStarted(start), onEachStepStoped(stop) {} -}; - -RecurrentGradientMachine::RecurrentGradientMachine( - const std::string& subModelName, NeuralNetwork* rootNetwork) - : NeuralNetwork(subModelName), - rootNetwork_(rootNetwork), - beamSearchCtrlCallbacks_(nullptr), - beamSearchStatistics_(nullptr) { - CHECK(!subModelName_.empty()); -} - -/** - * bias layer, as input of memory frame 0 will give vector of zeros - * if bias parameter is not set. - * - * boot bias layer create directly in recurrent gradient machine, because: - * - * 1. It is only one frame, so it should not be placed in layer group, - * which is one instance for every one frame. - * - * 2. It is no input layer, so it need resetHeight() before forward(), - * and resetHeight() must be called in recurrent gradient machine, - * so it's should not be placed in root network. - */ -class BootBiasLayer : public Layer { - protected: - std::unique_ptr biases_; - IVectorPtr cpuIds_; - - public: - explicit BootBiasLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override { - if (!Layer::init(layerMap, parameterMap)) return false; - - if (biasParameter_) { - biases_ = - std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - return true; - } - - void resetHeight(int height) { - if (config_.has_bos_id()) { // used as a constant id layerConfig - IVector::resizeOrCreate(output_.ids, height, useGpu_); - output_.ids->reset((int)config_.bos_id()); - } else { - resetOutput(height, getSize()); - } - } - - void forward(PassType passType) override { - if (biases_) { - MatrixPtr outV = getOutputValue(); - outV->addBias(*(biases_->getW()), 1); - forwardActivation(); - } - } - - void backward(const UpdateCallback& callback) override { - if (biases_ && biases_->getWGrad()) { - backwardActivation(); - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - biases_->getParameterPtr()->incUpdate(callback); - } - } -}; - -void RecurrentGradientMachine::init( - const ModelConfig& config, - ParamInitCallback callback, - const std::vector& parameterTypes, - bool useGpu) { - NeuralNetwork::init(config, callback, parameterTypes, useGpu); - useGpu_ = useGpu; - - auto subModelConfig = - std::find_if(config.sub_models().begin(), - config.sub_models().end(), - [this](const SubModelConfig& sub_model) { - return sub_model.name() == this->subModelName_; - }); - CHECK(subModelConfig != config.sub_models().end()); - reversed_ = subModelConfig->reversed(); - generating_ = subModelConfig->has_generator(); - - inFrameLines_.resize(subModelConfig->in_links_size()); - for (size_t i = 0; i < inFrameLines_.size(); ++i) { - inFrameLines_[i].linkName = subModelConfig->in_links(i).link_name(); - inFrameLines_[i].inLayer = - rootNetwork_->getLayer(subModelConfig->in_links(i).layer_name()); - } - - outFrameLines_.resize(subModelConfig->out_links_size()); - for (size_t i = 0; i < outFrameLines_.size(); ++i) { - auto& linkPair = subModelConfig->out_links(i); - outFrameLines_[i].layerName = linkPair.layer_name(); - outFrameLines_[i].agentLayer = rootNetwork_->getLayer(linkPair.link_name()); - } - - memoryFrameLines_.resize(subModelConfig->memories_size()); - for (size_t i = 0; i < memoryFrameLines_.size(); ++i) { - auto& memoryConfig = subModelConfig->memories(i); - memoryFrameLines_[i].layerName = memoryConfig.layer_name(); - memoryFrameLines_[i].linkName = memoryConfig.link_name(); - auto agentConfig = - std::find_if(config.layers().begin(), - config.layers().end(), - [&memoryConfig](const LayerConfig& layerConfig) { - return layerConfig.name() == memoryConfig.link_name(); - }); - CHECK(agentConfig != config.layers().end()); - if (memoryConfig.has_boot_layer_name()) { - memoryFrameLines_[i].rootLayer = - rootNetwork_->getLayer(memoryConfig.boot_layer_name()); - - LayerConfig scatterConfig = *agentConfig; - memoryFrameLines_[i].rootAgent.reset( - new ScatterAgentLayer(scatterConfig)); - memoryFrameLines_[i].rootAgent->init(LayerMap(), parameterMap_); - - memoryFrameLines_[i].bootLayer = memoryFrameLines_[i].rootAgent; - } else { - LayerConfig biasConfig = *agentConfig; - if (memoryConfig.has_boot_bias_parameter_name()) { - biasConfig.set_bias_parameter_name( - memoryConfig.boot_bias_parameter_name()); - biasConfig.set_active_type(memoryConfig.boot_bias_active_type()); - } else if (memoryConfig.has_boot_with_const_id()) { - biasConfig.set_bos_id(memoryConfig.boot_with_const_id()); - } - memoryFrameLines_[i].biasLayer.reset(new BootBiasLayer(biasConfig)); - memoryFrameLines_[i].biasLayer->init(LayerMap(), parameterMap_); - - memoryFrameLines_[i].bootLayer = memoryFrameLines_[i].biasLayer; - } - - if (subModelConfig->has_generator()) { - memoryFrameLines_[i].scatterAgents.resize(2); - for (auto& agent : memoryFrameLines_[i].scatterAgents) { - agent.reset(new ScatterAgentLayer(*agentConfig)); - agent->init(LayerMap(), parameterMap_); - } - } - } - - if (subModelConfig->has_generator()) { - generator_.config = subModelConfig->generator(); - eosFrameLine_.reset(new EosFrameLine); - maxSequenceLength_ = generator_.config.max_num_frames(); - } - - // get parameters actually used by this Layer Group - resizeOrCreateFrames(1); - for (auto& para : frames_[0]->getParameters()) { - if (para->getSharedCount() > 0) { - parameterIds_.push_back(para->getID()); - } - } - for (auto& para : parameters_) { // bias layer parameters - if (para->getSharedCount() > 0) { - parameterIds_.push_back(para->getID()); - } - } -} - -void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) { - if ((size_t)numFrames <= frames_.size()) { - return; - } - - frames_.reserve(numFrames); - for (auto& inFrameLine : inFrameLines_) { - inFrameLine.agents.reserve(numFrames); - } - for (auto& outFrameLine : outFrameLines_) { - outFrameLine.frames.reserve(numFrames); - } - for (auto& memoryFrameLine : memoryFrameLines_) { - memoryFrameLine.frames.reserve(numFrames); - memoryFrameLine.agents.reserve(numFrames); - } - if (eosFrameLine_) { - eosFrameLine_->layers.reserve(numFrames); - } - - ParamInitCallback subParamInitCb = [this](int paramId, Parameter* para) { - para->enableSharedType(PARAMETER_VALUE, - this->parameters_[paramId]->getBuf(PARAMETER_VALUE), - this->parameters_[paramId]->getMat(PARAMETER_VALUE)); - para->enableSharedType( - PARAMETER_GRADIENT, - this->parameters_[paramId]->getBuf(PARAMETER_GRADIENT), - this->parameters_[paramId]->getMat(PARAMETER_GRADIENT)); - }; - - for (int i = frames_.size(); i < numFrames; ++i) { - std::unique_ptr frame( - NeuralNetwork::newNeuralNetwork(subModelName_)); - frame->init(config_, subParamInitCb); - - for (auto& inFrameLine : inFrameLines_) { - inFrameLine.agents.push_back(frame->getLayer(inFrameLine.linkName)); - } - - for (auto& outFrameLine : outFrameLines_) { - outFrameLine.frames.push_back(frame->getLayer(outFrameLine.layerName)); - } - for (auto& memoryFrameLine : memoryFrameLines_) { - memoryFrameLine.frames.push_back( - frame->getLayer(memoryFrameLine.layerName)); - memoryFrameLine.agents.push_back( - frame->getLayer(memoryFrameLine.linkName)); - } - if (eosFrameLine_) { - eosFrameLine_->layers.push_back( - frame->getLayer(generator_.config.eos_layer_name())); - } - - frames_.emplace_back(std::move(frame)); - } -} - -void RecurrentGradientMachine::resizeBootFrame(int numSequences) { - for (auto& memoryFrameLine : memoryFrameLines_) { - if (memoryFrameLine.biasLayer) { - auto biasLayer = - dynamic_cast(memoryFrameLine.biasLayer.get()); - CHECK_NOTNULL(biasLayer); - biasLayer->resetHeight(numSequences); - } else { // check input root layer height - CHECK_EQ(numSequences, - memoryFrameLine.rootLayer->getOutput().getNumSequences()); - } - } -} - -void RecurrentGradientMachine::prefetch(const std::vector& inArgs) { - LOG(FATAL) << "should not use this function"; -} - -void RecurrentGradientMachine::checkInputConsistency( - int inlinkId, const std::vector& seqInfo) { - if (commonSeqInfo_.empty()) { - commonSeqInfo_.resize(seqInfo.size()); - for (size_t i = 0; i < seqInfo.size(); ++i) { - commonSeqInfo_[i].topLevelLength = seqInfo[i].topLevelLength; - commonSeqInfo_[i].seqId = seqInfo[i].seqId; - } - } else { - CHECK_EQ(commonSeqInfo_.size(), seqInfo.size()) - << " RecurrentGroup " << subModelName_ << " input " << inlinkId - << " has mismatched number of sequences"; - for (size_t i = 0; i < seqInfo.size(); ++i) { - CHECK_EQ(commonSeqInfo_[i].topLevelLength, seqInfo[i].topLevelLength) - << " RecurrentGroup " << subModelName_ << " input " << inlinkId - << " has mismatched sequence length"; - CHECK_EQ(commonSeqInfo_[i].seqId, seqInfo[i].seqId) - << " RecurrentGroup " << subModelName_ << " input " << inlinkId - << " has mismatched sequence length"; - } - } -} - -void RecurrentGradientMachine::calcNumSequencesAtEachStep() { - int numSequences = commonSeqInfo_.size(); - numSeqs_.resize(maxSequenceLength_); - for (int i = 0; i < numSequences; ++i) { - for (int j = 0; j < commonSeqInfo_[i].topLevelLength; ++j) { - numSeqs_[j] = i + 1; - } - } -} - -void RecurrentGradientMachine::reorganizeInput(PassType passType) { - info_.clear(); - info_.resize(inFrameLines_.size()); - - commonSeqInfo_.clear(); - seqInfos_.clear(); - seqInfos_.resize(inFrameLines_.size()); - - for (size_t i = 0; i < inFrameLines_.size(); i++) { - const Argument& input = inFrameLines_[i].inLayer->getOutput(); - if (!input.hasSeq()) { - continue; - } - input.getSeqInfo(&seqInfos_[i]); - checkInputConsistency(i, seqInfos_[i]); - } - CHECK(!commonSeqInfo_.empty()) - << "At least one input needs to be sequence or subsequence"; - maxSequenceLength_ = commonSeqInfo_[0].topLevelLength; - - calcNumSequencesAtEachStep(); - - for (size_t i = 0; i < inFrameLines_.size(); ++i) { - const Argument& input = inFrameLines_[i].inLayer->getOutput(); - if (!input.hasSeq()) { - seqInfos_[i] = commonSeqInfo_; - } - createInFrameInfo(i, input, passType); - } - - { - AsyncGpuBlock asyncGpuBlock; - - // inFrameLine select rows in real layer one time - for (size_t i = 0; i < inFrameLines_.size(); i++) { - selectRowsOneTime(inFrameLines_[i].inLayer, - info_[i].allIds, - &(inFrameLines_[i].outArg), - passType); - } - } -} - -void RecurrentGradientMachine::reorganizeOutput(PassType passType) { - calcSequenceStartPositions(); - for (size_t i = 0; i < outFrameLines_.size(); ++i) { - Info info; - auto& outFrameLine = outFrameLines_[i]; - ICpuGpuVectorPtr sequenceStartPositions; - ICpuGpuVectorPtr subSequenceStartPositions; - createOutFrameInfo( - outFrameLine, info, sequenceStartPositions, subSequenceStartPositions); - auto gatherAgent = - dynamic_cast(outFrameLine.agentLayer.get()); - CHECK_NOTNULL(gatherAgent); - gatherAgent->copyIdAndSequenceInfo(sequenceStartPositions, - subSequenceStartPositions, - info.allIds, - info.idIndex); - } -} - -void RecurrentGradientMachine::connectFrames(PassType passType) { - for (auto& memoryFrameLine : memoryFrameLines_) { - if (memoryFrameLine.rootAgent) { - auto scatterAgent = - dynamic_cast(memoryFrameLine.rootAgent.get()); - createMemoryFrameInfo(&memoryFrameLine, passType); - scatterAgent->setRealLayerAndOutput(memoryFrameLine.rootLayer, - memoryFrameLine.outArg, - memoryFrameLine.allIds, - /* idIndex */ 0, - memoryFrameLine.allIds->getSize(), - /* handleBackward */ true); - if (memoryFrameLine.sequenceStartPositions) { - int size = memoryFrameLine.sequenceStartPositions->getSize(); - scatterAgent->setSequenceStartPositions( - memoryFrameLine.sequenceStartPositions, - /* seqStartPosIndex */ 0, - size); - } - } - } - - for (auto& outFrameLine : outFrameLines_) { - auto gatherAgent = - dynamic_cast(outFrameLine.agentLayer.get()); - gatherAgent->clearRealLayers(); - } - for (int i = 0; i < maxSequenceLength_; ++i) { - // connect in_links - for (size_t j = 0; j < inFrameLines_.size(); ++j) { - Info& info = info_[j]; - // idSize denotes the sum number of tokens in each length i - int idIndex = info.idIndex.empty() ? 0 : info.idIndex[i]; - int idSize = info.idIndex.empty() ? numSeqs_[i] - : info.idIndex[i + 1] - info.idIndex[i]; - InFrameLine inFrameLine = inFrameLines_[j]; - auto scatterAgent = - dynamic_cast(inFrameLine.agents[i].get()); - scatterAgent->setRealLayerAndOutput(inFrameLine.inLayer, - inFrameLine.outArg, - info.allIds, - idIndex, - idSize, - i == 0); - if (info.sequenceStartPositions) { - // size: the length of subsequence - int size = info.seqStartPosIndex[i + 1] - info.seqStartPosIndex[i]; - scatterAgent->setSequenceStartPositions( - info.sequenceStartPositions, info.seqStartPosIndex[i], size); - } - } - - // connect out_links - for (auto& outFrameLine : outFrameLines_) { - auto gatherAgent = - dynamic_cast(outFrameLine.agentLayer.get()); - gatherAgent->addRealLayer(outFrameLine.frames[i]); - } - for (auto& memoryFrameLine : memoryFrameLines_) { - NeuralNetwork::connect( - memoryFrameLine.agents[i], - i == 0 ? memoryFrameLine.bootLayer : memoryFrameLine.frames[i - 1], - numSeqs_[i] /*height of agent*/); - } - } -} - -void RecurrentGradientMachine::forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType) { - /* inArgs and outArgs are not used. - The inputs are inFrameLines_[i].inLayer. - The outputs are outFramesLines_[i].agentLayer - */ - - if (generating_) { - generateSequence(); - return; - } // else forward.. - - reorganizeInput(passType); - int numSequences = commonSeqInfo_.size(); - - resizeOrCreateFrames(maxSequenceLength_); - resizeBootFrame(numSequences); - - connectFrames(passType); - - REGISTER_TIMER_INFO("RecurrentFwTime", "RecurrentFwTime"); - // forward - for (auto& memoryFrameLine : memoryFrameLines_) { - memoryFrameLine.bootLayer->forward(passType); - } - for (int i = 0; i < maxSequenceLength_; ++i) { - const std::vector inArgs; - std::vector outArgs; - frames_[i]->forward(inArgs, &outArgs, passType); - } - - reorganizeOutput(passType); -} - -void RecurrentGradientMachine::backward(const UpdateCallback& callback) { - if (generating_) { - return; - } - REGISTER_TIMER_INFO("RecurrentBwTime", "RecurrentBwTime"); - AsyncGpuBlock asyncGpuBlock; - for (int i = maxSequenceLength_ - 1; i >= 0; --i) { - frames_[i]->backward(nullptr); - } - for (auto& memoryFrameLine : memoryFrameLines_) { - memoryFrameLine.bootLayer->backward(nullptr); - } -} - -void RecurrentGradientMachine::forwardBackward( - const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback) { - LOG(FATAL) << "should not use this function"; -} - -void RecurrentGradientMachine::eval(Evaluator* evaluator) const { - // call printers frame by frame - for (int i = 0; i < maxSequenceLength_; ++i) { - VLOG(2) << "Recurrent Layer Group eval frame " << i << " begin"; - evaluator->eval(*(frames_[i].get())); - VLOG(2) << "Recurrent Layer Group eval frame " << i << " end"; - } -} - -void RecurrentGradientMachine::registerBeamSearchControlCallbacks( - const BeamSearchCandidatesAdjustCallback& adjustBeamSearch, - const NormOrDropNodeCallback& normOrDropNode, - const DropCallback& stopBeamSearch) { - this->removeBeamSearchControlCallbacks(); - //! for gcc 46, aggregate initialization is not supported. TAT - this->beamSearchCtrlCallbacks_ = new BeamSearchControlCallbacks( - adjustBeamSearch, normOrDropNode, stopBeamSearch); -} - -void RecurrentGradientMachine::removeBeamSearchControlCallbacks() { - if (this->beamSearchCtrlCallbacks_) { - delete this->beamSearchCtrlCallbacks_; - this->beamSearchCtrlCallbacks_ = nullptr; - } -} - -void RecurrentGradientMachine::registerBeamSearchStatisticsCallbacks( - const EachStepCallback& onEachStepStarted, - const EachStepCallback& onEachStepStoped) { - this->removeBeamSearchStatisticsCallbacks(); - this->beamSearchStatistics_ = - new BeamSearchStatisticsCallbacks(onEachStepStarted, onEachStepStoped); -} - -void RecurrentGradientMachine::removeBeamSearchStatisticsCallbacks() { - if (this->beamSearchStatistics_) { - delete this->beamSearchStatistics_; - this->beamSearchStatistics_ = nullptr; - } -} - -namespace { -void lenToStarts(std::vector& starts) { - int pos = 0; - starts.back() = 0; - for (auto& start : starts) { - int tmp = start; - start = pos; - pos += tmp; - } - starts.back() = pos; -} -} // namespace - -void RecurrentGradientMachine::calcSequenceStartPositions() { - std::vector starts(commonSeqInfo_.size() + 1); - for (auto& seqInfo : commonSeqInfo_) { - starts[seqInfo.seqId] = seqInfo.topLevelLength; - } - lenToStarts(starts); - ICpuGpuVector::resizeOrCreate(sequenceStartPositions_, starts.size(), false); - std::copy(starts.begin(), - starts.end(), - sequenceStartPositions_->getMutableData(false)); -} - -void RecurrentGradientMachine::checkOutputConsistency( - OutFrameLine& outFrameLine) { - bool hasSeq = outFrameLine.frames[0]->getOutput().hasSeq(); - for (int i = 0; i < maxSequenceLength_; ++i) { - LayerPtr frame = outFrameLine.frames[i]; - CHECK_EQ(hasSeq, frame->getOutput().hasSeq()); - int numSequences = frame->getOutput().getNumSequences(); - CHECK_EQ(numSeqs_[i], numSequences); - } -} - -void RecurrentGradientMachine::createOutFrameInfo( - OutFrameLine& outFrameLine, - Info& info, - ICpuGpuVectorPtr& sequenceStartPositions, - ICpuGpuVectorPtr& subSequenceStartPositions) { - checkOutputConsistency(outFrameLine); - - if (!outFrameLine.frames[0]->getOutput().hasSeq()) { - createOutFrameInfo_seq( - outFrameLine, info, sequenceStartPositions, subSequenceStartPositions); - } else { - createOutFrameInfo_subseq( - outFrameLine, info, sequenceStartPositions, subSequenceStartPositions); - } -} - -void RecurrentGradientMachine::createOutFrameInfo_seq( - OutFrameLine& outFrameLine, - Info& info, - ICpuGpuVectorPtr& sequenceStartPositions, - ICpuGpuVectorPtr& subSequenceStartPositions) { - std::vector allIds; - info.idIndex.resize(1, 0); // first idIndex = 0 - - const int* starts = sequenceStartPositions_->getData(false); - - for (int i = 0; i < maxSequenceLength_; ++i) { - LayerPtr frame = outFrameLine.frames[i]; - size_t numSequences = frame->getOutput().getNumSequences(); - for (size_t j = 0; j < numSequences; ++j) { - int seqStart = starts[commonSeqInfo_[j].seqId]; - int seqLength = commonSeqInfo_[j].topLevelLength; - allIds.push_back(reversed_ ? (seqStart + seqLength - 1 - i) - : (seqStart + i)); - } - info.idIndex.push_back(allIds.size()); - } - sequenceStartPositions = sequenceStartPositions_; - copyScattedId(allIds, &info.allIds, allIds.size()); - CHECK_EQ(info.idIndex.size(), static_cast(maxSequenceLength_ + 1)); -} - -void RecurrentGradientMachine::createOutFrameInfo_subseq( - OutFrameLine& outFrameLine, - Info& info, - ICpuGpuVectorPtr& sequenceStartPositions, - ICpuGpuVectorPtr& subSequenceStartPositions) { - size_t numSequences = commonSeqInfo_.size(); - std::vector allIds; - info.idIndex.resize(1, 0); // first idIndex = 0 - - const int* starts = sequenceStartPositions_->getData(false); - std::vector subStarts(starts[numSequences] + 1); - for (int i = 0; i < maxSequenceLength_; ++i) { - LayerPtr frame = outFrameLine.frames[i]; - size_t numSequences = frame->getOutput().getNumSequences(); - const int* seqStarts = - frame->getOutput().sequenceStartPositions->getData(false); - for (size_t j = 0; j < numSequences; ++j) { - subStarts[starts[commonSeqInfo_[j].seqId] + i] = - seqStarts[j + 1] - seqStarts[j]; - } - } - lenToStarts(subStarts); - - for (int i = 0; i < maxSequenceLength_; ++i) { - LayerPtr frame = outFrameLine.frames[i]; - size_t numSequences = frame->getOutput().getNumSequences(); - for (size_t j = 0; j < numSequences; ++j) { - int pos = starts[commonSeqInfo_[j].seqId] + i; - int subSeqStart = subStarts[pos]; - int subSeqEnd = subStarts[pos + 1]; - for (int k = subSeqStart; k < subSeqEnd; ++k) { - allIds.push_back(k); - } - } - info.idIndex.push_back(allIds.size()); - } - - ICpuGpuVector::resizeOrCreate( - subSequenceStartPositions, subStarts.size(), false); - int* cpuSubSequenceStartPositions = - subSequenceStartPositions->getMutableData(false); - std::copy(subStarts.begin(), subStarts.end(), cpuSubSequenceStartPositions); - ICpuGpuVector::resizeOrCreate( - sequenceStartPositions, numSequences + 1, false); - int* cpuSequenceStartPositions = - sequenceStartPositions->getMutableData(false); - for (size_t i = 0; i <= numSequences; ++i) { - cpuSequenceStartPositions[i] = subStarts[starts[i]]; - } - copyScattedId(allIds, &info.allIds, allIds.size()); - CHECK_EQ(info.idIndex.size(), static_cast(maxSequenceLength_ + 1)); -} - -/* create scattered id infomation for all realLayer of inFrameLines one time. - * If hasSubseq, will also create scattered sequenceStartPositions infomation - * for all realLayer of inFrameLines one time. - */ -void RecurrentGradientMachine::createInFrameInfo(int inlinkId, - const Argument& input, - PassType passType) { - if (!input.hasSeq()) { - createInFrameInfo_nonseq(inlinkId, input, passType); - } else if (!input.hasSubseq()) { - createInFrameInfo_seq(inlinkId, input, passType); - } else { - createInFrameInfo_subseq(inlinkId, input, passType); - } -} - -void RecurrentGradientMachine::createInFrameInfo_nonseq(int inlinkId, - const Argument& input, - PassType passType) { - std::vector allIds; - - auto& seqInfo = seqInfos_[inlinkId]; - Info* inlinkInfo = &info_[inlinkId]; - inlinkInfo->idIndex.clear(); - for (size_t i = 0; i < seqInfo.size(); ++i) { - allIds.push_back(seqInfo[i].seqId); - } - // copy and check scatterId - copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize()); -} - -void RecurrentGradientMachine::createInFrameInfo_seq(int inlinkId, - const Argument& input, - PassType passType) { - std::vector allIds; - auto& seqInfo = seqInfos_[inlinkId]; - Info* inlinkInfo = &info_[inlinkId]; - inlinkInfo->idIndex.resize(1, 0); // first idIndex = 0 - - for (int i = 0; i < maxSequenceLength_; ++i) { - for (int j = 0; j < numSeqs_[i]; ++j) { - int seqLength = seqInfo[j].topLevelLength; - int seqStart = seqInfo[j].seqStart; - allIds.push_back(reversed_ ? (seqStart + seqLength - 1 - i) - : (seqStart + i)); - } - inlinkInfo->idIndex.push_back(allIds.size()); - } - - // copy and check scatterId - copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize()); - CHECK_EQ(inlinkInfo->idIndex.size(), - static_cast(maxSequenceLength_ + 1)); -} -void RecurrentGradientMachine::createInFrameInfo_subseq(int inlinkId, - const Argument& input, - PassType passType) { - std::vector allIds; - - auto& seqInfo = seqInfos_[inlinkId]; - - Info* inlinkInfo = &info_[inlinkId]; - inlinkInfo->idIndex.resize(1, 0); // first idIndex = 0 - std::vector sequenceStartPositions; - const int* subSequenceStartPositions = nullptr; - - subSequenceStartPositions = input.subSequenceStartPositions->getData(false); - inlinkInfo->seqStartPosIndex.clear(); - inlinkInfo->seqStartPosIndex.push_back(0); // first seqStartPosIndex = 0 - for (int i = 0; i < maxSequenceLength_; ++i) { - sequenceStartPositions.push_back(0); // first element = 0 - for (int j = 0; j < numSeqs_[i]; ++j) { - int subSeqStart = subSequenceStartPositions[seqInfo[j].subSeqStart + i]; - int subSeqEnd = subSequenceStartPositions[seqInfo[j].subSeqStart + i + 1]; - for (int k = subSeqStart; k < subSeqEnd; ++k) { - allIds.push_back(k); - } - sequenceStartPositions.push_back(sequenceStartPositions.back() + - subSeqEnd - subSeqStart); - } - inlinkInfo->idIndex.push_back(allIds.size()); - inlinkInfo->seqStartPosIndex.push_back(sequenceStartPositions.size()); - } - // inFrameLine create sequenceStartPositions one time - CHECK_EQ( - sequenceStartPositions.size(), - static_cast(maxSequenceLength_ + input.getNumSubSequences())); - CHECK_EQ(inlinkInfo->seqStartPosIndex.size(), - static_cast(maxSequenceLength_ + 1)); - createSeqPos(sequenceStartPositions, &inlinkInfo->sequenceStartPositions); - - // copy and check scatterId - copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize()); - CHECK_EQ(inlinkInfo->idIndex.size(), - static_cast(maxSequenceLength_ + 1)); -} - -/* like createInFrameInfo, but for all realLayer of memoryFrameLines*/ -void RecurrentGradientMachine::createMemoryFrameInfo( - MemoryFrameLine* memoryFrameLine, PassType passType) { - const Argument& input = (*memoryFrameLine).rootLayer->getOutput(); - size_t numSequences = input.getNumSequences(); - std::vector allIds; - bool seqFlag = input.hasSeq(); - CHECK(!input.hasSubseq()) - << "Subsequence boot layer for memory is not supported"; - - if (seqFlag) { // for sequenceScatterAgentLayer - std::vector sequenceStartPositions; - sequenceStartPositions.push_back(0); // first element = 0 - const int* starts = input.sequenceStartPositions->getData(false); - for (size_t i = 0; i < numSequences; ++i) { - // memory info adopt info of inlinks[0] - int seqId = seqInfos_[0][i].seqId; - for (int k = starts[seqId]; k < starts[seqId + 1]; ++k) { - allIds.push_back(k); - } - sequenceStartPositions.push_back(sequenceStartPositions.back() + - starts[seqId + 1] - starts[seqId]); - } - createSeqPos(sequenceStartPositions, - &(*memoryFrameLine).sequenceStartPositions); - - } else { // for scatterAgentLayer - for (size_t i = 0; i < numSequences; ++i) { - allIds.push_back(seqInfos_[0][i].seqId); - } - } - // copy and check scatterId - copyScattedId(allIds, &(*memoryFrameLine).allIds, input.getBatchSize()); - // memoryFrameLine select rows in real layer one time - selectRowsOneTime((*memoryFrameLine).rootLayer, - (*memoryFrameLine).allIds, - &(*memoryFrameLine).outArg, - passType); -} - -void RecurrentGradientMachine::copyScattedId(std::vector& srcIds, - IVectorPtr* dstIds, - int size) { - int idSize = srcIds.size(); - CHECK_EQ(idSize, size); - IVector::resizeOrCreate(*dstIds, idSize, useGpu_); - (*dstIds)->copyFrom(srcIds.data(), idSize); - // check - std::sort(srcIds.begin(), srcIds.end()); - for (int i = 0; i < idSize; ++i) { - CHECK_EQ(srcIds[i], i); - } -} - -void RecurrentGradientMachine::selectRowsOneTime(LayerPtr layer, - const IVectorPtr& allIds, - Argument* arg, - PassType passType) { - Argument& src = layer->getOutput(); - if (src.value) { - const MatrixPtr& realV = src.value; - int height = realV->getHeight(); - int width = realV->getWidth(); - Matrix::resizeOrCreate( - arg->value, height, width, /* trans */ false, useGpu_); - arg->value->zeroMem(); - arg->value->selectRows(*realV, *allIds); - if (passType != PASS_TEST) { - Matrix::resizeOrCreate( - arg->grad, height, width, /* trans */ false, useGpu_); - arg->grad->zeroMem(); - } - } - if (src.ids) { - IVector::resizeOrCreate(arg->ids, src.ids->getSize(), useGpu_); - arg->ids->selectFrom(*src.ids, *allIds); - } -} - -void RecurrentGradientMachine::createSeqPos( - const std::vector& sequenceStartPosition, - ICpuGpuVectorPtr* sequenceStartPositions) { - int size = sequenceStartPosition.size(); - const int* data = sequenceStartPosition.data(); - ICpuGpuVector::resizeOrCreate(*sequenceStartPositions, size, false); - (*sequenceStartPositions)->copyFrom(data, size, false); -} - -size_t RecurrentGradientMachine::getGenBatchSize() { - size_t numSequences = 0; - for (auto& memoryFrameLine : memoryFrameLines_) { - if (!memoryFrameLine.rootLayer) continue; - Argument& bootArg = memoryFrameLine.rootLayer->getOutput(); - size_t batchSize = bootArg.getNumSequences(); - if (numSequences) { - CHECK_EQ(numSequences, batchSize); - } else { - numSequences = batchSize; - } - } - CHECK(numSequences) - << "Fail to get batch size in generation. " - "At least one of the Memory layer MUST have a layer that is NOT in " - "the layer group to boot it, and this boot layer is used to " - "decide batch_size in generation process."; - return numSequences; -} - -void RecurrentGradientMachine::generateSequence() { - CHECK_NOTNULL(eosFrameLine_.get()); - CHECK_GE(outFrameLines_.size(), 1UL); - size_t numSequences = getGenBatchSize(); - - resizeBootFrame(numSequences); - // We create only two sub-network in generation, one stores states of all - // layers in previous time step and the other storing the states at current - // time step. - resizeOrCreateFrames(2); - - // outFrameLines_.size() > 1UL - dataArgsSize_ = outFrameLines_.size() - 1; - dataArgs_.resize(dataArgsSize_); - dataArgsFrame_.clear(); - dataArgsFrame_.resize(dataArgsSize_); - - // connect boot frame memory links - std::vector ids(numSequences); - for (size_t i = 0; i < numSequences; ++i) { - ids[i] = i; - } - for (auto& memoryFrameLine : memoryFrameLines_) { - if (memoryFrameLine.rootAgent) { - auto scatterAgent = - dynamic_cast(memoryFrameLine.rootAgent.get()); - scatterAgent->setRealLayer(memoryFrameLine.rootLayer, ids); - } - NeuralNetwork::connect( - memoryFrameLine.agents[0], memoryFrameLine.bootLayer, ids.size()); - } - - // boot layer forward - AsyncGpuBlock asyncGpuBlock; - - for (auto& memoryFrameLine : memoryFrameLines_) { - memoryFrameLine.bootLayer->forward(PASS_TEST); - } - - // init outArg - size_t resultNum = generator_.config.num_results_per_sample(); - size_t maxGenWordCount = - generator_.config.max_num_frames() * numSequences * resultNum; - IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false); - if (resultNum > 1) { - CHECK_LE(resultNum, static_cast(generator_.config.beam_size())); - Matrix::resizeOrCreate(generator_.outArg.in, - /* height */ numSequences, - /* width */ resultNum, - false, - /* useGpu */ false); - } - ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions, - numSequences + 1, - /* useGpu */ false); - if (getBeamSize() > 1) { - beamSearch(numSequences); - } else { - oneWaySearch(numSequences); - } - if (dataArgsSize_) createDataOutlink(); - - size_t size = generator_.ids.size(); - generator_.outArg.ids->resize(size); - generator_.outArg.ids->copyFrom(generator_.ids.data(), size); - - OutFrameLine& outFrameLine = outFrameLines_[0]; - auto dataAgent = dynamic_cast(outFrameLine.agentLayer.get()); - CHECK_NOTNULL(dataAgent); - dataAgent->setData(generator_.outArg); - dataAgent->prefetch(); -} - -void RecurrentGradientMachine::oneWaySearch(size_t batchSize) { - OutFrameLine& outFrameLine = outFrameLines_[0]; - - // finalPaths_[0] stores the generated results of the - // entire batch, so its size exactly equals to batchSize. - finalPaths_.clear(); - finalPaths_.resize(1); - std::vector& finalPaths = finalPaths_[0]; - finalPaths.resize(batchSize); - - seqIds_.resize(batchSize); - std::vector scatterIds; - for (size_t i = 0; i < batchSize; ++i) { - finalPaths[i].seqId = i; - seqIds_[i] = i; - } - - // forward - for (int i = 0; i < maxSequenceLength_; ++i) { - if (i && scatterIds.empty()) break; - int machineCur = i % 2; - int machinePrev = (i - 1) % 2; - // connect memory links - if (i) { - seqIds_.clear(); - for (size_t j = 0; j < batchSize; ++j) { - if (finalPaths[j].seqId != -1) seqIds_.push_back(j); - } - - for (auto& memoryFrameLine : memoryFrameLines_) { - auto scatterAgent = dynamic_cast( - memoryFrameLine.scatterAgents[machineCur].get()); - scatterAgent->setRealLayer(memoryFrameLine.frames[machinePrev], - scatterIds); - scatterAgent->forward(PASS_TEST); - NeuralNetwork::connect(memoryFrameLine.agents[machineCur], - memoryFrameLine.scatterAgents[machineCur]); - } - } - const std::vector inArgs; - std::vector outArgs; - frames_[machineCur]->forward(inArgs, &outArgs, PASS_TEST); - - const IVectorPtr& idVec = outFrameLine.frames[machineCur]->getOutput().ids; - for (size_t j = 0; j < seqIds_.size(); ++j) { - finalPaths[seqIds_[j]].ids.push_back(idVec->getElement(j)); - finalPaths[seqIds_[j]].machineIdVec.push_back(j); - } - - copyDataOutlinkFrame(machineCur); - - // check eos - const IVectorPtr& eosVec = - eosFrameLine_->layers[machineCur]->getOutput().ids; - scatterIds.clear(); - for (size_t j = 0; j < seqIds_.size(); ++j) { - if (eosVec->getElement(j) == 1U) { - // path.seqId = -1 indicates end of generation - // of an input sequence - finalPaths[seqIds_[j]].seqId = -1; - } else { - scatterIds.push_back(j); - } - } - } - - batchMachineIdVec_.clear(); - batchMachineStartPos_.clear(); - int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); - starts[0] = 0; - generator_.ids.clear(); - for (size_t i = 0; i < batchSize; ++i) { - generator_.ids.insert(generator_.ids.end(), - finalPaths[i].ids.begin(), - finalPaths[i].ids.end()); - starts[i + 1] = generator_.ids.size(); - batchMachineIdVec_.insert(batchMachineIdVec_.end(), - finalPaths[i].machineIdVec.begin(), - finalPaths[i].machineIdVec.end()); - } -} - -void RecurrentGradientMachine::connectPrevFrame(int stepId, - std::vector& paths) { - int machineCur = stepId % 2; - int machinePrev = (stepId - 1) % 2; - int beam = getBeamSize(); - machineIds_.clear(); - topIds_.clear(); - seqIds_.clear(); - - for (size_t j = 0; j < paths.size(); ++j) { - machineIds_.push_back(paths[j].machineId); - topIds_.push_back(paths[j].machineId * beam + paths[j].topIndex); - seqIds_.push_back(paths[j].seqId); - } - - for (auto& memoryFrameLine : memoryFrameLines_) { - bool isOutIds = (memoryFrameLine.layerName == outFrameLines_[0].layerName); - auto scatterAgent = dynamic_cast( - memoryFrameLine.scatterAgents[machineCur].get()); - scatterAgent->setRealLayer(memoryFrameLine.frames[machinePrev], - isOutIds ? topIds_ : machineIds_); - scatterAgent->forward(PASS_TEST); - NeuralNetwork::connect(memoryFrameLine.agents[machineCur], - memoryFrameLine.scatterAgents[machineCur]); - } -} - -void RecurrentGradientMachine::forwardFrame(int machineCur) { - // forward - const std::vector inArgs; - std::vector outArgs; - frames_[machineCur]->forward(inArgs, &outArgs, PASS_TEST); - - copyDataOutlinkFrame(machineCur); - - IVectorPtr& ids = outFrameLines_[0].frames[machineCur]->getOutput().ids; - MatrixPtr in = outFrameLines_[0].frames[machineCur]->getOutput().in; - IVectorPtr& eos = eosFrameLine_->layers[machineCur]->getOutput().ids; - if (useGpu_) { - IVector::resizeOrCreate(cpuId_, ids->getSize(), false /* useGpu */); - cpuId_->copyFrom(*ids); - Matrix::resizeOrCreate(cpuProb_, - in->getHeight(), - in->getWidth(), - false /* trans */, - false /* useGpu */); - cpuProb_->copyFrom(*in); - IVector::resizeOrCreate(cpuEos_, eos->getSize(), false /* useGpu */); - cpuEos_->copyFrom(*eos); - } else { - cpuId_ = ids; - cpuProb_ = in; - cpuEos_ = eos; - } -} - -void RecurrentGradientMachine::singlePathExpand(Path& curPath, - size_t curPathId, - std::vector& newPaths, - size_t expandWidth) { - int calc_id = - gDiyProbStart ? gDiyProbStart(curPath.ids.size(), curPath.ids.data()) : 0; - - const int* idVec = cpuId_->getData(); - const real* probMat = cpuProb_->getData(); - const int* eosVec = cpuEos_->getData(); - - for (size_t k = 0; k < expandWidth; k++) { - int index = curPathId * expandWidth + k; - int id = idVec[index]; - real prob = probMat[index]; - /* - * Ordinarily, beam search greedily expands the most promising expandWidth - * paths that currently are ALWAYS returned by MaxIdLayer. - * In one condition, if user customizes the beam search procedure by - * restricting the expansion within a user defined subset, - * as a result, MaxIdLayer possibly COULD NOT return expandWidth - * vaild expansions, and it will use -1 to indicate the end of valid - * expansion candidates. - */ - if (id == -1) break; - - real newLogProb = generator_.config.log_prob() ? std::log(prob) : prob; - Path newPath( - curPath, id, newLogProb, curPathId /*machineId*/, k /*topIndex*/); - if (this->beamSearchCtrlCallbacks_) { - if (beamSearchCtrlCallbacks_->stopDetermineCandidates( - newPath.seqId, newPath.ids, newPath.probHistory)) - return; - } - // outFrameLines_.size() > 1UL - if (dataArgsSize_) { - newPath.machineIdVec = curPath.machineIdVec; - newPath.machineIdVec.push_back(curPathId); - } - bool atEos = - eosVec[index] == 1U || newPath.ids.size() >= (size_t)maxSequenceLength_; - // adjustNewPath - newPath.adjustProb(calc_id, atEos); - if (this->beamSearchCtrlCallbacks_) { - this->beamSearchCtrlCallbacks_->normOrDropNode( - newPath.seqId, newPath.ids, newPath.probHistory, &newPath.logProb); - } - if (!newPath.isDropable()) { - atEos ? finalPaths_[curPath.seqId].push_back(newPath) - : newPaths.push_back(newPath); - } - } // for expandWidth - - if (gDiyProbStop) { - gDiyProbStop(calc_id); - } -} - -void RecurrentGradientMachine::beamExpand(std::vector& paths, - std::vector& newPaths) { - size_t candidatePathCount = paths.size(); - // idVec.size() could be larger than candidatePathCount * beam, - // so user can drop some node customly. - CHECK_EQ(cpuId_->getSize() % candidatePathCount, 0UL); - size_t expandWidth = cpuId_->getSize() / candidatePathCount; - - // iterate over each sequence - size_t totalExpandCount = 0; - int prevSeqId = -1; - int curSeqId = 0; - for (size_t j = 0; j <= candidatePathCount; j++) { - // expansions of a single sequence are all processed - curSeqId = (j < candidatePathCount ? paths[j].seqId : curSeqId + 1); - if (prevSeqId != -1 && curSeqId != prevSeqId) { - totalExpandCount += beamShrink(newPaths, prevSeqId, totalExpandCount); - } - if (j == candidatePathCount) return; - singlePathExpand(paths[j], j, newPaths, expandWidth); - - prevSeqId = paths[j].seqId; - } // for paths -} - -// Drop extra nodes to beam size. -size_t RecurrentGradientMachine::beamShrink(std::vector& newPaths, - size_t seqId, - size_t totalExpandCount) { - size_t minNewPathSize = - std::min(getBeamSize(), newPaths.size() - totalExpandCount); - if (!minNewPathSize) { - return 0; - } - std::nth_element(newPaths.begin() + totalExpandCount, - newPaths.begin() + totalExpandCount + minNewPathSize, - newPaths.end(), - Path::greaterPath); - newPaths.resize(totalExpandCount + minNewPathSize); - - real minPathLogProb = - std::min_element(newPaths.end() - minNewPathSize, newPaths.end()) - ->logProb; - real maxPathLogProb = - std::max_element(newPaths.end() - minNewPathSize, newPaths.end()) - ->logProb; - - // Remove the already formed paths that are relatively short - finalPaths_[seqId].erase( - std::remove_if(finalPaths_[seqId].begin(), - finalPaths_[seqId].end(), - [&](Path& p) { return p.logProb < minPathLogProb; }), - finalPaths_[seqId].end()); - for (auto p : finalPaths_[seqId]) { - if (minFinalPathLogProb_[seqId] > p.logProb) { - minFinalPathLogProb_[seqId] = p.logProb; - } - } - - if (finalPaths_[seqId].size() >= getBeamSize() && - minFinalPathLogProb_[seqId] >= maxPathLogProb) { - newPaths.resize(totalExpandCount); - return 0; - } - return minNewPathSize; -} - -void RecurrentGradientMachine::fillGenOutputs() { - size_t numResults = generator_.config.num_results_per_sample(); - for (size_t i = 0; i < finalPaths_.size(); ++i) { - size_t minFinalPathsSize = std::min(numResults, finalPaths_[i].size()); - std::partial_sort(finalPaths_[i].begin(), - finalPaths_[i].begin() + minFinalPathsSize, - finalPaths_[i].end(), - Path::greaterPath); - finalPaths_[i].resize(minFinalPathsSize); - } - - generator_.ids.clear(); - int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); - starts[0] = 0; - if (numResults > 1) { - int idsProbSaveSize = 0; - for (auto inSeq : finalPaths_) { - for (auto path : inSeq) idsProbSaveSize += path.ids.size(); - idsProbSaveSize += inSeq.size(); - } - Matrix::resizeOrCreate( - generator_.outArg.value, idsProbSaveSize, 1, false, false); - real* idsProb = generator_.outArg.value->getData(); - - real* probs = generator_.outArg.in->getData(); - size_t curPos = 0; - for (size_t i = 0; i < finalPaths_.size(); ++i) { - for (size_t j = 0; j < finalPaths_[i].size(); ++j) { - Path& path = finalPaths_[i][j]; - size_t genLen = path.ids.size(); - generator_.ids.push_back(genLen); // sequence size - generator_.ids.insert( - generator_.ids.end(), path.ids.begin(), path.ids.end()); - generator_.ids.push_back(-1); // end of sequence - - memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen); - curPos += genLen; - idsProb[curPos++] = -1.0; - probs[i * numResults + j] = path.logProb; - } - starts[i + 1] = generator_.ids.size(); - } - } else { - for (size_t i = 0; i < finalPaths_.size(); ++i) { - CHECK(!finalPaths_[i].empty()); - Path& path = finalPaths_[i][0]; - generator_.ids.insert( - generator_.ids.end(), path.ids.begin(), path.ids.end()); - starts[i + 1] = starts[i] + path.ids.size(); - } - } -} - -void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) { - for (size_t i = 0; i < dataArgsSize_; i++) { - Argument outFrame; - outFrame.resizeAndCopyFrom( - outFrameLines_[i + 1].frames[machineCur]->getOutput(), useGpu_); - dataArgsFrame_[i].emplace_back(outFrame); - } -} - -void RecurrentGradientMachine::createDataOutlinkSelRowsInfo( - bool isSeq, std::vector& outArgs) { - batchMachineIdVec_.clear(); - - size_t seqIdx = 0; - for (size_t i = 0; i < finalPaths_.size(); ++i) { - for (size_t j = 0; j < finalPaths_[i].size(); ++j) { - std::vector& machineIdVec = finalPaths_[i][j].machineIdVec; - if (isSeq) { - for (size_t i = 0; i < machineIdVec.size(); ++i) { - size_t rowId = machineIdVec[i]; - int* seqPos = - outArgs[i].sequenceStartPositions->getMutableData(false); - batchMachineIdVec_.push_back(seqPos[rowId]); - } - } else { - batchMachineIdVec_.insert( - batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end()); - } - seqIdx++; - } - } -} - -void RecurrentGradientMachine::createDataOutlinkCopySizeInfo( - bool isSeq, std::vector& outArgs, std::vector& copySize) { - size_t totalSeqNum = std::accumulate( - finalPaths_.begin(), - finalPaths_.end(), - 0UL, - [](size_t a, const std::vector& b) { return a + b.size(); }); - copySize.resize(totalSeqNum, 1); - - batchMachineStartPos_.resize(totalSeqNum + 1, 0); - if (isSeq) { - ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions; - CHECK_EQ(static_cast(inputSeqStartPos->getSize() - 1), - getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size()); - int* starts = inputSeqStartPos->getMutableData(false); - int seqId = 0; - for (size_t i = 0; i < finalPaths_.size(); ++i) { - for (size_t j = 0; j < finalPaths_[i].size(); ++j) { - copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i] - : starts[j + 1] - starts[j]; - batchMachineStartPos_[seqId + 1] = - batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size(); - seqId++; - } - } - } else { - for (size_t i = 0; i < finalPaths_[0].size(); ++i) - batchMachineStartPos_[i + 1] = - batchMachineStartPos_[i] + finalPaths_[0][i].ids.size(); - } -} - -void RecurrentGradientMachine::createDataOutlink() { - for (size_t i = 0; i < dataArgsSize_; i++) { - bool isSeq = dataArgsFrame_[i][0].hasSeq(); - std::vector copySize; - createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize); - createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]); - - dataArgs_[i].concat(dataArgsFrame_[i], - batchMachineIdVec_, - batchMachineStartPos_, - copySize, - useGpu_, - HPPL_STREAM_1, - PASS_TEST); - auto dataAgent = - dynamic_cast(outFrameLines_[i + 1].agentLayer.get()); - CHECK_NOTNULL(dataAgent); - dataAgent->setData(dataArgs_[i]); - } -} - -void RecurrentGradientMachine::beamSearch(size_t batchSize) { - finalPaths_.clear(); - finalPaths_.resize(batchSize); - seqIds_.resize(batchSize); - minFinalPathLogProb_.clear(); - minFinalPathLogProb_.resize(batchSize, 0); - - std::vector paths; - std::vector newPaths; - for (size_t i = 0; i < batchSize; ++i) { - paths.push_back(Path(i)); - if (this->beamSearchCtrlCallbacks_) { - paths.back().recordHistory(); - } - } - - // restart beam search - stopBeamSearch_ = false; - for (int i = 0; i < maxSequenceLength_; ++i) { - int machineCur = i % 2; - std::unique_ptr< - ScopedCallbacks> - statisticsBlock; - if (this->beamSearchStatistics_) { - auto ptr = - new ScopedCallbacks(beamSearchStatistics_->onEachStepStarted, - beamSearchStatistics_->onEachStepStoped, - i); - statisticsBlock.reset(ptr); - } - if (stopBeamSearch_) break; - - if (i) connectPrevFrame(i, paths); - - if (this->beamSearchCtrlCallbacks_) { - std::vector*> prefixes; - prefixes.resize(paths.size()); - std::transform( - paths.begin(), paths.end(), prefixes.begin(), [](const Path& p) { - return const_cast*>(&p.ids); - }); - beamSearchCtrlCallbacks_->beamSearchCandidateAdjust( - prefixes, frames_[machineCur].get(), i); - } - - forwardFrame(machineCur); - beamExpand(paths, newPaths); - if (newPaths.empty()) break; - - paths = newPaths; - newPaths.clear(); - } // end for machineCur - fillGenOutputs(); -} - -void RecurrentGradientMachine::Path::adjustProb(int calc_id, bool atEos) { - if (gDiyProbMethod) { - logProb = gDiyProbMethod(calc_id, ids.size(), ids.data(), logProb, atEos); - } -} - -} // namespace paddle diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h deleted file mode 100644 index 7e943cebd35234ba7af357c9f64fde6b0a9546ce..0000000000000000000000000000000000000000 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h +++ /dev/null @@ -1,580 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "GradientMachine.h" -#include "NeuralNetwork.h" - -#include "paddle/utils/Locks.h" - -namespace paddle { - -/** - * Private data class declares. - * Used for user customized beam search. - */ -class BeamSearchControlCallbacks; -class BeamSearchStatisticsCallbacks; - -class RecurrentGradientMachine : public NeuralNetwork { - public: - RecurrentGradientMachine(const std::string& subModelName, - NeuralNetwork* rootNetwork); - - // Disable copy and assign. - RecurrentGradientMachine(const RecurrentGradientMachine& other) = delete; - RecurrentGradientMachine& operator=(const RecurrentGradientMachine& other) = - delete; - - virtual ~RecurrentGradientMachine() { - this->removeBeamSearchStatisticsCallbacks(); - this->removeBeamSearchControlCallbacks(); - } - - virtual void init(const ModelConfig& config, - ParamInitCallback callback, - const std::vector& parameterTypes, - bool useGpu); - - virtual void prefetch(const std::vector& inArgs); - - virtual void forward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType); - - virtual void backward(const UpdateCallback& callback = nullptr); - - void forwardBackward(const std::vector& inArgs, - std::vector* outArgs, - PassType passType, - const UpdateCallback& callback); - - virtual void resetState() {} - virtual void eval(Evaluator* evaluator) const; - - const std::vector& getParameterIds() { return parameterIds_; } - - /** - * @brief BeamSearchCandidatesAdjustCallback - * - * Adjust searching candidates to restrict beam search - * searching within a limited subset of all possibile paths. - * - * The first parameter is the prefixes of all formed paths in current - * beam search step, whose type is basically int[][]. - * - * The second parameter is a pointer to the network used to generate sequence, - * user can use this pointer to tranverse each layer in the network to - * modify behaivors of a particular layer. - * - * The third parameter is an integer to indicate the iteration number of - * beam search, so that user can customize different operations in different - * beam search iterations. - */ - typedef std::function*>&, NeuralNetwork*, const int)> - BeamSearchCandidatesAdjustCallback; - - /** - * @brief DropCallback - * - * Drop a whole prefix or one candidate in beam search or not. - * - * The first parameter is sequence index in a batch - * - * The second parameter is one path in beam search, - * which is made up of node indices. - * - * The third parameter is probabilites for each node in this path. - * - * Return true if this prefix or candidate is expected to be dropped. - */ - typedef std::function&, const std::vector&)> - DropCallback; - - /** - * @brief NormOrDropNodeCallback - * - * Normalize a path's probabilities or just drop it by modifying path.logProb - * - * The first parameter is sequence index in a batch - * - * The second parameter is path.ids - * - * The third parameter is probabilites for each node in this path. - * - * The fourth parameter is the probability of the whole path. - */ - typedef std::function&, std::vector&, real*)> - NormOrDropNodeCallback; - - /** - * @brief Register beam search control callbacks. Used for prediction. - * - * @param queryBeamSearch: Give the sequences already formed, return the - * nodes expected to be expanded. - * Input: A pointer to an array holding pathes which have been expanded - * Return: A pointer to an array holding nodes wanted to be expanded. - * - * @param dropOneNode: Early drop a node in one beam search step. - * Given the path formed and probability history, decide whether a node - * should be dropped or not. - * - * @param stopBeamSearch: Early stop a path in one beam search step. - * Given the path and probability history, decide whether a path - * should be dropped or not. - */ - void registerBeamSearchControlCallbacks( - const BeamSearchCandidatesAdjustCallback& adjustBeamSearch, - const NormOrDropNodeCallback& normOrDropNode, - const DropCallback& stopBeamSearch); - - /** - * @brief Remove user costumized beam search callbacks, - * - * make sequence generation acts like normal beam search. - */ - void removeBeamSearchControlCallbacks(); - - /** - * @brief EachStepCallback - * - * Invoke with beam search step. - */ - typedef std::function EachStepCallback; - - /** - * @brief register statistics methods for performance profile of beam search. - * - * @param onEachStepStarted: invoke once a beam search step starts. - * Its input is index of the beam search step. - * - * @param onEachStepStoped: invoke once a beam search step ends. - * Its input is index of the beam search step. - */ - void registerBeamSearchStatisticsCallbacks( - const EachStepCallback& onEachStepStarted, - const EachStepCallback& onEachStepStoped); - - /** - * @brief Remove beam search callbacks. - */ - void removeBeamSearchStatisticsCallbacks(); - - /** - * @brief Stop beam search for current source. - * - * Will restart beam search in the next forward - */ - void stopBeamSearch(); - - struct Path { - /** - * @brief ids, path of beam search. - */ - std::vector ids; - - /** - * @brief idsProb, log probability of each generated word. - */ - std::vector idsProb; - - /** - * @brief logProb, current probability of path. - */ - real logProb; - - int machineId; // index of sample in frame - int topIndex; // index of MaxIdLayer output in one sample - int seqId; // index of sequence in batch generation - std::vector machineIdVec; - - /** - * @brief A record of each node's probality in a formed path in beam search. - * - * @note It could be empty when history is not recorded. If the history is - * wanted to be recorded, recordHistory() MUST be invoked first. - */ - std::vector probHistory; - - /** - * @brief Path default ctor, first logProb is 0. - */ - Path() { - logProb = 0; - seqId = 0; - } - explicit Path(size_t seqId) : seqId(seqId) { logProb = 0; } - - /** - * @brief Create a new path based on an old path and - * a new node with probability. - * - * @param old old path - * @param newId index of the new node - * @param logProb probability of the new node. - * @param machineId sample index of a frame in RNN - * @param topIndex index of MaxIdLayer output in one sample - */ - Path(Path& old, int newId, real logProb, int machineId, int topIndex) - : ids(old.ids), - idsProb(old.idsProb), - logProb(old.logProb + logProb), - machineId(machineId), - topIndex(topIndex), - seqId(old.seqId) { - ids.push_back(newId); - idsProb.push_back(logProb); - if (!old.probHistory.empty()) { - this->probHistory = old.probHistory; - // probHistory store current prob, not sum - this->probHistory.push_back(logProb); - } - } - - /** - * @brief operator < - * - * Path a < Path b means log probability of a is smaller than that of b - */ - bool operator<(const Path& other) const { - return (logProb < other.logProb); - } - - static bool greaterPath(const Path& a, const Path& b) { return (b < a); } - - /** - * @brief Start recording history in this path. - */ - void recordHistory() { this->probHistory.push_back(this->logProb); } - - /** - * @brief Adjust probability for DIY beam search interface. - * In normal situation, it will do nothing. - * - * @param calc_id: the object id for DIY beam search interface. - * @param atEos: at end of sequence or not. - */ - void adjustProb(int calc_id, bool atEos = false); - - /** - * @brief isDropable indacating whether the current node will be - * dropped or not in beam search. - * - * @note: if logProb is -inf, current node will be dropped. - * @return true to drop the current node. - */ - bool isDropable() const { return std::isinf(logProb) && logProb < 0; } - }; - - /** - * @brief access beam search results. - * @return beam search results. - */ - const std::vector>& getFinalPaths() const { - return this->finalPaths_; - } - - protected: - std::vector commonSeqInfo_; - ICpuGpuVectorPtr sequenceStartPositions_; - void calcSequenceStartPositions(); - void checkInputConsistency(int inlinkId, - const std::vector& seqInfo); - void reorganizeInput(PassType passType); - void reorganizeOutput(PassType passType); - void connectFrames(PassType passType); - void calcNumSequencesAtEachStep(); - - void resizeOrCreateFrames(int numFrames); - void resizeBootFrame(int numSequences); - - void generateSequence(); - void oneWaySearch(size_t batchSize); - void beamSearch(size_t batchSize); - - struct InFrameLine { - std::string linkName; - LayerPtr inLayer; - std::vector agents; // Scatter Agents to reform batch input - Argument outArg; // scatter output argument - }; - std::vector inFrameLines_; - - struct OutFrameLine { - std::string layerName; - LayerPtr agentLayer; - std::vector frames; - }; - std::vector outFrameLines_; - - struct MemoryFrameLine { - std::string layerName; - std::string linkName; - LayerPtr bootLayer; // actually used biasLayer or rootAgent - LayerPtr biasLayer; - LayerPtr rootLayer; // layer in root network to boot this memory - LayerPtr rootAgent; // agent to link rootLayer - std::vector frames; - std::vector agents; - std::vector scatterAgents; // scatter agent used by beam search - Argument outArg; // scatter output argument - // Different memoryFrameLine have different element as follows - IVectorPtr allIds; // scattered id of realLayer - ICpuGpuVectorPtr - sequenceStartPositions; // scattered sequenceStartPositions - }; - std::vector memoryFrameLines_; - - // Each inFrameLines(inlinks) has its own info(elements) below, - // and all outFrameLines(outlinks) share the info with one inFrameLine, - // which is assigned by targetInfoInlinkId_. - struct Info { - // The original positions in the original batch - IVectorPtr allIds; // scattered id of realLayer [batchSize] - - // index of allIds for each step [maxSequenceLength_] - // idIndex[i] is the total length of the first i sequences - std::vector idIndex; - - ICpuGpuVectorPtr - sequenceStartPositions; // scattered sequenceStartPositions - std::vector seqStartPosIndex; // index of sequenceStartPositions - }; - std::vector info_; // for input - - // numSeqs_[i] is the number sequences which is longer than i (for sequence - // data) or has more than i subsequences (for subsequence data) - // Equivalently, numSeqs_[i] is the number of sequences at step i; - std::vector numSeqs_; - - std::vector> seqInfos_; - - void checkOutputConsistency(OutFrameLine& outFrameLine); - - /* create scattered id infomation for all realLayer of inFrameLines one time. - * If hasSubseq, will also create scattered sequenceStartPositions infomation - * for all realLayer of inFrameLines one time. - */ - void createInFrameInfo(int inlinks_id, - const Argument& input, - PassType passType); - void createInFrameInfo_nonseq(int inlinks_id, - const Argument& input, - PassType passType); - void createInFrameInfo_seq(int inlinks_id, - const Argument& input, - PassType passType); - void createInFrameInfo_subseq(int inlinks_id, - const Argument& input, - PassType passType); - - void createOutFrameInfo(OutFrameLine& outFrameLine, - Info& info, - ICpuGpuVectorPtr& sequenceStartPositions, - ICpuGpuVectorPtr& subSequenceStartPositions); - void createOutFrameInfo_seq(OutFrameLine& outFrameLine, - Info& info, - ICpuGpuVectorPtr& sequenceStartPositions, - ICpuGpuVectorPtr& subSequenceStartPositions); - void createOutFrameInfo_subseq(OutFrameLine& outFrameLine, - Info& info, - ICpuGpuVectorPtr& sequenceStartPositions, - ICpuGpuVectorPtr& subSequenceStartPositions); - - void createMemoryFrameInfo(MemoryFrameLine* memoryFrameLine, - PassType passType); - - void copyScattedId(std::vector& srcIds, IVectorPtr* dstIds, int size); - - void selectRowsOneTime(LayerPtr layer, - const IVectorPtr& allIds, - Argument* arg, - PassType passType); - - void createSeqPos(const std::vector& sequenceStartPosition, - ICpuGpuVectorPtr* sequenceStartPositions); - - // for generator - struct EosFrameLine { - std::vector layers; - }; - std::unique_ptr eosFrameLine_; - - struct Generator { - GeneratorConfig config; - std::vector ids; // store generated sequences - std::vector idsProb; // log probability of each generated word - Argument outArg; // final output argument - }; - bool generating_; - Generator generator_; - - std::vector> frames_; - - NeuralNetwork* rootNetwork_; - bool reversed_; - - int maxSequenceLength_; // Max top-level length - bool useGpu_; - bool stopBeamSearch_; - - std::vector - parameterIds_; // parameters actually used by this Layer Group - - // store final argument of outFrameLines_ - std::vector dataArgs_; - // store each frame's output argument of outFrameLines_ - std::vector> dataArgsFrame_; - size_t dataArgsSize_; // size of dataArgs_ = size of dataArgsFrame_ - - IVectorPtr cpuId_; - MatrixPtr cpuProb_; - IVectorPtr cpuEos_; - - private: - /* - * @return beam size in beam search - */ - size_t getBeamSize() { return generator_.config.beam_size(); } - - /* - * @return number of sequence in a batch in generation - */ - size_t getGenBatchSize(); - - /* - * @brief store output of the machineCur-th frame during generation, for - * creating the final outlink after the entire generation process is finished. - * - * In generation, if the layer group has more than 1 outlink, the first - * one is reserved to store the generated word indices, the others are data - * outlinks, that can be used like a common layer in the network. - * - * @param machineCur : index to access the layer group frame in - * currrent generation step. - */ - void copyDataOutlinkFrame(size_t machineCur); - - /* - * @brief In generation, if the layer group has more than 1 outlink, outlink - * except the first one is a data outlink. In RecurrentLayerGroup, each time - * step is a separate Network, outputs of a layer inside the - * RecurrentLayerGroup are stored in separate Arguments. If one layer is - * specified as an outlink of RecurrentLayerGroup. This function will - * collect outputs in each time step of each generated sequence which are - * dispersed in separate Arguments to form a new single Argument as output of - * RecurrentLayerGroup. - */ - void createDataOutlink(); - - /* - * @brief decide to select how many rows from the Matrix stored the forward - * pass results from a start position. - * - * @param isSeq: a flag indicating whetehr the layer to be output of the - * RecurrentGradientMachine is a sequence or not - * @param outArgs: all of the the returned Arguments of the forward pass - * during the generation process. - * @param copySize: the returned result, number of rows to select from the - * Matrix stored the forward pass results from a start position. - */ - void createDataOutlinkCopySizeInfo(bool isSeq, - std::vector& outArgs, - std::vector& copySize); - - /* - * @brief decide index of the start row for each time step of a generated - * sequence in Matrix stored the entire beam search batch's forward pass - * results. - * - * @param isSeq: a flag indicating whether the layer to be output of the - * RecurrentGradientMachine is a sequence or not - * @param outArgs: all of the returned Arguments of the forward pass - * during the generation process. - */ - void createDataOutlinkSelRowsInfo(bool isSeq, std::vector& outArgs); - - /* - * @brief used in beam search, connect previous frame to form recurrent link - * @param stepId : iteration number of generation process. - * It equals to the length of longest half-generated sequence. - * @param paths : half-generated paths that are going to be expanded - * in current beam search iteration. - */ - void connectPrevFrame(int stepId, std::vector& paths); - - /* - * @brief used in beam search, forward current recurrent frame - * @param machineCur : index to access the layer group frame in - * currrent generation step. - */ - void forwardFrame(int machineCur); - - /* - * @brief reduce all expanded paths to beam size. - * - * @param newPaths : newPaths[totalExpandCount : ] stores all expanded paths - * for the seqId-th sequence - * @param seqId : sequence index in a batch - * @param totalExpandCount : number of already shrinked paths in newPaths - * @return size of retained paths at the end of a beam search iteration - */ - size_t beamShrink(std::vector& newPaths, - size_t seqId, - size_t totalExpandCount); - - /* - * @brief expand a single path to expandWidth new paths - * with highest probability - * @param curPath : path to be expanded - * @param curPathId : index of curPath in member newPaths - * @param expandWidth : number of paths to be expanded - */ - void singlePathExpand(Path& curPath, - size_t curPathId, - std::vector& newPaths, - size_t expandWidth); - - /* - * @brief A new beam search iteration. Each half-generated paths in previous - * beam search iteration are further expanded to beam_size new paths - * with highest probabilities, and then all the expanded paths are again - * reduced to beam_size paths according to their log probabilities. - * @param paths : half-generated paths in previous iteration. - * @param newPaths : paths expanded and then reduces in current iteration. - */ - void beamExpand(std::vector& paths, std::vector& newPaths); - - /* - * @brief fill sequence start positions and some other information that are - * uesed by the "text_printer" evaluator. - */ - void fillGenOutputs(); - - std::vector machineIds_; - std::vector topIds_; - std::vector seqIds_; - std::vector batchMachineIdVec_; - std::vector batchMachineStartPos_; - std::vector> finalPaths_; - std::vector minFinalPathLogProb_; - BeamSearchControlCallbacks* beamSearchCtrlCallbacks_; - BeamSearchStatisticsCallbacks* beamSearchStatistics_; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/AddtoLayer.cpp b/paddle/gserver/layers/AddtoLayer.cpp deleted file mode 100644 index 75e17f52df64253232dc8fc042d0a1a8e7d98e26..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/AddtoLayer.cpp +++ /dev/null @@ -1,79 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "AddtoLayer.h" - -#include "paddle/utils/Logging.h" - -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(addto, AddtoLayer); - -bool AddtoLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - - return true; -} - -void AddtoLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(0)->getHeight(); - int size = getSize(); - - reserveOutput(batchSize, size); - - MatrixPtr outV = getOutputValue(); - for (size_t i = 0; i != inputLayers_.size(); ++i) { - MatrixPtr input = getInputValue(i); - i == 0 ? outV->assign(*input) : outV->add(*input); - } - /* add the bias-vector */ - if (biases_.get() != NULL) { - outV->addBias(*(biases_->getW()), 1); - } - - /* activation */ { forwardActivation(); } -} - -void AddtoLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { backwardActivation(); } - - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - /* Calculate the input layers error */ - MatrixPtr preGrad = getInputGrad(i); - if (NULL != preGrad) { - preGrad->add(*getOutputGrad()); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/AddtoLayer.h b/paddle/gserver/layers/AddtoLayer.h deleted file mode 100644 index 6ea54f4a53d466594055db2fb5167fa1a9d6c9da..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/AddtoLayer.h +++ /dev/null @@ -1,63 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * This layer just simply add all input layers together, then activate - * the sum inputs. Each input of this layer should be the same size, - * which is also the output size of this layer. - * \f[ - * y=f(\sum_{i}x_i + b) - * \f] - * where \f$y\f$ is output, \f$x\f$ is input, \f$b\f$ is bias, and \f$f\f$ is - * activation function. - * - * The config file api is addto_layer. - */ -class AddtoLayer : public Layer { - protected: - std::unique_ptr biases_; - - public: - explicit AddtoLayer(const LayerConfig& config) : Layer(config) {} - - ~AddtoLayer() {} - - /** - * Intialization of AddtoLayer. - */ - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - /** - * Forward propagation. - * @note There is no weight matrix for each input, - * because it just a simple add operation. - */ - void forward(PassType passType) override; - - /** - * Backward propagation. - */ - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/AgentLayer.cpp b/paddle/gserver/layers/AgentLayer.cpp deleted file mode 100644 index e2f73f88f59278c6e6e6f0a1fe8457393d53f44a..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/AgentLayer.cpp +++ /dev/null @@ -1,281 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "AgentLayer.h" - -#include "paddle/utils/Logging.h" - -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(agent, AgentLayer); - -bool AgentLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - CHECK_EQ(config_.inputs_size(), 0); - if (!Layer::init(layerMap, parameterMap)) { - return false; - } - setNeedGradient(true); - return true; -} - -void AgentLayer::forward(PassType passType) { - Layer::forward(passType); - - Argument& realOutput = realLayer_->getOutput(); - int realNumSequences = realOutput.getNumSequences(); - CHECK_LE(numSamples_, realNumSequences); - - // get Arguments from real layers - if (numSamples_ > 0 && numSamples_ < realNumSequences) { - if (realOutput.hasSeq()) { - int numRows = - realOutput.sequenceStartPositions->getData(false)[numSamples_]; - output_.subArgFrom(realOutput, - /* offset */ 0, - numRows, - getSize(), - useGpu_, - /* trans */ false, - /* seqFlag */ true, - /* seqStart */ 0, - /* seqSize */ numSamples_ + 1); - } else { - output_.subArgFrom( - realOutput, /* offset */ 0, numSamples_, getSize(), useGpu_); - } - } else { - output_ = realOutput; - } -} - -bool GatherAgentLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - CHECK_EQ(config_.inputs_size(), 0); - if (!Layer::init(layerMap, parameterMap)) { - return false; - } - setNeedGradient(true); - return true; -} - -void GatherAgentLayer::copyIdAndSequenceInfo( - ICpuGpuVectorPtr sequenceStartPositions, - ICpuGpuVectorPtr subSequenceStartPositions, - const IVectorPtr& ids, - const std::vector& idIndex) { - output_.sequenceStartPositions = sequenceStartPositions; - output_.subSequenceStartPositions = subSequenceStartPositions; - allIds_ = ids; - idIndex_ = idIndex; -} - -void GatherAgentLayer::forward(PassType passType) { - Layer::forward(passType); - forwardIds(passType); - forwardValue(passType); -} - -void GatherAgentLayer::forwardValue(PassType passType) { - MatrixPtr valueReal = realLayers_[0]->getOutputValue(); - if (!valueReal) return; - - int height = allIds_->getSize(); - int width = this->getSize(); - resetOutput(height, width); - idsVec_.resize(idIndex_.size()); - - const MatrixPtr& outV = getOutputValue(); - - for (size_t i = 0; i < realLayers_.size(); ++i) { - const MatrixPtr& realV = realLayers_[i]->getOutputValue(); - idsVec_[i] = IVector::create(allIds_->getData() + idIndex_[i], - /* size */ realV->getHeight(), - useGpu_); - realV->addToRows(*outV, *idsVec_[i]); - } -} - -namespace { - -// dest[index[i]] <- src[i] for each i -void copyElements(const IVector& srcVec, - const IVector& indexVec, - IVector& destVec) { - const int* src = srcVec.getData(); - const int* index = indexVec.getData(); - int* dest = destVec.getData(); - int len = indexVec.getSize(); - CHECK_EQ(srcVec.getSize(), indexVec.getSize()); - for (int i = 0; i < len; ++i) { - dest[index[i]] = src[i]; - } -} -} // namespace - -void GatherAgentLayer::forwardIds(PassType passType) { - IVectorPtr realId = realLayers_[0]->getOutputLabel(); - if (!realId) return; - - IVector::resizeOrCreate(output_.ids, allIds_->getSize(), useGpu_); - IVectorPtr outId = output_.ids; - idsVec_.resize(idIndex_.size()); - - for (size_t i = 0; i < realLayers_.size(); ++i) { - const IVectorPtr& realId = realLayers_[i]->getOutputLabel(); - idsVec_[i] = IVector::create(allIds_->getData() + idIndex_[i], - /* size */ realId->getSize(), - useGpu_); - execViaCpu(©Elements, *realId, *idsVec_[i], *outId); - } -} - -void GatherAgentLayer::backward(const UpdateCallback& callback) { - (void)callback; - const MatrixPtr& outputGrad = getOutputGrad(); - - for (size_t i = 0; i < realLayers_.size(); ++i) { - const MatrixPtr& realG = realLayers_[i]->getOutputGrad(); - if (realG) { - realG->selectRows(*outputGrad, *idsVec_[i]); - } - } -} - -bool ScatterAgentLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - CHECK_EQ(config_.inputs_size(), 0); - if (!Layer::init(layerMap, parameterMap)) { - return false; - } - setNeedGradient(true); - return true; -} - -void ScatterAgentLayer::forward(PassType passType) { - Layer::forward(passType); - CHECK_EQ(realLayer_->getDeviceId(), this->getDeviceId()); - - int width = this->getSize(); - if (selectionMode_) { - forwardWithSelection(passType); - } else { - if (realOutArg_.hasSeq()) { - output_.subArgFrom(realOutArg_, - /* offset */ idIndex_, - idSize_, - width, - useGpu_, - /* trans */ false, - /* seqFlag */ true, - /* seqStart */ seqStartPosIndex_, - /* seqSize */ numSequences_); - } else { - output_.subArgFrom( - realOutArg_, /* offset */ idIndex_, idSize_, width, useGpu_); - } - } -} - -void ScatterAgentLayer::backward(const UpdateCallback& callback) { - (void)callback; - - CHECK(!selectionMode_); - - const MatrixPtr& outputGrad = realOutArg_.grad; - const MatrixPtr& realGrad = realLayer_->getOutputGrad(); - if (realGrad) { - // for agent in inFrameLines and memoryFrameLines, - // only first scatterAgentLayer should do addToRows in backward - if (handleBackward_) { - outputGrad->addToRows(*realGrad, *ids_); - } - } -} - -REGISTER_LAYER(gather_agent, GatherAgentLayer); -REGISTER_LAYER(scatter_agent, ScatterAgentLayer); - -void ScatterAgentLayer::forwardWithSelection(PassType passType) { - Layer::forward(passType); - CHECK_EQ(realLayer_->getDeviceId(), this->getDeviceId()); - - const Argument& input = realLayer_->getOutput(); - CHECK_EQ(realLayer_->getSize(), this->getSize()); - int width = this->getSize(); - - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceAgentLayerForward", getName().c_str()); - - if (!input.hasSeq()) { - if (realLayer_->getOutput().ids) { - IVector::resizeOrCreate(output_.ids, ids_->getSize(), useGpu_); - output_.ids->selectFrom(*realLayer_->getOutput().ids, *ids_); - } - if (realLayer_->getOutput().value) { - int height = ids_->getSize(); - resetOutput(height, width); - - const MatrixPtr& outV = getOutputValue(); - const MatrixPtr& realV = realLayer_->getOutputValue(); - outV->selectRows(*realV, *ids_); - } - } else { - // Putting the generation logic here is really an ugly hack! - // used in generation - int height = 0; - size_t numSequences = ids_->getSize(); - const int* starts = input.getCpuStartPositions(); - size_t size = input.hasSubseq() ? input.getNumSubSequences() - : input.getNumSequences(); - const int* cpuIds = cpuIds_->getData(); - - for (size_t i = 0; i < numSequences; ++i) { - size_t seqId = cpuIds[i]; - CHECK_LT(seqId, size); - height += starts[seqId + 1] - starts[seqId]; - } - reserveOutput(height, width); - - const MatrixPtr& outputValue = getOutputValue(); - - CHECK_NE(input.sequenceStartPositions.get(), - output_.sequenceStartPositions.get()); - ICpuGpuVector::resizeOrCreate( - output_.sequenceStartPositions, numSequences + 1, false); - int* outStarts = output_.sequenceStartPositions->getMutableData(false); - - ICpuGpuVector::resizeOrCreate(inputStartPos_, height, false); - int* inStarts = inputStartPos_->getMutableData(false); - - size_t offsetOut = 0; - for (size_t i = 0; i < numSequences; ++i) { - outStarts[i] = offsetOut; - size_t seqId = cpuIds[i]; - int size = starts[seqId + 1] - starts[seqId]; - for (int j = 0; j < size; j++) { - inStarts[offsetOut + j] = starts[seqId] + j; - } - offsetOut += size; - } - outStarts[numSequences] = offsetOut; - - outputValue->copyByRowIndex(*input.value, - *inputStartPos_->getVector(useGpu_)); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/AgentLayer.h b/paddle/gserver/layers/AgentLayer.h deleted file mode 100644 index 51f346d5c9fdf9599cddf4b668c128035fd94187..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/AgentLayer.h +++ /dev/null @@ -1,177 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * AgentLayer use as a virtual input of another layer in config, - * before execute forward/backward, setRealLayer() should be - * called to set one and only one real layer - */ -class AgentLayer : public Layer { - protected: - LayerPtr realLayer_; - int numSamples_; - - public: - explicit AgentLayer(const LayerConfig& config) : Layer(config) {} - - ~AgentLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - // if *numSamples* set, - // real layer output will only use first *numSamples* rows - void setRealLayer(LayerPtr layer, int numSamples = 0) { - realLayer_ = layer; - numSamples_ = numSamples; - } - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override {} -}; - -/** - * Like AgentLayer, but it can gather many real layers. Each real - * layer give a few rows of a sequence, after gather all real layers, - * GatherAgentLayer collect a complete sequence. - */ -class GatherAgentLayer : public Layer { - protected: - std::vector realLayers_; - std::vector idsVec_; - // we don't clear idsVec_ vector to aviod IVector alloc/free - IVectorPtr allIds_; - std::vector idIndex_; - - public: - explicit GatherAgentLayer(const LayerConfig& config) : Layer(config) {} - - virtual ~GatherAgentLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - // call before addRealLayer - void clearRealLayers() { realLayers_.clear(); } - - void copyIdAndSequenceInfo(ICpuGpuVectorPtr sequenceStartPositions, - ICpuGpuVectorPtr subSequenceStartPositions, - const IVectorPtr& allIds, - const std::vector& idIndex); - - // add one real layer, can call many times - void addRealLayer(LayerPtr layer) { realLayers_.push_back(layer); } - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override; - void forwardValue(PassType passType); - void forwardIds(PassType passType); -}; - -/** - * Like AgentLayer, but only select a few rows in real layer. - * [idIndex, idIndex + idSize) of *ids* in setRealLayerAndOutput() - * are the selected row ids. It's used to scatter one layer's output - * to many small submodels. ScatterAgentLayer can support ids real layer, - * if it is, the agent will select a few ids in real layer. - */ -class ScatterAgentLayer : public Layer { - protected: - LayerPtr realLayer_; - IVectorPtr ids_; - IVectorPtr cpuIds_; - Argument realOutArg_; - int idIndex_; - int idSize_; - int seqStartPosIndex_; - int numSequences_; // number of sequences in this scatterAgentLayer - bool handleBackward_; - - // use to store expanded cpuStartPositions or subSequenceStartPositions - // of real layer. - ICpuGpuVectorPtr inputStartPos_; - - // true for setRealLayer, false for setRealLayerAndOutput - bool selectionMode_; - - public: - explicit ScatterAgentLayer(const LayerConfig& config) : Layer(config) {} - - virtual ~ScatterAgentLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - /** - * @brief set real layer in generation - * - * @param layer[input] realLayer - * @param ids[input] row id in real layer - * @param copyId[input] whether to copy a cpu version of ids, - * false(default) in ScatterAgentLayer, and - * true in SequenceScatterAgentLayer. - */ - void setRealLayer(LayerPtr layer, const std::vector& ids) { - realLayer_ = layer; - IVector::resizeOrCreate(ids_, ids.size(), useGpu_); - ids_->copyFrom(ids.data(), ids.size()); - if (useGpu_) { - IVector::resizeOrCreate(cpuIds_, ids.size(), false); - cpuIds_->copyFrom(ids.data(), ids.size()); - } else { - cpuIds_ = ids_; - } - selectionMode_ = true; - } - - // set real layer and output, [idIndex, idIndex + idSize) of *ids* - // are selected row for realOutArg in realLayer - void setRealLayerAndOutput(LayerPtr layer, - const Argument& outArg, - const IVectorPtr& ids, - int idIndex, - int idSize, - bool handleBackward) { - realLayer_ = layer; - realOutArg_ = outArg; - ids_ = ids; - idIndex_ = idIndex; - idSize_ = idSize; - handleBackward_ = handleBackward; - selectionMode_ = false; - } - - void setSequenceStartPositions(const ICpuGpuVectorPtr& sequenceStartPositions, - int seqStartPosIndex, - int numSequences) { - realOutArg_.sequenceStartPositions = sequenceStartPositions; - seqStartPosIndex_ = seqStartPosIndex; - numSequences_ = numSequences; - } - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override; - - void forwardWithSelection(PassType passType); -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/AverageLayer.cpp b/paddle/gserver/layers/AverageLayer.cpp deleted file mode 100644 index b3787b1448a272d2879b372d34406aacc6c0bbfb..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/AverageLayer.cpp +++ /dev/null @@ -1,67 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "AverageLayer.h" - -#include "paddle/utils/Logging.h" - -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(average, AverageLayer); - -bool AverageLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - SequencePoolLayer::init(layerMap, parameterMap); - - // average strategy - if (config_.average_strategy() == "average") { - mode_ = kAverage; - } else if (config_.average_strategy() == "sum") { - mode_ = kSum; - } else if (config_.average_strategy() == "squarerootn") { - mode_ = kAverageSquareRootN; - } else { - LOG(FATAL) << "Unknown average strategy: " << config_.average_strategy(); - } - return true; -} - -void AverageLayer::forward(PassType passType) { - SequencePoolLayer::forward(passType); - - MatrixPtr inputValue = getInputValue(0); - getOutputValue()->sequenceAvgForward( - *inputValue, *startPositions_->getVector(useGpu_), mode_); - - /* add the bias-vector AFTER average operation */ - if (biases_.get() != NULL) { - MatrixPtr outV = getOutputValue(); - outV->addBias(*(biases_->getW()), 1); - } - - /* activation */ { forwardActivation(); } -} - -void AverageLayer::backward(const UpdateCallback& callback) { - SequencePoolLayer::backward(callback); - - if (getInputGrad(0)) { - getInputGrad(0)->sequenceAvgBackward( - *getOutputGrad(), *startPositions_->getVector(useGpu_), mode_); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/AverageLayer.h b/paddle/gserver/layers/AverageLayer.h deleted file mode 100644 index 03e2673b55ceca7a698f1b858327ad6fad739087..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/AverageLayer.h +++ /dev/null @@ -1,54 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "SequencePoolLayer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * A layer for "internal average" for sequence input. - * Input: one or more sequences. Each sequence contains some instances. - * If SequenceLevel = kNonSeq: - * Output: output size is the number of input sequences (NOT input instances) - * output[i] = average_{for each instance in this sequence}{input[i]} - * If stride_ > 0: - * Output: a shorten sequence. Stride is the step size by which we slide a - * window upon the input sequence, and the average pooling - * operation is then applied to each interval independently. - * If SequenceLevel = kSeq: - * Check input sequence must has sub-sequence - * Output: output size is the number of input sub-sequences - * output[i] = average_{for each instance in this sub-sequence}{input[i]} - * - * The config file api is pooling_layer. - */ -class AverageLayer : public SequencePoolLayer { - public: - enum AverageStrategy { kAverage = 0, kSum = 1, kAverageSquareRootN = 2 }; - explicit AverageLayer(const LayerConfig& config) - : SequencePoolLayer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - protected: - int mode_; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/BatchNormBaseLayer.cpp b/paddle/gserver/layers/BatchNormBaseLayer.cpp deleted file mode 100644 index a3516f9423e62df0192485c4476357ac51dc27a4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BatchNormBaseLayer.cpp +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "BatchNormBaseLayer.h" -#include "BatchNormalizationLayer.h" -#include "Layer.h" -#include "paddle/utils/Stat.h" -#ifdef PADDLE_WITH_CUDA -#include "CudnnBatchNormLayer.h" -#endif - -namespace paddle { - -bool BatchNormBaseLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - if (!Layer::init(layerMap, parameterMap)) return false; - - /* initialize the weightList */ - // first is Input in configure - // other two is created in config_parser.py - CHECK_EQ(inputLayers_.size(), 3U); - CHECK_EQ(inputLayers_.size(), parameters_.size()); - CHECK_EQ(inputLayers_.size(), size_t(config_.inputs_size())); - const ImageConfig& conf = config_.inputs(0).image_conf(); - channels_ = conf.channels(); - calFeatureMapSize(); - - if (config_.has_use_global_stats()) { - useGlobalStats_ = config_.use_global_stats(); - } - movingAvgFraction_ = config_.moving_average_fraction(); - epsilon_ = config_.epsilon(); - - weight_.reset(new Weight(1, channels_, parameters_[0])); - movingMean_.reset(new Weight(1, channels_, parameters_[1])); - movingVar_.reset(new Weight(1, channels_, parameters_[2])); - - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, channels_, biasParameter_)); - } - - savedMean_ = Matrix::create(1, channels_, false, useGpu_); - savedInvVar_ = Matrix::create(1, channels_, false, useGpu_); - savedMean_->zeroMem(); - savedInvVar_->zeroMem(); - - return true; -} - -void BatchNormBaseLayer::calFeatureMapSize() { - const ImageConfig& conf = config_.inputs(0).image_conf(); - imageH_ = inputLayers_[0]->getOutput().getFrameHeight(); - imageW_ = inputLayers_[0]->getOutput().getFrameWidth(); - imageD_ = inputLayers_[0]->getOutput().getFrameDepth(); - - if (0 == imageD_) imageD_ = conf.img_size_z(); - if (imageH_ == 0 && imageW_ == 0) { - imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - imageW_ = conf.img_size(); - } else { - getOutput().setFrameHeight(imageH_); - getOutput().setFrameWidth(imageW_); - getOutput().setFrameDepth(imageD_); - } - imgPixels_ = imageH_ * imageW_ * imageD_; -} - -} // namespace paddle diff --git a/paddle/gserver/layers/BatchNormBaseLayer.h b/paddle/gserver/layers/BatchNormBaseLayer.h deleted file mode 100644 index 5a446c0843a22adecbaf2ae09fcd526b68865ae2..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BatchNormBaseLayer.h +++ /dev/null @@ -1,101 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief Batch normalization layer use to normalizes the input to across the - * batch. - * - * By default, calculating global mean and variance statistics via a running - * average in the training peroid. Then the pre-calculated global mean and - * variance are used for testing. - * - * Moving mean and variance are located in Parameter object when constructing - * and the calculation will change them. Now we only save global mean and - * variance of one thread in first node for GPU. - * But the calculation in CPU is different, because parameters are shared by - * multiple threads. Here using ShareCpuMatrix with lock to calculate. We - * still save global mean and variance in first node in CPU when multi machine. - * - * [1] S. Ioffe and C. Szegedy, "Batch Normalization: Accelerating Deep Network - * Training by Reducing Internal Covariate Shift." arXiv preprint - * arXiv:1502.03167 (2015). - */ - -class BatchNormBaseLayer : public Layer { - public: - explicit BatchNormBaseLayer(const LayerConfig& config) : Layer(config) {} - - ~BatchNormBaseLayer() {} - - /** - * @brief Create BatchNorm layer by norm_type, including batch_norm and - * cudnn_batch_norm. If do not set norm_type, it will automatically select - * cudnn_batch_norm for GPU and batch_norm for CPU. - */ - static Layer* create(const LayerConfig& config); - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - /** - * @brief Calculate feature map size. Some input uses frameHeight and - * frameWidth to store feature size - */ - void calFeatureMapSize(); - - protected: - /// Batch normalization scale parameter, which is referred to as gamma in - /// in original paper. - std::unique_ptr weight_; - /// Moving average of mean. - std::unique_ptr movingMean_; - /// Moving average of variance. - std::unique_ptr movingVar_; - /// Batch normalization bias parameter, which is referred to as beta in - /// in original paper. - std::unique_ptr biases_; - - /// Save intermediate results computed during the forward pass, - /// these can then be reused to speed up the backward pass. - MatrixPtr savedMean_; - MatrixPtr savedInvVar_; - - /// Height or width of input image feature. - /// Both of them are 1 if the input is fully-connected layer. - int imageD_; - int imageH_; - int imageW_; - /// Height * Width. - int imgPixels_; - /// Feature dimension. If the input layer is conv layer, it is the channels - /// of feature map of the conv layer. If the input layer is fully-connected - /// layer, it is the dimension of fc layer. - int channels_; - // if useGlobalStats_ is true, will use the loaded mean and variance. - // otherwise, calculate mean and variance in this mini-batch. - bool useGlobalStats_; - // use to compute moving mean and variance. - real movingAvgFraction_; - // Epsilon is a small random noise used in batch normalization for stability. - real epsilon_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/BatchNormalizationLayer.cpp b/paddle/gserver/layers/BatchNormalizationLayer.cpp deleted file mode 100644 index 59831dd9049d70198721989b4a515df39e015968..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BatchNormalizationLayer.cpp +++ /dev/null @@ -1,266 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Stat.h" -#ifdef PADDLE_WITH_CUDA -#include "hl_batch_transpose.h" -#endif -#include "BatchNormalizationLayer.h" - -namespace paddle { - -REGISTER_LAYER(batch_norm, BatchNormalizationLayer); - -bool BatchNormalizationLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - if (!BatchNormBaseLayer::init(layerMap, parameterMap)) return false; - - return true; -} - -void BatchNormalizationLayer::calMeanAndStd(const MatrixPtr& mat) { - int numSamples = mat->getHeight(); - Matrix::resizeOrCreate(tmpMat_, numSamples, channels_, false, useGpu_); - savedMean_->zeroMem(); - savedMean_->accumulateColSum(*mat); - savedMean_->mulScalar(1.0 / numSamples); // E[x] - - tmpMat_->assign(*mat); - tmpMat_->square2(); - savedInvVar_->zeroMem(); - savedInvVar_->accumulateColSum(*tmpMat_); - savedInvVar_->mulScalar(1.0 / numSamples); // E[x^2] - savedInvVar_->addSquare(*savedMean_, -1.0); // E[x^2] - E^2[x] - - // Variance may be small negative value - // because of the subtraction operation. - // Here using clipping. - savedInvVar_->downClip(real(0.0)); - - calMovingMeanAndVar(); - - savedInvVar_->subScalar(-epsilon_); - savedInvVar_->sqrt2(*savedInvVar_); -} - -void BatchNormalizationLayer::calMovingMeanAndVar() { - // calculating and saving moving mean and variance - auto& movingMean = movingMean_->getW(); - auto& movingVar = movingVar_->getW(); - // movingMean = movingMean * movingAvgFraction_ - // + savedMean_ * (1 - movingAvgFraction_) - movingMean->add(*savedMean_, movingAvgFraction_, 1.0 - movingAvgFraction_); - // movingVar = movingVar * movingAvgFraction_ - // + savedInvVar_ * (1 - movingAvgFraction_) - movingVar->add(*savedInvVar_, movingAvgFraction_, 1.0 - movingAvgFraction_); -} - -void BatchNormalizationLayer::setMeanAndStd() { - savedMean_->copyFrom(*(movingMean_->getW())); - savedInvVar_->copyFrom(*(movingVar_->getW())); - savedInvVar_->downClip(real(0.0)); - - savedInvVar_->subScalar(-epsilon_); - savedInvVar_->sqrt2(*savedInvVar_); -} - -void BatchNormalizationLayer::expandMat(const MatrixPtr& in, MatrixPtr& out) { - CHECK_EQ(in->getWidth(), static_cast(channels_ * imgPixels_)); - CHECK_EQ(out->getWidth(), static_cast(channels_)); - CHECK(!in->isTransposed()); - CHECK(!out->isTransposed()); - if (imgPixels_ == 1) { - out->assign(*in); - return; - } - size_t batchSize = in->getHeight(); - CHECK_EQ(out->getHeight(), batchSize * imgPixels_); - if (useGpu_) { -#ifndef PADDLE_WITH_CUDA - LOG(FATAL) << "paddle is compiled only for cpu"; -#else - batchTranspose( - in->getData(), out->getData(), imgPixels_, channels_, batchSize); -#endif - } else { - for (size_t i = 0; i < batchSize; i++) { - const MatrixPtr inTmp = - Matrix::create(in->getData() + i * imgPixels_ * channels_, - channels_, - imgPixels_, - false, - useGpu_); - MatrixPtr outTmp = - Matrix::create(out->getData() + i * imgPixels_ * channels_, - imgPixels_, - channels_, - false, - useGpu_); - inTmp->transpose(outTmp, false); - } - } -} - -void BatchNormalizationLayer::shrinkMat(const MatrixPtr& in, MatrixPtr& out) { - CHECK_EQ(in->getWidth(), static_cast(channels_)); - CHECK_EQ(out->getWidth(), static_cast(channels_ * imgPixels_)); - size_t batchSize = out->getHeight(); - CHECK(!in->isTransposed()); - CHECK(!out->isTransposed()); - if (imgPixels_ == 1) { - out->assign(*in); - return; - } - CHECK_EQ(in->getHeight(), static_cast(batchSize * imgPixels_)); - if (useGpu_) { -#ifndef PADDLE_WITH_CUDA - LOG(FATAL) << "paddle is compiled only for cpu"; -#else - batchTranspose( - in->getData(), out->getData(), channels_, imgPixels_, batchSize); -#endif - } else { - for (size_t i = 0; i < batchSize; i++) { - const MatrixPtr inTmp = - Matrix::create(in->getData() + i * channels_ * imgPixels_, - imgPixels_, - channels_, - false, - useGpu_); - MatrixPtr outTmp = - Matrix::create(out->getData() + i * imgPixels_ * channels_, - channels_, - imgPixels_, - useGpu_); - inTmp->transpose(outTmp, false); - } - } -} - -void BatchNormalizationLayer::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = getInputValue(0)->getHeight(); - calFeatureMapSize(); - resetOutput(batchSize, getInputValue(0)->getWidth()); - - // for testing in training peroid. - useGlobalStats_ = (passType == PASS_TEST); - if (passType == PASS_TEST && config_.has_use_global_stats()) { - useGlobalStats_ = config_.use_global_stats(); - } - - Matrix::resizeOrCreate( - expandedIn_, batchSize * imgPixels_, channels_, false, useGpu_); - Matrix::resizeOrCreate( - normIn_, batchSize * imgPixels_, channels_, false, useGpu_); - Matrix::resizeOrCreate( - expandedOut_, batchSize * imgPixels_, channels_, false, useGpu_); - expandMat(getInputValue(0), expandedIn_); - - if (useGlobalStats_) { - if (firstTest_) { - setMeanAndStd(); - firstTest_ = false; - } - } else { - calMeanAndStd(expandedIn_); - firstTest_ = true; - } - - normIn_->assign(*expandedIn_); - normIn_->addBias(*savedMean_, -1); // subtract mean. - normIn_->divRowVector(*savedInvVar_); // divide std. - - expandedOut_->assign(*normIn_); - expandedOut_->mulRowVector(*weight_->getW()); // multiple gamma. - if (biases_) { - expandedOut_->addBias(*(biases_->getW()), 1); // add beta. - } - MatrixPtr out = getOutputValue(); - shrinkMat(expandedOut_, out); - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void BatchNormalizationLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - int batchSize = getInputValue(0)->getHeight(); - - Matrix::resizeOrCreate(meanGrad_, 1, channels_, false, useGpu_); - Matrix::resizeOrCreate(stdGrad_, 1, channels_, false, useGpu_); - - Matrix::resizeOrCreate( - expandedInGrad_, batchSize * imgPixels_, channels_, false, useGpu_); - Matrix::resizeOrCreate( - inGrad_, batchSize, imgPixels_ * channels_, false, useGpu_); - Matrix::resizeOrCreate( - normInGrad_, batchSize * imgPixels_, channels_, false, useGpu_); - Matrix::resizeOrCreate( - expandedOutGrad_, batchSize * imgPixels_, channels_, false, useGpu_); - Matrix::resizeOrCreate( - tmpMat_, batchSize * imgPixels_, channels_, false, useGpu_); - Matrix::resizeOrCreate( - tmpGrad_, batchSize * imgPixels_, channels_, false, useGpu_); - - expandMat(getOutputGrad(), expandedOutGrad_); - - // compute derivatives. - if (biases_ && biases_->getWGrad()) { - REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); - biases_->getWGrad()->collectBias(*expandedOutGrad_, 1); - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - if (weight_->getWGrad()) { - tmpMat_->dotMul(*expandedOutGrad_, *normIn_); - weight_->getWGrad()->collectBias(*tmpMat_, 1); - } - - // compute input gradients. - normInGrad_->assign(*expandedOutGrad_); - normInGrad_->mulRowVector(*(weight_->getW())); // multiple gamma. - // normInGrad * (x - \mu)/ \sqrt(\delta^2) - tmpMat_->dotMul(*normInGrad_, *normIn_); - stdGrad_->zeroMem(); - stdGrad_->collectBias(*tmpMat_, -1.0 / (batchSize * imgPixels_)); - tmpGrad_->assign(*normIn_); - tmpGrad_->mulRowVector(*stdGrad_); - - meanGrad_->zeroMem(); - meanGrad_->collectBias(*normInGrad_, -1.0 / (batchSize * imgPixels_)); - - expandedInGrad_->zeroMem(); - expandedInGrad_->add(*normInGrad_, *tmpGrad_); - expandedInGrad_->addRowVector(*meanGrad_); - expandedInGrad_->divRowVector(*savedInvVar_); - - shrinkMat(expandedInGrad_, inGrad_); - if (getInputGrad(0)) { - getInputGrad(0)->add(*getInputGrad(0), *inGrad_); - } - { - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); - weight_->getParameterPtr()->incUpdate(callback); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/BilinearInterpLayer.cpp b/paddle/gserver/layers/BilinearInterpLayer.cpp deleted file mode 100644 index 9775914596ce3253aada71fbe7197410414fede5..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BilinearInterpLayer.cpp +++ /dev/null @@ -1,107 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "BilinearInterpLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(bilinear_interp, BilinearInterpLayer); - -size_t BilinearInterpLayer::getSize() { - inImgH_ = inputLayers_[0]->getOutput().getFrameHeight(); - inImgW_ = inputLayers_[0]->getOutput().getFrameWidth(); - - const BilinearInterpConfig& conf = config_.inputs(0).bilinear_interp_conf(); - if (inImgH_ == 0) { - inImgH_ = conf.image_conf().img_size_y(); - } - if (inImgW_ == 0) { - inImgW_ = conf.image_conf().img_size(); - } - - outImgH_ = conf.out_size_y(); - outImgW_ = conf.out_size_x(); - numChannels_ = conf.image_conf().channels(); - - CHECK(outImgH_ > 0 && outImgW_ > 0); - CHECK(inImgH_ > 0 && inImgW_ > 0); - CHECK(numChannels_); - - ratioH_ = - (outImgH_ > 1) ? static_cast(inImgH_ - 1) / (outImgH_ - 1) : 0.f; - ratioW_ = - (outImgW_ > 1) ? static_cast(inImgW_ - 1) / (outImgW_ - 1) : 0.f; - - getOutput().setFrameHeight(outImgH_); - getOutput().setFrameWidth(outImgW_); - return outImgH_ * outImgW_ * numChannels_; -} - -bool BilinearInterpLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(1, config_.inputs_size()); - - return true; -} - -void BilinearInterpLayer::forward(PassType passType) { - Layer::forward(passType); - - size_t batchSize = getInput(0).getBatchSize(); - size_t size = getSize(); - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - resetOutput(batchSize, size); - } - - MatrixPtr inV = getInputValue(0); - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwBilinearInterpTimer", getName().c_str()); - outV->bilinearForward(*inV, - inImgH_, - inImgW_, - outImgH_, - outImgW_, - numChannels_, - ratioH_, - ratioW_); - } -} - -void BilinearInterpLayer::backward(const UpdateCallback& callback) { - (void)callback; - - MatrixPtr inputG = getInputGrad(0); - MatrixPtr outG = getOutputGrad(); - { - REGISTER_TIMER_INFO("BwBilinearInterpTimer", getName().c_str()); - if (inputG) { - inputG->bilinearBackward(*outG, - outImgH_, - outImgW_, - inImgH_, - inImgW_, - numChannels_, - ratioH_, - ratioW_); - } - } -} -} // namespace paddle diff --git a/paddle/gserver/layers/BilinearInterpLayer.h b/paddle/gserver/layers/BilinearInterpLayer.h deleted file mode 100644 index 8e08c2e1ce80172f55c93d8242821f683fa1a731..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BilinearInterpLayer.h +++ /dev/null @@ -1,47 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief A layer for bilinear interpolation which is - * used on conv layer output. - * - * @note The config file api is bilinear_interp_layer. - */ -class BilinearInterpLayer : public Layer { - protected: - size_t outImgH_, outImgW_; - size_t inImgH_, inImgW_; - real ratioH_, ratioW_; - size_t numChannels_; - - public: - explicit BilinearInterpLayer(const LayerConfig& config) : Layer(config) {} - - virtual ~BilinearInterpLayer() {} - - size_t getSize(); - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/BlockExpandLayer.cpp b/paddle/gserver/layers/BlockExpandLayer.cpp deleted file mode 100644 index 793d24e884a6f76c2aa897b3d03f3adc3e201265..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BlockExpandLayer.cpp +++ /dev/null @@ -1,121 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "BlockExpandLayer.h" - -#include "paddle/utils/Logging.h" - -namespace paddle { - -REGISTER_LAYER(blockexpand, BlockExpandLayer); - -bool BlockExpandLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(config_.inputs_size(), 1); - const BlockExpandConfig& blockConf = config_.inputs(0).block_expand_conf(); - blockH_ = blockConf.block_y(); - blockW_ = blockConf.block_x(); - strideH_ = blockConf.stride_y(); - strideW_ = blockConf.stride_x(); - paddingH_ = blockConf.padding_y(); - paddingW_ = blockConf.padding_x(); - channels_ = blockConf.channels(); - imgSizeH_ = blockConf.img_size_y(); - imgSizeW_ = blockConf.img_size_x(); - - std::vector strides = {(size_t)strideH_, (size_t)strideW_}; - std::vector paddings = {(size_t)paddingH_, (size_t)paddingW_}; - std::vector blocks = {(size_t)blockH_, (size_t)blockW_}; - createFunction(forward_, - "BlockExpand", - FuncConfig() - .set("strides", strides) - .set("paddings", paddings) - .set("blocks", blocks)); - createFunction(backward_, - "BlockExpandGrad", - FuncConfig() - .set("strides", strides) - .set("paddings", paddings) - .set("blocks", blocks)); - - return true; -} - -size_t BlockExpandLayer::getBlockNum() { - CHECK_EQ(inputLayers_.size(), 1UL); - const BlockExpandConfig& blockConf = config_.inputs(0).block_expand_conf(); - imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight(); - imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth(); - if (imgSizeH_ == 0) { - imgSizeH_ = blockConf.img_size_y(); - } - if (imgSizeW_ == 0) { - imgSizeW_ = blockConf.img_size_x(); - } - size_t tmpH = 2 * paddingH_ + imgSizeH_ - blockH_; - outputH_ = (int)tmpH < 0 ? 1 : 1 + (tmpH + strideH_ - 1) / strideH_; - size_t tmpW = 2 * paddingW_ + imgSizeW_ - blockW_; - outputW_ = (int)tmpW < 0 ? 1 : 1 + (tmpW + strideW_ - 1) / strideW_; - - return outputH_ * outputW_; -} - -void BlockExpandLayer::forward(PassType passType) { - Layer::forward(passType); - - size_t batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - size_t blockNum = getBlockNum(); - size_t blockSize = blockH_ * blockW_ * channels_; - resetOutput(blockNum * batchSize, blockSize); - - // calculate output_.value - inputShape_ = TensorShape({batchSize, channels_, imgSizeH_, imgSizeW_}); - outputShape_ = TensorShape({batchSize, blockNum, blockSize}); - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), inputShape_); - outputs.addArg(*getOutputValue(), outputShape_, ASSIGN_TO); - forward_[0]->calc(inputs, outputs); - - // calculate output_.sequenceStartPositions and output_.cpuSequenceDims - Argument& out = getOutput(); - ICpuGpuVector::resizeOrCreate( - out.sequenceStartPositions, batchSize + 1, false); - IVector::resizeOrCreate(out.cpuSequenceDims, 2 * batchSize, false); - int* start = out.sequenceStartPositions->getMutableData(false); - int* dims = out.cpuSequenceDims->getData(); - for (size_t i = 0; i < batchSize; i++) { - start[i] = i * blockNum; - dims[2 * i] = outputH_; - dims[2 * i + 1] = outputW_; - } - start[batchSize] = batchSize * blockNum; -} - -void BlockExpandLayer::backward(const UpdateCallback& callback) { - /* Calculate the input layers error */ - if (getInputGrad(0)) { - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getOutputGrad(), outputShape_); - outputs.addArg(*getInputGrad(0), inputShape_, ADD_TO); - backward_[0]->calc(inputs, outputs); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/BlockExpandLayer.h b/paddle/gserver/layers/BlockExpandLayer.h deleted file mode 100644 index 9d76584f3a4eda19a9e8f806256a7b8da617cc37..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/BlockExpandLayer.h +++ /dev/null @@ -1,68 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief Expand feature map to minibatch matrix. - * - matrix width is: blockH_ * blockW_ * channels_ - * - matirx height is: outputH_ * outputW_ - * - * \f[ - * outputH\_ = 1 + (2 * paddingH\_ + imgSizeH\_ - blockH\_ + strideH\_ - 1) / - * strideH\_ \\ - * outputW\_ = 1 + (2 * paddingW\_ + imgSizeW\_ - blockW\_ + strideW\_ - 1) / - * strideW\_ - * \f] - * - * The expand method is the same with ExpandConvLayer, but saved the transposed - * value. After expanding, output_.sequenceStartPositions will store timeline. - * The number of time steps are outputH_ * outputW_ and the dimension of each - * time step is blockH_ * blockW_ * channels_. This layer can be used after - * convolution neural network, and before recurrent neural network. - * - * The config file api is block_expand_layer. - */ -class BlockExpandLayer : public Layer { - protected: - /** - * @brief Calculate outputH_ and outputW_ and return block number which - * actually is time steps. - * @return time steps, outoutH_ * outputW_. - */ - size_t getBlockNum(); - size_t blockH_, blockW_, strideH_, strideW_, paddingH_, paddingW_; - size_t imgSizeH_, imgSizeW_, outputH_, outputW_, channels_; - - TensorShape inputShape_; - TensorShape outputShape_; - - public: - explicit BlockExpandLayer(const LayerConfig& config) : Layer(config) {} - - ~BlockExpandLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConcatenateLayer.cpp b/paddle/gserver/layers/ConcatenateLayer.cpp deleted file mode 100644 index e6de329ff3f9ccfdd1cbe697c1de1a9cd8c7926a..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConcatenateLayer.cpp +++ /dev/null @@ -1,208 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "Projection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A concatenate layer has multiple input layers. It concatenates rows of - * each input as one row for the output of this layer and apply activation. - */ -class ConcatenateLayer : public Layer { - public: - explicit ConcatenateLayer(const LayerConfig& config) : Layer(config) {} - - ~ConcatenateLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(concat, ConcatenateLayer); - -bool ConcatenateLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - if (!Layer::init(layerMap, parameterMap)) return false; - - CHECK(!biasParameter_); - - return true; -} - -void ConcatenateLayer::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = getInput(0).getBatchSize(); - int size = getSize(); - reserveOutput(batchSize, size); - - const MatrixPtr& out = getOutputValue(); - int offset = 0; - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - const MatrixPtr& in = getInputValue(i); - size_t inSize = in->getWidth(); - out->assignAtOffset(*in, offset); - offset += inSize; - } - CHECK_EQ(size, offset); - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void ConcatenateLayer::backward(const UpdateCallback& callback) { - (void)callback; - - /* Do activation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - const MatrixPtr& out = getOutputGrad(); - int offset = 0; - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - const MatrixPtr& in = getInputGrad(i); - size_t inSize = getInputValue(i)->getWidth(); - if (in) { - in->addAtOffset(*out, offset); - } - offset += inSize; - } -} - -/** - * concat2 layer is like concat layer, but each input layer was - * processed by a Projection. - */ -class ConcatenateLayer2 : public Layer { - public: - explicit ConcatenateLayer2(const LayerConfig& config) : Layer(config) {} - - ~ConcatenateLayer2() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - protected: - std::vector> projections_; - std::vector projOutput_; - std::vector> projCol_; - bool sharedBias_; - std::unique_ptr biases_; -}; - -REGISTER_LAYER(concat2, ConcatenateLayer2); - -bool ConcatenateLayer2::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - if (!Layer::init(layerMap, parameterMap)) return false; - - CHECK_EQ(inputLayers_.size(), parameters_.size()); - projections_.reserve(inputLayers_.size()); - projCol_.reserve(inputLayers_.size()); - projOutput_.resize(inputLayers_.size()); - - size_t startCol = 0; - size_t endCol = 0; - for (size_t i = 0; i < inputLayers_.size(); i++) { - projections_.emplace_back(Projection::create( - config_.inputs(i).proj_conf(), parameters_[i], useGpu_)); - - endCol += projections_[i]->getOutputSize(); - projCol_.push_back(std::make_pair(startCol, endCol)); - startCol = endCol; - } - CHECK_EQ(getSize(), endCol); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - sharedBias_ = config_.shared_biases(); - size_t psize = config_.bias_size(); - biases_ = std::unique_ptr(new Weight(1, psize, biasParameter_)); - } - - return true; -} - -void ConcatenateLayer2::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = getInput(0).getBatchSize(); - int size = getSize(); - resetOutput(batchSize, size); - - for (size_t i = 0; i < projections_.size(); i++) { - size_t startCol = projCol_[i].first; - size_t endCol = projCol_[i].second; - projOutput_[i].value = output_.value->subColMatrix(startCol, endCol); - if (output_.grad) { - projOutput_[i].grad = output_.grad->subColMatrix(startCol, endCol); - } - } - - { - AsyncGpuBlock block; - for (size_t i = 0; i != inputLayers_.size(); ++i) { - projections_[i]->forward(&getInput(i), &projOutput_[i], passType); - } - } - - /* add the bias-vector */ - if (biases_) { - REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); - output_.value->addBias(*(biases_->getW()), 1, sharedBias_); - } - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void ConcatenateLayer2::backward(const UpdateCallback& callback) { - /* Do activation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - AsyncGpuBlock block; - if (biases_ && biases_->getWGrad()) { - REGISTER_TIMER_INFO("Concat2BpBiasTimer", getName().c_str()); - biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBias_); - biases_->getParameterPtr()->incUpdate(callback); - } - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - if (projections_[i]) { - projections_[i]->backward(callback); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ContextProjection.cpp b/paddle/gserver/layers/ContextProjection.cpp deleted file mode 100644 index 10c3cef0da61af76a6b0a207e4b914276a2fa39b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ContextProjection.cpp +++ /dev/null @@ -1,185 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ContextProjection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_PROJECTION(context, ContextProjection); - -ContextProjection::ContextProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu) - : Projection(config, parameter, useGpu) { - CHECK(config.has_context_start()); - CHECK(config.has_context_length()); - if (config.context_start() == 0 && config.context_length() == 1) { - config_.set_trainable_padding(false); - } - if (config_.trainable_padding()) { - CHECK(parameter); - beginPad_ = std::max(0, -config.context_start()); - endPad_ = std::max(0, config.context_start() + config.context_length() - 1); - size_t totalPad = beginPad_ + endPad_; - size_t inputDim = parameter->getSize() / totalPad; - CHECK_EQ(config.input_size(), inputDim); - CHECK_EQ(inputDim * totalPad, parameter->getSize()); - weight_.reset(new Weight(totalPad, inputDim, parameter)); - } - // init forward_ and backward_ functions - init(); -} - -bool ContextProjection::init() { - size_t context_length = config_.context_length(); - int context_start = config_.context_start(); - bool is_padding = config_.trainable_padding(); - size_t total_pad = is_padding ? beginPad_ + endPad_ : 0; - - createFunction(forward_, - "ContextProjectionForward", - FuncConfig() - .set("context_length", context_length) - .set("context_start", context_start) - .set("begin_pad", beginPad_)); - createFunction(backward_, - "ContextProjectionBackward", - FuncConfig() - .set("context_length", context_length) - .set("context_start", context_start) - .set("begin_pad", beginPad_) - .set("is_padding", is_padding) - .set("total_pad", total_pad)); - - return true; -} - -void ContextProjection::resetState() { - CHECK_LE(config_.context_start() + config_.context_length(), 1) - << "state is not allowed for future context"; - if (config_.context_start() >= 0) return; - Matrix::resizeOrCreate(state_, - -config_.context_start(), - config_.input_size(), - false, // trans - useGpu_); - Matrix::resizeOrCreate(state2_, - -config_.context_start(), - config_.input_size(), - false, // trans - useGpu_); - if (config_.trainable_padding()) { - state_->assign(*weight_->getW()->subMatrix(0, -config_.context_start())); - } else { - state_->zeroMem(); - } -} - -void ContextProjection::setState(LayerStatePtr state) { - CHECK(state->value.size() == 1) - << "one matrix is expected for ContextProjection state"; - state_->copyFrom(*(state->value[0])); -} - -LayerStatePtr ContextProjection::getState() { - if (state_ == nullptr) { - return nullptr; - } - LayerStatePtr res = std::make_shared(); - res->value.push_back(state_->clone(0, 0, false)); - res->value[0]->copyFrom(*state_); - return res; -} - -void ContextProjection::forward() { - CHECK(in_->value && out_->value); - CHECK(in_->sequenceStartPositions); - - size_t input_dim = in_->value->getWidth(); - size_t dim = out_->value->getWidth(); - CHECK_EQ(dim, input_dim * config_.context_length()); - // size_t batch_size = in_->value->getHeight(); - CHECK_EQ(forward_.size(), (size_t)1) << "Only one forward function here"; - - REGISTER_TIMER_INFO("ContextProjectionForward", getName().c_str()); - bool is_padding = config_.trainable_padding(); - /// first use state_, otherwise use weight_(padding false === w nullptr) - auto w_ptr = - state_ ? state_.get() : is_padding ? weight_->getW().get() : nullptr; - const auto start_pos = in_->sequenceStartPositions->getVector(useGpu_); - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*in_->value, *start_pos); - if (w_ptr) { - inputs.addArg(CpuMatrix(w_ptr->getData(), w_ptr->getHeight(), input_dim), - *start_pos); - } - outputs.addArg(*out_->value, *start_pos, ADD_TO); - forward_[0]->calc(inputs, outputs); - - if (state_ && config_.context_start() < 0) { - CHECK_EQ(1, in_->getNumSequences()); - const int* starts = in_->sequenceStartPositions->getData(false); - int length = starts[1] - starts[0]; - if (-config_.context_start() <= length) { - MatrixPtr sub = in_->value->subMatrix(starts[1] + config_.context_start(), - -config_.context_start()); - state_->copyFrom(*sub); - } else { - int prevLength = -config_.context_start() - length; - state2_->subMatrix(0, prevLength) - ->copyFrom(*state_->subMatrix(length, prevLength)); - state2_->subMatrix(prevLength, length) - ->copyFrom(*in_->value->subMatrix(starts[0], length)); - std::swap(state_, state2_); - } - } -} - -void ContextProjection::backward(const UpdateCallback& callback) { - CHECK(in_->value && out_->value && out_->grad); - size_t input_dim = in_->value->getWidth(); - size_t dim = out_->value->getWidth(); - CHECK_EQ(dim, input_dim * config_.context_length()); - size_t batch_size = in_->value->getHeight(); - CHECK_EQ(batch_size, out_->value->getHeight()); - CHECK_EQ(static_cast(backward_.size()), 1) - << "Only one backward function here"; - - REGISTER_TIMER_INFO("ContextProjectionBackward", getName().c_str()); - bool is_padding = config_.trainable_padding(); - auto start_pos = in_->sequenceStartPositions; - auto w_ptr = is_padding ? weight_->getWGrad() : nullptr; - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*out_->grad, *in_->sequenceStartPositions->getVector(useGpu_)); - outputs.addArg( - CpuMatrix( - in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim), - *in_->sequenceStartPositions->getVector(useGpu_), - ADD_TO); - outputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr, - w_ptr ? w_ptr->getHeight() : 0, - input_dim), - ADD_TO); - backward_[0]->calc(inputs, outputs); - - if (config_.trainable_padding()) { - weight_->getParameterPtr()->incUpdate(callback); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/Conv3DLayer.cpp b/paddle/gserver/layers/Conv3DLayer.cpp deleted file mode 100644 index b38de86b1591f987a63478d019019f87c88cee20..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Conv3DLayer.cpp +++ /dev/null @@ -1,253 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Conv3DLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(conv3d, Conv3DLayer); - -bool Conv3DLayer::init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) { - if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; - int index = 0; - for (auto &inputConfig : config_.inputs()) { - const ConvConfig &conf = inputConfig.conv_conf(); - M_.push_back(numFilters_ / conf.groups()); - K_.push_back(filterPixels_[index] * filterChannels_[index]); - - // create a new weight - size_t height, width; - width = filterPixels_[index] * filterChannels_[index]; - height = numFilters_; - CHECK_EQ(parameters_[index]->getSize(), width * height); - Weight *w = new Weight(height, width, parameters_[index]); - weights_.emplace_back(w); - ++index; - } - if (biasParameter_.get()) { - if (sharedBiases_) { - CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); - biases_ = - std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); - } else { - biases_ = - std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); - } - } - return true; -} - -size_t Conv3DLayer::getSize() { - CHECK_NE(inputLayers_.size(), 0UL); - outputH_.clear(); - outputW_.clear(); - outputD_.clear(); - N_.clear(); - size_t layerSize = 0; - for (size_t i = 0; i < inputLayers_.size(); ++i) { - outputW_.push_back(outputSize( - imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true)); - outputH_.push_back(outputSize( - imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true)); - outputD_.push_back(outputSize( - imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true)); - - N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]); - CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize); - layerSize += N_[i] * numFilters_; - } - getOutput().setFrameHeight(outputH_[0]); - getOutput().setFrameWidth(outputW_[0]); - getOutput().setFrameDepth(outputD_[0]); - return layerSize; -} - -void Conv3DLayer::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - int outWidth = getSize(); - resetOutput(batchSize, outWidth); - - REGISTER_TIMER_INFO("FwdConv3D", getName().c_str()); - for (size_t i = 0; i != inputLayers_.size(); ++i) { - const MatrixPtr &inMat = getInputValue(i); - const MatrixPtr &outMat = getOutputValue(); - int M = M_[i]; - int N = N_[i]; - int K = K_[i]; - Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); - MatrixPtr wMat = weights_[i]->getW(); - for (int n = 0; n < batchSize; ++n) { - colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(), - channels_[i], - imgSizeD_[i], - imgSizeH_[i], - imgSizeW_[i], - filterSizeZ_[i], - filterSizeY_[i], - filterSize_[i], - strideZ_[i], - strideY_[i], - stride_[i], - paddingZ_[i], - paddingY_[i], - padding_[i]); - - real *outData = outMat->getData() + n * outMat->getStride(); - MatrixPtr outMatSub = - Matrix::create(outData, groups_[i] * M, N, false, useGpu_); - for (int g = 0; g < groups_[i]; g++) { - MatrixPtr wMatSub = wMat->subMatrix(g * M, M); - MatrixPtr in = colBuf_->subMatrix(g * K, K); - MatrixPtr out = outMatSub->subMatrix(g * M, M); - out->mul(*wMatSub, *in, 1.0, 1.0); - } - } - } - if (nullptr != this->biasParameter_) { - this->addBias(); - } - forwardActivation(); -} - -void Conv3DLayer::backward(const UpdateCallback &callback) { - backwardActivation(); - - if (biases_ && biases_->getWGrad()) { - bpropBiases(); - biases_->getParameterPtr()->incUpdate(callback); - } - - REGISTER_TIMER_INFO("BwdConv3D", getName().c_str()); - for (size_t i = 0; i != inputLayers_.size(); ++i) { - if (weights_[i]->getWGrad()) { - bpropWeights(i); - } - if (getInputGrad(i)) { - bpropData(i); - } - weights_[i]->getParameterPtr()->incUpdate(callback); - } -} - -void Conv3DLayer::bpropWeights(int i) { - int M = M_[i]; - int N = N_[i]; - int K = K_[i]; - const MatrixPtr &inMat = getInputValue(i); - Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); - MatrixPtr wGradMat = weights_[i]->getWGrad(); - int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - for (int n = 0; n < batchSize; ++n) { - colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(), - channels_[i], - imgSizeD_[i], - imgSizeH_[i], - imgSizeW_[i], - filterSizeZ_[i], - filterSizeY_[i], - filterSize_[i], - strideZ_[i], - strideY_[i], - stride_[i], - paddingZ_[i], - paddingY_[i], - padding_[i]); - - real *outGradData = - getOutputGrad()->getData() + n * getOutputGrad()->getStride(); - MatrixPtr outGradSub = - Matrix::create(outGradData, groups_[i] * M, N, false, useGpu_); - for (int g = 0; g < groups_[i]; ++g) { - MatrixPtr inMatSub = colBuf_->subMatrix(g * K, K); - MatrixPtr outG = outGradSub->subMatrix(g * M, M); - MatrixPtr wGradSub = wGradMat->subMatrix(g * M, M); - wGradSub->mul(*outG, *(inMatSub->getTranspose()), 1.0, 1.0); - } - } -} - -void Conv3DLayer::bpropData(int i) { - int M = M_[i]; - int N = N_[i]; - int K = K_[i]; - Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); - MatrixPtr wMat = weights_[i]->getW(); - int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - for (int n = 0; n < batchSize; ++n) { - real *outGradData = - getOutputGrad()->getData() + n * getOutputGrad()->getStride(); - real *preGradData = - getInputGrad(i)->getData() + n * getInputGrad(i)->getStride(); - MatrixPtr outGradSub = - Matrix::create(outGradData, M * groups_[i], N, false, useGpu_); - for (int g = 0; g < groups_[i]; ++g) { - MatrixPtr wMatSub = wMat->subMatrix(g * M, M); - MatrixPtr outG = outGradSub->subMatrix(g * M, M); - MatrixPtr inGradMatSub = colBuf_->subMatrix(g * K, K); - inGradMatSub->mul(*(wMatSub->getTranspose()), *outG, 1.0, 0.0); - } - colBuf_->col2Vol(preGradData, - channels_[i], - imgSizeD_[i], - imgSizeH_[i], - imgSizeW_[i], - filterSizeZ_[i], - filterSizeY_[i], - filterSize_[i], - strideZ_[i], - strideY_[i], - stride_[i], - paddingZ_[i], - paddingY_[i], - padding_[i], - 1.0, - 1.0); - } -} - -void Conv3DLayer::bpropBiases() { - MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), - 1, - biases_->getWGrad()->getElementCnt(), - false, - useGpu_); - MatrixPtr outGradMat = getOutputGrad(); - - if (this->sharedBiases_) { - biases->collectSharedBias(*outGradMat, 1.0f); - } else { - biases->collectBias(*outGradMat, 1.0f); - } -} - -void Conv3DLayer::addBias() { - MatrixPtr outMat = getOutputValue(); - MatrixPtr bias = Matrix::create(biases_->getW()->getData(), - 1, - biases_->getW()->getElementCnt(), - false, - useGpu_); - if (this->sharedBiases_) { - outMat->addSharedBias(*(bias), 1.0f); - } else { - outMat->addBias(*(bias), 1.0f); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/Conv3DLayer.h b/paddle/gserver/layers/Conv3DLayer.h deleted file mode 100644 index 07b804bad02beb6ec9c3e9fd43c3cd3aa6d50b22..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Conv3DLayer.h +++ /dev/null @@ -1,51 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include "ConvBaseLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief A subclass of convolution layer. - * This layer expands input and use matrix multiplication to - * calculate convolution operation. - */ -class Conv3DLayer : public ConvBaseLayer { - public: - explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} - ~Conv3DLayer() {} - - bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); - - void forward(PassType passType); - void addBias(); - void backward(const UpdateCallback& callback); - void bpropBiases(); - void bpropData(int i); - void bpropWeights(int i); - size_t getSize(); - - protected: - // Figure out the dimensions for individual gemms. - IntV M_; /// numFilters_ / filter_group_; - IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ - IntV K_; /// outputD_ * outputH_ * outputW_ - MatrixPtr colBuf_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseLayer.cpp b/paddle/gserver/layers/ConvBaseLayer.cpp deleted file mode 100644 index 56bf4f9fcb187f73409076b826b738f62d19516a..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvBaseLayer.cpp +++ /dev/null @@ -1,120 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvBaseLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/Logging.h" -namespace paddle { - -bool ConvBaseLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - isDeconv_ = (config_.type() == "exconv" || config_.type() == "cudnn_conv") - ? false - : true; - - /* Initialize the convolutional layer parameter */ - numFilters_ = config_.num_filters(); - sharedBiases_ = config_.shared_biases(); - for (auto& inputConfig : config_.inputs()) { - const ConvConfig& conf = inputConfig.conv_conf(); - padding_.push_back(conf.padding()); - stride_.push_back(conf.stride()); - dilation_.push_back(conf.dilation()); - filterSize_.push_back(conf.filter_size()); - paddingY_.push_back(conf.padding_y()); - strideY_.push_back(conf.stride_y()); - dilationY_.push_back(conf.dilation_y()); - filterSizeY_.push_back(conf.filter_size_y()); - channels_.push_back(conf.channels()); - imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y() - : conf.img_size()); - imgSizeW_.push_back(conf.img_size()); - groups_.push_back(conf.groups()); - filterChannels_.push_back(conf.filter_channels()); - outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x()); - outputW_.push_back(conf.output_x()); - - paddingZ_.push_back(conf.padding_z()); - strideZ_.push_back(conf.stride_z()); - filterSizeZ_.push_back(conf.filter_size_z()); - imgSizeD_.push_back(conf.img_size_z()); - outputD_.push_back(conf.output_z()); - filterPixels_.push_back(filterSize_.back() * filterSizeY_.back() * - filterSizeZ_.back()); - } - - CHECK(inputLayers_.size() == parameters_.size()); - - // create new weights_ in derived class - // create new biases_ in derived class - - // default caffe model - caffeMode_ = true; - - return true; -} - -size_t ConvBaseLayer::calOutputSize() { - auto clearAndReserve = [this](IntV* vec) { - vec->clear(); - vec->reserve(this->inputLayers_.size()); - }; - clearAndReserve(&imgSizeH_); - clearAndReserve(&imgSizeW_); - clearAndReserve(&outputH_); - clearAndReserve(&outputW_); - size_t layerSize = 0; - - auto setLayerSize = [&](IntV& inH, IntV& inW, IntV& outH, IntV& outW) { - size_t filterSizeY; - size_t filterSize; - for (size_t i = 0; i < inputLayers_.size(); i++) { - filterSizeY = (filterSizeY_[i] - 1) * dilationY_[i] + 1; - filterSize = (filterSize_[i] - 1) * dilation_[i] + 1; - inH.push_back(inputLayers_[i]->getOutput().getFrameHeight()); - inW.push_back(inputLayers_[i]->getOutput().getFrameWidth()); - const ConvConfig& conf = config_.inputs(i).conv_conf(); - if (isDeconv_) { - if (inH[i] == 0) - inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x(); - if (inW[i] == 0) inW[i] = conf.output_x(); - outH.push_back(imageSize( - inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); - outW.push_back( - imageSize(inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); - } else { - if (inH[i] == 0) - inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - if (inW[i] == 0) inW[i] = conf.img_size(); - outH.push_back(outputSize( - inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); - outW.push_back(outputSize( - inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); - } - CHECK_EQ(outH[i], outH[0]); - CHECK_EQ(outW[i], outW[0]); - } - getOutput().setFrameHeight(outH[0]); - getOutput().setFrameWidth(outW[0]); - layerSize = outH[0] * outW[0] * size_t(numFilters_); - }; - - setLayerSize(imgSizeH_, imgSizeW_, outputH_, outputW_); - - return layerSize; -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseLayer.h b/paddle/gserver/layers/ConvBaseLayer.h deleted file mode 100644 index 801bc4f888c5a60e803c882dcf807678c64af20c..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvBaseLayer.h +++ /dev/null @@ -1,107 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/MathUtils.h" -namespace paddle { - -/** - * @brief A Base Convolution Layer, which convolves the input image - * with learned filters and (optionally) adds biases. - */ - -class ConvBaseLayer : public Layer { - protected: - typedef std::vector IntV; - - /// True if it's deconv layer, false if it's convolution layer - bool isDeconv_; - - /// The number of filters. - int numFilters_; - /// The x dimension of the padding. - IntV padding_; - /// The y dimension of the padding. - IntV paddingY_; - /// The x dimension of the stride. - IntV stride_; - /// The y dimension of the stride. - IntV strideY_; - /// The x dimension of the dilation. - IntV dilation_; - /// The y dimension of the dilation. - IntV dilationY_; - /// The x dimension of a filter kernel. - IntV filterSize_; - /// The y dimension of a filter kernel. - IntV filterSizeY_; - /// The spatial dimensions of the convolution input. - IntV channels_; - /// The spatial dimensions of input feature map height. - IntV imgSizeH_; - /// The spatial dimensions of input feature map width. - IntV imgSizeW_; - /// filterPixels_ = filterSizeX_ * filterSizeY_. - IntV filterPixels_; - /// filterChannels_ = channels_/groups_. - IntV filterChannels_; - /// The spatial dimensions of output feature map height. - IntV outputH_; - /// The spatial dimensions of output feature map width. - IntV outputW_; - - IntV outputD_; - IntV imgSizeD_; - IntV filterSizeZ_; - IntV strideZ_; - IntV paddingZ_; - - /// Group size, refer to grouped convolution in - /// Alex Krizhevsky's paper: when group=2, the first half of the - /// filters are only connected to the first half of the input channels, - /// and the second half only connected to the second half. - IntV groups_; - /// Whether the bias is shared for feature in each channel. - bool sharedBiases_; - - /// shape of weight: (numChannels * filterPixels_, numFilters) - WeightList weights_; - /// If shared_biases is false shape of bias: (numFilters_, 1) - /// If shared_biases is ture shape of bias: - /// (numFilters_ * outputX * outputY, 1) - std::unique_ptr biases_; - - /// True by default. The only difference is the calculation - /// of output size. - bool caffeMode_; - - public: - explicit ConvBaseLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - /** - * imgSizeH_ and imgSizeW_ will be set according to the previous input layers - * in this function. Then it will calculate outputH_ and outputW_ and set them - * into output argument. - */ - virtual size_t calOutputSize(); - - Weight& getWeight(int idx) { return *weights_[idx]; } -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseOperator.cpp b/paddle/gserver/layers/ConvBaseOperator.cpp deleted file mode 100644 index 317e7d5c607683efa1e93aba9bc9ba472d37d60d..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvBaseOperator.cpp +++ /dev/null @@ -1,151 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvBaseOperator.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief ConvBaseOperator takes two inputs to perform the convolution. - * The first input is the image, and the second input is the convolution kernel. - * The height of data for two inputs are the same. Each data of the first input - * is convolved with each data of the second input indepedently. - * - * The config file api is conv_operator. - */ - -ConvBaseOperator::ConvBaseOperator(const OperatorConfig &config, bool useGpu) - : Operator(config, useGpu) { - CHECK(useGpu); - CHECK_EQ(config_.input_indices_size(), 2L); - - caffeMode_ = true; - getConvParams(); - computeConvSizes(); - - // initialize all to default algorithms - fwdAlgo_ = 0; - bwdFilterAlgo_ = 0; - bwdDataAlgo_ = 0; - fwdLimitBytes_ = 0; - bwdDataLimitBytes_ = 0; - bwdFilterLimitBytes_ = 0; - workSpaceInBytes_ = 0; - workSpace_ = nullptr; - - isSelectAlgo_ = false; -} - -void ConvBaseOperator::allocConvWorkSpace() { - hl_conv_workspace(imageDesc_, - outputDesc_, - filterDesc_, - convDesc_, - &fwdAlgo_, - &fwdLimitBytes_, - &bwdDataAlgo_, - &bwdDataLimitBytes_, - &bwdFilterAlgo_, - &bwdFilterLimitBytes_, - /*useDilation*/ false); - - size_t maxWorkSpace = 0; - maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); - maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_); - - if (maxWorkSpace > workSpaceInBytes_) { - if (workSpaceInBytes_ != 0) { - hl_free_mem_device(workSpace_); - } - // total amount of storage needed - workSpace_ = hl_malloc_device(maxWorkSpace); - workSpaceInBytes_ = maxWorkSpace; - } -} - -void ConvBaseOperator::computeConvSizes() { - hl_create_filter_descriptor( - &filterDesc_, channels_, numFilters_, filterSizeY_, filterSize_); - hl_create_tensor_descriptor(&imageDesc_); - hl_create_tensor_descriptor(&outputDesc_); - hl_create_convolution_descriptor(&convDesc_, - imageDesc_, - filterDesc_, - paddingY_, - padding_, - strideY_, - stride_); -} - -void ConvBaseOperator::reshapeImageDescriptors() { - hl_tensor_reshape(imageDesc_, - 1, - channels_, - imageH_, - imageW_, - channels_ * imageH_ * imageW_, - imageH_ * imageW_, - imageW_, - 1); - hl_tensor_reshape(outputDesc_, - 1, - numFilters_, - outputH_, - outputW_, - numFilters_ * outputH_ * outputW_, - outputH_ * outputW_, - outputW_, - 1); - hl_reset_convolution_descriptor(convDesc_, - imageDesc_, - filterDesc_, - paddingY_, - padding_, - strideY_, - stride_); -} - -void ConvBaseOperator::getConvParams() { - configNumFilters_ = config_.num_filters(); - const ConvConfig &conf = config_.conv_conf(); - padding_ = conf.padding(); - stride_ = conf.stride(); - filterSize_ = conf.filter_size(); - paddingY_ = conf.padding_y(); - strideY_ = conf.stride_y(); - filterSizeY_ = conf.filter_size_y(); - filterPixels_ = filterSize_ * filterSizeY_; - configChannels_ = conf.channels(); - imgSize_ = conf.img_size(); - imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - imgPixels_ = imgSize_ * imgSizeY_; - CHECK_EQ(conf.groups(), 1U); - filterChannels_ = conf.filter_channels(); - outputX_ = conf.output_x(); - outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); - outputs_ = outputX_ * outputX_; - - isDeconv_ = (config_.type() == "conv") ? false : true; - if (isDeconv_) { - channels_ = configNumFilters_; - numFilters_ = configChannels_; - } else { - channels_ = configChannels_; - numFilters_ = configNumFilters_; - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseOperator.h b/paddle/gserver/layers/ConvBaseOperator.h deleted file mode 100644 index c3c647cb69da5a70eb5346737cc0092e2201c89e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvBaseOperator.h +++ /dev/null @@ -1,112 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#pragma once - -#include "Operator.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief ConvOperator takes two inputs to perform the convolution. - * The first input is the image, and the second input is the convolution kernel. - * The height of data for two inputs are the same. Each data of the first input - * is convolved with each data of the second input indepedently. - * - * The config file api is conv_operator. - */ - -class ConvBaseOperator : public Operator { - public: - ConvBaseOperator(const OperatorConfig &config, bool useGpu); - /** - * Free workspace in device and destroy cudnn tensor descriptor. - */ - virtual ~ConvBaseOperator() { - if (workSpaceInBytes_ != 0) { - hl_free_mem_device(workSpace_); - workSpaceInBytes_ = 0; - } - - hl_destroy_tensor_descriptor(imageDesc_); - hl_destroy_tensor_descriptor(outputDesc_); - hl_destroy_filter_descriptor(filterDesc_); - hl_destroy_convolution_descriptor(convDesc_); - } - - protected: - /** - * Get convolution parameters from layer config and - * initialize member variables. - */ - void getConvParams(); - - /** - * Allocate Gpu Memory for cudnn convolution algorithms. - */ - void allocConvWorkSpace(); - - /** - * Create cudnn tensor descriptor for convolution operation. - */ - void computeConvSizes(); - - /** - * Reshape cudnn tensor descriptor. - */ - void reshapeImageDescriptors(); - - /** - * Reshape cudnn tensor descriptor. - */ - virtual void reshape(int batchSize) = 0; - - /** - * Check filter size is equal to the size calculated by parameters from - * layer config. - */ - void checkFilterSize(const MatrixPtr &filter) { - CHECK_EQ(static_cast(filter->getWidth()), - filterSize_ * filterSizeY_ * channels_ * numFilters_); - } - - /// Most of member variables are same with CudnnConvLayer. - /// There is no explanation here. - bool isDeconv_; - int imageH_, imageW_, outputH_, outputW_; - hl_tensor_descriptor imageDesc_; - hl_tensor_descriptor outputDesc_; - hl_filter_descriptor filterDesc_; - hl_convolution_descriptor convDesc_; - bool caffeMode_; - int inputOffset_, outputOffset_, weightOffset_; - int numFilters_, channels_; - - /// from parsing config - int configNumFilters_, configChannels_; - int padding_, stride_, filterSize_, imgSize_, imgSizeY_; - int paddingY_, strideY_, filterSizeY_; - int imgPixels_, filterPixels_, filterChannels_, outputX_, outputY_, outputs_; - - /// Following member variables are same with CudnnConvLayer. - /// There is no explanation here. - int fwdAlgo_, bwdFilterAlgo_, bwdDataAlgo_; - size_t fwdLimitBytes_, bwdDataLimitBytes_, bwdFilterLimitBytes_; - size_t workSpaceInBytes_; - void *workSpace_; - bool isSelectAlgo_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseProjection.cpp b/paddle/gserver/layers/ConvBaseProjection.cpp deleted file mode 100644 index 39f433b78fe7ce22cc7f93b87d96ed19c10fc2e9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvBaseProjection.cpp +++ /dev/null @@ -1,199 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvBaseProjection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -ThreadLocalD> ConvBaseProjection::convMem_; - -ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config, - ParameterPtr parameter, - bool useGpu) - : Projection(config, parameter, useGpu) { - CHECK(useGpu); // only support GPU - getConvParams(); - initCudnn(); - - size_t height = filterH_ * filterW_ * channels_ / groups_; - size_t width = numFilters_; - weight_.reset(new Weight(height, width, parameter)); - weightOffset_ = height * width / groups_; -} - -void ConvBaseProjection::getConvParams() { - const ConvConfig &conf = config_.conv_conf(); - paddingH_ = conf.padding_y(); - paddingW_ = conf.padding(); - - strideH_ = conf.stride_y(); - strideW_ = conf.stride(); - - dilationH_ = conf.dilation_y(); - dilationW_ = conf.dilation(); - CHECK_GT(dilationH_, 0); - CHECK_GT(dilationW_, 0); - - filterH_ = conf.filter_size_y(); - filterW_ = conf.filter_size(); - - configImgH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - configImgW_ = conf.img_size(); - - configOutH_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); - configOutW_ = conf.output_x(); - - configChannels_ = conf.channels(); - configNumFilters_ = config_.num_filters(); - - isDeconv_ = (config_.type() == "conv") ? false : true; - - channels_ = (isDeconv_) ? configNumFilters_ : configChannels_; - numFilters_ = (isDeconv_) ? configChannels_ : configNumFilters_; - - groups_ = conf.groups(); - CHECK_EQ(channels_ % groups_, 0); - CHECK_EQ(numFilters_ % groups_, 0); -} - -void ConvBaseProjection::initCudnn() { - hl_create_filter_descriptor(&filterDesc_, - channels_ / groups_, - numFilters_ / groups_, - filterH_, - filterW_); - hl_create_tensor_descriptor(&imageDesc_); - hl_create_tensor_descriptor(&outputDesc_); - hl_create_convolution_descriptor(&convDesc_, - imageDesc_, - filterDesc_, - paddingH_, - paddingW_, - strideH_, - strideW_, - dilationH_, - dilationW_); - - // initialize all to default algorithms - fwdAlgo_ = 0; - bwdFilterAlgo_ = 0; - bwdDataAlgo_ = 0; - fwdLimitBytes_ = 0; - bwdDataLimitBytes_ = 0; - bwdFilterLimitBytes_ = 0; - workSpaceInBytes_ = 0; -} - -void ConvBaseProjection::reshapeTensorDesc(int batchSize) { - // The stride between two consecutive samples in the output of ConvProjection - // may not be numFilters_ * outputH_ * outputW_ (conv) or - // channels_ * imageH_ * imageW_ (deconv) - // for example, in the case of layer ConcatenateLayer2 with two - // ConvProjection, the stride is the output_size of layer ConcatenateLayer2. - // So the calculation of nStride is different from CudnnConvLayer. - size_t nStrideImage, nStrideOutput; - if (isDeconv_) { - nStrideImage = out_->value->getStride(); - nStrideOutput = numFilters_ * outputH_ * outputW_; - } else { - nStrideImage = channels_ * imageH_ * imageW_; - nStrideOutput = out_->value->getStride(); - } - - hl_tensor_reshape(imageDesc_, - batchSize, - channels_ / groups_, - imageH_, - imageW_, - nStrideImage, - imageH_ * imageW_, - imageW_, - 1); - - hl_tensor_reshape(outputDesc_, - batchSize, - numFilters_ / groups_, - outputH_, - outputW_, - nStrideOutput, - outputH_ * outputW_, - outputW_, - 1); - - hl_reset_convolution_descriptor(convDesc_, - imageDesc_, - filterDesc_, - paddingH_, - paddingW_, - strideH_, - strideW_, - dilationH_, - dilationW_); -} - -void ConvBaseProjection::reshape(int batchSize) { - size_t width = calOutputSize(); - CHECK_EQ(width, out_->value->getWidth()); - CHECK_EQ(calInputSize(), in_->value->getWidth()); - - reshapeTensorDesc(batchSize); - bool useDilation = false; - if (dilationH_ > 1 || dilationW_ > 1) { - useDilation = true; - } - hl_conv_workspace(imageDesc_, - outputDesc_, - filterDesc_, - convDesc_, - &fwdAlgo_, - &fwdLimitBytes_, - &bwdDataAlgo_, - &bwdDataLimitBytes_, - &bwdFilterAlgo_, - &bwdFilterLimitBytes_, - useDilation); - - size_t maxWorkSpace = 0; - maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); - maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_); - workSpaceInBytes_ = maxWorkSpace; - - VLOG(3) << getName() << " Fwd / BwdData / BwdFilter algo: " << fwdAlgo_ - << " / " << bwdDataAlgo_ << " / " << bwdFilterAlgo_; -} - -void *ConvBaseProjection::getSpaceBytes(size_t size) { - std::vector &convMem = *convMem_; - if (convMem.empty()) { - int numDevices = hl_get_device_count(); - convMem.resize(numDevices); - } - - int devId = hl_get_device(); - MemoryHandlePtr localMem = convMem[devId]; - if (NULL == localMem || size > localMem->getAllocSize()) { - localMem = std::make_shared(size); - } - return localMem->getBuf(); -} - -ConvBaseProjection::~ConvBaseProjection() { - hl_destroy_tensor_descriptor(imageDesc_); - hl_destroy_tensor_descriptor(outputDesc_); - hl_destroy_filter_descriptor(filterDesc_); - hl_destroy_convolution_descriptor(convDesc_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseProjection.h b/paddle/gserver/layers/ConvBaseProjection.h deleted file mode 100644 index f3266ae1ab945042cde9f24b7c2673c18d37bc11..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvBaseProjection.h +++ /dev/null @@ -1,111 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Projection.h" -#include "paddle/math/MathUtils.h" - -namespace paddle { - -/** - * @brief Base class for ConvProjection and ConvTransProjection. - */ -class ConvBaseProjection : public Projection { - public: - /** - * Constructor. - */ - ConvBaseProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu); - - ~ConvBaseProjection(); - - protected: - void getConvParams(); - void initCudnn(); - - void reshapeTensorDesc(int batchSize); - void reshape(int batchSize); - - virtual size_t calOutputSize() = 0; - virtual size_t calInputSize() = 0; - - static void* getSpaceBytes(size_t size); - - /// True if it's deconv projection layer, false if it's ConvProjection layer - bool isDeconv_; - /// imageH_ and imageW_ / outputH_ and outputW_ - /// is calculated from the input layer. - int imageH_, imageW_; - int outputH_, outputW_; - /// configImgH_ and configImgW_ / configOutH_ and configOutW_ - /// is obtained from config. - int configImgH_, configImgW_; - int configOutH_, configOutW_; - /// channels_ and numFilters_ are defined in terms of convolution semantics - int channels_, numFilters_; - /// configChannels and configNumFilters_ are obtained from config - /// For Conv they are the same as channels_ and numFilters - /// For ConvTrans they are opposite to channels_ and numFilters - int configChannels_, configNumFilters_; - int paddingH_, paddingW_; - int strideH_, strideW_; - int dilationH_, dilationW_; - int filterH_, filterW_; - /// One group offset of input data. - int inputOffset_; - /// One group offset of output data. - int outputOffset_; - /// One group offset of weight. - int weightOffset_; - int groups_; - - /// Cudnn tensor descriptor for input. - hl_tensor_descriptor imageDesc_; - /// Cudnn tensor descriptor for output. - hl_tensor_descriptor outputDesc_; - /// Cudnn tensor descriptor for filter. - hl_filter_descriptor filterDesc_; - /// Cudnn tensor descriptor for a convolution operation. - hl_convolution_descriptor convDesc_; - - /// Record the algorithm for forward convolution, which is obtained by cudnn - /// api to search the best suited algorithm. - int fwdAlgo_; - /// Record the algorithm for computing convolution gradient with respect to - /// filter coefficients. - int bwdFilterAlgo_; - /// Record the algorithm for computing convolution gradient with respect to - /// the output. - int bwdDataAlgo_; - /// Amount of GPU memory needed as workspace to be able to execute a - /// forward convolution with the specified algo. - size_t fwdLimitBytes_; - /// Amount of GPU memory needed as workspace to be able to execute a - /// backwardFilter with the specified algo. - size_t bwdDataLimitBytes_; - /// Amount of GPU memory needed as workspace to be able to execute a - /// backwardData with the specified algo. - size_t bwdFilterLimitBytes_; - /// Size of total work space. - size_t workSpaceInBytes_; - bool bias_; - - std::unique_ptr weight_; - static ThreadLocalD> convMem_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvOperator.cpp b/paddle/gserver/layers/ConvOperator.cpp deleted file mode 100644 index 45498b92d32e0fa72adbe95a98e8d30c7f8929e2..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvOperator.cpp +++ /dev/null @@ -1,128 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvOperator.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief ConvOperator takes two inputs to perform the convolution. - * The first input is the image, and the second input is the convolution kernel. - * The height of data for two inputs are the same. Each data of the first input - * is convolved with each data of the second input indepedently. - * - * The config file api is conv_operator. - */ - -REGISTER_OPERATOR(conv, ConvOperator); - -void ConvOperator::reshape(int batchSize) { - imageH_ = ins_[0]->getFrameHeight(); - imageW_ = ins_[0]->getFrameWidth(); - if (imageH_ == 0) imageH_ = imgSizeY_; - if (imageW_ == 0) imageW_ = imgSize_; - outputH_ = outputSize(imageH_, filterSizeY_, paddingY_, strideY_, caffeMode_); - outputW_ = outputSize(imageW_, filterSize_, padding_, stride_, caffeMode_); - /// Check that the outputSizes are consistent with config - CHECK_EQ(outputH_, outputY_); - CHECK_EQ(outputW_, outputX_); - out_->setFrameHeight(outputH_); - out_->setFrameWidth(outputW_); - - reshapeImageDescriptors(); - - inputOffset_ = channels_ * imageH_ * imageW_; - outputOffset_ = numFilters_ * outputH_ * outputW_; - weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSizeY_; - - if (!isSelectAlgo_) { - allocConvWorkSpace(); - } - - isSelectAlgo_ = true; -} - -void ConvOperator::forward() { - size_t batchSize = ins_[0]->value->getHeight(); - reshape(batchSize); - CHECK_EQ(ins_[1]->value->getHeight(), batchSize); - checkFilterSize(ins_[1]->value); - Matrix::resizeOrCreate(out_->value, - batchSize, - outputH_ * outputW_ * numFilters_, - false, - useGpu_); - { - AsyncGpuBlock block; - for (size_t batchId = 0; batchId < batchSize; ++batchId) { - real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; - real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; - real *outData = out_->value->getData() + outputOffset_ * batchId; - hl_convolution_forward(imageDesc_, - inputData, - outputDesc_, - outData, - filterDesc_, - wgtData, - convDesc_, - workSpace_, - workSpaceInBytes_, - fwdAlgo_); - } - } -} - -void ConvOperator::backward() { - size_t batchSize = ins_[0]->value->getHeight(); - { - AsyncGpuBlock block; - for (size_t batchId = 0; batchId < batchSize; ++batchId) { - real *outGrad = out_->grad->getData() + outputOffset_ * batchId; - if (ins_[1]->grad) { - real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; - real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId; - hl_convolution_backward_filter(imageDesc_, - inputData, - outputDesc_, - outGrad, - filterDesc_, - weightGrad, - convDesc_, - workSpace_, - workSpaceInBytes_, - bwdFilterAlgo_); - } - - MatrixPtr preGrad = ins_[0]->grad; - if (NULL != preGrad) { - real *inputGrad = preGrad->getData() + inputOffset_ * batchId; - real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; - hl_convolution_backward_data(imageDesc_, - inputGrad, - outputDesc_, - outGrad, - filterDesc_, - wgtData, - convDesc_, - workSpace_, - workSpaceInBytes_, - bwdDataAlgo_); - } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvOperator.h b/paddle/gserver/layers/ConvOperator.h deleted file mode 100644 index 527dbf8c270f35e19ca23acd8a3ba8197d03b988..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvOperator.h +++ /dev/null @@ -1,44 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#pragma once - -#include "ConvBaseOperator.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief ConvOperator takes two inputs to perform the convolution. - * The first input is the image, and the second input is the convolution kernel. - * The height of data for two inputs are the same. Each data of the first input - * is convolved with each data of the second input indepedently. - * - * The config file api is conv_operator. - */ - -class ConvOperator : public ConvBaseOperator { - public: - ConvOperator(const OperatorConfig &config, bool useGpu) - : ConvBaseOperator(config, useGpu) {} - /** - * Free workspace in device and destroy cudnn tensor descriptor. - */ - virtual ~ConvOperator() {} - void forward() override; - void backward() override; - void reshape(int batchSize) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvProjection.cpp b/paddle/gserver/layers/ConvProjection.cpp deleted file mode 100644 index f382e6cab12a833ce555c948f41e1086093bd78e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvProjection.cpp +++ /dev/null @@ -1,123 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvProjection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_PROJECTION(conv, ConvProjection); - -size_t ConvProjection::calOutputSize() { - imageH_ = in_->getFrameHeight(); - imageW_ = in_->getFrameWidth(); - if (imageH_ == 0) imageH_ = configImgH_; - if (imageW_ == 0) imageW_ = configImgW_; - outputH_ = outputSize(imageH_, - (filterH_ - 1) * dilationH_ + 1, - paddingH_, - strideH_, - /* caffeMode */ true); - outputW_ = outputSize(imageW_, - (filterW_ - 1) * dilationW_ + 1, - paddingW_, - strideW_, - /* caffeMode */ true); - - const_cast(out_)->setFrameHeight(outputH_); - const_cast(out_)->setFrameWidth(outputW_); - - inputOffset_ = (configChannels_ / groups_) * imageH_ * imageW_; - outputOffset_ = (configNumFilters_ / groups_) * outputH_ * outputW_; - return outputH_ * outputW_ * configNumFilters_; -} - -size_t ConvProjection::calInputSize() { - return static_cast(configChannels_ * imageH_ * imageW_); -} - -void ConvProjection::forward() { - int batchSize = in_->value->getHeight(); - reshape(batchSize); - - void *workSpace = NULL; - if (workSpaceInBytes_ > 0) { - workSpace = getSpaceBytes(workSpaceInBytes_); - } - - for (int g = 0; g < groups_; ++g) { - REGISTER_TIMER_INFO("CudnnConvFwTimer", getName().c_str()); - - real *inputData = in_->value->getData() + g * inputOffset_; - real *wgtData = weight_->getW()->getData() + g * weightOffset_; - real *outData = out_->value->getData() + g * outputOffset_; - hl_convolution_forward(imageDesc_, - inputData, - outputDesc_, - outData, - filterDesc_, - wgtData, - convDesc_, - workSpace, - fwdLimitBytes_, - fwdAlgo_); - } -} - -void ConvProjection::backward(const UpdateCallback &callback) { - REGISTER_TIMER_INFO("CudnnConvBpTimer", getName().c_str()); - - void *workSpace = NULL; - if (workSpaceInBytes_ > 0) { - workSpace = getSpaceBytes(workSpaceInBytes_); - } - - for (int g = 0; g < groups_; ++g) { - real *outGrad = out_->grad->getData() + g * outputOffset_; - if (weight_->getWGrad()) { - real *inputData = in_->value->getData() + g * inputOffset_; - real *weightGrad = weight_->getWGrad()->getData() + g * weightOffset_; - hl_convolution_backward_filter(imageDesc_, - inputData, - outputDesc_, - outGrad, - filterDesc_, - weightGrad, - convDesc_, - workSpace, - bwdFilterLimitBytes_, - bwdFilterAlgo_); - } - - MatrixPtr preGrad = in_->grad; - if (NULL != preGrad) { - real *inputGrad = preGrad->getData() + g * inputOffset_; - real *wgtData = weight_->getW()->getData() + g * weightOffset_; - hl_convolution_backward_data(imageDesc_, - inputGrad, - outputDesc_, - outGrad, - filterDesc_, - wgtData, - convDesc_, - workSpace, - bwdDataLimitBytes_, - bwdDataAlgo_); - } - } - - weight_->getParameterPtr()->incUpdate(callback); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvProjection.h b/paddle/gserver/layers/ConvProjection.h deleted file mode 100644 index 22a2202bb6cc256a4a5897724d8eb8a93fefb79f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvProjection.h +++ /dev/null @@ -1,43 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ConvBaseProjection.h" -#include "paddle/math/MathUtils.h" - -namespace paddle { - -/** - * @brief Convolution projection do the same calculation with CudnnConvLayer. - */ -class ConvProjection : public ConvBaseProjection { - public: - /** - * Constructor. - */ - ConvProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu) - : ConvBaseProjection(config, parameter, useGpu) {} - - ~ConvProjection() {} - - virtual void forward(); - virtual void backward(const UpdateCallback& callback); - virtual size_t calOutputSize(); - virtual size_t calInputSize(); -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvShiftLayer.cpp b/paddle/gserver/layers/ConvShiftLayer.cpp deleted file mode 100644 index 615c3478061b591ea30cbf0b3d27ef2551c0dd28..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvShiftLayer.cpp +++ /dev/null @@ -1,108 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief A layer for circular convluation of two vectors, - * which is used in NEURAL TURING MACHINE. - * - Input: two vectors, the first is data (batchSize x dataDim) - * the second is shift weights (batchSize x shiftDim) - * - Output: a vector (batchSize x dataDim) - * Assumed that: - * - a[in]: contains M elements. - * - b[in]: contains N elements (N should be odd). - * - c[out]: contains M elements. - * - * \f[ - * c[i] = \sum_{j=-(N-1)/2}^{(N-1)/2}a_{i+j} * b_{j} - * \f] - * - * In this formula: - * - a's index is computed modulo M. - * - b's index is comupted modulo N. - * - * The config file api is conv_shift_layer. - */ - -class ConvShiftLayer : public Layer { - public: - explicit ConvShiftLayer(const LayerConfig& config) : Layer(config) {} - - ~ConvShiftLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(conv_shift, ConvShiftLayer); - -bool ConvShiftLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2U); - - return true; -} - -void ConvShiftLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV0->getHeight(); - size_t dataDim = inV0->getWidth(); - - CHECK_EQ(batchSize, inV1->getHeight()); - CHECK_EQ(dataDim, getSize()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - resetOutput(batchSize, dataDim); - } - - MatrixPtr outV = getOutputValue(); - - REGISTER_TIMER_INFO("FwConvShiftTimer", getName().c_str()); - outV->circularConv(*inV0, *inV1); -} - -void ConvShiftLayer::backward(const UpdateCallback& callback) { - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr outG = getOutputGrad(); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - - REGISTER_TIMER_INFO("BwConvShiftTimer", getName().c_str()); - - if (inG0 && inG1) { - outG->circularConvDerivative(*outG, *inV0, *inV1, *inG0, *inG1); - } else { - CHECK(!inG0 || !inG1) << "Not supported"; - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvTransOperator.cpp b/paddle/gserver/layers/ConvTransOperator.cpp deleted file mode 100644 index ac41d6f9a4f86364930e27ee401406432e731b65..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvTransOperator.cpp +++ /dev/null @@ -1,125 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvTransOperator.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief ConvTransOperator takes two inputs to perform the convolution. - * The first input is the image, and the second input is the convolution kernel. - * The height of data for two inputs are the same. Each data of the first input - * is convolved with each data of the second input indepedently. - * - * The config file api is conv_operator. - */ - -REGISTER_OPERATOR(convt, ConvTransOperator); - -void ConvTransOperator::reshape(int batchSize) { - outputH_ = ins_[0]->getFrameHeight(); - outputW_ = ins_[0]->getFrameWidth(); - if (outputH_ == 0) outputH_ = outputY_; - if (outputW_ == 0) outputW_ = outputX_; - imageH_ = imageSize(outputH_, filterSizeY_, paddingY_, strideY_, caffeMode_); - imageW_ = imageSize(outputW_, filterSize_, padding_, stride_, caffeMode_); - /// Check that the imageSizes are consistent with config - CHECK_EQ(imageH_, imgSizeY_); - CHECK_EQ(imageW_, imgSize_); - out_->setFrameHeight(imageH_); - out_->setFrameWidth(imageW_); - - reshapeImageDescriptors(); - - inputOffset_ = numFilters_ * outputH_ * outputW_; - outputOffset_ = channels_ * imageH_ * imageW_; - weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSizeY_; - - if (!isSelectAlgo_) { - allocConvWorkSpace(); - } - - isSelectAlgo_ = true; -} - -void ConvTransOperator::forward() { - size_t batchSize = ins_[0]->value->getHeight(); - reshape(batchSize); - CHECK_EQ(ins_[1]->value->getHeight(), batchSize); - checkFilterSize(ins_[1]->value); - Matrix::resizeOrCreate( - out_->value, batchSize, imageH_ * imageW_ * channels_, false, useGpu_); - { - AsyncGpuBlock block; - for (size_t batchId = 0; batchId < batchSize; ++batchId) { - real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; - real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; - real *outData = out_->value->getData() + outputOffset_ * batchId; - hl_convolution_backward_data(imageDesc_, - outData, - outputDesc_, - inputData, - filterDesc_, - wgtData, - convDesc_, - workSpace_, - workSpaceInBytes_, - bwdDataAlgo_); - } - } -} - -void ConvTransOperator::backward() { - size_t batchSize = ins_[0]->value->getHeight(); - { - AsyncGpuBlock block; - for (size_t batchId = 0; batchId < batchSize; ++batchId) { - real *outGrad = out_->grad->getData() + outputOffset_ * batchId; - if (ins_[1]->grad) { - real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; - real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId; - hl_convolution_backward_filter(imageDesc_, - outGrad, - outputDesc_, - inputData, - filterDesc_, - weightGrad, - convDesc_, - workSpace_, - workSpaceInBytes_, - bwdFilterAlgo_); - } - - MatrixPtr preGrad = ins_[0]->grad; - if (NULL != preGrad) { - real *inputGrad = preGrad->getData() + inputOffset_ * batchId; - real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; - hl_convolution_forward(imageDesc_, - outGrad, - outputDesc_, - inputGrad, - filterDesc_, - wgtData, - convDesc_, - workSpace_, - workSpaceInBytes_, - fwdAlgo_); - } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvTransOperator.h b/paddle/gserver/layers/ConvTransOperator.h deleted file mode 100644 index 53cb7a21b49189898d09aa20cd46d04cc5c20198..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvTransOperator.h +++ /dev/null @@ -1,44 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#pragma once - -#include "ConvBaseOperator.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief ConvTransOperator takes two inputs to perform the convolution. - * The first input is the image, and the second input is the convolution kernel. - * The height of data for two inputs are the same. Each data of the first input - * is convolved with each data of the second input indepedently. - * - * The config file api is conv_operator. - */ - -class ConvTransOperator : public ConvBaseOperator { - public: - ConvTransOperator(const OperatorConfig &config, bool useGpu) - : ConvBaseOperator(config, useGpu) {} - /** - * Free workspace in device and destroy cudnn tensor descriptor. - */ - virtual ~ConvTransOperator() {} - void forward() override; - void backward() override; - void reshape(int batchSize) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvTransProjection.cpp b/paddle/gserver/layers/ConvTransProjection.cpp deleted file mode 100644 index 242ce34a607057069a4d0a31e9b70d56279d37ab..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvTransProjection.cpp +++ /dev/null @@ -1,123 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ConvTransProjection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_PROJECTION(convt, ConvTransProjection); -size_t ConvTransProjection::calOutputSize() { - outputH_ = in_->getFrameHeight(); - outputW_ = in_->getFrameWidth(); - if (outputH_ == 0) outputH_ = configOutH_; - if (outputW_ == 0) outputW_ = configOutW_; - imageH_ = imageSize(outputH_, - (filterH_ - 1) * dilationH_ + 1, - paddingH_, - strideH_, - /* caffeMode */ true); - - imageW_ = imageSize(outputW_, - (filterW_ - 1) * dilationW_ + 1, - paddingW_, - strideW_, - /* caffeMode */ true); - - const_cast(out_)->setFrameHeight(imageH_); - const_cast(out_)->setFrameWidth(imageW_); - - inputOffset_ = (configChannels_ / groups_) * outputH_ * outputW_; - outputOffset_ = (configNumFilters_ / groups_) * imageH_ * imageW_; - return imageH_ * imageW_ * configNumFilters_; -} - -size_t ConvTransProjection::calInputSize() { - return static_cast(configChannels_ * outputH_ * outputW_); -} - -void ConvTransProjection::forward() { - int batchSize = in_->value->getHeight(); - reshape(batchSize); - - void *workSpace = NULL; - if (workSpaceInBytes_ > 0) { - workSpace = getSpaceBytes(workSpaceInBytes_); - } - - for (int g = 0; g < groups_; ++g) { - REGISTER_TIMER_INFO("CudnnConvTransFwTimer", getName().c_str()); - - real *inData = in_->value->getData() + g * inputOffset_; - real *wgtData = weight_->getW()->getData() + g * weightOffset_; - real *outData = out_->value->getData() + g * outputOffset_; - hl_convolution_backward_data(imageDesc_, - outData, - outputDesc_, - inData, - filterDesc_, - wgtData, - convDesc_, - workSpace, - bwdDataLimitBytes_, - bwdDataAlgo_); - } -} - -void ConvTransProjection::backward(const UpdateCallback &callback) { - REGISTER_TIMER_INFO("CudnnConvTransBpTimer", getName().c_str()); - - void *workSpace = NULL; - if (workSpaceInBytes_ > 0) { - workSpace = getSpaceBytes(workSpaceInBytes_); - } - - for (int g = 0; g < groups_; ++g) { - real *outGrad = out_->grad->getData() + g * outputOffset_; - if (weight_->getWGrad()) { - real *inData = in_->value->getData() + g * inputOffset_; - real *weightGrad = weight_->getWGrad()->getData() + g * weightOffset_; - hl_convolution_backward_filter(imageDesc_, - outGrad, - outputDesc_, - inData, - filterDesc_, - weightGrad, - convDesc_, - workSpace, - bwdFilterLimitBytes_, - bwdFilterAlgo_); - } - - MatrixPtr preGrad = in_->grad; - if (NULL != preGrad) { - real *inGrad = preGrad->getData() + g * inputOffset_; - real *wgtData = weight_->getW()->getData() + g * weightOffset_; - hl_convolution_forward(imageDesc_, - outGrad, - outputDesc_, - inGrad, - filterDesc_, - wgtData, - convDesc_, - workSpace, - fwdLimitBytes_, - fwdAlgo_); - } - } - - weight_->getParameterPtr()->incUpdate(callback); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvTransProjection.h b/paddle/gserver/layers/ConvTransProjection.h deleted file mode 100644 index 0f9ed720d3b8855a3a24ac25a1c3917c4b98e81d..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvTransProjection.h +++ /dev/null @@ -1,43 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ConvBaseProjection.h" -#include "paddle/math/MathUtils.h" - -namespace paddle { - -/** - * @brief Convolution projection do the same calculation with CudnnConvLayer. - */ -class ConvTransProjection : public ConvBaseProjection { - public: - /** - * Constructor. - */ - ConvTransProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu) - : ConvBaseProjection(config, parameter, useGpu) {} - - ~ConvTransProjection() {} - - virtual void forward(); - virtual void backward(const UpdateCallback& callback); - virtual size_t calOutputSize(); - virtual size_t calInputSize(); -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ConvexCombinationLayer.cpp b/paddle/gserver/layers/ConvexCombinationLayer.cpp deleted file mode 100644 index 31363d97c4fd318ec2c6d48f9200f6ba1f49ba11..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ConvexCombinationLayer.cpp +++ /dev/null @@ -1,155 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief A layer for weighted sum of vectors, - * which is used in NEURAL MACHINE TRANSLATION BY JOINTLY LEARNING TO ALIGN AND - * TRANSLATE - * - Input: the the size of the first input is weightDim, - * and the size of the second input is weightdim * dataDim. - * - Output: the sizeof the output is dataDim - * \f[ - * out(j) = \sum_{i}(in0(i) * in1(i,j + i * dataDim)), - * i = 0,1,...,(weightDim-1); j = 0, 1,...,(dataDim-1) - * \f] - * Note that the above computation is for one sample. Multiple samples are - * processed in one batch. - * - * The config file api is linear_comb_layer. - */ -class ConvexCombinationLayer : public Layer { - protected: - /// A matrix pointer pointing to second input. - MatrixPtr tmpMtx0; - /// A matrix pointer pointing to first input. - MatrixPtr tmpRow0; - /// A matrix pointer pointing to output. - MatrixPtr tmpRow1; - - public: - explicit ConvexCombinationLayer(const LayerConfig& config) : Layer(config) {} - - ~ConvexCombinationLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(convex_comb, ConvexCombinationLayer); - -bool ConvexCombinationLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(2U, inputLayers_.size()); - size_t dataDim = getSize(); - size_t weightDim = inputLayers_[0]->getSize(); - - CHECK_EQ(weightDim * dataDim, inputLayers_[1]->getSize()) - << "Dimension mismatch"; - - tmpRow0 = Matrix::create(nullptr, - /* height= */ 1, - weightDim, - /* trans= */ false, - useGpu_); - tmpRow1 = Matrix::create(nullptr, - /* height= */ 1, - dataDim, - /* trans= */ false, - useGpu_); - tmpMtx0 = Matrix::create(nullptr, - /* height= */ weightDim, - dataDim, - /* trans= */ false, - useGpu_); - - return true; -} - -void ConvexCombinationLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV0->getHeight(); - size_t weightDim = inV0->getWidth(); - size_t dataDim = getSize(); - - CHECK_EQ(batchSize, inV1->getHeight()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, dataDim); - } - - MatrixPtr outV = getOutputValue(); - - REGISTER_TIMER_INFO("FwCvxCombTimer", getName().c_str()); - for (size_t i = 0; i < batchSize; i++) { - tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim); - tmpRow0->setData(inV0->getData() + i * weightDim); - tmpRow1->setData(outV->getData() + i * dataDim); - - tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 0); - } -} - -void ConvexCombinationLayer::backward(const UpdateCallback& callback) { - MatrixPtr outG = getOutputGrad(); - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - - size_t batchSize = inV0->getHeight(); - size_t weightDim = inV0->getWidth(); - size_t dataDim = getSize(); - - REGISTER_TIMER_INFO("BwCvxCombTimer", getName().c_str()); - - if (inG0) { - for (size_t i = 0; i < batchSize; i++) { - tmpRow0->setData(inG0->getData() + i * weightDim); - tmpRow1->setData(outG->getData() + i * dataDim); - tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim); - - tmpRow0->mul(*tmpRow1, *(tmpMtx0->getTranspose()), 1, 1); - } - } - - if (inG1) { - for (size_t i = 0; i < batchSize; i++) { - tmpRow0->setData(inV0->getData() + i * weightDim); - tmpRow1->setData(outG->getData() + i * dataDim); - tmpMtx0->setData(inG1->getData() + i * weightDim * dataDim); - - tmpMtx0->mul(*(tmpRow0->getTranspose()), *tmpRow1, 1, 1); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/CosSimLayer.cpp b/paddle/gserver/layers/CosSimLayer.cpp deleted file mode 100644 index 4e44a5e8dfdad98bff0cd0f405b4227340a45728..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CosSimLayer.cpp +++ /dev/null @@ -1,93 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CosSimLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(cos, CosSimLayer); - -bool CosSimLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2LU); - - createFunction(forward_, - "CosSimForward", - FuncConfig().set("scale", (real)config_.cos_scale())); - createFunction(backward_, - "CosSimBackward", - FuncConfig().set("scale", (real)config_.cos_scale())); - - return true; -} - -void CosSimLayer::forward(PassType passType) { - Layer::forward(passType); - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(0)->getHeight(); - int size = getSize(); - CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed"; - - { - REGISTER_TIMER_INFO("CosFwResetTimer", getName().c_str()); - reserveOutput(batchSize, size); - } - - MatrixPtr outV = getOutputValue(); - /* activation */ { - REGISTER_TIMER_INFO("CosFwAtvTimer", getName().c_str()); - MatrixPtr prevOut1 = getInputValue(0); - MatrixPtr prevOut2 = getInputValue(1); - - CHECK(outV && prevOut1 && prevOut2); - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*prevOut1); - inputs.addArg(*prevOut2); - outputs.addArg(*outV, ASSIGN_TO); - forward_[0]->calc(inputs, outputs); - } -} - -void CosSimLayer::backward(const UpdateCallback& callback) { - /* activation */ { - REGISTER_TIMER_INFO("CosBpAtvTimer", getName().c_str()); - CHECK_EQ(backward_.size(), 1UL) << "Only one backward function needed"; - - const auto outG = this->getOutputGrad(); - const auto outV = this->getOutputValue(); - const auto inV1 = this->getInputValue(0); - const auto inV2 = this->getInputValue(1); - auto inG1 = this->getInputGrad(0); - auto inG2 = this->getInputGrad(1); - CHECK(outG && outV && inV1 && inV2 && inG1 && inG2); - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*outG); - inputs.addArg(*outV); - inputs.addArg(*inV1); - inputs.addArg(*inV2); - outputs.addArg(*inG1, ADD_TO); - outputs.addArg(*inG2, ADD_TO); - - backward_[0]->calc(inputs, outputs); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/CosSimLayer.h b/paddle/gserver/layers/CosSimLayer.h deleted file mode 100644 index d9fe1ff270f1f76e3b246dca374ddf45445419f9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CosSimLayer.h +++ /dev/null @@ -1,48 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { -/** - * @brief A layer for calculating cosine similarity between two vector - * \f[ - * f(x,y)=scale\frac{x_1y_1+x_2y_2+...+x_ny_n}{\sqrt{x_1^2+x_2^2+... - * +x_n^2}\sqrt{y_1^2+y_2^2+...+y_n^2}} - * \f] - * - * - Input1: A vector (batchSize * dataDim) * - * - Input2: A vector (batchSize * dataDim) or (1 * dataDim) * - * - Output: A vector (batchSize * 1) - * - * The config file api is cos_sim. - */ -class CosSimLayer : public Layer { - public: - explicit CosSimLayer(const LayerConfig& config) : Layer(config) {} - - ~CosSimLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/CosSimVecMatLayer.cpp b/paddle/gserver/layers/CosSimVecMatLayer.cpp deleted file mode 100644 index 230ecc768b4d7314b21ac1d76899c3c3bab12309..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CosSimVecMatLayer.cpp +++ /dev/null @@ -1,182 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { -/** - * @brief A layer for computing cosine similarity between a vector - * and each row of a matrix - * out[i] = cos_scale * cos(in1, in2(i,:)); - * @note used in NEURAL TURING MACHINE - * - * Input1: a vector (batchSize * dataDim) - * - * Input2: a matrix in vector form (batchSize * (weightDim*dataDim)) - * - * Output: a vector (batchSize * weightDim) - */ - -class CosSimVecMatLayer : public Layer { - protected: - MatrixPtr tmpMtx0; - MatrixPtr tmpMtx1; - MatrixPtr tmpRow0; - MatrixPtr tmpRow1; - MatrixPtr tmpRow2; - MatrixPtr tmpRow3; - - public: - explicit CosSimVecMatLayer(const LayerConfig& config) : Layer(config) {} - - ~CosSimVecMatLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(cos_vm, CosSimVecMatLayer); - -bool CosSimVecMatLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2U); - - size_t dataDim = inputLayers_[0]->getSize(); - size_t numKeys = getSize(); - size_t memoryDim = inputLayers_[1]->getSize(); - - CHECK_EQ(dataDim * numKeys, memoryDim) << "Dimension mismatch"; - - tmpRow0 = Matrix::create(nullptr, - /* height= */ 1, - dataDim, - /* trans= */ false, - useGpu_); - tmpRow1 = Matrix::create(nullptr, - /* height= */ 1, - dataDim, - /* trans= */ false, - useGpu_); - tmpRow2 = Matrix::create(nullptr, - /* height= */ numKeys, - 1, - /* trans= */ false, - useGpu_); - tmpRow3 = Matrix::create(nullptr, - /* height= */ numKeys, - 1, - /* trans= */ false, - useGpu_); - - tmpMtx0 = Matrix::create(nullptr, - /* height= */ numKeys, - dataDim, - /* trans= */ false, - useGpu_); - tmpMtx1 = Matrix::create(nullptr, - /* height= */ numKeys, - dataDim, - /* trans= */ false, - useGpu_); - - CHECK(tmpRow0 && tmpRow1 && tmpRow2 && tmpRow3 && tmpMtx0 && tmpMtx1); - - createFunction(forward_, - "CosSimForward", - FuncConfig().set("scale", (real)config_.cos_scale())); - createFunction(backward_, - "CosSimBackward", - FuncConfig().set("scale", (real)config_.cos_scale())); - - return true; -} - -void CosSimVecMatLayer::forward(PassType passType) { - Layer::forward(passType); - CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed"; - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV0->getHeight(); - size_t numKeys = getSize(); - - CHECK_EQ(batchSize, inV1->getHeight()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, numKeys); - } - - MatrixPtr outV = getOutputValue(); - CHECK(outV && inV0 && inV1); - REGISTER_TIMER_INFO("FwCosVMTimer", getName().c_str()); - for (size_t i = 0; i < batchSize; i++) { - tmpRow0->setData(inV0->rowBuf(i)); - tmpMtx0->setData(inV1->rowBuf(i)); - tmpRow2->setData(outV->rowBuf(i)); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*tmpMtx0); - inputs.addArg(*tmpRow0); - outputs.addArg(*tmpRow2, ASSIGN_TO); - forward_[0]->calc(inputs, outputs); - } -} - -void CosSimVecMatLayer::backward(const UpdateCallback& callback) { - CHECK_EQ(backward_.size(), 1UL) << "Only one forward function needed"; - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - MatrixPtr outV = getOutputValue(); - MatrixPtr outG = getOutputGrad(); - - size_t batchSize = inV0->getHeight(); - CHECK(inV0 && inV1 && inG0 && inG1 && outV && outG); - REGISTER_TIMER_INFO("BwCosVMTimer", getName().c_str()); - - for (size_t i = 0; i < batchSize; i++) { - tmpRow0->setData(inV0->rowBuf(i)); - tmpRow1->setData(inG0->rowBuf(i)); - tmpMtx0->setData(inV1->rowBuf(i)); - tmpMtx1->setData(inG1->rowBuf(i)); - tmpRow2->setData(outV->rowBuf(i)); - tmpRow3->setData(outG->rowBuf(i)); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*tmpRow3); - inputs.addArg(*tmpRow2); - inputs.addArg(*tmpMtx0); - inputs.addArg(*tmpRow0); - outputs.addArg(*tmpMtx1, ADD_TO); - outputs.addArg(*tmpRow1, ADD_TO); - - backward_[0]->calc(inputs, outputs); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/CostLayer.cpp b/paddle/gserver/layers/CostLayer.cpp deleted file mode 100644 index 1327616950a8887efa2cba410fa7ae8b5bd97da4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CostLayer.cpp +++ /dev/null @@ -1,748 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CostLayer.h" -#include -#include -#include -#include "paddle/utils/Logging.h" - -#include "paddle/math/SparseMatrix.h" - -namespace paddle { - -bool CostLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - bool ret = Layer::init(layerMap, parameterMap); - coeff_ = config_.coeff(); - if (!ret) return ret; - CHECK_GE(inputLayers_.size(), 2UL); - CHECK_LE(inputLayers_.size(), 3UL); - if (inputLayers_.size() == 3) { - weightLayer_ = inputLayers_[2]; - } - return true; -} - -void CostLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(*getOutputLayer())->getHeight(); - int size = 1; - resetOutput(batchSize, size); - - const MatrixPtr& output = getInputValue(*getOutputLayer()); - Argument label = getInput(*getLabelLayer()); - - /* get the cost value for each sample*/ - forwardImp(*output, label, *getOutputValue()); - if (weightLayer_) { - const MatrixPtr& weight = getInputValue(*weightLayer_); - getOutputValue()->dotMul(*getOutputValue(), *weight); - } -} - -void CostLayer::backward(const UpdateCallback& callback) { - (void)callback; - - const Argument& output = getInput(*getOutputLayer()); - Argument label = getInput(*getLabelLayer()); - - bool support = true; - if (weightLayer_) { - support = output.grad->getAbsSum() == 0; - } - - backwardImp(*output.value, label, *output.grad); - - if (weightLayer_) { - CHECK(support) << "Weighted cost layer '" << getName() - << "' must be the last layer " - "connected to the output layer '" - << getOutputLayer()->getName() << "'"; - output.grad->rowScale(0, *output.grad, *getInputValue(*weightLayer_)); - } - if (coeff_ != real(1.0f)) { - output.grad->add(coeff_, 0); - } -} - -// -// class MultiClassCrossEntropy -// -bool MultiClassCrossEntropy::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return CostLayer::init(layerMap, parameterMap); -} - -void MultiClassCrossEntropy::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - target.oneHotCrossEntropy(output, *label.ids); -} - -void MultiClassCrossEntropy::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - outputG.oneHotCrossEntropyBp(output, *label.ids); -} - -// -// class MultiClassCrossEntropyWithSelfNorm -// -REGISTER_LAYER(multi_class_cross_entropy_with_selfnorm, - MultiClassCrossEntropyWithSelfNorm); - -bool MultiClassCrossEntropyWithSelfNorm::init( - const LayerMap& layerMap, const ParameterMap& parameterMap) { - return CostLayer::init(layerMap, parameterMap); -} - -void MultiClassCrossEntropyWithSelfNorm::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - Matrix::resizeOrCreate(sftMaxSum_, output.getHeight(), 1, false, useGpu_); - output.rowSum(*sftMaxSum_); - sftMaxSum_->log2(); - - target.oneHotCrossEntropy(output, *label.ids); - target.add(*sftMaxSum_); - - sftMaxSum_->square2(); - target.add(*sftMaxSum_, config_.softmax_selfnorm_alpha()); -} - -void MultiClassCrossEntropyWithSelfNorm::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - Matrix::resizeOrCreate(sftMaxSum_, output.getHeight(), 1, false, useGpu_); - output.rowSum(*sftMaxSum_); - - Matrix::resizeOrCreate(sumInv_, output.getHeight(), 1, false, useGpu_); - sftMaxSum_->reciprocal2(*sumInv_); - - outputG.oneHotCrossEntropyBp(output, *label.ids); - outputG.addColumnVector(*sumInv_); - - sftMaxSum_->log2(); - sumInv_->dotMul(*sumInv_, *sftMaxSum_); - sumInv_->mulScalar(2 * config_.softmax_selfnorm_alpha()); - - outputG.addColumnVector(*sumInv_); -} - -// -// class SoftBinaryClassCrossEntropy -// -REGISTER_LAYER(soft_binary_class_cross_entropy, SoftBinaryClassCrossEntropy); - -bool SoftBinaryClassCrossEntropy::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return CostLayer::init(layerMap, parameterMap); -} - -void SoftBinaryClassCrossEntropy::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - Matrix::resizeOrCreate( - targetPerDim_, output.getHeight(), output.getWidth(), false, useGpu_); - - targetPerDim_->softCrossEntropy(output, *label.value); - targetPerDim_->rowSum(target); -} - -void SoftBinaryClassCrossEntropy::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - outputG.softCrossEntropyBp(output, *label.value); -} - -// -// class SumOfSquaresCostLayer -// - -REGISTER_LAYER(square_error, SumOfSquaresCostLayer); - -bool SumOfSquaresCostLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return CostLayer::init(layerMap, parameterMap); -} - -void SumOfSquaresCostLayer::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - target.sumOfSquares(output, *label.value); -} - -void SumOfSquaresCostLayer::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - outputG.sumOfSquaresBp(output, *label.value); -} - -// -// class SmoothL1CostLayer -// - -REGISTER_LAYER(smooth_l1, SmoothL1CostLayer); - -bool SmoothL1CostLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return CostLayer::init(layerMap, parameterMap); -} - -void SmoothL1CostLayer::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - MatrixPtr targetCpu, outputCpu, labelCpu; - if (useGpu_) { - targetCpu = - Matrix::create(target.getHeight(), target.getWidth(), false, false); - outputCpu = - Matrix::create(output.getHeight(), output.getWidth(), false, false); - labelCpu = Matrix::create( - label.value->getHeight(), label.value->getWidth(), false, false); - targetCpu->copyFrom(target); - outputCpu->copyFrom(output); - labelCpu->copyFrom(*label.value); - targetCpu->smoothL1(*outputCpu, *labelCpu, 1.0); - target.copyFrom(*targetCpu); - } else { - target.smoothL1(output, *label.value, 1.0); - } -} - -void SmoothL1CostLayer::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - MatrixPtr outputGCpu, outputCpu, labelCpu; - if (useGpu_) { - outputGCpu = - Matrix::create(outputG.getHeight(), outputG.getWidth(), false, false); - outputCpu = - Matrix::create(output.getHeight(), output.getWidth(), false, false); - labelCpu = Matrix::create( - label.value->getHeight(), label.value->getWidth(), false, false); - outputGCpu->copyFrom(outputG); - outputCpu->copyFrom(output); - labelCpu->copyFrom(*label.value); - outputGCpu->smoothL1Bp(*outputCpu, *labelCpu, 1.0); - outputG.copyFrom(*outputGCpu); - } else { - outputG.smoothL1Bp(output, *label.value, 1.0); - } -} - -// -// class RankingCost -// -bool RankingCost::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - posPairCount_ = 0; - negPairCount_ = 0; - - bool ret = Layer::init(layerMap, parameterMap); - if (!ret) return ret; - CHECK_GE(inputLayers_.size(), 3UL); - CHECK_LE(inputLayers_.size(), 4UL); - if (inputLayers_.size() == 4) { - weightLayer_ = inputLayers_[3]; - } - return true; -} - -void RankingCost::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(*getOutputLayer(0))->getHeight(); - int size = 1; - resizeOutput(batchSize, size); - Matrix::resizeOrCreate(margin_, batchSize, size, /* trans= */ false, useGpu_); - MatrixPtr label = getInputValue(*getLabelLayer()); - if (!label) { - // input label is not in value, try ids - IVectorPtr idLabel = getInput(*getLabelLayer()).ids; - CHECK(idLabel) << "label layer has neither value nor ids"; - CHECK_EQ((size_t)batchSize, idLabel->getSize()); - Matrix::resizeOrCreate( - labelBuf_, batchSize, /*width*/ 1, /*trans*/ false, useGpu_); - labelBuf_->copyFrom(*idLabel); - label = labelBuf_; - } - - MatrixPtr output[] = {getInputValue(*getOutputLayer(0)), - getInputValue(*getOutputLayer(1))}; - MatrixPtr target = this->getOutputValue(); - margin_->sub(*output[0], *output[1]); - - // for validation - size_t height = output[0]->getHeight(); - target->biggerThan(*(output[0]), *(output[1]), *label); - double total = static_cast(height); - if (weightLayer_) { - const MatrixPtr& weight = getInputValue(*weightLayer_); - target->dotMul(*target, *weight); - total = weight->getSum(); - } - double pos = target->getSum(); - posPairCount_ += pos; - negPairCount_ += (total - pos); - - // forward - target->logisticRegressionLoss(*margin_, *label); - if (weightLayer_) { - const MatrixPtr& weight = getInputValue(*weightLayer_); - target->dotMul(*target, *weight); - } -} - -void RankingCost::backward(const UpdateCallback& callback) { - (void)callback; - - MatrixPtr label = getInputValue(*getLabelLayer()); - if (!label) { - // input label is not in value, but in ids - // use labelBuf_ (should already resized and copied during forward) - label = labelBuf_; - } - - Matrix::resizeOrCreate( - marginGrad_, label->getHeight(), 1, /* trans= */ false, useGpu_); - marginGrad_->zeroMem(); - marginGrad_->logisticRegressionLossBp(*margin_, *label); - if (weightLayer_) { - const MatrixPtr& weight = getInputValue(*weightLayer_); - marginGrad_->dotMul(*marginGrad_, *weight); - } - - getInputGrad(0)->add(*marginGrad_); - getInputGrad(1)->sub(*marginGrad_); -} - -void RankingCost::onPassEnd() { - double ratio = posPairCount_ / ((negPairCount_ <= 0) ? 1.0 : negPairCount_); - LOG(INFO) << "calc pos/neg: " << ratio << " pos= " << posPairCount_ - << " neg= " << negPairCount_; - - posPairCount_ = 0; - negPairCount_ = 0; -} - -// -// class LambdaCost -// -REGISTER_LAYER(lambda_cost, LambdaCost); - -bool LambdaCost::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - truncationSize_ = config_.ndcg_num(); - maxSortSize_ = config_.max_sort_size(); - if (maxSortSize_ != -1) { - CHECK_GE(maxSortSize_, truncationSize_) - << "maxSortSize must be greater than or equal to NDCG size!"; - } - LOG(INFO) << "LambdaRank v1.3, NDCG size = " << truncationSize_ - << ", Max partial sort size = " << maxSortSize_; - CHECK(!useGpu_) << "LambdaRank supports CPU only!"; - return Layer::init(layerMap, parameterMap); -} - -void LambdaCost::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(*getOutputLayer())->getHeight(); - resizeOutput(batchSize, 1); - - MatrixPtr score = getInputValue(*getScoreLayer()); - MatrixPtr output = getInputValue(*getOutputLayer()); - MatrixPtr target = this->getOutputValue(); - - real* scoreData = score->getData(); - real* outputData = output->getData(); - real* targetData = target->getData(); - - auto startPos = getInput(*getOutputLayer()).sequenceStartPositions; - const int* startPosData = startPos->getData(false); - size_t batchNum = startPos->getSize() - 1; - for (size_t i = 0; i < batchNum; ++i) { - int beginPos = startPosData[i]; - int endPos = startPosData[i + 1]; - real NDCG = calcNDCG( - outputData + beginPos, scoreData + beginPos, endPos - beginPos); - for (int j = beginPos; j < endPos; ++j) { - targetData[j] = NDCG; - } - } -} - -void LambdaCost::backward(const UpdateCallback& callback) { - (void)callback; - MatrixPtr score = getInputValue(*getScoreLayer()); - MatrixPtr output = getInputValue(*getOutputLayer()); - Matrix::resizeOrCreate(marginGrad_, - score->getHeight(), - 1, - /* trans= */ false, - useGpu_); - marginGrad_->zeroMem(); - - real* gradData = marginGrad_->getData(); - real* scoreData = score->getData(); - real* outputData = output->getData(); - - auto startPos = getInput(*getOutputLayer()).sequenceStartPositions; - const int* startPosData = startPos->getData(false); - size_t batchNum = startPos->getSize() - 1; - - for (size_t i = 0; i < batchNum; ++i) { - int beginPos = startPosData[i]; - int endPos = startPosData[i + 1]; - calcGrad(outputData + beginPos, - scoreData + beginPos, - gradData + beginPos, - endPos - beginPos); - } - - getInputGrad(0)->add(*marginGrad_); -} - -void LambdaCost::calcGrad(const real* outputScore, - const real* score, - real* gradData, - int size) { - CHECK_GE(size, truncationSize_) - << "Invalid: (Sample num in the same list) < (NDCG truncation num) !"; - int sortSize = maxSortSize_ == -1 ? size : std::min(maxSortSize_, size); - - scorePair_.clear(); - for (int i = 0; i < size; ++i) { - scorePair_.push_back(std::make_pair(score[i], i)); - } - if (size <= sortSize) { - std::sort(scorePair_.begin(), - scorePair_.end(), - [](const std::pair& a, const std::pair& b) { - return a.first > b.first; - }); - } else { - std::partial_sort( - scorePair_.begin(), - scorePair_.begin() + sortSize, - scorePair_.end(), - [](const std::pair& a, const std::pair& b) { - return a.first > b.first; - }); - } - - real maxDCG = 0; - for (int i = 0; i < truncationSize_; ++i) { - maxDCG += (std::pow(2, scorePair_[i].first) - 1) / std::log(i + 2); - } - CHECK_GT(maxDCG, 0) << "Invalid: max DCG = 0!"; - - for (int i = 0; i < sortSize; ++i) { - for (int j = i + 1; j < size; ++j) { - int index_i = scorePair_[i].second; - int index_j = scorePair_[j].second; - real score_i = score[index_i]; - real score_j = score[index_j]; - real dcgDif = 0; - if (j < sortSize) { - dcgDif = (std::pow(2, score_i) - std::pow(2, score_j)) * - (1 / std::log(i + 2) - 1 / std::log(j + 2)); - } else { - dcgDif = - (std::pow(2, score_i) - std::pow(2, score_j)) / std::log(i + 2); - } - - real lambda_ij = - -std::abs(dcgDif) / - (1 + std::exp(outputScore[index_i] - outputScore[index_j])); - gradData[index_i] += lambda_ij / maxDCG; - gradData[index_j] -= lambda_ij / maxDCG; - } - } -} - -real LambdaCost::calcNDCG(const real* outputScore, - const real* score, - int size) { - CHECK_GE(size, truncationSize_) - << "Invalid: (Sample num in the same list) < (NDCG truncation num) !"; - - outputScorePair_.clear(); - for (int i = 0; i < size; ++i) { - outputScorePair_.push_back(std::make_pair(outputScore[i], i)); - } - std::partial_sort( - outputScorePair_.begin(), - outputScorePair_.begin() + truncationSize_, - outputScorePair_.end(), - [](const std::pair& a, const std::pair& b) { - return a.first > b.first; - }); - - real DCG = 0; - for (int i = 0; i < truncationSize_; ++i) { - DCG += - (std::pow(2, score[outputScorePair_[i].second]) - 1) / std::log(i + 2); - } - - scoreVec_.resize(size); - std::copy(score, score + size, scoreVec_.begin()); - real maxDCG = 0; - std::partial_sort(scoreVec_.begin(), - scoreVec_.begin() + truncationSize_, - scoreVec_.end(), - std::greater()); - for (int i = 0; i < truncationSize_; ++i) { - maxDCG += (std::pow(2, scoreVec_[i]) - 1) / std::log(i + 2); - } - CHECK_GT(maxDCG, 0) << "Invalid: max DCG = 0!"; - - return DCG / maxDCG; -} - -// -// class MultiBinaryLabelCrossEntropy -// - -REGISTER_LAYER(multi_binary_label_cross_entropy, MultiBinaryLabelCrossEntropy); - -bool MultiBinaryLabelCrossEntropy::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return CostLayer::init(layerMap, parameterMap); -} - -void MultiBinaryLabelCrossEntropy::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - MatrixPtr value = nullptr; - if (label.ids) { - CHECK(!label.value); - value = label.ids->toOneHotSparseMatrix(output.getWidth(), useGpu_); - } else { - CHECK(label.value); - value = label.value; - } - - if (dynamic_cast(value.get()) || - dynamic_cast(value.get())) { - target.multiBinaryLabelCrossEntropy(output, *value); - } else { - Matrix::resizeOrCreate( - targetPerDim_, output.getHeight(), output.getWidth(), false, useGpu_); - - targetPerDim_->binaryLabelCrossEntropy(output, *value); - targetPerDim_->rowSum(target); - } -} - -void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - MatrixPtr value = nullptr; - if (label.ids) { - CHECK(!value); - value = label.ids->toOneHotSparseMatrix(output.getWidth(), useGpu_); - } else { - CHECK(label.value); - value = label.value; - } - - if (dynamic_cast(value.get()) || - dynamic_cast(value.get())) { - outputG.multiBinaryLabelCrossEntropyBp(output, *value); - } else { - outputG.binaryLabelCrossEntropyBp(output, *value); - } -} - -bool HuberCost::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - CostLayer::init(layerMap, parameterMap); - if (useGpu_) { - tmpCpuInput_.reserve(inputLayers_.size()); - for (size_t i = 0; i < inputLayers_.size(); i++) { - tmpCpuInput_.push_back(Argument()); - } - } - return true; -} - -void HuberCost::forwardImp(Matrix& output, Argument& label, Matrix& cost) { - if (useGpu_) { - for (size_t i = 0; i < inputLayers_.size(); i++) { - tmpCpuInput_[i].resizeAndCopyFrom( - getInput(i), false, HPPL_STREAM_DEFAULT); - } - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - } -} - -// -// Huber loss for robust regression. -// -REGISTER_LAYER(huber_regression, HuberRegressionLoss); - -bool HuberRegressionLoss::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - HuberCost::init(layerMap, parameterMap); - delta_ = config_.delta(); - return true; -} - -void HuberRegressionLoss::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - HuberCost::forwardImp(output, label, target); - size_t numSamples = target.getHeight(); - size_t dim = output.getWidth(); - CHECK(label.value); - CHECK_EQ((*label.value).getHeight(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(dim, (*label.value).getWidth()); - CHECK_EQ(target.getWidth(), (size_t)1); - - real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); - real* lbl = - useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); - std::vector cost(numSamples, 0); - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = 0; j < dim; ++j) { - int index = i * dim + j; - real a = std::abs(lbl[index] - out[index]); - if (a <= delta_) - cost[i] += a * a / 2; - else - cost[i] += delta_ * (a - delta_ / 2); - } - } - target.copyFrom(cost.data(), numSamples); -} - -void HuberRegressionLoss::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - size_t numSamples = output.getHeight(); - size_t dim = output.getWidth(); - real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); - real* lbl = - useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); - real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = 0; j < dim; ++j) { - int index = i * dim + j; - real a = lbl[index] - out[index]; - if (std::abs(a) <= delta_) - grad[index] += -a; - else - grad[index] += a > 0 ? -delta_ : delta_; - } - } - if (useGpu_) outputG.copyFrom(grad, numSamples * dim); -} - -// -// Huber loss for robust 2-classes classification -// -REGISTER_LAYER(huber_classification, HuberTwoClassification); - -bool HuberTwoClassification::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return HuberCost::init(layerMap, parameterMap); -} - -void HuberTwoClassification::forwardImp(Matrix& output, - Argument& label, - Matrix& target) { - HuberCost::forwardImp(output, label, target); - size_t numSamples = target.getHeight(); - CHECK(label.ids); - CHECK_EQ((*label.ids).getSize(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(output.getWidth(), (size_t)1); - CHECK_EQ(target.getWidth(), (size_t)1); - - real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); - int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); - std::vector cost(numSamples, 0); - for (size_t i = 0; i < numSamples; ++i) { - int y = 2 * lbl[i] - 1; - real a = out[i] * y; - if (a < -1) - cost[i] = -4 * a; - else if (a < 1) - cost[i] = (1 - a) * (1 - a); - } - target.copyFrom(cost.data(), numSamples); -} - -void HuberTwoClassification::backwardImp(Matrix& output, - Argument& label, - Matrix& outputG) { - size_t numSamples = output.getHeight(); - real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); - int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); - real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); - for (size_t i = 0; i < numSamples; ++i) { - int y = 2 * lbl[i] - 1; - real a = out[i] * y; - if (a < -1) - grad[i] += -4 * y; - else if (a < 1) - grad[i] += -2 * (1 - a) * y; - } - if (useGpu_) outputG.copyFrom(grad, numSamples); -} -/** - * This cost layer compute the sum of its input as loss. - * \f[ - * o(i) = \sum_{j=1}^D y_{ij} - * \f] - */ -class SumCostLayer : public Layer { - public: - explicit SumCostLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override { - bool ret = Layer::init(layerMap, parameterMap); - if (!ret) return ret; - CHECK_EQ(inputLayers_.size(), 1UL); - return true; - } - - void forward(PassType passType) override { - Layer::forward(passType); - const MatrixPtr& input = getInputValue(0); - - /* malloc memory for the output_ if necessary */ - int batchSize = input->getHeight(); - int size = 1; - resizeOutput(batchSize, size); - output_.value->sumRows(*input, /* scaleSum= */ 1, /* scaleDest= */ 0); - } - - void backward(const UpdateCallback& callback = nullptr) override { - getInputGrad(0)->add((real)1); - } -}; - -REGISTER_LAYER(sum_cost, SumCostLayer); - -} // namespace paddle diff --git a/paddle/gserver/layers/CropLayer.cpp b/paddle/gserver/layers/CropLayer.cpp deleted file mode 100644 index bc97ca2f9e0cdc86f82baa0ce3fbafde2db0c10f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CropLayer.cpp +++ /dev/null @@ -1,146 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CropLayer.h" -#include "paddle/utils/Stat.h" -namespace paddle { - -REGISTER_LAYER(crop, CropLayer); - -bool CropLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - CHECK_LE(static_cast(inputLayers_.size()), 2); - CHECK_GE(static_cast(inputLayers_.size()), 1); - crop_axis_ = config_.axis(); - for (int i = 0; i < config_.offset_size(); i++) { - crop_offsets_.push_back(config_.offset(i)); - } - - // 1. get input_0 shape - auto& input0_img_conf = config_.inputs(0).image_conf(); - inDims_ = TensorShape({0, - input0_img_conf.channels(), - input0_img_conf.has_img_size_y() - ? input0_img_conf.img_size_y() - : input0_img_conf.img_size(), - input0_img_conf.img_size()}); - // 2. get target dims from config - if (config_.inputs_size() == 1) { - targetDims_ = TensorShape({config_.shape(0), - config_.shape(1), - config_.shape(2), - config_.shape(3)}); - } else { - // 2. get input_1 shape - auto& input1_img_conf = config_.inputs(1).image_conf(); - targetDims_ = TensorShape({0, - input1_img_conf.channels(), - input1_img_conf.has_img_size_y() - ? input1_img_conf.img_size_y() - : input1_img_conf.img_size(), - input1_img_conf.img_size()}); - } - - // 3. get final crop corner - int dimSize = 4; - crop_corner_ = {0, 0, 0, 0}; - for (int i = 0; i < dimSize; i++) { - if (i >= crop_axis_) { - if (crop_offsets_.size() > 1) { - crop_corner_[i] = crop_offsets_[i - crop_axis_]; - } else { - crop_corner_[i] = crop_offsets_[0]; - } - } - } - - outDims_ = TensorShape(4); - - createFunction( - forward_, "Crop", FuncConfig().set("crop_corner", crop_corner_)); - createFunction( - backward_, "CropGrad", FuncConfig().set("crop_corner", crop_corner_)); - - return true; -} - -void CropLayer::setOutDims() { - MatrixPtr input = inputLayers_[1]->getOutputValue(); - size_t batchSize = input->getHeight(); - // get target dims from input_1 - if (config_.inputs_size() == 2) { - targetDims_.setDim(0, batchSize); - int ch = config_.inputs(0).image_conf().channels(); - if (ch != 0) targetDims_.setDim(1, ch); - int h = inputLayers_[1]->getOutput().getFrameHeight(); - if (h != 0) targetDims_.setDim(2, h); - int w = inputLayers_[1]->getOutput().getFrameWidth(); - if (w != 0) targetDims_.setDim(3, w); - } - // get final crop shape from target dims and crop axis - std::vector crop_shape; - int dimSize = 4; - for (int i = 0; i < dimSize; i++) { - if (i >= crop_axis_) { - crop_shape.push_back(targetDims_[i]); - } else { - crop_shape.push_back(inDims_[i]); - } - } - - outDims_.reshape( - {crop_shape[0], crop_shape[1], crop_shape[2], crop_shape[3]}); - output_.setFrameHeight(crop_shape[2]); - output_.setFrameWidth(crop_shape[3]); -} - -void CropLayer::setInDims() { - MatrixPtr input = inputLayers_[0]->getOutputValue(); - size_t batchSize = input->getHeight(); - inDims_.setDim(0, batchSize); - int h = inputLayers_[0]->getOutput().getFrameHeight(); - if (h != 0) inDims_.setDim(2, h); - int w = inputLayers_[0]->getOutput().getFrameWidth(); - if (w != 0) inDims_.setDim(3, w); -} - -void CropLayer::forward(PassType passType) { - Layer::forward(passType); - setInDims(); - setOutDims(); - int size = outDims_[1] * outDims_[2] * outDims_[3]; - resetOutput(outDims_[0], size); - MatrixPtr outV = getOutputValue(); - REGISTER_TIMER_INFO("CropForward", getName().c_str()); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), inDims_); - outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO); - forward_[0]->calc(inputs, outputs); -} - -void CropLayer::backward(const UpdateCallback& callback) { - (void)callback; - REGISTER_TIMER_INFO("CropBackward", getName().c_str()); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getOutputGrad(), outDims_); - outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); - backward_[0]->calc(inputs, outputs); -} -} // namespace paddle diff --git a/paddle/gserver/layers/CrossChannelNormLayer.cpp b/paddle/gserver/layers/CrossChannelNormLayer.cpp deleted file mode 100644 index 644450291ee8a308accf7a1fe096332cc8c241dc..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CrossChannelNormLayer.cpp +++ /dev/null @@ -1,137 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "NormLayer.h" -#include "paddle/math/BaseMatrix.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -MatrixPtr CrossChannelNormLayer::createSampleMatrix(MatrixPtr data, - size_t iter, - size_t spatialDim) { - return Matrix::create(data->getData() + iter * channels_ * spatialDim, - channels_, - spatialDim, - false, - useGpu_); -} - -MatrixPtr CrossChannelNormLayer::createSpatialMatrix(MatrixPtr data, - size_t iter, - size_t spatialDim) { - return Matrix::create( - data->getData() + iter * spatialDim, 1, spatialDim, false, useGpu_); -} - -bool CrossChannelNormLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - CHECK(parameters_[0]); - const NormConfig& conf = config_.inputs(0).norm_conf(); - channels_ = conf.channels(); - scale_.reset(new Weight(channels_, 1, parameters_[0])); - return true; -} - -void CrossChannelNormLayer::forward(PassType passType) { - Layer::forward(passType); - MatrixPtr inV = getInputValue(0); - - size_t batchSize = inV->getHeight(); - size_t dataDim = inV->getWidth(); - CHECK_EQ(getSize(), dataDim); - - reserveOutput(batchSize, dataDim); - MatrixPtr outV = getOutputValue(); - size_t spatialDim = dataDim / channels_; - - Matrix::resizeOrCreate(dataBuffer_, batchSize, dataDim, false, useGpu_); - Matrix::resizeOrCreate(spatialBuffer_, 1, spatialDim, false, useGpu_); - Matrix::resizeOrCreate(normBuffer_, batchSize, spatialDim, false, useGpu_); - - inV->square2(*dataBuffer_); - for (size_t i = 0; i < batchSize; i++) { - const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim); - const MatrixPtr dataTmp = createSampleMatrix(dataBuffer_, i, spatialDim); - MatrixPtr outVTmp = createSampleMatrix(outV, i, spatialDim); - MatrixPtr normTmp = createSpatialMatrix(normBuffer_, i, spatialDim); - - // compute norm. - spatialBuffer_->sumCols(*dataTmp, 1, 0); - // add eps to avoid overflow - spatialBuffer_->add(1e-6); - spatialBuffer_->sqrt2(*spatialBuffer_); - normTmp->copyFrom(*spatialBuffer_); - outVTmp->copyFrom(*inVTmp); - outVTmp->divRowVector(*spatialBuffer_); - // scale the layer. - outVTmp->mulColVector(*scale_->getW()); - } -} - -void CrossChannelNormLayer::backward(const UpdateCallback& callback) { - MatrixPtr inG = getInputGrad(0); - MatrixPtr inV = getInputValue(0); - MatrixPtr outG = getOutputGrad(); - MatrixPtr outV = getOutputValue(); - - size_t batchSize = inG->getHeight(); - size_t dataDim = inG->getWidth(); - size_t spatialDim = dataDim / channels_; - - MatrixPtr inGBuffer; - Matrix::resizeOrCreate(inGBuffer, channels_, spatialDim, false, useGpu_); - - dataBuffer_->dotMul(*outG, *outV); - Matrix::resizeOrCreate(scaleDiff_, channels_, 1, false, useGpu_); - Matrix::resizeOrCreate(channelBuffer_, channels_, 1, false, useGpu_); - Matrix::resizeOrCreate(sampleBuffer_, channels_, spatialDim, false, useGpu_); - scaleDiff_->zeroMem(); - for (size_t i = 0; i < batchSize; i++) { - MatrixPtr outGTmp = createSampleMatrix(outG, i, spatialDim); - const MatrixPtr dataTmp = createSampleMatrix(dataBuffer_, i, spatialDim); - const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim); - const MatrixPtr inGTmp = createSampleMatrix(inG, i, spatialDim); - const MatrixPtr normTmp = createSpatialMatrix(normBuffer_, i, spatialDim); - - channelBuffer_->sumRows(*dataTmp, 1, 0); - channelBuffer_->dotDiv(*channelBuffer_, *(scale_->getW())); - // store a / scale[i] in scaleDiff_ temporary - scaleDiff_->add(*channelBuffer_, 1.); - - sampleBuffer_->dotMul(*inVTmp, *outGTmp); - spatialBuffer_->sumCols(*sampleBuffer_, 1., 0.); - // scale the grad - inGBuffer->copyFrom(*inVTmp); - inGBuffer->mulRowVector(*spatialBuffer_); - // divide by square of norm - spatialBuffer_->dotMul(*normTmp, *normTmp); - inGBuffer->divRowVector(*spatialBuffer_); - // subtract - inGBuffer->add(*outGTmp, -1, 1); - // divide by norm - inGBuffer->divRowVector(*normTmp); - // scale the diff - inGBuffer->mulColVector(*scale_->getW()); - - inGTmp->add(*inGBuffer); - } - // updata scale - if (scale_->getWGrad()) scale_->getWGrad()->add(*scaleDiff_); - scale_->getParameterPtr()->incUpdate(callback); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp deleted file mode 100644 index 9a29e6a55e95334def2b83dc4a794e07a7fd5154..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp +++ /dev/null @@ -1,180 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CudnnBatchNormLayer.h" -#include "Layer.h" -#include "paddle/cuda/include/hl_batch_norm.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(cudnn_batch_norm, CudnnBatchNormLayer); - -bool CudnnBatchNormLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - if (!BatchNormBaseLayer::init(layerMap, parameterMap)) return false; - CHECK(useGpu_) << "CudnnBatchNorm only support GPU"; - - hl_create_tensor_descriptor(&ioDesc_); - hl_create_tensor_descriptor(&bnParamDesc_); - hl_tensor_reshape(bnParamDesc_, 1, channels_, 1, 1); - - return true; -} - -void CudnnBatchNormLayer::reshape(int batchSize) { - hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_); -} - -void CudnnBatchNormLayer::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = getInputValue(0)->getHeight(); - calFeatureMapSize(); - reshape(batchSize); - resetOutput(batchSize, getInputValue(0)->getWidth()); - - // for testing in training peroid. - useGlobalStats_ = (passType == PASS_TEST); - if (passType == PASS_TEST && config_.has_use_global_stats()) { - useGlobalStats_ = config_.use_global_stats(); - } - - real* input = getInputValue(0)->getData(); - real* output = getOutputValue()->getData(); - real* gamma = weight_->getW()->getData(); - real* beta = biases_->getW()->getData(); - real* movingMean = movingMean_->getW()->getData(); - real* movingVar = movingVar_->getW()->getData(); - - // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. - eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); - - if (!useGlobalStats_) { - REGISTER_TIMER_INFO("CudnnBatchFwTimer", getName().c_str()); - real* savedMean = savedMean_->getData(); - real* savedInvVar = savedInvVar_->getData(); - hl_batch_norm_forward_training(ioDesc_, - input, - ioDesc_, - output, - bnParamDesc_, - gamma, - beta, - 1.0 - movingAvgFraction_, - movingMean, - movingVar, - eps_, - savedMean, - savedInvVar); - } else { - // used movingMean and movingVar in testing - if (batchSize <= 1024) { - hl_batch_norm_forward_inference(ioDesc_, - input, - ioDesc_, - output, - bnParamDesc_, - gamma, - beta, - movingMean, - movingVar, - eps_); - } else { - // There is a limitation in cudnn library. - // When the batch size is larger than 1024 in cuDNN v5.1, - // the cudnnBatchNormalizationForwardInference will fail. - hl_batch_norm_cuda_inference(input, - output, - gamma, - beta, - movingMean, - movingVar, - eps_, - batchSize, - channels_, - imageH_ * imageD_, - imageW_); - } - } - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void CudnnBatchNormLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - real* input = getInputValue(0)->getData(); - real* outGrad = getOutputGrad()->getData(); - real* inGrad = getInputGrad(0)->getData(); - real* gamma = weight_->getW()->getData(); - real* savedMean = savedMean_->getData(); - real* savedInvVar = savedInvVar_->getData(); - - // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. - eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); - - auto create = [](MatrixPtr& m, size_t h, size_t w, real** p) { - Matrix::resizeOrCreate(m, h, w, false, true); - m->zeroMem(); - *p = m->getData(); - }; - - real* gammaGrad = nullptr; - real* betaGrad = nullptr; - if (weight_->getWGrad()) { - gammaGrad = weight_->getWGrad()->getData(); - } else { - create(tmpWGrad_, 1, channels_, &gammaGrad); - } - if (biases_ && biases_->getWGrad()) { - betaGrad = biases_->getWGrad()->getData(); - } else { - create(tmpBiasGrad_, 1, channels_, &betaGrad); - } - - hl_batch_norm_backward(ioDesc_, - input, - ioDesc_, - outGrad, - ioDesc_, - inGrad, - bnParamDesc_, - gamma, - gammaGrad, - betaGrad, - eps_, - savedMean, - savedInvVar); - - { - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); - biases_->getParameterPtr()->incUpdate(callback); - weight_->getParameterPtr()->incUpdate(callback); - } -} - -CudnnBatchNormLayer::~CudnnBatchNormLayer() { - hl_destroy_tensor_descriptor(ioDesc_); - hl_destroy_tensor_descriptor(bnParamDesc_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.h b/paddle/gserver/layers/CudnnBatchNormLayer.h deleted file mode 100644 index 1bb4eff8d2372660caa4ec4a4a20a27f365bebd0..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CudnnBatchNormLayer.h +++ /dev/null @@ -1,68 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "BatchNormBaseLayer.h" -#include "Layer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief Cudnn Batch normalization layer use to cuDNN lib to implentment. - * @note Cudnn version must >= v4.0, and better to use the latest version - * (v5.1). - * - * The config file api is batch_norm_layer. - */ - -class CudnnBatchNormLayer : public BatchNormBaseLayer { - public: - explicit CudnnBatchNormLayer(const LayerConfig& config) - : BatchNormBaseLayer(config) {} - - ~CudnnBatchNormLayer(); - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - /** - * reshape tensor of ioDesc_. - */ - void reshape(int batchSize); - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - protected: - /// Epsilon value used in the batch normalization formula. - /// Same epsilon value should be used in forward and backward functions. - double eps_; - - /// Input/output tensor descriptor desc - hl_tensor_descriptor ioDesc_; - /// Shared tensor descriptor desc for the 6 tenros: - /// bnScale, bnBias, running mean/var, save_mean/var - hl_tensor_descriptor bnParamDesc_; - - /** - * @brief The gradient of weight and bias in cudnn api can not be empty. - * If set is_static for weight or bias, it will not allocate memory for them, - * and the gradient is NULL. In this case, will use two matrix. - */ - MatrixPtr tmpWGrad_, tmpBiasGrad_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/CudnnConvBaseLayer.cpp b/paddle/gserver/layers/CudnnConvBaseLayer.cpp deleted file mode 100644 index 6d0a40a60710603900a9b89980d38b2d7638ad60..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CudnnConvBaseLayer.cpp +++ /dev/null @@ -1,135 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CudnnConvBaseLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { -REGISTER_LAYER(cudnn_conv, CudnnConvBaseLayer); -REGISTER_LAYER(cudnn_convt, CudnnConvBaseLayer); - -bool CudnnConvBaseLayer::init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) { - if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; - CHECK(useGpu_) << "CudnnConvLayer only support gpu"; - - CHECK_EQ(inputLayers_.size(), parameters_.size()); - projections_.reserve(inputLayers_.size()); - projConf_.reserve(inputLayers_.size()); - - numFilters_ = config_.num_filters(); - CHECK(config_.shared_biases()); - for (size_t i = 0; i < inputLayers_.size(); i++) { - ProjectionConfig *conf = new ProjectionConfig(); - if (isDeconv_) { - conf->set_type("convt"); - } else { - conf->set_type("conv"); - } - conf->set_num_filters(numFilters_); - ConvConfig *convConf = conf->mutable_conv_conf(); - *convConf = *(config_.mutable_inputs(i)->mutable_conv_conf()); - conf->set_input_size(getPrev(i)->getSize()); - conf->set_output_size(getSize()); - projConf_.emplace_back(conf); - projections_.emplace_back( - Projection::create(*projConf_[i], parameters_[i], useGpu_)); - - // create a new weight - size_t height, width; - height = filterPixels_[i] * filterChannels_[i]; - width = (!isDeconv_) ? numFilters_ : channels_[i]; - CHECK_EQ(parameters_[i]->getSize(), width * height); - Weight *w = new Weight(height, width, parameters_[i]); - weights_.emplace_back(w); - } - - if (biasParameter_.get()) { - if (sharedBiases_) { - CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); - biases_ = - std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); - } else { - biases_ = - std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); - } - } - if (biases_.get() && sharedBiases_) { - hl_create_tensor_descriptor(&biasDesc_); - hl_create_tensor_descriptor(&outputDesc_); - hl_tensor_reshape(biasDesc_, 1, numFilters_, 1, 1); - } - - return true; -} - -void CudnnConvBaseLayer::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = getInput(0).getBatchSize(); - resetOutput(batchSize, calOutputSize()); - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - projections_[i]->forward(&getInput(i), &getOutput(), passType); - } - - if (biases_) { - REGISTER_TIMER_INFO("CudnnConvBiasTimer", getName().c_str()); - int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - int outH = outputH_[0]; - int outW = outputW_[0]; - - hl_tensor_reshape(outputDesc_, - batchSize, - numFilters_, - outH, - outW, - numFilters_ * outH * outW, - outH * outW, - outW, - 1); - real *outData = getOutputValue()->getData(); - real *biasData = biases_->getW()->getData(); - hl_convolution_forward_add_bias(biasDesc_, biasData, outputDesc_, outData); - } - - forwardActivation(); -} - -void CudnnConvBaseLayer::backward(const UpdateCallback &callback) { - backwardActivation(); - - if (biases_ && biases_->getWGrad()) { - REGISTER_TIMER_INFO("CudnnConvBpBiasTimer", getName().c_str()); - real *biasGrad = biases_->getWGrad()->getData(); - real *outGrad = getOutputGrad()->getData(); - hl_convolution_backward_bias(biasDesc_, biasGrad, outputDesc_, outGrad); - - biases_->getParameterPtr()->incUpdate(callback); - } - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - projections_[i]->backward(callback); - } -} - -CudnnConvBaseLayer::~CudnnConvBaseLayer() { - if (biases_) { - hl_destroy_tensor_descriptor(biasDesc_); - hl_destroy_tensor_descriptor(outputDesc_); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/CudnnConvBaseLayer.h b/paddle/gserver/layers/CudnnConvBaseLayer.h deleted file mode 100644 index 1ee1aa100d8adaed04ce24ee12b5b9af52c14b13..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CudnnConvBaseLayer.h +++ /dev/null @@ -1,53 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "ConvBaseLayer.h" -#include "Projection.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief A 2-dimension conv layer implemented by cuDNN. It only - * supports GPU mode. We automatic select CudnnConvLayer for GPU - * mode and ExpandConvLayer for CPU mode if you set type of "conv". - * User also can specfiy type of "exconv" or "cudnn_conv" for - * particular type. - * - * The config file api is img_conv_layer. - */ -class CudnnConvBaseLayer : public ConvBaseLayer { - protected: - std::vector> projConf_; - std::vector> projections_; - - hl_tensor_descriptor biasDesc_; - hl_tensor_descriptor outputDesc_; - - public: - explicit CudnnConvBaseLayer(const LayerConfig& config) - : ConvBaseLayer(config) {} - - ~CudnnConvBaseLayer(); - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override; - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/CudnnPoolLayer.cpp b/paddle/gserver/layers/CudnnPoolLayer.cpp deleted file mode 100644 index ac6d2168f43590a6acd70f6641ff729327894ea0..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/CudnnPoolLayer.cpp +++ /dev/null @@ -1,139 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CudnnPoolLayer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -bool CudnnPoolLayer::typeCheck(const std::string &poolType, - hl_pooling_mode_t *mode) { - if (poolType == "cudnn-max-pool") { - if (mode) { - *mode = HL_POOLING_MAX; - } - } else if (poolType == "cudnn-avg-pool") { - if (mode) { - *mode = HL_POOLING_AVERAGE; - } - } else if (poolType == "cudnn-avg-incl-pad-pool") { - if (mode) { - *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING; - } - } else { - return false; - } - - return true; -} - -CudnnPoolLayer::CudnnPoolLayer(const LayerConfig &config) : PoolLayer(config) { - const std::string &pool_type = config.inputs(0).pool_conf().pool_type(); - CHECK_EQ(CudnnPoolLayer::typeCheck(pool_type, &mode_), true); -} - -bool CudnnPoolLayer::init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) { - PoolLayer::init(layerMap, parameterMap); - - CHECK(useGpu_) << "CudnnPoolLayer only support gpu"; - - hl_create_tensor_descriptor(&inputDesc_); - hl_create_tensor_descriptor(&outputDesc_); - - windowHeight = sizeY_; - windowWidth = sizeX_; - heightPadding = confPaddingY_; - widthPadding = confPadding_; - strideHeight = strideY_; - strideWidth = stride_; - - hl_create_pooling_descriptor(&poolingDesc_, - mode_, - windowHeight, - windowWidth, - heightPadding, - widthPadding, - strideHeight, - strideWidth); - - return true; -} - -void CudnnPoolLayer::reshape(int batchSize) { - imageH_ = inputLayers_[0]->getOutput().getFrameHeight(); - imageW_ = inputLayers_[0]->getOutput().getFrameWidth(); - if (imageH_ == 0) { - imageH_ = imgSizeY_; - } - if (imageW_ == 0) { - imageW_ = imgSize_; - } - CHECK_EQ(inputLayers_[0]->getOutput().value->getWidth(), - channels_ * imageH_ * imageW_); - outputH_ = outputSize(imageH_, - sizeY_, - confPaddingY_, - strideY_, - /* caffeMode */ false); - outputW_ = - outputSize(imageW_, sizeX_, confPadding_, stride_, /* caffeMode */ false); - getOutput().setFrameHeight(outputH_); - getOutput().setFrameWidth(outputW_); - - hl_tensor_reshape(inputDesc_, batchSize, channels_, imageH_, imageW_); - hl_tensor_reshape(outputDesc_, batchSize, channels_, outputH_, outputW_); -} - -void CudnnPoolLayer::forward(PassType passType) { - Layer::forward(passType); - - CHECK(inputLayers_[0]->getOutputValue()->useGpu()); - int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - reshape(batchSize); - resetOutput(batchSize, outputH_ * outputW_ * channels_); - - real *inputData = getInputValue(0)->getData(); - real *outData = getOutputValue()->getData(); - hl_pooling_forward(inputDesc_, inputData, outputDesc_, outData, poolingDesc_); -} - -void CudnnPoolLayer::backward(const UpdateCallback &callback) { - (void)callback; - if (NULL == getInputGrad(0)) { - return; - } - - real *inputData = getInputValue(0)->getData(); - real *inputGrad = getInputGrad(0)->getData(); - real *outData = getOutputValue()->getData(); - real *outGrad = getOutputGrad()->getData(); - hl_pooling_backward(inputDesc_, - inputData, - inputGrad, - outputDesc_, - outData, - outGrad, - poolingDesc_); -} - -CudnnPoolLayer::~CudnnPoolLayer() { - hl_destroy_tensor_descriptor(inputDesc_); - hl_destroy_tensor_descriptor(outputDesc_); - hl_destroy_pooling_descriptor(poolingDesc_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/DataNormLayer.cpp b/paddle/gserver/layers/DataNormLayer.cpp deleted file mode 100644 index 86da4d6f957e2ce0afc53d69f9d57c234f8f178f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/DataNormLayer.cpp +++ /dev/null @@ -1,140 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "DataNormLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(data_norm, DataNormLayer); - -bool DataNormLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* initialize the weight */ - CHECK(!biasParameter_) << "DataNormLayer does not need bias"; - CHECK(inputLayers_.size() == 1 && inputLayers_[0]->getType() == "data") - << "DataNormLayer accepts one and only one DataLayer as its input layer"; - CHECK_EQ(inputLayers_.size(), parameters_.size()); - CHECK_EQ(inputLayers_[0]->getSize(), getSize()); - CHECK_EQ(parameters_[0]->getSize(), 5 * getSize()); - CHECK(parameters_[0]->isStatic()) - << "The parameter of DataNormLayer must be static"; - - weight_ = std::unique_ptr(new Weight(5, getSize(), parameters_[0])); - min_ = Matrix::create( - nullptr, /* height= */ 1, getSize(), /* trans= */ false, useGpu_); - rangeReciprocal_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - mean_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - stdReciprocal_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - decimalReciprocal_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - - min_->setData(weight_->getW()->getData()); - rangeReciprocal_->setData(weight_->getW()->getData() + getSize()); - mean_->setData(weight_->getW()->getData() + 2 * getSize()); - stdReciprocal_->setData(weight_->getW()->getData() + 3 * getSize()); - decimalReciprocal_->setData(weight_->getW()->getData() + 4 * getSize()); - - /* normalization strategy */ - if (config_.data_norm_strategy() == "z-score") { - mode_ = kZScore; - } else if (config_.data_norm_strategy() == "min-max") { - mode_ = kMinMax; - } else if (config_.data_norm_strategy() == "decimal-scaling") { - mode_ = kDecimalScaling; - } else { - LOG(FATAL) << "Unknown data normalization strategy: " - << config_.data_norm_strategy(); - } - - return true; -} - -void DataNormLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInput(0).getBatchSize(); - int size = getSize(); - reserveOutput(batchSize, size); - - const MatrixPtr inValue = getInputValue(0); - MatrixPtr outValue = getOutputValue(); - outValue->copyFrom(*inValue); - switch (mode_) { - case kZScore: { - outValue->addBias(*mean_, -1.0); - outValue->colScale(0, *outValue, *stdReciprocal_); - break; - } - case kMinMax: { - outValue->addBias(*min_, -1.0); - outValue->colScale(0, *outValue, *rangeReciprocal_); - break; - } - case kDecimalScaling: { - outValue->colScale(0, *outValue, *decimalReciprocal_); - break; - } - default: - LOG(FATAL) << "should not reach here"; - } -} - -void DataNormLayer::backward(const UpdateCallback& callback) { - // The parameter for DataNormLayer is static, and does not need to be updated - (void)callback; - - /* Calculate the input layers error */ - const MatrixPtr& outGrad = getOutputGrad(); - MatrixPtr inGrad = getInputGrad(0); - if (inGrad) { - switch (mode_) { - case kZScore: { - inGrad->addColScale(0, *outGrad, *stdReciprocal_); - break; - } - case kMinMax: { - inGrad->addColScale(0, *outGrad, *rangeReciprocal_); - break; - } - case kDecimalScaling: { - inGrad->addColScale(0, *outGrad, *decimalReciprocal_); - break; - } - default: { LOG(FATAL) << "should not reach here"; } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/DataNormLayer.h b/paddle/gserver/layers/DataNormLayer.h deleted file mode 100644 index 7ae67a877b488c8d197896b8b1e3e90057fbe1c9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/DataNormLayer.h +++ /dev/null @@ -1,62 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * @brief A layer for data normalization - * - Input: One and only one input layer is accepted. The input layer must - * be DataLayer with dense data type. - * - Output: The normalization of the input data - * - * Reference: - * LA Shalabi, Z Shaaban, B Kasasbeh. Data mining: A preprocessing engine - * - * Three data normalization methoeds are considered - * - z-score: y = (x-mean)/std - * - min-max: y = (x-min)/(max-min) - * - decimal-scaling: y = x/10^j, where j is the smallest integer such that - *max(|y|)<1 - */ - -class DataNormLayer : public Layer { - public: - enum NormalizationStrategy { kZScore = 0, kMinMax = 1, kDecimalScaling = 2 }; - - explicit DataNormLayer(const LayerConfig& config) : Layer(config) {} - - ~DataNormLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - protected: - int mode_; - std::unique_ptr weight_; - MatrixPtr min_; - MatrixPtr rangeReciprocal_; // 1/(max-min) - MatrixPtr mean_; - MatrixPtr stdReciprocal_; // 1/std - MatrixPtr decimalReciprocal_; // 1/10^j -}; -} // namespace paddle diff --git a/paddle/gserver/layers/DeConv3DLayer.cpp b/paddle/gserver/layers/DeConv3DLayer.cpp deleted file mode 100644 index db6d6e073c08c35c5a71b2b18ab0103d42ccd318..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/DeConv3DLayer.cpp +++ /dev/null @@ -1,220 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "DeConv3DLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(deconv3d, DeConv3DLayer); - -bool DeConv3DLayer::init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) { - if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; - // for Deconv, the dimension of Kernel is - // channel * output * depth * height * weigth - // Matrix storage format: (output * depth * height * weigth) x channel - for (int index = 0; index < config_.inputs().size(); ++index) { - M_.push_back(filterChannels_[index]); - K_.push_back(filterPixels_[index] * (numFilters_ / groups_[index])); - - // create a new weight - size_t height, width; - height = filterPixels_[index] * numFilters_; - width = filterChannels_[index]; - CHECK_EQ(parameters_[index]->getSize(), width * height); - Weight *w = new Weight(height, width, parameters_[index]); - weights_.emplace_back(w); - } - if (biasParameter_.get()) { - if (sharedBiases_) { - CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); - biases_ = - std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); - } else { - biases_ = - std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); - } - } - return true; -} - -size_t DeConv3DLayer::getSize() { - CHECK_NE(inputLayers_.size(), 0UL); - imgSizeW_.clear(); - imgSizeH_.clear(); - imgSizeD_.clear(); - N_.clear(); - NOut_.clear(); - size_t layerSize = 0; - for (size_t i = 0; i < inputLayers_.size(); ++i) { - imgSizeW_.push_back( - imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i], true)); - imgSizeH_.push_back(imageSize( - outputH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true)); - imgSizeD_.push_back(imageSize( - outputD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true)); - NOut_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]); - N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]); - CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize); - layerSize += NOut_[i] * numFilters_; - } - getOutput().setFrameHeight(imgSizeH_[0]); - getOutput().setFrameWidth(imgSizeW_[0]); - getOutput().setFrameDepth(imgSizeD_[0]); - return layerSize; -} - -void DeConv3DLayer::forward(PassType passType) { - Layer::forward(passType); - int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - int outWidth = getSize(); - resetOutput(batchSize, outWidth); - const MatrixPtr outMat = getOutputValue(); - - REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str()); - for (size_t i = 0; i != inputLayers_.size(); ++i) { - const MatrixPtr &inMat = getInputValue(i); - int M = M_[i]; - int N = N_[i]; - int K = K_[i]; - MatrixPtr wMat = weights_[i]->getW(); - Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); - for (int n = 0; n < batchSize; ++n) { - real *inData = inMat->getData() + n * inMat->getStride(); - for (int g = 0; g < groups_[i]; ++g) { - MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_); - MatrixPtr wMatSub = wMat->subMatrix(g * K, K); - MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K); - colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0); - inData += M * N; - } - colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(), - numFilters_, - imgSizeD_[i], - imgSizeH_[i], - imgSizeW_[i], - filterSizeZ_[i], - filterSizeY_[i], - filterSize_[i], - strideZ_[i], - strideY_[i], - stride_[i], - paddingZ_[i], - paddingY_[i], - padding_[i], - 1.0, - 1.0); - } - } - if (nullptr != this->biasParameter_) { - this->addBias(); - } - forwardActivation(); -} - -void DeConv3DLayer::backward(const UpdateCallback &callback) { - backwardActivation(); - int batchSize = getOutputGrad()->getHeight(); - if (biases_ && biases_->getWGrad()) { - bpropBiases(); - biases_->getParameterPtr()->incUpdate(callback); - } - REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str()); - for (size_t i = 0; i < inputLayers_.size(); ++i) { - if (weights_[i]->getWGrad() || this->needGradient_) { - int M = M_[i]; - int N = N_[i]; - int K = K_[i]; - Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); - const MatrixPtr &inMat = getInputValue(i); - for (int n = 0; n < batchSize; ++n) { - colBuf_->vol2Col( - getOutputGrad()->getData() + n * getOutputGrad()->getStride(), - numFilters_, - imgSizeD_[i], - imgSizeH_[i], - imgSizeW_[i], - filterSizeZ_[i], - filterSizeY_[i], - filterSize_[i], - strideZ_[i], - strideY_[i], - stride_[i], - paddingZ_[i], - paddingY_[i], - padding_[i]); - if (weights_[i]->getWGrad()) { - real *inData = inMat->getData() + n * inMat->getStride(); - for (int g = 0; g < groups_[i]; ++g) { - MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K); - MatrixPtr wGradMatSub = - weights_[i]->getWGrad()->subMatrix(g * K, K); - MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_); - wGradMatSub->mul( - *colBufDataSub, *(inMatSub->getTranspose()), 1.0, 1.0); - inData += M * N; - } - } - if (getInputGrad(i)) { - real *preGrad = - getInputGrad(i)->getData() + n * getInputGrad(i)->getStride(); - for (int g = 0; g < groups_[i]; ++g) { - MatrixPtr w = weights_[i]->getW()->subMatrix(g * K, K); - MatrixPtr outGradMat = colBuf_->subMatrix(g * K, K); - MatrixPtr inGradMatSub = - Matrix::create(preGrad, M, N, false, useGpu_); - inGradMatSub->mul(*(w->getTranspose()), *outGradMat, 1.0, 1.0); - preGrad += M * N; - } - } - } - weights_[i]->getParameterPtr()->incUpdate(callback); - } - } -} -void DeConv3DLayer::bpropWeights(int i) {} -void DeConv3DLayer::bpropData(int i) {} - -void DeConv3DLayer::bpropBiases() { - MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), - 1, - biases_->getWGrad()->getElementCnt(), - false, - useGpu_); - const MatrixPtr &outGradMat = getOutputGrad(); - - if (this->sharedBiases_) { - biases->collectSharedBias(*outGradMat, 1.0f); - } else { - biases->collectBias(*outGradMat, 1.0f); - } -} - -void DeConv3DLayer::addBias() { - MatrixPtr outMat = getOutputValue(); - MatrixPtr bias = Matrix::create(biases_->getW()->getData(), - 1, - biases_->getW()->getElementCnt(), - false, - useGpu_); - if (this->sharedBiases_) { - outMat->addSharedBias(*(bias), 1.0f); - } else { - outMat->addBias(*(bias), 1.0f); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/DeConv3DLayer.h b/paddle/gserver/layers/DeConv3DLayer.h deleted file mode 100644 index 13d1d07cf5cc6e2a6ea89768e29b1fe8cda5e81c..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/DeConv3DLayer.h +++ /dev/null @@ -1,52 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "ConvBaseLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief A subclass of deconvolution3D layer. - * This layer expands input and use matrix multiplication to - * calculate deconvolution3D operation. - */ -class DeConv3DLayer : public ConvBaseLayer { - public: - explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} - ~DeConv3DLayer() {} - bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); - - void forward(PassType passType); - void addBias(); - void backward(const UpdateCallback& callback); - void bpropBiases(); - void bpropData(int i); - void bpropWeights(int i); - size_t getSize(); - - protected: - // Figure out the dimensions for individual gemms. - IntV M_; /// numFilters_ / filter_group_; - IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ - IntV K_; /// outputD_ * outputH_ * outputW_ - IntV NOut_; - MatrixPtr colBuf_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/DetectionUtil.h b/paddle/gserver/layers/DetectionUtil.h deleted file mode 100644 index d6502fcf8fb12a434632876c25ac3ca23b87e60e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/DetectionUtil.h +++ /dev/null @@ -1,307 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include -#include "paddle/math/Matrix.h" - -using std::vector; -using std::pair; -using std::map; - -namespace paddle { - -template -struct BBoxBase { - BBoxBase(T xMin, T yMin, T xMax, T yMax) - : xMin(xMin), yMin(yMin), xMax(xMax), yMax(yMax), isDifficult(false) {} - - BBoxBase() {} - - T getWidth() const { return xMax - xMin; } - - T getHeight() const { return yMax - yMin; } - - T getCenterX() const { return (xMin + xMax) / 2; } - - T getCenterY() const { return (yMin + yMax) / 2; } - - T getArea() const { return getWidth() * getHeight(); } - - // coordinate of bounding box - T xMin; - T yMin; - T xMax; - T yMax; - // whether difficult object (e.g. object with heavy occlusion is difficult) - bool isDifficult; -}; - -struct NormalizedBBox : BBoxBase { - NormalizedBBox() : BBoxBase() {} -}; - -enum PermMode { kNCHWToNHWC, kNHWCToNCHW }; - -/** - * @brief First permute input maxtrix then append to output matrix - */ -size_t appendWithPermute(const Matrix& inMatrix, - size_t height, - size_t width, - size_t outTotalSize, - size_t outOffset, - size_t batchSize, - Matrix& outMatrix, - PermMode permMode); - -/** - * @brief First permute input maxtrix then decompose to output - */ -size_t decomposeWithPermute(const Matrix& inMatrix, - size_t height, - size_t width, - size_t totalSize, - size_t offset, - size_t batchSize, - Matrix& outMatrix, - PermMode permMode); - -/** - * @brief Compute jaccard overlap between two bboxes. - * @param bbox1 The first bbox - * @param bbox2 The second bbox - */ -real jaccardOverlap(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2); - -/** - * @brief Compute offset parameters between prior bbox and ground truth bbox - * and variances of prior bbox are considered - * @param priorBBox Input prior bbox - * @param priorBBoxVar Variance parameters of prior bbox - * @param gtBBox Groundtruth bbox - * @param outVec Output vector - */ -void encodeBBoxWithVar(const NormalizedBBox& priorBBox, - const vector& priorBBoxVar, - const NormalizedBBox& gtBBox, - vector& outVec); - -/** - * @brief Decode prior bbox with offset parameters - * and variances of prior bbox are considered - * @param priorBBox Prior bbox to be decoded - * @param priorBBoxVar Variance parameters of prior bbox - * @param locPredData Offset parameters - */ -NormalizedBBox decodeBBoxWithVar(const NormalizedBBox& priorBBox, - const vector& priorBBoxVar, - const vector& locPredData); - -/** - * @brief Extract bboxes from prior matrix, the layout is - * xmin1 | ymin1 | xmax1 | ymax1 | xmin1Var | ymin1Var | xmax1Var | ymax1Var ... - * @param priorData Matrix of prior value - * @param numBBoxes Number of bbox to be extracted - * @param bboxVec Append to the vector - */ -void getBBoxFromPriorData(const real* priorData, - const size_t numBBoxes, - vector& bboxVec); - -/** - * @brief Extract labels, scores and bboxes from detection matrix, the layout is - * imageId | label | score | xmin | ymin | xmax | ymax - * @param detectData Matrix of detection value - * @param numBBoxes Number of bbox to be extracted - * @param labelVec Label of bbox - * @param scoreVec Score of bbox - * @param bboxVec Append to the vector - */ -void getBBoxFromDetectData(const real* detectData, - const size_t numBBoxes, - vector& labelVec, - vector& scoreVec, - vector& bboxVec); - -/** - * @brief Extract variances from prior matrix, the layout is - * xmin1 | ymin1 | xmax1 | ymax1 | xmin1Var | ymin1Var | xmax1Var | ymax1Var ... - * @param priorData Matrix of prior value - * @param num Number to be extracted - * @param varVec Append to the vector - */ -void getBBoxVarFromPriorData(const real* priorData, - const size_t num, - vector>& varVec); - -/** - * @brief Extract bboxes from label matrix, the layout is - * class1_1 | xmin1_1 | ymin1_1 | xmax1_1 | ymax1_1 | difficult1_1 | ... - * @param labelData Matrix of label value - * @param numBBoxes Number to be extracted - * @param bboxVec Append to the vector - */ -void getBBoxFromLabelData(const real* labelData, - const size_t numBBoxes, - vector& bboxVec); - -/** -* @brief Match prior bbox to groundtruth bbox, the strategy is: -1. Find the most overlaped bbox pair (prior and groundtruth) -2. For rest of prior bboxes find the most overlaped groundtruth bbox -* @param priorBBoxes prior bbox -* @param gtBBoxes groundtruth bbox -* @param overlapThreshold Low boundary of overlap (judge whether matched) -* @param matchIndices For each prior bbox, groundtruth bbox index if matched -otherwise -1 -* @param matchOverlaps For each prior bbox, overap with all groundtruth bboxes -*/ -void matchBBox(const vector& priorBBoxes, - const vector& gtBBoxes, - real overlapThreshold, - vector* matchIndices, - vector* matchOverlaps); - -/** -* @brief Generate positive bboxes and negative bboxes, -|positive bboxes|/|negative bboxes| is negPosRatio -* @param priorValue Prior value -* @param numPriorBBoxes Number of prior bbox -* @param gtValue Groundtruth value -* @param gtStartPosPtr Since groundtruth value stored as sequence type, -this parameter indicates start position of each record -* @param seqNum Number of sequence -* @param maxConfScore Classification score for prior bbox, used to mine -negative examples -* @param batchSize Image number -* @param overlapThreshold Low boundary of overap -* @param negOverlapThreshold Upper boundary of overap (judge negative example) -* @param negPosRatio Control number of negative bboxes -* @param matchIndicesVecPtr Save indices of matched prior bbox -* @param negIndicesVecPtr Save indices of negative prior bbox -*/ -pair generateMatchIndices( - const Matrix& priorValue, - const size_t numPriorBBoxes, - const Matrix& gtValue, - const int* gtStartPosPtr, - const size_t seqNum, - const vector>& maxConfScore, - const size_t batchSize, - const real overlapThreshold, - const real negOverlapThreshold, - const size_t negPosRatio, - vector>* matchIndicesVecPtr, - vector>* negIndicesVecPtr); - -/** - * @brief Get max confidence score for each prior bbox - * @param confData Confidence scores, layout is - * class1 score | class2 score | ... | classN score ... - * @param batchSize Image number - * @param numPriorBBoxes Prior bbox number - * @param numClasses Classes number - * @param backgroundId Background id - * @param maxConfScoreVecPtr Ouput - */ -void getMaxConfidenceScores(const real* confData, - const size_t batchSize, - const size_t numPriorBBoxes, - const size_t numClasses, - const size_t backgroundId, - vector>* maxConfScoreVecPtr); - -template -bool sortScorePairDescend(const pair& pair1, - const pair& pair2); - -template <> -bool sortScorePairDescend(const pair& pair1, - const pair& pair2); - -/** - * @brief Do NMS for bboxes to remove duplicated bboxes - * @param bboxes BBoxes to apply NMS - * @param confScoreData Confidence scores - * @param classIdx Class to do NMS - * @param topK Number to keep - * @param confThreshold Low boundary of confidence score - * @param nmsThreshold Threshold of overlap - * @param numPriorBBoxes Total number of prior bboxes - * @param numClasses Total class number - * @param indices Indices of high quality bboxes - */ -void applyNMSFast(const vector& bboxes, - const real* confScoreData, - size_t classIdx, - size_t topK, - real confThreshold, - real nmsThreshold, - size_t numPriorBBoxes, - size_t numClasses, - vector* indices); - -/** - * @brief Get detection results which satify requirements - * @param numPriorBBoxes Prior bbox number - * @param numClasses Class number - * @param backgroundId Background class - * @param batchSize Image number - * @param confThreshold Threshold of class confidence - * @param nmsTopK Used in NMS operation to keep top k bbox - * @param nmsThreshold Used in NMS, threshold of overlap - * @param keepTopK How many bboxes keeped in an image - * @param allDecodedBBoxes Decoded bboxes for all images - * @param allDetectionIndices Save detection bbox indices - */ -size_t getDetectionIndices( - const real* confData, - const size_t numPriorBBoxes, - const size_t numClasses, - const size_t backgroundId, - const size_t batchSize, - const real confThreshold, - const size_t nmsTopK, - const real nmsThreshold, - const size_t keepTopK, - const vector>& allDecodedBBoxes, - vector>>* allDetectionIndices); - -/** - * @brief Get detection results - * @param confData Confidence scores - * @param numPriorBBoxes Prior bbox number - * @param numClasses Class number - * @param batchSize Image number - * @param allIndices Indices of predicted bboxes - * @param allDecodedBBoxes BBoxes decoded - * @param out Output matrix - * image number | label | confidence score | xMin | yMin | xMax | yMax - */ -void getDetectionOutput(const real* confData, - const size_t numKept, - const size_t numPriorBBoxes, - const size_t numClasses, - const size_t batchSize, - const vector>>& allIndices, - const vector>& allDecodedBBoxes, - Matrix& out); - -NormalizedBBox clipBBox(const NormalizedBBox& bbox); - -} // namespace paddle diff --git a/paddle/gserver/layers/DotProdLayer.cpp b/paddle/gserver/layers/DotProdLayer.cpp deleted file mode 100644 index 72b0c707b2131dc275ba604cd20ae0007c34a9a9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/DotProdLayer.cpp +++ /dev/null @@ -1,97 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief A layer for computing the dot product of two vectors. - * Input1: vector (batchSize * dim) - * Input2: vector (batchSize * dim) - * Output: a matrix: (batchSize * 1) - */ - -class DotProdLayer : public Layer { - public: - explicit DotProdLayer(const LayerConfig& config) : Layer(config) {} - - ~DotProdLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(dot_prod, DotProdLayer); - -bool DotProdLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2U); - CHECK_EQ(1UL, getSize()) - << "The output dimensionality of this layer should be fixed to 1."; - - return true; -} - -void DotProdLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV0->getHeight(); - CHECK_EQ(inV1->getHeight(), batchSize); - CHECK_EQ(inV0->getWidth(), inV1->getWidth()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, 1); - } - - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwDotProdTimer", getName().c_str()); - outV->sumOfProducts(*inV0, *inV1, 1, 0); - } -} - -void DotProdLayer::backward(const UpdateCallback& callback) { - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr outG = getOutputGrad(); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - - { - REGISTER_TIMER_INFO("BwDotProdTimer", getName().c_str()); - - if (inG0) { - inG0->addRowScale(0, *inV1, *outG); - } - - if (inG1) { - inG1->addRowScale(0, *inV0, *outG); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/EosIdCheckLayer.cpp b/paddle/gserver/layers/EosIdCheckLayer.cpp deleted file mode 100644 index 04400f2836581179849a4dd1c256bbddcc82530f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/EosIdCheckLayer.cpp +++ /dev/null @@ -1,50 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/utils/Logging.h" - -namespace paddle { -/** - * A layer for checking EOS for each sample: - * - output_id = (input_id == conf.eos_id) - * - * The result is stored in output_.ids. - * It is used by recurrent layer group. - */ -class EosIdCheckLayer : public Layer { - public: - explicit EosIdCheckLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override { - bool ret = Layer::init(layerMap, parameterMap); - CHECK_EQ(1UL, inputLayers_.size()); - return ret; - } - - void forward(PassType passType) override { - Layer::forward(passType); - - const Argument& input = getInput(0); - IVector::resizeOrCreate(output_.ids, input.ids->getSize(), useGpu_); - output_.ids->isEqualTo(*input.ids, config_.eos_id()); - } - - void backward(const UpdateCallback& callback) override {} -}; - -REGISTER_LAYER(eos_id, EosIdCheckLayer); - -} // namespace paddle diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp deleted file mode 100644 index 3a8478658249bfb0886e904aec43e50fe3618f79..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ExpandConvLayer.cpp +++ /dev/null @@ -1,248 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ExpandConvLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -DEFINE_bool(use_nnpack, - false, - "Whether to use nnpack for convolution calculation."); - -namespace paddle { - -/* - * The calculation of the exconvt(convolution transpose (deconv) operation) - * is a swap of forward and backward of the calculation of exconv. - * */ -REGISTER_LAYER(exconv, ExpandConvLayer); -REGISTER_LAYER(exconvt, ExpandConvLayer); - -inline bool isDepthwiseConv(int channels, int groups) { - return channels == groups; -} - -bool ExpandConvLayer::init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) { - /* Initialize the basic convolutional parent class */ - ConvBaseLayer::init(layerMap, parameterMap); - - int index = 0; - for (auto &inputConfig : config_.inputs()) { - const ConvConfig &conf = inputConfig.conv_conf(); - /* Consistent caffe mode for multiple input */ - caffeMode_ = conf.caffe_mode(); - - // create a new weight - size_t height, width; - height = filterPixels_[index] * filterChannels_[index]; - width = (!isDeconv_) ? numFilters_ : channels_[index]; - CHECK_EQ(parameters_[index]->getSize(), width * height); - Weight *w = new Weight(height, width, parameters_[index]); - weights_.emplace_back(w); - index++; - } - - if (biasParameter_.get()) { - if (sharedBiases_) { - CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); - biases_ = std::unique_ptr( - new Weight(1, numFilters_, biasParameter_, 0)); - } else { - biases_ = - std::unique_ptr(new Weight(1, getSize(), biasParameter_, 0)); - } - } - - getOutputSize(); - - size_t numInputs = config_.inputs_size(); - inputShape_.resize(numInputs); - filterShape_.resize(numInputs); - outputShape_.resize(numInputs); - - std::string convType; - std::string convGradInputType; - std::string convGradFilterType; - - for (int i = 0; i < config_.inputs_size(); i++) { - std::vector paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; - std::vector strides = {(size_t)strideY_[i], (size_t)stride_[i]}; - std::vector dilations = {(size_t)dilationY_[i], - (size_t)dilation_[i]}; - - bool useDilation = ((size_t)dilationY_[i] > 1 || (size_t)dilation_[i] > 1); - - // Convolution Layer uses the GemmConv function by default. - convType = "GemmConv"; - convGradInputType = "GemmConvGradInput"; - convGradFilterType = "GemmConvGradFilter"; - - // If depth wise convolution and useGpu == true - if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { - convType = "DepthwiseConv"; - convGradInputType = "DepthwiseConvGradInput"; - convGradFilterType = "DepthwiseConvGradFilter"; - } - - // If depth wise convolution and useGpu == false and ARM-NEON - if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { -#if defined(__ARM_NEON__) || defined(__ARM_NEON) - if ((filterSize_[i] == filterSizeY_[i]) && - (filterSize_[i] == 3 || filterSize_[i] == 4) && - (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2) && - !useDilation) { - convType = "NeonDepthwiseConv"; - } -#endif - } - - if (FLAGS_use_nnpack && !isDeconv_ && !useDilation) { - createFunction(forward_, - "NNPACKConv", - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", (size_t)groups_[i]) - .set("algo", std::string("auto"))); - } else { - createFunction(forward_, - !isDeconv_ ? convType : convGradInputType, - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("dilations", dilations) - .set("groups", (size_t)groups_[i])); - - createFunction(backward_, - !isDeconv_ ? convGradInputType : convType, - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("dilations", dilations) - .set("groups", (size_t)groups_[i])); - - createFunction(backward_, - convGradFilterType, - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("dilations", dilations) - .set("groups", (size_t)groups_[i])); - } - } - return true; -} - -size_t ExpandConvLayer::getOutputSize() { - CHECK_NE(inputLayers_.size(), 0UL); - size_t layerSize = ConvBaseLayer::calOutputSize(); - return layerSize; -} - -// i is the index of input layers -#define BACKWARD_INPUT(i, inputs, outputs) \ - backward_[2 * i]->calc(inputs, outputs) -#define BACKWARD_FILTER(i, inputs, outputs) \ - backward_[2 * i + 1]->calc(inputs, outputs) - -void ExpandConvLayer::forward(PassType passType) { - Layer::forward(passType); - - size_t batchSize = inputLayers_[0]->getOutputValue()->getHeight(); - resetOutput(batchSize, getOutputSize()); - - // Calculate the shape of the input, output, and filter. - for (size_t i = 0; i < inputLayers_.size(); ++i) { - inputShape_[i] = TensorShape({(size_t)batchSize, - (size_t)channels_[i], - (size_t)imgSizeH_[i], - (size_t)imgSizeW_[i]}); - filterShape_[i] = - TensorShape({(size_t)groups_[i], - !isDeconv_ ? (size_t)numFilters_ / groups_[i] - : (size_t)channels_[i] / groups_[i], - !isDeconv_ ? (size_t)channels_[i] / groups_[i] - : (size_t)numFilters_ / groups_[i], - (size_t)filterSizeY_[i], - (size_t)filterSize_[i]}); - outputShape_[i] = TensorShape({(size_t)batchSize, - (size_t)numFilters_, - (size_t)outputH_[i], - (size_t)outputW_[i]}); - } - - // Calculate the output value. - for (size_t i = 0; i < inputLayers_.size(); ++i) { - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(i), inputShape_[i]); - inputs.addArg(*weights_[i]->getW(), filterShape_[i]); - outputs.addArg(*getOutputValue(), - outputShape_[i], - !isDeconv_ && i == 0 ? ASSIGN_TO : ADD_TO); - - forward_[i]->calc(inputs, outputs); - } - - /* add the bias-vector */ - if (biases_.get()) { - output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_); - } - - /* activation */ - forwardActivation(); -} - -void ExpandConvLayer::backward(const UpdateCallback &callback) { - backwardActivation(); - - MatrixPtr outGrad = getOutputGrad(); - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_); - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - // Calculate the input grad and filter grad. - for (size_t i = 0; i < inputLayers_.size(); ++i) { - if (getInputGrad(i)) { - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getOutputGrad(), outputShape_[i]); - inputs.addArg(*weights_[i]->getW(), filterShape_[i]); - outputs.addArg(*getInputGrad(i), inputShape_[i], ADD_TO); - BACKWARD_INPUT(i, inputs, outputs); - } - - if (weights_[i]->getWGrad()) { - BufferArgs inputs; - BufferArgs outputs; - if (!isDeconv_) { - inputs.addArg(*getOutputGrad(), outputShape_[i]); - inputs.addArg(*getInputValue(i), inputShape_[i]); - } else { - inputs.addArg(*getInputValue(i), inputShape_[i]); - inputs.addArg(*getOutputGrad(), outputShape_[i]); - } - outputs.addArg(*weights_[i]->getWGrad(), filterShape_[i], ADD_TO); - BACKWARD_FILTER(i, inputs, outputs); - - /* Increasing the number of gradient */ - weights_[i]->getParameterPtr()->incUpdate(callback); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ExpandConvLayer.h b/paddle/gserver/layers/ExpandConvLayer.h deleted file mode 100644 index 6919ef71355a4c660b9ddd60bff75fee399cfaa9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ExpandConvLayer.h +++ /dev/null @@ -1,51 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "ConvBaseLayer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief A subclass of convolution layer. - * This layer expands input and use matrix multiplication to - * calculate convolution operation. - * - * The config file api is img_conv_layer. - */ - -class ExpandConvLayer : public ConvBaseLayer { - public: - explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {} - - ~ExpandConvLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override; - - size_t getOutputSize(); - - protected: - std::vector inputShape_; - std::vector filterShape_; - std::vector outputShape_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ExpandLayer.cpp b/paddle/gserver/layers/ExpandLayer.cpp deleted file mode 100644 index 6b5776754017bca8f8c14170ecfb4faa4109e0b5..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ExpandLayer.cpp +++ /dev/null @@ -1,133 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ExpandLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(expand, ExpandLayer); - -bool ExpandLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - CHECK_EQ(inputLayers_.size(), 2UL); - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - // which sequence type of input[0] - if (config_.trans_type() == "non-seq") { - type_ = kNonSeq; - } else if (config_.trans_type() == "seq") { - type_ = kSeq; - } else { - LOG(FATAL) << "Unknown trans_type: " << config_.trans_type(); - } - setNeedSequenceInfo(false); - return true; -} - -void ExpandLayer::forward(PassType passType) { - Layer::forward(passType); - // Expand layer should have exactly 2 input, one for data, one for size - CHECK_EQ(2U, inputLayers_.size()); - - // using two input: - // * first one for data; - // * second one only for sequence info - const Argument& shapeInput = getInput(1); - const Argument& dataInput = getInput(0); - size_t outputBatchSize = shapeInput.getBatchSize(); - auto startPositions = type_ ? shapeInput.subSequenceStartPositions - : shapeInput.sequenceStartPositions; - size_t numSequences = startPositions->getSize() - 1; - const int* starts = startPositions->getData(false); - - CHECK_EQ(starts[numSequences], shapeInput.getBatchSize()); - if (type_) { - // when trans_type = seq, input[1] must hasSubseq - CHECK_EQ(shapeInput.hasSubseq(), 1UL); - CHECK_EQ(dataInput.getNumSequences(), shapeInput.getNumSequences()); - } else { - CHECK_EQ(dataInput.getBatchSize(), shapeInput.getNumSequences()); - } - - // set output sequence info as shape sequence - output_.sequenceStartPositions = shapeInput.sequenceStartPositions; - if (shapeInput.hasSubseq()) { - output_.subSequenceStartPositions = shapeInput.subSequenceStartPositions; - } - - // reserve output: Expand output to batchsize of sequence data. - reserveOutput(outputBatchSize, dataInput.value->getWidth()); - - MatrixPtr inputValue = getInputValue(0); - MatrixPtr outputValue = getOutputValue(); - - ICpuGpuVector::resizeOrCreate(expandStartsPos_, outputBatchSize, false); - int* expandStarts = expandStartsPos_->getMutableData(false); - for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) { - int sequenceLength = starts[sequenceId + 1] - starts[sequenceId]; - for (int j = 0; j < sequenceLength; j++) { - expandStarts[starts[sequenceId] + j] = sequenceId; - } - } - - outputValue->copyByRowIndex(*inputValue, - *expandStartsPos_->getVector(useGpu_)); - - if (biases_.get() != NULL) { - outputValue->addBias(*(biases_->getW()), 1); - } -} - -void ExpandLayer::backward(const UpdateCallback& callback) { - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - if (!getInputGrad(0)) return; - MatrixPtr inputGrad = getInputGrad(0); - MatrixPtr outputGrad = getOutputGrad(); - auto cpuSeqStartPos = type_ ? getInput(1).subSequenceStartPositions - : getInput(1).sequenceStartPositions; - size_t numSequences = cpuSeqStartPos->getSize() - 1; - const int* starts = cpuSeqStartPos->getData(false); - - CHECK_EQ(inputGrad->getWidth(), outputGrad->getWidth()); - CHECK_EQ(outputGrad->getHeight(), (size_t)starts[numSequences]); - - AsyncGpuBlock asyncGpuBlock; - - // sum to get the grad - real scale = 1; - for (size_t sequenceId = 0; sequenceId < numSequences; sequenceId++) { - // TODO(Dangqingqing) optimization for GPU - int sequenceLength = starts[sequenceId + 1] - starts[sequenceId]; - if (sequenceLength == 0) { - // empty sequence - continue; - } - MatrixPtr copyData = inputGrad->subMatrix(sequenceId, 1); - copyData->collectBias( - *outputGrad->subMatrix(starts[sequenceId], sequenceLength), scale); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ExpandLayer.h b/paddle/gserver/layers/ExpandLayer.h deleted file mode 100644 index 06bd4ef05ee206628d981fee8e7eec3c91b18b7a..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ExpandLayer.h +++ /dev/null @@ -1,63 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * A layer for "Expand Dense data or (sequence data where the length of each - * sequence is one) to sequence data." - * - * It should have exactly 2 input, one for data, one for size: - * - first one for data - * - If ExpandLevel = kNonSeq: dense data - * - If ExpandLevel = kSeq: sequence data where the length of each sequence is - * one - * - second one only for sequence info - * - should be sequence data with or without sub-sequence. - * - * And the output size is the batch size(not instances) of second input. - * - * The config file api is expand_layer. - */ - -class ExpandLayer : public Layer { - protected: - std::unique_ptr biases_; - /// if input[0] is dense data, ExpandLevel=kNonSeq; - /// if input[0] is sequence data, ExpandLevel=kSeq - enum ExpandLevel { kNonSeq = 0, kSeq = 1 }; - /// store the ExpandLevel - int type_; - /// expanded sequenceStartPositions or subSequenceStartPositions - /// of input[1] - ICpuGpuVectorPtr expandStartsPos_; - - public: - explicit ExpandLayer(const LayerConfig& config) : Layer(config) {} - - ~ExpandLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/FactorizationMachineLayer.cpp b/paddle/gserver/layers/FactorizationMachineLayer.cpp deleted file mode 100644 index 1744faada2ebd9f2c88ba9a3952b6b2646729e3b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/FactorizationMachineLayer.cpp +++ /dev/null @@ -1,158 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "FactorizationMachineLayer.h" -#include -#include -#include "paddle/math/SparseMatrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(factorization_machine, FactorizationMachineLayer); - -bool FactorizationMachineLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - factorSize_ = config_.factor_size(); - - /* initialize the latentVectors_ */ - CHECK_EQ(inputLayers_.size(), 1UL); - size_t inputSize = inputLayers_[0]->getSize(); - CHECK_EQ(parameters_[0]->getSize(), inputSize * factorSize_); - latentVectors_ = std::unique_ptr( - new Weight(inputSize, factorSize_, parameters_[0])); - - return true; -} - -void FactorizationMachineLayer::forward(PassType passType) { - Layer::forward(passType); - - const MatrixPtr& inputV = getInputValue(0); - - size_t batchSize = inputV->getHeight(); - size_t outputSize = getSize(); - size_t inputSize = inputLayers_[0]->getSize(); - reserveOutput(batchSize, outputSize); - - MatrixPtr outV = getOutputValue(); - - Matrix::resizeOrCreate( - latentVectorsSquare_, inputSize, factorSize_, false, useGpu_); - Matrix::resizeOrCreate( - inputMulFactor_, batchSize, factorSize_, false, useGpu_); - Matrix::resizeOrCreate(tmpOut_, batchSize, factorSize_, false, useGpu_); - - REGISTER_TIMER_INFO("FmInputMulFactorTimer", getName().c_str()); - inputMulFactor_->mul(*inputV, *latentVectors_->getW()); - inputMulFactor_->square2(*tmpOut_); - outV->sumRows(*tmpOut_, 0.5, 0); - - if (dynamic_cast(inputV.get())) { - Matrix::resizeOrCreateSparseMatrix(inputSquare_, - inputV->getHeight(), - inputV->getWidth(), - inputV->getElementCnt(), - inputV->getValueType()); - inputSquare_->copyFrom(*inputV); - (dynamic_cast(inputSquare_.get()))->square2(); - } else { - Matrix::resizeOrCreate( - inputSquare_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); - inputV->square2(*inputSquare_); - } - latentVectors_->getW()->square2(*latentVectorsSquare_); - tmpOut_->mul(*inputSquare_, *latentVectorsSquare_); - outV->sumRows(*tmpOut_, -0.5, 1.0); - - /* activation */ { - REGISTER_TIMER_INFO("FmFwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void FactorizationMachineLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { backwardActivation(); } - - const MatrixPtr& inputV = getInputValue(0); - const MatrixPtr& oGrad = getOutputGrad(); - - Matrix::resizeOrCreate( - tmpSum_, 1, latentVectors_->getW()->getHeight(), false, useGpu_); - MatrixPtr tmpSumTrans = Matrix::create(tmpSum_->getRowBuf(0), - latentVectors_->getW()->getHeight(), - 1, - false, - useGpu_); - - /* Calculate the gradients of the latentVectors_ matrix */ - if (latentVectors_->getWGrad()) { - if (dynamic_cast(inputV.get())) { - Matrix::resizeOrCreateSparseMatrix(tmpInput_, - inputV->getHeight(), - inputV->getWidth(), - inputV->getElementCnt()); - - CpuSparseMatrix* sparseInputV = - dynamic_cast(inputV.get()); - CpuSparseMatrix* sparseInputSquare = - dynamic_cast(inputSquare_.get()); - CpuSparseMatrix* sparseTmpInput = - dynamic_cast(tmpInput_.get()); - sparseTmpInput->copyFrom(*sparseInputV); - - sparseTmpInput->rowScale(0, *sparseInputV, *oGrad); - latentVectors_->getWGrad()->mul( - *sparseTmpInput->getTranspose(), *inputMulFactor_, 1, 1); - sparseTmpInput->rowScale(0, *sparseInputSquare, *oGrad); - - Matrix::resizeOrCreate(negOnes_, 1, inputV->getHeight(), false, useGpu_); - negOnes_->zeroMem(); - negOnes_->add(-1); - tmpSum_->mul(*negOnes_, *sparseTmpInput, 1, 0); - } else { - Matrix::resizeOrCreate( - tmpInput_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); - - tmpInput_->rowScale(0, *inputV, *oGrad); - latentVectors_->getWGrad()->mul( - *tmpInput_->getTranspose(), *inputMulFactor_, 1, 1); - tmpInput_->rowScale(0, *inputSquare_, *oGrad); - - tmpSum_->sumCols(*tmpInput_, -1, 0); - } - - latentVectors_->getWGrad()->addRowScale( - 0, *latentVectors_->getW(), *tmpSumTrans); - - /* Increasing the number of gradient */ - latentVectors_->getParameterPtr()->incUpdate(callback); - } - - /* Calculate the input layers gradient */ - MatrixPtr inGrad = getInputGrad(0); - if (inGrad != NULL) { - inGrad->mul( - *inputMulFactor_, *latentVectors_->getW()->getTranspose(), 1, 1); - tmpSumTrans->sumRows(*latentVectorsSquare_, -1, 0); - inGrad->addColScale(0, *inputV, *tmpSum_); - inGrad->rowScale(0, *inGrad, *oGrad); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/FactorizationMachineLayer.h b/paddle/gserver/layers/FactorizationMachineLayer.h deleted file mode 100644 index 148abe238173dd44cd0fcf3f5cda732f70078706..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/FactorizationMachineLayer.h +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { -/** - * @brief The Factorization Machine models pairwise (order-2) feature - * interactions as inner product of the learned latent vectors corresponding - * to each input feature. - * - * The Factorization Machine can effectively capture feature interactions - * especially when the input is sparse. While in principle FM can model higher - * order feature interaction, in practice usually only order-2 feature - * interactions are considered. The Factorization Machine Layer here only - * computes the order-2 interations with the formula: - * - * \f[ - * y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j - * \f] - * - * The detailed calculation for forward and backward can be found at this paper: - * - * Factorization machines. - * - * The config file api is factorization_machine. - */ - -class FactorizationMachineLayer : public Layer { - protected: - // The latent vectors, shape: (size, factorSize_) - // Each row of the latentVectors_ matrix is the latent vector - // corresponding to one input feature dimension - std::unique_ptr latentVectors_; - // The hyperparameter that defines the dimensionality of the factorization - size_t factorSize_; - - private: - // Store the square values of the letent vectors matrix - MatrixPtr latentVectorsSquare_; - // Store the square values of input matrix - MatrixPtr inputSquare_; - // The result of input matrix * latent vector matrix that will be used in - // both forward and backward step - MatrixPtr inputMulFactor_; - // Store temporary calculation result - MatrixPtr tmpOut_; - MatrixPtr tmpSum_; - MatrixPtr tmpInput_; - // Negative identity matrix - MatrixPtr negOnes_; - - public: - explicit FactorizationMachineLayer(const LayerConfig& config) - : Layer(config) {} - ~FactorizationMachineLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/FeatureMapExpandLayer.cpp b/paddle/gserver/layers/FeatureMapExpandLayer.cpp deleted file mode 100644 index d95f0b9b3d13e8bff635373cb4d5705c2351bd97..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/FeatureMapExpandLayer.cpp +++ /dev/null @@ -1,155 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief A layer for expanding a batch of images to feature maps. - * Each data of the input is a 2 dimensional matrix. Each element of the matrix - * is replicated num_filters times to create a feature map with num_filters - * channels. - * - Input: Input one should be dense image data. - * - Output: expanded fature maps. - * \f[ - * y.row[i] = x.row[i \mod x.width], i = 0,1,..., (x.width * num\_filters - 1) - * \f] - * For example, num_filters = 4: - * @code - * x = [a1,a2; - * b1,b2] - * y = [a1, a2, a1, a2, a1, a2, a1, a2; - * b1, b2, b1, b2, b1, b2, b1, b2;] - * @endcode - */ - -class FeatureMapExpandLayer : public Layer { - private: - int numFilters_; - bool asRowVector_; - - public: - explicit FeatureMapExpandLayer(const LayerConfig& config) : Layer(config) {} - - ~FeatureMapExpandLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(featmap_expand, FeatureMapExpandLayer); - -bool FeatureMapExpandLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 1UL); - numFilters_ = config_.num_filters(); - asRowVector_ = config_.user_arg() != "as_col_vec"; - return true; -} - -void FeatureMapExpandLayer::forward(PassType passType) { - Layer::forward(passType); - MatrixPtr inputV = getInputValue(0); - size_t batchSize = getInput(0).getBatchSize(); - int imgSize = inputV->getWidth(); - resetOutput(batchSize, imgSize * numFilters_); - - MatrixPtr outputV = getOutputValue(); - - { - AsyncGpuBlock asyncGpuBlock; - if (asRowVector_) { - for (size_t i = 0; i < batchSize; i++) { - MatrixPtr outVTmp = - Matrix::create(outputV->getData() + i * imgSize * numFilters_, - numFilters_, - imgSize, - false, - useGpu_); - MatrixPtr inVTmp = Matrix::create( - inputV->getData() + i * imgSize, 1, imgSize, false, useGpu_); - outVTmp->addRowVector(*inVTmp); - } - } else { - for (size_t i = 0; i < batchSize; i++) { - MatrixPtr outVTmp = - Matrix::create(outputV->getData() + i * imgSize * numFilters_, - imgSize, - numFilters_, - false, - useGpu_); - MatrixPtr inVTmp = Matrix::create( - inputV->getData() + i * imgSize, imgSize, 1, false, useGpu_); - outVTmp->addColVector(*inVTmp); - } - } - } - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void FeatureMapExpandLayer::backward(const UpdateCallback& callback) { - MatrixPtr inGrad = getInputGrad(0); - if (NULL == inGrad) { - return; - } - MatrixPtr outGrad = getOutputGrad(); - size_t batchSize = getInput(0).getBatchSize(); - int imgSize = inGrad->getWidth(); - /* Do activation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - { - AsyncGpuBlock asyncGpuBlock; - if (asRowVector_) { - for (size_t i = 0; i < batchSize; i++) { - MatrixPtr outGradTmp = - Matrix::create(outGrad->getData() + i * imgSize * numFilters_, - numFilters_, - imgSize, - false, - useGpu_); - MatrixPtr inGradTmp = Matrix::create( - inGrad->getData() + i * imgSize, 1, imgSize, false, useGpu_); - inGradTmp->collectBias(*outGradTmp, 1); - } - } else { - for (size_t i = 0; i < batchSize; i++) { - MatrixPtr outGradTmp = - Matrix::create(outGrad->getData() + i * imgSize * numFilters_, - imgSize, - numFilters_, - false, - useGpu_); - MatrixPtr inGradTmp = Matrix::create( - inGrad->getData() + i * imgSize, imgSize, 1, false, useGpu_); - inGradTmp->sumRows(*outGradTmp, 1, 1); - } - } - } -} - -} // namespace paddle. diff --git a/paddle/gserver/layers/FullMatrixProjection.h b/paddle/gserver/layers/FullMatrixProjection.h deleted file mode 100644 index a27aa4a12327ac39ec3418a849b1230e13f759ee..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/FullMatrixProjection.h +++ /dev/null @@ -1,42 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "paddle/utils/Stat.h" - -#include "Projection.h" - -namespace paddle { - -/** - * FullMatrixProjection performs full matrix multiplication: - * \f[ - * out.row[i] += in.row[i] * weight - * \f] - * - * The config file api is full_matrix_projection. - */ -class FullMatrixProjection : public Projection { - public: - FullMatrixProjection(const ProjectionConfig& config, - const ParameterPtr& parameter, - bool useGpu); - virtual void forward(); - virtual void backward(const UpdateCallback& callback); - - protected: - std::unique_ptr weight_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/FullyConnectedLayer.cpp b/paddle/gserver/layers/FullyConnectedLayer.cpp deleted file mode 100644 index 21ffa01d95a460b4b6edc2b02d63c19b32d0b070..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/FullyConnectedLayer.cpp +++ /dev/null @@ -1,150 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "FullyConnectedLayer.h" -#include -#include -#include "paddle/math/SparseMatrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(fc, FullyConnectedLayer); - -bool FullyConnectedLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* initialize the weightList */ - CHECK(inputLayers_.size() == parameters_.size()); - for (size_t i = 0; i < inputLayers_.size(); i++) { - // Option the parameters - size_t height = inputLayers_[i]->getSize(); - size_t width = getSize(); - - // create a new weight - if (parameters_[i]->isSparse()) { - CHECK_LE(parameters_[i]->getSize(), width * height); - } else { - CHECK_EQ(parameters_[i]->getSize(), width * height); - } - Weight* w = new Weight(height, width, parameters_[i]); - - // append the new weight to the list - weights_.emplace_back(w); - } - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - - return true; -} - -void FullyConnectedLayer::prefetch() { - for (size_t i = 0; i != inputLayers_.size(); ++i) { - auto* sparseParam = - dynamic_cast(weights_[i]->getW().get()); - if (sparseParam) { - MatrixPtr input = getInputValue(i); - sparseParam->addRows(input); - } - } -} - -void FullyConnectedLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInput(0).getBatchSize(); - int size = getSize(); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, size); - } - - MatrixPtr outV = getOutputValue(); - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - auto input = getInput(i); - CHECK(input.value) << "The input of 'fc' layer must be matrix"; - REGISTER_TIMER_INFO("FwMulTimer", getName().c_str()); - i == 0 ? outV->mul(*input.value, *weights_[i]->getW(), 1, 0) - : outV->mul(*input.value, *weights_[i]->getW(), 1, 1); - } - - /* add the bias-vector */ - if (biases_.get() != NULL) { - REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); - outV->addBias(*(biases_->getW()), 1); - } - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void FullyConnectedLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - if (biases_ && biases_->getWGrad()) { - REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - bool syncFlag = hl_get_sync_flag(); - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - /* Calculate the W-gradient for the current layer */ - if (weights_[i]->getWGrad()) { - MatrixPtr input_T = getInputValue(i)->getTranspose(); - MatrixPtr oGrad = getOutputGrad(); - { - REGISTER_TIMER_INFO("GradMulTimer", getName().c_str()); - weights_[i]->getWGrad()->mul(*input_T, *oGrad, 1, 1); - } - } - - // If callback does not change value, backprop error asynchronously so that - // we can do the callback concurrently. - hl_set_sync_flag(false); - - /* Calculate the input layers error */ - MatrixPtr preGrad = getInputGrad(i); - if (NULL != preGrad) { - MatrixPtr weights_T = weights_[i]->getW()->getTranspose(); - REGISTER_TIMER_INFO("BpMulTimer", getName().c_str()); - preGrad->mul(*getOutputGrad(), *weights_T, 1, 1); - } - - hl_set_sync_flag(syncFlag); - { - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); - weights_[i]->getParameterPtr()->incUpdate(callback); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/FullyConnectedLayer.h b/paddle/gserver/layers/FullyConnectedLayer.h deleted file mode 100644 index e0f9d6ce55fbdf73e5507032c108c735bf04597b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/FullyConnectedLayer.h +++ /dev/null @@ -1,49 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { -/** - * A layer has full connections to all neurons in the previous layer. - * It computes an inner product with a set of learned weights, and - * (optionally) adds biases. - * - * The config file api is fc_layer. - */ - -class FullyConnectedLayer : public Layer { - protected: - WeightList weights_; - std::unique_ptr biases_; - - public: - explicit FullyConnectedLayer(const LayerConfig& config) : Layer(config) {} - ~FullyConnectedLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - Weight& getWeight(int idx) { return *weights_[idx]; } - - void prefetch() override; - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/GatedRecurrentLayer.cpp b/paddle/gserver/layers/GatedRecurrentLayer.cpp deleted file mode 100644 index 9d38849fdf97e6099e39384dd7e6546de9180462..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/GatedRecurrentLayer.cpp +++ /dev/null @@ -1,414 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "GatedRecurrentLayer.h" -#include "Layer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(gated_recurrent, GatedRecurrentLayer); - -bool GatedRecurrentLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!Layer::init(layerMap, parameterMap)) return false; - CHECK_EQ(1U, inputLayers_.size()); - CHECK_EQ(1U, parameters_.size()); - CHECK_EQ(getSize() * getSize() * 3, parameters_[0]->getSize()); - CHECK_EQ(getSize() * 3, biasParameter_->getSize()); - weight_.reset(new Weight(getSize(), getSize() * 3, parameters_[0])); - gateWeight_.reset(new Weight(getSize(), getSize() * 2, parameters_[0], 0)); - stateWeight_.reset(new Weight( - getSize(), getSize(), parameters_[0], 2 * getSize() * getSize())); - if (biasParameter_.get() != NULL) { - bias_.reset(new Weight(1, getSize() * 3, biasParameter_)); - } - - reversed_ = config_.reversed(); - activationGate_.reset(ActivationFunction::create(config_.active_gate_type())); - - GruCompute::init(config_); - useBatch_ = true; - - return true; -} - -void GatedRecurrentLayer::resetState() { - CHECK(!reversed_) << "state is not allowed for reversed gated " - "recurrent layer"; - Matrix::resizeOrCreate( - prevOutput_, 1, getSize(), /* trans= */ false, useGpu_); - prevOutput_->zeroMem(); - - // TODO(hedaoyuan): support prev_batch_state - CHECK(!FLAGS_prev_batch_state) << "Not supported"; - - useBatch_ = false; -} - -void GatedRecurrentLayer::setState(LayerStatePtr state) { - CHECK(state->value.size() == 1) - << "one matrix is expected for GatedRecurrentLayer state"; - prevOutput_->copyFrom(*(state->value[0])); -} - -LayerStatePtr GatedRecurrentLayer::getState() { - LayerStatePtr res = std::make_shared(); - res->value.push_back(prevOutput_->clone(0, 0, useGpu_)); - res->value[0]->copyFrom(*prevOutput_); - return res; -} - -void GatedRecurrentLayer::forward(PassType passType) { - REGISTER_TIMER_INFO("GruFwTimer", getName().c_str()); - Layer::forward(passType); - - const Argument& input = getInput(0); - CHECK(input.sequenceStartPositions); - int batchSize = input.getBatchSize(); - size_t numSequences = input.getNumSequences(); - resetOutput(batchSize, getSize()); - CHECK_EQ(getSize() * 3, input.value->getWidth()); - const int* starts = input.sequenceStartPositions->getData(false); - // batchSize = length of total frames in a batch (NOT size of mini-batch) - CHECK_EQ(starts[numSequences], batchSize); - - Matrix::resizeOrCreate(gate_.value, - /* height= */ batchSize, - getSize() * 3, - /* trans= */ false, - useGpu_); - Matrix::resizeOrCreate(resetOutput_.value, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - - if (useBatch_) { - forwardBatch(batchSize, numSequences, starts, input.value); - } else { - forwardSequence(batchSize, numSequences, starts, input.value); - } -} - -void GatedRecurrentLayer::backward(const UpdateCallback& callback) { - REGISTER_TIMER_INFO("GruBwTimer", getName().c_str()); - const Argument& input = getInput(0); - CHECK(input.sequenceStartPositions); - int batchSize = input.getBatchSize(); - const int* starts = input.sequenceStartPositions->getData(false); - size_t numSequences = input.getNumSequences(); - - Matrix::resizeOrCreate(gate_.grad, - /* height= */ batchSize, - getSize() * 3, - /* trans= */ false, - useGpu_); - Matrix::resizeOrCreate(resetOutput_.grad, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - - if (useBatch_) { - backwardBatch(batchSize, input.grad); - } else { - backwardSequence(batchSize, numSequences, starts, input.grad); - } - - if (bias_) { - bias_->getParameterPtr()->incUpdate(callback); - } - - weight_->getParameterPtr()->incUpdate(callback); -} - -void GatedRecurrentLayer::forwardSequence(int batchSize, - size_t numSequences, - const int* starts, - MatrixPtr inputValue) { - REGISTER_TIMER_INFO("GruFwSequenceTime", getName().c_str()); - gate_.value->assign(*inputValue); - if (bias_) { - gate_.value->addBias(*(bias_->getW()), 1); - } - - hl_gru_value gruValue; - gruValue.gateWeight = (gateWeight_->getW())->getData(); - gruValue.stateWeight = (stateWeight_->getW())->getData(); - gruValue.gateValue = gate_.value->getData(); - gruValue.resetOutputValue = resetOutput_.value->getData(); - gruValue.outputValue = output_.value->getData(); - gruValue.prevOutValue = nullptr; - - if (reversed_) { - gruValue.gateValue += (batchSize - 1) * getSize() * 3; - gruValue.resetOutputValue += (batchSize - 1) * getSize(); - gruValue.outputValue += (batchSize - 1) * getSize(); - } - - auto nextFrame = [&gruValue](bool reversed, int frameSize) { - gruValue.prevOutValue = gruValue.outputValue; - if (!reversed) { - gruValue.gateValue += frameSize * 3; - gruValue.resetOutputValue += frameSize; - gruValue.outputValue += frameSize; - } else { - gruValue.gateValue -= frameSize * 3; - gruValue.resetOutputValue -= frameSize; - gruValue.outputValue -= frameSize; - } - }; - - if (!reversed_) { - if (prevOutput_) { - gruValue.prevOutValue = prevOutput_->getData(); - } - } - AsyncGpuBlock asyncGpuBlock; - for (size_t n = 0; n < numSequences; ++n) { - int length; - if (!reversed_) { - length = starts[n + 1] - starts[n]; - } else { - length = starts[numSequences - n] - starts[numSequences - n - 1]; - } - for (int l = 0; l < length; ++l) { - if (useGpu_) { - GruCompute::forward<1>(gruValue, getSize()); - } else { - GruCompute::forward<0>(gruValue, getSize()); - } - - nextFrame(reversed_, getSize()); - } - if (!reversed_) { - if (!prevOutput_) gruValue.prevOutValue = nullptr; - } else { - gruValue.prevOutValue = nullptr; - } - } - - if (!reversed_) { - if (prevOutput_) { - prevOutput_->assign(*output_.value->subMatrix(batchSize - 1, 1)); - } - } -} - -void GatedRecurrentLayer::backwardSequence(int batchSize, - size_t numSequences, - const int* starts, - MatrixPtr inputGrad) { - REGISTER_TIMER_INFO("GruBwSequenceTime", getName().c_str()); - - hl_gru_value gruValue; - gruValue.gateWeight = (gateWeight_->getW())->getData(); - gruValue.stateWeight = (stateWeight_->getW())->getData(); - gruValue.gateValue = gate_.value->getData(); - gruValue.resetOutputValue = resetOutput_.value->getData(); - gruValue.outputValue = output_.value->getData(); - - hl_gru_grad gruGrad; - gruGrad.gateWeightGrad = - (gateWeight_->getWGrad() ? gateWeight_->getWGrad()->getData() : nullptr); - gruGrad.stateWeightGrad = - (stateWeight_->getWGrad() ? stateWeight_->getWGrad()->getData() - : nullptr); - gruGrad.gateGrad = gate_.grad->getData(); - gruGrad.resetOutputGrad = resetOutput_.grad->getData(); - gruGrad.outputGrad = output_.grad->getData(); - - if (!reversed_) { - gruValue.gateValue += (batchSize - 1) * getSize() * 3; - gruValue.resetOutputValue += (batchSize - 1) * getSize(); - gruValue.outputValue += (batchSize - 1) * getSize(); - gruGrad.gateGrad += (batchSize - 1) * getSize() * 3; - gruGrad.resetOutputGrad += (batchSize - 1) * getSize(); - gruGrad.outputGrad += (batchSize - 1) * getSize(); - gruValue.prevOutValue = gruValue.outputValue - getSize(); - gruGrad.prevOutGrad = gruGrad.outputGrad - getSize(); - } else { - gruValue.prevOutValue = gruValue.outputValue + getSize(); - gruGrad.prevOutGrad = gruGrad.outputGrad + getSize(); - } - - auto nextFrame = [&gruValue, &gruGrad](bool reversed, int frameSize) { - if (reversed) { - gruValue.gateValue += frameSize * 3; - gruValue.resetOutputValue += frameSize; - gruValue.outputValue += frameSize; - gruGrad.gateGrad += frameSize * 3; - gruGrad.resetOutputGrad += frameSize; - gruGrad.outputGrad += frameSize; - gruValue.prevOutValue = gruValue.outputValue + frameSize; - gruGrad.prevOutGrad = gruGrad.outputGrad + frameSize; - } else { - gruValue.gateValue -= frameSize * 3; - gruValue.resetOutputValue -= frameSize; - gruValue.outputValue -= frameSize; - gruGrad.gateGrad -= frameSize * 3; - gruGrad.resetOutputGrad -= frameSize; - gruGrad.outputGrad -= frameSize; - gruValue.prevOutValue = gruValue.outputValue - frameSize; - gruGrad.prevOutGrad = gruGrad.outputGrad - frameSize; - } - }; - - { - AsyncGpuBlock asyncGpuBlock; - for (size_t n = 0; n < numSequences; ++n) { - int length; - if (reversed_) { - length = starts[n + 1] - starts[n]; - } else { - length = starts[numSequences - n] - starts[numSequences - n - 1]; - } - for (int l = 0; l < length; ++l) { - if (l == length - 1) { - gruValue.prevOutValue = nullptr; - gruGrad.prevOutGrad = nullptr; - } - if (useGpu_) { - GruCompute::backward<1>(gruValue, gruGrad, getSize()); - } else { - GruCompute::backward<0>(gruValue, gruGrad, getSize()); - } - nextFrame(reversed_, getSize()); - } - } - } - - if (inputGrad) { - inputGrad->add(*gate_.grad); - } - if (bias_ && bias_->getWGrad()) { - bias_->getWGrad()->collectBias(*gate_.grad, 1); - } -} - -void GatedRecurrentLayer::forwardBatch(int batchSize, - size_t numSequences, - const int* starts, - MatrixPtr inputValue) { - REGISTER_TIMER_INFO("GruFwBatchTime", getName().c_str()); - hl_gru_value gruValue; - gruValue.gateWeight = (gateWeight_->getW())->getData(); - gruValue.stateWeight = (stateWeight_->getW())->getData(); - - if (!batchValue_) { - batchValue_.reset(new SequenceToBatch(useGpu_)); - } - batchValue_->resizeOrCreateBatch(batchSize, numSequences, starts, reversed_); - - batchValue_->resizeOrCreate(*output_.value); - batchValue_->copy(*inputValue, *gate_.value, /* seq2batch */ true); - if (bias_) { - gate_.value->addBias(*(bias_->getW()), 1); - } - - { - int numBatch = batchValue_->getNumBatch(); - int curBatchSize = 0; - AsyncGpuBlock asyncGpuBlock; - for (int n = 0; n < numBatch; n++) { - MatrixPtr outputValueTmp = batchValue_->getBatchValue(n); - gruValue.outputValue = outputValueTmp->getData(); - gruValue.gateValue = - (batchValue_->getBatchValue(*gate_.value, n))->getData(); - gruValue.resetOutputValue = - (batchValue_->getBatchValue(*resetOutput_.value, n))->getData(); - - curBatchSize = outputValueTmp->getHeight(); - gruValue.prevOutValue = - (n == 0 - ? nullptr - : (batchValue_->getBatchValue(n - 1, curBatchSize))->getData()); - - { - if (useGpu_) { - GruCompute::forward<1>(gruValue, getSize(), curBatchSize); - } else { - GruCompute::forward<0>(gruValue, getSize(), curBatchSize); - } - } - } - } - { batchValue_->copyBackSeq(*output_.value); } -} - -void GatedRecurrentLayer::backwardBatch(int batchSize, MatrixPtr inputGrad) { - REGISTER_TIMER_INFO("GruBwBatchTime", getName().c_str()); - hl_gru_value gruValue; - gruValue.gateWeight = (gateWeight_->getW())->getData(); - gruValue.stateWeight = (stateWeight_->getW())->getData(); - - hl_gru_grad gruGrad; - gruGrad.gateWeightGrad = - (gateWeight_->getWGrad() ? gateWeight_->getWGrad()->getData() : nullptr); - gruGrad.stateWeightGrad = - (stateWeight_->getWGrad() ? stateWeight_->getWGrad()->getData() - : nullptr); - - if (!batchGrad_) { - batchGrad_.reset(new SequenceToBatch(useGpu_)); - } - batchGrad_->shareIndexWith(*batchValue_); - - { batchGrad_->copyFromSeq(*output_.grad); } - - { - int numBatch = batchGrad_->getNumBatch(); - int batchSize = 0; - AsyncGpuBlock asyncGpuBlock; - for (int n = (int)numBatch - 1; n >= 0; n--) { - gruValue.gateValue = - (batchGrad_->getBatchValue(*gate_.value, n))->getData(); - gruValue.resetOutputValue = - (batchGrad_->getBatchValue(*resetOutput_.value, n))->getData(); - - MatrixPtr outputGradTmp = batchGrad_->getBatchValue(n); - gruGrad.outputGrad = outputGradTmp->getData(); - gruGrad.gateGrad = (batchGrad_->getBatchValue(*gate_.grad, n))->getData(); - gruGrad.resetOutputGrad = - (batchGrad_->getBatchValue(*resetOutput_.grad, n))->getData(); - - { - batchSize = outputGradTmp->getHeight(); - gruValue.prevOutValue = - (n == 0 - ? nullptr - : (batchValue_->getBatchValue(n - 1, batchSize))->getData()); - gruGrad.prevOutGrad = - (n == 0 ? nullptr - : (batchGrad_->getBatchValue(n - 1, batchSize))->getData()); - - if (useGpu_) { - GruCompute::backward<1>(gruValue, gruGrad, getSize(), batchSize); - } else { - GruCompute::backward<0>(gruValue, gruGrad, getSize(), batchSize); - } - } - } - } - - if (inputGrad) { - batchGrad_->add(*inputGrad, *gate_.grad, /* seq2batch */ false); - } - if (bias_ && bias_->getWGrad()) { - bias_->getWGrad()->collectBias(*gate_.grad, /* scale */ 1); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/GatedRecurrentLayer.h b/paddle/gserver/layers/GatedRecurrentLayer.h deleted file mode 100644 index 46508dc977bf1a6fd33dc1fb024bd1aed36a0ff3..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/GatedRecurrentLayer.h +++ /dev/null @@ -1,100 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "GruCompute.h" -#include "Layer.h" -#include "SequenceToBatch.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief Please refer to "Junyoung Chung, Empirical Evaluation - * of Gated Recurrent Neural Networks on Sequence Modeling". - * - * GatedRecurrentLayer takes 1 input layer with size * 3. - * Input layer is diveded into 3 equal parts: (xz_t, xr_t, xi_t). - * parameter and biasParameter is also diveded into 3 equal parts: - * - parameter consists of (U_z, U_r, U) - * - baisParameter consists of (bias_z, bias_r, bias_o) - * - * \f[ - * update \ gate: z_t = actGate(xz_t + U_z * h_{t-1} + bias_z) \\ - * reset \ gate: r_t = actGate(xr_t + U_r * h_{t-1} + bias_r) \\ - * output \ candidate: {h}_t = actNode(xi_t + U * dot(r_t, h_{t-1}) + bias_o) \\ - * hidden \ activation: h_t = dot((1-z_t), h_{t-1}) + dot(z_t, {h}_t) \\ - * \f] - * - * @note - * - dot denotes "element-wise multiplication". - * - actNode is defined by config active_type - * - actGate is defined by config actvie_gate_type - * - * The config file is grumemory. - */ - -class GatedRecurrentLayer : public Layer, public GruCompute { - public: - explicit GatedRecurrentLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - - void backward(const UpdateCallback& callback) override; - - void resetState() override; - - void setState(LayerStatePtr state) override; - - LayerStatePtr getState() override; - - protected: - void forwardSequence(int batchSize, - size_t numSequences, - const int* starts, - MatrixPtr inputValue); - void backwardSequence(int batchSize, - size_t numSequences, - const int* starts, - MatrixPtr inputGrad); - - void forwardBatch(int batchSize, - size_t numSequences, - const int* starts, - MatrixPtr inputValue); - void backwardBatch(int batchSize, MatrixPtr inputGrad); - - protected: - std::unique_ptr weight_; - std::unique_ptr gateWeight_; - std::unique_ptr stateWeight_; - std::unique_ptr bias_; - - Argument gate_; - Argument resetOutput_; - - bool reversed_; - bool useBatch_; - std::unique_ptr batchValue_; - std::unique_ptr batchGrad_; - std::unique_ptr activationGate_; - - MatrixPtr prevOutput_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/GruCompute.cpp b/paddle/gserver/layers/GruCompute.cpp deleted file mode 100644 index 48ddbc413e6c915be6e86704f96e919932ca2970..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/GruCompute.cpp +++ /dev/null @@ -1,54 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "GruCompute.h" -#include "hl_recurrent_apply.cuh" -#include "paddle/function/GruFunctor.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -void GruCompute::init(LayerConfig &config) { - activeNode_ = hlActiveType(config.active_type()); - activeGate_ = hlActiveType(config.active_gate_type()); -} - -template <> -void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) { - GruFunctor::compute(hppl::forward::gru_resetOutput(), - hppl::forward::gru_finalOutput(), - value, - frameSize, - batchSize, - activeNode_, - activeGate_); -} - -template <> -void GruCompute::backward<0>(hl_gru_value value, - hl_gru_grad grad, - int frameSize, - int batchSize) { - GruGradFunctor::compute( - hppl::backward::gru_stateGrad(), - hppl::backward::gru_resetGrad(), - value, - grad, - frameSize, - batchSize, - activeNode_, - activeGate_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/GruCompute.h b/paddle/gserver/layers/GruCompute.h deleted file mode 100644 index 50006325ce9969c4941aaf28604260f0aeb9b97a..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/GruCompute.h +++ /dev/null @@ -1,41 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ModelConfig.pb.h" -#include "hl_gpu.h" -#include "paddle/utils/Common.h" - -namespace paddle { - -class GruCompute { - public: - void init(LayerConfig &config); - - template - void forward(hl_gru_value value, int frameSize, int batchSize = 1); - - template - void backward(hl_gru_value value, - hl_gru_grad grad, - int frameSize, - int batchSize = 1); - - public: - hl_activation_mode_t activeNode_; - hl_activation_mode_t activeGate_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/GruStepLayer.cpp b/paddle/gserver/layers/GruStepLayer.cpp deleted file mode 100644 index 114f287411c2fccbc08b7da4c05462967c81b268..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/GruStepLayer.cpp +++ /dev/null @@ -1,177 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "GruCompute.h" -#include "Layer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief GruStepLayer is like GatedRecurrentLayer, but used in recurrent - * layer group. GruStepLayer takes 2 input layer. - * - input[0] with size * 3 and diveded into 3 equal parts: (xz_t, xr_t, xi_t). - * - input[1] with size: {prev_out}. - * - * parameter and biasParameter is also diveded into 3 equal parts: - * - parameter consists of (U_z, U_r, U) - * - baisParameter consists of (bias_z, bias_r, bias_o) - * - * \f[ - * update \ gate: z_t = actGate(xz_t + U_z * prev_out + bias_z) \\ - * reset \ gate: r_t = actGate(xr_t + U_r * prev_out + bias_r) \\ - * output \ candidate: {h}_t = actNode(xi_t + U * dot(r_t, prev_out) + bias_o) - * \\ - * output: h_t = dot((1-z_t), prev_out) + dot(z_t, prev_out) - * \f] - * - * @note - * - dot denotes "element-wise multiplication". - * - actNode is defined by config active_type - * - actGate is defined by config actvie_gate_type - * - * The config file api if gru_step_layer. - */ -class GruStepLayer : public Layer, public GruCompute { - protected: - Argument gate_; - Argument resetOutput_; - std::unique_ptr weight_; - std::unique_ptr bias_; - - public: - explicit GruStepLayer(const LayerConfig& config) : Layer(config) {} - - ~GruStepLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(gru_step, GruStepLayer); - -bool GruStepLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!Layer::init(layerMap, parameterMap)) return false; - CHECK_EQ(2U, inputLayers_.size()); - - CHECK_EQ(getSize() * getSize() * 3, parameters_[0]->getSize()); - weight_.reset(new Weight(getSize(), getSize() * 3, parameters_[0])); - - if (biasParameter_.get() != NULL) { - CHECK_EQ(getSize() * 3, biasParameter_->getSize()); - bias_.reset(new Weight(1, getSize() * 3, biasParameter_)); - } - - GruCompute::init(config_); - return true; -} - -void GruStepLayer::forward(PassType passType) { - REGISTER_TIMER_INFO("GruStepFwTime", getName().c_str()); - Layer::forward(passType); - - const Argument& input = getInput(0); - const Argument& prevOutput = getInput(1); - CHECK_EQ(getSize() * 3, input.value->getWidth()); - CHECK_EQ(getSize(), prevOutput.value->getWidth()); - - int batchSize = input.getBatchSize(); - resetOutput(batchSize, getSize()); - resetSpecifyOutput(gate_, - batchSize, - getSize() * 3, - /* isValueClean */ false, - /* isGradClean */ false); - resetSpecifyOutput(resetOutput_, - batchSize, - getSize(), - /* isValueClean */ false, - /* isGradClean */ false); - gate_.value->assign(*input.value); - if (bias_) { - gate_.value->addBias(*(bias_->getW()), 1); - } - - hl_gru_value gruValue; - gruValue.gateWeight = weight_->getW()->getData(); - gruValue.stateWeight = weight_->getW()->getData() + getSize() * getSize() * 2; - gruValue.gateValue = gate_.value->getData(); - gruValue.resetOutputValue = resetOutput_.value->getData(); - gruValue.outputValue = output_.value->getData(); - gruValue.prevOutValue = prevOutput.value->getData(); - - if (useGpu_) { - GruCompute::forward<1>(gruValue, getSize(), batchSize); - } else { - GruCompute::forward<0>(gruValue, getSize(), batchSize); - } -} - -void GruStepLayer::backward(const UpdateCallback& callback) { - REGISTER_TIMER_INFO("GruStepBwTime", getName().c_str()); - - const Argument& input = getInput(0); - const Argument& prevOutput = getInput(1); - int batchSize = input.getBatchSize(); - - hl_gru_value gruValue; - gruValue.gateWeight = weight_->getW()->getData(); - gruValue.stateWeight = weight_->getW()->getData() + getSize() * getSize() * 2; - gruValue.gateValue = gate_.value->getData(); - gruValue.resetOutputValue = resetOutput_.value->getData(); - gruValue.outputValue = output_.value->getData(); - gruValue.prevOutValue = prevOutput.value->getData(); - - hl_gru_grad gruGrad; - gruGrad.gateWeightGrad = - (weight_->getWGrad() ? weight_->getWGrad()->getData() : nullptr); - gruGrad.stateWeightGrad = - (weight_->getWGrad() - ? weight_->getWGrad()->getData() + getSize() * getSize() * 2 - : nullptr); - - gruGrad.gateGrad = gate_.grad->getData(); - gruGrad.resetOutputGrad = resetOutput_.grad->getData(); - gruGrad.outputGrad = output_.grad->getData(); - if (prevOutput.grad) { - gruGrad.prevOutGrad = prevOutput.grad->getData(); - } else { - gruGrad.prevOutGrad = nullptr; - } - - if (useGpu_) { - GruCompute::backward<1>(gruValue, gruGrad, getSize(), batchSize); - } else { - GruCompute::backward<0>(gruValue, gruGrad, getSize(), batchSize); - } - - if (input.grad) { - input.grad->add(*gate_.grad); - } - - if (bias_ && bias_->getWGrad()) { - bias_->getWGrad()->collectBias(*gate_.grad, 1); - } - - if (bias_) { - bias_->getParameterPtr()->incUpdate(callback); - } - weight_->getParameterPtr()->incUpdate(callback); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp b/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp deleted file mode 100644 index 3e720f179ee66baa73f40b8f5f19bfb4090831c0..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp +++ /dev/null @@ -1,240 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "HierarchicalSigmoidLayer.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -REGISTER_LAYER(hsigmoid, HierarchicalSigmoidLayer); - -bool HierarchicalSigmoidLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK(config_.has_num_classes()) << "num_classes must be specifed in config"; - numClasses_ = config_.num_classes(); - CHECK_GE(numClasses_, (size_t)2); - codeLength_ = findLastSet(numClasses_ - 1); - - size_t height = numClasses_ - 1; - - /* initialize the weightList */ - // The last input layer is for label - CHECK(!parameters_.back()); - for (size_t i = 0; i < inputLayers_.size() - 1; i++) { - size_t width = inputLayers_[i]->getSize(); - // create a new weight - CHECK_EQ(parameters_[i]->getSize(), width * height); - Weight* w = new Weight(height, width, parameters_[i]); - - // append the new weight to the list - weights_.emplace_back(w); - } - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - CHECK_EQ(biasParameter_->getSize(), numClasses_ - 1); - biases_.reset(new Weight(1, numClasses_ - 1, biasParameter_)); - } - - return true; -} - -void HierarchicalSigmoidLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(0)->getHeight(); - int size = getSize(); - reserveOutput(batchSize, size); - Matrix::resizeOrCreate(preOutput_.value, - batchSize, - codeLength_, - /* trans */ false, - false); - Matrix::resizeOrCreate(preOutput_.grad, - batchSize, - codeLength_, - /* trans */ false, - false); - IVectorPtr label = getInput(*getLabelLayer()).ids; - preOutput_.value->zeroMem(); - - if (useGpu_) { - Matrix::resizeOrCreate(cpuOutput_, - output_.value->getHeight(), - output_.value->getWidth(), - /* trans */ false, - false); - IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); - cpuLabel_->copyFrom(*label); - cpuOutput_->copyFrom(*output_.value); - } else { - cpuOutput_ = output_.value; - cpuLabel_ = label; - } - /* add the bias-vector */ - if (biases_.get() != NULL) { - if (useGpu_) { - Matrix::resizeOrCreate(cpuBias_, - 1, - numClasses_ - 1, - /* trans */ false, - false); - cpuBias_->copyFrom(*biases_->getW()); - } else { - cpuBias_ = biases_->getW(); - } - preOutput_.value->addByBitCode(numClasses_, *cpuLabel_, *cpuBias_); - } - for (size_t i = 0; i < inputLayers_.size() - 1; ++i) { - MatrixPtr input = getInputValue(i); - if (useGpu_) { - Matrix::resizeOrCreate(cpuInput_, - input->getHeight(), - input->getWidth(), - /* trans */ false, - false); - Matrix::resizeOrCreate(cpuWeight_, - weights_[i]->getW()->getHeight(), - weights_[i]->getW()->getWidth(), - /* trans */ false, - false); - cpuInput_->copyFrom(*input); - cpuWeight_->copyFrom(*weights_[i]->getW()); - } else { - cpuInput_ = input; - cpuWeight_ = weights_[i]->getW(); - } - preOutput_.value->mulByBitCode( - numClasses_, *cpuLabel_, *cpuWeight_, *cpuInput_); - } - // keep consistent with the clipping in the following softrelu - preOutput_.value->clip(-40.0, 40.0); - preOutput_.value->sumByBitCode(numClasses_, - *cpuLabel_, - *cpuOutput_, - -1); // scaleSum - preOutput_.value->softrelu(*preOutput_.value); - MatrixPtr sum = Matrix::create(batchSize, 1, /* trans= */ false, false); - preOutput_.value->rowSum(*sum); - cpuOutput_->add(*sum); - if (useGpu_) { - output_.value->copyFrom(*cpuOutput_); - } else { - output_.value = cpuOutput_; - } -} - -void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { - IVectorPtr label = getInput(*getLabelLayer()).ids; - if (useGpu_) { - IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); - cpuLabel_->copyFrom(*label); - } else { - cpuLabel_ = label; - } - preOutput_.grad->one(); - preOutput_.grad->softreluDerivative(*preOutput_.value); - preOutput_.grad->subByBitCode(numClasses_, *cpuLabel_); - - if (biases_ && biases_->getWGrad()) { - MatrixPtr biases_grad = biases_->getWGrad(); - if (useGpu_) { - Matrix::resizeOrCreate(cpuBias_, - 1, - numClasses_ - 1, - /* trans */ false, - false); - cpuBias_->copyFrom(*biases_grad); - } else { - cpuBias_ = biases_grad; - } - preOutput_.grad->addByBitCodeBackward(numClasses_, *cpuLabel_, *cpuBias_); - if (useGpu_) { - biases_grad->copyFrom(*cpuBias_); - } else { - biases_grad = cpuBias_; - } - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - for (size_t i = 0; i < inputLayers_.size() - 1; ++i) { - /* Calculate the W-gradient for the current layer */ - MatrixPtr input = getInputValue(i); - if (weights_[i]->getWGrad()) { - MatrixPtr weights_grad = weights_[i]->getWGrad(); - if (useGpu_) { - Matrix::resizeOrCreate(cpuInput_, - input->getHeight(), - input->getWidth(), - /* trans */ false, - false); - Matrix::resizeOrCreate(cpuWeightGrad_, - weights_grad->getHeight(), - weights_grad->getWidth(), - /* trans */ false, - false); - cpuInput_->copyFrom(*input); - cpuWeightGrad_->copyFrom(*weights_grad); - } else { - cpuInput_ = input; - cpuWeightGrad_ = weights_grad; - } - preOutput_.grad->mulByBitCodeBackwardWeight( - numClasses_, *cpuLabel_, *cpuWeightGrad_, *cpuInput_); - if (useGpu_) { - weights_grad->copyFrom(*cpuWeightGrad_); - } else { - weights_grad = cpuWeightGrad_; - } - /* Increasing the number of gradient */ - weights_[i]->getParameterPtr()->incUpdate(callback); - } - - /* Calculate the input layers error */ - MatrixPtr inputGrad = getInputGrad(i); - if (inputGrad) { - if (useGpu_) { - Matrix::resizeOrCreate(cpuInputGrad_, - inputGrad->getHeight(), - inputGrad->getWidth(), - /* trans */ false, - false); - Matrix::resizeOrCreate(cpuWeight_, - weights_[i]->getW()->getHeight(), - weights_[i]->getW()->getWidth(), - /* trans */ false, - false); - cpuInputGrad_->copyFrom(*inputGrad); - cpuWeight_->copyFrom(*weights_[i]->getW()); - } else { - cpuInputGrad_ = inputGrad; - cpuWeight_ = weights_[i]->getW(); - } - preOutput_.grad->mulByBitCodeBackwardError( - numClasses_, *cpuLabel_, *cpuWeight_, *cpuInputGrad_); - if (useGpu_) { - inputGrad->copyFrom(*cpuInputGrad_); - } else { - inputGrad = cpuInputGrad_; - } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/IdentityProjection.cpp b/paddle/gserver/layers/IdentityProjection.cpp deleted file mode 100644 index 34e9eb90161f7942c528b70f177e30f301a8f53f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/IdentityProjection.cpp +++ /dev/null @@ -1,103 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Projection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * IdentityProjection performs addition: - * \f[ - * out.row[i] += in.row[i] - * \f] - * - * The config file api is identity_projection. - */ -class IdentityProjection : public Projection { - public: - IdentityProjection(const ProjectionConfig& config, - const ParameterPtr& parameter, - bool useGpu); - virtual void forward(); - virtual void backward(const UpdateCallback& callback); -}; - -REGISTER_PROJECTION(identity, IdentityProjection); - -/** - * Constructed function. - * @note IdentityProjection should not have any parameter. - */ -IdentityProjection::IdentityProjection(const ProjectionConfig& config, - const ParameterPtr& parameter, - bool useGpu) - : Projection(config, parameter, useGpu) { - CHECK(!parameter) << "'identity' projection should not have any parameter"; -} - -void IdentityProjection::forward() { out_->value->add(*in_->value); } - -void IdentityProjection::backward(const UpdateCallback& callback) { - if (in_->grad) { - in_->grad->add(*out_->grad); - } -} - -/** - * IdentityOffsetProjection likes IdentityProjection, but layer size may be - * smaller - * than input size. It selects dimensions [offset, offset+layer_size) from input - * to - * perform addition: - * \f[ - * out.row[i] += in.row[i + \textrm{offset}] - * \f] - * - * The config file api is identity_projection. - */ -class IdentityOffsetProjection : public Projection { - public: - IdentityOffsetProjection(const ProjectionConfig& config, - const ParameterPtr& parameter, - bool useGpu); - virtual void forward(); - virtual void backward(const UpdateCallback& callback); -}; - -REGISTER_PROJECTION(identity_offset, IdentityOffsetProjection); - -/** - * Constructed function. - * @note IdentityOffsetProjection should not have any parameter. - */ -IdentityOffsetProjection::IdentityOffsetProjection( - const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu) - : Projection(config, parameter, useGpu) { - CHECK(!parameter) << "'identity_offset' projection " - "should not have any parameter"; - CHECK_LE(config.output_size() + config.offset(), config.input_size()); -} - -void IdentityOffsetProjection::forward() { - out_->value->addAtOffset(*in_->value, config_.offset()); -} - -void IdentityOffsetProjection::backward(const UpdateCallback& callback) { - if (in_->grad) { - in_->grad->addAtOffset(*out_->grad, config_.offset()); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/InterpolationLayer.cpp b/paddle/gserver/layers/InterpolationLayer.cpp deleted file mode 100644 index 509c07cf22c9bcbe9283241b38540162b3dbe26b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/InterpolationLayer.cpp +++ /dev/null @@ -1,130 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for linear interpolation with two inputs, - * which is used in NEURAL TURING MACHINE. - * \f[ - * y.row[i] = w[i] * x_1.row[i] + (1 - w[i]) * x_2.row[i] - * \f] - * where \f$x_1\f$ and \f$x_2\f$ are two (batchSize x dataDim) inputs, - * \f$w\f$ is (batchSize x 1) weight vector, - * and \f$y\f$ is (batchSize x dataDim) output. - * - * The config file api is interpolation_layer. - */ - -class InterpolationLayer : public Layer { - protected: - /// weightLast = 1 - weight - MatrixPtr weightLast_; - MatrixPtr tmpMatrix; - - public: - explicit InterpolationLayer(const LayerConfig& config) : Layer(config) {} - - ~InterpolationLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(interpolation, InterpolationLayer); - -bool InterpolationLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(3U, inputLayers_.size()); - - return true; -} - -void InterpolationLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr weightV = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr inV2 = getInputValue(2); - - size_t batchSize = inV1->getHeight(); - size_t dataDim = inV1->getWidth(); - - CHECK_EQ(dataDim, getSize()); - CHECK_EQ(dataDim, inV2->getWidth()); - CHECK_EQ(batchSize, inV1->getHeight()); - CHECK_EQ(batchSize, inV2->getHeight()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - resetOutput(batchSize, dataDim); - } - - MatrixPtr outV = getOutputValue(); - - Matrix::resizeOrCreate(weightLast_, batchSize, 1, false, useGpu_); - weightLast_->one(); - weightLast_->sub(*weightV); - - REGISTER_TIMER_INFO("FwInterpTimer", getName().c_str()); - // outV = inV1 * weight + inV2 * weightLast - outV->addRowScale(0, *inV1, *weightV); - outV->addRowScale(0, *inV2, *weightLast_); -} - -void InterpolationLayer::backward(const UpdateCallback& callback) { - MatrixPtr outG = getOutputGrad(); - MatrixPtr weightV = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr inV2 = getInputValue(2); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - MatrixPtr inG2 = getInputGrad(2); - - size_t batchSize = inV1->getHeight(); - size_t dataDim = inV1->getWidth(); - - REGISTER_TIMER_INFO("BwInterpTimer", getName().c_str()); - - if (inG0) { - Matrix::resizeOrCreate(tmpMatrix, batchSize, dataDim, false, useGpu_); - - // inG0 += outG .* (inV1 - inV2) - tmpMatrix->sub(*inV1, *inV2); - inG0->rowDotMul(0, *outG, *tmpMatrix); - } - - if (inG1) { - // inG1 += outG * weight - inG1->addRowScale(0, *outG, *weightV); - } - - if (inG2) { - // inG2 += outG * weightLast - inG2->addRowScale(0, *outG, *weightLast_); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/L2DistanceLayer.cpp b/paddle/gserver/layers/L2DistanceLayer.cpp deleted file mode 100644 index c8cca3762cc3ecd6c04d7d2b804bc588c281bfb4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/L2DistanceLayer.cpp +++ /dev/null @@ -1,91 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "L2DistanceLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(l2_distance, L2DistanceLayer); - -bool L2DistanceLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2UL) << "The L2DistanceLayer accepts two and " - << "only two inputs."; - CHECK_EQ(getSize(), 1UL) << "The output dimensionality of L2DistanceLayer " - << "is fixed to be 1."; - - return true; -} - -void L2DistanceLayer::forward(PassType passType) { - Layer::forward(passType); - - const auto inV1 = getInputValue(0); - const auto inV2 = getInputValue(1); - - CHECK(inV1 && inV2); - CHECK_EQ(inV1->getHeight(), inV2->getHeight()) - << "The height of two inputs of this layer must be the same."; - CHECK_EQ(inV1->getWidth(), inV2->getWidth()) - << "The width of two inputs of this layer must be the same."; - - int batchSize = inV1->getHeight(); - int output_dim = getSize(); - { - REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str()); - reserveOutput(batchSize, output_dim); - auto outV = getOutputValue(); - CHECK(outV) << "The output matrix should not be null."; - - Matrix::resizeOrCreate( - inputSub_, inV1->getHeight(), inV1->getWidth(), false, useGpu_); - - inputSub_->assign(*inV1); - inputSub_->sub(*inV2); - outV->sumOfProducts(*inputSub_, *inputSub_, 1, 0); - outV->sqrt2(*outV); - } -} - -void L2DistanceLayer::backward(const UpdateCallback& callback) { - const auto outG = getOutputGrad(); - const auto outV = getOutputValue(); - CHECK(outG && outV); - - auto inGrad1 = getInputGrad(0); - auto inGrad2 = getInputGrad(1); - - { - REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str()); - - if (inGrad1 || inGrad2) { - outV->scalarDiv(*outV, 1.); - outV->dotMul(*outG, *outV); - } - - if (inGrad1) inGrad1->addRowScale(0, *inputSub_, *outV); - - if (inGrad2) { - inputSub_->mulScalar(-1.); - inGrad2->addRowScale(0, *inputSub_, *outV); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/L2DistanceLayer.h b/paddle/gserver/layers/L2DistanceLayer.h deleted file mode 100644 index 44e688e1377145845033d9d5cc3f31f5594a11f6..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/L2DistanceLayer.h +++ /dev/null @@ -1,52 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief The layer calculates the l2 distance between two input vectors. - * \f[ - * f(\bf{x}, \bf{y}) = \sqrt{\sum_{i=1}^D(x_i - y_i)} - * \f] - * - * - Input1: A vector (batchSize * dataDim) - * - Input2: A vector (batchSize * dataDim) - * - Output: A vector (batchSize * 1) - * - * The configuration api is: l2_distance_layer. - */ - -class L2DistanceLayer : public Layer { - public: - explicit L2DistanceLayer(const LayerConfig& config) : Layer(config) {} - ~L2DistanceLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - private: - // Store the result of subtracting Input2 from Input1 in forward computation, - // which will be reused in backward computation. - MatrixPtr inputSub_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/Layer.cpp b/paddle/gserver/layers/Layer.cpp deleted file mode 100644 index 32e2f4c9dd06e0ef7314b24719235c0be297961f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Layer.cpp +++ /dev/null @@ -1,410 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Util.h" - -#include "CostLayer.h" -#include "paddle/math/SparseMatrix.h" -#include "paddle/utils/Error.h" -#include "paddle/utils/Logging.h" - -#ifndef PADDLE_MOBILE_INFERENCE -#include "ValidationLayer.h" -#endif - -DEFINE_bool(log_error_clipping, false, "enable log error clipping or not"); - -namespace paddle { - -Layer::Layer(const LayerConfig& config, bool useGpu) - : config_(config), - useGpu_(useGpu), - deviceId_(CPU_DEVICE), - needSequenceInfo_(true) {} - -bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { - if (useGpu_ && FLAGS_parallel_nn) { - /* gpu environment is specified by device property */ - deviceId_ = config_.device(); - if (deviceId_ < 0) { - useGpu_ = false; - } - } - - output_.deviceId = deviceId_; - - for (auto& inputConfig : config_.inputs()) { - std::string inputName = inputConfig.input_layer_name(); - LayerPtr inputLayer; - CHECK(mapGet(inputName, layerMap, &inputLayer)) - << "Cannot find input layer " << inputName << " for layer " - << getName(); - this->addPrev(inputLayer); - - inputLayer->addOutputArgument(deviceId_); - - if (inputConfig.has_input_parameter_name()) { - ParameterPtr parameter; - CHECK( - mapGet(inputConfig.input_parameter_name(), parameterMap, ¶meter)) - << "Cannot find input parameter " - << inputConfig.input_parameter_name() << " for layer " << getName(); - parameter->incShared(); - CHECK_EQ(parameter->getDeviceId(), getDeviceId()); - parameters_.push_back(parameter); - } else { - parameters_.push_back(nullptr); - } - - if (inputConfig.has_input_layer_argument()) { - inputArgument_.push_back(inputConfig.input_layer_argument()); - } else { - inputArgument_.push_back(""); - } - } - - if (config_.has_bias_parameter_name()) { - CHECK(mapGet(config_.bias_parameter_name(), parameterMap, &biasParameter_)) - << "Cannot find bias parameter " << config_.bias_parameter_name() - << " for layer " << getName(); - biasParameter_->incShared(); - CHECK_EQ(biasParameter_->getDeviceId(), getDeviceId()); - } - - /* specify the activation function according to the configuration */ - std::string action_type = config_.active_type(); - activation_.reset(ActivationFunction::create(action_type)); - CHECK(activation_); - - initNeedFlags(); - markInBackward_.assign(inputLayers_.size(), false); - - return true; -} - -ClassRegistrar Layer::registrar_; - -LayerPtr Layer::create(const LayerConfig& config) { - std::string type = config.type(); - -#ifndef PADDLE_MOBILE_INFERENCE - // NOTE: As following types have illegal character '-', - // they can not use REGISTER_LAYER to registrar. - // Besides, to fit with old training models, - // they can not use '_' instead. - if (type == "multi-class-cross-entropy") - return LayerPtr(new MultiClassCrossEntropy(config)); - else if (type == "rank-cost") - return LayerPtr(new RankingCost(config)); - else if (type == "auc-validation") - return LayerPtr(new AucValidation(config)); - else if (type == "pnpair-validation") - return LayerPtr(new PnpairValidation(config)); -#endif - - return LayerPtr(registrar_.createByType(config.type(), config)); -} - -void Layer::resetSpecifyOutput(Argument& output, - size_t height, - size_t width, - bool isValueClean, - bool isGradClean) { - SetDevice device(output.deviceId); - - Matrix::resizeOrCreate( - output.value, height, width, /* trans */ false, useGpu(output.deviceId)); - if (isValueClean) { - output.value->zeroMem(); - } - - if (passType_ != PASS_TEST && needGradient()) { - Matrix::resizeOrCreate( - output.grad, height, width, /* trans */ false, useGpu(output.deviceId)); - if (isGradClean) { - output.grad->zeroMem(); - } - } -} - -void Layer::resizeOutput(size_t height, size_t width) { - resetSpecifyOutput(output_, height, width, false, false); - - for (size_t i = 0; i != outputOtherDevice_.size(); i++) { - resetSpecifyOutput(outputOtherDevice_[i], height, width, false, false); - } -} - -void Layer::reserveOutput(size_t height, size_t width) { - resetSpecifyOutput(output_, height, width, false, true); - - for (size_t i = 0; i != outputOtherDevice_.size(); i++) { - resetSpecifyOutput(outputOtherDevice_[i], height, width, false, true); - } -} - -void Layer::resetOutput(size_t height, size_t width) { - resetSpecifyOutput(output_, height, width, true, true); - - for (size_t i = 0; i != outputOtherDevice_.size(); i++) { - resetSpecifyOutput(outputOtherDevice_[i], height, width, true, true); - } -} - -void Layer::addOutputArgument(int deviceId) { - if (deviceId == deviceId_) { - output_.countIncrement(); - return; - } else { - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - if (outputOtherDevice_[i].deviceId == deviceId) { - outputOtherDevice_[i].countIncrement(); - return; - } - } - } - - Argument argu; - argu.deviceId = deviceId; - outputOtherDevice_.push_back(argu); - outputOtherDevice_.back().countIncrement(); -} - -void Layer::copyOutputToOtherDevice() { - for (size_t i = 0; i != outputOtherDevice_.size(); i++) { - SetDevice device(outputOtherDevice_[i].deviceId); - // If outputOtherDevice_[i].value is a CpuMatrix, - // the copyFrom is a synchronous interface. - // If outputOtherDevice_[i].value is a GpuMatrix, since subsequent - // calculations are all on HPPL_STREAM_DEFAULT, - // copyFrom can be an asynchronous interface. - outputOtherDevice_[i].value->copyFrom(*getOutputValue(), - HPPL_STREAM_DEFAULT); - outputOtherDevice_[i].sequenceStartPositions = - output_.sequenceStartPositions; - outputOtherDevice_[i].subSequenceStartPositions = - output_.subSequenceStartPositions; - outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; - - outputOtherDevice_[i].notifyValueReady(); - } -} - -void Layer::waitInputValue() { - for (size_t i = 0; i != inputLayers_.size(); i++) { - if (inputLayers_[i]->getDeviceId() != deviceId_) { - getInput(i).waitValueReady(); - } - } -} - -void Layer::waitAndMergeOutputGrad() { - if (!output_.grad || !outputOtherDevice_.size()) { - return; - } - - for (size_t i = 0; i != outputOtherDevice_.size(); i++) { - outputOtherDevice_[i].waitGradReady(); - } - - /* merge output grad */ - size_t i = 0; - if (!output_.getAllCount()) { - output_.grad->copyFrom(*outputOtherDevice_[0].grad, HPPL_STREAM_1); - hl_stream_synchronize(HPPL_STREAM_1); - - i++; - if (outputOtherDevice_.size() == 1) return; - } - - Matrix::resizeOrCreate(tmpGrad_, - output_.grad->getHeight(), - output_.grad->getWidth(), - /* trans */ false, - useGpu(output_.deviceId)); - - for (; i != outputOtherDevice_.size(); i++) { - tmpGrad_->copyFrom(*outputOtherDevice_[i].grad, HPPL_STREAM_1); - hl_stream_synchronize(HPPL_STREAM_1); - output_.grad->add(*tmpGrad_); - } -} - -void Layer::markAllInputGrad() { - for (size_t i = 0; i != inputLayers_.size(); ++i) { - if (!markInBackward_[i]) { - inputLayers_[i]->getOutput(deviceId_).notifyGradReady(); - } - markInBackward_[i] = false; - } -} - -void Layer::markInputGrad(int inputIndex) { - inputLayers_[inputIndex]->getOutput(deviceId_).notifyGradReady(); - markInBackward_[inputIndex] = true; -} - -void Layer::zeroGrad() { - CHECK(output_.grad.get() != NULL); - output_.grad->zeroMem(); -} - -void Layer::initNeedFlags() { - auto initFlag = [this]( - bool& flag, bool (Layer::*flagQueryFunc)() const, ParameterType type) { - flag = false; - if (biasParameter_ && biasParameter_->hasType(type)) { - flag = true; - } - if (!flag) { - for (auto& para : parameters_) { - if (para && para->hasType(type)) { - flag = true; - break; - } - } - } - if (!flag) { - for (auto& layer : inputLayers_) { - if ((layer.get()->*flagQueryFunc)()) { - flag = true; - } - } - } - }; - initFlag(needGradient_, &Layer::needGradient, PARAMETER_GRADIENT); -} - -void Layer::showOutputStats() { - MatrixPtr out = getOutputValue(); - if (!out) return; - if (!out->getElementCnt()) { - LOG(INFO) << "The number of output of " << config_.name() - << " is 0, skip to show the statistics"; - return; - } - MatrixPtr outSquare; - if (dynamic_cast(out.get())) { - GpuSparseMatrix* tmp = dynamic_cast(out.get()); - outSquare = std::make_shared(tmp->getHeight(), - tmp->getWidth(), - tmp->getElementCnt(), - tmp->getValueType(), - tmp->getFormat()); - } else { - outSquare = out->clone(); - } - outSquare->copyFrom(*out, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - - real mean = outSquare->getSum() / out->getElementCnt(); - real min; - real max; - if (dynamic_cast(outSquare.get())) { - auto tmpMat = dynamic_cast(outSquare.get()); - min = tmpMat->getMin(); - max = tmpMat->getMax(); - tmpMat->square2(); - LOG(INFO) << "show statistics of [none zero values] in sparse matrix"; - } else { - min = outSquare->getMin(); - max = outSquare->getMax(); - outSquare->square2(); - } - real std = (outSquare->getSum() / outSquare->getElementCnt()) - mean * mean; - std = std > 0 ? std : 0; - LOG(INFO) << "The output state of " << config_.name() << ": mean=" << mean - << ", " - << "std=" << std << ", " - << "min=" << min << ", " - << "max=" << max; -} - -void Layer::forwardActivation() { - /* activation */ - auto status = activation_->forward(output_); - status.check(); - - /* dropout */ - if (config_.drop_rate() > 0) { - forwardDropOut(); - CHECK_NE(activation_->getName(), "softmax") - << "Softmax activation cannot be used with Dropout"; - } - - if (FLAGS_show_layer_stat) { - showOutputStats(); - } -} - -void Layer::backwardActivation() { - /* Do error clipping */ - if (config_.error_clipping_threshold() > 0.0f) { - if (FLAGS_log_error_clipping) { - VectorPtr outGradVec = Vector::create( - output_.grad->getData(), output_.grad->getElementCnt(), useGpu_); - real maxAbsGrad = outGradVec->getAbsMax(); - if (maxAbsGrad > config_.error_clipping_threshold()) { - real avgAbsGrad = outGradVec->getAbsSum() / outGradVec->getSize(); - LOG(INFO) << " layer=" << config_.name() << " need clipping," - << " max error=" << maxAbsGrad << " avg error=" << avgAbsGrad; - } - } - output_.grad->clip(-config_.error_clipping_threshold(), - config_.error_clipping_threshold()); - } - - /* Do dropout for delta*/ - if (config_.drop_rate() > 0 && passType_ != PASS_TEST) { - MatrixPtr oGrad = getOutputGrad(); - oGrad->dotMul(*oGrad, *dropOutMask_); - } - - auto status = activation_->backward(output_); - status.check(); -} - -void Layer::forwardDropOut() { - auto& outV = getOutputValue(); - - if (passType_ == PASS_TRAIN) { - // new dropOutMask_ if dropOutMask_ is null ptr - Matrix::resizeOrCreate(dropOutMask_, - outV->getHeight(), - outV->getWidth(), - false, - useGpu(deviceId_)); - dropOutMask_->randomizeUniform(); // generate a uniform random matrix - dropOutMask_->biggerThanScalar(config_.drop_rate()); // random mask - outV->dotMul(*outV, *dropOutMask_); // dropout - } else if (passType_ == PASS_GC) { - // only initialize once - if (!dropOutMask_) { - dropOutMask_ = Matrix::create( - outV->getHeight(), outV->getWidth(), false, useGpu(deviceId_)); - // We use cpu matrix to generate mask so that the mask - // will be same for both gpu version and cpu version. - // This will help unittest to make sure they have same result. - MatrixPtr tmpMask = Matrix::create(outV->getHeight(), outV->getWidth()); - tmpMask->randomizeUniform(); // generate a uniform random matrix - tmpMask->biggerThanScalar(config_.drop_rate()); // random mask - dropOutMask_->copyFrom(*tmpMask); - } - outV->dotMul(*outV, *dropOutMask_); - } else { // passType == PASS_TEST - outV->mulScalar(1.0 - config_.drop_rate()); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/Layer.h b/paddle/gserver/layers/Layer.h deleted file mode 100644 index 13e20e8316323f9082a9615041584685853aa395..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Layer.h +++ /dev/null @@ -1,512 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/function/Function.h" -#include "paddle/gserver/activations/ActivationFunction.h" -#include "paddle/math/CpuSparseMatrix.h" -#include "paddle/parameter/Argument.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/parameter/Weight.h" -#include "paddle/utils/ClassRegistrar.h" -#include "paddle/utils/Util.h" - -/// Macro for registering a layer type. -/// Example: REGISTER_LAYER(crf_error, CRFDecodingErrorLayer); -#define REGISTER_LAYER(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name( \ - []() { Layer::registrar_.registerClass<__class_name>(#__type_name); }) - -#define REGISTER_LAYER_CREATE_FUNC(__type_name, createFunction) \ - static InitFunction __reg_type_##__type_name( \ - []() { Layer::registrar_.registerClass(#__type_name, createFunction); }) - -namespace paddle { - -class Layer; -typedef std::shared_ptr LayerPtr; -typedef std::map LayerMap; -class NeuralNetwork; - -/// layer state, used for RNN and LSTM layers -struct LayerState { - std::vector value; -}; -typedef std::shared_ptr LayerStatePtr; - -/// Paddle device ID, MKLDNN is -2, CPU is -1 -enum PADDLE_DEVICE_ID { - MKLDNN_DEVICE = -2, - CPU_DEVICE = -1, -}; - -/** - * @brief Base class for layer. - * Define necessary variables and functions for every layer. - */ -class Layer { - protected: - /// Layer config - LayerConfig config_; - /// whether to use GPU - bool useGpu_; - /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ... - int deviceId_; - /// Input layers - std::vector inputLayers_; - /// Argument of input layers - std::vector inputArgument_; - - /// Parameter for each input layer. - /// Parameters_[i] is nullptr if inputLayers_[i] does not need parameter. - std::vector parameters_; - - /// nullptr if bias is not needed. - ParameterPtr biasParameter_; - - /// Output - Argument output_; - /// Several outputs stored on different devices, used in 'parallel_nn' case, - /// and record them by deviceId_. - /// Also used in 'use_mkldnn' case. - std::vector outputOtherDevice_; - /// If there are several outputs, map them by each name. - /// MKLDNNLayer use it only to merge output grad - std::map outputMap_; - /// Used to merge grad on different devices. - MatrixPtr tmpGrad_; - - std::unique_ptr activation_; - - /// Current passType, PASS_TRAIN or PASS_TEST - PassType passType_; - - /// Random 0-1 matrix for dropOut - MatrixPtr dropOutMask_; - - /// Whether the layer need to compute gradient - bool needGradient_; - /// Whether the layer need to compute re-sequence information - bool needSequenceInfo_; - - /// Mark input grad in(true) or out(false) of backward function. - std::vector markInBackward_; - - /// Layer forward function - std::vector> forward_; - /// Layer backward function - std::vector> backward_; - - public: - /** - * Wait until all input value ready. - * Called before Layer::forward() function. - */ - virtual void waitInputValue(); - - /** - * Copy layer's output_ to other device. - * If output layer is in other device, called after Layer::forward() function. - */ - virtual void copyOutputToOtherDevice(); - - /** - * Wait until all output grad ready and merge them to output_.grad. - * Called before Layer::backward() function. - */ - virtual void waitAndMergeOutputGrad(); - - /** - * Notify previous layer the output grad ready. - * Called after Layer::backward() function. - */ - virtual void markAllInputGrad(); - - protected: - /** - * Create layer function. Function is called in forward or backward. - * \param function, Layer::forward_ or Layer::backward_ - * \param name, function name - * \param config, initialization configuration for the function - */ - void createFunction(std::vector>& function, - const std::string& name, - const FuncConfig& config) { - if (useGpu_) { - function.emplace_back( - FunctionBase::funcRegistrar_.createByType(name + "-GPU")); - } else { - function.emplace_back( - FunctionBase::funcRegistrar_.createByType(name + "-CPU")); - } - auto& func = function.back(); - func->init(config); - } - - /** - * Notify specified layer the output grad ready. - * Called in the backward function. - * If do mark input grad in the backward function, you should to ensure - * that all input grad will be marked in the backward function. - */ - void markInputGrad(int inputIndex); - - /** - * Get the argument of input layer. - */ - const Argument& getInput(size_t inputIndex) const { - return inputLayers_[inputIndex]->getOutput(deviceId_); - } - - /** - * Get the argument of input layer. - */ - const Argument& getInput(const Layer& inputLayer) const { - return inputLayer.getOutput(deviceId_); - } - - /** - * Get the argument of input layer with deviceId. - */ - const Argument& getInput(size_t inputIndex, int deviceId) const { - return inputLayers_[inputIndex]->getOutput(deviceId); - } - - /** - * Get the forward-input value. - */ - const MatrixPtr& getInputValue(int inputIndex) { - return inputLayers_[inputIndex]->getOutput(deviceId_).value; - } - - /** - * Get the forward-input value. - */ - const MatrixPtr& getInputValue(const Layer& inputLayer) { - return inputLayer.getOutput(deviceId_).value; - } - - /** - * Get the forward-input value with deviceId. - */ - const MatrixPtr& getInputValue(int inputIndex, int deviceId) { - return inputLayers_[inputIndex]->getOutput(deviceId).value; - } - - /** - * Get the forward-input grad. - */ - const MatrixPtr& getInputGrad(int inputIndex) { - return inputLayers_[inputIndex]->getOutput(deviceId_).grad; - } - - /** - * Get the forward-input grad. - */ - const MatrixPtr& getInputGrad(const Layer& inputLayer) { - return inputLayer.getOutput(deviceId_).grad; - } - - /** - * Get the forward-input grad. - */ - const MatrixPtr& getInputGrad(int inputIndex, int deviceId) { - return inputLayers_[inputIndex]->getOutput(deviceId).grad; - } - - /** - * Get the forward-input label. - */ - const IVectorPtr& getInputLabel(const Layer& inputLayer) { - return inputLayer.getOutput(deviceId_).ids; - } - - /** - * Change the size of output (value, grad). - * Reset to value zero if isValueClean = true, - * Reset to grad zero if isGradClean = true. - */ - void resetSpecifyOutput(Argument& output, - size_t height, - size_t width, - bool isValueClean, - bool isGradClean); - - /** - * Add output argument to other devices. - */ - void addOutputArgument(int deviceId); - - public: - explicit Layer(const LayerConfig& config, bool useGpu = FLAGS_use_gpu); - virtual ~Layer() {} - - /// Register a Layer - static ClassRegistrar registrar_; - - /** - * Get the flag whether layer need to compute gradient. - */ - bool needGradient() const { return needGradient_; } - - /** - * Set the flag whether layer need to compute gradient. - */ - void setNeedGradient(bool need) { needGradient_ = need; } - - /** - * Set the flag whether layer need to re-compute sequence information, - * which includes sequenceStartPositions or subSequenceStartPositions. - */ - void setNeedSequenceInfo(bool need) { needSequenceInfo_ = need; } - - /** - * Get layer's name. - */ - const std::string& getName() const { return config_.name(); } - - /** - * Get layer's type. - */ - const std::string& getType() const { return config_.type(); } - - /** - * Get layer's size. - */ - size_t getSize() const { return config_.size(); } - - /** - * Get layer's deviceId. - */ - int getDeviceId() const { return deviceId_; } - - /** - * Add the inputLayer. - */ - void addPrev(LayerPtr l) { inputLayers_.push_back(l); } - - /** - * Get the size of inputLayer[i]. - */ - const LayerPtr& getPrev(size_t i) { return inputLayers_[i]; } - - /** - * Get the forward-output value. - */ - const MatrixPtr& getOutputValue() { return output_.value; } - - /** - * Get the forward-output label. - */ - const IVectorPtr& getOutputLabel() { return output_.ids; } - - /** - * Get the backward-Loss value. - */ - const MatrixPtr& getOutputGrad() { return output_.grad; } - /** - * If layer has multi-output, set output into outputMap_. - */ - void setOutput(const std::string& name, Argument* output) { - outputMap_[name] = output; - } - - /** - * Get the output map size, if layer has multi-output. - */ - size_t getOutputMapSize() { return outputMap_.size(); } - - /** - * Get the output based on layer's name. - */ - Argument& getOutput(const std::string& str = "") { - if (str == "") { - return output_; - } else { - auto output = outputMap_.find(str); - if (output != outputMap_.end()) { - return *output->second; - } else { - LOG(FATAL) << "No specific output " << str; - return *((Argument*)nullptr); - } - } - } - - /** - * Get the output based on deviceId. - */ - const Argument& getOutput(int deviceId) const { - if (deviceId == getDeviceId()) { - return output_; - } else { - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - if (outputOtherDevice_[i].deviceId == deviceId) { - return outputOtherDevice_[i]; - } - } - - LOG(FATAL) << "No specific device output "; - return *((Argument*)nullptr); - } - } - - /** - * Get layer's parameters. - */ - const std::vector& getParameters() { return parameters_; } - - /** - * Get layer's bias-parameters. - */ - const ParameterPtr& getBiasParameter() { return biasParameter_; } - - /** - * Create pointer of layer. - */ - static LayerPtr create(const LayerConfig& config); - - /** - * Resize the output matrix size. - */ - void resizeOutput(size_t height, size_t width); - - /** - * Resize the output matrix size, - * and reset value to zero. - */ - void reserveOutput(size_t height, size_t width); - - /** - * Resize the output matrix size, - * and reset value and grad to zero. - */ - void resetOutput(size_t height, size_t width); - - /** - * Clear the gradient of output. - */ - void zeroGrad(); - - /** - * Intialization. - * For example, adding input layers from layerMap and parameterMap. - */ - virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); - - /** - * Intialization for sub network if there has sub network. - * @param rootNetwork root network - * @param config model config - * @param parameterTypes parameter's type - * @param useGpu whether to use gpu or not - */ - virtual void initSubNetwork(NeuralNetwork* rootNetwork, - const ModelConfig& config, - const std::vector& parameterTypes, - bool useGpu) {} - - /** - * @brief Access SubNetwork Object. - * If subnetwork exists, then invoke callback with subnetwrk. - * @param callback if sub-network is exist, the callback is invoked. - */ - virtual void accessSubNetwork( - const std::function& callback) {} - - /** - * If use sparse row matrix as parameter, - * prefetch feature ids in input label. - */ - virtual void prefetch() {} - - /** - * Forward propagation. - * All inherited implementation should call Layer::foward() function. - */ - virtual void forward(PassType passType) { - passType_ = passType; - if (!inputLayers_.empty() && needSequenceInfo_) { - const Argument& input = getInput(0); - output_.sequenceStartPositions = input.sequenceStartPositions; - output_.subSequenceStartPositions = input.subSequenceStartPositions; - output_.cpuSequenceDims = input.cpuSequenceDims; - } - } - - /** - * Reset the internal state variables. - * Allocate them if they have not been allocated. - * This function need to called before Layer::forward() for generating - * sequence. - * - * This is used for sequence generation. When generating sequence, the - * calculation at current timestamp depends on the state from previous - * timestamp. The model needs to keep the information about the previous - * timestamp in the state variables. Layers such as RecurrentLayer, - * LstmLayer and ContextLayer have state variables. - */ - virtual void resetState() {} - - /** - * Set layer state. - */ - virtual void setState(LayerStatePtr state) {} - - /** - * Get layer state. - * @return A copy of internal state. - */ - virtual LayerStatePtr getState() { return nullptr; } - - /** - * Show output state. - */ - void showOutputStats(); - - /** - * Backward propagation. - * Should only be called after Layer::forward() function. - */ - virtual void backward(const UpdateCallback& callback = nullptr) = 0; - - /** - * One pass is finished. - */ - virtual void onPassEnd() {} - - protected: - /** - * Forward of activation function. - */ - void forwardActivation(); - /** - * Backward of activation function. - */ - void backwardActivation(); - /** - * Forward of dropOut. - */ - void forwardDropOut(); - /** - * Initilize the needGradient_ flag. - */ - void initNeedFlags(); -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/LinearChainCRF.h b/paddle/gserver/layers/LinearChainCRF.h deleted file mode 100644 index e802b701d0237bed44adc83273fe53c3e18c92ec..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LinearChainCRF.h +++ /dev/null @@ -1,97 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/math/Matrix.h" - -namespace paddle { - -class LinearChainCRF { - public: - /** - * The size of para must be \f$(numClasses + 2) * numClasses\f$. - * The first numClasses values of para are for starting weights (\f$a\f$). - * The next numClasses values of para are for ending weights (\f$b\f$), - * The remaning values are for transition weights (\f$w\f$). - * - * The probability of a state sequence s of length \f$L\f$ is defined as: - * \f$P(s) = (1/Z) exp(a_{s_1} + b_{s_L} - * + \sum_{l=1}^L x_{s_l} - * + \sum_{l=2}^L w_{s_{l-1},s_l})\f$ - * where \f$Z\f$ is a normalization value so that the sum of \f$P(s)\f$ over - * all possible - * sequences is \f$1\f$, and \f$x\f$ is the input feature to the CRF. - */ - LinearChainCRF(int numClasses, real* para); - - /** - * Calculate the negative log likelihood of s given x. - * The size of x must be length * numClasses. Each consecutive numClasses - * values are the features for one time step. - */ - real forward(real* x, int* s, int length); - - /** - * Calculate the gradient with respect to x, a, b, and w. - * backward() can only be called after a corresponding call to forward() with - * the same x, s and length. - * The gradient with respect to a, b, and w will not be calculated if - * needWGrad is false. - * @note Please call getWGrad() and getXGrad() to get the gradient with - * respect to (a, b, w) and x respectively. - */ - void backward(real* x, int* s, int length, bool needWGrad); - - /** - * Find the most probable sequence given x. The result will be stored in s. - */ - void decode(real* x, int* s, int length); - - /* - * Return the gradient with respect to (a, b, w). It can only be called after - * a corresponding call to backward(). - */ - MatrixPtr getWGrad() { return matWGrad_; } - - /* - * Return the gradient with respect to x. It can only be called after a - * corresponding call to backward(). - */ - MatrixPtr getXGrad() { return matGrad_; } - - protected: - int numClasses_; - MatrixPtr a_; - MatrixPtr b_; - MatrixPtr w_; - MatrixPtr matWGrad_; - MatrixPtr da_; - MatrixPtr db_; - MatrixPtr dw_; - MatrixPtr ones_; - - MatrixPtr expX_; - MatrixPtr matGrad_; - MatrixPtr alpha_; - MatrixPtr beta_; - MatrixPtr maxX_; - MatrixPtr expW_; - - // track_(k,i) = j means that the best sequence at time k for class i comes - // from the sequence at time k-1 for class j - IVectorPtr track_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/LinearChainCTC.h b/paddle/gserver/layers/LinearChainCTC.h deleted file mode 100644 index 5b325a0deb0e9d8df241175159321e52f527f6c4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LinearChainCTC.h +++ /dev/null @@ -1,50 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "paddle/math/Matrix.h" - -namespace paddle { - -class LinearChainCTC { - public: - LinearChainCTC(int numClasses, bool normByTimes); - - // Calculate the negative log probability as loss - real forward(real* softmaxSeq, - int softmaxSeqLen, - int* labelSeq, - int labelSeqLen); - - // calculate the gradient - void backward(real* softmaxSeq, - real* softmaxSeqGrad, - int* labelSeq, - int labelSeqLen); - - protected: - int numClasses_, blank_, totalSegments_, totalTime_; - bool normByTimes_; - bool isInvalid_; - - MatrixPtr logActs_, forwardVars_, backwardVars_, gradTerms_; - - real logProb_; - - void segmentRange(int& start, int& end, int time); -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/LstmCompute.cpp b/paddle/gserver/layers/LstmCompute.cpp deleted file mode 100644 index ea30f6d6b1b8586569407af6baac2c14034e709c..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LstmCompute.cpp +++ /dev/null @@ -1,93 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "LstmCompute.h" -#include "hl_recurrent_apply.cuh" -#include "paddle/utils/Util.h" - -namespace paddle { - -void LstmCompute::init(LayerConfig &config) { - activeNode_ = hlActiveType(config.active_type()); - activeGate_ = hlActiveType(config.active_gate_type()); - activeState_ = hlActiveType(config.active_state_type()); -} - -template <> -void LstmCompute::forwardOneSequence<0>(hl_lstm_value value, int frameSize) { - hl_cpu_lstm_forward(hppl::forward::lstm(), - value, - frameSize, - activeNode_, - activeGate_, - activeState_); -} - -template <> -void LstmCompute::backwardOneSequence<0>(hl_lstm_value value, - hl_lstm_grad grad, - int frameSize) { - hl_cpu_lstm_backward(hppl::backward::lstm(), - value, - grad, - frameSize, - activeNode_, - activeGate_, - activeState_); -} - -template <> -void LstmCompute::forwardBatch<0>(hl_lstm_value value, - int frameSize, - int batchSize) { - for (int b = 0; b < batchSize; b++) { - forwardOneSequence<0>(value, frameSize); - - value.gateValue += frameSize * 4; - value.stateValue += frameSize; - value.stateActiveValue += frameSize; - value.outputValue += frameSize; - if (value.prevStateValue) { - value.prevStateValue += frameSize; - } - } -} - -template <> -void LstmCompute::backwardBatch<0>(hl_lstm_value value, - hl_lstm_grad grad, - int frameSize, - int batchSize) { - for (int b = 0; b < batchSize; b++) { - backwardOneSequence<0>(value, grad, frameSize); - - value.gateValue += frameSize * 4; - value.stateValue += frameSize; - value.stateActiveValue += frameSize; - value.outputValue += frameSize; - if (value.prevStateValue) { - value.prevStateValue += frameSize; - } - - grad.gateGrad += frameSize * 4; - grad.stateGrad += frameSize; - grad.stateActiveGrad += frameSize; - grad.outputGrad += frameSize; - if (grad.prevStateGrad) { - grad.prevStateGrad += frameSize; - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/LstmCompute.h b/paddle/gserver/layers/LstmCompute.h deleted file mode 100644 index 80fb01cd1885151c8d62a4b5dfdb4ba08327926d..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LstmCompute.h +++ /dev/null @@ -1,66 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ModelConfig.pb.h" -#include "hl_gpu.h" -#include "paddle/utils/Common.h" - -namespace paddle { - -class LstmCompute { - public: - void init(LayerConfig &config); - - /** - * LstmLayer batch compute API (forwardBatch, backwardBatch). - * If use batch compute api, lstm value(and grad) need to be batch structure. - * Compute order: - * forwardBatch: for 0 <= id < numBatch - * backwardBatch: for numBatch > id >= 0 - */ - template - void forwardBatch(hl_lstm_value value, int frameSize, int batchSize); - - template - void backwardBatch(hl_lstm_value value, - hl_lstm_grad grad, - int frameSize, - int batchSize); - - /** - * LstmLayer sequence compute API (forwardOneSequence, backwardOneSequence). - * Compute order(for each sequence): - * forwardOneSequence: - * if (!reversed) for 0 <= seqId < seqLength - * if (reversed) for seqLength > seqId >= 0 - * backwardOneSequence: - * if (!reversed) for seqLength > seqId >= 0 - * if (reversed) for 0 <= seqId < seqLength - */ - template - void forwardOneSequence(hl_lstm_value value, int frameSize); - template - void backwardOneSequence(hl_lstm_value value, - hl_lstm_grad grad, - int frameSize); - - public: - hl_activation_mode_t activeNode_; - hl_activation_mode_t activeGate_; - hl_activation_mode_t activeState_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/LstmLayer.cpp b/paddle/gserver/layers/LstmLayer.cpp deleted file mode 100644 index f65ae6a3e69cb5f0a7e6073d17bfd0beae91cd5d..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LstmLayer.cpp +++ /dev/null @@ -1,805 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "LstmLayer.h" -#include "paddle/math/BaseMatrix.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Stat.h" - -DECLARE_bool(prev_batch_state); - -namespace paddle { - -REGISTER_LAYER(lstmemory, LstmLayer); - -bool LstmLayer::init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) { - if (!Layer::init(layerMap, parameterMap)) return false; - CHECK_EQ(1U, inputLayers_.size()); - CHECK_EQ(1U, parameters_.size()); - CHECK_EQ(getSize() * getSize() * 4, parameters_[0]->getSize()); - CHECK_EQ(getSize() * 7, biasParameter_->getSize()); - weight_.reset(new Weight(getSize(), getSize() * 4, parameters_[0])); - if (biasParameter_.get() != NULL) { - bias_.reset(new Weight(1, getSize() * 7, biasParameter_)); - if (bias_->getW()) { - localBias_ = Matrix::create(nullptr, - /* height= */ 1, - getSize() * 4, - /* trans= */ false, - useGpu_); - checkIg_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkFg_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkOg_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - - localBias_->setData(bias_->getW()->getData()); - checkIg_->setData(bias_->getW()->getData() + getSize() * 4); - checkFg_->setData(bias_->getW()->getData() + getSize() * 5); - checkOg_->setData(bias_->getW()->getData() + getSize() * 6); - } - - if (bias_->getWGrad()) { - localBiasGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize() * 4, - /* trans= */ false, - useGpu_); - checkIgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkFgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkOgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - localBiasGrad_->setData(bias_->getWGrad()->getData()); - checkIgGrad_->setData(bias_->getWGrad()->getData() + getSize() * 4); - checkFgGrad_->setData(bias_->getWGrad()->getData() + getSize() * 5); - checkOgGrad_->setData(bias_->getWGrad()->getData() + getSize() * 6); - } - } else { - LOG(FATAL) << "Bias should be here."; - } - reversed_ = config_.reversed(); - - // create IdentityActivation for using drop_rate - activation_.reset(ActivationFunction::create("")); - - LstmCompute::init(config_); - useBatch_ = true; - useSeqParallel_ = false; - if (useGpu_ && (getSize() == 32 || getSize() == 64)) { - useSeqParallel_ = true; - } - - return true; -} - -void LstmLayer::resetState() { - CHECK(!reversed_) << "state is not allowed for reversed lstmemory layer"; - Matrix::resizeOrCreate( - prevOutput_, 1, getSize(), /* trans= */ false, useGpu_); - Matrix::resizeOrCreate(prevState_, 1, getSize(), /* trans= */ false, useGpu_); - prevOutput_->resize(0, getSize()); - prevState_->resize(0, getSize()); - if (FLAGS_prev_batch_state) { - useBatch_ = true; - } else { - useBatch_ = false; - } -} - -void LstmLayer::setState(LayerStatePtr state) { - CHECK(state->value.size() == 2) << "two matrices are expected for LSTM state"; - prevOutput_->resize(state->value[0]->getHeight(), - state->value[0]->getWidth()); - prevState_->resize(state->value[1]->getHeight(), state->value[1]->getWidth()); - prevOutput_->copyFrom(*(state->value[0])); - prevState_->copyFrom(*(state->value[1])); -} - -LayerStatePtr LstmLayer::getState() { - LayerStatePtr res = std::make_shared(); - if (prevOutput_->getHeight() && prevOutput_->getWidth()) { - res->value.push_back(prevOutput_->clone(0, 0, useGpu_)); - res->value[0]->copyFrom(*prevOutput_); - res->value.push_back(prevState_->clone(0, 0, useGpu_)); - res->value[1]->copyFrom(*prevState_); - } else { - MatrixPtr output = - Matrix::create(1, getSize(), /* trans= */ false, useGpu_); - MatrixPtr state = Matrix::create(1, getSize(), /* trans= */ false, useGpu_); - output->resize(0, getSize()); - state->resize(0, getSize()); - res->value.push_back(output); - res->value.push_back(state); - } - return res; -} - -void LstmLayer::forward(PassType passType) { - REGISTER_TIMER_INFO("LstmFwTimer", getName().c_str()); - Layer::forward(passType); - - const Argument &input = getInput(0); - CHECK(input.sequenceStartPositions); - int batchSize = input.getBatchSize(); - resetOutput(batchSize, getSize()); - CHECK_EQ(getSize() * 4, input.value->getWidth()); - size_t numSequences = input.getNumSequences(); - const int *starts = input.sequenceStartPositions->getData(false); - CHECK_EQ(starts[numSequences], batchSize); - - Matrix::resizeOrCreate(gate_.value, - /* height= */ batchSize, - getSize() * 4, - /* trans= */ false, - useGpu_); - if (prevOutput_) { - size_t prevNumSeq = useBatch_ ? numSequences : 1; - if (prevOutput_->getHeight() == 0) { - prevOutput_->resize(prevNumSeq, getSize()); - prevState_->resize(prevNumSeq, getSize()); - prevOutput_->zeroMem(); - prevState_->zeroMem(); - } else { - CHECK_EQ(prevOutput_->getHeight(), prevNumSeq) - << "the number of sequences must be the same"; - } - Matrix::resizeOrCreate(totalState_, - prevState_->getHeight() + batchSize, - getSize(), - /*trans*/ false, - useGpu_); - state_.value = Matrix::create(nullptr, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - state_.value->setData(totalState_->getData() + - prevState_->getHeight() * getSize()); - } else { - Matrix::resizeOrCreate(state_.value, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - } - Matrix::resizeOrCreate(preOutput_.value, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - - if (!useBatch_) { - forwardSequence(batchSize, numSequences, starts, input.value); - } else { - if (!useSeqParallel_) { - forwardBatch(batchSize, numSequences, starts, input.value); - } else { - const int *starts = input.sequenceStartPositions->getData(useGpu_); - forwardSeqParallel(batchSize, numSequences, starts, input.value); - } - } - /* activation */ { forwardActivation(); } -} - -void LstmLayer::backward(const UpdateCallback &callback) { - REGISTER_TIMER_INFO("LstmBwTimer", getName().c_str()); - /* Do derivation */ { backwardActivation(); } - - const Argument &input = getInput(0); - CHECK(input.sequenceStartPositions); - int batchSize = input.getBatchSize(); - size_t numSequences = input.getNumSequences(); - - Matrix::resizeOrCreate(gate_.grad, - /* height= */ batchSize, - getSize() * 4, - /* trans= */ false, - useGpu_); - Matrix::resizeOrCreate(state_.grad, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - Matrix::resizeOrCreate(preOutput_.grad, - /* height= */ batchSize, - getSize(), - /* trans= */ false, - useGpu_); - state_.grad->zero(); - - const int *starts = input.sequenceStartPositions->getData(false); - if (!useBatch_) { - backwardSequence(batchSize, numSequences, starts, input.grad); - } else { - if (!useSeqParallel_) { - backwardBatch(batchSize, numSequences, starts, input.grad); - } else { - const int *starts = input.sequenceStartPositions->getData(useGpu_); - backwardSeqParallel(batchSize, numSequences, starts, input.grad); - } - } - - if (bias_) { - bias_->getParameterPtr()->incUpdate(callback); - } - weight_->getParameterPtr()->incUpdate(callback); -} - -void LstmLayer::forwardSequence(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputValue) { - REGISTER_TIMER_INFO("LstmFwSequenceTime", getName().c_str()); - gate_.value->assign(*inputValue); - if (bias_) { - gate_.value->addBias(*localBias_, 1); - } - - hl_lstm_value lstmValue; - lstmValue.checkIg = checkIg_->getData(); - lstmValue.checkFg = checkFg_->getData(); - lstmValue.checkOg = checkOg_->getData(); - lstmValue.gateValue = gate_.value->getData(); - lstmValue.stateValue = state_.value->getData(); - lstmValue.stateActiveValue = preOutput_.value->getData(); - lstmValue.outputValue = output_.value->getData(); - lstmValue.prevStateValue = nullptr; - if (reversed_) { - lstmValue.gateValue += (batchSize - 1) * getSize() * 4; - lstmValue.stateValue += (batchSize - 1) * getSize(); - lstmValue.stateActiveValue += (batchSize - 1) * getSize(); - lstmValue.outputValue += (batchSize - 1) * getSize(); - } - - auto nextFrame = [&lstmValue](bool reversed, int frameSize) { - lstmValue.prevStateValue = lstmValue.stateValue; - if (!reversed) { - lstmValue.gateValue += frameSize * 4; - lstmValue.stateValue += frameSize; - lstmValue.stateActiveValue += frameSize; - lstmValue.outputValue += frameSize; - } else { - lstmValue.gateValue -= frameSize * 4; - lstmValue.stateValue -= frameSize; - lstmValue.stateActiveValue -= frameSize; - lstmValue.outputValue -= frameSize; - } - }; - - MatrixPtr frameGate = Matrix::create(nullptr, - /* height= */ 1, - getSize() * 4, - /* trans= */ false, - useGpu_); - MatrixPtr frameOutput = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - - if (!reversed_) { - if (prevState_) { - lstmValue.prevStateValue = prevState_->getData(); - } - if (prevOutput_) { - frameGate->setData(lstmValue.gateValue); - frameGate->mul(*prevOutput_, *weight_->getW(), 1, 1); - } - } - AsyncGpuBlock asyncGpuBlock; - for (size_t n = 0; n < numSequences; ++n) { - int length; - if (!reversed_) { - length = starts[n + 1] - starts[n]; - } else { - length = starts[numSequences - n] - starts[numSequences - n - 1]; - } - for (int l = 0; l < length; ++l) { - if (useGpu_) { - LstmCompute::forwardOneSequence<1>(lstmValue, getSize()); - } else { - LstmCompute::forwardOneSequence<0>(lstmValue, getSize()); - } - - if (l != length - 1) { - frameOutput->setData(lstmValue.outputValue); - nextFrame(reversed_, getSize()); - frameGate->setData(lstmValue.gateValue); - frameGate->mul(*frameOutput, *weight_->getW(), 1, 1); - } - } - if (n != numSequences - 1) { - frameOutput->setData(lstmValue.outputValue); - nextFrame(reversed_, getSize()); - frameGate->setData(lstmValue.gateValue); - if (!reversed_) { - if (!prevState_) lstmValue.prevStateValue = nullptr; - if (prevOutput_) { - frameGate->mul(*frameOutput, *weight_->getW(), 1, 1); - } - } else { - lstmValue.prevStateValue = nullptr; - } - } - } - - if (!reversed_) { - if (prevState_) { - prevState_->assign(*state_.value->subMatrix(batchSize - 1, 1)); - } - if (prevOutput_) { - prevOutput_->assign(*output_.value->subMatrix(batchSize - 1, 1)); - } - } -} - -void LstmLayer::backwardSequence(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputGrad) { - REGISTER_TIMER_INFO("LstmBwSequenceTime", getName().c_str()); - MatrixPtr weightT = weight_->getW()->getTranspose(); - - hl_lstm_value lstmValue; - hl_lstm_grad lstmGrad; - lstmValue.checkIg = checkIg_->getData(); - lstmValue.checkFg = checkFg_->getData(); - lstmValue.checkOg = checkOg_->getData(); - lstmValue.gateValue = gate_.value->getData(); - lstmValue.stateValue = state_.value->getData(); - lstmValue.stateActiveValue = preOutput_.value->getData(); - lstmValue.outputValue = nullptr; - - if (bias_->getWGrad()) { - lstmGrad.checkIgGrad = checkIgGrad_->getData(); - lstmGrad.checkFgGrad = checkFgGrad_->getData(); - lstmGrad.checkOgGrad = checkOgGrad_->getData(); - } else { - lstmGrad.checkIgGrad = nullptr; - lstmGrad.checkFgGrad = nullptr; - lstmGrad.checkOgGrad = nullptr; - } - lstmGrad.gateGrad = gate_.grad->getData(); - lstmGrad.stateGrad = state_.grad->getData(); - lstmGrad.stateActiveGrad = nullptr; - lstmGrad.outputGrad = output_.grad->getData(); - - if (!reversed_) { - lstmValue.gateValue += (batchSize - 1) * getSize() * 4; - lstmGrad.gateGrad += (batchSize - 1) * getSize() * 4; - lstmValue.stateValue += (batchSize - 1) * getSize(); - lstmGrad.stateGrad += (batchSize - 1) * getSize(); - lstmValue.stateActiveValue += (batchSize - 1) * getSize(); - lstmGrad.outputGrad += (batchSize - 1) * getSize(); - lstmValue.prevStateValue = lstmValue.stateValue - getSize(); - lstmGrad.prevStateGrad = lstmGrad.stateGrad - getSize(); - } else { - lstmValue.prevStateValue = lstmValue.stateValue + getSize(); - lstmGrad.prevStateGrad = lstmGrad.stateGrad + getSize(); - } - - auto nextFrame = [&lstmValue, &lstmGrad](bool reversed, int frameSize) { - if (reversed) { - lstmValue.gateValue += frameSize * 4; - lstmGrad.gateGrad += frameSize * 4; - lstmValue.stateValue += frameSize; - lstmGrad.stateGrad += frameSize; - lstmValue.stateActiveValue += frameSize; - lstmGrad.outputGrad += frameSize; - lstmValue.prevStateValue = lstmValue.stateValue + frameSize; - lstmGrad.prevStateGrad = lstmGrad.stateGrad + frameSize; - } else { - lstmValue.gateValue -= frameSize * 4; - lstmGrad.gateGrad -= frameSize * 4; - lstmValue.stateValue -= frameSize; - lstmGrad.stateGrad -= frameSize; - lstmValue.stateActiveValue -= frameSize; - lstmGrad.outputGrad -= frameSize; - lstmValue.prevStateValue = lstmValue.stateValue - frameSize; - lstmGrad.prevStateGrad = lstmGrad.stateGrad - frameSize; - } - }; - - MatrixPtr frameGate = Matrix::create(nullptr, - /* height= */ 1, - getSize() * 4, - /* trans= */ false, - useGpu_); - MatrixPtr frameOutput = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - - { - AsyncGpuBlock asyncGpuBlock; - for (size_t n = 0; n < numSequences; ++n) { - int length; - int start; - if (reversed_) { - length = starts[n + 1] - starts[n]; - start = starts[n]; - } else { - length = starts[numSequences - n] - starts[numSequences - n - 1]; - start = starts[numSequences - n - 1]; - } - for (int l = 0; l < length; ++l) { - if (l == length - 1) { - lstmValue.prevStateValue = nullptr; - lstmGrad.prevStateGrad = nullptr; - } - if (useGpu_) { - LstmCompute::backwardOneSequence<1>(lstmValue, lstmGrad, getSize()); - } else { - LstmCompute::backwardOneSequence<0>(lstmValue, lstmGrad, getSize()); - } - if (l != length - 1) { - frameGate->setData(lstmGrad.gateGrad); - nextFrame(reversed_, getSize()); - frameOutput->setData(lstmGrad.outputGrad); - frameOutput->mul(*frameGate, *weightT, 1, 1); - } else { - nextFrame(reversed_, getSize()); - } - } - - if (weight_->getWGrad()) { - if (!reversed_) { - weight_->getWGrad()->mul( - *output_.value->subMatrix(start, length - 1)->getTranspose(), - *gate_.grad->subMatrix(start + 1, length - 1), - 1, - 1); - } else { - weight_->getWGrad()->mul( - *output_.value->subMatrix(start + 1, length - 1)->getTranspose(), - *gate_.grad->subMatrix(start, length - 1), - 1, - 1); - } - } - } - } - - if (inputGrad) { - inputGrad->add(*gate_.grad); - } - if (bias_ && bias_->getWGrad()) { - localBiasGrad_->collectBias(*gate_.grad, 1); - } -} - -void LstmLayer::forwardBatch(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputValue) { - REGISTER_TIMER_INFO("LstmFwBatchTime", getName().c_str()); - - hl_lstm_value lstmValue; - lstmValue.checkIg = checkIg_->getData(); - lstmValue.checkFg = checkFg_->getData(); - lstmValue.checkOg = checkOg_->getData(); - - if (!batchValue_) { - batchValue_.reset(new SequenceToBatch(useGpu_)); - } - batchValue_->resizeOrCreateBatch( - batchSize, numSequences, starts, reversed_, prevOutput_ ? true : false); - - batchValue_->resizeOrCreate(*output_.value); - batchValue_->copy(*inputValue, *gate_.value, /* seq2batch */ true); - if (bias_) { - gate_.value->addBias(*localBias_, 1); - } - - { - int numBatch = batchValue_->getNumBatch(); - int batchSize = 0; - AsyncGpuBlock asyncGpuBlock; - if (prevState_) { - lstmValue.prevStateValue = totalState_->getData(); - } else { - lstmValue.prevStateValue = nullptr; - } - for (int n = 0; n < numBatch; n++) { - MatrixPtr outputValue = batchValue_->getBatchValue(n); - MatrixPtr gateValue = batchValue_->getBatchValue(*gate_.value, n); - batchSize = outputValue->getHeight(); - - if (n != 0) { - MatrixPtr batch1 = batchValue_->getBatchValue(n - 1, batchSize); - gateValue->mul(*batch1, *weight_->getW(), 1, 1); - } else if (prevOutput_) { - Matrix::resizeOrCreate(prevBatchOutput2_, - gateValue->getHeight(), - getSize(), - false, - useGpu_); - batchValue_->prevOutput2Batch(*prevOutput_, *prevBatchOutput2_); - gateValue->mul(*prevBatchOutput2_, *weight_->getW(), 1, 1); - - batchValue_->prevOutput2Batch(*prevState_, - *totalState_->subMatrix(0, numSequences)); - } - - lstmValue.gateValue = gateValue->getData(); - lstmValue.outputValue = outputValue->getData(); - lstmValue.stateValue = - batchValue_->getBatchValue(*state_.value, n)->getData(); - lstmValue.stateActiveValue = - batchValue_->getBatchValue(*preOutput_.value, n)->getData(); - { - if (useGpu_) { - LstmCompute::forwardBatch<1>(lstmValue, getSize(), batchSize); - } else { - LstmCompute::forwardBatch<0>(lstmValue, getSize(), batchSize); - } - } - lstmValue.prevStateValue = lstmValue.stateValue; - } - } - { - REGISTER_TIMER_INFO("batchToSeq", getName().c_str()); - batchValue_->copyBackSeq(*output_.value); - } - if (prevOutput_) { - getPrevBatchOutput(numSequences); - getPrevBatchState(numSequences); - } -} - -void LstmLayer::getPrevBatchOutput(size_t numSequences) { - prevOutput_->resize(numSequences, getSize()); - batchValue_->getSeqOutputFromBatch(*prevOutput_, - *batchValue_->getBatchValue()); -} - -void LstmLayer::getPrevBatchState(size_t numSequences) { - prevState_->resize(numSequences, getSize()); - batchValue_->getSeqOutputFromBatch(*prevState_, *state_.value); -} - -void LstmLayer::backwardBatch(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputGrad) { - REGISTER_TIMER_INFO("LstmBwBatchTime", getName().c_str()); - - hl_lstm_value lstmValue; - lstmValue.checkIg = checkIg_->getData(); - lstmValue.checkFg = checkFg_->getData(); - lstmValue.checkOg = checkOg_->getData(); - - hl_lstm_grad lstmGrad; - lstmGrad.stateActiveGrad = preOutput_.grad->getData(); - - if (bias_->getWGrad()) { - lstmGrad.checkIgGrad = checkIgGrad_->getData(); - lstmGrad.checkFgGrad = checkFgGrad_->getData(); - lstmGrad.checkOgGrad = checkOgGrad_->getData(); - } else { - lstmGrad.checkIgGrad = nullptr; - lstmGrad.checkFgGrad = nullptr; - lstmGrad.checkOgGrad = nullptr; - } - - if (!batchGrad_) { - batchGrad_.reset(new SequenceToBatch(useGpu_)); - } - batchGrad_->shareIndexWith(*batchValue_); - - { - REGISTER_TIMER_INFO("seqToBatch", getName().c_str()); - batchGrad_->copyFromSeq(*output_.grad); - } - - { - MatrixPtr weightT = weight_->getW()->getTranspose(); - int numBatch = batchGrad_->getNumBatch(); - int batchSize = 0; - AsyncGpuBlock asyncGpuBlock; - for (int n = (int)numBatch - 1; n >= 0; n--) { - MatrixPtr outputGrad = batchGrad_->getBatchValue(n); - MatrixPtr gateGrad = batchGrad_->getBatchValue(*gate_.grad, n); - - lstmValue.gateValue = - batchGrad_->getBatchValue(*gate_.value, n)->getData(); - lstmValue.stateValue = - batchGrad_->getBatchValue(*state_.value, n)->getData(); - lstmValue.stateActiveValue = - batchGrad_->getBatchValue(*preOutput_.value, n)->getData(); - lstmGrad.stateGrad = - batchGrad_->getBatchValue(*state_.grad, n)->getData(); - lstmGrad.gateGrad = gateGrad->getData(); - lstmGrad.outputGrad = outputGrad->getData(); - { - batchSize = outputGrad->getHeight(); - if (n != 0) { - lstmValue.prevStateValue = - batchGrad_->getBatchValue(*state_.value, n - 1)->getData(); - lstmGrad.prevStateGrad = - batchGrad_->getBatchValue(*state_.grad, n - 1)->getData(); - } else { - if (prevState_) { - lstmValue.prevStateValue = totalState_->getData(); - lstmGrad.prevStateGrad = nullptr; - } else { - lstmValue.prevStateValue = nullptr; - lstmGrad.prevStateGrad = nullptr; - } - } - if (useGpu_) { - LstmCompute::backwardBatch<1>( - lstmValue, lstmGrad, getSize(), batchSize); - } else { - LstmCompute::backwardBatch<0>( - lstmValue, lstmGrad, getSize(), batchSize); - } - } - - if (n != 0) { - MatrixPtr tmp = batchGrad_->getBatchValue(n - 1, batchSize); - tmp->mul(*gateGrad, *weightT, 1, 1); - } - - if (n != 0 && weight_->getWGrad()) { - /* backward weight */ - MatrixPtr outputValue = batchValue_->getBatchValue(n - 1, batchSize); - weight_->getWGrad()->mul(*outputValue->getTranspose(), *gateGrad, 1, 1); - } else if (prevOutput_ && weight_->getWGrad()) { - weight_->getWGrad()->mul( - *prevBatchOutput2_->getTranspose(), *gateGrad, 1, 1); - } - } - } - - if (inputGrad) { - batchGrad_->add(*inputGrad, *gate_.grad, /* seq2batch */ false); - } - if (bias_ && bias_->getWGrad()) { - localBiasGrad_->collectBias(*gate_.grad, /* scale */ 1); - } -} - -void LstmLayer::forwardSeqParallel(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputValue) { - REGISTER_TIMER_INFO("LstmFwSeqParallelTime", getName().c_str()); - gate_.value->assign(*inputValue); - if (bias_) { - gate_.value->addBias(*localBias_, /* scale */ 1); - } - - real *gateValue = gate_.value->getData(); - real *stateValue = state_.value->getData(); - real *outputValue = output_.value->getData(); - real *preOutputValue = preOutput_.value->getData(); - real *checkIg = checkIg_->getData(); - real *checkFg = checkFg_->getData(); - real *checkOg = checkOg_->getData(); - real *weight = weight_->getW()->getData(); - hl_lstm_parallel_forward(gateValue, - stateValue, - preOutputValue, - outputValue, - checkIg, - checkFg, - checkOg, - weight, - starts, - getSize(), - numSequences, - reversed_, - activeNode_, - activeGate_, - activeState_); -} - -void LstmLayer::backwardSeqParallel(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputGrad) { - REGISTER_TIMER_INFO("LstmBwSeqParallelTime", getName().c_str()); - real *gateValue = gate_.value->getData(); - real *gateGrad = gate_.grad->getData(); - real *stateValue = state_.value->getData(); - real *stateGrad = state_.grad->getData(); - real *preOutputValue = preOutput_.value->getData(); - real *preOutputGrad = preOutput_.grad->getData(); - real *checkIg = checkIg_->getData(); - real *checkFg = checkFg_->getData(); - real *checkOg = checkOg_->getData(); - real *outputGrad = output_.grad->getData(); - real *weight = weight_->getW()->getData(); - - real *checkIgGrad; - real *checkFgGrad; - real *checkOgGrad; - if (bias_->getWGrad()) { - checkIgGrad = checkIgGrad_->getData(); - checkFgGrad = checkFgGrad_->getData(); - checkOgGrad = checkOgGrad_->getData(); - } else { - checkIgGrad = nullptr; - checkFgGrad = nullptr; - checkOgGrad = nullptr; - } - - hl_lstm_parallel_backward_data(gateValue, - gateGrad, - stateValue, - stateGrad, - preOutputValue, - preOutputGrad, - outputGrad, - checkIg, - checkIgGrad, - checkFg, - checkFgGrad, - checkOg, - checkOgGrad, - weight, - starts, - getSize(), - numSequences, - reversed_, - activeNode_, - activeGate_, - activeState_); - - if (inputGrad) { - inputGrad->add(*gate_.grad); - } - if (bias_ && bias_->getWGrad()) { - localBiasGrad_->collectBias(*gate_.grad, 1); - } - - real *outputValue = output_.value->getData(); - if (weight_->getWGrad()) { - real *weightGrad = weight_->getWGrad()->getData(); - hl_lstm_parallel_backward_weight(weightGrad, - outputValue, - gateGrad, - starts, - getSize(), - batchSize, - numSequences, - reversed_); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/LstmLayer.h b/paddle/gserver/layers/LstmLayer.h deleted file mode 100644 index 76dfe8146bf67a0b7b4fd4835851fae6ac38d80f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LstmLayer.h +++ /dev/null @@ -1,221 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "LstmCompute.h" -#include "SequenceToBatch.h" -#include "paddle/math/BaseMatrix.h" -#include "paddle/math/Matrix.h" -namespace paddle { - -/** - * @brief LstmLayer takes 1 input layer with size * 4. - * Input layer is diveded into 4 equal parts: - * (input_s, input_ig, input_fg, input_og) - * - * For each sequence [start, end] it performs the following computation: - * @code - * output_{i} = actState(state_{i}) * actGate(outputGate_{i}) - * state_{i} = actInput(input_s_{i} + bias_s + - * output_{i-1} * recurrIW) * actGate(inputGate_{i}) + - * actGate(forgetGate_{i}) * state_{i-1} - * inputGate = input_ig_{i} + bias_ig + output_{i-1} * recurrIGW + - * state_{i-1} * inputCheck - * ouputGate = input_og_{i} + bias_og + output_{i-1} * recurrOGW + - * state_{i} * outputCheck - * forgetGate = input_fg_{i} + bias_fg + output_{i-1} * recurrFGW + - * state_{i-1} * forgetCheck - * @endcode - * - * - parameter[0] consists of (recurrIW, recurrIGW, recurrFGW, recurrOGW) - * - baisParameter consists of - * (bias_s, bias_ig, bias_og, bias_fg, inputCheck, forgetCheck, outputCheck) - * - * - actInput is defined by config active_type. - * - actState is defined by config active_state_type. - * - actGate is defined by config actvie_gate_type. - * - * There are two ways to compute, namely one sequence by one sequence or - * one batch by one batch. By default and no setting pre_batch_state true, - * it will compute batch by batch. - * - * The formula in the paper is as follows: - * \f[ - * i_t = \sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i) \\ - * f_t = \sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f) \\ - * \tilde{c_t} = tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c) \\ - * o_t = \sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o) \\ - * c_t = f_t * c_{t-1} + i_t * \tilde{c_t} \\ - * h_t = o_t tanh(c_t) - * \f] - * - * @note These \f$W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}\f$ - * operations on the input sequence were NOT included in LstmLayer. So - * users should use fc_layer or mixed_layer before lstm_later. - * - * The weight ([size, 4*size]) contains \f$W_{hi}, W_{hf}, W_{hc}, W_{ho}\f$. - * The bias contains \f$b_i, b_f, b_c, b_o\f$ and \f$W_{ci}, W_{cf}, W_{co}\f$. - */ - -class LstmLayer : public Layer, public LstmCompute { - public: - explicit LstmLayer(const LayerConfig &config) : Layer(config) {} - - bool init(const LayerMap &layerMap, - const ParameterMap ¶meterMap) override; - - void forward(PassType passType) override; - - void backward(const UpdateCallback &callback) override; - - void resetState() override; - - void setState(LayerStatePtr state) override; - - LayerStatePtr getState() override; - - protected: - /** - * @brief Compute lstm forward one sequence by one sequence. - * @param batchSize The batchSize is not equal to the batch_size in - * the config file. It is the total words number of all samples - * in this forward batch. - * @param numSequences The sample number. It is equal to the batch_size - * in the config file. - * @param starts Each start position of each samples. - * @param inputValue The input values. - */ - void forwardSequence(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputValue); - /** - * Compute lstm backward one sequence by one sequence. - */ - void backwardSequence(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputGrad); - - /** - * Compute lstm forward one batch by one batch. The batch value is - * reorganized by SequenceToBatch class. The batch output value will - * be convert into sequence value after finishing forward. Here, one - * batch contains one word of each sample. If the length of each sample - * is not equality, the batch will not pads zero and contains less words. - * The total batch numbers are the max length of the sequence. The details - * can refer to SequenceToBatch class. On GPU mode, it will launch GPU - * kernel for loop. - * - * @code - * for (int i = 0; i < numBatch(max_sequence_length); ++i) { - * compute one batch. - * } - * @endcode - */ - void forwardBatch(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputValue); - /** - * Compute lstm backward one batch by one batch. - */ - void backwardBatch(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputGrad); - - /** - * This function only supports GPU. It not need to reorganize input into - * batch value. It will launch one kernel to parallelly compute forward - * propagation in sequence level. - */ - void forwardSeqParallel(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputValue); - /** - * Backward propagation corresponding to forwardSeqParallel. - */ - void backwardSeqParallel(int batchSize, - size_t numSequences, - const int *starts, - MatrixPtr inputGrad); - /** - * This function is used for sequence generation and get output after - * forwardBatch. - */ - void getPrevBatchOutput(size_t numSequences); - /** - * This function is used for sequence generation and get state after - * forwardBatch. - */ - void getPrevBatchState(size_t numSequences); - - protected: - /// Learned parameters, shape: (size, 4*size). - /// The weight ([size, 4*size]) contains \f$W_{hi}, W_{hf}, W_{hc}, W_{ho}\f$. - std::unique_ptr weight_; - /// Learned bias parameter, shape: (1, 7 * size). - /// The bias contains \f$b_i, b_f, b_c, b_o\f$ and \f$W_{ci}, W_{cf}, - /// W_{co}\f$. - std::unique_ptr bias_; - /// The reeal bias, point to \f$b_i, b_f, b_c, b_o\f$. - MatrixPtr localBias_; - /// The peephole connection for input gate. - MatrixPtr checkIg_; - /// The peephole connection for forget gate. - MatrixPtr checkFg_; - /// The peephole connection for output gate. - MatrixPtr checkOg_; - /// The gradient of real bias - MatrixPtr localBiasGrad_; - /// The gradient of peephole connection for input gates. - MatrixPtr checkIgGrad_; - /// The gradient of peephole connection for forget gates. - MatrixPtr checkFgGrad_; - /// The gradient of peephole connection for output gates. - MatrixPtr checkOgGrad_; - - /// Stores the cell state of previous time step, namely \f$c_{t-1}\f$. - Argument state_; - /// Stores the hidden of previous time step, namely \f$h_{t-1}\f$. - Argument preOutput_; - /// Stores the value and gradient of four gates, namely - /// \f$i_t, f_t, o_t, c_t\f$. - Argument gate_; - /// Whether it is reversed lstm. - bool reversed_; - /// Whether to use batch method to compute. - bool useBatch_; - /// Whether to use sequence parallell method to compute. - bool useSeqParallel_; - /// batchValue_ is used in method of batch calculation. It stores the - /// batch value after reorganized input. - std::unique_ptr batchValue_; - /// The gradient of batchValue_. - std::unique_ptr batchGrad_; - - /// Used in generation and stores the state of previous time step. - MatrixPtr prevState_; - /// Used in generation and stores the output of previous time step. - MatrixPtr prevOutput_; - MatrixPtr prevBatchOutput2_; - /// The total state. - MatrixPtr totalState_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/LstmStepLayer.cpp b/paddle/gserver/layers/LstmStepLayer.cpp deleted file mode 100644 index c44768ddb2b903763288465325899d86176df73a..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/LstmStepLayer.cpp +++ /dev/null @@ -1,194 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "LstmCompute.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/* - * LstmStepLayer used in recurrent layer group. - */ -class LstmStepLayer : public Layer, public LstmCompute { - protected: - Argument state_; - Argument gate_; - Argument stateActive_; - MatrixPtr checkIg_, checkFg_, checkOg_; - MatrixPtr checkIgGrad_, checkFgGrad_, checkOgGrad_; - std::unique_ptr weight_; - - public: - explicit LstmStepLayer(const LayerConfig& config) : Layer(config) {} - - ~LstmStepLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(lstm_step, LstmStepLayer); - -bool LstmStepLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!Layer::init(layerMap, parameterMap)) return false; - CHECK_EQ(2U, inputLayers_.size()); - - checkIg_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkFg_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkOg_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkIgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkFgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - checkOgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - getSize(), - /* trans= */ false, - useGpu_); - - if (biasParameter_.get() != NULL) { - CHECK_EQ(getSize() * 3, biasParameter_->getSize()); - weight_.reset(new Weight(1, getSize() * 3, biasParameter_)); - if (weight_->getW()) { - real* data = weight_->getW()->getData(); - checkIg_->setData(data); - checkFg_->setData(data + getSize()); - checkOg_->setData(data + getSize() * 2); - } - - if (weight_->getWGrad()) { - real* data = weight_->getWGrad()->getData(); - checkIgGrad_->setData(data); - checkFgGrad_->setData(data + getSize()); - checkOgGrad_->setData(data + getSize() * 2); - } - } - - setOutput("state", &state_); - LstmCompute::init(config_); - return true; -} - -void LstmStepLayer::forward(PassType passType) { - REGISTER_TIMER_INFO("LstmRecurrentFwTime", getName().c_str()); - Layer::forward(passType); - - const Argument& input = getInput(0); - const Argument& prevState = getInput(1); - CHECK_EQ(getSize() * 4, input.value->getWidth()); - CHECK_EQ(getSize(), prevState.value->getWidth()); - int batchSize = input.getBatchSize(); - reserveOutput(batchSize, getSize()); - resetSpecifyOutput(state_, - batchSize, - getSize(), - /* isValueClean */ false, - /* isGradClean */ true); - resetSpecifyOutput(gate_, - batchSize, - getSize() * 4, - /* isValueClean */ false, - /* isGradClean */ false); - resetSpecifyOutput(stateActive_, - batchSize, - getSize(), - /* isValueClean */ false, - /* isGradClean */ false); - gate_.value->assign(*input.value); - - hl_lstm_value lstmValue; - lstmValue.checkIg = checkIg_->getData(); - lstmValue.checkFg = checkFg_->getData(); - lstmValue.checkOg = checkOg_->getData(); - lstmValue.gateValue = gate_.value->getData(); - lstmValue.stateValue = state_.value->getData(); - lstmValue.prevStateValue = prevState.value->getData(); - lstmValue.stateActiveValue = stateActive_.value->getData(); - lstmValue.outputValue = output_.value->getData(); - - if (useGpu_) { - LstmCompute::forwardBatch<1>(lstmValue, getSize(), batchSize); - } else { - LstmCompute::forwardBatch<0>(lstmValue, getSize(), batchSize); - } -} - -void LstmStepLayer::backward(const UpdateCallback& callback) { - REGISTER_TIMER_INFO("LstmRecurrentBwTime", getName().c_str()); - const Argument& input = getInput(0); - const Argument& prevState = getInput(1); - int batchSize = input.getBatchSize(); - - hl_lstm_value lstmValue; - hl_lstm_grad lstmGrad; - lstmValue.checkIg = checkIg_->getData(); - lstmValue.checkFg = checkFg_->getData(); - lstmValue.checkOg = checkOg_->getData(); - lstmValue.gateValue = gate_.value->getData(); - lstmValue.prevStateValue = prevState.value->getData(); - lstmValue.stateValue = state_.value->getData(); - lstmValue.stateActiveValue = stateActive_.value->getData(); - - lstmGrad.gateGrad = gate_.grad->getData(); - if (prevState.grad) { - lstmGrad.prevStateGrad = prevState.grad->getData(); - } else { - lstmGrad.prevStateGrad = nullptr; - } - lstmGrad.stateGrad = state_.grad->getData(); - lstmGrad.stateActiveGrad = stateActive_.grad->getData(); - lstmGrad.outputGrad = output_.grad->getData(); - lstmGrad.checkIgGrad = checkIgGrad_->getData(); - lstmGrad.checkFgGrad = checkFgGrad_->getData(); - lstmGrad.checkOgGrad = checkOgGrad_->getData(); - - if (useGpu_) { - LstmCompute::backwardBatch<1>(lstmValue, lstmGrad, getSize(), batchSize); - } else { - LstmCompute::backwardBatch<0>(lstmValue, lstmGrad, getSize(), batchSize); - } - - if (input.grad) { - input.grad->add(*gate_.grad); - } - - if (weight_) { - weight_->getParameterPtr()->incUpdate(callback); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MDLstmLayer.cpp b/paddle/gserver/layers/MDLstmLayer.cpp deleted file mode 100644 index 22c28157c5a5b19aa54b3151a6c9a4cdcfb01765..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MDLstmLayer.cpp +++ /dev/null @@ -1,769 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "LstmLayer.h" -#include "paddle/math/BaseMatrix.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -class CoordIterator { - public: - std::vector dims_; - std::vector directions_; - std::vector curPos_; - bool end_; - - void step(size_t d, bool reversed) { - if (directions_[d] ^ reversed) { - if (curPos_[d] == dims_[d] - 1) { - curPos_[d] = 0; - if (d) { - step(d - 1, reversed); - } else { - end_ = true; - } - } else { - curPos_[d]++; - } - } else { - if (curPos_[d] == 0) { - curPos_[d] = dims_[d] - 1; - if (d) { - step(d - 1, reversed); - } else { - end_ = true; - } - } else { - curPos_[d]--; - } - } - } - - public: - CoordIterator(std::vector dim, std::vector directions) - : dims_(dim), directions_(directions), end_(false) { - CHECK_EQ(dims_.size(), directions_.size()); - for (size_t i = 0; i < dims_.size(); i++) { - curPos_.push_back(-1); - } - } - CoordIterator& operator++() { - step(dims_.size() - 1, false); - return *this; - } - - CoordIterator& operator--() { - step(dims_.size() - 1, true); - return *this; - } - - std::vector& curPos() { return curPos_; } - - int offset() { - int offset = curPos_[0]; - for (size_t i = 1; i < dims_.size(); i++) { - offset = offset * dims_[i] + curPos_[i]; - } - return offset; - } - - int offset(const std::vector& pos) { - int offset = pos[0]; - for (size_t i = 1; i < dims_.size(); i++) { - offset = offset * dims_[i] + pos[i]; - } - return offset; - } - - std::vector& begin() { - for (size_t i = 0; i < dims_.size(); i++) { - curPos_[i] = directions_[i] ? 0 : dims_[i] - 1; - } - end_ = false; - return curPos_; - } - - std::vector& rbegin() { - for (size_t i = 0; i < dims_.size(); i++) { - curPos_[i] = directions_[i] ? dims_[i] - 1 : 0; - } - end_ = false; - return curPos_; - } - - bool end() { return end_; } - - bool getPrePos(const std::vector& delays, - int idx, - std::vector& prePos) { - bool isAvial = true; - prePos.clear(); - prePos.reserve(directions_.size()); - for (size_t i = 0; i < directions_.size(); i++) { - if (int(i) == idx) { - prePos.push_back(curPos_[i] + delays[i] * (directions_[i] ? 1 : -1)); - if (prePos[i] < 0) { - prePos[i] = 0; - isAvial = false; - } - if (prePos[i] >= dims_[i]) { - prePos[i] = dims_[i] - 1; - isAvial = false; - } - } else { - prePos.push_back(curPos_[i]); - } - } - return isAvial; - } - - bool getNextPos(const std::vector& delays, - int idx, - std::vector& nextPos) { - bool isAvial = true; - nextPos.clear(); - nextPos.reserve(directions_.size()); - for (size_t i = 0; i < directions_.size(); i++) { - if (int(i) == idx) { - nextPos.push_back(curPos_[i] - delays[i] * (directions_[i] ? 1 : -1)); - if (nextPos[i] < 0) { - nextPos[i] = 0; - isAvial = false; - } - if (nextPos[i] >= dims_[i]) { - nextPos[i] = dims_[i] - 1; - isAvial = false; - } - } else { - nextPos.push_back(curPos_[i]); - } - } - return isAvial; - } -}; -/* - * MDLstmLayer takes 1 input layer with size * (3+numDims). - * For each sequence [start, end] it performs the following computation: - * out_i = actState(state_i) * actGate(outputGate_i) - * - * For example the image with 2 dims, we take the scanning order from left-top - * to right-bottom, then the 2 previous states of the current pixels are the - * ones located at left and top. And each of them has a independent forget gate. - * - * state_i = actInput(input_i) * actGate(inputGate_i) + - * \sum{j}(actGate(forgetGate_i_j) * state_prev_i_j) - * - * inputGate = input_i * inputW + \sum{j}(output_prev_i_j * recurrInputW_j) + - * \sum{j}(state_prev_i_j * inputCheck_j) - * - * ouputGate = input_i * outputW + \sum{j}(output_prev_i_j * recurrOutputW_j) + - * state_i * outputCheck - * - * forgetGate_j = input_i * forgetW_j + \sum{j}(output_prev_i_j * - * recurrForgetW_j) + \sum{j}(state_prev_i_j * forgetCheck_j) - * - * IG Layer: (Input, InputGate, ForgetGates, OutputGate) * OutputSize - * */ - -class MDLstmLayer : public LstmLayer { - public: - explicit MDLstmLayer(const LayerConfig& config) : LstmLayer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - - void backward(const UpdateCallback& callback) override; - - protected: - void forwardOneSequence(int start, CoordIterator& coordIter); - void backwardOneSequence(int start, CoordIterator& coordIter); - void forwardGate2OutputSequence(int start, CoordIterator& coordIter); - void backwardGate2OutputSequence(int start, CoordIterator& coordIter); - - protected: - std::vector frameInputGate_; - std::vector frameForgetGate_; - std::vector frameOutputGate_; - std::vector frameInputNode_; - std::vector frameGate_; - std::vector frameState_; - std::vector framePreOutput_; - std::vector frameOutput_; - - // Activation - std::unique_ptr activationGate_; - std::unique_ptr activationState_; - - int numDims_; - size_t numBlocks_; - std::vector directions_; - std::vector delays_; - std::vector> dimsV_; -}; - -REGISTER_LAYER(mdlstmemory, MDLstmLayer); - -bool MDLstmLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!Layer::init(layerMap, parameterMap)) return false; - CHECK_EQ(1U, inputLayers_.size()); - CHECK_EQ(1U, parameters_.size()); - - numBlocks_ = getSize(); - numDims_ = config_.directions_size(); - CHECK_EQ(numBlocks_ * numBlocks_ * (3 + numDims_), parameters_[0]->getSize()); - - // inode(1), ig(1), fg(numDims_), og(1), peepIg(1), peepFg(numDims_), - // peepOg(1), then size of localBias_ is 3+numDims_ - CHECK_EQ(numBlocks_ * (5 + 2 * numDims_), biasParameter_->getSize()); - weight_.reset( - new Weight(numBlocks_, numBlocks_ * (3 + numDims_), parameters_[0])); - if (biasParameter_.get() != NULL) { - bias_.reset(new Weight(1, numBlocks_ * (5 + 2 * numDims_), biasParameter_)); - localBias_ = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_ * (3 + numDims_), - /* trans= */ false, - useGpu_); - checkIg_ = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - checkFg_ = Matrix::create(nullptr, - /* height= */ numDims_, - numBlocks_, - /* trans= */ false, - useGpu_); - checkOg_ = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - localBiasGrad_ = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_ * (3 + numDims_), - /* trans= */ false, - useGpu_); - checkIgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - checkFgGrad_ = Matrix::create(nullptr, - /* height= */ numDims_, - numBlocks_, - /* trans= */ false, - useGpu_); - checkOgGrad_ = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - - localBias_->setData(bias_->getW()->getData()); - checkIg_->setData(bias_->getW()->getData() + numBlocks_ * (3 + numDims_)); - checkFg_->setData(bias_->getW()->getData() + numBlocks_ * (4 + numDims_)); - checkOg_->setData(bias_->getW()->getData() + - numBlocks_ * (4 + 2 * numDims_)); - - if (bias_->getWGrad()) { - localBiasGrad_->setData(bias_->getWGrad()->getData()); - checkIgGrad_->setData(bias_->getWGrad()->getData() + - numBlocks_ * (3 + numDims_)); - checkFgGrad_->setData(bias_->getWGrad()->getData() + - numBlocks_ * (4 + numDims_)); - checkOgGrad_->setData(bias_->getWGrad()->getData() + - numBlocks_ * (4 + 2 * numDims_)); - } - } else { - LOG(FATAL) << "Bias should be here."; - } - for (int i = 0; i < numDims_; i++) { - directions_.push_back(config_.directions(i)); - } - for (int i = 0; i < numDims_; i++) { - delays_.push_back(-1); - } - activationGate_.reset(ActivationFunction::create(config_.active_gate_type())); - activationState_.reset( - ActivationFunction::create(config_.active_state_type())); - - return true; -} - -void MDLstmLayer::forward(PassType passType) { - Layer::forward(passType); - - const Argument& input = getInput(0); - CHECK(input.sequenceStartPositions); - int batchSize = input.getBatchSize(); - int numSequences = input.getNumSequences(); - resetOutput(batchSize, numBlocks_); - CHECK_EQ(numBlocks_ * (3 + numDims_), input.value->getWidth()); - const int* starts = input.sequenceStartPositions->getData(false); - CHECK_EQ(starts[numSequences], batchSize); - - int* dimsData = input.cpuSequenceDims->getData(); - CHECK_EQ(int(input.cpuSequenceDims->getSize()), numDims_* numSequences); - - for (int i = 0; i < numSequences; i++) { - std::vector dims; - for (int j = 0; j < numDims_; j++) { - dims.push_back(dimsData[i * numDims_ + j]); - } - dimsV_.push_back(dims); - } - - frameInputGate_.reserve(batchSize); - frameForgetGate_.reserve(batchSize); - frameOutputGate_.reserve(batchSize); - frameInputNode_.reserve(batchSize); - frameGate_.reserve(batchSize); - frameState_.reserve(batchSize); - framePreOutput_.reserve(batchSize); - frameOutput_.reserve(batchSize); - - Matrix::resizeOrCreate(gate_.value, - /* height= */ batchSize, - numBlocks_ * (3 + numDims_), - /* trans= */ false, - useGpu_); - - for (int i = frameGate_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_ * (3 + numDims_), - /* trans= */ false, - useGpu_); - arg.grad = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_ * (3 + numDims_), - /* trans= */ false, - useGpu_); - frameGate_.push_back(arg); - } - for (int i = frameInputGate_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - arg.grad = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - frameInputGate_.push_back(arg); - } - for (int i = frameForgetGate_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create(nullptr, - /* height= */ numDims_, - numBlocks_, - /* trans= */ false, - useGpu_); - arg.grad = Matrix::create(nullptr, - /* height= */ numDims_, - numBlocks_, - /* trans= */ false, - useGpu_); - frameForgetGate_.push_back(arg); - } - for (int i = frameOutputGate_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - arg.grad = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - frameOutputGate_.push_back(arg); - } - for (int i = frameInputNode_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - arg.grad = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - frameInputNode_.push_back(arg); - } - for (int i = frameState_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create( - /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); - frameState_.push_back(arg); - } - for (int i = framePreOutput_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create( - /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); - framePreOutput_.push_back(arg); - } - for (int i = frameOutput_.size(); i < batchSize; i++) { - Argument arg; - arg.value = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - arg.grad = Matrix::create(nullptr, - /* height= */ 1, - numBlocks_, - /* trans= */ false, - useGpu_); - frameOutput_.push_back(arg); - } - - for (int i = 0; i < batchSize; i++) { - frameOutput_[i].value->setData(output_.value->getData() + i * numBlocks_); - frameGate_[i].value->setData(gate_.value->getData() + - i * numBlocks_ * (3 + numDims_)); - frameInputNode_[i].value->setData(gate_.value->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * 0); - frameInputGate_[i].value->setData(gate_.value->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * 1); - frameForgetGate_[i].value->setData(gate_.value->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * 2); - frameOutputGate_[i].value->setData(gate_.value->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * (2 + numDims_)); - } - - AsyncGpuBlock asyncGpuBlock; - gate_.value->assign(*input.value); - - if (bias_) { - gate_.value->addBias(*localBias_, 1); - } - - for (int i = 0; i < numSequences; i++) { - CoordIterator coordIter(dimsV_[i], directions_); - forwardOneSequence(starts[i], coordIter); - } -} - -void MDLstmLayer::forwardGate2OutputSequence(int start, - CoordIterator& coordIter) { - int idxCurr = start + coordIter.offset(); - std::vector preOffsetV; - preOffsetV.reserve(numDims_); - for (int i = 0; i < numDims_; i++) { - std::vector prePos; - if (coordIter.getPrePos(delays_, i, prePos)) { - preOffsetV[i] = coordIter.offset(prePos); - } else { - preOffsetV[i] = -1; - } - } - - for (int i = 0; i < numDims_; i++) { - if (preOffsetV[i] >= 0) { - frameInputGate_[idxCurr].value->addDotMul( - *frameState_[start + preOffsetV[i]].value, *checkIg_, 1.0, 1.0); - - MatrixPtr fgGateOneDim = Matrix::create( - frameForgetGate_[idxCurr].value->getData() + i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - MatrixPtr checkFgOneDim = - Matrix::create(checkFg_->getData() + i * numBlocks_, - 1.0, - numBlocks_, - false, - useGpu_); - fgGateOneDim->addDotMul( - *frameState_[start + preOffsetV[i]].value, *checkFgOneDim, 1.0, 1.0); - } - } - auto status = activationGate_->forward(frameInputGate_[idxCurr]); - status.check(); - status = activationGate_->forward(frameForgetGate_[idxCurr]); - status.check(); - status = activation_->forward(frameInputNode_[idxCurr]); - status.check(); - - frameState_[idxCurr].value->zeroMem(); - for (int i = 0; i < numDims_; i++) { - if (preOffsetV[i] >= 0) { - MatrixPtr fgGateOneDim = Matrix::create( - frameForgetGate_[idxCurr].value->getData() + i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - frameState_[idxCurr].value->addDotMul( - *frameState_[start + preOffsetV[i]].value, *fgGateOneDim, 1.0, 1.0); - } - } - frameState_[idxCurr].value->addDotMul(*frameInputNode_[idxCurr].value, - *frameInputGate_[idxCurr].value, - 1.0, - 1.0); - - frameOutputGate_[idxCurr].value->addDotMul( - *frameState_[idxCurr].value, *checkOg_, 1.0, 1.0); - status = activationGate_->forward(frameOutputGate_[idxCurr]); - status.check(); - - framePreOutput_[idxCurr].value->copyFrom(*(frameState_[idxCurr].value)); - status = activationState_->forward(framePreOutput_[idxCurr]); - status.check(); - - frameOutput_[idxCurr].value->dotMul(*framePreOutput_[idxCurr].value, - *frameOutputGate_[idxCurr].value); -} - -void MDLstmLayer::forwardOneSequence(int start, CoordIterator& coordIter) { - for (coordIter.begin(); !coordIter.end(); ++coordIter) { - int offset = coordIter.offset(); - for (int i = 0; i < numDims_; i++) { - std::vector prePos; - if (coordIter.getPrePos(delays_, i, prePos)) { - int preOffset = coordIter.offset(prePos); - frameGate_[start + offset].value->mul( - *frameOutput_[start + preOffset].value, *weight_->getW(), 1.0, 1.0); - } - } - forwardGate2OutputSequence(start, coordIter); - } -} - -void MDLstmLayer::backward(const UpdateCallback& callback) { - const Argument& input = getInput(0); - CHECK(input.sequenceStartPositions); - int batchSize = input.getBatchSize(); - const int* starts = input.sequenceStartPositions->getData(false); - size_t numSequences = input.getNumSequences(); - - Matrix::resizeOrCreate(gate_.grad, - /* height= */ batchSize, - numBlocks_ * (3 + numDims_), - /* trans= */ false, - useGpu_); - - for (int i = 0; i < batchSize; i++) { - if (frameState_[i].grad == NULL) - frameState_[i].grad = Matrix::create( - /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); - } - for (int i = 0; i < batchSize; i++) { - if (framePreOutput_[i].grad == NULL) - framePreOutput_[i].grad = Matrix::create( - /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); - } - - for (int i = 0; i < batchSize; i++) { - frameOutput_[i].grad->setData(output_.grad->getData() + i * numBlocks_); - frameGate_[i].grad->setData(gate_.grad->getData() + - i * numBlocks_ * (3 + numDims_)); - frameInputNode_[i].grad->setData(gate_.grad->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * 0); - frameInputGate_[i].grad->setData(gate_.grad->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * 1); - frameForgetGate_[i].grad->setData(gate_.grad->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * 2); - frameOutputGate_[i].grad->setData(gate_.grad->getData() + - i * numBlocks_ * (3 + numDims_) + - numBlocks_ * (2 + numDims_)); - } - - { - AsyncGpuBlock asyncGpuBlock; - - for (size_t i = 0; i < numSequences; i++) { - CoordIterator coordIter(dimsV_[i], directions_); - backwardOneSequence(starts[i], coordIter); - } - } - - if (input.grad) { - input.grad->add(*gate_.grad); - } - if (bias_ && bias_->getWGrad()) { - localBiasGrad_->collectBias(*gate_.grad, 1); - bias_->getParameterPtr()->incUpdate(callback); - } - - weight_->getParameterPtr()->incUpdate(callback); -} - -void MDLstmLayer::backwardGate2OutputSequence(int start, - CoordIterator& coordIter) { - int idxCurr = start + coordIter.offset(); - std::vector preOffsetV; - std::vector nextOffsetV; - preOffsetV.reserve(numDims_); - nextOffsetV.reserve(numDims_); - for (int i = 0; i < numDims_; i++) { - std::vector prePos; - if (coordIter.getPrePos(delays_, i, prePos)) { - preOffsetV[i] = coordIter.offset(prePos); - } else { - preOffsetV[i] = -1; - } - std::vector nextPos; - if (coordIter.getNextPos(delays_, i, nextPos)) { - nextOffsetV[i] = coordIter.offset(nextPos); - } else { - nextOffsetV[i] = -1; - } - } - - framePreOutput_[idxCurr].grad->dotMul(*frameOutput_[idxCurr].grad, - *frameOutputGate_[idxCurr].value); - activationState_->backward(framePreOutput_[idxCurr]).check(); - frameState_[idxCurr].grad->copyFrom(*(framePreOutput_[idxCurr].grad)); - - frameOutputGate_[idxCurr].grad->dotMul(*frameOutput_[idxCurr].grad, - *framePreOutput_[idxCurr].value); - activationGate_->backward(frameOutputGate_[idxCurr]).check(); - - frameState_[idxCurr].grad->addDotMul( - *frameOutputGate_[idxCurr].grad, *checkOg_, 1.0, 1.0); - for (int i = 0; i < numDims_; i++) { - if (nextOffsetV[i] >= 0) { - frameState_[idxCurr].grad->addDotMul( - *frameInputGate_[start + nextOffsetV[i]].grad, *checkIg_, 1.0, 1.0); - - MatrixPtr fgGateOneDimGrad = Matrix::create( - frameForgetGate_[start + nextOffsetV[i]].grad->getData() + - i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - MatrixPtr fgGateOneDimVal = Matrix::create( - frameForgetGate_[start + nextOffsetV[i]].value->getData() + - i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - MatrixPtr checkFgOneDim = Matrix::create( - checkFg_->getData() + i * numBlocks_, 1, numBlocks_, false, useGpu_); - - frameState_[idxCurr].grad->addDotMul( - *fgGateOneDimGrad, *checkFgOneDim, 1.0, 1.0); - frameState_[idxCurr].grad->addDotMul( - *frameState_[start + nextOffsetV[i]].grad, - *fgGateOneDimVal, - 1.0, - 1.0); - } - } - - frameInputNode_[idxCurr].grad->dotMul(*frameState_[idxCurr].grad, - *frameInputGate_[idxCurr].value); - frameInputGate_[idxCurr].grad->dotMul(*frameState_[idxCurr].grad, - *frameInputNode_[idxCurr].value); - - frameForgetGate_[idxCurr].grad->zeroMem(); - for (int i = 0; i < numDims_; i++) { - if (preOffsetV[i] >= 0) { - MatrixPtr fgGateOneDimGrad = Matrix::create( - frameForgetGate_[idxCurr].grad->getData() + i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - fgGateOneDimGrad->addDotMul(*frameState_[idxCurr].grad, - *frameState_[start + preOffsetV[i]].value, - 1.0, - 1.0); - } - } - - activationGate_->backward(frameInputGate_[idxCurr]).check(); - activationGate_->backward(frameForgetGate_[idxCurr]).check(); - activation_->backward(frameInputNode_[idxCurr]).check(); - - if (bias_->getWGrad()) { - for (int i = 0; i < numDims_; i++) { - if (preOffsetV[i] >= 0) { - checkIgGrad_->addDotMul(*frameInputGate_[idxCurr].grad, - *frameState_[start + preOffsetV[i]].value, - 1.0, - 1.0); - - MatrixPtr fgGateOneDimGrad = Matrix::create( - frameForgetGate_[idxCurr].grad->getData() + i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - MatrixPtr checkFgOneDimGrad = - Matrix::create(checkFgGrad_->getData() + i * numBlocks_, - 1, - numBlocks_, - false, - useGpu_); - checkFgOneDimGrad->addDotMul(*fgGateOneDimGrad, - *frameState_[start + preOffsetV[i]].value, - 1.0, - 1.0); - } - } - checkOgGrad_->addDotMul( - *frameOutputGate_[idxCurr].grad, *frameState_[idxCurr].value, 1.0, 1.0); - } -} - -void MDLstmLayer::backwardOneSequence(int start, CoordIterator& coordIter) { - MatrixPtr weightT = weight_->getW()->getTranspose(); - for (coordIter.rbegin(); !coordIter.end(); --coordIter) { - int offset = coordIter.offset(); - backwardGate2OutputSequence(start, coordIter); - for (int i = 0; i < numDims_; i++) { - std::vector prePos; - if (coordIter.getPrePos(delays_, i, prePos)) { - int preOffset = coordIter.offset(prePos); - frameOutput_[start + preOffset].grad->mul( - *frameGate_[start + offset].grad, *weightT, 1.0, 1.0); - if (weight_->getWGrad()) { - weight_->getWGrad()->mul( - *frameOutput_[start + preOffset].value->getTranspose(), - *frameGate_[start + offset].grad, - 1.0, - 1.0); - } - } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp deleted file mode 100644 index a442a0a01369f4ceb27ba4a1976df7f6e25b832f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MKLDNNConvLayer.cpp +++ /dev/null @@ -1,388 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MKLDNNConvLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/Logging.h" - -using namespace mkldnn; // NOLINT -typedef memory::format format; - -namespace paddle { - -REGISTER_LAYER(mkldnn_conv, MKLDNNConvLayer); - -bool MKLDNNConvLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!MKLDNNLayer::init(layerMap, parameterMap)) { - return false; - } - CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; - CHECK_EQ(inputLayers_.size(), parameters_.size()); - CHECK(config_.shared_biases()) << "Only support shared biases yet"; - - oc_ = config_.num_filters(); - const ConvConfig& conf = config_.inputs(0).conv_conf(); - ic_ = conf.channels(); - fw_ = conf.filter_size(); - fh_ = conf.filter_size_y(); - pw_ = conf.padding(); - ph_ = conf.padding_y(); - dw_ = conf.dilation(); - dh_ = conf.dilation_y(); - sw_ = conf.stride(); - sh_ = conf.stride_y(); - gp_ = conf.groups(); - oh_ = conf.output_y(); - ow_ = conf.output_x(); - ih_ = conf.img_size_y(); - iw_ = conf.img_size(); - caffeMode_ = conf.caffe_mode(); - CHECK(caffeMode_) << "Only support caffe mode yet"; - CHECK(dh_ == 1 && dw_ == 1) << "Only support dilation 1 yet"; - // check group setting - CHECK_EQ((oc_ / gp_) * gp_, oc_) << "group is indivisible for oc"; - CHECK_EQ((ic_ / gp_) * gp_, ic_) << "group is indivisible for ic"; - - // create weight - size_t height = oc_ / gp_; - size_t width = ic_ * fh_ * fw_; - CHECK_EQ(parameters_[0]->getSize(), height * width); - weight_ = - std::unique_ptr(new Weight(height, width, parameters_[0], 0)); - - // create biases - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_, 0)); - } - return true; -} - -void MKLDNNConvLayer::convertWeightsFromPaddle() { - if (hasInitedWgt_) { - return; - } - - CHECK(wgtVal_) << "should have been initialized"; - // the paddle weight format is oihw or goihw - auto targetDim = wgtVal_->getDims(); - auto srcFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw; - wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim); - hasInitedWgt_ = true; -} - -void MKLDNNConvLayer::convertWeightsToPaddle() { - CHECK(wgtVal_) << "should have been initialized"; - auto targetDim = wgtVal_->getDims(); - auto dstFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw; - wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); -} - -void MKLDNNConvLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { - reshapeInput(bs, ih, iw); - - // cal output sizes - // oc can not be changed - int fh = (fh_ - 1) * dh_ + 1; - int fw = (fw_ - 1) * dw_ + 1; - oh = outputSize(ih, fh, ph_, sh_, caffeMode_); - ow = outputSize(iw, fw, pw_, sw_, caffeMode_); - - reshapeOutput(oh, ow); - resizeOutput(bs, oc * oh * ow); -} - -void MKLDNNConvLayer::resetFwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - resetFwdPD(fwdPD_); - - resetFwdBuffers(fwdPD_, inputs[0], wgtVal_, biasVal_, out); - - resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, biasVal_, out); -} - -void MKLDNNConvLayer::resetBwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - std::shared_ptr bwdWgtPD; - std::shared_ptr bwdDataPD; - - resetBwdWgtPD(bwdWgtPD); - - resetBwdDataPD(bwdDataPD); - - resetBwdBuffers(bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); - - resetBwdPipeline( - pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); -} - -void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) { - weight_->getParameterPtr()->incUpdate(callback); - if (biases_ && biases_->getWGrad()) { - biases_->getParameterPtr()->incUpdate(callback); - } -} - -void MKLDNNConvLayer::loadConvSettings(memory::dims& wgt, - memory::dims& bias, - memory::dims& stride, - memory::dims& dilation, - memory::dims& padL, - memory::dims& padR) { - wgt = (gp_ == 1) ? memory::dims{oc_, ic_, fh_, fw_} - : memory::dims{gp_, oc_ / gp_, ic_ / gp_, fh_, fw_}; - bias = memory::dims{oc_}; - stride = memory::dims{sh_, sw_}; - padL = memory::dims{ph_, pw_}; - padR = getPaddingR(); - // note: mkldnn dilation start from 0 - dilation = memory::dims{dh_ - 1, dw_ - 1}; -} - -void MKLDNNConvLayer::resetFwdPD( - std::shared_ptr& pd) { - // dims for conv - memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_}; - memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_}; - memory::dims wgtDims, biasDims, strides, dilations, padL, padR; - loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); - - prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring - : prop_kind::forward_training; - algorithm algo = algorithm::convolution_direct; - padding_kind padKind = padding_kind::zero; - conv_fwd::desc fwdDesc = - biases_ && biases_->getW() - ? conv_fwd::desc(pk, - algo, - MKLDNNMatrix::createMemoryDesc(inDims), - MKLDNNMatrix::createMemoryDesc(wgtDims), - MKLDNNMatrix::createMemoryDesc(biasDims), - MKLDNNMatrix::createMemoryDesc(outDims), - strides, - dilations, - padL, - padR, - padKind) - : conv_fwd::desc(pk, - algo, - MKLDNNMatrix::createMemoryDesc(inDims), - MKLDNNMatrix::createMemoryDesc(wgtDims), - MKLDNNMatrix::createMemoryDesc(outDims), - strides, - dilations, - padL, - padR, - padKind); - pd.reset(new conv_fwd::primitive_desc(fwdDesc, engine_)); -} - -void MKLDNNConvLayer::resetFwdBuffers( - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - CHECK(pd); - resetInValue( - in, std::make_shared(pd->src_primitive_desc())); - - resetOutValue(out, pd->dst_primitive_desc()); - - resetWithMatrix(wgt, weight_->getW(), pd->weights_primitive_desc()); - - if (biases_ && biases_->getW()) { - resetWithMatrix(bias, biases_->getW(), pd->bias_primitive_desc()); - } else { - bias = nullptr; - } -} - -void MKLDNNConvLayer::resetFwdPipeline( - std::vector& pipeline, - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - if (bias) { - fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out)); - } else { - fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out)); - } - pipeline.push_back(*fwd_); -} - -void MKLDNNConvLayer::resetBwdWgtPD( - std::shared_ptr& pd) { - memory::dims wgtDims, biasDims, strides, dilations, padL, padR; - loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); - - // create backward weight using input, output and weight value memory desc - CHECK(inVals_[0]) << "Should have internal input value"; - CHECK(outVal_) << "Should have internal output value"; - CHECK(wgtVal_) << "Should have weight value"; - algorithm algo = algorithm::convolution_direct; - padding_kind padKind = padding_kind::zero; - auto bwdWgtDesc = biasVal_ != nullptr - ? conv_bwdWgt::desc(algo, - inVals_[0]->getMemoryDesc(), - wgtVal_->getMemoryDesc(), - biasVal_->getMemoryDesc(), - outVal_->getMemoryDesc(), - strides, - padL, - padR, - padKind) - : conv_bwdWgt::desc(algo, - inVals_[0]->getMemoryDesc(), - wgtVal_->getMemoryDesc(), - outVal_->getMemoryDesc(), - strides, - padL, - padR, - padKind); - pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_)); - CHECK_PRIMITIVE_DESC_EQ(inVals_[0], pd->src_primitive_desc()); - CHECK_PRIMITIVE_DESC_EQ( - outVal_, - pd->diff_dst_primitive_desc(), - "primitive desc of out value and grad should be equal"); - CHECK_PRIMITIVE_DESC_EQ( - wgtVal_, - pd->diff_weights_primitive_desc(), - "primitive desc of weight value and grad should be equal"); -} - -void MKLDNNConvLayer::resetBwdDataPD( - std::shared_ptr& pd) { - pd = nullptr; - if (inputLayers_[0]->getOutput().grad == nullptr) { - return; - } - - memory::dims wgtDims, biasDims, strides, dilations, padL, padR; - loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); - CHECK(inVals_[0]) << "Should have internal input value"; - CHECK(outVal_) << "Should have internal output value"; - // create backward data using input and output value memory desc - // but using weight memory desc with any format - auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct, - inVals_[0]->getMemoryDesc(), - MKLDNNMatrix::createMemoryDesc(wgtDims), - outVal_->getMemoryDesc(), - strides, - padL, - padR, - padding_kind::zero); - pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_)); - CHECK_PRIMITIVE_DESC_EQ( - inVals_[0], - pd->diff_src_primitive_desc(), - "primitive desc of in value and grad should be equal"); - CHECK_PRIMITIVE_DESC_EQ( - outVal_, - pd->diff_dst_primitive_desc(), - "primitive desc of out value and grad should be equal"); -} - -void MKLDNNConvLayer::resetBwdBuffers( - std::shared_ptr& wgtPD, - std::shared_ptr& dataPD, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - CHECK(wgtPD); - resetOutGrad(out, wgtPD->diff_dst_primitive_desc()); - - resetWithMatrix( - wgt, weight_->getWGrad(), wgtPD->diff_weights_primitive_desc()); - CHECK_PRIMITIVE_DESC_EQ( - wgtVal_, - wgt->getPrimitiveDesc(), - "primitive desc of weight grad and value should be equal"); - - bias = nullptr; - if (biases_ && biases_->getWGrad()) { - resetWithMatrix( - bias, biases_->getWGrad(), wgtPD->diff_bias_primitive_desc()); - CHECK(bias); - CHECK_PRIMITIVE_DESC_EQ( - biasVal_, - bias->getPrimitiveDesc(), - "primitive desc of bias grad and value should be equal"); - } - - if (dataPD == nullptr) { - return; - } - resetInGrad(in, dataPD->diff_src_primitive_desc()); - resetWgtValBwdData(dataPD, wgtValBwdData_); -} - -void MKLDNNConvLayer::resetBwdPipeline( - std::vector& pipeline, - std::shared_ptr& wgtPD, - std::shared_ptr& dataPD, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - CHECK(inVals_[0]); - // add bwdWgt handle - if (bias) { - bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVals_[0], *out, *wgt, *bias)); - } else { - bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVals_[0], *out, *wgt)); - } - pipeline.push_back(*bwdWgt_); - - if (dataPD == nullptr) { - return; - } - if (cvtWgtVal_) { - pipeline.push_back(*cvtWgtVal_); - } - // add bwdData handle - CHECK(wgtValBwdData_) << "Should have weight memory"; - bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in)); - pipeline.push_back(*bwdData_); -} - -void MKLDNNConvLayer::resetWgtValBwdData( - std::shared_ptr& dataPD, - MKLDNNMatrixPtr& wgt) { - if (dataPD == nullptr) { - return; - } - - // create new weight value for backward data, and create reorder if necessary - // since the primitive_desc would be different with wgtVal_ - CHECK(wgtVal_) << "should have weight value"; - if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) { - wgtValBwdData_ = MKLDNNMatrix::create(dataPD->weights_primitive_desc()); - cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_); - CHECK(cvtWgtVal_); - } else { - wgtValBwdData_ = wgtVal_; - } - VLOG(MKLDNN_FMTS) << "weight value format for backward data: " - << wgtValBwdData_->getFormat(); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp deleted file mode 100644 index 0c7e6f16e24a65b552cebcbd2111926cefc211f4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MKLDNNFcLayer.cpp +++ /dev/null @@ -1,262 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MKLDNNFcLayer.h" -#include "paddle/utils/Logging.h" - -using namespace mkldnn; // NOLINT -typedef memory::format format; - -namespace paddle { - -REGISTER_LAYER(mkldnn_fc, MKLDNNFcLayer); - -bool MKLDNNFcLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!MKLDNNLayer::init(layerMap, parameterMap)) { - return false; - } - - CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; - CHECK_EQ(inputLayers_.size(), parameters_.size()); - CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet"; - - // output size, cat not be changed - oc_ = getSize(); - oh_ = 1; - ow_ = 1; - ih_ = 1; - iw_ = 1; - - // input size can not change in FC - iLayerSize_ = inputLayers_[0]->getSize(); - CHECK_EQ(parameters_[0]->getSize(), iLayerSize_ * oc_); - - // create weight - weight_ = - std::unique_ptr(new Weight(oc_, iLayerSize_, parameters_[0], 0)); - - // create biases - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_, 0)); - } - return true; -} - -void MKLDNNFcLayer::convertWeightsFromPaddle() { - if (hasInitedWgt_) { - return; - } - - CHECK(wgtVal_) << "should have been initialized"; - auto targetDim = wgtVal_->getDims(); - auto srcFmt = targetDim.size() == 2 ? format::io : format::ihwo; - wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim); - hasInitedWgt_ = true; -} - -void MKLDNNFcLayer::convertWeightsToPaddle() { - CHECK(wgtVal_) << "should have been initialized"; - auto targetDim = wgtVal_->getDims(); - auto dstFmt = targetDim.size() == 2 ? format::io : format::ihwo; - wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); -} - -void MKLDNNFcLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { - reshapeInput(bs, ih, iw); - - CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize()); - ic = iLayerSize_ / (ih * iw); - CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible"; - CHECK_EQ(size_t(oc), getSize()); - - reshapeOutput(oh, ow); - resizeOutput(bs, oc); -} - -void MKLDNNFcLayer::resetFwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - resetFwdBuffers(inputs[0], wgtVal_, biasVal_, out); - - resetFwdPD(fwdPD_, inputs[0], wgtVal_, biasVal_, out); - - resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, biasVal_, out); -} - -void MKLDNNFcLayer::resetBwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - std::shared_ptr bwdWgtPD; - std::shared_ptr bwdDataPD; - - resetBwdBuffers(inputs[0], wgtGrad_, biasGrad_, out); - - resetBwdWgtPD(bwdWgtPD, wgtGrad_, biasGrad_, out); - - resetBwdDataPD(bwdDataPD, inputs[0], out); - - resetBwdPipeline( - pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); -} - -void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) { - weight_->getParameterPtr()->incUpdate(callback); - if (biases_ && biases_->getWGrad()) { - biases_->getParameterPtr()->incUpdate(callback); - } -} - -void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - resetInValue(in); - CHECK(in); - in->downSpatial(); - - auto outPD = - MKLDNNMatrix::createPrimitiveDesc({bs_, oc_}, format::nc, engine_); - resetOutValue(out, outPD); - - format wgtFmt = format::oihw; - if (in->getFormat() == format::nChw8c) { - wgtFmt = format::oIhw8i; - } else if (in->getFormat() == format::nChw16c) { - wgtFmt = format::oIhw16i; - } - auto wgtPD = - MKLDNNMatrix::createPrimitiveDesc({oc_, ic_, ih_, iw_}, wgtFmt, engine_); - resetWithMatrix(wgt, weight_->getW(), wgtPD); - wgt->downSpatial(); - - if (biases_ && biases_->getW()) { - auto biasPD = MKLDNNMatrix::createPrimitiveDesc({oc_}, format::x, engine_); - resetWithMatrix(bias, biases_->getW(), biasPD); - } else { - bias = nullptr; - } -} - -void MKLDNNFcLayer::resetFwdPD(std::shared_ptr& pd, - MKLDNNMatrixPtr in, - MKLDNNMatrixPtr wgt, - MKLDNNMatrixPtr bias, - MKLDNNMatrixPtr out) { - CHECK(in); - CHECK(wgt); - CHECK(out); - prop_kind pk = prop_kind::forward; - fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk, - in->getMemoryDesc(), - wgt->getMemoryDesc(), - bias->getMemoryDesc(), - out->getMemoryDesc()) - : fc_fwd::desc(pk, - in->getMemoryDesc(), - wgt->getMemoryDesc(), - out->getMemoryDesc()); - pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_)); -} - -void MKLDNNFcLayer::resetFwdPipeline( - std::vector& pipeline, - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - if (bias) { - fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out)); - } else { - fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out)); - } - pipeline.push_back(*fwd_); -} - -void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - CHECK(inVals_[0] && outVal_); - resetOutGrad(out, outVal_->getPrimitiveDesc()); - resetInGrad(in, inVals_[0]->getPrimitiveDesc()); - - CHECK(wgtVal_); - resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc()); - - if (biasVal_) { - resetWithMatrix(bias, biases_->getWGrad(), biasVal_->getPrimitiveDesc()); - } else { - bias = nullptr; - } -} - -void MKLDNNFcLayer::resetBwdWgtPD( - std::shared_ptr& pd, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - CHECK(inVals_[0]); - fc_bwdWgt::desc bwdWgtDesc = - bias ? fc_bwdWgt::desc(inVals_[0]->getMemoryDesc(), - wgt->getMemoryDesc(), - bias->getMemoryDesc(), - out->getMemoryDesc()) - : fc_bwdWgt::desc(inVals_[0]->getMemoryDesc(), - wgt->getMemoryDesc(), - out->getMemoryDesc()); - pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_)); -} - -void MKLDNNFcLayer::resetBwdDataPD( - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - pd = nullptr; - if (in == nullptr) { - return; - } - CHECK(wgtVal_); - fc_bwdData::desc bwdDataDesc = fc_bwdData::desc( - in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc()); - pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_)); -} - -void MKLDNNFcLayer::resetBwdPipeline( - std::vector& pipeline, - std::shared_ptr& bwdWgtPD, - std::shared_ptr& bwdDataPD, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, - MKLDNNMatrixPtr& out) { - CHECK(inVals_[0]); - if (bias) { - bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVals_[0], *out, *wgt, *bias)); - } else { - bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVals_[0], *out, *wgt)); - } - pipeline.push_back(*bwdWgt_); - - if (bwdDataPD == nullptr) { - return; - } - CHECK(wgtVal_) << "Should have weight memory"; - bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in)); - pipeline.push_back(*bwdData_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLRNLayer.cpp b/paddle/gserver/layers/MKLDNNLRNLayer.cpp deleted file mode 100644 index 88513ab8bca3899775be7822083b51120a04d6e4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MKLDNNLRNLayer.cpp +++ /dev/null @@ -1,163 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MKLDNNLRNLayer.h" -#include "paddle/utils/Logging.h" - -using namespace mkldnn; // NOLINT -typedef memory::format format; - -namespace paddle { - -REGISTER_LAYER(mkldnn_lrn, MKLDNNLRNLayer); - -bool MKLDNNLRNLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!MKLDNNLayer::init(layerMap, parameterMap)) { - return false; - } - - /* the size of inputs for norm-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - const NormConfig& conf = config_.inputs(0).norm_conf(); - localSize_ = conf.size(); - alpha_ = conf.scale(); - beta_ = conf.pow(); - - ic_ = conf.channels(); - oc_ = ic_; - iw_ = conf.img_size(); - ow_ = conf.output_x(); - ih_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - oh_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); - CHECK_EQ(iw_, ow_); - CHECK_EQ(ih_, oh_); - return true; -} - -void MKLDNNLRNLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { - CHECK_EQ(inputLayers_.size(), 1UL); - reshapeInput(bs, ih, iw); - // ic_ and oc can not be changed - CHECK_EQ((size_t)ic, - inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) - << "Input channel can not be changed"; - oh = ih; - ow = iw; - reshapeOutput(oh, ow); - resizeOutput(bs, oc * oh * ow); -} - -void MKLDNNLRNLayer::resetFwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - resetFwdBuffers(inputs[0], out); - - resetFwdPD(fwdPD_, inputs[0], out); - - resetFwdPipeline(pipeline, fwdPD_, inputs[0], out); -} - -void MKLDNNLRNLayer::resetBwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - std::shared_ptr pd; - - resetBwdBuffers(inputs[0], out); - - resetBwdPD(pd, inputs[0], out); - - resetBwdPipeline(pipeline, pd, inputs[0], out); -} - -void MKLDNNLRNLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - resetInValue(in); - CHECK(in); - resetOutValue(out, in->getPrimitiveDesc()); -} - -void MKLDNNLRNLayer::resetFwdPD(std::shared_ptr& pd, - MKLDNNMatrixPtr in, - MKLDNNMatrixPtr out) { - prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring - : prop_kind::forward_training; - auto fwdDesc = lrn_fwd::desc(pk, - algorithm::lrn_across_channels, - in->getMemoryDesc(), - localSize_, - alpha_, - beta_, - 1.0f); - pd.reset(new lrn_fwd::primitive_desc(fwdDesc, engine_)); - // prepare workspace if necessary - workspace_ = - passType_ != PASS_TEST - ? std::make_shared(memory(pd->workspace_primitive_desc())) - : nullptr; -} - -void MKLDNNLRNLayer::resetFwdPipeline( - std::vector& pipeline, - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - fwd_ = workspace_ - ? std::make_shared(lrn_fwd(*pd, *in, *workspace_, *out)) - : std::make_shared(lrn_fwd(*pd, *in, *out)); - pipeline.push_back(*fwd_); -} - -void MKLDNNLRNLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - CHECK(inVals_[0] && outVal_); - resetOutGrad(out, outVal_->getPrimitiveDesc()); - resetInGrad(in, inVals_[0]->getPrimitiveDesc()); -} - -void MKLDNNLRNLayer::resetBwdPD(std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - pd = nullptr; - if (in == nullptr) { - return; - } - CHECK(out); - auto bwdDesc = lrn_bwd::desc(algorithm::lrn_across_channels, - in->getMemoryDesc(), - out->getMemoryDesc(), - localSize_, - alpha_, - beta_, - 1.0f); - pd.reset(new lrn_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); -} - -void MKLDNNLRNLayer::resetBwdPipeline( - std::vector& pipeline, - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - if (pd == nullptr) { - return; - } - CHECK(inVals_[0]); - CHECK(workspace_); - bwdData_ = std::make_shared( - lrn_bwd(*pd, *inVals_[0], *out, *workspace_, *in)); - pipeline.push_back(*bwdData_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h deleted file mode 100644 index 2b164d0d3bc0e1446d7e4d82bb8a713195dbd927..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MKLDNNLayer.h +++ /dev/null @@ -1,477 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Layer.h" -#include "MKLDNNBase.h" -#include "mkldnn.hpp" -#include "paddle/math/MKLDNNMatrix.h" -#include "paddle/utils/Stat.h" - -DECLARE_bool(use_mkldnn); - -namespace paddle { - -class MKLDNNLayer; -typedef std::shared_ptr MKLDNNLayerPtr; - -/** - * @brief Base class of MKLDNNlayer. - * - */ -class MKLDNNLayer : public Layer { - protected: - // batch size - int bs_; - // their sizes are always from the first input layer - // input image channel, height and width - int ic_, ih_, iw_; - // output image channel, height and width - int oc_, oh_, ow_; - - // the condition that forward need be reset - size_t condition_; - // backward also need reset after reset forward handle - bool needResetBwd_; - - // is output only mkldnn - bool outputOnlyMKLDNN_; - - // mkldnn engine, stream and primivtives - mkldnn::engine engine_; - std::shared_ptr stream_; - std::shared_ptr fwd_; - std::shared_ptr bwdWgt_; - std::shared_ptr bwdData_; - std::vector pipelineFwd_; - std::vector pipelineBwd_; - - /* Value and grad are seperated as internal and external buffers. - * Each MKLDNNLayer must init or reset internal buffer at least, - * and the external buffer format is always nchw of nc(when h==w==1), - * which is the same format as paddle. - * The output_.value and output_.grad always save the external data, - * when mixed with cpu device. - * When all layers are mkldnn layers, they could save internal data. - */ - // below MKLDNNMatrix buffers are all internal buffers - std::vector inVals_; - std::vector inGrads_; - MKLDNNMatrixPtr outVal_; - MKLDNNMatrixPtr outGrad_; - // below are external value and grad - std::vector extInVals_; - std::vector extInGrads_; - MKLDNNMatrixPtr extOutVal_; - MKLDNNMatrixPtr extOutGrad_; - // convert handle between external and internal buffers - std::vector> cvtInVals_; - std::vector> cvtInGrads_; - std::shared_ptr cvtOutVal_; - std::shared_ptr cvtOutGrad_; - - // weight and bias are always internal buffers - MKLDNNMatrixPtr wgtVal_; - MKLDNNMatrixPtr wgtGrad_; - MKLDNNMatrixPtr biasVal_; - MKLDNNMatrixPtr biasGrad_; - - // merge grad primitive - std::shared_ptr mergeGrad_; - std::vector pipelineMergeGrad_; - // tmp input argument to save input grad, only used to merge grad - Argument tmpInArg_; - - public: - explicit MKLDNNLayer(const LayerConfig& config) - : Layer(config), - ih_(0), - iw_(0), - condition_(0), - needResetBwd_(true), - outputOnlyMKLDNN_(false), - engine_(mkldnn::engine::cpu, 0), - stream_(nullptr), - fwd_(nullptr), - bwdWgt_(nullptr), - bwdData_(nullptr) {} - - ~MKLDNNLayer() {} - - virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); - virtual void forward(PassType passType); - virtual void backward(const UpdateCallback& callback); - - /** - * reshape the input and output channels and image sizes - * and reset output buffer size - */ - virtual void reshape( - int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) = 0; - - /** - * reset the mkldnn forward primitve and memories - * only would be called when input size changes - * weight and bias buffers should be coverd by child class itself - */ - virtual void resetFwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) = 0; - - /** - * reset the mkldnn backward primitve and memories - * only would be called when needed - * weight and bias buffers should be coverd by child class itself - */ - virtual void resetBwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) = 0; - - /** - * Update weights and biases if necessary. - */ - virtual void updateWeights(const UpdateCallback& callback) {} - - /** - * convert weight from paddle format to mkldnn format - * weight_ will be override - */ - virtual void convertWeightsFromPaddle() {} - - /** - * convert mkldnn weight to paddle format - * weight_ will be override - */ - virtual void convertWeightsToPaddle() {} - - /** - * add this interface as public for unit test - */ - void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); } - - protected: - /** - * Some layers may have different condition to reset the forward. - * The function returns the condition that do not need reset forward. - */ - inline virtual size_t keepCondition() { - // reset when the first input element size changed, not only the batchsize - return inputLayers_[0]->getOutputValue()->getElementCnt(); - } - - /** - * reshape the input image sizes and input batchsize - */ - void reshapeInput(int& batchsize, int& height, int& width, size_t idx = 0); - - /** - * reshape output image sizes - */ - void reshapeOutput(size_t height, size_t width); - - /** - * reset MKLDNNMatrix from Matrix and internal primitive desc. - * reset nullptr if matrix or primitive desc is empty - */ - void resetWithMatrix(MKLDNNMatrixPtr& dnn, - const MatrixPtr& mat, - mkldnn::memory::primitive_desc pd); - - /** - * reset input value from input MKLDNNMatrix and internal primitive desc. - * reset both internal and external buffer and create reorder if necessary. - * input channel may be different in concat. - */ - void resetInValue( - MKLDNNMatrixPtr& in, - const std::shared_ptr& intPD = nullptr, - size_t idx = 0, - int inputChannel = 0); - - /** - * reset output value from internal primitive desc. - * reset both internal and external buffer and create reorder if necessary. - */ - void resetOutValue(MKLDNNMatrixPtr& out, - mkldnn::memory::primitive_desc intPD); - - /** - * reset input grad from internal primitive desc. - * reset both internal and external buffer and create reorder if necessary. - */ - void resetInGrad(MKLDNNMatrixPtr& in, - mkldnn::memory::primitive_desc intPD, - size_t idx = 0); - - /** - * reset output grad from internal primitive desc. - * merge grad if necessary. - * reset both internal and external buffer and create reorder if necessary. - * note: about merge grad, when this layer has several outputs, - * it could not be mixed with cpu device, - * since it can not get memory desc from cpu device. - */ - void resetOutGrad(MKLDNNMatrixPtr& out, mkldnn::memory::primitive_desc intPD); - - /** - * reset the merge grad primitive if necessary. - * note: do not support the grads mixed with cpu device, - * since it can not get memory desc from cpu device. - */ - void resetMergeGrad(MKLDNNMatrixPtr& out); - - protected: - /** - * Set deviceId of this layer. - */ - void setDevice(int id) { deviceId_ = id; } - - /** - * check the format is nchw or nc, - * which is supported by Paddle default memory layout - */ - bool isPaddleFormat(mkldnn::memory::format fmt) { - if (fmt == mkldnn::memory::format::nchw || - fmt == mkldnn::memory::format::nc) { - return true; - } else { - return false; - } - } - - /** - * If input only has MKLDNN device. - * Otherwise, only support the previous layer using CPU device. - */ - bool inputIsOnlyMKLDNN(int index = 0) { - int prevDevice = getPrev(index)->getDeviceId(); - if (prevDevice == MKLDNN_DEVICE) { - return true; - } else { - CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet"; - return false; - } - } - - /** - * If output only has MKLDNN device. - * Otherwise, other devices should only using CPU device. - */ - bool outputIsOnlyMKLDNN() { - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE) - << "Only support other device is CPU yet"; - } - outputOnlyMKLDNN_ = outputOtherDevice_.size() == 0; - return outputOnlyMKLDNN_; - } - - /** - * print info about sizes - */ - virtual void printSizeInfo() { - VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_ - << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_ - << ", oh: " << oh_ << ", ow: " << ow_; - } - - /** - * print the mkldnn memory format of value - */ - virtual void printValueFormat() { - for (size_t i = 0; i < inVals_.size(); ++i) { - if (!inVals_[i]) { - continue; - } - VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() - << ": " << (extInVals_[i] ? extInVals_[i]->getFormat() - : inVals_[i]->getFormat()) - << " >>> " << inVals_[i]->getFormat() << " >>>"; - } - if (outVal_) { - VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> " - << (extOutVal_ ? extOutVal_->getFormat() - : outVal_->getFormat()); - } - if (wgtVal_) { - VLOG(MKLDNN_FMTS) << "Weight value format: " << wgtVal_->getFormat(); - } - if (biasVal_) { - VLOG(MKLDNN_FMTS) << "Bias value format: " << biasVal_->getFormat(); - } - } - - /** - * print the mkldnn memory format of grad - */ - virtual void printGradFormat() { - if (outGrad_) { - VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< " - << (extOutGrad_ ? extOutGrad_->getFormat() - : outGrad_->getFormat()); - } - for (size_t i = 0; i < inGrads_.size(); ++i) { - if (!inGrads_[i]) { - continue; - } - VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() - << ": " << (extInGrads_[i] ? extInGrads_[i]->getFormat() - : inGrads_[i]->getFormat()) - << " <<< " << inGrads_[i]->getFormat() << " <<<"; - } - if (wgtGrad_) { - VLOG(MKLDNN_FMTS) << "Weight grad format: " << wgtGrad_->getFormat(); - } - if (biasGrad_) { - VLOG(MKLDNN_FMTS) << "Bias grad format: " << biasGrad_->getFormat(); - } - } - - private: - /** - * clear all grad - */ - void clearGrads() { - if (output_.grad) { - output_.grad->zeroMem(); - } - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - if (outputOtherDevice_[i].grad) { - outputOtherDevice_[i].grad->zeroMem(); - } - } - } - - /** - * Set deviceId of the params used in this layer. - */ - void setParamsDevice(int id, const ParameterMap& parameterMap) { - for (auto& inputConfig : config_.inputs()) { - if (inputConfig.has_input_parameter_name()) { - ParameterPtr parameter; - std::string name = inputConfig.input_parameter_name(); - CHECK(mapGet(name, parameterMap, ¶meter)) - << "Cannot find input parameter " << name << " for layer " - << getName(); - parameter->setDevice(id); - } - } - if (config_.has_bias_parameter_name()) { - ParameterPtr parameter; - std::string name = config_.bias_parameter_name(); - CHECK(mapGet(name, parameterMap, ¶meter)) - << "Cannot find bias parameter " << name << " for layer " - << getName(); - parameter->setDevice(id); - } - } - - /** - * Set output map of prev layers. - */ - void setOutputMap() { - outputMap_.clear(); - for (size_t i = 0; i < inputLayers_.size(); ++i) { - inputLayers_[i]->setOutput(getName(), &tmpInArg_); - } - } - - /** - * if have cpu device, share value and grad data with output_ - */ - void shareCPUDevice() { - if (outputIsOnlyMKLDNN()) { - return; - } - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - outputOtherDevice_[i].value = output_.value; - outputOtherDevice_[i].grad = output_.grad; - } - } - - /** - * Check the cpu device number of outputOtherDevice_. - * should have only one at most. - */ - void checkCPUOutputsNumber(int max = 1) { - int cnt = 0; - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - if (outputOtherDevice_[i].deviceId == CPU_DEVICE) { - ++cnt; - } - } - CHECK_LE(cnt, max) << "too much CPU devies"; - } - - /** - * copy SeqInfo from input layer to this output and other output devices. - * @note: do not use getInput(0) since it used this deviceId_, - * use "inputLayers_[0]->getOutput()" instead. - */ - void copySeqInfoToOutputs() { - if (inputLayers_.empty() || !needSequenceInfo_) { - return; - } - const Argument& input = inputLayers_[0]->getOutput(); - output_.sequenceStartPositions = input.sequenceStartPositions; - output_.subSequenceStartPositions = input.subSequenceStartPositions; - output_.cpuSequenceDims = input.cpuSequenceDims; - for (size_t i = 0; i < outputOtherDevice_.size(); i++) { - outputOtherDevice_[i].sequenceStartPositions = - output_.sequenceStartPositions; - outputOtherDevice_[i].subSequenceStartPositions = - output_.subSequenceStartPositions; - outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; - } - } - - void prepareValueConversions(std::vector& pipeline) { - // MKLDNNLayer output value should be MKLDNNMatrix - // so external output value is necessary. - // Then external input value is not necessary, - // since input may be mkldnn internal buffer. - CHECK(extOutVal_) << "external output value is necessary"; - output_.value = std::dynamic_pointer_cast(extOutVal_); - CHECK(inVals_[0] && outVal_) << "internal memories are necessary"; - for (size_t i = 0; i < cvtInVals_.size(); ++i) { - if (cvtInVals_[i]) { - pipeline.insert(pipeline.begin(), *cvtInVals_[i]); - } - } - if (cvtOutVal_) { - pipeline.push_back(*cvtOutVal_); - } - } - void prepareGradConversions(std::vector& pipeline) { - // external output grad is not necessary - // since output may be mkldnn internal buffer or merge them directly. - CHECK(outGrad_) << "internal output grad is necessary"; - if (extOutGrad_) { - CHECK_EQ(extOutGrad_->getData(), output_.grad->getData()) - << "the external buffer should share the same data with output_.grad"; - } - if (cvtOutGrad_) { - pipeline.insert(pipeline.begin(), *cvtOutGrad_); - } - for (size_t i = 0; i < cvtInGrads_.size(); ++i) { - if (cvtInGrads_[i]) { - pipeline.push_back(*cvtInGrads_[i]); - } - } - } -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/gserver/layers/MKLDNNPoolLayer.cpp deleted file mode 100644 index 3be848c7496aac616903cb09844c5eadd320e91c..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MKLDNNPoolLayer.cpp +++ /dev/null @@ -1,195 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MKLDNNPoolLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/Logging.h" - -using namespace mkldnn; // NOLINT -typedef memory::format format; - -namespace paddle { - -REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer); - -bool MKLDNNPoolLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!MKLDNNLayer::init(layerMap, parameterMap)) { - return false; - } - - /* the size of inputs for pool-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - const PoolConfig& conf = config_.inputs(0).pool_conf(); - ic_ = conf.channels(); - ih_ = conf.img_size_y(); - iw_ = conf.img_size(); - oc_ = ic_; - oh_ = conf.output_y(); - ow_ = conf.output_x(); - fh_ = conf.size_y(); - fw_ = conf.size_x(); - ph_ = conf.padding_y(); - pw_ = conf.padding(); - sh_ = conf.stride_y(); - sw_ = conf.stride(); - - const std::string& type = conf.pool_type(); - if (type == "max-projection") { - poolAlgo_ = algorithm::pooling_max; - } else if (type == "avg-projection") { - // paddle only use exclude_padding - poolAlgo_ = algorithm::pooling_avg_exclude_padding; - } else { - LOG(FATAL) << "unknow pooling type!"; - } - return true; -} - -void MKLDNNPoolLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { - reshapeInput(bs, ih, iw); - // ic_ and oc can not be changed - CHECK_EQ((size_t)ic, - inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) - << "Input channel can not be changed"; - - // cal output sizes - // paddle used false caffeMode for pooling - oh = outputSize(ih, fh_, ph_, sh_, false); - ow = outputSize(iw, fw_, pw_, sw_, false); - reshapeOutput(oh, ow); - - resizeOutput(bs, oc * oh * ow); -} - -void MKLDNNPoolLayer::resetFwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - resetFwdBuffers(inputs[0], out); - - resetFwdPD(fwdPD_, inputs[0], out); - - resetFwdPipeline(pipeline, fwdPD_, inputs[0], out); -} - -void MKLDNNPoolLayer::resetBwd(std::vector& pipeline, - std::vector& inputs, - MKLDNNMatrixPtr& out) { - std::shared_ptr pd; - - resetBwdBuffers(inputs[0], out); - - resetBwdPD(pd, inputs[0], out); - - resetBwdPipeline(pipeline, pd, inputs[0], out); -} - -void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - resetInValue(in); - - memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_}; - CHECK(in); - auto outPD = - MKLDNNMatrix::createPrimitiveDesc(outDims, in->getFormat(), engine_); - resetOutValue(out, outPD); -} - -void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr& pd, - MKLDNNMatrixPtr in, - MKLDNNMatrixPtr out) { - memory::dims kernels = memory::dims{fh_, fw_}; - memory::dims strides = memory::dims{sh_, sw_}; - memory::dims padL = memory::dims{ph_, pw_}; - memory::dims padR = getPaddingR(); - padding_kind padKind = padding_kind::zero; - prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring - : prop_kind::forward_training; - auto fwdDesc = pool_fwd::desc(pk, - poolAlgo_, - in->getMemoryDesc(), - out->getMemoryDesc(), - strides, - kernels, - padL, - padR, - padKind); - pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_)); - - // prepare workspace if necessary - workspace_ = - (passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max) - ? std::make_shared(memory(pd->workspace_primitive_desc())) - : nullptr; -} - -void MKLDNNPoolLayer::resetFwdPipeline( - std::vector& pipeline, - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - fwd_ = workspace_ - ? std::make_shared(pool_fwd(*pd, *in, *out, *workspace_)) - : std::make_shared(pool_fwd(*pd, *in, *out)); - pipeline.push_back(*fwd_); -} - -void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - CHECK(inVals_[0] && outVal_); - resetOutGrad(out, outVal_->getPrimitiveDesc()); - resetInGrad(in, inVals_[0]->getPrimitiveDesc()); -} - -void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - pd = nullptr; - if (in == nullptr) { - return; - } - memory::dims kernels = memory::dims{fh_, fw_}; - memory::dims strides = memory::dims{sh_, sw_}; - memory::dims padL = memory::dims{ph_, pw_}; - memory::dims padR = getPaddingR(); - CHECK(out); - auto bwdDesc = pool_bwd::desc(poolAlgo_, - in->getMemoryDesc(), - out->getMemoryDesc(), - strides, - kernels, - padL, - padR, - padding_kind::zero); - pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); -} - -void MKLDNNPoolLayer::resetBwdPipeline( - std::vector& pipeline, - std::shared_ptr& pd, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& out) { - if (pd == nullptr) { - return; - } - - bwdData_ = - workspace_ - ? std::make_shared(pool_bwd(*pd, *out, *workspace_, *in)) - : std::make_shared(pool_bwd(*pd, *out, *in)); - pipeline.push_back(*bwdData_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MKLPackedWeight.h b/paddle/gserver/layers/MKLPackedWeight.h deleted file mode 100644 index b01a961d007a0e2e343db7b51e50fd3ee776435e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MKLPackedWeight.h +++ /dev/null @@ -1,86 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/math/MathFunctions.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/parameter/Weight.h" - -namespace paddle { - -class MKLPackedWeight { - protected: - /// The pointer of weight - real *weight_; - /// The pointer of cblas packed gemm to weight - real *packedWeight_; - size_t height_; - size_t width_; - bool transW_; - - public: - explicit MKLPackedWeight(MatrixPtr weight, bool transW = false) { - packedWeight_ = nullptr; - weight_ = weight->getData(); - height_ = weight->getHeight(); - width_ = weight->getWidth(); - transW_ = transW; - } - - ~MKLPackedWeight() { free_(); } - - void pack() { pack_(weight_); } - - void gemm_compute(const MatrixPtr src, MatrixPtr dst) { - cblas_sgemm_compute(CblasRowMajor, - CblasNoTrans, - CblasPacked, - src->getHeight(), - transW_ ? height_ : width_, - transW_ ? width_ : height_, - src->getData(), - src->getWidth(), - packedWeight_, - width_, - 1.0, - dst->getData(), - dst->getWidth()); - } - - protected: - void pack_(real *src) { - if (!packedWeight_) { - packedWeight_ = cblas_sgemm_alloc(CblasBMatrix, 1, width_, height_); - } - cblas_sgemm_pack(CblasRowMajor, - CblasBMatrix, - transW_ ? CblasTrans : CblasNoTrans, - 1, - transW_ ? height_ : width_, - transW_ ? width_ : height_, - 1.0, - src, - width_, - packedWeight_); - } - - void free_() { - if (packedWeight_) { - cblas_sgemm_free(packedWeight_); - } - } -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/MaxLayer.cpp b/paddle/gserver/layers/MaxLayer.cpp deleted file mode 100644 index 7ee2e0dd946d6f332f6b8454f977601b0ee8d249..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MaxLayer.cpp +++ /dev/null @@ -1,65 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MaxLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(max, MaxLayer); - -void MaxLayer::forward(PassType passType) { - SequencePoolLayer::forward(passType); - - IVector::resizeOrCreate( - maxIndex_, newBatchSize_ * getSize(), useGpu(deviceId_)); - maxIndex_->zeroMem(); - - MatrixPtr inputValue = getInputValue(0); - MatrixPtr outputValue = getOutputValue(); - - { - REGISTER_TIMER_INFO("MaxLayerForward", getName().c_str()); - outputValue->maxSequenceForward( - *inputValue, *startPositions_->getVector(useGpu_), *maxIndex_); - } - - if (config_.output_max_index()) { - // copy maxIndex_ to output - outputValue->copyFrom(*maxIndex_); - } else { - /* add the bias-vector AFTER max operation */ - if (biases_.get() != NULL) { - outputValue->addBias(*(biases_->getW()), 1); - } - /* activation */ { forwardActivation(); } - } -} - -void MaxLayer::backward(const UpdateCallback& callback) { - CHECK(!config_.output_max_index()) - << "backward is not available when output_max_index is set"; - SequencePoolLayer::backward(callback); - - MatrixPtr inputGrad = getInputGrad(0); - MatrixPtr outputGrad = getOutputGrad(); - if (inputGrad) { - REGISTER_TIMER_INFO("MaxLayerBackward", getName().c_str()); - inputGrad->maxSequenceBackward( - *outputGrad, *(startPositions_->getVector(useGpu_)), *maxIndex_); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MaxLayer.h b/paddle/gserver/layers/MaxLayer.h deleted file mode 100644 index e46f997c342ce5d6b724629dff6950c4f1680ce8..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MaxLayer.h +++ /dev/null @@ -1,58 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "SequencePoolLayer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * A layer for "internal max" for sequence input. - * Input: one or more sequences. Each sequence contains some instances. - * If SequenceLevel = kNonSeq: - * Output: output size is the number of input sequences (NOT input instances) - * output[i] = max_{for each instance in this sequence}{input[i]} - * If stride_ > 0: - * Output: a shorten sequence. Stride is the step size by which we slide a - * window upon the input sequence, and the max pooling operation is - * then applied to each interval independently. - * If SequenceLevel = kSeq: - * Check input sequence must has sub-sequence - * Output: output size is the number of input sub-sequences - * output[i] = max_{for each instance in this sub-sequence}{input[i]} - * - * The config file api is pooling_layer. - */ - -class MaxLayer : public SequencePoolLayer { - protected: - // maxIndex_[i][j] = k : the value at (i, j) is from input[k]. - IVectorPtr maxIndex_; - - public: - explicit MaxLayer(const LayerConfig& config) : SequencePoolLayer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override { - return SequencePoolLayer::init(layerMap, parameterMap); - } - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/MaxOutLayer.h b/paddle/gserver/layers/MaxOutLayer.h deleted file mode 100644 index 0eb8674b4c4f3f58b103c6b59ad13931a6992a1b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MaxOutLayer.h +++ /dev/null @@ -1,55 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * A layer to do max out on conv layer output. - * Input: output of a conv layer. - * Output: feature map size same as input. Channel is (input channel) / groups. - * So the num of channels should be able to devided by groups. - * - * The config file api is maxout_layer. - */ - -class MaxOutLayer : public Layer { - protected: - size_t groups_; - size_t imgSizeH_, imgSizeW_; - /// outputChannels_ = channels_ / groups_ - size_t channels_, outputChannels_; - /// feature length = imgSizeH_ * imgSizeW_ - size_t featLen_; - IVectorPtr maxoutId_; - - public: - /// return imgSizeH_ * imgSizeW_ * outputChannels_; - size_t getSize(); - - explicit MaxOutLayer(const LayerConfig& config) : Layer(config) {} - virtual ~MaxOutLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/MaxPoolWithMaskLayer.cpp b/paddle/gserver/layers/MaxPoolWithMaskLayer.cpp deleted file mode 100644 index e594e22b5eaa6027fdf5bbd09ab93774d9a798be..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MaxPoolWithMaskLayer.cpp +++ /dev/null @@ -1,109 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MaxPoolWithMaskLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -bool MaxPoolWithMaskLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - PoolLayer::init(layerMap, parameterMap); - setOutput("mask", &mask_); - return true; -} - -size_t MaxPoolWithMaskLayer::getSize() { - CHECK_EQ(inputLayers_.size(), 1UL); - size_t layerSize = 0; - - outputY_ = outputSize(imgSizeY_, - sizeY_, - confPaddingY_, - strideY_, - /* caffeMode */ false); - outputX_ = outputSize(imgSize_, - sizeX_, - confPadding_, - stride_, - /* caffeMode */ false); - - layerSize = outputX_ * outputY_ * channels_; - getOutput().setFrameHeight(outputY_); - getOutput().setFrameWidth(outputX_); - - return layerSize; -} - -void MaxPoolWithMaskLayer::forward(PassType passType) { - size_t size = getSize(); - MatrixPtr inputV = inputLayers_[0]->getOutputValue(); - int batchSize = inputV->getHeight(); - resetOutput(batchSize, size); - - MatrixPtr outV = getOutputValue(); - CHECK_EQ(size, outV->getWidth()); - - resetSpecifyOutput(mask_, - batchSize, - size, - /* isValueClean */ false, - /* isGradClean */ true); - - MatrixPtr maskV = mask_.value; - outV->maxPoolForward(*inputV, - imgSizeY_, - imgSize_, - channels_, - sizeX_, - sizeY_, - strideY_, - stride_, - outputY_, - outputX_, - confPaddingY_, - confPadding_, - maskV); -} - -void MaxPoolWithMaskLayer::backward(const UpdateCallback& callback) { - (void)callback; - if (NULL == getInputGrad(0)) { - return; - } - - MatrixPtr outGrad = getOutputGrad(); - MatrixPtr inputV = inputLayers_[0]->getOutputValue(); - MatrixPtr outV = getOutputValue(); - MatrixPtr inputGrad = inputLayers_[0]->getOutputGrad(); - - inputGrad->maxPoolBackward(*inputV, - imgSizeY_, - imgSize_, - *outGrad, - *outV, - sizeX_, - sizeY_, - strideY_, - stride_, - outputY_, - outputX_, - 1, - 1, - confPaddingY_, - confPadding_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MaxPoolWithMaskLayer.h b/paddle/gserver/layers/MaxPoolWithMaskLayer.h deleted file mode 100644 index c948364f6b83b0de1ee07cc185b69346f5cb1a7e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MaxPoolWithMaskLayer.h +++ /dev/null @@ -1,40 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "PoolLayer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * @brief Basic parent layer of different kinds of pooling - */ -class MaxPoolWithMaskLayer : public PoolLayer { - protected: - Argument mask_; - - public: - explicit MaxPoolWithMaskLayer(const LayerConfig& config) - : PoolLayer(config) {} - - size_t getSize(); - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/MixedLayer.cpp b/paddle/gserver/layers/MixedLayer.cpp deleted file mode 100644 index 7dcb30b98d6e6b08929d5fecba0833c8b1989725..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MixedLayer.cpp +++ /dev/null @@ -1,176 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MixedLayer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(mixed, MixedLayer); - -bool MixedLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - if (!Layer::init(layerMap, parameterMap)) return false; - - CHECK_EQ(inputLayers_.size(), parameters_.size()); - projections_.resize(inputLayers_.size()); - for (size_t i = 0; i < inputLayers_.size(); i++) { - if (config_.inputs(i).has_proj_conf()) { - projections_[i].reset(Projection::create( - config_.inputs(i).proj_conf(), parameters_[i], useGpu_)); - } else { - CHECK(!parameters_[i]) << "should no parameters for operators"; - } - } - for (auto& operator_conf : config_.operator_confs()) { - for (auto& input_index : operator_conf.input_indices()) { - CHECK(!config_.inputs(input_index).has_proj_conf()); - } - operators_.emplace_back(Operator::create(operator_conf, useGpu_)); - } - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - sharedBias_ = config_.shared_biases(); - size_t psize = config_.bias_size(); - biases_ = std::unique_ptr(new Weight(1, psize, biasParameter_)); - } - - return true; -} - -void MixedLayer::prefetch() { - for (size_t i = 0; i != inputLayers_.size(); ++i) { - if (projections_[i]) { - projections_[i]->prefetch(&getInput(i)); - } - } -} - -void MixedLayer::resetState() { - for (auto& proj : projections_) { - if (proj) { - proj->resetState(); - } - } -} - -void MixedLayer::setState(LayerStatePtr state) { - CHECK(projectionStateMatrixSize_.size() == projections_.size()) - << "projection size mis-match"; - - int start = 0; - LayerStatePtr statePtr = std::make_shared(); - for (int i = 0; i < (int)projectionStateMatrixSize_.size(); i++) { - if (projectionStateMatrixSize_[i] > 0) { - statePtr->value.clear(); - for (int j = start; j < start + projectionStateMatrixSize_[i]; j++) { - statePtr->value.push_back(state->value[j]); - } - projections_[i]->setState(statePtr); - start += projectionStateMatrixSize_[i]; - } - } - CHECK((int)state->value.size() == start) << "state matrix size mis-match"; -} - -// Return state which consists of all projections states -LayerStatePtr MixedLayer::getState() { - bool init = projectionStateMatrixSize_.size() == 0; - LayerStatePtr res = std::make_shared(); - for (int i = 0; i < (int)projections_.size(); i++) { - LayerStatePtr statePtr = - projections_[i] ? projections_[i]->getState() : nullptr; - int stateSize = statePtr == nullptr ? 0 : statePtr->value.size(); - if (init) { - projectionStateMatrixSize_.push_back(stateSize); - } else { - CHECK(projectionStateMatrixSize_[i] == stateSize) - << "state matrix size mis-match"; - } - if (statePtr != nullptr) { - for (auto& matrixPtr : statePtr->value) { - res->value.push_back(matrixPtr); - } - } - } - return res; -} - -void MixedLayer::forward(PassType passType) { - Layer::forward(passType); - - int batchSize = getInput(0).getBatchSize(); - int size = getSize(); - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - resetOutput(batchSize, size); - } - - MatrixPtr outV = getOutputValue(); - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - if (projections_[i]) { - projections_[i]->forward(&getInput(i), &output_, passType); - } - } - - std::vector ins; - for (auto& op : operators_) { - ins.clear(); - for (auto& input_index : op->getConfig().input_indices()) { - ins.push_back(&getInput(input_index)); - } - op->forward(ins, &output_, passType); - } - - /* add the bias-vector */ - if (biases_.get() != NULL) { - REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); - outV->addBias(*(biases_->getW()), 1, sharedBias_); - } - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void MixedLayer::backward(const UpdateCallback& callback) { - /* Do activation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - if (biases_ && biases_->getWGrad()) { - REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); - biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBias_); - - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - for (size_t i = 0; i != inputLayers_.size(); ++i) { - if (projections_[i]) { - projections_[i]->backward(callback); - } - } - - for (auto& op : operators_) { - op->backward(); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/MultinomialSampler.h b/paddle/gserver/layers/MultinomialSampler.h deleted file mode 100644 index 8cbb229f157c0904e63a696f860ec6739d5167c4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MultinomialSampler.h +++ /dev/null @@ -1,81 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "paddle/utils/Common.h" - -namespace paddle { - -/** - * @brief Given the probability of N objects, the sampler random select - * one of the object. - * @note: prob does not have to be unnormalized. - * - * The space requirement is O(N)=O(N * sizeof(Interval)). - * The computational complexity of generate one sample is O(1). - */ -class MultinomialSampler { - public: - MultinomialSampler(const real* prob, int size); - - //! protobuf always using double. - static MultinomialSampler* create(const double* prob, int size) { -#ifdef PADDLE_TYPE_DOUBLE - return new MultinomialSampler(prob, size); -#else - std::unique_ptr tmp(new real[size]); - std::copy(prob, prob + size, tmp.get()); - return new MultinomialSampler(tmp.get(), size); -#endif - } - - /** - * @brief Generate a random sample. - * @param g is a random number engine. See . - * @return Random integer. - */ - template - int gen(URNG& g) { - return gen1([&g, this]() { return rand_(g); }); - } - - protected: - /** - * @brief Generation - * @param[in] rand rand is a real random number distribution - * for the range [0, size). - * @return random int number or intervals_[random_int_number].otherId. - */ - template - int gen1(Rand rand) { - double r = rand(); // NOLINT - int i = (int)r; - r -= i; - return r < intervals_[i].thresh ? i : intervals_[i].otherId; - } - - struct Interval { - int otherId; - real thresh; - }; - - /// The probability of each interval will be 1./size - std::vector intervals_; - std::uniform_real_distribution rand_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/MultiplexLayer.cpp b/paddle/gserver/layers/MultiplexLayer.cpp deleted file mode 100644 index 43ecc48cd97fb54d8dc4eb1d87ebf60f5aa040d8..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/MultiplexLayer.cpp +++ /dev/null @@ -1,180 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - *@brief This layer multiplex multiple layers according to the index, - * which is provided by the first input layer. - * - Input[0]: the index of the layer to output of size batchSize. - * - Input[1:N]; the candidate output data. - * For each index i from 0 to batchSize -1, the output is the i-th row of the - * (index[i] + 1)-th layer. - * - * For each i-th row of output: - * - * \f[ - * y[i][j] = x_{x_{0}[i] + 1}[i][j], j = 0,1, ... , (x_{1}.width - 1) - * \f] - * where, y is output. \f$x_{k}\f$ is the k-th input layer and - * \f$k = x_{0}[i] + 1\f$. - */ - -class MultiplexLayer : public Layer { - protected: - /** - * @brief A struct is used to save the copy information, includes input - * layer index and copy size. - */ - struct CopyInfo { - CopyInfo(int inStartIdx, int inLength, int inCopyIdx) - : startIdx(inStartIdx), length(inLength), copyIdx(inCopyIdx) {} - - /// The start row of input. - int startIdx; - /// Number of rows. If the layer index in Input[0] is not consecutive, - /// the length is one. Otherwise, the length is > 1 and copy multi rows - /// once. - int length; - /// The copied layer index, which needs to add 1. - int copyIdx; - }; - - /// A list of CopyInfo used to save copy information. - std::vector copySchedule_; - - /// Temporary matrix pointer to point to input data. - MatrixPtr tmpSrc_; - /// Temporary matrix pointer to point to output data. - MatrixPtr tmpDest_; - - public: - explicit MultiplexLayer(const LayerConfig& config) : Layer(config) {} - - ~MultiplexLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - private: - /** - * @brief Calculate copy info for input layers. - */ - void calculateCopySchedule(const IVectorPtr& copyIds, size_t numIns); -}; - -REGISTER_LAYER(multiplex, MultiplexLayer); - -void MultiplexLayer::calculateCopySchedule(const IVectorPtr& copyIds, - size_t numIns) { - copySchedule_.clear(); - CopyInfo prevCopyInfo(0, 0, -1); - for (size_t i = 0; i < copyIds->getSize(); i++) { - int copyId = copyIds->getElement(i); - CHECK_GE(copyId, 0); - CHECK_LT(copyId, int(numIns)); - // copy same input layer with prevous and will copy consecutive. - if (copyId == prevCopyInfo.copyIdx) { - ++prevCopyInfo.length; - } else { - if (prevCopyInfo.copyIdx != -1) { - copySchedule_.emplace_back(prevCopyInfo); - } - prevCopyInfo.startIdx = i; - prevCopyInfo.length = 1; - prevCopyInfo.copyIdx = copyId; - } - } - if (prevCopyInfo.copyIdx != -1) { - copySchedule_.emplace_back(prevCopyInfo); - } -} - -bool MultiplexLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_GE(inputLayers_.size(), 2U); - - tmpSrc_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); - tmpDest_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); - return true; -} - -void MultiplexLayer::forward(PassType passType) { - Layer::forward(passType); - - IVectorPtr copyIds = getInput(0).ids; - MatrixPtr inV1 = getInputValue(1); - CHECK_EQ(copyIds->getSize(), inV1->getHeight()); - for (size_t i = 2; i < inputLayers_.size(); i++) { - CHECK_EQ(inV1->getHeight(), getInputValue(i)->getHeight()); - CHECK_EQ(inV1->getWidth(), getInputValue(i)->getWidth()); - } - - calculateCopySchedule(copyIds, inputLayers_.size() - 1); - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(inV1->getHeight(), inV1->getWidth()); - } - - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwLMultplexingTimer", getName().c_str()); - AsyncGpuBlock block; - for (const CopyInfo& info : copySchedule_) { - outV->subMatrix(info.startIdx, info.length, tmpDest_) - ->copyFrom(*getInputValue(info.copyIdx + 1) - ->subMatrix(info.startIdx, info.length, tmpSrc_)); - } - } - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void MultiplexLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - MatrixPtr outG = getOutputGrad(); - - { - REGISTER_TIMER_INFO("BwLMultiplexTimer", getName().c_str()); - AsyncGpuBlock block; - for (const CopyInfo& info : copySchedule_) { - if (getInputGrad(info.copyIdx + 1)) { - getInputGrad(info.copyIdx + 1) - ->subMatrix(info.startIdx, info.length, tmpDest_) - ->add(*outG->subMatrix(info.startIdx, info.length, tmpSrc_)); - } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/NCELayer.cpp b/paddle/gserver/layers/NCELayer.cpp deleted file mode 100644 index cc48fe100f12446f9522078119ae2ead039a82cc..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/NCELayer.cpp +++ /dev/null @@ -1,323 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "Layer.h" -#include "MultinomialSampler.h" -#include "paddle/math/MathFunctions.h" - -namespace paddle { - -/** - * Noise-contrastive estimation. - * Implements the method in the following paper: - * A fast and simple algorithm for training neural probabilistic language - * models. - * - * The config file api is nce_layer. - */ -class NCELayer : public Layer { - int numClasses_; - /// number of input layer besides labelLayer and weightLayer - int numInputs_; - LayerPtr labelLayer_; - /// weight layer, can be None - LayerPtr weightLayer_; - WeightList weights_; - std::unique_ptr biases_; - std::unique_ptr sampler_; - - std::uniform_int_distribution rand_; - - struct Sample { - int sampleId; - int labelId; - bool target; - real weight; - }; - std::vector samples_; - /// whether samples_ is prepared - bool prepared_; - Argument sampleOut_; - - IVectorPtr labelIds_; - - public: - explicit NCELayer(const LayerConfig& config) - : Layer(config), - numClasses_(config.num_classes()), - rand_(0, config.num_classes() - 1), - prepared_(false) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* initialize the weightList */ - size_t i; - for (i = 0; i < inputLayers_.size(); i++) { - if (!parameters_[i]) break; - size_t width = inputLayers_[i]->getSize(); - // create a new weight - CHECK_EQ(parameters_[i]->getSize(), width * numClasses_); - Weight* w = new Weight(numClasses_, width, parameters_[i]); - - // append the new weight to the list - weights_.emplace_back(w); - } - - CHECK_EQ(1U, getSize()); - - numInputs_ = i; - CHECK_GE(numInputs_, 1) - << "Must have at least one input besides label and weight"; - CHECK_LT(i, inputLayers_.size()) << "Missing label layer"; - labelLayer_ = inputLayers_[i]; - if (++i < inputLayers_.size()) { - weightLayer_ = inputLayers_[i]; - ++i; - } - CHECK_EQ(i, inputLayers_.size()); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - CHECK_EQ(biasParameter_->getSize(), (size_t)numClasses_); - biases_.reset(new Weight(1, numClasses_, biasParameter_)); - } - - if (config_.neg_sampling_dist_size()) { - CHECK_EQ(numClasses_, config_.neg_sampling_dist_size()); - sampler_.reset(MultinomialSampler::create( - config_.neg_sampling_dist().data(), numClasses_)); - } - - return true; - } - - void prepareSamples() { - CHECK(!useGpu_) << "GPU is not supported"; - - int batchSize = getInput(*labelLayer_).getBatchSize(); - IVectorPtr label = getInput(*labelLayer_).ids; - - CpuSparseMatrixPtr multiLabel = std::dynamic_pointer_cast( - getInput(*labelLayer_).value); - - CHECK(label || multiLabel) - << "The label layer must have ids or NonValueSparseMatrix value"; - - auto& randEngine = ThreadLocalRandomEngine::get(); - - samples_.clear(); - samples_.reserve(batchSize * (1 + config_.num_neg_samples())); - - real* weight = - weightLayer_ ? getInputValue(*weightLayer_)->getData() : nullptr; - - for (int i = 0; i < batchSize; ++i) { - real w = weight ? weight[i] : 1; - if (label) { - int* ids = label->getData(); - samples_.push_back({i, ids[i], true, w}); - } else { - const int* cols = multiLabel->getRowCols(i); - int n = multiLabel->getColNum(i); - for (int j = 0; j < n; ++j) { - samples_.push_back({i, cols[j], true, w}); - } - } - for (int j = 0; j < config_.num_neg_samples(); ++j) { - int id = sampler_ ? sampler_->gen(randEngine) : rand_(randEngine); - samples_.push_back({i, id, false, w}); - } - } - prepared_ = true; - } - - void prefetch() override { - prepareSamples(); - IVector::resizeOrCreate(labelIds_, samples_.size(), useGpu_); - int* ids = labelIds_->getData(); - for (size_t i = 0; i < samples_.size(); ++i) { - ids[i] = samples_[i].labelId; - } - - for (int i = 0; i < numInputs_; ++i) { - auto sparseParam = - dynamic_cast(weights_[i]->getW().get()); - if (sparseParam) { - sparseParam->addRows(labelIds_); - } - } - } - - void forward(PassType passType) override { - Layer::forward(passType); - - CHECK(!useGpu_) << "GPU is not supported"; - - if (!prepared_) { - if (passType == PASS_GC) { - ThreadLocalRandomEngine::get().seed(ThreadLocalRand::getDefaultSeed()); - } - prepareSamples(); - } - prepared_ = false; - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(0)->getHeight(); - int size = getSize(); - resetOutput(batchSize, size); - - Matrix::resizeOrCreate(sampleOut_.value, - 1, - samples_.size(), - /* trans= */ false, - useGpu_); - - forwardBias(); - - for (int l = 0; l < numInputs_; ++l) { - forwardOneInput(l); - } - - auto status = activation_->forward(sampleOut_); - status.check(); - - forwardCost(); - } - - void backward(const UpdateCallback& callback) override { - Matrix::resizeOrCreate(sampleOut_.grad, - 1, - samples_.size(), - /* trans= */ false, - useGpu_); - - backwardCost(); - - auto status = activation_->backward(sampleOut_); - status.check(); - - if (biases_->getWGrad()) { - backwardBias(callback); - } - - for (int l = 0; l < numInputs_; ++l) { - backwardOneInput(l, callback); - } - } - - void forwardBias() { - if (!biases_) { - sampleOut_.value->zeroMem(); - } else { - real* bias = biases_->getW()->getData(); - real* sampleOut = sampleOut_.value->getData(); - for (size_t i = 0; i < samples_.size(); ++i) { - sampleOut[i] = bias[samples_[i].labelId]; - } - } - } - - void backwardBias(const UpdateCallback& callback) { - if (!biases_) return; - real* bias = biases_->getWGrad()->getData(); - real* sampleOut = sampleOut_.grad->getData(); - for (size_t i = 0; i < samples_.size(); ++i) { - bias[samples_[i].labelId] += sampleOut[i]; - } - biases_->incUpdate(callback); - } - - void forwardOneInput(int layerId) { - const MatrixPtr& inputMat = getInputValue(layerId); - const MatrixPtr& weightMat = weights_[layerId]->getW(); - - int dim = inputMat->getWidth(); - real* sampleOut = sampleOut_.value->getData(); - - for (size_t i = 0; i < samples_.size(); ++i) { - sampleOut[i] += dotProduct(dim, - inputMat->getRowBuf(samples_[i].sampleId), - weightMat->getRowBuf(samples_[i].labelId)); - } - } - - void backwardOneInput(int layerId, const UpdateCallback& callback) { - const MatrixPtr& inputMat = getInputValue(layerId); - const MatrixPtr& inputGradMat = getInputGrad(layerId); - const MatrixPtr& weightMat = weights_[layerId]->getW(); - const MatrixPtr& weightGradMat = weights_[layerId]->getWGrad(); - - int dim = inputMat->getWidth(); - real* sampleGrad = sampleOut_.grad->getData(); - - if (weightGradMat) { - for (size_t i = 0; i < samples_.size(); ++i) { - axpy(dim, - sampleGrad[i], - inputMat->getRowBuf(samples_[i].sampleId), - weightGradMat->getRowBuf(samples_[i].labelId)); - } - weights_[layerId]->incUpdate(callback); - } - - if (inputGradMat) { - for (size_t i = 0; i < samples_.size(); ++i) { - axpy(dim, - sampleGrad[i], - weightMat->getRowBuf(samples_[i].labelId), - inputGradMat->getRowBuf(samples_[i].sampleId)); - } - } - } - - void forwardCost() { - real* out = output_.value->getData(); - real* sampleOut = sampleOut_.value->getData(); - real b = 1. / numClasses_ * config_.num_neg_samples(); - for (size_t i = 0; i < samples_.size(); ++i) { - real o = sampleOut[i]; - if (sampler_) { - b = config_.num_neg_samples() * - config_.neg_sampling_dist(samples_[i].labelId); - } - real cost = samples_[i].target ? -log(o / (o + b)) : -log(b / (o + b)); - out[samples_[i].sampleId] += samples_[i].weight * cost; - } - } - - void backwardCost() { - real* sampleOut = sampleOut_.value->getData(); - real* sampleGrad = sampleOut_.grad->getData(); - - real b = 1. / numClasses_ * config_.num_neg_samples(); - for (size_t i = 0; i < samples_.size(); ++i) { - real o = sampleOut[i]; - if (sampler_) { - b = config_.num_neg_samples() * - config_.neg_sampling_dist(samples_[i].labelId); - } - real w = samples_[i].weight; - sampleGrad[i] = samples_[i].target ? -w * b / (o * (o + b)) : w / (o + b); - } - } -}; - -REGISTER_LAYER(nce, NCELayer); - -} // namespace paddle diff --git a/paddle/gserver/layers/NormLayer.cpp b/paddle/gserver/layers/NormLayer.cpp deleted file mode 100644 index 4678f6fa9ab184870fc2651def18f47da9a0cc01..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/NormLayer.cpp +++ /dev/null @@ -1,59 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "NormLayer.h" -#include "NormProjectionLayer.h" -#include "paddle/utils/Logging.h" -namespace paddle { - -REGISTER_LAYER_CREATE_FUNC(norm, &NormLayer::create); - -Layer* NormLayer::create(const LayerConfig& config) { - CHECK_EQ(config.inputs_size(), 1); - const std::string& norm = config.inputs(0).norm_conf().norm_type(); - if (norm == "rnorm") { - return new ResponseNormLayer(config); - } else if (norm == "cmrnorm-projection") { - return new CMRProjectionNormLayer(config); - } else if (norm == "cross-channel-norm") { - return new CrossChannelNormLayer(config); - } else { - LOG(FATAL) << "Unknown norm type: " << norm; - return nullptr; - } -} - -bool ResponseNormLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - NormLayer::init(layerMap, parameterMap); - - /* the size of inputs for norm-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - - const NormConfig& conf = config_.inputs(0).norm_conf(); - channels_ = conf.channels(); - size_ = conf.size(); - scale_ = conf.scale(); - pow_ = conf.pow(); - outputX_ = conf.output_x(); - imgSize_ = conf.img_size(); - denoms_ = NULL; - - outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); - imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - return true; -} - -} // namespace paddle diff --git a/paddle/gserver/layers/NormLayer.h b/paddle/gserver/layers/NormLayer.h deleted file mode 100644 index 3807584415f99a7110170748501589dac85eac52..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/NormLayer.h +++ /dev/null @@ -1,99 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Layer.h" -#include "NormLayer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief Basic parent layer of normalization - * - * @note Normalize the input in local region - */ -class NormLayer : public Layer { - public: - explicit NormLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override { - Layer::init(layerMap, parameterMap); - return true; - } - - /** - * @brief create norm layer by norm_type - */ - static Layer* create(const LayerConfig& config); -}; - -/** - * @brief response normalization within feature maps - * namely normalize in independent channel - * When code refactoring, we delete the original implementation. - * Need to implement in the futrue. - */ -class ResponseNormLayer : public NormLayer { - protected: - size_t channels_, size_, outputX_, imgSize_, outputY_, imgSizeY_; - real scale_, pow_; - MatrixPtr denoms_; - - public: - explicit ResponseNormLayer(const LayerConfig& config) : NormLayer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - void forward(PassType passType) override { LOG(FATAL) << "Not implemented"; } - void backward(const UpdateCallback& callback = nullptr) override { - LOG(FATAL) << "Not implemented"; - } -}; - -/** - * This layer applys normalization across the channels of each sample to a - * conv layer's output, and scales the output by a group of trainable factors - * whose dimensions equal to the number of channels. - * - Input: One and only one input layer are accepted. - * - Output: The normalized data of the input data. - * Reference: - * Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, - * Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector - */ -class CrossChannelNormLayer : public NormLayer { - public: - explicit CrossChannelNormLayer(const LayerConfig& config) - : NormLayer(config) {} - bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); - void forward(PassType passType); - void backward(const UpdateCallback& callback); - MatrixPtr createSampleMatrix(MatrixPtr data, size_t iter, size_t spatialDim); - MatrixPtr createSpatialMatrix(MatrixPtr data, size_t iter, size_t spatialDim); - - protected: - size_t channels_; - std::unique_ptr scale_; - MatrixPtr scaleDiff_; - MatrixPtr normBuffer_; - MatrixPtr dataBuffer_; - MatrixPtr channelBuffer_; - MatrixPtr spatialBuffer_; - MatrixPtr sampleBuffer_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/NormProjectionLayer.cpp b/paddle/gserver/layers/NormProjectionLayer.cpp deleted file mode 100644 index 3013bbdbc791546897fca51e73a056f2c843e63f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/NormProjectionLayer.cpp +++ /dev/null @@ -1,101 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "NormProjectionLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { -size_t CMRProjectionNormLayer::getSize() { - CHECK_EQ(inputLayers_.size(), 1UL); - size_t layerSize = 0; - imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight(); - imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth(); - if (imgSizeH_ == 0) { - imgSizeH_ = imgSizeY_; - } - if (imgSizeW_ == 0) { - imgSizeW_ = imgSize_; - } - outputH_ = imgSizeH_; - outputW_ = imgSizeW_; - layerSize = outputH_ * outputW_ * channels_; - - getOutput().setFrameHeight(outputH_); - getOutput().setFrameWidth(outputW_); - return layerSize; -} - -bool CMRProjectionNormLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - ResponseNormLayer::init(layerMap, parameterMap); - - /* the size of inputs for norm-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - - createFunction( - forward_, - "CrossMapNormal", - FuncConfig().set("size", size_).set("scale", scale_).set("pow", pow_)); - createFunction( - backward_, - "CrossMapNormalGrad", - FuncConfig().set("size", size_).set("scale", scale_).set("pow", pow_)); - - return true; -} - -void CMRProjectionNormLayer::forward(PassType passType) { - Layer::forward(passType); - /* malloc memory for the output_ if necessary */ - /* note: one sample correspond to one row */ - MatrixPtr input = inputLayers_[0]->getOutputValue(); - size_t batchSize = input->getHeight(); - int size = getSize(); - resetOutput(batchSize, size); - - Matrix::resizeOrCreate(denoms_, batchSize, size, /* trans */ false, useGpu_); - - shape_ = TensorShape({batchSize, channels_, imgSizeH_, imgSizeW_}); - - // prepare forward arguments - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), shape_); - outputs.addArg(*getOutputValue(), shape_, ASSIGN_TO); - outputs.addArg(*denoms_, shape_, ASSIGN_TO); - - forward_[0]->calc(inputs, outputs); -} - -void CMRProjectionNormLayer::backward(const UpdateCallback& callback) { - (void)callback; - - if (NULL == getInputGrad(0)) { - return; - } - - // prepare backward arguments - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), shape_); - inputs.addArg(*getOutputValue(), shape_); - inputs.addArg(*getOutputGrad(), shape_); - inputs.addArg(*denoms_, shape_); - outputs.addArg(*getInputGrad(0), shape_, ADD_TO); - - backward_[0]->calc(inputs, outputs); -} -} // namespace paddle diff --git a/paddle/gserver/layers/NormProjectionLayer.h b/paddle/gserver/layers/NormProjectionLayer.h deleted file mode 100644 index 64803a1603599f2e393ec772a32d64f4d271fe71..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/NormProjectionLayer.h +++ /dev/null @@ -1,47 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "NormLayer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief response normalization across feature maps - * namely normalize in number of size_ channels - */ -class CMRProjectionNormLayer : public ResponseNormLayer { - size_t imgSizeH_, imgSizeW_; - size_t outputH_, outputW_; - - public: - explicit CMRProjectionNormLayer(const LayerConfig& config) - : ResponseNormLayer(config) {} - - ~CMRProjectionNormLayer() {} - - size_t getSize(); - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - protected: - TensorShape shape_; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/Operator.h b/paddle/gserver/layers/Operator.h deleted file mode 100644 index 42d525ef3e4534acea7512d5ecdbe8a0e1d110d9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Operator.h +++ /dev/null @@ -1,96 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ModelConfig.pb.h" -#include "paddle/parameter/Parameter.h" - -#include "Layer.h" -#include "paddle/parameter/Argument.h" - -namespace paddle { - -// Macro for registering a operator type -// Example: REGISTER_OPERATOR(dot_mul, DotMulOperator); -#define REGISTER_OPERATOR(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name([]() { \ - Operator::registrar_.registerClass<__class_name>(#__type_name); \ - }) - -/** - * Operator like Projection, but takes more than one Arguments as input. - * @note: Operator can't have parameters. - */ -class Operator { - public: - static Operator* create(const OperatorConfig& config, bool useGpu); - - Operator(const OperatorConfig& config, bool useGpu) - : config_(config), useGpu_(useGpu) {} - - virtual ~Operator() {} - - const OperatorConfig& getConfig() const { return config_; } - - static ClassRegistrar registrar_; - - /** - * Forward propagation. If backward() will be called, in and out must be kept - * valid until then. - * @param ins inputs of operator - * @param out output of operator - * @param passType PASS_TRAIN of PASS_TEST - */ - void forward(std::vector ins, - Argument* out, - PassType passType) { - ins_ = ins; - out_ = out; - passType_ = passType; - forward(); - } - - virtual void prefetch(const Argument* in) {} - virtual void forward() = 0; - virtual void backward() = 0; - - /** - * See comment in Layer.h for the function with the same name. - */ - virtual void resetState() {} - - /** - * Set layer state. - */ - virtual void setState(LayerStatePtr state) {} - - /** - * Set layer state. - */ - virtual LayerStatePtr getState() { return nullptr; } - - protected: - /// Config of operator - OperatorConfig config_; - bool useGpu_; - - /// Store `ins` passed to forward() - std::vector ins_; - /// Store `out` passed to forward() - Argument* out_; - /// Store `passType` passed to forward() - PassType passType_; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/OuterProdLayer.cpp b/paddle/gserver/layers/OuterProdLayer.cpp deleted file mode 100644 index 11a910f3316114b309efe9007a156e842b3d6229..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/OuterProdLayer.cpp +++ /dev/null @@ -1,141 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief A layer for computing the outer product of two vectors - * @note used in NEURAL TURING MACHINE - * Input1: vector (batchSize * dim1) - * Input2: vector (batchSize * dim2) - * Output: a matrix: (batchSize * (dim1*dim2)) - */ - -class OuterProdLayer : public Layer { - protected: - MatrixPtr tmpMtx0; - MatrixPtr tmpRow0; - MatrixPtr tmpRow1; - - public: - explicit OuterProdLayer(const LayerConfig& config) : Layer(config) {} - - ~OuterProdLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(out_prod, OuterProdLayer); - -bool OuterProdLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2U); - - size_t dim0 = inputLayers_[0]->getSize(); - size_t dim1 = inputLayers_[1]->getSize(); - - CHECK_EQ(dim0 * dim1, getSize()) << "Dimension mismatch"; - - tmpRow0 = Matrix::create( - nullptr, /* height= */ 1, dim0, /* trans= */ false, useGpu_); - tmpRow1 = Matrix::create( - nullptr, /* height= */ 1, dim1, /* trans= */ false, useGpu_); - tmpMtx0 = Matrix::create(nullptr, - /* height= */ dim0, - dim1, - /* trans= */ false, - useGpu_); - return true; -} - -void OuterProdLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV0->getHeight(); - size_t dim0 = inV0->getWidth(); - size_t dim1 = inV1->getWidth(); - - CHECK_EQ(dim0 * dim1, getSize()); - CHECK_EQ(inV1->getHeight(), batchSize); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, dim0 * dim1); - } - - MatrixPtr outV = getOutputValue(); - - { - REGISTER_TIMER_INFO("FwOutProdTimer", getName().c_str()); - for (size_t i = 0; i < batchSize; i++) { - tmpMtx0->setData(outV->getData() + i * dim0 * dim1); - tmpRow0->setData(inV0->getData() + i * dim0); - tmpRow1->setData(inV1->getData() + i * dim1); - - tmpMtx0->mul(*tmpRow0->getTranspose(), *tmpRow1); - } - } -} - -void OuterProdLayer::backward(const UpdateCallback& callback) { - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr outG = getOutputGrad(); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - - size_t batchSize = inV0->getHeight(); - size_t dim0 = inV0->getWidth(); - size_t dim1 = inV1->getWidth(); - - { - REGISTER_TIMER_INFO("BwOutProdTimer", getName().c_str()); - - if (inG0) { - for (size_t i = 0; i < batchSize; i++) { - tmpMtx0->setData(outG->getData() + i * dim0 * dim1); - tmpRow0->setData(inG0->getData() + i * dim0); - tmpRow1->setData(inV1->getData() + i * dim1); - - tmpRow0->mul(*tmpRow1, *tmpMtx0->getTranspose(), 1, 1); - } - } - - if (inG1) { - for (size_t i = 0; i < batchSize; i++) { - tmpMtx0->setData(outG->getData() + i * dim0 * dim1); - tmpRow0->setData(inV0->getData() + i * dim0); - tmpRow1->setData(inG1->getData() + i * dim1); - - tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 1); - } - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/PadLayer.cpp b/paddle/gserver/layers/PadLayer.cpp deleted file mode 100644 index b1910e108b5b2f7b55a2aa1527b96e6b8a16f348..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PadLayer.cpp +++ /dev/null @@ -1,106 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PadLayer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(pad, PadLayer); - -bool PadLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - auto& pad_conf = config_.inputs(0).pad_conf(); - auto& img_conf = pad_conf.image_conf(); - CHECK_EQ(config_.inputs_size(), 1); - inDims_ = TensorShape( - {0, - img_conf.channels(), - img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(), - img_conf.img_size()}); - - CHECK_EQ(2, pad_conf.pad_c_size()); - CHECK_EQ(2, pad_conf.pad_h_size()); - CHECK_EQ(2, pad_conf.pad_w_size()); - padc_ = {pad_conf.pad_c(0), pad_conf.pad_c(1)}; - padh_ = {pad_conf.pad_h(0), pad_conf.pad_h(1)}; - padw_ = {pad_conf.pad_w(0), pad_conf.pad_w(1)}; - - outDims_ = TensorShape(4); - setOutDims(0); - - createFunction(forward_, - "Pad", - FuncConfig() - .set("channel", padc_) - .set("height", padh_) - .set("width", padw_)); - createFunction(backward_, - "PadGrad", - FuncConfig() - .set("channel", padc_) - .set("height", padh_) - .set("width", padw_)); - - return true; -} - -void PadLayer::setOutDims(const size_t batchSize) { - outDims_.reshape({batchSize, - inDims_[1] + padc_[0] + padc_[1], - inDims_[2] + padh_[0] + padh_[1], - inDims_[3] + padw_[0] + padw_[1]}); -} - -void PadLayer::setTensorDim(const size_t batchSize) { - CHECK_EQ(static_cast(inputLayers_.size()), 1); - inDims_.setDim(0, batchSize); - int h = inputLayers_[0]->getOutput().getFrameHeight(); - if (h != 0) inDims_.setDim(2, h); - int w = inputLayers_[0]->getOutput().getFrameWidth(); - if (w != 0) inDims_.setDim(3, w); - setOutDims(batchSize); -} - -void PadLayer::forward(PassType passType) { - Layer::forward(passType); - MatrixPtr input = inputLayers_[0]->getOutputValue(); - size_t batchSize = input->getHeight(); - setTensorDim(batchSize); - int size = outDims_[1] * outDims_[2] * outDims_[3]; - resetOutput(batchSize, size); - MatrixPtr outV = getOutputValue(); - REGISTER_TIMER_INFO("PadForward", getName().c_str()); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), inDims_); - outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO); - forward_[0]->calc(inputs, outputs); -} - -void PadLayer::backward(const UpdateCallback& callback) { - (void)callback; - REGISTER_TIMER_INFO("PadBackward", getName().c_str()); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getOutputGrad(), outDims_); - outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); - backward_[0]->calc(inputs, outputs); -} -} // namespace paddle diff --git a/paddle/gserver/layers/ParameterReluLayer.cpp b/paddle/gserver/layers/ParameterReluLayer.cpp deleted file mode 100644 index 12d04fc1c3ca169179beafc372a07a2e6d0a1773..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ParameterReluLayer.cpp +++ /dev/null @@ -1,69 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ParameterReluLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(prelu, ParameterReluLayer); - -bool ParameterReluLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - CHECK_EQ(inputLayers_.size(), 1UL); - CHECK_EQ(inputLayers_.size(), parameters_.size()); - partialSum_ = config_.partial_sum(); - CHECK_GT(partialSum_, 0UL) << "partial_sum must be larger than zero."; - CHECK(!(inputLayers_[0]->getSize() % partialSum_)) - << "Incorrect value for partialSum: " << partialSum_ - << " must divide input size: " << inputLayers_[0]->getSize(); - CHECK_EQ(getSize() / partialSum_, parameters_[0]->getSize()); - weight_ = std::unique_ptr(new Weight( - 1UL, inputLayers_[0]->getSize() / partialSum_, parameters_[0])); - return true; -} - -void ParameterReluLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInput(0).getBatchSize(); - int size = getSize(); - reserveOutput(batchSize, size); - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - outV->paramReluForward(*(getInput(0).value), *(weight_->getW())); - } -} - -void ParameterReluLayer::backward(const UpdateCallback& callback) { - if (weight_->getWGrad()) { - weight_->getWGrad()->paramReluBackwardW(*getOutputGrad(), - *(getInputValue(0))); - } - - MatrixPtr preGrad = getInputGrad(0); - preGrad->paramReluBackwardDiff( - *getOutputGrad(), *(getInputValue(0)), *(weight_->getW())); - { - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); - weight_->getParameterPtr()->incUpdate(callback); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ParameterReluLayer.h b/paddle/gserver/layers/ParameterReluLayer.h deleted file mode 100644 index 4553413fcdbecbc83e1f50e8ffbe874fdf05d828..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ParameterReluLayer.h +++ /dev/null @@ -1,65 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * @brief ParameterReluLayer active inputs with learnable parameter weight_. - * forward: - * \f[ - * y = x > 0 ? x : w .* x - * \f] - * backward: - * \f[ - * dx = x > 0 ? dy : w .* dy \\ - * dw = x > 0 ? 0 : dy.*x - * \f] - * Here, x is the input, w is the weight, y is the output. - * dx, dw, dy is the gradient. - */ - -class ParameterReluLayer : public Layer { - protected: - std::unique_ptr weight_; - - /** - * @brief partialSum_ makes a group of inputs share same weights, - * - partialSum_ = 1: - * element wise activation: each element has a weight_, - * - partialSum_ = number of elements in one channel, - * channels wise parameter activation, elements in a channel - * share same weight_, - * - partialSum_ = number of outputs - * all elements share same weight_, - */ - size_t partialSum_; - - public: - explicit ParameterReluLayer(const LayerConfig& config) : Layer(config) {} - - ~ParameterReluLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/Pool3DLayer.cpp b/paddle/gserver/layers/Pool3DLayer.cpp deleted file mode 100644 index 3ac9eb0d8198814c9f01fe101a60ab1f1f431062..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Pool3DLayer.cpp +++ /dev/null @@ -1,178 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Pool3DLayer.h" -#include "PoolProjectionLayer.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -REGISTER_LAYER(pool3d, Pool3DLayer); - -bool Pool3DLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - /* the size of inputs for pool-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - - const PoolConfig& conf = config_.inputs(0).pool_conf(); - poolType_ = conf.pool_type(); - channels_ = conf.channels(); - - sizeX_ = conf.size_x(); - sizeY_ = conf.size_y(); - sizeZ_ = conf.size_z(); - - strideW_ = conf.stride(); - strideH_ = conf.stride_y(); - strideD_ = conf.stride_z(); - - imgSizeW_ = conf.img_size(); - imgSizeH_ = conf.img_size_y(); - imgSizeD_ = conf.img_size_z(); - - paddingW_ = conf.padding(); - paddingH_ = conf.padding_y(); - paddingD_ = conf.padding_z(); - - outputW_ = conf.output_x(); - outputH_ = conf.output_y(); - outputD_ = conf.output_z(); - - return true; -} - -size_t Pool3DLayer::getSize() { - CHECK_EQ(inputLayers_.size(), 1UL); - - size_t layerSize = 0; - outputD_ = outputSize(imgSizeD_, sizeZ_, paddingD_, strideD_, false); - outputH_ = outputSize(imgSizeH_, sizeY_, paddingH_, strideH_, false); - outputW_ = outputSize(imgSizeW_, sizeX_, paddingW_, strideW_, false); - - layerSize = outputD_ * outputH_ * outputW_ * channels_; - getOutput().setFrameHeight(outputH_); - getOutput().setFrameWidth(outputW_); - getOutput().setFrameDepth(outputD_); - return layerSize; -} - -void Pool3DLayer::forward(PassType passType) { - Layer::forward(passType); - const MatrixPtr& inMat = inputLayers_[0]->getOutputValue(); - size_t batchSize = inMat->getHeight(); - size_t outWidth = getSize(); - resetOutput(batchSize, outWidth); - Matrix::resizeOrCreate(maxPoolIdx_, batchSize, outWidth, false, useGpu_); - const MatrixPtr outMat = getOutputValue(); - - if (poolType_ == "avg") { - outMat->avgPool3DForward(*inMat, - channels_, - imgSizeD_, - imgSizeH_, - imgSizeW_, - outputD_, - outputH_, - outputW_, - sizeZ_, - sizeY_, - sizeX_, - strideD_, - strideH_, - strideW_, - paddingD_, - paddingH_, - paddingW_); - } else if (poolType_ == "max") { - outMat->maxPool3DForward(*inMat, - *maxPoolIdx_, - channels_, - imgSizeD_, - imgSizeH_, - imgSizeW_, - outputD_, - outputH_, - outputW_, - sizeZ_, - sizeY_, - sizeX_, - strideD_, - strideH_, - strideW_, - paddingD_, - paddingH_, - paddingW_); - } else { - LOG(FATAL) << "Unknown pool type: " << poolType_; - } - forwardActivation(); -} - -void Pool3DLayer::backward(const UpdateCallback& callback) { - backwardActivation(); - - (void)callback; - if (NULL == getInputGrad(0)) return; - MatrixPtr inMat = inputLayers_[0]->getOutputValue(); - MatrixPtr inGradMat = inputLayers_[0]->getOutputGrad(); - MatrixPtr outMat = getOutputValue(); - MatrixPtr outGradMat = getOutputGrad(); - - if (poolType_ == "avg") { - inGradMat->avgPool3DBackward(*outGradMat, - imgSizeD_, - imgSizeH_, - imgSizeW_, - outputD_, - outputH_, - outputW_, - sizeZ_, - sizeY_, - sizeZ_, - strideD_, - strideH_, - strideW_, - paddingD_, - paddingH_, - paddingW_, - 1.0, - 1.0); - } else if (poolType_ == "max") { - inGradMat->maxPool3DBackward(*outGradMat, - *maxPoolIdx_, - imgSizeD_, - imgSizeH_, - imgSizeW_, - outputD_, - outputH_, - outputW_, - sizeZ_, - sizeY_, - sizeZ_, - strideD_, - strideH_, - strideW_, - paddingD_, - paddingH_, - paddingW_, - 1.0, - 1.0); - } else { - LOG(FATAL) << "Unknown pool type: " << poolType_; - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/Pool3DLayer.h b/paddle/gserver/layers/Pool3DLayer.h deleted file mode 100644 index 32605f8b7028cfb4909c885e83017a8cffa79575..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Pool3DLayer.h +++ /dev/null @@ -1,49 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Layer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief Basic parent layer of pooling - * Pools the input within regions - */ -class Pool3DLayer : public Layer { - public: - explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {} - ~Pool3DLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override; - size_t getSize(); - - protected: - int channels_; - int sizeX_, sizeY_, sizeZ_; - int strideW_, strideH_, strideD_; - int paddingW_, paddingH_, paddingD_; - int imgSizeW_, imgSizeH_, imgSizeD_; - int outputW_, outputH_, outputD_; - std::string poolType_; - MatrixPtr maxPoolIdx_; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/PoolLayer.cpp b/paddle/gserver/layers/PoolLayer.cpp deleted file mode 100644 index ee589e6be51b1e66984f5a1d808b73aab962821d..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PoolLayer.cpp +++ /dev/null @@ -1,70 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PoolLayer.h" -#include "MaxPoolWithMaskLayer.h" -#include "PoolProjectionLayer.h" -#include "paddle/utils/Logging.h" -#ifdef PADDLE_WITH_CUDA -#include "CudnnPoolLayer.h" -#endif -namespace paddle { - -REGISTER_LAYER_CREATE_FUNC(pool, &PoolLayer::create); - -bool PoolLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* the size of inputs for pool-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - - const PoolConfig& conf = config_.inputs(0).pool_conf(); - poolType_ = conf.pool_type(); - channels_ = conf.channels(); - sizeX_ = conf.size_x(); - stride_ = conf.stride(); - outputX_ = conf.output_x(); - imgSize_ = conf.img_size(); - confPadding_ = conf.padding(); - - sizeY_ = conf.has_size_y() ? conf.size_y() : conf.size_x(); - imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); - strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride(); - confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding(); - outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); - - excludeMode_ = conf.has_exclude_mode() ? conf.exclude_mode() : true; - return true; -} - -Layer* PoolLayer::create(const LayerConfig& config) { - CHECK_EQ(config.inputs_size(), 1); - const std::string& pool = config.inputs(0).pool_conf().pool_type(); - if (pool == "max-projection" || pool == "avg-projection") { - return new PoolProjectionLayer(config); -#ifdef PADDLE_WITH_CUDA - } else if (CudnnPoolLayer::typeCheck(pool)) { - return new CudnnPoolLayer(config); -#endif - } else if (pool == "max-pool-with-mask") { - return new MaxPoolWithMaskLayer(config); - } else { - LOG(FATAL) << "Unknown pool type: " << pool; - return nullptr; - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/PoolLayer.h b/paddle/gserver/layers/PoolLayer.h deleted file mode 100644 index 99f8f148e2eb00f7e431e7d8c5acbf9e27574017..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PoolLayer.h +++ /dev/null @@ -1,55 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Layer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" - -namespace paddle { - -/** - * @brief Basic parent layer of pooling - * Pools the input within regions - */ -class PoolLayer : public Layer { - protected: - size_t channels_, sizeX_, stride_, outputX_, imgSize_; - int confPadding_; - - size_t sizeY_; - size_t imgSizeY_; - size_t strideY_; - size_t outputY_; - int confPaddingY_; - - std::string poolType_; - - bool excludeMode_; - - public: - explicit PoolLayer(const LayerConfig& config) : Layer(config) {} - - /** - * @brief create pooling layer by pool_type - */ - static Layer* create(const LayerConfig& config); - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/PoolProjection.h b/paddle/gserver/layers/PoolProjection.h deleted file mode 100644 index 8004cc1550337160b7f022c97a23ed8eb9d43ca4..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PoolProjection.h +++ /dev/null @@ -1,68 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Projection.h" -#include "paddle/math/MathUtils.h" - -namespace paddle { - -class PoolProjection : public Projection { - protected: - size_t imgSizeY_, imgSize_; - size_t outputY_, outputX_; - size_t strideY_, stride_; - size_t sizeY_, sizeX_; - int confPaddingY_, confPadding_; - size_t channels_; - std::string poolType_; - bool excludeMode_; - - public: - PoolProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu); - - static PoolProjection* create(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu); - - const std::string& getPoolType() const { return poolType_; } - - size_t getSize(); -}; - -class MaxPoolProjection : public PoolProjection { - public: - MaxPoolProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu) - : PoolProjection(config, parameter, useGpu) {} - - virtual void forward(); - virtual void backward(const UpdateCallback& callback = nullptr); -}; - -class AvgPoolProjection : public PoolProjection { - public: - AvgPoolProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu) - : PoolProjection(config, parameter, useGpu) {} - - virtual void forward(); - virtual void backward(const UpdateCallback& callback = nullptr); -}; -} // namespace paddle diff --git a/paddle/gserver/layers/PoolProjectionLayer.cpp b/paddle/gserver/layers/PoolProjectionLayer.cpp deleted file mode 100644 index 73d320e67ec09513f419ecdd45a57fc5c54df5ed..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PoolProjectionLayer.cpp +++ /dev/null @@ -1,65 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PoolProjectionLayer.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -size_t PoolProjectionLayer::getSize() { - CHECK_EQ(inputLayers_.size(), 1UL); - size_t layerSize = 0; - imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight(); - imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth(); - if (imgSizeH_ == 0) { - imgSizeH_ = imgSizeY_; - } - if (imgSizeW_ == 0) { - imgSizeW_ = imgSize_; - } - - outputH_ = outputSize(imgSizeH_, - sizeY_, - confPaddingY_, - strideY_, - /* caffeMode */ false); - outputW_ = outputSize(imgSizeW_, - sizeX_, - confPadding_, - stride_, - /* caffeMode */ false); - - layerSize = outputH_ * outputW_ * channels_; - - return layerSize; -} - -void PoolProjectionLayer::forward(PassType passType) { - Layer::forward(passType); - const Argument& in = getInput(0); - int batchSize = in.value->getHeight(); - int size = getSize(); - resetOutput(batchSize, size); - poolProjection_->forward(&in, &output_, passType); -} - -void PoolProjectionLayer::backward(const UpdateCallback& callback) { - (void)callback; - if (NULL == getInputGrad(0)) { - return; - } - poolProjection_->backward(callback); -} -} // namespace paddle diff --git a/paddle/gserver/layers/PoolProjectionLayer.h b/paddle/gserver/layers/PoolProjectionLayer.h deleted file mode 100644 index 9ad144cc2ad426caa522bf1061a750d47e64a755..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PoolProjectionLayer.h +++ /dev/null @@ -1,46 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "PoolLayer.h" -#include "PoolProjection.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * @brief Basic parent layer of different kinds of pooling - */ -class PoolProjectionLayer : public PoolLayer { - protected: - size_t imgSizeH_, imgSizeW_; - size_t outputH_, outputW_; - std::unique_ptr poolProjection_; - ProjectionConfig projectionConfig_; - - public: - explicit PoolProjectionLayer(const LayerConfig& config) : PoolLayer(config) { - PoolConfig* conf = projectionConfig_.mutable_pool_conf(); - *conf = config_.inputs(0).pool_conf(); - poolProjection_.reset( - PoolProjection::create(projectionConfig_, nullptr, useGpu_)); - } - - size_t getSize(); - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/PowerLayer.cpp b/paddle/gserver/layers/PowerLayer.cpp deleted file mode 100644 index 7e8d60db8fe588026c6040099745c3aefd7237b5..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PowerLayer.cpp +++ /dev/null @@ -1,120 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * This layer applys a power function to a vector element-wise, - * which is used in NEURAL TURING MACHINE. - * \f[ - * y = x^w - * \f] - * where \f$x\f$ is a input vector, \f$w\f$ is scalar weight, - * and output \f$y\f$ is a vector. - * - * The config file api is power_layer. - */ - -class PowerLayer : public Layer { - protected: - MatrixPtr tmpMtx; - - public: - explicit PowerLayer(const LayerConfig& config) : Layer(config) {} - - ~PowerLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(power, PowerLayer); - -bool PowerLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2U); - - return true; -} - -void PowerLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV1->getHeight(); - size_t dataDim = inV1->getWidth(); - - CHECK_EQ(getSize(), dataDim); - CHECK_EQ(1U, inV0->getWidth()); - CHECK_EQ(batchSize, inV0->getHeight()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, dataDim); - } - - MatrixPtr outV = getOutputValue(); - - { - REGISTER_TIMER_INFO("FwPowerTimer", getName().c_str()); - outV->rowPow(0, *inV1, *inV0); - } -} - -void PowerLayer::backward(const UpdateCallback& callback) { - MatrixPtr inV0 = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - MatrixPtr outV = getOutputValue(); - MatrixPtr outG = getOutputGrad(); - - size_t batchSize = inV1->getHeight(); - size_t dataDim = inV1->getWidth(); - - { - REGISTER_TIMER_INFO("BwPowerTimer", getName().c_str()); - Matrix::resizeOrCreate(tmpMtx, batchSize, dataDim, false, useGpu_); - - if (inG0) { - tmpMtx->log2(*inV1); - tmpMtx->dotMul(*tmpMtx, *outV); - - // inG0 += outG .* (log(inV1) * outV) - inG0->rowDotMul(0, *outG, *tmpMtx); - } - - if (inG1) { - // tmp = (outV / inV1) * inV0 - tmpMtx->dotDiv(*outV, *inV1); - tmpMtx->rowScale(0, *tmpMtx, *inV0); - - inG1->addDotMul(*outG, *tmpMtx, 1, 1); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/PriorBox.cpp b/paddle/gserver/layers/PriorBox.cpp deleted file mode 100644 index 39d2c2d737fa90737635efdb209610e156c8662f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/PriorBox.cpp +++ /dev/null @@ -1,159 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/BaseMatrix.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * @brief A layer for generating priorbox locations and variances. - * - Input: Two and only two input layer are accepted. The input layer must be - * be a data output layer and a convolution output layer. - * - Output: The priorbox locations and variances of the input data. - * Reference: - * Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, - * Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector - */ - -class PriorBoxLayer : public Layer { - public: // NOLINT - explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {} - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override {} - - protected: // NOLINT - int numPriors_; - std::vector minSize_; - std::vector maxSize_; - std::vector aspectRatio_; - std::vector variance_; - MatrixPtr buffer_; -}; - -REGISTER_LAYER(priorbox, PriorBoxLayer); - -bool PriorBoxLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - auto pbConf = config_.inputs(0).priorbox_conf(); - std::vector tmp; - aspectRatio_.push_back(1.); - std::copy(pbConf.min_size().begin(), - pbConf.min_size().end(), - std::back_inserter(minSize_)); - std::copy(pbConf.max_size().begin(), - pbConf.max_size().end(), - std::back_inserter(maxSize_)); - std::copy(pbConf.variance().begin(), - pbConf.variance().end(), - std::back_inserter(variance_)); - std::copy(pbConf.aspect_ratio().begin(), - pbConf.aspect_ratio().end(), - std::back_inserter(tmp)); - - if (maxSize_.size() > 0) CHECK_EQ(minSize_.size(), maxSize_.size()); - - // flip aspect ratios - for (unsigned index = 0; index < tmp.size(); index++) { - real ar = tmp[index]; - if (fabs(ar - 1.) < 1e-6) continue; - aspectRatio_.push_back(ar); - aspectRatio_.push_back(1. / ar); - } - - numPriors_ = aspectRatio_.size() * minSize_.size() + maxSize_.size(); - - return true; -} - -void PriorBoxLayer::forward(PassType passType) { - Layer::forward(passType); - auto input = getInput(0); - int layerWidth = input.getFrameWidth(); - int layerHeight = input.getFrameHeight(); - - auto image = getInput(1); - int imageWidth = image.getFrameWidth(); - int imageHeight = image.getFrameHeight(); - - real stepW = static_cast(imageWidth) / layerWidth; - real stepH = static_cast(imageHeight) / layerHeight; - int dim = layerHeight * layerWidth * numPriors_ * 4; - reserveOutput(1, dim * 2); - // use a cpu buffer to compute - Matrix::resizeOrCreate(buffer_, 1, dim * 2, false, false); - auto* tmpPtr = buffer_->getData(); - - int idx = 0; - for (int h = 0; h < layerHeight; ++h) { - for (int w = 0; w < layerWidth; ++w) { - real centerX = (w + 0.5) * stepW; - real centerY = (h + 0.5) * stepH; - for (size_t s = 0; s < minSize_.size(); s++) { - real minSize = minSize_[s]; - real boxWidth = minSize; - real boxHeight = minSize; - - // first prior: aspect_ratio == 1.0, compatible to old logic - tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; - tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; - tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; - tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; - // set the variance. - for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; - - if (maxSize_.size() > 0) { - // square prior with size sqrt(minSize * maxSize) - real maxSize = maxSize_[s]; - boxWidth = boxHeight = sqrt(minSize * maxSize); - tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; - tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; - tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; - tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; - // set the variance. - for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; - } - - // priors with different aspect ratios - for (size_t r = 0; r < aspectRatio_.size(); r++) { - real ar = aspectRatio_[r]; - if (fabs(ar - 1.0) < 1e-6) { - continue; - } - boxWidth = minSize * sqrt(ar); - boxHeight = minSize / sqrt(ar); - tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; - tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; - tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; - tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; - // set the variance. - for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; - } - } - } - } - - // clip the prior's coordidate such that it is within [0, 1] - for (int d = 0; d < dim * 2; ++d) - if ((d % 8) < 4) - tmpPtr[d] = std::min(std::max(tmpPtr[d], (real)0.), (real)1.); - MatrixPtr outV = getOutputValue(); - outV->copyFrom(buffer_->data_, dim * 2); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/Projection.h b/paddle/gserver/layers/Projection.h deleted file mode 100644 index 88a41355cfce711e1e9522655058d0f1198e4e76..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/Projection.h +++ /dev/null @@ -1,140 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "ModelConfig.pb.h" -#include "paddle/parameter/Parameter.h" - -namespace paddle { - -// Macro for registering a projection type -// Example: REGISTER_LAYER(fc, FullMatrixProjection); -#define REGISTER_PROJECTION(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name([]() { \ - Projection::registrar_.registerClass<__class_name>(#__type_name); \ - }) - -#define REGISTER_PROJECTION_CREATE_FUNC(__type_name, createFunction) \ - static InitFunction __reg_type_##__type_name([]() { \ - Projection::registrar_.registerClass(#__type_name, createFunction); \ - }) - -/** - * A projection takes one Argument as input, calculate the result and add it - * to output Argument. - */ -class Projection { - public: - static Projection* create(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu); - - Projection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGpu) - : config_(config), parameter_(parameter), useGpu_(useGpu) {} - - virtual ~Projection() {} - - const std::string& getName() const { return config_.name(); } - - /// Register a projection - static ClassRegistrar - registrar_; - - /** - * Forward propagation. If backward() will be called, in and out must be kept - * valid until then. - * @param in input of projection - * @param out output of projection - * @param passType PASS_TRAIN of PASS_TEST - */ - void forward(const Argument* in, const Argument* out, PassType passType) { - in_ = in; - out_ = out; - passType_ = passType; - forward(); - } - - virtual void prefetch(const Argument* in) {} - virtual void forward() = 0; - virtual void backward(const UpdateCallback& callback) = 0; - - /** - * See comment in Layer.h for the function with the same name. - */ - virtual void resetState() {} - - /** - * Set layer state. - */ - virtual void setState(LayerStatePtr state) {} - - /** - * Get layer state. A copy of internal state is returned. - */ - virtual LayerStatePtr getState() { return nullptr; } - - /** - * init forward_ and backward_ functions - */ - virtual bool init() { return true; } - - /** - * Get output size of projection. - */ - size_t getOutputSize() const { return config_.output_size(); } - - protected: - /** - * Create layer function. Function is called in forward or backward. - * \param function, Layer::forward_ or Layer::backward_ - * \param name, function name - * \param config, initialization configuration for the function - */ - void createFunction(std::vector>& function, - const std::string& name, - const FuncConfig& config) { - if (useGpu_) { - function.emplace_back( - FunctionBase::funcRegistrar_.createByType(name + "-GPU")); - } else { - function.emplace_back( - FunctionBase::funcRegistrar_.createByType(name + "-CPU")); - } - auto& func = function.back(); - func->init(config); - } - - protected: - /// Config of projection - ProjectionConfig config_; - /// Parameter of projection - ParameterPtr parameter_; - bool useGpu_; - - /// Store `in` passed to forward() - const Argument* in_; - /// Store `out` passed to forward() - const Argument* out_; - /// Store `passType` passed to forward() - PassType passType_; - /// Layer forward function - std::vector> forward_; - /// Layer backward function - std::vector> backward_; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/RecurrentLayer.h b/paddle/gserver/layers/RecurrentLayer.h deleted file mode 100644 index 94e633e65777aad540738ea67ea1b4e03dd75954..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/RecurrentLayer.h +++ /dev/null @@ -1,130 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ -#pragma once -#include -#include "Layer.h" -#include "SequenceToBatch.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief RecurrentLayer takes 1 input layer. The output size is the same with - * input layer. - * For each sequence [start, end] it performs the following computation: - * \f[ - * out_{i} = act(in_{i}) \ \ \text{for} \ i = start \\ - * out_{i} = act(in_{i} + out_{i-1} * W) \ \ \text{for} \ start < i <= end - * - * \f] - * If reversed is true, the order is reversed: - * \f[ - * out_{i} = act(in_{i}) \ \ \text{for} \ i = end \\ - * out_{i} = act(in_{i} + out_{i+1} * W) \ \ \text{for} \ start <= i < end - * \f] - * There are two methods to calculate rnn. One way is to compute rnn one - * sequence by one sequence. The other way is to reorganize the input - * into batches, then compute rnn one batch by one batch. Users can select - * them by rnn_use_batch flag. - */ - -class RecurrentLayer : public Layer { - public: - explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - - void backward(const UpdateCallback& callback) override; - - void resetState() override; - - void setState(LayerStatePtr state) override; - - LayerStatePtr getState() override; - - protected: - /** - * @brief If user do not set --rnn_use_batch=true, it will - * compute rnn forward one sequence by one sequence in default. - * @param batchSize Total words number of all samples in this batch. - * @param numSequences The sample number. - * @param starts Each start position of each samples. - */ - void forwardSequence(int batchSize, size_t numSequences, const int* starts); - /** - * @brief Compute rnn forward by one sequence. - * @param start The start position of this sequence (or sample). - * @param length The length of this sequence (or sample), namely the words - * number of this sequence. - */ - void forwardOneSequence(int start, int length); - /** - * @brief Compute rnn backward one sequence by onesequence. - * @param batchSize Total words number of all samples in this batch. - * @param numSequences The sample number. - * @param starts Each start position of each samples. - */ - void backwardSequence(int batchSize, size_t numSequences, const int* starts); - /** - * @brief Compute rnn backward by one sequence. - * @param start The start position of this sequence (or sample). - * @param length The length of this sequence (or sample), namely the words - * number of this sequence. - */ - void backwardOneSequence(int start, int length); - - /** - * @brief Reorganize input into batches and compute rnn forward batch - * by batch. It will convert batch shape to sequence after finishing forward. - * The batch info can refer to SequenceToBatch class. - * @param batchSize Total words number of all samples in this batch. - * @param numSequences The sample number. - * @param starts Each start position of each samples. - */ - virtual void forwardBatch(int batchSize, - size_t numSequences, - const int* starts); - - /** - * @brief Reorganize input into batches and compute rnn forward batch - * by batch. - * @param batchSize Total words number of all samples in this batch. - * @param numSequences The sample number. - * @param starts Each start position of each samples. - */ - virtual void backwardBatch(int batchSize, - size_t numSequences, - const int* starts); - - protected: - std::unique_ptr weight_; - std::unique_ptr bias_; - - /// frameOutput_[i] is used to hold the i-th sample of output_ - std::vector frameOutput_; - MatrixPtr prevOutput_; - /// Whether compute rnn by reverse. - bool reversed_; - /// If compute batch by batch, batchValue_ will be used to save the - /// reorganized input value. - std::unique_ptr batchValue_; - /// If compute batch by batch, batchGrad_ will be used to save the - /// gradient with respect to reorganized input value. - std::unique_ptr batchGrad_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/RecurrentLayerGroup.cpp b/paddle/gserver/layers/RecurrentLayerGroup.cpp deleted file mode 100644 index 6694e8f2996fdd2c98da1507e5fb3b90b271c850..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/RecurrentLayerGroup.cpp +++ /dev/null @@ -1,95 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/gserver/layers/Layer.h" - -#include "paddle/gserver/gradientmachines/RecurrentGradientMachine.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * Recurrent layer group is a group of layers, which forward/backward one frame - * after previous frame forward/backward through all layers in layer group. - * It's automatically added by config_parser if some layers are defined - * between RecurrentLayerGroupBegin and RecurrentLayerGroupEnd. - */ -class RecurrentLayerGroup : public Layer { - public: - explicit RecurrentLayerGroup(const LayerConfig& config) : Layer(config) {} - - void initSubNetwork(NeuralNetwork* rootNetwork, - const ModelConfig& config, - const std::vector& parameterTypes, - bool useGpu) override; - - void forward(PassType passType) override { - REGISTER_TIMER_INFO("RecurrentGroupFwTime", getName().c_str()); - const std::vector inArgs; - std::vector outArgs; - network_->forward(inArgs, &outArgs, passType); - } - void backward(const UpdateCallback& callback) override { - REGISTER_TIMER_INFO("RecurrentGroupBwTime", getName().c_str()); - network_->backward(nullptr); - - for (auto& para : parameters_) { - para->incUpdate(callback); - } - } - - /** - * @see Layer.accessSubNetwork - */ - void accessSubNetwork( - const std::function& callback) override { - callback(*network_); - } - - private: - std::unique_ptr network_; -}; - -REGISTER_LAYER(recurrent_layer_group, RecurrentLayerGroup); - -void RecurrentLayerGroup::initSubNetwork( - NeuralNetwork* rootNetwork, - const ModelConfig& config, - const std::vector& parameterTypes, - bool useGpu) { - setNeedGradient(true); - - network_.reset(new RecurrentGradientMachine(config_.name(), rootNetwork)); - ParamInitCallback cb = [rootNetwork](int paramId, Parameter* para) { - para->enableSharedType( - PARAMETER_VALUE, - rootNetwork->getParameters()[paramId]->getBuf(PARAMETER_VALUE), - rootNetwork->getParameters()[paramId]->getMat(PARAMETER_VALUE)); - para->enableSharedType( - PARAMETER_GRADIENT, - rootNetwork->getParameters()[paramId]->getBuf(PARAMETER_GRADIENT), - rootNetwork->getParameters()[paramId]->getMat(PARAMETER_GRADIENT)); - }; - network_->init(config, cb, parameterTypes, useGpu); - - for (auto paramId : network_->getParameterIds()) { - ParameterPtr parameter = rootNetwork->getParameters()[paramId]; - parameter->incShared(); - CHECK_EQ(parameter->getDeviceId(), getDeviceId()); - parameters_.push_back(parameter); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ResizeLayer.cpp b/paddle/gserver/layers/ResizeLayer.cpp deleted file mode 100644 index d4ae9945934a40719d253d4b53915530423448af..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ResizeLayer.cpp +++ /dev/null @@ -1,79 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/BaseMatrix.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * @brief A layer for resizing a minibatch matrix h*w to h'*w' - * @note - * origin matrix height * width) - * resize matrix: (height * width / size) * size - */ -class ResizeLayer : public Layer { - public: - explicit ResizeLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - - void backward(const UpdateCallback& callback) override; -}; - -REGISTER_LAYER(resize, ResizeLayer); - -bool ResizeLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - if (!Layer::init(layerMap, parameterMap)) return false; - CHECK_EQ(1U, inputLayers_.size()); - - setNeedSequenceInfo(false); - return true; -} - -void ResizeLayer::forward(PassType passType) { - Layer::forward(passType); - const Argument& input = getInput(0); - size_t height = input.value->getHeight(); - size_t width = input.value->getWidth(); - CHECK_EQ((height * width) % getSize(), 0UL); - - reserveOutput(height * width / getSize(), getSize()); - MatrixPtr tmp = - Matrix::create(output_.value->getData(), height, width, false, useGpu_); - tmp->assign(*input.value); -} - -void ResizeLayer::backward(const UpdateCallback& callback) { - const Argument& input = getInput(0); - size_t height = input.value->getHeight(); - size_t width = input.value->getWidth(); - - if (!input.grad) { - return; - } - - MatrixPtr tmp = Matrix::create(input.grad->getData(), - height * width / getSize(), - getSize(), - false, - useGpu_); - tmp->add(*output_.grad); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/RotateLayer.h b/paddle/gserver/layers/RotateLayer.h deleted file mode 100644 index 7ecbff20167dd95f782f2d61dc34697ab3273934..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/RotateLayer.h +++ /dev/null @@ -1,51 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * A layer for rotating a multi-channel feature map (M x N x C) in the spatial - * domain - * The rotation is 90 degrees in clock-wise for each channel - * \f[ - * y(j,i,:) = x(M-i-1,j,:) - * \f] - * where \f$x\f$ is (M x N x C) input, and \f$y\f$ is (N x M x C) output. - * - * The config file api is rotate_layer - * - */ - -class RotateLayer : public Layer { - public: - explicit RotateLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); - - void forward(PassType passType); - void backward(const UpdateCallback& callback = nullptr); - - private: - int batchSize_; - int size_; - int height_; - int width_; - int channels_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/RowConvLayer.cpp b/paddle/gserver/layers/RowConvLayer.cpp deleted file mode 100644 index 63b499e486fd24b5f816ee0e897b040ee5007581..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/RowConvLayer.cpp +++ /dev/null @@ -1,106 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "RowConvLayer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(row_conv, RowConvLayer); - -bool RowConvLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - contexLength_ = config_.inputs(0).row_conv_conf().context_length(); - - CHECK_EQ(inputLayers_.size(), 1UL); - weight_.reset(new Weight(contexLength_, getSize(), parameters_[0])); - createFunction(forward_, "RowConv", FuncConfig()); - createFunction(backward_, "RowConvGrad", FuncConfig()); - - return true; -} - -void RowConvLayer::forward(PassType passType) { - Layer::forward(passType); - MatrixPtr input = getInputValue(0); - size_t height = input->getHeight(); - size_t width = input->getWidth(); - CHECK_EQ(width, getSize()); - resetOutput(height, width); - - const auto startPos = getInput(0).sequenceStartPositions->getVector(useGpu_); - MatrixPtr w = weight_->getW(); - wDims_ = TensorShape({w->getHeight(), w->getWidth()}); - - MatrixPtr outV = getOutputValue(); - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), *startPos); - inputs.addArg(*w, wDims_); - outputs.addArg(*getOutputValue(), *startPos, ADD_TO); - - { - REGISTER_TIMER_INFO("RowConvForward", getName().c_str()); - forward_[0]->calc(inputs, outputs); - } - - /* activation */ { - REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); - forwardActivation(); - } -} - -void RowConvLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { - REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); - backwardActivation(); - } - - const auto startPos = getInput(0).sequenceStartPositions->getVector(useGpu_); - - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getOutputGrad(), *startPos); - inputs.addArg(*getInputValue(0), *startPos); - inputs.addArg(*weight_->getW(), wDims_); - - MatrixPtr inGrad = getInputGrad(0); - MatrixPtr wGrad = weight_->getWGrad(); - size_t h = getInputValue(0)->getHeight(); - size_t w = getInputValue(0)->getWidth(); - outputs.addArg( - inGrad ? (*inGrad) : *(Matrix::create(nullptr, h, w, false, useGpu_)), - *startPos, - ADD_TO); - outputs.addArg( - wGrad ? (*wGrad) - : *(Matrix::create(nullptr, contexLength_, w, false, useGpu_)), - wDims_, - ADD_TO); - - { - REGISTER_TIMER_INFO("RowConvBackward", getName().c_str()); - backward_[0]->calc(inputs, outputs); - } - - { - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); - weight_->getParameterPtr()->incUpdate(callback); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ScaleSubRegionLayer.cpp b/paddle/gserver/layers/ScaleSubRegionLayer.cpp deleted file mode 100644 index 68a0ff735844679df1393473355f54ee616c09bd..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ScaleSubRegionLayer.cpp +++ /dev/null @@ -1,78 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ScaleSubRegionLayer.h" -#include "paddle/utils/Stat.h" -namespace paddle { - -REGISTER_LAYER(scale_sub_region, ScaleSubRegionLayer); - -bool ScaleSubRegionLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - CHECK_EQ(static_cast(inputLayers_.size()), 2); - auto& conf = config_.inputs(0).scale_sub_region_conf(); - value_ = conf.value(); - - createFunction(forward_, "ScaleSubRegion", FuncConfig().set("value", value_)); - createFunction( - backward_, "ScaleSubRegionGrad", FuncConfig().set("value", value_)); - - return true; -} - -void ScaleSubRegionLayer::forward(PassType passType) { - Layer::forward(passType); - auto in0 = getInput(0); - imgH_ = in0.getFrameHeight(); - imgW_ = in0.getFrameWidth(); - if (imgH_ == 0 || imgW_ == 0) { - auto& conf = config_.inputs(0).scale_sub_region_conf(); - imgH_ = conf.image_conf().img_size_y(); - imgW_ = conf.image_conf().img_size(); - } - MatrixPtr imgV = in0.value; - size_t batchSize = imgV->getHeight(); - size_t spatialSize = imgH_ * imgW_; - channelsNum_ = imgV->getWidth() / spatialSize; - shape_ = TensorShape({batchSize, channelsNum_, imgH_, imgW_}); - - resetOutput(batchSize, imgV->getWidth()); - auto& out = getOutput(); - out.setFrameHeight(imgH_); - out.setFrameWidth(imgW_); - - MatrixPtr indicesV = getInputValue(1); - indicesShape_ = TensorShape({batchSize, 6}); - - REGISTER_TIMER_INFO("ScaleSubRegionForward", getName().c_str()); - BufferArgs inArgs; - BufferArgs outArgs; - inArgs.addArg(*imgV, shape_); - inArgs.addArg(*indicesV, indicesShape_); - outArgs.addArg(*out.value, shape_, ASSIGN_TO); - forward_[0]->calc(inArgs, outArgs); -} - -void ScaleSubRegionLayer::backward(const UpdateCallback& callback) { - REGISTER_TIMER_INFO("ScaleSubRegionBackward", getName().c_str()); - BufferArgs inArgs; - BufferArgs outArgs; - inArgs.addArg(*getOutputGrad(), shape_); - inArgs.addArg(*getInputValue(1), indicesShape_); - outArgs.addArg(*getInputGrad(0), shape_, ADD_TO); - backward_[0]->calc(inArgs, outArgs); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ScalingLayer.cpp b/paddle/gserver/layers/ScalingLayer.cpp deleted file mode 100644 index 15e07daebee194a789da52d37a192e031348300c..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ScalingLayer.cpp +++ /dev/null @@ -1,106 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for each row of a matrix, multiplying with a element of a vector, - * which is used in NEURAL TURING MACHINE. - * \f[ - * y.row[i] = w[i] * x.row[i] - * \f] - * where \f$x\f$ is (batchSize x dataDim) input, \f$w\f$ is - * (batchSize x 1) weight vector, and \f$y\f$ is (batchSize x dataDim) output. - * - * The config file api is scaling_layer. - */ - -class ScalingLayer : public Layer { - public: - explicit ScalingLayer(const LayerConfig& config) : Layer(config) {} - - ~ScalingLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(scaling, ScalingLayer); - -bool ScalingLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 2U); - - return true; -} - -void ScalingLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr weightV = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - - size_t batchSize = inV1->getHeight(); - size_t dataDim = inV1->getWidth(); - - CHECK_EQ(dataDim, getSize()); - CHECK_EQ(weightV->getWidth(), 1U); - CHECK_EQ(weightV->getHeight(), batchSize); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - resetOutput(batchSize, dataDim); - } - - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwScalingTimer", getName().c_str()); - // outV += inV1 * weight - outV->addRowScale(0, *inV1, *weightV); - } -} - -void ScalingLayer::backward(const UpdateCallback& callback) { - MatrixPtr weightV = getInputValue(0); - MatrixPtr inV1 = getInputValue(1); - MatrixPtr inG0 = getInputGrad(0); - MatrixPtr inG1 = getInputGrad(1); - MatrixPtr outG = getOutputGrad(); - - { - REGISTER_TIMER_INFO("BwScalingTimer", getName().c_str()); - - if (inG0) { - // inG0 += outG .* inV1 - inG0->rowDotMul(0, *outG, *inV1); - } - - if (inG1) { - // inG1 += outG * weight; - inG1->addRowScale(0, *outG, *weightV); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SelectiveFullyConnectedLayer.cpp b/paddle/gserver/layers/SelectiveFullyConnectedLayer.cpp deleted file mode 100644 index 43c98993f3f6f74c034c59176378c3ea97a9c19b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SelectiveFullyConnectedLayer.cpp +++ /dev/null @@ -1,336 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SelectiveFullyConnectedLayer.h" -#include -#include -#include "paddle/math/SparseMatrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(selective_fc, SelectiveFullyConnectedLayer); - -bool SelectiveFullyConnectedLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - inputNum_ = inputLayers_.size(); - if (config_.has_selected_colums()) { - inputNum_ -= 1; - } - for (size_t i = 0; i < inputNum_; i++) { - size_t height = inputLayers_[i]->getSize(); - size_t width = getSize(); - // NOTE weight is transpoed - weights_.emplace_back(new Weight(width, height, parameters_[i])); - } - - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - - fullOutput_ = false; - - return true; -} - -void SelectiveFullyConnectedLayer::prefetch() {} - -void SelectiveFullyConnectedLayer::reserveOutput(size_t height, - size_t width, - size_t nnz) { - bool flag = (passType_ == PASS_TEST && - config_.selective_fc_pass_generation() && !fullOutput_); - SetDevice device(output_.deviceId); - if (flag) { - // output_.value is sparse matrix - if (dynamic_cast(output_.value.get()) || - dynamic_cast(output_.value.get())) { - output_.value = nullptr; - } - Matrix::resizeOrCreateSparseMatrix(output_.value, - height, - width, - nnz, - FLOAT_VALUE, - SPARSE_CSR, - /*trans=*/false, - /*useGpu=*/useGpu_); - output_.value->copyFrom(*selCols_); - interOutput_ = output_.value; - } else { - if (fullOutput_) { - // output_.value is dense matrix - if (dynamic_cast(output_.value.get()) || - dynamic_cast(output_.value.get())) { - output_.value = nullptr; - } - Matrix::resizeOrCreate(output_.value, - height, - width, - /*trans=*/false, - /*useGpu=*/useGpu_); - interOutput_ = output_.value; - } else { - // output_.value is dense matrix, but width = nnz /height - CHECK_EQ(nnz % height, 0U); - CHECK(nnz / height); - Matrix::resizeOrCreate(output_.value, - height, - nnz / height, - /*trans=*/false, - /*useGpu=*/useGpu_); - interOutput_ = Matrix::createSparseMatrix(output_.value->getData(), - selCols_->getRows(), - selCols_->getCols(), - height, - width, - nnz, - FLOAT_VALUE, - SPARSE_CSR, - /*trans=*/false, - /*useGpu=*/useGpu_); - } - } - interOutput_->zeroMem(); - - if (passType_ != PASS_TEST && needGradient()) { - CHECK_EQ(nnz % height, 0U) << "during training, each sample must have a " - "same number of selected columns."; - CHECK(nnz / height) - << "during training, " - "each sample must have at least one column selected."; - Matrix::resizeOrCreate(output_.grad, - height, - nnz / height, - /*trans=*/false, - /*useGpu=*/useGpu_); - output_.grad->zeroMem(); - } -} - -void SelectiveFullyConnectedLayer::forward(PassType passType) { - REGISTER_TIMER("selective_fc.forward"); - Layer::forward(passType); - - getSelectiveCols(); - size_t height = getInput(0).getBatchSize(); - size_t width = getSize(); - size_t nnz = height * width; - if (!fullOutput_) { - CHECK(selCols_); - CHECK(height == selCols_->getHeight()); - CHECK(width == selCols_->getWidth()); - nnz = selCols_->getElementCnt(); - } - - // Layer::ResetOutput(), here we set outV/outG as SparseMatrix manually - // this outV should be used as input of MaxIdLayer and softmax activation - reserveOutput(height, width, nnz); - - bool flag = true; - for (size_t i = 0; i < inputNum_; i++) { - MatrixPtr input = getInputValue(i); - MatrixPtr weight = weights_[i]->getW(); - size_t hsize = input->getHeight(); - size_t wsize = weight->getHeight(); - real scaleT = i == 0 ? real(0) : real(1); - - flag = nnz < (hsize * wsize) * config_.selective_fc_full_mul_ratio() && - !fullOutput_; - if (flag) { - // if the indecies are highly sparse, - // manully compute the multiplication of - // the input vector and the selected rows. - REGISTER_TIMER("selective.plain"); - interOutput_->mul(*input, *weight->getTranspose(), 1, scaleT); - } else { - // if the indecies is not sparse enough, - // use full mul instead - REGISTER_TIMER("selective.mul"); - if (fullOutput_) { - interOutput_->mul(*input, *weight->getTranspose(), 1, scaleT); - } else { - Matrix::resizeOrCreate(mmat_, - hsize, - wsize, - /*trans=*/false, - /*useGpu=*/useGpu_); - mmat_->mul(*input, *weight->getTranspose()); - interOutput_->add3(mmat_); - } - } - } - - if (biases_) { - interOutput_->addBias(*(biases_->getW()), 1); - } - - flag = (passType_ == PASS_TEST && config_.selective_fc_pass_generation() && - !fullOutput_); - if (flag) { - // during generation, output of this layer is a sparse csr matrix, - // which is probably the input of maxid layer - // if the model is trained with multi-class-cross-entroy-with-selfnorm, - // activiation of this layer should be exponential, not softmax. - - Argument arg; - arg.value = Matrix::create(interOutput_->getData(), - 1, - nnz, - /*trans=*/false, - /*useGpu=*/useGpu_); - //! TODO(yuyang18): Why we cannot invoke forwardActivation here? - activation_->forward(arg).check(); - } else /* train and test in train, not generating */ { - // during training, this layer output value is *Matrix*, which is input of - // eg. multi-class-cross-entropy - - // while training, every sample has a equal number of selected - // columns to be activated. - // note indices of multi-class-cross-entropy need to be remapped - // to this index. - // e.g. sample = [1,3,5] and 3 is gold, then label is 1 - - forwardActivation(); - } -} - -void SelectiveFullyConnectedLayer::backward(const UpdateCallback& callback) { - backwardActivation(); - MatrixPtr oGrad = getOutputGrad(); - if (!fullOutput_) { - interOutGrad_ = Matrix::createSparseMatrix(oGrad->getData(), - interOutput_->getRows(), - interOutput_->getCols(), - interOutput_->getHeight(), - interOutput_->getWidth(), - interOutput_->getElementCnt(), - FLOAT_VALUE, - SPARSE_CSR, - /*trans=*/false, - /*useGpu=*/useGpu_); - } else { - interOutGrad_ = Matrix::create(oGrad->getData(), - oGrad->getHeight(), - oGrad->getWidth(), - /*trans=*/false, - /*useGpu=*/useGpu_); - } - - if (biases_ && biases_->getWGrad()) { - REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); - biases_->getWGrad()->collectBias(*interOutGrad_, 1); - biases_->getParameterPtr()->incUpdate(callback); - } - - // backward is different from FullyConnectedLayer - // because the weight is transposed - for (size_t i = 0; i < inputNum_; i++) { - AsyncGpuBlock block; - MatrixPtr preGrad = getInputGrad(i); - if (preGrad) { - REGISTER_TIMER_INFO("BpMulTimer", getName().c_str()); - preGrad->mul(*interOutGrad_, *weights_[i]->getW(), 1, 1); - } - - MatrixPtr wGrad = weights_[i]->getWGrad(); - if (wGrad) { - REGISTER_TIMER_INFO("GradMulTimer", getName().c_str()); - MatrixPtr input = getInputValue(i); - wGrad->mul(*interOutGrad_->getTranspose(), *input, 1, 1); - } - - { - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); - weights_[i]->getParameterPtr()->incUpdate(callback); - } - } -} - -void paddle::SelectiveFullyConnectedLayer::fillSelectiveData( - const std::shared_ptr>>& candidates) { - if (candidates == nullptr) { - fillFullySelectiveData(); - return; - } - - size_t sampleNum = candidates->size(); - size_t outputWidth = getSize(); - size_t nnz = - std::accumulate(candidates->begin(), - candidates->end(), - 0UL, - [](size_t a, const std::pair& arr) { - return a + arr.second; - }); - - Matrix::resizeOrCreateSparseMatrix(this->cpuSelCols_, - sampleNum, - outputWidth, - nnz, - NO_VALUE, - SPARSE_CSR, - false, - false); - CHECK(this->cpuSelCols_ != nullptr); - CpuSparseMatrixPtr selCols = - std::dynamic_pointer_cast(cpuSelCols_); - int* rowOffsets = selCols->getRows(); - int* colIndices = selCols->getCols(); - - rowOffsets[0] = 0; - int idx = 0; - for (size_t i = 0; i < sampleNum; ++i) { - if ((*candidates)[i].second > 0) { - rowOffsets[i + 1] = rowOffsets[i] + (*candidates)[i].second; - for (size_t j = 0; j < (*candidates)[i].second; ++j) { - colIndices[idx] = (*candidates)[i].first[j]; - idx++; - } - } else { - rowOffsets[i + 1] = rowOffsets[i]; - } - } - - CHECK_EQ(static_cast(rowOffsets[sampleNum]), nnz); - if (!useGpu_) { - this->selCols_ = this->cpuSelCols_; - } else { - Matrix::resizeOrCreateSparseMatrix(this->selCols_, - sampleNum, - outputWidth, - nnz, - NO_VALUE, - SPARSE_CSR, - false, - true); - this->selCols_->copyFrom(*cpuSelCols_, HPPL_STREAM_1); - hl_stream_synchronize(HPPL_STREAM_1); - } - - fullOutput_ = false; -} - -void paddle::SelectiveFullyConnectedLayer::getSelectiveCols() { - if (config_.has_selected_colums()) { - this->selCols_ = inputLayers_[inputNum_]->getOutputValue(); - fullOutput_ = false; - } else if (!config_.selective_fc_pass_generation() || selCols_ == nullptr) { - this->fillFullySelectiveData(); - } // else selCols_ is initialized by fillSelectiveData -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SelectiveFullyConnectedLayer.h b/paddle/gserver/layers/SelectiveFullyConnectedLayer.h deleted file mode 100644 index 4b32ce8b162c2a8b1a6c34adc0885a7701f5f91e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SelectiveFullyConnectedLayer.h +++ /dev/null @@ -1,103 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * @brief The SelectiveFullyConnectedLayer class - * - * SelectiveFullyConnectedLayer differs from FullyConnectedLayer by that it - * requires an additional input to indicate several selected columns, and only - * compute the multiplications between the input matrices and the selected - * columns of the parameter matrices of this layer. If the selected columns is - * not specified, SelectiveFullyConnected layer acts exactly like - * FullyConnectedLayer. - * - * The config file api is selective_fc_layer. - */ -class SelectiveFullyConnectedLayer : public Layer { - protected: - WeightList weights_; - std::unique_ptr biases_; - - private: - /** - * Get selected columns each forward. - */ - void getSelectiveCols(); - - MatrixPtr mmat_; - /// cpuSelCols_ is a CpuSparseMatrix, used to save selected columns. - MatrixPtr cpuSelCols_; - /// CpuSparseMatrix or GpuSparseMatrix. In CPU mode, selCols_ points - /// to cpuSelCols_. - MatrixPtr selCols_; - size_t inputNum_; - - /// interOutput_ shared same memory with output_.value. - MatrixPtr interOutput_; - - /// if fullOutput_ is false, interOutGrad_ sparse matrix - MatrixPtr interOutGrad_; - - /// if true, means output_.value is the same as Fc Layer - bool fullOutput_; - - public: - explicit SelectiveFullyConnectedLayer(const LayerConfig& config) - : Layer(config), selCols_(nullptr) {} - - ~SelectiveFullyConnectedLayer() {} - void prefetch() override; - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - Weight& getWeight(int idx) { return *weights_[idx]; } - - /** - * @brief Resize the output matrix size. - * And reset value to zero - */ - void reserveOutput(size_t height, size_t width, size_t nnz); - - /** - * @brief Fill candidates to select several activations as output. - * @param candidates specifies several selected columns of the parameter - * matrices of this layer. - * Multiplications only between the input matrices and the selected columns - * are computed. - * If the candidates is a nullptr, selective fc layer acts exactly like the - * fully connected layer. - * @note CURRENTLY, THIS METHOD IS ONLY USED FOR BEAM SEARCH - */ - void fillSelectiveData( - const std::shared_ptr>>& candidates); - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - private: - /** - * @brief Make SelectiveFC act as FullyConnectedLayer - */ - void fillFullySelectiveData() { fullOutput_ = true; } -}; -} // namespace paddle diff --git a/paddle/gserver/layers/SequenceConcatLayer.cpp b/paddle/gserver/layers/SequenceConcatLayer.cpp deleted file mode 100644 index c84c3ce4f080cc19f4937f04585accb5b2b347f9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequenceConcatLayer.cpp +++ /dev/null @@ -1,189 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for concatenating the first sequence with the second sequence - * Input: two sequences each containing the same number of instances - * seq1 = [a1, a2, ..., an] - * seq2 = [b1, b2, ..., bn] - * Output: a concatenated sequence of the two input sequences - * out = [a1, b1, a2, b2, ..., an, bn] - */ - -class SequenceConcatLayer : public Layer { - protected: - std::unique_ptr biases_; - - public: - explicit SequenceConcatLayer(const LayerConfig& config) : Layer(config) {} - - ~SequenceConcatLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(seqconcat, SequenceConcatLayer); - -bool SequenceConcatLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - // sequene concatenation layer should have exactly 2 inputs - CHECK_EQ(2U, inputLayers_.size()); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - - setNeedSequenceInfo(false); - return true; -} - -void SequenceConcatLayer::forward(PassType passType) { - Layer::forward(passType); - - size_t dim = getSize(); - - const Argument& input1 = getInput(0); - size_t numSequences1 = input1.getNumSequences(); - auto startPositions1 = input1.sequenceStartPositions->getVector(false); - - const Argument& input2 = getInput(1); - size_t numSequences2 = input2.getNumSequences(); - auto startPositions2 = input2.sequenceStartPositions->getVector(false); - - CHECK_EQ(dim, input1.value->getWidth()); - CHECK_EQ(startPositions1->getData()[numSequences1], input1.getBatchSize()); - CHECK_EQ(numSequences1, startPositions1->getSize() - 1); - - CHECK_EQ(dim, input2.value->getWidth()); - CHECK_EQ(startPositions2->getData()[numSequences2], input2.getBatchSize()); - CHECK_EQ(numSequences2, startPositions2->getSize() - 1); - - CHECK_EQ(numSequences1, numSequences2); - - MatrixPtr inputValue1 = getInputValue(0); - MatrixPtr inputValue2 = getInputValue(1); - - // reset output - reserveOutput(inputValue1->getHeight() + inputValue2->getHeight(), dim); - - MatrixPtr outputValue = getOutputValue(); - - const int* starts1 = startPositions1->getData(); - const int* starts2 = startPositions2->getData(); - - { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceConcatLayerForward", getName().c_str()); - - size_t offset = 0; - size_t leftNumIns = 0; - size_t rightNumIns = 0; - for (size_t seqId = 0; seqId < numSequences1; ++seqId) { - leftNumIns = starts1[seqId + 1] - starts1[seqId]; - outputValue->subMatrix(offset, leftNumIns) - ->assign(*(inputValue1->subMatrix(starts1[seqId], leftNumIns))); - offset += leftNumIns; - - rightNumIns = starts2[seqId + 1] - starts2[seqId]; - outputValue->subMatrix(offset, rightNumIns) - ->assign(*(inputValue2->subMatrix(starts2[seqId], rightNumIns))); - offset += rightNumIns; - } - - // modify the sequenceStartPositions - ICpuGpuVector::resizeOrCreate( - output_.sequenceStartPositions, numSequences1 + 1, false); - - int* tgtBuf = output_.sequenceStartPositions->getMutableData(false); - - for (size_t seqId = 0; seqId < numSequences1 + 1; ++seqId) { - tgtBuf[seqId] = starts1[seqId] + starts2[seqId]; - } - } - - if (biases_.get() != NULL) { - MatrixPtr outV = getOutputValue(); - outV->addBias(*(biases_->getW()), 1); - } - - /* activation */ - forwardActivation(); -} - -void SequenceConcatLayer::backward(const UpdateCallback& callback) { - /* activation */ - backwardActivation(); - - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - // Increasing the number of gradient - biases_->getParameterPtr()->incUpdate(callback); - } - - MatrixPtr inputGrad1 = getInputGrad(0); - MatrixPtr inputGrad2 = getInputGrad(1); - MatrixPtr outputGrad = getOutputGrad(); - auto startPositions1 = getInput(0).sequenceStartPositions->getVector(false); - auto startPositions2 = getInput(1).sequenceStartPositions->getVector(false); - - size_t numSequences1 = startPositions1->getSize() - 1; - size_t numSequences2 = startPositions2->getSize() - 1; - - CHECK_EQ(numSequences1, numSequences2); - - const int* starts1 = startPositions1->getData(); - const int* starts2 = startPositions2->getData(); - - { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceConcatLayerBackward", getName().c_str()); - - size_t offset = 0; - size_t leftNumIns = 0; - size_t rightNumIns = 0; - for (size_t seqId = 0; seqId < numSequences1; ++seqId) { - leftNumIns = starts1[seqId + 1] - starts1[seqId]; - if (inputGrad1) { - inputGrad1->subMatrix(starts1[seqId], leftNumIns) - ->add(*(outputGrad->subMatrix(offset, leftNumIns))); - } - offset += leftNumIns; - - rightNumIns = starts2[seqId + 1] - starts2[seqId]; - if (inputGrad2) { - inputGrad2->subMatrix(starts2[seqId], rightNumIns) - ->add(*(outputGrad->subMatrix(offset, rightNumIns))); - } - offset += rightNumIns; - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SequenceLastInstanceLayer.cpp b/paddle/gserver/layers/SequenceLastInstanceLayer.cpp deleted file mode 100644 index 28d0a9296d4accd4152e886ccae12a776fdb8f7f..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequenceLastInstanceLayer.cpp +++ /dev/null @@ -1,118 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Logging.h" - -#include "SequencePoolLayer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for extracting the last instance of the input sequence. - * Input: a sequence - * If SequenceLevel = kNonseq: - * Output: a sequence containing only the last instance of the input sequence - * If stride_ > 0: - * Output: a shorten sequence. Stride is the step size by which we slide a - * window upon the input sequence, and getting last instance - * operation is then applied to each interval independently. - * If SequenceLevel = kSeq: - * Check input sequence must has sub-sequence - * Output: a sequence containing only the last instance of each sub-sequence - * of the input sequence - * - * The config file api is last_seq and first_seq. - */ - -class SequenceLastInstanceLayer : public SequencePoolLayer { - protected: - MatrixPtr tmpSrc_; - MatrixPtr tmpDest_; - std::vector instanceIds_; - - public: - explicit SequenceLastInstanceLayer(const LayerConfig& config) - : SequencePoolLayer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer); - -bool SequenceLastInstanceLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - SequencePoolLayer::init(layerMap, parameterMap); - reversed_ = config_.select_first(); - - tmpSrc_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); - tmpDest_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); - - return true; -} - -void SequenceLastInstanceLayer::forward(PassType passType) { - SequencePoolLayer::forward(passType); - - auto starts = startPositions_->getData(false); - MatrixPtr inputValue = getInputValue(0); - MatrixPtr outputValue = getOutputValue(); - - { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceLastInstanceLayerForward", getName().c_str()); - - instanceIds_.clear(); - for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) { - int insId = reversed_ ? starts[seqId] : starts[seqId + 1] - 1; - instanceIds_.push_back(insId); - - outputValue->subMatrix(seqId, 1, tmpDest_) - ->assign(*(inputValue->subMatrix(insId, 1, tmpSrc_))); - } - } - - if (biases_.get() != NULL) { - outputValue->addBias(*(biases_->getW()), 1); - } - - /* activation, should set to 'linear' in most cases */ - forwardActivation(); -} - -void SequenceLastInstanceLayer::backward(const UpdateCallback& callback) { - SequencePoolLayer::backward(callback); - - MatrixPtr inputGrad = getInputGrad(0); - MatrixPtr outputGrad = getOutputGrad(); - - if (inputGrad) { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceLastInstanceLayerBackward", getName().c_str()); - - for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) { - inputGrad->subMatrix(instanceIds_[seqId], 1, tmpDest_) - ->add(*(outputGrad->subMatrix(seqId, 1, tmpSrc_))); - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SequencePoolLayer.cpp b/paddle/gserver/layers/SequencePoolLayer.cpp deleted file mode 100644 index 650ab425d1fcca56d8862200f37dd5bb36a67240..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequencePoolLayer.cpp +++ /dev/null @@ -1,93 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SequencePoolLayer.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -bool SequencePoolLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - // seqlastins/max/average layer should have exactly 1 input - CHECK_EQ(1U, inputLayers_.size()); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - // transform to which sequence type - if (config_.trans_type() == "non-seq") { - type_ = kNonSeq; - } else if (config_.trans_type() == "seq") { - type_ = kSeq; - } else { - LOG(FATAL) << "Unknown trans_type: " << config_.trans_type(); - } - stride_ = config_.seq_pool_stride(); - setNeedSequenceInfo(false); - return true; -} - -void SequencePoolLayer::forward(PassType passType) { - Layer::forward(passType); - - const Argument& input = getInput(0); - CHECK(input.hasSeq() || input.hasSubseq()) - << "Input should be a sequence or subsequence for layer " << getName(); - - newBatchSize_ = type_ ? input.getNumSubSequences() : input.getNumSequences(); - size_t dim = getSize(); - // check - CHECK_EQ(dim, input.value->getWidth()); - startPositions_ = - type_ ? input.subSequenceStartPositions : input.sequenceStartPositions; - auto starts = startPositions_->getVector(false); - CHECK_EQ(starts->getData()[newBatchSize_], input.getBatchSize()); - CHECK_EQ(newBatchSize_, starts->getSize() - 1); - - /* If type_ = kNonSeq, both seq has or not has sub-seq degrade to a non-seq, - * thus, in this case, output_ has no sequenceStartPositions. - * If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this - * case, we should compute the new sequenceStartPositions. - */ - if (type_) { - CHECK(input.subSequenceStartPositions) - << "when trans_type = seq, input must hasSubseq"; - output_.degradeSequence(input); - } - if (stride_ > 0) { - CHECK_EQ(input.hasSubseq(), 0UL) - << "sequence stride pooling is invalid for hasSubseq now"; - output_.poolSequenceWithStride(input, stride_, &startPositions_, reversed_); - newBatchSize_ = startPositions_->getSize() - 1; - } - - resetOutput(newBatchSize_, dim); -} - -void SequencePoolLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { backwardActivation(); } - - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - // Increasing the number of gradient - biases_->getParameterPtr()->incUpdate(callback); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SequencePoolLayer.h b/paddle/gserver/layers/SequencePoolLayer.h deleted file mode 100644 index 01183060afd58376bb718dda64d8106cce4899f9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequencePoolLayer.h +++ /dev/null @@ -1,64 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * A base layer for SequenceLastInstanceLayer/AverageLayer/MaxLayer. - * - * Input: one or more sequences. Each sequence contains some instances. - * If SequenceLevel = kNonSeq: - * Output: output size is the number of input sequences (NOT input instances) - * output[i] = seqlastin/average/max_{for each instance in this - * sequence}{input[i]} - * If stride_ > 0: - * Check input sequence must not have sub-sequence - * Output: a shorten sequence. Stride is the step size by which we slide - * a window upon the input sequence, and the pooling operation - * is then applied to each interval independently. - * If SequenceLevel = kSeq: - * Check input sequence must has sub-sequence - * Output: output size is the number of input sub-sequences - * output[i] = seqlastin/average/max_{for each instance in this - * sub-sequence}{input[i]} - * - * The config file api is pooling_layer. - */ - -class SequencePoolLayer : public Layer { - protected: - int type_; - std::unique_ptr biases_; - enum SequenceLevel { kNonSeq = 0, kSeq = 1 }; - size_t newBatchSize_; - ICpuGpuVectorPtr startPositions_; - int stride_; - // Whether the input sequence is reversed or not. - bool reversed_ = false; - - public: - explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/SequenceReshapeLayer.cpp b/paddle/gserver/layers/SequenceReshapeLayer.cpp deleted file mode 100644 index 319310af8c4ac3bdefd814ad05b7fde6070f2340..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequenceReshapeLayer.cpp +++ /dev/null @@ -1,157 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for reshaping the sequence. Assume the input sequence has - * T instances, the dimension of each instance is M, and the input - * reshape_dim is N, then the output sequence has T*M/N instances, - * the dimension of each instance is N. - * - * Note that T*M/N must be an integer. - */ - -class SequenceReshapeLayer : public Layer { - protected: - std::unique_ptr biases_; - - MatrixPtr reshapedOutputGrad; - - public: - explicit SequenceReshapeLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(seqreshape, SequenceReshapeLayer); - -bool SequenceReshapeLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - CHECK_EQ(1U, inputLayers_.size()); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - setNeedSequenceInfo(false); - return true; -} - -void SequenceReshapeLayer::forward(PassType passType) { - Layer::forward(passType); - - const Argument& input = getInput(0); - - size_t inDim = input.value->getWidth(); - size_t outDim = getSize(); - - size_t numSequences = input.getNumSequences(); - - // by default, we assume each instance as a sequence - IVectorPtr seqStarts; - IVector::resizeOrCreate(seqStarts, input.getBatchSize() + 1, false); - int* startsData = seqStarts->getData(); - for (int i = 0; i < input.getBatchSize() + 1; i++) { - startsData[i] = i; - } - const int* starts = startsData; - - // if there is sequence, then use start positions - if (input.sequenceStartPositions) { - auto startPositions = input.sequenceStartPositions->getVector(false); - starts = startPositions->getData(); - CHECK_EQ(starts[numSequences], input.getBatchSize()); - CHECK_EQ(numSequences, startPositions->getSize() - 1); - } - - for (size_t seqID = 0; seqID < numSequences; seqID++) { - size_t inNumIns = starts[seqID + 1] - starts[seqID]; - size_t outNumIns = inNumIns * inDim / outDim; - CHECK_EQ(outNumIns * outDim, inNumIns * inDim); - } - - MatrixPtr inputValue = getInputValue(0); - - // reset output - reserveOutput(inputValue->getHeight() * inDim / outDim, outDim); - MatrixPtr outputValue = getOutputValue(); - - { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceReshapeLayerForward", getName().c_str()); - - outputValue->copyFrom(*inputValue); - - // modify the sequenceStartPositions - ICpuGpuVector::resizeOrCreate( - output_.sequenceStartPositions, numSequences + 1, false); - - int* tgtBuf = output_.sequenceStartPositions->getMutableData(false); - - for (size_t seqId = 0; seqId < numSequences + 1; ++seqId) { - tgtBuf[seqId] = starts[seqId] * inDim / outDim; - } - } - - if (biases_.get() != NULL) { - MatrixPtr outV = getOutputValue(); - outV->addBias(*(biases_->getW()), 1); - } - - /* activation */ - forwardActivation(); -} - -void SequenceReshapeLayer::backward(const UpdateCallback& callback) { - /* activation */ - backwardActivation(); - - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - // Increasing the number of gradient - biases_->getParameterPtr()->incUpdate(callback); - } - - MatrixPtr inputGrad = getInputGrad(0); - MatrixPtr outputGrad = getOutputGrad(); - - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SequenceReshapeLayerBackward", getName().c_str()); - - if (inputGrad) { - Matrix::resizeOrCreate(reshapedOutputGrad, - inputGrad->getHeight(), - inputGrad->getWidth(), - false, - useGpu_); - reshapedOutputGrad->copyFrom(*outputGrad); - inputGrad->add(*reshapedOutputGrad); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SequenceSliceLayer.cpp b/paddle/gserver/layers/SequenceSliceLayer.cpp deleted file mode 100644 index a6d810b583aab6e44faa583795686f06e17beeb9..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequenceSliceLayer.cpp +++ /dev/null @@ -1,224 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -class SequenceSliceLayer : public Layer { - public: - explicit SequenceSliceLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - private: - /* - * TODO(caoying) - * In PaddePaddle, currently all matrices are real number types, - * but the second and the (optional) third input which are some - * selected indices of the give sequence to trim the sequence, are actually - * filled with int types so that storing int types information in real number - * matrices is very dangerous, since real numbers will be convered to int - * types. If a user fills this matrix himself, invalid data may occor. - */ - - MatrixPtr startIdsOnCpu_; - MatrixPtr endIdsOnCpu_; - - std::vector selectedRows_; - IVectorPtr rowIndice_; - std::vector> inputSeqInfoVec_; - std::vector outSubSeqStartPos_; - std::vector outSeqStartPos_; - - void checkInputs(); - void copySliceIdsToCpu(); - void calSelectedRows(const MatrixPtr starts, const MatrixPtr ends); -}; - -REGISTER_LAYER(seq_slice, SequenceSliceLayer); - -bool SequenceSliceLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - CHECK_GE(inputLayers_.size(), 2U); - CHECK_LE(inputLayers_.size(), 3U); - - setNeedSequenceInfo(false); - return true; -} - -void SequenceSliceLayer::checkInputs() { - const Argument& inputSeq = getInput(0); - CHECK(inputSeq.hasSeq()) << "The first input of sequence slice layer " - << "must be a sequence."; - const MatrixPtr indices1 = getInputValue(1); - CHECK_EQ( - indices1->getHeight(), - static_cast(inputSeq.hasSubseq() ? inputSeq.getNumSubSequences() - : inputSeq.getNumSequences())) - << "Height of the second input should be equal to number of sequence " - << "in the first input."; - if (inputLayers_.size() == 3) { - const MatrixPtr indices2 = getInputValue(2); - CHECK_EQ(indices2->getHeight(), indices1->getHeight()) - << "start indices and end indices should have the same height."; - CHECK_EQ(indices2->getWidth(), indices1->getWidth()) - << "start indices and end indices should have the same Width."; - } -} - -void SequenceSliceLayer::copySliceIdsToCpu() { - const MatrixPtr indices1 = getInputValue(1); - if (inputLayers_.size() == 2U) { - if (config_.select_first()) { - Matrix::resizeOrCreate(startIdsOnCpu_, - indices1->getHeight(), - indices1->getWidth(), - false /* trans */, - false /* useGpu */); - startIdsOnCpu_->copyFrom(*indices1); - endIdsOnCpu_ = nullptr; - } else { - Matrix::resizeOrCreate(endIdsOnCpu_, - indices1->getHeight(), - indices1->getWidth(), - false /* trans */, - false /* useGpu */); - endIdsOnCpu_->copyFrom(*indices1); - startIdsOnCpu_ = nullptr; - } - } else if (inputLayers_.size() == 3U) { - Matrix::resizeOrCreate(startIdsOnCpu_, - indices1->getHeight(), - indices1->getWidth(), - false /* trans */, - false /* useGpu */); - startIdsOnCpu_->copyFrom(*indices1); - - const MatrixPtr indices2 = getInputValue(2); - Matrix::resizeOrCreate(endIdsOnCpu_, - indices2->getHeight(), - indices2->getWidth(), - false /* trans */, - false /* useGpu */); - endIdsOnCpu_->copyFrom(*indices2); - } -} - -void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts, - const MatrixPtr ends) { - CHECK(starts || ends) << "At least one of the start or end indices " - << "should be given."; - - bool hasSubseq = getInput(0).hasSubseq(); - - outSeqStartPos_.resize(1, 0); - outSubSeqStartPos_.resize(1, 0); - selectedRows_.clear(); - - size_t beamSize = starts ? starts->getWidth() : ends->getWidth(); - size_t rowIdx = 0; - for (size_t i = 0; i < inputSeqInfoVec_.size(); ++i) { - for (size_t j = 0; j < inputSeqInfoVec_[i].size() - 1; ++j) { - for (size_t k = 0; k < beamSize; ++k) { - if (starts && starts->getElement(rowIdx, k) == -1.) break; - if (ends && ends->getElement(rowIdx, k) == -1.) break; - - int begPos = inputSeqInfoVec_[i][j]; - if (starts) begPos += starts->getElement(rowIdx, k); - - int endPos = inputSeqInfoVec_[i][j + 1] - 1; - if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k); - - int seqLen = endPos - begPos + 1; - CHECK_GT(seqLen, 0); - for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m); - hasSubseq - ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen) - : outSeqStartPos_.push_back(outSeqStartPos_.back() + seqLen); - } - rowIdx++; - } - if (hasSubseq) outSeqStartPos_.push_back(outSubSeqStartPos_.back()); - } - - if (useGpu_) { - rowIndice_ = IVector::create(selectedRows_.size(), useGpu_); - rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size()); - } else { - rowIndice_ = - IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_); - } - - // create the sequence information for the output. - ICpuGpuVector::resizeOrCreate( - output_.sequenceStartPositions, outSeqStartPos_.size(), false); - output_.sequenceStartPositions->copyFrom( - outSeqStartPos_.data(), outSeqStartPos_.size(), false); - - if (hasSubseq) { - ICpuGpuVector::resizeOrCreate( - output_.subSequenceStartPositions, outSubSeqStartPos_.size(), false); - output_.subSequenceStartPositions->copyFrom( - outSubSeqStartPos_.data(), outSubSeqStartPos_.size(), false); - } -} - -void SequenceSliceLayer::forward(PassType passType) { - Layer::forward(passType); - checkInputs(); - - const Argument& inputSeq = getInput(0); - inputSeqInfoVec_.clear(); - Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, - inputSeq.subSequenceStartPositions, - inputSeqInfoVec_); - if (!useGpu_) { - if (inputLayers_.size() == 2U) { - startIdsOnCpu_ = config_.select_first() ? getInputValue(1) : nullptr; - endIdsOnCpu_ = config_.select_first() ? nullptr : getInputValue(1); - } else if (inputLayers_.size() == 3U) { - startIdsOnCpu_ = getInputValue(1); - endIdsOnCpu_ = getInputValue(2); - } - } else { - copySliceIdsToCpu(); - } - - /* - * calculate the selected row indices in a batch, and build the output - * sequence information. - */ - calSelectedRows(startIdsOnCpu_, endIdsOnCpu_); - - resetOutput(selectedRows_.size(), getSize()); - - getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); -} - -void SequenceSliceLayer::backward(const UpdateCallback& callback) { - getOutputGrad()->addToRows(*getInputGrad(0), *rowIndice_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SequenceToBatch.h b/paddle/gserver/layers/SequenceToBatch.h deleted file mode 100644 index 5200e702d9bc947746567c19ca7d552750828131..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SequenceToBatch.h +++ /dev/null @@ -1,107 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" - -namespace paddle { - -/* - * This class can used to modify the matrix structure of sequence matrix into - * batch structure. - * sequence matrix: [C1_s ... Cn_s | ...... | C1_t ... Cn_t] - * batch matrix: [C1_s ... C1_t | ...... | Cn_s ... Cn_t] - * Cn_s is the state for sequence s at time n. - * - * Exampel: sequence matrix = {{0, 0, 0, 0}, {1, 1, 1, 1, 1}, {2, 2, 2}} - * s0: 0 0 0 0, s1: 1 1 1 1 1, s2: 2 2 2 - * batch matrix = {{1, 0, 2}, {1, 0, 2}, {1, 0, 2}, {1, 0}, {1}} - * b0: 1 0 2, b1: 1 0 2, b2: 1 0 2, b3: 1 0, b4: 1 - * - * Use: - * Input: seqMatrix, seqStarts(Sequence Start Positions) - * Output: batchMatrix - * 1. SequenceToBatch seq2batch; - * 2. seq2batch.resizeOrCreateBatch(seqStarts); // calculate seq2BatchIdx - * 3. seq2batch.copy(seqMatrix, batchMatrix, true); // copy seq to batch matrix - * - */ -class SequenceToBatch { - public: - explicit SequenceToBatch(bool useGpu) : useGpu_(useGpu) {} - - /* resize and calculate the batchIndex_ */ - void resizeOrCreateBatch(int batchSize, - size_t numSequences, - const int *seqStarts, - bool reversed, - bool prevBatchState = false); - - /* sequence matrix and batch matrix copy: - * seq2batch: copy(seqValue, batchValue, true); - * batch2seq: copy(seqValue, batchValue, false); - */ - void copy(Matrix &seqValue, Matrix &batchValue, bool seq2batch); - /* sequence/batch matrix add to batch/sequence matrix */ - void add(Matrix &seqValue, Matrix &batchValue, bool seq2batch); - MatrixPtr getBatchValue(Matrix &batchValue, int batchId, int numRows = 0); - - size_t getNumBatch() const { return numBatch_; } - - /* resize or create a batch matrix(batchValue_) */ - void resizeOrCreate(Matrix &seqValue); - /* copy seqValue to batchValue_ */ - void copyFromSeq(Matrix &seqValue); - /* copy batchValue_ to seqValue */ - void copyBackSeq(Matrix &seqValue); - MatrixPtr getBatchValue(int batchId, int numRows = 0); - MatrixPtr getBatchValue() { return batchValue_; } - /*tranfer preBatchOutput to batch struct*/ - void prevOutput2Batch(Matrix &src, Matrix &dst); - /*get sequence output from batch struct*/ - void getSeqOutputFromBatch(Matrix &sequence, Matrix &batch); - - /* Copy the index from another seq2batch. */ - void shareIndexWith(const SequenceToBatch &seq2batch) { - CHECK(useGpu_ == seq2batch.useGpu_); - batchStartPositions_ = seq2batch.batchStartPositions_; - seq2BatchIdx_ = seq2batch.seq2BatchIdx_; - cpuSeq2BatchIdx_ = seq2batch.cpuSeq2BatchIdx_; - numBatch_ = seq2batch.numBatch_; - } - - protected: - void sequence2BatchCopy(Matrix &batch, - Matrix &sequence, - IVector &seq2BatchIdx, - bool seq2batch); - void sequence2BatchAdd(Matrix &batch, - Matrix &sequence, - IVector &seq2BatchIdx, - bool seq2batch); - - IVectorPtr batchStartPositions_; - IVectorPtr seq2BatchIdx_; - IVectorPtr cpuSeq2BatchIdx_; - IVectorPtr cpuSeqIdx_; - IVectorPtr cpuSeqEndIdxInBatch_; - IVectorPtr seqIdx_; - IVectorPtr seqEndIdxInBatch_; - size_t numBatch_; - bool useGpu_; - MatrixPtr batchValue_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/SlopeInterceptLayer.cpp b/paddle/gserver/layers/SlopeInterceptLayer.cpp deleted file mode 100644 index f7f4735c1b72d4ac6540714573fd7e15ef99ea5b..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SlopeInterceptLayer.cpp +++ /dev/null @@ -1,94 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief A layer for applying a slope and an intercept to the input - * element-wise. - * This layer is used in NEURAL TURING MACHINE. - * @note There is no activation and weight in this layer. - * - * \f[ - * y = ax + b - * \f] - * - * Here, a is scale and b is offset, which are provided as attributes of the - * layer. - * - * The config file api is slope_intercept_layer. - */ - -class SlopeInterceptLayer : public Layer { - public: - explicit SlopeInterceptLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(slope_intercept, SlopeInterceptLayer); - -bool SlopeInterceptLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 1U); - - return true; -} - -void SlopeInterceptLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV = getInputValue(0); - - /* malloc memory for the output_ if necessary */ - size_t batchSize = inV->getHeight(); - size_t size = getSize(); - - CHECK_EQ(size, inV->getWidth()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - reserveOutput(batchSize, size); - } - - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwSlopeInterceptTimer", getName().c_str()); - outV->mulScalar(*inV, config_.slope()); - outV->add(config_.intercept()); - } -} - -void SlopeInterceptLayer::backward(const UpdateCallback& callback) { - MatrixPtr inG = getInputGrad(0); - MatrixPtr outG = getOutputGrad(); - - if (inG) { - REGISTER_TIMER_INFO("BwSlopeInterceptTimer", getName().c_str()); - inG->add(*outG, config_.slope()); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SpatialPyramidPoolLayer.h b/paddle/gserver/layers/SpatialPyramidPoolLayer.h deleted file mode 100644 index 421bdfe09c46f656f500daff195c755274bf8bb7..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SpatialPyramidPoolLayer.h +++ /dev/null @@ -1,59 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "PoolProjection.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/Logging.h" - -namespace paddle { -/** - * @brief A layer for spatial pyramid pooling on the input image by taking - * the max, average, etc. within regions, so that the result vector of - * different sized images are of the same size. - * - * The config file api is spp_layer. - */ - -class SpatialPyramidPoolLayer : public Layer { - protected: - size_t channels_; - size_t imgSizeW_; - size_t imgSizeH_; - size_t pyramidHeight_; - std::string poolType_; - - std::vector> poolProjections_; - std::vector projOutput_; - std::vector> projCol_; - - public: - explicit SpatialPyramidPoolLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - ProjectionConfig getConfig(size_t sizeX_, - size_t sizeY_, - size_t channels, - size_t pyamidLevel_, - std::string& poolType_); - size_t getSize(); - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/SubNestedSequenceLayer.cpp b/paddle/gserver/layers/SubNestedSequenceLayer.cpp deleted file mode 100644 index e2bb00bbfacb26dc736a63877119b379f22b5983..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SubNestedSequenceLayer.cpp +++ /dev/null @@ -1,187 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -class SubNestedSequenceLayer : public Layer { - public: - explicit SubNestedSequenceLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; - - private: - /* - * This functions generates the indices of rows in a batch according to the - * indices of selected sub-sequence in each sequence. - * - * Examples: - * selectedIndices: - * [ - * [0, 1, -1], - * [0, 1, 2], - * [0, -1, -1], - * [0, 2, 3], - * ] - * inputSeqInfo: - * [ - * [0,3,4], - * [4,5,7,10,15], - * [15,20], - * [20,22,23,25,28] - * ] - * - * ths output is saved to private member rowIndice_; - * [0,1,2,3,4,5,6,7,8,9,15,16,17,18,19,20,21,23,24,25,26,27] - */ - - void calSelectedRows(const MatrixPtr selectedIndices, - const std::vector>& inputSeqInfo); - - /* - * TODO(caoying) - * In PaddePaddle, currently all matrices are real number types, - * but the second is some selected indices of the give sequence to trim - * the nested sequence, are actually filled with int types so that storing - * int types information in real number matrices is very dangerous, since - * real numbers will be convered to int types. If a user fills this matrix - * himself, invalid data may occor. - * - * if the second input of this layer is on GPU memory, copy it to CPU memory. - */ - MatrixPtr selIdsCpu_; - - /* - * reorganize sequenceStartPositions and subSequenceStartPositions - * into a 2d vector to facilitate the sequence selection process. - */ - std::vector> inputSeqInfoVec_; - - /* store the final selected row indices in a batch */ - IVectorPtr rowIndice_; - /* rowIndice_ and selectedRows_ actually share a same memory. */ - std::vector selectedRows_; -}; - -REGISTER_LAYER(sub_nested_seq, SubNestedSequenceLayer); - -bool SubNestedSequenceLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - CHECK_EQ(2U, inputLayers_.size()); - setNeedSequenceInfo(false); - return true; -} - -void SubNestedSequenceLayer::calSelectedRows( - const MatrixPtr selectedIndices, - const std::vector>& inputSeqInfo) { - selectedRows_.clear(); - - std::vector outSeqStartInfo(1, 0); - std::vector outSubSeqStartInfo(1, 0); - - size_t seqNum = selectedIndices->getHeight(); - size_t beamSize = selectedIndices->getWidth(); - for (size_t i = 0; i < seqNum; ++i) { - for (size_t j = 0; j < beamSize; ++j) { - if (selectedIndices->getElement(i, j) == -1.) break; - size_t selSubSeqIdx = selectedIndices->getElement(i, j); - CHECK_GT(inputSeqInfoVec_[i].size() - 1, selSubSeqIdx); - - size_t subSeqLen = inputSeqInfoVec_[i][selSubSeqIdx + 1] - - inputSeqInfoVec_[i][selSubSeqIdx]; - for (size_t k = 0; k < subSeqLen; ++k) - selectedRows_.push_back(inputSeqInfoVec_[i][selSubSeqIdx] + k); - outSubSeqStartInfo.push_back(outSubSeqStartInfo.back() + subSeqLen); - } - outSeqStartInfo.push_back(outSubSeqStartInfo.back()); - } - - if (useGpu_) { - rowIndice_ = IVector::create(selectedRows_.size(), useGpu_); - rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size()); - } else { - rowIndice_ = - IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_); - } - - // create the sequence information for the output. - ICpuGpuVector::resizeOrCreate( - output_.sequenceStartPositions, outSeqStartInfo.size(), false); - output_.sequenceStartPositions->copyFrom( - outSeqStartInfo.data(), outSeqStartInfo.size(), false); - - ICpuGpuVector::resizeOrCreate( - output_.subSequenceStartPositions, outSubSeqStartInfo.size(), false); - output_.subSequenceStartPositions->copyFrom( - outSubSeqStartInfo.data(), outSubSeqStartInfo.size(), false); -} - -void SubNestedSequenceLayer::forward(PassType passType) { - Layer::forward(passType); - - const Argument& inputSeq = getInput(0); - CHECK(inputSeq.hasSubseq()) << "The first input of SubNestSequence layer " - << "must be a nested sequence."; - const MatrixPtr selectedIndices = getInputValue(1); - CHECK_EQ(size_t(inputSeq.getNumSequences()), selectedIndices->getHeight()); - - if (dynamic_cast(selectedIndices.get())) { - /* - * Currently, the second input for this layer is generated by - * kmax_sequence_score_layer whose output is always stored on CPU, - * or a data_layer which canbe on GPU. - * - * If the second input is on GPU, copy it to CPU memory, because this - * input always uses very few memory, and operations related to it are - * all logic control, not computations. - */ - Matrix::resizeOrCreate(selIdsCpu_, - selectedIndices->getHeight(), - selectedIndices->getWidth(), - false /* trans */, - false /* useGpu */); - selIdsCpu_->copyFrom(*selectedIndices); - } else { - selIdsCpu_ = selectedIndices; - } - - Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, - inputSeq.subSequenceStartPositions, - inputSeqInfoVec_); - calSelectedRows(selIdsCpu_, inputSeqInfoVec_); - - resetOutput(selectedRows_.size(), getSize()); - getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); -} - -void SubNestedSequenceLayer::backward(const UpdateCallback& callback) { - MatrixPtr inputSeqGrad = getInputGrad(0); - MatrixPtr outputGrad = getOutputGrad(); - - if (inputSeqGrad) outputGrad->addToRows(*inputSeqGrad, *rowIndice_); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SubSequenceLayer.cpp b/paddle/gserver/layers/SubSequenceLayer.cpp deleted file mode 100644 index ba49f5710f9d0bb985cf1e80d5c4a972d8f046a6..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SubSequenceLayer.cpp +++ /dev/null @@ -1,226 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for taking the subsequence according to given offset and size - * Input: original sequence, offset, size - * Output: subsequence - */ - -class SubSequenceLayer : public Layer { - protected: - std::unique_ptr biases_; - MatrixPtr tmpSrc_; - MatrixPtr tmpDest_; - - public: - explicit SubSequenceLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(subseq, SubSequenceLayer); - -bool SubSequenceLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - // sequene concatenation layer should have exactly 2 inputs - CHECK_EQ(3U, inputLayers_.size()); - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - - tmpSrc_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); - tmpDest_ = - Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); - - setNeedSequenceInfo(false); - return true; -} - -void SubSequenceLayer::forward(PassType passType) { - Layer::forward(passType); - - size_t dim = getSize(); - - const Argument& input = getInput(0); - size_t numSequences1 = input.getNumSequences(); - auto startPositions1 = input.sequenceStartPositions->getVector(false); - - const Argument& offsetSeq = getInput(1); - size_t numSequences2 = offsetSeq.getNumSequences(); - auto startPositions2 = offsetSeq.sequenceStartPositions->getVector(false); - - const Argument& sizeSeq = getInput(2); - size_t numSequences3 = sizeSeq.getNumSequences(); - auto startPositions3 = sizeSeq.sequenceStartPositions->getVector(false); - - CHECK_EQ(dim, input.value->getWidth()); - - CHECK_EQ(startPositions1->getData()[numSequences1], input.getBatchSize()); - CHECK_EQ(numSequences1, startPositions1->getSize() - 1); - - CHECK_EQ(startPositions2->getData()[numSequences2], offsetSeq.getBatchSize()); - CHECK_EQ(numSequences2, startPositions2->getSize() - 1); - - CHECK_EQ(startPositions3->getData()[numSequences3], sizeSeq.getBatchSize()); - CHECK_EQ(numSequences3, startPositions3->getSize() - 1); - - CHECK_EQ(numSequences1, numSequences2); - CHECK_EQ(numSequences2, numSequences3); - - MatrixPtr inputValue = input.value; - IVectorPtr offsetValue; - IVectorPtr sizeValue; - - if (useGpu_) { - // copy to cpu - IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false); - IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false); - offsetValue->copyFrom(*offsetSeq.ids); - sizeValue->copyFrom(*sizeSeq.ids); - } else { - offsetValue = offsetSeq.ids; - sizeValue = sizeSeq.ids; - } - - CHECK_EQ(offsetValue->getSize(), numSequences1); - CHECK_EQ(sizeValue->getSize(), numSequences1); - - int* offsets = offsetValue->getData(); - int* sizes = sizeValue->getData(); - - // get total height of output - size_t height = 0; - for (size_t seqId = 0; seqId < numSequences1; seqId++) { - height += sizes[seqId]; - } - - // reset output - resetOutput(height, dim); - - MatrixPtr outputValue = getOutputValue(); - - const int* starts1 = startPositions1->getData(); - - { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SubSequenceLayerForward", getName().c_str()); - - size_t offsetIn = 0; - size_t offsetOut = 0; - size_t size = 0; - for (size_t seqId = 0; seqId < numSequences1; ++seqId) { - offsetIn = starts1[seqId] + offsets[seqId]; - size = sizes[seqId]; - - outputValue->subMatrix(offsetOut, size, tmpDest_) - ->assign(*(inputValue->subMatrix(offsetIn, size, tmpSrc_))); - - offsetOut += size; - } - - // modify the sequenceStartPositions - ICpuGpuVector::resizeOrCreate( - output_.sequenceStartPositions, numSequences1 + 1, false); - - int* tgtBuf = output_.sequenceStartPositions->getMutableData(false); - int offset = 0; - for (size_t seqId = 0; seqId < numSequences1; ++seqId) { - tgtBuf[seqId] = offset; - offset += sizes[seqId]; - } - tgtBuf[numSequences1] = offset; - } - - if (biases_.get() != NULL) { - MatrixPtr outV = getOutputValue(); - outV->addBias(*(biases_->getW()), 1); - } - - /* activation */ - forwardActivation(); -} - -void SubSequenceLayer::backward(const UpdateCallback& callback) { - /* activation */ - backwardActivation(); - - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - // Increasing the number of gradient - biases_->getParameterPtr()->incUpdate(callback); - } - - MatrixPtr inputGrad1 = getInputGrad(0); - MatrixPtr outputGrad = getOutputGrad(); - auto startPositions1 = getInput(0).sequenceStartPositions->getVector(false); - size_t numSequences1 = startPositions1->getSize() - 1; - const int* starts1 = startPositions1->getData(); - - const Argument& offsetSeq = getInput(1); - const Argument& sizeSeq = getInput(2); - IVectorPtr offsetValue; - IVectorPtr sizeValue; - - if (useGpu_) { - // copy to cpu - IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false); - IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false); - offsetValue->copyFrom(*offsetSeq.ids); - sizeValue->copyFrom(*sizeSeq.ids); - } else { - offsetValue = offsetSeq.ids; - sizeValue = sizeSeq.ids; - } - - int* offsets = offsetValue->getData(); - int* sizes = sizeValue->getData(); - { - AsyncGpuBlock asyncGpuBlock; - REGISTER_TIMER_INFO("SubSequenceLayerBackward", getName().c_str()); - - int offsetIn = 0; - int offsetOut = 0; - int size = 0; - for (size_t seqId = 0; seqId < numSequences1; ++seqId) { - offsetIn = starts1[seqId] + offsets[seqId]; - size = sizes[seqId]; - - inputGrad1->subMatrix(offsetIn, size, tmpDest_) - ->add(*(outputGrad->subMatrix(offsetOut, size, tmpSrc_))); - offsetOut += size; - } - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SumToOneNormLayer.cpp b/paddle/gserver/layers/SumToOneNormLayer.cpp deleted file mode 100644 index 00764717e8b6be30230e44626974033e929352da..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SumToOneNormLayer.cpp +++ /dev/null @@ -1,120 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * A layer for sum-to-one normalization, - * which is used in NEURAL TURING MACHINE. - * \f[ - * out[i] = \frac {in[i]} {\sum_{k=1}^N in[k]} - * \f] - * where \f$in\f$ is a (batchSize x dataDim) input vector, - * and \f$out\f$ is a (batchSize x dataDim) output vector. - * - * The config file api is sum_to_one_norm_layer. - */ - -class SumToOneNormLayer : public Layer { - protected: - /// reciprocalRowSum_ = \f$1 / \sum_{k=1}^N in[k]\f$ - MatrixPtr reciprocalRowSum_; - /// dotSum = output_.grad \f$.*\f$ output_.value - MatrixPtr dotSum_; - - public: - explicit SumToOneNormLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; - -REGISTER_LAYER(sum_to_one_norm, SumToOneNormLayer); - -bool SumToOneNormLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - Layer::init(layerMap, parameterMap); - - CHECK_EQ(inputLayers_.size(), 1U); - - return true; -} - -void SumToOneNormLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr inV = getInputValue(0); - - /* malloc memory for the output_ if necessary */ - size_t batchSize = inV->getHeight(); - size_t dataDim = getSize(); - - CHECK_EQ(dataDim, inV->getWidth()); - - { - REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); - resetOutput(batchSize, dataDim); - } - - MatrixPtr outV = getOutputValue(); - { - REGISTER_TIMER_INFO("FwSumToOneNormTimer", getName().c_str()); - - Matrix::resizeOrCreate(reciprocalRowSum_, batchSize, 1, false, useGpu_); - inV->rowSum(*reciprocalRowSum_); - - // todo: matrix checks - CHECK_GT(reciprocalRowSum_->getMin(), 0.0); - - reciprocalRowSum_->scalarDiv(*reciprocalRowSum_, 1.0); - - // outV = inV * reciprocalRowSum - outV->rowScale(0, *inV, *reciprocalRowSum_); - } -} - -void SumToOneNormLayer::backward(const UpdateCallback& callback) { - MatrixPtr inV = getInputValue(0); - MatrixPtr inG = getInputGrad(0); - MatrixPtr outV = getOutputValue(); - MatrixPtr outG = getOutputGrad(); - - size_t batchSize = inV->getHeight(); - - if (inG) { - REGISTER_TIMER_INFO("BwSumToOneTimer", getName().c_str()); - - Matrix::resizeOrCreate(dotSum_, batchSize, 1, false, useGpu_); - - // dotSum = outG .* outV - dotSum_->zeroMem(); - dotSum_->rowDotMul(0, *outG, *outV); - - // inG += -1 * (dotSum / rowSum) - dotSum_->dotMul(*dotSum_, *reciprocalRowSum_); - inG->rowAdd(0, *inG, *dotSum_, -1.0); - // inG += outG * (1/rowSum) - inG->addRowScale(0, *outG, *reciprocalRowSum_); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/SwitchOrderLayer.cpp b/paddle/gserver/layers/SwitchOrderLayer.cpp deleted file mode 100644 index 704735de38bd373c0714de6bb4e139d1505c5451..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/SwitchOrderLayer.cpp +++ /dev/null @@ -1,109 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SwitchOrderLayer.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(switch_order, SwitchOrderLayer); - -bool SwitchOrderLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - auto& img_conf = config_.inputs(0).image_conf(); - size_t inD = img_conf.img_size_z(); - size_t inH = - img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(); - size_t inW = img_conf.img_size(); - size_t inC = img_conf.channels(); - inH = inH * inD; - inDims_ = TensorShape({0, inC, inH, inW}); - outDims_ = TensorShape(4); - - auto& reshape_conf = config_.reshape_conf(); - for (int i = 0; i < reshape_conf.height_axis_size(); i++) { - heightAxis_.push_back(reshape_conf.height_axis(i)); - } - for (int i = 0; i < reshape_conf.width_axis_size(); i++) { - widthAxis_.push_back(reshape_conf.width_axis(i)); - } - createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig()); - createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig()); - return true; -} - -void SwitchOrderLayer::setOutDims() { - outDims_.setDim(0, inDims_[0]); - outDims_.setDim(1, inDims_[2]); - outDims_.setDim(2, inDims_[3]); - outDims_.setDim(3, inDims_[1]); - reshapeHeight_ = 1; - for (size_t i = 0; i < heightAxis_.size(); i++) { - reshapeHeight_ *= outDims_[heightAxis_[i]]; - } - output_.setFrameHeight(reshapeHeight_); - reshapeWidth_ = 1; - for (size_t i = 0; i < widthAxis_.size(); i++) { - reshapeWidth_ *= outDims_[widthAxis_[i]]; - } - output_.setFrameWidth(reshapeWidth_); -} - -void SwitchOrderLayer::setInDims() { - MatrixPtr input = inputLayers_[0]->getOutputValue(); - size_t batchSize = input->getHeight(); - inDims_.setDim(0, batchSize); - int d = inputLayers_[0]->getOutput().getFrameDepth(); - d = (d == 0 ? 1 : d); - int h = inputLayers_[0]->getOutput().getFrameHeight(); - if (h != 0) inDims_.setDim(2, h * d); - int w = inputLayers_[0]->getOutput().getFrameWidth(); - if (w != 0) inDims_.setDim(3, w); - int totalCount = input->getElementCnt(); - int channels = totalCount / (inDims_[0] * inDims_[2] * inDims_[3]); - if (channels != 0) inDims_.setDim(1, channels); -} - -void SwitchOrderLayer::forward(PassType passType) { - Layer::forward(passType); - setInDims(); - setOutDims(); - resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]); - if (heightAxis_.size() > 0) { - resetOutput(reshapeHeight_, reshapeWidth_); - } - - // switch NCHW to NHWC - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getInputValue(0), inDims_); - outputs.addArg(*getOutputValue(), outDims_); - nchw2nhwc_[0]->calc(inputs, outputs); - forwardActivation(); -} - -void SwitchOrderLayer::backward(const UpdateCallback& callback) { - (void)callback; - backwardActivation(); - - // switch NHWC to NCHW - BufferArgs inputs; - BufferArgs outputs; - inputs.addArg(*getOutputGrad(), outDims_); - outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); - nhwc2nchw_[0]->calc(inputs, outputs); -} -} // namespace paddle diff --git a/paddle/gserver/layers/TensorLayer.cpp b/paddle/gserver/layers/TensorLayer.cpp deleted file mode 100644 index b2271c63ef76d85574cf7f71b18aef4239938b8e..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/TensorLayer.cpp +++ /dev/null @@ -1,145 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "TensorLayer.h" - -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -REGISTER_LAYER(tensor, TensorLayer); - -bool TensorLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* initialize the weightList */ - CHECK_EQ(inputLayers_.size(), 2LU); - CHECK(parameters_[0]); - CHECK(!parameters_[1]); - - // Option the parameters - size_t height = inputLayers_[0]->getSize(); - size_t width = inputLayers_[1]->getSize(); - CHECK_EQ(width * height * getSize(), parameters_[0]->getSize()); - - for (size_t i = 0; i < getSize(); ++i) { - // create a new weight - Weight* w = new Weight(height, width, parameters_[0], i * width * height); - - // append the new weight to the list - weights_.emplace_back(w); - } - - /* initialize biases_ */ - if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); - } - - return true; -} - -void TensorLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - int batchSize = getInputValue(0)->getHeight(); - int size = getSize(); - - { resetOutput(batchSize, size); } - - MatrixPtr outV = getOutputValue(); - /* add the bias-vector */ - if (biases_.get() != NULL) { - outV->addBias(*(biases_->getW()), 1); - } - - /* e1 * W * trans(e2) */ { - MatrixPtr input1 = getInputValue(0); - MatrixPtr input2 = getInputValue(1); - MatrixPtr tmpMat = Matrix::create(input2->getHeight(), - input2->getWidth(), - /* trans= */ false, - input2->useGpu()); - REGISTER_TIMER_INFO("TensorFwMulTimer", getName().c_str()); - for (size_t i = 0; i < getSize(); ++i) { - MatrixPtr weights = weights_[i]->getW(); - tmpMat->mul(*input1, *weights, 1, 0); - outV->rowDotMul(i, *tmpMat, *input2); - } - } - - /* activation */ { forwardActivation(); } -} - -void TensorLayer::backward(const UpdateCallback& callback) { - /* Do derivation */ { backwardActivation(); } - - if (biases_ && biases_->getWGrad()) { - biases_->getWGrad()->collectBias(*getOutputGrad(), 1); - - /* Increasing the number of gradient */ - biases_->getParameterPtr()->incUpdate(callback); - } - - bool syncFlag = hl_get_sync_flag(); - - /* Calculate the W-gradient for the current layer */ - MatrixPtr input1 = getInputValue(0); - MatrixPtr input2 = getInputValue(1); - MatrixPtr oGrad = getOutputGrad(); - MatrixPtr tmpMat = Matrix::create(input1->getHeight(), - input1->getWidth(), - /* trans= */ false, - input1->useGpu()); - - /* trans(grad * e1) * e2 */ { - REGISTER_TIMER_INFO("TensorGradMulTimer", getName().c_str()); - for (size_t i = 0; i < getSize(); ++i) { - if (weights_[i]->getWGrad()) { - tmpMat->rowScale(i, *input1, *oGrad); - MatrixPtr input1_T = tmpMat->getTranspose(); - weights_[i]->getWGrad()->mul(*input1_T, *input2, 1, 1); - } - } - } - - hl_set_sync_flag(false); - - /* Calculate the input layers error */ { - MatrixPtr preGrad1 = getInputGrad(0); - MatrixPtr preGrad2 = getInputGrad(1); - - REGISTER_TIMER_INFO("TensorBpMulTimer", getName().c_str()); - for (size_t i = 0; i < getSize(); ++i) { - MatrixPtr weights = weights_[i]->getW(); - - if (NULL != preGrad1) { /* (grad * e2) * trans(W) */ - tmpMat->rowScale(i, *input2, *oGrad); - MatrixPtr weights_T = weights->getTranspose(); - preGrad1->mul(*tmpMat, *weights_T, 1, 1); - } - if (NULL != preGrad2) { /* (grad * e1) * W */ - tmpMat->rowScale(i, *input1, *oGrad); - preGrad2->mul(*tmpMat, *weights, 1, 1); - } - } - } - hl_set_sync_flag(syncFlag); - parameters_[0]->incUpdate(callback); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/TensorLayer.h b/paddle/gserver/layers/TensorLayer.h deleted file mode 100644 index 5c1ee40ceda9387138a82368ec4edcbae4bd3419..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/TensorLayer.h +++ /dev/null @@ -1,55 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/** - * @brief TensorLayer takes two input vectors. - * \f[ - * y_{i} = x_{1} * W_{i} * x_{2}^{\rm T}, i=0, 1, ...,K-1 - * \f] - * - * - \f$x_{1}\f$: the first input, size is M. - * - \f$x_{2}\f$: the second input, size is N. - * - y: output, size is K. - * - \f$y_{i}\f$: i-th element of y. - * - \f$W_{i}\f$: the i-th learned weight, dimensions: [M, N]. - * - \f$x_{2}^{\rm T}\f$: the transpose of \f$x_{2}\f$. - * - * The config file api is tensor_layer. - */ - -class TensorLayer : public Layer { - protected: - WeightList weights_; - std::unique_ptr biases_; - - public: - explicit TensorLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - Weight& getWeight(int idx) { return *weights_[idx]; } - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/TransLayer.cpp b/paddle/gserver/layers/TransLayer.cpp deleted file mode 100644 index cf87ca53d1def32708400c507da673c3a6ec0a87..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/TransLayer.cpp +++ /dev/null @@ -1,69 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "TransLayer.h" -#include "paddle/utils/Logging.h" -namespace paddle { - -REGISTER_LAYER(trans, TransLayer); - -bool TransLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - /* Initialize the basic parent class */ - Layer::init(layerMap, parameterMap); - - /* the size of inputs for trans-layer is 1 */ - CHECK_EQ(config_.inputs_size(), 1); - - return true; -} - -void TransLayer::forward(PassType passType) { - Layer::forward(passType); - - /* malloc memory for the output_ if necessary */ - MatrixPtr input = getInputValue(0); - int height = input->getHeight(); - int width = input->getWidth(); - - resizeOutput(width, height); - - MatrixPtr outV = getOutputValue(); - - /* outV's memory has been allocated, so memAlloc = false */ - input->transpose(outV, false); - if (getInputGrad(0)) { - zeroGrad(); - } -} - -void TransLayer::backward(const UpdateCallback& callback) { - (void)callback; - - MatrixPtr outputGrad = getOutputGrad(); - if (outputGrad == NULL) { - return; - } - MatrixPtr preGrad = getInputGrad(0); - if (preGrad) { - MatrixPtr transGrad = Matrix::create(preGrad->getHeight(), - preGrad->getWidth(), - /* trans= */ false, - preGrad->useGpu()); - outputGrad->transpose(transGrad, false); - preGrad->add(*transGrad); - } -} - -} // namespace paddle diff --git a/paddle/gserver/layers/TransLayer.h b/paddle/gserver/layers/TransLayer.h deleted file mode 100644 index 1cd8fd91f785d5a43fc7d7663e657702b32fa534..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/TransLayer.h +++ /dev/null @@ -1,41 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Layer.h" -#include "paddle/math/Matrix.h" - -namespace paddle { -/** - * A layer for transposing a minibatch matrix. - * \f[ - y = x^\mathrm{T} - * \f] - * where \f$x\f$ is (M x N) input, and \f$y\f$ is (N x M) output. - * - * The config file api is trans_layer. - */ -class TransLayer : public Layer { - public: - explicit TransLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback = nullptr) override; -}; -} // namespace paddle diff --git a/paddle/gserver/layers/TransposedFullMatrixProjection.cpp b/paddle/gserver/layers/TransposedFullMatrixProjection.cpp deleted file mode 100644 index 45f59779896f993aface284e3485e1e3d801f4c5..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/TransposedFullMatrixProjection.cpp +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Projection.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * @brief TransposedFullMatrixProjection performs full matrix multiplication: - * out.row[i] += in.row[i] * weight.transpose - * - * The config file api is trans_full_matrix_projection. - */ -class TransposedFullMatrixProjection : public Projection { - public: - TransposedFullMatrixProjection(const ProjectionConfig& config, - ParameterPtr parameter, - bool useGPu); - virtual void forward(); - virtual void backward(const UpdateCallback& callback); - - protected: - std::unique_ptr weight_; -}; - -REGISTER_PROJECTION(trans_fc, TransposedFullMatrixProjection); - -TransposedFullMatrixProjection::TransposedFullMatrixProjection( - const ProjectionConfig& config, ParameterPtr parameter, bool useGpu) - : Projection(config, parameter, useGpu) { - weight_.reset( - new Weight(config.output_size(), config.input_size(), parameter)); -} - -void TransposedFullMatrixProjection::forward() { - REGISTER_TIMER_INFO("FwMulTimer", getName().c_str()); - out_->value->mul(*(in_->value), *(weight_->getW()->getTranspose()), 1, 1); -} - -void TransposedFullMatrixProjection::backward(const UpdateCallback& callback) { - bool syncFlag = hl_get_sync_flag(); - - /* Calculate the W-gradient for the current layer */ - if (weight_->getWGrad()) { - REGISTER_TIMER_INFO("GradMulTimer", getName().c_str()); - weight_->getWGrad()->mul( - *(out_->grad->getTranspose()), *(in_->value), 1, 1); - } - - // If callback does not change value, backprop error asynchronously so that - // we can do the callback concurrently. - // This is still a little bit dangerous since theoretically for - // SyncMultiGpuMachine it is possible that the value copyback can still - // happen at the same time as the error backprop where the value is being - // used. - hl_set_sync_flag(false); - - /* Calculate the input layers error */ - if (in_->grad) { - REGISTER_TIMER_INFO("BpMulTimer", getName().c_str()); - in_->grad->mul(*(out_->grad), *(weight_->getW()), 1, 1); - } - - hl_set_sync_flag(syncFlag); - parameter_->incUpdate(callback); -} - -} // namespace paddle diff --git a/paddle/gserver/layers/UpsampleLayer.h b/paddle/gserver/layers/UpsampleLayer.h deleted file mode 100644 index c9d079c3141c37517866bfdad10d9b2cdb89f7d5..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/UpsampleLayer.h +++ /dev/null @@ -1,53 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Layer.h" -#include "paddle/math/Matrix.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Stat.h" - -namespace paddle { - -/** - * This layer transpose the pooling process. - * It takes two input, the first input is the input data, and - * the second is the mask data from the max-pool-with-mask layer. - * - */ - -class UpsampleLayer : public Layer { - public: - explicit UpsampleLayer(const LayerConfig& config) : Layer(config) {} - ~UpsampleLayer() {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void forward(PassType passType) override; - void backward(const UpdateCallback& callback) override; - - size_t getOutputSize(); - - protected: - size_t scale_, scaleY_; - size_t upsampleSize_, upsampleSizeY_; - size_t padOutX_, padOutY_; - size_t imgSize_, imgSizeY_; - size_t channels_; -}; - -} // namespace paddle diff --git a/paddle/gserver/layers/ValidationLayer.cpp b/paddle/gserver/layers/ValidationLayer.cpp deleted file mode 100644 index b626825a7b45fdb09cd8f9e8cc6727e218ab2940..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ValidationLayer.cpp +++ /dev/null @@ -1,171 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include - -#include "ValidationLayer.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -bool ValidationLayer::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - return Layer::init(layerMap, parameterMap); -} - -void ValidationLayer::forward(PassType passType) { - Layer::forward(passType); - - MatrixPtr output = getInputValue(*getOutputLayer()); - CHECK(output); - IVectorPtr label = getInputLabel(*getLabelLayer()); - CHECK(label); - validationImp(output, label); -} - -void ValidationLayer::backward(const UpdateCallback& callback) { - (void)callback; -} - -bool AucValidation::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - bool ret = ValidationLayer::init(layerMap, parameterMap); - EvaluatorConfig config; - config.set_name(getName()); - config.set_type("last-column-auc"); - config.add_input_layers(inputLayers_[0]->getName()); - config.add_input_layers(inputLayers_[1]->getName()); - if (3 == inputLayers_.size()) { - config.add_input_layers(inputLayers_[2]->getName()); - } - evaluator_.reset(Evaluator::create(config)); - passBegin_ = false; - return ret; -} - -void AucValidation::validationImp(MatrixPtr output, IVectorPtr label) { - if (!passBegin_) { - passBegin_ = true; - evaluator_->start(); - } - - bool supportWeight = (3 == inputLayers_.size()) ? true : false; - MatrixPtr weight = supportWeight ? getInputValue(*inputLayers_[2]) : nullptr; - if (dynamic_cast(output.get())) { - size_t height = output->getHeight(); - size_t width = output->getWidth(); - Matrix::resizeOrCreate(cpuOutput_, - height, - width, - /* trans=*/false, - /* useGpu=*/false); - cpuOutput_->copyFrom(*output); - IVector::resizeOrCreate(cpuLabel_, height, false); - cpuLabel_->copyFrom(*label); - - if (supportWeight) { - Matrix::resizeOrCreate(cpuWeight_, height, (size_t)1, false, false); - cpuWeight_->copyFrom(*weight); - } - - output = cpuOutput_; - label = cpuLabel_; - weight = cpuWeight_; - } - - for (size_t i = 0; i < output->getHeight(); i++) { - float y1 = output->getData()[i * output->getWidth() + 1]; - int* labels = label->getData(); - predictArray_.push_back(PredictionResult(y1, labels[i])); - } - std::vector arguments; - if (3 == inputLayers_.size()) { - arguments.resize(3); - arguments[2].value = weight; - } else { - arguments.resize(2); - } - arguments[0].value = output; - arguments[1].ids = label; - evaluator_->evalImp(arguments); -} - -void AucValidation::onPassEnd() { - if (!FLAGS_predict_file.empty()) { - std::ofstream fs(FLAGS_predict_file); - CHECK(fs) << "Fail to open " << FLAGS_predict_file; - for (auto& res : predictArray_) { - fs << res.out << " " << res.label << std::endl; - } - } - - evaluator_->finish(); - LOG(INFO) << *evaluator_; - passBegin_ = false; - predictArray_.clear(); -} - -bool PnpairValidation::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { - bool ret = ValidationLayer::init(layerMap, parameterMap); - if (!ret) return ret; - CHECK_GE(inputLayers_.size(), 3UL); - CHECK_LE(inputLayers_.size(), 4UL); - EvaluatorConfig config; - config.set_name(getName()); - config.set_type("pnpair"); - config.add_input_layers(inputLayers_[0]->getName()); - config.add_input_layers(inputLayers_[1]->getName()); - config.add_input_layers(inputLayers_[2]->getName()); - if (4 == inputLayers_.size()) { - config.add_input_layers(inputLayers_[3]->getName()); - } - evaluator_.reset(Evaluator::create(config)); - passBegin_ = false; - return true; -} - -void PnpairValidation::validationImp(MatrixPtr output, IVectorPtr label) { - if (!passBegin_) { - passBegin_ = true; - evaluator_->start(); - } - MatrixPtr weight = - (4 == inputLayers_.size()) ? getInputValue(*inputLayers_[3]) : nullptr; - IVectorPtr info = getInputLabel(*getInfoLayer()); - std::vector arguments; - if (4 == inputLayers_.size()) { - arguments.resize(4); - arguments[3].value = weight; - } else { - arguments.resize(3); - } - arguments[0].value = output; - arguments[1].ids = label; - arguments[2].ids = info; - evaluator_->evalImp(arguments); -} - -void PnpairValidation::onPassEnd() { - if (!FLAGS_predict_file.empty()) { - (dynamic_cast(evaluator_.get()))->printPredictResults(); - } - evaluator_->finish(); - LOG(INFO) << *evaluator_; - passBegin_ = false; -} - -} // namespace paddle diff --git a/paddle/gserver/layers/ValidationLayer.h b/paddle/gserver/layers/ValidationLayer.h deleted file mode 100644 index be41128ef4530f32a63c757648c2f393fd118ea6..0000000000000000000000000000000000000000 --- a/paddle/gserver/layers/ValidationLayer.h +++ /dev/null @@ -1,104 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include - -#include "Layer.h" -#include "paddle/gserver/evaluators/Evaluator.h" - -DECLARE_int32(trainer_id); - -namespace paddle { - -class ValidationLayer : public Layer { - public: - explicit ValidationLayer(const LayerConfig& config) : Layer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - LayerPtr getOutputLayer() { return inputLayers_[0]; } - - LayerPtr getLabelLayer() { return inputLayers_[1]; } - - LayerPtr getInfoLayer() { - assert(inputLayers_.size() > 2); - return inputLayers_[2]; - } - - void forward(PassType passType) override; - - void backward(const UpdateCallback& callback = nullptr) override; - - virtual void validationImp(MatrixPtr outputValue, IVectorPtr label) = 0; - - void onPassEnd() override = 0; -}; - -/* - * AucValidation - */ -class AucValidation : public ValidationLayer { - public: - explicit AucValidation(const LayerConfig& config) - : ValidationLayer(config), - cpuOutput_(nullptr), - cpuLabel_(nullptr), - cpuWeight_(nullptr) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void validationImp(MatrixPtr outputValue, IVectorPtr label) override; - - void onPassEnd() override; - - struct PredictionResult { - PredictionResult(real __out, int __label) : out(__out), label(__label) {} - real out; - int label; - }; - std::vector predictArray_; - - private: - bool passBegin_; - std::unique_ptr evaluator_; - MatrixPtr cpuOutput_; - IVectorPtr cpuLabel_; - MatrixPtr cpuWeight_; -}; - -/* - * positive-negative pair rate Validation - */ -class PnpairValidation : public ValidationLayer { - public: - explicit PnpairValidation(const LayerConfig& config) - : ValidationLayer(config) {} - - bool init(const LayerMap& layerMap, - const ParameterMap& parameterMap) override; - - void validationImp(MatrixPtr outputValue, IVectorPtr label) override; - - void onPassEnd() override; - - private: - bool passBegin_; - std::unique_ptr evaluator_; -}; - -typedef std::shared_ptr ValidationLayerPtr; -} // namespace paddle diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt deleted file mode 100644 index 9d7cad7584d1defefe38bdd4d041b98bd9e45bf0..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/CMakeLists.txt +++ /dev/null @@ -1,103 +0,0 @@ -# gserver pacakge unittests -add_simple_unittest(test_LinearChainCRF) -add_simple_unittest(test_RecurrentLayer) - -if(NOT MOBILE_INFERENCE) - add_simple_unittest(test_MultinomialSampler) -endif() - -function(gserver_test TARGET) - add_unittest_without_exec(${TARGET} - ${TARGET}.cpp - LayerGradUtil.cpp) - add_test(NAME ${TARGET} - COMMAND ${TARGET}) -endfunction() - -add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/concat_dotmul_a.conf - COMMAND cp -r ${CMAKE_CURRENT_SOURCE_DIR}/* ${CMAKE_CURRENT_BINARY_DIR} -) -add_custom_target(copy_gserver_conf ALL DEPENDS concat_dotmul_a.conf) - -gserver_test(test_LayerGrad) -gserver_test(test_CRFLayerGrad) -gserver_test(test_CrossEntropyOverBeamGrad) -gserver_test(test_SeqSliceLayerGrad) -gserver_test(test_ActivationGrad) -gserver_test(test_ConvTrans) -gserver_test(test_PriorBox) -gserver_test(test_DetectionOutput) -gserver_test(test_ConvUnify) -gserver_test(test_BatchNorm) -gserver_test(test_KmaxSeqScore) -gserver_test(test_Expand) -gserver_test(test_MaxPoolingWithMaskOutput) -gserver_test(test_Upsample) - -set(PYTHON_PATH - ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d - ${PADDLE_BINARY_DIR}/python/:${PADDLE_BINARY_DIR}/paddle/gserver/tests) -function(gserver_test_with_python TARGET) - add_unittest_without_exec(${TARGET} ${TARGET}.cpp) - add_test(NAME ${TARGET} - COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET} - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) -endfunction() - -gserver_test_with_python(test_PyDataProvider2) -if(WITH_PYTHON) - gserver_test_with_python(test_PyDataProvider) -endif() -if(NOT MOBILE_INFERENCE) - gserver_test_with_python(test_CompareTwoNets) - # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine, I will fix it. - gserver_test_with_python(test_RecurrentGradientMachine) -endif() - -########## test_MKLDNN layers and activations ########## -if(WITH_MKLDNN) - add_unittest_without_exec(test_MKLDNN - test_MKLDNN.cpp - MKLDNNTester.cpp - LayerGradUtil.cpp) - add_test(NAME test_MKLDNN - COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/test_MKLDNN - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle) -endif() - -############### test_WarpCTCLayer ####################### -if(NOT WITH_DOUBLE AND NOT MOBILE_INFERENCE) - add_unittest_without_exec(test_WarpCTCLayer - test_WarpCTCLayer.cpp) - add_test(NAME test_WarpCTCLayer - COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_WarpCTCLayer --warpctc_dir=${WARPCTC_LIB_DIR} - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle) -endif() - -if(NOT MOBILE_INFERENCE) - ################## test_Evaluator ############# - add_unittest(test_Evaluator - test_Evaluator.cpp) - - ########### test_NetworkCompare ############### - add_unittest_without_exec(test_NetworkCompare - test_NetworkCompare.cpp) - if(WITH_GPU) - set(use_gpu true) - else() - set(use_gpu false) - endif() - add_test(NAME test_NetworkCompare - COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=${use_gpu} - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle) - - ############ test_CompareSparse ################ - add_unittest_without_exec(test_CompareSparse - test_CompareSparse.cpp) - if(NOT ON_TRAVIS) - add_test(NAME test_CompareSparse - COMMAND ${PYTHON_PATH} ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port -n 6 - ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) - endif() -endif() diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h deleted file mode 100644 index 1999b2204b1728bd60b1e107dfe7b10718e752a5..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/LayerGradUtil.h +++ /dev/null @@ -1,329 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" - -#include "paddle/testing/TestUtil.h" -using namespace std; // NOLINT - -namespace paddle { -enum InputType { - INPUT_DATA, // dense vector - INPUT_LABEL, // id - INPUT_DATA_TARGET, // dense vector, but no gradient - INPUT_SEQUENCE_DATA, - INPUT_HASSUB_SEQUENCE_DATA, // sequence has sub-sequence - INPUT_SEQUENCE_MDIM_DATA, - INPUT_SEQUENCE_LABEL, - INPUT_SPARSE_NON_VALUE_DATA, - INPUT_SPARSE_FLOAT_VALUE_DATA, - INPUT_DENSE_DIM_DATA, // using sequence length to init dense data - INPUT_SELF_DEFINE_DATA, // support customizing for input value -}; - -struct ParaSparse { - bool sparse; - string format; - // if equalNnzPerSample is set true, - // every row of the sparse matrix in a format of CSR has a same - // number of nnz values. Currently, this flag is only used for - // selective_fc layer - bool equalNnzPerSample; - ParaSparse(const string& formatIn = "") { // NOLINT - if (formatIn == "") { - sparse = false; - } else { - sparse = true; - } - equalNnzPerSample = false; - } - ParaSparse(const string& formatIn, bool equalNnz) { - format = formatIn; - sparse = true; - equalNnzPerSample = equalNnz; - } -}; - -struct InputDef { - InputType inputType; - string name; - size_t dim; - size_t paraSize; - ParaSparse sparse; - bool isStatic; - std::vector labelInitValue; - std::vector labelSeqStartPositions; - std::vector labelSubSeqStartPositions; - std::vector ids; - MatrixPtr selfDefinedData; - - InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) { - inputType = type; - name = nameIn; - dim = dimIn; - paraSize = sizeIn; - sparse = {""}; - isStatic = false; - } - - InputDef(InputType type, - string nameIn, - MatrixPtr selfDefinedData, - std::vector selfDefinedSeqStartPos = {}, - std::vector selfDefinedSubSeqStartPos = {}) - : labelSeqStartPositions(selfDefinedSeqStartPos), - labelSubSeqStartPositions(selfDefinedSubSeqStartPos), - selfDefinedData(selfDefinedData) { - inputType = type; - name = nameIn; - dim = 0; - sparse = {""}; - paraSize = 0; - isStatic = false; - } - - InputDef(InputType type, - string nameIn, - const std::vector& ids, - const std::vector& selfDefinedSeqStartPos = {}, - const std::vector& selfDefinedSubSeqStartPos = {}) - : labelSeqStartPositions(selfDefinedSeqStartPos), - labelSubSeqStartPositions(selfDefinedSubSeqStartPos), - ids(ids) { - selfDefinedData = nullptr; - inputType = type; - name = nameIn; - dim = 0; - sparse = {""}; - paraSize = 0; - isStatic = false; - } - - InputDef(InputType type, - string nameIn, - size_t dimIn, - size_t sizeIn, - const std::vector& labelInitValue, - const std::vector& labelSeqStartPositions) - : labelInitValue(labelInitValue), - labelSeqStartPositions(labelSeqStartPositions) { - inputType = type; - name = nameIn; - dim = dimIn; - paraSize = sizeIn; - sparse = {""}; - isStatic = false; - } - - InputDef(InputType type, - string nameIn, - size_t dimIn, - size_t sizeIn, - ParaSparse sparseIn) { - inputType = type; - name = nameIn; - dim = dimIn; - paraSize = sizeIn; - sparse = sparseIn; - } -}; - -struct TestConfig { - LayerConfig layerConfig; - std::vector inputDefs; - size_t biasSize; - real paramInitialMean; - real paramInitialStd; - bool testAccumulate; - bool testState; - bool staticBias; - bool testBatchState; - TestConfig() - : biasSize(0), - paramInitialMean(0.0), - paramInitialStd(1.0), - testAccumulate(true), - testState(false), - staticBias(false), - testBatchState(false) {} -}; - -real getCostSum(ParameterPtr& parameter, - CpuVector& cpuPara, - LayerPtr& testLayer, - MatrixPtr weights = nullptr); - -real getDiffAndPrint(real newCost1, - real newCost2, - real callbackCount, - char fill, - string testLayerName, - string name, - real step, - real delta); - -/** - * @brief verify that sequentially running forward() one timestamp at one time - * has same result as running forward() with one whole sequence - * - * @param testLayer[in/out] testLayer - * @param dataLayers[in/out] dataLayers - * @param datas[in/out] data of dataLayers - */ -void testState(LayerPtr testLayer, - vector& dataLayers, - vector& datas); - -/** - * @brief verify that sequentially running forward() with short sequences one - * time has same result as running forward() with long sequences. - * - * @param testLayer[in/out] testLayer - * @param dataLayers[in/out] dataLayers - * @param datas[in/out] data of dataLayers - */ -void testBatchState(LayerPtr testLayer, - vector& dataLayers, - vector& datas); - -/** - * @brief Generate a perturbation so that it is roughly aligned with the - * gradient direction. This is to make sure that change along this - * direction will make cost increase (or decrease) in a meaningful - * way so that the finite difference can be used to approximate the - * directional dirivative well. - * - * @param oldGrad[in] input gradient - * newGrad[out] output gradient - * dim dimension of oldGrad/newGrad - * - * @return sum_i(oldGrad[i] * newGrad[i]) - */ -double genPerturbation(const real* oldGrad, real* newGrad, size_t dim); - -void initWeight(MatrixPtr& weights); - -void initBatchState(LayerPtr dataLayer, - LayerPtr testLayer, - LayerStatePtr state, - bool useGpu); - -/** - * @brief initialize the dataLayer by its inputType - * - * @param testConf[in] test config - * dataLayers[out] dataLayers - * datas[out] initialized data of dataLayers - * layerMap[out] layerMap - */ -void initDataLayer(TestConfig testConf, - std::vector* dataLayers, - vector* datas, - LayerMap* layerMap, - string testLayerName, - size_t batchSize, - bool trans, - bool useGpu); - -/** - * @brief initialize the parameter of testLayer - * - * @param testConf[in/out] test config - * layerMap[out] layerMap - * parameters[out] parameters of testLayer - * testLayer[out] testLayer - */ -void initTestLayer(TestConfig testConf, - LayerMap* layerMap, - std::vector* parameters, - LayerPtr* testLayer); - -/** - * @brief Test whether the layer's forward calculation is stable by adding - * perturbation to its parameters - * - * @param testConf[in] test config - * weights[in] weights of testLayer - * state[in] state of testLayer - * cost[in] input cost - * callbackCount[in] number of done callback - * maxDiff[in/out] max of all previous diff - * testLayer[in/out] testLayer - * parameters[in/out] parameters of testLayer - */ -void testPerturbParameter(TestConfig testConf, - const MatrixPtr weights, - const LayerStatePtr state, - real cost, - real callbackCount, - real* maxDiff, - LayerPtr testLayer, - std::vector* parameters); - -/** - * @brief Test whether the layer's forward calculation is stable by adding - * perturbation to its input layers - * - * @param testConf[in] test config - * weights[in] weights of testLayer - * state[in] state of testLayer - * cost[in] input cost - * callbackCount[in] number of done callback - * maxDiff[in/out] max of all previous diff - * testLayer[in/out] testLayer - * dataLayers[in/out] dataLayers - */ -void testPerturbInput(TestConfig testConf, - const MatrixPtr weights, - const LayerStatePtr state, - real cost, - real callbackCount, - real* maxDiff, - LayerPtr testLayer, - std::vector dataLayers); - -void testLayerGradKernel(TestConfig testConf, - string testLayerName, - size_t batchSize, - bool trans, - bool useGpu, - bool useWeight = false, - float epsilon = 0.02); - -void testLayerGrad(TestConfig testConf, - string testLayerName, - size_t batchSize, - bool trans, - bool useGpu, - bool useWeight = false, - float epsilon = 0.02); - -void testProjectionGrad(ProjectionConfig conf, - InputType inputType, - size_t parameterSize, - size_t batchSize, - bool useGpu, - bool testState = false, - int biasSize = 0, - bool sharedBias = false); - -void testOperatorGrad(TestConfig& config, - OperatorConfig& operatorConf, - size_t batchSize, - bool useGpu, - bool testState = false); - -} // namespace paddle diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp deleted file mode 100644 index d2a9761a4e16832a0722d4375cc11adb42524a8c..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/MKLDNNTester.cpp +++ /dev/null @@ -1,580 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MKLDNNTester.h" -#include "paddle/gserver/layers/MKLDNNBase.h" -#include "paddle/gserver/layers/MKLDNNLayer.h" -#include "paddle/trainer/Trainer.h" - -namespace paddle { - -// init data layer and test layer of both dnn and reference -void MKLDNNTester::reset(const TestConfig& dnn, - const TestConfig& ref, - size_t batchSize) { - const bool trans = false; - const bool useGpu = false; - - // clear - configs_.clear(); - layerNames_.clear(); - dataLayers_.clear(); - datas_.clear(); - layerMaps_.clear(); - parameters_.clear(); - testLayers_.clear(); - - // resize - configs_.resize(NUM); - layerNames_.resize(NUM); - dataLayers_.resize(NUM); - datas_.resize(NUM); - layerMaps_.resize(NUM); - parameters_.resize(NUM); - testLayers_.resize(NUM); - - // reset configs and layer names - configs_[DNN] = dnn; - configs_[REF] = ref; - layerNames_[DNN] = "mkldnn"; // the first is mkldnn layer - layerNames_[REF] = "reference"; // second is reference layer - - // reset others - for (size_t i = 0; i < NUM; ++i) { - configs_[i].layerConfig.set_name(layerNames_[i]); - initDataLayer(configs_[i], - &(dataLayers_[i]), - &(datas_[i]), - &(layerMaps_[i]), - layerNames_[i], - batchSize, - trans, - useGpu); - initTestLayer( - configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i])); - } - refLayer_ = testLayers_[REF]; - dnnLayer_ = testLayers_[DNN]; - EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size()); - EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); - setInputImgSize(); - - // for comparison with Paddle reference results, - // need manually add cpu device output for test - MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(dnnLayer_); - if (dnnLayer) { - dnnLayer->addOutputArgument(CPU_DEVICE); - } -} - -void MKLDNNTester::setInputImgSize() { - for (size_t n = 0; n < dataLayers_.size(); ++n) { - for (size_t i = 0; i < dataLayers_[n].size(); ++i) { - // TODO(TJ): fix me when concat and elewise ready - dataLayers_[n][i]->getOutput().setFrameHeight(ih_); - dataLayers_[n][i]->getOutput().setFrameWidth(iw_); - } - } -} - -// init randome parameters of ref, and copy to mkldnn -void MKLDNNTester::randomWgtDatas() { - EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); - const bool isBN = refLayer_->getType() == "batch_norm"; - for (size_t i = 0; i < parameters_[REF].size(); ++i) { - const VectorPtr& dnnValue = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); - const VectorPtr& refValue = parameters_[REF][i]->getBuf(PARAMETER_VALUE); - parameters_[REF][i]->randomize(); - if (isBN && i == 2) { - // this param is moving average in batch norm, which must larger than 0 - real offset = fabs(refValue->getMin()) + 1.0; - refValue->add(offset); - } - dnnValue->copyFrom(*refValue); - - VLOG(MKLDNN_TESTS) << "Random weight " << parameters_[DNN][i]->getName(); - printVector(dnnValue); - } -} - -// random botdata of ref layer and copy same to mkldnn -void MKLDNNTester::randomBotDatas() { - CHECK_EQ(dataLayers_.size(), NUM); - for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) { - dataLayers_[REF][i]->getOutputValue()->randomizeUniform(); - dataLayers_[DNN][i]->getOutputValue()->copyFrom( - *(dataLayers_[REF][i]->getOutputValue())); - VLOG(MKLDNN_TESTS) << "Random Foward, InputValue " << i; - printMatrix(dataLayers_[REF][i]->getOutputValue()); - } -} - -void MKLDNNTester::randomTopDiffs() { - refLayer_->getOutputGrad()->randomizeUniform(); - dnnLayer_->getOutput(CPU_DEVICE) - .grad->copyFrom(*(refLayer_->getOutputGrad())); - VLOG(MKLDNN_TESTS) << "Random Backward, OutputGrad"; - printMatrix(refLayer_->getOutputGrad()); -} - -void MKLDNNTester::checkForward() { - VLOG(MKLDNN_TESTS) << "Check Forward"; - printTopDatas(); - double delta = - compareMatrix(refLayer_->getOutputValue(), dnnLayer_->getOutputValue()); - EXPECT_LE(fabs(delta), eps_); -} - -void MKLDNNTester::checkBackwardData() { - VLOG(MKLDNN_TESTS) << "Check Backward Data"; - const bool isBN = refLayer_->getType() == "batch_norm"; - for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) { - const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad(); - const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad(); - VLOG(MKLDNN_ALL) << "MKLDNN Backward Result: InputGrad " << i; - printMatrix(dnnDiff); - VLOG(MKLDNN_ALL) << "Reference Backward Result: InputGrad " << i; - printMatrix(refDiff); - - double delta = compareMatrix(refDiff, dnnDiff); - EXPECT_LE(fabs(delta), eps_); - if (isBN) { - // the other two inputs in batch norm are for moving mean and var - // do not have grad to compare - break; - } - } -} - -void MKLDNNTester::checkBackwardWgts() { - VLOG(MKLDNN_TESTS) << "Check Backward Weight"; - CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); - vector dnnWgts; // used to temply save mkldnn weights - saveWgt(parameters_[DNN], dnnWgts); - - MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(dnnLayer_); - if (dnnLayer) { - dnnLayer->convertWeightsToPaddle(); - } - for (size_t i = 0; i < parameters_[DNN].size(); ++i) { - const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); - const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE); - VLOG(MKLDNN_ALL) << "MKLDNN Result: weight value" - << parameters_[DNN][i]->getName(); - printVector(dnn); - VLOG(MKLDNN_ALL) << "Reference Result: weight value " - << parameters_[REF][i]->getName(); - printVector(ref); - - double delta = compareVector(ref, dnn); - EXPECT_LE(fabs(delta), eps_); - } - - VLOG(MKLDNN_ALL) << "Restore dnn weights before comapre"; - restoreWgt(dnnWgts, parameters_[DNN]); -} - -void MKLDNNTester::saveWgt(const vector& from, - vector& to) { - const bool useGpu = false; - to.resize(from.size()); - for (size_t i = 0; i < to.size(); ++i) { - const VectorPtr& wgt = from[i]->getBuf(PARAMETER_VALUE); - to[i] = Vector::create(wgt->getSize(), useGpu); - to[i]->copyFrom(*wgt); - } -} - -void MKLDNNTester::restoreWgt(const vector& from, - vector& to) { - CHECK_EQ(from.size(), to.size()); - for (size_t i = 0; i < from.size(); ++i) { - const VectorPtr& wgt = to[i]->getBuf(PARAMETER_VALUE); - wgt->copyFrom(*from[i]); - } -} - -// clear parameters grad -void MKLDNNTester::clearWgtDiffs(size_t id) { - CHECK_LE(id, parameters_.size()); - for (size_t n = 0; n < parameters_.size(); ++n) { - if (id == n || id == parameters_.size()) { - for (size_t i = 0; i < parameters_[n].size(); ++i) { - const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT); - if (grad) { - grad->zeroMem(); - } - } - } - } -} - -void MKLDNNTester::clearBotDiffs(size_t id) { - CHECK_LE(id, dataLayers_.size()); - for (size_t n = 0; n < dataLayers_.size(); ++n) { - if (id == n || id == dataLayers_.size()) { - // clear inputs layers of this specific layer - for (size_t i = 0; i < dataLayers_[n].size(); ++i) { - dataLayers_[n][i]->getOutputGrad()->zeroMem(); - } - } - } -} - -void MKLDNNTester::clearTopDatas(size_t id) { - CHECK_LE(id, testLayers_.size()); - for (size_t i = 0; i < testLayers_.size(); ++i) { - if (id == i || id == testLayers_.size()) { - testLayers_[i]->getOutputValue()->zeroMem(); - } - } -} - -void MKLDNNTester::printTopDatas() { - if (!log_) { - return; - } - - for (int n = 0; n < NUM; ++n) { - VLOG(MKLDNN_ALL) << testLayers_[n]->getType() - << " Forward Result: OutputValue"; - printMatrix(testLayers_[n]->getOutputValue()); - } -} - -void MKLDNNTester::printMatrix(const MatrixPtr& m) { - if (!log_) { - return; - } - - std::ostringstream ostr; - m->print(ostr); - VLOG(MKLDNN_ALL) << std::endl << ostr.str(); -} - -void MKLDNNTester::printVector(const VectorPtr& v) { - if (!log_) { - return; - } - - std::ostringstream ostr; - v->print(ostr, v->getSize()); - VLOG(MKLDNN_ALL) << std::endl << ostr.str(); -} - -double MKLDNNTester::getDelta(const real* refer, - const real* value, - size_t len, - const float failRate, - const float thres) { - double delta = 0, sum = 0; - int failCnt = 0; - const double eps = 1e-5; - double maxRatio = 0; - for (size_t i = 0; i < len; ++i) { - double ref = fabs(refer[i]); - double val = fabs(value[i]); - double diff = fabs(refer[i] - value[i]); - delta += diff; - sum += ref; - if (ref < eps && val < eps) { // both values are very small - continue; - } - double ratio = diff / ref; - if (ratio > thres) { - maxRatio = std::max(maxRatio, ratio); - failCnt++; - } - } - EXPECT_FALSE(std::isinf(sum)); - EXPECT_FALSE(std::isnan(sum)); - EXPECT_FALSE(std::isnan(delta)); - VLOG(MKLDNN_ALL) << "reference avg data: " << sum / len - << ", delta: " << delta / sum << ", failCnt:" << failCnt; - double res = sum > eps ? delta / sum : eps; - return (failCnt / (float)len) > failRate ? maxRatio : res; -} - -double MKLDNNTester::compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2) { - CHECK_EQ(m1->getElementCnt(), m2->getElementCnt()); - return getDelta(m1->getData(), m2->getData(), m1->getElementCnt()); -} - -double MKLDNNTester::compareVector(const VectorPtr& v1, const VectorPtr& v2) { - CHECK_EQ(v1->getSize(), v2->getSize()); - return getDelta(v1->getData(), v2->getData(), v1->getSize()); -} - -void MKLDNNTester::runOnce() { - // test forward - randomBotDatas(); - dnnLayer_->forward(passType_); - refLayer_->forward(passType_); - checkForward(); - - if (passType_ == PASS_TEST) { - return; - } - - // test backward - // simple updater - UpdateCallback updateCallback = [](Parameter* para) { - auto& grad = para->getBuf(PARAMETER_GRADIENT); - auto& value = para->getBuf(PARAMETER_VALUE); - real lr = 1e-2; - value->add(*grad, lr); - grad->zeroMem(); - }; - randomTopDiffs(); - dnnLayer_->backward(updateCallback); - refLayer_->backward(updateCallback); - checkBackwardData(); - checkBackwardWgts(); - - // clear buffers - // ref code will addto the diff, dnn code will writeto it - // and clearTopDatas(REF) should be coverd by ref layers - clearBotDiffs(REF); - clearWgtDiffs(REF); - // it is necessary to clear bottom diffs when only activation is dnn type - if (configs_[DNN].layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) { - clearBotDiffs(DNN); - } -} - -void MKLDNNTester::run(const TestConfig& dnn, - const TestConfig& ref, - size_t batchSize, - size_t inputImgH, - size_t inputImgW, - PassType passType, - bool printDetails, - size_t iter, - float epsilon) { - CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 || - dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) - << "should be MKLDNN layer or MKLDNN activation"; - if (dnn.layerConfig.type() == ref.layerConfig.type()) { - VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " - << dnn.layerConfig.active_type() << " vs " - << ref.layerConfig.active_type(); - } else { - VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " - << dnn.layerConfig.type() << " vs " - << ref.layerConfig.type(); - } - - ih_ = inputImgH; - iw_ = inputImgW; - passType_ = passType; - log_ = printDetails; - iter_ = iter; - eps_ = epsilon; - - // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight - reset(dnn, ref, batchSize); - randomWgtDatas(); - clearWgtDiffs(); - clearBotDiffs(); - for (size_t i = 0; i < iter_; ++i) { - VLOG(MKLDNN_TESTS) << "Check Iteration " << i; - runOnce(); - } - - if (parameters_[DNN].empty()) { - // has no paramters - return; - } - - // After run some iterations, the mkldnn weight has been stored in dnnLayer - // and we can also get the mkldnn weight parameter header format. - // Weight parameter should always be index 0 (and bias index 1). - // TODO(TJ): should also consider mean and var format when batchnorm ready - int dnnWgtFmt = parameters_[DNN][0]->getHeaderFormat(); - int refWgtFmt = parameters_[REF][0]->getHeaderFormat(); - if (dnnWgtFmt == refWgtFmt) { - // weight format are equal, so no need check more - return; - } - - // then save the weights and restart again - vector dnnWgts, refWgts; - CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); - saveWgt(parameters_[DNN], dnnWgts); - saveWgt(parameters_[REF], refWgts); - - // restart again with dnn weight format - reset(dnn, ref, batchSize); - // TODO(TJ): should also considerate mean and var format when batchnorm ready - parameters_[DNN][0]->setHeaderFormat(dnnWgtFmt); - - // restore wgt - restoreWgt(dnnWgts, parameters_[DNN]); - restoreWgt(refWgts, parameters_[REF]); - clearWgtDiffs(); - clearBotDiffs(); - - for (size_t i = 0; i < iter_; ++i) { - VLOG(MKLDNN_TESTS) << "Check Iteration " << i; - runOnce(); - } -} - -void MKLDNNTester::initArgument(DataIn& data, - const std::string& configPath, - const size_t iter) { - TrainerConfigHelper config(configPath); - size_t batchSize = config.getOptConfig().batch_size(); - data.inArgs.resize(iter); - data.outGrads.resize(iter); - data.paraValues.clear(); - for (const auto& layer_name : config.getModelConfig().input_layer_names()) { - auto layer_config = std::find_if(config.getModelConfig().layers().begin(), - config.getModelConfig().layers().end(), - [=](const LayerConfig& layer_config) { - return layer_config.name() == layer_name; - }); - CHECK(layer_config != config.getModelConfig().layers().end()); - - size_t layerSize = layer_config->size(); - for (size_t i = 0; i < iter; ++i) { - Argument arg; - arg.value = Matrix::create(batchSize, layerSize, false, false); - arg.grad = Matrix::create(batchSize, layerSize, false, false); - arg.value->randomizeUniform(); - arg.value->add(-0.5); - arg.value->sigmoid(*arg.value); - arg.grad->zeroMem(); - arg.ids = VectorT::create(batchSize, false); - arg.ids->rand(layerSize); - generateSequenceStartPositions(batchSize, arg.sequenceStartPositions); - data.inArgs[i].push_back(arg); - } - } - - for (const auto& layer_name : config.getModelConfig().output_layer_names()) { - auto layer_config = std::find_if(config.getModelConfig().layers().begin(), - config.getModelConfig().layers().end(), - [=](const LayerConfig& layer_config) { - return layer_config.name() == layer_name; - }); - CHECK(layer_config != config.getModelConfig().layers().end()); - - size_t layerSize = layer_config->size(); - for (size_t i = 0; i < iter; ++i) { - MatrixPtr grad = Matrix::create(batchSize, layerSize, false, false); - grad->randomizeUniform(); - data.outGrads[i].push_back(grad); - } - } - - for (const auto& para_config : config.getModelConfig().parameters()) { - VectorPtr value = Vector::create(para_config.size(), false); - value->randnorm(0, 2); - data.paraValues.push_back(value); - } -} - -void MKLDNNTester::getOutResult(const std::string& configPath, - DataIn& in, - DataOut& out, - bool use_mkldnn, - size_t iter) { - FLAGS_use_gpu = false; - FLAGS_use_mkldnn = use_mkldnn; - *ThreadLocalRand::getSeed() = 1; - srand(1); - - Trainer trainer; - auto config = std::make_shared(configPath); - trainer.init(config, false); - auto gradientMachine = trainer.getGradientMachine(); - std::vector parameters = gradientMachine->getParameters(); - for (size_t i = 0; i < in.paraValues.size(); i++) { - parameters[i]->getBuf(PARAMETER_VALUE)->copyFrom(*in.paraValues[i]); - } - UpdateCallback simpleUpdate = [](Parameter* para) { - auto& grad = para->getBuf(PARAMETER_GRADIENT); - auto& value = para->getBuf(PARAMETER_VALUE); - real lr = 1e-2; - value->add(*grad, lr); - grad->zeroMem(); - }; - - vector outArgs; - gradientMachine->start(); - out.outValues.clear(); - out.paraValues.clear(); - for (size_t i = 0; i < iter; ++i) { - VLOG(MKLDNN_TESTS) << "runing iteration " << i; - gradientMachine->forward(in.inArgs[i], &outArgs, PASS_TRAIN); - // save forward result - for (size_t k = 0; k < outArgs.size(); k++) { - const MatrixPtr& src = outArgs[k].value; - MatrixPtr dst = - Matrix::create(src->getHeight(), src->getWidth(), false, false); - if (typeid(*src) == typeid(MKLDNNMatrix)) { - MKLDNNMatrixPtr dnnSrc = std::dynamic_pointer_cast(src); - dnnSrc->copyTo(*dst); - } else { - dst->copyFrom(*src); - } - out.outValues.push_back(dst); - } - - // random backward input - for (size_t k = 0; k < outArgs.size(); k++) { - outArgs[k].grad->copyFrom(*in.outGrads[i][k]); - } - gradientMachine->backward(simpleUpdate); - } - gradientMachine->finish(); - - // save param value - for (size_t i = 0; i < in.paraValues.size(); i++) { - VectorPtr val = Vector::create( - parameters[i]->getBuf(PARAMETER_VALUE)->getSize(), false); - val->copyFrom(*parameters[i]->getBuf(PARAMETER_VALUE)); - out.paraValues.push_back(val); - } -} - -void MKLDNNTester::compareResult(DataOut& ref, DataOut& dnn, float eps) { - CHECK_EQ(ref.outValues.size(), dnn.outValues.size()); - CHECK_EQ(ref.paraValues.size(), dnn.paraValues.size()); - for (size_t i = 0; i < ref.outValues.size(); i++) { - VLOG(MKLDNN_TESTS) << "compare value index: " << i; - EXPECT_LE(fabs(compareMatrix(ref.outValues[i], dnn.outValues[i])), eps); - } - for (size_t i = 0; i < ref.paraValues.size(); i++) { - VLOG(MKLDNN_TESTS) << "compare param index: " << i; - EXPECT_LE(fabs(compareVector(ref.paraValues[i], dnn.paraValues[i])), eps); - } -} - -void MKLDNNTester::runNetTest(const std::string& configPath, - size_t iter, - float eps) { - DataIn in; - initArgument(in, configPath, iter); - DataOut outCpu, outDnn; - VLOG(MKLDNN_TESTS) << "runing cpu network"; - getOutResult(configPath, in, outCpu, false, iter); - VLOG(MKLDNN_TESTS) << "runing mkldnn network"; - getOutResult(configPath, in, outDnn, true, iter); - - compareResult(outCpu, outDnn, eps); -} - -} // namespace paddle diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h deleted file mode 100644 index 41ac46b70ab08d4071f4e6abfca94667268015d7..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/MKLDNNTester.h +++ /dev/null @@ -1,143 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "LayerGradUtil.h" -#include "paddle/gserver/layers/MKLDNNBase.h" -#include "paddle/gserver/layers/MKLDNNLayer.h" - -namespace paddle { - -/** - * @brief test the functionality of MKLDNNlayers and MKLDNNActivations - * refer to paddle original function - */ -class MKLDNNTester { - enum { - DNN = 0, // MKLDNN layer - REF = 1, // Reference layer - NUM = 2, // Number of total - }; - - struct DataIn { - std::vector> inArgs; - std::vector> outGrads; - std::vector paraValues; - }; - - struct DataOut { - std::vector outValues; - std::vector paraValues; - }; - - protected: - std::vector configs_; - vector layerNames_; - vector> dataLayers_; - vector> datas_; - vector layerMaps_; - vector> parameters_; - vector testLayers_; - LayerPtr refLayer_, dnnLayer_; - - /// run some iterations, all the result should pass - size_t iter_; - /// whether to print out the details - bool log_; - /// epsilon - float eps_; - /// input image size, default 1 - size_t ih_, iw_; - /// passType, PASS_TRAIN, PASS_TEST or PASS_GC (Gradient Check pass) - PassType passType_; - - public: - explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) { - iter_ = iter; - eps_ = epsilon; - log_ = false; - passType_ = PASS_TRAIN; - } - - ~MKLDNNTester() {} - - public: - void run(const TestConfig& dnn, - const TestConfig& ref, - size_t batchSize, - size_t inputImgH = 1, - size_t inputImgW = 1, - PassType passType = PASS_TRAIN, - bool printDetails = false, - size_t iter = 3, - float epsilon = 1e-4); - static void runNetTest(const std::string& configPath, - size_t iter = 2, - float eps = 1e-4); - static void initArgument(DataIn& data, - const std::string& configPath, - size_t iter = 2); - static void getOutResult(const std::string& configPath, - DataIn& in, - DataOut& out, - bool use_mkldnn, - size_t iter = 2); - - private: - void reset(const TestConfig& dnn, const TestConfig& ref, size_t batchSize); - void setInputImgSize(); - void runOnce(); - - void randomWgtDatas(); - void randomBotDatas(); - void randomTopDiffs(); - - void checkForward(); - void checkBackwardData(); - void checkBackwardWgts(); - - // clear specific layer, clear all when id equals NUM - void clearWgtDiffs(size_t id = NUM); - void clearBotDiffs(size_t id = NUM); - void clearTopDatas(size_t id = NUM); - - void printTopDatas(); - void printMatrix(const MatrixPtr& m); - void printVector(const VectorPtr& v); - - void saveWgt(const vector& from, vector& to); - void restoreWgt(const vector& from, vector& to); - - static double compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2); - static double compareVector(const VectorPtr& v1, const VectorPtr& v2); - static void compareResult(DataOut& ref, DataOut& dnn, float eps = 1e-4); - - /** - * Get delta percent - * if many(>failRate) wrong(abs(val-ref)/abs(ref) > thres) points - * return the max(diff/ref) - * else return sum(abs(diff)) / sum(abs(ref)) - * The return value should be smaller than eps when passing. - */ - static double getDelta(const real* refer, - const real* value, - size_t len, - const float failRate = 1e-3, - const float thres = 0.1); -}; - -} // namespace paddle diff --git a/paddle/gserver/tests/Sequence/train.list b/paddle/gserver/tests/Sequence/train.list deleted file mode 100644 index be27acb3a5411d8fe65797079a9a5977c1f0f90a..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/Sequence/train.list +++ /dev/null @@ -1 +0,0 @@ -gserver/tests/Sequence/tour_train_wdseg diff --git a/paddle/gserver/tests/Sequence/train.list.nest b/paddle/gserver/tests/Sequence/train.list.nest deleted file mode 100644 index 7683ebc68efbb07ce01d8faab14574109df99af9..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/Sequence/train.list.nest +++ /dev/null @@ -1 +0,0 @@ -gserver/tests/Sequence/tour_train_wdseg.nest diff --git a/paddle/gserver/tests/sequence_layer_group.conf b/paddle/gserver/tests/sequence_layer_group.conf deleted file mode 100644 index 50f2d89d0271b2eaa460e57636eb09b6d6aeda18..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_layer_group.conf +++ /dev/null @@ -1,62 +0,0 @@ -#!/usr/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict' -dict_file = dict() -for line_count, line in enumerate(open(dict_path, "r")): - dict_file[line.strip()] = line_count - -define_py_data_sources2( - train_list='gserver/tests/Sequence/train.list', - test_list=None, - module='sequenceGen', - obj='process', - args={"dict_file": dict_file}) - -settings(batch_size=5) -######################## network configure ################################ -dict_dim = len(open(dict_path, 'r').readlines()) -word_dim = 128 -hidden_dim = 256 -label_dim = 3 - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer(input=data, size=word_dim) - -# (lstm_input + lstm) is equal to lstmemory -with mixed_layer(size=hidden_dim * 4) as lstm_input: - lstm_input += full_matrix_projection(input=emb) - -lstm = lstmemory_group( - input=lstm_input, - size=hidden_dim, - act=TanhActivation(), - gate_act=SigmoidActivation(), - state_act=TanhActivation()) - -lstm_last = last_seq(input=lstm) - -with mixed_layer( - size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: - output += full_matrix_projection(input=lstm_last) - -outputs( - classification_cost( - input=output, label=data_layer( - name="label", size=1))) diff --git a/paddle/gserver/tests/sequence_lstm.conf b/paddle/gserver/tests/sequence_lstm.conf deleted file mode 100644 index f49a827f22edce056eaf9903e99b732cab7f3784..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_lstm.conf +++ /dev/null @@ -1,64 +0,0 @@ -#!/usr/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict' -dict_file = dict() -for line_count, line in enumerate(open(dict_path, "r")): - dict_file[line.strip()] = line_count - -define_py_data_sources2( - train_list='gserver/tests/Sequence/train.list', - test_list=None, - module='sequenceGen', - obj='process', - args={"dict_file": dict_file}) - -settings(batch_size=5) -######################## network configure ################################ -dict_dim = len(open(dict_path, 'r').readlines()) -word_dim = 128 -hidden_dim = 256 -label_dim = 3 -sparse_update = get_config_arg("sparse_update", bool, False) - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer( - input=data, - size=word_dim, - param_attr=ParamAttr(sparse_update=sparse_update)) - -with mixed_layer(size=hidden_dim * 4) as lstm_input: - lstm_input += full_matrix_projection(input=emb) - -lstm = lstmemory( - input=lstm_input, - act=TanhActivation(), - gate_act=SigmoidActivation(), - state_act=TanhActivation()) - -lstm_last = last_seq(input=lstm) - -with mixed_layer( - size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: - output += full_matrix_projection(input=lstm_last) - -outputs( - classification_cost( - input=output, label=data_layer( - name="label", size=1))) diff --git a/paddle/gserver/tests/sequence_nest_layer_group.conf b/paddle/gserver/tests/sequence_nest_layer_group.conf deleted file mode 100644 index 71ef53d08a2cea070806afb2c65ef15c4dd28f31..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_nest_layer_group.conf +++ /dev/null @@ -1,83 +0,0 @@ -#!/usr/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict' -dict_file = dict() -for line_count, line in enumerate(open(dict_path, "r")): - dict_file[line.strip()] = line_count - -define_py_data_sources2( - train_list='gserver/tests/Sequence/train.list.nest', - test_list=None, - module='sequenceGen', - obj='process2', - args={"dict_file": dict_file}) - -settings(batch_size=2) -######################## network configure ################################ -dict_dim = len(open(dict_path, 'r').readlines()) -word_dim = 128 -hidden_dim = 256 -label_dim = 3 - -data = data_layer(name="word", size=dict_dim) - -emb_group = embedding_layer(input=data, size=word_dim) - - -# (lstm_input + lstm) is equal to lstmemory -def lstm_group(lstm_group_input): - with mixed_layer(size=hidden_dim * 4) as group_input: - group_input += full_matrix_projection(input=lstm_group_input) - - lstm_output = lstmemory_group( - input=group_input, - name="lstm_group", - size=hidden_dim, - act=TanhActivation(), - gate_act=SigmoidActivation(), - state_act=TanhActivation()) - return lstm_output - - -lstm_nest_group = recurrent_group( - input=SubsequenceInput(emb_group), step=lstm_group, name="lstm_nest_group") -# hasSubseq ->(seqlastins) seq -lstm_last = last_seq( - input=lstm_nest_group, agg_level=AggregateLevel.TO_SEQUENCE) - -# seq ->(expand) hasSubseq -lstm_expand = expand_layer( - input=lstm_last, - expand_as=emb_group, - expand_level=ExpandLevel.FROM_SEQUENCE) - -# hasSubseq ->(average) seq -lstm_average = pooling_layer( - input=lstm_expand, - pooling_type=AvgPooling(), - agg_level=AggregateLevel.TO_SEQUENCE) - -with mixed_layer( - size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: - output += full_matrix_projection(input=lstm_average) - -outputs( - classification_cost( - input=output, label=data_layer( - name="label", size=1))) diff --git a/paddle/gserver/tests/sequence_nest_rnn.conf b/paddle/gserver/tests/sequence_nest_rnn.conf deleted file mode 100644 index 2873a599669b4281a53cd71e8bb56f0d18c26b5a..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_nest_rnn.conf +++ /dev/null @@ -1,74 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2(train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_subseq') - - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 8 -hidden_dim = 8 -label_dim = 3 - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer(input=data, size=word_dim) - -# This hierachical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn.conf - -def outer_step(x): - outer_mem = memory(name="outer_rnn_state", size=hidden_dim) - def inner_step(y): - inner_mem = memory(name="inner_rnn_state", - size=hidden_dim, - boot_layer=outer_mem) - out = fc_layer(input=[y, inner_mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name="inner_rnn_state") - return out - - inner_rnn_output = recurrent_group( - step=inner_step, - name="inner", - input=x) - last = last_seq(input=inner_rnn_output, name="outer_rnn_state") - - # "return last" won't work, because recurrent_group only support the input - # sequence type is same as return sequence type. - return inner_rnn_output - -out = recurrent_group( - name="outer", - step=outer_step, - input=SubsequenceInput(emb)) - -rep = last_seq(input=out) -prob = fc_layer(size=label_dim, - input=rep, - act=SoftmaxActivation(), - bias_attr=True) - -outputs(classification_cost(input=prob, - label=data_layer(name="label", size=label_dim))) diff --git a/paddle/gserver/tests/sequence_nest_rnn_multi_input.conf b/paddle/gserver/tests/sequence_nest_rnn_multi_input.conf deleted file mode 100644 index afdacfffd7aecfe2f4762f04a987126381bcea34..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_nest_rnn_multi_input.conf +++ /dev/null @@ -1,76 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2(train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_subseq') - - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 8 -hidden_dim = 8 -label_dim = 3 - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer(input=data, size=word_dim) - -# This hierachical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn.conf - -def outer_step(wid, x): - outer_mem = memory(name="outer_rnn_state", size=hidden_dim) - def inner_step(y, wid): - z = embedding_layer(input=wid, size=word_dim) - inner_mem = memory(name="inner_rnn_state", - size=hidden_dim, - boot_layer=outer_mem) - out = fc_layer(input=[y, z, inner_mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name="inner_rnn_state") - return out - - inner_rnn_output = recurrent_group( - step=inner_step, - name="inner", - input=[x, wid]) - last = last_seq(input=inner_rnn_output, name="outer_rnn_state") - - # "return last" should also work. But currently RecurrentGradientMachine - # does not handle it, and will report error: In hierachical RNN, all out - # links should be from sequences now. - return inner_rnn_output - -out = recurrent_group( - name="outer", - step=outer_step, - input=[SubsequenceInput(data), SubsequenceInput(emb)]) - -rep = last_seq(input=out) -prob = fc_layer(size=label_dim, - input=rep, - act=SoftmaxActivation(), - bias_attr=True) - -outputs(classification_cost(input=prob, - label=data_layer(name="label", size=label_dim))) diff --git a/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py b/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py deleted file mode 100644 index 569d3c094b6f5517dad0f1e04f98de12aaef9633..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py +++ /dev/null @@ -1,96 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2( - train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_unequalength_subseq') - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 8 -hidden_dim = 8 -label_dim = 2 - -speaker1 = data_layer(name="word1", size=dict_dim) -speaker2 = data_layer(name="word2", size=dict_dim) - -emb1 = embedding_layer(input=speaker1, size=word_dim) -emb2 = embedding_layer(input=speaker2, size=word_dim) - - -# This hierarchical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn_multi_unequalength_inputs.conf -def outer_step(x1, x2): - index = [0] - - def inner_step(ipt): - index[0] += 1 - i = index[0] - outer_mem = memory(name="outer_rnn_state_%d" % i, size=hidden_dim) - - def inner_step_impl(y): - inner_mem = memory( - name="inner_rnn_state_" + y.name, - size=hidden_dim, - boot_layer=outer_mem) - out = fc_layer( - input=[y, inner_mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name='inner_rnn_state_' + y.name) - return out - - encoder = recurrent_group( - step=inner_step_impl, name='inner_%d' % i, input=ipt) - last = last_seq(name="outer_rnn_state_%d" % i, input=encoder) - return encoder, last - - encoder1, sentence_last_state1 = inner_step(ipt=x1) - encoder2, sentence_last_state2 = inner_step(ipt=x2) - - encoder1_expand = expand_layer( - input=sentence_last_state1, expand_as=encoder2) - - return [encoder1_expand, encoder2] - - -encoder1_rep, encoder2_rep = recurrent_group( - name="outer", - step=outer_step, - input=[SubsequenceInput(emb1), SubsequenceInput(emb2)], - targetInlink=emb2) - -encoder1_last = last_seq(input=encoder1_rep) -encoder1_expandlast = expand_layer(input=encoder1_last, expand_as=encoder2_rep) -context = mixed_layer( - input=[ - identity_projection(encoder1_expandlast), - identity_projection(encoder2_rep) - ], - size=hidden_dim) - -rep = last_seq(input=context) -prob = fc_layer( - size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) - -outputs( - classification_cost( - input=prob, label=data_layer( - name="label", size=label_dim))) diff --git a/paddle/gserver/tests/sequence_recurrent.py b/paddle/gserver/tests/sequence_recurrent.py deleted file mode 100644 index b88c09084e1bc167a177b59566e9794ac4d616c7..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_recurrent.py +++ /dev/null @@ -1,55 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict' -dict_file = dict() -for line_count, line in enumerate(open(dict_path, "r")): - dict_file[line.strip()] = line_count - -define_py_data_sources2( - train_list='gserver/tests/Sequence/train.list', - test_list=None, - module='sequenceGen', - obj='process', - args={"dict_file": dict_file}) - -settings(batch_size=5) -######################## network configure ################################ -dict_dim = len(open(dict_path, 'r').readlines()) -word_dim = 128 -hidden_dim = 128 -label_dim = 3 - -# This config is designed to be equivalent with sequence_recurrent_group.py - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer( - input=data, size=word_dim, param_attr=ParamAttr(name="emb")) - -recurrent = recurrent_layer(input=emb, bias_attr=False, act=SoftmaxActivation()) - -recurrent_last = last_seq(input=recurrent) - -with mixed_layer( - size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: - output += full_matrix_projection(input=recurrent_last) - -outputs( - classification_cost( - input=output, label=data_layer( - name="label", size=1))) diff --git a/paddle/gserver/tests/sequence_recurrent_group.py b/paddle/gserver/tests/sequence_recurrent_group.py deleted file mode 100644 index 0daf746700231d302550004b1c10729e36807b8b..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_recurrent_group.py +++ /dev/null @@ -1,68 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -from paddle.trainer_config_helpers import * - -######################## data source ################################ -dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict' -dict_file = dict() -for line_count, line in enumerate(open(dict_path, "r")): - dict_file[line.strip()] = line_count - -define_py_data_sources2( - train_list='gserver/tests/Sequence/train.list', - test_list=None, - module='sequenceGen', - obj='process', - args={"dict_file": dict_file}) - -settings(batch_size=5) -######################## network configure ################################ -dict_dim = len(open(dict_path, 'r').readlines()) -word_dim = 128 -hidden_dim = 128 -label_dim = 3 - -# This config is designed to be equivalent with sequence_recurrent.py - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer( - input=data, size=word_dim, param_attr=ParamAttr(name="emb")) - - -def step(y): - mem = memory(name="rnn_state", size=hidden_dim) - with mixed_layer( - name="rnn_state", - size=hidden_dim, - bias_attr=False, - act=SoftmaxActivation()) as out: - out += identity_projection(input=y) - out += full_matrix_projection( - input=mem, param_attr=ParamAttr(name="___recurrent_layer_0__")) - return out - - -recurrent = recurrent_group(name="rnn", step=step, input=emb) - -recurrent_last = last_seq(input=recurrent) - -with mixed_layer( - size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: - output += full_matrix_projection(input=recurrent_last) - -outputs( - classification_cost( - input=output, label=data_layer( - name="label", size=1))) diff --git a/paddle/gserver/tests/sequence_rnn.conf b/paddle/gserver/tests/sequence_rnn.conf deleted file mode 100644 index 1084edfe708c3348d40b67e270f64d8cda3cee0f..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_rnn.conf +++ /dev/null @@ -1,57 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2(train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_seq') - - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 8 -hidden_dim = 8 -label_dim = 3 - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer(input=data, size=word_dim) - -def step(y): - mem = memory(name="rnn_state", size=hidden_dim) - out = fc_layer(input=[y, mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name="rnn_state") - return out - -out = recurrent_group( - name="rnn", - step=step, - input=emb) - -rep = last_seq(input=out) -prob = fc_layer(size=label_dim, - input=rep, - act=SoftmaxActivation(), - bias_attr=True) - -outputs(classification_cost(input=prob, - label=data_layer(name="label", size=label_dim))) diff --git a/paddle/gserver/tests/sequence_rnn_matched_inputs.py b/paddle/gserver/tests/sequence_rnn_matched_inputs.py deleted file mode 100644 index 41a581e0ccd59588d1bcce9345056bea9d80b73d..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_rnn_matched_inputs.py +++ /dev/null @@ -1,84 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2( - train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_mixed') - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 2 -hidden_dim = 2 -label_dim = 2 - -data1 = data_layer(name="word1", size=dict_dim) -data2 = data_layer(name="word2", size=dict_dim) -label = data_layer(name="label", size=label_dim) - -encoding = embedding_layer(input=data2, size=word_dim) - -subseq = embedding_layer(input=data1, size=word_dim) -seq = embedding_layer(input=data2, size=word_dim) -nonseq = embedding_layer(input=label, size=word_dim) - - -# This hierarchical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn_mixed_inputs.conf -def outer_step(subseq, seq, nonseq, encoding): - outer_mem = memory(name="outer_rnn_state", size=hidden_dim) - - def inner_step(subseq, seq, nonseq): - inner_mem = memory( - name="inner_rnn_state", size=hidden_dim, boot_layer=outer_mem) - - out = fc_layer( - input=[subseq, seq, nonseq, inner_mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name='inner_rnn_state') - return out - - decoder = recurrent_group( - step=inner_step, name='inner', input=[subseq, seq, nonseq]) - last = last_seq(name="outer_rnn_state", input=decoder) - context = simple_attention( - encoded_sequence=encoding, encoded_proj=encoding, decoder_state=last) - return context - - -out = recurrent_group( - name="outer", - step=outer_step, - input=[ - subseq, expand_layer( - seq, expand_as=subseq, - expand_level=ExpandLevel.FROM_SEQUENCE), expand_layer( - nonseq, - expand_as=subseq, - expand_level=ExpandLevel.FROM_NO_SEQUENCE), - StaticInput(encoding) - ]) - -rep = last_seq(input=out) -prob = fc_layer( - size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) - -outputs(classification_cost(input=prob, label=label)) diff --git a/paddle/gserver/tests/sequence_rnn_mixed_inputs.py b/paddle/gserver/tests/sequence_rnn_mixed_inputs.py deleted file mode 100644 index ae89d8e2bb6f672eaf697ae4d24895b89f76544f..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_rnn_mixed_inputs.py +++ /dev/null @@ -1,78 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2( - train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_mixed') - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 2 -hidden_dim = 2 -label_dim = 2 - -data1 = data_layer(name="word1", size=dict_dim) -data2 = data_layer(name="word2", size=dict_dim) -label = data_layer(name="label", size=label_dim) - -encoding = embedding_layer(input=data2, size=word_dim) - - -# This hierarchical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn_matched_inputs.conf -def outer_step(subseq, seq, nonseq, encoding): - outer_mem = memory(name="outer_rnn_state", size=hidden_dim) - - def inner_step(data1, data2, label): - inner_mem = memory( - name="inner_rnn_state", size=hidden_dim, boot_layer=outer_mem) - - subseq = embedding_layer(input=data1, size=word_dim) - seq = embedding_layer(input=data2, size=word_dim) - nonseq = embedding_layer(input=label, size=word_dim) - - print_layer(input=[data1, seq, label, inner_mem]) - out = fc_layer( - input=[subseq, seq, nonseq, inner_mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name='inner_rnn_state') - return out - - decoder = recurrent_group( - step=inner_step, name='inner', - input=[subseq, StaticInput(seq), nonseq]) - last = last_seq(name="outer_rnn_state", input=decoder) - context = simple_attention( - encoded_sequence=encoding, encoded_proj=encoding, decoder_state=last) - return context - - -out = recurrent_group( - name="outer", - step=outer_step, - input=[data1, data2, StaticInput(label), StaticInput(encoding)]) - -rep = last_seq(input=out) -prob = fc_layer( - size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) - -outputs(classification_cost(input=prob, label=label)) diff --git a/paddle/gserver/tests/sequence_rnn_multi_input.conf b/paddle/gserver/tests/sequence_rnn_multi_input.conf deleted file mode 100644 index 9fae974f3079c49ad03d6ba34e30190f325414e8..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_rnn_multi_input.conf +++ /dev/null @@ -1,58 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2(train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_seq') - - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 8 -hidden_dim = 8 -label_dim = 3 - -data = data_layer(name="word", size=dict_dim) - -emb = embedding_layer(input=data, size=word_dim) - -def step(y, wid): - z = embedding_layer(input=wid, size=word_dim) - mem = memory(name="rnn_state", size=hidden_dim) - out = fc_layer(input=[y, z, mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name="rnn_state") - return out - -out = recurrent_group( - name="rnn", - step=step, - input=[emb, data]) - -rep = last_seq(input=out) -prob = fc_layer(size=label_dim, - input=rep, - act=SoftmaxActivation(), - bias_attr=True) - -outputs(classification_cost(input=prob, - label=data_layer(name="label", size=label_dim))) diff --git a/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py b/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py deleted file mode 100644 index 6473fb3f3eddc803282911a156c489e4ba39aded..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py +++ /dev/null @@ -1,76 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. - -from paddle.trainer_config_helpers import * - -######################## data source ################################ -define_py_data_sources2( - train_list='gserver/tests/Sequence/dummy.list', - test_list=None, - module='rnn_data_provider', - obj='process_unequalength_seq') - -settings(batch_size=2, learning_rate=0.01) -######################## network configure ################################ -dict_dim = 10 -word_dim = 8 -hidden_dim = 8 -label_dim = 2 - -speaker1 = data_layer(name="word1", size=dict_dim) -speaker2 = data_layer(name="word2", size=dict_dim) - -emb1 = embedding_layer(input=speaker1, size=word_dim) -emb2 = embedding_layer(input=speaker2, size=word_dim) - -# This hierachical RNN is designed to be equivalent to the RNN in -# sequence_nest_rnn_multi_unequalength_inputs.conf - - -def step(x1, x2): - def calrnn(y): - mem = memory(name='rnn_state_' + y.name, size=hidden_dim) - out = fc_layer( - input=[y, mem], - size=hidden_dim, - act=TanhActivation(), - bias_attr=True, - name='rnn_state_' + y.name) - return out - - encoder1 = calrnn(x1) - encoder2 = calrnn(x2) - return [encoder1, encoder2] - - -encoder1_rep, encoder2_rep = recurrent_group( - name="stepout", step=step, input=[emb1, emb2]) - -encoder1_last = last_seq(input=encoder1_rep) -encoder1_expandlast = expand_layer(input=encoder1_last, expand_as=encoder2_rep) -context = mixed_layer( - input=[ - identity_projection(encoder1_expandlast), - identity_projection(encoder2_rep) - ], - size=hidden_dim) - -rep = last_seq(input=context) -prob = fc_layer( - size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) - -outputs( - classification_cost( - input=prob, label=data_layer( - name="label", size=label_dim))) diff --git a/paddle/gserver/tests/test_ActivationGrad.cpp b/paddle/gserver/tests/test_ActivationGrad.cpp deleted file mode 100644 index b5e4af26dc123be3748adb4faed5fe1656ca44b3..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_ActivationGrad.cpp +++ /dev/null @@ -1,98 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_bool(thread_local_rand_use_global_seed); - -void testActivation(const string& act) { - LOG(INFO) << "test activation: " << act; - size_t size = 10; - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("addto"); - config.layerConfig.set_size(size); - config.layerConfig.set_active_type(act); - config.inputDefs.push_back({INPUT_DATA, "layer_0", size, 0}); - config.layerConfig.add_inputs(); - for (auto useGpu : {false, true}) { - testLayerGrad(config, - act + "_activation", - 100, - /* trans= */ false, - useGpu, - /* useWeight */ true); - } -} - -TEST(Activation, activation) { - auto types = ActivationFunction::getAllRegisteredTypes(); - std::set excluded{"sequence_softmax"}; - for (auto type : types) { - if (excluded.count(type)) continue; - testActivation(type); - } -} - -void testSequenceSoftmaxAct(bool hasSubseq) { - LOG(INFO) << "test activation: sequence softmax"; - - const size_t size = 1; - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("addto"); - config.layerConfig.set_size(size); - config.layerConfig.set_active_type("sequence_softmax"); - config.inputDefs.push_back( - {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, - "layer_0", - 1, - 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "sequence_softmax", - 100, - /* trans= */ false, - useGpu, - /* useWeight */ true); - } -} - -TEST(SequenceSoftmaxActivation, activation) { - for (auto hasSubseq : {false, true}) { - LOG(INFO) << "hasSubseq = " << hasSubseq; - testSequenceSoftmaxAct(hasSubseq); - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_BatchNorm.cpp b/paddle/gserver/tests/test_BatchNorm.cpp deleted file mode 100644 index a3ec66c75829c5ef0ae834656ee82e40be76c892..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_BatchNorm.cpp +++ /dev/null @@ -1,195 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/utils/GlobalConstants.h" - -#include "LayerGradUtil.h" -#include "paddle/cuda/include/hl_batch_norm.h" -#include "paddle/math/tests/TensorCheck.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(gpu_id); -DECLARE_double(checkgrad_eps); -DECLARE_bool(thread_local_rand_use_global_seed); -DECLARE_bool(prev_batch_state); - -// Test that the batchNormLayer can be followed by a ConvLayer -TEST(Layer, batchNorm) { - FLAGS_use_gpu = false; - TestConfig configBN; - const int CHANNELS = 6272; - const int IMG_SIZE = 1; - configBN.layerConfig.set_type("batch_norm"); - configBN.layerConfig.set_name("bn"); - configBN.layerConfig.set_size(CHANNELS * IMG_SIZE * IMG_SIZE); - configBN.layerConfig.set_active_type("relu"); - configBN.biasSize = CHANNELS; - configBN.inputDefs.push_back({INPUT_DATA, - "layer_0", - /* dim= */ IMG_SIZE * IMG_SIZE * CHANNELS, - /* paraSize= */ CHANNELS}); - - configBN.inputDefs.push_back( - {INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); - configBN.inputDefs.back().isStatic = true; - configBN.inputDefs.push_back( - {INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); - configBN.inputDefs.back().isStatic = true; - - LayerInputConfig* input = configBN.layerConfig.add_inputs(); - configBN.layerConfig.add_inputs(); - configBN.layerConfig.add_inputs(); - - ImageConfig* img_conf = input->mutable_image_conf(); - img_conf->set_channels(CHANNELS); - img_conf->set_img_size(IMG_SIZE); - - // Setting up conv-layer config - TestConfig config; - config.biasSize = 64; - config.layerConfig.set_type("exconv"); - config.layerConfig.set_num_filters(64); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - config.inputDefs.push_back({INPUT_DATA, "bn", 6272, 204800}); - input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(5); - conv->set_filter_size_y(5); - conv->set_channels(128); - conv->set_padding(1); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_groups(1); - conv->set_filter_channels(conv->channels() / conv->groups()); - conv->set_img_size(7); - conv->set_output_x(3); - config.layerConfig.set_size(conv->output_x() * conv->output_x() * - config.layerConfig.num_filters()); - config.layerConfig.set_name("conv"); - - // data layer initialize - std::vector dataLayers; - LayerMap layerMap; - vector datas; - initDataLayer(configBN, - &dataLayers, - &datas, - &layerMap, - "batch_norm", - 100, - false, - false); - // test layer initialize - std::vector parameters; - LayerPtr bnLayer; - initTestLayer(configBN, &layerMap, ¶meters, &bnLayer); - - std::vector parameters2; - LayerPtr convLayer; - initTestLayer(config, &layerMap, ¶meters2, &convLayer); - - bnLayer->forward(PASS_GC); - convLayer->forward(PASS_GC); - - CHECK_EQ(static_cast(convLayer->getOutputValue()->getHeight()), 100); - CHECK_EQ(static_cast(convLayer->getOutputValue()->getWidth()), 576); -} - -#ifdef PADDLE_WITH_CUDA -void batchNormInference(int n, int c, int h, int w) { - MatrixPtr input = std::make_shared(n, c * h * w); - MatrixPtr cudnnOut = std::make_shared(n, c * h * w); - MatrixPtr cudaOut = std::make_shared(n, c * h * w); - MatrixPtr cudnnCheck = std::make_shared(n, c * h * w); - MatrixPtr cudaCheck = std::make_shared(n, c * h * w); - input->randomizeUniform(); - cudnnOut->zeroMem(); - cudaOut->zeroMem(); - - MatrixPtr scale = std::make_shared(1, c); - scale->randomizeUniform(); - MatrixPtr bias = std::make_shared(1, c); - bias->randomizeUniform(); - - MatrixPtr movingMean = std::make_shared(1, c); - movingMean->randomizeUniform(); - - MatrixPtr movingVar = std::make_shared(1, c); - movingVar->randomizeUniform(); - movingVar->clip(0.01, 50); - - hl_tensor_descriptor ioDesc; - hl_tensor_descriptor bnDesc; - hl_create_tensor_descriptor(&ioDesc); - hl_create_tensor_descriptor(&bnDesc); - hl_tensor_reshape(ioDesc, n, c, h, w); - hl_tensor_reshape(bnDesc, 1, c, 1, 1); - - double EPS = 1E-5; - hl_batch_norm_forward_inference(ioDesc, - input->getData(), - ioDesc, - cudnnOut->getData(), - bnDesc, - scale->getData(), - bias->getData(), - movingMean->getData(), - movingVar->getData(), - EPS); - - hl_batch_norm_cuda_inference(input->getData(), - cudaOut->getData(), - scale->getData(), - bias->getData(), - movingMean->getData(), - movingVar->getData(), - EPS, - n, - c, - h, - w); - - cudnnCheck->copyFrom(*cudnnOut); - cudaCheck->copyFrom(*cudaOut); - autotest::TensorCheckErr(*cudnnCheck, *cudaCheck); - - hl_destroy_tensor_descriptor(ioDesc); - hl_destroy_tensor_descriptor(bnDesc); -} - -TEST(BatchNorm, Inference) { - batchNormInference(33, 267, 1, 1); - batchNormInference(19, 105, 4, 4); -} -#endif - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_CRFLayerGrad.cpp b/paddle/gserver/tests/test_CRFLayerGrad.cpp deleted file mode 100644 index 9f3d2936569af8f1923a471f4d262e9a472649c0..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_CRFLayerGrad.cpp +++ /dev/null @@ -1,173 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/gserver/layers/LinearChainCRF.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT - -DECLARE_int32(gpu_id); -DECLARE_bool(thread_local_rand_use_global_seed); - -static inline bool getNextSequence(std::vector& seq, int numClasses) { - for (auto& v : seq) { - if (++v < numClasses) { - return true; - } - v = 0; - } - return false; -} - -// log(exp(x) + exp(y)) -static inline real logSum(real x, real y) { - real maxValue = std::max(x, y); - if (std::isinf(maxValue)) { - return -std::numeric_limits::infinity(); - } else { - return maxValue + log(exp(x - maxValue) + exp(y - maxValue)); - } -} - -static inline std::vector genRandLabels(int numClasses, int length) { - std::vector labels(length); - for (int i = 0; i < length; ++i) { - labels[i] = rand() % numClasses; // NOLINT - } - return labels; -} - -TEST(CRFLayer, cost) { - const int numClasses = 4; - CpuVector para(numClasses * (numClasses + 2)); - real* a = para.getData(); - real* b = para.getData() + numClasses; - real* w = para.getData() + 2 * numClasses; - LinearChainCRF crf(4, para.getData()); - for (int length : {1, 2, 3, 10}) { - for (int tries = 0; tries < 10; ++tries) { - CpuMatrix x(length, numClasses); - x.randomizeUniform(); - para.randnorm(0, 2); - - std::vector goldenLabels = genRandLabels(numClasses, length); - - real cost = crf.forward(x.getData(), goldenLabels.data(), length); - - real logZ = -std::numeric_limits::infinity(); - real logNominator = -std::numeric_limits::infinity(); - std::vector testResult(length, 0); - do { - real score = a[testResult.front()]; - score += x.getElement(0, testResult.front()); - for (int k = 1; k < length; ++k) { - score += x.getElement(k, testResult[k]) + - w[numClasses * testResult[k - 1] + testResult[k]]; - } - score += b[testResult.back()]; - logZ = logSum(logZ, score); - - if (goldenLabels == testResult) { - logNominator = score; - } - } while (getNextSequence(testResult, numClasses)); - - real trueCost = -logNominator + logZ; - - real diff = fabs(trueCost - cost); - diff /= fabs(cost) < fabs(trueCost) ? fabs(cost) : fabs(trueCost); - VLOG(1) << "cost=" << cost << " trueCost=" << trueCost << " diff=" << diff - << std::endl; - if (typeid(real) == typeid(double)) { // NOLINT - EXPECT_LE(diff, 1e-10); - } else { - EXPECT_LE(diff, 5e-3); - } - } - } -} - -inline real epsilon() { return typeid(real) == typeid(double) ? 1e-10 : 0.06; } - -TestConfig initTestConfig(size_t numClasses, bool withWeight) { - TestConfig config; - config.layerConfig.set_type("crf"); - config.layerConfig.set_size(numClasses); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, - "layer_0", - numClasses, - numClasses * (numClasses + 2)}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back( - {INPUT_SEQUENCE_LABEL, "layer_label", numClasses, 0}); - config.layerConfig.add_inputs(); - - if (withWeight) { - config.inputDefs.push_back({INPUT_DENSE_DIM_DATA, "layer_weight", 1, 0}); - config.layerConfig.add_inputs(); - } - - return config; -} - -TEST(Layer, CRFLayer) { - size_t numClasses = 10; - for (int tries = 0; tries < 5; ++tries) { - TestConfig config = initTestConfig(numClasses, /* withWeight= */ false); - for (int length : {1, 3, 100}) { - // Not support GPU now - testLayerGrad(config, - "crf", - length, - /* trans= */ false, - /* useGpu= */ false, - /* useWeight= */ false, - epsilon()); - } - } -} - -TEST(Layer, CRFLayerUseWeight) { - size_t numClasses = 10; - for (int tries = 0; tries < 5; ++tries) { - TestConfig config = initTestConfig(numClasses, /* withWeight= */ true); - for (int length : {1, 3, 100}) { - // Not support GPU now - testLayerGrad(config, - "crf", - length, - /* trans= */ false, - /* useGpu= */ false, - /* useWeight= */ false, - epsilon()); - } - } -} - -int main(int argc, char** argv) { - initMain(argc, argv); - hl_start(); - hl_init(FLAGS_gpu_id); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_CompareSparse.cpp b/paddle/gserver/tests/test_CompareSparse.cpp deleted file mode 100644 index 2fbc404125a9364ac44a990f8ec92962cf7d1298..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_CompareSparse.cpp +++ /dev/null @@ -1,228 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "paddle/trainer/Trainer.h" - -#include -#include - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static const string& configFile1 = "gserver/tests/sequence_lstm.conf"; - -DECLARE_bool(use_gpu); -DECLARE_string(config); -DECLARE_int32(gpu_id); -DECLARE_int32(seed); -DECLARE_int32(num_passes); -DECLARE_int32(saving_period); - -DECLARE_int32(num_gradient_servers); -DECLARE_int32(port); -DECLARE_bool(local); -DECLARE_bool(use_old_updater); -DECLARE_bool(parallel_nn); -DECLARE_string(config_args); -DEFINE_double(max_diff_ratio, - 0.0f, - "max diff ratio allowed for parameters value"); - -int gNumDevices = 0; - -std::vector trainerOnePassTest(const string& configFile, - bool sparseUpdate, - int trainerCount = 1, - bool useGpu = false) { - FLAGS_use_gpu = useGpu; - FLAGS_config = configFile; - FLAGS_trainer_count = trainerCount; - FLAGS_config_args = sparseUpdate ? "sparse_update=1" : "sparse_update=0"; - - LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount - << " configFile=" << configFile << " sparseUpdate=" << sparseUpdate; - srand(FLAGS_seed); - *ThreadLocalRand::getSeed() = FLAGS_seed; - ThreadLocalRandomEngine::get().seed(FLAGS_seed); - if (useGpu) { - CHECK_LE(trainerCount, gNumDevices); - } - - std::vector> pservers; - if (!FLAGS_local) { - int numPorts = FLAGS_ports_num + FLAGS_ports_num_for_sparse; - pservers.resize(numPorts); - - for (int i = 0; i < numPorts; ++i) { - pservers[i].reset(new ParameterServer2(std::string(), FLAGS_port + i)); - pservers[i]->init(); - pservers[i]->start(); - } - } - - Trainer trainer; - trainer.init(TrainerConfigHelper::createFromFlagConfig()); - trainer.train(); - return trainer.getGradientMachine()->getParameters(); -} - -std::vector& getDenseParameters() { - static std::vector denseParameters; - if (denseParameters.empty()) { - // use dense training as base - FLAGS_local = true; - denseParameters = trainerOnePassTest(configFile1, false); - } - - return denseParameters; -} - -void checkBuffer(real* A, - const char* desA, - real* B, - const char* desB, - size_t len, - double maxDiffRatio) { - double maxDiff = 0; - double maxValue = 0; - for (size_t i = 0; i < len; ++i) { - double diff = fabs(A[i] - B[i]); - maxValue = std::max(maxValue, std::max(fabs(A[i]), fabs(B[i]))); - maxDiff = std::max(maxDiff, diff); - } - EXPECT_LE(maxDiff / maxValue, maxDiffRatio); - LOG(INFO) << " maxDiff=" << maxDiff << " maxValue=" << maxValue - << " maxDiff/maxValue=" << maxDiff / maxValue << "\n\n"; -} - -void compareValue(const vector& parametersA, - const vector& parametersB, - double maxDiffRatio = 0.0) { - LOG(INFO) << "\n\n--------------------------------" - << " Check Gradient Machine Parameters:" - << " -------------------------------------\n"; - for (size_t i = 0; i < parametersA.size(); ++i) { - ParameterPtr parameterA, parameterB; - parameterA = parametersA[i]; - parameterB = parametersB[i]; - - CpuVector paraA(parameterA->getSize()); - CpuVector paraB(parameterB->getSize()); - paraA.copyFrom(*parameterA->getBuf(PARAMETER_VALUE)); - paraB.copyFrom(*parameterB->getBuf(PARAMETER_VALUE)); - - LOG(INFO) << "\n\n----------- PARAMETER_VALUE: " << parameterA->getName() - << " ; size : " << paraA.getSize() << " ------------"; - checkBuffer(paraA.getData(), - "para_A", - paraB.getData(), - "para_B", - paraA.getSize(), - maxDiffRatio); - } -} - -TEST(compareSparse, cpu) { - FLAGS_local = 1; // disable remote sparse update in parameter config - std::vector parameters = trainerOnePassTest(configFile1, true); - compareValue(getDenseParameters(), parameters); -} - -TEST(compareSparse, remote_cpu) { - FLAGS_local = 0; // will enable remote sparse update - FLAGS_ports_num_for_sparse = 5; - std::vector parameters = trainerOnePassTest(configFile1, true); - compareValue(getDenseParameters(), parameters); -} - -TEST(compareSparse, cpu10_local_vs_remote) { - FLAGS_local = 1; // disable remote sparse update in parameter config - std::vector localParameters = - trainerOnePassTest(configFile1, true, 2); - - FLAGS_local = 0; // will enable remote sparse update - FLAGS_ports_num_for_sparse = 5; - std::vector remoteParameters = - trainerOnePassTest(configFile1, true, 2); - - compareValue(localParameters, remoteParameters); -} - -TEST(compareSparse, multiGradientMachine) { - int numGpu; -#ifdef PADDLE_TYPE_DOUBLE - double eps = 1e-8; -#else - double eps = 1e-4; -#endif - numGpu = hl_get_device_count(); - for (bool local : {false, true}) { - FLAGS_local = local; - FLAGS_ports_num_for_sparse = 5; - for (bool useGpu : {false, true}) { -#ifndef PADDLE_WITH_CUDA - if (useGpu) continue; -#endif - FLAGS_parallel_nn = useGpu; - LOG(INFO) << " local=" << local << " useGpu=" << useGpu; - int trainerCount = useGpu ? numGpu : 2; - std::vector parameters = - trainerOnePassTest(configFile1, true, trainerCount, useGpu); - compareValue(getDenseParameters(), parameters, eps); - } - } - FLAGS_parallel_nn = false; -} - -TEST(compareSparse, NeuralNetwork) { -#ifdef PADDLE_TYPE_DOUBLE - double eps = 1e-8; -#else - double eps = 1e-4; -#endif - for (bool local : {false, true}) { - FLAGS_local = local; - FLAGS_ports_num_for_sparse = 5; - for (bool useGpu : {false, true}) { -#ifndef PADDLE_WITH_CUDA - if (useGpu) continue; -#endif - FLAGS_parallel_nn = useGpu; - LOG(INFO) << " local=" << local << " useGpu=" << useGpu; - int trainerCount = 1; - std::vector parameters = - trainerOnePassTest(configFile1, true, trainerCount, useGpu); - compareValue(getDenseParameters(), parameters, useGpu ? eps : 0); - } - } - FLAGS_parallel_nn = false; -} - -int main(int argc, char** argv) { - // FIXME(tonyyang-svail): - // Turn off this test due CI failure: - // https://paddleci.ngrok.io/viewLog.html?buildId=27608&buildTypeId=Paddle_PrCi&tab=buildLog&_focus=10430 - return 0; - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - initPython(argc, argv); - - gNumDevices = hl_get_device_count(); - FLAGS_num_passes = 1; // train one pass - FLAGS_saving_period = 100000; // do not save parameter - - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_CompareTwoNets.cpp b/paddle/gserver/tests/test_CompareTwoNets.cpp deleted file mode 100644 index 1c9b4002a34ca5a9b668be69bd0ad392eb763803..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_CompareTwoNets.cpp +++ /dev/null @@ -1,209 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include - -#include "paddle/trainer/Trainer.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_int32(gpu_id); - -DECLARE_bool(local); -DECLARE_bool(use_gpu); - -DECLARE_string(config); -DECLARE_string(nics); - -DEFINE_bool(need_high_accuracy, - false, - "whether need to run in double accuracy"); -DEFINE_double( - max_diff_ratio, - 0.0f, - "max diff ratio allowed for outputs and parameters (value/gradient)"); -DECLARE_bool(thread_local_rand_use_global_seed); -DECLARE_int32(seed); - -static const string& config_file_a = "gserver/tests/sequence_recurrent.py"; -static const string& config_file_b = - "gserver/tests/sequence_recurrent_group.py"; - -struct ComData { - vector outArgs; - vector parameters; -}; - -void calcGradient(ComData& data, const string configFile) { - FLAGS_config = configFile; - - FLAGS_local = true; - FLAGS_use_gpu = false; - - FLAGS_nics = ""; - - *ThreadLocalRand::getSeed() = FLAGS_seed; - srand(FLAGS_seed); - - Trainer trainer; - trainer.init(TrainerConfigHelper::createFromFlagConfig(), false); - - data.parameters = trainer.getGradientMachine()->getParameters(); - - DataBatch dataBatch; - int32_t batchSize = trainer.getConfig().opt_config().batch_size(); - - trainer.getDataProvider()->reset(); - trainer.getDataProvider()->setSkipShuffle(); - trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch); - - CHECK(dataBatch.getSize()) << "No data from data provider"; - vector& inArgs = dataBatch.getStreams(); - - trainer.getGradientMachine()->start(); - trainer.getGradientMachine()->forwardBackward( - inArgs, &data.outArgs, PASS_TRAIN); - - trainer.getGradientMachine()->finish(); -} - -void checkBuffer(real* A, - const char* desA, - real* B, - const char* desB, - size_t len, - size_t width = 1) { - int nNum = 0; - real maxVal = 0; - for (size_t i = 0; i < len; ++i) { - maxVal = std::max(maxVal, std::max(A[i], B[i])); - } - real maxDiff = 0; - for (size_t i = 0; i < len; ++i) { - real diff = fabs(A[i] - B[i]); - maxDiff = std::max(maxDiff, diff); - if (diff > maxVal * FLAGS_max_diff_ratio) { - nNum++; - VLOG(1) << "Row: " << i / width << ", " << desA << " : " << A[i] << " " - << desB << " : " << B[i] << " diff=" << diff; - } - } - EXPECT_EQ(0, nNum); - LOG(INFO) << "maxValue=" << maxVal << " maxDiff=" << maxDiff << "\n\n"; -} - -void compareGradient(ComData& comDataA, ComData& comDataB) { - vector outArgsA = comDataA.outArgs; - vector outArgsB = comDataB.outArgs; - - for (size_t i = 0; i < outArgsA.size(); ++i) { - CpuMatrix matA(outArgsA[i].value->getHeight(), - outArgsA[i].value->getWidth()); - CpuMatrix matB(outArgsB[i].value->getHeight(), - outArgsB[i].value->getWidth()); - - matA.copyFrom(*outArgsA[i].value); - matB.copyFrom(*outArgsB[i].value); - - LOG(INFO) << "\n--------------------------------" - << " Check Network Output_" << i << ":" - << " -------------------------------------\n"; - checkBuffer(matA.getData(), - "network A output", - matB.getData(), - "network B output", - matA.getElementCnt(), - matA.getWidth()); - } - - vector& parametersA = comDataA.parameters; - vector& parametersB = comDataB.parameters; - - LOG(INFO) << "\n\n--------------------------------" - << " Check Gradient Machine Parameters:" - << " -------------------------------------\n"; - for (size_t i = 0; i < parametersA.size(); ++i) { - ParameterPtr parameterA, parameterB; - parameterA = parametersA[i]; - parameterB = parametersB[i]; - - CpuVector paraA(parameterA->getSize()); - CpuVector paraB(parameterB->getSize()); - paraA.copyFrom(*parameterA->getBuf(PARAMETER_VALUE)); - paraB.copyFrom(*parameterB->getBuf(PARAMETER_VALUE)); - - LOG(INFO) << "\n\n----------- PARAMETER_VALUE: " << parameterA->getName() - << " ; size : " << paraA.getSize() << " ------------"; - checkBuffer(paraA.getData(), - "Network A", - paraB.getData(), - "Network B", - paraA.getSize()); - - CpuVector gradA(*parameterA->getBuf(PARAMETER_GRADIENT)); - CpuVector gradB(*parameterB->getBuf(PARAMETER_GRADIENT)); - - LOG(INFO) << "\n\n----------- PARAMETER_GRADIENT: " << parameterA->getName() - << " ; size : " << gradA.getSize() << " -----------"; - checkBuffer(gradA.getData(), - "Network A", - gradB.getData(), - "Network B", - gradA.getSize()); - } -} - -TEST(Trainer, create) { - ComData dataA; - calcGradient(dataA, config_file_a); - LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n"; - - ComData dataB; - calcGradient(dataB, config_file_b); - LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n"; - - compareGradient(dataA, dataB); -} - -int main(int argc, char** argv) { - FLAGS_thread_local_rand_use_global_seed = true; - paddle::initMain(argc, argv); - testing::InitGoogleTest(&argc, argv); - initPython(argc, argv); - -#ifndef PADDLE_TYPE_DOUBLE - if (FLAGS_need_high_accuracy) { - LOG(INFO) << "skip test due to it's need high accuracy"; - return 0; - } - if (FLAGS_max_diff_ratio == 0.0f) { - FLAGS_max_diff_ratio = 1e-5; - LOG(INFO) << "auto set max_diff_ratio " << FLAGS_max_diff_ratio - << " in low accuracy mode"; - } -#else - if (FLAGS_max_diff_ratio == 0.0f) { - FLAGS_max_diff_ratio = 1e-10; - LOG(INFO) << "auto set max_diff_ratio " << FLAGS_max_diff_ratio - << " in high accuracy mode"; - } -#endif - - int ret = RUN_ALL_TESTS(); - return ret; -} diff --git a/paddle/gserver/tests/test_ConvTrans.cpp b/paddle/gserver/tests/test_ConvTrans.cpp deleted file mode 100644 index 2e394a74b7d53fc53727d817c06479d545ade65d..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_ConvTrans.cpp +++ /dev/null @@ -1,244 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/GlobalConstants.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(gpu_id); -DECLARE_double(checkgrad_eps); -DECLARE_bool(thread_local_rand_use_global_seed); -DECLARE_bool(prev_batch_state); - -// Test that the convTrans forward is the same as conv backward -TEST(Layer, convTransLayerFwd) { - // Setting up conv-trans layer - TestConfig configt; - configt.biasSize = 3; - configt.layerConfig.set_type("exconvt"); - configt.layerConfig.set_num_filters(3); - configt.layerConfig.set_partial_sum(1); - configt.layerConfig.set_shared_biases(true); - - configt.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 384}); - LayerInputConfig* input = configt.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(2); - conv->set_filter_size_y(4); - conv->set_channels(16); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_groups(1); - conv->set_filter_channels(3 / conv->groups()); - conv->set_img_size(16); - conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - configt.layerConfig.set_size(conv->img_size() * conv->img_size() * - configt.layerConfig.num_filters()); - configt.layerConfig.set_name("convTrans"); - - // data layer initialize - std::vector dataLayers; - LayerMap layerMap; - vector datas; - initDataLayer( - configt, &dataLayers, &datas, &layerMap, "convTrans", 100, false, false); - // test layer initialize - std::vector parameters; - LayerPtr convtLayer; - initTestLayer(configt, &layerMap, ¶meters, &convtLayer); - convtLayer->getBiasParameter()->zeroMem(); - convtLayer->forward(PASS_GC); - - // Setting up conv-layer config - TestConfig config; - config.biasSize = 16; - config.layerConfig.set_type("exconv"); - config.layerConfig.set_num_filters(16); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - config.inputDefs.push_back({INPUT_DATA, "layer_1", 768, 384}); - input = config.layerConfig.add_inputs(); - conv = input->mutable_conv_conf(); - conv->set_filter_size(2); - conv->set_filter_size_y(4); - conv->set_channels(3); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_groups(1); - conv->set_filter_channels(conv->channels() / conv->groups()); - conv->set_img_size(16); - conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - config.layerConfig.set_size(conv->output_x() * conv->output_x() * - config.layerConfig.num_filters()); - config.layerConfig.set_name("conv"); - - // data layer initialize - std::vector dataLayers2; - LayerMap layerMap2; - vector datas2; - initDataLayer( - config, &dataLayers2, &datas2, &layerMap2, "conv", 100, false, false); - // test layer initialize - std::vector parameters2; - LayerPtr convLayer; - initTestLayer(config, &layerMap2, ¶meters2, &convLayer); - - // Sync convLayer and convtLayer parameter - convLayer->getBiasParameter()->zeroMem(); - convLayer->getParameters()[0] - ->getBuf(PARAMETER_VALUE) - ->copyFrom(*(convtLayer->getParameters()[0]->getBuf(PARAMETER_VALUE))); - - // Set convLayer outputGrad as convTransLayer input value - convLayer->forward(PASS_GC); - convLayer->getOutput().grad->copyFrom(*(dataLayers[0]->getOutputValue())); - - vector callbackFlags(parameters2.size(), 0); - auto callback = [&](Parameter* para) { ++callbackFlags[para->getID()]; }; - convLayer->backward(callback); - - // Check that the convLayer backward is the same as convTransLayer forward - checkMatrixEqual(convtLayer->getOutputValue(), - dataLayers2[0]->getOutputGrad()); -} - -// Do one forward pass of convTrans layer and check to see if its output -// matches the given result -void doOneConvtTest(size_t imgSize, - size_t output_x, - size_t stride, - size_t padding, - size_t filter_size, - MatrixPtr& result) { - TestConfig configt; - configt.biasSize = 1; - configt.layerConfig.set_type("exconvt"); - configt.layerConfig.set_num_filters(1); - configt.layerConfig.set_partial_sum(1); - configt.layerConfig.set_shared_biases(true); - - configt.inputDefs.push_back( - {INPUT_DATA, "layer_0", output_x * output_x, filter_size * filter_size}); - LayerInputConfig* input = configt.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(filter_size); - conv->set_filter_size_y(filter_size); - conv->set_channels(1); - conv->set_padding(padding); - conv->set_padding_y(padding); - conv->set_stride(stride); - conv->set_stride_y(stride); - conv->set_groups(1); - conv->set_filter_channels(1); - conv->set_img_size(imgSize); - conv->set_output_x(output_x); - - configt.layerConfig.set_size(conv->img_size() * conv->img_size() * - configt.layerConfig.num_filters()); - configt.layerConfig.set_name("convTrans"); - - std::vector dataLayers; - LayerMap layerMap; - vector datas; - initDataLayer( - configt, &dataLayers, &datas, &layerMap, "convTrans", 1, false, false); - dataLayers[0]->getOutputValue()->zeroMem(); - dataLayers[0]->getOutputValue()->add(1.0); - - // test layer initialize - std::vector parameters; - LayerPtr convtLayer; - initTestLayer(configt, &layerMap, ¶meters, &convtLayer); - convtLayer->getBiasParameter()->zeroMem(); - convtLayer->getParameters()[0]->zeroMem(); - convtLayer->getParameters()[0]->getBuf(PARAMETER_VALUE)->add(1.0); - convtLayer->forward(PASS_GC); - - checkMatrixEqual(convtLayer->getOutputValue(), result); -} - -TEST(Layer, convTransLayerFwd2) { - MatrixPtr result; - result = Matrix::create(1, 5 * 5, false, false); - result->zeroMem(); - result->add(1.0); - doOneConvtTest(/* imgSize */ 5, - /* output_x */ 1, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 5, - result); - - real resultData[] = {1, 2, 2, 2, 1, 2, 4, 4, 4, 2, 2, 4, 4, - 4, 2, 2, 4, 4, 4, 2, 1, 2, 2, 2, 1}; - result->setData(resultData); - doOneConvtTest(/* imgSize */ 5, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 4, - result); - - real resultData2[] = {1, 2, 2, 2, 1, 2, 4, 4, 4, 2, 2, 4, 4, - 4, 2, 2, 4, 4, 4, 2, 1, 2, 2, 2, 1}; - result->setData(resultData2); - doOneConvtTest(/* imgSize */ 5, - /* output_x */ 2, - /* stride */ 2, - /* padding */ 1, - /* filter_size */ 5, - result); - - real resultData3[] = {1, 1, 2, 1, 1, 1, 1, 2, 1, 1, 2, 2, 4, - 2, 2, 1, 1, 2, 1, 1, 1, 1, 2, 1, 1}; - result->setData(resultData3); - doOneConvtTest(/* imgSize */ 5, - /* output_x */ 2, - /* stride */ 2, - /* padding */ 0, - /* filter_size */ 3, - result); -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_ConvUnify.cpp b/paddle/gserver/tests/test_ConvUnify.cpp deleted file mode 100644 index ba820d9a2acabf95ff816705e4df124bb95da077..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_ConvUnify.cpp +++ /dev/null @@ -1,315 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/GlobalConstants.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(gpu_id); -DECLARE_double(checkgrad_eps); -DECLARE_bool(thread_local_rand_use_global_seed); -DECLARE_bool(prev_batch_state); - -// Do one forward pass of ConvLayer using either exconv or cudnn_conv -MatrixPtr doOneConvTest(size_t imgSize, - size_t output_x, - size_t stride, - size_t padding, - size_t filter_size, - size_t channel, - size_t numfilters, - size_t groups, - MatrixPtr& inputData, - real* param, - bool useGpu, - bool isDeconv = false) { - TestConfig config; - config.biasSize = numfilters; - string layerType; - if (useGpu) { - layerType = (isDeconv) ? "cudnn_convt" : "cudnn_conv"; - } else { - layerType = (isDeconv) ? "exconvt" : "exconv"; - } - config.layerConfig.set_type(layerType); - config.layerConfig.set_num_filters(numfilters); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - size_t weightSize = channel * filter_size * filter_size * - config.layerConfig.num_filters() / groups; - if (isDeconv) { - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", output_x * output_x * channel, weightSize}); - config.layerConfig.set_size(imgSize * imgSize * - config.layerConfig.num_filters()); - } else { - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", imgSize * imgSize * channel, weightSize}); - config.layerConfig.set_size(output_x * output_x * - config.layerConfig.num_filters()); - } - - LayerInputConfig* input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(filter_size); - conv->set_filter_size_y(filter_size); - conv->set_channels(channel); - conv->set_padding(padding); - conv->set_padding_y(padding); - conv->set_stride(stride); - conv->set_stride_y(stride); - conv->set_groups(groups); - conv->set_img_size(imgSize); - conv->set_output_x(output_x); - - if (isDeconv) { - conv->set_filter_channels(numfilters / groups); - } else { - conv->set_filter_channels(channel / groups); - } - - config.layerConfig.set_name("conv"); - - std::vector dataLayers; - LayerMap layerMap; - vector datas; - initDataLayer( - config, &dataLayers, &datas, &layerMap, "conv", 1, false, useGpu); - dataLayers[0]->getOutputValue()->zeroMem(); - dataLayers[0]->getOutputValue()->copyFrom(*inputData); - - // test layer initialize - std::vector parameters; - LayerPtr convLayer; - initTestLayer(config, &layerMap, ¶meters, &convLayer); - convLayer->getBiasParameter()->zeroMem(); - convLayer->getParameters()[0]->zeroMem(); - convLayer->getParameters()[0] - ->getBuf(PARAMETER_VALUE) - ->copyFrom(param, weightSize); - convLayer->forward(PASS_GC); - - return convLayer->getOutputValue(); -} - -TEST(Layer, convParaUnified) { -#ifdef PADDLE_WITH_CUDA - MatrixPtr input, resultCpu, resultGpu; - - /// TEST1 for conv /// - input = Matrix::create(1, 4 * 4, false, false); - real inputData[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; - real param[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 8, 7, 6, 5, 4, 3, 2, 1}; - - input->setData(inputData); - - resultCpu = doOneConvTest(/* imgSize */ 4, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 3, - /*channel*/ 1, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param, - /*useGpu*/ false); - - resultGpu = doOneConvTest(/* imgSize */ 4, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 3, - /*channel*/ 1, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param, - /*useGpu*/ true); - checkMatrixEqual(resultCpu, resultGpu); - - /// TEST1 for deconv /// - input = Matrix::create(1, 2 * 2, false, false); - real inputDataT[] = {1, 2, 3, 4}; - input->setData(inputDataT); - - resultCpu = doOneConvTest(/* imgSize */ 4, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 3, - /*channel*/ 1, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param, - /*useGpu*/ false, - /*isDeconv*/ true); - - resultGpu = doOneConvTest(/* imgSize */ 4, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 3, - /*channel*/ 1, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param, - /*useGpu*/ true, - /*isDeconv*/ true); - checkMatrixEqual(resultCpu, resultGpu); - - /// TEST2 for conv /// - input = Matrix::create(1, 3 * 3 * 2, false, false); - real inputData2[] = { - 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18}; - real param2[] = {1, 2, 3, 4, 5, 6, 7, 8, 8, 7, 6, 5, 4, 3, 2, 1}; - - input->setData(inputData2); - - resultCpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param2, - /*useGpu*/ false); - - resultGpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param2, - /*useGpu*/ true); - checkMatrixEqual(resultCpu, resultGpu); - - /// TEST3 for conv /// - real param3[] = {1, 2, 3, 4, 4, 3, 2, 1}; - - resultCpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 2, - input, - param3, - /*useGpu*/ false); - - resultGpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 2, - input, - param3, - /*useGpu*/ true); - checkMatrixEqual(resultCpu, resultGpu); - - /// TEST2 for deconv /// - input = Matrix::create(1, 2 * 2 * 2, false, false); - real inputData2T[] = {1, 2, 3, 4, 5, 6, 7, 8}; - input->setData(inputData2T); - - resultCpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param2, - /*useGpu*/ false, - /*isDeconv*/ true); - - resultGpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 1, - input, - param2, - /*useGpu*/ true, - /*isDeconv*/ true); - checkMatrixEqual(resultCpu, resultGpu); - - /// TEST3 for deconv /// - resultCpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 2, - input, - param3, - /*useGpu*/ false, - /*isDeconv*/ true); - - resultGpu = doOneConvTest(/* imgSize */ 3, - /* output_x */ 2, - /* stride */ 1, - /* padding */ 0, - /* filter_size */ 2, - /*channel*/ 2, - /*numfilters*/ 2, - /*groups*/ 2, - input, - param3, - /*useGpu*/ true, - /*isDeconv*/ true); - checkMatrixEqual(resultCpu, resultGpu); -#endif -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp deleted file mode 100644 index 0041ed30939d1a6111a2db753da6172bb65e374b..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp +++ /dev/null @@ -1,352 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT - -DECLARE_int32(gpu_id); -DECLARE_bool(thread_local_rand_use_global_seed); - -const size_t MAX_SEQ_NUM = 23; -const size_t MAX_SEQ_LEN = 50; -const size_t MAX_BEAM_SIZE = 27; - -const size_t SEED = (size_t)(time(NULL)); - -struct SingleBeamExpansion { - vector seqStartPos; - vector subSeqStartPos; - vector candidateScores; - - // TODO(caoying): store this into Argument.ids - vector selectedIndices; - - vector groundTruth; - vector inBeam; - vector rowIdxInBeam; - vector colIdxInBeam; - - void resetGroundTruth(size_t n) { - groundTruth.clear(); - groundTruth.resize(n, -1); - - inBeam.clear(); - inBeam.resize(n, 0); - - rowIdxInBeam.clear(); - rowIdxInBeam.resize(n, -1); - - colIdxInBeam.clear(); - colIdxInBeam.resize(n, -1); - } -}; - -inline float randFloat() { - return static_cast(rand()) / static_cast(RAND_MAX); -} - -void genRand(real* numbers, size_t n) { - default_random_engine generator; - uniform_real_distribution distribution(0.0, 1.0); - for (size_t i = 0; i < n; ++i) numbers[i] = distribution(generator); -} - -vector randSampling(real range, int n) { - CHECK_GE(range, n); - vector num(range); - iota(begin(num), end(num), 0.); - if (range == n) return num; - - random_shuffle(begin(num), end(num)); - num.resize(n); - sort(begin(num), end(num)); - return num; -} - -void genCandidateScores(bool hasSubseq, - size_t beamSize, - SingleBeamExpansion& prevBeam, - SingleBeamExpansion& curBeam) { - vector& seqStartPos = curBeam.seqStartPos; - seqStartPos.resize(1, 0); - vector& subSeqStartPos = curBeam.subSeqStartPos; - subSeqStartPos.resize(1, 0); - - srand(SEED); - if (prevBeam.selectedIndices.size()) { - if (prevBeam.subSeqStartPos.size() > 1) { - int seqIdx = 1; - // samples in previous beam are nested sequences. - for (size_t i = 1; i < prevBeam.subSeqStartPos.size(); ++i) { - for (size_t j = 0; j < beamSize; ++j) { - if (prevBeam.selectedIndices[(i - 1) * beamSize + j] == -1.) break; - subSeqStartPos.push_back(1 + (rand() % MAX_SEQ_LEN) + - subSeqStartPos.back()); - } - if (prevBeam.seqStartPos[seqIdx] == prevBeam.subSeqStartPos[i]) { - seqStartPos.push_back(subSeqStartPos.back()); - seqIdx++; - } - } - } else { - for (size_t i = 0; i <= prevBeam.selectedIndices.size(); ++i) { - if (i && i % beamSize == 0) { - seqStartPos.push_back(subSeqStartPos.back()); - if (i == prevBeam.selectedIndices.size()) break; - } - if (prevBeam.selectedIndices[i] == -1.) continue; - subSeqStartPos.push_back(subSeqStartPos.back() + - (1 + (rand() % MAX_SEQ_LEN))); - } - } - } else { - // the first beam expansion - int seqNum = 1 + (rand() % MAX_SEQ_NUM); - for (int i = 0; i < seqNum; ++i) { - if (hasSubseq) { - for (size_t j = 0; j < 1 + (rand() % MAX_SEQ_NUM); ++j) - subSeqStartPos.push_back(subSeqStartPos.back() + - (1 + (rand() % MAX_SEQ_LEN))); - seqStartPos.push_back(subSeqStartPos.back()); - } else { - seqStartPos.push_back(seqStartPos.back() + - (1 + (rand() % MAX_SEQ_LEN))); - } - } - } - - size_t totalSeqNum = hasSubseq ? subSeqStartPos.back() : seqStartPos.back(); - curBeam.candidateScores.resize(totalSeqNum, 0.); - genRand(curBeam.candidateScores.data(), totalSeqNum); -} - -void genSelectedIndices(size_t beamSize, - vector& seqStartPos, - vector& selectedIndices) { - size_t selectedIdsCount = beamSize * (seqStartPos.size() - 1); - selectedIndices.resize(selectedIdsCount, -1.); - - for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { - int seqLen = seqStartPos[i + 1] - seqStartPos[i]; - int n = min(seqLen, static_cast(beamSize)); - vector ids = randSampling(seqLen, n); - memcpy(selectedIndices.data() + i * beamSize, - ids.data(), - sizeof(real) * ids.size()); - } -} - -void genGroundTruth(vector& beamExpansions, - size_t beamSize) { - SingleBeamExpansion& beam = beamExpansions[1]; - size_t seqNum = beam.seqStartPos.size() - 1; - for (size_t i = 2; i < beamExpansions.size(); ++i) - CHECK_EQ(seqNum, beamExpansions[i].seqStartPos.size() - 1); - - srand(SEED); - - // initialize the first beam. - beam.resetGroundTruth(seqNum); - for (size_t i = 0; i < seqNum; ++i) { - if (randFloat() > 0.5) { - /* - * force the randomly generated label falls in the beam by chance 0.5. - * otherwise, when sequence length is relatively long and beam size is - * relatively small, the gold sequences falls off the beam at in the - * first search. - */ - real* begPos = beam.selectedIndices.data() + i * beamSize; - beam.colIdxInBeam[i] = - rand() % count_if(begPos, begPos + beamSize, [](const real& val) { - return val != -1.; - }); - beam.groundTruth[i] = - beam.selectedIndices[i * beamSize + beam.colIdxInBeam[i]]; - beam.inBeam[i] = 1; - } else { - int label = rand() % (beam.seqStartPos[i + 1] - beam.seqStartPos[i]); - beam.groundTruth[i] = label; - - real* begPos = beam.selectedIndices.data() + i * beamSize; - real* endPos = begPos + beamSize; - real* lblPos = find(begPos, endPos, real(label)); - if (lblPos != endPos) { - beam.inBeam[i] = 1; - beam.colIdxInBeam[i] = lblPos - begPos; - } - } - beam.rowIdxInBeam[i] = i; - } - - // iterate over each beam expansions - for (size_t i = 2; i < beamExpansions.size(); ++i) { - SingleBeamExpansion& curBeam = beamExpansions[i]; - SingleBeamExpansion& prevBeam = beamExpansions[i - 1]; - curBeam.resetGroundTruth(seqNum); - - // iterate over each sequence - for (size_t j = 0; j < seqNum; ++j) { - if (!prevBeam.inBeam[j]) continue; - - // gold sequence falls in the beam in previous search. - real* begPos = prevBeam.selectedIndices.data(); - int offset = - prevBeam.rowIdxInBeam[j] * beamSize + prevBeam.colIdxInBeam[j]; - curBeam.rowIdxInBeam[j] = count_if( - begPos, begPos + offset, [](const real& val) { return val != -1.; }); - - if (randFloat() > 0.5) { - // force the randomly generated label falls in the beam by chance 0.5. - - real* start = - curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize; - int n = rand() % count_if(start, start + beamSize, [](const real& val) { - return val != -1.; - }); - curBeam.colIdxInBeam[j] = n; - curBeam.groundTruth[j] = *(start + n); - curBeam.inBeam[j] = 1; - } else { - CHECK_LE((size_t)curBeam.rowIdxInBeam[j] + 1, - curBeam.subSeqStartPos.size() - 1); - int start = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j]]; - int end = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j] + 1]; - CHECK_GT(size_t(end), size_t(start)); - int label = rand() % (end - start); - - curBeam.groundTruth[j] = label; - real* findBeg = - curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize; - real* lblPos = - find(findBeg, findBeg + beamSize, static_cast(label)); - if (lblPos != (findBeg + beamSize)) { - curBeam.inBeam[j] = 1; - curBeam.colIdxInBeam[j] = lblPos - findBeg; - } - } - } - } -} - -void genOneBeam(size_t beamSize, - bool hasSubseq, - SingleBeamExpansion& prevBeam, - SingleBeamExpansion& curBeam) { - genCandidateScores(hasSubseq, beamSize, prevBeam, curBeam); - genSelectedIndices(beamSize, - hasSubseq ? curBeam.subSeqStartPos : curBeam.seqStartPos, - curBeam.selectedIndices); -} - -void genRandomBeamExpansion(size_t expansionCount, - size_t beamSize, - vector& beamExpansions) { - beamExpansions.clear(); - beamExpansions.resize(expansionCount + 1); - - // beamExpansions[0] is reserved. - for (size_t i = 1; i <= expansionCount; ++i) - genOneBeam(beamSize, bool(i - 1), beamExpansions[i - 1], beamExpansions[i]); - genGroundTruth(beamExpansions, beamSize); -} - -void testCrossEntropyOverBeam(bool useGpu, - size_t beamSize, - vector& beams) { - TestConfig config; - config.layerConfig.set_type("cross_entropy_over_beam"); - - size_t seqNum = 0; - for (size_t i = 1; i < beams.size(); ++i) { - const SingleBeamExpansion& beam = beams[i]; - // create scores for all the candidates - MatrixPtr candidateScorePtr = - Matrix::create(beam.candidateScores.size(), 1, false, false); - candidateScorePtr->copyFrom(beam.candidateScores.data(), - beam.candidateScores.size()); - - ostringstream paramName; - paramName << "candidate_scores_" << i; - - if (beam.subSeqStartPos.size() > 1) { - seqNum = beam.subSeqStartPos.size() - 1; - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, - paramName.str(), - candidateScorePtr, - beam.seqStartPos, - beam.subSeqStartPos}); - } else { - seqNum = beam.seqStartPos.size() - 1; - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, - paramName.str(), - candidateScorePtr, - beam.seqStartPos}); - } - config.layerConfig.add_inputs(); - - // create indices for the selected candidates - MatrixPtr selectedCandidates = - Matrix::create(seqNum, beamSize, false, false); - selectedCandidates->copyFrom(beam.selectedIndices.data(), - beam.selectedIndices.size()); - paramName.clear(); - paramName << "selected_candidates_" << i; - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, paramName.str(), selectedCandidates}); - config.layerConfig.add_inputs(); - - // create the ground truth - paramName.clear(); - paramName << "label_" << i; - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, paramName.str(), beam.groundTruth}); - config.layerConfig.add_inputs(); - } - - testLayerGrad( - config, "cross_entropy_over_beam", seqNum, false, useGpu, false); -} - -TEST(Layer, CrossEntropyOverBeam) { - LOG(INFO) << "SEED = " << SEED; - const size_t beamSize = 1 + rand() % MAX_BEAM_SIZE; - LOG(INFO) << "beamSize = " << beamSize; - - // TODO(caoying): test with random beam expansions. - const size_t expansionCount = 3; - vector beams; - genRandomBeamExpansion(expansionCount, beamSize, beams); - - for (bool useGpu : {false, true}) - testCrossEntropyOverBeam(useGpu, beamSize, beams); -} - -int main(int argc, char** argv) { - initMain(argc, argv); - hl_start(); - hl_init(FLAGS_gpu_id); - FLAGS_thread_local_rand_use_global_seed = true; - srand(SEED); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_Evaluator.cpp b/paddle/gserver/tests/test_Evaluator.cpp deleted file mode 100644 index 4a8843f3affe7b1d4f3172be733aefc085c9e7a5..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_Evaluator.cpp +++ /dev/null @@ -1,267 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/testing/TestUtil.h" -#include "paddle/trainer/Trainer.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(gpu_id); -DECLARE_bool(thread_local_rand_use_global_seed); - -enum InputType { - INPUT_DATA, // dense vector - INPUT_LABEL, // id - INPUT_DATA_TARGET, // dense vector, but no gradient - INPUT_SEQUENCE_DATA, - INPUT_SEQUENCE_LABEL, - INPUT_SPARSE_NON_VALUE_DATA -}; - -struct InputDef { - InputType inputType; - string name; - size_t dim; -}; - -struct TestConfig { - EvaluatorConfig evaluatorConfig; - std::vector inputDefs; - bool testAccumulate; - TestConfig() : testAccumulate(true) {} -}; - -void testEvaluator(TestConfig testConf, - string testEvaluatorName, - size_t batchSize, - bool useGpu) { -#ifndef PADDLE_WITH_CUDA - if (useGpu) return; -#endif - FLAGS_use_gpu = useGpu; - testConf.evaluatorConfig.set_name(testEvaluatorName); - LOG(INFO) << " evaluator_type=" << testConf.evaluatorConfig.type() - << " useGpu=" << useGpu; - - std::vector arguments; - for (size_t i = 0; i < testConf.inputDefs.size(); ++i) { - Argument data; - size_t dim = testConf.inputDefs[i].dim; - switch (testConf.inputDefs[i].inputType) { - case INPUT_DATA: - case INPUT_SEQUENCE_DATA: - case INPUT_DATA_TARGET: - data.value = Matrix::create(batchSize, dim, false, useGpu); - data.value->randomizeUniform(); - - // make sure output > 0 && output < 1 - data.value->add(-0.5); - data.value->sigmoid(*data.value); - break; - case INPUT_LABEL: - case INPUT_SEQUENCE_LABEL: - data.ids = VectorT::create(batchSize, useGpu); - data.ids->rand(dim); // now rand number can be 0 to inputDefs[i].dim. - break; - case INPUT_SPARSE_NON_VALUE_DATA: - data.value = makeRandomSparseMatrix(batchSize, - dim, - /* withValue= */ false, - useGpu); - break; - default: - LOG(FATAL) << " unknown inputType "; - return; - } - - ICpuGpuVectorPtr sequenceStartPositions; - if (testConf.inputDefs[i].inputType == INPUT_SEQUENCE_DATA || - testConf.inputDefs[i].inputType == INPUT_SEQUENCE_LABEL) { - if (!sequenceStartPositions) { - generateSequenceStartPositions(batchSize, sequenceStartPositions); - } - data.sequenceStartPositions = sequenceStartPositions; - } - - arguments.push_back(data); - } - - Evaluator* testEvaluator = Evaluator::create(testConf.evaluatorConfig); - double totalScore = 0.0; - testEvaluator->start(); - totalScore += testEvaluator->evalImp(arguments); - testEvaluator->updateSamplesNum(arguments); - testEvaluator->finish(); - LOG(INFO) << *testEvaluator; - - std::vector names; - testEvaluator->getNames(&names); - paddle::Error err; - for (auto& name : names) { - auto value = testEvaluator->getValue(name, &err); - ASSERT_TRUE(err.isOK()); - LOG(INFO) << name << " " << value; - auto tp = testEvaluator->getType(name, &err); - ASSERT_TRUE(err.isOK()); - ASSERT_EQ(testConf.evaluatorConfig.type(), tp); - } - - double totalScore2 = 0.0; - if (testConf.testAccumulate) { - testEvaluator->start(); - totalScore2 += testEvaluator->evalImp(arguments); - testEvaluator->finish(); - EXPECT_LE(fabs(totalScore - totalScore2), 1.0e-5); - } -} - -void testEvaluatorAll(TestConfig testConf, - string testEvaluatorName, - size_t batchSize) { - testEvaluator(testConf, testEvaluatorName, batchSize, true); - testEvaluator(testConf, testEvaluatorName, batchSize, false); -} - -TEST(Evaluator, detection_map) { - TestConfig config; - config.evaluatorConfig.set_type("detection_map"); - config.evaluatorConfig.set_overlap_threshold(0.5); - config.evaluatorConfig.set_background_id(0); - config.evaluatorConfig.set_ap_type("Integral"); - config.evaluatorConfig.set_evaluate_difficult(0); - - config.inputDefs.push_back({INPUT_DATA, "output", 7}); - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "label", 6}); - config.evaluatorConfig.set_evaluate_difficult(false); - testEvaluatorAll(config, "detection_map", 100); - - config.evaluatorConfig.set_evaluate_difficult(true); - testEvaluatorAll(config, "detection_map", 100); -} - -TEST(Evaluator, classification_error) { - TestConfig config; - config.evaluatorConfig.set_type("classification_error"); - config.evaluatorConfig.set_top_k(5); - - config.inputDefs.push_back({INPUT_DATA, "output", 50}); - config.inputDefs.push_back({INPUT_LABEL, "label", 50}); - testEvaluatorAll(config, "classification_error", 100); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - testEvaluatorAll(config, "classification_error_weight", 100); - - // multi binary labels - config.inputDefs.clear(); - config.inputDefs.push_back({INPUT_DATA, "output", 100}); - config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "label", 100}); - // Not support GPU - testEvaluator(config, "classification_error_multi_binary_label", 50, false); - - config.evaluatorConfig.set_classification_threshold(0.4); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - // Not support GPU - testEvaluator( - config, "classification_error_weight_multi_binary_label", 50, false); -} - -TEST(Evaluator, sum) { - TestConfig config; - config.evaluatorConfig.set_type("sum"); - - // sum of output - config.inputDefs.push_back({INPUT_DATA, "output", 10}); - testEvaluatorAll(config, "sum_output", 200); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - testEvaluatorAll(config, "sum_output_weight", 200); - - // sum of label - config.inputDefs.clear(); - config.inputDefs.push_back({INPUT_LABEL, "label", 10}); - testEvaluatorAll(config, "sum_label", 200); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - testEvaluatorAll(config, "sum_label_weight", 200); -} - -TEST(Evaluator, last_column_sum) { - TestConfig config; - config.evaluatorConfig.set_type("last-column-sum"); - - config.inputDefs.push_back({INPUT_DATA, "output", 50}); - testEvaluatorAll(config, "last-column-sum", 200); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - testEvaluatorAll(config, "last-column-sum_weight", 200); -} - -TEST(Evaluator, last_column_auc) { - TestConfig config; - config.evaluatorConfig.set_type("last-column-auc"); - - config.inputDefs.push_back({INPUT_DATA, "output", 2}); - config.inputDefs.push_back({INPUT_LABEL, "label", 2}); - testEvaluatorAll(config, "last-column-auc", 500); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - testEvaluatorAll(config, "last-column-auc_weight", 200); -} - -TEST(Evaluator, precision_recall) { - TestConfig config; - config.evaluatorConfig.set_type("precision_recall"); - - config.inputDefs.push_back({INPUT_DATA, "output", 10}); - config.inputDefs.push_back({INPUT_LABEL, "label", 10}); - testEvaluatorAll(config, "precision_recall", 200); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - testEvaluatorAll(config, "precision_recall_weight", 200); - - LOG(INFO) << "positive_label = 5"; - config.evaluatorConfig.set_positive_label(5); - testEvaluatorAll(config, "precision_recall_weight", 200); - - // multi binary labels - config.inputDefs.clear(); - config.evaluatorConfig.set_positive_label(-1); - config.inputDefs.push_back({INPUT_DATA, "output", 10}); - config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "label", 10}); - // Not support GPU - testEvaluator(config, "precision_recall_multi_binary_label", 100, false); - - LOG(INFO) << "classification_threshold = 0.4"; - config.evaluatorConfig.set_classification_threshold(0.4); - config.inputDefs.push_back({INPUT_DATA, "weight", 1}); - // Not support GPU - testEvaluator( - config, "precision_recall_weight_multi_binary_label", 100, false); -} - -TEST(Evaluator, ctc_error_evaluator) { - TestConfig config; - config.evaluatorConfig.set_type("ctc_edit_distance"); - - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "output", 32}); - config.inputDefs.push_back({INPUT_SEQUENCE_LABEL, "label", 1}); - testEvaluatorAll(config, "ctc_error_evaluator", 100); -} - -int main(int argc, char** argv) { - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_KmaxSeqScore.cpp b/paddle/gserver/tests/test_KmaxSeqScore.cpp deleted file mode 100644 index 168ffbdac8cd6fb0ee4fa62e3766905c30d1844b..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_KmaxSeqScore.cpp +++ /dev/null @@ -1,164 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/utils/GlobalConstants.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(gpu_id); -DECLARE_bool(thread_local_rand_use_global_seed); - -vector randSampling(int range, int n) { - CHECK_GE(range, n); - vector num(range); - iota(begin(num), end(num), 0); - if (range == n) return num; - - random_shuffle(begin(num), end(num)); - num.resize(n); - return num; -} - -void genRandomSeqInfo(vector& seqStartPosition, - vector& subSeqStartPosition) { - const int maxSeqNum = 100; - // generate random start position information - int seqNum = 1 + (rand() % maxSeqNum); - seqStartPosition.resize(seqNum + 1, 0); - subSeqStartPosition.resize(1, 0); - - for (int i = 0; i < seqNum; ++i) { - int subSeqLen = 1 + (rand() % maxSeqNum); - for (int j = 0; j < subSeqLen; ++j) - subSeqStartPosition.push_back(subSeqStartPosition.back() + subSeqLen); - seqStartPosition[i + 1] = subSeqStartPosition.back(); - } -} - -void genRandomGroundTruth(real* values, - vector>& groundTruth, - vector& startPos, - size_t beamSize) { - groundTruth.resize(startPos.size() - 1, vector(beamSize, -1)); - for (size_t i = 0; i < startPos.size() - 1; ++i) { - int seqLen = startPos[i + 1] - startPos[i]; - vector pos = - randSampling(seqLen, min(static_cast(beamSize), seqLen)); - for (size_t j = 0; j < pos.size(); ++j) { - groundTruth[i][j] = pos[j]; - values[startPos[i] + pos[j]] = 1.; - } - } -} - -void checkLayerOut(vector> groundTruth, - real* layerOut, - size_t beamSize) { - for (size_t i = 0; i < groundTruth.size(); ++i) { - int begPos = i * beamSize; - vector tmp(layerOut + begPos, layerOut + begPos + beamSize); - sort(begin(tmp), end(tmp)); - sort(begin(groundTruth[i]), end(groundTruth[i])); - for (size_t j = 0; j < beamSize; ++j) CHECK_EQ(tmp[j], groundTruth[i][j]); - } -} - -TEST(Layer, kmaxSeqScoreLayer) { - const size_t maxBeamSize = 100; - size_t beamSize = 1 + (rand() % maxBeamSize); - - vector seqStartPosition; - vector subSeqStartPosition; - genRandomSeqInfo(seqStartPosition, subSeqStartPosition); - MatrixPtr inValue = - Matrix::create(subSeqStartPosition.back(), 1, false, false); - - std::vector mode = {false}; -#ifdef PADDLE_WITH_CUDA - mode.push_back(true); -#endif - - for (auto hasSubseq : {false, true}) { - vector> groundTruth; - inValue->randomizeUniform(); - genRandomGroundTruth(inValue->getData(), - groundTruth, - hasSubseq ? subSeqStartPosition : seqStartPosition, - beamSize); - - for (auto useGpu : mode) { - TestConfig config; - config.layerConfig.set_type("kmax_seq_score"); - config.layerConfig.set_beam_size(beamSize); - - if (hasSubseq) { - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, - "scores", - inValue, - seqStartPosition, - subSeqStartPosition}); - } else { - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, "scores", inValue, seqStartPosition}); - } - config.layerConfig.add_inputs(); - - // data layer initialize - std::vector dataLayers; - LayerMap layerMap; - vector datas; - initDataLayer( - config, - &dataLayers, - &datas, - &layerMap, - "kmax_seq_score", - 100 /* actually this parameter is unused in self-defined input*/, - false, - useGpu); - // test layer initialize - std::vector parameters; - LayerPtr kmaxSeqScoreLayer; - FLAGS_use_gpu = useGpu; - initTestLayer(config, &layerMap, ¶meters, &kmaxSeqScoreLayer); - kmaxSeqScoreLayer->forward(PASS_TRAIN); - - const MatrixPtr outValue = kmaxSeqScoreLayer->getOutputValue(); - CHECK_EQ(outValue->getHeight(), - hasSubseq ? subSeqStartPosition.size() - 1 - : seqStartPosition.size() - 1); - CHECK_EQ(outValue->getWidth(), beamSize); - checkLayerOut(groundTruth, outValue->getData(), beamSize); - } - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand((size_t)(time(NULL))); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp deleted file mode 100644 index 1254d580505512dc8fd7e34a053a7538832d271f..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ /dev/null @@ -1,2532 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_CUDA -#include -#endif -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/math/MathUtils.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(gpu_id); -DECLARE_double(checkgrad_eps); -DECLARE_bool(thread_local_rand_use_global_seed); -DECLARE_bool(prev_batch_state); - -TEST(Operator, dot_mul) { - TestConfig config; - config.layerConfig.set_size(10); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs(); - operatorConf.set_type("dot_mul"); - operatorConf.set_dotmul_scale(-1); - - testOperatorGrad(config, operatorConf, 100, false, false); -} - -TEST(Projection, context) { - for (auto contextStart : {-5, -3, -1, 0, 3}) { - for (auto contextLength : {1, 2, 5, 7}) { - for (auto batchSize : {1, 2, 5, 20}) { - for (auto trainablePadding : {false, true}) { - LOG(INFO) << " contextStart=" << contextStart - << " contextLength=" << contextLength - << " batchSize=" << batchSize - << " trainablePadding=" << trainablePadding; - ProjectionConfig conf; - conf.set_type("context"); - conf.set_input_size(10); - conf.set_context_start(contextStart); - conf.set_context_length(contextLength); - conf.set_trainable_padding(trainablePadding); - conf.set_output_size(conf.context_length() * conf.input_size()); - int pad = - std::max(0, -conf.context_start()) + - std::max(0, conf.context_start() + conf.context_length() - 1); - for (auto useGpu : {false, true}) { - testProjectionGrad( - conf, - INPUT_SEQUENCE_DATA, - trainablePadding ? conf.input_size() * pad : 0, - batchSize, - useGpu, - contextStart + contextLength <= 1); // = testState - } - } - } - } - } -} - -TEST(Projection, trans_fc) { - ProjectionConfig conf; - conf.set_type("trans_fc"); - conf.set_input_size(50); - conf.set_output_size(20); - for (auto useGpu : {false, true}) { - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ 1000, - /* batchSize */ 100, - useGpu); - } -} - -TEST(Projection, fc) { - ProjectionConfig conf; - conf.set_type("fc"); - conf.set_input_size(10); - conf.set_output_size(20); - for (auto useGpu : {false, true}) { - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ 200, - /* batchSize */ 100, - useGpu); - } -} - -TEST(Projection, dot_mul) { - ProjectionConfig conf; - conf.set_type("dot_mul"); - conf.set_input_size(20); - conf.set_output_size(20); - for (auto useGpu : {false, true}) { - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ 20, - /* batchSize */ 100, - useGpu); - } -} - -TEST(Projection, table) { - ProjectionConfig conf; - conf.set_type("table"); - conf.set_input_size(10); - conf.set_output_size(20); - for (auto useGpu : {false, true}) { - testProjectionGrad(conf, - INPUT_LABEL, - /* parameterSize */ 200, - /* batchSize */ 100, - useGpu); - } -} - -TEST(Projection, identity) { - ProjectionConfig conf; - conf.set_type("identity"); - conf.set_input_size(10); - conf.set_output_size(10); - for (auto useGpu : {false, true}) { - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ 0, - /* batchSize */ 100, - useGpu); - } -} - -TEST(Projection, slice) { - ProjectionConfig conf; - conf.set_type("slice"); - conf.set_input_size(100); - SliceConfig& slice1 = *conf.add_slices(); - slice1.set_start(10); - slice1.set_end(20); - SliceConfig& slice2 = *conf.add_slices(); - slice2.set_start(50); - slice2.set_end(70); - conf.set_output_size(30); - for (auto useGpu : {false, true}) { - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ 0, - /* batchSize */ 10, - useGpu); - } -} - -TEST(Projection, scaling) { - ProjectionConfig conf; - conf.set_type("scaling"); - conf.set_input_size(10); - conf.set_output_size(10); - for (auto useGpu : {false}) { - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ 1, - /* batchSize */ 100, - useGpu); - } -} - -void testProjectionConv(size_t groups, bool isDeconv) { - const int NUM_FILTERS = 18; - const int FILTER_SIZE = 2; - const int FILTER_SIZE_Y = 2; - const int CHANNELS = 3; - const int IMAGE_SIZE = 16; - -#if CUDNN_VERSION >= 6000 - const int DILATION = 2; -#else - const int DILATION = 1; -#endif - - ProjectionConfig conf; - if (isDeconv) { - conf.set_type("convt"); - } else { - conf.set_type("conv"); - } - conf.set_num_filters(NUM_FILTERS); - - ConvConfig* conv = conf.mutable_conv_conf(); - conv->set_filter_size(FILTER_SIZE); - conv->set_filter_size_y(FILTER_SIZE_Y); - conv->set_channels(CHANNELS); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_dilation(DILATION); - conv->set_dilation_y(DILATION); - conv->set_groups(groups); - if (isDeconv) { - conv->set_filter_channels(NUM_FILTERS / conv->groups()); - } else { - conv->set_filter_channels(conv->channels() / conv->groups()); - } - conv->set_img_size(IMAGE_SIZE); - int output_x = outputSize(conv->img_size(), - (conv->filter_size() - 1) * DILATION + 1, - conv->padding(), - conv->stride(), - /* caffeMode */ true); - int output_y = outputSize(conv->img_size(), - (conv->filter_size_y() - 1) * DILATION + 1, - conv->padding_y(), - conv->stride_y(), - /* caffeMode */ true); - conv->set_output_x(output_x); - conv->set_output_y(output_y); - LOG(INFO) << "DILATION:" << DILATION << "; output_x: " << output_x - << "; output_y: " << output_y; - if (isDeconv) { - int deconv_image_x = imageSize(output_x, - (conv->filter_size() - 1) * DILATION + 1, - conv->padding(), - conv->stride(), - /* caffeMode */ true); - int deconv_image_y = imageSize(output_y, - (conv->filter_size_y() - 1) * DILATION + 1, - conv->padding_y(), - conv->stride_y(), - /* caffeMode */ true); - - LOG(INFO) << " deconv_image_x: " << deconv_image_x - << "; deconv_image_y: " << deconv_image_y; - conf.set_input_size(output_x * output_y * CHANNELS); - conf.set_output_size(deconv_image_x * deconv_image_y * NUM_FILTERS); - } else { - conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS); - conf.set_output_size(output_x * output_y * NUM_FILTERS); - } - - testProjectionGrad(conf, - INPUT_DATA, - /* parameterSize */ NUM_FILTERS * CHANNELS * FILTER_SIZE * - FILTER_SIZE_Y / groups, - /* batchSize */ 100, - true, - false, - NUM_FILTERS, - true); -} - -#ifdef PADDLE_WITH_CUDA -TEST(Projection, conv) { - /// test ConvProjection - testProjectionConv(1, false); - testProjectionConv(3, false); - /// test ConvTransProjection - testProjectionConv(1, true); - testProjectionConv(3, true); -} -#endif - -TEST(Layer, BilinearInterpLayer) { - TestConfig config; - config.layerConfig.set_type("bilinear_interp"); - config.biasSize = 0; - config.inputDefs.push_back({INPUT_DATA, "layer_0", 4096, 0}); - - LayerInputConfig* input = config.layerConfig.add_inputs(); - BilinearInterpConfig* bilinear = input->mutable_bilinear_interp_conf(); - ImageConfig* image = bilinear->mutable_image_conf(); - image->set_img_size(32); - image->set_img_size_y(32); - image->set_channels(4); - - for (auto useGpu : {false, true}) { - for (auto outSize : {32, 64}) { - bilinear->set_out_size_x(outSize); - bilinear->set_out_size_y(outSize); - testLayerGrad(config, "bilinear_interp", 10, false, useGpu); - } - } -} - -TEST(Layer, concat) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("concat"); - config.layerConfig.set_size(15); - config.layerConfig.set_active_type("sigmoid"); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 5, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "concat", 100, false, useGpu); - } -} - -TEST(Layer, AddtoLayer) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("addto"); - config.layerConfig.set_size(10); - config.layerConfig.set_active_type("sigmoid"); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "addto", 100, false, useGpu); - } -} - -TEST(Layer, CTCLayer) { - TestConfig config; - config.layerConfig.set_type("ctc"); - config.layerConfig.set_norm_by_times(false); - config.layerConfig.set_size(10); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_SEQUENCE_LABEL, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "ctc", - 100, - /* trans */ false, /* useGpu */ - useGpu); - } -} - -TEST(Layer, cosSimLayer) { - TestConfig config; - config.layerConfig.set_type("cos"); - config.layerConfig.set_size(1); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 50, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "cos", 100, false, useGpu); - } -} - -TEST(Layer, CosSimVecMatLayer) { - TestConfig config; - config.layerConfig.set_type("cos_vm"); - config.layerConfig.set_size(5); // output size - config.layerConfig.set_cos_scale(2.0); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 20, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 100, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "cos_vm", 100, false, useGpu); - } -} - -void testDepthwiseConvLayer(const string& type, bool useGpu) { - TestConfig config; - config.biasSize = 32; - config.layerConfig.set_type(type); - config.layerConfig.set_num_filters(32); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 2048, 192}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(2); - conv->set_filter_size_y(3); - conv->set_channels(16); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_groups(16); - conv->set_filter_channels(conv->channels() / conv->groups()); - conv->set_img_size(16); - conv->set_img_size_y(8); - conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - conv->set_output_y(outputSize(conv->img_size_y(), - conv->filter_size_y(), - conv->padding_y(), - conv->stride_y(), - /* caffeMode */ true)); - config.layerConfig.set_size(conv->output_x() * conv->output_y() * - config.layerConfig.num_filters()); - - testLayerGrad(config, "depthwise_conv", 100, false, useGpu); - // Use small batch_size and useWeight=true to test biasGrad - testLayerGrad(config, "depthwise_conv", 2, false, useGpu, true, 0.02); -} - -TEST(Layer, depthwiseConvLayer) { - // 'depthwise_conv' is a sepecial case of 'exconv' whose - // groups size equals to the input channels size. - testDepthwiseConvLayer("exconv", /* useGpu= */ false); -#ifdef PADDLE_WITH_CUDA - testDepthwiseConvLayer("exconv", /* useGpu= */ true); -#endif -} - -void testConvLayer(const string& type, bool trans, bool useGpu) { - TestConfig config; - config.biasSize = 16; - config.layerConfig.set_type(type); - config.layerConfig.set_num_filters(16); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - int dilation = 2; - if (type == "cudnn_conv") { -#if CUDNN_VERSION >= 6000 - dilation = 2; -#else - dilation = 1; -#endif - } - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 192}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(2); - conv->set_filter_size_y(2); - conv->set_channels(3); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_dilation(dilation); - conv->set_dilation_y(dilation); - conv->set_groups(1); - conv->set_filter_channels(conv->channels() / conv->groups()); - conv->set_img_size(16); - conv->set_img_size_y(16); - conv->set_output_x(outputSize(conv->img_size(), - (conv->filter_size() - 1) * dilation + 1, - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - conv->set_output_y(outputSize(conv->img_size_y(), - (conv->filter_size_y() - 1) * dilation + 1, - conv->padding_y(), - conv->stride_y(), - /* caffeMode */ true)); - config.layerConfig.set_size(conv->output_x() * conv->output_y() * - config.layerConfig.num_filters()); - - testLayerGrad(config, "conv", 100, trans, useGpu); - // Use small batch_size and useWeight=true to test biasGrad - testLayerGrad(config, "conv", 2, trans, useGpu, true, 0.02); -} - -TEST(Layer, convLayer) { - testConvLayer("exconv", /* trans= */ false, /* useGpu= */ false); -#ifdef PADDLE_WITH_CUDA - testConvLayer("exconv", /* trans= */ false, /* useGpu= */ true); - testConvLayer("cudnn_conv", /* trans= */ false, /* useGpu= */ true); -#endif -} - -void testConvTransLayer(const string& type, bool trans, bool useGpu) { - TestConfig config; - config.biasSize = 3; - config.layerConfig.set_type(type); - config.layerConfig.set_num_filters(3); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 384}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_filter_size(2); - conv->set_filter_size_y(4); - conv->set_channels(16); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_groups(1); - conv->set_filter_channels(3 / conv->groups()); - conv->set_img_size(16); - conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - - config.layerConfig.set_size(conv->img_size() * conv->img_size() * - config.layerConfig.num_filters()); - - testLayerGrad(config, "convTrans", 100, trans, useGpu); - // Use small batch_size and useWeight=true to test biasGrad - testLayerGrad(config, "convTrans", 2, trans, useGpu, true, 0.02); -} - -TEST(Layer, convTransLayer) { - for (auto useGpu : {false, true}) { - testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu); - } -#ifdef PADDLE_WITH_CUDA - testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true); -#endif -} - -TEST(Layer, blockExpandLayer) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("blockexpand"); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 6144, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - BlockExpandConfig* blockExpand = input->mutable_block_expand_conf(); - blockExpand->set_img_size_x(64); - blockExpand->set_img_size_y(32); - blockExpand->set_channels(3); - blockExpand->set_padding_x(0); - blockExpand->set_padding_y(0); - blockExpand->set_block_x(4); - blockExpand->set_block_y(32); - blockExpand->set_stride_x(2); - blockExpand->set_stride_y(2); - blockExpand->set_output_x(outputSize(blockExpand->img_size_x(), - blockExpand->block_x(), - blockExpand->padding_x(), - blockExpand->stride_x(), - /* caffeMode */ false)); - blockExpand->set_output_y(outputSize(blockExpand->img_size_y(), - blockExpand->block_y(), - blockExpand->padding_y(), - blockExpand->stride_y(), - /* caffeMode */ false)); - config.layerConfig.set_size(blockExpand->block_x() * blockExpand->block_y() * - blockExpand->channels()); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "blockexpand", 100, false, useGpu); - } -} - -TEST(Layer, maxoutLayer) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("maxout"); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 4096, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - MaxOutConfig* maxout = input->mutable_maxout_conf(); - ImageConfig* image = maxout->mutable_image_conf(); - - image->set_img_size(32); - image->set_img_size_y(32); - image->set_channels(4); - maxout->set_groups(2); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "maxout", 10, false, useGpu); - } -} - -void testFcLayer(string format, size_t nnz) { - TestConfig config; - config.biasSize = 1024; - config.layerConfig.set_type("fc"); - config.layerConfig.set_size(1024); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_drop_rate(0.1); - - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", 2048, nnz, ParaSparse(format)}); - config.layerConfig.add_inputs(); - - LOG(INFO) << config.inputDefs[0].sparse.sparse << " " - << config.inputDefs[0].sparse.format; - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "fc", - 100, - /* trans */ false, - useGpu, - /* weight */ true); - } -} - -TEST(Layer, fcLayer) { - testFcLayer("", 1024 * 1024 * 2); - testFcLayer("csc", 1024 * 10); - testFcLayer("csr", 1024 * 10); -} - -TEST(Layer, SelectiveFullyConnectedLayer) { - TestConfig config; - size_t nin = 16; - size_t nout = 256; - config.layerConfig.set_type("selective_fc"); - config.layerConfig.set_size(nout); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_has_selected_colums(true); - config.layerConfig.set_selective_fc_pass_generation(false); - config.biasSize = nout; - - config.inputDefs.push_back({INPUT_DATA, "input0", nin, nin * nout}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back( - {INPUT_SPARSE_NON_VALUE_DATA, "index", nout, 0, ParaSparse("csr", true)}); - config.layerConfig.add_inputs(); - - testLayerGrad(config, - "selective_fc", - 100, - /* trans= */ false, - /* useGup= */ false, - false); -#ifdef PADDLE_WITH_CUDA - testLayerGrad(config, - "selective_fc", - 100, - /* trans= */ false, - /* useGup= */ true, - false); -#endif -} - -TEST(Layer, DataNormLayer) { - TestConfig config; - config.layerConfig.set_type("data_norm"); - config.layerConfig.set_size(20); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 20, 100}); - config.inputDefs.back().isStatic = true; - config.layerConfig.add_inputs(); - - for (auto strategy : {"z-score", "min-max", "decimal-scaling"}) { - config.layerConfig.set_data_norm_strategy(strategy); - // The parameters are static, so not support GPU now - testLayerGrad(config, - "data_norm", - 200, - /* trans */ false, - /* useGpu */ false); - } -} - -TEST(Layer, hsigmoidLayer) { - TestConfig config; - config.layerConfig.set_type("hsigmoid"); - config.layerConfig.set_num_classes(5); - config.layerConfig.set_size(1); - config.biasSize = config.layerConfig.num_classes() - 1; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 200}); - config.inputDefs.push_back({INPUT_LABEL, "layer_1", 5, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "hsigmoid", - 100, - /* trans */ false, - /* useGpu */ useGpu); - } -} - -TEST(Layer, multi_cross) { - TestConfig config; - config.layerConfig.set_type("multi-class-cross-entropy"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_LABEL, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad( - config, "multi-class-cross-entropy", 100, /* trans */ false, useGpu); - } -} - -TEST(Layer, multi_binary_label_sparse_mat) { - TestConfig config; - config.layerConfig.set_type("multi_binary_label_cross_entropy"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "layer_1", 50, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "multi_binary_label_cross_entropy", - 100, - /* trans */ false, - useGpu); - } -} - -TEST(layer, multi_binary_label_id) { - TestConfig config; - config.layerConfig.set_type("multi_binary_label_cross_entropy"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_LABEL, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "multi_binary_label_cross_entropy", - 100, - /* trans */ false, - useGpu); - } -} - -TEST(Layer, multi_cross_with_selfnorm) { - TestConfig config; - config.layerConfig.set_type("multi_class_cross_entropy_with_selfnorm"); - config.layerConfig.set_softmax_selfnorm_alpha(0.1); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_LABEL, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - // Not support GPU now - testLayerGrad(config, - "multi_class_cross_entropy_with_selfnorm", - 100, - /* trans */ false, - /* useGpu */ false); -} - -TEST(Layer, multi_cross_soft) { - TestConfig config; - config.layerConfig.set_type("soft_binary_class_cross_entropy"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "soft_binary_class_cross_entropy", - 100, - /* trans */ false, - useGpu); - } -} - -TEST(Layer, square_error) { - TestConfig config; - config.layerConfig.set_type("square_error"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "square_error", 100, /* trans */ false, useGpu); - } -} - -TEST(Layer, sparse_square_error) { - TestConfig config; - config.layerConfig.set_type("square_error"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "layer_1", 50, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - // "GpuSparseMatrix" as label is not supported - testLayerGrad(config, - "square_error", - 100, - /* trans */ false, - /* useGpu */ false); -} - -TEST(Layer, sparse_float_square_error) { - TestConfig config; - config.layerConfig.set_type("square_error"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); - config.inputDefs.push_back({INPUT_SPARSE_FLOAT_VALUE_DATA, "layer_1", 50, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - // "GpuSparseMatrix" as label is not supported - testLayerGrad(config, - "square_error", - 100, - /* trans */ false, - /* useGpu */ false); -} - -TEST(Layer, square_error_weighted) { - TestConfig config; - config.layerConfig.set_type("square_error"); - config.biasSize = 0; - config.testAccumulate = false; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_2", 1, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "square_error", 100, /* trans */ false, useGpu); - } -} - -TEST(Layer, huber_regression_loss) { - TestConfig config; - config.layerConfig.set_type("huber_regression"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - for (auto delta : {1, 3, 5}) { - config.layerConfig.set_delta(delta); - testLayerGrad(config, "huber_regression", 100, /* trans */ false, useGpu); - } - } -} - -TEST(Layer, huber_two_class) { - TestConfig config; - config.layerConfig.set_type("huber_classification"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.inputDefs.push_back({INPUT_LABEL, "layer_1", 2, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "huber_two_class", 100, /* trans */ false, useGpu); - } -} - -void testExpandLayer(string trans_type, bool hasSubseq) { - TestConfig config; - config.layerConfig.set_type("expand"); - - config.inputDefs.push_back( - {trans_type == "non-seq" ? INPUT_DENSE_DIM_DATA : INPUT_SEQUENCE_DATA, - "layer_0", - 10, - 0}); - config.inputDefs.push_back( - {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, - "layer_1", - 10, - 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.set_trans_type(trans_type); - LOG(INFO) << " trans_type=" << trans_type << " hasSubseq=" << hasSubseq; - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "expand", 30, false, useGpu); - } -} - -TEST(Layer, ExpandLayer) { - testExpandLayer("non-seq", false); // non-seq expand to seq - testExpandLayer("non-seq", true); // non-seq expand to hasSubseq - testExpandLayer("seq", true); // seq expand to hasSubseq -} - -void testDegradeLayer(bool hasSubseq, - string layer_type, - string trans_type, - int stride) { - TestConfig config; - config.layerConfig.set_type(layer_type); - config.layerConfig.set_size(10); - config.layerConfig.set_seq_pool_stride(stride); - config.biasSize = 0; - - config.inputDefs.push_back( - {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, - "layer_0", - 10, - 0}); - config.layerConfig.add_inputs(); - config.layerConfig.set_trans_type(trans_type); - - auto testDegradeLayerGrad = [](TestConfig& config, string layer_type) { - for (auto useGpu : {false, true}) { - testLayerGrad(config, layer_type, 100, false, useGpu); - } - }; - - if (layer_type == "average") { - for (auto strategy : {"average", "sum", "squarerootn"}) { - LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type - << " average_strategy=" << strategy - << " seq_pool_stride=" << stride; - config.layerConfig.set_average_strategy(strategy); - testDegradeLayerGrad(config, layer_type); - } - } else { - LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type - << " seq_pool_stride=" << stride; - testDegradeLayerGrad(config, layer_type); - } -} - -TEST(Layer, MaxLayer) { - testDegradeLayer(false, "max", "non-seq", -1); // seq max to non-seq - testDegradeLayer(false, - "max", - "non-seq", - 5); // seq max to a shorten seq, stride window = 5 - testDegradeLayer(true, "max", "non-seq", -1); // hasSubseq max to non-seq - testDegradeLayer(true, "max", "seq", -1); // hasSubseq max to seq -} - -TEST(Layer, SequenceLastInstanceLayer) { - testDegradeLayer(false, - "seqlastins", - "non-seq", - -1); // seq seqlastins to non-seq - testDegradeLayer(false, - "seqlastins", - "non-seq", - 5); // seq seqlastins to a shorten seq, stride window = 5 - testDegradeLayer(true, - "seqlastins", - "non-seq", - -1); // hasSubseq seqlastins to non-seq - testDegradeLayer(true, - "seqlastins", - "seq", - -1); // hasSubseq seqlastins to seq -} - -TEST(Layer, AverageLayer) { - testDegradeLayer(false, "average", "non-seq", -1); // seq average to non-seq - testDegradeLayer(false, - "average", - "non-seq", - 5); // seq average to a shorten seq, stride window = 5 - testDegradeLayer(true, - "average", - "non-seq", - -1); // hasSubseq average to non-seq - testDegradeLayer(true, "average", "seq", -1); // hasSubseq average to seq -} - -TEST(Layer, SequenceConcatLayer) { - TestConfig config; - config.layerConfig.set_type("seqconcat"); - config.layerConfig.set_size(10); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_0", 10, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "seqconcat", 100, false, useGpu); - } -} - -TEST(Layer, SequenceReshapeLayer) { - TestConfig config; - config.layerConfig.set_type("seqreshape"); - config.layerConfig.set_size(10); - - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_0", 100, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "seqreshape", 100, false, useGpu); - } -} - -TEST(Layer, ConvShiftLayer) { - TestConfig config; - config.layerConfig.set_type("conv_shift"); - config.layerConfig.set_size(10); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 3, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - // Not support GPU now - testLayerGrad(config, "conv_shift", 100, false, false); -} - -TEST(Layer, PowerLayer) { - TestConfig config; - config.layerConfig.set_type("power"); - config.layerConfig.set_size(10); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "power", 100, false, useGpu); - } -} - -TEST(Layer, ConvexCombinationLayer) { - TestConfig config; - config.layerConfig.set_type("convex_comb"); - config.layerConfig.set_size(20); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 5, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 100, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "convex_comb", 100, false, useGpu); - } -} - -TEST(Layer, InterpolationLayer) { - TestConfig config; - config.layerConfig.set_type("interpolation"); - config.layerConfig.set_size(10); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_2", 10, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "interpolation", 100, false, useGpu); - } -} - -TEST(Layer, DotProdLayer) { - TestConfig config; - config.layerConfig.set_type("dot_prod"); - config.layerConfig.set_size(1); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "dot_prod", 10, false, useGpu); - } -} - -TEST(Layer, OuterProdLayer) { - TestConfig config; - config.layerConfig.set_type("out_prod"); - config.layerConfig.set_size(100); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "out_prod", 100, false, useGpu); - } -} - -TEST(Layer, SlopeInterceptLayer) { - TestConfig config; - config.layerConfig.set_type("slope_intercept"); - config.layerConfig.set_size(10); - config.layerConfig.set_slope(1.0); - config.layerConfig.set_intercept(0.1); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "slope_intercept", 100, false, useGpu); - } -} - -TEST(Layer, ScalingLayer) { - TestConfig config; - config.layerConfig.set_type("scaling"); - config.layerConfig.set_size(10); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "scaling", 100, false, useGpu); - } -} - -void testNormLayer(const string& normType, bool trans, bool useGpu) { - TestConfig config; - config.layerConfig.set_type("norm"); - config.layerConfig.set_active_type("relu"); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1568, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - NormConfig* norm = input->mutable_norm_conf(); - norm->set_norm_type(normType); - norm->set_channels(16); - norm->set_size(5); - norm->set_scale(0.001); - norm->set_pow(0.75); - norm->set_blocked(0); - norm->set_img_size(14); - norm->set_img_size_y(7); - norm->set_output_x(norm->img_size()); - norm->set_output_y(norm->img_size_y()); - if (norm->norm_type() == "cmrnorm" || - norm->norm_type() == "cmrnorm-projection") { - norm->set_scale(norm->scale() / norm->size()); - } else { - norm->set_scale(norm->scale() / (norm->size() * norm->size())); - } - - config.layerConfig.set_size(norm->output_x() * norm->output_y() * - norm->channels()); - config.biasSize = 0; - - testLayerGrad(config, "norm", 100, trans, useGpu); -} - -TEST(Layer, NormLayer) { - testNormLayer("cmrnorm-projection", - /* trans= */ false, /* useGpu= */ - true); - testNormLayer("cmrnorm-projection", - /* trans= */ false, /* useGpu= */ - false); -} - -void setPoolConfig(TestConfig* config, - PoolConfig* pool, - const string& poolType) { - (*config).biasSize = 0; - (*config).layerConfig.set_type("pool"); - (*config).layerConfig.set_num_filters(16); - - int kw = 3, kh = 3; - int pw = 0, ph = 0; - int sw = 2, sh = 2; - pool->set_pool_type(poolType); - pool->set_channels(16); - pool->set_size_x(kw); - pool->set_size_y(kh); - pool->set_start(0); - pool->set_padding(pw); - pool->set_padding_y(ph); - pool->set_stride(sw); - pool->set_stride_y(sh); - - int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); - int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); - pool->set_output_x(ow); - pool->set_output_y(oh); -} - -void testPoolLayer(const string& poolType, - bool trans, - bool useGpu, - bool excludeMode = true) { - TestConfig config; - config.inputDefs.push_back({INPUT_DATA, "layer_0", 3136, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - PoolConfig* pool = input->mutable_pool_conf(); - - pool->set_img_size(14); - pool->set_img_size_y(14); - pool->set_exclude_mode(excludeMode); - setPoolConfig(&config, pool, poolType); - config.layerConfig.set_size(pool->output_x() * pool->output_y() * - pool->channels()); - - testLayerGrad(config, "pool", 100, trans, useGpu); -} - -#ifdef PADDLE_WITH_CUDA -void testPoolLayer2(const string& poolType, bool trans, bool useGpu) { - TestConfig config; - config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - PoolConfig* pool = input->mutable_pool_conf(); - - pool->set_size_y(4); - pool->set_stride_y(3); - pool->set_img_size(10); - pool->set_img_size_y(20); - setPoolConfig(&config, pool, poolType); - pool->set_output_y((pool->img_size_y() - pool->start() - pool->size_y()) / - ((float)pool->stride_y()) + - 1.5); - config.layerConfig.set_size(pool->output_x() * pool->output_y() * - pool->channels()); - - testLayerGrad(config, "pool", 100, trans, useGpu); -} -#endif - -TEST(Layer, PoolLayer) { - testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false); - testPoolLayer("avg-projection", - /* trans= */ false, - /* useGpu= */ false, - /* excludeMode= */ false); - testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false); - testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ false); - -#ifdef PADDLE_WITH_CUDA - testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true); - testPoolLayer("avg-projection", - /* trans= */ false, - /* useGpu= */ true, - /* excludeMode= */ false); - testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true); - testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); - testPoolLayer("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true); - testPoolLayer2("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); - testPoolLayer2("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true); - testPoolLayer2("cudnn-avg-incl-pad-pool", - /* trans= */ false, - /* useGpu= */ true); - testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ true); -#endif -} - -void setPool3DConfig(TestConfig* config, - PoolConfig* pool, - const string& poolType) { - // filter size - const int NUM_FILTERS = 16; - const int FILTER_SIZE = 3; - const int FILTER_SIZE_Y = 3; - const int FILTER_SIZE_Z = 3; - const int CHANNELS = 16; - - (*config).biasSize = 0; - (*config).layerConfig.set_type("pool3d"); - (*config).layerConfig.set_num_filters(NUM_FILTERS); - - int kw = FILTER_SIZE, kh = FILTER_SIZE_Y, kd = FILTER_SIZE_Z; - int pw = 0, ph = 0, pd = 0; - int sw = 2, sh = 2, sd = 2; - - pool->set_pool_type(poolType); - pool->set_pool_type("avg"); - pool->set_channels(CHANNELS); - pool->set_size_x(kw); - pool->set_size_y(kh); - pool->set_size_z(kd); - pool->set_padding(0); - pool->set_padding_y(0); - pool->set_padding_z(0); - pool->set_stride(sw); - pool->set_stride_y(sh); - pool->set_stride_z(sd); - pool->set_start(0); - int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); - int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); - int od = outputSize(pool->img_size_z(), kd, pd, sd, /* caffeMode */ false); - pool->set_output_x(ow); - pool->set_output_y(oh); - pool->set_output_z(od); -} - -void testPool3DLayer(const string& poolType, bool trans, bool useGpu) { - TestConfig config; - config.inputDefs.push_back({INPUT_DATA, "layer_0", 11664, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - PoolConfig* pool = input->mutable_pool_conf(); - - const int IMAGE_SIZE = 9; - const int IMAGE_SIZE_Y = 9; - const int IMAGE_SIZE_Z = 9; - - pool->set_img_size(IMAGE_SIZE); - pool->set_img_size_y(IMAGE_SIZE_Y); - pool->set_img_size_z(IMAGE_SIZE_Z); - - setPool3DConfig(&config, pool, poolType); - config.layerConfig.set_size(pool->output_x() * pool->output_y() * - pool->channels()); - - testLayerGrad(config, "pool3d", 100, trans, useGpu); -} - -TEST(Layer, Pool3DLayer) { - testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false); - testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false); -#ifdef PADDLE_WITH_CUDA - testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true); - testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true); -#endif -} - -void testSppLayer(const string& poolType, - const int pyramidHeight, - bool trans, - bool useGpu) { - TestConfig config; - config.layerConfig.set_type("spp"); - config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - SppConfig* sppConfig = input->mutable_spp_conf(); - sppConfig->set_pool_type(poolType); - sppConfig->set_pyramid_height(pyramidHeight); - ImageConfig* imageConfig = sppConfig->mutable_image_conf(); - imageConfig->set_channels(16); - imageConfig->set_img_size(10); - imageConfig->set_img_size_y(20); - int outputSize = (std::pow(4, sppConfig->pyramid_height()) - 1) / (4 - 1); - config.layerConfig.set_size(outputSize * imageConfig->channels()); - testLayerGrad(config, "spp", 100, trans, useGpu); -} - -TEST(Layer, SpatialPyramidPoolLayer) { - for (auto useGpu : {false, true}) { - for (auto pyramidHeight : {1, 2, 3}) { - testSppLayer("avg-projection", pyramidHeight, false, useGpu); - testSppLayer("max-projection", pyramidHeight, false, useGpu); - } - } -} - -TEST(Layer, rankCostLayer) { - TestConfig config; - config.layerConfig.set_type("rank-cost"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 1, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_2", 1, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "rank-cost", 100, false, useGpu); - } -} - -TEST(Layer, sumCostLayer) { - TestConfig config; - config.layerConfig.set_type("sum_cost"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "sum_cost", 100, false, useGpu); - } -} - -TEST(Layer, weightedRankCostLayer) { - TestConfig config; - config.layerConfig.set_type("rank-cost"); - config.biasSize = 0; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 1, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_2", 1, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_3", 1, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "weighted-rank-cost", 100, false, useGpu); - } -} - -TEST(Layer, TensorLayer) { - TestConfig config; - config.layerConfig.set_type("tensor"); - config.layerConfig.set_size(10); - config.layerConfig.set_active_type("sigmoid"); - config.biasSize = config.layerConfig.size(); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 5, 250}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", 5, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "tensor", 100, false, useGpu); - } -} - -TEST(Layer, RecurrentLayer) { - TestConfig config; - config.layerConfig.set_type("recurrent"); - config.layerConfig.set_size(4); - config.layerConfig.set_active_type("tanh"); - config.biasSize = 4; - - config.inputDefs.push_back( - {INPUT_SEQUENCE_DATA, "layer_0", /* dim= */ 4, /* paraSize= */ 16}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - for (auto reversed : {false, true}) { - config.layerConfig.set_reversed(reversed); - config.testState = !reversed; - testLayerGrad( - config, "recurrent", 50, /* trans= */ false, useGpu, false, 1.0); - } - } -} - -TEST(Layer, LstmLayer) { - TestConfig config; - config.layerConfig.set_type("lstmemory"); - config.layerConfig.set_size(4); - config.layerConfig.set_active_type("tanh"); - config.layerConfig.set_active_state_type("sigmoid"); - config.layerConfig.set_active_gate_type("sigmoid"); - config.biasSize = 28; - - config.inputDefs.push_back( - {INPUT_SEQUENCE_DATA, "layer_0", /* dim= */ 16, /* paraSize= */ 64}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - for (auto reversed : {false, true}) { - config.layerConfig.set_reversed(reversed); - config.testState = !reversed; - testLayerGrad( - config, "lstmemory", 100, /* trans= */ false, useGpu, false, 0.02); - } - } - for (auto useGpu : {true}) { - config.testBatchState = true; - config.layerConfig.set_reversed(false); - testLayerGrad(config, "lstmemory", 10, /* trans= */ false, useGpu); - } -} - -TEST(Layer, MDLstmLayer) { - TestConfig config; - config.layerConfig.set_type("mdlstmemory"); - config.layerConfig.set_size(4); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_active_state_type("sigmoid"); - config.layerConfig.set_active_gate_type("sigmoid"); - config.biasSize = 4 * 9; - - config.inputDefs.push_back( - {INPUT_SEQUENCE_MDIM_DATA, "layer_0", 4 * 5, 4 * 4 * 5}); - config.layerConfig.add_inputs(); - config.layerConfig.add_directions(true); - config.layerConfig.add_directions(true); - - for (auto useGpu : {false, true}) { - for (int i = 0; i < 2; i++) { - for (int j = 0; j < 2; j++) { - config.layerConfig.set_directions(0, bool(i)); - config.layerConfig.set_directions(1, bool(j)); - testLayerGrad(config, "mdlstmemory", 100, false, useGpu); - } - } - } -} - -TEST(Layer, ParameterReluLayer) { - auto testParameterReluLayer = [&](size_t inputSize, size_t channels) { - TestConfig config; - config.layerConfig.set_type("prelu"); - config.inputDefs.push_back({INPUT_DATA, "layer_0", inputSize, channels}); - config.layerConfig.add_inputs(); - config.layerConfig.set_size(inputSize); - config.layerConfig.set_partial_sum(inputSize / - channels); // size of feature map - for (auto useGpu : {false, true}) { - testLayerGrad(config, "prelu", 100, false, useGpu); - } - }; - - testParameterReluLayer(192, 1); - testParameterReluLayer(192, 3); - testParameterReluLayer(192, 192); -} - -TEST(Layer, ResizeLayer) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("resize"); - config.layerConfig.set_size(64); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 16, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "resize", 100, false, useGpu); - } -} - -TEST(Layer, RotateLayer) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("rotate"); - const int CHANNEL = 2; - const int HEIGHT = 8; - const int WIDTH = 4; - const int INPUT_SIZE = HEIGHT * WIDTH * CHANNEL; - config.layerConfig.set_size(INPUT_SIZE); - config.layerConfig.set_height(HEIGHT); - config.layerConfig.set_width(WIDTH); - config.inputDefs.push_back({INPUT_DATA, "layer_0", INPUT_SIZE, 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "rotate", 100, false, useGpu); - } -} - -TEST(Layer, NCELayer) { - TestConfig config; - size_t numClasses = 4; - config.layerConfig.set_type("nce"); - config.layerConfig.set_size(1); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_num_classes(numClasses); - config.biasSize = numClasses; - - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", /* dim= */ 16, /* paraSize= */ 16 * numClasses}); - config.inputDefs.push_back( - {INPUT_LABEL, "label", /* dim= */ numClasses, /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto withWeight : {false, true}) { - if (withWeight) { - config.inputDefs.push_back( - {INPUT_DATA_TARGET, "weight", /* dim= */ 1, /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - } - - for (auto isIdLabel : {false, true}) { - config.inputDefs[1] = { - isIdLabel ? INPUT_LABEL : INPUT_SPARSE_NON_VALUE_DATA, - "label", - /* dim= */ numClasses, - /* paraSize= */ 0}; - - for (auto withDist : {false, true}) { - config.layerConfig.clear_neg_sampling_dist(); - if (withDist) { - double sum = 0; - for (size_t i = 0; i < numClasses; ++i) { - real p = rand(); // NOLINT use rand_r - config.layerConfig.add_neg_sampling_dist(p); - sum += p; - } - for (size_t i = 0; i < numClasses; ++i) { - real p = config.layerConfig.neg_sampling_dist(i) / sum; - config.layerConfig.set_neg_sampling_dist(i, p); - } - } - LOG(INFO) << "NCELayer " - << " isIdLabel=" << isIdLabel << " withWeight=" << withWeight - << " withDist=" << withDist; - // Not support GPU now - testLayerGrad(config, - "nce", - 100, - /* trans= */ false, - /* useGpu */ false); - } - } - } -} - -TEST(Layer, GatedRecurrentLayer) { - TestConfig config; - config.layerConfig.set_type("gated_recurrent"); - config.layerConfig.set_size(4); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_active_gate_type("sigmoid"); - config.biasSize = 12; - - config.inputDefs.push_back( - {INPUT_SEQUENCE_DATA, "layer_0", /* dim= */ 12, /* paraSize= */ 48}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - for (auto reversed : {false, true}) { - config.layerConfig.set_reversed(reversed); - config.testState = !reversed; - testLayerGrad(config, "gated_recurrent", 100, /* trans= */ false, useGpu); - } - } -} - -TEST(Layer, GruStepLayer) { - TestConfig config; - config.layerConfig.set_type("gru_step"); - config.layerConfig.set_size(4); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_active_gate_type("sigmoid"); - config.biasSize = 12; - - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", /* dim= */ 12, /* paraSize= */ 48}); - config.inputDefs.push_back( - {INPUT_DATA, "layer_1", /* dim= */ 4, /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "gruStep", 100, /* trans= */ false, useGpu); - } -} - -TEST(Layer, LstmStepLayer) { - TestConfig config; - config.layerConfig.set_type("lstm_step"); - config.layerConfig.set_size(4); - config.layerConfig.set_active_type("sigmoid"); - config.layerConfig.set_active_state_type("sigmoid"); - config.layerConfig.set_active_gate_type("sigmoid"); - config.biasSize = 12; - config.testAccumulate = false; - - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", /* dim= */ 16, /* paraSize= */ 0}); - config.inputDefs.push_back( - {INPUT_DATA, "layer_1", /* dim= */ 4, /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "lstmStep", 100, /* trans= */ false, useGpu); - } -} - -void testBatchNormLayer(const string& type, bool trans, bool useGpu) { - TestConfig config; - const int CHANNELS = 10; - const int IMG_SIZE = 16; - const int IMG_SIZE_Y = 8; - size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y; - config.layerConfig.set_type(type); - config.layerConfig.set_size(size); - config.layerConfig.set_active_type("sigmoid"); - config.biasSize = CHANNELS; - config.inputDefs.push_back({INPUT_DATA, - "layer_0", - /* dim= */ size, - /* paraSize= */ CHANNELS}); - - config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); - config.inputDefs.back().isStatic = true; - config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); - config.inputDefs.back().isStatic = true; - - LayerInputConfig* input = config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - ImageConfig* img_conf = input->mutable_image_conf(); - img_conf->set_channels(CHANNELS); - img_conf->set_img_size(IMG_SIZE); - img_conf->set_img_size_y(IMG_SIZE_Y); - - testLayerGrad(config, - "batch_norm", - 64, - /* trans= */ trans, - useGpu, - /* useWeight */ true); -} - -TEST(Layer, BatchNormalizationLayer) { - testBatchNormLayer("batch_norm", false, false); -#ifdef PADDLE_WITH_CUDA - testBatchNormLayer("batch_norm", false, true); - if (hl_get_cudnn_lib_version() >= int(4000)) { - testBatchNormLayer("cudnn_batch_norm", false, true); - } -#endif -} - -void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) { - TestConfig config; - const int CHANNELS = 10; - const int IMG_SIZE = 16; - const int IMG_SIZE_Y = 8; - const int IMG_SIZE_Z = 8; - size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y * IMG_SIZE_Z; - config.layerConfig.set_type(type); - config.layerConfig.set_size(size); - config.layerConfig.set_active_type("sigmoid"); - config.biasSize = CHANNELS; - config.inputDefs.push_back({INPUT_DATA, - "layer_0", - /* dim= */ size, - /* paraSize= */ CHANNELS}); - - config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); - config.inputDefs.back().isStatic = true; - config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); - config.inputDefs.back().isStatic = true; - - LayerInputConfig* input = config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - ImageConfig* img_conf = input->mutable_image_conf(); - img_conf->set_channels(CHANNELS); - img_conf->set_img_size(IMG_SIZE); - img_conf->set_img_size_y(IMG_SIZE_Y); - img_conf->set_img_size_z(IMG_SIZE_Z); - - testLayerGrad(config, - "batch_norm", - 64, - /* trans= */ trans, - useGpu, - /* useWeight */ true); -} - -TEST(Layer, testBatchNorm3DLayer) { - testBatchNorm3DLayer("batch_norm", false, false); -#ifdef PADDLE_WITH_CUDA - testBatchNorm3DLayer("batch_norm", false, true); - if (hl_get_cudnn_lib_version() >= int(4000)) { - testBatchNorm3DLayer("cudnn_batch_norm", false, true); - } -#endif -} - -void testConvOperator(bool isDeconv) { - TestConfig config; - const int NUM_FILTERS = 16; - const int FILTER_SIZE = 2; - const int FILTER_SIZE_Y = 3; - const int CHANNELS = 3; - const int IMAGE_SIZE = 16; - const int IMAGE_SIZE_Y = 9; - OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs(); - if (isDeconv) { - operatorConf.set_type("convt"); - } else { - operatorConf.set_type("conv"); - } - ConvConfig* conv = operatorConf.mutable_conv_conf(); - operatorConf.set_num_filters(NUM_FILTERS); - conv->set_filter_size(FILTER_SIZE); - conv->set_filter_size_y(FILTER_SIZE_Y); - conv->set_channels(CHANNELS); - conv->set_padding(0); - conv->set_padding_y(1); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_groups(1); - conv->set_img_size(IMAGE_SIZE); - conv->set_img_size_y(IMAGE_SIZE_Y); - conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - conv->set_output_y(outputSize(conv->img_size_y(), - conv->filter_size_y(), - conv->padding_y(), - conv->stride_y(), - /* caffeMode */ true)); - - if (isDeconv) { - conv->set_filter_channels(NUM_FILTERS / conv->groups()); - config.inputDefs.push_back({INPUT_DATA, - "layer_0", - conv->output_x() * conv->output_y() * CHANNELS, - 0}); - config.layerConfig.set_size(IMAGE_SIZE * IMAGE_SIZE_Y * NUM_FILTERS); - } else { - conv->set_filter_channels(conv->channels() / conv->groups()); - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE_Y * CHANNELS, 0}); - config.layerConfig.set_size(conv->output_x() * conv->output_y() * - NUM_FILTERS); - } - - config.inputDefs.push_back( - {INPUT_DATA, - "layer_1", - FILTER_SIZE * FILTER_SIZE_Y * CHANNELS * NUM_FILTERS, - 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - testOperatorGrad(config, operatorConf, 100, /*useGpu*/ true, false); -} - -TEST(Operator, conv) { - testConvOperator(/*isDeconv*/ true); - testConvOperator(/*isDeconv*/ false); -} - -TEST(Layer, FeatureMapExpandLayer) { - TestConfig config; - config.layerConfig.set_type("featmap_expand"); - const int CHANNELS = 10; - const int INPUT_SIZE = 100; - config.layerConfig.set_size(INPUT_SIZE * CHANNELS); - config.layerConfig.set_num_filters(CHANNELS); - config.inputDefs.push_back({INPUT_SEQUENCE_DATA, - "layer_0", - /* dim= */ INPUT_SIZE, - /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - for (auto useGpu : {false, true}) { - for (auto asRowVec : {false, true}) { - config.layerConfig.set_user_arg(asRowVec ? "as_row_vec" : "as_col_vec"); - testLayerGrad(config, - "featmap_expand", - /*batch_size*/ 100, - /* trans= */ false, - useGpu, - /* useWeight */ true); - } - } -} - -TEST(Layer, MultiplexLayer) { - TestConfig config; - const int LAYER_SIZE = 100; - config.layerConfig.set_type("multiplex"); - config.layerConfig.set_size(LAYER_SIZE); - - config.inputDefs.push_back({INPUT_LABEL, "layer_0", 2, 0}); - config.inputDefs.push_back( - {INPUT_DATA, "layer_1", /* dim= */ LAYER_SIZE, /* paraSize= */ 0}); - config.inputDefs.push_back( - {INPUT_DATA, "layer_2", /* dim= */ LAYER_SIZE, /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "multiplex", 512, /* trans= */ false, useGpu); - } -} - -TEST(Layer, PadLayer) { - TestConfig config; - config.biasSize = 0; - config.layerConfig.set_type("pad"); - - int c = 4; - int h = 31; - int w = 36; - size_t size = c * h * w; - config.inputDefs.push_back({INPUT_DATA, "layer_0", size, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - PadConfig* pad = input->mutable_pad_conf(); - ImageConfig* image = pad->mutable_image_conf(); - - image->set_channels(c); - image->set_img_size(h); - image->set_img_size_y(w); - pad->add_pad_c(1); - pad->add_pad_c(2); - pad->add_pad_h(2); - pad->add_pad_h(3); - pad->add_pad_w(3); - pad->add_pad_w(5); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "pad", 10, false, useGpu); - } -} - -TEST(Layer, CrossChannelNormLayer) { - TestConfig config; - config.paramInitialMean = 1.; - config.paramInitialStd = 0.; - config.layerConfig.set_type("norm"); - config.layerConfig.set_size(100); - LayerInputConfig* input = config.layerConfig.add_inputs(); - NormConfig* norm = input->mutable_norm_conf(); - norm->set_norm_type("cross-channel-norm"); - norm->set_channels(10); - norm->set_size(100); - norm->set_scale(0); - norm->set_pow(0); - norm->set_blocked(0); - config.inputDefs.push_back({INPUT_DATA, "layer_0", 100, 10}); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false); - } -} - -TEST(Layer, smooth_l1) { - TestConfig config; - config.layerConfig.set_type("smooth_l1"); - - config.inputDefs.push_back({INPUT_DATA, "layer_0", 200, 0}); - config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 200, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "smooth_l1", 100, false, useGpu, false); - } -} - -TEST(Layer, multibox_loss) { - TestConfig config; - config.layerConfig.set_type("multibox_loss"); - config.biasSize = 0; - LayerInputConfig* input = config.layerConfig.add_inputs(); - MultiBoxLossConfig* multiboxLoss = input->mutable_multibox_loss_conf(); - multiboxLoss->set_num_classes(21); - multiboxLoss->set_input_num(1); - multiboxLoss->set_overlap_threshold(0.5); - multiboxLoss->set_neg_pos_ratio(3); - multiboxLoss->set_neg_overlap(0.5); - multiboxLoss->set_background_id(0); - multiboxLoss->set_height(3); - multiboxLoss->set_width(3); - - size_t gtNum = 1; - MatrixPtr labelValue = Matrix::create(gtNum, 6, false, false); - labelValue->randomizeUniform(); - labelValue->add(-0.5); - labelValue->sigmoid(*labelValue); - real* labelData = labelValue->getData(); - size_t labelWidth = labelValue->getWidth(); - for (size_t i = 0; i < gtNum; ++i) { - *(labelData + i * labelWidth) = std::rand() % 20 + 1; - *(labelData + i * labelWidth + 1) = 0.400259; - *(labelData + i * labelWidth + 2) = 0.377857; - *(labelData + i * labelWidth + 3) = 0.525712; - *(labelData + i * labelWidth + 4) = 0.519368; - } - vector seqStartPositions(gtNum + 1, 0); - for (size_t i = 1; i <= gtNum; ++i) { - seqStartPositions[i] = i; - } - - // Ensure at lease one matched bbox - MatrixPtr priorValue = Matrix::create(1, 72, false, false); - priorValue->randomizeUniform(); - priorValue->add(-0.5); - priorValue->sigmoid(*priorValue); - real* priorData = priorValue->getData(); - *(priorData) = 0.424811; - *(priorData + 1) = 0.397059; - *(priorData + 2) = 0.538905; - *(priorData + 3) = 0.447091; - *(priorData + 4) = 0.425720; - *(priorData + 5) = 0.515228; - *(priorData + 6) = 0.519452; - *(priorData + 7) = 0.591065; - - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, "priorbox", priorValue, {}}); - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, "label", labelValue, seqStartPositions}); - config.inputDefs.push_back({INPUT_DATA, "locPred", 36, 0}); - config.inputDefs.push_back({INPUT_DATA, "confPred", 189, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "multibox_loss", 1, false, useGpu, false); - } -} - -TEST(Layer, TransLayer) { - TestConfig config; - const int height = 128; - const int width = 256; - config.layerConfig.set_type("trans"); - config.layerConfig.set_size(width); - - config.inputDefs.push_back( - {INPUT_DATA, "layer_0", /* dim= */ height * width, /* paraSize= */ 0}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "trans", height, /* trans= */ false, useGpu); - } -} - -TEST(Layer, RowConvLayer) { - const int context = 3; - const int size = 512; - - TestConfig config; - config.layerConfig.set_type("row_conv"); - config.layerConfig.set_size(size); - config.layerConfig.set_active_type("sigmoid"); - - config.inputDefs.push_back( - {INPUT_SEQUENCE_DATA, "layer_0", size, context * size}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - RowConvConfig* conv = input->mutable_row_conv_conf(); - conv->set_context_length(context); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "row_conv", 100, false, useGpu, false); - } -} - -TEST(Layer, CropLayer) { - TestConfig config; - // config input_0 - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ImageConfig* img = input->mutable_image_conf(); - img->set_channels(4); - img->set_img_size(16); - config.layerConfig.set_axis(2); - config.layerConfig.add_offset(0); - config.layerConfig.add_offset(0); - - // config input_1 - config.inputDefs.push_back({INPUT_DATA, "layer_1", 128, 0}); - input = config.layerConfig.add_inputs(); - img = input->mutable_image_conf(); - img->set_channels(2); - img->set_img_size(8); - - // config crop layer - config.layerConfig.set_type("crop"); - config.layerConfig.set_name("cropLayer"); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "crop", 100, false, useGpu, false); - } -} - -TEST(Layer, roi_pool) { - TestConfig config; - config.layerConfig.set_type("roi_pool"); - config.biasSize = 0; - LayerInputConfig* input = config.layerConfig.add_inputs(); - ROIPoolConfig* roiPoolConf = input->mutable_roi_pool_conf(); - roiPoolConf->set_pooled_width(7); - roiPoolConf->set_pooled_height(7); - roiPoolConf->set_spatial_scale(1. / 16); - roiPoolConf->set_width(14); - roiPoolConf->set_height(14); - - const size_t roiNum = 10; - const size_t roiDim = 10; - const size_t batchSize = 5; - MatrixPtr roiValue = Matrix::create(roiNum, roiDim, false, false); - roiValue->zeroMem(); - real* roiData = roiValue->getData(); - for (size_t i = 0; i < roiNum; ++i) { - roiData[i * roiDim + 0] = std::rand() % batchSize; - roiData[i * roiDim + 1] = std::rand() % 224; // xMin - roiData[i * roiDim + 2] = std::rand() % 224; // yMin - size_t xMin = static_cast(roiData[i * roiDim + 1]); - size_t yMin = static_cast(roiData[i * roiDim + 2]); - roiData[i * roiDim + 3] = xMin + std::rand() % (224 - xMin); // xMax - roiData[i * roiDim + 4] = yMin + std::rand() % (224 - yMin); // yMax - } - - config.inputDefs.push_back({INPUT_DATA, "input", 3 * 14 * 14, {}}); - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "rois", roiValue, {}}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "roi_pool", batchSize, false, useGpu, false); - } -} - -TEST(Layer, SwitchOrderLayer) { - TestConfig config; - // config input_0 - config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ImageConfig* img = input->mutable_image_conf(); - img->set_channels(4); - img->set_img_size(16); - img->set_img_size_y(16); - - ReshapeConfig* reshape = config.layerConfig.mutable_reshape_conf(); - reshape->add_height_axis(0); - reshape->add_height_axis(1); - reshape->add_height_axis(2); - reshape->add_width_axis(3); - - // config softmax layer - config.layerConfig.set_type("switch_order"); - config.layerConfig.set_name("switchOrderLayer"); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "switch_order", 100, false, useGpu, true); - } -} - -vector randSampling(real range, int n) { - CHECK_GE(range, n); - vector num(range); - iota(begin(num), end(num), 0.); - if (range == n) return num; - - random_shuffle(begin(num), end(num)); - num.resize(n); - sort(begin(num), end(num)); - return num; -} - -TEST(Layer, SubNestedSequenceLayer) { - // layer size is not crutial for this layer, - // so use a small layer size in unittest - const int layerSize = 4; - - const int maxSeqNum = 50; - const int maxSeqLen = 50; - const int maxBeamSize = 32; - - srand((size_t)(time(NULL))); - int beamSize = 1 + (rand() % maxBeamSize); - - TestConfig config; - config.layerConfig.set_type("sub_nested_seq"); - config.layerConfig.set_name("sub_nested_seq_layer"); - config.layerConfig.set_size(layerSize); - - int seqNum = 1 + (rand() % maxSeqNum); - - // sequence information for the first input, it is a nested sequence - vector seqStartPos(seqNum + 1, 0); - vector subSeqStartPos(1, 0); - - // selected indices - MatrixPtr selectedIndices = Matrix::create(seqNum, beamSize, false, false); - selectedIndices->one(); - selectedIndices->mulScalar(-1.); - real* indicesData = selectedIndices->getData(); - - for (int i = 0; i < seqNum; ++i) { - int subSeqNum = 1 + (rand() % maxSeqNum); - for (int j = 0; j < subSeqNum; ++j) { - subSeqStartPos.push_back(subSeqStartPos.back() + - (1 + (rand() % maxSeqLen))); - } - vector selSeqs = - randSampling(static_cast(subSeqNum), min(beamSize, subSeqNum)); - memcpy(indicesData + (i * beamSize), - selSeqs.data(), - selSeqs.size() * sizeof(real)); - seqStartPos[i + 1] = subSeqStartPos.back(); - } - - MatrixPtr seqInputPtr = - Matrix::create(seqStartPos.back(), layerSize, false, false); - seqInputPtr->randomizeUniform(); - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, - "nested_seq_input", - seqInputPtr, - seqStartPos, - subSeqStartPos}); - config.layerConfig.add_inputs(); - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, "selected_indices", selectedIndices}); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, - "sub_nested_seq", - /* batchSize */ seqNum, - /* trans */ false, - /* useGpu*/ useGpu, - /* useWeight */ false); - } -} - -TEST(Layer, ClipLayer) { - const size_t batchSize = 128; - const size_t size = 512; - TestConfig config; - config.layerConfig.set_type("clip"); - config.inputDefs.push_back({INPUT_DATA, "input", size, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ClipConfig* layerConf = input->mutable_clip_conf(); - double p1 = std::rand() / (double)RAND_MAX; - double p2 = std::rand() / (double)RAND_MAX; - layerConf->set_min(std::min(p1, p2)); - layerConf->set_max(std::max(p1, p2)); - for (auto useGpu : {false, true}) { - testLayerGrad(config, "clip", batchSize, false, useGpu, false); - } -} - -TEST(Layer, RowL2NormLayer) { - const size_t batchSize = 128; - const size_t size = 512; - TestConfig config; - config.layerConfig.set_type("row_l2_norm"); - config.layerConfig.set_size(size); - config.inputDefs.push_back({INPUT_DATA, "input", size, 0}); - config.layerConfig.add_inputs(); - for (auto useGpu : {false, true}) { - testLayerGrad(config, "row_l2_norm", batchSize, false, useGpu, false); - } -} - -void test3DConvLayer(const string& type, bool trans, bool useGpu) { - // filter size - const int NUM_FILTERS = 6; - // const int CHANNELS = 3; - const int FILTER_SIZE = 3; - const int FILTER_SIZE_Y = 3; - const int FILTER_SIZE_Z = 3; - - // input image - const int CHANNELS = 3; - const int IMAGE_SIZE = 9; - const int IMAGE_SIZE_Y = 9; - const int IMAGE_SIZE_Z = 9; - - TestConfig config; - config.biasSize = NUM_FILTERS; - config.layerConfig.set_type(type); - config.layerConfig.set_num_filters(NUM_FILTERS); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - // Setting up conv3D-trans layer - LayerInputConfig* input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - - conv->set_channels(CHANNELS); - conv->set_filter_size(FILTER_SIZE); - conv->set_filter_size_y(FILTER_SIZE_Y); - conv->set_filter_size_z(FILTER_SIZE_Z); - conv->set_padding(0); - conv->set_padding_y(0); - conv->set_padding_z(0); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_stride_z(2); - conv->set_img_size(IMAGE_SIZE); - conv->set_img_size_y(IMAGE_SIZE_Y); - conv->set_img_size_z(IMAGE_SIZE_Z); - conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), - conv->padding(), - conv->stride(), - /* caffeMode */ true)); - conv->set_output_y(outputSize(conv->img_size_y(), - conv->filter_size_y(), - conv->padding_y(), - conv->stride_y(), - /* caffeMode */ true)); - conv->set_output_z(outputSize(conv->img_size_z(), - conv->filter_size_z(), - conv->padding_z(), - conv->stride_z(), - /* caffeMode */ true)); - - config.layerConfig.set_size(conv->output_x() * conv->output_y() * - conv->output_z() * NUM_FILTERS); - conv->set_groups(1); - conv->set_filter_channels(conv->channels() / conv->groups()); - config.inputDefs.push_back( - {INPUT_DATA, - "layer_0", - CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z, - conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z * - NUM_FILTERS}); - - testLayerGrad(config, "conv3D", 10, trans, useGpu); - // Use small batch_size and useWeight=true to test biasGrad - testLayerGrad(config, "conv3D", 2, trans, useGpu, true, 0.02); -} - -TEST(Layer, test3DConvLayer) { - test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false); -#ifdef PADDLE_WITH_CUDA - test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true); -#endif -} - -void test3DDeConvLayer(const string& type, bool trans, bool useGpu) { - // filter size - const int NUM_FILTERS = 6; - // const int CHANNELS = 3; - const int FILTER_SIZE = 3; - const int FILTER_SIZE_Y = 3; - const int FILTER_SIZE_Z = 3; - - // input image - const int CHANNELS = 3; - const int IMAGE_SIZE = 4; - const int IMAGE_SIZE_Y = 6; - const int IMAGE_SIZE_Z = 6; - - // Setting up conv-trans layer - TestConfig config; - config.biasSize = NUM_FILTERS; - config.layerConfig.set_type("deconv3d"); - config.layerConfig.set_num_filters(NUM_FILTERS); - config.layerConfig.set_partial_sum(1); - config.layerConfig.set_shared_biases(true); - - LayerInputConfig* input = config.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - - conv->set_channels(CHANNELS); - conv->set_filter_size(FILTER_SIZE); - conv->set_filter_size_y(FILTER_SIZE_Y); - conv->set_filter_size_z(FILTER_SIZE_Z); - conv->set_padding(0); - conv->set_padding_y(0); - conv->set_padding_z(0); - conv->set_stride(2); - conv->set_stride_y(2); - conv->set_stride_z(2); - conv->set_output_x(IMAGE_SIZE); - conv->set_output_y(IMAGE_SIZE_Y); - conv->set_output_z(IMAGE_SIZE_Z); - - conv->set_img_size(imageSize(conv->output_x(), - conv->filter_size(), - conv->padding(), - conv->stride(), - true)); - conv->set_img_size_y(imageSize(conv->output_y(), - conv->filter_size_y(), - conv->padding_y(), - conv->stride_y(), - true)); - conv->set_img_size_z(imageSize(conv->output_z(), - conv->filter_size_z(), - conv->padding_z(), - conv->stride_z(), - true)); - config.layerConfig.set_size(conv->img_size() * conv->img_size_y() * - conv->img_size_z() * NUM_FILTERS); - conv->set_groups(1); - conv->set_filter_channels(conv->channels() / conv->groups()); - config.inputDefs.push_back( - {INPUT_DATA, - "layer_0", - CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z, - conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z * - NUM_FILTERS}); - - testLayerGrad(config, "deconv3D", 10, trans, useGpu); - // Use small batch_size and useWeight=true to test biasGrad - testLayerGrad(config, "deconv3D", 2, trans, useGpu, true, 0.02); -} - -TEST(Layer, test3DDeConvLayer) { - test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false); -#ifdef PADDLE_WITH_CUDA - test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true); -#endif -} - -TEST(Layer, ScaleShiftLayer) { - // FIXME: Disable ScaleShiftLayer because it is not stable. - // https://github.com/PaddlePaddle/Paddle/issues/7781 - return; - // const size_t batchSize = 16; - // const size_t size = 32; - // TestConfig config; - // config.layerConfig.set_type("scale_shift"); - // config.layerConfig.set_size(size); - // config.biasSize = 1; - // config.inputDefs.push_back( - // {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1}); - // config.layerConfig.add_inputs(); - // for (auto useGpu : {false, true}) { - // testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false); - // } -} - -TEST(Layer, ScaleSubRegionLayer) { - const size_t batchSize = 64; - const size_t size = 4096; - TestConfig config; - config.layerConfig.set_type("scale_sub_region"); - config.inputDefs.push_back({INPUT_DATA, "input", size, 0}); - MatrixPtr indicesV = Matrix::create(batchSize, 6, false, false); - auto* data = indicesV->getData(); - for (size_t i = 0; i < batchSize; ++i) { - data[i * 2] = 2; - data[i * 2 + 1] = 4; - data[i * 2 + 2] = 16; - data[i * 2 + 3] = 32; - data[i * 2 + 4] = 16; - data[i * 2 + 5] = 32; - } - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "indices", indicesV, {}}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - ScaleSubRegionConfig* scaleSubRegionConf = - input->mutable_scale_sub_region_conf(); - ImageConfig* imgConf = scaleSubRegionConf->mutable_image_conf(); - imgConf->set_img_size(32); - imgConf->set_img_size_y(32); - imgConf->set_channels(4); - scaleSubRegionConf->set_value(2.0); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "scale_sub_region", batchSize, false, useGpu, false); - } -} - -TEST(Layer, L2DistanceLayer) { - TestConfig config; - config.layerConfig.set_type("l2_distance"); - config.layerConfig.set_size(1); - config.biasSize = 0; - - const size_t input_dim = 27; - const size_t batch_size = 11; - - config.inputDefs.push_back({INPUT_DATA, "layer_0", input_dim, 0}); - config.inputDefs.push_back({INPUT_DATA, "layer_1", input_dim, 0}); - config.layerConfig.add_inputs(); - config.layerConfig.add_inputs(); - - for (auto useGpu : {false, true}) { - testLayerGrad(config, "l2_distance", batch_size, false, useGpu); - } -} - -void testFactorizationMachineLayer(InputType type, bool useGpu) { - const int FACTOR_SIZE = 10; - TestConfig config; - config.layerConfig.set_type("factorization_machine"); - config.layerConfig.set_factor_size(FACTOR_SIZE); - config.layerConfig.set_size(1); - config.biasSize = 0; - config.inputDefs.push_back({type, "layer_0", 128, 1280}); - config.layerConfig.add_inputs(); - testLayerGrad(config, "factorization_machine", 16, false, useGpu, false); -} - -TEST(Layer, FactorizationMachineLayer) { - for (auto useGpu : {false, true}) { - testFactorizationMachineLayer(INPUT_DATA, useGpu); - } - testFactorizationMachineLayer(INPUT_SPARSE_FLOAT_VALUE_DATA, false); -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_LinearChainCRF.cpp b/paddle/gserver/tests/test_LinearChainCRF.cpp deleted file mode 100644 index 423c31e27d7ca223f1cbff8f030b006d3889f0bb..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_LinearChainCRF.cpp +++ /dev/null @@ -1,67 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "paddle/gserver/layers/LinearChainCRF.h" -#include "paddle/utils/Util.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static inline bool getNextSequence(vector& seq, int numClasses) { - for (auto& v : seq) { - if (++v < numClasses) { - return true; - } - v = 0; - } - return false; -} - -TEST(LinearChainCRF, decoding) { - const int numClasses = 4; - CpuVector para(numClasses * (numClasses + 2)); - real* a = para.getData(); - real* b = para.getData() + numClasses; - real* w = para.getData() + 2 * numClasses; - LinearChainCRF crf(4, para.getData()); - for (int length : {1, 2, 3, 10}) { - for (int tries = 0; tries < 10; ++tries) { - CpuMatrix x(length, numClasses); - x.randomizeUniform(); - para.randnorm(0, 2); - vector decodingResult(length); - vector bestResult(length); - vector testResult(length, 0); - crf.decode(x.getData(), &decodingResult[0], length); - real bestScore = -std::numeric_limits::max(); - do { - real score = a[testResult.front()] + b[testResult.back()]; - score += x.getElement(0, testResult.front()); - for (int k = 1; k < length; ++k) { - score += x.getElement(k, testResult[k]) + - w[numClasses * testResult[k - 1] + testResult[k]]; - } - if (score > bestScore) { - bestScore = score; - bestResult = testResult; - } - } while (getNextSequence(testResult, numClasses)); - for (int k = 0; k < length; ++k) { - EXPECT_EQ(decodingResult[k], bestResult[k]); - } - } - } -} diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp deleted file mode 100644 index a34a3f6206171fb1e0563ab9ef8550bc890359ce..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_MKLDNN.cpp +++ /dev/null @@ -1,448 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include "MKLDNNTester.h" -#include "ModelConfig.pb.h" -#include "paddle/gserver/activations/MKLDNNActivation.h" -#include "paddle/math/MathUtils.h" - -using namespace paddle; // NOLINT - -DECLARE_bool(thread_local_rand_use_global_seed); -DECLARE_bool(use_gpu); -DECLARE_bool(use_mkldnn); - -#define RUN_MKLDNN_TEST(DNN_CONFIG, REF_CONFIG, DESC) \ - MKLDNNTester tester; \ - for (auto bs : {DESC.bs, 1}) { \ - tester.run(DNN_CONFIG, REF_CONFIG, bs, DESC.ih, DESC.iw); \ - } - -#define RUN_MKLDNN_TEST_LAYER(DNN_CONFIG, REF_TYPE, DESC) \ - TestConfig ref = DNN_CONFIG; \ - ref.layerConfig.set_type(REF_TYPE); \ - RUN_MKLDNN_TEST(DNN_CONFIG, ref, DESC) - -struct testFcDesc { - int bs; - int ic; - int ih, iw; // oh == ow == 1 - int oc; -}; - -static void getMKLDNNFcConfig(TestConfig& cfg, const testFcDesc& pm) { - cfg.layerConfig.set_type("mkldnn_fc"); - cfg.layerConfig.set_active_type("relu"); - cfg.layerConfig.set_size(pm.oc); - cfg.inputDefs.push_back( - {INPUT_DATA, - "layer_0", - /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), - /* size of weight= */ size_t(pm.oc * pm.ic * pm.ih * pm.iw)}); - cfg.layerConfig.add_inputs(); -} - -void testFcLayer(const testFcDesc& pm) { - TestConfig dnnConfig; - getMKLDNNFcConfig(dnnConfig, pm); - for (auto biasSize : {pm.oc, 0}) { - dnnConfig.biasSize = biasSize; - RUN_MKLDNN_TEST_LAYER(dnnConfig, "fc", pm) - } -} - -TEST(MKLDNNLayer, FcLayer) { - /* bs, ic, ih, iw, oc */ - testFcLayer({2, 2, 1, 1, 3}); - testFcLayer({3, 7, 1, 1, 19}); - testFcLayer({8, 16, 13, 13, 32}); - testFcLayer({4, 12, 13, 13, 18}); - testFcLayer({2, 64, 16, 16, 32}); - testFcLayer({15, 3, 16, 16, 6}); -} - -struct testConvDesc { - int bs, gp; - int ic, ih, iw; - int oc, oh, ow; - int fh, fw; - int ph, pw; - int sh, sw; - int dh, dw; -}; - -static void getMKLDNNConvConfig(TestConfig& cfg, const testConvDesc& pm) { - cfg.layerConfig.set_type("mkldnn_conv"); - cfg.layerConfig.set_active_type("relu"); - cfg.layerConfig.set_num_filters(pm.oc); - cfg.layerConfig.set_size(pm.oc * pm.oh * pm.ow); - cfg.layerConfig.set_shared_biases(true); - cfg.inputDefs.push_back( - {INPUT_DATA, - "layer_0", - /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), - /* size of weight= */ size_t(pm.oc * pm.ic * pm.fh * pm.fw / pm.gp)}); - LayerInputConfig* input = cfg.layerConfig.add_inputs(); - ConvConfig* conv = input->mutable_conv_conf(); - conv->set_groups(pm.gp); - conv->set_img_size(pm.iw); - conv->set_img_size_y(pm.ih); - conv->set_output_x(pm.ow); - conv->set_output_y(pm.oh); - conv->set_filter_size(pm.fw); - conv->set_filter_size_y(pm.fh); - conv->set_channels(pm.ic); - conv->set_padding(pm.pw); - conv->set_padding_y(pm.ph); - conv->set_stride(pm.sw); - conv->set_stride_y(pm.sh); - conv->set_dilation(pm.dw); - conv->set_dilation_y(pm.dh); - conv->set_caffe_mode(true); - conv->set_filter_channels(conv->channels() / conv->groups()); - CHECK_EQ(conv->filter_channels() * pm.gp, conv->channels()) - << "it is indivisible"; - - int fh = (pm.fh - 1) * pm.dh + 1; - int fw = (pm.fw - 1) * pm.dw + 1; - int ow = outputSize(pm.iw, fw, pm.pw, pm.sw, true); - int oh = outputSize(pm.ih, fh, pm.ph, pm.sh, true); - CHECK_EQ(ow, pm.ow) << "output size check failed"; - CHECK_EQ(oh, pm.oh) << "output size check failed"; -} - -void testConvLayer(const testConvDesc& pm) { - TestConfig dnnConfig; - getMKLDNNConvConfig(dnnConfig, pm); - for (auto biasSize : {pm.oc, 0}) { - dnnConfig.biasSize = biasSize; - RUN_MKLDNN_TEST_LAYER(dnnConfig, "exconv", pm) - } -} - -TEST(MKLDNNLayer, ConvLayer) { - /* bs, gp, ic, ih, iw, oc, oh, ow, fh, fw, ph, pw, sh, sw, dh, dw */ - testConvLayer({2, 1, 3, 32, 32, 16, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1}); - testConvLayer({2, 1, 8, 16, 16, 8, 16, 16, 3, 3, 1, 1, 1, 1, 1, 1}); - testConvLayer({3, 1, 16, 32, 32, 3, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1}); - testConvLayer({8, 1, 16, 18, 18, 32, 18, 18, 3, 3, 1, 1, 1, 1, 1, 1}); - testConvLayer({16, 1, 1, 42, 31, 32, 23, 11, 4, 5, 3, 2, 2, 3, 1, 1}); - testConvLayer({2, 1, 8, 16, 16, 8, 8, 8, 3, 3, 1, 1, 2, 2, 1, 1}); - testConvLayer({3, 1, 8, 13, 13, 8, 7, 7, 3, 3, 1, 1, 2, 2, 1, 1}); - // with groups - testConvLayer({2, 2, 4, 5, 5, 8, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1}); - testConvLayer({2, 3, 3, 5, 5, 3, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1}); - testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1}); -} - -struct testPoolDesc { - int bs, ic; // input channel and output channel are the same - int ih, iw; - int oh, ow; - int fh, fw; - int ph, pw; - int sh, sw; -}; - -static void getMKLDNNPoolConfig(TestConfig& cfg, const testPoolDesc& pm) { - cfg.layerConfig.set_type("mkldnn_pool"); - cfg.layerConfig.set_active_type("relu"); - cfg.layerConfig.set_size(pm.ic * pm.oh * pm.ow); - cfg.inputDefs.push_back( - {INPUT_DATA, - "layer_0", - /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), - 0}); - LayerInputConfig* input = cfg.layerConfig.add_inputs(); - PoolConfig* pool = input->mutable_pool_conf(); - pool->set_pool_type("avg-projection"); - pool->set_channels(pm.ic); - pool->set_img_size(pm.iw); - pool->set_img_size_y(pm.ih); - pool->set_output_x(pm.ow); - pool->set_output_y(pm.oh); - pool->set_size_x(pm.fw); - pool->set_size_y(pm.fh); - pool->set_padding(pm.pw); - pool->set_padding_y(pm.ph); - pool->set_stride(pm.sw); - pool->set_stride_y(pm.sh); - - int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false); - int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false); - CHECK_EQ(ow, pm.ow) << "output size check failed"; - CHECK_EQ(oh, pm.oh) << "output size check failed"; -} - -void testPoolLayer(const testPoolDesc& pm) { - TestConfig dnnConfig; - getMKLDNNPoolConfig(dnnConfig, pm); - LayerInputConfig* input = dnnConfig.layerConfig.mutable_inputs(0); - PoolConfig* pool = input->mutable_pool_conf(); - for (auto type : {"max-projection", "avg-projection"}) { - pool->set_pool_type(type); - RUN_MKLDNN_TEST_LAYER(dnnConfig, "pool", pm) - } -} - -TEST(MKLDNNLayer, PoolLayer) { - /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw */ - testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2}); - testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2}); - testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2}); - testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2}); - testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2}); - testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1}); - testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1}); - testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2}); -} - -struct testBatchNormDesc { - int bs; - int ic; - int ih, iw; -}; - -static void getMKLDNNBatchNormConfig(TestConfig& cfg, - const testBatchNormDesc& pm) { - cfg.layerConfig.set_size(pm.ic * pm.ih * pm.iw); - cfg.layerConfig.set_type("mkldnn_batch_norm"); - cfg.biasSize = pm.ic; - cfg.inputDefs.push_back( - {INPUT_DATA, - "layer_0", - /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), - /* size of weight= */ size_t(pm.ic)}); - cfg.inputDefs.push_back( - {INPUT_DATA, "layer_1_moving_mean", 1, size_t(pm.ic)}); - cfg.inputDefs.back().isStatic = true; - cfg.inputDefs.push_back({INPUT_DATA, "layer_2_moving_var", 1, size_t(pm.ic)}); - cfg.inputDefs.back().isStatic = true; - LayerInputConfig* input = cfg.layerConfig.add_inputs(); - cfg.layerConfig.set_active_type("relu"); - cfg.layerConfig.add_inputs(); - cfg.layerConfig.add_inputs(); - ImageConfig* img_conf = input->mutable_image_conf(); - img_conf->set_channels(pm.ic); - img_conf->set_img_size_y(pm.ih); - img_conf->set_img_size(pm.iw); -} - -void testBatchNormLayer(const testBatchNormDesc& pm) { - TestConfig dnnConfig; - getMKLDNNBatchNormConfig(dnnConfig, pm); - TestConfig refConfig = dnnConfig; - refConfig.layerConfig.set_type("batch_norm"); - // for PASS_TRAIN, use_global_stats always should be false, and batchsize != 1 - VLOG(MKLDNN_TESTS) << "check train phase"; - dnnConfig.layerConfig.set_use_global_stats(false); - refConfig.layerConfig.set_use_global_stats(false); - MKLDNNTester tester; - tester.run(dnnConfig, refConfig, pm.bs, pm.ih, pm.iw, PASS_TRAIN); - // for PASS_TEST, check use_global_stats true and false, and batchsize 1 - VLOG(MKLDNN_TESTS) << "check test phase"; - for (auto useGS : {false, true}) { - dnnConfig.layerConfig.set_use_global_stats(useGS); - refConfig.layerConfig.set_use_global_stats(useGS); - MKLDNNTester tester; - for (auto bs : {pm.bs, 1}) { - tester.run(dnnConfig, refConfig, bs, pm.ih, pm.iw, PASS_TEST); - } - } -} - -TEST(MKLDNNLayer, BatchNormLayer) { - testBatchNormLayer({4, 10, 6, 6}); - testBatchNormLayer({16, 32, 16, 16}); - testBatchNormLayer({4, 16, 8, 10}); -} - -struct testLRNDesc { - int bs, ic, ih, iw; - float scale, pow; - int localSize; -}; - -void getMKLDNNLRNConfig(TestConfig& cfg, const testLRNDesc& pm) { - cfg.layerConfig.set_type("mkldnn_lrn"); - cfg.layerConfig.set_active_type("relu"); - size_t layerSize = pm.ic * pm.ih * pm.iw; - cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0}); - LayerInputConfig* input = cfg.layerConfig.add_inputs(); - NormConfig* norm = input->mutable_norm_conf(); - norm->set_channels(pm.ic); - norm->set_size(pm.localSize); - norm->set_scale(pm.scale); - norm->set_pow(pm.pow); - norm->set_blocked(0); - norm->set_img_size(pm.iw); - norm->set_img_size_y(pm.ih); - norm->set_output_x(norm->img_size()); - norm->set_output_y(norm->img_size_y()); - cfg.layerConfig.set_size(layerSize); - cfg.biasSize = 0; -} - -void testLRNLayer(const testLRNDesc& pm) { - TestConfig dnnConfig; - getMKLDNNLRNConfig(dnnConfig, pm); - // mkldnn_lrn <==> norm with cmrnorm-projection type - TestConfig refConfig = dnnConfig; - refConfig.layerConfig.set_type("norm"); - LayerInputConfig* input = refConfig.layerConfig.mutable_inputs(0); - NormConfig* norm = input->mutable_norm_conf(); - norm->set_norm_type("cmrnorm-projection"); - norm->set_scale(norm->scale() / norm->size()); - RUN_MKLDNN_TEST(dnnConfig, refConfig, pm) -} - -TEST(MKLDNNLayer, LRNLayer) { - testLRNLayer({4, 10, 12, 12, 0.001f, 0.75f, 5}); - testLRNLayer({2, 32, 6, 6, 0.001f, 0.75f, 5}); - testLRNLayer({4, 16, 8, 10, 0.01f, 0.5f, 5}); -} - -struct testImageDesc { - int bs, ic, ih, iw; -}; - -static void getAddtoConfig(TestConfig& cfg, - const testImageDesc& pm, - const size_t nInputs = 1) { - cfg.biasSize = 0; - cfg.layerConfig.set_type("addto"); - size_t layerSize = pm.ic * pm.ih * pm.iw; - cfg.layerConfig.set_size(layerSize); - cfg.layerConfig.set_active_type("relu"); - for (size_t i = 0; i < nInputs; ++i) { - std::stringstream ss; - ss << "layer_" << i; - cfg.inputDefs.push_back({INPUT_DATA, ss.str(), layerSize, 0}); - LayerInputConfig* input = cfg.layerConfig.add_inputs(); - ImageConfig* img_conf = input->mutable_image_conf(); - img_conf->set_channels(pm.ic); - img_conf->set_img_size_y(pm.ih); - img_conf->set_img_size(pm.iw); - } -} - -void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) { - CHECK_GE(nInputs, 1UL); - TestConfig dnnConfig; - getAddtoConfig(dnnConfig, pm, nInputs); - dnnConfig.layerConfig.set_type("mkldnn_addto"); - for (auto withBias : {false, true}) { - dnnConfig.biasSize = withBias ? pm.ic * pm.ih * pm.iw : 0; - RUN_MKLDNN_TEST_LAYER(dnnConfig, "addto", pm) - } -} - -TEST(MKLDNNLayer, AddtoLayer) { - testAddtoLayer({16, 5, 14, 14}, 1); - testAddtoLayer({8, 10, 8, 8}, 2); - testAddtoLayer({4, 12, 1, 1}, 3); -} - -static void getMKLDNNConcatConfig(TestConfig& cfg, - const std::vector& inputs) { - CHECK_GE(inputs.size(), 2UL) << "at least two inputs"; - int oc = inputs[0].ic; - for (size_t i = 1; i < inputs.size(); ++i) { - CHECK_EQ(inputs[i].bs, inputs[0].bs); - CHECK_EQ(inputs[i].ih, inputs[0].ih); - CHECK_EQ(inputs[i].iw, inputs[0].iw); - oc += inputs[i].ic; - } - cfg.biasSize = 0; - cfg.layerConfig.set_type("mkldnn_concat"); - cfg.layerConfig.set_size(oc * inputs[0].ih * inputs[0].iw); - cfg.layerConfig.set_active_type("relu"); - for (size_t i = 0; i < inputs.size(); ++i) { - std::stringstream ss; - ss << "layer_" << i; - cfg.inputDefs.push_back( - {INPUT_DATA, - ss.str(), - (size_t)(inputs[i].ic) * inputs[i].ih * inputs[i].iw, - 0}); - LayerInputConfig* input = cfg.layerConfig.add_inputs(); - ImageConfig* img_conf = input->mutable_image_conf(); - img_conf->set_channels(inputs[i].ic); - img_conf->set_img_size_y(inputs[i].ih); - img_conf->set_img_size(inputs[i].iw); - } -} - -void testConcatLayer(const std::vector& inputs) { - TestConfig dnnConfig; - getMKLDNNConcatConfig(dnnConfig, inputs); - RUN_MKLDNN_TEST_LAYER(dnnConfig, "concat", inputs[0]) -} - -TEST(MKLDNNLayer, ConcatLayer) { - testConcatLayer({{64, 128, 1, 1}, {64, 32, 1, 1}, {64, 64, 1, 1}}); - testConcatLayer({{32, 100, 8, 8}, {32, 10, 8, 8}}); -} - -void testActivation(std::string actType, const testImageDesc& pm) { - // TODO(TJ): remove me when paddle support elu activation - if (actType == "mkldnn_elu") { - return; - } - const std::string compareTypes[] = {actType, actType.erase(0, 7)}; - TestConfig cfg; - getAddtoConfig(cfg, pm); - TestConfig ref = cfg; - cfg.layerConfig.set_active_type(compareTypes[0]); - ref.layerConfig.set_active_type(compareTypes[1]); - RUN_MKLDNN_TEST(cfg, ref, pm) -} - -TEST(MKLDNNActivation, Activations) { - auto types = MKLDNNActivation::getAllRegisteredTypes(); - for (auto type : types) { - /* bs, c, h, w*/ - testActivation(type, {16, 64, 32, 32}); - testActivation(type, {2, 8, 1, 1}); - } -} - -DECLARE_string(config_args); -TEST(MKLDNNNet, net) { - std::vector cases = {"simple", "branch"}; - for (auto name : cases) { - std::string config = "./gserver/tests/mkldnn_" + name + "_net.conf"; - for (auto channels : {2, 32}) { - std::ostringstream oss; - oss << "channels=" << channels; - FLAGS_config_args = oss.str(); - MKLDNNTester::runNetTest(config); - } - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - FLAGS_use_gpu = false; - FLAGS_use_mkldnn = true; - initMain(argc, argv); - initPython(argc, argv); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_MaxPoolingWithMaskOutput.cpp b/paddle/gserver/tests/test_MaxPoolingWithMaskOutput.cpp deleted file mode 100644 index 5188d2abed899a210de66084109034ee381cd078..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_MaxPoolingWithMaskOutput.cpp +++ /dev/null @@ -1,117 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include - -#include "LayerGradUtil.h" -#include "paddle/math/MathUtils.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; - -void setPoolConfig(TestConfig* config, - PoolConfig* pool, - const string& poolType) { - (*config).biasSize = 0; - (*config).layerConfig.set_type("pool"); - (*config).layerConfig.set_num_filters(1); - - int kw = 3, kh = 3; - int pw = 0, ph = 0; - int sw = 2, sh = 2; - pool->set_pool_type(poolType); - pool->set_channels(1); - pool->set_size_x(kw); - pool->set_size_y(kh); - pool->set_start(0); - pool->set_padding(pw); - pool->set_padding_y(ph); - pool->set_stride(sw); - pool->set_stride_y(sh); - - int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); - int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); - pool->set_output_x(ow); - pool->set_output_y(oh); -} - -void doOneMaxPoolingWithMaskOutputTest(MatrixPtr& inputMat, - const string& poolType, - bool use_gpu, - MatrixPtr& maskMat) { - TestConfig config; - config.inputDefs.push_back({INPUT_DATA, "layer_0", 25, 0}); - LayerInputConfig* input = config.layerConfig.add_inputs(); - PoolConfig* pool = input->mutable_pool_conf(); - - pool->set_img_size(5); - pool->set_img_size_y(5); - setPoolConfig(&config, pool, poolType); - config.layerConfig.set_size(pool->output_x() * pool->output_y() * - pool->channels()); - - config.layerConfig.set_name("MaxPoolWithMask"); - - std::vector dataLayers; - LayerMap layerMap; - vector datas; - - initDataLayer(config, - &dataLayers, - &datas, - &layerMap, - "MaxPoolWithMask", - 1, - false, - use_gpu); - - dataLayers[0]->getOutputValue()->copyFrom(*inputMat); - - FLAGS_use_gpu = use_gpu; - std::vector parameters; - LayerPtr maxPoolingWithMaskOutputLayer; - initTestLayer(config, &layerMap, ¶meters, &maxPoolingWithMaskOutputLayer); - maxPoolingWithMaskOutputLayer->forward(PASS_GC); - - checkMatrixEqual(maxPoolingWithMaskOutputLayer->getOutput("mask").value, - maskMat); -} - -TEST(Layer, maxPoolingWithMaskOutputLayerFwd) { - bool useGpu = false; - MatrixPtr inputMat; - MatrixPtr maskMat; - real inputData[] = {0.1, 0.1, 0.5, 0.5, 1.1, 0.2, 0.2, 0.6, 0.1, - 0.1, 0.3, 0.3, 0.7, 0.1, 0.1, 0.4, 0.4, 0.8, - 0.8, 0.1, 1.0, 2.0, 3.0, 0.0, 9.0}; - real maskData[] = {12, 4, 22, 24}; - - inputMat = Matrix::create(1, 25, false, useGpu); - maskMat = Matrix::create(1, 4, false, useGpu); - inputMat->setData(inputData); - maskMat->setData(maskData); - doOneMaxPoolingWithMaskOutputTest( - inputMat, "max-pool-with-mask", useGpu, maskMat); -#ifdef PADDLE_WITH_CUDA - useGpu = true; - inputMat = Matrix::create(1, 25, false, useGpu); - maskMat = Matrix::create(1, 4, false, useGpu); - inputMat->copyFrom(inputData, 25); - maskMat->copyFrom(maskData, 4); - doOneMaxPoolingWithMaskOutputTest( - inputMat, "max-pool-with-mask", useGpu, maskMat); -#endif -} diff --git a/paddle/gserver/tests/test_MultinomialSampler.cpp b/paddle/gserver/tests/test_MultinomialSampler.cpp deleted file mode 100644 index 043025239e744601cbef3ca5c241509872963bd8..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_MultinomialSampler.cpp +++ /dev/null @@ -1,147 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include -#include - -#undef PADDLE_DISABLE_TIMER -#include "paddle/utils/Stat.h" - -#include "paddle/gserver/layers/MultinomialSampler.h" -#include "paddle/utils/Util.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -class MultinomialSamplerTester : public MultinomialSampler { - public: - MultinomialSamplerTester(real* prob, int size) - : MultinomialSampler(prob, size) {} - - template - int testGen(Rand1 rand1) { - return gen1(rand1); - } -}; - -TEST(MultinomialSampler, gen) { - int numGrids = 1024 * 1024; - int size = 1024 * 4; - default_random_engine reng; - - for (size_t iter = 0; iter < 256; ++iter) { - uniform_int_distribution rand(1, numGrids / size * 1.8); - vector prob; - int sum = 0; - for (int i = 0; i < size; ++i) { - prob.push_back(rand(reng)); - sum += prob.back(); - } - - CHECK_LE(sum, numGrids); - prob.back() += numGrids - sum; - - vector counts(size); - MultinomialSamplerTester sampler(&prob[0], size); - counts.assign(size, 0); - { - double s = (double)size / (double)numGrids; - REGISTER_TIMER("MultinomialSampler"); - for (double i = 0; i < numGrids; ++i) { - int ret = sampler.testGen([i, s]() { return s * i; }); - if (ret < 0 || ret >= size) { - EXPECT_GE(ret, 0); - EXPECT_LT(ret, size); - break; - } - ++counts[ret]; - } - } - for (int i = 0; i < size; ++i) { - if (prob[i] != counts[i]) { - EXPECT_EQ(prob[i], counts[i]); - LOG(INFO) << iter; - break; - } - } - } -} - -void benchmarkRandom() { - int n = 1024 * 1024; - - int sum; - double sum1; - - sum = 0; - unsigned int seed = 1; - { - REGISTER_TIMER("crand"); - for (int i = 0; i < n; ++i) { - sum += rand_r(&seed) % 1000; - } - } - LOG(INFO) << "sum=" << sum; - - default_random_engine reng; - uniform_int_distribution rand(1, 1000); - sum = 0; - { - REGISTER_TIMER("stdrand"); - for (int i = 0; i < n; ++i) { - sum += rand(reng); - } - } - LOG(INFO) << "sum=" << sum; - - sum = 0; - { - REGISTER_TIMER("default_random_engine"); - for (int i = 0; i < n; ++i) { - sum += reng(); - } - } - LOG(INFO) << "sum=" << sum; - - uniform_real_distribution rand1(0, 1); - sum1 = 0; - { - REGISTER_TIMER("stdrand1"); - for (int i = 0; i < n; ++i) { - sum1 += rand1(reng); - } - } - LOG(INFO) << "sum1=" << sum1; - - sum1 = 0; - { - real a = 1.0f / (real)RAND_MAX; - REGISTER_TIMER("crand1"); - for (int i = 0; i < n; ++i) { - sum1 += a * rand_r(&seed); - } - } - LOG(INFO) << "sum1=" << sum1; -} - -int main(int argc, char** argv) { - initMain(argc, argv); - testing::InitGoogleTest(&argc, argv); - benchmarkRandom(); - int ret = RUN_ALL_TESTS(); - globalStat.printSegTimerStatus(); - return ret; -} diff --git a/paddle/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp deleted file mode 100644 index fda3f2f7934adde09303f443ca5e8de6a7d077cd..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_NetworkCompare.cpp +++ /dev/null @@ -1,294 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#undef PADDLE_DISABLE_TIMER -#include -#include -#include -#include - -#include "paddle/testing/TestUtil.h" -#include "paddle/trainer/Trainer.h" -#include "paddle/utils/Stat.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_int32(gpu_id); -DECLARE_double(checkgrad_eps); -DEFINE_bool(use_label, true, "input label or sequence label"); -DEFINE_bool(static_para, false, "static parameter"); - -struct DataIn { - std::vector inArgs; - std::vector outGrads; - std::vector paraValues; -}; - -struct DataOut { - std::vector outValues; - std::vector paraGrads; -}; - -void initArgument(DataIn& data, - const std::string& configPath, - bool useGpu = FLAGS_use_gpu) { - TrainerConfigHelper config(configPath); - size_t batchSize = config.getOptConfig().batch_size(); - - for (const auto& layer_name : config.getModelConfig().input_layer_names()) { - auto layer_config = std::find_if(config.getModelConfig().layers().begin(), - config.getModelConfig().layers().end(), - [=](const LayerConfig& layer_config) { - return layer_config.name() == layer_name; - }); - CHECK(layer_config != config.getModelConfig().layers().end()); - - size_t layerSize = layer_config->size(); - Argument arg; - arg.value = Matrix::create(batchSize, layerSize, false, useGpu); - arg.grad = Matrix::create(batchSize, layerSize, false, useGpu); - arg.value->randomizeUniform(); - arg.value->add(-0.5); - arg.value->sigmoid(*arg.value); - arg.grad->zeroMem(); - if (FLAGS_use_label) { - arg.ids = VectorT::create(batchSize, useGpu); - arg.ids->rand(layerSize); - } - generateSequenceStartPositions(batchSize, arg.sequenceStartPositions); - data.inArgs.push_back(arg); - } - - for (const auto& layer_name : config.getModelConfig().output_layer_names()) { - auto layer_config = std::find_if(config.getModelConfig().layers().begin(), - config.getModelConfig().layers().end(), - [=](const LayerConfig& layer_config) { - return layer_config.name() == layer_name; - }); - CHECK(layer_config != config.getModelConfig().layers().end()); - - size_t layerSize = layer_config->size(); - MatrixPtr grad = Matrix::create(batchSize, layerSize, false, useGpu); - grad->randomizeUniform(); - data.outGrads.push_back(grad); - } - - for (const auto& para_config : config.getModelConfig().parameters()) { - VectorPtr value = Vector::create(para_config.size(), useGpu); - value->randnorm(0, 2); - data.paraValues.push_back(value); - } -} - -void calcGradient(DataIn& in, DataOut& out, const std::string& configPath) { - *ThreadLocalRand::getSeed() = 0; - srand(0); - - Trainer trainer; - auto config = std::make_shared(configPath); - trainer.init(config, false); - - std::vector parameters; - vector outArgs; - - auto gradientMachine = trainer.getGradientMachine(); - parameters = gradientMachine->getParameters(); - if (FLAGS_static_para) { - for (size_t i = 0; i < parameters.size(); i++) { - parameters[i]->getBuf(PARAMETER_VALUE)->one(); - } - } else { - for (size_t i = 0; i < in.paraValues.size(); i++) { - parameters[i]->getBuf(PARAMETER_VALUE)->copyFrom(*in.paraValues[i]); - } - } - gradientMachine->start(); - gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN); - for (size_t i = 0; i < in.outGrads.size(); i++) { - // If the all the layers in the config have no parameters, also - // not set NeedGradient(), the outArgs[i] will be nullptr. - outArgs[i].grad->copyFrom(*in.outGrads[i]); - } - gradientMachine->backward(); - for (size_t i = 0; i < in.outGrads.size(); i++) { - MatrixPtr value = Matrix::create(outArgs[i].value->getHeight(), - outArgs[i].value->getWidth(), - false, - false); - value->copyFrom(*outArgs[i].value); - out.outValues.push_back(value); - } - for (size_t i = 0; i < in.paraValues.size(); i++) { - VectorPtr grad = Vector::create( - parameters[i]->getBuf(PARAMETER_GRADIENT)->getSize(), false); - grad->copyFrom(*parameters[i]->getBuf(PARAMETER_GRADIENT)); - out.paraGrads.push_back(grad); - } - - for (int i = 0; i < 20; i++) { - REGISTER_TIMER("forward"); - gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN); - } - for (int i = 0; i < 20; i++) { - REGISTER_TIMER("backward"); - gradientMachine->backward(); - } - - gradientMachine->finish(); -} - -void checkBuffer(real* A, - const char* desA, - real* B, - const char* desB, - size_t len, - size_t width = 1) { - int nNum = 0; - for (size_t i = 0; i < len; ++i) { - real diff = fabs(A[i] - B[i]); - if (diff > 0.0f && - diff / std::max(fabs(A[i]), fabs(B[i])) > FLAGS_checkgrad_eps) { - nNum++; - LOG(INFO) << "Row: " << i / width << ", " << desA << " : " << A[i] - << " " << desB << " : " << B[i]; - } - } - EXPECT_EQ(0, nNum); -} - -void compareGradient(DataOut& outA, DataOut& outB) { - LOG(INFO) << "------------------------------" - << " Check Network Output " - << "------------------------------"; - for (size_t i = 0; i < outA.outValues.size(); ++i) { - LOG(INFO) << "OUTPUT VALUE: " << i; - checkBuffer(outA.outValues[i]->getData(), - "network A output", - outB.outValues[i]->getData(), - "network B output", - outA.outValues[i]->getElementCnt(), - outA.outValues[i]->getWidth()); - } - - if (!FLAGS_static_para) { - LOG(INFO) << "------------------------------" - << " Check Parameters " - << "------------------------------"; - for (size_t i = 0; i < outA.paraGrads.size(); ++i) { - LOG(INFO) << "PARAMETER GRADIENT: " << i; - checkBuffer(outA.paraGrads[i]->getData(), - "Network A", - outB.paraGrads[i]->getData(), - "Network B", - outA.paraGrads[i]->getSize()); - } - } -} - -void compareNetwork(const std::string& config_file_a, - const std::string& config_file_b) { - DataIn in; - initArgument(in, config_file_a); - - DataOut dataA; - calcGradient(in, dataA, config_file_a); - LOG(INFO) << "forwardBackward of Network A is finished"; - globalStat.printSegTimerStatus(); - globalStat.reset(); - LOG(INFO) << "\n\n"; - - DataOut dataB; - calcGradient(in, dataB, config_file_b); - LOG(INFO) << "forwardBackward of the Network B is finished"; - globalStat.printSegTimerStatus(); - globalStat.reset(); - LOG(INFO) << "\n\n"; - - compareGradient(dataA, dataB); -} - -TEST(Compare, concat_dotmul) { - std::string config_file_a = "./gserver/tests/concat_dotmul_a.conf"; - std::string config_file_b = "./gserver/tests/concat_dotmul_b.conf"; - compareNetwork(config_file_a, config_file_b); -} - -TEST(Compare, concat_fullmatrix) { - std::string config_file_a = "./gserver/tests/concat_fullmatrix_a.conf"; - std::string config_file_b = "./gserver/tests/concat_fullmatrix_b.conf"; - compareNetwork(config_file_a, config_file_b); -} - -TEST(Compare, concat_table) { - std::string config_file_a = "./gserver/tests/concat_table_a.conf"; - std::string config_file_b = "./gserver/tests/concat_table_b.conf"; - compareNetwork(config_file_a, config_file_b); -} - -TEST(Compare, concat_slice) { - std::string config_file_a = "./gserver/tests/concat_slice_a.conf"; - std::string config_file_b = "./gserver/tests/concat_slice_b.conf"; - compareNetwork(config_file_a, config_file_b); -} - -#ifdef PADDLE_WITH_CUDA -TEST(Compare, img_pool) { - std::string config_file_a = "./gserver/tests/img_pool_a.conf"; - std::string config_file_b = "./gserver/tests/img_pool_b.conf"; - bool useGpu = FLAGS_use_gpu; - FLAGS_use_gpu = true; - compareNetwork(config_file_a, config_file_b); - FLAGS_use_gpu = useGpu; -} - -TEST(Compare, img_conv) { - std::string config_file_a = "./gserver/tests/img_conv_a.conf"; - std::string config_file_b = "./gserver/tests/img_conv_b.conf"; - bool useGpu = FLAGS_use_gpu; - FLAGS_use_gpu = true; - compareNetwork(config_file_a, config_file_b); - FLAGS_use_gpu = useGpu; -} - -// Test cudnn_conv and exconv give the same result -TEST(Compare, img_conv2) { - std::string config_file_a = "./gserver/tests/img_conv_cudnn.py"; - std::string config_file_b = "./gserver/tests/img_conv_exconv.py"; - bool useGpu = FLAGS_use_gpu; - double eps = FLAGS_checkgrad_eps; - FLAGS_use_gpu = true; - // Sometimes, this unit test will fail with 1e-2 - FLAGS_checkgrad_eps = 4e-2; - compareNetwork(config_file_a, config_file_b); - FLAGS_use_gpu = useGpu; - FLAGS_checkgrad_eps = eps; -} -#endif - -DEFINE_string(config_file_a, "", "config of one network to compare"); -DEFINE_string(config_file_b, "", "config of another network to compare"); -TEST(Compare, network) { - if (FLAGS_config_file_a != "" && FLAGS_config_file_b != "") { - compareNetwork(FLAGS_config_file_a, FLAGS_config_file_b); - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - paddle::initMain(argc, argv); - initPython(argc, argv); - int ret = RUN_ALL_TESTS(); - return ret; -} diff --git a/paddle/gserver/tests/test_PyDataProvider.cpp b/paddle/gserver/tests/test_PyDataProvider.cpp deleted file mode 100644 index a1dee9795077b835392469b5085e9728679a1664..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_PyDataProvider.cpp +++ /dev/null @@ -1,175 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include - -#include "paddle/gserver/dataproviders/PyDataProvider.h" -#include "paddle/utils/Util.h" - -#include "paddle/testing/TestUtil.h" - -using namespace std; // NOLINT -using namespace paddle; // NOLINT - -void simpleValueCheck(const vector& argumentList, bool useGpu); -void simpleSequenceCheck(const vector& argumentList, int sample_num); - -TEST(PyDataProvider, py_fill_slots) { - DataConfig config; - config.set_type("py"); - config.set_async_load_data(false); - config.set_load_data_module(std::string("pyDataProvider")); - config.set_load_data_object(std::string("SimpleDataProvider")); - config.clear_files(); - std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList"; - config.set_files(dataFile); -#ifndef PADDLE_WITH_CUDA - bool useGpu = false; -#else - bool useGpu = true; -#endif - unique_ptr dataProvider(DataProvider::create(config, useGpu)); - DataBatch dataBatch; - dataProvider->getNextBatchInternal(2, &dataBatch); - const std::vector& argumentList = dataBatch.getStreams(); - // Check size - EXPECT_EQ(argumentList.size(), 3UL); - EXPECT_EQ(argumentList[0].value->getWidth(), 3UL); - EXPECT_EQ(argumentList[0].value->getHeight(), 2UL); - EXPECT_EQ(argumentList[0].value->getElementCnt(), 6UL); - EXPECT_EQ(argumentList[1].value->getWidth(), 7UL); - EXPECT_EQ(argumentList[1].value->getHeight(), 2UL); - EXPECT_EQ(argumentList[1].value->getElementCnt(), 4UL); - EXPECT_EQ(argumentList[2].ids->getSize(), 2UL); - // Check value - simpleValueCheck(argumentList, useGpu); - // Check sequenceStartPositions - simpleSequenceCheck(argumentList, 2); -} - -TEST(PyDataProvider, py_fill_nest_slots) { - DataConfig config; - config.set_type("py"); - config.set_async_load_data(false); - config.set_load_data_module(std::string("pyDataProvider")); - config.set_load_data_object(std::string("SimpleNestDataProvider")); - config.clear_files(); - std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList"; - config.set_files(dataFile); - EXPECT_EQ(config.IsInitialized(), true); -#ifndef PADDLE_WITH_CUDA - bool useGpu = false; -#else - bool useGpu = true; -#endif - unique_ptr dataProvider(DataProvider::create(config, useGpu)); - DataBatch dataBatch; - dataProvider->getNextBatchInternal(2, &dataBatch); - const std::vector& argumentList = dataBatch.getStreams(); - // Check size - EXPECT_EQ(argumentList.size(), 3UL); - EXPECT_EQ(argumentList[0].value->getWidth(), 3UL); - EXPECT_EQ(argumentList[0].value->getHeight(), 4UL); - EXPECT_EQ(argumentList[0].value->getElementCnt(), 12UL); - EXPECT_EQ(argumentList[1].value->getWidth(), 7UL); - EXPECT_EQ(argumentList[1].value->getHeight(), 4UL); - EXPECT_EQ(argumentList[1].value->getElementCnt(), 8UL); - EXPECT_EQ(argumentList[2].ids->getSize(), 4UL); - // Check value - simpleValueCheck(argumentList, useGpu); - // Check sequenceStartPositions - simpleSequenceCheck(argumentList, 4); - // Check subSequenceStartPositions - EXPECT_EQ(argumentList[0].subSequenceStartPositions->getSize(), 4UL); - EXPECT_EQ(argumentList[1].subSequenceStartPositions->getSize(), 3UL); - EXPECT_EQ(argumentList[2].subSequenceStartPositions->getSize(), 4UL); - for (size_t i = 0; i < argumentList.size(); i++) { - EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(0), 0); - EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(1), 1); - if (i != 1) { - EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(2), 2); - EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(3), 4); - } else { - EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(2), 4); - } - } -} - -void simpleValueCheck(const vector& argumentList, bool useGpu) { - // Dense - real* data; - if (useGpu) { - MatrixPtr cpuMatrixPtr = Matrix::create(argumentList[0].value->getHeight(), - argumentList[0].value->getWidth(), - 0, - 0); - cpuMatrixPtr->copyFrom(*argumentList[0].value); - data = cpuMatrixPtr->getData(); - } else { - data = argumentList[0].value->getData(); - } - for (size_t i = 0; i < argumentList[0].value->getElementCnt(); ++i) { - EXPECT_EQ(*(data + i), (float)(i % 3 + 1)); - } - // Sparse without value - GpuSparseMatrixPtr matGpu; - CpuSparseMatrixPtr matCpu; - if (useGpu) { - matGpu = dynamic_pointer_cast(argumentList[1].value); - ASSERT_TRUE(matGpu != NULL); - } else { - data = argumentList[0].value->getData(); - matCpu = dynamic_pointer_cast(argumentList[1].value); - ASSERT_TRUE(matCpu != NULL); - } - for (size_t i = 0; i < argumentList[1].value->getHeight(); ++i) { - size_t colNum = useGpu ? matGpu->getColNum(i) : matCpu->getColNum(i); - EXPECT_EQ(colNum, (size_t)2); - const int* buf = useGpu ? matGpu->getRowCols(i) : matCpu->getRowCols(i); - for (size_t j = 0; j < colNum; ++j) { - EXPECT_EQ((size_t)buf[j], (size_t)(j + 1)); - } - } - // Index - for (size_t j = 0; j < argumentList[2].ids->getSize(); ++j) { - EXPECT_EQ((size_t)argumentList[2].ids->get(j), 0UL); - } -} - -void simpleSequenceCheck(const vector& argumentList, int sample_num) { - EXPECT_EQ(argumentList[0].sequenceStartPositions->getSize(), 3UL); - EXPECT_EQ(argumentList[1].sequenceStartPositions->getSize(), 2UL); - EXPECT_EQ(argumentList[2].sequenceStartPositions->getSize(), 3UL); - for (size_t i = 0; i < argumentList.size(); i++) { - EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(0), 0); - if (i != 1) { - EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(1), 1); - EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(2), - sample_num); - } else { - EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(1), - sample_num); - } - } -} - -int main(int argc, char** argv) { - initMain(argc, argv); - initPython(argc, argv); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_PyDataProvider2.cpp b/paddle/gserver/tests/test_PyDataProvider2.cpp deleted file mode 100644 index b39fb3534509ebde2702c02e35800fe3ed6016c3..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_PyDataProvider2.cpp +++ /dev/null @@ -1,409 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifndef PADDLE_NO_PYTHON -#include -#include -#include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Util.h" - -DEFINE_string(train_list, "unittest.list", "file list for unittest"); - -namespace paddle { -namespace unittest { -namespace pydp2 { -extern void setOnPoolFilledHook(const std::function &func); -extern void clearOnPoolFilledHook(); - -} // namespace pydp2 -} // namespace unittest -} // namespace paddle - -const paddle::real epsilon = 1e-5; - -static inline int64_t readDataBatch(paddle::DataBatch *batch, - const std::string &funcName, - int64_t batchSize = 65535) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object(funcName); - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->setSkipShuffle(); - provider->reset(); - return provider->getNextBatchInternal(batchSize, batch); -} - -TEST(PyDataProvider2, dense_no_seq) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_dense_no_seq"); - - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - - provider->setSkipShuffle(); // skip shuffle for unittest. - - paddle::DataBatch batch; - for (size_t pass = 0; pass < 2; ++pass) { // read 2 passes - provider->reset(); - int64_t num = provider->getNextBatchInternal(100, &batch); - ASSERT_NE(num, 0); - ASSERT_EQ((size_t)batch.getStreams().size(), (size_t)1); - ASSERT_EQ((size_t)batch.getSize(), (size_t)100); - // Check batch data. - for (size_t i = 0; i < 100; ++i) { - for (size_t j = 0; j < 200; ++j) { - paddle::real tmp = (paddle::real)((j - 100.0) * (i + 1) / 200.0); - ASSERT_NEAR( - batch.getStreams()[0].value->getData()[i * 200 + j], tmp, epsilon); - } - } - - num = provider->getNextBatchInternal(100, &batch); - ASSERT_NE(num, 0); - ASSERT_EQ(batch.getStreams().size(), (size_t)1); - ASSERT_EQ((size_t)batch.getSize(), (size_t)100); - // Check batch data. - for (size_t i = 0; i < 100; ++i) { - size_t ii = i + 100; - for (size_t j = 0; j < 200; ++j) { - paddle::real tmp = (paddle::real)((j - 100.0) * (ii + 1) / 200.0); - ASSERT_NEAR( - batch.getStreams()[0].value->getData()[i * 200 + j], tmp, epsilon); - } - } - num = provider->getNextBatchInternal(100, &batch); - ASSERT_EQ(num, 0); - } -} - -TEST(PyDataProvider2, index_no_seq) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_index_no_seq"); - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - - provider->setSkipShuffle(); // skip shuffle for unittest. - paddle::DataBatch batch; - for (size_t pass = 0; pass < 2; ++pass) { - provider->reset(); - int64_t num = provider->getNextBatchInternal(10000, &batch); - CHECK_EQ(num, 200); - for (int i = 0; i < 200; ++i) { - CHECK_EQ(i, batch.getStreams()[0].ids->getData()[i]); - } - } -} - -TEST(PyDataProvider2, init_hook) { - paddle::PyObjectPtr pickle = paddle::py::import("pickle"); - paddle::PyObjectPtr globals(PyModule_GetDict(PyImport_AddModule("__main__"))); - PyDict_SetItemString(globals.get(), "pickle", pickle.get()); - paddle::PyObjectPtr locals(PyDict_New()); - paddle::PyObjectPtr mdl(PyRun_String( - "dumps = pickle.dumps({'value':[float(x) for x in xrange(20)]})", - Py_file_input, - globals.get(), - locals.get())); - CHECK_PY(mdl) << "Error!"; - paddle::PyObjectPtr dps(PyDict_GetItemString(locals.get(), "dumps")); - CHECK_PY(dps) << "Error!"; - - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_init_hook"); - config.set_load_data_args(PyString_AsString(dps.get())); - - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->setSkipShuffle(); // skip shuffle for unittest. - provider->reset(); - paddle::DataBatch batch; - int64_t num = provider->getNextBatchInternal(100000, &batch); - ASSERT_EQ(num, 200); - auto &mat = batch.getStreams()[0].value; - ASSERT_EQ((size_t)mat->getWidth(), (size_t)20); - for (size_t i = 0; i < 200; ++i) { - for (size_t j = 0; j < 20; ++j) { - ASSERT_NEAR((paddle::real)j, mat->getData()[i * 20 + j], epsilon); - } - } -} - -TEST(PyDataProvider2, sparse_no_value_no_seq) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_sparse_non_value_no_seq"); - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->setSkipShuffle(); - provider->reset(); - paddle::DataBatch batch; - int64_t num = provider->getNextBatchInternal(10000, &batch); - CHECK_EQ(num, 200); - auto csm = std::dynamic_pointer_cast( - batch.getStreams()[0].value); - CHECK(csm != nullptr); - for (int i = 0; i < 200; ++i) { - CHECK_EQ(csm->getColNum(i), (size_t)10); - int *cols = csm->getRowCols(i); - for (int j = 0; j < 10; ++j) { - CHECK_EQ(cols[j], (i + 1) * (j + 1)); - } - } -} - -TEST(PyDataProvider2, sparse_value_no_seq) { - paddle::DataBatch batch; - CHECK_EQ(readDataBatch(&batch, "test_sparse_value_no_seq"), 200); - auto csm = std::dynamic_pointer_cast( - batch.getStreams()[0].value); - CHECK(csm != nullptr); - for (int i = 0; i < 200; ++i) { - CHECK_EQ(csm->getColNum(i), (size_t)10); - int *cols = csm->getRowCols(i); - real *dat = csm->getRowValues(i); - for (int j = 0; j < 10; ++j) { - EXPECT_EQ(cols[j], (i + 1) * (j + 1)); - EXPECT_EQ(dat[j], real(j) / real(i + 1)); - } - } -} - -TEST(PyDataProvider2, index_seq) { - paddle::DataBatch batch; - CHECK_EQ(readDataBatch(&batch, "test_index_seq"), 200); - auto &arg = batch.getStreams()[0]; - CHECK_EQ((int)arg.ids->getSize(), (200 + 1) * 200 / 2); - size_t tmp = 0; - for (size_t i = 0; i < 200; ++i) { // CHECK DATA CORRECT - for (size_t j = 0; j < i + 1; ++j) { - ASSERT_EQ((size_t)arg.ids->getData()[tmp], j); - ++tmp; - } - } - ASSERT_EQ(arg.sequenceStartPositions->getSize(), (size_t)201); - tmp = 0; - for (size_t i = 0; i < 200; ++i) { - tmp += i; - ASSERT_EQ((size_t)arg.sequenceStartPositions->getData(false)[i], tmp); - } - tmp += 200; - ASSERT_EQ((size_t)arg.sequenceStartPositions->getData(false)[200], tmp); -} - -TEST(PyDataProvider2, index_sub_seq) { - paddle::DataBatch batch; - ASSERT_EQ(readDataBatch(&batch, "test_index_sub_seq"), 200); - auto &arg = batch.getStreams()[0]; - size_t tmp = 0; - for (size_t i = 0; i < 200; ++i) { - for (size_t j = 0; j < i + 1; ++j) { - for (size_t k = 0; k < j + 1; ++k) { - CHECK_EQ((size_t)arg.ids->getData()[tmp++], k); - } - } - } - - CHECK_EQ(tmp, arg.ids->getSize()); - - ASSERT_EQ((size_t)arg.sequenceStartPositions->getSize(), (size_t)201); - ASSERT_EQ(arg.subSequenceStartPositions->getData(false)[0], 0); - ASSERT_EQ(arg.sequenceStartPositions->getData(false)[0], 0); - size_t idx = 1; - tmp = 0; - for (size_t i = 0; i < 200; ++i) { - for (size_t j = 0; j < i + 1; ++j) { - tmp += j + 1; - ASSERT_EQ((size_t)arg.subSequenceStartPositions->getData(false)[idx], - (size_t)tmp); - ++idx; - } - ASSERT_EQ((size_t)arg.sequenceStartPositions->getData(false)[i + 1], tmp); - } -} - -TEST(PyDataProvider2, min_pool_size) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_min_pool_size"); - config.set_load_data_args(""); - size_t totalData = 1 << 14; - constexpr size_t batchSize = 100; - constexpr size_t minPoolSize = 1000; - paddle::DataBatch batch; - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->reset(); - - paddle::unittest::pydp2::setOnPoolFilledHook([&](size_t poolSize) { - if (totalData > batchSize) { - CHECK_GE(poolSize, std::min(totalData - batchSize, minPoolSize)); - } - }); - while (true) { - int64_t realBatchSize = provider->getNextBatchInternal(batchSize, &batch); - if (realBatchSize) { - totalData -= realBatchSize; - } else { - break; - } - } - paddle::unittest::pydp2::clearOnPoolFilledHook(); -} - -TEST(PyDataProvider2, can_over_batch_size) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_can_over_batch_size"); - config.set_load_data_args(""); - paddle::DataBatch batch; - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->reset(); - constexpr size_t batchSize = 100; - while (true) { - int64_t realBatchSize = provider->getNextBatchInternal(batchSize, &batch); - if (realBatchSize) { - CHECK_LE(static_cast(realBatchSize), batchSize); - } else { - break; - } - } -} - -TEST(PyDataProvider2, input_order) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_input_order"); - config.set_load_data_args(""); - - paddle::ModelConfig modelConfig; - *modelConfig.add_input_layer_names() = "input1"; - *modelConfig.add_input_layer_names() = "input2"; - paddle::DataBatch batch; - std::unique_ptr provider( - paddle::DataProvider::create(config, modelConfig, false)); - provider->reset(); - constexpr size_t batchSize = 100; - while (true) { - int64_t realBatchSize = provider->getNextBatchInternal(batchSize, &batch); - if (!realBatchSize) { - break; - } - ASSERT_EQ(batch.getStreams().size(), static_cast(2)); - for (int64_t i = 0; i < realBatchSize; ++i) { - ASSERT_EQ(batch.getStream(0).ids->getData()[i], 0); - ASSERT_EQ(batch.getStream(1).ids->getData()[i], 1); - } - } -} - -TEST(PyDataProvider2, test_check) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_check"); - config.set_load_data_args(""); - paddle::DataBatch batch; - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->reset(); - while (true) { - int64_t realBatchSize = provider->getNextBatchInternal(100, &batch); - if (!realBatchSize) { - break; - } else { - auto &ivec = batch.getStream(0).ids; - for (size_t i = 0; i < ivec->getSize(); ++i) { - CHECK_LT(ivec->getData()[i], 10); - } - } - } -} - -TEST(PyDataProvider2, multiThread) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_dense_no_seq"); - config.set_async_load_data(true); - - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - provider->reset(); - paddle::DataBatch batch; - provider->getNextBatch(100, &batch); - provider->reset(); - provider.reset(); -} - -TEST(PyDataProvider2, minPoolSizeWithCache) { - paddle::DataConfig config; - config.set_type("py2"); - config.set_files(FLAGS_train_list.c_str()); - config.set_load_data_module("test_PyDataProvider2"); - config.set_load_data_object("test_min_pool_size_with_cache"); - config.set_async_load_data(true); - - std::unique_ptr provider( - paddle::DataProvider::create(config, false)); - - paddle::DataBatch batch; - - for (int i = 0; i < 10; ++i) { - provider->reset(); - int64_t sum = 0; - while (int64_t actualNum = provider->getNextBatch(100, &batch)) { - sum += actualNum; - } - ASSERT_EQ(1 << 20, sum); - } -} - -int main(int argc, char **argv) { - testing::InitGoogleTest(&argc, argv); - paddle::initMain(argc, argv); - paddle::initPython(argc, argv); - - std::ofstream fout(FLAGS_train_list); - CHECK(fout.is_open()); - fout << "stub file name" << std::endl; // in unittest, filename is not used. - fout.close(); - - return RUN_ALL_TESTS(); -} - -#endif diff --git a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp deleted file mode 100644 index 9770567b88a2af946b30439300540ed61694ba10..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp +++ /dev/null @@ -1,180 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include -#include -#include -#include - -DECLARE_int32(seed); - -using namespace paddle; // NOLINT -using namespace std; // NOLINT -class TrainerForTest : public paddle::Trainer { - public: - void startTrain() { - GradientMachine& gm = *this->trainerInternal_.getGradientMachine(); - gm.start(); - } - - void finishTrain() { - GradientMachine& gm = *this->trainerInternal_.getGradientMachine(); - gm.finish(); - } - - /** - * Get total dimension of all parameters. - * - * @return the total dimension of all parameters - */ - size_t getTotalParameterSize() const { - auto p = const_cast(this); - auto& params = p->getGradientMachine()->getParameters(); - return std::accumulate( - params.begin(), params.end(), 0UL, [](size_t a, const ParameterPtr& p) { - return a + p->getSize(); - }); - } -}; - -void CalCost(const string& conf, - const string& dir, - real* cost, - int num_passes) { - auto config = std::make_shared(conf); - TrainerForTest trainer; - trainer.init(config); - mkDir(dir.c_str()); - config->setSaveDir(dir); - auto dataProvider = trainer.getDataProvider(); - int32_t batchSize = config->getOptConfig().batch_size(); - real learningRate = config->getOptConfig().learning_rate(); - real momentum = 0; - real decayRate = 0; - int64_t dim = trainer.getTotalParameterSize(); - CpuVector vecW(dim); - CpuVector vecGradient(dim); - CpuVector vecMomentum(dim); - - // vecW needs to be assigned, otherwise the variable is an uncertain value. - - *ThreadLocalRand::getSeed() = FLAGS_seed; - vecW.randnorm(0, 0.1); - vecMomentum.randnorm(0, 0.1); - - trainer.startTrain(); - for (int i = 0; i < num_passes; ++i) { - real totalCost = 0; - dataProvider->reset(); - while (true) { - DataBatch dataBatch; - int num = dataProvider->getNextBatch(batchSize, &dataBatch); - if (num == 0) break; - totalCost += trainer.calcGradient(dataBatch, vecW, vecGradient); - sgdUpdate( - learningRate, momentum, decayRate, &vecW, &vecGradient, &vecMomentum); - } - cost[i] = totalCost; - } - trainer.finishTrain(); - rmDir(dir.c_str()); -} - -void test(const string& conf1, const string& conf2, double eps, bool useGpu) { - if (!paddle::version::isWithGpu() && useGpu) { - return; - } - FLAGS_use_gpu = useGpu; - int num_passes = 5; - real* cost1 = new real[num_passes]; - const string dir1 = "gserver/tests/t1"; - CalCost(conf1, dir1, cost1, num_passes); - - real* cost2 = new real[num_passes]; - const string dir2 = "gserver/tests/t2"; - CalCost(conf2, dir2, cost2, num_passes); - - for (int i = 0; i < num_passes; i++) { - LOG(INFO) << "num_passes: " << i << ", cost1=" << cost1[i] - << ", cost2=" << cost2[i] - << ", diff=" << std::abs(cost1[i] - cost2[i]); - ASSERT_NEAR(cost1[i], cost2[i], eps); - } - delete[] cost1; - delete[] cost2; -} - -TEST(RecurrentGradientMachine, HasSubSequence) { - for (bool useGpu : {false, true}) { - test("gserver/tests/sequence_layer_group.conf", - "gserver/tests/sequence_nest_layer_group.conf", - 1e-5, - useGpu); - } -} - -TEST(RecurrentGradientMachine, rnn) { - for (bool useGpu : {false, true}) { - test("gserver/tests/sequence_rnn.conf", - "gserver/tests/sequence_nest_rnn.conf", - 1e-6, - useGpu); - } -} - -TEST(RecurrentGradientMachine, rnn_multi_input) { - for (bool useGpu : {false, true}) { - test("gserver/tests/sequence_rnn_multi_input.conf", - "gserver/tests/sequence_nest_rnn_multi_input.conf", - 1e-6, - useGpu); - } -} - -TEST(RecurrentGradientMachine, rnn_multi_unequalength_input) { - for (bool useGpu : {false, true}) { - test("gserver/tests/sequence_rnn_multi_unequalength_inputs.py", - "gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py", - 1e-6, - useGpu); - } -} - -TEST(RecurrentGradientMachine, rnn_mixed_input) { - for (bool useGpu : {false, true}) { - test("gserver/tests/sequence_rnn_mixed_inputs.py", - "gserver/tests/sequence_rnn_matched_inputs.py", - 1e-6, - useGpu); - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - - if (paddle::version::isWithPyDataProvider()) { - if (!paddle::version::isWithGpu()) { - FLAGS_use_gpu = false; - } - initMain(argc, argv); - initPython(argc, argv); - return RUN_ALL_TESTS(); - } else { - return 0; - } -} diff --git a/paddle/gserver/tests/test_RecurrentLayer.cpp b/paddle/gserver/tests/test_RecurrentLayer.cpp deleted file mode 100644 index b54e37b7dbf8bffeb949f709e6a4f9ec86ea13c3..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_RecurrentLayer.cpp +++ /dev/null @@ -1,571 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/gserver/layers/Layer.h" - -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT -DECLARE_bool(use_gpu); -DECLARE_bool(rnn_use_batch); -DECLARE_int32(fixed_seq_length); - -void checkError(const Matrix& matrix1, const Matrix& matrix2) { - CHECK(matrix1.getHeight() == matrix2.getHeight()); - CHECK(matrix1.getWidth() == matrix2.getWidth()); -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - - int height = matrix1.getHeight(); - int width = matrix1.getWidth(); - const real* data1 = matrix1.getData(); - const real* data2 = matrix2.getData(); - int count = 0; - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - if (fabs(data1[i * width + j] - data2[i * width + j]) > err) { - count++; - } - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; -} - -void checkError(const CpuVector& vector1, const CpuVector& vector2) { - CHECK(vector1.getSize() == vector2.getSize()); -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - - int size = vector1.getSize(); - const real* data1 = vector1.getData(); - const real* data2 = vector2.getData(); - int count = 0; - for (int i = 0; i < size; i++) { - if (fabs(data1[i] - data2[i]) > err) { - count++; - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; -} - -LayerPtr creatDataLayer(string name, - size_t batchSize, - int layerSize, - bool useGpu) { - LayerConfig dataConfig; - dataConfig.set_name(name); - dataConfig.set_type("data"); - dataConfig.set_size(layerSize); - LayerPtr layer = LayerPtr(new DataLayer(dataConfig)); - - Argument data; - data.value = Matrix::create(batchSize, layer->getSize(), false, useGpu); - data.grad = Matrix::create(batchSize, layer->getSize(), false, useGpu); - data.value->randomizeUniform(); - data.value->add(-0.5); - data.value->sigmoid(*data.value); - data.grad->zeroMem(); - - generateSequenceStartPositions(batchSize, data.sequenceStartPositions); - - DataLayerPtr dataLayer = std::dynamic_pointer_cast(layer); - dataLayer->setData(data); - dataLayer->forward(PASS_GC); - - return layer; -} - -ParameterPtr creatParameter(string name, - int pid, - size_t paraSize, - bool useGpu) { - ParameterConfig paraConfig; - paraConfig.set_name(name); - paraConfig.set_size(paraSize); - - ParameterPtr parameter = - std::make_shared(paraConfig, useGpu, /*initialize */ false); - parameter->enableType(PARAMETER_VALUE); - parameter->enableType(PARAMETER_GRADIENT); - parameter->randomize(); - parameter->setID(pid); - - return parameter; -} - -ParameterPtr creatParameterBias(string name, - int pid, - size_t paraSize, - bool useGpu) { - ParameterConfig paraConfig; - paraConfig.set_name(name); - paraConfig.set_size(paraSize); - paraConfig.set_initial_std(1); - - ParameterPtr parameter = - std::make_shared(paraConfig, useGpu, /*initialize */ true); - parameter->randomize(); - parameter->setID(pid); - - return parameter; -} - -LayerPtr initRecurrentLayer(LayerConfig layerConfig, - size_t batchSize, - int layerSize, - bool useGpu) { - FLAGS_use_gpu = useGpu; - LayerMap layerMap; - ParameterMap parameterMap; - LayerPtr dataLayer = creatDataLayer("layer_0", batchSize, layerSize, useGpu); - layerMap[dataLayer->getName()] = dataLayer; - - ParameterPtr para = - creatParameter("para_0", 0, layerSize * layerSize, useGpu); - parameterMap[para->getName()] = para; - - layerConfig.add_inputs(); - LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); - input.set_input_layer_name("layer_0"); - input.set_input_parameter_name("para_0"); - LayerPtr testLayer = Layer::create(layerConfig); - layerMap[testLayer->getName()] = testLayer; - - testLayer->init(layerMap, parameterMap); - testLayer->setNeedGradient(true); - - return testLayer; -} - -void checkRecurrentLayer(LayerPtr testLayer) { - const VectorPtr& weightGrad = - (testLayer->getParameters()[0])->getBuf(PARAMETER_GRADIENT); - const MatrixPtr& inputGrad = testLayer->getPrev(0)->getOutputGrad(); - CpuVector seqPara(weightGrad->getSize()); - CpuVector batPara(weightGrad->getSize()); - CpuMatrix seqInputGrad(inputGrad->getHeight(), inputGrad->getWidth()); - CpuMatrix batInputGrad(inputGrad->getHeight(), inputGrad->getWidth()); - - CpuMatrix outputGrad(inputGrad->getHeight(), inputGrad->getWidth()); - outputGrad.randomizeUniform(); - - /* use sequence calculate */ - FLAGS_rnn_use_batch = false; - weightGrad->zero(); - inputGrad->zero(); - testLayer->forward(PASS_GC); - testLayer->getOutputGrad()->copyFrom(outputGrad); - testLayer->backward(); - seqPara.copyFrom(*weightGrad); - seqInputGrad.copyFrom(*inputGrad); - - /* use batch calculate */ - FLAGS_rnn_use_batch = true; - weightGrad->zero(); - inputGrad->zero(); - testLayer->forward(PASS_GC); - testLayer->getOutputGrad()->copyFrom(outputGrad); - testLayer->backward(); - batPara.copyFrom(*weightGrad); - batInputGrad.copyFrom(*inputGrad); - - /* check */ - checkError(seqInputGrad, batInputGrad); - checkError(seqPara, batPara); -} - -TEST(Layer, RecurrentLayer) { - LayerConfig layerConfig; - layerConfig.set_name("rnn"); - layerConfig.set_type("recurrent"); - layerConfig.set_active_type("tanh"); - for (auto layerSize : {1, 10, 64, 128, 256, 512}) { - for (auto batchSize : {1, 5, 20, 100, 128}) { - for (auto useGpu : {false, true}) { - for (auto reversed : {false, true}) { - LOG(INFO) << " layerSize=" << layerSize << " batchSize=" << batchSize - << " useGpu=" << useGpu << " reversed=" << reversed; - layerConfig.set_size(layerSize); - layerConfig.set_reversed(reversed); - LayerPtr testLayer = - initRecurrentLayer(layerConfig, batchSize, layerSize, useGpu); - checkRecurrentLayer(testLayer); - } - } - } - } -} - -#define protected public -#include "paddle/gserver/layers/GatedRecurrentLayer.h" -#include "paddle/gserver/layers/LstmLayer.h" -#include "paddle/gserver/layers/RecurrentLayer.h" -template -class TestRecurrentLayer { - public: - LayerConfig config_; - bool useGpu_; - bool useBatch_; - LayerPtr testLayer_; - LayerPtr dataLayer_; - ParameterPtr para_; - ParameterPtr bias_; - LayerMap layerMap_; - ParameterMap parameterMap_; - TestRecurrentLayer(const LayerConfig& config, - bool useGpu, - bool useBatch = false) - : config_(config), useGpu_(useGpu), useBatch_(useBatch) {} - void init(size_t batchSize) { - FLAGS_use_gpu = useGpu_; - testLayer_ = Layer::create(config_); - if (typeid(T) == typeid(GatedRecurrentLayer)) { - dataLayer_ = creatDataLayer(config_.mutable_inputs(0)->input_layer_name(), - batchSize, - config_.size() * 3, - useGpu_); - para_ = creatParameter(config_.mutable_inputs(0)->input_parameter_name(), - 0, - config_.size() * config_.size() * 3, - useGpu_); - bias_ = creatParameterBias( - config_.bias_parameter_name(), 1, config_.size() * 3, useGpu_); - } else if (typeid(T) == typeid(LstmLayer)) { - dataLayer_ = creatDataLayer(config_.mutable_inputs(0)->input_layer_name(), - batchSize, - config_.size() * 4, - useGpu_); - para_ = creatParameter(config_.mutable_inputs(0)->input_parameter_name(), - 0, - config_.size() * config_.size() * 4, - useGpu_); - bias_ = creatParameterBias( - config_.bias_parameter_name(), 1, config_.size() * 7, useGpu_); - } - layerMap_[dataLayer_->getName()] = dataLayer_; - parameterMap_[para_->getName()] = para_; - parameterMap_[bias_->getName()] = bias_; - - layerMap_[testLayer_->getName()] = testLayer_; - testLayer_->init(layerMap_, parameterMap_); - testLayer_->setNeedGradient(true); - (dynamic_cast(testLayer_.get()))->useBatch_ = useBatch_; - } - void forward() { - FLAGS_use_gpu = useGpu_; - testLayer_->forward(PASS_GC); - } - void backward() { - FLAGS_use_gpu = useGpu_; - testLayer_->backward(nullptr); - } -}; - -template -void checkRecurrentLayer(LayerConfig layerConfig, - size_t batchSize, - bool cpuBatch, - bool gpuBatch) { - TestRecurrentLayer testCpu(layerConfig, false, cpuBatch); - TestRecurrentLayer testGpu(layerConfig, true, gpuBatch); - testCpu.init(batchSize); - testGpu.init(batchSize); - auto checkError = []( - MatrixPtr cpu, MatrixPtr gpu, int numSequences, const char* str) { - CpuMatrix check(gpu->getHeight(), gpu->getWidth()); - check.copyFrom(*gpu); - int height = cpu->getHeight(); - int width = cpu->getWidth(); - const real* data1 = cpu->getData(); - const real* data2 = check.getData(); - int count = 0; - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - if (fabs(data1[i * width + j] - data2[i * width + j]) / numSequences > - 1e-4) { - count++; - } - } - } - EXPECT_EQ(count, 0) << "[" << str << "]" - << "There are " << count << " different element."; - }; - T* cpuLayer = dynamic_cast(testCpu.testLayer_.get()); - T* gpuLayer = dynamic_cast(testGpu.testLayer_.get()); - - Argument& cpuInput = testCpu.dataLayer_->getOutput(); - Argument& gpuInput = testGpu.dataLayer_->getOutput(); - gpuInput.resizeAndCopyFrom(cpuInput, true); - - const VectorPtr& cpuVec = testCpu.para_->getBuf(PARAMETER_VALUE); - const VectorPtr& gpuVec = testGpu.para_->getBuf(PARAMETER_VALUE); - gpuVec->copyFrom(*cpuVec); - - const VectorPtr& cpuBiasVec = testCpu.bias_->getBuf(PARAMETER_VALUE); - const VectorPtr& gpuBiasVec = testGpu.bias_->getBuf(PARAMETER_VALUE); - gpuBiasVec->copyFrom(*cpuBiasVec); - - /* check forward */ - testCpu.forward(); - testGpu.forward(); - - checkError( - cpuLayer->getOutputValue(), gpuLayer->getOutputValue(), 1, "outputValue"); - - /* check backward */ - cpuLayer->getOutputGrad()->randomizeUniform(); - gpuLayer->getOutputGrad()->copyFrom(*cpuLayer->getOutputGrad()); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - - testCpu.backward(); - testGpu.backward(); - - // check input grad - checkError(cpuInput.grad, gpuInput.grad, 1, "inputGrad"); - // check weight grad - int numSequences = cpuInput.getNumSequences(); - checkError(cpuLayer->weight_->getWGrad(), - gpuLayer->weight_->getWGrad(), - numSequences, - "weightGrad"); - // check bias grad - checkError(cpuLayer->bias_->getWGrad(), - gpuLayer->bias_->getWGrad(), - numSequences, - "biasGrad"); -} - -TEST(Layer, GatedRecurrentLayer) { - LayerConfig layerConfig; - layerConfig.set_type("gated_recurrent"); - layerConfig.set_active_type("sigmoid"); - layerConfig.set_active_gate_type("sigmoid"); - - layerConfig.add_inputs(); - LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); - input.set_input_layer_name("layer_0"); - input.set_input_parameter_name("para_0"); - layerConfig.set_bias_parameter_name("bias"); - - for (auto frameSize : {32, 64, 128, 256, 512}) { - for (auto batchSize : {1, 5, 100, 500}) { - for (auto reversed : {false, true}) { - for (auto cpuBatch : {false, true}) { - for (auto gpuBatch : {false, true}) { - LOG(INFO) << " batchSize=" << batchSize - << " frameSize=" << frameSize << " reversed=" << reversed - << " cpuBatch=" << cpuBatch << " gpuBatch=" << gpuBatch; - layerConfig.set_size(frameSize); - layerConfig.set_reversed(reversed); - checkRecurrentLayer( - layerConfig, batchSize, cpuBatch, gpuBatch); - } - } - } - } - } -} - -TEST(Layer, LstmLayer) { - LayerConfig layerConfig; - layerConfig.set_type("lstmemory"); - layerConfig.set_active_type("relu"); - layerConfig.set_active_state_type("tanh"); - layerConfig.set_active_gate_type("sigmoid"); - - layerConfig.add_inputs(); - LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); - input.set_input_layer_name("layer_0"); - input.set_input_parameter_name("para_0"); - layerConfig.set_bias_parameter_name("bias"); - - for (auto frameSize : {32, 64, 128, 256, 512}) { - for (auto batchSize : {1, 5, 100, 500}) { - for (auto reversed : {false, true}) { - for (auto cpuBatch : {false, true}) { - for (auto gpuBatch : {false, true}) { - LOG(INFO) << " batchSize=" << batchSize - << " frameSize=" << frameSize << " reversed=" << reversed - << " cpuBatch=" << cpuBatch << " gpuBatch=" << gpuBatch; - layerConfig.set_size(frameSize); - layerConfig.set_reversed(reversed); - checkRecurrentLayer( - layerConfig, batchSize, cpuBatch, gpuBatch); - } - } - } - } - } -} - -#ifdef PADDLE_WITH_MKLML - -#include "paddle/gserver/layers/MKLPackedRecurrentLayer.h" - -LayerPtr initMKLPackedLayer(LayerConfig layerConfig, - bool reversed, - int layerSize, - LayerPtr dataLayer, - ParameterPtr para, - ParameterPtr bias = nullptr) { - LayerMap layerMap; - ParameterMap parameterMap; - layerMap[dataLayer->getName()] = dataLayer; - parameterMap[para->getName()] = para; - if (bias) { - parameterMap[bias->getName()] = bias; - layerConfig.set_bias_parameter_name("bias_0"); - } - - layerConfig.set_size(layerSize); - layerConfig.set_reversed(reversed); - layerConfig.add_inputs(); - LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); - input.set_input_layer_name("layer_0"); - input.set_input_parameter_name("para_0"); - - LayerPtr testLayer = Layer::create(layerConfig); - layerMap[testLayer->getName()] = testLayer; - - testLayer->init(layerMap, parameterMap); - testLayer->setNeedGradient(true); - - return testLayer; -} - -void checkMKLPackedLayer(LayerConfig layerConfig1, - LayerConfig layerConfig2, - bool reversed, - int layerSize, - int batchSize, - bool useBatch1, - bool useBatch2) { - LayerPtr dataLayer; - ParameterPtr para, bias; - - if (layerConfig1.type() == "recurrent") { - dataLayer = creatDataLayer("layer_0", batchSize, layerSize, false); - para = creatParameter("para_0", 0, layerSize * layerSize, false); - bias = nullptr; - } else if (layerConfig1.type() == "gated_recurrent") { - dataLayer = creatDataLayer("layer_0", batchSize, layerSize * 3, false); - para = creatParameter("para_0", 0, layerSize * layerSize * 3, false); - bias = creatParameterBias("bias_0", 1, layerSize * 3, false); - } - - LayerPtr testLayer1 = initMKLPackedLayer( - layerConfig1, reversed, layerSize, dataLayer, para, bias); - LayerPtr testLayer2 = initMKLPackedLayer( - layerConfig2, reversed, layerSize, dataLayer, para, bias); - - const VectorPtr& weightGrad = - (testLayer1->getParameters()[0])->getBuf(PARAMETER_GRADIENT); - const MatrixPtr& inputGrad = testLayer1->getPrev(0)->getOutputGrad(); - CpuVector wgt_grad1(weightGrad->getSize()); - CpuVector wgt_grad2(weightGrad->getSize()); - CpuMatrix input_grad1(inputGrad->getHeight(), inputGrad->getWidth()); - CpuMatrix input_grad2(inputGrad->getHeight(), inputGrad->getWidth()); - - for (int i = 0; i < 2; i++) { - FLAGS_rnn_use_batch = useBatch1; - - testLayer1->forward(PASS_GC); - - FLAGS_rnn_use_batch = useBatch2; - testLayer2->forward(PASS_GC); - - testLayer1->getOutputGrad()->randomizeUniform(); - testLayer2->getOutputGrad()->copyFrom(*testLayer1->getOutputGrad()); - - weightGrad->zero(); - inputGrad->zero(); - FLAGS_rnn_use_batch = useBatch1; - testLayer1->backward(nullptr); - - wgt_grad1.copyFrom(*weightGrad); - input_grad1.copyFrom(*inputGrad); - - weightGrad->zero(); - inputGrad->zero(); - FLAGS_rnn_use_batch = useBatch2; - testLayer2->backward(nullptr); - - wgt_grad2.copyFrom(*weightGrad); - input_grad2.copyFrom(*inputGrad); - - checkError(*testLayer1->getOutputValue(), *testLayer2->getOutputValue()); - checkError(wgt_grad1, wgt_grad2); - checkError(input_grad1, input_grad2); - } -} - -TEST(MKLPackedLayer, RecurrentLayer) { - LayerConfig layerConfig1; - LayerConfig layerConfig2; - - layerConfig1.set_name("paddle-rnn"); - layerConfig1.set_type("recurrent"); - layerConfig1.set_active_type("relu"); - - layerConfig2.set_name("mkl-packed-rnn"); - layerConfig2.set_type("mkl_packed_recurrent"); - layerConfig2.set_active_type("relu"); - - FLAGS_use_gpu = false; - - for (auto layerSize : {32, 64, 128, 256, 512}) { - for (auto batchSize : {1, 5, 100, 500}) { - for (auto reversed : {true, false}) { - for (auto paddle_use_batch : {true, false}) { - for (auto MKLPacked_use_batch : {true, false}) { - LOG(INFO) << " layerSize=" << layerSize - << " batchSize=" << batchSize << " reversed=" << reversed - << " paddle_use_batch=" << paddle_use_batch - << " MKLPacked_use_batch=" << MKLPacked_use_batch; - - checkMKLPackedLayer(layerConfig1, - layerConfig2, - reversed, - layerSize, - batchSize, - paddle_use_batch, - MKLPacked_use_batch); - } - } - } - } - } -} -#endif - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - if (!version::isWithGpu()) { - testing::GTEST_FLAG(filter) = "-Layer.*"; - } - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_SelectiveFCLayer.cpp b/paddle/gserver/tests/test_SelectiveFCLayer.cpp deleted file mode 100644 index 583e3bc545a3b5eb158490a8ccc5ea7060c7c6ab..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_SelectiveFCLayer.cpp +++ /dev/null @@ -1,471 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/gserver/layers/FullyConnectedLayer.h" -#include "paddle/gserver/layers/Layer.h" -#include "paddle/gserver/layers/SelectiveFullyConnectedLayer.h" -#include "paddle/math/CpuSparseMatrix.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); -DECLARE_int32(num_passes); -DECLARE_string(config); -DECLARE_string(init_model_path); -DECLARE_string(config_args); - -size_t fcLayerWidth = 1024; - -struct ComData { - vector outArgs; - vector parameters; -}; - -int randint(int* data, size_t int_max, size_t size) { - srand((size_t)(time(NULL))); - if (int_max < size) { - return -1; - } - size_t count = 0; - std::map tmp; - int this_int = 0; - - while (count < size) { - this_int = std::rand() % int_max; // NOLINT - if (tmp.find(this_int) == tmp.end()) { - tmp[this_int] = 0; - count += 1; - } - } - - if (tmp.size() != size) { - return -1; - } - count = 0; - for (auto itr = tmp.begin(); itr != tmp.end(); ++itr) { - data[count] = itr->first; - count += 1; - } - return 0; -} - -void calcOutput(ComData& comData, - const string configFile, - const string configArgs, - bool useGpu) { - FLAGS_config = configFile; - FLAGS_config_args = configArgs; - FLAGS_use_gpu = useGpu; - FLAGS_init_model_path = "gserver/tests/SelectiveFcTest/model"; - *ThreadLocalRand::getSeed() = 0; - srand(0); - - Trainer trainer; - trainer.init(TrainerConfigHelper::createFromFlags(), false); - - comData.parameters = trainer.getGradientMachine()->getParameters(); - - auto dataProvider = trainer.getDataProvider(); - int32_t batchSize = trainer.getConfig().opt_config().batch_size(); - DataBatch dataBatch; - dataProvider->setSkipShuffle(); - dataProvider->reset(); - dataProvider->getNextBatch(batchSize, &dataBatch); - CHECK(dataBatch.getSize()) << "No data from data provider"; - - vector& inArgs = dataBatch.getStreams(); - trainer.getGradientMachine()->start(trainer.getConfig(), nullptr); - trainer.getGradientMachine()->forwardBackward( - inArgs, &comData.outArgs, PASS_TRAIN); - trainer.getGradientMachine()->finish(); -} - -void checkMatrix(real* A, real* B, size_t matSize) { -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - int diffNum = 0; - for (size_t i = 0; i < matSize; ++i) { - if (std::isinf(A[i]) || std::isnan(A[i]) || std::isinf(B[i]) || - std::isnan(B[i])) { - } else if (fabs(A[i] - B[i]) > err) { - diffNum++; - } - } - EXPECT_EQ(0, diffNum); -} - -void checkTranspose(real* matrix, - real* transpose, - size_t width, - size_t matSize) { -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - size_t height = matSize / width; - int diffNum = 0; - size_t rowId = 0; - size_t colId = 0; - for (size_t i = 0; i < matSize; ++i) { - if (i % width == 0 && i) { - rowId++; - } - colId = i % width; - if (fabs(matrix[i] - transpose[colId * height + rowId]) > err) { - diffNum++; - LOG(INFO) << i << " diff : " << matrix[i] << "\t" - << transpose[colId * height + rowId]; - } - } - EXPECT_EQ(0, diffNum); -} - -void compareOutput(ComData& fcData, ComData& selFcData) { - vector outArgsFc = fcData.outArgs; - vector outArgsSelfc = selFcData.outArgs; - - // check cost - LOG(INFO) << "Check cost"; - CpuMatrix fcCost(outArgsFc[0].value->getHeight(), - outArgsFc[0].value->getWidth()); - CpuMatrix selfcCost(outArgsSelfc[0].value->getHeight(), - outArgsSelfc[0].value->getWidth()); - fcCost.copyFrom(*outArgsFc[0].value); - selfcCost.copyFrom(*outArgsSelfc[0].value); - checkMatrix(fcCost.getData(), selfcCost.getData(), fcCost.getElementCnt()); - - // check selective fc output and fc output - LOG(INFO) << "Compare output of SelectiveFullyConectedLayer " - << "with FullyConectedLayer"; - CpuMatrix fcOut(outArgsFc[1].value->getHeight(), - outArgsFc[1].value->getWidth()); - CpuMatrix selfcOut(outArgsSelfc[1].value->getHeight(), - outArgsSelfc[1].value->getWidth()); - - fcOut.copyFrom(*outArgsFc[1].value); - selfcOut.copyFrom(*outArgsSelfc[1].value); - checkMatrix(fcOut.getData(), selfcOut.getData(), fcOut.getElementCnt()); - - // check gradient math - vector& fcParam = fcData.parameters; - vector& selfcParam = selFcData.parameters; - for (size_t i = 0; i < fcParam.size(); ++i) { - ParameterPtr p1, p2; - p1 = fcParam[i]; - p2 = selfcParam[i]; - - string paramName = p1->getName(); - LOG(INFO) << "check parameter : " << paramName; - - // check parameter value - CpuVector paraValue1(p1->getSize()); - CpuVector paraValue2(p2->getSize()); - paraValue1.copyFrom(*p1->getBuf(PARAMETER_VALUE)); - paraValue2.copyFrom(*p2->getBuf(PARAMETER_VALUE)); - - // check gradient - CpuVector paraGrad1(*p1->getBuf(PARAMETER_GRADIENT)); - CpuVector paraGrad2(*p2->getBuf(PARAMETER_GRADIENT)); - if (paramName == "rand_fc_param.bias") { - checkMatrix( - paraValue1.getData(), paraValue2.getData(), paraValue1.getSize()); - checkMatrix( - paraGrad1.getData(), paraGrad2.getData(), paraGrad1.getSize()); - } else { - checkTranspose(paraValue1.getData(), - paraValue2.getData(), - fcLayerWidth, - paraValue1.getSize()); - checkTranspose(paraGrad1.getData(), - paraGrad2.getData(), - fcLayerWidth, - paraGrad1.getSize()); - } - } -} - -void compareSparseMulOutput( - real* fcOutput, - real* selOutput, - size_t nnz, - const std::shared_ptr>>& selCols) { -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - size_t nnzCount = - std::accumulate(selCols->begin(), - selCols->end(), - 0UL, - [](size_t a, const std::pair& arr) { - return a + arr.second; - }); - EXPECT_EQ(nnz, nnzCount); - - size_t sampleNum = selCols->size(); - int diffNum = 0; - size_t count = 0; - for (size_t i = 0; i < sampleNum; ++i) { - for (size_t j = 0; j < (*selCols)[i].second; ++j) { - size_t selIdx = (*selCols)[i].first[j]; - if (fabs(fcOutput[i * fcLayerWidth + selIdx] - selOutput[count]) > err) { - diffNum++; - LOG(INFO) << count << " diff : " << fcOutput[i * fcLayerWidth + selIdx] - << "\t" << selOutput[count]; - } - count++; - } - } - EXPECT_EQ(0, diffNum); -} - -LayerPtr creatDataLayer(string name, - size_t batchSize, - size_t layerSize, - std::vector& values, - bool useGpu) { - LayerConfig dataConfig; - dataConfig.set_name(name); - dataConfig.set_type("data"); - dataConfig.set_size(layerSize); - LayerPtr layer = LayerPtr(new DataLayer(dataConfig)); - - Argument data; - data.value = Matrix::create(batchSize, layerSize, false, useGpu); - data.value->copyFrom(values.data(), batchSize * layerSize); - - DataLayerPtr dataLayer = std::dynamic_pointer_cast(layer); - dataLayer->setData(data); - dataLayer->forward(PASS_TEST); - return layer; -} - -ParameterPtr creatParameter( - string name, int pid, size_t paraSize, string paramFile, bool useGpu) { - ParameterConfig paraConfig; - paraConfig.set_name(name); - paraConfig.set_size(paraSize); - - ParameterPtr parameter = - std::make_shared(paraConfig, useGpu, /*initialize */ false); - parameter->enableType(PARAMETER_VALUE); - parameter->randomize(); - parameter->setID(pid); - parameter->load(paramFile); - return parameter; -} - -LayerPtr initFcLayer(LayerPtr dataLayer, - LayerConfig layerConfig, - int dataLayerSize, - int fcLayerSize, - string paraName, - string paraFile, - bool useGpu) { - LayerMap layerMap; - ParameterMap parameterMap; - - layerMap[dataLayer->getName()] = dataLayer; - ParameterPtr para = creatParameter( - paraName, 0, dataLayerSize * fcLayerSize, paraFile, useGpu); - parameterMap[para->getName()] = para; - - layerConfig.add_inputs(); - LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); - input.set_input_layer_name(dataLayer->getName()); - input.set_input_parameter_name(paraName); - - LayerPtr testLayer = Layer::create(layerConfig); - layerMap[testLayer->getName()] = testLayer; - - testLayer->setNeedGradient(false); - testLayer->init(layerMap, parameterMap); - return testLayer; -} - -#ifndef PADDLE_TYPE_DOUBLE -// The parameter file used in fc.conf and selective_fc.conf is float -TEST(Layer, SelectiveFcLayer_train_dense_mul) { - const string& fcConfig = "gserver/tests/SelectiveFcTest/conf/fc.conf"; - const string& fcConfigArgs = - "filelist=gserver/tests/SelectiveFcTest/dense_mul_list"; - const string& selFcConfig = - "gserver/tests/SelectiveFcTest/conf/selective_fc.conf"; - const string& selConfigArgs = - "filelist=gserver/tests/SelectiveFcTest/dense_mul_list"; - - for (auto useGpu : {false, true}) { -#ifndef PADDLE_WITH_CUDA - if (useGpu) { - break; - } -#endif - LOG(INFO) << "FullyConnectedLayer forwardBackward()"; - ComData fcData; - calcOutput(fcData, fcConfig, fcConfigArgs, useGpu); - - LOG(INFO) << "SelectiveFullyConnectedLayer forwardBackward()"; - ComData selFcData; - calcOutput(selFcData, selFcConfig, selConfigArgs, useGpu); - compareOutput(fcData, selFcData); - } -} -#endif // PADDLE_TYPE_DOUBLE - -void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config, - bool useGpu) { - FLAGS_use_gpu = useGpu; - size_t batchSize = 100; - size_t dataLayerSize = 512; - std::vector values(batchSize * dataLayerSize); - for (size_t j = 0; j < batchSize * dataLayerSize; ++j) { - values[j] = std::rand() / real(RAND_MAX); - } - LayerPtr dataLayer = - creatDataLayer("data", batchSize, dataLayerSize, values, useGpu); - - const string& selfcParaFile = - "gserver/tests/SelectiveFcTest/model/rand_fc_param.w.transpose"; - const string& selfcParaName = "rand_fc_param.w.transpose"; - - std::shared_ptr selfcLayer = - std::dynamic_pointer_cast( - initFcLayer(dataLayer, - config, - dataLayerSize, - fcLayerWidth, - selfcParaName, - selfcParaFile, - useGpu)); - - // create selected columns - std::shared_ptr>> selCols( - new std::vector>(batchSize)); - size_t maxNNZ = 30; - srand((size_t)(time(NULL))); - int total = 0; - while (total == 0) { - for (size_t i = 0; i < batchSize; ++i) { - size_t num = std::rand() % maxNNZ; - int* data = new int[num]; - randint(data, fcLayerWidth, num); - (*selCols)[i] = std::make_pair(data, num); - total += num; - } - } - selfcLayer->fillSelectiveData(selCols); - selfcLayer->forward(PASS_TEST); - - MatrixPtr outMatSelfc = selfcLayer->getOutputValue(); - CpuSparseMatrixPtr cpuOutMatSelfc( - new CpuSparseMatrix(outMatSelfc->getHeight(), - outMatSelfc->getWidth(), - outMatSelfc->getElementCnt())); - cpuOutMatSelfc->copyFrom(*outMatSelfc, HPPL_STREAM_DEFAULT); -#ifdef PADDLE_WITH_CUDA - if (useGpu) { - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - } -#endif - real* outValueSelfc = cpuOutMatSelfc->getValue(); - size_t nnz = cpuOutMatSelfc->getElementCnt(); - - const string& fcParaFile = - "gserver/tests/SelectiveFcTest/model/rand_fc_param.w"; - const string& fcParaName = "rand_fc_param.w"; - LayerConfig fcLayerConfig; - fcLayerConfig.set_name("fc_layer"); - fcLayerConfig.set_type("fc"); - fcLayerConfig.set_active_type("linear"); - fcLayerConfig.set_size(fcLayerWidth); - - LayerPtr fcLayer = initFcLayer(dataLayer, - fcLayerConfig, - dataLayerSize, - fcLayerWidth, - fcParaName, - fcParaFile, - useGpu); - fcLayer->forward(PASS_TEST); - - MatrixPtr outMatFc = fcLayer->getOutputValue(); - MatrixPtr cpuOutMatFc( - new CpuMatrix(outMatFc->getHeight(), outMatFc->getWidth())); - cpuOutMatFc->copyFrom(*outMatFc, HPPL_STREAM_DEFAULT); -#ifdef PADDLE_WITH_CUDA - if (useGpu) { - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - } -#endif - real* outValueFc = cpuOutMatFc->getData(); - - compareSparseMulOutput(outValueFc, outValueSelfc, nnz, selCols); - for (size_t i = 0; i < batchSize; ++i) { - delete[](*selCols)[i].first; - } -} - -#ifndef PADDLE_TYPE_DOUBLE -// The parameter file used in testSelectiveFcLayerTrainSparseMul is float -TEST(Layer, SelectiveFcLayer_train_sparse_mul) { - LayerConfig selLayerConfig; - selLayerConfig.set_name("sel_fc"); - selLayerConfig.set_type("selective_fc"); - selLayerConfig.set_active_type("linear"); - selLayerConfig.set_has_selected_colums(false); - selLayerConfig.set_selective_fc_pass_generation(true); - selLayerConfig.set_size(fcLayerWidth); - - testSelectiveFcLayerTrainSparseMul(selLayerConfig, false); -#ifdef PADDLE_WITH_CUDA - testSelectiveFcLayerTrainSparseMul(selLayerConfig, true); -#endif -} -#endif // PADDLE_TYPE_DOUBLE - -// TODO(dangqingqing) test multi threads after support in matrix -// TEST(Layer, SelectiveFcLayer_train_sparse_mul_parallel) { -// LayerConfig selLayerConfig; -// selLayerConfig.set_name("sel_fc"); -// selLayerConfig.set_type("selective_fc"); -// selLayerConfig.set_active_type("linear"); -// selLayerConfig.set_has_selected_colums(false); -// selLayerConfig.set_selective_fc_pass_generation(true); -// selLayerConfig.set_selective_fc_parallel_plain_mul_thread_num(10); -// selLayerConfig.set_selective_fc_full_mul_ratio(1000); -// selLayerConfig.set_size(fcLayerWidth); -// SelectiveFcLayer_test(selLayerConfig, false); -// } - -int main(int argc, char** argv) { - paddle::initMain(argc, argv); - testing::InitGoogleTest(&argc, argv); - initPython(argc, argv); - int ret = RUN_ALL_TESTS(); - return ret; -} diff --git a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp deleted file mode 100644 index 406ca63b6ee030a0882e05294d8d355d84531385..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp +++ /dev/null @@ -1,224 +0,0 @@ -/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/DataLayer.h" - -#include "LayerGradUtil.h" -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_int32(gpu_id); -DECLARE_bool(thread_local_rand_use_global_seed); - -const int MAX_SEQ_NUM = 17; -const int MAX_SEQ_LEN = 23; -const int MAX_BEAM_SIZE = 13; - -const size_t SEED = (size_t)(time(NULL)); - -vector randSampling(real range, int n) { - CHECK_GE(range, n); - vector num(range); - iota(begin(num), end(num), 0.); - if (range == n) return num; - - random_shuffle(begin(num), end(num)); - num.resize(n); - sort(begin(num), end(num)); - return num; -} - -void genSeqInfo(vector& seqStartPos, vector& subSeqStartPos) { - seqStartPos.resize(1, 0); - subSeqStartPos.resize(1, 0); - - srand(SEED); - int seqNum = 1 + (rand() % MAX_SEQ_NUM); - for (int i = 0; i < seqNum; ++i) { - int subSeqNum = 1 + (rand() % MAX_SEQ_NUM); - for (int j = 0; j < subSeqNum; ++j) - subSeqStartPos.push_back(subSeqStartPos.back() + - (1 + (rand() % MAX_SEQ_LEN))); - seqStartPos.push_back(subSeqStartPos.back()); - } -} - -/* - generate start indices according to sequence start positions. - */ -void genStarts(vector& seqStartPos, - vector>& starts, - size_t beamSize) { - starts.clear(); - starts.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); - - for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { - int seqLen = seqStartPos[i + 1] - seqStartPos[i]; - vector randStarts = - randSampling(seqLen, min(seqLen, static_cast(beamSize))); - copy(begin(randStarts), end(randStarts), begin(starts[i])); - } -} - -/* - generate end indices according to sequence start positions and start indices. - */ -void genEnds(vector& seqStartPos, - vector>& starts, - vector>& ends, - size_t beamSize) { - CHECK_EQ(seqStartPos.size() - 1, starts.size()); - ends.clear(); - ends.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); - - for (size_t i = 0; i < starts.size(); ++i) { - for (size_t j = 0; j < starts[i].size(); ++j) { - int seqLen = seqStartPos[i + 1] - seqStartPos[i]; - CHECK_GE(seqLen - 1, starts[i][j]); - if (starts[i][j] == -1.) break; - if (starts[i][j] == (seqLen - 1)) { - ends[i][j] = starts[i][j]; - } else { - ends[i][j] = starts[i][j] + randSampling(seqLen - starts[i][j], 1)[0]; - } - } - } -} - -void genTestData(vector& seqStartPos, - vector& subSeqStartPos, - vector>& starts, - vector>& ends, - bool hasSubseq) { - size_t beamSize = 1 + (rand() % MAX_BEAM_SIZE); - genSeqInfo(seqStartPos, subSeqStartPos); - - genStarts(hasSubseq ? subSeqStartPos : seqStartPos, starts, beamSize); - genEnds(hasSubseq ? subSeqStartPos : seqStartPos, starts, ends, beamSize); -} - -template -void flatten2dVector(vector>& inVec, vector& outVec) { - size_t totalSize{0}; - for (auto const& items : inVec) totalSize += items.size(); - outVec.reserve(totalSize); - - for (auto& items : inVec) - move(items.begin(), items.end(), back_inserter(outVec)); -} - -void testSeqSliceLayer(bool hasSubseq, - bool useGpu, - vector& seqStartPos, - vector& subSeqStartPos, - vector>& starts, - vector>& ends) { - // layer size is not crutial for this layer, - // so here use a small layer size in the unittest. - const size_t layerSize{4}; - TestConfig config; - config.layerConfig.set_type("seq_slice"); - config.layerConfig.set_size(layerSize); - - // add the first input - MatrixPtr seqInputPtr = - Matrix::create(hasSubseq ? subSeqStartPos.back() : seqStartPos.back(), - layerSize, - false, - false); - seqInputPtr->randomizeUniform(); - - if (hasSubseq) { - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, - "seq_input", - seqInputPtr, - seqStartPos, - subSeqStartPos}); - } else { - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, "seq_input", seqInputPtr, seqStartPos}); - } - config.layerConfig.add_inputs(); - - // add start indices - if (starts.size()) { - vector startsToVec; - flatten2dVector(starts, startsToVec); - - MatrixPtr startMatrixPtr = - Matrix::create(starts.size(), starts[0].size(), false, false); - startMatrixPtr->copyFrom(startsToVec.data(), startsToVec.size()); - - config.inputDefs.push_back( - {INPUT_SELF_DEFINE_DATA, "starts", startMatrixPtr}); - config.layerConfig.add_inputs(); - config.layerConfig.set_select_first(true); - } - - // add end indices - if (ends.size()) { - vector endsToVec; - flatten2dVector(ends, endsToVec); - - MatrixPtr endMatrixPtr = - Matrix::create(ends.size(), ends[0].size(), false, false); - endMatrixPtr->copyFrom(endsToVec.data(), endsToVec.size()); - - config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "ends", endMatrixPtr}); - config.layerConfig.add_inputs(); - config.layerConfig.set_select_first(false); - } - - testLayerGrad(config, "seq_slice", /*batchSize*/ 100, false, useGpu, false); -} - -TEST(Layer, SeqSliceLayer) { - vector seqStartPos; - vector subSeqStartPos; - vector> starts; - vector> ends; - - std::vector mode = {false}; -#ifdef PADDLE_WITH_CUDA - mode.push_back(true); -#endif - genSeqInfo(seqStartPos, subSeqStartPos); - for (bool hasSubseq : {true, false}) { - LOG(INFO) << "hasSubSeq : " << hasSubseq; - genTestData(seqStartPos, subSeqStartPos, starts, ends, hasSubseq); - for (bool useGpu : mode) { - vector> tmp; - testSeqSliceLayer( - hasSubseq, useGpu, seqStartPos, subSeqStartPos, tmp, ends); - testSeqSliceLayer( - hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, tmp); - testSeqSliceLayer( - hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, ends); - } - } -} - -int main(int argc, char** argv) { - initMain(argc, argv); - hl_start(); - hl_init(FLAGS_gpu_id); - FLAGS_thread_local_rand_use_global_seed = true; - srand(1); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/gserver/tests/test_Upsample.cpp b/paddle/gserver/tests/test_Upsample.cpp deleted file mode 100644 index 39b902fcc75e71007f855e4e258e54ed8d40f16b..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_Upsample.cpp +++ /dev/null @@ -1,153 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include - -#include "LayerGradUtil.h" -#include "paddle/math/MathUtils.h" -#include "paddle/testing/TestUtil.h" - -void setPoolConfig(paddle::TestConfig* config, - paddle::PoolConfig* pool, - const string& poolType) { - (*config).biasSize = 0; - (*config).layerConfig.set_type("pool"); - (*config).layerConfig.set_num_filters(1); - - int kw = 2, kh = 2; - int pw = 0, ph = 0; - int sw = 2, sh = 2; - pool->set_pool_type(poolType); - pool->set_channels(2); - pool->set_size_x(kw); - pool->set_size_y(kh); - pool->set_start(0); - pool->set_padding(pw); - pool->set_padding_y(ph); - pool->set_stride(sw); - pool->set_stride_y(sh); - - int ow = - paddle::outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); - int oh = - paddle::outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); - pool->set_output_x(ow); - pool->set_output_y(oh); -} - -paddle::LayerPtr doOneUpsampleTest(const paddle::MatrixPtr& inputMat, - const string& poolType, - bool use_gpu, - real* tempGradData) { - /* prepare maxPoolWithMaskLayer */ - paddle::TestConfig config; - config.inputDefs.push_back({paddle::INPUT_DATA, "layer_0", 128, 0}); - paddle::LayerInputConfig* input = config.layerConfig.add_inputs(); - paddle::PoolConfig* pool = input->mutable_pool_conf(); - - pool->set_img_size(8); - pool->set_img_size_y(8); - setPoolConfig(&config, pool, "max-pool-with-mask"); - config.layerConfig.set_size(pool->output_x() * pool->output_y() * - pool->channels()); - - config.layerConfig.set_name("MaxPoolWithMask"); - - std::vector dataLayers; - paddle::LayerMap layerMap; - vector datas; - - initDataLayer(config, - &dataLayers, - &datas, - &layerMap, - "MaxPoolWithMask", - 1, - false, - use_gpu); - - dataLayers[0]->getOutputValue()->copyFrom(*inputMat); - - FLAGS_use_gpu = use_gpu; - std::vector parameters; - paddle::LayerPtr maxPoolingWithMaskOutputLayer; - initTestLayer(config, &layerMap, ¶meters, &maxPoolingWithMaskOutputLayer); - maxPoolingWithMaskOutputLayer->forward(paddle::PASS_GC); - - /* prepare the upsample layer */ - paddle::LayerConfig upsampleLayerConfig; - upsampleLayerConfig.set_type("upsample"); - paddle::LayerInputConfig* input1 = upsampleLayerConfig.add_inputs(); - upsampleLayerConfig.add_inputs(); - - paddle::UpsampleConfig* upsampleConfig = input1->mutable_upsample_conf(); - upsampleConfig->set_scale(2); - paddle::ImageConfig* imageConfig = upsampleConfig->mutable_image_conf(); - imageConfig->set_channels(2); - imageConfig->set_img_size(4); - imageConfig->set_img_size_y(4); - upsampleLayerConfig.set_size(2 * 8 * 8); - upsampleLayerConfig.set_name("upsample"); - - for (size_t i = 0; i < 2; i++) { - paddle::LayerInputConfig& inputTemp = - *(upsampleLayerConfig.mutable_inputs(i)); - inputTemp.set_input_layer_name("MaxPoolWithMask"); - } - - paddle::LayerPtr upsampleLayer; - paddle::ParameterMap parameterMap; - upsampleLayer = paddle::Layer::create(upsampleLayerConfig); - layerMap[upsampleLayerConfig.name()] = upsampleLayer; - upsampleLayer->init(layerMap, parameterMap); - upsampleLayer->setNeedGradient(true); - upsampleLayer->forward(paddle::PASS_GC); - upsampleLayer->getOutputGrad()->copyFrom(tempGradData, 128); - upsampleLayer->backward(); - - return upsampleLayer; -} - -TEST(Layer, maxPoolingWithMaskOutputLayerFwd) { - bool useGpu = false; - paddle::MatrixPtr inputMat; - paddle::MatrixPtr inputGPUMat; - paddle::MatrixPtr tempGradMat; - - inputMat = paddle::Matrix::create(1, 128, false, useGpu); - inputMat->randomizeUniform(); - - tempGradMat = paddle::Matrix::create(1, 128, false, useGpu); - tempGradMat->randomizeUniform(); - real* tempGradData = tempGradMat->getData(); - - paddle::LayerPtr upsampleLayerCPU = - doOneUpsampleTest(inputMat, "max-pool-with-mask", useGpu, tempGradData); - -#ifdef PADDLE_WITH_CUDA - useGpu = true; - real* data = inputMat->getData(); - inputGPUMat = paddle::Matrix::create(1, 128, false, useGpu); - inputGPUMat->copyFrom(data, 128); - paddle::LayerPtr upsampleLayerGPU = doOneUpsampleTest( - inputGPUMat, "max-pool-with-mask", useGpu, tempGradData); - paddle::checkMatrixEqual(upsampleLayerCPU->getOutput("").value, - upsampleLayerGPU->getOutput("").value); - - paddle::checkMatrixEqual(upsampleLayerCPU->getPrev(0)->getOutputGrad(), - upsampleLayerGPU->getPrev(0)->getOutputGrad()); -#endif -} diff --git a/paddle/gserver/tests/test_WarpCTCLayer.cpp b/paddle/gserver/tests/test_WarpCTCLayer.cpp deleted file mode 100644 index f2299d7da2a51e4015793ae531af002aed1f6b2f..0000000000000000000000000000000000000000 --- a/paddle/gserver/tests/test_WarpCTCLayer.cpp +++ /dev/null @@ -1,244 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "ModelConfig.pb.h" -#include "paddle/gserver/layers/CTCLayer.h" -#include "paddle/gserver/layers/DataLayer.h" -#include "paddle/gserver/layers/Layer.h" -#include "paddle/gserver/layers/WarpCTCLayer.h" - -#include "paddle/testing/TestUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_bool(use_gpu); - -const real* getData(const Matrix& matrix) { - if (matrix.useGpu()) { - MatrixPtr cpuMatrix = Matrix::create( - matrix.getHeight(), matrix.getWidth(), matrix.isTransposed(), false); - cpuMatrix->copyFrom(matrix); - return cpuMatrix->getData(); - } else { - return matrix.getData(); - } -} - -int checkError(const Matrix& matrix1, const Matrix& matrix2) { - CHECK_EQ(matrix1.getHeight(), matrix2.getHeight()); - CHECK_EQ(matrix1.getWidth(), matrix2.getWidth()); - CHECK_EQ(matrix1.isTransposed(), matrix2.isTransposed()); -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - - int height = matrix1.getHeight(); - int width = matrix1.getWidth(); - - const real* data1 = getData(matrix1); - const real* data2 = getData(matrix2); - int count = 0; - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - if (fabs(data1[i * width + j] - data2[i * width + j]) > err) { - count++; - } - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; - return count; -} - -void initArgument(size_t batchSize, - int layerSize, - bool useGpu, - Argument& data) { - data.value = Matrix::create(batchSize, layerSize, false, useGpu); - data.grad = Matrix::create(batchSize, layerSize, false, useGpu); - data.value->randomizeUniform(); - data.value->add(-0.5); - data.grad->zeroMem(); - - generateSequenceStartPositions(batchSize, data.sequenceStartPositions); -} - -LayerPtr createDataLayer( - string name, size_t batchSize, int layerSize, bool useGpu, Argument& data) { - LayerConfig layerConfig; - layerConfig.set_name(name); - layerConfig.set_type("data"); - layerConfig.set_size(layerSize); - LayerPtr layer = LayerPtr(new DataLayer(layerConfig)); - - DataLayerPtr dataLayer = std::dynamic_pointer_cast(layer); - dataLayer->setData(data); - dataLayer->forward(PASS_GC); - - return layer; -} - -LayerPtr createLabelLayer(string name, - size_t batchSize, - size_t numClasses, - bool useGpu) { - LayerConfig layerConfig; - layerConfig.set_name(name); - layerConfig.set_type("data"); - layerConfig.set_size(1); - LayerPtr layer = LayerPtr(new DataLayer(layerConfig)); - - Argument data; - data.ids = IVector::create(batchSize, useGpu); - data.ids->rand(numClasses - 1); - - generateSequenceStartPositions(batchSize, data.sequenceStartPositions); - - DataLayerPtr labelLayer = std::dynamic_pointer_cast(layer); - labelLayer->setData(data); - labelLayer->forward(PASS_GC); - - return layer; -} - -LayerPtr createCTCLayer(string name, - size_t numClasses, - bool useGpu, - bool normByTimes, - LayerPtr dataLayer, - LayerPtr labelLayer) { - LayerMap layerMap; - layerMap[dataLayer->getName()] = dataLayer; - layerMap[labelLayer->getName()] = labelLayer; - - ParameterMap parameterMap; - - LayerConfig layerConfig; - layerConfig.set_name(name); - layerConfig.set_type("ctc"); - layerConfig.set_size(numClasses); - layerConfig.set_norm_by_times(normByTimes); - - layerConfig.add_inputs(); - LayerInputConfig& input0 = *(layerConfig.mutable_inputs(0)); - input0.set_input_layer_name(dataLayer->getName()); - - layerConfig.add_inputs(); - LayerInputConfig& input1 = *(layerConfig.mutable_inputs(1)); - input1.set_input_layer_name(labelLayer->getName()); - - LayerPtr layer = LayerPtr(new CTCLayer(layerConfig)); - layerMap[layer->getName()] = layer; - layer->init(layerMap, parameterMap); - - ActivationFunction* softmaxActivation = ActivationFunction::create("softmax"); - - softmaxActivation->forward(dataLayer->getOutput()).check(); - layer->forward(PASS_GC); - - layer->backward(); - softmaxActivation->backward(dataLayer->getOutput()).check(); - - return layer; -} - -LayerPtr createWarpCTCLayer(string name, - size_t numClasses, - bool useGpu, - bool normByTimes, - LayerPtr dataLayer, - LayerPtr labelLayer) { - LayerMap layerMap; - layerMap[dataLayer->getName()] = dataLayer; - layerMap[labelLayer->getName()] = labelLayer; - - ParameterMap parameterMap; - - LayerConfig layerConfig; - layerConfig.set_name(name); - layerConfig.set_type("warp_ctc"); - layerConfig.set_size(numClasses); - layerConfig.set_blank(numClasses - 1); - layerConfig.set_norm_by_times(normByTimes); - - layerConfig.add_inputs(); - LayerInputConfig& input0 = *(layerConfig.mutable_inputs(0)); - input0.set_input_layer_name(dataLayer->getName()); - - layerConfig.add_inputs(); - LayerInputConfig& input1 = *(layerConfig.mutable_inputs(1)); - input1.set_input_layer_name(labelLayer->getName()); - - LayerPtr layer = LayerPtr(new WarpCTCLayer(layerConfig)); - layerMap[layer->getName()] = layer; - layer->init(layerMap, parameterMap); - - layer->forward(PASS_GC); - layer->backward(); - - return layer; -} - -TEST(Layer, WarpCTCLayer) { - for (auto layerSize : {10, 64}) { - for (auto batchSize : {1, 10, 32}) { - for (auto normByTimes : {false, true}) { - for (auto useGpu : {false, true}) { -#ifndef PADDLE_WITH_CUDA - if (useGpu) continue; -#endif - LOG(INFO) << "layerSize=" << layerSize << " batchSize=" << batchSize - << " normByTimes = " << normByTimes << " useGpu=" << useGpu; - - FLAGS_use_gpu = useGpu; - - Argument data0; - initArgument(batchSize, layerSize, useGpu, data0); - - Argument data1; - data1.resizeAndCopyFrom(data0); - - LayerPtr dataLayer0 = - createDataLayer("data", batchSize, layerSize, useGpu, data0); - LayerPtr dataLayer1 = - createDataLayer("data", batchSize, layerSize, useGpu, data1); - - LayerPtr labelLayer = - createLabelLayer("label", batchSize, layerSize, useGpu); - - LayerPtr warpctcLayer = createWarpCTCLayer( - "cost", layerSize, useGpu, normByTimes, dataLayer0, labelLayer); - LayerPtr ctcLayer = createCTCLayer( - "cost", layerSize, useGpu, normByTimes, dataLayer1, labelLayer); - - /// Check cost - LOG(INFO) << "Check cost: " - << checkError(*(warpctcLayer->getOutput().value), - *(ctcLayer->getOutput().value)) - << " different elements."; - - /// Check gradients - LOG(INFO) << "Check gradients: " - << checkError(*(dataLayer0->getOutput().grad), - *(dataLayer1->getOutput().grad)) - << " different elements"; - } - } - } - } -} diff --git a/paddle/legacy/api/Arguments.cpp b/paddle/legacy/api/Arguments.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7bb5a6f75b9a8ab800fc74c6cc01c0b104ccdd5e --- /dev/null +++ b/paddle/legacy/api/Arguments.cpp @@ -0,0 +1,174 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" +#include "PaddleAPIPrivate.h" + +#include "paddle/legacy/parameter/Argument.h" + +size_t Arguments::getSlotNum() const { return m->outputs.size(); } + +Arguments* Arguments::createArguments(size_t slotNum) { + auto args = new Arguments(); + args->m->outputs.resize(slotNum); + return args; +} + +void Arguments::resize(size_t slotNum) { m->outputs.resize(slotNum); } + +Arguments::Arguments() : m(new ArgumentsPrivate()) {} + +Arguments::~Arguments() { delete m; } + +Arguments* Arguments::createByPaddleArgumentVector(void* ptr) { + auto p = (std::vector*)(ptr); + auto args = new Arguments(); + args->m->outputs = *p; + return args; +} + +Arguments* Arguments::createByPaddleArgument(const void* ptr) { + auto p = (paddle::Argument*)(ptr); + auto args = new Arguments(); + args->m->outputs.push_back(*p); + return args; +} + +Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return Matrix::createByPaddleMatrixPtr(&a.value); +} + +Matrix* Arguments::getSlotGrad(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return Matrix::createByPaddleMatrixPtr(&a.grad); +} + +IVector* Arguments::getSlotIds(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return IVector::createByPaddleVectorPtr(&a.ids); +} + +Matrix* Arguments::getSlotIn(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return Matrix::createByPaddleMatrixPtr(&a.in); +} + +void Arguments::setSlotValue(size_t idx, Matrix* mat) throw(RangeError) { + auto& a = m->getArg(idx); + a.value = m->cast(mat->getSharedPtr()); +} + +void Arguments::setSlotGrad(size_t idx, Matrix* mat) throw(RangeError) { + auto& a = m->getArg(idx); + a.grad = m->cast(mat->getSharedPtr()); +} + +void Arguments::setSlotIn(size_t idx, Matrix* mat) throw(RangeError) { + auto& a = m->getArg(idx); + a.in = m->cast(mat->getSharedPtr()); +} + +void Arguments::setSlotIds(size_t idx, IVector* vec) throw(RangeError) { + auto& a = m->getArg(idx); + auto& v = m->cast(vec->getSharedPtr()); + a.ids = v; +} + +template +static inline void doCopyFromSafely(std::shared_ptr& dest, + std::shared_ptr& src) { + if (src) { + if (dest) { + dest->copyFrom(*src); + } else { + dest = src; + } + } +} + +IVector* Arguments::getSlotSequenceStartPositions(size_t idx) const + throw(RangeError) { + auto& a = m->getArg(idx); + if (a.sequenceStartPositions) { + return IVector::createByPaddleVectorPtr( + &a.sequenceStartPositions->getMutableVector(false)); + } else { + return nullptr; + } +} + +IVector* Arguments::getSlotSubSequenceStartPositions(size_t idx) const + throw(RangeError) { + auto& a = m->getArg(idx); + if (a.subSequenceStartPositions) { + return IVector::createByPaddleVectorPtr( + &a.subSequenceStartPositions->getMutableVector(false)); + } else { + return nullptr; + } +} + +void Arguments::setSlotSequenceStartPositions(size_t idx, + IVector* vec) throw(RangeError) { + auto& a = m->getArg(idx); + auto& v = m->cast(vec->getSharedPtr()); + a.sequenceStartPositions = std::make_shared(v); +} + +void Arguments::setSlotSubSequenceStartPositions( + size_t idx, IVector* vec) throw(RangeError) { + auto& a = m->getArg(idx); + auto& v = m->cast(vec->getSharedPtr()); + a.subSequenceStartPositions = std::make_shared(v); +} + +IVector* Arguments::getSlotSequenceDim(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return IVector::createByPaddleVectorPtr(&a.cpuSequenceDims); +} + +void Arguments::setSlotSequenceDim(size_t idx, IVector* vec) throw(RangeError) { + auto& a = m->getArg(idx); + a.cpuSequenceDims = m->cast(vec->getSharedPtr()); +} + +float Arguments::sum() const { return paddle::Argument::sum(m->outputs); } + +int64_t Arguments::getBatchSize(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return a.getBatchSize(); +} + +void Arguments::setSlotFrameHeight(size_t idx, size_t h) throw(RangeError) { + auto& a = m->getArg(idx); + a.setFrameHeight(h); +} + +void Arguments::setSlotFrameWidth(size_t idx, size_t w) throw(RangeError) { + auto& a = m->getArg(idx); + a.setFrameWidth(w); +} + +size_t Arguments::getSlotFrameHeight(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return a.getFrameHeight(); +} + +size_t Arguments::getSlotFrameWidth(size_t idx) const throw(RangeError) { + auto& a = m->getArg(idx); + return a.getFrameWidth(); +} + +void* Arguments::getInternalArgumentsPtr() const { return &m->outputs; } diff --git a/paddle/api/CMakeLists.txt b/paddle/legacy/api/CMakeLists.txt similarity index 100% rename from paddle/api/CMakeLists.txt rename to paddle/legacy/api/CMakeLists.txt diff --git a/paddle/legacy/api/ConfigParser.cpp b/paddle/legacy/api/ConfigParser.cpp new file mode 100644 index 0000000000000000000000000000000000000000..016d6da4e2e4ce888527fe9b61a163056d7729eb --- /dev/null +++ b/paddle/legacy/api/ConfigParser.cpp @@ -0,0 +1,114 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" +#include "PaddleAPIPrivate.h" +#include "paddle/legacy/trainer/Trainer.h" + +struct ParameterConfigPrivate { + paddle::ParameterPtr parameter; + paddle::ParameterConfig config; + + inline paddle::ParameterConfig* getConfigPtr() { + if (parameter != nullptr) { + auto& conf = parameter->getConfig(); + return const_cast(&conf); + } else { + return &config; + } + } +}; + +TrainerConfig::TrainerConfig() : m(new TrainerConfigPrivate()) {} + +TrainerConfig::~TrainerConfig() { delete m; } + +TrainerConfig* TrainerConfig::createFromTrainerConfigFile( + const std::string& confPath) { + LOG(INFO) << "load trainer config from " << confPath; + auto conf = std::make_shared(confPath); + auto retv = new TrainerConfig(); + retv->m->conf = conf; + return retv; +} + +TrainerConfig* TrainerConfig::createFromProtoString(const std::string& str) { + auto retv = new TrainerConfig(); + paddle::TrainerConfig trainerConfigProto; + auto conf = std::make_shared(trainerConfigProto); + CHECK(conf->getMutableConfig().ParseFromString(str)); + retv->m->conf = conf; + return retv; +} + +ModelConfig::ModelConfig() : m(new ModelConfigPrivate()) {} + +ModelConfig::~ModelConfig() { delete m; } + +ModelConfig* TrainerConfig::getModelConfig() const { + auto retv = new ModelConfig(); + retv->m->conf = m->conf; + return retv; +} + +ParameterConfig::ParameterConfig() : m(new ParameterConfigPrivate()) {} + +ParameterConfig::~ParameterConfig() { delete m; } + +ParameterConfig* ParameterConfig::createParameterConfigFromParameterSharedPtr( + void* ptr) { + auto& p = *(paddle::ParameterPtr*)(ptr); + if (p != nullptr) { + auto conf = new ParameterConfig(); + conf->m->parameter = p; + return conf; + } else { + return nullptr; + } +} + +ParameterConfig* ParameterConfig::createParameterConfigFromParameterPtr( + void* ptr) { + auto& p = *(paddle::Parameter*)(ptr); + auto conf = new ParameterConfig(); + conf->m->config = p.getConfig(); + return conf; +} + +std::string ParameterConfig::toProtoString() const { + return m->getConfigPtr()->SerializeAsString(); +} + +void* ParameterConfig::getRawPtr() { return m->getConfigPtr(); } + +OptimizationConfig::OptimizationConfig() : m(new OptimizationConfigPrivate()) {} + +OptimizationConfig::~OptimizationConfig() { delete m; } + +std::string OptimizationConfig::toProtoString() { + return m->getConfig().SerializeAsString(); +} + +OptimizationConfig* TrainerConfig::getOptimizationConfig() const { + auto opt_config = new OptimizationConfig(); + opt_config->m->trainer_config = m->conf; + return opt_config; +} + +OptimizationConfig* OptimizationConfig::createFromProtoString( + const std::string& str) { + auto conf = new OptimizationConfig(); + conf->m->config.ParseFromString(str); + return conf; +} diff --git a/paddle/api/Evaluator.cpp b/paddle/legacy/api/Evaluator.cpp similarity index 100% rename from paddle/api/Evaluator.cpp rename to paddle/legacy/api/Evaluator.cpp diff --git a/paddle/legacy/api/GradientMachine.cpp b/paddle/legacy/api/GradientMachine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5ad2fe11a4c668a318f76492f57091f386183986 --- /dev/null +++ b/paddle/legacy/api/GradientMachine.cpp @@ -0,0 +1,196 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" +#include "PaddleAPIPrivate.h" + +#include "Internal.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" + +std::vector GradientMachine::defaultParamTypes = { + PARAMETER_VALUE, PARAMETER_GRADIENT, PARAMETER_MOMENTUM}; + +GradientMachine::GradientMachine() : m(new GradientMachinePrivate()) {} + +GradientMachine::~GradientMachine() { delete m; } + +GradientMachine* GradientMachine::createFromPaddleModelPtr( + const void* confPtr, + GradientMatchineCreateMode mode, + const std::vector& types) { + auto& conf = *(const paddle::ModelConfig*)(confPtr); + std::vector realTypes; + staticCastVector(&realTypes, types); + auto machineRawPtr = paddle::GradientMachine::create(conf, mode, realTypes); + auto machinePtr = std::shared_ptr(machineRawPtr); + if (machinePtr != nullptr) { + auto machine = new GradientMachine(); + machine->m->machine = machinePtr; + return machine; + } else { + return nullptr; + } +} + +GradientMachine* GradientMachine::createByConfigProtoStr( + const std::string& protoStr, + GradientMatchineCreateMode mode, + const std::vector& types) { + paddle::ModelConfig conf; + conf.ParseFromString(protoStr); + if (conf.IsInitialized()) { + return GradientMachine::createFromPaddleModelPtr(&conf, mode, types); + } else { + return nullptr; + } +} + +GradientMachine* GradientMachine::createByModelConfig( + ModelConfig* conf, + GradientMatchineCreateMode mode, + const std::vector& types) { + auto confPtr = &conf->m->conf->getModelConfig(); + return GradientMachine::createFromPaddleModelPtr(confPtr, mode, types); +} + +void GradientMachine::start() { m->machine->start(); } + +void GradientMachine::finish() { m->machine->finish(); } + +void GradientMachine::onPassEnd() { m->machine->onPassEnd(); } + +void GradientMachine::prefetch(const Arguments& inArgs) { + auto& in = + m->cast>(inArgs.getInternalArgumentsPtr()); + m->machine->prefetch(in); +} + +void GradientMachine::forward(const Arguments& inArgs, + Arguments* outArgs, + PassType passType) { + auto& in = + m->cast>(inArgs.getInternalArgumentsPtr()); + auto& out = m->cast>( + outArgs->getInternalArgumentsPtr()); + paddle::PassType pt = (paddle::PassType)(passType); + m->machine->forward(in, &out, pt); +} + +UpdateCallback::~UpdateCallback() {} + +void UpdateCallback::apply(Parameter* p) { + // UNUSED(p); +} + +class UpdateCallbackWrapper { + public: + explicit UpdateCallbackWrapper(const UpdateCallback& callback) + : callback(const_cast(callback)) {} + + void operator()(paddle::Parameter* param) { + auto p = Parameter::createFromRawPtr(¶m); + // @TODO Use Stack variable instead. + callback.apply(p); + delete p; + } + + private: + UpdateCallback& callback; +}; + +void GradientMachine::backward(const UpdateCallback& callback) { + m->machine->backward(UpdateCallbackWrapper(callback)); +} + +void GradientMachine::forwardBackward(const Arguments& inArgs, + Arguments* outArgs, + PassType passType, + const UpdateCallback& callback) { + auto& in = + m->cast>(inArgs.getInternalArgumentsPtr()); + auto& out = m->cast>( + outArgs->getInternalArgumentsPtr()); + paddle::PassType pt = (paddle::PassType)(passType); + m->machine->forwardBackward(in, &out, pt, UpdateCallbackWrapper(callback)); +} + +void GradientMachine::loadParameters(const std::string& path) { + m->machine->loadParameters(path); +} + +size_t GradientMachine::getParameterSize() const { + return m->machine->getParameters().size(); +} + +Parameter* GradientMachine::getParameter(size_t i) throw(RangeError) { + auto params = m->machine->getParameters(); + if (i < params.size()) { + return Parameter::createFromSharedPtr(&m->machine->getParameters()[i]); + } else { + throw RangeError(); + } +} + +size_t GradientMachine::getNonStaticParameterSize() const { + return m->machine->getNonStaticParameters().size(); +} + +Parameter* GradientMachine::getNonStaticParameter(size_t i) throw(RangeError) { + auto params = m->machine->getNonStaticParameters(); + if (i < params.size()) { + return Parameter::createFromSharedPtr( + &m->machine->getNonStaticParameters()[i]); + } else { + throw RangeError(); + } +} + +void GradientMachine::randParameters() { m->machine->randParameters(); } + +Arguments* GradientMachine::getLayerOutput(const std::string& layerName) const + throw(UnsupportError) { + auto nn = m->machine; + if (nn) { + auto arg = nn->getLayerOutput(layerName); + return Arguments::createByPaddleArgument(&arg); + } else { + throw UnsupportError(); + } +} + +SequenceGenerator* GradientMachine::asSequenceGenerator( + const std::vector& dict, + size_t begin_id, + size_t end_id, + size_t max_length, + size_t beam_size) { + SequenceGenerator* r = + SequenceGenerator::createByGradientMachineSharedPtr(&m->machine); + r->setDict(dict); + r->setBos(begin_id); + r->setEos(end_id); + r->setMaxLength(max_length); + r->setBeamSize(beam_size); + return r; +} + +Evaluator* GradientMachine::makeEvaluator() { + auto ev = new Evaluator(); + ev->m->rawPtr = m->machine->makeEvaluator(); + return ev; +} + +void GradientMachine::eval(Evaluator* evaluator) { + m->machine->eval(evaluator->m->rawPtr); +} diff --git a/paddle/api/Internal.h b/paddle/legacy/api/Internal.h similarity index 100% rename from paddle/api/Internal.h rename to paddle/legacy/api/Internal.h diff --git a/paddle/legacy/api/Matrix.cpp b/paddle/legacy/api/Matrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8862d0ea92c92a2608b49c6b1315badae9e9fd98 --- /dev/null +++ b/paddle/legacy/api/Matrix.cpp @@ -0,0 +1,317 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/math/Matrix.h" +#include +#include +#include "PaddleAPI.h" +#include "paddle/legacy/math/CpuSparseMatrix.h" +#include "paddle/legacy/math/SparseMatrix.h" + +struct MatrixPrivate { + std::shared_ptr mat; +}; + +Matrix::Matrix() : m(new MatrixPrivate()) {} + +Matrix* Matrix::createByPaddleMatrixPtr(void* sharedPtr) { + auto* mat = reinterpret_cast(sharedPtr); + if ((*mat) != nullptr) { + auto m = new Matrix(); + m->m->mat = *mat; + return m; + } else { + return nullptr; + } +} + +Matrix* Matrix::createZero(size_t height, size_t width, bool useGpu) { + auto m = new Matrix(); + m->m->mat = paddle::Matrix::create(height, width, useGpu); + m->m->mat->zero(); + return m; +} + +Matrix* Matrix::createDense(const std::vector& data, + size_t height, + size_t width, + bool useGpu) { + auto m = new Matrix(); + m->m->mat = paddle::Matrix::create(height, width, useGpu); + m->m->mat->copyFrom(data.data(), data.size()); + return m; +} + +Matrix* Matrix::createDenseFromNumpy(float* data, + int dim1, + int dim2, + bool copy, + bool useGpu) throw(UnsupportError) { + if (useGpu) { + /// Gpu mode only supports copy=True + if (!copy) { + throw UnsupportError("Gpu mode only supports copy=True"); + } + return Matrix::createGpuDenseFromNumpy(data, dim1, dim2); + } else { + return Matrix::createCpuDenseFromNumpy(data, dim1, dim2, copy); + } +} + +Matrix* Matrix::createCpuDenseFromNumpy(float* data, + int dim1, + int dim2, + bool copy) { + auto m = new Matrix(); + if (copy) { + m->m->mat = paddle::Matrix::create(dim1, dim2); + m->m->mat->copyFrom(data, dim1 * dim2); + } else { + m->m->mat = paddle::Matrix::create(data, dim1, dim2, false); + } + return m; +} + +Matrix* Matrix::createGpuDenseFromNumpy(float* data, int dim1, int dim2) { + auto m = new Matrix(); + m->m->mat = paddle::Matrix::create(dim1, dim2, false, true); + m->m->mat->copyFrom(data, dim1 * dim2); + return m; +} + +Matrix* Matrix::createSparse(size_t height, + size_t width, + size_t nnz, + bool isNonVal, + bool isTrans, + bool useGpu) { + auto m = new Matrix(); + m->m->mat = paddle::Matrix::createSparseMatrix( + height, + width, + nnz, + isNonVal ? paddle::NO_VALUE : paddle::FLOAT_VALUE, + isTrans, + useGpu); + return m; +} + +Matrix::~Matrix() { delete m; } + +size_t Matrix::getHeight() const { return m->mat->getHeight(); } + +size_t Matrix::getWidth() const { return m->mat->getWidth(); } + +float Matrix::get(size_t x, size_t y) const throw(RangeError) { + if (x > this->getWidth() || y > this->getHeight()) { + RangeError e; + throw e; + } + return m->mat->getElement(x, y); +} + +void Matrix::set(size_t x, size_t y, float val) throw(RangeError, + UnsupportError) { + if (x > this->getWidth() || y > this->getHeight()) { + RangeError e; + throw e; + } + auto rawMat = m->mat.get(); + if (auto cDenseMat = dynamic_cast(rawMat)) { + *(cDenseMat->getData() + x + y * cDenseMat->getWidth()) = val; + } else { + UnsupportError e; + throw e; + } +} + +bool Matrix::isSparse() const { + auto raw_mat = m->mat.get(); + return dynamic_cast(raw_mat) != nullptr || + dynamic_cast(raw_mat) != nullptr; +} + +SparseValueType Matrix::getSparseValueType() const throw(UnsupportError) { + auto cpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (cpuSparseMat != nullptr) { + return (SparseValueType)cpuSparseMat->getValueType(); + } else { + auto gpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (gpuSparseMat != nullptr) { + return (SparseValueType)gpuSparseMat->getValueType(); + } else { + UnsupportError e; + throw e; + } + } +} + +SparseFormatType Matrix::getSparseFormat() const throw(UnsupportError) { + auto cpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (cpuSparseMat != nullptr) { + return (SparseFormatType)cpuSparseMat->getFormat(); + } else { + auto gpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (gpuSparseMat != nullptr) { + return SPARSE_CSR; + } else { + UnsupportError e; + throw e; + } + } +} + +IntArray Matrix::getSparseRowCols(size_t i) const + throw(UnsupportError, RangeError) { + auto cpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (cpuSparseMat != nullptr && + cpuSparseMat->getFormat() == paddle::SPARSE_CSR) { + if (i < cpuSparseMat->getHeight()) { + // cpuSparseMat->print(std::cout); + size_t len = cpuSparseMat->getColNum(i); + return IntArray(cpuSparseMat->getRowCols(i), len); + } else { + RangeError e; + throw e; + } + } else { + UnsupportError e; + throw e; + } +} + +IntWithFloatArray Matrix::getSparseRowColsVal(size_t i) const + throw(UnsupportError, RangeError) { + auto cpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (cpuSparseMat != nullptr && + cpuSparseMat->getValueType() == paddle::FLOAT_VALUE) { + if (i < cpuSparseMat->getHeight()) { + return IntWithFloatArray(cpuSparseMat->getRowValues(i), + cpuSparseMat->getRowCols(i), + cpuSparseMat->getColNum(i)); + } else { + RangeError e; + throw e; + } + } else { + UnsupportError e; + throw e; + } +} + +FloatArray Matrix::getData() const { + auto rawMat = m->mat.get(); + if (dynamic_cast(rawMat->getMemoryHandle().get())) { + // is gpu. then copy data + float* data = rawMat->getData(); + size_t len = rawMat->getElementCnt(); + float* cpuData = new float[len]; + hl_memcpy_device2host(cpuData, data, len * sizeof(float)); + FloatArray ret_val(cpuData, len); + ret_val.needFree = true; + return ret_val; + } else { + FloatArray ret_val(rawMat->getData(), rawMat->getElementCnt()); + return ret_val; + } +} + +void Matrix::sparseCopyFrom( + const std::vector& rows, + const std::vector& cols, + const std::vector& vals) throw(UnsupportError) { + auto cpuSparseMat = + std::dynamic_pointer_cast(m->mat); + if (cpuSparseMat != nullptr) { + // LOG(INFO) <<"RowSize = "<isSparse()) { + throw UnsupportError(); + } else { + *dim1 = m->mat->getHeight(); + *dim2 = m->mat->getWidth(); + *view_m_data = new float[(*dim1) * (*dim2)]; + if (auto cpuMat = dynamic_cast(m->mat.get())) { + auto src = cpuMat->getData(); + auto dest = *view_m_data; + std::memcpy(dest, src, sizeof(paddle::real) * (*dim1) * (*dim2)); + } else if (auto gpuMat = dynamic_cast(m->mat.get())) { + auto src = gpuMat->getData(); + auto dest = *view_m_data; + hl_memcpy_device2host( + dest, src, sizeof(paddle::real) * (*dim1) * (*dim2)); + } else { + LOG(WARNING) << "Unexpected Situation"; + throw UnsupportError(); + } + } +} + +void Matrix::copyFromNumpyMat(float* data, + int dim1, + int dim2) throw(UnsupportError, RangeError) { + if (isSparse()) { + throw UnsupportError(); + } else { + if (this->getHeight() == (size_t)dim1 && this->getWidth() == (size_t)dim2) { + if (m->mat->getData() != data) { + m->mat->copyFrom(data, dim1 * dim2); + } + } else { + throw RangeError(); + } + } +} + +bool Matrix::isGpu() const { + auto rawPtr = m->mat.get(); + return dynamic_cast(rawPtr) != nullptr || + dynamic_cast(rawPtr) != nullptr; +} diff --git a/paddle/legacy/api/Paddle.i b/paddle/legacy/api/Paddle.i new file mode 100644 index 0000000000000000000000000000000000000000..7a1456a5c065821caa54fbf4a10f7ceda08780c0 --- /dev/null +++ b/paddle/legacy/api/Paddle.i @@ -0,0 +1,202 @@ +%module(directors="1") swig_paddle +%include "std_string.i" +%{ +#define SWIG_FILE_WITH_INIT +#include "legacy/api/PaddleAPI.h" +%} + +%include "exception.i" +%typemap(throws) UnsupportError %{ + SWIG_exception(SWIG_RuntimeError, $1.what()); + SWIG_fail; +%} + +%include "std_vector.i" +%include "std_pair.i" +#ifdef SWIGPYTHON +%include "numpy.i" +#endif + +%init %{ +#ifdef SWIGPYTHON +import_array(); +#endif +%} + + +namespace std { +%template(vector_int) vector; +%template(vector_uint) vector; +%template(vector_float) vector; +%template(vector_string) vector; +%template(vector_vec_star) vector; +} +#ifdef SWIGPYTHON +%typemap(in) (int argc, char** argv) { + int i = 0; + if (!PyList_Check($input)) { + PyErr_SetString(PyExc_ValueError, "Expecting a list"); + return NULL; + } + $1 = PyList_Size($input); + $2 = (char **) malloc(($1+1)*sizeof(char *)); + for (i = 0; i < $1; i++) { + PyObject *s = PyList_GetItem($input,i); + if (!PyString_Check(s)) { + free($2); + PyErr_SetString(PyExc_ValueError, "List items must be strings"); + return NULL; + } + $2[i] = PyString_AsString(s); + } + $2[i] = 0; +} +%typemap(freearg) (int argc, char** argv) { + if ($2) free($2); +} + +%typemap(out) FloatArray { + $result = PyList_New($1.length); + for (size_t i=0; i<$1.length; ++i) { + PyList_SetItem($result, i, PyFloat_FromDouble($1.buf[i])); + } + if($1.needFree) { + delete [] $1.buf; + } +} + +%typemap(out) IntArray { + $result = PyList_New($1.length); + for (size_t i=0; i<$1.length; ++i) { + PyList_SetItem($result, i, PyInt_FromLong($1.buf[i])); + } + if ($1.needFree) { + delete [] $1.buf; + } +} + +%typemap(out) IntWithFloatArray { + $result = PyList_New($1.length); + for (size_t i=0; i<$1.length; ++i) { + PyList_SetItem($result, i, PyTuple_Pack(2, + PyInt_FromLong($1.idxBuf[i]), + PyFloat_FromDouble($1.valBuf[i]) + )); + } + if ($1.needFree) { + delete [] $1.idxBuf; + delete [] $1.valBuf; + } +} + + +%rename(__getitem__) IVector::get; +%rename(__setitem__) IVector::set; +%rename(__len__) IVector::getSize; +%rename(__getitem__) Vector::get; +%rename(__setitem__) Vector::set; +%rename(__len__) Vector::getSize; +%rename(__len__) Parameter::getSize; +%rename(__call__) ParameterTraverseCallback::apply; +%rename(__repr__) Evaluator::toString; + +%apply (float* INPLACE_ARRAY2, int DIM1, int DIM2) { + (float* data, int dim1, int dim2) +} + +%apply (float** ARGOUTVIEW_ARRAY2, int* DIM1, int* DIM2) { + (float** view_data, int* dim1, int* dim2) +} + +%apply (float** ARGOUTVIEWM_ARRAY2, int* DIM1, int* DIM2) { + (float** view_m_data, int* dim1, int* dim2) +} + +%apply (int** ARGOUTVIEWM_ARRAY1, int* DIM1) { + (int** view_m_data, int* dim1) +} + +%apply (int* INPLACE_ARRAY1, int DIM1) { + (int* data, int dim) +} + +%apply (int** ARGOUTVIEW_ARRAY1, int* DIM1) { + (int** view_data, int* dim1) +} + +%apply (float* INPLACE_ARRAY1, int DIM1) { + (float* data, int dim) +} + +%apply (float** ARGOUTVIEW_ARRAY1, int* DIM1) { + (float** view_data, int* dim1) +} + +%apply (float** ARGOUTVIEWM_ARRAY1, int* DIM1) { + (float** view_m_data, int* dim1) +} + +#endif +// The below functions internally create object by "new", so it should use +// use SWIG to handle gc. There are hints for SWIG to handle GC. +%newobject Matrix::createZero; +%newobject Matrix::createSparse; +%newobject Matrix::createDense; +%newobject Matrix::createDenseFromNumpy; +%newobject Matrix::createCpuDenseFromNumpy; +%newobject Matrix::createGpuDenseFromNumpy; +%newobject Vector::createZero; +%newobject Vector::create; +%newobject Vector::createVectorFromNumpy; +%newobject Vector::createCpuVectorFromNumpy; +%newobject Vector::createGpuVectorFromNumpy; +%newobject IVector::createZero; +%newobject IVector::create; +%newobject IVector::createVectorFromNumpy; +%newobject IVector::createCpuVectorFromNumpy; +%newobject IVector::createGpuVectorFromNumpy; +%newobject Trainer::createByCommandLine; +%newobject Trainer::getForwardOutput; +%newobject Trainer::getLayerOutput; +%newobject Arguments::getSlotValue; +%newobject Arguments::getSlotIds; +%newobject Arguments::getSlotIn; +%newobject Arguments::getSlotSequenceStartPositions; +%newobject Arguments::getSlotSequenceDim; +%newobject Arguments::createArguments; +%newobject GradientMachine::createByConfigProtoStr; +%newobject GradientMachine::createByModelConfig; +%newobject GradientMachine::asSequenceGenerator; +%newobject GradientMachine::getParameter; +%newobject GradientMachine::getLayerOutput; +%newobject GradientMachine::makeEvaluator; +%newobject TrainerConfig::createFromTrainerConfigFile; +%newobject TrainerConfig::getModelConfig; +%newobject TrainerConfig::getOptimizationConfig; +%newobject Parameter::getBuf; +%newobject Parameter::getConfig; +%newobject ParameterOptimizer::create; +%newobject ParameterOptimizer::needSpecialTraversal; +%newobject ParameterUpdater::createLocalUpdater; +%newobject ParameterUpdater::createRemoteUpdater; +%newobject ParameterUpdater::createNewRemoteUpdater; + +%feature("director") UpdateCallback; +%feature("autodoc", 1); // To generate method stub, for code hint in ide + +// Ignore many private class, and method cannot be handled by swig. +%ignore MatrixPrivate; +%ignore TrainerPrivate; +%ignore IVector::operator[]; +%ignore ArgumentsPrivate; +%ignore GradientMachinePrivate; +%ignore TrainerConfigPrivate; +%ignore ModelConfigPrivate; +%ignore ParameterPrivate; +%ignore SequenceGeneratorPrivate; +%ignore VectorPrivate; +%ignore ParameterConfigPrivate; +%ignore OptimizationConfigPrivate; +%ignore ParameterTraverseCallbackPrivate; +%include "legacy/utils/GlobalConstants.h" +%include "legacy/api/PaddleAPI.h" diff --git a/paddle/legacy/api/PaddleAPI.h b/paddle/legacy/api/PaddleAPI.h new file mode 100644 index 0000000000000000000000000000000000000000..475984a3d57ebc25d5d071c33b7e6562ac78c503 --- /dev/null +++ b/paddle/legacy/api/PaddleAPI.h @@ -0,0 +1,1054 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/GlobalConstants.h" + +/// Import PaddlePaddle's enumeration into global namespace. +using namespace paddle::enumeration_wrapper; // NOLINT + +/** + * @brief Initialize paddle. + * + * In python, this method should be invoked as + * @code + * import sys + * import paddle + * paddle.initPaddle(sys.argv) + * or you can change arguments as any list of str. + * @endcode + */ +void initPaddle(int argc, char** argv); + +/// Return FLAGS_use_gpu +bool isUsingGpu(); + +/// Set the Flags_use_gpu to the given parameter +void setUseGpu(bool useGpu); + +/// Return true if this py_paddle is compiled in GPU Version +bool isGpuVersion(); + +/// Return FLAGS_trainer_count +int getTrainerCount(); + +/// The Error of IO Operation. Such as file not found, etc. +class IOError {}; + +/// Out of range error +class RangeError {}; + +/// Not support Error, such as access GPU memory directly, etc. +class UnsupportError : public std::runtime_error { + public: + UnsupportError() : std::runtime_error(" ") {} + explicit UnsupportError(const std::string& message) + : std::runtime_error(message) {} +}; + +/// This type will map to python's list of float. +struct FloatArray { + const float* buf; + const size_t length; + bool needFree; // true if the buf is dynamic alloced. + FloatArray(const float* b, const size_t l); +}; + +/// This type will map to python's list of int +struct IntArray { + const int* buf; + const size_t length; + bool needFree; + IntArray(const int* b, const size_t l, bool f = false); +}; + +/// This type will map to python's list of (int, float) +struct IntWithFloatArray { + const float* valBuf; + const int* idxBuf; + const size_t length; + bool needFree; + IntWithFloatArray(const float* v, const int* i, size_t l, bool f = false); +}; + +enum SparseValueType { SPARSE_NON_VALUE = 0, SPARSE_VALUE = 1 }; + +enum SparseFormatType { SPARSE_CSR = 0, SPARSE_CSC = 1 }; + +/** + * In Python, -1UL is hard to write. So define a const value used by python + * side. + */ +const size_t NO_SPARSE_ID = -1UL; + +struct MatrixPrivate; +class Matrix { + Matrix(); // User Cannot Create Matrix. + DISABLE_COPY(Matrix); + static Matrix* createByPaddleMatrixPtr(void* sharedPtr); + + public: + virtual ~Matrix(); + + /** + * Create A Matrix with height,width, which is filled by zero. + */ + static Matrix* createZero(size_t height, + size_t width, + bool useGpu = isUsingGpu()); + + /** + * Create Sparse Matrix. + * + * After create sparse, sparseCopyFrom can be used to fill matrix. + * + * @param nnz Number of non zero values. + * + * @note the default sparse type is SPARSE_CSR. + */ + static Matrix* createSparse(size_t height, + size_t width, + size_t nnz, + bool isNonVal = true, + bool trans = false, + bool useGpu = isUsingGpu()); + + /** + * Create Dense Matrix. + * + * @param data list of float should be passed in python. + * @note the value will be copy into a new matrix. + */ + static Matrix* createDense(const std::vector& data, + size_t height, + size_t width, + bool useGpu = isUsingGpu()); + + static Matrix* createDenseFromNumpy( + float* data, + int dim1, + int dim2, + bool copy = true, + bool useGpu = isUsingGpu()) throw(UnsupportError); + + /** + * Create Cpu Dense Matrix from numpy matrix, dtype=float32 + * + * @param data a numpy matrix. + * @param dim1 dimension of data. + * @param dim2 dimension of data. + * @param copy true if copy into a new matrix, false will create + * matrix inplace. copy = false should be used with extreme + * care because Matrix will share the memory with the given + * numpy array. If the numpy array object is no longer valid, + * the memory space will not be usable. + */ + static Matrix* createCpuDenseFromNumpy(float* data, + int dim1, + int dim2, + bool copy = true); + + /// Create Gpu Dense Matrix from numpy matrix, dtype=float32 + static Matrix* createGpuDenseFromNumpy(float* data, int dim1, int dim2); + + /** + * Cast to numpy matrix. + * + * @note This method take no parameter in python. + * @note This method in python will return a numpy matrix, not void. + * @note Only CpuDenseMatrix is supported. + * + * Example: + * @code + * import paddle + * m = paddle.Matrix.createZero(10,2) + * numpy_mat = m.toNumpyMat() + * @endcode + */ + void toNumpyMatInplace(float** view_data, + int* dim1, + int* dim2) throw(UnsupportError); + + /// Copy To numpy mat. + void copyToNumpyMat(float** view_m_data, + int* dim1, + int* dim2) throw(UnsupportError); + + /// Copy From Numpy Mat + void copyFromNumpyMat(float* data, int dim1, int dim2) throw(UnsupportError, + RangeError); + + /// return true if this matrix is sparse. + bool isSparse() const; + + SparseValueType getSparseValueType() const throw(UnsupportError); + + SparseFormatType getSparseFormat() const throw(UnsupportError); + + IntArray getSparseRowCols(size_t i) const throw(UnsupportError, RangeError); + + IntWithFloatArray getSparseRowColsVal(size_t i) const + throw(UnsupportError, RangeError); + + size_t getHeight() const; + + size_t getWidth() const; + + float get(size_t x, size_t y) const throw(RangeError); + + void set(size_t x, size_t y, float val) throw(RangeError, UnsupportError); + + /// return type is list of float + FloatArray getData() const; + + /** + * Copy from rows, cols, values. + * + * if sparse_nonvalue, the values should be [] + */ + void sparseCopyFrom(const std::vector& rows, + const std::vector& cols, + const std::vector& values = + std::vector()) throw(UnsupportError); + + bool isGpu() const; + + private: + void* getSharedPtr() const; + + MatrixPrivate* m; + friend class Trainer; + friend class GradientMachine; + friend class Arguments; +}; + +struct VectorPrivate; +class Vector { + DISABLE_COPY(Vector); + Vector(); + static Vector* createByPaddleVectorPtr(void* ptr); + + void* getSharedPtr(); + + public: + ~Vector(); + + /// Create Vector filled with zero. + static Vector* createZero(size_t sz, bool useGpu = isUsingGpu()); + + /** + * Create Vector from list of float. + * + * It will create a new vector, and copy data into it. + */ + static Vector* create(const std::vector& data, + bool useGpu = isUsingGpu()); + + static Vector* createVectorFromNumpy( + float* data, + int dim, + bool copy = true, + bool useGpu = isUsingGpu()) throw(UnsupportError); + /** + * Create Cpu Vector from numpy array, which dtype=float32 + * + * If copy is false, it will create vector inplace. + */ + static Vector* createCpuVectorFromNumpy(float* data, + int dim, + bool copy = true); + + /// Create Gpu Vector from numpy array, which dtype=float32 + static Vector* createGpuVectorFromNumpy(float* data, int dim); + + /** + * copy from another vector + * throw(RangeError) if size of src vector is different from size of this + * vector + */ + void copyFrom(Vector* src) throw(RangeError); + + /// Cast to numpy array inplace. + void toNumpyArrayInplace(float** view_data, int* dim1) throw(UnsupportError); + + /// Copy to numpy array. + void copyToNumpyArray(float** view_m_data, int* dim1); + + /// Copy from numpy array. + void copyFromNumpyArray(float* data, int dim); + + /// __getitem__ in python + float get(const size_t idx) const throw(RangeError, UnsupportError); + + /// __setitem__ in python + void set(const size_t idx, float val) throw(RangeError, UnsupportError); + + /// Return is GPU vector or not. + bool isGpu() const; + + /// Return a list of float, the memory is alloced and copied. + FloatArray getData() const; + + /// __len__ in python + size_t getSize() const; + + private: + VectorPrivate* m; + + private: + friend class Parameter; + friend class ParameterOptimizer; + friend struct ParameterTraverseCallbackPrivate; +}; + +struct IVectorPrivate; +class IVector { + IVector(); + DISABLE_COPY(IVector); + static IVector* createByPaddleVectorPtr(void* ptr); + + public: + /// Create IVector filled with zero + static IVector* createZero(size_t sz, bool useGpu = isUsingGpu()); + + /** + * Create IVector from list of int. + * It will create a new vector, and copy data into it. + */ + static IVector* create(const std::vector& data, + bool useGpu = isUsingGpu()); + + static IVector* createVectorFromNumpy( + int* data, + int dim, + bool copy = true, + bool useGpu = isUsingGpu()) throw(UnsupportError); + + /** + * Create Cpu IVector from numpy array, which dtype=int32 + * + * If copy is false, it will create vector inplace + */ + static IVector* createCpuVectorFromNumpy(int* data, + int dim, + bool copy = true); + /** + * Create Gpu IVector from numpy array, which dtype=int32 + */ + static IVector* createGpuVectorFromNumpy(int* data, int dim); + + /// Cast to numpy array inplace. + void toNumpyArrayInplace(int** view_data, int* dim1) throw(UnsupportError); + + /// Copy to numpy array. + void copyToNumpyArray(int** view_m_data, int* dim1); + + /// Copy from numpy array. + void copyFromNumpyArray(int* data, int dim); + + virtual ~IVector(); + + /// Return a list of int, the memory is alloced and copied. + IntArray getData() const; + + /// This method will map to python [] method. + int& operator[](const size_t idx) throw(RangeError, UnsupportError); + + const int& operator[](const size_t idx) const + throw(RangeError, UnsupportError); + + inline int get(const size_t idx) const throw(RangeError, UnsupportError) { + return (*this)[idx]; + } + + inline void set(const size_t idx, int val) throw(RangeError, UnsupportError) { + (*this)[idx] = val; + } + + /// Return true if it is gpu vector. + bool isGpu() const; + + /// This method will map to python __len__(); + size_t getSize() const; + + private: + void* getSharedPtr() const; + + friend class Arguments; + IVectorPrivate* m; +}; + +struct ArgumentsPrivate; + +/// The Arguments is actual a std::vector in paddle. +class Arguments { + private: + Arguments(); // Internal Create. + DISABLE_COPY(Arguments); + + public: + /** + * Create a arguments with size. + * Note that it can be zero. + */ + static Arguments* createArguments(size_t slotNum); + + void resize(size_t slotNum); + + virtual ~Arguments(); + + /** + * Return the slot number that aguments contains. + * + * It is actually the vector's size + */ + size_t getSlotNum() const; + + /** + * The get functions of Arguments + * + * the param idx is the slot id + */ + Matrix* getSlotValue(size_t idx) const throw(RangeError); + Matrix* getSlotGrad(size_t idx) const throw(RangeError); + IVector* getSlotIds(size_t idx) const throw(RangeError); + Matrix* getSlotIn(size_t idx) const throw(RangeError); + IVector* getSlotSequenceStartPositions(size_t idx) const throw(RangeError); + IVector* getSlotSubSequenceStartPositions(size_t idx) const throw(RangeError); + IVector* getSlotSequenceDim(size_t idx) const throw(RangeError); + // End Of get functions of Arguments + + int64_t getBatchSize(size_t idx = 0) const throw(RangeError); + + /** + * The set functions of Arguments. + * + * The param idx is the slot id. + * The other param is the input Matrix or vector. + */ + void setSlotValue(size_t idx, Matrix* mat) throw(RangeError); + void setSlotGrad(size_t idx, Matrix* mat) throw(RangeError); + void setSlotIn(size_t idx, Matrix* mat) throw(RangeError); + void setSlotIds(size_t idx, IVector* vec) throw(RangeError); + void setSlotSequenceStartPositions(size_t idx, + IVector* vec) throw(RangeError); + void setSlotSubSequenceStartPositions(size_t idx, + IVector* vec) throw(RangeError); + void setSlotSequenceDim(size_t idx, IVector* vec) throw(RangeError); + + /** + * Set the frame height of the idx-th Argument. + * + * @param ids The index of which Argument. + * @param h The height value. + */ + void setSlotFrameHeight(size_t idx, size_t h) throw(RangeError); + + /** + * Set the frame height of the idx-th Argument. + * + * @param ids The index of which Argument. + * @param h The height value. + */ + void setSlotFrameWidth(size_t idx, size_t w) throw(RangeError); + + size_t getSlotFrameHeight(size_t idx = 0) const throw(RangeError); + size_t getSlotFrameWidth(size_t idx = 0) const throw(RangeError); + + float sum() const; + + private: + static Arguments* createByPaddleArgumentVector(void* ptr); + static Arguments* createByPaddleArgument(const void* ptr); + void* getInternalArgumentsPtr() const; + + private: + ArgumentsPrivate* m; + friend class Trainer; + friend class GradientMachine; + friend class SequenceGenerator; +}; + +enum GradientMatchineCreateMode { + CREATE_MODE_NORMAL = paddle::GradientMachine::kNormal, + CREATE_MODE_SGD_SPARSE_CPU_TRAINING = + paddle::GradientMachine::kSgdSparseCpuTraining, + CREATE_MODE_TESTING = paddle::GradientMachine::kTesting +}; + +struct ParameterConfigPrivate; +class ParameterConfig { + DISABLE_COPY(ParameterConfig); + ParameterConfig(); + + /** + * Internal methods + */ + static ParameterConfig* createParameterConfigFromParameterSharedPtr( + void* ptr); + static ParameterConfig* createParameterConfigFromParameterPtr(void* ptr); + void* getRawPtr(); + + public: + ~ParameterConfig(); + + /** + * return proto buf string. + */ + std::string toProtoString() const; + + private: + ParameterConfigPrivate* m; + + private: + friend class Parameter; + friend class ParameterOptimizer; + friend struct ParameterTraverseCallbackPrivate; +}; + +struct OptimizationConfigPrivate; +class OptimizationConfig { + DISABLE_COPY(OptimizationConfig); + OptimizationConfig(); + + public: + static OptimizationConfig* createFromProtoString(const std::string& str); + ~OptimizationConfig(); + + /** + * return protobuf string. + */ + std::string toProtoString(); + + private: + OptimizationConfigPrivate* m; + + friend class TrainerConfig; + friend class ParameterOptimizer; + friend class ParameterUpdater; + friend class Trainer; +}; + +struct ParameterPrivate; +class Parameter { + private: + Parameter(); + DISABLE_COPY(Parameter); + + public: + virtual ~Parameter(); + + /** + * get parameter name + */ + std::string getName() const; + + /** + * get buf in Parameter + */ + Vector* getBuf(ParameterType type); + + /** + * get id + */ + size_t getID() const; + + ParameterConfig* getConfig(); + void setValueUpdated(); + + bool save(const std::string& filename) const; + + bool load(const std::string& filename) const; + + size_t getSize() const; + + private: + static Parameter* createFromRawPtr(void* ptr); + static Parameter* createFromSharedPtr(void* ptr); + + private: + ParameterPrivate* m; + friend class UpdateCallbackWrapper; + friend class GradientMachine; + friend class ParameterUpdater; +}; + +struct ModelConfigPrivate; +/** + * You can only get model config from TrainerConfig. + * + * It is used by GradientMachine. + */ +class ModelConfig { + private: + ModelConfig(); + DISABLE_COPY(ModelConfig); + + public: + virtual ~ModelConfig(); + + private: + ModelConfigPrivate* m; + friend class TrainerConfig; + friend struct TrainerConfigPrivate; + friend class GradientMachine; +}; + +struct TrainerConfigPrivate; +/** + * To get TrainerConfig from file. + * + * It is used by GradientMachine. + */ +class TrainerConfig { + private: + TrainerConfig(); + DISABLE_COPY(TrainerConfig); + + public: + virtual ~TrainerConfig(); + + static TrainerConfig* createFromTrainerConfigFile( + const std::string& configPath); + static TrainerConfig* createFromProtoString(const std::string& str); + + ModelConfig* getModelConfig() const; + + OptimizationConfig* getOptimizationConfig() const; + + private: + TrainerConfigPrivate* m; + friend class Trainer; +}; + +/** + * The callback in backword. + * + * You can inherit this class in python. + * + * @code + * class UpdateCallbackInPython(paddle.UpdateCallback): + * def __init__(self): + * paddle.UpdateCallback.__init__(self) + * + * def apply(self, param): + * assert isinstance(param, paddle.Parameter) + * @endcode + */ +class UpdateCallback { + public: + virtual ~UpdateCallback(); + virtual void apply(Parameter* p); +}; + +struct ParameterTraverseCallbackPrivate; +class ParameterTraverseCallback { + DISABLE_COPY(ParameterTraverseCallback); + ParameterTraverseCallback(); + + public: + ~ParameterTraverseCallback(); + + void apply(const std::vector& vecs, + const ParameterConfig& config, + size_t sparseId); + + private: + ParameterTraverseCallbackPrivate* m; + friend class ParameterOptimizer; +}; + +/** + * The ParameterOptimizer Wrapper Class. + * + * Basically same as common/ParameterOptimizer.h + */ +struct ParameterOptimizerPrivate; +class ParameterOptimizer { + DISABLE_COPY(ParameterOptimizer); + ParameterOptimizer(); + + public: + static ParameterOptimizer* create(OptimizationConfig* config); + + ~ParameterOptimizer(); + + void init(size_t numRows, const ParameterConfig* config); + + void startPass(); + + void finishPass(); + + void startBatch(size_t numSamplesProcessed); + + void finishBatch(); + + void update(const std::vector& vecs, + const ParameterConfig& conf, + size_t sparseId = NO_SPARSE_ID); + + std::vector getParameterTypes() const; + + ParameterTraverseCallback* needSpecialTraversal( + const ParameterConfig& config) const; + + private: + ParameterOptimizerPrivate* m; +}; + +class SequenceGenerator; +class Evaluator; +struct GradientMachinePrivate; +class GradientMachine { + private: + GradientMachine(); + DISABLE_COPY(GradientMachine); + + public: + virtual ~GradientMachine(); + + /** + * Create By ProtoStr. + * + * The ProtoStr can be generate by python's protobuf code. + */ + static GradientMachine* createByConfigProtoStr( + const std::string& protoStr, + GradientMatchineCreateMode mode = CREATE_MODE_NORMAL, + const std::vector& parameterTypes = defaultParamTypes); + + /** + * Create by ModelConfig object. + * + * To get ModelConfig, you can get TrainerConfig from config file, then get + * model config by TrainerConfig + */ + static GradientMachine* createByModelConfig( + ModelConfig* conf, + GradientMatchineCreateMode mode = CREATE_MODE_NORMAL, + const std::vector& parameterTypes = defaultParamTypes); + + /** + * @brief finish + */ + void finish(); + + void start(); + + /** + * Prefetch row ids of sparse parameter. + */ + void prefetch(const Arguments& inArgs); + + /** + * Do some thing when train pass ended. + */ + void onPassEnd(); + + /** + * The forward stage of GradientMachine. + * + * @note the outArgs could be zero length arguemnts. + * @note THIS METHOD IS VERY USEFULL FOR PREDICT FROM TRAINED MODEL. + */ + void forward(const Arguments& inArgs, Arguments* outArgs, PassType passType); + + /** + * The backward stage of GradientMachine. + * + * @note Currently the ParameterUpdater is not wrapped in SWIG, so backward + * cannot actually train a network. But you can write a update callback to + * change the parameter or implement a ParameterUpdater in python side. + */ + void backward(const UpdateCallback& callback = UpdateCallback()); + + /** + * Combine forward/backward + */ + void forwardBackward(const Arguments& inArgs, + Arguments* outArgs, + PassType passType, + const UpdateCallback& callback = UpdateCallback()); + + void loadParameters(const std::string& path); + + size_t getParameterSize() const; + Parameter* getParameter(size_t i) throw(RangeError); + + size_t getNonStaticParameterSize() const; + Parameter* getNonStaticParameter(size_t i) throw(RangeError); + + void randParameters(); + + Arguments* getLayerOutput(const std::string& layerName) const + throw(UnsupportError); + + /** + * Create a sequence generator. + * + * @note It just like a paddle_gen_sequence. + */ + SequenceGenerator* asSequenceGenerator( + const std::vector& dict = std::vector(), + size_t begin_id = 0UL, + size_t end_id = 0UL, + size_t max_length = 100UL, + size_t beam_size = -1UL); + + Evaluator* makeEvaluator(); + + void eval(Evaluator* evaluator); + + private: + GradientMachinePrivate* m; + + static GradientMachine* createFromPaddleModelPtr( + const void* confPtr, + GradientMatchineCreateMode mode, + const std::vector& types); + + // Not to use c++ 11 init-list, so we use static var as function default arg. + static std::vector defaultParamTypes; + friend class Trainer; + friend class ParameterUpdater; +}; + +struct ParameterUpdaterPrivate; +class ParameterUpdater { + private: + ParameterUpdater(); + + public: + static ParameterUpdater* createLocalUpdater(OptimizationConfig* config); + static ParameterUpdater* createRemoteUpdater(OptimizationConfig* config, + int passCount, + bool useSparseUpdater); + static ParameterUpdater* createNewRemoteUpdater( + OptimizationConfig* config, + const std::string pserverSpec, + const bool useEtcd) throw(UnsupportError); + ~ParameterUpdater(); + + /** + * @brief initialize Parameter Updater by GradientMachine. + * @param gm + */ + void init(const GradientMachine& gm); + + /** + * @brief begin of a training/testing of one pass. + */ + void startPass(); + + /** + * @brief end of a traning/testing of one pass. + */ + void finishPass(); + + /** + * @brief begin of a training/testing of one batch. + * @param data batch's size + * @return PassType, mostly will be training. + */ + PassType startBatch(size_t batchSize); + + /** + * @brief end of a traning/testing of one batch + * @param cost current batch cost. + */ + void finishBatch(float cost); + + /** + * @brief update a parameter (by local optimizer or by cluster pserver) + * @param param + */ + void update(Parameter* param); + + /** + * @breif only get required sparse rows by default. + * @param fullSize: get full matrix parameter if *fullSize* set + * @param apply: get PARAMETER_APPLY on pserver if *apply* set + */ + void getParametersRemote(bool fullSize = false, bool apply = false); + + /** + * @brief restore the average parameter. + * @note It is only used in AverageOptimizer. Restore will get the current + * PARAMETER_VALUE back. + */ + void restore(); + + /** + * @brief apply. Store the average parameter. + * @note It is only used in AverageOptimizer. Apply will store the current + * PARAMETER_VALUE to buffer, calcaualte current Average Parameter, and save + * it to PARAMETER_VALUE. + */ + void apply(); + + /** + * @brief catchUpWith The Regularization will be delayed in many situations( + * pserver, local sparse). Catch Up means catch the regularization up, apply + * regularization to all params. + */ + void catchUpWith(); + + private: + ParameterUpdaterPrivate* m; +}; + +struct EvaluatorPrivate; +class Evaluator { + private: + Evaluator(); + DISABLE_COPY(Evaluator); + + public: + ~Evaluator(); + + /** + * @brief begin an evaluate stage. + */ + void start(); + + /** + * @brief end an evaluate stage. + */ + void finish(); + + /** + * @brief toString will get a evaluate result. + * + * __repr__ method in python + */ + std::string toString(); + + std::vector getNames() const; + + double getValue(const std::string name) const; + + private: + EvaluatorPrivate* m; + + friend class GradientMachine; +}; + +struct TrainerPrivate; +class Trainer { + private: + TrainerPrivate* m; + Trainer(); + Trainer(TrainerConfig* optConfig, GradientMachine* gm); + DISABLE_COPY(Trainer); + + public: + virtual ~Trainer(); + + /// Create A Trainer By TrainerConfig. using paddle command line. + static Trainer* createByCommandLine() throw(IOError); + + static Trainer* create(TrainerConfig* optConfig, + GradientMachine* gm) throw(IOError); + + /// Start training + void startTrain(); + + /// Finish training + void finishTrain(); + + /// Start a pass. + void startTrainPass(); + + /// Finish a pass + void finishTrainPass(); + + /** + * Train one batch, + * + * @return true if all batch finished. + */ + bool trainOneBatch(size_t batchSize); + + void trainOneDataBatch(size_t batchSize, const Arguments& args); + + void startTestPeriod(); + void testOneDataBatch(size_t batchSize, const Arguments& args); + void finishTestPeriod(); + + void forwardOneBatch(size_t batchSize); + + Arguments* getForwardOutput(); + + Arguments* getLayerOutput(const std::string& layerName) const; +}; + +/// the N-Best results generated from one input sequence. +class ISequenceResults { + public: + virtual ~ISequenceResults(); + + /// Number of result. + virtual size_t getSize() const = 0; + + /** + * Get sentence from dictionary. + * + * @param id the index of result. + * @param split if true, the return sentence will be splited with ' ' by + * each word. Default is false. + */ + virtual std::string getSentence(size_t id, bool split = false) const + throw(RangeError) = 0; + virtual std::vector getSequence(size_t id) const throw(RangeError) = 0; + virtual float getScore(size_t id) const throw(RangeError) = 0; +}; + +struct SequenceGeneratorPrivate; +class SequenceGenerator { + DISABLE_COPY(SequenceGenerator); + SequenceGenerator(); + + public: + virtual ~SequenceGenerator(); + + /** + * Generate Sequence by input. + * + * @note The inArgs is just one sequence of data. + * @note The return will get a N-best generate result by inArgs. + * Sort by score. + */ + ISequenceResults* generateSequence(const Arguments& inArgs) const; + + void setDict(const std::vector& dict); + void setBos(size_t bos); + void setEos(size_t eos); + void setMaxLength(size_t maxlength); + void setBeamSize(size_t beamSize); + + private: + static SequenceGenerator* createByGradientMachineSharedPtr(void* ptr); + friend class GradientMachine; + + private: + SequenceGeneratorPrivate* m; +}; diff --git a/paddle/legacy/api/PaddleAPIPrivate.h b/paddle/legacy/api/PaddleAPIPrivate.h new file mode 100644 index 0000000000000000000000000000000000000000..3ee192c31d597c4b4575e4a53a4aece09e642831 --- /dev/null +++ b/paddle/legacy/api/PaddleAPIPrivate.h @@ -0,0 +1,97 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#pragma once +#include +#include "PaddleAPI.h" +#include "paddle/legacy/gserver/evaluators/Evaluator.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" +#include "paddle/legacy/parameter/ParameterUpdaterBase.h" +#include "paddle/legacy/trainer/TrainerConfigHelper.h" + +struct GradientMachinePrivate { + std::shared_ptr machine; + + template + inline T& cast(void* ptr) { + return *(T*)(ptr); + } +}; + +struct OptimizationConfigPrivate { + std::shared_ptr trainer_config; + paddle::OptimizationConfig config; + + const paddle::OptimizationConfig& getConfig() { + if (trainer_config != nullptr) { + return trainer_config->getOptConfig(); + } else { + return config; + } + } +}; + +struct TrainerConfigPrivate { + std::shared_ptr conf; + TrainerConfigPrivate() {} +}; + +struct ModelConfigPrivate { + std::shared_ptr conf; +}; + +struct ArgumentsPrivate { + std::vector outputs; + + inline paddle::Argument& getArg(size_t idx) throw(RangeError) { + if (idx < outputs.size()) { + return outputs[idx]; + } else { + RangeError e; + throw e; + } + } + + template + std::shared_ptr& cast(void* rawPtr) const { + return *(std::shared_ptr*)(rawPtr); + } +}; + +struct ParameterUpdaterPrivate { + std::unique_ptr updater; +}; + +struct ParameterPrivate { + std::shared_ptr sharedPtr; + paddle::Parameter* rawPtr; // rawPtr only used in ParameterUpdater, + // in other situation sharedPtr should + // contains value. + + ParameterPrivate() : sharedPtr(nullptr), rawPtr(nullptr) {} + + paddle::Parameter* getPtr() { + if (sharedPtr) { + return sharedPtr.get(); + } else { + return rawPtr; + } + } +}; + +struct EvaluatorPrivate { + paddle::Evaluator* rawPtr; + + EvaluatorPrivate() : rawPtr(nullptr) {} + ~EvaluatorPrivate() { delete rawPtr; } +}; diff --git a/paddle/legacy/api/Parameter.cpp b/paddle/legacy/api/Parameter.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f05740eb750cccd8cfb6cbc826a04585ec06822e --- /dev/null +++ b/paddle/legacy/api/Parameter.cpp @@ -0,0 +1,68 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/parameter/Parameter.h" +#include "PaddleAPI.h" +#include "PaddleAPIPrivate.h" + +Parameter::Parameter() : m(new ParameterPrivate()) {} + +Parameter::~Parameter() { delete m; } + +Parameter* Parameter::createFromRawPtr(void* ptr) { + auto p = new Parameter(); + p->m->rawPtr = *static_cast(ptr); + return p; +} + +Parameter* Parameter::createFromSharedPtr(void* ptr) { + auto& p = *(paddle::ParameterPtr*)(ptr); + if (p == nullptr) { + return nullptr; + } else { + auto retParam = new Parameter(); + retParam->m->sharedPtr = p; + return retParam; + } +} + +std::string Parameter::getName() const { return m->getPtr()->getName(); } + +Vector* Parameter::getBuf(ParameterType type) { + auto buf = m->getPtr()->getBuf(type); + return Vector::createByPaddleVectorPtr(&buf); +} + +ParameterConfig* Parameter::getConfig() { + if (m->sharedPtr) { + return ParameterConfig::createParameterConfigFromParameterSharedPtr( + &m->sharedPtr); + } else { + return ParameterConfig::createParameterConfigFromParameterPtr(m->rawPtr); + } +} + +size_t Parameter::getID() const { return m->getPtr()->getID(); } + +void Parameter::setValueUpdated() { m->getPtr()->setValueUpdated(); } + +bool Parameter::save(const std::string& filename) const { + return m->getPtr()->save(filename); +} + +bool Parameter::load(const std::string& filename) const { + return m->getPtr()->load(filename); +} + +size_t Parameter::getSize() const { return m->getPtr()->getSize(); } diff --git a/paddle/legacy/api/ParameterOptimizer.cpp b/paddle/legacy/api/ParameterOptimizer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..477d9dae44362f9073639093c3c4d1cf0ac12044 --- /dev/null +++ b/paddle/legacy/api/ParameterOptimizer.cpp @@ -0,0 +1,124 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/parameter/ParameterOptimizer.h" +#include +#include "Internal.h" +#include "PaddleAPI.h" +#include "PaddleAPIPrivate.h" + +struct ParameterOptimizerPrivate { + std::unique_ptr optimizer; +}; + +struct ParameterTraverseCallbackPrivate { + paddle::ParameterOptimizer::TraverseCallback callback; + + ParameterTraverseCallbackPrivate() {} + + ParameterTraverseCallbackPrivate( + const paddle::ParameterOptimizer::TraverseCallback& callback) + : callback(callback) {} + + void apply(const std::vector& vecs, + const ParameterConfig& conf, + size_t sparseId) { + std::vector real_vecs; + real_vecs.resize(vecs.size()); + std::transform(vecs.begin(), vecs.end(), real_vecs.begin(), [](Vector* v) { + if (v) { + return *(paddle::VectorPtr*)(v->getSharedPtr()); + } else { + return paddle::VectorPtr(); + } + }); + + paddle::ParameterConfig& real_conf = + *(paddle::ParameterConfig*)(const_cast(conf) + .getRawPtr()); + callback(real_vecs.data(), real_conf, sparseId); + } +}; + +ParameterOptimizer::ParameterOptimizer() : m(new ParameterOptimizerPrivate()) {} + +ParameterOptimizer::~ParameterOptimizer() { delete m; } + +ParameterOptimizer* ParameterOptimizer::create(OptimizationConfig* config) { + CHECK(config != nullptr); + auto retOptimizer = new ParameterOptimizer(); + retOptimizer->m->optimizer.reset( + paddle::ParameterOptimizer::create(config->m->getConfig(), false)); + return retOptimizer; +} + +void ParameterOptimizer::init(size_t numRows, const ParameterConfig* config) { + auto& conf = *(paddle::ParameterConfig*)(const_cast(config) + ->getRawPtr()); + m->optimizer->init(numRows, &conf); +} + +void ParameterOptimizer::startPass() { m->optimizer->startPass(); } + +void ParameterOptimizer::finishPass() { m->optimizer->finishPass(); } + +void ParameterOptimizer::startBatch(size_t numSamplesProcessed) { + constexpr size_t high_1 = 1UL << (sizeof(size_t) * 8 - 1); + CHECK_EQ(numSamplesProcessed & high_1, 0UL); // Safely cast. + m->optimizer->startBatch((int64_t)numSamplesProcessed); +} + +void ParameterOptimizer::finishBatch() { m->optimizer->finishBatch(); } + +void ParameterOptimizer::update(const std::vector& vecs, + const ParameterConfig& conf, + size_t sparseId) { + ParameterTraverseCallbackPrivate invoker( + [&](const paddle::VectorPtr _vecs[], + const paddle::ParameterConfig& config, + size_t sid = -1UL) { m->optimizer->update(_vecs, config, sid); }); + invoker.apply(vecs, conf, sparseId); +} + +std::vector ParameterOptimizer::getParameterTypes() const { + std::vector returnValue; + staticCastVector(&returnValue, m->optimizer->getParameterTypes()); + return returnValue; +} + +ParameterTraverseCallback::ParameterTraverseCallback() + : m(new ParameterTraverseCallbackPrivate()) {} + +ParameterTraverseCallback::~ParameterTraverseCallback() { delete m; } + +void ParameterTraverseCallback::apply(const std::vector& vecs, + const ParameterConfig& conf, + size_t sparseId) { + m->apply(vecs, conf, sparseId); +} + +ParameterTraverseCallback* ParameterOptimizer::needSpecialTraversal( + const ParameterConfig& config) const { + auto& param_config = + *(paddle::ParameterConfig*)const_cast(config) + .getRawPtr(); + auto callback = m->optimizer->needSpecialTraversal(param_config); + if (callback) { + auto retCallback = new ParameterTraverseCallback(); + retCallback->m->callback = callback; + return retCallback; + } else { + return nullptr; + } +} diff --git a/paddle/legacy/api/ParameterUpdater.cpp b/paddle/legacy/api/ParameterUpdater.cpp new file mode 100644 index 0000000000000000000000000000000000000000..44af3f4635f2bda07d0079faff0bbc1ec7ed3954 --- /dev/null +++ b/paddle/legacy/api/ParameterUpdater.cpp @@ -0,0 +1,99 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" + +#include "PaddleAPIPrivate.h" +#ifndef PADDLE_WITHOUT_GOLANG +#include "paddle/legacy/trainer/NewRemoteParameterUpdater.h" +#endif +#include "paddle/legacy/trainer/RemoteParameterUpdater.h" +#include "paddle/legacy/trainer/ThreadParameterUpdater.h" + +ParameterUpdater::ParameterUpdater() : m(new ParameterUpdaterPrivate()) {} + +ParameterUpdater *ParameterUpdater::createLocalUpdater( + OptimizationConfig *config) { + auto updater = new ParameterUpdater(); + updater->m->updater.reset( + new paddle::SgdThreadUpdater(config->m->getConfig())); + return updater; +} + +ParameterUpdater *ParameterUpdater::createNewRemoteUpdater( + OptimizationConfig *config, + const std::string pserverSpec, + const bool useEtcd) throw(UnsupportError) { +#ifndef PADDLE_WITHOUT_GOLANG + auto updater = new ParameterUpdater(); + updater->m->updater.reset(new paddle::NewRemoteParameterUpdater( + config->m->getConfig(), pserverSpec, useEtcd)); + return updater; +#else + throw UnsupportError("not compiled with WITH_GOLANG"); +#endif +} + +ParameterUpdater *ParameterUpdater::createRemoteUpdater( + OptimizationConfig *config, int passCount, bool useSparseUpdater) { + auto updater = new ParameterUpdater(); + auto remoteUpdater = new paddle::RemoteParameterUpdater( + config->m->getConfig(), passCount, nullptr); + if (useSparseUpdater) { + std::unique_ptr remoteUpdaterPtr(remoteUpdater); + auto sparseRemoteUpdater = + new paddle::SparseRemoteParameterUpdaterComposite( + config->m->getConfig(), + passCount, + false, + std::move(remoteUpdaterPtr)); + updater->m->updater.reset(sparseRemoteUpdater); + } else { + updater->m->updater.reset(remoteUpdater); + } + return updater; +} + +ParameterUpdater::~ParameterUpdater() { delete m; } + +void ParameterUpdater::init(const GradientMachine &gm) { + m->updater->init(gm.m->machine->getNonStaticParameters()); +} + +void ParameterUpdater::startPass() { m->updater->startPass(); } + +void ParameterUpdater::finishPass() { m->updater->finishPass(); } + +PassType ParameterUpdater::startBatch(size_t batchSize) { + return m->updater->startBatch((int64_t)batchSize); +} + +void ParameterUpdater::finishBatch(float cost) { + m->updater->finishBatch(cost); +} + +void ParameterUpdater::update(Parameter *param) { + auto paddleParam = param->m->getPtr(); + m->updater->update(paddleParam); +} + +void ParameterUpdater::getParametersRemote(bool fullSize, bool apply) { + m->updater->getParametersRemote(fullSize, apply); +} + +void ParameterUpdater::restore() { m->updater->restore(); } + +void ParameterUpdater::apply() { m->updater->apply(); } + +void ParameterUpdater::catchUpWith() { m->updater->catchUpWith(); } diff --git a/paddle/legacy/api/SequenceGenerator.cpp b/paddle/legacy/api/SequenceGenerator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2a73228f6d4770d9be31defd7a5dc217fc5c21f2 --- /dev/null +++ b/paddle/legacy/api/SequenceGenerator.cpp @@ -0,0 +1,242 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include "PaddleAPI.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" +#include "paddle/legacy/parameter/Argument.h" +#include "paddle/legacy/utils/Flags.h" + +// used to represent partial sequence +struct Path { + std::vector ids; + float logProb; + paddle::MachineState machineState; + + Path() { logProb = 0; } + + Path(std::vector& ids, float logProb, paddle::MachineState& machineState) + : ids(ids), logProb(logProb), machineState(machineState) {} + + bool operator<(const Path& other) const { return (logProb > other.logProb); } +}; + +// Return top k (k == beam_size) optimal paths using beam search. The last +// element of inArgs is the Argument of feedback. gradMachine has MaxIdLayer +// as output and outArgs thus stores top k labels and their probabilities per +// position +static void findNBest(paddle::GradientMachine* gradMachine, + std::vector& inArgs, + std::vector& finalPaths, + size_t bos_id, + size_t eos_id, + size_t max_length) { + std::vector paths; + Path emptyPath; + paths.push_back(emptyPath); + finalPaths.clear(); + gradMachine->resetState(); + paddle::Argument feedback = inArgs.back(); + feedback.ids->setElement(0, (int)(bos_id)); + float minFinalPathLogProb = 0; + size_t beam = 0; + int id; + std::vector outArgs; + while (true) { // iterate over each generated word + std::vector newPaths; + paddle::MachineState machineState; + for (size_t j = 0; j < paths.size(); j++) { + Path& path = paths[j]; + if (path.machineState.size() > 0) { + gradMachine->setState(path.machineState); + feedback.ids->setElement(0, path.ids.back()); + } + gradMachine->forward(inArgs, &outArgs, paddle::PASS_TEST); + gradMachine->getState(machineState); + beam = outArgs[0].ids->getSize(); + for (size_t k = 0; k < beam; k++) { + id = outArgs[0].ids->getElement(k); + float prob = outArgs[0].in->getElement(0, k); + std::vector nids(path.ids); + nids.push_back(id); + float newLogProb = path.logProb + log(prob); + Path newPath(nids, newLogProb, machineState); + if (id == (int)eos_id || nids.size() >= max_length) { + finalPaths.push_back(newPath); + if (minFinalPathLogProb > newPath.logProb) { + minFinalPathLogProb = newPath.logProb; + } + } else { + newPaths.push_back(newPath); + } + } + } + + if (newPaths.size() == 0) { + break; + } + std::nth_element(newPaths.begin(), + newPaths.begin() + std::min(beam, newPaths.size()), + newPaths.end()); + if (newPaths.size() > beam) { + newPaths.resize(beam); + } + // pathA < pathB means pathA.logProb > pathB.logProb + float maxPathLogProb = + std::min_element(newPaths.begin(), newPaths.end())->logProb; + if (finalPaths.size() >= beam && minFinalPathLogProb >= maxPathLogProb) { + break; + } + paths = newPaths; + } // end while + + std::partial_sort(finalPaths.begin(), + finalPaths.begin() + std::min(beam, finalPaths.size()), + finalPaths.end()); + if (finalPaths.size() > beam) { + finalPaths.resize(beam); + } +} + +struct SequenceGeneratorPrivate { + std::shared_ptr machine; + std::shared_ptr> dict; + size_t beginPos; + size_t endPos; + size_t maxLength; + + paddle::Argument feedback; + + template + inline T& cast(void* ptr) { + return *(T*)(ptr); + } + + inline void findNBest(std::vector& inArgs, + std::vector& path) { + ::findNBest(machine.get(), inArgs, path, beginPos, endPos, maxLength); + } + + SequenceGeneratorPrivate() + : dict(std::make_shared>()), + beginPos(0UL), + endPos(0UL), + maxLength(0UL), + feedback(__create_feedback__()) {} + + private: + static paddle::Argument __create_feedback__() { + paddle::Argument feedback; + feedback.ids = paddle::IVector::create(/* size= */ 1, FLAGS_use_gpu); + + feedback.sequenceStartPositions = + paddle::ICpuGpuVector::create(/* size= */ 2, /* useGpu= */ false); + feedback.sequenceStartPositions->getMutableData(false)[0] = 0; + feedback.sequenceStartPositions->getMutableData(false)[1] = 1; + return feedback; + } +}; + +SequenceGenerator::SequenceGenerator() : m(new SequenceGeneratorPrivate()) {} + +SequenceGenerator::~SequenceGenerator() { delete m; } + +class PathSequenceResults : public ISequenceResults { + // ISequenceResults interface + public: + PathSequenceResults(const std::shared_ptr>& path, + const std::shared_ptr>& dict) + : path_(path), dict_(dict) {} + + size_t getSize() const { return path_->size(); } + std::string getSentence(size_t id, bool split) const throw(RangeError) { + if (id < getSize()) { + Path& p = (*path_)[id]; + std::ostringstream sout; + std::transform(p.ids.begin(), + p.ids.end(), + std::ostream_iterator(sout, split ? " " : ""), + [&](int id) { return (*dict_)[id]; }); + return sout.str(); + } else { + RangeError e; + throw e; + } + } + std::vector getSequence(size_t id) const throw(RangeError) { + if (id < getSize()) { + Path& p = (*path_)[id]; + return p.ids; + } else { + RangeError e; + throw e; + } + } + float getScore(size_t id) const throw(RangeError) { + if (id < getSize()) { + Path& p = (*path_)[id]; + return p.logProb; + } else { + RangeError e; + throw e; + } + } + + private: + std::shared_ptr> path_; + std::shared_ptr> dict_; +}; + +ISequenceResults* SequenceGenerator::generateSequence( + const Arguments& inArgs) const { + auto& in_args = + m->cast>(inArgs.getInternalArgumentsPtr()); + for (auto& arg : in_args) { + arg.sequenceStartPositions = m->feedback.sequenceStartPositions; + } + in_args.push_back(m->feedback); + auto path = std::make_shared>(); + m->findNBest(in_args, *path); + return new PathSequenceResults(path, m->dict); +} + +SequenceGenerator* SequenceGenerator::createByGradientMachineSharedPtr( + void* ptr) { + SequenceGenerator* r = new SequenceGenerator(); + r->m->machine = r->m->cast>(ptr); + return r; +} + +void SequenceGenerator::setDict(const std::vector& dict) { + *m->dict = dict; +} + +void SequenceGenerator::setBos(size_t bos) { m->beginPos = bos; } + +void SequenceGenerator::setEos(size_t eos) { m->endPos = eos; } + +void SequenceGenerator::setMaxLength(size_t maxLength) { + m->maxLength = maxLength; +} + +void SequenceGenerator::setBeamSize(size_t beamSize) { + if (beamSize != -1UL) { + FLAGS_beam_size = beamSize; + } +} + +ISequenceResults::~ISequenceResults() {} diff --git a/paddle/legacy/api/Trainer.cpp b/paddle/legacy/api/Trainer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e7c607201b0b946a6d6b2f3da35356e2c4e5e15e --- /dev/null +++ b/paddle/legacy/api/Trainer.cpp @@ -0,0 +1,175 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" +#include "PaddleAPIPrivate.h" + +#include +#include +#include + +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/trainer/ParamUtil.h" +#include "paddle/legacy/trainer/Trainer.h" +#include "paddle/legacy/trainer/TrainerInternal.h" +#include "paddle/legacy/utils/Flags.h" + +using paddle::real; + +DECLARE_string(config); +DECLARE_string(init_model_path); +DECLARE_int32(start_pass); + +struct TrainerPrivate : public paddle::Trainer { + bool _trainOneBatch(size_t batchSize); + bool forwardOneBatch(size_t batchSize); + void forwardOneDataBatch(const std::vector& inArgs); + void setBatchSize(size_t batchSize); + std::vector& getForwardOutput(); + + void startTestPeriod(); + void finishTestPeriod(); + void testOneDataBatch(const paddle::DataBatch& dataBatch); + TrainerPrivate() : paddle::Trainer() {} +}; + +Trainer::Trainer() : m(new TrainerPrivate()) { + auto conf = paddle::TrainerConfigHelper::createFromFlags(); + if (conf != nullptr) { + m->init(conf); + } +} + +Trainer::~Trainer() { delete m; } + +Trainer* Trainer::createByCommandLine() throw(IOError) { + auto retv = new Trainer(); + if (retv->m->getConfig().IsInitialized()) { + return retv; + } else { + throw IOError(); + } +} + +Trainer::Trainer(TrainerConfig* config, GradientMachine* gm) + : m(new TrainerPrivate()) { + m->init(config->m->conf, /* testing= */ false, gm ? gm->m->machine : nullptr); +} + +Trainer* Trainer::create(TrainerConfig* config, + GradientMachine* gm) throw(IOError) { + auto retv = new Trainer(config, gm); + if (retv->m->getConfig().IsInitialized()) { + return retv; + } else { + retv->m->getConfig().CheckInitialized(); + throw IOError(); + } +} + +void Trainer::startTrain() { m->startTrain(); } + +void Trainer::finishTrain() { m->finishTrain(); } + +void Trainer::startTrainPass() { m->startTrainPass(); } + +void Trainer::finishTrainPass() { m->finishTrainPass(); } + +void Trainer::trainOneDataBatch(size_t batchSize, const Arguments& inArgs) { + paddle::DataBatch dataBatch; + dataBatch.getStreams() = inArgs.m->outputs; + dataBatch.setSize(batchSize); + m->trainOneDataBatch(dataBatch); +} + +bool Trainer::trainOneBatch(size_t batchSize) { + return m->_trainOneBatch(batchSize); +} + +bool TrainerPrivate::_trainOneBatch(size_t batchSize) { + paddle::DataBatch dataBatch; + CHECK(dataProvider_) << "data_provider is not specified"; + int num = dataProvider_->getNextBatch(batchSize, &dataBatch); + if (num == 0) { + return false; + } + trainOneDataBatch(dataBatch); + return false; +} + +void TrainerPrivate::startTestPeriod() { + if (!tester_) { + createTester(); + } + tester_->startTestPeriod(); +} + +void Trainer::startTestPeriod() { m->startTestPeriod(); } + +void TrainerPrivate::testOneDataBatch(const paddle::DataBatch& dataBatch) { + tester_->testOneDataBatch(dataBatch, &forwardOutput_); +} + +void Trainer::testOneDataBatch(size_t batchSize, const Arguments& args) { + paddle::DataBatch dataBatch; + dataBatch.getStreams() = args.m->outputs; + dataBatch.setSize(batchSize); + m->testOneDataBatch(dataBatch); +} + +void TrainerPrivate::finishTestPeriod() { tester_->finishTestPeriod(); } +void Trainer::finishTestPeriod() { m->finishTestPeriod(); } + +Arguments* Trainer::getLayerOutput(const std::string& layerName) const { + auto nn = this->m->getGradientMachine(); + CHECK(nn) << "trainerInternal_.getGradientMachine() is not NeuralNetwork"; + auto arg = nn->getLayerOutput(layerName); + return Arguments::createByPaddleArgument(&arg); +} + +void Trainer::forwardOneBatch(size_t batchSize) { + m->forwardOneBatch(batchSize); +} + +bool TrainerPrivate::forwardOneBatch(size_t batchSize) { + CHECK(dataProvider_) << "data_provider is not specified"; + paddle::DataBatch dataBatch; + int num = dataProvider_->getNextBatch(batchSize, &dataBatch); + if (num == 0) { + return false; + } + + forwardOneDataBatch(dataBatch.getStreams()); + return true; +} + +void TrainerPrivate::forwardOneDataBatch( + const std::vector& inArgs) { + std::vector& outArgs = forwardOutput_; + + if (config_->getOptConfig().use_sparse_remote_updater()) { + trainerInternal_.getGradientMachine()->prefetch(inArgs); + trainerInternal_.getParameterUpdater()->getParametersRemote(); + } + trainerInternal_.getGradientMachine()->forward( + inArgs, &outArgs, paddle::PASS_TEST); +} + +Arguments* Trainer::getForwardOutput() { + return Arguments::createByPaddleArgumentVector(&m->getForwardOutput()); +} + +std::vector& TrainerPrivate::getForwardOutput() { + return forwardOutput_; +} diff --git a/paddle/legacy/api/Util.cpp b/paddle/legacy/api/Util.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b458c4d90ecc7333066f887dcbc93c4da5c43853 --- /dev/null +++ b/paddle/legacy/api/Util.cpp @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" + +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Util.h" + +#include +#include +#include + +void initPaddle(int argc, char** argv) { + paddle::initMain(argc, argv); + paddle::initPython(argc, argv); + feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); +} + +FloatArray::FloatArray(const float* b, const size_t l) + : buf(b), length(l), needFree(false) {} + +IntArray::IntArray(const int* b, const size_t l, bool f) + : buf(b), length(l), needFree(f) {} + +IntWithFloatArray::IntWithFloatArray(const float* v, + const int* i, + size_t l, + bool f) + : valBuf(v), idxBuf(i), length(l), needFree(f) {} + +bool isUsingGpu() { return FLAGS_use_gpu; } + +void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; } + +bool isGpuVersion() { +#ifndef PADDLE_WITH_CUDA + return false; +#else + return true; +#endif +} + +int getTrainerCount() { return FLAGS_trainer_count; } + +static_assert(NUM_PARAMETER_TYPES == paddle::NUM_PARAMETER_TYPES, + "The Parameter Type should be same in core/api and core/common"); diff --git a/paddle/legacy/api/Vector.cpp b/paddle/legacy/api/Vector.cpp new file mode 100644 index 0000000000000000000000000000000000000000..73b6d3a15d6d0ddc80a17846604d9500d8f7e4e3 --- /dev/null +++ b/paddle/legacy/api/Vector.cpp @@ -0,0 +1,304 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PaddleAPI.h" + +#include "paddle/legacy/math/Vector.h" + +#include + +struct IVectorPrivate { + paddle::IVectorPtr vec; +}; + +IVector::IVector() : m(new IVectorPrivate()) {} + +IVector* IVector::createZero(size_t sz, bool useGpu) { + auto v = new IVector(); + v->m->vec = paddle::IVector::create(sz, useGpu); + v->m->vec->zeroMem(); + return v; +} + +IVector* IVector::create(const std::vector& data, bool useGpu) { + auto v = new IVector(); + v->m->vec = paddle::IVector::create(data.size(), useGpu); + v->m->vec->copyFrom(data.data(), data.size()); + return v; +} + +IVector* IVector::createVectorFromNumpy(int* data, + int dim, + bool copy, + bool useGpu) throw(UnsupportError) { + if (useGpu) { + /// if use gpu only copy=true is supported + if (!copy) { + throw UnsupportError("Gpu mode only supports copy=True"); + } + return IVector::createGpuVectorFromNumpy(data, dim); + } else { + return IVector::createCpuVectorFromNumpy(data, dim, copy); + } +} + +IVector* IVector::createCpuVectorFromNumpy(int* data, int dim, bool copy) { + auto v = new IVector(); + if (copy) { + v->m->vec = paddle::IVector::create(dim, false); + v->m->vec->copyFrom(data, dim); + } else { + v->m->vec = paddle::IVector::create(data, dim, false); + } + return v; +} + +IVector* IVector::createGpuVectorFromNumpy(int* data, int dim) { + auto v = new IVector(); + v->m->vec = paddle::IVector::create(dim, true); + v->m->vec->copyFrom(data, dim); + return v; +} + +bool IVector::isGpu() const { + return dynamic_cast(m->vec.get()) != nullptr; +} + +IntArray IVector::getData() const { + if (this->isGpu()) { + int* src = m->vec->getData(); + size_t len = m->vec->getSize(); + int* dest = new int[len]; + hl_memcpy_device2host(dest, src, len * sizeof(int)); + return IntArray(dest, len, true); + } else { + return IntArray(m->vec->getData(), m->vec->getSize()); + } +} + +int& IVector::operator[](const size_t idx) throw(RangeError, UnsupportError) { + if (this->isGpu()) { + UnsupportError e; + throw e; + } else { + if (idx >= m->vec->getSize()) { + RangeError e; + throw e; + } + } + return m->vec->getData()[idx]; +} + +const int& IVector::operator[](const size_t idx) const + throw(RangeError, UnsupportError) { + return (*const_cast(this))[idx]; +} + +IVector* IVector::createByPaddleVectorPtr(void* ptr) { + auto* p = (paddle::IVectorPtr*)ptr; + if ((*p) != nullptr) { + IVector* vec = new IVector(); + vec->m->vec = *p; + return vec; + } else { + return nullptr; + } +} + +IVector::~IVector() { delete m; } + +void* IVector::getSharedPtr() const { return &m->vec; } + +size_t IVector::getSize() const { return m->vec->getSize(); } + +void IVector::toNumpyArrayInplace(int** data, int* dim1) throw(UnsupportError) { + auto v = std::dynamic_pointer_cast(m->vec); + if (v) { + *data = v->getData(); + *dim1 = v->getSize(); + } else { + throw UnsupportError(); + } +} + +void IVector::copyToNumpyArray(int** view_m_data, int* dim1) { + *dim1 = m->vec->getSize(); + *view_m_data = new int[*dim1]; + if (auto cpuVec = dynamic_cast(m->vec.get())) { + std::memcpy(*view_m_data, cpuVec->getData(), sizeof(int) * (*dim1)); + } else if (auto gpuVec = dynamic_cast(m->vec.get())) { + hl_memcpy_device2host( + *view_m_data, gpuVec->getData(), sizeof(int) * (*dim1)); + } else { + LOG(INFO) << "Unexpected situation"; + } +} + +void IVector::copyFromNumpyArray(int* data, int dim) { + m->vec->resize(dim); + m->vec->copyFrom(data, dim); +} + +struct VectorPrivate { + paddle::VectorPtr vec; + + void safeAccessData(const size_t idx, + const std::function& func) const + throw(RangeError, UnsupportError) { + auto cpuVec = std::dynamic_pointer_cast(vec); + if (cpuVec != nullptr) { + if (idx < vec->getSize()) { + func(vec->getData()[idx]); + } else { + throw RangeError(); + } + } else { + throw UnsupportError(); + } + } +}; + +Vector::Vector() : m(new VectorPrivate()) {} + +Vector::~Vector() { delete m; } + +Vector* Vector::createZero(size_t sz, bool useGpu) { + auto retVec = new Vector(); + retVec->m->vec = paddle::Vector::create(sz, useGpu); + retVec->m->vec->zero(); + return retVec; +} + +Vector* Vector::create(const std::vector& data, bool useGpu) { + auto retVec = new Vector(); + retVec->m->vec = paddle::Vector::create(data.size(), useGpu); + retVec->m->vec->copyFrom(data.data(), data.size()); + return retVec; +} + +Vector* Vector::createByPaddleVectorPtr(void* ptr) { + auto& v = *(paddle::VectorPtr*)(ptr); + if (v == nullptr) { + return nullptr; + } else { + auto retVec = new Vector(); + retVec->m->vec = v; + return retVec; + } +} + +Vector* Vector::createVectorFromNumpy(float* data, + int dim, + bool copy, + bool useGpu) throw(UnsupportError) { + if (useGpu) { + /// if use gpu only copy=True is supported + if (!copy) { + throw UnsupportError("Gpu mode only supports copy=True"); + } + return Vector::createGpuVectorFromNumpy(data, dim); + } else { + return Vector::createCpuVectorFromNumpy(data, dim, copy); + } +} + +Vector* Vector::createCpuVectorFromNumpy(float* data, int dim, bool copy) { + CHECK_GT(dim, 0); + auto retVec = new Vector(); + if (copy) { + retVec->m->vec = paddle::Vector::create((size_t)dim, false); + retVec->m->vec->copyFrom(data, dim); + } else { + retVec->m->vec = paddle::Vector::create(data, (size_t)dim, false); + } + return retVec; +} + +Vector* Vector::createGpuVectorFromNumpy(float* data, int dim) { + CHECK_GT(dim, 0); + auto retVec = new Vector(); + retVec->m->vec = paddle::Vector::create((size_t)dim, true); + retVec->m->vec->copyFrom(data, (size_t)dim); + return retVec; +} + +void Vector::toNumpyArrayInplace(float** view_data, + int* dim1) throw(UnsupportError) { + auto v = std::dynamic_pointer_cast(m->vec); + if (v != nullptr) { + *view_data = v->getData(); + *dim1 = (int)v->getSize(); + } else { + throw UnsupportError(); + } +} + +void Vector::copyToNumpyArray(float** view_m_data, int* dim1) { + *dim1 = m->vec->getSize(); + *view_m_data = new float[*dim1]; + if (auto cpuVec = dynamic_cast(m->vec.get())) { + std::memcpy(*view_m_data, cpuVec->getData(), sizeof(float) * (*dim1)); + } else if (auto gpuVec = dynamic_cast(m->vec.get())) { + hl_memcpy_device2host( + *view_m_data, gpuVec->getData(), sizeof(float) * (*dim1)); + } else { + LOG(INFO) << "Unexpected situation"; + } +} + +void Vector::copyFromNumpyArray(float* data, int dim) { + m->vec->resize(dim); + m->vec->copyFrom(data, dim); +} + +FloatArray Vector::getData() const { + if (this->isGpu()) { + float* src = m->vec->getData(); + size_t len = m->vec->getSize(); + float* dest = new float[len]; + hl_memcpy_device2host(dest, src, len * sizeof(float)); + FloatArray ret_val(dest, len); + ret_val.needFree = true; + return ret_val; + } else { + FloatArray ret_val(m->vec->getData(), m->vec->getSize()); + return ret_val; + } +} + +void Vector::copyFrom(Vector* src) throw(RangeError) { + if (src->m->vec->getSize() != m->vec->getSize()) { + throw RangeError(); + } + m->vec->copyFrom(*src->m->vec); +} + +bool Vector::isGpu() const { + return std::dynamic_pointer_cast(m->vec) != nullptr; +} + +float Vector::get(const size_t idx) const throw(RangeError, UnsupportError) { + float r; + m->safeAccessData(idx, [&](float& o) { r = o; }); + return r; +} + +void Vector::set(const size_t idx, float val) throw(RangeError, + UnsupportError) { + m->safeAccessData(idx, [&](float& o) { o = val; }); +} + +size_t Vector::getSize() const { return m->vec->getSize(); } + +void* Vector::getSharedPtr() { return &m->vec; } diff --git a/paddle/api/__init__.py b/paddle/legacy/api/__init__.py similarity index 100% rename from paddle/api/__init__.py rename to paddle/legacy/api/__init__.py diff --git a/paddle/api/numpy.i b/paddle/legacy/api/numpy.i similarity index 100% rename from paddle/api/numpy.i rename to paddle/legacy/api/numpy.i diff --git a/paddle/api/test/.gitignore b/paddle/legacy/api/test/.gitignore similarity index 100% rename from paddle/api/test/.gitignore rename to paddle/legacy/api/test/.gitignore diff --git a/paddle/api/test/CMakeLists.txt b/paddle/legacy/api/test/CMakeLists.txt similarity index 100% rename from paddle/api/test/CMakeLists.txt rename to paddle/legacy/api/test/CMakeLists.txt diff --git a/paddle/api/test/testArguments.py b/paddle/legacy/api/test/testArguments.py similarity index 100% rename from paddle/api/test/testArguments.py rename to paddle/legacy/api/test/testArguments.py diff --git a/paddle/api/test/testGradientMachine.py b/paddle/legacy/api/test/testGradientMachine.py similarity index 100% rename from paddle/api/test/testGradientMachine.py rename to paddle/legacy/api/test/testGradientMachine.py diff --git a/paddle/api/test/testMatrix.py b/paddle/legacy/api/test/testMatrix.py similarity index 100% rename from paddle/api/test/testMatrix.py rename to paddle/legacy/api/test/testMatrix.py diff --git a/paddle/api/test/testTrain.py b/paddle/legacy/api/test/testTrain.py similarity index 100% rename from paddle/api/test/testTrain.py rename to paddle/legacy/api/test/testTrain.py diff --git a/paddle/api/test/testTrainConfig.py b/paddle/legacy/api/test/testTrainConfig.py similarity index 100% rename from paddle/api/test/testTrainConfig.py rename to paddle/legacy/api/test/testTrainConfig.py diff --git a/paddle/api/test/testTrainer.py b/paddle/legacy/api/test/testTrainer.py similarity index 100% rename from paddle/api/test/testTrainer.py rename to paddle/legacy/api/test/testTrainer.py diff --git a/paddle/api/test/testVector.py b/paddle/legacy/api/test/testVector.py similarity index 100% rename from paddle/api/test/testVector.py rename to paddle/legacy/api/test/testVector.py diff --git a/paddle/api/test/util.py b/paddle/legacy/api/test/util.py similarity index 100% rename from paddle/api/test/util.py rename to paddle/legacy/api/test/util.py diff --git a/paddle/legacy/capi/Arguments.cpp b/paddle/legacy/capi/Arguments.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0ce1770c76c2e145d0b2bf71332cc4593517f195 --- /dev/null +++ b/paddle/legacy/capi/Arguments.cpp @@ -0,0 +1,140 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "arguments.h" +#include "capi_private.h" + +using paddle::capi::cast; + +#define castArg(v) cast(v) +#define castIVec(v) cast(v) + +extern "C" { +paddle_arguments paddle_arguments_create_none() { + return new paddle::capi::CArguments(); +} + +paddle_error paddle_arguments_destroy(paddle_arguments args) { + if (args == nullptr) return kPD_NULLPTR; + delete castArg(args); + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_get_size(paddle_arguments args, uint64_t* size) { + if (args == nullptr || size == nullptr) return kPD_NULLPTR; + *size = castArg(args)->args.size(); + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_resize(paddle_arguments args, uint64_t size) { + if (args == nullptr) return kPD_NULLPTR; + castArg(args)->args.resize(size); + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_set_value(paddle_arguments args, + uint64_t ID, + paddle_matrix mat) { + if (args == nullptr || mat == nullptr) return kPD_NULLPTR; + auto m = paddle::capi::cast(mat); + if (m->mat == nullptr) return kPD_NULLPTR; + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + a->args[ID].value = m->mat; + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_get_value(paddle_arguments args, + uint64_t ID, + paddle_matrix mat) { + if (args == nullptr || mat == nullptr) return kPD_NULLPTR; + auto m = paddle::capi::cast(mat); + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + m->mat = a->args[ID].value; + return kPD_NO_ERROR; +} + +PD_API paddle_error paddle_arguments_get_prob(paddle_arguments args, + uint64_t ID, + paddle_matrix mat) { + if (args == nullptr || mat == nullptr) return kPD_NULLPTR; + auto m = paddle::capi::cast(mat); + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + m->mat = a->args[ID].in; + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_get_ids(paddle_arguments args, + uint64_t ID, + paddle_ivector ids) { + if (args == nullptr || ids == nullptr) return kPD_NULLPTR; + auto iv = castIVec(ids); + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + iv->vec = a->args[ID].ids; + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_set_ids(paddle_arguments args, + uint64_t ID, + paddle_ivector ids) { + //! TODO(lizhao): Complete this method. + if (args == nullptr || ids == nullptr) return kPD_NULLPTR; + auto iv = paddle::capi::cast(ids); + if (iv->vec == nullptr) return kPD_NULLPTR; + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + a->args[ID].ids = iv->vec; + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_set_frame_shape(paddle_arguments args, + uint64_t ID, + uint64_t frameHeight, + uint64_t frameWidth) { + if (args == nullptr) return kPD_NULLPTR; + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + a->args[ID].setFrameHeight(frameHeight); + a->args[ID].setFrameWidth(frameWidth); + return kPD_NO_ERROR; +} + +paddle_error paddle_arguments_set_sequence_start_pos(paddle_arguments args, + uint64_t ID, + uint32_t nestedLevel, + paddle_ivector seqPos) { + if (args == nullptr || seqPos == nullptr) return kPD_NULLPTR; + auto iv = paddle::capi::cast(seqPos); + if (iv->vec == nullptr) return kPD_NULLPTR; + auto a = castArg(args); + return a->accessSeqPos(ID, nestedLevel, [&iv](paddle::ICpuGpuVectorPtr& ptr) { + ptr = std::make_shared(iv->vec); + }); +} + +paddle_error paddle_arguments_get_sequence_start_pos(paddle_arguments args, + uint64_t ID, + uint32_t nestedLevel, + paddle_ivector seqPos) { + if (args == nullptr || seqPos == nullptr) return kPD_NULLPTR; + auto iv = paddle::capi::cast(seqPos); + auto a = castArg(args); + return a->accessSeqPos(ID, nestedLevel, [&iv](paddle::ICpuGpuVectorPtr& ptr) { + iv->vec = ptr->getMutableVector(false); + }); +} +} diff --git a/paddle/capi/CMakeLists.txt b/paddle/legacy/capi/CMakeLists.txt similarity index 100% rename from paddle/capi/CMakeLists.txt rename to paddle/legacy/capi/CMakeLists.txt diff --git a/paddle/legacy/capi/Main.cpp b/paddle/legacy/capi/Main.cpp new file mode 100644 index 0000000000000000000000000000000000000000..17d8f00a88a9fd0818e6b90f8f6888b7d793a46e --- /dev/null +++ b/paddle/legacy/capi/Main.cpp @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include "capi_private.h" +#include "main.h" +#include "paddle/legacy/trainer/TrainerConfigHelper.h" +#include "paddle/legacy/utils/Excepts.h" +#include "paddle/legacy/utils/PythonUtil.h" + +static void initPaddle(int argc, char** argv) { + paddle::initMain(argc, argv); + paddle::initPython(argc, argv); +} + +extern "C" { +paddle_error paddle_init(int argc, char** argv) { + static bool isInit = false; + if (isInit) return kPD_NO_ERROR; + + std::vector realArgv; + realArgv.reserve(argc + 1); + realArgv.push_back(strdup("")); + for (int i = 0; i < argc; ++i) { + realArgv.push_back(argv[i]); + } + initPaddle(argc + 1, realArgv.data()); + free(realArgv[0]); + isInit = true; + return kPD_NO_ERROR; +} + +paddle_error paddle_init_thread() { + if (FLAGS_use_gpu) { + hl_init(FLAGS_gpu_id); + } + return kPD_NO_ERROR; +} +} diff --git a/paddle/capi/Matrix.cpp b/paddle/legacy/capi/Matrix.cpp similarity index 100% rename from paddle/capi/Matrix.cpp rename to paddle/legacy/capi/Matrix.cpp diff --git a/paddle/capi/Vector.cpp b/paddle/legacy/capi/Vector.cpp similarity index 100% rename from paddle/capi/Vector.cpp rename to paddle/legacy/capi/Vector.cpp diff --git a/paddle/legacy/capi/arguments.h b/paddle/legacy/capi/arguments.h new file mode 100644 index 0000000000000000000000000000000000000000..ceb64ee6aa74a8ba4b5cb9045b366dcda8f8cc90 --- /dev/null +++ b/paddle/legacy/capi/arguments.h @@ -0,0 +1,171 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef __PADDLE_CAPI_ARGUMENTS_H__ +#define __PADDLE_CAPI_ARGUMENTS_H__ + +#include +#include "config.h" +#include "error.h" +#include "matrix.h" +#include "vector.h" + +/** + * Arguments functions. Each argument means layer output. Arguments means a + * array of arguemnt. + */ +typedef void* paddle_arguments; + +#ifdef __cplusplus +extern "C" { +#endif + +/** + * @brief paddle_arguments_create_none Create a array of arguments, which size + * is zero. + * @return Arguemnts + */ +PD_API paddle_arguments paddle_arguments_create_none(); + +/** + * @brief paddle_arguments_destroy Destroy the arguments + * @param args arguments to destroy + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_destroy(paddle_arguments args); + +/** + * @brief paddle_arguments_get_size Get size of arguments array + * @param [in] args arguments array + * @param [out] size array size + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_get_size(paddle_arguments args, + uint64_t* size); + +/** + * @brief PDArgsResize Resize a arguments array. + * @param args arguments array. + * @param size target size of array + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_resize(paddle_arguments args, + uint64_t size); + +/** + * @brief PDArgsSetValue Set value matrix of one argument in array, which index + * is `ID`. + * @param args arguments array + * @param ID array index + * @param mat matrix pointer + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_set_value(paddle_arguments args, + uint64_t ID, + paddle_matrix mat); + +/** + * @brief PDArgsGetValue Get value matrix of one argument in array, which index + * is `ID`. + * @param [in] args arguments array + * @param [in] ID array index + * @param [out] mat matrix pointer + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_get_value(paddle_arguments args, + uint64_t ID, + paddle_matrix mat); + +/** + * @brief paddle_arguments_get_prob Get the prob matrix of beam search, which + * slot ID is `ID` + * @param [in] args arguments array + * @param [in] ID array index + * @param [out] mat matrix pointer + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_get_prob(paddle_arguments args, + uint64_t ID, + paddle_matrix mat); + +/** + * @brief PDArgsGetIds Get the integer vector of one argument in array, which + * index is `ID`. + * @param args arguments array + * @param ID array index + * @param ids integer vector pointer + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_get_ids(paddle_arguments args, + uint64_t ID, + paddle_ivector ids); + +/** + * @brief PDArgsSetIds Set the integer vector of one argument in array, which + * index is `ID`. + * @param [in] args arguments array + * @param [in] ID array index + * @param [out] ids integer vector pointer + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_set_ids(paddle_arguments args, + uint64_t ID, + paddle_ivector ids); + +/** + * @brief paddle_arguments_set_frame_shape Set the fram size of one argument + * in array, which index is `ID`. + * @param [in] args arguments array + * @param [in] ID array index + * @param [in] frameHeight maximum height of input images + * @param [in] frameWidth maximum width of input images + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_set_frame_shape(paddle_arguments args, + uint64_t ID, + uint64_t frameHeight, + uint64_t frameWidth); + +/** + * @brief PDArgsSetSequenceStartPos Set sequence start position vector of one + * argument in array, which index is `ID`. + * @param args arguments array + * @param ID array index + * @param seqPos sequence position array. + * @return paddle_error + */ +PD_API paddle_error +paddle_arguments_set_sequence_start_pos(paddle_arguments args, + uint64_t ID, + uint32_t nestedLevel, + paddle_ivector seqPos); +/** + * @brief PDArgsGetSequenceStartPos Get sequence start position vector of one + * argument in array, which index is `ID`. + * @param [in] args arguments array + * @param [in] ID array index + * @param [out] seqPos sequence position array + * @return paddle_error + */ +PD_API paddle_error +paddle_arguments_get_sequence_start_pos(paddle_arguments args, + uint64_t ID, + uint32_t nestedLevel, + paddle_ivector seqPos); + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/paddle/capi/capi.h b/paddle/legacy/capi/capi.h similarity index 100% rename from paddle/capi/capi.h rename to paddle/legacy/capi/capi.h diff --git a/paddle/legacy/capi/capi_private.h b/paddle/legacy/capi/capi_private.h new file mode 100644 index 0000000000000000000000000000000000000000..e5f8c8c5c8bd506f9c8f49ee7d03f9b20460efdb --- /dev/null +++ b/paddle/legacy/capi/capi_private.h @@ -0,0 +1,82 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "capi.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/parameter/Argument.h" +#pragma once + +namespace paddle { +namespace capi { + +enum CType { kIVECTOR = 0, kMATRIX, kARGUMENTS, kGRADIENT_MACHINE }; + +#define STRUCT_HEADER CType type; + +struct CHeader { + STRUCT_HEADER +}; + +struct CIVector { + STRUCT_HEADER + IVectorPtr vec; + + CIVector() : type(kIVECTOR) {} +}; + +struct CMatrix { + STRUCT_HEADER + MatrixPtr mat; + + CMatrix() : type(kMATRIX) {} +}; + +struct CArguments { + STRUCT_HEADER + std::vector args; + + CArguments() : type(kARGUMENTS) {} + + template + paddle_error accessSeqPos(uint64_t ID, uint32_t nestedLevel, T callback) { + if (ID >= args.size()) return kPD_OUT_OF_RANGE; + switch (nestedLevel) { + case 0: + callback(args[ID].sequenceStartPositions); + break; + case 1: + callback(args[ID].subSequenceStartPositions); + break; + default: + return kPD_OUT_OF_RANGE; + } + return kPD_NO_ERROR; + } +}; + +struct CGradientMachine { + STRUCT_HEADER + paddle::GradientMachinePtr machine; + + CGradientMachine() : type(kGRADIENT_MACHINE) {} +}; + +template +inline T* cast(void* ptr) { + return reinterpret_cast(ptr); +} +} // namespace capi +} // namespace paddle diff --git a/paddle/capi/config.h.in b/paddle/legacy/capi/config.h.in similarity index 100% rename from paddle/capi/config.h.in rename to paddle/legacy/capi/config.h.in diff --git a/paddle/capi/error.cpp b/paddle/legacy/capi/error.cpp similarity index 100% rename from paddle/capi/error.cpp rename to paddle/legacy/capi/error.cpp diff --git a/paddle/capi/error.h b/paddle/legacy/capi/error.h similarity index 100% rename from paddle/capi/error.h rename to paddle/legacy/capi/error.h diff --git a/paddle/capi/examples/.gitignore b/paddle/legacy/capi/examples/.gitignore similarity index 100% rename from paddle/capi/examples/.gitignore rename to paddle/legacy/capi/examples/.gitignore diff --git a/paddle/capi/examples/README.md b/paddle/legacy/capi/examples/README.md similarity index 100% rename from paddle/capi/examples/README.md rename to paddle/legacy/capi/examples/README.md diff --git a/paddle/capi/examples/model_inference/README.md b/paddle/legacy/capi/examples/model_inference/README.md similarity index 100% rename from paddle/capi/examples/model_inference/README.md rename to paddle/legacy/capi/examples/model_inference/README.md diff --git a/paddle/capi/examples/model_inference/common/common.h b/paddle/legacy/capi/examples/model_inference/common/common.h similarity index 100% rename from paddle/capi/examples/model_inference/common/common.h rename to paddle/legacy/capi/examples/model_inference/common/common.h diff --git a/paddle/capi/examples/model_inference/dense/CMakeLists.txt b/paddle/legacy/capi/examples/model_inference/dense/CMakeLists.txt similarity index 100% rename from paddle/capi/examples/model_inference/dense/CMakeLists.txt rename to paddle/legacy/capi/examples/model_inference/dense/CMakeLists.txt diff --git a/paddle/capi/examples/model_inference/dense/convert_protobin.sh b/paddle/legacy/capi/examples/model_inference/dense/convert_protobin.sh similarity index 100% rename from paddle/capi/examples/model_inference/dense/convert_protobin.sh rename to paddle/legacy/capi/examples/model_inference/dense/convert_protobin.sh diff --git a/paddle/capi/examples/model_inference/dense/main.c b/paddle/legacy/capi/examples/model_inference/dense/main.c similarity index 100% rename from paddle/capi/examples/model_inference/dense/main.c rename to paddle/legacy/capi/examples/model_inference/dense/main.c diff --git a/paddle/capi/examples/model_inference/dense/merge_v2_model.py b/paddle/legacy/capi/examples/model_inference/dense/merge_v2_model.py similarity index 100% rename from paddle/capi/examples/model_inference/dense/merge_v2_model.py rename to paddle/legacy/capi/examples/model_inference/dense/merge_v2_model.py diff --git a/paddle/capi/examples/model_inference/dense/mnist_v2.py b/paddle/legacy/capi/examples/model_inference/dense/mnist_v2.py similarity index 100% rename from paddle/capi/examples/model_inference/dense/mnist_v2.py rename to paddle/legacy/capi/examples/model_inference/dense/mnist_v2.py diff --git a/paddle/capi/examples/model_inference/dense/trainer_config.py b/paddle/legacy/capi/examples/model_inference/dense/trainer_config.py similarity index 100% rename from paddle/capi/examples/model_inference/dense/trainer_config.py rename to paddle/legacy/capi/examples/model_inference/dense/trainer_config.py diff --git a/paddle/capi/examples/model_inference/multi_thread/.gitignore b/paddle/legacy/capi/examples/model_inference/multi_thread/.gitignore similarity index 100% rename from paddle/capi/examples/model_inference/multi_thread/.gitignore rename to paddle/legacy/capi/examples/model_inference/multi_thread/.gitignore diff --git a/paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt b/paddle/legacy/capi/examples/model_inference/multi_thread/CMakeLists.txt similarity index 100% rename from paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt rename to paddle/legacy/capi/examples/model_inference/multi_thread/CMakeLists.txt diff --git a/paddle/legacy/capi/examples/model_inference/multi_thread/convert_protobin.sh b/paddle/legacy/capi/examples/model_inference/multi_thread/convert_protobin.sh new file mode 100644 index 0000000000000000000000000000000000000000..b29f2cd21418ecbd2fb2ba626138e5aa11bf77f3 --- /dev/null +++ b/paddle/legacy/capi/examples/model_inference/multi_thread/convert_protobin.sh @@ -0,0 +1 @@ +../dense/convert_protobin.sh diff --git a/paddle/capi/examples/model_inference/multi_thread/main.c b/paddle/legacy/capi/examples/model_inference/multi_thread/main.c similarity index 100% rename from paddle/capi/examples/model_inference/multi_thread/main.c rename to paddle/legacy/capi/examples/model_inference/multi_thread/main.c diff --git a/paddle/capi/examples/model_inference/multi_thread/main_gpu.c b/paddle/legacy/capi/examples/model_inference/multi_thread/main_gpu.c similarity index 100% rename from paddle/capi/examples/model_inference/multi_thread/main_gpu.c rename to paddle/legacy/capi/examples/model_inference/multi_thread/main_gpu.c diff --git a/paddle/capi/examples/model_inference/multi_thread/trainer_config.py b/paddle/legacy/capi/examples/model_inference/multi_thread/trainer_config.py similarity index 100% rename from paddle/capi/examples/model_inference/multi_thread/trainer_config.py rename to paddle/legacy/capi/examples/model_inference/multi_thread/trainer_config.py diff --git a/paddle/capi/examples/model_inference/sequence/.gitignore b/paddle/legacy/capi/examples/model_inference/sequence/.gitignore similarity index 100% rename from paddle/capi/examples/model_inference/sequence/.gitignore rename to paddle/legacy/capi/examples/model_inference/sequence/.gitignore diff --git a/paddle/capi/examples/model_inference/sequence/CMakeLists.txt b/paddle/legacy/capi/examples/model_inference/sequence/CMakeLists.txt similarity index 100% rename from paddle/capi/examples/model_inference/sequence/CMakeLists.txt rename to paddle/legacy/capi/examples/model_inference/sequence/CMakeLists.txt diff --git a/paddle/legacy/capi/examples/model_inference/sequence/convert_protobin.sh b/paddle/legacy/capi/examples/model_inference/sequence/convert_protobin.sh new file mode 100644 index 0000000000000000000000000000000000000000..b29f2cd21418ecbd2fb2ba626138e5aa11bf77f3 --- /dev/null +++ b/paddle/legacy/capi/examples/model_inference/sequence/convert_protobin.sh @@ -0,0 +1 @@ +../dense/convert_protobin.sh diff --git a/paddle/capi/examples/model_inference/sequence/main.c b/paddle/legacy/capi/examples/model_inference/sequence/main.c similarity index 100% rename from paddle/capi/examples/model_inference/sequence/main.c rename to paddle/legacy/capi/examples/model_inference/sequence/main.c diff --git a/paddle/capi/examples/model_inference/sequence/trainer_config.py b/paddle/legacy/capi/examples/model_inference/sequence/trainer_config.py similarity index 100% rename from paddle/capi/examples/model_inference/sequence/trainer_config.py rename to paddle/legacy/capi/examples/model_inference/sequence/trainer_config.py diff --git a/paddle/capi/examples/model_inference/sparse_binary/.gitignore b/paddle/legacy/capi/examples/model_inference/sparse_binary/.gitignore similarity index 100% rename from paddle/capi/examples/model_inference/sparse_binary/.gitignore rename to paddle/legacy/capi/examples/model_inference/sparse_binary/.gitignore diff --git a/paddle/capi/examples/model_inference/sparse_binary/CMakeLists.txt b/paddle/legacy/capi/examples/model_inference/sparse_binary/CMakeLists.txt similarity index 100% rename from paddle/capi/examples/model_inference/sparse_binary/CMakeLists.txt rename to paddle/legacy/capi/examples/model_inference/sparse_binary/CMakeLists.txt diff --git a/paddle/legacy/capi/examples/model_inference/sparse_binary/convert_protobin.sh b/paddle/legacy/capi/examples/model_inference/sparse_binary/convert_protobin.sh new file mode 100644 index 0000000000000000000000000000000000000000..b29f2cd21418ecbd2fb2ba626138e5aa11bf77f3 --- /dev/null +++ b/paddle/legacy/capi/examples/model_inference/sparse_binary/convert_protobin.sh @@ -0,0 +1 @@ +../dense/convert_protobin.sh diff --git a/paddle/capi/examples/model_inference/sparse_binary/main.c b/paddle/legacy/capi/examples/model_inference/sparse_binary/main.c similarity index 100% rename from paddle/capi/examples/model_inference/sparse_binary/main.c rename to paddle/legacy/capi/examples/model_inference/sparse_binary/main.c diff --git a/paddle/capi/examples/model_inference/sparse_binary/trainer_config.py b/paddle/legacy/capi/examples/model_inference/sparse_binary/trainer_config.py similarity index 100% rename from paddle/capi/examples/model_inference/sparse_binary/trainer_config.py rename to paddle/legacy/capi/examples/model_inference/sparse_binary/trainer_config.py diff --git a/paddle/legacy/capi/gradient_machine.cpp b/paddle/legacy/capi/gradient_machine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0c5ddd856b5d374ae90d6c8ef898be52aa2e4e89 --- /dev/null +++ b/paddle/legacy/capi/gradient_machine.cpp @@ -0,0 +1,180 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "gradient_machine.h" +#include "capi_private.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" + +#define cast(v) paddle::capi::cast(v) + +enum GradientMatchineCreateMode { + CREATE_MODE_NORMAL = 0, + CREATE_MODE_TESTING = 4 +}; + +namespace paddle { + +class MyNeuralNetwork : public NeuralNetwork { + public: + MyNeuralNetwork(const std::string& name, NeuralNetwork* network) + : NeuralNetwork(name, network) {} +}; + +NeuralNetwork* newCustomNerualNetwork(const std::string& name, + NeuralNetwork* network) { + return new MyNeuralNetwork(name, network); +} +} // namespace paddle + +extern "C" { +paddle_error paddle_gradient_machine_create_for_inference( + paddle_gradient_machine* machine, void* modelConfigProtobuf, int size) { + if (modelConfigProtobuf == nullptr) return kPD_NULLPTR; + paddle::ModelConfig config; + if (!config.ParseFromArray(modelConfigProtobuf, size) || + !config.IsInitialized()) { + return kPD_PROTOBUF_ERROR; + } + + auto ptr = new paddle::capi::CGradientMachine(); + ptr->machine.reset(paddle::GradientMachine::create( + config, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE})); + *machine = ptr; + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_create_for_inference_with_parameters( + paddle_gradient_machine* machine, void* mergedModel, uint64_t size) { + if (mergedModel == nullptr) return kPD_NULLPTR; + std::istringstream is(std::string(static_cast(mergedModel), size)); + int64_t modelConfigSize = 0; + is.read((char*)(&modelConfigSize), sizeof(modelConfigSize)); + std::string modelConfigProtobuf; + modelConfigProtobuf.resize(modelConfigSize); + is.read(&modelConfigProtobuf[0], modelConfigSize); + paddle::TrainerConfig config; + paddle::ModelConfig modelConfig; + if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) { + if (!modelConfig.ParseFromString(modelConfigProtobuf) || + !modelConfig.IsInitialized()) { + return kPD_PROTOBUF_ERROR; + } + } else { + modelConfig = config.model_config(); + } + auto ptr = new paddle::capi::CGradientMachine(); + ptr->machine.reset(paddle::GradientMachine::create( + modelConfig, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE})); + std::vector& parameters = ptr->machine->getParameters(); + for (auto& para : parameters) { + para->load(is); + } + + *machine = ptr; + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine) { + delete cast(machine); + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_load_parameter_from_disk( + paddle_gradient_machine machine, const char* path) { + auto m = cast(machine); + if (m == nullptr || path == nullptr || m->machine == nullptr) + return kPD_NULLPTR; + m->machine->loadParameters(path); + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_forward(paddle_gradient_machine machine, + paddle_arguments inArgs, + paddle_arguments outArgs, + bool isTrain) { + auto m = cast(machine); + auto in = paddle::capi::cast(inArgs); + auto out = paddle::capi::cast(outArgs); + if (m == nullptr || in == nullptr || out == nullptr || m->machine == nullptr) + return kPD_NULLPTR; + m->machine->forward( + in->args, &out->args, isTrain ? paddle::PASS_TRAIN : paddle::PASS_TEST); + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_create_shared_param( + paddle_gradient_machine origin, + void* modelConfigProtobuf, + int size, + paddle_gradient_machine* slave) { + auto o = cast(origin); + if (origin == nullptr || slave == nullptr || o->machine == nullptr) { + return kPD_NULLPTR; + } + paddle::ModelConfig config; + if (!config.ParseFromArray(modelConfigProtobuf, size) || + !config.IsInitialized()) { + return kPD_PROTOBUF_ERROR; + } + + std::unique_ptr ptr( + new paddle::capi::CGradientMachine()); + auto nn = paddle::NeuralNetwork::create(config); + nn->init(config, + [&o](int paramId, paddle::Parameter* param) { + auto p = o->machine->getParameters()[paramId]; + param->enableSharedType(paddle::PARAMETER_VALUE, + p->getBuf(paddle::PARAMETER_VALUE)); + }, + {paddle::PARAMETER_VALUE}, + false); + ptr->machine.reset(nn); + *slave = ptr.release(); + return kPD_NO_ERROR; +} +} + +paddle_error paddle_gradient_machine_randomize_param( + paddle_gradient_machine machine) { + auto m = cast(machine); + if (m == nullptr || m->machine == nullptr) return kPD_NULLPTR; + m->machine->randParameters(); + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_get_layer_output( + paddle_gradient_machine machine, + const char* layerName, + paddle_arguments args) { + auto m = cast(machine); + auto out = paddle::capi::cast(args); + if (m == nullptr || layerName == nullptr || out == nullptr || + m->machine == nullptr) { + return kPD_NULLPTR; + } + + auto layerOutput = m->machine->getLayerOutput(layerName); + out->args.push_back(layerOutput); + return kPD_NO_ERROR; +} + +paddle_error paddle_gradient_machine_release_layer_output( + paddle_gradient_machine machine) { + auto m = cast(machine); + if (m == nullptr || m->machine == nullptr) { + return kPD_NULLPTR; + } + m->machine->releaseOutput(); + return kPD_NO_ERROR; +} diff --git a/paddle/capi/gradient_machine.h b/paddle/legacy/capi/gradient_machine.h similarity index 100% rename from paddle/capi/gradient_machine.h rename to paddle/legacy/capi/gradient_machine.h diff --git a/paddle/capi/main.h b/paddle/legacy/capi/main.h similarity index 100% rename from paddle/capi/main.h rename to paddle/legacy/capi/main.h diff --git a/paddle/capi/matrix.h b/paddle/legacy/capi/matrix.h similarity index 100% rename from paddle/capi/matrix.h rename to paddle/legacy/capi/matrix.h diff --git a/paddle/capi/paddle_capi.map b/paddle/legacy/capi/paddle_capi.map similarity index 100% rename from paddle/capi/paddle_capi.map rename to paddle/legacy/capi/paddle_capi.map diff --git a/paddle/capi/tests/.gitignore b/paddle/legacy/capi/tests/.gitignore similarity index 100% rename from paddle/capi/tests/.gitignore rename to paddle/legacy/capi/tests/.gitignore diff --git a/paddle/capi/tests/CMakeLists.txt b/paddle/legacy/capi/tests/CMakeLists.txt similarity index 100% rename from paddle/capi/tests/CMakeLists.txt rename to paddle/legacy/capi/tests/CMakeLists.txt diff --git a/paddle/legacy/capi/tests/test_Arguments.cpp b/paddle/legacy/capi/tests/test_Arguments.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6fb379719dc0f3230c0801752720703ad185216f --- /dev/null +++ b/paddle/legacy/capi/tests/test_Arguments.cpp @@ -0,0 +1,129 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "capi.h" +#include "gtest/gtest.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +static std::vector randomBuffer(size_t bufSize) { + auto& eng = paddle::ThreadLocalRandomEngine::get(); + std::uniform_real_distribution dist(-1.0, 1.0); + std::vector retv; + retv.reserve(bufSize); + for (size_t i = 0; i < bufSize; ++i) { + retv.push_back(dist(eng)); + } + return retv; +} + +TEST(CAPIArguments, create) { + //! TODO(yuyang18): Test GPU Code. + paddle_arguments args = paddle_arguments_create_none(); + uint64_t size; + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_size(args, &size)); + ASSERT_EQ(0UL, size); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); +} + +TEST(CAPIArguments, value) { + paddle_arguments args = paddle_arguments_create_none(); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1)); + + paddle_matrix mat = paddle_matrix_create(128, 64, false); + for (size_t i = 0; i < 128; ++i) { + std::vector sampleBuf = randomBuffer(64); + paddle_matrix_set_row(mat, i, sampleBuf.data()); + } + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(args, 0, mat)); + + paddle_matrix val = paddle_matrix_create_none(); + + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_value(args, 0, val)); + + for (size_t i = 0; i < 128; ++i) { + paddle_real* row1; + paddle_real* row2; + + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, i, &row1)); + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(val, i, &row2)); + ASSERT_EQ(row1, row2); + } + + paddle_ivector ivec = paddle_ivector_create_none(); + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(ivec)); + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(val)); + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat)); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); +} + +TEST(CAPIArguments, ids) { + paddle_arguments args = paddle_arguments_create_none(); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1)); + + paddle_ivector ivec; + int array[3] = {1, 2, 3}; + ivec = paddle_ivector_create(array, 3, true, false); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_ids(args, 0, ivec)); + + paddle_ivector val = paddle_ivector_create_none(); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_ids(args, 0, val)); + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(ivec)); + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(val)); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); +} + +template +void testSequenceHelper(T1 setter, T2 getter) { + paddle_arguments args = paddle_arguments_create_none(); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(args, 1)); + + paddle_ivector ivec; + int array[3] = {1, 2, 3}; + ivec = paddle_ivector_create(array, 3, true, false); + ASSERT_EQ(kPD_NO_ERROR, setter(args, 0, ivec)); + + paddle_ivector val = paddle_ivector_create_none(); + ASSERT_EQ(kPD_NO_ERROR, getter(args, 0, val)); + uint64_t size; + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_get_size(val, &size)); + + int* rawBuf; + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_get(val, &rawBuf)); + for (size_t i = 0; i < size; ++i) { + ASSERT_EQ(array[i], rawBuf[i]); + } + + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(ivec)); + ASSERT_EQ(kPD_NO_ERROR, paddle_ivector_destroy(val)); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(args)); +} + +TEST(CAPIArguments, Sequence) { + auto testSequence = [](uint32_t nestedLevel) { + testSequenceHelper(std::bind(paddle_arguments_set_sequence_start_pos, + std::placeholders::_1, + std::placeholders::_2, + nestedLevel, + std::placeholders::_3), + std::bind(paddle_arguments_get_sequence_start_pos, + std::placeholders::_1, + std::placeholders::_2, + nestedLevel, + std::placeholders::_3)); + }; + for (uint32_t i = 0; i < 2; ++i) { // test seq and sub-seq. + testSequence(i); + } +} diff --git a/paddle/legacy/capi/tests/test_GradientMachine.cpp b/paddle/legacy/capi/tests/test_GradientMachine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5d1b7cb6ca4073c0a489366e415f8f74d3c19bec --- /dev/null +++ b/paddle/legacy/capi/tests/test_GradientMachine.cpp @@ -0,0 +1,117 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include "capi.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +static std::vector randomBuffer(size_t bufSize) { + auto& eng = paddle::ThreadLocalRandomEngine::get(); + std::uniform_real_distribution dist(-1.0, 1.0); + std::vector retv; + retv.reserve(bufSize); + for (size_t i = 0; i < bufSize; ++i) { + retv.push_back(dist(eng)); + } + return retv; +} + +TEST(GradientMachine, testPredict) { + //! TODO(yuyang18): Test GPU Code. + paddle::TrainerConfigHelper config("./test_predict_network.py"); + std::string buffer; + ASSERT_TRUE(config.getModelConfig().SerializeToString(&buffer)); + paddle_gradient_machine machine; + + ASSERT_EQ(kPD_NO_ERROR, + paddle_gradient_machine_create_for_inference( + &machine, &buffer[0], (int)buffer.size())); + std::unique_ptr gm( + paddle::GradientMachine::create(config.getModelConfig())); + ASSERT_NE(nullptr, gm); + gm->randParameters(); + gm->saveParameters("./"); + + ASSERT_EQ(kPD_NO_ERROR, + paddle_gradient_machine_load_parameter_from_disk(machine, "./")); + + paddle_gradient_machine machineSlave; + ASSERT_EQ(kPD_NO_ERROR, + paddle_gradient_machine_create_shared_param( + machine, &buffer[0], (int)buffer.size(), &machineSlave)); + std::swap(machineSlave, machine); + paddle_arguments outArgs = paddle_arguments_create_none(); + + paddle_arguments inArgs = paddle_arguments_create_none(); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(inArgs, 1)); + paddle_matrix mat = paddle_matrix_create(1, 100, false); + static_assert(std::is_same::value, ""); + + auto data = randomBuffer(100); + paddle_real* rowPtr; + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr)); + memcpy(rowPtr, data.data(), data.size() * sizeof(paddle_real)); + + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(inArgs, 0, mat)); + ASSERT_EQ(kPD_NO_ERROR, + paddle_gradient_machine_forward(machine, inArgs, outArgs, false)); + + uint64_t sz; + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_size(outArgs, &sz)); + ASSERT_EQ(1UL, sz); + + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_value(outArgs, 0, mat)); + std::vector paddleInArgs; + std::vector paddleOutArgs; + paddleInArgs.resize(1); + paddleInArgs[0].value = + paddle::Matrix::create(data.data(), 1, 100, false, false); + + gm->forward(paddleInArgs, &paddleOutArgs, paddle::PASS_TEST); + + auto matPaddle = paddleOutArgs[0].value; + + uint64_t height, width; + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width)); + ASSERT_EQ(matPaddle->getHeight(), height); + ASSERT_EQ(matPaddle->getWidth(), width); + + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr)); + for (size_t i = 0; i < width; ++i) { + ASSERT_NEAR(matPaddle->getData()[i], rowPtr[i], 1e-5); + } + + ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat)); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(inArgs)); + ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(outArgs)); + std::swap(machineSlave, machine); + ASSERT_EQ(kPD_NO_ERROR, paddle_gradient_machine_destroy(machineSlave)); + ASSERT_EQ(kPD_NO_ERROR, paddle_gradient_machine_destroy(machine)); +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + std::vector argvs; + argvs.push_back(strdup("--use_gpu=false")); + paddle_init((int)argvs.size(), argvs.data()); + for (auto each : argvs) { + free(each); + } + return RUN_ALL_TESTS(); +} diff --git a/paddle/capi/tests/test_Matrix.cpp b/paddle/legacy/capi/tests/test_Matrix.cpp similarity index 100% rename from paddle/capi/tests/test_Matrix.cpp rename to paddle/legacy/capi/tests/test_Matrix.cpp diff --git a/paddle/capi/tests/test_Vector.cpp b/paddle/legacy/capi/tests/test_Vector.cpp similarity index 100% rename from paddle/capi/tests/test_Vector.cpp rename to paddle/legacy/capi/tests/test_Vector.cpp diff --git a/paddle/capi/tests/test_predict_network.py b/paddle/legacy/capi/tests/test_predict_network.py similarity index 100% rename from paddle/capi/tests/test_predict_network.py rename to paddle/legacy/capi/tests/test_predict_network.py diff --git a/paddle/capi/vector.h b/paddle/legacy/capi/vector.h similarity index 100% rename from paddle/capi/vector.h rename to paddle/legacy/capi/vector.h diff --git a/paddle/cuda/CMakeLists.txt b/paddle/legacy/cuda/CMakeLists.txt similarity index 100% rename from paddle/cuda/CMakeLists.txt rename to paddle/legacy/cuda/CMakeLists.txt diff --git a/paddle/cuda/include/hl_activation_functions.h b/paddle/legacy/cuda/include/hl_activation_functions.h similarity index 100% rename from paddle/cuda/include/hl_activation_functions.h rename to paddle/legacy/cuda/include/hl_activation_functions.h diff --git a/paddle/cuda/include/hl_aggregate.h b/paddle/legacy/cuda/include/hl_aggregate.h similarity index 100% rename from paddle/cuda/include/hl_aggregate.h rename to paddle/legacy/cuda/include/hl_aggregate.h diff --git a/paddle/cuda/include/hl_avx_functions.h b/paddle/legacy/cuda/include/hl_avx_functions.h similarity index 100% rename from paddle/cuda/include/hl_avx_functions.h rename to paddle/legacy/cuda/include/hl_avx_functions.h diff --git a/paddle/legacy/cuda/include/hl_base.h b/paddle/legacy/cuda/include/hl_base.h new file mode 100644 index 0000000000000000000000000000000000000000..bfe812a4387be72c3e73d6b45852e3a90b1926eb --- /dev/null +++ b/paddle/legacy/cuda/include/hl_base.h @@ -0,0 +1,250 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#ifdef PADDLE_TYPE_DOUBLE +#define HL_FLOAT_MAX 3.40282347e+38F +#define HL_FLOAT_MIN 1.17549435e-38F +using real = double; +#else +#define HL_FLOAT_MAX 1.7976931348623157e+308 +#define HL_FLOAT_MIN 2.2250738585072014e-308 +using real = float; +#endif + +/** + * The maximum input value for exp, used to avoid overflow problem. + * currently only used for tanh function. + */ +#define EXP_MAX_INPUT 40.0 + +/** + * @brief DIVUP(x, y) is similar to ceil(x / y). + * @note For CUDA, DIVUP will be used to specify + * the size of blockDim. + */ +#ifndef DIVUP +#define DIVUP(x, y) (((x) + (y)-1) / (y)) +#endif + +/** + * HPPL is an internal high performance parallel computing library + * for high-level neural network routines, which can support many + * heterogeneous compute architectures, such as GPU, FPGA, etc. + */ + +/** + * @brief HPPL CUDA Stream. + * + * @note Each thread can use HPPL_STREAM_* after calling hl_init. + * HPPL_STREAM_DEFAULT is HPPL default stream. + */ +typedef enum { + HPPL_STREAM_DEFAULT = 0, /* Thread Default Stream*/ + HPPL_STREAM_1 = 1, + HPPL_STREAM_2 = 2, + HPPL_STREAM_3 = 3, + HPPL_STREAM_4 = 4, + HPPL_THREAD_STREAM_1 = 5, + HPPL_THREAD_STREAM_2 = 6, + HPPL_THREAD_STREAM_3 = 7, + HPPL_THREAD_STREAM_4 = 8, + HPPL_STREAM_END +} hl_stream_t; + +/** + * @brief HPPL activation mode. + */ +typedef enum { + HL_ACTIVATION_SIGMOID = 0, + HL_ACTIVATION_RELU = 1, + HL_ACTIVATION_TANH = 2, + HL_ACTIVATION_LINEAR = 3, + HL_ACTIVATION_END +} hl_activation_mode_t; + +/** + * @brief Transpose type. + */ +typedef enum { + HPPL_OP_N = 0, /* transpose */ + HPPL_OP_T = 1, /* non transpose */ + HPPL_OP_END +} hl_trans_op_t; + +/** + * @brief Lstm value. + * + * @param gateValue input value. + * @param prevStateValue previous state value. + * @param stateValue state value. + * @param stateActiveValue state active value. + * @param outputValue output value. + */ +typedef struct { + real *gateValue; + real *prevStateValue; + real *stateValue; + real *stateActiveValue; + real *outputValue; + real *checkIg; + real *checkFg; + real *checkOg; +} hl_lstm_value; + +/** + * @brief Lstm gradient. + * + * @param gateGrad input gradient. + * @param prevStateGrad previous state gradient. + * @param stateGrad state gradient. + * @param stateActiveGrad state active gradient. + * @param outputGrad output gradient. + */ +typedef struct { + real *gateGrad; + real *prevStateGrad; + real *stateGrad; + real *stateActiveGrad; + real *outputGrad; + real *checkIgGrad; + real *checkFgGrad; + real *checkOgGrad; +} hl_lstm_grad; + +/** + * @brief Gru value. + * + * @param gateWeight gate weight (updateGate + resetGate). + * @param stateWeight frame state weight. + * @param gateValue gate value results. + * @param resetOutputValue resetOutput value. + * @param outputValue output value. + * @param prevOutValue previous output value. + * + */ +typedef struct { + real *gateWeight; + real *stateWeight; + real *gateValue; + real *resetOutputValue; + real *outputValue; + real *prevOutValue; +} hl_gru_value; + +/** + * @brief Gru gradient. + * + * @param gateWeightGrad gate weight gradient. + * @param stateWeightGrad frame state weight gradient. + * @param gateGrad gate gradient results. + * @param resetOutputGrad resetOutput gradient. + * @param outputGrad output gradient. + * @param prevOutGrad previous output gradient. + */ +typedef struct { + real *gateWeightGrad; + real *stateWeightGrad; + real *gateGrad; + real *resetOutputGrad; + real *outputGrad; + real *prevOutGrad; +} hl_gru_grad; + +/** + * @brief Sparse matrix value type. + */ +typedef enum { + HL_NO_VALUE = 0, /* matrix values only 0 or 1 */ + HL_FLOAT_VALUE = 1, + HL_VALUE_END +} hl_matrix_value_t; + +/** + * @brief HPPL matrix format. + */ +typedef enum { + HL_SPARSE_CSR = 0, + HL_SPARSE_CSC = 1, + HL_SPARSE_END +} hl_matrix_format_t; + +typedef struct _hl_matrix_s *hl_matrix_s; + +/** + * @brief HPPL sparse matrix. + * + * @param matrix sparse matrix. + * @param format matrix format. + * @param type the type of matrix values. + * @param rows matrix rows. + * @param cols matrix columns. + * @param nnz nonzero values of sparse matrix. + */ +typedef struct { + hl_matrix_s matrix; + hl_matrix_format_t format; + hl_matrix_value_t type; + int rows; + int cols; + size_t nnz; +} _hl_sparse_matrix_s, *hl_sparse_matrix_s; + +#ifdef __NVCC__ + +#include +#include "paddle/legacy/cuda/include/hl_cuda.h" +#include "paddle/legacy/utils/Logging.h" + +extern __thread bool g_sync_flag; +extern __thread cudaStream_t default_stream; +#define STREAM_DEFAULT default_stream + +/** + * @brief Check cuda kernel execution. + * @param msg error string + */ +#define CHECK_SYNC(msg) \ + if (true == g_sync_flag) { \ + hl_stream_synchronize(HPPL_STREAM_DEFAULT); \ + cudaError_t err = (cudaError_t)hl_get_device_last_error(); \ + CHECK_EQ(cudaSuccess, err) \ + << "[" << msg << "] " \ + << "CUDA error: " << hl_get_device_error_string((size_t)err); \ + } + +// __shfl has been deprecated as of CUDA 9.0. +#if CUDA_VERSION < 9000 +template +__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { + return __shfl_down(val, delta); +} + +template +__forceinline__ __device__ T +__shfl_sync(unsigned, T val, int src_line, int width) { + return __shfl(val, src_line, width); +} + +#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; +#else +#define FULL_WARP_MASK 0xFFFFFFFF +#define CREATE_SHFL_MASK(mask, predicate) \ + mask = __ballot_sync(FULL_WARP_MASK, (predicate)) +#endif + +#endif // __NVCC__ diff --git a/paddle/cuda/include/hl_batch_norm.h b/paddle/legacy/cuda/include/hl_batch_norm.h similarity index 100% rename from paddle/cuda/include/hl_batch_norm.h rename to paddle/legacy/cuda/include/hl_batch_norm.h diff --git a/paddle/cuda/include/hl_batch_transpose.h b/paddle/legacy/cuda/include/hl_batch_transpose.h similarity index 100% rename from paddle/cuda/include/hl_batch_transpose.h rename to paddle/legacy/cuda/include/hl_batch_transpose.h diff --git a/paddle/cuda/include/hl_cnn.h b/paddle/legacy/cuda/include/hl_cnn.h similarity index 100% rename from paddle/cuda/include/hl_cnn.h rename to paddle/legacy/cuda/include/hl_cnn.h diff --git a/paddle/cuda/include/hl_cpu_gru.cuh b/paddle/legacy/cuda/include/hl_cpu_gru.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_gru.cuh rename to paddle/legacy/cuda/include/hl_cpu_gru.cuh diff --git a/paddle/cuda/include/hl_cpu_lstm.cuh b/paddle/legacy/cuda/include/hl_cpu_lstm.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_lstm.cuh rename to paddle/legacy/cuda/include/hl_cpu_lstm.cuh diff --git a/paddle/cuda/include/hl_cpu_matrix_kernel.cuh b/paddle/legacy/cuda/include/hl_cpu_matrix_kernel.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_matrix_kernel.cuh rename to paddle/legacy/cuda/include/hl_cpu_matrix_kernel.cuh diff --git a/paddle/cuda/include/hl_cpu_matrix_kernel_detail.cuh b/paddle/legacy/cuda/include/hl_cpu_matrix_kernel_detail.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_matrix_kernel_detail.cuh rename to paddle/legacy/cuda/include/hl_cpu_matrix_kernel_detail.cuh diff --git a/paddle/cuda/include/hl_cpu_scalar.cuh b/paddle/legacy/cuda/include/hl_cpu_scalar.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_scalar.cuh rename to paddle/legacy/cuda/include/hl_cpu_scalar.cuh diff --git a/paddle/cuda/include/hl_cpu_simd_neon.cuh b/paddle/legacy/cuda/include/hl_cpu_simd_neon.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_simd_neon.cuh rename to paddle/legacy/cuda/include/hl_cpu_simd_neon.cuh diff --git a/paddle/cuda/include/hl_cpu_simd_sse.cuh b/paddle/legacy/cuda/include/hl_cpu_simd_sse.cuh similarity index 100% rename from paddle/cuda/include/hl_cpu_simd_sse.cuh rename to paddle/legacy/cuda/include/hl_cpu_simd_sse.cuh diff --git a/paddle/cuda/include/hl_cuda.h b/paddle/legacy/cuda/include/hl_cuda.h similarity index 100% rename from paddle/cuda/include/hl_cuda.h rename to paddle/legacy/cuda/include/hl_cuda.h diff --git a/paddle/cuda/include/hl_cuda.ph b/paddle/legacy/cuda/include/hl_cuda.ph similarity index 100% rename from paddle/cuda/include/hl_cuda.ph rename to paddle/legacy/cuda/include/hl_cuda.ph diff --git a/paddle/cuda/include/hl_cuda_cublas.h b/paddle/legacy/cuda/include/hl_cuda_cublas.h similarity index 100% rename from paddle/cuda/include/hl_cuda_cublas.h rename to paddle/legacy/cuda/include/hl_cuda_cublas.h diff --git a/paddle/cuda/include/hl_cuda_cudnn.h b/paddle/legacy/cuda/include/hl_cuda_cudnn.h similarity index 100% rename from paddle/cuda/include/hl_cuda_cudnn.h rename to paddle/legacy/cuda/include/hl_cuda_cudnn.h diff --git a/paddle/cuda/include/hl_cuda_cudnn.ph b/paddle/legacy/cuda/include/hl_cuda_cudnn.ph similarity index 100% rename from paddle/cuda/include/hl_cuda_cudnn.ph rename to paddle/legacy/cuda/include/hl_cuda_cudnn.ph diff --git a/paddle/cuda/include/hl_device_functions.cuh b/paddle/legacy/cuda/include/hl_device_functions.cuh similarity index 100% rename from paddle/cuda/include/hl_device_functions.cuh rename to paddle/legacy/cuda/include/hl_device_functions.cuh diff --git a/paddle/cuda/include/hl_functions.h b/paddle/legacy/cuda/include/hl_functions.h similarity index 100% rename from paddle/cuda/include/hl_functions.h rename to paddle/legacy/cuda/include/hl_functions.h diff --git a/paddle/cuda/include/hl_gpu.h b/paddle/legacy/cuda/include/hl_gpu.h similarity index 100% rename from paddle/cuda/include/hl_gpu.h rename to paddle/legacy/cuda/include/hl_gpu.h diff --git a/paddle/cuda/include/hl_gpu_functions.cuh b/paddle/legacy/cuda/include/hl_gpu_functions.cuh similarity index 100% rename from paddle/cuda/include/hl_gpu_functions.cuh rename to paddle/legacy/cuda/include/hl_gpu_functions.cuh diff --git a/paddle/legacy/cuda/include/hl_gpu_gru.cuh b/paddle/legacy/cuda/include/hl_gpu_gru.cuh new file mode 100644 index 0000000000000000000000000000000000000000..8d299572c73e879a3a1e9fb60608c4f3abd1f685 --- /dev/null +++ b/paddle/legacy/cuda/include/hl_gpu_gru.cuh @@ -0,0 +1,393 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + + +#ifndef HL_GPU_GRU_CUH_ +#define HL_GPU_GRU_CUH_ + +#ifdef __NVCC__ + +#include "paddle/legacy/utils/Logging.h" + +/* + * threads(framePerBlock, batchPerBlock) + * grid(frameBlocks, batchBlocks) + */ +template +__global__ void KeGruForwardResetOutput(OpResetOutput opResetOutput, + real *gateValue, + real *resetOutputValue, + real *prevOutputValue, + int frameSize, + int batchSize, + hl_activation_mode_t active_gate) { + const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (frameIdx >= frameSize) return; + + int batchIdx = 0; + if (isBatch) { + batchIdx = blockIdx.y * blockDim.y + threadIdx.y; + if (batchIdx >= batchSize) return; + gateValue += batchIdx * 3 * frameSize; + resetOutputValue += batchIdx * frameSize; + } + + real rPrevOut = 0; + real rValueResetOutput; + real rValueUpdateGate = gateValue[frameIdx + frameSize * 0]; + real rValueResetGate = gateValue[frameIdx + frameSize * 1]; + + if (prevOutputValue) { + if (isBatch) prevOutputValue += batchIdx * frameSize; + rPrevOut = prevOutputValue[frameIdx]; + } + + opResetOutput(rValueUpdateGate, + rValueResetGate, + rPrevOut, + rValueResetOutput, + hppl::gpu::forward[active_gate]); + + gateValue[frameIdx + frameSize * 0] = rValueUpdateGate; + gateValue[frameIdx + frameSize * 1] = rValueResetGate; + resetOutputValue[frameIdx] = rValueResetOutput; +} + +/* + * threads(framePerBlock, batchPerBlock) + * grid(frameBlocks, batchBlocks) + */ +template +__global__ void KeGruForwardFinalOutput(OpFinalOutput opFinalOutput, + real *gateValue, + real *prevOutputValue, + real *outputValue, + int frameSize, + int batchSize, + hl_activation_mode_t active_node) { + const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (frameIdx >= frameSize) return; + int batchIdx = 0; + if (isBatch) { + batchIdx = blockIdx.y * blockDim.y + threadIdx.y; + if (batchIdx >= batchSize) return; + gateValue += batchIdx * 3 * frameSize; + outputValue += batchIdx * frameSize; + } + + real rOutput; + real rPrevOut = 0; + real rValueUpdateGate = gateValue[frameIdx + frameSize * 0]; + real rValueFrameState = gateValue[frameIdx + frameSize * 2]; + + if (prevOutputValue) { + if (isBatch) prevOutputValue += batchIdx * frameSize; + rPrevOut = prevOutputValue[frameIdx]; + } + + opFinalOutput(rValueUpdateGate, + rValueFrameState, + rPrevOut, + rOutput, + hppl::gpu::forward[active_node]); + + gateValue[frameIdx + frameSize * 2] = rValueFrameState; + outputValue[frameIdx] = rOutput; +} + +template +void hl_gpu_gru_forward(OpResetOutput opResetOutput, + OpFinalOutput opFinalOutput, + hl_gru_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) { + dim3 threads; + dim3 grid; + if (batchSize == 1) { + int framePerBlock = frameSize <= 1024 ? frameSize : 1024; + int frameBlocks = (frameSize + 1024 - 1) / 1024; + threads = dim3(framePerBlock, 1); + grid = dim3(frameBlocks, 1); + } else { + threads = dim3(32, 32); + grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + } + + if (value.prevOutValue) { + hl_matrix_mul(value.prevOutValue, HPPL_OP_N, + value.gateWeight, HPPL_OP_N, + value.gateValue, + batchSize, 2*frameSize, frameSize, + /*alpha = */ 1, /*beta = */ 1, + frameSize, 2* frameSize, 3*frameSize); + } + + if (batchSize == 1) { + KeGruForwardResetOutput + <<>>(opResetOutput, + value.gateValue, value.resetOutputValue, value.prevOutValue, + frameSize, batchSize, active_gate); + } else { + KeGruForwardResetOutput + <<>>(opResetOutput, + value.gateValue, value.resetOutputValue, value.prevOutValue, + frameSize, batchSize, active_gate); + } + + if (value.prevOutValue) { + hl_matrix_mul(value.resetOutputValue, HPPL_OP_N, + value.stateWeight, HPPL_OP_N, + value.gateValue + 2*frameSize, + batchSize, frameSize, frameSize, + /*alpha = */ 1, /*beta = */ 1, + frameSize, frameSize, 3*frameSize); + } + + if (batchSize == 1) { + KeGruForwardFinalOutput + <<>>(opFinalOutput, + value.gateValue, value.prevOutValue, value.outputValue, + frameSize, batchSize, active_node); + } else { + KeGruForwardFinalOutput + <<>>(opFinalOutput, + value.gateValue, value.prevOutValue, value.outputValue, + frameSize, batchSize, active_node); + } + + CHECK_SYNC("hl_gpu_gru_forward failed"); +} + +/* + * threads(framePerBlock, batchPerBlock) + * grid(frameBlocks, batchBlocks) + */ +template +__global__ void KeGruBackwardStateGrad(OpStateGrad opStateGrad, + real *gateValue, + real *gateGrad, + real *prevOutValue, + real *prevOutGrad, + real *outputGrad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node) { + const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (frameIdx >= frameSize) return; + int batchIdx = 0; + if (isBatch) { + batchIdx = blockIdx.y * blockDim.y + threadIdx.y; + if (batchIdx >= batchSize) return; + gateValue += batchIdx * 3 * frameSize; + gateGrad += batchIdx * 3 * frameSize; + outputGrad += batchIdx * frameSize; + } + + real rUpdateGateGrad; + real rFrameStateGrad; + real rPrevOutValue = 0; + real rPrevOutGrad = 0; + real rUpdateGateValue = gateValue[frameIdx + frameSize * 0]; + real rFrameStateValue = gateValue[frameIdx + frameSize * 2]; + real rOutGrad = outputGrad[frameIdx]; + + if (prevOutValue && prevOutGrad) { + if (isBatch) prevOutValue += batchIdx * frameSize; + rPrevOutValue = prevOutValue[frameIdx]; + + if (isBatch) prevOutGrad += batchIdx * frameSize; + rPrevOutGrad = prevOutGrad[frameIdx]; + } + + opStateGrad(rUpdateGateValue, + rUpdateGateGrad, + rFrameStateValue, + rFrameStateGrad, + rPrevOutValue, + rPrevOutGrad, + rOutGrad, + hppl::gpu::backward[active_node]); + + gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad; + gateGrad[frameIdx + frameSize * 2] = rFrameStateGrad; + if (prevOutGrad) { + prevOutGrad[frameIdx] = rPrevOutGrad; + } +} + +/* + * threads(framePerBlock, batchPerBlock) + * grid(frameBlocks, batchBlocks) + */ +template +__global__ void KeGruBackwardResetGrad(OpResetGrad opResetGrad, + real *gateValue, + real *gateGrad, + real *prevOutValue, + real *prevOutGrad, + real *resetOutputGrad, + int frameSize, + int batchSize, + hl_activation_mode_t active_gate) { + const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (frameIdx >= frameSize) return; + int batchIdx = 0; + if (isBatch) { + batchIdx = blockIdx.y * blockDim.y + threadIdx.y; + if (batchIdx >= batchSize) return; + gateValue += batchIdx * 3 * frameSize; + gateGrad += batchIdx * 3 * frameSize; + resetOutputGrad += batchIdx * frameSize; + } + + real rResetGateGrad; + real rPrevOutValue = 0; + real rPrevOutGrad = 0; + real rResetOutputGrad = 0; + real rUpdateGateValue = gateValue[frameIdx + frameSize * 0]; + real rUpdateGateGrad = gateGrad[frameIdx + frameSize * 0]; + real rResetGateValue = gateValue[frameIdx + frameSize * 1]; + + if (prevOutValue && prevOutGrad) { + if (isBatch) prevOutValue += batchIdx * frameSize; + if (isBatch) prevOutGrad += batchIdx * frameSize; + rPrevOutValue = prevOutValue[frameIdx]; + rPrevOutGrad = prevOutGrad[frameIdx]; + rResetOutputGrad = resetOutputGrad[frameIdx]; + } + + opResetGrad(rUpdateGateValue, + rUpdateGateGrad, + rResetGateValue, + rResetGateGrad, + rPrevOutValue, + rPrevOutGrad, + rResetOutputGrad, + hppl::gpu::backward[active_gate]); + + gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad; + gateGrad[frameIdx + frameSize * 1] = rResetGateGrad; + if (prevOutGrad) { + prevOutGrad[frameIdx] = rPrevOutGrad; + } +} + +template +void hl_gpu_gru_backward(OpStateGrad opStateGrad, + OpResetGrad opResetGrad, + hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) { + dim3 threads; + dim3 grid; + if (batchSize == 1) { + int framePerBlock = frameSize <= 1024 ? frameSize : 1024; + int frameBlocks = (frameSize + 1024 - 1) / 1024; + threads = dim3(framePerBlock, 1); + grid = dim3(frameBlocks, 1); + } else { + threads = dim3(32, 32); + grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + } + + if (batchSize == 1) { + KeGruBackwardStateGrad + <<>>(opStateGrad, + value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, + grad.outputGrad, frameSize, batchSize, active_node); + } else { + KeGruBackwardStateGrad + <<>>(opStateGrad, + value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, + grad.outputGrad, frameSize, batchSize, active_node); + } + + if (value.prevOutValue && grad.prevOutGrad) { + hl_matrix_mul(grad.gateGrad + 2*frameSize, HPPL_OP_N, + value.stateWeight, HPPL_OP_T, + grad.resetOutputGrad, + batchSize, frameSize, frameSize, + /*alpha = */ 1, /*beta = */ 0, + 3*frameSize, frameSize, frameSize); + if (grad.stateWeightGrad) { + hl_matrix_mul(value.resetOutputValue, HPPL_OP_T, + grad.gateGrad + 2*frameSize, HPPL_OP_N, + grad.stateWeightGrad, + frameSize, frameSize, batchSize, + /*alpha = */ 1, /*beta = */ 1, + frameSize, 3*frameSize, frameSize); + } + } + + if (batchSize == 1) { + KeGruBackwardResetGrad + <<>>(opResetGrad, + value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, + grad.resetOutputGrad, frameSize, batchSize, active_gate); + } else { + KeGruBackwardResetGrad + <<>>(opResetGrad, + value.gateValue, grad.gateGrad, value.prevOutValue, grad.prevOutGrad, + grad.resetOutputGrad, frameSize, batchSize, active_gate); + } + + if (grad.prevOutGrad && value.prevOutValue) { + hl_matrix_mul(grad.gateGrad, HPPL_OP_N, + value.gateWeight, HPPL_OP_T, + grad.prevOutGrad, + batchSize, frameSize, 2*frameSize, + /*alpha = */ 1, /*beta = */ 1, + 3*frameSize, 2*frameSize, frameSize); + if (grad.gateWeightGrad) { + hl_matrix_mul(value.prevOutValue, HPPL_OP_T, + grad.gateGrad, HPPL_OP_N, + grad.gateWeightGrad, + frameSize, 2*frameSize, batchSize, + /*alpha = */ 1, /*beta = */ 1, + frameSize, 3*frameSize, 2*frameSize); + } + } + + CHECK_SYNC("hl_gpu_gru_backward failed"); +} + +#else + +template +void hl_gpu_gru_forward(OpResetOutput opResetOutput, + OpFinalOutput opFinalOutput, + hl_gru_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) {} + +template +void hl_gpu_gru_backward(OpStateGrad opStateGrad, + OpResetGrad opResetGrad, + hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) {} + +#endif + +#endif /* HL_GPU_GRU_CUH_ */ diff --git a/paddle/legacy/cuda/include/hl_gpu_lstm.cuh b/paddle/legacy/cuda/include/hl_gpu_lstm.cuh new file mode 100644 index 0000000000000000000000000000000000000000..aae011b838c0eca1197f55d236d759eab8ea993c --- /dev/null +++ b/paddle/legacy/cuda/include/hl_gpu_lstm.cuh @@ -0,0 +1,300 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + + +#ifndef HL_GPU_LSTM_CUH_ +#define HL_GPU_LSTM_CUH_ + +#ifdef __NVCC__ + +#include "paddle/legacy/utils/Logging.h" +#include "hl_device_functions.cuh" + +/* + * threads(framePerBlock, batchPerBlock) + * grid(frameBlocks, batchBlocks) + */ +template +__global__ void KeLstmForward(Op op, + hl_lstm_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (frameIdx >= frameSize) return; + + int batchIdx = 0; + if (isBatch) { + batchIdx = blockIdx.y * blockDim.y + threadIdx.y; + if (batchIdx >= batchSize) return; + value.gateValue += batchIdx * frameSize * 4; + value.outputValue += batchIdx * frameSize; + value.stateValue += batchIdx * frameSize; + value.stateActiveValue += batchIdx * frameSize; + } + + real rState; + real rPrevState = 0; + real rStateAtv; + real rOut; + real rValueIn; + real rValueIg; + real rValueFg; + real rValueOg; + real rCheckI = value.checkIg[frameIdx]; + real rCheckF = value.checkFg[frameIdx]; + real rCheckO = value.checkOg[frameIdx]; + + rValueIn = value.gateValue[frameIdx]; + rValueIg = value.gateValue[frameIdx + frameSize]; + rValueFg = value.gateValue[frameIdx + frameSize * 2]; + rValueOg = value.gateValue[frameIdx + frameSize * 3]; + + if (value.prevStateValue) { + if (isBatch) value.prevStateValue += batchIdx * frameSize; + rPrevState = value.prevStateValue[frameIdx]; + } + + op(rValueIn, + rValueIg, + rValueFg, + rValueOg, + rPrevState, + rState, + rStateAtv, + rOut, + rCheckI, + rCheckF, + rCheckO, + hppl::gpu::forward[active_node], + hppl::gpu::forward[active_gate], + hppl::gpu::forward[active_state]); + + value.gateValue[frameIdx] = rValueIn; + value.gateValue[frameIdx + frameSize] = rValueIg; + value.gateValue[frameIdx + frameSize * 2] = rValueFg; + value.gateValue[frameIdx + frameSize * 3] = rValueOg; + + value.stateValue[frameIdx] = rState; + value.stateActiveValue[frameIdx] = rStateAtv; + value.outputValue[frameIdx] = rOut; +} + +/* + * threads(framePerBlock, batchPerBlock) + * grid(frameBlocks, batchBlocks) + */ +template +__global__ void KeLstmBackward(Op op, + hl_lstm_value value, + hl_lstm_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (frameIdx >= frameSize) return; + + int batchIdx = 0; + if (isBatch) { + batchIdx = blockIdx.y * blockDim.y + threadIdx.y; + if (batchIdx >= batchSize) return; + value.gateValue += batchIdx * frameSize * 4; + value.stateValue += batchIdx * frameSize; + value.stateActiveValue += batchIdx * frameSize; + grad.gateGrad += batchIdx * frameSize * 4; + grad.stateGrad += batchIdx * frameSize; + grad.outputGrad += batchIdx * frameSize; + } + + real rValueIn; + real rValueIg; + real rValueFg; + real rValueOg; + real rGradIn; + real rGradIg; + real rGradFg; + real rGradOg; + real rPrevState = 0; + real rPrevStateGrad; + real rState; + real rStateGrad; + real rStateAtv; + real rOutputGrad; + real rCheckI = value.checkIg[frameIdx]; + real rCheckF = value.checkFg[frameIdx]; + real rCheckO = value.checkOg[frameIdx]; + real rCheckIGrad; + real rCheckFGrad; + real rCheckOGrad; + + rValueIn = value.gateValue[frameIdx]; + rValueIg = value.gateValue[frameIdx + frameSize]; + rValueFg = value.gateValue[frameIdx + frameSize * 2]; + rValueOg = value.gateValue[frameIdx + frameSize * 3]; + rState = value.stateValue[frameIdx]; + rStateAtv = value.stateActiveValue[frameIdx]; + rOutputGrad = grad.outputGrad[frameIdx]; + rStateGrad = grad.stateGrad[frameIdx]; + + if (value.prevStateValue) { + if (isBatch) value.prevStateValue += batchIdx * frameSize; + rPrevState = value.prevStateValue[frameIdx]; + } + + op(rValueIn, + rValueIg, + rValueFg, + rValueOg, + rGradIn, + rGradIg, + rGradFg, + rGradOg, + rPrevState, + rPrevStateGrad, + rState, + rStateGrad, + rStateAtv, + rOutputGrad, + rCheckI, + rCheckF, + rCheckO, + rCheckIGrad, + rCheckFGrad, + rCheckOGrad, + hppl::gpu::backward[active_node], + hppl::gpu::backward[active_gate], + hppl::gpu::backward[active_state]); + + grad.gateGrad[frameIdx] = rGradIn; + grad.gateGrad[frameIdx + frameSize ] = rGradIg; + grad.gateGrad[frameIdx + frameSize * 2] = rGradFg; + grad.gateGrad[frameIdx + frameSize * 3] = rGradOg; + grad.stateGrad[frameIdx] = rStateGrad; + if (grad.prevStateGrad) { + if (isBatch) grad.prevStateGrad += batchIdx * frameSize; + grad.prevStateGrad[frameIdx] = rPrevStateGrad; + } + + if (isBatch) { + if (value.prevStateValue) { + if (grad.checkIgGrad) paddle::paddleAtomicAdd(grad.checkIgGrad+frameIdx, rCheckIGrad); + if (grad.checkFgGrad) paddle::paddleAtomicAdd(grad.checkFgGrad+frameIdx, rCheckFGrad); + } + if (grad.checkOgGrad) paddle::paddleAtomicAdd(grad.checkOgGrad+frameIdx, rCheckOGrad); + } else { + if (value.prevStateValue) { + if (grad.checkIgGrad) grad.checkIgGrad[frameIdx] += rCheckIGrad; + if (grad.checkFgGrad) grad.checkFgGrad[frameIdx] += rCheckFGrad; + } + if (grad.checkOgGrad) grad.checkOgGrad[frameIdx] += rCheckOGrad; + } +} + +template +void hl_gpu_lstm_forward(Op op, + hl_lstm_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + dim3 threads; + dim3 grid; + if (batchSize == 1) { + int framePerBlock = frameSize <= 1024 ? frameSize : 1024; + int frameBlocks = (frameSize + 1024 - 1) / 1024; + threads = dim3(framePerBlock, 1); + grid = dim3(frameBlocks, 1); + } else { + /* framePerBlock = 32 batchPerBlock = 32 */ + threads = dim3(32, 32); + grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + } + + if (batchSize == 1) { + KeLstmForward + <<>>(op, value, + frameSize, batchSize, active_node, active_gate, active_state); + } else { + KeLstmForward + <<>>(op, value, + frameSize, batchSize, active_node, active_gate, active_state); + } + + CHECK_SYNC("hl_gpu_lstm_forward failed"); +} + +template +void hl_gpu_lstm_backward(Op op, + hl_lstm_value value, + hl_lstm_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + dim3 threads; + dim3 grid; + if (batchSize == 1) { + int framePerBlock = frameSize <= 1024 ? frameSize : 1024; + int frameBlocks = (frameSize + 1024 - 1) / 1024; + threads = dim3(framePerBlock, 1); + grid = dim3(frameBlocks, 1); + } else { + /* framePerBlock = 32 batchPerBlock = 32 */ + threads = dim3(32, 32); + grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + } + + if (batchSize == 1) { + KeLstmBackward + <<>>(op, value, grad, + frameSize, batchSize, active_node, active_gate, active_state); + } else { + KeLstmBackward + <<>>(op, value, grad, + frameSize, batchSize, active_node, active_gate, active_state); + } + + CHECK_SYNC("hl_gpu_lstm_backward failed"); +} + +#else + +template +void hl_gpu_lstm_forward(Op op, + hl_lstm_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) {} + +template +void hl_gpu_lstm_backward(Op op, + hl_lstm_value value, + hl_lstm_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) {} + +#endif + +#endif /* HL_GPU_LSTM_CUH_ */ diff --git a/paddle/legacy/cuda/include/hl_gpu_matrix_kernel.cuh b/paddle/legacy/cuda/include/hl_gpu_matrix_kernel.cuh new file mode 100644 index 0000000000000000000000000000000000000000..6177d23657fba5b2800041a3dd7b5f76bf35aa1a --- /dev/null +++ b/paddle/legacy/cuda/include/hl_gpu_matrix_kernel.cuh @@ -0,0 +1,629 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + + + +#ifndef HL_GPU_MATRIX_KERNEL_CUH_ +#define HL_GPU_MATRIX_KERNEL_CUH_ + +#include +#include "paddle/legacy/utils/Logging.h" +#include "hl_base.h" + +#ifdef __NVCC__ +/* gpu apply interface */ + +template +__global__ void KeEltWiseUnaryOp(T* A_d, const int border, Op op) { + const int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < border) { + op.gpuOperator(A_d[idx]); + } +} + +template +__global__ void KeEltWiseUnaryOp(T* A_d, + int dimM, + int dimN, + int lda, + Op op) { + const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; + const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; + for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { + for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { + op.gpuOperator(A_d[i * lda + j]); + } + } +} + +template +__global__ void KeEltWiseBinaryOp(T* A_d, T *B_d, const int border, Op op) { + const int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < border) { + op.gpuOperator(A_d[idx], B_d[idx]); + } +} + +template +__global__ void KeEltWiseBinaryOp(T *A_d, + T *B_d, + int dimM, + int dimN, + int lda, + int ldb, + Op op) { + const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; + const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; + for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { + for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { + if (BAsRowVector == 0 && BAsColVector == 0) { + op.gpuOperator(A_d[i * lda + j], B_d[i * ldb + j]); + } else if (BAsRowVector == 1 && BAsColVector == 0) { + op.gpuOperator(A_d[i * lda + j], B_d[j]); + } else if (BAsRowVector == 0 && BAsColVector == 1) { + op.gpuOperator(A_d[i * lda + j], B_d[i * ldb]); + } else { + op.gpuOperator(A_d[i * lda + j], B_d[0]); + } + } + } +} + +template +__global__ void KeEltWiseTernaryOp(T* A_d, + T *B_d, + T *C_d, + const int border, + Op op) { + const int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < border) { + op.gpuOperator(A_d[idx], B_d[idx], C_d[idx]); + } +} + +template +__global__ void KeEltWiseTernaryOp(T* A_d, + T* B_d, + T* C_d, + int dimM, + int dimN, + int lda, + int ldb, + int ldc, + Op op) { + const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; + const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; + for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { + for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { + if (CAsRowVector == 0 && CAsColVector == 0) { + op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[i*ldc + j]); + } else if (CAsRowVector == 1 && CAsColVector == 0) { + op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[j]); + } else if (CAsRowVector == 0 && CAsColVector == 1) { + op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[i*ldc]); + } else { + op.gpuOperator(A_d[i*lda + j], B_d[i*ldb + j], C_d[0]); + } + } + } +} + +template +__global__ void KeEltWiseQuaternaryOp(T* A_d, + T* B_d, + T* C_d, + T* D_d, + const int border, + Op op) { + const int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < border) { + op.gpuOperator(A_d[idx], B_d[idx], C_d[idx], D_d[idx]); + } +} + +template +__global__ void KeEltWiseQuaternaryOp(T* A_d, + T* B_d, + T* C_d, + T* D_d, + int dimM, + int dimN, + int lda, + int ldb, + int ldc, + int ldd, + Op op) { + const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; + const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; + for (int i = rowIdx; i < dimM; i += gridDim.y * blockDim.y) { + for (int j = colIdx; j < dimN; j += gridDim.x * blockDim.x) { + op.gpuOperator(A_d[i*lda + j], + B_d[i*ldb + j], C_d[i*ldc + j], D_d[i*ldd + j]); + } + } +} + +/** + * @brief gpu element wise unary operator. + */ +template +void hl_gpu_apply_unary_op(Op op, T* A_d, int dimM, int dimN, int lda) { + CHECK_NOTNULL(A_d); + + if (dimM == 1 || dimN == lda) { + int size = dimM * dimN; + int blockSize = size <= 1024 ? size : 1024; + int gridSize = (size + 1024 - 1) / 1024; + KeEltWiseUnaryOp<<>> + (A_d, size, op); + } else { + int blockSizeY = std::min(32, dimM); + int blockSizeX = (32 / blockSizeY) * 32; + int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); + int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); + dim3 threads(blockSizeX, blockSizeY); + dim3 grid(gridSizeX, gridSizeY); + KeEltWiseUnaryOp<<>> + (A_d, dimM, dimN, lda, op); + } + + CHECK_SYNC("hl_gpu_apply_unary_op failed"); +} + +/** + * @brief gpu element wise binary operator. + */ +template +void hl_gpu_apply_binary_op(Op op, + T* A_d, + T* B_d, + int dimM, + int dimN, + int lda, + int ldb) { + CHECK_NOTNULL(A_d); + + if ((BAsRowVector == 0 && BAsColVector == 0) && + ((dimM == 1) || (dimN == lda && dimN == ldb))) { + int size = dimM * dimN; + int blockSize = size <= 1024 ? size : 1024; + int gridSize = (size + 1024 - 1) / 1024; + KeEltWiseBinaryOp<<>> + (A_d, B_d, size, op); + } else { + int blockSizeY = std::min(32, dimM); + int blockSizeX = (32 / blockSizeY) * 32; + int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); + int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); + dim3 threads(blockSizeX, blockSizeY); + dim3 grid(gridSizeX, gridSizeY); + KeEltWiseBinaryOp + <<>> + (A_d, B_d, dimM, dimN, lda, ldb, op); + } + + CHECK_SYNC("hl_gpu_apply_binary_op failed"); +} + +/** + * @brief gpu element wise ternary operator. + */ +template +void hl_gpu_apply_ternary_op(Op op, + T* A_d, + T* B_d, + T* C_d, + int dimM, + int dimN, + int lda, + int ldb, + int ldc) { + CHECK_NOTNULL(A_d); + + if ((CAsRowVector == 0 && CAsColVector == 0) && + ((dimM == 1) || (dimN == lda && dimN == ldb && dimN == ldc))) { + int size = dimM * dimN; + int blockSize = size <= 1024 ? size : 1024; + int gridSize = (size + 1024 - 1) / 1024; + KeEltWiseTernaryOp<<>> + (A_d, B_d, C_d, size, op); + } else { + int blockSizeY = std::min(32, dimM); + int blockSizeX = (32 / blockSizeY) * 32; + int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); + int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); + dim3 threads(blockSizeX, blockSizeY); + dim3 grid(gridSizeX, gridSizeY); + KeEltWiseTernaryOp + <<>> + (A_d, B_d, C_d, dimM, dimN, lda, ldb, ldc, op); + } + + CHECK_SYNC("hl_gpu_apply_ternary_op failed"); +} + + +/** + * @brief gpu element wise quaternary operator. + */ +template +void hl_gpu_apply_quaternary_op(Op op, + T* A_d, + T* B_d, + T* C_d, + T* D_d, + int dimM, + int dimN, + int lda, + int ldb, + int ldc, + int ldd) { + CHECK_NOTNULL(A_d); + + if ((dimM == 1) || + (dimN == lda && dimN == ldb && dimN == ldc && dimN == ldd)) { + int size = dimM * dimN; + int blockSize = size <= 1024 ? size : 1024; + int gridSize = (size + 1024 - 1) / 1024; + KeEltWiseQuaternaryOp<<>> + (A_d, B_d, C_d, D_d, size, op); + } else { + int blockSizeY = std::min(32, dimM); + int blockSizeX = (32 / blockSizeY) * 32; + int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); + int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); + dim3 threads(blockSizeX, blockSizeY); + dim3 grid(gridSizeX, gridSizeY); + KeEltWiseQuaternaryOp<<>> + (A_d, B_d, C_d, D_d, dimM, dimN, lda, ldb, ldc, ldd, op); + } + + CHECK_SYNC("hl_gpu_apply_quaternary_op failed"); +} + +#else + +template +void hl_gpu_apply_unary_op(Op op, T* A_d, int dimM, int dimN, int lda) {} + +template +void hl_gpu_apply_binary_op(Op op, + T* A_d, + T* B_d, + int dimM, + int dimN, + int lda, + int ldb) {} + +template +void hl_gpu_apply_ternary_op(Op op, + T* A_d, + T* B_d, + T* C_d, + int dimM, + int dimN, + int lda, + int ldb, + int ldc) {} + +template +void hl_gpu_apply_quaternary_op(Op op, + T* A_d, + T* B_d, + T* C_d, + T* D_d, + int dimM, + int dimN, + int lda, + int ldb, + int ldc, + int ldd) {} +#endif + +#ifdef __NVCC__ +/** + * @brief matrix row operator. + */ + +template +__device__ __inline__ real sumRow(Agg agg, Op op, + int idx, int blockSize, + int dimN, real *A) { + real tmp = agg.init(); + int cnt = (dimN + blockSize -1) / blockSize; + for (int i = 0; i < cnt && idx < dimN; i++) { + tmp = agg(tmp, op(A[idx])); + idx += blockSize; + } + return tmp; +} + +template +__device__ __inline__ real sumRow(Agg agg, Op op, + int idx, int blockSize, + int dimN, real *A, real *B) { + real tmp = agg.init(); + int cnt = (dimN + blockSize -1) / blockSize; + for (int i = 0; i < cnt && idx < dimN; i++) { + tmp = agg(tmp, op(A[idx], B[idx])); + idx += blockSize; + } + return tmp; +} + +template +__device__ __inline__ void aggRow(Agg agg, real *row, int size, int tid) { + for (int stride = size/2; stride > 0; stride = stride/2) { + if (tid < stride) { + row[tid] = agg(row[tid], row[tid + stride]); + } + __syncthreads(); + } +} + +template +__global__ void KeMatrixRowOp(Agg agg, Op op, Saver sv, + int dimN, + real *dst, int ld, + real *A, int lda) { + __shared__ real row_s[blockSize]; + int rowId = blockIdx.x + blockIdx.y*gridDim.x; + int tid = threadIdx.x; + + A += rowId*lda; + row_s[tid] = sumRow(agg, op, tid, blockSize, dimN, A); + __syncthreads(); + + aggRow(agg, row_s, blockSize, tid); + __syncthreads(); + + if (tid == 0) { + dst[rowId*ld] = sv(dst[rowId*ld], row_s[0]); + } +} + +template +__global__ void KeMatrixRowOp(Agg agg, Op op, Saver sv, + int dimN, + real *dst, int ld, + real *A, int lda, + real *B, int ldb) { + __shared__ real row_s[blockSize]; + int rowId = blockIdx.x + blockIdx.y*gridDim.x; + int tid = threadIdx.x; + + A += rowId*lda; + B += rowId*ldb; + row_s[tid] = sumRow(agg, op, tid, blockSize, dimN, A, B); + __syncthreads(); + + aggRow(agg, row_s, blockSize, tid); + __syncthreads(); + + if (tid == 0) { + dst[rowId*ld] = sv(dst[rowId*ld], row_s[0]); + } +} + +/** + * @brief matrix column operator. + */ +template +__device__ __inline__ real sumCol(Agg agg, Op op, + int index, int stride, + int dimM, real *A, int lda) { + real tmp = agg.init(); + for (; index < dimM;) { + tmp = agg(tmp, op(A[index*lda])); + index += stride; + } + return tmp; +} + +template +__device__ __inline__ real sumCol(Agg agg, Op op, + int index, int stride, int dimM, + real *A, int lda, real *B, int ldb) { + real tmp = agg.init(); + for (; index < dimM;) { + tmp = agg(tmp, op(A[index*lda], B[index*ldb])); + index += stride; + } + return tmp; +} + +template +__global__ void KeMatrixColumnOp(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, + real *A, int lda) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (rowIdx < dimN) { + A += rowIdx; + real tmp = sumCol(agg, op, 0, 1, dimM, A, lda); + dst[rowIdx] = sv(dst[rowIdx], tmp); + } +} + +template +__global__ void KeMatrixColumnOp_S(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, + real *A, int lda) { + __shared__ real col_s[blockDimX*blockDimY]; + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + + if (rowIdx < dimN) { + A += rowIdx; + real tmp = sumCol(agg, op, threadIdx.y, blockDimY, dimM, A, lda); + col_s[threadIdx.x + threadIdx.y*blockDimX] = tmp; + } + __syncthreads(); + + if (rowIdx < dimN) { + if (threadIdx.y ==0) { + real tmp = agg.init(); + for (int i=0; i < blockDimY; i++) { + tmp = agg(tmp, col_s[threadIdx.x + i*blockDimX]); + } + dst[rowIdx] = sv(dst[rowIdx], tmp); + } + } +} + +template +__global__ void KeMatrixColumnOp(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, + real *A, int lda, + real *B, int ldb) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + if (rowIdx < dimN) { + A += rowIdx; + B += rowIdx; + real tmp = sumCol(agg, op, 0, 1, dimM, A, lda, B, ldb); + dst[rowIdx] = sv(dst[rowIdx], tmp); + } +} + +template +__global__ void KeMatrixColumnOp_S(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, + real *A, int lda, + real *B, int ldb) { + __shared__ real col_s[blockDimX*blockDimY]; + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + + if (rowIdx < dimN) { + A += rowIdx; + B += rowIdx; + real tmp = sumCol(agg, op, + threadIdx.y, blockDimY, dimM, A, lda, B, ldb); + col_s[threadIdx.x + threadIdx.y*blockDimX] = tmp; + } + __syncthreads(); + + if (rowIdx < dimN) { + if (threadIdx.y ==0) { + real tmp = agg.init(); + for (int i=0; i < blockDimY; i++) { + tmp = agg(tmp, col_s[threadIdx.x + i*blockDimX]); + } + dst[rowIdx] = sv(dst[rowIdx], tmp); + } + } +} + +#endif + +template +void hl_gpu_matrix_row_op(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, int ld, + real *A, int lda) { +#ifdef __NVCC__ + CHECK_NOTNULL(dst); + CHECK_NOTNULL(A); + + int blocksX = dimM; + int blocksY = 1; + dim3 threads(128, 1); + dim3 grid(blocksX, blocksY); + KeMatrixRowOp<<< grid, threads, 0, STREAM_DEFAULT >>> + (agg, op, sv, dimN, dst, ld, A, lda); + + CHECK_SYNC("hl_matrix_row_op failed"); +#endif +} + +template +void hl_gpu_matrix_row_op(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, int ld, + real *A, int lda, + real *B, int ldb) { +#ifdef __NVCC__ + CHECK_NOTNULL(dst); + CHECK_NOTNULL(A); + + int blocksX = dimM; + int blocksY = 1; + dim3 threads(128, 1); + dim3 grid(blocksX, blocksY); + KeMatrixRowOp<<< grid, threads, 0, STREAM_DEFAULT >>> + (agg, op, sv, dimN, dst, ld, A, lda, B, ldb); + + CHECK_SYNC("hl_matrix_row_op failed"); +#endif +} + +template +void hl_gpu_matrix_column_op(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, + real *A, int lda) { +#ifdef __NVCC__ + if (dimN >= 8192) { + int blocksX = (dimN + 128 -1) / 128; + int blocksY = 1; + dim3 threads(128, 1); + dim3 grid(blocksX, blocksY); + KeMatrixColumnOp + <<< grid, threads, 0, STREAM_DEFAULT >>> + (agg, op, sv, dimM, dimN, dst, A, lda); + } else { + int blocksX = (dimN + 32 -1) / 32; + int blocksY = 1; + dim3 threads(32, 32); + dim3 grid(blocksX, blocksY); + KeMatrixColumnOp_S + <<< grid, threads, 0, STREAM_DEFAULT>>> + (agg, op, sv, dimM, dimN, dst, A, lda); + } + + CHECK_SYNC("hl_matrix_column_op failed"); +#endif +} + +template +void hl_gpu_matrix_column_op(Agg agg, Op op, Saver sv, + int dimM, int dimN, + real *dst, + real *A, int lda, + real *B, int ldb) { +#ifdef __NVCC__ + if (dimN >= 8192) { + int blocksX = (dimN + 128 -1) / 128; + int blocksY = 1; + dim3 threads(128, 1); + dim3 grid(blocksX, blocksY); + KeMatrixColumnOp + <<< grid, threads, 0, STREAM_DEFAULT >>> + (agg, op, sv, dimM, dimN, dst, A, lda, B, ldb); + } else { + int blocksX = (dimN + 32 -1) / 32; + int blocksY = 1; + dim3 threads(32, 32); + dim3 grid(blocksX, blocksY); + KeMatrixColumnOp_S + <<< grid, threads, 0, STREAM_DEFAULT>>> + (agg, op, sv, dimM, dimN, dst, A, lda, B, ldb); + } + + CHECK_SYNC("hl_matrix_column_op failed"); +#endif +} + +#endif /* HL_GPU_MATRIX_KERNEL_CUH_ */ diff --git a/paddle/cuda/include/hl_gru_ops.cuh b/paddle/legacy/cuda/include/hl_gru_ops.cuh similarity index 100% rename from paddle/cuda/include/hl_gru_ops.cuh rename to paddle/legacy/cuda/include/hl_gru_ops.cuh diff --git a/paddle/cuda/include/hl_lstm.h b/paddle/legacy/cuda/include/hl_lstm.h similarity index 100% rename from paddle/cuda/include/hl_lstm.h rename to paddle/legacy/cuda/include/hl_lstm.h diff --git a/paddle/cuda/include/hl_lstm_ops.cuh b/paddle/legacy/cuda/include/hl_lstm_ops.cuh similarity index 100% rename from paddle/cuda/include/hl_lstm_ops.cuh rename to paddle/legacy/cuda/include/hl_lstm_ops.cuh diff --git a/paddle/cuda/include/hl_matrix.h b/paddle/legacy/cuda/include/hl_matrix.h similarity index 100% rename from paddle/cuda/include/hl_matrix.h rename to paddle/legacy/cuda/include/hl_matrix.h diff --git a/paddle/cuda/include/hl_matrix_apply.cuh b/paddle/legacy/cuda/include/hl_matrix_apply.cuh similarity index 100% rename from paddle/cuda/include/hl_matrix_apply.cuh rename to paddle/legacy/cuda/include/hl_matrix_apply.cuh diff --git a/paddle/cuda/include/hl_matrix_base.cuh b/paddle/legacy/cuda/include/hl_matrix_base.cuh similarity index 100% rename from paddle/cuda/include/hl_matrix_base.cuh rename to paddle/legacy/cuda/include/hl_matrix_base.cuh diff --git a/paddle/cuda/include/hl_matrix_base_detail.cuh b/paddle/legacy/cuda/include/hl_matrix_base_detail.cuh similarity index 100% rename from paddle/cuda/include/hl_matrix_base_detail.cuh rename to paddle/legacy/cuda/include/hl_matrix_base_detail.cuh diff --git a/paddle/cuda/include/hl_matrix_ops.cuh b/paddle/legacy/cuda/include/hl_matrix_ops.cuh similarity index 100% rename from paddle/cuda/include/hl_matrix_ops.cuh rename to paddle/legacy/cuda/include/hl_matrix_ops.cuh diff --git a/paddle/cuda/include/hl_matrix_type.cuh b/paddle/legacy/cuda/include/hl_matrix_type.cuh similarity index 100% rename from paddle/cuda/include/hl_matrix_type.cuh rename to paddle/legacy/cuda/include/hl_matrix_type.cuh diff --git a/paddle/cuda/include/hl_perturbation_util.cuh b/paddle/legacy/cuda/include/hl_perturbation_util.cuh similarity index 100% rename from paddle/cuda/include/hl_perturbation_util.cuh rename to paddle/legacy/cuda/include/hl_perturbation_util.cuh diff --git a/paddle/cuda/include/hl_recurrent_apply.cuh b/paddle/legacy/cuda/include/hl_recurrent_apply.cuh similarity index 100% rename from paddle/cuda/include/hl_recurrent_apply.cuh rename to paddle/legacy/cuda/include/hl_recurrent_apply.cuh diff --git a/paddle/cuda/include/hl_sequence.h b/paddle/legacy/cuda/include/hl_sequence.h similarity index 100% rename from paddle/cuda/include/hl_sequence.h rename to paddle/legacy/cuda/include/hl_sequence.h diff --git a/paddle/cuda/include/hl_sparse.h b/paddle/legacy/cuda/include/hl_sparse.h similarity index 100% rename from paddle/cuda/include/hl_sparse.h rename to paddle/legacy/cuda/include/hl_sparse.h diff --git a/paddle/cuda/include/hl_sparse.ph b/paddle/legacy/cuda/include/hl_sparse.ph similarity index 100% rename from paddle/cuda/include/hl_sparse.ph rename to paddle/legacy/cuda/include/hl_sparse.ph diff --git a/paddle/cuda/include/hl_table_apply.h b/paddle/legacy/cuda/include/hl_table_apply.h similarity index 100% rename from paddle/cuda/include/hl_table_apply.h rename to paddle/legacy/cuda/include/hl_table_apply.h diff --git a/paddle/cuda/include/hl_tensor_ops.h b/paddle/legacy/cuda/include/hl_tensor_ops.h similarity index 100% rename from paddle/cuda/include/hl_tensor_ops.h rename to paddle/legacy/cuda/include/hl_tensor_ops.h diff --git a/paddle/cuda/include/hl_thread.ph b/paddle/legacy/cuda/include/hl_thread.ph similarity index 100% rename from paddle/cuda/include/hl_thread.ph rename to paddle/legacy/cuda/include/hl_thread.ph diff --git a/paddle/cuda/include/hl_time.h b/paddle/legacy/cuda/include/hl_time.h similarity index 100% rename from paddle/cuda/include/hl_time.h rename to paddle/legacy/cuda/include/hl_time.h diff --git a/paddle/cuda/include/hl_top_k.h b/paddle/legacy/cuda/include/hl_top_k.h similarity index 100% rename from paddle/cuda/include/hl_top_k.h rename to paddle/legacy/cuda/include/hl_top_k.h diff --git a/paddle/cuda/include/hl_warpctc_wrap.h b/paddle/legacy/cuda/include/hl_warpctc_wrap.h similarity index 100% rename from paddle/cuda/include/hl_warpctc_wrap.h rename to paddle/legacy/cuda/include/hl_warpctc_wrap.h diff --git a/paddle/cuda/include/stub/hl_aggregate_stub.h b/paddle/legacy/cuda/include/stub/hl_aggregate_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_aggregate_stub.h rename to paddle/legacy/cuda/include/stub/hl_aggregate_stub.h diff --git a/paddle/cuda/include/stub/hl_cnn_stub.h b/paddle/legacy/cuda/include/stub/hl_cnn_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_cnn_stub.h rename to paddle/legacy/cuda/include/stub/hl_cnn_stub.h diff --git a/paddle/cuda/include/stub/hl_cuda_cublas_stub.h b/paddle/legacy/cuda/include/stub/hl_cuda_cublas_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_cuda_cublas_stub.h rename to paddle/legacy/cuda/include/stub/hl_cuda_cublas_stub.h diff --git a/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h b/paddle/legacy/cuda/include/stub/hl_cuda_cudnn_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_cuda_cudnn_stub.h rename to paddle/legacy/cuda/include/stub/hl_cuda_cudnn_stub.h diff --git a/paddle/cuda/include/stub/hl_cuda_stub.h b/paddle/legacy/cuda/include/stub/hl_cuda_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_cuda_stub.h rename to paddle/legacy/cuda/include/stub/hl_cuda_stub.h diff --git a/paddle/cuda/include/stub/hl_lstm_stub.h b/paddle/legacy/cuda/include/stub/hl_lstm_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_lstm_stub.h rename to paddle/legacy/cuda/include/stub/hl_lstm_stub.h diff --git a/paddle/cuda/include/stub/hl_matrix_stub.h b/paddle/legacy/cuda/include/stub/hl_matrix_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_matrix_stub.h rename to paddle/legacy/cuda/include/stub/hl_matrix_stub.h diff --git a/paddle/cuda/include/stub/hl_sequence_stub.h b/paddle/legacy/cuda/include/stub/hl_sequence_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_sequence_stub.h rename to paddle/legacy/cuda/include/stub/hl_sequence_stub.h diff --git a/paddle/cuda/include/stub/hl_sparse_stub.h b/paddle/legacy/cuda/include/stub/hl_sparse_stub.h similarity index 100% rename from paddle/cuda/include/stub/hl_sparse_stub.h rename to paddle/legacy/cuda/include/stub/hl_sparse_stub.h diff --git a/paddle/cuda/src/avx_mathfun.h b/paddle/legacy/cuda/src/avx_mathfun.h similarity index 100% rename from paddle/cuda/src/avx_mathfun.h rename to paddle/legacy/cuda/src/avx_mathfun.h diff --git a/paddle/cuda/src/hl_avx_functions.cc b/paddle/legacy/cuda/src/hl_avx_functions.cc similarity index 100% rename from paddle/cuda/src/hl_avx_functions.cc rename to paddle/legacy/cuda/src/hl_avx_functions.cc diff --git a/paddle/cuda/src/hl_batch_norm.cu b/paddle/legacy/cuda/src/hl_batch_norm.cu similarity index 100% rename from paddle/cuda/src/hl_batch_norm.cu rename to paddle/legacy/cuda/src/hl_batch_norm.cu diff --git a/paddle/cuda/src/hl_batch_transpose.cu b/paddle/legacy/cuda/src/hl_batch_transpose.cu similarity index 100% rename from paddle/cuda/src/hl_batch_transpose.cu rename to paddle/legacy/cuda/src/hl_batch_transpose.cu diff --git a/paddle/cuda/src/hl_cpu_functions.cc b/paddle/legacy/cuda/src/hl_cpu_functions.cc similarity index 100% rename from paddle/cuda/src/hl_cpu_functions.cc rename to paddle/legacy/cuda/src/hl_cpu_functions.cc diff --git a/paddle/legacy/cuda/src/hl_cuda_aggregate.cu b/paddle/legacy/cuda/src/hl_cuda_aggregate.cu new file mode 100644 index 0000000000000000000000000000000000000000..9831c5ecc340135c27b49d24715c63f8a8dfa8e9 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_aggregate.cu @@ -0,0 +1,293 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_aggregate.h" +#include "hl_base.h" +#include "hl_cuda.h" +#include "hl_cuda.ph" +#include "hl_matrix_base.cuh" +#include "hl_thread.ph" +#include "paddle/legacy/utils/Logging.h" + +/** + * @brief matrix row operator. + */ +template +__global__ void KeMatrixRowOp(Agg agg, real *E, real *Sum, int dimN) { + __shared__ real sum_s[blockSize]; + int cnt = (dimN + blockSize - 1) / blockSize; + int rowId = blockIdx.x + blockIdx.y * gridDim.x; + int index = rowId * dimN; + int tid = threadIdx.x; + int lmt = tid; + + real tmp = agg.init(); + for (int ii = 0; ii < cnt && lmt < dimN; ii++) { + tmp = agg(tmp, E[index + lmt]); + lmt += blockSize; + } + sum_s[tid] = tmp; + __syncthreads(); + + for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { + if (tid < stride) { + sum_s[tid] = agg(sum_s[tid], sum_s[tid + stride]); + } + __syncthreads(); + } + __syncthreads(); + + if (tid == 0) { + Sum[rowId] = sum_s[0]; + } +} + +template +void hl_matrix_row_op(Agg agg, real *A_d, real *C_d, int dimM, int dimN) { + int blocksX = dimM; + int blocksY = 1; + dim3 threads(128, 1); + dim3 grid(blocksX, blocksY); + + KeMatrixRowOp<<>>( + agg, A_d, C_d, dimN); +} + +void hl_matrix_row_sum(real *A_d, real *C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + hl_matrix_row_op(aggregate::sum(), A_d, C_d, dimM, dimN); + CHECK_SYNC("hl_matrix_row_sum failed"); +} + +void hl_matrix_row_max(real *A_d, real *C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + hl_matrix_row_op(aggregate::max(), A_d, C_d, dimM, dimN); + CHECK_SYNC("hl_matrix_row_max failed"); +} + +void hl_matrix_row_min(real *A_d, real *C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + hl_matrix_row_op(aggregate::min(), A_d, C_d, dimM, dimN); + CHECK_SYNC("hl_matrix_row_min failed"); +} + +/** + * @brief matrix column operator. + */ +template +__global__ void KeMatrixColumnOp( + Agg agg, real *E, real *Sum, int dimM, int dimN) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + real tmp = agg.init(); + if (rowIdx < dimN) { + for (int index = 0; index < dimM; index++) { + tmp = agg(tmp, E[dimN * index + rowIdx]); + } + Sum[rowIdx] = tmp; + } +} + +template +__global__ void KeMatrixColumnOp_S( + Agg agg, real *E, real *Sum, int dimM, int dimN) { + __shared__ real _sum[blockDimX * blockDimY]; + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + int index = threadIdx.y; + + real tmp = agg.init(); + if (rowIdx < dimN) { + for (; index < dimM;) { + tmp = agg(tmp, E[dimN * index + rowIdx]); + index += blockDimY; + } + } + _sum[threadIdx.x + threadIdx.y * blockDimX] = tmp; + __syncthreads(); + + if (rowIdx < dimN) { + if (threadIdx.y == 0) { + real tmp = agg.init(); + for (int i = 0; i < blockDimY; i++) { + tmp = agg(tmp, _sum[threadIdx.x + i * blockDimX]); + } + Sum[rowIdx] = tmp; + } + } +} + +template +void hl_matrix_column_op(Agg agg, real *A_d, real *C_d, int dimM, int dimN) { + if (dimN >= 8192) { + int blocksX = (dimN + 128 - 1) / 128; + int blocksY = 1; + dim3 threads(128, 1); + dim3 grid(blocksX, blocksY); + KeMatrixColumnOp<<>>( + agg, A_d, C_d, dimM, dimN); + } else { + int blocksX = (dimN + 32 - 1) / 32; + int blocksY = 1; + dim3 threads(32, 32); + dim3 grid(blocksX, blocksY); + KeMatrixColumnOp_S<<>>( + agg, A_d, C_d, dimM, dimN); + } + + return; +} + +void hl_matrix_column_sum(real *A_d, real *C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + hl_matrix_column_op(aggregate::sum(), A_d, C_d, dimM, dimN); + + CHECK_SYNC("hl_matrix_column_sum failed"); +} + +void hl_matrix_column_max(real *A_d, real *C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + hl_matrix_column_op(aggregate::max(), A_d, C_d, dimM, dimN); + + CHECK_SYNC("hl_matrix_column_max failed"); +} + +void hl_matrix_column_min(real *A_d, real *C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + hl_matrix_column_op(aggregate::min(), A_d, C_d, dimM, dimN); + + CHECK_SYNC("hl_matrix_column_min failed"); +} + +template +__global__ void KeVectorSum(real *E, real *Sum, int dimM) { + __shared__ double sum_s[blockSize]; + int tid = threadIdx.x; + int index = blockIdx.y * blockDim.x + threadIdx.x; + + sum_s[tid] = 0.0f; + while (index < dimM) { + sum_s[tid] += E[index]; + index += blockDim.x * gridDim.y; + } + __syncthreads(); + + for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { + if (tid < stride) { + sum_s[tid] += sum_s[tid + stride]; + } + __syncthreads(); + } + __syncthreads(); + + if (tid == 0) { + Sum[blockIdx.y] = sum_s[0]; + } +} + +void hl_vector_sum(real *A_d, real *C_h, int dimM) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_h); + + int blockSize = 128; + int gridSize = 128; + int blocksX = 1; + int blocksY = gridSize; + dim3 threads(blockSize, 1); + dim3 grid(blocksX, blocksY); + + struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; + hl_event_t hl_event = &hl_event_st; + while (!hl_cuda_event_is_ready(hl_event)) { + } + + KeVectorSum<128><<>>( + A_d, t_resource.gpu_mem, dimM); + KeVectorSum<128><<<1, threads, 0, STREAM_DEFAULT>>>( + t_resource.gpu_mem, t_resource.cpu_mem, 128); + + hl_memcpy_async(C_h, t_resource.cpu_mem, sizeof(real), HPPL_STREAM_DEFAULT); + hl_stream_record_event(HPPL_STREAM_DEFAULT, hl_event); + + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + cudaError_t err = (cudaError_t)hl_get_device_last_error(); + CHECK_EQ(cudaSuccess, err) << "CUDA error: " + << hl_get_device_error_string((size_t)err); +} + +template +__global__ void KeVectorAbsSum(real *E, real *Sum, int dimM) { + __shared__ double sum_s[blockSize]; + int tid = threadIdx.x; + int index = blockIdx.y * blockDim.x + threadIdx.x; + + sum_s[tid] = 0.0f; + while (index < dimM) { + sum_s[tid] += abs(E[index]); + index += blockDim.x * gridDim.y; + } + __syncthreads(); + + for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { + if (tid < stride) { + sum_s[tid] += sum_s[tid + stride]; + } + __syncthreads(); + } + __syncthreads(); + + if (tid == 0) { + Sum[blockIdx.y] = sum_s[0]; + } +} + +void hl_vector_abs_sum(real *A_d, real *C_h, int dimM) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_h); + + int blockSize = 128; + int gridSize = 128; + int blocksX = 1; + int blocksY = gridSize; + dim3 threads(blockSize, 1); + dim3 grid(blocksX, blocksY); + + struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; + hl_event_t hl_event = &hl_event_st; + while (!hl_cuda_event_is_ready(hl_event)) { + } + + KeVectorAbsSum<128><<>>( + A_d, t_resource.gpu_mem, dimM); + KeVectorAbsSum<128><<<1, threads, 0, STREAM_DEFAULT>>>( + t_resource.gpu_mem, t_resource.cpu_mem, 128); + + hl_memcpy_async(C_h, t_resource.cpu_mem, sizeof(real), HPPL_STREAM_DEFAULT); + hl_stream_record_event(HPPL_STREAM_DEFAULT, hl_event); + + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + cudaError_t err = (cudaError_t)hl_get_device_last_error(); + CHECK_EQ(cudaSuccess, err) << "CUDA error: " + << hl_get_device_error_string((size_t)err); +} diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/legacy/cuda/src/hl_cuda_cnn.cu similarity index 100% rename from paddle/cuda/src/hl_cuda_cnn.cu rename to paddle/legacy/cuda/src/hl_cuda_cnn.cu diff --git a/paddle/legacy/cuda/src/hl_cuda_cublas.cc b/paddle/legacy/cuda/src/hl_cuda_cublas.cc new file mode 100644 index 0000000000000000000000000000000000000000..283b8b6e9c8e7b843a8d28b940c6ef53b77ef655 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_cublas.cc @@ -0,0 +1,400 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_cuda_cublas.h" +#include +#include "hl_cuda.h" +#include "hl_thread.ph" +#include "paddle/legacy/utils/DynamicLoader.h" +#include "paddle/legacy/utils/Logging.h" + +namespace dynload { + +std::once_flag cublas_dso_flag; +void *cublas_dso_handle = nullptr; + +/** + * The following macro definition can generate structs + * (for each function) to dynamic load cublas routine + * via operator overloading. + * + * note: default dynamic linked libs + */ +#ifdef PADDLE_USE_DSO +#define DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + cublasStatus_t operator()(Args... args) { \ + typedef cublasStatus_t (*cublasFunc)(Args...); \ + std::call_once(cublas_dso_flag, GetCublasDsoHandle, &cublas_dso_handle); \ + void *p_##__name = dlsym(cublas_dso_handle, #__name); \ + return reinterpret_cast(p_##__name)(args...); \ + } \ + } __name; // struct DynLoad__##__name +#else +#define DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + cublasStatus_t operator()(Args... args) { \ + return __name(args...); \ + } \ + } __name; // struct DynLoad__##__name +#endif + +#define DYNAMIC_LOAD_CUBLAS_V2_WRAP(__name) DYNAMIC_LOAD_CUBLAS_WRAP(__name) + +// include all needed cublas functions in HPPL +// clang-format off +#define CUBLAS_BLAS_ROUTINE_EACH(__macro) \ + __macro(cublasSgemv) \ + __macro(cublasDgemv) \ + __macro(cublasSgemm) \ + __macro(cublasDgemm) \ + __macro(cublasSgeam) \ + __macro(cublasDgeam) \ + +DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasCreate) +DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasDestroy) +DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetStream) +DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetPointerMode) +DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasGetPointerMode) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgemmBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgemmBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasCgemmBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasZgemmBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetrfBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetriBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetrfBatched) +DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetriBatched) +CUBLAS_BLAS_ROUTINE_EACH(DYNAMIC_LOAD_CUBLAS_V2_WRAP) + +#undef DYNAMIC_LOAD_CUBLAS_WRAP +#undef DYNAMIC_LOAD_CUBLAS_V2_WRAP +#undef CUBLAS_BLAS_ROUTINE_EACH + +} /* namespace dynload */ + +// clang-format on +#ifndef PADDLE_TYPE_DOUBLE +#define CUBLAS_GEAM dynload::cublasSgeam +#define CUBLAS_GEMV dynload::cublasSgemv +#define CUBLAS_GEMM dynload::cublasSgemm +#define CUBLAS_GETRF dynload::cublasSgetrfBatched +#define CUBLAS_GETRI dynload::cublasSgetriBatched +#else +#define CUBLAS_GEAM dynload::cublasDgeam +#define CUBLAS_GEMV dynload::cublasDgemv +#define CUBLAS_GEMM dynload::cublasDgemm +#define CUBLAS_GETRF dynload::cublasDgetrfBatched +#define CUBLAS_GETRI dynload::cublasDgetriBatched +#endif + +const char *hl_cublas_get_error_string(cublasStatus_t status) { + switch (status) { + case CUBLAS_STATUS_NOT_INITIALIZED: + return "[cublas status]: not initialized"; + case CUBLAS_STATUS_ALLOC_FAILED: + return "[cublas status]: allocate failed"; + case CUBLAS_STATUS_INVALID_VALUE: + return "[cublas status]: invalid value"; + case CUBLAS_STATUS_ARCH_MISMATCH: + return "[cublas status]: arch mismatch"; + case CUBLAS_STATUS_MAPPING_ERROR: + return "[cublas status]: mapping error"; + case CUBLAS_STATUS_EXECUTION_FAILED: + return "[cublas status]: execution failed"; + case CUBLAS_STATUS_INTERNAL_ERROR: + return "[cublas status]: internal error"; + case CUBLAS_STATUS_SUCCESS: + return "[cublas status]: success"; + default: + return "[cublas status]: unknown error"; + } +} + +/** + * Check build-in cublas function using glog and it also + * support << operator for more details error info. + */ +cublasStatus_t g_cublasStat; +#define CHECK_CUBLAS(cublas_func) \ + g_cublasStat = cublas_func; \ + CHECK_EQ(CUBLAS_STATUS_SUCCESS, g_cublasStat) \ + << "Cublas Error: " << hl_cublas_get_error_string(g_cublasStat) << " " + +void hl_cublas_init(cublasHandle_t *cublas_handle, cudaStream_t stream) { + CHECK_CUBLAS(dynload::cublasCreate(cublas_handle)) + << "[cublas init] Cublas create handle faild!"; + + CHECK_CUBLAS(dynload::cublasSetStream(*cublas_handle, stream)) + << "[cublas init] Cublas set stream faild!"; +} + +void hl_matrix_transpose( + real *A_d, real *C_d, int dimM, int dimN, int lda, int ldc) { + real alpha = 1.0; + real beta = 0.0; + + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + CHECK_CUBLAS(CUBLAS_GEAM(t_resource.handle, + CUBLAS_OP_T, + CUBLAS_OP_N, + dimM, + dimN, + &alpha, + A_d, + lda, + &beta, + nullptr, + dimM, + C_d, + ldc)); + CHECK_SYNC("hl_matrix_transpose failed"); +} + +void hl_matrix_transpose(real *A_d, real *C_d, int dimM, int dimN) { + hl_matrix_transpose(A_d, C_d, dimM, dimN, dimN, dimM); +} + +void hl_matrix_inverse(real *A_d, real *C_d, int dimN, int lda, int ldc) { + /* Solve Ax = I */ + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + /* Step 1: Compute the LU decomposition of matrix A */ + real **inout_h = &A_d; + real **inout_d = (real **)hl_malloc_device(sizeof(real *)); + hl_memcpy(inout_d, inout_h, sizeof(real *)); + + int *pivot_d = (int *)hl_malloc_device(dimN * sizeof(int)); + int *info_d = (int *)t_resource.gpu_mem; + + /* Note: cublasSgetrfBatched is used to calculate a number of + small-sized matrices. There may be a better way to reconstruct + the API for better performance. + */ + CHECK_CUBLAS( + CUBLAS_GETRF(t_resource.handle, dimN, inout_d, lda, pivot_d, info_d, 1)); + + int info_h; + hl_memcpy(&info_h, info_d, sizeof(int)); + if (info_h != 0) { + LOG(FATAL) << "Factorization of matrix failed: matrix may be singular.\n"; + } + + /* Step 2: Compute the inverse of the matrix given its LU decomposition */ + real **out_h = &C_d; + real **out_d = (real **)hl_malloc_device(sizeof(real *)); + hl_memcpy(out_d, out_h, sizeof(real *)); + + CHECK_CUBLAS(CUBLAS_GETRI(t_resource.handle, + dimN, + (const real **)inout_d, + lda, + pivot_d, + out_d, + ldc, + info_d, + 1)); + + hl_memcpy(&info_h, info_d, sizeof(int)); + if (info_h != 0) { + LOG(FATAL) << "Inversion of matrix failed: matrix may be singular.\n"; + } + + hl_free_mem_device(inout_d); + hl_free_mem_device(pivot_d); + hl_free_mem_device(out_d); + + CHECK_SYNC("hl_matrix_inverse failed"); +} + +void hl_matrix_mul(real *A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, + real *C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta, + int lda, + int ldb, + int ldc) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + + if (dimN == 1 && dimM != 1 && dimK != 1 && transb == HPPL_OP_N) { + int m = (transa == HPPL_OP_N) ? dimM : dimK; + int n = (transa == HPPL_OP_N) ? dimK : dimM; + hl_matrix_mul_vector( + A_d, transa, B_d, C_d, m, n, alpha, beta, lda, ldb, ldc); + return; + } + + if (dimM == 1 && dimN != 1 && dimK != 1 && transa == HPPL_OP_N) { + int m = (transb == HPPL_OP_N) ? dimK : dimN; + int n = (transb == HPPL_OP_N) ? dimN : dimK; + hl_trans_op_t trans = (transb == HPPL_OP_N) ? HPPL_OP_T : HPPL_OP_N; + hl_matrix_mul_vector(B_d, trans, A_d, C_d, m, n, alpha, beta, ldb, 1, 1); + return; + } + + cublasStatus_t stat; + if ((HPPL_OP_N == transa) && (HPPL_OP_N == transb)) { + stat = CUBLAS_GEMM(t_resource.handle, + CUBLAS_OP_N, + CUBLAS_OP_N, + dimN, + dimM, + dimK, + &alpha, + B_d, + ldb, + A_d, + lda, + &beta, + C_d, + ldc); + } else if ((HPPL_OP_T == transa) && (HPPL_OP_N == transb)) { + stat = CUBLAS_GEMM(t_resource.handle, + CUBLAS_OP_N, + CUBLAS_OP_T, + dimN, + dimM, + dimK, + &alpha, + B_d, + ldb, + A_d, + lda, + &beta, + C_d, + ldc); + } else if ((HPPL_OP_N == transa) && (HPPL_OP_T == transb)) { + stat = CUBLAS_GEMM(t_resource.handle, + CUBLAS_OP_T, + CUBLAS_OP_N, + dimN, + dimM, + dimK, + &alpha, + B_d, + ldb, + A_d, + lda, + &beta, + C_d, + ldc); + } else { + LOG(FATAL) << "parameter transa error!"; + } + CHECK_EQ(stat, CUBLAS_STATUS_SUCCESS) << hl_cublas_get_error_string(stat); + CHECK_SYNC("hl_matrix_mul failed"); +} + +void hl_matrix_mul(real *A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, + real *C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { + int lda = (HPPL_OP_N == transa) ? dimK : dimM; + int ldb = (HPPL_OP_N == transb) ? dimN : dimK; + int ldc = dimN; + + hl_matrix_mul(A_d, + transa, + B_d, + transb, + C_d, + dimM, + dimN, + dimK, + alpha, + beta, + lda, + ldb, + ldc); +} + +void hl_matrix_mul_vector(real *A_d, + hl_trans_op_t trans, + real *B_d, + real *C_d, + int dimM, + int dimN, + real alpha, + real beta, + int lda, + int incb, + int incc) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + + cublasStatus_t stat; + if (HPPL_OP_N == trans) { + stat = CUBLAS_GEMV(t_resource.handle, + CUBLAS_OP_T, + dimN, + dimM, + &alpha, + A_d, + lda, + B_d, + incb, + &beta, + C_d, + incc); + } else if (HPPL_OP_T == trans) { + stat = CUBLAS_GEMV(t_resource.handle, + CUBLAS_OP_N, + dimN, + dimM, + &alpha, + A_d, + lda, + B_d, + incb, + &beta, + C_d, + incc); + } else { + LOG(FATAL) << "parameter transa error!"; + } + + CHECK_EQ(stat, CUBLAS_STATUS_SUCCESS) << hl_cublas_get_error_string(stat); + CHECK_SYNC("hl_matrix_mul_vector"); +} + +void hl_matrix_mul_vector(real *A_d, + hl_trans_op_t trans, + real *B_d, + real *C_d, + int dimM, + int dimN, + real alpha, + real beta) { + hl_matrix_mul_vector( + A_d, trans, B_d, C_d, dimM, dimN, alpha, beta, dimN, 1, 1); +} diff --git a/paddle/legacy/cuda/src/hl_cuda_cudnn.cc b/paddle/legacy/cuda/src/hl_cuda_cudnn.cc new file mode 100644 index 0000000000000000000000000000000000000000..b0ac5aaac284cd939fc46be6a7320242312674ab --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_cudnn.cc @@ -0,0 +1,1117 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_cuda_cudnn.h" +#include +#include +#include "hl_cuda_cudnn.ph" +#include "hl_thread.ph" +#include "paddle/legacy/utils/DynamicLoader.h" +#include "paddle/legacy/utils/Logging.h" + +DEFINE_int32(cudnn_conv_workspace_limit_in_mb, + 4096, + "Specify cuDNN max workspace limit, in units MB, " + "4096MB=4GB by default."); + +namespace dynload { + +std::once_flag cudnn_dso_flag; +void* cudnn_dso_handle = nullptr; + +/** + * The following macro definition can generate structs + * (for each function) to dynamic load cudbnn routine + * via operator overloading: operator () + * + * note: default dynamic linked libs + **/ + +#ifdef PADDLE_USE_DSO + +#define DYNAMIC_LOAD_CUDNN_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + using cudnn_func = decltype(__name(args...)) (*)(Args...); \ + std::call_once(cudnn_dso_flag, GetCudnnDsoHandle, &cudnn_dso_handle); \ + void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ + return reinterpret_cast(p_##__name)(args...); \ + } \ + } __name; /* struct DynLoad__##__name */ + +#else + +#define DYNAMIC_LOAD_CUDNN_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + return __name(args...); \ + } \ + } __name; /* struct DynLoad__##__name */ + +#endif + +/** + * include all needed cudnn functions in HPPL + * different cudnn version has different interfaces + **/ +// clang-format off +#define CUDNN_DNN_ROUTINE_EACH(__macro) \ + __macro(cudnnSetTensor4dDescriptor) \ + __macro(cudnnSetTensor4dDescriptorEx) \ + __macro(cudnnGetConvolutionNdForwardOutputDim) \ + __macro(cudnnGetConvolutionForwardAlgorithm) \ + __macro(cudnnCreateTensorDescriptor) \ + __macro(cudnnDestroyTensorDescriptor) \ + __macro(cudnnCreateFilterDescriptor) \ + __macro(cudnnSetFilter4dDescriptor) \ + __macro(cudnnSetPooling2dDescriptor) \ + __macro(cudnnDestroyFilterDescriptor) \ + __macro(cudnnCreateConvolutionDescriptor) \ + __macro(cudnnCreatePoolingDescriptor) \ + __macro(cudnnDestroyPoolingDescriptor) \ + __macro(cudnnSetConvolution2dDescriptor) \ + __macro(cudnnDestroyConvolutionDescriptor) \ + __macro(cudnnCreate) \ + __macro(cudnnDestroy) \ + __macro(cudnnSetStream) \ + __macro(cudnnActivationForward) \ + __macro(cudnnConvolutionForward) \ + __macro(cudnnConvolutionBackwardBias) \ + __macro(cudnnGetConvolutionForwardWorkspaceSize) \ + __macro(cudnnTransformTensor) \ + __macro(cudnnPoolingForward) \ + __macro(cudnnPoolingBackward) \ + __macro(cudnnSoftmaxBackward) \ + __macro(cudnnSoftmaxForward) \ + __macro(cudnnGetVersion) \ + __macro(cudnnGetErrorString) +CUDNN_DNN_ROUTINE_EACH(DYNAMIC_LOAD_CUDNN_WRAP) + +#define CUDNN_DNN_ROUTINE_EACH_R2(__macro) \ + __macro(cudnnAddTensor) \ + __macro(cudnnConvolutionBackwardData) \ + __macro(cudnnConvolutionBackwardFilter) +CUDNN_DNN_ROUTINE_EACH_R2(DYNAMIC_LOAD_CUDNN_WRAP) + +// APIs available after R3: +#if CUDNN_VERSION >= 3000 +#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro) \ + __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize) \ + __macro(cudnnGetConvolutionBackwardDataAlgorithm) \ + __macro(cudnnGetConvolutionBackwardFilterAlgorithm) \ + __macro(cudnnGetConvolutionBackwardDataWorkspaceSize) +CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DYNAMIC_LOAD_CUDNN_WRAP) +#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R3 +#endif + + +// APIs available after R4: +#if CUDNN_VERSION >= 4007 +#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro) \ + __macro(cudnnBatchNormalizationForwardTraining) \ + __macro(cudnnBatchNormalizationForwardInference) \ + __macro(cudnnBatchNormalizationBackward) +CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DYNAMIC_LOAD_CUDNN_WRAP) +#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R4 +#endif + +// APIs in R5 +#if CUDNN_VERSION >= 5000 +#define CUDNN_DNN_ROUTINE_EACH_R5(__macro) \ + __macro(cudnnCreateActivationDescriptor) \ + __macro(cudnnSetActivationDescriptor) \ + __macro(cudnnGetActivationDescriptor) \ + __macro(cudnnDestroyActivationDescriptor) +CUDNN_DNN_ROUTINE_EACH_R5(DYNAMIC_LOAD_CUDNN_WRAP) +#undef CUDNN_DNN_ROUTINE_EACH_R5 +#endif + +#undef CUDNN_DNN_ROUTINE_EACH +// clang-format on +} /* namespace dynload */ + +/** + * Check build-in cudnn function using glog and it **does not** + * support << operator for more details error info. + */ +#define CHECK_CUDNN(cudnnFunc) \ + do { \ + cudnnStatus_t cudnnStat = cudnnFunc; \ + CHECK_EQ(CUDNN_STATUS_SUCCESS, cudnnStat) \ + << "Cudnn Error: " << dynload::cudnnGetErrorString(cudnnStat); \ + } while (0) + +bool g_is_libcudnn_init = false; +int g_cudnn_lib_version = 0; + +void hl_cudnn_desc_init(cudnnTensorDescriptor_t* cudnn_desc) { + CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(cudnn_desc)); +} + +void hl_cudnn_init(cudnnHandle_t* cudnn_handle, cudaStream_t stream) { + size_t cudnn_dso_ver = dynload::cudnnGetVersion(); + size_t cudnn_dso_major = cudnn_dso_ver / 1000; + size_t cudnn_cuh_major = CUDNN_VERSION / 1000; + + // Compare cudnn header version with that of cudnn.so. + CHECK((cudnn_cuh_major < 4 && cudnn_dso_major < 4) || + (cudnn_cuh_major == cudnn_dso_major)) + << "[cudnn init] libcudnn v" << cudnn_dso_major << " with header v" + << cudnn_cuh_major << " unmatched!\n" + << "PaddlePaddle Requirement: " + << "(header v[2-3] with libcudnn v[2-3]) Or " + << "(header v4 with libcudnn v4) Or " + << "(header v5 with libcudnn v5) Or" + << "(header v6 with libcudnn v6)."; + + CHECK(!(CUDNN_VERSION < 6000 && CUDNN_VERSION >= 5000 && CUDA_VERSION < 7050)) + << "cudnn v5 requires cuda version >= 7.5"; + + CHECK(!(CUDNN_VERSION >= 6000 && CUDA_VERSION < 8000)) + << "cudnn v6 requires cuda version >= 8.0"; + + CHECK_CUDNN(dynload::cudnnCreate(cudnn_handle)); + CHECK_CUDNN(dynload::cudnnSetStream(*cudnn_handle, stream)); + + g_is_libcudnn_init = true; + g_cudnn_lib_version = cudnn_dso_ver; +} + +int hl_get_cudnn_lib_version() { return g_cudnn_lib_version; } + +void hl_conv_workspace(hl_tensor_descriptor input, + hl_tensor_descriptor output, + hl_filter_descriptor filter, + hl_convolution_descriptor conv, + int* convFwdAlgo, + size_t* fwdLimitBytes, + int* convBwdDataAlgo, + size_t* bwdDataLimitBytes, + int* convBwdFilterAlgo, + size_t* bwdFilterLimitBytes, + bool useDilation) { +#if CUDNN_VERSION >= 4000 + + CHECK_NOTNULL(input); + CHECK_NOTNULL(output); + CHECK_NOTNULL(filter); + CHECK_NOTNULL(conv); + + // Specify workspace limit directly + size_t memoryLimitBytes = + (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb; + + // For dilation + int algo = 0; + + // cudnn convolution forward configuration + cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter); + cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + // cudnn convolution backward data configuration + cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter); + cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnConvolutionDescriptor_t bwd_data_conv_desc = + GET_CONVOLUTION_DESCRIPTOR(conv); + // cudnn convolution backward filter configuration + cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnConvolutionDescriptor_t bwd_filter_conv_desc = + GET_CONVOLUTION_DESCRIPTOR(conv); + cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter); + + if (useDilation) { + convFwdAlgo = &algo; + convBwdDataAlgo = &algo; + convBwdFilterAlgo = &algo; + } else { + CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm( + t_resource.cudnn_handle, + fwd_src_desc, + fwd_filter_desc, + fwd_conv_desc, + fwd_dest_desc, + CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convFwdAlgo))); + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm( + t_resource.cudnn_handle, + bwd_data_filter_desc, + bwd_data_diff_desc, + bwd_data_conv_desc, + bwd_data_grad_desc, + CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convBwdDataAlgo))); + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm( + t_resource.cudnn_handle, + bwd_filter_src_desc, + bwd_filter_diff_desc, + bwd_filter_conv_desc, + bwd_filter_grad_desc, + CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convBwdFilterAlgo))); + } + + CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize( + t_resource.cudnn_handle, + fwd_src_desc, + fwd_filter_desc, + fwd_conv_desc, + fwd_dest_desc, + static_cast(*convFwdAlgo), + fwdLimitBytes)); + + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( + t_resource.cudnn_handle, + bwd_data_filter_desc, + bwd_data_diff_desc, + bwd_data_conv_desc, + bwd_data_grad_desc, + static_cast(*convBwdDataAlgo), + bwdDataLimitBytes)); + + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( + t_resource.cudnn_handle, + bwd_filter_src_desc, + bwd_filter_diff_desc, + bwd_filter_conv_desc, + bwd_filter_grad_desc, + static_cast(*convBwdFilterAlgo), + bwdFilterLimitBytes)); + +#endif +} + +void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc, + int batch_size, + int feature_maps, + int height, + int width) { + CHECK_NOTNULL(image_desc); + + cudnn_tensor_descriptor hl_desc = + (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor)); + CHECK_NOTNULL(hl_desc); + +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(&hl_desc->desc)); + + CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(hl_desc->desc, + CUDNN_TENSOR_NCHW, + data_type, + batch_size, + feature_maps, + height, + width)); + + hl_desc->format = CUDNN_TENSOR_NCHW; + hl_desc->data_type = data_type; + hl_desc->batch_size = batch_size; + hl_desc->feature_maps = feature_maps; + hl_desc->height = height; + hl_desc->width = width; + + *image_desc = (hl_tensor_descriptor)hl_desc; +} + +void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc) { + CHECK_NOTNULL(image_desc); + + cudnn_tensor_descriptor hl_desc = + (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor)); + CHECK_NOTNULL(hl_desc); + +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(&hl_desc->desc)); + + hl_desc->data_type = data_type; + + *image_desc = (hl_tensor_descriptor)hl_desc; +} + +void hl_tensor_reshape(hl_tensor_descriptor image_desc, + int batch_size, + int feature_maps, + int height, + int width) { + const int stride_w = 1; + const int stride_h = width * stride_w; + const int stride_c = height * stride_h; + const int stride_n = feature_maps * stride_c; + return hl_tensor_reshape(image_desc, + batch_size, + feature_maps, + height, + width, + stride_n, + stride_c, + stride_h, + stride_w); +} + +void hl_tensor_reshape(hl_tensor_descriptor image_desc, + int batch_size, + int feature_maps, + int height, + int width, + int nStride, + int cStride, + int hStride, + int wStride) { + CHECK_NOTNULL(image_desc); + + cudnn_tensor_descriptor hl_desc = (cudnn_tensor_descriptor)image_desc; + CHECK_NOTNULL(hl_desc->desc); + + CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptorEx(hl_desc->desc, + hl_desc->data_type, + batch_size, + feature_maps, + height, + width, + nStride, + cStride, + hStride, + wStride)); + + hl_desc->batch_size = batch_size; + hl_desc->feature_maps = feature_maps; + hl_desc->height = height; + hl_desc->width = width; +} + +void hl_destroy_tensor_descriptor(hl_tensor_descriptor image_desc) { + CHECK_NOTNULL(image_desc); + + cudnn_tensor_descriptor hl_desc = (cudnn_tensor_descriptor)image_desc; + CHECK_NOTNULL(hl_desc->desc); + + CHECK_CUDNN(dynload::cudnnDestroyTensorDescriptor(hl_desc->desc)); + + hl_desc->desc = NULL; + + free(image_desc); +} + +void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc, + hl_pooling_mode_t mode, + int height, + int width, + int height_padding, + int width_padding, + int stride_height, + int stride_width) { + cudnnPoolingMode_t cudnn_mode; + switch (mode) { + case HL_POOLING_MAX: + cudnn_mode = CUDNN_POOLING_MAX; + break; + case HL_POOLING_AVERAGE: + cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + break; + case HL_POOLING_AVERAGE_INCLUDE_PADDING: + cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; + break; + default: + LOG(FATAL) << "parameter mode error"; + } + + CHECK_NOTNULL(pooling_desc); + + cudnn_pooling_descriptor hl_pooling_desc = + (cudnn_pooling_descriptor)malloc(sizeof(_cudnn_pooling_descriptor)); + CHECK_NOTNULL(hl_pooling_desc); + + CHECK_CUDNN(dynload::cudnnCreatePoolingDescriptor(&hl_pooling_desc->desc)); + + CHECK_CUDNN(dynload::cudnnSetPooling2dDescriptor(hl_pooling_desc->desc, + cudnn_mode, +#if CUDNN_VERSION >= 5000 + CUDNN_PROPAGATE_NAN, +#endif + height, + width, + height_padding, + width_padding, + stride_height, + stride_width)); + + hl_pooling_desc->mode = cudnn_mode; + hl_pooling_desc->window_height = height; + hl_pooling_desc->window_width = width; + hl_pooling_desc->stride_height = stride_height; + hl_pooling_desc->stride_width = stride_width; + + *pooling_desc = (hl_pooling_descriptor)hl_pooling_desc; +} + +void hl_destroy_pooling_descriptor(hl_pooling_descriptor pooling_desc) { + CHECK_NOTNULL(pooling_desc); + + cudnn_pooling_descriptor hl_pooling = (cudnn_pooling_descriptor)pooling_desc; + + CHECK_NOTNULL(hl_pooling->desc); + CHECK_CUDNN(dynload::cudnnDestroyPoolingDescriptor(hl_pooling->desc)); + + hl_pooling->desc = NULL; + + free(pooling_desc); +} + +void hl_pooling_forward(hl_tensor_descriptor input, + real* input_image, + hl_tensor_descriptor output, + real* output_image, + hl_pooling_descriptor pooling) { + cudnnPoolingDescriptor_t pooling_desc; + cudnnTensorDescriptor_t input_desc; + cudnnTensorDescriptor_t output_desc; + + CHECK_NOTNULL(input); + CHECK_NOTNULL(output); + CHECK_NOTNULL(pooling); + CHECK_NOTNULL(input_image); + CHECK_NOTNULL(output_image); + + real alpha = 1.0f; + real beta = 1.0f; + input_desc = ((cudnn_tensor_descriptor)input)->desc; + output_desc = ((cudnn_tensor_descriptor)output)->desc; + pooling_desc = ((cudnn_pooling_descriptor)pooling)->desc; + CHECK_CUDNN(dynload::cudnnPoolingForward(t_resource.cudnn_handle, + pooling_desc, + &alpha, + input_desc, + input_image, + &beta, + output_desc, + output_image)); + CHECK_SYNC("hl_pooling_forward failed"); +} + +void hl_pooling_backward(hl_tensor_descriptor input, + real* input_image, + real* input_image_grad, + hl_tensor_descriptor output, + real* output_image, + real* output_image_grad, + hl_pooling_descriptor pooling) { + cudnnPoolingDescriptor_t pooling_desc; + cudnnTensorDescriptor_t input_desc; + cudnnTensorDescriptor_t output_desc; + + CHECK_NOTNULL(input); + CHECK_NOTNULL(output); + CHECK_NOTNULL(pooling); + CHECK_NOTNULL(input_image); + CHECK_NOTNULL(input_image_grad); + CHECK_NOTNULL(output_image); + CHECK_NOTNULL(output_image_grad); + + real alpha = 1.0f; + real beta = 1.0f; + input_desc = ((cudnn_tensor_descriptor)input)->desc; + output_desc = ((cudnn_tensor_descriptor)output)->desc; + pooling_desc = ((cudnn_pooling_descriptor)pooling)->desc; + CHECK_CUDNN(dynload::cudnnPoolingBackward(t_resource.cudnn_handle, + pooling_desc, + &alpha, + output_desc, + output_image, + output_desc, + output_image_grad, + input_desc, + input_image, + &beta, + input_desc, + input_image_grad)); + CHECK_SYNC("hl_pooling_backward failed"); +} + +void hl_create_filter_descriptor(hl_filter_descriptor* filter, + int input_feature_maps, + int output_feature_maps, + int height, + int width) { + CHECK_NOTNULL(filter); + + cudnn_filter_descriptor hl_filter = + (cudnn_filter_descriptor)malloc(sizeof(_cudnn_filter_descriptor)); + CHECK_NOTNULL(hl_filter); + + CHECK_CUDNN(dynload::cudnnCreateFilterDescriptor(&hl_filter->desc)); + +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnSetFilter4dDescriptor(hl_filter->desc, + data_type, +#if CUDNN_VERSION >= 5000 + CUDNN_TENSOR_NCHW, +#endif + output_feature_maps, + input_feature_maps, + height, + width)); + + hl_filter->data_type = data_type; + hl_filter->output_feature_maps = output_feature_maps; + hl_filter->input_feature_maps = input_feature_maps; + hl_filter->filter_height = height; + hl_filter->filter_width = width; + + *filter = (hl_filter_descriptor)hl_filter; +} + +void hl_destroy_filter_descriptor(hl_filter_descriptor filter) { + CHECK_NOTNULL(filter); + + cudnn_filter_descriptor hl_filter = (cudnn_filter_descriptor)filter; + CHECK_NOTNULL(hl_filter->desc); + + CHECK_CUDNN(dynload::cudnnDestroyFilterDescriptor(hl_filter->desc)); + + hl_filter->desc = NULL; + + free(filter); +} + +void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, + hl_tensor_descriptor image, + hl_filter_descriptor filter, + int padding_height, + int padding_width, + int stride_height, + int stride_width, + int dilation_h, + int dilation_w) { + CHECK_NOTNULL(conv); + + cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc( + sizeof(_cudnn_convolution_descriptor)); + + CHECK_NOTNULL(hl_conv); + CHECK_CUDNN(dynload::cudnnCreateConvolutionDescriptor(&hl_conv->desc)); + + cudnnConvolutionMode_t mode = CUDNN_CROSS_CORRELATION; + +#if CUDNN_VERSION >= 6000 +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc, + padding_height, + padding_width, + stride_height, + stride_width, + dilation_h, + dilation_w, + mode, + data_type)); +#else + if (dilation_h > 1 || dilation_w > 1) { + LOG(FATAL) + << "Current cuDNN version does't support for dilation convolution. " + << "The dilation convolution requires cuDNN >= v6.0."; + } + + CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc, + padding_height, + padding_width, + stride_height, + stride_width, + dilation_h, + dilation_w, + mode)); +#endif + + hl_conv->input_image = image; + hl_conv->filter = filter; + hl_conv->padding_height = padding_height; + hl_conv->padding_width = padding_width; + hl_conv->stride_height = stride_height; + hl_conv->stride_width = stride_width; + hl_conv->upscalex = 1; + hl_conv->upscaley = 1; + hl_conv->mode = mode; + + *conv = (hl_convolution_descriptor)hl_conv; +} + +void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, + hl_tensor_descriptor image, + hl_filter_descriptor filter, + int padding_height, + int padding_width, + int stride_height, + int stride_width, + int dilation_h, + int dilation_w) { + CHECK_NOTNULL(conv); + CHECK_NOTNULL(image); + CHECK_NOTNULL(filter); + + cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + cudnnConvolutionMode_t mode = CUDNN_CROSS_CORRELATION; + +#if CUDNN_VERSION >= 6000 +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(conv_desc, + padding_height, + padding_width, + stride_height, + stride_width, + dilation_h, + dilation_w, + mode, + data_type)); +#else + CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(conv_desc, + padding_height, + padding_width, + stride_height, + stride_width, + dilation_h, + dilation_w, + mode)); +#endif + + cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)conv; + hl_conv->input_image = image; + hl_conv->filter = filter; + hl_conv->padding_height = padding_height; + hl_conv->padding_width = padding_width; + hl_conv->stride_height = stride_height; + hl_conv->stride_width = stride_width; + hl_conv->upscalex = 1; + hl_conv->upscaley = 1; + hl_conv->mode = mode; +} + +void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) { + CHECK_NOTNULL(conv); + + cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)conv; + CHECK_NOTNULL(hl_conv->desc); + + CHECK_CUDNN(dynload::cudnnDestroyConvolutionDescriptor(hl_conv->desc)); + hl_conv->desc = NULL; + + free(conv); +} + +void hl_convolution_forward(hl_tensor_descriptor input, + real* input_data, + hl_tensor_descriptor output, + real* output_data, + hl_filter_descriptor filter, + real* filter_data, + hl_convolution_descriptor conv, + void* gpuWorkSpace, + size_t sizeInBytes, + int convFwdAlgo) { + CHECK_NOTNULL(input); + CHECK_NOTNULL(output); + CHECK_NOTNULL(filter); + CHECK_NOTNULL(conv); + CHECK_NOTNULL(input_data); + CHECK_NOTNULL(output_data); + CHECK_NOTNULL(filter_data); + cudnnTensorDescriptor_t src_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnTensorDescriptor_t dest_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnFilterDescriptor_t filter_desc = GET_FILTER_DESCRIPTOR(filter); + cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + real alpha = 1.0f; + real beta = 1.0f; + CHECK_CUDNN(dynload::cudnnConvolutionForward( + t_resource.cudnn_handle, + &alpha, + src_desc, + input_data, + filter_desc, + filter_data, + conv_desc, + static_cast(convFwdAlgo), + gpuWorkSpace, + sizeInBytes, + &beta, + dest_desc, + output_data)); + CHECK_SYNC("hl_convolution_forward failed"); +} + +void hl_convolution_forward_add_bias(hl_tensor_descriptor bias, + real* bias_data, + hl_tensor_descriptor output, + real* output_data) { + CHECK_NOTNULL(bias); + CHECK_NOTNULL(output); + CHECK_NOTNULL(bias_data); + CHECK_NOTNULL(output_data); + + cudnnTensorDescriptor_t output_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnTensorDescriptor_t bias_desc = GET_TENSOR_DESCRIPTOR(bias); + real alpha = 1.0f; + real beta = 1.0f; + + CHECK_CUDNN(dynload::cudnnAddTensor(t_resource.cudnn_handle, +#if CUDNN_VERSION < 4000 + CUDNN_ADD_SAME_C, +#endif + &alpha, + bias_desc, + bias_data, + &beta, + output_desc, + output_data)); + CHECK_SYNC("hl_convolution_forward_add_bias failed"); +} + +void hl_convolution_backward_bias(hl_tensor_descriptor bias, + real* bias_grad_data, + hl_tensor_descriptor output, + real* output_grad_data) { + CHECK_NOTNULL(bias); + CHECK_NOTNULL(output); + CHECK_NOTNULL(bias_grad_data); + CHECK_NOTNULL(output_grad_data); + + real alpha = 1.0f; + real beta = 1.0f; + cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnTensorDescriptor_t bias_desc = GET_TENSOR_DESCRIPTOR(bias); + CHECK_CUDNN(dynload::cudnnConvolutionBackwardBias(t_resource.cudnn_handle, + &alpha, + diff_desc, + output_grad_data, + &beta, + bias_desc, + bias_grad_data)); + CHECK_SYNC("hl_convolution_backward_bias failed"); +} + +void hl_convolution_backward_filter(hl_tensor_descriptor input, + real* input_data, + hl_tensor_descriptor output, + real* output_grad_data, + hl_filter_descriptor filter, + real* filter_grad_data, + hl_convolution_descriptor conv, + void* gpuWorkSpace, + size_t sizeInBytes, + int convBwdFilterAlgo) { + CHECK_NOTNULL(input); + CHECK_NOTNULL(output); + CHECK_NOTNULL(filter); + CHECK_NOTNULL(conv); + CHECK_NOTNULL(input_data); + CHECK_NOTNULL(output_grad_data); + CHECK_NOTNULL(filter_grad_data); + + real alpha = 1.0f; + real beta = 1.0f; + cudnnTensorDescriptor_t src_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + cudnnFilterDescriptor_t grad_desc = GET_FILTER_DESCRIPTOR(filter); + + CHECK_CUDNN(dynload::cudnnConvolutionBackwardFilter( + t_resource.cudnn_handle, + &alpha, + src_desc, + input_data, + diff_desc, + output_grad_data, + conv_desc, +#if CUDNN_VERSION >= 4000 + static_cast(convBwdFilterAlgo), + gpuWorkSpace, + sizeInBytes, +#endif + &beta, + grad_desc, + filter_grad_data)); + CHECK_SYNC("hl_convolution_backward_filter failed"); +} + +void hl_convolution_backward_data(hl_tensor_descriptor input, + real* input_data_grad, + hl_tensor_descriptor output, + real* output_grad_data, + hl_filter_descriptor filter, + real* filter_data, + hl_convolution_descriptor conv, + void* gpuWorkSpace, + size_t sizeInBytes, + int convBwdDataAlgo) { + real alpha = 1.0f; + real beta = 1.0f; + cudnnFilterDescriptor_t filter_desc = GET_FILTER_DESCRIPTOR(filter); + cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnTensorDescriptor_t grad_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + + CHECK_CUDNN(dynload::cudnnConvolutionBackwardData( + t_resource.cudnn_handle, + &alpha, + filter_desc, + filter_data, + diff_desc, + output_grad_data, + conv_desc, +#if CUDNN_VERSION >= 4000 + static_cast(convBwdDataAlgo), + gpuWorkSpace, + sizeInBytes, +#endif + &beta, + grad_desc, + input_data_grad)); + CHECK_SYNC("hl_convolution_backward_data failed"); +} + +void hl_softmax_forward(real* input, real* output, int height, int width) { +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(t_resource.cudnn_desc, + CUDNN_TENSOR_NCHW, + data_type, + height, + width, + 1, + 1)); + + real alpha = 1.0f; + real beta = 0.0f; + CHECK_CUDNN(dynload::cudnnSoftmaxForward(t_resource.cudnn_handle, + CUDNN_SOFTMAX_ACCURATE, + CUDNN_SOFTMAX_MODE_CHANNEL, + &alpha, + t_resource.cudnn_desc, + input, + &beta, + t_resource.cudnn_desc, + output)); + CHECK_SYNC("hl_softmax_forward failed"); +} + +void hl_softmax_backward(real* output_value, + real* output_grad, + int height, + int width) { +#ifndef PADDLE_TYPE_DOUBLE + cudnnDataType_t data_type = CUDNN_DATA_FLOAT; +#else + cudnnDataType_t data_type = CUDNN_DATA_DOUBLE; +#endif + CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(t_resource.cudnn_desc, + CUDNN_TENSOR_NCHW, + data_type, + height, + width, + 1, + 1)); + + real alpha = 1.0f; + real beta = 0.0f; + CHECK_CUDNN(dynload::cudnnSoftmaxBackward(t_resource.cudnn_handle, + CUDNN_SOFTMAX_ACCURATE, + CUDNN_SOFTMAX_MODE_CHANNEL, + &alpha, + t_resource.cudnn_desc, + output_value, + t_resource.cudnn_desc, + output_grad, + &beta, + t_resource.cudnn_desc, + output_grad)); + CHECK_SYNC("hl_softmax_backward failed"); +} + +void hl_batch_norm_forward_training(hl_tensor_descriptor inputDesc, + real* input, + hl_tensor_descriptor outputDesc, + real* output, + hl_tensor_descriptor bnParamDesc, + real* scale, + real* bias, + double factor, + real* runningMean, + real* runningInvVar, + double epsilon, + real* savedMean, + real* savedVar) { +#if CUDNN_VERSION >= 4007 + if ((NULL != runningMean && NULL == runningInvVar) || + (NULL == runningMean && NULL != runningInvVar)) { + LOG(FATAL) << "runningMean and runningInvVar can be NULL " + << "but only at the same time."; + } + if ((NULL != savedMean && NULL == savedVar) || + (NULL == savedMean && NULL != savedVar)) { + LOG(FATAL) << "savedMean and savedVar can be NULL " + << "but only at the same time."; + } + + cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc); + cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc); + cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc); + real alpha = 1.0f; + real beta = 1.0f; + cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; + CHECK_CUDNN( + dynload::cudnnBatchNormalizationForwardTraining(t_resource.cudnn_handle, + mode, + &alpha, + &beta, + xDesc, + input, + yDesc, + output, + bnDesc, + scale, + bias, + factor, + runningMean, + runningInvVar, + epsilon, + savedMean, + savedVar)); + + CHECK_SYNC("hl_batch_norm_forward_training failed"); +#else + LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. " + << "But cudnn lib version is " << g_cudnn_lib_version; +#endif +} + +void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc, + real* input, + hl_tensor_descriptor outputDesc, + real* output, + hl_tensor_descriptor bnParamDesc, + real* scale, + real* bias, + real* estimatedMean, + real* estimatedInvVar, + double epsilon) { +#if CUDNN_VERSION >= 4007 + cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc); + cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc); + cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc); + real alpha = 1.0f; + real beta = 1.0f; + cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; + + CHECK_CUDNN( + dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle, + mode, + &alpha, + &beta, + xDesc, + input, + yDesc, + output, + bnDesc, + scale, + bias, + estimatedMean, + estimatedInvVar, + epsilon)); + + CHECK_SYNC("hl_batch_norm_forward_inference failed"); +#else + LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. " + << "But cudnn lib version is " << g_cudnn_lib_version; +#endif +} + +void hl_batch_norm_backward(hl_tensor_descriptor inputDesc, + real* input, + hl_tensor_descriptor outGradDesc, + real* outGrad, + hl_tensor_descriptor inGradDesc, + real* inGrad, + hl_tensor_descriptor dBnParamDesc, + real* scale, + real* scaleGrad, + real* biasGrad, + double epsilon, + real* savedMean, + real* savedInvVar) { +#if CUDNN_VERSION >= 4007 + if ((NULL != savedMean && NULL == savedInvVar) || + (NULL == savedMean && NULL != savedInvVar)) { + LOG(FATAL) << "savedMean and savedVar can be NULL " + << "but only at the same time."; + } + + cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc); + cudnnTensorDescriptor_t dyDesc = GET_TENSOR_DESCRIPTOR(outGradDesc); + cudnnTensorDescriptor_t dxDesc = GET_TENSOR_DESCRIPTOR(inGradDesc); + cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(dBnParamDesc); + real alpha = 1.0f; + real beta = 1.0f; + cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; + CHECK_CUDNN(dynload::cudnnBatchNormalizationBackward(t_resource.cudnn_handle, + mode, + &alpha, + &beta, + &alpha, + &beta, + xDesc, + input, + dyDesc, + outGrad, + dxDesc, + inGrad, + bnDesc, + scale, + scaleGrad, + biasGrad, + epsilon, + savedMean, + savedInvVar)); + + CHECK_SYNC("hl_batch_norm_backward failed"); +#else + LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. " + << "But cudnn lib version is " << g_cudnn_lib_version; +#endif +} diff --git a/paddle/legacy/cuda/src/hl_cuda_device.cc b/paddle/legacy/cuda/src/hl_cuda_device.cc new file mode 100644 index 0000000000000000000000000000000000000000..501e3b0f3be02b9364f9182b2484d542f0f39889 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_device.cc @@ -0,0 +1,677 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +// clang-format off +// Because clang-format 4.X and clang-format 3.8+ format +// following lines in different. So disable clang-format. +#include "hl_cuda.h" +#include +#include +#include +#include +#include +#include "hl_cuda.ph" +#include "hl_thread.ph" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/DynamicLoader.h" +// clang-format on + +namespace dynload { + +std::once_flag curand_dso_flag; +void *curand_dso_handle = nullptr; + +/** + * The following macro definition can generate structs + * (for each function) to dynamic load curand routine + * via operator overloading. + * + * note: default dynamic linked libs + */ +#ifdef PADDLE_USE_DSO +#define DYNAMIC_LOAD_CURAND_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + curandStatus_t operator()(Args... args) { \ + typedef curandStatus_t (*curandFunc)(Args...); \ + std::call_once(curand_dso_flag, GetCurandDsoHandle, &curand_dso_handle); \ + void *p_##__name = dlsym(curand_dso_handle, #__name); \ + return reinterpret_cast(p_##__name)(args...); \ + } \ + } __name; /* struct DynLoad__##__name */ +#else +#define DYNAMIC_LOAD_CURAND_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + curandStatus_t operator()(Args... args) { \ + return __name(args...); \ + } \ + } __name; /* struct DynLoad__##__name */ +#endif + +/* include all needed curand functions in HPPL */ +// clang-format off +#define CURAND_RAND_ROUTINE_EACH(__macro) \ + __macro(curandCreateGenerator) \ + __macro(curandSetStream) \ + __macro(curandSetPseudoRandomGeneratorSeed)\ + __macro(curandGenerateUniform) \ + __macro(curandGenerateUniformDouble) +// clang-format on + +CURAND_RAND_ROUTINE_EACH(DYNAMIC_LOAD_CURAND_WRAP) + +#undef CURAND_RAND_ROUTINE_EACH +#undef DYNAMIC_LOAD_CURAND_WRAP + +} /* namespace dynload */ + +/** + * @brief global resource. + */ +int g_system_device_num = 0; /* system device number */ +int device_num = 0; /* use device number */ +hl_device_prop *g_device; /* device info table */ +__thread thread_device_resources *t_device; /* device resources table */ +int g_cuda_lib_version = 0; + +/* number of global stream */ +#define NUMBER_OF_GLOBAL_STREAM (HPPL_THREAD_STREAM_1) +/* number of thread stream */ +#define NUMBER_OF_THREAD_STREAM (HPPL_STREAM_END - HPPL_THREAD_STREAM_1) +/* sizeof of device memory */ +#define HPPL_GPU_MEMORY_SIZE (256 * 4) + +/** + * Check build-in cuda function using glog and it **does not** + * support << operator for more details error info. + */ +#define CHECK_CUDA(cudaFunc) \ + do { \ + cudaError_t cudaStat = cudaFunc; \ + CHECK_EQ(cudaSuccess, cudaStat) << "Cuda Error: " \ + << cudaGetErrorString(cudaStat); \ + } while (0) + +/** + * @brief thread resource. + */ +__thread _hl_thread_resource t_resource = {{0}, /* stream */ + 0, /* handle */ + 0, /* gen */ + 0, /* cudnn_handle */ + 0, /* cudnn_desc */ + NULL, /* gen_mutex */ + NULL, /* gpu_mem */ + NULL, /* cpu_mem */ + 0, /* event */ + -1, /* device */ + 0, /* major */ + false}; /* is_init */ + +__thread cudaStream_t default_stream = 0; +__thread bool g_sync_flag = true; +bool hl_start_flag = false; + +inline pid_t gettid() { +#if defined(__APPLE__) || defined(__OSX__) + // syscall is deprecated: first deprecated in macOS 10.12. + // syscall is unsupported; + // syscall pid_t tid = syscall(SYS_thread_selfid); + uint64_t tid; + pthread_threadid_np(NULL, &tid); +#else +#ifndef __NR_gettid +#define __NR_gettid 224 +#endif + pid_t tid = syscall(__NR_gettid); +#endif + CHECK_NE((int)tid, -1); + return tid; +} + +void hl_init(int device) { + CHECK(hl_start_flag) << "[Init failed] hl_start() did not succeed."; + + /* thread has been initialized */ + if (true == t_resource.is_init) { + hl_set_device(device); + return; + } + + /* create thread devcie resources */ + char *tmp; + thread_device_resources device_res; + tmp = (char *)malloc(g_system_device_num * sizeof(thread_device_resources *) + + device_num * sizeof(_thread_device_resources)); + CHECK_NOTNULL(tmp); + t_device = (thread_device_resources *)tmp; + device_res = (thread_device_resources)( + (char *)tmp + g_system_device_num * sizeof(thread_device_resources *)); + memset(t_device, 0, g_system_device_num * sizeof(thread_device_resources *)); + + char *tmp_stream = (char *)malloc(device_num * NUMBER_OF_THREAD_STREAM * + sizeof(cudaStream_t)); + CHECK_NOTNULL(tmp_stream); + + int num = 0; + for (int dev = 0; dev < g_system_device_num; dev++) { + if (!g_device[dev]) { + continue; + } + + t_device[dev] = &device_res[num]; + t_device[dev]->stream = + (cudaStream_t *)(tmp_stream + + num * NUMBER_OF_THREAD_STREAM * sizeof(cudaStream_t)); + + hl_create_thread_resources(dev, t_device[dev]); + num++; + } + + hl_cudnn_desc_init(&t_resource.cudnn_desc); + + /* thread initialization is complete */ + t_resource.is_init = true; + /* set device */ + t_resource.device = -1; + hl_set_device(device); +} + +void hl_fini() { + if (false == t_resource.is_init) { + return; + } + + /* hppl stream fini */ + t_resource.device = -1; + for (int i = NUMBER_OF_GLOBAL_STREAM; i < HPPL_STREAM_END; i++) { + t_resource.stream[i] = 0; + } + + char *tmp = (char *)t_device; + char *tmp_stream = NULL; + for (int dev = 0; dev < g_system_device_num; dev++) { + if (!t_device[dev]) { + continue; + } + if (!tmp_stream) { + tmp_stream = (char *)t_device[dev]->stream; + } + for (int j = 0; j < NUMBER_OF_THREAD_STREAM; j++) { + CHECK_CUDA(cudaStreamDestroy(t_device[dev]->stream[j])); + } + + /* free device memory */ + hl_free_mem_device(t_device[dev]->gpu_mem); + hl_free_mem_host(t_device[dev]->cpu_mem); + CHECK_CUDA(cudaEventDestroy(t_device[dev]->mem_event)); + } + + free(tmp); + free(tmp_stream); + t_resource.is_init = false; +} + +int hl_get_device_count() { return device_num; } + +void hl_set_device(int device) { + if (device == t_resource.device) { + return; + } + + CHECK(device >= 0 && device < g_system_device_num && g_device[device]) + << "Device: " << device << " is not specified in startup."; + + CHECK_CUDA(cudaSetDevice(device)); + + /* switch thread stream */ + for (int i = 0; i < NUMBER_OF_GLOBAL_STREAM; i++) { + t_resource.stream[i] = g_device[device]->device_resources->stream[i]; + } + + if (true == t_resource.is_init) { + for (int i = NUMBER_OF_GLOBAL_STREAM; i < HPPL_STREAM_END; i++) { + t_resource.stream[i] = + t_device[device]->stream[i - NUMBER_OF_GLOBAL_STREAM]; + } + t_resource.gpu_mem = t_device[device]->gpu_mem; + t_resource.cpu_mem = t_device[device]->cpu_mem; + t_resource.event = t_device[device]->mem_event; + } + + t_resource.handle = g_device[device]->device_resources->handle; + t_resource.gen = g_device[device]->device_resources->gen; + t_resource.cudnn_handle = g_device[device]->device_resources->cudnn_handle; + t_resource.gen_mutex = g_device[device]->device_resources->gen_mutex; + t_resource.device = device; + t_resource.major = g_device[device]->major; + default_stream = t_resource.stream[0]; +} + +int hl_get_device() { + int device; + CHECK_CUDA(cudaGetDevice(&device)); + return device; +} + +void *hl_malloc_device(size_t size) { + void *dest_d; + + CHECK(size) << __func__ << ": the size for device memory is 0, please check."; + CHECK_CUDA(cudaMalloc((void **)&dest_d, size)); + + return dest_d; +} + +void hl_free_mem_device(void *dest_d) { + CHECK_NOTNULL(dest_d); + + cudaError_t err = cudaFree(dest_d); + CHECK(cudaSuccess == err || cudaErrorCudartUnloading == err) + << hl_get_device_error_string(); +} + +void *hl_malloc_host(size_t size) { + void *dest_h; + + CHECK(size) << __func__ << ": the size for device memory is 0, please check."; + CHECK_CUDA(cudaHostAlloc((void **)&dest_h, size, cudaHostAllocDefault)); + + return dest_h; +} + +void hl_free_mem_host(void *dest_h) { + CHECK_NOTNULL(dest_h); + + cudaError_t err = cudaFreeHost(dest_h); + CHECK(cudaSuccess == err || cudaErrorCudartUnloading == err) + << hl_get_device_error_string(); +} + +void hl_memcpy(void *dst, void *src, size_t size) { + if (0 == size) { + return; + } + CHECK_NOTNULL(dst); + CHECK_NOTNULL(src); + CHECK_CUDA(cudaMemcpy(dst, src, size, cudaMemcpyDefault)); +} + +void hl_memset_device(void *dest_d, int value, size_t size) { + CHECK_CUDA(cudaMemset(dest_d, value, size)); +} + +void hl_memcpy_host2device(void *dest_d, void *src_h, size_t size) { + if (0 == size) { + return; + } + CHECK_NOTNULL(src_h); + CHECK_NOTNULL(dest_d); + CHECK_CUDA(cudaMemcpy(dest_d, src_h, size, cudaMemcpyHostToDevice)); +} + +void hl_memcpy_device2host(void *dest_h, void *src_d, size_t size) { + if (0 == size) { + return; + } + CHECK_NOTNULL(dest_h); + CHECK_NOTNULL(src_d); + CHECK_CUDA(cudaMemcpy(dest_h, src_d, size, cudaMemcpyDeviceToHost)); +} + +void hl_memcpy_device2device(void *dest_d, void *src_d, size_t size) { + if (0 == size) { + return; + } + CHECK_NOTNULL(dest_d); + CHECK_NOTNULL(src_d); + CHECK_CUDA(cudaMemcpy(dest_d, src_d, size, cudaMemcpyDeviceToDevice)); +} + +void hl_memcpy_async(void *dst, void *src, size_t size, hl_stream_t stream) { + cudaStream_t cu_stream; + + if (0 == size) { + return; + } + CHECK_NOTNULL(dst); + CHECK_NOTNULL(src); + CHECK_LT(stream, HPPL_STREAM_END); + cu_stream = t_resource.stream[stream]; + + CHECK_CUDA(cudaMemcpyAsync(dst, src, size, cudaMemcpyDefault, cu_stream)); +} + +void hl_start() { + hl_specify_devices_start(NULL, 0); + /* set default device */ + hl_set_device(0); +} + +bool hl_device_can_access_peer(int device, int peerDevice) { + int canAccessPeer; + CHECK_CUDA(cudaDeviceCanAccessPeer(&canAccessPeer, device, peerDevice)); + + if (canAccessPeer == 1) { + return true; + } else { + return false; + } +} + +void hl_device_enable_peer_access(int peerDevice) { + cudaError_t err = cudaDeviceEnablePeerAccess(peerDevice, 0); + if (cudaErrorPeerAccessAlreadyEnabled == err) { + cudaGetLastError(); + } else { + CHECK_CUDA(err); + } +} + +void hl_create_global_resources(hl_device_prop device_prop) { + struct cudaDeviceProp cu_prop; + int device = device_prop->device; + global_device_resources device_res = device_prop->device_resources; + + CHECK_CUDA(cudaSetDevice(device)); + /* device properties */ + CHECK_CUDA(cudaGetDeviceProperties(&cu_prop, device)); + + device_prop->major = cu_prop.major; + device_prop->minor = cu_prop.minor; + strncpy(device_prop->device_name, cu_prop.name, 256); + device_prop->device_mem = cu_prop.totalGlobalMem; + + /* create device stream */ + for (int j = 0; j < NUMBER_OF_GLOBAL_STREAM; j++) { + CHECK_CUDA(cudaStreamCreate(&device_res->stream[j])); + } + + /* cublas init */ + hl_cublas_init(&device_res->handle, device_res->stream[0]); + + /* create curand gen */ + CHECK_EQ(dynload::curandCreateGenerator(&device_res->gen, + CURAND_RNG_PSEUDO_DEFAULT), + CURAND_STATUS_SUCCESS) + << "[Start failed] Curand init failed."; + + CHECK_EQ(dynload::curandSetStream(device_res->gen, device_res->stream[0]), + CURAND_STATUS_SUCCESS) + << "[Start failed] Curand set stream failed!"; + + /* create cudnn handle */ + hl_cudnn_init(&device_res->cudnn_handle, device_res->stream[0]); + + int seed = gettid(); + CHECK_EQ(dynload::curandSetPseudoRandomGeneratorSeed(device_res->gen, + seed + device), + CURAND_STATUS_SUCCESS); + + device_res->gen_mutex = (pthread_mutex_t *)(malloc(sizeof(pthread_mutex_t))); + pthread_mutex_init(device_res->gen_mutex, NULL); + + CHECK_CUDA(cudaRuntimeGetVersion(&g_cuda_lib_version)); +} + +int hl_get_cuda_version() { return g_cuda_lib_version; } + +void hl_create_thread_resources(int device, + thread_device_resources device_res) { + CHECK_CUDA(cudaSetDevice(device)); + + /* create thread stream */ + for (int j = 0; j < NUMBER_OF_THREAD_STREAM; j++) { + CHECK_CUDA(cudaStreamCreate(&device_res->stream[j])); + } + + /* allocation device memory */ + device_res->gpu_mem = (real *)hl_malloc_device(HPPL_GPU_MEMORY_SIZE); + + /* allocation host memory */ + device_res->cpu_mem = (real *)hl_malloc_host(HPPL_GPU_MEMORY_SIZE); + + CHECK_CUDA(cudaEventCreate(&device_res->mem_event)); +} + +void hl_specify_devices_start(int *device, int number) { + if (hl_start_flag) return; + + /* 1. get the number of devices */ + CHECK_CUDA(cudaGetDeviceCount(&g_system_device_num)); + CHECK_NE(g_system_device_num, 0) << "[Start failed] there is no GPU device"; + if (device == NULL) { + number = g_system_device_num; + } + + /* 2. check device & create device property table */ + CHECK_LE(number, g_system_device_num) + << "[Start failed] System does not have enough device. " + << "Device number: " << g_system_device_num << "Input number: " << number; + + char *tmp; + hl_device_prop device_prop; + tmp = (char *)malloc(g_system_device_num * sizeof(hl_device_prop *) + + number * sizeof(_hl_device_prop)); + CHECK(tmp) << "[Start failed] System memory is not enough."; + + g_device = (hl_device_prop *)tmp; + device_prop = (hl_device_prop)( + (char *)tmp + g_system_device_num * sizeof(hl_device_prop *)); + memset(g_device, 0, g_system_device_num * sizeof(hl_device_prop *)); + int num = 0; + for (int i = 0; i < number; i++) { + int dev; + if (device == NULL) { + dev = i; + } else { + dev = device[i]; + } + + CHECK_LT(dev, g_system_device_num) + << "[Start failed] The specified device number is " + << "out of range. Max device number: " << g_system_device_num - 1 + << " Specified devcie number: " << dev; + + if (g_device[dev]) { + /* Warning */ + LOG(WARNING) << "[Warning] Repeat specify device: " << dev; + continue; + } + + g_device[dev] = &device_prop[num]; + g_device[dev]->device = dev; + num++; + } + device_num = num; + + /* 3. create global device resources */ + char *tmp_res = (char *)malloc(device_num * sizeof(_global_device_resources)); + CHECK_NOTNULL(tmp_res); + + char *tmp_stream = (char *)malloc(device_num * NUMBER_OF_GLOBAL_STREAM * + sizeof(cudaStream_t)); + CHECK_NOTNULL(tmp_stream); + + num = 0; + for (int i = 0; i < g_system_device_num; i++) { + if (!g_device[i]) { + continue; + } + + g_device[i]->device_resources = (global_device_resources)( + tmp_res + num * sizeof(_global_device_resources)); + g_device[i]->device_resources->stream = + (cudaStream_t *)(tmp_stream + + num * NUMBER_OF_GLOBAL_STREAM * sizeof(cudaStream_t)); + + hl_create_global_resources(g_device[i]); + num++; + } + + /* hl_start() is ok */ + hl_start_flag = true; + /* set default device */ + if (device == NULL) { + hl_set_device(0); + } else { + hl_set_device(device[0]); + } +} + +void hl_rand(real *dest_d, size_t num) { + pthread_mutex_lock(t_resource.gen_mutex); + CHECK_EQ( +#ifndef PADDLE_TYPE_DOUBLE + dynload::curandGenerateUniform(t_resource.gen, dest_d, num), +#else + dynload::curandGenerateUniformDouble(t_resource.gen, dest_d, num), +#endif + CURAND_STATUS_SUCCESS); + pthread_mutex_unlock(t_resource.gen_mutex); + CHECK_SYNC("hl_rand failed"); +} + +void hl_srand(unsigned int seed) { + pthread_mutex_lock(t_resource.gen_mutex); + CHECK_EQ(dynload::curandSetPseudoRandomGeneratorSeed(t_resource.gen, seed), + CURAND_STATUS_SUCCESS); + pthread_mutex_unlock(t_resource.gen_mutex); +} + +void hl_set_sync_flag(bool flag) { g_sync_flag = flag; } + +bool hl_get_sync_flag() { return g_sync_flag; } + +void hl_stream_synchronize(hl_stream_t stream) { + cudaStream_t cu_stream; + + CHECK_LT(stream, HPPL_STREAM_END) << __func__ + << ": the parameter stream is error."; + + cu_stream = t_resource.stream[stream]; + CHECK_CUDA(cudaStreamSynchronize(cu_stream)); +} + +void hl_create_event(hl_event_t *event) { + CHECK_NOTNULL(event); + + struct _hl_event_st *st_event = + (struct _hl_event_st *)malloc(sizeof(struct _hl_event_st)); + + CHECK_CUDA(cudaEventCreate(&st_event->cu_event)); + + *event = st_event; +} + +float hl_event_elapsed_time(hl_event_t start, hl_event_t end) { + float time; + CHECK_NOTNULL(start); + CHECK_NOTNULL(end); + + CHECK_CUDA(cudaEventElapsedTime(&time, start->cu_event, end->cu_event)); + return time; +} + +void hl_stream_record_event(hl_stream_t stream, hl_event_t event) { + cudaStream_t cu_stream; + + CHECK_NOTNULL(event); + CHECK_LT(stream, HPPL_STREAM_END) << __func__ + << ": the parameter stream is error."; + + cu_stream = t_resource.stream[stream]; + CHECK_CUDA(cudaEventRecord(event->cu_event, cu_stream)); +} + +void hl_stream_wait_event(hl_stream_t stream, hl_event_t event) { + cudaStream_t cu_stream; + + CHECK_NOTNULL(event); + CHECK_LT(stream, HPPL_STREAM_END) << __func__ + << ": the parameter stream is error."; + + cu_stream = t_resource.stream[stream]; + CHECK_CUDA(cudaStreamWaitEvent(cu_stream, event->cu_event, 0)); +} + +void hl_destroy_event(hl_event_t event) { + CHECK_NOTNULL(event); + CHECK_CUDA(cudaEventDestroy(event->cu_event)); + + free(event); + event = NULL; +} + +void hl_event_synchronize(hl_event_t event) { + CHECK_NOTNULL(event); + CHECK_CUDA(cudaEventSynchronize(event->cu_event)); +} + +void hl_get_device_name(char *name, int len, int device) { + CHECK_NOTNULL(name); + CHECK(device >= 0 && device < g_system_device_num && g_device[device]) + << "Device(" << device << ") is not specified in startup."; + + strncpy(name, g_device[device]->device_name, len); +} + +void hl_get_device_memory(size_t *mem_size, int device) { + CHECK_NOTNULL(mem_size); + CHECK(device >= 0 && device < g_system_device_num && g_device[device]) + << "Device(" << device << ") is not specified in startup."; + + *mem_size = g_device[device]->device_mem; +} + +void hl_get_device_compute_capability(int *major, int *minor, int device) { + CHECK_NOTNULL(major); + CHECK_NOTNULL(minor); + CHECK(device >= 0 && device < g_system_device_num && g_device[device]) + << "Device(" << device << ") is not specified in startup."; + + *major = g_device[device]->major; + *minor = g_device[device]->minor; +} + +int hl_get_device_last_error() { return (int)cudaGetLastError(); } + +const char *hl_get_device_error_string() { + cudaError_t err = cudaGetLastError(); + return cudaGetErrorString(err); +} + +const char *hl_get_device_error_string(size_t err) { + return cudaGetErrorString((cudaError_t)err); +} + +void hl_device_synchronize() { CHECK_CUDA(cudaDeviceSynchronize()); } +void hl_set_device_flags_block() { + CHECK_CUDA(cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync)); +} + +bool hl_cuda_event_is_ready(hl_event_t event) { + cudaError_t err = cudaEventQuery(event->cu_event); + CHECK(cudaSuccess == err || cudaErrorNotReady == err); + + if (cudaErrorNotReady == err) { + return false; + } + return true; +} + +void hl_profiler_start() { CHECK_CUDA(cudaProfilerStart()); } + +void hl_profiler_end() { CHECK_CUDA(cudaProfilerStop()); } diff --git a/paddle/legacy/cuda/src/hl_cuda_lstm.cu b/paddle/legacy/cuda/src/hl_cuda_lstm.cu new file mode 100644 index 0000000000000000000000000000000000000000..9ac564fd2548cc782bee2380350f4ab888670ca3 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_lstm.cu @@ -0,0 +1,876 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_activation_functions.h" +#include "hl_base.h" +#include "hl_cuda_cublas.h" +#include "hl_device_functions.cuh" +#include "paddle/legacy/utils/Logging.h" + +typedef hppl::Active::forward t_forward; +typedef hppl::Active::backward t_backward; + +bool hl_lstm_sequence_parallel(int frameSize) { + if (frameSize == 32 || frameSize == 64) { + return true; + } else { + return false; + } +} + +class frameValue { + public: + real *value_; + __device__ frameValue(real *value) : value_(value) {} + template + __device__ inline void init(int start, int length, int idx) { + if (reversed == 0) { + value_ += start * frameSize + idx; + } else { + value_ += (start + length - 1) * frameSize + idx; + } + } + __device__ inline real *getPtr() const { return value_; } + __device__ inline real getValue() { return *value_; } + __device__ inline void setValue(real value) { *value_ = value; } + template + __device__ inline void nextFrame() { + if (reversed == 0) { + value_ += frameSize; + } else { + value_ -= frameSize; + } + } +}; + +__device__ __forceinline__ void ptx_sync(const int id, const int barriers) { + asm volatile("bar.sync %0, %1;" : : "r"(id), "r"(barriers) : "memory"); +} + +__device__ __forceinline__ void ptx_arrive(const int id, const int barriers) { + asm volatile("bar.arrive %0, %1;" : : "r"(id), "r"(barriers) : "memory"); +} + +template +__device__ __forceinline__ real forward_sequence(real value, + real *shValue, + real *state, + real *preOutput, + real *output, + real check, + int index, + t_forward activeNode, + t_forward activeGate, + t_forward activeState) { + real out; + real prevOut; + real state_r; + const int idx = index % frameSize; + const int idy = index / frameSize; + // assert(index < valueSize); + + if (idy == 0) { + value = activeNode(value); + shValue[index] = value; + } + if (idy == 1 || idy == 2) { + state_r = state[idx]; + value += state_r * check; + value = activeGate(value); + shValue[index] = value; + } + ptx_sync(1, valueSize); + if (idy == 3) { + state_r = state[idx]; + state_r = state_r * shValue[idx + frameSize * 2]; + state_r += shValue[idx] * shValue[idx + frameSize]; + state[idx] = state_r; + ptx_arrive(2, frameSize * 2); + value += state_r * check; + value = activeGate(value); + shValue[index] = value; + ptx_sync(3, frameSize * 2); + prevOut = preOutput[idx]; + out = prevOut * value; + output[idx] = out; + } + if (idy == 0) { + ptx_sync(2, frameSize * 2); + prevOut = state[idx]; + prevOut = activeState(prevOut); + preOutput[idx] = prevOut; + ptx_arrive(3, frameSize * 2); + } + return value; +} + +#define OUTPUT_BARRIER_ID 10 +#define OUTPUT_BARRIER_ID2 11 +template +__global__ void KeLstmForward(real *gateValue, + real *state, + real *output, + real *preOutput, + real *checkIg, + real *checkFg, + real *checkOg, + real *weight, + const int *starts, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + __shared__ real shValue[valueSize]; + __shared__ real shState[frameSize]; + __shared__ real shPrevOutput[frameSize]; + __shared__ real shOutput[frameSize]; + + const int index = threadIdx.x; + int start = starts[blockIdx.x]; + int length = starts[blockIdx.x + 1] - start; + + /* init */ + real check; + real value; + frameValue frameGate(gateValue); + frameValue frameState(state); + frameValue frameOutput(output); + frameValue framePreOutput(preOutput); + if (index < valueSize) { + const int idx = index % frameSize; + const int idy = index / frameSize; + frameGate.init(start, length, index); + value = frameGate.getValue(); + if (idy == 0) { + shState[idx] = 0.0; + } else if (idy == 1) { + check = checkIg[idx]; + } else if (idy == 2) { + check = checkFg[idx]; + } else if (idy == 3) { + check = checkOg[idx]; + } + + if (idy == 3) { + frameState.init(start, length, idx); + frameOutput.init(start, length, idx); + framePreOutput.init(start, length, idx); + } + + ptx_sync(1, valueSize); + } + + for (int i = 0; i < length; ++i) { + if (index < valueSize) { + if (valueSize == 128) { + if (i != 0) { + ptx_sync(OUTPUT_BARRIER_ID2, blockSize); + value += shValue[index]; + } + } + value = forward_sequence( + value, + shValue, + shState, + shPrevOutput, + shOutput, + check, + index, + hppl::gpu::forward[active_node], + hppl::gpu::forward[active_gate], + hppl::gpu::forward[active_state]); + const int idx = index % frameSize; + const int idy = index / frameSize; + if (valueSize == 128) { + if (idy == 3) { + ptx_arrive(OUTPUT_BARRIER_ID, frameSize + 128); + } + } + if (valueSize == 256) { + ptx_sync(OUTPUT_BARRIER_ID, valueSize); + } + frameGate.setValue(value); + if (idy == 3) { + frameState.setValue(shState[idx]); + frameOutput.setValue(shOutput[idx]); + framePreOutput.setValue(shPrevOutput[idx]); + frameState.nextFrame(); + frameOutput.nextFrame(); + framePreOutput.nextFrame(); + } + if (i != length - 1) { + frameGate.nextFrame(); + value = frameGate.getValue(); + } + } + if (i != length - 1) { + if (valueSize == 128) { + if (valueSize <= index) { + real B_r[frameSize]; + const int computeIdx = index - valueSize; + if (i == 0) { +#pragma unroll + for (int n = 0; n < frameSize; n++) { + B_r[n] = weight[n * valueSize + computeIdx]; + } + } + ptx_sync(OUTPUT_BARRIER_ID, frameSize + 128); + real A_r[frameSize]; + for (int n = 0; n < frameSize; n++) { + A_r[n] = shOutput[n]; + } + real sum = 0.0f; + for (int n = 0; n < frameSize; n++) { + sum += A_r[n] * B_r[n]; + } + shValue[computeIdx] = sum; + ptx_arrive(OUTPUT_BARRIER_ID2, blockSize); + } + } + if (valueSize == 256) { + real B_r[frameSize]; + if (i == 0) { +#pragma unroll + for (int n = 0; n < frameSize; n++) { + B_r[n] = weight[n * valueSize + index]; + } + } + real sum = 0.0f; + for (int n = 0; n < frameSize; n++) { + sum += shOutput[n] * B_r[n]; + } + value += sum; + } + } + } +} + +void hl_lstm_parallel_forward(real *gateValue, + real *stateValue, + real *preOutputValue, + real *outputValue, + real *checkIg, + real *checkFg, + real *checkOg, + real *weight, + const int *sequence, + int frameSize, + int numSequences, + bool reversed, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + CHECK(frameSize == 32 || frameSize == 64); + dim3 grid(numSequences, 1); + if (!reversed) { + if (frameSize == 32) { + KeLstmForward<128, 32, 0, 128, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 64) { + KeLstmForward<256, 64, 0, 256, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); + } + } else { + if (frameSize == 32) { + KeLstmForward<128, 32, 1, 128, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 64) { + KeLstmForward<256, 64, 1, 256, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); + } + } + CHECK_SYNC("hl_lstm_parallel_forward failed"); +} + +__device__ __forceinline__ void transpose_32x32(real a[], const int idx) { + const int warp_size = 32; + int addr = idx % warp_size; + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, addr < warp_size); +#pragma unroll + for (int k = 1; k < 32; k++) { + // rSrc[k] = __shfl_sync(rSrc[k], (threadIdx.x + k) % 32, 32); + addr = __shfl_sync(mask, addr, (idx + 1) % 32, 32); + a[k] = __shfl_sync(mask, a[k], addr, 32); + } + +#pragma unroll + for (int tid = 0; tid < 31; tid++) { + real tmp = (idx > tid) ? a[0] : a[1]; +#pragma unroll + for (int k = 31; k > 0; k--) { + a[(k + 1) % 32] = (idx > tid) ? a[k] : a[(k + 1) % 32]; + } + a[1] = tmp; + } + + addr = (32 - idx) % 32; + CREATE_SHFL_MASK(mask, idx % 32 < warp_size); +#pragma unroll + for (int k = 0; k < 32; k++) { + a[k] = __shfl_sync(mask, a[k], addr, 32); + addr = __shfl_sync(mask, addr, (idx + 31) % 32, 32); + } +} + +template +__device__ void backward_sequence(real rGateValue, + real rOutputGrad, + real rPreOutputValue, + real &rGateGrad, + real &rStateGrad, + real *shStateGrad, + real *shStateValue, + real *shGateValue, + real rCheck, + real &rGateValuePrev, + int index, + t_backward activeNode, + t_backward activeGate, + t_backward activeState) { + const int frameIdx = index % frameSize; + const int frameIdy = index / frameSize; + if (frameIdy == 3) { + real rPrevOutputGrad; + rPrevOutputGrad = rOutputGrad * rGateValue; + rStateGrad = activeState(rPrevOutputGrad, rPreOutputValue); + rGateGrad = rOutputGrad * rPreOutputValue; + rGateGrad = activeGate(rGateGrad, rGateValue); + rStateGrad += rGateGrad * rCheck; + shStateGrad[index] = rStateGrad; + ptx_arrive(3, valueSize); + } else if (frameIdy == 1) { + shGateValue[frameIdx + frameSize] = rGateValue; + rStateGrad = rGateGrad * rCheck; + shStateGrad[index] = rStateGrad; + ptx_sync(3, valueSize); + rStateGrad += shStateGrad[frameIdx + frameSize * 2]; + rStateGrad += shStateGrad[frameIdx + frameSize * 3]; + rGateGrad = rStateGrad * shGateValue[frameIdx]; + rGateGrad = activeGate(rGateGrad, rGateValue); + } else if (frameIdy == 2) { + rStateGrad = rStateGrad * rGateValuePrev; + rStateGrad += rGateGrad * rCheck; + shStateGrad[index] = rStateGrad; + ptx_sync(3, valueSize); + rStateGrad += shStateGrad[frameIdx + frameSize]; + rStateGrad += shStateGrad[frameIdx + frameSize * 3]; + rGateValuePrev = rGateValue; + rGateGrad = rStateGrad * shStateValue[frameIdx]; + rGateGrad = activeGate(rGateGrad, rGateValue); + } else if (frameIdy == 0) { + shGateValue[frameIdx] = rGateValue; + ptx_sync(3, valueSize); + rStateGrad = shStateGrad[frameIdx + frameSize]; + rStateGrad += shStateGrad[frameIdx + frameSize * 2]; + rStateGrad += shStateGrad[frameIdx + frameSize * 3]; + rGateGrad = rStateGrad * shGateValue[frameIdx + frameSize]; + rGateGrad = activeNode(rGateGrad, rGateValue); + } +} + +template +__device__ void load_weight(real rWeight[], real *weight, const int index) { + if (valueSize == 128) { + weight += index; +#pragma unroll + for (int n = 0; n < frameSize; n++) { + rWeight[n] = weight[n * valueSize]; + } + transpose_32x32(rWeight, index % 32); + } + if (valueSize == 256) { + int id = (index / 32) % 2; + weight += index - id * 32 + id * 32 * valueSize; +#pragma unroll + for (int n = 0; n < 32; n++) { + rWeight[n] = weight[n * valueSize]; + rWeight[n + 32] = weight[n * valueSize + 32]; + } + transpose_32x32(rWeight, index % 32); + transpose_32x32(&rWeight[32], index % 32); + } +} + +template +__global__ void KeLstmBackward(real *gateValue, + real *gateGrad, + real *stateValue, + real *stateGrad, /* do not need save */ + real *preOutputValue, + real *preOutputGrad, /* do not need save */ + real *checkIg, + real *checkIgGrad, + real *checkFg, + real *checkFgGrad, + real *checkOg, + real *checkOgGrad, + real *outputGrad, + real *weightValue, + const int *starts, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + __shared__ real shGateValue[valueSize]; + __shared__ real shStateGrad[valueSize]; + __shared__ real shStateValue[frameSize]; + __shared__ real shGateGrad[4][frameSize]; + __shared__ real shOutputGrad[4][frameSize]; + const int index = threadIdx.x; + int start = starts[blockIdx.x]; + int length = starts[blockIdx.x + 1] - start; + + const int frameIdx = index % frameSize; + const int frameIdy = index / frameSize; + real rCheck; + real rCheckGrad; + real rGateGrad; + real rStateGrad; + real rGateValuePrev; + real rPreOutputValue; + real rOutputGrad; + real rGateValue; + real rStateValue; + + frameValue frameGateValue(gateValue); + frameValue frameGateGrad(gateGrad); + frameValue framePreOutputValue(preOutputValue); + frameValue frameStateValue(stateValue); + frameValue frameOutputGrad(outputGrad); + if (frameIdy == 0) { + } else if (frameIdy == 1) { + rCheck = checkIg[frameIdx]; + } else if (frameIdy == 2) { + rCheck = checkFg[frameIdx]; + rGateValuePrev = 0.0; + rStateGrad = 0.0; + } else if (frameIdy == 3) { + rCheck = checkOg[frameIdx]; + framePreOutputValue.init(start, length, frameIdx); + frameOutputGrad.init(start, length, frameIdx); + rOutputGrad = frameOutputGrad.getValue(); + rPreOutputValue = framePreOutputValue.getValue(); + frameStateValue.init(start, length, frameIdx); + rStateValue = frameStateValue.getValue(); + } + + frameGateValue.init(start, length, index); + frameGateGrad.init(start, length, index); + rGateValue = frameGateValue.getValue(); + rGateGrad = 0.0; + rCheckGrad = 0.0; + + real B_r[frameSize]; + load_weight(B_r, weightValue, index); + + for (int i = 0; i < length; ++i) { + if (frameIdy == 3) { + if (i != length - 1) { + frameStateValue.nextFrame(); + shStateValue[frameIdx] = frameStateValue.getValue(); + } else { + shStateValue[frameIdx] = 0.0; + } + } + backward_sequence(rGateValue, + rOutputGrad, + rPreOutputValue, + rGateGrad, + rStateGrad, + shStateGrad, + shStateValue, + shGateValue, + rCheck, + rGateValuePrev, + index, + hppl::gpu::backward[active_node], + hppl::gpu::backward[active_gate], + hppl::gpu::backward[active_state]); + if (frameIdy == 3) { + rCheckGrad += rGateGrad * rStateValue; + rStateValue = shStateValue[frameIdx]; + } + + frameGateGrad.setValue(rGateGrad); + frameGateGrad.nextFrame(); + + if (i != length - 1) { + if (frameIdy == 3) { + framePreOutputValue.nextFrame(); + rPreOutputValue = framePreOutputValue.getValue(); + frameOutputGrad.nextFrame(); + rOutputGrad = frameOutputGrad.getValue(); + } else if (frameIdy == 2) { + rCheckGrad += rGateGrad * shStateValue[frameIdx]; + } else if (frameIdy == 1) { + rCheckGrad += rGateGrad * shStateValue[frameIdx]; + } + + frameGateValue.nextFrame(); + rGateValue = frameGateValue.getValue(); + shGateGrad[frameIdy][frameIdx] = rGateGrad; + if (valueSize == 128) { + real sum = 0.0f; +#pragma unroll + for (int n = 0; n < frameSize; n++) { + sum += shGateGrad[frameIdy][n] * B_r[n]; + } + if (frameIdy == 3) { + rOutputGrad += sum; + } else { + shOutputGrad[frameIdy][frameIdx] = sum; + } + } + if (valueSize == 256) { + ptx_sync(5, valueSize); + real A_r[frameSize]; + for (int n = 0; n < frameSize; n++) { + A_r[n] = shGateGrad[frameIdy][n]; + } + real sum = 0.0f; + for (int n = 0; n < frameSize; n++) { + sum += A_r[n] * B_r[n]; + } + if (frameIdy == 3) { + rOutputGrad += sum; + } else { + shOutputGrad[frameIdy][frameIdx] = sum; + } + } + + if (frameIdy == 3) { + ptx_sync(6, valueSize); +#pragma unroll + for (int i = 0; i < 3; i++) { + rOutputGrad += shOutputGrad[i][frameIdx]; + } + } else { + ptx_arrive(6, valueSize); + } + } + } + + /* TODO: Temporary save & merger in another kernel */ + if (frameIdy == 1) { + if (checkIgGrad) + paddle::paddleAtomicAdd(checkIgGrad + frameIdx, rCheckGrad); + } else if (frameIdy == 2) { + if (checkFgGrad) + paddle::paddleAtomicAdd(checkFgGrad + frameIdx, rCheckGrad); + } else if (frameIdy == 3) { + if (checkOgGrad) + paddle::paddleAtomicAdd(checkOgGrad + frameIdx, rCheckGrad); + } +} + +void hl_lstm_parallel_backward_data(real *gateValue, + real *gateGrad, + real *stateValue, + real *stateGrad, + real *preOutputValue, + real *preOutputGrad, + real *outputGrad, + real *checkIg, + real *checkIgGrad, + real *checkFg, + real *checkFgGrad, + real *checkOg, + real *checkOgGrad, + real *weight, + const int *sequence, + int frameSize, + int numSequences, + bool reversed, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate, + hl_activation_mode_t active_state) { + CHECK(frameSize == 32 || frameSize == 64 || frameSize == 128 || + frameSize == 256); + dim3 grid(numSequences, 1); + if (!reversed) { + if (frameSize == 32) { + KeLstmBackward<128, 32, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 64) { + KeLstmBackward<256, 64, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 128) { + KeLstmBackward<512, 128, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 256) { + KeLstmBackward<1024, 256, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } + } else { + if (frameSize == 32) { + KeLstmBackward<128, 32, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 64) { + KeLstmBackward<256, 64, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 128) { + KeLstmBackward<512, 128, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } else if (frameSize == 256) { + KeLstmBackward<1024, 256, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); + } + } + CHECK_SYNC("hl_lstm_parallel_backward_data"); +} + +template +__global__ void KeSetGradZero(real *gateGrad, + const int *starts, + int valueSize, + int numSequences, + bool reversed) { + // const int tid = threadIdx.x; + + const int frameIdx = blockIdx.x * B_X + threadIdx.x; + const int numSeqId = blockIdx.y * B_Y + threadIdx.y; + + if (numSeqId >= numSequences || frameIdx >= valueSize) return; + + if (!reversed) { + int seqId = starts[numSeqId]; + gateGrad[seqId * valueSize + frameIdx] = 0.0; + } else { + int seqId = starts[numSeqId + 1] - 1; + gateGrad[seqId * valueSize + frameIdx] = 0.0; + } +} + +void hl_lstm_parallel_backward_weight(real *weightGrad, + real *outputValue, + real *gateGrad, + const int *sequence, + int frameSize, + int batchSize, + int numSequences, + bool reversed) { + int valueSize = 4 * frameSize; + dim3 threads(32, 32); + dim3 grid((valueSize + 32 - 1) / 32, (numSequences + 32 - 1) / 32); + KeSetGradZero<32, 32><<>>( + gateGrad, sequence, valueSize, numSequences, reversed); + + if (!reversed) { + hl_matrix_mul(outputValue, + HPPL_OP_T, + gateGrad + valueSize, + HPPL_OP_N, + weightGrad, + frameSize, + valueSize, + batchSize - 1, + 1.0, + 1.0); + } else { + hl_matrix_mul(outputValue + frameSize, + HPPL_OP_T, + gateGrad, + HPPL_OP_N, + weightGrad, + frameSize, + valueSize, + batchSize - 1, + 1.0, + 1.0); + } + CHECK_SYNC("hl_lstm_parallel_backward_weight"); +} diff --git a/paddle/legacy/cuda/src/hl_cuda_matrix.cu b/paddle/legacy/cuda/src/hl_cuda_matrix.cu new file mode 100644 index 0000000000000000000000000000000000000000..6fe460026bbd404e15b43bd221551094a7abeda2 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_matrix.cu @@ -0,0 +1,806 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_base.h" +#include "hl_device_functions.cuh" +#include "hl_gpu_matrix_kernel.cuh" +#include "hl_matrix.h" +#include "hl_matrix_apply.cuh" +#include "hl_matrix_ops.cuh" +#include "hl_sequence.h" +#include "hl_sparse.ph" +#include "paddle/legacy/utils/Logging.h" + +DEFINE_MATRIX_UNARY_OP(Zero, a = 0); +DEFINE_MATRIX_TERNARY_PARAMETER_OP(_add, TWO_PARAMETER, c = p1 * a + p2 * b); +void hl_matrix_add(real* A_d, + real* B_d, + real* C_d, + int dimM, + int dimN, + real alpha, + real beta) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + + hl_gpu_apply_ternary_op, 0, 0>( + ternary::_add(alpha, beta), + A_d, + B_d, + C_d, + dimM, + dimN, + dimN, + dimN, + dimN); + CHECK_SYNC("hl_matrix_add failed"); +} + +#ifdef PADDLE_TYPE_DOUBLE +#define THRESHOLD 128 +#else +#define THRESHOLD 64 +#endif +__device__ __forceinline__ void findMax(real* I, + real* dfMax_s, + int blockSize, + int base, + int curIdx, + int nextIdx, + int dimN, + real* max) { + dfMax_s[base] = -1.0e20; + while (curIdx < dimN) { + if (dfMax_s[base] < I[nextIdx]) { + dfMax_s[base] = I[nextIdx]; + } + nextIdx += blockSize; + curIdx += blockSize; + } + __syncthreads(); + + for (int stride = blockSize >> 1; stride > 0; stride >>= 1) { + __syncthreads(); + if (base < stride) { + nextIdx = base + stride; + if (dfMax_s[base] < dfMax_s[nextIdx]) { + dfMax_s[base] = dfMax_s[nextIdx]; + } + } + } + + if (0 == base) { + max[0] = dfMax_s[0]; + } + __syncthreads(); +} + +__device__ __forceinline__ void subMaxAndExp(real* I, + real* O, + int curIdx, + int nextIdx, + int blockSize, + int dimN, + real max) { + real val; + while (curIdx < dimN) { + val = I[nextIdx] - max; + if (val < -THRESHOLD) { + val = -THRESHOLD; + } + I[nextIdx] = val; +#ifndef PADDLE_TYPE_DOUBLE + O[nextIdx] = __expf(val); +#else + O[nextIdx] = exp(val); +#endif + nextIdx += blockSize; + curIdx += blockSize; + } + __syncthreads(); +} + +__device__ __forceinline__ void valueSum(real* O, + real* dfMax_s, + int blockSize, + int base, + int curIdx, + int nextIdx, + int dimN) { + dfMax_s[base] = 0; + while (curIdx < dimN) { + dfMax_s[base] += O[nextIdx]; + nextIdx += blockSize; + curIdx += blockSize; + } + __syncthreads(); + + for (int stride = blockSize >> 1; stride > 0; stride >>= 1) { + __syncthreads(); + if (base < stride) { + nextIdx = base + stride; + dfMax_s[base] += dfMax_s[nextIdx]; + } + } + __syncthreads(); +} + +__device__ __forceinline__ void divSum( + real* O, real sum, int curIdx, int nextIdx, int blockSize, int dimN) { + while (curIdx < dimN) { + O[nextIdx] /= sum; + nextIdx += blockSize; + curIdx += blockSize; + } +} + +__device__ __forceinline__ void softmax(real* I, + real* O, + real* dfMax_s, + int blockSize, + int base, + int curIdx, + int nextIdx, + int dimN) { + __shared__ real max; + + // find the max number + findMax(I, dfMax_s, blockSize, base, curIdx, nextIdx, dimN, &max); + + // sub max Value and do Exp operation + subMaxAndExp(I, O, base, nextIdx, blockSize, dimN, max); + + // add dimN values into blockDim.x buffer + // sum is in dfMax_s[0] + valueSum(O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); + + // divided by sum + divSum(O, dfMax_s[0], curIdx, nextIdx, blockSize, dimN); +} + +template +__global__ void KeMatrixSoftMax(real* O, real* I, int dimN) { + int base = threadIdx.x; + __shared__ real dfMax_s[blockSize]; + int nextIdx = blockIdx.x * dimN + base; + int curIdx = base; + + softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); +} + +void hl_matrix_softmax(real* A_d, real* C_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + dim3 block(512, 1); + dim3 grid(dimM, 1); + KeMatrixSoftMax<512><<>>(C_d, A_d, dimN); + CHECK_SYNC("hl_matrix_softmax failed"); +} + +template +__global__ void KeSequenceSoftMax(real* O, real* I, const int* index) { + int base = threadIdx.x; + int bid = blockIdx.x; + __shared__ real dfMax_s[blockSize]; + + int start = index[bid]; + int dimN = index[bid + 1] - start; + + int nextIdx = start + base; + int curIdx = base; + + softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); +} + +void hl_sequence_softmax_forward(real* A_d, + real* C_d, + const int* index, + int numSequence) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + dim3 block(512, 1); + dim3 grid(numSequence, 1); + KeSequenceSoftMax<512><<>>(C_d, A_d, index); + CHECK_SYNC("hl_sequence_softmax_forward failed"); +} + +__global__ void KeMatrixDerivative( + real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + int colIdx = blockIdx.y * blockDim.y + threadIdx.y; + int index; + + if (rowIdx < dimM && colIdx < dimN) { + index = rowIdx * dimN + colIdx; + grad_d[index] = output_d[index] * (grad_d[index] - sftmaxSum_d[rowIdx]); + } +} + +void hl_matrix_softmax_derivative( + real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) { + CHECK_NOTNULL(grad_d); + CHECK_NOTNULL(output_d); + CHECK_NOTNULL(sftmaxSum_d); + + int blocksX = (dimM + 0) / 1; + int blocksY = (dimN + 1024 - 1) / 1024; + dim3 threads(1, 1024); + dim3 grid(blocksX, blocksY); + + KeMatrixDerivative<<>>( + grad_d, output_d, sftmaxSum_d, dimM, dimN); + CHECK_SYNC("hl_matrix_softmax_derivative failed"); +} + +__global__ void KeMatrixMultiBinaryCrossEntropy( + real* output, real* entropy, int* row, int* col, int dimM, int dimN) { + int index = blockIdx.x * blockDim.x + threadIdx.x; + if (index < dimM) { + for (int i = 0; i < dimN; i++) { + entropy[index] -= log(1 - output[index * dimN + i]); + } + int* row_col = col + row[index]; + int col_num = row[index + 1] - row[index]; + for (int i = 0; i < col_num; i++) { + real o = output[index * dimN + row_col[i]]; + entropy[index] -= log(o / (1 - o)); + } + } +} + +void hl_matrix_multi_binary_cross_entropy(real* output, + real* entropy, + hl_sparse_matrix_s csr_mat, + int dimM, + int dimN) { + CHECK_NOTNULL(output); + CHECK_NOTNULL(entropy); + CHECK_NOTNULL(csr_mat); + CHECK_EQ(csr_mat->format, HL_SPARSE_CSR); + int n_threads = 1024; + int blocks = (dimM + n_threads - 1) / n_threads; + dim3 threads(n_threads); + dim3 grid(blocks); + hl_csr_matrix mat = (hl_csr_matrix)(csr_mat->matrix); + KeMatrixMultiBinaryCrossEntropy<<>>( + output, entropy, mat->csr_row, mat->csr_col, dimM, dimN); + CHECK_SYNC("hl_matrix_multi_binary_cross_entropy failed"); +} + +__global__ void KeMatrixMultiBinaryCrossEntropyBp( + real* output, real* grad, int* row, int* col, int dimM, int dimN) { + int row_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (row_idx < dimM) { + for (int i = 0; i < dimN; i++) { + int index = row_idx * dimN + i; + grad[index] += 1.0 / (1 - output[index]); + } + int col_num = row[row_idx + 1] - row[row_idx]; + int* row_col = col + row[row_idx]; + for (int i = 0; i < col_num; i++) { + int index = row_idx * dimN + row_col[i]; + grad[index] -= 1.0 / (output[index] * (1 - output[index])); + } + } +} + +void hl_matrix_multi_binary_cross_entropy_bp( + real* output, real* grad, hl_sparse_matrix_s csr_mat, int dimM, int dimN) { + CHECK_NOTNULL(output); + CHECK_NOTNULL(grad); + CHECK_NOTNULL(csr_mat); + CHECK_EQ(csr_mat->format, HL_SPARSE_CSR); + int n_threads = 1024; + int blocks = (dimM + n_threads - 1) / n_threads; + dim3 threads(n_threads); + dim3 grid(blocks); + hl_csr_matrix mat = (hl_csr_matrix)(csr_mat->matrix); + KeMatrixMultiBinaryCrossEntropyBp<<>>( + output, grad, mat->csr_row, mat->csr_col, dimM, dimN); + CHECK_SYNC("hl_matrix_multi_binary_cross_entropy_bp failed"); +} + +__global__ void KeMatrixCrossEntropy( + real* O, real* E, int* label, int dimM, int dimN) { + int index = blockIdx.x * blockDim.x + threadIdx.x; + int newBase; + if (index < dimM) { + newBase = label[index]; + newBase = newBase % dimN; + E[index] = -log(O[index * dimN + newBase]); + } +} + +void hl_matrix_cross_entropy( + real* A_d, real* C_d, int* label_d, int dimM, int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + + int blocks = (dimM + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + KeMatrixCrossEntropy<<>>( + A_d, C_d, label_d, dimM, dimN); + CHECK_SYNC("hl_matrix_cross_entropy failed"); +} + +__global__ void KeMatrixCrossEntropyBp( + real* grad_d, real* output_d, int* label_d, int dimM, int dimN) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + int colIdx = blockIdx.y * blockDim.y + threadIdx.y; + int index; + if (rowIdx < dimM && colIdx < dimN) { + index = rowIdx * dimN + colIdx; + if (label_d[rowIdx] == colIdx) { + grad_d[index] -= 1.0f / output_d[index]; + } + } +} + +void hl_matrix_cross_entropy_bp( + real* grad_d, real* output_d, int* label_d, int dimM, int dimN) { + CHECK_NOTNULL(grad_d); + CHECK_NOTNULL(output_d); + CHECK_NOTNULL(label_d); + + int blocksX = (dimM + 0) / 1; + int blocksY = (dimN + 1024 - 1) / 1024; + dim3 threads(1, 1024); + dim3 grid(blocksX, blocksY); + KeMatrixCrossEntropyBp<<>>( + grad_d, output_d, label_d, dimM, dimN); + CHECK_SYNC("hl_matrix_cross_entropy_bp failed"); +} + +void hl_matrix_zero_mem(real* data, int num) { + hl_gpu_apply_unary_op(unary::Zero(), data, 1, num, num); +} + +__global__ void KeParamReluForward(real* output, + real* input, + real* w, + int width, + int height, + int partial_sum) { + int tx = blockIdx.x * blockDim.x + threadIdx.x; + int ty = blockIdx.y * blockDim.y + threadIdx.y; + if (tx < width && ty < height) { + int index = ty * width + tx; + output[index] = + input[index] > 0 ? input[index] : input[index] * w[tx / partial_sum]; + } +} + +void hl_param_relu_forward(real* output, + real* input, + real* w, + int width, + int height, + int partial_sum) { + CHECK_NOTNULL(output); + CHECK_NOTNULL(input); + CHECK_NOTNULL(w); + dim3 threads(16, 16); + int blockX = (width + 16 - 1) / 16; + int blockY = (height + 16 - 1) / 16; + dim3 grid(blockX, blockY); + KeParamReluForward<<>>( + output, input, w, width, height, partial_sum); + CHECK_SYNC("hl_param_relu_forward failed"); +} + +template +__global__ void KeParamReluBackWardW(real* grad_w, + real* grad_o, + real* input, + int width, + int height, + int partial_sum) { + const int tid = threadIdx.x; + __shared__ real temp[blockSize]; + grad_o += partial_sum * blockIdx.x; + input += partial_sum * blockIdx.x; + real tmp = 0.0; + for (int index = tid; index < partial_sum * height; index += blockSize) { + int row = index / partial_sum; + int offset = row * width + (index - row * partial_sum); + if (input[offset] < 0) { + tmp += grad_o[offset] * input[offset]; + } + } + temp[tid] = tmp; + __syncthreads(); + for (int s = blockSize / 2; s > 0; s >>= 1) { + if (tid < s) { + temp[tid] += temp[tid + s]; + } + __syncthreads(); + } + if (tid == 0) { + grad_w[blockIdx.x] += temp[0]; + } +} + +void hl_param_relu_backward_w(real* grad_w, + real* grad_o, + real* input, + int width, + int height, + int partial_sum) { + CHECK_NOTNULL(grad_w); + CHECK_NOTNULL(grad_o); + CHECK_NOTNULL(input); + const int blockSize = 1024; + int grid_num = width / partial_sum; + dim3 threads(blockSize, 1); + dim3 grid(grid_num, 1); + KeParamReluBackWardW<<>>( + grad_w, grad_o, input, width, height, partial_sum); + CHECK_SYNC("hl_param_relu_backward_w failed"); +} + +__global__ void KeParamReluBackwardDiff(real* grad_o, + real* input, + real* w, + real* diff, + int width, + int height, + int partial_sum) { + int tx = blockIdx.x * blockDim.x + threadIdx.x; + int ty = blockIdx.y * blockDim.y + threadIdx.y; + if (tx < width && ty < height) { + int index = ty * width + tx; + diff[index] += grad_o[index] * (input[index] > 0 ? 1 : w[tx / partial_sum]); + } +} + +void hl_param_relu_backward_diff(real* grad_o, + real* data, + real* w, + real* diff, + int width, + int height, + int partial_sum) { + CHECK_NOTNULL(grad_o); + CHECK_NOTNULL(data); + CHECK_NOTNULL(w); + CHECK_NOTNULL(diff); + dim3 threads(16, 16); + int blockX = (width + 16 - 1) / 16; + int blockY = (height + 16 - 1) / 16; + dim3 grid(blockX, blockY); + KeParamReluBackwardDiff<<>>( + grad_o, data, w, diff, width, height, partial_sum); + CHECK_SYNC("hl_param_relu_backward_diff failed"); +} + +__global__ void KeMatrixAddSharedBias( + real* A, real* B, const int channel, const int M, const int N, real scale) { + int index = blockIdx.x * blockDim.x + threadIdx.x; + int dim = N / channel; + if (index < M * N) { + int i = index % N; + i = i / dim; + A[index] += scale * B[i]; + } +} + +void hl_matrix_add_shared_bias(real* A_d, + real* B_d, + const int channel, + const int dimM, + const int dimN, + real scale) { + const int blocks = 512; + const int grids = DIVUP(dimM * dimN, blocks); + KeMatrixAddSharedBias<<>>( + A_d, B_d, channel, dimM, dimN, scale); + CHECK_SYNC("hl_matrix_add_shared_bias failed"); +} + +template +__global__ void KeMatrixCollectSharedBias(real* B, + real* A, + const int channel, + const int M, + const int N, + const int dim, + const int limit, + real scale) { + if (dim < limit) { + int index = blockIdx.x * blockDim.x + threadIdx.x; + if (index < channel) { + real sum = 0.0; + for (int i = 0; i < M; ++i) { + for (int j = 0; j < dim; ++j) { + sum += A[i * N + index * dim + j]; + } + } + B[index] += scale * sum; + } + } else { + const int tid = threadIdx.x; + const int bid = blockIdx.x; + __shared__ real smem[blockSize]; + real sum = 0.0; + for (int j = 0; j < ((dim * M + blockSize - 1) / blockSize); ++j) { + int n = j * blockSize + tid; + int m = n / dim; + int w = n % dim; + smem[tid] = (m < M && w < dim) ? A[m * N + bid * dim + w] : 0.0; + __syncthreads(); + simpleReduce(smem, tid, blockSize); + sum += smem[0]; + } + if (tid == 0) { + B[bid] += scale * sum; + } + } +} + +void hl_matrix_collect_shared_bias(real* B_d, + real* A_d, + const int channel, + const int dimM, + const int dimN, + real scale) { + const int dim = dimN / channel; + const int blocks = 256; + const int limit = 64; + int grids = (dimM * dim) < limit ? DIVUP(channel, blocks) : channel; + + KeMatrixCollectSharedBias<<>>( + B_d, A_d, channel, dimM, dimN, dim, limit, scale); + CHECK_SYNC("hl_matrix_collect_shared_bias failed"); +} + +__global__ void keMatrixRotate( + real* mat, real* matRot, int dimM, int dimN, bool clockWise) { + int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < dimM * dimN) { + int i = idx / dimN; + int j = idx % dimN; + if (clockWise) { + matRot[j * dimM + i] = mat[(dimM - i - 1) * dimN + j]; + } else { + matRot[j * dimM + i] = mat[i * dimN + (dimN - j - 1)]; + } + } +} + +void hl_matrix_rotate( + real* mat, real* matRot, int dimM, int dimN, bool clockWise) { + CHECK_NOTNULL(mat); + CHECK_NOTNULL(matRot); + const int threads = 512; + const int blocks = DIVUP(dimM * dimN, threads); + keMatrixRotate<<>>( + mat, matRot, dimM, dimN, clockWise); + CHECK_SYNC("hl_matrix_rotate failed"); +} + +__global__ void keMatrixVol2Col(int num_kernels, + const real* dataSrc, + real* dataDst, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + int depth_col, + int height_col, + int width_col) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; + index += blockDim.x * gridDim.x) { + int w_out = index % width_col; + int h_out = (index / width_col) % height_col; + int d_out = (index / width_col / height_col) % depth_col; + int channel_in = index / width_col / height_col / depth_col; + int channel_out = channel_in * filterD * filterH * filterW; + int w_in = w_out * strideW - paddingW; + int h_in = h_out * strideH - paddingH; + int d_in = d_out * strideD - paddingD; + + dataDst += + ((channel_out * depth_col + d_out) * height_col + h_out) * width_col + + w_out; + dataSrc += ((channel_in * depth + d_in) * height + h_in) * width + w_in; + for (int k = 0; k < filterD; ++k) { + for (int i = 0; i < filterH; ++i) { + for (int j = 0; j < filterW; ++j) { + int d = d_in + k; + int h = h_in + i; + int w = w_in + j; + *dataDst = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 && + w < width) + ? dataSrc[(k * height + i) * width + j] + : 0; + dataDst += depth_col * height_col * width_col; + } + } + } + } +} + +void hl_matrix_vol2Col(const real* dataSrc, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real* dataDst) { + int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1; + int height_col = (height + 2 * paddingH - filterH) / strideH + 1; + int width_col = (width + 2 * paddingW - filterW) / strideW + 1; + int num_kernels = channels * depth_col * height_col * width_col; + + const int threads = 512; + const int blocks = DIVUP(num_kernels, threads); + + keMatrixVol2Col<<>>(num_kernels, + dataSrc, + dataDst, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + depth_col, + height_col, + width_col); + CHECK_SYNC("hl_matrix_vol2Col failed"); +} + +__global__ void keMatrixCol2Vol(int num_kernels, + real* dataDst, + const real* dataSrc, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + int depth_col, + int height_col, + int width_col, + real alpha, + real beta) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; + index += blockDim.x * gridDim.x) { + real srcVal = 0; + real dstVal = dataDst[index]; + int w = index % width + paddingW; + int h = (index / width) % height + paddingH; + int d = (index / width / height) % depth + paddingD; + int c = index / width / height / depth; + // compute the start and end of the output + int w_col_start = (w < filterW) ? 0 : (w - filterW) / strideW + 1; + int w_col_end = min(w / strideW + 1, width_col); + int h_col_start = (h < filterH) ? 0 : (h - filterH) / strideH + 1; + int h_col_end = min(h / strideH + 1, height_col); + int d_col_start = (d < filterD) ? 0 : (d - filterD) / strideD + 1; + int d_col_end = min(d / strideD + 1, depth_col); + + int offset = (c * filterD * filterW * filterH + d * filterW * filterH + + h * filterW + w) * + depth_col * height_col * width_col; + + int coeff_d_col = + (1 - strideD * filterW * filterH * depth_col) * height_col * width_col; + int coeff_h_col = + (1 - strideH * filterW * depth_col * height_col) * width_col; + int coeff_w_col = (1 - strideW * depth_col * height_col * width_col); + + for (int d_col = d_col_start; d_col < d_col_end; ++d_col) { + for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { + for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { + srcVal += dataSrc[offset + d_col * coeff_d_col + h_col * coeff_h_col + + w_col * coeff_w_col]; + } + } + } + dataDst[index] = alpha * srcVal + beta * dstVal; + } +} + +void hl_matrix_col2Vol(real* dataDst, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + const real* dataSrc, + real alpha, + real beta) { + int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1; + int height_col = (height + 2 * paddingH - filterH) / strideH + 1; + int width_col = (width + 2 * paddingW - filterW) / strideW + 1; + int num_kernels = channels * depth * height * width; + + const int threads = 512; + const int blocks = DIVUP(num_kernels, threads); + + keMatrixCol2Vol<<>>(num_kernels, + dataDst, + dataSrc, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + depth_col, + height_col, + width_col, + alpha, + beta); + + CHECK_SYNC("hl_matrix_col2Vol failed"); +} + +__global__ void keVectorCast2Int(int* out, real* vec, int size) { + for (int i = threadIdx.x; i < (size); i += blockDim.x) { + out[i] = int(vec[i]); + } +} + +void hl_vector_cast2int(int* out, real* vec, int size) { + keVectorCast2Int<<<1, 512, 0, STREAM_DEFAULT>>>(out, vec, size); + CHECK_SYNC("hl_vector_cast2int failed"); +} diff --git a/paddle/legacy/cuda/src/hl_cuda_sequence.cu b/paddle/legacy/cuda/src/hl_cuda_sequence.cu new file mode 100644 index 0000000000000000000000000000000000000000..1d772b5ce27615673d85231ec8fd3ab1d0aed523 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_sequence.cu @@ -0,0 +1,408 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_base.h" +#include "hl_device_functions.cuh" +#include "paddle/legacy/utils/Logging.h" + +__global__ void KeMaxSequenceForward(real* input, + const int* sequence, + real* output, + int* index, + int numSequences, + int dim) { + int dimIdx = threadIdx.x; + int sequenceId = blockIdx.x; + if (sequenceId >= numSequences) return; + int start = sequence[sequenceId]; + int end = sequence[sequenceId + 1]; + + for (int i = dimIdx; i < dim; i += blockDim.x) { + real tmp = -HL_FLOAT_MAX; + int tmpId = -1; + for (int insId = start; insId < end; insId++) { + if (tmp < input[insId * dim + i]) { + tmp = input[insId * dim + i]; + tmpId = insId; + } + } + output[sequenceId * dim + i] = tmp; + index[sequenceId * dim + i] = tmpId; + } +} + +void hl_max_sequence_forward(real* input, + const int* sequence, + real* output, + int* index, + int numSequences, + int dim) { + CHECK_NOTNULL(input); + CHECK_NOTNULL(sequence); + CHECK_NOTNULL(output); + CHECK_NOTNULL(index); + + dim3 threads(256, 1); + dim3 grid(numSequences, 1); + KeMaxSequenceForward<<>>( + input, sequence, output, index, numSequences, dim); + CHECK_SYNC("hl_max_sequence_forward failed"); +} + +__global__ void KeMaxSequenceBackward( + real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) { + int idx = threadIdx.x + blockIdx.x * blockDim.x; + int colIdx = idx % dim; + if (idx < numSequences * dim) { + int insId = index[idx]; + inputGrad[insId * dim + colIdx] += outputGrad[idx]; + } +} + +void hl_max_sequence_backward( + real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) { + CHECK_NOTNULL(outputGrad); + CHECK_NOTNULL(index); + CHECK_NOTNULL(inputGrad); + + unsigned int blocks = (numSequences * dim + 128 - 1) / 128; + dim3 threads(128, 1); + dim3 grid(blocks, 1); + KeMaxSequenceBackward<<>>( + outputGrad, index, inputGrad, numSequences, dim); + CHECK_SYNC("hl_max_sequence_backward failed"); +} + +template +__global__ void KeMatrixAddRows(real* output, + real* table, + int* ids, + int numSamples, + int tableSize, + int dim) { + int idx = threadIdx.x; + int idy = threadIdx.y; + int sampleId = blockIdx.x + idy * gridDimX; + + while (sampleId < numSamples) { + int tableId = ids[sampleId]; + if ((0 <= tableId) && (tableId < tableSize)) { + real* outputData = output + sampleId * dim; + real* tableData = table + tableId * dim; + for (int i = idx; i < dim; i += blockDimX) { + if (AddRow == 0) { + outputData[i] += tableData[i]; + } else { + paddle::paddleAtomicAdd(&tableData[i], outputData[i]); + } + } + } + sampleId += blockDimY * gridDimX; + } +} + +template +__global__ void KeSequence2Batch(real* batch, + real* sequence, + const int* batchIndex, + int seqWidth, + int batchCount) { + int idx = threadIdx.x; + int idy = threadIdx.y; + int id = blockIdx.x + idy * gridDimX; + while (id < batchCount) { + int seqId = batchIndex[id]; + real* batchData = batch + id * seqWidth; + real* seqData = sequence + seqId * seqWidth; + for (int i = idx; i < seqWidth; i += blockDimX) { + if (seq2batch) { + if (isAdd) { + batchData[i] += seqData[i]; + } else { + batchData[i] = seqData[i]; + } + } else { + if (isAdd) { + seqData[i] += batchData[i]; + } else { + seqData[i] = batchData[i]; + } + } + } + id += blockDimY * gridDimX; + } +} + +void hl_sequence2batch_copy(real* batch, + real* sequence, + const int* batchIndex, + int seqWidth, + int batchCount, + bool seq2batch) { + CHECK_NOTNULL(sequence); + CHECK_NOTNULL(batch); + CHECK_NOTNULL(batchIndex); + + dim3 threads(128, 8); + dim3 grid(8, 1); + if (seq2batch) { + KeSequence2Batch<128, 8, 8, 1, 0><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); + } else { + KeSequence2Batch<128, 8, 8, 0, 0><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); + } + CHECK_SYNC("hl_sequence2batch_copy failed"); +} + +void hl_sequence2batch_add(real* batch, + real* sequence, + int* batchIndex, + int seqWidth, + int batchCount, + bool seq2batch) { + CHECK_NOTNULL(sequence); + CHECK_NOTNULL(batch); + CHECK_NOTNULL(batchIndex); + + dim3 threads(128, 8); + dim3 grid(8, 1); + if (seq2batch) { + KeSequence2Batch<128, 8, 8, 1, 1><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); + } else { + KeSequence2Batch<128, 8, 8, 0, 1><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); + } + CHECK_SYNC("hl_sequence2batch_add failed"); +} + +template +__global__ void KeSequence2BatchPadding(real* batch, + real* sequence, + const int* sequenceStartPositions, + const size_t sequenceWidth, + const size_t maxSequenceLength, + const size_t numSequences) { + int batchIdx = blockIdx.y; + int sequenceStart = sequenceStartPositions[batchIdx]; + int sequenceLength = sequenceStartPositions[batchIdx + 1] - sequenceStart; + + int sequenceIdx = blockIdx.x * blockDim.y + threadIdx.y; + int batchBaseIdx = (sequenceIdx * numSequences + batchIdx) * sequenceWidth; + int sequenceBaseIdx = (sequenceStart + sequenceIdx) * sequenceWidth; + + real scale = normByTimes ? (1.0f / (real)sequenceLength) : 1.0f; + + if (sequenceIdx < sequenceLength) { + if (seq2batch) { + /* sequence -> batch */ + for (int i = threadIdx.x; i < sequenceWidth; i += blockDim.x) { + batch[batchBaseIdx + i] = scale * sequence[sequenceBaseIdx + i]; + } + } else { + /* batch -> sequence */ + for (int i = threadIdx.x; i < sequenceWidth; i += blockDim.x) { + sequence[sequenceBaseIdx + i] = scale * batch[batchBaseIdx + i]; + } + } + } else if (sequenceIdx < maxSequenceLength) { + if (seq2batch) { + /* sequence -> batch */ + for (int i = threadIdx.x; i < sequenceWidth; i += blockDim.x) { + batch[batchBaseIdx + i] = 0; + } + } + } +} + +void hl_sequence2batch_copy_padding(real* batch, + real* sequence, + const int* sequenceStartPositions, + const size_t sequenceWidth, + const size_t maxSequenceLength, + const size_t numSequences, + bool normByTimes, + bool seq2batch) { + CHECK_NOTNULL(batch); + CHECK_NOTNULL(sequence); + CHECK_NOTNULL(sequenceStartPositions); + + if (!normByTimes && numSequences == 1) { + size_t elementCount = maxSequenceLength * sequenceWidth; + if (seq2batch) { + /* sequence -> batch */ + hl_memcpy_device2device(batch, sequence, sizeof(real) * elementCount); + } else { + /* batch -> sequence */ + hl_memcpy_device2device(sequence, batch, sizeof(real) * elementCount); + } + return; + } + + const int CUDA_BLOCK_SIZE = 512; + + /* At least use 32 threads to copy sequenceWidth elements, + and at least 8 elements for each thread. */ + int blockDimX = ((((sequenceWidth + 7) >> 3) + 31) >> 5) << 5; + blockDimX = (blockDimX < CUDA_BLOCK_SIZE) ? blockDimX : CUDA_BLOCK_SIZE; + + int blockDimY = CUDA_BLOCK_SIZE / blockDimX; + dim3 threads(blockDimX, blockDimY); + + int gridDimX = (maxSequenceLength + blockDimY - 1) / blockDimY; + int gridDimY = numSequences; + dim3 grid(gridDimX, gridDimY); + + if (seq2batch) { + /* sequence -> batch */ + if (normByTimes) { + KeSequence2BatchPadding<1, 1><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); + } else { + KeSequence2BatchPadding<0, 1><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); + } + } else { + /* batch -> sequence */ + if (normByTimes) { + KeSequence2BatchPadding<1, 0><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); + } else { + KeSequence2BatchPadding<0, 0><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); + } + } + + CHECK_SYNC("hl_sequence2batch_copy_padding failed"); +} + +__device__ inline float my_rsqrt(float x) { return rsqrtf(x); } + +__device__ inline double my_rsqrt(double x) { return rsqrt(x); } + +__global__ void KeSequenceAvgForward(real* dst, + real* src, + const int* starts, + int height, + int width, + const int mode) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int row = gid / width; + int col = gid % width; + + if (gid < height * width) { + int start = starts[row]; + int end = starts[row + 1]; + int seqLength = end - start; + if (seqLength == 0) return; + real sum = 0.0; + for (int i = start; i < end; i++) { + sum += src[i * width + col]; + } + sum = mode == 1 ? sum : (mode == 0 ? sum / seqLength + : sum * my_rsqrt((real)seqLength)); + dst[gid] += sum; + } +} + +void hl_sequence_avg_forward(real* dst, + real* src, + const int* starts, + int height, + int width, + const int mode) { + CHECK_NOTNULL(dst); + CHECK_NOTNULL(src); + CHECK_NOTNULL(starts); + + int block = 512; + int grid = DIVUP(width * height, 512); + + CHECK(mode == 0 || mode == 1 || mode == 2) + << "mode error in hl_sequence_avg_forward!"; + + KeSequenceAvgForward<<>>( + dst, src, starts, height, width, mode); + CHECK_SYNC("hl_sequence_avg_forward failed"); +} + +__global__ void KeSequenceAvgBackward(real* dst, + real* src, + const int* starts, + int height, + int width, + const int mode) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int row = gid / width; + int col = gid % width; + + if (gid < height * width) { + int start = starts[row]; + int end = starts[row + 1]; + int seqLength = end - start; + if (seqLength == 0) return; + real grad = src[gid]; + grad = mode == 1 ? grad : (mode == 0 ? grad / seqLength + : grad * my_rsqrt((real)seqLength)); + for (int i = start; i < end; i++) { + dst[i * width + col] += grad; + } + } +} + +void hl_sequence_avg_backward(real* dst, + real* src, + const int* starts, + int height, + int width, + const int mode) { + CHECK_NOTNULL(dst); + CHECK_NOTNULL(src); + CHECK_NOTNULL(starts); + + int block = 512; + int grid = DIVUP(width * height, 512); + + CHECK(mode == 0 || mode == 1 || mode == 2) + << "mode error in hl_sequence_avg_backward!"; + + KeSequenceAvgBackward<<>>( + dst, src, starts, height, width, mode); + CHECK_SYNC("hl_sequence_avg_backward failed"); +} diff --git a/paddle/legacy/cuda/src/hl_cuda_sparse.cu b/paddle/legacy/cuda/src/hl_cuda_sparse.cu new file mode 100644 index 0000000000000000000000000000000000000000..8065a6f9f6f2ac4cacf9a63b7b80dd00391824a0 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_cuda_sparse.cu @@ -0,0 +1,1262 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_cuda.h" +#include "hl_cuda_sparse.cuh" +#include "hl_matrix_apply.cuh" +#include "hl_matrix_ops.cuh" +#include "hl_sparse.h" +#include "hl_sparse.ph" +#include "paddle/legacy/utils/Logging.h" + +DEFINE_MATRIX_UNARY_PARAMETER_OP(mul_scalar, ONE_PARAMETER, a = a * p); +DEFINE_MATRIX_UNARY_OP(Zero, a = 0); + +void hl_matrix_csr2dense(hl_sparse_matrix_s A_d, + real *C_d, + int dimM, + int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + CHECK(dimM > 0 && dimN > 0 && A_d->rows == dimM && A_d->cols == dimN); + CHECK(A_d->format == HL_SPARSE_CSR) << "matrix format error!"; + + if (A_d->nnz == 0) { + hl_gpu_apply_unary_op(unary::Zero(), C_d, dimM, dimN, dimN); + return; + } + + /* nnz != 0 */ + hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); + CHECK((A_d2->csr_val || A_d->type == HL_NO_VALUE) && A_d2->csr_row && + A_d2->csr_col) + << "parameter transa error!"; + + int blocksX = (dimN + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; + int blocksY = (dimM + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; + dim3 threads(CU_CSR2DENSE_THREAD_X, CU_CSR2DENSE_THREAD_X); + dim3 grid(blocksX, blocksY); + + if (A_d->type == HL_NO_VALUE) { + KeSMatrixCsr2Dense<0><<>>( + A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, C_d, dimM, dimN); + } else if (A_d->type == HL_FLOAT_VALUE) { + KeSMatrixCsr2Dense<1><<>>( + A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, C_d, dimM, dimN); + } else { + } + CHECK_SYNC("hl_matrix_csr2dense failed"); +} + +void hl_matrix_csc2dense(hl_sparse_matrix_s A_d, + real *C_d, + int dimM, + int dimN) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(C_d); + CHECK(dimM > 0 && dimN > 0 && A_d->rows == dimM && A_d->cols == dimN); + CHECK(A_d->format == HL_SPARSE_CSC) << "matrix format error!"; + + if (A_d->nnz == 0) { + hl_gpu_apply_unary_op(unary::Zero(), C_d, dimM, dimN, dimN); + return; + } + + /* nnz != 0 */ + hl_csc_matrix A_d2 = (hl_csc_matrix)(A_d->matrix); + CHECK((A_d2->csc_val || A_d->type == HL_NO_VALUE) && A_d2->csc_row && + A_d2->csc_col) + << "parameter transa error!"; + + int blocksX = (dimN + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; + int blocksY = (dimM + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; + dim3 threads(CU_CSR2DENSE_THREAD_X, CU_CSR2DENSE_THREAD_X); + dim3 grid(blocksX, blocksY); + + if (A_d->type == HL_NO_VALUE) { + KeSMatrixCsc2Dense<0><<>>( + A_d2->csc_val, A_d2->csc_row, A_d2->csc_col, C_d, dimM, dimN); + } else if (A_d->type == HL_FLOAT_VALUE) { + KeSMatrixCsc2Dense<1><<>>( + A_d2->csc_val, A_d2->csc_row, A_d2->csc_col, C_d, dimM, dimN); + } else { + } + CHECK_SYNC("hl_matrix_csc2dense failed"); +} + +void hl_malloc_sparse_matrix(hl_sparse_matrix_s *A_d, + hl_matrix_format_t format, + hl_matrix_value_t value_type, + int dimM, + int dimN, + int nnz) { + CHECK_NOTNULL(A_d); + CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) + << "sparse matrix format error!"; + CHECK(value_type == HL_FLOAT_VALUE || value_type == HL_NO_VALUE) + << "sparse matrix value type error!"; + /* avoid malloc 0 bytes */ + int nnz_s = (nnz == 0 ? 1 : nnz); + + if (format == HL_SPARSE_CSR) { + CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; + + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); + CHECK_NOTNULL(tmp); + + hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); + csr->sparsity = -1.0; + + if (value_type == HL_NO_VALUE) { + csr->csr_val = NULL; + csr->nnz_s = nnz_s; + csr->row_s = dimM + 1; + csr->csr_row = (int *)hl_malloc_device((dimM + 1) * sizeof(int)); + csr->csr_col = (int *)hl_malloc_device((nnz_s) * sizeof(int)); + + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csr; + } else if (value_type == HL_FLOAT_VALUE) { + csr->nnz_s = nnz_s; + csr->row_s = dimM + 1; + csr->csr_val = (real *)hl_malloc_device((nnz_s) * sizeof(real)); + csr->csr_row = (int *)hl_malloc_device((dimM + 1) * sizeof(int)); + csr->csr_col = (int *)hl_malloc_device((nnz_s) * sizeof(int)); + + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csr; + } + } else if (format == HL_SPARSE_CSC) { + CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; + + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); + CHECK_NOTNULL(tmp); + + hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); + csc->sparsity = -1.0f; + + if (value_type == HL_NO_VALUE) { + csc->csc_val = NULL; + csc->nnz_s = nnz_s; + csc->col_s = dimN + 1; + csc->csc_row = (int *)hl_malloc_device((nnz_s) * sizeof(int)); + csc->csc_col = (int *)hl_malloc_device((dimN + 1) * sizeof(int)); + + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csc; + } else if (value_type == HL_FLOAT_VALUE) { + csc->nnz_s = nnz_s; + csc->col_s = dimN + 1; + csc->csc_val = (real *)hl_malloc_device((nnz_s) * sizeof(real)); + csc->csc_row = (int *)hl_malloc_device((nnz_s) * sizeof(int)); + csc->csc_col = (int *)hl_malloc_device((dimN + 1) * sizeof(int)); + + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csc; + } + } + + (*A_d)->format = format; + (*A_d)->type = value_type; + (*A_d)->rows = dimM; + (*A_d)->cols = dimN; + (*A_d)->nnz = nnz; +} + +void hl_free_sparse_matrix(hl_sparse_matrix_s A_d) { + CHECK_NOTNULL(A_d); + CHECK(A_d->format == HL_SPARSE_CSR || A_d->format == HL_SPARSE_CSC) + << "sparse matrix format error!"; + + if (A_d->matrix == NULL) { + free(A_d); + return; + } + + if (A_d->format == HL_SPARSE_CSR) { + hl_csr_matrix csr = (hl_csr_matrix)A_d->matrix; + if (csr->csr_val != NULL) { + hl_free_mem_device(csr->csr_val); + csr->csr_val = NULL; + } + + if (csr->csr_row != NULL) { + hl_free_mem_device(csr->csr_row); + csr->csr_row = NULL; + } + + if (csr->csr_col != NULL) { + hl_free_mem_device(csr->csr_col); + csr->csr_col = NULL; + } + + A_d->matrix = NULL; + free(A_d); + } else if (A_d->format == HL_SPARSE_CSC) { + hl_csc_matrix csc = (hl_csc_matrix)A_d->matrix; + if (csc->csc_val != NULL) { + hl_free_mem_device(csc->csc_val); + csc->csc_val = NULL; + } + + if (csc->csc_row != NULL) { + hl_free_mem_device(csc->csc_row); + csc->csc_row = NULL; + } + + if (csc->csc_col != NULL) { + hl_free_mem_device(csc->csc_col); + csc->csc_col = NULL; + } + + A_d->matrix = NULL; + free(A_d); + } +} + +void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, + void *dest_d, + size_t size, + hl_matrix_format_t format, + hl_matrix_value_t value_type, + int dimM, + int dimN, + int nnz) { + CHECK_NOTNULL(A_d); + CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) + << "sparse matrix format error!"; + + if (format == HL_SPARSE_CSR) { + CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; + + size_t size_ = (dimM + 1) * sizeof(int) + nnz * sizeof(int); + if (value_type != HL_NO_VALUE) { + size_ += nnz * sizeof(real); + } + CHECK_LE(size_, size) << "dest_d size(" << size + << ") too small, should bigger than(" << size_ + << ")!"; + + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); + CHECK_NOTNULL(tmp); + + hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); + + if (value_type == HL_NO_VALUE) { + csr->csr_val = NULL; + csr->csr_row = (int *)dest_d; + csr->csr_col = (int *)((char *)dest_d + (dimM + 1) * sizeof(int)); + } else { + csr->csr_val = (real *)dest_d; + csr->csr_row = (int *)((char *)dest_d + nnz * sizeof(real)); + csr->csr_col = (int *)((char *)dest_d + nnz * sizeof(real) + + (dimM + 1) * sizeof(int)); + } + csr->nnz_s = nnz; + csr->row_s = dimM + 1; + csr->sparsity = -1.0; + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csr; + } else if (format == HL_SPARSE_CSC) { + CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; + + size_t size_ = (dimN + 1) * sizeof(int) + nnz * sizeof(int); + if (value_type != HL_NO_VALUE) { + size_ += nnz * sizeof(real); + } + CHECK_LE(size_, size) << "dest_d size(" << size + << ") too small, should bigger than(" << size_ + << ")!"; + + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); + CHECK_NOTNULL(tmp); + + hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); + if (value_type == HL_NO_VALUE) { + csc->csc_val = NULL; + csc->csc_col = (int *)dest_d; + csc->csc_row = (int *)((char *)dest_d + (dimN + 1) * sizeof(int)); + } else { + csc->csc_val = (real *)dest_d; + csc->csc_col = (int *)((char *)dest_d + nnz * sizeof(real)); + csc->csc_row = (int *)((char *)dest_d + nnz * sizeof(real) + + (dimN + 1) * sizeof(int)); + } + csc->nnz_s = nnz; + csc->col_s = dimN + 1; + csc->sparsity = -1.0f; + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csc; + } + + (*A_d)->format = format; + (*A_d)->type = value_type; + (*A_d)->rows = dimM; + (*A_d)->cols = dimN; + (*A_d)->nnz = nnz; +} + +void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, + real *value_d, + int *rows_d, + int *cols_d, + hl_matrix_format_t format, + hl_matrix_value_t value_type, + int dimM, + int dimN, + int nnz) { + CHECK_NOTNULL(A_d); + CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; + + CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) + << "sparse matrix format error!"; + + if (format == HL_SPARSE_CSR) { + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); + CHECK_NOTNULL(tmp); + + hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); + csr->csr_row = rows_d; + csr->csr_col = cols_d; + csr->csr_val = value_d; + csr->nnz_s = nnz; + csr->row_s = dimM + 1; + csr->sparsity = -1.0; + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csr; + } else if (format == HL_SPARSE_CSC) { + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); + CHECK_NOTNULL(tmp); + + hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); + csc->csc_row = rows_d; + csc->csc_col = cols_d; + csc->csc_val = value_d; + csc->nnz_s = nnz; + csc->col_s = dimN + 1; + csc->sparsity = -1.0f; + *A_d = (hl_sparse_matrix_s)tmp; + (*A_d)->matrix = (hl_matrix_s)csc; + } + + (*A_d)->format = format; + (*A_d)->type = value_type; + (*A_d)->rows = dimM; + (*A_d)->cols = dimN; + (*A_d)->nnz = nnz; +} + +void hl_destruct_sparse_matrix(hl_sparse_matrix_s A_d) { + CHECK_NOTNULL(A_d); + free(A_d); +} + +void hl_memcpy_csr_matrix(hl_sparse_matrix_s csr_matrix, + real *csr_val, + int *csr_row, + int *csr_col, + hl_stream_t stream) { + CHECK_NOTNULL(csr_matrix); + CHECK_EQ(csr_matrix->format, HL_SPARSE_CSR) + << "csr_matrix is not csr format!"; + CHECK_NOTNULL(csr_matrix->matrix); + + hl_csr_matrix csr = (hl_csr_matrix)(csr_matrix->matrix); + CHECK_LE(csr_matrix->nnz, csr->nnz_s) << "copy size " << csr_matrix->nnz + << " is big than alloc size " + << csr->nnz_s; + + CHECK_LE((csr_matrix->rows + 1), csr->row_s) + << "copy size " << (csr_matrix->rows + 1) << " is big than alloc size " + << csr->row_s; + + CHECK(csr_matrix->type == HL_FLOAT_VALUE || csr_matrix->type == HL_NO_VALUE) + << "sparse matrix value type error!"; + + if (csr_matrix->type == HL_NO_VALUE) { + if (csr_row == NULL && csr_col == NULL) { + return; + } else if (csr_row != NULL && csr_col != NULL) { + hl_memcpy_async( + csr->csr_row, csr_row, (csr_matrix->rows + 1) * sizeof(int), stream); + + hl_memcpy_async( + csr->csr_col, csr_col, (csr_matrix->nnz) * sizeof(int), stream); + } else { + LOG(FATAL) << "parameter csr_row or csr_col is null pointer!"; + } + } else if (csr_matrix->type == HL_FLOAT_VALUE) { + if (csr_val == NULL && csr_row == NULL && csr_col == NULL) { + return; + } else if (csr_val != NULL && csr_row == NULL && csr_col == NULL) { + hl_memcpy_async( + csr->csr_val, csr_val, (csr_matrix->nnz) * sizeof(real), stream); + } else if (csr_val != NULL && csr_row != NULL && csr_col != NULL) { + hl_memcpy_async( + csr->csr_val, csr_val, (csr_matrix->nnz) * sizeof(real), stream); + hl_memcpy_async( + csr->csr_row, csr_row, (csr_matrix->rows + 1) * sizeof(int), stream); + hl_memcpy_async( + csr->csr_col, csr_col, (csr_matrix->nnz) * sizeof(int), stream); + } else { + LOG(FATAL) << "parameter csr_row or csr_col is null pointer!"; + } + } + + csr->sparsity = ((float)csr_matrix->nnz) / ((float)csr_matrix->rows) / + ((float)csr_matrix->cols); +} + +void hl_memcpy_csc_matrix(hl_sparse_matrix_s csc_matrix, + real *csc_val, + int *csc_row, + int *csc_col, + hl_stream_t stream) { + CHECK_NOTNULL(csc_matrix); + CHECK_EQ(csc_matrix->format, HL_SPARSE_CSC) + << "csc_matrix is not csc format error!"; + + hl_csc_matrix csc = (hl_csc_matrix)(csc_matrix->matrix); + CHECK_LE(csc_matrix->nnz, csc->nnz_s) << "copy size " << csc_matrix->nnz + << " is big than alloc size " + << csc->nnz_s; + + CHECK_LE((csc_matrix->cols + 1), csc->col_s) + << "copy size " << (csc_matrix->cols + 1) << " is big than alloc size " + << csc->col_s; + + CHECK(csc_matrix->type == HL_FLOAT_VALUE || csc_matrix->type == HL_NO_VALUE) + << "sparse matrix value type error!"; + + if (csc_matrix->type == HL_NO_VALUE) { + if (csc_row == NULL && csc_col == NULL) { + return; + } else if (csc_row != NULL && csc_col != NULL) { + hl_memcpy_async( + csc->csc_row, csc_row, (csc_matrix->nnz) * sizeof(int), stream); + hl_memcpy_async( + csc->csc_col, csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); + } else { + LOG(FATAL) << "parameter csc_row or csc_col is null pointer!"; + } + } else if (csc_matrix->type == HL_FLOAT_VALUE) { + if (csc_val == NULL && csc_row == NULL && csc_col == NULL) { + return; + } else if (csc_val != NULL && csc_row == NULL && csc_col == NULL) { + hl_memcpy_async( + csc->csc_val, csc_val, (csc_matrix->nnz) * sizeof(real), stream); + } else if (csc_val != NULL && csc_row != NULL && csc_col != NULL) { + hl_memcpy_async( + csc->csc_val, csc_val, (csc_matrix->nnz) * sizeof(real), stream); + hl_memcpy_async( + csc->csc_row, csc_row, (csc_matrix->nnz) * sizeof(int), stream); + hl_memcpy_async( + csc->csc_col, csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); + } else { + LOG(FATAL) << "parameter csc_row or csc_col is null pointer!"; + } + } + + csc->sparsity = ((float)csc_matrix->nnz) / ((float)csc_matrix->rows) / + ((float)csc_matrix->cols); +} + +void hl_memcpy_sparse_matrix(hl_sparse_matrix_s dst, + hl_sparse_matrix_s src, + hl_stream_t stream) { + CHECK(dst && src && dst->matrix && src->matrix) + << "parameter dst or src is null pointer!"; + CHECK_EQ(dst->format, src->format) << "sparse matrix format does not match!"; + CHECK(dst->type != HL_FLOAT_VALUE || src->type != HL_NO_VALUE) + << "src sparse matrix is no value, dst sparse matrix has value!"; + + if (dst->format == HL_SPARSE_CSR) { + dst->rows = src->rows; + dst->cols = src->cols; + dst->nnz = src->nnz; + hl_csr_matrix csr = (hl_csr_matrix)src->matrix; + hl_memcpy_csr_matrix(dst, csr->csr_val, csr->csr_row, csr->csr_col, stream); + } else if (dst->format == HL_SPARSE_CSC) { + dst->rows = src->rows; + dst->cols = src->cols; + dst->nnz = src->nnz; + hl_csc_matrix csc = (hl_csc_matrix)src->matrix; + hl_memcpy_csc_matrix(dst, csc->csc_val, csc->csc_row, csc->csc_col, stream); + } else { + LOG(FATAL) << "sparse matrix format error!"; + } +} + +/** + * Calculate beta * C, if beta is zero, C does not have to be a valid input. + */ +static void _beta_mul_c(real *c, int dimM, int dimN, real beta) { + if (beta == 0.0) { + hl_gpu_apply_unary_op(unary::Zero(), c, dimM, dimN, dimN); + } else { + if (beta != 1.0) { + hl_gpu_apply_unary_op(unary::mul_scalar(beta), c, dimM, dimN, dimN); + } + } + + return; +} + +void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, + real *C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { + CHECK_EQ(transb, HPPL_OP_N); + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + CHECK(dimM > 0 && dimN > 0 && dimK > 0); + CHECK_EQ(A_d->format, HL_SPARSE_CSR) << "matrix format error!"; + + if ((HPPL_OP_N == transa && (A_d->rows != dimM || A_d->cols != dimK)) || + (HPPL_OP_T == transa && (A_d->rows != dimK || A_d->cols != dimM))) { + LOG(FATAL) << "parameter error!"; + } + + if (A_d->nnz == 0) { + _beta_mul_c(C_d, dimM, dimN, beta); + return; + } + + /* nnz != 0 */ + hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); + if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || + A_d2->csr_row == NULL || A_d2->csr_col == NULL) { + LOG(FATAL) << "parameter error!"; + } + + if (HPPL_OP_N == transa) { + int blocksX = (dimN + CU_CSRMM_BLOCK_N - 1) / CU_CSRMM_BLOCK_N; + int blocksY = (dimM + CU_CSRMM_THREAD_Y - 1) / CU_CSRMM_THREAD_Y; + dim3 threads(CU_CSRMM_THREAD_X, CU_CSRMM_THREAD_Y); + dim3 grid(blocksX, blocksY); + + /* sparsity pattern */ + // A_d->sparsity; + if (A_d->type == HL_NO_VALUE) { + KeSMatrixCsrMulDense<0><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixCsrMulDense<1><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else if (HPPL_OP_T == transa) { + _beta_mul_c(C_d, dimM, dimN, beta); + + int blocksX = + (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / CU_CSC_MUL_DENSE_BLOCK_N; + int blocksY = + (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / CU_CSC_MUL_DENSE_BLOCK_K; + dim3 threads(CU_CSC_MUL_DENSE_THREAD_X, CU_CSC_MUL_DENSE_THREAD_Y); + dim3 grid(blocksX, blocksY); + if (A_d->type == HL_NO_VALUE) { + KeSMatrixCscMulDense<0><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixCscMulDense<1><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else { + LOG(FATAL) << "parameter transa error!"; + } + + CHECK_SYNC("hl_matrix_csr_mul_dense failed"); +} + +void hl_matrix_dense_mul_csc(real *A_d, + hl_trans_op_t transa, + hl_sparse_matrix_s B_d, + hl_trans_op_t transb, + real *C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { + CHECK_EQ(transa, HPPL_OP_N); + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + + if (dimM <= 0 || dimN <= 0 || dimK <= 0 || + ((transb == HPPL_OP_N) && (B_d->rows != dimK || B_d->cols != dimN)) || + ((transb == HPPL_OP_T) && (B_d->rows != dimN || B_d->cols != dimK))) { + LOG(FATAL) << "parameter dims error!"; + } + + CHECK_EQ(B_d->format, HL_SPARSE_CSC) << "matrix format error!"; + + if (B_d->nnz == 0) { + _beta_mul_c(C_d, dimM, dimN, beta); + return; + } + + /* nnz != 0 */ + hl_csc_matrix B_d2 = (hl_csc_matrix)(B_d->matrix); + if ((B_d2->csc_val == NULL && B_d->type != HL_NO_VALUE) || + B_d2->csc_row == NULL || B_d2->csc_col == NULL) { + LOG(FATAL) << "parameter B is null!"; + } + + if (transb == HPPL_OP_N) { + int blocksX = (dimM + CU_CSCMM_BLOCK_M_BEST - 1) / CU_CSCMM_BLOCK_M_BEST; + int blocksY = (dimN + CU_CSCMM_BLOCK_N_BEST - 1) / CU_CSCMM_BLOCK_N_BEST; + dim3 threads(CU_CSCMM_THREAD_X_BEST, CU_CSCMM_THREAD_Y_BEST); + dim3 grid(blocksX, blocksY); + + if (B_d->type == HL_NO_VALUE) { + KeSMatrixDenseMulCsc<0><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_row, + B_d2->csc_col, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixDenseMulCsc<1><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_row, + B_d2->csc_col, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else if (transb == HPPL_OP_T) { + _beta_mul_c(C_d, dimM, dimN, beta); + int blocksX = 1 + (dimK - 1) / CU_DM_CSR_THREAD_X; + int blocksY = 1 + (dimM - 1) / CU_DM_CSR_BLOCK_M; + dim3 threads(CU_DM_CSR_THREAD_X, CU_DM_CSR_THREAD_Y); + dim3 grid(blocksX, blocksY); + if (B_d->type == HL_NO_VALUE) { + KeSMatrixDenseMulCsr<0><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_col, + B_d2->csc_row, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixDenseMulCsr<1><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_col, + B_d2->csc_row, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else { + LOG(FATAL) << "parameter transb error!"; + } + + CHECK_SYNC("hl_matrix_dense_mul_csc failed"); +} + +void hl_matrix_dense_mul_csr(real *A_d, + hl_trans_op_t transa, + hl_sparse_matrix_s B_d, + hl_trans_op_t transb, + real *C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { + CHECK_EQ(transa, HPPL_OP_N); + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + + if (dimM <= 0 || dimN <= 0 || dimK <= 0 || + (transb == HPPL_OP_N && (B_d->rows != dimK || B_d->cols != dimN)) || + (transb == HPPL_OP_T && (B_d->rows != dimN || B_d->cols != dimK))) { + LOG(FATAL) << "parameter dims error!"; + } + + CHECK_EQ(B_d->format, HL_SPARSE_CSR) << "matrix format error!"; + + if (B_d->nnz == 0) { + _beta_mul_c(C_d, dimM, dimN, beta); + return; + } + + /* nnz != 0 */ + hl_csr_matrix B_d2 = (hl_csr_matrix)(B_d->matrix); + if ((B_d2->csr_val == NULL && B_d->type != HL_NO_VALUE) || + B_d2->csr_row == NULL || B_d2->csr_col == NULL) { + LOG(FATAL) << "parameter transa error!"; + } + + if (transb == HPPL_OP_N) { + _beta_mul_c(C_d, dimM, dimN, beta); + int blocksX = 1 + (dimK - 1) / CU_DM_CSR_THREAD_X; + int blocksY = 1 + (dimM - 1) / CU_DM_CSR_BLOCK_M; + dim3 threads(CU_DM_CSR_THREAD_X, CU_DM_CSR_THREAD_Y); + dim3 grid(blocksX, blocksY); + if (B_d->type == HL_NO_VALUE) { + KeSMatrixDenseMulCsr<0><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_row, + B_d2->csr_col, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixDenseMulCsr<1><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_row, + B_d2->csr_col, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else if (transb == HPPL_OP_T) { + int blocksX = (dimM + CU_CSCMM_BLOCK_M_BEST - 1) / CU_CSCMM_BLOCK_M_BEST; + int blocksY = (dimN + CU_CSCMM_BLOCK_N_BEST - 1) / CU_CSCMM_BLOCK_N_BEST; + dim3 threads(CU_CSCMM_THREAD_X_BEST, CU_CSCMM_THREAD_Y_BEST); + dim3 grid(blocksX, blocksY); + if (B_d->type == HL_NO_VALUE) { + KeSMatrixDenseMulCsc<0><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_col, + B_d2->csr_row, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixDenseMulCsc<1><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_col, + B_d2->csr_row, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else { + LOG(FATAL) << "parameter transb error!"; + } + + CHECK_SYNC("hl_matrix_dense_mul_csr failed"); +} + +void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, + real *C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { + CHECK_EQ(transb, HPPL_OP_N); + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + CHECK(dimM > 0 && dimN > 0 && dimK > 0) << "parameter error!"; + CHECK_EQ(A_d->format, HL_SPARSE_CSC) << "matrix format error!"; + + if ((HPPL_OP_N == transa && (A_d->rows != dimM || A_d->cols != dimK)) || + (HPPL_OP_T == transa && (A_d->rows != dimK || A_d->cols != dimM))) { + LOG(FATAL) << "parameter error!"; + } + + if (A_d->nnz == 0) { + _beta_mul_c(C_d, dimM, dimN, beta); + return; + } + + /* nnz != 0 */ + hl_csc_matrix A_d2 = (hl_csc_matrix)(A_d->matrix); + if ((A_d2->csc_val == NULL && A_d->type != HL_NO_VALUE) || + A_d2->csc_row == NULL || A_d2->csc_col == NULL) { + LOG(FATAL) << "parameter error!"; + } + + if (HPPL_OP_N == transa) { + _beta_mul_c(C_d, dimM, dimN, beta); + + int blocksX = + (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / CU_CSC_MUL_DENSE_BLOCK_N; + int blocksY = + (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / CU_CSC_MUL_DENSE_BLOCK_K; + dim3 threads(CU_CSC_MUL_DENSE_THREAD_X, CU_CSC_MUL_DENSE_THREAD_Y); + dim3 grid(blocksX, blocksY); + if (A_d->type == HL_NO_VALUE) { + KeSMatrixCscMulDense<0><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixCscMulDense<1><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else if (HPPL_OP_T == transa) { + int blocksX = (dimN + CU_CSRMM_BLOCK_N - 1) / CU_CSRMM_BLOCK_N; + int blocksY = (dimM + CU_CSRMM_THREAD_Y - 1) / CU_CSRMM_THREAD_Y; + dim3 threads(CU_CSRMM_THREAD_X, CU_CSRMM_THREAD_Y); + dim3 grid(blocksX, blocksY); + + /* sparsity pattern */ + // A_d->sparsity; + if (A_d->type == HL_NO_VALUE) { + KeSMatrixCsrMulDense<0><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } else { + KeSMatrixCsrMulDense<1><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); + } + } else { + LOG(FATAL) << "parameter transa error!"; + } + + CHECK_SYNC("hl_matrix_csc_mul_dense failed"); +} + +void hl_sparse_matrix_mul(real *A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, + hl_sparse_matrix_s C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + CHECK_NOTNULL(C_d); + CHECK(dimM > 0 && dimN > 0 && dimK > 0) << "parameter error!"; + CHECK_NE(C_d->type, HL_NO_VALUE) << "C value type error!"; + + if (C_d->nnz == 0) return; + + if (C_d->format == HL_SPARSE_CSC) { + hl_csc_matrix C_d2 = (hl_csc_matrix)(C_d->matrix); + if (C_d2->csc_val == NULL || C_d2->csc_row == NULL || + C_d2->csc_col == NULL) { + LOG(FATAL) << "parameter error!"; + } + + if (beta != 1.0) { + hl_gpu_apply_unary_op( + unary::mul_scalar(beta), C_d2->csc_val, 1, C_d->nnz, C_d->nnz); + } + + int blocksX = dimN; + int blocksY = 1; + dim3 threads(CU_CSCMM_DMD2CSC_THREAD_X, 1); + dim3 grid(blocksX, blocksY); + bool transA = transa == HPPL_OP_T ? 1 : 0; + bool transB = transb == HPPL_OP_T ? 1 : 0; + KeSMatrixDenseMulDense2CSC<<>>( + C_d2->csc_val, + C_d2->csc_row, + C_d2->csc_col, + A_d, + B_d, + transA, + transB, + dimM, + dimN, + dimK, + alpha, + beta); + CHECK_SYNC("hl_sparse_matrix_mul failed"); + } else { + hl_csr_matrix C_d2 = (hl_csr_matrix)(C_d->matrix); + if ((C_d2->csr_val == NULL && C_d->type != HL_NO_VALUE) || + C_d2->csr_row == NULL || C_d2->csr_col == NULL) { + LOG(FATAL) << "parameter error!"; + } + + if (beta != 1.0) { + hl_gpu_apply_unary_op( + unary::mul_scalar(beta), C_d2->csr_val, 1, C_d->nnz, C_d->nnz); + } + + bool transA = transa == HPPL_OP_T ? 1 : 0; + bool transB = transb == HPPL_OP_T ? 1 : 0; + if (!transB) { + int blocksX = dimM; + int blocksY = 1; + dim3 threads(CU_CSCMM_DMD2CSR_THREAD_X, 1); + dim3 grid(blocksX, blocksY); + + KeSMatrixDenseMulDense2CSR<<>>( + C_d2->csr_val, + C_d2->csr_row, + C_d2->csr_col, + A_d, + B_d, + transA, + transB, + dimM, + dimN, + dimK, + alpha, + beta); + CHECK_SYNC("hl_sparse_matrix_mul failed"); + } else { + CHECK(!transA) << "Not supported A is trans and B is not trans!"; + + dim3 block(CU_BLOCK_SIZE, 1); + int avgNnzPerRow = C_d->nnz / dimM; + avgNnzPerRow = avgNnzPerRow > 0 ? avgNnzPerRow : 1; + int gridx = DIVUP(avgNnzPerRow, CU_BLOCK_SIZE); + dim3 grid(gridx, dimM); + KeSMatrixDenseMulDenseTrans2CSR<<>>( + C_d2->csr_val, + C_d2->csr_row, + C_d2->csr_col, + A_d, + B_d, + transA, + transB, + dimM, + dimN, + dimK, + alpha, + beta); + CHECK_SYNC("hl_sparse_matrix_mul failed"); + } + } +} + +void hl_memcpy_from_csc_matrix(real *csc_val, + size_t val_size, + int *csc_row, + size_t row_size, + int *csc_col, + size_t col_size, + hl_sparse_matrix_s csc_matrix, + hl_stream_t stream) { + CHECK_NOTNULL(csc_matrix); + CHECK_NOTNULL(csc_row); + CHECK_NOTNULL(csc_col); + + CHECK_EQ(csc_matrix->format, HL_SPARSE_CSC) + << "csc_matrix is not csc format error!"; + + if (csc_matrix->nnz > row_size || + csc_matrix->cols + 1 > static_cast(col_size)) { + LOG(FATAL) << "size not match!"; + } + + hl_csc_matrix csc = (hl_csc_matrix)(csc_matrix->matrix); + hl_memcpy_async((void *)csc_row, + (void *)csc->csc_row, + (csc_matrix->nnz) * sizeof(int), + stream); + hl_memcpy_async((void *)csc_col, + (void *)csc->csc_col, + (csc_matrix->cols + 1) * sizeof(int), + stream); + if (csc_matrix->type == HL_FLOAT_VALUE) { + if (csc_val != NULL) { + CHECK_LE(csc_matrix->nnz, val_size) << "size not match!"; + hl_memcpy_async((void *)csc_val, + (void *)csc->csc_val, + (csc_matrix->nnz) * sizeof(real), + stream); + } else { + LOG(FATAL) << "parameter csr_val is null pointer!"; + } + } +} + +void hl_memcpy_from_csr_matrix(real *csr_val, + size_t val_size, + int *csr_row, + size_t row_size, + int *csr_col, + size_t col_size, + hl_sparse_matrix_s csr_matrix, + hl_stream_t stream) { + CHECK_NOTNULL(csr_matrix); + CHECK_NOTNULL(csr_row); + CHECK_NOTNULL(csr_col); + CHECK_EQ(csr_matrix->format, HL_SPARSE_CSR) + << "csr_matrix is not csr format error!"; + + if (csr_matrix->nnz > col_size || + csr_matrix->rows + 1 > static_cast(row_size)) { + LOG(FATAL) << "size not match!"; + } + + hl_csr_matrix csr = (hl_csr_matrix)(csr_matrix->matrix); + hl_memcpy_async((void *)csr_row, + (void *)csr->csr_row, + (csr_matrix->rows + 1) * sizeof(int), + stream); + hl_memcpy_async((void *)csr_col, + (void *)csr->csr_col, + (csr_matrix->nnz) * sizeof(int), + stream); + if (csr_matrix->type == HL_FLOAT_VALUE) { + if (csr_val != NULL) { + CHECK_LE(csr_matrix->nnz, val_size) << "size not match!"; + hl_memcpy_async((void *)csr_val, + (void *)csr->csr_val, + (csr_matrix->nnz) * sizeof(real), + stream); + } else { + LOG(FATAL) << "parameter csr_val is null pointer!"; + } + } +} + +void hl_sparse_matrix_column_sum( + real *A_d, hl_sparse_matrix_s B_d, int dimM, int dimN, real scale) { + if (B_d->format == HL_SPARSE_CSR) { + hl_matrix_csr_column_sum(A_d, B_d, dimM, dimN, scale); + } else { + LOG(FATAL) << "Not support CSC format error!"; + } +} + +void hl_matrix_csr_column_sum( + real *A_d, hl_sparse_matrix_s B_d, int dimM, int dimN, real scale) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + + if (dimM <= 0 || dimN <= 0 || (B_d->rows != dimM || B_d->cols != dimN)) { + LOG(FATAL) << "parameter dims error!"; + } + + hl_csr_matrix B_d2 = (hl_csr_matrix)(B_d->matrix); + if ((B_d2->csr_val == NULL && B_d->type != HL_NO_VALUE) || + B_d2->csr_row == NULL || B_d2->csr_col == NULL) { + LOG(FATAL) << "parameter B is null!"; + } + + if (B_d->nnz == 0) return; + + int nnz = B_d->nnz; + int block = 512; + int grid = DIVUP(nnz, 512); + KeSMatrixCsrColumnSum<<>>( + A_d, B_d2->csr_val, B_d2->csr_col, nnz); + + CHECK_SYNC("hl_matrix_csr_column_sum failed"); +} + +void hl_sparse_matrix_add_bias(hl_sparse_matrix_s A_d, real *B_d, real scale) { + if (A_d->format == HL_SPARSE_CSR) { + hl_matrix_csr_add_bias(A_d, B_d, scale); + } else { + LOG(FATAL) << "Not support CSC format error!"; + } +} + +void hl_matrix_csr_add_bias(hl_sparse_matrix_s A_d, real *B_d, real scale) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + + hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); + if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || + A_d2->csr_row == NULL || A_d2->csr_col == NULL) { + LOG(FATAL) << "parameter A_d is null!"; + } + + if (A_d->nnz == 0) return; + + int nnz = A_d->nnz; + int block = 512; + int grid = DIVUP(nnz, 512); + KeSMatrixCsrAddBias<<>>( + A_d2->csr_val, A_d2->csr_col, B_d, scale, nnz); + + CHECK_SYNC("hl_sparse_matrix_add_bias failed"); +} + +void hl_sparse_matrix_add_dense(hl_sparse_matrix_s A_d, + real *B_d, + int dimM, + int dimN, + real alpha, + real beta) { + if (A_d->format == HL_SPARSE_CSR) { + hl_matrix_csr_add_dense(A_d, B_d, dimM, dimN, alpha, beta); + } else { + LOG(FATAL) << "Not support CSC format error!"; + } +} + +void hl_matrix_csr_add_dense(hl_sparse_matrix_s A_d, + real *B_d, + int dimM, + int dimN, + real alpha, + real beta) { + CHECK_NOTNULL(A_d); + CHECK_NOTNULL(B_d); + + if (dimM <= 0 || dimN <= 0 || A_d->rows != dimM || A_d->cols != dimN) { + LOG(FATAL) << "parameter dim error!"; + } + + hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); + if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || + A_d2->csr_row == NULL || A_d2->csr_col == NULL) { + LOG(FATAL) << "parameter A_d is null!"; + } + + if (A_d->nnz == 0) return; + + int gridX = DIVUP((A_d->nnz / dimM), 512); + gridX = gridX > 0 ? gridX : 1; + dim3 block(512, 1); + dim3 grid(gridX, dimM); + KeSMatrixCsrAddDense<<>>(A_d2->csr_val, + A_d2->csr_row, + A_d2->csr_col, + B_d, + alpha, + beta, + dimM, + dimN); + + CHECK_SYNC("hl_sparse_matrix_add_dense failed"); +} + +int *hl_sparse_matrix_get_rows(hl_sparse_matrix_s sMat) { + __sparse_get_return__(sMat, row); +} + +int *hl_sparse_matrix_get_cols(hl_sparse_matrix_s sMat) { + __sparse_get_return__(sMat, col); +} + +real *hl_sparse_matrix_get_value(hl_sparse_matrix_s sMat) { + __sparse_get_return__(sMat, val); +} diff --git a/paddle/cuda/src/hl_cuda_sparse.cuh b/paddle/legacy/cuda/src/hl_cuda_sparse.cuh similarity index 100% rename from paddle/cuda/src/hl_cuda_sparse.cuh rename to paddle/legacy/cuda/src/hl_cuda_sparse.cuh diff --git a/paddle/cuda/src/hl_math.cc b/paddle/legacy/cuda/src/hl_math.cc similarity index 100% rename from paddle/cuda/src/hl_math.cc rename to paddle/legacy/cuda/src/hl_math.cc diff --git a/paddle/cuda/src/hl_perturbation_util.cu b/paddle/legacy/cuda/src/hl_perturbation_util.cu similarity index 100% rename from paddle/cuda/src/hl_perturbation_util.cu rename to paddle/legacy/cuda/src/hl_perturbation_util.cu diff --git a/paddle/legacy/cuda/src/hl_table_apply.cu b/paddle/legacy/cuda/src/hl_table_apply.cu new file mode 100644 index 0000000000000000000000000000000000000000..7411ae35d382833253e3ceabe36b3a1938138028 --- /dev/null +++ b/paddle/legacy/cuda/src/hl_table_apply.cu @@ -0,0 +1,124 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_base.h" +#include "hl_cuda.h" +#include "hl_device_functions.cuh" +#include "paddle/legacy/utils/Logging.h" + +template +__global__ void KeMatrixAddRows(real* output, + int ldo, + real* table, + int ldt, + int* ids, + int numSamples, + int tableSize, + int dim) { + int idx = threadIdx.x; + int idy = blockIdx.x + threadIdx.y * gridDimX; + + while (idy < numSamples) { + int tableId = ids[idy]; + if ((0 <= tableId) && (tableId < tableSize)) { + real* out = output + idy * ldo; + real* tab = table + tableId * ldt; + for (int i = idx; i < dim; i += blockDimX) { + if (AddRow) { + paddle::paddleAtomicAdd(&tab[i], out[i]); + } else { + out[i] += tab[i]; + } + } + } + idy += blockDimY * gridDimX; + } +} + +void hl_matrix_select_rows(real* output, + int ldo, + real* table, + int ldt, + int* ids, + int numSamples, + int tableSize, + int dim) { + CHECK_NOTNULL(output); + CHECK_NOTNULL(table); + CHECK_NOTNULL(ids); + + dim3 threads(128, 8); + dim3 grid(8, 1); + KeMatrixAddRows<128, 8, 8, 0><<>>( + output, ldo, table, ldt, ids, numSamples, tableSize, dim); + + CHECK_SYNC("hl_matrix_select_rows failed"); +} + +void hl_matrix_add_to_rows(real* table, + int ldt, + real* input, + int ldi, + int* ids, + int numSamples, + int tableSize, + int dim) { + CHECK_NOTNULL(input); + CHECK_NOTNULL(table); + CHECK_NOTNULL(ids); + + dim3 threads(128, 8); + dim3 grid(8, 1); + KeMatrixAddRows<128, 8, 8, 1><<>>( + input, ldi, table, ldt, ids, numSamples, tableSize, dim); + + CHECK_SYNC("hl_matrix_add_to_rows failed"); +} + +template +__global__ void KeVectorSelect( + T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) { + int idx = threadIdx.x + blockDimX * blockIdx.x; + while (idx < sizei) { + int index = ids[idx]; + // check(index < sizes); + dst[idx] = src[index]; + idx += blockDimX * gridDimX; + } +} + +template +void hl_vector_select_from( + T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) { + CHECK_NOTNULL(dst); + CHECK_NOTNULL(src); + CHECK_NOTNULL(ids); + CHECK_EQ(sized, sizei); + + dim3 threads(512, 1); + dim3 grid(8, 1); + KeVectorSelect<<>>( + dst, sized, src, sizes, ids, sizei); + + CHECK_SYNC("hl_vector_select_from failed"); +} + +template void hl_vector_select_from(real* dst, + int sized, + const real* src, + int sizes, + const int* ids, + int sizei); +template void hl_vector_select_from( + int* dst, int sized, const int* src, int sizes, const int* ids, int sizei); diff --git a/paddle/cuda/src/hl_time.cc b/paddle/legacy/cuda/src/hl_time.cc similarity index 100% rename from paddle/cuda/src/hl_time.cc rename to paddle/legacy/cuda/src/hl_time.cc diff --git a/paddle/legacy/cuda/src/hl_top_k.cu b/paddle/legacy/cuda/src/hl_top_k.cu new file mode 100644 index 0000000000000000000000000000000000000000..041ac419f5addfa49148270b8a8b421eb8ada78c --- /dev/null +++ b/paddle/legacy/cuda/src/hl_top_k.cu @@ -0,0 +1,481 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/cuda/include/hl_base.h" +#include "paddle/legacy/cuda/include/hl_sparse.ph" +#include "paddle/legacy/cuda/include/hl_top_k.h" +#include "paddle/legacy/utils/Logging.h" + +// using namespace hppl; + +struct Pair { + __device__ __forceinline__ Pair() {} + + __device__ __forceinline__ Pair(real value, int id) : v_(value), id_(id) {} + + __device__ __forceinline__ void set(real value, int id) { + v_ = value; + id_ = id; + } + + __device__ __forceinline__ void operator=(const Pair& in) { + v_ = in.v_; + id_ = in.id_; + } + + __device__ __forceinline__ bool operator<(const real value) const { + return (v_ < value); + } + + __device__ __forceinline__ bool operator<(const Pair& in) const { + return (v_ < in.v_) || ((v_ == in.v_) && (id_ > in.id_)); + } + + __device__ __forceinline__ bool operator>(const Pair& in) const { + return (v_ > in.v_) || ((v_ == in.v_) && (id_ < in.id_)); + } + + real v_; + int id_; +}; + +__device__ __forceinline__ void addTo(Pair topK[], + const Pair& p, + int beamSize) { + for (int k = beamSize - 2; k >= 0; k--) { + if (topK[k] < p) { + topK[k + 1] = topK[k]; + } else { + topK[k + 1] = p; + return; + } + } + topK[0] = p; +} + +template +__device__ __forceinline__ void addTo(Pair topK[], const Pair& p) { + for (int k = beamSize - 2; k >= 0; k--) { + if (topK[k] < p) { + topK[k + 1] = topK[k]; + } else { + topK[k + 1] = p; + return; + } + } + topK[0] = p; +} + +template +__device__ __forceinline__ void getTopK( + Pair topK[], real* src, int idx, int dim, int beamSize) { + while (idx < dim) { + if (topK[beamSize - 1] < src[idx]) { + Pair tmp(src[idx], idx); + addTo(topK, tmp, beamSize); + } + idx += blockSize; + } +} + +template +__device__ __forceinline__ void getTopK( + Pair topK[], real* src, int idx, int dim, const Pair& max, int beamSize) { + while (idx < dim) { + if (topK[beamSize - 1] < src[idx]) { + Pair tmp(src[idx], idx); + if (tmp < max) { + addTo(topK, tmp, beamSize); + } + } + idx += blockSize; + } +} + +template +__device__ __forceinline__ void getTopK( + Pair topK[], real* val, int* col, int idx, int dim, int beamSize) { + while (idx < dim) { + if (topK[beamSize - 1] < val[idx]) { + Pair tmp(val[idx], col[idx]); + addTo(topK, tmp, beamSize); + } + idx += blockSize; + } +} + +template +__device__ __forceinline__ void getTopK(Pair topK[], + real* val, + int* col, + int idx, + int dim, + const Pair& max, + int beamSize) { + while (idx < dim) { + if (topK[beamSize - 1] < val[idx]) { + Pair tmp(val[idx], col[idx]); + if (tmp < max) { + addTo(topK, tmp, beamSize); + } + } + idx += blockSize; + } +} + +template +__device__ __forceinline__ void threadGetTopK(Pair topK[], + int& beam, + int beamSize, + real* src, + bool& firstStep, + bool& isEmpty, + Pair& max, + int dim, + const int tid) { + if (beam > 0) { + int length = beam < beamSize ? beam : beamSize; + if (firstStep) { + firstStep = false; + getTopK(topK, src, tid, dim, length); + } else { + for (int k = 0; k < maxLength; k++) { + if (k < maxLength - beam) { + topK[k] = topK[k + beam]; + } else { + topK[k].set(-HL_FLOAT_MAX, -1); + } + } + if (!isEmpty) { + getTopK(topK + maxLength - beam, src, tid, dim, max, length); + } + } + + max = topK[maxLength - 1]; + if (max.id_ == -1) isEmpty = true; + beam = 0; + } +} + +template +__device__ __forceinline__ void threadGetTopK(Pair topK[], + int& beam, + int beamSize, + real* val, + int* col, + bool& firstStep, + bool& isEmpty, + Pair& max, + int dim, + const int tid) { + if (beam > 0) { + int length = beam < beamSize ? beam : beamSize; + if (firstStep) { + firstStep = false; + getTopK(topK, val, col, tid, dim, length); + } else { + for (int k = 0; k < maxLength; k++) { + if (k < maxLength - beam) { + topK[k] = topK[k + beam]; + } else { + topK[k].set(-HL_FLOAT_MAX, -1); + } + } + if (!isEmpty) { + getTopK( + topK + maxLength - beam, val, col, tid, dim, max, length); + } + } + + max = topK[maxLength - 1]; + if (max.id_ == -1) isEmpty = true; + beam = 0; + } +} + +template +__device__ __forceinline__ void blockReduce(Pair* shTopK, + int* maxId, + Pair topK[], + real** topVal, + int** topIds, + int& beam, + int& beamSize, + const int tid, + const int warp) { + while (true) { + __syncthreads(); + if (tid < blockSize / 2) { + if (shTopK[tid] < shTopK[tid + blockSize / 2]) { + maxId[tid] = tid + blockSize / 2; + } else { + maxId[tid] = tid; + } + } + __syncthreads(); + for (int stride = blockSize / 4; stride > 0; stride = stride / 2) { + if (tid < stride) { + if (shTopK[maxId[tid]] < shTopK[maxId[tid + stride]]) { + maxId[tid] = maxId[tid + stride]; + } + } + __syncthreads(); + } + __syncthreads(); + + if (tid == 0) { + **topVal = shTopK[maxId[0]].v_; + **topIds = shTopK[maxId[0]].id_; + (*topVal)++; + (*topIds)++; + } + if (tid == maxId[0]) beam++; + if (--beamSize == 0) break; + __syncthreads(); + + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + + if (tid == maxId[0]) { + if (beam < maxLength) { + shTopK[tid] = topK[beam]; + } + } + if (maxId[0] / 32 == warp) { + if (__shfl_sync(mask, beam, (maxId[0]) % 32, 32) == maxLength) break; + } + } +} + +/** + * Each block compute one sample. + * In a block: + * 1. every thread get top maxLength value; + * 2. merge to shTopK, block reduce and get max value; + * 3. go to the second setp, until one thread's topK value is null; + * 4. go to the first setp, until get the topK value. + */ +template +__global__ void KeMatrixTopK(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, + int dim, + int beamSize) { + __shared__ Pair shTopK[blockSize]; + __shared__ int maxId[blockSize / 2]; + const int tid = threadIdx.x; + const int warp = threadIdx.x / 32; + src += blockIdx.x * lds; + topVal += blockIdx.x * ldv; + topIds += blockIdx.x * beamSize; + + Pair topK[maxLength]; // NOLINT + int beam = maxLength; + Pair max; + bool isEmpty = false; + bool firstStep = true; + + for (int k = 0; k < maxLength; k++) { + topK[k].set(-HL_FLOAT_MAX, -1); + } + while (beamSize) { + threadGetTopK( + topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); + + shTopK[tid] = topK[0]; + blockReduce( + shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); + } +} + +template +__global__ void KeSMatrixTopK(real* topVal, + int ldv, + int* topIds, + real* val, + int* row, + int* col, + int beamSize) { + __shared__ Pair shTopK[blockSize]; + __shared__ int maxId[blockSize / 2]; + const int tid = threadIdx.x; + const int warp = threadIdx.x / 32; + topVal += blockIdx.x * ldv; + topIds += blockIdx.x * beamSize; + + Pair topK[maxLength]; // NOLINT + int beam = maxLength; + Pair max; + bool isEmpty = false; + bool firstStep = true; + + int start = row[blockIdx.x]; + int end = row[blockIdx.x + 1]; + int dim = end - start; + val += start; + col += start; + + if (beamSize > dim) { + // if the number of values to sort are less than the output size, + // use -1 to indicate the end of valid sorted values. + if (tid == 0) { + topIds[dim] = -1; + } + + beamSize = dim; + } + + for (int k = 0; k < maxLength; k++) { + topK[k].set(-HL_FLOAT_MAX, -1); + } + while (beamSize) { + threadGetTopK( + topK, beam, beamSize, val, col, firstStep, isEmpty, max, dim, tid); + + shTopK[tid] = topK[0]; + blockReduce( + shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); + } +} + +void hl_matrix_top_k(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, + int dim, + int beamSize, + int numSamples) { + CHECK_NOTNULL(topVal); + CHECK_NOTNULL(topIds); + CHECK_NOTNULL(src); + + if (beamSize > dim) beamSize = dim; + + dim3 threads(256, 1); + dim3 grid(numSamples, 1); + KeMatrixTopK<5, 256><<>>( + topVal, ldv, topIds, src, lds, dim, beamSize); + + CHECK_SYNC("hl_matrix_top_k failed"); +} + +void hl_sparse_matrix_top_k(real* topVal, + int ldv, + int* topIds, + hl_sparse_matrix_s src, + int beamSize, + int numSamples) { + CHECK_NOTNULL(topVal); + CHECK_NOTNULL(topIds); + CHECK_NOTNULL(src); + CHECK_EQ(src->format, HL_SPARSE_CSR) << "sparse matrix format error!"; + + hl_csr_matrix csr = (hl_csr_matrix)src->matrix; + if (csr->csr_val == NULL || csr->csr_row == NULL || csr->csr_col == NULL) { + LOG(FATAL) << "parameter src is null!"; + } + + dim3 threads(256, 1); + dim3 grid(numSamples, 1); + KeSMatrixTopK<5, 256><<>>( + topVal, ldv, topIds, csr->csr_val, csr->csr_row, csr->csr_col, beamSize); + + CHECK_SYNC("hl_sparse_matrix_top_k failed"); +} + +/** + * Each block compute one sample. + * In a block: + * 1. every thread get top maxLength value; + * 2. merge to shTopK, block reduce and get max value; + * 3. go to the second setp, until one thread's topK value is null; + * 4. go to the first setp, until get the topK value. + */ +template +__global__ void KeMatrixTopKClassificationError(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, + int dim, + int beamSize, + int* label, + real* recResult) { + __shared__ Pair shTopK[blockSize]; + __shared__ int maxId[blockSize / 2]; + const int tid = threadIdx.x; + const int warp = threadIdx.x / 32; + src += blockIdx.x * lds; + topVal += blockIdx.x * ldv; + topIds += blockIdx.x * beamSize; + + Pair topK[maxLength]; // NOLINT + int beam = maxLength; + Pair max; + bool isEmpty = false; + bool firstStep = true; + int topkSize = beamSize; + + for (int k = 0; k < maxLength; k++) { + topK[k].set(-HL_FLOAT_MAX, -1); + } + + while (beamSize) { + threadGetTopK( + topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); + + shTopK[tid] = topK[0]; + blockReduce( + shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); + } + + __syncthreads(); + if (tid == 0) { + for (int i = 0; i < topkSize; i++) { + if (*--topIds == label[blockIdx.x]) { + recResult[blockIdx.x] = 0; + break; + } + recResult[blockIdx.x] = 1.0f; + } + } +} + +void hl_matrix_classification_error(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, + int dim, + int topkSize, + int numSamples, + int* label, + real* recResult) { + CHECK_NOTNULL(topVal); + CHECK_NOTNULL(topIds); + CHECK_NOTNULL(src); + + if (topkSize > dim) topkSize = dim; + + dim3 threads(256, 1); + dim3 grid(numSamples, 1); + KeMatrixTopKClassificationError<5, 256><<>>( + topVal, ldv, topIds, src, lds, dim, topkSize, label, recResult); + + CHECK_SYNC("hl_matrix_top_k classification error failed"); +} diff --git a/paddle/legacy/cuda/src/hl_warpctc_wrap.cc b/paddle/legacy/cuda/src/hl_warpctc_wrap.cc new file mode 100644 index 0000000000000000000000000000000000000000..31a8652f1f55387ae48cb516cd092442be784cbb --- /dev/null +++ b/paddle/legacy/cuda/src/hl_warpctc_wrap.cc @@ -0,0 +1,151 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "hl_warpctc_wrap.h" +#include +#include "paddle/legacy/utils/DynamicLoader.h" +#include "paddle/legacy/utils/Logging.h" + +namespace dynload { + +std::once_flag warpctc_dso_flag; +void* warpctc_dso_handle = nullptr; + +/** + * The following macro definition can generate structs + * (for each function) to dynamic load warpctc routine + * via operator overloading. When PADDLE_USE_DSO is + * false, you need to add the path of libwarp-ctc.so to + * the linked-libs of paddle or to LD_PRELOAD. + */ +#define DYNAMIC_LOAD_WARPCTC_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + using warpctcFunc = decltype(__name(args...)) (*)(Args...); \ + std::call_once( \ + warpctc_dso_flag, GetWarpCTCDsoHandle, &warpctc_dso_handle); \ + void* p_##_name = dlsym(warpctc_dso_handle, #__name); \ + return reinterpret_cast(p_##_name)(args...); \ + } \ + } __name; // struct DynLoad__##__name + +// include all needed warp-ctc functions +DYNAMIC_LOAD_WARPCTC_WRAP(get_warpctc_version) +DYNAMIC_LOAD_WARPCTC_WRAP(ctcGetStatusString) +DYNAMIC_LOAD_WARPCTC_WRAP(compute_ctc_loss) +DYNAMIC_LOAD_WARPCTC_WRAP(get_workspace_size) + +#undef DYNAMIC_LOAD_WARPCTC_WRAP + +} /* namespace dynload */ + +#define WARPCTC_GET_VERSION dynload::get_warpctc_version +#define WARPCTC_GET_STATUS_STRING dynload::ctcGetStatusString + +static int g_warpctcVersion = -1; +#ifndef PADDLE_TYPE_DOUBLE +#define WARPCTC_COMPUTE_LOSS dynload::compute_ctc_loss +#define WARPCTC_GET_WORKSPACE_SIZE dynload::get_workspace_size +#else +hl_warpctc_status_t fatal(...) { + LOG(FATAL) << "warp-ctc [version " << g_warpctcVersion + << "] Error: not support double precision."; + // both of get_warpctc_version() and get_workspace_size() return an ctcStatus + // type value + return CTC_STATUS_EXECUTION_FAILED; +} +#define WARPCTC_COMPUTE_LOSS fatal +#define WARPCTC_GET_WORKSPACE_SIZE fatal +#endif + +/** + * Check build-in warp-ctc function using glog and it also + * support << operator for more details error info. + */ +#define CHECK_WARPCTC(warpctcStat) \ + CHECK_EQ(CTC_STATUS_SUCCESS, warpctcStat) \ + << "warp-ctc [version " << g_warpctcVersion \ + << "] Error: " << WARPCTC_GET_STATUS_STRING(warpctcStat) << " " + +void hl_warpctc_init(const size_t blank, + bool useGpu, + hl_warpctc_options_t* options) { + CHECK_NOTNULL(options); + + g_warpctcVersion = WARPCTC_GET_VERSION(); + + if (useGpu) { +#ifdef __NVCC__ + options->loc = CTC_GPU; + options->stream = STREAM_DEFAULT; +#else + LOG(FATAL) << "[warpctc init] GPU is not enabled."; +#endif + } else { + options->loc = CTC_CPU; + options->num_threads = 1; + } + + options->blank_label = blank; +} + +void hl_warpctc_compute_loss(const real* batchInput, + real* batchGrad, + const int* cpuLabels, + const int* cpuLabelLengths, + const int* cpuInputLengths, + const size_t numClasses, + const size_t numSequences, + real* cpuCosts, + void* workspace, + hl_warpctc_options_t* options) { + CHECK_NOTNULL(batchInput); + CHECK_NOTNULL(cpuLabels); + CHECK_NOTNULL(cpuLabelLengths); + CHECK_NOTNULL(cpuInputLengths); + CHECK_NOTNULL(cpuCosts); + CHECK_NOTNULL(workspace); + CHECK_NOTNULL(options); + + CHECK_WARPCTC(WARPCTC_COMPUTE_LOSS(batchInput, + batchGrad, + cpuLabels, + cpuLabelLengths, + cpuInputLengths, + numClasses, + numSequences, + cpuCosts, + workspace, + *options)); +} + +void hl_warpctc_get_workspace_size(const int* cpuLabelLengths, + const int* cpuInputLengths, + const size_t numClasses, + const size_t numSequences, + hl_warpctc_options_t* options, + size_t* bytes) { + CHECK_NOTNULL(cpuLabelLengths); + CHECK_NOTNULL(cpuInputLengths); + CHECK_NOTNULL(options); + CHECK_NOTNULL(bytes); + + CHECK_WARPCTC(WARPCTC_GET_WORKSPACE_SIZE(cpuLabelLengths, + cpuInputLengths, + numClasses, + numSequences, + *options, + bytes)); +} diff --git a/paddle/function/BlockExpandOp.cpp b/paddle/legacy/function/BlockExpandOp.cpp similarity index 100% rename from paddle/function/BlockExpandOp.cpp rename to paddle/legacy/function/BlockExpandOp.cpp diff --git a/paddle/function/BlockExpandOpTest.cpp b/paddle/legacy/function/BlockExpandOpTest.cpp similarity index 100% rename from paddle/function/BlockExpandOpTest.cpp rename to paddle/legacy/function/BlockExpandOpTest.cpp diff --git a/paddle/legacy/function/BufferArg.cpp b/paddle/legacy/function/BufferArg.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1f3d505c31bf8d50503032a4baae6230b9f7241d --- /dev/null +++ b/paddle/legacy/function/BufferArg.cpp @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "BufferArg.h" +#include "paddle/legacy/math/SparseMatrix.h" + +namespace paddle { + +const SequenceArg& BufferArg::sequence() const { + CHECK_EQ(bufferType_, TENSOR_SEQUENCE_DATA); + return dynamic_cast(*this); +} + +const SparseMatrixArg& BufferArg::sparse() const { + CHECK_EQ(bufferType_, TENSOR_SPARSE); + return dynamic_cast(*this); +} + +SparseMatrixArg::SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType) + : BufferArg(sparse, argType), + row_(reinterpret_cast(sparse.getRows()), VALUE_TYPE_INT32), + col_(reinterpret_cast(sparse.getCols()), VALUE_TYPE_INT32), + nnz_(sparse.getElementCnt()), + format_(static_cast(sparse.getFormat())), + type_(static_cast(sparse.getValueType())) { + bufferType_ = TENSOR_SPARSE; +} + +SparseMatrixArg::SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType) + : BufferArg(sparse, argType), + row_(reinterpret_cast(sparse.getRows()), VALUE_TYPE_INT32), + col_(reinterpret_cast(sparse.getCols()), VALUE_TYPE_INT32), + nnz_(sparse.getElementCnt()), + format_(static_cast(sparse.getFormat())), + type_(static_cast(sparse.getValueType())) { + bufferType_ = TENSOR_SPARSE; +} + +} // namespace paddle diff --git a/paddle/legacy/function/BufferArg.h b/paddle/legacy/function/BufferArg.h new file mode 100644 index 0000000000000000000000000000000000000000..1f47ad556d29363d784fde718fdacdf0658ef010 --- /dev/null +++ b/paddle/legacy/function/BufferArg.h @@ -0,0 +1,364 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include "TensorShape.h" +#include "TensorType.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +enum BufferType { + TENSOR_UNKNOWN = 0, + TENSOR_NORMAL = 1, + TENSOR_SEQUENCE_ID = 2, + TENSOR_SEQUENCE_DATA = 3, + TENSOR_SPARSE = 4 +}; + +class BufferArg; +class SequenceArg; +class SparseMatrixArg; + +/** + * \brief BufferArg used as the argument type of Function. + * + * The arguments of the Paddle Function have four Buffer types. + * 1. BufferArg for a dense Buffer of any dimension. + * 2. SequenceIdArg for a Buffer of sequence start positions. + * 3. SequenceArg for a Buffer of sequence data. + * 4. SparseMatrixArg for a Buffer of sparse matrix. + * + * Buffer shape + * For most buffers, the first dimension `shape()[0]` represents + * the size of the mini-batch. + * + * Buffer argType + * There is an ArgType property for the BufferArg used as Function Output. + * Whether the result of the Function calculation is assigned to the + * output Buffer or added to the output Buffer is determined by the + * argType_ property of the output BufferArg. + */ + +// ArgType is only used by output BufferArg. +// For input argument, argType_ is ignored. +// For output argument, need to set the argType_ of the BufferArg. +enum ArgType { + UNSPECIFIED = 0, + ASSIGN_TO = 1, + ADD_TO = 2, +}; +class BufferArg { + public: + void setArgType(ArgType argType) { argType_ = argType; } + + ArgType getArgType() const { return argType_; } + + public: + BufferArg(ValueType valueType, + const TensorShape& shape, + ArgType argType = UNSPECIFIED) + : buf_(nullptr), valueType_(valueType), shape_(shape), argType_(argType) { + bufferType_ = TENSOR_NORMAL; + } + + BufferArg(void* buf, + ValueType valueType, + const TensorShape& shape, + ArgType argType = UNSPECIFIED) + : buf_(buf), valueType_(valueType), shape_(shape), argType_(argType) { + bufferType_ = TENSOR_NORMAL; + } + + BufferArg(void* buf, ValueType valueType) : buf_(buf), valueType_(valueType) { + bufferType_ = TENSOR_NORMAL; + } + + BufferArg(const Matrix& matrix, ArgType argType = UNSPECIFIED) + : buf_( + const_cast(reinterpret_cast(matrix.getData()))), + valueType_(DataType::value), + shape_(2), + argType_(argType) { + bufferType_ = TENSOR_NORMAL; + shape_.setDim(0, matrix.getHeight()); + shape_.setDim(1, matrix.getWidth()); + } + + BufferArg(const Matrix& matrix, + const TensorShape& shape, + ArgType argType = UNSPECIFIED) + : buf_( + const_cast(reinterpret_cast(matrix.getData()))), + valueType_(DataType::value), + shape_(shape), + argType_(argType) { + bufferType_ = TENSOR_NORMAL; + CHECK_EQ(matrix.getElementCnt(), shape.getElements()); + } + + BufferArg(const Vector& vector, ArgType argType = UNSPECIFIED) + : buf_( + const_cast(reinterpret_cast(vector.getData()))), + valueType_(DataType::value), + shape_(1), + argType_(argType) { + bufferType_ = TENSOR_NORMAL; + shape_.setDim(0, vector.getSize()); + } + + BufferArg(const IVector& vector, ArgType argType = UNSPECIFIED) + : buf_( + const_cast(reinterpret_cast(vector.getData()))), + valueType_(VALUE_TYPE_INT32), + shape_(1), + argType_(argType) { + bufferType_ = TENSOR_NORMAL; + shape_.setDim(0, vector.getSize()); + } + + template + typename Tensor::Matrix matrix() const { + CHECK(buf_); + CHECK(valueType_ == DataType::value); + // CHECK(deviceType_ == DType); + CHECK_EQ((size_t)2, shape_.ndims()); + return typename Tensor::Matrix( + reinterpret_cast(buf_), shape_[0], shape_[1]); + } + + template + typename Tensor::Vector vector() const { + CHECK(buf_); + CHECK(valueType_ == DataType::value); + // CHECK(deviceType_ == DType); + CHECK_EQ((size_t)1, shape_.ndims()); + return typename Tensor::Vector( + shape_[0], reinterpret_cast(buf_)); + } + + virtual ~BufferArg() {} + + template + T* data() const { + return reinterpret_cast(buf_); + } + + void* data() const { return buf_; } + ValueType valueType() const { return valueType_; } + BufferType bufferType() const { return bufferType_; } + const TensorShape& shape() const { return shape_; } + bool isSparseArg() const { return TENSOR_SPARSE == bufferType_; } + bool isSequenceArg() const { return TENSOR_SEQUENCE_DATA == bufferType_; } + virtual size_t numElements() const { return shape_.getElements(); } + + const SequenceArg& sequence() const; + const SparseMatrixArg& sparse() const; + + protected: + void* buf_; + ValueType valueType_; + TensorShape shape_; + BufferType bufferType_{TENSOR_UNKNOWN}; + ArgType argType_{UNSPECIFIED}; + // TODO(tianbing), add deviceType_ + // leading dimensions. The size is dims_.size() + // Dims lds_; +}; + +// sequence start positions in a mini-batch of sequences +// shape_.ndims() == 1 +// valueType_ = int32 +// if a < b then value_.buf_[a] < value_.buf_[b] +class SequenceIdArg : public BufferArg { + public: + SequenceIdArg(const TensorShape& shape, ArgType argType = UNSPECIFIED) + : BufferArg(VALUE_TYPE_INT32, shape, argType) { + bufferType_ = TENSOR_SEQUENCE_ID; + CHECK_EQ(shape_.ndims(), 1UL); + CHECK_GE(shape_[0], 1UL); + numSeqs_ = shape_[0] - 1; + } + + SequenceIdArg(void* buf, + const TensorShape& shape, + ArgType argType = UNSPECIFIED) + : BufferArg(buf, VALUE_TYPE_INT32, shape, argType) { + bufferType_ = TENSOR_SEQUENCE_ID; + CHECK_EQ(shape_.ndims(), 1UL); + numSeqs_ = shape_[0] - 1; + } + + SequenceIdArg(const IVector& vector) : BufferArg(vector) { + bufferType_ = TENSOR_SEQUENCE_ID; + numSeqs_ = shape_[0] - 1; + } + + ~SequenceIdArg() {} + + size_t numSeqs() const { return numSeqs_; } + + private: + size_t numSeqs_; +}; + +// sequences data +// For mini-batch calculate, +// one batch can contain more than one sequence of data. +// SequenceArg can be used to represent sequences that contain multiple +// unequal lengths. +class SequenceArg : public BufferArg { + public: + SequenceArg(ValueType valueType, + const TensorShape& shape, + ArgType argType = UNSPECIFIED) + : BufferArg(valueType, shape, argType), + startPositions_(TensorShape({shape[0]})) { + bufferType_ = TENSOR_SEQUENCE_DATA; + } + + SequenceArg(void* buf, + ValueType valueType, + const TensorShape& shape, + const SequenceIdArg& startPositions, + ArgType argType = UNSPECIFIED) + : BufferArg(buf, valueType, shape, argType), + startPositions_(startPositions) { + bufferType_ = TENSOR_SEQUENCE_DATA; + } + + SequenceArg(const Matrix& matrix, + const IVector& vector, + ArgType argType = UNSPECIFIED) + : BufferArg(matrix, argType), startPositions_(vector) { + bufferType_ = TENSOR_SEQUENCE_DATA; + } + + ~SequenceArg() {} + + void* getIdBuf() const { return startPositions_.data(); } + size_t numSeqs() const { return startPositions_.numSeqs(); } + SequenceIdArg& getSequenceId() { return startPositions_; } + const SequenceIdArg& getSequenceId() const { return startPositions_; } + + private: + SequenceIdArg startPositions_; +}; + +// sparse matrix +// valueType_ == float or double +// shape_.ndims() == 2 +class SparseMatrixArg : public BufferArg { + public: + SparseMatrixArg(void* buf, + ValueType valueType, + const TensorShape& shape, + const BufferArg& row, + const BufferArg& col, + size_t nnz, + SparseFormat format, + SparseValueType type, + ArgType argType = UNSPECIFIED) + : BufferArg(buf, valueType, shape, argType), + row_(row), + col_(col), + nnz_(nnz), + format_(static_cast(format)), + type_(static_cast(type)) { + bufferType_ = TENSOR_SPARSE; + CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE)); + CHECK_EQ(shape_.ndims(), 2UL); + CHECK_EQ(row_.shape().ndims(), 1UL); + CHECK_EQ(col_.shape().ndims(), 1UL); + if (format_ == T_SPARSE_CSR) { + CHECK_EQ(nnz, col.shape()[0]); + } else if (format_ == T_SPARSE_CSC) { + CHECK_EQ(nnz, row.shape()[0]); + } + } + + SparseMatrixArg(ValueType valueType, + const TensorShape& shape, + size_t nnz, + SparseFormat format, + SparseValueType type, + ArgType argType = UNSPECIFIED) + : BufferArg(valueType, shape, argType), + row_(BufferArg(nullptr, VALUE_TYPE_INT32)), + col_(BufferArg(nullptr, VALUE_TYPE_INT32)), + nnz_(nnz), + format_(static_cast(format)), + type_(static_cast(type)) { + bufferType_ = TENSOR_SPARSE; + CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE)); + CHECK_EQ(shape_.ndims(), 2UL); + + /// len of row_ : height + 1 (CSR) or nnz (CSC), buf_ == nullptr + row_ = (format_ == T_SPARSE_CSR + ? BufferArg(VALUE_TYPE_INT32, TensorShape{shape_[0] + 1}) + : BufferArg(VALUE_TYPE_INT32, TensorShape{nnz})); + /// len of col_ : width + 1 (CSC) or nnz (CSR), buf_ == nullptr + col_ = (format_ == T_SPARSE_CSR + ? BufferArg(VALUE_TYPE_INT32, TensorShape{nnz}) + : BufferArg(VALUE_TYPE_INT32, TensorShape{shape_[1] + 1})); + } + + SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType = UNSPECIFIED); + + SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType = UNSPECIFIED); + + template + typename Tensor::SparseMatrix SparseMatrix() const { + CHECK(buf_); + CHECK(valueType_ == DataType::value); + // CHECK(deviceType_ == DType); + CHECK_EQ(2UL, shape_.ndims()); + return typename Tensor::SparseMatrix( + reinterpret_cast(buf_), + reinterpret_cast(row_.data()), + reinterpret_cast(col_.data()), + shape_[0], + shape_[1], + nnz_, + static_cast(type_), + static_cast(format_), + false); + } + + ~SparseMatrixArg() {} + + void* getRowBuf() const { return row_.data(); } + + void* getColBuf() const { return col_.data(); } + + size_t nnz() const { return nnz_; } + + size_t numElements() const override { return nnz_; } + + SparseDataFormat dataFormat() const { return format_; } + + SparseDataType dataType() const { return type_; } + + private: + BufferArg row_; + BufferArg col_; + size_t nnz_; + SparseDataFormat format_; + SparseDataType type_; +}; + +} // namespace paddle diff --git a/paddle/legacy/function/BufferArgTest.cpp b/paddle/legacy/function/BufferArgTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1ec153bea89f25414b0df3088ab0c366c92ecbe0 --- /dev/null +++ b/paddle/legacy/function/BufferArgTest.cpp @@ -0,0 +1,38 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "BufferArg.h" +#include +#include "paddle/legacy/math/MemoryHandle.h" + +namespace paddle { + +TEST(BufferTest, BufferArg) { + TensorShape shape({8, 10}); + CpuMemoryHandle memory(shape.getElements() * + sizeOfValuType(VALUE_TYPE_FLOAT)); + BufferArg buffer(memory.getBuf(), VALUE_TYPE_FLOAT, shape); + EXPECT_EQ(buffer.data(), memory.getBuf()); +} + +TEST(BufferTest, SequenceIdArg) { + TensorShape shape({10}); + CpuMemoryHandle memory(shape.getElements() * + sizeOfValuType(VALUE_TYPE_INT32)); + SequenceIdArg buffer(memory.getBuf(), shape); + EXPECT_EQ(buffer.data(), memory.getBuf()); + EXPECT_EQ(buffer.numSeqs(), 9U); +} + +} // namespace paddle diff --git a/paddle/function/CMakeLists.txt b/paddle/legacy/function/CMakeLists.txt similarity index 100% rename from paddle/function/CMakeLists.txt rename to paddle/legacy/function/CMakeLists.txt diff --git a/paddle/legacy/function/ContextProjectionOp.cpp b/paddle/legacy/function/ContextProjectionOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..05a3f915862b6657fc0a4300cbbea36721219e10 --- /dev/null +++ b/paddle/legacy/function/ContextProjectionOp.cpp @@ -0,0 +1,412 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ContextProjectionOp.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { +/** + * Context Projection Forward with CPU Matrix Device. + * + */ +template <> +void ContextProjectionForward(CpuMatrix& out_mat, + const CpuMatrix& input_mat, + const CpuMatrix& weight_mat, + const CpuIVector& seq_vec, + size_t context_length, + int context_start, + size_t begin_pad) { + const int* starts = seq_vec.getData(); + const size_t num_sequences = seq_vec.getSize() - 1; + for (size_t i = 0; i < num_sequences; ++i) { + for (size_t j = 0; j < context_length; ++j) { + int begin = starts[i] + context_start + j; + int end = starts[i + 1] + context_start + j; + int dst_begin = starts[i]; + int dst_end = starts[i + 1]; + if (begin < starts[i]) { + int64_t pad_size = + std::min(starts[i] - begin, starts[i + 1] - starts[i]); + MatrixPtr mat = out_mat.subMatrix(starts[i], pad_size); + if (weight_mat) { + MatrixPtr sub = + const_cast(weight_mat).subMatrix(j, pad_size); + mat->addAtOffset(*sub, j * input_mat.getWidth()); + } + dst_begin = starts[i] + pad_size; + begin = starts[i]; + } + if (end > starts[i + 1]) { + int64_t pad_size = + std::min(end - starts[i + 1], starts[i + 1] - starts[i]); + MatrixPtr mat = out_mat.subMatrix(starts[i + 1] - pad_size, pad_size); + if (weight_mat) { + MatrixPtr sub = + const_cast(weight_mat) + .subMatrix(begin_pad + context_start + j - pad_size, + pad_size); + mat->addAtOffset(*sub, j * input_mat.getWidth()); + } + dst_end = starts[i + 1] - pad_size; + end = starts[i + 1]; + } + if (end <= begin) continue; + MatrixPtr src = + const_cast(input_mat).subMatrix(begin, end - begin); + MatrixPtr dst = out_mat.subMatrix(dst_begin, dst_end - dst_begin); + dst->addAtOffset(*src, j * input_mat.getWidth()); + } + } +} + +/** + * Paddle Function for Context Projection Forward. + * Calculate the output layer value sequence after context projection. + * + * What is Context Projection for a sequence? + * For example, assumed input (x) has 4 words and the dimension of each word + * representation is 2. If we use zero to pad instead of learned weight to pad, + * and the context_lenth is 3, the output (y) is: + * + * @code + * x = [a1, a2; + * b1, b2; + * c1, c2; + * d1, d2] + * y = [0, 0, a1, a2, b1, b2; + * a1, a2, b1, b2, c1, c2; + * b1, b2, c1, c2, d1, d2; + * c1, c2, d1, d2, 0, 0] + * @endcode + * + * \param outputs[0].matrix output layer value, n * (d * l) + * \param outputs[0].vector start position sequence, n * 1 + * \param inputs[0].matrix input layer value, n * d + * \param inputs[0].vector start position sequence, n * 1 + * \param inputs[1].matrix input layer weight, pad * d + */ +template +class ContextProjectionForwardFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + context_length_ = config.get("context_length"); + context_start_ = config.get("context_start"); + begin_pad_ = config.get("begin_pad"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK(1UL == inputs.size() || 2UL == inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) + << "SequenceArg required here"; + const auto val_seqs = dynamic_cast(inputs[0]); + auto out_seq = dynamic_cast(outputs[0]); + + CHECK(out_seq.data() && val_seqs.data() && val_seqs.getSequenceId().data()); + CHECK_EQ(out_seq.shape().ndims(), 2UL); + CHECK_EQ(val_seqs.shape().ndims(), 2UL); + /// dim of output = dim of input * context_length + CHECK_EQ(out_seq.shape()[1], val_seqs.shape()[1] * context_length_); + /// input and output has the same batch_size + CHECK_EQ(val_seqs.shape()[0], out_seq.shape()[0]); + if (2UL == inputs.size()) { + CHECK_EQ(inputs[1].shape().ndims(), 2UL); + /// dim of input == dim of weight + CHECK_EQ(val_seqs.shape()[1], inputs[1].shape()[1]); + } + + CHECK_EQ(out_seq.getArgType(), ADD_TO); + auto out_mat = out_seq.matrix(); + const auto in_mat = val_seqs.matrix(); + const auto w_mat = + (2UL == inputs.size() && inputs[1].data()) + ? inputs[1].matrix() + : typename Tensor::Matrix(nullptr, 0, 0); + const auto seq_vec = val_seqs.getSequenceId().vector(); + + ContextProjectionForward(out_mat, + in_mat, + w_mat, + seq_vec, + context_length_, + context_start_, + begin_pad_); + } + + private: + size_t context_length_; + int context_start_; + size_t begin_pad_; +}; + +/** + * Context Projection Backward with CPU Matrix Device. + * + */ +template <> +void ContextProjectionBackward(const CpuMatrix& out_grad_mat, + CpuMatrix& in_grad_mat, + CpuMatrix& w_grad_mat, + const CpuIVector& seq_vec, + size_t context_length, + int context_start, + size_t begin_pad, + bool is_padding, + size_t total_pad) { + size_t input_dim = in_grad_mat ? in_grad_mat.getWidth() + : w_grad_mat ? w_grad_mat.getWidth() : 0; + const int* starts = seq_vec.getData(); + size_t num_sequences = seq_vec.getSize() - 1; + for (size_t i = 0; i < num_sequences; ++i) { + for (size_t j = 0; j < context_length; ++j) { + int begin = starts[i] + context_start + j; + int end = starts[i + 1] + context_start + j; + int dst_begin = starts[i]; + int dst_end = starts[i + 1]; + if (begin < starts[i]) { + int64_t pad_size = + std::min(starts[i] - begin, starts[i + 1] - starts[i]); + if (is_padding && w_grad_mat) { + MatrixPtr mat = const_cast(out_grad_mat) + .subMatrix(starts[i], pad_size); + MatrixPtr sub = w_grad_mat.subMatrix(j, pad_size); + sub->addAtOffset(*mat, j * input_dim); + } + dst_begin = starts[i] + pad_size; + begin = starts[i]; + } + if (end > starts[i + 1]) { + int64_t pad_size = + std::min(end - starts[i + 1], starts[i + 1] - starts[i]); + if (is_padding && w_grad_mat) { + MatrixPtr mat = const_cast(out_grad_mat) + .subMatrix(starts[i + 1] - pad_size, pad_size); + MatrixPtr sub = w_grad_mat.subMatrix( + begin_pad + context_start + j - pad_size, pad_size); + sub->addAtOffset(*mat, j * input_dim); + } + dst_end = starts[i + 1] - pad_size; + end = starts[i + 1]; + } + if (end <= begin) continue; + if (!in_grad_mat) continue; + MatrixPtr src = in_grad_mat.subMatrix(begin, end - begin); + MatrixPtr dst = const_cast(out_grad_mat) + .subMatrix(dst_begin, dst_end - dst_begin); + src->addAtOffset(*dst, j * input_dim); + } + } +} + +/** + * Context Projection Backward Function. + * Update the weight gradient and input layer gradient with backprop + * + * \param inputs[0].matrix output layer grad, n * (d * l) + * \param inputs[0].vector start position sequence, n * 1 + * \param outputs[0].matrix input layer grad, n * d + * \param outputs[0].vector start position sequence, n * 1 + * \param outputs[1] weight grad, pad * d + */ +template +class ContextProjectionBackwardFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + context_length_ = config.get("context_length"); + context_start_ = config.get("context_start"); + begin_pad_ = config.get("begin_pad"); + is_padding_ = config.get("is_padding"); + total_pad_ = config.get("total_pad"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK(1UL == outputs.size() || 2UL == outputs.size()); + CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) + << "SequenceArg required here"; + const auto in_seq = dynamic_cast(inputs[0]); + auto out_seq = dynamic_cast(outputs[0]); + CHECK(in_seq.data() && in_seq.getSequenceId().data()); + CHECK_EQ(in_seq.shape().ndims(), 2UL); + CHECK_EQ(out_seq.shape().ndims(), 2UL); + CHECK_EQ(out_seq.getSequenceId().shape().ndims(), 1UL); + + /// input and output grad has the same batch_size + CHECK_EQ(out_seq.shape()[0], in_seq.shape()[0]); + /// dim of output grad = dim of input grad * context_length + CHECK_EQ(in_seq.shape()[1], out_seq.shape()[1] * context_length_); + CHECK_EQ(out_seq.getArgType(), ADD_TO); + + if (2UL == outputs.size()) { + CHECK_EQ(outputs[1].shape().ndims(), 2UL); + /// dim of input grad == dim of weight + CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]); + CHECK_EQ(outputs[1].getArgType(), ADD_TO); + } + + const auto seq_vec = in_seq.getSequenceId().vector(); + const auto out_grad_mat = in_seq.matrix(); + auto in_grad_mat = + !out_seq.data() ? typename Tensor::Matrix(nullptr, 0, 0) + : out_seq.matrix(); + auto w_grad_mat = + (2UL == outputs.size() && outputs[1].data()) + ? outputs[1].matrix() + : typename Tensor::Matrix(nullptr, 0, 0); + + ContextProjectionBackward(out_grad_mat, + in_grad_mat, + w_grad_mat, + seq_vec, + context_length_, + context_start_, + begin_pad_, + is_padding_, + total_pad_); + } + + private: + size_t context_length_; + int context_start_; + size_t begin_pad_; + bool is_padding_; + size_t total_pad_; +}; + +/** + * Context Projection Backward Data Function + * Update input layer grad + * input: sequence of output layer grad + * output: sequence of input layer grad + * + * \param outputs[0].matrix input layer grad, n * d + * \param outputs[0].vector start position sequence, n * 1 + * \param inputs[0].matrix output layer grad, n * (d * l) + * \param inputs[0].vector start positon sequence, n * 1 + */ +template +class ContextProjectionBackwardDataFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + context_length_ = config.get("context_length"); + context_start_ = config.get("context_start"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) + << "SequenceArg required here"; + const auto in_seq = dynamic_cast(inputs[0]); + const auto out_seq = dynamic_cast(outputs[0]); + + CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceId().data()); + CHECK_EQ(out_seq.shape().ndims(), 2UL); + CHECK_EQ(in_seq.shape().ndims(), 2UL); + CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL); + /// output layer grad dim == input layer grad dim * context_length_ + CHECK_EQ(in_seq.shape().ndims(), out_seq.shape().ndims() * context_length_); + /// input and output has the same batch_size + CHECK_EQ(in_seq.shape()[0], out_seq.shape()[0]); + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + + const auto out_grad_mat = in_seq.matrix(); + const auto seq_vec = in_seq.getSequenceId().vector(); + auto in_grad_mat = out_seq.matrix(); + + ContextProjectionBackwardData( + out_grad_mat, in_grad_mat, seq_vec, context_length_, context_start_); + } + + private: + size_t context_length_; + int context_start_; +}; + +/** + * Context Projection Backward Weight Function + * Update weight grad by backprop + * input: sequence of output layer grad + * output: weight grad + * + * \param outputs[0] weight grad, pad * d + * \param inputs[0].matrix output layer grad, n * (d * l) + * \param inputs[0].vecotr start positon sequence, n * 1 + */ +template +class ContextProjectionBackwardWeightFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + context_length_ = config.get("context_length"); + context_start_ = config.get("context_start"); + begin_pad_ = config.get("begin_pad"); + total_pad_ = config.get("total_pad"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK(inputs[0].isSequenceArg()) << "SequenceArg required here"; + const auto in_seq = dynamic_cast(inputs[0]); + CHECK(in_seq.data() && in_seq.getSequenceId().data() && outputs[0].data()); + CHECK_EQ(outputs[0].shape().ndims(), 2UL); + CHECK_EQ(in_seq.shape().ndims(), 2UL); + CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL); + CHECK_EQ(in_seq.shape()[0], outputs[0].shape()[0]); + /// output layer grad dim == weight dim * context_length_ + CHECK_EQ(in_seq.shape()[1], outputs[0].shape()[1] * context_length_); + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + + const auto seq_vec = in_seq.getSequenceId().vector(); + const auto out_grad_mat = in_seq.matrix(); + auto w_grad_mat = outputs[0].matrix(); + ContextProjectionBackwardWeight(out_grad_mat, + w_grad_mat, + seq_vec, + context_length_, + context_start_, + total_pad_, + begin_pad_); + } + + private: + size_t context_length_; + int context_start_; + size_t begin_pad_; + size_t total_pad_; +}; + +REGISTER_TYPED_FUNC(ContextProjectionForward, + CPU, + ContextProjectionForwardFunc); +REGISTER_TYPED_FUNC(ContextProjectionBackward, + CPU, + ContextProjectionBackwardFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(ContextProjectionForward, + GPU, + ContextProjectionForwardFunc); +REGISTER_TYPED_FUNC(ContextProjectionBackward, + GPU, + ContextProjectionBackwardFunc); +REGISTER_TYPED_FUNC(ContextProjectionBackwardData, + GPU, + ContextProjectionBackwardDataFunc); +REGISTER_TYPED_FUNC(ContextProjectionBackwardWeight, + GPU, + ContextProjectionBackwardWeightFunc); +#endif +} // namespace paddle diff --git a/paddle/function/ContextProjectionOp.h b/paddle/legacy/function/ContextProjectionOp.h similarity index 100% rename from paddle/function/ContextProjectionOp.h rename to paddle/legacy/function/ContextProjectionOp.h diff --git a/paddle/function/ContextProjectionOpGpu.cu b/paddle/legacy/function/ContextProjectionOpGpu.cu similarity index 100% rename from paddle/function/ContextProjectionOpGpu.cu rename to paddle/legacy/function/ContextProjectionOpGpu.cu diff --git a/paddle/legacy/function/ContextProjectionOpTest.cpp b/paddle/legacy/function/ContextProjectionOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3b0a34567fe17b466de6186e537243fe8166a77a --- /dev/null +++ b/paddle/legacy/function/ContextProjectionOpTest.cpp @@ -0,0 +1,114 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "FunctionTest.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT + +void testMatrixProjectionForward(int context_start, + size_t context_length, + bool is_padding, + size_t batch_size, + size_t input_dim) { + size_t pad = std::max(0, -context_start) + + std::max(0, (int)(context_start + context_length - 1)); + if (pad == 0) is_padding = false; + + CpuGpuFuncCompare test( + "ContextProjectionForward", + FuncConfig() + .set("context_length", context_length) + .set("context_start", context_start) + .set("begin_pad", (size_t)std::max(0, -context_start))); + + // prepare input arguments + test.addSequence(SequenceIdArg(TensorShape{batch_size})); + test.addInputs( + SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim})); + if (is_padding) { // weight + test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim})); + } + test.addOutputs( + SequenceArg(VALUE_TYPE_FLOAT, + TensorShape{batch_size, input_dim * context_length}), + ADD_TO); + + // run Function + test.run(); +} + +void testMatrixProjectionBackward(int context_start, + size_t context_length, + bool is_padding, + size_t batch_size, + size_t input_dim) { + size_t pad = std::max(0, -context_start) + + std::max(0, (int)(context_start + context_length - 1)); + if (pad == 0) is_padding = false; + + CpuGpuFuncCompare test( + "ContextProjectionBackward", + FuncConfig() + .set("context_length", context_length) + .set("context_start", context_start) + .set("begin_pad", (size_t)std::max(0, -context_start)) + .set("is_padding", is_padding) + .set("total_pad", pad)); + + // prepare input arguments + test.addSequence(SequenceIdArg(TensorShape{batch_size})); + test.addInputs(SequenceArg( + VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim * context_length})); + test.addOutputs( + SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim}), + ADD_TO); + if (is_padding) { // weight + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim}), + ADD_TO); + } + + // run Function + test.run(); +} + +TEST(ContextProjection, Projection) { + for (auto context_start : {-5, -3, -1, 0, 3}) { + for (auto context_length : {1, 2, 5, 7}) { + for (auto trainable_padding : {false, true}) { + for (auto batch_size : {1, 2, 5, 20, 100}) { + for (auto input_dim : {15, 32, 63, 128, 200}) { + VLOG(3) << " context_start=" << context_start + << " context_length=" << context_length + << " trainable_padding=" << trainable_padding + << " batch_size=" << batch_size + << " input_dim=" << input_dim; + testMatrixProjectionForward(context_start, + context_length, + trainable_padding, + batch_size, + input_dim); + testMatrixProjectionBackward(context_start, + context_length, + trainable_padding, + batch_size, + input_dim); + } + } + } + } + } +} diff --git a/paddle/function/ConvOp.h b/paddle/legacy/function/ConvOp.h similarity index 100% rename from paddle/function/ConvOp.h rename to paddle/legacy/function/ConvOp.h diff --git a/paddle/function/ConvOpTest.h b/paddle/legacy/function/ConvOpTest.h similarity index 100% rename from paddle/function/ConvOpTest.h rename to paddle/legacy/function/ConvOpTest.h diff --git a/paddle/legacy/function/CosSimOp.cpp b/paddle/legacy/function/CosSimOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d04f4396caade803aa846fa81388f95a194845e6 --- /dev/null +++ b/paddle/legacy/function/CosSimOp.cpp @@ -0,0 +1,240 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CosSimOp.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { +/** + * Cosine Similarity for CpuMatrix + * + * \param out_mat, output value, size: nSamples * 1. + * \param in1_mat, input value 1, size: nSamples * dim. + * \param in2_mat, input value 2, size: n2 * dim (n2 == 1 or n2 == nSamples). + * \param scale, default 1.0 + * + */ +template <> +void CosSimForward(CpuMatrix& out_mat, + const CpuMatrix& in1_mat, + const CpuMatrix& in2_mat, + real scale) { + CHECK(out_mat.getData() && in1_mat.getData() && in2_mat.getData()); + size_t num_samples = out_mat.getHeight(); + size_t dim = in1_mat.getWidth(); + /// column vector [nSamples, 1] + real* out = out_mat.getData(); + const real* x = in1_mat.getData(); + const real* y = in2_mat.getData(); + + /// in2 might only have one row or full rows + CHECK(in2_mat.getHeight() == 1LU || in2_mat.getHeight() == num_samples); + size_t inc = (in2_mat.getHeight() == 1LU) ? 0 : dim; + for (size_t i = 0; i < num_samples; ++i, x += dim, y += inc) { + real square_sum_x = 0; + real square_sum_y = 0; + real xy = 0; + for (size_t j = 0; j < dim; ++j) { + square_sum_x += x[j] * x[j]; + square_sum_y += y[j] * y[j]; + xy += x[j] * y[j]; + } + CHECK(square_sum_x > 0 && square_sum_y > 0); + out[i] = scale * xy / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y)); + } +} + +/** + * Cosine Similarity + * for each row i, + * out[i] = scale * cos(input1[i], input2[i]) + * = scale * /sqrt(|input1[i]|^2 * |input2[i]|^2) + * when input2 only has one row, then for each row i, + * out[i] = cos(input1[i], input2[0]) + * + * \param inputs[0] input matrix 1, size: nSamples * dim. + * \param inputs[1] input matrix 2, size: n2 * dim (n2 == 1 or n2 == nSamples). + * \param outputs[0] output matrix, size : nSamples * 1. + */ + +template +class CosSimForwardFunc : public FunctionBase { + void init(const FuncConfig& config) override { + scale_ = config.get("scale"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(inputs.size(), 2UL); + CHECK_EQ(outputs.size(), 1UL); + + CHECK_EQ(inputs[0].shape().ndims(), 2UL); + CHECK_EQ(inputs[1].shape().ndims(), 2UL); + CHECK_EQ(outputs[0].shape().ndims(), 2UL); + + CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]); + CHECK_EQ(inputs[0].shape()[1], inputs[1].shape()[1]); + CHECK_EQ(outputs[0].shape()[1], 1UL); + + CHECK(outputs[0].data() && inputs[0].data() && inputs[1].data()); + + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + auto out_mat = outputs[0].matrix(); + const auto in1_mat = inputs[0].matrix(); + const auto in2_mat = inputs[1].matrix(); + + CosSimForward(out_mat, in1_mat, in2_mat, scale_); + } + + private: + real scale_; +}; + +/** + * Cosine Similarity Derivative for CpuMatrix + * + * \param in1_grad forward input grad 1, size: nSamples * dim. + * \param in2_grad forward input grad 2, + * size: n2 * dim (n2 == 1 or n2 == nSamples). + * + * \param out_grad backward loss output grad, size : nSamples * 1. + * \param out_val forward output value, size: nSamples * 1. + * \param in1_val forward input value 1, size: nSamples * dim. + * \param in2_val forward input value 2, + * size: n2 * dim (n2 == 1 or n2 == nSamples). + * \param scale, default 1.0 + */ +template <> +void CosSimBackward(const CpuMatrix& out_grad, + const CpuMatrix& out_val, + const CpuMatrix& in1_val, + const CpuMatrix& in2_val, + CpuMatrix& in1_grad, + CpuMatrix& in2_grad, + real scale) { + CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() && + in2_val.getData() && in1_grad.getData() && in2_grad.getData()); + CHECK_EQ(out_val.useGpu_, false) << "Matrix type are GPU, CPU required"; + + const real* grad = out_grad.getData(); + const real* out = out_val.getData(); + const real* prev_out_x = in1_val.getData(); + const real* prev_out_y = in2_val.getData(); + real* prev_grad_x = in1_grad.getData(); + real* prev_grad_y = in2_grad.getData(); + + size_t num_samples = out_grad.getHeight(); + size_t dim = in1_val.getWidth(); + CHECK_EQ(in2_val.getHeight(), in2_grad.getHeight()); + CHECK(in2_val.getHeight() == 1LU || in2_val.getHeight() == num_samples); + size_t inc = (in2_val.getHeight() == 1LU) ? 0 : dim; + for (size_t i = 0; i < num_samples; ++i, + prev_out_x += dim, + prev_out_y += inc, + prev_grad_x += dim, + prev_grad_y += inc) { + real square_sum_x = 0; + real square_sum_y = 0; + real xy = 0; + for (size_t j = 0; j < dim; ++j) { + square_sum_x += prev_out_x[j] * prev_out_x[j]; + square_sum_y += prev_out_y[j] * prev_out_y[j]; + xy += prev_out_x[j] * prev_out_y[j]; + } + CHECK(square_sum_x > 0 && square_sum_y > 0); + if (xy == 0) { + real reciprocal = + 1.0f / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y)); + for (size_t j = 0; j < dim; ++j) { + prev_grad_x[j] += scale * grad[i] * prev_out_y[j] * reciprocal; + prev_grad_y[j] += scale * grad[i] * prev_out_x[j] * reciprocal; + } + } else { + real reciprocal_xy = 1.0f / xy; + real reciprocal_square_sum_x = 1.0f / square_sum_x; + real reciprocal_square_sum_y = 1.0f / square_sum_y; + for (size_t j = 0; j < dim; ++j) { + prev_grad_x[j] += + out[i] * grad[i] * (prev_out_y[j] * reciprocal_xy - + prev_out_x[j] * reciprocal_square_sum_x); + prev_grad_y[j] += + out[i] * grad[i] * (prev_out_x[j] * reciprocal_xy - + prev_out_y[j] * reciprocal_square_sum_y); + } + } + } +} + +/** + * Cosine Similarity backward Derivative + * + * \param outputs[0] forward input grad 1, size: nSamples * dim. + * \param outputs[1] forward input grad 2, + * size: n2 * dim (n2 == 1 or n2 == nSamples). + * + * \param inputs[0] backward loss output grad, size : nSamples * 1. + * \param inputs[1] forward output value, size: nSamples * 1. + * \param inputs[2] forward input value 1, size: nSamples * dim. + * \param inputs[3] forward input value 2, + * size: n2 * dim (n2 == 1 or n2 == nSamples). + */ +template +class CosSimBackwardFunc : public FunctionBase { + void init(const FuncConfig& config) override { + scale_ = config.get("scale"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(inputs.size(), 4UL); + CHECK_EQ(outputs.size(), 2UL); + /// dim of out_grad and out_val == 1, column vector + CHECK_EQ(inputs[0].shape()[1], 1UL); + CHECK_EQ(inputs[1].shape()[1], 1UL); + /// nSamples of out_grad == out_val == in_val1 == in_grad1 + CHECK_EQ(inputs[1].shape()[0], inputs[0].shape()[0]); + CHECK_EQ(inputs[0].shape()[0], inputs[0].shape()[0]); + CHECK_EQ(outputs[0].shape()[0], inputs[0].shape()[0]); + /// dim of in1_val1 == in_val2 == in_grad1 == in_grad2 + CHECK_EQ(inputs[3].shape()[1], inputs[2].shape()[1]); + CHECK_EQ(outputs[0].shape()[1], inputs[2].shape()[1]); + CHECK_EQ(outputs[1].shape()[1], inputs[2].shape()[1]); + + CHECK(inputs[0].data() && inputs[1].data() && inputs[2].data() && + inputs[3].data() && outputs[0].data() && outputs[1].data()); + + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + CHECK_EQ(outputs[1].getArgType(), ADD_TO); + + const auto out_grad = inputs[0].matrix(); + const auto out_val = inputs[1].matrix(); + const auto in1_val = inputs[2].matrix(); + const auto in2_val = inputs[3].matrix(); + auto in1_grad = outputs[0].matrix(); + auto in2_grad = outputs[1].matrix(); + + CosSimBackward( + out_grad, out_val, in1_val, in2_val, in1_grad, in2_grad, scale_); + } + + private: + real scale_; +}; + +REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc); +REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc); +REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc); +#endif +} // namespace paddle diff --git a/paddle/function/CosSimOp.h b/paddle/legacy/function/CosSimOp.h similarity index 100% rename from paddle/function/CosSimOp.h rename to paddle/legacy/function/CosSimOp.h diff --git a/paddle/function/CosSimOpGpu.cu b/paddle/legacy/function/CosSimOpGpu.cu similarity index 100% rename from paddle/function/CosSimOpGpu.cu rename to paddle/legacy/function/CosSimOpGpu.cu diff --git a/paddle/legacy/function/CosSimOpTest.cpp b/paddle/legacy/function/CosSimOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..31bb43e1baa9a6d890d1b8fe2abf15a07a7094c6 --- /dev/null +++ b/paddle/legacy/function/CosSimOpTest.cpp @@ -0,0 +1,64 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "FunctionTest.h" +#include "paddle/legacy/math/Matrix.h" + +using namespace paddle; // NOLINT + +void testCosSimForward(size_t height_x, + size_t height_y, + size_t width, + real scale) { + CpuGpuFuncCompare test("CosSimForward", FuncConfig().set("scale", scale)); + // prepare input arguments + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width})); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width})); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}), + ASSIGN_TO); + // run Function + test.run(); +} + +void testCosSimBackward(size_t height_x, + size_t height_y, + size_t width, + real scale) { + CpuGpuFuncCompare test("CosSimBackward", FuncConfig().set("scale", scale)); + // prepare input arguments + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1})); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1})); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width})); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width})); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}), + ADD_TO); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}), + ADD_TO); + // run Function + test.run(); +} + +TEST(Matrix, cosSim) { + for (auto height_x : {10, 100, 1000}) { + for (auto height_y : {1, height_x}) { + for (auto width : {10, 100, 1000}) { + for (auto scale : {1.0, 2.0}) { + testCosSimForward(height_x, height_y, width, scale); + testCosSimBackward(height_x, height_y, width, scale); + } + } + } + } +} diff --git a/paddle/legacy/function/CropOp.cpp b/paddle/legacy/function/CropOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e22678822f06a323d1e6c17dce63d44d143484a3 --- /dev/null +++ b/paddle/legacy/function/CropOp.cpp @@ -0,0 +1,177 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CropOp.h" +#include "paddle/legacy/function/TensorShape.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { + +template <> +void Crop(real* outputs, + const real* inputs, + const TensorShape inShape, + const TensorShape outShape, + const FuncConfig& conf) { + std::vector crop_corner = + conf.get>("crop_corner"); + int cCrop = crop_corner[1]; + int hCrop = crop_corner[2]; + int wCrop = crop_corner[3]; + + int num = inShape[0]; + int inC = inShape[1]; + int inH = inShape[2]; + int inW = inShape[3]; + + int outC = outShape[1]; + int outH = outShape[2]; + int outW = outShape[3]; + + for (int n = 0; n < num; n++) { + for (int c = 0; c < outC; c++) { + for (int h = 0; h < outH; h++) { + int outoff = ((n * outC + c) * outH + h) * outW; + int inoff = ((n * inC + c + cCrop) * inH + h + hCrop) * inW + wCrop; + memcpy(outputs + outoff, inputs + inoff, outW * sizeof(real)); + } + } + } +} + +template <> +void CropGrad(const real* inGrad, + real* outGrad, + const TensorShape inShape, + const TensorShape outShape, + const FuncConfig& conf) { + std::vector crop_corner = + conf.get>("crop_corner"); + int cCrop = crop_corner[1]; + int hCrop = crop_corner[2]; + int wCrop = crop_corner[3]; + + int num = outShape[0]; + int outC = outShape[1]; + int outH = outShape[2]; + int outW = outShape[3]; + + int inC = inShape[1]; + int inH = inShape[2]; + int inW = inShape[3]; + + for (int n = 0; n < num; n++) { + for (int c = 0; c < inC; c++) { + for (int h = 0; h < inH; h++) { + int outoff = ((n * outC + c + cCrop) * outH + h + hCrop) * outW + wCrop; + int inoff = ((n * inC + c) * inH + h) * inW; + CpuVector inG = CpuVector(inW, const_cast(inGrad + inoff)); + CpuVector outG = CpuVector(inW, outGrad + outoff); + outG += inG; + } + } + } +} + +/** + * \brief Crop input according to the specify corner and shape. + * The input and output is a 4D tensor. In CropFunc, we only + * crop the 2nd to 4th dimension. + * + * Argument in this Function: + * \param pad_ A struct object contains the cropping corner and shape. + * \param inputs A 4D tensor, only one input. + * \param outputs A 4D tensor, the output value after cropping. + * + * For example, + * Input(2,2,2,3) = [ + * [ [[1,2,3], [3,4,5]], + * [[2,3,5], [1,6,7]] ], + * [ [[4,3,1], [1,8,7]], + * [[3,8,9], [2,3,5]] ] + * ] # the input shape is (2,2,2,3) + * + * pad_: if corner = (0,1,1) and crop_shape = (2,1,2) + * Output(2,2,1,2) = [ + * [ [[4,5]], + * [[6,7]] ], + * [ [[8,7]], + * [[3,5]] ] + * ] # the input shape is (2,2,2,3) + */ +template +class CropFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { conf_ = config; } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + + TensorShape inShape = inputs[0].shape(); + TensorShape outShape = outputs[0].shape(); + + Crop(outputs[0].data(), + inputs[0].data(), + inShape, + outShape, + conf_); + } + + private: + FuncConfig conf_; +}; + +/** + * \brief The backward propagation of cropping Function. + * + * Argument in this Function: + * \param crop_ The same meaning as it in CropFunc. + * \param inputs The gradient with respect to the output value of CropFunc. + * \param outputs The gradient with respect to the input value of CropFunc. + */ + +template +class CropGradFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { conf_ = config; } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + + TensorShape outShape = outputs[0].shape(); + TensorShape inShape = inputs[0].shape(); + + CropGrad(inputs[0].data(), + outputs[0].data(), + inShape, + outShape, + conf_); + } + + private: + FuncConfig conf_; +}; + +REGISTER_TYPED_FUNC(Crop, CPU, CropFunc); +REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(Crop, GPU, CropFunc); +REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/CropOp.h b/paddle/legacy/function/CropOp.h similarity index 100% rename from paddle/function/CropOp.h rename to paddle/legacy/function/CropOp.h diff --git a/paddle/function/CropOpGpu.cu b/paddle/legacy/function/CropOpGpu.cu similarity index 100% rename from paddle/function/CropOpGpu.cu rename to paddle/legacy/function/CropOpGpu.cu diff --git a/paddle/function/CropOpTest.cpp b/paddle/legacy/function/CropOpTest.cpp similarity index 100% rename from paddle/function/CropOpTest.cpp rename to paddle/legacy/function/CropOpTest.cpp diff --git a/paddle/legacy/function/CrossMapNormalOp.cpp b/paddle/legacy/function/CrossMapNormalOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f28703af00fa4bd7bebd98839cb077798083b61f --- /dev/null +++ b/paddle/legacy/function/CrossMapNormalOp.cpp @@ -0,0 +1,344 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CrossMapNormalOp.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { + +template <> +void CrossMapNormal(real* outputs, + real* denoms, + const real* inputs, + size_t numSamples, + size_t channels, + size_t height, + size_t width, + size_t size, + real scale, + real pow) { + size_t oneImage = height * width; + size_t oneSample = channels * oneImage; + + CpuVector outputsV(numSamples * oneSample, outputs); + CpuVector inputsV(numSamples * oneSample, const_cast(inputs)); + CpuVector denomsV(numSamples * oneSample, denoms); + + // f(x) = x * ( 1 + scale * SUM((x)^2) )^(-pow) + // x represents inputs + // f(x) represents outputs + // denoms save the intermediate result for backward + denomsV = denomsV.constant(1.0); + const int start = -((int)size - 1) / 2; + const int end = (int)size + start; + for (size_t i = 0; i < numSamples; i++) { + real* oneDenom = denoms + i * oneSample; + real* oneInput = const_cast(inputs) + i * oneSample; + for (int c = 0; c < (int)channels; c++) { + CpuVector denom(oneImage, oneDenom + c * oneImage); + for (int s = start; s < end; s++) { + if (c + s >= 0 && c + s < (int)channels) { + CpuVector input(oneImage, oneInput + (c + s) * oneImage); + denom += input.square() * scale; + } + } + } + } + + outputsV = inputsV * denomsV.pow(-pow); +} + +template <> +void CrossMapNormalGrad(real* inputsGrad, + const real* inputsValue, + const real* outputsValue, + const real* outputsGrad, + const real* denoms, + size_t numSamples, + size_t channels, + size_t height, + size_t width, + size_t size, + real scale, + real pow) { + size_t oneSample = channels * height * width; + std::function oneImage = [=](real* data, + size_t offset) { + return CpuVector(height * width, data + offset); + }; + + const int start = -((int)size) / 2; + const int end = (int)size + start; + const real ratio = -(real)2 * scale * pow; + for (size_t i = 0; i < numSamples; i++) { + size_t sOffset = i * oneSample; + real* oneInputGrad = inputsGrad + sOffset; + real* oneInputValue = const_cast(inputsValue) + sOffset; + real* oneDenom = const_cast(denoms) + sOffset; + real* oneOutputGrad = const_cast(outputsGrad) + sOffset; + real* oneOutputValue = const_cast(outputsValue) + sOffset; + + for (int c = 0; c < (int)channels; c++) { + size_t cOffset = c * height * width; + CpuVector inputGrad = oneImage(oneInputGrad, cOffset); + CpuVector inputValue = oneImage(oneInputValue, cOffset); + CpuVector denom = oneImage(oneDenom, cOffset); + CpuVector outputGrad = oneImage(oneOutputGrad, cOffset); + + inputGrad = inputGrad + denom.pow(-pow) * outputGrad; + for (int s = start; s < end; s++) { + if (c + s >= 0 && c + s < (int)channels) { + size_t offset = (c + s) * height * width; + CpuVector output = oneImage(oneOutputValue, offset); + CpuVector outputGrad = oneImage(oneOutputGrad, offset); + CpuVector denom = oneImage(oneDenom, offset); + + inputGrad += ((outputGrad * output * ratio) / denom) * inputValue; + } + } + } + } +} + +/** + * \brief Normalization with across maps. + * + * This Function comes from the paper + * "ImageNet Classification with Deep Convolutional Neural Networks". + * + * The original formula is: + * + * Input(i, x, y) + * Output(i, x, y) = ---------------------------------------------- + * -- upper + * (k + alpha * > (Input(j, x, y))^2) ^ (beta) + * -- j = lower + * + * upper is `min(C, c + N/2)` + * lower if `max(0, c - N/2)` + * + * Function implementation: + * + * inputs and outpus is NCHW format, while input.shape.ndims() is equal 4. + * And the meaning of each dimension(0-3) is respectively batch size, + * feature maps, rows and columns. + * + * Input and Output in the above formula is for each map(i) of one image, and + * Input(i, x, y), Output(i, x, y) represents an element in an image. + * + * C is the number of feature maps of one image, and N is a hyper-parameters + * is configured when Function is initialized. The sum in the denominator + * is the sum of the same position in the neighboring maps. + * + * In the implementation of Function, k is equal to 1, + * so Function has no argument for k. + * + * Function Arguments: + * + * \param size_ represent N + * \param scale_ represent alpha + * \param pow_ represent beta + * \param inputs[0] represent Input + * \param outputs[0] represent Output + * \param outputs[1] represent The denominator in the formula(except beta) + * + * Note: + * Save output[1] is to simplify the backward calculation. + * TODO, if only consider the forward calculation, we can optimize to + * remove the output[1]. + */ +template +class CrossMapNormalFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + // function arguments + size_ = config.get("size"); + scale_ = config.get("scale"); + pow_ = config.get("pow"); + + // number of inputs and outputs + numInputs_ = 1; + numOutputs_ = 2; + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + check(inputs, outputs); + // ArgType check still on here, + // not sure whether it is better to put inside the check. + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + CHECK_EQ(outputs[1].getArgType(), ASSIGN_TO); + size_t batchSize = inputs[0].shape()[0]; + size_t maps = inputs[0].shape()[1]; + size_t rows = inputs[0].shape()[2]; + size_t columns = inputs[0].shape()[3]; + + CrossMapNormal(outputs[0].data(), + outputs[1].data(), + inputs[0].data(), + batchSize, + maps, + rows, + columns, + size_, + scale_, + pow_); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + + CHECK_EQ(inputs[0].shape().ndims(), (size_t)4); + CHECK(inputs[0].shape() == outputs[0].shape()); + CHECK(inputs[0].shape() == outputs[1].shape()); + } + + // Only need the shape of the input, can calculate the + // floating-point operation. + size_t ops(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ((size_t)numInputs_, inputs.size()); + size_t batchSize = inputs[0].shape()[0]; + size_t maps = inputs[0].shape()[1]; + size_t rows = inputs[0].shape()[2]; + size_t columns = inputs[0].shape()[3]; + + // number of floating-point operations + // an approximate value + size_t ops = batchSize * maps * rows * columns * (size_ * 2 + 3); + + return ops; + } + + private: + size_t size_; + real scale_; + real pow_; +}; + +/** + * \brief Backward calculation for normalization with across maps. + * + * Function implementation: + * + * The implementation of this Function is derived from the + * CrossMapNormalFunc implementation. + * + * InputGrad = OutputGrad * denoms ^ (-beta) + * -- upper + * + > (OutputGrad * OutputValue * (-2 * alpha * beta) / denoms) * InputValue + * -- lower + * + * The data of inputs/outputs format is the same as the forward interface + * and is NCHW. + * + * The upper and lower is the same as forward. The logic of the sum + * is also the same as forward. + * + * Function Arguments: + * + * \param size_ represent N + * \param scale_ represent alpha + * \param pow_ represent beta + * \param inputs[0] represent InputValue, inputs[0] of CrossMapNormalFunc + * \param inputs[1] represent OutputValue, outputs[0] of CrossMapNormalFunc + * \param inputs[2] represent OutputGrad + * \param inputs[3] represent denoms, outputs[1] of CrossMapNormalFunc + * This is the intermediate result that is + * preserved in the forward calculation. + * \param outputs[0] represent InputGrad + */ +template +class CrossMapNormalGradFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + // function arguments + size_ = config.get("size"); + scale_ = config.get("scale"); + pow_ = config.get("pow"); + + // number of inputs and outputs + numInputs_ = 4; + numOutputs_ = 1; + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + check(inputs, outputs); + if (outputs[0].getArgType() != ADD_TO) { + // Currently, some algorithm implementations are ASSIGN_TO mode, + // if need to support the ADD_TO calculation, need to clear the output. + typename Tensor::Vector tmp( + outputs[0].shape().getElements(), outputs[0].data()); + tmp.zero(); + } + + size_t batchSize = inputs[0].shape()[0]; + size_t maps = inputs[0].shape()[1]; + size_t rows = inputs[0].shape()[2]; + size_t columns = inputs[0].shape()[3]; + + CrossMapNormalGrad(outputs[0].data(), + inputs[0].data(), + inputs[1].data(), + inputs[2].data(), + inputs[3].data(), + batchSize, + maps, + rows, + columns, + size_, + scale_, + pow_); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + + CHECK_EQ(inputs[0].shape().ndims(), (size_t)4); + CHECK(inputs[0].shape() == inputs[1].shape()); + CHECK(inputs[0].shape() == inputs[2].shape()); + CHECK(inputs[0].shape() == inputs[3].shape()); + CHECK(inputs[0].shape() == outputs[0].shape()); + } + + // Only need the shape of one input, can calculate the + // floating-point operation. + size_t ops(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_LT((size_t)1, inputs.size()); + size_t batchSize = inputs[0].shape()[0]; + size_t maps = inputs[0].shape()[1]; + size_t rows = inputs[0].shape()[2]; + size_t columns = inputs[0].shape()[3]; + + // number of floating-point operations + // an approximate value + size_t ops = batchSize * maps * rows * columns * (size_ * 4 + 2); + + return ops; + } + + private: + size_t size_; + real scale_; + real pow_; +}; + +REGISTER_TYPED_FUNC(CrossMapNormal, CPU, CrossMapNormalFunc); +REGISTER_TYPED_FUNC(CrossMapNormalGrad, CPU, CrossMapNormalGradFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(CrossMapNormal, GPU, CrossMapNormalFunc); +REGISTER_TYPED_FUNC(CrossMapNormalGrad, GPU, CrossMapNormalGradFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/CrossMapNormalOp.h b/paddle/legacy/function/CrossMapNormalOp.h similarity index 100% rename from paddle/function/CrossMapNormalOp.h rename to paddle/legacy/function/CrossMapNormalOp.h diff --git a/paddle/function/CrossMapNormalOpGpu.cu b/paddle/legacy/function/CrossMapNormalOpGpu.cu similarity index 100% rename from paddle/function/CrossMapNormalOpGpu.cu rename to paddle/legacy/function/CrossMapNormalOpGpu.cu diff --git a/paddle/function/CrossMapNormalOpTest.cpp b/paddle/legacy/function/CrossMapNormalOpTest.cpp similarity index 100% rename from paddle/function/CrossMapNormalOpTest.cpp rename to paddle/legacy/function/CrossMapNormalOpTest.cpp diff --git a/paddle/function/DepthwiseConvOp.cpp b/paddle/legacy/function/DepthwiseConvOp.cpp similarity index 100% rename from paddle/function/DepthwiseConvOp.cpp rename to paddle/legacy/function/DepthwiseConvOp.cpp diff --git a/paddle/function/DepthwiseConvOp.h b/paddle/legacy/function/DepthwiseConvOp.h similarity index 100% rename from paddle/function/DepthwiseConvOp.h rename to paddle/legacy/function/DepthwiseConvOp.h diff --git a/paddle/legacy/function/DepthwiseConvOpGpu.cu b/paddle/legacy/function/DepthwiseConvOpGpu.cu new file mode 100644 index 0000000000000000000000000000000000000000..17138cc56390d0fcfb15d4b77a56eda466bcfd3c --- /dev/null +++ b/paddle/legacy/function/DepthwiseConvOpGpu.cu @@ -0,0 +1,376 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "DepthwiseConvOp.h" +#include "paddle/legacy/math/BaseMatrix.h" + +namespace paddle { + +// CUDA kernel to compute the depthwise convolution forward pass +template +__global__ void ConvolutionDepthwiseForward(const int nthreads, + const T* const inputData, + const T* const filterData, + const int batchSize, + const int outputChannels, + const int outputHeight, + const int outputWidth, + const int inputChannels, + const int inputHeight, + const int inputWidth, + const int filterMultiplier, + const int filterHeight, + const int filterWidth, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + T* const outputData) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + + if (index < nthreads) { + const int batch = index / outputChannels / outputHeight / outputWidth; + const int c_out = (index / outputHeight / outputWidth) % outputChannels; + const int h_out = (index / outputWidth) % outputHeight; + const int w_out = index % outputWidth; + + const int c_in = c_out / filterMultiplier; + const T* weight = filterData + c_out * filterHeight * filterWidth; + T value = 0; + const int h_in_start = -paddingH + h_out * strideH; + const int w_in_start = -paddingW + w_out * strideW; + const int h_in_end = -paddingH + h_out * strideH + filterHeight - 1; + const int w_in_end = -paddingW + w_out * strideW + filterWidth - 1; + if ((h_in_start >= 0) && (h_in_end < inputHeight) && (w_in_start >= 0) && + (w_in_end < inputWidth)) { + for (int kh = 0; kh < filterHeight; ++kh) { + for (int kw = 0; kw < filterWidth; ++kw) { + const int h_in = -paddingH + h_out * strideH + kh; + const int w_in = -paddingW + w_out * strideW + kw; + const int offset = + ((batch * inputChannels + c_in) * inputHeight + h_in) * + inputWidth + + w_in; + value += (*weight) * inputData[offset]; + ++weight; + } + } + } else { + for (int kh = 0; kh < filterHeight; ++kh) { + for (int kw = 0; kw < filterWidth; ++kw) { + const int h_in = -paddingH + h_out * strideH + kh; + const int w_in = -paddingW + w_out * strideW + kw; + if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) && + (w_in < inputWidth)) { + const int offset = + ((batch * inputChannels + c_in) * inputHeight + h_in) * + inputWidth + + w_in; + value += (*weight) * inputData[offset]; + } + ++weight; + } + } + } + outputData[index] = value; + } +} + +// CUDA kernel to compute the depthwise convolution backprop w.r.t input. +template +__global__ void ConvolutionDepthwiseInputBackward(const int nthreads, + const T* const top_diff, + const T* const weight_data, + const int num, + const int outputChannels, + const int outputHeight, + const int outputWidth, + const int inputChannels, + const int inputHeight, + const int inputWidth, + const int filterMultiplier, + const int filterHeight, + const int filterWidth, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + T* const bottom_diff) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + if (index < nthreads) { + const int batch = index / inputChannels / inputHeight / inputWidth; + const int c_in = (index / inputHeight / inputWidth) % inputChannels; + const int h_in = (index / inputWidth) % inputHeight; + const int w_in = index % inputWidth; + + const int c_out_start = c_in * filterMultiplier; + + int h_out_start = (h_in - filterHeight + paddingH + strideH) / strideH; + h_out_start = 0 > h_out_start ? 0 : h_out_start; + int h_out_end = (h_in + paddingH) / strideH; + h_out_end = outputHeight - 1 < h_out_end ? outputHeight - 1 : h_out_end; + int w_out_start = (w_in - filterWidth + paddingW + strideW) / strideW; + w_out_start = 0 > w_out_start ? 0 : w_out_start; + int w_out_end = (w_in + paddingW) / strideW; + w_out_end = outputWidth - 1 < w_out_end ? outputWidth - 1 : w_out_end; + + T value = 0; + + for (int c_out = c_out_start; c_out < c_out_start + filterMultiplier; + c_out++) { + for (int h_out = h_out_start; h_out <= h_out_end; ++h_out) { + const int filter_h = h_in + paddingH - h_out * strideH; + for (int w_out = w_out_start; w_out <= w_out_end; ++w_out) { + const int filter_w = w_in + paddingW - w_out * strideW; + const int filter_offset = c_out * filterHeight * filterWidth + + filter_h * filterWidth + filter_w; + const int top_diff_offset = + ((batch * outputChannels + c_out) * outputHeight + h_out) * + outputWidth + + w_out; + value += top_diff[top_diff_offset] * weight_data[filter_offset]; + } + } + } + bottom_diff[index] += value; + } +} + +// CUDA kernel to compute the depthwise convolution backprop w.r.t filter. +template +__global__ void ConvolutionDepthwiseFilterBackward(const int num_i, + const int nthreads, + const T* const top_diff, + const T* const inputData, + const int num, + const int outputChannels, + const int outputHeight, + const int outputWidth, + const int inputChannels, + const int inputHeight, + const int inputWidth, + const int filterMultiplier, + const int filterHeight, + const int filterWidth, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + T* const buffer_data) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + if (index < nthreads) { + const int h_out = (index / outputWidth) % outputHeight; + const int w_out = index % outputWidth; + const int kh = + (index / filterWidth / outputHeight / outputWidth) % filterHeight; + const int kw = (index / outputHeight / outputWidth) % filterWidth; + const int h_in = -paddingH + h_out * strideH + kh; + const int w_in = -paddingW + w_out * strideW + kw; + if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) && + (w_in < inputWidth)) { + const int c_out = + index / (filterHeight * filterWidth * outputHeight * outputWidth); + const int c_in = c_out / filterMultiplier; + const int batch = num_i; + const int top_offset = + ((batch * outputChannels + c_out) * outputHeight + h_out) * + outputWidth + + w_out; + const int bottom_offset = + ((batch * inputChannels + c_in) * inputHeight + h_in) * inputWidth + + w_in; + buffer_data[index] = top_diff[top_offset] * inputData[bottom_offset]; + } else { + buffer_data[index] = 0; + } + } +} + +template +class DepthwiseConvFunctor { + public: + void operator()(const T* inputData, + const T* filterData, + int batchSize, + int outputChannels, + int outputHeight, + int outputWidth, + int inputChannels, + int inputHeight, + int inputWidth, + int filterMultiplier, + int filterHeight, + int filterWidth, + int strideH, + int strideW, + int paddingH, + int paddingW, + T* outputData) { + int outputSize = batchSize * outputChannels * outputHeight * outputWidth; + + size_t blocks = (outputSize + 1024 - 1) / 1024; + size_t blockX = 512; + size_t blockY = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(blockX, blockY); + + ConvolutionDepthwiseForward<<>>( + outputSize, + inputData, + filterData, + batchSize, + outputChannels, + outputHeight, + outputWidth, + inputChannels, + inputHeight, + inputWidth, + filterMultiplier, + filterHeight, + filterWidth, + strideH, + strideW, + paddingH, + paddingW, + outputData); + } +}; + +template +class DepthwiseConvGradInputFunctor { + public: + void operator()(const T* outputGrad, + const T* filterData, + int batchSize, + int outputChannels, + int outputHeight, + int outputWidth, + int inputChannels, + int inputHeight, + int inputWidth, + int filterMultiplier, + int filterHeight, + int filterWidth, + int strideH, + int strideW, + int paddingH, + int paddingW, + T* inputGrad) { + int inputSize = batchSize * inputChannels * inputHeight * inputWidth; + + size_t blocks = (inputSize + 1024 - 1) / 1024; + size_t blockX = 512; + size_t blockY = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(blockX, blockY); + + ConvolutionDepthwiseInputBackward + // NOLINT_NEXT_LINE(whitespace/operators) + <<>>(inputSize, + outputGrad, + filterData, + batchSize, + outputChannels, + outputHeight, + outputWidth, + inputChannels, + inputHeight, + inputWidth, + filterMultiplier, + filterHeight, + filterWidth, + strideH, + strideW, + paddingH, + paddingW, + inputGrad); + } +}; + +template +class DepthwiseConvGradFilterFunctor { + public: + void operator()(const T* outputGrad, + const T* inputData, + int batchSize, + int outputChannels, + int outputHeight, + int outputWidth, + int inputChannels, + int inputHeight, + int inputWidth, + int filterMultiplier, + int filterHeight, + int filterWidth, + int strideH, + int strideW, + int paddingH, + int paddingW, + T* colData, + T* filterGrad) { + int colDataSize = outputChannels * filterHeight * filterWidth * + outputHeight * outputWidth; + + size_t blocks = (colDataSize + 1024 - 1) / 1024; + size_t blockX = 512; + size_t blockY = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(blockX, blockY); + BaseMatrix filterGradMatrix(outputChannels * filterHeight * filterWidth, + 1, + filterGrad, + false, + true); + + for (int i = 0; i < batchSize; i++) { + ConvolutionDepthwiseFilterBackward< + T><<>>(i, + colDataSize, + outputGrad, + inputData, + batchSize, + outputChannels, + outputHeight, + outputWidth, + inputChannels, + inputHeight, + inputWidth, + filterMultiplier, + filterHeight, + filterWidth, + strideH, + strideW, + paddingH, + paddingW, + colData); + int K = outputHeight * outputWidth; + int M = colDataSize / K; + + BaseMatrix colMatrix(M, K, colData, false, true); + filterGradMatrix.sumRows(colMatrix, (T)1.0, (T)1.0); + } + } +}; + +#ifdef PADDLE_TYPE_DOUBLE +template class DepthwiseConvGradInputFunctor; +template class DepthwiseConvFunctor; +template class DepthwiseConvGradFilterFunctor; +#else +template class DepthwiseConvGradInputFunctor; +template class DepthwiseConvFunctor; +template class DepthwiseConvGradFilterFunctor; +#endif + +} // namespace paddle diff --git a/paddle/function/DepthwiseConvOpTest.cpp b/paddle/legacy/function/DepthwiseConvOpTest.cpp similarity index 100% rename from paddle/function/DepthwiseConvOpTest.cpp rename to paddle/legacy/function/DepthwiseConvOpTest.cpp diff --git a/paddle/legacy/function/EigenGemm.cpp b/paddle/legacy/function/EigenGemm.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5929c5c68ec818c2307580b06f76c63f04e0db5f --- /dev/null +++ b/paddle/legacy/function/EigenGemm.cpp @@ -0,0 +1,102 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/function/EigenThreadDevice.h" + +namespace paddle { + +template +struct EigenBlasGemm { + typedef Eigen::TensorMap, + Eigen::Aligned> + EigenMatrix; + + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc) { + Eigen::array sizeA; + if (transA) { + sizeA[0] = K; + sizeA[1] = M; + CHECK_EQ(M, lda); + } else { + sizeA[0] = M; + sizeA[1] = K; + CHECK_EQ(K, lda); + } + Eigen::array sizeB; + if (transB) { + sizeB[0] = N; + sizeB[1] = K; + CHECK_EQ(K, ldb); + } else { + sizeB[0] = K; + sizeB[1] = N; + CHECK_EQ(N, ldb); + } + Eigen::array sizeC = {{M, ldc}}; + Eigen::array offsetC = {{0, 0}}; + Eigen::array extentC = {{M, N}}; + + const EigenMatrix a(const_cast(A), sizeA); + const EigenMatrix b(const_cast(B), sizeB); + EigenMatrix c(C, sizeC); + + typedef typename Eigen::Tensor::DimensionPair DimPair; + Eigen::array dims; + dims[0] = DimPair(1, 0); + dims[0].first = transA ? 0 : 1; + dims[0].second = transB ? 1 : 0; + + auto* device = EigenDeviceWarpper::device(); + if (N == ldc) { + if (alpha == T(1) && beta == T(0)) { + c.device(*device) = a.contract(b, dims); + } else if (alpha == T(1) && beta == T(1)) { + c.device(*device) += a.contract(b, dims); + } else { + c.device(*device) = alpha * a.contract(b, dims) + beta * c; + } + } else { + if (alpha == T(1) && beta == T(0)) { + c.slice(offsetC, extentC).device(*device) = a.contract(b, dims); + } else if (alpha == T(1) && beta == T(1)) { + c.slice(offsetC, extentC).device(*device) += a.contract(b, dims); + } else { + c.slice(offsetC, extentC).device(*device) = + alpha * a.contract(b, dims) + beta * c.slice(offsetC, extentC); + } + } + EigenDeviceWarpper::free_device(device); + } +}; + +#ifdef PADDLE_TYPE_DOUBLE +template struct EigenBlasGemm; +#else +template struct EigenBlasGemm; +#endif + +} // namespace paddle diff --git a/paddle/function/EigenThreadDevice.h b/paddle/legacy/function/EigenThreadDevice.h similarity index 100% rename from paddle/function/EigenThreadDevice.h rename to paddle/legacy/function/EigenThreadDevice.h diff --git a/paddle/function/Function.cpp b/paddle/legacy/function/Function.cpp similarity index 100% rename from paddle/function/Function.cpp rename to paddle/legacy/function/Function.cpp diff --git a/paddle/legacy/function/Function.h b/paddle/legacy/function/Function.h new file mode 100644 index 0000000000000000000000000000000000000000..bc5ef7e6f20b63a120a577ded876820aafecff19 --- /dev/null +++ b/paddle/legacy/function/Function.h @@ -0,0 +1,214 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "BufferArg.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Any.h" +#include "paddle/legacy/utils/ClassRegistrar.h" +#include "paddle/legacy/utils/Error.h" + +namespace paddle { + +/** + * Function Configuration. + * The argument type of Function::init. + */ +class FuncConfig { + public: + template + T get(const std::string& key, Error* err = nullptr) const { + try { + return any_cast(valueMap_.at(key)); + } catch (std::exception& e) { // could be cast or out of range exception. + if (err) { + *err = Error(e.what()); + } else { + LOG(FATAL) << "Cannot get key " << key << " with error " << e.what(); + } + return T(); + } + } + + template + FuncConfig& set(const std::string& key, T v, Error* err = nullptr) { + auto it = valueMap_.find(key); + if (it != valueMap_.end()) { // already contains key. + if (err) { + *err = Error("Key %s is already set in FuncConfig", key.c_str()); + } else { + LOG(FATAL) << "Key " << key << " is already set in FuncConfig."; + } + return *this; + } + valueMap_[key] = any(v); + return *this; + } + + protected: + mutable std::unordered_map valueMap_; +}; + +/** + * Argument type for Function::calc(). + * A BufferArgs contains a set of BufferArg, + * because Function can have multiple inputs and outputs. + * + * addArg() with Matix object used to adapt Layer Argument. + * Will create a BufferArg object in addArg(), + * and free in destructor of BufferArgs. + * + * addArg() with BufferArg object, just save BufferArg object address, + * and the caller needs to guarantee the validity of the BufferArg object + * in the BufferArgs life time. + */ +class BufferArgs { + public: + BufferArgs() {} + + ~BufferArgs() { + for (auto arg : _args_) { + delete arg; + } + } + + size_t size() const { return args_.size(); } + + // add argument into BufferArgs + // Tensor can be Matrix, Vector, IVector. + // For inputs, do not need argType. + // For outputs, the argType needs to be specified as ASSIGN_TO or ADD_TO. + void addArg(const Matrix& arg, ArgType argType = UNSPECIFIED) { + _args_.push_back(new BufferArg(arg, argType)); + addArg(*_args_.back()); + } + + void addArg(const Vector& arg, ArgType argType = UNSPECIFIED) { + _args_.push_back(new BufferArg(arg, argType)); + addArg(*_args_.back()); + } + + void addArg(const IVector& arg, ArgType argType = UNSPECIFIED) { + _args_.push_back(new BufferArg(arg, argType)); + addArg(*_args_.back()); + } + + // Add arg into BufferArgs and reshape the arg. + // + // For example, arg represents an image buffer, + // but Matrix can only represent a two-dimensional Tensor. + // So need an extra argument to describe the shape of the image buffer. + void addArg(const Matrix& arg, + const TensorShape& shape, + ArgType argType = UNSPECIFIED); + + void addArg(const CpuSparseMatrix& arg, ArgType argType = UNSPECIFIED); + void addArg(const GpuSparseMatrix& arg, ArgType argType = UNSPECIFIED); + + void addArg(const Matrix& matrix, + const IVector& vector, + ArgType argType = UNSPECIFIED); + + // get argument + const BufferArg& operator[](size_t num) const { + CHECK_LT(num, args_.size()); + return *args_[num]; + } + + void addArg(BufferArg& arg) { args_.push_back(&arg); } + + void addArg(SequenceIdArg& arg) { args_.push_back(&arg); } + + void addArg(SequenceArg& arg) { args_.push_back(&arg); } + + void addArg(SparseMatrixArg& arg) { args_.push_back(&arg); } + + private: + std::vector args_; + // The BufferArg object is constructed and freed by BufferArgs. + std::vector _args_; +}; + +/** + * \brief Base class for Function. + * The basic Function implementation requires override init and calc interfaces. + * + * The caller needs to ensure the validity of the arguments + * during Function execution. + * + * Function inputs are readonly, Function outputs have two modes: ASSIGN_TO + * and ADD_TO. + * If output.getArgType() == ASSIGN_TO, this is assign mode, and the calculation + * result of Function assigned to the output BufferArg. + * If output.getArgType() == ADD_TO, this is add mode, and the calculation + * result of Function need added to the output BufferArg. + * + * For example: + * ASSIGN_TO: output = Function(inputs) + * ADD_TO: output += Function(inputs) + * If Function has more than one output, each output can have different modes. + */ +class FunctionBase { + public: + virtual ~FunctionBase() {} + + virtual void init(const FuncConfig& config) {} + + virtual void calc(const BufferArgs& inputs, const BufferArgs& outputs) {} + + // This member function is used to check whether the BufferType and shape of + // the inputs and outputs arguments of the Function are correct. + // General calc function which will call this check to do arguments check. + // And before the calc called, the caller can also check their own arguments. + virtual void check(const BufferArgs& inputs, const BufferArgs& outputs) {} + + // Calculate the number of floating-point operations of this Function. + // The inputs and outputs arguments do not need to contain the actual data, + // only the shape. + // And some Functions have the same input and output shapes, + // so you may not need to enter the complete number of arguments. + // But entering the full arguments is always correct for this interface. + virtual size_t ops(const BufferArgs& inputs, const BufferArgs& outputs) { + return 0; + } + + int getNumInputs() const { return numInputs_; } + + int getNumOutputs() const { return numOutputs_; } + + static ClassRegistrar funcRegistrar_; + + protected: + // numInputs_ and numOutputs_ represents the maximum + // input and output supported by Function. + // Some functions are optimized for input and output, + // so when comparing the number of arguments, for these functions + // inputs.size() <= numInputs_ or outputs.size() <= numOutputs_ + size_t numInputs_; + size_t numOutputs_; +}; + +#define FUNC_NAME(typeName, deviceName) #typeName "-" #deviceName + +#define REGISTER_TYPED_FUNC(typeName, deviceName, className) \ + static InitFunction __reg_type_##typeName##deviceName([]() { \ + FunctionBase::funcRegistrar_ \ + .registerClass>( \ + FUNC_NAME(typeName, deviceName)); \ + }) + +} // namespace paddle diff --git a/paddle/legacy/function/FunctionTest.cpp b/paddle/legacy/function/FunctionTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1a0993e3135bcad9eb8a431e079ed56a267174ea --- /dev/null +++ b/paddle/legacy/function/FunctionTest.cpp @@ -0,0 +1,166 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Function.h" +#include +#include "paddle/legacy/math/SparseMatrix.h" + +namespace paddle { + +template +void FunctionApi(typename Tensor::Matrix& output, + const typename Tensor::Matrix& input); + +template <> +void FunctionApi(CpuMatrix& output, const CpuMatrix& input) { + EXPECT_EQ(output.getHeight(), 100U); + EXPECT_EQ(output.getWidth(), 200U); +} + +template <> +void FunctionApi(GpuMatrix& output, const GpuMatrix& input) { + EXPECT_EQ(output.getHeight(), 10U); + EXPECT_EQ(output.getWidth(), 20U); +} + +template +void Function(const BufferArgs& arguments) { + const auto input = arguments[0].matrix(); + auto output = arguments[1].matrix(); + FunctionApi(output, input); +} + +TEST(Function, BufferArgs) { + CpuMatrix cpuInput = CpuMatrix(100, 200); + CpuMatrix cpuOutput = CpuMatrix(100, 200); + BufferArgs cpuArgments; + cpuArgments.addArg(cpuInput); + cpuArgments.addArg(cpuOutput); + Function(cpuArgments); + + GpuMatrix gpuInput = GpuMatrix(10, 20); + GpuMatrix gpuOutput = GpuMatrix(10, 20); + BufferArgs gpuArgments; + gpuArgments.addArg(gpuInput); + gpuArgments.addArg(gpuOutput); + Function(gpuArgments); +} + +/** + * Some tests case are used to check the consistency between the BufferArg type + * argument received by Function and the original type argument. + * + * Use Case: + * TEST() { + * Matrix matrix(...); + * CheckBufferArg lambda = [=](const BufferArg& arg) { + * // check matrix and arg are equivalent + * EXPECT_EQ(matrix, arg); + * } + * + * BufferArgs argments{matrix...}; + * std::vector checkFunc{lambda...}; + * testBufferArgs(argments, checkFunc); + * } + */ +typedef std::function CheckBufferArg; + +void testBufferArgs(const BufferArgs& inputs, + const std::vector& check) { + EXPECT_EQ(inputs.size(), check.size()); + for (size_t i = 0; i < inputs.size(); i++) { + check[i](inputs[i]); + } +} + +void testBufferArgs(const BufferArgs& inputs, const CheckBufferArg& check) { + EXPECT_EQ(inputs.size(), 1U); + check(inputs[0]); +} + +TEST(Arguments, Matrix) { + MatrixPtr matrix = Matrix::create(100, 200); + CheckBufferArg check = [=](const BufferArg& arg) { + EXPECT_EQ(arg.shape().ndims(), 2U); + EXPECT_EQ(arg.shape()[0], 100U); + EXPECT_EQ(arg.shape()[1], 200U); + EXPECT_EQ(arg.data(), matrix->getData()); + + EXPECT_EQ(arg.matrix().getHeight(), matrix->getHeight()); + EXPECT_EQ(arg.matrix().getWidth(), matrix->getWidth()); + EXPECT_EQ(arg.matrix().getData(), matrix->getData()); + }; + + BufferArgs argments; + argments.addArg(*matrix); + std::vector checkFunc; + checkFunc.push_back(check); + testBufferArgs(argments, checkFunc); +} + +TEST(Arguments, Vector) { + VectorPtr vector = Vector::create(100, false); + CheckBufferArg check = [=](const BufferArg& arg) { + EXPECT_EQ(arg.shape().ndims(), 1U); + EXPECT_EQ(arg.shape()[0], 100U); + EXPECT_EQ(arg.data(), vector->getData()); + + CpuVector inVector = arg.vector(); + EXPECT_EQ(inVector.getSize(), vector->getSize()); + EXPECT_EQ(inVector.getData(), vector->getData()); + }; + + BufferArgs argments; + argments.addArg(*vector); + std::vector checkFunc; + checkFunc.push_back(check); + testBufferArgs(argments, checkFunc); +} + +TEST(Arguments, CpuSparseMatrix) { + CpuSparseMatrix sparse(200, 300, 50); + CheckBufferArg check = [=](const BufferArg& arg) { + EXPECT_EQ(arg.shape().ndims(), 2U); + EXPECT_EQ(arg.shape()[0], 200U); + EXPECT_EQ(arg.shape()[1], 300U); + EXPECT_EQ(arg.data(), sparse.getData()); + // CHECK_EQ(arg.sparse().nnz(), 50); + // CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT); + // CHECK_EQ(arg.sparse().dataType(), SPARSE_FLOAT_VALUE); + EXPECT_EQ(arg.sparse().getRowBuf(), sparse.getRows()); + EXPECT_EQ(arg.sparse().getColBuf(), sparse.getCols()); + }; + + BufferArgs argments; + argments.addArg(sparse); + std::vector checkFunc; + checkFunc.push_back(check); + testBufferArgs(argments, checkFunc); +} + +TEST(Arguments, BufferArg) { + BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3}); + CheckBufferArg check = [=](const BufferArg& arg) { + EXPECT_EQ(arg.shape().ndims(), 3U); + EXPECT_EQ(arg.shape()[0], 1U); + EXPECT_EQ(arg.shape()[1], 2U); + EXPECT_EQ(arg.shape()[2], 3U); + }; + + BufferArgs argments; + argments.addArg(arg); + testBufferArgs(argments, check); +} + +} // namespace paddle diff --git a/paddle/legacy/function/FunctionTest.h b/paddle/legacy/function/FunctionTest.h new file mode 100644 index 0000000000000000000000000000000000000000..6f01981a34bff0a7d9bb04d0a0012117ecf5f803 --- /dev/null +++ b/paddle/legacy/function/FunctionTest.h @@ -0,0 +1,410 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Function.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/math/tests/TensorCheck.h" +#include "paddle/testing/TestUtil.h" + +namespace paddle { + +typedef std::shared_ptr BufferArgPtr; + +namespace test { +template +struct Allocator; + +template <> +struct Allocator { + using type = CpuMemoryHandle; +}; + +template <> +struct Allocator { + using type = GpuMemoryHandle; +}; + +// Copy argument1 to argument2 +template +class CopyArgument { + public: + void operator()(const BufferArg& arg1, BufferArg& arg2) { + CHECK_EQ(arg1.valueType(), arg2.valueType()); + CHECK_LE(arg1.shape().getElements(), arg2.shape().getElements()); + + if (arg1.valueType() == VALUE_TYPE_INT32) { + IVectorPtr vector1 = + IVector::create((int*)arg1.data(), + arg1.shape().getElements(), + DType1 == DEVICE_TYPE_CPU ? false : true); + IVectorPtr vector2 = + IVector::create((int*)arg2.data(), + arg2.shape().getElements(), + DType2 == DEVICE_TYPE_CPU ? false : true); + vector2->copyFrom(*vector1); + } else { + VectorPtr vector1 = + Vector::create((real*)arg1.data(), + arg1.shape().getElements(), + DType1 == DEVICE_TYPE_CPU ? false : true); + VectorPtr vector2 = + Vector::create((real*)arg2.data(), + arg2.shape().getElements(), + DType2 == DEVICE_TYPE_CPU ? false : true); + vector2->copyFrom(*vector1); + } + } +}; +} // namespace test + +/** + * \brief A class for comparing two Functions of different implementations. + * For example, can be used to compare the CPU and GPU implementation + * of the function is consistent. + * + * Use case: + * // Initializes a test object, the corresponding cpu and gpu Function + * // are constructed according to FunctionName and FuncConfig. + * CpuGpuFuncCompare test(FunctionName, FuncConfig); + * // Prepare inputs and outputs arguments. + * // Here the input and output can not contain real data, + * // only contains the argument type and shape. + * test.addInputs(input1); + * test.addInputs(input2); + * test.addOutputs(output1); + * test.addOutputs(output2); + * // Run. + * // Will according to the type and shape of arguments(inputs_/outputs_), + * // automatic initialization cpu and gpu function required arguments + * // (cpuInputs_/cpuOutputs_/gpuInputs_/gpuOutputs_). + * // Call the CPU and GPU Function calculation results. + * // Compares CPU and GPU calculation results for consistency. + * test.run(); + */ +template +class Compare2Function { + public: + typedef typename test::Allocator::type Allocator1; + typedef typename test::Allocator::type Allocator2; + typedef typename Tensor::Vector Vector1; + typedef typename Tensor::Vector Vector2; + typedef typename Tensor::SparseMatrix SparseMatrix1; + typedef typename Tensor::SparseMatrix SparseMatrix2; + + Compare2Function(const std::string& name1, + const std::string& name2, + const FuncConfig& config) + : function1_(FunctionBase::funcRegistrar_.createByType(name1)), + function2_(FunctionBase::funcRegistrar_.createByType(name2)) { + function1_->init(config); + function2_->init(config); + initArgsCallback_ = nullptr; + } + + ~Compare2Function() {} + + // input need only contains shape, do not contains data. + void addInputs(const BufferArg& input) { + size_t size = + input.shape().getElements() * sizeOfValuType(input.valueType()); + func1Memory_.emplace_back(std::make_shared(size)); + func2Memory_.emplace_back(std::make_shared(size)); + + func1Inputs_.emplace_back(std::make_shared( + func1Memory_.back()->getBuf(), input.valueType(), input.shape())); + func2Inputs_.emplace_back(std::make_shared( + func2Memory_.back()->getBuf(), input.valueType(), input.shape())); + } + + // assume one copy of sequence is shared by different SequenceArgs + void addSequence(const SequenceIdArg& input) { + CHECK_EQ(input.shape().ndims(), 1UL); + size_t batchSize = input.shape()[0]; + size_t numSeqs = batchSize / 10 + 1; + size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32); + func1Memory_.emplace_back(std::make_shared(sizeId)); + func2Memory_.emplace_back(std::make_shared(sizeId)); + seq1_ = std::make_shared(func1Memory_.back()->getBuf(), + TensorShape{numSeqs + 1}); + seq2_ = std::make_shared(func2Memory_.back()->getBuf(), + TensorShape{numSeqs + 1}); + /// init sequence Id + initArg(*seq1_, batchSize); + + copyArg_(*seq1_, *seq2_); + } + + void addInputs(const SequenceArg& input) { + CHECK_EQ(input.shape().ndims(), 2UL); + size_t batchSize = input.shape()[0]; + if (!seq1_ || !seq2_) { // sequence not exist + addSequence(SequenceIdArg(TensorShape{batchSize})); + } + + size_t size = + input.shape().getElements() * sizeOfValuType(input.valueType()); + func1Memory_.emplace_back(std::make_shared(size)); + func2Memory_.emplace_back(std::make_shared(size)); + + /// SequenceArg + func1Inputs_.emplace_back( + std::make_shared(func1Memory_.back()->getBuf(), + input.valueType(), + input.shape(), + *seq1_)); + func2Inputs_.emplace_back( + std::make_shared(func2Memory_.back()->getBuf(), + input.valueType(), + input.shape(), + *seq2_)); + } + + void registerInitCallback(std::function callback) { + initArgsCallback_ = callback; + } + + // output need only contains shape, do not contains data. + void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) { + size_t size = + output.shape().getElements() * sizeOfValuType(output.valueType()); + func1Memory_.emplace_back(std::make_shared(size)); + func2Memory_.emplace_back(std::make_shared(size)); + + func1Outputs_.emplace_back( + std::make_shared(func1Memory_.back()->getBuf(), + output.valueType(), + output.shape(), + argType)); + func2Outputs_.emplace_back( + std::make_shared(func2Memory_.back()->getBuf(), + output.valueType(), + output.shape(), + argType)); + } + + /// add and init output sparse matrix + void addOutputs(const SparseMatrixArg& output, ArgType argType = ASSIGN_TO) { + sparse1_ = std::make_shared( + output.shape()[0], + output.shape()[1], + output.nnz(), + static_cast(output.dataType()), + static_cast(output.dataFormat())); + + sparse2_ = std::make_shared( + output.shape()[0], + output.shape()[1], + output.nnz(), + static_cast(output.dataType()), + static_cast(output.dataFormat())); + + /// init sparse matrix + hl_stream_t stream(HPPL_STREAM_1); + sparse1_->randomizeUniform(); + sparse2_->copyFrom(*sparse1_, stream); + hl_stream_synchronize(stream); + + func1Outputs_.emplace_back( + std::make_shared(*sparse1_, argType)); + func2Outputs_.emplace_back( + std::make_shared(*sparse2_, argType)); + } + + void addOutputs(const SequenceArg& output, ArgType argType = ASSIGN_TO) { + CHECK_EQ(output.shape().ndims(), 2UL); + size_t batchSize = output.shape()[0]; + + if (!seq1_ || !seq2_) { // sequence not exist + addSequence(SequenceIdArg(TensorShape{batchSize})); + } + size_t size = + output.shape().getElements() * sizeOfValuType(output.valueType()); + func1Memory_.emplace_back(std::make_shared(size)); + func2Memory_.emplace_back(std::make_shared(size)); + + /// SequenceArg + func1Outputs_.emplace_back( + std::make_shared(func1Memory_.back()->getBuf(), + output.valueType(), + output.shape(), + *seq1_, + argType)); + func2Outputs_.emplace_back( + std::make_shared(func2Memory_.back()->getBuf(), + output.valueType(), + output.shape(), + *seq2_, + argType)); + } + + void addInputs(const SparseMatrixArg& input) { + sparse1_ = std::make_shared( + input.shape()[0], + input.shape()[1], + input.nnz(), + static_cast(input.dataType()), + static_cast(input.dataFormat())); + + sparse2_ = std::make_shared( + input.shape()[0], + input.shape()[1], + input.nnz(), + static_cast(input.dataType()), + static_cast(input.dataFormat())); + + /// init sparse matrix + hl_stream_t stream(HPPL_STREAM_1); + sparse1_->randomizeUniform(); + sparse2_->copyFrom(*sparse1_, stream); + hl_stream_synchronize(stream); + + func1Inputs_.emplace_back(std::make_shared(*sparse1_)); + func2Inputs_.emplace_back(std::make_shared(*sparse2_)); + } + + void run() { + // prepare cpu/gpu arguments + initInputs(); + + initOutputs(); + // function calculate + auto callFunction = [](FunctionBase* function, + std::vector& inputs, + std::vector& outputs) { + BufferArgs inArgs; + BufferArgs outArgs; + for (auto arg : inputs) { + inArgs.addArg(*arg); + } + for (auto arg : outputs) { + outArgs.addArg(*arg); + } + function->calc(inArgs, outArgs); + }; + + callFunction(function1_.get(), func1Inputs_, func1Outputs_); + callFunction(function2_.get(), func2Inputs_, func2Outputs_); + + // check outputs + compareOutputs(); + } + + std::shared_ptr getFunction1() const { return function1_; } + + std::shared_ptr getFunction2() const { return function2_; } + + protected: + // only init cpu argument, gpu argument copy from cpu argument. + void initArg(BufferArg& arg) { + Vector1 vector(arg.shape().getElements(), (real*)arg.data()); + vector.uniform(0.001, 1); + } + + void initArg(SequenceArg& arg) { + /// init only matrix + Vector1 vector(arg.shape().getElements(), (real*)arg.data()); + vector.uniform(0.001, 1); + } + + void initArg(SequenceIdArg& arg, size_t batchSize) { + size_t numSeqs = arg.numSeqs(); + int* buf = reinterpret_cast(arg.data()); + int pos = 0; + size_t maxLen = 2 * batchSize / numSeqs; + for (int i = 0; i < (int)numSeqs; ++i) { + int len = 1 + uniformRandom(std::min( + maxLen, batchSize - pos - numSeqs + i)); + buf[i] = pos; + pos += len; + VLOG(1) << " len=" << len; + } + buf[numSeqs] = batchSize; + } + + void initInputs() { + for (size_t i = 0; i < func1Inputs_.size(); i++) { + if (func1Inputs_[i]->isSparseArg()) { + continue; /// sparse matrix already init + } + + if (func1Inputs_[i]->isSequenceArg()) { + initArg(dynamic_cast(*func1Inputs_[i])); + } else { + initArg(*func1Inputs_[i]); + } + + if (initArgsCallback_ != nullptr) { + initArgsCallback_(*func1Inputs_[i], i); + } + + copyArg_(*func1Inputs_[i], *func2Inputs_[i]); + } + } + + void initOutputs() { + for (size_t i = 0; i < func1Outputs_.size(); i++) { + if (func1Outputs_[i]->isSparseArg()) { + continue; /// sparse matrix already init + } + + if (func1Outputs_[i]->isSequenceArg()) { + initArg(dynamic_cast(*func1Outputs_[i])); + } else { + initArg(*func1Outputs_[i]); + } + + copyArg_(*func1Outputs_[i], *func2Outputs_[i]); + } + } + + void compareOutputs() { + for (size_t i = 0; i < func1Outputs_.size(); i++) { + // TODO, Need a BufferCheck used to compare the two buffers. + const auto cpu = func1Outputs_[i]; + const auto gpu = func2Outputs_[i]; + CHECK_EQ(cpu->numElements(), gpu->numElements()); + Vector1 cpuVector(cpu->numElements(), (real*)cpu->data()); + Vector2 gpuVector(gpu->numElements(), (real*)gpu->data()); + autotest::TensorCheckErr(cpuVector, gpuVector); + } + } + + protected: + std::shared_ptr function1_; + std::shared_ptr function2_; + std::vector> func1Memory_; + std::vector> func2Memory_; + std::vector func1Inputs_; + std::vector func1Outputs_; + std::vector func2Inputs_; + std::vector func2Outputs_; + std::shared_ptr sparse1_; + std::shared_ptr sparse2_; + std::shared_ptr seq1_; + std::shared_ptr seq2_; + test::CopyArgument copyArg_; + std::function initArgsCallback_; +}; + +class CpuGpuFuncCompare + : public Compare2Function { + public: + CpuGpuFuncCompare(const std::string& name, const FuncConfig& config) + : Compare2Function(name + "-CPU", name + "-GPU", config) {} + + ~CpuGpuFuncCompare() {} +}; + +} // namespace paddle diff --git a/paddle/legacy/function/GemmConvOp.cpp b/paddle/legacy/function/GemmConvOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5a81315661dc2843a648315ca4a6b590f217a657 --- /dev/null +++ b/paddle/legacy/function/GemmConvOp.cpp @@ -0,0 +1,522 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvOp.h" +#include "GemmFunctor.h" +#include "Im2Col.h" +#include "paddle/legacy/math/MemoryHandle.h" + +namespace paddle { + +/* + * \brief Forward calculation of convolution. + */ +template +class GemmConvFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + // TODO(hedaoyuan): Need to define some index macros, + // to avoid useing 0 and 1. + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + real beta; + if (outputs[0].getArgType() == ADD_TO) { + beta = 1.0; + } else { + beta = 0.0; + } + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + + real* inputData = inputs[0].data(); + real* filterData = inputs[1].data(); + real* outputData = outputs[0].data(); + bool needIm2col = isNeedIm2col(filter); + + TensorShape imShape = + TensorShape({inputChannels / groups_, inputHeight, inputWidth}); + + TensorShape colShape; + real* colData = NULL; + + if (needIm2col) { + colShape = TensorShape({inputChannels / groups_, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + resizeBuffer(colShape.getElements()); + colData = reinterpret_cast(memory_->getBuf()); + } + + Im2ColFunctor im2col; + size_t inputOffset = imShape.getElements(); + size_t outputOffset = + (outputChannels / groups_) * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + + for (size_t i = 0; i < batchSize; i++) { + for (size_t g = 0; g < groups_; g++) { + if (needIm2col) { + im2col(inputData + g * inputOffset, + imShape, + colData, + colShape, + strideH(), + strideW(), + paddingH(), + paddingW(), + dilationH(), + dilationW()); + } else { + colData = inputData + g * inputOffset; + } + int M = outputChannels / groups_; + int N = outputHeight * outputWidth; + int K = inputChannels / groups_ * filterHeight * filterWidth; + BlasGemm::compute(false, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + K, + colData, + N, + beta, + outputData + g * outputOffset, + N); + } + inputData += inputChannels * inputHeight * inputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifdef PADDLE_MOBILE_INFERENCE + +/* + * \brief Forward calculation of convolution, optimized for mobile. + */ +template +class GemmConvMobileFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + // TODO(hedaoyuan): Need to define some index macros, + // to avoid useing 0 and 1. + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + real beta; + if (outputs[0].getArgType() == ADD_TO) { + beta = 1.0; + } else { + beta = 0.0; + } + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + + real* inputData = inputs[0].data(); + real* filterData = inputs[1].data(); + real* outputData = outputs[0].data(); + real* colData = NULL; + bool needIm2col = isNeedIm2col(filter); + + TensorShape imShape = + TensorShape({inputChannels / groups_, inputHeight, inputWidth}); + TensorShape colShape; + + // Max col matrix width 4096, Max col matrix size 4M. + size_t outputHeightSteps = + std::min(std::max(4096 / outputWidth, (size_t)1), outputHeight); + size_t maxColWidth = outputHeightSteps * outputWidth; + size_t channelSteps = + std::min(std::max((1048576 / maxColWidth) / filterHeight * filterWidth, + (size_t)1), + inputChannels / groups_); + size_t maxColHeight = channelSteps * filterHeight * filterWidth; + + if (needIm2col) { + colShape = TensorShape({inputChannels / groups_, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + + resizeBuffer(maxColHeight * maxColWidth * sizeof(real)); + colData = reinterpret_cast(memory_->getBuf()); + } + + Im2ColMobileFunctor im2col; + size_t inputOffset = imShape.getElements(); + size_t outputOffset = + (outputChannels / groups_) * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + + int nStride = outputHeight * outputWidth; + int kStride = inputChannels / groups_ * filterHeight * filterWidth; + for (size_t i = 0; i < batchSize; i++) { + filterData = inputs[1].data(); + for (size_t g = 0; g < groups_; g++) { + if (needIm2col) { + real beta_ = beta; + for (size_t ic = 0; ic < inputChannels / groups_; + ic += channelSteps) { + int channels = std::min(inputChannels / groups_ - ic, channelSteps); + for (size_t oh = 0; oh < outputHeight; oh += outputHeightSteps) { + int height = std::min(outputHeight - oh, outputHeightSteps); + + int M = outputChannels / groups_; + int N = height * outputWidth; + int K = channels * filterHeight * filterWidth; + // im2col + im2col(inputData, + imShape, + colData, + colShape, + strideH(), + strideW(), + paddingH(), + paddingW(), + dilationH(), + dilationW(), + channels, + oh, + height, + N); + + // gemm + BlasGemm::compute( + false, + false, + M, + N, + K, + 1.0f, + filterData + ic * filterHeight * filterWidth, + kStride, + colData, + N, + beta_, + outputData + oh * outputWidth, + nStride); + } + beta_ = 1.0; + } + } else { + int M = outputChannels / groups_; + int N = outputHeight * outputWidth; + int K = inputChannels / groups_ * filterHeight * filterWidth; + BlasGemm::compute(false, + false, + M, + N, + K, + 1.0f, + filterData, + K, + inputData, + N, + beta, + outputData, + N); + } + inputData += inputOffset; + outputData += outputOffset; + filterData += filterOffset; + } + } + + memory_.reset(); + } +}; + +#endif + +/* + * \brief Backward input calculation of convolution. + */ +template +class GemmConvGradInputFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& output = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& input = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + // Since the implementation of Col2ImFunctor is ADD_TO, + // this function only supports ADD_TO mode. + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + const TensorShape& output = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& input = outputs[0].shape(); + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + + real* outputGrad = inputs[0].data(); + real* filterData = inputs[1].data(); + real* inputGrad = outputs[0].data(); + bool needIm2col = isNeedIm2col(filter); + + TensorShape imShape = + TensorShape({inputChannels / groups_, inputHeight, inputWidth}); + + TensorShape colShape; + real* colData = NULL; + + if (needIm2col) { + colShape = TensorShape({inputChannels / groups_, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + resizeBuffer(colShape.getElements()); + colData = reinterpret_cast(memory_->getBuf()); + } + + Col2ImFunctor col2im; + size_t inputOffset = imShape.getElements(); + size_t outputOffset = + (outputChannels / groups_) * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + + for (size_t i = 0; i < batchSize; i++) { + for (size_t g = 0; g < groups_; g++) { + int K = outputChannels / groups_; + int N = outputHeight * outputWidth; + int M = inputChannels / groups_ * filterHeight * filterWidth; + real scale = 0.0f; + if (!needIm2col) { + colData = inputGrad + g * inputOffset; + scale = 1.0f; + } + BlasGemm::compute(true, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + M, + outputGrad + g * outputOffset, + N, + scale, + colData, + N); + if (needIm2col) { + col2im(inputGrad + g * inputOffset, + imShape, + colData, + colShape, + strideH(), + strideW(), + paddingH(), + paddingW(), + dilationH(), + dilationW()); + } + } + inputGrad += inputChannels * inputHeight * inputWidth; + outputGrad += outputChannels * outputHeight * outputWidth; + } + } +}; + +/* + * \brief Backward filter calculation of convolution. + */ +template +class GemmConvGradFilterFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& output = inputs[0].shape(); + const TensorShape& input = inputs[1].shape(); + const TensorShape& filter = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + const TensorShape& output = inputs[0].shape(); + const TensorShape& input = inputs[1].shape(); + const TensorShape& filter = outputs[0].shape(); + + real beta; + if (outputs[0].getArgType() == ADD_TO) { + beta = 1.0; + } else { + beta = 0.0; + } + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + + real* outputGrad = inputs[0].data(); + real* inputData = inputs[1].data(); + real* filterGrad = outputs[0].data(); + bool needIm2col = isNeedIm2col(filter); + + TensorShape imShape = + TensorShape({inputChannels / groups_, inputHeight, inputWidth}); + + TensorShape colShape; + real* colData = NULL; + + if (needIm2col) { + colShape = TensorShape({inputChannels / groups_, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + resizeBuffer(colShape.getElements()); + colData = reinterpret_cast(memory_->getBuf()); + } + + Im2ColFunctor im2col; + size_t inputOffset = imShape.getElements(); + size_t outputOffset = + (outputChannels / groups_) * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + for (size_t i = 0; i < batchSize; i++) { + for (size_t g = 0; g < groups_; g++) { + if (needIm2col) { + im2col(inputData + g * inputOffset, + imShape, + colData, + colShape, + strideH(), + strideW(), + paddingH(), + paddingW(), + dilationH(), + dilationW()); + } else { + colData = inputData + g * inputOffset; + } + int M = outputChannels / groups_; + int K = outputHeight * outputWidth; + int N = inputChannels / groups_ * filterHeight * filterWidth; + BlasGemm::compute(false, + true, + M, + N, + K, + 1.0f, + outputGrad + g * outputOffset, + K, + colData, + K, + i == 0 ? beta : 1.0f, + filterGrad + g * filterOffset, + N); + } + inputData += inputChannels * inputHeight * inputWidth; + outputGrad += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifdef PADDLE_MOBILE_INFERENCE +REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvMobileFunction); +#else +REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction); +#endif +REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction); +REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction); +REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction); +REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction); +#endif + +} // namespace paddle diff --git a/paddle/function/GemmConvOpTest.cpp b/paddle/legacy/function/GemmConvOpTest.cpp similarity index 100% rename from paddle/function/GemmConvOpTest.cpp rename to paddle/legacy/function/GemmConvOpTest.cpp diff --git a/paddle/legacy/function/GemmFunctor.cpp b/paddle/legacy/function/GemmFunctor.cpp new file mode 100644 index 0000000000000000000000000000000000000000..450293dfeea170e287cfc90226dabad25c76e537 --- /dev/null +++ b/paddle/legacy/function/GemmFunctor.cpp @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GemmFunctor.h" +#include "paddle/legacy/math/MathFunctions.h" + +namespace paddle { + +template +struct BlasGemm { + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc) { +#ifdef PADDLE_USE_EIGEN_FOR_BLAS + EigenBlasGemm::compute( + transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc); +#else + gemm(transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, + M, + N, + K, + alpha, + A, + lda, + B, + ldb, + beta, + C, + ldc); +#endif + } +}; + +template +struct BlasGemm { + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc) { + hl_matrix_mul((T*)A, + transA == false ? HPPL_OP_N : HPPL_OP_T, + (T*)B, + transB == false ? HPPL_OP_N : HPPL_OP_T, + C, + M, + N, + K, + alpha, + beta, + lda, + ldb, + ldc); + } +}; + +template struct BlasGemm; +template struct BlasGemm; + +} // namespace paddle diff --git a/paddle/function/GemmFunctor.h b/paddle/legacy/function/GemmFunctor.h similarity index 100% rename from paddle/function/GemmFunctor.h rename to paddle/legacy/function/GemmFunctor.h diff --git a/paddle/function/GruFunctor.h b/paddle/legacy/function/GruFunctor.h similarity index 100% rename from paddle/function/GruFunctor.h rename to paddle/legacy/function/GruFunctor.h diff --git a/paddle/function/Im2Col.h b/paddle/legacy/function/Im2Col.h similarity index 100% rename from paddle/function/Im2Col.h rename to paddle/legacy/function/Im2Col.h diff --git a/paddle/function/Im2ColOp.cpp b/paddle/legacy/function/Im2ColOp.cpp similarity index 100% rename from paddle/function/Im2ColOp.cpp rename to paddle/legacy/function/Im2ColOp.cpp diff --git a/paddle/function/Im2ColOpGpu.cu b/paddle/legacy/function/Im2ColOpGpu.cu similarity index 100% rename from paddle/function/Im2ColOpGpu.cu rename to paddle/legacy/function/Im2ColOpGpu.cu diff --git a/paddle/legacy/function/Im2ColTest.cpp b/paddle/legacy/function/Im2ColTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2c5f06f38991497963cfbe1e12825f1bc39dffa6 --- /dev/null +++ b/paddle/legacy/function/Im2ColTest.cpp @@ -0,0 +1,223 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Im2Col.h" +#include +#include "Function.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/tests/TensorCheck.h" + +namespace paddle { + +template +void TestIm2ColFunctor() { + for (size_t channels : {1, 5, 32}) { + for (size_t inputHeight : {5, 33, 100}) { + for (size_t inputWidth : {5, 32, 96}) { + for (size_t filterHeight : {1, 5}) { + for (size_t filterWidth : {3, 7}) { + for (size_t stride : {1, 2}) { + for (size_t padding : {0, 1}) { + for (size_t dilation : {1, 3}) { + size_t filterSizeH = (filterHeight - 1) * dilation + 1; + size_t filterSizeW = (filterWidth - 1) * dilation + 1; + if (inputHeight + 2 * padding < filterSizeH || + inputWidth + 2 * padding < filterSizeW) + break; + if (padding >= filterSizeH || padding >= filterSizeW) break; + size_t outputHeight = + (inputHeight - filterSizeH + 2 * padding) / stride + 1; + size_t outputWidth = + (inputWidth - filterSizeW + 2 * padding) / stride + 1; + + TensorShape imShape = + TensorShape({channels, inputHeight, inputWidth}); + TensorShape colShape1 = TensorShape({channels, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + TensorShape colShape2 = TensorShape({outputHeight, + outputWidth, + channels, + filterHeight, + filterWidth}); + + size_t height = channels * filterHeight * filterWidth; + size_t width = outputHeight * outputWidth; + VectorPtr input1 = + Vector::create(imShape.getElements(), false); + VectorPtr input2 = + Vector::create(imShape.getElements(), false); + MatrixPtr output1 = + Matrix::create(height, width, false, false); + MatrixPtr output2 = + Matrix::create(width, height, false, false); + input1->uniform(0.001, 1); + input2->copyFrom(*input1); + + Im2ColFunctor im2Col1; + Im2ColFunctor im2Col2; + im2Col1(input1->getData(), + imShape, + output1->getData(), + colShape1, + stride, + stride, + padding, + padding, + dilation, + dilation); + im2Col2(input2->getData(), + imShape, + output2->getData(), + colShape2, + stride, + stride, + padding, + padding, + dilation, + dilation); + + // The transposition of the result of ColFormat == kCFO + // is equal to the result of ColFormat == kOCF. + MatrixPtr test; + output2->transpose(test, true); + autotest::TensorCheckErr(*output1, *test); + + Col2ImFunctor col2Im1; + Col2ImFunctor col2Im2; + + col2Im1(input1->getData(), + imShape, + output1->getData(), + colShape1, + stride, + stride, + padding, + padding, + dilation, + dilation); + col2Im2(input2->getData(), + imShape, + output2->getData(), + colShape2, + stride, + stride, + padding, + padding, + dilation, + dilation); + autotest::TensorCheckErr(*input1, *input2); + } + } + } + } + } + } + } + } +} + +TEST(Im2ColFunctor, CPU) { TestIm2ColFunctor(); } + +#ifdef PADDLE_WITH_CUDA + +TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor(); } + +#endif + +template +void TestIm2ColMobileFunctor() { + for (size_t channels : {32}) { + for (size_t inputHeight : {33, 100}) { + for (size_t inputWidth : {32, 96}) { + for (size_t filterHeight : {5}) { + for (size_t filterWidth : {7}) { + for (size_t stride : {2}) { + for (size_t padding : {1}) { + for (size_t dilation : {1, 3}) { + size_t filterSizeH = (filterHeight - 1) * dilation + 1; + size_t filterSizeW = (filterWidth - 1) * dilation + 1; + if (inputHeight + 2 * padding < filterSizeH || + inputWidth + 2 * padding < filterSizeW) + break; + if (padding >= filterSizeH || padding >= filterSizeW) break; + size_t outputHeight = + (inputHeight - filterSizeH + 2 * padding) / stride + 1; + size_t outputWidth = + (inputWidth - filterSizeW + 2 * padding) / stride + 1; + + TensorShape imShape = + TensorShape({channels, inputHeight, inputWidth}); + TensorShape colShape1 = TensorShape({channels, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + + size_t height = channels * filterHeight * filterWidth; + size_t width = outputHeight * outputWidth; + VectorPtr input1 = + Vector::create(imShape.getElements(), false); + VectorPtr input2 = + Vector::create(imShape.getElements(), false); + MatrixPtr output1 = + Matrix::create(height, width, false, false); + MatrixPtr output2 = + Matrix::create(height, width, false, false); + input1->uniform(0.001, 1); + input2->copyFrom(*input1); + + Im2ColFunctor im2Col1; + Im2ColMobileFunctor im2Col2; + im2Col1(input1->getData(), + imShape, + output1->getData(), + colShape1, + stride, + stride, + padding, + padding, + dilation, + dilation); + im2Col2(input2->getData(), + imShape, + output2->getData(), + colShape1, + stride, + stride, + padding, + padding, + dilation, + dilation, + channels, + 0, + outputHeight, + outputHeight * outputWidth); + + autotest::TensorCheckEqual(*output1, *output2); + } + } + } + } + } + } + } + } +} + +TEST(Im2ColFunctor, Mobile) { TestIm2ColMobileFunctor(); } + +} // namespace paddle diff --git a/paddle/legacy/function/MulOp.cpp b/paddle/legacy/function/MulOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..750978fc90201ccdc0a32f93fc01a2170d3f39d5 --- /dev/null +++ b/paddle/legacy/function/MulOp.cpp @@ -0,0 +1,347 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MulOp.h" +#include "GemmFunctor.h" +#include "paddle/legacy/math/SIMDFunctions.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace { +inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { + for (unsigned int i = 0; i < len; ++i) { + a[i] += (1.0 == scaleB) ? b[i] : scaleB * b[i]; + } +} + +inline void colVecAddTo( + real* a, real* b, real c, size_t len, size_t aWidth, size_t bWidth) { + for (unsigned int i = 0; i < len; ++i) { + a[i * aWidth] += (1.0 == c) ? b[i * bWidth] : b[i * bWidth] * c; + } +} +} // namespace + +namespace paddle { +/// sparse matrix (+)= dense matrix * dense matrix +template <> +void MulOp(CpuSparseMatrix& out, + const CpuMatrix& a, + const CpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + CHECK_EQ(out.getValueType(), FLOAT_VALUE); + if (scaleT == 0) { + out.zeroMem(); + } + const real* A = a.getData(); + const real* B = b.getData(); + real* C = out.getValue(); + int* rows = out.getRows(); + int* cols = out.getCols(); + size_t width = out.getWidth(); + size_t height = out.getHeight(); + + /// SPARSE_CSC, {a any, b not trans} + if (out.getFormat() == SPARSE_CSC) { + /// b not trans and a any + CHECK(!bTrans); + size_t m = !aTrans ? a.getWidth() : a.getHeight(); + for (size_t i = 0; i < width; i++) { + size_t start = out.getColStartIdx(i); + size_t end = out.getColStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + real sum = 0; + size_t rowIdx = rows[j]; + for (size_t k = 0; k < m; k++) { + sum += (!aTrans ? A[rowIdx * m + k] : A[k * height + rowIdx]) * + B[k * width + i]; + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + return; + } + + /// SPARSE_CSR, {a any, b not trans} or {a not trans, b trans} + if (out.getFormat() == SPARSE_CSR) { + /// a and b can not both transpose + CHECK(!(aTrans && bTrans)); + size_t m = a.getWidth(); + for (size_t i = 0; i < height; i++) { + size_t start = out.getRowStartIdx(i); + size_t end = out.getRowStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + real sum = 0; + size_t colIdx = cols[j]; + for (size_t k = 0; k < m; k++) { + sum += (!aTrans ? A[i * m + k] : A[k * height + i]) * + (!bTrans ? B[k * width + colIdx] : B[colIdx * m + k]); + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + return; + } +} + +/// dense matrix (+)= dense matrix * dense matrix +template <> +void MulOp(CpuMatrix& out, + const CpuMatrix& a, + const CpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + BlasGemm::compute( + aTrans, + bTrans, + out.getHeight(), + out.getWidth(), + !aTrans ? a.getWidth() : a.getHeight(), + scaleAB, + a.getData(), + a.getStride(), + b.getData(), + b.getStride(), + scaleT, + out.getData(), + out.getStride()); +} + +/// dense matrix (+)= sparse matrix * dense matrix +template <> +void MulOp(CpuMatrix& out, + const CpuSparseMatrix& a, + const CpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + if (scaleT == 0) { + out.zeroMem(); + } + const real* B = b.getData(); + real* C = out.getData(); + if (out.getWidth() % 32 == 0) { + CHECK_EQ((size_t)B % 32, 0UL); + CHECK_EQ((size_t)C % 32, 0UL); + } + + int* cols = a.getCols(); + real* values = a.getValue(); + for (size_t i = 0; i < a.getHeight(); ++i) { + const int start = a.getRowStartIdx(i); + const int end = a.getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + vecAddTo(!aTrans ? out.getRow(i) : out.getRow(cols[j]), + !aTrans ? const_cast(b).getRow(cols[j]) + : const_cast(b).getRow(i), + (a.getValueType() == FLOAT_VALUE) ? values[j] : (real)1.0, + out.getWidth()); + } + } +} + +/// dense matrix (+)= dense matrix * sparse matrix +template <> +void MulOp(CpuMatrix& out, + const CpuMatrix& a, + const CpuSparseMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + if (scaleT == 0) { + out.zeroMem(); + } + real* A = const_cast(a.getData()); + real* B = const_cast(b.getValue()); + real* C = out.getData(); + int* rows = b.getRows(); + int* cols = b.getCols(); + + /// SPARSE_CSC format + if (b.getFormat() == SPARSE_CSC) { + for (size_t j = 0; j < b.getWidth(); ++j) { + int start = b.getColStartIdx(j); + int end = b.getColStartIdx(j + 1); + for (int i = start; i < end; ++i) { + colVecAddTo(!bTrans ? C + j : C + rows[i], + !bTrans ? A + rows[i] : A + j, + (b.getValueType() == NO_VALUE) ? (real)1.0 : B[i], + out.getHeight(), + out.getWidth(), + a.getWidth()); + } + } + return; + } + + /// SPARSE_CSR format + if (b.getFormat() == SPARSE_CSR) { + for (size_t j = 0; j < b.getHeight(); ++j) { + int start = b.getRowStartIdx(j); + int end = b.getRowStartIdx(j + 1); + for (int i = start; i < end; ++i) { + colVecAddTo(!bTrans ? C + cols[i] : C + j, + !bTrans ? A + j : A + cols[i], + (b.getValueType() == NO_VALUE) ? (real)1.0 : B[i], + out.getHeight(), + out.getWidth(), + a.getWidth()); + } + } + return; + } +} + +/** + * mul operator + * out = scaleT * out + scaleAB * (A * B) + * here, scaleT in {0, 1}, scaleAB == 1, + * out = A * B, ASSIGN_TO + * out += A * B, ADD_TO + * + * + * \param outputs[0] output matrix (out), M * N, + * could be either Sparse or Dense Matrix + * M is num of rows, N is num of columns + * \param inputs[0] first input matrix (A), M * K (if non-trans) + * could be either Sparse or Dense Matrix + * M is num of rows, K is num of columns + * \param inputs[1] second input matrix (B), K * N (if non-trans) + * could be either Sparse or Dense Matrix + * K is num of rows, N is num of columns + * + * Support eight Mul operators, with both GPU and CPU devices + * For each device, four Mul operators are supported: + * 1. dense (out) = dense (A) * dense (B) + * 2. dense (out) = sparse (A) * dense (B) + * sparse matrix only support SPARSE_CSR format + * 3. dense (out) = dense (A) * sparse (B) + * sparse matrix support SPARSE_CSC and SPARSE_CSR formats + * 4. sparse (out) = dense (A) * dense (B) + * sparse matrix support SPARSE_CSC and SPARSE_CSR formats + * + */ +template +class MulFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { + aTrans_ = config.get("aTrans"); + bTrans_ = config.get("bTrans"); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK(!aTrans_ || !bTrans_) + << "Not support both a and b are transpose matrices"; + + CHECK_EQ((size_t)2, inputs.size()); + CHECK_EQ((size_t)1, outputs.size()); + CHECK(inputs[0].data() && inputs[1].data() && outputs[0].data()); + CHECK_EQ(inputs[0].shape().ndims(), (size_t)2); + CHECK_EQ(inputs[1].shape().ndims(), (size_t)2); + CHECK_EQ(outputs[0].shape().ndims(), (size_t)2); + + size_t aRow = !aTrans_ ? inputs[0].shape()[0] : inputs[0].shape()[1]; + size_t aCol = !aTrans_ ? inputs[0].shape()[1] : inputs[0].shape()[0]; + size_t bRow = !bTrans_ ? inputs[1].shape()[0] : inputs[1].shape()[1]; + size_t bCol = !bTrans_ ? inputs[1].shape()[1] : inputs[1].shape()[0]; + /// C = A * B, or C += A * B, for matrix format + CHECK_EQ(aCol, bRow); + CHECK_EQ(aRow, outputs[0].shape()[0]); + CHECK_EQ(bCol, outputs[0].shape()[1]); + + /// only support C = A * B (ASSIGN_TO) or C += A * B (ADD_TO) + real scaleT = (outputs[0].getArgType() == ADD_TO) ? 1.0 : 0.0; + + /// support dense = not both sparse * sparse + /// or sparse = dense * dense + CHECK((!outputs[0].isSparseArg() && + !(inputs[0].isSparseArg() && inputs[1].isSparseArg())) || + (outputs[0].isSparseArg() && !inputs[0].isSparseArg() && + !inputs[1].isSparseArg())); + + auto outMat = outputs[0].matrix(); + /// dense matrix = dense matrix * dense matrix + if (!inputs[0].isSparseArg() && !inputs[1].isSparseArg() && + !outputs[0].isSparseArg()) { + MulOp(outMat, + inputs[0].matrix(), + inputs[1].matrix(), + 1.0, // scaleAB + scaleT, + aTrans_, + bTrans_); + return; + } + + /// dense matrix = dense matrix * sparse matrix + if (!inputs[0].isSparseArg() && inputs[1].isSparseArg() && + !outputs[0].isSparseArg()) { + CHECK(!aTrans_) << "Not supported a transpose"; + MulOp(outMat, + inputs[0].matrix(), + inputs[1].sparse().SparseMatrix(), + 1.0, // scaleAB + scaleT, + aTrans_, + bTrans_); + return; + } + + /// dense matrix = sparse matrix * dense matrix + if (inputs[0].isSparseArg() && !inputs[1].isSparseArg() && + !outputs[0].isSparseArg()) { + CHECK(!bTrans_) << "Not supported b transpose"; + CHECK_EQ(inputs[0].sparse().dataFormat(), T_SPARSE_CSR) + << "Only supported SPARSE_CSR format for sparse matrix a"; + MulOp(outMat, + inputs[0].sparse().SparseMatrix(), + inputs[1].matrix(), + 1.0, // scaleAB + scaleT, + aTrans_, + bTrans_); + return; + } + + /// sparse matrix = dense matrix * dense matrix + auto outSparseMat = outputs[0].sparse().SparseMatrix(); + if (!inputs[0].isSparseArg() && !inputs[1].isSparseArg() && + outputs[0].isSparseArg()) { + MulOp(outSparseMat, + inputs[0].matrix(), + inputs[1].matrix(), + 1.0, // scaleAB + scaleT, + aTrans_, + bTrans_); + return; + } + } + + private: + bool aTrans_; + bool bTrans_; +}; + +REGISTER_TYPED_FUNC(MulOp, CPU, MulFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(MulOp, GPU, MulFunc); +#endif +} // namespace paddle diff --git a/paddle/legacy/function/MulOp.h b/paddle/legacy/function/MulOp.h new file mode 100644 index 0000000000000000000000000000000000000000..ab33bde17296cd2b17ac45c5a936cfd2727919a5 --- /dev/null +++ b/paddle/legacy/function/MulOp.h @@ -0,0 +1,102 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Function.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" + +namespace paddle { +/// CPU, dense matrix (+)= dense matrix * dense matrix +template +void MulOp(CpuMatrix& out, + const CpuMatrix& a, + const CpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// CPU, dense matrix (+)= sparse matrix * dense matrix +template +void MulOp(CpuMatrix& out, + const CpuSparseMatrix& a, + const CpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// CPU, dense matrix (+)= dense matrix * sparse matrix +template +void MulOp(CpuMatrix& out, + const CpuMatrix& a, + const CpuSparseMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// CPU, sparse matrix (+)= dense matrix * dense matrix +template +void MulOp(CpuSparseMatrix& out, + const CpuMatrix& a, + const CpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// GPU, dense matrix (+)= dense matrix * dense matrix +template +void MulOp(GpuMatrix& out, + const GpuMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// GPU, dense matrix (+)= sparse matrix * dense matrix +template +void MulOp(GpuMatrix& out, + const GpuSparseMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// GPU, dense matrix (+)= dense matrix * sparse matrix +template +void MulOp(GpuMatrix& out, + const GpuMatrix& a, + const GpuSparseMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +/// GPU, sparse matrix (+)= dense matrix * dense matrix +template +void MulOp(GpuSparseMatrix& out, + const GpuMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans); + +} // namespace paddle diff --git a/paddle/legacy/function/MulOpGpu.cu b/paddle/legacy/function/MulOpGpu.cu new file mode 100644 index 0000000000000000000000000000000000000000..217c983cb75dfcbc0e17f752a66847c5e92fcc91 --- /dev/null +++ b/paddle/legacy/function/MulOpGpu.cu @@ -0,0 +1,130 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MulOp.h" +#include "hl_base.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" + +namespace paddle { +/// dense matrix (+)= dense matrix * dense matrix +template <> +void MulOp(GpuMatrix& out, + const GpuMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; + hl_matrix_mul(const_cast(a.getData()), + !aTrans ? HPPL_OP_N : HPPL_OP_T, + const_cast(b.getData()), + !bTrans ? HPPL_OP_N : HPPL_OP_T, + const_cast(out.getData()), + out.getHeight(), + out.getWidth(), + !aTrans ? a.getWidth() : a.getHeight(), + scaleAB, + scaleT, + a.getStride(), + b.getStride(), + out.getStride()); +} + +/// dense matrix (+)= sparse matrix * dense matrix +template <> +void MulOp(GpuMatrix& out, + const GpuSparseMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + CHECK(out.isContiguous()); + CHECK(b.isContiguous()); + CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; + hl_matrix_csr_mul_dense(a.sMatrix_.get(), + aTrans ? HPPL_OP_T : HPPL_OP_N, + const_cast(b.getData()), + HPPL_OP_N, + const_cast(out.getData()), + out.getHeight(), + out.getWidth(), + b.getHeight(), + scaleAB, + scaleT); +} + +/// dense matrix (+)= dense matrix * sparse matrix +template <> +void MulOp(GpuMatrix& out, + const GpuMatrix& a, + const GpuSparseMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + CHECK(out.isContiguous()); + CHECK(a.isContiguous()); + CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; + + if (b.format_ == SPARSE_CSC) { + hl_matrix_dense_mul_csc(const_cast(a.getData()), + HPPL_OP_N, + b.sMatrix_.get(), + bTrans ? HPPL_OP_T : HPPL_OP_N, + const_cast(out.getData()), + out.getHeight(), + out.getWidth(), + a.getWidth(), + scaleAB, + scaleT); + } else { + hl_matrix_dense_mul_csr(const_cast(a.getData()), + HPPL_OP_N, + b.sMatrix_.get(), + bTrans ? HPPL_OP_T : HPPL_OP_N, + const_cast(out.getData()), + out.getHeight(), + out.getWidth(), + a.getWidth(), + scaleAB, + scaleT); + } +} + +/// sparse matrix (+)= dense matrix * dense matrix +template <> +void MulOp(GpuSparseMatrix& out, + const GpuMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT, + bool aTrans, + bool bTrans) { + CHECK(a.useGpu_ && b.useGpu_) << "matrix device type not match"; + hl_sparse_matrix_mul(const_cast(a.getData()), + aTrans ? HPPL_OP_T : HPPL_OP_N, + const_cast(b.getData()), + bTrans ? HPPL_OP_T : HPPL_OP_N, + out.sMatrix_.get(), + out.getHeight(), + out.getWidth(), + !bTrans ? b.getHeight() : b.getWidth(), + scaleAB, + scaleT); +} + +} // namespace paddle diff --git a/paddle/legacy/function/MulOpTest.cpp b/paddle/legacy/function/MulOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ab08b6f8696ff4aefd2dbdda591b20730b46898c --- /dev/null +++ b/paddle/legacy/function/MulOpTest.cpp @@ -0,0 +1,212 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "FunctionTest.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/math/tests/test_matrixUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT + +/** + * C += A * B, A, B, C dense matrix + * dense = dense * dense + */ +void testFuncDDDMatrix( + bool transa, bool transb, size_t dimM, size_t dimN, size_t dimK) { + real scaleT = 1.0; + size_t heightA = (transa == false) ? dimM : dimK; + size_t widthA = (transa == false) ? dimK : dimM; + size_t heightB = (transb == false) ? dimK : dimN; + size_t widthB = (transb == false) ? dimN : dimK; + size_t heightC = dimM; + size_t widthC = dimN; + // init Test object + CpuGpuFuncCompare test( + "MulOp", FuncConfig().set("aTrans", transa).set("bTrans", transb)); + // prepare input arguments + /// matrix A : HA * WA + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{heightA, widthA})); + /// matrix B: HB * WB + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{heightB, widthB})); + + /// output matrix C: HC * WC + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{heightC, widthC}), + scaleT == 1.0 ? ADD_TO : ASSIGN_TO); + // run Function + test.run(); +} + +TEST(MulOp, DDDMatrixMul) { + LOG(INFO) << "function test for dense = dense * dense matrix"; + for (const auto transa : {false, true}) { + for (const auto transb : {false, true}) { + for (const auto dimM : {1, 10, 100}) { + for (const auto dimN : {1, 10}) { + for (const auto dimK : {8}) { + if (transa && transb) { + continue; + } + VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') + << " transa=" << transa << " transb=" << transb + << " dimM=" << std::setw(5) << dimM + << " dimN=" << std::setw(5) << dimN + << " dimK=" << std::setw(5) << dimK; + testFuncDDDMatrix(transa, transb, dimM, dimN, dimK); + } + } + } + } + } +} + +/** + * C += A * B, B, C dense, A sparse + * dense = sparse * dense + */ +void testFuncDSparseDMatrix( + size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) { + real scaleT = 1.0; + // init Test object + CpuGpuFuncCompare test( + "MulOp", FuncConfig().set("aTrans", false).set("bTrans", false)); + // prepare input arguments + /// sparse matrix A : M * K + test.addInputs(SparseMatrixArg( + VALUE_TYPE_FLOAT, TensorShape{dimM, dimK}, nnz, FORMAT, FLOAT_VALUE)); + /// matrix B: K * N + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimK, dimN})); + + /// output matrix C: M * N + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimN}), + scaleT == 1.0 ? ADD_TO : ASSIGN_TO); + // run Function + test.run(); +} + +TEST(MuLOp, DSparseDMul) { + LOG(INFO) << "function test for dense = sparse * dense matrix"; + for (const auto dimM : {10, 100, 1000}) { + for (const auto dimN : {10, 100}) { + for (const auto dimK : {3, 10}) { + for (const auto nnz : {3, 10}) { + for (const auto FORMAT : {SPARSE_CSR}) { + VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') + << " dimM=" << std::setw(5) << dimM + << " dimN=" << std::setw(5) << dimN + << " dimK=" << std::setw(5) << dimK + << " nnz=" << std::setw(5) << nnz + << " format=" << std::setw(5) << FORMAT; + testFuncDSparseDMatrix(dimM, dimN, dimK, nnz, FORMAT); + } + } + } + } + } +} + +/** + * C += A * B, A, C dense, B sparse + * dense = dense * sparse + */ +void testFuncDDSparseMatrix( + size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) { + real scaleT = 1.0; + // init Test object + CpuGpuFuncCompare test( + "MulOp", FuncConfig().set("aTrans", false).set("bTrans", false)); + // prepare input arguments + /// matrix A : M * K + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimK})); + + /// matrix B: K * N + test.addInputs(SparseMatrixArg( + VALUE_TYPE_FLOAT, TensorShape{dimK, dimN}, nnz, FORMAT, FLOAT_VALUE)); + + /// output matrix C: M * N + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimN}), + scaleT == 1.0 ? ADD_TO : ASSIGN_TO); + // run Function + test.run(); +} + +TEST(MulOp, DDSparseMul) { + LOG(INFO) << "function test for dense = dense * sparse matrix"; + for (const auto dimM : {10, 100, 1000}) { + for (const auto dimN : {10, 100}) { + for (const auto dimK : {3, 10}) { + for (const auto nnz : {3, 10}) { + for (const auto FORMAT : {SPARSE_CSR, SPARSE_CSC}) { + VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') + << " dimM=" << std::setw(5) << dimM + << " dimN=" << std::setw(5) << dimN + << " dimK=" << std::setw(5) << dimK + << " nnz=" << std::setw(5) << nnz + << " format=" << std::setw(5) << FORMAT; + testFuncDDSparseMatrix(dimM, dimN, dimK, nnz, FORMAT); + } + } + } + } + } +} + +/** + * C += A * B, A sparse, B, C dense + * sparse = dense * dense + */ +void testFuncSparseDDMatrix( + size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) { + real scaleT = 1.0; + // init Test object + CpuGpuFuncCompare test( + "MulOp", FuncConfig().set("aTrans", false).set("bTrans", false)); + // prepare input arguments + /// matrix A : M * K + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimM, dimK})); + + /// matrix B: K * N + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{dimK, dimN})); + + /// output sparse matrix C: M * N + test.addOutputs( + SparseMatrixArg( + VALUE_TYPE_FLOAT, TensorShape{dimM, dimN}, nnz, FORMAT, FLOAT_VALUE), + scaleT == 1.0 ? ADD_TO : ASSIGN_TO); + // run Function + test.run(); +} + +TEST(MulOp, SparseDDMul) { + LOG(INFO) << "function test for sparse = dense * dense matrix"; + for (const auto dimM : {10, 100, 1000}) { + for (const auto dimN : {10, 100}) { + for (const auto dimK : {3, 10}) { + for (const auto nnz : {3, 10}) { + for (const auto FORMAT : {SPARSE_CSC, SPARSE_CSR}) { + VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ') + << " dimM=" << std::setw(5) << dimM + << " dimN=" << std::setw(5) << dimN + << " dimK=" << std::setw(5) << dimK + << " nnz=" << std::setw(5) << nnz + << " format=" << std::setw(5) << FORMAT; + testFuncSparseDDMatrix(dimM, dimN, dimK, nnz, FORMAT); + } + } + } + } + } +} diff --git a/paddle/function/NaiveConvOp.cpp b/paddle/legacy/function/NaiveConvOp.cpp similarity index 100% rename from paddle/function/NaiveConvOp.cpp rename to paddle/legacy/function/NaiveConvOp.cpp diff --git a/paddle/legacy/function/PadOp.cpp b/paddle/legacy/function/PadOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9d011d28e6938fac6980bed88f774abdbf3532d4 --- /dev/null +++ b/paddle/legacy/function/PadOp.cpp @@ -0,0 +1,215 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PadOp.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { + +template <> +void Pad(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const PadConf& pad) { + int cstart = pad.channel[0], cend = pad.channel[1]; + int hstart = pad.height[0], hend = pad.height[1]; + int wstart = pad.width[0], wend = pad.width[1]; + int outC = inC + cstart + cend; + int outH = inH + hstart + hend; + int outW = inW + wstart + wend; + for (int i = 0; i < num; i++) { + for (int c = 0; c < inC; c++) { + for (int h = 0; h < inH; h++) { + int inoff = ((i * inC + c) * inH + h) * inW; + int outoff = + ((i * outC + c + cstart) * outH + h + hstart) * outW + wstart; + memcpy(outputs + outoff, inputs + inoff, inW * sizeof(real)); + } + } + } +} + +template <> +void PadGrad(real* inGrad, + const real* outGrad, + const int num, + const int inC, + const int inH, + const int inW, + const PadConf& pad) { + int cstart = pad.channel[0], cend = pad.channel[1]; + int hstart = pad.height[0], hend = pad.height[1]; + int wstart = pad.width[0], wend = pad.width[1]; + int outC = inC + cstart + cend; + int outH = inH + hstart + hend; + int outW = inW + wstart + wend; + for (int i = 0; i < num; i++) { + for (int c = 0; c < inC; c++) { + for (int h = 0; h < inH; h++) { + int inoff = ((i * inC + c) * inH + h) * inW; + int outoff = + ((i * outC + c + cstart) * outH + h + hstart) * outW + wstart; + CpuVector inG = CpuVector(inW, inGrad + inoff); + CpuVector outG = CpuVector(inW, const_cast(outGrad + outoff)); + inG += outG; + } + } + } +} + +static inline PadConf castToPadConf(const FuncConfig& conf) { + return {conf.get>("channel"), + conf.get>("height"), + conf.get>("width")}; +} + +/** + * \brief Padding zeros to input according to the specify dimension. + * The struct pad_ contains the padding size in each dimension. + * The input and output is a 4D tensor. In PadFunc, we only + * pad zeros to the 2nd to 4th dimension. + * + * Argument in this Function: + * \param pad_ A struct object contains the padding size in each dimension. + * It has six integers. The channelStart and channelEnd indicate + * how many zeros to add before and after the input in channel + * dimension. And the heightStart and heightEnd indicate padding + * in height dimension. The widthStart and widthEnd indicate the + * padding in width dimension. + * \param inputs A 4D tensor, only one input. + * \param outputs A 4D tensor, the output value after padding. + * + * For example, + * Input(2,2,2,3) = [ + * [ [[1,2,3], [3,4,5]], + * [[2,3,5], [1,6,7]] ], + * [ [[4,3,1], [1,8,7]], + * [[3,8,9], [2,3,5]] ] + * ] # the shape is (1,2,2,3) + * + * pad_: if channelStart = channelEnd = 1, others are 0. + * Output(2,4,2,3) = [ + * [ [[0,0,0], [0,0,0]], + * [[1,2,3], [3,4,5]], + * [[2,3,5], [1,6,7]], + * [[0,0,0], [0,0,0]] ], + * [ [[0,0,0], [0,0,0]], + * [[4,3,1], [1,8,7]], + * [[3,8,9], [2,3,5]], + * [[0,0,0], [0,0,0]] ] + * ] # the shape is (2,4,2,3) + * + * pad_: if widthStart = 1, widthEnd = 2, others are 0. + * Output(2,2,2,6) = [ + * [ [[0,1,2,3,0,0], [0,3,4,5,0,0]], + * [[0,2,3,5,0,0], [0,1,6,7,0,0]] ], + * [ [[0,4,3,1,0,0], [0,1,8,7,0,0]], + * [[0,3,8,9,0,0], [0,2,3,5,0,0]] ], + * ] # the shape is (2,2,2,6) + * + * pad_: if heightStart = 1, heightEnd = 1, others are 0. + * Output(2,2,4,3) = [ + * [ [[0,0,0], [1,2,3], [3,4,5], [0,0,0]], + * [[0,0,0], [2,3,5], [1,6,7], [0,0,0]] ], + * [ [[0,0,0], [4,3,1], [1,8,7], [0,0,0]], + * [[0,0,0], [3,8,9], [2,3,5], [0,0,0]] ], + * ] # the shape is (2,2,4,3) + */ + +template +class PadFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { pad_ = castToPadConf(config); } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + + size_t num = inputs[0].shape()[0]; + size_t inC = inputs[0].shape()[1]; + size_t inH = inputs[0].shape()[2]; + size_t inW = inputs[0].shape()[3]; + typename Tensor::Vector vec(outputs[0].shape().getElements(), + outputs[0].data()); + vec.zero(); + + Pad(outputs[0].data(), + inputs[0].data(), + num, + inC, + inH, + inW, + pad_); + } + + private: + PadConf pad_; +}; + +/** + * \brief The backward propagation of padding Function. Remove the elements + * in the padding positions of forward. + * + * Argument in this Function: + * \param pad_ The same meaning as it in PadFunc. + * \param inputs The gradient with respect to the output value of PadFunc. + * \param outputs The gradient with respect to the input value of PadFunc. + */ + +template +class PadGradFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { pad_ = castToPadConf(config); } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = outputs[0].shape()[0]; + size_t inC = outputs[0].shape()[1]; + size_t inH = outputs[0].shape()[2]; + size_t inW = outputs[0].shape()[3]; + + if (outputs[0].getArgType() != ADD_TO) { + // for unit test + typename Tensor::Vector tmp( + outputs[0].shape().getElements(), outputs[0].data()); + tmp.zero(); + } + + PadGrad(outputs[0].data(), + inputs[0].data(), + num, + inC, + inH, + inW, + pad_); + } + + private: + PadConf pad_; +}; + +REGISTER_TYPED_FUNC(Pad, CPU, PadFunc); +REGISTER_TYPED_FUNC(PadGrad, CPU, PadGradFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(Pad, GPU, PadFunc); +REGISTER_TYPED_FUNC(PadGrad, GPU, PadGradFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/PadOp.h b/paddle/legacy/function/PadOp.h similarity index 100% rename from paddle/function/PadOp.h rename to paddle/legacy/function/PadOp.h diff --git a/paddle/function/PadOpGpu.cu b/paddle/legacy/function/PadOpGpu.cu similarity index 100% rename from paddle/function/PadOpGpu.cu rename to paddle/legacy/function/PadOpGpu.cu diff --git a/paddle/function/PadOpTest.cpp b/paddle/legacy/function/PadOpTest.cpp similarity index 100% rename from paddle/function/PadOpTest.cpp rename to paddle/legacy/function/PadOpTest.cpp diff --git a/paddle/legacy/function/RowConvOp.cpp b/paddle/legacy/function/RowConvOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3be50e80d71fabdb3e7a22bfc061da09412c132d --- /dev/null +++ b/paddle/legacy/function/RowConvOp.cpp @@ -0,0 +1,225 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "RowConvOp.h" +#include +#include "paddle/legacy/math/Vector.h" + +namespace paddle { + +template <> +void RowConv(CpuMatrix& out, + const CpuMatrix& in, + const CpuMatrix& filter, + const CpuIVector& seq) { + const int* starts = seq.getData(); + const size_t numSeq = seq.getSize() - 1; + const size_t contextLength = filter.getHeight(); + for (size_t i = 0; i < numSeq; ++i) { + size_t begin = starts[i]; + size_t end = starts[i + 1]; + for (size_t j = begin; j < end; ++j) { + MatrixPtr x; + MatrixPtr w; + if ((j + contextLength) < end) { + x = (const_cast(in)).subMatrix(j, contextLength); + w = (const_cast(filter)).subMatrix(0, contextLength); + } else { + x = (const_cast(in)).subMatrix(j, end - j); + w = (const_cast(filter)).subMatrix(0, end - j); + } + MatrixPtr y = out.subMatrix(j, 1); + y->addDotMulVMM(*x, *w); + } + } +} + +template <> +void RowConvGrad(const CpuMatrix& outG, + const CpuMatrix& in, + const CpuMatrix& filter, + CpuMatrix& inG, + CpuMatrix& filterG, + const CpuIVector& seq) { + // gradient w.r.t filter + const int* starts = seq.getData(); + const size_t numSeq = seq.getSize() - 1; + const size_t contextLength = filter.getHeight(); + if (filterG) { + for (size_t i = 0; i < numSeq; ++i) { + size_t begin = starts[i]; + size_t end = starts[i + 1]; + size_t steps = end - begin; + for (size_t j = 0; j < contextLength && (begin + j) < end; ++j) { + MatrixPtr x = + (const_cast(in)).subMatrix(begin + j, steps - j); + MatrixPtr dy = + (const_cast(outG)).subMatrix(begin, steps - j); + MatrixPtr dw = filterG.subMatrix(j, 1); + dw->addDotMulVMM(*dy, *x); + } + } + } + + // gradient w.r.t input feature + if (inG) { + for (size_t i = 0; i < numSeq; ++i) { + size_t begin = starts[i]; + size_t end = starts[i + 1]; + size_t steps = end - begin; + for (size_t j = 0; j < steps; ++j) { + MatrixPtr dx = inG.subMatrix(begin + j, 1); + for (size_t t = 0; t < contextLength; ++t) { + if (int(j - t) >= 0) { + MatrixPtr dy = + (const_cast(outG)).subMatrix(begin + j - t, 1); + MatrixPtr w = (const_cast(filter)).subMatrix(t, 1); + dx->addDotMul(*dy, *w, 1.0, 1.0); + } + } + } + } + } +} + +/** + * \brief The row convolution is called lookahead convolution. It is firstly + * introduced in deep-speech2 system. The bidirectional RNN that learns + * representation for a sequence by performing a forward and a backward pass + * through the entire sequence. However, unlike unidirectional RNNs, + * bidirectional RNNs are challenging to deploy in an online and low-latency + * setting. The lookahead convolution incorporates information from future + * subsequences in a computationally efficient manner to improve unidirectional + * recurrent neural networks. + * + * The connection of row convolution is different form the 1D sequence + * convolution. Assumed that, the future context-length is k, that is to say, + * it can get the output at timestep t by using the the input feature from t-th + * timestep to (t+k)-th timestep. Assumed that the hidden dim of input + * activations are d, the activations r_t for the new layer at time-step t are: + * + * + * -- k + 1 + * r(t,i) = > W(i,j) * h(t+j-1, i), for (1 <= i <= d) + * -- j = 1 + * + * + * The weight shape is: (k + 1) x d + * Function Arguments: + * + * \param inputs[0] The input activations. + * \param inputs[0] The filter (or weight) and shape is (k+1) x d. + * \param outputs[1] The output activations. + * + * [1] Dario Amodei, etc. Deep Speech 2 : End-to-End Speech Recognition in + * English + * and Mandarin. https://arxiv.org/abs/1512.02595 + */ + +template +class RowConvFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + // check + CHECK_EQ(2UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + // TODO(qingqing): support ASSIGN_TO. + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg()) + << "SequenceArg required here."; + const auto in = dynamic_cast(inputs[0]); + auto out = dynamic_cast(outputs[0]); + auto w = inputs[1]; + CHECK(in.data() && out.data() && in.getSequenceId().data()); + CHECK_EQ(in.shape().ndims(), 2UL); + CHECK(in.shape() == out.shape()); + CHECK_EQ(w.shape()[1], in.shape()[1]); + + auto outMat = out.matrix(); + const auto inMat = in.matrix(); + const auto wMat = w.matrix(); + const auto seqId = in.getSequenceId().vector(); + + RowConv(outMat, inMat, wMat, seqId); + } +}; + +/** + * \brief The backward of row convolution function. This function calculated + * the gradient w.r.t filter and the gradient w.r.t input activations(or data). + * + * Argument in this Function: + * + * \param inputs[0] The gradient w.r.t output activations. + * \param inputs[1] The input activations. + * \param inputs[2] The filter (or weight) and shape is (k+1) x d. + * \param outputs[0] The gradient w.r.t input activations. + * \param outputs[1] The gradient w.r.r filter. + * + * Abbreviation: + * w.r.t: with respect to. + */ + +template +class RowConvGradFunc : public FunctionBase { + // TODO(qingqing): split into RowConvDataFunc and RowConvWeightFunc + public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + // check + CHECK_EQ(3UL, inputs.size()); + CHECK_EQ(2UL, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + CHECK_EQ(outputs[1].getArgType(), ADD_TO); + CHECK(inputs[0].isSequenceArg() && inputs[1].isSequenceArg() && + outputs[0].isSequenceArg()) + << "SequenceArg required here."; + + const auto outGrad = dynamic_cast(inputs[0]); + const auto in = dynamic_cast(inputs[1]); + const auto w = inputs[2]; + auto inGrad = dynamic_cast(outputs[0]); + auto wGrad = outputs[1]; + + CHECK_EQ(in.shape().ndims(), 2UL); + CHECK(in.shape() == inGrad.shape()); + CHECK(in.shape() == outGrad.shape()); + CHECK_EQ(wGrad.shape()[1], in.shape()[1]); + + const auto outGMat = outGrad.matrix(); + const auto inMat = in.matrix(); + const auto wMat = w.matrix(); + auto inGMat = inGrad.data() + ? inGrad.matrix() + : typename Tensor::Matrix(nullptr, 0, 0); + auto wGMat = wGrad.data() + ? wGrad.matrix() + : typename Tensor::Matrix(nullptr, 0, 0); + const auto seqId = in.getSequenceId().vector(); + + RowConvGrad(outGMat, inMat, wMat, inGMat, wGMat, seqId); + } +}; + +REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc); +REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc); +REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/RowConvOp.h b/paddle/legacy/function/RowConvOp.h similarity index 100% rename from paddle/function/RowConvOp.h rename to paddle/legacy/function/RowConvOp.h diff --git a/paddle/legacy/function/RowConvOpGpu.cu b/paddle/legacy/function/RowConvOpGpu.cu new file mode 100644 index 0000000000000000000000000000000000000000..a6d2e4c7e38b12bcd448a85f9e74df226e6984af --- /dev/null +++ b/paddle/legacy/function/RowConvOpGpu.cu @@ -0,0 +1,373 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/cuda/include/hl_base.h" +#include "paddle/legacy/function/RowConvOp.h" + +namespace paddle { + +template +__global__ void KeRowConv(real* y, + const real* x, + const real* w, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { + const int tidx = threadIdx.x; + const int tidy = threadIdx.y; + const int blky = blockDim.y; + const int gidx = blockIdx.x * blockDim.x; + + __shared__ real sw[BLOCK_H][BLOCK_W]; + + for (int i = tidy; i < context; i += blky) { + sw[i][tidx] = gidx + tidx < width ? w[i * width + gidx + tidx] : 0.0; + } + + __syncthreads(); + + for (int i = 0; i < numSeq; ++i) { + const int start = starts[i]; + const int end = starts[i + 1]; + const int steps = end - start; + for (int j = tidy; j < steps; j += blky) { + real sum = 0; + int off = (start + j) * width; + for (int t = 0; t < context; ++t) { + if ((start + j + t) < end) { + int xoff = off + t * width; + real xVal = gidx + tidx < width ? x[xoff + gidx + tidx] : 0.0; + sum += sw[t][tidx] * xVal; + } + } + if (gidx + tidx < width) { + y[off + gidx + tidx] += sum; + } + } + } +} + +__global__ void KeRowConv2(real* y, + const real* x, + const real* w, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { + const int tidx = threadIdx.x; + const int tidy = threadIdx.y; + const int blky = blockDim.y; + const int gidx = blockIdx.x * blockDim.x; + + for (int i = 0; i < numSeq; ++i) { + const int start = starts[i]; + const int end = starts[i + 1]; + const int steps = end - start; + for (int j = tidy; j < steps; j += blky) { + int off = (start + j) * width; + real sum = 0; + for (int t = 0; t < context && (start + j + t) < end; ++t) { + int xoff = off + t * width; + real xd = gidx + tidx < width ? x[xoff + gidx + tidx] : 0.0; + real wd = gidx + tidx < width ? w[t * width + gidx + tidx] : 0.0; + sum += wd * xd; + } + if (gidx + tidx < width) { + y[off + gidx + tidx] += sum; + } + } + } +} + +template <> +void RowConv(GpuMatrix& out, // NOLINT + const GpuMatrix& in, + const GpuMatrix& filter, + const GpuIVector& seq) { + const size_t numSeq = seq.getSize() - 1; + const size_t contextLength = filter.getHeight(); + const size_t height = in.getHeight(); + const size_t width = in.getWidth(); + + real* y = out.getData(); + const real* x = in.getData(); + const real* w = filter.getData(); + const int* starts = seq.getData(); + + dim3 dimBlock(32, 32); + dim3 dimGrid(DIVUP(width, dimBlock.x), 1); + + if (contextLength <= 32) { + KeRowConv<32, 32><<>>( + y, x, w, starts, height, width, numSeq, contextLength); + } else { + KeRowConv2<<>>( + y, x, w, starts, height, width, numSeq, contextLength); + } + CHECK_SYNC("RowConv"); +} + +template +__global__ void KeRowConvBwWeight(real* dw, + const real* x, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { + const int tidx = threadIdx.x; + const int tidy = threadIdx.y; + const int blky = blockDim.y; + const int gidx = blockIdx.x * blockDim.x; + + __shared__ real sh_x[BLOCK_W][BLOCK_H]; + __shared__ real sh_dy[BLOCK_W][BLOCK_H + CONTEXT - 1]; + __shared__ real sh_dw[CONTEXT][BLOCK_W]; + + if (tidy < context) { + sh_dw[tidy][tidx] = 0.0; + } + __syncthreads(); + + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + + for (int i = 0; i < numSeq; ++i) { + const int start = starts[i]; + const int end = starts[i + 1]; + const int steps = end - start; + const int size = ((steps + BLOCK_H - 1) / BLOCK_H) * BLOCK_H; + for (int j = tidy; j < size; j += BLOCK_H) { + int xoff = gidx + tidx; + int yoff = start + j; + + // transpose + sh_x[tidx][tidy] = + (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0; + sh_dy[tidx][tidy + context - 1] = + (xoff < width && yoff < end) ? dy[yoff * width + xoff] : 0.0; + __syncthreads(); + if (tidy < (context - 1)) { + yoff = yoff - context + 1; + sh_dy[tidx][tidy] = + (xoff < width && yoff >= start) ? dy[yoff * width + xoff] : 0.0; + } + __syncthreads(); + + for (int t = 0; t < context; t++) { + real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx + context - 1 - t]; + __syncthreads(); + // warp size and blockDim.x is 32. + + for (int offset = 16; offset > 0; offset /= 2) + val += __shfl_down_sync(mask, val, offset); + + __syncthreads(); + if (tidx == 0) { + sh_dw[t][tidy] += val; + } + __syncthreads(); + } + } + } + + for (int t = tidy; (t < context) && ((gidx + tidx) < width); t += blky) { + dw[t * width + gidx + tidx] += sh_dw[t][tidx]; + } +} + +template +__global__ void KeRowConvBwWeight2(real* dw, + const real* x, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { + const int tidx = threadIdx.x; + const int tidy = threadIdx.y; + const int gidx = blockIdx.x * blockDim.x; + + __shared__ real sh_x[BLOCK_H][BLOCK_W]; + __shared__ real sh_dy[BLOCK_H][BLOCK_W]; + + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + + for (int i = 0; i < numSeq; ++i) { + const int start = starts[i]; + const int end = starts[i + 1]; + const int steps = end - start; + + const int size = ((steps + BLOCK_H - 1) / BLOCK_H) * BLOCK_H; + for (int j = tidy; j < size; j += BLOCK_H) { + int xoff = gidx + tidx; + int yoff = start + j; + + // transpose + sh_x[tidx][tidy] = + (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0; + __syncthreads(); + + for (int t = 0; t < context; t++) { + sh_dy[tidx][tidy] = + (xoff < width && (yoff - t) >= start && yoff - t < end) + ? dy[(yoff - t) * width + xoff] + : 0.0; + __syncthreads(); + + real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx]; + __syncthreads(); + // warp size and blockDim.x is 32. + for (int offset = 16; offset > 0; offset /= 2) + val += __shfl_down_sync(mask, val, offset); + + __syncthreads(); + + if (tidx == 0 && (gidx + tidy) < width) { + dw[t * width + gidx + tidy] += val; + } + } + } + } +} + +template +__global__ void KeRowConvBwData(real* dx, + const real* w, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { + const int tidx = threadIdx.x; + const int tidy = threadIdx.y; + const int blky = blockDim.y; + const int gidx = blockIdx.x * blockDim.x; + + __shared__ real sw[BLOCK_H][BLOCK_W]; + + for (int i = tidy; i < context; i += blky) { + sw[i][tidx] = gidx + tidx < width ? w[i * width + gidx + tidx] : 0.0; + } + + __syncthreads(); + + for (int i = 0; i < numSeq; ++i) { + const int start = starts[i]; + const int end = starts[i + 1]; + const int steps = end - start; + for (int j = tidy; j < steps; j += blky) { + real sum = 0; + int off = (start + j) * width; + for (int t = 0; t < context && (j - t) >= 0; ++t) { + int dyOff = off - t * width; + real dyVal = gidx + tidx < width ? dy[dyOff + gidx + tidx] : 0.0; + sum += sw[t][tidx] * dyVal; + } + if (gidx + tidx < width) { + dx[off + gidx + tidx] += sum; + } + } + } +} + +__global__ void KeRowConvBwData2(real* dx, + const real* w, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { + const int tidx = threadIdx.x; + const int tidy = threadIdx.y; + const int blky = blockDim.y; + const int gidx = blockIdx.x * blockDim.x; + + for (int i = 0; i < numSeq; ++i) { + const int start = starts[i]; + const int end = starts[i + 1]; + const int steps = end - start; + for (int j = tidy; j < steps; j += blky) { + real sum = 0; + int off = (start + j) * width; + for (int t = 0; t < context && (j - t) >= 0; ++t) { + int dyOff = off - t * width; + real dyVal = gidx + tidx < width ? dy[dyOff + gidx + tidx] : 0.0; + real wVal = gidx + tidx < width ? w[t * width + gidx + tidx] : 0.0; + sum += wVal * dyVal; + } + if (gidx + tidx < width) { + dx[off + gidx + tidx] += sum; + } + } + } +} + +template <> +void RowConvGrad(const GpuMatrix& outG, + const GpuMatrix& in, + const GpuMatrix& filter, + GpuMatrix& inG, // NOLINT + GpuMatrix& filterG, // NOLINT + const GpuIVector& seq) { + const size_t numSeq = seq.getSize() - 1; + const size_t contextLength = filter.getHeight(); + const size_t height = in.getHeight(); + const size_t width = in.getWidth(); + + const real* dy = outG.getData(); + const real* x = in.getData(); + const real* w = filter.getData(); + const int* starts = seq.getData(); + + if (filterG) { + dim3 dimBlock(32, 32); + dim3 dimGrid(DIVUP(width, dimBlock.x), 1); + real* dw = filterG.getData(); + if (contextLength <= 32) { + KeRowConvBwWeight<32, 32, 32><<>>( + dw, x, dy, starts, height, width, numSeq, contextLength); + } else { + KeRowConvBwWeight2<32, 32><<>>( + dw, x, dy, starts, height, width, numSeq, contextLength); + } + } + + if (inG) { + real* dx = inG.getData(); + dim3 dimBlock2(32, 32); + dim3 dimGrid2(DIVUP(width, dimBlock2.x), 1); + if (contextLength <= 64) { + KeRowConvBwData<32, 64><<>>( + dx, w, dy, starts, height, width, numSeq, contextLength); + } else { + KeRowConvBwData2<<>>( + dx, w, dy, starts, height, width, numSeq, contextLength); + } + } + + CHECK_SYNC("RowConvGrad"); +} + +} // namespace paddle diff --git a/paddle/function/RowConvOpTest.cpp b/paddle/legacy/function/RowConvOpTest.cpp similarity index 100% rename from paddle/function/RowConvOpTest.cpp rename to paddle/legacy/function/RowConvOpTest.cpp diff --git a/paddle/legacy/function/ScaleSubRegionOp.cpp b/paddle/legacy/function/ScaleSubRegionOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..03a422a740dca4499532cdb1bdfbf3d3ab272a9a --- /dev/null +++ b/paddle/legacy/function/ScaleSubRegionOp.cpp @@ -0,0 +1,155 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ScaleSubRegionOp.h" +#include "paddle/legacy/function/TensorShape.h" + +namespace paddle { + +template <> +void ScaleSubRegion(real* outputs, + const real* inputs, + const real* indices, + const TensorShape shape, + const FuncConfig& conf) { + real value = conf.get("value"); + + int number = shape[0]; + int channel = shape[1]; + int height = shape[2]; + int width = shape[3]; + + memcpy(outputs, inputs, number * channel * height * width * sizeof(real)); + + for (int n = 0; n < number; ++n) { + // indices start from 1 + int offset = n * 6; + for (int c = indices[offset] - 1; c < indices[offset + 1]; ++c) { + for (int h = indices[offset + 2] - 1; h < indices[offset + 3]; ++h) { + for (int w = indices[offset + 4] - 1; w < indices[offset + 5]; ++w) { + int idx = ((n * channel + c) * height + h) * width + w; + outputs[idx] *= value; + } + } + } + } +} + +template <> +void ScaleSubRegionGrad(const real* inGrad, + real* outGrad, + const real* indices, + const TensorShape shape, + const FuncConfig& conf) { + real value = conf.get("value"); + + int number = shape[0]; + int channel = shape[1]; + int height = shape[2]; + int width = shape[3]; + + for (int n = 0; n < number; ++n) { + for (int c = 0; c < channel; ++c) { + for (int h = 0; h < height; ++h) { + for (int w = 0; w < width; ++w) { + int idx = ((n * channel + c) * height + h) * width + w; + int offset = n * 6; + if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) && + h >= (indices[offset + 2] - 1) && + h <= (indices[offset + 3] - 1) && + w >= (indices[offset + 4] - 1) && + w <= (indices[offset + 5] - 1)) { + outGrad[idx] += inGrad[idx] * value; + } else { + outGrad[idx] += inGrad[idx]; + } + } + } + } + } +} + +/** + * \brief For each instance, ScaleSubRegion can be used to multiply a value to + * a specified sub continuous region. By providing start index and end + * index for C/H/W, you can specify the location and shape of the region. + * + * Argument in this Function: + * \param inputs A 4-D tensor with shape [N, C, H, W], only one input. + * \param indices A 2-D tensor with shape [N, 6], indicates the sub region. + * \param outputs A 4-D tensor with same shape as inputs, output value. + */ +template +class ScaleSubRegionFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { conf_ = config; } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(2UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + + TensorShape shape = inputs[0].shape(); + + ScaleSubRegion(outputs[0].data(), + inputs[0].data(), + inputs[1].data(), + shape, + conf_); + } + + private: + FuncConfig conf_; +}; + +/** + * \brief The backward propagation of ScaleSubRegion Function. + * + * Argument in this Function: + * \param inputs A 4-D tensor with shape [N, C, H, W], output gradient. + * \param indices A 2-D tensor with shape [N, 6], indicates the sub region. + * \param outputs A 4-D tensor with shape [N, C, H, W], gradient of input value. + */ + +template +class ScaleSubRegionGradFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override { conf_ = config; } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(2UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ADD_TO); + + TensorShape shape = inputs[0].shape(); + + ScaleSubRegionGrad(inputs[0].data(), + outputs[0].data(), + inputs[1].data(), + shape, + conf_); + } + + private: + FuncConfig conf_; +}; + +REGISTER_TYPED_FUNC(ScaleSubRegion, CPU, ScaleSubRegionFunc); +REGISTER_TYPED_FUNC(ScaleSubRegionGrad, CPU, ScaleSubRegionGradFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(ScaleSubRegion, GPU, ScaleSubRegionFunc); +REGISTER_TYPED_FUNC(ScaleSubRegionGrad, GPU, ScaleSubRegionGradFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/ScaleSubRegionOp.h b/paddle/legacy/function/ScaleSubRegionOp.h similarity index 100% rename from paddle/function/ScaleSubRegionOp.h rename to paddle/legacy/function/ScaleSubRegionOp.h diff --git a/paddle/function/ScaleSubRegionOpGpu.cu b/paddle/legacy/function/ScaleSubRegionOpGpu.cu similarity index 100% rename from paddle/function/ScaleSubRegionOpGpu.cu rename to paddle/legacy/function/ScaleSubRegionOpGpu.cu diff --git a/paddle/function/ScaleSubRegionOpTest.cpp b/paddle/legacy/function/ScaleSubRegionOpTest.cpp similarity index 100% rename from paddle/function/ScaleSubRegionOpTest.cpp rename to paddle/legacy/function/ScaleSubRegionOpTest.cpp diff --git a/paddle/legacy/function/SwitchOp.cpp b/paddle/legacy/function/SwitchOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c6accd18039180aa521c18193e576d22e11f5a97 --- /dev/null +++ b/paddle/legacy/function/SwitchOp.cpp @@ -0,0 +1,140 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SwitchOp.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { + +template <> +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argType) { + for (int n = 0; n < num; ++n) { + for (int c = 0; c < inC; ++c) { + for (int h = 0; h < inH; ++h) { + for (int w = 0; w < inW; ++w) { + if (argType == ADD_TO) { + outputs[((n * inH + h) * inW + w) * inC + c] += *(inputs++); + } else { + outputs[((n * inH + h) * inW + w) * inC + c] = *(inputs++); + } + } + } + } + } +} + +template <> +void NHWC2NCHW(real* outputs, + const real* inputs, + const int num, + const int inH, + const int inW, + const int inC, + const int argType) { + for (int n = 0; n < num; ++n) { + for (int h = 0; h < inH; ++h) { + for (int w = 0; w < inW; ++w) { + for (int c = 0; c < inC; ++c) { + if (argType == ADD_TO) { + outputs[((n * inC + c) * inH + h) * inW + w] += *(inputs++); + } else { + outputs[((n * inC + c) * inH + h) * inW + w] = *(inputs++); + } + } + } + } + } +} + +/** + * \brief Switch dimension order of image input. + * The input and output is a 4D tensor. Switch order + * 'batch_size,channels, height, width' to + * order 'batch_size, height, width, channels'. + * + * Argument in this Function: + * \param inputs input data with order 'batch_size,channels, height, width'. + * \param outputs output data with order 'batch_size, height, width, channels'. + */ +template +class NCHW2NHWCFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = inputs[0].shape()[0]; + size_t inC = inputs[0].shape()[1]; + size_t inH = inputs[0].shape()[2]; + size_t inW = inputs[0].shape()[3]; + NCHW2NHWC(outputs[0].data(), + inputs[0].data(), + num, + inC, + inH, + inW, + outputs[0].getArgType()); + } +}; + +/** + * \brief Switch dimension order of image input. + * The input and output is a 4D tensor. Switch order + * 'batch_size, height, width, channels' to + * order 'batch_size, channels, height, width'. + * + * Argument in this Function: + * \param inputs input data with order 'batch_size, height, width, channels'. + * \param outputs output data with order 'batch_size, channels, height, width'. + */ +template +class NHWC2NCHWFunc : public FunctionBase { + public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = inputs[0].shape()[0]; + size_t inH = inputs[0].shape()[1]; + size_t inW = inputs[0].shape()[2]; + size_t inC = inputs[0].shape()[3]; + + NHWC2NCHW(outputs[0].data(), + inputs[0].data(), + num, + inH, + inW, + inC, + outputs[0].getArgType()); + } +}; + +REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc); +REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc); +#ifdef PADDLE_WITH_CUDA +REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc); +REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/SwitchOp.h b/paddle/legacy/function/SwitchOp.h similarity index 100% rename from paddle/function/SwitchOp.h rename to paddle/legacy/function/SwitchOp.h diff --git a/paddle/function/SwitchOpGpu.cu b/paddle/legacy/function/SwitchOpGpu.cu similarity index 100% rename from paddle/function/SwitchOpGpu.cu rename to paddle/legacy/function/SwitchOpGpu.cu diff --git a/paddle/function/SwitchOpTest.cpp b/paddle/legacy/function/SwitchOpTest.cpp similarity index 100% rename from paddle/function/SwitchOpTest.cpp rename to paddle/legacy/function/SwitchOpTest.cpp diff --git a/paddle/function/TensorShape.h b/paddle/legacy/function/TensorShape.h similarity index 100% rename from paddle/function/TensorShape.h rename to paddle/legacy/function/TensorShape.h diff --git a/paddle/function/TensorShapeTest.cpp b/paddle/legacy/function/TensorShapeTest.cpp similarity index 100% rename from paddle/function/TensorShapeTest.cpp rename to paddle/legacy/function/TensorShapeTest.cpp diff --git a/paddle/legacy/function/TensorType.h b/paddle/legacy/function/TensorType.h new file mode 100644 index 0000000000000000000000000000000000000000..13994821be7ba7264f43d8550e6800cdc5b93875 --- /dev/null +++ b/paddle/legacy/function/TensorType.h @@ -0,0 +1,149 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +enum ValueType { + VALUE_TYPE_INT32 = 0, + VALUE_TYPE_FLOAT = 1, + VALUE_TYPE_DOUBLE = 2, + VALUE_TYPE_BYTE = 3 +}; + +enum DeviceType { + DEVICE_TYPE_UNSPECIFIED = 0, + DEVICE_TYPE_CPU = 1, + DEVICE_TYPE_GPU = 2 +}; + +enum SparseDataType { T_NO_VALUE = 0, T_FLOAT_VALUE = 1 }; + +enum SparseDataFormat { T_SPARSE_CSR = 0, T_SPARSE_CSC = 1 }; + +inline int sizeOfValuType(ValueType valueType) { + if (valueType == VALUE_TYPE_INT32) { + return 4; + } else if (valueType == VALUE_TYPE_FLOAT) { + return 4; + } else if (valueType == VALUE_TYPE_DOUBLE) { + return 8; + } else { + LOG(FATAL) << "Unknown type: " << valueType; + return 0; + } +} + +template +struct DataType; + +template <> +struct DataType { + static const ValueType value = VALUE_TYPE_FLOAT; +}; + +template <> +struct DataType { + static const ValueType value = VALUE_TYPE_DOUBLE; +}; + +template <> +struct DataType { + static const ValueType value = VALUE_TYPE_INT32; +}; + +namespace detail { + +template +struct MatrixT; + +template <> +struct MatrixT { + using type = CpuMatrix; +}; + +template <> +struct MatrixT { + using type = GpuMatrix; +}; + +template <> +struct MatrixT { + using type = void; // Not implemented +}; + +template <> +struct MatrixT { + using type = void; // Not implemented +}; + +template +struct SparseMatrixT; + +template <> +struct SparseMatrixT { + using type = CpuSparseMatrix; +}; + +template <> +struct SparseMatrixT { + using type = GpuSparseMatrix; +}; + +template <> +struct SparseMatrixT { + using type = void; // Not implemented +}; + +template <> +struct SparseMatrixT { + using type = void; // Not implemented +}; + +template +struct VectorT; + +template <> +struct VectorT { + using type = CpuVector; +}; + +template <> +struct VectorT { + using type = GpuVector; +}; + +template <> +struct VectorT { + using type = CpuIVector; +}; + +template <> +struct VectorT { + using type = GpuIVector; +}; + +} // namespace detail + +template +struct Tensor { + typedef typename detail::VectorT::type Vector; + typedef typename detail::MatrixT::type Matrix; + typedef typename detail::SparseMatrixT::type SparseMatrix; +}; + +} // namespace paddle diff --git a/paddle/function/TensorTypeTest.cpp b/paddle/legacy/function/TensorTypeTest.cpp similarity index 100% rename from paddle/function/TensorTypeTest.cpp rename to paddle/legacy/function/TensorTypeTest.cpp diff --git a/paddle/legacy/function/neon/NeonDepthwiseConv.cpp b/paddle/legacy/function/neon/NeonDepthwiseConv.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6179635a9fec4afecf53fabdc6a818588b54c808 --- /dev/null +++ b/paddle/legacy/function/neon/NeonDepthwiseConv.cpp @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "NeonDepthwiseConv.h" +#include "paddle/legacy/function/ConvOp.h" + +namespace paddle { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +class NeonDepthwiseConvFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + int batchSize = input[0]; + int inputChannels = input[1]; + int inputHeight = input[2]; + int inputWidth = input[3]; + int filterHeight = getFilterHeight(filter); + int filterWidth = getFilterWidth(filter); + int outputChannels = output[1]; + int outputHeight = output[2]; + int outputWidth = output[3]; + int filterMultiplier = outputChannels / groups_; + CHECK_EQ(static_cast(inputChannels), groups_); + + // only support strideH() == strideW() and filterHeight == filterWidth. + CHECK_EQ(strideH(), strideW()); + CHECK_EQ(filterHeight, filterWidth); + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + // padding the input + float* inputPadding = inputData; + int padInputHeight = inputHeight + 2 * paddingH(); + int padInputWidth = inputWidth + 2 * paddingW(); + int newSize = + batchSize * (inputChannels + 1) * padInputHeight * padInputWidth; + + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + + std::function + DepthWiseConv; + + if (filterWidth == 3 && strideW() == 1) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; + } else if (filterWidth == 3 && strideW() == 2) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 2>::run; + } else if (filterWidth == 4 && strideW() == 1) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; + } else if (filterWidth == 4 && strideW() == 2) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 2>::run; + } else { + LOG(FATAL) << "Not supported"; + } + + for (int i = 0; i < batchSize; i++) { + DepthWiseConv(inputPadding, + filterData, + padInputHeight, + padInputWidth, + outputChannels, + outputHeight, + outputWidth, + filterMultiplier, + outputData); + inputPadding += inputChannels * padInputHeight * padInputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifndef PADDLE_TYPE_DOUBLE +REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction); +#endif + +#endif + +} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/legacy/function/neon/NeonDepthwiseConv.h similarity index 100% rename from paddle/function/neon/NeonDepthwiseConv.h rename to paddle/legacy/function/neon/NeonDepthwiseConv.h diff --git a/paddle/legacy/function/neon/NeonDepthwiseConvTranspose.cpp b/paddle/legacy/function/neon/NeonDepthwiseConvTranspose.cpp new file mode 100644 index 0000000000000000000000000000000000000000..feb77e1ff9f591d63dbf86a05313d65025f7c65d --- /dev/null +++ b/paddle/legacy/function/neon/NeonDepthwiseConvTranspose.cpp @@ -0,0 +1,136 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "NeonDepthwiseConv.h" +#include "paddle/legacy/function/ConvOp.h" + +namespace paddle { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +class NeonDepthwiseConvTransposeFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + int batchSize = input[0]; + int inputChannels = input[1]; + int inputHeight = input[2]; + int inputWidth = input[3]; + int filterHeight = getFilterHeight(filter); + int filterWidth = getFilterWidth(filter); + int outputChannels = output[1]; + int outputHeight = output[2]; + int outputWidth = output[3]; + int filterMultiplier = outputChannels / groups_; + CHECK_EQ(inputChannels, groups_); + + // only support strideH() == strideW() and filterHeight == filterWidth. + CHECK_EQ(strideH(), strideW()); + CHECK_EQ(paddingH(), paddingW()); + CHECK_EQ(filterHeight, filterWidth); + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + // padding the input, input -> inputPadding + float* inputPadding = inputData; + int padInputHeight = + (inputHeight - 1) * strideH() + 2 * filterHeight - 1 - 2 * paddingH(); + int padInputWidth = + (inputWidth - 1) * strideW() + 2 * filterWidth - 1 - 2 * paddingW(); + + if (padInputHeight > inputHeight || padInputWidth > inputWidth) { + int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + if (strideH() == 1) { + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } else if (strideH() == 2) { + neon::StridePadding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } else { + LOG(FATAL) << "Not supported"; + } + } + + std::function + DepthWiseConv; + + if (filterWidth == 3) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; + } else if (filterWidth == 4) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; + } else { + LOG(FATAL) << "Not supported"; + } + + for (int i = 0; i < batchSize; i++) { + DepthWiseConv(inputPadding, + filterData, + padInputHeight, + padInputWidth, + outputChannels, + outputHeight, + outputWidth, + filterMultiplier, + outputData); + inputPadding += inputChannels * padInputHeight * padInputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifndef PADDLE_TYPE_DOUBLE + +REGISTER_TYPED_FUNC(NeonDepthwiseConvTranspose, + CPU, + NeonDepthwiseConvTransposeFunction); + +#endif + +#endif + +} // namespace paddle diff --git a/paddle/function/neon/neon_util.h b/paddle/legacy/function/neon/neon_util.h similarity index 100% rename from paddle/function/neon/neon_util.h rename to paddle/legacy/function/neon/neon_util.h diff --git a/paddle/legacy/function/nnpack/NNPACKConvOp.cpp b/paddle/legacy/function/nnpack/NNPACKConvOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..81c832e7747f8e75d322891476e08dacc435f5d4 --- /dev/null +++ b/paddle/legacy/function/nnpack/NNPACKConvOp.cpp @@ -0,0 +1,247 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "nnpack.h" +#include "paddle/legacy/function/ConvOp.h" + +DEFINE_bool(nnpack_allocate_outside, + true, + "Allocate and free workspace memory outside the NNPACK interface."); +DEFINE_int32(nnpack_num_threads, + 0, + "The number of nnpack threads" + "default: 0; 0 to disable threadpool."); + +namespace paddle { + +nnp_convolution_algorithm get_nnp_convolution_algorithm( + const std::string& algorithm) { + if (algorithm == "auto") { + return nnp_convolution_algorithm_auto; + } else if (algorithm == "ft8x8") { + return nnp_convolution_algorithm_ft8x8; + } else if (algorithm == "ft16x16") { + return nnp_convolution_algorithm_ft16x16; + } else if (algorithm == "wt8x8") { + return nnp_convolution_algorithm_wt8x8; + } else if (algorithm == "implicit-gemm") { + return nnp_convolution_algorithm_implicit_gemm; + } else if (algorithm == "direct") { + return nnp_convolution_algorithm_direct; + } else { + return nnp_convolution_algorithm_auto; + } +} + +template +class NNPACKConvFunction : public ConvFunctionBase { + public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + algorithm_ = get_nnp_convolution_algorithm(config.get("algo")); + transform_strategy_ = nnp_convolution_transform_strategy_compute; + nnp_status status = nnp_initialize(); + CHECK_EQ(status, nnp_status_success); + workspaceBuffer_ = nullptr; + workspaceSize_ = 0; + + create_nnpack_threadpool(); + } + + ~NNPACKConvFunction() { + if (workspaceBuffer_) { + free(workspaceBuffer_); + } + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + check(inputs, outputs); + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + + nnp_size inputSize = {.width = inputWidth, .height = inputHeight}; + nnp_padding padding = {.top = (size_t)paddingH(), + .right = (size_t)paddingW(), + .bottom = (size_t)paddingH(), + .left = (size_t)paddingW()}; + nnp_size kernelSize = {.width = filterWidth, .height = filterHeight}; + nnp_size outputSubsampling = {.width = (size_t)strideW(), + .height = (size_t)strideH()}; + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + void* bufferPtr = nullptr; + size_t* sizePtr = nullptr; + size_t needSize; + if (FLAGS_nnpack_allocate_outside) { + if (batchSize == 1) { + nnp_status status = nnp_convolution_inference(algorithm_, + transform_strategy_, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + outputSubsampling, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, + &needSize, + nnp_activation_identity, + nullptr, + nullptr, + nullptr); + CHECK_EQ(status, nnp_status_success); + } else { + // only supports stride = 1 + CHECK_EQ(strideH(), 1); + CHECK_EQ(strideW(), 1); + nnp_status status = nnp_convolution_output(algorithm_, + batchSize, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, + &needSize, + nnp_activation_identity, + nullptr, + nullptr, + nullptr); + CHECK_EQ(status, nnp_status_success); + } + + VLOG(3) << "workspace size is " << needSize; + if (needSize > workspaceSize_) { + workspaceSize_ = needSize; + if (workspaceBuffer_) { + free(workspaceBuffer_); + } else { + posix_memalign(&workspaceBuffer_, 64, needSize); + } + } + + if (needSize) { + bufferPtr = workspaceBuffer_; + sizePtr = &needSize; + } + } + + size_t inputOffset = inputChannels / groups_ * inputHeight * inputWidth; + size_t outputOffset = outputChannels / groups_ * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + + if (batchSize == 1) { + for (size_t g = 0; g < groups_; g++) { + nnp_status status = + nnp_convolution_inference(algorithm_, + transform_strategy_, + inputChannels / groups_, + outputChannels / groups_, + inputSize, + padding, + kernelSize, + outputSubsampling, + inputData + inputOffset * g, + filterData + filterOffset * g, + nullptr, /* bias */ + outputData + outputOffset * g, + bufferPtr, + sizePtr, + nnp_activation_identity, + nullptr, + threadpool_, /* threadpool */ + nullptr); + CHECK_EQ(status, nnp_status_success); + } + } else { + // only supports stride = 1 + CHECK_EQ(strideH(), 1); + CHECK_EQ(strideW(), 1); + + // TODO(hedaoyuan): There has some bug when batchSize > 1 and groups_ > 1. + CHECK_EQ(groups_, static_cast(1)); + nnp_status status = nnp_convolution_output(algorithm_, + batchSize, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + inputData, + filterData, + nullptr, /* bias */ + outputData, + bufferPtr, + sizePtr, + nnp_activation_identity, + nullptr, + threadpool_, /* threadpool */ + nullptr); + CHECK_EQ(status, nnp_status_success); + } + } + + static void create_nnpack_threadpool() { + if (FLAGS_nnpack_num_threads && threadpool_ == nullptr) { + threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads); + VLOG(3) << "Number of threads " + << pthreadpool_get_threads_count(threadpool_); + } + } + + private: + nnp_convolution_algorithm algorithm_; + nnp_convolution_transform_strategy transform_strategy_; + void* workspaceBuffer_; + size_t workspaceSize_; + static pthreadpool_t threadpool_; +}; + +template +pthreadpool_t NNPACKConvFunction::threadpool_ = nullptr; + +REGISTER_TYPED_FUNC(NNPACKConv, CPU, NNPACKConvFunction); + +} // namespace paddle diff --git a/paddle/legacy/function/nnpack/NNPACKConvOpTest.cpp b/paddle/legacy/function/nnpack/NNPACKConvOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a2db83f5a36310ca6f173d6e6501118b34060761 --- /dev/null +++ b/paddle/legacy/function/nnpack/NNPACKConvOpTest.cpp @@ -0,0 +1,30 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/function/ConvOpTest.h" + +namespace paddle { + +TEST(NNPACK, Forward) { + Convolution( + "GemmConv-CPU", "NNPACKConv-CPU", forward); +} + +TEST(NNPACK, Depthwise) { + DepthwiseConvolution( + "GemmConv-CPU", "NNPACKConv-CPU", forward); +} + +} // namespace paddle diff --git a/paddle/gserver/CMakeLists.txt b/paddle/legacy/gserver/CMakeLists.txt similarity index 100% rename from paddle/gserver/CMakeLists.txt rename to paddle/legacy/gserver/CMakeLists.txt diff --git a/paddle/legacy/gserver/activations/ActivationFunction.cpp b/paddle/legacy/gserver/activations/ActivationFunction.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ae07c7e6d7fd9fe28a00dd209ae834cd28a327f7 --- /dev/null +++ b/paddle/legacy/gserver/activations/ActivationFunction.cpp @@ -0,0 +1,509 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ActivationFunction.h" + +#include +#include +#include +#include +#include +#include +#include "paddle/legacy/parameter/Argument.h" +#include "paddle/legacy/utils/ClassRegistrar.h" +#include "paddle/legacy/utils/Logging.h" + +#ifdef PADDLE_WITH_MKLDNN +#include "MKLDNNActivation.h" +#endif + +namespace paddle { + +static ClassRegistrar gActivationRegistrar; +/** + * @def ACTIVATION_CLASS_NAME + * @brief Macro for getting derived activation class name + * @note ACTIVATION_CLASS_NAME(softmax) softmax_; + * means softmaxActivation softmax_; + */ +#define ACTIVATION_CLASS_NAME(ACTIVATION_NAME) ACTIVATION_NAME##Activation +/** + * @def BEGIN_DEFINE_ACTIVATION + * @brief Macro for defining a devried activation class + */ +#define BEGIN_DEFINE_ACTIVATION(ACTIVATION_NAME) \ + class ACTIVATION_CLASS_NAME(ACTIVATION_NAME) : public ActivationFunction { \ + private: \ + static const std::string name; \ + \ + public: \ + const std::string& getName() const { return name; } +/** + * @def END_DEFINE_ACTIVATION + * @brief Macro for registering a derived activation class + */ +#define END_DEFINE_ACTIVATION(ACTIVATION_NAME) \ + } \ + ; \ + const std::string ACTIVATION_CLASS_NAME(ACTIVATION_NAME)::name = \ + #ACTIVATION_NAME; \ + static InitFunction __reg_activation__##ACTIVATION_NAME([] { \ + gActivationRegistrar \ + .registerClass( \ + #ACTIVATION_NAME); \ + }); + +/** + * @brief The IdentityActivation class + * + * Do nothing when forward/backward. + */ +class IdentityActivation : public ActivationFunction { + public: + static const std::string name; + Error __must_check forward(Argument& act) { + (void)act; + return Error(); + } + Error __must_check backward(Argument& act) { + (void)act; + return Error(); + } + const std::string& getName() const { return name; } +}; +const std::string IdentityActivation::name = ""; +static InitFunction __reg_activation__identity([] { + gActivationRegistrar.registerClass(""); + gActivationRegistrar.registerClass("linear"); +}); + +/** + * @brief Sigmoid Activation + * \f[ + * f(z) = \frac{1}{1+exp(-z)} + * \f] + */ +BEGIN_DEFINE_ACTIVATION(sigmoid) +Error __must_check forward(Argument& act) { + act.value->sigmoid(*act.value); + return Error(); +} +Error __must_check backward(Argument& act) { + act.grad->sigmoidDerivative(*act.value); + return Error(); +} +END_DEFINE_ACTIVATION(sigmoid) + +/** + * @brief Softmax Activation + * \f[ + * P(y=j|x) = \frac{e^{x^Tw_j}}{\sum^K_{k=1}e^{x^Tw_k}} + * \f] + */ +BEGIN_DEFINE_ACTIVATION(softmax) +private: +MatrixPtr sftMaxSum_; +MatrixPtr sftMaxDot_; + +public: +Error __must_check forward(Argument& act) { + act.value->softmax(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + MatrixPtr outputV = act.value; + MatrixPtr outputG = act.grad; + + if (outputG->useGpu()) { + outputG->softmaxBackward(*outputV); + } else { + SetDevice device(act.deviceId); + Matrix::resizeOrCreate(sftMaxDot_, + outputG->getHeight(), + outputG->getWidth(), + /* trans */ false, + useGpu(act.deviceId)); + Matrix::resizeOrCreate(sftMaxSum_, + outputG->getHeight(), + 1, + /* trans */ false, + useGpu(act.deviceId)); + + sftMaxDot_->dotMul(*outputG, *outputV); + sftMaxSum_->colMerge(*sftMaxDot_); + + act.grad->softmaxDerivative(*act.value, *sftMaxSum_); + } + return Error(); +} +END_DEFINE_ACTIVATION(softmax) + +/** + * @brief Sequence_softmax Activation + * @note Softmax on all frames of one sequence. + * Width of frame must be one. + */ +BEGIN_DEFINE_ACTIVATION(sequence_softmax) +private: +ACTIVATION_CLASS_NAME(softmax) softmax_; +Argument argument_; + +public: +Error __must_check forward(Argument& act) { + if (act.value->getWidth() != 1UL) { + return Error( + "Input width for each timestep of sequence softmax should be 1"); + } + + if (!argument_.value) { + argument_.value = Matrix::create(nullptr, + /* height= */ 1, + 1, + /* trans= */ false, + useGpu(act.deviceId)); + argument_.grad = Matrix::create(nullptr, + /* height= */ 1, + 1, + /* trans= */ false, + useGpu(act.deviceId)); + } + + auto starts = + act.hasSubseq() + ? act.subSequenceStartPositions->getVector(useGpu(act.deviceId)) + : act.sequenceStartPositions->getVector(useGpu(act.deviceId)); + act.value->sequenceSoftmax(*act.value, *starts); + return Error(); +} + +Error __must_check backward(Argument& act) { + if (act.value->getWidth() != 1UL) { + return Error( + "Input width for each timestep of sequence softmax should be 1"); + } + + size_t numSequences = + act.hasSubseq() ? act.getNumSubSequences() : act.getNumSequences(); + const int* starts = act.getCpuStartPositions(); + + for (size_t i = 0; i < numSequences; ++i) { + // TODO(Dangqingqing) optimization for GPU + size_t offset = starts[i]; + size_t size = starts[i + 1] - starts[i]; + argument_.value->setData(act.value->getData() + offset, 1UL, size); + argument_.grad->setData(act.grad->getData() + offset, 1UL, size); + + Error err = softmax_.backward(argument_); + if (!err.isOK()) return err; + } + return Error(); +} +END_DEFINE_ACTIVATION(sequence_softmax) + +/* + * @brief SoftSign Activation. + * \f[ + * f(z) = \frac{z}{1 + |z|} + * \f] + */ +BEGIN_DEFINE_ACTIVATION(softsign) +private: +MatrixPtr denominator_; + +Error __must_check forward(Argument& act) { + size_t height = act.value->getHeight(); + size_t width = act.value->getWidth(); + Matrix::resizeOrCreate( + denominator_, height, width, false, useGpu(act.deviceId)); + denominator_->assign(*act.value); + denominator_->abs2(); + denominator_->add(1.); + + act.value->dotDiv(*act.value, *denominator_); + return Error(); +} + +Error __must_check backward(Argument& act) { + denominator_->square2(); + denominator_->scalarDiv(*denominator_, 1.); + act.grad->dotMul(*act.grad, *denominator_); + return Error(); +} +END_DEFINE_ACTIVATION(softsign) + +/** + * @brief Relu Activation. + * forward. y = max(0, z) + * + * derivative of relu is: + * + * 1 if z > 0 + * + * 0 otherwise. + */ +BEGIN_DEFINE_ACTIVATION(relu) +Error __must_check forward(Argument& act) { + act.value->relu(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->reluDerivative(*act.value); + return Error(); +} +END_DEFINE_ACTIVATION(relu) + +/** + * @brief BRelu Activation. + * + * forward. y = min(24, max(0, z)) + * + * derivative of brelu is: + * + * 1 if 0 < z < 24 + * + * 0 otherwise. + * + * TODO(yuyang18): Remove magic number 24 or make it configuable. + */ +BEGIN_DEFINE_ACTIVATION(brelu) +Error __must_check forward(Argument& act) { + act.value->brelu(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->breluDerivative(*act.value); + return Error(); +} +END_DEFINE_ACTIVATION(brelu) + +/** + * @brief Tanh Activation. + * \f[ + * f(z) = tanh(z)=\frac{e^z-e^{-z}}{e^z+e^{-z}} + * \f] + */ +BEGIN_DEFINE_ACTIVATION(tanh) +Error __must_check forward(Argument& act) { + act.value->tanh(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->tanhDerivative(*act.value); + return Error(); +} +END_DEFINE_ACTIVATION(tanh) + +/** + * @brief Scaled Tanh Activation + * \f[ + * f(z) = 1.7159 * tanh(2/3*z) + * \f] + */ +BEGIN_DEFINE_ACTIVATION(stanh) +private: +real a, b; + +public: +ACTIVATION_CLASS_NAME(stanh)() : a(1.7159), b(2. / 3.) {} +Error __must_check forward(Argument& act) { + act.value->scaledTanh(*act.value, a, b); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->scaledTanhDerivative(*act.value, a, b); + return Error(); +} +END_DEFINE_ACTIVATION(stanh) + +/** + * @brief Soft Relu Activation. + * \f[ + * f(z) = ln(1+e^z) + * \f] + */ +BEGIN_DEFINE_ACTIVATION(softrelu) +Error __must_check forward(Argument& act) { + act.value->softrelu(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->softreluDerivative(*act.value); + return Error(); +} +END_DEFINE_ACTIVATION(softrelu) + +/** + * @brief Abs Activation. + * Forward: f(z) = abs(z) + * + * Derivative: + * + * 1 if z>0 + * + * -1 if z<0 + * + * 0 if z=0 + */ +BEGIN_DEFINE_ACTIVATION(abs) +Error __must_check forward(Argument& act) { + SetDevice device(act.deviceId); + Matrix::resizeOrCreate(act.in, + act.value->getHeight(), + act.value->getWidth(), + /* trans */ false, + useGpu(act.deviceId)); + + act.in->copyFrom(*act.value); + act.value->abs2(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->absDerivative(*act.in); + return Error(); +} +END_DEFINE_ACTIVATION(abs) + +/** + * @brief Square Activation. + * \f[ + * f(z) = z^2. + * \f] + */ +BEGIN_DEFINE_ACTIVATION(square) +Error __must_check forward(Argument& act) { + SetDevice device(act.deviceId); + Matrix::resizeOrCreate(act.in, + act.value->getHeight(), + act.value->getWidth(), + /* trans */ false, + useGpu(act.deviceId)); + + act.in->copyFrom(*act.value); + act.value->square2(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->squareDerivative(*act.in); + return Error(); +} +END_DEFINE_ACTIVATION(square) + +/** + * @brief Exponential Activation. + * \f[ + * f(z) = e^z + * \f] + */ +BEGIN_DEFINE_ACTIVATION(exponential) +Error __must_check forward(Argument& act) { + act.value->exp2(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->expDerivative(*act.value); + return Error(); +} +END_DEFINE_ACTIVATION(exponential) + +/** + * @brief Reciprocal Activation. + * \f[ + * f(z) = 1/z + * \f] + */ +BEGIN_DEFINE_ACTIVATION(reciprocal) +Error __must_check forward(Argument& act) { + act.value->reciprocal2(); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->dotMulSquare(*act.value); + act.grad->neg(); + return Error(); +} +END_DEFINE_ACTIVATION(reciprocal) + +/** + * @brief Square Root Activation. + * \f[ + * f(z) = sqrt(z) + * \f] + */ +BEGIN_DEFINE_ACTIVATION(sqrt) +Error __must_check forward(Argument& act) { + act.value->sqrt2(); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->dotDiv(*act.grad, *act.value); + act.grad->mulScalar(0.5); + return Error(); +} +END_DEFINE_ACTIVATION(sqrt) + +/** + * @brief Logarithm Activation. + * \f[ + * f(z) = log(z) + * \f] + */ +BEGIN_DEFINE_ACTIVATION(log) +Error __must_check forward(Argument& act) { + SetDevice device(act.deviceId); + Matrix::resizeOrCreate(act.in, + act.value->getHeight(), + act.value->getWidth(), + /* trans */ false, + useGpu(act.deviceId)); + + act.in->copyFrom(*act.value); + act.value->log2(*act.value); + return Error(); +} + +Error __must_check backward(Argument& act) { + act.grad->dotDiv(*act.grad, *act.in); + return Error(); +} +END_DEFINE_ACTIVATION(log) + +ActivationFunction* ActivationFunction::create(const std::string& type) { +#ifdef PADDLE_WITH_MKLDNN + if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) { + return MKLDNNActivation::create(type); + } +#endif + + return gActivationRegistrar.createByType(type); +} + +std::vector ActivationFunction::getAllRegisteredTypes() { + std::vector types; + gActivationRegistrar.forEachType( + [&](const std::string& type) { types.push_back(type); }); + return types; +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/activations/ActivationFunction.h b/paddle/legacy/gserver/activations/ActivationFunction.h new file mode 100644 index 0000000000000000000000000000000000000000..8bc5b0f529a6358fba8b6c9d1e1f6ee2358dbbf9 --- /dev/null +++ b/paddle/legacy/gserver/activations/ActivationFunction.h @@ -0,0 +1,66 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/legacy/utils/Error.h" + +namespace paddle { + +struct Argument; +/** + * @brief Activation function is a function that transforms a set of input + * signals into an output signals. The purpose of the activation function + * is to introduce non-liearilty into the network. + * + * @note Common activation function are provieded, including linear, + * sigmoid, softmax, sequence_max, relu, brelu, tanh, stanh, + * softrelu, abs, square, exponential. + * + */ +class ActivationFunction { + public: + static ActivationFunction* create(const std::string& type); + static std::vector getAllRegisteredTypes(); + + ActivationFunction() {} + + virtual ~ActivationFunction() {} + + /** + * @brief Foward propagation + * + * act.value <- f(act.value), + * where f is the activation function. + * Suppose that before calling forward(), act.value is x and + * after forward() is called, act.value is y, then y = f(x). + * + * Usually, act is Layer::output_ + */ + virtual Error __must_check forward(Argument& act) = 0; + + /** + * @brief Backward propagaion + * + * x and y are defined in the above comment for forward(). + * - Before calling backward(), act.grad = dE / dy, where E is the error/cost + * - After backward() returns, act.grad = dE / dx = (dE/dy) * (dy/dx) + */ + virtual Error __must_check backward(Argument& act) = 0; + + virtual const std::string& getName() const = 0; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/activations/MKLDNNActivation.cpp b/paddle/legacy/gserver/activations/MKLDNNActivation.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2eed7af70a8a3cc305a79bbe23177ea71d15d252 --- /dev/null +++ b/paddle/legacy/gserver/activations/MKLDNNActivation.cpp @@ -0,0 +1,249 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNActivation.h" +#include "mkldnn.hpp" +#include "paddle/legacy/utils/ClassRegistrar.h" + +namespace paddle { + +static ClassRegistrar gMKLDNNActivationRegistrar; +/** + * @def MKLDNN_ACTIVATION_CLASS_NAME + * @note MKLDNN_ACTIVATION_CLASS_NAME(relu) relu_; + * means mkldnn_reluActivation relu_; + */ +#define MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) mkldnn_##ACT_TYPE##Activation + +/** + * @def BEGIN_MKLDNN_ACTIVATION + */ +#define BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + class MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) : public BASE_CLASS { +/** + * @def END_MKLDNN_ACTIVATION + */ +#define END_MKLDNN_ACTIVATION(ACT_TYPE) \ + private: \ + static const std::string name; \ + \ + public: \ + const std::string& getName() const { return name; } \ + } \ + ; \ + const std::string MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::name = \ + "mkldnn_" #ACT_TYPE; \ + static InitFunction __reg_activation__mkldnn_##ACT_TYPE([] { \ + gMKLDNNActivationRegistrar \ + .registerClass( \ + "mkldnn_" #ACT_TYPE); \ + }); + +/** + * @def DEFINE_MKLDNN_ACTIVATION + */ +#define DEFINE_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + END_MKLDNN_ACTIVATION(ACT_TYPE) + +/** + * @def DEFINE_MKLDNN_ELTWISE_ACTIVATION + */ +#define DEFINE_MKLDNN_ELTWISE_ACTIVATION( \ + ACT_TYPE, BASE_CLASS, ALPHA, BWD_ALPHA) \ + BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + private: \ + static const float alpha; \ + static const float bwdAlpha; \ + \ + public: \ + float getAlpha() const { return alpha; } \ + float getBwdAlpha() const { return bwdAlpha; } \ + END_MKLDNN_ACTIVATION(ACT_TYPE) \ + const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::alpha = ALPHA; \ + const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::bwdAlpha = BWD_ALPHA; + +/** + * @brief MKLDNN Relu Activation. + * Actually mkldnn_relu is Leaky Relu. + * f(x) = x (x >= 0) + * f(x) = negative_slope * x (x < 0) + * @note the negative_slope should be -0.f in forward + */ +DEFINE_MKLDNN_ELTWISE_ACTIVATION(relu, MKLDNNEltwiseActivation, -0.f, 0.f) + +/** + * @brief MKLDNN Tanh Activation. + */ +DEFINE_MKLDNN_ELTWISE_ACTIVATION(tanh, MKLDNNEltwiseActivation, 0.f, 0.f) + +/** + * @brief MKLDNN ELU(Exponential Linear Unit) Activation. + * f(x) = x (x >= 0) + * f(x) = negative_slope * (exp(x) - 1) (x < 0) + */ +DEFINE_MKLDNN_ELTWISE_ACTIVATION(elu, MKLDNNEltwiseActivation, 0.f, 0.f) + +mkldnn::algorithm MKLDNNEltwiseActivation::getAlgo(std::string type) const { + const std::map algoMap = { + {"relu", algorithm::eltwise_relu}, + {"tanh", algorithm::eltwise_tanh}, + {"elu", algorithm::eltwise_elu}}; + type.erase(0, 7); // remove mkldnn_ + algorithm algo = (algorithm)0; + mapGet(type, algoMap, &algo); + return algo; +} + +void MKLDNNEltwiseActivation::resetFwd(Argument& act) { + if (cnt_ == act.value->getElementCnt()) { + return; + } + MKLDNNActivation::resetFwd(act); + // note: alpha represents the NegativeSlope when used in relu. + float alpha = getAlpha(); + float beta = getBeta(); + algorithm algo = getAlgo(this->getName()); + auto fwdDesc = eltwise_fwd::desc(mkldnn::prop_kind::forward_training, + algo, + val_->getMemoryDesc(), + alpha, + beta); + fwdPD_.reset(new eltwise_fwd::primitive_desc(fwdDesc, *engine_)); + // use inplace for forward but save input value before submit + inVal_ = val_; + copyInVal_ = nullptr; + if (act.grad && algo == algorithm::eltwise_tanh) { + // tanh need save src input for backward + inVal_ = MKLDNNMatrix::create(val_->getPrimitiveDesc()); + copyInVal_ = std::make_shared(*val_, *inVal_); + CHECK(copyInVal_) << "should not be emptry"; + pipelineFwd_.push_back(*copyInVal_); + } + fwd_.reset(new eltwise_fwd(*fwdPD_, *val_, *val_)); + pipelineFwd_.push_back(*fwd_); + needResetBwd_ = true; +} + +void MKLDNNEltwiseActivation::resetBwd(Argument& act) { + if (!needResetBwd_) { + return; + } + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; + needResetBwd_ = false; + algorithm algo = getAlgo(this->getName()); + float alpha = getBwdAlpha(); + float beta = getBeta(); + grad_ = MKLDNNMatrix::create(val_->getPrimitiveDesc(), act.grad); + auto eng = CPUEngine::Instance().getEngine(); + auto bwdDesc = eltwise_bwd::desc( + algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta); + auto bwdPD = eltwise_bwd::primitive_desc(bwdDesc, eng, *fwdPD_); + CHECK(inVal_); + bwd_.reset(new eltwise_bwd(bwdPD, *inVal_, *grad_, *grad_)); + pipelineBwd_.clear(); + pipelineBwd_.push_back(*bwd_); +} + +/** + * @brief MKLDNN Softmax Activation + */ +DEFINE_MKLDNN_ACTIVATION(softmax, MKLDNNSoftmaxActivation) + +void MKLDNNSoftmaxActivation::resetFwd(Argument& act) { + if (cnt_ == act.value->getElementCnt()) { + return; + } + MKLDNNActivation::resetFwd(act); + int axis = 1; + auto fwdDesc = softmax_fwd::desc( + mkldnn::prop_kind::forward_scoring, val_->getMemoryDesc(), axis); + auto fwdPD = softmax_fwd::primitive_desc(fwdDesc, *engine_); + fwd_.reset(new softmax_fwd(fwdPD, *val_, *val_)); + pipelineFwd_.push_back(*fwd_); +} + +Error __must_check MKLDNNSoftmaxActivation::forward(Argument& act) { + resetFwd(act); + stream_->submit(pipelineFwd_); + real* v = act.value->getData(); + real threshold = exp(-64); +#pragma omp parallel for + for (size_t i = 0; i < act.value->getElementCnt(); ++i) { + v[i] = v[i] < threshold ? threshold : v[i]; + } + return Error(); +} + +Error __must_check MKLDNNSoftmaxActivation::backward(Argument& act) { + MatrixPtr outputV = act.value; + MatrixPtr outputG = act.grad; + Matrix::resizeOrCreate(sftMaxDot_, + outputG->getHeight(), + outputG->getWidth(), + /* trans */ false, + /* useGpu */ false); + Matrix::resizeOrCreate(sftMaxSum_, + outputG->getHeight(), + 1, + /* trans */ false, + /* useGpu */ false); + sftMaxDot_->dotMul(*outputG, *outputV); + sftMaxSum_->colMerge(*sftMaxDot_); + act.grad->softmaxDerivative(*act.value, *sftMaxSum_); + return Error(); +} + +ActivationFunction* MKLDNNActivation::create(const std::string& type) { + return gMKLDNNActivationRegistrar.createByType(type); +} + +std::vector MKLDNNActivation::getAllRegisteredTypes() { + std::vector types; + gMKLDNNActivationRegistrar.forEachType( + [&](const std::string& type) { types.push_back(type); }); + return types; +} + +void MKLDNNActivation::resetFwd(Argument& act) { + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; + cnt_ = act.value->getElementCnt(); + pipelineFwd_.clear(); + stream_.reset(new MKLDNNStream()); + engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0)); + val_ = std::dynamic_pointer_cast(act.value); + if (val_ == nullptr) { + int bs = act.getBatchSize(); + int ih = act.getFrameHeight() > 0 ? act.getFrameHeight() : 1; + int iw = act.getFrameWidth() > 0 ? act.getFrameWidth() : 1; + int ic = cnt_ / bs / ih / iw; + CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw); + val_ = MKLDNNMatrix::create( + {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_, act.value); + CHECK(val_); + val_->downSpatial(); + } +} + +Error __must_check MKLDNNActivation::forward(Argument& act) { + resetFwd(act); + stream_->submit(pipelineFwd_); + return Error(); +} +Error __must_check MKLDNNActivation::backward(Argument& act) { + resetBwd(act); + stream_->submit(pipelineBwd_); + return Error(); +} +} // namespace paddle diff --git a/paddle/legacy/gserver/activations/MKLDNNActivation.h b/paddle/legacy/gserver/activations/MKLDNNActivation.h new file mode 100644 index 0000000000000000000000000000000000000000..59c447ad07398c0b6ca7d78766dd533963744d1b --- /dev/null +++ b/paddle/legacy/gserver/activations/MKLDNNActivation.h @@ -0,0 +1,119 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "ActivationFunction.h" +#include "mkldnn.hpp" +#include "paddle/legacy/gserver/layers/MKLDNNBase.h" +#include "paddle/legacy/math/MKLDNNMatrix.h" +#include "paddle/legacy/parameter/Argument.h" + +namespace paddle { + +/** + * @brief Base class of MKLDNN Activation. + * Common activation function are provieded, + * including mkldnn_relu, mkldnn_elu, mkldnn_tanh, mkldnn_softmax + */ +class MKLDNNActivation : public ActivationFunction { + protected: + // input value element count + size_t cnt_; + // should not merge the resetBwd into resetFwd, + // because the grad data would be changing before backward. + bool needResetBwd_; + // mkldnn matrix, primitive, stream and pipeline + MKLDNNMatrixPtr val_; + MKLDNNMatrixPtr grad_; + std::shared_ptr engine_; + std::shared_ptr stream_; + std::shared_ptr fwd_; + std::shared_ptr bwd_; + std::vector pipelineFwd_; + std::vector pipelineBwd_; + + public: + MKLDNNActivation() : cnt_(0), needResetBwd_(true) {} + ~MKLDNNActivation() {} + static ActivationFunction* create(const std::string& type); + static std::vector getAllRegisteredTypes(); + virtual const std::string& getName() const = 0; + /** + * reset the forward primitives + */ + virtual void resetFwd(Argument& act); + /** + * reset the backward primitives, + * can not merge this functions into resetFwd as the grad data + * would be changing before backward. + */ + virtual void resetBwd(Argument& act) {} + virtual Error __must_check forward(Argument& act); + virtual Error __must_check backward(Argument& act); +}; + +/** + * @brief Base class of MKLDNN Eltwise Activation, + * includes mkldnn_relu, mkldnn_elu and mkldnn_tanh. + */ +class MKLDNNEltwiseActivation : public MKLDNNActivation { + typedef mkldnn::eltwise_forward eltwise_fwd; + typedef mkldnn::eltwise_backward eltwise_bwd; + typedef mkldnn::algorithm algorithm; + + protected: + // save the forward primitive desc, which can be used backward + std::shared_ptr fwdPD_; + // eltwise_bwd need src input value + MKLDNNMatrixPtr inVal_; + // use for copy data + std::shared_ptr copyInVal_; + + public: + MKLDNNEltwiseActivation() {} + ~MKLDNNEltwiseActivation() {} + virtual const std::string& getName() const = 0; + + // in common, the alpha of forward and backward should be equal. + // but for relu, to avoid negative value, they should be opposite + virtual float getAlpha() const = 0; + virtual float getBwdAlpha() const = 0; + virtual float getBeta() const { return 0.f; } + virtual algorithm getAlgo(std::string type) const; + void resetFwd(Argument& act) override; + void resetBwd(Argument& act) override; +}; + +/** + * @brief Base class of MKLDNN softmax Activation, + * only have mkldnn forward, use cpu implement for backward. + */ +class MKLDNNSoftmaxActivation : public MKLDNNActivation { + typedef mkldnn::softmax_forward softmax_fwd; + + private: + // for backward + MatrixPtr sftMaxSum_; + MatrixPtr sftMaxDot_; + + public: + MKLDNNSoftmaxActivation() {} + ~MKLDNNSoftmaxActivation() {} + virtual const std::string& getName() const = 0; + void resetFwd(Argument& act) override; + Error __must_check forward(Argument& act) override; + Error __must_check backward(Argument& act) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/dataproviders/DataProvider.cpp b/paddle/legacy/gserver/dataproviders/DataProvider.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b67af8a326bdfd211ee5720bf67828040b19e5c1 --- /dev/null +++ b/paddle/legacy/gserver/dataproviders/DataProvider.cpp @@ -0,0 +1,410 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "DataProvider.h" + +#include +#include +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/StringUtil.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +void BufferBatch::swap(BufferBatch* bufBatch) { + DataBatch* batchData = bufBatch->getDataBatch(); + hl_event_t hlEvent = bufBatch->getCuEvent(); + hl_stream_t hlStream = bufBatch->getCuStream(); + bufBatch->setDataBatch(batchData_); + bufBatch->setCuStream(hlStream_); + bufBatch->setCuEvent(hlEvent_); + + batchData_ = batchData; + hlEvent_ = hlEvent; + hlStream_ = hlStream; +} + +void BufferBatch::clone(DataBatch* srcBatch, bool useGpu) { + if (batchData_ == NULL) { + batchData_ = new DataBatch(); + } + std::vector& destData = batchData_->getStreams(); + int numStreams = srcBatch->getNumStreams(); + destData.resize(numStreams); + batchData_->setSize(srcBatch->getSize()); + if (useGpu) { + createCuEvent(); + } + + for (int i = 0; i < numStreams; i++) { + destData[i].resizeAndCopyFrom(srcBatch->getStream(i), useGpu, hlStream_); + } + if (useGpu) { + hl_stream_record_event(hlStream_, hlEvent_); + } +} + +DoubleBuffer::DoubleBuffer(DataProvider* dataPool, + bool useGpu, + int64_t batchSize) { + batchSize_ = batchSize; + dataPool_ = dataPool; + useGpu_ = useGpu; + dataQueue_ = new BufferBatchQueue(); + bufferQueue_ = new BufferBatchQueue(); + + // insert a empty buffer + bufferQueue_->enqueue(new BufferBatch()); + stopping_ = false; + pending_ = true; +} + +DoubleBuffer::~DoubleBuffer() { + finishAsyncLoad(); + while (dataQueue_->size()) { + BufferBatch* dataBtch = dataQueue_->dequeue(); + delete dataBtch; + dataBtch = NULL; + } + while (bufferQueue_->size()) { + BufferBatch* bufBtch = bufferQueue_->dequeue(); + delete bufBtch; + bufBtch = NULL; + } + delete dataQueue_; + dataQueue_ = NULL; + delete bufferQueue_; + bufferQueue_ = NULL; +} + +void DoubleBuffer::removeOneBatch(DataBatch* dataBatch) { + // get data + BufferBatch* batch = dataQueue_->dequeue(); + batch->syncEvent(); // when use GPU, need synchronized with the cuEvent + *dataBatch = *(batch->getDataBatch()); + + // push anothor buffer + if (*usingBatch_ == nullptr) { + *usingBatch_ = std::make_shared(); + } + + // Mark the using-batch + batch->swap((*usingBatch_).get()); + bufferQueue_->enqueue(batch); + + if (0 == dataBatch->getSize()) { + setPending(true); + } +} + +void DoubleBuffer::insertOneBatch(DataBatch* batch) { + while (!bufferQueue_->waitNotEmptyFor(2 /* seconds */)) { // time out + if (stopping_) return; + } + BufferBatch* bufBatch = bufferQueue_->dequeue(); + // clone and copy the data from an Threadlocal Variable + bufBatch->clone(batch, useGpu_); + dataQueue_->enqueue(bufBatch); +} + +void DoubleBuffer::asyncLoadBatch() { + int64_t actualSize = 0; + if (useGpu_) { + hl_set_device(FLAGS_gpu_id); + } + setPending(false); + + while (true) { + taskReadySem_.wait(); + if (stopping_) break; + + while (batchSize_ == 0 && !stopping_) { + usleep(5); + } + if (stopping_) break; + + do { + DataBatch newBatch; + { + REGISTER_TIMER("getNextBatchInternal"); + actualSize = dataPool_->getNextBatchInternal(batchSize_, &newBatch); + } + insertOneBatch(&newBatch); + } while (actualSize > 0 && !stopping_); + } +} + +void DoubleBuffer::startAsyncLoad() { + if (asyncLoader_ == nullptr) { + asyncLoader_.reset(new std::thread([this]() { this->asyncLoadBatch(); })); + } + taskReadySem_.post(); +} + +ClassRegistrar + DataProvider::registrar_; + +DataProvider* DataProvider::create(const DataConfig& config, + const ModelConfig& modelConfig, + bool useGpu) { + return registrar_.createByType(config.type(), config, modelConfig, useGpu); +} + +REGISTER_DATA_PROVIDER(simple, SimpleDataProvider); +REGISTER_DATA_PROVIDER(dummy, DummyDataProvider); + +int64_t DataProvider::getNextBatch(int64_t size, DataBatch* batch) { + int64_t batchSize = doubleBuffer_ ? getNextBatchFromBuffer(size, batch) + : getNextBatchInternal(size, batch); + + if (!batchSize) return 0; + + if (!config_.constant_slots_size()) return batchSize; + + auto& constantSlots = *constantSlots_; + constantSlots.resize(config_.constant_slots_size()); + + for (int i = 0; i < config_.constant_slots_size(); ++i) { + MemoryHandlePtr handle = + constantSlots[i] ? constantSlots[i]->getMemoryHandle() : nullptr; + Matrix::resizeOrCreate(constantSlots[i], + batchSize, + 1, // = width + false, // = trans + useGpu_); // = useGpu + if (handle != constantSlots[i]->getMemoryHandle()) { + // memory buf was reallocated. We need to initialize the value + constantSlots[i]->assign(config_.constant_slots(i)); + } + batch->appendData(constantSlots[i], + batch->getStream(0).sequenceStartPositions); + } + + return batchSize; +} + +int64_t DataProvider::getNextBatchFromBuffer(int64_t size, DataBatch* batch) { + CHECK(doubleBuffer_ != nullptr); + + if (doubleBuffer_->getBatchSize() != size) { + doubleBuffer_->setBatchSize(size); + } + + doubleBuffer_->removeOneBatch(batch); + return batch->getSize(); +} + +void DataProvider::initAsyncLoader() { + if (doubleBuffer_ == nullptr) { + doubleBuffer_.reset(new DoubleBuffer(this, useGpu_)); + } + useGpu_ = false; // Avoid D2D copy, it will delay the computing performance +} + +SimpleDataProviderBase::SimpleDataProviderBase(const DataConfig& config, + bool useGpu, + bool withInfo) + : DataProvider(config, useGpu) { + /* initialize the size of a sample, and the buffer */ + sampleDim_ = config_.feat_dim() * (2 * config_.context_len() + 1); + bufferCapacity_ = config_.buffer_capacity(); + withInfo_ = withInfo; + sampleNumInBuf_ = 0; + nextItemIndex_ = 0; + + /* malloc buffer in cpu */ + hInputDataBuf_ = std::make_shared(bufferCapacity_, sampleDim_); + hInputLabelBuf_ = std::make_shared(bufferCapacity_); + hInputInfoBuf_ = std::make_shared(bufferCapacity_); +} + +void SimpleDataProviderBase::shuffle() { + int i, t; + int len = sampleNumInBuf_; + std::vector temp(sampleDim_); + real* data = hInputDataBuf_->getData(); + int* label = hInputLabelBuf_->getData(); + int* info = hInputInfoBuf_->getData(); + int sampleSz = sizeof(real) * sampleDim_; + for (i = 0; i < len; i++) { + int randNum = rand(); // NOLINT TODO(yuyang18): Use rand_r instead? + t = randNum % (len - i) + i; + // swap + if (i != t) { + // swap data + memcpy(&temp[0], &data[i * sampleDim_], sampleSz); + memcpy(&data[i * sampleDim_], &data[t * sampleDim_], sampleSz); + memcpy(&data[t * sampleDim_], &temp[0], sampleSz); + std::swap(label[i], label[t]); + if (withInfo_) { + std::swap(info[i], info[t]); + } + } + } +} + +int64_t SimpleDataProviderBase::getNextBatchInternal(int64_t size, + DataBatch* batch) { + CHECK(batch != NULL); + batch->clear(); + + int64_t startIndex; + int64_t cpySize; + + std::lock_guard guard(lock_); + if (sampleNumInBuf_ - nextItemIndex_ < size) { + int64_t n = fillBuffer(); + VLOG(1) << "fillBuffer return " << n << " samples.\n"; + } + + startIndex = nextItemIndex_; + cpySize = std::min(size, sampleNumInBuf_ - nextItemIndex_); + nextItemIndex_ += cpySize; + + if (cpySize > 0) { + real* data = hInputDataBuf_->getData() + startIndex * sampleDim_; + int* label = hInputLabelBuf_->getData() + startIndex; + int* info = hInputInfoBuf_->getData() + startIndex; + + MatrixPtr& dataBatch = *dataBatch_; // get the thread local object + IVectorPtr& labelBatch = *labelBatch_; // get the thread local object + IVectorPtr& infoBatch = *infoBatch_; // get the thread local object + if (!dataBatch) { + dataBatch = Matrix::create(cpySize, sampleDim_, false, useGpu_); + labelBatch = IVector::create(cpySize, useGpu_); + if (withInfo_) { + infoBatch = IVector::create(cpySize, 0); + } + } else { + dataBatch->resize(cpySize, sampleDim_); + labelBatch->resize(cpySize); + if (withInfo_) { + infoBatch->resize(cpySize); + } + } + dataBatch->copyFrom(data, cpySize * sampleDim_); + labelBatch->copyFrom(label, cpySize); + batch->appendData(dataBatch); + batch->appendLabel(labelBatch); + if (withInfo_) { + infoBatch->copyFrom(info, cpySize); + batch->appendLabel(infoBatch); + } + } + + batch->setSize(cpySize); + return cpySize; +} + +void SimpleDataProviderBase::reset() { + sampleNumInBuf_ = 0; + nextItemIndex_ = 0; + DataProvider::reset(); +} + +int64_t SimpleDataProviderBase::getSize() { + LOG(FATAL) << "Currently, not implemented"; + return 0; +} + +int64_t SimpleDataProviderBase::fillBuffer() { + int64_t n = sampleNumInBuf_ - nextItemIndex_; + + /* flash the remaining data to the beginning of the buffer */ + if (n > 0) { + hInputDataBuf_->copyFrom( + hInputDataBuf_->getData() + nextItemIndex_ * sampleDim_, + n * sampleDim_); + hInputLabelBuf_->copyFrom(hInputLabelBuf_->getData() + nextItemIndex_, n); + if (withInfo_) { + hInputInfoBuf_->copyFrom(hInputInfoBuf_->getData() + nextItemIndex_, n); + } + } + + sampleNumInBuf_ = + n + fillBufferImp(hInputDataBuf_->getData() + n * sampleDim_, + hInputLabelBuf_->getData() + n, + hInputInfoBuf_->getData() + n, + bufferCapacity_ - n); + + /* for stachastic gradient training */ + if (!skipShuffle_) { + shuffle(); + } + + nextItemIndex_ = 0; + + return sampleNumInBuf_; +} + +SimpleDataProvider::SimpleDataProvider(const DataConfig& config, bool useGpu) + : SimpleDataProviderBase(config, useGpu, /* withInfo= */ false), + currentSampleIndex_(0) { + loadData(config_.files()); +} + +SimpleDataProvider::~SimpleDataProvider() {} + +int64_t SimpleDataProvider::fillBufferImp(real* data, + int* label, + int* info, + int64_t size) { + (void)info; + int64_t n = std::min(labels_.size() - currentSampleIndex_, size); + memcpy(data, + &data_[currentSampleIndex_ * sampleDim_], + n * sampleDim_ * sizeof(real)); + memcpy(label, &labels_[currentSampleIndex_], sizeof(int) * n); + currentSampleIndex_ += n; + + return n; +} + +void SimpleDataProvider::reset() { + currentSampleIndex_ = 0; + SimpleDataProviderBase::reset(); +} + +void SimpleDataProvider::loadData(const std::string& fileName) { + std::ifstream is(fileName); + CHECK(is) << "Fail to open " << fileName; + std::string line; + while (is) { + if (!getline(is, line)) break; + LOG(INFO) << "load data file " << line; + loadDataFile(line); + } + LOG(INFO) << "read done, num of instance=" << labels_.size() + << " data size=" << data_.size(); +} + +void SimpleDataProvider::loadDataFile(const std::string& fileName) { + std::ifstream is(fileName); + std::string line; + std::vector pieces; + while (is) { + if (!getline(is, line)) break; + str::split(line, ' ', &pieces); + CHECK_EQ((uint64_t)(sampleDim_ + 1), pieces.size()) + << " Dimension mismatch, " << pieces.size() - 1 << " in " << fileName + << " " << sampleDim_ << " from config"; + labels_.push_back(atoi(pieces[0].c_str())); + for (int i = 0; i < sampleDim_; ++i) { + data_.push_back(atof(pieces[i + 1].c_str())); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/dataproviders/DataProvider.h b/paddle/legacy/gserver/dataproviders/DataProvider.h new file mode 100644 index 0000000000000000000000000000000000000000..c2e1c5fdd6d504b77873aaeeba3611dff6d8f738 --- /dev/null +++ b/paddle/legacy/gserver/dataproviders/DataProvider.h @@ -0,0 +1,480 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "DataConfig.pb.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/parameter/Argument.h" +#include "paddle/legacy/utils/ClassRegistrar.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Queue.h" +#include "paddle/legacy/utils/ThreadLocal.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { +/** + * @def REGISTER_DATA_PROVIDER + * @brief Macro for registering a data provider. The class type should contain + * a consturctor with parameter (DataConfig, bool). + */ +#define REGISTER_DATA_PROVIDER(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name([]() { \ + DataProvider::registrar_.registerClass( \ + #__type_name, \ + [](DataConfig conf, ModelConfig, bool useGpu) -> DataProvider* { \ + DataProvider* dp = new __class_name(conf, useGpu); \ + return dp; \ + }); \ + }) + +/** + * @def REGISTER_DATA_PROVIDER_EX + * @brief Macro for registering a data provider, which contains a constructor + * with parameter (DataConfig, ModelConfig, bool). + */ +#define REGISTER_DATA_PROVIDER_EX(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name([] { \ + DataProvider::registrar_.registerClass<__class_name>(#__type_name); \ + }) + +class DataBatch; +class BufferBatch; +typedef std::shared_ptr DataBatchPtr; +typedef std::shared_ptr BufferBatchPtr; +/** + * @brief Data for batch training a neural network + */ +class DataBatch { + public: + DataBatch() : size_(0) { data_.clear(); } + /** + * @brief Get batch size + * @return batch size + */ + int64_t getSize() const { return size_; } + /** + * @brief Get num of sequences of sequence data + * @return num of sequences + */ + int64_t getNumSequences() const { + if (data_.empty()) return size_; + return data_[0].sequenceStartPositions + ? data_[0].sequenceStartPositions->getSize() - 1 + : size_; + } + /** + * @brief Set batch size + * @param[in] size size + */ + void setSize(int64_t size) { size_ = size; } + /** + * @brief Get size of argument vector + * @return size of argument vector + * @note For usual supervised learning, input data and label is needed, + * then there will be two argument. + */ + int64_t getNumStreams() const { return data_.size(); } + + /** + * @brief Get a argument with index i + * @param[in] i index in argument vector + * @return a argument with index i + */ + const Argument& getStream(int i) const { return data_[i]; } + /** + * @brief Get all argument + * @return an argument vector + */ + std::vector& getStreams() { return data_; } + /** + * @brief Get all argument const + * @return an argument vector + */ + std::vector getStreams() const { return data_; } + /** + * @brief Clear DataBatch + */ + void clear() { + data_.clear(); + size_ = 0; + } + + /** + * @brief Append data to DataBatch + * @param[in] data matrix data + * @note The order in which each data stream is appended must match the order + * specified in stream_names of DataConfig. The stream_names can be obtained + * using DataProvider::getStreamNames(). + */ + void appendData(MatrixPtr data) { + Argument argu; + argu.value = data; + data_.push_back(argu); + } + + /** + * @brief Append sequence data to DataBatch + * @param[in] data matrix data + * @param[in] sequenceStartPositions sequence data + * @note The order in which each data stream is appended must match the order + * specified in stream_names of DataConfig. The stream_names can be obtained + * using DataProvider::getStreamNames(). + */ + void appendData(const MatrixPtr& data, + const ICpuGpuVectorPtr& sequenceStartPositions) { + Argument argu; + argu.value = data; + argu.sequenceStartPositions = sequenceStartPositions; + data_.push_back(argu); + } + /** + * @brief Append label data + * @param[in] label label data + * @param[in] value matrix data, default null + */ + void appendLabel(IVectorPtr label, MatrixPtr value = nullptr) { + Argument argu; + argu.ids = label; + argu.value = value; + data_.push_back(argu); + } + + /* + * @brief Append argument + * @param[in] argus DataBatch.getStreams() + * @param[in] size DataBatch.getSize() + * @param[in] dataId sub dataprovider id (in MultiDataProvider) + */ + void appendArguments(const std::vector& argus, + int size, + int dataId) { + size_ += size; + for (const auto& argu : argus) { + data_.push_back(argu); + data_.back().dataId = dataId; + } + } + + protected: + /** + * @brief batch size + */ + int64_t size_; + /** + * @brief A batch data consist of a Argument vector, + * An argument corresponds to a type of input data. + */ + std::vector data_; +}; + +class BufferBatch { + public: + BufferBatch() { + hlStream_ = HPPL_STREAM_DEFAULT; + hlEvent_ = NULL; + batchData_ = NULL; + } + ~BufferBatch() { + if (hlEvent_) { + hl_destroy_event(hlEvent_); + hlEvent_ = NULL; + } + delete batchData_; + batchData_ = NULL; + } + + void setDataBatch(DataBatch* batchData) { batchData_ = batchData; } + DataBatch* getDataBatch() { return batchData_; } + + void setCuStream(hl_stream_t stream) { hlStream_ = stream; } + hl_stream_t getCuStream() const { return hlStream_; } + + void setCuEvent(hl_event_t event) { hlEvent_ = event; } + + hl_event_t getCuEvent() const { return hlEvent_; } + + void createCuEvent() { + if (!hlEvent_) { + hlStream_ = HPPL_STREAM_1; + hl_create_event(&hlEvent_); + } + } + + void syncEvent() { + if (hlEvent_) { + hl_stream_wait_event(hlStream_, hlEvent_); + } + } + + void swap(BufferBatch* bufBatch); + void clone(DataBatch* srcBatch, bool useGpu); + + protected: + DataBatch* batchData_; + hl_stream_t hlStream_; + hl_event_t hlEvent_; +}; + +class DataProvider; +typedef std::shared_ptr DataProviderPtr; + +typedef Queue BufferBatchQueue; + +class DoubleBuffer { + public: + DoubleBuffer(DataProvider* dataPool, bool useGpu, int64_t batchSize = 0); + virtual ~DoubleBuffer(); + void removeOneBatch(DataBatch* dataBatch); + + void setBatchSize(int64_t newBatchSize) { batchSize_ = newBatchSize; } + + int64_t getBatchSize() { return batchSize_; } + + void startAsyncLoad(); + void finishAsyncLoad() { + stopping_ = true; + taskReadySem_.post(); + if (asyncLoader_) { + asyncLoader_->join(); + } + } + + void setPending(bool pending) { pending_ = pending; } + + protected: + virtual void asyncLoadBatch(); + void insertOneBatch(DataBatch* batch); + + DataProvider* dataPool_; + bool useGpu_; + int32_t batchSize_; + ThreadLocal usingBatch_; + BufferBatchQueue* dataQueue_; + BufferBatchQueue* bufferQueue_; + std::unique_ptr asyncLoader_; + Semaphore taskReadySem_; + bool stopping_; + bool pending_; +}; + +/** + * @brief Base class for DataProvider, which supplies data for training + * @note It can supplies multiple streams of data. + * For typical supervised training, there are two streams: + * one is for input, one is for label. + */ +class DataProvider { + public: + static ClassRegistrar registrar_; + static DataProvider* create(const DataConfig& config, + const ModelConfig& modelConfig, + bool useGpu = FLAGS_use_gpu); + + /** + * @brief create only used for unittest. + */ + inline static DataProvider* create(const DataConfig& config, + bool useGpu = FLAGS_use_gpu) { + return create(config, ModelConfig(), useGpu); + } + + DataProvider(const DataConfig& config, bool useGpu) + : config_(config), + skipShuffle_(false), + usageRatio_(config.usage_ratio()), + useGpu_(useGpu) { + if (config_.async_load_data()) { + initAsyncLoader(); + } + } + virtual ~DataProvider() {} + + const DataConfig& getConfig() const { return config_; } + + void setSkipShuffle() { skipShuffle_ = true; } + + /** + * @brief Get next batch of training samples + * @param[in] size size of training samples to get + * @param[out] batch a batch of training samples + * @return actual size of obtained training samples + */ + int64_t getNextBatch(int64_t size, DataBatch* batch); + + /** + * @brief Shuffle the data set + */ + virtual void shuffle() = 0; + + /** + * @brief reset all the value of index + * @note reset() must be called before any calls to getNextBatch() + * IMPORTANT: subclass reset() should always call the base class reset() + * at the end of the function + */ + virtual void reset() { + if (doubleBuffer_ != nullptr) { + doubleBuffer_->startAsyncLoad(); + } + } + + /** + * @brief Get the size of training samples + * @return the number of training samples in the data set. + * @note return -1 to indicate unlimited number of samples. + */ + virtual int64_t getSize() = 0; + + /** + * @brief Get next batch training samples internally + * @param[in] size size of training samples to get + * @param[out] batch a batch of training samples + * @return actual size of obtained training samples + */ + virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch) = 0; + + protected: + DataConfig config_; + bool skipShuffle_; + float usageRatio_; + bool useGpu_; + std::unique_ptr doubleBuffer_; + ThreadLocal> constantSlots_; + /** + * @@brief Get next batch training samples from buffer + * @param[in] size size of training samples to get + * @param[out] batch a batch of training samples + * @return actual size of obtained training samples + */ + int64_t getNextBatchFromBuffer(int64_t size, DataBatch* batch); + + void initAsyncLoader(); +}; + +/** + * A data provider which does nothing. It only serves as providing + * necessary configurations such as stream_names + */ +class DummyDataProvider : public DataProvider { + public: + DummyDataProvider(const DataConfig& config, bool useGpu) + : DataProvider(config, useGpu) {} + virtual void shuffle() {} + virtual void reset() { DataProvider::reset(); } + virtual int64_t getSize() { return 0; } + virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch) { + (void)size; + (void)batch; + return 0; + } +}; + +/** + * Data provider for one input and one integer label. + */ +class SimpleDataProviderBase : public DataProvider { + protected: + /// sample feature dimension + int64_t sampleDim_; + /// the number of samples + int64_t bufferCapacity_; + int64_t sampleNumInBuf_; + /// next item to read in buffer + int64_t nextItemIndex_; + /// some user defined info for validation + bool withInfo_; + + /// data buffer: bufferCapacity_ * nDataDim_ + CpuMatrixPtr hInputDataBuf_; + + /// label buffer:bufferCapacity_ * 1 + CpuIVectorPtr hInputLabelBuf_; + + /// info buffer:bufferCapacity_ * 1 + CpuIVectorPtr hInputInfoBuf_; + + ThreadLocal dataBatch_; + ThreadLocal labelBatch_; + ThreadLocal infoBatch_; + + RWLock lock_; + + public: + SimpleDataProviderBase(const DataConfig& config, bool useGpu, bool withInfo); + ~SimpleDataProviderBase() {} + + void shuffle(); + + virtual void reset(); + + virtual int64_t getSize(); + + virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); + + /// return the number of samples in the buffer + int64_t fillBuffer(); + + protected: + /** + * @brief Fill at most size samples into data and label. + * + * Each input is stored in contiguous memory locations in data. + * + * data[n * sampleDim_] .. data[n * sampleDim_ + sampleDim_ - 1] is for + * the input of the n-th sample. + * + * label[n] is the label for the n-th sample. + */ + virtual int64_t fillBufferImp(real* data, + int* label, + int* info, + int64_t size) = 0; +}; + +class SimpleDataProvider : public SimpleDataProviderBase { + public: + SimpleDataProvider(const DataConfig& config, bool useGpu); + ~SimpleDataProvider(); + virtual void reset(); + + protected: + void loadData(const std::string& fileName); + void loadDataFile(const std::string& fileName); + virtual int64_t fillBufferImp(real* data, + int* label, + int* info, + int64_t size); + + protected: + size_t currentSampleIndex_; + std::vector labels_; + std::vector data_; +}; + +} // namespace paddle diff --git a/paddle/gserver/dataproviders/DataProviderGroup.h b/paddle/legacy/gserver/dataproviders/DataProviderGroup.h similarity index 100% rename from paddle/gserver/dataproviders/DataProviderGroup.h rename to paddle/legacy/gserver/dataproviders/DataProviderGroup.h diff --git a/paddle/legacy/gserver/dataproviders/MultiDataProvider.cpp b/paddle/legacy/gserver/dataproviders/MultiDataProvider.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e5fc6d8a88fe2c03cc74b4a38e999d11d676dfdf --- /dev/null +++ b/paddle/legacy/gserver/dataproviders/MultiDataProvider.cpp @@ -0,0 +1,122 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MultiDataProvider.h" +#include +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +using namespace std; + +MultiDataProvider::MultiDataProvider(const DataConfig& config, + const ModelConfig& modelConfig, + bool useGpu) + : DataProvider(config, useGpu) { + bool atLeastOneMainDataFlag = false; + totalDataRatio_ = 0; + LOG(INFO) << "MultiDataProvider: sub data provider size: " + << config.sub_data_configs_size(); + LOG(INFO) << "MultiDataProvider: for_test: " << config.for_test(); + isTestMode_ = config.for_test(); + for (int i = 0; i < config.sub_data_configs_size(); i++) { + LOG(INFO) << "dataRatio of sub(" << i + << ") is: " << config.sub_data_configs(i).data_ratio(); + totalDataRatio_ += config.sub_data_configs(i).data_ratio(); + if (config.sub_data_configs(i).is_main_data()) { + LOG(INFO) << "main data is [" << i << "]"; + atLeastOneMainDataFlag = true; + } + } + CHECK(atLeastOneMainDataFlag) << "all sub dataproviders in MultiData do not" + << " have is_main_data flag"; + LOG(INFO) << "totalDataRatio_=" << totalDataRatio_; + DataConfig subConfig; + int subDataProviderCount = config.sub_data_configs_size(); + if (isTestMode()) { + LOG(INFO) << "construct MultiDataProvider in test mode"; + } else { + LOG(INFO) << "construct MultiDataProvider in train mode"; + } + subDataProviders_.resize(subDataProviderCount); + for (int i = 0; i < subDataProviderCount; i++) { + subConfig = config.sub_data_configs(i); + if (subConfig.async_load_data()) { + LOG(INFO) << "can not use async_load_data in sub dataprovider of " + "MultiDataProvider"; + subConfig.set_async_load_data(false); + } + subDataProviders_[i] = std::unique_ptr( + DataProvider::create(subConfig, modelConfig, useGpu_)); + } +} + +void MultiDataProvider::reset() { + for (auto& elem : subDataProviders_) { + elem->reset(); + } + DataProvider::reset(); +} + +void MultiDataProvider::shuffle() { + for (auto& elem : subDataProviders_) { + elem->shuffle(); + } +} + +int64_t MultiDataProvider::getNextBatchInternal(int64_t size, + DataBatch* batch) { + batch->clear(); + for (size_t i = 0; i < subDataProviders_.size(); ++i) { + // calc size according to data ratio + int64_t subSize = + (int64_t)(1.0 * size * config_.sub_data_configs(i).data_ratio() / + totalDataRatio_); + DataBatch subBatch; + int64_t realSize = + subDataProviders_[i]->getNextBatchInternal(subSize, &subBatch); + if (realSize == 0) { + // current subDataProvider has no data + if (!isTestMode()) { + // in train mode + if (config_.sub_data_configs(i).is_main_data()) { + // is main data provider. then return 0 + batch->clear(); + return 0; + } else { + // not main data provider, reset current subDataProvider and try again + subDataProviders_[i]->reset(); + subBatch.clear(); + realSize = + subDataProviders_[i]->getNextBatchInternal(subSize, &subBatch); + CHECK_GT(realSize, 0); + } + } else { + // in test mode, make an empty argument + Argument emptyArgu; + std::vector argus; + argus.push_back(emptyArgu); + batch->appendArguments(argus, 0, -1); + continue; + } + } + batch->appendArguments(subBatch.getStreams(), subBatch.getSize(), i); + } + return batch->getSize(); +} + +REGISTER_DATA_PROVIDER_EX(multi, MultiDataProvider); + +} // namespace paddle diff --git a/paddle/gserver/dataproviders/MultiDataProvider.h b/paddle/legacy/gserver/dataproviders/MultiDataProvider.h similarity index 100% rename from paddle/gserver/dataproviders/MultiDataProvider.h rename to paddle/legacy/gserver/dataproviders/MultiDataProvider.h diff --git a/paddle/gserver/dataproviders/ProtoReader.h b/paddle/legacy/gserver/dataproviders/ProtoReader.h similarity index 100% rename from paddle/gserver/dataproviders/ProtoReader.h rename to paddle/legacy/gserver/dataproviders/ProtoReader.h diff --git a/paddle/legacy/gserver/dataproviders/PyDataProvider.cpp b/paddle/legacy/gserver/dataproviders/PyDataProvider.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0827bd39d4cc78ef5658d437b6502f2e60e90b4c --- /dev/null +++ b/paddle/legacy/gserver/dataproviders/PyDataProvider.cpp @@ -0,0 +1,498 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PyDataProvider.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +#ifndef PADDLE_NO_PYTHON +REGISTER_DATA_PROVIDER(py, PyDataProvider); +#endif + +PyDataProvider::PyDataProvider(const DataConfig& config, + bool useGpu, + bool loadDataAll) + : DataProvider(config, useGpu), batchSize_(0) { + PyGuard guard; + pyModuleName_ = config_.load_data_module(); + pyClassName_ = config_.load_data_object(); + if (config_.load_data_args() != "") { + pyUserArgs_["load_data_args"] = config_.load_data_args(); + } + + if (loadDataAll) { + std::vector fileList; + if (!config_.files().empty()) { + loadFileList(config_.files(), fileList); + } + loadData(fileList); + } +} + +void PyDataProvider::loadData(const std::vector& fileList) { + VLOG(1) << "module:" << pyModuleName_ << " class:" << pyClassName_; + classInstance_ = + createPythonClass(pyModuleName_, pyClassName_, fileList, pyUserArgs_); + CHECK(classInstance_) << "Create class instance failed."; + PyObjectPtr obj(PyObject_CallMethod( + classInstance_.get(), const_cast("getHeader"), NULL)); + CHECK_PY(obj) << "Call function getHeader failed."; + std::string headerInfo = + std::string(PyString_AsString(obj.get()), PyString_Size(obj.get())); + parseHeaderData(headerInfo); + feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); +} + +void PyDataProvider::parseHeaderData(const std::string& headerData) { + char* pHeader = const_cast(headerData.c_str()); + char* pHeaderEnd = pHeader + headerData.size(); + slotNum_ = readT(pHeader, pHeaderEnd); + unsigned int useSequenceFlag = readT(pHeader, pHeaderEnd); + isIID_ = useSequenceFlag != 1; + slots_.clear(); + slots_.reserve(slotNum_); + for (size_t i = 0; i < slotNum_; ++i) { + unsigned int slotType = readT(pHeader, pHeaderEnd); + unsigned int slotDim = readT(pHeader, pHeaderEnd); + slots_.emplace_back(); + slots_.back().dim = slotDim; + slots_.back().type = static_cast(slotType); + } +} + +void PyDataProvider::resetSlots() { + for (auto& slot : slots_) { + slot.indexData.clear(); + slot.denseData.clear(); + slot.sparseNonValueData.clear(); + slot.sparseFloatValueData.clear(); + slot.indices.clear(); + slot.sequenceStartPositions.clear(); + slot.sampleSequenceIdVec.clear(); + slot.subSequenceStartPositions.clear(); + slot.strData.clear(); + } +} + +void PyDataProvider::fillDenseSlot(ProtoSlot& slot, + char*& data, + const char* dataEnd) { + unsigned int dim = slot.dim; + slot.sampleNum = readT(data, dataEnd); + slot.denseData.resize(slot.sampleNum * dim); +#ifdef PADDLE_TYPE_DOUBLE + CHECK_LE(data + sizeof(real) * dim * slot.sampleNum, dataEnd) + << "std::copy data is out of range"; + // PyDataProvider always provide data in float + float* dat = reinterpret_cast(data); + std::copy(dat, dat + slot.sampleNum * dim, slot.denseData.begin()); +#else + memcpyWithCheck(slot.denseData.data(), + data, + sizeof(real) * dim * slot.sampleNum, + dataEnd); +#endif + // PyDataProvider always provide data in float + data += sizeof(float) * dim * slot.sampleNum; +} + +void PyDataProvider::fillSparseNonValueSlot(ProtoSlot& slot, + char*& data, + const char* dataEnd) { + slot.sampleNum = readT(data, dataEnd); + unsigned int* indexPtr = (unsigned int*)data; + CHECK_LE(data + sizeof(unsigned int) * slot.sampleNum, dataEnd) + << "Vector assign value is out of range"; + slot.indices.assign(indexPtr, indexPtr + slot.sampleNum); + data += sizeof(unsigned int) * slot.sampleNum; + unsigned int length = 0; + length = readT(data, dataEnd); + slot.indices.push_back(length); + slot.sparseNonValueData.resize(length); + memcpyWithCheck(slot.sparseNonValueData.data(), + data, + sizeof(unsigned int) * length, + dataEnd); + data += sizeof(unsigned int) * length; +} + +void PyDataProvider::fillSparseValueSlot(ProtoSlot& slot, + char*& data, + const char* dataEnd) { + slot.sampleNum = readT(data, dataEnd); + unsigned int* indexPtr = (unsigned int*)data; + CHECK_LE(data + sizeof(unsigned int) * slot.sampleNum, dataEnd) + << "Vector assign value is out of range"; + slot.indices.assign(indexPtr, indexPtr + slot.sampleNum); + data += sizeof(unsigned int) * slot.sampleNum; + unsigned int length = 0; + length = readT(data, dataEnd); + unsigned int* colPtr = reinterpret_cast(data); + CHECK_LE(data + sizeof(unsigned int) * length, dataEnd) + << "Data is out of range"; + data += sizeof(unsigned int) * length; + size_t colLen = readT(data, dataEnd); + CHECK_EQ(colLen, length); + float* valuePtr = reinterpret_cast(data); + CHECK_LE(data + sizeof(real) * length, dataEnd) << "Data is out of range"; + data += sizeof(real) * length; + slot.indices.push_back(length); + slot.sparseFloatValueData.resize(length); + for (unsigned int ii = 0; ii < length; ++ii) { + slot.sparseFloatValueData[ii].col = colPtr[ii]; + slot.sparseFloatValueData[ii].value = valuePtr[ii]; + } +} + +void PyDataProvider::fillIndexSlot(ProtoSlot& slot, + char*& data, + const char* dataEnd) { + slot.sampleNum = readT(data, dataEnd); + CHECK_LE(data + sizeof(unsigned int) * slot.sampleNum, dataEnd) + << "Vector assign is out of range"; + slot.indexData.assign(reinterpret_cast(data), + reinterpret_cast(data) + slot.sampleNum); + data += sizeof(unsigned int) * slot.sampleNum; +} + +void PyDataProvider::fillStringSlot(ProtoSlot& slot, + char*& data, + const char* dataEnd) { + slot.sampleNum = readT(data, dataEnd); + for (unsigned int i = 0; i < slot.sampleNum; ++i) { + size_t len = readT(data, dataEnd); + auto str_begin = data; + data += len; + CHECK_LE(data, dataEnd) << "Data is out of range"; + slot.strData.emplace_back(str_begin, len); + } +} + +void PyDataProvider::fillSlotsByStr(const std::string& samples) { + char* data = const_cast(samples.c_str()); + char* dataEnd = data + samples.size(); + batchSize_ = readT(data, dataEnd); + if (0 == batchSize_) { + return; + } + + for (size_t j = 0; j < slotNum_; ++j) { + auto& slot = slots_[j]; + CHECK(SlotDef::INDEX >= slot.type || SlotDef::STRING == slot.type) + << " Slot type:" << slot.type << " is out of range."; + CHECK_GE(slot.type, SlotDef::VECTOR_DENSE) << " Slot type:" << slot.type + << " is out of range."; + switch (slot.type) { + case SlotDef::VECTOR_DENSE: + fillDenseSlot(slot, data, dataEnd); + break; + case SlotDef::VECTOR_SPARSE_NON_VALUE: + fillSparseNonValueSlot(slot, data, dataEnd); + break; + case SlotDef::VECTOR_SPARSE_VALUE: + fillSparseValueSlot(slot, data, dataEnd); + break; + case SlotDef::INDEX: + fillIndexSlot(slot, data, dataEnd); + break; + case SlotDef::VAR_MDIM_DENSE: + LOG(FATAL) << "Not implemented"; + break; + case SlotDef::VAR_MDIM_INDEX: + LOG(FATAL) << "Not implemented"; + break; + case SlotDef::STRING: + fillStringSlot(slot, data, dataEnd); + break; + } + } + // read sequenceStartPositions + for (size_t j = 0; j < slotNum_; ++j) { + auto& slot = slots_[j]; + if (!iidData()) { + unsigned int sequenceNum = readT(data, dataEnd); + slot.sequenceNum = sequenceNum; + for (size_t i = 0; i < sequenceNum; ++i) { + slot.sequenceStartPositions.push_back( + readT(data, dataEnd)); + } + for (size_t i = 0; i < sequenceNum; ++i) { + size_t begin = slot.sequenceStartPositions[i]; + size_t end = (i < sequenceNum - 1) ? slot.sequenceStartPositions[i + 1] + : slot.sampleNum; + for (size_t ii = begin; ii < end; ++ii) { + slot.sampleSequenceIdVec.push_back(ii); + } + } + } else { + for (size_t i = 0; i < slot.sampleNum; ++i) { + slot.sampleSequenceIdVec.push_back(i); + } + } + } + // read subSequenceStartPositions, not all slots have this infomation. + for (size_t j = 0; j < slotNum_; ++j) { + auto& slot = slots_[j]; + if (!iidData() && data != dataEnd) { + unsigned int subSequenceNum = readT(data, dataEnd); + slot.subSequenceNum = subSequenceNum; + for (size_t i = 0; i < subSequenceNum; ++i) { + slot.subSequenceStartPositions.push_back( + readT(data, dataEnd)); + } + } + } +} + +void PyDataProvider::reset() { + { // Invoke PyDataProvider Reset + PyGuard guard; + PyObjectPtr obj(PyObject_CallMethod( + classInstance_.get(), const_cast("reset"), NULL)); + CHECK_PY(obj) << "Call function reset failed."; + } + + if (!skipShuffle_) { + // Invoke PyDataProvider Shuffle + shuffle(); + } + DataProvider::reset(); +} + +void PyDataProvider::shuffle() { + // py shuffle + PyGuard guard; + PyObjectPtr obj(PyObject_CallMethod( + classInstance_.get(), const_cast("shuffle"), NULL)); + CHECK_PY(obj) << "Call function shuffle failed."; +} + +void PyDataProvider::handleDenseSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments) { + unsigned int dim = slot.dim; + Matrix::resizeOrCreate(cpuArguments[slotIndex].value, + slot.sampleNum, + dim, + false, // trans = false + false); // useGpu = false + real* buf = cpuArguments[slotIndex].value->getData(); + for (size_t i = 0; i < slot.sampleNum; ++i) { + memcpyWithCheck(buf + i * dim, + slot.denseData.data() + slot.sampleSequenceIdVec[i] * dim, + sizeof(real) * dim, + slot.denseData.data() + slot.denseData.size()); + } +} + +void PyDataProvider::handleSparseNonValueSlot( + ProtoSlot& slot, size_t slotIndex, std::vector& cpuArguments) { + unsigned int dim = slot.dim; + if (!(cpuArguments[slotIndex].value)) { + cpuArguments[slotIndex].value = + Matrix::createSparseMatrix(slot.sampleNum, + dim, + slot.sampleNum /*DEFAULT_AVG_WIDTH = 1*/, + NO_VALUE, + SPARSE_CSR, + false, + useGpu_); + } + auto mat = cpuArguments[slotIndex].value; + mat->resize(slot.sampleNum, dim, slot.sampleNum, NO_VALUE, SPARSE_CSR); + if (std::dynamic_pointer_cast(mat)) { + std::dynamic_pointer_cast(mat)->copyFrom( + slot.sampleSequenceIdVec.data(), + slot.indices.data(), + slot.sparseNonValueData.data(), + HPPL_STREAM_1); + } else if (std::dynamic_pointer_cast(mat)) { + std::dynamic_pointer_cast(mat)->copyFrom( + slot.sampleSequenceIdVec.data(), + slot.indices.data(), + slot.sparseNonValueData.data()); + } else { + LOG(FATAL) << "Not Supported"; + } +} + +void PyDataProvider::handleSparseValueSlot( + ProtoSlot& slot, size_t slotIndex, std::vector& cpuArguments) { + unsigned int dim = slot.dim; + if (!(cpuArguments[slotIndex].value)) { + cpuArguments[slotIndex].value = + Matrix::createSparseMatrix(slot.sampleNum, + dim, + slot.sampleNum /*DEFAULT_AVG_WIDTH = 1*/, + FLOAT_VALUE, + SPARSE_CSR, + false, + useGpu_); + } + auto mat = cpuArguments[slotIndex].value; + mat->resize(slot.sampleNum, dim, slot.sampleNum, FLOAT_VALUE, SPARSE_CSR); + if (std::dynamic_pointer_cast(mat)) { + std::dynamic_pointer_cast(mat)->copyFrom( + slot.sampleSequenceIdVec.data(), + slot.indices.data(), + slot.sparseFloatValueData.data(), + HPPL_STREAM_DEFAULT); + } else if (std::dynamic_pointer_cast(mat)) { + std::dynamic_pointer_cast(mat)->copyFrom( + slot.sampleSequenceIdVec.data(), + slot.indices.data(), + slot.sparseFloatValueData.data()); + } else { + LOG(FATAL) << "Not Supported"; + } +} + +void PyDataProvider::handleIndexSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments) { + IVector::resizeOrCreate(cpuArguments[slotIndex].ids, + slot.sampleNum, + /*useGpu_*/ false); + int* buf = cpuArguments[slotIndex].ids->getData(); + for (size_t i = 0; i < slot.sampleNum; ++i) { + buf[i] = slot.indexData[slot.sampleSequenceIdVec[i]]; + } +} + +void PyDataProvider::handleStringSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments) { + if (cpuArguments[slotIndex].strs) { + cpuArguments[slotIndex].strs->resize(slot.sampleNum); + } else { + cpuArguments[slotIndex].strs = + std::make_shared>(slot.sampleNum); + } + for (size_t i = 0; i < slot.sampleNum; ++i) { + (*cpuArguments[slotIndex].strs)[i] = + slot.strData[slot.sampleSequenceIdVec[i]]; + } +} + +int64_t PyDataProvider::getNextBatchInternal(int64_t size, DataBatch* batch) { + PyGuard guard; + PyObjectPtr obj(PyObject_CallMethod(classInstance_.get(), + const_cast("getNextBatch"), + const_cast("i"), + size)); + CHECK_PY(obj) << "Call function getNextBatch failed."; + const std::string& samples = + std::string(PyString_AsString(obj.get()), PyString_Size(obj.get())); + resetSlots(); + fillSlotsByStr(samples); + size = batchSize_; + if (size <= 0) return 0; + + DataBatch& cpuBatch = *cpuBatch_; + std::vector& cpuArguments = cpuBatch.getStreams(); + cpuBatch.setSize(size); + cpuArguments.resize(slotNum_); + + if (!iidData()) { + for (size_t j = 0; j < slotNum_; ++j) { + auto& slot = slots_[j]; + ICpuGpuVector::resizeOrCreate(cpuArguments[j].sequenceStartPositions, + slot.sequenceNum + 1, + /* useGpu= */ false); + int* buf = cpuArguments[j].sequenceStartPositions->getMutableData(false); + std::copy(slot.sequenceStartPositions.begin(), + slot.sequenceStartPositions.end(), + buf); + buf[slot.sequenceStartPositions.size()] = slot.sampleNum; + + if (slot.subSequenceStartPositions.size()) { + ICpuGpuVector::resizeOrCreate(cpuArguments[j].subSequenceStartPositions, + slot.subSequenceNum + 1, + /* useGpu= */ false); + int* buf = + cpuArguments[j].subSequenceStartPositions->getMutableData(false); + std::copy(slot.subSequenceStartPositions.begin(), + slot.subSequenceStartPositions.end(), + buf); + buf[slot.subSequenceNum] = slot.sampleNum; + // check subSequenceStartPositions and sequenceStartPositions + cpuArguments[j].checkSubset(); + } + } + } + + for (size_t slotIndex = 0; slotIndex < slotNum_; ++slotIndex) { + auto& slot = slots_[slotIndex]; + SlotDef::SlotType slotType = slot.type; + switch (slotType) { + case SlotDef::VECTOR_DENSE: + handleDenseSlot(slot, slotIndex, cpuArguments); + break; + case SlotDef::VECTOR_SPARSE_NON_VALUE: + handleSparseNonValueSlot(slot, slotIndex, cpuArguments); + break; + case SlotDef::VECTOR_SPARSE_VALUE: + handleSparseValueSlot(slot, slotIndex, cpuArguments); + break; + case SlotDef::INDEX: + handleIndexSlot(slot, slotIndex, cpuArguments); + break; + case SlotDef::VAR_MDIM_DENSE: + LOG(FATAL) << "Not implemented"; + break; + case SlotDef::VAR_MDIM_INDEX: + LOG(FATAL) << "Not implemented"; + break; + case SlotDef::STRING: + handleStringSlot(slot, slotIndex, cpuArguments); + break; + } + } + + if (useGpu_) { + std::vector& cpuArguments = cpuBatch.getStreams(); + DataBatch& gpuBatch = *gpuBatch_; + std::vector& gpuArguments = gpuBatch.getStreams(); + gpuArguments.resize(cpuArguments.size()); + gpuBatch.setSize(size); + for (size_t i = 0; i < slotNum_; ++i) { + SlotDef::SlotType slotType = slots_[i].type; + if (SlotDef::VECTOR_SPARSE_VALUE == slotType || + SlotDef::VECTOR_SPARSE_NON_VALUE == slotType) { + gpuArguments[i] = cpuArguments[i]; + gpuArguments[i].sequenceStartPositions = + cpuArguments[i].sequenceStartPositions; + + if (slots_[i].subSequenceStartPositions.size()) { + gpuArguments[i].subSequenceStartPositions = + cpuArguments[i].subSequenceStartPositions; + } + } else { + gpuArguments[i].resizeAndCopyFrom( + cpuArguments[i], useGpu_, HPPL_STREAM_1); + } + } + hl_stream_synchronize(HPPL_STREAM_1); + *batch = gpuBatch; + } else { + *batch = cpuBatch; + } + + return batch->getSize(); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/dataproviders/PyDataProvider.h b/paddle/legacy/gserver/dataproviders/PyDataProvider.h new file mode 100644 index 0000000000000000000000000000000000000000..4b8bea04a1670c60d5a801ca950f59116ba50195 --- /dev/null +++ b/paddle/legacy/gserver/dataproviders/PyDataProvider.h @@ -0,0 +1,124 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "DataFormat.pb.h" +#include "DataProvider.h" + +#include + +namespace paddle { + +class PyDataProvider : public DataProvider { + public: + PyDataProvider(const DataConfig& config, + bool useGpu, + bool loadDataAll = true); + + virtual void reset(); + + // Note this size includes the sequences which are skipped because they + // are longer than the batch size + virtual int64_t getSize() { + LOG(FATAL) << "Not implement yet"; + return -1; + } + virtual void shuffle(); + + virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); + + protected: + struct ProtoSlot; + // return false if each each sample is one sequence, i.e., independent + // of other samples. + inline bool iidData() const { return isIID_; } + + void parseHeaderData(const std::string& headerData); + void fillDenseSlot(ProtoSlot& slot, char*& data, const char* dataEnd); + void fillSparseNonValueSlot(ProtoSlot& slot, + char*& data, + const char* dataEnd); + void fillSparseValueSlot(ProtoSlot& slot, char*& data, const char* dataEnd); + void fillIndexSlot(ProtoSlot& slot, char*& data, const char* dataEnd); + void fillStringSlot(ProtoSlot& slot, char*& data, const char* dataEnd); + void fillSlotsByStr(const std::string& samples); + void handleDenseSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments); + void handleSparseNonValueSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments); + void handleSparseValueSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments); + void handleIndexSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments); + void handleStringSlot(ProtoSlot& slot, + size_t slotIndex, + std::vector& cpuArguments); + void resetSlots(); + void loadData(const std::vector& fileList); + + protected: + struct ProtoSlot { + SlotDef::SlotType type; + int dim; + unsigned int sampleNum; + unsigned int sequenceNum; + unsigned int subSequenceNum; + // Store the data of index type slot + std::vector indexData; + // Store the data of dense type slot + std::vector denseData; + // Store the data of sparseNonValue type slot + std::vector sparseNonValueData; + // Store the data of sparseValue type slot + std::vector sparseFloatValueData; + // Used to store the index of each sample in slot values + std::vector indices; + // The starting position of each sequence in samples + // The last element should be the number of samples + // If empty, each sample is one sequence. + std::vector sequenceStartPositions; + // The index id of sequences in slot + std::vector sampleSequenceIdVec; + // The starting position of each subsequence in samples + // The last element should be the number of subsequence + // If empty, each sequence of sample has no subsequence. + std::vector subSequenceStartPositions; + // Store the data of string type slot + std::vector strData; + }; + std::vector slots_; + + PyObjectPtr classInstance_; + unsigned int batchSize_; + unsigned int slotNum_; + // if use sequence, isIID_ equals false, otherwise it is true. + bool isIID_; + // The name of python module name + std::string pyModuleName_; + // The name of python class name + std::string pyClassName_; + // User args set in config + std::map pyUserArgs_; + + ThreadLocalD cpuBatch_; + ThreadLocalD gpuBatch_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/dataproviders/PyDataProvider2.cpp b/paddle/legacy/gserver/dataproviders/PyDataProvider2.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8e931e40611e27caa43675c3567972384a4d9026 --- /dev/null +++ b/paddle/legacy/gserver/dataproviders/PyDataProvider2.cpp @@ -0,0 +1,1031 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef PADDLE_NO_PYTHON + +#include +#include +#include +#include +#include +#include +#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION +#include + +#include "DataProvider.h" + +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +namespace unittest { + +static std::unique_ptr> + OnPoolFilled; + +namespace pydp2 { + +void setOnPoolFilledHook(const std::function& callback) { + OnPoolFilled.reset(new std::function()); + *OnPoolFilled = callback; +} + +void clearOnPoolFilledHook() { OnPoolFilled.reset(); } + +} // namespace pydp2 +} // namespace unittest + +/** + * Slot type + */ +enum SlotType { + ST_DENSE = 0, + ST_NON_SPARSE_VALUE = 1, + ST_SPARSE_VALUE = 2, + ST_INDEX = 3 +}; + +/** + * Sequence type + */ +enum SeqType { SQT_NONE = 0, SQT_SEQ, SQT_SUBSEQ }; + +/** + * Cache Type. + */ +enum CacheType { + NO_CACHE = 0, // Each pass will load data from PyDataProvider2. + CACHE_PASS_IN_MEM = 1, // First pass will load data from PyDataProvider2, + // then cache all data in memory. Load data from + // memory in rest passes. +}; + +struct SlotHeader { // Slot Header will parse from python object's slots field. + size_t dim; + SlotType slotType; + SeqType seqType; +}; + +inline std::ostream& operator<<(std::ostream& os, const SlotHeader& header) { + os << "Dim = " << header.dim << " Type = " << header.slotType + << " SeqType = " << header.seqType; + return os; +} + +/** + * FieldScanner Interface. + * + * It will read python object, and fill to argument's each slot. + * There are two steps, prepare and fill. Scanner will alloc memory during + * prepare step, fill data into argument during fill step. + */ +class IFieldScanner { + public: + DISABLE_COPY(IFieldScanner); + /** + * Ctor. + * @param headerPtr slot header that scanner belong to. + */ + explicit IFieldScanner(SlotHeader* headerPtr) : headerPtr_(headerPtr) {} + virtual ~IFieldScanner() {} + + /** + * Start prepare step. + */ + virtual void startPrepare(Argument& argument) {} + + /** + * Prepare step. + * + * @note the obj could be a timestep of sample or whole sample. It depends + * what scanner it is. + */ + virtual void prepare(Argument& argument, PyObject* obj) {} + + /** + * Finish Prepare step. + */ + virtual void finishPrepare(Argument& argument) {} + + /** + * Start fill step. + */ + virtual void startFill(Argument& argument) {} + + /** + * Fill step. + * + * @note the obj could be a timestep of sample or whole sample. It depends + * what scanner it is. + */ + virtual void fill(Argument& argument, PyObject* obj) {} + + /** + * Finish fill step. + */ + virtual void finishFill(Argument& argument) {} + + /** + * Factory method. Create a scanner by header. The final scanner may be + * combine many scanners. + * + * @note Fatal if header is not support. + */ + static IFieldScanner* create(SlotHeader* header); + + protected: + SlotHeader* headerPtr_; +}; + +/** + * Py Data Provider Cache Interface. + */ +class IPyDataProviderCache { + public: + virtual ~IPyDataProviderCache() {} + + /** + * invoke when DataProvider::reset() + * @return true if read data from python. + */ + virtual bool reset() = 0; + + /** + * invoke when these data are used by DataProvider, and need to clear. + * @param [inout] data used data. + * + * @note The implemented class must clear these data array. Or if you want to + * delete the PyObjectPtr later, you should make sure the paddle process only + * have one active thread calling python code (use PyGuard otherwise). + */ + virtual void drop(std::deque* data) = 0; + + /** + * Return whole data in cache. + */ + virtual std::deque* load() = 0; + + /** + * Factory method. Convert CacheType to IPyDataProviderCache* + */ + static IPyDataProviderCache* create(CacheType ct); +}; + +/** + * PyDataProvider2. + * + * For usage, please refer python module 'paddle.trainer.PyDataProvider2' + * + * Here, we start a thread to read data. It is totally asynchronous for reading + * data. And it support cache strategies. + */ +class PyDataProvider2 : public DataProvider { + public: + /** + * Ctor + */ + PyDataProvider2(const DataConfig& config, + const ModelConfig& modelConfig, + bool useGpu) + : DataProvider(config, useGpu), callingContextCreated_(2) { + if (PyArray_API == NULL) import_array(); + auto& args = config.load_data_args(); + PyObjectPtr kwargs = PyObjectPtr(PyDict_New()); + if (!args.empty()) { + kwargs = callPythonFuncRetPyObj( + "paddle.trainer.PyDataProvider2", "deserialize_args", {args}); + } + + py::DictHelper kwargsDict(kwargs); + kwargsDict.setBool("is_train", !config.for_test()); + std::vector inputs; + inputs.reserve(modelConfig.input_layer_names().size()); + std::copy(modelConfig.input_layer_names().begin(), + modelConfig.input_layer_names().end(), + std::back_inserter(inputs)); + kwargsDict.setStringList("input_order", inputs); + + // kwargs is keyword arguemts to create object. + this->createPyDataObj(config.load_data_module(), + config.load_data_object(), + config.files(), + std::move(kwargs)); + DBG << "Instance " << instance_.get() << " loaded."; + this->readPyFields(config.for_test()); + DBG << "Py Field Done"; + } + + /** + * Dtor + * @note will stop loading thread when destructing + */ + virtual ~PyDataProvider2() { resetImpl(false); } + + private: + void createPyDataObj(const std::string& model, + const std::string& className, + const std::string& fileListName, + PyObjectPtr&& kwargs // NOLINT + ) { + LOG(INFO) << "loading dataprovider " << model << "::" << className; + + PyObjectPtr module = py::import(model); + PyObjectPtr moduleDict(PyModule_GetDict(module.get())); + CHECK_PY(moduleDict) << "Invoke module.__dict__ error"; + PyObjectPtr cls(PyDict_GetItemString(moduleDict.get(), className.c_str())); + CHECK_PY(cls) << "load class " << className.c_str() << "error"; + + // If there are multiple python instance share same module, the PyObjectPtr + // only for instance will make python reference-count error. + // + // So here, we increase reference count manually. + Py_XINCREF(module.get()); + Py_XINCREF(moduleDict.get()); + Py_XINCREF(cls.get()); + + PyObjectPtr fileListInPy = loadPyFileLists(fileListName); + PyDict_SetItemString(kwargs.get(), "file_list", fileListInPy.get()); + { + PyGuard guard; + instance_.reset(PyObject_Call(cls.get(), zeroTuple_.get(), kwargs.get())); + } + CHECK_PY(instance_) << "Cannot Create instance"; + } + + void readPyFields(bool testing) { + py::ObjectHelper self(this->instance_); + bool ok; + + this->skipShuffle_ = + !self.getBoolAttr("should_shuffle", &ok /*isBoolType*/); + if (!ok) { + this->skipShuffle_ = testing; // shuffle when is training, skip shuffle + // when is testing. + } + DBG << "Provider Skip Shuffle " << this->skipShuffle_; + + this->poolSize_ = self.getIntAttr("pool_size", &ok); + if (!ok) { + this->poolSize_ = -1UL; + } + this->minPoolSize_ = self.getIntAttr("min_pool_size", &ok); + if (!ok) { + this->minPoolSize_ = -1UL; + } + this->minPoolSize_ = std::min(this->poolSize_, this->minPoolSize_); + + this->canOverBatchSize_ = self.getBoolAttr("can_over_batch_size"); + + calcBatchSize_.reset(self.getAttr("calc_batch_size")); + if (this->calcBatchSize_ && !py::isCallable(this->calcBatchSize_)) { + this->calcBatchSize_.reset(); + } + + generator_.reset(self.getAttr("generator")); + CHECK(py::isCallable(generator_)); + + // Reading slots. + PyObjectPtr slotsPtr(self.getAttr("slots")); + py::SequenceHelper slots(slotsPtr); + headers_.reserve(slots.size()); + for (size_t i = 0; i < slots.size(); ++i) { + headers_.emplace_back(); + auto& header = headers_.back(); + PyObject* hdPtr = slots[i]; + CHECK(hdPtr != nullptr); + Py_XINCREF(hdPtr); + PyObjectPtr headerPtrWrap(hdPtr); + py::ObjectHelper hd(headerPtrWrap); + header.dim = hd.getIntAttrWithError("dim"); + header.seqType = (SeqType)hd.getIntAttrWithError("seq_type"); + header.slotType = (SlotType)hd.getIntAttrWithError("type"); + } + + DBG << "Data header size " << headers_.size(); + for (auto& header : headers_) { + DBG << header; + } + cache_.reset(IPyDataProviderCache::create( + (CacheType)self.getIntAttrWithError("cache"))); + } + + PyObjectPtr loadPyFileLists(const std::string& fileListName) { + loadFileList(fileListName, fileLists_); + PyObject* lst = PyList_New(fileLists_.size()); + for (size_t i = 0; i < fileLists_.size(); ++i) { + PyList_SET_ITEM(lst, i, PyString_FromString(fileLists_[i].c_str())); + } + return PyObjectPtr(lst); + } + + void loadThread() { + DBG << "Creating context"; + for (auto& filename : fileLists_) { + PyGuard g; + py::CallableHelper generator(this->generator_); + generator.setArgsSize(2); + generator.getArgs().set(0, instance_); + generator.getArgs().set(1, PyString_FromString(filename.c_str()), true); + callingContexts_.emplace_back(generator()); + CHECK_PY(callingContexts_.back()) << "Generator error."; + CHECK(PyIter_Check(callingContexts_.back())); + } + DBG << "Create context done"; + callingContextCreated_.wait(); + + PositionRandom p(skipShuffle_); + + while (!exit_ && !callingContexts_.empty()) { + PyObject* data = nullptr; + + { // Read data. + size_t cid = p(callingContexts_.size()); + bool atEnd; + data = py::iterNext(callingContexts_[cid], &atEnd); + if (atEnd || data == nullptr) { + if (cid != 0) { + std::swap(callingContexts_[cid], callingContexts_[0]); + cid = 0; + } + + PyObjectPtr front; + { + std::unique_lock l(mtx_); + front = pop_get_front(callingContexts_); + } + { + PyGuard g; + front.reset(); + } + this->pullCV_.notify_all(); + continue; + } + } + + size_t additionalBatchSize = 1; + if (calcBatchSize_) { + PyGuard guard; + py::CallableHelper calcBatchSize(this->calcBatchSize_); + calcBatchSize.setArgsSize(1); + calcBatchSize.getArgs().set(0, data); + PyObjectPtr bs(calcBatchSize()); + CHECK_PY(bs); + bool ok; + additionalBatchSize = py::castInt(bs.get(), &ok); + CHECK(ok) << "CalcBatchSize must return int or long"; + } + + if (this->loadThread_) { // wait poolActualSize < poolSize; + std::unique_lock l(mtx_); + pushCV_.wait(l, [this] { return this->poolActualSize_ < poolSize_; }); + } + + { + std::lock_guard guard(mtx_); + poolActualSize_ += additionalBatchSize; + dataPool_.emplace_back(data); + } + pullCV_.notify_all(); + } + DBG << "load thread end"; + } + + inline void resetImpl(bool startNewThread) { + DBG << "Reseting " << startNewThread; + exit_.store(true); + if (loadThread_) { // is loading. + loadThread_->join(); + loadThread_.reset(); + } + { + PyGuard g; + callingContexts_.clear(); + this->pullCV_.notify_one(); + } + + std::lock_guard guard(mutexForReset_); + { + PyGuard g; + dataPool_.clear(); + } + poolActualSize_ = 0; + + if (startNewThread && cache_->reset()) { + DBG << "Start new thread."; + loadThread_.reset(new std::thread([this] { + exit_ = false; + loadThread(); + })); + callingContextCreated_.wait(); + } + DBG << "Reset done"; + exit_ = false; + } + + private: + std::unique_ptr loadThread_; + std::atomic exit_; + std::deque callingContexts_; + std::deque dataPool_; + size_t poolActualSize_; + std::condition_variable pushCV_; + std::condition_variable pullCV_; + std::mutex mtx_; + + std::mutex mutexForReset_; + + ThreadBarrier callingContextCreated_; + std::unique_ptr cache_; + + PyObjectPtr instance_; + size_t poolSize_; + size_t minPoolSize_; + bool canOverBatchSize_; + PyObjectPtr calcBatchSize_; + PyObjectPtr generator_; + std::vector fileLists_; + std::vector headers_; + static PyObjectPtr zeroTuple_; + + class PositionRandom { + public: + inline explicit PositionRandom(bool skipRand) + : eng_(ThreadLocalRandomEngine::get()), skipRand_(skipRand) {} + + inline size_t operator()(size_t len) { + if (!skipRand_) { + if (!dist_ || dist_->b() != len - 1) { + dist_.reset(new std::uniform_int_distribution(0, len - 1)); + } + return (*dist_)(eng_); + } else { + return 0; + } + } + + private: + std::default_random_engine& eng_; + std::unique_ptr> dist_; + bool skipRand_; + }; + + // DataProvider interface + public: + /** + * Resetting the PyDataProvider. May start reading thread here. + */ + virtual void reset() { + resetImpl(true); + DataProvider::reset(); + } + + /** + * Shuffle. Do nothing because PyDataProvider do shuffle implicitly by random + * select data from datapool. + */ + void shuffle() {} + + /** + * Not limited size. + */ + int64_t getSize() { return -1; } + + /** + * Loading a batch of data. + */ + int64_t getNextBatchInternal(int64_t size_, DataBatch* batch) { + std::lock_guard guard(mutexForReset_); + REGISTER_TIMER("PyDP2.getNextBatchInternal") + CHECK_GE(size_, 0); + size_t size = (size_t)size_; + if (loadThread_) { // loading from thread should wait for data pool ready. + // but, loading from cache, cache object should ensure + // data pool ready. + std::unique_lock l(mtx_); + pullCV_.wait(l, [this, &size] { + return this->poolActualSize_ >= std::max(size, this->minPoolSize_) || + callingContexts_.empty(); + }); + + if (unittest::OnPoolFilled) { + (*unittest::OnPoolFilled)(this->poolActualSize_); + } + } + std::deque data; + size_t bsize = 0; + std::deque* poolPtr = nullptr; + + if (this->loadThread_) { // loading from thread. + poolPtr = &this->dataPool_; + } else { // loading from cache. + poolPtr = this->cache_->load(); + } + if (exit_) { + // PyDataProvider is destructing. + return 0; + } + CHECK(poolPtr != nullptr); + + std::deque& pool = *poolPtr; + + while (bsize < size && !pool.empty()) { + { + // move data from pool to data + std::lock_guard guard(mtx_); + if (skipShuffle_) { + size_t i = 0; + CHECK(pool[i] != nullptr); + data.emplace_back(std::move(pool[i])); + pool.pop_front(); + } else { // when shuffle, use swap to drop only last pool element. + size_t i = ThreadLocalRand::rand() % pool.size(); + CHECK(pool[i] != nullptr); + if (i != 0) { + std::swap(pool[i], pool.front()); + } + data.emplace_back(std::move(pool.front())); + pool.pop_front(); + } + + if (calcBatchSize_) { // custom calc batch size. + PyGuard guard; + Py_INCREF(data.back().get()); + py::CallableHelper calcBatchSize(calcBatchSize_); + calcBatchSize.setArgsSize(1); + calcBatchSize.getArgs().set(0, data.back()); + PyObjectPtr customBatchSize(calcBatchSize()); + bool ok; + size_t tmp = py::castInt(customBatchSize.get(), &ok); + CHECK(ok) << "calc_batch_size must return int"; + + if (bsize + tmp > size && !canOverBatchSize_) { + // Put data back. + pool.push_front(std::move(data.back())); + data.pop_back(); + break; + } else { + bsize += tmp; + } + } else { + bsize += 1; + } + } + } + + if (this->loadThread_) { + { + std::lock_guard g(mtx_); + poolActualSize_ -= bsize; + } + this->pushCV_.notify_all(); + } + + if (bsize == 0) { // end of pass. In data pool, cannot get any data. + return 0; + } + + DataBatch cpuBatch; + cpuBatch.setSize(bsize); + auto& inArgs = cpuBatch.getStreams(); + inArgs.resize(headers_.size()); + std::vector> scanners; + scanners.reserve(headers_.size()); + for (auto& header : headers_) { + scanners.emplace_back(IFieldScanner::create(&header)); + } + DBG << "Scanner created."; + for (size_t i = 0; i < headers_.size(); ++i) { + scanners[i]->startPrepare(inArgs[i]); + } + for (auto& d : data) { + py::SequenceHelper s(d); + for (size_t i = 0; i < headers_.size(); ++i) { + scanners[i]->prepare(inArgs[i], s[i]); + } + } + for (size_t i = 0; i < headers_.size(); ++i) { + scanners[i]->finishPrepare(inArgs[i]); + } + for (size_t i = 0; i < headers_.size(); ++i) { + scanners[i]->startFill(inArgs[i]); + } + for (auto& d : data) { + py::SequenceHelper s(d); + for (size_t i = 0; i < headers_.size(); ++i) { + scanners[i]->fill(inArgs[i], s[i]); + } + } + + for (size_t i = 0; i < headers_.size(); ++i) { + scanners[i]->finishFill(inArgs[i]); + } + + { + PyGuard g; + cache_->drop(&data); + } + + DBG << "Reading CPU Batch Done."; + + if (useGpu_) { + std::vector& cpuArguments = cpuBatch.getStreams(); + DataBatch& gpuBatch = *batch; + std::vector& gpuArguments = gpuBatch.getStreams(); + gpuArguments.resize(cpuArguments.size()); + gpuBatch.setSize(bsize); + for (size_t i = 0; i < headers_.size(); ++i) { + gpuArguments[i].resizeAndCopyFrom( + cpuArguments[i], useGpu_, HPPL_STREAM_1); + } + hl_stream_synchronize(HPPL_STREAM_1); + } else { + *batch = cpuBatch; + } + return bsize; + } +}; + +PyObjectPtr PyDataProvider2::zeroTuple_(PyTuple_New(0)); + +REGISTER_DATA_PROVIDER_EX(py2, PyDataProvider2); + +/** + * Scanner for dense slot. + */ +class DenseScanner : public IFieldScanner { + public: + explicit DenseScanner(SlotHeader* ptr) : IFieldScanner(ptr), height_(0) {} + + /** + * Prepare. + * @param argument target argument + * @param obj each timestep of a sample. + */ + virtual void prepare(Argument& argument, PyObject* obj) { ++height_; } + + virtual void finishPrepare(Argument& argument) { + Matrix::resizeOrCreate( + argument.value, height_, headerPtr_->dim, false, false); + height_ = 0; + } + + /** + * Fill argument from obj. + * @param argument + * @param obj + */ + virtual void fill(Argument& argument, PyObject* obj) { + real* dat = argument.value->getData() + height_ * headerPtr_->dim; + if (PyArray_Check(obj)) { + auto dtype = PyArray_DTYPE((PyArrayObject*)obj); + if (dtype->type == 'f' && dtype->elsize == sizeof(real)) { + real* data = (real*)PyArray_DATA((PyArrayObject*)obj); + auto sz = PyArray_SIZE((PyArrayObject*)obj); + std::copy(data, data + sz, dat); + } else { + LOG(FATAL) << "You should yield float" << sizeof(real) * 8 << " array"; + } + } else { + py::SequenceHelper s(obj); + // TODO(yuyang18): Here we can use AVX or SSE to accelerate memory copy. + for (size_t i = 0; i < headerPtr_->dim; ++i) { + dat[i] = (real)s.getDouble(i); + } + } + ++height_; + } + + private: + size_t height_; +}; + +/** + * Scanner for index slot + */ +class IndexScanner : public IFieldScanner { + public: + explicit IndexScanner(SlotHeader* ptr) : IFieldScanner(ptr), cnt_(0) {} + + /** + * Prepare memory space. + * + * @note obj is a single timestep of sample + */ + virtual void prepare(Argument& argument, PyObject* obj) { ++cnt_; } + + virtual void finishPrepare(Argument& argument) { + IVector::resizeOrCreate(argument.ids, cnt_, false); + cnt_ = 0; + } + + /** + * Fill one index to argument. + */ + virtual void fill(Argument& argument, PyObject* obj) { + bool ok; + argument.ids->getData()[cnt_++] = py::castInt(obj, &ok); + CHECK(ok) << "Cannot cast int " << py::repr(obj); + } + + private: + size_t cnt_; +}; + +class SparseNonValueScanner : public IFieldScanner { + public: + explicit SparseNonValueScanner(SlotHeader* ptr) + : IFieldScanner(ptr), nnz_(0), height_(0) {} + + /** + * Prepare memory space + * @note obj is a timestep of one sample. + */ + virtual void prepare(Argument& argument, PyObject* obj) { + ++height_; + nnz_ += py::SequenceHelper(obj).size(); + } + + virtual void finishPrepare(Argument& argument) { + Matrix::resizeOrCreateSparseMatrix( + argument.value, height_, headerPtr_->dim, nnz_, NO_VALUE); + } + + virtual void startFill(Argument& argument) { + auto smat = (CpuSparseMatrix*)(argument.value.get()); + smat->getRows()[0] = 0; + nnz_ = 0; + height_ = 1; + } + + /** + * Fill one sparse vector to argument. + * @note obj is a timestep of one sample. + */ + virtual void fill(Argument& argument, PyObject* obj) { + py::SequenceHelper s(obj); + auto sz = s.size(); + auto smat = (CpuSparseMatrix*)(argument.value.get()); + int* row = smat->getRows(); + int* col = smat->getCols(); + real* dat = smat->getData(); + row[height_] = row[height_ - 1] + (int)sz; + + for (decltype(sz) i = 0; i < sz; ++i) { + setData(col + nnz_, dat + nnz_, s[i]); + ++nnz_; + } + ++height_; + } + + protected: + /** + * Set a single sparse index and value. + * @param [out] col sparse index + * @param [out] dat sparse value + * @param [in] obj Python Object. For sparse_non_value is a PyInt or PyLong. + * For sparse_value is a Tuple (int, float). + */ + virtual void setData(int* col, real* dat, PyObject* obj) { + bool ok; + *col = py::castInt(obj, &ok); + CHECK(ok); + } + + size_t nnz_; + size_t height_; +}; + +class SparseValueScanner : public SparseNonValueScanner { + public: + explicit SparseValueScanner(SlotHeader* ptr) : SparseNonValueScanner(ptr) {} + + virtual void finishPrepare(Argument& argument) { + Matrix::resizeOrCreateSparseMatrix( + argument.value, height_, headerPtr_->dim, nnz_, FLOAT_VALUE); + } + + protected: + virtual void setData(int* col, real* dat, PyObject* obj) { + py::SequenceHelper s(obj); + SparseNonValueScanner::setData(col, dat, s[0]); + *dat = (real)s.getDouble(1); + } +}; + +/** + * Sequence Scanner. Scanner for sequence or sub-sequence. + */ +class SequenceScanner : public IFieldScanner { + public: + /** + * Ctor + * @param innerScanner inner scanner for each timestep or sub-sequence. + * @param getSeqStartPos A callback, (Argument) => ICpuGpuVectorPtr. + * return a sequence start position or a sub-sequence + * start position. + */ + SequenceScanner( + std::unique_ptr&& innerScanner, + const std::function& getSeqStartPos) + : IFieldScanner(nullptr), + inner_(std::move(innerScanner)), + cnt_(0), + getSeqStartPos_(getSeqStartPos) {} + + /** + * Start prepare. Invoke inner->startPrepare too. + */ + virtual void startPrepare(Argument& argument) { + inner_->startPrepare(argument); + } + + /** + * Prepare. obj is a list or tuple. it will invoke inner_->prepare for each + * element of sequence obj. + */ + virtual void prepare(Argument& argument, PyObject* obj) { + py::SequenceHelper s(obj); + ++cnt_; + for (size_t i = 0; i < s.size(); ++i) { + inner_->prepare(argument, s[i]); + } + } + + /** + * Finish prepare. invoke inner_->finishPrepare too. + */ + virtual void finishPrepare(Argument& argument) { + ICpuGpuVector::resizeOrCreate(getSeqStartPos_(argument), cnt_ + 1, false); + inner_->finishPrepare(argument); + } + + /** + * Start fill. invoke inner->startFill too. + */ + virtual void startFill(Argument& argument) { + getSeqStartPos_(argument)->getMutableData(false)[0] = 0; + cnt_ = 1; + inner_->startFill(argument); + } + + /** + * Fill. Obj is a tuple or list. invoke inner->fill for each element of + * sequence obj. And set seqStartPos at same time. The seqStartPos will be + * calculated by getSeqStartPos callback passed in ctor. + */ + virtual void fill(Argument& argument, PyObject* obj) { + getSeqStartPos_(argument)->getMutableData(false)[cnt_] = + getSeqStartPos_(argument)->getMutableData(false)[cnt_ - 1] + + (int)getSize(obj); + py::SequenceHelper s(obj); + ++cnt_; + for (size_t i = 0; i < s.size(); ++i) { + inner_->fill(argument, s[i]); + } + } + + /** + * Finish fill. will invoke inner->finishFill too. + */ + virtual void finishFill(Argument& argument) { inner_->finishFill(argument); } + + protected: + size_t getSize(PyObject* obj) { + py::SequenceHelper s(obj); + auto sc = dynamic_cast(inner_.get()); + if (sc) { + size_t sum = 0; + for (size_t i = 0; i < s.size(); ++i) { + sum += sc->getSize(s[i]); + } + return sum; + } else { + return s.size(); + } + } + + private: + std::unique_ptr inner_; + size_t cnt_; + std::function getSeqStartPos_; +}; + +IFieldScanner* IFieldScanner::create(SlotHeader* header) { + IFieldScanner* retv = nullptr; + switch (header->slotType) { + case ST_DENSE: + retv = new DenseScanner(header); + break; + case ST_INDEX: + retv = new IndexScanner(header); + break; + case ST_NON_SPARSE_VALUE: + retv = new SparseNonValueScanner(header); + break; + case ST_SPARSE_VALUE: + retv = new SparseValueScanner(header); + break; + default: + LOG(FATAL) << "Not implemented " << header->slotType; + } + + switch (header->seqType) { + case SQT_NONE: + break; + case SQT_SUBSEQ: + retv = new SequenceScanner(std::unique_ptr(retv), + [](Argument& arg) -> ICpuGpuVectorPtr& { + return arg.subSequenceStartPositions; + }); + // fall through, not break; + case SQT_SEQ: + retv = new SequenceScanner(std::unique_ptr(retv), + [](Argument& arg) -> ICpuGpuVectorPtr& { + return arg.sequenceStartPositions; + }); + break; + default: + LOG(FATAL) << "Not implemented"; + } + + return retv; +} + +/** + * No Cache Strategy. Will destruct old data immediately and load data from + * python every pass. + */ +class NoCacheStrategy : public IPyDataProviderCache { + public: + virtual bool reset() { return true; } + + virtual void drop(std::deque* data) { data->clear(); } + + virtual std::deque* load() { return nullptr; } +}; + +/** + * Cache One Pass In Memory strategy. + * + * In first pass, will load data from python and store them in memory. + * The rest passes, will load data from memory. + */ +class CacheOnePassInMemory : public IPyDataProviderCache { + public: + CacheOnePassInMemory() + : objPool_(new std::deque()), + droppedPool_(new std::deque()) {} + + virtual bool reset() { + if (objPool_->empty() && droppedPool_->empty()) { + return true; + } else if (objPool_->empty()) { + std::swap(objPool_, droppedPool_); + return false; + } else { + LOG(FATAL) << "Unexpected branch"; + } + } + + virtual void drop(std::deque* data) { + size_t orgSize = droppedPool_->size(); + droppedPool_->resize(orgSize + data->size()); + for (size_t i = 0; i < data->size(); ++i) { + std::swap((*droppedPool_)[orgSize + i], (*data)[i]); + } + data->clear(); + } + + virtual std::deque* load() { return objPool_.get(); } + + private: + std::unique_ptr> objPool_; + std::unique_ptr> droppedPool_; +}; + +IPyDataProviderCache* IPyDataProviderCache::create(CacheType ct) { + switch (ct) { + case NO_CACHE: + return new NoCacheStrategy(); + case CACHE_PASS_IN_MEM: + return new CacheOnePassInMemory(); + default: + LOG(FATAL) << "Not implemented"; + } +} +} // namespace paddle + +#endif diff --git a/paddle/legacy/gserver/evaluators/CTCErrorEvaluator.cpp b/paddle/legacy/gserver/evaluators/CTCErrorEvaluator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c145adda5e04fb4a35df480fd3d0cf93ad453e0d --- /dev/null +++ b/paddle/legacy/gserver/evaluators/CTCErrorEvaluator.cpp @@ -0,0 +1,320 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Evaluator.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/utils/StringUtil.h" + +namespace paddle { + +/** + * calculate sequence-to-sequence edit distance + */ +class CTCErrorEvaluator : public Evaluator { + private: + MatrixPtr outActivations_; + int numTimes_, numClasses_, numSequences_, blank_; + real deletions_, insertions_, substitutions_; + int seqClassficationError_; + mutable std::unordered_map evalResults_; + + std::vector path2String(const std::vector& path) { + std::vector str; + str.clear(); + int prevLabel = -1; + for (std::vector::const_iterator label = path.begin(); + label != path.end(); + label++) { + if (*label != blank_ && + (str.empty() || *label != str.back() || prevLabel == blank_)) { + str.push_back(*label); + } + prevLabel = *label; + } + return str; + } + + std::vector bestLabelSeq() { + std::vector path; + path.clear(); + real* acts = outActivations_->getData(); + for (int i = 0; i < numTimes_; ++i) { + path.push_back(std::max_element(acts + i * numClasses_, + acts + (i + 1) * numClasses_) - + (acts + i * numClasses_)); + } + return path2String(path); + } + + /* "sp, dp, ip" is the weighting parameter of "substitution, deletion, + * insertion" + * in edit-distance error */ + real stringAlignment(std::vector& gtStr, + std::vector& recogStr, + bool backtrace = true, + real sp = 1.0, + real dp = 1.0, + real ip = 1.0) { + std::vector> matrix; + int substitutions, deletions, insertions; + real distance; + int n = gtStr.size(); + int m = recogStr.size(); + + if (n == 0) { + substitutions = 0; + deletions = 0; + insertions = m; + distance = m; + } else if (m == 0) { + substitutions = 0; + deletions = n; + insertions = 0; + distance = n; + } else { + substitutions = 0; + deletions = 0; + insertions = 0; + distance = 0; + // initialize the matrix + matrix.resize(n + 1); + for (int i = 0; i < n + 1; ++i) { + matrix[i].resize(m + 1); + for (int j = 0; j < m + 1; ++j) { + matrix[i][j] = 0; + } + } + for (int i = 0; i < n + 1; ++i) { + matrix[i][0] = i; + } + for (int j = 0; j < m + 1; ++j) { + matrix[0][j] = j; + } + + // calculate the insertions, substitutions and deletions + for (int i = 1; i < n + 1; ++i) { + int s_i = gtStr[i - 1]; + for (int j = 1; j < m + 1; ++j) { + int t_j = recogStr[j - 1]; + int cost = (s_i == t_j) ? 0 : 1; + const int above = matrix[i - 1][j]; + const int left = matrix[i][j - 1]; + const int diag = matrix[i - 1][j - 1]; + const int cell = std::min(above + 1, std::min(left + 1, diag + cost)); + matrix[i][j] = cell; + } + } + + if (backtrace) { + size_t i = n; + size_t j = m; + substitutions = 0; + deletions = 0; + insertions = 0; + + while (i != 0 && j != 0) { + if (matrix[i][j] == matrix[i - 1][j - 1]) { + --i; + --j; + } else if (matrix[i][j] == matrix[i - 1][j - 1] + 1) { + ++substitutions; + --i; + --j; + } else if (matrix[i][j] == matrix[i - 1][j] + 1) { + ++deletions; + --i; + } else { + ++insertions; + --j; + } + } + while (i != 0) { + ++deletions; + --i; + } + while (j != 0) { + ++insertions; + --j; + } + int diff = substitutions + deletions + insertions; + if (diff != matrix[n][m]) { + LOG(ERROR) << "Found path with distance " << diff + << " but Levenshtein distance is " << matrix[n][m]; + } + + distance = (sp * substitutions) + (dp * deletions) + (ip * insertions); + } else { + distance = (real)matrix[n][m]; + } + } + real maxLen = std::max(m, n); + deletions_ += deletions / maxLen; + insertions_ += insertions / maxLen; + substitutions_ += substitutions / maxLen; + + if (distance != 0) { + seqClassficationError_ += 1; + } + + return distance / maxLen; + } + + real editDistance( + real* output, int numTimes, int numClasses, int* labels, int labelsLen) { + numTimes_ = numTimes; + numClasses_ = numClasses; + blank_ = numClasses_ - 1; + outActivations_ = Matrix::create(output, numTimes, numClasses); + std::vector recogStr, gtStr; + recogStr = bestLabelSeq(); + for (int i = 0; i < labelsLen; ++i) { + gtStr.push_back(labels[i]); + } + + return stringAlignment(gtStr, recogStr); + } + + void storeLocalValues() const { + evalResults_["error"] = numSequences_ ? totalScore_ / numSequences_ : 0; + evalResults_["deletion_error"] = + numSequences_ ? deletions_ / numSequences_ : 0; + evalResults_["insertion_error"] = + numSequences_ ? insertions_ / numSequences_ : 0; + evalResults_["substitution_error"] = + numSequences_ ? substitutions_ / numSequences_ : 0; + evalResults_["sequence_error"] = + (real)seqClassficationError_ / numSequences_; + } + + public: + CTCErrorEvaluator() + : numTimes_(0), + numClasses_(0), + numSequences_(0), + blank_(0), + deletions_(0), + insertions_(0), + substitutions_(0), + seqClassficationError_(0) {} + + virtual real evalImp(std::vector& arguments) { + CHECK_EQ(arguments.size(), (size_t)2); + Argument output, label; + output.resizeAndCopyFrom(arguments[0], false, HPPL_STREAM_DEFAULT); + label.resizeAndCopyFrom(arguments[1], false, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + CHECK(label.sequenceStartPositions); + CHECK(label.ids); + size_t numSequences = label.sequenceStartPositions->getSize() - 1; + const int* labelStarts = label.sequenceStartPositions->getData(false); + const int* outputStarts = output.sequenceStartPositions->getData(false); + real totalErr = 0; + for (size_t i = 0; i < numSequences; ++i) { + real err = 0; + err = editDistance( + output.value->getData() + output.value->getWidth() * outputStarts[i], + outputStarts[i + 1] - outputStarts[i], + output.value->getWidth(), + label.ids->getData() + labelStarts[i], + labelStarts[i + 1] - labelStarts[i]); + + totalErr += err; + } + + return totalErr; + } + + virtual void eval(const NeuralNetwork& nn) { + Evaluator::eval(nn); + std::vector arguments; + arguments.reserve(config_.input_layers_size()); + for (const std::string& name : config_.input_layers()) { + arguments.push_back(nn.getLayer(name)->getOutput()); + } + } + + virtual void updateSamplesNum(const std::vector& arguments) { + numSequences_ += arguments[1].getNumSequences(); + } + + virtual void start() { + Evaluator::start(); + numSequences_ = 0; + blank_ = 0; + deletions_ = 0; + insertions_ = 0; + substitutions_ = 0; + seqClassficationError_ = 0; + } + + virtual void printStats(std::ostream& os) const { + storeLocalValues(); + os << config_.name() << " error = " << evalResults_["error"]; + os << " deletions error = " << evalResults_["deletion_error"]; + os << " insertions error = " << evalResults_["insertion_error"]; + os << " substitution error = " << evalResults_["substitution_error"]; + os << " sequence error = " << evalResults_["sequence_error"]; + } + + virtual void distributeEval(ParameterClient2* client) { + double buf[6] = {totalScore_, + (double)deletions_, + (double)insertions_, + (double)substitutions_, + (double)seqClassficationError_, + (double)numSequences_}; + client->reduce(buf, buf, 6, FLAGS_trainer_id, 0); + totalScore_ = buf[0]; + deletions_ = (real)buf[1]; + insertions_ = (real)buf[2]; + substitutions_ = (real)buf[3]; + seqClassficationError_ = (int)buf[4]; + numSequences_ = (int)buf[5]; + } + + void getNames(std::vector* names) { + storeLocalValues(); + names->reserve(names->size() + evalResults_.size()); + for (auto it = evalResults_.begin(); it != evalResults_.end(); ++it) { + names->push_back(config_.name() + "." + it->first); + } + } + + real getValue(const std::string& name, Error* err) const { + storeLocalValues(); + + std::vector buffers; + paddle::str::split(name, '.', &buffers); + auto it = evalResults_.find(buffers[buffers.size() - 1]); + + if (it == evalResults_.end()) { + *err = Error("Evaluator does not have the key %s", name.c_str()); + return 0.0f; + } + + return it->second; + } + + std::string getType(const std::string& name, Error* err) const { + this->getValue(name, err); + if (!err->isOK()) { + return ""; + } + return "ctc_edit_distance"; + } +}; + +REGISTER_EVALUATOR(ctc_edit_distance, CTCErrorEvaluator); + +} // namespace paddle diff --git a/paddle/legacy/gserver/evaluators/ChunkEvaluator.cpp b/paddle/legacy/gserver/evaluators/ChunkEvaluator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0ff3f2fa8cf06c13ef327aa7ae2511bfc0d028be --- /dev/null +++ b/paddle/legacy/gserver/evaluators/ChunkEvaluator.cpp @@ -0,0 +1,296 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/StringUtil.h" + +#include "Evaluator.h" + +namespace paddle { + +/** + * Chunk evaluator is used to evaluate segment labelling accuracy for a + * sequence. It calculates the chunk detection F1 score. + * + * A chunk is correctly detected if its beginning, end and type are correct. + * Other chunk type is ignored. + * For each label in the label sequence, we have + * + * @code + * tagType = label % numTagType + * chunkType = label / numTagType + * otherChunkType = numChunkTypes + * @endcode + * + * The total number of different labels is numTagType*numChunkTypes+1 + * We support 4 labelling scheme + * The tag type for each of the scheme is shown as follows: + * + * @code + * Scheme Begin Inside End Single + * plain 0 - - - + * IOB 0 1 - - + * IOE - 0 1 - + * IOBES 0 1 2 3 + * @endcode + * + * 'plain' means the whole chunk must contain exactly the same chunk label. + */ +class ChunkEvaluator : public Evaluator { + int otherChunkType_; + int numChunkTypes_; // number of chunk types besides other chunk type + int numTagTypes_; + int tagBegin_; + int tagInside_; + int tagEnd_; + int tagSingle_; + + int64_t numLabelSegments_; + int64_t numOutputSegments_; + int64_t numCorrect_; + + struct Segment { + int begin; + int end; + int type; + bool operator==(const Segment& y) const { + return begin == y.begin && end == y.end && type == y.type; + } + }; + + std::vector labelSegments_; + std::vector outputSegments_; + std::set excludedChunkTypes_; + mutable std::unordered_map values_; + + public: + virtual void init(const EvaluatorConfig& config) { + Evaluator::init(config); + if (config.chunk_scheme() == "IOB") { + numTagTypes_ = 2; + tagBegin_ = 0; + tagInside_ = 1; + tagEnd_ = -1; + tagSingle_ = -1; + } else if (config.chunk_scheme() == "IOE") { + numTagTypes_ = 2; + tagBegin_ = -1; + tagInside_ = 0; + tagEnd_ = 1; + tagSingle_ = -1; + } else if (config.chunk_scheme() == "IOBES") { + numTagTypes_ = 4; + tagBegin_ = 0; + tagInside_ = 1; + tagEnd_ = 2; + tagSingle_ = 3; + } else if (config.chunk_scheme() == "plain") { + numTagTypes_ = 1; + tagBegin_ = -1; + tagInside_ = -1; + tagEnd_ = -1; + tagSingle_ = -1; + } else { + LOG(FATAL) << "Unknown chunk scheme: " << config.chunk_scheme(); + } + CHECK(config.has_num_chunk_types()) << "Missing num_chunk_types in config"; + otherChunkType_ = numChunkTypes_ = config.num_chunk_types(); + + // the chunks of types in excludedChunkTypes_ will not be counted + auto& tmp = config.excluded_chunk_types(); + excludedChunkTypes_.insert(tmp.begin(), tmp.end()); + } + + virtual void start() { + Evaluator::start(); + numLabelSegments_ = 0; + numOutputSegments_ = 0; + numCorrect_ = 0; + } + + virtual void printStats(std::ostream& os) const { + storeLocalValues(); + os << config_.name() << "=" << values_["F1-score"] + << " true_chunks=" << numLabelSegments_ + << " result_chunks=" << numOutputSegments_ + << " correct_chunks=" << numCorrect_; + } + + virtual void distributeEval(ParameterClient2* client) { + int64_t buf[3] = {numLabelSegments_, numOutputSegments_, numCorrect_}; + client->reduce(buf, buf, 3, FLAGS_trainer_id, 0); + numLabelSegments_ = buf[0]; + numOutputSegments_ = buf[1]; + numCorrect_ = buf[2]; + } + + virtual real evalImp(std::vector& arguments) { + CHECK_EQ(arguments.size(), (size_t)2); + IVectorPtr& output = arguments[0].ids; + IVectorPtr& label = arguments[1].ids; + CHECK(!output->useGpu() && !label->useGpu()) << "Not supported"; + auto sequenceStartPositions = + arguments[1].sequenceStartPositions->getVector(false); + CHECK_EQ(output->getSize(), label->getSize()); + CHECK(sequenceStartPositions); + size_t numSequences = sequenceStartPositions->getSize() - 1; + const int* starts = sequenceStartPositions->getData(); + for (size_t i = 0; i < numSequences; ++i) { + eval1(output->getData() + starts[i], + label->getData() + starts[i], + starts[i + 1] - starts[i]); + } + return 0; + } + + void eval1(int* output, int* label, int length) { + getSegments(output, length, outputSegments_); + getSegments(label, length, labelSegments_); + size_t i = 0, j = 0; + while (i < outputSegments_.size() && j < labelSegments_.size()) { + if (outputSegments_[i] == labelSegments_[j] && + excludedChunkTypes_.count(outputSegments_[i].type) != 1) { + ++numCorrect_; + } + if (outputSegments_[i].end < labelSegments_[j].end) { + ++i; + } else if (outputSegments_[i].end > labelSegments_[j].end) { + ++j; + } else { + ++i; + ++j; + } + } + for (auto& segment : labelSegments_) { + if (excludedChunkTypes_.count(segment.type) != 1) ++numLabelSegments_; + } + for (auto& segment : outputSegments_) { + if (excludedChunkTypes_.count(segment.type) != 1) ++numOutputSegments_; + } + } + + void getSegments(int* label, int length, std::vector& segments) { + segments.clear(); + segments.reserve(length); + int chunkStart = 0; + bool inChunk = false; + int tag = -1; + int type = otherChunkType_; + for (int i = 0; i < length; ++i) { + int prevTag = tag; + int prevType = type; + CHECK_LE(label[i], numChunkTypes_ * numTagTypes_); + tag = label[i] % numTagTypes_; + type = label[i] / numTagTypes_; + if (inChunk && isChunkEnd(prevTag, prevType, tag, type)) { + Segment segment{ + chunkStart, // begin + i - 1, // end + prevType, + }; + segments.push_back(segment); + inChunk = false; + } + if (isChunkBegin(prevTag, prevType, tag, type)) { + chunkStart = i; + inChunk = true; + } + } + if (inChunk) { + Segment segment{ + chunkStart, // begin + length - 1, // end + type, + }; + segments.push_back(segment); + } + } + + // whether (prevTag, prevType) is the end of a chunk + bool isChunkEnd(int prevTag, int prevType, int tag, int type) { + if (prevType == otherChunkType_) return false; + if (type == otherChunkType_) return true; + if (type != prevType) return true; + if (prevTag == tagBegin_) return tag == tagBegin_ || tag == tagSingle_; + if (prevTag == tagInside_) return tag == tagBegin_ || tag == tagSingle_; + if (prevTag == tagEnd_) return true; + if (prevTag == tagSingle_) return true; + return false; + } + + // whether (tag, type) is the beginning of a chunk + bool isChunkBegin(int prevTag, int prevType, int tag, int type) { + if (prevType == otherChunkType_) return type != otherChunkType_; + if (type == otherChunkType_) return false; + if (type != prevType) return true; + if (tag == tagBegin_) return true; + if (tag == tagInside_) return prevTag == tagEnd_ || prevTag == tagSingle_; + if (tag == tagEnd_) return prevTag == tagEnd_ || prevTag == tagSingle_; + if (tag == tagSingle_) return true; + return false; + } + + // three metrics: precision, recall and F1-score + void getNames(std::vector* names) { + storeLocalValues(); + names->reserve(names->size() + values_.size()); + for (auto it = values_.begin(); it != values_.end(); ++it) { + names->push_back(config_.name() + "." + it->first); + } + } + + // get value by field name + real getValue(const std::string& name, Error* err) const { + storeLocalValues(); + std::vector buffers; + paddle::str::split(name, '.', &buffers); + auto it = values_.find(buffers.back()); + if (it == values_.end()) { // not found + *err = Error("No such key %s", name.c_str()); + return 0.0f; + } + + return it->second; + } + + // get type of evaluator + std::string getType(const std::string& name, Error* err) const { + this->getValue(name, err); + if (!err->isOK()) { + return ""; + } + return "chunk"; + } + + private: + void storeLocalValues() const { + CHECK_GE(numOutputSegments_, 0); + CHECK_GE(numLabelSegments_, 0); + double precision = + !numOutputSegments_ ? 0 : (double)numCorrect_ / numOutputSegments_; + double recall = + !numLabelSegments_ ? 0 : (double)numCorrect_ / numLabelSegments_; + values_["precision"] = precision; + values_["recall"] = recall; + values_["F1-score"] = + !numCorrect_ ? 0 : 2 * precision * recall / (precision + recall); + } +}; + +REGISTER_EVALUATOR(chunk, ChunkEvaluator); + +} // namespace paddle diff --git a/paddle/legacy/gserver/evaluators/DetectionMAPEvaluator.cpp b/paddle/legacy/gserver/evaluators/DetectionMAPEvaluator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..57657241f8c1517f674670d34cb984b85996bfc7 --- /dev/null +++ b/paddle/legacy/gserver/evaluators/DetectionMAPEvaluator.cpp @@ -0,0 +1,308 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Evaluator.h" +#include "paddle/legacy/gserver/layers/DetectionUtil.h" + +using std::map; +using std::vector; +using std::pair; +using std::make_pair; + +namespace paddle { + +/** + * @brief detection map Evaluator + * + * The config file api is detection_map_evaluator. + */ +class DetectionMAPEvaluator : public Evaluator { + public: + DetectionMAPEvaluator() + : evaluateDifficult_(false), cpuOutput_(nullptr), cpuLabel_(nullptr) {} + + virtual void start() { + Evaluator::start(); + allTruePos_.clear(); + allFalsePos_.clear(); + numPos_.clear(); + } + + virtual real evalImp(std::vector& arguments) { + overlapThreshold_ = config_.overlap_threshold(); + backgroundId_ = config_.background_id(); + evaluateDifficult_ = config_.evaluate_difficult(); + apType_ = config_.ap_type(); + + MatrixPtr detectTmpValue = arguments[0].value; + Matrix::resizeOrCreate(cpuOutput_, + detectTmpValue->getHeight(), + detectTmpValue->getWidth(), + false, + false); + + MatrixPtr labelTmpValue = arguments[1].value; + Matrix::resizeOrCreate(cpuLabel_, + labelTmpValue->getHeight(), + labelTmpValue->getWidth(), + false, + false); + + cpuOutput_->copyFrom(*detectTmpValue); + cpuLabel_->copyFrom(*labelTmpValue); + + Argument label = arguments[1]; + const int* labelIndex = label.sequenceStartPositions->getData(false); + size_t batchSize = label.getNumSequences(); + + vector>> allGTBBoxes; + vector>>> allDetectBBoxes; + + for (size_t n = 0; n < batchSize; ++n) { + map> bboxes; + for (int i = labelIndex[n]; i < labelIndex[n + 1]; ++i) { + vector bbox; + getBBoxFromLabelData(cpuLabel_->getData() + i * 6, 1, bbox); + int c = cpuLabel_->getData()[i * 6]; + bboxes[c].push_back(bbox[0]); + } + allGTBBoxes.push_back(bboxes); + } + + size_t n = 0; + const real* cpuOutputData = cpuOutput_->getData(); + for (size_t imgId = 0; imgId < batchSize; ++imgId) { + map>> bboxes; + size_t curImgId = static_cast((cpuOutputData + n * 7)[0]); + while (curImgId == imgId && n < cpuOutput_->getHeight()) { + vector label; + vector score; + vector bbox; + getBBoxFromDetectData(cpuOutputData + n * 7, 1, label, score, bbox); + bboxes[label[0]].push_back(make_pair(score[0], bbox[0])); + ++n; + curImgId = static_cast((cpuOutputData + n * 7)[0]); + } + allDetectBBoxes.push_back(bboxes); + } + + for (size_t n = 0; n < batchSize; ++n) { + for (map>::iterator it = + allGTBBoxes[n].begin(); + it != allGTBBoxes[n].end(); + ++it) { + size_t count = 0; + if (evaluateDifficult_) { + count = it->second.size(); + } else { + for (size_t i = 0; i < it->second.size(); ++i) + if (!(it->second[i].isDifficult)) ++count; + } + if (numPos_.find(it->first) == numPos_.end() && count != 0) { + numPos_[it->first] = count; + } else { + numPos_[it->first] += count; + } + } + } + + // calcTFPos + calcTFPos(batchSize, allGTBBoxes, allDetectBBoxes); + + return 0; + } + + virtual void printStats(std::ostream& os) const { + real mAP = calcMAP(); + os << "Detection mAP=" << mAP; + } + + virtual void distributeEval(ParameterClient2* client) { + LOG(FATAL) << "Distribute detection evaluation not implemented."; + } + + protected: + void calcTFPos(const size_t batchSize, + const vector>>& allGTBBoxes, + const vector>>>& + allDetectBBoxes) { + for (size_t n = 0; n < allDetectBBoxes.size(); ++n) { + if (allGTBBoxes[n].size() == 0) { + for (map>>::const_iterator + it = allDetectBBoxes[n].begin(); + it != allDetectBBoxes[n].end(); + ++it) { + size_t label = it->first; + for (size_t i = 0; i < it->second.size(); ++i) { + allTruePos_[label].push_back(make_pair(it->second[i].first, 0)); + allFalsePos_[label].push_back(make_pair(it->second[i].first, 1)); + } + } + } else { + for (map>>::const_iterator + it = allDetectBBoxes[n].begin(); + it != allDetectBBoxes[n].end(); + ++it) { + size_t label = it->first; + vector> predBBoxes = it->second; + if (allGTBBoxes[n].find(label) == allGTBBoxes[n].end()) { + for (size_t i = 0; i < predBBoxes.size(); ++i) { + allTruePos_[label].push_back(make_pair(predBBoxes[i].first, 0)); + allFalsePos_[label].push_back(make_pair(predBBoxes[i].first, 1)); + } + } else { + vector gtBBoxes = + allGTBBoxes[n].find(label)->second; + vector visited(gtBBoxes.size(), false); + // Sort detections in descend order based on scores + std::sort(predBBoxes.begin(), + predBBoxes.end(), + sortScorePairDescend); + for (size_t i = 0; i < predBBoxes.size(); ++i) { + real maxOverlap = -1.0; + size_t maxIdx = 0; + for (size_t j = 0; j < gtBBoxes.size(); ++j) { + real overlap = + jaccardOverlap(predBBoxes[i].second, gtBBoxes[j]); + if (overlap > maxOverlap) { + maxOverlap = overlap; + maxIdx = j; + } + } + if (maxOverlap > overlapThreshold_) { + if (evaluateDifficult_ || + (!evaluateDifficult_ && !gtBBoxes[maxIdx].isDifficult)) { + if (!visited[maxIdx]) { + allTruePos_[label].push_back( + make_pair(predBBoxes[i].first, 1)); + allFalsePos_[label].push_back( + make_pair(predBBoxes[i].first, 0)); + visited[maxIdx] = true; + } else { + allTruePos_[label].push_back( + make_pair(predBBoxes[i].first, 0)); + allFalsePos_[label].push_back( + make_pair(predBBoxes[i].first, 1)); + } + } + } else { + allTruePos_[label].push_back(make_pair(predBBoxes[i].first, 0)); + allFalsePos_[label].push_back( + make_pair(predBBoxes[i].first, 1)); + } + } + } + } + } + } + } + + real calcMAP() const { + real mAP = 0.0; + size_t count = 0; + for (map::const_iterator it = numPos_.begin(); + it != numPos_.end(); + ++it) { + size_t label = it->first; + size_t labelNumPos = it->second; + if (labelNumPos == 0 || allTruePos_.find(label) == allTruePos_.end()) + continue; + vector> labelTruePos = allTruePos_.find(label)->second; + vector> labelFalsePos = + allFalsePos_.find(label)->second; + // Compute average precision. + vector tpCumSum; + getAccumulation(labelTruePos, &tpCumSum); + vector fpCumSum; + getAccumulation(labelFalsePos, &fpCumSum); + std::vector precision, recall; + size_t num = tpCumSum.size(); + // Compute Precision. + for (size_t i = 0; i < num; ++i) { + CHECK_LE(tpCumSum[i], labelNumPos); + precision.push_back(static_cast(tpCumSum[i]) / + static_cast(tpCumSum[i] + fpCumSum[i])); + recall.push_back(static_cast(tpCumSum[i]) / labelNumPos); + } + // VOC2007 style + if (apType_ == "11point") { + vector maxPrecisions(11, 0.0); + int startIdx = num - 1; + for (int j = 10; j >= 0; --j) + for (int i = startIdx; i >= 0; --i) { + if (recall[i] < j / 10.) { + startIdx = i; + if (j > 0) maxPrecisions[j - 1] = maxPrecisions[j]; + break; + } else { + if (maxPrecisions[j] < precision[i]) + maxPrecisions[j] = precision[i]; + } + } + for (int j = 10; j >= 0; --j) mAP += maxPrecisions[j] / 11; + ++count; + } else if (apType_ == "Integral") { + // Nature integral + real averagePrecisions = 0.; + real prevRecall = 0.; + for (size_t i = 0; i < num; ++i) { + if (fabs(recall[i] - prevRecall) > 1e-6) + averagePrecisions += precision[i] * fabs(recall[i] - prevRecall); + prevRecall = recall[i]; + } + mAP += averagePrecisions; + ++count; + } else { + LOG(FATAL) << "Unkown ap version: " << apType_; + } + } + if (count != 0) mAP /= count; + return mAP * 100; + } + + void getAccumulation(vector> inPairs, + vector* accuVec) const { + std::stable_sort( + inPairs.begin(), inPairs.end(), sortScorePairDescend); + accuVec->clear(); + size_t sum = 0; + for (size_t i = 0; i < inPairs.size(); ++i) { + sum += inPairs[i].second; + accuVec->push_back(sum); + } + } + + std::string getTypeImpl() const { return "detection_map"; } + + real getValueImpl() const { return calcMAP(); } + + private: + real overlapThreshold_; // overlap threshold when determining whether matched + bool evaluateDifficult_; // whether evaluate difficult ground truth + size_t backgroundId_; // class index of background + std::string apType_; // how to calculate mAP (Integral or 11point) + + MatrixPtr cpuOutput_; + MatrixPtr cpuLabel_; + + map numPos_; // counts of true objects each classification + map>> + allTruePos_; // true positive prediction + map>> + allFalsePos_; // false positive prediction +}; + +REGISTER_EVALUATOR(detection_map, DetectionMAPEvaluator); + +} // namespace paddle diff --git a/paddle/legacy/gserver/evaluators/Evaluator.cpp b/paddle/legacy/gserver/evaluators/Evaluator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a956f40d02e39ac57ca745988491c2b54741dca3 --- /dev/null +++ b/paddle/legacy/gserver/evaluators/Evaluator.cpp @@ -0,0 +1,1361 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/gserver/evaluators/Evaluator.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/StringUtil.h" + +DECLARE_int32(trainer_id); + +namespace paddle { + +void Evaluator::eval(const NeuralNetwork& nn) { + std::vector arguments; + arguments.reserve(config_.input_layers_size()); + for (const std::string& name : config_.input_layers()) { + arguments.push_back(nn.getLayer(name)->getOutput()); + } + SetDevice device(arguments[0].deviceId); + real score = evalImp(arguments); + totalScore_ += score; + updateSamplesNum(arguments); +} +/** + * @brief classification error Evaluator + * + * The config file api is classification_error_evaluator. + */ +class ClassificationErrorEvaluator : public Evaluator { + public: + /* + ClassificationErrorEvaluator() : totalScore2_(0) {} + + virtual void start() { + Evaluator::start(); + totalScore2_ = 0; + } */ + + virtual void updateSamplesNum(const std::vector& arguments) { + if (3 == arguments.size()) { + numSamples_ += arguments[2].value->getSum(); + } else { + numSamples_ += arguments[0].getBatchSize(); + } + } + + MatrixPtr calcError(std::vector& arguments) { + CHECK_GE(arguments.size(), (size_t)2); + CHECK_LE(arguments.size(), (size_t)3); + MatrixPtr& output = arguments[0].value; + IVectorPtr& label = arguments[1].ids; + MatrixPtr& multiBinaryLabel = arguments[1].value; // For multi binary label + bool supportWeight = (3 == arguments.size()) ? true : false; + MatrixPtr weight = supportWeight ? arguments[2].value : nullptr; + if (nullptr == output || + (nullptr == label && nullptr == multiBinaryLabel) || + (supportWeight && nullptr == weight)) { + return 0; + } + + if (label != nullptr) { + CHECK_EQ(label->getSize(), output->getHeight()); + } else { + CHECK_EQ(multiBinaryLabel->getHeight(), output->getHeight()); + CHECK_EQ(multiBinaryLabel->getWidth(), output->getWidth()); + } + if (supportWeight) { + CHECK_EQ(output->getHeight(), weight->getHeight()); + CHECK_EQ((size_t)1, weight->getWidth()); + } + + const MatrixPtr errorMat = Matrix::create(output->getHeight(), + 1, + /* trans= */ false, + useGpu(arguments[0].deviceId)); + + errorMat->zeroMem(); + + if (label != nullptr) { + errorMat->classificationError(*output, *label, config_.top_k()); + } else if (dynamic_cast(multiBinaryLabel.get()) || + dynamic_cast(multiBinaryLabel.get())) { + errorMat->classificationErrorMulti( + *output, *multiBinaryLabel, config_.classification_threshold()); + } else { + errorMat->binaryClassificationError( + 0, *output, *multiBinaryLabel, config_.classification_threshold()); + } + + if (supportWeight) { + errorMat->dotMul(*errorMat, *weight); + } + return errorMat; + } + + void printStats(std::ostream& os) const { + if (config_.top_k() == 1) { + os << config_.name() << "=" + << (numSamples_ ? totalScore_ / numSamples_ : 0); + } else { + os << " top_" << config_.top_k() + << "_error=" << (numSamples_ ? totalScore_ / numSamples_ : 0); + } + } + + virtual real evalImp(std::vector& arguments) { + MatrixPtr errorMat = calcError(arguments); + return errorMat->getSum(); + } + + virtual void distributeEval(ParameterClient2* client) { + mergeResultsOfAllClients(client); + } + + // Evaluator interface + protected: + std::string getTypeImpl() const { return "classification_error"; } +}; + +/** + * @brief sequence classification error Evaluator + * @note sequence level classification error stats, + * if any frame in one sequence has error, the sequence is error + */ +class SequenceClassificationErrorEvaluator + : public ClassificationErrorEvaluator { + public: + virtual void updateSamplesNum(const std::vector& arguments) { + numSamples_ += arguments[0].getNumSequences(); + } + + virtual real evalImp(std::vector& arguments) { + auto sequenceStartPositions = + arguments[0].sequenceStartPositions->getVector(false); + CHECK(sequenceStartPositions != nullptr); + const int* starts = sequenceStartPositions->getData(); + + MatrixPtr errorMat = calcError(arguments); + + int errCounter = 0; + CpuVector errorVec(0, nullptr); + for (size_t i = 0; i < sequenceStartPositions->getSize() - 1; ++i) { + errorVec.subVecFrom( + errorMat->getData(), starts[i], starts[i + 1] - starts[i]); + if (errorVec.getSum() > 0) { + errCounter += 1; + } + } + + return static_cast(errCounter); + } + + virtual void distributeEval(ParameterClient2* client) { + mergeResultsOfAllClients(client); + } + + // Evaluator interface + protected: + std::string getTypeImpl() const { return "seq_classification_error"; } +}; +REGISTER_EVALUATOR(seq_classification_error, + SequenceClassificationErrorEvaluator); +/** + * @brief sum Evaluator + * Calculate the sum of output or label + * + * The config file api is sum_evaluator. + */ +class SumEvaluator : public Evaluator { + public: + SumEvaluator() : cpuLabel_(nullptr), cpuWeight_(nullptr) {} + + virtual void updateSamplesNum(const std::vector& arguments) { + if (2 == arguments.size()) { + numSamples_ += arguments[1].value->getSum(); + } else { + numSamples_ += arguments[0].getBatchSize(); + } + } + + virtual real evalImp(std::vector& arguments) { + REGISTER_TIMER("SumEvaluator"); + CHECK_GE(arguments.size(), (size_t)1); + CHECK_LE(arguments.size(), (size_t)2); + bool supportWeight = (2 == arguments.size()) ? true : false; + if (supportWeight) { + if (nullptr == arguments[1].value) { + return 0; + } + CHECK_EQ(arguments[1].value->getWidth(), (size_t)1); + } + + // The sum of output + if (arguments[0].value) { + if (supportWeight) { + CHECK_EQ(arguments[0].value->getHeight(), + arguments[1].value->getHeight()); + MatrixPtr tmpMat = Matrix::create(arguments[0].value->getHeight(), + arguments[0].value->getWidth(), + /* trans= */ false, + arguments[0].value->useGpu()); + tmpMat->copyFrom(*arguments[0].value); + tmpMat->rowScale(0, *tmpMat, *arguments[1].value); + return tmpMat->getSum(); + } else { + return arguments[0].value->getSum(); + } + // The sum of label + } else if (arguments[0].ids) { + size_t insNum = arguments[0].ids->getSize(); + IVectorPtr label = arguments[0].ids; + MatrixPtr weight = supportWeight ? arguments[1].value : nullptr; + if (dynamic_cast(label.get())) { + IVector::resizeOrCreate(cpuLabel_, insNum, false); + cpuLabel_->copyFrom(*arguments[0].ids); + + if (supportWeight) { + CHECK_EQ(insNum, arguments[1].value->getHeight()); + Matrix::resizeOrCreate(cpuWeight_, insNum, (size_t)1, false, false); + cpuWeight_->copyFrom(*arguments[1].value); + } + + label = cpuLabel_; + weight = cpuWeight_; + } + + if (supportWeight) { + real score = 0.0; + int* labelD = label->getData(); + real* weightD = weight->getData(); + for (size_t i = 0; i < insNum; ++i) { + score += (labelD[i] * weightD[i]); + } + return score; + } else { + return label->getSum(); + } + } else { + return 0; + } + } + + virtual void distributeEval(ParameterClient2* client) { + mergeResultsOfAllClients(client); + } + + private: + IVectorPtr cpuLabel_; + MatrixPtr cpuWeight_; + + // Evaluator interface + protected: + std::string getTypeImpl() const { return "sum"; } +}; +/** + * @brief column sum Evaluator + * @note column sum for the colIdx-th column * + * - colIdx = 0: the 0-th column. + * - colIdx > 0: the colIdx-th column. + * - colIdx < 0: the last colIdx-th column. + * + * The config file api is column_sum_evaluator. + * + */ +class ColumnSumEvaluator : public Evaluator { + public: + explicit ColumnSumEvaluator(int32_t colIdx) + : colIdx_(colIdx), colNum_(0), sum_(nullptr) {} + + virtual void start() { + Evaluator::start(); + if (nullptr != sum_) { + sum_->zeroMem(); + } + } + + virtual void updateSamplesNum(const std::vector& arguments) { + if (2 == arguments.size()) { + numSamples_ += arguments[1].value->getSum(); + } else { + numSamples_ += arguments[0].getBatchSize(); + } + } + + virtual real evalImp(std::vector& arguments) { + REGISTER_TIMER("ColumnSumEvaluator"); + CHECK_GE(arguments.size(), (size_t)1); + CHECK_LE(arguments.size(), (size_t)2); + bool supportWeight = (2 == arguments.size()) ? true : false; + if (nullptr == arguments[0].value || + (supportWeight && nullptr == arguments[1].value)) { + return 0; + } + + size_t insNum = arguments[0].value->getHeight(); + size_t colNum = arguments[0].value->getWidth(); + if (nullptr == sum_) { + sum_ = Matrix::create((size_t)1, colNum, false, /* useGpu */ false); + colNum_ = colNum; + sum_->zeroMem(); + } else { + CHECK_EQ(colNum, sum_->getWidth()); + } + + if (supportWeight) { + CHECK_EQ(insNum, arguments[1].value->getHeight()); + CHECK_EQ((size_t)1, arguments[1].value->getWidth()); + MatrixPtr tmpMat = Matrix::create(insNum, colNum); + if (arguments[0].value->useGpu()) { + tmpMat->copyFrom(*arguments[0].value); + } + if (!arguments[1].value->useGpu()) { + if (!arguments[0].value->useGpu()) { + tmpMat->rowScale(0, *arguments[0].value, *arguments[1].value); + } else { + tmpMat->rowScale(0, *tmpMat, *arguments[1].value); + } + } else { + MatrixPtr tmp2 = Matrix::create(insNum, 1); + tmp2->copyFrom(*arguments[1].value); + if (!arguments[0].value->useGpu()) { + tmpMat->rowScale(0, *arguments[0].value, *tmp2); + } else { + tmpMat->rowScale(0, *tmpMat, *tmp2); + } + } + sum_->accumulateColSum(*tmpMat); + } else { + if (!arguments[0].value->useGpu()) { + sum_->accumulateColSum(*arguments[0].value); + } else { + MatrixPtr tmpMat = Matrix::create(insNum, colNum); + tmpMat->copyFrom(*arguments[0].value); + sum_->accumulateColSum(*tmpMat); + } + } + return 0; + } + + virtual void printStats(std::ostream& os) const { + CHECK(colIdx_ + (int32_t)colNum_ >= 0 && colIdx_ - (int32_t)colNum_ < 0) + << "column index [" << colIdx_ << "] out of range [-" << colNum_ << ", " + << colNum_ << ")"; + size_t colIdx = 0; + if (colIdx_ >= 0) { + colIdx = colIdx_; + } else { + colIdx = colNum_ + colIdx_; + } + os << config_.name() << "=" + << (numSamples_ ? sum_->getElement(0, colIdx) / numSamples_ : 0); + } + + void distributeEval(ParameterClient2* client) { + client->reduce( + sum_->getData(), sum_->getData(), colNum_, FLAGS_trainer_id, 0); + client->reduce(&numSamples_, &numSamples_, 1, FLAGS_trainer_id, 0); + } + + private: + int32_t colIdx_; + size_t colNum_; + MatrixPtr sum_; /* cpu matrix */ + + // Evaluator interface + protected: + std::string getTypeImpl() const { + if (colIdx_ == -1) + return "last-column-sum"; + else + return "column-sum"; + } +}; + +void AucEvaluator::start() { + Evaluator::start(); + memset(statPos_, 0, sizeof(statPos_)); + memset(statNeg_, 0, sizeof(statNeg_)); +} + +real AucEvaluator::evalImp(std::vector& arguments) { + REGISTER_TIMER("AucEvaluator"); + CHECK_GE(arguments.size(), (size_t)2); + CHECK_LE(arguments.size(), (size_t)3); + MatrixPtr output = arguments[0].value; + IVectorPtr label = arguments[1].ids; + MatrixPtr labelval = arguments[1].value; + bool supportWeight = (3 == arguments.size()) ? true : false; + MatrixPtr weight = supportWeight ? arguments[2].value : nullptr; + + if (nullptr == output || (supportWeight && nullptr == weight)) { + return 0; + } + size_t insNum = output->getHeight(); + size_t outputDim = output->getWidth(); + // Copy label from value to a vector. + if (nullptr == label && nullptr != labelval) { + // label width is 1 + CHECK_EQ(1U, labelval->getWidth()); + VectorPtr vec = + Vector::create(labelval->getData(), insNum, output->useGpu()); + label = vec->castToInt(); + } + + CHECK_EQ(insNum, label->getSize()); + if (supportWeight) { + CHECK_EQ(insNum, weight->getHeight()); + CHECK_EQ((size_t)1, weight->getWidth()); + } + + CHECK(colIdx_ + (int32_t)outputDim >= 0 && colIdx_ - (int32_t)outputDim < 0) + << "column index [" << colIdx_ << "] out of range [-" << outputDim << ", " + << outputDim << ")"; + realColumnIdx_ = 0; + if (colIdx_ >= 0) { + realColumnIdx_ = colIdx_; + } else { + realColumnIdx_ = outputDim + colIdx_; + } + + if (dynamic_cast(output.get())) { + Matrix::resizeOrCreate(cpuOutput_, + insNum, + outputDim, + /* trans=*/false, + /* useGpu=*/false); + cpuOutput_->copyFrom(*output); + IVector::resizeOrCreate(cpuLabel_, insNum, false); + cpuLabel_->copyFrom(*label); + + if (supportWeight) { + Matrix::resizeOrCreate(cpuWeight_, insNum, (size_t)1, false, false); + cpuWeight_->copyFrom(*weight); + } + + output = cpuOutput_; + label = cpuLabel_; + weight = cpuWeight_; + } + + real* outputD = output->getData(); + int* labelD = label->getData(); + real* weightD = supportWeight ? weight->getData() : nullptr; + size_t pos = realColumnIdx_; + + for (size_t i = 0; i < insNum; ++i) { + real value = outputD[pos]; + uint32_t binIdx = static_cast(value * kBinNum_); + CHECK(binIdx <= kBinNum_) << "bin index [" << binIdx + << "] out of range, predict value[" << value + << "]"; + real w = supportWeight ? weightD[i] : 1.0; + if (labelD[i] == kNegativeLabel_) { + statNeg_[binIdx] += w; + } else { + statPos_[binIdx] += w; + } + pos += outputDim; + } + return 0; +} + +void AucEvaluator::distributeEval(ParameterClient2* client) { + client->reduce(statPos_, statPos_, kBinNum_ + 1, FLAGS_trainer_id, 0); + client->reduce(statNeg_, statNeg_, kBinNum_ + 1, FLAGS_trainer_id, 0); +} + +double AucEvaluator::calcAuc() const { + double totPos = 0.0; + double totNeg = 0.0; + double totPosPrev = 0.0; + double totNegPrev = 0.0; + double auc = 0.0; + + int64_t idx = kBinNum_; + while (idx >= 0) { + totPosPrev = totPos; + totNegPrev = totNeg; + totPos += statPos_[idx]; + totNeg += statNeg_[idx]; + auc += trapezoidArea(totNeg, totNegPrev, totPos, totPosPrev); + --idx; + } + + if (totPos > 0.0 && totNeg > 0.0) { + return auc / totPos / totNeg; + } else { + return 0.0; + } +} + +real AucEvaluator::getValueImpl() const { return calcAuc(); } + +std::string AucEvaluator::getTypeImpl() const { + if (colIdx_ == -1) { + return "last-column-auc"; + } else { + return "auc"; + } +} + +// class RankAucEvaluator +REGISTER_EVALUATOR(rankauc, RankAucEvaluator); + +void RankAucEvaluator::start() { Evaluator::start(); } +void RankAucEvaluator::updateSamplesNum( + const std::vector& arguments) { + numSamples_ += arguments[0].getNumSequences(); +} +real RankAucEvaluator::evalImp(std::vector& arguments) { + CHECK_GE(arguments.size(), 2U); + CHECK_LE(arguments.size(), 3U); + double batchAuc = 0.0; + output_ = arguments[0].value; + click_ = arguments[1].value; + size_t batchSize = output_->getHeight(); + CHECK(!output_->useGpu()) << "RankAUC evaluator does not support GPU!"; + + if (arguments.size() == 3U) { + pv_ = arguments[2].value; + } else { + Matrix::resizeOrCreate(pv_, batchSize, 1, false, false); + std::fill(pv_->getData(), pv_->getData() + batchSize, 1.0); + } + + real* outputData = output_->getData(); + real* clickData = click_->getData(); + real* pvData = pv_->getData(); + + auto startPos = arguments[0].sequenceStartPositions->getVector(false); + const int* startPosData = startPos->getData(); + size_t batchNum = startPos->getSize() - 1; + for (size_t i = 0; i < batchNum; ++i) { + int beginPos = startPosData[i]; + int endPos = startPosData[i + 1]; + batchAuc += calcRankAuc(outputData + beginPos, + clickData + beginPos, + pvData + beginPos, + endPos - beginPos); + } + return batchAuc; +} + +double RankAucEvaluator::calcRankAuc(real* outputData, + real* clickData, + real* pvData, + size_t size) { + outputPair_.clear(); + for (size_t i = 0; i < size; ++i) { + outputPair_.push_back(std::make_pair(outputData[i], i)); + } + std::sort(outputPair_.begin(), + outputPair_.end(), + [](const std::pair& a, const std::pair& b) { + return a.first > b.first; + }); + double aucTmp = 0.0; + double clickSum = 0.0; + double oldClickSum = 0.0; + double noClick = 0.0; + double noClickSum = 0.0; + + double lastScore = outputPair_[0].first + 1.0; + for (size_t i = 0; i < size; ++i) { + if (lastScore != outputPair_[i].first) { + aucTmp += (clickSum + oldClickSum) * noClick / 2.0; + oldClickSum = clickSum; + noClick = 0.0; + lastScore = outputPair_[i].first; + } + size_t id = outputPair_[i].second; + noClick += pvData[id] - clickData[id]; + noClickSum += noClick; + clickSum += clickData[id]; + } + aucTmp += (clickSum + oldClickSum) * noClick / 2.0; + return (clickSum * noClickSum) == 0.0 ? 0.0 + : aucTmp / (clickSum * noClickSum); +} + +std::string RankAucEvaluator::getTypeImpl() const { return "rankauc"; } + +// class PrecisionRecallEvaluator +REGISTER_EVALUATOR(precision_recall, PrecisionRecallEvaluator); + +void PrecisionRecallEvaluator::start() { + Evaluator::start(); + statsInfo_.clear(); + values_.clear(); +} + +real PrecisionRecallEvaluator::evalImp(std::vector& arguments) { + REGISTER_TIMER("PrecisionRecallEvaluator"); + CHECK_GE(arguments.size(), (size_t)2); + CHECK_LE(arguments.size(), (size_t)3); + MatrixPtr output = arguments[0].value; + IVectorPtr label = arguments[1].ids; + MatrixPtr multiBinaryLabel = arguments[1].value; + bool supportWeight = (3 == arguments.size()) ? true : false; + MatrixPtr weight = supportWeight ? arguments[2].value : nullptr; + if (nullptr == output || (nullptr == label && nullptr == multiBinaryLabel) || + (supportWeight && nullptr == weight)) { + return 0; + } + + size_t insNum = output->getHeight(); + size_t outputDim = output->getWidth(); + if (label != nullptr) { + CHECK_EQ(insNum, label->getSize()); + } else { + CHECK_EQ(insNum, multiBinaryLabel->getHeight()); + CHECK_EQ(outputDim, multiBinaryLabel->getWidth()); + } + if (supportWeight) { + CHECK_EQ(insNum, weight->getHeight()); + CHECK_EQ((size_t)1, weight->getWidth()); + } + + if (statsInfo_.size() != outputDim) { + statsInfo_.clear(); + statsInfo_.resize(outputDim); + } + + isMultiBinaryLabel_ = (nullptr == label) ? true : false; + if (label != nullptr) { + if (dynamic_cast(output.get())) { + Matrix::resizeOrCreate(cpuOutput_, insNum, outputDim, false, false); + cpuOutput_->copyFrom(*output); + IVector::resizeOrCreate(cpuLabel_, insNum, false); + cpuLabel_->copyFrom(*label); + if (supportWeight) { + Matrix::resizeOrCreate(cpuWeight_, insNum, (size_t)1, false, false); + cpuWeight_->copyFrom(*weight); + } + + output = cpuOutput_; + label = cpuLabel_; + weight = cpuWeight_; + } + calcStatsInfo(output, label, weight); + } else { + // Not support GPU for multi binary labels + CHECK(dynamic_cast(multiBinaryLabel.get())); + calcStatsInfoMulti(output, multiBinaryLabel, weight); + } + return 0; +} + +void PrecisionRecallEvaluator::printStats(std::ostream& os) const { + PrintStatsInfo info; + bool containMacroMicroInfo = getStatsInfo(&info); + os << "positive_label=" << config_.positive_label() + << " precision=" << info.precision << " recall=" << info.recall + << " F1-score=" << info.f1; + if (containMacroMicroInfo) { + os << "macro-average-precision=" << info.macroAvgPrecision + << " macro-average-recall=" << info.macroAvgRecall + << " macro-average-F1-score=" << info.macroAvgF1Score; + if (!isMultiBinaryLabel_) { + // precision and recall are equal in this case + os << " micro-average-precision=" << info.microAvgPrecision; + } else { + os << " micro-average-precision=" << info.microAvgPrecision + << " micro-average-recall=" << info.microAvgRecall + << " micro-average-F1-score=" << info.microAvgF1Score; + } + } +} + +void PrecisionRecallEvaluator::calcStatsInfo(const MatrixPtr& output, + const IVectorPtr& label, + const MatrixPtr& weight) { + size_t insNum = output->getHeight(); + size_t dim = output->getWidth(); + real* outputD = output->getData(); + int* labelD = label->getData(); + real* weightD = (weight != nullptr) ? weight->getData() : nullptr; + for (size_t i = 0; i < insNum; ++i) { + CHECK_GE(labelD[i], 0); + CHECK_LT((size_t)labelD[i], dim); + size_t maxIdx = 0; + real maxValue = outputD[i * dim]; + for (size_t j = 1; j < dim; ++j) { + size_t idx = i * dim + j; + if (maxValue < outputD[idx]) { + maxIdx = j; + maxValue = outputD[idx]; + } + } + + real w = (weightD != nullptr) ? weightD[i] : 1.0; + if (maxIdx == (size_t)labelD[i]) { + statsInfo_[maxIdx].TP += w; // true positive for labelD[i] + // true negative for all labels except for labelD[i] + for (size_t j = 0; j < dim; ++j) { + statsInfo_[j].TN += w; + } + statsInfo_[maxIdx].TN -= w; + } else { + statsInfo_[labelD[i]].FN += w; // false negative for labelD[i] + statsInfo_[maxIdx].FP += w; // false positive for maxIdx + // true negatives for all labels except for maxIdx and labelD[i] + for (size_t j = 0; j < dim; ++j) { + statsInfo_[j].TN += w; + } + statsInfo_[maxIdx].TN -= w; + statsInfo_[labelD[i]].TN -= w; + } + } +} + +void PrecisionRecallEvaluator::calcStatsInfoMulti(const MatrixPtr& output, + const MatrixPtr& label, + const MatrixPtr& weight) { + size_t insNum = output->getHeight(); + size_t dim = output->getWidth(); + real* outputD = output->getData(); + auto labelD = dynamic_cast(label.get()); + real* weightD = (weight != nullptr) ? weight->getData() : nullptr; + real threshold = config_.classification_threshold(); + for (size_t i = 0; i < insNum; ++i) { + for (size_t j = 0; j < dim; ++j) { + real w = (weightD != nullptr) ? weightD[i] : 1.0; + size_t idx = i * dim + j; + if (outputD[idx] < threshold) { + statsInfo_[j].TN += w; // true negative + } else { + statsInfo_[j].FP += w; // false positive + } + } + + const int* cols = labelD->getRowCols(i); + for (size_t j = 0; j < labelD->getColNum(i); ++j) { + CHECK_LT(size_t(cols[j]), dim); + real w = (weightD != nullptr) ? weightD[i] : 1.0; + size_t idx = i * dim + cols[j]; + if (outputD[idx] < threshold) { + statsInfo_[cols[j]].FN += w; // false negative + statsInfo_[cols[j]].TN -= w; // true negative + } else { + statsInfo_[cols[j]].TP += w; // true positive + statsInfo_[cols[j]].FP -= w; // false positive + } + } + } +} + +void PrecisionRecallEvaluator::storeLocalValues() const { + if (this->values_.size() == 0) { + PrintStatsInfo info; + bool containMacroMicroInfo = getStatsInfo(&info); + values_["precision"] = info.precision; + values_["recal"] = info.recall; + values_["F1-score"] = info.f1; + if (containMacroMicroInfo) { + values_["macro-average-precision"] = info.macroAvgPrecision; + values_["macro-average-recall"] = info.macroAvgRecall; + values_["macro-average-F1-score"] = info.macroAvgF1Score; + if (!isMultiBinaryLabel_) { + // precision and recall are equal in this case + values_["micro-average-precision"] = info.microAvgPrecision; + } else { + values_["micro-average-precision"] = info.microAvgPrecision; + values_["micro-average-recall"] = info.microAvgRecall; + values_["micro-average-F1-score"] = info.microAvgF1Score; + } + } + } +} + +void PrecisionRecallEvaluator::getNames(std::vector* names) { + this->storeLocalValues(); + names->reserve(this->values_.size()); + for (auto it = this->values_.begin(); it != this->values_.end(); ++it) { + names->push_back(this->config_.name() + "." + it->first); + } +} + +real PrecisionRecallEvaluator::getValue(const std::string& name, + Error* err) const { + this->storeLocalValues(); + std::vector buffers; + paddle::str::split(name, '.', &buffers); + auto it = this->values_.find(buffers[buffers.size() - 1]); + if (it == this->values_.end()) { // not found + *err = Error("No such key %s", name.c_str()); + return .0f; + } + + return it->second; +} + +std::string PrecisionRecallEvaluator::getType(const std::string& name, + Error* err) const { + this->getValue(name, err); + if (!err->isOK()) { + return ""; + } + return "precision_recall"; +} + +void PrecisionRecallEvaluator::distributeEval(ParameterClient2* client) { + size_t size = 4 * statsInfo_.size(); + double* buf = new double[size]; + for (size_t i = 0; i < statsInfo_.size(); ++i) { + buf[4 * i + 0] = statsInfo_[i].TP; + buf[4 * i + 1] = statsInfo_[i].TN; + buf[4 * i + 2] = statsInfo_[i].FP; + buf[4 * i + 3] = statsInfo_[i].FN; + } + client->reduce(buf, buf, size, FLAGS_trainer_id, 0); + for (size_t i = 0; i < statsInfo_.size(); ++i) { + statsInfo_[i].TP = buf[4 * i + 0]; + statsInfo_[i].TN = buf[4 * i + 1]; + statsInfo_[i].FP = buf[4 * i + 2]; + statsInfo_[i].FN = buf[4 * i + 3]; + } + delete[] buf; +} + +bool PrecisionRecallEvaluator::getStatsInfo( + PrecisionRecallEvaluator::PrintStatsInfo* info) const { + int label = config_.positive_label(); + if (label != -1) { + CHECK(label >= 0 && label < (int)statsInfo_.size()) + << "positive_label [" << label << "] should be in range [0, " + << statsInfo_.size() << ")"; + info->precision = calcPrecision(statsInfo_[label].TP, statsInfo_[label].FP); + info->recall = calcRecall(statsInfo_[label].TP, statsInfo_[label].FN); + info->f1 = calcF1Score(info->precision, info->recall); + return false; + } + + // micro average method: precision = (TP1+TP2)/(TP1+FP1+TP2+FP2) + // macro average method: precision = (precision1+precision2)/2 + double microTotalTP = 0; + double microTotalFP = 0; + double microTotalFN = 0; + info->macroAvgPrecision = 0; + info->macroAvgRecall = 0; + size_t numLabels = statsInfo_.size(); + for (size_t i = 0; i < numLabels; ++i) { + microTotalTP += statsInfo_[i].TP; + microTotalFP += statsInfo_[i].FP; + microTotalFN += statsInfo_[i].FN; + info->macroAvgPrecision += + calcPrecision(statsInfo_[i].TP, statsInfo_[i].FP); + info->macroAvgRecall += calcRecall(statsInfo_[i].TP, statsInfo_[i].FN); + } + info->macroAvgPrecision /= numLabels; + info->macroAvgRecall /= numLabels; + info->macroAvgF1Score = + calcF1Score(info->macroAvgPrecision, info->macroAvgRecall); + + info->microAvgPrecision = calcPrecision(microTotalTP, microTotalFP); + info->microAvgRecall = calcPrecision(microTotalTP, microTotalFN); + info->microAvgF1Score = + calcF1Score(info->microAvgPrecision, info->microAvgRecall); + return true; +} + +REGISTER_EVALUATOR(pnpair, PnpairEvaluator); +void PnpairEvaluator::start() { + Evaluator::start(); + memset(pairArray_, 0, sizeof(pairArray_)); + predictArray_.clear(); +} + +real PnpairEvaluator::evalImp(std::vector& arguments) { + CHECK_GE(arguments.size(), 3UL); + CHECK_LE(arguments.size(), 4UL); + MatrixPtr output = arguments[0].value; + IVectorPtr label = arguments[1].ids; + IVectorPtr info = arguments[2].ids; + bool supportWeight = (4 == arguments.size()) ? true : false; + MatrixPtr weight = supportWeight ? arguments[3].value : nullptr; + if (nullptr == output || nullptr == label || + (supportWeight && nullptr == weight)) { + return 0; + } + size_t height = output->getHeight(); + size_t width = output->getWidth(); + CHECK_EQ(height, label->getSize()); + CHECK_EQ(height, info->getSize()); + if (supportWeight) { + CHECK_EQ(height, weight->getHeight()); + CHECK_EQ((size_t)1, weight->getWidth()); + } + + if (dynamic_cast(output.get())) { + Matrix::resizeOrCreate(cpuOutput_, height, width, false, false); + IVector::resizeOrCreate(cpuLabel_, height, false); + IVector::resizeOrCreate(cpuInfo_, height, false); + cpuOutput_->copyFrom(*output); + cpuLabel_->copyFrom(*label); + cpuInfo_->copyFrom(*info); + + output = cpuOutput_; + label = cpuLabel_; + info = cpuInfo_; + + if (supportWeight) { + Matrix::resizeOrCreate(cpuWeight_, height, (size_t)1, false, false); + cpuWeight_->copyFrom(*weight); + weight = cpuWeight_; + } + } + + real* outputs = output->getData(); + int* labels = label->getData(); + int* infos = info->getData(); + real* weights = supportWeight ? weight->getData() : nullptr; + for (size_t i = 0; i < output->getHeight(); i++) { + real y1 = outputs[i * width + (width - 1)]; + real w = supportWeight ? weights[i] : 1.0; + predictArray_.push_back(PredictionResult(y1, labels[i], infos[i], w)); + } + return 0; +} + +void PnpairEvaluator::stat(size_t start, + size_t end, + PredictionResult* answers, + double& pos, + double& neg, + double& spe) { + for (size_t i = start; i < end; i++) { + for (size_t j = i + 1; j < end; j++) { + CHECK_EQ(answers[i].queryid, answers[j].queryid); + // The pair weight is the mean of the two samples' weight + double weight = (answers[i].weight + answers[j].weight) / 2.0; + if (answers[i].label != answers[j].label) { + if ((answers[i].out > answers[j].out && + answers[i].label > answers[j].label) || + (answers[i].out < answers[j].out && + answers[i].label < answers[j].label)) { + pos += weight; + } else if ((answers[i].out > answers[j].out && + answers[i].label < answers[j].label) || + (answers[i].out < answers[j].out && + answers[i].label > answers[j].label)) { + neg += weight; + } else { + spe += weight; + } + } + } + } +} + +void PnpairEvaluator::calc(std::vector& predictArray) { + std::sort(predictArray.begin(), + predictArray.end(), + [](const PredictionResult& x, const PredictionResult& y) { + return x.queryid < y.queryid; + }); + + double pos = 0; + double neg = 0; + double special = 0; + auto start = predictArray.begin(); + while (start != predictArray.end()) { + auto end = std::find_if( + start + 1, predictArray.end(), [=](const PredictionResult& x) { + return x.queryid != start->queryid; + }); + CHECK(end != start); + stat(start - predictArray.begin(), + end - predictArray.begin(), + predictArray.data(), + pos, + neg, + special); + + start = end; + } + + pairArray_[0] += pos; + pairArray_[1] += neg; + + LOG(INFO) << " calc total pos pair: " << pos + << " calc total neg pair: " << neg + << " calc total special pair: " << special; +} + +std::string PnpairEvaluator::getTypeImpl() const { return "pnpair"; } + +ClassRegistrar Evaluator::registrar_; +Evaluator* Evaluator::create(const EvaluatorConfig& config) { + Evaluator* evaluator = registrar_.createByType(config.type()); + evaluator->init(config); + return evaluator; +} + +REGISTER_EVALUATOR(classification_error, ClassificationErrorEvaluator); +REGISTER_EVALUATOR(sum, SumEvaluator); +static InitFunction __reg_type_auc_sum__([]() { + Evaluator::registrar_.registerClass( + "last-column-sum", [] { return new ColumnSumEvaluator(-1); }); + Evaluator::registrar_.registerClass("last-column-auc", + [] { return new AucEvaluator(-1); }); +}); + +/** + * @brief print value of each layer. + * + * The config file api is value_printer_evaluator. + */ +class ValuePrinter : public NotGetableEvaluator { + public: + virtual void eval(const NeuralNetwork& nn) { + for (const std::string& name : config_.input_layers()) { + nn.getLayer(name)->getOutput().printValueString(LOG(INFO), + "layer=" + name + " "); + } + } + + virtual void updateSamplesNum(const std::vector& arguments) {} + + virtual real evalImp(std::vector& arguments) { return 0; } +}; +REGISTER_EVALUATOR(value_printer, ValuePrinter); + +/** + * @brief print gradient of each layer. + * + * The config file api is gradient_printer_evaluator. + */ +class GradientPrinter : public NotGetableEvaluator { + public: + virtual void eval(const NeuralNetwork& nn) { + for (const std::string& name : config_.input_layers()) { + const Argument& argu = nn.getLayer(name)->getOutput(); + if (argu.grad) { + std::ostringstream os; + argu.grad->print(os); + LOG(INFO) << "layer=" << name << " grad matrix:\n" << os.str(); + } + } + } + + virtual void updateSamplesNum(const std::vector& arguments) {} + + virtual real evalImp(std::vector& arguments) { return 0; } +}; +REGISTER_EVALUATOR(gradient_printer, GradientPrinter); +/** + * @brief print row max id vctor of each layer + * + * The config file api is maxid_printer_evaluator. + */ +class MaxIdPrinter : public NotGetableEvaluator { + private: + IVectorPtr maxIds_; + MatrixPtr maxValues_; + + public: + MaxIdPrinter() {} + + virtual void eval(const NeuralNetwork& nn) { + for (const std::string& name : config_.input_layers()) { + const Argument& argu = nn.getLayer(name)->getOutput(); + if (argu.value) { + size_t height = argu.value->getHeight(); + size_t width = config_.num_results(); + IVector::resizeOrCreate(maxIds_, height * width, false); + Matrix::resizeOrCreate(maxValues_, height, width, false); + argu.value->rowMax(*maxIds_, *maxValues_); + std::ostringstream os; + int* ids = maxIds_->getData(); + real* values = maxValues_->getData(); + for (size_t i = 0; i < height; ++i) { + for (size_t j = 0; j < width; ++j) { + size_t pos = i * width + j; + os << ids[pos] << " : " << values[pos] << ", "; + } + os << std::endl; + } + LOG(INFO) << "layer=" << name << " row max id vector:\n" << os.str(); + } + } + } + + virtual void updateSamplesNum(const std::vector& arguments) {} + + virtual real evalImp(std::vector& arguments) { return 0; } +}; +REGISTER_EVALUATOR(max_id_printer, MaxIdPrinter); +/** + * @brief print sequence max frames of each layer + * + * The config file api is maxframe_printer_evaluator. + */ +class MaxFramePrinter : public NotGetableEvaluator { + private: + IVectorPtr maxIds_; + MatrixPtr maxValues_; + MatrixPtr value_; + + public: + MaxFramePrinter() { + value_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, false); + } + + virtual void eval(const NeuralNetwork& nn) { + for (const std::string& name : config_.input_layers()) { + const Argument& argu = nn.getLayer(name)->getOutput(); + + CHECK_EQ(argu.value->getWidth(), 1LU); + size_t numSequences = argu.getNumSequences(); + const int* starts = argu.sequenceStartPositions->getData(false); + + std::ostringstream os; + for (size_t i = 0; i < numSequences; ++i) { + size_t offset = starts[i]; + size_t size = starts[i + 1] - starts[i]; + value_->setData(argu.value->getData() + offset, 1LU, size); + + size_t height = 1LU; + size_t width = std::min((size_t)config_.num_results(), size); + IVector::resizeOrCreate(maxIds_, height * width, false); + Matrix::resizeOrCreate(maxValues_, height, width, false); + + value_->rowMax(*maxIds_, *maxValues_); + + int* ids = maxIds_->getData(); + real* values = maxValues_->getData(); + for (size_t j = 0; j < width; ++j) { + os << ids[j] << " : " << values[j] << ", "; + } + os << "total " << size << " frames" << std::endl; + } + LOG(INFO) << "layer=" << name << " sequence max frames:\n" << os.str(); + } + } + + virtual void updateSamplesNum(const std::vector& arguments) {} + + virtual real evalImp(std::vector& arguments) { return 0; } +}; +REGISTER_EVALUATOR(max_frame_printer, MaxFramePrinter); + +/** + * @brief print text according to index matrix and a dictionary. + * + * There can be multiple input to this layer: + * - If there is only one input, the input must be a matrix containing + * the sequence of indices; + * - If there are more than one input, the first input should be ids, + * and are interpreted as sample ids. + * + * The output format will be: + * + * - sequence without sub-sequence, and there is probability. + * + * @code + * id \t prob space_seperated_tokens_from_dictionary_according_to_seq + * @endcode + * + * - sequence without sub-sequence, and there is not probability. + * + * @code + * id \t space_seperated_tokens_from_dictionary_according_to_seq + * @endcode + * + * - sequence with sub-sequence, and there is not probability. + * + * @code + * id \t space_seperated_tokens_from_dictionary_according_to_sub_seq + * \t \t space_seperated_tokens_from_dictionary_according_to_sub_seq + * ... + * @endcode + * + * Typically SequenceTextPrinter layer takes output of maxid or RecurrentGroup + * with maxid (when generating) as an input. + * + * The config file api is seqtext_printer_evaluator. + * + */ +class SequenceTextPrinter : public NotGetableEvaluator { + private: + /// dict_file, which contains a list of tokens + std::vector dict_; + /// result_file, which is the output file + std::ofstream os_; + /// True/False, to indicate whether to use space to separate output tokens. + /// Default is True. No space is added if set to False. + bool delimited_; + /// store the cpu version of argument.ids + std::vector cpuIds_; + /// store the probability associated with each sequence + std::vector cpuIn_; + + public: + SequenceTextPrinter() {} + + virtual void init(const EvaluatorConfig& config) { + Evaluator::init(config); + if (!config.dict_file().empty()) { + loadFileList(config.dict_file(), dict_); + } + + os_.open(config.result_file(), std::ofstream::trunc); + CHECK(os_.is_open()) << "Failed to open file " << config.result_file(); + delimited_ = config.delimited(); + } + + virtual void updateSamplesNum(const std::vector& arguments) {} + + virtual real evalImp(std::vector& arguments) { + CHECK_GE(arguments.size(), 1LU); + bool hasId = arguments.size() > 1; + size_t numSequences = arguments[0].getNumSequences(); + if (hasId) { + CHECK_EQ(arguments[0].ids->getSize(), numSequences) + << "first input must be sample id."; + } + for (size_t i = hasId ? 1 : 0; i < arguments.size(); ++i) { + CHECK_EQ((size_t)arguments[i].getNumSequences(), numSequences); + } + + auto resizeVector = [](IVectorPtr& dest, const IVectorPtr& src) { + if (src && src->useGpu()) { + IVector::resizeOrCreate(dest, src->getSize(), false); + dest->copyFrom(*src); + } else { + dest = src; + } + }; + + auto resizeMatrix = [](MatrixPtr& dest, const MatrixPtr& src) { + if (src && src->useGpu()) { + Matrix::resizeOrCreate( + dest, src->getHeight(), src->getWidth(), false, false); + dest->copyFrom(*src); + } else { + dest = src; + } + }; + + cpuIds_.resize(arguments.size()); + cpuIn_.resize(arguments.size()); + for (size_t i = 0; i < arguments.size(); ++i) { + resizeVector(cpuIds_[i], arguments[i].ids); + resizeMatrix(cpuIn_[i], arguments[i].in); + } + + int* sampleIds = nullptr; + if (hasId) { + sampleIds = cpuIds_[0]->getData(); + } + + for (size_t i = 0; i < numSequences; ++i) { + os_ << (hasId ? sampleIds[i] : i); + for (size_t j = hasId ? 1 : 0; j < arguments.size(); ++j) { + int* output = cpuIds_[j]->getData(); + const int* starts = arguments[j].sequenceStartPositions->getData(false); + + auto seqPrint = [&](int start, int end) { + os_ << "\t"; + for (int k = start; k < end; k++) { + int id = output[k]; + os_ << (delimited_ ? " " : ""); + if (!dict_.empty()) { + CHECK_LT((size_t)id, dict_.size()); + os_ << dict_[id]; + } else { + os_ << id; + } + } + }; + + if (arguments[j].hasSubseq()) { + // print sequence with sub-sequence + const int* subStarts = + arguments[j].subSequenceStartPositions->getData(false); + int subSeqId_start = 0; + int subSeqId_end = 0; + for (size_t k = 0; k < (size_t)arguments[j].getNumSubSequences() + 1; + ++k) { + if (starts[i] == subStarts[k]) subSeqId_start = k; + if (starts[i + 1] == subStarts[k]) subSeqId_end = k; + } + for (int k = subSeqId_start; k < subSeqId_end; k++) { + seqPrint(subStarts[k], subStarts[k + 1]); + os_ << std::endl; + } + + } else { + // print sequence without sub-sequence + if (arguments[j].in) { // beam print + real* probs = cpuIn_[j]->rowBuf(i); + os_ << std::endl; + int start = starts[i]; + int seqEnd = starts[i + 1]; + for (size_t k = 0; k < arguments[j].in->getWidth(); ++k) { + if (start == seqEnd) { + break; + } + int end = start + output[start] + 2; + CHECK_LE(end, seqEnd); + CHECK_EQ(output[end - 1], -1); + os_ << k << "\t" << probs[k]; + seqPrint(start + 1, end - 1); + os_ << std::endl; + start = end; + } + } else { + seqPrint(starts[i], starts[i + 1]); + } + } + } + os_ << std::endl; + } + return 0; + } +}; +REGISTER_EVALUATOR(seq_text_printer, SequenceTextPrinter); +/** + * @brief print classification error. + * + * The config file api is classification_error_printer_evaluator. + */ +class ClassificationErrorPrinter : public ClassificationErrorEvaluator { + public: + virtual void updateSamplesNum(const std::vector& arguments) {} + + virtual real evalImp(std::vector& arguments) { + MatrixPtr errorMat = calcError(arguments); + + std::ostringstream os; + errorMat->print(os); + LOG(INFO) << "Printer=" << config_.name() << " Classification Error:\n" + << os.str(); + + if (auto startPos = arguments[0].sequenceStartPositions) { + std::ostringstream os; + startPos->getVector(false)->print(os, startPos->getSize()); + LOG(INFO) << "Printer=" << config_.name() << " sequence pos vector:\n" + << os.str(); + } + return 0; + } +}; +REGISTER_EVALUATOR(classification_error_printer, ClassificationErrorPrinter); + +std::string DummyEvaluator::getTypeImpl() const { return "dummy"; } + +} // namespace paddle diff --git a/paddle/legacy/gserver/evaluators/Evaluator.h b/paddle/legacy/gserver/evaluators/Evaluator.h new file mode 100644 index 0000000000000000000000000000000000000000..b3462819b1244e9f2d1a463cb44e7c550406c000 --- /dev/null +++ b/paddle/legacy/gserver/evaluators/Evaluator.h @@ -0,0 +1,510 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/parameter/Argument.h" +#include "paddle/legacy/pserver/ParameterClient2.h" +#include "paddle/legacy/utils/ClassRegistrar.h" +#include "paddle/legacy/utils/Error.h" + +namespace paddle { + +class NeuralNetwork; +/** + * @def REGISTER_EVALUATOR + * @brief Macro for registering evaluator class + */ + +#define REGISTER_EVALUATOR(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name([]() { \ + Evaluator::registrar_.registerClass<__class_name>(#__type_name); \ + }) +/** + * @brief Base class for Evaluator + * Evaluating the performance of a model is very important. + * It indicates how successful the scores(predictions) of a datasets + * has been by a trained model. + */ +class Evaluator { + public: + static Evaluator* create(const EvaluatorConfig& config); + + Evaluator() : numSamples_(0), totalScore_(0) {} + + virtual ~Evaluator() {} + + virtual void init(const EvaluatorConfig& config) { config_ = config; } + + /** + * @brief start to evaluate some data + */ + virtual void start() { + numSamples_ = 0; + totalScore_ = 0; + } + + /** + * @brief Process a batch of data. + */ + virtual void eval(const NeuralNetwork& nn); + + /** + * @brief Process a batch of data. + * @return the score for the batch if it make sense to sum the score across + * batches. + * @note Otherwise evaluator should return 0 and override finish() and + * printStats() to do the right calculation. + */ + virtual real evalImp(std::vector& arguments) = 0; + + /** + * @brief Update the number of processed samples + */ + virtual void updateSamplesNum(const std::vector& arguments) { + numSamples_ += arguments[0].getBatchSize(); + } + + /// finish() should be called before distributeEval + virtual void distributeEval(ParameterClient2* client) { + LOG(FATAL) << "Not implemeted"; + } + + void mergeResultsOfAllClients(ParameterClient2* client) { + double data[2] = {totalScore_, numSamples_}; + client->reduce(data, data, 2, FLAGS_trainer_id, 0); + totalScore_ = data[0]; + numSamples_ = data[1]; + } + + /** + * @brief finish the evaluation. + */ + virtual void finish() {} + + /** + * @brief print the statistics of evaluate result + * @note finish() should be called before printStats + */ + virtual void printStats(std::ostream& os) const { + os << config_.name() << "=" + << (numSamples_ ? totalScore_ / numSamples_ : 0); + } + + friend std::ostream& operator<<(std::ostream& os, + const Evaluator& evaluator) { + evaluator.printStats(os); + return os; + } + + friend std::ostream&& operator<<(std::ostream&& os, // NOLINT + const Evaluator& evaluator) { + evaluator.printStats(os); + return std::move(os); + } + + static ClassRegistrar registrar_; + + /** + * @brief getNames will return all field names of current evaluator. + * + * The format of name is `evaluator_name.evaluator_fields`. If the evaluator + * has multiple field, the name could be `evaluator_name.field1`. For example + * the PrecisionRecallEvaluator contains `precision`, `recall` fields. The get + * names will return `precision_recall_evaluator.precision`, + * `precision_recall_evaluator.recal`, etc. + * + * Also, if current Evaluator is a combined evaluator. getNames will return + * all names of all evaluators inside the combined evaluator. + * + * @param names [out]: the field names of current evaluator. + * @note Never clear the names parameter inside getNames. + */ + virtual void getNames(std::vector* names) { + names->push_back(config_.name()); + } + + /** + * @brief getValue will return the current evaluate value of one field. + * + * @param name: The field name of current evaluator. + * @param err [out]: The error state. + * + * @return The evaluate value(metric). + */ + virtual real getValue(const std::string& name, Error* err) const { + if (name != config_.name()) { + *err = Error("no such name of evaluator %s", name.c_str()); + return .0f; + } + return this->getValueImpl(); + } + + /** + * @brief getType will return the evaluator type by field name. + * + * Evaluate Type is the current type of evaluator in string. Such as 'auc', + * 'precision_recall'. In combined evaluator, different name may get different + * evaluate type because it could be evaluated by different evaluator inside. + * + * @param name: The field name of current Evaluator. + * @param err: The error state. nullptr means don't care. + * @return the evaluator type string. + */ + virtual std::string getType(const std::string& name, Error* err) const { + if (name != config_.name()) { + *err = Error("no such name of evaluator %s", name.c_str()); + return std::string(); + } + return this->getTypeImpl(); + } + + protected: + /** + * @brief getValueImpl The simplest way to define getValue result. If this + * evaluator doesn't contain multiple fields, and do not throw any error, just + * implemented this method to get the evaluate result(metric). + * @return Evaluate result(metric). + */ + virtual real getValueImpl() const { + return numSamples_ != .0 ? totalScore_ / numSamples_ : .0; + } + + /** + * @brief getTypeImpl The simplest way to define getType result. If this + * evaluator doesn't combine many evaluators, the get type should only return + * itself type. + * @return Evaluator type. + */ + virtual std::string getTypeImpl() const { return "base"; } + + protected: + EvaluatorConfig config_; + double numSamples_; + double totalScore_; +}; + +/** + * @brief The NotGetableEvaluator class is the base class of evaluator that + * cannot get value in runtime. The most NotGetableEvaluator is Printer + * Evaluator, which is only used to debug network configuration. + */ +class NotGetableEvaluator : public Evaluator { + // Evaluator interface + public: + void getNames(std::vector* names) {} + + real getValue(const std::string& name, Error* err) const { + *err = Error("Not implemented"); + return .0f; + } + + std::string getType(const std::string& name, Error* err) const { + *err = Error("Not implemented"); + return ""; + } +}; + +class DummyEvaluator : public Evaluator { + public: + DummyEvaluator() {} + virtual void init(const EvaluatorConfig&) {} + virtual void start() {} + virtual void eval(const NeuralNetwork&) {} + virtual real evalImp(std::vector& arguments) { + (void)arguments; + return -1; + } + virtual void finish() {} + virtual void printStats(std::ostream&) const {} + + // Evaluator interface + protected: + std::string getTypeImpl() const; +}; +/** + * @brief evaluate AUC using colIdx-th column as prediction. + * The AUC(Area Under the Curve) is a common evaluation metric + * for binary classification problems. It computes the area under + * the receiver operating characteristic(ROC) curve. + * + * @note colIdx-th column + * + * - colIdx = 0: the 0-th column. + * - colIdx > 0: the colIdx-th column. + * - colIdx < 0: the last colIdx-th column. + * + * The config file api is auc_evaluator. + * + */ +class AucEvaluator : public Evaluator { + public: + AucEvaluator(int32_t colIdx) + : colIdx_(colIdx), + realColumnIdx_(0), + cpuOutput_(nullptr), + cpuLabel_(nullptr), + cpuWeight_(nullptr) {} + + virtual void start(); + + virtual real evalImp(std::vector& arguments); + + virtual void printStats(std::ostream& os) const { + os << config_.name() << "=" << calcAuc(); + } + + virtual void distributeEval(ParameterClient2* client); + + private: + static const uint32_t kBinNum_ = (1 << 24) - 1; + static const int kNegativeLabel_ = 0; + double statPos_[kBinNum_ + 1]; + double statNeg_[kBinNum_ + 1]; + int32_t colIdx_; + uint32_t realColumnIdx_; + MatrixPtr cpuOutput_; + IVectorPtr cpuLabel_; + MatrixPtr cpuWeight_; + + AucEvaluator() {} + + inline static double trapezoidArea(double X1, + double X2, + double Y1, + double Y2) { + return (X1 > X2 ? (X1 - X2) : (X2 - X1)) * (Y1 + Y2) / 2.0; + } + + double calcAuc() const; + + // Evaluator interface + protected: + real getValueImpl() const; + std::string getTypeImpl() const; +}; + +/** + * @brief RankAucEvaluator calculates the AUC of each list (i.e., titles + * under the same query), and averages them. Each list should be organized + * as a sequence. The inputs of this evaluator is [output, click, pv]. If pv + * is not provided, it will be set to 1. The types of click and pv are + * dense value. + */ +class RankAucEvaluator : public Evaluator { + public: + // evaluate ranking AUC + virtual void start(); + + virtual void updateSamplesNum(const std::vector& arguments); + + virtual real evalImp(std::vector& arguments); + + virtual void distributeEval(ParameterClient2* client) { + mergeResultsOfAllClients(client); + } + + private: + MatrixPtr output_; + MatrixPtr click_; + MatrixPtr pv_; + std::vector> outputPair_; + + double calcRankAuc(real* outputData, + real* clickData, + real* pvData, + size_t size); + + // Evaluator interface + protected: + std::string getTypeImpl() const; +}; + +/** + * @brief precision, recall and f1 score Evaluator + * \f[ + * precision = \frac{tp}{tp+tn} \\ + * recall=\frac{tp}{tp+fn} \\ + * f1=2*\frac{precsion*recall}{precision+recall} + * \f] + * + * The config file api is precision_recall_evaluator. + */ +class PrecisionRecallEvaluator : public Evaluator { + public: + // Evaluate precision, recall and F1 score + PrecisionRecallEvaluator() + : isMultiBinaryLabel_(false), + cpuOutput_(nullptr), + cpuLabel_(nullptr), + cpuWeight_(nullptr) {} + + virtual void start(); + + virtual real evalImp(std::vector& arguments); + + virtual void printStats(std::ostream& os) const; + + virtual void distributeEval(ParameterClient2* client); + + void getNames(std::vector* names); + + real getValue(const std::string& name, Error* err) const; + + std::string getType(const std::string& name, Error* err) const; + + struct StatsInfo { + /// numbers of true positives + double TP; + /// numbers of true negatives + double TN; + /// numbers of false positives + double FP; + /// numbers of false negatives + double FN; + + StatsInfo() : TP(0.0), TN(0.0), FP(0.0), FN(0.0) {} + }; + + private: + bool isMultiBinaryLabel_; + std::vector statsInfo_; + + MatrixPtr cpuOutput_; + IVectorPtr cpuLabel_; + MatrixPtr cpuWeight_; + + struct PrintStatsInfo { + double precision; + double recall; + double f1; + double macroAvgPrecision; + double macroAvgRecall; + double macroAvgF1Score; + double microAvgPrecision; + double microAvgRecall; + double microAvgF1Score; + }; + + bool getStatsInfo(PrintStatsInfo* info) const; + + void calcStatsInfo(const MatrixPtr& output, + const IVectorPtr& label, + const MatrixPtr& weight); + + void calcStatsInfoMulti(const MatrixPtr& output, + const MatrixPtr& label, + const MatrixPtr& weight); + + inline static double calcPrecision(double TP, double FP) { + if (TP > 0.0 || FP > 0.0) { + return TP / (TP + FP); + } else { + return 1.0; + } + } + + inline static double calcRecall(double TP, double FN) { + if (TP > 0.0 || FN > 0.0) { + return TP / (TP + FN); + } else { + return 1.0; + } + } + + inline static double calcF1Score(double precision, double recall) { + if (precision > 0.0 || recall > 0.0) { + return 2 * precision * recall / (precision + recall); + } else { + return 0; + } + } + + mutable std::unordered_map values_; + + void storeLocalValues() const; +}; + +/* + * @brief positive-negative pair rate Evaluator + * + * The config file api is pnpair_evaluator. + */ +class PnpairEvaluator : public Evaluator { + public: + PnpairEvaluator() + : cpuOutput_(nullptr), + cpuLabel_(nullptr), + cpuInfo_(nullptr), + cpuWeight_(nullptr) {} + + virtual void start(); + virtual real evalImp(std::vector& arguments); + + struct PredictionResult { + PredictionResult(real __out, int __label, int __queryid, real __weight) + : out(__out), label(__label), queryid(__queryid), weight(__weight) {} + real out; + int label; + int queryid; + real weight; + }; + std::vector predictArray_; + void printPredictResults() { + std::ofstream fs(FLAGS_predict_file); + CHECK(fs) << "Fail to open " << FLAGS_predict_file; + for (auto& res : predictArray_) { + fs << res.out << " " << res.label << " " << res.queryid << std::endl; + } + } + + void stat(size_t start, + size_t end, + PredictionResult* answers, + double& pos, + double& neg, + double& spe); + void calc(std::vector& predictArray); + + virtual void finish() { calc(predictArray_); } + + virtual void printStats(std::ostream& os) const { + os << " pos/neg=" << this->getValueImpl(); + } + + virtual void distributeEval(ParameterClient2* client) { + client->reduce(pairArray_, pairArray_, kPairArrayNum_, FLAGS_trainer_id, 0); + LOG(INFO) << " distribute eval calc total pos pair: " << pairArray_[0] + << " calc total neg pair: " << pairArray_[1]; + } + + private: + static const uint32_t kPairArrayNum_ = 2; + double pairArray_[kPairArrayNum_]; + MatrixPtr cpuOutput_; + IVectorPtr cpuLabel_; + IVectorPtr cpuInfo_; + MatrixPtr cpuWeight_; + + // Evaluator interface + protected: + real getValueImpl() const { + return pairArray_[0] / ((pairArray_[1] <= 0) ? 1.0 : pairArray_[1]); + } + std::string getTypeImpl() const; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/GradientMachine.cpp b/paddle/legacy/gserver/gradientmachines/GradientMachine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1c4034d8bba59dbae0a1059b96ac2b6f18c5971b --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/GradientMachine.cpp @@ -0,0 +1,104 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GradientMachine.h" + +#include +#include "paddle/legacy/utils/Logging.h" + +#include "NeuralNetwork.h" +#include "hl_gpu.h" + +#ifndef PADDLE_MOBILE_INFERENCE +#include "GradientMachineMode.h" +#include "MultiGradientMachine.h" +#include "MultiNetwork.h" +#include "ParallelNeuralNetwork.h" +#endif + +namespace paddle { + +GradientMachine* GradientMachine::create( + const ModelConfig& config, + int mode, + const std::vector& parameterTypes) { +#ifndef PADDLE_MOBILE_INFERENCE + if (auto gm = IGradientMachineMode::tryCreateGradientMachine(mode, config)) { + return gm; + } + if (FLAGS_trainer_count > 1) { + return new MultiGradientMachine(config, FLAGS_use_gpu); + } +#endif + if (FLAGS_trainer_count == 1) { // single +#ifndef PADDLE_MOBILE_INFERENCE + NeuralNetwork* nn; + if (config.type() == "multi_nn") { + /* multi submodel calculate, thread(s) will be initialized inside */ + nn = new MultiNetwork("root"); + } else if (FLAGS_parallel_nn) { + /* multi threads calculate */ + nn = new ParallelNeuralNetwork(); + } else { + /* single thread calculate */ + nn = NeuralNetwork::create(config); + } +#else + NeuralNetwork* nn = NeuralNetwork::create(config); +#endif + ParamInitCallback testParamInitCb = [](int paramId, Parameter* para) { + para->enableType(PARAMETER_VALUE); + }; + nn->init( + config, mode == kTesting ? testParamInitCb : nullptr, parameterTypes); + return nn; + } + LOG(FATAL) << "Unknown model type: " << config.type(); + return nullptr; +} + +void GradientMachine::saveParameters(const std::string& dir) const { + LOG(INFO) << "Saving parameters to " << dir; + + for (auto& para : parameters_) { + std::string filename = dir + "/" + para->getName(); + if (para->isFullSize()) { + para->save(filename); + } + } +} + +void GradientMachine::loadParameters(const std::string& dir) { + LOG(INFO) << "Loading parameters from " << dir; + + for (auto& para : parameters_) { + std::string filename = dir + "/" + para->getName(); + if (para->isFullSize()) { + para->load(filename); + } + } +} + +void GradientMachine::randParameters() { + LOG(INFO) << "Initing parameters.."; + + for (auto& para : parameters_) { + if (para->isFullSize()) { + para->randomize(); + } + } + LOG(INFO) << "Init parameters done."; +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/GradientMachine.h b/paddle/legacy/gserver/gradientmachines/GradientMachine.h new file mode 100644 index 0000000000000000000000000000000000000000..d4f754a9f4dc3175f5000774c77a0e7334df7d85 --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/GradientMachine.h @@ -0,0 +1,250 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "ModelConfig.pb.h" +#include "TrainerConfig.pb.h" +#include "paddle/legacy/gserver/dataproviders/DataProvider.h" +#include "paddle/legacy/gserver/layers/Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/parameter/ParameterUpdaterBase.h" +#include "paddle/legacy/utils/Thread.h" + +#ifndef PADDLE_MOBILE_INFERENCE +#include "paddle/legacy/gserver/evaluators/Evaluator.h" +#endif + +namespace paddle { +/** + * @brief A gradient machine is capable of calculating some outputs given + * some inputs and performing gradient calculation based on the + * derivative from the outputs. + * + * A gradient machine can be either a full neural network or part of a neural + * network. + * + * Usage for training: + * + * 1. Prepare inArgs. Put your input data into inArgs[i].value. + * + * 2. Call forward(inArgs, &outArgs) + * + * 3. Calculate gradient with respect to outArgs[i]->value + * and fill them into outArgs[i]->grad. + * This step can be skipped if your the outputs are from cost layers. + * + * 4. Call backward(). After backward, gradient of each parameter is + * accumulated to getParameters()[i]->getBuf(PARAMETER_GRADIENT) + * + * 5. Update parameter value getParameters()[i]->getBuf(PARAMETER_VALUE) using + * gradients. + * + * 6. Clear gradients to zero. + * + * Usage for prediction: + * + * 1. Prepare inArgs. Put your input data into inArgs[i].value. + * + * 2. Call forward(inArgs, &outArgs) + * + * 3. Obtain the prediction result from outArgs[i] + */ + +typedef std::vector MachineState; + +class GradientMachine; + +typedef std::shared_ptr GradientMachinePtr; + +class GradientMachine { + public: + enum CreateMode { + kNormal = 0, + kSgdSparseCpuTraining = 3, + kTesting = 4, + kCustom = 10 + }; + + /** + * Create a gradient machine from ModelConfig + * Parameter will have parameterTypes + */ + static GradientMachine* create( + const ModelConfig& config, + int mode = kNormal, + const std::vector& parameterTypes = + std::vector{ + PARAMETER_VALUE, PARAMETER_GRADIENT, PARAMETER_MOMENTUM}); + + virtual ~GradientMachine() {} + + /** + * Prefetch row ids of sparse parameter. + */ + virtual void prefetch(const std::vector& inArgs) { (void)inArgs; } + + /** + * @brief Forward propagation. + * + * Calculate outputs (outArgs) based the inputs (inArgs) + * + * @note: if passType==PASS_TEST, then backward() should not be called + */ + virtual void forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType) = 0; + + /** + * @brief Backward propagation. + * + * Calculate the gradient of inArgs and parameter. + * + * This function should only be called after a corresponding forward() call. + * The caller is responsible for filling the correct grad for the outArgs + * obtained using forward(). + * + * It may also change the grad field for the inArgs supplied at forward() + */ + virtual void backward(const UpdateCallback& callback = nullptr) = 0; + + /** + * Combine forward() and backward(). For multithread training, this + * may be faster. + * + * @note: passType PASS_TEST is not allowed for forwardBackward(). + */ + virtual void forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback = nullptr) { + forward(inArgs, outArgs, passType); + backward(callback); + } + + virtual Argument getLayerOutput(const std::string& layerName) = 0; + + // see comment in Layer.h for the function with the same name + virtual void resetState() {} + + // set machine state + virtual void setState(const MachineState& machineState) {} + + // save machine state + virtual void getState(MachineState& machineState) {} + + virtual void onPassEnd() = 0; + +#ifndef PADDLE_MOBILE_INFERENCE + /** + * Create an evaluator which can be used for eval() + */ + virtual Evaluator* makeEvaluator() const = 0; + + /** + * evaluate using the given evaluator + */ + virtual void eval(Evaluator* evaluator) const = 0; +#endif + + std::vector& getParameters() { return parameters_; } + + std::vector& getNonStaticParameters() { + if (nonStaticParameters_.empty()) { + for (auto para : parameters_) { + if (!para->isStatic()) { + nonStaticParameters_.push_back(para); + } + } + } + return nonStaticParameters_; + } + + inline bool hasStaticParameters() { + return parameters_.size() != getNonStaticParameters().size(); + } + + /** + * @brief Used before formal training, start work-threads and set + * trainer Parameters; + * + * @note This function will only been implemented and used in a + * multithreaded environment. + */ + virtual void start() {} + + /** + * @brief check each work-thread whether is failed/error/finish, + * if not, return ture, and yes return false. + * + * @note This function will only been implemented and used in a + * multithreaded environment. + */ + virtual void finish() {} + + /** + * @brief set the training status a "finished" value, the sub_work_threads + * will option the change, and then exit. + * + * @note This function will only been implemented and used in a + * multithreaded environment. + */ + virtual bool trainIsOn() { return true; } + + /** + * @brief when all or some of the sub-workThreads are suspended to waiting + * controller's instructions, and after some processing done in the + * controller, it will call this function to wake up all the pending + * thread. + * + * @note This function will only been implemented and used in a + * multithreaded environment. + */ + virtual void restart() {} + + /// Set the gradient of the output from outside. + virtual void setOutputGrad(const std::vector& args) { + LOG(FATAL) << "Not implemented!"; + } + + void saveParameters(const std::string& dir) const; + + void loadParameters(const std::string& dir); + + void randParameters(); + + virtual void getStats(real& cost, int64_t& numProcessed) { + (void)cost; + (void)numProcessed; + } + + /** + * @brief Release the middle layer's output memory. + * + * @note This function is used for memory optimization in inference. + */ + virtual void releaseOutput() {} + + protected: + virtual void onLoadParameter() {} + + std::vector parameters_; + std::vector nonStaticParameters_; +}; + +} // namespace paddle diff --git a/paddle/gserver/gradientmachines/GradientMachineMode.cpp b/paddle/legacy/gserver/gradientmachines/GradientMachineMode.cpp similarity index 100% rename from paddle/gserver/gradientmachines/GradientMachineMode.cpp rename to paddle/legacy/gserver/gradientmachines/GradientMachineMode.cpp diff --git a/paddle/gserver/gradientmachines/GradientMachineMode.h b/paddle/legacy/gserver/gradientmachines/GradientMachineMode.h similarity index 100% rename from paddle/gserver/gradientmachines/GradientMachineMode.h rename to paddle/legacy/gserver/gradientmachines/GradientMachineMode.h diff --git a/paddle/legacy/gserver/gradientmachines/MultiGradientMachine.cpp b/paddle/legacy/gserver/gradientmachines/MultiGradientMachine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3ef0dfbfe2e5842918500a3b0706c1a55024ce46 --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/MultiGradientMachine.cpp @@ -0,0 +1,898 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MultiGradientMachine.h" + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/utils/Stat.h" + +#include "NeuralNetwork.h" +#include "ParallelNeuralNetwork.h" + +DEFINE_bool(allow_only_one_model_on_one_gpu, + true, + "If true, do not allow multiple models on one GPU device"); + +namespace paddle { + +// get types of the parameters which need to be merged after backward() +static void fillMergeTypes(PassType passType, + std::vector* mergeTypes) { + mergeTypes->clear(); + if (passType != PASS_TEST) { + mergeTypes->push_back(PARAMETER_GRADIENT); + } +} + +MultiGradientMachine::MultiGradientMachine(const ModelConfig& config, + bool useGpu) + : useGpu_(useGpu), + trainerBarrier_(FLAGS_trainer_count), + allBarrier_(FLAGS_trainer_count + 1), + inArgsCopied_(false) { + isPassGrad_ = false; + numThreads_ = FLAGS_trainer_count; + if (useGpu) { + //! TODO(yuyang18): When useGpu=false && paddle is not compiled with gpu, + //! the hl_get_device_count will get an error result. It seems should return + //! 0 when hppl is not compiled as gpu version. + numDevices_ = hl_get_device_count(); + } else { + numDevices_ = 0; + } + ParamInitCallback mainParamInitCb = [](int paramId, Parameter* para) { + // only create buf for CPU parameters + // GPU parameters will be created in each thread + if (para->useGpu()) return; + + if (para->isSparseRemoteUpdate()) { + para->enableType(PARAMETER_VALUE, + FLAGS_loadsave_parameters_in_pserver + ? Parameter::MAT_SPARSE_ROW_PREFETCH + : Parameter::MAT_SPARSE_ROW_PREFETCH_FULL_SIZE); + para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW); + } else if (para->isGradSparseUpdate()) { + para->enableType(PARAMETER_VALUE); + para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW_IDS); + SparseRowIdsCpuMatrix* mat = dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get()); + mat->setNumOfThreads(FLAGS_trainer_count); + } else if (para->isValueShared()) { + para->enableType(PARAMETER_VALUE, Parameter::MAT_VALUE_SHARED); + if (!para->isStatic()) { + para->enableType(PARAMETER_GRADIENT); + } + } else { + para->enableType(PARAMETER_VALUE); + if (!para->isStatic()) { + para->enableType(PARAMETER_GRADIENT); + } + } + }; + + NeuralNetwork* nn = NeuralNetwork::create(config); + nn->init(config, mainParamInitCb); + gradientMachine_.reset(nn); + parameters_ = gradientMachine_->getParameters(); + + numLogicalDevices_ = 0; + if (useGpu_) { + numLogicalDevices_ = 1; + + for (size_t pid = 0; pid < parameters_.size(); pid++) { + if (parameters_[pid]->getConfig().device() + 1 > numLogicalDevices_) { + numLogicalDevices_ = parameters_[pid]->getConfig().device() + 1; + } + } + LOG(INFO) << "numLogicalDevices=" << numLogicalDevices_ + << " numThreads=" << numThreads_ << " numDevices=" << numDevices_; + + if (numLogicalDevices_ * numThreads_ > numDevices_ && + FLAGS_allow_only_one_model_on_one_gpu) { + LOG(FATAL) << "trainer_count * num_devices_in_model " + << "(" << numThreads_ << "*" << numLogicalDevices_ << ")" + << "=" << numThreads_ * numLogicalDevices_ + << " exceeds number of GPU devices(" << numDevices_ << ")"; + } + numLogicalDevices_ = std::min(numLogicalDevices_, numDevices_); + + /* Enables direct access to memory allocations on a peer device */ + for (int i = 0; i < numThreads_; i++) { + for (int d = 0; d < numLogicalDevices_; ++d) { + enablePeerAccess(logicalDeviceId2RealDeviceId(d, i), + logicalDeviceId2RealDeviceId(d, i + 1)); + enablePeerAccess(logicalDeviceId2RealDeviceId(d, i), + logicalDeviceId2RealDeviceId(d, i - 1)); + } + } + } + + for (int i = 0; i < numThreads_; ++i) { + threads_.emplace_back(new TrainerThread(config, i, this)); + } + + bufferSizes_.resize(numLogicalDevices_, 0); + paraMainThread_.reserve(parameters_.size()); + int pid = 0; + for (auto& para : parameters_) { + if (para->isStatic() || !para->useGpu()) { + paraMainThread_.push_back(0); + } else { + int end = pid++ % numThreads_; + paraMainThread_.push_back(end); + int paraDeviceId = para->getDeviceId(); + if (paraDeviceId == -1) paraDeviceId = 0; + paraDeviceId = paraDeviceId % numLogicalDevices_; + if (para->getSize() > bufferSizes_[paraDeviceId]) { + bufferSizes_[paraDeviceId] = para->getSize(); + VLOG(1) << "bufferSize[" << paraDeviceId << "]" << para->getSize(); + } + } + } + + // TODO(xuwei06) Instead of using maximal buffer size, we may use a smaller + // fixed buffer size and use pipeline to dispatch parameter value and merge + // parameter gradient, which may be faster. + + // combination of all trainers mainPara into GradientMachine parameters + hasNonstaticCpuParamters_ = false; + for (size_t pid = 0; pid < parameters_.size(); pid++) { + if (parameters_[pid]->useGpu()) { + parameters_[pid] = threads_[paraMainThread_[pid]]->getParameters()[pid]; + } else if (!parameters_[pid]->isStatic()) { + hasNonstaticCpuParamters_ = true; + } + } + + gradBufs_.resize(numThreads_); + for (int i = 0; i < numThreads_; ++i) { + gradBufs_[i].resize(numLogicalDevices_); + for (int d = 0; d < numLogicalDevices_; ++d) { + gradBufs_[i][d].sem.post(); + } + } + + outArgStream_ = HPPL_STREAM_1; + + start(); +} + +void MultiGradientMachine::start() { + for (auto& thread : threads_) { + thread->start(); + } +} + +void MultiGradientMachine::finish() { + for (auto& thread : threads_) { + thread->stop(); + } +} + +std::vector*> +MultiGradientMachine::getSlaveParameters() { + std::vector*> vec; + vec.reserve(threads_.size()); + for (auto& thread : threads_) { + vec.push_back(&thread->getParameters()); + } + return vec; +} + +void MultiGradientMachine::notifyGradientTransfer(int paramId) { + gradQueue_.enqueue(paramId); +} + +void MultiGradientMachine::allocGradBufs() { + if (numLogicalDevices_ == 0) return; + if (gradBufs_[0][0].bufs.size() >= mergeTypes_.size()) return; + + for (int i = 0; i < numThreads_; i++) { + for (int d = 0; d < numLogicalDevices_; ++d) { + if (bufferSizes_[d] == 0) continue; + SetDevice device(logicalDeviceId2RealDeviceId(d, i)); + for (size_t j = 0; j < mergeTypes_.size(); j++) { + gradBufs_[i][d].bufs.push_back( + Vector::create(bufferSizes_[d], /* useGpu= */ true)); + } + } + } +} + +void MultiGradientMachine::prefetch(const std::vector& inArgs) { + // Each gradient machine in threads needs to do prefetch on its own + // part of inArgs. So we need to first divide inArgs to each thread + inArgs_ = inArgs; + startTask(TASK_COPY_IN_ARGS); + + for (auto& para : parameters_) { + if (para->isSparseRemoteUpdate()) { + auto mat = dynamic_cast( + para->getMat(PARAMETER_VALUE).get()); + mat->clearIndices(); + } + } + + waitForCopyInArgs(); + + // Because SparsePrefetchRowCpuMatrix can only be changed by ONE thread + // at one time, we need to do prefetch sequentially + for (auto& thread : threads_) { + thread->prefetch(); + } + + for (auto& para : parameters_) { + if (para->isSparseRemoteUpdate()) { + auto mat = dynamic_cast( + para->getMat(PARAMETER_VALUE).get()); + mat->setupIndices(); + auto matGrad = dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get()); + matGrad->reserveStore(); + } + } +} + +void MultiGradientMachine::forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType) { + forwardImp(inArgs, outArgs, passType, TASK_FORWARD); +} + +void MultiGradientMachine::forwardImp(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + TaskType taskType) { + updateThreadParameters(); + passType_ = passType; + + if (!inArgsCopied_) { + inArgs_ = inArgs; + inArgsCopied_ = false; + } + + fillMergeTypes(passType, &mergeTypes_); + allocGradBufs(); + startTask(taskType); + + getOutArgs(outArgs, passType); +} + +void MultiGradientMachine::backward(const UpdateCallback& callback) { + backwardCallback_ = callback; + startTask(TASK_BACKWARD); + backwardImp(callback); +} + +void MultiGradientMachine::forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback) { + backwardCallback_ = callback; + forwardImp(inArgs, outArgs, passType, TASK_FORWARD_BACKWARD); + backwardImp(callback); +} + +Argument MultiGradientMachine::getLayerOutput(const std::string& layerName) { + std::vector args; + args.reserve(threads_.size()); + + for (auto& thread : threads_) { + args.push_back(thread->getGradientMachine()->getLayerOutput(layerName)); + } + outLayerArgs_.concat(args, false /* use_gpu */, outArgStream_, passType_); + + return outLayerArgs_; +} + +void MultiGradientMachine::backwardImp(const UpdateCallback& callback) { + for (size_t i = 0; i < parameters_.size(); i++) { + if (!parameters_[i]->useGpu() || parameters_[i]->isStatic()) continue; + REGISTER_TIMER("controller_dequeue"); + gradQueue_.dequeue(); + } + if (hasNonstaticCpuParamters()) { + waitAfterMerge(); + if (backwardCallback_) { + for (auto& para : parameters_) { + if (!para->useGpu() && !para->isStatic()) { + backwardCallback_(para.get()); + } + } + } + } +} + +void MultiGradientMachine::updateThreadParameters() { + for (size_t pid = 0; pid < parameters_.size(); ++pid) { + if (!parameters_[pid]->useGpu()) continue; + if (!parameters_[pid]->isValueUpdated()) continue; + parameters_[pid]->clearValueUpdated(); + for (int i = 0; i < (int)threads_.size(); i++) { + threads_[i]->incUpdateCounter(); + } + // NotifyValueReady should happen after that all threads' incUpdateCounter() + // are called so that the counters are correct when notifyValueReady() + // is called. + threads_[paraMainThread_[pid]]->notifyValueReady(pid); + } +} + +void MultiGradientMachine::onPassEnd() { + for (auto& thread : threads_) { + thread->onPassEnd(); + } +} + +Evaluator* MultiGradientMachine::makeEvaluator() const { + return threads_[0]->getGradientMachine()->makeEvaluator(); +} + +void MultiGradientMachine::eval(Evaluator* evaluator) const { + for (auto& thread : threads_) { + SetDevice device(thread->getDeviceId()); + if (thread->hasInputData()) { + thread->getGradientMachine()->eval(evaluator); + } + } +} + +void MultiGradientMachine::getOutArgs(std::vector* outArgs, + PassType passType) { + for (auto& thread : threads_) { + REGISTER_TIMER("waitOutArgs"); + thread->waitOutArgsReady(); + } + + outArgs_.resize(threads_[threads_.size() - 1]->getOutArgs().size()); + + REGISTER_TIMER("copyOutArgs"); + for (size_t i = 0; i < outArgs_.size(); ++i) { + std::vector args; + args.reserve(threads_.size()); + for (auto& thread : threads_) { + // If the thread input is empty, then the output is empty. + auto tmp = thread->getOutArgs(); + if (tmp.size() > 0) { + args.push_back(tmp[i]); + } + } + outArgs_[i].concat(args, useGpu_, outArgStream_, passType); + } + + if (useGpu_) { + hl_stream_synchronize(outArgStream_); + } + + *outArgs = outArgs_; +} + +void MultiGradientMachine::setOutputGrad(const std::vector& args) { + CHECK_EQ(args.size(), outArgs_.size()); + for (size_t i = 0; i < args.size(); i++) { + outArgs_[i].grad = args[i].grad; + } +} + +void MultiGradientMachine::startTask(TaskType taskType) { + taskType_ = taskType; + for (auto& thread : threads_) { + thread->notifyTaskReady(); + } +} + +TrainerThread::TrainerThread(const ModelConfig& config, + int threadId, + MultiGradientMachine* multiMachine) + : multiMachine_(multiMachine), + config_(config), + threadId_(threadId), + inArgsCopied_(false) { + int numThreads = multiMachine->getNumThreads(); + + auto& mainParas = multiMachine->getParameters(); + + using std::placeholders::_1; + using std::placeholders::_2; + + partnerId_ = mod(threadId_ - 1, numThreads); + + deviceId_ = !multiMachine_->useGpu() + ? -1 + : multiMachine_->logicalDeviceId2RealDeviceId(0, threadId_); + SetDevice gpuDevice(deviceId_); + + NeuralNetwork* nn = nullptr; + if (!multiMachine->useGpu() || !FLAGS_parallel_nn) { + nn = NeuralNetwork::create(config); + } else { + nn = new ParallelNeuralNetwork(); + for (auto& paraConfig : *config_.mutable_parameters()) { + if (paraConfig.device() != -1) { + paraConfig.set_device(multiMachine_->logicalDeviceId2RealDeviceId( + paraConfig.device(), threadId_)); + } + } + for (auto& layerConfig : *config_.mutable_layers()) { + if (layerConfig.device() != -1) { + layerConfig.set_device(multiMachine_->logicalDeviceId2RealDeviceId( + layerConfig.device(), threadId_)); + } + } + } + // Only GPU do not share parameter values with main paramters. + ParamInitCallback slaveParamInitCb = + std::bind(parameterInitNN, _1, _2, &mainParas); + nn->init(config_, slaveParamInitCb); + gradientMachine_.reset(nn); + parameters_ = gradientMachine_->getParameters(); + if (!FLAGS_parallel_nn) { + for (auto& para : parameters_) { + para->setDevice(deviceId_); + } + } + + backwardCallback_ = + std::bind(&TrainerThread::backwardCallback, this, std::placeholders::_1); + + gradStream_ = HPPL_STREAM_2; + valueStream_ = HPPL_STREAM_3; + stopping_ = true; + updateCounter_ = 0; + parameterUpdated_ = false; +} + +TrainerThread::~TrainerThread() { stop(); } + +void TrainerThread::start() { + if (!stopping_) return; + + stopping_ = false; + + gradientMachine_->start(); + + computeThread_.reset(new std::thread([this]() { computeThread(); })); + + if (multiMachine_->useGpu()) { + gradCollectThread_.reset( + new std::thread([this]() { gradCollectThread(); })); + + valueDispatchThread_.reset( + new std::thread([this]() { valueDispatchThread(); })); + + copyThread_.reset(new std::thread([this]() { copyGradToBufferThread(); })); + } +} + +void TrainerThread::stop() { + if (stopping_) return; + + stopping_ = true; + + if (computeThread_) { + taskReadySem_.post(); + computeThread_->join(); + } + if (gradCollectThread_) { + gradQueue_.enqueue(0); + gradCollectThread_->join(); + } + if (copyThread_) { + gradBufQueue_.enqueue(0); + copyThread_->join(); + } + if (valueDispatchThread_) { + valueReadyQueue_.enqueue(0); + valueDispatchThread_->join(); + } +} + +void TrainerThread::computeThread() { + VLOG(1) << "gradComputeThread " << threadId_; + + if (deviceId_ >= 0) { + hl_init(deviceId_); + } + + while (true) { + { + REGISTER_TIMER("taskSem_wait"); + taskReadySem_.wait(); + } + + if (stopping_) break; + + switch (multiMachine_->getTaskType()) { + case MultiGradientMachine::TASK_FORWARD_BACKWARD: + forward(); + backward(); + break; + case MultiGradientMachine::TASK_FORWARD: + forward(); + break; + case MultiGradientMachine::TASK_BACKWARD: + backward(); + break; + case MultiGradientMachine::TASK_COPY_IN_ARGS: + batchSize_ = copyInArgs(); + inArgsCopied_ = true; + multiMachine_->waitForCopyInArgs(); + break; + } + } + hl_fini(); +} + +void TrainerThread::prefetch() { + SetDevice setDevice(deviceId_); + gradientMachine_->prefetch(inArgs_); +} + +void TrainerThread::forward() { + if (!inArgsCopied_) { + REGISTER_TIMER("copyInArgs"); + batchSize_ = copyInArgs(); + } else { + inArgsCopied_ = false; + } + + if (multiMachine_->getPassType() != PASS_TEST) { + REGISTER_TIMER("clearGradient"); + // For main parameter, the user of MultiGpuSyncMachine is responsible + // for setting the gradient to zero + for (size_t i = 0; i < parameters_.size(); i++) { + if (parameters_[i]->useGpu()) { + if (multiMachine_->paraMainThread(i) != threadId_) { + SetDevice device(parameters_[i]->getDeviceId()); + parameters_[i]->clearGradient(); + } + } else { + parameters_[i]->clearGradient(); + } + } + } + + { + REGISTER_TIMER("wait_value"); + valueReadyCond_.wait([this]() { return !parameterUpdated_; }); + } + + { fillMergeTypes(multiMachine_->getPassType(), &mergeTypes_); } + + { + REGISTER_TIMER("thread_forward"); + if (batchSize_ > 0) { + gradientMachine_->forward( + inArgs_, &outArgs_, multiMachine_->getPassType()); + } else { + outArgs_.clear(); + } + } + outArgsReadySem_.post(); +} + +void TrainerThread::backward() { + REGISTER_TIMER("thread_backward"); + if (multiMachine_->isPassGrad()) { + copyOutputGrad(); + } + if (batchSize_ > 0) { + gradientMachine_->backward(backwardCallback_); + } else { + for (size_t i = parameters_.size(); i > 0; i--) { + backwardCallback(parameters_[i - 1].get()); + } + } + if (multiMachine_->hasNonstaticCpuParamters()) { + mergeCpuGradients(); + } +} + +void TrainerThread::backwardCallback(Parameter* para) { + // CPU parameters are merged in the end + if (!para->useGpu() || para->isStatic()) return; + + int paramId = para->getID(); + if (multiMachine_->getNumThreads() == 1) { + // no need to do merge if there is only one thread + doCallback(paramId); + } else if (threadId_ == mod(multiMachine_->paraMainThread(paramId) - 1, + multiMachine_->getNumThreads())) { + notifyCopyGradToBuffer(paramId); + } else { + notifyGradientCollect(paramId); + } +} + +void TrainerThread::copyGradToBufferThread() { + VLOG(1) << "copyGradToBufferThread " << threadId_; + + if (deviceId_ >= 0) { + hl_init(deviceId_); + } + auto& partnerThread = multiMachine_->getThread(partnerId_); + auto& gradBufs = multiMachine_->getGradBuf(partnerId_); + + while (true) { + int pid = gradBufQueue_.dequeue(); + if (stopping_) break; + + int pdeviceId = multiMachine_->realDeviceId2LogicalDeviceId( + parameters_[pid]->getDeviceId(), threadId_); + + auto& gradBuf = gradBufs[pdeviceId]; + + { + REGISTER_TIMER("waitBufferReady"); + gradBuf.sem.wait(); + } + + { + REGISTER_TIMER("copyGradToBuffer"); + SetDevice setDevice(parameters_[pid]->getDeviceId()); + for (size_t i = 0; i < mergeTypes_.size(); ++i) { + gradBuf.bufs[i]->resize( + parameters_[pid]->getBuf(mergeTypes_[i])->getSize()); + gradBuf.bufs[i]->copyFrom(*parameters_[pid]->getBuf(mergeTypes_[i]), + gradStream_); + } + hl_stream_synchronize(gradStream_); + } + partnerThread->notifyGradientCollect(pid); + } + hl_fini(); +} + +void TrainerThread::gradCollectThread() { + VLOG(1) << "gradCollectThread " << threadId_; + + if (deviceId_ >= 0) { + hl_init(deviceId_); + } + + std::vector gradReadyCount(parameters_.size(), 0); + + auto& gradBufs = multiMachine_->getGradBuf(threadId_); + + while (true) { + int pid = gradQueue_.dequeue(); + if (stopping_) break; + + if (++gradReadyCount[pid] < 2) continue; + gradReadyCount[pid] = 0; + int pdeviceId = multiMachine_->realDeviceId2LogicalDeviceId( + parameters_[pid]->getDeviceId(), threadId_); + + auto& gradBuf = gradBufs[pdeviceId]; + + { + REGISTER_TIMER("mergeGrad"); + for (size_t i = 0; i < mergeTypes_.size(); ++i) { + ParameterType type = mergeTypes_[i]; + const VectorPtr& localGrad = parameters_[pid]->getBuf(type); + SetDevice setDevice(parameters_[pid]->getDeviceId()); + localGrad->add(*gradBuf.bufs[i]); + } + } + + gradBuf.sem.post(); + + if (multiMachine_->paraMainThread(pid) == threadId_) { + doCallback(pid); + } else { + notifyCopyGradToBuffer(pid); + } + } + hl_fini(); +} + +void TrainerThread::doCallback(int pid) { + REGISTER_TIMER("callback"); + auto& gpuThreads = multiMachine_->getAllThreads(); + if (multiMachine_->getBackwardCallback()) { + // The callback supplied by the user of MultiGradientMachine may handle + // the parameter update using the gradient. + multiMachine_->getBackwardCallback()(parameters_[pid].get()); + if (parameters_[pid]->isValueUpdated()) { + parameters_[pid]->clearValueUpdated(); + for (auto& thread : gpuThreads) { + thread->incUpdateCounter(); + } + notifyValueReady(pid); + } + } + multiMachine_->notifyGradientTransfer(pid); +} + +void TrainerThread::valueDispatchThread() { + VLOG(1) << "valueDispatchThread " << threadId_; + + if (deviceId_ >= 0) { + hl_init(deviceId_); + } + + auto& thread = multiMachine_->getThread(partnerId_); + + while (true) { + int pid; + { + REGISTER_TIMER("value_dequeue"); + pid = valueReadyQueue_.dequeue(); + } + if (stopping_) break; + + if (multiMachine_->paraMainThread(pid) == partnerId_) continue; + + { + REGISTER_TIMER("copyValue"); + SetDevice setDevice(parameters_[pid]->getDeviceId()); + thread->getValueBuf(pid)->copyFrom(*getValueBuf(pid), valueStream_); + hl_stream_synchronize(valueStream_); + } + + thread->notifyValueReady(pid); + } + hl_fini(); +} + +void TrainerThread::notifyValueReady(int paramId) { + if (--updateCounter_ == 0) { + valueReadyCond_.notify_all([this] { parameterUpdated_ = false; }); + } + + notifyValueDispatch(paramId); +} + +int TrainerThread::copyInArgs() { + const std::vector& fullInArgs = multiMachine_->getInArgs(); + int numThreads = multiMachine_->getAllThreads().size(); + int32_t numSequences = fullInArgs[0].getNumSequences(); + int32_t startSeq = numSequences * threadId_ / numThreads; + int32_t endSeq = numSequences * (threadId_ + 1) / numThreads; + int32_t copySize = endSeq - startSeq; + + /** + * For the first copy, need to allocate space here + */ + if (inArgs_.size() == 0) { + inArgs_.resize(fullInArgs.size()); + } + + if (copySize == 0) { + return 0; + } + + for (size_t i = 0; i < fullInArgs.size(); i++) { + inArgs_[i].resizeAndCopyFrom( + fullInArgs[i], + startSeq, + copySize, + FLAGS_parallel_nn ? false : multiMachine_->useGpu()); + } + return copySize; +} + +void TrainerThread::mergeCpuGradients() { + CHECK_EQ(mergeTypes_.size(), 1UL); + CHECK_EQ(mergeTypes_[0], PARAMETER_GRADIENT); + + { + REGISTER_TIMER("waitbeforeMerge"); + multiMachine_->waitBeforeMerge(); + } + std::vector*> slaveParameters = + multiMachine_->getSlaveParameters(); + + CHECK(slaveParameters.size()); + for (auto& para : multiMachine_->getNonStaticParameters()) { + if (para->useGpu()) continue; + if (para->isSparseRemoteUpdate()) { + REGISTER_TIMER("mergeRemoteGradSparse"); + mergeGradSparseRemote(para.get(), slaveParameters); + } else if (para->isGradSparseUpdate()) { + REGISTER_TIMER("mergeGradSparse"); + mergeGradSparse(para.get(), slaveParameters); + } else { + REGISTER_TIMER("mergeGradDense"); + mergeGradDense(para.get(), slaveParameters); + } + } + { + REGISTER_TIMER("waitbeforeMerge"); + multiMachine_->waitAfterMerge(); + } +} + +void TrainerThread::mergeGradSparse( + Parameter* para, + std::vector*>& slaveParameters) { + size_t pid = para->getID(); + SparseRowIdsCpuMatrix* mainMat = dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get()); + std::vector& ids = mainMat->getIds(threadId_); + + for (auto slaveParams : slaveParameters) { + SparseRowCpuMatrix* mat = dynamic_cast( + (*slaveParams)[pid]->getMat(PARAMETER_GRADIENT).get()); + mat->addTo(*mainMat, ids, threadId_, multiMachine_->getNumThreads()); + // we use a sample hash method(%) instead of range partition, + // because range partition has balance issue sometimes, + // when feature ids are not generated from hashcode. + } + uniqueIds(ids); +} + +void TrainerThread::mergeGradSparseRemote( + Parameter* para, + std::vector*>& slaveParameters) { + size_t pid = para->getID(); + SparseRowCpuMatrix* mainMat = + dynamic_cast(para->getMat(PARAMETER_GRADIENT).get()); + + mainMat->checkIndices(); + mainMat->zeroMemThread(threadId_, multiMachine_->getNumThreads()); + + for (auto slaveParams : slaveParameters) { + SparseRowCpuMatrix* mat = dynamic_cast( + (*slaveParams)[pid]->getMat(PARAMETER_GRADIENT).get()); + mat->addTo(*mainMat, threadId_, multiMachine_->getNumThreads()); + } +} + +void TrainerThread::mergeGradDense( + Parameter* para, + std::vector*>& slaveParameters) { + size_t pid = para->getID(); + auto interval = calcSplitArrayInterval(para->getSize(), + (size_t)threadId_, + multiMachine_->getNumThreads(), + 8LU /*for avx*/); + size_t startSeq = interval.first; + size_t copySize = interval.second - interval.first; + + // setup sub bufs + CpuVector destGrad(0, nullptr); + destGrad.subVecFrom(*para->getBuf(PARAMETER_GRADIENT), startSeq, copySize); + + // merge + CpuVector slaveGradSub(0, nullptr); + for (auto slaveParams : slaveParameters) { + slaveGradSub.subVecFrom( + *(*slaveParams)[pid]->getBuf(PARAMETER_GRADIENT), startSeq, copySize); + destGrad.add(slaveGradSub); + } +} + +void TrainerThread::copyOutputGrad() { + const std::vector& outputGradArgs = multiMachine_->outArgs_; + int numThreads = multiMachine_->getAllThreads().size(); + int32_t numSequences = outputGradArgs[0].getNumSequences(); + int32_t startSeq = numSequences * threadId_ / numThreads; + int32_t endSeq = numSequences * (threadId_ + 1) / numThreads; + int32_t copySize = endSeq - startSeq; + outArgs_.resize(outputGradArgs.size()); + for (size_t i = 0; i < outputGradArgs.size(); i++) { + outArgs_[i].resizeAndCopyFrom(outputGradArgs[i], + startSeq, + copySize, + multiMachine_->useGpu(), + HPPL_STREAM_DEFAULT); + } + if (multiMachine_->useGpu()) { + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + } + gradientMachine_->setOutputGrad(outArgs_); +} +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/MultiGradientMachine.h b/paddle/legacy/gserver/gradientmachines/MultiGradientMachine.h new file mode 100644 index 0000000000000000000000000000000000000000..674acd4124981face13b21aee02f031ea775ffec --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/MultiGradientMachine.h @@ -0,0 +1,478 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include "GradientMachine.h" + +#include "hl_gpu.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Queue.h" + +namespace paddle { + +class TrainerThread; + +typedef Queue PidQueue; +typedef std::unique_ptr TrainerThreadPtr; + +struct GradBuffer { + /// GradBuffer is used for gathering gradient for GPU parameters + int paramId; + + /// sem is used to notify that the local gradient merge of the current thread + /// finished for the current thread. + Semaphore sem; + + // bufs[mergeIndex] + std::vector bufs; +}; + +/** + * A MultiGradientMachine is a synchronous GradientMachine which devides + * one data batch into several smaller batches and assign each one small batch + * to one computint thread for computation. After each thread finishes + * computation, it merges result (including output Argument and gradient during + * backward()). It basically is the same as single thread gradient machine, + * except that it uses multi-thread to do the computation. + * + * It handles GPU and Cpu parameters differently. In GPU, one computing thread + * generally corresponds to one GPU device. Thus, each thread keeps a separate + * copy of the parameter in its own device's memory. In CPU, we only need to + keep + * one copy of the parameters in the main memory. After, each computing thread + * computes its own parameter gradient, the update process needs to accumulate + * the parameter gradients from all the computing threads, and update the + * accumulated parameter gradient to the corresponding parameter value. + * + * Each GPU parameter is assigned to a thread called its main thread. For each + * parameter, the accumulation of its gradients and the update of its value + * happens in its main thread. The main thread first gather the parameter + * gradients from all the computing thread. Then, it performs parameter update. + * After a gradient is updated by the main thread, it is scattered to all the + * computing thread so that the parameters in all the computing threads are + * synchronized. The scatter and gather process are implemented by ring-style + * communication. Assume we have N computing threads, its thread ids will be + * 0, 1, ..., N-1. For each parameter, the id of the main thread is specified + in + * paraMainThread_[pid], where pid is the id of the parameter. Each thread i + only + * sends data to its partner thread (i - 1) % N. For example, for a parameter + * gradient that is computed in thread 4, and its main thread is 2. Its + * traveling process would be 4, 5,..., N-1, 0, 1, 2. In each step, the + gradient + * buffer is added to the local gradient, and the local gradient is then copied + * to the gradient buffer of the next thread. At last, its main thread 2 will + * get the accumulated parameter gradient. For the same parameter, after its + * value is updated, the value's traveling process would be 2, 1, 0, N-1, ... + 3. + * At the end, all the computing threads would have the updated parameter + value. + * + * A computing thread (TrainerThread) uses 4 threads to do different jobs: + * + * 1. computeThread(): performing forward(), backward(), prefetch(). + * + * 2. valueDispatchThread(): copying parameter values to partner thread. + * + * 3. copyGradToBufferThread(): copying parameter gradient to partner thread. + * + * 4. gradCollectThread(): merging the gradient from step 3 with local gradient + * and call the callback supplied by the user to update parameter value. + * + * CPU parameter value has only one copy. And their gradients are merged at the + * end of backward(). + * + * * Handling of sparse update + * Currently, sparse update is only supported for CPU parameters. + + * Sparse updates refers to gradient caculation where the gradient is sparse. + For + * example, if the input argument to a 'fc' layer is sparse, the gradient of + the + * weight matrix of this layer will be sparse. It is usually more efficient to + * treat the gradient explicitly as sparse vector during the parameter update. + + * There are two types of sparse updates called local sparse update and remote + * sparse update. + + * For both types of sparse updates, there is one copy of parameter value and + * gradient called main parameter value and gradient, and there is a copy of + * parameter value and gradient for each computing thread called slave + parameter + * value and gradient. The slave parameter values are always shared with the + * corresponding main parameter value. The slave parameter grad is a sparse row + * matrix. The sparse pattern for slave parameter grads are different, because + * the small batches for each computing thread might have different sparsity + * pattern. + + * 1. Local sparse update + * + * Main parameter value type is MAT_NORMAL. It is a dense matrix. + * + * Main parameter grad type is MAT_SPARSE_ROW_IDS (SparseRowIdsCpuMatrix) + * It is also a dense matrix, but the updated values are specified by IDS. + * + * Slave parameter value shares with main parameter value. + * + * Slave parameter grad type is MAT_SPARSE_ROW_AUTO_GROW + * (SparseAutoGrowRowCpuMatrix). It is a sparse row matrix. + * + * During backward() of each TrainerThread, SparseAutoGrowRowCpuMatrix will + * gather all the non-zero gradient. And After backward(), they will be + merged + * into main parameter grad (SparseRowIdsCpuMatrix), with indices indicating + * which rows have nonzero gradient. + * + * 2. Remote sparse update + * + * Main parameter value type is MAT_SPARSE_ROW_PREFETCH(_FULL_SIZE) + * (SparsePrefetchRowCpuMatrix). MAT_SPARSE_ROW_PREFETCH is a sparse matrix. + * MAT_SPARSE_ROW_PREFETCH_FULL_SIZE is a dense matrix. However, only the + * parameter values that are prefetched is up-to-date. + * + * Main parameter grad type is MAT_SPARSE_ROW (SparseRowCpuMatrix). + * And it shares sparse pattern with value by sharing indexDictHandle_, + which + * is an internal data structure used by SparseRowCpuMatrixto specify the + * sparsity pattern of Slave parameter value shares with main parameter + value. + * + * Slave parameter grad type is MAT_SPARSE_ROW_AUTO_GROW + * (SparsePrefetchRowCpuMatrix). It is a sparse row matrix + * + * During prefetch(), all the layers will indicates which rows of each + * parameter are needed. Then the framework will retrieve those rows from + * parameter server. + * + * During backward() of each TrainerThread, SparseAutoGrowRowCpuMatrix will + * gather all the non-zero gradient. And After backward(), they will be + merged + * into main parameter grad (SparseRowCpuMatrix). And the framework will + send + * the merged gradient to parameter server. + */ +class MultiGradientMachine : public GradientMachine { + public: + enum TaskType { + TASK_FORWARD_BACKWARD = 0, + TASK_FORWARD = 1, + TASK_BACKWARD = 2, + TASK_COPY_IN_ARGS = 3, + }; + + explicit MultiGradientMachine(const ModelConfig& config, bool useGpu); + + virtual void start(); + + virtual void finish(); + + virtual void prefetch(const std::vector& inArgs); + + virtual void forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType); + + virtual void backward(const UpdateCallback& callback = nullptr); + + void forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback); + + virtual Argument getLayerOutput(const std::string& layerName); + + virtual void onPassEnd(); + + virtual Evaluator* makeEvaluator() const; + + virtual void eval(Evaluator* evaluator) const; + + bool useGpu() const { return useGpu_; } + + /// @return whether to pass the gradients in outArgs_ to each threads. + bool isPassGrad() { return isPassGrad_; } + + /// @brief set whether to pass the gradient in outArgs_ to each threads. + void setPassGrad(bool isPass) { isPassGrad_ = isPass; } + + /// Set the gradients of the outputs. + /// The gradietns will be copied to each thread in the computing threads. + virtual void setOutputGrad(const std::vector& args); + + protected: + friend class TrainerThread; + + std::vector& getAllThreads() { return threads_; } + /// Calculate the real device id based on the logical device id and the + /// thread id. + int logicalDeviceId2RealDeviceId(int logicalId, int threadId = 0) const { + if (logicalId == -1) { + logicalId = 0; + } + return mod(logicalId + FLAGS_gpu_id + threadId * numLogicalDevices_, + numDevices_); + } + + /// Calculate the logical device id based on the real device id and the + /// thread id. + int realDeviceId2LogicalDeviceId(int realId, int threadId = 0) const { + if (realId == -1) { + return 0; + } else { + return mod(realId - FLAGS_gpu_id - threadId * numLogicalDevices_, + numDevices_); + } + } + + std::vector*> getSlaveParameters(); + + bool hasNonstaticCpuParamters() const { return hasNonstaticCpuParamters_; } + + /// Called TrainerThread to wait before merging CPU parameter gradients. + void waitBeforeMerge() { trainerBarrier_.wait(); } + + /// called by MultiGradientMachine and TrainerThread to wait after merging + /// CPU parameter graidents. + void waitAfterMerge() { allBarrier_.wait(); } + + /// called by MultiGradientMachine and TrainerThread to wait for copyInArgs() + /// finishing + void waitForCopyInArgs() { allBarrier_.wait(); } + + TrainerThreadPtr& getThread(int threadId) { return threads_[threadId]; } + + std::vector& getGradBuf(int threadId) { + return gradBufs_[threadId]; + } + + PassType getPassType() const { return passType_; } + + /// Called by TrainerThread to notify MultiGradientMachine that the gradient + /// for paramId is ready + void notifyGradientTransfer(int paramId); + + const std::vector& getInArgs() { return inArgs_; } + + TaskType getTaskType() const { return taskType_; } + + const UpdateCallback& getBackwardCallback() const { + return backwardCallback_; + } + + int getNumDevices() const { return numDevices_; } + + int getNumLogicalDevices() const { return numLogicalDevices_; } + + int getNumThreads() const { return numThreads_; } + + int paraMainThread(int pid) const { return paraMainThread_[pid]; } + + protected: + virtual void forwardImp(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + TaskType taskType); + + virtual void backwardImp(const UpdateCallback& callback = NULL); + + /// update all parameters + void updateThreadParameters(); + + void startTask(TaskType taskType); + + void getOutArgs(std::vector* outArgs, PassType passType); + + void allocGradBufs(); + + protected: + bool useGpu_; + + bool hasNonstaticCpuParamters_; + + /// store main parameter only + std::unique_ptr gradientMachine_; + + std::vector threads_; + std::vector paraMainThread_; + std::vector> gradBufs_; // [threadId][deviceId] + std::vector bufferSizes_; + + PassType passType_; + TaskType taskType_; + PidQueue gradQueue_; + std::vector inArgs_; + std::vector outArgs_; + hl_stream_t outArgStream_; + + Argument outLayerArgs_; + + /// ParameterType which needs to be merged from each GPU + std::vector mergeTypes_; + int numDevices_; /* number of gpu devices */ + int numLogicalDevices_; // number of GPU used by one NN + int numThreads_; /* number of train threads */ + + UpdateCallback backwardCallback_; + + /// barrrier for threads_ + ThreadBarrier trainerBarrier_; + + /// barrier for both MultiGradientMachine and threds_ + ThreadBarrier allBarrier_; + + /// indicate whether inArgs is copied before forward() + bool inArgsCopied_; + + /// Whether to copy the gradient back from an external input. + bool isPassGrad_; +}; + +class TrainerThread { + public: + TrainerThread(const ModelConfig& config, + int threadId, + MultiGradientMachine* multiMachine); + + ~TrainerThread(); + + void start(); + + void onPassEnd() { gradientMachine_->onPassEnd(); } + + void waitOutArgsReady() { outArgsReadySem_.wait(); } + + void notifyTaskReady() { taskReadySem_.post(); } + + int getDeviceId() const { return deviceId_; } + + GradientMachine* getGradientMachine() { return gradientMachine_.get(); } + + const std::vector& getParameters() { return parameters_; } + + void stop(); + + void notifyValueReady(int paramId); + + const VectorPtr& getValueBuf(int paramId) { + return parameters_[paramId]->getBuf(PARAMETER_VALUE); + } + + const std::vector& getOutArgs() { return outArgs_; } + + void incUpdateCounter(int n = 1) { + updateCounter_ += n; + parameterUpdated_ = true; + } + + void notifyGradientCollect(int paramId) { gradQueue_.enqueue(paramId); } + + void notifyCopyGradToBuffer(int paramId) { gradBufQueue_.enqueue(paramId); } + + void notifyValueDispatch(int paramId) { valueReadyQueue_.enqueue(paramId); } + + void prefetch(); + + /// copy the output gradient from the main GradientMachine. + void copyOutputGrad(); + + /// Whether the thread has input data. + bool hasInputData() { return batchSize_ != 0; } + + protected: + void mergeCpuGradients(); + + void mergeGradSparse( + Parameter* para, + std::vector*>& slaveParameters); + + void mergeGradSparseRemote( + Parameter* para, + std::vector*>& slaveParameters); + + void mergeGradDense( + Parameter* para, + std::vector*>& slaveParameters); + + void computeThread(); + void valueDispatchThread(); + void copyGradToBufferThread(); + void gradCollectThread(); + + int copyInArgs(); + void forward(); + void backward(); + void backwardCallback(Parameter* para); + + /// call the actuall callback supplied by the caller of + /// GradientMachine::backward + void doCallback(int pid); + + protected: + MultiGradientMachine* multiMachine_; + ModelConfig config_; + /// whether the thread should stop + bool stopping_; + /// the threads form which to collect gradient + int partnerId_; + /// from 0 to threads-1 + int threadId_; + int deviceId_; + std::unique_ptr gradientMachine_; + std::vector parameters_; + + /// ParameterType which needs to be merged from each GPU + std::vector mergeTypes_; + + /// compute thread + std::unique_ptr computeThread_; + std::vector inArgs_; + std::vector outArgs_; + Semaphore taskReadySem_; + Semaphore outArgsReadySem_; + + /// copy thread + std::unique_ptr copyThread_; + /// queue of gradient needs to be copied to partner + PidQueue gradBufQueue_; + hl_stream_t gradStream_; + + /// grad merge thread + std::unique_ptr gradCollectThread_; + /// queue of gradient needs to be merged with gradient coopied by + /// copyGradToBufferThread + PidQueue gradQueue_; + UpdateCallback backwardCallback_; + + /// value dispatch thread + std::unique_ptr valueDispatchThread_; + /// queue of the parameter whose the vale are ready for copy + PidQueue valueReadyQueue_; + + /// used to notify all the parameter values are ready + LockedCondition valueReadyCond_; + + hl_stream_t valueStream_; + /// how many parameters are updated + std::atomic updateCounter_; + bool parameterUpdated_; + + /// indicate whether inArgs is copied before forward() + bool inArgsCopied_; + int batchSize_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/MultiNetwork.cpp b/paddle/legacy/gserver/gradientmachines/MultiNetwork.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1245c441036a601025192ab23a6d2899b688a9dc --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/MultiNetwork.cpp @@ -0,0 +1,185 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +#include "MultiNetwork.h" + +#include "NeuralNetwork.h" +#include "ParallelNeuralNetwork.h" + +namespace paddle { + +void MultiNetwork::init(const ModelConfig& config, + ParamInitCallback callback, + const std::vector& parameterTypes, + bool useGpu) { + CHECK_GT(config.sub_models_size(), 1) << "sub_models_size should GT 1"; + // check submodel[0] is root + CHECK_EQ("root", config.sub_models(0).name()) + << "sub_models(0) should be root"; + // ignore root + subNetworks_.resize(config.sub_models_size() - 1); + // base class + NeuralNetwork::init(config, callback, parameterTypes, useGpu); + // sub networks + for (int i = 1; i < config.sub_models_size(); ++i) { + std::string subModelName = config.sub_models(i).name(); + if (FLAGS_parallel_nn) { + subNetworks_[i - 1] = std::unique_ptr( + new ParallelNeuralNetwork(subModelName, this)); + } else { + subNetworks_[i - 1] = std::unique_ptr( + NeuralNetwork::newNeuralNetwork(subModelName, this)); + } + subNetworks_[i - 1]->init(config); + } +} + +void MultiNetwork::prefetch(const std::vector& inArgs) { + std::vector> argumentGroups; + Argument::splitByDataId(inArgs, &argumentGroups); + // check group size is equal to sub network size + CHECK_EQ(argumentGroups.size(), subNetworks_.size()); + for (size_t i = 0; i < subNetworks_.size(); i++) { + if (argumentGroups[i].size() == 1 && argumentGroups[i][0].dataId == -1) { + // check input args: if dataId is -1, then skip this sub network + continue; + } + subNetworks_[i]->prefetch(argumentGroups[i]); + } +} + +void MultiNetwork::forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType) { + // split inArgs to several vectors + std::vector> argumentGroups; + Argument::splitByDataId(inArgs, &argumentGroups); + + // check group size is equal to sub network size + CHECK_EQ(argumentGroups.size(), subNetworks_.size()); + std::vector tempOutArgs; + outArgs->clear(); + + for (size_t i = 0; i < subNetworks_.size(); i++) { + tempOutArgs.clear(); + if (argumentGroups[i].size() == 1 && argumentGroups[i][0].dataId == -1) { + // check input args: if dataId is -1, then skip this sub network + continue; + } + subNetworks_[i]->forward(argumentGroups[i], &tempOutArgs, passType); + for (const auto& elem : tempOutArgs) { + outArgs->push_back(elem); + outArgs->back().dataId = i; + } + } +} + +void MultiNetwork::backward(const UpdateCallback& callback) { + for (size_t i = 0; i < subNetworks_.size(); i++) { + subNetworks_[i]->backward(callback); + } +} + +void MultiNetwork::forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback) { + forward(inArgs, outArgs, passType); + backward(callback); +} + +void MultiNetwork::onPassEnd() { + for (size_t i = 0; i < subNetworks_.size(); i++) { + subNetworks_[i]->onPassEnd(); + } +} + +void MultiNetwork::start() { + for (auto& subNetwork : subNetworks_) { + subNetwork->start(); + } +} + +void MultiNetwork::finish() { + for (size_t i = 0; i < subNetworks_.size(); i++) { + subNetworks_[i]->finish(); + } +} + +class MultiCombinedEvaluator : public Evaluator { + public: + MultiCombinedEvaluator() {} + void addEvaluator(std::unique_ptr&& evaluator) { + evaluators_.emplace_back(std::move(evaluator)); + } + virtual void start() { + for (auto& evaluator : evaluators_) { + evaluator->start(); + } + } + + virtual void finish() { + for (auto& evaluator : evaluators_) { + evaluator->finish(); + } + } + + virtual void eval(const NeuralNetwork& nn) { + const MultiNetwork& multiNetwork = dynamic_cast(nn); + CHECK_EQ(evaluators_.size(), multiNetwork.getSubNetworks().size()); + int size = evaluators_.size(); + for (int i = 0; i < size; i++) { + // one evaluator for one subNetwork + evaluators_[i]->eval(*multiNetwork.getSubNetworks()[i]); + } + } + + virtual real evalImp(std::vector& arguments) { + (void)arguments; + return -1; + } + + virtual void printStats(std::ostream& os) const { + for (auto& evaluator : evaluators_) { + evaluator->printStats(os); + os << ' '; + } + } + + virtual void distributeEval(ParameterClient2* client) { + for (auto& evaluator : evaluators_) { + evaluator->distributeEval(client); + } + } + + protected: + std::vector> evaluators_; +}; + +Evaluator* MultiNetwork::makeEvaluator() const { + MultiCombinedEvaluator* multiCombinedEvaluator = new MultiCombinedEvaluator(); + for (size_t i = 0; i < subNetworks_.size(); i++) { + std::unique_ptr evaluator(subNetworks_[i]->makeEvaluator()); + multiCombinedEvaluator->addEvaluator(std::move(evaluator)); + } + return multiCombinedEvaluator; +} + +void MultiNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); } + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/MultiNetwork.h b/paddle/legacy/gserver/gradientmachines/MultiNetwork.h new file mode 100644 index 0000000000000000000000000000000000000000..afe15cb020ebe3bbe051800a72562c9543f3faa4 --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/MultiNetwork.h @@ -0,0 +1,64 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "GradientMachine.h" +#include "NeuralNetwork.h" + +#include "paddle/legacy/utils/Locks.h" + +namespace paddle { + +class MultiNetwork : public NeuralNetwork { + public: + explicit MultiNetwork(std::string subModelName = "") + : NeuralNetwork(subModelName) {} + + virtual void init(const ModelConfig& config, + ParamInitCallback callback, + const std::vector& parameterTypes, + bool useGpu); + + virtual void prefetch(const std::vector& inArgs); + + virtual void forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType); + + virtual void backward(const UpdateCallback& callback = nullptr); + + void forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback); + + virtual void onPassEnd(); + + virtual Evaluator* makeEvaluator() const; + + virtual void eval(Evaluator* evaluator) const; + + const std::vector>& getSubNetworks() const { + return subNetworks_; + } + + virtual void start(); + + virtual void finish(); + + protected: + std::vector> subNetworks_; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/legacy/gserver/gradientmachines/NeuralNetwork.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0f8048152ff317a1e445249fa7093158d2d4a5c5 --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/NeuralNetwork.cpp @@ -0,0 +1,548 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Util.h" + +#include "NeuralNetwork.h" +#include "hl_gpu.h" +#include "paddle/legacy/utils/CustomStackTrace.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/legacy/gserver/layers/MKLDNNLayer.h" +#endif + +#ifndef PADDLE_MOBILE_INFERENCE +#include "MultiNetwork.h" +#include "RecurrentGradientMachine.h" +#include "paddle/legacy/gserver/layers/AgentLayer.h" +#endif + +namespace paddle { +void parameterInitNN(int paramId, + Parameter* para, + std::vector* sharedParams) { + // Create parameters values. + if (!para->useGpu() && sharedParams) { + para->enableSharedType(PARAMETER_VALUE, + (*sharedParams)[paramId]->getBuf(PARAMETER_VALUE), + (*sharedParams)[paramId]->getMat(PARAMETER_VALUE)); + } else { + if (para->isSparseRemoteUpdate()) { + para->enableType(PARAMETER_VALUE, + FLAGS_loadsave_parameters_in_pserver + ? Parameter::MAT_SPARSE_ROW_PREFETCH + : Parameter::MAT_SPARSE_ROW_PREFETCH_FULL_SIZE); + } else { + para->enableType(PARAMETER_VALUE); + } + } + // Create parameter gradients. + if (para->isSparseRemoteUpdate() && !sharedParams) { + para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW); + } else if (para->isGradSparseUpdate()) { + para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW_AUTO_GROW); + } else if (!para->isStatic()) { + para->enableType(PARAMETER_GRADIENT); + } +} + +NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) { +#ifndef PADDLE_MOBILE_INFERENCE + if (config.type() == "recurrent_nn") { + return newNeuralNetwork("root"); + } else if (config.type() == "multi_nn") { + return new MultiNetwork("root"); + } else { + return newNeuralNetwork(); + } +#else + return new NeuralNetwork(); +#endif +} + +std::map NeuralNetwork::dllInitMap; + +void NeuralNetwork::init(const ModelConfig& config, + ParamInitCallback callback, + const std::vector& parameterTypes, + bool useGpu) { + using std::placeholders::_1; + using std::placeholders::_2; + ParamInitCallback paramCallback = nullptr; + if (callback != nullptr) { + paramSelfInited_ = false; + paramCallback = callback; + } else { + paramSelfInited_ = true; + paramCallback = std::bind(parameterInitNN, _1, _2, nullptr); + } + config_ = config; + + if (rootNetwork_ != nullptr) { + // direct use parameters_ and parameterMap_ from base network + CHECK_EQ((size_t)config.parameters_size(), + rootNetwork_->getParameters().size()); + parameters_ = rootNetwork_->getParameters(); + parameterMap_ = *(rootNetwork_->getParameterMap()); + } else { + parameters_.reserve(config.parameters_size()); + for (const auto& para_config : config.parameters()) { + auto parameter = std::make_shared(para_config, + useGpu, + /*initialize=*/false); + paramCallback(parameters_.size(), parameter.get()); + if (!callback) { + for (ParameterType type : + (parameter->isStatic() + ? std::vector{PARAMETER_VALUE} + : parameterTypes)) { + if (type != PARAMETER_VALUE && type != PARAMETER_GRADIENT) { + parameter->enableType(type); + } + } + } + parameter->setID(parameters_.size()); + parameters_.push_back(parameter); + CHECK(!parameterMap_.count(parameter->getName())); + parameterMap_[parameter->getName()] = parameter; + } + } + + auto layerCreate = [&](const LayerConfig& layer_config) { + auto layer = Layer::create(layer_config); + CHECK(layer) << "Create layer failed. Layer name:" << layer->getName(); + layers_.push_back(layer); + CHECK(!layerMap_.count(layer->getName())); + layerMap_[layer->getName()] = layer; + }; + + auto subModelConfig = std::find_if(config.sub_models().begin(), + config.sub_models().end(), + [=](const SubModelConfig& sub_model) { + return sub_model.name() == subModelName_; + }); + bool useSubModel = (subModelConfig != config.sub_models().end()); + CHECK_EQ(useSubModel, !subModelName_.empty()); + if (useSubModel) { + layers_.reserve(subModelConfig->layer_names_size()); + for (const auto& layer_name : subModelConfig->layer_names()) { + auto layer_config = + std::find_if(config.layers().begin(), + config.layers().end(), + [=](const LayerConfig& layer_config) { + return layer_config.name() == layer_name; + }); + CHECK(layer_config != config.layers().end()); + layerCreate(*layer_config); + } + } else { + layers_.reserve(config.layers_size()); + for (const auto& layer_config : config.layers()) { + bool useLayer = true; + if (config.has_external_config()) { + useLayer = true; + for (const auto& name : config.external_config().layer_names()) { + if (layer_config.name() == name) { + useLayer = false; + break; + } + } + } + if (useLayer) { + layerCreate(layer_config); + } + } + } + + for (const auto& layer : layers_) { + layer->init(layerMap_, parameterMap_); + layer->initSubNetwork(this /*root*/, config_, parameterTypes, useGpu); + } + + for (const auto& layer_name : + (useSubModel ? subModelConfig->input_layer_names() + : config.input_layer_names())) { + auto it = layerMap_.find(layer_name); + CHECK(it != layerMap_.end()); + dataLayers_.push_back(std::dynamic_pointer_cast(it->second)); + } + + for (const auto& layer_name : + (useSubModel ? subModelConfig->output_layer_names() + : config.output_layer_names())) { + auto it = layerMap_.find(layer_name); + CHECK(it != layerMap_.end()); + outputLayers_.push_back(it->second); + } + + for (const auto& layer : layers_) { + const auto& name = layer->getName(); + bool isMiddleLayer = true; + + // if data layer + for (const auto& dataLayer : dataLayers_) { + if (name == dataLayer->getName()) { + isMiddleLayer = false; + break; + } + } + + // if output layer + for (const auto& dataLayer : outputLayers_) { + if (name == dataLayer->getName()) { + isMiddleLayer = false; + break; + } + } + + if (isMiddleLayer) { + middleLayers_.push_back(layer); + } + } +} + +void NeuralNetwork::connect(LayerPtr agentLayer, + LayerPtr realLayer, + int height) { +#ifndef PADDLE_MOBILE_INFERENCE + AgentLayer* agent = dynamic_cast(agentLayer.get()); + CHECK_NOTNULL(agent); + agent->setRealLayer(realLayer, height); +#endif +} + +void NeuralNetwork::connect(std::string agentLayerName, + NeuralNetwork* srcNN, + std::string realLayerName) { + connect(this->getLayer(agentLayerName), srcNN->getLayer(realLayerName)); +} + +void NeuralNetwork::prefetch(const std::vector& inArgs) { + CHECK_EQ(inArgs.size(), dataLayers_.size()); + + if (paramSelfInited_) { + for (auto& para : parameters_) { + if (para->isSparseRemoteUpdate()) { + auto mat = dynamic_cast( + para->getMat(PARAMETER_VALUE).get()); + para->clearGradient(); + if (mat) mat->clearIndices(); + } + } + } + + for (size_t i = 0; i != dataLayers_.size(); ++i) { + if (FLAGS_parallel_nn) { + const_cast(inArgs[i]).deviceId = -1; + } + dataLayers_[i]->setData(inArgs[i]); + } + + for (auto& layer : layers_) { + layer->prefetch(); + } + + if (paramSelfInited_) { + for (auto& para : parameters_) { + if (para->isSparseRemoteUpdate()) { + auto mat = dynamic_cast( + para->getMat(PARAMETER_VALUE).get()); + mat->setupIndices(); + auto matGrad = dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get()); + matGrad->reserveStore(); + } + } + } +} + +void NeuralNetwork::forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType) { + CHECK_EQ(inArgs.size(), dataLayers_.size()); + outArgs->resize(outputLayers_.size()); + for (size_t i = 0; i != dataLayers_.size(); ++i) { + dataLayers_[i]->setData(inArgs[i]); + } + + gLayerStackTrace.set_stage(true); + + { + for (auto& layer : layers_) { + REGISTER_TIMER_INFO("ForwardTimer", layer->getName().c_str()); + gLayerStackTrace.push(layer->getName()); + layer->forward(passType); + gLayerStackTrace.pop(layer->getName()); + } + } + + outArgs->clear(); + outArgs->reserve(outputLayers_.size()); + for (auto& layer : outputLayers_) { + outArgs->push_back(layer->getOutput()); + } +} + +void NeuralNetwork::resetState() { + for (auto& layer : layers_) { + layer->resetState(); + } +} + +void NeuralNetwork::setState(const MachineState& machineState) { + for (size_t i = 0; i < layers_.size(); i++) { + if (machineState[i] != nullptr) { + layers_[i]->setState(machineState[i]); + } + } +} + +void NeuralNetwork::getState(MachineState& machineState) { + machineState.clear(); + machineState.reserve(layers_.size()); + for (auto& layer : layers_) { + LayerStatePtr p = layer->getState(); + machineState.push_back(p); + } +} + +void NeuralNetwork::backward(const UpdateCallback& callback) { + gLayerStackTrace.set_stage(false); + FOR_EACH_R(layer, layers_) { + REGISTER_TIMER_INFO("BackwardTimer", (*layer)->getName().c_str()); + gLayerStackTrace.push((*layer)->getName()); + if ((*layer)->needGradient()) { + (*layer)->backward(callback); + } + gLayerStackTrace.pop((*layer)->getName()); + } +} + +void NeuralNetwork::finish() { +#ifdef PADDLE_WITH_MKLDNN + FOR_EACH_R(layer, layers_) { + MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(*layer); + if (dnnLayer) { + dnnLayer->convertWeightsToPaddle(); + } + } +#endif +} + +Argument NeuralNetwork::getLayerOutput(const std::string& layerName) { + return getLayer(layerName)->getOutput(); +} + +void NeuralNetwork::onPassEnd() { + for (auto& layer : layers_) { + layer->onPassEnd(); + } +} + +void NeuralNetwork::releaseOutput() { + for (auto& layer : middleLayers_) { + Argument& arg = layer->getOutput(); + arg.value.reset(); + } +} + +#ifndef PADDLE_MOBILE_INFERENCE + +class CombinedEvaluator : public Evaluator { + public: + void addEvaluator(std::unique_ptr&& evaluator) { + evaluators_.emplace_back(std::move(evaluator)); + } + void start() override { + for (auto& evaluator : evaluators_) { + evaluator->start(); + } + } + + void finish() override { + for (auto& evaluator : evaluators_) { + evaluator->finish(); + } + } + + void eval(const NeuralNetwork& nn) override { + for (auto& evaluator : evaluators_) { + evaluator->eval(nn); + } + } + real evalImp(std::vector& arguments) override { + (void)arguments; + return -1; + } + void printStats(std::ostream& os) const override { + for (auto& evaluator : evaluators_) { + evaluator->printStats(os); + os << ' '; + } + } + + void distributeEval(ParameterClient2* client) override { + for (auto& evaluator : evaluators_) { + evaluator->distributeEval(client); + } + } + + protected: + std::vector> evaluators_; + + // Evaluator interface + public: + /** + * @brief getNames will return all inside evaluators' names. + * @param names [out]: return names. + */ + void getNames(std::vector* names) override { + for (auto& eval : evaluators_) { + eval->getNames(names); + } + } + + /** + * @brief getValue could get all inside evaluators' value. + */ + real getValue(const std::string& name, Error* err) const override { + return this->getMethodHelper( + name, err, [&name, err](const std::unique_ptr& eval) { + return eval->getValue(name, err); + }); + } + + /** + * @brief getType could get all inside evaluators' type. + */ + std::string getType(const std::string& name, Error* err) const override { + return this->getMethodHelper( + name, err, [&name, err](const std::unique_ptr& eval) { + return eval->getType(name, err); + }); + } + + private: + template + T getMethodHelper(const std::string& name, + Error* err, + const std::function&)>& + callback) const { + for (auto& eval : evaluators_) { + std::vector names; + eval->getNames(&names); + if (std::find(names.begin(), names.end(), name) != names.end()) { + return callback(eval); + } + } + *err = Error("No such key %s", name.c_str()); + return T(); + } +}; + +class SubnetEvaluator : public CombinedEvaluator { + public: + SubnetEvaluator(const std::string& layerName, + std::unique_ptr&& evaluator) + : layerName_(layerName) { + addEvaluator(std::move(evaluator)); + } + void eval(const NeuralNetwork& nn) override { + const LayerPtr& layer = nn.getLayer(layerName_); + CHECK(layer) << "Nonexisted layer: " << layerName_ << " in submodel " + << nn.getName(); + bool accessed = false; + layer->accessSubNetwork([this, &accessed](NeuralNetwork& subnet) { + subnet.eval(evaluators_[0].get()); + accessed = true; + }); + CHECK(accessed) << "There is no subnetwork for layer " << layerName_ + << " in submodel " << nn.getName(); + } + + protected: + std::string layerName_; +}; + +Evaluator* NeuralNetwork::makeEvaluator() const { + CombinedEvaluator* combinedEvaluator = new CombinedEvaluator(); + auto subModelConfig = std::find_if(config_.sub_models().begin(), + config_.sub_models().end(), + [=](const SubModelConfig& sub_model) { + return sub_model.name() == subModelName_; + }); + bool useSubModel = (subModelConfig != config_.sub_models().end()); + CHECK_EQ(useSubModel, !subModelName_.empty()); + if (useSubModel) { + // create the evaluators that belong to CURRENT submodel + for (int i = 0; i < subModelConfig->evaluator_names_size(); ++i) { + // find evaluator by name + auto thisEvalConfig = std::find_if( + config_.evaluators().begin(), + config_.evaluators().end(), + [=](const EvaluatorConfig& ecfg) { + return ecfg.name() == subModelConfig->evaluator_names(i); + }); + bool validConfig = (thisEvalConfig != config_.evaluators().end()); + if (validConfig) { + std::unique_ptr evaluator( + Evaluator::create(*thisEvalConfig)); + combinedEvaluator->addEvaluator(std::move(evaluator)); + } + } + for (auto& layer : layers_) { + layer->accessSubNetwork( + [layer, combinedEvaluator](NeuralNetwork& subnet) { + std::unique_ptr subEvaluator(new SubnetEvaluator( + layer->getName(), + std::unique_ptr(subnet.makeEvaluator()))); + combinedEvaluator->addEvaluator(std::move(subEvaluator)); + }); + } + } else { + for (const EvaluatorConfig& evalConfig : config_.evaluators()) { + std::unique_ptr evaluator(Evaluator::create(evalConfig)); + combinedEvaluator->addEvaluator(std::move(evaluator)); + } + } + return combinedEvaluator; +} + +void NeuralNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); } + +#endif + +void NeuralNetwork::setOutputGrad(const std::vector& args) { + CHECK_GE(outputLayers_.size(), args.size()); + for (size_t i = 0; i < args.size(); ++i) { + outputLayers_[i]->getOutput().grad = args[i].grad; + } +} + +extern NeuralNetwork* newCustomNerualNetwork(const std::string& name, + NeuralNetwork* network) + __attribute__((weak)); + +NeuralNetwork* NeuralNetwork::newNeuralNetwork(const std::string& name, + NeuralNetwork* rootNetwork) { + if (newCustomNerualNetwork) { + return newCustomNerualNetwork(name, rootNetwork); + } else { + return new NeuralNetwork(name, rootNetwork); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/NeuralNetwork.h b/paddle/legacy/gserver/gradientmachines/NeuralNetwork.h new file mode 100644 index 0000000000000000000000000000000000000000..566157c8998a38aef4a3620a4dca7246c6e66391 --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/NeuralNetwork.h @@ -0,0 +1,179 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/dataproviders/DataProvider.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" +#include "paddle/legacy/gserver/layers/CostLayer.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/gserver/layers/Layer.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/utils/ClassRegistrar.h" + +namespace paddle { +/* + * @brief Init function for the parameters. + * @param paramId: the id of the parameter to init. + * @param para: the pointer to the parameter to init. + * @param sharedParams: the pointer to an array of the parameter to be shared. + * If it is null, no parameter sharing is used. + * Only CPU paramters can be shared. + * It handles CPU, CPU sparse, CPU sparse remote, + * and GPU parameters differently. If the type + * of a parameter is NORMAL. Basically nothing need to be done. + * CPU value: NORMAL. + * CPU param: NORMAL. + * + * CPU sparse value: NORMAL. + * CPU sparse gradient: MAT_SPARSE_ROW_AUTO_GROW. + * + * CPU sparse remote value: MAT_SPARSE_ROW_PREFETCH(_FULL_SIZE). + * CPU sparse remote gradient: MAT_SPARSE_ROW (!sharedParams) + * MAT_SPARSE_ROW_AUTO_GROW (sharedParams) + * + * GPU value: NORMAL + * GPU param: NORMAL + */ +void parameterInitNN(int paramId, + Parameter* para, + std::vector* sharedParams); + +class NeuralNetwork : public GradientMachine { + public: + virtual void init(const ModelConfig& config, + ParamInitCallback callback = nullptr, + const std::vector& parameterTypes = + std::vector{PARAMETER_VALUE, + PARAMETER_GRADIENT, + PARAMETER_MOMENTUM}, + bool useGpu = FLAGS_use_gpu); + + /** + * Connect two submodels and + * down-submodel's output become up-submodel's input. + * By default, connection is one by one, + * If the agent height is smaller than real layer, *height* has to be filled. + * + * @param realLayer The down-submodel's output layer. + * @param agentLayer The up-submodel's input agent layer. + */ + static void connect(LayerPtr agentLayer, LayerPtr realLayer, int height = 0); + void connect(std::string agentLayerName, + NeuralNetwork* srcNN, + std::string realLayerName); + + virtual void prefetch(const std::vector& inArgs); + + virtual void forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType); + + virtual void backward(const UpdateCallback& callback = nullptr); + + virtual Argument getLayerOutput(const std::string& layerName); + + const LayerPtr& getLayer(const std::string& layerName) const { + auto it = layerMap_.find(layerName); + CHECK(it != layerMap_.end()) << "Unknown layer " << layerName; + return it->second; + } + + virtual void onPassEnd(); + +#ifndef PADDLE_MOBILE_INFERENCE + virtual Evaluator* makeEvaluator() const; + + virtual void eval(Evaluator* evaluator) const; +#endif + + virtual void resetState(); + virtual void setOutputGrad(const std::vector& args); + + /// set machine state + virtual void setState(const MachineState& machineState); + + /// get machine state + virtual void getState(MachineState& machineState); + + static NeuralNetwork* create(const ModelConfig& config); + + ParameterMap* getParameterMap() { return ¶meterMap_; } + + /** + * @brief Access each layer as a for each loop. + * @param callback invoke with each layer. + */ + template + void forEachLayer(T callback) { + for (auto& l : layers_) { + if (callback(l)) { + break; + } + } + } + + static NeuralNetwork* newNeuralNetwork(const std::string& name = "", + NeuralNetwork* rootNetwork = nullptr); + + const std::string& getName() const { return subModelName_; } + + /// some finish work, like convert the weight format of MKLDNNLayers + void finish(); + + /** + * @brief Release the middle layer's output memory. + * + * @note This function is used for memory optimization in inference. + */ + void releaseOutput(); + + protected: + /** + * The constructor of NeuralNetwork. + * The sub networks can get parameters_ and parameterMap_ + * from base NeuralNetwork. + * + * @param subModelName The name of sub-model. + * @param rootNetwork It used in MultiNetwork. + */ + NeuralNetwork(std::string subModelName = "", + NeuralNetwork* rootNetwork = nullptr) + : subModelName_(subModelName), rootNetwork_(rootNetwork) {} + + std::string subModelName_; + ModelConfig config_; + std::vector layers_; + ParameterMap parameterMap_; + LayerMap layerMap_; + + std::vector dataLayers_; + std::vector outputLayers_; + std::vector middleLayers_; + + static std::map dllInitMap; + + NeuralNetwork* rootNetwork_; + + /// Whether parameter of this NN is initialized by its own + /// (i.e., not by callback supplied with the caller) + bool paramSelfInited_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/ParallelNeuralNetwork.cpp b/paddle/legacy/gserver/gradientmachines/ParallelNeuralNetwork.cpp new file mode 100644 index 0000000000000000000000000000000000000000..33d24b5b832fe9011591606860e0f50361367790 --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/ParallelNeuralNetwork.cpp @@ -0,0 +1,214 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +#include "ParallelNeuralNetwork.h" + +#include +#include + +namespace paddle { + +void ParallelNeuralNetwork::init( + const ModelConfig& config, + ParamInitCallback callback, + const std::vector& parameterTypes, + bool useGpu) { + NeuralNetwork::init(config, callback, parameterTypes, useGpu); + + if (config.type() == "recurrent_nn") { + LOG(FATAL) + << "You can not add `--parallel_nn=true` on the command line, " + << "parallel_nn training mode does not support the recurrent_nn model."; + } + + useGpu_ = useGpu; + numDevices_ = 0; + if (useGpu_) { + numDevices_ = hl_get_device_count(); + } + + for (auto& layer : layers_) { + int deviceId = layer->getDeviceId(); + CHECK_LT(deviceId, numDevices_); + addComputeThread(deviceId); + } +} + +void ParallelNeuralNetwork::addComputeThread(int deviceId) { + for (auto& thread : threads_) { + if (thread->getDeviceId() == deviceId) { + return; + } + } + + threads_.emplace_back(new ParallelThread( + threads_.size(), deviceId, deviceId >= 0 ? useGpu_ : false)); +} + +void ParallelNeuralNetwork::waitAllThread() { + for (auto& thread : threads_) { + thread->jobEnqueue(NULL, TASK_END_LAYER); + } + + for (size_t i = 0; i < threads_.size(); i++) { + threads_[i]->queue_.waitEmpty(); + } +} + +void ParallelNeuralNetwork::dispatchByDeviceId(int deviceId, + LayerPtr layer, + TaskType task) { + for (auto& thread : threads_) { + if (thread->getDeviceId() == deviceId) { + thread->jobEnqueue(layer, task); + return; + } + } + LOG(FATAL) << "No specific device thread "; +} + +void ParallelNeuralNetwork::forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType) { + for (auto& thread : threads_) { + thread->setForwardPassType(passType); + } + CHECK_EQ(inArgs.size(), dataLayers_.size()); + outArgs->resize(outputLayers_.size()); + for (size_t i = 0; i != dataLayers_.size(); ++i) { + const_cast(inArgs[i]).deviceId = -1; + dataLayers_[i]->setData(inArgs[i]); + } + + for (auto& layer : layers_) { + dispatchByDeviceId(layer->getDeviceId(), layer, TASK_FORWARD); + } + + { + REGISTER_TIMER("forwardTime"); + waitAllThread(); + } + outArgs->clear(); + outArgs->reserve(outputLayers_.size()); + for (auto& layer : outputLayers_) { + outArgs->push_back(layer->getOutput()); + } +} + +void ParallelNeuralNetwork::backward(const UpdateCallback& callback) { + for (auto& thread : threads_) { + thread->setBackwardCallback(callback); + } + + FOR_EACH_R(layer, layers_) { + dispatchByDeviceId((*layer)->getDeviceId(), *layer, TASK_BACKWARD); + } + { + REGISTER_TIMER("backwardTime"); + waitAllThread(); + } +} + +void ParallelNeuralNetwork::forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback) { + forward(inArgs, outArgs, passType); + backward(callback); +} + +void ParallelNeuralNetwork::start() { + for (auto& thread : threads_) { + thread->start(); + } +} + +ParallelThread::ParallelThread(int threadId, int deviceId, bool useGpu) + : threadId_(threadId), deviceId_(deviceId), useGpu_(useGpu) {} + +ParallelThread::~ParallelThread() { stop(); } + +void ParallelThread::stop() { + if (computeThread_) { + jobEnqueue(NULL, TASK_THREAD_FINISH); + computeThread_->join(); + computeThread_.reset(nullptr); + } +} + +void ParallelThread::computeThread() { + LOG(INFO) << "gradComputeThread " << threadId_; + + if (useGpu_) { + hl_init(deviceId_); + } + + while (true) { + struct Job job_work = queue_.dequeue(); + + if (job_work.task_ == TASK_END_LAYER) { + continue; + } else if (job_work.task_ == TASK_THREAD_FINISH) { + break; + } + + if (TASK_FORWARD == job_work.task_) { + { + REGISTER_TIMER_INFO("waitInputValue", + job_work.layer_->getName().c_str()); + job_work.layer_->waitInputValue(); + } + { + REGISTER_TIMER_INFO("threadForwardTimer", + job_work.layer_->getName().c_str()); + job_work.layer_->forward(passType_); + } + { + REGISTER_TIMER_INFO("copyOutputToOtherDevice", + job_work.layer_->getName().c_str()); + job_work.layer_->copyOutputToOtherDevice(); + } + } else { + { + REGISTER_TIMER_INFO("waitAndMergeOutputGrad", + job_work.layer_->getName().c_str()); + job_work.layer_->waitAndMergeOutputGrad(); + } + { + REGISTER_TIMER_INFO("threadBackwardTimer", + job_work.layer_->getName().c_str()); + job_work.layer_->backward(backwardCallback_); + } + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + job_work.layer_->markAllInputGrad(); + } + } + hl_fini(); +} + +void ParallelThread::start() { + computeThread_.reset(new std::thread([this]() { computeThread(); })); +} + +void ParallelThread::jobEnqueue(LayerPtr layer, TaskType task) { + struct Job job_work; + job_work.layer_ = layer; + job_work.task_ = task; + queue_.enqueue(job_work); +} + +} // namespace paddle diff --git a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h b/paddle/legacy/gserver/gradientmachines/ParallelNeuralNetwork.h similarity index 100% rename from paddle/gserver/gradientmachines/ParallelNeuralNetwork.h rename to paddle/legacy/gserver/gradientmachines/ParallelNeuralNetwork.h diff --git a/paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e49f042404f80a21293545023efa3e68417c1edb --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.cpp @@ -0,0 +1,1501 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "RecurrentGradientMachine.h" +#include +#include +#include +#include +#include +#include "NeuralNetwork.h" +#include "paddle/legacy/gserver/layers/AgentLayer.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +DEFINE_string(diy_beam_search_prob_so, "", "the diy beam search cost so"); + +static const char* DIY_CALC_PROB_SYMBOL_NAME = "calc_prob"; +static const char* DIY_START_CALC_PROB_SYMBOL_NAME = "start_calc_prob"; +static const char* DIY_FINISH_CALC_PROB_SYMBOL_NAME = "finish_calc_prob"; + +namespace paddle { + +/** + * Start Custom Calculate Probability callback type. + * + * @param nNode, nodes: the path will be explored. nNodes is array size. + * nodes is array elements. + * + * @return: A custom handler id that will passed to another callback. + */ +typedef int (*DiyStartCalcProbCallback)(size_t nNodes, int* nodes); + +/** + * Doing Custom Calculation of Probability callback type. + * + * @param handler: User custom handler. The return value from start calc prob. + * @param nNode, nodes: Array. The current path. + * @param curProb: The current log probability that neural network returns. + * + * @return: Log probability which user calculated, it will be updated to this + * path. + * @NOTE: Return -INFINITY will DROP this path IMMEDIATELY!! + */ +typedef real (*DiyCalcProbCallback)( + int handler, size_t nNodes, int* nodes, real curProb, bool atEos); + +/** + * Finish Custom Calculation of Probability callback type. + * + * @param handler: User custom handler. The return value from start calc prob. + */ +typedef void (*DiyStopCalcProbCallback)(int handler); + +static DiyCalcProbCallback gDiyProbMethod = nullptr; +static DiyStartCalcProbCallback gDiyProbStart = nullptr; +static DiyStopCalcProbCallback gDiyProbStop = nullptr; +static void* gDiyProbHandle = nullptr; + +static void exit_diy_prob() { dlclose(gDiyProbHandle); } + +template +static inline SymbolType loadDiySymbol(const char* symbolName) { + void* sym = dlsym(gDiyProbHandle, symbolName); + CHECK(sym) << "Cannot load symbol " << symbolName << " from " + << FLAGS_diy_beam_search_prob_so; + return reinterpret_cast(sym); +} + +static InitFunction __init__diy_prob_method( + [] { + std::string soName = FLAGS_diy_beam_search_prob_so; + if (!soName.empty()) { + gDiyProbHandle = dlopen(soName.c_str(), RTLD_LAZY); + CHECK(gDiyProbHandle) << "Cannot Open DIY Prob So " << soName; + atexit(exit_diy_prob); + gDiyProbMethod = + loadDiySymbol(DIY_CALC_PROB_SYMBOL_NAME); + gDiyProbStart = loadDiySymbol( + DIY_START_CALC_PROB_SYMBOL_NAME); + gDiyProbStop = loadDiySymbol( + DIY_FINISH_CALC_PROB_SYMBOL_NAME); + } + }, + std::numeric_limits::max()); + +class BeamSearchControlCallbacks { + public: + RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback + beamSearchCandidateAdjust; + RecurrentGradientMachine::NormOrDropNodeCallback normOrDropNode; + RecurrentGradientMachine::DropCallback stopDetermineCandidates; + + //! for gcc46 aggregate initialization is not very well, so we need to + //! explicit + BeamSearchControlCallbacks( + const RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback& + candidateAdjust, + const RecurrentGradientMachine::NormOrDropNodeCallback& norm, + const RecurrentGradientMachine::DropCallback& stop) + : beamSearchCandidateAdjust(candidateAdjust), + normOrDropNode(norm), + stopDetermineCandidates(stop) {} +}; + +class BeamSearchStatisticsCallbacks { + public: + RecurrentGradientMachine::EachStepCallback onEachStepStarted; + RecurrentGradientMachine::EachStepCallback onEachStepStoped; + + BeamSearchStatisticsCallbacks( + const RecurrentGradientMachine::EachStepCallback& start, + const RecurrentGradientMachine::EachStepCallback& stop) + : onEachStepStarted(start), onEachStepStoped(stop) {} +}; + +RecurrentGradientMachine::RecurrentGradientMachine( + const std::string& subModelName, NeuralNetwork* rootNetwork) + : NeuralNetwork(subModelName), + rootNetwork_(rootNetwork), + beamSearchCtrlCallbacks_(nullptr), + beamSearchStatistics_(nullptr) { + CHECK(!subModelName_.empty()); +} + +/** + * bias layer, as input of memory frame 0 will give vector of zeros + * if bias parameter is not set. + * + * boot bias layer create directly in recurrent gradient machine, because: + * + * 1. It is only one frame, so it should not be placed in layer group, + * which is one instance for every one frame. + * + * 2. It is no input layer, so it need resetHeight() before forward(), + * and resetHeight() must be called in recurrent gradient machine, + * so it's should not be placed in root network. + */ +class BootBiasLayer : public Layer { + protected: + std::unique_ptr biases_; + IVectorPtr cpuIds_; + + public: + explicit BootBiasLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override { + if (!Layer::init(layerMap, parameterMap)) return false; + + if (biasParameter_) { + biases_ = + std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + return true; + } + + void resetHeight(int height) { + if (config_.has_bos_id()) { // used as a constant id layerConfig + IVector::resizeOrCreate(output_.ids, height, useGpu_); + output_.ids->reset((int)config_.bos_id()); + } else { + resetOutput(height, getSize()); + } + } + + void forward(PassType passType) override { + if (biases_) { + MatrixPtr outV = getOutputValue(); + outV->addBias(*(biases_->getW()), 1); + forwardActivation(); + } + } + + void backward(const UpdateCallback& callback) override { + if (biases_ && biases_->getWGrad()) { + backwardActivation(); + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + biases_->getParameterPtr()->incUpdate(callback); + } + } +}; + +void RecurrentGradientMachine::init( + const ModelConfig& config, + ParamInitCallback callback, + const std::vector& parameterTypes, + bool useGpu) { + NeuralNetwork::init(config, callback, parameterTypes, useGpu); + useGpu_ = useGpu; + + auto subModelConfig = + std::find_if(config.sub_models().begin(), + config.sub_models().end(), + [this](const SubModelConfig& sub_model) { + return sub_model.name() == this->subModelName_; + }); + CHECK(subModelConfig != config.sub_models().end()); + reversed_ = subModelConfig->reversed(); + generating_ = subModelConfig->has_generator(); + + inFrameLines_.resize(subModelConfig->in_links_size()); + for (size_t i = 0; i < inFrameLines_.size(); ++i) { + inFrameLines_[i].linkName = subModelConfig->in_links(i).link_name(); + inFrameLines_[i].inLayer = + rootNetwork_->getLayer(subModelConfig->in_links(i).layer_name()); + } + + outFrameLines_.resize(subModelConfig->out_links_size()); + for (size_t i = 0; i < outFrameLines_.size(); ++i) { + auto& linkPair = subModelConfig->out_links(i); + outFrameLines_[i].layerName = linkPair.layer_name(); + outFrameLines_[i].agentLayer = rootNetwork_->getLayer(linkPair.link_name()); + } + + memoryFrameLines_.resize(subModelConfig->memories_size()); + for (size_t i = 0; i < memoryFrameLines_.size(); ++i) { + auto& memoryConfig = subModelConfig->memories(i); + memoryFrameLines_[i].layerName = memoryConfig.layer_name(); + memoryFrameLines_[i].linkName = memoryConfig.link_name(); + auto agentConfig = + std::find_if(config.layers().begin(), + config.layers().end(), + [&memoryConfig](const LayerConfig& layerConfig) { + return layerConfig.name() == memoryConfig.link_name(); + }); + CHECK(agentConfig != config.layers().end()); + if (memoryConfig.has_boot_layer_name()) { + memoryFrameLines_[i].rootLayer = + rootNetwork_->getLayer(memoryConfig.boot_layer_name()); + + LayerConfig scatterConfig = *agentConfig; + memoryFrameLines_[i].rootAgent.reset( + new ScatterAgentLayer(scatterConfig)); + memoryFrameLines_[i].rootAgent->init(LayerMap(), parameterMap_); + + memoryFrameLines_[i].bootLayer = memoryFrameLines_[i].rootAgent; + } else { + LayerConfig biasConfig = *agentConfig; + if (memoryConfig.has_boot_bias_parameter_name()) { + biasConfig.set_bias_parameter_name( + memoryConfig.boot_bias_parameter_name()); + biasConfig.set_active_type(memoryConfig.boot_bias_active_type()); + } else if (memoryConfig.has_boot_with_const_id()) { + biasConfig.set_bos_id(memoryConfig.boot_with_const_id()); + } + memoryFrameLines_[i].biasLayer.reset(new BootBiasLayer(biasConfig)); + memoryFrameLines_[i].biasLayer->init(LayerMap(), parameterMap_); + + memoryFrameLines_[i].bootLayer = memoryFrameLines_[i].biasLayer; + } + + if (subModelConfig->has_generator()) { + memoryFrameLines_[i].scatterAgents.resize(2); + for (auto& agent : memoryFrameLines_[i].scatterAgents) { + agent.reset(new ScatterAgentLayer(*agentConfig)); + agent->init(LayerMap(), parameterMap_); + } + } + } + + if (subModelConfig->has_generator()) { + generator_.config = subModelConfig->generator(); + eosFrameLine_.reset(new EosFrameLine); + maxSequenceLength_ = generator_.config.max_num_frames(); + } + + // get parameters actually used by this Layer Group + resizeOrCreateFrames(1); + for (auto& para : frames_[0]->getParameters()) { + if (para->getSharedCount() > 0) { + parameterIds_.push_back(para->getID()); + } + } + for (auto& para : parameters_) { // bias layer parameters + if (para->getSharedCount() > 0) { + parameterIds_.push_back(para->getID()); + } + } +} + +void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) { + if ((size_t)numFrames <= frames_.size()) { + return; + } + + frames_.reserve(numFrames); + for (auto& inFrameLine : inFrameLines_) { + inFrameLine.agents.reserve(numFrames); + } + for (auto& outFrameLine : outFrameLines_) { + outFrameLine.frames.reserve(numFrames); + } + for (auto& memoryFrameLine : memoryFrameLines_) { + memoryFrameLine.frames.reserve(numFrames); + memoryFrameLine.agents.reserve(numFrames); + } + if (eosFrameLine_) { + eosFrameLine_->layers.reserve(numFrames); + } + + ParamInitCallback subParamInitCb = [this](int paramId, Parameter* para) { + para->enableSharedType(PARAMETER_VALUE, + this->parameters_[paramId]->getBuf(PARAMETER_VALUE), + this->parameters_[paramId]->getMat(PARAMETER_VALUE)); + para->enableSharedType( + PARAMETER_GRADIENT, + this->parameters_[paramId]->getBuf(PARAMETER_GRADIENT), + this->parameters_[paramId]->getMat(PARAMETER_GRADIENT)); + }; + + for (int i = frames_.size(); i < numFrames; ++i) { + std::unique_ptr frame( + NeuralNetwork::newNeuralNetwork(subModelName_)); + frame->init(config_, subParamInitCb); + + for (auto& inFrameLine : inFrameLines_) { + inFrameLine.agents.push_back(frame->getLayer(inFrameLine.linkName)); + } + + for (auto& outFrameLine : outFrameLines_) { + outFrameLine.frames.push_back(frame->getLayer(outFrameLine.layerName)); + } + for (auto& memoryFrameLine : memoryFrameLines_) { + memoryFrameLine.frames.push_back( + frame->getLayer(memoryFrameLine.layerName)); + memoryFrameLine.agents.push_back( + frame->getLayer(memoryFrameLine.linkName)); + } + if (eosFrameLine_) { + eosFrameLine_->layers.push_back( + frame->getLayer(generator_.config.eos_layer_name())); + } + + frames_.emplace_back(std::move(frame)); + } +} + +void RecurrentGradientMachine::resizeBootFrame(int numSequences) { + for (auto& memoryFrameLine : memoryFrameLines_) { + if (memoryFrameLine.biasLayer) { + auto biasLayer = + dynamic_cast(memoryFrameLine.biasLayer.get()); + CHECK_NOTNULL(biasLayer); + biasLayer->resetHeight(numSequences); + } else { // check input root layer height + CHECK_EQ(numSequences, + memoryFrameLine.rootLayer->getOutput().getNumSequences()); + } + } +} + +void RecurrentGradientMachine::prefetch(const std::vector& inArgs) { + LOG(FATAL) << "should not use this function"; +} + +void RecurrentGradientMachine::checkInputConsistency( + int inlinkId, const std::vector& seqInfo) { + if (commonSeqInfo_.empty()) { + commonSeqInfo_.resize(seqInfo.size()); + for (size_t i = 0; i < seqInfo.size(); ++i) { + commonSeqInfo_[i].topLevelLength = seqInfo[i].topLevelLength; + commonSeqInfo_[i].seqId = seqInfo[i].seqId; + } + } else { + CHECK_EQ(commonSeqInfo_.size(), seqInfo.size()) + << " RecurrentGroup " << subModelName_ << " input " << inlinkId + << " has mismatched number of sequences"; + for (size_t i = 0; i < seqInfo.size(); ++i) { + CHECK_EQ(commonSeqInfo_[i].topLevelLength, seqInfo[i].topLevelLength) + << " RecurrentGroup " << subModelName_ << " input " << inlinkId + << " has mismatched sequence length"; + CHECK_EQ(commonSeqInfo_[i].seqId, seqInfo[i].seqId) + << " RecurrentGroup " << subModelName_ << " input " << inlinkId + << " has mismatched sequence length"; + } + } +} + +void RecurrentGradientMachine::calcNumSequencesAtEachStep() { + int numSequences = commonSeqInfo_.size(); + numSeqs_.resize(maxSequenceLength_); + for (int i = 0; i < numSequences; ++i) { + for (int j = 0; j < commonSeqInfo_[i].topLevelLength; ++j) { + numSeqs_[j] = i + 1; + } + } +} + +void RecurrentGradientMachine::reorganizeInput(PassType passType) { + info_.clear(); + info_.resize(inFrameLines_.size()); + + commonSeqInfo_.clear(); + seqInfos_.clear(); + seqInfos_.resize(inFrameLines_.size()); + + for (size_t i = 0; i < inFrameLines_.size(); i++) { + const Argument& input = inFrameLines_[i].inLayer->getOutput(); + if (!input.hasSeq()) { + continue; + } + input.getSeqInfo(&seqInfos_[i]); + checkInputConsistency(i, seqInfos_[i]); + } + CHECK(!commonSeqInfo_.empty()) + << "At least one input needs to be sequence or subsequence"; + maxSequenceLength_ = commonSeqInfo_[0].topLevelLength; + + calcNumSequencesAtEachStep(); + + for (size_t i = 0; i < inFrameLines_.size(); ++i) { + const Argument& input = inFrameLines_[i].inLayer->getOutput(); + if (!input.hasSeq()) { + seqInfos_[i] = commonSeqInfo_; + } + createInFrameInfo(i, input, passType); + } + + { + AsyncGpuBlock asyncGpuBlock; + + // inFrameLine select rows in real layer one time + for (size_t i = 0; i < inFrameLines_.size(); i++) { + selectRowsOneTime(inFrameLines_[i].inLayer, + info_[i].allIds, + &(inFrameLines_[i].outArg), + passType); + } + } +} + +void RecurrentGradientMachine::reorganizeOutput(PassType passType) { + calcSequenceStartPositions(); + for (size_t i = 0; i < outFrameLines_.size(); ++i) { + Info info; + auto& outFrameLine = outFrameLines_[i]; + ICpuGpuVectorPtr sequenceStartPositions; + ICpuGpuVectorPtr subSequenceStartPositions; + createOutFrameInfo( + outFrameLine, info, sequenceStartPositions, subSequenceStartPositions); + auto gatherAgent = + dynamic_cast(outFrameLine.agentLayer.get()); + CHECK_NOTNULL(gatherAgent); + gatherAgent->copyIdAndSequenceInfo(sequenceStartPositions, + subSequenceStartPositions, + info.allIds, + info.idIndex); + } +} + +void RecurrentGradientMachine::connectFrames(PassType passType) { + for (auto& memoryFrameLine : memoryFrameLines_) { + if (memoryFrameLine.rootAgent) { + auto scatterAgent = + dynamic_cast(memoryFrameLine.rootAgent.get()); + createMemoryFrameInfo(&memoryFrameLine, passType); + scatterAgent->setRealLayerAndOutput(memoryFrameLine.rootLayer, + memoryFrameLine.outArg, + memoryFrameLine.allIds, + /* idIndex */ 0, + memoryFrameLine.allIds->getSize(), + /* handleBackward */ true); + if (memoryFrameLine.sequenceStartPositions) { + int size = memoryFrameLine.sequenceStartPositions->getSize(); + scatterAgent->setSequenceStartPositions( + memoryFrameLine.sequenceStartPositions, + /* seqStartPosIndex */ 0, + size); + } + } + } + + for (auto& outFrameLine : outFrameLines_) { + auto gatherAgent = + dynamic_cast(outFrameLine.agentLayer.get()); + gatherAgent->clearRealLayers(); + } + for (int i = 0; i < maxSequenceLength_; ++i) { + // connect in_links + for (size_t j = 0; j < inFrameLines_.size(); ++j) { + Info& info = info_[j]; + // idSize denotes the sum number of tokens in each length i + int idIndex = info.idIndex.empty() ? 0 : info.idIndex[i]; + int idSize = info.idIndex.empty() ? numSeqs_[i] + : info.idIndex[i + 1] - info.idIndex[i]; + InFrameLine inFrameLine = inFrameLines_[j]; + auto scatterAgent = + dynamic_cast(inFrameLine.agents[i].get()); + scatterAgent->setRealLayerAndOutput(inFrameLine.inLayer, + inFrameLine.outArg, + info.allIds, + idIndex, + idSize, + i == 0); + if (info.sequenceStartPositions) { + // size: the length of subsequence + int size = info.seqStartPosIndex[i + 1] - info.seqStartPosIndex[i]; + scatterAgent->setSequenceStartPositions( + info.sequenceStartPositions, info.seqStartPosIndex[i], size); + } + } + + // connect out_links + for (auto& outFrameLine : outFrameLines_) { + auto gatherAgent = + dynamic_cast(outFrameLine.agentLayer.get()); + gatherAgent->addRealLayer(outFrameLine.frames[i]); + } + for (auto& memoryFrameLine : memoryFrameLines_) { + NeuralNetwork::connect( + memoryFrameLine.agents[i], + i == 0 ? memoryFrameLine.bootLayer : memoryFrameLine.frames[i - 1], + numSeqs_[i] /*height of agent*/); + } + } +} + +void RecurrentGradientMachine::forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType) { + /* inArgs and outArgs are not used. + The inputs are inFrameLines_[i].inLayer. + The outputs are outFramesLines_[i].agentLayer + */ + + if (generating_) { + generateSequence(); + return; + } // else forward.. + + reorganizeInput(passType); + int numSequences = commonSeqInfo_.size(); + + resizeOrCreateFrames(maxSequenceLength_); + resizeBootFrame(numSequences); + + connectFrames(passType); + + REGISTER_TIMER_INFO("RecurrentFwTime", "RecurrentFwTime"); + // forward + for (auto& memoryFrameLine : memoryFrameLines_) { + memoryFrameLine.bootLayer->forward(passType); + } + for (int i = 0; i < maxSequenceLength_; ++i) { + const std::vector inArgs; + std::vector outArgs; + frames_[i]->forward(inArgs, &outArgs, passType); + } + + reorganizeOutput(passType); +} + +void RecurrentGradientMachine::backward(const UpdateCallback& callback) { + if (generating_) { + return; + } + REGISTER_TIMER_INFO("RecurrentBwTime", "RecurrentBwTime"); + AsyncGpuBlock asyncGpuBlock; + for (int i = maxSequenceLength_ - 1; i >= 0; --i) { + frames_[i]->backward(nullptr); + } + for (auto& memoryFrameLine : memoryFrameLines_) { + memoryFrameLine.bootLayer->backward(nullptr); + } +} + +void RecurrentGradientMachine::forwardBackward( + const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback) { + LOG(FATAL) << "should not use this function"; +} + +void RecurrentGradientMachine::eval(Evaluator* evaluator) const { + // call printers frame by frame + for (int i = 0; i < maxSequenceLength_; ++i) { + VLOG(2) << "Recurrent Layer Group eval frame " << i << " begin"; + evaluator->eval(*(frames_[i].get())); + VLOG(2) << "Recurrent Layer Group eval frame " << i << " end"; + } +} + +void RecurrentGradientMachine::registerBeamSearchControlCallbacks( + const BeamSearchCandidatesAdjustCallback& adjustBeamSearch, + const NormOrDropNodeCallback& normOrDropNode, + const DropCallback& stopBeamSearch) { + this->removeBeamSearchControlCallbacks(); + //! for gcc 46, aggregate initialization is not supported. TAT + this->beamSearchCtrlCallbacks_ = new BeamSearchControlCallbacks( + adjustBeamSearch, normOrDropNode, stopBeamSearch); +} + +void RecurrentGradientMachine::removeBeamSearchControlCallbacks() { + if (this->beamSearchCtrlCallbacks_) { + delete this->beamSearchCtrlCallbacks_; + this->beamSearchCtrlCallbacks_ = nullptr; + } +} + +void RecurrentGradientMachine::registerBeamSearchStatisticsCallbacks( + const EachStepCallback& onEachStepStarted, + const EachStepCallback& onEachStepStoped) { + this->removeBeamSearchStatisticsCallbacks(); + this->beamSearchStatistics_ = + new BeamSearchStatisticsCallbacks(onEachStepStarted, onEachStepStoped); +} + +void RecurrentGradientMachine::removeBeamSearchStatisticsCallbacks() { + if (this->beamSearchStatistics_) { + delete this->beamSearchStatistics_; + this->beamSearchStatistics_ = nullptr; + } +} + +namespace { +void lenToStarts(std::vector& starts) { + int pos = 0; + starts.back() = 0; + for (auto& start : starts) { + int tmp = start; + start = pos; + pos += tmp; + } + starts.back() = pos; +} +} // namespace + +void RecurrentGradientMachine::calcSequenceStartPositions() { + std::vector starts(commonSeqInfo_.size() + 1); + for (auto& seqInfo : commonSeqInfo_) { + starts[seqInfo.seqId] = seqInfo.topLevelLength; + } + lenToStarts(starts); + ICpuGpuVector::resizeOrCreate(sequenceStartPositions_, starts.size(), false); + std::copy(starts.begin(), + starts.end(), + sequenceStartPositions_->getMutableData(false)); +} + +void RecurrentGradientMachine::checkOutputConsistency( + OutFrameLine& outFrameLine) { + bool hasSeq = outFrameLine.frames[0]->getOutput().hasSeq(); + for (int i = 0; i < maxSequenceLength_; ++i) { + LayerPtr frame = outFrameLine.frames[i]; + CHECK_EQ(hasSeq, frame->getOutput().hasSeq()); + int numSequences = frame->getOutput().getNumSequences(); + CHECK_EQ(numSeqs_[i], numSequences); + } +} + +void RecurrentGradientMachine::createOutFrameInfo( + OutFrameLine& outFrameLine, + Info& info, + ICpuGpuVectorPtr& sequenceStartPositions, + ICpuGpuVectorPtr& subSequenceStartPositions) { + checkOutputConsistency(outFrameLine); + + if (!outFrameLine.frames[0]->getOutput().hasSeq()) { + createOutFrameInfo_seq( + outFrameLine, info, sequenceStartPositions, subSequenceStartPositions); + } else { + createOutFrameInfo_subseq( + outFrameLine, info, sequenceStartPositions, subSequenceStartPositions); + } +} + +void RecurrentGradientMachine::createOutFrameInfo_seq( + OutFrameLine& outFrameLine, + Info& info, + ICpuGpuVectorPtr& sequenceStartPositions, + ICpuGpuVectorPtr& subSequenceStartPositions) { + std::vector allIds; + info.idIndex.resize(1, 0); // first idIndex = 0 + + const int* starts = sequenceStartPositions_->getData(false); + + for (int i = 0; i < maxSequenceLength_; ++i) { + LayerPtr frame = outFrameLine.frames[i]; + size_t numSequences = frame->getOutput().getNumSequences(); + for (size_t j = 0; j < numSequences; ++j) { + int seqStart = starts[commonSeqInfo_[j].seqId]; + int seqLength = commonSeqInfo_[j].topLevelLength; + allIds.push_back(reversed_ ? (seqStart + seqLength - 1 - i) + : (seqStart + i)); + } + info.idIndex.push_back(allIds.size()); + } + sequenceStartPositions = sequenceStartPositions_; + copyScattedId(allIds, &info.allIds, allIds.size()); + CHECK_EQ(info.idIndex.size(), static_cast(maxSequenceLength_ + 1)); +} + +void RecurrentGradientMachine::createOutFrameInfo_subseq( + OutFrameLine& outFrameLine, + Info& info, + ICpuGpuVectorPtr& sequenceStartPositions, + ICpuGpuVectorPtr& subSequenceStartPositions) { + size_t numSequences = commonSeqInfo_.size(); + std::vector allIds; + info.idIndex.resize(1, 0); // first idIndex = 0 + + const int* starts = sequenceStartPositions_->getData(false); + std::vector subStarts(starts[numSequences] + 1); + for (int i = 0; i < maxSequenceLength_; ++i) { + LayerPtr frame = outFrameLine.frames[i]; + size_t numSequences = frame->getOutput().getNumSequences(); + const int* seqStarts = + frame->getOutput().sequenceStartPositions->getData(false); + for (size_t j = 0; j < numSequences; ++j) { + subStarts[starts[commonSeqInfo_[j].seqId] + i] = + seqStarts[j + 1] - seqStarts[j]; + } + } + lenToStarts(subStarts); + + for (int i = 0; i < maxSequenceLength_; ++i) { + LayerPtr frame = outFrameLine.frames[i]; + size_t numSequences = frame->getOutput().getNumSequences(); + for (size_t j = 0; j < numSequences; ++j) { + int pos = starts[commonSeqInfo_[j].seqId] + i; + int subSeqStart = subStarts[pos]; + int subSeqEnd = subStarts[pos + 1]; + for (int k = subSeqStart; k < subSeqEnd; ++k) { + allIds.push_back(k); + } + } + info.idIndex.push_back(allIds.size()); + } + + ICpuGpuVector::resizeOrCreate( + subSequenceStartPositions, subStarts.size(), false); + int* cpuSubSequenceStartPositions = + subSequenceStartPositions->getMutableData(false); + std::copy(subStarts.begin(), subStarts.end(), cpuSubSequenceStartPositions); + ICpuGpuVector::resizeOrCreate( + sequenceStartPositions, numSequences + 1, false); + int* cpuSequenceStartPositions = + sequenceStartPositions->getMutableData(false); + for (size_t i = 0; i <= numSequences; ++i) { + cpuSequenceStartPositions[i] = subStarts[starts[i]]; + } + copyScattedId(allIds, &info.allIds, allIds.size()); + CHECK_EQ(info.idIndex.size(), static_cast(maxSequenceLength_ + 1)); +} + +/* create scattered id infomation for all realLayer of inFrameLines one time. + * If hasSubseq, will also create scattered sequenceStartPositions infomation + * for all realLayer of inFrameLines one time. + */ +void RecurrentGradientMachine::createInFrameInfo(int inlinkId, + const Argument& input, + PassType passType) { + if (!input.hasSeq()) { + createInFrameInfo_nonseq(inlinkId, input, passType); + } else if (!input.hasSubseq()) { + createInFrameInfo_seq(inlinkId, input, passType); + } else { + createInFrameInfo_subseq(inlinkId, input, passType); + } +} + +void RecurrentGradientMachine::createInFrameInfo_nonseq(int inlinkId, + const Argument& input, + PassType passType) { + std::vector allIds; + + auto& seqInfo = seqInfos_[inlinkId]; + Info* inlinkInfo = &info_[inlinkId]; + inlinkInfo->idIndex.clear(); + for (size_t i = 0; i < seqInfo.size(); ++i) { + allIds.push_back(seqInfo[i].seqId); + } + // copy and check scatterId + copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize()); +} + +void RecurrentGradientMachine::createInFrameInfo_seq(int inlinkId, + const Argument& input, + PassType passType) { + std::vector allIds; + auto& seqInfo = seqInfos_[inlinkId]; + Info* inlinkInfo = &info_[inlinkId]; + inlinkInfo->idIndex.resize(1, 0); // first idIndex = 0 + + for (int i = 0; i < maxSequenceLength_; ++i) { + for (int j = 0; j < numSeqs_[i]; ++j) { + int seqLength = seqInfo[j].topLevelLength; + int seqStart = seqInfo[j].seqStart; + allIds.push_back(reversed_ ? (seqStart + seqLength - 1 - i) + : (seqStart + i)); + } + inlinkInfo->idIndex.push_back(allIds.size()); + } + + // copy and check scatterId + copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize()); + CHECK_EQ(inlinkInfo->idIndex.size(), + static_cast(maxSequenceLength_ + 1)); +} +void RecurrentGradientMachine::createInFrameInfo_subseq(int inlinkId, + const Argument& input, + PassType passType) { + std::vector allIds; + + auto& seqInfo = seqInfos_[inlinkId]; + + Info* inlinkInfo = &info_[inlinkId]; + inlinkInfo->idIndex.resize(1, 0); // first idIndex = 0 + std::vector sequenceStartPositions; + const int* subSequenceStartPositions = nullptr; + + subSequenceStartPositions = input.subSequenceStartPositions->getData(false); + inlinkInfo->seqStartPosIndex.clear(); + inlinkInfo->seqStartPosIndex.push_back(0); // first seqStartPosIndex = 0 + for (int i = 0; i < maxSequenceLength_; ++i) { + sequenceStartPositions.push_back(0); // first element = 0 + for (int j = 0; j < numSeqs_[i]; ++j) { + int subSeqStart = subSequenceStartPositions[seqInfo[j].subSeqStart + i]; + int subSeqEnd = subSequenceStartPositions[seqInfo[j].subSeqStart + i + 1]; + for (int k = subSeqStart; k < subSeqEnd; ++k) { + allIds.push_back(k); + } + sequenceStartPositions.push_back(sequenceStartPositions.back() + + subSeqEnd - subSeqStart); + } + inlinkInfo->idIndex.push_back(allIds.size()); + inlinkInfo->seqStartPosIndex.push_back(sequenceStartPositions.size()); + } + // inFrameLine create sequenceStartPositions one time + CHECK_EQ( + sequenceStartPositions.size(), + static_cast(maxSequenceLength_ + input.getNumSubSequences())); + CHECK_EQ(inlinkInfo->seqStartPosIndex.size(), + static_cast(maxSequenceLength_ + 1)); + createSeqPos(sequenceStartPositions, &inlinkInfo->sequenceStartPositions); + + // copy and check scatterId + copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize()); + CHECK_EQ(inlinkInfo->idIndex.size(), + static_cast(maxSequenceLength_ + 1)); +} + +/* like createInFrameInfo, but for all realLayer of memoryFrameLines*/ +void RecurrentGradientMachine::createMemoryFrameInfo( + MemoryFrameLine* memoryFrameLine, PassType passType) { + const Argument& input = (*memoryFrameLine).rootLayer->getOutput(); + size_t numSequences = input.getNumSequences(); + std::vector allIds; + bool seqFlag = input.hasSeq(); + CHECK(!input.hasSubseq()) + << "Subsequence boot layer for memory is not supported"; + + if (seqFlag) { // for sequenceScatterAgentLayer + std::vector sequenceStartPositions; + sequenceStartPositions.push_back(0); // first element = 0 + const int* starts = input.sequenceStartPositions->getData(false); + for (size_t i = 0; i < numSequences; ++i) { + // memory info adopt info of inlinks[0] + int seqId = seqInfos_[0][i].seqId; + for (int k = starts[seqId]; k < starts[seqId + 1]; ++k) { + allIds.push_back(k); + } + sequenceStartPositions.push_back(sequenceStartPositions.back() + + starts[seqId + 1] - starts[seqId]); + } + createSeqPos(sequenceStartPositions, + &(*memoryFrameLine).sequenceStartPositions); + + } else { // for scatterAgentLayer + for (size_t i = 0; i < numSequences; ++i) { + allIds.push_back(seqInfos_[0][i].seqId); + } + } + // copy and check scatterId + copyScattedId(allIds, &(*memoryFrameLine).allIds, input.getBatchSize()); + // memoryFrameLine select rows in real layer one time + selectRowsOneTime((*memoryFrameLine).rootLayer, + (*memoryFrameLine).allIds, + &(*memoryFrameLine).outArg, + passType); +} + +void RecurrentGradientMachine::copyScattedId(std::vector& srcIds, + IVectorPtr* dstIds, + int size) { + int idSize = srcIds.size(); + CHECK_EQ(idSize, size); + IVector::resizeOrCreate(*dstIds, idSize, useGpu_); + (*dstIds)->copyFrom(srcIds.data(), idSize); + // check + std::sort(srcIds.begin(), srcIds.end()); + for (int i = 0; i < idSize; ++i) { + CHECK_EQ(srcIds[i], i); + } +} + +void RecurrentGradientMachine::selectRowsOneTime(LayerPtr layer, + const IVectorPtr& allIds, + Argument* arg, + PassType passType) { + Argument& src = layer->getOutput(); + if (src.value) { + const MatrixPtr& realV = src.value; + int height = realV->getHeight(); + int width = realV->getWidth(); + Matrix::resizeOrCreate( + arg->value, height, width, /* trans */ false, useGpu_); + arg->value->zeroMem(); + arg->value->selectRows(*realV, *allIds); + if (passType != PASS_TEST) { + Matrix::resizeOrCreate( + arg->grad, height, width, /* trans */ false, useGpu_); + arg->grad->zeroMem(); + } + } + if (src.ids) { + IVector::resizeOrCreate(arg->ids, src.ids->getSize(), useGpu_); + arg->ids->selectFrom(*src.ids, *allIds); + } +} + +void RecurrentGradientMachine::createSeqPos( + const std::vector& sequenceStartPosition, + ICpuGpuVectorPtr* sequenceStartPositions) { + int size = sequenceStartPosition.size(); + const int* data = sequenceStartPosition.data(); + ICpuGpuVector::resizeOrCreate(*sequenceStartPositions, size, false); + (*sequenceStartPositions)->copyFrom(data, size, false); +} + +size_t RecurrentGradientMachine::getGenBatchSize() { + size_t numSequences = 0; + for (auto& memoryFrameLine : memoryFrameLines_) { + if (!memoryFrameLine.rootLayer) continue; + Argument& bootArg = memoryFrameLine.rootLayer->getOutput(); + size_t batchSize = bootArg.getNumSequences(); + if (numSequences) { + CHECK_EQ(numSequences, batchSize); + } else { + numSequences = batchSize; + } + } + CHECK(numSequences) + << "Fail to get batch size in generation. " + "At least one of the Memory layer MUST have a layer that is NOT in " + "the layer group to boot it, and this boot layer is used to " + "decide batch_size in generation process."; + return numSequences; +} + +void RecurrentGradientMachine::generateSequence() { + CHECK_NOTNULL(eosFrameLine_.get()); + CHECK_GE(outFrameLines_.size(), 1UL); + size_t numSequences = getGenBatchSize(); + + resizeBootFrame(numSequences); + // We create only two sub-network in generation, one stores states of all + // layers in previous time step and the other storing the states at current + // time step. + resizeOrCreateFrames(2); + + // outFrameLines_.size() > 1UL + dataArgsSize_ = outFrameLines_.size() - 1; + dataArgs_.resize(dataArgsSize_); + dataArgsFrame_.clear(); + dataArgsFrame_.resize(dataArgsSize_); + + // connect boot frame memory links + std::vector ids(numSequences); + for (size_t i = 0; i < numSequences; ++i) { + ids[i] = i; + } + for (auto& memoryFrameLine : memoryFrameLines_) { + if (memoryFrameLine.rootAgent) { + auto scatterAgent = + dynamic_cast(memoryFrameLine.rootAgent.get()); + scatterAgent->setRealLayer(memoryFrameLine.rootLayer, ids); + } + NeuralNetwork::connect( + memoryFrameLine.agents[0], memoryFrameLine.bootLayer, ids.size()); + } + + // boot layer forward + AsyncGpuBlock asyncGpuBlock; + + for (auto& memoryFrameLine : memoryFrameLines_) { + memoryFrameLine.bootLayer->forward(PASS_TEST); + } + + // init outArg + size_t resultNum = generator_.config.num_results_per_sample(); + size_t maxGenWordCount = + generator_.config.max_num_frames() * numSequences * resultNum; + IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false); + if (resultNum > 1) { + CHECK_LE(resultNum, static_cast(generator_.config.beam_size())); + Matrix::resizeOrCreate(generator_.outArg.in, + /* height */ numSequences, + /* width */ resultNum, + false, + /* useGpu */ false); + } + ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions, + numSequences + 1, + /* useGpu */ false); + if (getBeamSize() > 1) { + beamSearch(numSequences); + } else { + oneWaySearch(numSequences); + } + if (dataArgsSize_) createDataOutlink(); + + size_t size = generator_.ids.size(); + generator_.outArg.ids->resize(size); + generator_.outArg.ids->copyFrom(generator_.ids.data(), size); + + OutFrameLine& outFrameLine = outFrameLines_[0]; + auto dataAgent = dynamic_cast(outFrameLine.agentLayer.get()); + CHECK_NOTNULL(dataAgent); + dataAgent->setData(generator_.outArg); + dataAgent->prefetch(); +} + +void RecurrentGradientMachine::oneWaySearch(size_t batchSize) { + OutFrameLine& outFrameLine = outFrameLines_[0]; + + // finalPaths_[0] stores the generated results of the + // entire batch, so its size exactly equals to batchSize. + finalPaths_.clear(); + finalPaths_.resize(1); + std::vector& finalPaths = finalPaths_[0]; + finalPaths.resize(batchSize); + + seqIds_.resize(batchSize); + std::vector scatterIds; + for (size_t i = 0; i < batchSize; ++i) { + finalPaths[i].seqId = i; + seqIds_[i] = i; + } + + // forward + for (int i = 0; i < maxSequenceLength_; ++i) { + if (i && scatterIds.empty()) break; + int machineCur = i % 2; + int machinePrev = (i - 1) % 2; + // connect memory links + if (i) { + seqIds_.clear(); + for (size_t j = 0; j < batchSize; ++j) { + if (finalPaths[j].seqId != -1) seqIds_.push_back(j); + } + + for (auto& memoryFrameLine : memoryFrameLines_) { + auto scatterAgent = dynamic_cast( + memoryFrameLine.scatterAgents[machineCur].get()); + scatterAgent->setRealLayer(memoryFrameLine.frames[machinePrev], + scatterIds); + scatterAgent->forward(PASS_TEST); + NeuralNetwork::connect(memoryFrameLine.agents[machineCur], + memoryFrameLine.scatterAgents[machineCur]); + } + } + const std::vector inArgs; + std::vector outArgs; + frames_[machineCur]->forward(inArgs, &outArgs, PASS_TEST); + + const IVectorPtr& idVec = outFrameLine.frames[machineCur]->getOutput().ids; + for (size_t j = 0; j < seqIds_.size(); ++j) { + finalPaths[seqIds_[j]].ids.push_back(idVec->getElement(j)); + finalPaths[seqIds_[j]].machineIdVec.push_back(j); + } + + copyDataOutlinkFrame(machineCur); + + // check eos + const IVectorPtr& eosVec = + eosFrameLine_->layers[machineCur]->getOutput().ids; + scatterIds.clear(); + for (size_t j = 0; j < seqIds_.size(); ++j) { + if (eosVec->getElement(j) == 1U) { + // path.seqId = -1 indicates end of generation + // of an input sequence + finalPaths[seqIds_[j]].seqId = -1; + } else { + scatterIds.push_back(j); + } + } + } + + batchMachineIdVec_.clear(); + batchMachineStartPos_.clear(); + int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); + starts[0] = 0; + generator_.ids.clear(); + for (size_t i = 0; i < batchSize; ++i) { + generator_.ids.insert(generator_.ids.end(), + finalPaths[i].ids.begin(), + finalPaths[i].ids.end()); + starts[i + 1] = generator_.ids.size(); + batchMachineIdVec_.insert(batchMachineIdVec_.end(), + finalPaths[i].machineIdVec.begin(), + finalPaths[i].machineIdVec.end()); + } +} + +void RecurrentGradientMachine::connectPrevFrame(int stepId, + std::vector& paths) { + int machineCur = stepId % 2; + int machinePrev = (stepId - 1) % 2; + int beam = getBeamSize(); + machineIds_.clear(); + topIds_.clear(); + seqIds_.clear(); + + for (size_t j = 0; j < paths.size(); ++j) { + machineIds_.push_back(paths[j].machineId); + topIds_.push_back(paths[j].machineId * beam + paths[j].topIndex); + seqIds_.push_back(paths[j].seqId); + } + + for (auto& memoryFrameLine : memoryFrameLines_) { + bool isOutIds = (memoryFrameLine.layerName == outFrameLines_[0].layerName); + auto scatterAgent = dynamic_cast( + memoryFrameLine.scatterAgents[machineCur].get()); + scatterAgent->setRealLayer(memoryFrameLine.frames[machinePrev], + isOutIds ? topIds_ : machineIds_); + scatterAgent->forward(PASS_TEST); + NeuralNetwork::connect(memoryFrameLine.agents[machineCur], + memoryFrameLine.scatterAgents[machineCur]); + } +} + +void RecurrentGradientMachine::forwardFrame(int machineCur) { + // forward + const std::vector inArgs; + std::vector outArgs; + frames_[machineCur]->forward(inArgs, &outArgs, PASS_TEST); + + copyDataOutlinkFrame(machineCur); + + IVectorPtr& ids = outFrameLines_[0].frames[machineCur]->getOutput().ids; + MatrixPtr in = outFrameLines_[0].frames[machineCur]->getOutput().in; + IVectorPtr& eos = eosFrameLine_->layers[machineCur]->getOutput().ids; + if (useGpu_) { + IVector::resizeOrCreate(cpuId_, ids->getSize(), false /* useGpu */); + cpuId_->copyFrom(*ids); + Matrix::resizeOrCreate(cpuProb_, + in->getHeight(), + in->getWidth(), + false /* trans */, + false /* useGpu */); + cpuProb_->copyFrom(*in); + IVector::resizeOrCreate(cpuEos_, eos->getSize(), false /* useGpu */); + cpuEos_->copyFrom(*eos); + } else { + cpuId_ = ids; + cpuProb_ = in; + cpuEos_ = eos; + } +} + +void RecurrentGradientMachine::singlePathExpand(Path& curPath, + size_t curPathId, + std::vector& newPaths, + size_t expandWidth) { + int calc_id = + gDiyProbStart ? gDiyProbStart(curPath.ids.size(), curPath.ids.data()) : 0; + + const int* idVec = cpuId_->getData(); + const real* probMat = cpuProb_->getData(); + const int* eosVec = cpuEos_->getData(); + + for (size_t k = 0; k < expandWidth; k++) { + int index = curPathId * expandWidth + k; + int id = idVec[index]; + real prob = probMat[index]; + /* + * Ordinarily, beam search greedily expands the most promising expandWidth + * paths that currently are ALWAYS returned by MaxIdLayer. + * In one condition, if user customizes the beam search procedure by + * restricting the expansion within a user defined subset, + * as a result, MaxIdLayer possibly COULD NOT return expandWidth + * vaild expansions, and it will use -1 to indicate the end of valid + * expansion candidates. + */ + if (id == -1) break; + + real newLogProb = generator_.config.log_prob() ? std::log(prob) : prob; + Path newPath( + curPath, id, newLogProb, curPathId /*machineId*/, k /*topIndex*/); + if (this->beamSearchCtrlCallbacks_) { + if (beamSearchCtrlCallbacks_->stopDetermineCandidates( + newPath.seqId, newPath.ids, newPath.probHistory)) + return; + } + // outFrameLines_.size() > 1UL + if (dataArgsSize_) { + newPath.machineIdVec = curPath.machineIdVec; + newPath.machineIdVec.push_back(curPathId); + } + bool atEos = + eosVec[index] == 1U || newPath.ids.size() >= (size_t)maxSequenceLength_; + // adjustNewPath + newPath.adjustProb(calc_id, atEos); + if (this->beamSearchCtrlCallbacks_) { + this->beamSearchCtrlCallbacks_->normOrDropNode( + newPath.seqId, newPath.ids, newPath.probHistory, &newPath.logProb); + } + if (!newPath.isDropable()) { + atEos ? finalPaths_[curPath.seqId].push_back(newPath) + : newPaths.push_back(newPath); + } + } // for expandWidth + + if (gDiyProbStop) { + gDiyProbStop(calc_id); + } +} + +void RecurrentGradientMachine::beamExpand(std::vector& paths, + std::vector& newPaths) { + size_t candidatePathCount = paths.size(); + // idVec.size() could be larger than candidatePathCount * beam, + // so user can drop some node customly. + CHECK_EQ(cpuId_->getSize() % candidatePathCount, 0UL); + size_t expandWidth = cpuId_->getSize() / candidatePathCount; + + // iterate over each sequence + size_t totalExpandCount = 0; + int prevSeqId = -1; + int curSeqId = 0; + for (size_t j = 0; j <= candidatePathCount; j++) { + // expansions of a single sequence are all processed + curSeqId = (j < candidatePathCount ? paths[j].seqId : curSeqId + 1); + if (prevSeqId != -1 && curSeqId != prevSeqId) { + totalExpandCount += beamShrink(newPaths, prevSeqId, totalExpandCount); + } + if (j == candidatePathCount) return; + singlePathExpand(paths[j], j, newPaths, expandWidth); + + prevSeqId = paths[j].seqId; + } // for paths +} + +// Drop extra nodes to beam size. +size_t RecurrentGradientMachine::beamShrink(std::vector& newPaths, + size_t seqId, + size_t totalExpandCount) { + size_t minNewPathSize = + std::min(getBeamSize(), newPaths.size() - totalExpandCount); + if (!minNewPathSize) { + return 0; + } + std::nth_element(newPaths.begin() + totalExpandCount, + newPaths.begin() + totalExpandCount + minNewPathSize, + newPaths.end(), + Path::greaterPath); + newPaths.resize(totalExpandCount + minNewPathSize); + + real minPathLogProb = + std::min_element(newPaths.end() - minNewPathSize, newPaths.end()) + ->logProb; + real maxPathLogProb = + std::max_element(newPaths.end() - minNewPathSize, newPaths.end()) + ->logProb; + + // Remove the already formed paths that are relatively short + finalPaths_[seqId].erase( + std::remove_if(finalPaths_[seqId].begin(), + finalPaths_[seqId].end(), + [&](Path& p) { return p.logProb < minPathLogProb; }), + finalPaths_[seqId].end()); + for (auto p : finalPaths_[seqId]) { + if (minFinalPathLogProb_[seqId] > p.logProb) { + minFinalPathLogProb_[seqId] = p.logProb; + } + } + + if (finalPaths_[seqId].size() >= getBeamSize() && + minFinalPathLogProb_[seqId] >= maxPathLogProb) { + newPaths.resize(totalExpandCount); + return 0; + } + return minNewPathSize; +} + +void RecurrentGradientMachine::fillGenOutputs() { + size_t numResults = generator_.config.num_results_per_sample(); + for (size_t i = 0; i < finalPaths_.size(); ++i) { + size_t minFinalPathsSize = std::min(numResults, finalPaths_[i].size()); + std::partial_sort(finalPaths_[i].begin(), + finalPaths_[i].begin() + minFinalPathsSize, + finalPaths_[i].end(), + Path::greaterPath); + finalPaths_[i].resize(minFinalPathsSize); + } + + generator_.ids.clear(); + int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); + starts[0] = 0; + if (numResults > 1) { + int idsProbSaveSize = 0; + for (auto inSeq : finalPaths_) { + for (auto path : inSeq) idsProbSaveSize += path.ids.size(); + idsProbSaveSize += inSeq.size(); + } + Matrix::resizeOrCreate( + generator_.outArg.value, idsProbSaveSize, 1, false, false); + real* idsProb = generator_.outArg.value->getData(); + + real* probs = generator_.outArg.in->getData(); + size_t curPos = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + Path& path = finalPaths_[i][j]; + size_t genLen = path.ids.size(); + generator_.ids.push_back(genLen); // sequence size + generator_.ids.insert( + generator_.ids.end(), path.ids.begin(), path.ids.end()); + generator_.ids.push_back(-1); // end of sequence + + memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen); + curPos += genLen; + idsProb[curPos++] = -1.0; + probs[i * numResults + j] = path.logProb; + } + starts[i + 1] = generator_.ids.size(); + } + } else { + for (size_t i = 0; i < finalPaths_.size(); ++i) { + CHECK(!finalPaths_[i].empty()); + Path& path = finalPaths_[i][0]; + generator_.ids.insert( + generator_.ids.end(), path.ids.begin(), path.ids.end()); + starts[i + 1] = starts[i] + path.ids.size(); + } + } +} + +void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) { + for (size_t i = 0; i < dataArgsSize_; i++) { + Argument outFrame; + outFrame.resizeAndCopyFrom( + outFrameLines_[i + 1].frames[machineCur]->getOutput(), useGpu_); + dataArgsFrame_[i].emplace_back(outFrame); + } +} + +void RecurrentGradientMachine::createDataOutlinkSelRowsInfo( + bool isSeq, std::vector& outArgs) { + batchMachineIdVec_.clear(); + + size_t seqIdx = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + std::vector& machineIdVec = finalPaths_[i][j].machineIdVec; + if (isSeq) { + for (size_t i = 0; i < machineIdVec.size(); ++i) { + size_t rowId = machineIdVec[i]; + int* seqPos = + outArgs[i].sequenceStartPositions->getMutableData(false); + batchMachineIdVec_.push_back(seqPos[rowId]); + } + } else { + batchMachineIdVec_.insert( + batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end()); + } + seqIdx++; + } + } +} + +void RecurrentGradientMachine::createDataOutlinkCopySizeInfo( + bool isSeq, std::vector& outArgs, std::vector& copySize) { + size_t totalSeqNum = std::accumulate( + finalPaths_.begin(), + finalPaths_.end(), + 0UL, + [](size_t a, const std::vector& b) { return a + b.size(); }); + copySize.resize(totalSeqNum, 1); + + batchMachineStartPos_.resize(totalSeqNum + 1, 0); + if (isSeq) { + ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions; + CHECK_EQ(static_cast(inputSeqStartPos->getSize() - 1), + getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size()); + int* starts = inputSeqStartPos->getMutableData(false); + int seqId = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i] + : starts[j + 1] - starts[j]; + batchMachineStartPos_[seqId + 1] = + batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size(); + seqId++; + } + } + } else { + for (size_t i = 0; i < finalPaths_[0].size(); ++i) + batchMachineStartPos_[i + 1] = + batchMachineStartPos_[i] + finalPaths_[0][i].ids.size(); + } +} + +void RecurrentGradientMachine::createDataOutlink() { + for (size_t i = 0; i < dataArgsSize_; i++) { + bool isSeq = dataArgsFrame_[i][0].hasSeq(); + std::vector copySize; + createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize); + createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]); + + dataArgs_[i].concat(dataArgsFrame_[i], + batchMachineIdVec_, + batchMachineStartPos_, + copySize, + useGpu_, + HPPL_STREAM_1, + PASS_TEST); + auto dataAgent = + dynamic_cast(outFrameLines_[i + 1].agentLayer.get()); + CHECK_NOTNULL(dataAgent); + dataAgent->setData(dataArgs_[i]); + } +} + +void RecurrentGradientMachine::beamSearch(size_t batchSize) { + finalPaths_.clear(); + finalPaths_.resize(batchSize); + seqIds_.resize(batchSize); + minFinalPathLogProb_.clear(); + minFinalPathLogProb_.resize(batchSize, 0); + + std::vector paths; + std::vector newPaths; + for (size_t i = 0; i < batchSize; ++i) { + paths.push_back(Path(i)); + if (this->beamSearchCtrlCallbacks_) { + paths.back().recordHistory(); + } + } + + // restart beam search + stopBeamSearch_ = false; + for (int i = 0; i < maxSequenceLength_; ++i) { + int machineCur = i % 2; + std::unique_ptr< + ScopedCallbacks> + statisticsBlock; + if (this->beamSearchStatistics_) { + auto ptr = + new ScopedCallbacks(beamSearchStatistics_->onEachStepStarted, + beamSearchStatistics_->onEachStepStoped, + i); + statisticsBlock.reset(ptr); + } + if (stopBeamSearch_) break; + + if (i) connectPrevFrame(i, paths); + + if (this->beamSearchCtrlCallbacks_) { + std::vector*> prefixes; + prefixes.resize(paths.size()); + std::transform( + paths.begin(), paths.end(), prefixes.begin(), [](const Path& p) { + return const_cast*>(&p.ids); + }); + beamSearchCtrlCallbacks_->beamSearchCandidateAdjust( + prefixes, frames_[machineCur].get(), i); + } + + forwardFrame(machineCur); + beamExpand(paths, newPaths); + if (newPaths.empty()) break; + + paths = newPaths; + newPaths.clear(); + } // end for machineCur + fillGenOutputs(); +} + +void RecurrentGradientMachine::Path::adjustProb(int calc_id, bool atEos) { + if (gDiyProbMethod) { + logProb = gDiyProbMethod(calc_id, ids.size(), ids.data(), logProb, atEos); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.h new file mode 100644 index 0000000000000000000000000000000000000000..0a13d4f6f84eb5309a1b25f039357cb8af02c35e --- /dev/null +++ b/paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.h @@ -0,0 +1,580 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "GradientMachine.h" +#include "NeuralNetwork.h" + +#include "paddle/legacy/utils/Locks.h" + +namespace paddle { + +/** + * Private data class declares. + * Used for user customized beam search. + */ +class BeamSearchControlCallbacks; +class BeamSearchStatisticsCallbacks; + +class RecurrentGradientMachine : public NeuralNetwork { + public: + RecurrentGradientMachine(const std::string& subModelName, + NeuralNetwork* rootNetwork); + + // Disable copy and assign. + RecurrentGradientMachine(const RecurrentGradientMachine& other) = delete; + RecurrentGradientMachine& operator=(const RecurrentGradientMachine& other) = + delete; + + virtual ~RecurrentGradientMachine() { + this->removeBeamSearchStatisticsCallbacks(); + this->removeBeamSearchControlCallbacks(); + } + + virtual void init(const ModelConfig& config, + ParamInitCallback callback, + const std::vector& parameterTypes, + bool useGpu); + + virtual void prefetch(const std::vector& inArgs); + + virtual void forward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType); + + virtual void backward(const UpdateCallback& callback = nullptr); + + void forwardBackward(const std::vector& inArgs, + std::vector* outArgs, + PassType passType, + const UpdateCallback& callback); + + virtual void resetState() {} + virtual void eval(Evaluator* evaluator) const; + + const std::vector& getParameterIds() { return parameterIds_; } + + /** + * @brief BeamSearchCandidatesAdjustCallback + * + * Adjust searching candidates to restrict beam search + * searching within a limited subset of all possibile paths. + * + * The first parameter is the prefixes of all formed paths in current + * beam search step, whose type is basically int[][]. + * + * The second parameter is a pointer to the network used to generate sequence, + * user can use this pointer to tranverse each layer in the network to + * modify behaivors of a particular layer. + * + * The third parameter is an integer to indicate the iteration number of + * beam search, so that user can customize different operations in different + * beam search iterations. + */ + typedef std::function*>&, NeuralNetwork*, const int)> + BeamSearchCandidatesAdjustCallback; + + /** + * @brief DropCallback + * + * Drop a whole prefix or one candidate in beam search or not. + * + * The first parameter is sequence index in a batch + * + * The second parameter is one path in beam search, + * which is made up of node indices. + * + * The third parameter is probabilites for each node in this path. + * + * Return true if this prefix or candidate is expected to be dropped. + */ + typedef std::function&, const std::vector&)> + DropCallback; + + /** + * @brief NormOrDropNodeCallback + * + * Normalize a path's probabilities or just drop it by modifying path.logProb + * + * The first parameter is sequence index in a batch + * + * The second parameter is path.ids + * + * The third parameter is probabilites for each node in this path. + * + * The fourth parameter is the probability of the whole path. + */ + typedef std::function&, std::vector&, real*)> + NormOrDropNodeCallback; + + /** + * @brief Register beam search control callbacks. Used for prediction. + * + * @param queryBeamSearch: Give the sequences already formed, return the + * nodes expected to be expanded. + * Input: A pointer to an array holding pathes which have been expanded + * Return: A pointer to an array holding nodes wanted to be expanded. + * + * @param dropOneNode: Early drop a node in one beam search step. + * Given the path formed and probability history, decide whether a node + * should be dropped or not. + * + * @param stopBeamSearch: Early stop a path in one beam search step. + * Given the path and probability history, decide whether a path + * should be dropped or not. + */ + void registerBeamSearchControlCallbacks( + const BeamSearchCandidatesAdjustCallback& adjustBeamSearch, + const NormOrDropNodeCallback& normOrDropNode, + const DropCallback& stopBeamSearch); + + /** + * @brief Remove user costumized beam search callbacks, + * + * make sequence generation acts like normal beam search. + */ + void removeBeamSearchControlCallbacks(); + + /** + * @brief EachStepCallback + * + * Invoke with beam search step. + */ + typedef std::function EachStepCallback; + + /** + * @brief register statistics methods for performance profile of beam search. + * + * @param onEachStepStarted: invoke once a beam search step starts. + * Its input is index of the beam search step. + * + * @param onEachStepStoped: invoke once a beam search step ends. + * Its input is index of the beam search step. + */ + void registerBeamSearchStatisticsCallbacks( + const EachStepCallback& onEachStepStarted, + const EachStepCallback& onEachStepStoped); + + /** + * @brief Remove beam search callbacks. + */ + void removeBeamSearchStatisticsCallbacks(); + + /** + * @brief Stop beam search for current source. + * + * Will restart beam search in the next forward + */ + void stopBeamSearch(); + + struct Path { + /** + * @brief ids, path of beam search. + */ + std::vector ids; + + /** + * @brief idsProb, log probability of each generated word. + */ + std::vector idsProb; + + /** + * @brief logProb, current probability of path. + */ + real logProb; + + int machineId; // index of sample in frame + int topIndex; // index of MaxIdLayer output in one sample + int seqId; // index of sequence in batch generation + std::vector machineIdVec; + + /** + * @brief A record of each node's probality in a formed path in beam search. + * + * @note It could be empty when history is not recorded. If the history is + * wanted to be recorded, recordHistory() MUST be invoked first. + */ + std::vector probHistory; + + /** + * @brief Path default ctor, first logProb is 0. + */ + Path() { + logProb = 0; + seqId = 0; + } + explicit Path(size_t seqId) : seqId(seqId) { logProb = 0; } + + /** + * @brief Create a new path based on an old path and + * a new node with probability. + * + * @param old old path + * @param newId index of the new node + * @param logProb probability of the new node. + * @param machineId sample index of a frame in RNN + * @param topIndex index of MaxIdLayer output in one sample + */ + Path(Path& old, int newId, real logProb, int machineId, int topIndex) + : ids(old.ids), + idsProb(old.idsProb), + logProb(old.logProb + logProb), + machineId(machineId), + topIndex(topIndex), + seqId(old.seqId) { + ids.push_back(newId); + idsProb.push_back(logProb); + if (!old.probHistory.empty()) { + this->probHistory = old.probHistory; + // probHistory store current prob, not sum + this->probHistory.push_back(logProb); + } + } + + /** + * @brief operator < + * + * Path a < Path b means log probability of a is smaller than that of b + */ + bool operator<(const Path& other) const { + return (logProb < other.logProb); + } + + static bool greaterPath(const Path& a, const Path& b) { return (b < a); } + + /** + * @brief Start recording history in this path. + */ + void recordHistory() { this->probHistory.push_back(this->logProb); } + + /** + * @brief Adjust probability for DIY beam search interface. + * In normal situation, it will do nothing. + * + * @param calc_id: the object id for DIY beam search interface. + * @param atEos: at end of sequence or not. + */ + void adjustProb(int calc_id, bool atEos = false); + + /** + * @brief isDropable indacating whether the current node will be + * dropped or not in beam search. + * + * @note: if logProb is -inf, current node will be dropped. + * @return true to drop the current node. + */ + bool isDropable() const { return std::isinf(logProb) && logProb < 0; } + }; + + /** + * @brief access beam search results. + * @return beam search results. + */ + const std::vector>& getFinalPaths() const { + return this->finalPaths_; + } + + protected: + std::vector commonSeqInfo_; + ICpuGpuVectorPtr sequenceStartPositions_; + void calcSequenceStartPositions(); + void checkInputConsistency(int inlinkId, + const std::vector& seqInfo); + void reorganizeInput(PassType passType); + void reorganizeOutput(PassType passType); + void connectFrames(PassType passType); + void calcNumSequencesAtEachStep(); + + void resizeOrCreateFrames(int numFrames); + void resizeBootFrame(int numSequences); + + void generateSequence(); + void oneWaySearch(size_t batchSize); + void beamSearch(size_t batchSize); + + struct InFrameLine { + std::string linkName; + LayerPtr inLayer; + std::vector agents; // Scatter Agents to reform batch input + Argument outArg; // scatter output argument + }; + std::vector inFrameLines_; + + struct OutFrameLine { + std::string layerName; + LayerPtr agentLayer; + std::vector frames; + }; + std::vector outFrameLines_; + + struct MemoryFrameLine { + std::string layerName; + std::string linkName; + LayerPtr bootLayer; // actually used biasLayer or rootAgent + LayerPtr biasLayer; + LayerPtr rootLayer; // layer in root network to boot this memory + LayerPtr rootAgent; // agent to link rootLayer + std::vector frames; + std::vector agents; + std::vector scatterAgents; // scatter agent used by beam search + Argument outArg; // scatter output argument + // Different memoryFrameLine have different element as follows + IVectorPtr allIds; // scattered id of realLayer + ICpuGpuVectorPtr + sequenceStartPositions; // scattered sequenceStartPositions + }; + std::vector memoryFrameLines_; + + // Each inFrameLines(inlinks) has its own info(elements) below, + // and all outFrameLines(outlinks) share the info with one inFrameLine, + // which is assigned by targetInfoInlinkId_. + struct Info { + // The original positions in the original batch + IVectorPtr allIds; // scattered id of realLayer [batchSize] + + // index of allIds for each step [maxSequenceLength_] + // idIndex[i] is the total length of the first i sequences + std::vector idIndex; + + ICpuGpuVectorPtr + sequenceStartPositions; // scattered sequenceStartPositions + std::vector seqStartPosIndex; // index of sequenceStartPositions + }; + std::vector info_; // for input + + // numSeqs_[i] is the number sequences which is longer than i (for sequence + // data) or has more than i subsequences (for subsequence data) + // Equivalently, numSeqs_[i] is the number of sequences at step i; + std::vector numSeqs_; + + std::vector> seqInfos_; + + void checkOutputConsistency(OutFrameLine& outFrameLine); + + /* create scattered id infomation for all realLayer of inFrameLines one time. + * If hasSubseq, will also create scattered sequenceStartPositions infomation + * for all realLayer of inFrameLines one time. + */ + void createInFrameInfo(int inlinks_id, + const Argument& input, + PassType passType); + void createInFrameInfo_nonseq(int inlinks_id, + const Argument& input, + PassType passType); + void createInFrameInfo_seq(int inlinks_id, + const Argument& input, + PassType passType); + void createInFrameInfo_subseq(int inlinks_id, + const Argument& input, + PassType passType); + + void createOutFrameInfo(OutFrameLine& outFrameLine, + Info& info, + ICpuGpuVectorPtr& sequenceStartPositions, + ICpuGpuVectorPtr& subSequenceStartPositions); + void createOutFrameInfo_seq(OutFrameLine& outFrameLine, + Info& info, + ICpuGpuVectorPtr& sequenceStartPositions, + ICpuGpuVectorPtr& subSequenceStartPositions); + void createOutFrameInfo_subseq(OutFrameLine& outFrameLine, + Info& info, + ICpuGpuVectorPtr& sequenceStartPositions, + ICpuGpuVectorPtr& subSequenceStartPositions); + + void createMemoryFrameInfo(MemoryFrameLine* memoryFrameLine, + PassType passType); + + void copyScattedId(std::vector& srcIds, IVectorPtr* dstIds, int size); + + void selectRowsOneTime(LayerPtr layer, + const IVectorPtr& allIds, + Argument* arg, + PassType passType); + + void createSeqPos(const std::vector& sequenceStartPosition, + ICpuGpuVectorPtr* sequenceStartPositions); + + // for generator + struct EosFrameLine { + std::vector layers; + }; + std::unique_ptr eosFrameLine_; + + struct Generator { + GeneratorConfig config; + std::vector ids; // store generated sequences + std::vector idsProb; // log probability of each generated word + Argument outArg; // final output argument + }; + bool generating_; + Generator generator_; + + std::vector> frames_; + + NeuralNetwork* rootNetwork_; + bool reversed_; + + int maxSequenceLength_; // Max top-level length + bool useGpu_; + bool stopBeamSearch_; + + std::vector + parameterIds_; // parameters actually used by this Layer Group + + // store final argument of outFrameLines_ + std::vector dataArgs_; + // store each frame's output argument of outFrameLines_ + std::vector> dataArgsFrame_; + size_t dataArgsSize_; // size of dataArgs_ = size of dataArgsFrame_ + + IVectorPtr cpuId_; + MatrixPtr cpuProb_; + IVectorPtr cpuEos_; + + private: + /* + * @return beam size in beam search + */ + size_t getBeamSize() { return generator_.config.beam_size(); } + + /* + * @return number of sequence in a batch in generation + */ + size_t getGenBatchSize(); + + /* + * @brief store output of the machineCur-th frame during generation, for + * creating the final outlink after the entire generation process is finished. + * + * In generation, if the layer group has more than 1 outlink, the first + * one is reserved to store the generated word indices, the others are data + * outlinks, that can be used like a common layer in the network. + * + * @param machineCur : index to access the layer group frame in + * currrent generation step. + */ + void copyDataOutlinkFrame(size_t machineCur); + + /* + * @brief In generation, if the layer group has more than 1 outlink, outlink + * except the first one is a data outlink. In RecurrentLayerGroup, each time + * step is a separate Network, outputs of a layer inside the + * RecurrentLayerGroup are stored in separate Arguments. If one layer is + * specified as an outlink of RecurrentLayerGroup. This function will + * collect outputs in each time step of each generated sequence which are + * dispersed in separate Arguments to form a new single Argument as output of + * RecurrentLayerGroup. + */ + void createDataOutlink(); + + /* + * @brief decide to select how many rows from the Matrix stored the forward + * pass results from a start position. + * + * @param isSeq: a flag indicating whetehr the layer to be output of the + * RecurrentGradientMachine is a sequence or not + * @param outArgs: all of the the returned Arguments of the forward pass + * during the generation process. + * @param copySize: the returned result, number of rows to select from the + * Matrix stored the forward pass results from a start position. + */ + void createDataOutlinkCopySizeInfo(bool isSeq, + std::vector& outArgs, + std::vector& copySize); + + /* + * @brief decide index of the start row for each time step of a generated + * sequence in Matrix stored the entire beam search batch's forward pass + * results. + * + * @param isSeq: a flag indicating whether the layer to be output of the + * RecurrentGradientMachine is a sequence or not + * @param outArgs: all of the returned Arguments of the forward pass + * during the generation process. + */ + void createDataOutlinkSelRowsInfo(bool isSeq, std::vector& outArgs); + + /* + * @brief used in beam search, connect previous frame to form recurrent link + * @param stepId : iteration number of generation process. + * It equals to the length of longest half-generated sequence. + * @param paths : half-generated paths that are going to be expanded + * in current beam search iteration. + */ + void connectPrevFrame(int stepId, std::vector& paths); + + /* + * @brief used in beam search, forward current recurrent frame + * @param machineCur : index to access the layer group frame in + * currrent generation step. + */ + void forwardFrame(int machineCur); + + /* + * @brief reduce all expanded paths to beam size. + * + * @param newPaths : newPaths[totalExpandCount : ] stores all expanded paths + * for the seqId-th sequence + * @param seqId : sequence index in a batch + * @param totalExpandCount : number of already shrinked paths in newPaths + * @return size of retained paths at the end of a beam search iteration + */ + size_t beamShrink(std::vector& newPaths, + size_t seqId, + size_t totalExpandCount); + + /* + * @brief expand a single path to expandWidth new paths + * with highest probability + * @param curPath : path to be expanded + * @param curPathId : index of curPath in member newPaths + * @param expandWidth : number of paths to be expanded + */ + void singlePathExpand(Path& curPath, + size_t curPathId, + std::vector& newPaths, + size_t expandWidth); + + /* + * @brief A new beam search iteration. Each half-generated paths in previous + * beam search iteration are further expanded to beam_size new paths + * with highest probabilities, and then all the expanded paths are again + * reduced to beam_size paths according to their log probabilities. + * @param paths : half-generated paths in previous iteration. + * @param newPaths : paths expanded and then reduces in current iteration. + */ + void beamExpand(std::vector& paths, std::vector& newPaths); + + /* + * @brief fill sequence start positions and some other information that are + * uesed by the "text_printer" evaluator. + */ + void fillGenOutputs(); + + std::vector machineIds_; + std::vector topIds_; + std::vector seqIds_; + std::vector batchMachineIdVec_; + std::vector batchMachineStartPos_; + std::vector> finalPaths_; + std::vector minFinalPathLogProb_; + BeamSearchControlCallbacks* beamSearchCtrlCallbacks_; + BeamSearchStatisticsCallbacks* beamSearchStatistics_; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/AddtoLayer.cpp b/paddle/legacy/gserver/layers/AddtoLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..39c5603d9389b433b77e2876f34b3061c62f68f0 --- /dev/null +++ b/paddle/legacy/gserver/layers/AddtoLayer.cpp @@ -0,0 +1,79 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "AddtoLayer.h" + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(addto, AddtoLayer); + +bool AddtoLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + + return true; +} + +void AddtoLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(0)->getHeight(); + int size = getSize(); + + reserveOutput(batchSize, size); + + MatrixPtr outV = getOutputValue(); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + MatrixPtr input = getInputValue(i); + i == 0 ? outV->assign(*input) : outV->add(*input); + } + /* add the bias-vector */ + if (biases_.get() != NULL) { + outV->addBias(*(biases_->getW()), 1); + } + + /* activation */ { forwardActivation(); } +} + +void AddtoLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { backwardActivation(); } + + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + /* Calculate the input layers error */ + MatrixPtr preGrad = getInputGrad(i); + if (NULL != preGrad) { + preGrad->add(*getOutputGrad()); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/AddtoLayer.h b/paddle/legacy/gserver/layers/AddtoLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..ad3cefe1a4d27953b2fef535e1b865175a2cadc2 --- /dev/null +++ b/paddle/legacy/gserver/layers/AddtoLayer.h @@ -0,0 +1,63 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * This layer just simply add all input layers together, then activate + * the sum inputs. Each input of this layer should be the same size, + * which is also the output size of this layer. + * \f[ + * y=f(\sum_{i}x_i + b) + * \f] + * where \f$y\f$ is output, \f$x\f$ is input, \f$b\f$ is bias, and \f$f\f$ is + * activation function. + * + * The config file api is addto_layer. + */ +class AddtoLayer : public Layer { + protected: + std::unique_ptr biases_; + + public: + explicit AddtoLayer(const LayerConfig& config) : Layer(config) {} + + ~AddtoLayer() {} + + /** + * Intialization of AddtoLayer. + */ + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + /** + * Forward propagation. + * @note There is no weight matrix for each input, + * because it just a simple add operation. + */ + void forward(PassType passType) override; + + /** + * Backward propagation. + */ + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/AgentLayer.cpp b/paddle/legacy/gserver/layers/AgentLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..bae89b2fa34d156adae1305d78d6c1465ccdd0ae --- /dev/null +++ b/paddle/legacy/gserver/layers/AgentLayer.cpp @@ -0,0 +1,281 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "AgentLayer.h" + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(agent, AgentLayer); + +bool AgentLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + CHECK_EQ(config_.inputs_size(), 0); + if (!Layer::init(layerMap, parameterMap)) { + return false; + } + setNeedGradient(true); + return true; +} + +void AgentLayer::forward(PassType passType) { + Layer::forward(passType); + + Argument& realOutput = realLayer_->getOutput(); + int realNumSequences = realOutput.getNumSequences(); + CHECK_LE(numSamples_, realNumSequences); + + // get Arguments from real layers + if (numSamples_ > 0 && numSamples_ < realNumSequences) { + if (realOutput.hasSeq()) { + int numRows = + realOutput.sequenceStartPositions->getData(false)[numSamples_]; + output_.subArgFrom(realOutput, + /* offset */ 0, + numRows, + getSize(), + useGpu_, + /* trans */ false, + /* seqFlag */ true, + /* seqStart */ 0, + /* seqSize */ numSamples_ + 1); + } else { + output_.subArgFrom( + realOutput, /* offset */ 0, numSamples_, getSize(), useGpu_); + } + } else { + output_ = realOutput; + } +} + +bool GatherAgentLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + CHECK_EQ(config_.inputs_size(), 0); + if (!Layer::init(layerMap, parameterMap)) { + return false; + } + setNeedGradient(true); + return true; +} + +void GatherAgentLayer::copyIdAndSequenceInfo( + ICpuGpuVectorPtr sequenceStartPositions, + ICpuGpuVectorPtr subSequenceStartPositions, + const IVectorPtr& ids, + const std::vector& idIndex) { + output_.sequenceStartPositions = sequenceStartPositions; + output_.subSequenceStartPositions = subSequenceStartPositions; + allIds_ = ids; + idIndex_ = idIndex; +} + +void GatherAgentLayer::forward(PassType passType) { + Layer::forward(passType); + forwardIds(passType); + forwardValue(passType); +} + +void GatherAgentLayer::forwardValue(PassType passType) { + MatrixPtr valueReal = realLayers_[0]->getOutputValue(); + if (!valueReal) return; + + int height = allIds_->getSize(); + int width = this->getSize(); + resetOutput(height, width); + idsVec_.resize(idIndex_.size()); + + const MatrixPtr& outV = getOutputValue(); + + for (size_t i = 0; i < realLayers_.size(); ++i) { + const MatrixPtr& realV = realLayers_[i]->getOutputValue(); + idsVec_[i] = IVector::create(allIds_->getData() + idIndex_[i], + /* size */ realV->getHeight(), + useGpu_); + realV->addToRows(*outV, *idsVec_[i]); + } +} + +namespace { + +// dest[index[i]] <- src[i] for each i +void copyElements(const IVector& srcVec, + const IVector& indexVec, + IVector& destVec) { + const int* src = srcVec.getData(); + const int* index = indexVec.getData(); + int* dest = destVec.getData(); + int len = indexVec.getSize(); + CHECK_EQ(srcVec.getSize(), indexVec.getSize()); + for (int i = 0; i < len; ++i) { + dest[index[i]] = src[i]; + } +} +} // namespace + +void GatherAgentLayer::forwardIds(PassType passType) { + IVectorPtr realId = realLayers_[0]->getOutputLabel(); + if (!realId) return; + + IVector::resizeOrCreate(output_.ids, allIds_->getSize(), useGpu_); + IVectorPtr outId = output_.ids; + idsVec_.resize(idIndex_.size()); + + for (size_t i = 0; i < realLayers_.size(); ++i) { + const IVectorPtr& realId = realLayers_[i]->getOutputLabel(); + idsVec_[i] = IVector::create(allIds_->getData() + idIndex_[i], + /* size */ realId->getSize(), + useGpu_); + execViaCpu(©Elements, *realId, *idsVec_[i], *outId); + } +} + +void GatherAgentLayer::backward(const UpdateCallback& callback) { + (void)callback; + const MatrixPtr& outputGrad = getOutputGrad(); + + for (size_t i = 0; i < realLayers_.size(); ++i) { + const MatrixPtr& realG = realLayers_[i]->getOutputGrad(); + if (realG) { + realG->selectRows(*outputGrad, *idsVec_[i]); + } + } +} + +bool ScatterAgentLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + CHECK_EQ(config_.inputs_size(), 0); + if (!Layer::init(layerMap, parameterMap)) { + return false; + } + setNeedGradient(true); + return true; +} + +void ScatterAgentLayer::forward(PassType passType) { + Layer::forward(passType); + CHECK_EQ(realLayer_->getDeviceId(), this->getDeviceId()); + + int width = this->getSize(); + if (selectionMode_) { + forwardWithSelection(passType); + } else { + if (realOutArg_.hasSeq()) { + output_.subArgFrom(realOutArg_, + /* offset */ idIndex_, + idSize_, + width, + useGpu_, + /* trans */ false, + /* seqFlag */ true, + /* seqStart */ seqStartPosIndex_, + /* seqSize */ numSequences_); + } else { + output_.subArgFrom( + realOutArg_, /* offset */ idIndex_, idSize_, width, useGpu_); + } + } +} + +void ScatterAgentLayer::backward(const UpdateCallback& callback) { + (void)callback; + + CHECK(!selectionMode_); + + const MatrixPtr& outputGrad = realOutArg_.grad; + const MatrixPtr& realGrad = realLayer_->getOutputGrad(); + if (realGrad) { + // for agent in inFrameLines and memoryFrameLines, + // only first scatterAgentLayer should do addToRows in backward + if (handleBackward_) { + outputGrad->addToRows(*realGrad, *ids_); + } + } +} + +REGISTER_LAYER(gather_agent, GatherAgentLayer); +REGISTER_LAYER(scatter_agent, ScatterAgentLayer); + +void ScatterAgentLayer::forwardWithSelection(PassType passType) { + Layer::forward(passType); + CHECK_EQ(realLayer_->getDeviceId(), this->getDeviceId()); + + const Argument& input = realLayer_->getOutput(); + CHECK_EQ(realLayer_->getSize(), this->getSize()); + int width = this->getSize(); + + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceAgentLayerForward", getName().c_str()); + + if (!input.hasSeq()) { + if (realLayer_->getOutput().ids) { + IVector::resizeOrCreate(output_.ids, ids_->getSize(), useGpu_); + output_.ids->selectFrom(*realLayer_->getOutput().ids, *ids_); + } + if (realLayer_->getOutput().value) { + int height = ids_->getSize(); + resetOutput(height, width); + + const MatrixPtr& outV = getOutputValue(); + const MatrixPtr& realV = realLayer_->getOutputValue(); + outV->selectRows(*realV, *ids_); + } + } else { + // Putting the generation logic here is really an ugly hack! + // used in generation + int height = 0; + size_t numSequences = ids_->getSize(); + const int* starts = input.getCpuStartPositions(); + size_t size = input.hasSubseq() ? input.getNumSubSequences() + : input.getNumSequences(); + const int* cpuIds = cpuIds_->getData(); + + for (size_t i = 0; i < numSequences; ++i) { + size_t seqId = cpuIds[i]; + CHECK_LT(seqId, size); + height += starts[seqId + 1] - starts[seqId]; + } + reserveOutput(height, width); + + const MatrixPtr& outputValue = getOutputValue(); + + CHECK_NE(input.sequenceStartPositions.get(), + output_.sequenceStartPositions.get()); + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, numSequences + 1, false); + int* outStarts = output_.sequenceStartPositions->getMutableData(false); + + ICpuGpuVector::resizeOrCreate(inputStartPos_, height, false); + int* inStarts = inputStartPos_->getMutableData(false); + + size_t offsetOut = 0; + for (size_t i = 0; i < numSequences; ++i) { + outStarts[i] = offsetOut; + size_t seqId = cpuIds[i]; + int size = starts[seqId + 1] - starts[seqId]; + for (int j = 0; j < size; j++) { + inStarts[offsetOut + j] = starts[seqId] + j; + } + offsetOut += size; + } + outStarts[numSequences] = offsetOut; + + outputValue->copyByRowIndex(*input.value, + *inputStartPos_->getVector(useGpu_)); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/AgentLayer.h b/paddle/legacy/gserver/layers/AgentLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..a05eac5e704466df02a74ce6e5364ab6f03f7446 --- /dev/null +++ b/paddle/legacy/gserver/layers/AgentLayer.h @@ -0,0 +1,177 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * AgentLayer use as a virtual input of another layer in config, + * before execute forward/backward, setRealLayer() should be + * called to set one and only one real layer + */ +class AgentLayer : public Layer { + protected: + LayerPtr realLayer_; + int numSamples_; + + public: + explicit AgentLayer(const LayerConfig& config) : Layer(config) {} + + ~AgentLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + // if *numSamples* set, + // real layer output will only use first *numSamples* rows + void setRealLayer(LayerPtr layer, int numSamples = 0) { + realLayer_ = layer; + numSamples_ = numSamples; + } + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override {} +}; + +/** + * Like AgentLayer, but it can gather many real layers. Each real + * layer give a few rows of a sequence, after gather all real layers, + * GatherAgentLayer collect a complete sequence. + */ +class GatherAgentLayer : public Layer { + protected: + std::vector realLayers_; + std::vector idsVec_; + // we don't clear idsVec_ vector to aviod IVector alloc/free + IVectorPtr allIds_; + std::vector idIndex_; + + public: + explicit GatherAgentLayer(const LayerConfig& config) : Layer(config) {} + + virtual ~GatherAgentLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + // call before addRealLayer + void clearRealLayers() { realLayers_.clear(); } + + void copyIdAndSequenceInfo(ICpuGpuVectorPtr sequenceStartPositions, + ICpuGpuVectorPtr subSequenceStartPositions, + const IVectorPtr& allIds, + const std::vector& idIndex); + + // add one real layer, can call many times + void addRealLayer(LayerPtr layer) { realLayers_.push_back(layer); } + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + void forwardValue(PassType passType); + void forwardIds(PassType passType); +}; + +/** + * Like AgentLayer, but only select a few rows in real layer. + * [idIndex, idIndex + idSize) of *ids* in setRealLayerAndOutput() + * are the selected row ids. It's used to scatter one layer's output + * to many small submodels. ScatterAgentLayer can support ids real layer, + * if it is, the agent will select a few ids in real layer. + */ +class ScatterAgentLayer : public Layer { + protected: + LayerPtr realLayer_; + IVectorPtr ids_; + IVectorPtr cpuIds_; + Argument realOutArg_; + int idIndex_; + int idSize_; + int seqStartPosIndex_; + int numSequences_; // number of sequences in this scatterAgentLayer + bool handleBackward_; + + // use to store expanded cpuStartPositions or subSequenceStartPositions + // of real layer. + ICpuGpuVectorPtr inputStartPos_; + + // true for setRealLayer, false for setRealLayerAndOutput + bool selectionMode_; + + public: + explicit ScatterAgentLayer(const LayerConfig& config) : Layer(config) {} + + virtual ~ScatterAgentLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + /** + * @brief set real layer in generation + * + * @param layer[input] realLayer + * @param ids[input] row id in real layer + * @param copyId[input] whether to copy a cpu version of ids, + * false(default) in ScatterAgentLayer, and + * true in SequenceScatterAgentLayer. + */ + void setRealLayer(LayerPtr layer, const std::vector& ids) { + realLayer_ = layer; + IVector::resizeOrCreate(ids_, ids.size(), useGpu_); + ids_->copyFrom(ids.data(), ids.size()); + if (useGpu_) { + IVector::resizeOrCreate(cpuIds_, ids.size(), false); + cpuIds_->copyFrom(ids.data(), ids.size()); + } else { + cpuIds_ = ids_; + } + selectionMode_ = true; + } + + // set real layer and output, [idIndex, idIndex + idSize) of *ids* + // are selected row for realOutArg in realLayer + void setRealLayerAndOutput(LayerPtr layer, + const Argument& outArg, + const IVectorPtr& ids, + int idIndex, + int idSize, + bool handleBackward) { + realLayer_ = layer; + realOutArg_ = outArg; + ids_ = ids; + idIndex_ = idIndex; + idSize_ = idSize; + handleBackward_ = handleBackward; + selectionMode_ = false; + } + + void setSequenceStartPositions(const ICpuGpuVectorPtr& sequenceStartPositions, + int seqStartPosIndex, + int numSequences) { + realOutArg_.sequenceStartPositions = sequenceStartPositions; + seqStartPosIndex_ = seqStartPosIndex; + numSequences_ = numSequences; + } + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + + void forwardWithSelection(PassType passType); +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/AverageLayer.cpp b/paddle/legacy/gserver/layers/AverageLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0539da793712527c72792603ae28a1d0aa903bcc --- /dev/null +++ b/paddle/legacy/gserver/layers/AverageLayer.cpp @@ -0,0 +1,67 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "AverageLayer.h" + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(average, AverageLayer); + +bool AverageLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + SequencePoolLayer::init(layerMap, parameterMap); + + // average strategy + if (config_.average_strategy() == "average") { + mode_ = kAverage; + } else if (config_.average_strategy() == "sum") { + mode_ = kSum; + } else if (config_.average_strategy() == "squarerootn") { + mode_ = kAverageSquareRootN; + } else { + LOG(FATAL) << "Unknown average strategy: " << config_.average_strategy(); + } + return true; +} + +void AverageLayer::forward(PassType passType) { + SequencePoolLayer::forward(passType); + + MatrixPtr inputValue = getInputValue(0); + getOutputValue()->sequenceAvgForward( + *inputValue, *startPositions_->getVector(useGpu_), mode_); + + /* add the bias-vector AFTER average operation */ + if (biases_.get() != NULL) { + MatrixPtr outV = getOutputValue(); + outV->addBias(*(biases_->getW()), 1); + } + + /* activation */ { forwardActivation(); } +} + +void AverageLayer::backward(const UpdateCallback& callback) { + SequencePoolLayer::backward(callback); + + if (getInputGrad(0)) { + getInputGrad(0)->sequenceAvgBackward( + *getOutputGrad(), *startPositions_->getVector(useGpu_), mode_); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/AverageLayer.h b/paddle/legacy/gserver/layers/AverageLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..a0d457d35f4bce99860cf45e94525f323f45e286 --- /dev/null +++ b/paddle/legacy/gserver/layers/AverageLayer.h @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "SequencePoolLayer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * A layer for "internal average" for sequence input. + * Input: one or more sequences. Each sequence contains some instances. + * If SequenceLevel = kNonSeq: + * Output: output size is the number of input sequences (NOT input instances) + * output[i] = average_{for each instance in this sequence}{input[i]} + * If stride_ > 0: + * Output: a shorten sequence. Stride is the step size by which we slide a + * window upon the input sequence, and the average pooling + * operation is then applied to each interval independently. + * If SequenceLevel = kSeq: + * Check input sequence must has sub-sequence + * Output: output size is the number of input sub-sequences + * output[i] = average_{for each instance in this sub-sequence}{input[i]} + * + * The config file api is pooling_layer. + */ +class AverageLayer : public SequencePoolLayer { + public: + enum AverageStrategy { kAverage = 0, kSum = 1, kAverageSquareRootN = 2 }; + explicit AverageLayer(const LayerConfig& config) + : SequencePoolLayer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + protected: + int mode_; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/BatchNormBaseLayer.cpp b/paddle/legacy/gserver/layers/BatchNormBaseLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4dcbd8dc270d5e5329b33b366ac937894833085f --- /dev/null +++ b/paddle/legacy/gserver/layers/BatchNormBaseLayer.cpp @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "BatchNormBaseLayer.h" +#include "BatchNormalizationLayer.h" +#include "Layer.h" +#include "paddle/legacy/utils/Stat.h" +#ifdef PADDLE_WITH_CUDA +#include "CudnnBatchNormLayer.h" +#endif + +namespace paddle { + +bool BatchNormBaseLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + if (!Layer::init(layerMap, parameterMap)) return false; + + /* initialize the weightList */ + // first is Input in configure + // other two is created in config_parser.py + CHECK_EQ(inputLayers_.size(), 3U); + CHECK_EQ(inputLayers_.size(), parameters_.size()); + CHECK_EQ(inputLayers_.size(), size_t(config_.inputs_size())); + const ImageConfig& conf = config_.inputs(0).image_conf(); + channels_ = conf.channels(); + calFeatureMapSize(); + + if (config_.has_use_global_stats()) { + useGlobalStats_ = config_.use_global_stats(); + } + movingAvgFraction_ = config_.moving_average_fraction(); + epsilon_ = config_.epsilon(); + + weight_.reset(new Weight(1, channels_, parameters_[0])); + movingMean_.reset(new Weight(1, channels_, parameters_[1])); + movingVar_.reset(new Weight(1, channels_, parameters_[2])); + + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, channels_, biasParameter_)); + } + + savedMean_ = Matrix::create(1, channels_, false, useGpu_); + savedInvVar_ = Matrix::create(1, channels_, false, useGpu_); + savedMean_->zeroMem(); + savedInvVar_->zeroMem(); + + return true; +} + +void BatchNormBaseLayer::calFeatureMapSize() { + const ImageConfig& conf = config_.inputs(0).image_conf(); + imageH_ = inputLayers_[0]->getOutput().getFrameHeight(); + imageW_ = inputLayers_[0]->getOutput().getFrameWidth(); + imageD_ = inputLayers_[0]->getOutput().getFrameDepth(); + + if (0 == imageD_) imageD_ = conf.img_size_z(); + if (imageH_ == 0 && imageW_ == 0) { + imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + imageW_ = conf.img_size(); + } else { + getOutput().setFrameHeight(imageH_); + getOutput().setFrameWidth(imageW_); + getOutput().setFrameDepth(imageD_); + } + imgPixels_ = imageH_ * imageW_ * imageD_; +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/BatchNormBaseLayer.h b/paddle/legacy/gserver/layers/BatchNormBaseLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..8dc1d7883767b4aabc8501531996036c2def9481 --- /dev/null +++ b/paddle/legacy/gserver/layers/BatchNormBaseLayer.h @@ -0,0 +1,101 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief Batch normalization layer use to normalizes the input to across the + * batch. + * + * By default, calculating global mean and variance statistics via a running + * average in the training peroid. Then the pre-calculated global mean and + * variance are used for testing. + * + * Moving mean and variance are located in Parameter object when constructing + * and the calculation will change them. Now we only save global mean and + * variance of one thread in first node for GPU. + * But the calculation in CPU is different, because parameters are shared by + * multiple threads. Here using ShareCpuMatrix with lock to calculate. We + * still save global mean and variance in first node in CPU when multi machine. + * + * [1] S. Ioffe and C. Szegedy, "Batch Normalization: Accelerating Deep Network + * Training by Reducing Internal Covariate Shift." arXiv preprint + * arXiv:1502.03167 (2015). + */ + +class BatchNormBaseLayer : public Layer { + public: + explicit BatchNormBaseLayer(const LayerConfig& config) : Layer(config) {} + + ~BatchNormBaseLayer() {} + + /** + * @brief Create BatchNorm layer by norm_type, including batch_norm and + * cudnn_batch_norm. If do not set norm_type, it will automatically select + * cudnn_batch_norm for GPU and batch_norm for CPU. + */ + static Layer* create(const LayerConfig& config); + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + /** + * @brief Calculate feature map size. Some input uses frameHeight and + * frameWidth to store feature size + */ + void calFeatureMapSize(); + + protected: + /// Batch normalization scale parameter, which is referred to as gamma in + /// in original paper. + std::unique_ptr weight_; + /// Moving average of mean. + std::unique_ptr movingMean_; + /// Moving average of variance. + std::unique_ptr movingVar_; + /// Batch normalization bias parameter, which is referred to as beta in + /// in original paper. + std::unique_ptr biases_; + + /// Save intermediate results computed during the forward pass, + /// these can then be reused to speed up the backward pass. + MatrixPtr savedMean_; + MatrixPtr savedInvVar_; + + /// Height or width of input image feature. + /// Both of them are 1 if the input is fully-connected layer. + int imageD_; + int imageH_; + int imageW_; + /// Height * Width. + int imgPixels_; + /// Feature dimension. If the input layer is conv layer, it is the channels + /// of feature map of the conv layer. If the input layer is fully-connected + /// layer, it is the dimension of fc layer. + int channels_; + // if useGlobalStats_ is true, will use the loaded mean and variance. + // otherwise, calculate mean and variance in this mini-batch. + bool useGlobalStats_; + // use to compute moving mean and variance. + real movingAvgFraction_; + // Epsilon is a small random noise used in batch normalization for stability. + real epsilon_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/BatchNormalizationLayer.cpp b/paddle/legacy/gserver/layers/BatchNormalizationLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0297bd44c7b0485f34598f6926e5337da452460d --- /dev/null +++ b/paddle/legacy/gserver/layers/BatchNormalizationLayer.cpp @@ -0,0 +1,266 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Stat.h" +#ifdef PADDLE_WITH_CUDA +#include "hl_batch_transpose.h" +#endif +#include "BatchNormalizationLayer.h" + +namespace paddle { + +REGISTER_LAYER(batch_norm, BatchNormalizationLayer); + +bool BatchNormalizationLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + if (!BatchNormBaseLayer::init(layerMap, parameterMap)) return false; + + return true; +} + +void BatchNormalizationLayer::calMeanAndStd(const MatrixPtr& mat) { + int numSamples = mat->getHeight(); + Matrix::resizeOrCreate(tmpMat_, numSamples, channels_, false, useGpu_); + savedMean_->zeroMem(); + savedMean_->accumulateColSum(*mat); + savedMean_->mulScalar(1.0 / numSamples); // E[x] + + tmpMat_->assign(*mat); + tmpMat_->square2(); + savedInvVar_->zeroMem(); + savedInvVar_->accumulateColSum(*tmpMat_); + savedInvVar_->mulScalar(1.0 / numSamples); // E[x^2] + savedInvVar_->addSquare(*savedMean_, -1.0); // E[x^2] - E^2[x] + + // Variance may be small negative value + // because of the subtraction operation. + // Here using clipping. + savedInvVar_->downClip(real(0.0)); + + calMovingMeanAndVar(); + + savedInvVar_->subScalar(-epsilon_); + savedInvVar_->sqrt2(*savedInvVar_); +} + +void BatchNormalizationLayer::calMovingMeanAndVar() { + // calculating and saving moving mean and variance + auto& movingMean = movingMean_->getW(); + auto& movingVar = movingVar_->getW(); + // movingMean = movingMean * movingAvgFraction_ + // + savedMean_ * (1 - movingAvgFraction_) + movingMean->add(*savedMean_, movingAvgFraction_, 1.0 - movingAvgFraction_); + // movingVar = movingVar * movingAvgFraction_ + // + savedInvVar_ * (1 - movingAvgFraction_) + movingVar->add(*savedInvVar_, movingAvgFraction_, 1.0 - movingAvgFraction_); +} + +void BatchNormalizationLayer::setMeanAndStd() { + savedMean_->copyFrom(*(movingMean_->getW())); + savedInvVar_->copyFrom(*(movingVar_->getW())); + savedInvVar_->downClip(real(0.0)); + + savedInvVar_->subScalar(-epsilon_); + savedInvVar_->sqrt2(*savedInvVar_); +} + +void BatchNormalizationLayer::expandMat(const MatrixPtr& in, MatrixPtr& out) { + CHECK_EQ(in->getWidth(), static_cast(channels_ * imgPixels_)); + CHECK_EQ(out->getWidth(), static_cast(channels_)); + CHECK(!in->isTransposed()); + CHECK(!out->isTransposed()); + if (imgPixels_ == 1) { + out->assign(*in); + return; + } + size_t batchSize = in->getHeight(); + CHECK_EQ(out->getHeight(), batchSize * imgPixels_); + if (useGpu_) { +#ifndef PADDLE_WITH_CUDA + LOG(FATAL) << "paddle is compiled only for cpu"; +#else + batchTranspose( + in->getData(), out->getData(), imgPixels_, channels_, batchSize); +#endif + } else { + for (size_t i = 0; i < batchSize; i++) { + const MatrixPtr inTmp = + Matrix::create(in->getData() + i * imgPixels_ * channels_, + channels_, + imgPixels_, + false, + useGpu_); + MatrixPtr outTmp = + Matrix::create(out->getData() + i * imgPixels_ * channels_, + imgPixels_, + channels_, + false, + useGpu_); + inTmp->transpose(outTmp, false); + } + } +} + +void BatchNormalizationLayer::shrinkMat(const MatrixPtr& in, MatrixPtr& out) { + CHECK_EQ(in->getWidth(), static_cast(channels_)); + CHECK_EQ(out->getWidth(), static_cast(channels_ * imgPixels_)); + size_t batchSize = out->getHeight(); + CHECK(!in->isTransposed()); + CHECK(!out->isTransposed()); + if (imgPixels_ == 1) { + out->assign(*in); + return; + } + CHECK_EQ(in->getHeight(), static_cast(batchSize * imgPixels_)); + if (useGpu_) { +#ifndef PADDLE_WITH_CUDA + LOG(FATAL) << "paddle is compiled only for cpu"; +#else + batchTranspose( + in->getData(), out->getData(), channels_, imgPixels_, batchSize); +#endif + } else { + for (size_t i = 0; i < batchSize; i++) { + const MatrixPtr inTmp = + Matrix::create(in->getData() + i * channels_ * imgPixels_, + imgPixels_, + channels_, + false, + useGpu_); + MatrixPtr outTmp = + Matrix::create(out->getData() + i * imgPixels_ * channels_, + channels_, + imgPixels_, + useGpu_); + inTmp->transpose(outTmp, false); + } + } +} + +void BatchNormalizationLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = getInputValue(0)->getHeight(); + calFeatureMapSize(); + resetOutput(batchSize, getInputValue(0)->getWidth()); + + // for testing in training peroid. + useGlobalStats_ = (passType == PASS_TEST); + if (passType == PASS_TEST && config_.has_use_global_stats()) { + useGlobalStats_ = config_.use_global_stats(); + } + + Matrix::resizeOrCreate( + expandedIn_, batchSize * imgPixels_, channels_, false, useGpu_); + Matrix::resizeOrCreate( + normIn_, batchSize * imgPixels_, channels_, false, useGpu_); + Matrix::resizeOrCreate( + expandedOut_, batchSize * imgPixels_, channels_, false, useGpu_); + expandMat(getInputValue(0), expandedIn_); + + if (useGlobalStats_) { + if (firstTest_) { + setMeanAndStd(); + firstTest_ = false; + } + } else { + calMeanAndStd(expandedIn_); + firstTest_ = true; + } + + normIn_->assign(*expandedIn_); + normIn_->addBias(*savedMean_, -1); // subtract mean. + normIn_->divRowVector(*savedInvVar_); // divide std. + + expandedOut_->assign(*normIn_); + expandedOut_->mulRowVector(*weight_->getW()); // multiple gamma. + if (biases_) { + expandedOut_->addBias(*(biases_->getW()), 1); // add beta. + } + MatrixPtr out = getOutputValue(); + shrinkMat(expandedOut_, out); + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void BatchNormalizationLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + int batchSize = getInputValue(0)->getHeight(); + + Matrix::resizeOrCreate(meanGrad_, 1, channels_, false, useGpu_); + Matrix::resizeOrCreate(stdGrad_, 1, channels_, false, useGpu_); + + Matrix::resizeOrCreate( + expandedInGrad_, batchSize * imgPixels_, channels_, false, useGpu_); + Matrix::resizeOrCreate( + inGrad_, batchSize, imgPixels_ * channels_, false, useGpu_); + Matrix::resizeOrCreate( + normInGrad_, batchSize * imgPixels_, channels_, false, useGpu_); + Matrix::resizeOrCreate( + expandedOutGrad_, batchSize * imgPixels_, channels_, false, useGpu_); + Matrix::resizeOrCreate( + tmpMat_, batchSize * imgPixels_, channels_, false, useGpu_); + Matrix::resizeOrCreate( + tmpGrad_, batchSize * imgPixels_, channels_, false, useGpu_); + + expandMat(getOutputGrad(), expandedOutGrad_); + + // compute derivatives. + if (biases_ && biases_->getWGrad()) { + REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); + biases_->getWGrad()->collectBias(*expandedOutGrad_, 1); + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + if (weight_->getWGrad()) { + tmpMat_->dotMul(*expandedOutGrad_, *normIn_); + weight_->getWGrad()->collectBias(*tmpMat_, 1); + } + + // compute input gradients. + normInGrad_->assign(*expandedOutGrad_); + normInGrad_->mulRowVector(*(weight_->getW())); // multiple gamma. + // normInGrad * (x - \mu)/ \sqrt(\delta^2) + tmpMat_->dotMul(*normInGrad_, *normIn_); + stdGrad_->zeroMem(); + stdGrad_->collectBias(*tmpMat_, -1.0 / (batchSize * imgPixels_)); + tmpGrad_->assign(*normIn_); + tmpGrad_->mulRowVector(*stdGrad_); + + meanGrad_->zeroMem(); + meanGrad_->collectBias(*normInGrad_, -1.0 / (batchSize * imgPixels_)); + + expandedInGrad_->zeroMem(); + expandedInGrad_->add(*normInGrad_, *tmpGrad_); + expandedInGrad_->addRowVector(*meanGrad_); + expandedInGrad_->divRowVector(*savedInvVar_); + + shrinkMat(expandedInGrad_, inGrad_); + if (getInputGrad(0)) { + getInputGrad(0)->add(*getInputGrad(0), *inGrad_); + } + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + weight_->getParameterPtr()->incUpdate(callback); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/BatchNormalizationLayer.h b/paddle/legacy/gserver/layers/BatchNormalizationLayer.h similarity index 100% rename from paddle/gserver/layers/BatchNormalizationLayer.h rename to paddle/legacy/gserver/layers/BatchNormalizationLayer.h diff --git a/paddle/legacy/gserver/layers/BilinearInterpLayer.cpp b/paddle/legacy/gserver/layers/BilinearInterpLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a091f51bc20e219c3111fb07058b5adea5a3fc38 --- /dev/null +++ b/paddle/legacy/gserver/layers/BilinearInterpLayer.cpp @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "BilinearInterpLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(bilinear_interp, BilinearInterpLayer); + +size_t BilinearInterpLayer::getSize() { + inImgH_ = inputLayers_[0]->getOutput().getFrameHeight(); + inImgW_ = inputLayers_[0]->getOutput().getFrameWidth(); + + const BilinearInterpConfig& conf = config_.inputs(0).bilinear_interp_conf(); + if (inImgH_ == 0) { + inImgH_ = conf.image_conf().img_size_y(); + } + if (inImgW_ == 0) { + inImgW_ = conf.image_conf().img_size(); + } + + outImgH_ = conf.out_size_y(); + outImgW_ = conf.out_size_x(); + numChannels_ = conf.image_conf().channels(); + + CHECK(outImgH_ > 0 && outImgW_ > 0); + CHECK(inImgH_ > 0 && inImgW_ > 0); + CHECK(numChannels_); + + ratioH_ = + (outImgH_ > 1) ? static_cast(inImgH_ - 1) / (outImgH_ - 1) : 0.f; + ratioW_ = + (outImgW_ > 1) ? static_cast(inImgW_ - 1) / (outImgW_ - 1) : 0.f; + + getOutput().setFrameHeight(outImgH_); + getOutput().setFrameWidth(outImgW_); + return outImgH_ * outImgW_ * numChannels_; +} + +bool BilinearInterpLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(1, config_.inputs_size()); + + return true; +} + +void BilinearInterpLayer::forward(PassType passType) { + Layer::forward(passType); + + size_t batchSize = getInput(0).getBatchSize(); + size_t size = getSize(); + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + resetOutput(batchSize, size); + } + + MatrixPtr inV = getInputValue(0); + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwBilinearInterpTimer", getName().c_str()); + outV->bilinearForward(*inV, + inImgH_, + inImgW_, + outImgH_, + outImgW_, + numChannels_, + ratioH_, + ratioW_); + } +} + +void BilinearInterpLayer::backward(const UpdateCallback& callback) { + (void)callback; + + MatrixPtr inputG = getInputGrad(0); + MatrixPtr outG = getOutputGrad(); + { + REGISTER_TIMER_INFO("BwBilinearInterpTimer", getName().c_str()); + if (inputG) { + inputG->bilinearBackward(*outG, + outImgH_, + outImgW_, + inImgH_, + inImgW_, + numChannels_, + ratioH_, + ratioW_); + } + } +} +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/BilinearInterpLayer.h b/paddle/legacy/gserver/layers/BilinearInterpLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..c585a5ed10d9c8f241b5a5ff3a671752fda6d432 --- /dev/null +++ b/paddle/legacy/gserver/layers/BilinearInterpLayer.h @@ -0,0 +1,47 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief A layer for bilinear interpolation which is + * used on conv layer output. + * + * @note The config file api is bilinear_interp_layer. + */ +class BilinearInterpLayer : public Layer { + protected: + size_t outImgH_, outImgW_; + size_t inImgH_, inImgW_; + real ratioH_, ratioW_; + size_t numChannels_; + + public: + explicit BilinearInterpLayer(const LayerConfig& config) : Layer(config) {} + + virtual ~BilinearInterpLayer() {} + + size_t getSize(); + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/BlockExpandLayer.cpp b/paddle/legacy/gserver/layers/BlockExpandLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..24b5af67d40958c940eb0864994e7e81464f6c70 --- /dev/null +++ b/paddle/legacy/gserver/layers/BlockExpandLayer.cpp @@ -0,0 +1,121 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "BlockExpandLayer.h" + +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +REGISTER_LAYER(blockexpand, BlockExpandLayer); + +bool BlockExpandLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(config_.inputs_size(), 1); + const BlockExpandConfig& blockConf = config_.inputs(0).block_expand_conf(); + blockH_ = blockConf.block_y(); + blockW_ = blockConf.block_x(); + strideH_ = blockConf.stride_y(); + strideW_ = blockConf.stride_x(); + paddingH_ = blockConf.padding_y(); + paddingW_ = blockConf.padding_x(); + channels_ = blockConf.channels(); + imgSizeH_ = blockConf.img_size_y(); + imgSizeW_ = blockConf.img_size_x(); + + std::vector strides = {(size_t)strideH_, (size_t)strideW_}; + std::vector paddings = {(size_t)paddingH_, (size_t)paddingW_}; + std::vector blocks = {(size_t)blockH_, (size_t)blockW_}; + createFunction(forward_, + "BlockExpand", + FuncConfig() + .set("strides", strides) + .set("paddings", paddings) + .set("blocks", blocks)); + createFunction(backward_, + "BlockExpandGrad", + FuncConfig() + .set("strides", strides) + .set("paddings", paddings) + .set("blocks", blocks)); + + return true; +} + +size_t BlockExpandLayer::getBlockNum() { + CHECK_EQ(inputLayers_.size(), 1UL); + const BlockExpandConfig& blockConf = config_.inputs(0).block_expand_conf(); + imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight(); + imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth(); + if (imgSizeH_ == 0) { + imgSizeH_ = blockConf.img_size_y(); + } + if (imgSizeW_ == 0) { + imgSizeW_ = blockConf.img_size_x(); + } + size_t tmpH = 2 * paddingH_ + imgSizeH_ - blockH_; + outputH_ = (int)tmpH < 0 ? 1 : 1 + (tmpH + strideH_ - 1) / strideH_; + size_t tmpW = 2 * paddingW_ + imgSizeW_ - blockW_; + outputW_ = (int)tmpW < 0 ? 1 : 1 + (tmpW + strideW_ - 1) / strideW_; + + return outputH_ * outputW_; +} + +void BlockExpandLayer::forward(PassType passType) { + Layer::forward(passType); + + size_t batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + size_t blockNum = getBlockNum(); + size_t blockSize = blockH_ * blockW_ * channels_; + resetOutput(blockNum * batchSize, blockSize); + + // calculate output_.value + inputShape_ = TensorShape({batchSize, channels_, imgSizeH_, imgSizeW_}); + outputShape_ = TensorShape({batchSize, blockNum, blockSize}); + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), inputShape_); + outputs.addArg(*getOutputValue(), outputShape_, ASSIGN_TO); + forward_[0]->calc(inputs, outputs); + + // calculate output_.sequenceStartPositions and output_.cpuSequenceDims + Argument& out = getOutput(); + ICpuGpuVector::resizeOrCreate( + out.sequenceStartPositions, batchSize + 1, false); + IVector::resizeOrCreate(out.cpuSequenceDims, 2 * batchSize, false); + int* start = out.sequenceStartPositions->getMutableData(false); + int* dims = out.cpuSequenceDims->getData(); + for (size_t i = 0; i < batchSize; i++) { + start[i] = i * blockNum; + dims[2 * i] = outputH_; + dims[2 * i + 1] = outputW_; + } + start[batchSize] = batchSize * blockNum; +} + +void BlockExpandLayer::backward(const UpdateCallback& callback) { + /* Calculate the input layers error */ + if (getInputGrad(0)) { + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outputShape_); + outputs.addArg(*getInputGrad(0), inputShape_, ADD_TO); + backward_[0]->calc(inputs, outputs); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/BlockExpandLayer.h b/paddle/legacy/gserver/layers/BlockExpandLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..8b90249bfb0958f0081e7c668cd3b38a53c39951 --- /dev/null +++ b/paddle/legacy/gserver/layers/BlockExpandLayer.h @@ -0,0 +1,68 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief Expand feature map to minibatch matrix. + * - matrix width is: blockH_ * blockW_ * channels_ + * - matirx height is: outputH_ * outputW_ + * + * \f[ + * outputH\_ = 1 + (2 * paddingH\_ + imgSizeH\_ - blockH\_ + strideH\_ - 1) / + * strideH\_ \\ + * outputW\_ = 1 + (2 * paddingW\_ + imgSizeW\_ - blockW\_ + strideW\_ - 1) / + * strideW\_ + * \f] + * + * The expand method is the same with ExpandConvLayer, but saved the transposed + * value. After expanding, output_.sequenceStartPositions will store timeline. + * The number of time steps are outputH_ * outputW_ and the dimension of each + * time step is blockH_ * blockW_ * channels_. This layer can be used after + * convolution neural network, and before recurrent neural network. + * + * The config file api is block_expand_layer. + */ +class BlockExpandLayer : public Layer { + protected: + /** + * @brief Calculate outputH_ and outputW_ and return block number which + * actually is time steps. + * @return time steps, outoutH_ * outputW_. + */ + size_t getBlockNum(); + size_t blockH_, blockW_, strideH_, strideW_, paddingH_, paddingW_; + size_t imgSizeH_, imgSizeW_, outputH_, outputW_, channels_; + + TensorShape inputShape_; + TensorShape outputShape_; + + public: + explicit BlockExpandLayer(const LayerConfig& config) : Layer(config) {} + + ~BlockExpandLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/CRFDecodingLayer.cpp b/paddle/legacy/gserver/layers/CRFDecodingLayer.cpp similarity index 100% rename from paddle/gserver/layers/CRFDecodingLayer.cpp rename to paddle/legacy/gserver/layers/CRFDecodingLayer.cpp diff --git a/paddle/gserver/layers/CRFDecodingLayer.h b/paddle/legacy/gserver/layers/CRFDecodingLayer.h similarity index 100% rename from paddle/gserver/layers/CRFDecodingLayer.h rename to paddle/legacy/gserver/layers/CRFDecodingLayer.h diff --git a/paddle/gserver/layers/CRFLayer.cpp b/paddle/legacy/gserver/layers/CRFLayer.cpp similarity index 100% rename from paddle/gserver/layers/CRFLayer.cpp rename to paddle/legacy/gserver/layers/CRFLayer.cpp diff --git a/paddle/gserver/layers/CRFLayer.h b/paddle/legacy/gserver/layers/CRFLayer.h similarity index 100% rename from paddle/gserver/layers/CRFLayer.h rename to paddle/legacy/gserver/layers/CRFLayer.h diff --git a/paddle/gserver/layers/CTCLayer.cpp b/paddle/legacy/gserver/layers/CTCLayer.cpp similarity index 100% rename from paddle/gserver/layers/CTCLayer.cpp rename to paddle/legacy/gserver/layers/CTCLayer.cpp diff --git a/paddle/gserver/layers/CTCLayer.h b/paddle/legacy/gserver/layers/CTCLayer.h similarity index 100% rename from paddle/gserver/layers/CTCLayer.h rename to paddle/legacy/gserver/layers/CTCLayer.h diff --git a/paddle/gserver/layers/ClipLayer.cpp b/paddle/legacy/gserver/layers/ClipLayer.cpp similarity index 100% rename from paddle/gserver/layers/ClipLayer.cpp rename to paddle/legacy/gserver/layers/ClipLayer.cpp diff --git a/paddle/legacy/gserver/layers/ConcatenateLayer.cpp b/paddle/legacy/gserver/layers/ConcatenateLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ce3f2ca950bf87e287163f1cfc8b15d815a68cf4 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConcatenateLayer.cpp @@ -0,0 +1,208 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "Projection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A concatenate layer has multiple input layers. It concatenates rows of + * each input as one row for the output of this layer and apply activation. + */ +class ConcatenateLayer : public Layer { + public: + explicit ConcatenateLayer(const LayerConfig& config) : Layer(config) {} + + ~ConcatenateLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(concat, ConcatenateLayer); + +bool ConcatenateLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + if (!Layer::init(layerMap, parameterMap)) return false; + + CHECK(!biasParameter_); + + return true; +} + +void ConcatenateLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = getInput(0).getBatchSize(); + int size = getSize(); + reserveOutput(batchSize, size); + + const MatrixPtr& out = getOutputValue(); + int offset = 0; + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + const MatrixPtr& in = getInputValue(i); + size_t inSize = in->getWidth(); + out->assignAtOffset(*in, offset); + offset += inSize; + } + CHECK_EQ(size, offset); + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void ConcatenateLayer::backward(const UpdateCallback& callback) { + (void)callback; + + /* Do activation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + const MatrixPtr& out = getOutputGrad(); + int offset = 0; + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + const MatrixPtr& in = getInputGrad(i); + size_t inSize = getInputValue(i)->getWidth(); + if (in) { + in->addAtOffset(*out, offset); + } + offset += inSize; + } +} + +/** + * concat2 layer is like concat layer, but each input layer was + * processed by a Projection. + */ +class ConcatenateLayer2 : public Layer { + public: + explicit ConcatenateLayer2(const LayerConfig& config) : Layer(config) {} + + ~ConcatenateLayer2() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + protected: + std::vector> projections_; + std::vector projOutput_; + std::vector> projCol_; + bool sharedBias_; + std::unique_ptr biases_; +}; + +REGISTER_LAYER(concat2, ConcatenateLayer2); + +bool ConcatenateLayer2::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + if (!Layer::init(layerMap, parameterMap)) return false; + + CHECK_EQ(inputLayers_.size(), parameters_.size()); + projections_.reserve(inputLayers_.size()); + projCol_.reserve(inputLayers_.size()); + projOutput_.resize(inputLayers_.size()); + + size_t startCol = 0; + size_t endCol = 0; + for (size_t i = 0; i < inputLayers_.size(); i++) { + projections_.emplace_back(Projection::create( + config_.inputs(i).proj_conf(), parameters_[i], useGpu_)); + + endCol += projections_[i]->getOutputSize(); + projCol_.push_back(std::make_pair(startCol, endCol)); + startCol = endCol; + } + CHECK_EQ(getSize(), endCol); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + sharedBias_ = config_.shared_biases(); + size_t psize = config_.bias_size(); + biases_ = std::unique_ptr(new Weight(1, psize, biasParameter_)); + } + + return true; +} + +void ConcatenateLayer2::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = getInput(0).getBatchSize(); + int size = getSize(); + resetOutput(batchSize, size); + + for (size_t i = 0; i < projections_.size(); i++) { + size_t startCol = projCol_[i].first; + size_t endCol = projCol_[i].second; + projOutput_[i].value = output_.value->subColMatrix(startCol, endCol); + if (output_.grad) { + projOutput_[i].grad = output_.grad->subColMatrix(startCol, endCol); + } + } + + { + AsyncGpuBlock block; + for (size_t i = 0; i != inputLayers_.size(); ++i) { + projections_[i]->forward(&getInput(i), &projOutput_[i], passType); + } + } + + /* add the bias-vector */ + if (biases_) { + REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); + output_.value->addBias(*(biases_->getW()), 1, sharedBias_); + } + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void ConcatenateLayer2::backward(const UpdateCallback& callback) { + /* Do activation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + AsyncGpuBlock block; + if (biases_ && biases_->getWGrad()) { + REGISTER_TIMER_INFO("Concat2BpBiasTimer", getName().c_str()); + biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBias_); + biases_->getParameterPtr()->incUpdate(callback); + } + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (projections_[i]) { + projections_[i]->backward(callback); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ContextProjection.cpp b/paddle/legacy/gserver/layers/ContextProjection.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8bcf32663eb381a7d7700270efcaa08f9ff86356 --- /dev/null +++ b/paddle/legacy/gserver/layers/ContextProjection.cpp @@ -0,0 +1,185 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ContextProjection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_PROJECTION(context, ContextProjection); + +ContextProjection::ContextProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu) + : Projection(config, parameter, useGpu) { + CHECK(config.has_context_start()); + CHECK(config.has_context_length()); + if (config.context_start() == 0 && config.context_length() == 1) { + config_.set_trainable_padding(false); + } + if (config_.trainable_padding()) { + CHECK(parameter); + beginPad_ = std::max(0, -config.context_start()); + endPad_ = std::max(0, config.context_start() + config.context_length() - 1); + size_t totalPad = beginPad_ + endPad_; + size_t inputDim = parameter->getSize() / totalPad; + CHECK_EQ(config.input_size(), inputDim); + CHECK_EQ(inputDim * totalPad, parameter->getSize()); + weight_.reset(new Weight(totalPad, inputDim, parameter)); + } + // init forward_ and backward_ functions + init(); +} + +bool ContextProjection::init() { + size_t context_length = config_.context_length(); + int context_start = config_.context_start(); + bool is_padding = config_.trainable_padding(); + size_t total_pad = is_padding ? beginPad_ + endPad_ : 0; + + createFunction(forward_, + "ContextProjectionForward", + FuncConfig() + .set("context_length", context_length) + .set("context_start", context_start) + .set("begin_pad", beginPad_)); + createFunction(backward_, + "ContextProjectionBackward", + FuncConfig() + .set("context_length", context_length) + .set("context_start", context_start) + .set("begin_pad", beginPad_) + .set("is_padding", is_padding) + .set("total_pad", total_pad)); + + return true; +} + +void ContextProjection::resetState() { + CHECK_LE(config_.context_start() + config_.context_length(), 1) + << "state is not allowed for future context"; + if (config_.context_start() >= 0) return; + Matrix::resizeOrCreate(state_, + -config_.context_start(), + config_.input_size(), + false, // trans + useGpu_); + Matrix::resizeOrCreate(state2_, + -config_.context_start(), + config_.input_size(), + false, // trans + useGpu_); + if (config_.trainable_padding()) { + state_->assign(*weight_->getW()->subMatrix(0, -config_.context_start())); + } else { + state_->zeroMem(); + } +} + +void ContextProjection::setState(LayerStatePtr state) { + CHECK(state->value.size() == 1) + << "one matrix is expected for ContextProjection state"; + state_->copyFrom(*(state->value[0])); +} + +LayerStatePtr ContextProjection::getState() { + if (state_ == nullptr) { + return nullptr; + } + LayerStatePtr res = std::make_shared(); + res->value.push_back(state_->clone(0, 0, false)); + res->value[0]->copyFrom(*state_); + return res; +} + +void ContextProjection::forward() { + CHECK(in_->value && out_->value); + CHECK(in_->sequenceStartPositions); + + size_t input_dim = in_->value->getWidth(); + size_t dim = out_->value->getWidth(); + CHECK_EQ(dim, input_dim * config_.context_length()); + // size_t batch_size = in_->value->getHeight(); + CHECK_EQ(forward_.size(), (size_t)1) << "Only one forward function here"; + + REGISTER_TIMER_INFO("ContextProjectionForward", getName().c_str()); + bool is_padding = config_.trainable_padding(); + /// first use state_, otherwise use weight_(padding false === w nullptr) + auto w_ptr = + state_ ? state_.get() : is_padding ? weight_->getW().get() : nullptr; + const auto start_pos = in_->sequenceStartPositions->getVector(useGpu_); + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*in_->value, *start_pos); + if (w_ptr) { + inputs.addArg(CpuMatrix(w_ptr->getData(), w_ptr->getHeight(), input_dim), + *start_pos); + } + outputs.addArg(*out_->value, *start_pos, ADD_TO); + forward_[0]->calc(inputs, outputs); + + if (state_ && config_.context_start() < 0) { + CHECK_EQ(1, in_->getNumSequences()); + const int* starts = in_->sequenceStartPositions->getData(false); + int length = starts[1] - starts[0]; + if (-config_.context_start() <= length) { + MatrixPtr sub = in_->value->subMatrix(starts[1] + config_.context_start(), + -config_.context_start()); + state_->copyFrom(*sub); + } else { + int prevLength = -config_.context_start() - length; + state2_->subMatrix(0, prevLength) + ->copyFrom(*state_->subMatrix(length, prevLength)); + state2_->subMatrix(prevLength, length) + ->copyFrom(*in_->value->subMatrix(starts[0], length)); + std::swap(state_, state2_); + } + } +} + +void ContextProjection::backward(const UpdateCallback& callback) { + CHECK(in_->value && out_->value && out_->grad); + size_t input_dim = in_->value->getWidth(); + size_t dim = out_->value->getWidth(); + CHECK_EQ(dim, input_dim * config_.context_length()); + size_t batch_size = in_->value->getHeight(); + CHECK_EQ(batch_size, out_->value->getHeight()); + CHECK_EQ(static_cast(backward_.size()), 1) + << "Only one backward function here"; + + REGISTER_TIMER_INFO("ContextProjectionBackward", getName().c_str()); + bool is_padding = config_.trainable_padding(); + auto start_pos = in_->sequenceStartPositions; + auto w_ptr = is_padding ? weight_->getWGrad() : nullptr; + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*out_->grad, *in_->sequenceStartPositions->getVector(useGpu_)); + outputs.addArg( + CpuMatrix( + in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim), + *in_->sequenceStartPositions->getVector(useGpu_), + ADD_TO); + outputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr, + w_ptr ? w_ptr->getHeight() : 0, + input_dim), + ADD_TO); + backward_[0]->calc(inputs, outputs); + + if (config_.trainable_padding()) { + weight_->getParameterPtr()->incUpdate(callback); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/ContextProjection.h b/paddle/legacy/gserver/layers/ContextProjection.h similarity index 100% rename from paddle/gserver/layers/ContextProjection.h rename to paddle/legacy/gserver/layers/ContextProjection.h diff --git a/paddle/legacy/gserver/layers/Conv3DLayer.cpp b/paddle/legacy/gserver/layers/Conv3DLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d072a74234b43e06c1194acc2ec2b3f961b4a97e --- /dev/null +++ b/paddle/legacy/gserver/layers/Conv3DLayer.cpp @@ -0,0 +1,253 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Conv3DLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(conv3d, Conv3DLayer); + +bool Conv3DLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; + int index = 0; + for (auto &inputConfig : config_.inputs()) { + const ConvConfig &conf = inputConfig.conv_conf(); + M_.push_back(numFilters_ / conf.groups()); + K_.push_back(filterPixels_[index] * filterChannels_[index]); + + // create a new weight + size_t height, width; + width = filterPixels_[index] * filterChannels_[index]; + height = numFilters_; + CHECK_EQ(parameters_[index]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[index]); + weights_.emplace_back(w); + ++index; + } + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } + return true; +} + +size_t Conv3DLayer::getSize() { + CHECK_NE(inputLayers_.size(), 0UL); + outputH_.clear(); + outputW_.clear(); + outputD_.clear(); + N_.clear(); + size_t layerSize = 0; + for (size_t i = 0; i < inputLayers_.size(); ++i) { + outputW_.push_back(outputSize( + imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true)); + outputH_.push_back(outputSize( + imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true)); + outputD_.push_back(outputSize( + imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true)); + + N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]); + CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize); + layerSize += N_[i] * numFilters_; + } + getOutput().setFrameHeight(outputH_[0]); + getOutput().setFrameWidth(outputW_[0]); + getOutput().setFrameDepth(outputD_[0]); + return layerSize; +} + +void Conv3DLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + int outWidth = getSize(); + resetOutput(batchSize, outWidth); + + REGISTER_TIMER_INFO("FwdConv3D", getName().c_str()); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + const MatrixPtr &inMat = getInputValue(i); + const MatrixPtr &outMat = getOutputValue(); + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + MatrixPtr wMat = weights_[i]->getW(); + for (int n = 0; n < batchSize; ++n) { + colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(), + channels_[i], + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i]); + + real *outData = outMat->getData() + n * outMat->getStride(); + MatrixPtr outMatSub = + Matrix::create(outData, groups_[i] * M, N, false, useGpu_); + for (int g = 0; g < groups_[i]; g++) { + MatrixPtr wMatSub = wMat->subMatrix(g * M, M); + MatrixPtr in = colBuf_->subMatrix(g * K, K); + MatrixPtr out = outMatSub->subMatrix(g * M, M); + out->mul(*wMatSub, *in, 1.0, 1.0); + } + } + } + if (nullptr != this->biasParameter_) { + this->addBias(); + } + forwardActivation(); +} + +void Conv3DLayer::backward(const UpdateCallback &callback) { + backwardActivation(); + + if (biases_ && biases_->getWGrad()) { + bpropBiases(); + biases_->getParameterPtr()->incUpdate(callback); + } + + REGISTER_TIMER_INFO("BwdConv3D", getName().c_str()); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (weights_[i]->getWGrad()) { + bpropWeights(i); + } + if (getInputGrad(i)) { + bpropData(i); + } + weights_[i]->getParameterPtr()->incUpdate(callback); + } +} + +void Conv3DLayer::bpropWeights(int i) { + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + const MatrixPtr &inMat = getInputValue(i); + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + MatrixPtr wGradMat = weights_[i]->getWGrad(); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + for (int n = 0; n < batchSize; ++n) { + colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(), + channels_[i], + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i]); + + real *outGradData = + getOutputGrad()->getData() + n * getOutputGrad()->getStride(); + MatrixPtr outGradSub = + Matrix::create(outGradData, groups_[i] * M, N, false, useGpu_); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr inMatSub = colBuf_->subMatrix(g * K, K); + MatrixPtr outG = outGradSub->subMatrix(g * M, M); + MatrixPtr wGradSub = wGradMat->subMatrix(g * M, M); + wGradSub->mul(*outG, *(inMatSub->getTranspose()), 1.0, 1.0); + } + } +} + +void Conv3DLayer::bpropData(int i) { + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + MatrixPtr wMat = weights_[i]->getW(); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + for (int n = 0; n < batchSize; ++n) { + real *outGradData = + getOutputGrad()->getData() + n * getOutputGrad()->getStride(); + real *preGradData = + getInputGrad(i)->getData() + n * getInputGrad(i)->getStride(); + MatrixPtr outGradSub = + Matrix::create(outGradData, M * groups_[i], N, false, useGpu_); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr wMatSub = wMat->subMatrix(g * M, M); + MatrixPtr outG = outGradSub->subMatrix(g * M, M); + MatrixPtr inGradMatSub = colBuf_->subMatrix(g * K, K); + inGradMatSub->mul(*(wMatSub->getTranspose()), *outG, 1.0, 0.0); + } + colBuf_->col2Vol(preGradData, + channels_[i], + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i], + 1.0, + 1.0); + } +} + +void Conv3DLayer::bpropBiases() { + MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), + 1, + biases_->getWGrad()->getElementCnt(), + false, + useGpu_); + MatrixPtr outGradMat = getOutputGrad(); + + if (this->sharedBiases_) { + biases->collectSharedBias(*outGradMat, 1.0f); + } else { + biases->collectBias(*outGradMat, 1.0f); + } +} + +void Conv3DLayer::addBias() { + MatrixPtr outMat = getOutputValue(); + MatrixPtr bias = Matrix::create(biases_->getW()->getData(), + 1, + biases_->getW()->getElementCnt(), + false, + useGpu_); + if (this->sharedBiases_) { + outMat->addSharedBias(*(bias), 1.0f); + } else { + outMat->addBias(*(bias), 1.0f); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/Conv3DLayer.h b/paddle/legacy/gserver/layers/Conv3DLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..cb42a2f36d31365b473d7f593fd27dc063c83c47 --- /dev/null +++ b/paddle/legacy/gserver/layers/Conv3DLayer.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "ConvBaseLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief A subclass of convolution layer. + * This layer expands input and use matrix multiplication to + * calculate convolution operation. + */ +class Conv3DLayer : public ConvBaseLayer { + public: + explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} + ~Conv3DLayer() {} + + bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + + void forward(PassType passType); + void addBias(); + void backward(const UpdateCallback& callback); + void bpropBiases(); + void bpropData(int i); + void bpropWeights(int i); + size_t getSize(); + + protected: + // Figure out the dimensions for individual gemms. + IntV M_; /// numFilters_ / filter_group_; + IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ + IntV K_; /// outputD_ * outputH_ * outputW_ + MatrixPtr colBuf_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvBaseLayer.cpp b/paddle/legacy/gserver/layers/ConvBaseLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..76120915e48661a9b14fb6b9bb99e9ec9dd71e4b --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvBaseLayer.cpp @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvBaseLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/Logging.h" +namespace paddle { + +bool ConvBaseLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + isDeconv_ = (config_.type() == "exconv" || config_.type() == "cudnn_conv") + ? false + : true; + + /* Initialize the convolutional layer parameter */ + numFilters_ = config_.num_filters(); + sharedBiases_ = config_.shared_biases(); + for (auto& inputConfig : config_.inputs()) { + const ConvConfig& conf = inputConfig.conv_conf(); + padding_.push_back(conf.padding()); + stride_.push_back(conf.stride()); + dilation_.push_back(conf.dilation()); + filterSize_.push_back(conf.filter_size()); + paddingY_.push_back(conf.padding_y()); + strideY_.push_back(conf.stride_y()); + dilationY_.push_back(conf.dilation_y()); + filterSizeY_.push_back(conf.filter_size_y()); + channels_.push_back(conf.channels()); + imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y() + : conf.img_size()); + imgSizeW_.push_back(conf.img_size()); + groups_.push_back(conf.groups()); + filterChannels_.push_back(conf.filter_channels()); + outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x()); + outputW_.push_back(conf.output_x()); + + paddingZ_.push_back(conf.padding_z()); + strideZ_.push_back(conf.stride_z()); + filterSizeZ_.push_back(conf.filter_size_z()); + imgSizeD_.push_back(conf.img_size_z()); + outputD_.push_back(conf.output_z()); + filterPixels_.push_back(filterSize_.back() * filterSizeY_.back() * + filterSizeZ_.back()); + } + + CHECK(inputLayers_.size() == parameters_.size()); + + // create new weights_ in derived class + // create new biases_ in derived class + + // default caffe model + caffeMode_ = true; + + return true; +} + +size_t ConvBaseLayer::calOutputSize() { + auto clearAndReserve = [this](IntV* vec) { + vec->clear(); + vec->reserve(this->inputLayers_.size()); + }; + clearAndReserve(&imgSizeH_); + clearAndReserve(&imgSizeW_); + clearAndReserve(&outputH_); + clearAndReserve(&outputW_); + size_t layerSize = 0; + + auto setLayerSize = [&](IntV& inH, IntV& inW, IntV& outH, IntV& outW) { + size_t filterSizeY; + size_t filterSize; + for (size_t i = 0; i < inputLayers_.size(); i++) { + filterSizeY = (filterSizeY_[i] - 1) * dilationY_[i] + 1; + filterSize = (filterSize_[i] - 1) * dilation_[i] + 1; + inH.push_back(inputLayers_[i]->getOutput().getFrameHeight()); + inW.push_back(inputLayers_[i]->getOutput().getFrameWidth()); + const ConvConfig& conf = config_.inputs(i).conv_conf(); + if (isDeconv_) { + if (inH[i] == 0) + inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x(); + if (inW[i] == 0) inW[i] = conf.output_x(); + outH.push_back(imageSize( + inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); + outW.push_back( + imageSize(inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); + } else { + if (inH[i] == 0) + inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + if (inW[i] == 0) inW[i] = conf.img_size(); + outH.push_back(outputSize( + inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); + outW.push_back(outputSize( + inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); + } + CHECK_EQ(outH[i], outH[0]); + CHECK_EQ(outW[i], outW[0]); + } + getOutput().setFrameHeight(outH[0]); + getOutput().setFrameWidth(outW[0]); + layerSize = outH[0] * outW[0] * size_t(numFilters_); + }; + + setLayerSize(imgSizeH_, imgSizeW_, outputH_, outputW_); + + return layerSize; +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvBaseLayer.h b/paddle/legacy/gserver/layers/ConvBaseLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..01e90e999625f986b0f13d2b73a883297c097841 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvBaseLayer.h @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/MathUtils.h" +namespace paddle { + +/** + * @brief A Base Convolution Layer, which convolves the input image + * with learned filters and (optionally) adds biases. + */ + +class ConvBaseLayer : public Layer { + protected: + typedef std::vector IntV; + + /// True if it's deconv layer, false if it's convolution layer + bool isDeconv_; + + /// The number of filters. + int numFilters_; + /// The x dimension of the padding. + IntV padding_; + /// The y dimension of the padding. + IntV paddingY_; + /// The x dimension of the stride. + IntV stride_; + /// The y dimension of the stride. + IntV strideY_; + /// The x dimension of the dilation. + IntV dilation_; + /// The y dimension of the dilation. + IntV dilationY_; + /// The x dimension of a filter kernel. + IntV filterSize_; + /// The y dimension of a filter kernel. + IntV filterSizeY_; + /// The spatial dimensions of the convolution input. + IntV channels_; + /// The spatial dimensions of input feature map height. + IntV imgSizeH_; + /// The spatial dimensions of input feature map width. + IntV imgSizeW_; + /// filterPixels_ = filterSizeX_ * filterSizeY_. + IntV filterPixels_; + /// filterChannels_ = channels_/groups_. + IntV filterChannels_; + /// The spatial dimensions of output feature map height. + IntV outputH_; + /// The spatial dimensions of output feature map width. + IntV outputW_; + + IntV outputD_; + IntV imgSizeD_; + IntV filterSizeZ_; + IntV strideZ_; + IntV paddingZ_; + + /// Group size, refer to grouped convolution in + /// Alex Krizhevsky's paper: when group=2, the first half of the + /// filters are only connected to the first half of the input channels, + /// and the second half only connected to the second half. + IntV groups_; + /// Whether the bias is shared for feature in each channel. + bool sharedBiases_; + + /// shape of weight: (numChannels * filterPixels_, numFilters) + WeightList weights_; + /// If shared_biases is false shape of bias: (numFilters_, 1) + /// If shared_biases is ture shape of bias: + /// (numFilters_ * outputX * outputY, 1) + std::unique_ptr biases_; + + /// True by default. The only difference is the calculation + /// of output size. + bool caffeMode_; + + public: + explicit ConvBaseLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + /** + * imgSizeH_ and imgSizeW_ will be set according to the previous input layers + * in this function. Then it will calculate outputH_ and outputW_ and set them + * into output argument. + */ + virtual size_t calOutputSize(); + + Weight& getWeight(int idx) { return *weights_[idx]; } +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvBaseOperator.cpp b/paddle/legacy/gserver/layers/ConvBaseOperator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e8e59b3bfe9d8a9e54e5c11906707d10ec346a4d --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvBaseOperator.cpp @@ -0,0 +1,151 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvBaseOperator.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief ConvBaseOperator takes two inputs to perform the convolution. + * The first input is the image, and the second input is the convolution kernel. + * The height of data for two inputs are the same. Each data of the first input + * is convolved with each data of the second input indepedently. + * + * The config file api is conv_operator. + */ + +ConvBaseOperator::ConvBaseOperator(const OperatorConfig &config, bool useGpu) + : Operator(config, useGpu) { + CHECK(useGpu); + CHECK_EQ(config_.input_indices_size(), 2L); + + caffeMode_ = true; + getConvParams(); + computeConvSizes(); + + // initialize all to default algorithms + fwdAlgo_ = 0; + bwdFilterAlgo_ = 0; + bwdDataAlgo_ = 0; + fwdLimitBytes_ = 0; + bwdDataLimitBytes_ = 0; + bwdFilterLimitBytes_ = 0; + workSpaceInBytes_ = 0; + workSpace_ = nullptr; + + isSelectAlgo_ = false; +} + +void ConvBaseOperator::allocConvWorkSpace() { + hl_conv_workspace(imageDesc_, + outputDesc_, + filterDesc_, + convDesc_, + &fwdAlgo_, + &fwdLimitBytes_, + &bwdDataAlgo_, + &bwdDataLimitBytes_, + &bwdFilterAlgo_, + &bwdFilterLimitBytes_, + /*useDilation*/ false); + + size_t maxWorkSpace = 0; + maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); + maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_); + + if (maxWorkSpace > workSpaceInBytes_) { + if (workSpaceInBytes_ != 0) { + hl_free_mem_device(workSpace_); + } + // total amount of storage needed + workSpace_ = hl_malloc_device(maxWorkSpace); + workSpaceInBytes_ = maxWorkSpace; + } +} + +void ConvBaseOperator::computeConvSizes() { + hl_create_filter_descriptor( + &filterDesc_, channels_, numFilters_, filterSizeY_, filterSize_); + hl_create_tensor_descriptor(&imageDesc_); + hl_create_tensor_descriptor(&outputDesc_); + hl_create_convolution_descriptor(&convDesc_, + imageDesc_, + filterDesc_, + paddingY_, + padding_, + strideY_, + stride_); +} + +void ConvBaseOperator::reshapeImageDescriptors() { + hl_tensor_reshape(imageDesc_, + 1, + channels_, + imageH_, + imageW_, + channels_ * imageH_ * imageW_, + imageH_ * imageW_, + imageW_, + 1); + hl_tensor_reshape(outputDesc_, + 1, + numFilters_, + outputH_, + outputW_, + numFilters_ * outputH_ * outputW_, + outputH_ * outputW_, + outputW_, + 1); + hl_reset_convolution_descriptor(convDesc_, + imageDesc_, + filterDesc_, + paddingY_, + padding_, + strideY_, + stride_); +} + +void ConvBaseOperator::getConvParams() { + configNumFilters_ = config_.num_filters(); + const ConvConfig &conf = config_.conv_conf(); + padding_ = conf.padding(); + stride_ = conf.stride(); + filterSize_ = conf.filter_size(); + paddingY_ = conf.padding_y(); + strideY_ = conf.stride_y(); + filterSizeY_ = conf.filter_size_y(); + filterPixels_ = filterSize_ * filterSizeY_; + configChannels_ = conf.channels(); + imgSize_ = conf.img_size(); + imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + imgPixels_ = imgSize_ * imgSizeY_; + CHECK_EQ(conf.groups(), 1U); + filterChannels_ = conf.filter_channels(); + outputX_ = conf.output_x(); + outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + outputs_ = outputX_ * outputX_; + + isDeconv_ = (config_.type() == "conv") ? false : true; + if (isDeconv_) { + channels_ = configNumFilters_; + numFilters_ = configChannels_; + } else { + channels_ = configChannels_; + numFilters_ = configNumFilters_; + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvBaseOperator.h b/paddle/legacy/gserver/layers/ConvBaseOperator.h new file mode 100644 index 0000000000000000000000000000000000000000..4ac77f2d743abd6f01e8e3f1e2f4e730c0e6fb39 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvBaseOperator.h @@ -0,0 +1,112 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#pragma once + +#include "Operator.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief ConvOperator takes two inputs to perform the convolution. + * The first input is the image, and the second input is the convolution kernel. + * The height of data for two inputs are the same. Each data of the first input + * is convolved with each data of the second input indepedently. + * + * The config file api is conv_operator. + */ + +class ConvBaseOperator : public Operator { + public: + ConvBaseOperator(const OperatorConfig &config, bool useGpu); + /** + * Free workspace in device and destroy cudnn tensor descriptor. + */ + virtual ~ConvBaseOperator() { + if (workSpaceInBytes_ != 0) { + hl_free_mem_device(workSpace_); + workSpaceInBytes_ = 0; + } + + hl_destroy_tensor_descriptor(imageDesc_); + hl_destroy_tensor_descriptor(outputDesc_); + hl_destroy_filter_descriptor(filterDesc_); + hl_destroy_convolution_descriptor(convDesc_); + } + + protected: + /** + * Get convolution parameters from layer config and + * initialize member variables. + */ + void getConvParams(); + + /** + * Allocate Gpu Memory for cudnn convolution algorithms. + */ + void allocConvWorkSpace(); + + /** + * Create cudnn tensor descriptor for convolution operation. + */ + void computeConvSizes(); + + /** + * Reshape cudnn tensor descriptor. + */ + void reshapeImageDescriptors(); + + /** + * Reshape cudnn tensor descriptor. + */ + virtual void reshape(int batchSize) = 0; + + /** + * Check filter size is equal to the size calculated by parameters from + * layer config. + */ + void checkFilterSize(const MatrixPtr &filter) { + CHECK_EQ(static_cast(filter->getWidth()), + filterSize_ * filterSizeY_ * channels_ * numFilters_); + } + + /// Most of member variables are same with CudnnConvLayer. + /// There is no explanation here. + bool isDeconv_; + int imageH_, imageW_, outputH_, outputW_; + hl_tensor_descriptor imageDesc_; + hl_tensor_descriptor outputDesc_; + hl_filter_descriptor filterDesc_; + hl_convolution_descriptor convDesc_; + bool caffeMode_; + int inputOffset_, outputOffset_, weightOffset_; + int numFilters_, channels_; + + /// from parsing config + int configNumFilters_, configChannels_; + int padding_, stride_, filterSize_, imgSize_, imgSizeY_; + int paddingY_, strideY_, filterSizeY_; + int imgPixels_, filterPixels_, filterChannels_, outputX_, outputY_, outputs_; + + /// Following member variables are same with CudnnConvLayer. + /// There is no explanation here. + int fwdAlgo_, bwdFilterAlgo_, bwdDataAlgo_; + size_t fwdLimitBytes_, bwdDataLimitBytes_, bwdFilterLimitBytes_; + size_t workSpaceInBytes_; + void *workSpace_; + bool isSelectAlgo_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvBaseProjection.cpp b/paddle/legacy/gserver/layers/ConvBaseProjection.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ff5d3412de1c2940cdd9dcf9397370153c24b0c6 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvBaseProjection.cpp @@ -0,0 +1,199 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvBaseProjection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +ThreadLocalD> ConvBaseProjection::convMem_; + +ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config, + ParameterPtr parameter, + bool useGpu) + : Projection(config, parameter, useGpu) { + CHECK(useGpu); // only support GPU + getConvParams(); + initCudnn(); + + size_t height = filterH_ * filterW_ * channels_ / groups_; + size_t width = numFilters_; + weight_.reset(new Weight(height, width, parameter)); + weightOffset_ = height * width / groups_; +} + +void ConvBaseProjection::getConvParams() { + const ConvConfig &conf = config_.conv_conf(); + paddingH_ = conf.padding_y(); + paddingW_ = conf.padding(); + + strideH_ = conf.stride_y(); + strideW_ = conf.stride(); + + dilationH_ = conf.dilation_y(); + dilationW_ = conf.dilation(); + CHECK_GT(dilationH_, 0); + CHECK_GT(dilationW_, 0); + + filterH_ = conf.filter_size_y(); + filterW_ = conf.filter_size(); + + configImgH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + configImgW_ = conf.img_size(); + + configOutH_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + configOutW_ = conf.output_x(); + + configChannels_ = conf.channels(); + configNumFilters_ = config_.num_filters(); + + isDeconv_ = (config_.type() == "conv") ? false : true; + + channels_ = (isDeconv_) ? configNumFilters_ : configChannels_; + numFilters_ = (isDeconv_) ? configChannels_ : configNumFilters_; + + groups_ = conf.groups(); + CHECK_EQ(channels_ % groups_, 0); + CHECK_EQ(numFilters_ % groups_, 0); +} + +void ConvBaseProjection::initCudnn() { + hl_create_filter_descriptor(&filterDesc_, + channels_ / groups_, + numFilters_ / groups_, + filterH_, + filterW_); + hl_create_tensor_descriptor(&imageDesc_); + hl_create_tensor_descriptor(&outputDesc_); + hl_create_convolution_descriptor(&convDesc_, + imageDesc_, + filterDesc_, + paddingH_, + paddingW_, + strideH_, + strideW_, + dilationH_, + dilationW_); + + // initialize all to default algorithms + fwdAlgo_ = 0; + bwdFilterAlgo_ = 0; + bwdDataAlgo_ = 0; + fwdLimitBytes_ = 0; + bwdDataLimitBytes_ = 0; + bwdFilterLimitBytes_ = 0; + workSpaceInBytes_ = 0; +} + +void ConvBaseProjection::reshapeTensorDesc(int batchSize) { + // The stride between two consecutive samples in the output of ConvProjection + // may not be numFilters_ * outputH_ * outputW_ (conv) or + // channels_ * imageH_ * imageW_ (deconv) + // for example, in the case of layer ConcatenateLayer2 with two + // ConvProjection, the stride is the output_size of layer ConcatenateLayer2. + // So the calculation of nStride is different from CudnnConvLayer. + size_t nStrideImage, nStrideOutput; + if (isDeconv_) { + nStrideImage = out_->value->getStride(); + nStrideOutput = numFilters_ * outputH_ * outputW_; + } else { + nStrideImage = channels_ * imageH_ * imageW_; + nStrideOutput = out_->value->getStride(); + } + + hl_tensor_reshape(imageDesc_, + batchSize, + channels_ / groups_, + imageH_, + imageW_, + nStrideImage, + imageH_ * imageW_, + imageW_, + 1); + + hl_tensor_reshape(outputDesc_, + batchSize, + numFilters_ / groups_, + outputH_, + outputW_, + nStrideOutput, + outputH_ * outputW_, + outputW_, + 1); + + hl_reset_convolution_descriptor(convDesc_, + imageDesc_, + filterDesc_, + paddingH_, + paddingW_, + strideH_, + strideW_, + dilationH_, + dilationW_); +} + +void ConvBaseProjection::reshape(int batchSize) { + size_t width = calOutputSize(); + CHECK_EQ(width, out_->value->getWidth()); + CHECK_EQ(calInputSize(), in_->value->getWidth()); + + reshapeTensorDesc(batchSize); + bool useDilation = false; + if (dilationH_ > 1 || dilationW_ > 1) { + useDilation = true; + } + hl_conv_workspace(imageDesc_, + outputDesc_, + filterDesc_, + convDesc_, + &fwdAlgo_, + &fwdLimitBytes_, + &bwdDataAlgo_, + &bwdDataLimitBytes_, + &bwdFilterAlgo_, + &bwdFilterLimitBytes_, + useDilation); + + size_t maxWorkSpace = 0; + maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); + maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_); + workSpaceInBytes_ = maxWorkSpace; + + VLOG(3) << getName() << " Fwd / BwdData / BwdFilter algo: " << fwdAlgo_ + << " / " << bwdDataAlgo_ << " / " << bwdFilterAlgo_; +} + +void *ConvBaseProjection::getSpaceBytes(size_t size) { + std::vector &convMem = *convMem_; + if (convMem.empty()) { + int numDevices = hl_get_device_count(); + convMem.resize(numDevices); + } + + int devId = hl_get_device(); + MemoryHandlePtr localMem = convMem[devId]; + if (NULL == localMem || size > localMem->getAllocSize()) { + localMem = std::make_shared(size); + } + return localMem->getBuf(); +} + +ConvBaseProjection::~ConvBaseProjection() { + hl_destroy_tensor_descriptor(imageDesc_); + hl_destroy_tensor_descriptor(outputDesc_); + hl_destroy_filter_descriptor(filterDesc_); + hl_destroy_convolution_descriptor(convDesc_); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvBaseProjection.h b/paddle/legacy/gserver/layers/ConvBaseProjection.h new file mode 100644 index 0000000000000000000000000000000000000000..dcf5ce0f48daac396bab0ec7620303f6c1236fc2 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvBaseProjection.h @@ -0,0 +1,111 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Projection.h" +#include "paddle/legacy/math/MathUtils.h" + +namespace paddle { + +/** + * @brief Base class for ConvProjection and ConvTransProjection. + */ +class ConvBaseProjection : public Projection { + public: + /** + * Constructor. + */ + ConvBaseProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu); + + ~ConvBaseProjection(); + + protected: + void getConvParams(); + void initCudnn(); + + void reshapeTensorDesc(int batchSize); + void reshape(int batchSize); + + virtual size_t calOutputSize() = 0; + virtual size_t calInputSize() = 0; + + static void* getSpaceBytes(size_t size); + + /// True if it's deconv projection layer, false if it's ConvProjection layer + bool isDeconv_; + /// imageH_ and imageW_ / outputH_ and outputW_ + /// is calculated from the input layer. + int imageH_, imageW_; + int outputH_, outputW_; + /// configImgH_ and configImgW_ / configOutH_ and configOutW_ + /// is obtained from config. + int configImgH_, configImgW_; + int configOutH_, configOutW_; + /// channels_ and numFilters_ are defined in terms of convolution semantics + int channels_, numFilters_; + /// configChannels and configNumFilters_ are obtained from config + /// For Conv they are the same as channels_ and numFilters + /// For ConvTrans they are opposite to channels_ and numFilters + int configChannels_, configNumFilters_; + int paddingH_, paddingW_; + int strideH_, strideW_; + int dilationH_, dilationW_; + int filterH_, filterW_; + /// One group offset of input data. + int inputOffset_; + /// One group offset of output data. + int outputOffset_; + /// One group offset of weight. + int weightOffset_; + int groups_; + + /// Cudnn tensor descriptor for input. + hl_tensor_descriptor imageDesc_; + /// Cudnn tensor descriptor for output. + hl_tensor_descriptor outputDesc_; + /// Cudnn tensor descriptor for filter. + hl_filter_descriptor filterDesc_; + /// Cudnn tensor descriptor for a convolution operation. + hl_convolution_descriptor convDesc_; + + /// Record the algorithm for forward convolution, which is obtained by cudnn + /// api to search the best suited algorithm. + int fwdAlgo_; + /// Record the algorithm for computing convolution gradient with respect to + /// filter coefficients. + int bwdFilterAlgo_; + /// Record the algorithm for computing convolution gradient with respect to + /// the output. + int bwdDataAlgo_; + /// Amount of GPU memory needed as workspace to be able to execute a + /// forward convolution with the specified algo. + size_t fwdLimitBytes_; + /// Amount of GPU memory needed as workspace to be able to execute a + /// backwardFilter with the specified algo. + size_t bwdDataLimitBytes_; + /// Amount of GPU memory needed as workspace to be able to execute a + /// backwardData with the specified algo. + size_t bwdFilterLimitBytes_; + /// Size of total work space. + size_t workSpaceInBytes_; + bool bias_; + + std::unique_ptr weight_; + static ThreadLocalD> convMem_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvOperator.cpp b/paddle/legacy/gserver/layers/ConvOperator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5276b2c3920eee923f13a47d40b4498c6846f94b --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvOperator.cpp @@ -0,0 +1,128 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvOperator.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief ConvOperator takes two inputs to perform the convolution. + * The first input is the image, and the second input is the convolution kernel. + * The height of data for two inputs are the same. Each data of the first input + * is convolved with each data of the second input indepedently. + * + * The config file api is conv_operator. + */ + +REGISTER_OPERATOR(conv, ConvOperator); + +void ConvOperator::reshape(int batchSize) { + imageH_ = ins_[0]->getFrameHeight(); + imageW_ = ins_[0]->getFrameWidth(); + if (imageH_ == 0) imageH_ = imgSizeY_; + if (imageW_ == 0) imageW_ = imgSize_; + outputH_ = outputSize(imageH_, filterSizeY_, paddingY_, strideY_, caffeMode_); + outputW_ = outputSize(imageW_, filterSize_, padding_, stride_, caffeMode_); + /// Check that the outputSizes are consistent with config + CHECK_EQ(outputH_, outputY_); + CHECK_EQ(outputW_, outputX_); + out_->setFrameHeight(outputH_); + out_->setFrameWidth(outputW_); + + reshapeImageDescriptors(); + + inputOffset_ = channels_ * imageH_ * imageW_; + outputOffset_ = numFilters_ * outputH_ * outputW_; + weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSizeY_; + + if (!isSelectAlgo_) { + allocConvWorkSpace(); + } + + isSelectAlgo_ = true; +} + +void ConvOperator::forward() { + size_t batchSize = ins_[0]->value->getHeight(); + reshape(batchSize); + CHECK_EQ(ins_[1]->value->getHeight(), batchSize); + checkFilterSize(ins_[1]->value); + Matrix::resizeOrCreate(out_->value, + batchSize, + outputH_ * outputW_ * numFilters_, + false, + useGpu_); + { + AsyncGpuBlock block; + for (size_t batchId = 0; batchId < batchSize; ++batchId) { + real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; + real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; + real *outData = out_->value->getData() + outputOffset_ * batchId; + hl_convolution_forward(imageDesc_, + inputData, + outputDesc_, + outData, + filterDesc_, + wgtData, + convDesc_, + workSpace_, + workSpaceInBytes_, + fwdAlgo_); + } + } +} + +void ConvOperator::backward() { + size_t batchSize = ins_[0]->value->getHeight(); + { + AsyncGpuBlock block; + for (size_t batchId = 0; batchId < batchSize; ++batchId) { + real *outGrad = out_->grad->getData() + outputOffset_ * batchId; + if (ins_[1]->grad) { + real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; + real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId; + hl_convolution_backward_filter(imageDesc_, + inputData, + outputDesc_, + outGrad, + filterDesc_, + weightGrad, + convDesc_, + workSpace_, + workSpaceInBytes_, + bwdFilterAlgo_); + } + + MatrixPtr preGrad = ins_[0]->grad; + if (NULL != preGrad) { + real *inputGrad = preGrad->getData() + inputOffset_ * batchId; + real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; + hl_convolution_backward_data(imageDesc_, + inputGrad, + outputDesc_, + outGrad, + filterDesc_, + wgtData, + convDesc_, + workSpace_, + workSpaceInBytes_, + bwdDataAlgo_); + } + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvOperator.h b/paddle/legacy/gserver/layers/ConvOperator.h new file mode 100644 index 0000000000000000000000000000000000000000..8f31620111c8ff3818d83145e16012d22b067a12 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvOperator.h @@ -0,0 +1,44 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#pragma once + +#include "ConvBaseOperator.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief ConvOperator takes two inputs to perform the convolution. + * The first input is the image, and the second input is the convolution kernel. + * The height of data for two inputs are the same. Each data of the first input + * is convolved with each data of the second input indepedently. + * + * The config file api is conv_operator. + */ + +class ConvOperator : public ConvBaseOperator { + public: + ConvOperator(const OperatorConfig &config, bool useGpu) + : ConvBaseOperator(config, useGpu) {} + /** + * Free workspace in device and destroy cudnn tensor descriptor. + */ + virtual ~ConvOperator() {} + void forward() override; + void backward() override; + void reshape(int batchSize) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvProjection.cpp b/paddle/legacy/gserver/layers/ConvProjection.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b40cdac2587d1fc0fec00801414560d2a27bd34a --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvProjection.cpp @@ -0,0 +1,123 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvProjection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_PROJECTION(conv, ConvProjection); + +size_t ConvProjection::calOutputSize() { + imageH_ = in_->getFrameHeight(); + imageW_ = in_->getFrameWidth(); + if (imageH_ == 0) imageH_ = configImgH_; + if (imageW_ == 0) imageW_ = configImgW_; + outputH_ = outputSize(imageH_, + (filterH_ - 1) * dilationH_ + 1, + paddingH_, + strideH_, + /* caffeMode */ true); + outputW_ = outputSize(imageW_, + (filterW_ - 1) * dilationW_ + 1, + paddingW_, + strideW_, + /* caffeMode */ true); + + const_cast(out_)->setFrameHeight(outputH_); + const_cast(out_)->setFrameWidth(outputW_); + + inputOffset_ = (configChannels_ / groups_) * imageH_ * imageW_; + outputOffset_ = (configNumFilters_ / groups_) * outputH_ * outputW_; + return outputH_ * outputW_ * configNumFilters_; +} + +size_t ConvProjection::calInputSize() { + return static_cast(configChannels_ * imageH_ * imageW_); +} + +void ConvProjection::forward() { + int batchSize = in_->value->getHeight(); + reshape(batchSize); + + void *workSpace = NULL; + if (workSpaceInBytes_ > 0) { + workSpace = getSpaceBytes(workSpaceInBytes_); + } + + for (int g = 0; g < groups_; ++g) { + REGISTER_TIMER_INFO("CudnnConvFwTimer", getName().c_str()); + + real *inputData = in_->value->getData() + g * inputOffset_; + real *wgtData = weight_->getW()->getData() + g * weightOffset_; + real *outData = out_->value->getData() + g * outputOffset_; + hl_convolution_forward(imageDesc_, + inputData, + outputDesc_, + outData, + filterDesc_, + wgtData, + convDesc_, + workSpace, + fwdLimitBytes_, + fwdAlgo_); + } +} + +void ConvProjection::backward(const UpdateCallback &callback) { + REGISTER_TIMER_INFO("CudnnConvBpTimer", getName().c_str()); + + void *workSpace = NULL; + if (workSpaceInBytes_ > 0) { + workSpace = getSpaceBytes(workSpaceInBytes_); + } + + for (int g = 0; g < groups_; ++g) { + real *outGrad = out_->grad->getData() + g * outputOffset_; + if (weight_->getWGrad()) { + real *inputData = in_->value->getData() + g * inputOffset_; + real *weightGrad = weight_->getWGrad()->getData() + g * weightOffset_; + hl_convolution_backward_filter(imageDesc_, + inputData, + outputDesc_, + outGrad, + filterDesc_, + weightGrad, + convDesc_, + workSpace, + bwdFilterLimitBytes_, + bwdFilterAlgo_); + } + + MatrixPtr preGrad = in_->grad; + if (NULL != preGrad) { + real *inputGrad = preGrad->getData() + g * inputOffset_; + real *wgtData = weight_->getW()->getData() + g * weightOffset_; + hl_convolution_backward_data(imageDesc_, + inputGrad, + outputDesc_, + outGrad, + filterDesc_, + wgtData, + convDesc_, + workSpace, + bwdDataLimitBytes_, + bwdDataAlgo_); + } + } + + weight_->getParameterPtr()->incUpdate(callback); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvProjection.h b/paddle/legacy/gserver/layers/ConvProjection.h new file mode 100644 index 0000000000000000000000000000000000000000..890a17e2f8d2d05001f825f374e8ab6420f7b3ea --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvProjection.h @@ -0,0 +1,43 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ConvBaseProjection.h" +#include "paddle/legacy/math/MathUtils.h" + +namespace paddle { + +/** + * @brief Convolution projection do the same calculation with CudnnConvLayer. + */ +class ConvProjection : public ConvBaseProjection { + public: + /** + * Constructor. + */ + ConvProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu) + : ConvBaseProjection(config, parameter, useGpu) {} + + ~ConvProjection() {} + + virtual void forward(); + virtual void backward(const UpdateCallback& callback); + virtual size_t calOutputSize(); + virtual size_t calInputSize(); +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvShiftLayer.cpp b/paddle/legacy/gserver/layers/ConvShiftLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b7ecbe556c59b32cc5833617717b40c730392506 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvShiftLayer.cpp @@ -0,0 +1,108 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief A layer for circular convluation of two vectors, + * which is used in NEURAL TURING MACHINE. + * - Input: two vectors, the first is data (batchSize x dataDim) + * the second is shift weights (batchSize x shiftDim) + * - Output: a vector (batchSize x dataDim) + * Assumed that: + * - a[in]: contains M elements. + * - b[in]: contains N elements (N should be odd). + * - c[out]: contains M elements. + * + * \f[ + * c[i] = \sum_{j=-(N-1)/2}^{(N-1)/2}a_{i+j} * b_{j} + * \f] + * + * In this formula: + * - a's index is computed modulo M. + * - b's index is comupted modulo N. + * + * The config file api is conv_shift_layer. + */ + +class ConvShiftLayer : public Layer { + public: + explicit ConvShiftLayer(const LayerConfig& config) : Layer(config) {} + + ~ConvShiftLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(conv_shift, ConvShiftLayer); + +bool ConvShiftLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2U); + + return true; +} + +void ConvShiftLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV0->getHeight(); + size_t dataDim = inV0->getWidth(); + + CHECK_EQ(batchSize, inV1->getHeight()); + CHECK_EQ(dataDim, getSize()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + resetOutput(batchSize, dataDim); + } + + MatrixPtr outV = getOutputValue(); + + REGISTER_TIMER_INFO("FwConvShiftTimer", getName().c_str()); + outV->circularConv(*inV0, *inV1); +} + +void ConvShiftLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr outG = getOutputGrad(); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + + REGISTER_TIMER_INFO("BwConvShiftTimer", getName().c_str()); + + if (inG0 && inG1) { + outG->circularConvDerivative(*outG, *inV0, *inV1, *inG0, *inG1); + } else { + CHECK(!inG0 || !inG1) << "Not supported"; + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvTransOperator.cpp b/paddle/legacy/gserver/layers/ConvTransOperator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f4ce2affb144152ed41a9d4be9fa87f800c83dbb --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvTransOperator.cpp @@ -0,0 +1,125 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvTransOperator.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief ConvTransOperator takes two inputs to perform the convolution. + * The first input is the image, and the second input is the convolution kernel. + * The height of data for two inputs are the same. Each data of the first input + * is convolved with each data of the second input indepedently. + * + * The config file api is conv_operator. + */ + +REGISTER_OPERATOR(convt, ConvTransOperator); + +void ConvTransOperator::reshape(int batchSize) { + outputH_ = ins_[0]->getFrameHeight(); + outputW_ = ins_[0]->getFrameWidth(); + if (outputH_ == 0) outputH_ = outputY_; + if (outputW_ == 0) outputW_ = outputX_; + imageH_ = imageSize(outputH_, filterSizeY_, paddingY_, strideY_, caffeMode_); + imageW_ = imageSize(outputW_, filterSize_, padding_, stride_, caffeMode_); + /// Check that the imageSizes are consistent with config + CHECK_EQ(imageH_, imgSizeY_); + CHECK_EQ(imageW_, imgSize_); + out_->setFrameHeight(imageH_); + out_->setFrameWidth(imageW_); + + reshapeImageDescriptors(); + + inputOffset_ = numFilters_ * outputH_ * outputW_; + outputOffset_ = channels_ * imageH_ * imageW_; + weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSizeY_; + + if (!isSelectAlgo_) { + allocConvWorkSpace(); + } + + isSelectAlgo_ = true; +} + +void ConvTransOperator::forward() { + size_t batchSize = ins_[0]->value->getHeight(); + reshape(batchSize); + CHECK_EQ(ins_[1]->value->getHeight(), batchSize); + checkFilterSize(ins_[1]->value); + Matrix::resizeOrCreate( + out_->value, batchSize, imageH_ * imageW_ * channels_, false, useGpu_); + { + AsyncGpuBlock block; + for (size_t batchId = 0; batchId < batchSize; ++batchId) { + real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; + real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; + real *outData = out_->value->getData() + outputOffset_ * batchId; + hl_convolution_backward_data(imageDesc_, + outData, + outputDesc_, + inputData, + filterDesc_, + wgtData, + convDesc_, + workSpace_, + workSpaceInBytes_, + bwdDataAlgo_); + } + } +} + +void ConvTransOperator::backward() { + size_t batchSize = ins_[0]->value->getHeight(); + { + AsyncGpuBlock block; + for (size_t batchId = 0; batchId < batchSize; ++batchId) { + real *outGrad = out_->grad->getData() + outputOffset_ * batchId; + if (ins_[1]->grad) { + real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId; + real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId; + hl_convolution_backward_filter(imageDesc_, + outGrad, + outputDesc_, + inputData, + filterDesc_, + weightGrad, + convDesc_, + workSpace_, + workSpaceInBytes_, + bwdFilterAlgo_); + } + + MatrixPtr preGrad = ins_[0]->grad; + if (NULL != preGrad) { + real *inputGrad = preGrad->getData() + inputOffset_ * batchId; + real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId; + hl_convolution_forward(imageDesc_, + outGrad, + outputDesc_, + inputGrad, + filterDesc_, + wgtData, + convDesc_, + workSpace_, + workSpaceInBytes_, + fwdAlgo_); + } + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvTransOperator.h b/paddle/legacy/gserver/layers/ConvTransOperator.h new file mode 100644 index 0000000000000000000000000000000000000000..206335a01ff7509eaa5528002c6c9686f05c931b --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvTransOperator.h @@ -0,0 +1,44 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#pragma once + +#include "ConvBaseOperator.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief ConvTransOperator takes two inputs to perform the convolution. + * The first input is the image, and the second input is the convolution kernel. + * The height of data for two inputs are the same. Each data of the first input + * is convolved with each data of the second input indepedently. + * + * The config file api is conv_operator. + */ + +class ConvTransOperator : public ConvBaseOperator { + public: + ConvTransOperator(const OperatorConfig &config, bool useGpu) + : ConvBaseOperator(config, useGpu) {} + /** + * Free workspace in device and destroy cudnn tensor descriptor. + */ + virtual ~ConvTransOperator() {} + void forward() override; + void backward() override; + void reshape(int batchSize) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvTransProjection.cpp b/paddle/legacy/gserver/layers/ConvTransProjection.cpp new file mode 100644 index 0000000000000000000000000000000000000000..00e34c8f2dcd2ea9698779f8b4425561f979cfef --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvTransProjection.cpp @@ -0,0 +1,123 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ConvTransProjection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_PROJECTION(convt, ConvTransProjection); +size_t ConvTransProjection::calOutputSize() { + outputH_ = in_->getFrameHeight(); + outputW_ = in_->getFrameWidth(); + if (outputH_ == 0) outputH_ = configOutH_; + if (outputW_ == 0) outputW_ = configOutW_; + imageH_ = imageSize(outputH_, + (filterH_ - 1) * dilationH_ + 1, + paddingH_, + strideH_, + /* caffeMode */ true); + + imageW_ = imageSize(outputW_, + (filterW_ - 1) * dilationW_ + 1, + paddingW_, + strideW_, + /* caffeMode */ true); + + const_cast(out_)->setFrameHeight(imageH_); + const_cast(out_)->setFrameWidth(imageW_); + + inputOffset_ = (configChannels_ / groups_) * outputH_ * outputW_; + outputOffset_ = (configNumFilters_ / groups_) * imageH_ * imageW_; + return imageH_ * imageW_ * configNumFilters_; +} + +size_t ConvTransProjection::calInputSize() { + return static_cast(configChannels_ * outputH_ * outputW_); +} + +void ConvTransProjection::forward() { + int batchSize = in_->value->getHeight(); + reshape(batchSize); + + void *workSpace = NULL; + if (workSpaceInBytes_ > 0) { + workSpace = getSpaceBytes(workSpaceInBytes_); + } + + for (int g = 0; g < groups_; ++g) { + REGISTER_TIMER_INFO("CudnnConvTransFwTimer", getName().c_str()); + + real *inData = in_->value->getData() + g * inputOffset_; + real *wgtData = weight_->getW()->getData() + g * weightOffset_; + real *outData = out_->value->getData() + g * outputOffset_; + hl_convolution_backward_data(imageDesc_, + outData, + outputDesc_, + inData, + filterDesc_, + wgtData, + convDesc_, + workSpace, + bwdDataLimitBytes_, + bwdDataAlgo_); + } +} + +void ConvTransProjection::backward(const UpdateCallback &callback) { + REGISTER_TIMER_INFO("CudnnConvTransBpTimer", getName().c_str()); + + void *workSpace = NULL; + if (workSpaceInBytes_ > 0) { + workSpace = getSpaceBytes(workSpaceInBytes_); + } + + for (int g = 0; g < groups_; ++g) { + real *outGrad = out_->grad->getData() + g * outputOffset_; + if (weight_->getWGrad()) { + real *inData = in_->value->getData() + g * inputOffset_; + real *weightGrad = weight_->getWGrad()->getData() + g * weightOffset_; + hl_convolution_backward_filter(imageDesc_, + outGrad, + outputDesc_, + inData, + filterDesc_, + weightGrad, + convDesc_, + workSpace, + bwdFilterLimitBytes_, + bwdFilterAlgo_); + } + + MatrixPtr preGrad = in_->grad; + if (NULL != preGrad) { + real *inGrad = preGrad->getData() + g * inputOffset_; + real *wgtData = weight_->getW()->getData() + g * weightOffset_; + hl_convolution_forward(imageDesc_, + outGrad, + outputDesc_, + inGrad, + filterDesc_, + wgtData, + convDesc_, + workSpace, + fwdLimitBytes_, + fwdAlgo_); + } + } + + weight_->getParameterPtr()->incUpdate(callback); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvTransProjection.h b/paddle/legacy/gserver/layers/ConvTransProjection.h new file mode 100644 index 0000000000000000000000000000000000000000..9b63dd47352b9f24810d9406b314fbfa15ae13c3 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvTransProjection.h @@ -0,0 +1,43 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ConvBaseProjection.h" +#include "paddle/legacy/math/MathUtils.h" + +namespace paddle { + +/** + * @brief Convolution projection do the same calculation with CudnnConvLayer. + */ +class ConvTransProjection : public ConvBaseProjection { + public: + /** + * Constructor. + */ + ConvTransProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu) + : ConvBaseProjection(config, parameter, useGpu) {} + + ~ConvTransProjection() {} + + virtual void forward(); + virtual void backward(const UpdateCallback& callback); + virtual size_t calOutputSize(); + virtual size_t calInputSize(); +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ConvexCombinationLayer.cpp b/paddle/legacy/gserver/layers/ConvexCombinationLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c38ab251f18728425d01479b82630550d29e9b61 --- /dev/null +++ b/paddle/legacy/gserver/layers/ConvexCombinationLayer.cpp @@ -0,0 +1,155 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief A layer for weighted sum of vectors, + * which is used in NEURAL MACHINE TRANSLATION BY JOINTLY LEARNING TO ALIGN AND + * TRANSLATE + * - Input: the the size of the first input is weightDim, + * and the size of the second input is weightdim * dataDim. + * - Output: the sizeof the output is dataDim + * \f[ + * out(j) = \sum_{i}(in0(i) * in1(i,j + i * dataDim)), + * i = 0,1,...,(weightDim-1); j = 0, 1,...,(dataDim-1) + * \f] + * Note that the above computation is for one sample. Multiple samples are + * processed in one batch. + * + * The config file api is linear_comb_layer. + */ +class ConvexCombinationLayer : public Layer { + protected: + /// A matrix pointer pointing to second input. + MatrixPtr tmpMtx0; + /// A matrix pointer pointing to first input. + MatrixPtr tmpRow0; + /// A matrix pointer pointing to output. + MatrixPtr tmpRow1; + + public: + explicit ConvexCombinationLayer(const LayerConfig& config) : Layer(config) {} + + ~ConvexCombinationLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(convex_comb, ConvexCombinationLayer); + +bool ConvexCombinationLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(2U, inputLayers_.size()); + size_t dataDim = getSize(); + size_t weightDim = inputLayers_[0]->getSize(); + + CHECK_EQ(weightDim * dataDim, inputLayers_[1]->getSize()) + << "Dimension mismatch"; + + tmpRow0 = Matrix::create(nullptr, + /* height= */ 1, + weightDim, + /* trans= */ false, + useGpu_); + tmpRow1 = Matrix::create(nullptr, + /* height= */ 1, + dataDim, + /* trans= */ false, + useGpu_); + tmpMtx0 = Matrix::create(nullptr, + /* height= */ weightDim, + dataDim, + /* trans= */ false, + useGpu_); + + return true; +} + +void ConvexCombinationLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV0->getHeight(); + size_t weightDim = inV0->getWidth(); + size_t dataDim = getSize(); + + CHECK_EQ(batchSize, inV1->getHeight()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, dataDim); + } + + MatrixPtr outV = getOutputValue(); + + REGISTER_TIMER_INFO("FwCvxCombTimer", getName().c_str()); + for (size_t i = 0; i < batchSize; i++) { + tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim); + tmpRow0->setData(inV0->getData() + i * weightDim); + tmpRow1->setData(outV->getData() + i * dataDim); + + tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 0); + } +} + +void ConvexCombinationLayer::backward(const UpdateCallback& callback) { + MatrixPtr outG = getOutputGrad(); + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + + size_t batchSize = inV0->getHeight(); + size_t weightDim = inV0->getWidth(); + size_t dataDim = getSize(); + + REGISTER_TIMER_INFO("BwCvxCombTimer", getName().c_str()); + + if (inG0) { + for (size_t i = 0; i < batchSize; i++) { + tmpRow0->setData(inG0->getData() + i * weightDim); + tmpRow1->setData(outG->getData() + i * dataDim); + tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim); + + tmpRow0->mul(*tmpRow1, *(tmpMtx0->getTranspose()), 1, 1); + } + } + + if (inG1) { + for (size_t i = 0; i < batchSize; i++) { + tmpRow0->setData(inV0->getData() + i * weightDim); + tmpRow1->setData(outG->getData() + i * dataDim); + tmpMtx0->setData(inG1->getData() + i * weightDim * dataDim); + + tmpMtx0->mul(*(tmpRow0->getTranspose()), *tmpRow1, 1, 1); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CosSimLayer.cpp b/paddle/legacy/gserver/layers/CosSimLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ab8d7cc1f61823890676e8f647f784cfa9a0775e --- /dev/null +++ b/paddle/legacy/gserver/layers/CosSimLayer.cpp @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CosSimLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(cos, CosSimLayer); + +bool CosSimLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2LU); + + createFunction(forward_, + "CosSimForward", + FuncConfig().set("scale", (real)config_.cos_scale())); + createFunction(backward_, + "CosSimBackward", + FuncConfig().set("scale", (real)config_.cos_scale())); + + return true; +} + +void CosSimLayer::forward(PassType passType) { + Layer::forward(passType); + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(0)->getHeight(); + int size = getSize(); + CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed"; + + { + REGISTER_TIMER_INFO("CosFwResetTimer", getName().c_str()); + reserveOutput(batchSize, size); + } + + MatrixPtr outV = getOutputValue(); + /* activation */ { + REGISTER_TIMER_INFO("CosFwAtvTimer", getName().c_str()); + MatrixPtr prevOut1 = getInputValue(0); + MatrixPtr prevOut2 = getInputValue(1); + + CHECK(outV && prevOut1 && prevOut2); + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*prevOut1); + inputs.addArg(*prevOut2); + outputs.addArg(*outV, ASSIGN_TO); + forward_[0]->calc(inputs, outputs); + } +} + +void CosSimLayer::backward(const UpdateCallback& callback) { + /* activation */ { + REGISTER_TIMER_INFO("CosBpAtvTimer", getName().c_str()); + CHECK_EQ(backward_.size(), 1UL) << "Only one backward function needed"; + + const auto outG = this->getOutputGrad(); + const auto outV = this->getOutputValue(); + const auto inV1 = this->getInputValue(0); + const auto inV2 = this->getInputValue(1); + auto inG1 = this->getInputGrad(0); + auto inG2 = this->getInputGrad(1); + CHECK(outG && outV && inV1 && inV2 && inG1 && inG2); + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*outG); + inputs.addArg(*outV); + inputs.addArg(*inV1); + inputs.addArg(*inV2); + outputs.addArg(*inG1, ADD_TO); + outputs.addArg(*inG2, ADD_TO); + + backward_[0]->calc(inputs, outputs); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CosSimLayer.h b/paddle/legacy/gserver/layers/CosSimLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..b08e2c6a35369832732706d64f209f85a5292a6f --- /dev/null +++ b/paddle/legacy/gserver/layers/CosSimLayer.h @@ -0,0 +1,48 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { +/** + * @brief A layer for calculating cosine similarity between two vector + * \f[ + * f(x,y)=scale\frac{x_1y_1+x_2y_2+...+x_ny_n}{\sqrt{x_1^2+x_2^2+... + * +x_n^2}\sqrt{y_1^2+y_2^2+...+y_n^2}} + * \f] + * + * - Input1: A vector (batchSize * dataDim) * + * - Input2: A vector (batchSize * dataDim) or (1 * dataDim) * + * - Output: A vector (batchSize * 1) + * + * The config file api is cos_sim. + */ +class CosSimLayer : public Layer { + public: + explicit CosSimLayer(const LayerConfig& config) : Layer(config) {} + + ~CosSimLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CosSimVecMatLayer.cpp b/paddle/legacy/gserver/layers/CosSimVecMatLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..03de0be815a1fb5eeb7ffab31b1721dc5951a469 --- /dev/null +++ b/paddle/legacy/gserver/layers/CosSimVecMatLayer.cpp @@ -0,0 +1,182 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { +/** + * @brief A layer for computing cosine similarity between a vector + * and each row of a matrix + * out[i] = cos_scale * cos(in1, in2(i,:)); + * @note used in NEURAL TURING MACHINE + * + * Input1: a vector (batchSize * dataDim) + * + * Input2: a matrix in vector form (batchSize * (weightDim*dataDim)) + * + * Output: a vector (batchSize * weightDim) + */ + +class CosSimVecMatLayer : public Layer { + protected: + MatrixPtr tmpMtx0; + MatrixPtr tmpMtx1; + MatrixPtr tmpRow0; + MatrixPtr tmpRow1; + MatrixPtr tmpRow2; + MatrixPtr tmpRow3; + + public: + explicit CosSimVecMatLayer(const LayerConfig& config) : Layer(config) {} + + ~CosSimVecMatLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(cos_vm, CosSimVecMatLayer); + +bool CosSimVecMatLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2U); + + size_t dataDim = inputLayers_[0]->getSize(); + size_t numKeys = getSize(); + size_t memoryDim = inputLayers_[1]->getSize(); + + CHECK_EQ(dataDim * numKeys, memoryDim) << "Dimension mismatch"; + + tmpRow0 = Matrix::create(nullptr, + /* height= */ 1, + dataDim, + /* trans= */ false, + useGpu_); + tmpRow1 = Matrix::create(nullptr, + /* height= */ 1, + dataDim, + /* trans= */ false, + useGpu_); + tmpRow2 = Matrix::create(nullptr, + /* height= */ numKeys, + 1, + /* trans= */ false, + useGpu_); + tmpRow3 = Matrix::create(nullptr, + /* height= */ numKeys, + 1, + /* trans= */ false, + useGpu_); + + tmpMtx0 = Matrix::create(nullptr, + /* height= */ numKeys, + dataDim, + /* trans= */ false, + useGpu_); + tmpMtx1 = Matrix::create(nullptr, + /* height= */ numKeys, + dataDim, + /* trans= */ false, + useGpu_); + + CHECK(tmpRow0 && tmpRow1 && tmpRow2 && tmpRow3 && tmpMtx0 && tmpMtx1); + + createFunction(forward_, + "CosSimForward", + FuncConfig().set("scale", (real)config_.cos_scale())); + createFunction(backward_, + "CosSimBackward", + FuncConfig().set("scale", (real)config_.cos_scale())); + + return true; +} + +void CosSimVecMatLayer::forward(PassType passType) { + Layer::forward(passType); + CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed"; + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV0->getHeight(); + size_t numKeys = getSize(); + + CHECK_EQ(batchSize, inV1->getHeight()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, numKeys); + } + + MatrixPtr outV = getOutputValue(); + CHECK(outV && inV0 && inV1); + REGISTER_TIMER_INFO("FwCosVMTimer", getName().c_str()); + for (size_t i = 0; i < batchSize; i++) { + tmpRow0->setData(inV0->rowBuf(i)); + tmpMtx0->setData(inV1->rowBuf(i)); + tmpRow2->setData(outV->rowBuf(i)); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*tmpMtx0); + inputs.addArg(*tmpRow0); + outputs.addArg(*tmpRow2, ASSIGN_TO); + forward_[0]->calc(inputs, outputs); + } +} + +void CosSimVecMatLayer::backward(const UpdateCallback& callback) { + CHECK_EQ(backward_.size(), 1UL) << "Only one forward function needed"; + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + MatrixPtr outV = getOutputValue(); + MatrixPtr outG = getOutputGrad(); + + size_t batchSize = inV0->getHeight(); + CHECK(inV0 && inV1 && inG0 && inG1 && outV && outG); + REGISTER_TIMER_INFO("BwCosVMTimer", getName().c_str()); + + for (size_t i = 0; i < batchSize; i++) { + tmpRow0->setData(inV0->rowBuf(i)); + tmpRow1->setData(inG0->rowBuf(i)); + tmpMtx0->setData(inV1->rowBuf(i)); + tmpMtx1->setData(inG1->rowBuf(i)); + tmpRow2->setData(outV->rowBuf(i)); + tmpRow3->setData(outG->rowBuf(i)); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*tmpRow3); + inputs.addArg(*tmpRow2); + inputs.addArg(*tmpMtx0); + inputs.addArg(*tmpRow0); + outputs.addArg(*tmpMtx1, ADD_TO); + outputs.addArg(*tmpRow1, ADD_TO); + + backward_[0]->calc(inputs, outputs); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CostLayer.cpp b/paddle/legacy/gserver/layers/CostLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..18b5b77bde9dee97cb6971624007307ff06411c7 --- /dev/null +++ b/paddle/legacy/gserver/layers/CostLayer.cpp @@ -0,0 +1,748 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CostLayer.h" +#include +#include +#include +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/math/SparseMatrix.h" + +namespace paddle { + +bool CostLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + bool ret = Layer::init(layerMap, parameterMap); + coeff_ = config_.coeff(); + if (!ret) return ret; + CHECK_GE(inputLayers_.size(), 2UL); + CHECK_LE(inputLayers_.size(), 3UL); + if (inputLayers_.size() == 3) { + weightLayer_ = inputLayers_[2]; + } + return true; +} + +void CostLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(*getOutputLayer())->getHeight(); + int size = 1; + resetOutput(batchSize, size); + + const MatrixPtr& output = getInputValue(*getOutputLayer()); + Argument label = getInput(*getLabelLayer()); + + /* get the cost value for each sample*/ + forwardImp(*output, label, *getOutputValue()); + if (weightLayer_) { + const MatrixPtr& weight = getInputValue(*weightLayer_); + getOutputValue()->dotMul(*getOutputValue(), *weight); + } +} + +void CostLayer::backward(const UpdateCallback& callback) { + (void)callback; + + const Argument& output = getInput(*getOutputLayer()); + Argument label = getInput(*getLabelLayer()); + + bool support = true; + if (weightLayer_) { + support = output.grad->getAbsSum() == 0; + } + + backwardImp(*output.value, label, *output.grad); + + if (weightLayer_) { + CHECK(support) << "Weighted cost layer '" << getName() + << "' must be the last layer " + "connected to the output layer '" + << getOutputLayer()->getName() << "'"; + output.grad->rowScale(0, *output.grad, *getInputValue(*weightLayer_)); + } + if (coeff_ != real(1.0f)) { + output.grad->add(coeff_, 0); + } +} + +// +// class MultiClassCrossEntropy +// +bool MultiClassCrossEntropy::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return CostLayer::init(layerMap, parameterMap); +} + +void MultiClassCrossEntropy::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + target.oneHotCrossEntropy(output, *label.ids); +} + +void MultiClassCrossEntropy::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + outputG.oneHotCrossEntropyBp(output, *label.ids); +} + +// +// class MultiClassCrossEntropyWithSelfNorm +// +REGISTER_LAYER(multi_class_cross_entropy_with_selfnorm, + MultiClassCrossEntropyWithSelfNorm); + +bool MultiClassCrossEntropyWithSelfNorm::init( + const LayerMap& layerMap, const ParameterMap& parameterMap) { + return CostLayer::init(layerMap, parameterMap); +} + +void MultiClassCrossEntropyWithSelfNorm::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + Matrix::resizeOrCreate(sftMaxSum_, output.getHeight(), 1, false, useGpu_); + output.rowSum(*sftMaxSum_); + sftMaxSum_->log2(); + + target.oneHotCrossEntropy(output, *label.ids); + target.add(*sftMaxSum_); + + sftMaxSum_->square2(); + target.add(*sftMaxSum_, config_.softmax_selfnorm_alpha()); +} + +void MultiClassCrossEntropyWithSelfNorm::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + Matrix::resizeOrCreate(sftMaxSum_, output.getHeight(), 1, false, useGpu_); + output.rowSum(*sftMaxSum_); + + Matrix::resizeOrCreate(sumInv_, output.getHeight(), 1, false, useGpu_); + sftMaxSum_->reciprocal2(*sumInv_); + + outputG.oneHotCrossEntropyBp(output, *label.ids); + outputG.addColumnVector(*sumInv_); + + sftMaxSum_->log2(); + sumInv_->dotMul(*sumInv_, *sftMaxSum_); + sumInv_->mulScalar(2 * config_.softmax_selfnorm_alpha()); + + outputG.addColumnVector(*sumInv_); +} + +// +// class SoftBinaryClassCrossEntropy +// +REGISTER_LAYER(soft_binary_class_cross_entropy, SoftBinaryClassCrossEntropy); + +bool SoftBinaryClassCrossEntropy::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return CostLayer::init(layerMap, parameterMap); +} + +void SoftBinaryClassCrossEntropy::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + Matrix::resizeOrCreate( + targetPerDim_, output.getHeight(), output.getWidth(), false, useGpu_); + + targetPerDim_->softCrossEntropy(output, *label.value); + targetPerDim_->rowSum(target); +} + +void SoftBinaryClassCrossEntropy::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + outputG.softCrossEntropyBp(output, *label.value); +} + +// +// class SumOfSquaresCostLayer +// + +REGISTER_LAYER(square_error, SumOfSquaresCostLayer); + +bool SumOfSquaresCostLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return CostLayer::init(layerMap, parameterMap); +} + +void SumOfSquaresCostLayer::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + target.sumOfSquares(output, *label.value); +} + +void SumOfSquaresCostLayer::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + outputG.sumOfSquaresBp(output, *label.value); +} + +// +// class SmoothL1CostLayer +// + +REGISTER_LAYER(smooth_l1, SmoothL1CostLayer); + +bool SmoothL1CostLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return CostLayer::init(layerMap, parameterMap); +} + +void SmoothL1CostLayer::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + MatrixPtr targetCpu, outputCpu, labelCpu; + if (useGpu_) { + targetCpu = + Matrix::create(target.getHeight(), target.getWidth(), false, false); + outputCpu = + Matrix::create(output.getHeight(), output.getWidth(), false, false); + labelCpu = Matrix::create( + label.value->getHeight(), label.value->getWidth(), false, false); + targetCpu->copyFrom(target); + outputCpu->copyFrom(output); + labelCpu->copyFrom(*label.value); + targetCpu->smoothL1(*outputCpu, *labelCpu, 1.0); + target.copyFrom(*targetCpu); + } else { + target.smoothL1(output, *label.value, 1.0); + } +} + +void SmoothL1CostLayer::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + MatrixPtr outputGCpu, outputCpu, labelCpu; + if (useGpu_) { + outputGCpu = + Matrix::create(outputG.getHeight(), outputG.getWidth(), false, false); + outputCpu = + Matrix::create(output.getHeight(), output.getWidth(), false, false); + labelCpu = Matrix::create( + label.value->getHeight(), label.value->getWidth(), false, false); + outputGCpu->copyFrom(outputG); + outputCpu->copyFrom(output); + labelCpu->copyFrom(*label.value); + outputGCpu->smoothL1Bp(*outputCpu, *labelCpu, 1.0); + outputG.copyFrom(*outputGCpu); + } else { + outputG.smoothL1Bp(output, *label.value, 1.0); + } +} + +// +// class RankingCost +// +bool RankingCost::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + posPairCount_ = 0; + negPairCount_ = 0; + + bool ret = Layer::init(layerMap, parameterMap); + if (!ret) return ret; + CHECK_GE(inputLayers_.size(), 3UL); + CHECK_LE(inputLayers_.size(), 4UL); + if (inputLayers_.size() == 4) { + weightLayer_ = inputLayers_[3]; + } + return true; +} + +void RankingCost::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(*getOutputLayer(0))->getHeight(); + int size = 1; + resizeOutput(batchSize, size); + Matrix::resizeOrCreate(margin_, batchSize, size, /* trans= */ false, useGpu_); + MatrixPtr label = getInputValue(*getLabelLayer()); + if (!label) { + // input label is not in value, try ids + IVectorPtr idLabel = getInput(*getLabelLayer()).ids; + CHECK(idLabel) << "label layer has neither value nor ids"; + CHECK_EQ((size_t)batchSize, idLabel->getSize()); + Matrix::resizeOrCreate( + labelBuf_, batchSize, /*width*/ 1, /*trans*/ false, useGpu_); + labelBuf_->copyFrom(*idLabel); + label = labelBuf_; + } + + MatrixPtr output[] = {getInputValue(*getOutputLayer(0)), + getInputValue(*getOutputLayer(1))}; + MatrixPtr target = this->getOutputValue(); + margin_->sub(*output[0], *output[1]); + + // for validation + size_t height = output[0]->getHeight(); + target->biggerThan(*(output[0]), *(output[1]), *label); + double total = static_cast(height); + if (weightLayer_) { + const MatrixPtr& weight = getInputValue(*weightLayer_); + target->dotMul(*target, *weight); + total = weight->getSum(); + } + double pos = target->getSum(); + posPairCount_ += pos; + negPairCount_ += (total - pos); + + // forward + target->logisticRegressionLoss(*margin_, *label); + if (weightLayer_) { + const MatrixPtr& weight = getInputValue(*weightLayer_); + target->dotMul(*target, *weight); + } +} + +void RankingCost::backward(const UpdateCallback& callback) { + (void)callback; + + MatrixPtr label = getInputValue(*getLabelLayer()); + if (!label) { + // input label is not in value, but in ids + // use labelBuf_ (should already resized and copied during forward) + label = labelBuf_; + } + + Matrix::resizeOrCreate( + marginGrad_, label->getHeight(), 1, /* trans= */ false, useGpu_); + marginGrad_->zeroMem(); + marginGrad_->logisticRegressionLossBp(*margin_, *label); + if (weightLayer_) { + const MatrixPtr& weight = getInputValue(*weightLayer_); + marginGrad_->dotMul(*marginGrad_, *weight); + } + + getInputGrad(0)->add(*marginGrad_); + getInputGrad(1)->sub(*marginGrad_); +} + +void RankingCost::onPassEnd() { + double ratio = posPairCount_ / ((negPairCount_ <= 0) ? 1.0 : negPairCount_); + LOG(INFO) << "calc pos/neg: " << ratio << " pos= " << posPairCount_ + << " neg= " << negPairCount_; + + posPairCount_ = 0; + negPairCount_ = 0; +} + +// +// class LambdaCost +// +REGISTER_LAYER(lambda_cost, LambdaCost); + +bool LambdaCost::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + truncationSize_ = config_.ndcg_num(); + maxSortSize_ = config_.max_sort_size(); + if (maxSortSize_ != -1) { + CHECK_GE(maxSortSize_, truncationSize_) + << "maxSortSize must be greater than or equal to NDCG size!"; + } + LOG(INFO) << "LambdaRank v1.3, NDCG size = " << truncationSize_ + << ", Max partial sort size = " << maxSortSize_; + CHECK(!useGpu_) << "LambdaRank supports CPU only!"; + return Layer::init(layerMap, parameterMap); +} + +void LambdaCost::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(*getOutputLayer())->getHeight(); + resizeOutput(batchSize, 1); + + MatrixPtr score = getInputValue(*getScoreLayer()); + MatrixPtr output = getInputValue(*getOutputLayer()); + MatrixPtr target = this->getOutputValue(); + + real* scoreData = score->getData(); + real* outputData = output->getData(); + real* targetData = target->getData(); + + auto startPos = getInput(*getOutputLayer()).sequenceStartPositions; + const int* startPosData = startPos->getData(false); + size_t batchNum = startPos->getSize() - 1; + for (size_t i = 0; i < batchNum; ++i) { + int beginPos = startPosData[i]; + int endPos = startPosData[i + 1]; + real NDCG = calcNDCG( + outputData + beginPos, scoreData + beginPos, endPos - beginPos); + for (int j = beginPos; j < endPos; ++j) { + targetData[j] = NDCG; + } + } +} + +void LambdaCost::backward(const UpdateCallback& callback) { + (void)callback; + MatrixPtr score = getInputValue(*getScoreLayer()); + MatrixPtr output = getInputValue(*getOutputLayer()); + Matrix::resizeOrCreate(marginGrad_, + score->getHeight(), + 1, + /* trans= */ false, + useGpu_); + marginGrad_->zeroMem(); + + real* gradData = marginGrad_->getData(); + real* scoreData = score->getData(); + real* outputData = output->getData(); + + auto startPos = getInput(*getOutputLayer()).sequenceStartPositions; + const int* startPosData = startPos->getData(false); + size_t batchNum = startPos->getSize() - 1; + + for (size_t i = 0; i < batchNum; ++i) { + int beginPos = startPosData[i]; + int endPos = startPosData[i + 1]; + calcGrad(outputData + beginPos, + scoreData + beginPos, + gradData + beginPos, + endPos - beginPos); + } + + getInputGrad(0)->add(*marginGrad_); +} + +void LambdaCost::calcGrad(const real* outputScore, + const real* score, + real* gradData, + int size) { + CHECK_GE(size, truncationSize_) + << "Invalid: (Sample num in the same list) < (NDCG truncation num) !"; + int sortSize = maxSortSize_ == -1 ? size : std::min(maxSortSize_, size); + + scorePair_.clear(); + for (int i = 0; i < size; ++i) { + scorePair_.push_back(std::make_pair(score[i], i)); + } + if (size <= sortSize) { + std::sort(scorePair_.begin(), + scorePair_.end(), + [](const std::pair& a, const std::pair& b) { + return a.first > b.first; + }); + } else { + std::partial_sort( + scorePair_.begin(), + scorePair_.begin() + sortSize, + scorePair_.end(), + [](const std::pair& a, const std::pair& b) { + return a.first > b.first; + }); + } + + real maxDCG = 0; + for (int i = 0; i < truncationSize_; ++i) { + maxDCG += (std::pow(2, scorePair_[i].first) - 1) / std::log(i + 2); + } + CHECK_GT(maxDCG, 0) << "Invalid: max DCG = 0!"; + + for (int i = 0; i < sortSize; ++i) { + for (int j = i + 1; j < size; ++j) { + int index_i = scorePair_[i].second; + int index_j = scorePair_[j].second; + real score_i = score[index_i]; + real score_j = score[index_j]; + real dcgDif = 0; + if (j < sortSize) { + dcgDif = (std::pow(2, score_i) - std::pow(2, score_j)) * + (1 / std::log(i + 2) - 1 / std::log(j + 2)); + } else { + dcgDif = + (std::pow(2, score_i) - std::pow(2, score_j)) / std::log(i + 2); + } + + real lambda_ij = + -std::abs(dcgDif) / + (1 + std::exp(outputScore[index_i] - outputScore[index_j])); + gradData[index_i] += lambda_ij / maxDCG; + gradData[index_j] -= lambda_ij / maxDCG; + } + } +} + +real LambdaCost::calcNDCG(const real* outputScore, + const real* score, + int size) { + CHECK_GE(size, truncationSize_) + << "Invalid: (Sample num in the same list) < (NDCG truncation num) !"; + + outputScorePair_.clear(); + for (int i = 0; i < size; ++i) { + outputScorePair_.push_back(std::make_pair(outputScore[i], i)); + } + std::partial_sort( + outputScorePair_.begin(), + outputScorePair_.begin() + truncationSize_, + outputScorePair_.end(), + [](const std::pair& a, const std::pair& b) { + return a.first > b.first; + }); + + real DCG = 0; + for (int i = 0; i < truncationSize_; ++i) { + DCG += + (std::pow(2, score[outputScorePair_[i].second]) - 1) / std::log(i + 2); + } + + scoreVec_.resize(size); + std::copy(score, score + size, scoreVec_.begin()); + real maxDCG = 0; + std::partial_sort(scoreVec_.begin(), + scoreVec_.begin() + truncationSize_, + scoreVec_.end(), + std::greater()); + for (int i = 0; i < truncationSize_; ++i) { + maxDCG += (std::pow(2, scoreVec_[i]) - 1) / std::log(i + 2); + } + CHECK_GT(maxDCG, 0) << "Invalid: max DCG = 0!"; + + return DCG / maxDCG; +} + +// +// class MultiBinaryLabelCrossEntropy +// + +REGISTER_LAYER(multi_binary_label_cross_entropy, MultiBinaryLabelCrossEntropy); + +bool MultiBinaryLabelCrossEntropy::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return CostLayer::init(layerMap, parameterMap); +} + +void MultiBinaryLabelCrossEntropy::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + MatrixPtr value = nullptr; + if (label.ids) { + CHECK(!label.value); + value = label.ids->toOneHotSparseMatrix(output.getWidth(), useGpu_); + } else { + CHECK(label.value); + value = label.value; + } + + if (dynamic_cast(value.get()) || + dynamic_cast(value.get())) { + target.multiBinaryLabelCrossEntropy(output, *value); + } else { + Matrix::resizeOrCreate( + targetPerDim_, output.getHeight(), output.getWidth(), false, useGpu_); + + targetPerDim_->binaryLabelCrossEntropy(output, *value); + targetPerDim_->rowSum(target); + } +} + +void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + MatrixPtr value = nullptr; + if (label.ids) { + CHECK(!value); + value = label.ids->toOneHotSparseMatrix(output.getWidth(), useGpu_); + } else { + CHECK(label.value); + value = label.value; + } + + if (dynamic_cast(value.get()) || + dynamic_cast(value.get())) { + outputG.multiBinaryLabelCrossEntropyBp(output, *value); + } else { + outputG.binaryLabelCrossEntropyBp(output, *value); + } +} + +bool HuberCost::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + CostLayer::init(layerMap, parameterMap); + if (useGpu_) { + tmpCpuInput_.reserve(inputLayers_.size()); + for (size_t i = 0; i < inputLayers_.size(); i++) { + tmpCpuInput_.push_back(Argument()); + } + } + return true; +} + +void HuberCost::forwardImp(Matrix& output, Argument& label, Matrix& cost) { + if (useGpu_) { + for (size_t i = 0; i < inputLayers_.size(); i++) { + tmpCpuInput_[i].resizeAndCopyFrom( + getInput(i), false, HPPL_STREAM_DEFAULT); + } + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + } +} + +// +// Huber loss for robust regression. +// +REGISTER_LAYER(huber_regression, HuberRegressionLoss); + +bool HuberRegressionLoss::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + HuberCost::init(layerMap, parameterMap); + delta_ = config_.delta(); + return true; +} + +void HuberRegressionLoss::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + HuberCost::forwardImp(output, label, target); + size_t numSamples = target.getHeight(); + size_t dim = output.getWidth(); + CHECK(label.value); + CHECK_EQ((*label.value).getHeight(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(dim, (*label.value).getWidth()); + CHECK_EQ(target.getWidth(), (size_t)1); + + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + real* lbl = + useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); + std::vector cost(numSamples, 0); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + int index = i * dim + j; + real a = std::abs(lbl[index] - out[index]); + if (a <= delta_) + cost[i] += a * a / 2; + else + cost[i] += delta_ * (a - delta_ / 2); + } + } + target.copyFrom(cost.data(), numSamples); +} + +void HuberRegressionLoss::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + size_t numSamples = output.getHeight(); + size_t dim = output.getWidth(); + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + real* lbl = + useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); + real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + int index = i * dim + j; + real a = lbl[index] - out[index]; + if (std::abs(a) <= delta_) + grad[index] += -a; + else + grad[index] += a > 0 ? -delta_ : delta_; + } + } + if (useGpu_) outputG.copyFrom(grad, numSamples * dim); +} + +// +// Huber loss for robust 2-classes classification +// +REGISTER_LAYER(huber_classification, HuberTwoClassification); + +bool HuberTwoClassification::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return HuberCost::init(layerMap, parameterMap); +} + +void HuberTwoClassification::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + HuberCost::forwardImp(output, label, target); + size_t numSamples = target.getHeight(); + CHECK(label.ids); + CHECK_EQ((*label.ids).getSize(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(output.getWidth(), (size_t)1); + CHECK_EQ(target.getWidth(), (size_t)1); + + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); + std::vector cost(numSamples, 0); + for (size_t i = 0; i < numSamples; ++i) { + int y = 2 * lbl[i] - 1; + real a = out[i] * y; + if (a < -1) + cost[i] = -4 * a; + else if (a < 1) + cost[i] = (1 - a) * (1 - a); + } + target.copyFrom(cost.data(), numSamples); +} + +void HuberTwoClassification::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + size_t numSamples = output.getHeight(); + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); + real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); + for (size_t i = 0; i < numSamples; ++i) { + int y = 2 * lbl[i] - 1; + real a = out[i] * y; + if (a < -1) + grad[i] += -4 * y; + else if (a < 1) + grad[i] += -2 * (1 - a) * y; + } + if (useGpu_) outputG.copyFrom(grad, numSamples); +} +/** + * This cost layer compute the sum of its input as loss. + * \f[ + * o(i) = \sum_{j=1}^D y_{ij} + * \f] + */ +class SumCostLayer : public Layer { + public: + explicit SumCostLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override { + bool ret = Layer::init(layerMap, parameterMap); + if (!ret) return ret; + CHECK_EQ(inputLayers_.size(), 1UL); + return true; + } + + void forward(PassType passType) override { + Layer::forward(passType); + const MatrixPtr& input = getInputValue(0); + + /* malloc memory for the output_ if necessary */ + int batchSize = input->getHeight(); + int size = 1; + resizeOutput(batchSize, size); + output_.value->sumRows(*input, /* scaleSum= */ 1, /* scaleDest= */ 0); + } + + void backward(const UpdateCallback& callback = nullptr) override { + getInputGrad(0)->add((real)1); + } +}; + +REGISTER_LAYER(sum_cost, SumCostLayer); + +} // namespace paddle diff --git a/paddle/gserver/layers/CostLayer.h b/paddle/legacy/gserver/layers/CostLayer.h similarity index 100% rename from paddle/gserver/layers/CostLayer.h rename to paddle/legacy/gserver/layers/CostLayer.h diff --git a/paddle/legacy/gserver/layers/CropLayer.cpp b/paddle/legacy/gserver/layers/CropLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d891375ecce0371503ba3034f0584f3b1e553a55 --- /dev/null +++ b/paddle/legacy/gserver/layers/CropLayer.cpp @@ -0,0 +1,146 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CropLayer.h" +#include "paddle/legacy/utils/Stat.h" +namespace paddle { + +REGISTER_LAYER(crop, CropLayer); + +bool CropLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_LE(static_cast(inputLayers_.size()), 2); + CHECK_GE(static_cast(inputLayers_.size()), 1); + crop_axis_ = config_.axis(); + for (int i = 0; i < config_.offset_size(); i++) { + crop_offsets_.push_back(config_.offset(i)); + } + + // 1. get input_0 shape + auto& input0_img_conf = config_.inputs(0).image_conf(); + inDims_ = TensorShape({0, + input0_img_conf.channels(), + input0_img_conf.has_img_size_y() + ? input0_img_conf.img_size_y() + : input0_img_conf.img_size(), + input0_img_conf.img_size()}); + // 2. get target dims from config + if (config_.inputs_size() == 1) { + targetDims_ = TensorShape({config_.shape(0), + config_.shape(1), + config_.shape(2), + config_.shape(3)}); + } else { + // 2. get input_1 shape + auto& input1_img_conf = config_.inputs(1).image_conf(); + targetDims_ = TensorShape({0, + input1_img_conf.channels(), + input1_img_conf.has_img_size_y() + ? input1_img_conf.img_size_y() + : input1_img_conf.img_size(), + input1_img_conf.img_size()}); + } + + // 3. get final crop corner + int dimSize = 4; + crop_corner_ = {0, 0, 0, 0}; + for (int i = 0; i < dimSize; i++) { + if (i >= crop_axis_) { + if (crop_offsets_.size() > 1) { + crop_corner_[i] = crop_offsets_[i - crop_axis_]; + } else { + crop_corner_[i] = crop_offsets_[0]; + } + } + } + + outDims_ = TensorShape(4); + + createFunction( + forward_, "Crop", FuncConfig().set("crop_corner", crop_corner_)); + createFunction( + backward_, "CropGrad", FuncConfig().set("crop_corner", crop_corner_)); + + return true; +} + +void CropLayer::setOutDims() { + MatrixPtr input = inputLayers_[1]->getOutputValue(); + size_t batchSize = input->getHeight(); + // get target dims from input_1 + if (config_.inputs_size() == 2) { + targetDims_.setDim(0, batchSize); + int ch = config_.inputs(0).image_conf().channels(); + if (ch != 0) targetDims_.setDim(1, ch); + int h = inputLayers_[1]->getOutput().getFrameHeight(); + if (h != 0) targetDims_.setDim(2, h); + int w = inputLayers_[1]->getOutput().getFrameWidth(); + if (w != 0) targetDims_.setDim(3, w); + } + // get final crop shape from target dims and crop axis + std::vector crop_shape; + int dimSize = 4; + for (int i = 0; i < dimSize; i++) { + if (i >= crop_axis_) { + crop_shape.push_back(targetDims_[i]); + } else { + crop_shape.push_back(inDims_[i]); + } + } + + outDims_.reshape( + {crop_shape[0], crop_shape[1], crop_shape[2], crop_shape[3]}); + output_.setFrameHeight(crop_shape[2]); + output_.setFrameWidth(crop_shape[3]); +} + +void CropLayer::setInDims() { + MatrixPtr input = inputLayers_[0]->getOutputValue(); + size_t batchSize = input->getHeight(); + inDims_.setDim(0, batchSize); + int h = inputLayers_[0]->getOutput().getFrameHeight(); + if (h != 0) inDims_.setDim(2, h); + int w = inputLayers_[0]->getOutput().getFrameWidth(); + if (w != 0) inDims_.setDim(3, w); +} + +void CropLayer::forward(PassType passType) { + Layer::forward(passType); + setInDims(); + setOutDims(); + int size = outDims_[1] * outDims_[2] * outDims_[3]; + resetOutput(outDims_[0], size); + MatrixPtr outV = getOutputValue(); + REGISTER_TIMER_INFO("CropForward", getName().c_str()); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), inDims_); + outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO); + forward_[0]->calc(inputs, outputs); +} + +void CropLayer::backward(const UpdateCallback& callback) { + (void)callback; + REGISTER_TIMER_INFO("CropBackward", getName().c_str()); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outDims_); + outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); + backward_[0]->calc(inputs, outputs); +} +} // namespace paddle diff --git a/paddle/gserver/layers/CropLayer.h b/paddle/legacy/gserver/layers/CropLayer.h similarity index 100% rename from paddle/gserver/layers/CropLayer.h rename to paddle/legacy/gserver/layers/CropLayer.h diff --git a/paddle/legacy/gserver/layers/CrossChannelNormLayer.cpp b/paddle/legacy/gserver/layers/CrossChannelNormLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0fe100a96c01713f6c8d10d4eff428e7e743b002 --- /dev/null +++ b/paddle/legacy/gserver/layers/CrossChannelNormLayer.cpp @@ -0,0 +1,137 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "NormLayer.h" +#include "paddle/legacy/math/BaseMatrix.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +MatrixPtr CrossChannelNormLayer::createSampleMatrix(MatrixPtr data, + size_t iter, + size_t spatialDim) { + return Matrix::create(data->getData() + iter * channels_ * spatialDim, + channels_, + spatialDim, + false, + useGpu_); +} + +MatrixPtr CrossChannelNormLayer::createSpatialMatrix(MatrixPtr data, + size_t iter, + size_t spatialDim) { + return Matrix::create( + data->getData() + iter * spatialDim, 1, spatialDim, false, useGpu_); +} + +bool CrossChannelNormLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + CHECK(parameters_[0]); + const NormConfig& conf = config_.inputs(0).norm_conf(); + channels_ = conf.channels(); + scale_.reset(new Weight(channels_, 1, parameters_[0])); + return true; +} + +void CrossChannelNormLayer::forward(PassType passType) { + Layer::forward(passType); + MatrixPtr inV = getInputValue(0); + + size_t batchSize = inV->getHeight(); + size_t dataDim = inV->getWidth(); + CHECK_EQ(getSize(), dataDim); + + reserveOutput(batchSize, dataDim); + MatrixPtr outV = getOutputValue(); + size_t spatialDim = dataDim / channels_; + + Matrix::resizeOrCreate(dataBuffer_, batchSize, dataDim, false, useGpu_); + Matrix::resizeOrCreate(spatialBuffer_, 1, spatialDim, false, useGpu_); + Matrix::resizeOrCreate(normBuffer_, batchSize, spatialDim, false, useGpu_); + + inV->square2(*dataBuffer_); + for (size_t i = 0; i < batchSize; i++) { + const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim); + const MatrixPtr dataTmp = createSampleMatrix(dataBuffer_, i, spatialDim); + MatrixPtr outVTmp = createSampleMatrix(outV, i, spatialDim); + MatrixPtr normTmp = createSpatialMatrix(normBuffer_, i, spatialDim); + + // compute norm. + spatialBuffer_->sumCols(*dataTmp, 1, 0); + // add eps to avoid overflow + spatialBuffer_->add(1e-6); + spatialBuffer_->sqrt2(*spatialBuffer_); + normTmp->copyFrom(*spatialBuffer_); + outVTmp->copyFrom(*inVTmp); + outVTmp->divRowVector(*spatialBuffer_); + // scale the layer. + outVTmp->mulColVector(*scale_->getW()); + } +} + +void CrossChannelNormLayer::backward(const UpdateCallback& callback) { + MatrixPtr inG = getInputGrad(0); + MatrixPtr inV = getInputValue(0); + MatrixPtr outG = getOutputGrad(); + MatrixPtr outV = getOutputValue(); + + size_t batchSize = inG->getHeight(); + size_t dataDim = inG->getWidth(); + size_t spatialDim = dataDim / channels_; + + MatrixPtr inGBuffer; + Matrix::resizeOrCreate(inGBuffer, channels_, spatialDim, false, useGpu_); + + dataBuffer_->dotMul(*outG, *outV); + Matrix::resizeOrCreate(scaleDiff_, channels_, 1, false, useGpu_); + Matrix::resizeOrCreate(channelBuffer_, channels_, 1, false, useGpu_); + Matrix::resizeOrCreate(sampleBuffer_, channels_, spatialDim, false, useGpu_); + scaleDiff_->zeroMem(); + for (size_t i = 0; i < batchSize; i++) { + MatrixPtr outGTmp = createSampleMatrix(outG, i, spatialDim); + const MatrixPtr dataTmp = createSampleMatrix(dataBuffer_, i, spatialDim); + const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim); + const MatrixPtr inGTmp = createSampleMatrix(inG, i, spatialDim); + const MatrixPtr normTmp = createSpatialMatrix(normBuffer_, i, spatialDim); + + channelBuffer_->sumRows(*dataTmp, 1, 0); + channelBuffer_->dotDiv(*channelBuffer_, *(scale_->getW())); + // store a / scale[i] in scaleDiff_ temporary + scaleDiff_->add(*channelBuffer_, 1.); + + sampleBuffer_->dotMul(*inVTmp, *outGTmp); + spatialBuffer_->sumCols(*sampleBuffer_, 1., 0.); + // scale the grad + inGBuffer->copyFrom(*inVTmp); + inGBuffer->mulRowVector(*spatialBuffer_); + // divide by square of norm + spatialBuffer_->dotMul(*normTmp, *normTmp); + inGBuffer->divRowVector(*spatialBuffer_); + // subtract + inGBuffer->add(*outGTmp, -1, 1); + // divide by norm + inGBuffer->divRowVector(*normTmp); + // scale the diff + inGBuffer->mulColVector(*scale_->getW()); + + inGTmp->add(*inGBuffer); + } + // updata scale + if (scale_->getWGrad()) scale_->getWGrad()->add(*scaleDiff_); + scale_->getParameterPtr()->incUpdate(callback); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/CrossEntropyOverBeam.cpp b/paddle/legacy/gserver/layers/CrossEntropyOverBeam.cpp similarity index 100% rename from paddle/gserver/layers/CrossEntropyOverBeam.cpp rename to paddle/legacy/gserver/layers/CrossEntropyOverBeam.cpp diff --git a/paddle/gserver/layers/CrossEntropyOverBeam.h b/paddle/legacy/gserver/layers/CrossEntropyOverBeam.h similarity index 100% rename from paddle/gserver/layers/CrossEntropyOverBeam.h rename to paddle/legacy/gserver/layers/CrossEntropyOverBeam.h diff --git a/paddle/legacy/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/legacy/gserver/layers/CudnnBatchNormLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..051155e0d2c1b4910c6627a902a4150cbfb15800 --- /dev/null +++ b/paddle/legacy/gserver/layers/CudnnBatchNormLayer.cpp @@ -0,0 +1,180 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CudnnBatchNormLayer.h" +#include "Layer.h" +#include "paddle/legacy/cuda/include/hl_batch_norm.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(cudnn_batch_norm, CudnnBatchNormLayer); + +bool CudnnBatchNormLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + if (!BatchNormBaseLayer::init(layerMap, parameterMap)) return false; + CHECK(useGpu_) << "CudnnBatchNorm only support GPU"; + + hl_create_tensor_descriptor(&ioDesc_); + hl_create_tensor_descriptor(&bnParamDesc_); + hl_tensor_reshape(bnParamDesc_, 1, channels_, 1, 1); + + return true; +} + +void CudnnBatchNormLayer::reshape(int batchSize) { + hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_); +} + +void CudnnBatchNormLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = getInputValue(0)->getHeight(); + calFeatureMapSize(); + reshape(batchSize); + resetOutput(batchSize, getInputValue(0)->getWidth()); + + // for testing in training peroid. + useGlobalStats_ = (passType == PASS_TEST); + if (passType == PASS_TEST && config_.has_use_global_stats()) { + useGlobalStats_ = config_.use_global_stats(); + } + + real* input = getInputValue(0)->getData(); + real* output = getOutputValue()->getData(); + real* gamma = weight_->getW()->getData(); + real* beta = biases_->getW()->getData(); + real* movingMean = movingMean_->getW()->getData(); + real* movingVar = movingVar_->getW()->getData(); + + // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. + eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); + + if (!useGlobalStats_) { + REGISTER_TIMER_INFO("CudnnBatchFwTimer", getName().c_str()); + real* savedMean = savedMean_->getData(); + real* savedInvVar = savedInvVar_->getData(); + hl_batch_norm_forward_training(ioDesc_, + input, + ioDesc_, + output, + bnParamDesc_, + gamma, + beta, + 1.0 - movingAvgFraction_, + movingMean, + movingVar, + eps_, + savedMean, + savedInvVar); + } else { + // used movingMean and movingVar in testing + if (batchSize <= 1024) { + hl_batch_norm_forward_inference(ioDesc_, + input, + ioDesc_, + output, + bnParamDesc_, + gamma, + beta, + movingMean, + movingVar, + eps_); + } else { + // There is a limitation in cudnn library. + // When the batch size is larger than 1024 in cuDNN v5.1, + // the cudnnBatchNormalizationForwardInference will fail. + hl_batch_norm_cuda_inference(input, + output, + gamma, + beta, + movingMean, + movingVar, + eps_, + batchSize, + channels_, + imageH_ * imageD_, + imageW_); + } + } + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void CudnnBatchNormLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + real* input = getInputValue(0)->getData(); + real* outGrad = getOutputGrad()->getData(); + real* inGrad = getInputGrad(0)->getData(); + real* gamma = weight_->getW()->getData(); + real* savedMean = savedMean_->getData(); + real* savedInvVar = savedInvVar_->getData(); + + // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. + eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); + + auto create = [](MatrixPtr& m, size_t h, size_t w, real** p) { + Matrix::resizeOrCreate(m, h, w, false, true); + m->zeroMem(); + *p = m->getData(); + }; + + real* gammaGrad = nullptr; + real* betaGrad = nullptr; + if (weight_->getWGrad()) { + gammaGrad = weight_->getWGrad()->getData(); + } else { + create(tmpWGrad_, 1, channels_, &gammaGrad); + } + if (biases_ && biases_->getWGrad()) { + betaGrad = biases_->getWGrad()->getData(); + } else { + create(tmpBiasGrad_, 1, channels_, &betaGrad); + } + + hl_batch_norm_backward(ioDesc_, + input, + ioDesc_, + outGrad, + ioDesc_, + inGrad, + bnParamDesc_, + gamma, + gammaGrad, + betaGrad, + eps_, + savedMean, + savedInvVar); + + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + biases_->getParameterPtr()->incUpdate(callback); + weight_->getParameterPtr()->incUpdate(callback); + } +} + +CudnnBatchNormLayer::~CudnnBatchNormLayer() { + hl_destroy_tensor_descriptor(ioDesc_); + hl_destroy_tensor_descriptor(bnParamDesc_); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CudnnBatchNormLayer.h b/paddle/legacy/gserver/layers/CudnnBatchNormLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..3b33b983b31173ab941df5f2e66eac51aabc6315 --- /dev/null +++ b/paddle/legacy/gserver/layers/CudnnBatchNormLayer.h @@ -0,0 +1,68 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "BatchNormBaseLayer.h" +#include "Layer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief Cudnn Batch normalization layer use to cuDNN lib to implentment. + * @note Cudnn version must >= v4.0, and better to use the latest version + * (v5.1). + * + * The config file api is batch_norm_layer. + */ + +class CudnnBatchNormLayer : public BatchNormBaseLayer { + public: + explicit CudnnBatchNormLayer(const LayerConfig& config) + : BatchNormBaseLayer(config) {} + + ~CudnnBatchNormLayer(); + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + /** + * reshape tensor of ioDesc_. + */ + void reshape(int batchSize); + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + protected: + /// Epsilon value used in the batch normalization formula. + /// Same epsilon value should be used in forward and backward functions. + double eps_; + + /// Input/output tensor descriptor desc + hl_tensor_descriptor ioDesc_; + /// Shared tensor descriptor desc for the 6 tenros: + /// bnScale, bnBias, running mean/var, save_mean/var + hl_tensor_descriptor bnParamDesc_; + + /** + * @brief The gradient of weight and bias in cudnn api can not be empty. + * If set is_static for weight or bias, it will not allocate memory for them, + * and the gradient is NULL. In this case, will use two matrix. + */ + MatrixPtr tmpWGrad_, tmpBiasGrad_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CudnnConvBaseLayer.cpp b/paddle/legacy/gserver/layers/CudnnConvBaseLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9353cca9c83bd90a454b2be56dc08b8eadee0bf7 --- /dev/null +++ b/paddle/legacy/gserver/layers/CudnnConvBaseLayer.cpp @@ -0,0 +1,135 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CudnnConvBaseLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { +REGISTER_LAYER(cudnn_conv, CudnnConvBaseLayer); +REGISTER_LAYER(cudnn_convt, CudnnConvBaseLayer); + +bool CudnnConvBaseLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; + CHECK(useGpu_) << "CudnnConvLayer only support gpu"; + + CHECK_EQ(inputLayers_.size(), parameters_.size()); + projections_.reserve(inputLayers_.size()); + projConf_.reserve(inputLayers_.size()); + + numFilters_ = config_.num_filters(); + CHECK(config_.shared_biases()); + for (size_t i = 0; i < inputLayers_.size(); i++) { + ProjectionConfig *conf = new ProjectionConfig(); + if (isDeconv_) { + conf->set_type("convt"); + } else { + conf->set_type("conv"); + } + conf->set_num_filters(numFilters_); + ConvConfig *convConf = conf->mutable_conv_conf(); + *convConf = *(config_.mutable_inputs(i)->mutable_conv_conf()); + conf->set_input_size(getPrev(i)->getSize()); + conf->set_output_size(getSize()); + projConf_.emplace_back(conf); + projections_.emplace_back( + Projection::create(*projConf_[i], parameters_[i], useGpu_)); + + // create a new weight + size_t height, width; + height = filterPixels_[i] * filterChannels_[i]; + width = (!isDeconv_) ? numFilters_ : channels_[i]; + CHECK_EQ(parameters_[i]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[i]); + weights_.emplace_back(w); + } + + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } + if (biases_.get() && sharedBiases_) { + hl_create_tensor_descriptor(&biasDesc_); + hl_create_tensor_descriptor(&outputDesc_); + hl_tensor_reshape(biasDesc_, 1, numFilters_, 1, 1); + } + + return true; +} + +void CudnnConvBaseLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = getInput(0).getBatchSize(); + resetOutput(batchSize, calOutputSize()); + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + projections_[i]->forward(&getInput(i), &getOutput(), passType); + } + + if (biases_) { + REGISTER_TIMER_INFO("CudnnConvBiasTimer", getName().c_str()); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + int outH = outputH_[0]; + int outW = outputW_[0]; + + hl_tensor_reshape(outputDesc_, + batchSize, + numFilters_, + outH, + outW, + numFilters_ * outH * outW, + outH * outW, + outW, + 1); + real *outData = getOutputValue()->getData(); + real *biasData = biases_->getW()->getData(); + hl_convolution_forward_add_bias(biasDesc_, biasData, outputDesc_, outData); + } + + forwardActivation(); +} + +void CudnnConvBaseLayer::backward(const UpdateCallback &callback) { + backwardActivation(); + + if (biases_ && biases_->getWGrad()) { + REGISTER_TIMER_INFO("CudnnConvBpBiasTimer", getName().c_str()); + real *biasGrad = biases_->getWGrad()->getData(); + real *outGrad = getOutputGrad()->getData(); + hl_convolution_backward_bias(biasDesc_, biasGrad, outputDesc_, outGrad); + + biases_->getParameterPtr()->incUpdate(callback); + } + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + projections_[i]->backward(callback); + } +} + +CudnnConvBaseLayer::~CudnnConvBaseLayer() { + if (biases_) { + hl_destroy_tensor_descriptor(biasDesc_); + hl_destroy_tensor_descriptor(outputDesc_); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CudnnConvBaseLayer.h b/paddle/legacy/gserver/layers/CudnnConvBaseLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..d050183eb7838bed803995985383e0ee4e9731a1 --- /dev/null +++ b/paddle/legacy/gserver/layers/CudnnConvBaseLayer.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "ConvBaseLayer.h" +#include "Projection.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief A 2-dimension conv layer implemented by cuDNN. It only + * supports GPU mode. We automatic select CudnnConvLayer for GPU + * mode and ExpandConvLayer for CPU mode if you set type of "conv". + * User also can specfiy type of "exconv" or "cudnn_conv" for + * particular type. + * + * The config file api is img_conv_layer. + */ +class CudnnConvBaseLayer : public ConvBaseLayer { + protected: + std::vector> projConf_; + std::vector> projections_; + + hl_tensor_descriptor biasDesc_; + hl_tensor_descriptor outputDesc_; + + public: + explicit CudnnConvBaseLayer(const LayerConfig& config) + : ConvBaseLayer(config) {} + + ~CudnnConvBaseLayer(); + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/CudnnPoolLayer.cpp b/paddle/legacy/gserver/layers/CudnnPoolLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c790dfd71efbee1a2a0afa69e6c336c4330737d0 --- /dev/null +++ b/paddle/legacy/gserver/layers/CudnnPoolLayer.cpp @@ -0,0 +1,139 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CudnnPoolLayer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +bool CudnnPoolLayer::typeCheck(const std::string &poolType, + hl_pooling_mode_t *mode) { + if (poolType == "cudnn-max-pool") { + if (mode) { + *mode = HL_POOLING_MAX; + } + } else if (poolType == "cudnn-avg-pool") { + if (mode) { + *mode = HL_POOLING_AVERAGE; + } + } else if (poolType == "cudnn-avg-incl-pad-pool") { + if (mode) { + *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING; + } + } else { + return false; + } + + return true; +} + +CudnnPoolLayer::CudnnPoolLayer(const LayerConfig &config) : PoolLayer(config) { + const std::string &pool_type = config.inputs(0).pool_conf().pool_type(); + CHECK_EQ(CudnnPoolLayer::typeCheck(pool_type, &mode_), true); +} + +bool CudnnPoolLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + PoolLayer::init(layerMap, parameterMap); + + CHECK(useGpu_) << "CudnnPoolLayer only support gpu"; + + hl_create_tensor_descriptor(&inputDesc_); + hl_create_tensor_descriptor(&outputDesc_); + + windowHeight = sizeY_; + windowWidth = sizeX_; + heightPadding = confPaddingY_; + widthPadding = confPadding_; + strideHeight = strideY_; + strideWidth = stride_; + + hl_create_pooling_descriptor(&poolingDesc_, + mode_, + windowHeight, + windowWidth, + heightPadding, + widthPadding, + strideHeight, + strideWidth); + + return true; +} + +void CudnnPoolLayer::reshape(int batchSize) { + imageH_ = inputLayers_[0]->getOutput().getFrameHeight(); + imageW_ = inputLayers_[0]->getOutput().getFrameWidth(); + if (imageH_ == 0) { + imageH_ = imgSizeY_; + } + if (imageW_ == 0) { + imageW_ = imgSize_; + } + CHECK_EQ(inputLayers_[0]->getOutput().value->getWidth(), + channels_ * imageH_ * imageW_); + outputH_ = outputSize(imageH_, + sizeY_, + confPaddingY_, + strideY_, + /* caffeMode */ false); + outputW_ = + outputSize(imageW_, sizeX_, confPadding_, stride_, /* caffeMode */ false); + getOutput().setFrameHeight(outputH_); + getOutput().setFrameWidth(outputW_); + + hl_tensor_reshape(inputDesc_, batchSize, channels_, imageH_, imageW_); + hl_tensor_reshape(outputDesc_, batchSize, channels_, outputH_, outputW_); +} + +void CudnnPoolLayer::forward(PassType passType) { + Layer::forward(passType); + + CHECK(inputLayers_[0]->getOutputValue()->useGpu()); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + reshape(batchSize); + resetOutput(batchSize, outputH_ * outputW_ * channels_); + + real *inputData = getInputValue(0)->getData(); + real *outData = getOutputValue()->getData(); + hl_pooling_forward(inputDesc_, inputData, outputDesc_, outData, poolingDesc_); +} + +void CudnnPoolLayer::backward(const UpdateCallback &callback) { + (void)callback; + if (NULL == getInputGrad(0)) { + return; + } + + real *inputData = getInputValue(0)->getData(); + real *inputGrad = getInputGrad(0)->getData(); + real *outData = getOutputValue()->getData(); + real *outGrad = getOutputGrad()->getData(); + hl_pooling_backward(inputDesc_, + inputData, + inputGrad, + outputDesc_, + outData, + outGrad, + poolingDesc_); +} + +CudnnPoolLayer::~CudnnPoolLayer() { + hl_destroy_tensor_descriptor(inputDesc_); + hl_destroy_tensor_descriptor(outputDesc_); + hl_destroy_pooling_descriptor(poolingDesc_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/CudnnPoolLayer.h b/paddle/legacy/gserver/layers/CudnnPoolLayer.h similarity index 100% rename from paddle/gserver/layers/CudnnPoolLayer.h rename to paddle/legacy/gserver/layers/CudnnPoolLayer.h diff --git a/paddle/gserver/layers/DataLayer.cpp b/paddle/legacy/gserver/layers/DataLayer.cpp similarity index 100% rename from paddle/gserver/layers/DataLayer.cpp rename to paddle/legacy/gserver/layers/DataLayer.cpp diff --git a/paddle/gserver/layers/DataLayer.h b/paddle/legacy/gserver/layers/DataLayer.h similarity index 100% rename from paddle/gserver/layers/DataLayer.h rename to paddle/legacy/gserver/layers/DataLayer.h diff --git a/paddle/legacy/gserver/layers/DataNormLayer.cpp b/paddle/legacy/gserver/layers/DataNormLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6820dfa4d4dcf90b2318a190ad4cc082c26fc180 --- /dev/null +++ b/paddle/legacy/gserver/layers/DataNormLayer.cpp @@ -0,0 +1,140 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "DataNormLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(data_norm, DataNormLayer); + +bool DataNormLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* initialize the weight */ + CHECK(!biasParameter_) << "DataNormLayer does not need bias"; + CHECK(inputLayers_.size() == 1 && inputLayers_[0]->getType() == "data") + << "DataNormLayer accepts one and only one DataLayer as its input layer"; + CHECK_EQ(inputLayers_.size(), parameters_.size()); + CHECK_EQ(inputLayers_[0]->getSize(), getSize()); + CHECK_EQ(parameters_[0]->getSize(), 5 * getSize()); + CHECK(parameters_[0]->isStatic()) + << "The parameter of DataNormLayer must be static"; + + weight_ = std::unique_ptr(new Weight(5, getSize(), parameters_[0])); + min_ = Matrix::create( + nullptr, /* height= */ 1, getSize(), /* trans= */ false, useGpu_); + rangeReciprocal_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + mean_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + stdReciprocal_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + decimalReciprocal_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + + min_->setData(weight_->getW()->getData()); + rangeReciprocal_->setData(weight_->getW()->getData() + getSize()); + mean_->setData(weight_->getW()->getData() + 2 * getSize()); + stdReciprocal_->setData(weight_->getW()->getData() + 3 * getSize()); + decimalReciprocal_->setData(weight_->getW()->getData() + 4 * getSize()); + + /* normalization strategy */ + if (config_.data_norm_strategy() == "z-score") { + mode_ = kZScore; + } else if (config_.data_norm_strategy() == "min-max") { + mode_ = kMinMax; + } else if (config_.data_norm_strategy() == "decimal-scaling") { + mode_ = kDecimalScaling; + } else { + LOG(FATAL) << "Unknown data normalization strategy: " + << config_.data_norm_strategy(); + } + + return true; +} + +void DataNormLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInput(0).getBatchSize(); + int size = getSize(); + reserveOutput(batchSize, size); + + const MatrixPtr inValue = getInputValue(0); + MatrixPtr outValue = getOutputValue(); + outValue->copyFrom(*inValue); + switch (mode_) { + case kZScore: { + outValue->addBias(*mean_, -1.0); + outValue->colScale(0, *outValue, *stdReciprocal_); + break; + } + case kMinMax: { + outValue->addBias(*min_, -1.0); + outValue->colScale(0, *outValue, *rangeReciprocal_); + break; + } + case kDecimalScaling: { + outValue->colScale(0, *outValue, *decimalReciprocal_); + break; + } + default: + LOG(FATAL) << "should not reach here"; + } +} + +void DataNormLayer::backward(const UpdateCallback& callback) { + // The parameter for DataNormLayer is static, and does not need to be updated + (void)callback; + + /* Calculate the input layers error */ + const MatrixPtr& outGrad = getOutputGrad(); + MatrixPtr inGrad = getInputGrad(0); + if (inGrad) { + switch (mode_) { + case kZScore: { + inGrad->addColScale(0, *outGrad, *stdReciprocal_); + break; + } + case kMinMax: { + inGrad->addColScale(0, *outGrad, *rangeReciprocal_); + break; + } + case kDecimalScaling: { + inGrad->addColScale(0, *outGrad, *decimalReciprocal_); + break; + } + default: { LOG(FATAL) << "should not reach here"; } + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/DataNormLayer.h b/paddle/legacy/gserver/layers/DataNormLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..7bb8e928248355cb7ae78dc16e467b77a42e02fc --- /dev/null +++ b/paddle/legacy/gserver/layers/DataNormLayer.h @@ -0,0 +1,62 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * @brief A layer for data normalization + * - Input: One and only one input layer is accepted. The input layer must + * be DataLayer with dense data type. + * - Output: The normalization of the input data + * + * Reference: + * LA Shalabi, Z Shaaban, B Kasasbeh. Data mining: A preprocessing engine + * + * Three data normalization methoeds are considered + * - z-score: y = (x-mean)/std + * - min-max: y = (x-min)/(max-min) + * - decimal-scaling: y = x/10^j, where j is the smallest integer such that + *max(|y|)<1 + */ + +class DataNormLayer : public Layer { + public: + enum NormalizationStrategy { kZScore = 0, kMinMax = 1, kDecimalScaling = 2 }; + + explicit DataNormLayer(const LayerConfig& config) : Layer(config) {} + + ~DataNormLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + protected: + int mode_; + std::unique_ptr weight_; + MatrixPtr min_; + MatrixPtr rangeReciprocal_; // 1/(max-min) + MatrixPtr mean_; + MatrixPtr stdReciprocal_; // 1/std + MatrixPtr decimalReciprocal_; // 1/10^j +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/DeConv3DLayer.cpp b/paddle/legacy/gserver/layers/DeConv3DLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2cd635564c4cd9f42d27cd58694cff381d1ce224 --- /dev/null +++ b/paddle/legacy/gserver/layers/DeConv3DLayer.cpp @@ -0,0 +1,220 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "DeConv3DLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(deconv3d, DeConv3DLayer); + +bool DeConv3DLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; + // for Deconv, the dimension of Kernel is + // channel * output * depth * height * weigth + // Matrix storage format: (output * depth * height * weigth) x channel + for (int index = 0; index < config_.inputs().size(); ++index) { + M_.push_back(filterChannels_[index]); + K_.push_back(filterPixels_[index] * (numFilters_ / groups_[index])); + + // create a new weight + size_t height, width; + height = filterPixels_[index] * numFilters_; + width = filterChannels_[index]; + CHECK_EQ(parameters_[index]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[index]); + weights_.emplace_back(w); + } + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } + return true; +} + +size_t DeConv3DLayer::getSize() { + CHECK_NE(inputLayers_.size(), 0UL); + imgSizeW_.clear(); + imgSizeH_.clear(); + imgSizeD_.clear(); + N_.clear(); + NOut_.clear(); + size_t layerSize = 0; + for (size_t i = 0; i < inputLayers_.size(); ++i) { + imgSizeW_.push_back( + imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i], true)); + imgSizeH_.push_back(imageSize( + outputH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true)); + imgSizeD_.push_back(imageSize( + outputD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true)); + NOut_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]); + N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]); + CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize); + layerSize += NOut_[i] * numFilters_; + } + getOutput().setFrameHeight(imgSizeH_[0]); + getOutput().setFrameWidth(imgSizeW_[0]); + getOutput().setFrameDepth(imgSizeD_[0]); + return layerSize; +} + +void DeConv3DLayer::forward(PassType passType) { + Layer::forward(passType); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + int outWidth = getSize(); + resetOutput(batchSize, outWidth); + const MatrixPtr outMat = getOutputValue(); + + REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str()); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + const MatrixPtr &inMat = getInputValue(i); + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + MatrixPtr wMat = weights_[i]->getW(); + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + for (int n = 0; n < batchSize; ++n) { + real *inData = inMat->getData() + n * inMat->getStride(); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_); + MatrixPtr wMatSub = wMat->subMatrix(g * K, K); + MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K); + colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0); + inData += M * N; + } + colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(), + numFilters_, + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i], + 1.0, + 1.0); + } + } + if (nullptr != this->biasParameter_) { + this->addBias(); + } + forwardActivation(); +} + +void DeConv3DLayer::backward(const UpdateCallback &callback) { + backwardActivation(); + int batchSize = getOutputGrad()->getHeight(); + if (biases_ && biases_->getWGrad()) { + bpropBiases(); + biases_->getParameterPtr()->incUpdate(callback); + } + REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str()); + for (size_t i = 0; i < inputLayers_.size(); ++i) { + if (weights_[i]->getWGrad() || this->needGradient_) { + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + const MatrixPtr &inMat = getInputValue(i); + for (int n = 0; n < batchSize; ++n) { + colBuf_->vol2Col( + getOutputGrad()->getData() + n * getOutputGrad()->getStride(), + numFilters_, + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i]); + if (weights_[i]->getWGrad()) { + real *inData = inMat->getData() + n * inMat->getStride(); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K); + MatrixPtr wGradMatSub = + weights_[i]->getWGrad()->subMatrix(g * K, K); + MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_); + wGradMatSub->mul( + *colBufDataSub, *(inMatSub->getTranspose()), 1.0, 1.0); + inData += M * N; + } + } + if (getInputGrad(i)) { + real *preGrad = + getInputGrad(i)->getData() + n * getInputGrad(i)->getStride(); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr w = weights_[i]->getW()->subMatrix(g * K, K); + MatrixPtr outGradMat = colBuf_->subMatrix(g * K, K); + MatrixPtr inGradMatSub = + Matrix::create(preGrad, M, N, false, useGpu_); + inGradMatSub->mul(*(w->getTranspose()), *outGradMat, 1.0, 1.0); + preGrad += M * N; + } + } + } + weights_[i]->getParameterPtr()->incUpdate(callback); + } + } +} +void DeConv3DLayer::bpropWeights(int i) {} +void DeConv3DLayer::bpropData(int i) {} + +void DeConv3DLayer::bpropBiases() { + MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), + 1, + biases_->getWGrad()->getElementCnt(), + false, + useGpu_); + const MatrixPtr &outGradMat = getOutputGrad(); + + if (this->sharedBiases_) { + biases->collectSharedBias(*outGradMat, 1.0f); + } else { + biases->collectBias(*outGradMat, 1.0f); + } +} + +void DeConv3DLayer::addBias() { + MatrixPtr outMat = getOutputValue(); + MatrixPtr bias = Matrix::create(biases_->getW()->getData(), + 1, + biases_->getW()->getElementCnt(), + false, + useGpu_); + if (this->sharedBiases_) { + outMat->addSharedBias(*(bias), 1.0f); + } else { + outMat->addBias(*(bias), 1.0f); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/DeConv3DLayer.h b/paddle/legacy/gserver/layers/DeConv3DLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..9931bccb1284111e299206883847045edaae4ded --- /dev/null +++ b/paddle/legacy/gserver/layers/DeConv3DLayer.h @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "ConvBaseLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief A subclass of deconvolution3D layer. + * This layer expands input and use matrix multiplication to + * calculate deconvolution3D operation. + */ +class DeConv3DLayer : public ConvBaseLayer { + public: + explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} + ~DeConv3DLayer() {} + bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + + void forward(PassType passType); + void addBias(); + void backward(const UpdateCallback& callback); + void bpropBiases(); + void bpropData(int i); + void bpropWeights(int i); + size_t getSize(); + + protected: + // Figure out the dimensions for individual gemms. + IntV M_; /// numFilters_ / filter_group_; + IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ + IntV K_; /// outputD_ * outputH_ * outputW_ + IntV NOut_; + MatrixPtr colBuf_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/DetectionOutputLayer.cpp b/paddle/legacy/gserver/layers/DetectionOutputLayer.cpp similarity index 100% rename from paddle/gserver/layers/DetectionOutputLayer.cpp rename to paddle/legacy/gserver/layers/DetectionOutputLayer.cpp diff --git a/paddle/gserver/layers/DetectionOutputLayer.h b/paddle/legacy/gserver/layers/DetectionOutputLayer.h similarity index 100% rename from paddle/gserver/layers/DetectionOutputLayer.h rename to paddle/legacy/gserver/layers/DetectionOutputLayer.h diff --git a/paddle/gserver/layers/DetectionUtil.cpp b/paddle/legacy/gserver/layers/DetectionUtil.cpp similarity index 100% rename from paddle/gserver/layers/DetectionUtil.cpp rename to paddle/legacy/gserver/layers/DetectionUtil.cpp diff --git a/paddle/legacy/gserver/layers/DetectionUtil.h b/paddle/legacy/gserver/layers/DetectionUtil.h new file mode 100644 index 0000000000000000000000000000000000000000..c1e0bb809ad290613159f558e9b1860476b3b5f2 --- /dev/null +++ b/paddle/legacy/gserver/layers/DetectionUtil.h @@ -0,0 +1,307 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include "paddle/legacy/math/Matrix.h" + +using std::vector; +using std::pair; +using std::map; + +namespace paddle { + +template +struct BBoxBase { + BBoxBase(T xMin, T yMin, T xMax, T yMax) + : xMin(xMin), yMin(yMin), xMax(xMax), yMax(yMax), isDifficult(false) {} + + BBoxBase() {} + + T getWidth() const { return xMax - xMin; } + + T getHeight() const { return yMax - yMin; } + + T getCenterX() const { return (xMin + xMax) / 2; } + + T getCenterY() const { return (yMin + yMax) / 2; } + + T getArea() const { return getWidth() * getHeight(); } + + // coordinate of bounding box + T xMin; + T yMin; + T xMax; + T yMax; + // whether difficult object (e.g. object with heavy occlusion is difficult) + bool isDifficult; +}; + +struct NormalizedBBox : BBoxBase { + NormalizedBBox() : BBoxBase() {} +}; + +enum PermMode { kNCHWToNHWC, kNHWCToNCHW }; + +/** + * @brief First permute input maxtrix then append to output matrix + */ +size_t appendWithPermute(const Matrix& inMatrix, + size_t height, + size_t width, + size_t outTotalSize, + size_t outOffset, + size_t batchSize, + Matrix& outMatrix, + PermMode permMode); + +/** + * @brief First permute input maxtrix then decompose to output + */ +size_t decomposeWithPermute(const Matrix& inMatrix, + size_t height, + size_t width, + size_t totalSize, + size_t offset, + size_t batchSize, + Matrix& outMatrix, + PermMode permMode); + +/** + * @brief Compute jaccard overlap between two bboxes. + * @param bbox1 The first bbox + * @param bbox2 The second bbox + */ +real jaccardOverlap(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2); + +/** + * @brief Compute offset parameters between prior bbox and ground truth bbox + * and variances of prior bbox are considered + * @param priorBBox Input prior bbox + * @param priorBBoxVar Variance parameters of prior bbox + * @param gtBBox Groundtruth bbox + * @param outVec Output vector + */ +void encodeBBoxWithVar(const NormalizedBBox& priorBBox, + const vector& priorBBoxVar, + const NormalizedBBox& gtBBox, + vector& outVec); + +/** + * @brief Decode prior bbox with offset parameters + * and variances of prior bbox are considered + * @param priorBBox Prior bbox to be decoded + * @param priorBBoxVar Variance parameters of prior bbox + * @param locPredData Offset parameters + */ +NormalizedBBox decodeBBoxWithVar(const NormalizedBBox& priorBBox, + const vector& priorBBoxVar, + const vector& locPredData); + +/** + * @brief Extract bboxes from prior matrix, the layout is + * xmin1 | ymin1 | xmax1 | ymax1 | xmin1Var | ymin1Var | xmax1Var | ymax1Var ... + * @param priorData Matrix of prior value + * @param numBBoxes Number of bbox to be extracted + * @param bboxVec Append to the vector + */ +void getBBoxFromPriorData(const real* priorData, + const size_t numBBoxes, + vector& bboxVec); + +/** + * @brief Extract labels, scores and bboxes from detection matrix, the layout is + * imageId | label | score | xmin | ymin | xmax | ymax + * @param detectData Matrix of detection value + * @param numBBoxes Number of bbox to be extracted + * @param labelVec Label of bbox + * @param scoreVec Score of bbox + * @param bboxVec Append to the vector + */ +void getBBoxFromDetectData(const real* detectData, + const size_t numBBoxes, + vector& labelVec, + vector& scoreVec, + vector& bboxVec); + +/** + * @brief Extract variances from prior matrix, the layout is + * xmin1 | ymin1 | xmax1 | ymax1 | xmin1Var | ymin1Var | xmax1Var | ymax1Var ... + * @param priorData Matrix of prior value + * @param num Number to be extracted + * @param varVec Append to the vector + */ +void getBBoxVarFromPriorData(const real* priorData, + const size_t num, + vector>& varVec); + +/** + * @brief Extract bboxes from label matrix, the layout is + * class1_1 | xmin1_1 | ymin1_1 | xmax1_1 | ymax1_1 | difficult1_1 | ... + * @param labelData Matrix of label value + * @param numBBoxes Number to be extracted + * @param bboxVec Append to the vector + */ +void getBBoxFromLabelData(const real* labelData, + const size_t numBBoxes, + vector& bboxVec); + +/** +* @brief Match prior bbox to groundtruth bbox, the strategy is: +1. Find the most overlaped bbox pair (prior and groundtruth) +2. For rest of prior bboxes find the most overlaped groundtruth bbox +* @param priorBBoxes prior bbox +* @param gtBBoxes groundtruth bbox +* @param overlapThreshold Low boundary of overlap (judge whether matched) +* @param matchIndices For each prior bbox, groundtruth bbox index if matched +otherwise -1 +* @param matchOverlaps For each prior bbox, overap with all groundtruth bboxes +*/ +void matchBBox(const vector& priorBBoxes, + const vector& gtBBoxes, + real overlapThreshold, + vector* matchIndices, + vector* matchOverlaps); + +/** +* @brief Generate positive bboxes and negative bboxes, +|positive bboxes|/|negative bboxes| is negPosRatio +* @param priorValue Prior value +* @param numPriorBBoxes Number of prior bbox +* @param gtValue Groundtruth value +* @param gtStartPosPtr Since groundtruth value stored as sequence type, +this parameter indicates start position of each record +* @param seqNum Number of sequence +* @param maxConfScore Classification score for prior bbox, used to mine +negative examples +* @param batchSize Image number +* @param overlapThreshold Low boundary of overap +* @param negOverlapThreshold Upper boundary of overap (judge negative example) +* @param negPosRatio Control number of negative bboxes +* @param matchIndicesVecPtr Save indices of matched prior bbox +* @param negIndicesVecPtr Save indices of negative prior bbox +*/ +pair generateMatchIndices( + const Matrix& priorValue, + const size_t numPriorBBoxes, + const Matrix& gtValue, + const int* gtStartPosPtr, + const size_t seqNum, + const vector>& maxConfScore, + const size_t batchSize, + const real overlapThreshold, + const real negOverlapThreshold, + const size_t negPosRatio, + vector>* matchIndicesVecPtr, + vector>* negIndicesVecPtr); + +/** + * @brief Get max confidence score for each prior bbox + * @param confData Confidence scores, layout is + * class1 score | class2 score | ... | classN score ... + * @param batchSize Image number + * @param numPriorBBoxes Prior bbox number + * @param numClasses Classes number + * @param backgroundId Background id + * @param maxConfScoreVecPtr Ouput + */ +void getMaxConfidenceScores(const real* confData, + const size_t batchSize, + const size_t numPriorBBoxes, + const size_t numClasses, + const size_t backgroundId, + vector>* maxConfScoreVecPtr); + +template +bool sortScorePairDescend(const pair& pair1, + const pair& pair2); + +template <> +bool sortScorePairDescend(const pair& pair1, + const pair& pair2); + +/** + * @brief Do NMS for bboxes to remove duplicated bboxes + * @param bboxes BBoxes to apply NMS + * @param confScoreData Confidence scores + * @param classIdx Class to do NMS + * @param topK Number to keep + * @param confThreshold Low boundary of confidence score + * @param nmsThreshold Threshold of overlap + * @param numPriorBBoxes Total number of prior bboxes + * @param numClasses Total class number + * @param indices Indices of high quality bboxes + */ +void applyNMSFast(const vector& bboxes, + const real* confScoreData, + size_t classIdx, + size_t topK, + real confThreshold, + real nmsThreshold, + size_t numPriorBBoxes, + size_t numClasses, + vector* indices); + +/** + * @brief Get detection results which satify requirements + * @param numPriorBBoxes Prior bbox number + * @param numClasses Class number + * @param backgroundId Background class + * @param batchSize Image number + * @param confThreshold Threshold of class confidence + * @param nmsTopK Used in NMS operation to keep top k bbox + * @param nmsThreshold Used in NMS, threshold of overlap + * @param keepTopK How many bboxes keeped in an image + * @param allDecodedBBoxes Decoded bboxes for all images + * @param allDetectionIndices Save detection bbox indices + */ +size_t getDetectionIndices( + const real* confData, + const size_t numPriorBBoxes, + const size_t numClasses, + const size_t backgroundId, + const size_t batchSize, + const real confThreshold, + const size_t nmsTopK, + const real nmsThreshold, + const size_t keepTopK, + const vector>& allDecodedBBoxes, + vector>>* allDetectionIndices); + +/** + * @brief Get detection results + * @param confData Confidence scores + * @param numPriorBBoxes Prior bbox number + * @param numClasses Class number + * @param batchSize Image number + * @param allIndices Indices of predicted bboxes + * @param allDecodedBBoxes BBoxes decoded + * @param out Output matrix + * image number | label | confidence score | xMin | yMin | xMax | yMax + */ +void getDetectionOutput(const real* confData, + const size_t numKept, + const size_t numPriorBBoxes, + const size_t numClasses, + const size_t batchSize, + const vector>>& allIndices, + const vector>& allDecodedBBoxes, + Matrix& out); + +NormalizedBBox clipBBox(const NormalizedBBox& bbox); + +} // namespace paddle diff --git a/paddle/gserver/layers/DotMulOperator.cpp b/paddle/legacy/gserver/layers/DotMulOperator.cpp similarity index 100% rename from paddle/gserver/layers/DotMulOperator.cpp rename to paddle/legacy/gserver/layers/DotMulOperator.cpp diff --git a/paddle/gserver/layers/DotMulProjection.cpp b/paddle/legacy/gserver/layers/DotMulProjection.cpp similarity index 100% rename from paddle/gserver/layers/DotMulProjection.cpp rename to paddle/legacy/gserver/layers/DotMulProjection.cpp diff --git a/paddle/legacy/gserver/layers/DotProdLayer.cpp b/paddle/legacy/gserver/layers/DotProdLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..06060d93f76c18d893852a5f5c99c36fe5641b2e --- /dev/null +++ b/paddle/legacy/gserver/layers/DotProdLayer.cpp @@ -0,0 +1,97 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief A layer for computing the dot product of two vectors. + * Input1: vector (batchSize * dim) + * Input2: vector (batchSize * dim) + * Output: a matrix: (batchSize * 1) + */ + +class DotProdLayer : public Layer { + public: + explicit DotProdLayer(const LayerConfig& config) : Layer(config) {} + + ~DotProdLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(dot_prod, DotProdLayer); + +bool DotProdLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2U); + CHECK_EQ(1UL, getSize()) + << "The output dimensionality of this layer should be fixed to 1."; + + return true; +} + +void DotProdLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV0->getHeight(); + CHECK_EQ(inV1->getHeight(), batchSize); + CHECK_EQ(inV0->getWidth(), inV1->getWidth()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, 1); + } + + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwDotProdTimer", getName().c_str()); + outV->sumOfProducts(*inV0, *inV1, 1, 0); + } +} + +void DotProdLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr outG = getOutputGrad(); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + + { + REGISTER_TIMER_INFO("BwDotProdTimer", getName().c_str()); + + if (inG0) { + inG0->addRowScale(0, *inV1, *outG); + } + + if (inG1) { + inG1->addRowScale(0, *inV0, *outG); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/EosIdCheckLayer.cpp b/paddle/legacy/gserver/layers/EosIdCheckLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..38671126c62ba36e22496dcbe1ff3c8d6dcea742 --- /dev/null +++ b/paddle/legacy/gserver/layers/EosIdCheckLayer.cpp @@ -0,0 +1,50 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { +/** + * A layer for checking EOS for each sample: + * - output_id = (input_id == conf.eos_id) + * + * The result is stored in output_.ids. + * It is used by recurrent layer group. + */ +class EosIdCheckLayer : public Layer { + public: + explicit EosIdCheckLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override { + bool ret = Layer::init(layerMap, parameterMap); + CHECK_EQ(1UL, inputLayers_.size()); + return ret; + } + + void forward(PassType passType) override { + Layer::forward(passType); + + const Argument& input = getInput(0); + IVector::resizeOrCreate(output_.ids, input.ids->getSize(), useGpu_); + output_.ids->isEqualTo(*input.ids, config_.eos_id()); + } + + void backward(const UpdateCallback& callback) override {} +}; + +REGISTER_LAYER(eos_id, EosIdCheckLayer); + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ExpandConvLayer.cpp b/paddle/legacy/gserver/layers/ExpandConvLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8a53db380686cea2ad121c948c45a0fa1154381e --- /dev/null +++ b/paddle/legacy/gserver/layers/ExpandConvLayer.cpp @@ -0,0 +1,248 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ExpandConvLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +DEFINE_bool(use_nnpack, + false, + "Whether to use nnpack for convolution calculation."); + +namespace paddle { + +/* + * The calculation of the exconvt(convolution transpose (deconv) operation) + * is a swap of forward and backward of the calculation of exconv. + * */ +REGISTER_LAYER(exconv, ExpandConvLayer); +REGISTER_LAYER(exconvt, ExpandConvLayer); + +inline bool isDepthwiseConv(int channels, int groups) { + return channels == groups; +} + +bool ExpandConvLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + /* Initialize the basic convolutional parent class */ + ConvBaseLayer::init(layerMap, parameterMap); + + int index = 0; + for (auto &inputConfig : config_.inputs()) { + const ConvConfig &conf = inputConfig.conv_conf(); + /* Consistent caffe mode for multiple input */ + caffeMode_ = conf.caffe_mode(); + + // create a new weight + size_t height, width; + height = filterPixels_[index] * filterChannels_[index]; + width = (!isDeconv_) ? numFilters_ : channels_[index]; + CHECK_EQ(parameters_[index]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[index]); + weights_.emplace_back(w); + index++; + } + + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = std::unique_ptr( + new Weight(1, numFilters_, biasParameter_, 0)); + } else { + biases_ = + std::unique_ptr(new Weight(1, getSize(), biasParameter_, 0)); + } + } + + getOutputSize(); + + size_t numInputs = config_.inputs_size(); + inputShape_.resize(numInputs); + filterShape_.resize(numInputs); + outputShape_.resize(numInputs); + + std::string convType; + std::string convGradInputType; + std::string convGradFilterType; + + for (int i = 0; i < config_.inputs_size(); i++) { + std::vector paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; + std::vector strides = {(size_t)strideY_[i], (size_t)stride_[i]}; + std::vector dilations = {(size_t)dilationY_[i], + (size_t)dilation_[i]}; + + bool useDilation = ((size_t)dilationY_[i] > 1 || (size_t)dilation_[i] > 1); + + // Convolution Layer uses the GemmConv function by default. + convType = "GemmConv"; + convGradInputType = "GemmConvGradInput"; + convGradFilterType = "GemmConvGradFilter"; + + // If depth wise convolution and useGpu == true + if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { + convType = "DepthwiseConv"; + convGradInputType = "DepthwiseConvGradInput"; + convGradFilterType = "DepthwiseConvGradFilter"; + } + + // If depth wise convolution and useGpu == false and ARM-NEON + if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + if ((filterSize_[i] == filterSizeY_[i]) && + (filterSize_[i] == 3 || filterSize_[i] == 4) && + (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2) && + !useDilation) { + convType = "NeonDepthwiseConv"; + } +#endif + } + + if (FLAGS_use_nnpack && !isDeconv_ && !useDilation) { + createFunction(forward_, + "NNPACKConv", + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)groups_[i]) + .set("algo", std::string("auto"))); + } else { + createFunction(forward_, + !isDeconv_ ? convType : convGradInputType, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("dilations", dilations) + .set("groups", (size_t)groups_[i])); + + createFunction(backward_, + !isDeconv_ ? convGradInputType : convType, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("dilations", dilations) + .set("groups", (size_t)groups_[i])); + + createFunction(backward_, + convGradFilterType, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("dilations", dilations) + .set("groups", (size_t)groups_[i])); + } + } + return true; +} + +size_t ExpandConvLayer::getOutputSize() { + CHECK_NE(inputLayers_.size(), 0UL); + size_t layerSize = ConvBaseLayer::calOutputSize(); + return layerSize; +} + +// i is the index of input layers +#define BACKWARD_INPUT(i, inputs, outputs) \ + backward_[2 * i]->calc(inputs, outputs) +#define BACKWARD_FILTER(i, inputs, outputs) \ + backward_[2 * i + 1]->calc(inputs, outputs) + +void ExpandConvLayer::forward(PassType passType) { + Layer::forward(passType); + + size_t batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + resetOutput(batchSize, getOutputSize()); + + // Calculate the shape of the input, output, and filter. + for (size_t i = 0; i < inputLayers_.size(); ++i) { + inputShape_[i] = TensorShape({(size_t)batchSize, + (size_t)channels_[i], + (size_t)imgSizeH_[i], + (size_t)imgSizeW_[i]}); + filterShape_[i] = + TensorShape({(size_t)groups_[i], + !isDeconv_ ? (size_t)numFilters_ / groups_[i] + : (size_t)channels_[i] / groups_[i], + !isDeconv_ ? (size_t)channels_[i] / groups_[i] + : (size_t)numFilters_ / groups_[i], + (size_t)filterSizeY_[i], + (size_t)filterSize_[i]}); + outputShape_[i] = TensorShape({(size_t)batchSize, + (size_t)numFilters_, + (size_t)outputH_[i], + (size_t)outputW_[i]}); + } + + // Calculate the output value. + for (size_t i = 0; i < inputLayers_.size(); ++i) { + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(i), inputShape_[i]); + inputs.addArg(*weights_[i]->getW(), filterShape_[i]); + outputs.addArg(*getOutputValue(), + outputShape_[i], + !isDeconv_ && i == 0 ? ASSIGN_TO : ADD_TO); + + forward_[i]->calc(inputs, outputs); + } + + /* add the bias-vector */ + if (biases_.get()) { + output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_); + } + + /* activation */ + forwardActivation(); +} + +void ExpandConvLayer::backward(const UpdateCallback &callback) { + backwardActivation(); + + MatrixPtr outGrad = getOutputGrad(); + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_); + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + // Calculate the input grad and filter grad. + for (size_t i = 0; i < inputLayers_.size(); ++i) { + if (getInputGrad(i)) { + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outputShape_[i]); + inputs.addArg(*weights_[i]->getW(), filterShape_[i]); + outputs.addArg(*getInputGrad(i), inputShape_[i], ADD_TO); + BACKWARD_INPUT(i, inputs, outputs); + } + + if (weights_[i]->getWGrad()) { + BufferArgs inputs; + BufferArgs outputs; + if (!isDeconv_) { + inputs.addArg(*getOutputGrad(), outputShape_[i]); + inputs.addArg(*getInputValue(i), inputShape_[i]); + } else { + inputs.addArg(*getInputValue(i), inputShape_[i]); + inputs.addArg(*getOutputGrad(), outputShape_[i]); + } + outputs.addArg(*weights_[i]->getWGrad(), filterShape_[i], ADD_TO); + BACKWARD_FILTER(i, inputs, outputs); + + /* Increasing the number of gradient */ + weights_[i]->getParameterPtr()->incUpdate(callback); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ExpandConvLayer.h b/paddle/legacy/gserver/layers/ExpandConvLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..c0eff3ab061949bd583e0deaf121912ed993be76 --- /dev/null +++ b/paddle/legacy/gserver/layers/ExpandConvLayer.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "ConvBaseLayer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief A subclass of convolution layer. + * This layer expands input and use matrix multiplication to + * calculate convolution operation. + * + * The config file api is img_conv_layer. + */ + +class ExpandConvLayer : public ConvBaseLayer { + public: + explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {} + + ~ExpandConvLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + + size_t getOutputSize(); + + protected: + std::vector inputShape_; + std::vector filterShape_; + std::vector outputShape_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ExpandLayer.cpp b/paddle/legacy/gserver/layers/ExpandLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..074fbab8ef9d1453160058031be370e991459fa5 --- /dev/null +++ b/paddle/legacy/gserver/layers/ExpandLayer.cpp @@ -0,0 +1,133 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ExpandLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(expand, ExpandLayer); + +bool ExpandLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_EQ(inputLayers_.size(), 2UL); + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + // which sequence type of input[0] + if (config_.trans_type() == "non-seq") { + type_ = kNonSeq; + } else if (config_.trans_type() == "seq") { + type_ = kSeq; + } else { + LOG(FATAL) << "Unknown trans_type: " << config_.trans_type(); + } + setNeedSequenceInfo(false); + return true; +} + +void ExpandLayer::forward(PassType passType) { + Layer::forward(passType); + // Expand layer should have exactly 2 input, one for data, one for size + CHECK_EQ(2U, inputLayers_.size()); + + // using two input: + // * first one for data; + // * second one only for sequence info + const Argument& shapeInput = getInput(1); + const Argument& dataInput = getInput(0); + size_t outputBatchSize = shapeInput.getBatchSize(); + auto startPositions = type_ ? shapeInput.subSequenceStartPositions + : shapeInput.sequenceStartPositions; + size_t numSequences = startPositions->getSize() - 1; + const int* starts = startPositions->getData(false); + + CHECK_EQ(starts[numSequences], shapeInput.getBatchSize()); + if (type_) { + // when trans_type = seq, input[1] must hasSubseq + CHECK_EQ(shapeInput.hasSubseq(), 1UL); + CHECK_EQ(dataInput.getNumSequences(), shapeInput.getNumSequences()); + } else { + CHECK_EQ(dataInput.getBatchSize(), shapeInput.getNumSequences()); + } + + // set output sequence info as shape sequence + output_.sequenceStartPositions = shapeInput.sequenceStartPositions; + if (shapeInput.hasSubseq()) { + output_.subSequenceStartPositions = shapeInput.subSequenceStartPositions; + } + + // reserve output: Expand output to batchsize of sequence data. + reserveOutput(outputBatchSize, dataInput.value->getWidth()); + + MatrixPtr inputValue = getInputValue(0); + MatrixPtr outputValue = getOutputValue(); + + ICpuGpuVector::resizeOrCreate(expandStartsPos_, outputBatchSize, false); + int* expandStarts = expandStartsPos_->getMutableData(false); + for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) { + int sequenceLength = starts[sequenceId + 1] - starts[sequenceId]; + for (int j = 0; j < sequenceLength; j++) { + expandStarts[starts[sequenceId] + j] = sequenceId; + } + } + + outputValue->copyByRowIndex(*inputValue, + *expandStartsPos_->getVector(useGpu_)); + + if (biases_.get() != NULL) { + outputValue->addBias(*(biases_->getW()), 1); + } +} + +void ExpandLayer::backward(const UpdateCallback& callback) { + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + if (!getInputGrad(0)) return; + MatrixPtr inputGrad = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + auto cpuSeqStartPos = type_ ? getInput(1).subSequenceStartPositions + : getInput(1).sequenceStartPositions; + size_t numSequences = cpuSeqStartPos->getSize() - 1; + const int* starts = cpuSeqStartPos->getData(false); + + CHECK_EQ(inputGrad->getWidth(), outputGrad->getWidth()); + CHECK_EQ(outputGrad->getHeight(), (size_t)starts[numSequences]); + + AsyncGpuBlock asyncGpuBlock; + + // sum to get the grad + real scale = 1; + for (size_t sequenceId = 0; sequenceId < numSequences; sequenceId++) { + // TODO(Dangqingqing) optimization for GPU + int sequenceLength = starts[sequenceId + 1] - starts[sequenceId]; + if (sequenceLength == 0) { + // empty sequence + continue; + } + MatrixPtr copyData = inputGrad->subMatrix(sequenceId, 1); + copyData->collectBias( + *outputGrad->subMatrix(starts[sequenceId], sequenceLength), scale); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ExpandLayer.h b/paddle/legacy/gserver/layers/ExpandLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..75a1ec75688cdbc61a117da7d4be47848c30425a --- /dev/null +++ b/paddle/legacy/gserver/layers/ExpandLayer.h @@ -0,0 +1,63 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * A layer for "Expand Dense data or (sequence data where the length of each + * sequence is one) to sequence data." + * + * It should have exactly 2 input, one for data, one for size: + * - first one for data + * - If ExpandLevel = kNonSeq: dense data + * - If ExpandLevel = kSeq: sequence data where the length of each sequence is + * one + * - second one only for sequence info + * - should be sequence data with or without sub-sequence. + * + * And the output size is the batch size(not instances) of second input. + * + * The config file api is expand_layer. + */ + +class ExpandLayer : public Layer { + protected: + std::unique_ptr biases_; + /// if input[0] is dense data, ExpandLevel=kNonSeq; + /// if input[0] is sequence data, ExpandLevel=kSeq + enum ExpandLevel { kNonSeq = 0, kSeq = 1 }; + /// store the ExpandLevel + int type_; + /// expanded sequenceStartPositions or subSequenceStartPositions + /// of input[1] + ICpuGpuVectorPtr expandStartsPos_; + + public: + explicit ExpandLayer(const LayerConfig& config) : Layer(config) {} + + ~ExpandLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/FactorizationMachineLayer.cpp b/paddle/legacy/gserver/layers/FactorizationMachineLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6cf269fa3ffb3f4a2864aea4225d9401930e73b1 --- /dev/null +++ b/paddle/legacy/gserver/layers/FactorizationMachineLayer.cpp @@ -0,0 +1,158 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "FactorizationMachineLayer.h" +#include +#include +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(factorization_machine, FactorizationMachineLayer); + +bool FactorizationMachineLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + factorSize_ = config_.factor_size(); + + /* initialize the latentVectors_ */ + CHECK_EQ(inputLayers_.size(), 1UL); + size_t inputSize = inputLayers_[0]->getSize(); + CHECK_EQ(parameters_[0]->getSize(), inputSize * factorSize_); + latentVectors_ = std::unique_ptr( + new Weight(inputSize, factorSize_, parameters_[0])); + + return true; +} + +void FactorizationMachineLayer::forward(PassType passType) { + Layer::forward(passType); + + const MatrixPtr& inputV = getInputValue(0); + + size_t batchSize = inputV->getHeight(); + size_t outputSize = getSize(); + size_t inputSize = inputLayers_[0]->getSize(); + reserveOutput(batchSize, outputSize); + + MatrixPtr outV = getOutputValue(); + + Matrix::resizeOrCreate( + latentVectorsSquare_, inputSize, factorSize_, false, useGpu_); + Matrix::resizeOrCreate( + inputMulFactor_, batchSize, factorSize_, false, useGpu_); + Matrix::resizeOrCreate(tmpOut_, batchSize, factorSize_, false, useGpu_); + + REGISTER_TIMER_INFO("FmInputMulFactorTimer", getName().c_str()); + inputMulFactor_->mul(*inputV, *latentVectors_->getW()); + inputMulFactor_->square2(*tmpOut_); + outV->sumRows(*tmpOut_, 0.5, 0); + + if (dynamic_cast(inputV.get())) { + Matrix::resizeOrCreateSparseMatrix(inputSquare_, + inputV->getHeight(), + inputV->getWidth(), + inputV->getElementCnt(), + inputV->getValueType()); + inputSquare_->copyFrom(*inputV); + (dynamic_cast(inputSquare_.get()))->square2(); + } else { + Matrix::resizeOrCreate( + inputSquare_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); + inputV->square2(*inputSquare_); + } + latentVectors_->getW()->square2(*latentVectorsSquare_); + tmpOut_->mul(*inputSquare_, *latentVectorsSquare_); + outV->sumRows(*tmpOut_, -0.5, 1.0); + + /* activation */ { + REGISTER_TIMER_INFO("FmFwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void FactorizationMachineLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { backwardActivation(); } + + const MatrixPtr& inputV = getInputValue(0); + const MatrixPtr& oGrad = getOutputGrad(); + + Matrix::resizeOrCreate( + tmpSum_, 1, latentVectors_->getW()->getHeight(), false, useGpu_); + MatrixPtr tmpSumTrans = Matrix::create(tmpSum_->getRowBuf(0), + latentVectors_->getW()->getHeight(), + 1, + false, + useGpu_); + + /* Calculate the gradients of the latentVectors_ matrix */ + if (latentVectors_->getWGrad()) { + if (dynamic_cast(inputV.get())) { + Matrix::resizeOrCreateSparseMatrix(tmpInput_, + inputV->getHeight(), + inputV->getWidth(), + inputV->getElementCnt()); + + CpuSparseMatrix* sparseInputV = + dynamic_cast(inputV.get()); + CpuSparseMatrix* sparseInputSquare = + dynamic_cast(inputSquare_.get()); + CpuSparseMatrix* sparseTmpInput = + dynamic_cast(tmpInput_.get()); + sparseTmpInput->copyFrom(*sparseInputV); + + sparseTmpInput->rowScale(0, *sparseInputV, *oGrad); + latentVectors_->getWGrad()->mul( + *sparseTmpInput->getTranspose(), *inputMulFactor_, 1, 1); + sparseTmpInput->rowScale(0, *sparseInputSquare, *oGrad); + + Matrix::resizeOrCreate(negOnes_, 1, inputV->getHeight(), false, useGpu_); + negOnes_->zeroMem(); + negOnes_->add(-1); + tmpSum_->mul(*negOnes_, *sparseTmpInput, 1, 0); + } else { + Matrix::resizeOrCreate( + tmpInput_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); + + tmpInput_->rowScale(0, *inputV, *oGrad); + latentVectors_->getWGrad()->mul( + *tmpInput_->getTranspose(), *inputMulFactor_, 1, 1); + tmpInput_->rowScale(0, *inputSquare_, *oGrad); + + tmpSum_->sumCols(*tmpInput_, -1, 0); + } + + latentVectors_->getWGrad()->addRowScale( + 0, *latentVectors_->getW(), *tmpSumTrans); + + /* Increasing the number of gradient */ + latentVectors_->getParameterPtr()->incUpdate(callback); + } + + /* Calculate the input layers gradient */ + MatrixPtr inGrad = getInputGrad(0); + if (inGrad != NULL) { + inGrad->mul( + *inputMulFactor_, *latentVectors_->getW()->getTranspose(), 1, 1); + tmpSumTrans->sumRows(*latentVectorsSquare_, -1, 0); + inGrad->addColScale(0, *inputV, *tmpSum_); + inGrad->rowScale(0, *inGrad, *oGrad); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/FactorizationMachineLayer.h b/paddle/legacy/gserver/layers/FactorizationMachineLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..fc015ed727bbd8781bb50a22b8e745d8896837e1 --- /dev/null +++ b/paddle/legacy/gserver/layers/FactorizationMachineLayer.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { +/** + * @brief The Factorization Machine models pairwise (order-2) feature + * interactions as inner product of the learned latent vectors corresponding + * to each input feature. + * + * The Factorization Machine can effectively capture feature interactions + * especially when the input is sparse. While in principle FM can model higher + * order feature interaction, in practice usually only order-2 feature + * interactions are considered. The Factorization Machine Layer here only + * computes the order-2 interations with the formula: + * + * \f[ + * y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j + * \f] + * + * The detailed calculation for forward and backward can be found at this paper: + * + * Factorization machines. + * + * The config file api is factorization_machine. + */ + +class FactorizationMachineLayer : public Layer { + protected: + // The latent vectors, shape: (size, factorSize_) + // Each row of the latentVectors_ matrix is the latent vector + // corresponding to one input feature dimension + std::unique_ptr latentVectors_; + // The hyperparameter that defines the dimensionality of the factorization + size_t factorSize_; + + private: + // Store the square values of the letent vectors matrix + MatrixPtr latentVectorsSquare_; + // Store the square values of input matrix + MatrixPtr inputSquare_; + // The result of input matrix * latent vector matrix that will be used in + // both forward and backward step + MatrixPtr inputMulFactor_; + // Store temporary calculation result + MatrixPtr tmpOut_; + MatrixPtr tmpSum_; + MatrixPtr tmpInput_; + // Negative identity matrix + MatrixPtr negOnes_; + + public: + explicit FactorizationMachineLayer(const LayerConfig& config) + : Layer(config) {} + ~FactorizationMachineLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/FeatureMapExpandLayer.cpp b/paddle/legacy/gserver/layers/FeatureMapExpandLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a3fe1433e4b5fd7bd77f8d6bb73378243d391dd5 --- /dev/null +++ b/paddle/legacy/gserver/layers/FeatureMapExpandLayer.cpp @@ -0,0 +1,155 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief A layer for expanding a batch of images to feature maps. + * Each data of the input is a 2 dimensional matrix. Each element of the matrix + * is replicated num_filters times to create a feature map with num_filters + * channels. + * - Input: Input one should be dense image data. + * - Output: expanded fature maps. + * \f[ + * y.row[i] = x.row[i \mod x.width], i = 0,1,..., (x.width * num\_filters - 1) + * \f] + * For example, num_filters = 4: + * @code + * x = [a1,a2; + * b1,b2] + * y = [a1, a2, a1, a2, a1, a2, a1, a2; + * b1, b2, b1, b2, b1, b2, b1, b2;] + * @endcode + */ + +class FeatureMapExpandLayer : public Layer { + private: + int numFilters_; + bool asRowVector_; + + public: + explicit FeatureMapExpandLayer(const LayerConfig& config) : Layer(config) {} + + ~FeatureMapExpandLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(featmap_expand, FeatureMapExpandLayer); + +bool FeatureMapExpandLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 1UL); + numFilters_ = config_.num_filters(); + asRowVector_ = config_.user_arg() != "as_col_vec"; + return true; +} + +void FeatureMapExpandLayer::forward(PassType passType) { + Layer::forward(passType); + MatrixPtr inputV = getInputValue(0); + size_t batchSize = getInput(0).getBatchSize(); + int imgSize = inputV->getWidth(); + resetOutput(batchSize, imgSize * numFilters_); + + MatrixPtr outputV = getOutputValue(); + + { + AsyncGpuBlock asyncGpuBlock; + if (asRowVector_) { + for (size_t i = 0; i < batchSize; i++) { + MatrixPtr outVTmp = + Matrix::create(outputV->getData() + i * imgSize * numFilters_, + numFilters_, + imgSize, + false, + useGpu_); + MatrixPtr inVTmp = Matrix::create( + inputV->getData() + i * imgSize, 1, imgSize, false, useGpu_); + outVTmp->addRowVector(*inVTmp); + } + } else { + for (size_t i = 0; i < batchSize; i++) { + MatrixPtr outVTmp = + Matrix::create(outputV->getData() + i * imgSize * numFilters_, + imgSize, + numFilters_, + false, + useGpu_); + MatrixPtr inVTmp = Matrix::create( + inputV->getData() + i * imgSize, imgSize, 1, false, useGpu_); + outVTmp->addColVector(*inVTmp); + } + } + } + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void FeatureMapExpandLayer::backward(const UpdateCallback& callback) { + MatrixPtr inGrad = getInputGrad(0); + if (NULL == inGrad) { + return; + } + MatrixPtr outGrad = getOutputGrad(); + size_t batchSize = getInput(0).getBatchSize(); + int imgSize = inGrad->getWidth(); + /* Do activation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + { + AsyncGpuBlock asyncGpuBlock; + if (asRowVector_) { + for (size_t i = 0; i < batchSize; i++) { + MatrixPtr outGradTmp = + Matrix::create(outGrad->getData() + i * imgSize * numFilters_, + numFilters_, + imgSize, + false, + useGpu_); + MatrixPtr inGradTmp = Matrix::create( + inGrad->getData() + i * imgSize, 1, imgSize, false, useGpu_); + inGradTmp->collectBias(*outGradTmp, 1); + } + } else { + for (size_t i = 0; i < batchSize; i++) { + MatrixPtr outGradTmp = + Matrix::create(outGrad->getData() + i * imgSize * numFilters_, + imgSize, + numFilters_, + false, + useGpu_); + MatrixPtr inGradTmp = Matrix::create( + inGrad->getData() + i * imgSize, imgSize, 1, false, useGpu_); + inGradTmp->sumRows(*outGradTmp, 1, 1); + } + } + } +} + +} // namespace paddle. diff --git a/paddle/gserver/layers/FullMatrixProjection.cpp b/paddle/legacy/gserver/layers/FullMatrixProjection.cpp similarity index 100% rename from paddle/gserver/layers/FullMatrixProjection.cpp rename to paddle/legacy/gserver/layers/FullMatrixProjection.cpp diff --git a/paddle/legacy/gserver/layers/FullMatrixProjection.h b/paddle/legacy/gserver/layers/FullMatrixProjection.h new file mode 100644 index 0000000000000000000000000000000000000000..c33d02a3aeac8e83f613e61320cb6cd63baeae83 --- /dev/null +++ b/paddle/legacy/gserver/layers/FullMatrixProjection.h @@ -0,0 +1,42 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/legacy/utils/Stat.h" + +#include "Projection.h" + +namespace paddle { + +/** + * FullMatrixProjection performs full matrix multiplication: + * \f[ + * out.row[i] += in.row[i] * weight + * \f] + * + * The config file api is full_matrix_projection. + */ +class FullMatrixProjection : public Projection { + public: + FullMatrixProjection(const ProjectionConfig& config, + const ParameterPtr& parameter, + bool useGpu); + virtual void forward(); + virtual void backward(const UpdateCallback& callback); + + protected: + std::unique_ptr weight_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/FullyConnectedLayer.cpp b/paddle/legacy/gserver/layers/FullyConnectedLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..07f4dfbe39c6b9bc233b3c75b4b5891a1ec9b2ec --- /dev/null +++ b/paddle/legacy/gserver/layers/FullyConnectedLayer.cpp @@ -0,0 +1,150 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "FullyConnectedLayer.h" +#include +#include +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(fc, FullyConnectedLayer); + +bool FullyConnectedLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* initialize the weightList */ + CHECK(inputLayers_.size() == parameters_.size()); + for (size_t i = 0; i < inputLayers_.size(); i++) { + // Option the parameters + size_t height = inputLayers_[i]->getSize(); + size_t width = getSize(); + + // create a new weight + if (parameters_[i]->isSparse()) { + CHECK_LE(parameters_[i]->getSize(), width * height); + } else { + CHECK_EQ(parameters_[i]->getSize(), width * height); + } + Weight* w = new Weight(height, width, parameters_[i]); + + // append the new weight to the list + weights_.emplace_back(w); + } + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + + return true; +} + +void FullyConnectedLayer::prefetch() { + for (size_t i = 0; i != inputLayers_.size(); ++i) { + auto* sparseParam = + dynamic_cast(weights_[i]->getW().get()); + if (sparseParam) { + MatrixPtr input = getInputValue(i); + sparseParam->addRows(input); + } + } +} + +void FullyConnectedLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInput(0).getBatchSize(); + int size = getSize(); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, size); + } + + MatrixPtr outV = getOutputValue(); + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + auto input = getInput(i); + CHECK(input.value) << "The input of 'fc' layer must be matrix"; + REGISTER_TIMER_INFO("FwMulTimer", getName().c_str()); + i == 0 ? outV->mul(*input.value, *weights_[i]->getW(), 1, 0) + : outV->mul(*input.value, *weights_[i]->getW(), 1, 1); + } + + /* add the bias-vector */ + if (biases_.get() != NULL) { + REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); + outV->addBias(*(biases_->getW()), 1); + } + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void FullyConnectedLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + if (biases_ && biases_->getWGrad()) { + REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + bool syncFlag = hl_get_sync_flag(); + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + /* Calculate the W-gradient for the current layer */ + if (weights_[i]->getWGrad()) { + MatrixPtr input_T = getInputValue(i)->getTranspose(); + MatrixPtr oGrad = getOutputGrad(); + { + REGISTER_TIMER_INFO("GradMulTimer", getName().c_str()); + weights_[i]->getWGrad()->mul(*input_T, *oGrad, 1, 1); + } + } + + // If callback does not change value, backprop error asynchronously so that + // we can do the callback concurrently. + hl_set_sync_flag(false); + + /* Calculate the input layers error */ + MatrixPtr preGrad = getInputGrad(i); + if (NULL != preGrad) { + MatrixPtr weights_T = weights_[i]->getW()->getTranspose(); + REGISTER_TIMER_INFO("BpMulTimer", getName().c_str()); + preGrad->mul(*getOutputGrad(), *weights_T, 1, 1); + } + + hl_set_sync_flag(syncFlag); + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + weights_[i]->getParameterPtr()->incUpdate(callback); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/FullyConnectedLayer.h b/paddle/legacy/gserver/layers/FullyConnectedLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..7e29cac0437a8ae735ffb71e5ee901edd79fa7f3 --- /dev/null +++ b/paddle/legacy/gserver/layers/FullyConnectedLayer.h @@ -0,0 +1,49 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { +/** + * A layer has full connections to all neurons in the previous layer. + * It computes an inner product with a set of learned weights, and + * (optionally) adds biases. + * + * The config file api is fc_layer. + */ + +class FullyConnectedLayer : public Layer { + protected: + WeightList weights_; + std::unique_ptr biases_; + + public: + explicit FullyConnectedLayer(const LayerConfig& config) : Layer(config) {} + ~FullyConnectedLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + Weight& getWeight(int idx) { return *weights_[idx]; } + + void prefetch() override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/GatedRecurrentLayer.cpp b/paddle/legacy/gserver/layers/GatedRecurrentLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..bdcd445cb47de346a8ca496fdaecf7d1f841f51e --- /dev/null +++ b/paddle/legacy/gserver/layers/GatedRecurrentLayer.cpp @@ -0,0 +1,414 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GatedRecurrentLayer.h" +#include "Layer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(gated_recurrent, GatedRecurrentLayer); + +bool GatedRecurrentLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!Layer::init(layerMap, parameterMap)) return false; + CHECK_EQ(1U, inputLayers_.size()); + CHECK_EQ(1U, parameters_.size()); + CHECK_EQ(getSize() * getSize() * 3, parameters_[0]->getSize()); + CHECK_EQ(getSize() * 3, biasParameter_->getSize()); + weight_.reset(new Weight(getSize(), getSize() * 3, parameters_[0])); + gateWeight_.reset(new Weight(getSize(), getSize() * 2, parameters_[0], 0)); + stateWeight_.reset(new Weight( + getSize(), getSize(), parameters_[0], 2 * getSize() * getSize())); + if (biasParameter_.get() != NULL) { + bias_.reset(new Weight(1, getSize() * 3, biasParameter_)); + } + + reversed_ = config_.reversed(); + activationGate_.reset(ActivationFunction::create(config_.active_gate_type())); + + GruCompute::init(config_); + useBatch_ = true; + + return true; +} + +void GatedRecurrentLayer::resetState() { + CHECK(!reversed_) << "state is not allowed for reversed gated " + "recurrent layer"; + Matrix::resizeOrCreate( + prevOutput_, 1, getSize(), /* trans= */ false, useGpu_); + prevOutput_->zeroMem(); + + // TODO(hedaoyuan): support prev_batch_state + CHECK(!FLAGS_prev_batch_state) << "Not supported"; + + useBatch_ = false; +} + +void GatedRecurrentLayer::setState(LayerStatePtr state) { + CHECK(state->value.size() == 1) + << "one matrix is expected for GatedRecurrentLayer state"; + prevOutput_->copyFrom(*(state->value[0])); +} + +LayerStatePtr GatedRecurrentLayer::getState() { + LayerStatePtr res = std::make_shared(); + res->value.push_back(prevOutput_->clone(0, 0, useGpu_)); + res->value[0]->copyFrom(*prevOutput_); + return res; +} + +void GatedRecurrentLayer::forward(PassType passType) { + REGISTER_TIMER_INFO("GruFwTimer", getName().c_str()); + Layer::forward(passType); + + const Argument& input = getInput(0); + CHECK(input.sequenceStartPositions); + int batchSize = input.getBatchSize(); + size_t numSequences = input.getNumSequences(); + resetOutput(batchSize, getSize()); + CHECK_EQ(getSize() * 3, input.value->getWidth()); + const int* starts = input.sequenceStartPositions->getData(false); + // batchSize = length of total frames in a batch (NOT size of mini-batch) + CHECK_EQ(starts[numSequences], batchSize); + + Matrix::resizeOrCreate(gate_.value, + /* height= */ batchSize, + getSize() * 3, + /* trans= */ false, + useGpu_); + Matrix::resizeOrCreate(resetOutput_.value, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + + if (useBatch_) { + forwardBatch(batchSize, numSequences, starts, input.value); + } else { + forwardSequence(batchSize, numSequences, starts, input.value); + } +} + +void GatedRecurrentLayer::backward(const UpdateCallback& callback) { + REGISTER_TIMER_INFO("GruBwTimer", getName().c_str()); + const Argument& input = getInput(0); + CHECK(input.sequenceStartPositions); + int batchSize = input.getBatchSize(); + const int* starts = input.sequenceStartPositions->getData(false); + size_t numSequences = input.getNumSequences(); + + Matrix::resizeOrCreate(gate_.grad, + /* height= */ batchSize, + getSize() * 3, + /* trans= */ false, + useGpu_); + Matrix::resizeOrCreate(resetOutput_.grad, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + + if (useBatch_) { + backwardBatch(batchSize, input.grad); + } else { + backwardSequence(batchSize, numSequences, starts, input.grad); + } + + if (bias_) { + bias_->getParameterPtr()->incUpdate(callback); + } + + weight_->getParameterPtr()->incUpdate(callback); +} + +void GatedRecurrentLayer::forwardSequence(int batchSize, + size_t numSequences, + const int* starts, + MatrixPtr inputValue) { + REGISTER_TIMER_INFO("GruFwSequenceTime", getName().c_str()); + gate_.value->assign(*inputValue); + if (bias_) { + gate_.value->addBias(*(bias_->getW()), 1); + } + + hl_gru_value gruValue; + gruValue.gateWeight = (gateWeight_->getW())->getData(); + gruValue.stateWeight = (stateWeight_->getW())->getData(); + gruValue.gateValue = gate_.value->getData(); + gruValue.resetOutputValue = resetOutput_.value->getData(); + gruValue.outputValue = output_.value->getData(); + gruValue.prevOutValue = nullptr; + + if (reversed_) { + gruValue.gateValue += (batchSize - 1) * getSize() * 3; + gruValue.resetOutputValue += (batchSize - 1) * getSize(); + gruValue.outputValue += (batchSize - 1) * getSize(); + } + + auto nextFrame = [&gruValue](bool reversed, int frameSize) { + gruValue.prevOutValue = gruValue.outputValue; + if (!reversed) { + gruValue.gateValue += frameSize * 3; + gruValue.resetOutputValue += frameSize; + gruValue.outputValue += frameSize; + } else { + gruValue.gateValue -= frameSize * 3; + gruValue.resetOutputValue -= frameSize; + gruValue.outputValue -= frameSize; + } + }; + + if (!reversed_) { + if (prevOutput_) { + gruValue.prevOutValue = prevOutput_->getData(); + } + } + AsyncGpuBlock asyncGpuBlock; + for (size_t n = 0; n < numSequences; ++n) { + int length; + if (!reversed_) { + length = starts[n + 1] - starts[n]; + } else { + length = starts[numSequences - n] - starts[numSequences - n - 1]; + } + for (int l = 0; l < length; ++l) { + if (useGpu_) { + GruCompute::forward<1>(gruValue, getSize()); + } else { + GruCompute::forward<0>(gruValue, getSize()); + } + + nextFrame(reversed_, getSize()); + } + if (!reversed_) { + if (!prevOutput_) gruValue.prevOutValue = nullptr; + } else { + gruValue.prevOutValue = nullptr; + } + } + + if (!reversed_) { + if (prevOutput_) { + prevOutput_->assign(*output_.value->subMatrix(batchSize - 1, 1)); + } + } +} + +void GatedRecurrentLayer::backwardSequence(int batchSize, + size_t numSequences, + const int* starts, + MatrixPtr inputGrad) { + REGISTER_TIMER_INFO("GruBwSequenceTime", getName().c_str()); + + hl_gru_value gruValue; + gruValue.gateWeight = (gateWeight_->getW())->getData(); + gruValue.stateWeight = (stateWeight_->getW())->getData(); + gruValue.gateValue = gate_.value->getData(); + gruValue.resetOutputValue = resetOutput_.value->getData(); + gruValue.outputValue = output_.value->getData(); + + hl_gru_grad gruGrad; + gruGrad.gateWeightGrad = + (gateWeight_->getWGrad() ? gateWeight_->getWGrad()->getData() : nullptr); + gruGrad.stateWeightGrad = + (stateWeight_->getWGrad() ? stateWeight_->getWGrad()->getData() + : nullptr); + gruGrad.gateGrad = gate_.grad->getData(); + gruGrad.resetOutputGrad = resetOutput_.grad->getData(); + gruGrad.outputGrad = output_.grad->getData(); + + if (!reversed_) { + gruValue.gateValue += (batchSize - 1) * getSize() * 3; + gruValue.resetOutputValue += (batchSize - 1) * getSize(); + gruValue.outputValue += (batchSize - 1) * getSize(); + gruGrad.gateGrad += (batchSize - 1) * getSize() * 3; + gruGrad.resetOutputGrad += (batchSize - 1) * getSize(); + gruGrad.outputGrad += (batchSize - 1) * getSize(); + gruValue.prevOutValue = gruValue.outputValue - getSize(); + gruGrad.prevOutGrad = gruGrad.outputGrad - getSize(); + } else { + gruValue.prevOutValue = gruValue.outputValue + getSize(); + gruGrad.prevOutGrad = gruGrad.outputGrad + getSize(); + } + + auto nextFrame = [&gruValue, &gruGrad](bool reversed, int frameSize) { + if (reversed) { + gruValue.gateValue += frameSize * 3; + gruValue.resetOutputValue += frameSize; + gruValue.outputValue += frameSize; + gruGrad.gateGrad += frameSize * 3; + gruGrad.resetOutputGrad += frameSize; + gruGrad.outputGrad += frameSize; + gruValue.prevOutValue = gruValue.outputValue + frameSize; + gruGrad.prevOutGrad = gruGrad.outputGrad + frameSize; + } else { + gruValue.gateValue -= frameSize * 3; + gruValue.resetOutputValue -= frameSize; + gruValue.outputValue -= frameSize; + gruGrad.gateGrad -= frameSize * 3; + gruGrad.resetOutputGrad -= frameSize; + gruGrad.outputGrad -= frameSize; + gruValue.prevOutValue = gruValue.outputValue - frameSize; + gruGrad.prevOutGrad = gruGrad.outputGrad - frameSize; + } + }; + + { + AsyncGpuBlock asyncGpuBlock; + for (size_t n = 0; n < numSequences; ++n) { + int length; + if (reversed_) { + length = starts[n + 1] - starts[n]; + } else { + length = starts[numSequences - n] - starts[numSequences - n - 1]; + } + for (int l = 0; l < length; ++l) { + if (l == length - 1) { + gruValue.prevOutValue = nullptr; + gruGrad.prevOutGrad = nullptr; + } + if (useGpu_) { + GruCompute::backward<1>(gruValue, gruGrad, getSize()); + } else { + GruCompute::backward<0>(gruValue, gruGrad, getSize()); + } + nextFrame(reversed_, getSize()); + } + } + } + + if (inputGrad) { + inputGrad->add(*gate_.grad); + } + if (bias_ && bias_->getWGrad()) { + bias_->getWGrad()->collectBias(*gate_.grad, 1); + } +} + +void GatedRecurrentLayer::forwardBatch(int batchSize, + size_t numSequences, + const int* starts, + MatrixPtr inputValue) { + REGISTER_TIMER_INFO("GruFwBatchTime", getName().c_str()); + hl_gru_value gruValue; + gruValue.gateWeight = (gateWeight_->getW())->getData(); + gruValue.stateWeight = (stateWeight_->getW())->getData(); + + if (!batchValue_) { + batchValue_.reset(new SequenceToBatch(useGpu_)); + } + batchValue_->resizeOrCreateBatch(batchSize, numSequences, starts, reversed_); + + batchValue_->resizeOrCreate(*output_.value); + batchValue_->copy(*inputValue, *gate_.value, /* seq2batch */ true); + if (bias_) { + gate_.value->addBias(*(bias_->getW()), 1); + } + + { + int numBatch = batchValue_->getNumBatch(); + int curBatchSize = 0; + AsyncGpuBlock asyncGpuBlock; + for (int n = 0; n < numBatch; n++) { + MatrixPtr outputValueTmp = batchValue_->getBatchValue(n); + gruValue.outputValue = outputValueTmp->getData(); + gruValue.gateValue = + (batchValue_->getBatchValue(*gate_.value, n))->getData(); + gruValue.resetOutputValue = + (batchValue_->getBatchValue(*resetOutput_.value, n))->getData(); + + curBatchSize = outputValueTmp->getHeight(); + gruValue.prevOutValue = + (n == 0 + ? nullptr + : (batchValue_->getBatchValue(n - 1, curBatchSize))->getData()); + + { + if (useGpu_) { + GruCompute::forward<1>(gruValue, getSize(), curBatchSize); + } else { + GruCompute::forward<0>(gruValue, getSize(), curBatchSize); + } + } + } + } + { batchValue_->copyBackSeq(*output_.value); } +} + +void GatedRecurrentLayer::backwardBatch(int batchSize, MatrixPtr inputGrad) { + REGISTER_TIMER_INFO("GruBwBatchTime", getName().c_str()); + hl_gru_value gruValue; + gruValue.gateWeight = (gateWeight_->getW())->getData(); + gruValue.stateWeight = (stateWeight_->getW())->getData(); + + hl_gru_grad gruGrad; + gruGrad.gateWeightGrad = + (gateWeight_->getWGrad() ? gateWeight_->getWGrad()->getData() : nullptr); + gruGrad.stateWeightGrad = + (stateWeight_->getWGrad() ? stateWeight_->getWGrad()->getData() + : nullptr); + + if (!batchGrad_) { + batchGrad_.reset(new SequenceToBatch(useGpu_)); + } + batchGrad_->shareIndexWith(*batchValue_); + + { batchGrad_->copyFromSeq(*output_.grad); } + + { + int numBatch = batchGrad_->getNumBatch(); + int batchSize = 0; + AsyncGpuBlock asyncGpuBlock; + for (int n = (int)numBatch - 1; n >= 0; n--) { + gruValue.gateValue = + (batchGrad_->getBatchValue(*gate_.value, n))->getData(); + gruValue.resetOutputValue = + (batchGrad_->getBatchValue(*resetOutput_.value, n))->getData(); + + MatrixPtr outputGradTmp = batchGrad_->getBatchValue(n); + gruGrad.outputGrad = outputGradTmp->getData(); + gruGrad.gateGrad = (batchGrad_->getBatchValue(*gate_.grad, n))->getData(); + gruGrad.resetOutputGrad = + (batchGrad_->getBatchValue(*resetOutput_.grad, n))->getData(); + + { + batchSize = outputGradTmp->getHeight(); + gruValue.prevOutValue = + (n == 0 + ? nullptr + : (batchValue_->getBatchValue(n - 1, batchSize))->getData()); + gruGrad.prevOutGrad = + (n == 0 ? nullptr + : (batchGrad_->getBatchValue(n - 1, batchSize))->getData()); + + if (useGpu_) { + GruCompute::backward<1>(gruValue, gruGrad, getSize(), batchSize); + } else { + GruCompute::backward<0>(gruValue, gruGrad, getSize(), batchSize); + } + } + } + } + + if (inputGrad) { + batchGrad_->add(*inputGrad, *gate_.grad, /* seq2batch */ false); + } + if (bias_ && bias_->getWGrad()) { + bias_->getWGrad()->collectBias(*gate_.grad, /* scale */ 1); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/GatedRecurrentLayer.h b/paddle/legacy/gserver/layers/GatedRecurrentLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..8bbf01ce200c9922f49508b0499aa9422745f474 --- /dev/null +++ b/paddle/legacy/gserver/layers/GatedRecurrentLayer.h @@ -0,0 +1,100 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "GruCompute.h" +#include "Layer.h" +#include "SequenceToBatch.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief Please refer to "Junyoung Chung, Empirical Evaluation + * of Gated Recurrent Neural Networks on Sequence Modeling". + * + * GatedRecurrentLayer takes 1 input layer with size * 3. + * Input layer is diveded into 3 equal parts: (xz_t, xr_t, xi_t). + * parameter and biasParameter is also diveded into 3 equal parts: + * - parameter consists of (U_z, U_r, U) + * - baisParameter consists of (bias_z, bias_r, bias_o) + * + * \f[ + * update \ gate: z_t = actGate(xz_t + U_z * h_{t-1} + bias_z) \\ + * reset \ gate: r_t = actGate(xr_t + U_r * h_{t-1} + bias_r) \\ + * output \ candidate: {h}_t = actNode(xi_t + U * dot(r_t, h_{t-1}) + bias_o) \\ + * hidden \ activation: h_t = dot((1-z_t), h_{t-1}) + dot(z_t, {h}_t) \\ + * \f] + * + * @note + * - dot denotes "element-wise multiplication". + * - actNode is defined by config active_type + * - actGate is defined by config actvie_gate_type + * + * The config file is grumemory. + */ + +class GatedRecurrentLayer : public Layer, public GruCompute { + public: + explicit GatedRecurrentLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + + void backward(const UpdateCallback& callback) override; + + void resetState() override; + + void setState(LayerStatePtr state) override; + + LayerStatePtr getState() override; + + protected: + void forwardSequence(int batchSize, + size_t numSequences, + const int* starts, + MatrixPtr inputValue); + void backwardSequence(int batchSize, + size_t numSequences, + const int* starts, + MatrixPtr inputGrad); + + void forwardBatch(int batchSize, + size_t numSequences, + const int* starts, + MatrixPtr inputValue); + void backwardBatch(int batchSize, MatrixPtr inputGrad); + + protected: + std::unique_ptr weight_; + std::unique_ptr gateWeight_; + std::unique_ptr stateWeight_; + std::unique_ptr bias_; + + Argument gate_; + Argument resetOutput_; + + bool reversed_; + bool useBatch_; + std::unique_ptr batchValue_; + std::unique_ptr batchGrad_; + std::unique_ptr activationGate_; + + MatrixPtr prevOutput_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/GetOutputLayer.cpp b/paddle/legacy/gserver/layers/GetOutputLayer.cpp similarity index 100% rename from paddle/gserver/layers/GetOutputLayer.cpp rename to paddle/legacy/gserver/layers/GetOutputLayer.cpp diff --git a/paddle/legacy/gserver/layers/GruCompute.cpp b/paddle/legacy/gserver/layers/GruCompute.cpp new file mode 100644 index 0000000000000000000000000000000000000000..adad6285b7d5acd8780444ffeab6627531683cb7 --- /dev/null +++ b/paddle/legacy/gserver/layers/GruCompute.cpp @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GruCompute.h" +#include "hl_recurrent_apply.cuh" +#include "paddle/legacy/function/GruFunctor.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +void GruCompute::init(LayerConfig &config) { + activeNode_ = hlActiveType(config.active_type()); + activeGate_ = hlActiveType(config.active_gate_type()); +} + +template <> +void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) { + GruFunctor::compute(hppl::forward::gru_resetOutput(), + hppl::forward::gru_finalOutput(), + value, + frameSize, + batchSize, + activeNode_, + activeGate_); +} + +template <> +void GruCompute::backward<0>(hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize) { + GruGradFunctor::compute( + hppl::backward::gru_stateGrad(), + hppl::backward::gru_resetGrad(), + value, + grad, + frameSize, + batchSize, + activeNode_, + activeGate_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/GruCompute.cu b/paddle/legacy/gserver/layers/GruCompute.cu similarity index 100% rename from paddle/gserver/layers/GruCompute.cu rename to paddle/legacy/gserver/layers/GruCompute.cu diff --git a/paddle/legacy/gserver/layers/GruCompute.h b/paddle/legacy/gserver/layers/GruCompute.h new file mode 100644 index 0000000000000000000000000000000000000000..6feea7aca81b8618071893581a4e16d8ad38101c --- /dev/null +++ b/paddle/legacy/gserver/layers/GruCompute.h @@ -0,0 +1,41 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ModelConfig.pb.h" +#include "hl_gpu.h" +#include "paddle/legacy/utils/Common.h" + +namespace paddle { + +class GruCompute { + public: + void init(LayerConfig &config); + + template + void forward(hl_gru_value value, int frameSize, int batchSize = 1); + + template + void backward(hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize = 1); + + public: + hl_activation_mode_t activeNode_; + hl_activation_mode_t activeGate_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/GruStepLayer.cpp b/paddle/legacy/gserver/layers/GruStepLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2480e42d68b87ee406efc2b220b9ad6bf5cacbd6 --- /dev/null +++ b/paddle/legacy/gserver/layers/GruStepLayer.cpp @@ -0,0 +1,177 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GruCompute.h" +#include "Layer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief GruStepLayer is like GatedRecurrentLayer, but used in recurrent + * layer group. GruStepLayer takes 2 input layer. + * - input[0] with size * 3 and diveded into 3 equal parts: (xz_t, xr_t, xi_t). + * - input[1] with size: {prev_out}. + * + * parameter and biasParameter is also diveded into 3 equal parts: + * - parameter consists of (U_z, U_r, U) + * - baisParameter consists of (bias_z, bias_r, bias_o) + * + * \f[ + * update \ gate: z_t = actGate(xz_t + U_z * prev_out + bias_z) \\ + * reset \ gate: r_t = actGate(xr_t + U_r * prev_out + bias_r) \\ + * output \ candidate: {h}_t = actNode(xi_t + U * dot(r_t, prev_out) + bias_o) + * \\ + * output: h_t = dot((1-z_t), prev_out) + dot(z_t, prev_out) + * \f] + * + * @note + * - dot denotes "element-wise multiplication". + * - actNode is defined by config active_type + * - actGate is defined by config actvie_gate_type + * + * The config file api if gru_step_layer. + */ +class GruStepLayer : public Layer, public GruCompute { + protected: + Argument gate_; + Argument resetOutput_; + std::unique_ptr weight_; + std::unique_ptr bias_; + + public: + explicit GruStepLayer(const LayerConfig& config) : Layer(config) {} + + ~GruStepLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(gru_step, GruStepLayer); + +bool GruStepLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!Layer::init(layerMap, parameterMap)) return false; + CHECK_EQ(2U, inputLayers_.size()); + + CHECK_EQ(getSize() * getSize() * 3, parameters_[0]->getSize()); + weight_.reset(new Weight(getSize(), getSize() * 3, parameters_[0])); + + if (biasParameter_.get() != NULL) { + CHECK_EQ(getSize() * 3, biasParameter_->getSize()); + bias_.reset(new Weight(1, getSize() * 3, biasParameter_)); + } + + GruCompute::init(config_); + return true; +} + +void GruStepLayer::forward(PassType passType) { + REGISTER_TIMER_INFO("GruStepFwTime", getName().c_str()); + Layer::forward(passType); + + const Argument& input = getInput(0); + const Argument& prevOutput = getInput(1); + CHECK_EQ(getSize() * 3, input.value->getWidth()); + CHECK_EQ(getSize(), prevOutput.value->getWidth()); + + int batchSize = input.getBatchSize(); + resetOutput(batchSize, getSize()); + resetSpecifyOutput(gate_, + batchSize, + getSize() * 3, + /* isValueClean */ false, + /* isGradClean */ false); + resetSpecifyOutput(resetOutput_, + batchSize, + getSize(), + /* isValueClean */ false, + /* isGradClean */ false); + gate_.value->assign(*input.value); + if (bias_) { + gate_.value->addBias(*(bias_->getW()), 1); + } + + hl_gru_value gruValue; + gruValue.gateWeight = weight_->getW()->getData(); + gruValue.stateWeight = weight_->getW()->getData() + getSize() * getSize() * 2; + gruValue.gateValue = gate_.value->getData(); + gruValue.resetOutputValue = resetOutput_.value->getData(); + gruValue.outputValue = output_.value->getData(); + gruValue.prevOutValue = prevOutput.value->getData(); + + if (useGpu_) { + GruCompute::forward<1>(gruValue, getSize(), batchSize); + } else { + GruCompute::forward<0>(gruValue, getSize(), batchSize); + } +} + +void GruStepLayer::backward(const UpdateCallback& callback) { + REGISTER_TIMER_INFO("GruStepBwTime", getName().c_str()); + + const Argument& input = getInput(0); + const Argument& prevOutput = getInput(1); + int batchSize = input.getBatchSize(); + + hl_gru_value gruValue; + gruValue.gateWeight = weight_->getW()->getData(); + gruValue.stateWeight = weight_->getW()->getData() + getSize() * getSize() * 2; + gruValue.gateValue = gate_.value->getData(); + gruValue.resetOutputValue = resetOutput_.value->getData(); + gruValue.outputValue = output_.value->getData(); + gruValue.prevOutValue = prevOutput.value->getData(); + + hl_gru_grad gruGrad; + gruGrad.gateWeightGrad = + (weight_->getWGrad() ? weight_->getWGrad()->getData() : nullptr); + gruGrad.stateWeightGrad = + (weight_->getWGrad() + ? weight_->getWGrad()->getData() + getSize() * getSize() * 2 + : nullptr); + + gruGrad.gateGrad = gate_.grad->getData(); + gruGrad.resetOutputGrad = resetOutput_.grad->getData(); + gruGrad.outputGrad = output_.grad->getData(); + if (prevOutput.grad) { + gruGrad.prevOutGrad = prevOutput.grad->getData(); + } else { + gruGrad.prevOutGrad = nullptr; + } + + if (useGpu_) { + GruCompute::backward<1>(gruValue, gruGrad, getSize(), batchSize); + } else { + GruCompute::backward<0>(gruValue, gruGrad, getSize(), batchSize); + } + + if (input.grad) { + input.grad->add(*gate_.grad); + } + + if (bias_ && bias_->getWGrad()) { + bias_->getWGrad()->collectBias(*gate_.grad, 1); + } + + if (bias_) { + bias_->getParameterPtr()->incUpdate(callback); + } + weight_->getParameterPtr()->incUpdate(callback); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/HierarchicalSigmoidLayer.cpp b/paddle/legacy/gserver/layers/HierarchicalSigmoidLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..34495994096a87640bdeef777feb5cd783cd4598 --- /dev/null +++ b/paddle/legacy/gserver/layers/HierarchicalSigmoidLayer.cpp @@ -0,0 +1,240 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "HierarchicalSigmoidLayer.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +REGISTER_LAYER(hsigmoid, HierarchicalSigmoidLayer); + +bool HierarchicalSigmoidLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK(config_.has_num_classes()) << "num_classes must be specifed in config"; + numClasses_ = config_.num_classes(); + CHECK_GE(numClasses_, (size_t)2); + codeLength_ = findLastSet(numClasses_ - 1); + + size_t height = numClasses_ - 1; + + /* initialize the weightList */ + // The last input layer is for label + CHECK(!parameters_.back()); + for (size_t i = 0; i < inputLayers_.size() - 1; i++) { + size_t width = inputLayers_[i]->getSize(); + // create a new weight + CHECK_EQ(parameters_[i]->getSize(), width * height); + Weight* w = new Weight(height, width, parameters_[i]); + + // append the new weight to the list + weights_.emplace_back(w); + } + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + CHECK_EQ(biasParameter_->getSize(), numClasses_ - 1); + biases_.reset(new Weight(1, numClasses_ - 1, biasParameter_)); + } + + return true; +} + +void HierarchicalSigmoidLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(0)->getHeight(); + int size = getSize(); + reserveOutput(batchSize, size); + Matrix::resizeOrCreate(preOutput_.value, + batchSize, + codeLength_, + /* trans */ false, + false); + Matrix::resizeOrCreate(preOutput_.grad, + batchSize, + codeLength_, + /* trans */ false, + false); + IVectorPtr label = getInput(*getLabelLayer()).ids; + preOutput_.value->zeroMem(); + + if (useGpu_) { + Matrix::resizeOrCreate(cpuOutput_, + output_.value->getHeight(), + output_.value->getWidth(), + /* trans */ false, + false); + IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); + cpuLabel_->copyFrom(*label); + cpuOutput_->copyFrom(*output_.value); + } else { + cpuOutput_ = output_.value; + cpuLabel_ = label; + } + /* add the bias-vector */ + if (biases_.get() != NULL) { + if (useGpu_) { + Matrix::resizeOrCreate(cpuBias_, + 1, + numClasses_ - 1, + /* trans */ false, + false); + cpuBias_->copyFrom(*biases_->getW()); + } else { + cpuBias_ = biases_->getW(); + } + preOutput_.value->addByBitCode(numClasses_, *cpuLabel_, *cpuBias_); + } + for (size_t i = 0; i < inputLayers_.size() - 1; ++i) { + MatrixPtr input = getInputValue(i); + if (useGpu_) { + Matrix::resizeOrCreate(cpuInput_, + input->getHeight(), + input->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeight_, + weights_[i]->getW()->getHeight(), + weights_[i]->getW()->getWidth(), + /* trans */ false, + false); + cpuInput_->copyFrom(*input); + cpuWeight_->copyFrom(*weights_[i]->getW()); + } else { + cpuInput_ = input; + cpuWeight_ = weights_[i]->getW(); + } + preOutput_.value->mulByBitCode( + numClasses_, *cpuLabel_, *cpuWeight_, *cpuInput_); + } + // keep consistent with the clipping in the following softrelu + preOutput_.value->clip(-40.0, 40.0); + preOutput_.value->sumByBitCode(numClasses_, + *cpuLabel_, + *cpuOutput_, + -1); // scaleSum + preOutput_.value->softrelu(*preOutput_.value); + MatrixPtr sum = Matrix::create(batchSize, 1, /* trans= */ false, false); + preOutput_.value->rowSum(*sum); + cpuOutput_->add(*sum); + if (useGpu_) { + output_.value->copyFrom(*cpuOutput_); + } else { + output_.value = cpuOutput_; + } +} + +void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { + IVectorPtr label = getInput(*getLabelLayer()).ids; + if (useGpu_) { + IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); + cpuLabel_->copyFrom(*label); + } else { + cpuLabel_ = label; + } + preOutput_.grad->one(); + preOutput_.grad->softreluDerivative(*preOutput_.value); + preOutput_.grad->subByBitCode(numClasses_, *cpuLabel_); + + if (biases_ && biases_->getWGrad()) { + MatrixPtr biases_grad = biases_->getWGrad(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuBias_, + 1, + numClasses_ - 1, + /* trans */ false, + false); + cpuBias_->copyFrom(*biases_grad); + } else { + cpuBias_ = biases_grad; + } + preOutput_.grad->addByBitCodeBackward(numClasses_, *cpuLabel_, *cpuBias_); + if (useGpu_) { + biases_grad->copyFrom(*cpuBias_); + } else { + biases_grad = cpuBias_; + } + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + for (size_t i = 0; i < inputLayers_.size() - 1; ++i) { + /* Calculate the W-gradient for the current layer */ + MatrixPtr input = getInputValue(i); + if (weights_[i]->getWGrad()) { + MatrixPtr weights_grad = weights_[i]->getWGrad(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuInput_, + input->getHeight(), + input->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeightGrad_, + weights_grad->getHeight(), + weights_grad->getWidth(), + /* trans */ false, + false); + cpuInput_->copyFrom(*input); + cpuWeightGrad_->copyFrom(*weights_grad); + } else { + cpuInput_ = input; + cpuWeightGrad_ = weights_grad; + } + preOutput_.grad->mulByBitCodeBackwardWeight( + numClasses_, *cpuLabel_, *cpuWeightGrad_, *cpuInput_); + if (useGpu_) { + weights_grad->copyFrom(*cpuWeightGrad_); + } else { + weights_grad = cpuWeightGrad_; + } + /* Increasing the number of gradient */ + weights_[i]->getParameterPtr()->incUpdate(callback); + } + + /* Calculate the input layers error */ + MatrixPtr inputGrad = getInputGrad(i); + if (inputGrad) { + if (useGpu_) { + Matrix::resizeOrCreate(cpuInputGrad_, + inputGrad->getHeight(), + inputGrad->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeight_, + weights_[i]->getW()->getHeight(), + weights_[i]->getW()->getWidth(), + /* trans */ false, + false); + cpuInputGrad_->copyFrom(*inputGrad); + cpuWeight_->copyFrom(*weights_[i]->getW()); + } else { + cpuInputGrad_ = inputGrad; + cpuWeight_ = weights_[i]->getW(); + } + preOutput_.grad->mulByBitCodeBackwardError( + numClasses_, *cpuLabel_, *cpuWeight_, *cpuInputGrad_); + if (useGpu_) { + inputGrad->copyFrom(*cpuInputGrad_); + } else { + inputGrad = cpuInputGrad_; + } + } + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.h b/paddle/legacy/gserver/layers/HierarchicalSigmoidLayer.h similarity index 100% rename from paddle/gserver/layers/HierarchicalSigmoidLayer.h rename to paddle/legacy/gserver/layers/HierarchicalSigmoidLayer.h diff --git a/paddle/legacy/gserver/layers/IdentityProjection.cpp b/paddle/legacy/gserver/layers/IdentityProjection.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f707642e09b86721a88142ab8b745bb3492e820c --- /dev/null +++ b/paddle/legacy/gserver/layers/IdentityProjection.cpp @@ -0,0 +1,103 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Projection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * IdentityProjection performs addition: + * \f[ + * out.row[i] += in.row[i] + * \f] + * + * The config file api is identity_projection. + */ +class IdentityProjection : public Projection { + public: + IdentityProjection(const ProjectionConfig& config, + const ParameterPtr& parameter, + bool useGpu); + virtual void forward(); + virtual void backward(const UpdateCallback& callback); +}; + +REGISTER_PROJECTION(identity, IdentityProjection); + +/** + * Constructed function. + * @note IdentityProjection should not have any parameter. + */ +IdentityProjection::IdentityProjection(const ProjectionConfig& config, + const ParameterPtr& parameter, + bool useGpu) + : Projection(config, parameter, useGpu) { + CHECK(!parameter) << "'identity' projection should not have any parameter"; +} + +void IdentityProjection::forward() { out_->value->add(*in_->value); } + +void IdentityProjection::backward(const UpdateCallback& callback) { + if (in_->grad) { + in_->grad->add(*out_->grad); + } +} + +/** + * IdentityOffsetProjection likes IdentityProjection, but layer size may be + * smaller + * than input size. It selects dimensions [offset, offset+layer_size) from input + * to + * perform addition: + * \f[ + * out.row[i] += in.row[i + \textrm{offset}] + * \f] + * + * The config file api is identity_projection. + */ +class IdentityOffsetProjection : public Projection { + public: + IdentityOffsetProjection(const ProjectionConfig& config, + const ParameterPtr& parameter, + bool useGpu); + virtual void forward(); + virtual void backward(const UpdateCallback& callback); +}; + +REGISTER_PROJECTION(identity_offset, IdentityOffsetProjection); + +/** + * Constructed function. + * @note IdentityOffsetProjection should not have any parameter. + */ +IdentityOffsetProjection::IdentityOffsetProjection( + const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu) + : Projection(config, parameter, useGpu) { + CHECK(!parameter) << "'identity_offset' projection " + "should not have any parameter"; + CHECK_LE(config.output_size() + config.offset(), config.input_size()); +} + +void IdentityOffsetProjection::forward() { + out_->value->addAtOffset(*in_->value, config_.offset()); +} + +void IdentityOffsetProjection::backward(const UpdateCallback& callback) { + if (in_->grad) { + in_->grad->addAtOffset(*out_->grad, config_.offset()); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/InterpolationLayer.cpp b/paddle/legacy/gserver/layers/InterpolationLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ed2294e8a397edfee6ad3c1f52235970d6ad48a9 --- /dev/null +++ b/paddle/legacy/gserver/layers/InterpolationLayer.cpp @@ -0,0 +1,130 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for linear interpolation with two inputs, + * which is used in NEURAL TURING MACHINE. + * \f[ + * y.row[i] = w[i] * x_1.row[i] + (1 - w[i]) * x_2.row[i] + * \f] + * where \f$x_1\f$ and \f$x_2\f$ are two (batchSize x dataDim) inputs, + * \f$w\f$ is (batchSize x 1) weight vector, + * and \f$y\f$ is (batchSize x dataDim) output. + * + * The config file api is interpolation_layer. + */ + +class InterpolationLayer : public Layer { + protected: + /// weightLast = 1 - weight + MatrixPtr weightLast_; + MatrixPtr tmpMatrix; + + public: + explicit InterpolationLayer(const LayerConfig& config) : Layer(config) {} + + ~InterpolationLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(interpolation, InterpolationLayer); + +bool InterpolationLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(3U, inputLayers_.size()); + + return true; +} + +void InterpolationLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr weightV = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr inV2 = getInputValue(2); + + size_t batchSize = inV1->getHeight(); + size_t dataDim = inV1->getWidth(); + + CHECK_EQ(dataDim, getSize()); + CHECK_EQ(dataDim, inV2->getWidth()); + CHECK_EQ(batchSize, inV1->getHeight()); + CHECK_EQ(batchSize, inV2->getHeight()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + resetOutput(batchSize, dataDim); + } + + MatrixPtr outV = getOutputValue(); + + Matrix::resizeOrCreate(weightLast_, batchSize, 1, false, useGpu_); + weightLast_->one(); + weightLast_->sub(*weightV); + + REGISTER_TIMER_INFO("FwInterpTimer", getName().c_str()); + // outV = inV1 * weight + inV2 * weightLast + outV->addRowScale(0, *inV1, *weightV); + outV->addRowScale(0, *inV2, *weightLast_); +} + +void InterpolationLayer::backward(const UpdateCallback& callback) { + MatrixPtr outG = getOutputGrad(); + MatrixPtr weightV = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr inV2 = getInputValue(2); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + MatrixPtr inG2 = getInputGrad(2); + + size_t batchSize = inV1->getHeight(); + size_t dataDim = inV1->getWidth(); + + REGISTER_TIMER_INFO("BwInterpTimer", getName().c_str()); + + if (inG0) { + Matrix::resizeOrCreate(tmpMatrix, batchSize, dataDim, false, useGpu_); + + // inG0 += outG .* (inV1 - inV2) + tmpMatrix->sub(*inV1, *inV2); + inG0->rowDotMul(0, *outG, *tmpMatrix); + } + + if (inG1) { + // inG1 += outG * weight + inG1->addRowScale(0, *outG, *weightV); + } + + if (inG2) { + // inG2 += outG * weightLast + inG2->addRowScale(0, *outG, *weightLast_); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp b/paddle/legacy/gserver/layers/KmaxSeqScoreLayer.cpp similarity index 100% rename from paddle/gserver/layers/KmaxSeqScoreLayer.cpp rename to paddle/legacy/gserver/layers/KmaxSeqScoreLayer.cpp diff --git a/paddle/legacy/gserver/layers/L2DistanceLayer.cpp b/paddle/legacy/gserver/layers/L2DistanceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a3e627e57047b790b4f74089a352f06b55e48664 --- /dev/null +++ b/paddle/legacy/gserver/layers/L2DistanceLayer.cpp @@ -0,0 +1,91 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "L2DistanceLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(l2_distance, L2DistanceLayer); + +bool L2DistanceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2UL) << "The L2DistanceLayer accepts two and " + << "only two inputs."; + CHECK_EQ(getSize(), 1UL) << "The output dimensionality of L2DistanceLayer " + << "is fixed to be 1."; + + return true; +} + +void L2DistanceLayer::forward(PassType passType) { + Layer::forward(passType); + + const auto inV1 = getInputValue(0); + const auto inV2 = getInputValue(1); + + CHECK(inV1 && inV2); + CHECK_EQ(inV1->getHeight(), inV2->getHeight()) + << "The height of two inputs of this layer must be the same."; + CHECK_EQ(inV1->getWidth(), inV2->getWidth()) + << "The width of two inputs of this layer must be the same."; + + int batchSize = inV1->getHeight(); + int output_dim = getSize(); + { + REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str()); + reserveOutput(batchSize, output_dim); + auto outV = getOutputValue(); + CHECK(outV) << "The output matrix should not be null."; + + Matrix::resizeOrCreate( + inputSub_, inV1->getHeight(), inV1->getWidth(), false, useGpu_); + + inputSub_->assign(*inV1); + inputSub_->sub(*inV2); + outV->sumOfProducts(*inputSub_, *inputSub_, 1, 0); + outV->sqrt2(*outV); + } +} + +void L2DistanceLayer::backward(const UpdateCallback& callback) { + const auto outG = getOutputGrad(); + const auto outV = getOutputValue(); + CHECK(outG && outV); + + auto inGrad1 = getInputGrad(0); + auto inGrad2 = getInputGrad(1); + + { + REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str()); + + if (inGrad1 || inGrad2) { + outV->scalarDiv(*outV, 1.); + outV->dotMul(*outG, *outV); + } + + if (inGrad1) inGrad1->addRowScale(0, *inputSub_, *outV); + + if (inGrad2) { + inputSub_->mulScalar(-1.); + inGrad2->addRowScale(0, *inputSub_, *outV); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/L2DistanceLayer.h b/paddle/legacy/gserver/layers/L2DistanceLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..aa8aabd9ca5702e3ebdccbe7bb4f98fa087dd238 --- /dev/null +++ b/paddle/legacy/gserver/layers/L2DistanceLayer.h @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief The layer calculates the l2 distance between two input vectors. + * \f[ + * f(\bf{x}, \bf{y}) = \sqrt{\sum_{i=1}^D(x_i - y_i)} + * \f] + * + * - Input1: A vector (batchSize * dataDim) + * - Input2: A vector (batchSize * dataDim) + * - Output: A vector (batchSize * 1) + * + * The configuration api is: l2_distance_layer. + */ + +class L2DistanceLayer : public Layer { + public: + explicit L2DistanceLayer(const LayerConfig& config) : Layer(config) {} + ~L2DistanceLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + private: + // Store the result of subtracting Input2 from Input1 in forward computation, + // which will be reused in backward computation. + MatrixPtr inputSub_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/Layer.cpp b/paddle/legacy/gserver/layers/Layer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..890d33552dd31a8fd348a36d44fb0824ac9b32b5 --- /dev/null +++ b/paddle/legacy/gserver/layers/Layer.cpp @@ -0,0 +1,410 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Util.h" + +#include "CostLayer.h" +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/utils/Error.h" +#include "paddle/legacy/utils/Logging.h" + +#ifndef PADDLE_MOBILE_INFERENCE +#include "ValidationLayer.h" +#endif + +DEFINE_bool(log_error_clipping, false, "enable log error clipping or not"); + +namespace paddle { + +Layer::Layer(const LayerConfig& config, bool useGpu) + : config_(config), + useGpu_(useGpu), + deviceId_(CPU_DEVICE), + needSequenceInfo_(true) {} + +bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { + if (useGpu_ && FLAGS_parallel_nn) { + /* gpu environment is specified by device property */ + deviceId_ = config_.device(); + if (deviceId_ < 0) { + useGpu_ = false; + } + } + + output_.deviceId = deviceId_; + + for (auto& inputConfig : config_.inputs()) { + std::string inputName = inputConfig.input_layer_name(); + LayerPtr inputLayer; + CHECK(mapGet(inputName, layerMap, &inputLayer)) + << "Cannot find input layer " << inputName << " for layer " + << getName(); + this->addPrev(inputLayer); + + inputLayer->addOutputArgument(deviceId_); + + if (inputConfig.has_input_parameter_name()) { + ParameterPtr parameter; + CHECK( + mapGet(inputConfig.input_parameter_name(), parameterMap, ¶meter)) + << "Cannot find input parameter " + << inputConfig.input_parameter_name() << " for layer " << getName(); + parameter->incShared(); + CHECK_EQ(parameter->getDeviceId(), getDeviceId()); + parameters_.push_back(parameter); + } else { + parameters_.push_back(nullptr); + } + + if (inputConfig.has_input_layer_argument()) { + inputArgument_.push_back(inputConfig.input_layer_argument()); + } else { + inputArgument_.push_back(""); + } + } + + if (config_.has_bias_parameter_name()) { + CHECK(mapGet(config_.bias_parameter_name(), parameterMap, &biasParameter_)) + << "Cannot find bias parameter " << config_.bias_parameter_name() + << " for layer " << getName(); + biasParameter_->incShared(); + CHECK_EQ(biasParameter_->getDeviceId(), getDeviceId()); + } + + /* specify the activation function according to the configuration */ + std::string action_type = config_.active_type(); + activation_.reset(ActivationFunction::create(action_type)); + CHECK(activation_); + + initNeedFlags(); + markInBackward_.assign(inputLayers_.size(), false); + + return true; +} + +ClassRegistrar Layer::registrar_; + +LayerPtr Layer::create(const LayerConfig& config) { + std::string type = config.type(); + +#ifndef PADDLE_MOBILE_INFERENCE + // NOTE: As following types have illegal character '-', + // they can not use REGISTER_LAYER to registrar. + // Besides, to fit with old training models, + // they can not use '_' instead. + if (type == "multi-class-cross-entropy") + return LayerPtr(new MultiClassCrossEntropy(config)); + else if (type == "rank-cost") + return LayerPtr(new RankingCost(config)); + else if (type == "auc-validation") + return LayerPtr(new AucValidation(config)); + else if (type == "pnpair-validation") + return LayerPtr(new PnpairValidation(config)); +#endif + + return LayerPtr(registrar_.createByType(config.type(), config)); +} + +void Layer::resetSpecifyOutput(Argument& output, + size_t height, + size_t width, + bool isValueClean, + bool isGradClean) { + SetDevice device(output.deviceId); + + Matrix::resizeOrCreate( + output.value, height, width, /* trans */ false, useGpu(output.deviceId)); + if (isValueClean) { + output.value->zeroMem(); + } + + if (passType_ != PASS_TEST && needGradient()) { + Matrix::resizeOrCreate( + output.grad, height, width, /* trans */ false, useGpu(output.deviceId)); + if (isGradClean) { + output.grad->zeroMem(); + } + } +} + +void Layer::resizeOutput(size_t height, size_t width) { + resetSpecifyOutput(output_, height, width, false, false); + + for (size_t i = 0; i != outputOtherDevice_.size(); i++) { + resetSpecifyOutput(outputOtherDevice_[i], height, width, false, false); + } +} + +void Layer::reserveOutput(size_t height, size_t width) { + resetSpecifyOutput(output_, height, width, false, true); + + for (size_t i = 0; i != outputOtherDevice_.size(); i++) { + resetSpecifyOutput(outputOtherDevice_[i], height, width, false, true); + } +} + +void Layer::resetOutput(size_t height, size_t width) { + resetSpecifyOutput(output_, height, width, true, true); + + for (size_t i = 0; i != outputOtherDevice_.size(); i++) { + resetSpecifyOutput(outputOtherDevice_[i], height, width, true, true); + } +} + +void Layer::addOutputArgument(int deviceId) { + if (deviceId == deviceId_) { + output_.countIncrement(); + return; + } else { + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + if (outputOtherDevice_[i].deviceId == deviceId) { + outputOtherDevice_[i].countIncrement(); + return; + } + } + } + + Argument argu; + argu.deviceId = deviceId; + outputOtherDevice_.push_back(argu); + outputOtherDevice_.back().countIncrement(); +} + +void Layer::copyOutputToOtherDevice() { + for (size_t i = 0; i != outputOtherDevice_.size(); i++) { + SetDevice device(outputOtherDevice_[i].deviceId); + // If outputOtherDevice_[i].value is a CpuMatrix, + // the copyFrom is a synchronous interface. + // If outputOtherDevice_[i].value is a GpuMatrix, since subsequent + // calculations are all on HPPL_STREAM_DEFAULT, + // copyFrom can be an asynchronous interface. + outputOtherDevice_[i].value->copyFrom(*getOutputValue(), + HPPL_STREAM_DEFAULT); + outputOtherDevice_[i].sequenceStartPositions = + output_.sequenceStartPositions; + outputOtherDevice_[i].subSequenceStartPositions = + output_.subSequenceStartPositions; + outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; + + outputOtherDevice_[i].notifyValueReady(); + } +} + +void Layer::waitInputValue() { + for (size_t i = 0; i != inputLayers_.size(); i++) { + if (inputLayers_[i]->getDeviceId() != deviceId_) { + getInput(i).waitValueReady(); + } + } +} + +void Layer::waitAndMergeOutputGrad() { + if (!output_.grad || !outputOtherDevice_.size()) { + return; + } + + for (size_t i = 0; i != outputOtherDevice_.size(); i++) { + outputOtherDevice_[i].waitGradReady(); + } + + /* merge output grad */ + size_t i = 0; + if (!output_.getAllCount()) { + output_.grad->copyFrom(*outputOtherDevice_[0].grad, HPPL_STREAM_1); + hl_stream_synchronize(HPPL_STREAM_1); + + i++; + if (outputOtherDevice_.size() == 1) return; + } + + Matrix::resizeOrCreate(tmpGrad_, + output_.grad->getHeight(), + output_.grad->getWidth(), + /* trans */ false, + useGpu(output_.deviceId)); + + for (; i != outputOtherDevice_.size(); i++) { + tmpGrad_->copyFrom(*outputOtherDevice_[i].grad, HPPL_STREAM_1); + hl_stream_synchronize(HPPL_STREAM_1); + output_.grad->add(*tmpGrad_); + } +} + +void Layer::markAllInputGrad() { + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (!markInBackward_[i]) { + inputLayers_[i]->getOutput(deviceId_).notifyGradReady(); + } + markInBackward_[i] = false; + } +} + +void Layer::markInputGrad(int inputIndex) { + inputLayers_[inputIndex]->getOutput(deviceId_).notifyGradReady(); + markInBackward_[inputIndex] = true; +} + +void Layer::zeroGrad() { + CHECK(output_.grad.get() != NULL); + output_.grad->zeroMem(); +} + +void Layer::initNeedFlags() { + auto initFlag = [this]( + bool& flag, bool (Layer::*flagQueryFunc)() const, ParameterType type) { + flag = false; + if (biasParameter_ && biasParameter_->hasType(type)) { + flag = true; + } + if (!flag) { + for (auto& para : parameters_) { + if (para && para->hasType(type)) { + flag = true; + break; + } + } + } + if (!flag) { + for (auto& layer : inputLayers_) { + if ((layer.get()->*flagQueryFunc)()) { + flag = true; + } + } + } + }; + initFlag(needGradient_, &Layer::needGradient, PARAMETER_GRADIENT); +} + +void Layer::showOutputStats() { + MatrixPtr out = getOutputValue(); + if (!out) return; + if (!out->getElementCnt()) { + LOG(INFO) << "The number of output of " << config_.name() + << " is 0, skip to show the statistics"; + return; + } + MatrixPtr outSquare; + if (dynamic_cast(out.get())) { + GpuSparseMatrix* tmp = dynamic_cast(out.get()); + outSquare = std::make_shared(tmp->getHeight(), + tmp->getWidth(), + tmp->getElementCnt(), + tmp->getValueType(), + tmp->getFormat()); + } else { + outSquare = out->clone(); + } + outSquare->copyFrom(*out, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + + real mean = outSquare->getSum() / out->getElementCnt(); + real min; + real max; + if (dynamic_cast(outSquare.get())) { + auto tmpMat = dynamic_cast(outSquare.get()); + min = tmpMat->getMin(); + max = tmpMat->getMax(); + tmpMat->square2(); + LOG(INFO) << "show statistics of [none zero values] in sparse matrix"; + } else { + min = outSquare->getMin(); + max = outSquare->getMax(); + outSquare->square2(); + } + real std = (outSquare->getSum() / outSquare->getElementCnt()) - mean * mean; + std = std > 0 ? std : 0; + LOG(INFO) << "The output state of " << config_.name() << ": mean=" << mean + << ", " + << "std=" << std << ", " + << "min=" << min << ", " + << "max=" << max; +} + +void Layer::forwardActivation() { + /* activation */ + auto status = activation_->forward(output_); + status.check(); + + /* dropout */ + if (config_.drop_rate() > 0) { + forwardDropOut(); + CHECK_NE(activation_->getName(), "softmax") + << "Softmax activation cannot be used with Dropout"; + } + + if (FLAGS_show_layer_stat) { + showOutputStats(); + } +} + +void Layer::backwardActivation() { + /* Do error clipping */ + if (config_.error_clipping_threshold() > 0.0f) { + if (FLAGS_log_error_clipping) { + VectorPtr outGradVec = Vector::create( + output_.grad->getData(), output_.grad->getElementCnt(), useGpu_); + real maxAbsGrad = outGradVec->getAbsMax(); + if (maxAbsGrad > config_.error_clipping_threshold()) { + real avgAbsGrad = outGradVec->getAbsSum() / outGradVec->getSize(); + LOG(INFO) << " layer=" << config_.name() << " need clipping," + << " max error=" << maxAbsGrad << " avg error=" << avgAbsGrad; + } + } + output_.grad->clip(-config_.error_clipping_threshold(), + config_.error_clipping_threshold()); + } + + /* Do dropout for delta*/ + if (config_.drop_rate() > 0 && passType_ != PASS_TEST) { + MatrixPtr oGrad = getOutputGrad(); + oGrad->dotMul(*oGrad, *dropOutMask_); + } + + auto status = activation_->backward(output_); + status.check(); +} + +void Layer::forwardDropOut() { + auto& outV = getOutputValue(); + + if (passType_ == PASS_TRAIN) { + // new dropOutMask_ if dropOutMask_ is null ptr + Matrix::resizeOrCreate(dropOutMask_, + outV->getHeight(), + outV->getWidth(), + false, + useGpu(deviceId_)); + dropOutMask_->randomizeUniform(); // generate a uniform random matrix + dropOutMask_->biggerThanScalar(config_.drop_rate()); // random mask + outV->dotMul(*outV, *dropOutMask_); // dropout + } else if (passType_ == PASS_GC) { + // only initialize once + if (!dropOutMask_) { + dropOutMask_ = Matrix::create( + outV->getHeight(), outV->getWidth(), false, useGpu(deviceId_)); + // We use cpu matrix to generate mask so that the mask + // will be same for both gpu version and cpu version. + // This will help unittest to make sure they have same result. + MatrixPtr tmpMask = Matrix::create(outV->getHeight(), outV->getWidth()); + tmpMask->randomizeUniform(); // generate a uniform random matrix + tmpMask->biggerThanScalar(config_.drop_rate()); // random mask + dropOutMask_->copyFrom(*tmpMask); + } + outV->dotMul(*outV, *dropOutMask_); + } else { // passType == PASS_TEST + outV->mulScalar(1.0 - config_.drop_rate()); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/Layer.h b/paddle/legacy/gserver/layers/Layer.h new file mode 100644 index 0000000000000000000000000000000000000000..a7ff76decea9a448acfcdef1c81a68b5a823cc56 --- /dev/null +++ b/paddle/legacy/gserver/layers/Layer.h @@ -0,0 +1,512 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/function/Function.h" +#include "paddle/legacy/gserver/activations/ActivationFunction.h" +#include "paddle/legacy/math/CpuSparseMatrix.h" +#include "paddle/legacy/parameter/Argument.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/parameter/Weight.h" +#include "paddle/legacy/utils/ClassRegistrar.h" +#include "paddle/legacy/utils/Util.h" + +/// Macro for registering a layer type. +/// Example: REGISTER_LAYER(crf_error, CRFDecodingErrorLayer); +#define REGISTER_LAYER(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name( \ + []() { Layer::registrar_.registerClass<__class_name>(#__type_name); }) + +#define REGISTER_LAYER_CREATE_FUNC(__type_name, createFunction) \ + static InitFunction __reg_type_##__type_name( \ + []() { Layer::registrar_.registerClass(#__type_name, createFunction); }) + +namespace paddle { + +class Layer; +typedef std::shared_ptr LayerPtr; +typedef std::map LayerMap; +class NeuralNetwork; + +/// layer state, used for RNN and LSTM layers +struct LayerState { + std::vector value; +}; +typedef std::shared_ptr LayerStatePtr; + +/// Paddle device ID, MKLDNN is -2, CPU is -1 +enum PADDLE_DEVICE_ID { + MKLDNN_DEVICE = -2, + CPU_DEVICE = -1, +}; + +/** + * @brief Base class for layer. + * Define necessary variables and functions for every layer. + */ +class Layer { + protected: + /// Layer config + LayerConfig config_; + /// whether to use GPU + bool useGpu_; + /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ... + int deviceId_; + /// Input layers + std::vector inputLayers_; + /// Argument of input layers + std::vector inputArgument_; + + /// Parameter for each input layer. + /// Parameters_[i] is nullptr if inputLayers_[i] does not need parameter. + std::vector parameters_; + + /// nullptr if bias is not needed. + ParameterPtr biasParameter_; + + /// Output + Argument output_; + /// Several outputs stored on different devices, used in 'parallel_nn' case, + /// and record them by deviceId_. + /// Also used in 'use_mkldnn' case. + std::vector outputOtherDevice_; + /// If there are several outputs, map them by each name. + /// MKLDNNLayer use it only to merge output grad + std::map outputMap_; + /// Used to merge grad on different devices. + MatrixPtr tmpGrad_; + + std::unique_ptr activation_; + + /// Current passType, PASS_TRAIN or PASS_TEST + PassType passType_; + + /// Random 0-1 matrix for dropOut + MatrixPtr dropOutMask_; + + /// Whether the layer need to compute gradient + bool needGradient_; + /// Whether the layer need to compute re-sequence information + bool needSequenceInfo_; + + /// Mark input grad in(true) or out(false) of backward function. + std::vector markInBackward_; + + /// Layer forward function + std::vector> forward_; + /// Layer backward function + std::vector> backward_; + + public: + /** + * Wait until all input value ready. + * Called before Layer::forward() function. + */ + virtual void waitInputValue(); + + /** + * Copy layer's output_ to other device. + * If output layer is in other device, called after Layer::forward() function. + */ + virtual void copyOutputToOtherDevice(); + + /** + * Wait until all output grad ready and merge them to output_.grad. + * Called before Layer::backward() function. + */ + virtual void waitAndMergeOutputGrad(); + + /** + * Notify previous layer the output grad ready. + * Called after Layer::backward() function. + */ + virtual void markAllInputGrad(); + + protected: + /** + * Create layer function. Function is called in forward or backward. + * \param function, Layer::forward_ or Layer::backward_ + * \param name, function name + * \param config, initialization configuration for the function + */ + void createFunction(std::vector>& function, + const std::string& name, + const FuncConfig& config) { + if (useGpu_) { + function.emplace_back( + FunctionBase::funcRegistrar_.createByType(name + "-GPU")); + } else { + function.emplace_back( + FunctionBase::funcRegistrar_.createByType(name + "-CPU")); + } + auto& func = function.back(); + func->init(config); + } + + /** + * Notify specified layer the output grad ready. + * Called in the backward function. + * If do mark input grad in the backward function, you should to ensure + * that all input grad will be marked in the backward function. + */ + void markInputGrad(int inputIndex); + + /** + * Get the argument of input layer. + */ + const Argument& getInput(size_t inputIndex) const { + return inputLayers_[inputIndex]->getOutput(deviceId_); + } + + /** + * Get the argument of input layer. + */ + const Argument& getInput(const Layer& inputLayer) const { + return inputLayer.getOutput(deviceId_); + } + + /** + * Get the argument of input layer with deviceId. + */ + const Argument& getInput(size_t inputIndex, int deviceId) const { + return inputLayers_[inputIndex]->getOutput(deviceId); + } + + /** + * Get the forward-input value. + */ + const MatrixPtr& getInputValue(int inputIndex) { + return inputLayers_[inputIndex]->getOutput(deviceId_).value; + } + + /** + * Get the forward-input value. + */ + const MatrixPtr& getInputValue(const Layer& inputLayer) { + return inputLayer.getOutput(deviceId_).value; + } + + /** + * Get the forward-input value with deviceId. + */ + const MatrixPtr& getInputValue(int inputIndex, int deviceId) { + return inputLayers_[inputIndex]->getOutput(deviceId).value; + } + + /** + * Get the forward-input grad. + */ + const MatrixPtr& getInputGrad(int inputIndex) { + return inputLayers_[inputIndex]->getOutput(deviceId_).grad; + } + + /** + * Get the forward-input grad. + */ + const MatrixPtr& getInputGrad(const Layer& inputLayer) { + return inputLayer.getOutput(deviceId_).grad; + } + + /** + * Get the forward-input grad. + */ + const MatrixPtr& getInputGrad(int inputIndex, int deviceId) { + return inputLayers_[inputIndex]->getOutput(deviceId).grad; + } + + /** + * Get the forward-input label. + */ + const IVectorPtr& getInputLabel(const Layer& inputLayer) { + return inputLayer.getOutput(deviceId_).ids; + } + + /** + * Change the size of output (value, grad). + * Reset to value zero if isValueClean = true, + * Reset to grad zero if isGradClean = true. + */ + void resetSpecifyOutput(Argument& output, + size_t height, + size_t width, + bool isValueClean, + bool isGradClean); + + /** + * Add output argument to other devices. + */ + void addOutputArgument(int deviceId); + + public: + explicit Layer(const LayerConfig& config, bool useGpu = FLAGS_use_gpu); + virtual ~Layer() {} + + /// Register a Layer + static ClassRegistrar registrar_; + + /** + * Get the flag whether layer need to compute gradient. + */ + bool needGradient() const { return needGradient_; } + + /** + * Set the flag whether layer need to compute gradient. + */ + void setNeedGradient(bool need) { needGradient_ = need; } + + /** + * Set the flag whether layer need to re-compute sequence information, + * which includes sequenceStartPositions or subSequenceStartPositions. + */ + void setNeedSequenceInfo(bool need) { needSequenceInfo_ = need; } + + /** + * Get layer's name. + */ + const std::string& getName() const { return config_.name(); } + + /** + * Get layer's type. + */ + const std::string& getType() const { return config_.type(); } + + /** + * Get layer's size. + */ + size_t getSize() const { return config_.size(); } + + /** + * Get layer's deviceId. + */ + int getDeviceId() const { return deviceId_; } + + /** + * Add the inputLayer. + */ + void addPrev(LayerPtr l) { inputLayers_.push_back(l); } + + /** + * Get the size of inputLayer[i]. + */ + const LayerPtr& getPrev(size_t i) { return inputLayers_[i]; } + + /** + * Get the forward-output value. + */ + const MatrixPtr& getOutputValue() { return output_.value; } + + /** + * Get the forward-output label. + */ + const IVectorPtr& getOutputLabel() { return output_.ids; } + + /** + * Get the backward-Loss value. + */ + const MatrixPtr& getOutputGrad() { return output_.grad; } + /** + * If layer has multi-output, set output into outputMap_. + */ + void setOutput(const std::string& name, Argument* output) { + outputMap_[name] = output; + } + + /** + * Get the output map size, if layer has multi-output. + */ + size_t getOutputMapSize() { return outputMap_.size(); } + + /** + * Get the output based on layer's name. + */ + Argument& getOutput(const std::string& str = "") { + if (str == "") { + return output_; + } else { + auto output = outputMap_.find(str); + if (output != outputMap_.end()) { + return *output->second; + } else { + LOG(FATAL) << "No specific output " << str; + return *((Argument*)nullptr); + } + } + } + + /** + * Get the output based on deviceId. + */ + const Argument& getOutput(int deviceId) const { + if (deviceId == getDeviceId()) { + return output_; + } else { + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + if (outputOtherDevice_[i].deviceId == deviceId) { + return outputOtherDevice_[i]; + } + } + + LOG(FATAL) << "No specific device output "; + return *((Argument*)nullptr); + } + } + + /** + * Get layer's parameters. + */ + const std::vector& getParameters() { return parameters_; } + + /** + * Get layer's bias-parameters. + */ + const ParameterPtr& getBiasParameter() { return biasParameter_; } + + /** + * Create pointer of layer. + */ + static LayerPtr create(const LayerConfig& config); + + /** + * Resize the output matrix size. + */ + void resizeOutput(size_t height, size_t width); + + /** + * Resize the output matrix size, + * and reset value to zero. + */ + void reserveOutput(size_t height, size_t width); + + /** + * Resize the output matrix size, + * and reset value and grad to zero. + */ + void resetOutput(size_t height, size_t width); + + /** + * Clear the gradient of output. + */ + void zeroGrad(); + + /** + * Intialization. + * For example, adding input layers from layerMap and parameterMap. + */ + virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + + /** + * Intialization for sub network if there has sub network. + * @param rootNetwork root network + * @param config model config + * @param parameterTypes parameter's type + * @param useGpu whether to use gpu or not + */ + virtual void initSubNetwork(NeuralNetwork* rootNetwork, + const ModelConfig& config, + const std::vector& parameterTypes, + bool useGpu) {} + + /** + * @brief Access SubNetwork Object. + * If subnetwork exists, then invoke callback with subnetwrk. + * @param callback if sub-network is exist, the callback is invoked. + */ + virtual void accessSubNetwork( + const std::function& callback) {} + + /** + * If use sparse row matrix as parameter, + * prefetch feature ids in input label. + */ + virtual void prefetch() {} + + /** + * Forward propagation. + * All inherited implementation should call Layer::foward() function. + */ + virtual void forward(PassType passType) { + passType_ = passType; + if (!inputLayers_.empty() && needSequenceInfo_) { + const Argument& input = getInput(0); + output_.sequenceStartPositions = input.sequenceStartPositions; + output_.subSequenceStartPositions = input.subSequenceStartPositions; + output_.cpuSequenceDims = input.cpuSequenceDims; + } + } + + /** + * Reset the internal state variables. + * Allocate them if they have not been allocated. + * This function need to called before Layer::forward() for generating + * sequence. + * + * This is used for sequence generation. When generating sequence, the + * calculation at current timestamp depends on the state from previous + * timestamp. The model needs to keep the information about the previous + * timestamp in the state variables. Layers such as RecurrentLayer, + * LstmLayer and ContextLayer have state variables. + */ + virtual void resetState() {} + + /** + * Set layer state. + */ + virtual void setState(LayerStatePtr state) {} + + /** + * Get layer state. + * @return A copy of internal state. + */ + virtual LayerStatePtr getState() { return nullptr; } + + /** + * Show output state. + */ + void showOutputStats(); + + /** + * Backward propagation. + * Should only be called after Layer::forward() function. + */ + virtual void backward(const UpdateCallback& callback = nullptr) = 0; + + /** + * One pass is finished. + */ + virtual void onPassEnd() {} + + protected: + /** + * Forward of activation function. + */ + void forwardActivation(); + /** + * Backward of activation function. + */ + void backwardActivation(); + /** + * Forward of dropOut. + */ + void forwardDropOut(); + /** + * Initilize the needGradient_ flag. + */ + void initNeedFlags(); +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/LinearChainCRF.cpp b/paddle/legacy/gserver/layers/LinearChainCRF.cpp similarity index 100% rename from paddle/gserver/layers/LinearChainCRF.cpp rename to paddle/legacy/gserver/layers/LinearChainCRF.cpp diff --git a/paddle/legacy/gserver/layers/LinearChainCRF.h b/paddle/legacy/gserver/layers/LinearChainCRF.h new file mode 100644 index 0000000000000000000000000000000000000000..65e23905435da24a1a7554c30e33d303b05aef69 --- /dev/null +++ b/paddle/legacy/gserver/layers/LinearChainCRF.h @@ -0,0 +1,97 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +class LinearChainCRF { + public: + /** + * The size of para must be \f$(numClasses + 2) * numClasses\f$. + * The first numClasses values of para are for starting weights (\f$a\f$). + * The next numClasses values of para are for ending weights (\f$b\f$), + * The remaning values are for transition weights (\f$w\f$). + * + * The probability of a state sequence s of length \f$L\f$ is defined as: + * \f$P(s) = (1/Z) exp(a_{s_1} + b_{s_L} + * + \sum_{l=1}^L x_{s_l} + * + \sum_{l=2}^L w_{s_{l-1},s_l})\f$ + * where \f$Z\f$ is a normalization value so that the sum of \f$P(s)\f$ over + * all possible + * sequences is \f$1\f$, and \f$x\f$ is the input feature to the CRF. + */ + LinearChainCRF(int numClasses, real* para); + + /** + * Calculate the negative log likelihood of s given x. + * The size of x must be length * numClasses. Each consecutive numClasses + * values are the features for one time step. + */ + real forward(real* x, int* s, int length); + + /** + * Calculate the gradient with respect to x, a, b, and w. + * backward() can only be called after a corresponding call to forward() with + * the same x, s and length. + * The gradient with respect to a, b, and w will not be calculated if + * needWGrad is false. + * @note Please call getWGrad() and getXGrad() to get the gradient with + * respect to (a, b, w) and x respectively. + */ + void backward(real* x, int* s, int length, bool needWGrad); + + /** + * Find the most probable sequence given x. The result will be stored in s. + */ + void decode(real* x, int* s, int length); + + /* + * Return the gradient with respect to (a, b, w). It can only be called after + * a corresponding call to backward(). + */ + MatrixPtr getWGrad() { return matWGrad_; } + + /* + * Return the gradient with respect to x. It can only be called after a + * corresponding call to backward(). + */ + MatrixPtr getXGrad() { return matGrad_; } + + protected: + int numClasses_; + MatrixPtr a_; + MatrixPtr b_; + MatrixPtr w_; + MatrixPtr matWGrad_; + MatrixPtr da_; + MatrixPtr db_; + MatrixPtr dw_; + MatrixPtr ones_; + + MatrixPtr expX_; + MatrixPtr matGrad_; + MatrixPtr alpha_; + MatrixPtr beta_; + MatrixPtr maxX_; + MatrixPtr expW_; + + // track_(k,i) = j means that the best sequence at time k for class i comes + // from the sequence at time k-1 for class j + IVectorPtr track_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/LinearChainCTC.cpp b/paddle/legacy/gserver/layers/LinearChainCTC.cpp similarity index 100% rename from paddle/gserver/layers/LinearChainCTC.cpp rename to paddle/legacy/gserver/layers/LinearChainCTC.cpp diff --git a/paddle/legacy/gserver/layers/LinearChainCTC.h b/paddle/legacy/gserver/layers/LinearChainCTC.h new file mode 100644 index 0000000000000000000000000000000000000000..e6c4c7bfe0cdb1bbcafbf5b847ea592eef02794a --- /dev/null +++ b/paddle/legacy/gserver/layers/LinearChainCTC.h @@ -0,0 +1,50 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +class LinearChainCTC { + public: + LinearChainCTC(int numClasses, bool normByTimes); + + // Calculate the negative log probability as loss + real forward(real* softmaxSeq, + int softmaxSeqLen, + int* labelSeq, + int labelSeqLen); + + // calculate the gradient + void backward(real* softmaxSeq, + real* softmaxSeqGrad, + int* labelSeq, + int labelSeqLen); + + protected: + int numClasses_, blank_, totalSegments_, totalTime_; + bool normByTimes_; + bool isInvalid_; + + MatrixPtr logActs_, forwardVars_, backwardVars_, gradTerms_; + + real logProb_; + + void segmentRange(int& start, int& end, int time); +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/LstmCompute.cpp b/paddle/legacy/gserver/layers/LstmCompute.cpp new file mode 100644 index 0000000000000000000000000000000000000000..70f08e1d4efd2223e7ddec1b104e4ee63fc34de5 --- /dev/null +++ b/paddle/legacy/gserver/layers/LstmCompute.cpp @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "LstmCompute.h" +#include "hl_recurrent_apply.cuh" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +void LstmCompute::init(LayerConfig &config) { + activeNode_ = hlActiveType(config.active_type()); + activeGate_ = hlActiveType(config.active_gate_type()); + activeState_ = hlActiveType(config.active_state_type()); +} + +template <> +void LstmCompute::forwardOneSequence<0>(hl_lstm_value value, int frameSize) { + hl_cpu_lstm_forward(hppl::forward::lstm(), + value, + frameSize, + activeNode_, + activeGate_, + activeState_); +} + +template <> +void LstmCompute::backwardOneSequence<0>(hl_lstm_value value, + hl_lstm_grad grad, + int frameSize) { + hl_cpu_lstm_backward(hppl::backward::lstm(), + value, + grad, + frameSize, + activeNode_, + activeGate_, + activeState_); +} + +template <> +void LstmCompute::forwardBatch<0>(hl_lstm_value value, + int frameSize, + int batchSize) { + for (int b = 0; b < batchSize; b++) { + forwardOneSequence<0>(value, frameSize); + + value.gateValue += frameSize * 4; + value.stateValue += frameSize; + value.stateActiveValue += frameSize; + value.outputValue += frameSize; + if (value.prevStateValue) { + value.prevStateValue += frameSize; + } + } +} + +template <> +void LstmCompute::backwardBatch<0>(hl_lstm_value value, + hl_lstm_grad grad, + int frameSize, + int batchSize) { + for (int b = 0; b < batchSize; b++) { + backwardOneSequence<0>(value, grad, frameSize); + + value.gateValue += frameSize * 4; + value.stateValue += frameSize; + value.stateActiveValue += frameSize; + value.outputValue += frameSize; + if (value.prevStateValue) { + value.prevStateValue += frameSize; + } + + grad.gateGrad += frameSize * 4; + grad.stateGrad += frameSize; + grad.stateActiveGrad += frameSize; + grad.outputGrad += frameSize; + if (grad.prevStateGrad) { + grad.prevStateGrad += frameSize; + } + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/LstmCompute.cu b/paddle/legacy/gserver/layers/LstmCompute.cu similarity index 100% rename from paddle/gserver/layers/LstmCompute.cu rename to paddle/legacy/gserver/layers/LstmCompute.cu diff --git a/paddle/legacy/gserver/layers/LstmCompute.h b/paddle/legacy/gserver/layers/LstmCompute.h new file mode 100644 index 0000000000000000000000000000000000000000..ac40c35ef1b0a11e61b5d1b11476ffe7daff6d5e --- /dev/null +++ b/paddle/legacy/gserver/layers/LstmCompute.h @@ -0,0 +1,66 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ModelConfig.pb.h" +#include "hl_gpu.h" +#include "paddle/legacy/utils/Common.h" + +namespace paddle { + +class LstmCompute { + public: + void init(LayerConfig &config); + + /** + * LstmLayer batch compute API (forwardBatch, backwardBatch). + * If use batch compute api, lstm value(and grad) need to be batch structure. + * Compute order: + * forwardBatch: for 0 <= id < numBatch + * backwardBatch: for numBatch > id >= 0 + */ + template + void forwardBatch(hl_lstm_value value, int frameSize, int batchSize); + + template + void backwardBatch(hl_lstm_value value, + hl_lstm_grad grad, + int frameSize, + int batchSize); + + /** + * LstmLayer sequence compute API (forwardOneSequence, backwardOneSequence). + * Compute order(for each sequence): + * forwardOneSequence: + * if (!reversed) for 0 <= seqId < seqLength + * if (reversed) for seqLength > seqId >= 0 + * backwardOneSequence: + * if (!reversed) for seqLength > seqId >= 0 + * if (reversed) for 0 <= seqId < seqLength + */ + template + void forwardOneSequence(hl_lstm_value value, int frameSize); + template + void backwardOneSequence(hl_lstm_value value, + hl_lstm_grad grad, + int frameSize); + + public: + hl_activation_mode_t activeNode_; + hl_activation_mode_t activeGate_; + hl_activation_mode_t activeState_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/LstmLayer.cpp b/paddle/legacy/gserver/layers/LstmLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..43a55d8d490faf0049d47bbca6ae1947d13e6be8 --- /dev/null +++ b/paddle/legacy/gserver/layers/LstmLayer.cpp @@ -0,0 +1,805 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "LstmLayer.h" +#include "paddle/legacy/math/BaseMatrix.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Stat.h" + +DECLARE_bool(prev_batch_state); + +namespace paddle { + +REGISTER_LAYER(lstmemory, LstmLayer); + +bool LstmLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + if (!Layer::init(layerMap, parameterMap)) return false; + CHECK_EQ(1U, inputLayers_.size()); + CHECK_EQ(1U, parameters_.size()); + CHECK_EQ(getSize() * getSize() * 4, parameters_[0]->getSize()); + CHECK_EQ(getSize() * 7, biasParameter_->getSize()); + weight_.reset(new Weight(getSize(), getSize() * 4, parameters_[0])); + if (biasParameter_.get() != NULL) { + bias_.reset(new Weight(1, getSize() * 7, biasParameter_)); + if (bias_->getW()) { + localBias_ = Matrix::create(nullptr, + /* height= */ 1, + getSize() * 4, + /* trans= */ false, + useGpu_); + checkIg_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkFg_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkOg_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + + localBias_->setData(bias_->getW()->getData()); + checkIg_->setData(bias_->getW()->getData() + getSize() * 4); + checkFg_->setData(bias_->getW()->getData() + getSize() * 5); + checkOg_->setData(bias_->getW()->getData() + getSize() * 6); + } + + if (bias_->getWGrad()) { + localBiasGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize() * 4, + /* trans= */ false, + useGpu_); + checkIgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkFgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkOgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + localBiasGrad_->setData(bias_->getWGrad()->getData()); + checkIgGrad_->setData(bias_->getWGrad()->getData() + getSize() * 4); + checkFgGrad_->setData(bias_->getWGrad()->getData() + getSize() * 5); + checkOgGrad_->setData(bias_->getWGrad()->getData() + getSize() * 6); + } + } else { + LOG(FATAL) << "Bias should be here."; + } + reversed_ = config_.reversed(); + + // create IdentityActivation for using drop_rate + activation_.reset(ActivationFunction::create("")); + + LstmCompute::init(config_); + useBatch_ = true; + useSeqParallel_ = false; + if (useGpu_ && (getSize() == 32 || getSize() == 64)) { + useSeqParallel_ = true; + } + + return true; +} + +void LstmLayer::resetState() { + CHECK(!reversed_) << "state is not allowed for reversed lstmemory layer"; + Matrix::resizeOrCreate( + prevOutput_, 1, getSize(), /* trans= */ false, useGpu_); + Matrix::resizeOrCreate(prevState_, 1, getSize(), /* trans= */ false, useGpu_); + prevOutput_->resize(0, getSize()); + prevState_->resize(0, getSize()); + if (FLAGS_prev_batch_state) { + useBatch_ = true; + } else { + useBatch_ = false; + } +} + +void LstmLayer::setState(LayerStatePtr state) { + CHECK(state->value.size() == 2) << "two matrices are expected for LSTM state"; + prevOutput_->resize(state->value[0]->getHeight(), + state->value[0]->getWidth()); + prevState_->resize(state->value[1]->getHeight(), state->value[1]->getWidth()); + prevOutput_->copyFrom(*(state->value[0])); + prevState_->copyFrom(*(state->value[1])); +} + +LayerStatePtr LstmLayer::getState() { + LayerStatePtr res = std::make_shared(); + if (prevOutput_->getHeight() && prevOutput_->getWidth()) { + res->value.push_back(prevOutput_->clone(0, 0, useGpu_)); + res->value[0]->copyFrom(*prevOutput_); + res->value.push_back(prevState_->clone(0, 0, useGpu_)); + res->value[1]->copyFrom(*prevState_); + } else { + MatrixPtr output = + Matrix::create(1, getSize(), /* trans= */ false, useGpu_); + MatrixPtr state = Matrix::create(1, getSize(), /* trans= */ false, useGpu_); + output->resize(0, getSize()); + state->resize(0, getSize()); + res->value.push_back(output); + res->value.push_back(state); + } + return res; +} + +void LstmLayer::forward(PassType passType) { + REGISTER_TIMER_INFO("LstmFwTimer", getName().c_str()); + Layer::forward(passType); + + const Argument &input = getInput(0); + CHECK(input.sequenceStartPositions); + int batchSize = input.getBatchSize(); + resetOutput(batchSize, getSize()); + CHECK_EQ(getSize() * 4, input.value->getWidth()); + size_t numSequences = input.getNumSequences(); + const int *starts = input.sequenceStartPositions->getData(false); + CHECK_EQ(starts[numSequences], batchSize); + + Matrix::resizeOrCreate(gate_.value, + /* height= */ batchSize, + getSize() * 4, + /* trans= */ false, + useGpu_); + if (prevOutput_) { + size_t prevNumSeq = useBatch_ ? numSequences : 1; + if (prevOutput_->getHeight() == 0) { + prevOutput_->resize(prevNumSeq, getSize()); + prevState_->resize(prevNumSeq, getSize()); + prevOutput_->zeroMem(); + prevState_->zeroMem(); + } else { + CHECK_EQ(prevOutput_->getHeight(), prevNumSeq) + << "the number of sequences must be the same"; + } + Matrix::resizeOrCreate(totalState_, + prevState_->getHeight() + batchSize, + getSize(), + /*trans*/ false, + useGpu_); + state_.value = Matrix::create(nullptr, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + state_.value->setData(totalState_->getData() + + prevState_->getHeight() * getSize()); + } else { + Matrix::resizeOrCreate(state_.value, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + } + Matrix::resizeOrCreate(preOutput_.value, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + + if (!useBatch_) { + forwardSequence(batchSize, numSequences, starts, input.value); + } else { + if (!useSeqParallel_) { + forwardBatch(batchSize, numSequences, starts, input.value); + } else { + const int *starts = input.sequenceStartPositions->getData(useGpu_); + forwardSeqParallel(batchSize, numSequences, starts, input.value); + } + } + /* activation */ { forwardActivation(); } +} + +void LstmLayer::backward(const UpdateCallback &callback) { + REGISTER_TIMER_INFO("LstmBwTimer", getName().c_str()); + /* Do derivation */ { backwardActivation(); } + + const Argument &input = getInput(0); + CHECK(input.sequenceStartPositions); + int batchSize = input.getBatchSize(); + size_t numSequences = input.getNumSequences(); + + Matrix::resizeOrCreate(gate_.grad, + /* height= */ batchSize, + getSize() * 4, + /* trans= */ false, + useGpu_); + Matrix::resizeOrCreate(state_.grad, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + Matrix::resizeOrCreate(preOutput_.grad, + /* height= */ batchSize, + getSize(), + /* trans= */ false, + useGpu_); + state_.grad->zero(); + + const int *starts = input.sequenceStartPositions->getData(false); + if (!useBatch_) { + backwardSequence(batchSize, numSequences, starts, input.grad); + } else { + if (!useSeqParallel_) { + backwardBatch(batchSize, numSequences, starts, input.grad); + } else { + const int *starts = input.sequenceStartPositions->getData(useGpu_); + backwardSeqParallel(batchSize, numSequences, starts, input.grad); + } + } + + if (bias_) { + bias_->getParameterPtr()->incUpdate(callback); + } + weight_->getParameterPtr()->incUpdate(callback); +} + +void LstmLayer::forwardSequence(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputValue) { + REGISTER_TIMER_INFO("LstmFwSequenceTime", getName().c_str()); + gate_.value->assign(*inputValue); + if (bias_) { + gate_.value->addBias(*localBias_, 1); + } + + hl_lstm_value lstmValue; + lstmValue.checkIg = checkIg_->getData(); + lstmValue.checkFg = checkFg_->getData(); + lstmValue.checkOg = checkOg_->getData(); + lstmValue.gateValue = gate_.value->getData(); + lstmValue.stateValue = state_.value->getData(); + lstmValue.stateActiveValue = preOutput_.value->getData(); + lstmValue.outputValue = output_.value->getData(); + lstmValue.prevStateValue = nullptr; + if (reversed_) { + lstmValue.gateValue += (batchSize - 1) * getSize() * 4; + lstmValue.stateValue += (batchSize - 1) * getSize(); + lstmValue.stateActiveValue += (batchSize - 1) * getSize(); + lstmValue.outputValue += (batchSize - 1) * getSize(); + } + + auto nextFrame = [&lstmValue](bool reversed, int frameSize) { + lstmValue.prevStateValue = lstmValue.stateValue; + if (!reversed) { + lstmValue.gateValue += frameSize * 4; + lstmValue.stateValue += frameSize; + lstmValue.stateActiveValue += frameSize; + lstmValue.outputValue += frameSize; + } else { + lstmValue.gateValue -= frameSize * 4; + lstmValue.stateValue -= frameSize; + lstmValue.stateActiveValue -= frameSize; + lstmValue.outputValue -= frameSize; + } + }; + + MatrixPtr frameGate = Matrix::create(nullptr, + /* height= */ 1, + getSize() * 4, + /* trans= */ false, + useGpu_); + MatrixPtr frameOutput = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + + if (!reversed_) { + if (prevState_) { + lstmValue.prevStateValue = prevState_->getData(); + } + if (prevOutput_) { + frameGate->setData(lstmValue.gateValue); + frameGate->mul(*prevOutput_, *weight_->getW(), 1, 1); + } + } + AsyncGpuBlock asyncGpuBlock; + for (size_t n = 0; n < numSequences; ++n) { + int length; + if (!reversed_) { + length = starts[n + 1] - starts[n]; + } else { + length = starts[numSequences - n] - starts[numSequences - n - 1]; + } + for (int l = 0; l < length; ++l) { + if (useGpu_) { + LstmCompute::forwardOneSequence<1>(lstmValue, getSize()); + } else { + LstmCompute::forwardOneSequence<0>(lstmValue, getSize()); + } + + if (l != length - 1) { + frameOutput->setData(lstmValue.outputValue); + nextFrame(reversed_, getSize()); + frameGate->setData(lstmValue.gateValue); + frameGate->mul(*frameOutput, *weight_->getW(), 1, 1); + } + } + if (n != numSequences - 1) { + frameOutput->setData(lstmValue.outputValue); + nextFrame(reversed_, getSize()); + frameGate->setData(lstmValue.gateValue); + if (!reversed_) { + if (!prevState_) lstmValue.prevStateValue = nullptr; + if (prevOutput_) { + frameGate->mul(*frameOutput, *weight_->getW(), 1, 1); + } + } else { + lstmValue.prevStateValue = nullptr; + } + } + } + + if (!reversed_) { + if (prevState_) { + prevState_->assign(*state_.value->subMatrix(batchSize - 1, 1)); + } + if (prevOutput_) { + prevOutput_->assign(*output_.value->subMatrix(batchSize - 1, 1)); + } + } +} + +void LstmLayer::backwardSequence(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputGrad) { + REGISTER_TIMER_INFO("LstmBwSequenceTime", getName().c_str()); + MatrixPtr weightT = weight_->getW()->getTranspose(); + + hl_lstm_value lstmValue; + hl_lstm_grad lstmGrad; + lstmValue.checkIg = checkIg_->getData(); + lstmValue.checkFg = checkFg_->getData(); + lstmValue.checkOg = checkOg_->getData(); + lstmValue.gateValue = gate_.value->getData(); + lstmValue.stateValue = state_.value->getData(); + lstmValue.stateActiveValue = preOutput_.value->getData(); + lstmValue.outputValue = nullptr; + + if (bias_->getWGrad()) { + lstmGrad.checkIgGrad = checkIgGrad_->getData(); + lstmGrad.checkFgGrad = checkFgGrad_->getData(); + lstmGrad.checkOgGrad = checkOgGrad_->getData(); + } else { + lstmGrad.checkIgGrad = nullptr; + lstmGrad.checkFgGrad = nullptr; + lstmGrad.checkOgGrad = nullptr; + } + lstmGrad.gateGrad = gate_.grad->getData(); + lstmGrad.stateGrad = state_.grad->getData(); + lstmGrad.stateActiveGrad = nullptr; + lstmGrad.outputGrad = output_.grad->getData(); + + if (!reversed_) { + lstmValue.gateValue += (batchSize - 1) * getSize() * 4; + lstmGrad.gateGrad += (batchSize - 1) * getSize() * 4; + lstmValue.stateValue += (batchSize - 1) * getSize(); + lstmGrad.stateGrad += (batchSize - 1) * getSize(); + lstmValue.stateActiveValue += (batchSize - 1) * getSize(); + lstmGrad.outputGrad += (batchSize - 1) * getSize(); + lstmValue.prevStateValue = lstmValue.stateValue - getSize(); + lstmGrad.prevStateGrad = lstmGrad.stateGrad - getSize(); + } else { + lstmValue.prevStateValue = lstmValue.stateValue + getSize(); + lstmGrad.prevStateGrad = lstmGrad.stateGrad + getSize(); + } + + auto nextFrame = [&lstmValue, &lstmGrad](bool reversed, int frameSize) { + if (reversed) { + lstmValue.gateValue += frameSize * 4; + lstmGrad.gateGrad += frameSize * 4; + lstmValue.stateValue += frameSize; + lstmGrad.stateGrad += frameSize; + lstmValue.stateActiveValue += frameSize; + lstmGrad.outputGrad += frameSize; + lstmValue.prevStateValue = lstmValue.stateValue + frameSize; + lstmGrad.prevStateGrad = lstmGrad.stateGrad + frameSize; + } else { + lstmValue.gateValue -= frameSize * 4; + lstmGrad.gateGrad -= frameSize * 4; + lstmValue.stateValue -= frameSize; + lstmGrad.stateGrad -= frameSize; + lstmValue.stateActiveValue -= frameSize; + lstmGrad.outputGrad -= frameSize; + lstmValue.prevStateValue = lstmValue.stateValue - frameSize; + lstmGrad.prevStateGrad = lstmGrad.stateGrad - frameSize; + } + }; + + MatrixPtr frameGate = Matrix::create(nullptr, + /* height= */ 1, + getSize() * 4, + /* trans= */ false, + useGpu_); + MatrixPtr frameOutput = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + + { + AsyncGpuBlock asyncGpuBlock; + for (size_t n = 0; n < numSequences; ++n) { + int length; + int start; + if (reversed_) { + length = starts[n + 1] - starts[n]; + start = starts[n]; + } else { + length = starts[numSequences - n] - starts[numSequences - n - 1]; + start = starts[numSequences - n - 1]; + } + for (int l = 0; l < length; ++l) { + if (l == length - 1) { + lstmValue.prevStateValue = nullptr; + lstmGrad.prevStateGrad = nullptr; + } + if (useGpu_) { + LstmCompute::backwardOneSequence<1>(lstmValue, lstmGrad, getSize()); + } else { + LstmCompute::backwardOneSequence<0>(lstmValue, lstmGrad, getSize()); + } + if (l != length - 1) { + frameGate->setData(lstmGrad.gateGrad); + nextFrame(reversed_, getSize()); + frameOutput->setData(lstmGrad.outputGrad); + frameOutput->mul(*frameGate, *weightT, 1, 1); + } else { + nextFrame(reversed_, getSize()); + } + } + + if (weight_->getWGrad()) { + if (!reversed_) { + weight_->getWGrad()->mul( + *output_.value->subMatrix(start, length - 1)->getTranspose(), + *gate_.grad->subMatrix(start + 1, length - 1), + 1, + 1); + } else { + weight_->getWGrad()->mul( + *output_.value->subMatrix(start + 1, length - 1)->getTranspose(), + *gate_.grad->subMatrix(start, length - 1), + 1, + 1); + } + } + } + } + + if (inputGrad) { + inputGrad->add(*gate_.grad); + } + if (bias_ && bias_->getWGrad()) { + localBiasGrad_->collectBias(*gate_.grad, 1); + } +} + +void LstmLayer::forwardBatch(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputValue) { + REGISTER_TIMER_INFO("LstmFwBatchTime", getName().c_str()); + + hl_lstm_value lstmValue; + lstmValue.checkIg = checkIg_->getData(); + lstmValue.checkFg = checkFg_->getData(); + lstmValue.checkOg = checkOg_->getData(); + + if (!batchValue_) { + batchValue_.reset(new SequenceToBatch(useGpu_)); + } + batchValue_->resizeOrCreateBatch( + batchSize, numSequences, starts, reversed_, prevOutput_ ? true : false); + + batchValue_->resizeOrCreate(*output_.value); + batchValue_->copy(*inputValue, *gate_.value, /* seq2batch */ true); + if (bias_) { + gate_.value->addBias(*localBias_, 1); + } + + { + int numBatch = batchValue_->getNumBatch(); + int batchSize = 0; + AsyncGpuBlock asyncGpuBlock; + if (prevState_) { + lstmValue.prevStateValue = totalState_->getData(); + } else { + lstmValue.prevStateValue = nullptr; + } + for (int n = 0; n < numBatch; n++) { + MatrixPtr outputValue = batchValue_->getBatchValue(n); + MatrixPtr gateValue = batchValue_->getBatchValue(*gate_.value, n); + batchSize = outputValue->getHeight(); + + if (n != 0) { + MatrixPtr batch1 = batchValue_->getBatchValue(n - 1, batchSize); + gateValue->mul(*batch1, *weight_->getW(), 1, 1); + } else if (prevOutput_) { + Matrix::resizeOrCreate(prevBatchOutput2_, + gateValue->getHeight(), + getSize(), + false, + useGpu_); + batchValue_->prevOutput2Batch(*prevOutput_, *prevBatchOutput2_); + gateValue->mul(*prevBatchOutput2_, *weight_->getW(), 1, 1); + + batchValue_->prevOutput2Batch(*prevState_, + *totalState_->subMatrix(0, numSequences)); + } + + lstmValue.gateValue = gateValue->getData(); + lstmValue.outputValue = outputValue->getData(); + lstmValue.stateValue = + batchValue_->getBatchValue(*state_.value, n)->getData(); + lstmValue.stateActiveValue = + batchValue_->getBatchValue(*preOutput_.value, n)->getData(); + { + if (useGpu_) { + LstmCompute::forwardBatch<1>(lstmValue, getSize(), batchSize); + } else { + LstmCompute::forwardBatch<0>(lstmValue, getSize(), batchSize); + } + } + lstmValue.prevStateValue = lstmValue.stateValue; + } + } + { + REGISTER_TIMER_INFO("batchToSeq", getName().c_str()); + batchValue_->copyBackSeq(*output_.value); + } + if (prevOutput_) { + getPrevBatchOutput(numSequences); + getPrevBatchState(numSequences); + } +} + +void LstmLayer::getPrevBatchOutput(size_t numSequences) { + prevOutput_->resize(numSequences, getSize()); + batchValue_->getSeqOutputFromBatch(*prevOutput_, + *batchValue_->getBatchValue()); +} + +void LstmLayer::getPrevBatchState(size_t numSequences) { + prevState_->resize(numSequences, getSize()); + batchValue_->getSeqOutputFromBatch(*prevState_, *state_.value); +} + +void LstmLayer::backwardBatch(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputGrad) { + REGISTER_TIMER_INFO("LstmBwBatchTime", getName().c_str()); + + hl_lstm_value lstmValue; + lstmValue.checkIg = checkIg_->getData(); + lstmValue.checkFg = checkFg_->getData(); + lstmValue.checkOg = checkOg_->getData(); + + hl_lstm_grad lstmGrad; + lstmGrad.stateActiveGrad = preOutput_.grad->getData(); + + if (bias_->getWGrad()) { + lstmGrad.checkIgGrad = checkIgGrad_->getData(); + lstmGrad.checkFgGrad = checkFgGrad_->getData(); + lstmGrad.checkOgGrad = checkOgGrad_->getData(); + } else { + lstmGrad.checkIgGrad = nullptr; + lstmGrad.checkFgGrad = nullptr; + lstmGrad.checkOgGrad = nullptr; + } + + if (!batchGrad_) { + batchGrad_.reset(new SequenceToBatch(useGpu_)); + } + batchGrad_->shareIndexWith(*batchValue_); + + { + REGISTER_TIMER_INFO("seqToBatch", getName().c_str()); + batchGrad_->copyFromSeq(*output_.grad); + } + + { + MatrixPtr weightT = weight_->getW()->getTranspose(); + int numBatch = batchGrad_->getNumBatch(); + int batchSize = 0; + AsyncGpuBlock asyncGpuBlock; + for (int n = (int)numBatch - 1; n >= 0; n--) { + MatrixPtr outputGrad = batchGrad_->getBatchValue(n); + MatrixPtr gateGrad = batchGrad_->getBatchValue(*gate_.grad, n); + + lstmValue.gateValue = + batchGrad_->getBatchValue(*gate_.value, n)->getData(); + lstmValue.stateValue = + batchGrad_->getBatchValue(*state_.value, n)->getData(); + lstmValue.stateActiveValue = + batchGrad_->getBatchValue(*preOutput_.value, n)->getData(); + lstmGrad.stateGrad = + batchGrad_->getBatchValue(*state_.grad, n)->getData(); + lstmGrad.gateGrad = gateGrad->getData(); + lstmGrad.outputGrad = outputGrad->getData(); + { + batchSize = outputGrad->getHeight(); + if (n != 0) { + lstmValue.prevStateValue = + batchGrad_->getBatchValue(*state_.value, n - 1)->getData(); + lstmGrad.prevStateGrad = + batchGrad_->getBatchValue(*state_.grad, n - 1)->getData(); + } else { + if (prevState_) { + lstmValue.prevStateValue = totalState_->getData(); + lstmGrad.prevStateGrad = nullptr; + } else { + lstmValue.prevStateValue = nullptr; + lstmGrad.prevStateGrad = nullptr; + } + } + if (useGpu_) { + LstmCompute::backwardBatch<1>( + lstmValue, lstmGrad, getSize(), batchSize); + } else { + LstmCompute::backwardBatch<0>( + lstmValue, lstmGrad, getSize(), batchSize); + } + } + + if (n != 0) { + MatrixPtr tmp = batchGrad_->getBatchValue(n - 1, batchSize); + tmp->mul(*gateGrad, *weightT, 1, 1); + } + + if (n != 0 && weight_->getWGrad()) { + /* backward weight */ + MatrixPtr outputValue = batchValue_->getBatchValue(n - 1, batchSize); + weight_->getWGrad()->mul(*outputValue->getTranspose(), *gateGrad, 1, 1); + } else if (prevOutput_ && weight_->getWGrad()) { + weight_->getWGrad()->mul( + *prevBatchOutput2_->getTranspose(), *gateGrad, 1, 1); + } + } + } + + if (inputGrad) { + batchGrad_->add(*inputGrad, *gate_.grad, /* seq2batch */ false); + } + if (bias_ && bias_->getWGrad()) { + localBiasGrad_->collectBias(*gate_.grad, /* scale */ 1); + } +} + +void LstmLayer::forwardSeqParallel(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputValue) { + REGISTER_TIMER_INFO("LstmFwSeqParallelTime", getName().c_str()); + gate_.value->assign(*inputValue); + if (bias_) { + gate_.value->addBias(*localBias_, /* scale */ 1); + } + + real *gateValue = gate_.value->getData(); + real *stateValue = state_.value->getData(); + real *outputValue = output_.value->getData(); + real *preOutputValue = preOutput_.value->getData(); + real *checkIg = checkIg_->getData(); + real *checkFg = checkFg_->getData(); + real *checkOg = checkOg_->getData(); + real *weight = weight_->getW()->getData(); + hl_lstm_parallel_forward(gateValue, + stateValue, + preOutputValue, + outputValue, + checkIg, + checkFg, + checkOg, + weight, + starts, + getSize(), + numSequences, + reversed_, + activeNode_, + activeGate_, + activeState_); +} + +void LstmLayer::backwardSeqParallel(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputGrad) { + REGISTER_TIMER_INFO("LstmBwSeqParallelTime", getName().c_str()); + real *gateValue = gate_.value->getData(); + real *gateGrad = gate_.grad->getData(); + real *stateValue = state_.value->getData(); + real *stateGrad = state_.grad->getData(); + real *preOutputValue = preOutput_.value->getData(); + real *preOutputGrad = preOutput_.grad->getData(); + real *checkIg = checkIg_->getData(); + real *checkFg = checkFg_->getData(); + real *checkOg = checkOg_->getData(); + real *outputGrad = output_.grad->getData(); + real *weight = weight_->getW()->getData(); + + real *checkIgGrad; + real *checkFgGrad; + real *checkOgGrad; + if (bias_->getWGrad()) { + checkIgGrad = checkIgGrad_->getData(); + checkFgGrad = checkFgGrad_->getData(); + checkOgGrad = checkOgGrad_->getData(); + } else { + checkIgGrad = nullptr; + checkFgGrad = nullptr; + checkOgGrad = nullptr; + } + + hl_lstm_parallel_backward_data(gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + outputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + weight, + starts, + getSize(), + numSequences, + reversed_, + activeNode_, + activeGate_, + activeState_); + + if (inputGrad) { + inputGrad->add(*gate_.grad); + } + if (bias_ && bias_->getWGrad()) { + localBiasGrad_->collectBias(*gate_.grad, 1); + } + + real *outputValue = output_.value->getData(); + if (weight_->getWGrad()) { + real *weightGrad = weight_->getWGrad()->getData(); + hl_lstm_parallel_backward_weight(weightGrad, + outputValue, + gateGrad, + starts, + getSize(), + batchSize, + numSequences, + reversed_); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/LstmLayer.h b/paddle/legacy/gserver/layers/LstmLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..8c8b382f505d791fb1ef4265dcfe95046aa832fb --- /dev/null +++ b/paddle/legacy/gserver/layers/LstmLayer.h @@ -0,0 +1,221 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "LstmCompute.h" +#include "SequenceToBatch.h" +#include "paddle/legacy/math/BaseMatrix.h" +#include "paddle/legacy/math/Matrix.h" +namespace paddle { + +/** + * @brief LstmLayer takes 1 input layer with size * 4. + * Input layer is diveded into 4 equal parts: + * (input_s, input_ig, input_fg, input_og) + * + * For each sequence [start, end] it performs the following computation: + * @code + * output_{i} = actState(state_{i}) * actGate(outputGate_{i}) + * state_{i} = actInput(input_s_{i} + bias_s + + * output_{i-1} * recurrIW) * actGate(inputGate_{i}) + + * actGate(forgetGate_{i}) * state_{i-1} + * inputGate = input_ig_{i} + bias_ig + output_{i-1} * recurrIGW + + * state_{i-1} * inputCheck + * ouputGate = input_og_{i} + bias_og + output_{i-1} * recurrOGW + + * state_{i} * outputCheck + * forgetGate = input_fg_{i} + bias_fg + output_{i-1} * recurrFGW + + * state_{i-1} * forgetCheck + * @endcode + * + * - parameter[0] consists of (recurrIW, recurrIGW, recurrFGW, recurrOGW) + * - baisParameter consists of + * (bias_s, bias_ig, bias_og, bias_fg, inputCheck, forgetCheck, outputCheck) + * + * - actInput is defined by config active_type. + * - actState is defined by config active_state_type. + * - actGate is defined by config actvie_gate_type. + * + * There are two ways to compute, namely one sequence by one sequence or + * one batch by one batch. By default and no setting pre_batch_state true, + * it will compute batch by batch. + * + * The formula in the paper is as follows: + * \f[ + * i_t = \sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i) \\ + * f_t = \sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f) \\ + * \tilde{c_t} = tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c) \\ + * o_t = \sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o) \\ + * c_t = f_t * c_{t-1} + i_t * \tilde{c_t} \\ + * h_t = o_t tanh(c_t) + * \f] + * + * @note These \f$W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}\f$ + * operations on the input sequence were NOT included in LstmLayer. So + * users should use fc_layer or mixed_layer before lstm_later. + * + * The weight ([size, 4*size]) contains \f$W_{hi}, W_{hf}, W_{hc}, W_{ho}\f$. + * The bias contains \f$b_i, b_f, b_c, b_o\f$ and \f$W_{ci}, W_{cf}, W_{co}\f$. + */ + +class LstmLayer : public Layer, public LstmCompute { + public: + explicit LstmLayer(const LayerConfig &config) : Layer(config) {} + + bool init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) override; + + void forward(PassType passType) override; + + void backward(const UpdateCallback &callback) override; + + void resetState() override; + + void setState(LayerStatePtr state) override; + + LayerStatePtr getState() override; + + protected: + /** + * @brief Compute lstm forward one sequence by one sequence. + * @param batchSize The batchSize is not equal to the batch_size in + * the config file. It is the total words number of all samples + * in this forward batch. + * @param numSequences The sample number. It is equal to the batch_size + * in the config file. + * @param starts Each start position of each samples. + * @param inputValue The input values. + */ + void forwardSequence(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputValue); + /** + * Compute lstm backward one sequence by one sequence. + */ + void backwardSequence(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputGrad); + + /** + * Compute lstm forward one batch by one batch. The batch value is + * reorganized by SequenceToBatch class. The batch output value will + * be convert into sequence value after finishing forward. Here, one + * batch contains one word of each sample. If the length of each sample + * is not equality, the batch will not pads zero and contains less words. + * The total batch numbers are the max length of the sequence. The details + * can refer to SequenceToBatch class. On GPU mode, it will launch GPU + * kernel for loop. + * + * @code + * for (int i = 0; i < numBatch(max_sequence_length); ++i) { + * compute one batch. + * } + * @endcode + */ + void forwardBatch(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputValue); + /** + * Compute lstm backward one batch by one batch. + */ + void backwardBatch(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputGrad); + + /** + * This function only supports GPU. It not need to reorganize input into + * batch value. It will launch one kernel to parallelly compute forward + * propagation in sequence level. + */ + void forwardSeqParallel(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputValue); + /** + * Backward propagation corresponding to forwardSeqParallel. + */ + void backwardSeqParallel(int batchSize, + size_t numSequences, + const int *starts, + MatrixPtr inputGrad); + /** + * This function is used for sequence generation and get output after + * forwardBatch. + */ + void getPrevBatchOutput(size_t numSequences); + /** + * This function is used for sequence generation and get state after + * forwardBatch. + */ + void getPrevBatchState(size_t numSequences); + + protected: + /// Learned parameters, shape: (size, 4*size). + /// The weight ([size, 4*size]) contains \f$W_{hi}, W_{hf}, W_{hc}, W_{ho}\f$. + std::unique_ptr weight_; + /// Learned bias parameter, shape: (1, 7 * size). + /// The bias contains \f$b_i, b_f, b_c, b_o\f$ and \f$W_{ci}, W_{cf}, + /// W_{co}\f$. + std::unique_ptr bias_; + /// The reeal bias, point to \f$b_i, b_f, b_c, b_o\f$. + MatrixPtr localBias_; + /// The peephole connection for input gate. + MatrixPtr checkIg_; + /// The peephole connection for forget gate. + MatrixPtr checkFg_; + /// The peephole connection for output gate. + MatrixPtr checkOg_; + /// The gradient of real bias + MatrixPtr localBiasGrad_; + /// The gradient of peephole connection for input gates. + MatrixPtr checkIgGrad_; + /// The gradient of peephole connection for forget gates. + MatrixPtr checkFgGrad_; + /// The gradient of peephole connection for output gates. + MatrixPtr checkOgGrad_; + + /// Stores the cell state of previous time step, namely \f$c_{t-1}\f$. + Argument state_; + /// Stores the hidden of previous time step, namely \f$h_{t-1}\f$. + Argument preOutput_; + /// Stores the value and gradient of four gates, namely + /// \f$i_t, f_t, o_t, c_t\f$. + Argument gate_; + /// Whether it is reversed lstm. + bool reversed_; + /// Whether to use batch method to compute. + bool useBatch_; + /// Whether to use sequence parallell method to compute. + bool useSeqParallel_; + /// batchValue_ is used in method of batch calculation. It stores the + /// batch value after reorganized input. + std::unique_ptr batchValue_; + /// The gradient of batchValue_. + std::unique_ptr batchGrad_; + + /// Used in generation and stores the state of previous time step. + MatrixPtr prevState_; + /// Used in generation and stores the output of previous time step. + MatrixPtr prevOutput_; + MatrixPtr prevBatchOutput2_; + /// The total state. + MatrixPtr totalState_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/LstmStepLayer.cpp b/paddle/legacy/gserver/layers/LstmStepLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f02f8ad62fe4d4cb4bb580923200b398c8483a99 --- /dev/null +++ b/paddle/legacy/gserver/layers/LstmStepLayer.cpp @@ -0,0 +1,194 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "LstmCompute.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/* + * LstmStepLayer used in recurrent layer group. + */ +class LstmStepLayer : public Layer, public LstmCompute { + protected: + Argument state_; + Argument gate_; + Argument stateActive_; + MatrixPtr checkIg_, checkFg_, checkOg_; + MatrixPtr checkIgGrad_, checkFgGrad_, checkOgGrad_; + std::unique_ptr weight_; + + public: + explicit LstmStepLayer(const LayerConfig& config) : Layer(config) {} + + ~LstmStepLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(lstm_step, LstmStepLayer); + +bool LstmStepLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!Layer::init(layerMap, parameterMap)) return false; + CHECK_EQ(2U, inputLayers_.size()); + + checkIg_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkFg_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkOg_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkIgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkFgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + checkOgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + getSize(), + /* trans= */ false, + useGpu_); + + if (biasParameter_.get() != NULL) { + CHECK_EQ(getSize() * 3, biasParameter_->getSize()); + weight_.reset(new Weight(1, getSize() * 3, biasParameter_)); + if (weight_->getW()) { + real* data = weight_->getW()->getData(); + checkIg_->setData(data); + checkFg_->setData(data + getSize()); + checkOg_->setData(data + getSize() * 2); + } + + if (weight_->getWGrad()) { + real* data = weight_->getWGrad()->getData(); + checkIgGrad_->setData(data); + checkFgGrad_->setData(data + getSize()); + checkOgGrad_->setData(data + getSize() * 2); + } + } + + setOutput("state", &state_); + LstmCompute::init(config_); + return true; +} + +void LstmStepLayer::forward(PassType passType) { + REGISTER_TIMER_INFO("LstmRecurrentFwTime", getName().c_str()); + Layer::forward(passType); + + const Argument& input = getInput(0); + const Argument& prevState = getInput(1); + CHECK_EQ(getSize() * 4, input.value->getWidth()); + CHECK_EQ(getSize(), prevState.value->getWidth()); + int batchSize = input.getBatchSize(); + reserveOutput(batchSize, getSize()); + resetSpecifyOutput(state_, + batchSize, + getSize(), + /* isValueClean */ false, + /* isGradClean */ true); + resetSpecifyOutput(gate_, + batchSize, + getSize() * 4, + /* isValueClean */ false, + /* isGradClean */ false); + resetSpecifyOutput(stateActive_, + batchSize, + getSize(), + /* isValueClean */ false, + /* isGradClean */ false); + gate_.value->assign(*input.value); + + hl_lstm_value lstmValue; + lstmValue.checkIg = checkIg_->getData(); + lstmValue.checkFg = checkFg_->getData(); + lstmValue.checkOg = checkOg_->getData(); + lstmValue.gateValue = gate_.value->getData(); + lstmValue.stateValue = state_.value->getData(); + lstmValue.prevStateValue = prevState.value->getData(); + lstmValue.stateActiveValue = stateActive_.value->getData(); + lstmValue.outputValue = output_.value->getData(); + + if (useGpu_) { + LstmCompute::forwardBatch<1>(lstmValue, getSize(), batchSize); + } else { + LstmCompute::forwardBatch<0>(lstmValue, getSize(), batchSize); + } +} + +void LstmStepLayer::backward(const UpdateCallback& callback) { + REGISTER_TIMER_INFO("LstmRecurrentBwTime", getName().c_str()); + const Argument& input = getInput(0); + const Argument& prevState = getInput(1); + int batchSize = input.getBatchSize(); + + hl_lstm_value lstmValue; + hl_lstm_grad lstmGrad; + lstmValue.checkIg = checkIg_->getData(); + lstmValue.checkFg = checkFg_->getData(); + lstmValue.checkOg = checkOg_->getData(); + lstmValue.gateValue = gate_.value->getData(); + lstmValue.prevStateValue = prevState.value->getData(); + lstmValue.stateValue = state_.value->getData(); + lstmValue.stateActiveValue = stateActive_.value->getData(); + + lstmGrad.gateGrad = gate_.grad->getData(); + if (prevState.grad) { + lstmGrad.prevStateGrad = prevState.grad->getData(); + } else { + lstmGrad.prevStateGrad = nullptr; + } + lstmGrad.stateGrad = state_.grad->getData(); + lstmGrad.stateActiveGrad = stateActive_.grad->getData(); + lstmGrad.outputGrad = output_.grad->getData(); + lstmGrad.checkIgGrad = checkIgGrad_->getData(); + lstmGrad.checkFgGrad = checkFgGrad_->getData(); + lstmGrad.checkOgGrad = checkOgGrad_->getData(); + + if (useGpu_) { + LstmCompute::backwardBatch<1>(lstmValue, lstmGrad, getSize(), batchSize); + } else { + LstmCompute::backwardBatch<0>(lstmValue, lstmGrad, getSize(), batchSize); + } + + if (input.grad) { + input.grad->add(*gate_.grad); + } + + if (weight_) { + weight_->getParameterPtr()->incUpdate(callback); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MDLstmLayer.cpp b/paddle/legacy/gserver/layers/MDLstmLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4838183e8ccb213aa249fddf5102026198e98d3c --- /dev/null +++ b/paddle/legacy/gserver/layers/MDLstmLayer.cpp @@ -0,0 +1,769 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "LstmLayer.h" +#include "paddle/legacy/math/BaseMatrix.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +class CoordIterator { + public: + std::vector dims_; + std::vector directions_; + std::vector curPos_; + bool end_; + + void step(size_t d, bool reversed) { + if (directions_[d] ^ reversed) { + if (curPos_[d] == dims_[d] - 1) { + curPos_[d] = 0; + if (d) { + step(d - 1, reversed); + } else { + end_ = true; + } + } else { + curPos_[d]++; + } + } else { + if (curPos_[d] == 0) { + curPos_[d] = dims_[d] - 1; + if (d) { + step(d - 1, reversed); + } else { + end_ = true; + } + } else { + curPos_[d]--; + } + } + } + + public: + CoordIterator(std::vector dim, std::vector directions) + : dims_(dim), directions_(directions), end_(false) { + CHECK_EQ(dims_.size(), directions_.size()); + for (size_t i = 0; i < dims_.size(); i++) { + curPos_.push_back(-1); + } + } + CoordIterator& operator++() { + step(dims_.size() - 1, false); + return *this; + } + + CoordIterator& operator--() { + step(dims_.size() - 1, true); + return *this; + } + + std::vector& curPos() { return curPos_; } + + int offset() { + int offset = curPos_[0]; + for (size_t i = 1; i < dims_.size(); i++) { + offset = offset * dims_[i] + curPos_[i]; + } + return offset; + } + + int offset(const std::vector& pos) { + int offset = pos[0]; + for (size_t i = 1; i < dims_.size(); i++) { + offset = offset * dims_[i] + pos[i]; + } + return offset; + } + + std::vector& begin() { + for (size_t i = 0; i < dims_.size(); i++) { + curPos_[i] = directions_[i] ? 0 : dims_[i] - 1; + } + end_ = false; + return curPos_; + } + + std::vector& rbegin() { + for (size_t i = 0; i < dims_.size(); i++) { + curPos_[i] = directions_[i] ? dims_[i] - 1 : 0; + } + end_ = false; + return curPos_; + } + + bool end() { return end_; } + + bool getPrePos(const std::vector& delays, + int idx, + std::vector& prePos) { + bool isAvial = true; + prePos.clear(); + prePos.reserve(directions_.size()); + for (size_t i = 0; i < directions_.size(); i++) { + if (int(i) == idx) { + prePos.push_back(curPos_[i] + delays[i] * (directions_[i] ? 1 : -1)); + if (prePos[i] < 0) { + prePos[i] = 0; + isAvial = false; + } + if (prePos[i] >= dims_[i]) { + prePos[i] = dims_[i] - 1; + isAvial = false; + } + } else { + prePos.push_back(curPos_[i]); + } + } + return isAvial; + } + + bool getNextPos(const std::vector& delays, + int idx, + std::vector& nextPos) { + bool isAvial = true; + nextPos.clear(); + nextPos.reserve(directions_.size()); + for (size_t i = 0; i < directions_.size(); i++) { + if (int(i) == idx) { + nextPos.push_back(curPos_[i] - delays[i] * (directions_[i] ? 1 : -1)); + if (nextPos[i] < 0) { + nextPos[i] = 0; + isAvial = false; + } + if (nextPos[i] >= dims_[i]) { + nextPos[i] = dims_[i] - 1; + isAvial = false; + } + } else { + nextPos.push_back(curPos_[i]); + } + } + return isAvial; + } +}; +/* + * MDLstmLayer takes 1 input layer with size * (3+numDims). + * For each sequence [start, end] it performs the following computation: + * out_i = actState(state_i) * actGate(outputGate_i) + * + * For example the image with 2 dims, we take the scanning order from left-top + * to right-bottom, then the 2 previous states of the current pixels are the + * ones located at left and top. And each of them has a independent forget gate. + * + * state_i = actInput(input_i) * actGate(inputGate_i) + + * \sum{j}(actGate(forgetGate_i_j) * state_prev_i_j) + * + * inputGate = input_i * inputW + \sum{j}(output_prev_i_j * recurrInputW_j) + + * \sum{j}(state_prev_i_j * inputCheck_j) + * + * ouputGate = input_i * outputW + \sum{j}(output_prev_i_j * recurrOutputW_j) + + * state_i * outputCheck + * + * forgetGate_j = input_i * forgetW_j + \sum{j}(output_prev_i_j * + * recurrForgetW_j) + \sum{j}(state_prev_i_j * forgetCheck_j) + * + * IG Layer: (Input, InputGate, ForgetGates, OutputGate) * OutputSize + * */ + +class MDLstmLayer : public LstmLayer { + public: + explicit MDLstmLayer(const LayerConfig& config) : LstmLayer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + + void backward(const UpdateCallback& callback) override; + + protected: + void forwardOneSequence(int start, CoordIterator& coordIter); + void backwardOneSequence(int start, CoordIterator& coordIter); + void forwardGate2OutputSequence(int start, CoordIterator& coordIter); + void backwardGate2OutputSequence(int start, CoordIterator& coordIter); + + protected: + std::vector frameInputGate_; + std::vector frameForgetGate_; + std::vector frameOutputGate_; + std::vector frameInputNode_; + std::vector frameGate_; + std::vector frameState_; + std::vector framePreOutput_; + std::vector frameOutput_; + + // Activation + std::unique_ptr activationGate_; + std::unique_ptr activationState_; + + int numDims_; + size_t numBlocks_; + std::vector directions_; + std::vector delays_; + std::vector> dimsV_; +}; + +REGISTER_LAYER(mdlstmemory, MDLstmLayer); + +bool MDLstmLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!Layer::init(layerMap, parameterMap)) return false; + CHECK_EQ(1U, inputLayers_.size()); + CHECK_EQ(1U, parameters_.size()); + + numBlocks_ = getSize(); + numDims_ = config_.directions_size(); + CHECK_EQ(numBlocks_ * numBlocks_ * (3 + numDims_), parameters_[0]->getSize()); + + // inode(1), ig(1), fg(numDims_), og(1), peepIg(1), peepFg(numDims_), + // peepOg(1), then size of localBias_ is 3+numDims_ + CHECK_EQ(numBlocks_ * (5 + 2 * numDims_), biasParameter_->getSize()); + weight_.reset( + new Weight(numBlocks_, numBlocks_ * (3 + numDims_), parameters_[0])); + if (biasParameter_.get() != NULL) { + bias_.reset(new Weight(1, numBlocks_ * (5 + 2 * numDims_), biasParameter_)); + localBias_ = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_ * (3 + numDims_), + /* trans= */ false, + useGpu_); + checkIg_ = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + checkFg_ = Matrix::create(nullptr, + /* height= */ numDims_, + numBlocks_, + /* trans= */ false, + useGpu_); + checkOg_ = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + localBiasGrad_ = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_ * (3 + numDims_), + /* trans= */ false, + useGpu_); + checkIgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + checkFgGrad_ = Matrix::create(nullptr, + /* height= */ numDims_, + numBlocks_, + /* trans= */ false, + useGpu_); + checkOgGrad_ = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + + localBias_->setData(bias_->getW()->getData()); + checkIg_->setData(bias_->getW()->getData() + numBlocks_ * (3 + numDims_)); + checkFg_->setData(bias_->getW()->getData() + numBlocks_ * (4 + numDims_)); + checkOg_->setData(bias_->getW()->getData() + + numBlocks_ * (4 + 2 * numDims_)); + + if (bias_->getWGrad()) { + localBiasGrad_->setData(bias_->getWGrad()->getData()); + checkIgGrad_->setData(bias_->getWGrad()->getData() + + numBlocks_ * (3 + numDims_)); + checkFgGrad_->setData(bias_->getWGrad()->getData() + + numBlocks_ * (4 + numDims_)); + checkOgGrad_->setData(bias_->getWGrad()->getData() + + numBlocks_ * (4 + 2 * numDims_)); + } + } else { + LOG(FATAL) << "Bias should be here."; + } + for (int i = 0; i < numDims_; i++) { + directions_.push_back(config_.directions(i)); + } + for (int i = 0; i < numDims_; i++) { + delays_.push_back(-1); + } + activationGate_.reset(ActivationFunction::create(config_.active_gate_type())); + activationState_.reset( + ActivationFunction::create(config_.active_state_type())); + + return true; +} + +void MDLstmLayer::forward(PassType passType) { + Layer::forward(passType); + + const Argument& input = getInput(0); + CHECK(input.sequenceStartPositions); + int batchSize = input.getBatchSize(); + int numSequences = input.getNumSequences(); + resetOutput(batchSize, numBlocks_); + CHECK_EQ(numBlocks_ * (3 + numDims_), input.value->getWidth()); + const int* starts = input.sequenceStartPositions->getData(false); + CHECK_EQ(starts[numSequences], batchSize); + + int* dimsData = input.cpuSequenceDims->getData(); + CHECK_EQ(int(input.cpuSequenceDims->getSize()), numDims_* numSequences); + + for (int i = 0; i < numSequences; i++) { + std::vector dims; + for (int j = 0; j < numDims_; j++) { + dims.push_back(dimsData[i * numDims_ + j]); + } + dimsV_.push_back(dims); + } + + frameInputGate_.reserve(batchSize); + frameForgetGate_.reserve(batchSize); + frameOutputGate_.reserve(batchSize); + frameInputNode_.reserve(batchSize); + frameGate_.reserve(batchSize); + frameState_.reserve(batchSize); + framePreOutput_.reserve(batchSize); + frameOutput_.reserve(batchSize); + + Matrix::resizeOrCreate(gate_.value, + /* height= */ batchSize, + numBlocks_ * (3 + numDims_), + /* trans= */ false, + useGpu_); + + for (int i = frameGate_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_ * (3 + numDims_), + /* trans= */ false, + useGpu_); + arg.grad = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_ * (3 + numDims_), + /* trans= */ false, + useGpu_); + frameGate_.push_back(arg); + } + for (int i = frameInputGate_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + arg.grad = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + frameInputGate_.push_back(arg); + } + for (int i = frameForgetGate_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create(nullptr, + /* height= */ numDims_, + numBlocks_, + /* trans= */ false, + useGpu_); + arg.grad = Matrix::create(nullptr, + /* height= */ numDims_, + numBlocks_, + /* trans= */ false, + useGpu_); + frameForgetGate_.push_back(arg); + } + for (int i = frameOutputGate_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + arg.grad = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + frameOutputGate_.push_back(arg); + } + for (int i = frameInputNode_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + arg.grad = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + frameInputNode_.push_back(arg); + } + for (int i = frameState_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create( + /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); + frameState_.push_back(arg); + } + for (int i = framePreOutput_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create( + /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); + framePreOutput_.push_back(arg); + } + for (int i = frameOutput_.size(); i < batchSize; i++) { + Argument arg; + arg.value = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + arg.grad = Matrix::create(nullptr, + /* height= */ 1, + numBlocks_, + /* trans= */ false, + useGpu_); + frameOutput_.push_back(arg); + } + + for (int i = 0; i < batchSize; i++) { + frameOutput_[i].value->setData(output_.value->getData() + i * numBlocks_); + frameGate_[i].value->setData(gate_.value->getData() + + i * numBlocks_ * (3 + numDims_)); + frameInputNode_[i].value->setData(gate_.value->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * 0); + frameInputGate_[i].value->setData(gate_.value->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * 1); + frameForgetGate_[i].value->setData(gate_.value->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * 2); + frameOutputGate_[i].value->setData(gate_.value->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * (2 + numDims_)); + } + + AsyncGpuBlock asyncGpuBlock; + gate_.value->assign(*input.value); + + if (bias_) { + gate_.value->addBias(*localBias_, 1); + } + + for (int i = 0; i < numSequences; i++) { + CoordIterator coordIter(dimsV_[i], directions_); + forwardOneSequence(starts[i], coordIter); + } +} + +void MDLstmLayer::forwardGate2OutputSequence(int start, + CoordIterator& coordIter) { + int idxCurr = start + coordIter.offset(); + std::vector preOffsetV; + preOffsetV.reserve(numDims_); + for (int i = 0; i < numDims_; i++) { + std::vector prePos; + if (coordIter.getPrePos(delays_, i, prePos)) { + preOffsetV[i] = coordIter.offset(prePos); + } else { + preOffsetV[i] = -1; + } + } + + for (int i = 0; i < numDims_; i++) { + if (preOffsetV[i] >= 0) { + frameInputGate_[idxCurr].value->addDotMul( + *frameState_[start + preOffsetV[i]].value, *checkIg_, 1.0, 1.0); + + MatrixPtr fgGateOneDim = Matrix::create( + frameForgetGate_[idxCurr].value->getData() + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + MatrixPtr checkFgOneDim = + Matrix::create(checkFg_->getData() + i * numBlocks_, + 1.0, + numBlocks_, + false, + useGpu_); + fgGateOneDim->addDotMul( + *frameState_[start + preOffsetV[i]].value, *checkFgOneDim, 1.0, 1.0); + } + } + auto status = activationGate_->forward(frameInputGate_[idxCurr]); + status.check(); + status = activationGate_->forward(frameForgetGate_[idxCurr]); + status.check(); + status = activation_->forward(frameInputNode_[idxCurr]); + status.check(); + + frameState_[idxCurr].value->zeroMem(); + for (int i = 0; i < numDims_; i++) { + if (preOffsetV[i] >= 0) { + MatrixPtr fgGateOneDim = Matrix::create( + frameForgetGate_[idxCurr].value->getData() + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + frameState_[idxCurr].value->addDotMul( + *frameState_[start + preOffsetV[i]].value, *fgGateOneDim, 1.0, 1.0); + } + } + frameState_[idxCurr].value->addDotMul(*frameInputNode_[idxCurr].value, + *frameInputGate_[idxCurr].value, + 1.0, + 1.0); + + frameOutputGate_[idxCurr].value->addDotMul( + *frameState_[idxCurr].value, *checkOg_, 1.0, 1.0); + status = activationGate_->forward(frameOutputGate_[idxCurr]); + status.check(); + + framePreOutput_[idxCurr].value->copyFrom(*(frameState_[idxCurr].value)); + status = activationState_->forward(framePreOutput_[idxCurr]); + status.check(); + + frameOutput_[idxCurr].value->dotMul(*framePreOutput_[idxCurr].value, + *frameOutputGate_[idxCurr].value); +} + +void MDLstmLayer::forwardOneSequence(int start, CoordIterator& coordIter) { + for (coordIter.begin(); !coordIter.end(); ++coordIter) { + int offset = coordIter.offset(); + for (int i = 0; i < numDims_; i++) { + std::vector prePos; + if (coordIter.getPrePos(delays_, i, prePos)) { + int preOffset = coordIter.offset(prePos); + frameGate_[start + offset].value->mul( + *frameOutput_[start + preOffset].value, *weight_->getW(), 1.0, 1.0); + } + } + forwardGate2OutputSequence(start, coordIter); + } +} + +void MDLstmLayer::backward(const UpdateCallback& callback) { + const Argument& input = getInput(0); + CHECK(input.sequenceStartPositions); + int batchSize = input.getBatchSize(); + const int* starts = input.sequenceStartPositions->getData(false); + size_t numSequences = input.getNumSequences(); + + Matrix::resizeOrCreate(gate_.grad, + /* height= */ batchSize, + numBlocks_ * (3 + numDims_), + /* trans= */ false, + useGpu_); + + for (int i = 0; i < batchSize; i++) { + if (frameState_[i].grad == NULL) + frameState_[i].grad = Matrix::create( + /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); + } + for (int i = 0; i < batchSize; i++) { + if (framePreOutput_[i].grad == NULL) + framePreOutput_[i].grad = Matrix::create( + /* height= */ 1, numBlocks_, /* trans= */ false, useGpu_); + } + + for (int i = 0; i < batchSize; i++) { + frameOutput_[i].grad->setData(output_.grad->getData() + i * numBlocks_); + frameGate_[i].grad->setData(gate_.grad->getData() + + i * numBlocks_ * (3 + numDims_)); + frameInputNode_[i].grad->setData(gate_.grad->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * 0); + frameInputGate_[i].grad->setData(gate_.grad->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * 1); + frameForgetGate_[i].grad->setData(gate_.grad->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * 2); + frameOutputGate_[i].grad->setData(gate_.grad->getData() + + i * numBlocks_ * (3 + numDims_) + + numBlocks_ * (2 + numDims_)); + } + + { + AsyncGpuBlock asyncGpuBlock; + + for (size_t i = 0; i < numSequences; i++) { + CoordIterator coordIter(dimsV_[i], directions_); + backwardOneSequence(starts[i], coordIter); + } + } + + if (input.grad) { + input.grad->add(*gate_.grad); + } + if (bias_ && bias_->getWGrad()) { + localBiasGrad_->collectBias(*gate_.grad, 1); + bias_->getParameterPtr()->incUpdate(callback); + } + + weight_->getParameterPtr()->incUpdate(callback); +} + +void MDLstmLayer::backwardGate2OutputSequence(int start, + CoordIterator& coordIter) { + int idxCurr = start + coordIter.offset(); + std::vector preOffsetV; + std::vector nextOffsetV; + preOffsetV.reserve(numDims_); + nextOffsetV.reserve(numDims_); + for (int i = 0; i < numDims_; i++) { + std::vector prePos; + if (coordIter.getPrePos(delays_, i, prePos)) { + preOffsetV[i] = coordIter.offset(prePos); + } else { + preOffsetV[i] = -1; + } + std::vector nextPos; + if (coordIter.getNextPos(delays_, i, nextPos)) { + nextOffsetV[i] = coordIter.offset(nextPos); + } else { + nextOffsetV[i] = -1; + } + } + + framePreOutput_[idxCurr].grad->dotMul(*frameOutput_[idxCurr].grad, + *frameOutputGate_[idxCurr].value); + activationState_->backward(framePreOutput_[idxCurr]).check(); + frameState_[idxCurr].grad->copyFrom(*(framePreOutput_[idxCurr].grad)); + + frameOutputGate_[idxCurr].grad->dotMul(*frameOutput_[idxCurr].grad, + *framePreOutput_[idxCurr].value); + activationGate_->backward(frameOutputGate_[idxCurr]).check(); + + frameState_[idxCurr].grad->addDotMul( + *frameOutputGate_[idxCurr].grad, *checkOg_, 1.0, 1.0); + for (int i = 0; i < numDims_; i++) { + if (nextOffsetV[i] >= 0) { + frameState_[idxCurr].grad->addDotMul( + *frameInputGate_[start + nextOffsetV[i]].grad, *checkIg_, 1.0, 1.0); + + MatrixPtr fgGateOneDimGrad = Matrix::create( + frameForgetGate_[start + nextOffsetV[i]].grad->getData() + + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + MatrixPtr fgGateOneDimVal = Matrix::create( + frameForgetGate_[start + nextOffsetV[i]].value->getData() + + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + MatrixPtr checkFgOneDim = Matrix::create( + checkFg_->getData() + i * numBlocks_, 1, numBlocks_, false, useGpu_); + + frameState_[idxCurr].grad->addDotMul( + *fgGateOneDimGrad, *checkFgOneDim, 1.0, 1.0); + frameState_[idxCurr].grad->addDotMul( + *frameState_[start + nextOffsetV[i]].grad, + *fgGateOneDimVal, + 1.0, + 1.0); + } + } + + frameInputNode_[idxCurr].grad->dotMul(*frameState_[idxCurr].grad, + *frameInputGate_[idxCurr].value); + frameInputGate_[idxCurr].grad->dotMul(*frameState_[idxCurr].grad, + *frameInputNode_[idxCurr].value); + + frameForgetGate_[idxCurr].grad->zeroMem(); + for (int i = 0; i < numDims_; i++) { + if (preOffsetV[i] >= 0) { + MatrixPtr fgGateOneDimGrad = Matrix::create( + frameForgetGate_[idxCurr].grad->getData() + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + fgGateOneDimGrad->addDotMul(*frameState_[idxCurr].grad, + *frameState_[start + preOffsetV[i]].value, + 1.0, + 1.0); + } + } + + activationGate_->backward(frameInputGate_[idxCurr]).check(); + activationGate_->backward(frameForgetGate_[idxCurr]).check(); + activation_->backward(frameInputNode_[idxCurr]).check(); + + if (bias_->getWGrad()) { + for (int i = 0; i < numDims_; i++) { + if (preOffsetV[i] >= 0) { + checkIgGrad_->addDotMul(*frameInputGate_[idxCurr].grad, + *frameState_[start + preOffsetV[i]].value, + 1.0, + 1.0); + + MatrixPtr fgGateOneDimGrad = Matrix::create( + frameForgetGate_[idxCurr].grad->getData() + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + MatrixPtr checkFgOneDimGrad = + Matrix::create(checkFgGrad_->getData() + i * numBlocks_, + 1, + numBlocks_, + false, + useGpu_); + checkFgOneDimGrad->addDotMul(*fgGateOneDimGrad, + *frameState_[start + preOffsetV[i]].value, + 1.0, + 1.0); + } + } + checkOgGrad_->addDotMul( + *frameOutputGate_[idxCurr].grad, *frameState_[idxCurr].value, 1.0, 1.0); + } +} + +void MDLstmLayer::backwardOneSequence(int start, CoordIterator& coordIter) { + MatrixPtr weightT = weight_->getW()->getTranspose(); + for (coordIter.rbegin(); !coordIter.end(); --coordIter) { + int offset = coordIter.offset(); + backwardGate2OutputSequence(start, coordIter); + for (int i = 0; i < numDims_; i++) { + std::vector prePos; + if (coordIter.getPrePos(delays_, i, prePos)) { + int preOffset = coordIter.offset(prePos); + frameOutput_[start + preOffset].grad->mul( + *frameGate_[start + offset].grad, *weightT, 1.0, 1.0); + if (weight_->getWGrad()) { + weight_->getWGrad()->mul( + *frameOutput_[start + preOffset].value->getTranspose(), + *frameGate_[start + offset].grad, + 1.0, + 1.0); + } + } + } + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNAddtoLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNAddtoLayer.cpp similarity index 100% rename from paddle/gserver/layers/MKLDNNAddtoLayer.cpp rename to paddle/legacy/gserver/layers/MKLDNNAddtoLayer.cpp diff --git a/paddle/gserver/layers/MKLDNNAddtoLayer.h b/paddle/legacy/gserver/layers/MKLDNNAddtoLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNAddtoLayer.h rename to paddle/legacy/gserver/layers/MKLDNNAddtoLayer.h diff --git a/paddle/gserver/layers/MKLDNNBase.h b/paddle/legacy/gserver/layers/MKLDNNBase.h similarity index 100% rename from paddle/gserver/layers/MKLDNNBase.h rename to paddle/legacy/gserver/layers/MKLDNNBase.h diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNBatchNormLayer.cpp similarity index 100% rename from paddle/gserver/layers/MKLDNNBatchNormLayer.cpp rename to paddle/legacy/gserver/layers/MKLDNNBatchNormLayer.cpp diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.h b/paddle/legacy/gserver/layers/MKLDNNBatchNormLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNBatchNormLayer.h rename to paddle/legacy/gserver/layers/MKLDNNBatchNormLayer.h diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNConcatLayer.cpp similarity index 100% rename from paddle/gserver/layers/MKLDNNConcatLayer.cpp rename to paddle/legacy/gserver/layers/MKLDNNConcatLayer.cpp diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.h b/paddle/legacy/gserver/layers/MKLDNNConcatLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNConcatLayer.h rename to paddle/legacy/gserver/layers/MKLDNNConcatLayer.h diff --git a/paddle/legacy/gserver/layers/MKLDNNConvLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNConvLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b47bf14821fed4057227c80bb77e584649ab3145 --- /dev/null +++ b/paddle/legacy/gserver/layers/MKLDNNConvLayer.cpp @@ -0,0 +1,388 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNConvLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/Logging.h" + +using namespace mkldnn; // NOLINT +typedef memory::format format; + +namespace paddle { + +REGISTER_LAYER(mkldnn_conv, MKLDNNConvLayer); + +bool MKLDNNConvLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!MKLDNNLayer::init(layerMap, parameterMap)) { + return false; + } + CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), parameters_.size()); + CHECK(config_.shared_biases()) << "Only support shared biases yet"; + + oc_ = config_.num_filters(); + const ConvConfig& conf = config_.inputs(0).conv_conf(); + ic_ = conf.channels(); + fw_ = conf.filter_size(); + fh_ = conf.filter_size_y(); + pw_ = conf.padding(); + ph_ = conf.padding_y(); + dw_ = conf.dilation(); + dh_ = conf.dilation_y(); + sw_ = conf.stride(); + sh_ = conf.stride_y(); + gp_ = conf.groups(); + oh_ = conf.output_y(); + ow_ = conf.output_x(); + ih_ = conf.img_size_y(); + iw_ = conf.img_size(); + caffeMode_ = conf.caffe_mode(); + CHECK(caffeMode_) << "Only support caffe mode yet"; + CHECK(dh_ == 1 && dw_ == 1) << "Only support dilation 1 yet"; + // check group setting + CHECK_EQ((oc_ / gp_) * gp_, oc_) << "group is indivisible for oc"; + CHECK_EQ((ic_ / gp_) * gp_, ic_) << "group is indivisible for ic"; + + // create weight + size_t height = oc_ / gp_; + size_t width = ic_ * fh_ * fw_; + CHECK_EQ(parameters_[0]->getSize(), height * width); + weight_ = + std::unique_ptr(new Weight(height, width, parameters_[0], 0)); + + // create biases + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_, 0)); + } + return true; +} + +void MKLDNNConvLayer::convertWeightsFromPaddle() { + if (hasInitedWgt_) { + return; + } + + CHECK(wgtVal_) << "should have been initialized"; + // the paddle weight format is oihw or goihw + auto targetDim = wgtVal_->getDims(); + auto srcFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw; + wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim); + hasInitedWgt_ = true; +} + +void MKLDNNConvLayer::convertWeightsToPaddle() { + CHECK(wgtVal_) << "should have been initialized"; + auto targetDim = wgtVal_->getDims(); + auto dstFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw; + wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); +} + +void MKLDNNConvLayer::reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { + reshapeInput(bs, ih, iw); + + // cal output sizes + // oc can not be changed + int fh = (fh_ - 1) * dh_ + 1; + int fw = (fw_ - 1) * dw_ + 1; + oh = outputSize(ih, fh, ph_, sh_, caffeMode_); + ow = outputSize(iw, fw, pw_, sw_, caffeMode_); + + reshapeOutput(oh, ow); + resizeOutput(bs, oc * oh * ow); +} + +void MKLDNNConvLayer::resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + resetFwdPD(fwdPD_); + + resetFwdBuffers(fwdPD_, inputs[0], wgtVal_, biasVal_, out); + + resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, biasVal_, out); +} + +void MKLDNNConvLayer::resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + std::shared_ptr bwdWgtPD; + std::shared_ptr bwdDataPD; + + resetBwdWgtPD(bwdWgtPD); + + resetBwdDataPD(bwdDataPD); + + resetBwdBuffers(bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); + + resetBwdPipeline( + pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); +} + +void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) { + weight_->getParameterPtr()->incUpdate(callback); + if (biases_ && biases_->getWGrad()) { + biases_->getParameterPtr()->incUpdate(callback); + } +} + +void MKLDNNConvLayer::loadConvSettings(memory::dims& wgt, + memory::dims& bias, + memory::dims& stride, + memory::dims& dilation, + memory::dims& padL, + memory::dims& padR) { + wgt = (gp_ == 1) ? memory::dims{oc_, ic_, fh_, fw_} + : memory::dims{gp_, oc_ / gp_, ic_ / gp_, fh_, fw_}; + bias = memory::dims{oc_}; + stride = memory::dims{sh_, sw_}; + padL = memory::dims{ph_, pw_}; + padR = getPaddingR(); + // note: mkldnn dilation start from 0 + dilation = memory::dims{dh_ - 1, dw_ - 1}; +} + +void MKLDNNConvLayer::resetFwdPD( + std::shared_ptr& pd) { + // dims for conv + memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_}; + memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_}; + memory::dims wgtDims, biasDims, strides, dilations, padL, padR; + loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); + + prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring + : prop_kind::forward_training; + algorithm algo = algorithm::convolution_direct; + padding_kind padKind = padding_kind::zero; + conv_fwd::desc fwdDesc = + biases_ && biases_->getW() + ? conv_fwd::desc(pk, + algo, + MKLDNNMatrix::createMemoryDesc(inDims), + MKLDNNMatrix::createMemoryDesc(wgtDims), + MKLDNNMatrix::createMemoryDesc(biasDims), + MKLDNNMatrix::createMemoryDesc(outDims), + strides, + dilations, + padL, + padR, + padKind) + : conv_fwd::desc(pk, + algo, + MKLDNNMatrix::createMemoryDesc(inDims), + MKLDNNMatrix::createMemoryDesc(wgtDims), + MKLDNNMatrix::createMemoryDesc(outDims), + strides, + dilations, + padL, + padR, + padKind); + pd.reset(new conv_fwd::primitive_desc(fwdDesc, engine_)); +} + +void MKLDNNConvLayer::resetFwdBuffers( + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + CHECK(pd); + resetInValue( + in, std::make_shared(pd->src_primitive_desc())); + + resetOutValue(out, pd->dst_primitive_desc()); + + resetWithMatrix(wgt, weight_->getW(), pd->weights_primitive_desc()); + + if (biases_ && biases_->getW()) { + resetWithMatrix(bias, biases_->getW(), pd->bias_primitive_desc()); + } else { + bias = nullptr; + } +} + +void MKLDNNConvLayer::resetFwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + if (bias) { + fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out)); + } else { + fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out)); + } + pipeline.push_back(*fwd_); +} + +void MKLDNNConvLayer::resetBwdWgtPD( + std::shared_ptr& pd) { + memory::dims wgtDims, biasDims, strides, dilations, padL, padR; + loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); + + // create backward weight using input, output and weight value memory desc + CHECK(inVals_[0]) << "Should have internal input value"; + CHECK(outVal_) << "Should have internal output value"; + CHECK(wgtVal_) << "Should have weight value"; + algorithm algo = algorithm::convolution_direct; + padding_kind padKind = padding_kind::zero; + auto bwdWgtDesc = biasVal_ != nullptr + ? conv_bwdWgt::desc(algo, + inVals_[0]->getMemoryDesc(), + wgtVal_->getMemoryDesc(), + biasVal_->getMemoryDesc(), + outVal_->getMemoryDesc(), + strides, + padL, + padR, + padKind) + : conv_bwdWgt::desc(algo, + inVals_[0]->getMemoryDesc(), + wgtVal_->getMemoryDesc(), + outVal_->getMemoryDesc(), + strides, + padL, + padR, + padKind); + pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_)); + CHECK_PRIMITIVE_DESC_EQ(inVals_[0], pd->src_primitive_desc()); + CHECK_PRIMITIVE_DESC_EQ( + outVal_, + pd->diff_dst_primitive_desc(), + "primitive desc of out value and grad should be equal"); + CHECK_PRIMITIVE_DESC_EQ( + wgtVal_, + pd->diff_weights_primitive_desc(), + "primitive desc of weight value and grad should be equal"); +} + +void MKLDNNConvLayer::resetBwdDataPD( + std::shared_ptr& pd) { + pd = nullptr; + if (inputLayers_[0]->getOutput().grad == nullptr) { + return; + } + + memory::dims wgtDims, biasDims, strides, dilations, padL, padR; + loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); + CHECK(inVals_[0]) << "Should have internal input value"; + CHECK(outVal_) << "Should have internal output value"; + // create backward data using input and output value memory desc + // but using weight memory desc with any format + auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct, + inVals_[0]->getMemoryDesc(), + MKLDNNMatrix::createMemoryDesc(wgtDims), + outVal_->getMemoryDesc(), + strides, + padL, + padR, + padding_kind::zero); + pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_)); + CHECK_PRIMITIVE_DESC_EQ( + inVals_[0], + pd->diff_src_primitive_desc(), + "primitive desc of in value and grad should be equal"); + CHECK_PRIMITIVE_DESC_EQ( + outVal_, + pd->diff_dst_primitive_desc(), + "primitive desc of out value and grad should be equal"); +} + +void MKLDNNConvLayer::resetBwdBuffers( + std::shared_ptr& wgtPD, + std::shared_ptr& dataPD, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + CHECK(wgtPD); + resetOutGrad(out, wgtPD->diff_dst_primitive_desc()); + + resetWithMatrix( + wgt, weight_->getWGrad(), wgtPD->diff_weights_primitive_desc()); + CHECK_PRIMITIVE_DESC_EQ( + wgtVal_, + wgt->getPrimitiveDesc(), + "primitive desc of weight grad and value should be equal"); + + bias = nullptr; + if (biases_ && biases_->getWGrad()) { + resetWithMatrix( + bias, biases_->getWGrad(), wgtPD->diff_bias_primitive_desc()); + CHECK(bias); + CHECK_PRIMITIVE_DESC_EQ( + biasVal_, + bias->getPrimitiveDesc(), + "primitive desc of bias grad and value should be equal"); + } + + if (dataPD == nullptr) { + return; + } + resetInGrad(in, dataPD->diff_src_primitive_desc()); + resetWgtValBwdData(dataPD, wgtValBwdData_); +} + +void MKLDNNConvLayer::resetBwdPipeline( + std::vector& pipeline, + std::shared_ptr& wgtPD, + std::shared_ptr& dataPD, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0]); + // add bwdWgt handle + if (bias) { + bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVals_[0], *out, *wgt, *bias)); + } else { + bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVals_[0], *out, *wgt)); + } + pipeline.push_back(*bwdWgt_); + + if (dataPD == nullptr) { + return; + } + if (cvtWgtVal_) { + pipeline.push_back(*cvtWgtVal_); + } + // add bwdData handle + CHECK(wgtValBwdData_) << "Should have weight memory"; + bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in)); + pipeline.push_back(*bwdData_); +} + +void MKLDNNConvLayer::resetWgtValBwdData( + std::shared_ptr& dataPD, + MKLDNNMatrixPtr& wgt) { + if (dataPD == nullptr) { + return; + } + + // create new weight value for backward data, and create reorder if necessary + // since the primitive_desc would be different with wgtVal_ + CHECK(wgtVal_) << "should have weight value"; + if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) { + wgtValBwdData_ = MKLDNNMatrix::create(dataPD->weights_primitive_desc()); + cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_); + CHECK(cvtWgtVal_); + } else { + wgtValBwdData_ = wgtVal_; + } + VLOG(MKLDNN_FMTS) << "weight value format for backward data: " + << wgtValBwdData_->getFormat(); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNConvLayer.h b/paddle/legacy/gserver/layers/MKLDNNConvLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNConvLayer.h rename to paddle/legacy/gserver/layers/MKLDNNConvLayer.h diff --git a/paddle/legacy/gserver/layers/MKLDNNFcLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNFcLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f3747c7db84ef53fdcfa3741525a754fab63bca5 --- /dev/null +++ b/paddle/legacy/gserver/layers/MKLDNNFcLayer.cpp @@ -0,0 +1,262 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNFcLayer.h" +#include "paddle/legacy/utils/Logging.h" + +using namespace mkldnn; // NOLINT +typedef memory::format format; + +namespace paddle { + +REGISTER_LAYER(mkldnn_fc, MKLDNNFcLayer); + +bool MKLDNNFcLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!MKLDNNLayer::init(layerMap, parameterMap)) { + return false; + } + + CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), parameters_.size()); + CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet"; + + // output size, cat not be changed + oc_ = getSize(); + oh_ = 1; + ow_ = 1; + ih_ = 1; + iw_ = 1; + + // input size can not change in FC + iLayerSize_ = inputLayers_[0]->getSize(); + CHECK_EQ(parameters_[0]->getSize(), iLayerSize_ * oc_); + + // create weight + weight_ = + std::unique_ptr(new Weight(oc_, iLayerSize_, parameters_[0], 0)); + + // create biases + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_, 0)); + } + return true; +} + +void MKLDNNFcLayer::convertWeightsFromPaddle() { + if (hasInitedWgt_) { + return; + } + + CHECK(wgtVal_) << "should have been initialized"; + auto targetDim = wgtVal_->getDims(); + auto srcFmt = targetDim.size() == 2 ? format::io : format::ihwo; + wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim); + hasInitedWgt_ = true; +} + +void MKLDNNFcLayer::convertWeightsToPaddle() { + CHECK(wgtVal_) << "should have been initialized"; + auto targetDim = wgtVal_->getDims(); + auto dstFmt = targetDim.size() == 2 ? format::io : format::ihwo; + wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); +} + +void MKLDNNFcLayer::reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { + reshapeInput(bs, ih, iw); + + CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize()); + ic = iLayerSize_ / (ih * iw); + CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible"; + CHECK_EQ(size_t(oc), getSize()); + + reshapeOutput(oh, ow); + resizeOutput(bs, oc); +} + +void MKLDNNFcLayer::resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + resetFwdBuffers(inputs[0], wgtVal_, biasVal_, out); + + resetFwdPD(fwdPD_, inputs[0], wgtVal_, biasVal_, out); + + resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, biasVal_, out); +} + +void MKLDNNFcLayer::resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + std::shared_ptr bwdWgtPD; + std::shared_ptr bwdDataPD; + + resetBwdBuffers(inputs[0], wgtGrad_, biasGrad_, out); + + resetBwdWgtPD(bwdWgtPD, wgtGrad_, biasGrad_, out); + + resetBwdDataPD(bwdDataPD, inputs[0], out); + + resetBwdPipeline( + pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); +} + +void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) { + weight_->getParameterPtr()->incUpdate(callback); + if (biases_ && biases_->getWGrad()) { + biases_->getParameterPtr()->incUpdate(callback); + } +} + +void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + resetInValue(in); + CHECK(in); + in->downSpatial(); + + auto outPD = + MKLDNNMatrix::createPrimitiveDesc({bs_, oc_}, format::nc, engine_); + resetOutValue(out, outPD); + + format wgtFmt = format::oihw; + if (in->getFormat() == format::nChw8c) { + wgtFmt = format::oIhw8i; + } else if (in->getFormat() == format::nChw16c) { + wgtFmt = format::oIhw16i; + } + auto wgtPD = + MKLDNNMatrix::createPrimitiveDesc({oc_, ic_, ih_, iw_}, wgtFmt, engine_); + resetWithMatrix(wgt, weight_->getW(), wgtPD); + wgt->downSpatial(); + + if (biases_ && biases_->getW()) { + auto biasPD = MKLDNNMatrix::createPrimitiveDesc({oc_}, format::x, engine_); + resetWithMatrix(bias, biases_->getW(), biasPD); + } else { + bias = nullptr; + } +} + +void MKLDNNFcLayer::resetFwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr in, + MKLDNNMatrixPtr wgt, + MKLDNNMatrixPtr bias, + MKLDNNMatrixPtr out) { + CHECK(in); + CHECK(wgt); + CHECK(out); + prop_kind pk = prop_kind::forward; + fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk, + in->getMemoryDesc(), + wgt->getMemoryDesc(), + bias->getMemoryDesc(), + out->getMemoryDesc()) + : fc_fwd::desc(pk, + in->getMemoryDesc(), + wgt->getMemoryDesc(), + out->getMemoryDesc()); + pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_)); +} + +void MKLDNNFcLayer::resetFwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + if (bias) { + fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out)); + } else { + fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out)); + } + pipeline.push_back(*fwd_); +} + +void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0] && outVal_); + resetOutGrad(out, outVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); + + CHECK(wgtVal_); + resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc()); + + if (biasVal_) { + resetWithMatrix(bias, biases_->getWGrad(), biasVal_->getPrimitiveDesc()); + } else { + bias = nullptr; + } +} + +void MKLDNNFcLayer::resetBwdWgtPD( + std::shared_ptr& pd, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0]); + fc_bwdWgt::desc bwdWgtDesc = + bias ? fc_bwdWgt::desc(inVals_[0]->getMemoryDesc(), + wgt->getMemoryDesc(), + bias->getMemoryDesc(), + out->getMemoryDesc()) + : fc_bwdWgt::desc(inVals_[0]->getMemoryDesc(), + wgt->getMemoryDesc(), + out->getMemoryDesc()); + pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_)); +} + +void MKLDNNFcLayer::resetBwdDataPD( + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + pd = nullptr; + if (in == nullptr) { + return; + } + CHECK(wgtVal_); + fc_bwdData::desc bwdDataDesc = fc_bwdData::desc( + in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc()); + pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_)); +} + +void MKLDNNFcLayer::resetBwdPipeline( + std::vector& pipeline, + std::shared_ptr& bwdWgtPD, + std::shared_ptr& bwdDataPD, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& wgt, + MKLDNNMatrixPtr& bias, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0]); + if (bias) { + bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVals_[0], *out, *wgt, *bias)); + } else { + bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVals_[0], *out, *wgt)); + } + pipeline.push_back(*bwdWgt_); + + if (bwdDataPD == nullptr) { + return; + } + CHECK(wgtVal_) << "Should have weight memory"; + bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in)); + pipeline.push_back(*bwdData_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/legacy/gserver/layers/MKLDNNFcLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNFcLayer.h rename to paddle/legacy/gserver/layers/MKLDNNFcLayer.h diff --git a/paddle/legacy/gserver/layers/MKLDNNLRNLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNLRNLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..739482348f71bf144551cd1d881f1f1d7d69201f --- /dev/null +++ b/paddle/legacy/gserver/layers/MKLDNNLRNLayer.cpp @@ -0,0 +1,163 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNLRNLayer.h" +#include "paddle/legacy/utils/Logging.h" + +using namespace mkldnn; // NOLINT +typedef memory::format format; + +namespace paddle { + +REGISTER_LAYER(mkldnn_lrn, MKLDNNLRNLayer); + +bool MKLDNNLRNLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!MKLDNNLayer::init(layerMap, parameterMap)) { + return false; + } + + /* the size of inputs for norm-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + const NormConfig& conf = config_.inputs(0).norm_conf(); + localSize_ = conf.size(); + alpha_ = conf.scale(); + beta_ = conf.pow(); + + ic_ = conf.channels(); + oc_ = ic_; + iw_ = conf.img_size(); + ow_ = conf.output_x(); + ih_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + oh_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + CHECK_EQ(iw_, ow_); + CHECK_EQ(ih_, oh_); + return true; +} + +void MKLDNNLRNLayer::reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { + CHECK_EQ(inputLayers_.size(), 1UL); + reshapeInput(bs, ih, iw); + // ic_ and oc can not be changed + CHECK_EQ((size_t)ic, + inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) + << "Input channel can not be changed"; + oh = ih; + ow = iw; + reshapeOutput(oh, ow); + resizeOutput(bs, oc * oh * ow); +} + +void MKLDNNLRNLayer::resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + resetFwdBuffers(inputs[0], out); + + resetFwdPD(fwdPD_, inputs[0], out); + + resetFwdPipeline(pipeline, fwdPD_, inputs[0], out); +} + +void MKLDNNLRNLayer::resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + std::shared_ptr pd; + + resetBwdBuffers(inputs[0], out); + + resetBwdPD(pd, inputs[0], out); + + resetBwdPipeline(pipeline, pd, inputs[0], out); +} + +void MKLDNNLRNLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + resetInValue(in); + CHECK(in); + resetOutValue(out, in->getPrimitiveDesc()); +} + +void MKLDNNLRNLayer::resetFwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr in, + MKLDNNMatrixPtr out) { + prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring + : prop_kind::forward_training; + auto fwdDesc = lrn_fwd::desc(pk, + algorithm::lrn_across_channels, + in->getMemoryDesc(), + localSize_, + alpha_, + beta_, + 1.0f); + pd.reset(new lrn_fwd::primitive_desc(fwdDesc, engine_)); + // prepare workspace if necessary + workspace_ = + passType_ != PASS_TEST + ? std::make_shared(memory(pd->workspace_primitive_desc())) + : nullptr; +} + +void MKLDNNLRNLayer::resetFwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + fwd_ = workspace_ + ? std::make_shared(lrn_fwd(*pd, *in, *workspace_, *out)) + : std::make_shared(lrn_fwd(*pd, *in, *out)); + pipeline.push_back(*fwd_); +} + +void MKLDNNLRNLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0] && outVal_); + resetOutGrad(out, outVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); +} + +void MKLDNNLRNLayer::resetBwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + pd = nullptr; + if (in == nullptr) { + return; + } + CHECK(out); + auto bwdDesc = lrn_bwd::desc(algorithm::lrn_across_channels, + in->getMemoryDesc(), + out->getMemoryDesc(), + localSize_, + alpha_, + beta_, + 1.0f); + pd.reset(new lrn_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); +} + +void MKLDNNLRNLayer::resetBwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + if (pd == nullptr) { + return; + } + CHECK(inVals_[0]); + CHECK(workspace_); + bwdData_ = std::make_shared( + lrn_bwd(*pd, *inVals_[0], *out, *workspace_, *in)); + pipeline.push_back(*bwdData_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLRNLayer.h b/paddle/legacy/gserver/layers/MKLDNNLRNLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNLRNLayer.h rename to paddle/legacy/gserver/layers/MKLDNNLRNLayer.h diff --git a/paddle/gserver/layers/MKLDNNLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNLayer.cpp similarity index 100% rename from paddle/gserver/layers/MKLDNNLayer.cpp rename to paddle/legacy/gserver/layers/MKLDNNLayer.cpp diff --git a/paddle/legacy/gserver/layers/MKLDNNLayer.h b/paddle/legacy/gserver/layers/MKLDNNLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..94dc8625f68985a16bd68a6e36a1ad607d77a7cb --- /dev/null +++ b/paddle/legacy/gserver/layers/MKLDNNLayer.h @@ -0,0 +1,477 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "MKLDNNBase.h" +#include "mkldnn.hpp" +#include "paddle/legacy/math/MKLDNNMatrix.h" +#include "paddle/legacy/utils/Stat.h" + +DECLARE_bool(use_mkldnn); + +namespace paddle { + +class MKLDNNLayer; +typedef std::shared_ptr MKLDNNLayerPtr; + +/** + * @brief Base class of MKLDNNlayer. + * + */ +class MKLDNNLayer : public Layer { + protected: + // batch size + int bs_; + // their sizes are always from the first input layer + // input image channel, height and width + int ic_, ih_, iw_; + // output image channel, height and width + int oc_, oh_, ow_; + + // the condition that forward need be reset + size_t condition_; + // backward also need reset after reset forward handle + bool needResetBwd_; + + // is output only mkldnn + bool outputOnlyMKLDNN_; + + // mkldnn engine, stream and primivtives + mkldnn::engine engine_; + std::shared_ptr stream_; + std::shared_ptr fwd_; + std::shared_ptr bwdWgt_; + std::shared_ptr bwdData_; + std::vector pipelineFwd_; + std::vector pipelineBwd_; + + /* Value and grad are seperated as internal and external buffers. + * Each MKLDNNLayer must init or reset internal buffer at least, + * and the external buffer format is always nchw of nc(when h==w==1), + * which is the same format as paddle. + * The output_.value and output_.grad always save the external data, + * when mixed with cpu device. + * When all layers are mkldnn layers, they could save internal data. + */ + // below MKLDNNMatrix buffers are all internal buffers + std::vector inVals_; + std::vector inGrads_; + MKLDNNMatrixPtr outVal_; + MKLDNNMatrixPtr outGrad_; + // below are external value and grad + std::vector extInVals_; + std::vector extInGrads_; + MKLDNNMatrixPtr extOutVal_; + MKLDNNMatrixPtr extOutGrad_; + // convert handle between external and internal buffers + std::vector> cvtInVals_; + std::vector> cvtInGrads_; + std::shared_ptr cvtOutVal_; + std::shared_ptr cvtOutGrad_; + + // weight and bias are always internal buffers + MKLDNNMatrixPtr wgtVal_; + MKLDNNMatrixPtr wgtGrad_; + MKLDNNMatrixPtr biasVal_; + MKLDNNMatrixPtr biasGrad_; + + // merge grad primitive + std::shared_ptr mergeGrad_; + std::vector pipelineMergeGrad_; + // tmp input argument to save input grad, only used to merge grad + Argument tmpInArg_; + + public: + explicit MKLDNNLayer(const LayerConfig& config) + : Layer(config), + ih_(0), + iw_(0), + condition_(0), + needResetBwd_(true), + outputOnlyMKLDNN_(false), + engine_(mkldnn::engine::cpu, 0), + stream_(nullptr), + fwd_(nullptr), + bwdWgt_(nullptr), + bwdData_(nullptr) {} + + ~MKLDNNLayer() {} + + virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + virtual void forward(PassType passType); + virtual void backward(const UpdateCallback& callback); + + /** + * reshape the input and output channels and image sizes + * and reset output buffer size + */ + virtual void reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) = 0; + + /** + * reset the mkldnn forward primitve and memories + * only would be called when input size changes + * weight and bias buffers should be coverd by child class itself + */ + virtual void resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) = 0; + + /** + * reset the mkldnn backward primitve and memories + * only would be called when needed + * weight and bias buffers should be coverd by child class itself + */ + virtual void resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) = 0; + + /** + * Update weights and biases if necessary. + */ + virtual void updateWeights(const UpdateCallback& callback) {} + + /** + * convert weight from paddle format to mkldnn format + * weight_ will be override + */ + virtual void convertWeightsFromPaddle() {} + + /** + * convert mkldnn weight to paddle format + * weight_ will be override + */ + virtual void convertWeightsToPaddle() {} + + /** + * add this interface as public for unit test + */ + void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); } + + protected: + /** + * Some layers may have different condition to reset the forward. + * The function returns the condition that do not need reset forward. + */ + inline virtual size_t keepCondition() { + // reset when the first input element size changed, not only the batchsize + return inputLayers_[0]->getOutputValue()->getElementCnt(); + } + + /** + * reshape the input image sizes and input batchsize + */ + void reshapeInput(int& batchsize, int& height, int& width, size_t idx = 0); + + /** + * reshape output image sizes + */ + void reshapeOutput(size_t height, size_t width); + + /** + * reset MKLDNNMatrix from Matrix and internal primitive desc. + * reset nullptr if matrix or primitive desc is empty + */ + void resetWithMatrix(MKLDNNMatrixPtr& dnn, + const MatrixPtr& mat, + mkldnn::memory::primitive_desc pd); + + /** + * reset input value from input MKLDNNMatrix and internal primitive desc. + * reset both internal and external buffer and create reorder if necessary. + * input channel may be different in concat. + */ + void resetInValue( + MKLDNNMatrixPtr& in, + const std::shared_ptr& intPD = nullptr, + size_t idx = 0, + int inputChannel = 0); + + /** + * reset output value from internal primitive desc. + * reset both internal and external buffer and create reorder if necessary. + */ + void resetOutValue(MKLDNNMatrixPtr& out, + mkldnn::memory::primitive_desc intPD); + + /** + * reset input grad from internal primitive desc. + * reset both internal and external buffer and create reorder if necessary. + */ + void resetInGrad(MKLDNNMatrixPtr& in, + mkldnn::memory::primitive_desc intPD, + size_t idx = 0); + + /** + * reset output grad from internal primitive desc. + * merge grad if necessary. + * reset both internal and external buffer and create reorder if necessary. + * note: about merge grad, when this layer has several outputs, + * it could not be mixed with cpu device, + * since it can not get memory desc from cpu device. + */ + void resetOutGrad(MKLDNNMatrixPtr& out, mkldnn::memory::primitive_desc intPD); + + /** + * reset the merge grad primitive if necessary. + * note: do not support the grads mixed with cpu device, + * since it can not get memory desc from cpu device. + */ + void resetMergeGrad(MKLDNNMatrixPtr& out); + + protected: + /** + * Set deviceId of this layer. + */ + void setDevice(int id) { deviceId_ = id; } + + /** + * check the format is nchw or nc, + * which is supported by Paddle default memory layout + */ + bool isPaddleFormat(mkldnn::memory::format fmt) { + if (fmt == mkldnn::memory::format::nchw || + fmt == mkldnn::memory::format::nc) { + return true; + } else { + return false; + } + } + + /** + * If input only has MKLDNN device. + * Otherwise, only support the previous layer using CPU device. + */ + bool inputIsOnlyMKLDNN(int index = 0) { + int prevDevice = getPrev(index)->getDeviceId(); + if (prevDevice == MKLDNN_DEVICE) { + return true; + } else { + CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet"; + return false; + } + } + + /** + * If output only has MKLDNN device. + * Otherwise, other devices should only using CPU device. + */ + bool outputIsOnlyMKLDNN() { + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE) + << "Only support other device is CPU yet"; + } + outputOnlyMKLDNN_ = outputOtherDevice_.size() == 0; + return outputOnlyMKLDNN_; + } + + /** + * print info about sizes + */ + virtual void printSizeInfo() { + VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_ + << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_ + << ", oh: " << oh_ << ", ow: " << ow_; + } + + /** + * print the mkldnn memory format of value + */ + virtual void printValueFormat() { + for (size_t i = 0; i < inVals_.size(); ++i) { + if (!inVals_[i]) { + continue; + } + VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() + << ": " << (extInVals_[i] ? extInVals_[i]->getFormat() + : inVals_[i]->getFormat()) + << " >>> " << inVals_[i]->getFormat() << " >>>"; + } + if (outVal_) { + VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> " + << (extOutVal_ ? extOutVal_->getFormat() + : outVal_->getFormat()); + } + if (wgtVal_) { + VLOG(MKLDNN_FMTS) << "Weight value format: " << wgtVal_->getFormat(); + } + if (biasVal_) { + VLOG(MKLDNN_FMTS) << "Bias value format: " << biasVal_->getFormat(); + } + } + + /** + * print the mkldnn memory format of grad + */ + virtual void printGradFormat() { + if (outGrad_) { + VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< " + << (extOutGrad_ ? extOutGrad_->getFormat() + : outGrad_->getFormat()); + } + for (size_t i = 0; i < inGrads_.size(); ++i) { + if (!inGrads_[i]) { + continue; + } + VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() + << ": " << (extInGrads_[i] ? extInGrads_[i]->getFormat() + : inGrads_[i]->getFormat()) + << " <<< " << inGrads_[i]->getFormat() << " <<<"; + } + if (wgtGrad_) { + VLOG(MKLDNN_FMTS) << "Weight grad format: " << wgtGrad_->getFormat(); + } + if (biasGrad_) { + VLOG(MKLDNN_FMTS) << "Bias grad format: " << biasGrad_->getFormat(); + } + } + + private: + /** + * clear all grad + */ + void clearGrads() { + if (output_.grad) { + output_.grad->zeroMem(); + } + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + if (outputOtherDevice_[i].grad) { + outputOtherDevice_[i].grad->zeroMem(); + } + } + } + + /** + * Set deviceId of the params used in this layer. + */ + void setParamsDevice(int id, const ParameterMap& parameterMap) { + for (auto& inputConfig : config_.inputs()) { + if (inputConfig.has_input_parameter_name()) { + ParameterPtr parameter; + std::string name = inputConfig.input_parameter_name(); + CHECK(mapGet(name, parameterMap, ¶meter)) + << "Cannot find input parameter " << name << " for layer " + << getName(); + parameter->setDevice(id); + } + } + if (config_.has_bias_parameter_name()) { + ParameterPtr parameter; + std::string name = config_.bias_parameter_name(); + CHECK(mapGet(name, parameterMap, ¶meter)) + << "Cannot find bias parameter " << name << " for layer " + << getName(); + parameter->setDevice(id); + } + } + + /** + * Set output map of prev layers. + */ + void setOutputMap() { + outputMap_.clear(); + for (size_t i = 0; i < inputLayers_.size(); ++i) { + inputLayers_[i]->setOutput(getName(), &tmpInArg_); + } + } + + /** + * if have cpu device, share value and grad data with output_ + */ + void shareCPUDevice() { + if (outputIsOnlyMKLDNN()) { + return; + } + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + outputOtherDevice_[i].value = output_.value; + outputOtherDevice_[i].grad = output_.grad; + } + } + + /** + * Check the cpu device number of outputOtherDevice_. + * should have only one at most. + */ + void checkCPUOutputsNumber(int max = 1) { + int cnt = 0; + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + if (outputOtherDevice_[i].deviceId == CPU_DEVICE) { + ++cnt; + } + } + CHECK_LE(cnt, max) << "too much CPU devies"; + } + + /** + * copy SeqInfo from input layer to this output and other output devices. + * @note: do not use getInput(0) since it used this deviceId_, + * use "inputLayers_[0]->getOutput()" instead. + */ + void copySeqInfoToOutputs() { + if (inputLayers_.empty() || !needSequenceInfo_) { + return; + } + const Argument& input = inputLayers_[0]->getOutput(); + output_.sequenceStartPositions = input.sequenceStartPositions; + output_.subSequenceStartPositions = input.subSequenceStartPositions; + output_.cpuSequenceDims = input.cpuSequenceDims; + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + outputOtherDevice_[i].sequenceStartPositions = + output_.sequenceStartPositions; + outputOtherDevice_[i].subSequenceStartPositions = + output_.subSequenceStartPositions; + outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; + } + } + + void prepareValueConversions(std::vector& pipeline) { + // MKLDNNLayer output value should be MKLDNNMatrix + // so external output value is necessary. + // Then external input value is not necessary, + // since input may be mkldnn internal buffer. + CHECK(extOutVal_) << "external output value is necessary"; + output_.value = std::dynamic_pointer_cast(extOutVal_); + CHECK(inVals_[0] && outVal_) << "internal memories are necessary"; + for (size_t i = 0; i < cvtInVals_.size(); ++i) { + if (cvtInVals_[i]) { + pipeline.insert(pipeline.begin(), *cvtInVals_[i]); + } + } + if (cvtOutVal_) { + pipeline.push_back(*cvtOutVal_); + } + } + void prepareGradConversions(std::vector& pipeline) { + // external output grad is not necessary + // since output may be mkldnn internal buffer or merge them directly. + CHECK(outGrad_) << "internal output grad is necessary"; + if (extOutGrad_) { + CHECK_EQ(extOutGrad_->getData(), output_.grad->getData()) + << "the external buffer should share the same data with output_.grad"; + } + if (cvtOutGrad_) { + pipeline.insert(pipeline.begin(), *cvtOutGrad_); + } + for (size_t i = 0; i < cvtInGrads_.size(); ++i) { + if (cvtInGrads_[i]) { + pipeline.push_back(*cvtInGrads_[i]); + } + } + } +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/legacy/gserver/layers/MKLDNNPoolLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..83d980538d2b1b7351bf858ab391c14f6e7170bd --- /dev/null +++ b/paddle/legacy/gserver/layers/MKLDNNPoolLayer.cpp @@ -0,0 +1,195 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNPoolLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/Logging.h" + +using namespace mkldnn; // NOLINT +typedef memory::format format; + +namespace paddle { + +REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer); + +bool MKLDNNPoolLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!MKLDNNLayer::init(layerMap, parameterMap)) { + return false; + } + + /* the size of inputs for pool-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + const PoolConfig& conf = config_.inputs(0).pool_conf(); + ic_ = conf.channels(); + ih_ = conf.img_size_y(); + iw_ = conf.img_size(); + oc_ = ic_; + oh_ = conf.output_y(); + ow_ = conf.output_x(); + fh_ = conf.size_y(); + fw_ = conf.size_x(); + ph_ = conf.padding_y(); + pw_ = conf.padding(); + sh_ = conf.stride_y(); + sw_ = conf.stride(); + + const std::string& type = conf.pool_type(); + if (type == "max-projection") { + poolAlgo_ = algorithm::pooling_max; + } else if (type == "avg-projection") { + // paddle only use exclude_padding + poolAlgo_ = algorithm::pooling_avg_exclude_padding; + } else { + LOG(FATAL) << "unknow pooling type!"; + } + return true; +} + +void MKLDNNPoolLayer::reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { + reshapeInput(bs, ih, iw); + // ic_ and oc can not be changed + CHECK_EQ((size_t)ic, + inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) + << "Input channel can not be changed"; + + // cal output sizes + // paddle used false caffeMode for pooling + oh = outputSize(ih, fh_, ph_, sh_, false); + ow = outputSize(iw, fw_, pw_, sw_, false); + reshapeOutput(oh, ow); + + resizeOutput(bs, oc * oh * ow); +} + +void MKLDNNPoolLayer::resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + resetFwdBuffers(inputs[0], out); + + resetFwdPD(fwdPD_, inputs[0], out); + + resetFwdPipeline(pipeline, fwdPD_, inputs[0], out); +} + +void MKLDNNPoolLayer::resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + std::shared_ptr pd; + + resetBwdBuffers(inputs[0], out); + + resetBwdPD(pd, inputs[0], out); + + resetBwdPipeline(pipeline, pd, inputs[0], out); +} + +void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + resetInValue(in); + + memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_}; + CHECK(in); + auto outPD = + MKLDNNMatrix::createPrimitiveDesc(outDims, in->getFormat(), engine_); + resetOutValue(out, outPD); +} + +void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr in, + MKLDNNMatrixPtr out) { + memory::dims kernels = memory::dims{fh_, fw_}; + memory::dims strides = memory::dims{sh_, sw_}; + memory::dims padL = memory::dims{ph_, pw_}; + memory::dims padR = getPaddingR(); + padding_kind padKind = padding_kind::zero; + prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring + : prop_kind::forward_training; + auto fwdDesc = pool_fwd::desc(pk, + poolAlgo_, + in->getMemoryDesc(), + out->getMemoryDesc(), + strides, + kernels, + padL, + padR, + padKind); + pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_)); + + // prepare workspace if necessary + workspace_ = + (passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max) + ? std::make_shared(memory(pd->workspace_primitive_desc())) + : nullptr; +} + +void MKLDNNPoolLayer::resetFwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + fwd_ = workspace_ + ? std::make_shared(pool_fwd(*pd, *in, *out, *workspace_)) + : std::make_shared(pool_fwd(*pd, *in, *out)); + pipeline.push_back(*fwd_); +} + +void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0] && outVal_); + resetOutGrad(out, outVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); +} + +void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + pd = nullptr; + if (in == nullptr) { + return; + } + memory::dims kernels = memory::dims{fh_, fw_}; + memory::dims strides = memory::dims{sh_, sw_}; + memory::dims padL = memory::dims{ph_, pw_}; + memory::dims padR = getPaddingR(); + CHECK(out); + auto bwdDesc = pool_bwd::desc(poolAlgo_, + in->getMemoryDesc(), + out->getMemoryDesc(), + strides, + kernels, + padL, + padR, + padding_kind::zero); + pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); +} + +void MKLDNNPoolLayer::resetBwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + if (pd == nullptr) { + return; + } + + bwdData_ = + workspace_ + ? std::make_shared(pool_bwd(*pd, *out, *workspace_, *in)) + : std::make_shared(pool_bwd(*pd, *out, *in)); + pipeline.push_back(*bwdData_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.h b/paddle/legacy/gserver/layers/MKLDNNPoolLayer.h similarity index 100% rename from paddle/gserver/layers/MKLDNNPoolLayer.h rename to paddle/legacy/gserver/layers/MKLDNNPoolLayer.h diff --git a/paddle/gserver/layers/MKLPackedRecurrentLayer.cpp b/paddle/legacy/gserver/layers/MKLPackedRecurrentLayer.cpp similarity index 100% rename from paddle/gserver/layers/MKLPackedRecurrentLayer.cpp rename to paddle/legacy/gserver/layers/MKLPackedRecurrentLayer.cpp diff --git a/paddle/gserver/layers/MKLPackedRecurrentLayer.h b/paddle/legacy/gserver/layers/MKLPackedRecurrentLayer.h similarity index 100% rename from paddle/gserver/layers/MKLPackedRecurrentLayer.h rename to paddle/legacy/gserver/layers/MKLPackedRecurrentLayer.h diff --git a/paddle/legacy/gserver/layers/MKLPackedWeight.h b/paddle/legacy/gserver/layers/MKLPackedWeight.h new file mode 100644 index 0000000000000000000000000000000000000000..47f225bd03c3ccb594db952483d3b8397b61e1ec --- /dev/null +++ b/paddle/legacy/gserver/layers/MKLPackedWeight.h @@ -0,0 +1,86 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/math/MathFunctions.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/parameter/Weight.h" + +namespace paddle { + +class MKLPackedWeight { + protected: + /// The pointer of weight + real *weight_; + /// The pointer of cblas packed gemm to weight + real *packedWeight_; + size_t height_; + size_t width_; + bool transW_; + + public: + explicit MKLPackedWeight(MatrixPtr weight, bool transW = false) { + packedWeight_ = nullptr; + weight_ = weight->getData(); + height_ = weight->getHeight(); + width_ = weight->getWidth(); + transW_ = transW; + } + + ~MKLPackedWeight() { free_(); } + + void pack() { pack_(weight_); } + + void gemm_compute(const MatrixPtr src, MatrixPtr dst) { + cblas_sgemm_compute(CblasRowMajor, + CblasNoTrans, + CblasPacked, + src->getHeight(), + transW_ ? height_ : width_, + transW_ ? width_ : height_, + src->getData(), + src->getWidth(), + packedWeight_, + width_, + 1.0, + dst->getData(), + dst->getWidth()); + } + + protected: + void pack_(real *src) { + if (!packedWeight_) { + packedWeight_ = cblas_sgemm_alloc(CblasBMatrix, 1, width_, height_); + } + cblas_sgemm_pack(CblasRowMajor, + CblasBMatrix, + transW_ ? CblasTrans : CblasNoTrans, + 1, + transW_ ? height_ : width_, + transW_ ? width_ : height_, + 1.0, + src, + width_, + packedWeight_); + } + + void free_() { + if (packedWeight_) { + cblas_sgemm_free(packedWeight_); + } + } +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/MaxIdLayer.cpp b/paddle/legacy/gserver/layers/MaxIdLayer.cpp similarity index 100% rename from paddle/gserver/layers/MaxIdLayer.cpp rename to paddle/legacy/gserver/layers/MaxIdLayer.cpp diff --git a/paddle/legacy/gserver/layers/MaxLayer.cpp b/paddle/legacy/gserver/layers/MaxLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b51251b663cf818fbe662a96b7c0d55a615640d4 --- /dev/null +++ b/paddle/legacy/gserver/layers/MaxLayer.cpp @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MaxLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(max, MaxLayer); + +void MaxLayer::forward(PassType passType) { + SequencePoolLayer::forward(passType); + + IVector::resizeOrCreate( + maxIndex_, newBatchSize_ * getSize(), useGpu(deviceId_)); + maxIndex_->zeroMem(); + + MatrixPtr inputValue = getInputValue(0); + MatrixPtr outputValue = getOutputValue(); + + { + REGISTER_TIMER_INFO("MaxLayerForward", getName().c_str()); + outputValue->maxSequenceForward( + *inputValue, *startPositions_->getVector(useGpu_), *maxIndex_); + } + + if (config_.output_max_index()) { + // copy maxIndex_ to output + outputValue->copyFrom(*maxIndex_); + } else { + /* add the bias-vector AFTER max operation */ + if (biases_.get() != NULL) { + outputValue->addBias(*(biases_->getW()), 1); + } + /* activation */ { forwardActivation(); } + } +} + +void MaxLayer::backward(const UpdateCallback& callback) { + CHECK(!config_.output_max_index()) + << "backward is not available when output_max_index is set"; + SequencePoolLayer::backward(callback); + + MatrixPtr inputGrad = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + if (inputGrad) { + REGISTER_TIMER_INFO("MaxLayerBackward", getName().c_str()); + inputGrad->maxSequenceBackward( + *outputGrad, *(startPositions_->getVector(useGpu_)), *maxIndex_); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MaxLayer.h b/paddle/legacy/gserver/layers/MaxLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..12d0128e39f2113d0e156813f9b3657cae145eed --- /dev/null +++ b/paddle/legacy/gserver/layers/MaxLayer.h @@ -0,0 +1,58 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "SequencePoolLayer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * A layer for "internal max" for sequence input. + * Input: one or more sequences. Each sequence contains some instances. + * If SequenceLevel = kNonSeq: + * Output: output size is the number of input sequences (NOT input instances) + * output[i] = max_{for each instance in this sequence}{input[i]} + * If stride_ > 0: + * Output: a shorten sequence. Stride is the step size by which we slide a + * window upon the input sequence, and the max pooling operation is + * then applied to each interval independently. + * If SequenceLevel = kSeq: + * Check input sequence must has sub-sequence + * Output: output size is the number of input sub-sequences + * output[i] = max_{for each instance in this sub-sequence}{input[i]} + * + * The config file api is pooling_layer. + */ + +class MaxLayer : public SequencePoolLayer { + protected: + // maxIndex_[i][j] = k : the value at (i, j) is from input[k]. + IVectorPtr maxIndex_; + + public: + explicit MaxLayer(const LayerConfig& config) : SequencePoolLayer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override { + return SequencePoolLayer::init(layerMap, parameterMap); + } + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/MaxOutLayer.cpp b/paddle/legacy/gserver/layers/MaxOutLayer.cpp similarity index 100% rename from paddle/gserver/layers/MaxOutLayer.cpp rename to paddle/legacy/gserver/layers/MaxOutLayer.cpp diff --git a/paddle/legacy/gserver/layers/MaxOutLayer.h b/paddle/legacy/gserver/layers/MaxOutLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..e56f34b8e02bf1dd48c6b5b6ea135cc1009c25b5 --- /dev/null +++ b/paddle/legacy/gserver/layers/MaxOutLayer.h @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * A layer to do max out on conv layer output. + * Input: output of a conv layer. + * Output: feature map size same as input. Channel is (input channel) / groups. + * So the num of channels should be able to devided by groups. + * + * The config file api is maxout_layer. + */ + +class MaxOutLayer : public Layer { + protected: + size_t groups_; + size_t imgSizeH_, imgSizeW_; + /// outputChannels_ = channels_ / groups_ + size_t channels_, outputChannels_; + /// feature length = imgSizeH_ * imgSizeW_ + size_t featLen_; + IVectorPtr maxoutId_; + + public: + /// return imgSizeH_ * imgSizeW_ * outputChannels_; + size_t getSize(); + + explicit MaxOutLayer(const LayerConfig& config) : Layer(config) {} + virtual ~MaxOutLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.cpp b/paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a1cc59a719e43453a8919a5827369982ac355480 --- /dev/null +++ b/paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.cpp @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MaxPoolWithMaskLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +bool MaxPoolWithMaskLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + PoolLayer::init(layerMap, parameterMap); + setOutput("mask", &mask_); + return true; +} + +size_t MaxPoolWithMaskLayer::getSize() { + CHECK_EQ(inputLayers_.size(), 1UL); + size_t layerSize = 0; + + outputY_ = outputSize(imgSizeY_, + sizeY_, + confPaddingY_, + strideY_, + /* caffeMode */ false); + outputX_ = outputSize(imgSize_, + sizeX_, + confPadding_, + stride_, + /* caffeMode */ false); + + layerSize = outputX_ * outputY_ * channels_; + getOutput().setFrameHeight(outputY_); + getOutput().setFrameWidth(outputX_); + + return layerSize; +} + +void MaxPoolWithMaskLayer::forward(PassType passType) { + size_t size = getSize(); + MatrixPtr inputV = inputLayers_[0]->getOutputValue(); + int batchSize = inputV->getHeight(); + resetOutput(batchSize, size); + + MatrixPtr outV = getOutputValue(); + CHECK_EQ(size, outV->getWidth()); + + resetSpecifyOutput(mask_, + batchSize, + size, + /* isValueClean */ false, + /* isGradClean */ true); + + MatrixPtr maskV = mask_.value; + outV->maxPoolForward(*inputV, + imgSizeY_, + imgSize_, + channels_, + sizeX_, + sizeY_, + strideY_, + stride_, + outputY_, + outputX_, + confPaddingY_, + confPadding_, + maskV); +} + +void MaxPoolWithMaskLayer::backward(const UpdateCallback& callback) { + (void)callback; + if (NULL == getInputGrad(0)) { + return; + } + + MatrixPtr outGrad = getOutputGrad(); + MatrixPtr inputV = inputLayers_[0]->getOutputValue(); + MatrixPtr outV = getOutputValue(); + MatrixPtr inputGrad = inputLayers_[0]->getOutputGrad(); + + inputGrad->maxPoolBackward(*inputV, + imgSizeY_, + imgSize_, + *outGrad, + *outV, + sizeX_, + sizeY_, + strideY_, + stride_, + outputY_, + outputX_, + 1, + 1, + confPaddingY_, + confPadding_); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.h b/paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..fcd5388abe3f8229dfa418e6917a8a73c93900a7 --- /dev/null +++ b/paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.h @@ -0,0 +1,40 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "PoolLayer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * @brief Basic parent layer of different kinds of pooling + */ +class MaxPoolWithMaskLayer : public PoolLayer { + protected: + Argument mask_; + + public: + explicit MaxPoolWithMaskLayer(const LayerConfig& config) + : PoolLayer(config) {} + + size_t getSize(); + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MixedLayer.cpp b/paddle/legacy/gserver/layers/MixedLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..63e658c09c2b3bae30c8b2890e4d67f72266dd4d --- /dev/null +++ b/paddle/legacy/gserver/layers/MixedLayer.cpp @@ -0,0 +1,176 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MixedLayer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(mixed, MixedLayer); + +bool MixedLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + if (!Layer::init(layerMap, parameterMap)) return false; + + CHECK_EQ(inputLayers_.size(), parameters_.size()); + projections_.resize(inputLayers_.size()); + for (size_t i = 0; i < inputLayers_.size(); i++) { + if (config_.inputs(i).has_proj_conf()) { + projections_[i].reset(Projection::create( + config_.inputs(i).proj_conf(), parameters_[i], useGpu_)); + } else { + CHECK(!parameters_[i]) << "should no parameters for operators"; + } + } + for (auto& operator_conf : config_.operator_confs()) { + for (auto& input_index : operator_conf.input_indices()) { + CHECK(!config_.inputs(input_index).has_proj_conf()); + } + operators_.emplace_back(Operator::create(operator_conf, useGpu_)); + } + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + sharedBias_ = config_.shared_biases(); + size_t psize = config_.bias_size(); + biases_ = std::unique_ptr(new Weight(1, psize, biasParameter_)); + } + + return true; +} + +void MixedLayer::prefetch() { + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (projections_[i]) { + projections_[i]->prefetch(&getInput(i)); + } + } +} + +void MixedLayer::resetState() { + for (auto& proj : projections_) { + if (proj) { + proj->resetState(); + } + } +} + +void MixedLayer::setState(LayerStatePtr state) { + CHECK(projectionStateMatrixSize_.size() == projections_.size()) + << "projection size mis-match"; + + int start = 0; + LayerStatePtr statePtr = std::make_shared(); + for (int i = 0; i < (int)projectionStateMatrixSize_.size(); i++) { + if (projectionStateMatrixSize_[i] > 0) { + statePtr->value.clear(); + for (int j = start; j < start + projectionStateMatrixSize_[i]; j++) { + statePtr->value.push_back(state->value[j]); + } + projections_[i]->setState(statePtr); + start += projectionStateMatrixSize_[i]; + } + } + CHECK((int)state->value.size() == start) << "state matrix size mis-match"; +} + +// Return state which consists of all projections states +LayerStatePtr MixedLayer::getState() { + bool init = projectionStateMatrixSize_.size() == 0; + LayerStatePtr res = std::make_shared(); + for (int i = 0; i < (int)projections_.size(); i++) { + LayerStatePtr statePtr = + projections_[i] ? projections_[i]->getState() : nullptr; + int stateSize = statePtr == nullptr ? 0 : statePtr->value.size(); + if (init) { + projectionStateMatrixSize_.push_back(stateSize); + } else { + CHECK(projectionStateMatrixSize_[i] == stateSize) + << "state matrix size mis-match"; + } + if (statePtr != nullptr) { + for (auto& matrixPtr : statePtr->value) { + res->value.push_back(matrixPtr); + } + } + } + return res; +} + +void MixedLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = getInput(0).getBatchSize(); + int size = getSize(); + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + resetOutput(batchSize, size); + } + + MatrixPtr outV = getOutputValue(); + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (projections_[i]) { + projections_[i]->forward(&getInput(i), &output_, passType); + } + } + + std::vector ins; + for (auto& op : operators_) { + ins.clear(); + for (auto& input_index : op->getConfig().input_indices()) { + ins.push_back(&getInput(input_index)); + } + op->forward(ins, &output_, passType); + } + + /* add the bias-vector */ + if (biases_.get() != NULL) { + REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); + outV->addBias(*(biases_->getW()), 1, sharedBias_); + } + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void MixedLayer::backward(const UpdateCallback& callback) { + /* Do activation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + if (biases_ && biases_->getWGrad()) { + REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); + biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBias_); + + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (projections_[i]) { + projections_[i]->backward(callback); + } + } + + for (auto& op : operators_) { + op->backward(); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MixedLayer.h b/paddle/legacy/gserver/layers/MixedLayer.h similarity index 100% rename from paddle/gserver/layers/MixedLayer.h rename to paddle/legacy/gserver/layers/MixedLayer.h diff --git a/paddle/gserver/layers/MultiBoxLossLayer.cpp b/paddle/legacy/gserver/layers/MultiBoxLossLayer.cpp similarity index 100% rename from paddle/gserver/layers/MultiBoxLossLayer.cpp rename to paddle/legacy/gserver/layers/MultiBoxLossLayer.cpp diff --git a/paddle/gserver/layers/MultiBoxLossLayer.h b/paddle/legacy/gserver/layers/MultiBoxLossLayer.h similarity index 100% rename from paddle/gserver/layers/MultiBoxLossLayer.h rename to paddle/legacy/gserver/layers/MultiBoxLossLayer.h diff --git a/paddle/gserver/layers/MultinomialSampler.cpp b/paddle/legacy/gserver/layers/MultinomialSampler.cpp similarity index 100% rename from paddle/gserver/layers/MultinomialSampler.cpp rename to paddle/legacy/gserver/layers/MultinomialSampler.cpp diff --git a/paddle/legacy/gserver/layers/MultinomialSampler.h b/paddle/legacy/gserver/layers/MultinomialSampler.h new file mode 100644 index 0000000000000000000000000000000000000000..ed445352418f8504e52a6139492e3577a95eecb1 --- /dev/null +++ b/paddle/legacy/gserver/layers/MultinomialSampler.h @@ -0,0 +1,81 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/legacy/utils/Common.h" + +namespace paddle { + +/** + * @brief Given the probability of N objects, the sampler random select + * one of the object. + * @note: prob does not have to be unnormalized. + * + * The space requirement is O(N)=O(N * sizeof(Interval)). + * The computational complexity of generate one sample is O(1). + */ +class MultinomialSampler { + public: + MultinomialSampler(const real* prob, int size); + + //! protobuf always using double. + static MultinomialSampler* create(const double* prob, int size) { +#ifdef PADDLE_TYPE_DOUBLE + return new MultinomialSampler(prob, size); +#else + std::unique_ptr tmp(new real[size]); + std::copy(prob, prob + size, tmp.get()); + return new MultinomialSampler(tmp.get(), size); +#endif + } + + /** + * @brief Generate a random sample. + * @param g is a random number engine. See . + * @return Random integer. + */ + template + int gen(URNG& g) { + return gen1([&g, this]() { return rand_(g); }); + } + + protected: + /** + * @brief Generation + * @param[in] rand rand is a real random number distribution + * for the range [0, size). + * @return random int number or intervals_[random_int_number].otherId. + */ + template + int gen1(Rand rand) { + double r = rand(); // NOLINT + int i = (int)r; + r -= i; + return r < intervals_[i].thresh ? i : intervals_[i].otherId; + } + + struct Interval { + int otherId; + real thresh; + }; + + /// The probability of each interval will be 1./size + std::vector intervals_; + std::uniform_real_distribution rand_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/MultiplexLayer.cpp b/paddle/legacy/gserver/layers/MultiplexLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9ca2b2417596e7978ea6b84ec76bcb8a305a4f5d --- /dev/null +++ b/paddle/legacy/gserver/layers/MultiplexLayer.cpp @@ -0,0 +1,180 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + *@brief This layer multiplex multiple layers according to the index, + * which is provided by the first input layer. + * - Input[0]: the index of the layer to output of size batchSize. + * - Input[1:N]; the candidate output data. + * For each index i from 0 to batchSize -1, the output is the i-th row of the + * (index[i] + 1)-th layer. + * + * For each i-th row of output: + * + * \f[ + * y[i][j] = x_{x_{0}[i] + 1}[i][j], j = 0,1, ... , (x_{1}.width - 1) + * \f] + * where, y is output. \f$x_{k}\f$ is the k-th input layer and + * \f$k = x_{0}[i] + 1\f$. + */ + +class MultiplexLayer : public Layer { + protected: + /** + * @brief A struct is used to save the copy information, includes input + * layer index and copy size. + */ + struct CopyInfo { + CopyInfo(int inStartIdx, int inLength, int inCopyIdx) + : startIdx(inStartIdx), length(inLength), copyIdx(inCopyIdx) {} + + /// The start row of input. + int startIdx; + /// Number of rows. If the layer index in Input[0] is not consecutive, + /// the length is one. Otherwise, the length is > 1 and copy multi rows + /// once. + int length; + /// The copied layer index, which needs to add 1. + int copyIdx; + }; + + /// A list of CopyInfo used to save copy information. + std::vector copySchedule_; + + /// Temporary matrix pointer to point to input data. + MatrixPtr tmpSrc_; + /// Temporary matrix pointer to point to output data. + MatrixPtr tmpDest_; + + public: + explicit MultiplexLayer(const LayerConfig& config) : Layer(config) {} + + ~MultiplexLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + private: + /** + * @brief Calculate copy info for input layers. + */ + void calculateCopySchedule(const IVectorPtr& copyIds, size_t numIns); +}; + +REGISTER_LAYER(multiplex, MultiplexLayer); + +void MultiplexLayer::calculateCopySchedule(const IVectorPtr& copyIds, + size_t numIns) { + copySchedule_.clear(); + CopyInfo prevCopyInfo(0, 0, -1); + for (size_t i = 0; i < copyIds->getSize(); i++) { + int copyId = copyIds->getElement(i); + CHECK_GE(copyId, 0); + CHECK_LT(copyId, int(numIns)); + // copy same input layer with prevous and will copy consecutive. + if (copyId == prevCopyInfo.copyIdx) { + ++prevCopyInfo.length; + } else { + if (prevCopyInfo.copyIdx != -1) { + copySchedule_.emplace_back(prevCopyInfo); + } + prevCopyInfo.startIdx = i; + prevCopyInfo.length = 1; + prevCopyInfo.copyIdx = copyId; + } + } + if (prevCopyInfo.copyIdx != -1) { + copySchedule_.emplace_back(prevCopyInfo); + } +} + +bool MultiplexLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_GE(inputLayers_.size(), 2U); + + tmpSrc_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); + tmpDest_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); + return true; +} + +void MultiplexLayer::forward(PassType passType) { + Layer::forward(passType); + + IVectorPtr copyIds = getInput(0).ids; + MatrixPtr inV1 = getInputValue(1); + CHECK_EQ(copyIds->getSize(), inV1->getHeight()); + for (size_t i = 2; i < inputLayers_.size(); i++) { + CHECK_EQ(inV1->getHeight(), getInputValue(i)->getHeight()); + CHECK_EQ(inV1->getWidth(), getInputValue(i)->getWidth()); + } + + calculateCopySchedule(copyIds, inputLayers_.size() - 1); + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(inV1->getHeight(), inV1->getWidth()); + } + + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwLMultplexingTimer", getName().c_str()); + AsyncGpuBlock block; + for (const CopyInfo& info : copySchedule_) { + outV->subMatrix(info.startIdx, info.length, tmpDest_) + ->copyFrom(*getInputValue(info.copyIdx + 1) + ->subMatrix(info.startIdx, info.length, tmpSrc_)); + } + } + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void MultiplexLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + MatrixPtr outG = getOutputGrad(); + + { + REGISTER_TIMER_INFO("BwLMultiplexTimer", getName().c_str()); + AsyncGpuBlock block; + for (const CopyInfo& info : copySchedule_) { + if (getInputGrad(info.copyIdx + 1)) { + getInputGrad(info.copyIdx + 1) + ->subMatrix(info.startIdx, info.length, tmpDest_) + ->add(*outG->subMatrix(info.startIdx, info.length, tmpSrc_)); + } + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/NCELayer.cpp b/paddle/legacy/gserver/layers/NCELayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ae4d6408168d1597760fe0094bc04f9cef657da4 --- /dev/null +++ b/paddle/legacy/gserver/layers/NCELayer.cpp @@ -0,0 +1,323 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "Layer.h" +#include "MultinomialSampler.h" +#include "paddle/legacy/math/MathFunctions.h" + +namespace paddle { + +/** + * Noise-contrastive estimation. + * Implements the method in the following paper: + * A fast and simple algorithm for training neural probabilistic language + * models. + * + * The config file api is nce_layer. + */ +class NCELayer : public Layer { + int numClasses_; + /// number of input layer besides labelLayer and weightLayer + int numInputs_; + LayerPtr labelLayer_; + /// weight layer, can be None + LayerPtr weightLayer_; + WeightList weights_; + std::unique_ptr biases_; + std::unique_ptr sampler_; + + std::uniform_int_distribution rand_; + + struct Sample { + int sampleId; + int labelId; + bool target; + real weight; + }; + std::vector samples_; + /// whether samples_ is prepared + bool prepared_; + Argument sampleOut_; + + IVectorPtr labelIds_; + + public: + explicit NCELayer(const LayerConfig& config) + : Layer(config), + numClasses_(config.num_classes()), + rand_(0, config.num_classes() - 1), + prepared_(false) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* initialize the weightList */ + size_t i; + for (i = 0; i < inputLayers_.size(); i++) { + if (!parameters_[i]) break; + size_t width = inputLayers_[i]->getSize(); + // create a new weight + CHECK_EQ(parameters_[i]->getSize(), width * numClasses_); + Weight* w = new Weight(numClasses_, width, parameters_[i]); + + // append the new weight to the list + weights_.emplace_back(w); + } + + CHECK_EQ(1U, getSize()); + + numInputs_ = i; + CHECK_GE(numInputs_, 1) + << "Must have at least one input besides label and weight"; + CHECK_LT(i, inputLayers_.size()) << "Missing label layer"; + labelLayer_ = inputLayers_[i]; + if (++i < inputLayers_.size()) { + weightLayer_ = inputLayers_[i]; + ++i; + } + CHECK_EQ(i, inputLayers_.size()); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + CHECK_EQ(biasParameter_->getSize(), (size_t)numClasses_); + biases_.reset(new Weight(1, numClasses_, biasParameter_)); + } + + if (config_.neg_sampling_dist_size()) { + CHECK_EQ(numClasses_, config_.neg_sampling_dist_size()); + sampler_.reset(MultinomialSampler::create( + config_.neg_sampling_dist().data(), numClasses_)); + } + + return true; + } + + void prepareSamples() { + CHECK(!useGpu_) << "GPU is not supported"; + + int batchSize = getInput(*labelLayer_).getBatchSize(); + IVectorPtr label = getInput(*labelLayer_).ids; + + CpuSparseMatrixPtr multiLabel = std::dynamic_pointer_cast( + getInput(*labelLayer_).value); + + CHECK(label || multiLabel) + << "The label layer must have ids or NonValueSparseMatrix value"; + + auto& randEngine = ThreadLocalRandomEngine::get(); + + samples_.clear(); + samples_.reserve(batchSize * (1 + config_.num_neg_samples())); + + real* weight = + weightLayer_ ? getInputValue(*weightLayer_)->getData() : nullptr; + + for (int i = 0; i < batchSize; ++i) { + real w = weight ? weight[i] : 1; + if (label) { + int* ids = label->getData(); + samples_.push_back({i, ids[i], true, w}); + } else { + const int* cols = multiLabel->getRowCols(i); + int n = multiLabel->getColNum(i); + for (int j = 0; j < n; ++j) { + samples_.push_back({i, cols[j], true, w}); + } + } + for (int j = 0; j < config_.num_neg_samples(); ++j) { + int id = sampler_ ? sampler_->gen(randEngine) : rand_(randEngine); + samples_.push_back({i, id, false, w}); + } + } + prepared_ = true; + } + + void prefetch() override { + prepareSamples(); + IVector::resizeOrCreate(labelIds_, samples_.size(), useGpu_); + int* ids = labelIds_->getData(); + for (size_t i = 0; i < samples_.size(); ++i) { + ids[i] = samples_[i].labelId; + } + + for (int i = 0; i < numInputs_; ++i) { + auto sparseParam = + dynamic_cast(weights_[i]->getW().get()); + if (sparseParam) { + sparseParam->addRows(labelIds_); + } + } + } + + void forward(PassType passType) override { + Layer::forward(passType); + + CHECK(!useGpu_) << "GPU is not supported"; + + if (!prepared_) { + if (passType == PASS_GC) { + ThreadLocalRandomEngine::get().seed(ThreadLocalRand::getDefaultSeed()); + } + prepareSamples(); + } + prepared_ = false; + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(0)->getHeight(); + int size = getSize(); + resetOutput(batchSize, size); + + Matrix::resizeOrCreate(sampleOut_.value, + 1, + samples_.size(), + /* trans= */ false, + useGpu_); + + forwardBias(); + + for (int l = 0; l < numInputs_; ++l) { + forwardOneInput(l); + } + + auto status = activation_->forward(sampleOut_); + status.check(); + + forwardCost(); + } + + void backward(const UpdateCallback& callback) override { + Matrix::resizeOrCreate(sampleOut_.grad, + 1, + samples_.size(), + /* trans= */ false, + useGpu_); + + backwardCost(); + + auto status = activation_->backward(sampleOut_); + status.check(); + + if (biases_->getWGrad()) { + backwardBias(callback); + } + + for (int l = 0; l < numInputs_; ++l) { + backwardOneInput(l, callback); + } + } + + void forwardBias() { + if (!biases_) { + sampleOut_.value->zeroMem(); + } else { + real* bias = biases_->getW()->getData(); + real* sampleOut = sampleOut_.value->getData(); + for (size_t i = 0; i < samples_.size(); ++i) { + sampleOut[i] = bias[samples_[i].labelId]; + } + } + } + + void backwardBias(const UpdateCallback& callback) { + if (!biases_) return; + real* bias = biases_->getWGrad()->getData(); + real* sampleOut = sampleOut_.grad->getData(); + for (size_t i = 0; i < samples_.size(); ++i) { + bias[samples_[i].labelId] += sampleOut[i]; + } + biases_->incUpdate(callback); + } + + void forwardOneInput(int layerId) { + const MatrixPtr& inputMat = getInputValue(layerId); + const MatrixPtr& weightMat = weights_[layerId]->getW(); + + int dim = inputMat->getWidth(); + real* sampleOut = sampleOut_.value->getData(); + + for (size_t i = 0; i < samples_.size(); ++i) { + sampleOut[i] += dotProduct(dim, + inputMat->getRowBuf(samples_[i].sampleId), + weightMat->getRowBuf(samples_[i].labelId)); + } + } + + void backwardOneInput(int layerId, const UpdateCallback& callback) { + const MatrixPtr& inputMat = getInputValue(layerId); + const MatrixPtr& inputGradMat = getInputGrad(layerId); + const MatrixPtr& weightMat = weights_[layerId]->getW(); + const MatrixPtr& weightGradMat = weights_[layerId]->getWGrad(); + + int dim = inputMat->getWidth(); + real* sampleGrad = sampleOut_.grad->getData(); + + if (weightGradMat) { + for (size_t i = 0; i < samples_.size(); ++i) { + axpy(dim, + sampleGrad[i], + inputMat->getRowBuf(samples_[i].sampleId), + weightGradMat->getRowBuf(samples_[i].labelId)); + } + weights_[layerId]->incUpdate(callback); + } + + if (inputGradMat) { + for (size_t i = 0; i < samples_.size(); ++i) { + axpy(dim, + sampleGrad[i], + weightMat->getRowBuf(samples_[i].labelId), + inputGradMat->getRowBuf(samples_[i].sampleId)); + } + } + } + + void forwardCost() { + real* out = output_.value->getData(); + real* sampleOut = sampleOut_.value->getData(); + real b = 1. / numClasses_ * config_.num_neg_samples(); + for (size_t i = 0; i < samples_.size(); ++i) { + real o = sampleOut[i]; + if (sampler_) { + b = config_.num_neg_samples() * + config_.neg_sampling_dist(samples_[i].labelId); + } + real cost = samples_[i].target ? -log(o / (o + b)) : -log(b / (o + b)); + out[samples_[i].sampleId] += samples_[i].weight * cost; + } + } + + void backwardCost() { + real* sampleOut = sampleOut_.value->getData(); + real* sampleGrad = sampleOut_.grad->getData(); + + real b = 1. / numClasses_ * config_.num_neg_samples(); + for (size_t i = 0; i < samples_.size(); ++i) { + real o = sampleOut[i]; + if (sampler_) { + b = config_.num_neg_samples() * + config_.neg_sampling_dist(samples_[i].labelId); + } + real w = samples_[i].weight; + sampleGrad[i] = samples_[i].target ? -w * b / (o * (o + b)) : w / (o + b); + } + } +}; + +REGISTER_LAYER(nce, NCELayer); + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/NormLayer.cpp b/paddle/legacy/gserver/layers/NormLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..443e26dbc859b1c51c5fb93077178ac45bdeaff3 --- /dev/null +++ b/paddle/legacy/gserver/layers/NormLayer.cpp @@ -0,0 +1,59 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "NormLayer.h" +#include "NormProjectionLayer.h" +#include "paddle/legacy/utils/Logging.h" +namespace paddle { + +REGISTER_LAYER_CREATE_FUNC(norm, &NormLayer::create); + +Layer* NormLayer::create(const LayerConfig& config) { + CHECK_EQ(config.inputs_size(), 1); + const std::string& norm = config.inputs(0).norm_conf().norm_type(); + if (norm == "rnorm") { + return new ResponseNormLayer(config); + } else if (norm == "cmrnorm-projection") { + return new CMRProjectionNormLayer(config); + } else if (norm == "cross-channel-norm") { + return new CrossChannelNormLayer(config); + } else { + LOG(FATAL) << "Unknown norm type: " << norm; + return nullptr; + } +} + +bool ResponseNormLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + NormLayer::init(layerMap, parameterMap); + + /* the size of inputs for norm-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + + const NormConfig& conf = config_.inputs(0).norm_conf(); + channels_ = conf.channels(); + size_ = conf.size(); + scale_ = conf.scale(); + pow_ = conf.pow(); + outputX_ = conf.output_x(); + imgSize_ = conf.img_size(); + denoms_ = NULL; + + outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + return true; +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/NormLayer.h b/paddle/legacy/gserver/layers/NormLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..5ac00034d086a5952b30576268c72af326e3ebf9 --- /dev/null +++ b/paddle/legacy/gserver/layers/NormLayer.h @@ -0,0 +1,99 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "NormLayer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief Basic parent layer of normalization + * + * @note Normalize the input in local region + */ +class NormLayer : public Layer { + public: + explicit NormLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override { + Layer::init(layerMap, parameterMap); + return true; + } + + /** + * @brief create norm layer by norm_type + */ + static Layer* create(const LayerConfig& config); +}; + +/** + * @brief response normalization within feature maps + * namely normalize in independent channel + * When code refactoring, we delete the original implementation. + * Need to implement in the futrue. + */ +class ResponseNormLayer : public NormLayer { + protected: + size_t channels_, size_, outputX_, imgSize_, outputY_, imgSizeY_; + real scale_, pow_; + MatrixPtr denoms_; + + public: + explicit ResponseNormLayer(const LayerConfig& config) : NormLayer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override { LOG(FATAL) << "Not implemented"; } + void backward(const UpdateCallback& callback = nullptr) override { + LOG(FATAL) << "Not implemented"; + } +}; + +/** + * This layer applys normalization across the channels of each sample to a + * conv layer's output, and scales the output by a group of trainable factors + * whose dimensions equal to the number of channels. + * - Input: One and only one input layer are accepted. + * - Output: The normalized data of the input data. + * Reference: + * Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, + * Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector + */ +class CrossChannelNormLayer : public NormLayer { + public: + explicit CrossChannelNormLayer(const LayerConfig& config) + : NormLayer(config) {} + bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + void forward(PassType passType); + void backward(const UpdateCallback& callback); + MatrixPtr createSampleMatrix(MatrixPtr data, size_t iter, size_t spatialDim); + MatrixPtr createSpatialMatrix(MatrixPtr data, size_t iter, size_t spatialDim); + + protected: + size_t channels_; + std::unique_ptr scale_; + MatrixPtr scaleDiff_; + MatrixPtr normBuffer_; + MatrixPtr dataBuffer_; + MatrixPtr channelBuffer_; + MatrixPtr spatialBuffer_; + MatrixPtr sampleBuffer_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/NormProjectionLayer.cpp b/paddle/legacy/gserver/layers/NormProjectionLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..72affaa1ce618a841f8040c84467a46b77531958 --- /dev/null +++ b/paddle/legacy/gserver/layers/NormProjectionLayer.cpp @@ -0,0 +1,101 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "NormProjectionLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { +size_t CMRProjectionNormLayer::getSize() { + CHECK_EQ(inputLayers_.size(), 1UL); + size_t layerSize = 0; + imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight(); + imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth(); + if (imgSizeH_ == 0) { + imgSizeH_ = imgSizeY_; + } + if (imgSizeW_ == 0) { + imgSizeW_ = imgSize_; + } + outputH_ = imgSizeH_; + outputW_ = imgSizeW_; + layerSize = outputH_ * outputW_ * channels_; + + getOutput().setFrameHeight(outputH_); + getOutput().setFrameWidth(outputW_); + return layerSize; +} + +bool CMRProjectionNormLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + ResponseNormLayer::init(layerMap, parameterMap); + + /* the size of inputs for norm-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + + createFunction( + forward_, + "CrossMapNormal", + FuncConfig().set("size", size_).set("scale", scale_).set("pow", pow_)); + createFunction( + backward_, + "CrossMapNormalGrad", + FuncConfig().set("size", size_).set("scale", scale_).set("pow", pow_)); + + return true; +} + +void CMRProjectionNormLayer::forward(PassType passType) { + Layer::forward(passType); + /* malloc memory for the output_ if necessary */ + /* note: one sample correspond to one row */ + MatrixPtr input = inputLayers_[0]->getOutputValue(); + size_t batchSize = input->getHeight(); + int size = getSize(); + resetOutput(batchSize, size); + + Matrix::resizeOrCreate(denoms_, batchSize, size, /* trans */ false, useGpu_); + + shape_ = TensorShape({batchSize, channels_, imgSizeH_, imgSizeW_}); + + // prepare forward arguments + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), shape_); + outputs.addArg(*getOutputValue(), shape_, ASSIGN_TO); + outputs.addArg(*denoms_, shape_, ASSIGN_TO); + + forward_[0]->calc(inputs, outputs); +} + +void CMRProjectionNormLayer::backward(const UpdateCallback& callback) { + (void)callback; + + if (NULL == getInputGrad(0)) { + return; + } + + // prepare backward arguments + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), shape_); + inputs.addArg(*getOutputValue(), shape_); + inputs.addArg(*getOutputGrad(), shape_); + inputs.addArg(*denoms_, shape_); + outputs.addArg(*getInputGrad(0), shape_, ADD_TO); + + backward_[0]->calc(inputs, outputs); +} +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/NormProjectionLayer.h b/paddle/legacy/gserver/layers/NormProjectionLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..492d1fcb72343a54577a459aaa5de53596f43f42 --- /dev/null +++ b/paddle/legacy/gserver/layers/NormProjectionLayer.h @@ -0,0 +1,47 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "NormLayer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief response normalization across feature maps + * namely normalize in number of size_ channels + */ +class CMRProjectionNormLayer : public ResponseNormLayer { + size_t imgSizeH_, imgSizeW_; + size_t outputH_, outputW_; + + public: + explicit CMRProjectionNormLayer(const LayerConfig& config) + : ResponseNormLayer(config) {} + + ~CMRProjectionNormLayer() {} + + size_t getSize(); + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + protected: + TensorShape shape_; +}; +} // namespace paddle diff --git a/paddle/gserver/layers/Operator.cpp b/paddle/legacy/gserver/layers/Operator.cpp similarity index 100% rename from paddle/gserver/layers/Operator.cpp rename to paddle/legacy/gserver/layers/Operator.cpp diff --git a/paddle/legacy/gserver/layers/Operator.h b/paddle/legacy/gserver/layers/Operator.h new file mode 100644 index 0000000000000000000000000000000000000000..20a248985eb6b3aba016b28bca4c0eea44baa868 --- /dev/null +++ b/paddle/legacy/gserver/layers/Operator.h @@ -0,0 +1,96 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ModelConfig.pb.h" +#include "paddle/legacy/parameter/Parameter.h" + +#include "Layer.h" +#include "paddle/legacy/parameter/Argument.h" + +namespace paddle { + +// Macro for registering a operator type +// Example: REGISTER_OPERATOR(dot_mul, DotMulOperator); +#define REGISTER_OPERATOR(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name([]() { \ + Operator::registrar_.registerClass<__class_name>(#__type_name); \ + }) + +/** + * Operator like Projection, but takes more than one Arguments as input. + * @note: Operator can't have parameters. + */ +class Operator { + public: + static Operator* create(const OperatorConfig& config, bool useGpu); + + Operator(const OperatorConfig& config, bool useGpu) + : config_(config), useGpu_(useGpu) {} + + virtual ~Operator() {} + + const OperatorConfig& getConfig() const { return config_; } + + static ClassRegistrar registrar_; + + /** + * Forward propagation. If backward() will be called, in and out must be kept + * valid until then. + * @param ins inputs of operator + * @param out output of operator + * @param passType PASS_TRAIN of PASS_TEST + */ + void forward(std::vector ins, + Argument* out, + PassType passType) { + ins_ = ins; + out_ = out; + passType_ = passType; + forward(); + } + + virtual void prefetch(const Argument* in) {} + virtual void forward() = 0; + virtual void backward() = 0; + + /** + * See comment in Layer.h for the function with the same name. + */ + virtual void resetState() {} + + /** + * Set layer state. + */ + virtual void setState(LayerStatePtr state) {} + + /** + * Set layer state. + */ + virtual LayerStatePtr getState() { return nullptr; } + + protected: + /// Config of operator + OperatorConfig config_; + bool useGpu_; + + /// Store `ins` passed to forward() + std::vector ins_; + /// Store `out` passed to forward() + Argument* out_; + /// Store `passType` passed to forward() + PassType passType_; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/OuterProdLayer.cpp b/paddle/legacy/gserver/layers/OuterProdLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d0928be9d4d52532503987af8e29fdf5c7fb16a5 --- /dev/null +++ b/paddle/legacy/gserver/layers/OuterProdLayer.cpp @@ -0,0 +1,141 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief A layer for computing the outer product of two vectors + * @note used in NEURAL TURING MACHINE + * Input1: vector (batchSize * dim1) + * Input2: vector (batchSize * dim2) + * Output: a matrix: (batchSize * (dim1*dim2)) + */ + +class OuterProdLayer : public Layer { + protected: + MatrixPtr tmpMtx0; + MatrixPtr tmpRow0; + MatrixPtr tmpRow1; + + public: + explicit OuterProdLayer(const LayerConfig& config) : Layer(config) {} + + ~OuterProdLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(out_prod, OuterProdLayer); + +bool OuterProdLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2U); + + size_t dim0 = inputLayers_[0]->getSize(); + size_t dim1 = inputLayers_[1]->getSize(); + + CHECK_EQ(dim0 * dim1, getSize()) << "Dimension mismatch"; + + tmpRow0 = Matrix::create( + nullptr, /* height= */ 1, dim0, /* trans= */ false, useGpu_); + tmpRow1 = Matrix::create( + nullptr, /* height= */ 1, dim1, /* trans= */ false, useGpu_); + tmpMtx0 = Matrix::create(nullptr, + /* height= */ dim0, + dim1, + /* trans= */ false, + useGpu_); + return true; +} + +void OuterProdLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV0->getHeight(); + size_t dim0 = inV0->getWidth(); + size_t dim1 = inV1->getWidth(); + + CHECK_EQ(dim0 * dim1, getSize()); + CHECK_EQ(inV1->getHeight(), batchSize); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, dim0 * dim1); + } + + MatrixPtr outV = getOutputValue(); + + { + REGISTER_TIMER_INFO("FwOutProdTimer", getName().c_str()); + for (size_t i = 0; i < batchSize; i++) { + tmpMtx0->setData(outV->getData() + i * dim0 * dim1); + tmpRow0->setData(inV0->getData() + i * dim0); + tmpRow1->setData(inV1->getData() + i * dim1); + + tmpMtx0->mul(*tmpRow0->getTranspose(), *tmpRow1); + } + } +} + +void OuterProdLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr outG = getOutputGrad(); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + + size_t batchSize = inV0->getHeight(); + size_t dim0 = inV0->getWidth(); + size_t dim1 = inV1->getWidth(); + + { + REGISTER_TIMER_INFO("BwOutProdTimer", getName().c_str()); + + if (inG0) { + for (size_t i = 0; i < batchSize; i++) { + tmpMtx0->setData(outG->getData() + i * dim0 * dim1); + tmpRow0->setData(inG0->getData() + i * dim0); + tmpRow1->setData(inV1->getData() + i * dim1); + + tmpRow0->mul(*tmpRow1, *tmpMtx0->getTranspose(), 1, 1); + } + } + + if (inG1) { + for (size_t i = 0; i < batchSize; i++) { + tmpMtx0->setData(outG->getData() + i * dim0 * dim1); + tmpRow0->setData(inV0->getData() + i * dim0); + tmpRow1->setData(inG1->getData() + i * dim1); + + tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 1); + } + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/PadLayer.cpp b/paddle/legacy/gserver/layers/PadLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7b92b3de2d839f240ec8cbe07ed7685295568809 --- /dev/null +++ b/paddle/legacy/gserver/layers/PadLayer.cpp @@ -0,0 +1,106 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PadLayer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(pad, PadLayer); + +bool PadLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + auto& pad_conf = config_.inputs(0).pad_conf(); + auto& img_conf = pad_conf.image_conf(); + CHECK_EQ(config_.inputs_size(), 1); + inDims_ = TensorShape( + {0, + img_conf.channels(), + img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(), + img_conf.img_size()}); + + CHECK_EQ(2, pad_conf.pad_c_size()); + CHECK_EQ(2, pad_conf.pad_h_size()); + CHECK_EQ(2, pad_conf.pad_w_size()); + padc_ = {pad_conf.pad_c(0), pad_conf.pad_c(1)}; + padh_ = {pad_conf.pad_h(0), pad_conf.pad_h(1)}; + padw_ = {pad_conf.pad_w(0), pad_conf.pad_w(1)}; + + outDims_ = TensorShape(4); + setOutDims(0); + + createFunction(forward_, + "Pad", + FuncConfig() + .set("channel", padc_) + .set("height", padh_) + .set("width", padw_)); + createFunction(backward_, + "PadGrad", + FuncConfig() + .set("channel", padc_) + .set("height", padh_) + .set("width", padw_)); + + return true; +} + +void PadLayer::setOutDims(const size_t batchSize) { + outDims_.reshape({batchSize, + inDims_[1] + padc_[0] + padc_[1], + inDims_[2] + padh_[0] + padh_[1], + inDims_[3] + padw_[0] + padw_[1]}); +} + +void PadLayer::setTensorDim(const size_t batchSize) { + CHECK_EQ(static_cast(inputLayers_.size()), 1); + inDims_.setDim(0, batchSize); + int h = inputLayers_[0]->getOutput().getFrameHeight(); + if (h != 0) inDims_.setDim(2, h); + int w = inputLayers_[0]->getOutput().getFrameWidth(); + if (w != 0) inDims_.setDim(3, w); + setOutDims(batchSize); +} + +void PadLayer::forward(PassType passType) { + Layer::forward(passType); + MatrixPtr input = inputLayers_[0]->getOutputValue(); + size_t batchSize = input->getHeight(); + setTensorDim(batchSize); + int size = outDims_[1] * outDims_[2] * outDims_[3]; + resetOutput(batchSize, size); + MatrixPtr outV = getOutputValue(); + REGISTER_TIMER_INFO("PadForward", getName().c_str()); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), inDims_); + outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO); + forward_[0]->calc(inputs, outputs); +} + +void PadLayer::backward(const UpdateCallback& callback) { + (void)callback; + REGISTER_TIMER_INFO("PadBackward", getName().c_str()); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outDims_); + outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); + backward_[0]->calc(inputs, outputs); +} +} // namespace paddle diff --git a/paddle/gserver/layers/PadLayer.h b/paddle/legacy/gserver/layers/PadLayer.h similarity index 100% rename from paddle/gserver/layers/PadLayer.h rename to paddle/legacy/gserver/layers/PadLayer.h diff --git a/paddle/legacy/gserver/layers/ParameterReluLayer.cpp b/paddle/legacy/gserver/layers/ParameterReluLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..23715d1975d7a3606a9418d54bc69ae6f036a93a --- /dev/null +++ b/paddle/legacy/gserver/layers/ParameterReluLayer.cpp @@ -0,0 +1,69 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ParameterReluLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(prelu, ParameterReluLayer); + +bool ParameterReluLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_EQ(inputLayers_.size(), 1UL); + CHECK_EQ(inputLayers_.size(), parameters_.size()); + partialSum_ = config_.partial_sum(); + CHECK_GT(partialSum_, 0UL) << "partial_sum must be larger than zero."; + CHECK(!(inputLayers_[0]->getSize() % partialSum_)) + << "Incorrect value for partialSum: " << partialSum_ + << " must divide input size: " << inputLayers_[0]->getSize(); + CHECK_EQ(getSize() / partialSum_, parameters_[0]->getSize()); + weight_ = std::unique_ptr(new Weight( + 1UL, inputLayers_[0]->getSize() / partialSum_, parameters_[0])); + return true; +} + +void ParameterReluLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInput(0).getBatchSize(); + int size = getSize(); + reserveOutput(batchSize, size); + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + outV->paramReluForward(*(getInput(0).value), *(weight_->getW())); + } +} + +void ParameterReluLayer::backward(const UpdateCallback& callback) { + if (weight_->getWGrad()) { + weight_->getWGrad()->paramReluBackwardW(*getOutputGrad(), + *(getInputValue(0))); + } + + MatrixPtr preGrad = getInputGrad(0); + preGrad->paramReluBackwardDiff( + *getOutputGrad(), *(getInputValue(0)), *(weight_->getW())); + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + weight_->getParameterPtr()->incUpdate(callback); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ParameterReluLayer.h b/paddle/legacy/gserver/layers/ParameterReluLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..3aac4b42f60531b5856ddef208b8356898e42859 --- /dev/null +++ b/paddle/legacy/gserver/layers/ParameterReluLayer.h @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * @brief ParameterReluLayer active inputs with learnable parameter weight_. + * forward: + * \f[ + * y = x > 0 ? x : w .* x + * \f] + * backward: + * \f[ + * dx = x > 0 ? dy : w .* dy \\ + * dw = x > 0 ? 0 : dy.*x + * \f] + * Here, x is the input, w is the weight, y is the output. + * dx, dw, dy is the gradient. + */ + +class ParameterReluLayer : public Layer { + protected: + std::unique_ptr weight_; + + /** + * @brief partialSum_ makes a group of inputs share same weights, + * - partialSum_ = 1: + * element wise activation: each element has a weight_, + * - partialSum_ = number of elements in one channel, + * channels wise parameter activation, elements in a channel + * share same weight_, + * - partialSum_ = number of outputs + * all elements share same weight_, + */ + size_t partialSum_; + + public: + explicit ParameterReluLayer(const LayerConfig& config) : Layer(config) {} + + ~ParameterReluLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/Pool3DLayer.cpp b/paddle/legacy/gserver/layers/Pool3DLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ae3f55c27f2d7bd3ab47d834d5b6f274ff558310 --- /dev/null +++ b/paddle/legacy/gserver/layers/Pool3DLayer.cpp @@ -0,0 +1,178 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Pool3DLayer.h" +#include "PoolProjectionLayer.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +REGISTER_LAYER(pool3d, Pool3DLayer); + +bool Pool3DLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + /* the size of inputs for pool-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + + const PoolConfig& conf = config_.inputs(0).pool_conf(); + poolType_ = conf.pool_type(); + channels_ = conf.channels(); + + sizeX_ = conf.size_x(); + sizeY_ = conf.size_y(); + sizeZ_ = conf.size_z(); + + strideW_ = conf.stride(); + strideH_ = conf.stride_y(); + strideD_ = conf.stride_z(); + + imgSizeW_ = conf.img_size(); + imgSizeH_ = conf.img_size_y(); + imgSizeD_ = conf.img_size_z(); + + paddingW_ = conf.padding(); + paddingH_ = conf.padding_y(); + paddingD_ = conf.padding_z(); + + outputW_ = conf.output_x(); + outputH_ = conf.output_y(); + outputD_ = conf.output_z(); + + return true; +} + +size_t Pool3DLayer::getSize() { + CHECK_EQ(inputLayers_.size(), 1UL); + + size_t layerSize = 0; + outputD_ = outputSize(imgSizeD_, sizeZ_, paddingD_, strideD_, false); + outputH_ = outputSize(imgSizeH_, sizeY_, paddingH_, strideH_, false); + outputW_ = outputSize(imgSizeW_, sizeX_, paddingW_, strideW_, false); + + layerSize = outputD_ * outputH_ * outputW_ * channels_; + getOutput().setFrameHeight(outputH_); + getOutput().setFrameWidth(outputW_); + getOutput().setFrameDepth(outputD_); + return layerSize; +} + +void Pool3DLayer::forward(PassType passType) { + Layer::forward(passType); + const MatrixPtr& inMat = inputLayers_[0]->getOutputValue(); + size_t batchSize = inMat->getHeight(); + size_t outWidth = getSize(); + resetOutput(batchSize, outWidth); + Matrix::resizeOrCreate(maxPoolIdx_, batchSize, outWidth, false, useGpu_); + const MatrixPtr outMat = getOutputValue(); + + if (poolType_ == "avg") { + outMat->avgPool3DForward(*inMat, + channels_, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeX_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_); + } else if (poolType_ == "max") { + outMat->maxPool3DForward(*inMat, + *maxPoolIdx_, + channels_, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeX_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_); + } else { + LOG(FATAL) << "Unknown pool type: " << poolType_; + } + forwardActivation(); +} + +void Pool3DLayer::backward(const UpdateCallback& callback) { + backwardActivation(); + + (void)callback; + if (NULL == getInputGrad(0)) return; + MatrixPtr inMat = inputLayers_[0]->getOutputValue(); + MatrixPtr inGradMat = inputLayers_[0]->getOutputGrad(); + MatrixPtr outMat = getOutputValue(); + MatrixPtr outGradMat = getOutputGrad(); + + if (poolType_ == "avg") { + inGradMat->avgPool3DBackward(*outGradMat, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeZ_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_, + 1.0, + 1.0); + } else if (poolType_ == "max") { + inGradMat->maxPool3DBackward(*outGradMat, + *maxPoolIdx_, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeZ_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_, + 1.0, + 1.0); + } else { + LOG(FATAL) << "Unknown pool type: " << poolType_; + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/Pool3DLayer.h b/paddle/legacy/gserver/layers/Pool3DLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..6851c44ab22a39bebe3592b8e5f6384a393947f2 --- /dev/null +++ b/paddle/legacy/gserver/layers/Pool3DLayer.h @@ -0,0 +1,49 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief Basic parent layer of pooling + * Pools the input within regions + */ +class Pool3DLayer : public Layer { + public: + explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {} + ~Pool3DLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + size_t getSize(); + + protected: + int channels_; + int sizeX_, sizeY_, sizeZ_; + int strideW_, strideH_, strideD_; + int paddingW_, paddingH_, paddingD_; + int imgSizeW_, imgSizeH_, imgSizeD_; + int outputW_, outputH_, outputD_; + std::string poolType_; + MatrixPtr maxPoolIdx_; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/PoolLayer.cpp b/paddle/legacy/gserver/layers/PoolLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..df172d95757e0842328caa508042f3613bc72232 --- /dev/null +++ b/paddle/legacy/gserver/layers/PoolLayer.cpp @@ -0,0 +1,70 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PoolLayer.h" +#include "MaxPoolWithMaskLayer.h" +#include "PoolProjectionLayer.h" +#include "paddle/legacy/utils/Logging.h" +#ifdef PADDLE_WITH_CUDA +#include "CudnnPoolLayer.h" +#endif +namespace paddle { + +REGISTER_LAYER_CREATE_FUNC(pool, &PoolLayer::create); + +bool PoolLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* the size of inputs for pool-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + + const PoolConfig& conf = config_.inputs(0).pool_conf(); + poolType_ = conf.pool_type(); + channels_ = conf.channels(); + sizeX_ = conf.size_x(); + stride_ = conf.stride(); + outputX_ = conf.output_x(); + imgSize_ = conf.img_size(); + confPadding_ = conf.padding(); + + sizeY_ = conf.has_size_y() ? conf.size_y() : conf.size_x(); + imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride(); + confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding(); + outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + + excludeMode_ = conf.has_exclude_mode() ? conf.exclude_mode() : true; + return true; +} + +Layer* PoolLayer::create(const LayerConfig& config) { + CHECK_EQ(config.inputs_size(), 1); + const std::string& pool = config.inputs(0).pool_conf().pool_type(); + if (pool == "max-projection" || pool == "avg-projection") { + return new PoolProjectionLayer(config); +#ifdef PADDLE_WITH_CUDA + } else if (CudnnPoolLayer::typeCheck(pool)) { + return new CudnnPoolLayer(config); +#endif + } else if (pool == "max-pool-with-mask") { + return new MaxPoolWithMaskLayer(config); + } else { + LOG(FATAL) << "Unknown pool type: " << pool; + return nullptr; + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/PoolLayer.h b/paddle/legacy/gserver/layers/PoolLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..0808dfae8497008f974730b65977c85e914a7a27 --- /dev/null +++ b/paddle/legacy/gserver/layers/PoolLayer.h @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { + +/** + * @brief Basic parent layer of pooling + * Pools the input within regions + */ +class PoolLayer : public Layer { + protected: + size_t channels_, sizeX_, stride_, outputX_, imgSize_; + int confPadding_; + + size_t sizeY_; + size_t imgSizeY_; + size_t strideY_; + size_t outputY_; + int confPaddingY_; + + std::string poolType_; + + bool excludeMode_; + + public: + explicit PoolLayer(const LayerConfig& config) : Layer(config) {} + + /** + * @brief create pooling layer by pool_type + */ + static Layer* create(const LayerConfig& config); + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/PoolProjection.cpp b/paddle/legacy/gserver/layers/PoolProjection.cpp similarity index 100% rename from paddle/gserver/layers/PoolProjection.cpp rename to paddle/legacy/gserver/layers/PoolProjection.cpp diff --git a/paddle/legacy/gserver/layers/PoolProjection.h b/paddle/legacy/gserver/layers/PoolProjection.h new file mode 100644 index 0000000000000000000000000000000000000000..d01b6a13f0a5fd2283f1f216ef419b9ccc7308f9 --- /dev/null +++ b/paddle/legacy/gserver/layers/PoolProjection.h @@ -0,0 +1,68 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Projection.h" +#include "paddle/legacy/math/MathUtils.h" + +namespace paddle { + +class PoolProjection : public Projection { + protected: + size_t imgSizeY_, imgSize_; + size_t outputY_, outputX_; + size_t strideY_, stride_; + size_t sizeY_, sizeX_; + int confPaddingY_, confPadding_; + size_t channels_; + std::string poolType_; + bool excludeMode_; + + public: + PoolProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu); + + static PoolProjection* create(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu); + + const std::string& getPoolType() const { return poolType_; } + + size_t getSize(); +}; + +class MaxPoolProjection : public PoolProjection { + public: + MaxPoolProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu) + : PoolProjection(config, parameter, useGpu) {} + + virtual void forward(); + virtual void backward(const UpdateCallback& callback = nullptr); +}; + +class AvgPoolProjection : public PoolProjection { + public: + AvgPoolProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu) + : PoolProjection(config, parameter, useGpu) {} + + virtual void forward(); + virtual void backward(const UpdateCallback& callback = nullptr); +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/PoolProjectionLayer.cpp b/paddle/legacy/gserver/layers/PoolProjectionLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e44b1d7ba1494e43db81f998c2818bbbf7779d6f --- /dev/null +++ b/paddle/legacy/gserver/layers/PoolProjectionLayer.cpp @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PoolProjectionLayer.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +size_t PoolProjectionLayer::getSize() { + CHECK_EQ(inputLayers_.size(), 1UL); + size_t layerSize = 0; + imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight(); + imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth(); + if (imgSizeH_ == 0) { + imgSizeH_ = imgSizeY_; + } + if (imgSizeW_ == 0) { + imgSizeW_ = imgSize_; + } + + outputH_ = outputSize(imgSizeH_, + sizeY_, + confPaddingY_, + strideY_, + /* caffeMode */ false); + outputW_ = outputSize(imgSizeW_, + sizeX_, + confPadding_, + stride_, + /* caffeMode */ false); + + layerSize = outputH_ * outputW_ * channels_; + + return layerSize; +} + +void PoolProjectionLayer::forward(PassType passType) { + Layer::forward(passType); + const Argument& in = getInput(0); + int batchSize = in.value->getHeight(); + int size = getSize(); + resetOutput(batchSize, size); + poolProjection_->forward(&in, &output_, passType); +} + +void PoolProjectionLayer::backward(const UpdateCallback& callback) { + (void)callback; + if (NULL == getInputGrad(0)) { + return; + } + poolProjection_->backward(callback); +} +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/PoolProjectionLayer.h b/paddle/legacy/gserver/layers/PoolProjectionLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..fcd35bbba4dff612fba827cdf545de71127c560e --- /dev/null +++ b/paddle/legacy/gserver/layers/PoolProjectionLayer.h @@ -0,0 +1,46 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "PoolLayer.h" +#include "PoolProjection.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * @brief Basic parent layer of different kinds of pooling + */ +class PoolProjectionLayer : public PoolLayer { + protected: + size_t imgSizeH_, imgSizeW_; + size_t outputH_, outputW_; + std::unique_ptr poolProjection_; + ProjectionConfig projectionConfig_; + + public: + explicit PoolProjectionLayer(const LayerConfig& config) : PoolLayer(config) { + PoolConfig* conf = projectionConfig_.mutable_pool_conf(); + *conf = config_.inputs(0).pool_conf(); + poolProjection_.reset( + PoolProjection::create(projectionConfig_, nullptr, useGpu_)); + } + + size_t getSize(); + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/PowerLayer.cpp b/paddle/legacy/gserver/layers/PowerLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5e94c64db6098dbc1ed13bdcbd573f95024713bc --- /dev/null +++ b/paddle/legacy/gserver/layers/PowerLayer.cpp @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * This layer applys a power function to a vector element-wise, + * which is used in NEURAL TURING MACHINE. + * \f[ + * y = x^w + * \f] + * where \f$x\f$ is a input vector, \f$w\f$ is scalar weight, + * and output \f$y\f$ is a vector. + * + * The config file api is power_layer. + */ + +class PowerLayer : public Layer { + protected: + MatrixPtr tmpMtx; + + public: + explicit PowerLayer(const LayerConfig& config) : Layer(config) {} + + ~PowerLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(power, PowerLayer); + +bool PowerLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2U); + + return true; +} + +void PowerLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV1->getHeight(); + size_t dataDim = inV1->getWidth(); + + CHECK_EQ(getSize(), dataDim); + CHECK_EQ(1U, inV0->getWidth()); + CHECK_EQ(batchSize, inV0->getHeight()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, dataDim); + } + + MatrixPtr outV = getOutputValue(); + + { + REGISTER_TIMER_INFO("FwPowerTimer", getName().c_str()); + outV->rowPow(0, *inV1, *inV0); + } +} + +void PowerLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV0 = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + MatrixPtr outV = getOutputValue(); + MatrixPtr outG = getOutputGrad(); + + size_t batchSize = inV1->getHeight(); + size_t dataDim = inV1->getWidth(); + + { + REGISTER_TIMER_INFO("BwPowerTimer", getName().c_str()); + Matrix::resizeOrCreate(tmpMtx, batchSize, dataDim, false, useGpu_); + + if (inG0) { + tmpMtx->log2(*inV1); + tmpMtx->dotMul(*tmpMtx, *outV); + + // inG0 += outG .* (log(inV1) * outV) + inG0->rowDotMul(0, *outG, *tmpMtx); + } + + if (inG1) { + // tmp = (outV / inV1) * inV0 + tmpMtx->dotDiv(*outV, *inV1); + tmpMtx->rowScale(0, *tmpMtx, *inV0); + + inG1->addDotMul(*outG, *tmpMtx, 1, 1); + } + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/PrintLayer.cpp b/paddle/legacy/gserver/layers/PrintLayer.cpp similarity index 100% rename from paddle/gserver/layers/PrintLayer.cpp rename to paddle/legacy/gserver/layers/PrintLayer.cpp diff --git a/paddle/legacy/gserver/layers/PriorBox.cpp b/paddle/legacy/gserver/layers/PriorBox.cpp new file mode 100644 index 0000000000000000000000000000000000000000..83aab6e36662855a5867463757bc5a92e6e83e07 --- /dev/null +++ b/paddle/legacy/gserver/layers/PriorBox.cpp @@ -0,0 +1,159 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/BaseMatrix.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * @brief A layer for generating priorbox locations and variances. + * - Input: Two and only two input layer are accepted. The input layer must be + * be a data output layer and a convolution output layer. + * - Output: The priorbox locations and variances of the input data. + * Reference: + * Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, + * Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector + */ + +class PriorBoxLayer : public Layer { + public: // NOLINT + explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {} + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override {} + + protected: // NOLINT + int numPriors_; + std::vector minSize_; + std::vector maxSize_; + std::vector aspectRatio_; + std::vector variance_; + MatrixPtr buffer_; +}; + +REGISTER_LAYER(priorbox, PriorBoxLayer); + +bool PriorBoxLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + auto pbConf = config_.inputs(0).priorbox_conf(); + std::vector tmp; + aspectRatio_.push_back(1.); + std::copy(pbConf.min_size().begin(), + pbConf.min_size().end(), + std::back_inserter(minSize_)); + std::copy(pbConf.max_size().begin(), + pbConf.max_size().end(), + std::back_inserter(maxSize_)); + std::copy(pbConf.variance().begin(), + pbConf.variance().end(), + std::back_inserter(variance_)); + std::copy(pbConf.aspect_ratio().begin(), + pbConf.aspect_ratio().end(), + std::back_inserter(tmp)); + + if (maxSize_.size() > 0) CHECK_EQ(minSize_.size(), maxSize_.size()); + + // flip aspect ratios + for (unsigned index = 0; index < tmp.size(); index++) { + real ar = tmp[index]; + if (fabs(ar - 1.) < 1e-6) continue; + aspectRatio_.push_back(ar); + aspectRatio_.push_back(1. / ar); + } + + numPriors_ = aspectRatio_.size() * minSize_.size() + maxSize_.size(); + + return true; +} + +void PriorBoxLayer::forward(PassType passType) { + Layer::forward(passType); + auto input = getInput(0); + int layerWidth = input.getFrameWidth(); + int layerHeight = input.getFrameHeight(); + + auto image = getInput(1); + int imageWidth = image.getFrameWidth(); + int imageHeight = image.getFrameHeight(); + + real stepW = static_cast(imageWidth) / layerWidth; + real stepH = static_cast(imageHeight) / layerHeight; + int dim = layerHeight * layerWidth * numPriors_ * 4; + reserveOutput(1, dim * 2); + // use a cpu buffer to compute + Matrix::resizeOrCreate(buffer_, 1, dim * 2, false, false); + auto* tmpPtr = buffer_->getData(); + + int idx = 0; + for (int h = 0; h < layerHeight; ++h) { + for (int w = 0; w < layerWidth; ++w) { + real centerX = (w + 0.5) * stepW; + real centerY = (h + 0.5) * stepH; + for (size_t s = 0; s < minSize_.size(); s++) { + real minSize = minSize_[s]; + real boxWidth = minSize; + real boxHeight = minSize; + + // first prior: aspect_ratio == 1.0, compatible to old logic + tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; + tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; + // set the variance. + for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; + + if (maxSize_.size() > 0) { + // square prior with size sqrt(minSize * maxSize) + real maxSize = maxSize_[s]; + boxWidth = boxHeight = sqrt(minSize * maxSize); + tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; + tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; + // set the variance. + for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; + } + + // priors with different aspect ratios + for (size_t r = 0; r < aspectRatio_.size(); r++) { + real ar = aspectRatio_[r]; + if (fabs(ar - 1.0) < 1e-6) { + continue; + } + boxWidth = minSize * sqrt(ar); + boxHeight = minSize / sqrt(ar); + tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; + tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; + // set the variance. + for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; + } + } + } + } + + // clip the prior's coordidate such that it is within [0, 1] + for (int d = 0; d < dim * 2; ++d) + if ((d % 8) < 4) + tmpPtr[d] = std::min(std::max(tmpPtr[d], (real)0.), (real)1.); + MatrixPtr outV = getOutputValue(); + outV->copyFrom(buffer_->data_, dim * 2); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/Projection.cpp b/paddle/legacy/gserver/layers/Projection.cpp similarity index 100% rename from paddle/gserver/layers/Projection.cpp rename to paddle/legacy/gserver/layers/Projection.cpp diff --git a/paddle/legacy/gserver/layers/Projection.h b/paddle/legacy/gserver/layers/Projection.h new file mode 100644 index 0000000000000000000000000000000000000000..974f5a2cacd10a965adcb4accf6ca00c26044b64 --- /dev/null +++ b/paddle/legacy/gserver/layers/Projection.h @@ -0,0 +1,140 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "ModelConfig.pb.h" +#include "paddle/legacy/parameter/Parameter.h" + +namespace paddle { + +// Macro for registering a projection type +// Example: REGISTER_LAYER(fc, FullMatrixProjection); +#define REGISTER_PROJECTION(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name([]() { \ + Projection::registrar_.registerClass<__class_name>(#__type_name); \ + }) + +#define REGISTER_PROJECTION_CREATE_FUNC(__type_name, createFunction) \ + static InitFunction __reg_type_##__type_name([]() { \ + Projection::registrar_.registerClass(#__type_name, createFunction); \ + }) + +/** + * A projection takes one Argument as input, calculate the result and add it + * to output Argument. + */ +class Projection { + public: + static Projection* create(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu); + + Projection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGpu) + : config_(config), parameter_(parameter), useGpu_(useGpu) {} + + virtual ~Projection() {} + + const std::string& getName() const { return config_.name(); } + + /// Register a projection + static ClassRegistrar + registrar_; + + /** + * Forward propagation. If backward() will be called, in and out must be kept + * valid until then. + * @param in input of projection + * @param out output of projection + * @param passType PASS_TRAIN of PASS_TEST + */ + void forward(const Argument* in, const Argument* out, PassType passType) { + in_ = in; + out_ = out; + passType_ = passType; + forward(); + } + + virtual void prefetch(const Argument* in) {} + virtual void forward() = 0; + virtual void backward(const UpdateCallback& callback) = 0; + + /** + * See comment in Layer.h for the function with the same name. + */ + virtual void resetState() {} + + /** + * Set layer state. + */ + virtual void setState(LayerStatePtr state) {} + + /** + * Get layer state. A copy of internal state is returned. + */ + virtual LayerStatePtr getState() { return nullptr; } + + /** + * init forward_ and backward_ functions + */ + virtual bool init() { return true; } + + /** + * Get output size of projection. + */ + size_t getOutputSize() const { return config_.output_size(); } + + protected: + /** + * Create layer function. Function is called in forward or backward. + * \param function, Layer::forward_ or Layer::backward_ + * \param name, function name + * \param config, initialization configuration for the function + */ + void createFunction(std::vector>& function, + const std::string& name, + const FuncConfig& config) { + if (useGpu_) { + function.emplace_back( + FunctionBase::funcRegistrar_.createByType(name + "-GPU")); + } else { + function.emplace_back( + FunctionBase::funcRegistrar_.createByType(name + "-CPU")); + } + auto& func = function.back(); + func->init(config); + } + + protected: + /// Config of projection + ProjectionConfig config_; + /// Parameter of projection + ParameterPtr parameter_; + bool useGpu_; + + /// Store `in` passed to forward() + const Argument* in_; + /// Store `out` passed to forward() + const Argument* out_; + /// Store `passType` passed to forward() + PassType passType_; + /// Layer forward function + std::vector> forward_; + /// Layer backward function + std::vector> backward_; +}; +} // namespace paddle diff --git a/paddle/gserver/layers/ROIPoolLayer.cpp b/paddle/legacy/gserver/layers/ROIPoolLayer.cpp similarity index 100% rename from paddle/gserver/layers/ROIPoolLayer.cpp rename to paddle/legacy/gserver/layers/ROIPoolLayer.cpp diff --git a/paddle/gserver/layers/ROIPoolLayer.h b/paddle/legacy/gserver/layers/ROIPoolLayer.h similarity index 100% rename from paddle/gserver/layers/ROIPoolLayer.h rename to paddle/legacy/gserver/layers/ROIPoolLayer.h diff --git a/paddle/gserver/layers/RecurrentLayer.cpp b/paddle/legacy/gserver/layers/RecurrentLayer.cpp similarity index 100% rename from paddle/gserver/layers/RecurrentLayer.cpp rename to paddle/legacy/gserver/layers/RecurrentLayer.cpp diff --git a/paddle/legacy/gserver/layers/RecurrentLayer.h b/paddle/legacy/gserver/layers/RecurrentLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..287ea27a0984729fde5b35aa0807e9f2b29f993f --- /dev/null +++ b/paddle/legacy/gserver/layers/RecurrentLayer.h @@ -0,0 +1,130 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#pragma once +#include +#include "Layer.h" +#include "SequenceToBatch.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief RecurrentLayer takes 1 input layer. The output size is the same with + * input layer. + * For each sequence [start, end] it performs the following computation: + * \f[ + * out_{i} = act(in_{i}) \ \ \text{for} \ i = start \\ + * out_{i} = act(in_{i} + out_{i-1} * W) \ \ \text{for} \ start < i <= end + * + * \f] + * If reversed is true, the order is reversed: + * \f[ + * out_{i} = act(in_{i}) \ \ \text{for} \ i = end \\ + * out_{i} = act(in_{i} + out_{i+1} * W) \ \ \text{for} \ start <= i < end + * \f] + * There are two methods to calculate rnn. One way is to compute rnn one + * sequence by one sequence. The other way is to reorganize the input + * into batches, then compute rnn one batch by one batch. Users can select + * them by rnn_use_batch flag. + */ + +class RecurrentLayer : public Layer { + public: + explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + + void backward(const UpdateCallback& callback) override; + + void resetState() override; + + void setState(LayerStatePtr state) override; + + LayerStatePtr getState() override; + + protected: + /** + * @brief If user do not set --rnn_use_batch=true, it will + * compute rnn forward one sequence by one sequence in default. + * @param batchSize Total words number of all samples in this batch. + * @param numSequences The sample number. + * @param starts Each start position of each samples. + */ + void forwardSequence(int batchSize, size_t numSequences, const int* starts); + /** + * @brief Compute rnn forward by one sequence. + * @param start The start position of this sequence (or sample). + * @param length The length of this sequence (or sample), namely the words + * number of this sequence. + */ + void forwardOneSequence(int start, int length); + /** + * @brief Compute rnn backward one sequence by onesequence. + * @param batchSize Total words number of all samples in this batch. + * @param numSequences The sample number. + * @param starts Each start position of each samples. + */ + void backwardSequence(int batchSize, size_t numSequences, const int* starts); + /** + * @brief Compute rnn backward by one sequence. + * @param start The start position of this sequence (or sample). + * @param length The length of this sequence (or sample), namely the words + * number of this sequence. + */ + void backwardOneSequence(int start, int length); + + /** + * @brief Reorganize input into batches and compute rnn forward batch + * by batch. It will convert batch shape to sequence after finishing forward. + * The batch info can refer to SequenceToBatch class. + * @param batchSize Total words number of all samples in this batch. + * @param numSequences The sample number. + * @param starts Each start position of each samples. + */ + virtual void forwardBatch(int batchSize, + size_t numSequences, + const int* starts); + + /** + * @brief Reorganize input into batches and compute rnn forward batch + * by batch. + * @param batchSize Total words number of all samples in this batch. + * @param numSequences The sample number. + * @param starts Each start position of each samples. + */ + virtual void backwardBatch(int batchSize, + size_t numSequences, + const int* starts); + + protected: + std::unique_ptr weight_; + std::unique_ptr bias_; + + /// frameOutput_[i] is used to hold the i-th sample of output_ + std::vector frameOutput_; + MatrixPtr prevOutput_; + /// Whether compute rnn by reverse. + bool reversed_; + /// If compute batch by batch, batchValue_ will be used to save the + /// reorganized input value. + std::unique_ptr batchValue_; + /// If compute batch by batch, batchGrad_ will be used to save the + /// gradient with respect to reorganized input value. + std::unique_ptr batchGrad_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/RecurrentLayerGroup.cpp b/paddle/legacy/gserver/layers/RecurrentLayerGroup.cpp new file mode 100644 index 0000000000000000000000000000000000000000..39321245995fce2f2bd671593c028fd6038865de --- /dev/null +++ b/paddle/legacy/gserver/layers/RecurrentLayerGroup.cpp @@ -0,0 +1,95 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/gserver/layers/Layer.h" + +#include "paddle/legacy/gserver/gradientmachines/RecurrentGradientMachine.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * Recurrent layer group is a group of layers, which forward/backward one frame + * after previous frame forward/backward through all layers in layer group. + * It's automatically added by config_parser if some layers are defined + * between RecurrentLayerGroupBegin and RecurrentLayerGroupEnd. + */ +class RecurrentLayerGroup : public Layer { + public: + explicit RecurrentLayerGroup(const LayerConfig& config) : Layer(config) {} + + void initSubNetwork(NeuralNetwork* rootNetwork, + const ModelConfig& config, + const std::vector& parameterTypes, + bool useGpu) override; + + void forward(PassType passType) override { + REGISTER_TIMER_INFO("RecurrentGroupFwTime", getName().c_str()); + const std::vector inArgs; + std::vector outArgs; + network_->forward(inArgs, &outArgs, passType); + } + void backward(const UpdateCallback& callback) override { + REGISTER_TIMER_INFO("RecurrentGroupBwTime", getName().c_str()); + network_->backward(nullptr); + + for (auto& para : parameters_) { + para->incUpdate(callback); + } + } + + /** + * @see Layer.accessSubNetwork + */ + void accessSubNetwork( + const std::function& callback) override { + callback(*network_); + } + + private: + std::unique_ptr network_; +}; + +REGISTER_LAYER(recurrent_layer_group, RecurrentLayerGroup); + +void RecurrentLayerGroup::initSubNetwork( + NeuralNetwork* rootNetwork, + const ModelConfig& config, + const std::vector& parameterTypes, + bool useGpu) { + setNeedGradient(true); + + network_.reset(new RecurrentGradientMachine(config_.name(), rootNetwork)); + ParamInitCallback cb = [rootNetwork](int paramId, Parameter* para) { + para->enableSharedType( + PARAMETER_VALUE, + rootNetwork->getParameters()[paramId]->getBuf(PARAMETER_VALUE), + rootNetwork->getParameters()[paramId]->getMat(PARAMETER_VALUE)); + para->enableSharedType( + PARAMETER_GRADIENT, + rootNetwork->getParameters()[paramId]->getBuf(PARAMETER_GRADIENT), + rootNetwork->getParameters()[paramId]->getMat(PARAMETER_GRADIENT)); + }; + network_->init(config, cb, parameterTypes, useGpu); + + for (auto paramId : network_->getParameterIds()) { + ParameterPtr parameter = rootNetwork->getParameters()[paramId]; + parameter->incShared(); + CHECK_EQ(parameter->getDeviceId(), getDeviceId()); + parameters_.push_back(parameter); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ResizeLayer.cpp b/paddle/legacy/gserver/layers/ResizeLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8f8aad820f7d6d2be0af74d607d763912c3c0f2a --- /dev/null +++ b/paddle/legacy/gserver/layers/ResizeLayer.cpp @@ -0,0 +1,79 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/BaseMatrix.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * @brief A layer for resizing a minibatch matrix h*w to h'*w' + * @note + * origin matrix height * width) + * resize matrix: (height * width / size) * size + */ +class ResizeLayer : public Layer { + public: + explicit ResizeLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + + void backward(const UpdateCallback& callback) override; +}; + +REGISTER_LAYER(resize, ResizeLayer); + +bool ResizeLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!Layer::init(layerMap, parameterMap)) return false; + CHECK_EQ(1U, inputLayers_.size()); + + setNeedSequenceInfo(false); + return true; +} + +void ResizeLayer::forward(PassType passType) { + Layer::forward(passType); + const Argument& input = getInput(0); + size_t height = input.value->getHeight(); + size_t width = input.value->getWidth(); + CHECK_EQ((height * width) % getSize(), 0UL); + + reserveOutput(height * width / getSize(), getSize()); + MatrixPtr tmp = + Matrix::create(output_.value->getData(), height, width, false, useGpu_); + tmp->assign(*input.value); +} + +void ResizeLayer::backward(const UpdateCallback& callback) { + const Argument& input = getInput(0); + size_t height = input.value->getHeight(); + size_t width = input.value->getWidth(); + + if (!input.grad) { + return; + } + + MatrixPtr tmp = Matrix::create(input.grad->getData(), + height * width / getSize(), + getSize(), + false, + useGpu_); + tmp->add(*output_.grad); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/RotateLayer.cpp b/paddle/legacy/gserver/layers/RotateLayer.cpp similarity index 100% rename from paddle/gserver/layers/RotateLayer.cpp rename to paddle/legacy/gserver/layers/RotateLayer.cpp diff --git a/paddle/legacy/gserver/layers/RotateLayer.h b/paddle/legacy/gserver/layers/RotateLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..498e24372b8ca17c21ebecbe6a8c8b40217ab259 --- /dev/null +++ b/paddle/legacy/gserver/layers/RotateLayer.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * A layer for rotating a multi-channel feature map (M x N x C) in the spatial + * domain + * The rotation is 90 degrees in clock-wise for each channel + * \f[ + * y(j,i,:) = x(M-i-1,j,:) + * \f] + * where \f$x\f$ is (M x N x C) input, and \f$y\f$ is (N x M x C) output. + * + * The config file api is rotate_layer + * + */ + +class RotateLayer : public Layer { + public: + explicit RotateLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + + void forward(PassType passType); + void backward(const UpdateCallback& callback = nullptr); + + private: + int batchSize_; + int size_; + int height_; + int width_; + int channels_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/RowConvLayer.cpp b/paddle/legacy/gserver/layers/RowConvLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1961557dc2d2601091bb0e56fcd884d76d49bc0e --- /dev/null +++ b/paddle/legacy/gserver/layers/RowConvLayer.cpp @@ -0,0 +1,106 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "RowConvLayer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(row_conv, RowConvLayer); + +bool RowConvLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + contexLength_ = config_.inputs(0).row_conv_conf().context_length(); + + CHECK_EQ(inputLayers_.size(), 1UL); + weight_.reset(new Weight(contexLength_, getSize(), parameters_[0])); + createFunction(forward_, "RowConv", FuncConfig()); + createFunction(backward_, "RowConvGrad", FuncConfig()); + + return true; +} + +void RowConvLayer::forward(PassType passType) { + Layer::forward(passType); + MatrixPtr input = getInputValue(0); + size_t height = input->getHeight(); + size_t width = input->getWidth(); + CHECK_EQ(width, getSize()); + resetOutput(height, width); + + const auto startPos = getInput(0).sequenceStartPositions->getVector(useGpu_); + MatrixPtr w = weight_->getW(); + wDims_ = TensorShape({w->getHeight(), w->getWidth()}); + + MatrixPtr outV = getOutputValue(); + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), *startPos); + inputs.addArg(*w, wDims_); + outputs.addArg(*getOutputValue(), *startPos, ADD_TO); + + { + REGISTER_TIMER_INFO("RowConvForward", getName().c_str()); + forward_[0]->calc(inputs, outputs); + } + + /* activation */ { + REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void RowConvLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { + REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str()); + backwardActivation(); + } + + const auto startPos = getInput(0).sequenceStartPositions->getVector(useGpu_); + + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), *startPos); + inputs.addArg(*getInputValue(0), *startPos); + inputs.addArg(*weight_->getW(), wDims_); + + MatrixPtr inGrad = getInputGrad(0); + MatrixPtr wGrad = weight_->getWGrad(); + size_t h = getInputValue(0)->getHeight(); + size_t w = getInputValue(0)->getWidth(); + outputs.addArg( + inGrad ? (*inGrad) : *(Matrix::create(nullptr, h, w, false, useGpu_)), + *startPos, + ADD_TO); + outputs.addArg( + wGrad ? (*wGrad) + : *(Matrix::create(nullptr, contexLength_, w, false, useGpu_)), + wDims_, + ADD_TO); + + { + REGISTER_TIMER_INFO("RowConvBackward", getName().c_str()); + backward_[0]->calc(inputs, outputs); + } + + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + weight_->getParameterPtr()->incUpdate(callback); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/RowConvLayer.h b/paddle/legacy/gserver/layers/RowConvLayer.h similarity index 100% rename from paddle/gserver/layers/RowConvLayer.h rename to paddle/legacy/gserver/layers/RowConvLayer.h diff --git a/paddle/gserver/layers/RowL2NormLayer.cpp b/paddle/legacy/gserver/layers/RowL2NormLayer.cpp similarity index 100% rename from paddle/gserver/layers/RowL2NormLayer.cpp rename to paddle/legacy/gserver/layers/RowL2NormLayer.cpp diff --git a/paddle/gserver/layers/SamplingIdLayer.cpp b/paddle/legacy/gserver/layers/SamplingIdLayer.cpp similarity index 100% rename from paddle/gserver/layers/SamplingIdLayer.cpp rename to paddle/legacy/gserver/layers/SamplingIdLayer.cpp diff --git a/paddle/gserver/layers/ScaleShiftLayer.cpp b/paddle/legacy/gserver/layers/ScaleShiftLayer.cpp similarity index 100% rename from paddle/gserver/layers/ScaleShiftLayer.cpp rename to paddle/legacy/gserver/layers/ScaleShiftLayer.cpp diff --git a/paddle/legacy/gserver/layers/ScaleSubRegionLayer.cpp b/paddle/legacy/gserver/layers/ScaleSubRegionLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..70d44d2a7ef25df64beb2c861692436d842dac02 --- /dev/null +++ b/paddle/legacy/gserver/layers/ScaleSubRegionLayer.cpp @@ -0,0 +1,78 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ScaleSubRegionLayer.h" +#include "paddle/legacy/utils/Stat.h" +namespace paddle { + +REGISTER_LAYER(scale_sub_region, ScaleSubRegionLayer); + +bool ScaleSubRegionLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + CHECK_EQ(static_cast(inputLayers_.size()), 2); + auto& conf = config_.inputs(0).scale_sub_region_conf(); + value_ = conf.value(); + + createFunction(forward_, "ScaleSubRegion", FuncConfig().set("value", value_)); + createFunction( + backward_, "ScaleSubRegionGrad", FuncConfig().set("value", value_)); + + return true; +} + +void ScaleSubRegionLayer::forward(PassType passType) { + Layer::forward(passType); + auto in0 = getInput(0); + imgH_ = in0.getFrameHeight(); + imgW_ = in0.getFrameWidth(); + if (imgH_ == 0 || imgW_ == 0) { + auto& conf = config_.inputs(0).scale_sub_region_conf(); + imgH_ = conf.image_conf().img_size_y(); + imgW_ = conf.image_conf().img_size(); + } + MatrixPtr imgV = in0.value; + size_t batchSize = imgV->getHeight(); + size_t spatialSize = imgH_ * imgW_; + channelsNum_ = imgV->getWidth() / spatialSize; + shape_ = TensorShape({batchSize, channelsNum_, imgH_, imgW_}); + + resetOutput(batchSize, imgV->getWidth()); + auto& out = getOutput(); + out.setFrameHeight(imgH_); + out.setFrameWidth(imgW_); + + MatrixPtr indicesV = getInputValue(1); + indicesShape_ = TensorShape({batchSize, 6}); + + REGISTER_TIMER_INFO("ScaleSubRegionForward", getName().c_str()); + BufferArgs inArgs; + BufferArgs outArgs; + inArgs.addArg(*imgV, shape_); + inArgs.addArg(*indicesV, indicesShape_); + outArgs.addArg(*out.value, shape_, ASSIGN_TO); + forward_[0]->calc(inArgs, outArgs); +} + +void ScaleSubRegionLayer::backward(const UpdateCallback& callback) { + REGISTER_TIMER_INFO("ScaleSubRegionBackward", getName().c_str()); + BufferArgs inArgs; + BufferArgs outArgs; + inArgs.addArg(*getOutputGrad(), shape_); + inArgs.addArg(*getInputValue(1), indicesShape_); + outArgs.addArg(*getInputGrad(0), shape_, ADD_TO); + backward_[0]->calc(inArgs, outArgs); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/ScaleSubRegionLayer.h b/paddle/legacy/gserver/layers/ScaleSubRegionLayer.h similarity index 100% rename from paddle/gserver/layers/ScaleSubRegionLayer.h rename to paddle/legacy/gserver/layers/ScaleSubRegionLayer.h diff --git a/paddle/legacy/gserver/layers/ScalingLayer.cpp b/paddle/legacy/gserver/layers/ScalingLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a8286b6614c3cdfbd720d0719f939018f6ae9579 --- /dev/null +++ b/paddle/legacy/gserver/layers/ScalingLayer.cpp @@ -0,0 +1,106 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for each row of a matrix, multiplying with a element of a vector, + * which is used in NEURAL TURING MACHINE. + * \f[ + * y.row[i] = w[i] * x.row[i] + * \f] + * where \f$x\f$ is (batchSize x dataDim) input, \f$w\f$ is + * (batchSize x 1) weight vector, and \f$y\f$ is (batchSize x dataDim) output. + * + * The config file api is scaling_layer. + */ + +class ScalingLayer : public Layer { + public: + explicit ScalingLayer(const LayerConfig& config) : Layer(config) {} + + ~ScalingLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(scaling, ScalingLayer); + +bool ScalingLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 2U); + + return true; +} + +void ScalingLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr weightV = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + + size_t batchSize = inV1->getHeight(); + size_t dataDim = inV1->getWidth(); + + CHECK_EQ(dataDim, getSize()); + CHECK_EQ(weightV->getWidth(), 1U); + CHECK_EQ(weightV->getHeight(), batchSize); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + resetOutput(batchSize, dataDim); + } + + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwScalingTimer", getName().c_str()); + // outV += inV1 * weight + outV->addRowScale(0, *inV1, *weightV); + } +} + +void ScalingLayer::backward(const UpdateCallback& callback) { + MatrixPtr weightV = getInputValue(0); + MatrixPtr inV1 = getInputValue(1); + MatrixPtr inG0 = getInputGrad(0); + MatrixPtr inG1 = getInputGrad(1); + MatrixPtr outG = getOutputGrad(); + + { + REGISTER_TIMER_INFO("BwScalingTimer", getName().c_str()); + + if (inG0) { + // inG0 += outG .* inV1 + inG0->rowDotMul(0, *outG, *inV1); + } + + if (inG1) { + // inG1 += outG * weight; + inG1->addRowScale(0, *outG, *weightV); + } + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/ScalingProjection.cpp b/paddle/legacy/gserver/layers/ScalingProjection.cpp similarity index 100% rename from paddle/gserver/layers/ScalingProjection.cpp rename to paddle/legacy/gserver/layers/ScalingProjection.cpp diff --git a/paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.cpp b/paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..72fb06814884cc2bcca2c600105077d8cf1459c5 --- /dev/null +++ b/paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.cpp @@ -0,0 +1,336 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SelectiveFullyConnectedLayer.h" +#include +#include +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(selective_fc, SelectiveFullyConnectedLayer); + +bool SelectiveFullyConnectedLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + inputNum_ = inputLayers_.size(); + if (config_.has_selected_colums()) { + inputNum_ -= 1; + } + for (size_t i = 0; i < inputNum_; i++) { + size_t height = inputLayers_[i]->getSize(); + size_t width = getSize(); + // NOTE weight is transpoed + weights_.emplace_back(new Weight(width, height, parameters_[i])); + } + + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + + fullOutput_ = false; + + return true; +} + +void SelectiveFullyConnectedLayer::prefetch() {} + +void SelectiveFullyConnectedLayer::reserveOutput(size_t height, + size_t width, + size_t nnz) { + bool flag = (passType_ == PASS_TEST && + config_.selective_fc_pass_generation() && !fullOutput_); + SetDevice device(output_.deviceId); + if (flag) { + // output_.value is sparse matrix + if (dynamic_cast(output_.value.get()) || + dynamic_cast(output_.value.get())) { + output_.value = nullptr; + } + Matrix::resizeOrCreateSparseMatrix(output_.value, + height, + width, + nnz, + FLOAT_VALUE, + SPARSE_CSR, + /*trans=*/false, + /*useGpu=*/useGpu_); + output_.value->copyFrom(*selCols_); + interOutput_ = output_.value; + } else { + if (fullOutput_) { + // output_.value is dense matrix + if (dynamic_cast(output_.value.get()) || + dynamic_cast(output_.value.get())) { + output_.value = nullptr; + } + Matrix::resizeOrCreate(output_.value, + height, + width, + /*trans=*/false, + /*useGpu=*/useGpu_); + interOutput_ = output_.value; + } else { + // output_.value is dense matrix, but width = nnz /height + CHECK_EQ(nnz % height, 0U); + CHECK(nnz / height); + Matrix::resizeOrCreate(output_.value, + height, + nnz / height, + /*trans=*/false, + /*useGpu=*/useGpu_); + interOutput_ = Matrix::createSparseMatrix(output_.value->getData(), + selCols_->getRows(), + selCols_->getCols(), + height, + width, + nnz, + FLOAT_VALUE, + SPARSE_CSR, + /*trans=*/false, + /*useGpu=*/useGpu_); + } + } + interOutput_->zeroMem(); + + if (passType_ != PASS_TEST && needGradient()) { + CHECK_EQ(nnz % height, 0U) << "during training, each sample must have a " + "same number of selected columns."; + CHECK(nnz / height) + << "during training, " + "each sample must have at least one column selected."; + Matrix::resizeOrCreate(output_.grad, + height, + nnz / height, + /*trans=*/false, + /*useGpu=*/useGpu_); + output_.grad->zeroMem(); + } +} + +void SelectiveFullyConnectedLayer::forward(PassType passType) { + REGISTER_TIMER("selective_fc.forward"); + Layer::forward(passType); + + getSelectiveCols(); + size_t height = getInput(0).getBatchSize(); + size_t width = getSize(); + size_t nnz = height * width; + if (!fullOutput_) { + CHECK(selCols_); + CHECK(height == selCols_->getHeight()); + CHECK(width == selCols_->getWidth()); + nnz = selCols_->getElementCnt(); + } + + // Layer::ResetOutput(), here we set outV/outG as SparseMatrix manually + // this outV should be used as input of MaxIdLayer and softmax activation + reserveOutput(height, width, nnz); + + bool flag = true; + for (size_t i = 0; i < inputNum_; i++) { + MatrixPtr input = getInputValue(i); + MatrixPtr weight = weights_[i]->getW(); + size_t hsize = input->getHeight(); + size_t wsize = weight->getHeight(); + real scaleT = i == 0 ? real(0) : real(1); + + flag = nnz < (hsize * wsize) * config_.selective_fc_full_mul_ratio() && + !fullOutput_; + if (flag) { + // if the indecies are highly sparse, + // manully compute the multiplication of + // the input vector and the selected rows. + REGISTER_TIMER("selective.plain"); + interOutput_->mul(*input, *weight->getTranspose(), 1, scaleT); + } else { + // if the indecies is not sparse enough, + // use full mul instead + REGISTER_TIMER("selective.mul"); + if (fullOutput_) { + interOutput_->mul(*input, *weight->getTranspose(), 1, scaleT); + } else { + Matrix::resizeOrCreate(mmat_, + hsize, + wsize, + /*trans=*/false, + /*useGpu=*/useGpu_); + mmat_->mul(*input, *weight->getTranspose()); + interOutput_->add3(mmat_); + } + } + } + + if (biases_) { + interOutput_->addBias(*(biases_->getW()), 1); + } + + flag = (passType_ == PASS_TEST && config_.selective_fc_pass_generation() && + !fullOutput_); + if (flag) { + // during generation, output of this layer is a sparse csr matrix, + // which is probably the input of maxid layer + // if the model is trained with multi-class-cross-entroy-with-selfnorm, + // activiation of this layer should be exponential, not softmax. + + Argument arg; + arg.value = Matrix::create(interOutput_->getData(), + 1, + nnz, + /*trans=*/false, + /*useGpu=*/useGpu_); + //! TODO(yuyang18): Why we cannot invoke forwardActivation here? + activation_->forward(arg).check(); + } else /* train and test in train, not generating */ { + // during training, this layer output value is *Matrix*, which is input of + // eg. multi-class-cross-entropy + + // while training, every sample has a equal number of selected + // columns to be activated. + // note indices of multi-class-cross-entropy need to be remapped + // to this index. + // e.g. sample = [1,3,5] and 3 is gold, then label is 1 + + forwardActivation(); + } +} + +void SelectiveFullyConnectedLayer::backward(const UpdateCallback& callback) { + backwardActivation(); + MatrixPtr oGrad = getOutputGrad(); + if (!fullOutput_) { + interOutGrad_ = Matrix::createSparseMatrix(oGrad->getData(), + interOutput_->getRows(), + interOutput_->getCols(), + interOutput_->getHeight(), + interOutput_->getWidth(), + interOutput_->getElementCnt(), + FLOAT_VALUE, + SPARSE_CSR, + /*trans=*/false, + /*useGpu=*/useGpu_); + } else { + interOutGrad_ = Matrix::create(oGrad->getData(), + oGrad->getHeight(), + oGrad->getWidth(), + /*trans=*/false, + /*useGpu=*/useGpu_); + } + + if (biases_ && biases_->getWGrad()) { + REGISTER_TIMER_INFO("BpBiasTimer", getName().c_str()); + biases_->getWGrad()->collectBias(*interOutGrad_, 1); + biases_->getParameterPtr()->incUpdate(callback); + } + + // backward is different from FullyConnectedLayer + // because the weight is transposed + for (size_t i = 0; i < inputNum_; i++) { + AsyncGpuBlock block; + MatrixPtr preGrad = getInputGrad(i); + if (preGrad) { + REGISTER_TIMER_INFO("BpMulTimer", getName().c_str()); + preGrad->mul(*interOutGrad_, *weights_[i]->getW(), 1, 1); + } + + MatrixPtr wGrad = weights_[i]->getWGrad(); + if (wGrad) { + REGISTER_TIMER_INFO("GradMulTimer", getName().c_str()); + MatrixPtr input = getInputValue(i); + wGrad->mul(*interOutGrad_->getTranspose(), *input, 1, 1); + } + + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + weights_[i]->getParameterPtr()->incUpdate(callback); + } + } +} + +void paddle::SelectiveFullyConnectedLayer::fillSelectiveData( + const std::shared_ptr>>& candidates) { + if (candidates == nullptr) { + fillFullySelectiveData(); + return; + } + + size_t sampleNum = candidates->size(); + size_t outputWidth = getSize(); + size_t nnz = + std::accumulate(candidates->begin(), + candidates->end(), + 0UL, + [](size_t a, const std::pair& arr) { + return a + arr.second; + }); + + Matrix::resizeOrCreateSparseMatrix(this->cpuSelCols_, + sampleNum, + outputWidth, + nnz, + NO_VALUE, + SPARSE_CSR, + false, + false); + CHECK(this->cpuSelCols_ != nullptr); + CpuSparseMatrixPtr selCols = + std::dynamic_pointer_cast(cpuSelCols_); + int* rowOffsets = selCols->getRows(); + int* colIndices = selCols->getCols(); + + rowOffsets[0] = 0; + int idx = 0; + for (size_t i = 0; i < sampleNum; ++i) { + if ((*candidates)[i].second > 0) { + rowOffsets[i + 1] = rowOffsets[i] + (*candidates)[i].second; + for (size_t j = 0; j < (*candidates)[i].second; ++j) { + colIndices[idx] = (*candidates)[i].first[j]; + idx++; + } + } else { + rowOffsets[i + 1] = rowOffsets[i]; + } + } + + CHECK_EQ(static_cast(rowOffsets[sampleNum]), nnz); + if (!useGpu_) { + this->selCols_ = this->cpuSelCols_; + } else { + Matrix::resizeOrCreateSparseMatrix(this->selCols_, + sampleNum, + outputWidth, + nnz, + NO_VALUE, + SPARSE_CSR, + false, + true); + this->selCols_->copyFrom(*cpuSelCols_, HPPL_STREAM_1); + hl_stream_synchronize(HPPL_STREAM_1); + } + + fullOutput_ = false; +} + +void paddle::SelectiveFullyConnectedLayer::getSelectiveCols() { + if (config_.has_selected_colums()) { + this->selCols_ = inputLayers_[inputNum_]->getOutputValue(); + fullOutput_ = false; + } else if (!config_.selective_fc_pass_generation() || selCols_ == nullptr) { + this->fillFullySelectiveData(); + } // else selCols_ is initialized by fillSelectiveData +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.h b/paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..3ba04d9b2ae208eda021a451e94856d9993dc126 --- /dev/null +++ b/paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.h @@ -0,0 +1,103 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * @brief The SelectiveFullyConnectedLayer class + * + * SelectiveFullyConnectedLayer differs from FullyConnectedLayer by that it + * requires an additional input to indicate several selected columns, and only + * compute the multiplications between the input matrices and the selected + * columns of the parameter matrices of this layer. If the selected columns is + * not specified, SelectiveFullyConnected layer acts exactly like + * FullyConnectedLayer. + * + * The config file api is selective_fc_layer. + */ +class SelectiveFullyConnectedLayer : public Layer { + protected: + WeightList weights_; + std::unique_ptr biases_; + + private: + /** + * Get selected columns each forward. + */ + void getSelectiveCols(); + + MatrixPtr mmat_; + /// cpuSelCols_ is a CpuSparseMatrix, used to save selected columns. + MatrixPtr cpuSelCols_; + /// CpuSparseMatrix or GpuSparseMatrix. In CPU mode, selCols_ points + /// to cpuSelCols_. + MatrixPtr selCols_; + size_t inputNum_; + + /// interOutput_ shared same memory with output_.value. + MatrixPtr interOutput_; + + /// if fullOutput_ is false, interOutGrad_ sparse matrix + MatrixPtr interOutGrad_; + + /// if true, means output_.value is the same as Fc Layer + bool fullOutput_; + + public: + explicit SelectiveFullyConnectedLayer(const LayerConfig& config) + : Layer(config), selCols_(nullptr) {} + + ~SelectiveFullyConnectedLayer() {} + void prefetch() override; + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + Weight& getWeight(int idx) { return *weights_[idx]; } + + /** + * @brief Resize the output matrix size. + * And reset value to zero + */ + void reserveOutput(size_t height, size_t width, size_t nnz); + + /** + * @brief Fill candidates to select several activations as output. + * @param candidates specifies several selected columns of the parameter + * matrices of this layer. + * Multiplications only between the input matrices and the selected columns + * are computed. + * If the candidates is a nullptr, selective fc layer acts exactly like the + * fully connected layer. + * @note CURRENTLY, THIS METHOD IS ONLY USED FOR BEAM SEARCH + */ + void fillSelectiveData( + const std::shared_ptr>>& candidates); + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + private: + /** + * @brief Make SelectiveFC act as FullyConnectedLayer + */ + void fillFullySelectiveData() { fullOutput_ = true; } +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SequenceConcatLayer.cpp b/paddle/legacy/gserver/layers/SequenceConcatLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7b598e11acde533564f6eda49d78ea8df99a5056 --- /dev/null +++ b/paddle/legacy/gserver/layers/SequenceConcatLayer.cpp @@ -0,0 +1,189 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for concatenating the first sequence with the second sequence + * Input: two sequences each containing the same number of instances + * seq1 = [a1, a2, ..., an] + * seq2 = [b1, b2, ..., bn] + * Output: a concatenated sequence of the two input sequences + * out = [a1, b1, a2, b2, ..., an, bn] + */ + +class SequenceConcatLayer : public Layer { + protected: + std::unique_ptr biases_; + + public: + explicit SequenceConcatLayer(const LayerConfig& config) : Layer(config) {} + + ~SequenceConcatLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(seqconcat, SequenceConcatLayer); + +bool SequenceConcatLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + // sequene concatenation layer should have exactly 2 inputs + CHECK_EQ(2U, inputLayers_.size()); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + + setNeedSequenceInfo(false); + return true; +} + +void SequenceConcatLayer::forward(PassType passType) { + Layer::forward(passType); + + size_t dim = getSize(); + + const Argument& input1 = getInput(0); + size_t numSequences1 = input1.getNumSequences(); + auto startPositions1 = input1.sequenceStartPositions->getVector(false); + + const Argument& input2 = getInput(1); + size_t numSequences2 = input2.getNumSequences(); + auto startPositions2 = input2.sequenceStartPositions->getVector(false); + + CHECK_EQ(dim, input1.value->getWidth()); + CHECK_EQ(startPositions1->getData()[numSequences1], input1.getBatchSize()); + CHECK_EQ(numSequences1, startPositions1->getSize() - 1); + + CHECK_EQ(dim, input2.value->getWidth()); + CHECK_EQ(startPositions2->getData()[numSequences2], input2.getBatchSize()); + CHECK_EQ(numSequences2, startPositions2->getSize() - 1); + + CHECK_EQ(numSequences1, numSequences2); + + MatrixPtr inputValue1 = getInputValue(0); + MatrixPtr inputValue2 = getInputValue(1); + + // reset output + reserveOutput(inputValue1->getHeight() + inputValue2->getHeight(), dim); + + MatrixPtr outputValue = getOutputValue(); + + const int* starts1 = startPositions1->getData(); + const int* starts2 = startPositions2->getData(); + + { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceConcatLayerForward", getName().c_str()); + + size_t offset = 0; + size_t leftNumIns = 0; + size_t rightNumIns = 0; + for (size_t seqId = 0; seqId < numSequences1; ++seqId) { + leftNumIns = starts1[seqId + 1] - starts1[seqId]; + outputValue->subMatrix(offset, leftNumIns) + ->assign(*(inputValue1->subMatrix(starts1[seqId], leftNumIns))); + offset += leftNumIns; + + rightNumIns = starts2[seqId + 1] - starts2[seqId]; + outputValue->subMatrix(offset, rightNumIns) + ->assign(*(inputValue2->subMatrix(starts2[seqId], rightNumIns))); + offset += rightNumIns; + } + + // modify the sequenceStartPositions + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, numSequences1 + 1, false); + + int* tgtBuf = output_.sequenceStartPositions->getMutableData(false); + + for (size_t seqId = 0; seqId < numSequences1 + 1; ++seqId) { + tgtBuf[seqId] = starts1[seqId] + starts2[seqId]; + } + } + + if (biases_.get() != NULL) { + MatrixPtr outV = getOutputValue(); + outV->addBias(*(biases_->getW()), 1); + } + + /* activation */ + forwardActivation(); +} + +void SequenceConcatLayer::backward(const UpdateCallback& callback) { + /* activation */ + backwardActivation(); + + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + // Increasing the number of gradient + biases_->getParameterPtr()->incUpdate(callback); + } + + MatrixPtr inputGrad1 = getInputGrad(0); + MatrixPtr inputGrad2 = getInputGrad(1); + MatrixPtr outputGrad = getOutputGrad(); + auto startPositions1 = getInput(0).sequenceStartPositions->getVector(false); + auto startPositions2 = getInput(1).sequenceStartPositions->getVector(false); + + size_t numSequences1 = startPositions1->getSize() - 1; + size_t numSequences2 = startPositions2->getSize() - 1; + + CHECK_EQ(numSequences1, numSequences2); + + const int* starts1 = startPositions1->getData(); + const int* starts2 = startPositions2->getData(); + + { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceConcatLayerBackward", getName().c_str()); + + size_t offset = 0; + size_t leftNumIns = 0; + size_t rightNumIns = 0; + for (size_t seqId = 0; seqId < numSequences1; ++seqId) { + leftNumIns = starts1[seqId + 1] - starts1[seqId]; + if (inputGrad1) { + inputGrad1->subMatrix(starts1[seqId], leftNumIns) + ->add(*(outputGrad->subMatrix(offset, leftNumIns))); + } + offset += leftNumIns; + + rightNumIns = starts2[seqId + 1] - starts2[seqId]; + if (inputGrad2) { + inputGrad2->subMatrix(starts2[seqId], rightNumIns) + ->add(*(outputGrad->subMatrix(offset, rightNumIns))); + } + offset += rightNumIns; + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SequenceLastInstanceLayer.cpp b/paddle/legacy/gserver/layers/SequenceLastInstanceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8735d71ba372de894c9852229ed8c77537792ea0 --- /dev/null +++ b/paddle/legacy/gserver/layers/SequenceLastInstanceLayer.cpp @@ -0,0 +1,118 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Logging.h" + +#include "SequencePoolLayer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for extracting the last instance of the input sequence. + * Input: a sequence + * If SequenceLevel = kNonseq: + * Output: a sequence containing only the last instance of the input sequence + * If stride_ > 0: + * Output: a shorten sequence. Stride is the step size by which we slide a + * window upon the input sequence, and getting last instance + * operation is then applied to each interval independently. + * If SequenceLevel = kSeq: + * Check input sequence must has sub-sequence + * Output: a sequence containing only the last instance of each sub-sequence + * of the input sequence + * + * The config file api is last_seq and first_seq. + */ + +class SequenceLastInstanceLayer : public SequencePoolLayer { + protected: + MatrixPtr tmpSrc_; + MatrixPtr tmpDest_; + std::vector instanceIds_; + + public: + explicit SequenceLastInstanceLayer(const LayerConfig& config) + : SequencePoolLayer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer); + +bool SequenceLastInstanceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + SequencePoolLayer::init(layerMap, parameterMap); + reversed_ = config_.select_first(); + + tmpSrc_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); + tmpDest_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); + + return true; +} + +void SequenceLastInstanceLayer::forward(PassType passType) { + SequencePoolLayer::forward(passType); + + auto starts = startPositions_->getData(false); + MatrixPtr inputValue = getInputValue(0); + MatrixPtr outputValue = getOutputValue(); + + { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceLastInstanceLayerForward", getName().c_str()); + + instanceIds_.clear(); + for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) { + int insId = reversed_ ? starts[seqId] : starts[seqId + 1] - 1; + instanceIds_.push_back(insId); + + outputValue->subMatrix(seqId, 1, tmpDest_) + ->assign(*(inputValue->subMatrix(insId, 1, tmpSrc_))); + } + } + + if (biases_.get() != NULL) { + outputValue->addBias(*(biases_->getW()), 1); + } + + /* activation, should set to 'linear' in most cases */ + forwardActivation(); +} + +void SequenceLastInstanceLayer::backward(const UpdateCallback& callback) { + SequencePoolLayer::backward(callback); + + MatrixPtr inputGrad = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + + if (inputGrad) { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceLastInstanceLayerBackward", getName().c_str()); + + for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) { + inputGrad->subMatrix(instanceIds_[seqId], 1, tmpDest_) + ->add(*(outputGrad->subMatrix(seqId, 1, tmpSrc_))); + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SequencePoolLayer.cpp b/paddle/legacy/gserver/layers/SequencePoolLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..243b795db428ede1fbb39a5054485a198a14e00c --- /dev/null +++ b/paddle/legacy/gserver/layers/SequencePoolLayer.cpp @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SequencePoolLayer.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +bool SequencePoolLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + // seqlastins/max/average layer should have exactly 1 input + CHECK_EQ(1U, inputLayers_.size()); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + // transform to which sequence type + if (config_.trans_type() == "non-seq") { + type_ = kNonSeq; + } else if (config_.trans_type() == "seq") { + type_ = kSeq; + } else { + LOG(FATAL) << "Unknown trans_type: " << config_.trans_type(); + } + stride_ = config_.seq_pool_stride(); + setNeedSequenceInfo(false); + return true; +} + +void SequencePoolLayer::forward(PassType passType) { + Layer::forward(passType); + + const Argument& input = getInput(0); + CHECK(input.hasSeq() || input.hasSubseq()) + << "Input should be a sequence or subsequence for layer " << getName(); + + newBatchSize_ = type_ ? input.getNumSubSequences() : input.getNumSequences(); + size_t dim = getSize(); + // check + CHECK_EQ(dim, input.value->getWidth()); + startPositions_ = + type_ ? input.subSequenceStartPositions : input.sequenceStartPositions; + auto starts = startPositions_->getVector(false); + CHECK_EQ(starts->getData()[newBatchSize_], input.getBatchSize()); + CHECK_EQ(newBatchSize_, starts->getSize() - 1); + + /* If type_ = kNonSeq, both seq has or not has sub-seq degrade to a non-seq, + * thus, in this case, output_ has no sequenceStartPositions. + * If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this + * case, we should compute the new sequenceStartPositions. + */ + if (type_) { + CHECK(input.subSequenceStartPositions) + << "when trans_type = seq, input must hasSubseq"; + output_.degradeSequence(input); + } + if (stride_ > 0) { + CHECK_EQ(input.hasSubseq(), 0UL) + << "sequence stride pooling is invalid for hasSubseq now"; + output_.poolSequenceWithStride(input, stride_, &startPositions_, reversed_); + newBatchSize_ = startPositions_->getSize() - 1; + } + + resetOutput(newBatchSize_, dim); +} + +void SequencePoolLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { backwardActivation(); } + + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + // Increasing the number of gradient + biases_->getParameterPtr()->incUpdate(callback); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SequencePoolLayer.h b/paddle/legacy/gserver/layers/SequencePoolLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..1c019b313093f4ac717e0fc57a9aa798e2951580 --- /dev/null +++ b/paddle/legacy/gserver/layers/SequencePoolLayer.h @@ -0,0 +1,64 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * A base layer for SequenceLastInstanceLayer/AverageLayer/MaxLayer. + * + * Input: one or more sequences. Each sequence contains some instances. + * If SequenceLevel = kNonSeq: + * Output: output size is the number of input sequences (NOT input instances) + * output[i] = seqlastin/average/max_{for each instance in this + * sequence}{input[i]} + * If stride_ > 0: + * Check input sequence must not have sub-sequence + * Output: a shorten sequence. Stride is the step size by which we slide + * a window upon the input sequence, and the pooling operation + * is then applied to each interval independently. + * If SequenceLevel = kSeq: + * Check input sequence must has sub-sequence + * Output: output size is the number of input sub-sequences + * output[i] = seqlastin/average/max_{for each instance in this + * sub-sequence}{input[i]} + * + * The config file api is pooling_layer. + */ + +class SequencePoolLayer : public Layer { + protected: + int type_; + std::unique_ptr biases_; + enum SequenceLevel { kNonSeq = 0, kSeq = 1 }; + size_t newBatchSize_; + ICpuGpuVectorPtr startPositions_; + int stride_; + // Whether the input sequence is reversed or not. + bool reversed_ = false; + + public: + explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SequenceReshapeLayer.cpp b/paddle/legacy/gserver/layers/SequenceReshapeLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e3d40cab50af1d6eafe28331cdd481ee2b187a56 --- /dev/null +++ b/paddle/legacy/gserver/layers/SequenceReshapeLayer.cpp @@ -0,0 +1,157 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for reshaping the sequence. Assume the input sequence has + * T instances, the dimension of each instance is M, and the input + * reshape_dim is N, then the output sequence has T*M/N instances, + * the dimension of each instance is N. + * + * Note that T*M/N must be an integer. + */ + +class SequenceReshapeLayer : public Layer { + protected: + std::unique_ptr biases_; + + MatrixPtr reshapedOutputGrad; + + public: + explicit SequenceReshapeLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(seqreshape, SequenceReshapeLayer); + +bool SequenceReshapeLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + CHECK_EQ(1U, inputLayers_.size()); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + setNeedSequenceInfo(false); + return true; +} + +void SequenceReshapeLayer::forward(PassType passType) { + Layer::forward(passType); + + const Argument& input = getInput(0); + + size_t inDim = input.value->getWidth(); + size_t outDim = getSize(); + + size_t numSequences = input.getNumSequences(); + + // by default, we assume each instance as a sequence + IVectorPtr seqStarts; + IVector::resizeOrCreate(seqStarts, input.getBatchSize() + 1, false); + int* startsData = seqStarts->getData(); + for (int i = 0; i < input.getBatchSize() + 1; i++) { + startsData[i] = i; + } + const int* starts = startsData; + + // if there is sequence, then use start positions + if (input.sequenceStartPositions) { + auto startPositions = input.sequenceStartPositions->getVector(false); + starts = startPositions->getData(); + CHECK_EQ(starts[numSequences], input.getBatchSize()); + CHECK_EQ(numSequences, startPositions->getSize() - 1); + } + + for (size_t seqID = 0; seqID < numSequences; seqID++) { + size_t inNumIns = starts[seqID + 1] - starts[seqID]; + size_t outNumIns = inNumIns * inDim / outDim; + CHECK_EQ(outNumIns * outDim, inNumIns * inDim); + } + + MatrixPtr inputValue = getInputValue(0); + + // reset output + reserveOutput(inputValue->getHeight() * inDim / outDim, outDim); + MatrixPtr outputValue = getOutputValue(); + + { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceReshapeLayerForward", getName().c_str()); + + outputValue->copyFrom(*inputValue); + + // modify the sequenceStartPositions + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, numSequences + 1, false); + + int* tgtBuf = output_.sequenceStartPositions->getMutableData(false); + + for (size_t seqId = 0; seqId < numSequences + 1; ++seqId) { + tgtBuf[seqId] = starts[seqId] * inDim / outDim; + } + } + + if (biases_.get() != NULL) { + MatrixPtr outV = getOutputValue(); + outV->addBias(*(biases_->getW()), 1); + } + + /* activation */ + forwardActivation(); +} + +void SequenceReshapeLayer::backward(const UpdateCallback& callback) { + /* activation */ + backwardActivation(); + + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + // Increasing the number of gradient + biases_->getParameterPtr()->incUpdate(callback); + } + + MatrixPtr inputGrad = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SequenceReshapeLayerBackward", getName().c_str()); + + if (inputGrad) { + Matrix::resizeOrCreate(reshapedOutputGrad, + inputGrad->getHeight(), + inputGrad->getWidth(), + false, + useGpu_); + reshapedOutputGrad->copyFrom(*outputGrad); + inputGrad->add(*reshapedOutputGrad); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SequenceSliceLayer.cpp b/paddle/legacy/gserver/layers/SequenceSliceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3ed51c4ef2f6e91da94f302c14d1c0cc555886aa --- /dev/null +++ b/paddle/legacy/gserver/layers/SequenceSliceLayer.cpp @@ -0,0 +1,224 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +class SequenceSliceLayer : public Layer { + public: + explicit SequenceSliceLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + private: + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but the second and the (optional) third input which are some + * selected indices of the give sequence to trim the sequence, are actually + * filled with int types so that storing int types information in real number + * matrices is very dangerous, since real numbers will be convered to int + * types. If a user fills this matrix himself, invalid data may occor. + */ + + MatrixPtr startIdsOnCpu_; + MatrixPtr endIdsOnCpu_; + + std::vector selectedRows_; + IVectorPtr rowIndice_; + std::vector> inputSeqInfoVec_; + std::vector outSubSeqStartPos_; + std::vector outSeqStartPos_; + + void checkInputs(); + void copySliceIdsToCpu(); + void calSelectedRows(const MatrixPtr starts, const MatrixPtr ends); +}; + +REGISTER_LAYER(seq_slice, SequenceSliceLayer); + +bool SequenceSliceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_GE(inputLayers_.size(), 2U); + CHECK_LE(inputLayers_.size(), 3U); + + setNeedSequenceInfo(false); + return true; +} + +void SequenceSliceLayer::checkInputs() { + const Argument& inputSeq = getInput(0); + CHECK(inputSeq.hasSeq()) << "The first input of sequence slice layer " + << "must be a sequence."; + const MatrixPtr indices1 = getInputValue(1); + CHECK_EQ( + indices1->getHeight(), + static_cast(inputSeq.hasSubseq() ? inputSeq.getNumSubSequences() + : inputSeq.getNumSequences())) + << "Height of the second input should be equal to number of sequence " + << "in the first input."; + if (inputLayers_.size() == 3) { + const MatrixPtr indices2 = getInputValue(2); + CHECK_EQ(indices2->getHeight(), indices1->getHeight()) + << "start indices and end indices should have the same height."; + CHECK_EQ(indices2->getWidth(), indices1->getWidth()) + << "start indices and end indices should have the same Width."; + } +} + +void SequenceSliceLayer::copySliceIdsToCpu() { + const MatrixPtr indices1 = getInputValue(1); + if (inputLayers_.size() == 2U) { + if (config_.select_first()) { + Matrix::resizeOrCreate(startIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + startIdsOnCpu_->copyFrom(*indices1); + endIdsOnCpu_ = nullptr; + } else { + Matrix::resizeOrCreate(endIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + endIdsOnCpu_->copyFrom(*indices1); + startIdsOnCpu_ = nullptr; + } + } else if (inputLayers_.size() == 3U) { + Matrix::resizeOrCreate(startIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + startIdsOnCpu_->copyFrom(*indices1); + + const MatrixPtr indices2 = getInputValue(2); + Matrix::resizeOrCreate(endIdsOnCpu_, + indices2->getHeight(), + indices2->getWidth(), + false /* trans */, + false /* useGpu */); + endIdsOnCpu_->copyFrom(*indices2); + } +} + +void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts, + const MatrixPtr ends) { + CHECK(starts || ends) << "At least one of the start or end indices " + << "should be given."; + + bool hasSubseq = getInput(0).hasSubseq(); + + outSeqStartPos_.resize(1, 0); + outSubSeqStartPos_.resize(1, 0); + selectedRows_.clear(); + + size_t beamSize = starts ? starts->getWidth() : ends->getWidth(); + size_t rowIdx = 0; + for (size_t i = 0; i < inputSeqInfoVec_.size(); ++i) { + for (size_t j = 0; j < inputSeqInfoVec_[i].size() - 1; ++j) { + for (size_t k = 0; k < beamSize; ++k) { + if (starts && starts->getElement(rowIdx, k) == -1.) break; + if (ends && ends->getElement(rowIdx, k) == -1.) break; + + int begPos = inputSeqInfoVec_[i][j]; + if (starts) begPos += starts->getElement(rowIdx, k); + + int endPos = inputSeqInfoVec_[i][j + 1] - 1; + if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k); + + int seqLen = endPos - begPos + 1; + CHECK_GT(seqLen, 0); + for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m); + hasSubseq + ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen) + : outSeqStartPos_.push_back(outSeqStartPos_.back() + seqLen); + } + rowIdx++; + } + if (hasSubseq) outSeqStartPos_.push_back(outSubSeqStartPos_.back()); + } + + if (useGpu_) { + rowIndice_ = IVector::create(selectedRows_.size(), useGpu_); + rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size()); + } else { + rowIndice_ = + IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_); + } + + // create the sequence information for the output. + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, outSeqStartPos_.size(), false); + output_.sequenceStartPositions->copyFrom( + outSeqStartPos_.data(), outSeqStartPos_.size(), false); + + if (hasSubseq) { + ICpuGpuVector::resizeOrCreate( + output_.subSequenceStartPositions, outSubSeqStartPos_.size(), false); + output_.subSequenceStartPositions->copyFrom( + outSubSeqStartPos_.data(), outSubSeqStartPos_.size(), false); + } +} + +void SequenceSliceLayer::forward(PassType passType) { + Layer::forward(passType); + checkInputs(); + + const Argument& inputSeq = getInput(0); + inputSeqInfoVec_.clear(); + Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, + inputSeq.subSequenceStartPositions, + inputSeqInfoVec_); + if (!useGpu_) { + if (inputLayers_.size() == 2U) { + startIdsOnCpu_ = config_.select_first() ? getInputValue(1) : nullptr; + endIdsOnCpu_ = config_.select_first() ? nullptr : getInputValue(1); + } else if (inputLayers_.size() == 3U) { + startIdsOnCpu_ = getInputValue(1); + endIdsOnCpu_ = getInputValue(2); + } + } else { + copySliceIdsToCpu(); + } + + /* + * calculate the selected row indices in a batch, and build the output + * sequence information. + */ + calSelectedRows(startIdsOnCpu_, endIdsOnCpu_); + + resetOutput(selectedRows_.size(), getSize()); + + getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); +} + +void SequenceSliceLayer::backward(const UpdateCallback& callback) { + getOutputGrad()->addToRows(*getInputGrad(0), *rowIndice_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/SequenceToBatch.cpp b/paddle/legacy/gserver/layers/SequenceToBatch.cpp similarity index 100% rename from paddle/gserver/layers/SequenceToBatch.cpp rename to paddle/legacy/gserver/layers/SequenceToBatch.cpp diff --git a/paddle/legacy/gserver/layers/SequenceToBatch.h b/paddle/legacy/gserver/layers/SequenceToBatch.h new file mode 100644 index 0000000000000000000000000000000000000000..7ed517937d4a015b6b11de16412cac7599f5f8b9 --- /dev/null +++ b/paddle/legacy/gserver/layers/SequenceToBatch.h @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" + +namespace paddle { + +/* + * This class can used to modify the matrix structure of sequence matrix into + * batch structure. + * sequence matrix: [C1_s ... Cn_s | ...... | C1_t ... Cn_t] + * batch matrix: [C1_s ... C1_t | ...... | Cn_s ... Cn_t] + * Cn_s is the state for sequence s at time n. + * + * Exampel: sequence matrix = {{0, 0, 0, 0}, {1, 1, 1, 1, 1}, {2, 2, 2}} + * s0: 0 0 0 0, s1: 1 1 1 1 1, s2: 2 2 2 + * batch matrix = {{1, 0, 2}, {1, 0, 2}, {1, 0, 2}, {1, 0}, {1}} + * b0: 1 0 2, b1: 1 0 2, b2: 1 0 2, b3: 1 0, b4: 1 + * + * Use: + * Input: seqMatrix, seqStarts(Sequence Start Positions) + * Output: batchMatrix + * 1. SequenceToBatch seq2batch; + * 2. seq2batch.resizeOrCreateBatch(seqStarts); // calculate seq2BatchIdx + * 3. seq2batch.copy(seqMatrix, batchMatrix, true); // copy seq to batch matrix + * + */ +class SequenceToBatch { + public: + explicit SequenceToBatch(bool useGpu) : useGpu_(useGpu) {} + + /* resize and calculate the batchIndex_ */ + void resizeOrCreateBatch(int batchSize, + size_t numSequences, + const int *seqStarts, + bool reversed, + bool prevBatchState = false); + + /* sequence matrix and batch matrix copy: + * seq2batch: copy(seqValue, batchValue, true); + * batch2seq: copy(seqValue, batchValue, false); + */ + void copy(Matrix &seqValue, Matrix &batchValue, bool seq2batch); + /* sequence/batch matrix add to batch/sequence matrix */ + void add(Matrix &seqValue, Matrix &batchValue, bool seq2batch); + MatrixPtr getBatchValue(Matrix &batchValue, int batchId, int numRows = 0); + + size_t getNumBatch() const { return numBatch_; } + + /* resize or create a batch matrix(batchValue_) */ + void resizeOrCreate(Matrix &seqValue); + /* copy seqValue to batchValue_ */ + void copyFromSeq(Matrix &seqValue); + /* copy batchValue_ to seqValue */ + void copyBackSeq(Matrix &seqValue); + MatrixPtr getBatchValue(int batchId, int numRows = 0); + MatrixPtr getBatchValue() { return batchValue_; } + /*tranfer preBatchOutput to batch struct*/ + void prevOutput2Batch(Matrix &src, Matrix &dst); + /*get sequence output from batch struct*/ + void getSeqOutputFromBatch(Matrix &sequence, Matrix &batch); + + /* Copy the index from another seq2batch. */ + void shareIndexWith(const SequenceToBatch &seq2batch) { + CHECK(useGpu_ == seq2batch.useGpu_); + batchStartPositions_ = seq2batch.batchStartPositions_; + seq2BatchIdx_ = seq2batch.seq2BatchIdx_; + cpuSeq2BatchIdx_ = seq2batch.cpuSeq2BatchIdx_; + numBatch_ = seq2batch.numBatch_; + } + + protected: + void sequence2BatchCopy(Matrix &batch, + Matrix &sequence, + IVector &seq2BatchIdx, + bool seq2batch); + void sequence2BatchAdd(Matrix &batch, + Matrix &sequence, + IVector &seq2BatchIdx, + bool seq2batch); + + IVectorPtr batchStartPositions_; + IVectorPtr seq2BatchIdx_; + IVectorPtr cpuSeq2BatchIdx_; + IVectorPtr cpuSeqIdx_; + IVectorPtr cpuSeqEndIdxInBatch_; + IVectorPtr seqIdx_; + IVectorPtr seqEndIdxInBatch_; + size_t numBatch_; + bool useGpu_; + MatrixPtr batchValue_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/SliceProjection.cpp b/paddle/legacy/gserver/layers/SliceProjection.cpp similarity index 100% rename from paddle/gserver/layers/SliceProjection.cpp rename to paddle/legacy/gserver/layers/SliceProjection.cpp diff --git a/paddle/legacy/gserver/layers/SlopeInterceptLayer.cpp b/paddle/legacy/gserver/layers/SlopeInterceptLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9168fd7dda6dcdcd9e272acbf6337f1c8468e6f0 --- /dev/null +++ b/paddle/legacy/gserver/layers/SlopeInterceptLayer.cpp @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief A layer for applying a slope and an intercept to the input + * element-wise. + * This layer is used in NEURAL TURING MACHINE. + * @note There is no activation and weight in this layer. + * + * \f[ + * y = ax + b + * \f] + * + * Here, a is scale and b is offset, which are provided as attributes of the + * layer. + * + * The config file api is slope_intercept_layer. + */ + +class SlopeInterceptLayer : public Layer { + public: + explicit SlopeInterceptLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(slope_intercept, SlopeInterceptLayer); + +bool SlopeInterceptLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 1U); + + return true; +} + +void SlopeInterceptLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV = getInputValue(0); + + /* malloc memory for the output_ if necessary */ + size_t batchSize = inV->getHeight(); + size_t size = getSize(); + + CHECK_EQ(size, inV->getWidth()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + reserveOutput(batchSize, size); + } + + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwSlopeInterceptTimer", getName().c_str()); + outV->mulScalar(*inV, config_.slope()); + outV->add(config_.intercept()); + } +} + +void SlopeInterceptLayer::backward(const UpdateCallback& callback) { + MatrixPtr inG = getInputGrad(0); + MatrixPtr outG = getOutputGrad(); + + if (inG) { + REGISTER_TIMER_INFO("BwSlopeInterceptTimer", getName().c_str()); + inG->add(*outG, config_.slope()); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/SpatialPyramidPoolLayer.cpp b/paddle/legacy/gserver/layers/SpatialPyramidPoolLayer.cpp similarity index 100% rename from paddle/gserver/layers/SpatialPyramidPoolLayer.cpp rename to paddle/legacy/gserver/layers/SpatialPyramidPoolLayer.cpp diff --git a/paddle/legacy/gserver/layers/SpatialPyramidPoolLayer.h b/paddle/legacy/gserver/layers/SpatialPyramidPoolLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..6d8ed9c87889a93664f09dbaf2a84bd00b1757ad --- /dev/null +++ b/paddle/legacy/gserver/layers/SpatialPyramidPoolLayer.h @@ -0,0 +1,59 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "PoolProjection.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { +/** + * @brief A layer for spatial pyramid pooling on the input image by taking + * the max, average, etc. within regions, so that the result vector of + * different sized images are of the same size. + * + * The config file api is spp_layer. + */ + +class SpatialPyramidPoolLayer : public Layer { + protected: + size_t channels_; + size_t imgSizeW_; + size_t imgSizeH_; + size_t pyramidHeight_; + std::string poolType_; + + std::vector> poolProjections_; + std::vector projOutput_; + std::vector> projCol_; + + public: + explicit SpatialPyramidPoolLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + ProjectionConfig getConfig(size_t sizeX_, + size_t sizeY_, + size_t channels, + size_t pyamidLevel_, + std::string& poolType_); + size_t getSize(); + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SubNestedSequenceLayer.cpp b/paddle/legacy/gserver/layers/SubNestedSequenceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f363c2ac8dd22fc8b8e1d7fca27e5beb935d42de --- /dev/null +++ b/paddle/legacy/gserver/layers/SubNestedSequenceLayer.cpp @@ -0,0 +1,187 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +class SubNestedSequenceLayer : public Layer { + public: + explicit SubNestedSequenceLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + + private: + /* + * This functions generates the indices of rows in a batch according to the + * indices of selected sub-sequence in each sequence. + * + * Examples: + * selectedIndices: + * [ + * [0, 1, -1], + * [0, 1, 2], + * [0, -1, -1], + * [0, 2, 3], + * ] + * inputSeqInfo: + * [ + * [0,3,4], + * [4,5,7,10,15], + * [15,20], + * [20,22,23,25,28] + * ] + * + * ths output is saved to private member rowIndice_; + * [0,1,2,3,4,5,6,7,8,9,15,16,17,18,19,20,21,23,24,25,26,27] + */ + + void calSelectedRows(const MatrixPtr selectedIndices, + const std::vector>& inputSeqInfo); + + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but the second is some selected indices of the give sequence to trim + * the nested sequence, are actually filled with int types so that storing + * int types information in real number matrices is very dangerous, since + * real numbers will be convered to int types. If a user fills this matrix + * himself, invalid data may occor. + * + * if the second input of this layer is on GPU memory, copy it to CPU memory. + */ + MatrixPtr selIdsCpu_; + + /* + * reorganize sequenceStartPositions and subSequenceStartPositions + * into a 2d vector to facilitate the sequence selection process. + */ + std::vector> inputSeqInfoVec_; + + /* store the final selected row indices in a batch */ + IVectorPtr rowIndice_; + /* rowIndice_ and selectedRows_ actually share a same memory. */ + std::vector selectedRows_; +}; + +REGISTER_LAYER(sub_nested_seq, SubNestedSequenceLayer); + +bool SubNestedSequenceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_EQ(2U, inputLayers_.size()); + setNeedSequenceInfo(false); + return true; +} + +void SubNestedSequenceLayer::calSelectedRows( + const MatrixPtr selectedIndices, + const std::vector>& inputSeqInfo) { + selectedRows_.clear(); + + std::vector outSeqStartInfo(1, 0); + std::vector outSubSeqStartInfo(1, 0); + + size_t seqNum = selectedIndices->getHeight(); + size_t beamSize = selectedIndices->getWidth(); + for (size_t i = 0; i < seqNum; ++i) { + for (size_t j = 0; j < beamSize; ++j) { + if (selectedIndices->getElement(i, j) == -1.) break; + size_t selSubSeqIdx = selectedIndices->getElement(i, j); + CHECK_GT(inputSeqInfoVec_[i].size() - 1, selSubSeqIdx); + + size_t subSeqLen = inputSeqInfoVec_[i][selSubSeqIdx + 1] - + inputSeqInfoVec_[i][selSubSeqIdx]; + for (size_t k = 0; k < subSeqLen; ++k) + selectedRows_.push_back(inputSeqInfoVec_[i][selSubSeqIdx] + k); + outSubSeqStartInfo.push_back(outSubSeqStartInfo.back() + subSeqLen); + } + outSeqStartInfo.push_back(outSubSeqStartInfo.back()); + } + + if (useGpu_) { + rowIndice_ = IVector::create(selectedRows_.size(), useGpu_); + rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size()); + } else { + rowIndice_ = + IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_); + } + + // create the sequence information for the output. + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, outSeqStartInfo.size(), false); + output_.sequenceStartPositions->copyFrom( + outSeqStartInfo.data(), outSeqStartInfo.size(), false); + + ICpuGpuVector::resizeOrCreate( + output_.subSequenceStartPositions, outSubSeqStartInfo.size(), false); + output_.subSequenceStartPositions->copyFrom( + outSubSeqStartInfo.data(), outSubSeqStartInfo.size(), false); +} + +void SubNestedSequenceLayer::forward(PassType passType) { + Layer::forward(passType); + + const Argument& inputSeq = getInput(0); + CHECK(inputSeq.hasSubseq()) << "The first input of SubNestSequence layer " + << "must be a nested sequence."; + const MatrixPtr selectedIndices = getInputValue(1); + CHECK_EQ(size_t(inputSeq.getNumSequences()), selectedIndices->getHeight()); + + if (dynamic_cast(selectedIndices.get())) { + /* + * Currently, the second input for this layer is generated by + * kmax_sequence_score_layer whose output is always stored on CPU, + * or a data_layer which canbe on GPU. + * + * If the second input is on GPU, copy it to CPU memory, because this + * input always uses very few memory, and operations related to it are + * all logic control, not computations. + */ + Matrix::resizeOrCreate(selIdsCpu_, + selectedIndices->getHeight(), + selectedIndices->getWidth(), + false /* trans */, + false /* useGpu */); + selIdsCpu_->copyFrom(*selectedIndices); + } else { + selIdsCpu_ = selectedIndices; + } + + Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, + inputSeq.subSequenceStartPositions, + inputSeqInfoVec_); + calSelectedRows(selIdsCpu_, inputSeqInfoVec_); + + resetOutput(selectedRows_.size(), getSize()); + getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); +} + +void SubNestedSequenceLayer::backward(const UpdateCallback& callback) { + MatrixPtr inputSeqGrad = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + + if (inputSeqGrad) outputGrad->addToRows(*inputSeqGrad, *rowIndice_); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SubSequenceLayer.cpp b/paddle/legacy/gserver/layers/SubSequenceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..36796f04739054bb19d4a3ce656e248898ba4b17 --- /dev/null +++ b/paddle/legacy/gserver/layers/SubSequenceLayer.cpp @@ -0,0 +1,226 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for taking the subsequence according to given offset and size + * Input: original sequence, offset, size + * Output: subsequence + */ + +class SubSequenceLayer : public Layer { + protected: + std::unique_ptr biases_; + MatrixPtr tmpSrc_; + MatrixPtr tmpDest_; + + public: + explicit SubSequenceLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(subseq, SubSequenceLayer); + +bool SubSequenceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + // sequene concatenation layer should have exactly 2 inputs + CHECK_EQ(3U, inputLayers_.size()); + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + + tmpSrc_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); + tmpDest_ = + Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); + + setNeedSequenceInfo(false); + return true; +} + +void SubSequenceLayer::forward(PassType passType) { + Layer::forward(passType); + + size_t dim = getSize(); + + const Argument& input = getInput(0); + size_t numSequences1 = input.getNumSequences(); + auto startPositions1 = input.sequenceStartPositions->getVector(false); + + const Argument& offsetSeq = getInput(1); + size_t numSequences2 = offsetSeq.getNumSequences(); + auto startPositions2 = offsetSeq.sequenceStartPositions->getVector(false); + + const Argument& sizeSeq = getInput(2); + size_t numSequences3 = sizeSeq.getNumSequences(); + auto startPositions3 = sizeSeq.sequenceStartPositions->getVector(false); + + CHECK_EQ(dim, input.value->getWidth()); + + CHECK_EQ(startPositions1->getData()[numSequences1], input.getBatchSize()); + CHECK_EQ(numSequences1, startPositions1->getSize() - 1); + + CHECK_EQ(startPositions2->getData()[numSequences2], offsetSeq.getBatchSize()); + CHECK_EQ(numSequences2, startPositions2->getSize() - 1); + + CHECK_EQ(startPositions3->getData()[numSequences3], sizeSeq.getBatchSize()); + CHECK_EQ(numSequences3, startPositions3->getSize() - 1); + + CHECK_EQ(numSequences1, numSequences2); + CHECK_EQ(numSequences2, numSequences3); + + MatrixPtr inputValue = input.value; + IVectorPtr offsetValue; + IVectorPtr sizeValue; + + if (useGpu_) { + // copy to cpu + IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false); + IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false); + offsetValue->copyFrom(*offsetSeq.ids); + sizeValue->copyFrom(*sizeSeq.ids); + } else { + offsetValue = offsetSeq.ids; + sizeValue = sizeSeq.ids; + } + + CHECK_EQ(offsetValue->getSize(), numSequences1); + CHECK_EQ(sizeValue->getSize(), numSequences1); + + int* offsets = offsetValue->getData(); + int* sizes = sizeValue->getData(); + + // get total height of output + size_t height = 0; + for (size_t seqId = 0; seqId < numSequences1; seqId++) { + height += sizes[seqId]; + } + + // reset output + resetOutput(height, dim); + + MatrixPtr outputValue = getOutputValue(); + + const int* starts1 = startPositions1->getData(); + + { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SubSequenceLayerForward", getName().c_str()); + + size_t offsetIn = 0; + size_t offsetOut = 0; + size_t size = 0; + for (size_t seqId = 0; seqId < numSequences1; ++seqId) { + offsetIn = starts1[seqId] + offsets[seqId]; + size = sizes[seqId]; + + outputValue->subMatrix(offsetOut, size, tmpDest_) + ->assign(*(inputValue->subMatrix(offsetIn, size, tmpSrc_))); + + offsetOut += size; + } + + // modify the sequenceStartPositions + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, numSequences1 + 1, false); + + int* tgtBuf = output_.sequenceStartPositions->getMutableData(false); + int offset = 0; + for (size_t seqId = 0; seqId < numSequences1; ++seqId) { + tgtBuf[seqId] = offset; + offset += sizes[seqId]; + } + tgtBuf[numSequences1] = offset; + } + + if (biases_.get() != NULL) { + MatrixPtr outV = getOutputValue(); + outV->addBias(*(biases_->getW()), 1); + } + + /* activation */ + forwardActivation(); +} + +void SubSequenceLayer::backward(const UpdateCallback& callback) { + /* activation */ + backwardActivation(); + + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + // Increasing the number of gradient + biases_->getParameterPtr()->incUpdate(callback); + } + + MatrixPtr inputGrad1 = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + auto startPositions1 = getInput(0).sequenceStartPositions->getVector(false); + size_t numSequences1 = startPositions1->getSize() - 1; + const int* starts1 = startPositions1->getData(); + + const Argument& offsetSeq = getInput(1); + const Argument& sizeSeq = getInput(2); + IVectorPtr offsetValue; + IVectorPtr sizeValue; + + if (useGpu_) { + // copy to cpu + IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false); + IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false); + offsetValue->copyFrom(*offsetSeq.ids); + sizeValue->copyFrom(*sizeSeq.ids); + } else { + offsetValue = offsetSeq.ids; + sizeValue = sizeSeq.ids; + } + + int* offsets = offsetValue->getData(); + int* sizes = sizeValue->getData(); + { + AsyncGpuBlock asyncGpuBlock; + REGISTER_TIMER_INFO("SubSequenceLayerBackward", getName().c_str()); + + int offsetIn = 0; + int offsetOut = 0; + int size = 0; + for (size_t seqId = 0; seqId < numSequences1; ++seqId) { + offsetIn = starts1[seqId] + offsets[seqId]; + size = sizes[seqId]; + + inputGrad1->subMatrix(offsetIn, size, tmpDest_) + ->add(*(outputGrad->subMatrix(offsetOut, size, tmpSrc_))); + offsetOut += size; + } + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SumToOneNormLayer.cpp b/paddle/legacy/gserver/layers/SumToOneNormLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..410f4dd7c90e67488bc3dda6dfad551032890d65 --- /dev/null +++ b/paddle/legacy/gserver/layers/SumToOneNormLayer.cpp @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * A layer for sum-to-one normalization, + * which is used in NEURAL TURING MACHINE. + * \f[ + * out[i] = \frac {in[i]} {\sum_{k=1}^N in[k]} + * \f] + * where \f$in\f$ is a (batchSize x dataDim) input vector, + * and \f$out\f$ is a (batchSize x dataDim) output vector. + * + * The config file api is sum_to_one_norm_layer. + */ + +class SumToOneNormLayer : public Layer { + protected: + /// reciprocalRowSum_ = \f$1 / \sum_{k=1}^N in[k]\f$ + MatrixPtr reciprocalRowSum_; + /// dotSum = output_.grad \f$.*\f$ output_.value + MatrixPtr dotSum_; + + public: + explicit SumToOneNormLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(sum_to_one_norm, SumToOneNormLayer); + +bool SumToOneNormLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + CHECK_EQ(inputLayers_.size(), 1U); + + return true; +} + +void SumToOneNormLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV = getInputValue(0); + + /* malloc memory for the output_ if necessary */ + size_t batchSize = inV->getHeight(); + size_t dataDim = getSize(); + + CHECK_EQ(dataDim, inV->getWidth()); + + { + REGISTER_TIMER_INFO("FwResetTimer", getName().c_str()); + resetOutput(batchSize, dataDim); + } + + MatrixPtr outV = getOutputValue(); + { + REGISTER_TIMER_INFO("FwSumToOneNormTimer", getName().c_str()); + + Matrix::resizeOrCreate(reciprocalRowSum_, batchSize, 1, false, useGpu_); + inV->rowSum(*reciprocalRowSum_); + + // todo: matrix checks + CHECK_GT(reciprocalRowSum_->getMin(), 0.0); + + reciprocalRowSum_->scalarDiv(*reciprocalRowSum_, 1.0); + + // outV = inV * reciprocalRowSum + outV->rowScale(0, *inV, *reciprocalRowSum_); + } +} + +void SumToOneNormLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV = getInputValue(0); + MatrixPtr inG = getInputGrad(0); + MatrixPtr outV = getOutputValue(); + MatrixPtr outG = getOutputGrad(); + + size_t batchSize = inV->getHeight(); + + if (inG) { + REGISTER_TIMER_INFO("BwSumToOneTimer", getName().c_str()); + + Matrix::resizeOrCreate(dotSum_, batchSize, 1, false, useGpu_); + + // dotSum = outG .* outV + dotSum_->zeroMem(); + dotSum_->rowDotMul(0, *outG, *outV); + + // inG += -1 * (dotSum / rowSum) + dotSum_->dotMul(*dotSum_, *reciprocalRowSum_); + inG->rowAdd(0, *inG, *dotSum_, -1.0); + // inG += outG * (1/rowSum) + inG->addRowScale(0, *outG, *reciprocalRowSum_); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/SwitchOrderLayer.cpp b/paddle/legacy/gserver/layers/SwitchOrderLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..513f3df7bcaf854835ec0e500d47c23469d5aa46 --- /dev/null +++ b/paddle/legacy/gserver/layers/SwitchOrderLayer.cpp @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SwitchOrderLayer.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(switch_order, SwitchOrderLayer); + +bool SwitchOrderLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + auto& img_conf = config_.inputs(0).image_conf(); + size_t inD = img_conf.img_size_z(); + size_t inH = + img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(); + size_t inW = img_conf.img_size(); + size_t inC = img_conf.channels(); + inH = inH * inD; + inDims_ = TensorShape({0, inC, inH, inW}); + outDims_ = TensorShape(4); + + auto& reshape_conf = config_.reshape_conf(); + for (int i = 0; i < reshape_conf.height_axis_size(); i++) { + heightAxis_.push_back(reshape_conf.height_axis(i)); + } + for (int i = 0; i < reshape_conf.width_axis_size(); i++) { + widthAxis_.push_back(reshape_conf.width_axis(i)); + } + createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig()); + createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig()); + return true; +} + +void SwitchOrderLayer::setOutDims() { + outDims_.setDim(0, inDims_[0]); + outDims_.setDim(1, inDims_[2]); + outDims_.setDim(2, inDims_[3]); + outDims_.setDim(3, inDims_[1]); + reshapeHeight_ = 1; + for (size_t i = 0; i < heightAxis_.size(); i++) { + reshapeHeight_ *= outDims_[heightAxis_[i]]; + } + output_.setFrameHeight(reshapeHeight_); + reshapeWidth_ = 1; + for (size_t i = 0; i < widthAxis_.size(); i++) { + reshapeWidth_ *= outDims_[widthAxis_[i]]; + } + output_.setFrameWidth(reshapeWidth_); +} + +void SwitchOrderLayer::setInDims() { + MatrixPtr input = inputLayers_[0]->getOutputValue(); + size_t batchSize = input->getHeight(); + inDims_.setDim(0, batchSize); + int d = inputLayers_[0]->getOutput().getFrameDepth(); + d = (d == 0 ? 1 : d); + int h = inputLayers_[0]->getOutput().getFrameHeight(); + if (h != 0) inDims_.setDim(2, h * d); + int w = inputLayers_[0]->getOutput().getFrameWidth(); + if (w != 0) inDims_.setDim(3, w); + int totalCount = input->getElementCnt(); + int channels = totalCount / (inDims_[0] * inDims_[2] * inDims_[3]); + if (channels != 0) inDims_.setDim(1, channels); +} + +void SwitchOrderLayer::forward(PassType passType) { + Layer::forward(passType); + setInDims(); + setOutDims(); + resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]); + if (heightAxis_.size() > 0) { + resetOutput(reshapeHeight_, reshapeWidth_); + } + + // switch NCHW to NHWC + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), inDims_); + outputs.addArg(*getOutputValue(), outDims_); + nchw2nhwc_[0]->calc(inputs, outputs); + forwardActivation(); +} + +void SwitchOrderLayer::backward(const UpdateCallback& callback) { + (void)callback; + backwardActivation(); + + // switch NHWC to NCHW + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outDims_); + outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); + nhwc2nchw_[0]->calc(inputs, outputs); +} +} // namespace paddle diff --git a/paddle/gserver/layers/SwitchOrderLayer.h b/paddle/legacy/gserver/layers/SwitchOrderLayer.h similarity index 100% rename from paddle/gserver/layers/SwitchOrderLayer.h rename to paddle/legacy/gserver/layers/SwitchOrderLayer.h diff --git a/paddle/gserver/layers/TableProjection.cpp b/paddle/legacy/gserver/layers/TableProjection.cpp similarity index 100% rename from paddle/gserver/layers/TableProjection.cpp rename to paddle/legacy/gserver/layers/TableProjection.cpp diff --git a/paddle/gserver/layers/TableProjection.h b/paddle/legacy/gserver/layers/TableProjection.h similarity index 100% rename from paddle/gserver/layers/TableProjection.h rename to paddle/legacy/gserver/layers/TableProjection.h diff --git a/paddle/legacy/gserver/layers/TensorLayer.cpp b/paddle/legacy/gserver/layers/TensorLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7f874bce0f2bdf7ab4771e470e2e4535693ecf68 --- /dev/null +++ b/paddle/legacy/gserver/layers/TensorLayer.cpp @@ -0,0 +1,145 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "TensorLayer.h" + +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(tensor, TensorLayer); + +bool TensorLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* initialize the weightList */ + CHECK_EQ(inputLayers_.size(), 2LU); + CHECK(parameters_[0]); + CHECK(!parameters_[1]); + + // Option the parameters + size_t height = inputLayers_[0]->getSize(); + size_t width = inputLayers_[1]->getSize(); + CHECK_EQ(width * height * getSize(), parameters_[0]->getSize()); + + for (size_t i = 0; i < getSize(); ++i) { + // create a new weight + Weight* w = new Weight(height, width, parameters_[0], i * width * height); + + // append the new weight to the list + weights_.emplace_back(w); + } + + /* initialize biases_ */ + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + } + + return true; +} + +void TensorLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + int batchSize = getInputValue(0)->getHeight(); + int size = getSize(); + + { resetOutput(batchSize, size); } + + MatrixPtr outV = getOutputValue(); + /* add the bias-vector */ + if (biases_.get() != NULL) { + outV->addBias(*(biases_->getW()), 1); + } + + /* e1 * W * trans(e2) */ { + MatrixPtr input1 = getInputValue(0); + MatrixPtr input2 = getInputValue(1); + MatrixPtr tmpMat = Matrix::create(input2->getHeight(), + input2->getWidth(), + /* trans= */ false, + input2->useGpu()); + REGISTER_TIMER_INFO("TensorFwMulTimer", getName().c_str()); + for (size_t i = 0; i < getSize(); ++i) { + MatrixPtr weights = weights_[i]->getW(); + tmpMat->mul(*input1, *weights, 1, 0); + outV->rowDotMul(i, *tmpMat, *input2); + } + } + + /* activation */ { forwardActivation(); } +} + +void TensorLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { backwardActivation(); } + + if (biases_ && biases_->getWGrad()) { + biases_->getWGrad()->collectBias(*getOutputGrad(), 1); + + /* Increasing the number of gradient */ + biases_->getParameterPtr()->incUpdate(callback); + } + + bool syncFlag = hl_get_sync_flag(); + + /* Calculate the W-gradient for the current layer */ + MatrixPtr input1 = getInputValue(0); + MatrixPtr input2 = getInputValue(1); + MatrixPtr oGrad = getOutputGrad(); + MatrixPtr tmpMat = Matrix::create(input1->getHeight(), + input1->getWidth(), + /* trans= */ false, + input1->useGpu()); + + /* trans(grad * e1) * e2 */ { + REGISTER_TIMER_INFO("TensorGradMulTimer", getName().c_str()); + for (size_t i = 0; i < getSize(); ++i) { + if (weights_[i]->getWGrad()) { + tmpMat->rowScale(i, *input1, *oGrad); + MatrixPtr input1_T = tmpMat->getTranspose(); + weights_[i]->getWGrad()->mul(*input1_T, *input2, 1, 1); + } + } + } + + hl_set_sync_flag(false); + + /* Calculate the input layers error */ { + MatrixPtr preGrad1 = getInputGrad(0); + MatrixPtr preGrad2 = getInputGrad(1); + + REGISTER_TIMER_INFO("TensorBpMulTimer", getName().c_str()); + for (size_t i = 0; i < getSize(); ++i) { + MatrixPtr weights = weights_[i]->getW(); + + if (NULL != preGrad1) { /* (grad * e2) * trans(W) */ + tmpMat->rowScale(i, *input2, *oGrad); + MatrixPtr weights_T = weights->getTranspose(); + preGrad1->mul(*tmpMat, *weights_T, 1, 1); + } + if (NULL != preGrad2) { /* (grad * e1) * W */ + tmpMat->rowScale(i, *input1, *oGrad); + preGrad2->mul(*tmpMat, *weights, 1, 1); + } + } + } + hl_set_sync_flag(syncFlag); + parameters_[0]->incUpdate(callback); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/TensorLayer.h b/paddle/legacy/gserver/layers/TensorLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..fc491a7c9f223cf0dff6d878c6ec27a858c7c7b7 --- /dev/null +++ b/paddle/legacy/gserver/layers/TensorLayer.h @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/** + * @brief TensorLayer takes two input vectors. + * \f[ + * y_{i} = x_{1} * W_{i} * x_{2}^{\rm T}, i=0, 1, ...,K-1 + * \f] + * + * - \f$x_{1}\f$: the first input, size is M. + * - \f$x_{2}\f$: the second input, size is N. + * - y: output, size is K. + * - \f$y_{i}\f$: i-th element of y. + * - \f$W_{i}\f$: the i-th learned weight, dimensions: [M, N]. + * - \f$x_{2}^{\rm T}\f$: the transpose of \f$x_{2}\f$. + * + * The config file api is tensor_layer. + */ + +class TensorLayer : public Layer { + protected: + WeightList weights_; + std::unique_ptr biases_; + + public: + explicit TensorLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + Weight& getWeight(int idx) { return *weights_[idx]; } + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/TransLayer.cpp b/paddle/legacy/gserver/layers/TransLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..fd1d435ea5f53785c9c416146c642637adc786a8 --- /dev/null +++ b/paddle/legacy/gserver/layers/TransLayer.cpp @@ -0,0 +1,69 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "TransLayer.h" +#include "paddle/legacy/utils/Logging.h" +namespace paddle { + +REGISTER_LAYER(trans, TransLayer); + +bool TransLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + /* the size of inputs for trans-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + + return true; +} + +void TransLayer::forward(PassType passType) { + Layer::forward(passType); + + /* malloc memory for the output_ if necessary */ + MatrixPtr input = getInputValue(0); + int height = input->getHeight(); + int width = input->getWidth(); + + resizeOutput(width, height); + + MatrixPtr outV = getOutputValue(); + + /* outV's memory has been allocated, so memAlloc = false */ + input->transpose(outV, false); + if (getInputGrad(0)) { + zeroGrad(); + } +} + +void TransLayer::backward(const UpdateCallback& callback) { + (void)callback; + + MatrixPtr outputGrad = getOutputGrad(); + if (outputGrad == NULL) { + return; + } + MatrixPtr preGrad = getInputGrad(0); + if (preGrad) { + MatrixPtr transGrad = Matrix::create(preGrad->getHeight(), + preGrad->getWidth(), + /* trans= */ false, + preGrad->useGpu()); + outputGrad->transpose(transGrad, false); + preGrad->add(*transGrad); + } +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/TransLayer.h b/paddle/legacy/gserver/layers/TransLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..0a6b13933f83f30a07ed63d722dbb612c64edae7 --- /dev/null +++ b/paddle/legacy/gserver/layers/TransLayer.h @@ -0,0 +1,41 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" + +namespace paddle { +/** + * A layer for transposing a minibatch matrix. + * \f[ + y = x^\mathrm{T} + * \f] + * where \f$x\f$ is (M x N) input, and \f$y\f$ is (N x M) output. + * + * The config file api is trans_layer. + */ +class TransLayer : public Layer { + public: + explicit TransLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/TransposedFullMatrixProjection.cpp b/paddle/legacy/gserver/layers/TransposedFullMatrixProjection.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c8533dc7d78ec4fd3629e29e6c1c3e73c6acdc17 --- /dev/null +++ b/paddle/legacy/gserver/layers/TransposedFullMatrixProjection.cpp @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Projection.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * @brief TransposedFullMatrixProjection performs full matrix multiplication: + * out.row[i] += in.row[i] * weight.transpose + * + * The config file api is trans_full_matrix_projection. + */ +class TransposedFullMatrixProjection : public Projection { + public: + TransposedFullMatrixProjection(const ProjectionConfig& config, + ParameterPtr parameter, + bool useGPu); + virtual void forward(); + virtual void backward(const UpdateCallback& callback); + + protected: + std::unique_ptr weight_; +}; + +REGISTER_PROJECTION(trans_fc, TransposedFullMatrixProjection); + +TransposedFullMatrixProjection::TransposedFullMatrixProjection( + const ProjectionConfig& config, ParameterPtr parameter, bool useGpu) + : Projection(config, parameter, useGpu) { + weight_.reset( + new Weight(config.output_size(), config.input_size(), parameter)); +} + +void TransposedFullMatrixProjection::forward() { + REGISTER_TIMER_INFO("FwMulTimer", getName().c_str()); + out_->value->mul(*(in_->value), *(weight_->getW()->getTranspose()), 1, 1); +} + +void TransposedFullMatrixProjection::backward(const UpdateCallback& callback) { + bool syncFlag = hl_get_sync_flag(); + + /* Calculate the W-gradient for the current layer */ + if (weight_->getWGrad()) { + REGISTER_TIMER_INFO("GradMulTimer", getName().c_str()); + weight_->getWGrad()->mul( + *(out_->grad->getTranspose()), *(in_->value), 1, 1); + } + + // If callback does not change value, backprop error asynchronously so that + // we can do the callback concurrently. + // This is still a little bit dangerous since theoretically for + // SyncMultiGpuMachine it is possible that the value copyback can still + // happen at the same time as the error backprop where the value is being + // used. + hl_set_sync_flag(false); + + /* Calculate the input layers error */ + if (in_->grad) { + REGISTER_TIMER_INFO("BpMulTimer", getName().c_str()); + in_->grad->mul(*(out_->grad), *(weight_->getW()), 1, 1); + } + + hl_set_sync_flag(syncFlag); + parameter_->incUpdate(callback); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/UpsampleLayer.cpp b/paddle/legacy/gserver/layers/UpsampleLayer.cpp similarity index 100% rename from paddle/gserver/layers/UpsampleLayer.cpp rename to paddle/legacy/gserver/layers/UpsampleLayer.cpp diff --git a/paddle/legacy/gserver/layers/UpsampleLayer.h b/paddle/legacy/gserver/layers/UpsampleLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..2fe5938244c81ab25c66083cc1ad63ba15618aa1 --- /dev/null +++ b/paddle/legacy/gserver/layers/UpsampleLayer.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Stat.h" + +namespace paddle { + +/** + * This layer transpose the pooling process. + * It takes two input, the first input is the input data, and + * the second is the mask data from the max-pool-with-mask layer. + * + */ + +class UpsampleLayer : public Layer { + public: + explicit UpsampleLayer(const LayerConfig& config) : Layer(config) {} + ~UpsampleLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + + size_t getOutputSize(); + + protected: + size_t scale_, scaleY_; + size_t upsampleSize_, upsampleSizeY_; + size_t padOutX_, padOutY_; + size_t imgSize_, imgSizeY_; + size_t channels_; +}; + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ValidationLayer.cpp b/paddle/legacy/gserver/layers/ValidationLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9956fd2ed41464eae096911620e160f5ecd89da3 --- /dev/null +++ b/paddle/legacy/gserver/layers/ValidationLayer.cpp @@ -0,0 +1,171 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include + +#include "ValidationLayer.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +bool ValidationLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return Layer::init(layerMap, parameterMap); +} + +void ValidationLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr output = getInputValue(*getOutputLayer()); + CHECK(output); + IVectorPtr label = getInputLabel(*getLabelLayer()); + CHECK(label); + validationImp(output, label); +} + +void ValidationLayer::backward(const UpdateCallback& callback) { + (void)callback; +} + +bool AucValidation::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + bool ret = ValidationLayer::init(layerMap, parameterMap); + EvaluatorConfig config; + config.set_name(getName()); + config.set_type("last-column-auc"); + config.add_input_layers(inputLayers_[0]->getName()); + config.add_input_layers(inputLayers_[1]->getName()); + if (3 == inputLayers_.size()) { + config.add_input_layers(inputLayers_[2]->getName()); + } + evaluator_.reset(Evaluator::create(config)); + passBegin_ = false; + return ret; +} + +void AucValidation::validationImp(MatrixPtr output, IVectorPtr label) { + if (!passBegin_) { + passBegin_ = true; + evaluator_->start(); + } + + bool supportWeight = (3 == inputLayers_.size()) ? true : false; + MatrixPtr weight = supportWeight ? getInputValue(*inputLayers_[2]) : nullptr; + if (dynamic_cast(output.get())) { + size_t height = output->getHeight(); + size_t width = output->getWidth(); + Matrix::resizeOrCreate(cpuOutput_, + height, + width, + /* trans=*/false, + /* useGpu=*/false); + cpuOutput_->copyFrom(*output); + IVector::resizeOrCreate(cpuLabel_, height, false); + cpuLabel_->copyFrom(*label); + + if (supportWeight) { + Matrix::resizeOrCreate(cpuWeight_, height, (size_t)1, false, false); + cpuWeight_->copyFrom(*weight); + } + + output = cpuOutput_; + label = cpuLabel_; + weight = cpuWeight_; + } + + for (size_t i = 0; i < output->getHeight(); i++) { + float y1 = output->getData()[i * output->getWidth() + 1]; + int* labels = label->getData(); + predictArray_.push_back(PredictionResult(y1, labels[i])); + } + std::vector arguments; + if (3 == inputLayers_.size()) { + arguments.resize(3); + arguments[2].value = weight; + } else { + arguments.resize(2); + } + arguments[0].value = output; + arguments[1].ids = label; + evaluator_->evalImp(arguments); +} + +void AucValidation::onPassEnd() { + if (!FLAGS_predict_file.empty()) { + std::ofstream fs(FLAGS_predict_file); + CHECK(fs) << "Fail to open " << FLAGS_predict_file; + for (auto& res : predictArray_) { + fs << res.out << " " << res.label << std::endl; + } + } + + evaluator_->finish(); + LOG(INFO) << *evaluator_; + passBegin_ = false; + predictArray_.clear(); +} + +bool PnpairValidation::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + bool ret = ValidationLayer::init(layerMap, parameterMap); + if (!ret) return ret; + CHECK_GE(inputLayers_.size(), 3UL); + CHECK_LE(inputLayers_.size(), 4UL); + EvaluatorConfig config; + config.set_name(getName()); + config.set_type("pnpair"); + config.add_input_layers(inputLayers_[0]->getName()); + config.add_input_layers(inputLayers_[1]->getName()); + config.add_input_layers(inputLayers_[2]->getName()); + if (4 == inputLayers_.size()) { + config.add_input_layers(inputLayers_[3]->getName()); + } + evaluator_.reset(Evaluator::create(config)); + passBegin_ = false; + return true; +} + +void PnpairValidation::validationImp(MatrixPtr output, IVectorPtr label) { + if (!passBegin_) { + passBegin_ = true; + evaluator_->start(); + } + MatrixPtr weight = + (4 == inputLayers_.size()) ? getInputValue(*inputLayers_[3]) : nullptr; + IVectorPtr info = getInputLabel(*getInfoLayer()); + std::vector arguments; + if (4 == inputLayers_.size()) { + arguments.resize(4); + arguments[3].value = weight; + } else { + arguments.resize(3); + } + arguments[0].value = output; + arguments[1].ids = label; + arguments[2].ids = info; + evaluator_->evalImp(arguments); +} + +void PnpairValidation::onPassEnd() { + if (!FLAGS_predict_file.empty()) { + (dynamic_cast(evaluator_.get()))->printPredictResults(); + } + evaluator_->finish(); + LOG(INFO) << *evaluator_; + passBegin_ = false; +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/layers/ValidationLayer.h b/paddle/legacy/gserver/layers/ValidationLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..fbc94e8ef570e2eec1d3737aca97bbf91c1392b2 --- /dev/null +++ b/paddle/legacy/gserver/layers/ValidationLayer.h @@ -0,0 +1,104 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include + +#include "Layer.h" +#include "paddle/legacy/gserver/evaluators/Evaluator.h" + +DECLARE_int32(trainer_id); + +namespace paddle { + +class ValidationLayer : public Layer { + public: + explicit ValidationLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + LayerPtr getOutputLayer() { return inputLayers_[0]; } + + LayerPtr getLabelLayer() { return inputLayers_[1]; } + + LayerPtr getInfoLayer() { + assert(inputLayers_.size() > 2); + return inputLayers_[2]; + } + + void forward(PassType passType) override; + + void backward(const UpdateCallback& callback = nullptr) override; + + virtual void validationImp(MatrixPtr outputValue, IVectorPtr label) = 0; + + void onPassEnd() override = 0; +}; + +/* + * AucValidation + */ +class AucValidation : public ValidationLayer { + public: + explicit AucValidation(const LayerConfig& config) + : ValidationLayer(config), + cpuOutput_(nullptr), + cpuLabel_(nullptr), + cpuWeight_(nullptr) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void validationImp(MatrixPtr outputValue, IVectorPtr label) override; + + void onPassEnd() override; + + struct PredictionResult { + PredictionResult(real __out, int __label) : out(__out), label(__label) {} + real out; + int label; + }; + std::vector predictArray_; + + private: + bool passBegin_; + std::unique_ptr evaluator_; + MatrixPtr cpuOutput_; + IVectorPtr cpuLabel_; + MatrixPtr cpuWeight_; +}; + +/* + * positive-negative pair rate Validation + */ +class PnpairValidation : public ValidationLayer { + public: + explicit PnpairValidation(const LayerConfig& config) + : ValidationLayer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void validationImp(MatrixPtr outputValue, IVectorPtr label) override; + + void onPassEnd() override; + + private: + bool passBegin_; + std::unique_ptr evaluator_; +}; + +typedef std::shared_ptr ValidationLayerPtr; +} // namespace paddle diff --git a/paddle/gserver/layers/WarpCTCLayer.cpp b/paddle/legacy/gserver/layers/WarpCTCLayer.cpp similarity index 100% rename from paddle/gserver/layers/WarpCTCLayer.cpp rename to paddle/legacy/gserver/layers/WarpCTCLayer.cpp diff --git a/paddle/gserver/layers/WarpCTCLayer.h b/paddle/legacy/gserver/layers/WarpCTCLayer.h similarity index 100% rename from paddle/gserver/layers/WarpCTCLayer.h rename to paddle/legacy/gserver/layers/WarpCTCLayer.h diff --git a/paddle/gserver/tests/.gitignore b/paddle/legacy/gserver/tests/.gitignore similarity index 100% rename from paddle/gserver/tests/.gitignore rename to paddle/legacy/gserver/tests/.gitignore diff --git a/paddle/legacy/gserver/tests/CMakeLists.txt b/paddle/legacy/gserver/tests/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..93ddf5aa233017d4f5139a8add6c69ef3a4682b4 --- /dev/null +++ b/paddle/legacy/gserver/tests/CMakeLists.txt @@ -0,0 +1,103 @@ +# gserver pacakge unittests +add_simple_unittest(test_LinearChainCRF) +add_simple_unittest(test_RecurrentLayer) + +if(NOT MOBILE_INFERENCE) + add_simple_unittest(test_MultinomialSampler) +endif() + +function(gserver_test TARGET) + add_unittest_without_exec(${TARGET} + ${TARGET}.cpp + LayerGradUtil.cpp) + add_test(NAME ${TARGET} + COMMAND ${TARGET}) +endfunction() + +add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/concat_dotmul_a.conf + COMMAND cp -r ${CMAKE_CURRENT_SOURCE_DIR}/* ${CMAKE_CURRENT_BINARY_DIR} +) +add_custom_target(copy_gserver_conf ALL DEPENDS concat_dotmul_a.conf) + +gserver_test(test_LayerGrad) +gserver_test(test_CRFLayerGrad) +gserver_test(test_CrossEntropyOverBeamGrad) +gserver_test(test_SeqSliceLayerGrad) +gserver_test(test_ActivationGrad) +gserver_test(test_ConvTrans) +gserver_test(test_PriorBox) +gserver_test(test_DetectionOutput) +gserver_test(test_ConvUnify) +gserver_test(test_BatchNorm) +gserver_test(test_KmaxSeqScore) +gserver_test(test_Expand) +gserver_test(test_MaxPoolingWithMaskOutput) +gserver_test(test_Upsample) + +set(PYTHON_PATH + ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d + ${PADDLE_BINARY_DIR}/python/:${PADDLE_BINARY_DIR}/paddle/legacy/gserver/tests) +function(gserver_test_with_python TARGET) + add_unittest_without_exec(${TARGET} ${TARGET}.cpp) + add_test(NAME ${TARGET} + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET} + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) +endfunction() + +gserver_test_with_python(test_PyDataProvider2) +if(WITH_PYTHON) + gserver_test_with_python(test_PyDataProvider) +endif() +if(NOT MOBILE_INFERENCE) + gserver_test_with_python(test_CompareTwoNets) + # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine, I will fix it. + gserver_test_with_python(test_RecurrentGradientMachine) +endif() + +########## test_MKLDNN layers and activations ########## +if(WITH_MKLDNN) + add_unittest_without_exec(test_MKLDNN + test_MKLDNN.cpp + MKLDNNTester.cpp + LayerGradUtil.cpp) + add_test(NAME test_MKLDNN + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/test_MKLDNN + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle) +endif() + +############### test_WarpCTCLayer ####################### +if(NOT WITH_DOUBLE AND NOT MOBILE_INFERENCE) + add_unittest_without_exec(test_WarpCTCLayer + test_WarpCTCLayer.cpp) + add_test(NAME test_WarpCTCLayer + COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_WarpCTCLayer --warpctc_dir=${WARPCTC_LIB_DIR} + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle) +endif() + +if(NOT MOBILE_INFERENCE) + ################## test_Evaluator ############# + add_unittest(test_Evaluator + test_Evaluator.cpp) + + ########### test_NetworkCompare ############### + add_unittest_without_exec(test_NetworkCompare + test_NetworkCompare.cpp) + if(WITH_GPU) + set(use_gpu true) + else() + set(use_gpu false) + endif() + add_test(NAME test_NetworkCompare + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=${use_gpu} + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle) + + ############ test_CompareSparse ################ + add_unittest_without_exec(test_CompareSparse + test_CompareSparse.cpp) + if(NOT ON_TRAVIS) + add_test(NAME test_CompareSparse + COMMAND ${PYTHON_PATH} ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port -n 6 + ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) + endif() +endif() diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/legacy/gserver/tests/LayerGradUtil.cpp similarity index 100% rename from paddle/gserver/tests/LayerGradUtil.cpp rename to paddle/legacy/gserver/tests/LayerGradUtil.cpp diff --git a/paddle/legacy/gserver/tests/LayerGradUtil.h b/paddle/legacy/gserver/tests/LayerGradUtil.h new file mode 100644 index 0000000000000000000000000000000000000000..941989a1da49d215b9ed4af72e732d6a62fd225d --- /dev/null +++ b/paddle/legacy/gserver/tests/LayerGradUtil.h @@ -0,0 +1,329 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" + +#include "paddle/testing/TestUtil.h" +using namespace std; // NOLINT + +namespace paddle { +enum InputType { + INPUT_DATA, // dense vector + INPUT_LABEL, // id + INPUT_DATA_TARGET, // dense vector, but no gradient + INPUT_SEQUENCE_DATA, + INPUT_HASSUB_SEQUENCE_DATA, // sequence has sub-sequence + INPUT_SEQUENCE_MDIM_DATA, + INPUT_SEQUENCE_LABEL, + INPUT_SPARSE_NON_VALUE_DATA, + INPUT_SPARSE_FLOAT_VALUE_DATA, + INPUT_DENSE_DIM_DATA, // using sequence length to init dense data + INPUT_SELF_DEFINE_DATA, // support customizing for input value +}; + +struct ParaSparse { + bool sparse; + string format; + // if equalNnzPerSample is set true, + // every row of the sparse matrix in a format of CSR has a same + // number of nnz values. Currently, this flag is only used for + // selective_fc layer + bool equalNnzPerSample; + ParaSparse(const string& formatIn = "") { // NOLINT + if (formatIn == "") { + sparse = false; + } else { + sparse = true; + } + equalNnzPerSample = false; + } + ParaSparse(const string& formatIn, bool equalNnz) { + format = formatIn; + sparse = true; + equalNnzPerSample = equalNnz; + } +}; + +struct InputDef { + InputType inputType; + string name; + size_t dim; + size_t paraSize; + ParaSparse sparse; + bool isStatic; + std::vector labelInitValue; + std::vector labelSeqStartPositions; + std::vector labelSubSeqStartPositions; + std::vector ids; + MatrixPtr selfDefinedData; + + InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) { + inputType = type; + name = nameIn; + dim = dimIn; + paraSize = sizeIn; + sparse = {""}; + isStatic = false; + } + + InputDef(InputType type, + string nameIn, + MatrixPtr selfDefinedData, + std::vector selfDefinedSeqStartPos = {}, + std::vector selfDefinedSubSeqStartPos = {}) + : labelSeqStartPositions(selfDefinedSeqStartPos), + labelSubSeqStartPositions(selfDefinedSubSeqStartPos), + selfDefinedData(selfDefinedData) { + inputType = type; + name = nameIn; + dim = 0; + sparse = {""}; + paraSize = 0; + isStatic = false; + } + + InputDef(InputType type, + string nameIn, + const std::vector& ids, + const std::vector& selfDefinedSeqStartPos = {}, + const std::vector& selfDefinedSubSeqStartPos = {}) + : labelSeqStartPositions(selfDefinedSeqStartPos), + labelSubSeqStartPositions(selfDefinedSubSeqStartPos), + ids(ids) { + selfDefinedData = nullptr; + inputType = type; + name = nameIn; + dim = 0; + sparse = {""}; + paraSize = 0; + isStatic = false; + } + + InputDef(InputType type, + string nameIn, + size_t dimIn, + size_t sizeIn, + const std::vector& labelInitValue, + const std::vector& labelSeqStartPositions) + : labelInitValue(labelInitValue), + labelSeqStartPositions(labelSeqStartPositions) { + inputType = type; + name = nameIn; + dim = dimIn; + paraSize = sizeIn; + sparse = {""}; + isStatic = false; + } + + InputDef(InputType type, + string nameIn, + size_t dimIn, + size_t sizeIn, + ParaSparse sparseIn) { + inputType = type; + name = nameIn; + dim = dimIn; + paraSize = sizeIn; + sparse = sparseIn; + } +}; + +struct TestConfig { + LayerConfig layerConfig; + std::vector inputDefs; + size_t biasSize; + real paramInitialMean; + real paramInitialStd; + bool testAccumulate; + bool testState; + bool staticBias; + bool testBatchState; + TestConfig() + : biasSize(0), + paramInitialMean(0.0), + paramInitialStd(1.0), + testAccumulate(true), + testState(false), + staticBias(false), + testBatchState(false) {} +}; + +real getCostSum(ParameterPtr& parameter, + CpuVector& cpuPara, + LayerPtr& testLayer, + MatrixPtr weights = nullptr); + +real getDiffAndPrint(real newCost1, + real newCost2, + real callbackCount, + char fill, + string testLayerName, + string name, + real step, + real delta); + +/** + * @brief verify that sequentially running forward() one timestamp at one time + * has same result as running forward() with one whole sequence + * + * @param testLayer[in/out] testLayer + * @param dataLayers[in/out] dataLayers + * @param datas[in/out] data of dataLayers + */ +void testState(LayerPtr testLayer, + vector& dataLayers, + vector& datas); + +/** + * @brief verify that sequentially running forward() with short sequences one + * time has same result as running forward() with long sequences. + * + * @param testLayer[in/out] testLayer + * @param dataLayers[in/out] dataLayers + * @param datas[in/out] data of dataLayers + */ +void testBatchState(LayerPtr testLayer, + vector& dataLayers, + vector& datas); + +/** + * @brief Generate a perturbation so that it is roughly aligned with the + * gradient direction. This is to make sure that change along this + * direction will make cost increase (or decrease) in a meaningful + * way so that the finite difference can be used to approximate the + * directional dirivative well. + * + * @param oldGrad[in] input gradient + * newGrad[out] output gradient + * dim dimension of oldGrad/newGrad + * + * @return sum_i(oldGrad[i] * newGrad[i]) + */ +double genPerturbation(const real* oldGrad, real* newGrad, size_t dim); + +void initWeight(MatrixPtr& weights); + +void initBatchState(LayerPtr dataLayer, + LayerPtr testLayer, + LayerStatePtr state, + bool useGpu); + +/** + * @brief initialize the dataLayer by its inputType + * + * @param testConf[in] test config + * dataLayers[out] dataLayers + * datas[out] initialized data of dataLayers + * layerMap[out] layerMap + */ +void initDataLayer(TestConfig testConf, + std::vector* dataLayers, + vector* datas, + LayerMap* layerMap, + string testLayerName, + size_t batchSize, + bool trans, + bool useGpu); + +/** + * @brief initialize the parameter of testLayer + * + * @param testConf[in/out] test config + * layerMap[out] layerMap + * parameters[out] parameters of testLayer + * testLayer[out] testLayer + */ +void initTestLayer(TestConfig testConf, + LayerMap* layerMap, + std::vector* parameters, + LayerPtr* testLayer); + +/** + * @brief Test whether the layer's forward calculation is stable by adding + * perturbation to its parameters + * + * @param testConf[in] test config + * weights[in] weights of testLayer + * state[in] state of testLayer + * cost[in] input cost + * callbackCount[in] number of done callback + * maxDiff[in/out] max of all previous diff + * testLayer[in/out] testLayer + * parameters[in/out] parameters of testLayer + */ +void testPerturbParameter(TestConfig testConf, + const MatrixPtr weights, + const LayerStatePtr state, + real cost, + real callbackCount, + real* maxDiff, + LayerPtr testLayer, + std::vector* parameters); + +/** + * @brief Test whether the layer's forward calculation is stable by adding + * perturbation to its input layers + * + * @param testConf[in] test config + * weights[in] weights of testLayer + * state[in] state of testLayer + * cost[in] input cost + * callbackCount[in] number of done callback + * maxDiff[in/out] max of all previous diff + * testLayer[in/out] testLayer + * dataLayers[in/out] dataLayers + */ +void testPerturbInput(TestConfig testConf, + const MatrixPtr weights, + const LayerStatePtr state, + real cost, + real callbackCount, + real* maxDiff, + LayerPtr testLayer, + std::vector dataLayers); + +void testLayerGradKernel(TestConfig testConf, + string testLayerName, + size_t batchSize, + bool trans, + bool useGpu, + bool useWeight = false, + float epsilon = 0.02); + +void testLayerGrad(TestConfig testConf, + string testLayerName, + size_t batchSize, + bool trans, + bool useGpu, + bool useWeight = false, + float epsilon = 0.02); + +void testProjectionGrad(ProjectionConfig conf, + InputType inputType, + size_t parameterSize, + size_t batchSize, + bool useGpu, + bool testState = false, + int biasSize = 0, + bool sharedBias = false); + +void testOperatorGrad(TestConfig& config, + OperatorConfig& operatorConf, + size_t batchSize, + bool useGpu, + bool testState = false); + +} // namespace paddle diff --git a/paddle/legacy/gserver/tests/MKLDNNTester.cpp b/paddle/legacy/gserver/tests/MKLDNNTester.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b550ba9c72d85830dbf12485a6a645a6b5360026 --- /dev/null +++ b/paddle/legacy/gserver/tests/MKLDNNTester.cpp @@ -0,0 +1,580 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNTester.h" +#include "paddle/legacy/gserver/layers/MKLDNNBase.h" +#include "paddle/legacy/gserver/layers/MKLDNNLayer.h" +#include "paddle/legacy/trainer/Trainer.h" + +namespace paddle { + +// init data layer and test layer of both dnn and reference +void MKLDNNTester::reset(const TestConfig& dnn, + const TestConfig& ref, + size_t batchSize) { + const bool trans = false; + const bool useGpu = false; + + // clear + configs_.clear(); + layerNames_.clear(); + dataLayers_.clear(); + datas_.clear(); + layerMaps_.clear(); + parameters_.clear(); + testLayers_.clear(); + + // resize + configs_.resize(NUM); + layerNames_.resize(NUM); + dataLayers_.resize(NUM); + datas_.resize(NUM); + layerMaps_.resize(NUM); + parameters_.resize(NUM); + testLayers_.resize(NUM); + + // reset configs and layer names + configs_[DNN] = dnn; + configs_[REF] = ref; + layerNames_[DNN] = "mkldnn"; // the first is mkldnn layer + layerNames_[REF] = "reference"; // second is reference layer + + // reset others + for (size_t i = 0; i < NUM; ++i) { + configs_[i].layerConfig.set_name(layerNames_[i]); + initDataLayer(configs_[i], + &(dataLayers_[i]), + &(datas_[i]), + &(layerMaps_[i]), + layerNames_[i], + batchSize, + trans, + useGpu); + initTestLayer( + configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i])); + } + refLayer_ = testLayers_[REF]; + dnnLayer_ = testLayers_[DNN]; + EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size()); + EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); + setInputImgSize(); + + // for comparison with Paddle reference results, + // need manually add cpu device output for test + MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(dnnLayer_); + if (dnnLayer) { + dnnLayer->addOutputArgument(CPU_DEVICE); + } +} + +void MKLDNNTester::setInputImgSize() { + for (size_t n = 0; n < dataLayers_.size(); ++n) { + for (size_t i = 0; i < dataLayers_[n].size(); ++i) { + // TODO(TJ): fix me when concat and elewise ready + dataLayers_[n][i]->getOutput().setFrameHeight(ih_); + dataLayers_[n][i]->getOutput().setFrameWidth(iw_); + } + } +} + +// init randome parameters of ref, and copy to mkldnn +void MKLDNNTester::randomWgtDatas() { + EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); + const bool isBN = refLayer_->getType() == "batch_norm"; + for (size_t i = 0; i < parameters_[REF].size(); ++i) { + const VectorPtr& dnnValue = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); + const VectorPtr& refValue = parameters_[REF][i]->getBuf(PARAMETER_VALUE); + parameters_[REF][i]->randomize(); + if (isBN && i == 2) { + // this param is moving average in batch norm, which must larger than 0 + real offset = fabs(refValue->getMin()) + 1.0; + refValue->add(offset); + } + dnnValue->copyFrom(*refValue); + + VLOG(MKLDNN_TESTS) << "Random weight " << parameters_[DNN][i]->getName(); + printVector(dnnValue); + } +} + +// random botdata of ref layer and copy same to mkldnn +void MKLDNNTester::randomBotDatas() { + CHECK_EQ(dataLayers_.size(), NUM); + for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) { + dataLayers_[REF][i]->getOutputValue()->randomizeUniform(); + dataLayers_[DNN][i]->getOutputValue()->copyFrom( + *(dataLayers_[REF][i]->getOutputValue())); + VLOG(MKLDNN_TESTS) << "Random Foward, InputValue " << i; + printMatrix(dataLayers_[REF][i]->getOutputValue()); + } +} + +void MKLDNNTester::randomTopDiffs() { + refLayer_->getOutputGrad()->randomizeUniform(); + dnnLayer_->getOutput(CPU_DEVICE) + .grad->copyFrom(*(refLayer_->getOutputGrad())); + VLOG(MKLDNN_TESTS) << "Random Backward, OutputGrad"; + printMatrix(refLayer_->getOutputGrad()); +} + +void MKLDNNTester::checkForward() { + VLOG(MKLDNN_TESTS) << "Check Forward"; + printTopDatas(); + double delta = + compareMatrix(refLayer_->getOutputValue(), dnnLayer_->getOutputValue()); + EXPECT_LE(fabs(delta), eps_); +} + +void MKLDNNTester::checkBackwardData() { + VLOG(MKLDNN_TESTS) << "Check Backward Data"; + const bool isBN = refLayer_->getType() == "batch_norm"; + for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) { + const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad(); + const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad(); + VLOG(MKLDNN_ALL) << "MKLDNN Backward Result: InputGrad " << i; + printMatrix(dnnDiff); + VLOG(MKLDNN_ALL) << "Reference Backward Result: InputGrad " << i; + printMatrix(refDiff); + + double delta = compareMatrix(refDiff, dnnDiff); + EXPECT_LE(fabs(delta), eps_); + if (isBN) { + // the other two inputs in batch norm are for moving mean and var + // do not have grad to compare + break; + } + } +} + +void MKLDNNTester::checkBackwardWgts() { + VLOG(MKLDNN_TESTS) << "Check Backward Weight"; + CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); + vector dnnWgts; // used to temply save mkldnn weights + saveWgt(parameters_[DNN], dnnWgts); + + MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(dnnLayer_); + if (dnnLayer) { + dnnLayer->convertWeightsToPaddle(); + } + for (size_t i = 0; i < parameters_[DNN].size(); ++i) { + const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); + const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE); + VLOG(MKLDNN_ALL) << "MKLDNN Result: weight value" + << parameters_[DNN][i]->getName(); + printVector(dnn); + VLOG(MKLDNN_ALL) << "Reference Result: weight value " + << parameters_[REF][i]->getName(); + printVector(ref); + + double delta = compareVector(ref, dnn); + EXPECT_LE(fabs(delta), eps_); + } + + VLOG(MKLDNN_ALL) << "Restore dnn weights before comapre"; + restoreWgt(dnnWgts, parameters_[DNN]); +} + +void MKLDNNTester::saveWgt(const vector& from, + vector& to) { + const bool useGpu = false; + to.resize(from.size()); + for (size_t i = 0; i < to.size(); ++i) { + const VectorPtr& wgt = from[i]->getBuf(PARAMETER_VALUE); + to[i] = Vector::create(wgt->getSize(), useGpu); + to[i]->copyFrom(*wgt); + } +} + +void MKLDNNTester::restoreWgt(const vector& from, + vector& to) { + CHECK_EQ(from.size(), to.size()); + for (size_t i = 0; i < from.size(); ++i) { + const VectorPtr& wgt = to[i]->getBuf(PARAMETER_VALUE); + wgt->copyFrom(*from[i]); + } +} + +// clear parameters grad +void MKLDNNTester::clearWgtDiffs(size_t id) { + CHECK_LE(id, parameters_.size()); + for (size_t n = 0; n < parameters_.size(); ++n) { + if (id == n || id == parameters_.size()) { + for (size_t i = 0; i < parameters_[n].size(); ++i) { + const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT); + if (grad) { + grad->zeroMem(); + } + } + } + } +} + +void MKLDNNTester::clearBotDiffs(size_t id) { + CHECK_LE(id, dataLayers_.size()); + for (size_t n = 0; n < dataLayers_.size(); ++n) { + if (id == n || id == dataLayers_.size()) { + // clear inputs layers of this specific layer + for (size_t i = 0; i < dataLayers_[n].size(); ++i) { + dataLayers_[n][i]->getOutputGrad()->zeroMem(); + } + } + } +} + +void MKLDNNTester::clearTopDatas(size_t id) { + CHECK_LE(id, testLayers_.size()); + for (size_t i = 0; i < testLayers_.size(); ++i) { + if (id == i || id == testLayers_.size()) { + testLayers_[i]->getOutputValue()->zeroMem(); + } + } +} + +void MKLDNNTester::printTopDatas() { + if (!log_) { + return; + } + + for (int n = 0; n < NUM; ++n) { + VLOG(MKLDNN_ALL) << testLayers_[n]->getType() + << " Forward Result: OutputValue"; + printMatrix(testLayers_[n]->getOutputValue()); + } +} + +void MKLDNNTester::printMatrix(const MatrixPtr& m) { + if (!log_) { + return; + } + + std::ostringstream ostr; + m->print(ostr); + VLOG(MKLDNN_ALL) << std::endl << ostr.str(); +} + +void MKLDNNTester::printVector(const VectorPtr& v) { + if (!log_) { + return; + } + + std::ostringstream ostr; + v->print(ostr, v->getSize()); + VLOG(MKLDNN_ALL) << std::endl << ostr.str(); +} + +double MKLDNNTester::getDelta(const real* refer, + const real* value, + size_t len, + const float failRate, + const float thres) { + double delta = 0, sum = 0; + int failCnt = 0; + const double eps = 1e-5; + double maxRatio = 0; + for (size_t i = 0; i < len; ++i) { + double ref = fabs(refer[i]); + double val = fabs(value[i]); + double diff = fabs(refer[i] - value[i]); + delta += diff; + sum += ref; + if (ref < eps && val < eps) { // both values are very small + continue; + } + double ratio = diff / ref; + if (ratio > thres) { + maxRatio = std::max(maxRatio, ratio); + failCnt++; + } + } + EXPECT_FALSE(std::isinf(sum)); + EXPECT_FALSE(std::isnan(sum)); + EXPECT_FALSE(std::isnan(delta)); + VLOG(MKLDNN_ALL) << "reference avg data: " << sum / len + << ", delta: " << delta / sum << ", failCnt:" << failCnt; + double res = sum > eps ? delta / sum : eps; + return (failCnt / (float)len) > failRate ? maxRatio : res; +} + +double MKLDNNTester::compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2) { + CHECK_EQ(m1->getElementCnt(), m2->getElementCnt()); + return getDelta(m1->getData(), m2->getData(), m1->getElementCnt()); +} + +double MKLDNNTester::compareVector(const VectorPtr& v1, const VectorPtr& v2) { + CHECK_EQ(v1->getSize(), v2->getSize()); + return getDelta(v1->getData(), v2->getData(), v1->getSize()); +} + +void MKLDNNTester::runOnce() { + // test forward + randomBotDatas(); + dnnLayer_->forward(passType_); + refLayer_->forward(passType_); + checkForward(); + + if (passType_ == PASS_TEST) { + return; + } + + // test backward + // simple updater + UpdateCallback updateCallback = [](Parameter* para) { + auto& grad = para->getBuf(PARAMETER_GRADIENT); + auto& value = para->getBuf(PARAMETER_VALUE); + real lr = 1e-2; + value->add(*grad, lr); + grad->zeroMem(); + }; + randomTopDiffs(); + dnnLayer_->backward(updateCallback); + refLayer_->backward(updateCallback); + checkBackwardData(); + checkBackwardWgts(); + + // clear buffers + // ref code will addto the diff, dnn code will writeto it + // and clearTopDatas(REF) should be coverd by ref layers + clearBotDiffs(REF); + clearWgtDiffs(REF); + // it is necessary to clear bottom diffs when only activation is dnn type + if (configs_[DNN].layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) { + clearBotDiffs(DNN); + } +} + +void MKLDNNTester::run(const TestConfig& dnn, + const TestConfig& ref, + size_t batchSize, + size_t inputImgH, + size_t inputImgW, + PassType passType, + bool printDetails, + size_t iter, + float epsilon) { + CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 || + dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) + << "should be MKLDNN layer or MKLDNN activation"; + if (dnn.layerConfig.type() == ref.layerConfig.type()) { + VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " + << dnn.layerConfig.active_type() << " vs " + << ref.layerConfig.active_type(); + } else { + VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " + << dnn.layerConfig.type() << " vs " + << ref.layerConfig.type(); + } + + ih_ = inputImgH; + iw_ = inputImgW; + passType_ = passType; + log_ = printDetails; + iter_ = iter; + eps_ = epsilon; + + // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight + reset(dnn, ref, batchSize); + randomWgtDatas(); + clearWgtDiffs(); + clearBotDiffs(); + for (size_t i = 0; i < iter_; ++i) { + VLOG(MKLDNN_TESTS) << "Check Iteration " << i; + runOnce(); + } + + if (parameters_[DNN].empty()) { + // has no paramters + return; + } + + // After run some iterations, the mkldnn weight has been stored in dnnLayer + // and we can also get the mkldnn weight parameter header format. + // Weight parameter should always be index 0 (and bias index 1). + // TODO(TJ): should also consider mean and var format when batchnorm ready + int dnnWgtFmt = parameters_[DNN][0]->getHeaderFormat(); + int refWgtFmt = parameters_[REF][0]->getHeaderFormat(); + if (dnnWgtFmt == refWgtFmt) { + // weight format are equal, so no need check more + return; + } + + // then save the weights and restart again + vector dnnWgts, refWgts; + CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); + saveWgt(parameters_[DNN], dnnWgts); + saveWgt(parameters_[REF], refWgts); + + // restart again with dnn weight format + reset(dnn, ref, batchSize); + // TODO(TJ): should also considerate mean and var format when batchnorm ready + parameters_[DNN][0]->setHeaderFormat(dnnWgtFmt); + + // restore wgt + restoreWgt(dnnWgts, parameters_[DNN]); + restoreWgt(refWgts, parameters_[REF]); + clearWgtDiffs(); + clearBotDiffs(); + + for (size_t i = 0; i < iter_; ++i) { + VLOG(MKLDNN_TESTS) << "Check Iteration " << i; + runOnce(); + } +} + +void MKLDNNTester::initArgument(DataIn& data, + const std::string& configPath, + const size_t iter) { + TrainerConfigHelper config(configPath); + size_t batchSize = config.getOptConfig().batch_size(); + data.inArgs.resize(iter); + data.outGrads.resize(iter); + data.paraValues.clear(); + for (const auto& layer_name : config.getModelConfig().input_layer_names()) { + auto layer_config = std::find_if(config.getModelConfig().layers().begin(), + config.getModelConfig().layers().end(), + [=](const LayerConfig& layer_config) { + return layer_config.name() == layer_name; + }); + CHECK(layer_config != config.getModelConfig().layers().end()); + + size_t layerSize = layer_config->size(); + for (size_t i = 0; i < iter; ++i) { + Argument arg; + arg.value = Matrix::create(batchSize, layerSize, false, false); + arg.grad = Matrix::create(batchSize, layerSize, false, false); + arg.value->randomizeUniform(); + arg.value->add(-0.5); + arg.value->sigmoid(*arg.value); + arg.grad->zeroMem(); + arg.ids = VectorT::create(batchSize, false); + arg.ids->rand(layerSize); + generateSequenceStartPositions(batchSize, arg.sequenceStartPositions); + data.inArgs[i].push_back(arg); + } + } + + for (const auto& layer_name : config.getModelConfig().output_layer_names()) { + auto layer_config = std::find_if(config.getModelConfig().layers().begin(), + config.getModelConfig().layers().end(), + [=](const LayerConfig& layer_config) { + return layer_config.name() == layer_name; + }); + CHECK(layer_config != config.getModelConfig().layers().end()); + + size_t layerSize = layer_config->size(); + for (size_t i = 0; i < iter; ++i) { + MatrixPtr grad = Matrix::create(batchSize, layerSize, false, false); + grad->randomizeUniform(); + data.outGrads[i].push_back(grad); + } + } + + for (const auto& para_config : config.getModelConfig().parameters()) { + VectorPtr value = Vector::create(para_config.size(), false); + value->randnorm(0, 2); + data.paraValues.push_back(value); + } +} + +void MKLDNNTester::getOutResult(const std::string& configPath, + DataIn& in, + DataOut& out, + bool use_mkldnn, + size_t iter) { + FLAGS_use_gpu = false; + FLAGS_use_mkldnn = use_mkldnn; + *ThreadLocalRand::getSeed() = 1; + srand(1); + + Trainer trainer; + auto config = std::make_shared(configPath); + trainer.init(config, false); + auto gradientMachine = trainer.getGradientMachine(); + std::vector parameters = gradientMachine->getParameters(); + for (size_t i = 0; i < in.paraValues.size(); i++) { + parameters[i]->getBuf(PARAMETER_VALUE)->copyFrom(*in.paraValues[i]); + } + UpdateCallback simpleUpdate = [](Parameter* para) { + auto& grad = para->getBuf(PARAMETER_GRADIENT); + auto& value = para->getBuf(PARAMETER_VALUE); + real lr = 1e-2; + value->add(*grad, lr); + grad->zeroMem(); + }; + + vector outArgs; + gradientMachine->start(); + out.outValues.clear(); + out.paraValues.clear(); + for (size_t i = 0; i < iter; ++i) { + VLOG(MKLDNN_TESTS) << "runing iteration " << i; + gradientMachine->forward(in.inArgs[i], &outArgs, PASS_TRAIN); + // save forward result + for (size_t k = 0; k < outArgs.size(); k++) { + const MatrixPtr& src = outArgs[k].value; + MatrixPtr dst = + Matrix::create(src->getHeight(), src->getWidth(), false, false); + if (typeid(*src) == typeid(MKLDNNMatrix)) { + MKLDNNMatrixPtr dnnSrc = std::dynamic_pointer_cast(src); + dnnSrc->copyTo(*dst); + } else { + dst->copyFrom(*src); + } + out.outValues.push_back(dst); + } + + // random backward input + for (size_t k = 0; k < outArgs.size(); k++) { + outArgs[k].grad->copyFrom(*in.outGrads[i][k]); + } + gradientMachine->backward(simpleUpdate); + } + gradientMachine->finish(); + + // save param value + for (size_t i = 0; i < in.paraValues.size(); i++) { + VectorPtr val = Vector::create( + parameters[i]->getBuf(PARAMETER_VALUE)->getSize(), false); + val->copyFrom(*parameters[i]->getBuf(PARAMETER_VALUE)); + out.paraValues.push_back(val); + } +} + +void MKLDNNTester::compareResult(DataOut& ref, DataOut& dnn, float eps) { + CHECK_EQ(ref.outValues.size(), dnn.outValues.size()); + CHECK_EQ(ref.paraValues.size(), dnn.paraValues.size()); + for (size_t i = 0; i < ref.outValues.size(); i++) { + VLOG(MKLDNN_TESTS) << "compare value index: " << i; + EXPECT_LE(fabs(compareMatrix(ref.outValues[i], dnn.outValues[i])), eps); + } + for (size_t i = 0; i < ref.paraValues.size(); i++) { + VLOG(MKLDNN_TESTS) << "compare param index: " << i; + EXPECT_LE(fabs(compareVector(ref.paraValues[i], dnn.paraValues[i])), eps); + } +} + +void MKLDNNTester::runNetTest(const std::string& configPath, + size_t iter, + float eps) { + DataIn in; + initArgument(in, configPath, iter); + DataOut outCpu, outDnn; + VLOG(MKLDNN_TESTS) << "runing cpu network"; + getOutResult(configPath, in, outCpu, false, iter); + VLOG(MKLDNN_TESTS) << "runing mkldnn network"; + getOutResult(configPath, in, outDnn, true, iter); + + compareResult(outCpu, outDnn, eps); +} + +} // namespace paddle diff --git a/paddle/legacy/gserver/tests/MKLDNNTester.h b/paddle/legacy/gserver/tests/MKLDNNTester.h new file mode 100644 index 0000000000000000000000000000000000000000..086846ce537857eb76ffca492246677eb7982a42 --- /dev/null +++ b/paddle/legacy/gserver/tests/MKLDNNTester.h @@ -0,0 +1,143 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "LayerGradUtil.h" +#include "paddle/legacy/gserver/layers/MKLDNNBase.h" +#include "paddle/legacy/gserver/layers/MKLDNNLayer.h" + +namespace paddle { + +/** + * @brief test the functionality of MKLDNNlayers and MKLDNNActivations + * refer to paddle original function + */ +class MKLDNNTester { + enum { + DNN = 0, // MKLDNN layer + REF = 1, // Reference layer + NUM = 2, // Number of total + }; + + struct DataIn { + std::vector> inArgs; + std::vector> outGrads; + std::vector paraValues; + }; + + struct DataOut { + std::vector outValues; + std::vector paraValues; + }; + + protected: + std::vector configs_; + vector layerNames_; + vector> dataLayers_; + vector> datas_; + vector layerMaps_; + vector> parameters_; + vector testLayers_; + LayerPtr refLayer_, dnnLayer_; + + /// run some iterations, all the result should pass + size_t iter_; + /// whether to print out the details + bool log_; + /// epsilon + float eps_; + /// input image size, default 1 + size_t ih_, iw_; + /// passType, PASS_TRAIN, PASS_TEST or PASS_GC (Gradient Check pass) + PassType passType_; + + public: + explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) { + iter_ = iter; + eps_ = epsilon; + log_ = false; + passType_ = PASS_TRAIN; + } + + ~MKLDNNTester() {} + + public: + void run(const TestConfig& dnn, + const TestConfig& ref, + size_t batchSize, + size_t inputImgH = 1, + size_t inputImgW = 1, + PassType passType = PASS_TRAIN, + bool printDetails = false, + size_t iter = 3, + float epsilon = 1e-4); + static void runNetTest(const std::string& configPath, + size_t iter = 2, + float eps = 1e-4); + static void initArgument(DataIn& data, + const std::string& configPath, + size_t iter = 2); + static void getOutResult(const std::string& configPath, + DataIn& in, + DataOut& out, + bool use_mkldnn, + size_t iter = 2); + + private: + void reset(const TestConfig& dnn, const TestConfig& ref, size_t batchSize); + void setInputImgSize(); + void runOnce(); + + void randomWgtDatas(); + void randomBotDatas(); + void randomTopDiffs(); + + void checkForward(); + void checkBackwardData(); + void checkBackwardWgts(); + + // clear specific layer, clear all when id equals NUM + void clearWgtDiffs(size_t id = NUM); + void clearBotDiffs(size_t id = NUM); + void clearTopDatas(size_t id = NUM); + + void printTopDatas(); + void printMatrix(const MatrixPtr& m); + void printVector(const VectorPtr& v); + + void saveWgt(const vector& from, vector& to); + void restoreWgt(const vector& from, vector& to); + + static double compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2); + static double compareVector(const VectorPtr& v1, const VectorPtr& v2); + static void compareResult(DataOut& ref, DataOut& dnn, float eps = 1e-4); + + /** + * Get delta percent + * if many(>failRate) wrong(abs(val-ref)/abs(ref) > thres) points + * return the max(diff/ref) + * else return sum(abs(diff)) / sum(abs(ref)) + * The return value should be smaller than eps when passing. + */ + static double getDelta(const real* refer, + const real* value, + size_t len, + const float failRate = 1e-3, + const float thres = 0.1); +}; + +} // namespace paddle diff --git a/paddle/gserver/tests/Sequence/dummy.list b/paddle/legacy/gserver/tests/Sequence/dummy.list similarity index 100% rename from paddle/gserver/tests/Sequence/dummy.list rename to paddle/legacy/gserver/tests/Sequence/dummy.list diff --git a/paddle/gserver/tests/Sequence/tour_dict_phrase.dict b/paddle/legacy/gserver/tests/Sequence/tour_dict_phrase.dict similarity index 100% rename from paddle/gserver/tests/Sequence/tour_dict_phrase.dict rename to paddle/legacy/gserver/tests/Sequence/tour_dict_phrase.dict diff --git a/paddle/gserver/tests/Sequence/tour_train_wdseg b/paddle/legacy/gserver/tests/Sequence/tour_train_wdseg similarity index 100% rename from paddle/gserver/tests/Sequence/tour_train_wdseg rename to paddle/legacy/gserver/tests/Sequence/tour_train_wdseg diff --git a/paddle/gserver/tests/Sequence/tour_train_wdseg.nest b/paddle/legacy/gserver/tests/Sequence/tour_train_wdseg.nest similarity index 100% rename from paddle/gserver/tests/Sequence/tour_train_wdseg.nest rename to paddle/legacy/gserver/tests/Sequence/tour_train_wdseg.nest diff --git a/paddle/legacy/gserver/tests/Sequence/train.list b/paddle/legacy/gserver/tests/Sequence/train.list new file mode 100644 index 0000000000000000000000000000000000000000..1109a2449252cb9bfcb10ece4cf9a96e655e5a25 --- /dev/null +++ b/paddle/legacy/gserver/tests/Sequence/train.list @@ -0,0 +1 @@ +legacy/gserver/tests/Sequence/tour_train_wdseg diff --git a/paddle/legacy/gserver/tests/Sequence/train.list.nest b/paddle/legacy/gserver/tests/Sequence/train.list.nest new file mode 100644 index 0000000000000000000000000000000000000000..a67df35024f456d517899f37272b0f74d822f03d --- /dev/null +++ b/paddle/legacy/gserver/tests/Sequence/train.list.nest @@ -0,0 +1 @@ +legacy/gserver/tests/Sequence/tour_train_wdseg.nest diff --git a/paddle/gserver/tests/__init__.py b/paddle/legacy/gserver/tests/__init__.py similarity index 100% rename from paddle/gserver/tests/__init__.py rename to paddle/legacy/gserver/tests/__init__.py diff --git a/paddle/gserver/tests/concat_dotmul_a.conf b/paddle/legacy/gserver/tests/concat_dotmul_a.conf similarity index 100% rename from paddle/gserver/tests/concat_dotmul_a.conf rename to paddle/legacy/gserver/tests/concat_dotmul_a.conf diff --git a/paddle/gserver/tests/concat_dotmul_b.conf b/paddle/legacy/gserver/tests/concat_dotmul_b.conf similarity index 100% rename from paddle/gserver/tests/concat_dotmul_b.conf rename to paddle/legacy/gserver/tests/concat_dotmul_b.conf diff --git a/paddle/gserver/tests/concat_fullmatrix_a.conf b/paddle/legacy/gserver/tests/concat_fullmatrix_a.conf similarity index 100% rename from paddle/gserver/tests/concat_fullmatrix_a.conf rename to paddle/legacy/gserver/tests/concat_fullmatrix_a.conf diff --git a/paddle/gserver/tests/concat_fullmatrix_b.conf b/paddle/legacy/gserver/tests/concat_fullmatrix_b.conf similarity index 100% rename from paddle/gserver/tests/concat_fullmatrix_b.conf rename to paddle/legacy/gserver/tests/concat_fullmatrix_b.conf diff --git a/paddle/gserver/tests/concat_slice_a.conf b/paddle/legacy/gserver/tests/concat_slice_a.conf similarity index 100% rename from paddle/gserver/tests/concat_slice_a.conf rename to paddle/legacy/gserver/tests/concat_slice_a.conf diff --git a/paddle/gserver/tests/concat_slice_b.conf b/paddle/legacy/gserver/tests/concat_slice_b.conf similarity index 100% rename from paddle/gserver/tests/concat_slice_b.conf rename to paddle/legacy/gserver/tests/concat_slice_b.conf diff --git a/paddle/gserver/tests/concat_table_a.conf b/paddle/legacy/gserver/tests/concat_table_a.conf similarity index 100% rename from paddle/gserver/tests/concat_table_a.conf rename to paddle/legacy/gserver/tests/concat_table_a.conf diff --git a/paddle/gserver/tests/concat_table_b.conf b/paddle/legacy/gserver/tests/concat_table_b.conf similarity index 100% rename from paddle/gserver/tests/concat_table_b.conf rename to paddle/legacy/gserver/tests/concat_table_b.conf diff --git a/paddle/gserver/tests/img_conv_a.conf b/paddle/legacy/gserver/tests/img_conv_a.conf similarity index 100% rename from paddle/gserver/tests/img_conv_a.conf rename to paddle/legacy/gserver/tests/img_conv_a.conf diff --git a/paddle/gserver/tests/img_conv_b.conf b/paddle/legacy/gserver/tests/img_conv_b.conf similarity index 100% rename from paddle/gserver/tests/img_conv_b.conf rename to paddle/legacy/gserver/tests/img_conv_b.conf diff --git a/paddle/gserver/tests/img_conv_c.conf b/paddle/legacy/gserver/tests/img_conv_c.conf similarity index 100% rename from paddle/gserver/tests/img_conv_c.conf rename to paddle/legacy/gserver/tests/img_conv_c.conf diff --git a/paddle/gserver/tests/img_conv_cudnn.py b/paddle/legacy/gserver/tests/img_conv_cudnn.py similarity index 100% rename from paddle/gserver/tests/img_conv_cudnn.py rename to paddle/legacy/gserver/tests/img_conv_cudnn.py diff --git a/paddle/gserver/tests/img_conv_exconv.py b/paddle/legacy/gserver/tests/img_conv_exconv.py similarity index 100% rename from paddle/gserver/tests/img_conv_exconv.py rename to paddle/legacy/gserver/tests/img_conv_exconv.py diff --git a/paddle/gserver/tests/img_pool_a.conf b/paddle/legacy/gserver/tests/img_pool_a.conf similarity index 100% rename from paddle/gserver/tests/img_pool_a.conf rename to paddle/legacy/gserver/tests/img_pool_a.conf diff --git a/paddle/gserver/tests/img_pool_b.conf b/paddle/legacy/gserver/tests/img_pool_b.conf similarity index 100% rename from paddle/gserver/tests/img_pool_b.conf rename to paddle/legacy/gserver/tests/img_pool_b.conf diff --git a/paddle/gserver/tests/mkldnn_branch_net.conf b/paddle/legacy/gserver/tests/mkldnn_branch_net.conf similarity index 100% rename from paddle/gserver/tests/mkldnn_branch_net.conf rename to paddle/legacy/gserver/tests/mkldnn_branch_net.conf diff --git a/paddle/gserver/tests/mkldnn_simple_net.conf b/paddle/legacy/gserver/tests/mkldnn_simple_net.conf similarity index 100% rename from paddle/gserver/tests/mkldnn_simple_net.conf rename to paddle/legacy/gserver/tests/mkldnn_simple_net.conf diff --git a/paddle/gserver/tests/pyDataProvider.py b/paddle/legacy/gserver/tests/pyDataProvider.py similarity index 100% rename from paddle/gserver/tests/pyDataProvider.py rename to paddle/legacy/gserver/tests/pyDataProvider.py diff --git a/paddle/gserver/tests/pyDataProvider/pyDataProviderList b/paddle/legacy/gserver/tests/pyDataProvider/pyDataProviderList similarity index 100% rename from paddle/gserver/tests/pyDataProvider/pyDataProviderList rename to paddle/legacy/gserver/tests/pyDataProvider/pyDataProviderList diff --git a/paddle/gserver/tests/pyDataProvider/trainer.conf b/paddle/legacy/gserver/tests/pyDataProvider/trainer.conf similarity index 100% rename from paddle/gserver/tests/pyDataProvider/trainer.conf rename to paddle/legacy/gserver/tests/pyDataProvider/trainer.conf diff --git a/paddle/gserver/tests/rnn_data_provider.py b/paddle/legacy/gserver/tests/rnn_data_provider.py similarity index 100% rename from paddle/gserver/tests/rnn_data_provider.py rename to paddle/legacy/gserver/tests/rnn_data_provider.py diff --git a/paddle/gserver/tests/sequenceGen.py b/paddle/legacy/gserver/tests/sequenceGen.py similarity index 100% rename from paddle/gserver/tests/sequenceGen.py rename to paddle/legacy/gserver/tests/sequenceGen.py diff --git a/paddle/legacy/gserver/tests/sequence_layer_group.conf b/paddle/legacy/gserver/tests/sequence_layer_group.conf new file mode 100644 index 0000000000000000000000000000000000000000..ad1b61d5821fd20135e61bb95abdea16d27a6a9a --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_layer_group.conf @@ -0,0 +1,62 @@ +#!/usr/bin/env python +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +dict_path = 'legacy/gserver/tests/Sequence/tour_dict_phrase.dict' +dict_file = dict() +for line_count, line in enumerate(open(dict_path, "r")): + dict_file[line.strip()] = line_count + +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/train.list', + test_list=None, + module='sequenceGen', + obj='process', + args={"dict_file": dict_file}) + +settings(batch_size=5) +######################## network configure ################################ +dict_dim = len(open(dict_path, 'r').readlines()) +word_dim = 128 +hidden_dim = 256 +label_dim = 3 + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer(input=data, size=word_dim) + +# (lstm_input + lstm) is equal to lstmemory +with mixed_layer(size=hidden_dim * 4) as lstm_input: + lstm_input += full_matrix_projection(input=emb) + +lstm = lstmemory_group( + input=lstm_input, + size=hidden_dim, + act=TanhActivation(), + gate_act=SigmoidActivation(), + state_act=TanhActivation()) + +lstm_last = last_seq(input=lstm) + +with mixed_layer( + size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: + output += full_matrix_projection(input=lstm_last) + +outputs( + classification_cost( + input=output, label=data_layer( + name="label", size=1))) diff --git a/paddle/legacy/gserver/tests/sequence_lstm.conf b/paddle/legacy/gserver/tests/sequence_lstm.conf new file mode 100644 index 0000000000000000000000000000000000000000..6ab70e70713f31de31b5cd544cf132e7d0af0f2f --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_lstm.conf @@ -0,0 +1,64 @@ +#!/usr/bin/env python +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +dict_path = 'legacy/gserver/tests/Sequence/tour_dict_phrase.dict' +dict_file = dict() +for line_count, line in enumerate(open(dict_path, "r")): + dict_file[line.strip()] = line_count + +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/train.list', + test_list=None, + module='sequenceGen', + obj='process', + args={"dict_file": dict_file}) + +settings(batch_size=5) +######################## network configure ################################ +dict_dim = len(open(dict_path, 'r').readlines()) +word_dim = 128 +hidden_dim = 256 +label_dim = 3 +sparse_update = get_config_arg("sparse_update", bool, False) + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer( + input=data, + size=word_dim, + param_attr=ParamAttr(sparse_update=sparse_update)) + +with mixed_layer(size=hidden_dim * 4) as lstm_input: + lstm_input += full_matrix_projection(input=emb) + +lstm = lstmemory( + input=lstm_input, + act=TanhActivation(), + gate_act=SigmoidActivation(), + state_act=TanhActivation()) + +lstm_last = last_seq(input=lstm) + +with mixed_layer( + size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: + output += full_matrix_projection(input=lstm_last) + +outputs( + classification_cost( + input=output, label=data_layer( + name="label", size=1))) diff --git a/paddle/legacy/gserver/tests/sequence_nest_layer_group.conf b/paddle/legacy/gserver/tests/sequence_nest_layer_group.conf new file mode 100644 index 0000000000000000000000000000000000000000..75c36b118979760e034f81e3127a748651f53347 --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_nest_layer_group.conf @@ -0,0 +1,83 @@ +#!/usr/bin/env python +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +dict_path = 'legacy/gserver/tests/Sequence/tour_dict_phrase.dict' +dict_file = dict() +for line_count, line in enumerate(open(dict_path, "r")): + dict_file[line.strip()] = line_count + +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/train.list.nest', + test_list=None, + module='sequenceGen', + obj='process2', + args={"dict_file": dict_file}) + +settings(batch_size=2) +######################## network configure ################################ +dict_dim = len(open(dict_path, 'r').readlines()) +word_dim = 128 +hidden_dim = 256 +label_dim = 3 + +data = data_layer(name="word", size=dict_dim) + +emb_group = embedding_layer(input=data, size=word_dim) + + +# (lstm_input + lstm) is equal to lstmemory +def lstm_group(lstm_group_input): + with mixed_layer(size=hidden_dim * 4) as group_input: + group_input += full_matrix_projection(input=lstm_group_input) + + lstm_output = lstmemory_group( + input=group_input, + name="lstm_group", + size=hidden_dim, + act=TanhActivation(), + gate_act=SigmoidActivation(), + state_act=TanhActivation()) + return lstm_output + + +lstm_nest_group = recurrent_group( + input=SubsequenceInput(emb_group), step=lstm_group, name="lstm_nest_group") +# hasSubseq ->(seqlastins) seq +lstm_last = last_seq( + input=lstm_nest_group, agg_level=AggregateLevel.TO_SEQUENCE) + +# seq ->(expand) hasSubseq +lstm_expand = expand_layer( + input=lstm_last, + expand_as=emb_group, + expand_level=ExpandLevel.FROM_SEQUENCE) + +# hasSubseq ->(average) seq +lstm_average = pooling_layer( + input=lstm_expand, + pooling_type=AvgPooling(), + agg_level=AggregateLevel.TO_SEQUENCE) + +with mixed_layer( + size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: + output += full_matrix_projection(input=lstm_average) + +outputs( + classification_cost( + input=output, label=data_layer( + name="label", size=1))) diff --git a/paddle/legacy/gserver/tests/sequence_nest_rnn.conf b/paddle/legacy/gserver/tests/sequence_nest_rnn.conf new file mode 100644 index 0000000000000000000000000000000000000000..bc3b22c2a946a62c7a9d3163d3863a090d63539c --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_nest_rnn.conf @@ -0,0 +1,74 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2(train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_subseq') + + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 8 +hidden_dim = 8 +label_dim = 3 + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer(input=data, size=word_dim) + +# This hierachical RNN is designed to be equivalent to the simple RNN in +# sequence_rnn.conf + +def outer_step(x): + outer_mem = memory(name="outer_rnn_state", size=hidden_dim) + def inner_step(y): + inner_mem = memory(name="inner_rnn_state", + size=hidden_dim, + boot_layer=outer_mem) + out = fc_layer(input=[y, inner_mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name="inner_rnn_state") + return out + + inner_rnn_output = recurrent_group( + step=inner_step, + name="inner", + input=x) + last = last_seq(input=inner_rnn_output, name="outer_rnn_state") + + # "return last" won't work, because recurrent_group only support the input + # sequence type is same as return sequence type. + return inner_rnn_output + +out = recurrent_group( + name="outer", + step=outer_step, + input=SubsequenceInput(emb)) + +rep = last_seq(input=out) +prob = fc_layer(size=label_dim, + input=rep, + act=SoftmaxActivation(), + bias_attr=True) + +outputs(classification_cost(input=prob, + label=data_layer(name="label", size=label_dim))) diff --git a/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_input.conf b/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_input.conf new file mode 100644 index 0000000000000000000000000000000000000000..165ab229897d32ce2cae1d483b3ffd81392a355a --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_input.conf @@ -0,0 +1,76 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2(train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_subseq') + + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 8 +hidden_dim = 8 +label_dim = 3 + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer(input=data, size=word_dim) + +# This hierachical RNN is designed to be equivalent to the simple RNN in +# sequence_rnn.conf + +def outer_step(wid, x): + outer_mem = memory(name="outer_rnn_state", size=hidden_dim) + def inner_step(y, wid): + z = embedding_layer(input=wid, size=word_dim) + inner_mem = memory(name="inner_rnn_state", + size=hidden_dim, + boot_layer=outer_mem) + out = fc_layer(input=[y, z, inner_mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name="inner_rnn_state") + return out + + inner_rnn_output = recurrent_group( + step=inner_step, + name="inner", + input=[x, wid]) + last = last_seq(input=inner_rnn_output, name="outer_rnn_state") + + # "return last" should also work. But currently RecurrentGradientMachine + # does not handle it, and will report error: In hierachical RNN, all out + # links should be from sequences now. + return inner_rnn_output + +out = recurrent_group( + name="outer", + step=outer_step, + input=[SubsequenceInput(data), SubsequenceInput(emb)]) + +rep = last_seq(input=out) +prob = fc_layer(size=label_dim, + input=rep, + act=SoftmaxActivation(), + bias_attr=True) + +outputs(classification_cost(input=prob, + label=data_layer(name="label", size=label_dim))) diff --git a/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py b/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py new file mode 100644 index 0000000000000000000000000000000000000000..9a48b7f25c454b492d20e807f09f6d788af44681 --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py @@ -0,0 +1,96 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_unequalength_subseq') + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 8 +hidden_dim = 8 +label_dim = 2 + +speaker1 = data_layer(name="word1", size=dict_dim) +speaker2 = data_layer(name="word2", size=dict_dim) + +emb1 = embedding_layer(input=speaker1, size=word_dim) +emb2 = embedding_layer(input=speaker2, size=word_dim) + + +# This hierarchical RNN is designed to be equivalent to the simple RNN in +# sequence_rnn_multi_unequalength_inputs.conf +def outer_step(x1, x2): + index = [0] + + def inner_step(ipt): + index[0] += 1 + i = index[0] + outer_mem = memory(name="outer_rnn_state_%d" % i, size=hidden_dim) + + def inner_step_impl(y): + inner_mem = memory( + name="inner_rnn_state_" + y.name, + size=hidden_dim, + boot_layer=outer_mem) + out = fc_layer( + input=[y, inner_mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name='inner_rnn_state_' + y.name) + return out + + encoder = recurrent_group( + step=inner_step_impl, name='inner_%d' % i, input=ipt) + last = last_seq(name="outer_rnn_state_%d" % i, input=encoder) + return encoder, last + + encoder1, sentence_last_state1 = inner_step(ipt=x1) + encoder2, sentence_last_state2 = inner_step(ipt=x2) + + encoder1_expand = expand_layer( + input=sentence_last_state1, expand_as=encoder2) + + return [encoder1_expand, encoder2] + + +encoder1_rep, encoder2_rep = recurrent_group( + name="outer", + step=outer_step, + input=[SubsequenceInput(emb1), SubsequenceInput(emb2)], + targetInlink=emb2) + +encoder1_last = last_seq(input=encoder1_rep) +encoder1_expandlast = expand_layer(input=encoder1_last, expand_as=encoder2_rep) +context = mixed_layer( + input=[ + identity_projection(encoder1_expandlast), + identity_projection(encoder2_rep) + ], + size=hidden_dim) + +rep = last_seq(input=context) +prob = fc_layer( + size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) + +outputs( + classification_cost( + input=prob, label=data_layer( + name="label", size=label_dim))) diff --git a/paddle/legacy/gserver/tests/sequence_recurrent.py b/paddle/legacy/gserver/tests/sequence_recurrent.py new file mode 100644 index 0000000000000000000000000000000000000000..e2c6a7935c28838fb12fc6e44d99dd59636bf7dd --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_recurrent.py @@ -0,0 +1,55 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +dict_path = 'legacy/gserver/tests/Sequence/tour_dict_phrase.dict' +dict_file = dict() +for line_count, line in enumerate(open(dict_path, "r")): + dict_file[line.strip()] = line_count + +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/train.list', + test_list=None, + module='sequenceGen', + obj='process', + args={"dict_file": dict_file}) + +settings(batch_size=5) +######################## network configure ################################ +dict_dim = len(open(dict_path, 'r').readlines()) +word_dim = 128 +hidden_dim = 128 +label_dim = 3 + +# This config is designed to be equivalent with sequence_recurrent_group.py + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer( + input=data, size=word_dim, param_attr=ParamAttr(name="emb")) + +recurrent = recurrent_layer(input=emb, bias_attr=False, act=SoftmaxActivation()) + +recurrent_last = last_seq(input=recurrent) + +with mixed_layer( + size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: + output += full_matrix_projection(input=recurrent_last) + +outputs( + classification_cost( + input=output, label=data_layer( + name="label", size=1))) diff --git a/paddle/legacy/gserver/tests/sequence_recurrent_group.py b/paddle/legacy/gserver/tests/sequence_recurrent_group.py new file mode 100644 index 0000000000000000000000000000000000000000..b4638bd9075ff5cdd4a5ed1bc0e0d133f9a9ab86 --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_recurrent_group.py @@ -0,0 +1,68 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from paddle.trainer_config_helpers import * + +######################## data source ################################ +dict_path = 'legacy/gserver/tests/Sequence/tour_dict_phrase.dict' +dict_file = dict() +for line_count, line in enumerate(open(dict_path, "r")): + dict_file[line.strip()] = line_count + +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/train.list', + test_list=None, + module='sequenceGen', + obj='process', + args={"dict_file": dict_file}) + +settings(batch_size=5) +######################## network configure ################################ +dict_dim = len(open(dict_path, 'r').readlines()) +word_dim = 128 +hidden_dim = 128 +label_dim = 3 + +# This config is designed to be equivalent with sequence_recurrent.py + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer( + input=data, size=word_dim, param_attr=ParamAttr(name="emb")) + + +def step(y): + mem = memory(name="rnn_state", size=hidden_dim) + with mixed_layer( + name="rnn_state", + size=hidden_dim, + bias_attr=False, + act=SoftmaxActivation()) as out: + out += identity_projection(input=y) + out += full_matrix_projection( + input=mem, param_attr=ParamAttr(name="___recurrent_layer_0__")) + return out + + +recurrent = recurrent_group(name="rnn", step=step, input=emb) + +recurrent_last = last_seq(input=recurrent) + +with mixed_layer( + size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output: + output += full_matrix_projection(input=recurrent_last) + +outputs( + classification_cost( + input=output, label=data_layer( + name="label", size=1))) diff --git a/paddle/legacy/gserver/tests/sequence_rnn.conf b/paddle/legacy/gserver/tests/sequence_rnn.conf new file mode 100644 index 0000000000000000000000000000000000000000..3133595c9ce4c25683c06d326a5ebe9d2bf13077 --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_rnn.conf @@ -0,0 +1,57 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2(train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_seq') + + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 8 +hidden_dim = 8 +label_dim = 3 + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer(input=data, size=word_dim) + +def step(y): + mem = memory(name="rnn_state", size=hidden_dim) + out = fc_layer(input=[y, mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name="rnn_state") + return out + +out = recurrent_group( + name="rnn", + step=step, + input=emb) + +rep = last_seq(input=out) +prob = fc_layer(size=label_dim, + input=rep, + act=SoftmaxActivation(), + bias_attr=True) + +outputs(classification_cost(input=prob, + label=data_layer(name="label", size=label_dim))) diff --git a/paddle/legacy/gserver/tests/sequence_rnn_matched_inputs.py b/paddle/legacy/gserver/tests/sequence_rnn_matched_inputs.py new file mode 100644 index 0000000000000000000000000000000000000000..921cef04dda0da396a79592b09d7a7e7177462d5 --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_rnn_matched_inputs.py @@ -0,0 +1,84 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_mixed') + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 2 +hidden_dim = 2 +label_dim = 2 + +data1 = data_layer(name="word1", size=dict_dim) +data2 = data_layer(name="word2", size=dict_dim) +label = data_layer(name="label", size=label_dim) + +encoding = embedding_layer(input=data2, size=word_dim) + +subseq = embedding_layer(input=data1, size=word_dim) +seq = embedding_layer(input=data2, size=word_dim) +nonseq = embedding_layer(input=label, size=word_dim) + + +# This hierarchical RNN is designed to be equivalent to the simple RNN in +# sequence_rnn_mixed_inputs.conf +def outer_step(subseq, seq, nonseq, encoding): + outer_mem = memory(name="outer_rnn_state", size=hidden_dim) + + def inner_step(subseq, seq, nonseq): + inner_mem = memory( + name="inner_rnn_state", size=hidden_dim, boot_layer=outer_mem) + + out = fc_layer( + input=[subseq, seq, nonseq, inner_mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name='inner_rnn_state') + return out + + decoder = recurrent_group( + step=inner_step, name='inner', input=[subseq, seq, nonseq]) + last = last_seq(name="outer_rnn_state", input=decoder) + context = simple_attention( + encoded_sequence=encoding, encoded_proj=encoding, decoder_state=last) + return context + + +out = recurrent_group( + name="outer", + step=outer_step, + input=[ + subseq, expand_layer( + seq, expand_as=subseq, + expand_level=ExpandLevel.FROM_SEQUENCE), expand_layer( + nonseq, + expand_as=subseq, + expand_level=ExpandLevel.FROM_NO_SEQUENCE), + StaticInput(encoding) + ]) + +rep = last_seq(input=out) +prob = fc_layer( + size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) + +outputs(classification_cost(input=prob, label=label)) diff --git a/paddle/legacy/gserver/tests/sequence_rnn_mixed_inputs.py b/paddle/legacy/gserver/tests/sequence_rnn_mixed_inputs.py new file mode 100644 index 0000000000000000000000000000000000000000..c7bcaf6c4b21272e1c95d6de7e69e4558d52b9c6 --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_rnn_mixed_inputs.py @@ -0,0 +1,78 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_mixed') + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 2 +hidden_dim = 2 +label_dim = 2 + +data1 = data_layer(name="word1", size=dict_dim) +data2 = data_layer(name="word2", size=dict_dim) +label = data_layer(name="label", size=label_dim) + +encoding = embedding_layer(input=data2, size=word_dim) + + +# This hierarchical RNN is designed to be equivalent to the simple RNN in +# sequence_rnn_matched_inputs.conf +def outer_step(subseq, seq, nonseq, encoding): + outer_mem = memory(name="outer_rnn_state", size=hidden_dim) + + def inner_step(data1, data2, label): + inner_mem = memory( + name="inner_rnn_state", size=hidden_dim, boot_layer=outer_mem) + + subseq = embedding_layer(input=data1, size=word_dim) + seq = embedding_layer(input=data2, size=word_dim) + nonseq = embedding_layer(input=label, size=word_dim) + + print_layer(input=[data1, seq, label, inner_mem]) + out = fc_layer( + input=[subseq, seq, nonseq, inner_mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name='inner_rnn_state') + return out + + decoder = recurrent_group( + step=inner_step, name='inner', + input=[subseq, StaticInput(seq), nonseq]) + last = last_seq(name="outer_rnn_state", input=decoder) + context = simple_attention( + encoded_sequence=encoding, encoded_proj=encoding, decoder_state=last) + return context + + +out = recurrent_group( + name="outer", + step=outer_step, + input=[data1, data2, StaticInput(label), StaticInput(encoding)]) + +rep = last_seq(input=out) +prob = fc_layer( + size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) + +outputs(classification_cost(input=prob, label=label)) diff --git a/paddle/legacy/gserver/tests/sequence_rnn_multi_input.conf b/paddle/legacy/gserver/tests/sequence_rnn_multi_input.conf new file mode 100644 index 0000000000000000000000000000000000000000..bf4be779a23e081cef33ce2b2734ad91cfa33c0d --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_rnn_multi_input.conf @@ -0,0 +1,58 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2(train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_seq') + + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 8 +hidden_dim = 8 +label_dim = 3 + +data = data_layer(name="word", size=dict_dim) + +emb = embedding_layer(input=data, size=word_dim) + +def step(y, wid): + z = embedding_layer(input=wid, size=word_dim) + mem = memory(name="rnn_state", size=hidden_dim) + out = fc_layer(input=[y, z, mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name="rnn_state") + return out + +out = recurrent_group( + name="rnn", + step=step, + input=[emb, data]) + +rep = last_seq(input=out) +prob = fc_layer(size=label_dim, + input=rep, + act=SoftmaxActivation(), + bias_attr=True) + +outputs(classification_cost(input=prob, + label=data_layer(name="label", size=label_dim))) diff --git a/paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py b/paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py new file mode 100644 index 0000000000000000000000000000000000000000..3612b49c2279874a378d4aaed81623f7d0d2ea2f --- /dev/null +++ b/paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py @@ -0,0 +1,76 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. + +from paddle.trainer_config_helpers import * + +######################## data source ################################ +define_py_data_sources2( + train_list='legacy/gserver/tests/Sequence/dummy.list', + test_list=None, + module='rnn_data_provider', + obj='process_unequalength_seq') + +settings(batch_size=2, learning_rate=0.01) +######################## network configure ################################ +dict_dim = 10 +word_dim = 8 +hidden_dim = 8 +label_dim = 2 + +speaker1 = data_layer(name="word1", size=dict_dim) +speaker2 = data_layer(name="word2", size=dict_dim) + +emb1 = embedding_layer(input=speaker1, size=word_dim) +emb2 = embedding_layer(input=speaker2, size=word_dim) + +# This hierachical RNN is designed to be equivalent to the RNN in +# sequence_nest_rnn_multi_unequalength_inputs.conf + + +def step(x1, x2): + def calrnn(y): + mem = memory(name='rnn_state_' + y.name, size=hidden_dim) + out = fc_layer( + input=[y, mem], + size=hidden_dim, + act=TanhActivation(), + bias_attr=True, + name='rnn_state_' + y.name) + return out + + encoder1 = calrnn(x1) + encoder2 = calrnn(x2) + return [encoder1, encoder2] + + +encoder1_rep, encoder2_rep = recurrent_group( + name="stepout", step=step, input=[emb1, emb2]) + +encoder1_last = last_seq(input=encoder1_rep) +encoder1_expandlast = expand_layer(input=encoder1_last, expand_as=encoder2_rep) +context = mixed_layer( + input=[ + identity_projection(encoder1_expandlast), + identity_projection(encoder2_rep) + ], + size=hidden_dim) + +rep = last_seq(input=context) +prob = fc_layer( + size=label_dim, input=rep, act=SoftmaxActivation(), bias_attr=True) + +outputs( + classification_cost( + input=prob, label=data_layer( + name="label", size=label_dim))) diff --git a/paddle/legacy/gserver/tests/test_ActivationGrad.cpp b/paddle/legacy/gserver/tests/test_ActivationGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f468d229a889e02bf79baa29576c638acbd8eb08 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_ActivationGrad.cpp @@ -0,0 +1,98 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_bool(thread_local_rand_use_global_seed); + +void testActivation(const string& act) { + LOG(INFO) << "test activation: " << act; + size_t size = 10; + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("addto"); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type(act); + config.inputDefs.push_back({INPUT_DATA, "layer_0", size, 0}); + config.layerConfig.add_inputs(); + for (auto useGpu : {false, true}) { + testLayerGrad(config, + act + "_activation", + 100, + /* trans= */ false, + useGpu, + /* useWeight */ true); + } +} + +TEST(Activation, activation) { + auto types = ActivationFunction::getAllRegisteredTypes(); + std::set excluded{"sequence_softmax"}; + for (auto type : types) { + if (excluded.count(type)) continue; + testActivation(type); + } +} + +void testSequenceSoftmaxAct(bool hasSubseq) { + LOG(INFO) << "test activation: sequence softmax"; + + const size_t size = 1; + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("addto"); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type("sequence_softmax"); + config.inputDefs.push_back( + {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, + "layer_0", + 1, + 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "sequence_softmax", + 100, + /* trans= */ false, + useGpu, + /* useWeight */ true); + } +} + +TEST(SequenceSoftmaxActivation, activation) { + for (auto hasSubseq : {false, true}) { + LOG(INFO) << "hasSubseq = " << hasSubseq; + testSequenceSoftmaxAct(hasSubseq); + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_BatchNorm.cpp b/paddle/legacy/gserver/tests/test_BatchNorm.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e21fa16074406645be88eeb454d743531f825041 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_BatchNorm.cpp @@ -0,0 +1,195 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/utils/GlobalConstants.h" + +#include "LayerGradUtil.h" +#include "paddle/legacy/cuda/include/hl_batch_norm.h" +#include "paddle/legacy/math/tests/TensorCheck.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_double(checkgrad_eps); +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_bool(prev_batch_state); + +// Test that the batchNormLayer can be followed by a ConvLayer +TEST(Layer, batchNorm) { + FLAGS_use_gpu = false; + TestConfig configBN; + const int CHANNELS = 6272; + const int IMG_SIZE = 1; + configBN.layerConfig.set_type("batch_norm"); + configBN.layerConfig.set_name("bn"); + configBN.layerConfig.set_size(CHANNELS * IMG_SIZE * IMG_SIZE); + configBN.layerConfig.set_active_type("relu"); + configBN.biasSize = CHANNELS; + configBN.inputDefs.push_back({INPUT_DATA, + "layer_0", + /* dim= */ IMG_SIZE * IMG_SIZE * CHANNELS, + /* paraSize= */ CHANNELS}); + + configBN.inputDefs.push_back( + {INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); + configBN.inputDefs.back().isStatic = true; + configBN.inputDefs.push_back( + {INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); + configBN.inputDefs.back().isStatic = true; + + LayerInputConfig* input = configBN.layerConfig.add_inputs(); + configBN.layerConfig.add_inputs(); + configBN.layerConfig.add_inputs(); + + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(CHANNELS); + img_conf->set_img_size(IMG_SIZE); + + // Setting up conv-layer config + TestConfig config; + config.biasSize = 64; + config.layerConfig.set_type("exconv"); + config.layerConfig.set_num_filters(64); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + config.inputDefs.push_back({INPUT_DATA, "bn", 6272, 204800}); + input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(5); + conv->set_filter_size_y(5); + conv->set_channels(128); + conv->set_padding(1); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + conv->set_img_size(7); + conv->set_output_x(3); + config.layerConfig.set_size(conv->output_x() * conv->output_x() * + config.layerConfig.num_filters()); + config.layerConfig.set_name("conv"); + + // data layer initialize + std::vector dataLayers; + LayerMap layerMap; + vector datas; + initDataLayer(configBN, + &dataLayers, + &datas, + &layerMap, + "batch_norm", + 100, + false, + false); + // test layer initialize + std::vector parameters; + LayerPtr bnLayer; + initTestLayer(configBN, &layerMap, ¶meters, &bnLayer); + + std::vector parameters2; + LayerPtr convLayer; + initTestLayer(config, &layerMap, ¶meters2, &convLayer); + + bnLayer->forward(PASS_GC); + convLayer->forward(PASS_GC); + + CHECK_EQ(static_cast(convLayer->getOutputValue()->getHeight()), 100); + CHECK_EQ(static_cast(convLayer->getOutputValue()->getWidth()), 576); +} + +#ifdef PADDLE_WITH_CUDA +void batchNormInference(int n, int c, int h, int w) { + MatrixPtr input = std::make_shared(n, c * h * w); + MatrixPtr cudnnOut = std::make_shared(n, c * h * w); + MatrixPtr cudaOut = std::make_shared(n, c * h * w); + MatrixPtr cudnnCheck = std::make_shared(n, c * h * w); + MatrixPtr cudaCheck = std::make_shared(n, c * h * w); + input->randomizeUniform(); + cudnnOut->zeroMem(); + cudaOut->zeroMem(); + + MatrixPtr scale = std::make_shared(1, c); + scale->randomizeUniform(); + MatrixPtr bias = std::make_shared(1, c); + bias->randomizeUniform(); + + MatrixPtr movingMean = std::make_shared(1, c); + movingMean->randomizeUniform(); + + MatrixPtr movingVar = std::make_shared(1, c); + movingVar->randomizeUniform(); + movingVar->clip(0.01, 50); + + hl_tensor_descriptor ioDesc; + hl_tensor_descriptor bnDesc; + hl_create_tensor_descriptor(&ioDesc); + hl_create_tensor_descriptor(&bnDesc); + hl_tensor_reshape(ioDesc, n, c, h, w); + hl_tensor_reshape(bnDesc, 1, c, 1, 1); + + double EPS = 1E-5; + hl_batch_norm_forward_inference(ioDesc, + input->getData(), + ioDesc, + cudnnOut->getData(), + bnDesc, + scale->getData(), + bias->getData(), + movingMean->getData(), + movingVar->getData(), + EPS); + + hl_batch_norm_cuda_inference(input->getData(), + cudaOut->getData(), + scale->getData(), + bias->getData(), + movingMean->getData(), + movingVar->getData(), + EPS, + n, + c, + h, + w); + + cudnnCheck->copyFrom(*cudnnOut); + cudaCheck->copyFrom(*cudaOut); + autotest::TensorCheckErr(*cudnnCheck, *cudaCheck); + + hl_destroy_tensor_descriptor(ioDesc); + hl_destroy_tensor_descriptor(bnDesc); +} + +TEST(BatchNorm, Inference) { + batchNormInference(33, 267, 1, 1); + batchNormInference(19, 105, 4, 4); +} +#endif + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_CRFLayerGrad.cpp b/paddle/legacy/gserver/tests/test_CRFLayerGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1dafd1de4d82f1d306626090c30cf9203fa24dd0 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_CRFLayerGrad.cpp @@ -0,0 +1,173 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/gserver/layers/LinearChainCRF.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +static inline bool getNextSequence(std::vector& seq, int numClasses) { + for (auto& v : seq) { + if (++v < numClasses) { + return true; + } + v = 0; + } + return false; +} + +// log(exp(x) + exp(y)) +static inline real logSum(real x, real y) { + real maxValue = std::max(x, y); + if (std::isinf(maxValue)) { + return -std::numeric_limits::infinity(); + } else { + return maxValue + log(exp(x - maxValue) + exp(y - maxValue)); + } +} + +static inline std::vector genRandLabels(int numClasses, int length) { + std::vector labels(length); + for (int i = 0; i < length; ++i) { + labels[i] = rand() % numClasses; // NOLINT + } + return labels; +} + +TEST(CRFLayer, cost) { + const int numClasses = 4; + CpuVector para(numClasses * (numClasses + 2)); + real* a = para.getData(); + real* b = para.getData() + numClasses; + real* w = para.getData() + 2 * numClasses; + LinearChainCRF crf(4, para.getData()); + for (int length : {1, 2, 3, 10}) { + for (int tries = 0; tries < 10; ++tries) { + CpuMatrix x(length, numClasses); + x.randomizeUniform(); + para.randnorm(0, 2); + + std::vector goldenLabels = genRandLabels(numClasses, length); + + real cost = crf.forward(x.getData(), goldenLabels.data(), length); + + real logZ = -std::numeric_limits::infinity(); + real logNominator = -std::numeric_limits::infinity(); + std::vector testResult(length, 0); + do { + real score = a[testResult.front()]; + score += x.getElement(0, testResult.front()); + for (int k = 1; k < length; ++k) { + score += x.getElement(k, testResult[k]) + + w[numClasses * testResult[k - 1] + testResult[k]]; + } + score += b[testResult.back()]; + logZ = logSum(logZ, score); + + if (goldenLabels == testResult) { + logNominator = score; + } + } while (getNextSequence(testResult, numClasses)); + + real trueCost = -logNominator + logZ; + + real diff = fabs(trueCost - cost); + diff /= fabs(cost) < fabs(trueCost) ? fabs(cost) : fabs(trueCost); + VLOG(1) << "cost=" << cost << " trueCost=" << trueCost << " diff=" << diff + << std::endl; + if (typeid(real) == typeid(double)) { // NOLINT + EXPECT_LE(diff, 1e-10); + } else { + EXPECT_LE(diff, 5e-3); + } + } + } +} + +inline real epsilon() { return typeid(real) == typeid(double) ? 1e-10 : 0.06; } + +TestConfig initTestConfig(size_t numClasses, bool withWeight) { + TestConfig config; + config.layerConfig.set_type("crf"); + config.layerConfig.set_size(numClasses); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, + "layer_0", + numClasses, + numClasses * (numClasses + 2)}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back( + {INPUT_SEQUENCE_LABEL, "layer_label", numClasses, 0}); + config.layerConfig.add_inputs(); + + if (withWeight) { + config.inputDefs.push_back({INPUT_DENSE_DIM_DATA, "layer_weight", 1, 0}); + config.layerConfig.add_inputs(); + } + + return config; +} + +TEST(Layer, CRFLayer) { + size_t numClasses = 10; + for (int tries = 0; tries < 5; ++tries) { + TestConfig config = initTestConfig(numClasses, /* withWeight= */ false); + for (int length : {1, 3, 100}) { + // Not support GPU now + testLayerGrad(config, + "crf", + length, + /* trans= */ false, + /* useGpu= */ false, + /* useWeight= */ false, + epsilon()); + } + } +} + +TEST(Layer, CRFLayerUseWeight) { + size_t numClasses = 10; + for (int tries = 0; tries < 5; ++tries) { + TestConfig config = initTestConfig(numClasses, /* withWeight= */ true); + for (int length : {1, 3, 100}) { + // Not support GPU now + testLayerGrad(config, + "crf", + length, + /* trans= */ false, + /* useGpu= */ false, + /* useWeight= */ false, + epsilon()); + } + } +} + +int main(int argc, char** argv) { + initMain(argc, argv); + hl_start(); + hl_init(FLAGS_gpu_id); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_CompareSparse.cpp b/paddle/legacy/gserver/tests/test_CompareSparse.cpp new file mode 100644 index 0000000000000000000000000000000000000000..11b633a5885180ae227f6e93330117b567d4a4ab --- /dev/null +++ b/paddle/legacy/gserver/tests/test_CompareSparse.cpp @@ -0,0 +1,228 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "paddle/legacy/trainer/Trainer.h" + +#include +#include + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static const string& configFile1 = "legacy/gserver/tests/sequence_lstm.conf"; + +DECLARE_bool(use_gpu); +DECLARE_string(config); +DECLARE_int32(gpu_id); +DECLARE_int32(seed); +DECLARE_int32(num_passes); +DECLARE_int32(saving_period); + +DECLARE_int32(num_gradient_servers); +DECLARE_int32(port); +DECLARE_bool(local); +DECLARE_bool(use_old_updater); +DECLARE_bool(parallel_nn); +DECLARE_string(config_args); +DEFINE_double(max_diff_ratio, + 0.0f, + "max diff ratio allowed for parameters value"); + +int gNumDevices = 0; + +std::vector trainerOnePassTest(const string& configFile, + bool sparseUpdate, + int trainerCount = 1, + bool useGpu = false) { + FLAGS_use_gpu = useGpu; + FLAGS_config = configFile; + FLAGS_trainer_count = trainerCount; + FLAGS_config_args = sparseUpdate ? "sparse_update=1" : "sparse_update=0"; + + LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount + << " configFile=" << configFile << " sparseUpdate=" << sparseUpdate; + srand(FLAGS_seed); + *ThreadLocalRand::getSeed() = FLAGS_seed; + ThreadLocalRandomEngine::get().seed(FLAGS_seed); + if (useGpu) { + CHECK_LE(trainerCount, gNumDevices); + } + + std::vector> pservers; + if (!FLAGS_local) { + int numPorts = FLAGS_ports_num + FLAGS_ports_num_for_sparse; + pservers.resize(numPorts); + + for (int i = 0; i < numPorts; ++i) { + pservers[i].reset(new ParameterServer2(std::string(), FLAGS_port + i)); + pservers[i]->init(); + pservers[i]->start(); + } + } + + Trainer trainer; + trainer.init(TrainerConfigHelper::createFromFlagConfig()); + trainer.train(); + return trainer.getGradientMachine()->getParameters(); +} + +std::vector& getDenseParameters() { + static std::vector denseParameters; + if (denseParameters.empty()) { + // use dense training as base + FLAGS_local = true; + denseParameters = trainerOnePassTest(configFile1, false); + } + + return denseParameters; +} + +void checkBuffer(real* A, + const char* desA, + real* B, + const char* desB, + size_t len, + double maxDiffRatio) { + double maxDiff = 0; + double maxValue = 0; + for (size_t i = 0; i < len; ++i) { + double diff = fabs(A[i] - B[i]); + maxValue = std::max(maxValue, std::max(fabs(A[i]), fabs(B[i]))); + maxDiff = std::max(maxDiff, diff); + } + EXPECT_LE(maxDiff / maxValue, maxDiffRatio); + LOG(INFO) << " maxDiff=" << maxDiff << " maxValue=" << maxValue + << " maxDiff/maxValue=" << maxDiff / maxValue << "\n\n"; +} + +void compareValue(const vector& parametersA, + const vector& parametersB, + double maxDiffRatio = 0.0) { + LOG(INFO) << "\n\n--------------------------------" + << " Check Gradient Machine Parameters:" + << " -------------------------------------\n"; + for (size_t i = 0; i < parametersA.size(); ++i) { + ParameterPtr parameterA, parameterB; + parameterA = parametersA[i]; + parameterB = parametersB[i]; + + CpuVector paraA(parameterA->getSize()); + CpuVector paraB(parameterB->getSize()); + paraA.copyFrom(*parameterA->getBuf(PARAMETER_VALUE)); + paraB.copyFrom(*parameterB->getBuf(PARAMETER_VALUE)); + + LOG(INFO) << "\n\n----------- PARAMETER_VALUE: " << parameterA->getName() + << " ; size : " << paraA.getSize() << " ------------"; + checkBuffer(paraA.getData(), + "para_A", + paraB.getData(), + "para_B", + paraA.getSize(), + maxDiffRatio); + } +} + +TEST(compareSparse, cpu) { + FLAGS_local = 1; // disable remote sparse update in parameter config + std::vector parameters = trainerOnePassTest(configFile1, true); + compareValue(getDenseParameters(), parameters); +} + +TEST(compareSparse, remote_cpu) { + FLAGS_local = 0; // will enable remote sparse update + FLAGS_ports_num_for_sparse = 5; + std::vector parameters = trainerOnePassTest(configFile1, true); + compareValue(getDenseParameters(), parameters); +} + +TEST(compareSparse, cpu10_local_vs_remote) { + FLAGS_local = 1; // disable remote sparse update in parameter config + std::vector localParameters = + trainerOnePassTest(configFile1, true, 2); + + FLAGS_local = 0; // will enable remote sparse update + FLAGS_ports_num_for_sparse = 5; + std::vector remoteParameters = + trainerOnePassTest(configFile1, true, 2); + + compareValue(localParameters, remoteParameters); +} + +TEST(compareSparse, multiGradientMachine) { + int numGpu; +#ifdef PADDLE_TYPE_DOUBLE + double eps = 1e-8; +#else + double eps = 1e-4; +#endif + numGpu = hl_get_device_count(); + for (bool local : {false, true}) { + FLAGS_local = local; + FLAGS_ports_num_for_sparse = 5; + for (bool useGpu : {false, true}) { +#ifndef PADDLE_WITH_CUDA + if (useGpu) continue; +#endif + FLAGS_parallel_nn = useGpu; + LOG(INFO) << " local=" << local << " useGpu=" << useGpu; + int trainerCount = useGpu ? numGpu : 2; + std::vector parameters = + trainerOnePassTest(configFile1, true, trainerCount, useGpu); + compareValue(getDenseParameters(), parameters, eps); + } + } + FLAGS_parallel_nn = false; +} + +TEST(compareSparse, NeuralNetwork) { +#ifdef PADDLE_TYPE_DOUBLE + double eps = 1e-8; +#else + double eps = 1e-4; +#endif + for (bool local : {false, true}) { + FLAGS_local = local; + FLAGS_ports_num_for_sparse = 5; + for (bool useGpu : {false, true}) { +#ifndef PADDLE_WITH_CUDA + if (useGpu) continue; +#endif + FLAGS_parallel_nn = useGpu; + LOG(INFO) << " local=" << local << " useGpu=" << useGpu; + int trainerCount = 1; + std::vector parameters = + trainerOnePassTest(configFile1, true, trainerCount, useGpu); + compareValue(getDenseParameters(), parameters, useGpu ? eps : 0); + } + } + FLAGS_parallel_nn = false; +} + +int main(int argc, char** argv) { + // FIXME(tonyyang-svail): + // Turn off this test due CI failure: + // https://paddleci.ngrok.io/viewLog.html?buildId=27608&buildTypeId=Paddle_PrCi&tab=buildLog&_focus=10430 + return 0; + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + initPython(argc, argv); + + gNumDevices = hl_get_device_count(); + FLAGS_num_passes = 1; // train one pass + FLAGS_saving_period = 100000; // do not save parameter + + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_CompareTwoNets.cpp b/paddle/legacy/gserver/tests/test_CompareTwoNets.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e19c34abbd8a84660a9e79bcbf602437bfc92832 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_CompareTwoNets.cpp @@ -0,0 +1,210 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include + +#include "paddle/legacy/trainer/Trainer.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_int32(gpu_id); + +DECLARE_bool(local); +DECLARE_bool(use_gpu); + +DECLARE_string(config); +DECLARE_string(nics); + +DEFINE_bool(need_high_accuracy, + false, + "whether need to run in double accuracy"); +DEFINE_double( + max_diff_ratio, + 0.0f, + "max diff ratio allowed for outputs and parameters (value/gradient)"); +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_int32(seed); + +static const string& config_file_a = + "legacy/gserver/tests/sequence_recurrent.py"; +static const string& config_file_b = + "legacy/gserver/tests/sequence_recurrent_group.py"; + +struct ComData { + vector outArgs; + vector parameters; +}; + +void calcGradient(ComData& data, const string configFile) { + FLAGS_config = configFile; + + FLAGS_local = true; + FLAGS_use_gpu = false; + + FLAGS_nics = ""; + + *ThreadLocalRand::getSeed() = FLAGS_seed; + srand(FLAGS_seed); + + Trainer trainer; + trainer.init(TrainerConfigHelper::createFromFlagConfig(), false); + + data.parameters = trainer.getGradientMachine()->getParameters(); + + DataBatch dataBatch; + int32_t batchSize = trainer.getConfig().opt_config().batch_size(); + + trainer.getDataProvider()->reset(); + trainer.getDataProvider()->setSkipShuffle(); + trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch); + + CHECK(dataBatch.getSize()) << "No data from data provider"; + vector& inArgs = dataBatch.getStreams(); + + trainer.getGradientMachine()->start(); + trainer.getGradientMachine()->forwardBackward( + inArgs, &data.outArgs, PASS_TRAIN); + + trainer.getGradientMachine()->finish(); +} + +void checkBuffer(real* A, + const char* desA, + real* B, + const char* desB, + size_t len, + size_t width = 1) { + int nNum = 0; + real maxVal = 0; + for (size_t i = 0; i < len; ++i) { + maxVal = std::max(maxVal, std::max(A[i], B[i])); + } + real maxDiff = 0; + for (size_t i = 0; i < len; ++i) { + real diff = fabs(A[i] - B[i]); + maxDiff = std::max(maxDiff, diff); + if (diff > maxVal * FLAGS_max_diff_ratio) { + nNum++; + VLOG(1) << "Row: " << i / width << ", " << desA << " : " << A[i] << " " + << desB << " : " << B[i] << " diff=" << diff; + } + } + EXPECT_EQ(0, nNum); + LOG(INFO) << "maxValue=" << maxVal << " maxDiff=" << maxDiff << "\n\n"; +} + +void compareGradient(ComData& comDataA, ComData& comDataB) { + vector outArgsA = comDataA.outArgs; + vector outArgsB = comDataB.outArgs; + + for (size_t i = 0; i < outArgsA.size(); ++i) { + CpuMatrix matA(outArgsA[i].value->getHeight(), + outArgsA[i].value->getWidth()); + CpuMatrix matB(outArgsB[i].value->getHeight(), + outArgsB[i].value->getWidth()); + + matA.copyFrom(*outArgsA[i].value); + matB.copyFrom(*outArgsB[i].value); + + LOG(INFO) << "\n--------------------------------" + << " Check Network Output_" << i << ":" + << " -------------------------------------\n"; + checkBuffer(matA.getData(), + "network A output", + matB.getData(), + "network B output", + matA.getElementCnt(), + matA.getWidth()); + } + + vector& parametersA = comDataA.parameters; + vector& parametersB = comDataB.parameters; + + LOG(INFO) << "\n\n--------------------------------" + << " Check Gradient Machine Parameters:" + << " -------------------------------------\n"; + for (size_t i = 0; i < parametersA.size(); ++i) { + ParameterPtr parameterA, parameterB; + parameterA = parametersA[i]; + parameterB = parametersB[i]; + + CpuVector paraA(parameterA->getSize()); + CpuVector paraB(parameterB->getSize()); + paraA.copyFrom(*parameterA->getBuf(PARAMETER_VALUE)); + paraB.copyFrom(*parameterB->getBuf(PARAMETER_VALUE)); + + LOG(INFO) << "\n\n----------- PARAMETER_VALUE: " << parameterA->getName() + << " ; size : " << paraA.getSize() << " ------------"; + checkBuffer(paraA.getData(), + "Network A", + paraB.getData(), + "Network B", + paraA.getSize()); + + CpuVector gradA(*parameterA->getBuf(PARAMETER_GRADIENT)); + CpuVector gradB(*parameterB->getBuf(PARAMETER_GRADIENT)); + + LOG(INFO) << "\n\n----------- PARAMETER_GRADIENT: " << parameterA->getName() + << " ; size : " << gradA.getSize() << " -----------"; + checkBuffer(gradA.getData(), + "Network A", + gradB.getData(), + "Network B", + gradA.getSize()); + } +} + +TEST(Trainer, create) { + ComData dataA; + calcGradient(dataA, config_file_a); + LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n"; + + ComData dataB; + calcGradient(dataB, config_file_b); + LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n"; + + compareGradient(dataA, dataB); +} + +int main(int argc, char** argv) { + FLAGS_thread_local_rand_use_global_seed = true; + paddle::initMain(argc, argv); + testing::InitGoogleTest(&argc, argv); + initPython(argc, argv); + +#ifndef PADDLE_TYPE_DOUBLE + if (FLAGS_need_high_accuracy) { + LOG(INFO) << "skip test due to it's need high accuracy"; + return 0; + } + if (FLAGS_max_diff_ratio == 0.0f) { + FLAGS_max_diff_ratio = 1e-5; + LOG(INFO) << "auto set max_diff_ratio " << FLAGS_max_diff_ratio + << " in low accuracy mode"; + } +#else + if (FLAGS_max_diff_ratio == 0.0f) { + FLAGS_max_diff_ratio = 1e-10; + LOG(INFO) << "auto set max_diff_ratio " << FLAGS_max_diff_ratio + << " in high accuracy mode"; + } +#endif + + int ret = RUN_ALL_TESTS(); + return ret; +} diff --git a/paddle/legacy/gserver/tests/test_ConvTrans.cpp b/paddle/legacy/gserver/tests/test_ConvTrans.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4ea0a3d379b010fcb6ccb91a28e653a53cfe66d8 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_ConvTrans.cpp @@ -0,0 +1,244 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/GlobalConstants.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_double(checkgrad_eps); +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_bool(prev_batch_state); + +// Test that the convTrans forward is the same as conv backward +TEST(Layer, convTransLayerFwd) { + // Setting up conv-trans layer + TestConfig configt; + configt.biasSize = 3; + configt.layerConfig.set_type("exconvt"); + configt.layerConfig.set_num_filters(3); + configt.layerConfig.set_partial_sum(1); + configt.layerConfig.set_shared_biases(true); + + configt.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 384}); + LayerInputConfig* input = configt.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(2); + conv->set_filter_size_y(4); + conv->set_channels(16); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_groups(1); + conv->set_filter_channels(3 / conv->groups()); + conv->set_img_size(16); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + configt.layerConfig.set_size(conv->img_size() * conv->img_size() * + configt.layerConfig.num_filters()); + configt.layerConfig.set_name("convTrans"); + + // data layer initialize + std::vector dataLayers; + LayerMap layerMap; + vector datas; + initDataLayer( + configt, &dataLayers, &datas, &layerMap, "convTrans", 100, false, false); + // test layer initialize + std::vector parameters; + LayerPtr convtLayer; + initTestLayer(configt, &layerMap, ¶meters, &convtLayer); + convtLayer->getBiasParameter()->zeroMem(); + convtLayer->forward(PASS_GC); + + // Setting up conv-layer config + TestConfig config; + config.biasSize = 16; + config.layerConfig.set_type("exconv"); + config.layerConfig.set_num_filters(16); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + config.inputDefs.push_back({INPUT_DATA, "layer_1", 768, 384}); + input = config.layerConfig.add_inputs(); + conv = input->mutable_conv_conf(); + conv->set_filter_size(2); + conv->set_filter_size_y(4); + conv->set_channels(3); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + conv->set_img_size(16); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + config.layerConfig.set_size(conv->output_x() * conv->output_x() * + config.layerConfig.num_filters()); + config.layerConfig.set_name("conv"); + + // data layer initialize + std::vector dataLayers2; + LayerMap layerMap2; + vector datas2; + initDataLayer( + config, &dataLayers2, &datas2, &layerMap2, "conv", 100, false, false); + // test layer initialize + std::vector parameters2; + LayerPtr convLayer; + initTestLayer(config, &layerMap2, ¶meters2, &convLayer); + + // Sync convLayer and convtLayer parameter + convLayer->getBiasParameter()->zeroMem(); + convLayer->getParameters()[0] + ->getBuf(PARAMETER_VALUE) + ->copyFrom(*(convtLayer->getParameters()[0]->getBuf(PARAMETER_VALUE))); + + // Set convLayer outputGrad as convTransLayer input value + convLayer->forward(PASS_GC); + convLayer->getOutput().grad->copyFrom(*(dataLayers[0]->getOutputValue())); + + vector callbackFlags(parameters2.size(), 0); + auto callback = [&](Parameter* para) { ++callbackFlags[para->getID()]; }; + convLayer->backward(callback); + + // Check that the convLayer backward is the same as convTransLayer forward + checkMatrixEqual(convtLayer->getOutputValue(), + dataLayers2[0]->getOutputGrad()); +} + +// Do one forward pass of convTrans layer and check to see if its output +// matches the given result +void doOneConvtTest(size_t imgSize, + size_t output_x, + size_t stride, + size_t padding, + size_t filter_size, + MatrixPtr& result) { + TestConfig configt; + configt.biasSize = 1; + configt.layerConfig.set_type("exconvt"); + configt.layerConfig.set_num_filters(1); + configt.layerConfig.set_partial_sum(1); + configt.layerConfig.set_shared_biases(true); + + configt.inputDefs.push_back( + {INPUT_DATA, "layer_0", output_x * output_x, filter_size * filter_size}); + LayerInputConfig* input = configt.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(filter_size); + conv->set_filter_size_y(filter_size); + conv->set_channels(1); + conv->set_padding(padding); + conv->set_padding_y(padding); + conv->set_stride(stride); + conv->set_stride_y(stride); + conv->set_groups(1); + conv->set_filter_channels(1); + conv->set_img_size(imgSize); + conv->set_output_x(output_x); + + configt.layerConfig.set_size(conv->img_size() * conv->img_size() * + configt.layerConfig.num_filters()); + configt.layerConfig.set_name("convTrans"); + + std::vector dataLayers; + LayerMap layerMap; + vector datas; + initDataLayer( + configt, &dataLayers, &datas, &layerMap, "convTrans", 1, false, false); + dataLayers[0]->getOutputValue()->zeroMem(); + dataLayers[0]->getOutputValue()->add(1.0); + + // test layer initialize + std::vector parameters; + LayerPtr convtLayer; + initTestLayer(configt, &layerMap, ¶meters, &convtLayer); + convtLayer->getBiasParameter()->zeroMem(); + convtLayer->getParameters()[0]->zeroMem(); + convtLayer->getParameters()[0]->getBuf(PARAMETER_VALUE)->add(1.0); + convtLayer->forward(PASS_GC); + + checkMatrixEqual(convtLayer->getOutputValue(), result); +} + +TEST(Layer, convTransLayerFwd2) { + MatrixPtr result; + result = Matrix::create(1, 5 * 5, false, false); + result->zeroMem(); + result->add(1.0); + doOneConvtTest(/* imgSize */ 5, + /* output_x */ 1, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 5, + result); + + real resultData[] = {1, 2, 2, 2, 1, 2, 4, 4, 4, 2, 2, 4, 4, + 4, 2, 2, 4, 4, 4, 2, 1, 2, 2, 2, 1}; + result->setData(resultData); + doOneConvtTest(/* imgSize */ 5, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 4, + result); + + real resultData2[] = {1, 2, 2, 2, 1, 2, 4, 4, 4, 2, 2, 4, 4, + 4, 2, 2, 4, 4, 4, 2, 1, 2, 2, 2, 1}; + result->setData(resultData2); + doOneConvtTest(/* imgSize */ 5, + /* output_x */ 2, + /* stride */ 2, + /* padding */ 1, + /* filter_size */ 5, + result); + + real resultData3[] = {1, 1, 2, 1, 1, 1, 1, 2, 1, 1, 2, 2, 4, + 2, 2, 1, 1, 2, 1, 1, 1, 1, 2, 1, 1}; + result->setData(resultData3); + doOneConvtTest(/* imgSize */ 5, + /* output_x */ 2, + /* stride */ 2, + /* padding */ 0, + /* filter_size */ 3, + result); +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_ConvUnify.cpp b/paddle/legacy/gserver/tests/test_ConvUnify.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d4ca158352d9e4bf859b31b7c7410518bdc20ac6 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_ConvUnify.cpp @@ -0,0 +1,315 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/GlobalConstants.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_double(checkgrad_eps); +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_bool(prev_batch_state); + +// Do one forward pass of ConvLayer using either exconv or cudnn_conv +MatrixPtr doOneConvTest(size_t imgSize, + size_t output_x, + size_t stride, + size_t padding, + size_t filter_size, + size_t channel, + size_t numfilters, + size_t groups, + MatrixPtr& inputData, + real* param, + bool useGpu, + bool isDeconv = false) { + TestConfig config; + config.biasSize = numfilters; + string layerType; + if (useGpu) { + layerType = (isDeconv) ? "cudnn_convt" : "cudnn_conv"; + } else { + layerType = (isDeconv) ? "exconvt" : "exconv"; + } + config.layerConfig.set_type(layerType); + config.layerConfig.set_num_filters(numfilters); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + size_t weightSize = channel * filter_size * filter_size * + config.layerConfig.num_filters() / groups; + if (isDeconv) { + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", output_x * output_x * channel, weightSize}); + config.layerConfig.set_size(imgSize * imgSize * + config.layerConfig.num_filters()); + } else { + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", imgSize * imgSize * channel, weightSize}); + config.layerConfig.set_size(output_x * output_x * + config.layerConfig.num_filters()); + } + + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(filter_size); + conv->set_filter_size_y(filter_size); + conv->set_channels(channel); + conv->set_padding(padding); + conv->set_padding_y(padding); + conv->set_stride(stride); + conv->set_stride_y(stride); + conv->set_groups(groups); + conv->set_img_size(imgSize); + conv->set_output_x(output_x); + + if (isDeconv) { + conv->set_filter_channels(numfilters / groups); + } else { + conv->set_filter_channels(channel / groups); + } + + config.layerConfig.set_name("conv"); + + std::vector dataLayers; + LayerMap layerMap; + vector datas; + initDataLayer( + config, &dataLayers, &datas, &layerMap, "conv", 1, false, useGpu); + dataLayers[0]->getOutputValue()->zeroMem(); + dataLayers[0]->getOutputValue()->copyFrom(*inputData); + + // test layer initialize + std::vector parameters; + LayerPtr convLayer; + initTestLayer(config, &layerMap, ¶meters, &convLayer); + convLayer->getBiasParameter()->zeroMem(); + convLayer->getParameters()[0]->zeroMem(); + convLayer->getParameters()[0] + ->getBuf(PARAMETER_VALUE) + ->copyFrom(param, weightSize); + convLayer->forward(PASS_GC); + + return convLayer->getOutputValue(); +} + +TEST(Layer, convParaUnified) { +#ifdef PADDLE_WITH_CUDA + MatrixPtr input, resultCpu, resultGpu; + + /// TEST1 for conv /// + input = Matrix::create(1, 4 * 4, false, false); + real inputData[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; + real param[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 8, 7, 6, 5, 4, 3, 2, 1}; + + input->setData(inputData); + + resultCpu = doOneConvTest(/* imgSize */ 4, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 3, + /*channel*/ 1, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param, + /*useGpu*/ false); + + resultGpu = doOneConvTest(/* imgSize */ 4, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 3, + /*channel*/ 1, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param, + /*useGpu*/ true); + checkMatrixEqual(resultCpu, resultGpu); + + /// TEST1 for deconv /// + input = Matrix::create(1, 2 * 2, false, false); + real inputDataT[] = {1, 2, 3, 4}; + input->setData(inputDataT); + + resultCpu = doOneConvTest(/* imgSize */ 4, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 3, + /*channel*/ 1, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param, + /*useGpu*/ false, + /*isDeconv*/ true); + + resultGpu = doOneConvTest(/* imgSize */ 4, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 3, + /*channel*/ 1, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param, + /*useGpu*/ true, + /*isDeconv*/ true); + checkMatrixEqual(resultCpu, resultGpu); + + /// TEST2 for conv /// + input = Matrix::create(1, 3 * 3 * 2, false, false); + real inputData2[] = { + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18}; + real param2[] = {1, 2, 3, 4, 5, 6, 7, 8, 8, 7, 6, 5, 4, 3, 2, 1}; + + input->setData(inputData2); + + resultCpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param2, + /*useGpu*/ false); + + resultGpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param2, + /*useGpu*/ true); + checkMatrixEqual(resultCpu, resultGpu); + + /// TEST3 for conv /// + real param3[] = {1, 2, 3, 4, 4, 3, 2, 1}; + + resultCpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 2, + input, + param3, + /*useGpu*/ false); + + resultGpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 2, + input, + param3, + /*useGpu*/ true); + checkMatrixEqual(resultCpu, resultGpu); + + /// TEST2 for deconv /// + input = Matrix::create(1, 2 * 2 * 2, false, false); + real inputData2T[] = {1, 2, 3, 4, 5, 6, 7, 8}; + input->setData(inputData2T); + + resultCpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param2, + /*useGpu*/ false, + /*isDeconv*/ true); + + resultGpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 1, + input, + param2, + /*useGpu*/ true, + /*isDeconv*/ true); + checkMatrixEqual(resultCpu, resultGpu); + + /// TEST3 for deconv /// + resultCpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 2, + input, + param3, + /*useGpu*/ false, + /*isDeconv*/ true); + + resultGpu = doOneConvTest(/* imgSize */ 3, + /* output_x */ 2, + /* stride */ 1, + /* padding */ 0, + /* filter_size */ 2, + /*channel*/ 2, + /*numfilters*/ 2, + /*groups*/ 2, + input, + param3, + /*useGpu*/ true, + /*isDeconv*/ true); + checkMatrixEqual(resultCpu, resultGpu); +#endif +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_CrossEntropyOverBeamGrad.cpp b/paddle/legacy/gserver/tests/test_CrossEntropyOverBeamGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..34eb0dedffeba46c662a0e69ce9ba82f474a8358 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_CrossEntropyOverBeamGrad.cpp @@ -0,0 +1,352 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +const size_t MAX_SEQ_NUM = 23; +const size_t MAX_SEQ_LEN = 50; +const size_t MAX_BEAM_SIZE = 27; + +const size_t SEED = (size_t)(time(NULL)); + +struct SingleBeamExpansion { + vector seqStartPos; + vector subSeqStartPos; + vector candidateScores; + + // TODO(caoying): store this into Argument.ids + vector selectedIndices; + + vector groundTruth; + vector inBeam; + vector rowIdxInBeam; + vector colIdxInBeam; + + void resetGroundTruth(size_t n) { + groundTruth.clear(); + groundTruth.resize(n, -1); + + inBeam.clear(); + inBeam.resize(n, 0); + + rowIdxInBeam.clear(); + rowIdxInBeam.resize(n, -1); + + colIdxInBeam.clear(); + colIdxInBeam.resize(n, -1); + } +}; + +inline float randFloat() { + return static_cast(rand()) / static_cast(RAND_MAX); +} + +void genRand(real* numbers, size_t n) { + default_random_engine generator; + uniform_real_distribution distribution(0.0, 1.0); + for (size_t i = 0; i < n; ++i) numbers[i] = distribution(generator); +} + +vector randSampling(real range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0.); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + sort(begin(num), end(num)); + return num; +} + +void genCandidateScores(bool hasSubseq, + size_t beamSize, + SingleBeamExpansion& prevBeam, + SingleBeamExpansion& curBeam) { + vector& seqStartPos = curBeam.seqStartPos; + seqStartPos.resize(1, 0); + vector& subSeqStartPos = curBeam.subSeqStartPos; + subSeqStartPos.resize(1, 0); + + srand(SEED); + if (prevBeam.selectedIndices.size()) { + if (prevBeam.subSeqStartPos.size() > 1) { + int seqIdx = 1; + // samples in previous beam are nested sequences. + for (size_t i = 1; i < prevBeam.subSeqStartPos.size(); ++i) { + for (size_t j = 0; j < beamSize; ++j) { + if (prevBeam.selectedIndices[(i - 1) * beamSize + j] == -1.) break; + subSeqStartPos.push_back(1 + (rand() % MAX_SEQ_LEN) + + subSeqStartPos.back()); + } + if (prevBeam.seqStartPos[seqIdx] == prevBeam.subSeqStartPos[i]) { + seqStartPos.push_back(subSeqStartPos.back()); + seqIdx++; + } + } + } else { + for (size_t i = 0; i <= prevBeam.selectedIndices.size(); ++i) { + if (i && i % beamSize == 0) { + seqStartPos.push_back(subSeqStartPos.back()); + if (i == prevBeam.selectedIndices.size()) break; + } + if (prevBeam.selectedIndices[i] == -1.) continue; + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + } + } + } else { + // the first beam expansion + int seqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int i = 0; i < seqNum; ++i) { + if (hasSubseq) { + for (size_t j = 0; j < 1 + (rand() % MAX_SEQ_NUM); ++j) + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + seqStartPos.push_back(subSeqStartPos.back()); + } else { + seqStartPos.push_back(seqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + } + } + } + + size_t totalSeqNum = hasSubseq ? subSeqStartPos.back() : seqStartPos.back(); + curBeam.candidateScores.resize(totalSeqNum, 0.); + genRand(curBeam.candidateScores.data(), totalSeqNum); +} + +void genSelectedIndices(size_t beamSize, + vector& seqStartPos, + vector& selectedIndices) { + size_t selectedIdsCount = beamSize * (seqStartPos.size() - 1); + selectedIndices.resize(selectedIdsCount, -1.); + + for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + int n = min(seqLen, static_cast(beamSize)); + vector ids = randSampling(seqLen, n); + memcpy(selectedIndices.data() + i * beamSize, + ids.data(), + sizeof(real) * ids.size()); + } +} + +void genGroundTruth(vector& beamExpansions, + size_t beamSize) { + SingleBeamExpansion& beam = beamExpansions[1]; + size_t seqNum = beam.seqStartPos.size() - 1; + for (size_t i = 2; i < beamExpansions.size(); ++i) + CHECK_EQ(seqNum, beamExpansions[i].seqStartPos.size() - 1); + + srand(SEED); + + // initialize the first beam. + beam.resetGroundTruth(seqNum); + for (size_t i = 0; i < seqNum; ++i) { + if (randFloat() > 0.5) { + /* + * force the randomly generated label falls in the beam by chance 0.5. + * otherwise, when sequence length is relatively long and beam size is + * relatively small, the gold sequences falls off the beam at in the + * first search. + */ + real* begPos = beam.selectedIndices.data() + i * beamSize; + beam.colIdxInBeam[i] = + rand() % count_if(begPos, begPos + beamSize, [](const real& val) { + return val != -1.; + }); + beam.groundTruth[i] = + beam.selectedIndices[i * beamSize + beam.colIdxInBeam[i]]; + beam.inBeam[i] = 1; + } else { + int label = rand() % (beam.seqStartPos[i + 1] - beam.seqStartPos[i]); + beam.groundTruth[i] = label; + + real* begPos = beam.selectedIndices.data() + i * beamSize; + real* endPos = begPos + beamSize; + real* lblPos = find(begPos, endPos, real(label)); + if (lblPos != endPos) { + beam.inBeam[i] = 1; + beam.colIdxInBeam[i] = lblPos - begPos; + } + } + beam.rowIdxInBeam[i] = i; + } + + // iterate over each beam expansions + for (size_t i = 2; i < beamExpansions.size(); ++i) { + SingleBeamExpansion& curBeam = beamExpansions[i]; + SingleBeamExpansion& prevBeam = beamExpansions[i - 1]; + curBeam.resetGroundTruth(seqNum); + + // iterate over each sequence + for (size_t j = 0; j < seqNum; ++j) { + if (!prevBeam.inBeam[j]) continue; + + // gold sequence falls in the beam in previous search. + real* begPos = prevBeam.selectedIndices.data(); + int offset = + prevBeam.rowIdxInBeam[j] * beamSize + prevBeam.colIdxInBeam[j]; + curBeam.rowIdxInBeam[j] = count_if( + begPos, begPos + offset, [](const real& val) { return val != -1.; }); + + if (randFloat() > 0.5) { + // force the randomly generated label falls in the beam by chance 0.5. + + real* start = + curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize; + int n = rand() % count_if(start, start + beamSize, [](const real& val) { + return val != -1.; + }); + curBeam.colIdxInBeam[j] = n; + curBeam.groundTruth[j] = *(start + n); + curBeam.inBeam[j] = 1; + } else { + CHECK_LE((size_t)curBeam.rowIdxInBeam[j] + 1, + curBeam.subSeqStartPos.size() - 1); + int start = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j]]; + int end = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j] + 1]; + CHECK_GT(size_t(end), size_t(start)); + int label = rand() % (end - start); + + curBeam.groundTruth[j] = label; + real* findBeg = + curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize; + real* lblPos = + find(findBeg, findBeg + beamSize, static_cast(label)); + if (lblPos != (findBeg + beamSize)) { + curBeam.inBeam[j] = 1; + curBeam.colIdxInBeam[j] = lblPos - findBeg; + } + } + } + } +} + +void genOneBeam(size_t beamSize, + bool hasSubseq, + SingleBeamExpansion& prevBeam, + SingleBeamExpansion& curBeam) { + genCandidateScores(hasSubseq, beamSize, prevBeam, curBeam); + genSelectedIndices(beamSize, + hasSubseq ? curBeam.subSeqStartPos : curBeam.seqStartPos, + curBeam.selectedIndices); +} + +void genRandomBeamExpansion(size_t expansionCount, + size_t beamSize, + vector& beamExpansions) { + beamExpansions.clear(); + beamExpansions.resize(expansionCount + 1); + + // beamExpansions[0] is reserved. + for (size_t i = 1; i <= expansionCount; ++i) + genOneBeam(beamSize, bool(i - 1), beamExpansions[i - 1], beamExpansions[i]); + genGroundTruth(beamExpansions, beamSize); +} + +void testCrossEntropyOverBeam(bool useGpu, + size_t beamSize, + vector& beams) { + TestConfig config; + config.layerConfig.set_type("cross_entropy_over_beam"); + + size_t seqNum = 0; + for (size_t i = 1; i < beams.size(); ++i) { + const SingleBeamExpansion& beam = beams[i]; + // create scores for all the candidates + MatrixPtr candidateScorePtr = + Matrix::create(beam.candidateScores.size(), 1, false, false); + candidateScorePtr->copyFrom(beam.candidateScores.data(), + beam.candidateScores.size()); + + ostringstream paramName; + paramName << "candidate_scores_" << i; + + if (beam.subSeqStartPos.size() > 1) { + seqNum = beam.subSeqStartPos.size() - 1; + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + paramName.str(), + candidateScorePtr, + beam.seqStartPos, + beam.subSeqStartPos}); + } else { + seqNum = beam.seqStartPos.size() - 1; + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + paramName.str(), + candidateScorePtr, + beam.seqStartPos}); + } + config.layerConfig.add_inputs(); + + // create indices for the selected candidates + MatrixPtr selectedCandidates = + Matrix::create(seqNum, beamSize, false, false); + selectedCandidates->copyFrom(beam.selectedIndices.data(), + beam.selectedIndices.size()); + paramName.clear(); + paramName << "selected_candidates_" << i; + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, paramName.str(), selectedCandidates}); + config.layerConfig.add_inputs(); + + // create the ground truth + paramName.clear(); + paramName << "label_" << i; + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, paramName.str(), beam.groundTruth}); + config.layerConfig.add_inputs(); + } + + testLayerGrad( + config, "cross_entropy_over_beam", seqNum, false, useGpu, false); +} + +TEST(Layer, CrossEntropyOverBeam) { + LOG(INFO) << "SEED = " << SEED; + const size_t beamSize = 1 + rand() % MAX_BEAM_SIZE; + LOG(INFO) << "beamSize = " << beamSize; + + // TODO(caoying): test with random beam expansions. + const size_t expansionCount = 3; + vector beams; + genRandomBeamExpansion(expansionCount, beamSize, beams); + + for (bool useGpu : {false, true}) + testCrossEntropyOverBeam(useGpu, beamSize, beams); +} + +int main(int argc, char** argv) { + initMain(argc, argv); + hl_start(); + hl_init(FLAGS_gpu_id); + FLAGS_thread_local_rand_use_global_seed = true; + srand(SEED); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/gserver/tests/test_DetectionOutput.cpp b/paddle/legacy/gserver/tests/test_DetectionOutput.cpp similarity index 100% rename from paddle/gserver/tests/test_DetectionOutput.cpp rename to paddle/legacy/gserver/tests/test_DetectionOutput.cpp diff --git a/paddle/legacy/gserver/tests/test_Evaluator.cpp b/paddle/legacy/gserver/tests/test_Evaluator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8aab50d23e56e449d86f22a315c45432253cdd07 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_Evaluator.cpp @@ -0,0 +1,267 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/trainer/Trainer.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +enum InputType { + INPUT_DATA, // dense vector + INPUT_LABEL, // id + INPUT_DATA_TARGET, // dense vector, but no gradient + INPUT_SEQUENCE_DATA, + INPUT_SEQUENCE_LABEL, + INPUT_SPARSE_NON_VALUE_DATA +}; + +struct InputDef { + InputType inputType; + string name; + size_t dim; +}; + +struct TestConfig { + EvaluatorConfig evaluatorConfig; + std::vector inputDefs; + bool testAccumulate; + TestConfig() : testAccumulate(true) {} +}; + +void testEvaluator(TestConfig testConf, + string testEvaluatorName, + size_t batchSize, + bool useGpu) { +#ifndef PADDLE_WITH_CUDA + if (useGpu) return; +#endif + FLAGS_use_gpu = useGpu; + testConf.evaluatorConfig.set_name(testEvaluatorName); + LOG(INFO) << " evaluator_type=" << testConf.evaluatorConfig.type() + << " useGpu=" << useGpu; + + std::vector arguments; + for (size_t i = 0; i < testConf.inputDefs.size(); ++i) { + Argument data; + size_t dim = testConf.inputDefs[i].dim; + switch (testConf.inputDefs[i].inputType) { + case INPUT_DATA: + case INPUT_SEQUENCE_DATA: + case INPUT_DATA_TARGET: + data.value = Matrix::create(batchSize, dim, false, useGpu); + data.value->randomizeUniform(); + + // make sure output > 0 && output < 1 + data.value->add(-0.5); + data.value->sigmoid(*data.value); + break; + case INPUT_LABEL: + case INPUT_SEQUENCE_LABEL: + data.ids = VectorT::create(batchSize, useGpu); + data.ids->rand(dim); // now rand number can be 0 to inputDefs[i].dim. + break; + case INPUT_SPARSE_NON_VALUE_DATA: + data.value = makeRandomSparseMatrix(batchSize, + dim, + /* withValue= */ false, + useGpu); + break; + default: + LOG(FATAL) << " unknown inputType "; + return; + } + + ICpuGpuVectorPtr sequenceStartPositions; + if (testConf.inputDefs[i].inputType == INPUT_SEQUENCE_DATA || + testConf.inputDefs[i].inputType == INPUT_SEQUENCE_LABEL) { + if (!sequenceStartPositions) { + generateSequenceStartPositions(batchSize, sequenceStartPositions); + } + data.sequenceStartPositions = sequenceStartPositions; + } + + arguments.push_back(data); + } + + Evaluator* testEvaluator = Evaluator::create(testConf.evaluatorConfig); + double totalScore = 0.0; + testEvaluator->start(); + totalScore += testEvaluator->evalImp(arguments); + testEvaluator->updateSamplesNum(arguments); + testEvaluator->finish(); + LOG(INFO) << *testEvaluator; + + std::vector names; + testEvaluator->getNames(&names); + paddle::Error err; + for (auto& name : names) { + auto value = testEvaluator->getValue(name, &err); + ASSERT_TRUE(err.isOK()); + LOG(INFO) << name << " " << value; + auto tp = testEvaluator->getType(name, &err); + ASSERT_TRUE(err.isOK()); + ASSERT_EQ(testConf.evaluatorConfig.type(), tp); + } + + double totalScore2 = 0.0; + if (testConf.testAccumulate) { + testEvaluator->start(); + totalScore2 += testEvaluator->evalImp(arguments); + testEvaluator->finish(); + EXPECT_LE(fabs(totalScore - totalScore2), 1.0e-5); + } +} + +void testEvaluatorAll(TestConfig testConf, + string testEvaluatorName, + size_t batchSize) { + testEvaluator(testConf, testEvaluatorName, batchSize, true); + testEvaluator(testConf, testEvaluatorName, batchSize, false); +} + +TEST(Evaluator, detection_map) { + TestConfig config; + config.evaluatorConfig.set_type("detection_map"); + config.evaluatorConfig.set_overlap_threshold(0.5); + config.evaluatorConfig.set_background_id(0); + config.evaluatorConfig.set_ap_type("Integral"); + config.evaluatorConfig.set_evaluate_difficult(0); + + config.inputDefs.push_back({INPUT_DATA, "output", 7}); + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "label", 6}); + config.evaluatorConfig.set_evaluate_difficult(false); + testEvaluatorAll(config, "detection_map", 100); + + config.evaluatorConfig.set_evaluate_difficult(true); + testEvaluatorAll(config, "detection_map", 100); +} + +TEST(Evaluator, classification_error) { + TestConfig config; + config.evaluatorConfig.set_type("classification_error"); + config.evaluatorConfig.set_top_k(5); + + config.inputDefs.push_back({INPUT_DATA, "output", 50}); + config.inputDefs.push_back({INPUT_LABEL, "label", 50}); + testEvaluatorAll(config, "classification_error", 100); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + testEvaluatorAll(config, "classification_error_weight", 100); + + // multi binary labels + config.inputDefs.clear(); + config.inputDefs.push_back({INPUT_DATA, "output", 100}); + config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "label", 100}); + // Not support GPU + testEvaluator(config, "classification_error_multi_binary_label", 50, false); + + config.evaluatorConfig.set_classification_threshold(0.4); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + // Not support GPU + testEvaluator( + config, "classification_error_weight_multi_binary_label", 50, false); +} + +TEST(Evaluator, sum) { + TestConfig config; + config.evaluatorConfig.set_type("sum"); + + // sum of output + config.inputDefs.push_back({INPUT_DATA, "output", 10}); + testEvaluatorAll(config, "sum_output", 200); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + testEvaluatorAll(config, "sum_output_weight", 200); + + // sum of label + config.inputDefs.clear(); + config.inputDefs.push_back({INPUT_LABEL, "label", 10}); + testEvaluatorAll(config, "sum_label", 200); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + testEvaluatorAll(config, "sum_label_weight", 200); +} + +TEST(Evaluator, last_column_sum) { + TestConfig config; + config.evaluatorConfig.set_type("last-column-sum"); + + config.inputDefs.push_back({INPUT_DATA, "output", 50}); + testEvaluatorAll(config, "last-column-sum", 200); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + testEvaluatorAll(config, "last-column-sum_weight", 200); +} + +TEST(Evaluator, last_column_auc) { + TestConfig config; + config.evaluatorConfig.set_type("last-column-auc"); + + config.inputDefs.push_back({INPUT_DATA, "output", 2}); + config.inputDefs.push_back({INPUT_LABEL, "label", 2}); + testEvaluatorAll(config, "last-column-auc", 500); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + testEvaluatorAll(config, "last-column-auc_weight", 200); +} + +TEST(Evaluator, precision_recall) { + TestConfig config; + config.evaluatorConfig.set_type("precision_recall"); + + config.inputDefs.push_back({INPUT_DATA, "output", 10}); + config.inputDefs.push_back({INPUT_LABEL, "label", 10}); + testEvaluatorAll(config, "precision_recall", 200); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + testEvaluatorAll(config, "precision_recall_weight", 200); + + LOG(INFO) << "positive_label = 5"; + config.evaluatorConfig.set_positive_label(5); + testEvaluatorAll(config, "precision_recall_weight", 200); + + // multi binary labels + config.inputDefs.clear(); + config.evaluatorConfig.set_positive_label(-1); + config.inputDefs.push_back({INPUT_DATA, "output", 10}); + config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "label", 10}); + // Not support GPU + testEvaluator(config, "precision_recall_multi_binary_label", 100, false); + + LOG(INFO) << "classification_threshold = 0.4"; + config.evaluatorConfig.set_classification_threshold(0.4); + config.inputDefs.push_back({INPUT_DATA, "weight", 1}); + // Not support GPU + testEvaluator( + config, "precision_recall_weight_multi_binary_label", 100, false); +} + +TEST(Evaluator, ctc_error_evaluator) { + TestConfig config; + config.evaluatorConfig.set_type("ctc_edit_distance"); + + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "output", 32}); + config.inputDefs.push_back({INPUT_SEQUENCE_LABEL, "label", 1}); + testEvaluatorAll(config, "ctc_error_evaluator", 100); +} + +int main(int argc, char** argv) { + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/gserver/tests/test_Expand.cpp b/paddle/legacy/gserver/tests/test_Expand.cpp similarity index 100% rename from paddle/gserver/tests/test_Expand.cpp rename to paddle/legacy/gserver/tests/test_Expand.cpp diff --git a/paddle/legacy/gserver/tests/test_KmaxSeqScore.cpp b/paddle/legacy/gserver/tests/test_KmaxSeqScore.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e15b4e5038cddda00acdd06b7748984b03094e6e --- /dev/null +++ b/paddle/legacy/gserver/tests/test_KmaxSeqScore.cpp @@ -0,0 +1,164 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/utils/GlobalConstants.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +vector randSampling(int range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + return num; +} + +void genRandomSeqInfo(vector& seqStartPosition, + vector& subSeqStartPosition) { + const int maxSeqNum = 100; + // generate random start position information + int seqNum = 1 + (rand() % maxSeqNum); + seqStartPosition.resize(seqNum + 1, 0); + subSeqStartPosition.resize(1, 0); + + for (int i = 0; i < seqNum; ++i) { + int subSeqLen = 1 + (rand() % maxSeqNum); + for (int j = 0; j < subSeqLen; ++j) + subSeqStartPosition.push_back(subSeqStartPosition.back() + subSeqLen); + seqStartPosition[i + 1] = subSeqStartPosition.back(); + } +} + +void genRandomGroundTruth(real* values, + vector>& groundTruth, + vector& startPos, + size_t beamSize) { + groundTruth.resize(startPos.size() - 1, vector(beamSize, -1)); + for (size_t i = 0; i < startPos.size() - 1; ++i) { + int seqLen = startPos[i + 1] - startPos[i]; + vector pos = + randSampling(seqLen, min(static_cast(beamSize), seqLen)); + for (size_t j = 0; j < pos.size(); ++j) { + groundTruth[i][j] = pos[j]; + values[startPos[i] + pos[j]] = 1.; + } + } +} + +void checkLayerOut(vector> groundTruth, + real* layerOut, + size_t beamSize) { + for (size_t i = 0; i < groundTruth.size(); ++i) { + int begPos = i * beamSize; + vector tmp(layerOut + begPos, layerOut + begPos + beamSize); + sort(begin(tmp), end(tmp)); + sort(begin(groundTruth[i]), end(groundTruth[i])); + for (size_t j = 0; j < beamSize; ++j) CHECK_EQ(tmp[j], groundTruth[i][j]); + } +} + +TEST(Layer, kmaxSeqScoreLayer) { + const size_t maxBeamSize = 100; + size_t beamSize = 1 + (rand() % maxBeamSize); + + vector seqStartPosition; + vector subSeqStartPosition; + genRandomSeqInfo(seqStartPosition, subSeqStartPosition); + MatrixPtr inValue = + Matrix::create(subSeqStartPosition.back(), 1, false, false); + + std::vector mode = {false}; +#ifdef PADDLE_WITH_CUDA + mode.push_back(true); +#endif + + for (auto hasSubseq : {false, true}) { + vector> groundTruth; + inValue->randomizeUniform(); + genRandomGroundTruth(inValue->getData(), + groundTruth, + hasSubseq ? subSeqStartPosition : seqStartPosition, + beamSize); + + for (auto useGpu : mode) { + TestConfig config; + config.layerConfig.set_type("kmax_seq_score"); + config.layerConfig.set_beam_size(beamSize); + + if (hasSubseq) { + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "scores", + inValue, + seqStartPosition, + subSeqStartPosition}); + } else { + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "scores", inValue, seqStartPosition}); + } + config.layerConfig.add_inputs(); + + // data layer initialize + std::vector dataLayers; + LayerMap layerMap; + vector datas; + initDataLayer( + config, + &dataLayers, + &datas, + &layerMap, + "kmax_seq_score", + 100 /* actually this parameter is unused in self-defined input*/, + false, + useGpu); + // test layer initialize + std::vector parameters; + LayerPtr kmaxSeqScoreLayer; + FLAGS_use_gpu = useGpu; + initTestLayer(config, &layerMap, ¶meters, &kmaxSeqScoreLayer); + kmaxSeqScoreLayer->forward(PASS_TRAIN); + + const MatrixPtr outValue = kmaxSeqScoreLayer->getOutputValue(); + CHECK_EQ(outValue->getHeight(), + hasSubseq ? subSeqStartPosition.size() - 1 + : seqStartPosition.size() - 1); + CHECK_EQ(outValue->getWidth(), beamSize); + checkLayerOut(groundTruth, outValue->getData(), beamSize); + } + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand((size_t)(time(NULL))); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_LayerGrad.cpp b/paddle/legacy/gserver/tests/test_LayerGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..979cf8ee673291d66f8704f2deda6c7160f4b228 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_LayerGrad.cpp @@ -0,0 +1,2532 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +#include +#endif +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/math/MathUtils.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_double(checkgrad_eps); +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_bool(prev_batch_state); + +TEST(Operator, dot_mul) { + TestConfig config; + config.layerConfig.set_size(10); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs(); + operatorConf.set_type("dot_mul"); + operatorConf.set_dotmul_scale(-1); + + testOperatorGrad(config, operatorConf, 100, false, false); +} + +TEST(Projection, context) { + for (auto contextStart : {-5, -3, -1, 0, 3}) { + for (auto contextLength : {1, 2, 5, 7}) { + for (auto batchSize : {1, 2, 5, 20}) { + for (auto trainablePadding : {false, true}) { + LOG(INFO) << " contextStart=" << contextStart + << " contextLength=" << contextLength + << " batchSize=" << batchSize + << " trainablePadding=" << trainablePadding; + ProjectionConfig conf; + conf.set_type("context"); + conf.set_input_size(10); + conf.set_context_start(contextStart); + conf.set_context_length(contextLength); + conf.set_trainable_padding(trainablePadding); + conf.set_output_size(conf.context_length() * conf.input_size()); + int pad = + std::max(0, -conf.context_start()) + + std::max(0, conf.context_start() + conf.context_length() - 1); + for (auto useGpu : {false, true}) { + testProjectionGrad( + conf, + INPUT_SEQUENCE_DATA, + trainablePadding ? conf.input_size() * pad : 0, + batchSize, + useGpu, + contextStart + contextLength <= 1); // = testState + } + } + } + } + } +} + +TEST(Projection, trans_fc) { + ProjectionConfig conf; + conf.set_type("trans_fc"); + conf.set_input_size(50); + conf.set_output_size(20); + for (auto useGpu : {false, true}) { + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ 1000, + /* batchSize */ 100, + useGpu); + } +} + +TEST(Projection, fc) { + ProjectionConfig conf; + conf.set_type("fc"); + conf.set_input_size(10); + conf.set_output_size(20); + for (auto useGpu : {false, true}) { + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ 200, + /* batchSize */ 100, + useGpu); + } +} + +TEST(Projection, dot_mul) { + ProjectionConfig conf; + conf.set_type("dot_mul"); + conf.set_input_size(20); + conf.set_output_size(20); + for (auto useGpu : {false, true}) { + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ 20, + /* batchSize */ 100, + useGpu); + } +} + +TEST(Projection, table) { + ProjectionConfig conf; + conf.set_type("table"); + conf.set_input_size(10); + conf.set_output_size(20); + for (auto useGpu : {false, true}) { + testProjectionGrad(conf, + INPUT_LABEL, + /* parameterSize */ 200, + /* batchSize */ 100, + useGpu); + } +} + +TEST(Projection, identity) { + ProjectionConfig conf; + conf.set_type("identity"); + conf.set_input_size(10); + conf.set_output_size(10); + for (auto useGpu : {false, true}) { + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ 0, + /* batchSize */ 100, + useGpu); + } +} + +TEST(Projection, slice) { + ProjectionConfig conf; + conf.set_type("slice"); + conf.set_input_size(100); + SliceConfig& slice1 = *conf.add_slices(); + slice1.set_start(10); + slice1.set_end(20); + SliceConfig& slice2 = *conf.add_slices(); + slice2.set_start(50); + slice2.set_end(70); + conf.set_output_size(30); + for (auto useGpu : {false, true}) { + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ 0, + /* batchSize */ 10, + useGpu); + } +} + +TEST(Projection, scaling) { + ProjectionConfig conf; + conf.set_type("scaling"); + conf.set_input_size(10); + conf.set_output_size(10); + for (auto useGpu : {false}) { + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ 1, + /* batchSize */ 100, + useGpu); + } +} + +void testProjectionConv(size_t groups, bool isDeconv) { + const int NUM_FILTERS = 18; + const int FILTER_SIZE = 2; + const int FILTER_SIZE_Y = 2; + const int CHANNELS = 3; + const int IMAGE_SIZE = 16; + +#if CUDNN_VERSION >= 6000 + const int DILATION = 2; +#else + const int DILATION = 1; +#endif + + ProjectionConfig conf; + if (isDeconv) { + conf.set_type("convt"); + } else { + conf.set_type("conv"); + } + conf.set_num_filters(NUM_FILTERS); + + ConvConfig* conv = conf.mutable_conv_conf(); + conv->set_filter_size(FILTER_SIZE); + conv->set_filter_size_y(FILTER_SIZE_Y); + conv->set_channels(CHANNELS); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_dilation(DILATION); + conv->set_dilation_y(DILATION); + conv->set_groups(groups); + if (isDeconv) { + conv->set_filter_channels(NUM_FILTERS / conv->groups()); + } else { + conv->set_filter_channels(conv->channels() / conv->groups()); + } + conv->set_img_size(IMAGE_SIZE); + int output_x = outputSize(conv->img_size(), + (conv->filter_size() - 1) * DILATION + 1, + conv->padding(), + conv->stride(), + /* caffeMode */ true); + int output_y = outputSize(conv->img_size(), + (conv->filter_size_y() - 1) * DILATION + 1, + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true); + conv->set_output_x(output_x); + conv->set_output_y(output_y); + LOG(INFO) << "DILATION:" << DILATION << "; output_x: " << output_x + << "; output_y: " << output_y; + if (isDeconv) { + int deconv_image_x = imageSize(output_x, + (conv->filter_size() - 1) * DILATION + 1, + conv->padding(), + conv->stride(), + /* caffeMode */ true); + int deconv_image_y = imageSize(output_y, + (conv->filter_size_y() - 1) * DILATION + 1, + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true); + + LOG(INFO) << " deconv_image_x: " << deconv_image_x + << "; deconv_image_y: " << deconv_image_y; + conf.set_input_size(output_x * output_y * CHANNELS); + conf.set_output_size(deconv_image_x * deconv_image_y * NUM_FILTERS); + } else { + conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS); + conf.set_output_size(output_x * output_y * NUM_FILTERS); + } + + testProjectionGrad(conf, + INPUT_DATA, + /* parameterSize */ NUM_FILTERS * CHANNELS * FILTER_SIZE * + FILTER_SIZE_Y / groups, + /* batchSize */ 100, + true, + false, + NUM_FILTERS, + true); +} + +#ifdef PADDLE_WITH_CUDA +TEST(Projection, conv) { + /// test ConvProjection + testProjectionConv(1, false); + testProjectionConv(3, false); + /// test ConvTransProjection + testProjectionConv(1, true); + testProjectionConv(3, true); +} +#endif + +TEST(Layer, BilinearInterpLayer) { + TestConfig config; + config.layerConfig.set_type("bilinear_interp"); + config.biasSize = 0; + config.inputDefs.push_back({INPUT_DATA, "layer_0", 4096, 0}); + + LayerInputConfig* input = config.layerConfig.add_inputs(); + BilinearInterpConfig* bilinear = input->mutable_bilinear_interp_conf(); + ImageConfig* image = bilinear->mutable_image_conf(); + image->set_img_size(32); + image->set_img_size_y(32); + image->set_channels(4); + + for (auto useGpu : {false, true}) { + for (auto outSize : {32, 64}) { + bilinear->set_out_size_x(outSize); + bilinear->set_out_size_y(outSize); + testLayerGrad(config, "bilinear_interp", 10, false, useGpu); + } + } +} + +TEST(Layer, concat) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("concat"); + config.layerConfig.set_size(15); + config.layerConfig.set_active_type("sigmoid"); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 5, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "concat", 100, false, useGpu); + } +} + +TEST(Layer, AddtoLayer) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("addto"); + config.layerConfig.set_size(10); + config.layerConfig.set_active_type("sigmoid"); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "addto", 100, false, useGpu); + } +} + +TEST(Layer, CTCLayer) { + TestConfig config; + config.layerConfig.set_type("ctc"); + config.layerConfig.set_norm_by_times(false); + config.layerConfig.set_size(10); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_SEQUENCE_LABEL, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "ctc", + 100, + /* trans */ false, /* useGpu */ + useGpu); + } +} + +TEST(Layer, cosSimLayer) { + TestConfig config; + config.layerConfig.set_type("cos"); + config.layerConfig.set_size(1); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 50, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "cos", 100, false, useGpu); + } +} + +TEST(Layer, CosSimVecMatLayer) { + TestConfig config; + config.layerConfig.set_type("cos_vm"); + config.layerConfig.set_size(5); // output size + config.layerConfig.set_cos_scale(2.0); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 20, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 100, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "cos_vm", 100, false, useGpu); + } +} + +void testDepthwiseConvLayer(const string& type, bool useGpu) { + TestConfig config; + config.biasSize = 32; + config.layerConfig.set_type(type); + config.layerConfig.set_num_filters(32); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 2048, 192}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(2); + conv->set_filter_size_y(3); + conv->set_channels(16); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_groups(16); + conv->set_filter_channels(conv->channels() / conv->groups()); + conv->set_img_size(16); + conv->set_img_size_y(8); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + conv->set_output_y(outputSize(conv->img_size_y(), + conv->filter_size_y(), + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true)); + config.layerConfig.set_size(conv->output_x() * conv->output_y() * + config.layerConfig.num_filters()); + + testLayerGrad(config, "depthwise_conv", 100, false, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "depthwise_conv", 2, false, useGpu, true, 0.02); +} + +TEST(Layer, depthwiseConvLayer) { + // 'depthwise_conv' is a sepecial case of 'exconv' whose + // groups size equals to the input channels size. + testDepthwiseConvLayer("exconv", /* useGpu= */ false); +#ifdef PADDLE_WITH_CUDA + testDepthwiseConvLayer("exconv", /* useGpu= */ true); +#endif +} + +void testConvLayer(const string& type, bool trans, bool useGpu) { + TestConfig config; + config.biasSize = 16; + config.layerConfig.set_type(type); + config.layerConfig.set_num_filters(16); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + int dilation = 2; + if (type == "cudnn_conv") { +#if CUDNN_VERSION >= 6000 + dilation = 2; +#else + dilation = 1; +#endif + } + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 192}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(2); + conv->set_filter_size_y(2); + conv->set_channels(3); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_dilation(dilation); + conv->set_dilation_y(dilation); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + conv->set_img_size(16); + conv->set_img_size_y(16); + conv->set_output_x(outputSize(conv->img_size(), + (conv->filter_size() - 1) * dilation + 1, + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + conv->set_output_y(outputSize(conv->img_size_y(), + (conv->filter_size_y() - 1) * dilation + 1, + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true)); + config.layerConfig.set_size(conv->output_x() * conv->output_y() * + config.layerConfig.num_filters()); + + testLayerGrad(config, "conv", 100, trans, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "conv", 2, trans, useGpu, true, 0.02); +} + +TEST(Layer, convLayer) { + testConvLayer("exconv", /* trans= */ false, /* useGpu= */ false); +#ifdef PADDLE_WITH_CUDA + testConvLayer("exconv", /* trans= */ false, /* useGpu= */ true); + testConvLayer("cudnn_conv", /* trans= */ false, /* useGpu= */ true); +#endif +} + +void testConvTransLayer(const string& type, bool trans, bool useGpu) { + TestConfig config; + config.biasSize = 3; + config.layerConfig.set_type(type); + config.layerConfig.set_num_filters(3); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 384}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_filter_size(2); + conv->set_filter_size_y(4); + conv->set_channels(16); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_groups(1); + conv->set_filter_channels(3 / conv->groups()); + conv->set_img_size(16); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + + config.layerConfig.set_size(conv->img_size() * conv->img_size() * + config.layerConfig.num_filters()); + + testLayerGrad(config, "convTrans", 100, trans, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "convTrans", 2, trans, useGpu, true, 0.02); +} + +TEST(Layer, convTransLayer) { + for (auto useGpu : {false, true}) { + testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu); + } +#ifdef PADDLE_WITH_CUDA + testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true); +#endif +} + +TEST(Layer, blockExpandLayer) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("blockexpand"); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 6144, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + BlockExpandConfig* blockExpand = input->mutable_block_expand_conf(); + blockExpand->set_img_size_x(64); + blockExpand->set_img_size_y(32); + blockExpand->set_channels(3); + blockExpand->set_padding_x(0); + blockExpand->set_padding_y(0); + blockExpand->set_block_x(4); + blockExpand->set_block_y(32); + blockExpand->set_stride_x(2); + blockExpand->set_stride_y(2); + blockExpand->set_output_x(outputSize(blockExpand->img_size_x(), + blockExpand->block_x(), + blockExpand->padding_x(), + blockExpand->stride_x(), + /* caffeMode */ false)); + blockExpand->set_output_y(outputSize(blockExpand->img_size_y(), + blockExpand->block_y(), + blockExpand->padding_y(), + blockExpand->stride_y(), + /* caffeMode */ false)); + config.layerConfig.set_size(blockExpand->block_x() * blockExpand->block_y() * + blockExpand->channels()); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "blockexpand", 100, false, useGpu); + } +} + +TEST(Layer, maxoutLayer) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("maxout"); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 4096, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + MaxOutConfig* maxout = input->mutable_maxout_conf(); + ImageConfig* image = maxout->mutable_image_conf(); + + image->set_img_size(32); + image->set_img_size_y(32); + image->set_channels(4); + maxout->set_groups(2); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "maxout", 10, false, useGpu); + } +} + +void testFcLayer(string format, size_t nnz) { + TestConfig config; + config.biasSize = 1024; + config.layerConfig.set_type("fc"); + config.layerConfig.set_size(1024); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_drop_rate(0.1); + + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", 2048, nnz, ParaSparse(format)}); + config.layerConfig.add_inputs(); + + LOG(INFO) << config.inputDefs[0].sparse.sparse << " " + << config.inputDefs[0].sparse.format; + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "fc", + 100, + /* trans */ false, + useGpu, + /* weight */ true); + } +} + +TEST(Layer, fcLayer) { + testFcLayer("", 1024 * 1024 * 2); + testFcLayer("csc", 1024 * 10); + testFcLayer("csr", 1024 * 10); +} + +TEST(Layer, SelectiveFullyConnectedLayer) { + TestConfig config; + size_t nin = 16; + size_t nout = 256; + config.layerConfig.set_type("selective_fc"); + config.layerConfig.set_size(nout); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_has_selected_colums(true); + config.layerConfig.set_selective_fc_pass_generation(false); + config.biasSize = nout; + + config.inputDefs.push_back({INPUT_DATA, "input0", nin, nin * nout}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back( + {INPUT_SPARSE_NON_VALUE_DATA, "index", nout, 0, ParaSparse("csr", true)}); + config.layerConfig.add_inputs(); + + testLayerGrad(config, + "selective_fc", + 100, + /* trans= */ false, + /* useGup= */ false, + false); +#ifdef PADDLE_WITH_CUDA + testLayerGrad(config, + "selective_fc", + 100, + /* trans= */ false, + /* useGup= */ true, + false); +#endif +} + +TEST(Layer, DataNormLayer) { + TestConfig config; + config.layerConfig.set_type("data_norm"); + config.layerConfig.set_size(20); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 20, 100}); + config.inputDefs.back().isStatic = true; + config.layerConfig.add_inputs(); + + for (auto strategy : {"z-score", "min-max", "decimal-scaling"}) { + config.layerConfig.set_data_norm_strategy(strategy); + // The parameters are static, so not support GPU now + testLayerGrad(config, + "data_norm", + 200, + /* trans */ false, + /* useGpu */ false); + } +} + +TEST(Layer, hsigmoidLayer) { + TestConfig config; + config.layerConfig.set_type("hsigmoid"); + config.layerConfig.set_num_classes(5); + config.layerConfig.set_size(1); + config.biasSize = config.layerConfig.num_classes() - 1; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 200}); + config.inputDefs.push_back({INPUT_LABEL, "layer_1", 5, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "hsigmoid", + 100, + /* trans */ false, + /* useGpu */ useGpu); + } +} + +TEST(Layer, multi_cross) { + TestConfig config; + config.layerConfig.set_type("multi-class-cross-entropy"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_LABEL, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad( + config, "multi-class-cross-entropy", 100, /* trans */ false, useGpu); + } +} + +TEST(Layer, multi_binary_label_sparse_mat) { + TestConfig config; + config.layerConfig.set_type("multi_binary_label_cross_entropy"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "layer_1", 50, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "multi_binary_label_cross_entropy", + 100, + /* trans */ false, + useGpu); + } +} + +TEST(layer, multi_binary_label_id) { + TestConfig config; + config.layerConfig.set_type("multi_binary_label_cross_entropy"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_LABEL, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "multi_binary_label_cross_entropy", + 100, + /* trans */ false, + useGpu); + } +} + +TEST(Layer, multi_cross_with_selfnorm) { + TestConfig config; + config.layerConfig.set_type("multi_class_cross_entropy_with_selfnorm"); + config.layerConfig.set_softmax_selfnorm_alpha(0.1); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_LABEL, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + // Not support GPU now + testLayerGrad(config, + "multi_class_cross_entropy_with_selfnorm", + 100, + /* trans */ false, + /* useGpu */ false); +} + +TEST(Layer, multi_cross_soft) { + TestConfig config; + config.layerConfig.set_type("soft_binary_class_cross_entropy"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "soft_binary_class_cross_entropy", + 100, + /* trans */ false, + useGpu); + } +} + +TEST(Layer, square_error) { + TestConfig config; + config.layerConfig.set_type("square_error"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "square_error", 100, /* trans */ false, useGpu); + } +} + +TEST(Layer, sparse_square_error) { + TestConfig config; + config.layerConfig.set_type("square_error"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_SPARSE_NON_VALUE_DATA, "layer_1", 50, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + // "GpuSparseMatrix" as label is not supported + testLayerGrad(config, + "square_error", + 100, + /* trans */ false, + /* useGpu */ false); +} + +TEST(Layer, sparse_float_square_error) { + TestConfig config; + config.layerConfig.set_type("square_error"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 50, 0}); + config.inputDefs.push_back({INPUT_SPARSE_FLOAT_VALUE_DATA, "layer_1", 50, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + // "GpuSparseMatrix" as label is not supported + testLayerGrad(config, + "square_error", + 100, + /* trans */ false, + /* useGpu */ false); +} + +TEST(Layer, square_error_weighted) { + TestConfig config; + config.layerConfig.set_type("square_error"); + config.biasSize = 0; + config.testAccumulate = false; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_2", 1, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "square_error", 100, /* trans */ false, useGpu); + } +} + +TEST(Layer, huber_regression_loss) { + TestConfig config; + config.layerConfig.set_type("huber_regression"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + for (auto delta : {1, 3, 5}) { + config.layerConfig.set_delta(delta); + testLayerGrad(config, "huber_regression", 100, /* trans */ false, useGpu); + } + } +} + +TEST(Layer, huber_two_class) { + TestConfig config; + config.layerConfig.set_type("huber_classification"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.inputDefs.push_back({INPUT_LABEL, "layer_1", 2, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "huber_two_class", 100, /* trans */ false, useGpu); + } +} + +void testExpandLayer(string trans_type, bool hasSubseq) { + TestConfig config; + config.layerConfig.set_type("expand"); + + config.inputDefs.push_back( + {trans_type == "non-seq" ? INPUT_DENSE_DIM_DATA : INPUT_SEQUENCE_DATA, + "layer_0", + 10, + 0}); + config.inputDefs.push_back( + {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, + "layer_1", + 10, + 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.set_trans_type(trans_type); + LOG(INFO) << " trans_type=" << trans_type << " hasSubseq=" << hasSubseq; + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "expand", 30, false, useGpu); + } +} + +TEST(Layer, ExpandLayer) { + testExpandLayer("non-seq", false); // non-seq expand to seq + testExpandLayer("non-seq", true); // non-seq expand to hasSubseq + testExpandLayer("seq", true); // seq expand to hasSubseq +} + +void testDegradeLayer(bool hasSubseq, + string layer_type, + string trans_type, + int stride) { + TestConfig config; + config.layerConfig.set_type(layer_type); + config.layerConfig.set_size(10); + config.layerConfig.set_seq_pool_stride(stride); + config.biasSize = 0; + + config.inputDefs.push_back( + {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, + "layer_0", + 10, + 0}); + config.layerConfig.add_inputs(); + config.layerConfig.set_trans_type(trans_type); + + auto testDegradeLayerGrad = [](TestConfig& config, string layer_type) { + for (auto useGpu : {false, true}) { + testLayerGrad(config, layer_type, 100, false, useGpu); + } + }; + + if (layer_type == "average") { + for (auto strategy : {"average", "sum", "squarerootn"}) { + LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type + << " average_strategy=" << strategy + << " seq_pool_stride=" << stride; + config.layerConfig.set_average_strategy(strategy); + testDegradeLayerGrad(config, layer_type); + } + } else { + LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type + << " seq_pool_stride=" << stride; + testDegradeLayerGrad(config, layer_type); + } +} + +TEST(Layer, MaxLayer) { + testDegradeLayer(false, "max", "non-seq", -1); // seq max to non-seq + testDegradeLayer(false, + "max", + "non-seq", + 5); // seq max to a shorten seq, stride window = 5 + testDegradeLayer(true, "max", "non-seq", -1); // hasSubseq max to non-seq + testDegradeLayer(true, "max", "seq", -1); // hasSubseq max to seq +} + +TEST(Layer, SequenceLastInstanceLayer) { + testDegradeLayer(false, + "seqlastins", + "non-seq", + -1); // seq seqlastins to non-seq + testDegradeLayer(false, + "seqlastins", + "non-seq", + 5); // seq seqlastins to a shorten seq, stride window = 5 + testDegradeLayer(true, + "seqlastins", + "non-seq", + -1); // hasSubseq seqlastins to non-seq + testDegradeLayer(true, + "seqlastins", + "seq", + -1); // hasSubseq seqlastins to seq +} + +TEST(Layer, AverageLayer) { + testDegradeLayer(false, "average", "non-seq", -1); // seq average to non-seq + testDegradeLayer(false, + "average", + "non-seq", + 5); // seq average to a shorten seq, stride window = 5 + testDegradeLayer(true, + "average", + "non-seq", + -1); // hasSubseq average to non-seq + testDegradeLayer(true, "average", "seq", -1); // hasSubseq average to seq +} + +TEST(Layer, SequenceConcatLayer) { + TestConfig config; + config.layerConfig.set_type("seqconcat"); + config.layerConfig.set_size(10); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_0", 10, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "seqconcat", 100, false, useGpu); + } +} + +TEST(Layer, SequenceReshapeLayer) { + TestConfig config; + config.layerConfig.set_type("seqreshape"); + config.layerConfig.set_size(10); + + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "layer_0", 100, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "seqreshape", 100, false, useGpu); + } +} + +TEST(Layer, ConvShiftLayer) { + TestConfig config; + config.layerConfig.set_type("conv_shift"); + config.layerConfig.set_size(10); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 3, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + // Not support GPU now + testLayerGrad(config, "conv_shift", 100, false, false); +} + +TEST(Layer, PowerLayer) { + TestConfig config; + config.layerConfig.set_type("power"); + config.layerConfig.set_size(10); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "power", 100, false, useGpu); + } +} + +TEST(Layer, ConvexCombinationLayer) { + TestConfig config; + config.layerConfig.set_type("convex_comb"); + config.layerConfig.set_size(20); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 5, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 100, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "convex_comb", 100, false, useGpu); + } +} + +TEST(Layer, InterpolationLayer) { + TestConfig config; + config.layerConfig.set_type("interpolation"); + config.layerConfig.set_size(10); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_2", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "interpolation", 100, false, useGpu); + } +} + +TEST(Layer, DotProdLayer) { + TestConfig config; + config.layerConfig.set_type("dot_prod"); + config.layerConfig.set_size(1); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "dot_prod", 10, false, useGpu); + } +} + +TEST(Layer, OuterProdLayer) { + TestConfig config; + config.layerConfig.set_type("out_prod"); + config.layerConfig.set_size(100); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "out_prod", 100, false, useGpu); + } +} + +TEST(Layer, SlopeInterceptLayer) { + TestConfig config; + config.layerConfig.set_type("slope_intercept"); + config.layerConfig.set_size(10); + config.layerConfig.set_slope(1.0); + config.layerConfig.set_intercept(0.1); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "slope_intercept", 100, false, useGpu); + } +} + +TEST(Layer, ScalingLayer) { + TestConfig config; + config.layerConfig.set_type("scaling"); + config.layerConfig.set_size(10); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "scaling", 100, false, useGpu); + } +} + +void testNormLayer(const string& normType, bool trans, bool useGpu) { + TestConfig config; + config.layerConfig.set_type("norm"); + config.layerConfig.set_active_type("relu"); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1568, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + NormConfig* norm = input->mutable_norm_conf(); + norm->set_norm_type(normType); + norm->set_channels(16); + norm->set_size(5); + norm->set_scale(0.001); + norm->set_pow(0.75); + norm->set_blocked(0); + norm->set_img_size(14); + norm->set_img_size_y(7); + norm->set_output_x(norm->img_size()); + norm->set_output_y(norm->img_size_y()); + if (norm->norm_type() == "cmrnorm" || + norm->norm_type() == "cmrnorm-projection") { + norm->set_scale(norm->scale() / norm->size()); + } else { + norm->set_scale(norm->scale() / (norm->size() * norm->size())); + } + + config.layerConfig.set_size(norm->output_x() * norm->output_y() * + norm->channels()); + config.biasSize = 0; + + testLayerGrad(config, "norm", 100, trans, useGpu); +} + +TEST(Layer, NormLayer) { + testNormLayer("cmrnorm-projection", + /* trans= */ false, /* useGpu= */ + true); + testNormLayer("cmrnorm-projection", + /* trans= */ false, /* useGpu= */ + false); +} + +void setPoolConfig(TestConfig* config, + PoolConfig* pool, + const string& poolType) { + (*config).biasSize = 0; + (*config).layerConfig.set_type("pool"); + (*config).layerConfig.set_num_filters(16); + + int kw = 3, kh = 3; + int pw = 0, ph = 0; + int sw = 2, sh = 2; + pool->set_pool_type(poolType); + pool->set_channels(16); + pool->set_size_x(kw); + pool->set_size_y(kh); + pool->set_start(0); + pool->set_padding(pw); + pool->set_padding_y(ph); + pool->set_stride(sw); + pool->set_stride_y(sh); + + int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); + int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); + pool->set_output_x(ow); + pool->set_output_y(oh); +} + +void testPoolLayer(const string& poolType, + bool trans, + bool useGpu, + bool excludeMode = true) { + TestConfig config; + config.inputDefs.push_back({INPUT_DATA, "layer_0", 3136, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + PoolConfig* pool = input->mutable_pool_conf(); + + pool->set_img_size(14); + pool->set_img_size_y(14); + pool->set_exclude_mode(excludeMode); + setPoolConfig(&config, pool, poolType); + config.layerConfig.set_size(pool->output_x() * pool->output_y() * + pool->channels()); + + testLayerGrad(config, "pool", 100, trans, useGpu); +} + +#ifdef PADDLE_WITH_CUDA +void testPoolLayer2(const string& poolType, bool trans, bool useGpu) { + TestConfig config; + config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + PoolConfig* pool = input->mutable_pool_conf(); + + pool->set_size_y(4); + pool->set_stride_y(3); + pool->set_img_size(10); + pool->set_img_size_y(20); + setPoolConfig(&config, pool, poolType); + pool->set_output_y((pool->img_size_y() - pool->start() - pool->size_y()) / + ((float)pool->stride_y()) + + 1.5); + config.layerConfig.set_size(pool->output_x() * pool->output_y() * + pool->channels()); + + testLayerGrad(config, "pool", 100, trans, useGpu); +} +#endif + +TEST(Layer, PoolLayer) { + testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false); + testPoolLayer("avg-projection", + /* trans= */ false, + /* useGpu= */ false, + /* excludeMode= */ false); + testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false); + testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ false); + +#ifdef PADDLE_WITH_CUDA + testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true); + testPoolLayer("avg-projection", + /* trans= */ false, + /* useGpu= */ true, + /* excludeMode= */ false); + testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true); + testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); + testPoolLayer("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true); + testPoolLayer2("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); + testPoolLayer2("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true); + testPoolLayer2("cudnn-avg-incl-pad-pool", + /* trans= */ false, + /* useGpu= */ true); + testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ true); +#endif +} + +void setPool3DConfig(TestConfig* config, + PoolConfig* pool, + const string& poolType) { + // filter size + const int NUM_FILTERS = 16; + const int FILTER_SIZE = 3; + const int FILTER_SIZE_Y = 3; + const int FILTER_SIZE_Z = 3; + const int CHANNELS = 16; + + (*config).biasSize = 0; + (*config).layerConfig.set_type("pool3d"); + (*config).layerConfig.set_num_filters(NUM_FILTERS); + + int kw = FILTER_SIZE, kh = FILTER_SIZE_Y, kd = FILTER_SIZE_Z; + int pw = 0, ph = 0, pd = 0; + int sw = 2, sh = 2, sd = 2; + + pool->set_pool_type(poolType); + pool->set_pool_type("avg"); + pool->set_channels(CHANNELS); + pool->set_size_x(kw); + pool->set_size_y(kh); + pool->set_size_z(kd); + pool->set_padding(0); + pool->set_padding_y(0); + pool->set_padding_z(0); + pool->set_stride(sw); + pool->set_stride_y(sh); + pool->set_stride_z(sd); + pool->set_start(0); + int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); + int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); + int od = outputSize(pool->img_size_z(), kd, pd, sd, /* caffeMode */ false); + pool->set_output_x(ow); + pool->set_output_y(oh); + pool->set_output_z(od); +} + +void testPool3DLayer(const string& poolType, bool trans, bool useGpu) { + TestConfig config; + config.inputDefs.push_back({INPUT_DATA, "layer_0", 11664, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + PoolConfig* pool = input->mutable_pool_conf(); + + const int IMAGE_SIZE = 9; + const int IMAGE_SIZE_Y = 9; + const int IMAGE_SIZE_Z = 9; + + pool->set_img_size(IMAGE_SIZE); + pool->set_img_size_y(IMAGE_SIZE_Y); + pool->set_img_size_z(IMAGE_SIZE_Z); + + setPool3DConfig(&config, pool, poolType); + config.layerConfig.set_size(pool->output_x() * pool->output_y() * + pool->channels()); + + testLayerGrad(config, "pool3d", 100, trans, useGpu); +} + +TEST(Layer, Pool3DLayer) { + testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false); + testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false); +#ifdef PADDLE_WITH_CUDA + testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true); + testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true); +#endif +} + +void testSppLayer(const string& poolType, + const int pyramidHeight, + bool trans, + bool useGpu) { + TestConfig config; + config.layerConfig.set_type("spp"); + config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + SppConfig* sppConfig = input->mutable_spp_conf(); + sppConfig->set_pool_type(poolType); + sppConfig->set_pyramid_height(pyramidHeight); + ImageConfig* imageConfig = sppConfig->mutable_image_conf(); + imageConfig->set_channels(16); + imageConfig->set_img_size(10); + imageConfig->set_img_size_y(20); + int outputSize = (std::pow(4, sppConfig->pyramid_height()) - 1) / (4 - 1); + config.layerConfig.set_size(outputSize * imageConfig->channels()); + testLayerGrad(config, "spp", 100, trans, useGpu); +} + +TEST(Layer, SpatialPyramidPoolLayer) { + for (auto useGpu : {false, true}) { + for (auto pyramidHeight : {1, 2, 3}) { + testSppLayer("avg-projection", pyramidHeight, false, useGpu); + testSppLayer("max-projection", pyramidHeight, false, useGpu); + } + } +} + +TEST(Layer, rankCostLayer) { + TestConfig config; + config.layerConfig.set_type("rank-cost"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 1, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_2", 1, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "rank-cost", 100, false, useGpu); + } +} + +TEST(Layer, sumCostLayer) { + TestConfig config; + config.layerConfig.set_type("sum_cost"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "sum_cost", 100, false, useGpu); + } +} + +TEST(Layer, weightedRankCostLayer) { + TestConfig config; + config.layerConfig.set_type("rank-cost"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 1, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_2", 1, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_3", 1, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "weighted-rank-cost", 100, false, useGpu); + } +} + +TEST(Layer, TensorLayer) { + TestConfig config; + config.layerConfig.set_type("tensor"); + config.layerConfig.set_size(10); + config.layerConfig.set_active_type("sigmoid"); + config.biasSize = config.layerConfig.size(); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 5, 250}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", 5, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "tensor", 100, false, useGpu); + } +} + +TEST(Layer, RecurrentLayer) { + TestConfig config; + config.layerConfig.set_type("recurrent"); + config.layerConfig.set_size(4); + config.layerConfig.set_active_type("tanh"); + config.biasSize = 4; + + config.inputDefs.push_back( + {INPUT_SEQUENCE_DATA, "layer_0", /* dim= */ 4, /* paraSize= */ 16}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + for (auto reversed : {false, true}) { + config.layerConfig.set_reversed(reversed); + config.testState = !reversed; + testLayerGrad( + config, "recurrent", 50, /* trans= */ false, useGpu, false, 1.0); + } + } +} + +TEST(Layer, LstmLayer) { + TestConfig config; + config.layerConfig.set_type("lstmemory"); + config.layerConfig.set_size(4); + config.layerConfig.set_active_type("tanh"); + config.layerConfig.set_active_state_type("sigmoid"); + config.layerConfig.set_active_gate_type("sigmoid"); + config.biasSize = 28; + + config.inputDefs.push_back( + {INPUT_SEQUENCE_DATA, "layer_0", /* dim= */ 16, /* paraSize= */ 64}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + for (auto reversed : {false, true}) { + config.layerConfig.set_reversed(reversed); + config.testState = !reversed; + testLayerGrad( + config, "lstmemory", 100, /* trans= */ false, useGpu, false, 0.02); + } + } + for (auto useGpu : {true}) { + config.testBatchState = true; + config.layerConfig.set_reversed(false); + testLayerGrad(config, "lstmemory", 10, /* trans= */ false, useGpu); + } +} + +TEST(Layer, MDLstmLayer) { + TestConfig config; + config.layerConfig.set_type("mdlstmemory"); + config.layerConfig.set_size(4); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_active_state_type("sigmoid"); + config.layerConfig.set_active_gate_type("sigmoid"); + config.biasSize = 4 * 9; + + config.inputDefs.push_back( + {INPUT_SEQUENCE_MDIM_DATA, "layer_0", 4 * 5, 4 * 4 * 5}); + config.layerConfig.add_inputs(); + config.layerConfig.add_directions(true); + config.layerConfig.add_directions(true); + + for (auto useGpu : {false, true}) { + for (int i = 0; i < 2; i++) { + for (int j = 0; j < 2; j++) { + config.layerConfig.set_directions(0, bool(i)); + config.layerConfig.set_directions(1, bool(j)); + testLayerGrad(config, "mdlstmemory", 100, false, useGpu); + } + } + } +} + +TEST(Layer, ParameterReluLayer) { + auto testParameterReluLayer = [&](size_t inputSize, size_t channels) { + TestConfig config; + config.layerConfig.set_type("prelu"); + config.inputDefs.push_back({INPUT_DATA, "layer_0", inputSize, channels}); + config.layerConfig.add_inputs(); + config.layerConfig.set_size(inputSize); + config.layerConfig.set_partial_sum(inputSize / + channels); // size of feature map + for (auto useGpu : {false, true}) { + testLayerGrad(config, "prelu", 100, false, useGpu); + } + }; + + testParameterReluLayer(192, 1); + testParameterReluLayer(192, 3); + testParameterReluLayer(192, 192); +} + +TEST(Layer, ResizeLayer) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("resize"); + config.layerConfig.set_size(64); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 16, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "resize", 100, false, useGpu); + } +} + +TEST(Layer, RotateLayer) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("rotate"); + const int CHANNEL = 2; + const int HEIGHT = 8; + const int WIDTH = 4; + const int INPUT_SIZE = HEIGHT * WIDTH * CHANNEL; + config.layerConfig.set_size(INPUT_SIZE); + config.layerConfig.set_height(HEIGHT); + config.layerConfig.set_width(WIDTH); + config.inputDefs.push_back({INPUT_DATA, "layer_0", INPUT_SIZE, 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "rotate", 100, false, useGpu); + } +} + +TEST(Layer, NCELayer) { + TestConfig config; + size_t numClasses = 4; + config.layerConfig.set_type("nce"); + config.layerConfig.set_size(1); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_num_classes(numClasses); + config.biasSize = numClasses; + + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", /* dim= */ 16, /* paraSize= */ 16 * numClasses}); + config.inputDefs.push_back( + {INPUT_LABEL, "label", /* dim= */ numClasses, /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto withWeight : {false, true}) { + if (withWeight) { + config.inputDefs.push_back( + {INPUT_DATA_TARGET, "weight", /* dim= */ 1, /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + } + + for (auto isIdLabel : {false, true}) { + config.inputDefs[1] = { + isIdLabel ? INPUT_LABEL : INPUT_SPARSE_NON_VALUE_DATA, + "label", + /* dim= */ numClasses, + /* paraSize= */ 0}; + + for (auto withDist : {false, true}) { + config.layerConfig.clear_neg_sampling_dist(); + if (withDist) { + double sum = 0; + for (size_t i = 0; i < numClasses; ++i) { + real p = rand(); // NOLINT use rand_r + config.layerConfig.add_neg_sampling_dist(p); + sum += p; + } + for (size_t i = 0; i < numClasses; ++i) { + real p = config.layerConfig.neg_sampling_dist(i) / sum; + config.layerConfig.set_neg_sampling_dist(i, p); + } + } + LOG(INFO) << "NCELayer " + << " isIdLabel=" << isIdLabel << " withWeight=" << withWeight + << " withDist=" << withDist; + // Not support GPU now + testLayerGrad(config, + "nce", + 100, + /* trans= */ false, + /* useGpu */ false); + } + } + } +} + +TEST(Layer, GatedRecurrentLayer) { + TestConfig config; + config.layerConfig.set_type("gated_recurrent"); + config.layerConfig.set_size(4); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_active_gate_type("sigmoid"); + config.biasSize = 12; + + config.inputDefs.push_back( + {INPUT_SEQUENCE_DATA, "layer_0", /* dim= */ 12, /* paraSize= */ 48}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + for (auto reversed : {false, true}) { + config.layerConfig.set_reversed(reversed); + config.testState = !reversed; + testLayerGrad(config, "gated_recurrent", 100, /* trans= */ false, useGpu); + } + } +} + +TEST(Layer, GruStepLayer) { + TestConfig config; + config.layerConfig.set_type("gru_step"); + config.layerConfig.set_size(4); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_active_gate_type("sigmoid"); + config.biasSize = 12; + + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", /* dim= */ 12, /* paraSize= */ 48}); + config.inputDefs.push_back( + {INPUT_DATA, "layer_1", /* dim= */ 4, /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "gruStep", 100, /* trans= */ false, useGpu); + } +} + +TEST(Layer, LstmStepLayer) { + TestConfig config; + config.layerConfig.set_type("lstm_step"); + config.layerConfig.set_size(4); + config.layerConfig.set_active_type("sigmoid"); + config.layerConfig.set_active_state_type("sigmoid"); + config.layerConfig.set_active_gate_type("sigmoid"); + config.biasSize = 12; + config.testAccumulate = false; + + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", /* dim= */ 16, /* paraSize= */ 0}); + config.inputDefs.push_back( + {INPUT_DATA, "layer_1", /* dim= */ 4, /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "lstmStep", 100, /* trans= */ false, useGpu); + } +} + +void testBatchNormLayer(const string& type, bool trans, bool useGpu) { + TestConfig config; + const int CHANNELS = 10; + const int IMG_SIZE = 16; + const int IMG_SIZE_Y = 8; + size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y; + config.layerConfig.set_type(type); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type("sigmoid"); + config.biasSize = CHANNELS; + config.inputDefs.push_back({INPUT_DATA, + "layer_0", + /* dim= */ size, + /* paraSize= */ CHANNELS}); + + config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); + config.inputDefs.back().isStatic = true; + config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); + config.inputDefs.back().isStatic = true; + + LayerInputConfig* input = config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(CHANNELS); + img_conf->set_img_size(IMG_SIZE); + img_conf->set_img_size_y(IMG_SIZE_Y); + + testLayerGrad(config, + "batch_norm", + 64, + /* trans= */ trans, + useGpu, + /* useWeight */ true); +} + +TEST(Layer, BatchNormalizationLayer) { + testBatchNormLayer("batch_norm", false, false); +#ifdef PADDLE_WITH_CUDA + testBatchNormLayer("batch_norm", false, true); + if (hl_get_cudnn_lib_version() >= int(4000)) { + testBatchNormLayer("cudnn_batch_norm", false, true); + } +#endif +} + +void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) { + TestConfig config; + const int CHANNELS = 10; + const int IMG_SIZE = 16; + const int IMG_SIZE_Y = 8; + const int IMG_SIZE_Z = 8; + size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y * IMG_SIZE_Z; + config.layerConfig.set_type(type); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type("sigmoid"); + config.biasSize = CHANNELS; + config.inputDefs.push_back({INPUT_DATA, + "layer_0", + /* dim= */ size, + /* paraSize= */ CHANNELS}); + + config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); + config.inputDefs.back().isStatic = true; + config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); + config.inputDefs.back().isStatic = true; + + LayerInputConfig* input = config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(CHANNELS); + img_conf->set_img_size(IMG_SIZE); + img_conf->set_img_size_y(IMG_SIZE_Y); + img_conf->set_img_size_z(IMG_SIZE_Z); + + testLayerGrad(config, + "batch_norm", + 64, + /* trans= */ trans, + useGpu, + /* useWeight */ true); +} + +TEST(Layer, testBatchNorm3DLayer) { + testBatchNorm3DLayer("batch_norm", false, false); +#ifdef PADDLE_WITH_CUDA + testBatchNorm3DLayer("batch_norm", false, true); + if (hl_get_cudnn_lib_version() >= int(4000)) { + testBatchNorm3DLayer("cudnn_batch_norm", false, true); + } +#endif +} + +void testConvOperator(bool isDeconv) { + TestConfig config; + const int NUM_FILTERS = 16; + const int FILTER_SIZE = 2; + const int FILTER_SIZE_Y = 3; + const int CHANNELS = 3; + const int IMAGE_SIZE = 16; + const int IMAGE_SIZE_Y = 9; + OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs(); + if (isDeconv) { + operatorConf.set_type("convt"); + } else { + operatorConf.set_type("conv"); + } + ConvConfig* conv = operatorConf.mutable_conv_conf(); + operatorConf.set_num_filters(NUM_FILTERS); + conv->set_filter_size(FILTER_SIZE); + conv->set_filter_size_y(FILTER_SIZE_Y); + conv->set_channels(CHANNELS); + conv->set_padding(0); + conv->set_padding_y(1); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_groups(1); + conv->set_img_size(IMAGE_SIZE); + conv->set_img_size_y(IMAGE_SIZE_Y); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + conv->set_output_y(outputSize(conv->img_size_y(), + conv->filter_size_y(), + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true)); + + if (isDeconv) { + conv->set_filter_channels(NUM_FILTERS / conv->groups()); + config.inputDefs.push_back({INPUT_DATA, + "layer_0", + conv->output_x() * conv->output_y() * CHANNELS, + 0}); + config.layerConfig.set_size(IMAGE_SIZE * IMAGE_SIZE_Y * NUM_FILTERS); + } else { + conv->set_filter_channels(conv->channels() / conv->groups()); + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE_Y * CHANNELS, 0}); + config.layerConfig.set_size(conv->output_x() * conv->output_y() * + NUM_FILTERS); + } + + config.inputDefs.push_back( + {INPUT_DATA, + "layer_1", + FILTER_SIZE * FILTER_SIZE_Y * CHANNELS * NUM_FILTERS, + 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + testOperatorGrad(config, operatorConf, 100, /*useGpu*/ true, false); +} + +TEST(Operator, conv) { + testConvOperator(/*isDeconv*/ true); + testConvOperator(/*isDeconv*/ false); +} + +TEST(Layer, FeatureMapExpandLayer) { + TestConfig config; + config.layerConfig.set_type("featmap_expand"); + const int CHANNELS = 10; + const int INPUT_SIZE = 100; + config.layerConfig.set_size(INPUT_SIZE * CHANNELS); + config.layerConfig.set_num_filters(CHANNELS); + config.inputDefs.push_back({INPUT_SEQUENCE_DATA, + "layer_0", + /* dim= */ INPUT_SIZE, + /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + for (auto useGpu : {false, true}) { + for (auto asRowVec : {false, true}) { + config.layerConfig.set_user_arg(asRowVec ? "as_row_vec" : "as_col_vec"); + testLayerGrad(config, + "featmap_expand", + /*batch_size*/ 100, + /* trans= */ false, + useGpu, + /* useWeight */ true); + } + } +} + +TEST(Layer, MultiplexLayer) { + TestConfig config; + const int LAYER_SIZE = 100; + config.layerConfig.set_type("multiplex"); + config.layerConfig.set_size(LAYER_SIZE); + + config.inputDefs.push_back({INPUT_LABEL, "layer_0", 2, 0}); + config.inputDefs.push_back( + {INPUT_DATA, "layer_1", /* dim= */ LAYER_SIZE, /* paraSize= */ 0}); + config.inputDefs.push_back( + {INPUT_DATA, "layer_2", /* dim= */ LAYER_SIZE, /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "multiplex", 512, /* trans= */ false, useGpu); + } +} + +TEST(Layer, PadLayer) { + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("pad"); + + int c = 4; + int h = 31; + int w = 36; + size_t size = c * h * w; + config.inputDefs.push_back({INPUT_DATA, "layer_0", size, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + PadConfig* pad = input->mutable_pad_conf(); + ImageConfig* image = pad->mutable_image_conf(); + + image->set_channels(c); + image->set_img_size(h); + image->set_img_size_y(w); + pad->add_pad_c(1); + pad->add_pad_c(2); + pad->add_pad_h(2); + pad->add_pad_h(3); + pad->add_pad_w(3); + pad->add_pad_w(5); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "pad", 10, false, useGpu); + } +} + +TEST(Layer, CrossChannelNormLayer) { + TestConfig config; + config.paramInitialMean = 1.; + config.paramInitialStd = 0.; + config.layerConfig.set_type("norm"); + config.layerConfig.set_size(100); + LayerInputConfig* input = config.layerConfig.add_inputs(); + NormConfig* norm = input->mutable_norm_conf(); + norm->set_norm_type("cross-channel-norm"); + norm->set_channels(10); + norm->set_size(100); + norm->set_scale(0); + norm->set_pow(0); + norm->set_blocked(0); + config.inputDefs.push_back({INPUT_DATA, "layer_0", 100, 10}); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false); + } +} + +TEST(Layer, smooth_l1) { + TestConfig config; + config.layerConfig.set_type("smooth_l1"); + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 200, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 200, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "smooth_l1", 100, false, useGpu, false); + } +} + +TEST(Layer, multibox_loss) { + TestConfig config; + config.layerConfig.set_type("multibox_loss"); + config.biasSize = 0; + LayerInputConfig* input = config.layerConfig.add_inputs(); + MultiBoxLossConfig* multiboxLoss = input->mutable_multibox_loss_conf(); + multiboxLoss->set_num_classes(21); + multiboxLoss->set_input_num(1); + multiboxLoss->set_overlap_threshold(0.5); + multiboxLoss->set_neg_pos_ratio(3); + multiboxLoss->set_neg_overlap(0.5); + multiboxLoss->set_background_id(0); + multiboxLoss->set_height(3); + multiboxLoss->set_width(3); + + size_t gtNum = 1; + MatrixPtr labelValue = Matrix::create(gtNum, 6, false, false); + labelValue->randomizeUniform(); + labelValue->add(-0.5); + labelValue->sigmoid(*labelValue); + real* labelData = labelValue->getData(); + size_t labelWidth = labelValue->getWidth(); + for (size_t i = 0; i < gtNum; ++i) { + *(labelData + i * labelWidth) = std::rand() % 20 + 1; + *(labelData + i * labelWidth + 1) = 0.400259; + *(labelData + i * labelWidth + 2) = 0.377857; + *(labelData + i * labelWidth + 3) = 0.525712; + *(labelData + i * labelWidth + 4) = 0.519368; + } + vector seqStartPositions(gtNum + 1, 0); + for (size_t i = 1; i <= gtNum; ++i) { + seqStartPositions[i] = i; + } + + // Ensure at lease one matched bbox + MatrixPtr priorValue = Matrix::create(1, 72, false, false); + priorValue->randomizeUniform(); + priorValue->add(-0.5); + priorValue->sigmoid(*priorValue); + real* priorData = priorValue->getData(); + *(priorData) = 0.424811; + *(priorData + 1) = 0.397059; + *(priorData + 2) = 0.538905; + *(priorData + 3) = 0.447091; + *(priorData + 4) = 0.425720; + *(priorData + 5) = 0.515228; + *(priorData + 6) = 0.519452; + *(priorData + 7) = 0.591065; + + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "priorbox", priorValue, {}}); + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "label", labelValue, seqStartPositions}); + config.inputDefs.push_back({INPUT_DATA, "locPred", 36, 0}); + config.inputDefs.push_back({INPUT_DATA, "confPred", 189, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "multibox_loss", 1, false, useGpu, false); + } +} + +TEST(Layer, TransLayer) { + TestConfig config; + const int height = 128; + const int width = 256; + config.layerConfig.set_type("trans"); + config.layerConfig.set_size(width); + + config.inputDefs.push_back( + {INPUT_DATA, "layer_0", /* dim= */ height * width, /* paraSize= */ 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "trans", height, /* trans= */ false, useGpu); + } +} + +TEST(Layer, RowConvLayer) { + const int context = 3; + const int size = 512; + + TestConfig config; + config.layerConfig.set_type("row_conv"); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type("sigmoid"); + + config.inputDefs.push_back( + {INPUT_SEQUENCE_DATA, "layer_0", size, context * size}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + RowConvConfig* conv = input->mutable_row_conv_conf(); + conv->set_context_length(context); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "row_conv", 100, false, useGpu, false); + } +} + +TEST(Layer, CropLayer) { + TestConfig config; + // config input_0 + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ImageConfig* img = input->mutable_image_conf(); + img->set_channels(4); + img->set_img_size(16); + config.layerConfig.set_axis(2); + config.layerConfig.add_offset(0); + config.layerConfig.add_offset(0); + + // config input_1 + config.inputDefs.push_back({INPUT_DATA, "layer_1", 128, 0}); + input = config.layerConfig.add_inputs(); + img = input->mutable_image_conf(); + img->set_channels(2); + img->set_img_size(8); + + // config crop layer + config.layerConfig.set_type("crop"); + config.layerConfig.set_name("cropLayer"); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "crop", 100, false, useGpu, false); + } +} + +TEST(Layer, roi_pool) { + TestConfig config; + config.layerConfig.set_type("roi_pool"); + config.biasSize = 0; + LayerInputConfig* input = config.layerConfig.add_inputs(); + ROIPoolConfig* roiPoolConf = input->mutable_roi_pool_conf(); + roiPoolConf->set_pooled_width(7); + roiPoolConf->set_pooled_height(7); + roiPoolConf->set_spatial_scale(1. / 16); + roiPoolConf->set_width(14); + roiPoolConf->set_height(14); + + const size_t roiNum = 10; + const size_t roiDim = 10; + const size_t batchSize = 5; + MatrixPtr roiValue = Matrix::create(roiNum, roiDim, false, false); + roiValue->zeroMem(); + real* roiData = roiValue->getData(); + for (size_t i = 0; i < roiNum; ++i) { + roiData[i * roiDim + 0] = std::rand() % batchSize; + roiData[i * roiDim + 1] = std::rand() % 224; // xMin + roiData[i * roiDim + 2] = std::rand() % 224; // yMin + size_t xMin = static_cast(roiData[i * roiDim + 1]); + size_t yMin = static_cast(roiData[i * roiDim + 2]); + roiData[i * roiDim + 3] = xMin + std::rand() % (224 - xMin); // xMax + roiData[i * roiDim + 4] = yMin + std::rand() % (224 - yMin); // yMax + } + + config.inputDefs.push_back({INPUT_DATA, "input", 3 * 14 * 14, {}}); + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "rois", roiValue, {}}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "roi_pool", batchSize, false, useGpu, false); + } +} + +TEST(Layer, SwitchOrderLayer) { + TestConfig config; + // config input_0 + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ImageConfig* img = input->mutable_image_conf(); + img->set_channels(4); + img->set_img_size(16); + img->set_img_size_y(16); + + ReshapeConfig* reshape = config.layerConfig.mutable_reshape_conf(); + reshape->add_height_axis(0); + reshape->add_height_axis(1); + reshape->add_height_axis(2); + reshape->add_width_axis(3); + + // config softmax layer + config.layerConfig.set_type("switch_order"); + config.layerConfig.set_name("switchOrderLayer"); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "switch_order", 100, false, useGpu, true); + } +} + +vector randSampling(real range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0.); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + sort(begin(num), end(num)); + return num; +} + +TEST(Layer, SubNestedSequenceLayer) { + // layer size is not crutial for this layer, + // so use a small layer size in unittest + const int layerSize = 4; + + const int maxSeqNum = 50; + const int maxSeqLen = 50; + const int maxBeamSize = 32; + + srand((size_t)(time(NULL))); + int beamSize = 1 + (rand() % maxBeamSize); + + TestConfig config; + config.layerConfig.set_type("sub_nested_seq"); + config.layerConfig.set_name("sub_nested_seq_layer"); + config.layerConfig.set_size(layerSize); + + int seqNum = 1 + (rand() % maxSeqNum); + + // sequence information for the first input, it is a nested sequence + vector seqStartPos(seqNum + 1, 0); + vector subSeqStartPos(1, 0); + + // selected indices + MatrixPtr selectedIndices = Matrix::create(seqNum, beamSize, false, false); + selectedIndices->one(); + selectedIndices->mulScalar(-1.); + real* indicesData = selectedIndices->getData(); + + for (int i = 0; i < seqNum; ++i) { + int subSeqNum = 1 + (rand() % maxSeqNum); + for (int j = 0; j < subSeqNum; ++j) { + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % maxSeqLen))); + } + vector selSeqs = + randSampling(static_cast(subSeqNum), min(beamSize, subSeqNum)); + memcpy(indicesData + (i * beamSize), + selSeqs.data(), + selSeqs.size() * sizeof(real)); + seqStartPos[i + 1] = subSeqStartPos.back(); + } + + MatrixPtr seqInputPtr = + Matrix::create(seqStartPos.back(), layerSize, false, false); + seqInputPtr->randomizeUniform(); + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "nested_seq_input", + seqInputPtr, + seqStartPos, + subSeqStartPos}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "selected_indices", selectedIndices}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "sub_nested_seq", + /* batchSize */ seqNum, + /* trans */ false, + /* useGpu*/ useGpu, + /* useWeight */ false); + } +} + +TEST(Layer, ClipLayer) { + const size_t batchSize = 128; + const size_t size = 512; + TestConfig config; + config.layerConfig.set_type("clip"); + config.inputDefs.push_back({INPUT_DATA, "input", size, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ClipConfig* layerConf = input->mutable_clip_conf(); + double p1 = std::rand() / (double)RAND_MAX; + double p2 = std::rand() / (double)RAND_MAX; + layerConf->set_min(std::min(p1, p2)); + layerConf->set_max(std::max(p1, p2)); + for (auto useGpu : {false, true}) { + testLayerGrad(config, "clip", batchSize, false, useGpu, false); + } +} + +TEST(Layer, RowL2NormLayer) { + const size_t batchSize = 128; + const size_t size = 512; + TestConfig config; + config.layerConfig.set_type("row_l2_norm"); + config.layerConfig.set_size(size); + config.inputDefs.push_back({INPUT_DATA, "input", size, 0}); + config.layerConfig.add_inputs(); + for (auto useGpu : {false, true}) { + testLayerGrad(config, "row_l2_norm", batchSize, false, useGpu, false); + } +} + +void test3DConvLayer(const string& type, bool trans, bool useGpu) { + // filter size + const int NUM_FILTERS = 6; + // const int CHANNELS = 3; + const int FILTER_SIZE = 3; + const int FILTER_SIZE_Y = 3; + const int FILTER_SIZE_Z = 3; + + // input image + const int CHANNELS = 3; + const int IMAGE_SIZE = 9; + const int IMAGE_SIZE_Y = 9; + const int IMAGE_SIZE_Z = 9; + + TestConfig config; + config.biasSize = NUM_FILTERS; + config.layerConfig.set_type(type); + config.layerConfig.set_num_filters(NUM_FILTERS); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + // Setting up conv3D-trans layer + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + + conv->set_channels(CHANNELS); + conv->set_filter_size(FILTER_SIZE); + conv->set_filter_size_y(FILTER_SIZE_Y); + conv->set_filter_size_z(FILTER_SIZE_Z); + conv->set_padding(0); + conv->set_padding_y(0); + conv->set_padding_z(0); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_stride_z(2); + conv->set_img_size(IMAGE_SIZE); + conv->set_img_size_y(IMAGE_SIZE_Y); + conv->set_img_size_z(IMAGE_SIZE_Z); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + conv->set_output_y(outputSize(conv->img_size_y(), + conv->filter_size_y(), + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true)); + conv->set_output_z(outputSize(conv->img_size_z(), + conv->filter_size_z(), + conv->padding_z(), + conv->stride_z(), + /* caffeMode */ true)); + + config.layerConfig.set_size(conv->output_x() * conv->output_y() * + conv->output_z() * NUM_FILTERS); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + config.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z, + conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z * + NUM_FILTERS}); + + testLayerGrad(config, "conv3D", 10, trans, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "conv3D", 2, trans, useGpu, true, 0.02); +} + +TEST(Layer, test3DConvLayer) { + test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false); +#ifdef PADDLE_WITH_CUDA + test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true); +#endif +} + +void test3DDeConvLayer(const string& type, bool trans, bool useGpu) { + // filter size + const int NUM_FILTERS = 6; + // const int CHANNELS = 3; + const int FILTER_SIZE = 3; + const int FILTER_SIZE_Y = 3; + const int FILTER_SIZE_Z = 3; + + // input image + const int CHANNELS = 3; + const int IMAGE_SIZE = 4; + const int IMAGE_SIZE_Y = 6; + const int IMAGE_SIZE_Z = 6; + + // Setting up conv-trans layer + TestConfig config; + config.biasSize = NUM_FILTERS; + config.layerConfig.set_type("deconv3d"); + config.layerConfig.set_num_filters(NUM_FILTERS); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + + conv->set_channels(CHANNELS); + conv->set_filter_size(FILTER_SIZE); + conv->set_filter_size_y(FILTER_SIZE_Y); + conv->set_filter_size_z(FILTER_SIZE_Z); + conv->set_padding(0); + conv->set_padding_y(0); + conv->set_padding_z(0); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_stride_z(2); + conv->set_output_x(IMAGE_SIZE); + conv->set_output_y(IMAGE_SIZE_Y); + conv->set_output_z(IMAGE_SIZE_Z); + + conv->set_img_size(imageSize(conv->output_x(), + conv->filter_size(), + conv->padding(), + conv->stride(), + true)); + conv->set_img_size_y(imageSize(conv->output_y(), + conv->filter_size_y(), + conv->padding_y(), + conv->stride_y(), + true)); + conv->set_img_size_z(imageSize(conv->output_z(), + conv->filter_size_z(), + conv->padding_z(), + conv->stride_z(), + true)); + config.layerConfig.set_size(conv->img_size() * conv->img_size_y() * + conv->img_size_z() * NUM_FILTERS); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + config.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z, + conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z * + NUM_FILTERS}); + + testLayerGrad(config, "deconv3D", 10, trans, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "deconv3D", 2, trans, useGpu, true, 0.02); +} + +TEST(Layer, test3DDeConvLayer) { + test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false); +#ifdef PADDLE_WITH_CUDA + test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true); +#endif +} + +TEST(Layer, ScaleShiftLayer) { + // FIXME: Disable ScaleShiftLayer because it is not stable. + // https://github.com/PaddlePaddle/Paddle/issues/7781 + return; + // const size_t batchSize = 16; + // const size_t size = 32; + // TestConfig config; + // config.layerConfig.set_type("scale_shift"); + // config.layerConfig.set_size(size); + // config.biasSize = 1; + // config.inputDefs.push_back( + // {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1}); + // config.layerConfig.add_inputs(); + // for (auto useGpu : {false, true}) { + // testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false); + // } +} + +TEST(Layer, ScaleSubRegionLayer) { + const size_t batchSize = 64; + const size_t size = 4096; + TestConfig config; + config.layerConfig.set_type("scale_sub_region"); + config.inputDefs.push_back({INPUT_DATA, "input", size, 0}); + MatrixPtr indicesV = Matrix::create(batchSize, 6, false, false); + auto* data = indicesV->getData(); + for (size_t i = 0; i < batchSize; ++i) { + data[i * 2] = 2; + data[i * 2 + 1] = 4; + data[i * 2 + 2] = 16; + data[i * 2 + 3] = 32; + data[i * 2 + 4] = 16; + data[i * 2 + 5] = 32; + } + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "indices", indicesV, {}}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ScaleSubRegionConfig* scaleSubRegionConf = + input->mutable_scale_sub_region_conf(); + ImageConfig* imgConf = scaleSubRegionConf->mutable_image_conf(); + imgConf->set_img_size(32); + imgConf->set_img_size_y(32); + imgConf->set_channels(4); + scaleSubRegionConf->set_value(2.0); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "scale_sub_region", batchSize, false, useGpu, false); + } +} + +TEST(Layer, L2DistanceLayer) { + TestConfig config; + config.layerConfig.set_type("l2_distance"); + config.layerConfig.set_size(1); + config.biasSize = 0; + + const size_t input_dim = 27; + const size_t batch_size = 11; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", input_dim, 0}); + config.inputDefs.push_back({INPUT_DATA, "layer_1", input_dim, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "l2_distance", batch_size, false, useGpu); + } +} + +void testFactorizationMachineLayer(InputType type, bool useGpu) { + const int FACTOR_SIZE = 10; + TestConfig config; + config.layerConfig.set_type("factorization_machine"); + config.layerConfig.set_factor_size(FACTOR_SIZE); + config.layerConfig.set_size(1); + config.biasSize = 0; + config.inputDefs.push_back({type, "layer_0", 128, 1280}); + config.layerConfig.add_inputs(); + testLayerGrad(config, "factorization_machine", 16, false, useGpu, false); +} + +TEST(Layer, FactorizationMachineLayer) { + for (auto useGpu : {false, true}) { + testFactorizationMachineLayer(INPUT_DATA, useGpu); + } + testFactorizationMachineLayer(INPUT_SPARSE_FLOAT_VALUE_DATA, false); +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_LinearChainCRF.cpp b/paddle/legacy/gserver/tests/test_LinearChainCRF.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7082c1363a4cdadfd0e4a4497c20ae5c513bc7f1 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_LinearChainCRF.cpp @@ -0,0 +1,67 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/legacy/gserver/layers/LinearChainCRF.h" +#include "paddle/legacy/utils/Util.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static inline bool getNextSequence(vector& seq, int numClasses) { + for (auto& v : seq) { + if (++v < numClasses) { + return true; + } + v = 0; + } + return false; +} + +TEST(LinearChainCRF, decoding) { + const int numClasses = 4; + CpuVector para(numClasses * (numClasses + 2)); + real* a = para.getData(); + real* b = para.getData() + numClasses; + real* w = para.getData() + 2 * numClasses; + LinearChainCRF crf(4, para.getData()); + for (int length : {1, 2, 3, 10}) { + for (int tries = 0; tries < 10; ++tries) { + CpuMatrix x(length, numClasses); + x.randomizeUniform(); + para.randnorm(0, 2); + vector decodingResult(length); + vector bestResult(length); + vector testResult(length, 0); + crf.decode(x.getData(), &decodingResult[0], length); + real bestScore = -std::numeric_limits::max(); + do { + real score = a[testResult.front()] + b[testResult.back()]; + score += x.getElement(0, testResult.front()); + for (int k = 1; k < length; ++k) { + score += x.getElement(k, testResult[k]) + + w[numClasses * testResult[k - 1] + testResult[k]]; + } + if (score > bestScore) { + bestScore = score; + bestResult = testResult; + } + } while (getNextSequence(testResult, numClasses)); + for (int k = 0; k < length; ++k) { + EXPECT_EQ(decodingResult[k], bestResult[k]); + } + } + } +} diff --git a/paddle/legacy/gserver/tests/test_MKLDNN.cpp b/paddle/legacy/gserver/tests/test_MKLDNN.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c79ccd1956c5c68e5c97c2a185230b8ea9c3dea0 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_MKLDNN.cpp @@ -0,0 +1,448 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include "MKLDNNTester.h" +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/activations/MKLDNNActivation.h" +#include "paddle/legacy/math/MathUtils.h" + +using namespace paddle; // NOLINT + +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_bool(use_gpu); +DECLARE_bool(use_mkldnn); + +#define RUN_MKLDNN_TEST(DNN_CONFIG, REF_CONFIG, DESC) \ + MKLDNNTester tester; \ + for (auto bs : {DESC.bs, 1}) { \ + tester.run(DNN_CONFIG, REF_CONFIG, bs, DESC.ih, DESC.iw); \ + } + +#define RUN_MKLDNN_TEST_LAYER(DNN_CONFIG, REF_TYPE, DESC) \ + TestConfig ref = DNN_CONFIG; \ + ref.layerConfig.set_type(REF_TYPE); \ + RUN_MKLDNN_TEST(DNN_CONFIG, ref, DESC) + +struct testFcDesc { + int bs; + int ic; + int ih, iw; // oh == ow == 1 + int oc; +}; + +static void getMKLDNNFcConfig(TestConfig& cfg, const testFcDesc& pm) { + cfg.layerConfig.set_type("mkldnn_fc"); + cfg.layerConfig.set_active_type("relu"); + cfg.layerConfig.set_size(pm.oc); + cfg.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), + /* size of weight= */ size_t(pm.oc * pm.ic * pm.ih * pm.iw)}); + cfg.layerConfig.add_inputs(); +} + +void testFcLayer(const testFcDesc& pm) { + TestConfig dnnConfig; + getMKLDNNFcConfig(dnnConfig, pm); + for (auto biasSize : {pm.oc, 0}) { + dnnConfig.biasSize = biasSize; + RUN_MKLDNN_TEST_LAYER(dnnConfig, "fc", pm) + } +} + +TEST(MKLDNNLayer, FcLayer) { + /* bs, ic, ih, iw, oc */ + testFcLayer({2, 2, 1, 1, 3}); + testFcLayer({3, 7, 1, 1, 19}); + testFcLayer({8, 16, 13, 13, 32}); + testFcLayer({4, 12, 13, 13, 18}); + testFcLayer({2, 64, 16, 16, 32}); + testFcLayer({15, 3, 16, 16, 6}); +} + +struct testConvDesc { + int bs, gp; + int ic, ih, iw; + int oc, oh, ow; + int fh, fw; + int ph, pw; + int sh, sw; + int dh, dw; +}; + +static void getMKLDNNConvConfig(TestConfig& cfg, const testConvDesc& pm) { + cfg.layerConfig.set_type("mkldnn_conv"); + cfg.layerConfig.set_active_type("relu"); + cfg.layerConfig.set_num_filters(pm.oc); + cfg.layerConfig.set_size(pm.oc * pm.oh * pm.ow); + cfg.layerConfig.set_shared_biases(true); + cfg.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), + /* size of weight= */ size_t(pm.oc * pm.ic * pm.fh * pm.fw / pm.gp)}); + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + conv->set_groups(pm.gp); + conv->set_img_size(pm.iw); + conv->set_img_size_y(pm.ih); + conv->set_output_x(pm.ow); + conv->set_output_y(pm.oh); + conv->set_filter_size(pm.fw); + conv->set_filter_size_y(pm.fh); + conv->set_channels(pm.ic); + conv->set_padding(pm.pw); + conv->set_padding_y(pm.ph); + conv->set_stride(pm.sw); + conv->set_stride_y(pm.sh); + conv->set_dilation(pm.dw); + conv->set_dilation_y(pm.dh); + conv->set_caffe_mode(true); + conv->set_filter_channels(conv->channels() / conv->groups()); + CHECK_EQ(conv->filter_channels() * pm.gp, conv->channels()) + << "it is indivisible"; + + int fh = (pm.fh - 1) * pm.dh + 1; + int fw = (pm.fw - 1) * pm.dw + 1; + int ow = outputSize(pm.iw, fw, pm.pw, pm.sw, true); + int oh = outputSize(pm.ih, fh, pm.ph, pm.sh, true); + CHECK_EQ(ow, pm.ow) << "output size check failed"; + CHECK_EQ(oh, pm.oh) << "output size check failed"; +} + +void testConvLayer(const testConvDesc& pm) { + TestConfig dnnConfig; + getMKLDNNConvConfig(dnnConfig, pm); + for (auto biasSize : {pm.oc, 0}) { + dnnConfig.biasSize = biasSize; + RUN_MKLDNN_TEST_LAYER(dnnConfig, "exconv", pm) + } +} + +TEST(MKLDNNLayer, ConvLayer) { + /* bs, gp, ic, ih, iw, oc, oh, ow, fh, fw, ph, pw, sh, sw, dh, dw */ + testConvLayer({2, 1, 3, 32, 32, 16, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1}); + testConvLayer({2, 1, 8, 16, 16, 8, 16, 16, 3, 3, 1, 1, 1, 1, 1, 1}); + testConvLayer({3, 1, 16, 32, 32, 3, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1}); + testConvLayer({8, 1, 16, 18, 18, 32, 18, 18, 3, 3, 1, 1, 1, 1, 1, 1}); + testConvLayer({16, 1, 1, 42, 31, 32, 23, 11, 4, 5, 3, 2, 2, 3, 1, 1}); + testConvLayer({2, 1, 8, 16, 16, 8, 8, 8, 3, 3, 1, 1, 2, 2, 1, 1}); + testConvLayer({3, 1, 8, 13, 13, 8, 7, 7, 3, 3, 1, 1, 2, 2, 1, 1}); + // with groups + testConvLayer({2, 2, 4, 5, 5, 8, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1}); + testConvLayer({2, 3, 3, 5, 5, 3, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1}); + testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1}); +} + +struct testPoolDesc { + int bs, ic; // input channel and output channel are the same + int ih, iw; + int oh, ow; + int fh, fw; + int ph, pw; + int sh, sw; +}; + +static void getMKLDNNPoolConfig(TestConfig& cfg, const testPoolDesc& pm) { + cfg.layerConfig.set_type("mkldnn_pool"); + cfg.layerConfig.set_active_type("relu"); + cfg.layerConfig.set_size(pm.ic * pm.oh * pm.ow); + cfg.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), + 0}); + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + PoolConfig* pool = input->mutable_pool_conf(); + pool->set_pool_type("avg-projection"); + pool->set_channels(pm.ic); + pool->set_img_size(pm.iw); + pool->set_img_size_y(pm.ih); + pool->set_output_x(pm.ow); + pool->set_output_y(pm.oh); + pool->set_size_x(pm.fw); + pool->set_size_y(pm.fh); + pool->set_padding(pm.pw); + pool->set_padding_y(pm.ph); + pool->set_stride(pm.sw); + pool->set_stride_y(pm.sh); + + int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false); + int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false); + CHECK_EQ(ow, pm.ow) << "output size check failed"; + CHECK_EQ(oh, pm.oh) << "output size check failed"; +} + +void testPoolLayer(const testPoolDesc& pm) { + TestConfig dnnConfig; + getMKLDNNPoolConfig(dnnConfig, pm); + LayerInputConfig* input = dnnConfig.layerConfig.mutable_inputs(0); + PoolConfig* pool = input->mutable_pool_conf(); + for (auto type : {"max-projection", "avg-projection"}) { + pool->set_pool_type(type); + RUN_MKLDNN_TEST_LAYER(dnnConfig, "pool", pm) + } +} + +TEST(MKLDNNLayer, PoolLayer) { + /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw */ + testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2}); + testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2}); + testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2}); + testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2}); + testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2}); + testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1}); + testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1}); + testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2}); +} + +struct testBatchNormDesc { + int bs; + int ic; + int ih, iw; +}; + +static void getMKLDNNBatchNormConfig(TestConfig& cfg, + const testBatchNormDesc& pm) { + cfg.layerConfig.set_size(pm.ic * pm.ih * pm.iw); + cfg.layerConfig.set_type("mkldnn_batch_norm"); + cfg.biasSize = pm.ic; + cfg.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), + /* size of weight= */ size_t(pm.ic)}); + cfg.inputDefs.push_back( + {INPUT_DATA, "layer_1_moving_mean", 1, size_t(pm.ic)}); + cfg.inputDefs.back().isStatic = true; + cfg.inputDefs.push_back({INPUT_DATA, "layer_2_moving_var", 1, size_t(pm.ic)}); + cfg.inputDefs.back().isStatic = true; + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + cfg.layerConfig.set_active_type("relu"); + cfg.layerConfig.add_inputs(); + cfg.layerConfig.add_inputs(); + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(pm.ic); + img_conf->set_img_size_y(pm.ih); + img_conf->set_img_size(pm.iw); +} + +void testBatchNormLayer(const testBatchNormDesc& pm) { + TestConfig dnnConfig; + getMKLDNNBatchNormConfig(dnnConfig, pm); + TestConfig refConfig = dnnConfig; + refConfig.layerConfig.set_type("batch_norm"); + // for PASS_TRAIN, use_global_stats always should be false, and batchsize != 1 + VLOG(MKLDNN_TESTS) << "check train phase"; + dnnConfig.layerConfig.set_use_global_stats(false); + refConfig.layerConfig.set_use_global_stats(false); + MKLDNNTester tester; + tester.run(dnnConfig, refConfig, pm.bs, pm.ih, pm.iw, PASS_TRAIN); + // for PASS_TEST, check use_global_stats true and false, and batchsize 1 + VLOG(MKLDNN_TESTS) << "check test phase"; + for (auto useGS : {false, true}) { + dnnConfig.layerConfig.set_use_global_stats(useGS); + refConfig.layerConfig.set_use_global_stats(useGS); + MKLDNNTester tester; + for (auto bs : {pm.bs, 1}) { + tester.run(dnnConfig, refConfig, bs, pm.ih, pm.iw, PASS_TEST); + } + } +} + +TEST(MKLDNNLayer, BatchNormLayer) { + testBatchNormLayer({4, 10, 6, 6}); + testBatchNormLayer({16, 32, 16, 16}); + testBatchNormLayer({4, 16, 8, 10}); +} + +struct testLRNDesc { + int bs, ic, ih, iw; + float scale, pow; + int localSize; +}; + +void getMKLDNNLRNConfig(TestConfig& cfg, const testLRNDesc& pm) { + cfg.layerConfig.set_type("mkldnn_lrn"); + cfg.layerConfig.set_active_type("relu"); + size_t layerSize = pm.ic * pm.ih * pm.iw; + cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0}); + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + NormConfig* norm = input->mutable_norm_conf(); + norm->set_channels(pm.ic); + norm->set_size(pm.localSize); + norm->set_scale(pm.scale); + norm->set_pow(pm.pow); + norm->set_blocked(0); + norm->set_img_size(pm.iw); + norm->set_img_size_y(pm.ih); + norm->set_output_x(norm->img_size()); + norm->set_output_y(norm->img_size_y()); + cfg.layerConfig.set_size(layerSize); + cfg.biasSize = 0; +} + +void testLRNLayer(const testLRNDesc& pm) { + TestConfig dnnConfig; + getMKLDNNLRNConfig(dnnConfig, pm); + // mkldnn_lrn <==> norm with cmrnorm-projection type + TestConfig refConfig = dnnConfig; + refConfig.layerConfig.set_type("norm"); + LayerInputConfig* input = refConfig.layerConfig.mutable_inputs(0); + NormConfig* norm = input->mutable_norm_conf(); + norm->set_norm_type("cmrnorm-projection"); + norm->set_scale(norm->scale() / norm->size()); + RUN_MKLDNN_TEST(dnnConfig, refConfig, pm) +} + +TEST(MKLDNNLayer, LRNLayer) { + testLRNLayer({4, 10, 12, 12, 0.001f, 0.75f, 5}); + testLRNLayer({2, 32, 6, 6, 0.001f, 0.75f, 5}); + testLRNLayer({4, 16, 8, 10, 0.01f, 0.5f, 5}); +} + +struct testImageDesc { + int bs, ic, ih, iw; +}; + +static void getAddtoConfig(TestConfig& cfg, + const testImageDesc& pm, + const size_t nInputs = 1) { + cfg.biasSize = 0; + cfg.layerConfig.set_type("addto"); + size_t layerSize = pm.ic * pm.ih * pm.iw; + cfg.layerConfig.set_size(layerSize); + cfg.layerConfig.set_active_type("relu"); + for (size_t i = 0; i < nInputs; ++i) { + std::stringstream ss; + ss << "layer_" << i; + cfg.inputDefs.push_back({INPUT_DATA, ss.str(), layerSize, 0}); + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(pm.ic); + img_conf->set_img_size_y(pm.ih); + img_conf->set_img_size(pm.iw); + } +} + +void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) { + CHECK_GE(nInputs, 1UL); + TestConfig dnnConfig; + getAddtoConfig(dnnConfig, pm, nInputs); + dnnConfig.layerConfig.set_type("mkldnn_addto"); + for (auto withBias : {false, true}) { + dnnConfig.biasSize = withBias ? pm.ic * pm.ih * pm.iw : 0; + RUN_MKLDNN_TEST_LAYER(dnnConfig, "addto", pm) + } +} + +TEST(MKLDNNLayer, AddtoLayer) { + testAddtoLayer({16, 5, 14, 14}, 1); + testAddtoLayer({8, 10, 8, 8}, 2); + testAddtoLayer({4, 12, 1, 1}, 3); +} + +static void getMKLDNNConcatConfig(TestConfig& cfg, + const std::vector& inputs) { + CHECK_GE(inputs.size(), 2UL) << "at least two inputs"; + int oc = inputs[0].ic; + for (size_t i = 1; i < inputs.size(); ++i) { + CHECK_EQ(inputs[i].bs, inputs[0].bs); + CHECK_EQ(inputs[i].ih, inputs[0].ih); + CHECK_EQ(inputs[i].iw, inputs[0].iw); + oc += inputs[i].ic; + } + cfg.biasSize = 0; + cfg.layerConfig.set_type("mkldnn_concat"); + cfg.layerConfig.set_size(oc * inputs[0].ih * inputs[0].iw); + cfg.layerConfig.set_active_type("relu"); + for (size_t i = 0; i < inputs.size(); ++i) { + std::stringstream ss; + ss << "layer_" << i; + cfg.inputDefs.push_back( + {INPUT_DATA, + ss.str(), + (size_t)(inputs[i].ic) * inputs[i].ih * inputs[i].iw, + 0}); + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(inputs[i].ic); + img_conf->set_img_size_y(inputs[i].ih); + img_conf->set_img_size(inputs[i].iw); + } +} + +void testConcatLayer(const std::vector& inputs) { + TestConfig dnnConfig; + getMKLDNNConcatConfig(dnnConfig, inputs); + RUN_MKLDNN_TEST_LAYER(dnnConfig, "concat", inputs[0]) +} + +TEST(MKLDNNLayer, ConcatLayer) { + testConcatLayer({{64, 128, 1, 1}, {64, 32, 1, 1}, {64, 64, 1, 1}}); + testConcatLayer({{32, 100, 8, 8}, {32, 10, 8, 8}}); +} + +void testActivation(std::string actType, const testImageDesc& pm) { + // TODO(TJ): remove me when paddle support elu activation + if (actType == "mkldnn_elu") { + return; + } + const std::string compareTypes[] = {actType, actType.erase(0, 7)}; + TestConfig cfg; + getAddtoConfig(cfg, pm); + TestConfig ref = cfg; + cfg.layerConfig.set_active_type(compareTypes[0]); + ref.layerConfig.set_active_type(compareTypes[1]); + RUN_MKLDNN_TEST(cfg, ref, pm) +} + +TEST(MKLDNNActivation, Activations) { + auto types = MKLDNNActivation::getAllRegisteredTypes(); + for (auto type : types) { + /* bs, c, h, w*/ + testActivation(type, {16, 64, 32, 32}); + testActivation(type, {2, 8, 1, 1}); + } +} + +DECLARE_string(config_args); +TEST(MKLDNNNet, net) { + std::vector cases = {"simple", "branch"}; + for (auto name : cases) { + std::string config = "./legacy/gserver/tests/mkldnn_" + name + "_net.conf"; + for (auto channels : {2, 32}) { + std::ostringstream oss; + oss << "channels=" << channels; + FLAGS_config_args = oss.str(); + MKLDNNTester::runNetTest(config); + } + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + FLAGS_use_gpu = false; + FLAGS_use_mkldnn = true; + initMain(argc, argv); + initPython(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_MaxPoolingWithMaskOutput.cpp b/paddle/legacy/gserver/tests/test_MaxPoolingWithMaskOutput.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2bc261b4a87ce7f1f4ce1c936ee4151d75e17f3f --- /dev/null +++ b/paddle/legacy/gserver/tests/test_MaxPoolingWithMaskOutput.cpp @@ -0,0 +1,117 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include + +#include "LayerGradUtil.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; + +void setPoolConfig(TestConfig* config, + PoolConfig* pool, + const string& poolType) { + (*config).biasSize = 0; + (*config).layerConfig.set_type("pool"); + (*config).layerConfig.set_num_filters(1); + + int kw = 3, kh = 3; + int pw = 0, ph = 0; + int sw = 2, sh = 2; + pool->set_pool_type(poolType); + pool->set_channels(1); + pool->set_size_x(kw); + pool->set_size_y(kh); + pool->set_start(0); + pool->set_padding(pw); + pool->set_padding_y(ph); + pool->set_stride(sw); + pool->set_stride_y(sh); + + int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); + int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); + pool->set_output_x(ow); + pool->set_output_y(oh); +} + +void doOneMaxPoolingWithMaskOutputTest(MatrixPtr& inputMat, + const string& poolType, + bool use_gpu, + MatrixPtr& maskMat) { + TestConfig config; + config.inputDefs.push_back({INPUT_DATA, "layer_0", 25, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + PoolConfig* pool = input->mutable_pool_conf(); + + pool->set_img_size(5); + pool->set_img_size_y(5); + setPoolConfig(&config, pool, poolType); + config.layerConfig.set_size(pool->output_x() * pool->output_y() * + pool->channels()); + + config.layerConfig.set_name("MaxPoolWithMask"); + + std::vector dataLayers; + LayerMap layerMap; + vector datas; + + initDataLayer(config, + &dataLayers, + &datas, + &layerMap, + "MaxPoolWithMask", + 1, + false, + use_gpu); + + dataLayers[0]->getOutputValue()->copyFrom(*inputMat); + + FLAGS_use_gpu = use_gpu; + std::vector parameters; + LayerPtr maxPoolingWithMaskOutputLayer; + initTestLayer(config, &layerMap, ¶meters, &maxPoolingWithMaskOutputLayer); + maxPoolingWithMaskOutputLayer->forward(PASS_GC); + + checkMatrixEqual(maxPoolingWithMaskOutputLayer->getOutput("mask").value, + maskMat); +} + +TEST(Layer, maxPoolingWithMaskOutputLayerFwd) { + bool useGpu = false; + MatrixPtr inputMat; + MatrixPtr maskMat; + real inputData[] = {0.1, 0.1, 0.5, 0.5, 1.1, 0.2, 0.2, 0.6, 0.1, + 0.1, 0.3, 0.3, 0.7, 0.1, 0.1, 0.4, 0.4, 0.8, + 0.8, 0.1, 1.0, 2.0, 3.0, 0.0, 9.0}; + real maskData[] = {12, 4, 22, 24}; + + inputMat = Matrix::create(1, 25, false, useGpu); + maskMat = Matrix::create(1, 4, false, useGpu); + inputMat->setData(inputData); + maskMat->setData(maskData); + doOneMaxPoolingWithMaskOutputTest( + inputMat, "max-pool-with-mask", useGpu, maskMat); +#ifdef PADDLE_WITH_CUDA + useGpu = true; + inputMat = Matrix::create(1, 25, false, useGpu); + maskMat = Matrix::create(1, 4, false, useGpu); + inputMat->copyFrom(inputData, 25); + maskMat->copyFrom(maskData, 4); + doOneMaxPoolingWithMaskOutputTest( + inputMat, "max-pool-with-mask", useGpu, maskMat); +#endif +} diff --git a/paddle/legacy/gserver/tests/test_MultinomialSampler.cpp b/paddle/legacy/gserver/tests/test_MultinomialSampler.cpp new file mode 100644 index 0000000000000000000000000000000000000000..25b1a1191d0100c8ee625d3f5f36d1513164b23b --- /dev/null +++ b/paddle/legacy/gserver/tests/test_MultinomialSampler.cpp @@ -0,0 +1,147 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include +#include + +#undef PADDLE_DISABLE_TIMER +#include "paddle/legacy/utils/Stat.h" + +#include "paddle/legacy/gserver/layers/MultinomialSampler.h" +#include "paddle/legacy/utils/Util.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +class MultinomialSamplerTester : public MultinomialSampler { + public: + MultinomialSamplerTester(real* prob, int size) + : MultinomialSampler(prob, size) {} + + template + int testGen(Rand1 rand1) { + return gen1(rand1); + } +}; + +TEST(MultinomialSampler, gen) { + int numGrids = 1024 * 1024; + int size = 1024 * 4; + default_random_engine reng; + + for (size_t iter = 0; iter < 256; ++iter) { + uniform_int_distribution rand(1, numGrids / size * 1.8); + vector prob; + int sum = 0; + for (int i = 0; i < size; ++i) { + prob.push_back(rand(reng)); + sum += prob.back(); + } + + CHECK_LE(sum, numGrids); + prob.back() += numGrids - sum; + + vector counts(size); + MultinomialSamplerTester sampler(&prob[0], size); + counts.assign(size, 0); + { + double s = (double)size / (double)numGrids; + REGISTER_TIMER("MultinomialSampler"); + for (double i = 0; i < numGrids; ++i) { + int ret = sampler.testGen([i, s]() { return s * i; }); + if (ret < 0 || ret >= size) { + EXPECT_GE(ret, 0); + EXPECT_LT(ret, size); + break; + } + ++counts[ret]; + } + } + for (int i = 0; i < size; ++i) { + if (prob[i] != counts[i]) { + EXPECT_EQ(prob[i], counts[i]); + LOG(INFO) << iter; + break; + } + } + } +} + +void benchmarkRandom() { + int n = 1024 * 1024; + + int sum; + double sum1; + + sum = 0; + unsigned int seed = 1; + { + REGISTER_TIMER("crand"); + for (int i = 0; i < n; ++i) { + sum += rand_r(&seed) % 1000; + } + } + LOG(INFO) << "sum=" << sum; + + default_random_engine reng; + uniform_int_distribution rand(1, 1000); + sum = 0; + { + REGISTER_TIMER("stdrand"); + for (int i = 0; i < n; ++i) { + sum += rand(reng); + } + } + LOG(INFO) << "sum=" << sum; + + sum = 0; + { + REGISTER_TIMER("default_random_engine"); + for (int i = 0; i < n; ++i) { + sum += reng(); + } + } + LOG(INFO) << "sum=" << sum; + + uniform_real_distribution rand1(0, 1); + sum1 = 0; + { + REGISTER_TIMER("stdrand1"); + for (int i = 0; i < n; ++i) { + sum1 += rand1(reng); + } + } + LOG(INFO) << "sum1=" << sum1; + + sum1 = 0; + { + real a = 1.0f / (real)RAND_MAX; + REGISTER_TIMER("crand1"); + for (int i = 0; i < n; ++i) { + sum1 += a * rand_r(&seed); + } + } + LOG(INFO) << "sum1=" << sum1; +} + +int main(int argc, char** argv) { + initMain(argc, argv); + testing::InitGoogleTest(&argc, argv); + benchmarkRandom(); + int ret = RUN_ALL_TESTS(); + globalStat.printSegTimerStatus(); + return ret; +} diff --git a/paddle/legacy/gserver/tests/test_NetworkCompare.cpp b/paddle/legacy/gserver/tests/test_NetworkCompare.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c9f9f3e61be11fa33ab37e27065fdf275f86453a --- /dev/null +++ b/paddle/legacy/gserver/tests/test_NetworkCompare.cpp @@ -0,0 +1,294 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#undef PADDLE_DISABLE_TIMER +#include +#include +#include +#include + +#include "paddle/legacy/trainer/Trainer.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_double(checkgrad_eps); +DEFINE_bool(use_label, true, "input label or sequence label"); +DEFINE_bool(static_para, false, "static parameter"); + +struct DataIn { + std::vector inArgs; + std::vector outGrads; + std::vector paraValues; +}; + +struct DataOut { + std::vector outValues; + std::vector paraGrads; +}; + +void initArgument(DataIn& data, + const std::string& configPath, + bool useGpu = FLAGS_use_gpu) { + TrainerConfigHelper config(configPath); + size_t batchSize = config.getOptConfig().batch_size(); + + for (const auto& layer_name : config.getModelConfig().input_layer_names()) { + auto layer_config = std::find_if(config.getModelConfig().layers().begin(), + config.getModelConfig().layers().end(), + [=](const LayerConfig& layer_config) { + return layer_config.name() == layer_name; + }); + CHECK(layer_config != config.getModelConfig().layers().end()); + + size_t layerSize = layer_config->size(); + Argument arg; + arg.value = Matrix::create(batchSize, layerSize, false, useGpu); + arg.grad = Matrix::create(batchSize, layerSize, false, useGpu); + arg.value->randomizeUniform(); + arg.value->add(-0.5); + arg.value->sigmoid(*arg.value); + arg.grad->zeroMem(); + if (FLAGS_use_label) { + arg.ids = VectorT::create(batchSize, useGpu); + arg.ids->rand(layerSize); + } + generateSequenceStartPositions(batchSize, arg.sequenceStartPositions); + data.inArgs.push_back(arg); + } + + for (const auto& layer_name : config.getModelConfig().output_layer_names()) { + auto layer_config = std::find_if(config.getModelConfig().layers().begin(), + config.getModelConfig().layers().end(), + [=](const LayerConfig& layer_config) { + return layer_config.name() == layer_name; + }); + CHECK(layer_config != config.getModelConfig().layers().end()); + + size_t layerSize = layer_config->size(); + MatrixPtr grad = Matrix::create(batchSize, layerSize, false, useGpu); + grad->randomizeUniform(); + data.outGrads.push_back(grad); + } + + for (const auto& para_config : config.getModelConfig().parameters()) { + VectorPtr value = Vector::create(para_config.size(), useGpu); + value->randnorm(0, 2); + data.paraValues.push_back(value); + } +} + +void calcGradient(DataIn& in, DataOut& out, const std::string& configPath) { + *ThreadLocalRand::getSeed() = 0; + srand(0); + + Trainer trainer; + auto config = std::make_shared(configPath); + trainer.init(config, false); + + std::vector parameters; + vector outArgs; + + auto gradientMachine = trainer.getGradientMachine(); + parameters = gradientMachine->getParameters(); + if (FLAGS_static_para) { + for (size_t i = 0; i < parameters.size(); i++) { + parameters[i]->getBuf(PARAMETER_VALUE)->one(); + } + } else { + for (size_t i = 0; i < in.paraValues.size(); i++) { + parameters[i]->getBuf(PARAMETER_VALUE)->copyFrom(*in.paraValues[i]); + } + } + gradientMachine->start(); + gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN); + for (size_t i = 0; i < in.outGrads.size(); i++) { + // If the all the layers in the config have no parameters, also + // not set NeedGradient(), the outArgs[i] will be nullptr. + outArgs[i].grad->copyFrom(*in.outGrads[i]); + } + gradientMachine->backward(); + for (size_t i = 0; i < in.outGrads.size(); i++) { + MatrixPtr value = Matrix::create(outArgs[i].value->getHeight(), + outArgs[i].value->getWidth(), + false, + false); + value->copyFrom(*outArgs[i].value); + out.outValues.push_back(value); + } + for (size_t i = 0; i < in.paraValues.size(); i++) { + VectorPtr grad = Vector::create( + parameters[i]->getBuf(PARAMETER_GRADIENT)->getSize(), false); + grad->copyFrom(*parameters[i]->getBuf(PARAMETER_GRADIENT)); + out.paraGrads.push_back(grad); + } + + for (int i = 0; i < 20; i++) { + REGISTER_TIMER("forward"); + gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN); + } + for (int i = 0; i < 20; i++) { + REGISTER_TIMER("backward"); + gradientMachine->backward(); + } + + gradientMachine->finish(); +} + +void checkBuffer(real* A, + const char* desA, + real* B, + const char* desB, + size_t len, + size_t width = 1) { + int nNum = 0; + for (size_t i = 0; i < len; ++i) { + real diff = fabs(A[i] - B[i]); + if (diff > 0.0f && + diff / std::max(fabs(A[i]), fabs(B[i])) > FLAGS_checkgrad_eps) { + nNum++; + LOG(INFO) << "Row: " << i / width << ", " << desA << " : " << A[i] + << " " << desB << " : " << B[i]; + } + } + EXPECT_EQ(0, nNum); +} + +void compareGradient(DataOut& outA, DataOut& outB) { + LOG(INFO) << "------------------------------" + << " Check Network Output " + << "------------------------------"; + for (size_t i = 0; i < outA.outValues.size(); ++i) { + LOG(INFO) << "OUTPUT VALUE: " << i; + checkBuffer(outA.outValues[i]->getData(), + "network A output", + outB.outValues[i]->getData(), + "network B output", + outA.outValues[i]->getElementCnt(), + outA.outValues[i]->getWidth()); + } + + if (!FLAGS_static_para) { + LOG(INFO) << "------------------------------" + << " Check Parameters " + << "------------------------------"; + for (size_t i = 0; i < outA.paraGrads.size(); ++i) { + LOG(INFO) << "PARAMETER GRADIENT: " << i; + checkBuffer(outA.paraGrads[i]->getData(), + "Network A", + outB.paraGrads[i]->getData(), + "Network B", + outA.paraGrads[i]->getSize()); + } + } +} + +void compareNetwork(const std::string& config_file_a, + const std::string& config_file_b) { + DataIn in; + initArgument(in, config_file_a); + + DataOut dataA; + calcGradient(in, dataA, config_file_a); + LOG(INFO) << "forwardBackward of Network A is finished"; + globalStat.printSegTimerStatus(); + globalStat.reset(); + LOG(INFO) << "\n\n"; + + DataOut dataB; + calcGradient(in, dataB, config_file_b); + LOG(INFO) << "forwardBackward of the Network B is finished"; + globalStat.printSegTimerStatus(); + globalStat.reset(); + LOG(INFO) << "\n\n"; + + compareGradient(dataA, dataB); +} + +TEST(Compare, concat_dotmul) { + std::string config_file_a = "./legacy/gserver/tests/concat_dotmul_a.conf"; + std::string config_file_b = "./legacy/gserver/tests/concat_dotmul_b.conf"; + compareNetwork(config_file_a, config_file_b); +} + +TEST(Compare, concat_fullmatrix) { + std::string config_file_a = "./legacy/gserver/tests/concat_fullmatrix_a.conf"; + std::string config_file_b = "./legacy/gserver/tests/concat_fullmatrix_b.conf"; + compareNetwork(config_file_a, config_file_b); +} + +TEST(Compare, concat_table) { + std::string config_file_a = "./legacy/gserver/tests/concat_table_a.conf"; + std::string config_file_b = "./legacy/gserver/tests/concat_table_b.conf"; + compareNetwork(config_file_a, config_file_b); +} + +TEST(Compare, concat_slice) { + std::string config_file_a = "./legacy/gserver/tests/concat_slice_a.conf"; + std::string config_file_b = "./legacy/gserver/tests/concat_slice_b.conf"; + compareNetwork(config_file_a, config_file_b); +} + +#ifdef PADDLE_WITH_CUDA +TEST(Compare, img_pool) { + std::string config_file_a = "./legacy/gserver/tests/img_pool_a.conf"; + std::string config_file_b = "./legacy/gserver/tests/img_pool_b.conf"; + bool useGpu = FLAGS_use_gpu; + FLAGS_use_gpu = true; + compareNetwork(config_file_a, config_file_b); + FLAGS_use_gpu = useGpu; +} + +TEST(Compare, img_conv) { + std::string config_file_a = "./legacy/gserver/tests/img_conv_a.conf"; + std::string config_file_b = "./legacy/gserver/tests/img_conv_b.conf"; + bool useGpu = FLAGS_use_gpu; + FLAGS_use_gpu = true; + compareNetwork(config_file_a, config_file_b); + FLAGS_use_gpu = useGpu; +} + +// Test cudnn_conv and exconv give the same result +TEST(Compare, img_conv2) { + std::string config_file_a = "./legacy/gserver/tests/img_conv_cudnn.py"; + std::string config_file_b = "./legacy/gserver/tests/img_conv_exconv.py"; + bool useGpu = FLAGS_use_gpu; + double eps = FLAGS_checkgrad_eps; + FLAGS_use_gpu = true; + // Sometimes, this unit test will fail with 1e-2 + FLAGS_checkgrad_eps = 4e-2; + compareNetwork(config_file_a, config_file_b); + FLAGS_use_gpu = useGpu; + FLAGS_checkgrad_eps = eps; +} +#endif + +DEFINE_string(config_file_a, "", "config of one network to compare"); +DEFINE_string(config_file_b, "", "config of another network to compare"); +TEST(Compare, network) { + if (FLAGS_config_file_a != "" && FLAGS_config_file_b != "") { + compareNetwork(FLAGS_config_file_a, FLAGS_config_file_b); + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + paddle::initMain(argc, argv); + initPython(argc, argv); + int ret = RUN_ALL_TESTS(); + return ret; +} diff --git a/paddle/gserver/tests/test_PriorBox.cpp b/paddle/legacy/gserver/tests/test_PriorBox.cpp similarity index 100% rename from paddle/gserver/tests/test_PriorBox.cpp rename to paddle/legacy/gserver/tests/test_PriorBox.cpp diff --git a/paddle/legacy/gserver/tests/test_PyDataProvider.cpp b/paddle/legacy/gserver/tests/test_PyDataProvider.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0209e6818a8340fe128146909b9e8ec610e310a3 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_PyDataProvider.cpp @@ -0,0 +1,177 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include + +#include "paddle/legacy/gserver/dataproviders/PyDataProvider.h" +#include "paddle/legacy/utils/Util.h" + +#include "paddle/testing/TestUtil.h" + +using namespace std; // NOLINT +using namespace paddle; // NOLINT + +void simpleValueCheck(const vector& argumentList, bool useGpu); +void simpleSequenceCheck(const vector& argumentList, int sample_num); + +TEST(PyDataProvider, py_fill_slots) { + DataConfig config; + config.set_type("py"); + config.set_async_load_data(false); + config.set_load_data_module(std::string("pyDataProvider")); + config.set_load_data_object(std::string("SimpleDataProvider")); + config.clear_files(); + std::string dataFile = + "legacy/gserver/tests/pyDataProvider/pyDataProviderList"; + config.set_files(dataFile); +#ifndef PADDLE_WITH_CUDA + bool useGpu = false; +#else + bool useGpu = true; +#endif + unique_ptr dataProvider(DataProvider::create(config, useGpu)); + DataBatch dataBatch; + dataProvider->getNextBatchInternal(2, &dataBatch); + const std::vector& argumentList = dataBatch.getStreams(); + // Check size + EXPECT_EQ(argumentList.size(), 3UL); + EXPECT_EQ(argumentList[0].value->getWidth(), 3UL); + EXPECT_EQ(argumentList[0].value->getHeight(), 2UL); + EXPECT_EQ(argumentList[0].value->getElementCnt(), 6UL); + EXPECT_EQ(argumentList[1].value->getWidth(), 7UL); + EXPECT_EQ(argumentList[1].value->getHeight(), 2UL); + EXPECT_EQ(argumentList[1].value->getElementCnt(), 4UL); + EXPECT_EQ(argumentList[2].ids->getSize(), 2UL); + // Check value + simpleValueCheck(argumentList, useGpu); + // Check sequenceStartPositions + simpleSequenceCheck(argumentList, 2); +} + +TEST(PyDataProvider, py_fill_nest_slots) { + DataConfig config; + config.set_type("py"); + config.set_async_load_data(false); + config.set_load_data_module(std::string("pyDataProvider")); + config.set_load_data_object(std::string("SimpleNestDataProvider")); + config.clear_files(); + std::string dataFile = + "legacy/gserver/tests/pyDataProvider/pyDataProviderList"; + config.set_files(dataFile); + EXPECT_EQ(config.IsInitialized(), true); +#ifndef PADDLE_WITH_CUDA + bool useGpu = false; +#else + bool useGpu = true; +#endif + unique_ptr dataProvider(DataProvider::create(config, useGpu)); + DataBatch dataBatch; + dataProvider->getNextBatchInternal(2, &dataBatch); + const std::vector& argumentList = dataBatch.getStreams(); + // Check size + EXPECT_EQ(argumentList.size(), 3UL); + EXPECT_EQ(argumentList[0].value->getWidth(), 3UL); + EXPECT_EQ(argumentList[0].value->getHeight(), 4UL); + EXPECT_EQ(argumentList[0].value->getElementCnt(), 12UL); + EXPECT_EQ(argumentList[1].value->getWidth(), 7UL); + EXPECT_EQ(argumentList[1].value->getHeight(), 4UL); + EXPECT_EQ(argumentList[1].value->getElementCnt(), 8UL); + EXPECT_EQ(argumentList[2].ids->getSize(), 4UL); + // Check value + simpleValueCheck(argumentList, useGpu); + // Check sequenceStartPositions + simpleSequenceCheck(argumentList, 4); + // Check subSequenceStartPositions + EXPECT_EQ(argumentList[0].subSequenceStartPositions->getSize(), 4UL); + EXPECT_EQ(argumentList[1].subSequenceStartPositions->getSize(), 3UL); + EXPECT_EQ(argumentList[2].subSequenceStartPositions->getSize(), 4UL); + for (size_t i = 0; i < argumentList.size(); i++) { + EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(0), 0); + EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(1), 1); + if (i != 1) { + EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(2), 2); + EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(3), 4); + } else { + EXPECT_EQ(argumentList[i].subSequenceStartPositions->getElement(2), 4); + } + } +} + +void simpleValueCheck(const vector& argumentList, bool useGpu) { + // Dense + real* data; + if (useGpu) { + MatrixPtr cpuMatrixPtr = Matrix::create(argumentList[0].value->getHeight(), + argumentList[0].value->getWidth(), + 0, + 0); + cpuMatrixPtr->copyFrom(*argumentList[0].value); + data = cpuMatrixPtr->getData(); + } else { + data = argumentList[0].value->getData(); + } + for (size_t i = 0; i < argumentList[0].value->getElementCnt(); ++i) { + EXPECT_EQ(*(data + i), (float)(i % 3 + 1)); + } + // Sparse without value + GpuSparseMatrixPtr matGpu; + CpuSparseMatrixPtr matCpu; + if (useGpu) { + matGpu = dynamic_pointer_cast(argumentList[1].value); + ASSERT_TRUE(matGpu != NULL); + } else { + data = argumentList[0].value->getData(); + matCpu = dynamic_pointer_cast(argumentList[1].value); + ASSERT_TRUE(matCpu != NULL); + } + for (size_t i = 0; i < argumentList[1].value->getHeight(); ++i) { + size_t colNum = useGpu ? matGpu->getColNum(i) : matCpu->getColNum(i); + EXPECT_EQ(colNum, (size_t)2); + const int* buf = useGpu ? matGpu->getRowCols(i) : matCpu->getRowCols(i); + for (size_t j = 0; j < colNum; ++j) { + EXPECT_EQ((size_t)buf[j], (size_t)(j + 1)); + } + } + // Index + for (size_t j = 0; j < argumentList[2].ids->getSize(); ++j) { + EXPECT_EQ((size_t)argumentList[2].ids->get(j), 0UL); + } +} + +void simpleSequenceCheck(const vector& argumentList, int sample_num) { + EXPECT_EQ(argumentList[0].sequenceStartPositions->getSize(), 3UL); + EXPECT_EQ(argumentList[1].sequenceStartPositions->getSize(), 2UL); + EXPECT_EQ(argumentList[2].sequenceStartPositions->getSize(), 3UL); + for (size_t i = 0; i < argumentList.size(); i++) { + EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(0), 0); + if (i != 1) { + EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(1), 1); + EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(2), + sample_num); + } else { + EXPECT_EQ(argumentList[i].sequenceStartPositions->getElement(1), + sample_num); + } + } +} + +int main(int argc, char** argv) { + initMain(argc, argv); + initPython(argc, argv); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_PyDataProvider2.cpp b/paddle/legacy/gserver/tests/test_PyDataProvider2.cpp new file mode 100644 index 0000000000000000000000000000000000000000..de313ba82cf2697c13d6eae17056240b6272ca1c --- /dev/null +++ b/paddle/legacy/gserver/tests/test_PyDataProvider2.cpp @@ -0,0 +1,409 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef PADDLE_NO_PYTHON +#include +#include +#include "paddle/legacy/gserver/dataproviders/DataProvider.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Util.h" + +DEFINE_string(train_list, "unittest.list", "file list for unittest"); + +namespace paddle { +namespace unittest { +namespace pydp2 { +extern void setOnPoolFilledHook(const std::function &func); +extern void clearOnPoolFilledHook(); + +} // namespace pydp2 +} // namespace unittest +} // namespace paddle + +const paddle::real epsilon = 1e-5; + +static inline int64_t readDataBatch(paddle::DataBatch *batch, + const std::string &funcName, + int64_t batchSize = 65535) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object(funcName); + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->setSkipShuffle(); + provider->reset(); + return provider->getNextBatchInternal(batchSize, batch); +} + +TEST(PyDataProvider2, dense_no_seq) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_dense_no_seq"); + + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + + provider->setSkipShuffle(); // skip shuffle for unittest. + + paddle::DataBatch batch; + for (size_t pass = 0; pass < 2; ++pass) { // read 2 passes + provider->reset(); + int64_t num = provider->getNextBatchInternal(100, &batch); + ASSERT_NE(num, 0); + ASSERT_EQ((size_t)batch.getStreams().size(), (size_t)1); + ASSERT_EQ((size_t)batch.getSize(), (size_t)100); + // Check batch data. + for (size_t i = 0; i < 100; ++i) { + for (size_t j = 0; j < 200; ++j) { + paddle::real tmp = (paddle::real)((j - 100.0) * (i + 1) / 200.0); + ASSERT_NEAR( + batch.getStreams()[0].value->getData()[i * 200 + j], tmp, epsilon); + } + } + + num = provider->getNextBatchInternal(100, &batch); + ASSERT_NE(num, 0); + ASSERT_EQ(batch.getStreams().size(), (size_t)1); + ASSERT_EQ((size_t)batch.getSize(), (size_t)100); + // Check batch data. + for (size_t i = 0; i < 100; ++i) { + size_t ii = i + 100; + for (size_t j = 0; j < 200; ++j) { + paddle::real tmp = (paddle::real)((j - 100.0) * (ii + 1) / 200.0); + ASSERT_NEAR( + batch.getStreams()[0].value->getData()[i * 200 + j], tmp, epsilon); + } + } + num = provider->getNextBatchInternal(100, &batch); + ASSERT_EQ(num, 0); + } +} + +TEST(PyDataProvider2, index_no_seq) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_index_no_seq"); + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + + provider->setSkipShuffle(); // skip shuffle for unittest. + paddle::DataBatch batch; + for (size_t pass = 0; pass < 2; ++pass) { + provider->reset(); + int64_t num = provider->getNextBatchInternal(10000, &batch); + CHECK_EQ(num, 200); + for (int i = 0; i < 200; ++i) { + CHECK_EQ(i, batch.getStreams()[0].ids->getData()[i]); + } + } +} + +TEST(PyDataProvider2, init_hook) { + paddle::PyObjectPtr pickle = paddle::py::import("pickle"); + paddle::PyObjectPtr globals(PyModule_GetDict(PyImport_AddModule("__main__"))); + PyDict_SetItemString(globals.get(), "pickle", pickle.get()); + paddle::PyObjectPtr locals(PyDict_New()); + paddle::PyObjectPtr mdl(PyRun_String( + "dumps = pickle.dumps({'value':[float(x) for x in xrange(20)]})", + Py_file_input, + globals.get(), + locals.get())); + CHECK_PY(mdl) << "Error!"; + paddle::PyObjectPtr dps(PyDict_GetItemString(locals.get(), "dumps")); + CHECK_PY(dps) << "Error!"; + + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_init_hook"); + config.set_load_data_args(PyString_AsString(dps.get())); + + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->setSkipShuffle(); // skip shuffle for unittest. + provider->reset(); + paddle::DataBatch batch; + int64_t num = provider->getNextBatchInternal(100000, &batch); + ASSERT_EQ(num, 200); + auto &mat = batch.getStreams()[0].value; + ASSERT_EQ((size_t)mat->getWidth(), (size_t)20); + for (size_t i = 0; i < 200; ++i) { + for (size_t j = 0; j < 20; ++j) { + ASSERT_NEAR((paddle::real)j, mat->getData()[i * 20 + j], epsilon); + } + } +} + +TEST(PyDataProvider2, sparse_no_value_no_seq) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_sparse_non_value_no_seq"); + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->setSkipShuffle(); + provider->reset(); + paddle::DataBatch batch; + int64_t num = provider->getNextBatchInternal(10000, &batch); + CHECK_EQ(num, 200); + auto csm = std::dynamic_pointer_cast( + batch.getStreams()[0].value); + CHECK(csm != nullptr); + for (int i = 0; i < 200; ++i) { + CHECK_EQ(csm->getColNum(i), (size_t)10); + int *cols = csm->getRowCols(i); + for (int j = 0; j < 10; ++j) { + CHECK_EQ(cols[j], (i + 1) * (j + 1)); + } + } +} + +TEST(PyDataProvider2, sparse_value_no_seq) { + paddle::DataBatch batch; + CHECK_EQ(readDataBatch(&batch, "test_sparse_value_no_seq"), 200); + auto csm = std::dynamic_pointer_cast( + batch.getStreams()[0].value); + CHECK(csm != nullptr); + for (int i = 0; i < 200; ++i) { + CHECK_EQ(csm->getColNum(i), (size_t)10); + int *cols = csm->getRowCols(i); + real *dat = csm->getRowValues(i); + for (int j = 0; j < 10; ++j) { + EXPECT_EQ(cols[j], (i + 1) * (j + 1)); + EXPECT_EQ(dat[j], real(j) / real(i + 1)); + } + } +} + +TEST(PyDataProvider2, index_seq) { + paddle::DataBatch batch; + CHECK_EQ(readDataBatch(&batch, "test_index_seq"), 200); + auto &arg = batch.getStreams()[0]; + CHECK_EQ((int)arg.ids->getSize(), (200 + 1) * 200 / 2); + size_t tmp = 0; + for (size_t i = 0; i < 200; ++i) { // CHECK DATA CORRECT + for (size_t j = 0; j < i + 1; ++j) { + ASSERT_EQ((size_t)arg.ids->getData()[tmp], j); + ++tmp; + } + } + ASSERT_EQ(arg.sequenceStartPositions->getSize(), (size_t)201); + tmp = 0; + for (size_t i = 0; i < 200; ++i) { + tmp += i; + ASSERT_EQ((size_t)arg.sequenceStartPositions->getData(false)[i], tmp); + } + tmp += 200; + ASSERT_EQ((size_t)arg.sequenceStartPositions->getData(false)[200], tmp); +} + +TEST(PyDataProvider2, index_sub_seq) { + paddle::DataBatch batch; + ASSERT_EQ(readDataBatch(&batch, "test_index_sub_seq"), 200); + auto &arg = batch.getStreams()[0]; + size_t tmp = 0; + for (size_t i = 0; i < 200; ++i) { + for (size_t j = 0; j < i + 1; ++j) { + for (size_t k = 0; k < j + 1; ++k) { + CHECK_EQ((size_t)arg.ids->getData()[tmp++], k); + } + } + } + + CHECK_EQ(tmp, arg.ids->getSize()); + + ASSERT_EQ((size_t)arg.sequenceStartPositions->getSize(), (size_t)201); + ASSERT_EQ(arg.subSequenceStartPositions->getData(false)[0], 0); + ASSERT_EQ(arg.sequenceStartPositions->getData(false)[0], 0); + size_t idx = 1; + tmp = 0; + for (size_t i = 0; i < 200; ++i) { + for (size_t j = 0; j < i + 1; ++j) { + tmp += j + 1; + ASSERT_EQ((size_t)arg.subSequenceStartPositions->getData(false)[idx], + (size_t)tmp); + ++idx; + } + ASSERT_EQ((size_t)arg.sequenceStartPositions->getData(false)[i + 1], tmp); + } +} + +TEST(PyDataProvider2, min_pool_size) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_min_pool_size"); + config.set_load_data_args(""); + size_t totalData = 1 << 14; + constexpr size_t batchSize = 100; + constexpr size_t minPoolSize = 1000; + paddle::DataBatch batch; + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->reset(); + + paddle::unittest::pydp2::setOnPoolFilledHook([&](size_t poolSize) { + if (totalData > batchSize) { + CHECK_GE(poolSize, std::min(totalData - batchSize, minPoolSize)); + } + }); + while (true) { + int64_t realBatchSize = provider->getNextBatchInternal(batchSize, &batch); + if (realBatchSize) { + totalData -= realBatchSize; + } else { + break; + } + } + paddle::unittest::pydp2::clearOnPoolFilledHook(); +} + +TEST(PyDataProvider2, can_over_batch_size) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_can_over_batch_size"); + config.set_load_data_args(""); + paddle::DataBatch batch; + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->reset(); + constexpr size_t batchSize = 100; + while (true) { + int64_t realBatchSize = provider->getNextBatchInternal(batchSize, &batch); + if (realBatchSize) { + CHECK_LE(static_cast(realBatchSize), batchSize); + } else { + break; + } + } +} + +TEST(PyDataProvider2, input_order) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_input_order"); + config.set_load_data_args(""); + + paddle::ModelConfig modelConfig; + *modelConfig.add_input_layer_names() = "input1"; + *modelConfig.add_input_layer_names() = "input2"; + paddle::DataBatch batch; + std::unique_ptr provider( + paddle::DataProvider::create(config, modelConfig, false)); + provider->reset(); + constexpr size_t batchSize = 100; + while (true) { + int64_t realBatchSize = provider->getNextBatchInternal(batchSize, &batch); + if (!realBatchSize) { + break; + } + ASSERT_EQ(batch.getStreams().size(), static_cast(2)); + for (int64_t i = 0; i < realBatchSize; ++i) { + ASSERT_EQ(batch.getStream(0).ids->getData()[i], 0); + ASSERT_EQ(batch.getStream(1).ids->getData()[i], 1); + } + } +} + +TEST(PyDataProvider2, test_check) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_check"); + config.set_load_data_args(""); + paddle::DataBatch batch; + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->reset(); + while (true) { + int64_t realBatchSize = provider->getNextBatchInternal(100, &batch); + if (!realBatchSize) { + break; + } else { + auto &ivec = batch.getStream(0).ids; + for (size_t i = 0; i < ivec->getSize(); ++i) { + CHECK_LT(ivec->getData()[i], 10); + } + } + } +} + +TEST(PyDataProvider2, multiThread) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_dense_no_seq"); + config.set_async_load_data(true); + + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + provider->reset(); + paddle::DataBatch batch; + provider->getNextBatch(100, &batch); + provider->reset(); + provider.reset(); +} + +TEST(PyDataProvider2, minPoolSizeWithCache) { + paddle::DataConfig config; + config.set_type("py2"); + config.set_files(FLAGS_train_list.c_str()); + config.set_load_data_module("test_PyDataProvider2"); + config.set_load_data_object("test_min_pool_size_with_cache"); + config.set_async_load_data(true); + + std::unique_ptr provider( + paddle::DataProvider::create(config, false)); + + paddle::DataBatch batch; + + for (int i = 0; i < 10; ++i) { + provider->reset(); + int64_t sum = 0; + while (int64_t actualNum = provider->getNextBatch(100, &batch)) { + sum += actualNum; + } + ASSERT_EQ(1 << 20, sum); + } +} + +int main(int argc, char **argv) { + testing::InitGoogleTest(&argc, argv); + paddle::initMain(argc, argv); + paddle::initPython(argc, argv); + + std::ofstream fout(FLAGS_train_list); + CHECK(fout.is_open()); + fout << "stub file name" << std::endl; // in unittest, filename is not used. + fout.close(); + + return RUN_ALL_TESTS(); +} + +#endif diff --git a/paddle/gserver/tests/test_PyDataProvider2.py b/paddle/legacy/gserver/tests/test_PyDataProvider2.py similarity index 100% rename from paddle/gserver/tests/test_PyDataProvider2.py rename to paddle/legacy/gserver/tests/test_PyDataProvider2.py diff --git a/paddle/legacy/gserver/tests/test_RecurrentGradientMachine.cpp b/paddle/legacy/gserver/tests/test_RecurrentGradientMachine.cpp new file mode 100644 index 0000000000000000000000000000000000000000..153c3e7f36a30a70d0c5870144a0091b1e5f7237 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_RecurrentGradientMachine.cpp @@ -0,0 +1,180 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include +#include + +DECLARE_int32(seed); + +using namespace paddle; // NOLINT +using namespace std; // NOLINT +class TrainerForTest : public paddle::Trainer { + public: + void startTrain() { + GradientMachine& gm = *this->trainerInternal_.getGradientMachine(); + gm.start(); + } + + void finishTrain() { + GradientMachine& gm = *this->trainerInternal_.getGradientMachine(); + gm.finish(); + } + + /** + * Get total dimension of all parameters. + * + * @return the total dimension of all parameters + */ + size_t getTotalParameterSize() const { + auto p = const_cast(this); + auto& params = p->getGradientMachine()->getParameters(); + return std::accumulate( + params.begin(), params.end(), 0UL, [](size_t a, const ParameterPtr& p) { + return a + p->getSize(); + }); + } +}; + +void CalCost(const string& conf, + const string& dir, + real* cost, + int num_passes) { + auto config = std::make_shared(conf); + TrainerForTest trainer; + trainer.init(config); + mkDir(dir.c_str()); + config->setSaveDir(dir); + auto dataProvider = trainer.getDataProvider(); + int32_t batchSize = config->getOptConfig().batch_size(); + real learningRate = config->getOptConfig().learning_rate(); + real momentum = 0; + real decayRate = 0; + int64_t dim = trainer.getTotalParameterSize(); + CpuVector vecW(dim); + CpuVector vecGradient(dim); + CpuVector vecMomentum(dim); + + // vecW needs to be assigned, otherwise the variable is an uncertain value. + + *ThreadLocalRand::getSeed() = FLAGS_seed; + vecW.randnorm(0, 0.1); + vecMomentum.randnorm(0, 0.1); + + trainer.startTrain(); + for (int i = 0; i < num_passes; ++i) { + real totalCost = 0; + dataProvider->reset(); + while (true) { + DataBatch dataBatch; + int num = dataProvider->getNextBatch(batchSize, &dataBatch); + if (num == 0) break; + totalCost += trainer.calcGradient(dataBatch, vecW, vecGradient); + sgdUpdate( + learningRate, momentum, decayRate, &vecW, &vecGradient, &vecMomentum); + } + cost[i] = totalCost; + } + trainer.finishTrain(); + rmDir(dir.c_str()); +} + +void test(const string& conf1, const string& conf2, double eps, bool useGpu) { + if (!paddle::version::isWithGpu() && useGpu) { + return; + } + FLAGS_use_gpu = useGpu; + int num_passes = 5; + real* cost1 = new real[num_passes]; + const string dir1 = "legacy/gserver/tests/t1"; + CalCost(conf1, dir1, cost1, num_passes); + + real* cost2 = new real[num_passes]; + const string dir2 = "legacy/gserver/tests/t2"; + CalCost(conf2, dir2, cost2, num_passes); + + for (int i = 0; i < num_passes; i++) { + LOG(INFO) << "num_passes: " << i << ", cost1=" << cost1[i] + << ", cost2=" << cost2[i] + << ", diff=" << std::abs(cost1[i] - cost2[i]); + ASSERT_NEAR(cost1[i], cost2[i], eps); + } + delete[] cost1; + delete[] cost2; +} + +TEST(RecurrentGradientMachine, HasSubSequence) { + for (bool useGpu : {false, true}) { + test("legacy/gserver/tests/sequence_layer_group.conf", + "legacy/gserver/tests/sequence_nest_layer_group.conf", + 1e-5, + useGpu); + } +} + +TEST(RecurrentGradientMachine, rnn) { + for (bool useGpu : {false, true}) { + test("legacy/gserver/tests/sequence_rnn.conf", + "legacy/gserver/tests/sequence_nest_rnn.conf", + 1e-6, + useGpu); + } +} + +TEST(RecurrentGradientMachine, rnn_multi_input) { + for (bool useGpu : {false, true}) { + test("legacy/gserver/tests/sequence_rnn_multi_input.conf", + "legacy/gserver/tests/sequence_nest_rnn_multi_input.conf", + 1e-6, + useGpu); + } +} + +TEST(RecurrentGradientMachine, rnn_multi_unequalength_input) { + for (bool useGpu : {false, true}) { + test("legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py", + "legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py", + 1e-6, + useGpu); + } +} + +TEST(RecurrentGradientMachine, rnn_mixed_input) { + for (bool useGpu : {false, true}) { + test("legacy/gserver/tests/sequence_rnn_mixed_inputs.py", + "legacy/gserver/tests/sequence_rnn_matched_inputs.py", + 1e-6, + useGpu); + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + + if (paddle::version::isWithPyDataProvider()) { + if (!paddle::version::isWithGpu()) { + FLAGS_use_gpu = false; + } + initMain(argc, argv); + initPython(argc, argv); + return RUN_ALL_TESTS(); + } else { + return 0; + } +} diff --git a/paddle/legacy/gserver/tests/test_RecurrentLayer.cpp b/paddle/legacy/gserver/tests/test_RecurrentLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..71198cb6a1d29433ed0e315378f5aee51b921766 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_RecurrentLayer.cpp @@ -0,0 +1,571 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/gserver/layers/Layer.h" + +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT +DECLARE_bool(use_gpu); +DECLARE_bool(rnn_use_batch); +DECLARE_int32(fixed_seq_length); + +void checkError(const Matrix& matrix1, const Matrix& matrix2) { + CHECK(matrix1.getHeight() == matrix2.getHeight()); + CHECK(matrix1.getWidth() == matrix2.getWidth()); +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + + int height = matrix1.getHeight(); + int width = matrix1.getWidth(); + const real* data1 = matrix1.getData(); + const real* data2 = matrix2.getData(); + int count = 0; + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + if (fabs(data1[i * width + j] - data2[i * width + j]) > err) { + count++; + } + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; +} + +void checkError(const CpuVector& vector1, const CpuVector& vector2) { + CHECK(vector1.getSize() == vector2.getSize()); +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + + int size = vector1.getSize(); + const real* data1 = vector1.getData(); + const real* data2 = vector2.getData(); + int count = 0; + for (int i = 0; i < size; i++) { + if (fabs(data1[i] - data2[i]) > err) { + count++; + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; +} + +LayerPtr creatDataLayer(string name, + size_t batchSize, + int layerSize, + bool useGpu) { + LayerConfig dataConfig; + dataConfig.set_name(name); + dataConfig.set_type("data"); + dataConfig.set_size(layerSize); + LayerPtr layer = LayerPtr(new DataLayer(dataConfig)); + + Argument data; + data.value = Matrix::create(batchSize, layer->getSize(), false, useGpu); + data.grad = Matrix::create(batchSize, layer->getSize(), false, useGpu); + data.value->randomizeUniform(); + data.value->add(-0.5); + data.value->sigmoid(*data.value); + data.grad->zeroMem(); + + generateSequenceStartPositions(batchSize, data.sequenceStartPositions); + + DataLayerPtr dataLayer = std::dynamic_pointer_cast(layer); + dataLayer->setData(data); + dataLayer->forward(PASS_GC); + + return layer; +} + +ParameterPtr creatParameter(string name, + int pid, + size_t paraSize, + bool useGpu) { + ParameterConfig paraConfig; + paraConfig.set_name(name); + paraConfig.set_size(paraSize); + + ParameterPtr parameter = + std::make_shared(paraConfig, useGpu, /*initialize */ false); + parameter->enableType(PARAMETER_VALUE); + parameter->enableType(PARAMETER_GRADIENT); + parameter->randomize(); + parameter->setID(pid); + + return parameter; +} + +ParameterPtr creatParameterBias(string name, + int pid, + size_t paraSize, + bool useGpu) { + ParameterConfig paraConfig; + paraConfig.set_name(name); + paraConfig.set_size(paraSize); + paraConfig.set_initial_std(1); + + ParameterPtr parameter = + std::make_shared(paraConfig, useGpu, /*initialize */ true); + parameter->randomize(); + parameter->setID(pid); + + return parameter; +} + +LayerPtr initRecurrentLayer(LayerConfig layerConfig, + size_t batchSize, + int layerSize, + bool useGpu) { + FLAGS_use_gpu = useGpu; + LayerMap layerMap; + ParameterMap parameterMap; + LayerPtr dataLayer = creatDataLayer("layer_0", batchSize, layerSize, useGpu); + layerMap[dataLayer->getName()] = dataLayer; + + ParameterPtr para = + creatParameter("para_0", 0, layerSize * layerSize, useGpu); + parameterMap[para->getName()] = para; + + layerConfig.add_inputs(); + LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); + input.set_input_layer_name("layer_0"); + input.set_input_parameter_name("para_0"); + LayerPtr testLayer = Layer::create(layerConfig); + layerMap[testLayer->getName()] = testLayer; + + testLayer->init(layerMap, parameterMap); + testLayer->setNeedGradient(true); + + return testLayer; +} + +void checkRecurrentLayer(LayerPtr testLayer) { + const VectorPtr& weightGrad = + (testLayer->getParameters()[0])->getBuf(PARAMETER_GRADIENT); + const MatrixPtr& inputGrad = testLayer->getPrev(0)->getOutputGrad(); + CpuVector seqPara(weightGrad->getSize()); + CpuVector batPara(weightGrad->getSize()); + CpuMatrix seqInputGrad(inputGrad->getHeight(), inputGrad->getWidth()); + CpuMatrix batInputGrad(inputGrad->getHeight(), inputGrad->getWidth()); + + CpuMatrix outputGrad(inputGrad->getHeight(), inputGrad->getWidth()); + outputGrad.randomizeUniform(); + + /* use sequence calculate */ + FLAGS_rnn_use_batch = false; + weightGrad->zero(); + inputGrad->zero(); + testLayer->forward(PASS_GC); + testLayer->getOutputGrad()->copyFrom(outputGrad); + testLayer->backward(); + seqPara.copyFrom(*weightGrad); + seqInputGrad.copyFrom(*inputGrad); + + /* use batch calculate */ + FLAGS_rnn_use_batch = true; + weightGrad->zero(); + inputGrad->zero(); + testLayer->forward(PASS_GC); + testLayer->getOutputGrad()->copyFrom(outputGrad); + testLayer->backward(); + batPara.copyFrom(*weightGrad); + batInputGrad.copyFrom(*inputGrad); + + /* check */ + checkError(seqInputGrad, batInputGrad); + checkError(seqPara, batPara); +} + +TEST(Layer, RecurrentLayer) { + LayerConfig layerConfig; + layerConfig.set_name("rnn"); + layerConfig.set_type("recurrent"); + layerConfig.set_active_type("tanh"); + for (auto layerSize : {1, 10, 64, 128, 256, 512}) { + for (auto batchSize : {1, 5, 20, 100, 128}) { + for (auto useGpu : {false, true}) { + for (auto reversed : {false, true}) { + LOG(INFO) << " layerSize=" << layerSize << " batchSize=" << batchSize + << " useGpu=" << useGpu << " reversed=" << reversed; + layerConfig.set_size(layerSize); + layerConfig.set_reversed(reversed); + LayerPtr testLayer = + initRecurrentLayer(layerConfig, batchSize, layerSize, useGpu); + checkRecurrentLayer(testLayer); + } + } + } + } +} + +#define protected public +#include "paddle/legacy/gserver/layers/GatedRecurrentLayer.h" +#include "paddle/legacy/gserver/layers/LstmLayer.h" +#include "paddle/legacy/gserver/layers/RecurrentLayer.h" +template +class TestRecurrentLayer { + public: + LayerConfig config_; + bool useGpu_; + bool useBatch_; + LayerPtr testLayer_; + LayerPtr dataLayer_; + ParameterPtr para_; + ParameterPtr bias_; + LayerMap layerMap_; + ParameterMap parameterMap_; + TestRecurrentLayer(const LayerConfig& config, + bool useGpu, + bool useBatch = false) + : config_(config), useGpu_(useGpu), useBatch_(useBatch) {} + void init(size_t batchSize) { + FLAGS_use_gpu = useGpu_; + testLayer_ = Layer::create(config_); + if (typeid(T) == typeid(GatedRecurrentLayer)) { + dataLayer_ = creatDataLayer(config_.mutable_inputs(0)->input_layer_name(), + batchSize, + config_.size() * 3, + useGpu_); + para_ = creatParameter(config_.mutable_inputs(0)->input_parameter_name(), + 0, + config_.size() * config_.size() * 3, + useGpu_); + bias_ = creatParameterBias( + config_.bias_parameter_name(), 1, config_.size() * 3, useGpu_); + } else if (typeid(T) == typeid(LstmLayer)) { + dataLayer_ = creatDataLayer(config_.mutable_inputs(0)->input_layer_name(), + batchSize, + config_.size() * 4, + useGpu_); + para_ = creatParameter(config_.mutable_inputs(0)->input_parameter_name(), + 0, + config_.size() * config_.size() * 4, + useGpu_); + bias_ = creatParameterBias( + config_.bias_parameter_name(), 1, config_.size() * 7, useGpu_); + } + layerMap_[dataLayer_->getName()] = dataLayer_; + parameterMap_[para_->getName()] = para_; + parameterMap_[bias_->getName()] = bias_; + + layerMap_[testLayer_->getName()] = testLayer_; + testLayer_->init(layerMap_, parameterMap_); + testLayer_->setNeedGradient(true); + (dynamic_cast(testLayer_.get()))->useBatch_ = useBatch_; + } + void forward() { + FLAGS_use_gpu = useGpu_; + testLayer_->forward(PASS_GC); + } + void backward() { + FLAGS_use_gpu = useGpu_; + testLayer_->backward(nullptr); + } +}; + +template +void checkRecurrentLayer(LayerConfig layerConfig, + size_t batchSize, + bool cpuBatch, + bool gpuBatch) { + TestRecurrentLayer testCpu(layerConfig, false, cpuBatch); + TestRecurrentLayer testGpu(layerConfig, true, gpuBatch); + testCpu.init(batchSize); + testGpu.init(batchSize); + auto checkError = []( + MatrixPtr cpu, MatrixPtr gpu, int numSequences, const char* str) { + CpuMatrix check(gpu->getHeight(), gpu->getWidth()); + check.copyFrom(*gpu); + int height = cpu->getHeight(); + int width = cpu->getWidth(); + const real* data1 = cpu->getData(); + const real* data2 = check.getData(); + int count = 0; + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + if (fabs(data1[i * width + j] - data2[i * width + j]) / numSequences > + 1e-4) { + count++; + } + } + } + EXPECT_EQ(count, 0) << "[" << str << "]" + << "There are " << count << " different element."; + }; + T* cpuLayer = dynamic_cast(testCpu.testLayer_.get()); + T* gpuLayer = dynamic_cast(testGpu.testLayer_.get()); + + Argument& cpuInput = testCpu.dataLayer_->getOutput(); + Argument& gpuInput = testGpu.dataLayer_->getOutput(); + gpuInput.resizeAndCopyFrom(cpuInput, true); + + const VectorPtr& cpuVec = testCpu.para_->getBuf(PARAMETER_VALUE); + const VectorPtr& gpuVec = testGpu.para_->getBuf(PARAMETER_VALUE); + gpuVec->copyFrom(*cpuVec); + + const VectorPtr& cpuBiasVec = testCpu.bias_->getBuf(PARAMETER_VALUE); + const VectorPtr& gpuBiasVec = testGpu.bias_->getBuf(PARAMETER_VALUE); + gpuBiasVec->copyFrom(*cpuBiasVec); + + /* check forward */ + testCpu.forward(); + testGpu.forward(); + + checkError( + cpuLayer->getOutputValue(), gpuLayer->getOutputValue(), 1, "outputValue"); + + /* check backward */ + cpuLayer->getOutputGrad()->randomizeUniform(); + gpuLayer->getOutputGrad()->copyFrom(*cpuLayer->getOutputGrad()); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + + testCpu.backward(); + testGpu.backward(); + + // check input grad + checkError(cpuInput.grad, gpuInput.grad, 1, "inputGrad"); + // check weight grad + int numSequences = cpuInput.getNumSequences(); + checkError(cpuLayer->weight_->getWGrad(), + gpuLayer->weight_->getWGrad(), + numSequences, + "weightGrad"); + // check bias grad + checkError(cpuLayer->bias_->getWGrad(), + gpuLayer->bias_->getWGrad(), + numSequences, + "biasGrad"); +} + +TEST(Layer, GatedRecurrentLayer) { + LayerConfig layerConfig; + layerConfig.set_type("gated_recurrent"); + layerConfig.set_active_type("sigmoid"); + layerConfig.set_active_gate_type("sigmoid"); + + layerConfig.add_inputs(); + LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); + input.set_input_layer_name("layer_0"); + input.set_input_parameter_name("para_0"); + layerConfig.set_bias_parameter_name("bias"); + + for (auto frameSize : {32, 64, 128, 256, 512}) { + for (auto batchSize : {1, 5, 100, 500}) { + for (auto reversed : {false, true}) { + for (auto cpuBatch : {false, true}) { + for (auto gpuBatch : {false, true}) { + LOG(INFO) << " batchSize=" << batchSize + << " frameSize=" << frameSize << " reversed=" << reversed + << " cpuBatch=" << cpuBatch << " gpuBatch=" << gpuBatch; + layerConfig.set_size(frameSize); + layerConfig.set_reversed(reversed); + checkRecurrentLayer( + layerConfig, batchSize, cpuBatch, gpuBatch); + } + } + } + } + } +} + +TEST(Layer, LstmLayer) { + LayerConfig layerConfig; + layerConfig.set_type("lstmemory"); + layerConfig.set_active_type("relu"); + layerConfig.set_active_state_type("tanh"); + layerConfig.set_active_gate_type("sigmoid"); + + layerConfig.add_inputs(); + LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); + input.set_input_layer_name("layer_0"); + input.set_input_parameter_name("para_0"); + layerConfig.set_bias_parameter_name("bias"); + + for (auto frameSize : {32, 64, 128, 256, 512}) { + for (auto batchSize : {1, 5, 100, 500}) { + for (auto reversed : {false, true}) { + for (auto cpuBatch : {false, true}) { + for (auto gpuBatch : {false, true}) { + LOG(INFO) << " batchSize=" << batchSize + << " frameSize=" << frameSize << " reversed=" << reversed + << " cpuBatch=" << cpuBatch << " gpuBatch=" << gpuBatch; + layerConfig.set_size(frameSize); + layerConfig.set_reversed(reversed); + checkRecurrentLayer( + layerConfig, batchSize, cpuBatch, gpuBatch); + } + } + } + } + } +} + +#ifdef PADDLE_WITH_MKLML + +#include "paddle/legacy/gserver/layers/MKLPackedRecurrentLayer.h" + +LayerPtr initMKLPackedLayer(LayerConfig layerConfig, + bool reversed, + int layerSize, + LayerPtr dataLayer, + ParameterPtr para, + ParameterPtr bias = nullptr) { + LayerMap layerMap; + ParameterMap parameterMap; + layerMap[dataLayer->getName()] = dataLayer; + parameterMap[para->getName()] = para; + if (bias) { + parameterMap[bias->getName()] = bias; + layerConfig.set_bias_parameter_name("bias_0"); + } + + layerConfig.set_size(layerSize); + layerConfig.set_reversed(reversed); + layerConfig.add_inputs(); + LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); + input.set_input_layer_name("layer_0"); + input.set_input_parameter_name("para_0"); + + LayerPtr testLayer = Layer::create(layerConfig); + layerMap[testLayer->getName()] = testLayer; + + testLayer->init(layerMap, parameterMap); + testLayer->setNeedGradient(true); + + return testLayer; +} + +void checkMKLPackedLayer(LayerConfig layerConfig1, + LayerConfig layerConfig2, + bool reversed, + int layerSize, + int batchSize, + bool useBatch1, + bool useBatch2) { + LayerPtr dataLayer; + ParameterPtr para, bias; + + if (layerConfig1.type() == "recurrent") { + dataLayer = creatDataLayer("layer_0", batchSize, layerSize, false); + para = creatParameter("para_0", 0, layerSize * layerSize, false); + bias = nullptr; + } else if (layerConfig1.type() == "gated_recurrent") { + dataLayer = creatDataLayer("layer_0", batchSize, layerSize * 3, false); + para = creatParameter("para_0", 0, layerSize * layerSize * 3, false); + bias = creatParameterBias("bias_0", 1, layerSize * 3, false); + } + + LayerPtr testLayer1 = initMKLPackedLayer( + layerConfig1, reversed, layerSize, dataLayer, para, bias); + LayerPtr testLayer2 = initMKLPackedLayer( + layerConfig2, reversed, layerSize, dataLayer, para, bias); + + const VectorPtr& weightGrad = + (testLayer1->getParameters()[0])->getBuf(PARAMETER_GRADIENT); + const MatrixPtr& inputGrad = testLayer1->getPrev(0)->getOutputGrad(); + CpuVector wgt_grad1(weightGrad->getSize()); + CpuVector wgt_grad2(weightGrad->getSize()); + CpuMatrix input_grad1(inputGrad->getHeight(), inputGrad->getWidth()); + CpuMatrix input_grad2(inputGrad->getHeight(), inputGrad->getWidth()); + + for (int i = 0; i < 2; i++) { + FLAGS_rnn_use_batch = useBatch1; + + testLayer1->forward(PASS_GC); + + FLAGS_rnn_use_batch = useBatch2; + testLayer2->forward(PASS_GC); + + testLayer1->getOutputGrad()->randomizeUniform(); + testLayer2->getOutputGrad()->copyFrom(*testLayer1->getOutputGrad()); + + weightGrad->zero(); + inputGrad->zero(); + FLAGS_rnn_use_batch = useBatch1; + testLayer1->backward(nullptr); + + wgt_grad1.copyFrom(*weightGrad); + input_grad1.copyFrom(*inputGrad); + + weightGrad->zero(); + inputGrad->zero(); + FLAGS_rnn_use_batch = useBatch2; + testLayer2->backward(nullptr); + + wgt_grad2.copyFrom(*weightGrad); + input_grad2.copyFrom(*inputGrad); + + checkError(*testLayer1->getOutputValue(), *testLayer2->getOutputValue()); + checkError(wgt_grad1, wgt_grad2); + checkError(input_grad1, input_grad2); + } +} + +TEST(MKLPackedLayer, RecurrentLayer) { + LayerConfig layerConfig1; + LayerConfig layerConfig2; + + layerConfig1.set_name("paddle-rnn"); + layerConfig1.set_type("recurrent"); + layerConfig1.set_active_type("relu"); + + layerConfig2.set_name("mkl-packed-rnn"); + layerConfig2.set_type("mkl_packed_recurrent"); + layerConfig2.set_active_type("relu"); + + FLAGS_use_gpu = false; + + for (auto layerSize : {32, 64, 128, 256, 512}) { + for (auto batchSize : {1, 5, 100, 500}) { + for (auto reversed : {true, false}) { + for (auto paddle_use_batch : {true, false}) { + for (auto MKLPacked_use_batch : {true, false}) { + LOG(INFO) << " layerSize=" << layerSize + << " batchSize=" << batchSize << " reversed=" << reversed + << " paddle_use_batch=" << paddle_use_batch + << " MKLPacked_use_batch=" << MKLPacked_use_batch; + + checkMKLPackedLayer(layerConfig1, + layerConfig2, + reversed, + layerSize, + batchSize, + paddle_use_batch, + MKLPacked_use_batch); + } + } + } + } + } +} +#endif + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + if (!version::isWithGpu()) { + testing::GTEST_FLAG(filter) = "-Layer.*"; + } + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_SelectiveFCLayer.cpp b/paddle/legacy/gserver/tests/test_SelectiveFCLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..1975d9196d61dbb80667b2ba86c09d56bc568064 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_SelectiveFCLayer.cpp @@ -0,0 +1,471 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/gserver/layers/FullyConnectedLayer.h" +#include "paddle/legacy/gserver/layers/Layer.h" +#include "paddle/legacy/gserver/layers/SelectiveFullyConnectedLayer.h" +#include "paddle/legacy/math/CpuSparseMatrix.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(num_passes); +DECLARE_string(config); +DECLARE_string(init_model_path); +DECLARE_string(config_args); + +size_t fcLayerWidth = 1024; + +struct ComData { + vector outArgs; + vector parameters; +}; + +int randint(int* data, size_t int_max, size_t size) { + srand((size_t)(time(NULL))); + if (int_max < size) { + return -1; + } + size_t count = 0; + std::map tmp; + int this_int = 0; + + while (count < size) { + this_int = std::rand() % int_max; // NOLINT + if (tmp.find(this_int) == tmp.end()) { + tmp[this_int] = 0; + count += 1; + } + } + + if (tmp.size() != size) { + return -1; + } + count = 0; + for (auto itr = tmp.begin(); itr != tmp.end(); ++itr) { + data[count] = itr->first; + count += 1; + } + return 0; +} + +void calcOutput(ComData& comData, + const string configFile, + const string configArgs, + bool useGpu) { + FLAGS_config = configFile; + FLAGS_config_args = configArgs; + FLAGS_use_gpu = useGpu; + FLAGS_init_model_path = "legacy/gserver/tests/SelectiveFcTest/model"; + *ThreadLocalRand::getSeed() = 0; + srand(0); + + Trainer trainer; + trainer.init(TrainerConfigHelper::createFromFlags(), false); + + comData.parameters = trainer.getGradientMachine()->getParameters(); + + auto dataProvider = trainer.getDataProvider(); + int32_t batchSize = trainer.getConfig().opt_config().batch_size(); + DataBatch dataBatch; + dataProvider->setSkipShuffle(); + dataProvider->reset(); + dataProvider->getNextBatch(batchSize, &dataBatch); + CHECK(dataBatch.getSize()) << "No data from data provider"; + + vector& inArgs = dataBatch.getStreams(); + trainer.getGradientMachine()->start(trainer.getConfig(), nullptr); + trainer.getGradientMachine()->forwardBackward( + inArgs, &comData.outArgs, PASS_TRAIN); + trainer.getGradientMachine()->finish(); +} + +void checkMatrix(real* A, real* B, size_t matSize) { +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + int diffNum = 0; + for (size_t i = 0; i < matSize; ++i) { + if (std::isinf(A[i]) || std::isnan(A[i]) || std::isinf(B[i]) || + std::isnan(B[i])) { + } else if (fabs(A[i] - B[i]) > err) { + diffNum++; + } + } + EXPECT_EQ(0, diffNum); +} + +void checkTranspose(real* matrix, + real* transpose, + size_t width, + size_t matSize) { +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + size_t height = matSize / width; + int diffNum = 0; + size_t rowId = 0; + size_t colId = 0; + for (size_t i = 0; i < matSize; ++i) { + if (i % width == 0 && i) { + rowId++; + } + colId = i % width; + if (fabs(matrix[i] - transpose[colId * height + rowId]) > err) { + diffNum++; + LOG(INFO) << i << " diff : " << matrix[i] << "\t" + << transpose[colId * height + rowId]; + } + } + EXPECT_EQ(0, diffNum); +} + +void compareOutput(ComData& fcData, ComData& selFcData) { + vector outArgsFc = fcData.outArgs; + vector outArgsSelfc = selFcData.outArgs; + + // check cost + LOG(INFO) << "Check cost"; + CpuMatrix fcCost(outArgsFc[0].value->getHeight(), + outArgsFc[0].value->getWidth()); + CpuMatrix selfcCost(outArgsSelfc[0].value->getHeight(), + outArgsSelfc[0].value->getWidth()); + fcCost.copyFrom(*outArgsFc[0].value); + selfcCost.copyFrom(*outArgsSelfc[0].value); + checkMatrix(fcCost.getData(), selfcCost.getData(), fcCost.getElementCnt()); + + // check selective fc output and fc output + LOG(INFO) << "Compare output of SelectiveFullyConectedLayer " + << "with FullyConectedLayer"; + CpuMatrix fcOut(outArgsFc[1].value->getHeight(), + outArgsFc[1].value->getWidth()); + CpuMatrix selfcOut(outArgsSelfc[1].value->getHeight(), + outArgsSelfc[1].value->getWidth()); + + fcOut.copyFrom(*outArgsFc[1].value); + selfcOut.copyFrom(*outArgsSelfc[1].value); + checkMatrix(fcOut.getData(), selfcOut.getData(), fcOut.getElementCnt()); + + // check gradient math + vector& fcParam = fcData.parameters; + vector& selfcParam = selFcData.parameters; + for (size_t i = 0; i < fcParam.size(); ++i) { + ParameterPtr p1, p2; + p1 = fcParam[i]; + p2 = selfcParam[i]; + + string paramName = p1->getName(); + LOG(INFO) << "check parameter : " << paramName; + + // check parameter value + CpuVector paraValue1(p1->getSize()); + CpuVector paraValue2(p2->getSize()); + paraValue1.copyFrom(*p1->getBuf(PARAMETER_VALUE)); + paraValue2.copyFrom(*p2->getBuf(PARAMETER_VALUE)); + + // check gradient + CpuVector paraGrad1(*p1->getBuf(PARAMETER_GRADIENT)); + CpuVector paraGrad2(*p2->getBuf(PARAMETER_GRADIENT)); + if (paramName == "rand_fc_param.bias") { + checkMatrix( + paraValue1.getData(), paraValue2.getData(), paraValue1.getSize()); + checkMatrix( + paraGrad1.getData(), paraGrad2.getData(), paraGrad1.getSize()); + } else { + checkTranspose(paraValue1.getData(), + paraValue2.getData(), + fcLayerWidth, + paraValue1.getSize()); + checkTranspose(paraGrad1.getData(), + paraGrad2.getData(), + fcLayerWidth, + paraGrad1.getSize()); + } + } +} + +void compareSparseMulOutput( + real* fcOutput, + real* selOutput, + size_t nnz, + const std::shared_ptr>>& selCols) { +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + size_t nnzCount = + std::accumulate(selCols->begin(), + selCols->end(), + 0UL, + [](size_t a, const std::pair& arr) { + return a + arr.second; + }); + EXPECT_EQ(nnz, nnzCount); + + size_t sampleNum = selCols->size(); + int diffNum = 0; + size_t count = 0; + for (size_t i = 0; i < sampleNum; ++i) { + for (size_t j = 0; j < (*selCols)[i].second; ++j) { + size_t selIdx = (*selCols)[i].first[j]; + if (fabs(fcOutput[i * fcLayerWidth + selIdx] - selOutput[count]) > err) { + diffNum++; + LOG(INFO) << count << " diff : " << fcOutput[i * fcLayerWidth + selIdx] + << "\t" << selOutput[count]; + } + count++; + } + } + EXPECT_EQ(0, diffNum); +} + +LayerPtr creatDataLayer(string name, + size_t batchSize, + size_t layerSize, + std::vector& values, + bool useGpu) { + LayerConfig dataConfig; + dataConfig.set_name(name); + dataConfig.set_type("data"); + dataConfig.set_size(layerSize); + LayerPtr layer = LayerPtr(new DataLayer(dataConfig)); + + Argument data; + data.value = Matrix::create(batchSize, layerSize, false, useGpu); + data.value->copyFrom(values.data(), batchSize * layerSize); + + DataLayerPtr dataLayer = std::dynamic_pointer_cast(layer); + dataLayer->setData(data); + dataLayer->forward(PASS_TEST); + return layer; +} + +ParameterPtr creatParameter( + string name, int pid, size_t paraSize, string paramFile, bool useGpu) { + ParameterConfig paraConfig; + paraConfig.set_name(name); + paraConfig.set_size(paraSize); + + ParameterPtr parameter = + std::make_shared(paraConfig, useGpu, /*initialize */ false); + parameter->enableType(PARAMETER_VALUE); + parameter->randomize(); + parameter->setID(pid); + parameter->load(paramFile); + return parameter; +} + +LayerPtr initFcLayer(LayerPtr dataLayer, + LayerConfig layerConfig, + int dataLayerSize, + int fcLayerSize, + string paraName, + string paraFile, + bool useGpu) { + LayerMap layerMap; + ParameterMap parameterMap; + + layerMap[dataLayer->getName()] = dataLayer; + ParameterPtr para = creatParameter( + paraName, 0, dataLayerSize * fcLayerSize, paraFile, useGpu); + parameterMap[para->getName()] = para; + + layerConfig.add_inputs(); + LayerInputConfig& input = *(layerConfig.mutable_inputs(0)); + input.set_input_layer_name(dataLayer->getName()); + input.set_input_parameter_name(paraName); + + LayerPtr testLayer = Layer::create(layerConfig); + layerMap[testLayer->getName()] = testLayer; + + testLayer->setNeedGradient(false); + testLayer->init(layerMap, parameterMap); + return testLayer; +} + +#ifndef PADDLE_TYPE_DOUBLE +// The parameter file used in fc.conf and selective_fc.conf is float +TEST(Layer, SelectiveFcLayer_train_dense_mul) { + const string& fcConfig = "legacy/gserver/tests/SelectiveFcTest/conf/fc.conf"; + const string& fcConfigArgs = + "filelist=legacy/gserver/tests/SelectiveFcTest/dense_mul_list"; + const string& selFcConfig = + "legacy/gserver/tests/SelectiveFcTest/conf/selective_fc.conf"; + const string& selConfigArgs = + "filelist=legacy/gserver/tests/SelectiveFcTest/dense_mul_list"; + + for (auto useGpu : {false, true}) { +#ifndef PADDLE_WITH_CUDA + if (useGpu) { + break; + } +#endif + LOG(INFO) << "FullyConnectedLayer forwardBackward()"; + ComData fcData; + calcOutput(fcData, fcConfig, fcConfigArgs, useGpu); + + LOG(INFO) << "SelectiveFullyConnectedLayer forwardBackward()"; + ComData selFcData; + calcOutput(selFcData, selFcConfig, selConfigArgs, useGpu); + compareOutput(fcData, selFcData); + } +} +#endif // PADDLE_TYPE_DOUBLE + +void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config, + bool useGpu) { + FLAGS_use_gpu = useGpu; + size_t batchSize = 100; + size_t dataLayerSize = 512; + std::vector values(batchSize * dataLayerSize); + for (size_t j = 0; j < batchSize * dataLayerSize; ++j) { + values[j] = std::rand() / real(RAND_MAX); + } + LayerPtr dataLayer = + creatDataLayer("data", batchSize, dataLayerSize, values, useGpu); + + const string& selfcParaFile = + "legacy/gserver/tests/SelectiveFcTest/model/rand_fc_param.w.transpose"; + const string& selfcParaName = "rand_fc_param.w.transpose"; + + std::shared_ptr selfcLayer = + std::dynamic_pointer_cast( + initFcLayer(dataLayer, + config, + dataLayerSize, + fcLayerWidth, + selfcParaName, + selfcParaFile, + useGpu)); + + // create selected columns + std::shared_ptr>> selCols( + new std::vector>(batchSize)); + size_t maxNNZ = 30; + srand((size_t)(time(NULL))); + int total = 0; + while (total == 0) { + for (size_t i = 0; i < batchSize; ++i) { + size_t num = std::rand() % maxNNZ; + int* data = new int[num]; + randint(data, fcLayerWidth, num); + (*selCols)[i] = std::make_pair(data, num); + total += num; + } + } + selfcLayer->fillSelectiveData(selCols); + selfcLayer->forward(PASS_TEST); + + MatrixPtr outMatSelfc = selfcLayer->getOutputValue(); + CpuSparseMatrixPtr cpuOutMatSelfc( + new CpuSparseMatrix(outMatSelfc->getHeight(), + outMatSelfc->getWidth(), + outMatSelfc->getElementCnt())); + cpuOutMatSelfc->copyFrom(*outMatSelfc, HPPL_STREAM_DEFAULT); +#ifdef PADDLE_WITH_CUDA + if (useGpu) { + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + } +#endif + real* outValueSelfc = cpuOutMatSelfc->getValue(); + size_t nnz = cpuOutMatSelfc->getElementCnt(); + + const string& fcParaFile = + "legacy/gserver/tests/SelectiveFcTest/model/rand_fc_param.w"; + const string& fcParaName = "rand_fc_param.w"; + LayerConfig fcLayerConfig; + fcLayerConfig.set_name("fc_layer"); + fcLayerConfig.set_type("fc"); + fcLayerConfig.set_active_type("linear"); + fcLayerConfig.set_size(fcLayerWidth); + + LayerPtr fcLayer = initFcLayer(dataLayer, + fcLayerConfig, + dataLayerSize, + fcLayerWidth, + fcParaName, + fcParaFile, + useGpu); + fcLayer->forward(PASS_TEST); + + MatrixPtr outMatFc = fcLayer->getOutputValue(); + MatrixPtr cpuOutMatFc( + new CpuMatrix(outMatFc->getHeight(), outMatFc->getWidth())); + cpuOutMatFc->copyFrom(*outMatFc, HPPL_STREAM_DEFAULT); +#ifdef PADDLE_WITH_CUDA + if (useGpu) { + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + } +#endif + real* outValueFc = cpuOutMatFc->getData(); + + compareSparseMulOutput(outValueFc, outValueSelfc, nnz, selCols); + for (size_t i = 0; i < batchSize; ++i) { + delete[](*selCols)[i].first; + } +} + +#ifndef PADDLE_TYPE_DOUBLE +// The parameter file used in testSelectiveFcLayerTrainSparseMul is float +TEST(Layer, SelectiveFcLayer_train_sparse_mul) { + LayerConfig selLayerConfig; + selLayerConfig.set_name("sel_fc"); + selLayerConfig.set_type("selective_fc"); + selLayerConfig.set_active_type("linear"); + selLayerConfig.set_has_selected_colums(false); + selLayerConfig.set_selective_fc_pass_generation(true); + selLayerConfig.set_size(fcLayerWidth); + + testSelectiveFcLayerTrainSparseMul(selLayerConfig, false); +#ifdef PADDLE_WITH_CUDA + testSelectiveFcLayerTrainSparseMul(selLayerConfig, true); +#endif +} +#endif // PADDLE_TYPE_DOUBLE + +// TODO(dangqingqing) test multi threads after support in matrix +// TEST(Layer, SelectiveFcLayer_train_sparse_mul_parallel) { +// LayerConfig selLayerConfig; +// selLayerConfig.set_name("sel_fc"); +// selLayerConfig.set_type("selective_fc"); +// selLayerConfig.set_active_type("linear"); +// selLayerConfig.set_has_selected_colums(false); +// selLayerConfig.set_selective_fc_pass_generation(true); +// selLayerConfig.set_selective_fc_parallel_plain_mul_thread_num(10); +// selLayerConfig.set_selective_fc_full_mul_ratio(1000); +// selLayerConfig.set_size(fcLayerWidth); +// SelectiveFcLayer_test(selLayerConfig, false); +// } + +int main(int argc, char** argv) { + paddle::initMain(argc, argv); + testing::InitGoogleTest(&argc, argv); + initPython(argc, argv); + int ret = RUN_ALL_TESTS(); + return ret; +} diff --git a/paddle/legacy/gserver/tests/test_SeqSliceLayerGrad.cpp b/paddle/legacy/gserver/tests/test_SeqSliceLayerGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..05acd714219fa5964b5b3595543682825ea67d84 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_SeqSliceLayerGrad.cpp @@ -0,0 +1,224 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +const int MAX_SEQ_NUM = 17; +const int MAX_SEQ_LEN = 23; +const int MAX_BEAM_SIZE = 13; + +const size_t SEED = (size_t)(time(NULL)); + +vector randSampling(real range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0.); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + sort(begin(num), end(num)); + return num; +} + +void genSeqInfo(vector& seqStartPos, vector& subSeqStartPos) { + seqStartPos.resize(1, 0); + subSeqStartPos.resize(1, 0); + + srand(SEED); + int seqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int i = 0; i < seqNum; ++i) { + int subSeqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int j = 0; j < subSeqNum; ++j) + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + seqStartPos.push_back(subSeqStartPos.back()); + } +} + +/* + generate start indices according to sequence start positions. + */ +void genStarts(vector& seqStartPos, + vector>& starts, + size_t beamSize) { + starts.clear(); + starts.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); + + for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + vector randStarts = + randSampling(seqLen, min(seqLen, static_cast(beamSize))); + copy(begin(randStarts), end(randStarts), begin(starts[i])); + } +} + +/* + generate end indices according to sequence start positions and start indices. + */ +void genEnds(vector& seqStartPos, + vector>& starts, + vector>& ends, + size_t beamSize) { + CHECK_EQ(seqStartPos.size() - 1, starts.size()); + ends.clear(); + ends.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); + + for (size_t i = 0; i < starts.size(); ++i) { + for (size_t j = 0; j < starts[i].size(); ++j) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + CHECK_GE(seqLen - 1, starts[i][j]); + if (starts[i][j] == -1.) break; + if (starts[i][j] == (seqLen - 1)) { + ends[i][j] = starts[i][j]; + } else { + ends[i][j] = starts[i][j] + randSampling(seqLen - starts[i][j], 1)[0]; + } + } + } +} + +void genTestData(vector& seqStartPos, + vector& subSeqStartPos, + vector>& starts, + vector>& ends, + bool hasSubseq) { + size_t beamSize = 1 + (rand() % MAX_BEAM_SIZE); + genSeqInfo(seqStartPos, subSeqStartPos); + + genStarts(hasSubseq ? subSeqStartPos : seqStartPos, starts, beamSize); + genEnds(hasSubseq ? subSeqStartPos : seqStartPos, starts, ends, beamSize); +} + +template +void flatten2dVector(vector>& inVec, vector& outVec) { + size_t totalSize{0}; + for (auto const& items : inVec) totalSize += items.size(); + outVec.reserve(totalSize); + + for (auto& items : inVec) + move(items.begin(), items.end(), back_inserter(outVec)); +} + +void testSeqSliceLayer(bool hasSubseq, + bool useGpu, + vector& seqStartPos, + vector& subSeqStartPos, + vector>& starts, + vector>& ends) { + // layer size is not crutial for this layer, + // so here use a small layer size in the unittest. + const size_t layerSize{4}; + TestConfig config; + config.layerConfig.set_type("seq_slice"); + config.layerConfig.set_size(layerSize); + + // add the first input + MatrixPtr seqInputPtr = + Matrix::create(hasSubseq ? subSeqStartPos.back() : seqStartPos.back(), + layerSize, + false, + false); + seqInputPtr->randomizeUniform(); + + if (hasSubseq) { + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "seq_input", + seqInputPtr, + seqStartPos, + subSeqStartPos}); + } else { + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "seq_input", seqInputPtr, seqStartPos}); + } + config.layerConfig.add_inputs(); + + // add start indices + if (starts.size()) { + vector startsToVec; + flatten2dVector(starts, startsToVec); + + MatrixPtr startMatrixPtr = + Matrix::create(starts.size(), starts[0].size(), false, false); + startMatrixPtr->copyFrom(startsToVec.data(), startsToVec.size()); + + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "starts", startMatrixPtr}); + config.layerConfig.add_inputs(); + config.layerConfig.set_select_first(true); + } + + // add end indices + if (ends.size()) { + vector endsToVec; + flatten2dVector(ends, endsToVec); + + MatrixPtr endMatrixPtr = + Matrix::create(ends.size(), ends[0].size(), false, false); + endMatrixPtr->copyFrom(endsToVec.data(), endsToVec.size()); + + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "ends", endMatrixPtr}); + config.layerConfig.add_inputs(); + config.layerConfig.set_select_first(false); + } + + testLayerGrad(config, "seq_slice", /*batchSize*/ 100, false, useGpu, false); +} + +TEST(Layer, SeqSliceLayer) { + vector seqStartPos; + vector subSeqStartPos; + vector> starts; + vector> ends; + + std::vector mode = {false}; +#ifdef PADDLE_WITH_CUDA + mode.push_back(true); +#endif + genSeqInfo(seqStartPos, subSeqStartPos); + for (bool hasSubseq : {true, false}) { + LOG(INFO) << "hasSubSeq : " << hasSubseq; + genTestData(seqStartPos, subSeqStartPos, starts, ends, hasSubseq); + for (bool useGpu : mode) { + vector> tmp; + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, tmp, ends); + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, tmp); + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, ends); + } + } +} + +int main(int argc, char** argv) { + initMain(argc, argv); + hl_start(); + hl_init(FLAGS_gpu_id); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/gserver/tests/test_Upsample.cpp b/paddle/legacy/gserver/tests/test_Upsample.cpp new file mode 100644 index 0000000000000000000000000000000000000000..940d46baf73f2d600cff6edc37c29a3a36bf5d90 --- /dev/null +++ b/paddle/legacy/gserver/tests/test_Upsample.cpp @@ -0,0 +1,153 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include + +#include "LayerGradUtil.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/testing/TestUtil.h" + +void setPoolConfig(paddle::TestConfig* config, + paddle::PoolConfig* pool, + const string& poolType) { + (*config).biasSize = 0; + (*config).layerConfig.set_type("pool"); + (*config).layerConfig.set_num_filters(1); + + int kw = 2, kh = 2; + int pw = 0, ph = 0; + int sw = 2, sh = 2; + pool->set_pool_type(poolType); + pool->set_channels(2); + pool->set_size_x(kw); + pool->set_size_y(kh); + pool->set_start(0); + pool->set_padding(pw); + pool->set_padding_y(ph); + pool->set_stride(sw); + pool->set_stride_y(sh); + + int ow = + paddle::outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); + int oh = + paddle::outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); + pool->set_output_x(ow); + pool->set_output_y(oh); +} + +paddle::LayerPtr doOneUpsampleTest(const paddle::MatrixPtr& inputMat, + const string& poolType, + bool use_gpu, + real* tempGradData) { + /* prepare maxPoolWithMaskLayer */ + paddle::TestConfig config; + config.inputDefs.push_back({paddle::INPUT_DATA, "layer_0", 128, 0}); + paddle::LayerInputConfig* input = config.layerConfig.add_inputs(); + paddle::PoolConfig* pool = input->mutable_pool_conf(); + + pool->set_img_size(8); + pool->set_img_size_y(8); + setPoolConfig(&config, pool, "max-pool-with-mask"); + config.layerConfig.set_size(pool->output_x() * pool->output_y() * + pool->channels()); + + config.layerConfig.set_name("MaxPoolWithMask"); + + std::vector dataLayers; + paddle::LayerMap layerMap; + vector datas; + + initDataLayer(config, + &dataLayers, + &datas, + &layerMap, + "MaxPoolWithMask", + 1, + false, + use_gpu); + + dataLayers[0]->getOutputValue()->copyFrom(*inputMat); + + FLAGS_use_gpu = use_gpu; + std::vector parameters; + paddle::LayerPtr maxPoolingWithMaskOutputLayer; + initTestLayer(config, &layerMap, ¶meters, &maxPoolingWithMaskOutputLayer); + maxPoolingWithMaskOutputLayer->forward(paddle::PASS_GC); + + /* prepare the upsample layer */ + paddle::LayerConfig upsampleLayerConfig; + upsampleLayerConfig.set_type("upsample"); + paddle::LayerInputConfig* input1 = upsampleLayerConfig.add_inputs(); + upsampleLayerConfig.add_inputs(); + + paddle::UpsampleConfig* upsampleConfig = input1->mutable_upsample_conf(); + upsampleConfig->set_scale(2); + paddle::ImageConfig* imageConfig = upsampleConfig->mutable_image_conf(); + imageConfig->set_channels(2); + imageConfig->set_img_size(4); + imageConfig->set_img_size_y(4); + upsampleLayerConfig.set_size(2 * 8 * 8); + upsampleLayerConfig.set_name("upsample"); + + for (size_t i = 0; i < 2; i++) { + paddle::LayerInputConfig& inputTemp = + *(upsampleLayerConfig.mutable_inputs(i)); + inputTemp.set_input_layer_name("MaxPoolWithMask"); + } + + paddle::LayerPtr upsampleLayer; + paddle::ParameterMap parameterMap; + upsampleLayer = paddle::Layer::create(upsampleLayerConfig); + layerMap[upsampleLayerConfig.name()] = upsampleLayer; + upsampleLayer->init(layerMap, parameterMap); + upsampleLayer->setNeedGradient(true); + upsampleLayer->forward(paddle::PASS_GC); + upsampleLayer->getOutputGrad()->copyFrom(tempGradData, 128); + upsampleLayer->backward(); + + return upsampleLayer; +} + +TEST(Layer, maxPoolingWithMaskOutputLayerFwd) { + bool useGpu = false; + paddle::MatrixPtr inputMat; + paddle::MatrixPtr inputGPUMat; + paddle::MatrixPtr tempGradMat; + + inputMat = paddle::Matrix::create(1, 128, false, useGpu); + inputMat->randomizeUniform(); + + tempGradMat = paddle::Matrix::create(1, 128, false, useGpu); + tempGradMat->randomizeUniform(); + real* tempGradData = tempGradMat->getData(); + + paddle::LayerPtr upsampleLayerCPU = + doOneUpsampleTest(inputMat, "max-pool-with-mask", useGpu, tempGradData); + +#ifdef PADDLE_WITH_CUDA + useGpu = true; + real* data = inputMat->getData(); + inputGPUMat = paddle::Matrix::create(1, 128, false, useGpu); + inputGPUMat->copyFrom(data, 128); + paddle::LayerPtr upsampleLayerGPU = doOneUpsampleTest( + inputGPUMat, "max-pool-with-mask", useGpu, tempGradData); + paddle::checkMatrixEqual(upsampleLayerCPU->getOutput("").value, + upsampleLayerGPU->getOutput("").value); + + paddle::checkMatrixEqual(upsampleLayerCPU->getPrev(0)->getOutputGrad(), + upsampleLayerGPU->getPrev(0)->getOutputGrad()); +#endif +} diff --git a/paddle/legacy/gserver/tests/test_WarpCTCLayer.cpp b/paddle/legacy/gserver/tests/test_WarpCTCLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b1697e1616484ec5389cdb5b59ba413a9615cf2e --- /dev/null +++ b/paddle/legacy/gserver/tests/test_WarpCTCLayer.cpp @@ -0,0 +1,244 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/legacy/gserver/layers/CTCLayer.h" +#include "paddle/legacy/gserver/layers/DataLayer.h" +#include "paddle/legacy/gserver/layers/Layer.h" +#include "paddle/legacy/gserver/layers/WarpCTCLayer.h" + +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); + +const real* getData(const Matrix& matrix) { + if (matrix.useGpu()) { + MatrixPtr cpuMatrix = Matrix::create( + matrix.getHeight(), matrix.getWidth(), matrix.isTransposed(), false); + cpuMatrix->copyFrom(matrix); + return cpuMatrix->getData(); + } else { + return matrix.getData(); + } +} + +int checkError(const Matrix& matrix1, const Matrix& matrix2) { + CHECK_EQ(matrix1.getHeight(), matrix2.getHeight()); + CHECK_EQ(matrix1.getWidth(), matrix2.getWidth()); + CHECK_EQ(matrix1.isTransposed(), matrix2.isTransposed()); +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + + int height = matrix1.getHeight(); + int width = matrix1.getWidth(); + + const real* data1 = getData(matrix1); + const real* data2 = getData(matrix2); + int count = 0; + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + if (fabs(data1[i * width + j] - data2[i * width + j]) > err) { + count++; + } + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; + return count; +} + +void initArgument(size_t batchSize, + int layerSize, + bool useGpu, + Argument& data) { + data.value = Matrix::create(batchSize, layerSize, false, useGpu); + data.grad = Matrix::create(batchSize, layerSize, false, useGpu); + data.value->randomizeUniform(); + data.value->add(-0.5); + data.grad->zeroMem(); + + generateSequenceStartPositions(batchSize, data.sequenceStartPositions); +} + +LayerPtr createDataLayer( + string name, size_t batchSize, int layerSize, bool useGpu, Argument& data) { + LayerConfig layerConfig; + layerConfig.set_name(name); + layerConfig.set_type("data"); + layerConfig.set_size(layerSize); + LayerPtr layer = LayerPtr(new DataLayer(layerConfig)); + + DataLayerPtr dataLayer = std::dynamic_pointer_cast(layer); + dataLayer->setData(data); + dataLayer->forward(PASS_GC); + + return layer; +} + +LayerPtr createLabelLayer(string name, + size_t batchSize, + size_t numClasses, + bool useGpu) { + LayerConfig layerConfig; + layerConfig.set_name(name); + layerConfig.set_type("data"); + layerConfig.set_size(1); + LayerPtr layer = LayerPtr(new DataLayer(layerConfig)); + + Argument data; + data.ids = IVector::create(batchSize, useGpu); + data.ids->rand(numClasses - 1); + + generateSequenceStartPositions(batchSize, data.sequenceStartPositions); + + DataLayerPtr labelLayer = std::dynamic_pointer_cast(layer); + labelLayer->setData(data); + labelLayer->forward(PASS_GC); + + return layer; +} + +LayerPtr createCTCLayer(string name, + size_t numClasses, + bool useGpu, + bool normByTimes, + LayerPtr dataLayer, + LayerPtr labelLayer) { + LayerMap layerMap; + layerMap[dataLayer->getName()] = dataLayer; + layerMap[labelLayer->getName()] = labelLayer; + + ParameterMap parameterMap; + + LayerConfig layerConfig; + layerConfig.set_name(name); + layerConfig.set_type("ctc"); + layerConfig.set_size(numClasses); + layerConfig.set_norm_by_times(normByTimes); + + layerConfig.add_inputs(); + LayerInputConfig& input0 = *(layerConfig.mutable_inputs(0)); + input0.set_input_layer_name(dataLayer->getName()); + + layerConfig.add_inputs(); + LayerInputConfig& input1 = *(layerConfig.mutable_inputs(1)); + input1.set_input_layer_name(labelLayer->getName()); + + LayerPtr layer = LayerPtr(new CTCLayer(layerConfig)); + layerMap[layer->getName()] = layer; + layer->init(layerMap, parameterMap); + + ActivationFunction* softmaxActivation = ActivationFunction::create("softmax"); + + softmaxActivation->forward(dataLayer->getOutput()).check(); + layer->forward(PASS_GC); + + layer->backward(); + softmaxActivation->backward(dataLayer->getOutput()).check(); + + return layer; +} + +LayerPtr createWarpCTCLayer(string name, + size_t numClasses, + bool useGpu, + bool normByTimes, + LayerPtr dataLayer, + LayerPtr labelLayer) { + LayerMap layerMap; + layerMap[dataLayer->getName()] = dataLayer; + layerMap[labelLayer->getName()] = labelLayer; + + ParameterMap parameterMap; + + LayerConfig layerConfig; + layerConfig.set_name(name); + layerConfig.set_type("warp_ctc"); + layerConfig.set_size(numClasses); + layerConfig.set_blank(numClasses - 1); + layerConfig.set_norm_by_times(normByTimes); + + layerConfig.add_inputs(); + LayerInputConfig& input0 = *(layerConfig.mutable_inputs(0)); + input0.set_input_layer_name(dataLayer->getName()); + + layerConfig.add_inputs(); + LayerInputConfig& input1 = *(layerConfig.mutable_inputs(1)); + input1.set_input_layer_name(labelLayer->getName()); + + LayerPtr layer = LayerPtr(new WarpCTCLayer(layerConfig)); + layerMap[layer->getName()] = layer; + layer->init(layerMap, parameterMap); + + layer->forward(PASS_GC); + layer->backward(); + + return layer; +} + +TEST(Layer, WarpCTCLayer) { + for (auto layerSize : {10, 64}) { + for (auto batchSize : {1, 10, 32}) { + for (auto normByTimes : {false, true}) { + for (auto useGpu : {false, true}) { +#ifndef PADDLE_WITH_CUDA + if (useGpu) continue; +#endif + LOG(INFO) << "layerSize=" << layerSize << " batchSize=" << batchSize + << " normByTimes = " << normByTimes << " useGpu=" << useGpu; + + FLAGS_use_gpu = useGpu; + + Argument data0; + initArgument(batchSize, layerSize, useGpu, data0); + + Argument data1; + data1.resizeAndCopyFrom(data0); + + LayerPtr dataLayer0 = + createDataLayer("data", batchSize, layerSize, useGpu, data0); + LayerPtr dataLayer1 = + createDataLayer("data", batchSize, layerSize, useGpu, data1); + + LayerPtr labelLayer = + createLabelLayer("label", batchSize, layerSize, useGpu); + + LayerPtr warpctcLayer = createWarpCTCLayer( + "cost", layerSize, useGpu, normByTimes, dataLayer0, labelLayer); + LayerPtr ctcLayer = createCTCLayer( + "cost", layerSize, useGpu, normByTimes, dataLayer1, labelLayer); + + /// Check cost + LOG(INFO) << "Check cost: " + << checkError(*(warpctcLayer->getOutput().value), + *(ctcLayer->getOutput().value)) + << " different elements."; + + /// Check gradients + LOG(INFO) << "Check gradients: " + << checkError(*(dataLayer0->getOutput().grad), + *(dataLayer1->getOutput().grad)) + << " different elements"; + } + } + } + } +} diff --git a/paddle/legacy/math/Allocator.h b/paddle/legacy/math/Allocator.h new file mode 100644 index 0000000000000000000000000000000000000000..ffb5ec1cad4113c2035daad8c385bbe57a161079 --- /dev/null +++ b/paddle/legacy/math/Allocator.h @@ -0,0 +1,137 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "hl_gpu.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +/** + * @brief Allocator base class. + * + * This is the base class of all Allocator class. + */ +class Allocator { + public: + virtual ~Allocator() {} + virtual void* alloc(size_t size) = 0; + virtual void free(void* ptr) = 0; + virtual std::string getName() = 0; +}; + +/** + * @brief CPU allocator implementation. + */ +class CpuAllocator : public Allocator { + public: + ~CpuAllocator() {} + + /** + * @brief Aligned allocation on CPU. + * @param size Size to be allocated. + * @return Pointer to the allocated memory + */ + virtual void* alloc(size_t size) { + void* ptr; +#ifdef PADDLE_WITH_MKLDNN + // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp + // memory alignment + CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0); +#else + CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0); +#endif + CHECK(ptr) << "Fail to allocate CPU memory: size=" << size; + return ptr; + } + + /** + * @brief Free the memory space. + * @param ptr Pointer to be free. + */ + virtual void free(void* ptr) { + if (ptr) { + ::free(ptr); + } + } + + virtual std::string getName() { return "cpu_alloc"; } +}; + +/** + * @brief GPU allocator implementation. + */ +class GpuAllocator : public Allocator { + public: + ~GpuAllocator() {} + + /** + * @brief Allocate GPU memory. + * @param size Size to be allocated. + * @return Pointer to the allocated memory + */ + virtual void* alloc(size_t size) { + void* ptr = hl_malloc_device(size); + CHECK(ptr) << "Fail to allocate GPU memory " << size << " bytes"; + return ptr; + } + + /** + * @brief Free the GPU memory. + * @param ptr Pointer to be free. + */ + virtual void free(void* ptr) { + if (ptr) { + hl_free_mem_device(ptr); + } + } + + virtual std::string getName() { return "gpu_alloc"; } +}; + +/** + * @brief CPU pinned memory allocator implementation. + */ +class CudaHostAllocator : public Allocator { + public: + ~CudaHostAllocator() {} + + /** + * @brief Allocate pinned memory. + * @param size Size to be allocated. + * @return Pointer to the allocated memory + */ + virtual void* alloc(size_t size) { + void* ptr = hl_malloc_host(size); + CHECK(ptr) << "Fail to allocate pinned memory " << size << " bytes"; + return ptr; + } + + /** + * @brief Free the pinned memory. + * @param ptr Pointer to be free. + */ + virtual void free(void* ptr) { + if (ptr) { + hl_free_mem_host(ptr); + } + } + + virtual std::string getName() { return "cuda_host_alloc"; } +}; + +} // namespace paddle diff --git a/paddle/legacy/math/BaseMatrix.cu b/paddle/legacy/math/BaseMatrix.cu new file mode 100644 index 0000000000000000000000000000000000000000..7e7cdc57a9887152ecd9e0bbd9fe14fcba56799d --- /dev/null +++ b/paddle/legacy/math/BaseMatrix.cu @@ -0,0 +1,1953 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include "BaseMatrix.h" +#include "MathFunctions.h" +#include "NEONFunctions.h" +#include "SIMDFunctions.h" +#include "hl_matrix_apply.cuh" +#include "hl_matrix_base.cuh" +#include "hl_matrix_ops.cuh" + +namespace paddle { + +const char* SPARSE_SUPPORT_ERROR = "Sparse Matrix/Vector is not supported."; + +template +template +int BaseMatrixT::applyUnary(Op op) { + MatrixOffset offset(0, 0); + applyUnary(op, height_, width_, offset); + return 0; +} + +template +template +int BaseMatrixT::applyUnary(Op op, + int numRows, + int numCols, + MatrixOffset& offset) { + CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; + int dimM = numRows; + int dimN = numCols; + int lda = stride_; + + T* A = data_; + CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); + + CHECK_LE(dimM + offset.aRow_, this->height_); + CHECK_LE(dimN + offset.aCol_, this->width_); + if (true == useGpu_) { + hl_gpu_apply_unary_op(op, A, dimM, dimN, lda); + } else { + hl_cpu_apply_unary_op(op, A, dimM, dimN, lda); + } + return 0; +} + +template +template +int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b) { + CHECK(height_ == b.height_ && width_ == b.width_) + << "Matrix dimensions are not equal"; + + MatrixOffset offset(0, 0, 0, 0); + applyBinary(op, b, height_, width_, offset); + return 0; +} + +template +template +int BaseMatrixT::applyBinary( + Op op, BaseMatrixT& b, int numRows, int numCols, MatrixOffset& offset) { + applyBinary(op, b, numRows, numCols, offset, false_type(), false_type()); + return 0; +} + +template +template +int BaseMatrixT::applyBinary(Op op, + BaseMatrixT& b, + int numRows, + int numCols, + MatrixOffset& offset, + bAsRowVector, + bAsColVector) { + CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(useGpu_ == b.useGpu_) << "Matrix type mismatch"; + + int dimM = numRows; + int dimN = numCols; + int lda = stride_; + int ldb = b.stride_; + + T* A = data_; + T* B = b.data_; + CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CHECK_LE(dimM + offset.aRow_, this->height_); + CHECK_LE(dimN + offset.aCol_, this->width_); + if (!bAsRowVector::value && !bAsColVector::value) { + CHECK_LE(dimM + offset.bRow_, b.height_); + CHECK_LE(dimN + offset.bCol_, b.width_); + } else if (bAsRowVector::value && !bAsColVector::value) { + CHECK_LE(dimN + offset.bCol_, b.width_); + } else if (!bAsRowVector::value && bAsColVector::value) { + CHECK_LE(dimM + offset.bRow_, b.height_); + } else { + } + if (true == useGpu_) { + hl_gpu_apply_binary_op( + op, A, B, dimM, dimN, lda, ldb); + } else { + hl_cpu_apply_binary_op( + op, A, B, dimM, dimN, lda, ldb); + } + + return 0; +} + +template +template +int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c) { + CHECK_EQ(height_, b.height_); + CHECK_EQ(width_, b.width_); + CHECK_EQ(height_, c.height_); + CHECK_EQ(width_, c.width_); + + MatrixOffset offset(0, 0, 0, 0, 0, 0); + applyTernary(op, b, c, height_, width_, offset); + + return 0; +} + +template +template +int BaseMatrixT::applyTernary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset) { + applyTernary(op, b, c, numRows, numCols, offset, false_type(), false_type()); + + return 0; +} + +template +template +int BaseMatrixT::applyTernary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset, + cAsRowVector, + cAsColVector) { + CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(!c.isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK_EQ(useGpu_, b.useGpu_); + CHECK_EQ(useGpu_, c.useGpu_); + + int dimM = numRows; + int dimN = numCols; + int lda = stride_; + int ldb = b.stride_; + int ldc = c.stride_; + + T* A = data_; + T* B = b.data_; + T* C = c.data_; + CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CAL_MATRIX_START_ADDRESS( + C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); + + CHECK_LE(dimM + offset.aRow_, this->height_); + CHECK_LE(dimN + offset.aCol_, this->width_); + CHECK_LE(dimM + offset.bRow_, b.height_); + CHECK_LE(dimN + offset.bCol_, b.width_); + if (!cAsRowVector::value && !cAsColVector::value) { + CHECK_LE(dimM + offset.cRow_, c.height_); + CHECK_LE(dimN + offset.cCol_, c.width_); + } else if (cAsRowVector::value && !cAsColVector::value) { + CHECK_LE(dimN + offset.cCol_, c.width_); + } else if (!cAsRowVector::value && cAsColVector::value) { + CHECK_LE(dimM + offset.cRow_, c.height_); + } else { + } + + if (true == useGpu_) { + hl_gpu_apply_ternary_op( + op, A, B, C, dimM, dimN, lda, ldb, ldc); + } else { + hl_cpu_apply_ternary_op( + op, A, B, C, dimM, dimN, lda, ldb, ldc); + } + + return 0; +} + +template +template +int BaseMatrixT::applyQuaternary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d) { + CHECK_EQ(height_, b.height_); + CHECK_EQ(width_, b.width_); + CHECK_EQ(height_, c.height_); + CHECK_EQ(width_, c.width_); + CHECK_EQ(height_, d.height_); + CHECK_EQ(width_, d.width_); + + MatrixOffset offset(0, 0, 0, 0, 0, 0, 0, 0); + applyQuaternary(op, b, c, d, height_, width_, offset); + + return 0; +} + +template +template +int BaseMatrixT::applyQuaternary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d, + int numRows, + int numCols, + MatrixOffset& offset) { + CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(!c.isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK(!d.isSparse()) << SPARSE_SUPPORT_ERROR; + CHECK_EQ(useGpu_, b.useGpu_); + CHECK_EQ(useGpu_, c.useGpu_); + CHECK_EQ(useGpu_, d.useGpu_); + + int dimM = numRows; + int dimN = numCols; + int lda = stride_; + int ldb = b.stride_; + int ldc = c.stride_; + int ldd = d.stride_; + + T* A = data_; + T* B = b.data_; + T* C = c.data_; + T* D = d.data_; + CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CAL_MATRIX_START_ADDRESS( + C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); + CAL_MATRIX_START_ADDRESS( + D, d.height_, d.width_, ldd, offset.dCol_, offset.dRow_); + + CHECK_LE(dimM + offset.aRow_, this->height_); + CHECK_LE(dimN + offset.aCol_, this->width_); + CHECK_LE(dimM + offset.bRow_, b.height_); + CHECK_LE(dimN + offset.bCol_, b.width_); + CHECK_LE(dimM + offset.cRow_, c.height_); + CHECK_LE(dimN + offset.cCol_, c.width_); + CHECK_LE(dimM + offset.dRow_, d.height_); + CHECK_LE(dimN + offset.dCol_, d.width_); + if (true == useGpu_) { + hl_gpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, ldc, ldd); + } else { + hl_cpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, ldc, ldd); + } + + return 0; +} + +template +template +int BaseMatrixT::aggregate(Agg agg, + Op op, + Saver sv, + BaseMatrixT& b, + int numRows, + int numCols, + MatrixOffset& offset, + aAsRowVector, + aAsColVector) { + CHECK_EQ(useGpu_, b.useGpu_); + + int ld = stride_; + int ldb = b.stride_; + + T* dst = data_; + T* B = b.data_; + CAL_MATRIX_START_ADDRESS( + dst, height_, width_, ld, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + + if (aAsRowVector::value && !aAsColVector::value) { + if (useGpu_) { + hl_gpu_matrix_column_op(agg, op, sv, numRows, numCols, dst, B, ldb); + } else { + hl_cpu_matrix_column_op(agg, op, sv, numRows, numCols, dst, B, ldb); + } + } else if (!aAsRowVector::value && aAsColVector::value) { + if (useGpu_) { + hl_gpu_matrix_row_op(agg, op, sv, numRows, numCols, dst, ld, B, ldb); + } else { + hl_cpu_matrix_row_op(agg, op, sv, numRows, numCols, dst, ld, B, ldb); + } + } else { + LOG(FATAL) << "not supported"; + } + + return 0; +} + +template +template +int BaseMatrixT::aggregate(Agg agg, + Op op, + Saver sv, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset, + aAsRowVector, + aAsColVector) { + CHECK_EQ(useGpu_, b.useGpu_); + CHECK_EQ(useGpu_, c.useGpu_); + + int ld = stride_; + int ldb = b.stride_; + int ldc = c.stride_; + + T* dst = data_; + T* B = b.data_; + T* C = c.data_; + CAL_MATRIX_START_ADDRESS( + dst, height_, width_, ld, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CAL_MATRIX_START_ADDRESS( + C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); + + if (aAsRowVector::value && !aAsColVector::value) { + if (useGpu_) { + hl_gpu_matrix_column_op( + agg, op, sv, numRows, numCols, dst, B, ldb, C, ldc); + } else { + hl_cpu_matrix_column_op( + agg, op, sv, numRows, numCols, dst, B, ldb, C, ldc); + } + } else if (!aAsRowVector::value && aAsColVector::value) { + if (useGpu_) { + hl_gpu_matrix_row_op( + agg, op, sv, numRows, numCols, dst, ld, B, ldb, C, ldc); + } else { + hl_cpu_matrix_row_op( + agg, op, sv, numRows, numCols, dst, ld, B, ldb, C, ldc); + } + } else { + LOG(FATAL) << "not supported"; + } + + return 0; +} + +/** + * @brief unary operator. + * + */ + +DEFINE_MATRIX_UNARY_OP(Neg, a = -a); +template +void BaseMatrixT::neg() { + applyUnary(unary::Neg()); +} + +DEFINE_MATRIX_UNARY_OP(Exp, a = exp(a)); +template <> +void BaseMatrixT::exp2() { + applyUnary(unary::Exp()); +} + +DEFINE_MATRIX_UNARY_OP(Log, a = log(a)); +template <> +void BaseMatrixT::log2() { + if (useGpu_) { + applyUnary(unary::Log()); + } else { + vLog(height_ * width_, data_, data_); + } +} + +DEFINE_MATRIX_UNARY_OP(Sqrt, a = sqrt(a)); +template <> +void BaseMatrixT::sqrt2() { + applyUnary(unary::Sqrt()); +} + +DEFINE_MATRIX_UNARY_OP(Square, a = a * a); +template +void BaseMatrixT::square2() { + applyUnary(unary::Square()); +} + +DEFINE_MATRIX_UNARY_OP(Reciprocal, a = 1.0f / a); +template +void BaseMatrixT::reciprocal2() { + applyUnary(unary::Reciprocal()); +} + +DEFINE_MATRIX_UNARY_OP(Abs, a = a > 0 ? a : -a); +template +void BaseMatrixT::abs2() { + applyUnary(unary::Abs()); +} + +DEFINE_MATRIX_UNARY_OP(Sign, a = (a > 0) - (a < 0)); +template +void BaseMatrixT::sign2() { + applyUnary(unary::Sign()); +} + +DEFINE_MATRIX_UNARY_OP(Zero, a = 0); +template +void BaseMatrixT::zero() { + applyUnary(unary::Zero()); +} + +template +void BaseMatrixT::zeroAtOffset(int64_t columnOffset, int64_t numColumns) { + int numRows = height_; + int numCols = numColumns; + MatrixOffset offset(columnOffset, 0); + applyUnary(unary::Zero(), numRows, numCols, offset); +} + +DEFINE_MATRIX_UNARY_OP(One, a = 1); +template +void BaseMatrixT::one() { + applyUnary(unary::One()); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(Pow, ONE_PARAMETER, a = pow(a, p)); +template <> +void BaseMatrixT::pow2(real p) { + if (useGpu_) { + applyUnary(unary::Pow(p)); + } else { + vPow(height_ * width_, data_, p, data_); + } +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(SubScalar, ONE_PARAMETER, a -= p); +template +void BaseMatrixT::subScalar(T p) { + applyUnary(unary::SubScalar(p)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(MulScalar, ONE_PARAMETER, a *= p); +template +void BaseMatrixT::mulScalar(T p) { + applyUnary(unary::MulScalar(p)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(DivScalar, ONE_PARAMETER, a /= p); +template +void BaseMatrixT::divScalar(T p) { + applyUnary(unary::DivScalar(p)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(Assign, ONE_PARAMETER, a = p); +template +void BaseMatrixT::assign(T p) { + applyUnary(unary::Assign(p)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(Add, ONE_PARAMETER, a += p); +template +void BaseMatrixT::add(T p) { + applyUnary(unary::Add(p)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(Add2, TWO_PARAMETER, a = a * p1 + p2); +template +void BaseMatrixT::add(T p1, T p2) { + applyUnary(unary::Add2(p1, p2)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, + TWO_PARAMETER, + a = a < p1 ? p1 : (a > p2 ? p2 : a)); +template +void BaseMatrixT::clip(T p1, T p2) { + applyUnary(unary::Clip(p1, p2)); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, + TWO_PARAMETER, + a = b < p1 ? 0 : (b > p2 ? 0 : 1)); +template +void BaseMatrixT::clipDerivative(BaseMatrixT& b, T p1, T p2) { + applyBinary(binary::ClipDerivative(p1, p2), b); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, + ONE_PARAMETER, + a = a > p ? 1.0f : 0.0f); +template +void BaseMatrixT::biggerThanScalar(T p) { + applyUnary(unary::BiggerThanScalar(p)); +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(DownClip, ONE_PARAMETER, a = a > p ? a : p); +template +void BaseMatrixT::downClip(T p) { + applyUnary(unary::DownClip(p)); +} + +/** + * @brief binary operator. + * + */ + +DEFINE_MATRIX_BINARY_OP(Add, a += b); +template +void BaseMatrixT::add(BaseMatrixT& b) { + applyBinary(binary::Add(), b); +} + +template <> +void BaseMatrixT::add(BaseMatrixT& b) { + if (useGpu_) { + applyBinary(binary::Add(), b); + } else { // cpu branch + CHECK_EQ(height_, b.height_); + CHECK_EQ(width_, b.width_); + vAdd(height_ * width_, data_, b.data_, data_); + } +} + +template +void BaseMatrixT::addAtOffset(BaseMatrixT& b, int64_t columnOffset) { + if (columnOffset + b.width_ <= width_) { + int numRows = height_; + int numCols = b.width_; + MatrixOffset offset(columnOffset, 0, 0, 0); + applyBinary(binary::Add(), b, numRows, numCols, offset); + } else if (columnOffset + width_ <= b.width_) { + int numRows = height_; + int numCols = width_; + MatrixOffset offset(0, 0, columnOffset, 0); + applyBinary(binary::Add(), b, numRows, numCols, offset); + } else { + LOG(FATAL) << "Wrong argument " + << " a.width=" << width_ << " b.width=" << b.width_ + << " columnOffset=" << columnOffset; + } +} + +template +void BaseMatrixT::addP2P(BaseMatrixT& b) { + T* A = data_; + T* B = b.data_; + int dimM = height_; + int dimN = width_; + + hl_gpu_apply_binary_op, 0, 0>( + binary::Add(), A, B, dimM, dimN, dimN, dimN); +} + +template +void BaseMatrixT::addColVector(BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::Add(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /* bAsColVector */); +} + +template +void BaseMatrixT::addRowVector(BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::Add(), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(Add1, ONE_PARAMETER, a += b * p); +template +void BaseMatrixT::add(BaseMatrixT& b, T p) { + applyBinary(binary::Add1(p), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(Pow, ONE_PARAMETER, a = pow(b, p)); +template <> +void BaseMatrixT::pow2(BaseMatrixT& b, real p) { + if (useGpu_) { + applyBinary(binary::Pow(p), b); + } else { + vPow(height_ * width_, b.data_, p, data_); + } +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(Add2, TWO_PARAMETER, a = p1 * a + p2 * b); +template +void BaseMatrixT::add(BaseMatrixT& b, T p1, T p2) { + applyBinary(binary::Add2(p1, p2), b); +} + +template +void BaseMatrixT::addBias(BaseMatrixT& b, T scale) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::Add1(scale), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); +} + +DEFINE_MATRIX_BINARY_OP(Sub, a -= b); +template +void BaseMatrixT::sub(BaseMatrixT& b) { + applyBinary(binary::Sub(), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(Sub1, ONE_PARAMETER, a -= b * p); +template +void BaseMatrixT::sub(BaseMatrixT& b, T p) { + applyBinary(binary::Sub1(p), b); +} + +DEFINE_MATRIX_BINARY_OP(Relu, b = a > 0.0f ? a : 0.0f); +template +void BaseMatrixT::relu(BaseMatrixT& b) { + applyBinary(binary::Relu(), b); +} + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) +template <> +void BaseMatrixT::relu(BaseMatrixT& b) { + neon::relu(data_, b.data_, height_ * width_); +} +#endif + +DEFINE_MATRIX_BINARY_OP(ReluDerivative, a *= (b > 0.0f ? 1.0f : 0.0f)); +template +void BaseMatrixT::reluDerivative(BaseMatrixT& b) { + applyBinary(binary::ReluDerivative(), b); +} + +DEFINE_MATRIX_BINARY_OP(Softrelu, const T THRESHOLD = 40.0; + b = log(1.0 + exp((a > THRESHOLD) + ? THRESHOLD + : ((a < -THRESHOLD) ? (-THRESHOLD) + : a)))); +template <> +void BaseMatrixT::softrelu(BaseMatrixT& b) { + applyBinary(binary::Softrelu(), b); +} + +DEFINE_MATRIX_BINARY_OP( + SoftreluDerivative, const T THRESHOLD = 40.0; + a *= (1.0 - exp(-1.0 * ((b > THRESHOLD) + ? THRESHOLD + : ((b < -THRESHOLD) ? (-THRESHOLD) : b))))); +template <> +void BaseMatrixT::softreluDerivative(BaseMatrixT& b) { + applyBinary(binary::SoftreluDerivative(), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(Brelu, TWO_PARAMETER, b = a > p1 ? a : p1; + b = b < p2 ? b : p2); +template +void BaseMatrixT::brelu(BaseMatrixT& b) { + int p1 = 0, p2 = 24; //! TODO(yuyang18): Make p1,p2 configuable. + applyBinary(binary::Brelu(p1, p2), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(BreluDerivative, + TWO_PARAMETER, + a *= (b > p1 && b < p2) ? 1.0 : 0.0); +template +void BaseMatrixT::breluDerivative(BaseMatrixT& b) { + int p1 = 0, p2 = 24; + applyBinary(binary::BreluDerivative(p1, p2), b); +} + +DEFINE_MATRIX_BINARY_OP(Square, b = a * a); +template +void BaseMatrixT::square2(BaseMatrixT& b) { + applyBinary(binary::Square(), b); +} + +DEFINE_MATRIX_BINARY_OP(SquareDerivative, a *= 2.0 * b); +template +void BaseMatrixT::squareDerivative(BaseMatrixT& b) { + applyBinary(binary::SquareDerivative(), b); +} + +DEFINE_MATRIX_BINARY_OP(Tanh, T tmp = -2.0 * a; + tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp; + b = 2.0 / (1.0 + std::exp(tmp)) - 1.0); +template <> +void BaseMatrixT::tanh(BaseMatrixT& b) { + applyBinary(binary::Tanh(), b); +} + +DEFINE_MATRIX_BINARY_OP(TanhDerivative, a *= 1 - b * b); +template +void BaseMatrixT::tanhDerivative(BaseMatrixT& b) { + applyBinary(binary::TanhDerivative(), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP( + ScaledTanh, TWO_PARAMETER, b = p1 * (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0)); +template <> +void BaseMatrixT::scaledTanh(BaseMatrixT& b, real p1, real p2) { + applyBinary(binary::ScaledTanh(p1, p2), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(ScaledTanhDerivative, + TWO_PARAMETER, + a *= p2 * (p1 - b * b)); +template +void BaseMatrixT::scaledTanhDerivative(BaseMatrixT& b, T p1, T p2) { + applyBinary(binary::ScaledTanhDerivative(p1 * p1, p2 / p1), b); +} + +DEFINE_MATRIX_BINARY_OP(Reciprocal, b = 1.0f / a); +template +void BaseMatrixT::reciprocal2(BaseMatrixT& b) { + applyBinary(binary::Reciprocal(), b); +} + +DEFINE_MATRIX_BINARY_OP(ReciprocalDerivative, a *= -b * b); +template +void BaseMatrixT::reciprocalDerivative(BaseMatrixT& b) { + applyBinary(binary::ReciprocalDerivative(), b); +} + +DEFINE_MATRIX_BINARY_OP(Abs, b = a > 0.0f ? a : -a); +template +void BaseMatrixT::abs2(BaseMatrixT& b) { + applyBinary(binary::Abs(), b); +} + +DEFINE_MATRIX_BINARY_OP(AbsDerivative, a = (b > 0) ? a : (b < 0) ? -a : 0); +template +void BaseMatrixT::absDerivative(BaseMatrixT& b) { + applyBinary(binary::AbsDerivative(), b); +} + +DEFINE_MATRIX_BINARY_OP(Sigmoid, const T THRESHOLD_MIN = -40.0; + const T THRESHOLD_MAX = 13.0; + T tmp = (a < THRESHOLD_MIN) + ? THRESHOLD_MIN + : ((a > THRESHOLD_MAX) ? THRESHOLD_MAX : a); + b = 1.0f / (1.0f + exp(-tmp))); +template <> +void BaseMatrixT::sigmoid(BaseMatrixT& b) { + if (useGpu_) { + applyBinary(binary::Sigmoid(), b); + } else { // cpu versioni + size_t numSamples = this->height_; + size_t dim = this->width_; + CHECK_EQ(b.height_, numSamples); + CHECK_EQ(b.width_, dim); + const real* in = this->data_; + real* out = b.data_; + + // out = - in + const float THRESHOLD_MIN = -40.0; // make sure sigmoid(x) > 0 + const float THRESHOLD_MAX = 13.0; // make sure sigmoid(x) < 1 + for (size_t i = 0; i < numSamples * dim; ++i) { + real tmp = in[i]; + tmp = (tmp < THRESHOLD_MIN) + ? THRESHOLD_MIN + : ((tmp > THRESHOLD_MAX) ? THRESHOLD_MAX : tmp); + out[i] = -tmp; + } + + // out = exp(out) + vExp(numSamples * dim, out, out); + + // out = 1 / (1 + out) + for (size_t i = 0; i < numSamples * dim; ++i) { + out[i] = 1 / (1 + out[i]); + } + } +} + +DEFINE_MATRIX_BINARY_OP(SigmoidDerivative, a *= b * (1 - b)); +template +void BaseMatrixT::sigmoidDerivative(BaseMatrixT& b) { + applyBinary(binary::SigmoidDerivative(), b); +} + +DEFINE_MATRIX_BINARY_OP(ExpDerivative, a *= b); +template +void BaseMatrixT::expDerivative(BaseMatrixT& b) { + applyBinary(binary::ExpDerivative(), b); +} + +DEFINE_MATRIX_BINARY_OP(Sign, b = a > 0.0f ? 1.0f : -1.0f); +template +void BaseMatrixT::sign2(BaseMatrixT& b) { + applyBinary(binary::Sign(), b); +} + +DEFINE_MATRIX_BINARY_OP(Exp, a = exp(b)); +template <> +void BaseMatrixT::exp2(BaseMatrixT& b) { + applyBinary(binary::Exp(), b); +} + +DEFINE_MATRIX_BINARY_OP(Log, a = log(b)); +template <> +void BaseMatrixT::log2(BaseMatrixT& b) { + if (useGpu_) { + applyBinary(binary::Log(), b); + } else { + vLog(height_ * width_, b.data_, data_); + } +} + +DEFINE_MATRIX_BINARY_OP(Sqrt, a = sqrt(b)); +template <> +void BaseMatrixT::sqrt2(BaseMatrixT& b) { + applyBinary(binary::Sqrt(), b); +} + +DEFINE_MATRIX_BINARY_OP(InvSqrt, a = 1.0f / sqrt(b)); +template <> +void BaseMatrixT::invSqrt(BaseMatrixT& b) { + if (useGpu_) { + applyBinary(binary::InvSqrt(), b); + } else { // cpu branch + CHECK_EQ(height_, b.height_); + CHECK_EQ(width_, b.width_); + vInvSqrt(height_ * width_, b.data_, data_); + } +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(IsEqual, ONE_PARAMETER, a = (b == p)); +template +void BaseMatrixT::isEqualTo(BaseMatrixT& b, T value) { + applyBinary(binary::IsEqual(value), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(AddScalar, ONE_PARAMETER, a = b + p); +template +void BaseMatrixT::addScalar(BaseMatrixT& b, T p) { + applyBinary(binary::AddScalar(p), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(SubScalar, ONE_PARAMETER, a = b - p); +template +void BaseMatrixT::subScalar(BaseMatrixT& b, T p) { + applyBinary(binary::SubScalar(p), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(MulScalar, ONE_PARAMETER, a = b * p); +template +void BaseMatrixT::mulScalar(BaseMatrixT& b, T p) { + applyBinary(binary::MulScalar(p), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(DivScalar, ONE_PARAMETER, a = b / p); +template +void BaseMatrixT::divScalar(BaseMatrixT& b, T p) { + applyBinary(binary::DivScalar(p), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(ScalarDiv, ONE_PARAMETER, a = p / b); +template +void BaseMatrixT::scalarDiv(BaseMatrixT& b, T p) { + applyBinary(binary::ScalarDiv(p), b); +} + +/** + * @brief ternary operator. + * + */ + +DEFINE_MATRIX_TERNARY_OP(SoftCrossEntropy, + a = -c * log(b) - (1 - c) * log(1 - b)); +template <> +void BaseMatrixT::softCrossEntropy(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::SoftCrossEntropy(), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(SoftCrossEntropyBp, a += (b - c) / (b * (1 - b))); +template +void BaseMatrixT::softCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::SoftCrossEntropyBp(), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(BinaryCrossEntropy, + a = c > 0.5 ? -log(b) : -log(1.0 - b)); +template <> +void BaseMatrixT::binaryLabelCrossEntropy(BaseMatrixT& b, + BaseMatrixT& c) { + if (useGpu_) { + applyTernary(ternary::BinaryCrossEntropy(), b, c); + } else { + CHECK_EQ(height_, b.height_); + CHECK_EQ(height_, c.height_); + CHECK_EQ(width_, b.width_); + CHECK_EQ(width_, c.width_); + + size_t size = height_ * width_; + real* out = b.data_; + real* label = c.data_; + real* cost = data_; + + for (size_t i = 0; i < size; ++i) { + cost[i] = label[i] > 0.5 ? out[i] : 1.0 - out[i]; + } + vLog(size, cost, cost); + for (size_t i = 0; i < size; ++i) { + cost[i] *= -1.0; + } + } +} + +DEFINE_MATRIX_TERNARY_OP(BinaryCrossEntropyBp, + a += c > 0.5 ? -1.0 / b : 1.0 / (1.0 - b)); +template +void BaseMatrixT::binaryLabelCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::BinaryCrossEntropyBp(), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(Add, a = b + c); +template +void BaseMatrixT::add(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::Add(), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add1, TWO_PARAMETER, a = p1 * b + p2 * c); +template +void BaseMatrixT::add(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2) { + applyTernary(ternary::Add1(p1, p2), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(Sub, a = b - c); +template +void BaseMatrixT::sub(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::Sub(), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(Sub1, TWO_PARAMETER, a = p1 * b - p2 * c); +template +void BaseMatrixT::sub(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2) { + applyTernary(ternary::Sub1(p1, p2), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(Add2, a = a + b + c); +template +void BaseMatrixT::add2(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::Add2(), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add3, + THREE_PARAMETER, + a = p1 * a + p2 * b + p3 * c); +template +void BaseMatrixT::add2(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3) { + applyTernary(ternary::Add3(p1, p2, p3), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(SgdUpdate, + THREE_PARAMETER, + c = p2 * c - p1 * (b + p3 * a); + a = a + c); +template +void BaseMatrixT::sgdUpdate(BaseMatrixT& b, // grad + BaseMatrixT& c, // mom + T p1, // learningRate, + T p2, // momentum, + T p3) { // decayRate + applyTernary(ternary::SgdUpdate(p1, p2, p3), b, c); +} + +DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(SgdUpdate, + THREE_PARAMETER, + c = p2 * c - p1 * d * (b + p3 * a); + a += c); +template +void BaseMatrixT::sgdUpdate(BaseMatrixT& b, // grad, + BaseMatrixT& c, // mom, + BaseMatrixT& d, // lr, + T p1, // learningRate, + T p2, // momentum, + T p3) { // decayRate + applyQuaternary(quaternary::SgdUpdate(p1, p2, p3), b, c, d); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL1, ONE_PARAMETER, T lambda = p * b; + a = (a > lambda) + ? (a - lambda) + : (a < -lambda) ? (a + lambda) : 0); +template +void BaseMatrixT::applyL1(BaseMatrixT& lr, T learningRate, T decayRate) { + applyBinary(binary::ApplyL1(learningRate * decayRate), lr); +} + +template <> +void BaseMatrixT::applyL1(BaseMatrixT& lr, + real learningRate, + real decayRate) { + if (useGpu_) { + applyBinary(binary::ApplyL1(learningRate * decayRate), lr); + } else { + simd::decayL1(this->data_, + this->data_, + lr.data_, + learningRate * decayRate, + height_ * width_); + } +} + +DEFINE_MATRIX_UNARY_PARAMETER_OP(ApplyL1, ONE_PARAMETER, T lambda = p; + a = (a > lambda) + ? (a - lambda) + : (a < -lambda) ? (a + lambda) : 0); +template +void BaseMatrixT::applyL1(T learningRate, T decayRate) { + applyUnary(unary::ApplyL1(learningRate * decayRate)); +} + +template <> +void BaseMatrixT::applyL1(real learningRate, real decayRate) { + if (useGpu_) { + applyUnary(unary::ApplyL1(learningRate * decayRate)); + } else { + simd::decayL1( + this->data_, this->data_, learningRate * decayRate, height_ * width_); + } +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL2, + ONE_PARAMETER, + a *= (1.0f / (1.0f + p * b))); +template +void BaseMatrixT::applyL2(BaseMatrixT& lr, T learningRate, T decayRate) { + if (useGpu_) { + applyBinary(binary::ApplyL2(learningRate * decayRate), lr); + } else { + size_t size = this->height_ * this->width_; + T decay = learningRate * decayRate; + for (size_t j = 0; j < size; ++j) { + this->data_[j] *= 1.0f / (1.0f + decay * lr.data_[j]); + } + } +} + +template +void BaseMatrixT::applyL2(T learningRate, T decayRate) { + BaseMatrixT::mulScalar(1.0f / (1.0f + learningRate * decayRate)); +} + +DEFINE_MATRIX_BINARY_OP(DotMul, a *= b); +template +void BaseMatrixT::dotMul(BaseMatrixT& b) { + applyBinary(binary::DotMul(), b); +} + +DEFINE_MATRIX_TERNARY_OP(DotMul, a = b * c); +template +void BaseMatrixT::dotMul(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::DotMul(), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(DotDiv, a = (b == 0.0) ? 0.0 : b / c); +template +void BaseMatrixT::dotDiv(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::DotDiv(), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotDiv2P, + TWO_PARAMETER, + a = (b + p1) / (c + p2)); +template +void BaseMatrixT::dotDiv(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { + applyTernary(ternary::DotDiv2P(p1, p2), b, c); +} + +DEFINE_MATRIX_QUATERNARY_OP(RankLoss, const T THRESHOLD = 40.0; a = b - c; + a = (a > THRESHOLD) + ? THRESHOLD + : ((a < -THRESHOLD) ? (-THRESHOLD) : a); + a = log(1 + exp(a)) - a * d); +template <> +void BaseMatrixT::rankLoss(BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d) { + applyQuaternary(quaternary::RankLoss(), b, c, d); +} + +DEFINE_MATRIX_QUATERNARY_OP(RankLossBp, const T THRESHOLD = 40.0; a = b - c; + a = (a > THRESHOLD) + ? THRESHOLD + : ((a < -THRESHOLD) ? (-THRESHOLD) : a); + a = exp(a); + a = (a / (1 + a) - d)); +template <> +void BaseMatrixT::rankLossBp(BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d) { + applyQuaternary(quaternary::RankLossBp(), b, c, d); +} + +/* this = log(1 + exp(b)) - c * b */ +DEFINE_MATRIX_TERNARY_OP(LogisticRegressionLoss, const T THRESHOLD = 40.0; + T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) + ? -THRESHOLD + : b; + a = log(1 + exp(x)) - c * x); +template <> +void BaseMatrixT::logisticRegressionLoss(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::LogisticRegressionLoss(), b, c); +} + +/* this = exp(b)/(1+exp(b)) - c */ +DEFINE_MATRIX_TERNARY_OP(LogisticRegressionLossBp, const T THRESHOLD = 40.0; + T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) + ? -THRESHOLD + : b; + x = exp(x); + a = x / (1 + x) - c); +template <> +void BaseMatrixT::logisticRegressionLossBp(BaseMatrixT& b, + BaseMatrixT& c) { + applyTernary(ternary::LogisticRegressionLossBp(), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(BiggerThan, a = (b > c) ? 1.0f : 0.0f); +template +void BaseMatrixT::biggerThan(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::BiggerThan(), b, c); +} + +DEFINE_MATRIX_QUATERNARY_OP( + BiggerThan, a = ((b > c && d > 0.5f) || (b < c && d < 0.5f)) ? 1.0f : 0.0f); +template +void BaseMatrixT::biggerThan(BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d) { + applyQuaternary(quaternary::BiggerThan(), b, c, d); +} + +DEFINE_MATRIX_TERNARY_OP(Max, a = (b > c) ? b : c); +template +void BaseMatrixT::max2(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::Max(), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(BinaryClassificationError, + ONE_PARAMETER, + c += ((a > p) == (b > p)) ? 0.0f : 1.0f); +template +void BaseMatrixT::binaryClassificationError2(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c, + T p) { + CHECK(!useGpu_) << "do not support gpu"; + MatrixOffset offset(0, 0, 0, 0, destCol, 0); + int numRows = b.height_; + int numCols = b.width_; + b.applyTernary(ternary::BinaryClassificationError(p), + c, + *this, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); +} + +template <> +void BaseMatrixT::binaryClassificationError(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c, + real p) { + MatrixOffset offset(destCol, 0, 0, 0, 0, 0); + int numRows = b.height_; + int numCols = b.width_; + aggregate(aggregate::sum(), + base::binary::classificationError(p), + base::binary::add(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); +} + +DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(Add3, + THREE_PARAMETER, + a = p1 * b + p2 * c + p3 * d); +template +void BaseMatrixT::add3( + BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d, T p1, T p2, T p3) { + applyQuaternary(quaternary::Add3(p1, p2, p3), b, c, d); +} + +DEFINE_MATRIX_TERNARY_OP(DotMulSquare, a = b * c * c); +template +void BaseMatrixT::dotMulSquare(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::DotMulSquare(), b, c); +} + +DEFINE_MATRIX_TERNARY_OP(DotSquareSquare, a = b * b * c * c); +template +void BaseMatrixT::dotSquareSquare(BaseMatrixT& b, BaseMatrixT& c) { + applyTernary(ternary::DotSquareSquare(), b, c); +} + +DEFINE_MATRIX_BINARY_OP(DotMulSquare, a *= b * b); +template +void BaseMatrixT::dotMulSquare(BaseMatrixT& b) { + applyBinary(binary::DotMulSquare(), b); +} + +DEFINE_MATRIX_BINARY_OP(DotSquareMul, a = a * a * b); +template +void BaseMatrixT::dotSquareMul(BaseMatrixT& b) { + applyBinary(binary::DotSquareMul(), b); +} + +DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(AddSquareSum, + THREE_PARAMETER, + T tmp = p1 * b + p2 * c + p3 * d; + a += tmp * tmp); +template +void BaseMatrixT::addSquareSum( + BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT d, T p1, T p2, T p3) { + applyQuaternary(quaternary::AddSquareSum(p1, p2, p3), b, c, d); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(AddSquare, ONE_PARAMETER, a += p * b * b); +template +void BaseMatrixT::addSquare(BaseMatrixT& b, T p) { + applyBinary(binary::AddSquare(p), b); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(DecayAddSquare, + TWO_PARAMETER, + a = p1 * a + p2 * b * b); +template +void BaseMatrixT::decayAddSquare(BaseMatrixT& b, T p1, T p2) { + applyBinary(binary::DecayAddSquare(p1, p2), b); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DecayAddSquareMul, + TWO_PARAMETER, + a = p1 * a + p2 * b * b * c * c); +template +void BaseMatrixT::decayAddSquareMul(BaseMatrixT& b, + BaseMatrixT& c, + T p1, + T p2) { + applyTernary(ternary::DecayAddSquareMul(p1, p2), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(ReciprocalSum, + THREE_PARAMETER, + a = 1 / (p1 * b + p2 * c + p3)); +template +void BaseMatrixT::reciprocalSum( + BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3) { + applyTernary(ternary::ReciprocalSum(p1, p2, p3), b, c); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(Reciprocal2, + TWO_PARAMETER, + a = 1 / (p1 * b + p2)); +template +void BaseMatrixT::reciprocal2(BaseMatrixT& b, T p1, T p2) { + applyBinary(binary::Reciprocal2(p1, p2), b); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSquareSum, + TWO_PARAMETER, + T tmp = p1 * b + p2 * c; + a *= tmp * tmp); +template +void BaseMatrixT::dotMulSquareSum(BaseMatrixT& b, + BaseMatrixT& c, + T p1, + T p2) { + applyTernary(ternary::DotMulSquareSum(p1, p2), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotSquareSum, + TWO_PARAMETER, + T tmp = p1 * b + p2 * c; + a = tmp * tmp); +template +void BaseMatrixT::dotSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { + applyTernary(ternary::DotSquareSum(p1, p2), b, c); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSum, + TWO_PARAMETER, + a *= p1 * b + p2 * c); +template +void BaseMatrixT::dotMulSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { + applyTernary(ternary::DotMulSum(p1, p2), b, c); +} + +DEFINE_MATRIX_BINARY_OP(CopyAndClear, b = a; a = 0); +template +void BaseMatrixT::copyAndClear(BaseMatrixT& b) { + applyBinary(binary::CopyAndClear(), b); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(AddDotMul, + TWO_PARAMETER, + a = p1 * a + p2 * b * c); +template +void BaseMatrixT::addDotMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { + applyTernary(ternary::AddDotMul(p1, p2), b, c); +} + +DEFINE_MATRIX_BINARY_OP(Assign, a = b;); +template +void BaseMatrixT::assign(BaseMatrixT& b) { + if (useGpu_) { + applyBinary(binary::Assign(), b); + } else { // cpu version + CHECK_EQ(this->height_, b.height_); + CHECK_EQ(this->width_, b.width_); + memcpy(data_, b.data_, sizeof(T) * height_ * width_); + } +} + +template +void BaseMatrixT::assignAtOffset(BaseMatrixT& b, int64_t columnOffset) { + if (columnOffset + b.width_ <= width_) { + int numRows = height_; + int numCols = b.width_; + MatrixOffset offset(columnOffset, 0, 0, 0); + applyBinary(binary::Assign(), b, numRows, numCols, offset); + } else if (columnOffset + width_ <= b.width_) { + int numRows = height_; + int numCols = width_; + MatrixOffset offset(0, 0, columnOffset, 0); + applyBinary(binary::Assign(), b, numRows, numCols, offset); + } else { + LOG(FATAL) << "Wrong argument " + << " a.width=" << width_ << " b.width=" << b.width_ + << " columnOffset=" << columnOffset; + } +} + +DEFINE_MATRIX_BINARY_OP(DeepSwap, T tmp = a; a = b; b = tmp); +template +void BaseMatrixT::deepSwap(BaseMatrixT& b) { + applyBinary(binary::DeepSwap(), b); +} + +template <> +void BaseMatrixT::rowDotMul(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c) { + int numRows = b.height_; + int numCols = b.width_; + MatrixOffset offset(destCol, 0, 0, 0, 0, 0); + aggregate(aggregate::sum(), + base::binary::mul(), + base::binary::add(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); +} + +template +void BaseMatrixT::rowDotMul2(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c) { + CHECK(!useGpu_) << "do not support gpu"; + + size_t height = this->height_; + CHECK_LT(destCol, this->width_); + CHECK_EQ(height, b.height_); + CHECK_EQ(height, c.height_); + CHECK_EQ(b.width_, c.width_); + size_t width = b.width_; + T* A = this->data_; + const T* B = b.data_; + const T* C = c.data_; + for (size_t i = 0; i < height; + ++i, A += this->width_, B += width, C += width) { + for (size_t j = 0; j < width; ++j) { + A[destCol] += B[j] * C[j]; + } + } +} + +template <> +void BaseMatrixT::addDotMulVMM(BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + int numRows = b.height_; + int numCols = b.width_; + aggregate(aggregate::sum(), + base::binary::mul(), + base::binary::add(), + b, + c, + numRows, + numCols, + offset, + true_type() /*aAsRowVector*/, + false_type()); +} + +template +void BaseMatrixT::addDotMulVMM2(BaseMatrixT& b, BaseMatrixT& c) { + CHECK(!useGpu_) << "do not support gpu"; + + CHECK_EQ(height_, 1LU); + CHECK_EQ(b.height_, c.height_); + CHECK_EQ(width_, b.width_); + CHECK_EQ(width_, c.width_); + size_t height = b.height_; + size_t width = b.width_; + T* A = this->data_; + const T* B = b.data_; + const T* C = c.data_; + for (size_t i = 0; i < height; ++i, B += width, C += width) { + for (size_t j = 0; j < width; ++j) { + A[j] += B[j] * C[j]; + } + } +} + +DEFINE_MATRIX_TERNARY_OP(addDotMulMMV, a += b * c); +template +void BaseMatrixT::addDotMulMMV(BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::addDotMulMMV(), + b, + c, + numRows, + numCols, + offset, + true_type() /*cAsRowVector*/, + false_type()); +} + +template +void BaseMatrixT::addDotMulMMV2(BaseMatrixT& b, BaseMatrixT& c) { + CHECK(!useGpu_) << "do not support gpu"; + + CHECK_EQ(c.height_, 1LU); + CHECK_EQ(height_, b.height_); + CHECK_EQ(width_, b.width_); + CHECK_EQ(width_, c.width_); + size_t height = height_; + size_t width = width_; + T* A = this->data_; + const T* B = b.data_; + const T* C = c.data_; + for (size_t i = 0; i < height; ++i, A += width, B += width) { + for (size_t j = 0; j < width; ++j) { + A[j] += B[j] * C[j]; + } + } +} + +template +void BaseMatrixT::rowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, cCol, 0); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::DotMul(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); +} + +template +void BaseMatrixT::rowScale2(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { + CHECK(!useGpu_) << "do not support gpu"; + + size_t height = this->height_; + size_t width = this->width_; + CHECK_EQ(height, b.height_); + CHECK_EQ(width, b.width_); + CHECK_LT(cCol, c.width_); + CHECK_EQ(height, c.height_); + T* A = this->data_; + const T* B = b.data_; + const T* C = c.data_; + for (size_t i = 0; i < height; ++i, A += width, B += width, C += c.width_) { + for (size_t j = 0; j < width; ++j) { + A[j] = B[j] * C[cCol]; + } + } +} + +template +void BaseMatrixT::colScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, 0, cRow); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::DotMul(), + b, + c, + numRows, + numCols, + offset, + true_type() /* cAsRowVector */, + false_type() /* cAsColVector */); +} + +template +void BaseMatrixT::addColScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, 0, cRow); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::addDotMulMMV(), + b, + c, + numRows, + numCols, + offset, + true_type() /* cAsRowVector */, + false_type() /* cAsColVector */); +} + +template +void BaseMatrixT::addRowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, cCol, 0); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::addDotMulMMV(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); +} + +DEFINE_MATRIX_TERNARY_PARAMETER_OP(RowAdd, ONE_PARAMETER, a = b + p * c); +template +void BaseMatrixT::rowAdd(size_t cCol, BaseMatrixT& b, BaseMatrixT& c, T p) { + MatrixOffset offset(0, 0, 0, 0, cCol, 0); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::RowAdd(p), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); +} + +DEFINE_MATRIX_TERNARY_OP(RowPow, a = pow(b, c)); +template <> +void BaseMatrixT::rowPow(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { + if (useGpu_) { + MatrixOffset offset(0, 0, 0, 0, cCol, 0); + int numRows = height_; + int numCols = width_; + applyTernary(ternary::RowPow(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); + } else { + size_t height = this->height_; + size_t width = this->width_; + CHECK_EQ(height, b.height_); + CHECK_EQ(width, b.width_); + CHECK_LT(cCol, c.width_); + CHECK_EQ(height, c.height_); + real* A = this->data_; + const real* B = b.data_; + const real* C = c.data_; + for (size_t i = 0; i < height; ++i, A += width, B += width, C += c.width_) { + vPow(width, B, C[cCol], A); + } + } +} + +template +void BaseMatrixT::mulRowVector(BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::DotMul(), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); +} + +DEFINE_MATRIX_BINARY_OP(DotDiv, a /= b); +template +void BaseMatrixT::divRowVector(BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::DotDiv(), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); +} + +template +void BaseMatrixT::mulColVector(BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::DotMul(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /* bAsColVector */); +} + +template +void BaseMatrixT::divColVector(BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0); + int numRows = height_; + int numCols = width_; + applyBinary(binary::DotDiv(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /* bAsColVector */); +} + +template <> +template +int BaseMatrixT::applyRow(Agg agg, BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + size_t numRows = b.height_; + size_t numCols = b.width_; + CHECK_EQ(height_, numRows); + CHECK_EQ(width_, 1UL); + aggregate(agg, + base::unary::identity(), + base::binary::second(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); + + return 0; +} + +template <> +template +int BaseMatrixT::applyRow(Agg agg, Saver sv, BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + size_t numRows = b.height_; + size_t numCols = b.width_; + CHECK_EQ(height_, numRows); + CHECK_EQ(width_, 1UL); + aggregate(agg, + base::unary::identity(), + sv, + b, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); + + return 0; +} + +template <> +template +int BaseMatrixT::applyRow(Agg agg, + real scaleDest, + real scaleAgg, + BaseMatrixT& b) { + if (scaleDest != 0) { + applyRow(agg, base::binary::add2(scaleDest, scaleAgg), b); + } else { + applyRow(agg, base::binary::second(), b); + if (scaleAgg != 1) { + mulScalar(scaleAgg); + } + } + return 0; +} + +template <> +template +int BaseMatrixT::applyRow( + Agg agg, Op op, Saver sv, BaseMatrixT& b, BaseMatrixT& c) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + size_t numRows = b.height_; + size_t numCols = b.width_; + CHECK_EQ(height_, numRows); + CHECK_EQ(width_, 1UL); + CHECK_EQ(c.height_, numRows); + CHECK_EQ(c.width_, numCols); + aggregate(agg, + op, + sv, + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); + return 0; +} + +template <> +template +int BaseMatrixT::applyRow(Agg agg, + Op op, + real scaleDest, + real scaleAgg, + BaseMatrixT& b, + BaseMatrixT& c) { + if (scaleDest != 0) { + applyRow(agg, op, base::binary::add2(scaleDest, scaleAgg), b, c); + } else { + applyRow(agg, op, base::binary::second(), b, c); + if (scaleAgg != 1) { + mulScalar(scaleAgg); + } + } + return 0; +} + +template <> +template +int BaseMatrixT::applyCol(Agg agg, BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + size_t numRows = b.height_; + size_t numCols = b.width_; + CHECK_EQ(width_, numCols); + CHECK_EQ(height_, 1UL); + aggregate(agg, + base::unary::identity(), + base::binary::second(), + b, + numRows, + numCols, + offset, + true_type() /*aAsRowVector*/, + false_type()); + + return 0; +} + +template <> +template +int BaseMatrixT::applyCol(Agg agg, Saver sv, BaseMatrixT& b) { + MatrixOffset offset(0, 0, 0, 0, 0, 0); + size_t numRows = b.height_; + size_t numCols = b.width_; + CHECK_EQ(width_, numCols); + CHECK_EQ(height_, 1UL); + aggregate(agg, + base::unary::identity(), + sv, + b, + numRows, + numCols, + offset, + true_type() /*aAsRowVector*/, + false_type()); + + return 0; +} + +template <> +template +int BaseMatrixT::applyCol(Agg agg, + real scaleDest, + real scaleAgg, + BaseMatrixT& b) { + if (scaleDest != 0) { + applyCol(agg, base::binary::add2(scaleDest, scaleAgg), b); + } else { + applyCol(agg, base::binary::second(), b); + if (scaleAgg != 1) { + mulScalar(scaleAgg); + } + } + return 0; +} + +template <> +void BaseMatrixT::sumRows(BaseMatrixT& b, real scaleSum, real scaleDest) { + applyRow(aggregate::sum(), scaleDest, scaleSum, b); +} + +template <> +void BaseMatrixT::maxRows(BaseMatrixT& b) { + applyRow(aggregate::max(), b); +} + +template <> +void BaseMatrixT::minRows(BaseMatrixT& b) { + applyRow(aggregate::min(), b); +} + +template <> +void BaseMatrixT::maxCols(BaseMatrixT& b) { + applyCol(aggregate::max(), b); +} + +template <> +void BaseMatrixT::minCols(BaseMatrixT& b) { + applyCol(aggregate::min(), b); +} + +template <> +void BaseMatrixT::sumCols(BaseMatrixT& b, real scaleSum, real scaleDest) { + applyCol(aggregate::sum(), scaleDest, scaleSum, b); +} + +template <> +void BaseMatrixT::sumOfSquaredDiffs(BaseMatrixT& b, + BaseMatrixT& c, + real scaleSum, + real scaleDest) { + applyRow( + aggregate::sum(), base::binary::squaredDiff(), scaleDest, scaleSum, b, c); +} + +template <> +void BaseMatrixT::sumOfProducts(BaseMatrixT& b, + BaseMatrixT& c, + real scaleSum, + real scaleDest) { + applyRow(aggregate::sum(), base::binary::mul(), scaleDest, scaleSum, b, c); +} + +template class BaseMatrixT; + +#ifndef PADDLE_MOBILE_INFERENCE + +template class BaseMatrixT; + +#else + +template <> +void BaseMatrixT::zero() { + applyUnary(unary::Zero()); +} + +template <> +void BaseMatrixT::assign(int p) { + applyUnary(unary::Assign(p)); +} + +template <> +void BaseMatrixT::isEqualTo(BaseMatrixT& b, int value) { + applyBinary(binary::IsEqual(value), b); +} + +template <> +void BaseMatrixT::neg() { + applyUnary(unary::Neg()); +} + +template <> +void BaseMatrixT::abs2() { + applyUnary(unary::Abs()); +} + +template <> +void BaseMatrixT::add(int p) { + applyUnary(unary::Add(p)); +} + +template <> +void BaseMatrixT::add(int p1, int p2) { + applyUnary(unary::Add2(p1, p2)); +} + +template <> +void BaseMatrixT::applyL1(int learningRate, int decayRate) { + applyUnary(unary::ApplyL1(learningRate * decayRate)); +} + +#endif +} // namespace paddle diff --git a/paddle/legacy/math/BaseMatrix.h b/paddle/legacy/math/BaseMatrix.h new file mode 100644 index 0000000000000000000000000000000000000000..4627f847d356f07600edae8cadcb02302e19381c --- /dev/null +++ b/paddle/legacy/math/BaseMatrix.h @@ -0,0 +1,1095 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "TensorExpression.h" +#include "paddle/legacy/utils/Common.h" + +namespace paddle { + +/* + * nvcc currently does not support C++11, + * so I realized false_type and true_type. + */ +template +struct bool_constant { + static const T value = v; +}; +typedef bool_constant false_type; +typedef bool_constant true_type; + +/** + * @brief Calculate matrix element address. + * + * For instance, address of A[i][j] = i * ld + j. + * + */ +#define CAL_MATRIX_START_ADDRESS(address, height, width, ld, col, row) \ + CHECK_LE(col, width); \ + CHECK_LE(row, height); \ + address += row * ld + col; + +class MatrixOffset { + public: + size_t aCol_; + size_t aRow_; + size_t bCol_; + size_t bRow_; + size_t cCol_; + size_t cRow_; + size_t dCol_; + size_t dRow_; + MatrixOffset(size_t aCol = 0, + size_t aRow = 0, + size_t bCol = 0, + size_t bRow = 0, + size_t cCol = 0, + size_t cRow = 0, + size_t dCol = 0, + size_t dRow = 0) + : aCol_(aCol), + aRow_(aRow), + bCol_(bCol), + bRow_(bRow), + cCol_(cCol), + cRow_(cRow), + dCol_(dCol), + dRow_(dRow) {} +}; + +template +class BaseMatrixT : public TensorExpression, T> { + public: + size_t height_, width_; + size_t stride_; + T* data_; + bool trans_; + bool useGpu_; + + public: + virtual ~BaseMatrixT() {} + BaseMatrixT(size_t height, size_t width, T* data, bool trans, bool useGpu) + : height_(height), + width_(width), + stride_(width), + data_(data), + trans_(trans), + useGpu_(useGpu) {} + + /** + * @note This constructor is for temporarily making a matrix with different + * useGpu flag as the original matrix so that mixed gpu/cpu operations + * can be performed successfully. + */ + BaseMatrixT(BaseMatrixT& mat, bool useGpu) + : height_(mat.height_), + width_(mat.width_), + stride_(mat.stride_), + data_(mat.data_), + trans_(mat.trans_), + useGpu_(useGpu) {} + + BaseMatrixT(size_t height, + size_t width, + size_t stride, + T* data, + bool trans, + bool use_gpu) + : height_(height), + width_(width), + stride_(stride), + data_(data), + trans_(trans), + useGpu_(use_gpu) { + /* CHECK_LE(width_, stride_); */ + } + + /// caller should make sure that the size of data is at least height*width + void setData(T* data) { data_ = data; } + + /** + * unary operator: element wise op(a). + * + * @code + * for 0 <= i < this->height_ & for 0 <= j < this->width_. + * @endcode + */ + template + int applyUnary(Op op); + + /** + * unary operator: element wise op(a). + * + * @code + * for 0 <= i < numRows & for 0 <= j < numCols. + * While matrix start address is: + * A = this->data_ + offset.aRow_*ld + offset.aCol_; + * @endcode + */ + template + int applyUnary(Op op, int numRows, int numCols, MatrixOffset& offset); + + /** + * binary operator: element wise op(a, b). + * + * @code + * for 0 <= i < this->height_ & for 0 <= j < this->width_. + * While this->height_ == b.height_ && this->width_ == b.width_. + * @endcode + */ + template + int applyBinary(Op op, BaseMatrixT& b); + + /** + * binary operator: element wise op(a, b) + * + * @code + * for 0 <= i < numRows & for 0 <= j < numCols. + * While matrix start address is: + * A = this->data_ + offset.aRow_*lda + offset.aCol_; + * B = b->data_ + offset.bRow_*ldb + offset.bCol_; + * + * if (bAsRowVector == false_type && bAsColVector == false_type) + * op(A[i * lda + j], B[i * ldb + j]) + * + * if (bAsRowVector == true_type && bAsColVector == false_type) + * op(A[i * lda + j], B[j]) + * + * if (bAsRowVector == false_type && bAsColVector == true_type) + * op(A[i * lda + j], B[i * ldb]) + * + * if (bAsRowVector == true_type && bAsColVector == true_type) + * op(A[i * lda + j], B[0]) + * @endcode + */ + template + int applyBinary(Op op, + BaseMatrixT& b, + int numRows, + int numCols, + MatrixOffset& offset, + bAsRowVector, + bAsColVector); + + template + int applyBinary( + Op op, BaseMatrixT& b, int numRows, int numCols, MatrixOffset& offset); + + /** + * ternary operator: element wise op(a, b, c). + * + * @code + * for 0 <= i < this->height_ & for 0 <= j < this->width_. + * + * While this->height_ == b.height_ && this->width_ == b.width_ + * && this->height_ == c.height_ && this->width_ == c.width_ + * @endcode + */ + template + int applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c); + + /** + * ternary operator: element wise op(a, b, c). + * + * @code + * for 0 <= i < numRows & for 0 <= j < numCols. + * While matrix start address is: + * + * A = this->data_ + offset.aRow_*lda + offset.aCol_; + * B = b->data_ + offset.bRow_*ldb + offset.bCol_; + * C = c->data_ + offset.cRow_*ldc + offset.cCol_; + * + * if (cAsRowVector == false_type && cAsColVector == false_type) + * op(A[i*lda + j], B[i*ldb + j], C[i*ldc + j]) + * + * if (cAsRowVector == true_type && cAsColVector == false_type) + * op(A[i*lda + j], B[i*ldb + j], C[j]) + * + * if (cAsRowVector == false_type && cAsColVector == true_type) + * op(A[i*lda + j], B[i*ldb + j], C[i*ldc]) + * + * if (cAsRowVector == 1 && cAsColVector == 1) + * op(A[i*lda + j], B[i*ldb + j], C[0]) + * @endcode + */ + template + int applyTernary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset, + cAsRowVector, + cAsColVector); + + template + int applyTernary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset); + + /** + * quaternary operator: element wise op(a, b, c, d). + * + * @code + * for 0 <= i < this->height_ & for 0 <= j < this->width_. + * + * While this->height_ == b.height_ && this->width_ == b.width_ + * && this->height_ == c.height_ && this->width_ == c.width_ + * && this->height_ == d.height_ && this->width_ == d.width_ + * @endcode + */ + template + int applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d); + + /** + * quaternary operator: element wise op(a, b, c, d). + * + * @code + * for 0 <= i < numRows & for 0 <= j < numCols. + * While matrix start address is: + * A = this->data_ + offset.aRow_*lda + offset.aCol_; + * B = b->data_ + offset.bRow_*ldb + offset.bCol_; + * C = c->data_ + offset.cRow_*ldc + offset.cCol_; + * D = d->data_ + offset.dRow_*ldd + offset.dCol_; + * @endcode + */ + template + int applyQuaternary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d, + int numRows, + int numCols, + MatrixOffset& offset); + + /** + * a aggregate expression that apply each row(or column) of matrix b. + * op and sv is element wise operator. + * + * @code + * if (aAsRowVector == true_type && aAsColVector == false_type) + * for each column j & 0 <= i < numRows, do: + * dst = agg(op(b[i*ldb + j])) + * a[j] = sv(a[j], dst) + * + * if (aAsRowVector == false_type && aAsColVector == true_type) + * for each row i & 0 <= j < numCols, do: + * dst = agg(op(b[i*ldb + j])) + * a[i] = sv(a[i], dst) + * @endcode + */ + template + int aggregate(Agg agg, + Op op, + Saver sv, + BaseMatrixT& b, + int numRows, + int numCols, + MatrixOffset& offset, + aAsRowVector, + aAsColVector); + + /** + * a aggregate expression that apply each row(or column) of matrix b and c. + * + * op and sv is element wise operator. + * + * @code + * if (aAsRowVector == true_type && aAsColVector == false_type) + * for each column j & 0 <= i < numRows, do: + * dst = agg(op(b[i*ldb + j], c[i*ldc + j])) + * a[j] = sv(a[j], dst) + * + * if (aAsRowVector == false_type && aAsColVector == true_type) + * for each row i & 0 <= j < numCols, do: + * dst = agg(op(b[i*ldb + j], c[i*ldc + j])) + * a[i] = sv(a[i], dst) + * @endcode + */ + template + int aggregate(Agg agg, + Op op, + Saver sv, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset, + aAsRowVector, + aAsColVector); + + /** + * a aggregate expression that apply each row of matrix b. + * + * @code + * for each row i & 0 <= j < b.width_, do: + * this[i] = agg(b[i*ldb + j]) + * @endcode + */ + template + int applyRow(Agg agg, BaseMatrixT& b); + + /** + * a aggregate expression that apply each row of matrix b. + * + * @code + * for each row i & 0 <= j < b.width_, do: + * dst = agg(op(b[i*ldb + j], c[i*ldc + j]) + * this[i] = sv(this[i], dst) + * @endcode + */ + template + int applyRow(Agg agg, Op op, Saver sv, BaseMatrixT& b, BaseMatrixT& c); + + // Same as the above with the special handing of sv=add2(scaleDest, scaleAgg) + template + int applyRow(Agg agg, + Op op, + real scaleDest, + real scaleAgg, + BaseMatrixT& b, + BaseMatrixT& c); + + /** + * a aggregate expression that apply each row of matrix b. + * + * @code + * for each row i & 0 <= j < b.width_, do: + * dst = agg(b[i*ldb + j]) + * this[i] = sv(this[i], dst) + * @endcode + */ + template + int applyRow(Agg agg, Saver sv, BaseMatrixT& b); + + // Same as the above with the special handing of sv=add2(scaleDest, scaleAgg) + template + int applyRow(Agg agg, real scaleDest, real scaleAgg, BaseMatrixT& b); + + /** + * a aggregate expression that apply each column of matrix b. + * + * @code + * for each column j & 0 <= i < b.height_, do: + * this[j] = agg(b[i*ldb + j]) + * @endcode + */ + template + int applyCol(Agg agg, BaseMatrixT& b); + + /** + * a aggregate expression that apply each column of matrix b. + * + * @code + * for each column j & 0 <= i < b.height_, do: + * dst = agg(b[i*ldb + j]) + * this[j] = sv(this[j], dst) + * @endcode + */ + template + int applyCol(Agg agg, Saver sv, BaseMatrixT& b); + + // Same as the above with the special handing of sv=add2(scaleDest, scaleAgg) + template + int applyCol(Agg agg, real scaleDest, real scaleAgg, BaseMatrixT& b); + + bool useGpu() const { return useGpu_; } + + const T* rowBuf(size_t row) const { return data_ + width_ * row; } + + T* rowBuf(size_t row) { return data_ + width_ * row; } + + /** + * @brief unary operator. + * + */ + void neg(); + void exp2(); + void pow2(T p); + void log2(); + void sqrt2(); + void square2(); + void reciprocal2(); + void abs2(); + void sign2(); + void zero(); + + /** + * @code + * this(row, col + columnOffset) = 0 for 0 <= col < numColumns + * @endcode + */ + void zeroAtOffset(int64_t columnOffset, int64_t numColumns); + void one(); + void subScalar(T p); + void mulScalar(T p); + void divScalar(T p); + + /** + * @code + * this = p + * @endcode + */ + void assign(T p); + + /** + * @code + * swap(this, b) + * example: swap two Matrices + * MatrixPtr cpuA = std::make_shared(height, width); + * MatrixPtr cpuB = std::make_shared(height, width); + * cpuA->deepSwap(*cpuB); + * @endcode + */ + void deepSwap(BaseMatrixT& b); + + /** + * @code + * this = this + p + * @endcode + */ + void add(T p); + + /** + * @code + * this = this*p1 + p2 + * @endcode + */ + void add(T p1, T p2); + + /** + * this = this < low ? low : this + * + * this = this > high ? high : this + */ + void clip(T p1, T p2); + + /** + * this = b < low ? 0 : 1 + * + * this = b > high ? 0 : 1 + */ + void clipDerivative(BaseMatrixT& b, T p1, T p2); + + /** + * @code + * a = a > p ? 1.0f : 0.0f + * @endcode + */ + void biggerThanScalar(T p); + + /** + * @code + * a = a > p ? a : p + * @endcode + */ + void downClip(T p); + + /** + * @code + * this = b + * @endcode + */ + void assign(BaseMatrixT& b); + + /** + * @code + * If b.width + columOffset <= this.width + * this(row, col + columnOffset) = b(row, col) for 0 <= col < b.width + * + * If this.width + columnOffset <= b.width + * this(row, col) = b(row, col + columnOffset) for 0 <= col < this.width + * + * Otherwise, FATAL + * @endcode + */ + void assignAtOffset(BaseMatrixT& b, int64_t columnOffset); + + /// this = this + b + void add(BaseMatrixT& b); + + /** + * @code + * If b.width + columOffset <= this.width + * this(row, col + columnOffset) += b(row, col) for 0 <= col < b.width + * + * If this.width + columnOffset <= b.width + * this(row, col) += b(row, col + columnOffset) for 0 <= col < this.width + * + * Otherwise, FATAL + * @endcode + */ + void addAtOffset(BaseMatrixT& b, int64_t columnOffset); + + void addColVector(BaseMatrixT& b); + void addRowVector(BaseMatrixT& b); + void addBias(BaseMatrixT& b, T scale); + + void mulRowVector(BaseMatrixT& b); + void divRowVector(BaseMatrixT& b); + + void mulColVector(BaseMatrixT& b); + void divColVector(BaseMatrixT& b); + + void addP2P(BaseMatrixT& b); + + /** + * @code + * this = this + b*p + * @endcode + */ + void add(BaseMatrixT& b, T p); + + /** + * @code + * this = p1*this + p2*b + * @endcode + */ + void add(BaseMatrixT& b, T p1, T p2); + + /** + * @code + * this = this - b + * @endcode + */ + void sub(BaseMatrixT& b); + + /** + * @code + * this = this - b*p + * @endcode + */ + void sub(BaseMatrixT& b, T p); + + /** + * @code + * b = max(0, this) + * @endcode + */ + void relu(BaseMatrixT& b); + void reluDerivative(BaseMatrixT& b); + + /** + * @code + * b = log(1.0 + exp(this)) + * @endcode + */ + void softrelu(BaseMatrixT& b); + void softreluDerivative(BaseMatrixT& b); + + /** + * @code + * b = min(max(this, p1), p2) + * @endcode + */ + void brelu(BaseMatrixT& b); + void breluDerivative(BaseMatrixT& b); + + /** + * @code + * b = this * this + * @endcode + */ + void square2(BaseMatrixT& b); + void squareDerivative(BaseMatrixT& b); + + /** + * @code + * b = tanh(this) + * @endcode + */ + void tanh(BaseMatrixT& b); + void tanhDerivative(BaseMatrixT& b); + + /** + * @code + * b = p1 * tanh(p2 * this) + * @endcode + */ + void scaledTanh(BaseMatrixT& b, T p1, T p2); + void scaledTanhDerivative(BaseMatrixT& b, T p1, T p2); + + /** + * @code + * b = 1.0f / this + * @endcode + */ + void reciprocal2(BaseMatrixT& b); + void reciprocalDerivative(BaseMatrixT& b); + + /** + * @code + * b = this > 0.0f ? this : -this + * @endcode + */ + void abs2(BaseMatrixT& b); + void absDerivative(BaseMatrixT& b); + + /** + * @code + * b = 1.0f / (1.0f + exp(-this)) + * @endcode + */ + void sigmoid(BaseMatrixT& b); + void sigmoidDerivative(BaseMatrixT& b); + + /** + * @code + * b = a + * @endcode + */ + void expDerivative(BaseMatrixT& b); + + void sign2(BaseMatrixT& b); + + void exp2(BaseMatrixT& b); + void pow2(BaseMatrixT& b, T p); + void log2(BaseMatrixT& b); + void sqrt2(BaseMatrixT& b); + void addScalar(BaseMatrixT& b, T p); + void subScalar(BaseMatrixT& b, T p); + void mulScalar(BaseMatrixT& b, T p); + void divScalar(BaseMatrixT& b, T p); + void scalarDiv(BaseMatrixT& b, T p); + + /** + * @code + * this = 1.0f / sqrt(b) + * @endcode + */ + void invSqrt(BaseMatrixT& b); + + /// this = (b == value) + void isEqualTo(BaseMatrixT& b, T value); + + /** + * @brief ternary operator. + */ + void softCrossEntropy(BaseMatrixT& b, BaseMatrixT& c); + void softCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c); + void binaryLabelCrossEntropy(BaseMatrixT& b, BaseMatrixT& c); + void binaryLabelCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = b + c + * @endcode + */ + void add(BaseMatrixT& b, BaseMatrixT& c); + /** + * @code + * this = b*p1 + c*p2 + * @endcode + */ + void add(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2); + /** + * @code + * this = b - c + * @endcode + */ + void sub(BaseMatrixT& b, BaseMatrixT& c); + /** + * @code + * this = b*p1 - c*p2 + * @endcode + */ + void sub(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2); + + /** + * @code + * this = this + b + c + * @endcode + */ + void add2(BaseMatrixT& b, BaseMatrixT& c); + /** + * @code + * this = this*p1 + b*p2 + c*p3 + * @endcode + */ + void add2(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3); + + /** + * @code + * this = a*p1 + b*p2 + c*p3 + * @endcode + */ + void add3(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d, T p1, T p2, T p3); + + /** + * @code + * c = p2 * c - p1 * (b + p3 * this) + * this += mom + * @endcode + */ + void sgdUpdate(BaseMatrixT& b, // grad + BaseMatrixT& c, // mom + T p1, // learningRate, + T p2, // momentum, + T p3); // decayRate + + /** + * @code + * c = p2 * c - p1 * d * (b + p3 * this) + * this += mom + * @endcode + */ + void sgdUpdate(BaseMatrixT& b, // grad, + BaseMatrixT& c, // mom, + BaseMatrixT& d, // lr, + T p1, // learningRate, + T p2, // momentum, + T p3); // decayRate + + /// apply L1/L2 to *this* + virtual void applyL1(T learningRate, T decayRate); + void applyL1(BaseMatrixT& lr, T learningRate, T decayRate); + void applyL2(T learningRate, T decayRate); + void applyL2(BaseMatrixT& lr, T learningRate, T decayRate); + + /** + * @code + * this *= b + * @endcode + */ + void dotMul(BaseMatrixT& b); + + /** + * @code + * this = b * c + * @endcode + */ + void dotMul(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = b / c + * @endcode + */ + void dotDiv(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = (b + p1) / (c + p2) + * @endcode + */ + void dotDiv(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); + + /** + * @code + * this = log(1 + exp(b - c)) - d * (b - c) + * @endcode + */ + void rankLoss(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d); + void rankLossBp(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d); + + /** + * @code + * this = log(1 + exp(b)) - c * b + * @endcode + */ + void logisticRegressionLoss(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this += exp(b)/(1+exp(b)) - c + * @endcode + */ + void logisticRegressionLossBp(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = b > c ? 1.0 : 0.0 + * @endcode + */ + void biggerThan(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = ((b>c && d>0.5) || (bc ? b : c + * @endcode + */ + void max2(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this[destCol] += (b>p1 == c>p1) ? 0 : 1) + * @endcode + */ + void binaryClassificationError(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c, + T p); + void binaryClassificationError2(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c, + T p); + + /** + * @code + * this = this * b * b + * @endcode + */ + void dotMulSquare(BaseMatrixT& b); + + /** + * @code + * this = this * this * b + * @endcode + */ + void dotSquareMul(BaseMatrixT& b); + + /** + * @code + * this = b * c * c + * @endcode + */ + void dotMulSquare(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = b * b * c * c + * @endcode + */ + void dotSquareSquare(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = this * (p1*b + p2*c)^2 + * @endcode + */ + void dotMulSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); + + /** + * @code + * this = (p1*b + p2*c)^2 + * @endcode + */ + void dotSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); + + /** + * @code + * this= this * (p1*b + p2*c) + * @endcode + */ + void dotMulSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); + + /** + * @code + * this += sqr(p1*b + p2*c + p3*d) + * @endcode + */ + void addSquareSum( + BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT d, T p1, T p2, T p3); + + /** + * @code + * this += p * sqr(b) + * @endcode + */ + void addSquare(BaseMatrixT& b, T p); + + /** + * @code + * this = p1 * this + p2 * sqr(b) + * @endcode + */ + void decayAddSquare(BaseMatrixT& b, T p1, T p2); + + /** + * @code + * this = p1 * this + p2 * sqr(b * c) + * @endcode + */ + void decayAddSquareMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); + + /** + * @code + * this = 1 / (p1 * b + p2) + * @endcode + */ + void reciprocal2(BaseMatrixT& b, T p1, T p2); + + /** + * @code + * this = 1 / (p1 * b + p2 * c + p3) + * @endcode + */ + void reciprocalSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3); + + /** + * @code + * b = this; this = 0 + * @endcode + */ + void copyAndClear(BaseMatrixT& b); + + /** + * @code + * this_row[destCol] += dotprod(b_row, c_row) + * @endcode + */ + void rowDotMul(size_t destCol, BaseMatrixT& b, BaseMatrixT& c); + void rowDotMul2(size_t destCol, BaseMatrixT& b, BaseMatrixT& c); + + /** + * this is vector (one row matrix) + * + * @code + * for each row i, do: + * this_row += dotmul(b_row_i, c_row_i) + * @endcode + */ + void addDotMulVMM(BaseMatrixT& b, BaseMatrixT& c); + void addDotMulVMM2(BaseMatrixT& b, BaseMatrixT& c); + + /** + * c is vector (one row matrix) + * + * @code + * for each row i, do: + * this_row_i += dotmul(b_row_i, c_row) + * @endcode + */ + void addDotMulMMV(BaseMatrixT& b, BaseMatrixT& c); + void addDotMulMMV2(BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this = p1 * this + p2 * b * c + * @endcode + */ + void addDotMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); + + /** + * @code + * this_row = b_row * c_row[cCol] + * @endcode + */ + void rowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); + void rowScale2(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this_col = b_col * c_col[cRow] + * @endcode + */ + void colScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this_col += b_col * c_col[cRow] + * @endcode + */ + void addColScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c); + + /** + * @code + * this_row += b_row * c_row[cCol] + * @endcode + */ + void addRowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); + + /// calculate the sum of each row of the matrix b. + /// this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} + void sumRows(BaseMatrixT& b, T scaleSum, T scaleDest); + + /// calculate the maximum value of each row of the matrix b. + void maxRows(BaseMatrixT& b); + /// calculate the minimum value of each row of the matrix b. + void minRows(BaseMatrixT& b); + + /// calculate the maximum value of each column of the matrix b. + void maxCols(BaseMatrixT& b); + /// calculate the minimum value of each column of the matrix b. + void minCols(BaseMatrixT& b); + + /// calculate the sum of each column of the matrix b. + /// this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji} + void sumCols(BaseMatrixT& b, T scaleSum, T scaleDest); + + /// this_i = scaleDest * this_i + scaleSum * \sum_j (b_{ij} - c_{ij})^2 + void sumOfSquaredDiffs(BaseMatrixT& b, + BaseMatrixT& c, + T scaleSum, + T scaleDest); + + /// this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij} + void sumOfProducts(BaseMatrixT& b, BaseMatrixT& c, T scaleSum, T scaleDest); + + /** + * @code + * this_row = b_row + p * ones * c_row[cCol] + * @endcode + */ + void rowAdd(size_t cCol, BaseMatrixT& b, BaseMatrixT& c, T p); + /** + * @code + * this_row = pow(b_row, c_row[cCol]) + * @endcode + */ + void rowPow(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); + + virtual bool isSparse() const { return false; } + + template + void operator=(const ExpressionType& expr) { + if (useGpu_) { + TensorGpuApply(*this, expr); + } else { + TensorCpuApply(*this, expr); + } + } + + template + void operator+=(const ExpressionType& expr) { + (*this) = (*this) + expr; + } + template + void operator-=(const ExpressionType& expr) { + (*this) = (*this) - expr; + } + template + void operator*=(const ExpressionType& expr) { + (*this) = (*this) * expr; + } + template + void operator/=(const ExpressionType& expr) { + (*this) = (*this) / expr; + } +}; + +typedef BaseMatrixT BaseMatrix; +typedef BaseMatrixT IBaseMatrix; + +} // namespace paddle diff --git a/paddle/legacy/math/CMakeLists.txt b/paddle/legacy/math/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..9992ec71f45b592e0a73e1cc9c655e773fa18e86 --- /dev/null +++ b/paddle/legacy/math/CMakeLists.txt @@ -0,0 +1,57 @@ +# common package contains: +# * the utilities: +# * Thread Libs +# * Memory Manage libs +# * CommandLine Parser +# * Logging +# * Timer/Stats +# * the math libraries: +# * Matrix/Vector +# * the parameter optimizers. +# * the parameter updater functions. +# +# TODO(yuyang18): separate libs. +# +file(GLOB MATH_HEADERS . *.h) +file(GLOB MATH_SOURCES . *.cpp) + +if(NOT WITH_MKLDNN) + set(DNN_HEADER "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.h") + set(DNN_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.cpp") + list(REMOVE_ITEM MATH_HEADERS "${DNN_HEADER}") + list(REMOVE_ITEM MATH_SOURCES "${DNN_SOURCE}") + message(STATUS "Skip compiling with MKLDNNMatrix") +else() + message(STATUS "Compile with MKLDNNMatrix") +endif() + +if(MOBILE_INFERENCE) + # Remove sparse + list(REMOVE_ITEM MATH_HEADERS + ${CMAKE_CURRENT_SOURCE_DIR}/CpuSparseMatrix.h + ${CMAKE_CURRENT_SOURCE_DIR}/SparseMatrix.h + ${CMAKE_CURRENT_SOURCE_DIR}/SparseRowMatrix.h) + list(REMOVE_ITEM MATH_SOURCES + ${CMAKE_CURRENT_SOURCE_DIR}/CpuSparseMatrix.cpp + ${CMAKE_CURRENT_SOURCE_DIR}/SparseMatrix.cpp + ${CMAKE_CURRENT_SOURCE_DIR}/SparseRowMatrix.cpp) +endif() +set(MATH_SOURCES + "${PADDLE_SOURCE_DIR}/paddle/legacy/math/BaseMatrix.cu" + "${PADDLE_SOURCE_DIR}/paddle/legacy/math/TrainingAlgorithmOp.cu" + ${MATH_SOURCES}) +if(NOT WITH_GPU) + # then compile BaseMatrix.cu as c++ file + compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/legacy/math/BaseMatrix.cu") + compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/legacy/math/TrainingAlgorithmOp.cu") + add_library(paddle_math STATIC + ${MATH_SOURCES}) +else() + cuda_add_library(paddle_math ${MATH_SOURCES}) +endif() + + +add_dependencies(paddle_math paddle_proto ${external_project_dependencies}) # depends +if(WITH_TESTING) + add_subdirectory(tests) +endif() diff --git a/paddle/legacy/math/CpuSparseMatrix.cpp b/paddle/legacy/math/CpuSparseMatrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..20c65a3a1d7099a73d8b3c490cd42e721e60823b --- /dev/null +++ b/paddle/legacy/math/CpuSparseMatrix.cpp @@ -0,0 +1,787 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CpuSparseMatrix.h" +#include "SparseMatrix.h" +#include "float.h" +#include "hl_gpu.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +const size_t CpuSparseMatrix::DEFAULT_AVG_WIDTH; + +CpuSparseMatrix::CpuSparseMatrix(size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans) + : Matrix(NULL, height, width, trans, false) { + resize(height, width, nnz, valueType, format); +} + +CpuSparseMatrix::CpuSparseMatrix(CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans) + : Matrix(dataHandle, height, width, trans, false) { + resize(height, width, nnz, valueType, format); +} + +CpuSparseMatrix::CpuSparseMatrix(real* data, + int* rows, + int* cols, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans) + : Matrix(NULL, height, width, trans, false) { + cols_ = cols; + rows_ = rows; + value_ = data; + height_ = height; + width_ = width; + elementCnt_ = nnz; + valueType_ = valueType; + format_ = format; +} + +void CpuSparseMatrix::resize(size_t newHeight, + size_t newWidth, + size_t newNnz, + SparseValueType valueType, + SparseFormat format) { + CHECK_LE(newNnz, newHeight * newWidth); + size_t newSize = 0; + if (format == SPARSE_CSR) { + newSize = (newHeight + 1) * sizeof(int) + newNnz * sizeof(int); + } else { + newSize = (newWidth + 1) * sizeof(int) + newNnz * sizeof(int); + } + + if (NO_VALUE != valueType) { + newSize += newNnz * sizeof(real); + } + + if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) { + memoryHandle_ = std::make_shared(newSize); + } + + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newNnz; + valueType_ = valueType; + format_ = format; + sparseResize(); +} +void CpuSparseMatrix::sparseResize() { + if (format_ == SPARSE_CSR) { + rows_ = reinterpret_cast( + reinterpret_cast(memoryHandle_->getBuf())); + cols_ = reinterpret_cast( + reinterpret_cast(memoryHandle_->getBuf()) + + (height_ + 1) * sizeof(int)); + if (NO_VALUE != valueType_) { + value_ = reinterpret_cast( + reinterpret_cast(memoryHandle_->getBuf()) + + (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); + } else { + value_ = NULL; + } + } else { + cols_ = reinterpret_cast( + reinterpret_cast(memoryHandle_->getBuf())); + rows_ = reinterpret_cast( + reinterpret_cast(memoryHandle_->getBuf()) + + (width_ + 1) * sizeof(int)); + if (NO_VALUE != valueType_) { + value_ = reinterpret_cast( + reinterpret_cast(memoryHandle_->getBuf()) + + (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); + } else { + value_ = NULL; + } + } +} + +void CpuSparseMatrix::resize(size_t newHeight, size_t newWidth) { + resize(newHeight, + newWidth, + newHeight * std::min(DEFAULT_AVG_WIDTH, newWidth), + valueType_, + format_); +} + +MatrixPtr CpuSparseMatrix::getTranspose() { + if (!memoryHandle_ && !value_) { + MatrixPtr dest(new CpuSparseMatrix( + height_, width_, elementCnt_, valueType_, format_, true)); + return dest; + } else if (memoryHandle_) { + MatrixPtr dest(new CpuSparseMatrix( + std::dynamic_pointer_cast(memoryHandle_), + height_, + width_, + elementCnt_, + valueType_, + format_, + true)); + return dest; + } else if (value_) { + MatrixPtr dest(new CpuSparseMatrix(value_, + rows_, + cols_, + height_, + width_, + elementCnt_, + valueType_, + format_, + true)); + return dest; + } else { + return NULL; + } +} + +SparseValueType CpuSparseMatrix::getValueType() { return valueType_; } + +void CpuSparseMatrix::mul(const Matrix& a, + const Matrix& b, + real scaleAB, + real scaleT) { + CHECK(!isTransposed()) << "Not supported"; + const auto a_ptr = dynamic_cast(&a); + const auto b_ptr = dynamic_cast(&b); + + if (a_ptr && b_ptr) { + CpuMatrix::mul((CpuMatrix*)a_ptr, (CpuMatrix*)b_ptr, this, scaleAB, scaleT); + } else { + LOG(FATAL) << "not supported"; + } +} + +void CpuSparseMatrix::add3(CpuMatrix* b) { + CHECK(getFormat() != SPARSE_CSC) << "Not supported"; + CHECK(height_ == b->getHeight()); + CHECK(width_ == b->getWidth()); + real* A = getValue(); + real* B = b->getData(); + int* cols = getCols(); + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + A[j] = B[i * width_ + cols[j]]; + } + } +} + +void CpuSparseMatrix::add3(MatrixPtr b) { + if (dynamic_cast(b.get())) { + add3(dynamic_cast(b.get())); + } else { + LOG(FATAL) << "not supported"; + } +} + +void CpuSparseMatrix::addBias(Matrix& b, real scale) { + CHECK_EQ(b.getHeight(), (size_t)1); + CHECK_EQ(width_, b.getWidth()); + real* A = getValue(); + real* B = b.getData(); + int* cols = getCols(); + size_t nnz = getElementCnt(); + for (size_t i = 0; i < nnz; i++) { + A[i] += scale * B[cols[i]]; + } +} + +template +void printBuf(std::ostream& os, T* a, size_t len, const char* name) { + os << "\n: " << name << " ["; + for (size_t i = 0; i < len; i++) { + os << a[i] << " "; + } + os << "]\n"; +} + +void CpuSparseMatrix::print(std::ostream& os) const { + size_t rowSize = format_ == SPARSE_CSC ? elementCnt_ : height_ + 1; + size_t colSize = format_ == SPARSE_CSC ? width_ + 1 : elementCnt_; + printBuf(os, rows_, rowSize, "row"); + printBuf(os, cols_, colSize, "col"); + if (valueType_ == FLOAT_VALUE) { + printBuf(os, value_, elementCnt_, "value"); + } + return; +} + +void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const { + CHECK_LT(idx, height_); + if (format_ == SPARSE_CSC) { + LOG(FATAL) << "SPARSE_CSC not supported"; + return; + } + + const int* col = getRowCols(idx); + size_t num = getColNum(idx); + if (num > 0) { + if (valueType_ == FLOAT_VALUE) { + const real* data = getRowValues(idx); + os << col[0] << ":" << data[0]; + for (size_t i = 1; i < num; ++i) { + os << " " << col[i] << ":" << data[i]; + } + } else { + os << col[0]; + for (size_t i = 1; i < num; ++i) { + os << " " << col[i]; + } + } + } + os << ";"; +} + +void CpuSparseMatrix::rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c) { + CHECK(getFormat() != SPARSE_CSC) << "Not supported"; + CHECK_EQ(height_, b.getHeight()); + CHECK_EQ(width_, b.getWidth()); + real* A = getValue(); + real* B = b.getValue(); + if (b.getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + CHECK_EQ(start, b.getRowStartIdx(i)); + CHECK_EQ(end, b.getRowStartIdx(i + 1)); + for (size_t j = start; j < end; j++) { + A[j] = B[j] * c.getElement(i, cCol); + } + } + } else if (b.getValueType() == NO_VALUE) { + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + CHECK_EQ(start, b.getRowStartIdx(i)); + CHECK_EQ(end, b.getRowStartIdx(i + 1)); + for (size_t j = start; j < end; j++) { + A[j] = c.getElement(i, cCol); + } + } + } +} + +void CpuSparseMatrix::randomizeUniform() { + CHECK_LE(elementCnt_, height_ * width_); + if (valueType_ == FLOAT_VALUE) { + real* data = getValue(); + for (size_t i = 0; i < elementCnt_; ++i) { + *data++ = rand() / static_cast(RAND_MAX); // NOLINT + } + } + if (format_ == SPARSE_CSR) { + sparseRand(rows_, cols_, elementCnt_, height_ + 1, width_, false); + } else { + sparseRand(cols_, rows_, elementCnt_, width_ + 1, height_, false); + } +} + +void CpuSparseMatrix::copyFrom(std::vector& rows, + std::vector& cols, + std::vector& values) { + size_t size = format_ == SPARSE_CSR ? cols.size() : rows.size(); + resize(height_, width_, size, valueType_, format_); + if (valueType_ == FLOAT_VALUE) { + memcpy(&value_[0], &values[0], sizeof(real) * values.size()); + } + memcpy(&cols_[0], &cols[0], sizeof(int) * cols.size()); + memcpy(&rows_[0], &rows[0], sizeof(int) * rows.size()); +} + +// Copy from a CpuMatrix, only supported in sparse_float_value_t +// SparseMatrix. +void CpuSparseMatrix::copyFrom(const CpuMatrix& src) { + CHECK_EQ(getHeight(), src.getHeight()); + CHECK_EQ(getWidth(), src.getWidth()); + CHECK(!src.trans_ && !trans_); + if (format_ == SPARSE_CSR) { + std::vector rows(getHeight() + 1); + std::vector cols; + std::vector values; + rows[0] = 0; + for (size_t r = 0; r < getHeight(); ++r) { + for (size_t c = 0; c < getWidth(); ++c) { + real v = src.getElement(r, c); + if (fabs(v) > FLT_EPSILON) { + cols.push_back(c); + values.push_back(v); + } + } + rows[r + 1] = values.size(); + } + copyFrom(rows, cols, values); + } else { + std::vector cols(getWidth() + 1); + std::vector rows; + std::vector values; + cols[0] = 0; + for (size_t r = 0; r < getWidth(); ++r) { + for (size_t c = 0; c < getHeight(); ++c) { + real v = src.getElement(c, r); + if (fabs(v) > FLT_EPSILON) { + rows.push_back(c); + values.push_back(v); + } + } + cols[r + 1] = values.size(); + } + copyFrom(rows, cols, values); + } +} + +MatrixPtr CpuSparseMatrix::clone(size_t height, size_t width, bool useGpu) { + if (height == 0 && width == 0) { + height = height_; + width = width_; + } + CHECK(width && height); + if (!useGpu) { + return std::make_shared( + height, width, 0, valueType_, format_); + } else { + return std::make_shared( + height, width, elementCnt_, valueType_, format_); + } +} + +MatrixPtr CpuSparseMatrix::subMatrix(size_t startRow, size_t numRows) { + CHECK_LE(startRow + numRows, height_); + CHECK_EQ(format_, SPARSE_CSR); + if (valueType_ == NO_VALUE) { + return std::make_shared( + nullptr, + rows_ + startRow, + cols_, + numRows, + width_, + rows_[startRow + numRows] - rows_[startRow], + valueType_, + format_, + trans_); + } else { + return std::make_shared( + value_, + rows_ + startRow, + cols_, + numRows, + width_, + rows_[startRow + numRows] - rows_[startRow], + valueType_, + format_, + trans_); + } +} + +/* mem MUST be alloced outside (memAlloc=false) */ +void CpuSparseMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { + CHECK(!memAlloc); + CpuSparseMatrix* mat = dynamic_cast(matTrans.get()); + if (format_ == SPARSE_CSR) { + /*statistic element number in each col*/ + int* colCounters = mat->getRows() + 1; + memset(colCounters, 0, sizeof(int) * width_); + for (size_t i = 0; i < elementCnt_; ++i) { + int col = cols_[i]; + colCounters[col]++; + } + /*fill mat rows */ + mat->getRows()[0] = 0; + for (size_t i = 1; i < width_ + 1; i++) { + mat->getRows()[i] = mat->getRows()[i - 1] + mat->getRows()[i]; + } + /*fill mat values and cols*/ + std::vector colNumVec(width_, 0); + if (valueType_ == FLOAT_VALUE) { + for (size_t i = 0; i < height_; i++) { + for (int j = rows_[i]; j < rows_[i + 1]; j++) { + int colIdx = cols_[j]; + int index = mat->getRows()[colIdx] + colNumVec[colIdx]; + mat->getCols()[index] = i; + mat->getValue()[index] = value_[j]; + colNumVec[colIdx]++; + } + } + } else { + for (size_t i = 0; i < height_; i++) { + for (int j = rows_[i]; j < rows_[i + 1]; j++) { + int colIdx = cols_[j]; + int index = mat->getRows()[colIdx] + colNumVec[colIdx]; + mat->getCols()[index] = i; + colNumVec[colIdx]++; + } + } + } + } else { + /*statistic element number in each row*/ + int* rowCounters = mat->getCols() + 1; + memset(rowCounters, 0, sizeof(int) * height_); + for (size_t i = 0; i < elementCnt_; ++i) { + int row = rows_[i]; + rowCounters[row]++; + } + + /*fill mat cols */ + mat->getCols()[0] = 0; + for (size_t i = 1; i < height_ + 1; i++) { + mat->getCols()[i] = mat->getCols()[i - 1] + mat->getCols()[i]; + } + /*fill mat values and rows*/ + std::vector rowNumVec(height_, 0); + if (valueType_ == FLOAT_VALUE) { + for (size_t i = 0; i < width_; i++) { + for (int j = cols_[i]; j < cols_[i + 1]; j++) { + int rowIdx = rows_[j]; + int index = mat->getCols()[rowIdx] + rowNumVec[rowIdx]; + mat->getRows()[index] = i; + mat->getValue()[index] = value_[j]; + rowNumVec[rowIdx]++; + } + } + } else { + for (size_t i = 0; i < width_; i++) { + for (int j = cols_[i]; j < cols_[i + 1]; j++) { + int rowIdx = rows_[j]; + int index = mat->getCols()[rowIdx] + rowNumVec[rowIdx]; + mat->getRows()[index] = i; + rowNumVec[rowIdx]++; + } + } + } + } +} + +void CpuSparseMatrix::setRow(size_t row, + size_t colNum, + const unsigned int* cols, + const real* values) { + if (format_ == SPARSE_CSR) { + CHECK_LT(row, height_); + CHECK(NULL != cols); + if (0 == row) { + rows_[row] = 0; + } + rows_[row + 1] = rows_[row] + colNum; + for (size_t i = 0; i < colNum; ++i) { + cols_[rows_[row] + i] = cols[i]; + } + if (valueType_ == NO_VALUE) { + CHECK(!values); + } else { + for (size_t i = 0; i < colNum; ++i) { + value_[rows_[row] + i] = values[i]; + } + } + } else { + LOG(FATAL) << "not supported"; + } +} + +void CpuSparseMatrix::fillRowIndices(IVectorPtr& outVec) const { + if (format_ == SPARSE_CSR) { + auto nnz = getElementCnt(); + IVector::resizeOrCreate(outVec, nnz, false); + auto out = outVec->getData(); + int* rows = getRows(); + for (size_t i = 0; i < height_; i++) { + for (int j = rows[i]; j < rows[i + 1]; j++) { + out[j] = i; + } + } + } else { + LOG(FATAL) << "SPARSE_CSC not supported"; + } +} + +ThreadLocal> CpuSparseMatrix::cpuLocalMats_; + +CpuSparseMatrixPtr CpuSparseMatrix::getTmpSparseMatrix(size_t height, + size_t width) { + std::vector* localMats = cpuLocalMats_.get(); + auto it = localMats->begin(); + while (it != localMats->end()) { + if (it->unique()) { + (*it)->resize(height, width, elementCnt_, valueType_, format_); + return *it; + } + } + localMats->emplace_back(std::make_shared( + height, width, elementCnt_, valueType_, format_, false)); + return localMats->back(); +} + +void CpuSparseMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { + if (dynamic_cast(&src)) { + auto tmpSrc = dynamic_cast(&src); + copyFrom(*tmpSrc, stream); + } else if (dynamic_cast(&src)) { + auto tmpSrc = dynamic_cast(&src); + copyFrom(*tmpSrc); + } else if (dynamic_cast(&src)) { + auto tmpSrc = dynamic_cast(&src); + copyFrom(*tmpSrc); + } else { + LOG(FATAL) << "not implemented"; + } +} + +void CpuSparseMatrix::copyFrom(const Matrix& src) { + if (dynamic_cast(&src)) { + auto tmpSrc = dynamic_cast(&src); + copyFrom(*tmpSrc); + } else if (dynamic_cast(&src)) { + auto tmpSrc = dynamic_cast(&src); + copyFrom(*tmpSrc); + } else { + LOG(FATAL) << "not implemented"; + } +} + +void CpuSparseMatrix::copyFrom(const GpuSparseMatrix& src, hl_stream_t stream) { + CHECK_EQ(height_, src.getHeight()); + CHECK_EQ(width_, src.getWidth()); + CHECK_EQ(size_t(elementCnt_), src.getElementCnt()); + size_t valSize = valueType_ == NO_VALUE ? 0 : elementCnt_; + if (format_ == SPARSE_CSC) + hl_memcpy_from_csc_matrix(value_, + valSize, + rows_, + elementCnt_, + cols_, + width_ + 1, + src.sMatrix_.get(), + stream); + else + hl_memcpy_from_csr_matrix(value_, + valSize, + rows_, + height_ + 1, + cols_, + elementCnt_, + src.sMatrix_.get(), + stream); +} + +void CpuSparseMatrix::copyFrom(const CpuSparseMatrix& src) { + CHECK_EQ(height_, src.getHeight()); + CHECK_EQ(width_, src.getWidth()); + CHECK_EQ(format_, src.getFormat()); + int start = format_ == SPARSE_CSR ? src.getRows()[0] : src.getCols()[0]; + if (format_ == SPARSE_CSR) { + size_t totalColNum = 0; + for (size_t i = 0; i < height_; ++i) { + totalColNum += src.getColNum(i); + } + resize(height_, width_, totalColNum, valueType_, format_); + rows_[0] = 0; + for (size_t i = 0; i < height_; ++i) { + rows_[i + 1] = rows_[i] + src.getColNum(i); + } + memcpy(cols_, src.getCols() + start, totalColNum * sizeof(int)); + } else { + size_t totalColNum = 0; + for (size_t i = 0; i < width_; ++i) { + totalColNum += src.getRowNum(i); + } + resize(height_, width_, totalColNum, valueType_, format_); + cols_[0] = 0; + for (size_t i = 0; i < width_; ++i) { + cols_[i + 1] = cols_[i] + src.getRowNum(i); + } + memcpy(rows_, src.getRows() + start, totalColNum * sizeof(int)); + } + + // if have different value type, only copy rows and cols + if (valueType_ == FLOAT_VALUE && src.getValueType() == FLOAT_VALUE) { + memcpy(value_, src.getValue() + start, elementCnt_ * sizeof(real)); + } +} + +void CpuSparseMatrix::copyRow(int offsets, + size_t colNum, + const sparse_non_value_t* row) { + for (size_t j = 0; j < colNum; j++) { + cols_[offsets + j] = row[j].col; + } +} + +void CpuSparseMatrix::copyRow(int offsets, + size_t colNum, + const sparse_float_value_t* row) { + for (size_t j = 0; j < colNum; j++) { + cols_[offsets + j] = row[j].col; + value_[offsets + j] = row[j].value; + } +} + +template +void CpuSparseMatrix::copyFrom(int64_t* ids, int64_t* indices, T* data) { + size_t totalColNum = 0; + for (size_t i = 0; i < height_; ++i) { + int64_t id = ids[i]; + totalColNum += indices[id + 1] - indices[id]; + } + valueType_ = typeid(T) == typeid(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE; + + resize(height_, width_, totalColNum, valueType_, format_); + + rows_[0] = 0; + for (size_t i = 0; i < height_; ++i) { + int64_t id = ids[i]; + T* row = data + indices[id]; + size_t colNum = indices[id + 1] - indices[id]; + rows_[i + 1] = rows_[i] + colNum; + copyRow(rows_[i], colNum, row); + } +} + +template +void CpuSparseMatrix::copyFrom(int64_t* indices, T* data) { + CHECK(format_ == SPARSE_CSR); + size_t totalColNum = indices[height_] - indices[0]; + valueType_ = typeid(T) == typeid(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE; + resize(height_, width_, totalColNum, valueType_, format_); + + rows_[0] = 0; + for (size_t i = 0; i < height_; ++i) { + T* row = data + indices[i]; + size_t colNum = indices[i + 1] - indices[i]; + rows_[i + 1] = rows_[i] + colNum; + copyRow(rows_[i], colNum, row); + } +} + +void CpuSparseMatrix::trimFrom(const CpuSparseMatrix& src) { + CHECK_EQ(height_, src.getHeight()); + CHECK_LE(width_, src.getWidth()); + CHECK_EQ(format_, src.getFormat()); + CHECK_EQ(valueType_, src.getValueType()); + if (format_ == SPARSE_CSR) { + int* srcCols = src.getCols(); + size_t numLessWidth = + std::count_if(srcCols, srcCols + src.getElementCnt(), [this](size_t n) { + return n < this->width_; + }); + resize(height_, width_, numLessWidth, valueType_, format_); + rows_[0] = 0; + size_t index = 0; + for (size_t r = 0; r < height_; ++r) { + for (int i = src.getRows()[r]; i < src.getRows()[r + 1]; ++i) { + if (srcCols[i] < static_cast(width_)) { + cols_[index] = srcCols[i]; + if (valueType_ == FLOAT_VALUE) { + value_[index] = src.getValue()[i]; + } + ++index; + } + } + rows_[r + 1] = index; + } + CHECK_EQ(index, numLessWidth); + } else { + size_t numLessWidth = src.getCols()[width_] - src.getCols()[0]; + resize(height_, width_, numLessWidth, valueType_, format_); + cols_[0] = 0; + size_t index = 0; + // note: c < width_, not src.getWidth(); + for (size_t c = 0; c < width_; ++c) { + for (int i = src.getCols()[c]; i < src.getCols()[c + 1]; ++i) { + rows_[index] = src.getRows()[i]; + if (valueType_ == FLOAT_VALUE) { + value_[index] = src.getValue()[i]; + } + ++index; + } + cols_[c + 1] = index; + } + CHECK_EQ(index, numLessWidth); + } +} + +void CpuSparseMatrix::zeroMem() { + CHECK(valueType_ == FLOAT_VALUE); + memset(value_, 0, elementCnt_ * sizeof(real)); +} + +template void CpuSparseMatrix::copyFrom(int64_t* ids, + int64_t* indices, + sparse_non_value_t* data); + +template void CpuSparseMatrix::copyFrom(int64_t* ids, + int64_t* indices, + sparse_float_value_t* data); + +template void CpuSparseMatrix::copyFrom(int64_t* indices, + sparse_non_value_t* data); + +template void CpuSparseMatrix::copyFrom(int64_t* indices, + sparse_float_value_t* data); + +void CpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { + size_t numSamples = getHeight(); + size_t beam = maxVal.getWidth(); + CHECK_EQ(maxIds.getSize(), numSamples * beam); + CHECK_EQ(maxVal.getHeight(), numSamples); + maxVal.zeroMem(); + int* outids = maxIds.getData(); + real* outvalues = maxVal.getData(); + + typedef std::pair valuepair; + std::vector vec; + for (size_t i = 0; i < numSamples; i++) { + vec.clear(); + + auto num = getColNum(i); + auto ids = getRowCols(i); + auto values = getRowValues(i); + for (size_t j = 0; j < num; j++) { + vec.push_back(std::make_pair(values[j], ids[j])); + } + + size_t outsize = std::min(num, beam); + std::partial_sort(vec.begin(), + vec.begin() + outsize, + vec.end(), + [](const valuepair& a, const valuepair& b) { + return a.first > b.first; + }); + for (size_t j = 0; j < outsize; j++) { + outids[i * beam + j] = vec[j].second; + outvalues[i * beam + j] = vec[j].first; + } + if (outsize < beam) { + // if the number of values to sort are less than the output size, + // use -1 to indicate the end of valid sorted values. + outids[i * beam + outsize] = -1; + } + } +} + +} // namespace paddle diff --git a/paddle/math/CpuSparseMatrix.h b/paddle/legacy/math/CpuSparseMatrix.h similarity index 100% rename from paddle/math/CpuSparseMatrix.h rename to paddle/legacy/math/CpuSparseMatrix.h diff --git a/paddle/math/ExecViaCpu.h b/paddle/legacy/math/ExecViaCpu.h similarity index 100% rename from paddle/math/ExecViaCpu.h rename to paddle/legacy/math/ExecViaCpu.h diff --git a/paddle/math/MKLDNNMatrix.cpp b/paddle/legacy/math/MKLDNNMatrix.cpp similarity index 100% rename from paddle/math/MKLDNNMatrix.cpp rename to paddle/legacy/math/MKLDNNMatrix.cpp diff --git a/paddle/legacy/math/MKLDNNMatrix.h b/paddle/legacy/math/MKLDNNMatrix.h new file mode 100644 index 0000000000000000000000000000000000000000..5a0e5f85923dfd822dad4c63679acde63719f217 --- /dev/null +++ b/paddle/legacy/math/MKLDNNMatrix.h @@ -0,0 +1,256 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Matrix.h" +#include "mkldnn.hpp" +#include "paddle/legacy/parameter/Parameter.h" + +namespace paddle { + +class MKLDNNMatrix; +typedef std::shared_ptr MKLDNNMatrixPtr; + +#define CHECK_PRIMITIVE_DESC_EQ(MAT, PD, ...) \ + CHECK(MAT) << " can not be empty."; \ + CHECK(MAT->getPrimitiveDesc() == PD) \ + << #MAT "->getPrimitiveDesc() and " #PD " should be equal.\n " \ + << "" __VA_ARGS__; + +/** + * @brief MKLDNN Matrix. + * + */ +class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory { + public: + MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd) + : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false), + mkldnn::memory(pd, m->getData()), + m_(m) {} + + ~MKLDNNMatrix() {} + + /** + * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc + */ + static MKLDNNMatrixPtr create(mkldnn::memory::primitive_desc pd, + MatrixPtr m = nullptr); + + /** + * Create MKLDNNMatrix from a MatrixPtr and memory details info + */ + static MKLDNNMatrixPtr create( + mkldnn::memory::dims dims, + mkldnn::memory::format fmt, + mkldnn::engine& eg, + MatrixPtr m = nullptr, + mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32); + + /** + * Create primitive descriptor. + * default with f32 dtype + */ + static mkldnn::memory::primitive_desc createPrimitiveDesc( + const mkldnn::memory::dims dims, + const mkldnn::memory::format& fmt, + const mkldnn::engine& eg, + const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) { + return mkldnn::memory::primitive_desc(memory::desc(dims, dtype, fmt), eg); + } + + /** + * Create Memory descriptor. + * default with any format and f32 dtype + */ + static mkldnn::memory::desc createMemoryDesc( + const mkldnn::memory::dims dims, + const mkldnn::memory::format& fmt = mkldnn::memory::format::any, + const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) { + return mkldnn::memory::desc(dims, dtype, fmt); + } + + /** + * Create reorder primitive. + * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst. + * checkData: whether to check the data handle of src and dst. + * if true, it will check the data and do not allow them equal; + * otherwise, it will not check them, then the reorder created + * may have inplace buffer. + * Do not set false, if you can not guarantee the inplace logical + * would work with your reorder. + */ + static std::shared_ptr createReorder( + const MKLDNNMatrixPtr& src, + const MKLDNNMatrixPtr& dst, + bool checkData = true); + + void copyFrom(const Matrix& src) { + // TODO(TJ): reorder data if this format is not nchw or x + m_->copyFrom(src); + } + + void copyTo(Matrix& dst) { + // TODO(TJ): reorder data if this format is not nchw or x + dst.copyFrom(*m_); + } + + public: + /** + * Reorder this MKLDNNMatrix from other format. + * Support inplace reorder. + * @note: this function would only reorder the data layout. + * will NOT change this original dim or format info + */ + void reorderDataFrom(const MKLDNNMatrixPtr& m, + memory::format srcFmt, + memory::dims targetDim); + + /** + * Reorder this MKLDNNMatrix to other format. + * Support inplace reorder. + * @note: this function would only reorder the data layout. + * will NOT change the dst dim or format info + */ + void reorderDataTo(const MKLDNNMatrixPtr& m, + memory::format dstFmt, + memory::dims targetDim); + + /** + * Dimensionality reduction. + * Change format "nchw --> nc" or "oihw --> oi" if the h and w are both 1 + */ + void downSpatial(); + + /** + * set the memory data handle. + * Caution: This will not check the buffer size of the data, + * it should be coverd by user. + */ + void setData(real* data) { + set_data_handle(data); + CpuMatrix::setData(data); + m_.reset(); + } + + /** + * override the CpuMatrix::resize + */ + void resize(size_t newHeight, size_t newWidth) override { + m_->resize(newHeight, newWidth); + if (data_ == m_->getData() && elementCnt_ == newHeight * newWidth) { + return; + } + CpuMatrix::setData(data_); + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newHeight * newWidth; + stride_ = width_; + auto pd = mkldnn::memory::primitive_desc( + mkldnn::memory::desc({(int)newHeight, (int)newWidth}, + getDtype(), + mkldnn::memory::format::nc), + getEngine()); + resetMKLDNNMemory(pd, data_); + } + + /** + * override Matrix::getData + * check data before return + */ + real* getData() override { + CHECK_EQ((void*)data_, get_data_handle()); + return data_; + } + + const real* getData() const override { + CHECK_EQ((void*)data_, get_data_handle()); + return data_; + } + + /** + * Get primitive descriptor. + */ + mkldnn::memory::primitive_desc getPrimitiveDesc() { + return this->get_primitive_desc(); + } + + /** + * Get memory descriptor. + */ + mkldnn::memory::desc getMemoryDesc() { return getPrimitiveDesc().desc(); } + + /** + * Get dimensions. + */ + mkldnn::memory::dims getDims() { + mkldnn::memory::desc md = getMemoryDesc(); + const int* src = md.data.dims; + int ndims = md.data.ndims; + mkldnn::memory::dims dst; + dst.resize(ndims); + for (int i = 0; i < ndims; ++i) { + dst[i] = src[i]; + } + return dst; + } + + /** + * Get format. + */ + mkldnn::memory::format getFormat() { + return (mkldnn::memory::format)(getMemoryDesc().data.format); + } + + /** + * Get memory data type. + */ + mkldnn::memory::data_type getDtype() { + return (mkldnn::memory::data_type)(getMemoryDesc().data.data_type); + } + + /** + * Get engine. + */ + mkldnn::engine getEngine() { return getPrimitiveDesc().get_engine(); } + + protected: + /** + * Do reorder once. + * Can support inplace. + */ + void reorderOnce(void* srcData, + void* dstData, + memory::format srcFmt, + memory::format dstFmt, + memory::dims dm); + /** + * reset this MKLDNN Memory from primitve desc + */ + void resetMKLDNNMemory(memory::primitive_desc pd, real* data) { + mkldnn_primitive_t result; + mkldnn::error::wrap_c_api( + mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr), + "could not create a memory primitive"); + reset(result); + set_data_handle(data); + } + + private: + // save the CpuMatrixPtr in case the buffer released outside + CpuMatrixPtr m_; +}; + +} // namespace paddle diff --git a/paddle/legacy/math/MathFunctions.cpp b/paddle/legacy/math/MathFunctions.cpp new file mode 100644 index 0000000000000000000000000000000000000000..bbf34a32f36fa7988058f8d3bb7f91eaf2bc1ba0 --- /dev/null +++ b/paddle/legacy/math/MathFunctions.cpp @@ -0,0 +1,348 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/math/MathFunctions.h" +#include "hl_matrix_apply.cuh" +#include "hl_matrix_ops.cuh" +#include "paddle/legacy/utils/DynamicLoader.h" + +namespace dynload { + +std::once_flag lapack_dso_flag; +void* lapack_dso_handle = nullptr; + +/** + * The following macro definition can generate structs + * (for each function) to dynamic load lapack routine + * via operator overloading. + * + * note: default dynamic linked libs + */ + +// The argument for stringizing operator is not macro-expanded first. +// We have to use two levels of macro to do the expansion. +// See https://gcc.gnu.org/onlinedocs/cpp/Stringizing.html +#define STR(x) #x + +// clang-format off +#ifndef LAPACK_FOUND +#define DYNAMIC_LOAD_LAPACK_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + using lapack_func = decltype(__name(args...)) (*)(Args...); \ + std::call_once(lapack_dso_flag, GetLapackDsoHandle, &lapack_dso_handle); \ + void* p_##__name = dlsym(lapack_dso_handle, STR(__name)); \ + CHECK(p_##__name) << "Cannot find symbol " << STR(__name) \ + << " in liblapack.so"; \ + return reinterpret_cast(p_##__name)(args...); \ + } \ + } __name; // struct DynLoad__##__name +#else +#define DYNAMIC_LOAD_LAPACK_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + return __name(args...); \ + } \ + } __name; // struct DynLoad__##__name +#endif + +#define PADDLE_SGETRF LAPACKE_sgetrf +#define PADDLE_DGETRF LAPACKE_dgetrf +#define PADDLE_SGETRI LAPACKE_sgetri +#define PADDLE_DGETRI LAPACKE_dgetri + +#define LAPACK_ROUTINE_EACH(__macro) \ + __macro(PADDLE_SGETRF) \ + __macro(PADDLE_DGETRF) \ + __macro(PADDLE_SGETRI) \ + __macro(PADDLE_DGETRI) +// clang-format on + +LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP) + +} // namespace dynload + +namespace paddle { + +#ifndef PADDLE_USE_EIGEN_FOR_BLAS +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, + const int M, + const int N, + const int K, + const float alpha, + const float* A, + const int lda, + const float* B, + const int ldb, + const float beta, + float* C, + const int ldc) { + cblas_sgemm(CblasRowMajor, + transA, + transB, + M, + N, + K, + alpha, + A, + lda, + B, + ldb, + beta, + C, + ldc); +} + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, + const int M, + const int N, + const int K, + const double alpha, + const double* A, + const int lda, + const double* B, + const int ldb, + const double beta, + double* C, + const int ldc) { + cblas_dgemm(CblasRowMajor, + transA, + transB, + M, + N, + K, + alpha, + A, + lda, + B, + ldb, + beta, + C, + ldc); +} +#endif + +template <> +int getrf(const CBLAS_ORDER order, + const int M, + const int N, + float* A, + const int lda, + int* ipiv) { + return dynload::PADDLE_SGETRF(order, M, N, A, lda, ipiv); +} + +template <> +int getrf(const CBLAS_ORDER order, + const int M, + const int N, + double* A, + const int lda, + int* ipiv) { + return dynload::PADDLE_DGETRF(order, M, N, A, lda, ipiv); +} + +template <> +int getri(const CBLAS_ORDER order, + const int N, + float* A, + const int lda, + const int* ipiv) { + return dynload::PADDLE_SGETRI(order, N, A, lda, ipiv); +} + +template <> +int getri(const CBLAS_ORDER order, + const int N, + double* A, + const int lda, + const int* ipiv) { + return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv); +} + +#ifndef PADDLE_USE_EIGEN_FOR_BLAS +template <> +void axpy(const int n, const float alpha, const float* x, float* y) { + cblas_saxpy(n, alpha, x, 1, y, 1); +} + +template <> +void axpy(const int n, const double alpha, const double* x, double* y) { + cblas_daxpy(n, alpha, x, 1, y, 1); +} + +template <> +float dotProduct(const int n, const float* x, const float* y) { + return cblas_sdot(n, x, 1, y, 1); +} + +template <> +double dotProduct(const int n, const double* x, const double* y) { + return cblas_ddot(n, x, 1, y, 1); +} +#endif + +#if defined(PADDLE_WITH_MKLML) + +template <> +void vExp(const int n, const float* a, float* r) { + vsExp(n, a, r); +} + +template <> +void vExp(const int n, const double* a, double* r) { + vdExp(n, a, r); +} + +template <> +void vPow(const int n, const float* a, const float b, float* r) { + vsPowx(n, a, b, r); +} + +template <> +void vPow(const int n, const double* a, const double b, double* r) { + vdPowx(n, a, b, r); +} + +template <> +void vLog(const int n, const float* a, float* r) { + vsLn(n, a, r); +} + +template <> +void vLog(const int n, const double* a, double* r) { + vdLn(n, a, r); +} + +template <> +void vAdd(const int n, const float* a, const float* b, float* r) { + vsAdd(n, a, b, r); +} + +template <> +void vAdd(const int n, const double* a, const double* b, double* r) { + vdAdd(n, a, b, r); +} + +template <> +void vTanh(const int n, const float* a, float* r) { + vsTanh(n, a, r); +} + +template <> +void vTanh(const int n, const double* a, double* r) { + vdTanh(n, a, r); +} + +template <> +void vInvSqrt(const int n, const float* a, float* r) { + vsInvSqrt(n, a, r); +} + +template <> +void vInvSqrt(const int n, const double* a, double* r) { + vdInvSqrt(n, a, r); +} + +template <> +void vLog1p(const int n, const float* a, float* r) { + vsLog1p(n, a, r); +} + +template <> +void vLog1p(const int n, const double* a, double* r) { + vdLog1p(n, a, r); +} +#else + +DEFINE_MATRIX_BINARY_OP(vExp, b = std::exp(a)); +template +void vExp(const int n, const T* a, T* r) { + hl_cpu_apply_binary_op, 0, 0>( + binary::vExp(), const_cast(a), r, 1, n, n, n); +} + +DEFINE_MATRIX_BINARY_OP(vLog, b = std::log(a)); +template +void vLog(const int n, const T* a, T* r) { + hl_cpu_apply_binary_op, 0, 0>( + binary::vLog(), const_cast(a), r, 1, n, n, n); +} + +DEFINE_MATRIX_BINARY_PARAMETER_OP(vPow, ONE_PARAMETER, b = std::pow(a, p)); +template +void vPow(const int n, const T* a, const T b, T* r) { + hl_cpu_apply_binary_op, 0, 0>( + binary::vPow(b), const_cast(a), r, 1, n, n, n); +} + +DEFINE_MATRIX_TERNARY_OP(vAdd, c = a + b); +template +void vAdd(const int n, const T* a, const T* b, T* r) { + hl_cpu_apply_ternary_op, 0, 0>(ternary::vAdd(), + const_cast(a), + const_cast(b), + r, + 1, + n, + n, + n, + n); +} + +DEFINE_MATRIX_BINARY_OP(vInvSqrt, b = 1.0f / std::sqrt(a)); +template +void vInvSqrt(const int n, const T* a, T* r) { + hl_cpu_apply_binary_op, 0, 0>( + binary::vInvSqrt(), const_cast(a), r, 1, n, n, n); +} + +DEFINE_MATRIX_BINARY_OP(vLog1p, b = std::log(1.0f + a)); +template +void vLog1p(const int n, const T* a, T* r) { + hl_cpu_apply_binary_op, 0, 0>( + binary::vLog1p(), const_cast(a), r, 1, n, n, n); +} + +DEFINE_MATRIX_BINARY_OP(vTanh, T tmp = -2.0 * a; + tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp; + b = 2.0 / (1.0 + std::exp(tmp)) - 1.0); +template +void vTanh(const int n, const T* a, T* r) { + hl_cpu_apply_binary_op, 0, 0>( + binary::vTanh(), const_cast(a), r, 1, n, n, n); +} + +template void vExp(const int n, const float* a, float* r); +template void vExp(const int n, const double* a, double* r); +template void vLog(const int n, const float* a, float* r); +template void vLog(const int n, const double* a, double* r); +template void vPow(const int n, const float* a, const float b, float* r); +template void vPow(const int n, const double* a, const double b, double* r); +template void vAdd(const int n, const float* a, const float* b, float* r); +template void vAdd(const int n, const double* a, const double* b, double* r); +template void vInvSqrt(const int n, const double* a, double* r); +template void vInvSqrt(const int n, const float* a, float* r); +template void vLog1p(const int n, const float* a, float* r); +template void vLog1p(const int n, const double* a, double* r); +template void vTanh(const int n, const float* a, float* r); +template void vTanh(const int n, const double* a, double* r); +#endif +} // namespace paddle diff --git a/paddle/legacy/math/MathFunctions.h b/paddle/legacy/math/MathFunctions.h new file mode 100644 index 0000000000000000000000000000000000000000..854e4baa3987f61353038c7b26acf43943c89636 --- /dev/null +++ b/paddle/legacy/math/MathFunctions.h @@ -0,0 +1,129 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#ifdef PADDLE_WITH_MKLML +#include +#include +#include +#endif + +#ifdef PADDLE_USE_VECLIB +extern "C" { +#include +#include +} +#endif + +#ifdef PADDLE_USE_OPENBLAS +#include +#ifdef LAPACK_FOUND +#include +#endif +#endif + +#ifndef LAPACK_FOUND +extern "C" { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS +#include +#else +typedef enum CBLAS_ORDER { + CblasRowMajor = 101, + CblasColMajor = 102 +} CBLAS_ORDER; +#endif +int LAPACKE_sgetrf( + int matrix_layout, int m, int n, float* a, int lda, int* ipiv); +int LAPACKE_dgetrf( + int matrix_layout, int m, int n, double* a, int lda, int* ipiv); +int LAPACKE_sgetri( + int matrix_layout, int n, float* a, int lda, const int* ipiv); +int LAPACKE_dgetri( + int matrix_layout, int n, double* a, int lda, const int* ipiv); +} +#endif + +#include + +namespace paddle { + +#ifndef PADDLE_USE_EIGEN_FOR_BLAS +template +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc); +#endif + +template +int getrf(const CBLAS_ORDER Order, + const int M, + const int N, + T* A, + const int lda, + int* ipiv); + +template +int getri( + const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv); + +template +void axpy(const int n, const T alpha, const T* x, T* y) { + /// y = y + alpha * x + for (int i = 0; i < n; i++) { + y[i] = y[i] + alpha * x[i]; + } +} + +template +T dotProduct(const int n, const T* x, const T* y) { + T result = static_cast(0); + for (int i = 0; i < n; i++) { + result += x[i] * y[i]; + } + return result; +} + +template +void vExp(const int n, const T* a, T* r); + +template +void vPow(const int n, const T* a, const T b, T* r); + +template +void vLog(const int n, const T* a, T* r); + +template +void vAdd(const int n, const T* a, const T* b, T* r); + +template +void vInvSqrt(const int n, const T* a, T* r); + +template +void vLog1p(const int n, const T* a, T* r); + +template +void vTanh(const int n, const T* a, T* r); + +} // namespace paddle diff --git a/paddle/legacy/math/MathUtils.cpp b/paddle/legacy/math/MathUtils.cpp new file mode 100644 index 0000000000000000000000000000000000000000..47ac9c187ca731c98c755501ff3633eabf095186 --- /dev/null +++ b/paddle/legacy/math/MathUtils.cpp @@ -0,0 +1,97 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MathUtils.h" +#include +#include "Vector.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +/*if csc, major is cols and minor is rows, else + * major is rows and minor is cols, according to + * major value to initialize minor value" + */ +void sparseRand( + int* major, int* minor, int nnz, int majorLen, int minorMax, bool useGpu) { + CHECK(size_t(nnz) >= size_t(1)); + int* cpuMajor; + int* cpuMinor; + CpuIVector cpuMinorVec(nnz); + CpuIVector cpuMajorVec(majorLen); + if (useGpu) { + cpuMajor = cpuMajorVec.getData(); + cpuMinor = cpuMinorVec.getData(); + } else { + cpuMajor = major; + cpuMinor = minor; + } + + /*major value init*/ + for (int i = 0; i < majorLen - 1; i++) { + cpuMajor[i] = 1.0 * i * nnz / (majorLen - 1); + } + cpuMajor[majorLen - 1] = nnz; + + /*minor value init according to major value*/ + std::vector used(minorMax, 0); + for (int i = 0; i < majorLen - 1; i++) { + CHECK_LE(cpuMajor[i + 1] - cpuMajor[i], minorMax); + used.assign(minorMax, 0); + for (int j = cpuMajor[i]; j < cpuMajor[i + 1]; j++) { + int idx = ::rand() % minorMax; + while (used[idx]) { + idx = ::rand() % minorMax; + } + cpuMinor[j] = idx; + used[idx] = 1; + } + std::sort(cpuMinor + cpuMajor[i], + cpuMinor + cpuMajor[i + 1], + [](int a, int b) { return a < b; }); + } + /*memcpy result to gpu*/ + if (useGpu) { + hl_memcpy_host2device(major, cpuMajor, sizeof(int) * majorLen); + hl_memcpy_host2device(minor, cpuMinor, sizeof(int) * nnz); + } +} + +int outputSize( + int imageSize, int filterSize, int padding, int stride, bool caffeMode) { + int outputSize; + if (!caffeMode) { + outputSize = + (imageSize - filterSize + 2 * padding + stride - 1) / stride + 1; + } else { + outputSize = (imageSize - filterSize + 2 * padding) / stride + 1; + } + CHECK_GE(outputSize, 1); + return outputSize; +} + +int imageSize( + int outputSize, int filterSize, int padding, int stride, bool caffeMode) { + int imageSize; + if (!caffeMode) { + imageSize = + (outputSize - 1) * stride + filterSize - 2 * padding - stride + 1; + } else { + imageSize = (outputSize - 1) * stride + filterSize - 2 * padding; + } + CHECK_GE(imageSize, 1); + return imageSize; +} + +} // namespace paddle diff --git a/paddle/math/MathUtils.h b/paddle/legacy/math/MathUtils.h similarity index 100% rename from paddle/math/MathUtils.h rename to paddle/legacy/math/MathUtils.h diff --git a/paddle/legacy/math/Matrix.cpp b/paddle/legacy/math/Matrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e53f95006c36bfce5df8e57e9efc249f56098b70 --- /dev/null +++ b/paddle/legacy/math/Matrix.cpp @@ -0,0 +1,4787 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Matrix.h" +#include "MathFunctions.h" +#include "SparseMatrix.h" +#include "SparseRowMatrix.h" + +#include +#include +#include + +#include +#include "hl_cnn.h" +#include "hl_gpu.h" +#include "hl_table_apply.h" +#include "hl_top_k.h" +#include "paddle/legacy/utils/Logging.h" + +#include "NEONFunctions.h" +#include "paddle/legacy/function/GemmFunctor.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +#include "SIMDFunctions.h" + +namespace paddle { + +inline real _pow(real a, real beta) { return std::pow(a, beta); } + +inline real _square(real a) { return a * a; } + +inline real _safelog(real a) { return a > 0.0f ? std::log(a) : -40.0f; } + +Matrix::Matrix(MemoryHandlePtr memHandle, + size_t height, + size_t width, + bool trans, + bool use_gpu) + : BaseMatrix( + height, + width, + memHandle ? (reinterpret_cast(memHandle->getBuf())) : nullptr, + trans, + use_gpu) { + elementCnt_ = width * height; + memoryHandle_ = memHandle; +} + +Matrix::Matrix( + real* data, size_t height, size_t width, bool trans, bool use_gpu) + : BaseMatrix(height, width, data, trans, use_gpu) { + elementCnt_ = width * height; +} + +Matrix::Matrix(real* data, + size_t height, + size_t width, + size_t stride, + bool trans, + bool use_gpu) + : BaseMatrix(height, width, stride, data, trans, use_gpu) { + elementCnt_ = width * height; +} + +MatrixPtr Matrix::createSparseMatrix(real* data, + int* row, + int* col, + size_t height, + size_t width, + size_t nnz, /* used to allocate space */ + SparseValueType valueType, /*value type*/ + SparseFormat format, + bool trans, + bool useGpu) { + if (useGpu) { + return std::make_shared( + data, row, col, height, width, nnz, valueType, format, trans); + } else { + return std::make_shared( + data, row, col, height, width, nnz, valueType, format, trans); + } +} + +MatrixPtr Matrix::createSparseMatrix(size_t height, + size_t width, + size_t nnz, /* used to allocate space */ + SparseValueType valueType, /*value type*/ + SparseFormat format, + bool trans, + bool useGpu) { + if (useGpu) { + return std::make_shared( + height, width, nnz, valueType, format, trans); + } else { + return std::make_shared( + height, width, nnz, valueType, format, trans); + } +} + +MatrixPtr Matrix::create(MemoryHandlePtr memHandle, + size_t height, + size_t width, + bool trans) { + if (auto gpuHandle = std::dynamic_pointer_cast(memHandle)) { + return std::make_shared(gpuHandle, height, width, trans); + } else if (auto cpuHandle = + std::dynamic_pointer_cast(memHandle)) { + return std::make_shared(cpuHandle, height, width, trans); + } else { + LOG(FATAL) << "Wrong"; + return nullptr; + } +} + +MatrixPtr Matrix::create(size_t height, size_t width, bool trans, bool useGpu) { + if (useGpu) { + return std::make_shared(height, width, trans); + } else { + return std::make_shared(height, width, trans); + } +} + +MatrixPtr Matrix::create( + real* data, size_t height, size_t width, bool trans, bool useGpu) { + if (useGpu) { + return std::make_shared(data, height, width, trans); + } else { + return std::make_shared(data, height, width, trans); + } +} + +MatrixPtr Matrix::create(real* data, + size_t height, + size_t width, + size_t stride, + bool trans, + bool useGpu) { + if (useGpu) { + return std::make_shared(data, height, width, stride, trans); + } else { + return std::make_shared(data, height, width, stride, trans); + } +} + +MatrixPtr Matrix::createSparseMatrix(size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + bool trans, + bool useGpu) { + if (useGpu) { + return std::make_shared( + height, width, nnz, valueType, SPARSE_CSR, trans); + } else { + return std::make_shared( + height, width, nnz, valueType, SPARSE_CSR, trans); + } +} + +void Matrix::resizeOrCreate( + MatrixPtr& matrix, size_t height, size_t width, bool trans, bool useGpu) { + if (!matrix) { + matrix = Matrix::create(height, width, trans, useGpu); + } else { + CHECK_EQ(matrix->useGpu(), useGpu); + matrix->resize(height, width); + } +} + +void Matrix::resizeOrCreateSparseMatrix(MatrixPtr& matrix, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans, + bool useGpu) { + if (!matrix) { + matrix = Matrix::createSparseMatrix( + height, width, nnz, valueType, format, trans, useGpu); + } else { + CHECK(dynamic_cast(matrix.get()) || + dynamic_cast(matrix.get())); + CHECK_EQ(matrix->useGpu(), useGpu); + matrix->resize(height, width, nnz, valueType, format); + } +} + +void Matrix::reshape(size_t height, size_t width) { + CHECK(isContiguous()); + CHECK(height_ * width_ == height * width); + height_ = height; + width_ = width; + stride_ = width_; +} + +MatrixPtr Matrix::subMatrix(size_t startRow, + size_t endRow, + size_t startCol, + size_t endCol) { + CHECK_LE(startRow, endRow); + CHECK_LE(endRow, getHeight()); + CHECK_LE(startCol, endCol); + CHECK_LE(endCol, getWidth()); + + return Matrix::create(getData() + startRow * getStride() + startCol, + endRow - startRow, + endCol - startCol, + getStride(), + trans_, + useGpu_); +} + +void Matrix::setDiag(real value) { + CHECK(data_ != NULL); + CHECK_EQ(height_, width_); + + zeroMem(); + BaseMatrix diag(height_, 1, stride_ + 1, data_, false, useGpu_); + diag.assign(value); +} + +GpuMatrix::GpuMatrix(size_t height, size_t width, bool trans) + : Matrix(std::make_shared(height * width * sizeof(real)), + height, + width, + trans, + true) {} + +GpuMatrix::~GpuMatrix() {} + +void GpuMatrix::zeroMem() { + CHECK(data_ != NULL); + zero(); +} + +void GpuMatrix::resetOne() { + CHECK(data_ != NULL); + one(); +} + +void GpuMatrix::resize(size_t newHeight, size_t newWidth) { + size_t newSize = newHeight * newWidth; + if (NULL == memoryHandle_.get() || + newSize * sizeof(real) > memoryHandle_->getAllocSize()) { + memoryHandle_ = std::make_shared(newSize * sizeof(real)); + data_ = reinterpret_cast(memoryHandle_->getBuf()); + } + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newSize; + stride_ = width_; +} + +real GpuMatrix::getElement(size_t x, size_t y) const { + real elem = 0; + hl_memcpy_device2host(&elem, &data_[x * stride_ + y], sizeof(real)); + return elem; +} + +real GpuMatrix::getSum() { + CHECK(isContiguous()); + real sum = 0.0f; + hl_vector_sum(data_, &sum, height_ * width_); + return sum; +} + +real GpuMatrix::getMin() { + CHECK(isContiguous()); + auto vec = GpuVector(height_ * width_, data_); + return vec.getMin(); +} + +real GpuMatrix::getMax() { + CHECK(isContiguous()); + auto vec = GpuVector(height_ * width_, data_); + return vec.getMax(); +} + +void GpuMatrix::accumulateColSum(Matrix& src) { + CHECK_EQ(getWidth(), src.getWidth()); + CHECK_EQ(getHeight(), (size_t)1); + sumCols(src, 1.0, 1.0); +} + +real GpuMatrix::getAbsSum() { + CHECK(isContiguous()); + real sum = 0.0f; + hl_vector_abs_sum(data_, &sum, height_ * width_); + return sum; +} + +void GpuMatrix::copyFrom(const Matrix& src) { + CHECK(isContiguous()); + CHECK(src.isContiguous()); + CHECK(elementCnt_ == src.getElementCnt()); + + if (typeid(src) == typeid(CpuMatrix)) { + hl_memcpy_host2device( + data_, const_cast(src.getData()), sizeof(real) * elementCnt_); + } else if (typeid(src) == typeid(GpuMatrix)) { + hl_memcpy_device2device( + data_, const_cast(src.getData()), sizeof(real) * elementCnt_); + } else { + LOG(FATAL) << "Wrong"; + } +} + +void GpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { + CHECK(isContiguous()); + CHECK(src.isContiguous()); + CHECK(elementCnt_ == src.getElementCnt()); + hl_memcpy_async(this->getData(), + const_cast(src.getData()), + sizeof(real) * elementCnt_, + stream); +} + +void GpuMatrix::copyFrom(const real* hostSrc, size_t size) { + CHECK(isContiguous()); + CHECK(size <= elementCnt_); + hl_memcpy_host2device(data_, const_cast(hostSrc), sizeof(real) * size); +} + +void GpuMatrix::copyFrom(const real* hostSrc, const int64_t* seq) { + LOG(FATAL) << "not implemented"; +} + +void GpuMatrix::copyFrom(const IVector& src) { + CHECK(isContiguous()); + CpuMatrix matrix(src.getSize(), 1, false); + matrix.copyFrom(src); + copyFrom(matrix); +} + +void GpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) { + size_t height = getHeight(); + size_t width = getWidth(); + CHECK_EQ(b.getWidth(), width); + real* dst = getData(); + real* src = b.getData(); + const int* index = rowIndex.getData(); + hl_sequence2batch_copy(dst, src, index, width, height, true); +} + +MatrixPtr GpuMatrix::clone(size_t height, size_t width, bool useGpu) { + CHECK(isContiguous()); + + if (height == 0 && width == 0) { + height = height_; + width = width_; + } + + CHECK(width && height); + + if (useGpu) { + return std::make_shared(height, width); + } else { + return std::make_shared(height, width); + } +} + +MatrixPtr GpuMatrix::getTranspose() { + if (memoryHandle_.get() != NULL) { + MatrixPtr copy_T( + new GpuMatrix(std::dynamic_pointer_cast(memoryHandle_), + height_, + width_, + true)); + return copy_T; + } else { + MatrixPtr copy_T(new GpuMatrix(data_, height_, width_, true)); + return copy_T; + } +} + +void GpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { + if (memAlloc) { + matTrans = std::make_shared(width_, height_); + } else { + CHECK(matTrans != NULL); + CHECK_EQ(matTrans->getHeight(), width_); + CHECK_EQ(matTrans->getWidth(), height_); + } + real* dataTrans = matTrans->getData(); + real* data = getData(); + int lda = getStride(); + int ldc = matTrans->getStride(); + + hl_matrix_transpose(data, dataTrans, height_, width_, lda, ldc); +} + +void GpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) { + if (memAlloc) { + matRot = std::make_shared(width_, height_); + } else { + CHECK(matRot != NULL); + CHECK_EQ(matRot->getHeight(), width_); + CHECK_EQ(matRot->getWidth(), height_); + } + + real* dataRot = matRot->getData(); + real* data = getData(); + hl_matrix_rotate(data, dataRot, height_, width_, clockWise); +} + +MatrixPtr GpuMatrix::getInverse() { + MatrixPtr matInv; + inverse(matInv, true); + return matInv; +} + +void GpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) { + CHECK_EQ(height_, width_); + + if (memAlloc) { + matInv = std::make_shared(height_, width_); + } else { + CHECK(matInv != NULL); + } + + real* data = getData(); + real* dataInv = matInv->getData(); + int lda = getStride(); + int ldc = matInv->getStride(); + + hl_matrix_inverse(data, dataInv, height_, lda, ldc); +} + +void GpuMatrix::addBias(Matrix& b, real scale) { + CHECK(b.getHeight() == 1) << "the Bias should be a vector"; + BaseMatrix::addBias(b, scale); +} + +void GpuMatrix::addSharedBias(Matrix& b, real scale) { + CHECK(b.getHeight() == 1) << "the Bias should be a vector"; + CHECK_LE(b.getWidth(), getWidth()); + CHECK_EQ(getWidth() % b.getWidth(), 0UL); + hl_matrix_add_shared_bias( + getData(), b.getData(), b.getWidth(), getHeight(), getWidth(), scale); +} + +void GpuMatrix::collectBias(Matrix& a, real scale) { +#ifdef PADDLE_WITH_CUDA + CHECK_EQ(getHeight(), (size_t)1); + CHECK_EQ(width_, a.getWidth()); + GpuSparseMatrix* sMatPtr = dynamic_cast(&a); + if (!sMatPtr) { + sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1); + } else { + real* data = getData(); + hl_sparse_matrix_s A_d = sMatPtr->sMatrix_.get(); + hl_sparse_matrix_column_sum(data, A_d, sMatPtr->getHeight(), width_, scale); + } +#endif +} + +void GpuMatrix::collectSharedBias(Matrix& a, real scale) { + CHECK_EQ(getHeight(), (size_t)1); + CHECK_EQ(a.getWidth() % getWidth(), 0UL); + hl_matrix_collect_shared_bias( + getData(), a.getData(), getWidth(), a.getHeight(), a.getWidth(), scale); +} + +void GpuMatrix::sequenceAvgForward(Matrix& a, + const IVector& startsPos, + int mode) { + size_t height = getHeight(); + size_t width = getWidth(); + CHECK_EQ(height, startsPos.getSize() - 1); + CHECK_EQ(width, a.getWidth()); + real* dst = getData(); + real* src = a.getData(); + const int* starts = startsPos.getData(); + + hl_sequence_avg_forward(dst, src, starts, height, width, mode); +} + +void GpuMatrix::sequenceAvgBackward(Matrix& a, + const IVector& startsPos, + int mode) { + size_t height = a.getHeight(); + size_t width = getWidth(); + CHECK_EQ(height, startsPos.getSize() - 1); + CHECK_EQ(width, a.getWidth()); + real* dst = getData(); + real* src = a.getData(); + const int* starts = startsPos.getData(); + + hl_sequence_avg_backward(dst, src, starts, height, width, mode); +} + +/* this = scaleAB*(a*b) + scaleT*this */ +void GpuMatrix::mul(const GpuMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT) { + CHECK(!isTransposed()) << "Not supported"; + + if (!a.isTransposed() && !b.isTransposed()) { + CHECK_EQ(width_, b.width_); + CHECK_EQ(height_, a.height_); + CHECK_EQ(a.width_, b.height_); + } else if (a.isTransposed() && !b.isTransposed()) { + CHECK_EQ(width_, b.width_); + CHECK_EQ(height_, a.width_); + CHECK_EQ(a.height_, b.height_); + } else if (!a.isTransposed() && b.isTransposed()) { + CHECK_EQ(width_, b.height_); + CHECK_EQ(height_, a.height_); + CHECK_EQ(a.width_, b.width_); + } else { + LOG(FATAL) << "Is not supported"; + } + + real* A_d = a.data_; + real* B_d = b.data_; + real* C_d = data_; + int dimM = getHeight(); + int dimN = getWidth(); + int dimK = !a.isTransposed() ? a.width_ : a.height_; + int lda = a.getStride(); + int ldb = b.getStride(); + int ldc = getStride(); + hl_trans_op_t transa = !a.isTransposed() ? HPPL_OP_N : HPPL_OP_T; + hl_trans_op_t transb = !b.isTransposed() ? HPPL_OP_N : HPPL_OP_T; + + hl_matrix_mul(A_d, + transa, + B_d, + transb, + C_d, + dimM, + dimN, + dimK, + scaleAB, + scaleT, + lda, + ldb, + ldc); +} + +void GpuMatrix::mul(const GpuSparseMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT) { +#ifdef PADDLE_WITH_CUDA + CHECK(isContiguous()); + CHECK(b.isContiguous()); + CHECK(b.useGpu_ == true) << "Matrix type are not equal"; + CHECK(!trans_ && !b.trans_) << "not supported"; + + if (!a.trans_) { + CHECK(width_ == b.width_ && height_ == a.height_ && a.width_ == b.height_) + << "Matrix dimensions are not equal"; + } else { + CHECK(width_ == b.width_ && height_ == a.width_ && a.height_ == b.height_) + << "Matrix dimensions are not equal"; + } + hl_trans_op_t transA = a.trans_ ? HPPL_OP_T : HPPL_OP_N; + hl_sparse_matrix_s A_d = a.sMatrix_.get(); + real* B_d = b.data_; + real* C_d = data_; + hl_matrix_csr_mul_dense(A_d, + transA, + B_d, + HPPL_OP_N, + C_d, + height_, + width_, + b.height_, + scaleAB, + scaleT); +#endif +} + +void GpuMatrix::mul(const GpuMatrix& a, + const GpuSparseMatrix& b, + real scaleAB, + real scaleT) { +#ifdef PADDLE_WITH_CUDA + CHECK(isContiguous()); + CHECK(a.isContiguous()); + CHECK(a.useGpu_ == true) << "Matrix type are not equal"; + + hl_sparse_matrix_s B_d = b.sMatrix_.get(); + real* A_d = a.data_; + real* C_d = data_; + hl_trans_op_t transB = b.trans_ ? HPPL_OP_T : HPPL_OP_N; + if (!b.trans_) { + CHECK(width_ == b.width_ && height_ == a.height_ && a.width_ == b.height_) + << "Matrix dimensions are not equal"; + } else { + CHECK(width_ == b.height_ && height_ == a.height_ && a.width_ == b.width_) + << "Matrix dimensions are not equal"; + } + if (b.format_ == SPARSE_CSC) { + hl_matrix_dense_mul_csc(A_d, + HPPL_OP_N, + B_d, + transB, + C_d, + height_, + width_, + a.width_, + scaleAB, + scaleT); + } else { + hl_matrix_dense_mul_csr(A_d, + HPPL_OP_N, + B_d, + transB, + C_d, + height_, + width_, + a.width_, + scaleAB, + scaleT); + } +#endif +} + +/* this = a*b */ +void GpuMatrix::mul(const Matrix& a, const Matrix& b) { mul(a, b, 1.0, 0.0); } + +void GpuMatrix::mul(const Matrix& a, + const Matrix& b, + real scaleAB, + real scaleT) { + const auto a_ptr = dynamic_cast(&a); + const auto b_ptr = dynamic_cast(&b); + const auto a_ptr_s = dynamic_cast(&a); + const auto b_ptr_s = dynamic_cast(&b); + + if (a_ptr && b_ptr) { + mul(*a_ptr, *b_ptr, scaleAB, scaleT); + } else if (a_ptr_s && b_ptr) { + mul(*a_ptr_s, *b_ptr, scaleAB, scaleT); + } else if (a_ptr && b_ptr_s) { + mul(*a_ptr, *b_ptr_s, scaleAB, scaleT); + } else { + LOG(FATAL) << "Not supported"; + } +} + +/* this = this* b */ +void GpuMatrix::rightMul(Matrix& b) { rightMul(b, 1.0, 0.0); } + +/* this = scaleAB*(this*b) + scaleT*this */ +void GpuMatrix::rightMul(Matrix& b, real scaleAB, real scaleT) { + CHECK(dynamic_cast(&b)); + CHECK(!isTransposed()) << "Not supported"; + CHECK(!b.isTransposed()) << "Not supported"; + mul(*this, *dynamic_cast(&b), scaleAB, scaleT); +} + +/* this = a*this */ +void GpuMatrix::leftMul(Matrix& a) { leftMul(a, 1.0, 0.0); } + +/* this = scaleAB*(a*this) + scaleT*this */ +void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) { + CHECK(dynamic_cast(&a)); + CHECK(!isTransposed()) << "Not supported"; + CHECK(!a.isTransposed()) << "Not supported"; + mul(*dynamic_cast(&a), *this, scaleAB, scaleT); +} + +void GpuMatrix::selectRows(Matrix& table, IVector& ids) { +#ifdef PADDLE_WITH_CUDA + CHECK(dynamic_cast(&table)); + CHECK(table.useGpu()); + CHECK(ids.useGpu()); + CHECK_EQ(getHeight(), ids.getSize()); + CHECK_EQ(getWidth(), table.getWidth()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + real* a = getData(); + size_t tableSize = table.getHeight(); + int* index = ids.getData(); + + hl_matrix_select_rows(a, + stride_, + table.getData(), + table.stride_, + index, + numSamples, + tableSize, + dim); +#endif +} + +void GpuMatrix::addToRows(Matrix& table, IVector& ids) { +#ifdef PADDLE_WITH_CUDA + CHECK(dynamic_cast(&table)); + CHECK(table.useGpu()); + CHECK(ids.useGpu()); + CHECK_EQ(getHeight(), ids.getSize()); + CHECK_EQ(getWidth(), table.getWidth()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + real* a = getData(); + size_t tableSize = table.getHeight(); + int* index = ids.getData(); + + hl_matrix_add_to_rows(table.getData(), + table.stride_, + a, + stride_, + index, + numSamples, + tableSize, + dim); +#endif +} + +void GpuMatrix::colMerge(Matrix& src) { + CHECK(src.height_ == height_); + if (!trans_ && !src.trans_) { + sumRows(src, /* scaleSum= */ 1, /* scaleDest= */ 0); + } else { + LOG(FATAL) << "Is not supported"; + } +} + +void GpuMatrix::rowSum(Matrix& sum) { + CHECK_EQ(sum.getHeight(), getHeight()); + CHECK_EQ(sum.getWidth(), (size_t)1); + + sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0); +} + +void GpuMatrix::rowMax(Matrix& max) { + CHECK_EQ(max.getHeight(), getHeight()); + CHECK_EQ(max.getWidth(), (size_t)1); + + max.maxRows(*this); +} + +void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { +#ifdef PADDLE_WITH_CUDA + CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; + size_t numSamples = getHeight(); + size_t beam = maxVal.getWidth(); + CHECK_EQ(maxIds.getSize(), numSamples * beam); + CHECK_EQ(maxVal.getHeight(), numSamples); + CHECK_EQ(maxVal.getWidth(), beam); + + hl_matrix_top_k(maxVal.getData(), + maxVal.getStride(), + maxIds.getData(), + this->getData(), + this->getStride(), + this->getWidth(), + beam, + numSamples); +#endif +} + +void GpuMatrix::colMax(Matrix& max) { + CHECK_EQ(max.getWidth(), getWidth()); + CHECK_EQ(max.getHeight(), (size_t)1); + + max.maxCols(*this); +} + +void GpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) { + LOG(FATAL) << "Is not supported"; +} + +void GpuMatrix::maxoutForward(Matrix& a, + IVector& id, + size_t channels, + size_t groups) { + CHECK(dynamic_cast(&a)); + CHECK(dynamic_cast(&id)); + CHECK_EQ(a.getHeight(), getHeight()); + + size_t size = getWidth(); + size_t batchSize = getHeight(); + const real* input = a.getData(); + real* output = getData(); + int* idForGpu = id.getData(); + + hl_maxout_forward( + input, output, idForGpu, batchSize, size, size / channels, groups); +} + +void GpuMatrix::maxoutBackward(Matrix& a, + IVector& id, + size_t channels, + size_t groups) { + CHECK(dynamic_cast(&a)); + CHECK(dynamic_cast(&id)); + CHECK_EQ(a.getHeight(), getHeight()); + + size_t size = a.getWidth(); + size_t batchSize = getHeight(); + real* input = getData(); + const real* output = a.getData(); + const int* idForGpu = id.getData(); + + hl_maxout_backward( + input, output, idForGpu, batchSize, size, size / channels, groups); +} + +/*calulate the error of classification */ +void GpuMatrix::classificationError(Matrix& output, + IVector& label, + size_t topkSize) { + auto gpuOutput = dynamic_cast(&output); + auto gpuLabel = dynamic_cast(&label); + size_t numSamples = this->getHeight(); + GpuMatrixPtr gpuTopVal = std::make_shared(numSamples, topkSize); + GpuIVectorPtr gpuTopIds = std::make_shared(numSamples * topkSize); + + CHECK(gpuOutput && gpuLabel) << "Invalid argument pointer"; + CHECK(gpuTopVal && gpuTopIds) << "Allocate GPU memory failed"; + CHECK(gpuLabel->getSize() == numSamples) << "Vector size is not equal"; + CHECK(numSamples == gpuOutput->getHeight() && this->getWidth() == 1) + << "Matrix dimensions are not equal"; + + size_t dim = gpuOutput->getWidth(); + hl_matrix_classification_error(gpuTopVal->getData(), + gpuTopVal->getStride(), + gpuTopIds->getData(), + gpuOutput->getData(), + gpuOutput->getStride(), + dim, + topkSize, + numSamples, + gpuLabel->getData(), + this->getData()); +} + +/* copy -log(output[i * width + label]) to this->data[i] */ +void GpuMatrix::oneHotCrossEntropy(Matrix& output, IVector& label) { + GpuMatrix* output_ptr = dynamic_cast(&output); + GpuIVector* label_ptr = dynamic_cast(&label); + + CHECK(output_ptr && label_ptr) << "Invalid argument pointer"; + + CHECK(height_ == label.getSize() && width_ == 1 && height_ == output.height_) + << "Matrix dimensions are not equal"; + + real* A_d = output_ptr->data_; + real* C_d = data_; + int* label_d = label_ptr->getData(); + + hl_matrix_cross_entropy(A_d, C_d, label_d, height_, output.width_); +} + +/* calculate the error of outputV according to label */ +void GpuMatrix::oneHotCrossEntropyBp(Matrix& outputV, IVector& label) { + GpuMatrix* output_ptr = dynamic_cast(&outputV); + GpuIVector* label_ptr = dynamic_cast(&label); + + CHECK(output_ptr && label_ptr) << "Invalid argument pointer"; + + CHECK(height_ == output_ptr->height_ && width_ == output_ptr->width_) + << "Matrix dimensions are not equal"; + + real* output_d = output_ptr->data_; + real* grad_d = data_; + int* label_d = label_ptr->getData(); + + hl_matrix_cross_entropy_bp(grad_d, output_d, label_d, height_, width_); +} + +void GpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output, + IVector& label, + real alpha) { + LOG(FATAL) << "Not implemented"; +} + +void GpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, + IVector& label, + real alpha) { + LOG(FATAL) << "Not implemented"; +} + +void GpuMatrix::softmax(Matrix& output) { + CHECK(output.useGpu()) << "Matrix type are not equal"; + + size_t height = getHeight(); + size_t width = getWidth(); + CHECK(height == output.getHeight() && width == output.getWidth()) + << "Matrix dimensions are not equal"; + + real* inputData = getData(); + real* outputData = output.getData(); + hl_matrix_softmax(inputData, outputData, height, width); +} + +void GpuMatrix::sequenceSoftmax(Matrix& output, const IVector& index) { + CHECK_EQ(getWidth(), 1UL); + CHECK_EQ(output.getWidth(), 1UL); + CHECK(isContiguous()); + + real* inputData = getData(); + real* outputData = output.getData(); + auto starts = index.getData(); + int numSequences = index.getSize() - 1; + hl_sequence_softmax_forward(inputData, outputData, starts, numSequences); +} + +void GpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) { + CHECK(output.useGpu_ == true && sftmaxSum.useGpu_ == true) + << "Matrix type are not equal"; + + CHECK(height_ == output.height_ && width_ == output.width_ && + height_ == sftmaxSum.height_) + << "Matrix dimensions are not equal"; + + real* output_d = output.data_; + real* sftmaxSum_d = sftmaxSum.data_; + real* grad_d = data_; + hl_matrix_softmax_derivative(grad_d, output_d, sftmaxSum_d, height_, width_); +} + +void GpuMatrix::softmaxBackward(Matrix& outputV) { + CHECK(outputV.useGpu()) << "Matrix type are not equal"; + + size_t height = getHeight(); + size_t width = getWidth(); + CHECK(height == outputV.getHeight() && width == outputV.getWidth()) + << "Matrix dimensions are not equal"; + + real* output_grad = getData(); + real* output_value = outputV.getData(); + hl_softmax_backward(output_value, output_grad, height, width); +} + +void GpuMatrix::sumOfSquares(Matrix& output, Matrix& label) { + CHECK_EQ(label.getHeight(), height_); + CHECK_EQ(output.getHeight(), height_); + CHECK_EQ(label.getWidth(), output.getWidth()); + CHECK_EQ((size_t)1, width_); + + auto labelptr = dynamic_cast(&label); + if (labelptr) { + LOG(FATAL) << "not supported: GpuSparseMatrix as label"; + } + + BaseMatrix::sumOfSquaredDiffs(output, + label, + /* scaleSum= */ 1, + /* scaleDest= */ 1); +} + +void GpuMatrix::sumOfSquaresBp(Matrix& outputV, Matrix& label) { + add2(outputV, label, 1, 2, -2); +} + +void GpuMatrix::tanh(Matrix& output) { BaseMatrix::tanh(output); } + +void GpuMatrix::tanhDerivative(Matrix& output) { + BaseMatrix::tanhDerivative(output); +} + +void GpuMatrix::softrelu(Matrix& output) { BaseMatrix::softrelu(output); } + +void GpuMatrix::softreluDerivative(Matrix& output) { + BaseMatrix::softreluDerivative(output); +} + +void GpuMatrix::scaledTanh(Matrix& output, real p1, real p2) { + BaseMatrix::scaledTanh(output, p1, p2); +} + +void GpuMatrix::randomizeUniform() { + CHECK(isContiguous()); + real* data = data_; + size_t size = height_ * width_; + + hl_rand(data, size); +} + +void GpuMatrix::print(std::ostream& os) const { + CHECK(isContiguous()); + CpuMatrix cpuMat(getHeight(), getWidth()); + cpuMat.copyFrom(*this); + cpuMat.print(os); +} + +void GpuMatrix::print(std::ostream& os, size_t height, size_t width) const { + CHECK(isContiguous()); + CpuMatrix cpuMat(getHeight(), getWidth()); + cpuMat.copyFrom(*this); + cpuMat.print(os, height, width); +} + +void GpuMatrix::check(std::ostream& os, Matrix& refMat, bool printDiff) { + CHECK(isContiguous()); + CHECK(height_ == refMat.getHeight()); + CHECK(width_ == refMat.getWidth()); + CpuMatrix cpuRef(height_, width_); + GpuMatrix gpuRef(height_, width_); + cpuRef.copyFrom(refMat); + gpuRef.copyFrom(*this); + size_t diffCnt = 0; + for (size_t i = 0; i < height_; ++i) { + for (size_t j = 0; j < width_; ++j) { + real a = gpuRef.getElement(i, j); + real b = cpuRef.getElement(i, j); + if (fabs(a - b) > 0.00001) { + ++diffCnt; + if (printDiff) { + os << "ref= " << a << " check= " << b << std::endl; + } + } + } + } + LOG(INFO) << "the diffCnt is " << diffCnt; +} + +void GpuMatrix::upsampleForward(Matrix& input, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW) { + CHECK(input.useGpu_ == true) << "Matrix type are not equal"; + CHECK(mask.useGpu_ == true) << "Matrix type are not equal"; + + real* inputData = input.getData(); + real* maskData = mask.getData(); + real* outData = data_; + + size_t batch = input.getHeight(); + + CHECK(imgSizeH * imgSizeW * channels == input.getWidth()); + CHECK(imgSizeH * imgSizeW * channels == mask.getWidth()); + CHECK_EQ(batch, this->getHeight()); + CHECK(width_ == outputH * outputW * channels); + hl_upsample_forward(inputData, + maskData, + batch, + imgSizeH, + imgSizeW, + channels, + outputH, + outputW, + outData); +} + +void GpuMatrix::upsampleBackward(Matrix& outputGrad, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW) { + CHECK(outputGrad.useGpu_ == true) << "Matrix type are not equal"; + CHECK(mask.useGpu_ == true) << "Matrix type are not equal"; + + real* outputGradData = outputGrad.getData(); + real* maskData = mask.getData(); + real* inputGradData = data_; + size_t batch = outputGrad.getHeight(); + + CHECK(imgSizeH * imgSizeW == this->getWidth() / channels); + CHECK_EQ(batch, this->getHeight()); + CHECK_EQ(channels * outputH * outputW, outputGrad.getWidth()); + hl_upsample_backward(outputGradData, + maskData, + batch, + imgSizeH, + imgSizeW, + channels, + outputH, + outputW, + inputGradData); +} + +void GpuMatrix::maxPoolForward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + MatrixPtr maskMatP) { + CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal"; + + real* inputData = inputMat.getData(); + real* maskData = NULL; + size_t frameNum = inputMat.getHeight(); + CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(width_ == outputH * outputW * channels); + + if (maskMatP != NULL) { + CHECK(maskMatP->useGpu_ == true) << "Matrix type are not equal"; + CHECK(outputH * outputW * channels == maskMatP->getWidth()); + maskData = maskMatP->getData(); + } + + hl_maxpool_forward(frameNum, + inputData, + channels, + imgSizeH, + imgSizeW, + outputH, + outputW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + data_, + getStride(), + maskData); +} + +void GpuMatrix::maxPoolBackward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + Matrix& outGrad, + Matrix& outV, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW) { + CHECK(inputMat.useGpu_ == true && outGrad.useGpu_ == true && + outV.useGpu_ == true) + << "Matrix type are not equal"; + + real* inputData = inputMat.getData(); + real* outData = outV.getData(); + real* outDiff = outGrad.getData(); + size_t frameNum = inputMat.getHeight(); + size_t channels = outV.getWidth() / outputH / outputW; + CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(outGrad.getHeight() == outV.getHeight() && + outGrad.getWidth() == outV.getWidth()); + + hl_maxpool_backward(frameNum, + inputData, + outData, + outDiff, + channels, + imgSizeH, + imgSizeW, + outputH, + outputW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + scaleTargets, + scaleOutput, + data_, + outGrad.getStride()); +} + +void GpuMatrix::avgPoolForward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + bool excludeMode) { + CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal"; + + real* inputData = inputMat.getData(); + size_t frameNum = inputMat.getHeight(); + CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(width_ == outputH * outputW * channels); + + hl_avgpool_forward(frameNum, + inputData, + channels, + imgSizeH, + imgSizeW, + outputH, + outputW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + data_, + getStride(), + excludeMode); +} + +void GpuMatrix::avgPoolBackward(Matrix& outGrad, + size_t imgSizeH, + size_t imgSizeW, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW, + bool excludeMode) { + CHECK(outGrad.useGpu_ == true) << "Matrix type are not equal"; + + real* outDiff = outGrad.getData(); + size_t frameNum = outGrad.getHeight(); + size_t channels = outGrad.getWidth() / outputH / outputW; + CHECK(imgSizeH * imgSizeW * channels == width_); + CHECK(height_ == outGrad.getHeight()); + CHECK(outGrad.getWidth() == outputH * outputW * channels); + + hl_avgpool_backward(frameNum, + outDiff, + channels, + imgSizeH, + imgSizeW, + outputH, + outputW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + scaleTargets, + scaleOutput, + data_, + outGrad.getStride(), + excludeMode); +} + +void GpuMatrix::maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + CHECK(inputMat.useGpu_) << "Matrix type are not correct"; + + real* inputData = inputMat.getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t num = inputMat.getHeight(); + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(width_ == outputD * outputH * outputW * channels); + + hl_maxpool3D_forward(num, + inputData, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData(), + maxPoolIdxData, + getStride()); +} + +void GpuMatrix::maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + CHECK(outGrad.useGpu_ && maxPoolIdx.useGpu_) << "Matrix type are not equal"; + + real* outDiff = outGrad.getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t frameNum = getHeight(); + size_t channels = outGrad.getWidth() / outputD / outputH / outputW; + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth()); + CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() && + outGrad.getWidth() == maxPoolIdx.getWidth()); + + hl_maxpool3D_backward(frameNum, + outDiff, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + scaleTargets, + scaleOutput, + getData(), + maxPoolIdxData, + outGrad.getStride()); +} + +void GpuMatrix::avgPool3DForward(Matrix& inputMat, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + CHECK(inputMat.useGpu_) << "Matrix type are not equal"; + + real* inputData = inputMat.getData(); + size_t frameNum = inputMat.getHeight(); + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(width_ == outputD * outputH * outputW * channels); + + hl_avgpool3D_forward(frameNum, + inputData, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData(), + getStride()); +} + +void GpuMatrix::avgPool3DBackward(Matrix& outGrad, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + CHECK(outGrad.useGpu_) << "Matrix type are not equal"; + + real* outDiff = outGrad.getData(); + size_t frameNum = outGrad.getHeight(); + size_t channels = outGrad.getWidth() / outputD / outputH / outputW; + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_); + CHECK(height_ == outGrad.getHeight()); + CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels); + + hl_avgpool3D_backward(frameNum, + outDiff, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + scaleTargets, + scaleOutput, + getData(), + outGrad.getStride()); +} + +void GpuMatrix::maxSequenceForward(Matrix& input, + const IVector& sequence, + IVector& index) { + CHECK(dynamic_cast(&input)); + CHECK(dynamic_cast(&sequence)); + CHECK(dynamic_cast(&index)); + + real* outData = getData(); + real* inputData = input.getData(); + const int* starts = sequence.getData(); + int* maxIndex = index.getData(); + size_t numSequences = getHeight(); + size_t dim = getWidth(); + + CHECK_EQ(dim, input.getWidth()); + CHECK_EQ(numSequences, sequence.getSize() - 1); + CHECK_EQ(numSequences * dim, index.getSize()); + + hl_max_sequence_forward( + inputData, starts, outData, maxIndex, numSequences, dim); +} + +void GpuMatrix::maxSequenceBackward(Matrix& outputGrad, + const IVector& sequence, + IVector& index) { + CHECK(dynamic_cast(&outputGrad)); + CHECK(dynamic_cast(&sequence)); + CHECK(dynamic_cast(&index)); + + real* inputGrad = getData(); + real* outGrad = outputGrad.getData(); + int* maxIndex = index.getData(); + size_t dim = getWidth(); + size_t numSequences = sequence.getSize() - 1; + + CHECK_EQ(dim, outputGrad.getWidth()); + CHECK_EQ(numSequences, outputGrad.getHeight()); + CHECK_EQ(numSequences * dim, index.getSize()); + + hl_max_sequence_backward(outGrad, maxIndex, inputGrad, numSequences, dim); +} + +void GpuMatrix::paramReluForward(Matrix& data, Matrix& W) { + CHECK(data.useGpu_ == true && W.useGpu_ == true) + << "Matrix type are not equal"; + real* input = data.getData(); + real* w = W.getData(); + size_t numElements = data.getWidth(); + size_t numSamples = data.getHeight(); + size_t paraSize = W.getHeight() * W.getWidth(); + CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + size_t partial_sum = numElements / paraSize; + real* output = getData(); + hl_param_relu_forward(output, input, w, numElements, numSamples, partial_sum); +} + +void GpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) { + CHECK(oGrad.useGpu_ == true && data.useGpu_ == true) + << "Matrix type are not equal"; + real* ograd = oGrad.getData(); + real* input = data.getData(); + real* wgrad = data_; + size_t numElements = data.getWidth(); + size_t numSamples = data.getHeight(); + size_t paraSize = this->getHeight() * this->getWidth(); + CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + size_t partial_sum = numElements / paraSize; + hl_param_relu_backward_w( + wgrad, ograd, input, numElements, numSamples, partial_sum); +} + +void GpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) { + real* diff = data_; + real* input = data.getData(); + real* ograd = oGrad.getData(); + real* w = W.getData(); + size_t numElements = data.getWidth(); + size_t numSamples = data.getHeight(); + size_t paraSize = W.getHeight() * W.getWidth(); + CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + size_t partial_sum = numElements / paraSize; + hl_param_relu_backward_diff( + ograd, input, w, diff, numElements, numSamples, partial_sum); +} + +void GpuMatrix::addColumnVector(const Matrix& b) { + BaseMatrix::addColVector(const_cast(b)); +} + +void GpuMatrix::bilinearForward(const Matrix& in, + const size_t inImgH, + const size_t inImgW, + const size_t outImgH, + const size_t outImgW, + const size_t numChannels, + const real ratioH, + const real ratioW) { + CHECK(dynamic_cast(&in)); + + const size_t outputW = getWidth(); + const size_t outputH = getHeight(); + const size_t inputW = in.getWidth(); + const size_t inputH = in.getHeight(); + + real* outData = getData(); + const real* inData = in.getData(); + + if (inImgH == outImgW && inImgW == outImgW) { + this->copyFrom(in); + } else { + hl_bilinear_forward(inData, + inImgH, + inImgW, + inputH, + inputW, + outData, + outImgH, + outImgW, + outputH, + outputW, + numChannels, + ratioH, + ratioW); + } +} + +void GpuMatrix::bilinearBackward(const Matrix& out, + const size_t outImgH, + const size_t outImgW, + const size_t inImgH, + const size_t inImgW, + const size_t numChannels, + const real ratioH, + const real ratioW) { + CHECK(dynamic_cast(&out)); + + const size_t inputW = getWidth(); + const size_t inputH = getHeight(); + const size_t outputW = out.getWidth(); + const size_t outputH = out.getHeight(); + + real* inGrad = getData(); + const real* outGrad = out.getData(); + + if (outImgH == inImgH && outImgW == inImgW) { + this->add(const_cast(out)); + } else { + hl_bilinear_backward(inGrad, + inImgH, + inImgW, + inputH, + inputW, + outGrad, + outImgH, + outImgW, + outputH, + outputW, + numChannels, + ratioH, + ratioW); + } +} + +void GpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) { +#ifdef PADDLE_WITH_CUDA + GpuMatrix* outputPtr = dynamic_cast(&output); + auto labelPtr = dynamic_cast(&label); + + CHECK(outputPtr && labelPtr) << "Invalid argument pointer"; + CHECK(labelPtr->format_ == SPARSE_CSR) << "Matrix format not supported"; + CHECK(height_ == outputPtr->height_ && width_ == 1 && + outputPtr->width_ == labelPtr->getWidth() && + outputPtr->height_ == labelPtr->getHeight()) + << "Matrix dimensions are not equal"; + + real* output_d = outputPtr->data_; + real* entropy_d = data_; + hl_sparse_matrix_s mat_d = labelPtr->sMatrix_.get(); + hl_matrix_multi_binary_cross_entropy( + output_d, entropy_d, mat_d, height_, outputPtr->width_); +#endif +} + +void GpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { +#ifdef PADDLE_WITH_CUDA + GpuMatrix* outputPtr = dynamic_cast(&output); + auto labelPtr = dynamic_cast(&label); + + CHECK(outputPtr && labelPtr) << "Invalid argument pointer"; + CHECK(labelPtr->format_ == SPARSE_CSR) << "Matrix format not supported"; + CHECK(height_ == outputPtr->height_ && width_ == outputPtr->width_ && + outputPtr->width_ == labelPtr->getWidth() && + outputPtr->height_ == labelPtr->getHeight()) + << "Matrix dimensions are not equal"; + + real* output_d = outputPtr->data_; + real* grad_d = data_; + hl_sparse_matrix_s mat_d = labelPtr->sMatrix_.get(); + hl_matrix_multi_binary_cross_entropy_bp( + output_d, grad_d, mat_d, height_, width_); +#endif +} + +void GpuMatrix::vol2Col(real* dataSrc, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW) { + hl_matrix_vol2Col(dataSrc, + channels, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData()); +} + +void GpuMatrix::col2Vol(real* dataDst, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta) { + hl_matrix_col2Vol(dataDst, + channels, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData(), + alpha, + beta); +} + +/** + * CpuMatrix + */ + +CpuMatrix::CpuMatrix(size_t height, size_t width, bool trans) + : Matrix(std::make_shared(height * width * sizeof(real)), + height, + width, + trans, + false) {} + +CpuMatrix::~CpuMatrix() {} + +void CpuMatrix::zeroMem() { + CHECK(data_ != NULL); + if (isContiguous()) { + memset(data_, 0, height_ * width_ * sizeof(real)); + } else { + BaseMatrix::zero(); + } +} +void CpuMatrix::resetOne() { + CHECK(data_ != NULL); + BaseMatrix::one(); +} + +void CpuMatrix::copyFrom(const Matrix& src) { + CHECK(isContiguous()); + if (typeid(src) == typeid(GpuMatrix)) { + CHECK(src.isContiguous()); + CHECK(elementCnt_ == src.getElementCnt()); + hl_memcpy_device2host( + data_, const_cast(src.getData()), sizeof(real) * elementCnt_); + } else if (typeid(src) == typeid(CpuMatrix) || + typeid(src) == typeid(SharedCpuMatrix)) { + CHECK(src.isContiguous()); + CHECK(elementCnt_ == src.getElementCnt()); + memcpy(data_, src.getData(), sizeof(real) * elementCnt_); + } else if (typeid(src) == typeid(CpuSparseMatrix)) { + CHECK_GE(elementCnt_, src.getElementCnt()); + copyFrom(dynamic_cast(const_cast(src))); + } else { + LOG(FATAL) << "Wrong"; + } +} + +void CpuMatrix::copyFrom(CpuSparseMatrix& src) { + CHECK(isContiguous()); + CHECK(height_ == src.getHeight()); + CHECK(width_ == src.getWidth()); + memset(data_, 0, sizeof(real) * height_ * width_); + if (src.getValueType() == FLOAT_VALUE) { + if (src.getFormat() == SPARSE_CSC) { + int* rows = src.getRows(); + real* vals = src.getValue(); + for (size_t i = 0; i < width_; i++) { + for (size_t j = src.getColStartIdx(i); j < src.getColStartIdx(i + 1); + j++) { + data_[rows[j] * width_ + i] = vals[j]; + } + } + } else { + int* cols = src.getCols(); + real* vals = src.getValue(); + for (size_t i = 0; i < height_; i++) { + for (size_t j = src.getRowStartIdx(i); j < src.getRowStartIdx(i + 1); + j++) { + data_[i * width_ + cols[j]] = vals[j]; + } + } + } + } else { + if (src.getFormat() == SPARSE_CSC) { + int* rows = src.getRows(); + for (size_t i = 0; i < width_; i++) { + for (size_t j = src.getColStartIdx(i); j < src.getColStartIdx(i + 1); + j++) { + data_[rows[j] * width_ + i] = 1.0; + } + } + } else { + int* cols = src.getCols(); + for (size_t i = 0; i < height_; i++) { + for (size_t j = src.getRowStartIdx(i); j < src.getRowStartIdx(i + 1); + j++) { + data_[i * width_ + cols[j]] = 1.0; + } + } + } + } +} + +void CpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { + CHECK(isContiguous()); + CHECK(src.isContiguous()); + CHECK(elementCnt_ == src.getElementCnt()); + if (typeid(src) == typeid(GpuMatrix)) { + hl_memcpy_async(this->getData(), + const_cast(src.getData()), + sizeof(real) * elementCnt_, + stream); + // There is a need to add synchronization to ensure that the data is copied. + hl_stream_synchronize(stream); + } else if (typeid(src) == typeid(CpuMatrix)) { + memcpy(data_, src.getData(), sizeof(real) * elementCnt_); + } else { + LOG(FATAL) << "Wrong"; + } +} + +void CpuMatrix::copyFrom(const real* cpuSrc, size_t size) { + CHECK(isContiguous()); + CHECK(size <= elementCnt_); + memcpy(data_, cpuSrc, sizeof(real) * size); +} + +void CpuMatrix::copyFrom(const real* cpuSrc, const int64_t* seq) { + CHECK(isContiguous()); + for (size_t i = 0; i < height_; i++) { + memcpy(data_ + i * width_, cpuSrc + seq[i] * width_, sizeof(real) * width_); + } +} + +void CpuMatrix::copyFrom(const IVector& src) { + CHECK(isContiguous()); + CHECK(elementCnt_ == src.getSize()) + << "the src and dst should have same size."; + const int* cpuSrc = NULL; + IVectorPtr tmp; + if (src.useGpu()) { + CpuIVector tmp(src.getSize()); + tmp.copyFrom(src); + cpuSrc = tmp.getData(); + } else { + cpuSrc = src.getData(); + } + for (size_t i = 0; i < elementCnt_; ++i) { + data_[i] = cpuSrc[i]; + } +} + +void CpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) { + size_t height = getHeight(); + size_t width = getWidth(); + CHECK_EQ(b.getWidth(), width); + const int* index = rowIndex.getData(); + for (size_t i = 0; i < height; i++) { + CHECK_LT(static_cast(index[i]), b.getHeight()); + real* src = b.getData() + index[i] * width; + real* dst = getData() + i * width; + memcpy(dst, src, sizeof(real) * width); + } +} + +MatrixPtr CpuMatrix::clone(size_t height, size_t width, bool useGpu) { + CHECK(isContiguous()); + + if (height == 0 && width == 0) { + height = height_; + width = width_; + } + + CHECK(width && height); + + if (useGpu) { + return std::make_shared(height, width); + } else { + return std::make_shared(height, width); + } +} + +void CpuMatrix::resize(size_t newHeight, size_t newWidth) { + size_t newSize = newHeight * newWidth; + if (NULL == memoryHandle_.get() || + newSize * sizeof(real) > memoryHandle_->getAllocSize()) { + memoryHandle_ = std::make_shared(newSize * sizeof(real)); + data_ = reinterpret_cast(memoryHandle_->getBuf()); + } + + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newSize; + stride_ = width_; +} + +real CpuMatrix::getElement(size_t x, size_t y) const { + return data_[x * stride_ + y]; +} + +real CpuMatrix::getSum() { + CHECK(isContiguous()); + double sum = 0; + for (size_t i = 0; i < height_; ++i) { + for (size_t j = 0; j < width_; ++j) { + sum += data_[i * width_ + j]; + } + } + return sum; +} + +void CpuMatrix::accumulateColSum(Matrix& src) { + CHECK_EQ(getWidth(), src.getWidth()); + CHECK_EQ(getHeight(), (size_t)1); + + sumCols(src, /* scaleSum= */ 1, /* scaleDest= */ 1); +} + +real CpuMatrix::getAbsSum() { + CHECK(isContiguous()); + double sum = 0; + for (size_t i = 0; i < height_; ++i) { + for (size_t j = 0; j < width_; ++j) { + sum += fabs(data_[i * width_ + j]); + } + } + return sum; +} + +MatrixPtr CpuMatrix::getTranspose() { + if (memoryHandle_.get() != NULL) { + return std::make_shared( + std::dynamic_pointer_cast(memoryHandle_), + height_, + width_, + true); + } else { + MatrixPtr copy_T(new CpuMatrix(data_, height_, width_, true)); + return copy_T; + } +} + +void CpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { + if (memAlloc) { + matTrans = std::make_shared(width_, height_); + } else { + CHECK(matTrans != NULL); + CHECK_EQ(matTrans->getHeight(), width_); + CHECK_EQ(matTrans->getWidth(), height_); + } + real* dataTrans = matTrans->getData(); + real* data = getData(); + int lda = getStride(); + int ldc = matTrans->getStride(); + + for (size_t i = 0; i < height_; i++) { + for (size_t j = 0; j < width_; j++) { + dataTrans[j * ldc + i] = data[i * lda + j]; + } + } +} + +void CpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) { + if (memAlloc) { + matRot = std::make_shared(width_, height_); + } else { + CHECK(matRot != NULL); + CHECK_EQ(matRot->getHeight(), width_); + CHECK_EQ(matRot->getWidth(), height_); + } + real* dataRot = matRot->getData(); + real* data = getData(); + + for (size_t i = 0; i < height_; i++) { + for (size_t j = 0; j < width_; j++) { + if (clockWise) { + dataRot[j * height_ + i] = data[(height_ - i - 1) * width_ + j]; + } else { + dataRot[j * height_ + i] = data[i * width_ + (width_ - j - 1)]; + } + } + } +} + +MatrixPtr CpuMatrix::getInverse() { + MatrixPtr matInv; + inverse(matInv, true); + return matInv; +} + +void CpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) { + CHECK_EQ(height_, width_); + + if (memAlloc) { + matInv = std::make_shared(height_, width_); + } else { + CHECK(matInv != NULL); + } + + CHECK_EQ(height_, matInv->getHeight()); + CHECK_EQ(width_, matInv->getWidth()); + matInv->copyFrom(*this); + + real* data = getData(); + real* dataInv = matInv->getData(); + int ldc = matInv->getStride(); + + if (height_ == 1) { + CHECK_NE(*data, 0); + *dataInv = 1.0 / (*data); + return; + } + + /* Compute the LU decomposition of the matrix */ + std::vector ipiv(height_); + CBLAS_ORDER order = (matInv->isTransposed() ? CblasColMajor : CblasRowMajor); + int info = getrf(order, height_, height_, dataInv, ldc, ipiv.data()); + CHECK_EQ(info, 0); + + /* Compute the inverse of the matrix given its LU decompsotion */ + info = getri(order, height_, dataInv, ldc, ipiv.data()); + CHECK_EQ(info, 0); +} + +void CpuMatrix::upsampleForward(Matrix& input, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW) { + real* inputData = input.getData(); + real* maskData = mask.getData(); + real* outData = data_; + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + size_t batch = input.getHeight(); + CHECK(inLength == input.getWidth() / channels); + CHECK_EQ(batch, this->getHeight()); + CHECK_EQ(channels * outLength, this->getWidth()); + + for (size_t k = 0; k < batch; k++) { + for (size_t c = 0; c < channels; c++) { + for (size_t i = 0; i < inLength; i++) { + size_t out_index = static_cast(maskData[i]); + if (out_index >= outLength) { + LOG(FATAL) << "upsample index " << out_index << " out of range."; + } + outData[out_index] = inputData[i]; + } + inputData += inLength; + maskData += inLength; + outData += outLength; + } + } +} + +void CpuMatrix::upsampleBackward(Matrix& outputGrad, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW) { + real* outputGradData = outputGrad.getData(); + real* maskData = mask.getData(); + real* inputGradData = data_; + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + size_t batch = outputGrad.getHeight(); + CHECK(inLength == this->getWidth() / channels); + CHECK_EQ(batch, this->getHeight()); + CHECK_EQ(channels * outLength, outputGrad.getWidth()); + + for (size_t k = 0; k < batch; k++) { + for (size_t c = 0; c < channels; c++) { + for (size_t i = 0; i < inLength; i++) { + size_t out_index = static_cast(maskData[i]); + if (out_index >= outLength) { + LOG(FATAL) << "upsample index " << out_index << " out of range."; + } + inputGradData[i] = outputGradData[out_index]; + } + inputGradData += inLength; + maskData += inLength; + outputGradData += outLength; + } + } +} + +void CpuMatrix::maxPoolForward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + MatrixPtr maskMatP) { + real* inputData = inputMat.getData(); + real* outData = data_; + real* maskData = NULL; + size_t num = inputMat.getHeight(); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + CHECK(inLength == inputMat.getWidth() / channels); + CHECK_EQ(num, this->getHeight()); + CHECK_EQ(channels * outLength, this->getWidth()); + size_t outStride = getStride(); + + if (maskMatP != NULL) { + maskData = maskMatP->getData(); + CHECK_EQ(channels * outLength, maskMatP->getWidth()); + } + + /* pool max one by one */ + for (size_t n = 0; n < num; ++n) { // frame by frame + if (!isContiguous()) { + outData = data_ + n * outStride; + } + for (size_t c = 0; c < channels; ++c) { // channel by channel + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = hstart + sizeY; + hstart = hstart < 0 ? 0 : hstart; + hend = hend < (int)imgSizeH ? hend : (int)imgSizeH; + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = wstart + sizeX; + wstart = wstart < 0 ? 0 : wstart; + wend = wend < (int)imgSizeW ? wend : (int)imgSizeW; + + real maxval = -(real)FLT_MAX; + int max_index = -1; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (maxval < inputData[h * imgSizeW + w]) { + maxval = inputData[h * imgSizeW + w]; + max_index = h * imgSizeW + w; + } + } + } + + outData[ph * outputW + pw] = maxval; + if (maskData != NULL) maskData[ph * outputW + pw] = max_index; + } + } + // compute offset + inputData += inLength; + outData += outLength; + + if (maskData != NULL) maskData += outLength; + } + } +} + +void CpuMatrix::maxPoolBackward(Matrix& image, + size_t imgSizeH, + size_t imgSizeW, + Matrix& outGrad, + Matrix& outV, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW) { + size_t num = image.getHeight(); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + size_t channels = size_t(width_ / inLength); + CHECK(image.getWidth() == inLength * channels); + CHECK(image.getHeight() == height_ && image.getWidth() == width_); + CHECK(outV.getHeight() == outGrad.getHeight() && + outV.getWidth() == outGrad.getWidth()); + + real* tgtGrad = data_; + real* inData = image.getData(); + real* otData = outV.getData(); + real* otGrad = outGrad.getData(); + + size_t outStride = outV.getStride(); + real* origOutData = otData; + real* origOutGrad = otGrad; + + for (size_t n = 0; n < num; ++n) { + if (!outV.isContiguous()) { + otData = origOutData + n * outStride; + otGrad = origOutGrad + n * outStride; + } + for (size_t c = 0; c < channels; ++c) { + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = std::min(wstart + sizeX, imgSizeW); + wstart = std::max(wstart, 0); + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + tgtGrad[h * imgSizeW + w] = + scaleTargets * tgtGrad[h * imgSizeW + w] + + scaleOutput * otGrad[ph * outputW + pw] * + (inData[h * imgSizeW + w] == otData[ph * outputW + pw]); + } + } + } + } + // offset + inData += inLength; + tgtGrad += inLength; + otData += outLength; + otGrad += outLength; + } + } +} + +void CpuMatrix::avgPoolForward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + bool excludeMode) { + // The main loop + size_t num = input.getHeight(); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + CHECK(inLength * channels == input.getWidth()); + CHECK(outLength * channels * num == height_ * width_); + real* tgtData = data_; + real* inData = input.getData(); + + for (size_t n = 0; n < num; ++n) { + if (!isContiguous()) { + tgtData = data_ + n * getStride(); + } + for (size_t c = 0; c < channels; ++c) { + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = std::min(wstart + sizeX, imgSizeW); + wstart = std::max(wstart, 0); + tgtData[ph * outputW + pw] = 0; // clear + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + tgtData[ph * outputW + pw] += inData[h * imgSizeW + w]; + } + } + int poolSize = + excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; + CHECK(poolSize); + tgtData[ph * outputW + pw] /= poolSize; + } + } + // compute offset + inData += inLength; + tgtData += outLength; + } + } +} + +void CpuMatrix::avgPoolBackward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW, + bool excludeMode) { + size_t num = input.getHeight(); + size_t channels = input.getWidth() / outputH / outputW; + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + CHECK(inLength * channels == getWidth()); + real* inData = input.getData(); + real* outData = getData(); + + for (size_t n = 0; n < num; ++n) { + if (!input.isContiguous()) { + inData = input.getData() + n * input.getStride(); + } + for (size_t c = 0; c < channels; ++c) { + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = std::min(wstart + sizeX, imgSizeW); + wstart = std::max(wstart, 0); + int poolSize = + excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; + CHECK(poolSize); + + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + outData[h * imgSizeW + w] += inData[ph * outputW + pw] / poolSize; + } + } + } + } + // offset + outData += inLength; + inData += outLength; + } + } +} + +void CpuMatrix::maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + real* inputData = inputMat.getData(); + real* outData = getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t num = inputMat.getHeight(); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + CHECK(inLength == inputMat.getWidth() / channels); + CHECK_EQ(num, this->getHeight()); + CHECK_EQ(channels * outLength, this->getWidth()); + size_t outStride = getStride(); + + /* initialize the data_ */ + for (size_t i = 0; i < height_; i++) { + for (size_t j = 0; j < width_; j++) { + outData[(i)*outStride + j] = -(real)FLT_MAX; + maxPoolIdxData[(i)*outStride + j] = -1; + } + } + + /* pool max one by one */ + for (size_t n = 0; n < num; ++n) { // frame by frame + if (!isContiguous()) { + outData = getData() + n * outStride; + maxPoolIdxData = maxPoolIdx.getData() + n * outStride; + } + for (size_t c = 0; c < channels; ++c) { // channel by channel + for (size_t pd = 0; pd < outputD; ++pd) { + int dstart = pd * strideD - paddingD; + int dend = std::min(dstart + sizeZ, imgSizeD); + dstart = std::max(dstart, 0); + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = std::min(wstart + sizeX, imgSizeW); + wstart = std::max(wstart, 0); + int maxIdx = -1; + real maxOutData = outData[(pd * outputH + ph) * outputW + pw]; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (maxOutData < + inputData[(d * imgSizeH + h) * imgSizeW + w]) { + maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w]; + maxIdx = (d * imgSizeH + h) * imgSizeW + w; + } + } + } + } + outData[(pd * outputH + ph) * outputW + pw] = maxOutData; + maxPoolIdxData[(pd * outputH + ph) * outputW + pw] = maxIdx; + } + } + } + // compute offset + inputData += inLength; + outData += outLength; + maxPoolIdxData += outLength; + } + } +} + +void CpuMatrix::maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + size_t num = getHeight(); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + size_t channels = size_t(width_ / inLength); + CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() && + maxPoolIdx.getWidth() == outGrad.getWidth()); + + real* tgtGrad = getData(); + real* otGrad = outGrad.getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t outStride = outGrad.getStride(); + + for (size_t n = 0; n < num; ++n) { + if (!outGrad.isContiguous()) { + otGrad = outGrad.getData() + n * outStride; + maxPoolIdxData = maxPoolIdx.getData() + n * outStride; + } + for (size_t c = 0; c < channels; ++c) { + for (size_t pd = 0; pd < outputD; ++pd) { + for (size_t ph = 0; ph < outputH; ++ph) { + for (size_t pw = 0; pw < outputW; ++pw) { + const size_t index = (pd * outputH + ph) * outputW + pw; + const size_t tgtIdx = static_cast(maxPoolIdxData[index]); + tgtGrad[tgtIdx] = + scaleTargets * tgtGrad[tgtIdx] + scaleOutput * otGrad[index]; + } + } + } + // offset + tgtGrad += inLength; + otGrad += outLength; + maxPoolIdxData += outLength; + } + } +} + +void CpuMatrix::avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + // The main loop + size_t num = input.getHeight(); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + CHECK(inLength * channels == input.getWidth()); + CHECK(outLength * channels * num == height_ * width_); + real* tgtData = getData(); + real* inData = input.getData(); + + for (size_t n = 0; n < num; ++n) { + if (!isContiguous()) { + tgtData = data_ + n * getStride(); + } + for (size_t c = 0; c < channels; ++c) { + for (size_t pd = 0; pd < outputD; ++pd) { + int dstart = pd * strideD - paddingD; + int dend = std::min(dstart + sizeZ, imgSizeD); + dstart = std::max(dstart, 0); + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = std::min(wstart + sizeX, imgSizeW); + wstart = std::max(wstart, 0); + + tgtData[(pd * outputH + ph) * outputW + pw] = 0; // clear + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + tgtData[(pd * outputH + ph) * outputW + pw] += + inData[(d * imgSizeH + h) * imgSizeW + w]; + } + } + } + int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); + CHECK(poolSize); + tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize; + } + } + } + // compute offset + inData += inLength; + tgtData += outLength; + } + } +} + +void CpuMatrix::avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + size_t num = input.getHeight(); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + size_t channels = input.getWidth() / outLength; + CHECK(inLength * channels == getWidth()); + real* inData = input.getData(); + real* outData = getData(); + + for (size_t n = 0; n < num; ++n) { + if (!input.isContiguous()) { + inData = input.getData() + n * input.getStride(); + } + for (size_t c = 0; c < channels; ++c) { + for (size_t pd = 0; pd < outputD; ++pd) { + int dstart = pd * strideD - paddingD; + int dend = std::min(dstart + sizeZ, imgSizeD); + dstart = std::max(dstart, 0); + for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); + for (size_t pw = 0; pw < outputW; ++pw) { + int wstart = pw * strideW - paddingW; + int wend = std::min(wstart + sizeX, imgSizeW); + wstart = std::max(wstart, 0); + int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); + CHECK(poolSize); + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + outData[(d * imgSizeH + h) * imgSizeW + w] += + inData[(pd * outputH + ph) * outputW + pw] / poolSize; + } + } + } + } + } + } + // offset + outData += inLength; + inData += outLength; + } + } +} + +/** + * Input: one or more sequences. Each sequence contains some instances. + * Output: output size is the number of input sequences (NOT input instances). + * output[i] is set to max_{for each instance in this sequence}{input[i]} + */ +void CpuMatrix::maxSequenceForward(Matrix& input, + const IVector& sequence, + IVector& index) { + CHECK(dynamic_cast(&input)); + CHECK(dynamic_cast(&sequence)); + CHECK(dynamic_cast(&index)); + + real* outData = getData(); + real* inputData = input.getData(); + const int* starts = sequence.getData(); + int* maxIndex = index.getData(); + size_t numSequences = getHeight(); + size_t dim = getWidth(); + + CHECK_EQ(dim, input.getWidth()); + CHECK_EQ(numSequences, sequence.getSize() - 1); + CHECK_EQ(starts[numSequences], (int)input.getHeight()); + CHECK_EQ(numSequences * dim, index.getSize()); + + for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) { + // current sequence, loop for each input instance + // (1) first instance: do not need compare, copy value to outV directly + for (size_t k = 0; k < dim; ++k) { + outData[sequenceId * dim + k] = inputData[starts[sequenceId] * dim + k]; + maxIndex[sequenceId * dim + k] = starts[sequenceId]; + } + // (2) other instance in same sequence + for (int insId = starts[sequenceId] + 1; insId < starts[sequenceId + 1]; + ++insId) { + // insId is the index on all instances + for (size_t k = 0; k < dim; ++k) { + // for each dim + if (inputData[insId * dim + k] > outData[sequenceId * dim + k]) { + // update max value and record index + outData[sequenceId * dim + k] = inputData[insId * dim + k]; + maxIndex[sequenceId * dim + k] = insId; + } + } + } + } +} + +void CpuMatrix::maxSequenceBackward(Matrix& outputGrad, + const IVector& sequence, + IVector& index) { + CHECK(dynamic_cast(&outputGrad)); + CHECK(dynamic_cast(&sequence)); + CHECK(dynamic_cast(&index)); + + real* inputGrad = getData(); + real* outGrad = outputGrad.getData(); + int* maxIndex = index.getData(); + size_t dim = getWidth(); + size_t numSequences = sequence.getSize() - 1; + + CHECK_EQ(dim, outputGrad.getWidth()); + CHECK_EQ(numSequences, outputGrad.getHeight()); + CHECK_EQ(numSequences * dim, index.getSize()); + + for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) { + // current sequence + for (size_t j = 0; j < dim; ++j) { + // each dim + int insId = maxIndex[sequenceId * dim + j]; + inputGrad[insId * dim + j] += outGrad[sequenceId * dim + j]; + } + } +} + +inline void vecAddTo(real* a, const real* b, size_t len) { + for (unsigned int i = 0; i < len; ++i) { + a[i] += b[i]; + } +} + +inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { + for (unsigned int i = 0; i < len; ++i) { + a[i] += scaleB * b[i]; + } +} + +inline void colVecAddTo( + real* a, const real* b, size_t len, size_t aWidth, size_t bWidth) { + for (unsigned int i = 0; i < len; ++i) { + a[i * aWidth] += b[i * bWidth]; + } +} + +inline void colVecAddTo( + real* a, real* b, real c, size_t len, size_t aWidth, size_t bWidth) { + for (unsigned int i = 0; i < len; ++i) { + a[i * aWidth] += b[i * bWidth] * c; + } +} + +void CpuMatrix::addBias(Matrix& b, real scale) { + CHECK(b.useGpu_ == false) << "Matrix type are not equal"; + + CHECK_EQ(b.getHeight(), (size_t)1); + CHECK_EQ(width_, b.getWidth()); + real* aData = getData(); + real* bData = b.getData(); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + + if (scale == 1 && getStride() % 32 == 0) { // use libaddto + // @TODO(yuyang18) Make input addr can be unaligned. + // So merge this if and else + CHECK_EQ((size_t)aData % 32, 0UL); + CHECK_EQ((size_t)bData % 32, 0UL); + for (size_t i = 0; i < numSamples; i++) { + simd::addTo(aData + i * getStride(), bData, dim); + } + } else { + for (size_t i = 0; i < numSamples; i++) { + for (size_t j = 0; j < dim; j++) { + aData[i * getStride() + j] += scale * bData[j]; + } + } + } +} + +void CpuMatrix::addSharedBias(Matrix& b, real scale) { + CHECK_EQ(b.getHeight(), (size_t)1); + real* aData = getData(); + real* bData = b.getData(); + size_t numSamples = getHeight(); + size_t channel = b.getWidth(); + CHECK_EQ(getWidth() % channel, 0UL); + size_t dim = getWidth() / channel; + + for (size_t i = 0; i < numSamples; i++) { + for (size_t c = 0; c < channel; c++) { + for (size_t j = 0; j < dim; j++) { + aData[i * getStride() + c * dim + j] += scale * bData[c]; + } + } + } +} + +void CpuMatrix::collectBias(Matrix& a, real scale) { + CHECK_EQ(getHeight(), (size_t)1); + CHECK_EQ(width_, a.getWidth()); + CpuSparseMatrix* aptr = dynamic_cast(&a); + if (!aptr) { + sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1); + } else { + size_t nnz = aptr->getElementCnt(); + int* cols = aptr->getCols(); + real* A = aptr->getValue(); + real* B = getData(); + for (size_t i = 0; i < nnz; i++) { + B[cols[i]] += scale * A[i]; + } + } +} + +void CpuMatrix::collectSharedBias(Matrix& a, real scale) { + CHECK_EQ(getHeight(), (size_t)1); + real* B = getData(); + real* A = a.getData(); + size_t numSamples = a.getHeight(); + size_t channel = getWidth(); + CHECK_EQ(a.getWidth() % channel, 0UL); + size_t dim = a.getWidth() / channel; + for (size_t i = 0; i < numSamples; i++) { + for (size_t c = 0; c < channel; c++) { + for (size_t j = 0; j < dim; j++) { + B[c] += scale * A[i * channel * dim + c * dim + j]; + } + } + } +} + +void CpuMatrix::sequenceAvgForward(Matrix& a, + const IVector& startsPos, + int mode) { + size_t height = getHeight(); + size_t width = getWidth(); + CHECK_EQ(height, startsPos.getSize() - 1); + CHECK_EQ(width, a.getWidth()); + real* dst = getData(); + real* src = a.getData(); + const int* starts = startsPos.getData(); + MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false); + MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false); + for (size_t i = 0; i < height; i++) { + int sequenceLength = starts[i + 1] - starts[i]; + if (0 == sequenceLength) { + // empty sequence + continue; + } + outMtx->setData(dst + i * width); + dataMtx->setData(src + starts[i] * width, sequenceLength, width); + if (mode == 0) { + // plain average + outMtx->sumCols(*dataMtx, + (real)1 / (real)sequenceLength, + /* scaleDest= */ 1); + } else if (mode == 1) { + // sum instead of average + outMtx->sumCols(*dataMtx, /* scaleSum= */ 1, /* scaleDest= */ 1); + } else if (mode == 2) { + // divide by square root of sequenceLength + outMtx->sumCols(*dataMtx, + (real)1 / std::sqrt(sequenceLength), + /* scaleDest= */ 1); + } else { + LOG(FATAL) << "should not reach here"; + } + } +} + +void CpuMatrix::sequenceAvgBackward(Matrix& a, + const IVector& startsPos, + int mode) { + size_t height = a.getHeight(); + size_t width = getWidth(); + CHECK_EQ(height, startsPos.getSize() - 1); + CHECK_EQ(width, a.getWidth()); + real* dst = getData(); + real* src = a.getData(); + const int* starts = startsPos.getData(); + MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false); + MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false); + for (size_t i = 0; i < height; ++i) { + int sequenceLength = starts[i + 1] - starts[i]; + if (0 == sequenceLength) { + // empty sequence + continue; + } + outMtx->setData(dst + starts[i] * width, sequenceLength, width); + dataMtx->setData(src + i * width); + if (mode == 0) { + // plain average + outMtx->addBias(*dataMtx, 1.0f / sequenceLength); + } else if (mode == 1) { + // sum instead of average + outMtx->addBias(*dataMtx, 1.0f); + } else if (mode == 2) { + // divide by square root of sequenceLength + outMtx->addBias(*dataMtx, 1.0f / std::sqrt(sequenceLength)); + } else { + LOG(FATAL) << "should not reach here"; + } + } +} + +/* this = scaleAB*(a*b) + scaleT*this*/ +void CpuMatrix::mul(const Matrix& a, + const Matrix& b, + real scaleAB, + real scaleT) { + CHECK(!isTransposed()) << "Not supported"; + const auto a_ptr = dynamic_cast(&a); + const auto b_ptr = dynamic_cast(&b); + const auto a_ptr_s = dynamic_cast(&a); + const auto b_ptr_s = dynamic_cast(&b); + + if (a_ptr && b_ptr) { + mul((CpuMatrix*)a_ptr, (CpuMatrix*)b_ptr, scaleAB, scaleT); + } else if (a_ptr_s && b_ptr) { + mul((CpuSparseMatrix*)a_ptr_s, (CpuMatrix*)b_ptr, scaleAB, scaleT); + } else if (a_ptr && b_ptr_s) { + mul((CpuMatrix*)a_ptr, (CpuSparseMatrix*)b_ptr_s, scaleAB, scaleT); + } else { + LOG(FATAL) << "Not supported"; + } +} + +void CpuMatrix::mul(CpuSparseMatrix* a, + CpuMatrix* b, + real scaleAB, + real scaleT) { + if (dynamic_cast(b)) { + return mul(a, dynamic_cast(b), this, scaleAB, scaleT); + } else if (dynamic_cast(b)) { + return mul(a, dynamic_cast(b), this, scaleAB, scaleT); + } else { + return mul(a, b, this, scaleAB, scaleT); + } +} + +void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { + CHECK(!isTransposed()) << "Not supported"; + + size_t a_col, b_col, a_row, b_row; + bool a_trans, b_trans; + if (!a->isTransposed()) { + a_col = a->getWidth(); + a_row = a->getHeight(); + a_trans = false; + } else { + a_col = a->getHeight(); + a_row = a->getWidth(); + a_trans = true; + } + if (!b->isTransposed()) { + b_col = b->getWidth(); + b_row = b->getHeight(); + b_trans = false; + } else { + b_col = b->getHeight(); + b_row = b->getWidth(); + b_trans = true; + } + + CHECK_EQ(a_col, b_row); + CHECK_EQ(a_row, getHeight()); + CHECK_EQ(b_col, getWidth()); + + real* A = a->getData(); + real* B = b->getData(); + real* C = getData(); + + int M = getHeight(); + int N = getWidth(); + int K = a_col; + int lda = a->getStride(); + int ldb = b->getStride(); + int ldc = getStride(); + BlasGemm::compute( + a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc); +} + +void CpuMatrix::mul( + CpuMatrix* a, CpuMatrix* b, CpuSparseMatrix* c, real scaleAB, real scaleT) { + CHECK(!c->isTransposed()) << "Not supported"; + CHECK_EQ(c->getValueType(), FLOAT_VALUE); + + real* A = a->getData(); + real* B = b->getData(); + real* C = c->getValue(); + int* rows = c->getRows(); + int* cols = c->getCols(); + size_t height = c->getHeight(); + size_t width = c->getWidth(); + if (scaleT == 0) { + c->zeroMem(); + } + + if (!a->isTransposed() && !b->isTransposed()) { + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), m); + CHECK_EQ(a->getHeight(), height); + CHECK_EQ(b->getWidth(), width); + if (c->getFormat() == SPARSE_CSC) { + for (size_t i = 0; i < width; i++) { + size_t start = c->getColStartIdx(i); + size_t end = c->getColStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + real sum = 0; + size_t rowIdx = rows[j]; + for (size_t k = 0; k < m; k++) { + sum += A[rowIdx * m + k] * B[k * width + i]; + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + } else { + for (size_t i = 0; i < height; i++) { + size_t start = c->getRowStartIdx(i); + size_t end = c->getRowStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + real sum = 0; + size_t colIdx = cols[j]; + for (size_t k = 0; k < m; k++) { + sum += A[i * m + k] * B[k * width + colIdx]; + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + } + } else if (a->isTransposed() && !b->isTransposed()) { + size_t m = a->getHeight(); + CHECK_EQ(m, b->getHeight()); + CHECK_EQ(b->getWidth(), width); + CHECK_EQ(a->getWidth(), height); + + if (c->getFormat() == SPARSE_CSC) { + for (size_t i = 0; i < width; i++) { + size_t start = c->getColStartIdx(i); + size_t end = c->getColStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + real sum = 0; + size_t rowIdx = rows[j]; + for (size_t k = 0; k < m; k++) { + sum += A[k * height + rowIdx] * B[k * width + i]; + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + } else { + for (size_t i = 0; i < height; i++) { + int start = c->getRowStartIdx(i); + int end = c->getRowStartIdx(i + 1); + for (int j = start; j < end; j++) { + real sum = 0; + size_t colIdx = cols[j]; + for (size_t k = 0; k < m; k++) { + sum += A[k * height + i] * B[k * width + colIdx]; + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + } + } else if (!a->isTransposed() && b->isTransposed()) { + size_t m = a->getWidth(); + CHECK_EQ(b->getWidth(), m); + CHECK_EQ(a->getHeight(), height); + CHECK_EQ(b->getHeight(), width); + if (c->getFormat() == SPARSE_CSR) { + for (size_t i = 0; i < height; i++) { + size_t start = c->getRowStartIdx(i); + size_t end = c->getRowStartIdx(i + 1); + for (size_t j = start; j < end; j++) { + real sum = 0; + size_t colIdx = cols[j]; + for (size_t k = 0; k < m; k++) { + sum += A[i * m + k] * B[colIdx * m + k]; + } + C[j] = scaleAB * sum + scaleT * C[j]; + } + } + } else { + LOG(FATAL) << "Not supported csc format " + "when a is not trans and b is trans"; + } + } else { + LOG(FATAL) << "Not supported"; + } +} + +void CpuMatrix::mul(CpuMatrix* a, + CpuSparseMatrix* b, + real scaleAB, + real scaleT) { + CHECK(!trans_) << "Not supported"; + CHECK(!a->isTransposed()) << "Not supported"; + CHECK(scaleT == 0 || scaleT == 1); + + // TODO(yuyang18): Maybe bug implementation here + CHECK_EQ(scaleAB, static_cast(1.0)); + + real* A = a->getData(); + real* B = b->getValue(); + real* C = getData(); + int* rows = b->getRows(); + int* cols = b->getCols(); + + if (scaleT == 0) { + zeroMem(); + } + if (b->getFormat() == SPARSE_CSC) { + if (!b->isTransposed()) { + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), m); + CHECK_EQ(a->getHeight(), height_); + CHECK_EQ(b->getWidth(), width_); + + if (b->getValueType() == NO_VALUE) { + for (size_t j = 0; j < b->getWidth(); ++j) { + int start = b->getColStartIdx(j); + int end = b->getColStartIdx(j + 1); + for (int i = start; i < end; ++i) { + colVecAddTo(C + j, A + rows[i], height_, width_, a->getWidth()); + } + } + } else if (b->getValueType() == FLOAT_VALUE) { + for (size_t j = 0; j < b->getWidth(); ++j) { + int start = b->getColStartIdx(j); + int end = b->getColStartIdx(j + 1); + for (int i = start; i < end; ++i) { + colVecAddTo( + C + j, A + rows[i], B[i], height_, width_, a->getWidth()); + } + } + } + } else /*if (b->isTransposed())*/ { + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), width_); + CHECK_EQ(a->getHeight(), height_); + CHECK_EQ(b->getWidth(), m); + if (b->getValueType() == NO_VALUE) { + for (size_t i = 0; i < b->getWidth(); ++i) { + int start = b->getColStartIdx(i); + int end = b->getColStartIdx(i + 1); + for (int j = start; j < end; ++j) { + colVecAddTo(C + rows[j], A + i, height_, width_, a->getWidth()); + } + } + } else if (b->getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < b->getWidth(); ++i) { + int start = b->getColStartIdx(i); + int end = b->getColStartIdx(i + 1); + for (int j = start; j < end; ++j) { + colVecAddTo( + C + rows[j], A + i, B[j], height_, width_, a->getWidth()); + } + } + } + } + } else { + if (!b->isTransposed()) { + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), m); + CHECK_EQ(a->getHeight(), height_); + CHECK_EQ(b->getWidth(), width_); + + if (b->getValueType() == NO_VALUE) { + for (size_t j = 0; j < b->getHeight(); ++j) { + int start = b->getRowStartIdx(j); + int end = b->getRowStartIdx(j + 1); + for (int i = start; i < end; ++i) { + colVecAddTo(C + cols[i], A + j, height_, width_, a->getWidth()); + } + } + } else if (b->getValueType() == FLOAT_VALUE) { + for (size_t j = 0; j < b->getHeight(); ++j) { + int start = b->getRowStartIdx(j); + int end = b->getRowStartIdx(j + 1); + for (int i = start; i < end; ++i) { + colVecAddTo( + C + cols[i], A + j, B[i], height_, width_, a->getWidth()); + } + } + } + } else /*if (b->isTransposed())*/ { + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), width_); + CHECK_EQ(a->getHeight(), height_); + CHECK_EQ(b->getWidth(), m); + if (b->getValueType() == NO_VALUE) { + for (size_t i = 0; i < b->getHeight(); ++i) { + int start = b->getRowStartIdx(i); + int end = b->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + colVecAddTo(C + i, A + cols[j], height_, width_, a->getWidth()); + } + } + } else if (b->getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < b->getHeight(); ++i) { + int start = b->getRowStartIdx(i); + int end = b->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + colVecAddTo( + C + i, A + cols[j], B[j], height_, width_, a->getWidth()); + } + } + } + } + } +} + +void CpuMatrix::selectRows(Matrix& table, IVector& ids) { + if (dynamic_cast(&table)) { + selectRowsImp(*dynamic_cast(&table), ids); + } else if (dynamic_cast(&table)) { + selectRowsImp(*dynamic_cast(&table), ids); + } else { + CHECK(table.isContiguous()); + selectRowsImp(*dynamic_cast(&table), ids); + } +} + +void CpuMatrix::selectElements(Matrix& table, IVector& ids) { + CHECK_EQ(table.getHeight(), ids.getSize()); + CHECK_EQ(getHeight(), ids.getSize()); + CHECK_EQ(getWidth(), 1U); + real* tableData = table.getData(); + int* idsData = ids.getData(); + for (size_t i = 0; i < table.getHeight(); i++) { + data_[i] += tableData[i * table.getWidth() + idsData[i]]; + } +} + +void CpuMatrix::addElements(Matrix& table, IVector& ids) { + CHECK_EQ(table.getHeight(), ids.getSize()); + CHECK_EQ(getHeight(), ids.getSize()); + CHECK_EQ(getWidth(), 1U); + real* tableData = table.getData(); + int* idsData = ids.getData(); + for (size_t i = 0; i < table.getHeight(); i++) { + tableData[i * table.getWidth() + idsData[i]] += data_[i]; + } +} + +// this.row[i] += table.row[ids[i]] +template +void CpuMatrix::selectRowsImp(TableMatType& table, IVector& ids) { + CHECK(!table.useGpu()); + CHECK(!ids.useGpu()); + CHECK_EQ(getHeight(), ids.getSize()); + CHECK_EQ(getWidth(), table.getWidth()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + real* a = getData(); + size_t tableSize = table.getHeight(); + int* index = ids.getData(); + + for (size_t i = 0; i < numSamples; ++i) { + if (index[i] == -1) continue; + CHECK_LT(index[i], (int)tableSize); + CHECK_GE(index[i], 0); + vecAddTo(a + i * stride_, table.getRow(index[i]), dim); + } +} + +void CpuMatrix::addToRows(Matrix& table, IVector& ids) { + if (dynamic_cast(&table)) { + addToRowsImp(*dynamic_cast(&table), ids); + } else if (dynamic_cast(&table)) { + addToRowsImp(*dynamic_cast(&table), ids); + } else if (dynamic_cast(&table)) { + addToRowsImp(*dynamic_cast(&table), ids); + } else { + CHECK(table.isContiguous()); + addToRowsImp(*dynamic_cast(&table), ids); + } +} + +// table.row[ids[i]] += this.row[i] +template +void CpuMatrix::addToRowsImp(TableMatType& table, IVector& ids) { + CHECK(!table.useGpu()); + CHECK(!ids.useGpu()); + CHECK_EQ(getHeight(), ids.getSize()); + CHECK_EQ(getWidth(), table.getWidth()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + real* a = getData(); + size_t tableSize = table.getHeight(); + int* index = ids.getData(); + + for (size_t i = 0; i < numSamples; ++i) { + if (index[i] == -1) continue; + CHECK_LT(index[i], (int)tableSize); + CHECK_GE(index[i], 0); + vecAddTo(table.getRow(index[i]), a + i * stride_, dim); + } +} + +static ThreadLocal> threadLocalColArray; + +template +void CpuMatrix::mul( + CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB, real scaleT) { + CHECK(!c->isTransposed()) << "Not supported"; + CHECK(!b->isTransposed()) << "Not supported"; + // TODO(yuyang18): Maybe bug implementation here. + CHECK(scaleAB == 1) << "Not supported"; + CHECK(scaleT == 0 || scaleT == 1) << "Not supported"; + CHECK_EQ(a->getFormat(), SPARSE_CSR) << "Not supported"; + + real* B = b->getData(); + real* C = c->getData(); + size_t height = c->getHeight(); + size_t width = c->getWidth(); + int* cols = a->getCols(); + real* values = a->getValue(); + + if (scaleT == 0) { + c->zeroMem(); + } + + if (!a->isTransposed()) { + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), m); + CHECK_EQ(a->getHeight(), height); + CHECK_EQ(b->getWidth(), width); + + if (a->getValueType() == NO_VALUE) { + if (width % 32 == 0) { // use libaddto + // @TODO(yuyang18) Make input addr can be unaligned. + // So merge this if and else + CHECK_EQ((size_t)B % 32, 0UL); + CHECK_EQ((size_t)C % 32, 0UL); + auto& colArray = *threadLocalColArray; + for (size_t i = 0; i < a->getHeight(); ++i) { + const int start = a->getRowStartIdx(i); + const int end = a->getRowStartIdx(i + 1); + size_t colNum = end - start; + colArray.resize(colNum); + for (int j = 0; j < end - start; ++j) { + colArray[j] = b->getRow(cols[j + start]); + } + simd::batchAddTo(c->getRow(i), &colArray[0], colNum, width); + } + + } else { + for (size_t i = 0; i < a->getHeight(); ++i) { + const int start = a->getRowStartIdx(i); + const int end = a->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + vecAddTo(c->getRow(i), b->getRow(cols[j]), width); + } + } + } + } else if (a->getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < a->getHeight(); ++i) { + const int start = a->getRowStartIdx(i); + const int end = a->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + vecAddTo(c->getRow(i), b->getRow(cols[j]), values[j], width); + } + } + } + } else /*if (a->isTransposed())*/ { + size_t m = a->getHeight(); + CHECK_EQ(b->getHeight(), m); + CHECK_EQ(a->getWidth(), height); + CHECK_EQ(b->getWidth(), width); + if (a->getValueType() == NO_VALUE) { + if (width % 32 == 0) { // use libaddto + // @TODO(yuyang18) Make input addr can be unaligned. + // So merge this if and else + CHECK_EQ((size_t)B % 32, 0UL); + CHECK_EQ((size_t)C % 32, 0UL); + for (size_t i = 0; i < a->getHeight(); ++i) { + const int start = a->getRowStartIdx(i); + const int end = a->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + simd::addTo(c->getRow(cols[j]), b->getRow(i), width); + } + } + + } else { + for (size_t i = 0; i < a->getHeight(); ++i) { + const int start = a->getRowStartIdx(i); + const int end = a->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + vecAddTo(c->getRow(cols[j]), b->getRow(i), width); + } + } + } + } else if (a->getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < a->getHeight(); ++i) { + const int start = a->getRowStartIdx(i); + const int end = a->getRowStartIdx(i + 1); + for (int j = start; j < end; ++j) { + vecAddTo(c->getRow(cols[j]), b->getRow(i), values[j], width); + } + } + } + } +} + +// instantiation mul() called in SparseRowMatrix.cpp +template void CpuMatrix::mul( + CpuSparseMatrix* a, + CpuMatrix* b, + SparseRowCpuMatrix* c, + real scaleAB, + real scaleT); +template void CpuMatrix::mul( + CpuSparseMatrix* a, + CpuMatrix* b, + SparseAutoGrowRowCpuMatrix* c, + real scaleAB, + real scaleT); +template void CpuMatrix::mul(CpuSparseMatrix* a, + CpuMatrix* b, + CacheRowCpuMatrix* c, + real scaleAB, + real scaleT); + +#ifndef PADDLE_MOBILE_INFERENCE +void SharedCpuMatrix::mul(CpuSparseMatrix* a, + CpuMatrix* b, + real scaleAB, + real scaleT) { + CHECK(!isTransposed()) << "Not supported"; + CHECK(!b->isTransposed()) << "Not supported"; + CHECK_EQ(scaleAB, 1) << "Not supported"; + CHECK_EQ(scaleT, 1) << "Not supported"; + CHECK_EQ(a->getFormat(), SPARSE_CSR) << "not supported"; + + real* B = b->getData(); + real* C = getData(); + size_t height = getHeight(); + size_t width = getWidth(); + + // get real trans + MatrixPtr aTrans; + if (a->isTransposed()) { + aTrans = a->getTmpSparseMatrix(a->getWidth(), a->getHeight()); + a->transpose(aTrans, false); + } + a = dynamic_cast(aTrans.get()); + + size_t m = a->getWidth(); + CHECK_EQ(b->getHeight(), m); + CHECK_EQ(a->getHeight(), height); + CHECK_EQ(b->getWidth(), width); + + size_t blockSize = (height / blockNum_) + 1; + CpuMatrixPtr localBuf = *localBuf_; + if (!localBuf) { + localBuf = std::make_shared(blockSize, width); + } else { + localBuf->resize(blockSize, width); + } + localBuf->zeroMem(); + real* localC = localBuf->getData(); + std::vector& blockSeq = *blockSeq_; + if (blockSeq.size() == 0) { + for (int k = 0; k < blockNum_; ++k) { + blockSeq.push_back(k); + } + std::shuffle( + blockSeq.begin(), blockSeq.end(), ThreadLocalRandomEngine::get()); + } + std::vector& localBufRows = *localBufRows_; + int* cols = a->getCols(); + real* value = a->getValue(); + + for (int k = 0; k < blockNum_; ++k) { + int blockId = blockSeq[k]; + size_t blockBegin = blockId * blockSize; + size_t blockEnd = (blockId + 1) * blockSize; + if (blockId == blockNum_ - 1) { + blockEnd = height; + } + if (a->getValueType() == NO_VALUE) { + for (size_t i = blockBegin; i < blockEnd; ++i) { + int start = a->getRowStartIdx(i); + int end = a->getRowStartIdx(i); + size_t colNum = a->getColNum(i); + if (colNum == 0) { + continue; + } // skip empty row + localBufRows.push_back(i); + size_t bufPos = localBufRows.size() - 1; + for (int j = start; j < end; ++j) { + vecAddTo(localC + bufPos * width, B + cols[j] * width, width); + } + } + } else if (a->getValueType() == FLOAT_VALUE) { + for (size_t i = blockBegin; i < blockEnd; ++i) { + int start = a->getRowStartIdx(i); + int end = a->getRowStartIdx(i); + size_t colNum = a->getColNum(i); + if (colNum == 0) { + continue; + } // skip empty row + localBufRows.push_back(i); + size_t bufPos = localBufRows.size() - 1; + for (int j = start; j < end; ++j) { + vecAddTo( + localC + bufPos * width, B + cols[j] * width, value[j], width); + } + } + } + + { + std::lock_guard guard(*blockLocks_[blockId]); + for (size_t i = 0; i < localBufRows.size(); ++i) { + vecAddTo(C + localBufRows[i] * width, localC + i * width, width); + } + } + memset(localC, 0, localBufRows.size() * width * sizeof(real)); + localBufRows.clear(); + } + + VLOG(2) << " B[0]=" << B[0] << " B[1]=" << B[1] << " C[0]=" << C[0] + << " C[1]=" << C[1]; +} + +void SharedCpuMatrix::add(Matrix& b, real p1, real p2) { + CHECK_EQ(blockNum_, 1); + std::lock_guard guard(*blockLocks_[0]); + CpuMatrix::add(b, p1, p2); +} + +void SharedCpuMatrix::add(real p1, real p2) { + CHECK_EQ(blockNum_, 1); + std::lock_guard guard(*blockLocks_[0]); + CpuMatrix::add(p1, p2); +} + +void SharedCpuMatrix::initShared(int blockNum) { + CHECK_GT(height_ * width_, 1UL * 1024 * 1024) + << "should not share small matrix"; + initBlock(blockNum); +} + +void SharedCpuMatrix::initBlock(int blockNum) { + CHECK_LE(blockNum, 200) << "should not use large block number"; + blockNum_ = blockNum; + blockLocks_.resize(blockNum); + for (auto& locker : blockLocks_) { + locker.reset(new std::mutex); + } +} + +#endif +/* Add a (column) vector b to matrix a, column by column */ +void CpuMatrix::addColumnVector(const Matrix& b) { + BaseMatrix::addColVector(const_cast(b)); +} + +/* this = a*b */ +void CpuMatrix::mul(const Matrix& a, const Matrix& b) { + return mul(a, b, 1.0, 0.0); +} + +/* this = scaleAB*(this*b) + scaleT*this */ +void CpuMatrix::rightMul(Matrix& b, real scaleAB, real scaleT) { + (void)b; + (void)scaleAB; + (void)scaleT; + LOG(FATAL) << "Not implemented"; +} + +/* this = this* b */ +void CpuMatrix::rightMul(Matrix& b) { return rightMul(b, 1.0, 0.0); } + +/* this = scaleAB*(a*this) + scaleT*this */ +void CpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) { + (void)a; + (void)scaleAB; + (void)scaleT; + LOG(FATAL) << "Not implemented"; +} + +/* this = a*this) */ +void CpuMatrix::leftMul(Matrix& a) { return leftMul(a, 1.0, 0.0); } + +void CpuMatrix::colMerge(Matrix& src) { src.rowSum(*this); } + +void CpuMatrix::rowSum(Matrix& sum) { + CHECK_EQ(sum.getHeight(), getHeight()); + CHECK_EQ(sum.getWidth(), (size_t)1); + + sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0); +} + +void CpuMatrix::rowMaxId(IVector& maxIds) { + CHECK(!maxIds.useGpu()) << "Matrix type are not equal"; + + size_t numSamples = getHeight(); + CHECK_EQ(maxIds.getSize(), numSamples); + + real* a = getData(); + int* s = maxIds.getData(); + size_t dim = getWidth(); + + for (size_t i = 0; i < numSamples; i++) { + real sm = a[i * dim]; + int maxId = 0; + for (size_t j = 1; j < dim; j++) { + if (a[i * dim + j] > sm) { + maxId = j; + sm = a[i * dim + j]; + } + } + s[i] = maxId; + } +} + +void CpuMatrix::rowMax(Matrix& max) { + CHECK_EQ(max.getHeight(), getHeight()); + CHECK_EQ(max.getWidth(), (size_t)1); + max.maxRows(*this); +} + +/* Get the top k elements of each row of this matrix */ +void CpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { + CHECK(isContiguous()); + CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal"; + size_t numSamples = getHeight(); + size_t beam = maxVal.getWidth(); + CHECK_EQ(maxIds.getSize(), numSamples * beam); + CHECK_EQ(maxVal.getHeight(), numSamples); + CHECK_EQ(maxVal.getWidth(), beam); + + real* a = getData(); + int* s = maxIds.getData(); + real* t = maxVal.getData(); + size_t dim = getWidth(); + for (size_t i = 0; i < numSamples; i++) { + std::vector> vec; + for (size_t j = 0; j < dim; j++) { + vec.push_back(std::pair(a[i * dim + j], j)); + } + + std::partial_sort( + vec.begin(), + vec.begin() + beam, + vec.end(), + [](const std::pair& l, const std::pair& r) { + return l.first > r.first; + }); + for (size_t j = 0; j < beam; j++) { + t[i * beam + j] = vec[j].first; + s[i * beam + j] = vec[j].second; + } + } +} + +void CpuMatrix::colMax(Matrix& max) { + CHECK_EQ(max.getWidth(), getWidth()); + CHECK_EQ(max.getHeight(), (size_t)1); + max.maxCols(*this); +} + +void CpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) { + CHECK(isContiguous()); + CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal"; + size_t numSamples = getWidth(); + size_t beam = maxVal.getHeight(); + CHECK_EQ(maxIds.getSize(), numSamples * beam); + CHECK_EQ(maxVal.getWidth(), numSamples); + + real* a = getData(); + int* s = maxIds.getData(); + real* t = maxVal.getData(); + size_t dim = getHeight(); + for (size_t i = 0; i < numSamples; i++) { + std::vector> vec; + for (size_t j = 0; j < dim; j++) { + vec.push_back(std::pair(a[i + j * numSamples], j)); + } + + std::partial_sort( + vec.begin(), + vec.begin() + beam, + vec.end(), + [](const std::pair& l, const std::pair& r) { + return l.first > r.first; + }); + for (size_t j = 0; j < beam; j++) { + t[i + j * numSamples] = vec[j].first; + s[i + j * numSamples] = vec[j].second; + } + } +} + +void CpuMatrix::maxoutForward(Matrix& a, + IVector& id, + size_t channels, + size_t groups) { + CHECK(dynamic_cast(&a)); + CHECK(dynamic_cast(&id)); + CHECK_EQ(a.getHeight(), getHeight()); + + size_t size = getWidth(); + size_t batchSize = getHeight(); + size_t featLen = size / channels; + const real* input = a.getData(); + int* idForCpu = id.getData(); + + MatrixPtr maxInMat, maxOutMat; + Matrix::resizeOrCreate(maxInMat, groups, size, false, false); + Matrix::resizeOrCreate(maxOutMat, 1, size, false, false); + + for (size_t batch_idx = 0; batch_idx < batchSize; ++batch_idx) { + size_t newIndex = batch_idx * size; + IVectorPtr tmpId = IVector::create(idForCpu + newIndex, size, false); + + for (size_t i = 0; i < channels; ++i) { + size_t newFeatLen = i * featLen; + for (size_t j = 0; j < groups; ++j) { + maxInMat->subMatrix(j, j + 1, newFeatLen, newFeatLen + featLen) + ->copyFrom(input + (newIndex + newFeatLen) * groups + j * featLen, + featLen); + } + } + maxInMat->colMax(*tmpId, *maxOutMat); + this->subRowMatrix(batch_idx, batch_idx + 1)->copyFrom(*maxOutMat); + } +} + +void CpuMatrix::maxoutBackward(Matrix& a, + IVector& id, + size_t channels, + size_t groups) { + CHECK(dynamic_cast(&a)); + CHECK(dynamic_cast(&id)); + CHECK_EQ(a.getHeight(), getHeight()); + + size_t size = a.getWidth(); + size_t batchSize = getHeight(); + size_t featLen = size / channels; + size_t newFeatLen = groups * featLen; + real* inputG = getData(); + const real* outG = a.getData(); + int* idForCpu = id.getData(); + + for (size_t batch_idx = 0; batch_idx < batchSize; ++batch_idx) { + size_t newIndex = batch_idx * size; + int* idData = idForCpu + newIndex; + + for (size_t i = 0; i < size; ++i) { + int gradIdx = + idData[i] * featLen + (i / featLen) * newFeatLen + i % featLen; + (inputG + newIndex * groups)[gradIdx] += (outG + newIndex)[i]; + } + } +} + +void CpuMatrix::rowNormalizeL1(Matrix& out) { + CHECK(!out.useGpu()); + + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(out.getHeight(), numSamples); + CHECK_EQ(out.getWidth(), dim); + real* a = getData(); + real* b = out.getData(); + for (size_t i = 0; i < numSamples; ++i) { + real s = 0; + for (size_t j = 0; j < dim; ++j) { + s += a[i * dim + j]; + } + // Right now, we just bet that sum won't be zero. If this really happens, + // we will figure out what should be done then. + CHECK_GT(s, 0); + s = 1 / s; + for (size_t j = 0; j < dim; ++j) { + b[i * dim + j] = s * a[i * dim + j]; + } + } +} + +/* calulate classification error */ +void CpuMatrix::classificationError(Matrix& output, + IVector& label, + size_t topkSize) { + size_t numSamples = this->getHeight(); + auto cpuOutput = dynamic_cast(&output); + auto cpuLabel = dynamic_cast(&label); + IVectorPtr cpuTopIds = std::make_shared(numSamples * topkSize); + MatrixPtr cpuTopVal = std::make_shared(numSamples, topkSize); + + CHECK(cpuOutput && cpuLabel) << "Invalid argument pointer"; + CHECK(cpuTopIds && cpuTopVal) << "Allocate cpu memory failed"; + CHECK(cpuLabel->getSize() == numSamples) << "Vector size is not equal"; + CHECK(cpuOutput->getHeight() == numSamples && this->getWidth() == 1) + << "Matrix dimensions are not equal"; + + // top k matrix classification + cpuOutput->rowMax(*cpuTopIds, *cpuTopVal); + + size_t dim = cpuOutput->getWidth(); + real* result = this->getData(); + int* ids = cpuTopIds->getData(); + int* lbl = cpuLabel->getData(); + for (size_t i = 0; i < numSamples; ++i) { + CHECK_GE(lbl[i], 0); + CHECK_LT((size_t)lbl[i], dim); + + for (size_t j = 0; j < topkSize; ++j) { + if (ids[j + i * topkSize] == lbl[i]) { + result[i] = 0; + break; + } + result[i] = 1.0f; + } + } +} + +/* copy -log(output[label]) to this->data[i] */ +void CpuMatrix::oneHotCrossEntropy(Matrix& output, IVector& label) { + CHECK(dynamic_cast(&output)); + CHECK(dynamic_cast(&label)); + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(label.getSize(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(getWidth(), (size_t)1); + + real* out = output.getData(); + real* cost = getData(); + int* lbl = label.getData(); + for (size_t i = 0; i < numSamples; ++i, out += dim) { + CHECK_GE(lbl[i], 0); + CHECK_LT((size_t)lbl[i], dim); + cost[i] = -std::log(out[lbl[i]]); + } +} + +/* calculate the error of outputV according to label */ +void CpuMatrix::oneHotCrossEntropyBp(Matrix& output, IVector& label) { + CHECK(dynamic_cast(&output)); + CHECK(dynamic_cast(&label)); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(output.getWidth(), dim); + real* out = output.getData(); + real* grad = getData(); + int* lbl = label.getData(); + for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) { + grad[lbl[i]] -= 1 / out[lbl[i]]; + } +} + +/* + We implement the matrix functionality in CostLayer.cpp, + but we define the scalar function here for sanity check + deletion of the function does not affect anything neverthelss +*/ +void CpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output, + IVector& label, + real alpha) { + CHECK(dynamic_cast(&output)); + CHECK(dynamic_cast(&label)); + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(label.getSize(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(getWidth(), (size_t)1); + + real* out = output.getData(); + real* cost = getData(); + int* lbl = label.getData(); + for (size_t i = 0; i < numSamples; ++i, out += dim) { + CHECK_GE(lbl[i], 0); + CHECK_LT((size_t)lbl[i], dim); + real sum = 0; + for (size_t j = 0; j < dim; ++j) { + sum += out[j]; + } + sum = _safelog(sum); + cost[i] = -_safelog(out[lbl[i]]) + sum + alpha * _square(sum); + } +} + +/* + We implement the matrix functionality in CostLayer.cpp, + but we define the scalar function here for sanity check + deletion of the function does not affect anything neverthelss +*/ +void CpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& output, + IVector& label, + real alpha) { + CHECK(dynamic_cast(&output)); + CHECK(dynamic_cast(&label)); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(output.getWidth(), dim); + real* out = output.getData(); + real* grad = getData(); + int* lbl = label.getData(); + + for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) { + grad[lbl[i]] -= 1 / out[lbl[i]]; + real sum = 0; + for (size_t j = 0; j < dim; ++j) { + sum += out[j]; + } + for (size_t j = 0; j < dim; ++j) { + if (j == (size_t)lbl[i]) { + grad[j] += -1 / out[j]; + } + grad[j] += 1 / sum + 2 * alpha * _safelog(sum) / sum; + } + } +} + +#define FORWARD_LOOP() \ + size_t numSamples = getHeight(); \ + size_t dim = getWidth(); \ + CHECK_EQ(output.getHeight(), numSamples); \ + CHECK_EQ(output.getWidth(), dim); \ + const real* in = getData(); \ + real* out = output.getData(); \ + for (size_t i = 0; i < numSamples; ++i, in += dim, out += dim) + +#define BACKWARD_LOOP() \ + size_t numSamples = getHeight(); \ + size_t dim = getWidth(); \ + CHECK_EQ(output.getHeight(), numSamples); \ + CHECK_EQ(output.getWidth(), dim); \ + real* grad = getData(); \ + real* out = output.getData(); \ + for (size_t i = 0; i < numSamples; ++i, grad += dim, out += dim) + +void CpuMatrix::softmax(Matrix& output) { + CHECK(!output.useGpu()); + + const float THRESHOLD = -64.0; + + FORWARD_LOOP() { + real max = -1.0e20; + for (size_t j = 0; j < dim; ++j) { + if (in[j] > max) { + max = in[j]; + } + } + for (size_t j = 0; j < dim; ++j) { + real a = in[j] - max; + if (a < THRESHOLD) { + a = THRESHOLD; + } + out[j] = a; + } + vExp(dim, out, out); + + real sum = 0; + for (size_t j = 0; j < dim; ++j) { + sum += out[j]; + } + sum = 1 / sum; + for (size_t j = 0; j < dim; ++j) { + out[j] *= sum; + } + } +} + +void CpuMatrix::sequenceSoftmax(Matrix& output, const IVector& index) { + CHECK_EQ(getWidth(), 1UL); + CHECK_EQ(output.getWidth(), 1UL); + CHECK(isContiguous()); + + MatrixPtr inTmp = Matrix::create(nullptr, + /* height= */ 1, + 1, + /* trans= */ false, + false); + MatrixPtr outTmp = Matrix::create(nullptr, + /* height= */ 1, + 1, + /* trans= */ false, + false); + size_t numSequences = index.getSize() - 1; + auto starts = index.getData(); + for (size_t i = 0; i < numSequences; ++i) { + size_t offset = starts[i]; + size_t size = starts[i + 1] - starts[i]; + inTmp->setData(getData() + offset, 1UL, size); + outTmp->setData(output.getData() + offset, 1UL, size); + inTmp->softmax(*outTmp); + } +} + +void CpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) { + CHECK(output.useGpu_ == false) << "Matrix type are not equal"; + CHECK_EQ(getHeight(), sftmaxSum.getHeight()); + + real* sums = sftmaxSum.getData(); + + BACKWARD_LOOP() { + real sum = sums[i]; + for (size_t j = 0; j < dim; ++j) { + grad[j] = out[j] * (grad[j] - sum); + } + } +} + +void CpuMatrix::sumOfSquares(Matrix& output, Matrix& label) { + CHECK(output.useGpu_ == false && label.useGpu_ == false) + << "Matrix type are not equal"; + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(label.getHeight(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(label.getWidth(), dim); + CHECK_EQ(getWidth(), (size_t)1); + real* out = output.getData(); + real* cost = getData(); + + auto labelptr = dynamic_cast(&label); + if (labelptr) { + // it is a CpuSparseMatrix + if (labelptr->getFormat() == SPARSE_CSR) { + // treat label as a SparseMatrix + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + cost[i] += _square(out[i * dim + j]); + } + } + if (labelptr->getValueType() == NO_VALUE) { + int* cols = labelptr->getCols(); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = labelptr->getRowStartIdx(i); + j < labelptr->getRowStartIdx(i + 1); + ++j) { + cost[i] += 1.0 - 2.0 * out[i * dim + cols[j]]; + /* + * explanation of above line: original codes are follows: + * cost[i] -= _square(out[i * dim + feature.col]); + * cost[i] += _square(1.0 - out[i * dim + feature.col]); + */ + } + } + } else if (labelptr->getValueType() == FLOAT_VALUE) { + int* cols = labelptr->getCols(); + real* values = labelptr->getValue(); + for (size_t i = 0; i < numSamples; ++i) { + real sum1 = 0; + real sum2 = 0; + for (size_t j = labelptr->getRowStartIdx(i); + j < labelptr->getRowStartIdx(i + 1); + ++j) { + sum1 += values[j] * values[j]; + sum2 += values[j] * out[i * dim + cols[j]]; + /* + * explanation of above line: original codes are follows: + * cost[i] -= _square(out[i * dim + feature.col]); + * cost[i] += _square(value.col - out[i * dim + feature.col]); + */ + } + cost[i] += sum1 - 2.0 * sum2; + } + } else { + LOG(FATAL) << "unsupported sparse matrix value type in sumOfSquares"; + return; + } + return; + } else { + LOG(FATAL) << "unsupported sparse matrix format in sumOfSquares"; + return; + } + } + + BaseMatrix::sumOfSquaredDiffs(output, + label, + /* scaleSum= */ 1, + /* scaleDest= */ 1); +} + +/* calculate the error of outputV according to label */ +void CpuMatrix::sumOfSquaresBp(Matrix& output, Matrix& label) { + CHECK(output.useGpu_ == false && label.useGpu_ == false) + << "Matrix type are not equal"; + + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(output.getWidth(), dim); + CHECK_EQ(label.getWidth(), dim); + + real* out = output.getData(); + real* grad = getData(); + + auto labelptr = dynamic_cast(&label); + if (labelptr) { + // it is a CpuSparseMatrix + if (labelptr->getFormat() == SPARSE_CSR) { + // treat label as a SparseMatrix + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + grad[i * dim + j] += 2.0 * out[i * dim + j]; + } + } + if (labelptr->getValueType() == NO_VALUE) { + int* cols = labelptr->getCols(); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = labelptr->getRowStartIdx(i); + j < labelptr->getRowStartIdx(i + 1); + ++j) { + grad[i * dim + cols[j]] -= 2.0; + /* + * explanation of above line: original codes are follows: + * grad[i * dim + feature.col] -= 2.0 * out[i * dim + feature.col]; + * grad[i * dim + feature.col] += 2.0 * (out[i * dim + feature.col] + * - 1); + */ + } + } + } else if (labelptr->getValueType() == FLOAT_VALUE) { + int* cols = labelptr->getCols(); + real* values = labelptr->getValue(); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = labelptr->getRowStartIdx(i); + j < labelptr->getRowStartIdx(i + 1); + ++j) { + grad[i * dim + cols[j]] -= 2.0 * values[j]; + /* + * explanation of above line: original codes are follows: + * grad[i * dim + feature.col] -= 2.0 * out[i * dim + feature.col]; + * grad[i * dim + feature.col] += 2.0 * (out[i * dim + feature.col] + * - value.col); + */ + } + } + } else { + LOG(FATAL) << "unsupported sparse matrix value type in sumOfSquares"; + return; + } + return; + } else { + LOG(FATAL) << "unsupported sparse matrix format in sumOfSquares"; + return; + } + } + + real* lbl = label.getData(); + size_t ld = getStride(); + size_t outLd = output.getStride(); + size_t lblLd = label.getStride(); + CHECK(lbl); + for (size_t i = 0; i < numSamples; + ++i, out += outLd, lbl += lblLd, grad += ld) { + for (size_t j = 0; j < dim; ++j) { + grad[j] += 2.0 * (out[j] - lbl[j]); // positive gradient; + } + } +} + +void CpuMatrix::smoothL1(Matrix& output, Matrix& label, real destScale) { + CHECK(output.useGpu_ == false && label.useGpu_ == false) + << "Matrix type are not equal"; + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(label.getHeight(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(label.getWidth(), dim); + CHECK_EQ(getWidth(), (size_t)1); + + real* cost = getData(); + real* out = output.getData(); + real* lbl = label.getData(); + + for (size_t i = 0; i < numSamples; ++i, out += dim, lbl += dim) { + for (size_t j = 0; j < dim; ++j) { + real absVal = std::fabs(out[j] - lbl[j]); + cost[i] *= destScale; + if (absVal < 1.0) + cost[i] += 0.5 * absVal * absVal; + else + cost[i] += absVal - 0.5; + } + } +} + +void CpuMatrix::smoothL1Bp(Matrix& output, Matrix& label, real destScale) { + CHECK(output.useGpu_ == false && label.useGpu_ == false) + << "Matrix type are not equal"; + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(label.getHeight(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(label.getWidth(), dim); + CHECK_EQ(getWidth(), dim); + + real* out = output.getData(); + real* lbl = label.getData(); + real* grad = getData(); + + for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim, lbl += dim) { + for (size_t j = 0; j < dim; ++j) { + real val = out[j] - lbl[j]; + grad[j] *= destScale; + if (std::fabs(val) < 1) { + grad[j] += val; + } else { + grad[j] += (real(0) < val) - (val < real(0)); + } + } + } +} + +void CpuMatrix::tanh(Matrix& output) { + CHECK(isContiguous()); + CHECK(output.isContiguous()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(output.getWidth(), dim); + vTanh(numSamples * dim, getData(), output.getData()); +} + +void CpuMatrix::tanhDerivative(Matrix& output) { + BaseMatrix::tanhDerivative(output); +} + +void CpuMatrix::softrelu(Matrix& output) { + CHECK(isContiguous()); + CHECK(output.isContiguous()); + const real THRESHOLD = 40.0; + FORWARD_LOOP() { // TODO(yuyang18): SIMD it? + for (size_t j = 0; j < dim; ++j) { + real x = in[j]; + if (x > THRESHOLD) { + x = THRESHOLD; + } else if (x < -THRESHOLD) { + x = -THRESHOLD; + } + out[j] = x; + } + } + vExp(numSamples * dim, output.getData(), output.getData()); + vLog1p(numSamples * dim, output.getData(), output.getData()); +} + +void CpuMatrix::softreluDerivative(Matrix& output) { + CHECK(isContiguous()); + CHECK(output.isContiguous()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + size_t size = numSamples * dim; + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(output.getWidth(), dim); + real* grad = getData(); + MatrixPtr tmpMat = Matrix::create(numSamples, dim); + real* tmp = tmpMat->getData(); + + vExp(size, output.getData(), tmpMat->getData()); + + for (size_t i = 0; i < size; ++i) { + grad[i] *= (1.0 - 1.0 / tmp[i]); + } +} + +void CpuMatrix::scaledTanh(Matrix& output, real p1, real p2) { + CHECK(isContiguous()); + CHECK(output.isContiguous()); + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(output.getWidth(), dim); + + const real* in = getData(); + real* out = output.getData(); + + // out = p2*in + for (size_t i = 0; i < numSamples * dim; ++i) { + out[i] = p2 * in[i]; + } + + vTanh(numSamples * dim, out, out); + + // out = p1 * out + for (size_t i = 0; i < numSamples * dim; ++i) { + out[i] = p1 * out[i]; + } +} + +/* uniform randomization, minimize precision = 1e-5 */ +void CpuMatrix::randomizeUniform() { + CHECK(isContiguous()); + real* data = getData(); + unsigned int* randSeed = ThreadLocalRand::getSeed(); + real recipRandMax = 1.0f / (real)RAND_MAX; + for (size_t i = 0; i < elementCnt_; ++i) { + *data++ = rand_r(randSeed) * recipRandMax; + } +} + +void CpuMatrix::print(std::ostream& os) const { + CHECK(isContiguous()); + for (size_t i = 0; i < height_; ++i) { + for (size_t j = 0; j < width_; ++j) { + os << data_[i * width_ + j] << " "; + } + os << std::endl; + } +} + +void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) { + real* input = data.getData(); + real* w = W.getData(); + real* output = data_; + size_t numElements = data.getWidth(); + size_t numSamples = data.getHeight(); + size_t paraSize = W.getHeight() * W.getWidth(); + CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + + size_t partial_sum = numElements / paraSize; + if (paraSize == numElements) { + for (size_t n = 0; n < numSamples * numElements; ++n) { + output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements]; + } + return; + } + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + for (size_t n = 0; n < numSamples; ++n) { + for (size_t i = 0; i < paraSize; i++) { + neon::prelu( + input + i * partial_sum, w[i], output + i * partial_sum, partial_sum); + } + input = input + numElements; + output = output + numElements; + } +#else + for (size_t n = 0, k = 0; n < numSamples; ++n) { + for (size_t i = 0; i < numElements; ++i, ++k) { + output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum]; + } + } +#endif +} + +void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) { + real* ograd = oGrad.getData(); + real* input = data.getData(); + real* wgrad = data_; + size_t numElements = data.getWidth(); + size_t numSamples = data.getHeight(); + size_t paraSize = this->getHeight() * this->getWidth(); + CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + size_t partial_sum = numElements / paraSize; + for (size_t n = 0, k = 0; n < numSamples; ++n) { + for (size_t i = 0; i < numElements; ++i, ++k) { + wgrad[i / partial_sum] += ograd[k] * (input[k] > 0 ? 0 : input[k]); + } + } +} + +void CpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) { + real* diff = data_; + real* input = data.getData(); + real* ograd = oGrad.getData(); + real* w = W.getData(); + size_t numElements = data.getWidth(); + size_t numSamples = data.getHeight(); + size_t paraSize = W.getHeight() * W.getWidth(); + CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + size_t partial_sum = numElements / paraSize; + for (size_t n = 0, k = 0; n < numSamples; ++n) { + for (size_t i = 0; i < numElements; ++i, ++k) { + diff[k] += ograd[k] * (input[k] > 0 ? 1 : w[i / partial_sum]); + } + } +} + +void CpuMatrix::print(std::ostream& os, size_t height, size_t width) const { + CHECK(isContiguous()); + size_t h = height_ < height ? height_ : height; + size_t w = width_ < width ? width_ : width; + os.setf(std::ostream::scientific); + os << "["; + for (size_t i = 0; i < h; ++i) { + for (size_t j = 0; j < w; ++j) { + os << data_[i * width_ + j] << " "; + } + if (i == h - 1) { + os << "]"; + } + os << std::endl; + } +} + +void CpuMatrix::printOneRow(std::ostream& os, size_t idx) const { + CHECK_LT(idx, height_); + size_t offset = idx * stride_; + os << data_[offset]; + for (size_t i = 1; i < width_; ++i) { + os << " " << data_[offset + i]; + } + os << ";"; +} + +void CpuMatrix::check(std::ostream& os, Matrix& refMat, bool printDiff) { + CHECK(isContiguous()); + CHECK(height_ == refMat.getHeight()); + CHECK(width_ == refMat.getWidth()); + CpuMatrix cpuRef(height_, width_); + cpuRef.copyFrom(refMat); + size_t diffCnt = 0; + for (size_t i = 0; i < height_; ++i) { + for (size_t j = 0; j < width_; ++j) { + real a = getElement(i, j); + real b = cpuRef.getElement(i, j); + if (fabs(a - b) > 0.00001) { + ++diffCnt; + if (printDiff) { + os << "ref= " << a << " check= " << b << std::endl; + } + } + } + } + LOG(INFO) << "the diffCnt is " << diffCnt; +} + +real CpuMatrix::getMin() { + size_t size = getHeight() * getWidth(); + real* data = getData(); + real res = data[0]; + for (size_t i = 1; i < size; ++i) { + if (res > data[i]) { + res = data[i]; + } + } + return res; +} + +real CpuMatrix::getMax() { + size_t size = getHeight() * getWidth(); + real* data = getData(); + real res = data[0]; + for (size_t i = 1; i < size; ++i) { + if (res < data[i]) { + res = data[i]; + } + } + return res; +} + +void CpuMatrix::circularConv(Matrix& in0, Matrix& in1) { + size_t height = this->getHeight(); + size_t width0 = this->getWidth(); + size_t width1 = in1.getWidth(); + + CHECK_EQ(height, in0.getHeight()); + CHECK_EQ(width0, in0.getWidth()); + CHECK_EQ(height, in1.getHeight()); + + CHECK_EQ(width1 % 2, 1U); + + real* outV = this->getData(); + real* inV0 = in0.getData(); + real* inV1 = in1.getData(); + + int leftCtxLen = (width1 - 1) / 2; + for (size_t x = 0; x < height; + ++x, outV += width0, inV0 += width0, inV1 += width1) { + for (size_t i = 0; i < width0; ++i) { // each dimension of output + for (size_t j = 0; j < width1; ++j) { + // iterate over all dimentions of inV1 + int index = i + j - leftCtxLen; + index = (index + width0) % width0; + outV[i] += inV0[index] * inV1[j]; + } + } + } +} + +void CpuMatrix::circularConvDerivative( + Matrix& outG, Matrix& in0, Matrix& in1, Matrix& inG0, Matrix& inG1) { + size_t height = in0.getHeight(); + size_t width0 = in0.getWidth(); + size_t width1 = in1.getWidth(); + + CHECK_EQ(height, in1.getHeight()); + CHECK_EQ(height, inG0.getHeight()); + CHECK_EQ(width0, inG0.getWidth()); + CHECK_EQ(height, inG1.getHeight()); + CHECK_EQ(width1, inG1.getWidth()); + CHECK_EQ(height, outG.getHeight()); + CHECK_EQ(width0, outG.getWidth()); + + real* outGV = outG.getData(); + real* inV0 = in0.getData(); + real* inV1 = in1.getData(); + real* inGV0 = inG0.getData(); + real* inGV1 = inG1.getData(); + + int leftCtxLen = (width1 - 1) / 2; + for (size_t x = 0; x < height; ++x, + outGV += width0, + inV0 += width0, + inV1 += width1, + inGV0 += width0, + inGV1 += width1) { + for (size_t j = 0; j < width1; ++j) { // iterate over width1 + for (size_t i = 0; i < width0; ++i) { + // such over all dimensions of outG + int index = i + j - leftCtxLen; + index = (index + width0) % width0; + inGV0[index] += outGV[i] * inV1[j]; + inGV1[j] += outGV[i] * inV0[index]; + } + } + } +} + +void CpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) { + CHECK(dynamic_cast(&output)); + auto labelPtr = dynamic_cast(&label); + CHECK(labelPtr); + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(numSamples, output.getHeight()); + CHECK_EQ(numSamples, labelPtr->getHeight()); + CHECK_EQ(dim, labelPtr->getWidth()); + + real* out = output.getData(); + real* cost = getData(); + for (size_t i = 0; i < numSamples; ++i, out += dim) { + for (size_t j = 0; j < dim; ++j) { + CHECK(out[j] > 0 && out[j] < 1.0); + cost[i] -= std::log(1 - out[j]); + } + + const int* cols = labelPtr->getRowCols(i); + for (size_t j = 0; j < labelPtr->getColNum(i); ++j) { + CHECK_LT(size_t(cols[j]), dim); + cost[i] -= std::log(out[cols[j]] / (1 - out[cols[j]])); + } + } +} + +void CpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { + CHECK(dynamic_cast(&output)); + auto labelPtr = dynamic_cast(&label); + CHECK(labelPtr); + + size_t numSamples = getHeight(); + size_t dim = getWidth(); + CHECK_EQ(numSamples, output.getHeight()); + CHECK_EQ(numSamples, labelPtr->getHeight()); + CHECK_EQ(dim, output.getWidth()); + CHECK_EQ(dim, labelPtr->getWidth()); + + real* out = output.getData(); + real* grad = getData(); + for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) { + for (size_t j = 0; j < dim; ++j) { + CHECK(out[j] > 0 && out[j] < 1.0); + grad[j] += 1.0 / (1 - out[j]); + } + + const int* cols = labelPtr->getRowCols(i); + for (size_t j = 0; j < labelPtr->getColNum(i); ++j) { + CHECK_LT(size_t(cols[j]), dim); + grad[cols[j]] -= 1.0 / (out[cols[j]] * (1 - out[cols[j]])); + } + } +} + +/* calculate the classification error for multi binary label */ +void CpuMatrix::classificationErrorMulti(Matrix& output, + Matrix& label, + real threshold) { + CHECK(dynamic_cast(&output)); + auto labelPtr = dynamic_cast(&label); + CHECK(labelPtr); + + size_t numSamples = getHeight(); + size_t dim = output.getWidth(); + CHECK_EQ(numSamples, output.getHeight()); + CHECK_EQ(numSamples, labelPtr->getHeight()); + CHECK_EQ(dim, labelPtr->getWidth()); + + real* out = output.getData(); + real* result = getData(); + for (size_t i = 0; i < numSamples; ++i, out += dim) { + real sum = 0.0; + for (size_t j = 0; j < dim; ++j) { + if (out[j] >= threshold) { + sum += 1.0; + } + } + + const int* cols = labelPtr->getRowCols(i); + for (size_t j = 0; j < labelPtr->getColNum(i); ++j) { + CHECK_LT(size_t(cols[j]), dim); + if (out[cols[j]] < threshold) { + sum += 1.0; + } else { + sum -= 1.0; + } + } + result[i] = sum / dim; + } +} + +void CpuMatrix::bilinearForward(const Matrix& in, + const size_t inImgH, + const size_t inImgW, + const size_t outImgH, + const size_t outImgW, + const size_t numChannels, + const real ratioH, + const real ratioW) { + CHECK(dynamic_cast(&in)); + + size_t outputW = getWidth(); + size_t batchSize = getHeight(); + size_t inputW = in.getWidth(); + size_t inputH = in.getHeight(); + size_t inPosOffset = inImgH * inImgW; + size_t outPosOffset = outImgH * outImgW; + (void)(inputH); + + real* outData = getData(); + const real* inData = in.getData(); + + if (inImgH == outImgH && inImgW == outImgW) { + this->copyFrom(in); + } else { + for (size_t k = 0; k < batchSize; ++k) { // loop for batches + for (size_t i = 0; i < outImgH; ++i) { // loop for images + size_t h = ratioH * i; + size_t hid = (h < inImgH - 1) ? 1 : 0; + real h1lambda = ratioH * i - h; + real h2lambda = 1 - h1lambda; + + for (size_t j = 0; j < outImgW; ++j) { + size_t w = ratioW * j; + size_t wid = (w < inImgW - 1) ? 1 : 0; + real w1lambda = ratioW * j - w; + real w2lambda = 1 - w1lambda; + // calculate four position for bilinear interpolation + const real* inPos = &inData[k * inputW + h * inImgW + w]; + real* outPos = &outData[k * outputW + i * outImgW + j]; + for (size_t c = 0; c < numChannels; ++c) { // loop for channels + // bilinear interpolation + outPos[0] = + h2lambda * (w2lambda * inPos[0] + w1lambda * inPos[wid]) + + h1lambda * (w2lambda * inPos[hid * inImgW] + + w1lambda * inPos[hid * inImgW + wid]); + inPos += inPosOffset; + outPos += outPosOffset; + } + } + } + } + } +} + +void CpuMatrix::bilinearBackward(const Matrix& out, + const size_t outImgH, + const size_t outImgW, + const size_t inImgH, + const size_t inImgW, + const size_t numChannels, + const real ratioH, + const real ratioW) { + CHECK(dynamic_cast(&out)); + + size_t inputW = getWidth(); + size_t inputH = getHeight(); + size_t outputW = out.getWidth(); + size_t batchSize = out.getHeight(); + size_t inPosOffset = inImgH * inImgW; + size_t outPosOffset = outImgH * outImgW; + (void)(inputH); + + real* inGrad = getData(); + const real* outGrad = out.getData(); + + if (inImgH == outImgH && inImgW == outImgW) { + this->add(const_cast(out)); + } else { + for (size_t k = 0; k < batchSize; ++k) { // loop for batches + for (size_t i = 0; i < outImgH; ++i) { // loop for images + size_t h = ratioH * i; + size_t hid = (h < inImgH - 1) ? 1 : 0; + real h1lambda = ratioH * i - h; + real h2lambda = 1 - h1lambda; + for (size_t j = 0; j < outImgW; ++j) { + size_t w = ratioW * j; + size_t wid = (w < inImgW - 1) ? 1 : 0; + real w1lambda = ratioW * j - w; + real w2lambda = 1 - w1lambda; + + real* inPos = &inGrad[k * inputW + h * inImgW + w]; + const real* outPos = &outGrad[k * outputW + i * outImgW + j]; + for (size_t c = 0; c < numChannels; ++c) { // loop for channels + inPos[0] += h2lambda * w2lambda * outPos[0]; + inPos[wid] += h2lambda * w1lambda * outPos[0]; + inPos[hid * inImgW] += h1lambda * w2lambda * outPos[0]; + inPos[hid * inImgW + wid] += h1lambda * w1lambda * outPos[0]; + inPos += inPosOffset; + outPos += outPosOffset; + } + } + } + } + } +} + +void CpuMatrix::vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW) { + real* outData = getData(); + int outHeight = (height + 2 * paddingH - filterH) / strideH + 1; + int outWidth = (width + 2 * paddingW - filterW) / strideW + 1; + int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1; + + int channelsCol = channels * filterD * filterH * filterW; + for (int c = 0; c < channelsCol; ++c) { + int wOffset = c % filterW; + int hOffset = (c / filterW) % filterH; + int dOffset = (c / filterW / filterH) % filterD; + int cIn = c / filterW / filterH / filterD; + for (int d = 0; d < outDepth; ++d) { + for (int h = 0; h < outHeight; ++h) { + for (int w = 0; w < outWidth; ++w) { + int dPad = d * strideD - paddingD + dOffset; + int hPad = h * strideH - paddingH + hOffset; + int wPad = w * strideW - paddingW + wOffset; + + if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width && + dPad >= 0 && dPad < depth) + outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = + data[((cIn * depth + dPad) * height + hPad) * width + wPad]; + else + outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = 0; + } + } + } + } +} + +void CpuMatrix::col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta) { + real* src = getData(); + int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1; + int outHeight = (height + 2 * paddingH - filterH) / strideH + 1; + int outWidth = (width + 2 * paddingW - filterW) / strideW + 1; + int channelsCol = channels * filterD * filterH * filterW; + for (int c = 0; c < channelsCol; ++c) { + int wOffset = c % filterW; + int hOffset = (c / filterW) % filterH; + int dOffset = (c / filterW / filterH) % filterD; + int cIm = c / filterW / filterH / filterD; + for (int d = 0; d < outDepth; ++d) { + for (int h = 0; h < outHeight; ++h) { + for (int w = 0; w < outWidth; ++w) { + int dPad = d * strideD - paddingD + dOffset; + int hPad = h * strideH - paddingH + hOffset; + int wPad = w * strideW - paddingW + wOffset; + if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width && + dPad >= 0 && dPad < depth) + trg[((cIm * depth + dPad) * height + hPad) * width + wPad] = + alpha * + src[((c * outDepth + d) * outHeight + h) * outWidth + w] + + beta * + trg[((cIm * depth + dPad) * height + hPad) * width + wPad]; + } + } + } + } +} + +//////////////////////////////////////////////////////////////// +// functions executed via cpu // +//////////////////////////////////////////////////////////////// + +void GpuMatrix::selectElements(Matrix& table, IVector& ids) { + execViaCpu2(&CpuMatrix::selectElements, *this, table, ids); +} +} // namespace paddle diff --git a/paddle/legacy/math/Matrix.h b/paddle/legacy/math/Matrix.h new file mode 100644 index 0000000000000000000000000000000000000000..ff4f4cfc2a41add1a06308556b38aba5bbdac884 --- /dev/null +++ b/paddle/legacy/math/Matrix.h @@ -0,0 +1,2189 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +#include + +#include "BaseMatrix.h" +#include "MemoryHandle.h" +#include "Vector.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +/// TODO(tianbing), move to paddle/legacy/function/TensorType.h +enum SparseValueType { NO_VALUE = 0, FLOAT_VALUE = 1 }; + +/** + * @brief matrix sparse_format . + * + * nnz represents nonzero number in sparse matrix. + * + * SPARSE_CSR: row major matrix. length of row is height_ + 1, each element + * represents row start index in Matrix. length of col and value are nnz. + * + * SPARSE_CSC: col major matrix. length of col is width_ + 1, each element + * represents col start index in Matrix. length of col and value are nnz. + * + * @code + * for example: [0, 1, 0, 2, 0; + * 1, 0, 0, 0, 0; + * 0, 0, 0, 2, 5]; + * SPARSE_CSR row [0, 2, 3, 5]; + * col [1, 3, 0, 3, 4]; + * value [1, 2, 1, 2, 5] + * SPARSE_CSC col [0, 1, 2, 2, 4, 5]; + * row [1, 0, 0, 2, 2]; + * value [1, 1, 2, 2, 5] + * @endcode + */ +/// TODO(tianbing), move to paddle/legacy/function/TensorType.h +enum SparseFormat { SPARSE_CSR = 0, SPARSE_CSC = 1 }; + +class Matrix; +class GpuMatrix; +class CpuMatrix; +class CpuSparseMatrix; +class GpuSparseMatrix; +typedef std::shared_ptr MatrixPtr; +typedef std::shared_ptr GpuMatrixPtr; +typedef std::shared_ptr CpuMatrixPtr; +typedef std::shared_ptr GpuSparseMatrixPtr; +typedef std::shared_ptr CpuSparseMatrixPtr; + +/** + * Copy or assignemnt constructor will share the data as opposed to making a + * copy of the original data. To make a copy of the orinal data, use copyFrom() + * instead. + */ +class Matrix : public BaseMatrix { + protected: + Matrix(MemoryHandlePtr memHandle, + size_t height, + size_t width, + bool trans, + bool use_gpu); + + Matrix(real* data, size_t height, size_t width, bool trans, bool use_gpu); + + Matrix(real* data, + size_t height, + size_t width, + size_t stride, + bool trans, + bool use_gpu); + + static ThreadLocal tmpMat_; + + public: + size_t elementCnt_; // maximal number of elements which can be held in data_ + MemoryHandlePtr memoryHandle_; + + public: + virtual ~Matrix() {} + + static MatrixPtr create(MemoryHandlePtr memHandle, + size_t height, + size_t width, + bool trans = false); + static MatrixPtr create(size_t height, + size_t width, + bool trans = false, + bool useGpu = false); + static MatrixPtr create(real* data, + size_t height, + size_t width, + bool trans = false, + bool useGpu = false); + static MatrixPtr create(real* data, + size_t height, + size_t width, + size_t stride, + bool trans = false, + bool useGpu = false); + + static MatrixPtr createSparseMatrix(size_t height, + size_t width, + size_t nnz, + SparseValueType valueType = FLOAT_VALUE, + bool trans = false, + bool useGpu = false); + static MatrixPtr createSparseMatrix(size_t height, + size_t width, + size_t nnz, + SparseValueType valueType = FLOAT_VALUE, + SparseFormat foramt = SPARSE_CSR, + bool trans = false, + bool useGpu = false); + + static MatrixPtr createSparseMatrix(real* data, + int* row, + int* col, + size_t height, + size_t width, + size_t nnz, /* used to allocate space */ + SparseValueType valueType, /*value type*/ + SparseFormat format, + bool trans, + bool useGpu); + + static void resizeOrCreateSparseMatrix( + MatrixPtr& matrix, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType = FLOAT_VALUE, + SparseFormat foramt = SPARSE_CSR, + bool trans = false, + bool useGpu = false); + + static void resizeOrCreate(MatrixPtr& a, + size_t height, + size_t width, + bool trans = false, + bool useGpu = false); + + /** + * @brief set the data buffer used to hold the matrix data. + * + * caller should make sure that the size of data is at least + * sizeof(real)*height*width. + */ + void setData(real* data) { + BaseMatrix::setData(data); + memoryHandle_.reset(); + } + + /// the data should be contiguous + void setData(real* data, size_t newHeight, size_t newWidth) { + setData(data); + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newHeight * newWidth; + stride_ = width_; + } + + size_t getWidth() const { return width_; } + size_t getHeight() const { return height_; } + size_t getStride() const { return stride_; } + size_t getElementCnt() const { return elementCnt_; } + virtual real* getData() { return data_; } + virtual const real* getData() const { return data_; } + bool isTransposed() const { return trans_; } + bool isContiguous() const { return stride_ == width_ || height_ == 1; } + + // If sparse matrix, need to dynamic_cast to CpuSparseMatrix/GpuSparseMatrix + // befor call the following functions. + // Declare these functions in the base class just easy to call them. + // And these declarations should be moved to base class of sparse matrix + // if refactor sparse matrix + virtual int* getRows() const { + LOG(FATAL) << "Not implemented"; + return nullptr; //! suppress warning for no return value. + } + + virtual int* getCols() const { + LOG(FATAL) << "Not implemented"; + return nullptr; //! suppress warning for no return value. + } + + virtual SparseFormat getFormat() const { + LOG(FATAL) << "Not implemented"; + return SPARSE_CSR; //! suppress warning for no return value. + } + + virtual SparseValueType getValueType() const { + LOG(FATAL) << "Not implemented"; + return NO_VALUE; //! suppress warning for no return value. + } + + /** + * @brief matrix elment-wise add + * + * Named add3 just because add/add2 has been used in BaseMatrix.cu + * and they are not virtual function. + */ + virtual void add3(MatrixPtr b) { LOG(FATAL) << "Not implemented"; } + + MemoryHandlePtr getMemoryHandle() const { return memoryHandle_; } + + virtual void zeroMem() { LOG(FATAL) << "Not implemented"; } + + virtual void resetOne() { LOG(FATAL) << "Not implemented"; } + + void setDiag(real value); + + virtual void copyFrom(const Matrix& src) { LOG(FATAL) << "Not implemented"; } + + virtual void trimFrom(const CpuSparseMatrix& src) { + LOG(FATAL) << "Not implemented"; + } + + // For GpuMatrix this is an asynchronous copy interface + // For CpuMatrix this is an synchronous copy interface + virtual void copyFrom(const Matrix& src, hl_stream_t stream) { + LOG(FATAL) << "Not implemented"; + } + + MatrixPtr subMatrix(size_t startRow, + size_t endRow, + size_t startCol, + size_t endCol); + + MatrixPtr subRowMatrix(size_t startRow, size_t endRow) { + return subMatrix(startRow, endRow, 0, getWidth()); + } + + MatrixPtr subColMatrix(size_t startCol, size_t endCol) { + return subMatrix(0, getHeight(), startCol, endCol); + } + + virtual MatrixPtr subMatrix(size_t startRow, size_t numRows) { + CHECK_LE(startRow + numRows, getHeight()); + return Matrix::create(getData() + startRow * getWidth(), + numRows, + getWidth(), + trans_, + useGpu_); + } + virtual MatrixPtr subMatrix(size_t startRow, size_t numRows, MatrixPtr dest) { + CHECK_LE(startRow + numRows, getHeight()); + CHECK_EQ(useGpu_, dest->useGpu_); + dest->setData(this->rowBuf(startRow), numRows, getWidth()); + return dest; + } + + /** + * If this is GpuMatrix, src is assumed to be CPU memory + * + * If this is CpuMatrix, src is assumed to be CPU memory + */ + virtual void copyFrom(const real* src, size_t size) { + LOG(FATAL) << "Not implemented"; + } + + virtual void copyFrom(const real* src, const int64_t* seq) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @brief convert a int vector to a real matrix. + * + * (1) source and dest are both in CPU. + * + * (2) sizes are exactly match. + */ + virtual void copyFrom(const IVector& src) { + LOG(FATAL) << "copy data from int vector only available on CpuMatrix."; + } + + virtual void copyByRowIndex(Matrix& b, const IVector& rowIndex) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @brief Create a matrix with the same type (GpuMatrix, CpuMatrix, + * NonValueSparseMatrix, etc.) as this. + * + * If height and width is zero, the new matrix will have the same size + * as this, otherwise the new matrix will have the specified size. + * + */ + virtual MatrixPtr clone(size_t height = 0, + size_t width = 0, + bool useGpu = false) { + LOG(FATAL) << "Not implemented"; + return nullptr; + } + + virtual real* getRowBuf(size_t row) { + LOG(FATAL) << "Not implemented"; + return nullptr; + } + + virtual real getElement(size_t x, size_t y) const { + LOG(FATAL) << "Not implemented"; + return 0; + } + + virtual real getSum() { + LOG(FATAL) << "Not implemented"; + return 0; + } + + virtual void accumulateColSum(Matrix& src) { + LOG(FATAL) << "Not implemented"; + } + + virtual real getAbsSum() { + LOG(FATAL) << "Not implemented"; + return 0; + } + + /** + * @note Original data may not be preserved after resize(). + */ + virtual void resize(size_t newHeight, size_t newWidth) = 0; + + /** + * @note This should only be used for sparse matrix. + */ + virtual void resize(size_t newHeight, + size_t newWidth, + size_t newNnz, /* total item used to allocate space */ + SparseValueType valueType, + SparseFormat format) = 0; + + /** + * @brief This should only be used for sparse matrix. + * + * Currently must be called for each row in order. + * The matrix is not valid until setRow is called for the last row. + */ + virtual void setRow(size_t row, + size_t colNum, + const unsigned int* cols, + const real* values) = 0; + + virtual MatrixPtr getTranspose() = 0; + + /** + * @brief hard transpose. + * + * allocate matTrans' memory outside, then set memAlloc as false; + * else set as true. + */ + virtual void transpose(MatrixPtr& matTrans, bool memAlloc) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @brief rotate 90 degrees in clock-wise if clockWise=true; + * otherwise rotate in anti clock-wise + * clock-wise: + * \f[ + * y(j,i) = x(M-i-1,j) + * \f] + * anti clock-wise: + * \f[ + * y(j,i) = x(i, N-1-j) + * \f] + * where \f$x\f$ is (M x N) input, and \f$y\f$ is (N x M) output. + * + * allocate matRot' memory outside, then set memAlloc as false; + * else set as true. + */ + virtual void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) { + LOG(FATAL) << "Not implemented"; + } + + virtual MatrixPtr getInverse() { + LOG(FATAL) << "Not implemented"; + return nullptr; + } + + /** + * @brief inverse. + * + * if allocate matInv's memory outside, then set memAlloc as false; + * else set as true. + */ + virtual void inverse(MatrixPtr& matInv, bool memAlloc) { + LOG(FATAL) << "Not implemented"; + } + + public: + /// Only set all variables to 0 or NULL but not free them. + virtual void clear() { + height_ = 0; + width_ = 0; + data_ = NULL; + } + + void reshape(size_t height, size_t width); + + /// add b to each sample of this. + virtual void addBias(Matrix& b, real scale) { + LOG(FATAL) << "Not implemented"; + } + + virtual void addSharedBias(Matrix& b, real scale) { + LOG(FATAL) << "Not implemented"; + } + + void addBias(Matrix& b, real scale, bool sharedBias) { + if (!sharedBias) { + addBias(b, scale); + } else { + addSharedBias(b, scale); + } + } + + /// add each sample from a to this. + virtual void collectBias(Matrix& a, real scale) { + LOG(FATAL) << "Not implemented"; + } + + virtual void collectSharedBias(Matrix& a, real scale) { + LOG(FATAL) << "Not implemented"; + } + + void collectBias(Matrix& a, real scale, bool sharedBias) { + if (!sharedBias) { + collectBias(a, scale); + } else { + collectSharedBias(a, scale); + } + } + + virtual void sequenceAvgForward(Matrix& a, + const IVector& startsPos, + int mode) { + LOG(FATAL) << "Not implemented"; + } + + virtual void sequenceAvgBackward(Matrix& a, + const IVector& startsPos, + int mode) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * this = scaleAB*(a*b) + scaleT*this + * @endcode + */ + virtual void mul(const Matrix& a, + const Matrix& b, + real scaleAB, + real scaleT) { + LOG(FATAL) << "Not implemented"; + } + + /// Add a vector (column) b to matrix a, column by column. + virtual void addColumnVector(const Matrix& b) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * For j < codeLength: + * this(i, j) += vec(index(i, j), 0) + * where index(i, j) = ((codes(i) + numClasses) >> (j + 1)) - 1 + * @endcode + */ + virtual void addByBitCode(size_t numClasses, + const IVector& codes, + const Matrix& vec) { + (void)numClasses; + (void)codes; + (void)vec; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * For j < codeLength: + * vec(index(i, j), 0) += this(i, j) + * where index is same as the index for addByBitCode + * @endcode + */ + virtual void addByBitCodeBackward(size_t numClasses, + const IVector& codes, + Matrix& vec) { + (void)numClasses; + (void)codes; + (void)vec; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * For j < codeLength: + * this(i, j) += + * where index is same as the index for addByBitCode + * @endcode + */ + virtual void mulByBitCode(size_t numClasses, + const IVector& codes, + const Matrix& mat, + const Matrix& input) { + (void)numClasses; + (void)codes; + (void)mat; + (void)input; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * For j < codeLength: + * mat.row(index(i, j)) += this(i, j) * input.row(i) + * where index is same as the index for addByBitCode + * @endcode + */ + virtual void mulByBitCodeBackwardWeight(size_t numClasses, + const IVector& codes, + Matrix& mat, + const Matrix& input) { + (void)numClasses; + (void)codes; + (void)mat; + (void)input; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * For j < codeLength: + * input.row(i) += this(i, j) * mat.row(index(i, j)) + * where index is same as the index for addByBitCode + * @endcode + */ + virtual void mulByBitCodeBackwardError(size_t numClasses, + const IVector& codes, + const Matrix& mat, + Matrix& input) { + (void)numClasses; + (void)codes; + (void)mat; + (void)input; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * For j < codeLength + * sum(i, 0) = scaleSum * \sum_j bit(i, j) * this(i, j) + * where bit(i, j) = ((codes(i) + numClasses) & 2^j) ? 1 : 0 + * @endcode + */ + virtual void sumByBitCode(size_t numClasses, + IVector& codes, + Matrix& sum, + real scaleSum) { + (void)numClasses; + (void)codes; + (void)sum; + (void)scaleSum; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * For j < codeLength + * this(i, j) -= bit(i, j) + * where bit(i, j) is same as that for sumByBitCode + * @endcode + */ + virtual void subByBitCode(size_t numClasses_, IVector& codes) { + (void)numClasses_; + (void)codes; + LOG(FATAL) << "Not implemeted"; + } + + /** + * add the sum of each row of this to mat + */ + virtual void rowSum(Matrix& sum) { + (void)sum; + LOG(FATAL) << "Not implemeted"; + } + + /** + * set the max of each row of this to mat + */ + virtual void rowMax(Matrix& max) { + (void)max; + LOG(FATAL) << "Not implemeted"; + } + + /** + * set the max of each column of this to mat + */ + virtual void colMax(Matrix& max) { LOG(FATAL) << "not implemented"; } + + /** + * @brief Get the top k elements of each column of this matrix. + * + * The row ids and values of these elements are stored in + * maxIds and max respectively. where k is the size of maxIds. + * And note that the top k elements are not sorted. + */ + virtual void colMax(IVector& maxIds, Matrix& maxVal) { + LOG(FATAL) << "not implemented"; + } + + virtual void maxoutForward(Matrix& a, + IVector& id, + size_t channels, + size_t groups) { + LOG(FATAL) << "not implemented"; + } + + virtual void maxoutBackward(Matrix& a, + IVector& id, + size_t channels, + size_t groups) { + LOG(FATAL) << "not implemented"; + } + + virtual void rowMaxId(IVector& maxIds) { LOG(FATAL) << "Not implemented"; } + + /** + * @brief Get the top k elements of each row of this matrix. + * + * The column ids and values of these elements are stored in + * maxIds and max respectively. where k is the size of maxIds. + * And note that the top k elements are not sorted. + */ + virtual void rowMax(IVector& maxIds, Matrix& max) { + LOG(FATAL) << "Not implemented"; + } + + /// normalize each row so that the sum of each row is 1. + virtual void rowNormalizeL1(Matrix& out) { + (void)out; + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * this = a*b + * @endcode + */ + virtual void mul(const Matrix& a, const Matrix& b) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * this = scaleAB*(this*b) + scaleT*this + * @endcode + */ + virtual void rightMul(Matrix& b, real scaleAB, real scaleT) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * this = this* b + * @endcode + */ + virtual void rightMul(Matrix& b) { LOG(FATAL) << "Not implemented"; } + + /** + * @code + * this = scaleAB*(a*this) + scaleT*this + * @endcode + */ + virtual void leftMul(Matrix& a, real scaleAB, real scaleT) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * this = a*this) + * @endcode + */ + virtual void leftMul(Matrix& a) { LOG(FATAL) << "Not implemented"; } + + /// merge the element for each col. + virtual void colMerge(Matrix& src) { LOG(FATAL) << "Not implemented"; } + + /// copy -log(output[label]) to this->data[i]. + virtual void oneHotCrossEntropy(Matrix& output, IVector& label) { + LOG(FATAL) << "Not implemented"; + } + + /// calculate the error of outputV according to label. + virtual void oneHotCrossEntropyBp(Matrix& outputV, IVector& label) { + LOG(FATAL) << "Not implemented"; + } + + /// copy -log(output[label]) to this->data[i]. + virtual void oneHotCrossEntropyWithSelfNorm(Matrix& output, + IVector& label, + real alpha) { + LOG(FATAL) << "Not implemented"; + } + + /// calculate the error of outputV according to label. + virtual void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, + IVector& label, + real alpha) { + LOG(FATAL) << "Not implemented"; + } + + /** + * \f[ + * a[i] = \sum_{j=-(N-1)/2}^{(N-1)/2} b_{i+j} * c_{j} + * \f] + * + * b contains M elements, + * c contains N elements (N is odd), + * b's index arithmetic is computed modulo M, + * c's index arithmetic is computed modulo N. + */ + virtual void circularConv(Matrix& b, Matrix& c) { + LOG(FATAL) << "Not implemented"; + } + + virtual void circularConvDerivative(Matrix& output, + Matrix& prevOut1, + Matrix& prevOut2, + Matrix& prevGrad1, + Matrix& prevGrad2) { + LOG(FATAL) << "Not implemented"; + } + + /* output_ij = exp(this_{ij}) / (sum_j exp(this_ij)) */ + virtual void softmax(Matrix& output) { + (void)output; + LOG(FATAL) << "Not implemeted"; + } + virtual void sequenceSoftmax(Matrix& output, const IVector& index) { + (void)output; + LOG(FATAL) << "Not implemeted"; + } + + virtual void softmaxBackward(Matrix& outputV) { + (void)outputV; + LOG(FATAL) << "Not implemeted"; + } + + /* + sum_i = sum_j this_ij * output_ij + this_ij = output_ij* (this_ij - sum_i) + */ + virtual void softmaxDerivative(Matrix& output, Matrix& sftmaxSum) { + LOG(FATAL) << "Not implemented"; + } + + /// calculate the sum of squares diff cost. + virtual void sumOfSquares(Matrix& output, Matrix& label) { + LOG(FATAL) << "Not implemented"; + } + + /// gradient of sumOfSquares. + virtual void sumOfSquaresBp(Matrix& outputV, Matrix& label) { + LOG(FATAL) << "Not implemented"; + } + + virtual void smoothL1(Matrix& output, Matrix& label, real destScale) { + LOG(FATAL) << "Not implemented"; + } + + virtual void smoothL1Bp(Matrix& outputV, Matrix& label, real destScale) { + LOG(FATAL) << "Not implemented"; + } + + virtual void tanh(Matrix& output) { LOG(FATAL) << "Not implemented"; } + + virtual void tanhDerivative(Matrix& output) { + LOG(FATAL) << "Not implemented"; + } + + virtual void softrelu(Matrix& output) { LOG(FATAL) << "Not implemented"; } + + virtual void softreluDerivative(Matrix& output) { + LOG(FATAL) << "Not implemented"; + } + + virtual void scaledTanh(Matrix& output, real p1, real p2) { + LOG(FATAL) << "Not implemented"; + } + + /// print out the values of elements to os + virtual void print(std::ostream& os) const { + LOG(FATAL) << "Not implemented"; + } + + /** + * print a part of the matrix + * from the (top,left) value to the (height, width) value (not included) + */ + virtual void print(std::ostream& os, size_t height, size_t width) const { + LOG(FATAL) << "Not implemented"; + } + + /// print one row to os + virtual void printOneRow(std::ostream& os, size_t idx) const { + LOG(FATAL) << "Not implemented"; + } + + virtual void check(std::ostream& os, Matrix& refMat, bool printDiff = true) {} + + virtual real getMin() { + LOG(FATAL) << "Not implemented"; + return 0; + } + virtual real getMax() { + LOG(FATAL) << "Not implemented"; + return 0; + } + + virtual void randomizeUniform() { LOG(FATAL) << "Not implemented"; } + + /** + * @brief calulate the error of classification + * + * output[i] = 1 if row i is an error. + * + * output[i] = 0 if row i is correct. + * + */ + virtual void classificationError(Matrix& output, + IVector& label, + size_t topkSize = 1) { + LOG(FATAL) << "Not implemented"; + } + + virtual void upsampleForward(Matrix& input, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void upsampleBackward(Matrix& outputGrad, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW) { + LOG(FATAL) << "Not implemeted"; + } + + /** + * Pooling forward operation, pick out the largest element + * in the sizeX of value, if the maskMatP is not NULL, it will + * also caculate the location indices. + */ + virtual void maxPoolForward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + MatrixPtr maskMatP = NULL) { + LOG(FATAL) << "Not implemeted"; + } + + /// Pooling backward operation. + virtual void maxPoolBackward(Matrix& image, + size_t imgSizeH, + size_t imgSizeW, + Matrix& outGrad, + Matrix& outV, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + /// Pooling forward operation, caculate the average of sizeX elements. + virtual void avgPoolForward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + bool excludeMode = true) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void avgPoolBackward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW, + bool excludeMode = true) { + LOG(FATAL) << "Not implemeted"; + } + + /** + * Pooling 3D forward operation, pick out the largest element + * in the sizeX of value + */ + virtual void maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + LOG(FATAL) << "Not implemeted"; + } + + /** + * Input: one or more sequences. Each sequence contains some instances. + * + * Output: output size is the number of input sequences (NOT input + * instances). + * + * output[i] is set to max_input[i]. + */ + virtual void maxSequenceForward(Matrix& input, + const IVector& sequence, + IVector& index) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void maxSequenceBackward(Matrix& outputGrad, + const IVector& sequence, + IVector& index) { + LOG(FATAL) << "Not implemeted"; + } + + /** + * @code + * this.row[i] += table.row[ids[i]] + * if ids[i] == -1, it will be ignored + * @endcode + */ + virtual void selectRows(Matrix& table, IVector& ids) { + (void)table; + (void)ids; + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * this[i] = table[i, id[i]] + * @endcode + */ + virtual void selectElements(Matrix& table, IVector& ids) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * table.row[ids[i]] += this.row[i] + * if ids[i] == -1, it will be ignored + * @endcode + */ + virtual void addToRows(Matrix& table, IVector& ids) { + (void)table; + (void)ids; + LOG(FATAL) << "Not implemented"; + } + + /** + * @code + * table[i, id[i]] += this[i] + * @endcode + */ + virtual void addElements(Matrix& table, IVector& ids) { + LOG(FATAL) << "Not implemented"; + } + /** + * @brief cross entropy for multi binary labels + * + * @code + * this[i] = -sum(label[i][j]*log(output[i][j]) + * + (1-label[i][j])*log(1-output[i][j])) + * @endcode + */ + virtual void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @brief The gradient of cross entropy for multi binary labels on output + * + * @code + * this[i][j] = -label[i][j]/output[i][j] + * + (1-label[i][j])/(1-output[i][j]) + * @endcode + */ + virtual void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { + LOG(FATAL) << "Not implemented"; + } + + /** + * @brief Calculate the classification error for multi binary labels + * + * @code + * this[i] = sum((output[i][j] >= threshold && label[i][j] == 0) + * || (output[i][j] < threshold && label[i][j] == 1)) + * / output->getWidth() + * @endcode + */ + virtual void classificationErrorMulti(Matrix& output, + Matrix& label, + real threshold) { + LOG(FATAL) << "Not implemented"; + } + + virtual void paramReluForward(Matrix& data, Matrix& W) { + LOG(FATAL) << "Not implemented"; + } + virtual void paramReluBackwardW(Matrix& oGrad, Matrix& data) { + LOG(FATAL) << "Not implemented"; + } + virtual void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) { + LOG(FATAL) << "Not implemented"; + } + + virtual void vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void bilinearForward(const Matrix& in, + const size_t inImgH, + const size_t inImgW, + const size_t outImgH, + const size_t outImgW, + const size_t numChannels, + const real ratioH, + const real ratioW) { + LOG(FATAL) << "Not implemented"; + } + virtual void bilinearBackward(const Matrix& out, + const size_t outImgH, + const size_t outImgW, + const size_t inImgH, + const size_t inImgW, + const size_t numChannels, + const real ratioH, + const real ratioW) { + LOG(FATAL) << "Not implemented"; + } + + template + void operator=(const ExpressionType& expr) { + if (useGpu_) { + TensorGpuApply(*this, expr); + } else { + TensorCpuApply(*this, expr); + } + } + + bool isEmpty() const { return data_ == nullptr; } + + explicit operator bool() const { return !isEmpty(); } +}; + +inline std::ostream& operator<<(std::ostream& os, const Matrix& mat) { + mat.print(os); + return os; +} + +class GpuMatrix : public Matrix { + public: + GpuMatrix(); + + GpuMatrix(size_t height, size_t width, bool trans = false); + GpuMatrix(real* data, size_t height, size_t width, bool trans = false) + : Matrix(data, height, width, trans, true) {} + GpuMatrix(real* data, + size_t height, + size_t width, + size_t stride, + bool trans = false) + : Matrix(data, height, width, stride, trans, true) {} + GpuMatrix(GpuMemHandlePtr dataHandle, + size_t height, + size_t width, + bool trans = false) + : Matrix(dataHandle, height, width, trans, true) {} + ~GpuMatrix(); + + void zeroMem(); + void resetOne(); + void setDiag(real value); + + void resize(size_t newHeight, size_t newWidth); + void resize(size_t newHeight, + size_t newWidth, + size_t newNnz, /* used to allocate space */ + SparseValueType valueType, + SparseFormat format) { + LOG(FATAL) << "Only Support Sparse Matrix"; + } + void setRow(size_t row, + size_t colNum, + const unsigned int* cols, + const real* values) { + LOG(FATAL) << "Only Support Sparse Matrix"; + } + + /** + * Copy the data from cpu_memory buffer + */ + void copyFrom(const real* hostSrc, size_t size); + + void copyFrom(const real* hostSrc, const int64_t* seq); + + void copyFrom(const Matrix& src, hl_stream_t stream); + + void copyFrom(const Matrix& src); + + void copyFrom(const IVector& src); + + void copyByRowIndex(Matrix& b, const IVector& rowIndex); + + MatrixPtr clone(size_t height, size_t width, bool useGpu = false); + + real getElement(size_t x, size_t y) const; + + real* getRow(size_t row) { return BaseMatrix::rowBuf(row); } + virtual real* getRowBuf(size_t row) { return getRow(row); } + + real getSum(); + void accumulateColSum(Matrix& src); + real getAbsSum(); + + real getMin(); + real getMax(); + + MatrixPtr getTranspose(); + void transpose(MatrixPtr& matTrans, bool memAlloc); + void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise); + + MatrixPtr getInverse(); + void inverse(MatrixPtr& matInv, bool memAlloc); + + /// add b to each sample of this. + void addBias(Matrix& b, real scale); + void addSharedBias(Matrix& b, real scale); + + /** + * @code + * add each sample from a to this. + * @endcode + */ + void collectBias(Matrix& a, real scale); + void collectSharedBias(Matrix& a, real scale); + + void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode); + void sequenceAvgBackward(Matrix& a, const IVector& startsPos, int mode); + + /** + * @code + * this.row[i] += table.row[ids[i]] + * @endcode + */ + virtual void selectRows(Matrix& table, IVector& ids); + + /** + * @code + * this[i] = table[i, id[i]] + * @endcode + */ + virtual void selectElements(Matrix& table, IVector& ids); + + /** + * @code + * table.row[ids[i]] += this.row[i] + * @endcode + */ + virtual void addToRows(Matrix& table, IVector& ids); + + void addColumnVector(const Matrix& b); + + /** + * @code + * this = scaleAB*(a*b) + scaleT*this + * @endcode + */ + void mul(const Matrix& a, const Matrix& b, real scaleAB, real scaleT); + + /** + * @code + * this = a*b + * @endcode + */ + void mul(const Matrix& a, const Matrix& b); + + void mul(const GpuMatrix& a, const GpuMatrix& b, real scaleAB, real scaleT); + + void mul(const GpuSparseMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT); + + void mul(const GpuMatrix& a, + const GpuSparseMatrix& b, + real scaleAB, + real scaleT); + + /** + * @code + * this = scaleAB*(this*b) + scaleT*this + * @endcode + */ + void rightMul(Matrix& b, real scaleAB, real scaleT); + + /** + * @code + * this = this* b + * @endcode + */ + void rightMul(Matrix& b); + + /** + * @code + * this = scaleAB*(a*this) + scaleT*this + * @endcode + */ + void leftMul(Matrix& a, real scaleAB, real scaleT); + + /** + * @code + * this = a*this + * @endcode + */ + void leftMul(Matrix& a); + + void colMerge(Matrix& src); + void rowSum(Matrix& sum); + void rowMax(Matrix& max); + void rowMax(IVector& maxIds, Matrix& max); + void colMax(Matrix& max); + void colMax(IVector& maxIds, Matrix& max); + void maxoutForward(Matrix& a, IVector& id, size_t channels, size_t groups); + void maxoutBackward(Matrix& a, IVector& id, size_t channels, size_t groups); + + void oneHotCrossEntropy(Matrix& output, IVector& label); + void oneHotCrossEntropyBp(Matrix& outputV, IVector& label); + void oneHotCrossEntropyWithSelfNorm(Matrix& output, + IVector& label, + real alpha); + void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, + IVector& label, + real alpha); + + void softmax(Matrix& output); + void sequenceSoftmax(Matrix& output, const IVector& index); + void softmaxBackward(Matrix& outputV); + void softmaxDerivative(Matrix& output, Matrix& sftmaxSum); + + /// calculate the sum of squares diff cost. + void sumOfSquares(Matrix& output, Matrix& label); + + /// gradient of sumOfSquares. + void sumOfSquaresBp(Matrix& outputV, Matrix& label); + void tanh(Matrix& output); + void tanhDerivative(Matrix& output); + void softrelu(Matrix& output); + void softreluDerivative(Matrix& output); + void scaledTanh(Matrix& output, real p1, real p2); + + virtual void print(std::ostream& os) const; + virtual void print(std::ostream& os, size_t height, size_t width) const; + + void paramReluForward(Matrix& data, Matrix& W); + void paramReluBackwardW(Matrix& oGrad, Matrix& data); + void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W); + + void check(std::ostream& os, Matrix& refMat, bool printDiff = true); + void randomizeUniform(); + + void classificationError(Matrix& output, IVector& label, size_t topkSize = 1); + + void upsampleForward(Matrix& input, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW); + + void upsampleBackward(Matrix& outputGrad, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW); + + void maxPoolForward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + MatrixPtr maskMatP); + + void maxPoolBackward(Matrix& image, + size_t imgSizeH, + size_t imgSizeW, + Matrix& outGrad, + Matrix& outV, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW); + + void avgPoolForward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + bool excludeMode = true); + + void avgPoolBackward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW, + bool excludeMode = true); + + void maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + + void avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + + void maxSequenceForward(Matrix& input, + const IVector& sequence, + IVector& index); + + void maxSequenceBackward(Matrix& outputGrad, + const IVector& sequence, + IVector& index); + + void bilinearForward(const Matrix& in, + const size_t inImgH, + const size_t inImgW, + const size_t outImgH, + const size_t outImgW, + const size_t numChannels, + const real ratioH, + const real ratioW); + + void bilinearBackward(const Matrix& out, + const size_t outImgH, + const size_t outImgW, + const size_t inImgH, + const size_t inImgW, + const size_t numChannels, + const real ratioH, + const real ratioW); + + void vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW); + + void col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta); + + void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label); + + void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label); + + template + void operator=(const ExpressionType& expr) { + TensorGpuApply(*this, expr); + } +}; + +class CpuMatrix : public Matrix { + private: + MatrixPtr sftmaxSum_; + MatrixPtr sftmaxDot_; + + public: + CpuMatrix(size_t height, size_t width, bool trans = false); + CpuMatrix(real* data, size_t height, size_t width, bool trans = false) + : Matrix(data, height, width, trans, false) {} + CpuMatrix(real* data, + size_t height, + size_t width, + size_t stride, + bool trans = false) + : Matrix(data, height, width, stride, trans, false) {} + + CpuMatrix(CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + bool trans = false) + : Matrix(dataHandle, height, width, trans, false) {} + + ~CpuMatrix(); + + void zeroMem(); + void resetOne(); + void setDiag(real value); + + void resize(size_t newHeight, size_t newWidth); + void resize(size_t newHeight, + size_t newWidth, + size_t newNnz, /* used to allocate space */ + SparseValueType valueType, + SparseFormat format) { + LOG(FATAL) << "Only Support Sparse Matrix"; + } + void setRow(size_t row, + size_t colNum, + const unsigned int* cols, + const real* values) { + LOG(FATAL) << "Only Support Sparse Matrix"; + } + + real getElement(size_t x, size_t y) const; + real getSum(); + void accumulateColSum(Matrix& src); + real getAbsSum(); + + MatrixPtr getTranspose(); + void transpose(MatrixPtr& matTrans, bool memAlloc); + void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise); + + MatrixPtr getInverse(); + void inverse(MatrixPtr& matInv, bool memAlloc); + + void copyFrom(const Matrix& src); + + void copyFrom(const Matrix& src, hl_stream_t stream); + + void copyFrom(const real* cpuSrc, size_t size); + + void copyFrom(const real* cpuSrc, const int64_t* seq); + + void copyFrom(const IVector& src); + + void copyFrom(CpuSparseMatrix& src); + + void copyByRowIndex(Matrix& b, const IVector& rowIndex); + + MatrixPtr clone(size_t height, size_t width, bool useGpu = false); + + void upsampleForward(Matrix& input, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW); + + void upsampleBackward(Matrix& outputGrad, + Matrix& mask, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t outputH, + size_t outputW); + + void maxPoolForward(Matrix& inputMat, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + MatrixPtr maskMatP); + + void maxPoolBackward(Matrix& image, + size_t imgSizeH, + size_t imgSizeW, + Matrix& outGrad, + Matrix& outV, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW); + + void avgPoolForward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t channels, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + size_t paddingH, + size_t paddingW, + bool excludeMode = true); + + void avgPoolBackward(Matrix& input, + size_t imgSizeH, + size_t imgSizeW, + size_t sizeX, + size_t sizeY, + size_t strideH, + size_t strideW, + size_t outputH, + size_t outputW, + real scaleTargets, + real scaleOutput, + size_t paddingH, + size_t paddingW, + bool excludeMode = true); + + void maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + + void avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + + void maxSequenceForward(Matrix& input, + const IVector& sequence, + IVector& index); + + void maxSequenceBackward(Matrix& outputGrad, + const IVector& sequence, + IVector& index); + + real* getRow(size_t row) { return BaseMatrix::rowBuf(row); } + virtual real* getRowBuf(size_t row) { return getRow(row); } + + public: + /// add b to each sample of this. + void addBias(Matrix& b, real scale); + void addSharedBias(Matrix& b, real scale); + + /// add each sample of a to this. + void collectBias(Matrix& a, real scale); + void collectSharedBias(Matrix& a, real scale); + + void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode); + void sequenceAvgBackward(Matrix& a, const IVector& startsPos, int mode); + + /** + * @code + * this.row[i] += table.row[ids[i]] + * @endcode + */ + virtual void selectRows(Matrix& table, IVector& ids); + + /** + * @code + * table.row[ids[i]] += this.row[i] + * @endcode + */ + virtual void addToRows(Matrix& table, IVector& ids); + + /** + * @code + * this[i] = table[i, id[i]] + * @endcode + */ + virtual void selectElements(Matrix& table, IVector& ids); + + /** + * @code + * table[i, id[i]] += this[i] + * @endcode + */ + virtual void addElements(Matrix& table, IVector& ids); + + /** + * use abstract getRow() to get row from table. + * + * Define table as template instead of virtual class for performance sake. + * internal used by above two virtual funcs. + */ + template + void selectRowsImp(TableMatType& table, IVector& ids); + template + void addToRowsImp(TableMatType& table, IVector& ids); + + void addColumnVector(const Matrix& b); + + void mul(const Matrix& a, const Matrix& b, real scaleAB, real scaleT); + void mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); + + void mul(CpuMatrix* a, CpuSparseMatrix* b, real scaleAB, real scaleT); + + static void mul(CpuMatrix* a, + CpuMatrix* b, + CpuSparseMatrix* c, + real scaleAB, + real scaleT); + + /** + * c = a * b + * + * use abstract getRow() to get row from B,C. + * Define B,C as template instead of virtual class for performance sake. + */ + template + static void mul( + CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB, real scaleT); + + virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); + + void mul(const Matrix& a, const Matrix& b); + + void rightMul(Matrix& b, real scaleAB, real scaleT); + void rightMul(Matrix& b); + + void leftMul(Matrix& a, real scaleAB, real scaleT); + void leftMul(Matrix& a); + void colMerge(Matrix& src); + void rowSum(Matrix& sum); + void rowMaxId(IVector& maxIds); + void rowMax(Matrix& max); + void rowMax(IVector& maxIds, Matrix& maxVal); + void colMax(Matrix& max); + void colMax(IVector& maxIds, Matrix& maxVal); + void maxoutForward(Matrix& a, IVector& id, size_t channels, size_t groups); + void maxoutBackward(Matrix& a, IVector& id, size_t channels, size_t groups); + void rowNormalizeL1(Matrix& out); + + void oneHotCrossEntropy(Matrix& output, IVector& label); + void oneHotCrossEntropyBp(Matrix& outputV, IVector& label); + void oneHotCrossEntropyWithSelfNorm(Matrix& output, + IVector& label, + real alpha); + void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, + IVector& label, + real alpha); + + void circularConv(Matrix& b, Matrix& c); + void circularConvDerivative(Matrix& output, + Matrix& prevOut1, + Matrix& prevOut2, + Matrix& prevGrad1, + Matrix& prevGrad2); + + void softmax(Matrix& output); + void sequenceSoftmax(Matrix& output, const IVector& index); + void softmaxDerivative(Matrix& output, Matrix& sftmaxSum); + + /// calculate the sum of squares diff cost. + void sumOfSquares(Matrix& output, Matrix& label); + + /// gradient of sumOfSquares. + void sumOfSquaresBp(Matrix& outputV, Matrix& label); + + void smoothL1(Matrix& output, Matrix& label, real destScale); + void smoothL1Bp(Matrix& output, Matrix& label, real destScale); + + void tanh(Matrix& output); + void tanhDerivative(Matrix& output); + + void softrelu(Matrix& output); + void softreluDerivative(Matrix& output); + void scaledTanh(Matrix& output, real p1, real p2); + + void print(std::ostream& os) const; + void print(std::ostream& os, size_t height, size_t width) const; + void printOneRow(std::ostream& os, size_t idx) const; + + void paramReluForward(Matrix& data, Matrix& W); + void paramReluBackwardW(Matrix& oGrad, Matrix& data); + void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W); + + void check(std::ostream& os, Matrix& refMat, bool printDiff = true); + + real getMin(); + real getMax(); + + void randomizeUniform(); + + void classificationError(Matrix& output, IVector& label, size_t topkSize = 1); + + void addByBitCode(size_t numClasses, const IVector& codes, const Matrix& vec); + + void addByBitCodeBackward(size_t numClasses, + const IVector& codes, + Matrix& vec); + + void mulByBitCode(size_t numClasses, + const IVector& codes, + const Matrix& mat, + const Matrix& input); + + void mulByBitCodeBackwardWeight(size_t numClasses, + const IVector& codes, + Matrix& mat, + const Matrix& input); + + void mulByBitCodeBackwardError(size_t numClasses, + const IVector& codes, + const Matrix& mat, + Matrix& input); + + void sumByBitCode(size_t numClasses, + IVector& codes, + Matrix& sum, + real scaleSum); + + void subByBitCode(size_t numClasses_, IVector& codes); + + void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label); + void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label); + void classificationErrorMulti(Matrix& output, Matrix& label, real threshold); + + void bilinearForward(const Matrix& in, + const size_t inImgH, + const size_t inImgW, + const size_t outImgH, + const size_t outImgW, + const size_t numChannels, + const real ratioH, + const real ratioW); + + void bilinearBackward(const Matrix& out, + const size_t outImgH, + const size_t outImgW, + const size_t inImgH, + const size_t inImgW, + const size_t numChannels, + const real ratioH, + const real ratioW); + + void vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW); + + void col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta); + + template + void operator=(const ExpressionType& expr) { + TensorCpuApply(*this, expr); + } +}; + +class SharedCpuMatrix : public CpuMatrix { + public: +#ifndef PADDLE_MOBILE_INFERENCE + /* blockNum is number of partitions of the matrix */ + SharedCpuMatrix(int blockNum, size_t height, size_t width, bool trans = false) + : CpuMatrix(height, width, trans) { + initShared(blockNum); + } + SharedCpuMatrix( + int blockNum, real* data, size_t height, size_t width, bool trans = false) + : CpuMatrix(data, height, width, trans) { + initShared(blockNum); + } + + SharedCpuMatrix(int blockNum, + CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + bool trans = false) + : CpuMatrix(dataHandle, height, width, trans) { + initShared(blockNum); + } + + SharedCpuMatrix(CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + bool trans = false) + : CpuMatrix(dataHandle, height, width, trans) { + initBlock(1); + } + + ~SharedCpuMatrix() {} + + public: + virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); + virtual void add(Matrix& b, real p1, real p2); + virtual void add(real p1, real p2); + + private: + using Matrix::mul; + void initShared(int blockNum); + void initBlock(int blockNum); + + int blockNum_; + std::vector> blockLocks_; + ThreadLocal localBuf_; + ThreadLocal> localBufRows_; + ThreadLocal> blockSeq_; +#endif +}; + +typedef struct { unsigned int col; } sparse_non_value_t; + +typedef struct { + unsigned int col; + float value; +} sparse_float_value_t; + +} // namespace paddle +#include "ExecViaCpu.h" diff --git a/paddle/legacy/math/MatrixBitCode.cpp b/paddle/legacy/math/MatrixBitCode.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f35f266a30506110eb6c656f7b631d12d8f6ae90 --- /dev/null +++ b/paddle/legacy/math/MatrixBitCode.cpp @@ -0,0 +1,291 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Matrix.h" +#include "hl_gpu.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +namespace { + +struct SimpleCode { + SimpleCode(size_t code, size_t numClasses) : c_(code + numClasses) {} + inline size_t calcIndex(int bit) const { return (c_ >> (bit + 1)) - 1; } + inline bool calcBit(int bit) const { return c_ & (1 << bit); } + inline int getLength() const { return findLastSet(c_) - 1; } + + private: + size_t c_; +}; + +struct SimpleCodeTable { + explicit SimpleCodeTable(size_t numClasses) : numClasses_(numClasses) {} + SimpleCode operator()(size_t code) const { + return SimpleCode(code, numClasses_); + } + size_t size() const { return numClasses_; } + int getMaxCodeLength() const { return findLastSet(numClasses_ - 1); } + + private: + size_t numClasses_; + int maxCodeLength_; +}; + +} // namespace + +/** + * CodeTable class should support 3 functions: + * + * size_t size() + * return the number of codes + * + * int getMaxCodeLength() + * return the maximal code length + * + * Code operator()(size_t i) + * return the i-th code. Code class is descriebed below. + * + * Code class should support 3 functions: + * + * int getLength() + * return the length of the code + * + * bool calcIndex(int bit) + * bit ranges from 0 to getLength() - 1 + * return the index for the (1+bit) level parent + * + * bool calcBit(int bit) + * return true if the bit level parent is the right child of (1+bit) level + * parent + * + */ + +/* + for i: + for j < codeLength: + op(tmat(i, j), vec(0, index(i, j))) +*/ +template +static void addByBitCodeT( + Op op, CodeTable codeTable, const IVector& codes, TMat& tmat, Mat& vec) { + CHECK(!vec.useGpu()); + + size_t numClasses = codeTable.size(); + size_t maxCodeLength = codeTable.getMaxCodeLength(); + size_t numSamples = tmat.getHeight(); + size_t oWidth = tmat.getWidth(); + CHECK_EQ(tmat.getWidth(), maxCodeLength); + CHECK_EQ(codes.getSize(), numSamples); + CHECK_EQ(vec.getHeight(), (size_t)1); + CHECK_EQ(vec.getWidth(), numClasses - 1); + + auto data = tmat.getData(); + auto v = vec.getData(); + const int* c = codes.getData(); + for (size_t i = 0; i < numSamples; ++i) { + auto code = codeTable(c[i]); + int codeLength = code.getLength(); + for (int j = 0; j < codeLength; ++j) { + size_t index = code.calcIndex(j); + op(data[i * oWidth + j], v[index]); + } + } +} + +/* For j < codeLength: + this(i, j) += vec(0, index(i, j)) +*/ +void CpuMatrix::addByBitCode(size_t numClasses, + const IVector& codes, + const Matrix& vec) { + auto op = [](real& t, real v) { t += v; }; + addByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, vec); +} + +/* For j < codeLength: + vec(0, index(i, j)) += this(i, j) +*/ +void CpuMatrix::addByBitCodeBackward(size_t numClasses, + const IVector& codes, + Matrix& vec) { + auto op = [](real t, real& v) { v += t; }; + addByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, vec); +} + +/* + for i: + for j < codeLength: + op(tmat(i, j), mat.row(index(i, j)), input.row(i)) +*/ +template +void mulByBitCodeT(Op op, + CodeTable codeTable, + IVec& codes, + TMat& tmat, + WMat& weight, + InMat& input) { + CHECK(!tmat.useGpu() && !weight.useGpu() && !input.useGpu()); + + size_t numClasses = codeTable.size(); + size_t maxCodeLength = codeTable.getMaxCodeLength(); + size_t numSamples = tmat.getHeight(); + size_t inputDim = input.getWidth(); + size_t oWidth = tmat.getWidth(); + CHECK_EQ(tmat.getWidth(), maxCodeLength); + CHECK_EQ(codes.getSize(), numSamples); + CHECK_EQ(input.getHeight(), numSamples); + CHECK_EQ(weight.getHeight(), numClasses - 1); + CHECK_EQ(weight.getWidth(), inputDim); + + real* data = tmat.getData(); + const int* c = codes.getData(); + for (size_t i = 0; i < numSamples; ++i) { + auto code = codeTable(c[i]); + int codeLength = code.getLength(); + for (int j = 0; j < codeLength; ++j) { + size_t index = code.calcIndex(j); + op(data[i * oWidth + j], weight.rowBuf(index), input.rowBuf(i), inputDim); + } + } +} + +/* For j < codeLength: + this(i, j) += +*/ +void CpuMatrix::mulByBitCode(size_t numClasses, + const IVector& codes, + const Matrix& weight, + const Matrix& input) { + auto op = []( + real& t, const real* weightRow, const real* inputRow, size_t inputDim) { + real sum = 0; + for (size_t k = 0; k < inputDim; ++k) { + sum += weightRow[k] * inputRow[k]; + } + t += sum; + }; + + mulByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, weight, input); +} + +/* For index(i, j) >= 0: + weight.row(index(i, j)) += this(i, j) * input.row(i) +*/ +void CpuMatrix::mulByBitCodeBackwardWeight(size_t numClasses, + const IVector& codes, + Matrix& weight, + const Matrix& input) { + auto op = []( + const real t, real* weightRow, const real* inputRow, size_t inputDim) { + for (size_t k = 0; k < inputDim; ++k) { + weightRow[k] += t * inputRow[k]; + } + }; + + mulByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, weight, input); +} + +/* For j < codeLength: + input.row(i) += this(i, j) * weight.row(index(i, j)) +*/ +void CpuMatrix::mulByBitCodeBackwardError(size_t numClasses, + const IVector& codes, + const Matrix& weight, + Matrix& input) { + auto op = []( + const real t, const real* weightRow, real* inputRow, size_t inputDim) { + for (size_t k = 0; k < inputDim; ++k) { + inputRow[k] += t * weightRow[k]; + } + }; + + mulByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, weight, input); +} + +template +void sumByBitCodeT(CodeTable codeTable, + IVector& codes, + const CpuMatrix& tmat, + Matrix& sum, + real scaleSum) { + size_t maxCodeLength = codeTable.getMaxCodeLength(); + size_t numSamples = tmat.getHeight(); + size_t oWidth = tmat.getWidth(); + CHECK_EQ(tmat.getWidth(), maxCodeLength); + CHECK_EQ(codes.getSize(), numSamples); + CHECK_EQ(sum.getHeight(), numSamples); + CHECK_EQ(sum.getWidth(), (size_t)1); + + const real* data = tmat.getData(); + real* s = sum.getData(); + int* c = codes.getData(); + for (size_t i = 0; i < numSamples; ++i) { + real sm = 0; + auto code = codeTable(c[i]); + int codeLength = code.getLength(); + for (int j = 0; j < codeLength; ++j) { + if (code.calcBit(j)) { + sm += data[i * oWidth + j]; + } + } + s[i] = scaleSum * sm; + } +} + +/* For j < codeLength: + sum(i, 0) = \sum_j bit(i, j) * this(i, j) +*/ +void CpuMatrix::sumByBitCode(size_t numClasses, + IVector& codes, + Matrix& sum, + real scaleSum) { + sumByBitCodeT(SimpleCodeTable(numClasses), codes, *this, sum, scaleSum); +} + +template +void subByBitCodeT(CodeTable codeTable, IVector& codes, CpuMatrix& tmat) { + size_t maxCodeLength = codeTable.getMaxCodeLength(); + size_t numSamples = tmat.getHeight(); + size_t oWidth = tmat.getWidth(); + CHECK_EQ(tmat.getWidth(), maxCodeLength); + CHECK_EQ(codes.getSize(), numSamples); + + real* data = tmat.getData(); + int* c = codes.getData(); + for (size_t i = 0; i < numSamples; ++i) { + auto code = codeTable(c[i]); + int codeLength = code.getLength(); + for (int j = 0; j < codeLength; ++j) { + if (code.calcBit(j)) { + data[i * oWidth + j] -= 1; + } + } + } +} + +/* For j < codeLength + this(i, j) -= bit(i, j) +*/ +void CpuMatrix::subByBitCode(size_t numClasses, IVector& codes) { + subByBitCodeT(SimpleCodeTable(numClasses), codes, *this); +} + +} // namespace paddle diff --git a/paddle/math/MemoryHandle.cpp b/paddle/legacy/math/MemoryHandle.cpp similarity index 100% rename from paddle/math/MemoryHandle.cpp rename to paddle/legacy/math/MemoryHandle.cpp diff --git a/paddle/math/MemoryHandle.h b/paddle/legacy/math/MemoryHandle.h similarity index 100% rename from paddle/math/MemoryHandle.h rename to paddle/legacy/math/MemoryHandle.h diff --git a/paddle/math/NEONFunctions.cpp b/paddle/legacy/math/NEONFunctions.cpp similarity index 100% rename from paddle/math/NEONFunctions.cpp rename to paddle/legacy/math/NEONFunctions.cpp diff --git a/paddle/math/NEONFunctions.h b/paddle/legacy/math/NEONFunctions.h similarity index 100% rename from paddle/math/NEONFunctions.h rename to paddle/legacy/math/NEONFunctions.h diff --git a/paddle/math/PoolAllocator.cpp b/paddle/legacy/math/PoolAllocator.cpp similarity index 100% rename from paddle/math/PoolAllocator.cpp rename to paddle/legacy/math/PoolAllocator.cpp diff --git a/paddle/math/PoolAllocator.h b/paddle/legacy/math/PoolAllocator.h similarity index 100% rename from paddle/math/PoolAllocator.h rename to paddle/legacy/math/PoolAllocator.h diff --git a/paddle/legacy/math/RowBuffer.h b/paddle/legacy/math/RowBuffer.h new file mode 100644 index 0000000000000000000000000000000000000000..9dfd5eff06a39494cea6a8ce0b1f5ead6490b148 --- /dev/null +++ b/paddle/legacy/math/RowBuffer.h @@ -0,0 +1,139 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "MemoryHandle.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +/** + * @brief The RowBuffer class + * Represent the SparseRow Matrix Data. + * + * If not set memory handler, then the data could be auto growth. + */ +class RowBuffer { + public: + /** + * @brief RowBuffer create a auto-growth row buffer. The row length is width. + * @param width the length of each row, a.k.a matrix width. + */ + explicit RowBuffer(size_t width) : width_(width) {} + + /** + * @brief RowBuffer create a row buffer, which cannot be auto-growth. + * @param mem the pre-allocated memory. + * @param width the length of each row, a.k.a matrix width. + */ + RowBuffer(const CpuMemHandlePtr& mem, size_t width) + : preallocatedBuf_(mem), width_(width) {} + + /** + * @brief resize resize the buffer with rowCount + * @param rowCnt number of row. matrix height. + */ + inline void resize(int rowCnt) { + if (preallocatedBuf_) { + CHECK(preallocatedBuf_->getSize() >= rowCnt * width_ * sizeof(real)); + } else { + rowStore_.resize(rowCnt * width_); + } + } + + /** + * @brief get a row buffer with row index. + * @param row the index of row. + * @return row buffer. + */ + inline real* get(int row) const { + if (preallocatedBuf_) { + CHECK_LE((row)*width_ * sizeof(real), preallocatedBuf_->getSize()); + return reinterpret_cast(preallocatedBuf_->getBuf()) + row * width_; + } else { + CHECK_LE((row + 1) * width_, rowStore_.size()); + return const_cast(rowStore_.data() + row * width_); + } + } + + /** + * @brief get a row buffer with row index. If row index is larger than local + * buffer, the size of local buffer will grow. + * @param row the index of row. + * @return row buffer. + */ + inline real* getWithAutoGrowth(int row) { + if (preallocatedBuf_) { + return get(row); + } else { + if ((rowStore_.size() <= row * width_)) { + rowStore_.resize((row + 1) * width_); + } + return rowStore_.data() + row * width_; + } + } + + /** + * @return raw data buffer. + */ + inline real* data() { + if (preallocatedBuf_) { + return reinterpret_cast(preallocatedBuf_->getBuf()); + } else { + return rowStore_.data(); + } + } + + /** + * @brief clear local buffer. It only affect auto-growth buffer. + */ + inline void clear() { + // swap an empty vector to it to free the memory. + std::vector> empty; + rowStore_.swap(empty); + } + + /** + * @brief get current number of rows. + * @return number of rows. + */ + inline size_t getRowCount() const { + if (preallocatedBuf_) { + return preallocatedBuf_->getSize() / sizeof(real) / width_; + } else { + return rowStore_.size() / width_; + } + } + + /** + * @brief get is this buffer can automatically grow or not. + * @return ture if can automacitally grow. + */ + inline bool isAutoGrowth() const { return !preallocatedBuf_; } + + /** + * @brief return the width of matrix. a.k.a length of row. + * @return width of matrix + */ + inline size_t getWidth() const { return width_; } + + private: + //! TODO(yuyang18): Add resize method to CpuMemHandlePtr, then we can get rid + //! of std::vector here. + CpuMemHandlePtr preallocatedBuf_; + std::vector> rowStore_; + size_t width_; +}; +} // namespace paddle diff --git a/paddle/math/SIMDFunctions.cpp b/paddle/legacy/math/SIMDFunctions.cpp similarity index 100% rename from paddle/math/SIMDFunctions.cpp rename to paddle/legacy/math/SIMDFunctions.cpp diff --git a/paddle/math/SIMDFunctions.h b/paddle/legacy/math/SIMDFunctions.h similarity index 100% rename from paddle/math/SIMDFunctions.h rename to paddle/legacy/math/SIMDFunctions.h diff --git a/paddle/legacy/math/SparseMatrix.cpp b/paddle/legacy/math/SparseMatrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6f68252b0a74802946e899e6e13e1da681d76986 --- /dev/null +++ b/paddle/legacy/math/SparseMatrix.cpp @@ -0,0 +1,864 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SparseMatrix.h" +#include +#include +#include +#include "hl_gpu.h" +#include "hl_top_k.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +GpuSparseMatrix::GpuSparseMatrix(size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans) + : Matrix(NULL, height, width, trans, true) { + resize(height, width, nnz, valueType, format); +} + +GpuSparseMatrix::GpuSparseMatrix(GpuMemHandlePtr dataHandle, + hl_sparse_matrix_s_ptr sMatrix, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans, + MemoryHandlePtr sMemoryHandle) + : Matrix(dataHandle, height, width, trans, true) { + CHECK(dataHandle && sMatrix) << "Invalid argument pointer"; + + size_t size = 0; + if (format == SPARSE_CSR) { + size = (height + 1) * sizeof(int) + nnz * sizeof(int); + } else { + size = (width + 1) * sizeof(int) + nnz * sizeof(int); + } + + if (NO_VALUE != valueType) { + size += nnz * sizeof(real); + } + CHECK_LE(size, dataHandle->getSize()); + + sMatrix_ = sMatrix; + + if (sMemoryHandle == NULL) { + sMemoryHandle_ = std::make_shared(dataHandle->getSize()); + } else { + CHECK_EQ(sMemoryHandle->getSize(), dataHandle->getSize()); + sMemoryHandle_ = sMemoryHandle; + } + + elementCnt_ = nnz; + valueType_ = valueType; + format_ = format; + if (format_ == SPARSE_CSR) + sparseResizeCSR(); + else + sparseResizeCSC(); +} + +GpuSparseMatrix::GpuSparseMatrix(hl_sparse_matrix_s_ptr sMatrix, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans, + MemoryHandlePtr sMemoryHandle) + : Matrix(NULL, height, width, trans, true) { + CHECK(sMatrix) << "Invalid argument pointer"; + sMatrix_ = sMatrix; + sMemoryHandle_ = sMemoryHandle; + elementCnt_ = nnz; + format_ = format; + valueType_ = valueType; +} + +GpuSparseMatrix::GpuSparseMatrix(real* value, + int* rows, + int* cols, + size_t height, + size_t width, + size_t nnz, + SparseValueType valueType, + SparseFormat format, + bool trans) + : Matrix(NULL, height, width, trans, true) { + size_t size = 0; + if (format == SPARSE_CSR) { + size = (height + 1) * sizeof(int) + nnz * sizeof(int); + } else { + size = (width + 1) * sizeof(int) + nnz * sizeof(int); + } + + if (NO_VALUE != valueType) { + size += nnz * sizeof(real); + } + elementCnt_ = nnz; + valueType_ = valueType; + format_ = format; + + sMemoryHandle_ = std::make_shared(size); + if (format_ == SPARSE_CSR) { + rows_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf())); + cols_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf()) + + (height_ + 1) * sizeof(int)); + if (NO_VALUE != valueType_) { + value_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf()) + + (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); + } else { + value_ = NULL; + } + + if (sMatrix_ == NULL) { + /* construct hl_sparse_matrix_s */ + hl_sparse_matrix_s tmp; + hl_construct_sparse_matrix( + &tmp, + value, + rows, + cols, + HL_SPARSE_CSR, + valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, + height_, + width_, + elementCnt_); + hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); + sMatrix_ = tmp2; + } + + } else { + cols_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf())); + rows_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf()) + + (width_ + 1) * sizeof(int)); + if (NO_VALUE != valueType_) { + value_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf()) + + (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); + } else { + value_ = NULL; + } + + if (sMatrix_ == NULL) { + /* construct hl_sparse_matrix_s */ + hl_sparse_matrix_s tmp; + hl_construct_sparse_matrix( + &tmp, + value, + rows, + cols, + HL_SPARSE_CSC, + valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, + height_, + width_, + elementCnt_); + hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); + sMatrix_ = tmp2; + } + } +} + +void GpuSparseMatrix::sparseResizeCSR() { + rows_ = + reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf())); + cols_ = + reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf()) + + (height_ + 1) * sizeof(int)); + if (NO_VALUE != valueType_) { + value_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf()) + + (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); + } else { + value_ = NULL; + } + + if (sMatrix_ == NULL) { + /* construct hl_sparse_matrix_s */ + hl_sparse_matrix_s tmp; + hl_construct_sparse_matrix( + &tmp, + data_, + memoryHandle_->getSize(), + HL_SPARSE_CSR, + valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, + height_, + width_, + elementCnt_); + hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); + sMatrix_ = tmp2; + } +} + +void GpuSparseMatrix::sparseResizeCSC() { + cols_ = + reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf())); + rows_ = + reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf()) + + (width_ + 1) * sizeof(int)); + if (NO_VALUE != valueType_) { + value_ = reinterpret_cast( + reinterpret_cast(sMemoryHandle_->getBuf()) + + (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); + } else { + value_ = NULL; + } + + if (sMatrix_ == NULL) { + /* construct hl_sparse_matrix_s */ + hl_sparse_matrix_s tmp; + hl_construct_sparse_matrix( + &tmp, + memoryHandle_->getBuf(), + memoryHandle_->getSize(), + HL_SPARSE_CSC, + valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, + height_, + width_, + elementCnt_); + hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); + sMatrix_ = tmp2; + } +} + +void GpuSparseMatrix::resize(size_t newHeight, + size_t newWidth, + size_t newNnz, + SparseValueType valueType, + SparseFormat format) { + if (format == SPARSE_CSR) { + resizeCSR(newHeight, newWidth, newNnz, valueType); + } else { + resizeCSC(newHeight, newWidth, newNnz, valueType); + } +} + +void GpuSparseMatrix::resizeCSR(size_t newHeight, + size_t newWidth, + size_t newNnz, + SparseValueType valueType) { + size_t newSize = (newHeight + 1) * sizeof(int) + newNnz * sizeof(int); + if (NO_VALUE != valueType) { + newSize += newNnz * sizeof(real); + } + + if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) { + memoryHandle_ = std::make_shared(newSize); + data_ = reinterpret_cast(memoryHandle_->getBuf()); + sMemoryHandle_ = std::make_shared(newSize); + end_ = reinterpret_cast(sMemoryHandle_->getBuf()) + + sMemoryHandle_->getSize(); + sMatrix_ = NULL; + } else if (valueType != valueType_) { + sMatrix_ = NULL; + } else { + /* + * newNnz > elementCnt_ is necessary for the following condition: + * Firstly, height_ is 9 elementCnt_ is 56 + * Secondly, height_ is 11 elementCnt_ is 44 + * ==> height_ is bigger, sMatrix_ will resize, and total item is 44 now + * Then, height_ is 10 elementCnt_ is 52 + * ==> Without newNnz > elementCnt_ condition, sMatrix_ will fail + */ + if ((ssize_t)((newHeight + 1) * sizeof(int)) > + ((char*)cols_ - (char*)rows_) || + newNnz > static_cast(sMatrix_->nnz)) { + sMatrix_ = NULL; + } else if (NO_VALUE == valueType) { + if ((ssize_t)(newNnz * sizeof(int)) > (end_ - (char*)cols_)) { + sMatrix_ = NULL; + } + } else { + if ((ssize_t)(newNnz * sizeof(int)) > ((char*)value_ - (char*)cols_) || + (ssize_t)(newNnz * sizeof(real)) > (end_ - (char*)value_)) { + sMatrix_ = NULL; + } + } + } + + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newNnz; + valueType_ = valueType; + format_ = SPARSE_CSR; + + if (sMatrix_ == NULL) { + sparseResizeCSR(); + } +} + +void GpuSparseMatrix::resizeCSC(size_t newHeight, + size_t newWidth, + size_t newNnz, + SparseValueType valueType) { + size_t newSize = (newWidth + 1) * sizeof(int) + newNnz * sizeof(int); + if (NO_VALUE != valueType) { + newSize += newNnz * sizeof(real); + } + + if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) { + memoryHandle_ = std::make_shared(newSize); + data_ = reinterpret_cast(memoryHandle_->getBuf()); + sMemoryHandle_ = std::make_shared(newSize); + end_ = reinterpret_cast(sMemoryHandle_->getBuf()) + + sMemoryHandle_->getSize(); + sMatrix_ = NULL; + } else if (valueType != valueType_) { + sMatrix_ = NULL; + } else { + /* + * newNnz > elementCnt_ is necessary for the following condition: + * Firstly, height_ is 9 elementCnt_ is 56 + * Secondly, height_ is 11 elementCnt_ is 44 + * ==> height_ is bigger, sMatrix_ will resize, + * and total item is 44 now + * Then, height_ is 10 elementCnt_ is 52 + * ==> Without newNnz > elementCnt_ condition, sMatrix_ will fail + */ + if ((ssize_t)((newWidth + 1) * sizeof(int)) > + ((char*)rows_ - (char*)cols_) || + newNnz > static_cast(sMatrix_->nnz)) { + sMatrix_ = NULL; + } else if (NO_VALUE == valueType) { + if ((ssize_t)(newNnz * sizeof(int)) > (end_ - (char*)rows_)) { + sMatrix_ = NULL; + } + } else { + if ((ssize_t)(newNnz * sizeof(int)) > ((char*)value_ - (char*)rows_) || + (ssize_t)(newNnz * sizeof(real)) > (end_ - (char*)value_)) { + sMatrix_ = NULL; + } + } + } + + height_ = newHeight; + width_ = newWidth; + elementCnt_ = newNnz; + valueType_ = valueType; + format_ = SPARSE_CSC; + + if (sMatrix_ == NULL) { + sparseResizeCSC(); + } +} + +void GpuSparseMatrix::resize(size_t newHeight, size_t newWidth) { + resize(newHeight, newWidth, elementCnt_, valueType_, format_); +} + +MatrixPtr GpuSparseMatrix::getTranspose() { + CHECK(memoryHandle_.get() || sMatrix_) << "not supported"; + if (memoryHandle_.get()) { + MatrixPtr copy_T(new GpuSparseMatrix( + std::dynamic_pointer_cast(memoryHandle_), + sMatrix_, + height_, + width_, + elementCnt_, + valueType_, + format_, + true, + sMemoryHandle_)); + return copy_T; + } else { + MatrixPtr copy_T(new GpuSparseMatrix(sMatrix_, + height_, + width_, + elementCnt_, + valueType_, + format_, + true, + sMemoryHandle_)); + return copy_T; + } +} + +void GpuSparseMatrix::copyRow(int offsets, + size_t colNum, + const sparse_non_value_t* row) { + memcpy(cols_ + offsets, row, sizeof(int) * colNum); +} + +void GpuSparseMatrix::copyRow(int offsets, + size_t colNum, + const sparse_float_value_t* row) { + for (size_t j = 0; j < colNum; j++) { + cols_[offsets + j] = row[j].col; + value_[offsets + j] = row[j].value; + } +} + +void GpuSparseMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { + if (auto mat = dynamic_cast(&src)) { + copyFrom(*(const_cast(mat)), stream); + } else if (auto mat = dynamic_cast(&src)) { + copyFrom(*(const_cast(mat)), stream); + } else { + LOG(FATAL) << "Not implemented"; + } +} + +void GpuSparseMatrix::copyFrom(const Matrix& src) { + copyFrom(src, HPPL_STREAM_1); + hl_stream_synchronize(HPPL_STREAM_1); +} + +template +void GpuSparseMatrix::copyFrom(int64_t* ids, + int64_t* indices, + T* data, + hl_stream_t stream) { + CHECK_EQ(format_, SPARSE_CSR); + size_t nnz = 0; + for (size_t i = 0; i < height_; i++) { + int64_t id = ids[i]; + nnz += indices[id + 1] - indices[id]; + } + + resize(height_, + width_, + nnz, + sizeof(T) == sizeof(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE, + format_); + + rows_[0] = 0; + for (size_t i = 0; i < height_; i++) { + int64_t id = ids[i]; + size_t colNum = indices[id + 1] - indices[id]; + rows_[i + 1] = rows_[i] + colNum; + + T* row = data + indices[id]; + copyRow(rows_[i], colNum, row); + } + + sMatrix_->format = HL_SPARSE_CSR; + sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; + sMatrix_->rows = height_; + sMatrix_->cols = width_; + sMatrix_->nnz = nnz; + hl_memcpy_csr_matrix(sMatrix_.get(), value_, rows_, cols_, stream); +} + +void GpuSparseMatrix::setRow(size_t row, + size_t colNum, + const unsigned int* cols, + const real* values) { + CHECK_EQ(format_, SPARSE_CSR); + if (NO_VALUE == valueType_) { + CHECK_LT(row, height_); + CHECK(NULL != cols); + CHECK(NULL == values); + } else { + CHECK_LT(row, height_); + CHECK(NULL != cols); + CHECK(NULL != values); + } + if (0 == row) { + rows_[row] = 0; + } + rows_[row + 1] = rows_[row] + colNum; + + memcpy(cols_ + rows_[row], cols, sizeof(*cols) * colNum); + if (FLOAT_VALUE == valueType_) { + memcpy(value_ + rows_[row], values, sizeof(*values) * colNum); + } + + if (height_ - 1 == row) { + sMatrix_->format = HL_SPARSE_CSR; + sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; + sMatrix_->rows = height_; + sMatrix_->cols = width_; + sMatrix_->nnz = elementCnt_; + hl_memcpy_csr_matrix( + sMatrix_.get(), value_, rows_, cols_, HPPL_STREAM_DEFAULT); + } +} + +SparseValueType GpuSparseMatrix::getValueType() const { return valueType_; } + +void GpuSparseMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { + CHECK_EQ(format_, SPARSE_CSC); + int nnz = sMatrix_->nnz; + if (memAlloc) { + matTrans = std::make_shared( + width_, height_, nnz, valueType_, format_, false); + } else { + CHECK(matTrans != nullptr); + } + + CpuIVector rows(nnz); + CpuIVector cols(width_ + 1); + CpuIVector cols_full(nnz); + CpuVector value(nnz); + hl_stream_t stream = HPPL_STREAM_1; + hl_memcpy_from_csc_matrix(value.getData(), + nnz, + rows.getData(), + nnz, + cols.getData(), + width_ + 1, + sMatrix_.get(), + stream); + + hl_stream_synchronize(stream); + + /*for every non zero number, get its column index*/ + std::vector dataVec; + for (size_t i = 0; i < width_; i++) { + for (int j = cols.getData()[i]; j < cols.getData()[i + 1]; j++) { + cols_full.getData()[j] = i; + } + } + + /*sort row index and column index by the ascending order*/ + for (int i = 0; i < nnz; i++) { + dataVec.emplace_back( + rows.getData()[i], cols_full.getData()[i], value.getData()[i]); + } + std::sort(dataVec.begin(), dataVec.end(), [](Element a, Element b) { + return a.row < b.row || (a.row == b.row && a.col < b.col); + }); + + /*get sorted data, row index, and col index, put them in the right place*/ + cols.resize(height_ + 1); + rows.resize(nnz); + value.resize(nnz); + + cols.getData()[0] = 0; + rows.getData()[0] = dataVec[0].col; + value.getData()[0] = dataVec[0].val; + for (int i = 1; i < nnz; i++) { + if (dataVec[i].row != dataVec[i - 1].row) { + for (int j = dataVec[i - 1].row + 1; j <= dataVec[i].row; j++) { + cols.getData()[j] = i; + } + } + rows.getData()[i] = dataVec[i].col; + value.getData()[i] = dataVec[i].val; + } + cols.getData()[height_] = nnz; + + /*copy back from cpu*/ + GpuSparseMatrixPtr dest = + std::dynamic_pointer_cast(matTrans); + hl_memcpy_csc_matrix((dest->sMatrix_).get(), + value.getData(), + rows.getData(), + cols.getData(), + stream); + hl_stream_synchronize(stream); +} + +void GpuSparseMatrix::mul(const GpuMatrix& a, + const GpuMatrix& b, + real scaleAB, + real scaleT) { + CHECK(a.useGpu_ && b.useGpu_) << "type not match"; + CHECK(!trans_) << "trans not supported"; + real* A_d = (real*)a.getData(); + real* B_d = (real*)b.getData(); + hl_sparse_matrix_s C_d = sMatrix_.get(); + hl_trans_op_t a_trans = a.trans_ ? HPPL_OP_T : HPPL_OP_N; + hl_trans_op_t b_trans = b.trans_ ? HPPL_OP_T : HPPL_OP_N; + + if (!a.trans_ && !b.trans_) { + CHECK(height_ == a.getHeight()); + CHECK(width_ == b.getWidth()); + CHECK(a.getWidth() == b.getHeight()); + } else if (a.trans_ && !b.trans_) { + CHECK(height_ == a.getWidth()); + CHECK(width_ == b.getWidth()); + CHECK(a.getHeight() == b.getHeight()); + } else if (!a.trans_ && b.trans_) { + CHECK(height_ == a.getHeight()); + CHECK(width_ == b.getHeight()); + CHECK(a.getWidth() == b.getWidth()); + } else { + LOG(INFO) << "Not support"; + } + int dimM = height_; + int dimN = width_; + int dimK = !b.trans_ ? b.getHeight() : b.getWidth(); + hl_sparse_matrix_mul( + A_d, a_trans, B_d, b_trans, C_d, dimM, dimN, dimK, scaleAB, scaleT); +} + +void GpuSparseMatrix::mul(const Matrix& a, + const Matrix& b, + real scaleAB, + real scaleT) { + const auto a_ptr = dynamic_cast(&a); + const auto b_ptr = dynamic_cast(&b); + if (a_ptr && b_ptr) { + mul(*a_ptr, *b_ptr, scaleAB, scaleT); + } else { + LOG(FATAL) << "not supported"; + } +} + +template +void printBuf(std::ostream& os, T* a, size_t len, const char* name) { + os << "\n: " << name << " ["; + for (size_t i = 0; i < len; i++) { + os << a[i] << " "; + } + os << "]\n"; +} + +void GpuSparseMatrix::print(std::ostream& os) const { + if (format_ == SPARSE_CSC) { + int nnz = sMatrix_->nnz; + IVectorPtr rows = IVector::create(nnz, false); + IVectorPtr cols = IVector::create(width_ + 1, false); + VectorPtr value = Vector::create(nnz, false); + hl_stream_t stream = HPPL_STREAM_DEFAULT; + hl_memcpy_from_csc_matrix(value->getData(), + value->getSize(), + rows->getData(), + rows->getSize(), + cols->getData(), + cols->getSize(), + sMatrix_.get(), + stream); + hl_stream_synchronize(stream); + + printBuf(os, cols->getData(), width_ + 1, "col idx"); + printBuf(os, rows->getData(), elementCnt_, "row idx"); + printBuf(os, value->getData(), elementCnt_, "value"); + } +} + +void GpuSparseMatrix::copyFromCSR(CpuSparseMatrix& src, hl_stream_t stream) { + trans_ = src.trans_; + size_t nnz = src.getElementCnt(); + + resize(src.getHeight(), src.getWidth(), nnz, valueType_, src.getFormat()); + // if have different value type, only copy rows and cols + SparseValueType vType = + valueType_ != src.getValueType() ? NO_VALUE : valueType_; + + sMatrix_->format = HL_SPARSE_CSR; + sMatrix_->type = vType == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; + sMatrix_->rows = height_; + sMatrix_->cols = width_; + sMatrix_->nnz = nnz; + + hl_memcpy_csr_matrix(sMatrix_.get(), + vType == NO_VALUE ? NULL : src.getValue(), + src.getRows(), + src.getCols(), + stream); + + // restore type of sMatrix_ + sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; +} + +void GpuSparseMatrix::copyFromCSC(CpuSparseMatrix& src, hl_stream_t stream) { + trans_ = src.trans_; + size_t nnz = src.getElementCnt(); + + resize(src.getHeight(), src.getWidth(), nnz, valueType_, src.getFormat()); + + // if have different value type, only copy rows and cols + SparseValueType vType = + valueType_ != src.getValueType() ? NO_VALUE : valueType_; + + sMatrix_->format = HL_SPARSE_CSC; + sMatrix_->type = vType == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; + sMatrix_->rows = height_; + sMatrix_->cols = width_; + sMatrix_->nnz = nnz; + + hl_memcpy_csc_matrix(sMatrix_.get(), + vType == NO_VALUE ? NULL : src.getValue(), + src.getRows(), + src.getCols(), + stream); + + // restore type of sMatrix_ + sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; +} + +void GpuSparseMatrix::copyFrom(GpuSparseMatrix& src, hl_stream_t stream) { + CHECK(trans_ == src.trans_); + CHECK(format_ == src.getFormat()); + resize(src.getHeight(), + src.getWidth(), + elementCnt_, + valueType_, + src.getFormat()); + + size_t rowSize = format_ == SPARSE_CSC ? elementCnt_ : height_ + 1; + size_t colSize = format_ == SPARSE_CSC ? width_ + 1 : elementCnt_; + + if (valueType_ == FLOAT_VALUE && src.getValueType() == FLOAT_VALUE) { + hl_memcpy_async( + getValue(), src.getValue(), sizeof(real) * elementCnt_, stream); + } + CHECK(getRows()); + CHECK(src.getRows()); + + hl_memcpy_async(getRows(), src.getRows(), sizeof(int) * rowSize, stream); + hl_memcpy_async(getCols(), src.getCols(), sizeof(int) * colSize, stream); +} + +void GpuSparseMatrix::copyFrom(CpuSparseMatrix& src, hl_stream_t stream) { + if (format_ == SPARSE_CSR) { + copyFromCSR(src, stream); + } else { + copyFromCSC(src, stream); + } +} + +void GpuSparseMatrix::trimFromCSR(const CpuSparseMatrix& src) { + trans_ = src.trans_; + int* srcCols = src.getCols(); + size_t nnz = std::count_if(srcCols, + srcCols + src.getElementCnt(), + [this](size_t n) { return n < this->width_; }); + resize(height_, width_, nnz, valueType_, format_); + + rows_[0] = 0; + size_t index = 0; + for (size_t r = 0; r < height_; ++r) { + for (int i = src.getRows()[r]; i < src.getRows()[r + 1]; ++i) { + if (srcCols[i] < (int)width_) { + cols_[index] = srcCols[i]; + if (valueType_ == FLOAT_VALUE) { + value_[index] = src.getValue()[i]; + } + ++index; + } + } + rows_[r + 1] = index; + } + CHECK_EQ(index, nnz); + + sMatrix_->format = HL_SPARSE_CSR; + sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; + sMatrix_->rows = height_; + sMatrix_->cols = width_; + sMatrix_->nnz = nnz; + + hl_memcpy_csr_matrix(sMatrix_.get(), + valueType_ == NO_VALUE ? NULL : value_, + rows_, + cols_, + /*default stream = */ HPPL_STREAM_DEFAULT); +} + +void GpuSparseMatrix::trimFromCSC(const CpuSparseMatrix& src) { + trans_ = src.trans_; + size_t nnz = src.getCols()[width_] - src.getCols()[0]; + resize(height_, width_, nnz, valueType_, format_); + + cols_[0] = 0; + for (size_t i = 0; i < width_; i++) { + cols_[i + 1] = cols_[i] + (int)(src.getRowNum(i)); + } + memcpy(rows_, src.getRows() + src.getCols()[0], sizeof(int) * nnz); + if (valueType_ == FLOAT_VALUE) { + memcpy(value_, src.getValue() + src.getCols()[0], sizeof(real) * nnz); + } + + sMatrix_->format = HL_SPARSE_CSC; + sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; + sMatrix_->rows = height_; + sMatrix_->cols = width_; + sMatrix_->nnz = nnz; + + hl_memcpy_csc_matrix(sMatrix_.get(), + valueType_ == NO_VALUE ? NULL : value_, + rows_, + cols_, + /*default stream = */ HPPL_STREAM_DEFAULT); +} + +void GpuSparseMatrix::trimFrom(const CpuSparseMatrix& src) { + if (format_ == SPARSE_CSR) { + trimFromCSR(src); + } else { + trimFromCSC(src); + } +} + +void GpuSparseMatrix::addBias(Matrix& b, real scale) { + CHECK(b.getHeight() == 1) << "the Bias should be a vector"; + hl_sparse_matrix_s A_d = sMatrix_.get(); + hl_sparse_matrix_add_bias(A_d, b.getData(), scale); +} + +void GpuSparseMatrix::add3(GpuMatrix* b) { + CHECK(getFormat() != SPARSE_CSC) << "Not supported"; + CHECK(height_ == b->getHeight()); + CHECK(width_ == b->getWidth()); + real* B_d = b->getData(); + hl_sparse_matrix_s A_d = sMatrix_.get(); + hl_sparse_matrix_add_dense(A_d, B_d, height_, width_, 1, 0); +} + +void GpuSparseMatrix::add3(MatrixPtr b) { + if (dynamic_cast(b.get())) { + add3(dynamic_cast(b.get())); + } else { + LOG(FATAL) << "not supported"; + } +} + +void GpuSparseMatrix::zeroMem() { + CHECK(valueType_ == FLOAT_VALUE); + real* value = getValue(); + if (value == NULL) { + LOG(FATAL) << "value is nullptr"; + } + hl_matrix_zero_mem(value, elementCnt_); +} + +void GpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { +#ifdef PADDLE_WITH_CUDA + CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; + size_t numSamples = getHeight(); + size_t beam = maxVal.getWidth(); + CHECK_EQ(maxIds.getSize(), numSamples * beam); + CHECK_EQ(maxVal.getHeight(), numSamples); + CHECK_EQ(format_, SPARSE_CSR) << "Only support SPARSE_CSR"; + + hl_sparse_matrix_top_k(maxVal.getData(), + maxVal.getStride(), + maxIds.getData(), + sMatrix_.get(), + beam, + numSamples); +#endif +} + +template void GpuSparseMatrix::copyFrom(int64_t* ids, + int64_t* indices, + sparse_non_value_t* data, + hl_stream_t stream); +template void GpuSparseMatrix::copyFrom(int64_t* ids, + int64_t* indices, + sparse_float_value_t* data, + hl_stream_t stream); +} // namespace paddle diff --git a/paddle/math/SparseMatrix.h b/paddle/legacy/math/SparseMatrix.h similarity index 100% rename from paddle/math/SparseMatrix.h rename to paddle/legacy/math/SparseMatrix.h diff --git a/paddle/legacy/math/SparseRowMatrix.cpp b/paddle/legacy/math/SparseRowMatrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..39bcdf22984db766283a3b4fbf56f224f730c5f8 --- /dev/null +++ b/paddle/legacy/math/SparseRowMatrix.cpp @@ -0,0 +1,282 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SparseRowMatrix.h" +#include "CpuSparseMatrix.h" + +#include + +#include "paddle/legacy/utils/Logging.h" + +#include "SIMDFunctions.h" + +#include "paddle/legacy/utils/Thread.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +const unsigned int SparseRowCpuMatrix::kUnusedId_ = -1U; + +void SparseRowCpuMatrix::init(size_t height, size_t width) { + height_ = height; + if (!indexDictHandle_) { + indexDictHandle_.reset(new IndexDict); + indexDictHandle_->globalIndices.assign(height, kUnusedId_); + } + localIndices_ = &indexDictHandle_->localIndices; + globalIndices_ = indexDictHandle_->globalIndices.data(); +} + +void SparseRowCpuMatrix::mul(CpuSparseMatrix* a, + CpuMatrix* b, + real scaleAB, + real scaleT) { + CpuMatrix::mul(a, b, this, scaleAB, scaleT); +} + +void SparseRowCpuMatrix::copyFrom(const real* src, size_t size) { + LOG(FATAL) << "This should not be called"; +} + +void SparseRowCpuMatrix::zeroMem() { + apply([](real* buf, size_t len) { memset(buf, 0, sizeof(real) * len); }); + clearRows(); +} + +void SparseRowCpuMatrix::applyL1(real learningRate, real decayRate) { + apply([=](real* buf, size_t len) { + CpuVector value(0, nullptr); + value.subVecFrom(buf, 0, len); + value.applyL1(learningRate, decayRate); + }); +} + +void SparseRowCpuMatrix::sgdUpdate(BaseMatrix& value, + IVector& t0, + real learningRate, + int currentTime, + real decayRate, + bool useL1, + bool fini) { + std::vector& localIndices = indexDictHandle_->localIndices; + + // t0 and value are vectors + CHECK_EQ(t0.getSize(), this->height_); + CHECK_EQ(value.width_, this->height_ * this->width_); + + if (decayRate == 0.0f) { + if (fini) { + return; + } + + for (size_t i = 0; i < localIndices.size(); ++i) { + real* g = getLocalRow(i); + real* v = value.rowBuf(localIndices[i]); + for (size_t j = 0; j < this->width_; ++j) { + v[j] -= learningRate * g[j]; + } + } + return; + } // else + + if (useL1) { // L1 decay + if (fini) { + for (size_t i = 0; i < this->height_; ++i) { + real* v = value.rowBuf(i); + int* t = t0.getData() + i; + if (t[0] < currentTime) { + // W(t0) -> W(t+1) + int tDiff = currentTime - t[0]; + real delta = tDiff * learningRate * decayRate; + simd::decayL1(v, v, delta, this->width_); + } + } + return; + } // else + + for (size_t i = 0; i < localIndices.size(); ++i) { + real* g = getLocalRow(i); + real* v = value.rowBuf(localIndices[i]); + int* t = t0.getData() + localIndices[i]; + if (t[0] < currentTime) { + // W(t0) -> W(t) + int tDiff = currentTime - t[0]; + real delta = tDiff * learningRate * decayRate; + simd::decayL1(v, v, delta, this->width_); + } + + // W(t) -> W(t+1) + for (size_t j = 0; j < this->width_; ++j) { + v[j] -= learningRate * g[j]; + } + simd::decayL1(v, v, learningRate * decayRate, this->width_); + + // state update to t+1 + t[0] = currentTime + 1; + } + + } else { // L2 decay + if (fini) { + for (size_t i = 0; i < this->height_; ++i) { + real* v = value.rowBuf(i); + int* t = t0.getData() + i; + if (t[0] < currentTime) { + // W(t0) -> W(t+1) + int tDiff = currentTime - t[0]; + real recip = 1.0f / (1.0f + tDiff * learningRate * decayRate); + for (size_t j = 0; j < this->width_; ++j) { + v[j] *= recip; + } + } + } + return; + } // else + + real recipDecay = 1.0f / (1.0f + learningRate * decayRate); + + for (size_t i = 0; i < localIndices.size(); ++i) { + real* g = getLocalRow(i); + real* v = value.rowBuf(localIndices[i]); + int* t = t0.getData() + localIndices[i]; + if (t[0] < currentTime) { + // W(t0) -> W(t) + int tDiff = currentTime - t[0]; + real recip = 1.0f / (1.0f + tDiff * learningRate * decayRate); + for (size_t j = 0; j < this->width_; ++j) { + v[j] *= recip; + } + } + + // W(t) -> W(t+1) + for (size_t j = 0; j < this->width_; ++j) { + v[j] = recipDecay * (v[j] - learningRate * g[j]); + } + + // state update to t+1 + t[0] = currentTime + 1; + } + } +} + +void SparseRowCpuMatrix::addTo(BaseMatrix& dest, + std::vector& ids, + size_t tid, + size_t numThreads) { + CHECK(!dest.useGpu_); + CHECK_EQ(dest.height_ * dest.width_, this->height_ * this->width_); + + std::vector& localIndices = indexDictHandle_->localIndices; + for (size_t i = 0; i < localIndices.size(); ++i) { + uint32_t id = localIndices[i]; + if (id % numThreads == tid) { + simd::addTo(dest.rowBuf(id), getLocalRow(i), this->width_); + ids.push_back(id); + } + } +} + +void SparseRowCpuMatrix::addTo(SparseRowCpuMatrix& dest, + size_t tid, + size_t numThreads) { + CHECK(!dest.useGpu_); + CHECK_EQ(dest.height_ * dest.width_, this->height_ * this->width_); + + std::vector& localIndices = indexDictHandle_->localIndices; + for (size_t i = 0; i < localIndices.size(); ++i) { + uint32_t id = localIndices[i]; + if (id % numThreads == tid) { + dest.checkIndex(id); + simd::addTo(dest.getRow(id), getLocalRow(i), this->width_); + } + } +} + +void SparseRowCpuMatrix::zeroMemThread(size_t tid, size_t numThreads) { + std::vector& localIndices = indexDictHandle_->localIndices; + for (size_t i = 0; i < localIndices.size(); ++i) { + uint32_t id = localIndices[i]; + if (id % numThreads == tid) { + memset(this->getLocalRow(i), 0, this->width_ * sizeof(real)); + } + } +} + +void SparseAutoGrowRowCpuMatrix::mul(CpuSparseMatrix* a, + CpuMatrix* b, + real scaleAB, + real scaleT) { + CpuMatrix::mul( + a, b, this, scaleAB, scaleT); +} + +void CacheRowCpuMatrix::mul(CpuSparseMatrix* a, + CpuMatrix* b, + real scaleAB, + real scaleT) { + CpuMatrix::mul(a, b, this, scaleAB, scaleT); +} + +void SparsePrefetchRowCpuMatrix::addRows(const unsigned int* ids, size_t len) { + std::vector& localIndices = indexDictHandle_->localIndices; + for (size_t i = 0; i < len; i++) { + CHECK_LT(*(ids + i), this->getHeight()) + << "id:" << *(ids + i) << "Height:" << this->getHeight() + << "sparse id value exceeds the max input dimension, " + << "it could be caused invalid input data samples"; + } + localIndices.insert(localIndices.end(), ids, ids + len); +} + +void SparsePrefetchRowCpuMatrix::addRows(MatrixPtr input) { + CpuSparseMatrix* mat = dynamic_cast(input.get()); + CHECK(mat) << "only support sparse matrix"; + addRows(reinterpret_cast(mat->getCols()), + mat->getElementCnt()); +} + +void SparsePrefetchRowCpuMatrix::addRows(IVectorPtr ids) { + std::vector& localIndices = indexDictHandle_->localIndices; + size_t numSamples = ids->getSize(); + int* index = ids->getData(); + for (size_t i = 0; i < numSamples; ++i) { + if (index[i] == -1) continue; + + unsigned int id = (unsigned int)index[i]; + CHECK_LT(id, this->getHeight()) + << "id:" << id << "Height:" << this->getHeight() + << "sparse id value exceeds the max input dimension, " + << "it could be caused invalid input data samples"; + localIndices.push_back(id); + } +} + +void SparsePrefetchRowCpuMatrix::setupIndices() { + auto& localIndices = indexDictHandle_->localIndices; + uniqueIds(localIndices); + // for each sparse row + for (size_t id = 0; id < localIndices.size(); ++id) { + globalIndices_[localIndices[id]] = id; // sparse row -> local id + } + checkStoreSize(); +} + +void SparseRowCpuMatrix::checkIndices() { + std::vector& localIndices = indexDictHandle_->localIndices; + for (size_t i = 0; i < localIndices.size(); ++i) { + CHECK_EQ(globalIndices_[localIndices[i]], i); + } + checkStoreSize(); +} + +} // namespace paddle diff --git a/paddle/legacy/math/SparseRowMatrix.h b/paddle/legacy/math/SparseRowMatrix.h new file mode 100644 index 0000000000000000000000000000000000000000..e206747a41c9f3a0f058bf3b0a94472bf4b2c349 --- /dev/null +++ b/paddle/legacy/math/SparseRowMatrix.h @@ -0,0 +1,341 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#ifndef PADDLE_MOBILE_INFERENCE + +#include +#include +#include +#include "Matrix.h" +#include "RowBuffer.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +/** + * Sparse Row + */ +class SparseRowCpuMatrix : public CpuMatrix { + public: + struct IndexDict { + // In the following, global id means the row id in the original matrix. + // Local id means the row id in the local storage which only contains + // the sparse rows. + std::vector localIndices; // local id -> global id + std::vector globalIndices; // global id -> local id + }; + typedef std::shared_ptr IndexDictPtr; + + /// heightStore is max number of rows of the sparse matrix. + SparseRowCpuMatrix(CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + IndexDictPtr indexDictHandle = nullptr, + bool trans = false) + : CpuMatrix(nullptr, height, width, trans), + indexDictHandle_(indexDictHandle) { + init(height, width); + buf_.reset(new RowBuffer(dataHandle, width)); + } + + virtual ~SparseRowCpuMatrix() {} + + public: + /** + * Get the row buf + * + * @param row row id in the original matrix + */ + real* getRow(size_t row) { + CHECK_NE(globalIndices_[row], kUnusedId_); + return getLocalRow(globalIndices_[row]); + } + + /** + * Get the row buf + * + * @param row row id in local storage + */ + real* getLocalRow(size_t row) { return buf_->getWithAutoGrowth(row); } + + /** + * reserve the storage for rows according to current size of + * indexDictHandle. + * + * This is only used when SparseRowCpuMatrix is constructed with + * indexDictHandle. + */ + void reserveStore() { buf_->resize(localIndices_->size()); } + + // row is the row id in the original matrix + virtual real* getRowBuf(size_t row) { return getRow(row); } + + virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); + + /** + * Fill data according to row indexs added, setup indices inside. + * + * *src* and *size* are data and size of normal dense CpuMatrix. + */ + virtual void copyFrom(const real* src, size_t size); + virtual void zeroMem(); + + /** + * apply L1 to all sparse rows, should be apply after indices ready. + */ + virtual void applyL1(real learningRate, real decayRate); + + void clearIndices() { clearRows(); } + void zeroMemThread(size_t tid, size_t numThreads); + + /** + * value -= grad * learningRate, this is gradient. + * + * If L1 decay set use L1, else if L2 set use L2, otherwise no decay atall. + * + * t0 is a int vector used by L1/L2 decay, size = height of parameter + * matrix, + * store the time that each weight row last updated. + * + * Time is batchId, currentTime is current batchId. + * + * While pass finished, caller should call this func one more time + * with (fini=true) to let weight decay catch up current time. + */ + void sgdUpdate(BaseMatrix& value, + IVector& t0, + real learningRate, + int currentTime, + real decayRate, + bool useL1, + bool fini = false); + + /** + * merge rows in *this* to *dest* for designated thread + * + * values add to *dest* matrix + * + * ids occured in *this* append to *ids* + * filtered by (id % numThreads == tid) + */ + void addTo(BaseMatrix& dest, + std::vector& ids, + size_t tid, + size_t numThreads); + + /** + * the second version addTo(), *dest* is a SparseRowCpuMatrix. + * + * The dest's indices should be setup already, addTo() will + * check src ids is exist in dest's indices. + */ + void addTo(SparseRowCpuMatrix& dest, size_t tid, size_t numThreads); + + const IndexDictPtr& getIndexDictHandle() const { return indexDictHandle_; } + + /** + * check all local and global indices consistency + */ + void checkIndices(); + /** + * check whether row *i* exist in indices + */ + void checkIndex(size_t i) { + size_t localId = globalIndices_[i]; + CHECK_LT(localId, localIndices_->size()); + CHECK_EQ((*localIndices_)[localId], i); + } + + std::vector& getLocalIndices() const { + return indexDictHandle_->localIndices; + } + + protected: + template + void apply(Func f) { + f(buf_->data(), localIndices_->size() * width_); + } + + void init(size_t height, size_t width); + + /// clear row indices. + void clearRows() { + for (auto id : *localIndices_) { + globalIndices_[id] = kUnusedId_; + } + localIndices_->clear(); + buf_->clear(); + } + + inline void checkStoreSize() { + if (buf_->isAutoGrowth()) { + if (buf_->getRowCount() > 0.5 * height_) { + LOG(WARNING) << "There are more than 0.5*height (" + << localIndices_->size() << ") rows are used for sparse " + << "update, which is not efficient. Considering not use " + << "sparse_update."; + } + } else { + CHECK_LE(localIndices_->size(), buf_->getRowCount()); + } + } + + std::unique_ptr buf_; + IndexDictPtr indexDictHandle_; + std::vector* localIndices_; // =&indexDictHandle_->localIndices + unsigned int* globalIndices_; // =indexDictHandle_->globalIndices.data(); + static const unsigned int kUnusedId_; +}; + +class SyncThreadPool; + +/// For prefetching parameters from remote Parameter server +class SparsePrefetchRowCpuMatrix : public SparseRowCpuMatrix { + public: + SparsePrefetchRowCpuMatrix(CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + IndexDictPtr indexDictHandle = nullptr, + SyncThreadPool* pool = nullptr, + bool trans = false) + : SparseRowCpuMatrix(dataHandle, height, width, indexDictHandle, trans), + pool_(pool) {} + + /** + * Extract feature ids from *input*, to fill row indexs. + * + * *input* must be sparse matrix. + * + * Can call many times before setup. + */ + void addRows(MatrixPtr input); + void addRows(IVectorPtr ids); + + /** + * setup global indices of SparseRowMatrix after finish add rows. + */ + void setupIndices(); + + protected: + void addRows(const unsigned int* ids, size_t len); + SyncThreadPool* pool_; +}; + +class SparseAutoGrowRowCpuMatrix : public SparseRowCpuMatrix { + public: + SparseAutoGrowRowCpuMatrix(size_t height, + size_t width, + IndexDictPtr indexDictHandle = nullptr, + bool trans = false) + : SparseRowCpuMatrix(nullptr, height, width, indexDictHandle, trans) {} + + real* getRow(size_t row) { + auto id = globalIndices_[row]; + if (id == kUnusedId_) { + id = globalIndices_[row] = localIndices_->size(); + localIndices_->push_back(row); + checkStoreSize(); + } + return getLocalRow(id); + } + + virtual real* getRowBuf(size_t row) { return getRow(row); } + + virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); +}; + +class CacheRowCpuMatrix : public SparseAutoGrowRowCpuMatrix { + public: + CacheRowCpuMatrix(size_t height, + size_t width, + IndexDictPtr indexDictHandle = nullptr, + bool trans = false) + : SparseAutoGrowRowCpuMatrix(height, width, indexDictHandle, trans), + sourceData_(nullptr) {} + + void setSourceData(CpuVectorPtr sourceVec) { + sourceDataVec_ = sourceVec; + sourceData_ = sourceVec->getData(); + } + + real* getRow(size_t row) { + auto id = globalIndices_[row]; + if (id == kUnusedId_) { + id = globalIndices_[row] = localIndices_->size(); + localIndices_->push_back(row); + checkStoreSize(); + memcpy( + getLocalRow(id), sourceData_ + width_ * row, sizeof(float) * width_); + } + return getLocalRow(id); + } + + virtual real* getRowBuf(size_t row) { return getRow(row); } + + virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); + + public: + CpuVectorPtr sourceDataVec_; + real* sourceData_; +}; + +/** + * Sparse Row Ids Matrix. + * + * mostly same as CpuMatrix, but maintain sparse row ids occured, + * ids are hashed by worker thread id. + */ +class SparseRowIdsCpuMatrix : public CpuMatrix { + public: + SparseRowIdsCpuMatrix(CpuMemHandlePtr dataHandle, + size_t height, + size_t width, + bool trans = false) + : CpuMatrix(dataHandle, height, width, trans) {} + + void setNumOfThreads(size_t numOfThreads) { idsArray_.resize(numOfThreads); } + + std::vector& getIds(size_t threadId) { return idsArray_[threadId]; } + + private: + std::vector> idsArray_; +}; + +} // namespace paddle + +#else +namespace paddle { + +class SparseRowCpuMatrix : public CpuMatrix { + public: + void reserveStore() {} + void clearIndices() {} +}; + +class SparsePrefetchRowCpuMatrix : public SparseRowCpuMatrix { + public: + void setupIndices() {} + void addRows(MatrixPtr input) {} + void addRows(IVectorPtr ids) {} +}; + +class SparseAutoGrowRowCpuMatrix : public SparseRowCpuMatrix {}; +class CacheRowCpuMatrix : public SparseAutoGrowRowCpuMatrix {}; +class SparseRowIdsCpuMatrix : public CpuMatrix {}; + +} // namespace paddle + +#endif diff --git a/paddle/legacy/math/Storage.cpp b/paddle/legacy/math/Storage.cpp new file mode 100644 index 0000000000000000000000000000000000000000..65d53aeaa926690c7fe9e6fcac7affdfb68fede9 --- /dev/null +++ b/paddle/legacy/math/Storage.cpp @@ -0,0 +1,101 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Storage.h" +#include "Allocator.h" +#include "paddle/legacy/utils/StringUtil.h" +#include "paddle/legacy/utils/Util.h" + +#ifndef PADDLE_MOBILE_INFERENCE +DEFINE_int32(pool_limit_size, + 536870912, + "maximum memory size managed by a memory pool, default is 512M"); +#else +DEFINE_int32(pool_limit_size, 0, "default is 0"); +#endif + +namespace paddle { + +// Initialization StorageEngine singleton. +// Other modules may rely on storage management, +// so StorageEngine need to be initialized before other modules. +static InitFunction __init_storage_engine([]() { StorageEngine::singleton(); }, + std::numeric_limits::max()); + +StorageEngine::StorageEngine() : cpuAllocator_(nullptr) {} + +StorageEngine::~StorageEngine() { + delete cpuAllocator_; + for (auto it : gpuAllocator_) { + delete it; + } +} + +StorageEngine* StorageEngine::singleton() { + static StorageEngine storage; + return &storage; +} + +PoolAllocator* StorageEngine::getGpuAllocator(int deviceId) { + { + // if gpuAllocator_ has been constructed + ReadLockGuard guard(lock_); + if (deviceId < static_cast(gpuAllocator_.size()) && + (gpuAllocator_[deviceId] != nullptr)) { + return gpuAllocator_[deviceId]; + } + } + + { + // Construct gpuAllocator_ + std::lock_guard guard(lock_); + if (deviceId >= static_cast(gpuAllocator_.size())) { + gpuAllocator_.resize(deviceId + 1); + } + if (gpuAllocator_[deviceId] == nullptr) { + std::string name = + "gpu" + str::to_string(deviceId) + std::string("_pool"); + gpuAllocator_[deviceId] = + new PoolAllocator(new GpuAllocator(), FLAGS_pool_limit_size, name); + } + return gpuAllocator_[deviceId]; + } +} + +PoolAllocator* StorageEngine::getCpuAllocator() { + { + // if cpuAllocator_ has been constructed + ReadLockGuard guard(lock_); + if (cpuAllocator_ != nullptr) { + return cpuAllocator_; + } + } + + { + // Construct cpuAllocator_ + std::lock_guard guard(lock_); + if (cpuAllocator_ == nullptr) { + if (FLAGS_use_gpu) { + cpuAllocator_ = new PoolAllocator( + new CudaHostAllocator(), FLAGS_pool_limit_size, "cuda_host_pool"); + } else { + cpuAllocator_ = new PoolAllocator( + new CpuAllocator(), FLAGS_pool_limit_size, "cpu_pool"); + } + } + return cpuAllocator_; + } +} + +} // namespace paddle diff --git a/paddle/legacy/math/Storage.h b/paddle/legacy/math/Storage.h new file mode 100644 index 0000000000000000000000000000000000000000..bd22dde2c85be5ba432cb3a259211c1900a17b6c --- /dev/null +++ b/paddle/legacy/math/Storage.h @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "PoolAllocator.h" +#include "paddle/legacy/utils/Locks.h" + +namespace paddle { + +/** + * @brief Storage manager for multiple devices. + */ +class StorageEngine { + public: + /** + * @return Storage singleton + */ + static StorageEngine* singleton(); + + /** + * @return return one gpu allocator by deviceId + */ + PoolAllocator* getGpuAllocator(int deviceId); + + /** + * @return return cpu allocator + */ + PoolAllocator* getCpuAllocator(); + + protected: + StorageEngine(); + ~StorageEngine(); + RWLock lock_; + std::vector gpuAllocator_; + PoolAllocator* cpuAllocator_; +}; + +} // namespace paddle diff --git a/paddle/math/TensorApply.h b/paddle/legacy/math/TensorApply.h similarity index 100% rename from paddle/math/TensorApply.h rename to paddle/legacy/math/TensorApply.h diff --git a/paddle/legacy/math/TensorAssign.h b/paddle/legacy/math/TensorAssign.h new file mode 100644 index 0000000000000000000000000000000000000000..efbfce6c4f88197f18285e3679698b8bbb1ed3b8 --- /dev/null +++ b/paddle/legacy/math/TensorAssign.h @@ -0,0 +1,158 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +/** + * \brief Tensor Assign Expression(return by lazyAssign, + * and evaluated by AssignEvaluate) + */ +template +class TensorAssignOp { + public: + explicit TensorAssignOp(const LhsType& lhs, const RhsType& rhs) + : lhs_(lhs), rhs_(rhs) { +#ifndef __CUDA_ARCH__ + CHECK_EQ(lhs_.getWidth(), rhs_.getWidth()); + CHECK_EQ(lhs_.getHeight(), rhs_.getHeight()); + CHECK_EQ(lhs_.useGpu(), rhs_.useGpu()); +#endif + } + + INLINE void apply(const int i, const int j) { + lhs_.applyRef(i, j) = rhs_.apply(i, j); + } + INLINE void apply(const int index) { + lhs_.applyRef(index) = rhs_.apply(index); + } + + INLINE size_t getWidth() const { return lhs_.getWidth(); } + INLINE size_t getHeight() const { return rhs_.getHeight(); } + INLINE bool isContiguous() const { + return lhs_.isContiguous() && rhs_.isContiguous(); + } + INLINE bool useGpu() const { return lhs_.useGpu(); } + + private: + TensorApply lhs_; + TensorApply rhs_; +}; + +template +void AssignCpuEvaluate(int height, + int width, + bool isContiguous, + Assign&& assign, + AssignOp&&... args) { + if (isContiguous) { + int size = height * width; + for (int index = 0; index < size; index++) { + assign.apply(index); + __attribute__((unused)) int dummy[] = {(((args)).apply(index), 0)...}; + } + } else { + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + assign.apply(i, j); + __attribute__((unused)) int dummy[] = {(((args)).apply(i, j), 0)...}; + } + } + } +} + +#ifdef __NVCC__ +template +__global__ void AssignGpuEvaluate1(const int border, + Assign assign, + AssignOp... args) { + const int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < border) { + assign.apply(idx); + __attribute__((unused)) int dummy[] = {(((args)).apply(idx), 0)...}; + } +} + +template +__global__ void AssignGpuEvaluate2(const int height, + const int width, + Assign assign, + AssignOp... args) { + const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; + const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; + for (int i = rowIdx; i < height; i += gridDim.y * blockDim.y) { + for (int j = colIdx; j < width; j += gridDim.x * blockDim.x) { + assign.apply(i, j); + __attribute__((unused)) int dummy[] = {(((args)).apply(i, j), 0)...}; + } + } +} +#endif + +/** + * \brief Evaluate one or more TensorAssignOp objects. + * + * \note At least one assignment expression is required + */ +template +void AssignEvaluate(Assign&& assign, AssignOp&&... args) { + const bool useGpu_ = assign.useGpu(); + bool isContiguous_ = assign.isContiguous(); + const size_t height = assign.getHeight(); + const size_t width = assign.getWidth(); + + const int packSize = sizeof...(args); + const bool packUseGpu[] = {((args)).useGpu()...}; + const bool packIsContiguous[] = {((args)).isContiguous()...}; + const size_t packHeight[] = {((args)).getHeight()...}; + const size_t packWidth[] = {((args)).getWidth()...}; + + for (int i = 0; i < packSize; i++) { + CHECK_EQ(useGpu_, packUseGpu[i]); + CHECK_EQ(height, packHeight[i]); + CHECK_EQ(width, packWidth[i]); + isContiguous_ = isContiguous_ && packIsContiguous[i]; + } + + if (useGpu_) { +#ifdef __NVCC__ + if (isContiguous_) { + int size = height * width; + int blockSize = size <= 1024 ? size : 1024; + int gridSize = (size + 1024 - 1) / 1024; + AssignGpuEvaluate1<<>>( + size, assign, args...); + } else { + int blockSizeY = std::min(32, (int)height); + int blockSizeX = (32 / blockSizeY) * 32; + int gridSizeX = std::min(32, (int)(width + blockSizeX - 1) / blockSizeX); + int gridSizeY = std::min(32, (int)(height + blockSizeY - 1) / blockSizeY); + dim3 threads(blockSizeX, blockSizeY); + dim3 grid(gridSizeX, gridSizeY); + AssignGpuEvaluate2<<>>( + height, width, assign, args...); + } + + CHECK_SYNC("AssignEvaluate failed"); +#endif + } else { + AssignCpuEvaluate(height, width, isContiguous_, assign, args...); + } +} + +} // namespace paddle diff --git a/paddle/legacy/math/TensorEvaluate.h b/paddle/legacy/math/TensorEvaluate.h new file mode 100644 index 0000000000000000000000000000000000000000..3029dd35fb05c893f99cde0689f816f4257f21c4 --- /dev/null +++ b/paddle/legacy/math/TensorEvaluate.h @@ -0,0 +1,112 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "hl_base.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +/** + * \brief The tensor cpu evaluate api. + */ +template +inline void TensorCpuApply(LeftType& lhs, const RightType& rhs) { + TensorApply lhs_(lhs); + TensorApply rhs_(rhs); + CHECK_EQ(lhs_.getWidth(), rhs_.getWidth()); + CHECK_EQ(lhs_.getHeight(), rhs_.getHeight()); + CHECK_EQ(lhs_.useGpu(), rhs_.useGpu()); + + int height = lhs_.getHeight(); + int width = lhs_.getWidth(); + if (lhs_.isContiguous() && rhs_.isContiguous()) { + int size = height * width; + for (int index = 0; index < size; index++) { + lhs_.applyRef(index) = rhs_.apply(index); + } + } else { + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + lhs_.applyRef(i, j) = rhs_.apply(i, j); + } + } + } +} + +#ifdef __NVCC__ +template +__global__ void TensorElementWiseOp(LeftType lhs, + RightType rhs, + const int border) { + const int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < border) { + lhs.applyRef(idx) = rhs.apply(idx); + } +} + +template +__global__ void TensorElementWiseOp(LeftType lhs, RightType rhs) { + const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; + const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; + for (int i = rowIdx; i < lhs.getHeight(); i += gridDim.y * blockDim.y) { + for (int j = colIdx; j < lhs.getWidth(); j += gridDim.x * blockDim.x) { + lhs.applyRef(i, j) = rhs.apply(i, j); + } + } +} + +/** + * \brief The tensor gpu evaluate api. + */ +template +inline void TensorGpuApply(LeftType& lhs, const RightType& rhs) { + TensorApply lhs_(lhs); + TensorApply rhs_(rhs); + CHECK_EQ(lhs_.getWidth(), rhs_.getWidth()); + CHECK_EQ(lhs_.getHeight(), rhs_.getHeight()); + CHECK_EQ(lhs_.useGpu(), rhs_.useGpu()); + + int dimM = lhs_.getHeight(); + int dimN = lhs_.getWidth(); + + if (lhs_.isContiguous() && rhs_.isContiguous()) { + int size = dimM * dimN; + int blockSize = size <= 1024 ? size : 1024; + int gridSize = (size + 1024 - 1) / 1024; + TensorElementWiseOp<<>>( + lhs_, rhs_, size); + } else { + int blockSizeY = std::min(32, dimM); + int blockSizeX = (32 / blockSizeY) * 32; + int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); + int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); + dim3 threads(blockSizeX, blockSizeY); + dim3 grid(gridSizeX, gridSizeY); + TensorElementWiseOp<<>>(lhs_, rhs_); + } + + CHECK_SYNC("TensorGpuApply failed"); +} +#else +template +inline void TensorGpuApply(LeftType& lhs, RightType& rhs) { + LOG(FATAL) << "Since it is gcc compiled, " + "this calculation does not support GPU implementation."; +} +#endif + +} // namespace paddle diff --git a/paddle/legacy/math/TensorExpression.h b/paddle/legacy/math/TensorExpression.h new file mode 100644 index 0000000000000000000000000000000000000000..1c6cf07831487165445a3f59931c4ca9196375b9 --- /dev/null +++ b/paddle/legacy/math/TensorExpression.h @@ -0,0 +1,446 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "hl_tensor_ops.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +template +class TensorConstant; +template +class TensorUnaryOp; +template +class TensorBinaryOp; +template +class TensorTernaryOp; + +template +class TensorAssignOp; + +/** + * \brief Tensor base class. + * + * This is the base class of all Tensor and Expression class. + */ +template +class TensorExpression { + public: + /** + * Element wise unary expression. + */ + template + const TensorUnaryOp unaryExpression( + const UnaryOp& op) const { + return TensorUnaryOp(op, derived()); + } + + const TensorUnaryOp, const Derived, T> operator+( + T p) const { + return unaryExpression(hppl::unary::add_scale(p)); + } + + const TensorUnaryOp, const Derived, T> operator-( + T p) const { + return unaryExpression(hppl::unary::sub_scale(p)); + } + + const TensorUnaryOp, const Derived, T> operator*( + T p) const { + return unaryExpression(hppl::unary::mul_scale(p)); + } + + const TensorUnaryOp, const Derived, T> operator/( + T p) const { + return unaryExpression(hppl::unary::div_scale(p)); + } + + const TensorUnaryOp, const Derived, T> operator-() const { + return unaryExpression(hppl::unary::neg()); + } + + const TensorUnaryOp, const Derived, T> exp() const { + return unaryExpression(hppl::unary::exp_op()); + } + + const TensorUnaryOp, const Derived, T> log() const { + return unaryExpression(hppl::unary::log_op()); + } + + const TensorUnaryOp, const Derived, T> sqrt() const { + return unaryExpression(hppl::unary::sqrt_op()); + } + + const TensorUnaryOp, const Derived, T> square() const { + return unaryExpression(hppl::unary::square()); + } + + const TensorUnaryOp, const Derived, T> reciprocal() + const { + return unaryExpression(hppl::unary::reciprocal()); + } + + const TensorUnaryOp, const Derived, T> abs() const { + return unaryExpression(hppl::unary::abs()); + } + + const TensorUnaryOp, const Derived, T> sign() const { + return unaryExpression(hppl::unary::sign()); + } + + const TensorUnaryOp, const Derived, T> pow(T p) const { + return unaryExpression(hppl::unary::pow_op(p)); + } + + const TensorUnaryOp, const Derived, T> min(T p) const { + return unaryExpression(hppl::unary::min(p)); + } + + const TensorUnaryOp, const Derived, T> max(T p) const { + return unaryExpression(hppl::unary::max(p)); + } + + const TensorUnaryOp, const Derived, T> operator==( + T p) const { + return unaryExpression(hppl::unary::cmp_eq(p)); + } + + const TensorUnaryOp, const Derived, T> operator!=( + T p) const { + return unaryExpression(hppl::unary::cmp_ne(p)); + } + + const TensorUnaryOp, const Derived, T> operator<=( + T p) const { + return unaryExpression(hppl::unary::cmp_le(p)); + } + + const TensorUnaryOp, const Derived, T> operator<( + T p) const { + return unaryExpression(hppl::unary::cmp_lt(p)); + } + + const TensorUnaryOp, const Derived, T> operator>=( + T p) const { + return unaryExpression(hppl::unary::cmp_ge(p)); + } + + const TensorUnaryOp, const Derived, T> operator>( + T p) const { + return unaryExpression(hppl::unary::cmp_gt(p)); + } + + const TensorUnaryOp, const Derived, T> operator&&( + T p) const { + return unaryExpression(hppl::unary::and_op(p)); + } + + const TensorUnaryOp, const Derived, T> operator||( + T p) const { + return unaryExpression(hppl::unary::or_op(p)); + } + + /** + * Element wise binary expression. + */ + template + const TensorBinaryOp + binaryExpression(const BinaryOp& op, const ExpressionType& expr) const { + return TensorBinaryOp( + op, derived(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator==(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::cmp_eq(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator!=(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::cmp_ne(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator<=(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::cmp_le(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator<(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::cmp_lt(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator>=(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::cmp_ge(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator>(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::cmp_gt(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator&&(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::and_op(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator||(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::or_op(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator+(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::add(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator-(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::sub(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator*(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::mul(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + operator/(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::div(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + min(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::min(), expr); + } + + template + const TensorBinaryOp, + const Derived, + const ExpressionType, + T> + max(const ExpressionType& expr) const { + return binaryExpression(hppl::binary::max(), expr); + } + + /** + * Element wise ternary expression. + * + * ternary conditional operator(?: operator). + * The conditional expression returns one of two values depending on + * the result of derived expression. + * If derived expression evaluates to true, then expression1 is evaluated. + * If derived expression evaluates to false, then expression2 is evaluated. + */ + template + const TensorTernaryOp + condition(const ExprType1& expr1, const ExprType2& expr2) const { + return TensorTernaryOp( + derived(), expr1, expr2); + } + + template + const TensorTernaryOp< + const Derived, + const TensorConstant, const Derived, T>, + const ExprType, + T> + condition(T p, const ExprType& expr) const { + return condition(constant(p), expr); + } + + template + const TensorTernaryOp< + const Derived, + const ExprType, + const TensorConstant, const Derived, T>, + T> + condition(const ExprType& expr, T p) const { + return condition(expr, constant(p)); + } + + const TensorTernaryOp< + const Derived, + const TensorConstant, const Derived, T>, + const TensorConstant, const Derived, T>, + T> + condition(T p1, T p2) const { + return condition(constant(p1), constant(p2)); + } + + /** + * return a TensorConstant. A TensorConstant object hold a constant value. + */ + const TensorConstant, const Derived, T> constant( + T p) const { + return TensorConstant, const Derived, T>( + hppl::unary::constant(p), derived()); + } + + /** + * return a TensorAssignOp, and use AssignEvaluate to evaluate one or more + * TensorAssignOp objects. + */ + template + TensorAssignOp lazyAssign( + const ExpressionType& expr) const { + return TensorAssignOp(derived(), expr); + } + + protected: + const Derived& derived() const { return *static_cast(this); } +}; + +/** + * \brief Unary Operator Expression + */ +template +class TensorUnaryOp + : public TensorExpression, T> { + public: + explicit TensorUnaryOp(const OP op, const ExprType& expr) + : op_(op), expr_(expr) {} + + const OP op_; + const ExprType expr_; +}; + +/** + * \brief Binary Operator Expression + */ +template +class TensorBinaryOp + : public TensorExpression, T> { + public: + explicit TensorBinaryOp(const OP op, const LhsType& lhs, const RhsType& rhs) + : op_(op), lhs_(lhs), rhs_(rhs) {} + + const OP op_; + const LhsType lhs_; + const RhsType rhs_; +}; + +/** + * \brief Ternary Operator Expression + */ +template +class TensorTernaryOp : public TensorExpression< + TensorTernaryOp, + T> { + public: + explicit TensorTernaryOp(const ExprType1& expr1, + const ExprType2& expr2, + const ExprType3& expr3) + : expr1_(expr1), expr2_(expr2), expr3_(expr3) {} + + const ExprType1 expr1_; + const ExprType2 expr2_; + const ExprType3 expr3_; +}; + +/** + * \brief Constant Expression + */ +template +class TensorConstant + : public TensorExpression, T> { + public: + explicit TensorConstant(const OP op, const ExprType& expr) + : op_(op), expr_(expr) {} + + const OP op_; + const ExprType expr_; +}; + +/** + * \brief operator+ overload + * \return a unary operator expression + */ +template +const TensorUnaryOp, const Derived, T> operator+( + T p, const TensorExpression& expr) { + return expr + p; +} + +/** + * \brief operator* overload + * \return a unary operator expression + */ +template +const TensorUnaryOp, const Derived, T> operator*( + T p, const TensorExpression& expr) { + return expr * p; +} + +} // namespace paddle + +#include "TensorApply.h" +#include "TensorEvaluate.h" diff --git a/paddle/legacy/math/TrainingAlgorithmOp.cu b/paddle/legacy/math/TrainingAlgorithmOp.cu new file mode 100644 index 0000000000000000000000000000000000000000..9e1eaa0f45ae94d12cf7763bbaff632fc473bcc8 --- /dev/null +++ b/paddle/legacy/math/TrainingAlgorithmOp.cu @@ -0,0 +1,356 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "BaseMatrix.h" +#include "TrainingAlgorithmOp.h" +#include "paddle/legacy/utils/Logging.h" + +#if __cplusplus > 199711L + +#include "TensorAssign.h" + +namespace paddle { + +void sparseMomentumApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& momU, + BaseMatrix& momV, + real alpha, + real beta, + real gamma, + real tau, + real learningRate) { + auto expr1 = momU.lazyAssign(momU - (alpha * gamma * learningRate) * grad); + auto expr2 = + momV.lazyAssign(momV + (tau * alpha * gamma * learningRate) * grad); + auto expr3 = value.lazyAssign((tau / beta + (real)1 / alpha) * momU + + ((real)1 / beta) * momV); + + AssignEvaluate(expr1, expr2, expr3); +} + +void adadeltaApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum, + BaseMatrix& accum_update, + BaseMatrix& lr, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate) { + auto expr1 = accum.lazyAssign(rou * accum + ((real)1 - rou) * grad.square()); + auto expr2 = + lr.lazyAssign(((accum_update + epsilon) / (accum + epsilon)).sqrt()); + auto expr3 = accum_update.lazyAssign(rou * accum_update + + ((real)1 - rou) * (grad * lr).square()); + auto expr4 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); + auto expr5 = value.lazyAssign(value + mom); + + AssignEvaluate(expr1, expr2, expr3, expr4, expr5); +} + +void adagradApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum_buffer, + BaseMatrix& accum, + BaseMatrix& lr, + real epsilon, + real learningRate, + real momentum, + real decayRate) { + auto expr1 = accum.lazyAssign(accum + grad.square()); + auto expr2 = + lr.lazyAssign((accum_buffer + accum + epsilon).sqrt().reciprocal()); + auto expr3 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); + auto expr4 = value.lazyAssign(value + mom); + + AssignEvaluate(expr1, expr2, expr3, expr4); +} + +void rmspropApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& g, + BaseMatrix& f, + BaseMatrix& lr, + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime) { + auto expr2 = f.lazyAssign(accumulatedRou * f + ((real)1 - rou) * grad); + auto expr3 = lr.lazyAssign((g - f.square() + epsilon).sqrt().reciprocal()); + auto expr4 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); + auto expr5 = value.lazyAssign(value + mom); + + if (firstTime) { + auto expr1 = g.lazyAssign(accumulatedRou * g + grad.square()); + + AssignEvaluate(expr1, expr2, expr3, expr4, expr5); + } else { + auto expr1 = + g.lazyAssign(accumulatedRou * g + ((real)1 - rou) * grad.square()); + + AssignEvaluate(expr1, expr2, expr3, expr4, expr5); + } +} + +void decayedAdagradApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum, + BaseMatrix& lr, + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime) { + auto expr2 = lr.lazyAssign((accum + epsilon).sqrt().reciprocal()); + auto expr3 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); + auto expr4 = value.lazyAssign(value + mom); + + if (firstTime) { + auto expr1 = accum.lazyAssign(accumulatedRou * accum + grad.square()); + + AssignEvaluate(expr1, expr2, expr3, expr4); + } else { + auto expr1 = accum.lazyAssign(accumulatedRou * accum + + ((real)1 - rou) * grad.square()); + + AssignEvaluate(expr1, expr2, expr3, expr4); + } +} + +void adamApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, // firse moment + BaseMatrix& v, // second moment + real beta1, + real beta2, + real beta1_power, + real beta2_power, + real epsilon, + real learningRate) { + real alpha = + learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); + + auto expr1 = mom.lazyAssign(beta1 * mom + ((real)1 - beta1) * grad); + auto expr2 = v.lazyAssign(beta2 * v + ((real)1 - beta2) * grad.square()); + auto expr3 = value.lazyAssign(value - (mom * alpha) / (v.sqrt() + epsilon)); + + AssignEvaluate(expr1, expr2, expr3); +} + +void adamaxApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, // firse moment + BaseMatrix& u, // weighted infinity norm + real beta1, + real beta2, + int64_t step, + real alpha) { + auto expr1 = mom.lazyAssign(beta1 * mom + ((real)1 - beta1) * grad); + auto expr2 = + u.lazyAssign((beta2 * u > grad.abs()).condition(beta2 * u, grad.abs())); + auto expr3 = value.lazyAssign( + value - (alpha / ((real)1 - (real)std::pow(beta1, step))) * (mom / u)); + + AssignEvaluate(expr1, expr2, expr3); +} + +} // namespace paddle + +#else + +namespace paddle { + +void sparseMomentumApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& momU, + BaseMatrix& momV, + real alpha, + real beta, + real gamma, + real tau, + real learningRate) { + /** + * \alpha_t = \alpha_{t-1} / k + * \beta_t = \beta_{t-1} / (1 + \lambda\gamma_t) + * u_t = u_{t-1} - \alpha_t \gamma_t g_t + * v_t = v_{t-1} + \tau_{t-1} \alpha_t \gamma_t g_t + * \tau_t = \tau_{t-1} + \beta_t / \alpha_t + */ + momU -= (alpha * gamma * learningRate) * grad; + momV += (tau * alpha * gamma * learningRate) * grad; + value = (tau / beta + (real)1 / alpha) * momU + ((real)1 / beta) * momV; +} + +void adadeltaApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum, + BaseMatrix& accum_update, + BaseMatrix& lr, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate) { + // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 + accum = rou * accum + ((real)1 - rou) * grad.square(); + + // learn_rate: sqrt(( E(dx_{t-1}^2) + epsilon ) / ( E(g_t^2) + epsilon )) + lr = ((accum_update + epsilon) / (accum + epsilon)).sqrt(); + + // E(dx_t^2) = \rou * E(dx_{t-1}^2) + (1-\rou) * (-g*learn_rate)^2 + accum_update = rou * accum_update + ((real)1 - rou) * (grad * lr).square(); + + mom = mom * momentum - learningRate * lr * (grad + value * decayRate); + value += mom; +} + +void adagradApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum_buffer, + BaseMatrix& accum, + BaseMatrix& lr, + real epsilon, + real learningRate, + real momentum, + real decayRate) { + accum += grad.square(); + lr = (accum_buffer + accum + epsilon).sqrt().reciprocal(); + mom = mom * momentum - learningRate * lr * (grad + value * decayRate); + value += mom; +} + +void rmspropApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& g, + BaseMatrix& f, + BaseMatrix& lr, + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime) { + // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 + // For the first time update, make the sum be the current square + // so that the initial estimation of E(g_t^2) will not be too small. + if (firstTime) { + g = accumulatedRou * g + grad.square(); + } else { + g = accumulatedRou * g + ((real)1 - rou) * grad.square(); + } + + // E(f_t) = \rou * E(f_{t-1}) + (1-\rou) * g + f = accumulatedRou * f + ((real)1 - rou) * grad; + + // learn_rate = 1/sqrt( ( E(g_t^2) - (E(f_t))^2 + epsilon ) + // Basiclly if the sign of the gradient changes more often, + // the learning rate will be decreased. + lr = (g - f.square() + epsilon).sqrt().reciprocal(); + + mom = mom * momentum - learningRate * lr * (grad + value * decayRate); + value += mom; +} + +void decayedAdagradApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum, + BaseMatrix& lr, + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime) { + // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 + // For the first time update, make the sum be the current square + // so that the initial estimation of E(g_t^2) will not be too small. + if (firstTime) { + accum = accumulatedRou * accum + grad.square(); + } else { + accum = accumulatedRou * accum + ((real)1 - rou) * grad.square(); + } + + // learn_rate = 1/sqrt( ( E(g_t^2) + epsilon ) + // Basiclly if the bigger the magnitude gradient is, + // the smaller the learning rate will be. + lr = (accum + epsilon).sqrt().reciprocal(); + + mom = mom * momentum - learningRate * lr * (grad + value * decayRate); + value += mom; +} + +void adamApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, // firse moment + BaseMatrix& v, // second moment + real beta1, + real beta2, + real beta1_power, + real beta2_power, + real epsilon, + real learningRate) { + real alpha = + learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); + + // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; + mom = beta1 * mom + ((real)1 - beta1) * grad; + + // v_t = \beta_2 * v_{t-1} + (1-\beta_2)* g_{t-1}^2 + v = beta2 * v + ((real)1 - beta2) * grad.square(); + + value -= (mom * alpha) / (v.sqrt() + epsilon); +} + +void adamaxApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, // firse moment + BaseMatrix& u, // weighted infinity norm + real beta1, + real beta2, + int64_t step, + real alpha) { + // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; + mom = beta1 * mom + ((real)1 - beta1) * grad; + + // u_t = max(\beta_2*u_{t-1}, abs(g_t)) + u = (beta2 * u > grad.abs()).condition(beta2 * u, grad.abs()); + + // \theta_t = \theta_{t-1} - (\alpha/(1-\beta_1^t))*m_t/u_t + value -= (alpha / ((real)1 - (real)std::pow(beta1, step))) * (mom / u); +} + +} // namespace paddle + +#endif diff --git a/paddle/legacy/math/TrainingAlgorithmOp.h b/paddle/legacy/math/TrainingAlgorithmOp.h new file mode 100644 index 0000000000000000000000000000000000000000..921c2742cfe2576785768da40ab11c94234be966 --- /dev/null +++ b/paddle/legacy/math/TrainingAlgorithmOp.h @@ -0,0 +1,122 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "BaseMatrix.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +/** + * \brief Sparse Momentum optimizer. + */ +extern void sparseMomentumApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& momU, + BaseMatrix& momV, + real alpha, + real beta, + real gamma, + real tau, + real learningRate); + +/** + * \brief AdaDelta optimizer. + */ +extern void adadeltaApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& sum, + BaseMatrix& sum1, + BaseMatrix& mom, + BaseMatrix& lr, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate); + +/** + * \brief AdaGrad optimizer. + */ +extern void adagradApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& sum, + BaseMatrix& sum1, + BaseMatrix& mom, + BaseMatrix& lr, + real epsilon, + real learningRate, + real momentum, + real decayRate); + +/** + * \brief RMSProp optimizer. + */ +extern void rmspropApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& g, + BaseMatrix& f, + BaseMatrix& mom, + BaseMatrix& lr, + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime); + +/** + * \brief Decayed AdaGrad optimizer. + */ +extern void decayedAdagradApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& accum, + BaseMatrix& lr, + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime); + +/** + * \brief Adam optimizer. + */ +extern void adamApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, + BaseMatrix& v, + real beta1, + real beta2, + real beta1_power, + real beta2_power, + real epsilon, + real learningRate); + +/** + * \brief AdaMax optimizer. + */ +extern void adamaxApply(BaseMatrix& value, + BaseMatrix& grad, + BaseMatrix& mom, // firse moment + BaseMatrix& u, // weighted infinity norm + real beta1, + real beta2, + int64_t step, + real alpha); +} // namespace paddle diff --git a/paddle/legacy/math/Vector.cpp b/paddle/legacy/math/Vector.cpp new file mode 100644 index 0000000000000000000000000000000000000000..87f48bb1622f28f8cb53e5afc924f5cadb14c528 --- /dev/null +++ b/paddle/legacy/math/Vector.cpp @@ -0,0 +1,1091 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Vector.h" +#include "paddle/legacy/utils/Util.h" + +#include +#include "Matrix.h" +#include "hl_gpu.h" +#include "hl_matrix.h" +#include "hl_table_apply.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Thread.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +namespace paddle { + +template +std::shared_ptr> VectorT::create(size_t size, bool useGpu) { + if (useGpu) { + return std::make_shared>(size); + } else { + return std::make_shared>(size); + } +} + +template +std::shared_ptr> VectorT::createParallelVector( + size_t size, bool useGpu, SyncThreadPool* pool) { + if (!useGpu && FLAGS_trainer_count > 1 && FLAGS_enable_parallel_vector && + size >= (size_t)FLAGS_enable_parallel_vector) { + return std::make_shared>( + size, pool ? pool : getGlobalSyncThreadPool()); + } else { + return create(size, useGpu); + } +} + +template +std::shared_ptr> VectorT::create(T* data, + size_t size, + bool useGpu) { + if (useGpu) { + return std::make_shared>(size, data); + } else { + return std::make_shared>(size, data); + } +} + +template +std::shared_ptr> VectorT::create(size_t size, + MemoryHandlePtr memoryHandle, + size_t offset) { + if (auto cpuMemHandle = + std::dynamic_pointer_cast(memoryHandle)) { + return std::make_shared>(size, cpuMemHandle, offset); + } else if (auto gpuMemHandle = + std::dynamic_pointer_cast(memoryHandle)) { + return std::make_shared>(size, gpuMemHandle, offset); + } else { + LOG(FATAL) << "Wrong"; + return NULL; + } +} + +template <> +MatrixPtr VectorT::toOneHotSparseMatrix(size_t idRange, bool useGpu) { + LOG(FATAL) << "Wrong for real vector"; + return nullptr; +} + +template <> +MatrixPtr VectorT::toOneHotSparseMatrix(size_t idRange, bool useGpu) { + size_t height = getSize(); + size_t width = idRange; + MatrixPtr mat = Matrix::createSparseMatrix( + height, idRange, height, NO_VALUE, SPARSE_CSR, false, useGpu); + + CpuIVector cpuIds(height); + cpuIds.copyFrom(*this); + int* idData = cpuIds.getData(); + + for (decltype(height) i = 0; i < height; i++) { + const unsigned int id = idData[i]; + CHECK_LT(id, width); + mat->setRow(i, 1, &id, nullptr); + } + return mat; +} + +template <> +std::shared_ptr> VectorT::castToInt() { + std::shared_ptr> ret = IVector::create(this->getSize(), useGpu_); + if (useGpu_) { + hl_vector_cast2int(ret->getData(), this->getData(), this->getSize()); + } else { + for (size_t i = 0; i < getSize(); ++i) { + ret->getData()[i] = int(this->getData()[i]); + } + } + return ret; +} + +template +GpuVectorT::GpuVectorT(size_t size) + : VectorT(size, + std::make_shared(sizeof(T) * size), + 0, /* offset = 0 */ + true /* useGpu = true */) {} + +template +T GpuVectorT::getElement(size_t i) const { + T elem = 0; + hl_memcpy_device2host(&elem, const_cast(&this->getData()[i]), sizeof(T)); + return elem; +} +template +void GpuVectorT::setElement(size_t i, const T& value) { + hl_memcpy_host2device(&this->getData()[i], const_cast(&value), sizeof(T)); +} + +template +T* GpuVectorT::getPoint(const uint64_t beginPos) { + LOG(FATAL) << "Not implemented" << beginPos; + return NULL; +} + +template <> +int GpuVectorT::getAbsSum() { + LOG(FATAL) << "Not implemented"; + return 0; +} + +template <> +int GpuVectorT::getSum() { + LOG(FATAL) << "Not implemented"; + return 0; +} + +template <> +real GpuVectorT::getAbsSum() { + real* A = this->getData(); + real sum = 0; + hl_vector_abs_sum(A, &sum, this->getSize()); + return sum; +} + +template <> +real GpuVectorT::getSum() { + real* A = this->getData(); + real sum = 0; + hl_vector_sum(A, &sum, this->getSize()); + return sum; +} + +template <> +int GpuVectorT::getMax() { + CpuIVector cpuIVec = CpuIVector(this->getSize()); + copyTo(&cpuIVec); + return cpuIVec.getMax(); +} + +template <> +int GpuVectorT::getAbsMax() { + CpuIVector cpuIVec = CpuIVector(this->getSize()); + copyTo(&cpuIVec); + return cpuIVec.getAbsMax(); +} + +template +void GpuVectorT::isEqualTo(const VectorT& b, const T& value) { + BaseMatrixT::isEqualTo((BaseMatrixT&)b, value); +} + +template +void GpuVectorT::selectFrom(const VectorT& src, const VectorT& ids) { +#ifdef PADDLE_WITH_CUDA + hl_vector_select_from(this->getData(), + this->getSize(), + src.getData(), + src.getSize(), + ids.getData(), + ids.getSize()); +#endif +} + +template +real gpuRowFunc(Func f, GpuVector& v) { + static ThreadLocal>> local; + if (!*local) { + (*local).reset(new CpuVector(1)); + } + real* A = v.getData(); + f(A, (*local)->getData(), 1, v.getSize()); + return (*local)->getData()[0]; +} + +template <> +real GpuVectorT::getMax() { + return gpuRowFunc(hl_matrix_row_max, *this); +} + +template <> +real GpuVectorT::getAbsMax() { + return std::max(gpuRowFunc(hl_matrix_row_max, *this), + -gpuRowFunc(hl_matrix_row_min, *this)); +} + +template <> +int GpuVectorT::getMin() { + LOG(FATAL) << "Not implemented"; + return 0; +} + +template <> +real GpuVectorT::getMin() { + return gpuRowFunc(hl_matrix_row_min, *this); +} + +template +T GpuVectorT::get(size_t pos) { + T val = (T)0; + hl_memcpy_device2host((void*)&val, (void*)(this->getData() + pos), sizeof(T)); + return val; +} + +template +void GpuVectorT::histogram(std::ostream& os, int type) { + LOG(FATAL) << "Not implemented"; +} + +template +void GpuVectorT::zeroMem() { + BaseMatrixT::zero(); +} + +template +void GpuVectorT::reset(const T& value) { + BaseMatrixT::assign(value); +} + +template +void GpuVectorT::fillSequence() { + LOG(FATAL) << "not implemented"; +} + +template +void GpuVectorT::copyFrom(const VectorT& src) { + src.copyTo(this); +} + +template +void GpuVectorT::copyFrom(const VectorT& src, hl_stream_t stream) { + CHECK_EQ(src.getSize(), this->getSize()); + hl_memcpy_async((void*)this->getData(), + (void*)src.getData(), + sizeof(T) * this->getSize(), + stream); +} + +template +void GpuVectorT::copyFrom(const T* gpuSrc, size_t size) { + CHECK(gpuSrc != NULL); + CHECK_LE(size, this->size_); + + hl_memcpy((void*)this->getData(), (void*)gpuSrc, sizeof(T) * size); +} + +template +void GpuVectorT::copyFrom(const T* gpuSrc, size_t size, hl_stream_t stream) { + CHECK(gpuSrc != NULL); + CHECK_LE(size, this->size_); + + hl_memcpy_async( + (void*)this->getData(), (void*)gpuSrc, sizeof(T) * size, stream); +} + +template +void GpuVectorT::copyTo(CpuVectorT* dest) const { + CHECK_EQ(this->getSize(), dest->getSize()); + + hl_memcpy_device2host((void*)dest->getData(), + (void*)this->getData(), + sizeof(T) * this->getSize()); +} + +template +void GpuVectorT::copyTo(GpuVectorT* dest) const { + CHECK_EQ(this->getSize(), dest->getSize()); + + hl_memcpy_device2device((void*)dest->getData(), + (void*)this->getData(), + sizeof(T) * this->getSize()); +} + +template <> +void GpuVectorT::rand() { + LOG(FATAL) << "Not implemented"; +} + +template <> +void GpuVectorT::print(std::ostream& os, size_t num) const { + IVectorPtr dest = IVector::create(this->size_, false); + hl_memcpy_device2host((void*)dest->getData(), + (void*)this->getData(), + sizeof(int) * this->getSize()); + dest->print(os, num); +} + +template <> +void GpuVectorT::print(std::ostream& os, size_t num) const { + VectorPtr dest = Vector::create(this->size_, false); + hl_memcpy_device2host((void*)dest->getData(), + (void*)this->getData(), + sizeof(int) * this->getSize()); + dest->print(os, num); +} + +template <> +void GpuVectorT::printOneElement(std::ostream& os, size_t idx) const { + LOG(FATAL) << "Not implemented"; +} + +template <> +void GpuVectorT::printOneElement(std::ostream& os, size_t idx) const { + LOG(FATAL) << "Not implemented"; +} + +template <> +void CpuVectorT::rand() { + LOG(FATAL) << "Not implemented"; +} +template <> +void GpuVectorT::rand(size_t classNum) { + LOG(FATAL) << "Not implemented"; +} + +template <> +void CpuVectorT::rand(size_t classNum) { + LOG(FATAL) << "Not implemented"; +} + +template <> +void GpuVectorT::rand() { + VectorPtr cPtr = Vector::create(this->size_, false); + cPtr->rand(); + + hl_memcpy_host2device(data_, cPtr->getData(), this->size_ * sizeof(real)); +} + +template <> +void GpuVectorT::rand(size_t classNum) { + IVectorPtr cPtr = IVector::create(this->size_, false); + cPtr->rand(classNum); + + hl_memcpy_host2device(data_, cPtr->getData(), this->size_ * sizeof(int)); +} + +template <> +void CpuVectorT::rand(size_t classNum) { + size_t size = this->getSize(); + int* data = this->getData(); + for (size_t i = 0; i < size; i++) { + data[i] = + std::min(classNum - 1, + size_t(::rand() * (1. / ((double)RAND_MAX + 1)) * classNum)); + } +} + +template <> +void CpuVectorT::rand() { + size_t size = this->getSize(); + real* data = this->getData(); + for (size_t i = 0; i < size; i++) { + data[i] = ::rand() * (1. / (double)RAND_MAX); + // data[ii] = ((temp > RAND_MAX/2)? 1 : -1) * + // sqrt( abs((temp-RAND_MAX/2))/(double(RAND_MAX))/2048 ); + } +} + +template +void CpuVectorT::randnorm(real, real) { + LOG(FATAL) << "Not implemented"; +} + +template +void CpuVectorT::uniform(real, real) { + LOG(FATAL) << "Not implemented"; +} + +template +void GpuVectorT::randnorm(real, real) { + LOG(FATAL) << "Not implemented"; +} + +template +void GpuVectorT::uniform(real, real) { + LOG(FATAL) << "Not implemented"; +} + +template <> +void CpuVectorT::randnorm(real mean, real std) { + size_t size = this->getSize(); + real* data = this->getData(); + unsigned int* seed = ThreadLocalRand::getSeed(); + auto rand1 = [&]() { return (1. + ::rand_r(seed)) * (1. / (1. + RAND_MAX)); }; + for (size_t i = 0; i < size - 1; i += 2) { + real r1 = rand1(); + r1 = std::sqrt(-2 * std::log(r1)); + real r2 = rand1(); + data[i] = mean + std * r1 * cos(2 * M_PI * r2); + data[i + 1] = mean + std * r1 * sin(2 * M_PI * r2); + } + real r1 = rand1(); + r1 = std::sqrt(-2 * std::log(r1)); + real r2 = rand1(); + data[size - 1] = mean + std * r1 * cos(2 * M_PI * r2); +} + +template <> +void CpuVectorT::uniform(real left, real right) { + size_t size = this->getSize(); + real* data = this->getData(); + real range = right - left; + unsigned int* seed = ThreadLocalRand::getSeed(); + auto rand1 = [&]() { return ::rand_r(seed) * (1. / (1. + RAND_MAX)); }; + for (size_t i = 0; i < size; ++i) { + data[i] = rand1() * range + left; + } +} + +template <> +void GpuVectorT::randnorm(real mean, real std) { + CpuVector cpuVec = CpuVector(this->getSize()); + cpuVec.randnorm(mean, std); + + hl_memcpy_host2device( + data_, cpuVec.getData(), this->getSize() * sizeof(real)); +} + +template <> +void GpuVectorT::uniform(real left, real right) { + CpuVector cpuVec = CpuVector(this->getSize()); + cpuVec.uniform(left, right); + + hl_memcpy_host2device( + data_, cpuVec.getData(), this->getSize() * sizeof(real)); +} + +template +CpuVectorT::CpuVectorT(size_t size) + : VectorT(size, + std::make_shared(sizeof(T) * size), + 0, /* offset = 0 */ + false /* useGpu = false */) {} + +template +CpuVectorT::CpuVectorT(const VectorT& src) + : VectorT(src.getSize(), + src.getMemoryHandle(), + 0, /* offset = 0 */ + false /* useGpu = false */) { + if (typeid(*this->memoryHandle_.get()) != typeid(CpuMemoryHandle)) { + this->memoryHandle_ = + std::make_shared(sizeof(T) * this->getSize()); + this->data_ = reinterpret_cast(this->memoryHandle_->getBuf()); + } + src.copyTo(this); +} + +template +T CpuVectorT::getAbsSum() { + const T* A = this->getData(); + size_t size = this->getSize(); + T sum = 0; + for (size_t i = 0; i < size; i++) { + sum += (A[i] > 0) ? A[i] : -A[i]; + } + return sum; +} + +// cannot use above version, due to precision issue of float +template <> +real CpuVectorT::getAbsSum() { + const real* A = this->getData(); + size_t size = this->getSize(); + double sum = 0; + for (size_t i = 0; i < size; i++) { + sum += (A[i] > 0) ? A[i] : -A[i]; + } + return sum; +} + +template +T CpuVectorT::getSum() { + const T* A = this->getData(); + size_t size = this->getSize(); + T sum = 0; + for (size_t i = 0; i < size; i++) { + sum += A[i]; + } + return sum; +} + +template <> +real CpuVectorT::getSum() { + const real* A = this->getData(); + size_t size = this->getSize(); + double sum = 0; + for (size_t i = 0; i < size; i++) { + sum += A[i]; + } + return sum; +} + +template +T CpuVectorT::get(size_t pos) { + return this->getData()[pos]; +} + +template +T CpuVectorT::getMax() { + const T* A = this->getData(); + size_t size = this->getSize(); + T res = A[0]; + for (size_t i = 1; i < size; i++) { + if (res < A[i]) res = A[i]; + } + return res; +} + +template +T CpuVectorT::getAbsMax() { + const T* A = this->getData(); + size_t size = this->getSize(); + T res = std::abs(A[0]); + for (size_t i = 1; i < size; i++) { + if (res < std::abs(A[i])) res = std::abs(A[i]); + } + return res; +} + +template +T CpuVectorT::getMin() { + const T* A = this->getData(); + size_t size = this->getSize(); + T res = A[0]; + for (size_t i = 1; i < size; i++) { + if (res > A[i]) res = A[i]; + } + return res; +} + +template +void CpuVectorT::isEqualTo(const VectorT& b, const T& value) { + size_t size = this->getSize(); + CHECK_EQ(b.getSize(), size); + + const T* B = b.getData(); + T* A = this->getData(); + for (size_t i = 0; i < size; i++) { + A[i] = (B[i] == value); + } +} + +template +void CpuVectorT::selectFrom(const VectorT& src, const VectorT& ids) { + size_t size = this->getSize(); + CHECK_EQ(ids.getSize(), size); + + const int* indices = ids.getData(); + const T* B = src.getData(); + T* A = this->getData(); + for (size_t i = 0; i < size; i++) { + int index = indices[i]; + CHECK_LT(index, (int)src.getSize()); + A[i] = B[index]; + } +} + +static int getSignAndExponentOfFloat(float a) { + uint32_t* pa = reinterpret_cast(&a); + return *pa >> 23; +} + +template +void CpuVectorT::histogram(std::ostream& os, int type) { + LOG(FATAL) << "Not implemented"; +} + +template <> +void CpuVectorT::histogram(std::ostream& os, int type) { + int counters[512]; + memset(counters, 0, sizeof(counters)); + int counterZero = 0; + + const real* A = this->getData(); + size_t size = this->getSize(); + for (size_t i = 0; i < size; i++) { + if (A[i] == 0.0f) { + ++counterZero; + } else { + ++counters[getSignAndExponentOfFloat(A[i])]; + } + } + + int64_t sum = 0; + float sizeNonZero = size - counterZero; + os << "zero:" << counterZero; + for (int i = 0; i < 256; i++) { + int counter = counters[i]; + if (counter) { + os << " 2^" << i - 127 << ":" << counter / sizeNonZero * 100 << "%"; + sum += counter * (i - 127); + } + } + for (int i = 0; i < 256; i++) { + int counter = counters[i + 256]; + if (counter) { + os << " -2^" << i - 127 << ":" << counter / sizeNonZero * 100 << "%"; + sum += counter * (i - 127); + } + } + os << ", nonzero_exponent_avg=" << sum / sizeNonZero; +} + +template +void CpuVectorT::zeroMem() { + memset(this->getData(), 0, sizeof(T) * this->getSize()); +} + +template +void CpuVectorT::reset(const T& value) { + T* A = this->getData(); + size_t size = this->getSize(); + for (size_t i = 0; i < size; i++) { + A[i] = value; + } +} + +template +void CpuVectorT::fillSequence() { + T* A = this->getData(); + size_t size = this->getSize(); + for (size_t i = 0; i < size; i++) { + A[i] = i; + } +} + +template +void CpuVectorT::copyFrom(const VectorT& src) { + src.copyTo(this); +} + +template +void CpuVectorT::copyFrom(const VectorT& src, hl_stream_t stream) { + if (typeid(src) == typeid(GpuVectorT)) { + hl_memcpy_async((void*)this->getData(), + (void*)src.getData(), + sizeof(T) * this->getSize(), + stream); + // There is a need to add synchronization to ensure that the data is copied. + hl_stream_synchronize(stream); + } else { + src.copyTo(this); + } +} + +template +void CpuVectorT::copyFrom(const T* hostSrc, size_t size) { + CHECK(hostSrc != NULL); + CHECK_LE(size, this->size_); + memcpy(this->data_, hostSrc, sizeof(T) * size); +} + +template +void CpuVectorT::copyFrom(const T* hostSrc, + size_t size, + hl_stream_t stream) { + (void)stream; + + CHECK(hostSrc != NULL); + CHECK_LE(size, this->size_); + memcpy(this->data_, hostSrc, sizeof(T) * size); +} + +template +void CpuVectorT::copyTo(CpuVectorT* dest) const { + CHECK_EQ(this->getSize(), dest->getSize()); + memcpy(dest->getData(), this->getData(), sizeof(T) * this->getSize()); +} + +template +void CpuVectorT::copyTo(GpuVectorT* dest) const { + CHECK_EQ(this->getSize(), dest->getSize()); + hl_memcpy_host2device((void*)dest->getData(), + (void*)this->getData(), + sizeof(T) * this->getSize()); +} + +template <> +void CpuVectorT::print(std::ostream& os, size_t num) const { + size_t w = size_ < num ? size_ : num; + os << "["; + for (size_t i = 0; i < w; ++i) { + os << data_[i] << " "; + } + os << "]" << std::endl; +} + +template <> +void CpuVectorT::print(std::ostream& os, size_t num) const { + size_t w = size_ < num ? size_ : num; + os << "["; + for (size_t i = 0; i < w; ++i) { + os << (int)data_[i] << " "; + } + os << "]" << std::endl; +} + +template <> +void CpuVectorT::printOneElement(std::ostream& os, size_t idx) const { + CHECK_LT(idx, size_); + os << data_[idx] << ";"; +} + +template <> +void CpuVectorT::printOneElement(std::ostream& os, size_t idx) const { + CHECK_LT(idx, size_); + os << (int)data_[idx] << ";"; +} + +template +void ParallelCpuVectorT::parallelExec(ExecFunc func) { + LOG(FATAL) << "Not implemented"; +} + +template <> +void ParallelCpuVectorT::parallelExec(ExecFunc func) { + pool_->exec([this, func](int tid, size_t numThreads) { + auto interval = calcSplitArrayInterval( + this->getSize(), (size_t)tid, numThreads, 8LU /*for avx*/); + // setup sub bufs + CpuVector subVec(0, nullptr); + subVec.subVecFrom(*this, interval); + func(subVec); + }); +} + +template +void ParallelCpuVectorT::exec(SyncThreadPool::JobFunc func) { + LOG(FATAL) << "Not implemented"; +} + +template <> +void ParallelCpuVectorT::exec(SyncThreadPool::JobFunc func) { + pool_->exec(func); +} + +template +CpuGpuVectorT::CpuGpuVectorT(size_t size, bool useGpu) : sync_(nullptr) { + if (!useGpu) { + cpuVectorT_ = std::make_shared>(size); + } else { + gpuVectorT_ = std::make_shared>(size); + } + setSync(useGpu); +} + +template +CpuGpuVectorT::CpuGpuVectorT(const std::shared_ptr>& src) + : sync_(nullptr) { + bool useGpu = src->useGpu(); + if (useGpu) { + gpuVectorT_ = src; + } else { + cpuVectorT_ = src; + } + setSync(useGpu); +} + +template +CpuGpuVectorT::CpuGpuVectorT(size_t size, T* data, bool useGpu) + : sync_(nullptr) { + if (!useGpu) { + cpuVectorT_ = std::make_shared>(size, data); + setSync(DATA_AT_CPU); + } else { + gpuVectorT_ = std::make_shared>(size, data); + setSync(DATA_AT_GPU); + } +} + +template +std::shared_ptr> CpuGpuVectorT::create(size_t size, + bool useGpu) { + return std::make_shared>(size, useGpu); +} + +template +void CpuGpuVectorT::resize(size_t size, bool useGpu) { + if (useGpu) { + CHECK(gpuVectorT_) << "gpuVectorT_ is null"; + // If memoryHandle_ is nullptr, + // the data may be owned by the caller when it was constructed. + // It should not resize for this case. + if (gpuVectorT_->getMemoryHandle()) { + gpuVectorT_->resize(size); + } else { + CHECK_EQ(gpuVectorT_->getSize(), size); + } + } else { + CHECK(cpuVectorT_) << "cpuVectorT_ is null"; + // If memoryHandle_ is nullptr, + // the data may be owned by the caller when it was constructed. + // It should not resize for this case. + if (cpuVectorT_->getMemoryHandle()) { + cpuVectorT_->resize(size); + } else { + CHECK_EQ(cpuVectorT_->getSize(), size); + } + } + setSync(useGpu); +} + +template +void CpuGpuVectorT::resizeOrCreate(std::shared_ptr>& vec, + size_t size, + bool useGpu) { + if (vec) { + vec->resize(size, useGpu); + } else { + vec = create(size, useGpu); + } +} + +template +void CpuGpuVectorT::resizeOrCreate(size_t size, bool useGpu) { + if (useGpu && (!gpuVectorT_)) { + gpuVectorT_ = VectorT::create(size, true); + } else if ((!useGpu) && (!cpuVectorT_)) { + cpuVectorT_ = VectorT::create(size, false); + } else { + CHECK((useGpu && gpuVectorT_) || (!useGpu && cpuVectorT_)); + this->resize(size, useGpu); + } +} + +template +CpuGpuVectorT::CpuGpuVectorT(CpuGpuVectorT& src, + size_t offset, + size_t size) + : sync_(nullptr) { + CHECK_LE(offset + size, static_cast(src.getSize())); +#ifdef PADDLE_WITH_CUDA + SyncedFlag* flag = src.getSync(); + if (*flag == DATA_AT_CPU) { + src.copyToGpu(); // will set synchronous data between CPU and GPU + } else if (*flag == DATA_AT_GPU) { + src.copyToCpu(); // will set synchronous data between CPU and GPU + } +#endif + auto cMemHandle = (src.getVector(false))->getMemoryHandle(); + cpuVectorT_ = std::make_shared>( + size, std::dynamic_pointer_cast(cMemHandle), offset); +#ifdef PADDLE_WITH_CUDA + auto gMemHandle = (src.getVector(true))->getMemoryHandle(); + gpuVectorT_ = std::make_shared>( + size, std::dynamic_pointer_cast(gMemHandle), offset); + src.setSync(SYNCED); +#endif + setSync(src.getSync()); +} + +template +std::shared_ptr> CpuGpuVectorT::getVector( + bool useGpu) const { + auto* self = const_cast*>(this); + if (useGpu) { + self->copyToGpu(); + return std::const_pointer_cast>(gpuVectorT_); + } else { + self->copyToCpu(); + return std::const_pointer_cast>(cpuVectorT_); + } +} + +template +std::shared_ptr>& CpuGpuVectorT::getMutableVector(bool useGpu) { + setSync(useGpu); + if (useGpu) { + copyToGpu(); + return gpuVectorT_; + } else { + copyToCpu(); + return cpuVectorT_; + } +} + +template +const T* CpuGpuVectorT::getData(bool useGpu) const { + auto self = const_cast*>(this); + if (useGpu) { + self->copyToGpu(); + return gpuVectorT_->getData(); + } else { + self->copyToCpu(); + return cpuVectorT_->getData(); + } +} + +// Operation will change data and need to reset sync_ & syncFlag_. +#define MUTABLE_VECTOR_OP(OP, useGpu, args...) \ + do { \ + if (useGpu) { \ + copyToGpu(); \ + setSync(useGpu); \ + return gpuVectorT_->OP(args); \ + } else { \ + copyToCpu(); \ + setSync(useGpu); \ + return cpuVectorT_->OP(args); \ + } \ + } while (0) + +template +T* CpuGpuVectorT::getMutableData(bool useGpu) { + MUTABLE_VECTOR_OP(getData, useGpu); +} + +template +void CpuGpuVectorT::zeroMem(bool useGpu) { + MUTABLE_VECTOR_OP(zeroMem, useGpu); +} + +template +void CpuGpuVectorT::fillSequence(bool useGpu) { + MUTABLE_VECTOR_OP(fillSequence, useGpu); +} + +template +void CpuGpuVectorT::setElement(size_t i, const T& value, bool useGpu) { + MUTABLE_VECTOR_OP(setElement, useGpu, i, value); +} + +template +T CpuGpuVectorT::getElement(size_t i) const { + switch (*this->getSync()) { + case SYNCED: + case DATA_AT_CPU: + return cpuVectorT_->getElement(i); + break; + case DATA_AT_GPU: + return gpuVectorT_->getElement(i); + break; + default: + LOG(FATAL) << "Not support"; + break; + } +} + +template +void CpuGpuVectorT::copyFrom(const VectorT& src, hl_stream_t stream) { + auto cVec = dynamic_cast*>(&src); + auto gVec = dynamic_cast*>(&src); + if (cVec) { + copyToCpu(cVec->getData(), cVec->getSize(), stream); + } else if (gVec) { + copyToGpu(gVec->getData(), gVec->getSize(), stream); + } else { + LOG(FATAL) << "Invalid type of src"; + } +} + +template +void CpuGpuVectorT::copyFrom(const T* data, size_t size, bool useGpu) { + if (useGpu) { + copyToGpu(data, size); + } else { + copyToCpu(data, size); + } +} + +template +void CpuGpuVectorT::copyFrom(const T* data, + size_t size, + hl_stream_t stream, + bool useGpu) { + if (useGpu) { + copyToGpu(data, size, stream); + } else { + copyToCpu(data, size, stream); + } +} + +template +void CpuGpuVectorT::copyFrom(CpuGpuVectorT& src, + size_t offset, + size_t size, + bool useGpu, + hl_stream_t stream) { + if (useGpu) { + VectorT::resizeOrCreate(gpuVectorT_, size, true); + gpuVectorT_->copyFrom(src.getData(true) + offset, size, stream); + } else { + VectorT::resizeOrCreate(cpuVectorT_, size, false); + cpuVectorT_->copyFrom(src.getData(false) + offset, size, stream); + } + setSync(useGpu); +} + +template +void CpuGpuVectorT::copyFrom(CpuGpuVectorT& src, hl_stream_t stream) { + switch (*src.getSync()) { + case DATA_AT_CPU: + copyFrom(*(src.getVector(false)), stream); + break; + case DATA_AT_GPU: + copyFrom(*(src.getVector(true)), stream); + break; + case SYNCED: + copyFrom(*(src.getVector(false)), stream); + copyFrom(*(src.getVector(true)), stream); + setSync(SYNCED); + break; + default: + LOG(FATAL) << "Not support"; + break; + } +} + +template +void CpuGpuVectorT::copyToCpu() { + switch (*this->getSync()) { + case DATA_AT_GPU: + CHECK(gpuVectorT_); + this->resizeOrCreate(gpuVectorT_->getSize(), false); + cpuVectorT_->copyFrom(*gpuVectorT_); + setSync(SYNCED); + break; + case DATA_AT_CPU: + case SYNCED: + CHECK(cpuVectorT_); + break; + default: + LOG(FATAL) << "Not support"; + break; + } +} + +template +void CpuGpuVectorT::copyToGpu() { + switch (*this->getSync()) { + case DATA_AT_CPU: + CHECK(cpuVectorT_); + this->resizeOrCreate(cpuVectorT_->getSize(), true); + gpuVectorT_->copyFrom(*cpuVectorT_); + setSync(SYNCED); + break; + case DATA_AT_GPU: + case SYNCED: + CHECK(gpuVectorT_); + break; + default: + LOG(FATAL) << "Not support"; + break; + } +} + +template class VectorT; +template class VectorT; +template class CpuVectorT; +template class CpuVectorT; +template class GpuVectorT; +template class GpuVectorT; +template class CpuGpuVectorT; +template class CpuGpuVectorT; + +} // namespace paddle diff --git a/paddle/legacy/math/Vector.h b/paddle/legacy/math/Vector.h new file mode 100644 index 0000000000000000000000000000000000000000..63cb4651c52219807e11e778db9c42667759a055 --- /dev/null +++ b/paddle/legacy/math/Vector.h @@ -0,0 +1,726 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include + +#include "BaseMatrix.h" +#include "MemoryHandle.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Thread.h" + +namespace paddle { + +template +class GpuVectorT; +template +class CpuVectorT; + +template +class BaseVector; + +class SyncThreadPool; + +class Matrix; + +template +class BaseVector : public BaseMatrixT { + public: + BaseVector(size_t size, T* data, bool useGpu) + : BaseMatrixT(1, size, data, false, useGpu), size_(this->width_) {} + + ~BaseVector() {} + + protected: + size_t& size_; +}; + +/** + * Copy or assignemnt constructor will share the data as opposed to making a + * copy of the original data. To make a copy of the orinal data, use copyFrom() + * instead. + */ +template +class VectorT : public BaseVector { + protected: + VectorT(size_t size, MemoryHandlePtr memoryHandle, size_t offset, bool useGpu) + : BaseVector(size, + reinterpret_cast(memoryHandle->getBuf()) + offset, + useGpu) { + memoryHandle_ = memoryHandle; + } + + // data is still owned by the caller. + // data should be valid during the life of this vector. + // Caller is responsible for release the memory. + VectorT(size_t size, T* data, bool useGpu) + : BaseVector(size, data, useGpu) {} + + public: + virtual ~VectorT() {} + + static std::shared_ptr> create(size_t size, bool useGpu); + + static std::shared_ptr> create(T* data, size_t size, bool useGpu); + + static std::shared_ptr> create(size_t size, + MemoryHandlePtr memoryHandle, + size_t offset = 0); + + // owner can set SyncThreadPool, + // if not set, will use globalSyncThreadPool, + // which can be used in main thread only. + static std::shared_ptr> createParallelVector( + size_t size, bool useGpu, SyncThreadPool* pool = nullptr); + + size_t getSize() const { return this->size_; } + const T* getData() const { return this->data_; } + T* getData() { return this->data_; } + + virtual void zeroMem() = 0; + // set all elements to value + virtual void reset(const T& value) = 0; + // fill data by 0, 1, 2, ... + virtual void fillSequence() = 0; + + MemoryHandlePtr getMemoryHandle() const { return memoryHandle_; } + + /** + * resizing to a big vector will not preserve old values. + */ + void resize(size_t newSize) { + if (!memoryHandle_ || newSize * sizeof(T) > memoryHandle_->getAllocSize()) { + memoryHandle_ = newMemory(newSize * sizeof(T)); + this->data_ = reinterpret_cast(memoryHandle_->getBuf()); + } + this->size_ = newSize; + } + + static void resizeOrCreate(std::shared_ptr>& vec, + size_t size, + bool useGpu) { + if (vec) { + vec->resize(size); + } else { + vec = create(size, useGpu); + } + } + + virtual MemoryHandlePtr newMemory(size_t size) = 0; + + /** + * form sub vector from *src*, shallow copy + */ + void subVecFrom(const VectorT& src, size_t start, size_t size) { + CHECK_EQ(BaseVector::useGpu_, src.useGpu_); + CHECK_LT(start, src.size_); + CHECK_LE(start + size, src.size_); + + BaseVector::size_ = size; + BaseVector::data_ = const_cast(src.data_) + start; + } + + std::shared_ptr> subVec(size_t start, size_t size) { + CHECK_LE(start + size, static_cast(getSize())); + return VectorT::create(getData() + start, size, BaseVector::useGpu_); + } + + /** + * form sub vector from *src*, shallow copy + */ + void subVecFrom(const T* src, size_t start, size_t size) { + BaseVector::size_ = size; + BaseVector::data_ = const_cast(src) + start; + } + + /** + * form sub vector from *src*, shallow copy + * in *interval* [interval.first, interval.second) + */ + void subVecFrom(const VectorT& src, std::pair interval) { + subVecFrom(src, interval.first, interval.second - interval.first); + } + + /** + * convert the vector to a sparse one_hot matrix of width idRange + * only applies to IVector + */ + std::shared_ptr toOneHotSparseMatrix(size_t idRange, bool useGpu); + + /** + * @brief cast vector of "real" elements to "int" elements. + * + * @note: float -> int must be casted, or you'll get wrong data. + */ + std::shared_ptr> castToInt(); + + /** + * This function will crash if the size of src and dest is different. + */ + virtual void copyFrom(const VectorT& src) = 0; + + /** + * If GpuVector, this function is an asynchronous interface, + * will push the copy-task to the specifed-stream and return immediately. + * + * If CpuVector, this function is an synchronous interface, + * same as the copyFrom(const VectorT& src). + */ + virtual void copyFrom(const VectorT& src, hl_stream_t stream) = 0; + + /** + * copy size elements from src + * + * If this is GpuVector, src can be cpu or gpu memory + * + * If this is CpuVector, src is assumed to be cpu memory + */ + virtual void copyFrom(const T* src, size_t size) = 0; + + /** + * copy size elements from src + * + * If this is GpuVector, src can be cpu or gpu memory + * + * If this is CpuVector, src is assumed to be cpu memory, + */ + virtual void copyFrom(const T* src, size_t size, hl_stream_t stream) = 0; + + /** + * exec a func in single/multi thread + */ + virtual void exec(SyncThreadPool::JobFunc func) { func(0, 1); } + + /// Get the buffer point with beginPos + virtual T* getPoint(const uint64_t beginPos) = 0; + + /// Get the value for the i'th element + virtual T getElement(size_t i) const = 0; + virtual void setElement(size_t i, const T& value) = 0; + + //---------- math operations ---------------- + + // sum of the absolute value of each elements + virtual T getAbsSum() = 0; + + virtual T getSum() = 0; + virtual T getMax() = 0; + virtual T getAbsMax() = 0; + virtual T getMin() = 0; + + /// element-wise calc: this = (b == value) + virtual void isEqualTo(const VectorT& b, const T& value) = 0; + + /// select elements indexed by *ids* from vector *src* + virtual void selectFrom(const VectorT& src, const VectorT& ids) = 0; + + enum HistogramType { + HISTOGRAM_EXPONENT = 0, + }; + + /** + * @brief print histogram of vector values + * + * @note only exponent histogram supported currently + */ + virtual void histogram(std::ostream& os, int type = HISTOGRAM_EXPONENT) = 0; + + /// generate uniform random value for each element + virtual void rand() = 0; + /** + * generate uniform random value for each element, + * data range is from 0 to (classes - 1). + */ + virtual void rand(size_t classes) = 0; + + /** + * Debug use only. Very inefficient for GPU vector. + * get the value at pos. + */ + virtual T get(size_t pos) = 0; + + /** + * generate univariate Gaussian distributed random numbers + * with given mean and standardDeviation. + */ + virtual void randnorm(real mean, real standardDeviation) = 0; + + /** + * generate uniform distributed random numbers + * with given range. + */ + virtual void uniform(real left, real right) = 0; + + /// print the first "num" elements of the Vector + virtual void print(std::ostream& os, size_t num) const = 0; + + /// print the "idx" element of the Vector + virtual void printOneElement(std::ostream& os, size_t idx) const = 0; + + template + void operator=(const ExpressionType& expr) { + if (BaseVector::useGpu_) { + TensorGpuApply(*this, expr); + } else { + TensorCpuApply(*this, expr); + } + } + + protected: + friend class GpuVectorT; + friend class CpuVectorT; + virtual void copyTo(CpuVectorT* dest) const = 0; + virtual void copyTo(GpuVectorT* dest) const = 0; + MemoryHandlePtr memoryHandle_; +}; + +template +std::ostream& operator<<(std::ostream& os, const VectorT& vec) { + vec.print(os, vec.getSize()); + return os; +} + +template +class GpuVectorT : public VectorT { + public: + explicit GpuVectorT(size_t size); + GpuVectorT(size_t size, GpuMemHandlePtr memHandle, size_t offset) + : VectorT(size, memHandle, offset, true) {} + + // data is still owned by the caller. + // data should be valid during the life of this vector. + // Caller is responsible for release the memory. + GpuVectorT(size_t size, T* data) : VectorT(size, data, true) {} + + virtual MemoryHandlePtr newMemory(size_t size) { + return std::make_shared(size); + } + virtual void zeroMem(); + virtual void reset(const T& value); + virtual void fillSequence(); + + virtual void copyFrom(const T* src, size_t size); + virtual void copyFrom(const T* src, size_t size, hl_stream_t stream); + virtual void copyFrom(const VectorT& src); + virtual void copyFrom(const VectorT& src, hl_stream_t stream); + virtual T getElement(size_t i) const; + virtual void setElement(size_t i, const T& value); + virtual T* getPoint(const uint64_t beginPos); + + virtual T getAbsSum(); + virtual T getSum(); + virtual T getMax(); + virtual T getAbsMax(); + virtual T getMin(); + virtual void isEqualTo(const VectorT& b, const T& value); + virtual void selectFrom(const VectorT& src, const VectorT& ids); + virtual void histogram(std::ostream& os, int type); + virtual void rand(); + virtual void rand(size_t classes); + virtual void randnorm(real mean, real standardDeviation); + virtual void uniform(real left, real right); + virtual T get(size_t pos); + virtual void print(std::ostream& os, size_t num) const; + virtual void printOneElement(std::ostream& os, size_t idx) const; + + template + void operator=(const ExpressionType& expr) { + TensorGpuApply(*this, expr); + } + + protected: + virtual void copyTo(CpuVectorT* dest) const; + virtual void copyTo(GpuVectorT* dest) const; +}; + +template +class CpuVectorT : public VectorT { + public: + explicit CpuVectorT(size_t size); + CpuVectorT(size_t size, MemoryHandlePtr memoryHandle, size_t offset) + : VectorT(size, memoryHandle, offset, false) {} + + // data is still owned by the caller. + // data should be valid during the life of this vector. + // Caller is responsible for release the memory. + CpuVectorT(size_t size, T* data) : VectorT(size, data, false) {} + + /** + * If src is a CpuVector, the new CpuVector will share the data with src + * + * If src is a GpuVector, the new CpuVector will copy data from src + */ + explicit CpuVectorT(const VectorT& src); + + virtual MemoryHandlePtr newMemory(size_t size) { + return std::make_shared(size); + } + + virtual void zeroMem(); + virtual void reset(const T& value); + virtual void fillSequence(); + virtual void copyFrom(const T* src, size_t size); + virtual void copyFrom(const T* src, size_t size, hl_stream_t stream); + virtual void copyFrom(const VectorT& src); + virtual void copyFrom(const VectorT& src, hl_stream_t stream); + virtual void copyTo(CpuVectorT* dest) const; + virtual void copyTo(GpuVectorT* dest) const; + + /// Get the buffer point with beginPos + virtual T* getPoint(const uint64_t beginPos) { + return this->getData() + beginPos; + } + + virtual T getElement(size_t i) const { return this->getData()[i]; } + virtual void setElement(size_t i, const T& value) { + this->getData()[i] = value; + } + + virtual T getAbsSum(); + virtual T getSum(); + virtual T getMax(); + virtual T getAbsMax(); + virtual T getMin(); + virtual void isEqualTo(const VectorT& b, const T& value); + virtual void selectFrom(const VectorT& src, const VectorT& ids); + virtual void histogram(std::ostream& os, int type); + virtual void rand(); + virtual void rand(size_t classes); + virtual void randnorm(real mean, real standardDeviation); + virtual void uniform(real left, real right); + virtual T get(size_t pos); + virtual void print(std::ostream& os, size_t num) const; + virtual void printOneElement(std::ostream& os, size_t idx) const; + + template + void operator=(const ExpressionType& expr) { + TensorCpuApply(*this, expr); + } +}; + +template +class ParallelCpuVectorT : public CpuVectorT { + public: + ParallelCpuVectorT(size_t size, SyncThreadPool* pool) + : CpuVectorT(size), pool_(pool) {} + + virtual void zeroMem() { + parallelExec([](CpuVectorT& vec) { vec.CpuVectorT::zeroMem(); }); + } + virtual void randnorm(real mean, real standardDeviation) { + parallelExec([=](CpuVectorT& vec) { + vec.CpuVectorT::randnorm(mean, standardDeviation); + }); + } + virtual void uniform(real left, real right) { + parallelExec( + [=](CpuVectorT& vec) { vec.CpuVectorT::uniform(left, right); }); + } + + virtual void exec(SyncThreadPool::JobFunc jobFunc); + + private: + typedef std::function& vec)> ExecFunc; + void parallelExec(ExecFunc func); + SyncThreadPool* pool_; +}; + +/** + * A class to do conversion between CpuVector and GpuVector automatically. + */ +template +class CpuGpuVectorT { + public: + /** + * @brief An enum type of SyncedFlag using to + * mark data memory is in CPU or GPU. + * + * DATA_AT_CPU: data is located in CPU. + * + * DATA_AT_GPU: data is located in GPU. + * + * SYNCED: data is located in CPU and GPU simultaneously. + */ + enum SyncedFlag { DATA_AT_CPU = 0, DATA_AT_GPU = 1, SYNCED = 2 }; + + /** + * @brief A constructor, create cpuVectorT_ or gpuVectorT_. + * + * @param[in] size data size. + * @param[in] useGpu use gpu or not. + */ + explicit CpuGpuVectorT(size_t size, bool useGpu); + + /** + * @brief A constructor, create CpuGpuVectorT by VectorT. + * + * If src is CpuVector, cpuVectorT_ is shared data with src. + * + * If src is GpuVector, gpuVectorT_ is shared data with src. + */ + explicit CpuGpuVectorT(const std::shared_ptr>& src); + + /** + * @brief A constructor. + * + * If useGpu is true, data should be located in device and + * create gpuVectorT_ with data. + * + * If useGpu is false, data should be located in host and + * create cpuVectorT_ with data. + * + * @note Data is owned by the caller and should be valid during + * the life of this vector. + * Caller is responsible for release the memory. + */ + CpuGpuVectorT(size_t size, T* data, bool useGpu); + + CpuGpuVectorT(CpuGpuVectorT& src, size_t offset, size_t size); + + virtual ~CpuGpuVectorT() {} + + static std::shared_ptr> create(size_t size, bool useGpu); + + /** + * @brief resize vector. + * + * If useGpu is true, resize gpuVectorT_ and set syncFlag_ to DATA_AT_GPU, + * + * otherwise resize cpuVectorT_ and set syncFlag_ to DATA_AT_CPU. + */ + void resize(size_t size, bool useGpu); + + /** + * @brief resize or create CpuGpuVectorT. + */ + static void resizeOrCreate(std::shared_ptr>& vec, + size_t size, + bool useGpu); + + /** + * @brief return a const cpuVectorT_ or gpuVectorT_. + * + * If useGpu is true, return gpuVectorT_. + * + * If useGpu is false, return cpuVectorT_. + * + * @note Caller should not change the data. + * If caller changes const attribute, + * should set syncFlag_. + */ + std::shared_ptr> getVector(bool useGpu) const; + + /** + * @brief return a const cpuVectorT_ or gpuVectorT_. + * + * @note: This interface will change syncFlag_, so if you will + * not change the data, you should call getVector. + */ + std::shared_ptr>& getMutableVector(bool useGpu); + + /** + * @brief return const T* data. + * + * If useGpu is true, return device data. + * + * If useGpu is false, return host data. + */ + const T* getData(bool useGpu) const; + + // TODO(yuyang18): Make getData more c++ style. + // inline T* getData(bool useGpu) { + // return getMutableData(useGpu); + // } + + T* getMutableData(bool useGpu); + + /** + * If useGpu is true, gpuVectorT_->Op(). + * + * If useGpu is false, cpuVectorT_->Op(). + * + * Op is zeroMem, fillSequence, ... + */ + void zeroMem(bool useGpu); + void fillSequence(bool useGpu); + void setElement(size_t i, const T& value, bool useGpu); + + /** + * @brief return i-th element. + */ + T getElement(size_t i) const; + + /** + * @brief return vector size. + */ + size_t getSize() const { + size_t size = 0; + switch (*sync_) { + case SYNCED: + case DATA_AT_CPU: + size = cpuVectorT_->getSize(); + break; + case DATA_AT_GPU: + size = gpuVectorT_->getSize(); + break; + default: + LOG(FATAL) << "Not support"; + break; + } + return size; + } + + /// copy data to cpuVectorT_. + inline void copyToCpu(const T* data, size_t size) { + this->resizeOrCreate(size, false); + cpuVectorT_->copyFrom(data, size); + setSync(DATA_AT_CPU); + } + /// copy data to cpuVectorT_ using specifed-stream. + inline void copyToCpu(const T* data, size_t size, hl_stream_t stream) { + this->resizeOrCreate(size, false); + cpuVectorT_->copyFrom(data, size, stream); + setSync(DATA_AT_CPU); + } + + /// copy data to gpuVectorT_. + inline void copyToGpu(const T* data, size_t size) { + this->resizeOrCreate(size, true); + gpuVectorT_->copyFrom(data, size); + setSync(DATA_AT_GPU); + } + /// copy data to gpuVectorT_ using specifed-stream. + inline void copyToGpu(const T* data, size_t size, hl_stream_t stream) { + this->resizeOrCreate(size, true); + gpuVectorT_->copyFrom(data, size, stream); + setSync(DATA_AT_GPU); + } + + /** + * @brief copy from src using specifed-stream. + * + * If src is CpuVectorT, copy to cpuVectorT_. + * + * If src is GpuVectorT, copy to gpuVectorT_. + */ + void copyFrom(const VectorT& src, hl_stream_t stream); + + /** + * @brief copy data. + * + * If useGpu is false, copy host data to cpuVectorT_. + * + * If useGpu is true, copy device data to gpuVectorT_. + * + * @note data address should consistent with useGpu. + */ + void copyFrom(const T* data, size_t size, bool useGpu); + void copyFrom(const T* data, size_t size, hl_stream_t stream, bool useGpu); + + /** + * @brief copy from (src + offset) using specifed-stream. + */ + void copyFrom(CpuGpuVectorT& src, + size_t offset, + size_t size, + bool useGpu, + hl_stream_t stream); + + /** + * @brief copy from src using specifed-stream. + */ + void copyFrom(CpuGpuVectorT& src, hl_stream_t stream); + + /** + * @brief return sync_. + */ + inline SyncedFlag* getSync() const { return sync_; } + + /** + * @brief set sync_. + */ + inline void setSync(SyncedFlag* sync) { sync_ = sync; } + + inline void setSync(SyncedFlag syncFlag) { + if (sync_) { + *sync_ = syncFlag; + } else { + syncFlag_ = syncFlag; + sync_ = &syncFlag_; + } + } + + inline void setSync(bool useGpu) { + SyncedFlag flag = useGpu ? DATA_AT_GPU : DATA_AT_CPU; + setSync(flag); + } + + protected: + void resizeOrCreate(size_t size, bool useGpu); + + /** + * @brief copy between cpuVectorT_ and gpuVectorT_. + * + * If syncFlag_ is DATA_AT_CPU and SYNCED, do nothing. + * + * If syncFlag_ is DATA_AT_GPU, copy gpuVectorT_ to cpuVectorT_ + * and set syncFlag_ to SYNCED. + */ + void copyToCpu(); + + /** + * @brief copy between cpuVectorT_ and gpuVectorT_. + * + * If syncFlag_ is DATA_AT_GPU and SYNCED, do nothing. + * + * If syncFlag_ is DATA_AT_CPU, copy cpuVectorT_ to gpuVectorT_ + * and set syncFlag_ to SYNCED. + */ + void copyToGpu(); + + /// host pointer. + std::shared_ptr> cpuVectorT_; + /// device pointer. + std::shared_ptr> gpuVectorT_; + /// specify current data address. + SyncedFlag syncFlag_; + SyncedFlag* sync_; +}; + +typedef VectorT Vector; +typedef CpuVectorT CpuVector; +typedef GpuVectorT GpuVector; + +typedef VectorT IVector; +typedef CpuVectorT CpuIVector; +typedef GpuVectorT GpuIVector; + +typedef std::shared_ptr VectorPtr; +typedef std::shared_ptr CpuVectorPtr; +typedef std::shared_ptr GpuVectorPtr; + +typedef std::shared_ptr IVectorPtr; +typedef std::shared_ptr CpuIVectorPtr; +typedef std::shared_ptr GpuIVectorPtr; + +typedef CpuGpuVectorT CpuGpuVector; +typedef CpuGpuVectorT ICpuGpuVector; +typedef std::shared_ptr CpuGpuVectorPtr; +typedef std::shared_ptr ICpuGpuVectorPtr; + +} // namespace paddle diff --git a/paddle/math/tests/CMakeLists.txt b/paddle/legacy/math/tests/CMakeLists.txt similarity index 100% rename from paddle/math/tests/CMakeLists.txt rename to paddle/legacy/math/tests/CMakeLists.txt diff --git a/paddle/legacy/math/tests/OriginalOptimizerApi.h b/paddle/legacy/math/tests/OriginalOptimizerApi.h new file mode 100644 index 0000000000000000000000000000000000000000..f386e19958a21214151776e6d0ae7bb2a4530b6c --- /dev/null +++ b/paddle/legacy/math/tests/OriginalOptimizerApi.h @@ -0,0 +1,201 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/GlobalConstants.h" + +using namespace paddle; // NOLINT + +void SparseMomentumParameterOptimizer(const VectorPtr vecs[], + real alpha, + real beta, + real gamma, + real tau, + real learningRate) { + vecs[PARAMETER_MOMENTUM_UT]->add(*vecs[PARAMETER_GRADIENT], + -alpha * gamma * learningRate); + vecs[PARAMETER_MOMENTUM_VT]->add(*vecs[PARAMETER_GRADIENT], + tau * alpha * gamma * learningRate); + vecs[PARAMETER_VALUE]->add(*vecs[PARAMETER_MOMENTUM_UT], + tau / beta + 1.0 / alpha, + *vecs[PARAMETER_MOMENTUM_VT], + 1.0 / beta); +} + +void AdagradParameterOptimizer(const VectorPtr vecs[], + real epsilon, + real learningRate, + real momentum, + real decayRate) { + vecs[PARAMETER_GRADIENT_SQURESUM1]->addSquare(*vecs[PARAMETER_GRADIENT], + 1.0f); + vecs[PARAMETER_LEARNING_RATE]->add(*vecs[PARAMETER_GRADIENT_SQURESUM], + *vecs[PARAMETER_GRADIENT_SQURESUM1]); + vecs[PARAMETER_LEARNING_RATE]->add(epsilon); + vecs[PARAMETER_LEARNING_RATE]->invSqrt(*vecs[PARAMETER_LEARNING_RATE]); + + vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], + *vecs[PARAMETER_MOMENTUM], + *vecs[PARAMETER_LEARNING_RATE], + learningRate, + momentum, + decayRate); +} + +void AdaDeltaParameterOptimizer(const VectorPtr vecs[], + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate) { + // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 + vecs[PARAMETER_GRADIENT_SQURESUM]->decayAddSquare( + *vecs[PARAMETER_GRADIENT], rou, 1.0f - rou); + + // learn_rate = sqrt( ( E(dx_{t-1}^2) + epsilon ) / ( E(g_t^2) + epsilon ) ) + vecs[PARAMETER_LEARNING_RATE]->dotDiv(*vecs[PARAMETER_GRADIENT_SQURESUM1], + *vecs[PARAMETER_GRADIENT_SQURESUM], + epsilon, + epsilon); + vecs[PARAMETER_LEARNING_RATE]->sqrt2(); + + // E(dx_t^2) = \rou * E(dx_{t-1}^2) + (1-\rou) * (-g*learn_rate)^2 + vecs[PARAMETER_GRADIENT_SQURESUM1]->decayAddSquareMul( + *vecs[PARAMETER_GRADIENT], + *vecs[PARAMETER_LEARNING_RATE], + rou, + 1.0f - rou); + + vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], + *vecs[PARAMETER_MOMENTUM], + *vecs[PARAMETER_LEARNING_RATE], + learningRate, + momentum, + decayRate); +} + +void RMSPropParameterOptimizer(const VectorPtr vecs[], + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime) { + // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 + // For the first time update, make the sum be the current square + // so that the initial estimation of E(g_t^2) will not be too small. + vecs[PARAMETER_GRADIENT_SQURESUM]->decayAddSquare( + *vecs[PARAMETER_GRADIENT], accumulatedRou, firstTime ? 1.0f : 1.0f - rou); + + // E(g_t) = \rou * E(g_{t-1}) + (1-\rou) * g + vecs[PARAMETER_GRADIENT_SQURESUM1]->add( + *vecs[PARAMETER_GRADIENT], accumulatedRou, 1.0f - rou); + + // learn_rate = 1/sqrt( ( E(g_t^2) - (E(g_t))^2 + epsilon ) + // Basiclly if the sign of the gradient changes more often, + // the learning rate will be decreased. + vecs[PARAMETER_LEARNING_RATE]->assign(*vecs[PARAMETER_GRADIENT_SQURESUM]); + vecs[PARAMETER_LEARNING_RATE]->addSquare(*vecs[PARAMETER_GRADIENT_SQURESUM1], + -1.0f); + vecs[PARAMETER_LEARNING_RATE]->add(epsilon); + vecs[PARAMETER_LEARNING_RATE]->invSqrt(*vecs[PARAMETER_LEARNING_RATE]); + + vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], + *vecs[PARAMETER_MOMENTUM], + *vecs[PARAMETER_LEARNING_RATE], + learningRate, + momentum, + decayRate); +} + +void DecayedAdagradParameterOptimizer(const VectorPtr vecs[], + real accumulatedRou, + real rou, + real epsilon, + real learningRate, + real momentum, + real decayRate, + bool firstTime) { + // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 + // For the first time update, make the sum be the current square + // so that the initial estimation of E(g_t^2) will not be too small. + vecs[PARAMETER_GRADIENT_SQURESUM]->decayAddSquare( + *vecs[PARAMETER_GRADIENT], accumulatedRou, firstTime ? 1.0f : 1.0f - rou); + + // learn_rate = 1/sqrt( ( E(g_t^2) + epsilon ) + // Basiclly if the bigger the magnitude gradient is, + // the smaller the learning rate will be. + vecs[PARAMETER_LEARNING_RATE]->assign(epsilon); + vecs[PARAMETER_LEARNING_RATE]->add(*vecs[PARAMETER_GRADIENT_SQURESUM]); + vecs[PARAMETER_LEARNING_RATE]->invSqrt(*vecs[PARAMETER_LEARNING_RATE]); + + vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], + *vecs[PARAMETER_MOMENTUM], + *vecs[PARAMETER_LEARNING_RATE], + learningRate, + momentum, + decayRate); +} + +void AdamParameterOptimizer(const VectorPtr vecs[], + real beta1, + real beta2, + real beta1_power, + real beta2_power, + real epsilon, + real learningRate) { + Vector* m = vecs[PARAMETER_MOMENTUM].get(); + Vector* g = vecs[PARAMETER_GRADIENT].get(); + Vector* v = vecs[PARAMETER_SECOND_MOMENTUM].get(); + Vector* theta = vecs[PARAMETER_VALUE].get(); + + // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; + m->add(*g, beta1, 1 - beta1); + + // v_t = \beta_2 * v_{t-1} + (1-\beta_2)* g_{t-1}^2 + g->square2(); + v->add(*g, beta2, 1 - beta2); + + // tmp = m_t / ( \sqrt{v_t} + \epsilon ) + // \theta_t = \theta_{t-1} - \alpha * \sqrt(1-\beta_2^t) / (1-\beta_1^t) * tmp + g->sqrt2(*v); + g->dotDiv(*m, *g, 0., epsilon); + real alpha = + learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); + theta->add(*theta, 1.0, *g, -alpha); +} + +void AdamaxParameterOptimizer( + const VectorPtr vecs[], real beta1, real beta2, int64_t step, real alpha) { + Vector* m = vecs[PARAMETER_MOMENTUM].get(); + Vector* g = vecs[PARAMETER_GRADIENT].get(); + Vector* u = vecs[PARAMETER_WEIGHTED_INFINITY_NORM].get(); + Vector* theta = vecs[PARAMETER_VALUE].get(); + + // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; + m->add(*g, beta1, 1 - beta1); + + // u_t = max(\beta_2*u_{t-1}, abs(g_t)) + u->mulScalar(beta2); + g->abs2(); + u->max2(*u, *g); + + // \theta_t = \theta_{t-1} - (\alpha/(1-\beta_1^t))*m_t/u_t + g->dotDiv(*m, *u); + real learningRate = alpha / (1 - std::pow(beta1, step)); + theta->add(*theta, 1.0, *g, -learningRate); +} diff --git a/paddle/legacy/math/tests/PerfUtils.h b/paddle/legacy/math/tests/PerfUtils.h new file mode 100644 index 0000000000000000000000000000000000000000..eaf4869e4c994e5ec739fe650d0228687d24853f --- /dev/null +++ b/paddle/legacy/math/tests/PerfUtils.h @@ -0,0 +1,46 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +// Performance Check +#ifdef PADDLE_DISABLE_TIMER + +#define EXPRESSION_PERFORMANCE(expression) expression; + +#else + +#include "paddle/legacy/utils/Stat.h" +using namespace paddle; // NOLINT + +#define EXPRESSION_PERFORMANCE(expression) \ + do { \ + char expr[30]; \ + strncpy(expr, #expression, 30); \ + if (expr[29] != '\0') { \ + expr[27] = '.'; \ + expr[28] = '.'; \ + expr[29] = '\0'; \ + } \ + expression; \ + for (int i = 0; i < 20; i++) { \ + REGISTER_TIMER(expr); \ + expression; \ + } \ + LOG(INFO) << std::setiosflags(std::ios::left) << std::setfill(' ') \ + << *globalStat.getStat(expr); \ + globalStat.reset(); \ + } while (0) + +#endif diff --git a/paddle/legacy/math/tests/TensorCheck.h b/paddle/legacy/math/tests/TensorCheck.h new file mode 100644 index 0000000000000000000000000000000000000000..41c8ece282e05f55d063e6ad0d8805629c847d34 --- /dev/null +++ b/paddle/legacy/math/tests/TensorCheck.h @@ -0,0 +1,216 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +/** + * This file provides a TensorCheck template function, which can be used to + * compare CpuMatrix and GpuMatrix, CpuVector and GpuVector, and so on. + */ + +#include +#include "paddle/legacy/math/Matrix.h" + +namespace autotest { + +using paddle::Matrix; +using paddle::CpuMatrix; +using paddle::GpuMatrix; +using paddle::VectorT; +using paddle::CpuVectorT; +using paddle::GpuVectorT; + +class AssertEqual { + public: + AssertEqual(real err = 0) : err_(err) {} + + inline bool operator()(real a, real b) { + if (err_ == 0) { + if (a != b) { + return false; + } + } else { + if (std::fabs(a - b) > err_) { + if ((std::fabs(a - b) / std::fabs(a)) > (err_ / 10.0f)) { + return false; + } + } + } + + return true; + } + + private: + real err_; +}; + +template +class CopyToCpu; + +template <> +class CopyToCpu { + public: + explicit CopyToCpu(const CpuMatrix& arg) : arg_(arg) {} + const CpuMatrix& copiedArg() const { return arg_; } + + private: + const CpuMatrix& arg_; +}; + +template <> +class CopyToCpu { + public: + explicit CopyToCpu(const GpuMatrix& arg) + : arg_(arg.getHeight(), arg.getWidth()) { + arg_.copyFrom(arg); + } + CpuMatrix& copiedArg() { return arg_; } + + private: + CpuMatrix arg_; +}; + +template <> +class CopyToCpu { + public: + explicit CopyToCpu(const Matrix& arg) + : arg_(arg.getHeight(), arg.getWidth()) { + arg_.copyFrom(arg); + } + CpuMatrix& copiedArg() { return arg_; } + + private: + CpuMatrix arg_; +}; + +template +class CopyToCpu> { + public: + explicit CopyToCpu(const CpuVectorT& arg) : arg_(arg) {} + const CpuVectorT& copiedArg() const { return arg_; } + + private: + const CpuVectorT& arg_; +}; + +template +class CopyToCpu> { + public: + explicit CopyToCpu(const GpuVectorT& arg) : arg_(arg.getSize()) { + arg_.copyFrom(arg); + } + CpuVectorT& copiedArg() { return arg_; } + + private: + CpuVectorT arg_; +}; + +template +class CopyToCpu> { + public: + explicit CopyToCpu(const VectorT& arg) : arg_(arg.getSize()) { + arg_.copyFrom(arg); + } + CpuVectorT& copiedArg() { return arg_; } + + private: + CpuVectorT arg_; +}; + +template +void TensorCheck(AssertEq compare, + const CpuMatrix& matrix1, + const CpuMatrix& matrix2) { + CHECK(matrix1.getHeight() == matrix2.getHeight()); + CHECK(matrix1.getWidth() == matrix2.getWidth()); + + int height = matrix1.getHeight(); + int width = matrix1.getWidth(); + const real* data1 = matrix1.getData(); + const real* data2 = matrix2.getData(); + int count = 0; + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + real a = data1[i * width + j]; + real b = data2[i * width + j]; + if (!compare(a, b)) { + count++; + } + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; +} + +template +void TensorCheck(AssertEq compare, + const CpuVectorT& vector1, + const CpuVectorT& vector2) { + CHECK(vector1.getSize() == vector2.getSize()); + + const T* data1 = vector1.getData(); + const T* data2 = vector2.getData(); + size_t size = vector1.getSize(); + int count = 0; + for (size_t i = 0; i < size; i++) { + real a = data1[i]; + real b = data2[i]; + if (!compare(a, b)) { + count++; + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different elements."; +} + +template +void TensorCheck(AssertEq compare, + const Tensor1& tensor1, + const Tensor2& tensor2) { + TensorCheck(compare, + CopyToCpu(tensor1).copiedArg(), + CopyToCpu(tensor2).copiedArg()); +} + +template +void TensorCheck(AssertEq compare, real args1, real args2) { + EXPECT_EQ(compare(args1, args2), true) << "[Test error] args1 = " << args1 + << ", args2 = " << args2; +} + +template +void TensorCheck(AssertEq compare, size_t args1, size_t args2) { + EXPECT_EQ(args1, args2) << "[Test error] args1 = " << args1 + << ", args2 = " << args2; +} + +template +void TensorCheckEqual(const Tensor1& tensor1, const Tensor2& tensor2) { + AssertEqual compare(0); + TensorCheck(compare, + CopyToCpu(tensor1).copiedArg(), + CopyToCpu(tensor2).copiedArg()); +} + +template +void TensorCheckErr(const Tensor1& tensor1, const Tensor2& tensor2) { +#ifndef PADDLE_TYPE_DOUBLE + AssertEqual compare(1e-3); +#else + AssertEqual compare(1e-10); +#endif + TensorCheck(compare, + CopyToCpu(tensor1).copiedArg(), + CopyToCpu(tensor2).copiedArg()); +} + +} // namespace autotest diff --git a/paddle/legacy/math/tests/TestUtils.h b/paddle/legacy/math/tests/TestUtils.h new file mode 100644 index 0000000000000000000000000000000000000000..60e76359da61ac32346b093d9a9ff69104bfc494 --- /dev/null +++ b/paddle/legacy/math/tests/TestUtils.h @@ -0,0 +1,294 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +/** + * This file provides a AutoCompare calss to simplify the comparison + * of CPU and GPU member functions. + * + * This takes two steps + * 1. Construct an AutoCompare object. + * When constructing an AutoCompare object, you can set the err argument + * to specify the maximum error for CPU and GPU functions. + * + * 2. Use the template functions cmpWithArg or cmpWithoutArg. + * A. [cmpWithArg] Requires the caller construct the cpu arguments. + * + * AutoCompare test; + * Init Argument arg1,arg2... + * test.cmpWithArg(function, arg1, arg2....) + * + * B. [cmpWithoutArg] The caller do not need construct arguments. + * If matrix used in these functions arguments is the same size. + * Such as the element wise function and the aggregate function + * defined in the BaseMatrix.cpp. + * + * AutoCompare test; + * test.cmpWithoutArg(function, height, width) + */ + +#include +#include "TensorCheck.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" + +namespace autotest { + +using paddle::BaseMatrix; +using paddle::CpuMatrix; +using paddle::GpuMatrix; +using paddle::CpuIVector; +using paddle::GpuIVector; +using paddle::CpuSparseMatrix; +using paddle::GpuSparseMatrix; + +template +class ReplaceType { + public: + typedef T1 type; +}; + +template <> +class ReplaceType { + public: + typedef CpuMatrix type; +}; + +template <> +class ReplaceType { + public: + typedef GpuMatrix type; +}; + +template <> +class ReplaceType { + public: + typedef CpuMatrix type; +}; + +template <> +class ReplaceType { + public: + typedef GpuMatrix type; +}; + +// construct a argument +template +T construct(int height, int width); + +template <> +float construct(int height, int width) { + return 0.5; +} + +template <> +double construct(int height, int width) { + return 0.5; +} + +template <> +size_t construct(int height, int width) { + size_t offset = std::rand() % (height < width ? height : width); + return offset; +} + +template <> +CpuMatrix construct(int height, int width) { + CpuMatrix a(height, width); + return a; +} + +template <> +GpuMatrix construct(int height, int width) { + GpuMatrix a(height, width); + return a; +} + +// init a argument +template +void init(T& v) { + return; +} + +template <> +void init(CpuMatrix& v) { + v.randomizeUniform(); +} + +template <> +void init(GpuMatrix& v) { + v.randomizeUniform(); +} + +// init a tuple which contains a set of arguments. +template +inline typename std::enable_if::type initTuple( + std::tuple& t) {} + +template + inline typename std::enable_if < + I::type initTuple(std::tuple& t) { + init(std::get(t)); + initTuple(t); +} + +// copy a argument, copy src to dest +template +void copy(T1& dest, T2& src) { + dest = src; +} + +template <> +void copy(GpuMatrix& dest, CpuMatrix& src) { + dest.copyFrom(src); +} + +// copy a tuple, copy src to dest +template +inline typename std::enable_if::type copyTuple( + std::tuple& dest, std::tuple& src) {} + +template + inline typename std::enable_if < + I::type copyTuple(std::tuple& dest, + std::tuple& src) { + copy(std::get(dest), std::get(src)); + copyTuple(dest, src); +} + +// call member function +template +R call(C& obj, R (FC::*f)(FArgs...), Args&&... args) { + return (obj.*f)(args...); +} + +template +class ReturnType { + public: + typedef T type; +}; + +template <> +class ReturnType { + public: + typedef GpuMatrix type; +}; + +template <> +class ReturnType { + public: + typedef GpuIVector type; +}; + +template <> +class ReturnType { + public: + typedef GpuSparseMatrix type; +}; + +template +typename ReturnType::type autoArgs(T& v) { + return v; +} + +template <> +GpuMatrix autoArgs(CpuMatrix& v) { + GpuMatrix a(v.getHeight(), v.getWidth()); + a.copyFrom(v); + return a; +} + +template <> +GpuIVector autoArgs(CpuIVector& v) { + GpuIVector a(v.getSize()); + a.copyFrom(v); + return a; +} + +template <> +GpuSparseMatrix autoArgs(CpuSparseMatrix& v) { + GpuSparseMatrix a(v.getHeight(), + v.getWidth(), + v.getElementCnt(), + v.getValueType(), + v.getFormat()); + a.copyFrom(v, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + return a; +} + +class AutoCompare { + public: + /** + * err is the allowed calculation error. + * The smaller the value of err, + * the stricter the comparison is between CPU and GPU calculations. + */ + AutoCompare(size_t height, size_t width, real err = 1e-3) + : cpu(height, width), gpu(height, width), compare(err) { + init(cpu); + copy(gpu, cpu); + } + + template + void cmpWithArg(R (C::*f)(FArgs...), Args&&... args) { + static_assert(sizeof...(FArgs) == sizeof...(Args), + "size of parameter packs are not equal"); + call(cpu, f, args...); + call(gpu, f, autoArgs(args)...); + + TensorCheck(compare, cpu, gpu); + } + + template + void cmpWithoutArg(R (C::*f)(Args...), size_t height, size_t width) { + static_assert(sizeof...(I) == sizeof...(Args), + "size of parameter packs are not equal"); + (void)height; + (void)width; + auto tuple1 = std::make_tuple( + construct>::type>::type, + CpuMatrix>::type>(height, width)...); + + auto tuple2 = std::make_tuple( + construct>::type>::type, + GpuMatrix>::type>(height, width)...); + + initTuple(tuple1); + copyTuple(tuple2, tuple1); + + call(cpu, f, std::get(tuple1)...); + call(gpu, f, std::get(tuple2)...); + + TensorCheck(compare, cpu, gpu); + } + + protected: + CpuMatrix cpu; + GpuMatrix gpu; + AssertEqual compare; +}; + +} // namespace autotest diff --git a/paddle/legacy/math/tests/test_Allocator.cpp b/paddle/legacy/math/tests/test_Allocator.cpp new file mode 100644 index 0000000000000000000000000000000000000000..122be9082a8db33caf55661091caad115f575099 --- /dev/null +++ b/paddle/legacy/math/tests/test_Allocator.cpp @@ -0,0 +1,122 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Util.h" +#define private public +#include "paddle/legacy/math/Allocator.h" +#include "paddle/legacy/math/MemoryHandle.h" +#include "paddle/legacy/math/PoolAllocator.h" + +using namespace paddle; // NOLINT + +template +void testPoolAllocator() { + PoolAllocator* pool = + new PoolAllocator(new Allocator(), /* sizeLimit */ 1024); + + /* alloc from system memory */ + void* ptr1 = pool->alloc(10); + void* ptr2 = pool->alloc(200); + void* ptr3 = pool->alloc(200); + pool->free(ptr1, 10); + pool->free(ptr2, 200); + pool->free(ptr3, 200); + pool->printAll(); + EXPECT_EQ((size_t)2, pool->pool_.size()); + EXPECT_EQ((size_t)1, pool->pool_[10].size()); + EXPECT_EQ((size_t)2, pool->pool_[200].size()); + EXPECT_EQ(ptr1, pool->pool_[10][0]); + EXPECT_EQ(ptr2, pool->pool_[200][0]); + EXPECT_EQ(ptr3, pool->pool_[200][1]); + + /* alloc from pool */ + void* ptr4 = pool->alloc(10); + void* ptr5 = pool->alloc(200); + pool->printAll(); + EXPECT_EQ((size_t)0, pool->pool_[10].size()); + EXPECT_EQ((size_t)1, pool->pool_[200].size()); + EXPECT_EQ(ptr1, ptr4); + EXPECT_EQ(ptr3, ptr5); + pool->free(ptr4, 10); + pool->free(ptr5, 200); + + /* alloc size > sizeLimit */ + void* ptr6 = pool->alloc(1024); + pool->free(ptr6, 1024); + EXPECT_LE((size_t)1024, pool->poolMemorySize_); + + void* ptr7 = pool->alloc(1); + EXPECT_EQ((size_t)0, pool->poolMemorySize_); + EXPECT_EQ((size_t)0, pool->pool_.size()); + pool->free(ptr7, 1); + + delete pool; +} + +TEST(Allocator, Pool) { + testPoolAllocator(); +#ifdef PADDLE_WITH_CUDA + testPoolAllocator(); +#endif +} + +TEST(MemoryHandle, Cpu) { + for (auto size : {10, 30, 50, 100, 200, 512, 1000, 1023, 1024, 1025, 8193}) { + CpuMemoryHandle handle(size); + EXPECT_LE(handle.getSize(), handle.getAllocSize()); + } + + void* ptr1; + void* ptr2; + { + CpuMemoryHandle handle(256); + ptr1 = handle.getBuf(); + } + { + CpuMemoryHandle handle(256); + ptr2 = handle.getBuf(); + } + EXPECT_EQ(ptr1, ptr2); +} + +#ifdef PADDLE_WITH_CUDA +TEST(MemoryHandle, Gpu) { + int numGpu = hl_get_device_count(); + + /* alloc from system memory */ + void* ptr3[numGpu]; + void* ptr4[numGpu]; + for (int i = 0; i < numGpu; i++) { + SetDevice device(i); + GpuMemoryHandle handle1(30); + GpuMemoryHandle handle2(30); + GpuMemoryHandle handle3(4000); + GpuMemoryHandle handle4(500); + ptr3[i] = handle3.getBuf(); + ptr4[i] = handle4.getBuf(); + } + + /* alloc from pool */ + for (int i = 0; i < numGpu; i++) { + SetDevice device(i); + GpuMemoryHandle handle1(30); + GpuMemoryHandle handle3(4000); + GpuMemoryHandle handle4(500); + EXPECT_EQ(ptr3[i], handle3.getBuf()); + EXPECT_EQ(ptr4[i], handle4.getBuf()); + } +} +#endif diff --git a/paddle/legacy/math/tests/test_BaseMatrix.cpp b/paddle/legacy/math/tests/test_BaseMatrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..488765c6ac203ad064146faaab7b8c423d53cf0b --- /dev/null +++ b/paddle/legacy/math/tests/test_BaseMatrix.cpp @@ -0,0 +1,247 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +/** + * This test file use autotest::AutoCompare and cmpWithoutArg to compares the + * implementation of CPU and GPU member function in + * BaseMatrix.cpp and Matrix.cpp. + */ + +#include +#include "TestUtils.h" +#include "paddle/legacy/math/BaseMatrix.h" + +using paddle::BaseMatrix; +using paddle::Matrix; +using autotest::AutoCompare; + +// Test all void (BaseMatrix::*)() function +TEST(BaseMatrix, void) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + auto compare = [height, width](void (BaseMatrix::*f)()) { + AutoCompare test(height, width, 1e-5); + test.cmpWithoutArg(f, height, width); + }; + + compare(&BaseMatrix::neg); + compare(&BaseMatrix::exp2); + compare(&BaseMatrix::log2); + compare(&BaseMatrix::sqrt2); + compare(&BaseMatrix::square2); + compare(&BaseMatrix::reciprocal2); + compare(&BaseMatrix::abs2); + compare(&BaseMatrix::sign2); + compare(&BaseMatrix::zero); + compare(&BaseMatrix::one); + } + } +} + +// Test all void (BaseMatrix::*)(real) function +TEST(BaseMatrix, real) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + auto compare = [height, width](void (BaseMatrix::*f)(real)) { + AutoCompare test(height, width, 1e-5); + test.cmpWithoutArg<0>(f, height, width); + }; + + compare(&BaseMatrix::pow2); + compare(&BaseMatrix::subScalar); + compare(&BaseMatrix::mulScalar); + compare(&BaseMatrix::divScalar); + compare(&BaseMatrix::assign); + compare(&BaseMatrix::add); + compare(&BaseMatrix::biggerThanScalar); + compare(&BaseMatrix::downClip); + } + } +} + +// Test all void (BaseMatrix::*)(BaseMatrix&) function +TEST(BaseMatrix, BaseMatrix) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + auto compare = [height, width](void (BaseMatrix::*f)(BaseMatrix&)) { + AutoCompare test(height, width, 1e-5); + test.cmpWithoutArg<0>(f, height, width); + }; + + compare(&BaseMatrix::assign); + compare(&BaseMatrix::add); + compare(&BaseMatrix::relu); + compare(&BaseMatrix::reluDerivative); + compare(&BaseMatrix::softrelu); + compare(&BaseMatrix::softreluDerivative); + compare(&BaseMatrix::brelu); + compare(&BaseMatrix::breluDerivative); + compare(&BaseMatrix::square2); + compare(&BaseMatrix::squareDerivative); + compare(&BaseMatrix::tanh); + compare(&BaseMatrix::tanhDerivative); + compare(&BaseMatrix::reciprocal2); + compare(&BaseMatrix::reciprocalDerivative); + compare(&BaseMatrix::abs2); + compare(&BaseMatrix::absDerivative); + compare(&BaseMatrix::sigmoid); + compare(&BaseMatrix::sigmoidDerivative); + compare(&BaseMatrix::expDerivative); + compare(&BaseMatrix::sign2); + compare(&BaseMatrix::exp2); + compare(&BaseMatrix::log2); + compare(&BaseMatrix::sqrt2); + compare(&BaseMatrix::dotMul); + compare(&BaseMatrix::dotMulSquare); + compare(&BaseMatrix::dotSquareMul); + compare(&BaseMatrix::addColVector); + compare(&BaseMatrix::addRowVector); + compare(&BaseMatrix::mulRowVector); + compare(&BaseMatrix::divRowVector); + compare(&BaseMatrix::mulColVector); + compare(&BaseMatrix::divColVector); + compare(&BaseMatrix::addP2P); + compare(&BaseMatrix::invSqrt); + } + } +} + +// Test all void (BaseMatrix::*)(real, real) function +TEST(BaseMatrix, real_real) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + auto compare = [height, width](void (BaseMatrix::*f)(real, real)) { + AutoCompare test(height, width, 1e-5); + test.cmpWithoutArg<0, 1>(f, height, width); + }; + + compare(&BaseMatrix::add); + compare(&BaseMatrix::clip); + } + } +} + +// Test all void (BaseMatrix::*)(BaseMatrix&, real) function +TEST(BaseMatrix, BaseMatrix_real) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + auto compare = [height, width](void (BaseMatrix::*f)(BaseMatrix&, real)) { + AutoCompare test(height, width, 1e-5); + test.cmpWithoutArg<0, 1>(f, height, width); + }; + + compare(&BaseMatrix::addBias); + compare(&BaseMatrix::add); + compare(&BaseMatrix::sub); + compare(&BaseMatrix::pow2); + compare(&BaseMatrix::addScalar); + compare(&BaseMatrix::subScalar); + compare(&BaseMatrix::mulScalar); + compare(&BaseMatrix::divScalar); + compare(&BaseMatrix::scalarDiv); + compare(&BaseMatrix::addSquare); + compare(&BaseMatrix::isEqualTo); + } + } +} + +// Test all void (BaseMatrix::*)(BaseMatrix&, BaseMatrix&) function +TEST(BaseMatrix, BaseMatrix_BaseMatrix) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + auto compare = [height, + width](void (BaseMatrix::*f)(BaseMatrix&, BaseMatrix&)) { + AutoCompare test(height, width, 1e-5); + test.cmpWithoutArg<0, 1>(f, height, width); + }; + + compare(&BaseMatrix::softCrossEntropy); + compare(&BaseMatrix::softCrossEntropyBp); + compare(&BaseMatrix::binaryLabelCrossEntropy); + compare(&BaseMatrix::binaryLabelCrossEntropyBp); + compare(&BaseMatrix::sub); + compare(&BaseMatrix::add2); + compare(&BaseMatrix::dotMul); + compare(&BaseMatrix::dotDiv); + compare(&BaseMatrix::logisticRegressionLoss); + compare(&BaseMatrix::logisticRegressionLossBp); + compare(&BaseMatrix::biggerThan); + compare(&BaseMatrix::max2); + compare(&BaseMatrix::dotMulSquare); + compare(&BaseMatrix::dotSquareSquare); + } + } +} + +void TestEelementWise(size_t height, size_t width) { + AutoCompare rowScale(height, width); + rowScale.cmpWithoutArg<0, 1, 2>(&BaseMatrix::rowScale, height, width); + + AutoCompare rowDotMul(height, width); + rowDotMul.cmpWithoutArg<0, 1, 2>(&BaseMatrix::rowDotMul, height, width); + + AutoCompare binaryClassificationError(height, width); + binaryClassificationError.cmpWithoutArg<0, 1, 2, 3>( + &BaseMatrix::binaryClassificationError, height, width); + + AutoCompare sumOfSquaresBp(height, width); + sumOfSquaresBp.cmpWithoutArg<0, 1>(&Matrix::sumOfSquaresBp, height, width); +} + +void TestAggregateToRow(size_t height, size_t width) { + AutoCompare maxCols(1, width); + maxCols.cmpWithoutArg<0>(&BaseMatrix::maxCols, height, width); + + AutoCompare minCols(1, width); + minCols.cmpWithoutArg<0>(&BaseMatrix::minCols, height, width); + + AutoCompare addDotMulVMM(1, width); + addDotMulVMM.cmpWithoutArg<0, 1>(&BaseMatrix::addDotMulVMM, height, width); + + AutoCompare sumCols(1, width); + sumCols.cmpWithoutArg<0, 1, 2>(&BaseMatrix::sumCols, height, width); + + AutoCompare collectBias(1, width); + collectBias.cmpWithoutArg<0, 1>( + static_cast(&Matrix::collectBias), + height, + width); +} + +void TestAggregateToCol(size_t height, size_t width) { + AutoCompare maxRows(height, 1); + maxRows.cmpWithoutArg<0>(&BaseMatrix::maxRows, height, width); + + AutoCompare minRows(height, 1); + minRows.cmpWithoutArg<0>(&BaseMatrix::minRows, height, width); + + AutoCompare sumRows(height, 1); + sumRows.cmpWithoutArg<0, 1, 2>(&BaseMatrix::sumRows, height, width); + + AutoCompare sumOfSquares(height, 1); + sumOfSquares.cmpWithoutArg<0, 1>(&Matrix::sumOfSquares, height, width); +} + +TEST(BaseMatrix, Other) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + TestEelementWise(height, width); + TestAggregateToRow(height, width); + TestAggregateToCol(height, width); + } + } +} + +#endif diff --git a/paddle/legacy/math/tests/test_CpuGpuVector.cpp b/paddle/legacy/math/tests/test_CpuGpuVector.cpp new file mode 100644 index 0000000000000000000000000000000000000000..010fef534d1e19d2d7d134298eb97aa1b56e2270 --- /dev/null +++ b/paddle/legacy/math/tests/test_CpuGpuVector.cpp @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA + +#include +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Util.h" +#include "test_matrixUtil.h" + +using namespace paddle; // NOLINT + +TEST(CpuGpuVector, getData) { + size_t size = 500; + hl_stream_t stream(HPPL_STREAM_DEFAULT); + CpuVectorPtr cpuVec = std::make_shared(size); + GpuVectorPtr gpuVec = std::make_shared(size); + cpuVec->uniform(0.0, 10.0); + gpuVec->copyFrom(*cpuVec, stream); + hl_stream_synchronize(stream); + + CpuGpuVectorPtr vec = std::make_shared(gpuVec); + auto a = vec->getData(false); + auto b = cpuVec->getData(); + hl_stream_synchronize(stream); + checkDataEqual(a, b, size); +} + +TEST(CpuGpuVector, subCreate) { + size_t size1 = 1024; + size_t offset = 100; + size_t size2 = 500; + hl_stream_t stream(HPPL_STREAM_DEFAULT); + CpuGpuVectorPtr v1 = std::make_shared(size1, /*useGpu*/ false); + auto vec = v1->getMutableVector(false); + vec->uniform(0.0, 10.0); + auto v2 = std::make_shared(*v1, offset, size2); + CHECK_EQ(*v1->getSync(), *v2->getSync()); + + // check subVec equal + checkDataEqual(v1->getData(false) + offset, v2->getData(false), size2); + + CpuVectorPtr v1Check = std::make_shared(size1); + CpuVectorPtr v2Check = std::make_shared(size2); + v1Check->copyFrom(*(v1->getVector(true)), stream); + v2Check->copyFrom(*(v2->getVector(true)), stream); + hl_stream_synchronize(stream); + + checkDataEqual(v2->getData(false), v2Check->getData(), size2); + checkDataEqual(v1Check->getData() + offset, v2Check->getData(), size2); + + CpuVectorPtr noise = std::make_shared(size2); + noise->uniform(0.0, 1.0); + auto v = v2->getMutableVector(false); // will change header + // add noise to subVec + v->add(*noise); + + // check v1_cpu_data == v2_cpu_data + checkDataEqual(v1->getData(false) + offset, v2->getData(false), size2); + + v1Check->copyFrom(*(v1->getVector(true)), stream); + v2Check->copyFrom(*(v2->getVector(true)), stream); + hl_stream_synchronize(stream); + + // check v1_gpu_data == v2_gpu_data + checkDataEqual(v1Check->getData() + offset, v2Check->getData(), size2); +} + +#endif diff --git a/paddle/legacy/math/tests/test_ExecViaCpu.cpp b/paddle/legacy/math/tests/test_ExecViaCpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b2ce0bc7ede133028fff8a855ff336ff83f55d82 --- /dev/null +++ b/paddle/legacy/math/tests/test_ExecViaCpu.cpp @@ -0,0 +1,116 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include "paddle/legacy/math/SparseMatrix.h" + +using namespace paddle; // NOLINT + +const int height = 10; +const int width = 16; + +real f(Matrix& mat1, + const Matrix& mat2, + IVector& vec1, + const IVector& vec2, + real scalar) { + CHECK(!mat1.useGpu()); + CHECK(!mat2.useGpu()); + CHECK(!vec1.useGpu()); + CHECK(!vec2.useGpu()); + mat1.copyFrom(mat2); + vec1.copyFrom(vec2); + + return scalar; +} + +class Functor { + public: + real operator()(Matrix& mat1, + const Matrix& mat2, + IVector& vec1, + const IVector& vec2, + real scalar) { + a_ = f(mat1, mat2, vec1, vec2, scalar); + return a_; + } + + private: + real a_; +}; + +template +void testWrapper(F&& f) { + MatrixPtr cpumat1 = Matrix::create(height, width, false, /*useGpu=*/false); + MatrixPtr cpumat2 = Matrix::create(height, width, false, /*useGpu=*/false); + + IVectorPtr cpuvec1 = IVector::create(height, /*useGpu=*/false); + IVectorPtr cpuvec2 = IVector::create(height, /*useGpu=*/false); + + const real scalar = 1.23456; + + MatrixPtr gpumat1 = Matrix::create(height, width, false, /*useGpu=*/true); + MatrixPtr gpumat2 = Matrix::create(height, width, false, /*useGpu=*/true); + IVectorPtr gpuvec1 = IVector::create(height, /*useGpu=*/true); + IVectorPtr gpuvec2 = IVector::create(height, /*useGpu=*/true); + + cpumat2->randomizeUniform(); + cpuvec2->rand(width); + gpumat2->copyFrom(*cpumat2); + gpuvec2->copyFrom(*cpuvec2); + + real ret = execViaCpu(f, *gpumat1, *gpumat2, *gpuvec1, *gpuvec2, 1.23456); + EXPECT_EQ(ret, scalar); + cpumat1->copyFrom(*gpumat1); + cpuvec1->copyFrom(*gpuvec1); + + for (int i = 0; i < height; ++i) { + EXPECT_EQ(cpuvec1->getElement(i), cpuvec2->getElement(i)); + for (int j = 0; j < width; ++j) { + EXPECT_EQ(cpumat1->getElement(i, j), cpumat2->getElement(i, j)); + } + } + gpumat1->resize(height, 1); + execViaCpu2(&CpuMatrix::selectElements, *gpumat1, *gpumat2, *gpuvec1); + + cpumat1->resize(height, 1); + cpumat1->selectElements(*cpumat2, *cpuvec1); + for (int i = 0; i < height; ++i) { + EXPECT_EQ(cpumat1->getElement(i, 0), gpumat1->getElement(i, 0)); + } +} + +#ifdef PADDLE_WITH_CUDA +TEST(ExecViaCpu, test1) { + testWrapper(f); + testWrapper(&f); + + auto lambda = [](Matrix& mat1, + const Matrix& mat2, + IVector& vec1, + const IVector& vec2, + real scalar) -> real { + return f(mat1, mat2, vec1, vec2, scalar); + }; + LOG(INFO) << "lambda is_class=" << std::is_class::value + << " is_function=" << std::is_function::value; + testWrapper(lambda); + + Functor functor; + testWrapper(functor); +} +#endif diff --git a/paddle/legacy/math/tests/test_FPException.cpp b/paddle/legacy/math/tests/test_FPException.cpp new file mode 100644 index 0000000000000000000000000000000000000000..aa6aea71c8d959834ff11c04969e13bb36b630ff --- /dev/null +++ b/paddle/legacy/math/tests/test_FPException.cpp @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/** + * This test is about floating point calculation exception. + * Paddle catches FE_INVALID, FE DIVBYZERO and FE_OVERFLOW exceptions. + * + * Some exceptions occur in the middle of a set of formulas, + * that can be circumvented by some tricks. + * For example, + * calculate tanh + * b = 2.0 / (1.0 + exp(-2 * a)) - 1.0 + * + * If the result of (-2 * a) is too large, + * a FE_OVERFLOW exception occurs when calculating exp. + * But the result of tanh is no overflow problem, + * so we can add some tricks to prevent exp calculate an excessive value. + * + */ + +#include +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Common.h" + +using namespace paddle; // NOLINT + +void SetTensorValue(Matrix& matrix, real value) { + int height = matrix.getHeight(); + int width = matrix.getWidth(); + int stride = matrix.getStride(); + real* data = matrix.getData(); + for (int i = 0; i < height; i++) { + int j = rand() % width; // NOLINT + if (typeid(matrix) == typeid(CpuMatrix)) { + data[i * stride + j] = value; + } else if (typeid(matrix) == typeid(GpuMatrix)) { + hl_memcpy(&data[i * stride + j], &value, sizeof(real)); + } else { + LOG(FATAL) << "should not reach here"; + } + } +} + +template +void testTanh(real illegal) { + MatrixPtr A = std::make_shared(10, 10); + MatrixPtr B = std::make_shared(10, 10); + A->randomizeUniform(); + B->randomizeUniform(); + + SetTensorValue(*A, illegal); + + A->tanh(*B); +} + +template +void testSigmoid(real illegal) { + MatrixPtr A = std::make_shared(10, 10); + MatrixPtr B = std::make_shared(10, 10); + A->randomizeUniform(); + B->randomizeUniform(); + + SetTensorValue(*A, illegal); + + A->sigmoid(*B); +} + +TEST(fp, overflow) { + for (auto illegal : {-90.0, 90.0}) { + LOG(INFO) << " illegal=" << illegal; + testTanh(illegal); + testSigmoid(illegal); + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + + feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/math/tests/test_GpuProfiler.cpp b/paddle/legacy/math/tests/test_GpuProfiler.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ee27109f218ca56df8f42ca6395b22621f5fbc11 --- /dev/null +++ b/paddle/legacy/math/tests/test_GpuProfiler.cpp @@ -0,0 +1,165 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA + +#include +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +void MatrixCheckErr(const Matrix& matrix1, const Matrix& matrix2) { + CHECK(matrix1.getHeight() == matrix2.getHeight()); + CHECK(matrix1.getWidth() == matrix2.getWidth()); +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + + int height = matrix1.getHeight(); + int width = matrix1.getWidth(); + const real* data1 = matrix1.getData(); + const real* data2 = matrix2.getData(); + int count = 0; + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + real a = data1[i * width + j]; + real b = data2[i * width + j]; + if (fabs(a - b) > err) { + if ((fabsf(a - b) / fabsf(a)) > (err / 10.0f)) { + count++; + } + } + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; +} + +void testBilinearFwdBwd(int numSamples, + int imgSizeH, + int imgSizeW, + int channels) { + int inWidth = imgSizeH * imgSizeW * channels; + int outWidth = 2 * imgSizeH * 2 * imgSizeW * channels; + real ratioH = 0.5; + real ratioW = 0.5; + + // forward + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); + + input->randomizeUniform(); + inputGpu->copyFrom(*input); + + { + // nvprof: GPU Proflier + REGISTER_GPU_PROFILER("testBilinearFwdBwd"); + target->bilinearForward(*input, + imgSizeH, + imgSizeW, + 2 * imgSizeH, + 2 * imgSizeW, + channels, + ratioH, + ratioW); + targetGpu->bilinearForward(*inputGpu, + imgSizeH, + imgSizeW, + 2 * imgSizeH, + 2 * imgSizeW, + channels, + ratioH, + ratioW); + } + + // check + targetCheck->copyFrom(*targetGpu); + MatrixCheckErr(*target, *targetCheck); + + // backward + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + MatrixPtr targetCheckGrad = + CpuMatrix::create(numSamples, inWidth, false, false); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->bilinearBackward(*targetGrad, + 2 * imgSizeH, + 2 * imgSizeW, + imgSizeH, + imgSizeW, + channels, + ratioH, + ratioW); + inputGpuGrad->bilinearBackward(*targetGpuGrad, + 2 * imgSizeH, + 2 * imgSizeW, + imgSizeH, + imgSizeW, + channels, + ratioH, + ratioW); + + // check + targetCheckGrad->copyFrom(*inputGpuGrad); + MatrixCheckErr(*inputGrad, *targetCheckGrad); +} + +TEST(Profiler, testBilinearFwdBwd) { + auto numSamples = 10; + auto channels = 16; + auto imgSize = 64; + { + // nvprof: GPU Proflier + REGISTER_GPU_PROFILER("testBilinearFwdBwd"); + // Paddle built-in timer + REGISTER_TIMER_INFO( + "testBilinearFwdBwd", + "numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"); + testBilinearFwdBwd(numSamples, imgSize, imgSize, channels); + } + globalStat.printAllStatus(); +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + + // nvprof: GPU Proflier + REGISTER_GPU_PROFILER( + "RecursiveProfilingTest", + "numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"); + + return RUN_ALL_TESTS(); +} + +#endif diff --git a/paddle/math/tests/test_Matrix.cpp b/paddle/legacy/math/tests/test_Matrix.cpp similarity index 100% rename from paddle/math/tests/test_Matrix.cpp rename to paddle/legacy/math/tests/test_Matrix.cpp diff --git a/paddle/legacy/math/tests/test_RowBuffer.cpp b/paddle/legacy/math/tests/test_RowBuffer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2ef8cd303d65f50cd18adb7f80fa18a665b67340 --- /dev/null +++ b/paddle/legacy/math/tests/test_RowBuffer.cpp @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/math/RowBuffer.h" + +TEST(RowBuffer, testAutoGrow) { + paddle::RowBuffer buf(128); + ASSERT_EQ(128UL, buf.getWidth()); + ASSERT_TRUE(buf.isAutoGrowth()); + buf.resize(2); + ASSERT_EQ(2UL, buf.getRowCount()); + for (size_t i = 0; i < buf.getWidth() * 2; ++i) { + buf.data()[i] = i; + } + for (size_t i = 0; i < buf.getRowCount(); ++i) { + for (size_t j = 0; j < buf.getWidth(); ++j) { + ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5); + } + } + + auto data = buf.getWithAutoGrowth(2); + for (size_t i = 0; i < buf.getWidth(); ++i) { + data[i] = i; + } + + ASSERT_EQ(3UL, buf.getRowCount()); + for (size_t i = 0; i < buf.getRowCount() - 1; ++i) { + for (size_t j = 0; j < buf.getWidth(); ++j) { + ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5); + } + } + for (size_t i = 0; i < buf.getWidth(); ++i) { + ASSERT_NEAR(i, buf.get(2)[i], 1e-5); + } +} + +TEST(RowBuffer, testWithMemBuf) { + paddle::CpuMemHandlePtr mem = + std::make_shared(128 * 2 * sizeof(real)); + paddle::RowBuffer buf(mem, 128); + ASSERT_TRUE(!buf.isAutoGrowth()); + ASSERT_EQ(2UL, buf.getRowCount()); + for (size_t i = 0; i < buf.getWidth() * 2; ++i) { + buf.data()[i] = i; + } + for (size_t i = 0; i < buf.getRowCount(); ++i) { + for (size_t j = 0; j < buf.getWidth(); ++j) { + ASSERT_NEAR(i * buf.getWidth() + j, buf.getWithAutoGrowth(i)[j], 1e-5); + } + } + + ASSERT_DEATH_IF_SUPPORTED(buf.getWithAutoGrowth(3), ".*"); +} diff --git a/paddle/legacy/math/tests/test_SIMDFunctions.cpp b/paddle/legacy/math/tests/test_SIMDFunctions.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c6490f70e336dadcf6710c83ced2afddc13b7812 --- /dev/null +++ b/paddle/legacy/math/tests/test_SIMDFunctions.cpp @@ -0,0 +1,171 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/math/SIMDFunctions.h" +#include "paddle/legacy/utils/Util.h" + +#include + +#include +#include +#include +#include + +#include +#include + +static constexpr size_t VECTOR_LEN = 3072; +static constexpr size_t BATCH_SIZE = 64; +static constexpr size_t ALIGN = 32; +static_assert(VECTOR_LEN % ALIGN == 0, "VECTOR_LEN % ALIGN == 0"); +static_assert(BATCH_SIZE % ALIGN == 0, "BATCH_SIZE % ALIGN == 0"); +static constexpr float EPSILON = 1e-5; +static std::mt19937 RandomEngine(time(0)); + +inline static std::unique_ptr NewVector(size_t len = VECTOR_LEN, + size_t align = ALIGN) { + float* ptr; + CHECK_EQ(posix_memalign((void**)&ptr, align, len * sizeof(float)), 0); + return std::unique_ptr(ptr); +} + +inline static std::unique_ptr NewRandomVector(size_t len = VECTOR_LEN, + size_t align = ALIGN) { + std::uniform_real_distribution dist(-100.0f, 100.0f); + auto generator = std::bind(dist, RandomEngine); + auto retv = NewVector(len, align); + std::generate_n(retv.get(), len, generator); + return retv; +} + +TEST(SIMDFunction, addTo) { + typedef std::function AddToMethodType; + + AddToMethodType naive = paddle::simd::naive::addTo; + AddToMethodType simd = paddle::simd::addTo; + + auto A = NewRandomVector(); + auto B = NewRandomVector(); + + auto ACopy = NewVector(); + memcpy(ACopy.get(), A.get(), VECTOR_LEN * sizeof(float)); + + naive(A.get(), B.get(), VECTOR_LEN); + simd(ACopy.get(), B.get(), VECTOR_LEN); + + for (size_t i = 0; i < VECTOR_LEN; ++i) { + ASSERT_NEAR(A[i], ACopy[i], EPSILON); + } +} + +TEST(SIMDFunction, batchAddTo) { + auto A = NewRandomVector(); + auto ACopy = NewVector(); + memcpy(ACopy.get(), A.get(), sizeof(float) * VECTOR_LEN); + + std::vector> B; + for (size_t i = 0; i < BATCH_SIZE; ++i) { + B.emplace_back(NewRandomVector()); + } + std::unique_ptr BRaw(new float*[BATCH_SIZE]); + for (size_t i = 0; i < BATCH_SIZE; ++i) { + BRaw[i] = B[i].get(); + } + + typedef std::function + BatchAddToMethodType; + + BatchAddToMethodType naive = paddle::simd::naive::batchAddTo; + BatchAddToMethodType simd = paddle::simd::batchAddTo; + + naive(A.get(), (const float**)BRaw.get(), BATCH_SIZE, VECTOR_LEN); + simd(ACopy.get(), (const float**)BRaw.get(), BATCH_SIZE, VECTOR_LEN); + + for (size_t i = 0; i < VECTOR_LEN; ++i) { + ASSERT_NEAR(A[i], ACopy[i], EPSILON); + } +} + +TEST(SIMDFunction, colMax) { + auto A = NewRandomVector(VECTOR_LEN * BATCH_SIZE); + auto naiveResult = NewVector(BATCH_SIZE); + auto simdResult = NewVector(BATCH_SIZE); + + typedef std::function ColMaxMethodType; + ColMaxMethodType naive = paddle::simd::naive::colMax; + ColMaxMethodType simd = paddle::simd::colMax; + + naive(naiveResult.get(), A.get(), BATCH_SIZE, VECTOR_LEN); + simd(simdResult.get(), A.get(), BATCH_SIZE, VECTOR_LEN); + + for (size_t i = 0; i < BATCH_SIZE; ++i) { + ASSERT_NEAR(naiveResult[i], simdResult[i], EPSILON); + } +} + +TEST(SIMDFunction, decayL1_WithLR) { + auto dest = NewRandomVector(); + auto src = NewRandomVector(); + auto lr = NewRandomVector(); + auto lambda = 0.23f; + + auto simd_dest = NewVector(); + memcpy(simd_dest.get(), dest.get(), sizeof(float) * VECTOR_LEN); + + typedef std::function + DecayL1MethodType; + + DecayL1MethodType naive = []( + float* d, float* s, float* lr, float l, size_t len) { + paddle::simd::naive::decayL1(d, s, lr, l, len); + }; + + DecayL1MethodType simd = []( + float* d, float* s, float* lr, float l, size_t len) { + paddle::simd::decayL1(d, s, lr, l, len); + }; + + naive(dest.get(), src.get(), lr.get(), lambda, VECTOR_LEN); + simd(simd_dest.get(), src.get(), lr.get(), lambda, VECTOR_LEN); + + for (size_t i = 0; i < VECTOR_LEN; ++i) { + ASSERT_NEAR(dest[i], simd_dest[i], EPSILON); + } +} + +TEST(SIMDFunction, decayL1_WithoutLR) { + auto dest = NewRandomVector(); + auto src = NewRandomVector(); + auto lambda = 0.23; + + auto simd_dest = NewVector(); + memcpy(simd_dest.get(), dest.get(), sizeof(float) * VECTOR_LEN); + + typedef std::function DecayL1MethodType; + + DecayL1MethodType naive = [](float* d, float* s, float l, size_t len) { + paddle::simd::naive::decayL1(d, s, l, len); + }; + + DecayL1MethodType simd = [](float* d, float* s, float l, size_t len) { + paddle::simd::decayL1(d, s, l, len); + }; + + naive(dest.get(), src.get(), lambda, VECTOR_LEN); + simd(simd_dest.get(), src.get(), lambda, VECTOR_LEN); + + for (size_t i = 0; i < VECTOR_LEN; ++i) { + ASSERT_NEAR(dest[i], simd_dest[i], EPSILON); + } +} diff --git a/paddle/legacy/math/tests/test_SparseMatrix.cpp b/paddle/legacy/math/tests/test_SparseMatrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..30896a945ec6d111c35eea94d8008a62593d2893 --- /dev/null +++ b/paddle/legacy/math/tests/test_SparseMatrix.cpp @@ -0,0 +1,565 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "test_matrixUtil.h" + +using namespace paddle; // NOLINT + +TEST(Matrix, CopyCpuMatrixToSparseMatrix) { + const size_t HEIGHT = 20; + const size_t WIDTH = 10; + const size_t WIDTH_TEST = 15; + MatrixPtr testMatrix( + new CpuSparseMatrix(HEIGHT, WIDTH, HEIGHT * 5, FLOAT_VALUE, SPARSE_CSR)); + MatrixPtr testCpuMatrix(new CpuMatrix(HEIGHT, WIDTH)); + testCpuMatrix->randomizeUniform(); + testMatrix->copyFrom(*testCpuMatrix, HPPL_STREAM_DEFAULT); + MatrixPtr mulCpuMatrix(new CpuMatrix(WIDTH, WIDTH_TEST)); + mulCpuMatrix->randomizeUniform(); + MatrixPtr ret1(new CpuMatrix(HEIGHT, WIDTH_TEST)), + ret2(new CpuMatrix(HEIGHT, WIDTH_TEST)); + ret1->zeroMem(); + ret2->zeroMem(); + ret1->mul(*testMatrix, *mulCpuMatrix, 1.0, 1.0); + ret2->mul(*testCpuMatrix, *mulCpuMatrix, 1.0, 1.0); + checkMatrixEqual(ret1, ret2); +} + +struct MatrixPara { + size_t height; + size_t width; + bool trans; + bool sparse; + size_t nnz; + SparseFormat format; +}; + +#ifdef PADDLE_WITH_CUDA +void test_sparse_matrix_mul(MatrixPara paraA, + MatrixPara paraB, + MatrixPara paraC) { + // for cpu sparse matrix mul + MatrixPtr cpuMatrixA, cpuMatrixB, cpuMatrixC, gpuMatrixC_d2h; + // for gpu sparse matrix mul + MatrixPtr gpuMatrixA, gpuMatrixB, gpuMatrixC; + // for cpu dense matrix mul + MatrixPtr cpuDenseA, cpuDenseB, cpuDenseC; + + if (paraA.sparse) { + cpuMatrixA = Matrix::createSparseMatrix(paraA.height, + paraA.width, + paraA.nnz, + FLOAT_VALUE, + paraA.format, + paraA.trans, + false); + gpuMatrixA = Matrix::createSparseMatrix(paraA.height, + paraA.width, + paraA.nnz, + FLOAT_VALUE, + paraA.format, + paraA.trans, + true); + } else { + cpuMatrixA = Matrix::create(paraA.height, paraA.width, paraA.trans, false); + gpuMatrixA = Matrix::create(paraA.height, paraA.width, paraA.trans, true); + } + cpuDenseA = Matrix::create(paraA.height, paraA.width, paraA.trans, false); + + if (paraB.sparse) { + cpuMatrixB = Matrix::createSparseMatrix(paraB.height, + paraB.width, + paraB.nnz, + FLOAT_VALUE, + paraB.format, + paraB.trans, + false); + gpuMatrixB = Matrix::createSparseMatrix(paraB.height, + paraB.width, + paraB.nnz, + FLOAT_VALUE, + paraB.format, + paraB.trans, + true); + } else { + cpuMatrixB = Matrix::create(paraB.height, paraB.width, paraB.trans, false); + gpuMatrixB = Matrix::create(paraB.height, paraB.width, paraB.trans, true); + } + cpuDenseB = Matrix::create(paraB.height, paraB.width, paraB.trans, false); + + if (paraC.sparse) { + cpuMatrixC = Matrix::createSparseMatrix(paraC.height, + paraC.width, + paraC.nnz, + FLOAT_VALUE, + paraC.format, + paraC.trans, + false); + gpuMatrixC = Matrix::createSparseMatrix(paraC.height, + paraC.width, + paraC.nnz, + FLOAT_VALUE, + paraC.format, + paraC.trans, + true); + gpuMatrixC_d2h = Matrix::createSparseMatrix(paraC.height, + paraC.width, + paraC.nnz, + FLOAT_VALUE, + paraC.format, + paraC.trans, + false); + } else { + cpuMatrixC = Matrix::create(paraC.height, paraC.width, paraC.trans, false); + gpuMatrixC = Matrix::create(paraC.height, paraC.width, paraC.trans, true); + gpuMatrixC_d2h = + Matrix::create(paraC.height, paraC.width, paraC.trans, false); + } + cpuDenseC = Matrix::create(paraC.height, paraC.width, paraC.trans, false); + + /*matrix init*/ + hl_stream_t stream(HPPL_STREAM_1); + cpuMatrixA->randomizeUniform(); + cpuMatrixB->randomizeUniform(); + cpuMatrixC->randomizeUniform(); + + gpuMatrixA->copyFrom(*cpuMatrixA, stream); + gpuMatrixB->copyFrom(*cpuMatrixB, stream); + gpuMatrixC->copyFrom(*cpuMatrixC, stream); + + cpuDenseA->copyFrom(*cpuMatrixA); + cpuDenseB->copyFrom(*cpuMatrixB); + cpuDenseC->copyFrom(*cpuMatrixC); + + hl_stream_synchronize(stream); + + /*matrix mul*/ + cpuMatrixC->mul(*cpuMatrixA, *cpuMatrixB, 1.0, 1.0); + gpuMatrixC->mul(*gpuMatrixA, *gpuMatrixB, 1.0, 1.0); + cpuDenseC->mul(*cpuDenseA, *cpuDenseB, 1.0, 1.0); + + gpuMatrixC_d2h->copyFrom(*gpuMatrixC, stream); + hl_stream_synchronize(stream); + + /*check result*/ + if (paraC.sparse) { + checkSMatrixEqual( + std::dynamic_pointer_cast(cpuMatrixC), + std::dynamic_pointer_cast(gpuMatrixC_d2h)); + checkSMatrixEqual2Dense( + std::dynamic_pointer_cast(cpuMatrixC), + std::dynamic_pointer_cast(cpuDenseC)); + } else { + checkMatrixEqual(cpuMatrixC, gpuMatrixC_d2h); + checkMatrixEqual(cpuMatrixC, cpuDenseC); + } +} + +TEST(Matrix, SparseMatrixMul) { + const size_t DIM_M = 4; + const size_t DIM_N = 4; + const size_t DIM_K = 8; + const size_t NNZ = 5; + for (auto format : {SPARSE_CSC, SPARSE_CSR}) { + std::string str_format = format == SPARSE_CSC ? "CSC" : "CSR"; + LOG(INFO) << "test dense mul " << str_format; + test_sparse_matrix_mul( + {DIM_M, DIM_K, /*trans*/ false, /*sparse*/ false, NNZ, format}, + {DIM_K, DIM_N, /*trans*/ false, /*sparse*/ true, NNZ, format}, + {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}); + + LOG(INFO) << "test dense mul " << str_format << " trans"; + test_sparse_matrix_mul( + {DIM_M, DIM_K, /*trans*/ false, /*sparse*/ false, NNZ, format}, + {DIM_N, DIM_K, /*trans*/ true, /*sparse*/ true, NNZ, format}, + {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}); + + LOG(INFO) << "test dense mul dense 2 " << str_format; + test_sparse_matrix_mul( + {DIM_M, DIM_K, /*trans*/ false, /*sparse*/ false, NNZ, format}, + {DIM_K, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}, + {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ true, NNZ, format}); + + LOG(INFO) << "test denseT mul dense 2 " << str_format; + test_sparse_matrix_mul( + {DIM_K, DIM_M, /*trans*/ true, /*sparse*/ false, NNZ, format}, + {DIM_K, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}, + {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ true, NNZ, format}); + } +} + +TEST(Matrix, CopySparseMatrixToGpuSparseMatrix) { + const size_t HEIGHT = 20; + const size_t WIDTH = 10; + const size_t WIDTH_TEST = 15; + MatrixPtr testMatrix( + new CpuSparseMatrix(HEIGHT, WIDTH, HEIGHT * 2, FLOAT_VALUE, SPARSE_CSR)); + MatrixPtr testCpuMatrix(new CpuMatrix(HEIGHT, WIDTH)); + testCpuMatrix->randomizeUniform(); + testMatrix->copyFrom(*testCpuMatrix, HPPL_STREAM_DEFAULT); + + MatrixPtr testGpuMatrix = testMatrix->clone(HEIGHT, WIDTH, true); + hl_stream_t gpuStream(HPPL_STREAM_3); + testGpuMatrix->copyFrom(*testMatrix, gpuStream); + hl_stream_synchronize(gpuStream); + + MatrixPtr mulCpuMatrix(new CpuMatrix(WIDTH, WIDTH_TEST)); + mulCpuMatrix->randomizeUniform(); + MatrixPtr mulGpuMatrix(new GpuMatrix(WIDTH, WIDTH_TEST)); + mulGpuMatrix->copyFrom(*mulCpuMatrix); + MatrixPtr ret1(new CpuMatrix(HEIGHT, WIDTH_TEST)); + MatrixPtr ret2(new GpuMatrix(HEIGHT, WIDTH_TEST)); + ret1->zeroMem(); + ret2->zeroMem(); + ret1->mul(*testMatrix, *mulCpuMatrix, 1.0, 1.0); + ret2->mul(*testGpuMatrix, *mulGpuMatrix, 1.0, 1.0); + checkMatrixEqual(ret1, ret2); +} + +#endif + +TEST(Matrix, SparseMatrixTranspose) { + for (auto height : {10, 50, 100}) { + for (auto width : {10, 50, 100}) { + auto nnz = height * width; + for (auto valueType : {FLOAT_VALUE, NO_VALUE}) { + for (auto format : {SPARSE_CSR, SPARSE_CSC}) { + for (auto sparseRate : {0.1, 0.2, 0.5}) { + MatrixPtr matA = Matrix::createSparseMatrix( + height, width, size_t(nnz * sparseRate), valueType, format); + MatrixPtr matB(new CpuSparseMatrix( + width, height, size_t(nnz * sparseRate), valueType, format)); + matA->randomizeUniform(); + matA->transpose(matB, false); + + /*dense matrix transpose*/ + CpuMatrixPtr matC(new CpuMatrix(height, width)); + matC->copyFrom(*matA); + MatrixPtr matD(new CpuMatrix(width, height)); + matC->transpose(matD, false); + + /*check result*/ + checkSMatrixEqual2Dense( + std::dynamic_pointer_cast(matB), + std::dynamic_pointer_cast(matD)); + } + } + } + } + } +} + +TEST(Matrix, CpuSparseMatrixSubMatrix) { + const size_t HEIGHT = 10; + const size_t WIDTH = 10; + const size_t NNZ = HEIGHT * WIDTH; + for (auto valueType : {FLOAT_VALUE, NO_VALUE}) { + size_t startRow = 3; + size_t rowNum = 2; + real sparseRate = 0.1; + /*sparse matrix init and get subMatrix*/ + CpuSparseMatrixPtr matA = std::make_shared( + HEIGHT, WIDTH, size_t(NNZ * sparseRate), valueType, SPARSE_CSR); + matA->randomizeUniform(); + CpuSparseMatrixPtr matB = std::dynamic_pointer_cast( + matA->subMatrix(startRow, rowNum)); + + int start = matA->getRows()[startRow]; + int end = matA->getRows()[startRow + rowNum]; + + /*compare two matrix*/ + ASSERT_EQ(matB->getElementCnt(), size_t(end - start)); + if (valueType == FLOAT_VALUE) { + for (size_t i = 0; i < matB->getElementCnt(); i++) { + ASSERT_FLOAT_EQ(matB->getValue()[start + i], + matA->getValue()[start + i]); + } + } + + for (size_t i = 0; i < matB->getElementCnt(); i++) { + ASSERT_EQ(matB->getCols()[start + i], matA->getCols()[start + i]); + } + for (size_t i = 0; i < rowNum; i++) { + ASSERT_EQ(matB->getRows()[i], matA->getRows()[startRow + i]); + } + } +} + +void sparseValid( + int* major, int* minor, size_t nnz, size_t majorLen, size_t minorLen) { + CHECK_EQ(nnz, size_t(major[majorLen - 1])); + CHECK_EQ(nnz, minorLen); + for (size_t i = 0; i < majorLen - 1; i++) { + EXPECT_LE(major[i], major[i + 1]); + for (int j = major[i]; j < major[i + 1] - 1; j++) { + EXPECT_LE(minor[j], minor[j + 1]); + } + } +} + +TEST(Matrix, CpuSparseMatrixRandUniform) { + const size_t HEIGHT = 5; + const size_t WIDTH = 10; + const size_t NNZ = HEIGHT * WIDTH; + int* major = nullptr; + int* minor = nullptr; + size_t majorLen = 0; + size_t minorLen = 0; + size_t nnz = 0; + for (auto valueType : {NO_VALUE, FLOAT_VALUE}) { + for (auto format : {SPARSE_CSR, SPARSE_CSC}) { + CpuSparseMatrixPtr matA = std::make_shared( + HEIGHT, WIDTH, size_t(NNZ * 0.1), valueType, format); + matA->randomizeUniform(); + nnz = matA->getElementCnt(); + if (format == SPARSE_CSR) { + majorLen = matA->getHeight() + 1; + minorLen = matA->getElementCnt(); + major = matA->getRows(); + minor = matA->getCols(); + } else { + majorLen = matA->getWidth() + 1; + minorLen = matA->getElementCnt(); + major = matA->getCols(); + minor = matA->getRows(); + } + sparseValid(major, minor, nnz, majorLen, minorLen); + } + } +} + +TEST(Matrix, CpuSparseMatrixCopyFrom) { + size_t height = 10; + size_t width = 8; + int64_t indices[11] = {0, 1, 5, 5, 9, 13, 15, 17, 19, 30, 32}; + sparse_non_value_t data[32]; + for (size_t i = 0; i < 32; i++) { + data[i].col = ::rand() % width; + } + CpuSparseMatrixPtr mat = std::make_shared( + height, width, 32, NO_VALUE, SPARSE_CSR, false); + mat->copyFrom(indices, data); + + /*compare indices*/ + size_t sum = 0; + CHECK_EQ(sum, size_t(mat->getRows()[0])); + for (size_t i = 1; i < height + 1; i++) { + sum += indices[i] - indices[i - 1]; + CHECK_EQ(sum, size_t(mat->getRows()[i])); + } + CHECK_EQ(mat->getElementCnt(), size_t(indices[height] - indices[0])); + for (size_t i = 0; i < mat->getElementCnt(); i++) { + CHECK_EQ(size_t(mat->getCols()[i]), size_t(data[i].col)); + } +} + +TEST(Matrix, SparseMatrixCSRFormatTrimFrom) { + size_t height = 10; + size_t width = 8; + int64_t indices[11] = {0, 1, 5, 5, 9, 13, 15, 17, 19, 27, 32}; + sparse_float_value_t data[32]; + int value[32] = { + 1, // row_0 : 1 + 5, 3, 1, 6, // row_1 : 4 + 0, 1, 2, 3, // row_3 : 4 + 4, 5, 6, 7, // row_4 : 4 + 2, 3, // row_5 : 2 + 3, 5, // row_6 : 2 + 0, 1, // row_7 : 2 + 0, 1, 2, 3, 4, 5, 6, 7, // row_8 : 8 + 2, 4, 7, 3, 1 // row_9 : 5 + }; + for (size_t i = 0; i < 32; i++) { + data[i].col = value[i]; + data[i].value = float(value[i]); + } + CpuSparseMatrixPtr mat = std::make_shared( + height, width, 32, FLOAT_VALUE, SPARSE_CSR, false); + mat->copyFrom(indices, data); + + /*compare indices*/ + size_t sum = 0; + CHECK_EQ(sum, size_t(mat->getRows()[0])); + for (size_t i = 1; i < height + 1; i++) { + sum += indices[i] - indices[i - 1]; + CHECK_EQ(sum, size_t(mat->getRows()[i])); + } + CHECK_EQ(mat->getElementCnt(), size_t(indices[height] - indices[0])); + for (size_t i = 0; i < mat->getElementCnt(); i++) { + CHECK_EQ(size_t(mat->getCols()[i]), size_t(data[i].col)); + } + + size_t trimedWidth = 4; + int64_t trimedIndices[11] = {0, 1, 3, 3, 7, 7, 9, 10, 12, 16, 19}; + sparse_float_value_t trimedData[19]; + int trimedValue[19] = { + 1, // row_0 : 1 + 3, + 1, // row_1 : 2 + 0, + 1, + 2, + 3, // row_3 : 4 + 2, + 3, // row_5 : 2 + 3, // row_6 : 1 + 0, + 1, // row_7 : 2 + 0, + 1, + 2, + 3, // row_8 : 4 + 2, + 3, + 1 // row_9 : 3 + }; + for (size_t i = 0; i < 19; i++) { + trimedData[i].col = trimedValue[i]; + trimedData[i].value = float(trimedValue[i]); + } + CpuSparseMatrixPtr matA = std::make_shared( + height, trimedWidth, 19, FLOAT_VALUE, SPARSE_CSR, false); + matA->copyFrom(trimedIndices, trimedData); + + /*compare indices*/ + sum = 0; + CHECK_EQ(sum, size_t(matA->getRows()[0])); + for (size_t i = 1; i < height + 1; i++) { + sum += trimedIndices[i] - trimedIndices[i - 1]; + CHECK_EQ(sum, size_t(matA->getRows()[i])); + } + CHECK_EQ(matA->getElementCnt(), + size_t(trimedIndices[height] - trimedIndices[0])); + for (size_t i = 0; i < matA->getElementCnt(); i++) { + CHECK_EQ(size_t(matA->getCols()[i]), size_t(trimedData[i].col)); + } + + CpuSparseMatrixPtr matB = std::make_shared( + height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, false); + matB->trimFrom(*mat); + checkSMatrixEqual2(matA, matB); + +#ifdef PADDLE_WITH_CUDA + GpuSparseMatrixPtr matC = std::make_shared( + height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, true); + matC->trimFrom(*mat); + + CpuSparseMatrixPtr matD = + std::make_shared(height, + trimedWidth, + matC->getElementCnt(), + FLOAT_VALUE, + SPARSE_CSR, + false); + matD->copyFrom(*matC, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + checkSMatrixEqual2(matA, matD); +#endif +} + +TEST(Matrix, SparseMatrixCSCFormatTrimFrom) { + size_t height = 8; + size_t width = 10; + int indices[11] = {0, 1, 5, 5, 9, 13, 15, 17, 19, 27, 32}; + int value[32] = { + 1, // col_0 : 1 + 5, 3, 1, 6, // col_1 : 4 + 0, 1, 2, 3, // col_3 : 4 + 4, 5, 6, 7, // col_4 : 4 + 2, 3, // col_5 : 2 + 3, 5, // col_6 : 2 + 0, 1, // col_7 : 2 + 0, 1, 2, 3, 4, 5, 6, 7, // col_8 : 8 + 2, 4, 7, 3, 1 // col_9 : 5 + }; + std::vector rows(value, value + 32); + std::vector cols(indices, indices + 11); + std::vector values(value, value + 32); + CpuSparseMatrixPtr mat = std::make_shared( + height, width, 32, FLOAT_VALUE, SPARSE_CSC, false); + mat->copyFrom(rows, cols, values); + + /*compare indices*/ + size_t sum = 0; + CHECK_EQ(sum, size_t(mat->getCols()[0])); + for (size_t i = 1; i < width + 1; i++) { + sum += indices[i] - indices[i - 1]; + CHECK_EQ(sum, size_t(mat->getCols()[i])); + } + CHECK_EQ(mat->getElementCnt(), size_t(indices[width] - indices[0])); + for (size_t i = 0; i < mat->getElementCnt(); i++) { + CHECK_EQ(size_t(mat->getRows()[i]), size_t(value[i])); + } + + size_t trimedWidth = 5; + int trimedIndices[6] = {0, 1, 5, 5, 9, 13}; + int trimedValue[13] = { + 1, // col_0 : 1 + 5, + 3, + 1, + 6, // col_1 : 4 + 0, + 1, + 2, + 3, // col_3 : 4 + 4, + 5, + 6, + 7 // col_4 : 4 + }; + std::vector rowsA(trimedValue, trimedValue + 13); + std::vector colsA(trimedIndices, trimedIndices + 6); + std::vector valuesA(trimedValue, trimedValue + 13); + CpuSparseMatrixPtr matA = std::make_shared( + height, trimedWidth, 13, FLOAT_VALUE, SPARSE_CSC, false); + matA->copyFrom(rowsA, colsA, valuesA); + + /*compare indices*/ + sum = 0; + CHECK_EQ(sum, size_t(matA->getCols()[0])); + for (size_t i = 1; i < trimedWidth + 1; i++) { + sum += trimedIndices[i] - trimedIndices[i - 1]; + CHECK_EQ(sum, size_t(matA->getCols()[i])); + } + CHECK_EQ(matA->getElementCnt(), + size_t(trimedIndices[trimedWidth] - trimedIndices[0])); + for (size_t i = 0; i < matA->getElementCnt(); i++) { + CHECK_EQ(size_t(matA->getRows()[i]), size_t(rowsA[i])); + } + + CpuSparseMatrixPtr matB = std::make_shared( + height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, false); + matB->trimFrom(*mat); + checkSMatrixEqual2(matA, matB); + +#ifdef PADDLE_WITH_CUDA + GpuSparseMatrixPtr matC = std::make_shared( + height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, true); + matC->trimFrom(*mat); + + CpuSparseMatrixPtr matD = + std::make_shared(height, + trimedWidth, + matC->getElementCnt(), + FLOAT_VALUE, + SPARSE_CSC, + false); + matD->copyFrom(*matC, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + checkSMatrixEqual2(matA, matD); +#endif +} diff --git a/paddle/legacy/math/tests/test_Tensor.cu b/paddle/legacy/math/tests/test_Tensor.cu new file mode 100644 index 0000000000000000000000000000000000000000..3ce056d66140059be8145f7f49bb80cbff4686eb --- /dev/null +++ b/paddle/legacy/math/tests/test_Tensor.cu @@ -0,0 +1,1162 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "TensorCheck.h" +#include "paddle/legacy/math/Matrix.h" + +using paddle::Matrix; +using paddle::CpuMatrix; +using paddle::GpuMatrix; +using paddle::CpuVector; +using paddle::GpuVector; +using paddle::CpuIVector; +using paddle::GpuIVector; +using autotest::TensorCheckEqual; +using autotest::TensorCheckErr; + +#define INIT_UNARY(A1, A2) \ + Tensor A1(height, width); \ + Tensor A2(height, width); \ + A1.randomizeUniform(); \ + A2.copyFrom(A1) +#define INIT_BINARY(A1, A2, B) \ + INIT_UNARY(A1, A2); \ + Tensor B(height, width); \ + B.randomizeUniform() +#define INIT_TERNARY(A1, A2, B, C) \ + INIT_BINARY(A1, A2, B); \ + Tensor C(height, width); \ + C.randomizeUniform() +#define INIT_QUATERNARY(A1, A2, B, C, D) \ + INIT_TERNARY(A1, A2, B, C); \ + Tensor D(height, width); \ + D.randomizeUniform() + +template +struct TestUnaryMatrix { + typedef std::function UnaryFunc; + + explicit TestUnaryMatrix(UnaryFunc testUnaryFunc) { + for (auto height : {1, 11, 73, 128, 200, 330}) { + for (auto width : {1, 32, 100, 512, 1000, 3210}) { + LOG(INFO) << " height=" << height << " width=" << width; + INIT_UNARY(A1, A2); + testUnaryFunc(A1, A2); + } + } + } +}; + +template +struct TestBinaryMatrix { + typedef std::function BinaryFunc; + + explicit TestBinaryMatrix(BinaryFunc testBinaryFunc) { + for (auto height : {1, 11, 73, 128, 200, 330}) { + for (auto width : {1, 32, 100, 512, 1000, 3210}) { + LOG(INFO) << " height=" << height << " width=" << width; + INIT_BINARY(A1, A2, B); + testBinaryFunc(A1, A2, B); + } + } + } +}; + +template +struct TestTernaryMatrix { + typedef std::function + TernaryFunc; + + explicit TestTernaryMatrix(TernaryFunc testTernaryFunc) { + for (auto height : {1, 11, 73, 128, 200, 330}) { + for (auto width : {1, 32, 100, 512, 1000, 3210}) { + LOG(INFO) << " height=" << height << " width=" << width; + INIT_TERNARY(A1, A2, B, C); + testTernaryFunc(A1, A2, B, C); + } + } + } +}; + +template +struct TestQuaternaryMatrix { + typedef std::function + QuaternaryFunc; + + explicit TestQuaternaryMatrix(QuaternaryFunc testQuaternaryFunc) { + for (auto height : {1, 11, 73, 128, 200, 330}) { + for (auto width : {1, 32, 100, 512, 1000, 3210}) { + LOG(INFO) << " height=" << height << " width=" << width; + INIT_QUATERNARY(A1, A2, B, C, D); + testQuaternaryFunc(A1, A2, B, C, D); + } + } + } +}; + +template +struct TestUnaryVectorT { + typedef std::function UnaryFunc; + + explicit TestUnaryVectorT(UnaryFunc testUnaryFunc) { + for (auto size : {1, 11, 73, 128, 200, 330, 512, 1000, 4210}) { + LOG(INFO) << " size=" << size; + Tensor A1(size); + Tensor A2(size); + if (typeid(T) == typeid(real)) { + A1.rand(); + } else { + A1.rand(1000); + } + A2.copyFrom(A1); + testUnaryFunc(A1, A2); + } + } +}; + +void SetTensorValue(Matrix& matrix, real value) { + int height = matrix.getHeight(); + int width = matrix.getWidth(); + int stride = matrix.getStride(); + real* data = matrix.getData(); + for (int i = 0; i < height; i++) { + int j = rand() % width; // NOLINT + if (typeid(matrix) == typeid(CpuMatrix)) { + data[i * stride + j] = value; + } else if (typeid(matrix) == typeid(GpuMatrix)) { + hl_memcpy(&data[i * stride + j], &value, sizeof(real)); + } else { + } + } +} + +template +void testTensorAddScalar(Tensor& A1, Tensor& A2) { + real p1 = 2.5; + real p2 = 3.0; + A1.add(p1); // a += p + A2 += p1; + TensorCheckEqual(A1, A2); + + A1.add(p1, p2); // a = a * p1 + p2 + A2 = A2 * p1 + p2; + TensorCheckEqual(A1, A2); +} + +template +void testTensorSubScalar(Tensor& A1, Tensor& A2) { + real p = 2.5; + A1.subScalar(p); // a -= p + A2 -= p; + TensorCheckEqual(A1, A2); +} + +template +void testTensorMulScalar(Tensor& A1, Tensor& A2) { + real p = 2.5; + A1.mulScalar(p); // a *= p + A2 *= p; + TensorCheckEqual(A1, A2); + + real learningRate = 0.7f; + real decayRate = 1.2f; + A1.applyL2(learningRate, decayRate); + A2 = A2 * (1.0f / (1.0f + learningRate * decayRate)); + TensorCheckEqual(A1, A2); +} + +template +void testTensorDivScalar(Tensor& A1, Tensor& A2) { + real p = 2.5; + A1.divScalar(p); // a /= p + A2 /= p; + TensorCheckEqual(A1, A2); +} + +template +void testTensorNeg(Tensor& A1, Tensor& A2) { + A1.neg(); // a = -a + A2 = -A2; + TensorCheckEqual(A1, A2); +} + +template +void testTensorAbs(Tensor& A1, Tensor& A2) { + A1.abs2(); // a = a > 0 ? a : -a + A2 = A2.abs(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorSquare(Tensor& A1, Tensor& A2) { + A1.square2(); // a = a * a + A2 = A2.square(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorReciprocal(Tensor& A1, Tensor& A2) { + A1.reciprocal2(); // a = 1.0f / a + A2 = A2.reciprocal(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorSign(Tensor& A1, Tensor& A2) { + A1.sign2(); // a = (a > 0) - (a < 0) + A2 = A2.sign(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorAssign(Tensor& A1, Tensor& A2) { + A1.assign(1.5); // a = p + A2 = A2.constant(1.5); + TensorCheckEqual(A1, A2); + + A1.one(); // a = 1 + A2 = A2.constant(1.0); + TensorCheckEqual(A1, A2); + + A1.zero(); // a = 0 + A2 = A2.constant(0.0); + TensorCheckEqual(A1, A2); +} + +template +void testUnaryBaseOp(Tensor& A1, Tensor& A2) { + testTensorAddScalar(A1, A2); + testTensorSubScalar(A1, A2); + testTensorMulScalar(A1, A2); + testTensorDivScalar(A1, A2); + testTensorNeg(A1, A2); + testTensorAbs(A1, A2); + testTensorSquare(A1, A2); + testTensorReciprocal(A1, A2); + testTensorSign(A1, A2); + testTensorAssign(A1, A2); +} + +template +void testUnaryBaseOpInt(Tensor& A1, Tensor& A2) { + A1.add(2); // a += p + A2 += 2; + TensorCheckEqual(A1, A2); + + A1.add(3, 2); // a = a * p1 + p2 + A2 = A2 * 3 + 2; + TensorCheckEqual(A1, A2); + + testTensorNeg(A1, A2); + testTensorAbs(A1, A2); +} + +TEST(Unary, BaseOp) { + TestUnaryMatrix testCpuMatrix(testUnaryBaseOp); + TestUnaryVectorT testCpuVector(testUnaryBaseOp); + TestUnaryVectorT testCpuIVector( + testUnaryBaseOpInt); + +#ifdef PADDLE_WITH_GPU + TestUnaryMatrix testGpuMatrix(testUnaryBaseOp); + TestUnaryVectorT testGpuVector(testUnaryBaseOp); + TestUnaryVectorT testGpuIVector( + testUnaryBaseOpInt); +#endif +} + +template +void testTensorExp(Tensor& A1, Tensor& A2) { + A1.exp2(); // a = exp(a) + A2 = A2.exp(); + TensorCheckErr(A1, A2); +} + +template +void testTensorLog(Tensor& A1, Tensor& A2) { + A1.log2(); // a = log(a) + A2 = A2.log(); + TensorCheckErr(A1, A2); +} + +template +void testTensorSqrt(Tensor& A1, Tensor& A2) { + A1.sqrt2(); // a = sqrt(a) + A2 = A2.sqrt(); + TensorCheckErr(A1, A2); +} + +template +void testTensorPow(Tensor& A1, Tensor& A2) { + A1.pow2(3.2); // a = pow(a, p) + A2 = A2.pow(3.2); + TensorCheckErr(A1, A2); +} + +template +void testUnayrMathOp(Tensor& A1, Tensor& A2) { + testTensorExp(A1, A2); + testTensorLog(A1, A2); + testTensorSqrt(A1, A2); + testTensorPow(A1, A2); +} + +TEST(Unary, MathOp) { + TestUnaryMatrix testCpu(testUnayrMathOp); + +#ifdef PADDLE_WITH_GPU + TestUnaryMatrix testGpu(testUnayrMathOp); +#endif +} + +template +void testTensorClip(Tensor& A1, Tensor& A2) { + real p1 = 0.003f; + real p2 = 0.877f; + A1.clip(p1, p2); // a = a < p1 ? p1 : (a > p2 ? p2 : a) + // A2 = A2.min(0.877f).max(0.003f); + A2 = (A2 < p1).condition(p1, (A2 > p2).condition(p2, A2)); + TensorCheckEqual(A1, A2); +} + +template +void testTensorBiggerThanScalar(Tensor& A1, Tensor& A2) { + real p = 0.5f; + A1.biggerThanScalar(p); // a = a > p ? 1.0f : 0.0f + A2 = (A2 > p).condition((real)1.0, (real)0.0); + TensorCheckEqual(A1, A2); +} + +template +void testTensorapplyL1(Tensor& A1, Tensor& A2) { + /** + * T lambda = p; + * a = (a > lambda) ? (a - lambda) + * : (a < -lambda) ? (a + lambda) : 0 + * + * p = learningRate * decayRate; + */ + real learningRate = 0.7f; + real decayRate = 0.6f; + A1.applyL1(learningRate, decayRate); + A2 = (A2 > (learningRate * decayRate)) + .condition( + (A2 - (learningRate * decayRate)), + (A2 < -(learningRate * decayRate)) + .condition((A2 + (learningRate * decayRate)), (real)0.0)); + TensorCheckEqual(A1, A2); +} + +template +void testUnayrCompareOp(Tensor& A1, Tensor& A2) { + testTensorClip(A1, A2); + testTensorBiggerThanScalar(A1, A2); + + A1.randomizeUniform(); + A1.subScalar(0.5f); + A2.copyFrom(A1); + testTensorapplyL1(A1, A2); +} + +TEST(Unary, CompareOp) { + TestUnaryMatrix testCpu(testUnayrCompareOp); + +#ifdef PADDLE_WITH_GPU + TestUnaryMatrix testGpu(testUnayrCompareOp); +#endif +} + +template +void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B) { + real p1 = 2.5; + real p2 = 3.2; + A1.add(B); // a += b + A2 += B; + TensorCheckEqual(A1, A2); + + A1.add(B, p1); // a += b * p + A2 += B * p1; + TensorCheckEqual(A1, A2); + + A1.add(B, p1, p2); // a = p1 * a + p2 * b + A2 = A2 * p1 + B * p2; + TensorCheckEqual(A1, A2); + + A1.addScalar(B, p1); // a = b + p + A2 = B + p1; + TensorCheckEqual(A1, A2); + + A1.addSquare(B, p1); // a += p * b * b + A2 += B.constant(p1) * B * B; + TensorCheckEqual(A1, A2); + + A1.decayAddSquare(B, p1, p2); // a = p1 * a + p2 * b * b + A2 = A2 * p1 + B.constant(p2) * B * B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B) { + real p = 2.5; + A1.sub(B); // a -= b + A2 -= B; + TensorCheckEqual(A1, A2); + + A1.sub(B, p); // a -= b * p + A2 -= B * p; + TensorCheckEqual(A1, A2); + + A1.subScalar(B, p); // a = b - p + A2 = B - p; + TensorCheckEqual(A1, A2); +} + +template +void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B) { + real p = 2.5; + A1.mulScalar(B, p); // a = b * p + A2 = B * p; + TensorCheckEqual(A1, A2); + + A1.dotMulSquare(B); // a *= b * b + A2 *= B * B; + TensorCheckEqual(A1, A2); + + A1.dotSquareMul(B); // a = a * a * b + A2 = A2 * A2 * B; + TensorCheckEqual(A1, A2); + + A1.dotMul(B); // a *= b + A2 *= B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B) { + real p = 2.5; + A1.divScalar(B, p); // a = b / p + A2 = B / p; + TensorCheckEqual(A1, A2); + + A1.scalarDiv(B, p); // a = p / b + A2 = B.constant(p) / B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorAssign(Tensor& A1, Tensor& A2, Tensor& B) { + A1.assign(B); // a = b + A2 = B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorSquare(Tensor& A1, Tensor& A2, Tensor& B) { + B.square2(A1); // b = a * a + A2 = B.square(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorSquareDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.squareDerivative(B); // a *= 2.0 * b + A2 = A2 * (real)2.0 * B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B) { + B.reciprocal2(A1); // b = 1.0f / a + A2 = B.reciprocal(); + TensorCheckEqual(A1, A2); + + real p1 = 0.58; + real p2 = 0.32; + A1.reciprocal2(B, p1, p2); // a = 1 / (p1 * b + p2) + A2 = (B * p1 + p2).reciprocal(); + TensorCheckEqual(A1, A2); + + real learningRate = 0.7f; + real decayRate = 1.2f; + A1.applyL2(B, learningRate, decayRate); // a *= (1.0f / (1.0f + p * b)) + A2 *= (B.constant(1.0f) + B.constant(learningRate * decayRate) * B) + .reciprocal(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorReciprocalDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.reciprocalDerivative(B); // a *= -b * b + A2 *= (-B) * B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorSign(Tensor& A1, Tensor& A2, Tensor& B) { + B.sign2(A1); // b = a > 0.0f ? 1.0f : -1.0f + A2 = B.sign(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorAbs(Tensor& A1, Tensor& A2, Tensor& B) { + B.abs2(A1); // b = a > 0.0f ? a : -a + A2 = B.abs(); + TensorCheckEqual(A1, A2); +} + +template +void testBinaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B) { + testTensorAdd(A1, A2, B); + testTensorSub(A1, A2, B); + testTensorMul(A1, A2, B); + testTensorDiv(A1, A2, B); + testTensorSquare(A1, A2, B); + testTensorSquareDerivative(A1, A2, B); + testTensorReciprocal(A1, A2, B); + testTensorReciprocalDerivative(A1, A2, B); + testTensorAbs(A1, A2, B); + testTensorSign(A1, A2, B); + testTensorAssign(A1, A2, B); +} + +TEST(Binary, BaseOp) { + TestBinaryMatrix testCpu(testBinaryBaseOp); + +#ifdef PADDLE_WITH_GPU + TestBinaryMatrix testGpu(testBinaryBaseOp); +#endif +} + +template +void testTensorExp(Tensor& A1, Tensor& A2, Tensor& B) { + // a = exp(b) + A1.exp2(B); + A2 = B.exp(); + TensorCheckErr(A1, A2); +} + +template +void testTensorExpDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.expDerivative(B); // a *= b + A2 *= B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorLog(Tensor& A1, Tensor& A2, Tensor& B) { + // a = log(b) + A1.log2(B); + A2 = B.log(); + TensorCheckErr(A1, A2); +} + +template +void testTensorSqrt(Tensor& A1, Tensor& A2, Tensor& B) { + // a = sqrt(b) + A1.sqrt2(B); + A2 = B.sqrt(); + TensorCheckErr(A1, A2); +} + +template +void testTensorInvSqrt(Tensor& A1, Tensor& A2, Tensor& B) { + // a = 1.0f / sqrt(b) + A1.invSqrt(B); + A2 = B.sqrt().reciprocal(); + TensorCheckErr(A1, A2); +} + +template +void testTensorPow(Tensor& A1, Tensor& A2, Tensor& B) { + A1.pow2(B, 2.5f); // a = pow(b, p) + A2 = B.pow(2.5f); + TensorCheckErr(A1, A2); +} + +template +void testTensorSoftrelu(Tensor& A1, Tensor& A2, Tensor& B) { + /* + * const T THRESHOLD = 40.0; + * b = log(1.0 + + * exp((a > THRESHOLD) ? THRESHOLD + * : ((a < -THRESHOLD) ? (-THRESHOLD) : a))) + */ + B.softrelu(A1); + + real THRESHOLD = 40.0; + A2 = (B.constant(1.0f) + + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)) + .exp()) + .log(); + TensorCheckErr(A1, A2); +} + +template +void testTensorSoftreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + /* + * const T THRESHOLD = 40.0; + * a *= (1.0 - exp(-1.0 * ((b > THRESHOLD) + * ? THRESHOLD + * : ((b < -THRESHOLD) ? (-THRESHOLD) : b))))); + */ + A1.softreluDerivative(B); + real THRESHOLD = 40.0; + A2 = A2 * + (B.constant(1.0f) - + (B.constant(-1.0f) * + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B))) + .exp()); + TensorCheckErr(A1, A2); +} + +template +void testTensorSigmoid(Tensor& A1, Tensor& A2, Tensor& B) { + /* + const T THRESHOLD_MIN = -40.0; + const T THRESHOLD_MAX = 13.0; + T tmp = (a < THRESHOLD_MIN) ? THRESHOLD_MIN + : ((a > THRESHOLD_MAX) ? THRESHOLD_MAX : a); + b = 1.0f / (1.0f + exp(-tmp))) + */ + B.sigmoid(A1); + + const real THRESHOLD_MIN = -40.0; + const real THRESHOLD_MAX = 13.0; + auto tmp = (B < THRESHOLD_MIN) + .condition(THRESHOLD_MIN, + (B > THRESHOLD_MAX).condition(THRESHOLD_MAX, B)); + A2 = (B.constant(1.0f) + (-tmp).exp()).reciprocal(); + TensorCheckErr(A1, A2); +} + +template +void testTensorSigmoidDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.sigmoidDerivative(B); // a *= b * (1 - b) + A2 *= B * (B.constant(1.0f) - B); + TensorCheckEqual(A1, A2); +} + +template +void testTensorTanh(Tensor& A1, Tensor& A2, Tensor& B) { + B.tanh(A1); // b = 2.0 / (1.0 + exp(-2 * a)) - 1.0 + A2 = B.constant(2.0f) / ((B * ((real)-2.0f)).exp() + (real)1.0f) - (real)1.0f; + TensorCheckErr(A1, A2); +} + +template +void testTensorTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.tanhDerivative(B); // a *= 1 - b * b + A2 *= B.constant(1.0f) - B * B; + TensorCheckEqual(A1, A2); +} + +template +void testTensorScaledTanh(Tensor& A1, Tensor& A2, Tensor& B) { + real p1 = 2.5; + real p2 = 3.1; + // b = p1 * (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0) + B.scaledTanh(A1, p1, p2); + A2 = B.constant(p1) * + (B.constant(2.0f) / ((B.constant(-2.0f) * p2 * B).exp() + (real)1.0) - + (real)1.0); + TensorCheckErr(A1, A2); +} + +template +void testTensorScaledTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + real p1 = 2.5; + real p2 = 3.1; + // a *= (p2 / p1) * (p1 * p1 - b * b)); + A1.scaledTanhDerivative(B, p1, p2); + A2 = A2 * (B.constant(p2 / p1) * (B.constant(p1 * p1) - B * B)); + TensorCheckEqual(A1, A2); +} + +template +void testBinaryMathOp(Tensor& A1, Tensor& A2, Tensor& B) { + testTensorTanhDerivative(A1, A2, B); + testTensorScaledTanhDerivative(A1, A2, B); + testTensorSigmoidDerivative(A1, A2, B); + testTensorExpDerivative(A1, A2, B); + testTensorScaledTanh(A1, A2, B); + testTensorTanh(A1, A2, B); + testTensorExp(A1, A2, B); + testTensorLog(A1, A2, B); + testTensorSqrt(A1, A2, B); + testTensorInvSqrt(A1, A2, B); + testTensorPow(A1, A2, B); + + testTensorSoftrelu(A1, A2, B); + testTensorSoftreluDerivative(A1, A2, B); + testTensorSigmoid(A1, A2, B); +} + +TEST(Binary, MathOp) { + TestBinaryMatrix testCpu(testBinaryMathOp); + +#ifdef PADDLE_WITH_GPU + TestBinaryMatrix testGpu(testBinaryMathOp); +#endif +} + +template +void testTensorRelu(Tensor& A1, Tensor& A2, Tensor& B) { + B.relu(A1); // b = a > 0.0f ? a : 0.0f + A2 = (B > (real)0.0f).condition(B, (real)0.0f); + TensorCheckEqual(A1, A2); +} + +template +void testTensorReluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.reluDerivative(B); // a *= (b > 0.0f ? 1.0f : 0.0f) + A2 *= (B > (real)0.0).condition((real)1.0, (real)0.0); + TensorCheckEqual(A1, A2); +} + +template +void testTensorBrelu(Tensor& A1, Tensor& A2, Tensor& B) { + /* + * b = a > p1 ? a : p1 + * b = b < p2 ? b : p2 + * int p1 = 0, p2 = 24; + */ + SetTensorValue(B, 32.0f); + B.brelu(A1); + auto tmp = (B > (real)0.0f).condition(B, (real)0.0f); + A2 = (tmp < (real)24.0f).condition(tmp, (real)24.0f); + TensorCheckEqual(A1, A2); +} + +template +void testTensorBreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + SetTensorValue(B, 32.0f); + /* + * a *= (b > p1 && b < p2) ? 1.0 : 0.0 + * int p1 = 0, p2 = 24; + */ + A1.breluDerivative(B); + A2 *= (B > (real)0.0f && B < (real)24.0f).condition((real)1.0f, (real)0.0f); + TensorCheckEqual(A1, A2); +} + +template +void testTensorAbsDerivative(Tensor& A1, Tensor& A2, Tensor& B) { + A1.absDerivative(B); // a = (b > 0) ? a : (b < 0) ? -a : 0 + A2 = (B > (real)0.0f) + .condition(A2, (B < (real)0.0f).condition(-A2, (real)0.0f)); + TensorCheckEqual(A1, A2); +} + +template +void testTensorIsEqualTo(Tensor& A1, Tensor& A2, Tensor& B) { + real p = 0.613; + SetTensorValue(B, p); + A1.isEqualTo(B, p); // a = (b == p) + A2 = (B == p); + TensorCheckEqual(A1, A2); +} + +template +void testTensorapplyL1(Tensor& A1, Tensor& A2, Tensor& B) { + /** + * T lambda = p * b; + * a = (a > lambda) ? (a - lambda) + * : (a < -lambda) ? (a + lambda) : 0 + * + * p = learningRate * decayRate; + */ + real learningRate = 0.7f; + real decayRate = 0.6f; + A1.applyL1(B, learningRate, decayRate); + auto lambda = B.constant(learningRate * decayRate) * B; + A2 = (A2 > lambda) + .condition((A2 - lambda), + (A2 < -lambda).condition((A2 + lambda), (real)0.0f)); + TensorCheckEqual(A1, A2); +} + +template +void testBinaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B) { + B.subScalar(0.5f); + SetTensorValue(B, 0.0f); + testTensorReluDerivative(A1, A2, B); + + A1.randomizeUniform(); + A2.copyFrom(A1); + testTensorBreluDerivative(A1, A2, B); + + testTensorAbsDerivative(A1, A2, B); + testTensorRelu(A1, A2, B); + testTensorBrelu(A1, A2, B); + testTensorIsEqualTo(A1, A2, B); +} + +TEST(Binary, CompareOp) { + TestBinaryMatrix testCpu(testBinaryCompareOp); + +#ifdef PADDLE_WITH_GPU + TestBinaryMatrix testGpu(testBinaryCompareOp); +#endif +} + +template +void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.add(B, C); // a = b + c + A2 = B + C; + TensorCheckEqual(A1, A2); + + real p1 = 1.5; + real p2 = 2.5; + real p3 = 3.8; + A1.add(B, p1, C, p2); // a = p1 * b + p2 * c + A2 = B * p1 + C * p2; + TensorCheckEqual(A1, A2); + + A1.add2(B, C); // a = a + b + c + A2 = A2 + B + C; + TensorCheckEqual(A1, A2); + + A1.add2(B, C, p1, p2, p3); // a = p1 * a + p2 * b + p3 * c + A2 = A2 * p1 + B * p2 + C * p3; + TensorCheckEqual(A1, A2); + + A1.decayAddSquareMul(B, C, p1, p2); // a = p1 * a + p2 * b * b * c * c + A2 = A2 * p1 + B.constant(p2) * B * B * C * C; + TensorCheckEqual(A1, A2); +} + +template +void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.sub(B, C); // a = b - c + A2 = B - C; + TensorCheckEqual(A1, A2); + + real p1 = 1.5; + real p2 = 2.5; + A1.sub(B, p1, C, p2); // a = p1 * b - p2 * c + A2 = B * p1 - C * p2; + TensorCheckEqual(A1, A2); +} + +template +void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.dotMul(B, C); // a = b * c + A2 = B * C; + TensorCheckEqual(A1, A2); + + A1.dotMulSquare(B, C); // a = b * c * c + A2 = B * C * C; + TensorCheckEqual(A1, A2); + + A1.dotSquareSquare(B, C); // a = b * b * c * c + A2 = B * B * C * C; + TensorCheckEqual(A1, A2); + + real p1 = 1.5; + real p2 = 2.5; + + /* + * T tmp = p1 * b + p2 * c; + * a *= tmp * tmp + */ + A1.dotMulSquareSum(B, C, p1, p2); + auto tmp = B * p1 + C * p2; + A2 *= tmp * tmp; + TensorCheckEqual(A1, A2); + + /* + * T tmp = p1 * b + p2 * c; + * a = tmp * tmp + */ + A1.dotSquareSum(B, C, p1, p2); + auto tmp2 = B * p1 + C * p2; + A2 = tmp2 * tmp2; + TensorCheckEqual(A1, A2); + + // a *= p1 * b + p2 * c + A1.dotMulSum(B, C, p1, p2); + A2 *= B * p1 + C * p2; + TensorCheckEqual(A1, A2); + + // a = p1 * a + p2 * b * c + A1.addDotMul(B, C, p1, p2); + A2 = A2 * p1 + B.constant(p2) * B * C; + TensorCheckEqual(A1, A2); +} + +template +void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.dotDiv(B, C); // a = (b == 0.0) ? 0.0 : b / c + A2 = (B == (real)0.0).condition((real)0.0, B / C); + TensorCheckEqual(A1, A2); + + real p1 = 1.5; + real p2 = 2.5; + A1.dotDiv(B, C, p1, p2); // a = (b + p1) / (c + p2) + A2 = (B + p1) / (C + p2); + TensorCheckEqual(A1, A2); +} + +template +void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + real p1 = 1.5; + real p2 = 2.5; + real p3 = 3.5; + A1.reciprocalSum(B, C, p1, p2, p3); // a = 1 / (p1 * b + p2 * c + p3) + A2 = (B * p1 + C * p2 + p3).reciprocal(); + TensorCheckEqual(A1, A2); +} + +template +void testTensorSoftCrossEntropy(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.softCrossEntropy(B, C); // a = -c * log(b) - (1 - c) * log(1 - b) + A2 = -C * B.log() - (C.constant(1.0f) - C) * (B.constant(1.0f) - B).log(); + TensorCheckErr(A1, A2); +} + +template +void testTensorSoftCrossEntropyBp(Tensor& A1, + Tensor& A2, + Tensor& B, + Tensor& C) { + A1.softCrossEntropyBp(B, C); // a += (b - c) / (b * (1 - b)) + A2 += (B - C) / (B * (B.constant(1.0f) - B)); + TensorCheckEqual(A1, A2); +} + +template +void testTernaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + testTensorAdd(A1, A2, B, C); + testTensorSub(A1, A2, B, C); + testTensorMul(A1, A2, B, C); + testTensorDiv(A1, A2, B, C); + testTensorReciprocal(A1, A2, B, C); + testTensorSoftCrossEntropyBp(A1, A2, B, C); + + testTensorSoftCrossEntropy(A1, A2, B, C); +} + +TEST(Ternary, BaseOp) { + TestTernaryMatrix testCpu(testTernaryBaseOp); + +#ifdef PADDLE_WITH_GPU + TestTernaryMatrix testGpu(testTernaryBaseOp); +#endif +} + +template +void testTensorBinaryLabelCrossEntropy(Tensor& A1, + Tensor& A2, + Tensor& B, + Tensor& C) { + A1.binaryLabelCrossEntropy(B, C); // a = c > 0.5 ? -log(b) : -log(1.0 - b) + A2 = (C > (real)0.5).condition(-(B.log()), -((B.constant(1.0f) - B).log())); + TensorCheckErr(A1, A2); +} + +template +void testTensorBinaryLabelCrossEntropyBp(Tensor& A1, + Tensor& A2, + Tensor& B, + Tensor& C) { + // a += c > 0.5 ? -1.0 / b : 1.0 / (1.0 - b) + A1.binaryLabelCrossEntropyBp(B, C); + A2 += (C > (real)0.5) + .condition((B.constant(-1.0f) / B), + (B.constant(1.0f) - B).reciprocal()); + TensorCheckErr(A1, A2); +} + +template +void testTensorLogisticRegressionLoss(Tensor& A1, + Tensor& A2, + Tensor& B, + Tensor& C) { + SetTensorValue(B, 50.0f); + SetTensorValue(B, -50.0f); + /** + * const T THRESHOLD = 40.0; + * T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) + * ? -THRESHOLD + * : b; + * a = log(1 + exp(x)) - c * x + */ + A1.logisticRegressionLoss(B, C); + real THRESHOLD = 40.0; + auto tmp = + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); + A2 = (C.constant(1.0f) + tmp.exp()).log() - C * tmp; + TensorCheckErr(A1, A2); +} + +template +void testTensorLogisticRegressionLossBp(Tensor& A1, + Tensor& A2, + Tensor& B, + Tensor& C) { + SetTensorValue(B, 50.0f); + SetTensorValue(B, -50.0f); + /** + * const T THRESHOLD = 40.0; + * T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) + * ? -THRESHOLD + * : b; + * x = exp(x); a = x / (1 + x) - c + */ + A1.logisticRegressionLossBp(B, C); + real THRESHOLD = 40.0; + auto tmp = + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); + auto tmp2 = tmp.exp(); + A2 = tmp2 / (C.constant(1.0) + tmp2) - C; + TensorCheckErr(A1, A2); +} + +template +void testTensorBiggerThan(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.biggerThan(B, C); // a = (b > c) ? 1.0f : 0.0f + A2 = (B > C).condition((real)1.0f, (real)0.0f); + TensorCheckEqual(A1, A2); +} + +template +void testTensorMax(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + A1.max2(B, C); // a = (b > c) ? b : c + A2 = (B > C).condition(B, C); + TensorCheckEqual(A1, A2); +} + +template +void testTernaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { + testTensorBinaryLabelCrossEntropyBp(A1, A2, B, C); + testTensorBinaryLabelCrossEntropy(A1, A2, B, C); + testTensorBiggerThan(A1, A2, B, C); + testTensorMax(A1, A2, B, C); + + testTensorLogisticRegressionLoss(A1, A2, B, C); + testTensorLogisticRegressionLossBp(A1, A2, B, C); +} + +TEST(Ternary, CompareOp) { + TestTernaryMatrix testCpu(testTernaryCompareOp); + +#ifdef PADDLE_WITH_GPU + TestTernaryMatrix testGpu(testTernaryCompareOp); +#endif +} + +template +void testQuaternaryAdd( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { + // A1.add3(B, C, D, 1.5f, 2.5f, 3.5f); // a = p1 * b + p2 * c + p3 * d + // A2 = B * 1.5f + C * 2.5f + D * 3.5f; + // TensorCheckEqual(A1, A2); + + /* + * T tmp = p1 * b + p2 * c + p3 * d; + * a += tmp * tmp + */ + real p1 = 1.5f; + real p2 = 2.5f; + real p3 = 3.5f; + A1.addSquareSum(B, C, D, p1, p2, p3); + auto tmp = B * p1 + C * p2 + D * p3; + A2 += tmp * tmp; + TensorCheckEqual(A1, A2); +} + +TEST(Quaternary, BaseOp) { + TestQuaternaryMatrix testCpu(testQuaternaryAdd); + +#ifdef PADDLE_WITH_GPU + TestQuaternaryMatrix testGpu(testQuaternaryAdd); +#endif +} + +template +void testTensorBiggerThan( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { + // a = ((b > c && d > 0.5f) || (b < c && d < 0.5f)) ? 1.0f : 0.0f); + A1.biggerThan(B, C, D); + A2 = ((B > C && D > (real)0.5) || (B < C && D < (real)0.5)) + .condition((real)1.0, (real)0.0); + TensorCheckEqual(A1, A2); +} + +template +void testTensorRankLoss( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { + /** + * const T THRESHOLD = 40.0; a = b - c; + * a = (a > THRESHOLD) + * ? THRESHOLD + * : ((a < -THRESHOLD) ? (-THRESHOLD) : a); + * a = log(1 + exp(a)) - a * d + */ + A1.rankLoss(B, C, D); + + real THRESHOLD = 40.0; + auto tmp = B - C; + auto tmp2 = + (tmp > THRESHOLD) + .condition(THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); + A2 = (D.constant(1.0f) + tmp2.exp()).log() - tmp2 * D; + + TensorCheckErr(A1, A2); +} + +template +void testTensorRankLossBp( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { + /** + * const T THRESHOLD = 40.0; a = b - c; + * a = (a > THRESHOLD) + * ? THRESHOLD + * : ((a < -THRESHOLD) ? (-THRESHOLD) : a); + * a = exp(a); a = (a / (1 + a) - d) + */ + A1.rankLossBp(B, C, D); + real THRESHOLD = 40.0; + auto tmp = B - C; + auto tmp2 = + (tmp > THRESHOLD) + .condition(THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); + auto tmp3 = tmp2.exp(); + A2 = tmp3 / (D.constant(1.0f) + tmp3) - D; + + TensorCheckErr(A1, A2); +} + +template +void testQuaternaryCompareOp( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { + testTensorBiggerThan(A1, A2, B, C, D); + testTensorRankLoss(A1, A2, B, C, D); + testTensorRankLossBp(A1, A2, B, C, D); +} + +TEST(Quaternary, CompareOp) { + TestQuaternaryMatrix testCpu(testQuaternaryCompareOp); + +#ifdef PADDLE_WITH_GPU + TestQuaternaryMatrix testGpu(testQuaternaryCompareOp); +#endif +} diff --git a/paddle/legacy/math/tests/test_TrainingAlgorithm.cpp b/paddle/legacy/math/tests/test_TrainingAlgorithm.cpp new file mode 100644 index 0000000000000000000000000000000000000000..214ae8971ae953ce0266f03dc3bba8c6160f1cf6 --- /dev/null +++ b/paddle/legacy/math/tests/test_TrainingAlgorithm.cpp @@ -0,0 +1,461 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "OriginalOptimizerApi.h" +#include "PerfUtils.h" +#include "TensorCheck.h" +#include "paddle/legacy/math/TrainingAlgorithmOp.h" +#include "paddle/legacy/utils/Util.h" + +using namespace paddle; // NOLINT + +#ifndef PADDLE_TYPE_DOUBLE +DEFINE_double(max_diff, 1e-5, "max diff allowed"); +#else +DEFINE_double(max_diff, 1e-13, "max diff allowed"); +#endif + +class SetMaxDiff { + public: + explicit SetMaxDiff(double max_diff) { + max_diff_ = FLAGS_max_diff; + FLAGS_max_diff = max_diff; + } + ~SetMaxDiff() { FLAGS_max_diff = max_diff_; } + + private: + double max_diff_; +}; + +#define COPY_VECTOR_TO_CPU(cpuVec, vector) \ + do { \ + if (vector->useGpu()) { \ + cpuVec = Vector::create(vector->getSize(), false); \ + cpuVec->copyFrom(*vector); \ + } else { \ + cpuVec = vector; \ + } \ + } while (0) + +int VectorCheckErr(const Vector& vector1, const Vector& vector2) { + CHECK(vector1.getSize() == vector2.getSize()); + + const real* data1 = vector1.getData(); + const real* data2 = vector2.getData(); + size_t size = vector1.getSize(); + int count = 0; + for (size_t i = 0; i < size; i++) { + real a = data1[i]; + real b = data2[i]; + if (fabs(a - b) > FLAGS_max_diff) { + if ((fabsf(a - b) / fabsf(a)) > (FLAGS_max_diff / 10.0f)) { + count++; + } + } + } + + return count; +} + +int VectorCheckErr(const VectorPtr& vector1, const VectorPtr& vector2) { + VectorPtr tmp1; + VectorPtr tmp2; + COPY_VECTOR_TO_CPU(tmp1, vector1); + COPY_VECTOR_TO_CPU(tmp2, vector2); + return VectorCheckErr(*tmp1, *tmp2); +} + +#ifdef PADDLE_DISABLE_TIMER + +#define CHECK_VECTORPTR(vector1, vector2) \ + EXPECT_EQ(VectorCheckErr(vector1, vector2), 0) + +#else + +#define CHECK_VECTORPTR(vector1, vector2) + +#endif + +typedef std::function testMatrixFunc; + +void testCase(testMatrixFunc matrixFunc) { +#ifdef PADDLE_WITH_CUDA + for (auto useGpu : {false, true}) { +#else + for (auto useGpu : {false}) { +#endif + for (auto size : {1, + 32, + 64, + 128, + 512, + 1024, + 4096, + 32768, + 65536, + 131072, + 262144, + 524288, + 1048576, + 2097152}) { + LOG(INFO) << " size=" << size << " useGpu=" << useGpu; + matrixFunc(size, useGpu); + } + } +} + +#define INIT_VECTOR(vec1, vec2, type, size, useGpu) \ + vec1[type] = Vector::create(size, useGpu); \ + vec2[type] = Vector::create(size, useGpu); \ + vec1[type]->rand(); \ + vec2[type]->copyFrom(*vec1[type]); + +void testAdagrad(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM1, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); + + real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT + real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT + real momentum = (real)rand() / (real)RAND_MAX; // NOLINT + real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT + + EXPRESSION_PERFORMANCE(AdagradParameterOptimizer( + bufs1, epsilon, learningRate, momentum, decayRate)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; + BaseMatrix& accum_buffer = *bufs2[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& accum = *bufs2[PARAMETER_GRADIENT_SQURESUM1]; + BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; + + EXPRESSION_PERFORMANCE(adagradApply(value, + grad, + mom, + accum_buffer, + accum, + lr, + epsilon, + learningRate, + momentum, + decayRate)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM1], + bufs2[PARAMETER_GRADIENT_SQURESUM1]); + CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], + bufs2[PARAMETER_LEARNING_RATE]); +} + +TEST(Training, Adagrad) { testCase(testAdagrad); } + +void testAdaDelta(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM1, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); + + real rou = (real)rand() / (real)RAND_MAX; // NOLINT + real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT + real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT + real momentum = (real)rand() / (real)RAND_MAX; // NOLINT + real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT + + EXPRESSION_PERFORMANCE(AdaDeltaParameterOptimizer( + bufs1, rou, epsilon, learningRate, momentum, decayRate)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; + BaseMatrix& accum = *bufs2[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& accum_update = *bufs2[PARAMETER_GRADIENT_SQURESUM1]; + BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; + + EXPRESSION_PERFORMANCE(adadeltaApply(value, + grad, + mom, + accum, + accum_update, + lr, + rou, + epsilon, + learningRate, + momentum, + decayRate)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM], + bufs2[PARAMETER_GRADIENT_SQURESUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM1], + bufs2[PARAMETER_GRADIENT_SQURESUM1]); + CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], + bufs2[PARAMETER_LEARNING_RATE]); +} + +TEST(Training, AdaDelta) { testCase(testAdaDelta); } + +template +void testRMSProp(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM1, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); + + /* make sure 'g - f.square()' greater than 0 */ + bufs1[PARAMETER_GRADIENT_SQURESUM]->add(1.0); + bufs2[PARAMETER_GRADIENT_SQURESUM]->copyFrom( + *bufs1[PARAMETER_GRADIENT_SQURESUM]); + + real rou = (real)rand() / (real)RAND_MAX; // NOLINT + real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT + real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT + real momentum = (real)rand() / (real)RAND_MAX; // NOLINT + real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT + real accumulatedRou = rou; + + EXPRESSION_PERFORMANCE(RMSPropParameterOptimizer(bufs1, + accumulatedRou, + rou, + epsilon, + learningRate, + momentum, + decayRate, + isFirstTime)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; + BaseMatrix& sum = *bufs2[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& sum1 = *bufs2[PARAMETER_GRADIENT_SQURESUM1]; + BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; + + EXPRESSION_PERFORMANCE(rmspropApply(value, + grad, + mom, + sum, + sum1, + lr, + accumulatedRou, + rou, + epsilon, + learningRate, + momentum, + decayRate, + isFirstTime)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM], + bufs2[PARAMETER_GRADIENT_SQURESUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM1], + bufs2[PARAMETER_GRADIENT_SQURESUM1]); + CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], + bufs2[PARAMETER_LEARNING_RATE]); +} + +TEST(Training, RMSProp) { + testCase(testRMSProp); + testCase(testRMSProp); +} + +template +void testDecayedAdagrad(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); + + real rou = (real)rand() / (real)RAND_MAX; // NOLINT + real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT + real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT + real momentum = (real)rand() / (real)RAND_MAX; // NOLINT + real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT + real accumulatedRou = rou; + + if (isFirstTime) { + bufs1[PARAMETER_GRADIENT_SQURESUM]->zeroMem(); + bufs2[PARAMETER_GRADIENT_SQURESUM]->zeroMem(); + } + + EXPRESSION_PERFORMANCE(DecayedAdagradParameterOptimizer(bufs1, + accumulatedRou, + rou, + epsilon, + learningRate, + momentum, + decayRate, + isFirstTime)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; + BaseMatrix& sum = *bufs2[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; + + EXPRESSION_PERFORMANCE(decayedAdagradApply(value, + grad, + mom, + sum, + lr, + accumulatedRou, + rou, + epsilon, + learningRate, + momentum, + decayRate, + isFirstTime)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM], + bufs2[PARAMETER_GRADIENT_SQURESUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], + bufs2[PARAMETER_LEARNING_RATE]); +} + +TEST(Training, DecayedAdagrad) { + testCase(testDecayedAdagrad); + testCase(testDecayedAdagrad); +} + +void testAdam(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_SECOND_MOMENTUM, size, useGpu); + + real beta1 = (real)rand() / (real)RAND_MAX; // NOLINT + real beta2 = (real)rand() / (real)RAND_MAX; // NOLINT + real beta1_power = (real)rand() / (real)RAND_MAX; // NOLINT + real beta2_power = (real)rand() / (real)RAND_MAX; // NOLINT + real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT + real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT + + EXPRESSION_PERFORMANCE(AdamParameterOptimizer( + bufs1, beta1, beta2, beta1_power, beta2_power, epsilon, learningRate)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; + BaseMatrix& v = *bufs2[PARAMETER_SECOND_MOMENTUM]; + + EXPRESSION_PERFORMANCE(adamApply(value, + grad, + mom, + v, + beta1, + beta2, + beta1_power, + beta2_power, + epsilon, + learningRate)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_SECOND_MOMENTUM], + bufs2[PARAMETER_SECOND_MOMENTUM]); +} + +TEST(Training, Adam) { testCase(testAdam); } + +void testAdamax(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_WEIGHTED_INFINITY_NORM, size, useGpu); + + real beta1 = (real)rand() / (real)RAND_MAX; // NOLINT + real beta2 = (real)rand() / (real)RAND_MAX; // NOLINT + real alpha = (real)rand() / (real)RAND_MAX; // NOLINT + int64_t step = 2; + + EXPRESSION_PERFORMANCE( + AdamaxParameterOptimizer(bufs1, beta1, beta2, step, alpha)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; + BaseMatrix& u = *bufs2[PARAMETER_WEIGHTED_INFINITY_NORM]; + + EXPRESSION_PERFORMANCE( + adamaxApply(value, grad, mom, u, beta1, beta2, step, alpha)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); + CHECK_VECTORPTR(bufs1[PARAMETER_WEIGHTED_INFINITY_NORM], + bufs2[PARAMETER_WEIGHTED_INFINITY_NORM]); +} + +TEST(Training, Adamax) { +#ifndef PADDLE_TYPE_DOUBLE + SetMaxDiff diff(1e-4); +#endif + testCase(testAdamax); +} + +void testSparseMomentum(size_t size, bool useGpu) { + VectorPtr bufs1[NUM_PARAMETER_TYPES]; + VectorPtr bufs2[NUM_PARAMETER_TYPES]; + INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM_UT, size, useGpu); + INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM_VT, size, useGpu); + + real alpha = (real)rand() / (real)RAND_MAX; // NOLINT + real beta = (real)rand() / (real)RAND_MAX; // NOLINT + real gamma = (real)rand() / (real)RAND_MAX; // NOLINT + real tau = (real)rand() / (real)RAND_MAX; // NOLINT + real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT + + EXPRESSION_PERFORMANCE(SparseMomentumParameterOptimizer( + bufs1, alpha, beta, gamma, tau, learningRate)); + + BaseMatrix& value = *bufs2[PARAMETER_VALUE]; + BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; + BaseMatrix& momU = *bufs2[PARAMETER_MOMENTUM_UT]; + BaseMatrix& momV = *bufs2[PARAMETER_MOMENTUM_VT]; + + EXPRESSION_PERFORMANCE(sparseMomentumApply( + value, grad, momU, momV, alpha, beta, gamma, tau, learningRate)); + + CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM_UT], bufs2[PARAMETER_MOMENTUM_UT]); + CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM_VT], bufs2[PARAMETER_MOMENTUM_VT]); +} + +TEST(Training, SparseMomentum) { testCase(testSparseMomentum); } diff --git a/paddle/math/tests/test_batchTranspose.cpp b/paddle/legacy/math/tests/test_batchTranspose.cpp similarity index 100% rename from paddle/math/tests/test_batchTranspose.cpp rename to paddle/legacy/math/tests/test_batchTranspose.cpp diff --git a/paddle/legacy/math/tests/test_lazyAssign.cu b/paddle/legacy/math/tests/test_lazyAssign.cu new file mode 100644 index 0000000000000000000000000000000000000000..cf8c3d77199571dff314446a1e1b14e9b746e947 --- /dev/null +++ b/paddle/legacy/math/tests/test_lazyAssign.cu @@ -0,0 +1,147 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "PerfUtils.h" +#include "TensorCheck.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/TensorAssign.h" + +using paddle::BaseMatrix; +using paddle::CpuMatrix; +using paddle::GpuMatrix; +using autotest::TensorCheckEqual; +using autotest::TensorCheckErr; + +typedef std::function testMatrixFunc; +void testMatrixCase(testMatrixFunc matrixFunc) { + for (auto height : {1}) { + for (auto width : {1, + 32, + 64, + 128, + 512, + 1024, + 4096, + 32768, + 65536, + 131072, + 262144, + 524288, + 1048576, + 2097152, + 4194304, + 8388608}) { + matrixFunc(height, width); + } + } +} + +template +void testLazyAssign(int height, int width) { + Tensor A1(height, width); + Tensor A2(height, width); + Tensor B(height, width); + Tensor C(height, width); + Tensor D(height, width); + A1.randomizeUniform(); + B.randomizeUniform(); + C.randomizeUniform(); + D.randomizeUniform(); + A2.copyFrom(A1); + + EXPRESSION_PERFORMANCE(A1 = B + C; A1 = A1 * D;); + + EXPRESSION_PERFORMANCE(auto expr1 = A2.lazyAssign(B + C); + auto expr2 = A2.lazyAssign(A2 * D); + AssignEvaluate(expr1, expr2);); + + TensorCheckErr(A1, A2); +} + +TEST(lazyAssign, CPU) { testMatrixCase(testLazyAssign); } + +#ifdef PADDLE_WITH_GPU +TEST(lazyAssign, GPU) { testMatrixCase(testLazyAssign); } +#endif + +template +void sgdUpdateTensor( + Tensor& A, Tensor& B, Tensor& C, Tensor& D, real p1, real p2, real p3) { + C = C * p2 - D * (B + A * p3) * p1; + A += C; +} + +void sgdUpdateLazyAssign(BaseMatrix& A, + BaseMatrix& B, + BaseMatrix& C, + BaseMatrix& D, + real p1, + real p2, + real p3) { + auto expr1 = C.lazyAssign(C * p2 - D * (B + A * p3) * p1); + auto expr2 = A.lazyAssign(A + C); + AssignEvaluate(expr1, expr2); +} + +template +void testSgdUpdate(int height, int width) { + Tensor A1(height, width); + Tensor A2(height, width); + Tensor A3(height, width); + A1.randomizeUniform(); + A2.copyFrom(A1); + A3.copyFrom(A1); + + Tensor B(height, width); + B.randomizeUniform(); + + Tensor C1(height, width); + Tensor C2(height, width); + Tensor C3(height, width); + C1.randomizeUniform(); + C2.copyFrom(C1); + C3.copyFrom(C1); + + Tensor D(height, width); + D.randomizeUniform(); + + real p1 = 0.2; + real p2 = 0.3; + real p3 = 0.5; + + /** + * c = p2 * c - p1 * (b + p3 * a); + * a = a + c; + */ + // BaseMatrix API + EXPRESSION_PERFORMANCE(A1.sgdUpdate(B, C1, D, p1, p2, p3);); + + // Tensor expression + EXPRESSION_PERFORMANCE(sgdUpdateTensor(A2, B, C2, D, p1, p2, p3)); + + // lazyAssign + EXPRESSION_PERFORMANCE(sgdUpdateLazyAssign(A3, B, C3, D, p1, p2, p3)); + + TensorCheckErr(A1, A2); + TensorCheckErr(A1, A3); + TensorCheckErr(C1, C2); + TensorCheckErr(C1, C3); +} + +TEST(sgdUpdate, CPU) { testMatrixCase(testSgdUpdate); } + +#ifdef PADDLE_WITH_GPU +TEST(sgdUpdate, GPU) { testMatrixCase(testSgdUpdate); } +#endif diff --git a/paddle/legacy/math/tests/test_matrixCompare.cpp b/paddle/legacy/math/tests/test_matrixCompare.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a43adde46fc6526cc3ff5affec2ce1c7c3a44214 --- /dev/null +++ b/paddle/legacy/math/tests/test_matrixCompare.cpp @@ -0,0 +1,1698 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +/// This unittest checks GpuMatrix/CpuMatrix get same result, so disable when +/// only cpu version. + +#include +#include "TensorCheck.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseMatrix.h" +#include "paddle/legacy/utils/DynamicLoader.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT +using autotest::TensorCheckEqual; +using autotest::TensorCheckErr; + +void testMatrixMaxSequence(int batchSize, int inputDim) { + // forward + MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); + MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + + IVectorPtr cpuSequence; + generateSequenceStartPositions(batchSize, cpuSequence); + IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); + gpuSequence->copyFrom(*cpuSequence); + + int newBatchSize = cpuSequence->getSize() - 1; + MatrixPtr cpuOutput = std::make_shared(newBatchSize, inputDim); + MatrixPtr gpuOutput = std::make_shared(newBatchSize, inputDim); + cpuOutput->zero(); + gpuOutput->zero(); + + IVectorPtr cpuIndex = nullptr; + IVectorPtr gpuIndex = nullptr; + IVector::resizeOrCreate(cpuIndex, newBatchSize * inputDim, false); + IVector::resizeOrCreate(gpuIndex, newBatchSize * inputDim, true); + cpuIndex->zeroMem(); + gpuIndex->zeroMem(); + + cpuOutput->maxSequenceForward(*cpuInput, *cpuSequence, *cpuIndex); + gpuOutput->maxSequenceForward(*gpuInput, *gpuSequence, *gpuIndex); + + TensorCheckEqual(*cpuOutput, *gpuOutput); + TensorCheckEqual(*cpuIndex, *gpuIndex); + + // backward + MatrixPtr cpuOutputGrad = std::make_shared(newBatchSize, inputDim); + MatrixPtr gpuOutputGrad = std::make_shared(newBatchSize, inputDim); + cpuOutputGrad->randomizeUniform(); + gpuOutputGrad->copyFrom(*cpuOutputGrad); + + MatrixPtr cpuInputGrad = std::make_shared(batchSize, inputDim); + MatrixPtr gpuInputGrad = std::make_shared(batchSize, inputDim); + cpuInputGrad->randomizeUniform(); + gpuInputGrad->copyFrom(*cpuInputGrad); + + cpuInputGrad->maxSequenceBackward(*cpuOutputGrad, *cpuSequence, *cpuIndex); + gpuInputGrad->maxSequenceBackward(*gpuOutputGrad, *gpuSequence, *gpuIndex); + + TensorCheckEqual(*cpuInputGrad, *gpuInputGrad); +} + +TEST(Matrix, maxSequence) { + for (auto batchSize : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 + for (auto inputDim : {1, 7, 131}) { // prime numbers close to 1, 8, 128 + VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim; + testMatrixMaxSequence(batchSize, inputDim); + } + } +} + +void testMatrixGetSum(int height, int width) { + MatrixPtr cpuInput = std::make_shared(height, width); + MatrixPtr gpuInput = std::make_shared(height, width); + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + +#ifndef PADDLE_TYPE_DOUBLE + int x = log10(height * width); + real err = 1e-6 * pow(10, x); +#else + real err = 1e-8; +#endif + + real cpuSum = cpuInput->getSum(); + real gpuSum = gpuInput->getSum(); + + EXPECT_LE(fabs(cpuSum - gpuSum), err); +} + +void testMatrixGetMinMax(int height, int width) { + MatrixPtr cpuInput = std::make_shared(height, width); + MatrixPtr gpuInput = std::make_shared(height, width); + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + + real cpuMin = cpuInput->getMin(); + real gpuMin = gpuInput->getMin(); + real cpuMax = cpuInput->getMax(); + real gpuMax = gpuInput->getMax(); + + EXPECT_EQ(cpuMin, gpuMin); + EXPECT_EQ(cpuMax, gpuMax); +} + +void testMatrixZeroAtOffset(int height, int width) { + MatrixPtr cpuA = std::make_shared(height, width); + MatrixPtr gpuA = std::make_shared(height, width); + MatrixPtr cpuTest = std::make_shared(height, width); + + cpuA->randomizeUniform(); + gpuA->copyFrom(*cpuA); + cpuTest->copyFrom(*cpuA); + + int columnOffset = rand() % width; // NOLINT we just use rand() for test. + int numColumns = rand() % (width - columnOffset); // NOLINT + + if (numColumns == 0) return; + + cpuA->zeroAtOffset(columnOffset, numColumns); + gpuA->zeroAtOffset(columnOffset, numColumns); + + /* cpuTest */ + real* a = cpuTest->getData() + columnOffset; + for (int64_t i = 0; i < height; ++i) { + for (int64_t j = 0; j < numColumns; ++j) { + a[i * width + j] = 0; + } + } + + TensorCheckEqual(*cpuA, *gpuA); + TensorCheckEqual(*cpuA, *cpuTest); +} + +void testMatrixDeepSwap(int height, int width) { + MatrixPtr cpuA = std::make_shared(height, width); + MatrixPtr cpuB = std::make_shared(height, width); + MatrixPtr cpuCopyA = std::make_shared(height, width); + MatrixPtr cpuCopyB = std::make_shared(height, width); + + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + cpuCopyA->copyFrom(*cpuA); + cpuCopyB->copyFrom(*cpuB); + + // swap matrix cpuA and cpuB + cpuA->deepSwap(*cpuB); + + TensorCheckEqual(*cpuA, *cpuCopyB); + TensorCheckEqual(*cpuB, *cpuCopyA); +} + +void testMatrixTranspose(int height, int width) { + MatrixPtr cpu = std::make_shared(height, width); + MatrixPtr gpu = std::make_shared(height, width); + MatrixPtr cpuT = std::make_shared(width, height); + MatrixPtr gpuT = std::make_shared(width, height); + + cpu->randomizeUniform(); + gpu->copyFrom(*cpu); + cpu->transpose(cpuT, false); + gpu->transpose(gpuT, true); + + TensorCheckEqual(*cpuT, *gpuT); +} + +void testMatrixRotate(int height, int width) { + MatrixPtr cpu = std::make_shared(height, width); + MatrixPtr gpu = std::make_shared(height, width); + MatrixPtr cpuR = std::make_shared(width, height); + MatrixPtr gpuR = std::make_shared(width, height); + + cpu->randomizeUniform(); + gpu->copyFrom(*cpu); + + cpu->rotate(cpuR, false, true); + gpu->rotate(gpuR, true, true); + TensorCheckEqual(*cpuR, *gpuR); + + cpu->rotate(cpuR, true, false); + gpu->rotate(gpuR, false, false); + TensorCheckEqual(*cpuR, *gpuR); +} + +void testMatrixInverse(int height) { + MatrixPtr cpu = std::make_shared(height, height); + MatrixPtr gpu = std::make_shared(height, height); + MatrixPtr cpuI = std::make_shared(height, height); + MatrixPtr gpuI = std::make_shared(height, height); + + /* Make matrix well conditioned: cpu * cpuT + Identity */ + cpu->randomizeUniform(); + MatrixPtr cpuT = cpu->getTranspose(); + MatrixPtr outputCheck = std::make_shared(height, height); + outputCheck->mul(*cpu, *cpuT); + cpu->setDiag(1.0); + cpu->add(*outputCheck); + + gpu->copyFrom(*cpu); + cpu->inverse(cpuI, true); + gpu->inverse(gpuI, false); + + TensorCheckErr(*cpuI, *gpuI); + + outputCheck->mul(*cpu, *cpuI); + cpu->setDiag(1.0); + TensorCheckErr(*cpu, *outputCheck); +} + +TEST(Matrix, unary) { + for (auto height : {1, 3, 11, 73, 128, 200, 330}) { + for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { + VLOG(3) << " height=" << height << " width=" << width; + + testMatrixDeepSwap(height, width); + testMatrixZeroAtOffset(height, width); + testMatrixGetSum(height, width); + testMatrixTranspose(height, width); + testMatrixRotate(height, width); + } +#ifdef LAPACK_FOUND + // inverse matrix + testMatrixInverse(height); +#else + LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK" + << "support so we cannot test matrix inverse. To test " + << "matrix inverse, please install LAPACKE " + << "and MKL/Openblas, and re-build PaddlePaddle."; +#endif + } +} + +void testMatrixSoftmax(int height, int width) { + MatrixPtr cpuInput = std::make_shared(height, width); + MatrixPtr cpuOutput = std::make_shared(height, width); + MatrixPtr gpuInput = std::make_shared(height, width); + MatrixPtr gpuOutput = std::make_shared(height, width); + + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + cpuOutput->zero(); + gpuOutput->zero(); + cpuInput->softmax(*cpuOutput); + gpuInput->softmax(*gpuOutput); + + TensorCheckErr(*cpuOutput, *gpuOutput); +} + +void testSequenceSoftmax(int batchSize) { + // forward + int inputDim = 1; + MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); + MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + + IVectorPtr cpuSequence; + generateSequenceStartPositions(batchSize, cpuSequence); + IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); + gpuSequence->copyFrom(*cpuSequence); + + cpuInput->sequenceSoftmax(*cpuInput, *cpuSequence); + gpuInput->sequenceSoftmax(*gpuInput, *gpuSequence); + + TensorCheckErr(*cpuInput, *gpuInput); +} + +void testMatrixSoftmaxThreshold(int height, int width) { + MatrixPtr cpuInput = std::make_shared(height, width); + MatrixPtr cpuOutput = std::make_shared(height, width); + MatrixPtr gpuInput = std::make_shared(height, width); + MatrixPtr gpuOutput = std::make_shared(height, width); + + cpuInput->randomizeUniform(); + cpuInput->getData()[0] = 100.0; + gpuInput->copyFrom(*cpuInput); + cpuOutput->zero(); + gpuOutput->zero(); + cpuInput->softmax(*cpuOutput); + gpuInput->softmax(*gpuOutput); + + MatrixPtr outputCheck = std::make_shared(height, width); + outputCheck->copyFrom(*gpuOutput); + // check output zero + int cpuCount = 0; + int gpuCount = 0; + auto zeroNum = [](MatrixPtr out, int& count) { + for (size_t i = 0; i < out->getHeight(); i++) { + for (size_t j = 0; j < out->getWidth(); j++) { + if (out->getElement(i, j) == 0) count++; + } + } + }; + zeroNum(cpuOutput, cpuCount); + zeroNum(outputCheck, gpuCount); + EXPECT_EQ(cpuCount, 0) << "Cpu softmax output value 0"; + EXPECT_EQ(gpuCount, 0) << "Gpu softmax output value 0"; +} + +void testMatrixSoftmaxBp(int height, int width) { + MatrixPtr cpuInput = std::make_shared(height, width); + MatrixPtr cpuOutput = std::make_shared(height, width); + MatrixPtr gpuInput = std::make_shared(height, width); + MatrixPtr gpuOutput = std::make_shared(height, width); + + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + cpuOutput->randomizeUniform(); + gpuOutput->copyFrom(*cpuOutput); + gpuOutput->softmaxBackward(*gpuInput); + + MatrixPtr sftMaxSum = std::make_shared(height, 1); + MatrixPtr sftMaxDot = std::make_shared(height, width); + sftMaxDot->dotMul(*cpuOutput, *cpuInput); + sftMaxSum->colMerge(*sftMaxDot); + cpuOutput->softmaxDerivative(*cpuInput, *sftMaxSum); + + TensorCheckErr(*cpuOutput, *gpuOutput); +} + +TEST(Matrix, softmax) { + for (auto height : {1, 3, 131}) { // prime numbers close to 1, 4, 127 + for (auto width : {1, 17, 251}) { // prime numbers close to 1, 16, 256 + VLOG(3) << " height=" << height << " width=" << width; + + testMatrixSoftmax(height, width); + testMatrixSoftmaxBp(height, width); + testMatrixSoftmaxThreshold(height, width); + } + testSequenceSoftmax(height); + } +} + +void testMatrixAddToRows(int numSamples, int tableSize, int inputDim) { + MatrixPtr cpuTable = std::make_shared(tableSize, inputDim); + MatrixPtr gpuTable = std::make_shared(tableSize, inputDim); + cpuTable->randomizeUniform(); + gpuTable->copyFrom(*cpuTable); + + IVectorPtr cpuIds; + IVectorPtr gpuIds; + cpuIds = VectorT::create(numSamples, false); + gpuIds = VectorT::create(numSamples, true); + cpuIds->rand(tableSize); + gpuIds->copyFrom(*cpuIds); + + MatrixPtr cpuOutput = std::make_shared(numSamples, inputDim); + MatrixPtr gpuOutput = std::make_shared(numSamples, inputDim); + cpuOutput->randomizeUniform(); + gpuOutput->copyFrom(*cpuOutput); + + cpuOutput->addToRows(*cpuTable, *cpuIds); + gpuOutput->addToRows(*gpuTable, *gpuIds); + + TensorCheckErr(*cpuTable, *gpuTable); +} + +TEST(Matrix, tableProjection) { + for (auto numSamples : {10, 100, 1000, 10000, 80000}) { + for (auto tableSize : {10, 100}) { + for (auto inputDim : {20, 50}) { + VLOG(3) << " numSamples=" << numSamples << " tableSize=" << tableSize + << " inputDim=" << inputDim; + testMatrixAddToRows(numSamples, tableSize, inputDim); + } + } + } +} + +void testMatrixMul(bool transa, bool transb, int dimM, int dimN, int dimK) { + int heightA = transa == false ? dimM : dimK; + int widthA = transa == false ? dimK : dimM; + int heightB = transb == false ? dimK : dimN; + int widthB = transb == false ? dimN : dimK; + int heightC = dimM; + int widthC = dimN; + + MatrixPtr cpuA = std::make_shared(heightA, widthA, transa); + MatrixPtr cpuB = std::make_shared(heightB, widthB, transb); + MatrixPtr cpuC = std::make_shared(heightC, widthC); + MatrixPtr gpuA = std::make_shared(heightA, widthA, transa); + MatrixPtr gpuB = std::make_shared(heightB, widthB, transb); + MatrixPtr gpuC = std::make_shared(heightC, widthC); + + real alpha = 1.5; + real beta = 2.0; + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + cpuC->randomizeUniform(); + gpuA->copyFrom(*cpuA); + gpuB->copyFrom(*cpuB); + gpuC->copyFrom(*cpuC); + + cpuC->mul(*cpuA, *cpuB, alpha, beta); + gpuC->mul(*gpuA, *gpuB, alpha, beta); + + TensorCheckErr(*cpuC, *gpuC); +} + +void testSubMatrixMul(bool transa, bool transb, int dimM, int dimN, int dimK) { + int heightA = transa == false ? dimM : dimK; + int widthA = transa == false ? dimK : dimM; + int heightB = transb == false ? dimK : dimN; + int widthB = transb == false ? dimN : dimK; + int heightC = dimM; + int widthC = dimN; + + MatrixPtr cpuA = std::make_shared(heightA, widthA, transa); + MatrixPtr cpuB = std::make_shared(heightB, widthB, transb); + MatrixPtr cpuC = std::make_shared(heightC, widthC); + MatrixPtr gpuA = std::make_shared(heightA, widthA, transa); + MatrixPtr gpuB = std::make_shared(heightB, widthB, transb); + MatrixPtr gpuC = std::make_shared(heightC, widthC); + + real alpha = 1.5; + real beta = 2.0; + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + cpuC->randomizeUniform(); + gpuA->copyFrom(*cpuA); + gpuB->copyFrom(*cpuB); + gpuC->copyFrom(*cpuC); + + auto subSize = [](int& start, int& end, int dim) { + if (dim == 1) { + start = 0; + end = dim; + } else { + int subDim = rand() % (dim - 1) + 1; // NOLINT + start = rand() % (dim - subDim); // NOLINT + end = start + subDim; + } + }; + + auto subMatrix = [](MatrixPtr& sub, + MatrixPtr matrix, + size_t startRow, + size_t endRow, + size_t startCol, + size_t endCol) { + if (!matrix->isTransposed()) { + sub = matrix->subMatrix(startRow, endRow, startCol, endCol); + } else { + sub = matrix->subMatrix(startCol, endCol, startRow, endRow); + } + }; + + int startM, endM; + int startN, endN; + int startK, endK; + subSize(startM, endM, dimM); + subSize(startN, endN, dimN); + subSize(startK, endK, dimK); + + MatrixPtr subCpuA; + MatrixPtr subCpuB; + MatrixPtr subGpuA; + MatrixPtr subGpuB; + subMatrix(subCpuA, cpuA, startM, endM, startK, endK); + subMatrix(subGpuA, gpuA, startM, endM, startK, endK); + subMatrix(subCpuB, cpuB, startK, endK, startN, endN); + subMatrix(subGpuB, gpuB, startK, endK, startN, endN); + MatrixPtr subCpuC = cpuC->subMatrix(startM, endM, startN, endN); + MatrixPtr subGpuC = gpuC->subMatrix(startM, endM, startN, endN); + + subCpuC->mul(*subCpuA, *subCpuB, alpha, beta); + subGpuC->mul(*subGpuA, *subGpuB, alpha, beta); + + TensorCheckErr(*cpuC, *gpuC); +} + +TEST(Matrix, mul) { + for (auto transa : {false, true}) { + for (auto transb : {false, true}) { + for (auto dimM : {1, 9, 53, 127, 345, 1023, 2135}) { + for (auto dimN : {1, 5, 37, 256, 1024}) { + for (auto dimK : {8, 45, 346, 784, 1025}) { + if (true == transa && true == transb) { + continue; + } + VLOG(3) << setiosflags(ios::left) << setfill(' ') + << " transa=" << transa << " transb=" << transb + << " dimM=" << setw(5) << dimM << " dimN=" << setw(5) + << dimN << " dimK=" << setw(5) << dimK; + + testMatrixMul(transa, transb, dimM, dimN, dimK); + testSubMatrixMul(transa, transb, dimM, dimN, dimK); + } + } + } + } + } +} + +void testVectorRowFunc(int size) { + CpuVectorPtr cpu = std::make_shared>(size); + GpuVectorPtr gpu = std::make_shared>(size); + + cpu->rand(); + gpu->copyFrom(*cpu); + + EXPECT_EQ(cpu->getMax(), gpu->getMax()); + EXPECT_EQ(cpu->getMin(), gpu->getMin()); + EXPECT_EQ(cpu->getAbsMax(), gpu->getAbsMax()); +} + +TEST(Vector, rowFunc) { + for (auto size : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 + VLOG(3) << " size=" << size; + testVectorRowFunc(size); + } +} + +template +void testVectorReset(int size) { + std::shared_ptr> cpu = std::make_shared>(size); + std::shared_ptr> gpu = std::make_shared>(size); + + T value = (T)((int)rand() % 100 + 1.0f / ((int)rand() % 100)); + cpu->reset(value); + gpu->reset(value); + + TensorCheckEqual(*cpu, *gpu); +} + +template +void testVecortSelectFrom(int size) { + std::shared_ptr> cpuDst = std::make_shared>(size); + std::shared_ptr> gpuDst = std::make_shared>(size); + std::shared_ptr> cpuSrc = + std::make_shared>(size * 2); + std::shared_ptr> gpuSrc = + std::make_shared>(size * 2); + CpuIVectorPtr cpuIds = std::make_shared>(size); + GpuIVectorPtr gpuIds = std::make_shared>(size); + + if (std::is_same::value) { + cpuSrc->rand(); + } else { + cpuSrc->rand(100000); + } + gpuSrc->copyFrom(*cpuSrc); + cpuIds->rand(size); + gpuIds->copyFrom(*cpuIds); + + cpuDst->selectFrom(*cpuSrc, *cpuIds); + gpuDst->selectFrom(*gpuSrc, *gpuIds); + + TensorCheckEqual(*cpuDst, *gpuDst); +} + +template +void testVecotrZeroMem(int size) { + std::shared_ptr> cpu = std::make_shared>(size); + std::shared_ptr> gpu = std::make_shared>(size); + + cpu->zeroMem(); + gpu->zeroMem(); + + TensorCheckEqual(*cpu, *gpu); +} + +template +void testVectorIsEqual(int size) { + std::shared_ptr> cpuA = std::make_shared>(size); + std::shared_ptr> cpuB = std::make_shared>(size); + std::shared_ptr> gpuA = std::make_shared>(size); + std::shared_ptr> gpuB = std::make_shared>(size); + + if (std::is_same::value) { + cpuB->rand(); + } else { + cpuB->rand(100000); + } + gpuB->copyFrom(*cpuB); + + T value = (T)((int)rand() % 100 + 1.0f / ((int)rand() % 100)); + cpuA->isEqualTo(*cpuB, value); + gpuA->isEqualTo(*gpuB, value); + + TensorCheckEqual(*cpuA, *gpuA); +} + +TEST(Vector, Equal) { + for (auto size : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 + VLOG(3) << " size=" << size; + testVectorReset(size); + testVectorReset(size); + testVecortSelectFrom(size); + testVecortSelectFrom(size); + testVecotrZeroMem(size); + testVecotrZeroMem(size); + testVectorIsEqual(size); + testVectorIsEqual(size); + } +} + +void testMatrixTopK(int samples, int dim, int beamSize) { + MatrixPtr cpuSrc = std::make_shared(samples, dim); + MatrixPtr gpuSrc = std::make_shared(samples, dim); + MatrixPtr cpuVal = std::make_shared(samples, beamSize); + MatrixPtr gpuVal = std::make_shared(samples, beamSize); + IVectorPtr cpuIds = std::make_shared(samples * beamSize); + IVectorPtr gpuIds = std::make_shared(samples * beamSize); + + cpuSrc->randomizeUniform(); + gpuSrc->copyFrom(*cpuSrc); + + cpuSrc->rowMax(*cpuIds, *cpuVal); + gpuSrc->rowMax(*gpuIds, *gpuVal); + + TensorCheckEqual(*cpuVal, *gpuVal); +} + +TEST(Matrix, topK) { + for (auto samples : {1, 17, 131}) { // prime numbers close to 1, 16, 127 + for (auto dim : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 + for (auto beamSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) { + if (beamSize > dim) continue; + VLOG(3) << " samples=" << samples << " beamSize=" << beamSize + << " dim=" << dim; + testMatrixTopK(samples, dim, beamSize); + } + } + } +} + +void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) { + int nnz = samples * dim * ratio; + if (nnz < 1) nnz = 1; // Because sparseRand in MathUtil.cpp requires this. + MatrixPtr cpuSrc = std::make_shared(samples, dim, nnz); + MatrixPtr gpuSrc = std::make_shared(samples, dim, nnz); + MatrixPtr cpuVal = std::make_shared(samples, beamSize); + MatrixPtr gpuVal = std::make_shared(samples, beamSize); + IVectorPtr cpuIds = std::make_shared(samples * beamSize); + IVectorPtr gpuIds = std::make_shared(samples * beamSize); + + cpuSrc->randomizeUniform(); + gpuSrc->copyFrom(*cpuSrc); + cpuVal->zero(); + cpuIds->zero(); + gpuVal->zero(); + gpuIds->zero(); + + cpuSrc->rowMax(*cpuIds, *cpuVal); + gpuSrc->rowMax(*gpuIds, *gpuVal); + + TensorCheckEqual(*cpuVal, *gpuVal); + + IVectorPtr outCheckIds = std::make_shared(samples * beamSize); + outCheckIds->copyFrom(*gpuIds); + + const int* data1 = cpuIds->getData(); + const int* data2 = outCheckIds->getData(); + size_t size = cpuIds->getSize(); + for (size_t i = 0; i < size; i++) { + if (data1[i] == -1 && data1[i] != data2[i]) { + EXPECT_EQ(data1[i], data2[i]); + } + } +} + +TEST(SMatrix, topK) { + for (auto samples : {1, 3, 61}) { + for (auto dim : {1, 3, 61}) { + for (auto beamSize : {1, 3, 61}) { + for (auto ratio : {0.01, 0.001}) { + if (beamSize > dim) continue; + VLOG(3) << " samples=" << samples << " beamSize=" << beamSize + << " dim=" << dim << " ratio=" << ratio; + testSMatrixTopK(samples, dim, beamSize, ratio); + } + } + } + } +} + +void testMatrixSequenceAvg(int batchSize, int inputDim, int mode) { + MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); + MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + + IVectorPtr cpuSequence; + generateSequenceStartPositions(batchSize, cpuSequence); + IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); + gpuSequence->copyFrom(*cpuSequence); + + int newBatchSize = cpuSequence->getSize() - 1; + MatrixPtr cpuOutput = std::make_shared(newBatchSize, inputDim); + MatrixPtr gpuOutput = std::make_shared(newBatchSize, inputDim); + cpuOutput->zero(); + gpuOutput->zero(); + + cpuOutput->sequenceAvgForward(*cpuInput, *cpuSequence, mode); + gpuOutput->sequenceAvgForward(*gpuInput, *gpuSequence, mode); + + TensorCheckErr(*cpuOutput, *gpuOutput); + + MatrixPtr cpuInGrad = std::make_shared(batchSize, inputDim); + MatrixPtr gpuInGrad = std::make_shared(batchSize, inputDim); + cpuInGrad->randomizeUniform(); + gpuInGrad->copyFrom(*cpuInGrad); + + cpuInGrad->sequenceAvgBackward(*cpuOutput, *cpuSequence, mode); + gpuInGrad->sequenceAvgBackward(*gpuOutput, *gpuSequence, mode); + + TensorCheckErr(*cpuInGrad, *gpuInGrad); +} + +TEST(Matrix, sequenceAvg) { + for (auto batchSize : {10, 128, 6000}) { + for (auto inputDim : {32, 100, 512}) { + for (auto mode : {0, 1, 2}) { + VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim + << " mode=" << mode; + testMatrixSequenceAvg(batchSize, inputDim, mode); + } + } + } +} + +void testParamReluBackwardDiff(int height, + int width, + int w_height, + int w_width) { + MatrixPtr oGrad = CpuMatrix::create(height, width, false, false); + MatrixPtr input = CpuMatrix::create(height, width, false, false); + MatrixPtr diff = CpuMatrix::create(height, width, false, false); + MatrixPtr w = CpuMatrix::create(w_height, w_width, false, false); + + oGrad->randomizeUniform(); + input->randomizeUniform(); + w->randomizeUniform(); + diff->randomizeUniform(); + input->add(-0.5); + + MatrixPtr oGradGpu = GpuMatrix::create(height, width, false, true); + MatrixPtr inputGpu = GpuMatrix::create(height, width, false, true); + MatrixPtr diffGpu = CpuMatrix::create(height, width, false, true); + MatrixPtr wGpu = GpuMatrix::create(w_height, w_width, false, true); + + oGradGpu->copyFrom(*oGrad); + inputGpu->copyFrom(*input); + wGpu->copyFrom(*w); + diffGpu->copyFrom(*diff); + + diff->paramReluBackwardDiff(*oGrad, *input, *w); + diffGpu->paramReluBackwardDiff(*oGradGpu, *inputGpu, *wGpu); + + TensorCheckErr(*diff, *diffGpu); +} + +TEST(Matrix, paramReluBackwardDiff) { + for (auto height : {10, 40, 100}) { + for (auto width : {10, 40, 100}) { + for (auto w_height : {1, 2}) { + for (auto w_width : {1, 2}) { + if (width % (w_height * w_width)) continue; + testParamReluBackwardDiff(height, width, w_height, w_width); + } + } + } + } +} + +void testClassificationError(int numSamples, int dim, int topkSize) { + MatrixPtr cpuError = std::make_shared(numSamples, 1); + MatrixPtr gpuError = std::make_shared(numSamples, 1); + MatrixPtr cpuOutput = std::make_shared(numSamples, dim); + MatrixPtr gpuOutput = std::make_shared(numSamples, dim); + IVectorPtr cpuLabel = std::make_shared(numSamples); + IVectorPtr gpuLabel = std::make_shared(numSamples); + + cpuOutput->randomizeUniform(); + cpuLabel->rand(dim); + gpuOutput->copyFrom(*cpuOutput); + gpuLabel->copyFrom(*cpuLabel); + + cpuError->classificationError(*cpuOutput, *cpuLabel, topkSize); + gpuError->classificationError(*gpuOutput, *gpuLabel, topkSize); + + TensorCheckEqual(*cpuError, *gpuError); +} + +TEST(Matrix, classificationError) { + for (auto numSamples : {1, 3, 31}) { + for (auto dim : {1, 3, 31}) { + for (auto topkSize : {1, 3, (int)rand() % dim + 1}) { + if (topkSize > dim) continue; + VLOG(3) << " sample= " << numSamples << " topkSize= " << topkSize + << " dim= " << dim; + testClassificationError(numSamples, dim, topkSize); + } + } + } +} + +void testMaxPoolFwdBwd(int numSamples, + int channels, + int imgSizeH, + int imgSizeW, + int ksizeH, + int ksizeW, + int strideH, + int strideW, + int padH, + int padW) { + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); + + int inWidth = imgSizeH * imgSizeW * channels; + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + int outWidth = channels * outH * outW; + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + input->randomizeUniform(); + target->randomizeUniform(); + inputGpu->copyFrom(*input); + targetGpu->copyFrom(*target); + + target->maxPoolForward(*input, + imgSizeH, + imgSizeW, + channels, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + padH, + padW); + targetGpu->maxPoolForward(*inputGpu, + imgSizeH, + imgSizeW, + channels, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + padH, + padW); + MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); + targetCheck->copyFrom(*targetGpu); + checkMatrixEqual(target, targetCheck); + + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->maxPoolBackward(*input, + imgSizeH, + imgSizeW, + *targetGrad, + *target, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + 1.0, + 1.0, + padH, + padW); + inputGpuGrad->maxPoolBackward(*inputGpu, + imgSizeH, + imgSizeW, + *targetGpuGrad, + *targetGpu, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + 1.0, + 1.0, + padH, + padW); + MatrixPtr targetBwdCheck = + CpuMatrix::create(numSamples, inWidth, false, false); + targetBwdCheck->copyFrom(*inputGpuGrad); + checkMatrixEqual(inputGrad, targetBwdCheck); +} + +void testAvgPoolFwdBwd(int numSamples, + int channels, + int imgSizeH, + int imgSizeW, + int ksizeH, + int ksizeW, + int strideH, + int strideW, + int padH, + int padW) { + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); + + int inWidth = imgSizeH * imgSizeW * channels; + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + int outWidth = channels * outH * outW; + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + input->randomizeUniform(); + target->randomizeUniform(); + inputGpu->copyFrom(*input); + targetGpu->copyFrom(*target); + + target->avgPoolForward(*input, + imgSizeH, + imgSizeW, + channels, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + padH, + padW); + targetGpu->avgPoolForward(*inputGpu, + imgSizeH, + imgSizeW, + channels, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + padH, + padW); + + TensorCheckErr(*target, *targetGpu); + + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->avgPoolBackward(*targetGrad, + imgSizeH, + imgSizeW, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + 1.0, + 1.0, + padH, + padW); + inputGpuGrad->avgPoolBackward(*targetGpuGrad, + imgSizeH, + imgSizeW, + ksizeW, + ksizeH, + strideH, + strideW, + outH, + outW, + 1.0, + 1.0, + padH, + padW); + + TensorCheckErr(*inputGrad, *inputGpuGrad); +} + +// TODO(yi): I noticed many such blindly combinatorial tests in this +// file. They are no help to locate defects at all. +TEST(Matrix, PoolFwdBwd) { + for (auto numSamples : {1, 3}) { + for (auto channels : {1, 3}) { + for (auto imgSizeH : {13, 17}) { + for (auto imgSizeW : {17, 19}) { + for (auto sizeX : {2, 3}) { + for (auto sizeY : {2, 3}) { + for (auto sH : {1, 2}) { + for (auto sW : {1, 2}) { + for (auto pH : {0, (sizeY - 1) / 2}) { + for (auto pW : {0, (sizeX - 1) / 2}) { + VLOG(3) << " numSamples=" << numSamples + << " channels=" << channels + << " imgSizeH=" << imgSizeH + << " imgSizeW=" << imgSizeW << " sizeX=" << sizeX + << " sizeY=" << sizeY << " strideH=" << sH + << " strideW=" << sW << " padingH=" << pH + << " padingW=" << pW; + testMaxPoolFwdBwd(numSamples, + channels, + imgSizeH, + imgSizeW, + sizeX, + sizeY, + sH, + sW, + pH, + pW); + testAvgPoolFwdBwd(numSamples, + channels, + imgSizeH, + imgSizeW, + sizeX, + sizeY, + sH, + sW, + pH, + pW); + } + } + } + } + } + } + } + } + } + } +} + +void testMaxOutFwdBwd( + int numSamples, int imgSizeH, int imgSizeW, int channels, int groups) { + int inWidth = imgSizeH * imgSizeW * channels; + int outChannels = channels / groups; + int outWidth = imgSizeH * imgSizeW * outChannels; + + // forward + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + IVectorPtr id = CpuIVector::create(numSamples * outWidth, false); + IVectorPtr idGpu = GpuIVector::create(numSamples * outWidth, true); + + input->randomizeUniform(); + inputGpu->copyFrom(*input); + + target->maxoutForward(*input, *id, outChannels, groups); + targetGpu->maxoutForward(*inputGpu, *idGpu, outChannels, groups); + + TensorCheckErr(*target, *targetGpu); + TensorCheckEqual(*id, *idGpu); + + // backward + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->maxoutBackward(*targetGrad, *id, outChannels, groups); + inputGpuGrad->maxoutBackward(*targetGpuGrad, *idGpu, outChannels, groups); + + TensorCheckErr(*inputGrad, *inputGpuGrad); +} + +TEST(Matrix, MaxOutFwdBwd) { + for (auto numSamples : {5, 10}) { + for (auto channels : {8, 16}) { + for (auto imgSizeH : {14, 28}) { + for (auto imgSizeW : {16, 30}) { + for (auto groups : {2, 4}) { + VLOG(3) << " numSamples=" << numSamples << " channels=" << channels + << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW + << " groups=" << groups; + testMaxOutFwdBwd(numSamples, imgSizeH, imgSizeW, channels, groups); + } + } + } + } + } +} + +TEST(CpuMatrix, copyFrom) { + const size_t height = 31; + const size_t width = 53; + CpuMatrix cpu(height, width); + GpuMatrix gpu(height, width); + CpuMatrix copy(height, width); + + cpu.randomizeUniform(); + gpu.copyFrom(cpu); + copy.copyFrom(gpu, HPPL_STREAM_DEFAULT); + + TensorCheckEqual(cpu, copy); +} + +void testBatch2seqPadding(int batchSize, int inputDim) { + MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); + MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); + cpuInput->randomizeUniform(); + gpuInput->copyFrom(*cpuInput); + + IVectorPtr cpuSequence; + generateSequenceStartPositions(batchSize, cpuSequence); + for (int i = 0; i < int(cpuSequence->getSize()); ++i) { + (cpuSequence->getData())[i] += 1; // so no way that maxSeqLen is 0; + } + + IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); + gpuSequence->copyFrom(*cpuSequence); + + size_t numSeq = cpuSequence->getSize() - 1; + size_t maxSeqLen = *std::max_element(cpuSequence->getData(), + cpuSequence->getData() + numSeq); + + printf("numSeq = %ld, maxSeqLen = %ld\n", numSeq, maxSeqLen); + MatrixPtr cBatch = std::make_shared(numSeq * maxSeqLen, inputDim); + MatrixPtr gBatch = std::make_shared(numSeq * maxSeqLen, inputDim); + MatrixPtr cCheck = std::make_shared(numSeq * maxSeqLen, inputDim); + + // hl_sequence2batch_copy_padding(gBatch->getData(), + // gpuInput->getData(), + // cpuSequence->getData(), + // inputDim, + // maxSeqLen, + // numSeq, + // false, + // true); + // cCheck->copyFrom(*gBatch); + + // int* seqStart = cpuSequence->getData(); + // float* batchData = cBatch->getData(); + // float* seqData = cpuInput->getData(); + // for (size_t i = 0; i < maxSeqLen; i++) { + // for (size_t j = 0; j < numSeq; j++) { + // size_t sequenceStart = seqStart[j]; + // size_t sequenceLength = seqStart[j + 1] - seqStart[j]; + // if (i < sequenceLength) { + // memcpy(batchData + (i * numSeq + j) * inputDim, + // seqData + (sequenceStart + i) * inputDim, + // inputDim * sizeof(real)); + // } else { + // memset(batchData + (i * numSeq + j) * inputDim, + // 0, + // inputDim * sizeof(real)); + // } + // } + // } + + // TensorCheckErr(*cBatch, *cCheck); +} + +TEST(Matrix, warpCTC) { + for (auto batchSize : {1, 3, 17}) { + for (auto inputDim : {1, 3, 31}) { + VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim; + testBatch2seqPadding(batchSize, inputDim); + } + } +} + +void testMaxPool3DFwdBwd(int numSamples, + int channels, + int imgSizeD, + int imgSizeH, + int imgSizeW, + int ksizeD, + int ksizeH, + int ksizeW, + int strideD, + int strideH, + int strideW, + int padD, + int padH, + int padW) { + int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true); + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); + + int inWidth = channels * imgSizeD * imgSizeH * imgSizeW; + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + int outWidth = channels * outD * outH * outW; + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + MatrixPtr maxIdx = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr maxIdxGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + input->randomizeUniform(); + target->randomizeUniform(); + inputGpu->copyFrom(*input); + targetGpu->copyFrom(*target); + + target->maxPool3DForward(*input, + *maxIdx, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + targetGpu->maxPool3DForward(*inputGpu, + *maxIdxGpu, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); + targetCheck->copyFrom(*targetGpu); + checkMatrixEqual(target, targetCheck); + + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->maxPool3DBackward(*targetGrad, + *maxIdx, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + inputGpuGrad->maxPool3DBackward(*targetGpuGrad, + *maxIdxGpu, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + MatrixPtr targetBwdCheck = + CpuMatrix::create(numSamples, inWidth, false, false); + targetBwdCheck->copyFrom(*inputGpuGrad); + checkMatrixEqual(inputGrad, targetBwdCheck); +} + +void testAvgPool3DFwdBwd(int numSamples, + int channels, + int imgSizeD, + int imgSizeH, + int imgSizeW, + int ksizeD, + int ksizeH, + int ksizeW, + int strideD, + int strideH, + int strideW, + int padD, + int padH, + int padW) { + int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true); + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); + + int inWidth = imgSizeD * imgSizeH * imgSizeW * channels; + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + int outWidth = channels * outD * outH * outW; + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + input->randomizeUniform(); + target->randomizeUniform(); + inputGpu->copyFrom(*input); + targetGpu->copyFrom(*target); + + target->avgPool3DForward(*input, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + + targetGpu->avgPool3DForward(*inputGpu, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + + TensorCheckErr(*target, *targetGpu); + + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->avgPool3DBackward(*targetGrad, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + + inputGpuGrad->avgPool3DBackward(*targetGpuGrad, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + TensorCheckErr(*inputGrad, *inputGpuGrad); +} + +// TODO(yi): I noticed many such blindly combinatorial tests in this +// file. They are no help to locate defects at all. +TEST(Matrix, Pool3DFwdBwd) { + for (auto numSamples : {1, 3}) { + for (auto channels : {3}) { + for (auto imgSizeD : {9, 16}) { + for (auto imgSizeH : {9, 32}) { + for (auto imgSizeW : {9, 32}) { + for (auto sizeX : {3}) { + for (auto sizeY : {3}) { + for (auto sizeZ : {3}) { + for (auto sD : {2}) { + for (auto sH : {2}) { + for (auto sW : {2}) { + for (auto pD : {0, (sizeZ - 1) / 2}) { + for (auto pH : {0, (sizeY - 1) / 2}) { + for (auto pW : {0, (sizeX - 1) / 2}) { + VLOG(3) << " numSamples=" << numSamples + << " channels=" << channels + << " imgSizeD=" << imgSizeD + << " imgSizeH=" << imgSizeH + << " imgSizeW=" << imgSizeW + << " sizeX=" << sizeX + << " sizeY=" << sizeY + << " sizeZ=" << sizeZ << " strideD=" << sD + << " strideH=" << sH << " strideW=" << sW + << " padingD=" << pD << " padingH=" << pH + << " padingW=" << pW; + + testMaxPool3DFwdBwd(numSamples, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + sizeX, + sizeY, + sizeZ, + sD, + sH, + sW, + pD, + pH, + pW); + testAvgPool3DFwdBwd(numSamples, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + sizeX, + sizeY, + sizeZ, + sD, + sH, + sW, + pD, + pH, + pW); + } + } + } + } + } + } + } + } + } + } + } + } + } + } + + // for (auto numSamples : {1, 3}) { + // for (auto channels : {1, 3}) { + // for (auto imgSizeD : {9,16}) { + // for (auto imgSizeH : {9, 32}) { + // for (auto imgSizeW : {9, 32}) { + // for (auto sizeX : {2, 3}) { + // for (auto sizeY : {2, 3}) { + // for (auto sizeZ : {2,3}){ + // for (auto sD : {1, 2}) { + // for (auto sH : {1, 2}) { + // for (auto sW : {1, 2}) { + // for (auto pD : {0, (sizeZ - 1) / 2}){ + // for (auto pH : {0, (sizeY - 1) / 2}) { + // for (auto pW : {0, (sizeX - 1) / 2}) { + // VLOG(3) << " numSamples=" << numSamples + // << " channels=" << channels + // << " imgSizeD=" << imgSizeD + // << " imgSizeH=" << imgSizeH + // << " imgSizeW=" << imgSizeW + // << " sizeX=" << sizeX + // << " sizeY=" << sizeY + // << " sizeZ=" << sizeZ + // << " strideD=" << sD + // << " strideH=" << sH + // << " strideW=" << sW + // << " padingD=" << pD + // << " padingH=" << pH + // << " padingW=" << pW; + // + // testMaxPool3DFwdBwd(numSamples, + // channels, + // imgSizeD, + // imgSizeH, + // imgSizeW, + // sizeX, + // sizeY, + // sizeZ, + // sD, + // sH, + // sW, + // pD, + // pH, + // pW); + // testAvgPool3DFwdBwd(numSamples, + // channels, + // imgSizeD, + // imgSizeH, + // imgSizeW, + // sizeX, + // sizeY, + // sizeZ, + // sD, + // sH, + // sW, + // pD, + // pH, + // pW); + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } +} + +void testMatrixCol2Vol(int depth, int height, int width) { + int channel = 3; + int filterX = 3, filterY = 4, filterZ = 5; + int strideX = 2, strideY = 2, strideZ = 2; + int padX = 1, padY = 1, padZ = 1; + + MatrixPtr cpuImage = + std::make_shared(channel, depth * height * width); + MatrixPtr gpuImage = + std::make_shared(channel, depth * height * width); + cpuImage->randomizeUniform(); + gpuImage->copyFrom(*cpuImage); + + int outD = outputSize(depth, filterZ, padZ, strideZ, true); + int outH = outputSize(height, filterY, padY, strideY, true); + int outW = outputSize(width, filterX, padX, strideX, true); + + int colBufHeight = channel * filterZ * filterY * filterX; + int colBufWidth = outD * outH * outW; + MatrixPtr cpuColBuf = std::make_shared(colBufHeight, colBufWidth); + MatrixPtr gpuColBuf = std::make_shared(colBufHeight, colBufWidth); + cpuColBuf->vol2Col(cpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX); + gpuColBuf->vol2Col(gpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX); + TensorCheckEqual(*cpuColBuf, *gpuColBuf); + + cpuColBuf->randomizeUniform(); + gpuColBuf->copyFrom(*cpuColBuf); + cpuColBuf->col2Vol(cpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX, + 1.0, + 1.0); + gpuColBuf->col2Vol(gpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX, + 1.0, + 1.0); + TensorCheckErr(*cpuImage, *gpuImage); +} + +TEST(Matrix, col2Vol) { + for (auto depth : {9, 16, 64}) { + for (auto height : {9, 11, 128}) { + for (auto width : {9, 32, 128}) { + VLOG(3) << "depth=" << depth << " height=" << height + << " width=" << width; + testMatrixCol2Vol(depth, height, width); + } + } + } +} + +#endif diff --git a/paddle/legacy/math/tests/test_matrixUtil.h b/paddle/legacy/math/tests/test_matrixUtil.h new file mode 100644 index 0000000000000000000000000000000000000000..58c93f746e7ef4e2f2f98d4f410c74909a723812 --- /dev/null +++ b/paddle/legacy/math/tests/test_matrixUtil.h @@ -0,0 +1,233 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/legacy/math/SparseMatrix.h" + +namespace paddle { + +void checkMatrixEqual(const MatrixPtr& a, const MatrixPtr& b) { + ASSERT_EQ(a->getWidth(), b->getWidth()); + ASSERT_EQ(a->getHeight(), b->getHeight()); + ASSERT_EQ(a->isTransposed(), b->isTransposed()); + for (size_t r = 0; r < a->getHeight(); ++r) { + for (size_t c = 0; c < a->getWidth(); ++c) { + ASSERT_FLOAT_EQ(a->getElement(r, c), b->getElement(r, c)); + } + } +} + +void checkSMatrixEqual(const CpuSparseMatrix& a, const CpuSparseMatrix& b) { + ASSERT_EQ(a.getWidth(), b.getWidth()); + ASSERT_EQ(a.getHeight(), b.getHeight()); + ASSERT_EQ(a.isTransposed(), b.isTransposed()); + ASSERT_EQ(a.getFormat(), b.getFormat()); + ASSERT_EQ(a.getElementCnt(), b.getElementCnt()); + for (size_t r = 0; r < a.getElementCnt(); ++r) { + ASSERT_FLOAT_EQ(a.getValue()[r], b.getValue()[r]); + } +} + +void checkSMatrixEqual(const CpuSparseMatrixPtr& a, + const CpuSparseMatrixPtr& b) { + ASSERT_EQ(a->getWidth(), b->getWidth()); + ASSERT_EQ(a->getHeight(), b->getHeight()); + ASSERT_EQ(a->isTransposed(), b->isTransposed()); + ASSERT_EQ(a->getFormat(), b->getFormat()); + ASSERT_EQ(a->getElementCnt(), b->getElementCnt()); + for (size_t r = 0; r < a->getElementCnt(); ++r) { + ASSERT_FLOAT_EQ(a->getValue()[r], b->getValue()[r]); + } +} + +void checkSMatrixEqual2(const CpuSparseMatrixPtr& a, + const CpuSparseMatrixPtr& b) { + ASSERT_EQ(a->getWidth(), b->getWidth()); + ASSERT_EQ(a->getHeight(), b->getHeight()); + ASSERT_EQ(a->isTransposed(), b->isTransposed()); + ASSERT_EQ(a->getFormat(), b->getFormat()); + ASSERT_EQ(a->getValueType(), b->getValueType()); + ASSERT_EQ(a->getElementCnt(), b->getElementCnt()); + if (a->getFormat() == SPARSE_CSR) { + for (size_t r = 0; r < a->getElementCnt(); ++r) { + ASSERT_EQ(a->getCols()[r], b->getCols()[r]); + if (a->getValueType() == FLOAT_VALUE) { + ASSERT_FLOAT_EQ(a->getValue()[r], b->getValue()[r]); + } + } + for (size_t r = 0; r <= a->getHeight(); r++) { + ASSERT_EQ(a->getRows()[r], b->getRows()[r]); + } + } else { + for (size_t r = 0; r < a->getElementCnt(); ++r) { + ASSERT_EQ(a->getRows()[r], b->getRows()[r]); + if (a->getValueType() == FLOAT_VALUE) { + ASSERT_FLOAT_EQ(a->getValue()[r], b->getValue()[r]); + } + } + for (size_t r = 0; r <= a->getWidth(); r++) { + ASSERT_EQ(a->getCols()[r], b->getCols()[r]); + } + } +} + +void checkSMatrixEqual2Dense(const CpuSparseMatrix& a, const CpuMatrix& b) { + ASSERT_EQ(a.getWidth(), b.getWidth()); + ASSERT_EQ(a.getHeight(), b.getHeight()); + ASSERT_EQ(a.isTransposed(), b.isTransposed()); + + if (a.getFormat() == SPARSE_CSC) { + int* rows = a.getRows(); + for (size_t i = 0; i < a.getWidth(); i++) { + for (size_t j = a.getColStartIdx(i); j < a.getColStartIdx(i + 1); j++) { + if (a.getValueType() == FLOAT_VALUE) { + ASSERT_FLOAT_EQ(a.getValue()[j], b.getElement(rows[j], i)); + } else { + ASSERT_FLOAT_EQ(1.0, b.getElement(rows[j], i)); + } + } + } + } else { + int* cols = a.getCols(); + for (size_t i = 0; i < a.getHeight(); i++) { + for (size_t j = a.getRowStartIdx(i); j < a.getRowStartIdx(i + 1); j++) { + if (a.getValueType() == FLOAT_VALUE) { + ASSERT_FLOAT_EQ(a.getValue()[j], b.getElement(i, cols[j])); + } else { + ASSERT_FLOAT_EQ(1.0, b.getElement(i, cols[j])); + } + } + } + } +} + +void checkSMatrixEqual2Dense(const CpuSparseMatrixPtr& a, + const CpuMatrixPtr& b) { + ASSERT_EQ(a->getWidth(), b->getWidth()); + ASSERT_EQ(a->getHeight(), b->getHeight()); + ASSERT_EQ(a->isTransposed(), b->isTransposed()); + + if (a->getFormat() == SPARSE_CSC) { + int* rows = a->getRows(); + for (size_t i = 0; i < a->getWidth(); i++) { + for (size_t j = a->getColStartIdx(i); j < a->getColStartIdx(i + 1); j++) { + if (a->getValueType() == FLOAT_VALUE) { + ASSERT_FLOAT_EQ(a->getValue()[j], b->getElement(rows[j], i)); + } else { + ASSERT_FLOAT_EQ(1.0, b->getElement(rows[j], i)); + } + } + } + } else { + int* cols = a->getCols(); + for (size_t i = 0; i < a->getHeight(); i++) { + for (size_t j = a->getRowStartIdx(i); j < a->getRowStartIdx(i + 1); j++) { + if (a->getValueType() == FLOAT_VALUE) { + ASSERT_FLOAT_EQ(a->getValue()[j], b->getElement(i, cols[j])); + } else { + ASSERT_FLOAT_EQ(1.0, b->getElement(i, cols[j])); + } + } + } + } +} + +void checkSMatrixErr(const CpuSparseMatrixPtr& a, const CpuSparseMatrixPtr& b) { +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + ASSERT_EQ(a->getWidth(), b->getWidth()); + ASSERT_EQ(a->getHeight(), b->getHeight()); + ASSERT_EQ(a->isTransposed(), b->isTransposed()); + ASSERT_EQ(a->getFormat(), b->getFormat()); + ASSERT_EQ(a->getValueType(), b->getValueType()); + ASSERT_EQ(a->getElementCnt(), b->getElementCnt()); + int count = 0; + if (a->getFormat() == SPARSE_CSR) { + for (size_t r = 0; r < a->getElementCnt(); ++r) { + ASSERT_EQ(a->getCols()[r], b->getCols()[r]); + if (a->getValueType() == FLOAT_VALUE) { + real aVal = a->getValue()[r]; + real bVal = b->getValue()[r]; + if (std::abs(aVal - bVal) > err) { + if ((std::abs(aVal - bVal) / std::abs(aVal)) > (err / 10.0f)) { + LOG(INFO) << "a=" << aVal << "\t" + << "b=" << bVal; + count++; + } + } + } + } + for (size_t r = 0; r <= a->getHeight(); r++) { + ASSERT_EQ(a->getRows()[r], b->getRows()[r]); + } + } else { + for (size_t r = 0; r < a->getElementCnt(); ++r) { + ASSERT_EQ(a->getRows()[r], b->getRows()[r]); + if (a->getValueType() == FLOAT_VALUE) { + real aVal = a->getValue()[r]; + real bVal = b->getValue()[r]; + if (std::abs(aVal - bVal) > err) { + if ((std::abs(aVal - bVal) / std::abs(aVal)) > (err / 10.0f)) { + count++; + } + } + } + } + for (size_t r = 0; r <= a->getWidth(); r++) { + ASSERT_EQ(a->getCols()[r], b->getCols()[r]); + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; +} + +void checkMatrixErr(const Matrix& matrix1, const Matrix& matrix2) { + CHECK(matrix1.getHeight() == matrix2.getHeight()); + CHECK(matrix1.getWidth() == matrix2.getWidth()); +#ifndef PADDLE_TYPE_DOUBLE + real err = 1e-3; +#else + real err = 1e-10; +#endif + + int height = matrix1.getHeight(); + int width = matrix1.getWidth(); + const real* data1 = matrix1.getData(); + const real* data2 = matrix2.getData(); + int count = 0; + for (int i = 0; i < height; i++) { + for (int j = 0; j < width; j++) { + real a = data1[i * width + j]; + real b = data2[i * width + j]; + if (std::abs(a - b) > err) { + if ((std::abs(a - b) / std::abs(a)) > (err / 10.0f)) { + count++; + } + } + } + } + EXPECT_EQ(count, 0) << "There are " << count << " different element."; +} + +void checkDataEqual(const real* a, const real* b, size_t size) { + for (size_t i = 0; i < size; ++i) { + ASSERT_FLOAT_EQ(a[i], b[i]); + } +} + +} // namespace paddle diff --git a/paddle/math/tests/test_perturbation.cpp b/paddle/legacy/math/tests/test_perturbation.cpp similarity index 100% rename from paddle/math/tests/test_perturbation.cpp rename to paddle/legacy/math/tests/test_perturbation.cpp diff --git a/paddle/legacy/math/tests/test_sparseMatrixCompare.cpp b/paddle/legacy/math/tests/test_sparseMatrixCompare.cpp new file mode 100644 index 0000000000000000000000000000000000000000..492aa0a689540dbb2c687326ff8a2919d89d2e6f --- /dev/null +++ b/paddle/legacy/math/tests/test_sparseMatrixCompare.cpp @@ -0,0 +1,174 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +/// This unittest checks GpuSparseMatrix/CpuSparseMatrix get same result, +// so disable when +/// only cpu version. + +#include +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/utils/Util.h" +#include "test_matrixUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static inline int uniformRandom(int n) { return n == 0 ? 0 : rand() % n; } + +void testSpMatrixAddBias(int M, int N, real rate, real scale) { + int nnz = M * N * rate; + + MatrixPtr cpuA(new CpuSparseMatrix(M, N, nnz)); + MatrixPtr cpuB = std::make_shared(1, N); + + MatrixPtr gpuA(new GpuSparseMatrix(M, N, nnz)); + MatrixPtr gpuB = std::make_shared(1, N); + + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + + hl_stream_t stream(HPPL_STREAM_1); + gpuA->copyFrom(*cpuA, stream); + gpuB->copyFrom(*cpuB, stream); + hl_stream_synchronize(stream); + + cpuA->addBias(*cpuB, scale); + gpuA->addBias(*gpuB, scale); + + MatrixPtr outputCheck(new CpuSparseMatrix(M, N, nnz)); + outputCheck->copyFrom(*gpuA, stream); + hl_stream_synchronize(stream); + checkSMatrixEqual2(std::dynamic_pointer_cast(cpuA), + std::dynamic_pointer_cast(outputCheck)); +} + +void testSpMatrixAddDense(int M, int N, real rate) { // add3 + int nnz = M * N * rate; + + MatrixPtr cpuA(new CpuSparseMatrix(M, N, nnz)); + MatrixPtr cpuB = std::make_shared(M, N); + + MatrixPtr gpuA(new GpuSparseMatrix(M, N, nnz)); + MatrixPtr gpuB = std::make_shared(M, N); + + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + + hl_stream_t stream(HPPL_STREAM_3); + gpuA->copyFrom(*cpuA, stream); + gpuB->copyFrom(*cpuB, stream); + hl_stream_synchronize(stream); + + cpuA->add3(cpuB); + gpuA->add3(gpuB); + + MatrixPtr outputCheck(new CpuSparseMatrix(M, N, nnz)); + outputCheck->copyFrom(*gpuA, stream); + hl_stream_synchronize(stream); + checkSMatrixEqual2(std::dynamic_pointer_cast(cpuA), + std::dynamic_pointer_cast(outputCheck)); +} + +void testSpMatrixMul(int M, int N, int K, real rate) { + int nnz = M * N * rate; + + MatrixPtr cpuA = std::make_shared(M, K); + MatrixPtr cpuB = std::make_shared(N, K); + MatrixPtr cpuC(new CpuSparseMatrix(M, N, nnz)); + + MatrixPtr gpuA = std::make_shared(M, K); + MatrixPtr gpuB = std::make_shared(N, K); + MatrixPtr gpuC(new GpuSparseMatrix(M, N, nnz)); + + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + cpuC->randomizeUniform(); + + hl_stream_t stream(HPPL_STREAM_3); + gpuA->copyFrom(*cpuA, stream); + gpuB->copyFrom(*cpuB, stream); + gpuC->copyFrom(*cpuC, stream); + hl_stream_synchronize(stream); + + cpuC->mul(*cpuA, *cpuB->getTranspose(), 1, 1); + gpuC->mul(*gpuA, *gpuB->getTranspose(), 1, 1); + + MatrixPtr outputCheck(new CpuSparseMatrix(M, N, nnz)); + outputCheck->copyFrom(*gpuC, stream); + hl_stream_synchronize(stream); + checkSMatrixErr(std::dynamic_pointer_cast(cpuC), + std::dynamic_pointer_cast(outputCheck)); +} + +void testSpMatrixCollectBias(int M, int N, real rate) { + int nnz = M * N * rate; + LOG(INFO) << "nnz=" << nnz; + + MatrixPtr cpuA(new CpuSparseMatrix(M, N, nnz)); + MatrixPtr cpuB = std::make_shared(1, N); + + MatrixPtr gpuA(new GpuSparseMatrix(M, N, nnz)); + MatrixPtr gpuB = std::make_shared(1, N); + + cpuA->randomizeUniform(); + cpuB->randomizeUniform(); + + hl_stream_t stream(HPPL_STREAM_3); + gpuA->copyFrom(*cpuA, stream); + gpuB->copyFrom(*cpuB, stream); + hl_stream_synchronize(stream); + + cpuB->collectBias(*cpuA, 1); + gpuB->collectBias(*gpuA, 1); + + MatrixPtr outputCheck = std::make_shared(1, N); + outputCheck->copyFrom(*gpuB, stream); + hl_stream_synchronize(stream); + checkMatrixErr(*cpuB, *outputCheck); +} + +TEST(SMatrix, sMatrixOp) { + for (auto height : {1, 11, 200}) { + for (auto width : {200, 2048, 20480}) { + VLOG(3) << " height=" << height << " width=" << width; + for (auto rate : {0.02, 0.1}) { + testSpMatrixAddDense(height, width, rate); + testSpMatrixAddBias(height, width, rate, 1.0); + } + } + } +} + +TEST(SMatrix, sMatrixMul) { + for (auto M : {1, 40, 128, 200}) { + for (auto N : {100, 2000, 20480}) { + for (auto K : {100, 512, 1024}) { + VLOG(3) << " M=" << M << " N=" << N << " K=" << K; + testSpMatrixMul(M, N, K, 0.05); + } + } + } +} + +TEST(SMatrix, sMatrixCollectBias) { + for (auto height : {1, 128, 200}) { + for (auto width : {100, 2048, 20480}) { + VLOG(3) << " height=" << height << " width=" << width; + testSpMatrixCollectBias(height, width, 0.1); + } + } +} + +#endif diff --git a/paddle/optimizer/CMakeLists.txt b/paddle/legacy/optimizer/CMakeLists.txt similarity index 100% rename from paddle/optimizer/CMakeLists.txt rename to paddle/legacy/optimizer/CMakeLists.txt diff --git a/paddle/optimizer/adadelta_optimizer.cc b/paddle/legacy/optimizer/adadelta_optimizer.cc similarity index 100% rename from paddle/optimizer/adadelta_optimizer.cc rename to paddle/legacy/optimizer/adadelta_optimizer.cc diff --git a/paddle/optimizer/adadelta_optimizer.h b/paddle/legacy/optimizer/adadelta_optimizer.h similarity index 100% rename from paddle/optimizer/adadelta_optimizer.h rename to paddle/legacy/optimizer/adadelta_optimizer.h diff --git a/paddle/optimizer/adagrad_optimizer.cc b/paddle/legacy/optimizer/adagrad_optimizer.cc similarity index 100% rename from paddle/optimizer/adagrad_optimizer.cc rename to paddle/legacy/optimizer/adagrad_optimizer.cc diff --git a/paddle/optimizer/adagrad_optimizer.h b/paddle/legacy/optimizer/adagrad_optimizer.h similarity index 100% rename from paddle/optimizer/adagrad_optimizer.h rename to paddle/legacy/optimizer/adagrad_optimizer.h diff --git a/paddle/optimizer/adam_optimizer.cc b/paddle/legacy/optimizer/adam_optimizer.cc similarity index 100% rename from paddle/optimizer/adam_optimizer.cc rename to paddle/legacy/optimizer/adam_optimizer.cc diff --git a/paddle/optimizer/adam_optimizer.h b/paddle/legacy/optimizer/adam_optimizer.h similarity index 100% rename from paddle/optimizer/adam_optimizer.h rename to paddle/legacy/optimizer/adam_optimizer.h diff --git a/paddle/optimizer/lr_policy.h b/paddle/legacy/optimizer/lr_policy.h similarity index 100% rename from paddle/optimizer/lr_policy.h rename to paddle/legacy/optimizer/lr_policy.h diff --git a/paddle/optimizer/optimizer.cc b/paddle/legacy/optimizer/optimizer.cc similarity index 100% rename from paddle/optimizer/optimizer.cc rename to paddle/legacy/optimizer/optimizer.cc diff --git a/paddle/optimizer/optimizer.h b/paddle/legacy/optimizer/optimizer.h similarity index 100% rename from paddle/optimizer/optimizer.h rename to paddle/legacy/optimizer/optimizer.h diff --git a/paddle/optimizer/parameter_optimizer.cc b/paddle/legacy/optimizer/parameter_optimizer.cc similarity index 100% rename from paddle/optimizer/parameter_optimizer.cc rename to paddle/legacy/optimizer/parameter_optimizer.cc diff --git a/paddle/optimizer/parameter_optimizer.h b/paddle/legacy/optimizer/parameter_optimizer.h similarity index 100% rename from paddle/optimizer/parameter_optimizer.h rename to paddle/legacy/optimizer/parameter_optimizer.h diff --git a/paddle/optimizer/parameter_optimizer_test.cc b/paddle/legacy/optimizer/parameter_optimizer_test.cc similarity index 100% rename from paddle/optimizer/parameter_optimizer_test.cc rename to paddle/legacy/optimizer/parameter_optimizer_test.cc diff --git a/paddle/legacy/optimizer/serialization.h b/paddle/legacy/optimizer/serialization.h new file mode 100644 index 0000000000000000000000000000000000000000..2067a8d8cff23bff975d23a4df4d0aa7df20b00f --- /dev/null +++ b/paddle/legacy/optimizer/serialization.h @@ -0,0 +1,49 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include +#include "OptimizerConfig.pb.h" +#include "paddle/legacy/utils/Logging.h" +#include "tensor.h" + +namespace paddle { +namespace optimizer { + +static void TensorToProto(const Tensor& tensor, TensorProto* proto) { + proto->set_data_type(TensorProto::PADDLE_ELEMENT_TYPE_FLOAT32); + std::stringstream os; + for (size_t i = 0; i < tensor.size(); ++i) { + os << tensor[i]; + proto->add_content(os.str()); + os.str(std::string()); + } +} + +static void ProtoToTensor(const TensorProto& proto, Tensor* tensor) { + std::stringstream sin; + for (auto i = 0; i < proto.content_size(); ++i) { + sin << proto.content(i); + sin >> (*tensor)[i]; + sin.str(std::string()); + sin.clear(); + } +} + +} // namespace optimizer +} // namespace paddle diff --git a/paddle/optimizer/serialization_test.cc b/paddle/legacy/optimizer/serialization_test.cc similarity index 100% rename from paddle/optimizer/serialization_test.cc rename to paddle/legacy/optimizer/serialization_test.cc diff --git a/paddle/optimizer/sgd_optimizer.cc b/paddle/legacy/optimizer/sgd_optimizer.cc similarity index 100% rename from paddle/optimizer/sgd_optimizer.cc rename to paddle/legacy/optimizer/sgd_optimizer.cc diff --git a/paddle/optimizer/sgd_optimizer.h b/paddle/legacy/optimizer/sgd_optimizer.h similarity index 100% rename from paddle/optimizer/sgd_optimizer.h rename to paddle/legacy/optimizer/sgd_optimizer.h diff --git a/paddle/legacy/optimizer/tensor.h b/paddle/legacy/optimizer/tensor.h new file mode 100644 index 0000000000000000000000000000000000000000..2e58577d4df7aabd8cd218dc13837461cc681ac6 --- /dev/null +++ b/paddle/legacy/optimizer/tensor.h @@ -0,0 +1,68 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once +/** + * @brief tensor used by optimizer + */ + +#include +#include +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { +namespace optimizer { + +template +class TensorT { + public: + TensorT(size_t size) : height_(1), width_(size) { + // new T[size]() initializes all element to zero value. + data_ptr_ = std::shared_ptr(new T[size](), std::default_delete()); + data_ = data_ptr_.get(); + } + + TensorT(T* data, size_t size) + : height_(1), width_(size), data_ptr_(nullptr), data_(data) {} + + TensorT(T* data, size_t h, size_t w) + : height_(h), width_(w), data_ptr_(nullptr), data_(data) {} + + virtual ~TensorT() {} + + T* get_buffer() { return this->data_; } + + T& operator[](const size_t idx) { + CHECK(idx >= 0 && idx < this->width_) << "out of index range"; + return data_[idx]; + } + T& operator[](const size_t idx) const { + CHECK(idx >= 0 && idx < this->width_) << "out of index range"; + return data_[idx]; + } + // TODO: replace with tensorshape + size_t size() const { return this->width_ * this->height_; } + + protected: + size_t height_; + size_t width_; + std::shared_ptr data_ptr_; + T* data_; +}; + +// TODO(zhihong): design problem of dynamic datatype, need to fix it +typedef TensorT Tensor; + +} // namespace optimizer +} // namespace paddle diff --git a/paddle/legacy/parameter/Argument.cpp b/paddle/legacy/parameter/Argument.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3f1d599e901110a1c9390d76c45f8b4b1f4cab2a --- /dev/null +++ b/paddle/legacy/parameter/Argument.cpp @@ -0,0 +1,707 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Argument.h" +#include "paddle/legacy/math/SparseMatrix.h" + +#include + +namespace paddle { +static void resizeAndCopy(MatrixPtr& dest, + const MatrixPtr& src, + bool useGpu, + hl_stream_t stream) { + if (src) { + if (!dest) { + dest = src->clone(0, 0, useGpu); + } else { + CHECK_EQ(dest->useGpu(), useGpu); + dest->resize(src->getHeight(), src->getWidth()); + } + dest->copyFrom(*src, stream); + } else { + dest.reset(); + } +} + +static void resizeAndCopy(IVectorPtr& dest, + const IVectorPtr& src, + bool useGpu, + hl_stream_t stream) { + if (src) { + IVector::resizeOrCreate(dest, src->getSize(), useGpu); + dest->copyFrom(*src, stream); + } else { + dest.reset(); + } +} + +static void resizeAndCopy(ICpuGpuVectorPtr& dest, + const ICpuGpuVectorPtr& src, + bool useGpu, + hl_stream_t stream) { + if (src) { + ICpuGpuVector::resizeOrCreate(dest, src->getSize(), useGpu); + dest->copyFrom(*src, stream); + } else { + dest.reset(); + } +} + +static void resizeAndCopy(MatrixPtr& dest, + const MatrixPtr& src, + int32_t startRow, + int32_t copySize, + bool useGpu, + hl_stream_t stream = HPPL_STREAM_DEFAULT) { + if (src) { + CHECK_LE((size_t)startRow + copySize, src->getHeight()); + int height = copySize; + int width = src->getWidth(); + if (!dest) { + dest = src->clone(height, width, useGpu); + } else { + CHECK_EQ(dest->useGpu(), useGpu); + dest->resize(height, width); + } + MatrixPtr submat = src->subMatrix(startRow, copySize); + if (dynamic_cast(dest.get())) { + // copy a subMatrix of CpuSparseMatrix to GpuSparseMatrix. + // First copy it to CPU, and then copy it to the GPU. + MatrixPtr tmp = src->clone(height, width, false); + tmp->copyFrom(*submat, stream); + dest->copyFrom(*tmp, stream); + } else { + dest->copyFrom(*submat, stream); + } + } else { + dest.reset(); + } +} + +static void resizeAndCopy(IVectorPtr& dest, + const IVectorPtr& src, + int32_t startPos, + int32_t copySize, + bool useGpu, + hl_stream_t stream = HPPL_STREAM_DEFAULT) { + if (src) { + CHECK_LE((size_t)startPos + copySize, src->getSize()); + + int height = copySize; + IVector::resizeOrCreate(dest, height, useGpu); + dest->copyFrom(src->getData() + startPos, height, stream); + } else { + dest.reset(); + } +} + +static void resizeAndCopy(ICpuGpuVectorPtr& dest, + const ICpuGpuVectorPtr& src, + int32_t startPos, + int32_t copySize, + bool useGpu, + hl_stream_t stream = HPPL_STREAM_DEFAULT) { + if (src) { + CHECK_LE((size_t)startPos + copySize, src->getSize()); + + ICpuGpuVector::resizeOrCreate(dest, copySize, useGpu); + dest->copyFrom(*src, startPos, copySize, useGpu, stream); + } else { + dest.reset(); + } +} + +static void resizeAndCopy(SVectorPtr& dest, + const SVectorPtr& src, + bool useGpu, + hl_stream_t stream) { + if (src) { + size_t height = src->size(); + if (!dest) { + dest = std::make_shared>(height); + } else { + dest->resize(height); + } + std::copy_n(src->begin(), height, dest->begin()); + } else { + dest.reset(); + } +} + +static void resizeAndCopy(SVectorPtr& dest, + const SVectorPtr& src, + int32_t startPos, + int32_t copySize, + bool useGpu, + hl_stream_t stream = HPPL_STREAM_DEFAULT) { + if (src) { + CHECK_LE((size_t)startPos + copySize, src->size()); + size_t height = copySize; + if (!dest) { + dest = std::make_shared>(height); + } else { + dest->resize(height); + } + std::copy_n(src->begin() + startPos, height, dest->begin()); + } else { + dest.reset(); + } +} + +void Argument::resizeAndCopyFrom(const Argument& src, bool useGpu) { + resizeAndCopyFrom(src, useGpu, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); +} + +void Argument::resizeAndCopyFrom(const Argument& src, + bool useGpu, + hl_stream_t stream) { + dataId = src.dataId; + resizeAndCopy(value, src.value, useGpu, stream); + resizeAndCopy(grad, src.grad, useGpu, stream); + resizeAndCopy(in, src.in, useGpu, stream); + resizeAndCopy(ids, src.ids, useGpu, stream); + resizeAndCopy(sequenceStartPositions, + src.sequenceStartPositions, + false /* useGpu */, + stream); + if (src.hasSubseq()) { + resizeAndCopy(subSequenceStartPositions, + src.subSequenceStartPositions, + false /* useGpu */, + stream); + } + resizeAndCopy(strs, src.strs, useGpu, stream); + frameWidth = src.frameWidth; + frameHeight = src.frameHeight; + frameDepth = src.frameDepth; +} + +int32_t Argument::resizeAndCopyFrom(const Argument& src, + int32_t startSeq, + int32_t copySize, + bool useGpu) { + int32_t size = + resizeAndCopyFrom(src, startSeq, copySize, useGpu, HPPL_STREAM_DEFAULT); + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + return size; +} + +int32_t Argument::resizeAndCopyFrom(const Argument& src, + int32_t startSeq, + int32_t copySize, + bool useGpu, + hl_stream_t stream) { + dataId = src.dataId; + frameWidth = src.frameWidth; + frameHeight = src.frameHeight; + frameDepth = src.frameDepth; + + if (!src.sequenceStartPositions) { + // non-sequence input, copy samples directly + int32_t startRow = startSeq; + resizeAndCopy(in, src.in, startRow, copySize, useGpu, stream); + resizeAndCopy(value, src.value, startRow, copySize, useGpu, stream); + resizeAndCopy(grad, src.grad, startRow, copySize, useGpu, stream); + resizeAndCopy(ids, src.ids, startRow, copySize, useGpu, stream); + resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream); + return copySize; + } else { + // sequence input + const int* sequence = src.sequenceStartPositions->getData(false); + int32_t startRow = sequence[startSeq]; // sample start from here + int32_t endRow = sequence[startSeq + copySize]; // sample end + int32_t copyFeatureSize = endRow - startRow; // num of samples + resizeAndCopy(in, src.in, startRow, copyFeatureSize, useGpu, stream); + resizeAndCopy(value, src.value, startRow, copyFeatureSize, useGpu, stream); + resizeAndCopy(grad, src.grad, startRow, copyFeatureSize, useGpu, stream); + resizeAndCopy(ids, src.ids, startRow, copyFeatureSize, useGpu, stream); + resizeAndCopy(sequenceStartPositions, + src.sequenceStartPositions, + startSeq, + copySize + 1, + false, + stream); + // modify new sequenceStartPositions + int* destSequences = sequenceStartPositions->getMutableData(false); + for (int i = 0; i < copySize + 1; i++) { + destSequences[i] -= startRow; + } + CHECK_EQ(destSequences[0], 0); + CHECK_EQ(destSequences[copySize], copyFeatureSize); + if (src.hasSubseq()) { + // sequence has sub-sequence + int* subSequence = src.subSequenceStartPositions->getMutableData(false); + int32_t subStartSeq = 0; + int32_t subEndSeq = 0; + int numSubSequences = src.getNumSubSequences(); + for (int i = 0; i < numSubSequences + 1; i++) { + if (subSequence[i] == startRow) { + subStartSeq = i; + } else if (subSequence[i] == endRow) { + subEndSeq = i; + break; + } + } + int32_t copySubSize = subEndSeq - subStartSeq; + resizeAndCopy(subSequenceStartPositions, + src.subSequenceStartPositions, + subStartSeq, + copySubSize + 1, + false, + stream); + // modify new subSequenceStartPositions + int* destSubSequences = subSequenceStartPositions->getMutableData(false); + for (int i = 0; i < copySubSize + 1; i++) { + destSubSequences[i] -= startRow; + } + CHECK_EQ(destSubSequences[0], 0); + CHECK_EQ(destSubSequences[copySubSize], copyFeatureSize); + } + resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream); + return copyFeatureSize; + } +} + +void Argument::concat(const std::vector& args, + const std::vector& selectRows, + const std::vector& seqStartPos, + const std::vector& copySize, + bool useGpu, + hl_stream_t stream, + PassType passType) { + CHECK(!subSequenceStartPositions) + << "undefined behavior for subsequence positions"; + + size_t batchSize = 0; + for (size_t i = 0; i < copySize.size(); ++i) + batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]); + + auto copyArg = [batchSize, stream](MatrixPtr& dst, + MatrixPtr src, + int desStartRow, + int srcStartRow, + int size, + bool useGpu) { + if (!src) { + dst.reset(); + return; + } + size_t width = src->getWidth(); + if (!dst) { + dst = src->clone(batchSize, width, useGpu); + } else { + dst->resize(batchSize, width); + } + + MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size); + tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream); + }; + + auto copyIds = [batchSize, stream](IVectorPtr& dst, + const IVectorPtr& src, + int desStartRow, + int srcStartRow, + int size, + bool useGpu) { + if (!src) { + dst.reset(); + return; + } + IVector::resizeOrCreate(dst, batchSize, useGpu); + dst->subVec(desStartRow, size) + ->copyFrom(*src->subVec(srcStartRow, size), stream); + }; + + auto copyStrs = [batchSize](SVectorPtr& dst, + const SVectorPtr& src, + int desStartRow, + int srcStartRow, + int size, + bool useGpu) { + if (!src) { + dst.reset(); + return; + } + if (!dst) { + dst = std::make_shared>(batchSize); + } else { + dst->resize(batchSize); + } + std::copy(src->begin() + srcStartRow, + src->begin() + srcStartRow + size, + dst->begin() + desStartRow); + }; + + dataId = args[0].dataId; + CHECK_NE(seqStartPos.size(), 0UL); + int desStartRow = 0; + for (size_t i = 0; i < copySize.size(); ++i) { + int startPos = seqStartPos[i]; + int endPos = seqStartPos[i + 1]; + CHECK_GE(args.size(), static_cast(endPos - startPos)); + for (int j = startPos; j < endPos; ++j) { + const Argument& arg = args[j - startPos]; + CHECK_EQ(arg.dataId, dataId) << "Arguments to concatenate should have " + << "the same dataId."; + const int srcStartRow = selectRows[j]; + copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu); + copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu); + if (passType != PASS_TEST) { + copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu); + } + copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu); + copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu); + desStartRow += copySize[i]; + } + } + ICpuGpuVector::resizeOrCreate( + sequenceStartPositions, seqStartPos.size(), useGpu); + sequenceStartPositions->copyFrom( + seqStartPos.data(), seqStartPos.size(), useGpu); +} + +void Argument::concat(const std::vector& args, + bool useGpu, + hl_stream_t stream, + PassType passType) { + int32_t batchSize = 0; + int64_t numSequences = 0; + int64_t numSubSequences = 0; + for (auto& arg : args) { + batchSize += arg.getBatchSize(); + numSequences += arg.getNumSequences(); + numSubSequences += arg.getNumSubSequences(); + } + + auto copyArg = [batchSize, stream]( + MatrixPtr& dst, MatrixPtr src, int startRow, bool useGpu) { + if (!src) { + dst.reset(); + return; + } + size_t width = src->getWidth(); + if (!dst) { + dst = src->clone(batchSize, width, useGpu); + } else { + dst->resize(batchSize, width); + } + + MatrixPtr tmpMatrix = dst->subMatrix(startRow, src->getHeight()); + tmpMatrix->copyFrom(*src, stream); + }; + + auto copyIds = [batchSize, stream]( + IVectorPtr& dst, const IVectorPtr& src, int startRow, bool useGpu) { + if (!src) { + dst.reset(); + return; + } + IVector::resizeOrCreate(dst, batchSize, useGpu); + dst->subVec(startRow, src->getSize())->copyFrom(*src, stream); + }; + + auto copyStrs = [batchSize]( + SVectorPtr& dst, const SVectorPtr& src, int startRow, bool useGpu) { + if (!src) { + dst.reset(); + return; + } + if (!dst) { + dst = std::make_shared>(batchSize); + } else { + dst->resize(batchSize); + } + std::copy(src->begin(), src->end(), dst->begin() + startRow); + }; + + auto copySequencePos = [](ICpuGpuVectorPtr& dstSeq, + const ICpuGpuVectorPtr& srcSeq, + int dstNumSequences, + int srcNumSequences, + int& startSequences, + int startRow) { + if (srcSeq) { + ICpuGpuVector::resizeOrCreate(dstSeq, dstNumSequences + 1, false); + const int* src = srcSeq->getData(false); + int* dest = dstSeq->getMutableData(false); + for (int i = 0; i < srcNumSequences + 1; ++i) { + dest[i + startSequences] = src[i] + startRow; + } + startSequences += srcNumSequences; + } else { + dstSeq.reset(); + } + }; + + int startRow = 0; + int startSequences = 0; + int startSubSequences = 0; + dataId = args[0].dataId; + for (auto& arg : args) { + CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have" + << " same dataId"; + copyArg(in, arg.in, startRow, useGpu); + copyArg(value, arg.value, startRow, useGpu); + if (passType != PASS_TEST) copyArg(grad, arg.grad, startRow, useGpu); + copyIds(ids, arg.ids, startRow, useGpu); + copySequencePos(sequenceStartPositions, + arg.sequenceStartPositions, + numSequences, + arg.getNumSequences(), + startSequences, + startRow); + copySequencePos(subSequenceStartPositions, + arg.subSequenceStartPositions, + numSubSequences, + arg.getNumSubSequences(), + startSubSequences, + startRow); + copyStrs(strs, arg.strs, startRow, useGpu); + startRow += arg.getBatchSize(); + } +} + +void Argument::splitByDataId(const std::vector& argus, + std::vector>* arguGroups) { + arguGroups->clear(); + int lastDataId = -1; + for (const auto& argu : argus) { + if (argu.dataId == -1) { + // is -1, then create a new group + arguGroups->emplace_back(); + lastDataId = -1; + } else if (argu.dataId != lastDataId) { + // not -1, also not equal to last Argument, then create a new group + arguGroups->emplace_back(); + lastDataId = argu.dataId; + } else { + // not -1, and equal to last Argument, do nothing + } + arguGroups->back().push_back(argu); + } +} + +void Argument::getSeqInfo(std::vector* seqInfo) const { + const int* starts = sequenceStartPositions->getData(false); + const int* subStarts = + hasSubseq() ? subSequenceStartPositions->getData(false) : nullptr; + size_t numSequences = getNumSequences(); + seqInfo->reserve(numSequences); + int subSeqEnd = 0; + for (size_t i = 0; i < numSequences; ++i) { + SeqInfo info; + info.seqStart = starts[i]; + info.subLevelLength = starts[i + 1] - starts[i]; + info.seqId = i; + if (hasSubseq()) { + info.subSeqStart = subSeqEnd; + while (subStarts[subSeqEnd] < starts[i + 1]) { + ++subSeqEnd; + } + info.topLevelLength = subSeqEnd - info.subSeqStart; + } else { + info.topLevelLength = info.subLevelLength; + info.subSeqStart = 0; // not used + } + seqInfo->push_back(info); + } + std::sort( + seqInfo->begin(), seqInfo->end(), [](const SeqInfo& a, const SeqInfo& b) { + return a.topLevelLength > b.topLevelLength; + }); +} + +void Argument::checkSubset() const { + if (getNumSequences() > getNumSubSequences()) { + LOG(FATAL) << "numSubSequences is less than numSequences (" + << getNumSubSequences() << " vs. " << getNumSequences() << ")"; + } + const int* start = sequenceStartPositions->getData(false); + const int* subStart = subSequenceStartPositions->getData(false); + int seqId = 0; + int subSeqId = 0; + while (seqId < getNumSequences() && subSeqId < getNumSubSequences()) { + if (start[seqId] > subStart[subSeqId]) { + ++subSeqId; + } else if (start[seqId] == subStart[subSeqId]) { + ++subSeqId; + ++seqId; + } else { + LOG(FATAL) << "seqStartPositions is not subset of subSeqStartPositions"; + } + } + if (seqId < getNumSequences()) { + LOG(FATAL) << "seqStartPositions is not subset of subSeqStartPositions"; + } +} + +void Argument::degradeSequence(const Argument& input) { + CHECK_EQ(input.hasSubseq(), 1UL); + size_t numSequences = input.getNumSequences(); + size_t numSubSequences = input.getNumSubSequences(); + ICpuGpuVector::resizeOrCreate( + sequenceStartPositions, numSequences + 1, false); + int* tgtBuf = sequenceStartPositions->getMutableData(false); + const int* starts = input.sequenceStartPositions->getData(false); + const int* subStarts = input.subSequenceStartPositions->getData(false); + int seqId = 0; + for (size_t subSeqId = 0; subSeqId < numSubSequences; ++subSeqId) { + if (subStarts[subSeqId] == starts[seqId]) { + tgtBuf[seqId] = subSeqId; + seqId++; + } + } + tgtBuf[numSequences] = numSubSequences; +} + +void Argument::poolSequenceWithStride(const Argument& input, + size_t stride, + ICpuGpuVectorPtr* stridePostions, + bool reversed) { + // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5, + // then sequenceStartPositions = [0, 2, 3, 4, 7]. + // If reversed = false, stridePostions = [0, 5, 9, 14, 17, 22, 27, 30]; + // else reversed = true, stridePostions = [0, 4, 9, 14, 17, 20, 25, 30] + + CHECK(input.sequenceStartPositions); + CHECK_EQ(input.hasSubseq(), 0UL); + CHECK_GT(stride, 0UL) << "stride must larger than 0"; + size_t numSequences = input.getNumSequences(); + ICpuGpuVector::resizeOrCreate( + sequenceStartPositions, numSequences + 1, false); + const int* starts = input.sequenceStartPositions->getData(false); + int* tgtBuf = sequenceStartPositions->getMutableData(false); + // first index of target sequence and stride positions are both 0 + tgtBuf[0] = 0; + std::vector stridePos; + for (size_t seqId = 0; seqId < numSequences; ++seqId) { + size_t seqLength = starts[seqId + 1] - starts[seqId]; + stridePos.emplace_back(starts[seqId]); + if (seqLength == 0) { + // empty sequence + tgtBuf[seqId + 1] = tgtBuf[seqId]; + } else { + int size = ceil((float)seqLength / stride); + tgtBuf[seqId + 1] = tgtBuf[seqId] + size; + for (int i = 0; i < size - 1; ++i) { + int cur = reversed ? starts[seqId + 1] - (size - 1 - i) * stride + : stridePos.back() + stride; + stridePos.emplace_back(cur); + } + } + } + stridePos.emplace_back(starts[numSequences]); + int size = stridePos.size(); + CHECK_EQ(size - 1, tgtBuf[numSequences]); + ICpuGpuVector::resizeOrCreate(*stridePostions, size, false); + (*stridePostions)->getMutableVector(false)->copyFrom(stridePos.data(), size); +} + +void Argument::getValueString( + std::unordered_map* out) const { + if (value) { + std::ostringstream os; + value->print(os); + out->insert({"value", os.str()}); + } + if (ids) { + std::ostringstream os; + ids->print(os, ids->getSize()); + out->insert({"ids", os.str()}); + } + if (sequenceStartPositions) { + std::ostringstream os; + sequenceStartPositions->getVector(false)->print( + os, sequenceStartPositions->getSize()); + out->insert({"sequence pos", os.str()}); + } + if (subSequenceStartPositions) { + std::ostringstream os; + subSequenceStartPositions->getVector(false)->print( + os, subSequenceStartPositions->getSize()); + out->insert({"sub-sequence pos", os.str()}); + } +} + +void Argument::printValueString(std::ostream& stream, + const std::string& prefix) const { + std::unordered_map out; + getValueString(&out); + for (auto field : {"value", "ids", "sequence pos", "sub-sequence pos"}) { + auto it = out.find(field); + if (it != out.end()) { + stream << prefix << field << ":\n" << it->second; + } + } +} + +void Argument::subArgFrom(const Argument& input, + size_t offset, + size_t height, + size_t width, + bool useGpu, + bool trans, + bool seqFlag, + size_t seqStart, + size_t seqSize) { + if (input.value) { + value = Matrix::create( + input.value->getData() + offset * width, height, width, trans, useGpu); + } + if (input.ids) { + ids = IVector::create(input.ids->getData() + offset, height, useGpu); + } + if (input.grad) { + grad = Matrix::create( + input.grad->getData() + offset * width, height, width, trans, useGpu); + } + if (seqFlag) { + sequenceStartPositions = std::make_shared( + *(input.sequenceStartPositions), seqStart, seqSize); + } +} + +void Argument::reorganizeSeqInfo( + const ICpuGpuVectorPtr seqStartPos, + const ICpuGpuVectorPtr subSeqStartPos, + std::vector>& reorganizedSeqInfo) { + CHECK(seqStartPos); + reorganizedSeqInfo.clear(); + + int seqNum = seqStartPos->getSize() - 1; + int* seqStarts = seqStartPos->getMutableData(false); + + if (subSeqStartPos) { + int* subSeqStarts = subSeqStartPos->getMutableData(false); + reorganizedSeqInfo.resize(seqNum, std::vector()); + int seqIdx = 0; + for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) { + reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + if (subSeqStarts[i] == seqStarts[seqIdx + 1]) { + seqIdx++; + if (seqIdx == seqNum) return; + reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + } + } + } else { + reorganizedSeqInfo.resize(1, std::vector(seqNum + 1, 0)); + memcpy(reorganizedSeqInfo[0].data(), + seqStarts, + sizeof(int) * seqStartPos->getSize()); + } +} + +} // namespace paddle diff --git a/paddle/legacy/parameter/Argument.h b/paddle/legacy/parameter/Argument.h new file mode 100644 index 0000000000000000000000000000000000000000..ea8634896c18c7c3516c0d584aec4b475d626e61 --- /dev/null +++ b/paddle/legacy/parameter/Argument.h @@ -0,0 +1,349 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "hl_gpu.h" + +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +typedef std::shared_ptr> SVectorPtr; + +struct Argument { + Argument() + : in(nullptr), + value(nullptr), + ids(nullptr), + grad(nullptr), + strs(nullptr), + frameHeight(0), + frameWidth(0), + frameDepth(0), + sequenceStartPositions(nullptr), + subSequenceStartPositions(nullptr), + cpuSequenceDims(nullptr), + deviceId(-1), + allCount(0), + valueCount(0), + gradCount(0), + dataId(0) {} + Argument(const Argument& argument) { + *this = argument; + valueCount = 0; + gradCount = 0; + dataId = argument.dataId; + } + ~Argument() {} + + void operator=(const Argument& argument) { + in = argument.in; + value = argument.value; + ids = argument.ids; + grad = argument.grad; + strs = argument.strs; + sequenceStartPositions = argument.sequenceStartPositions; + subSequenceStartPositions = argument.subSequenceStartPositions; + cpuSequenceDims = argument.cpuSequenceDims; + deviceId = argument.deviceId; + allCount = argument.allCount; + frameHeight = argument.frameHeight; + frameWidth = argument.frameWidth; + frameDepth = argument.frameDepth; + dataId = argument.dataId; + } + + MatrixPtr in; // used if needed + MatrixPtr value; + IVectorPtr ids; // a sequence of ids. Can be use for class id for costLayer + MatrixPtr grad; // If empty, gradient is not needed. + SVectorPtr strs; + + // A dataBatch includes batchSize frames, one frame maybe not only vector + size_t frameHeight; + size_t frameWidth; + size_t frameDepth; + + // If NULL, each position is treated independently. + // Otherwise, its size should be #NumberOfSequences + 1. + // The first position is always 0 and + // the last position should be equal to batchSize. + ICpuGpuVectorPtr sequenceStartPositions; + + // If NULL, each sequence has no subsequence. + // Otherwise, its size should be #NumberOfSubSequences + 1. + // The first position is always 0 and + // the last position should be equal to batchSize. + ICpuGpuVectorPtr subSequenceStartPositions; + + // dimension of sequence, stored only in CPU + IVectorPtr cpuSequenceDims; + + int deviceId; // the GPU device id which the argument in + int allCount; // the number of output layers using this argument + mutable int valueCount; // waiting this member when layer do forward + mutable int gradCount; // waiting this member when layer do backward + mutable LockedCondition valueReadyCond; + mutable LockedCondition gradReadyCond; + + int dataId; // dataProvider id + + /* Increase the reference count of the argument. */ + void countIncrement() { allCount++; } + + int getAllCount() const { return allCount; } + + void waitValueReady() const { + valueReadyCond.wait([this] { return (valueCount != 0); }); + + std::lock_guard guard(*valueReadyCond.mutex()); + valueCount--; + } + + void notifyValueReady() const { + valueReadyCond.notify_all([this] { valueCount = allCount; }); + } + + void waitGradReady() const { + gradReadyCond.wait([this] { return (gradCount == allCount); }); + gradCount = 0; + } + + void notifyGradReady() const { + gradReadyCond.notify_all([this] { gradCount++; }); + } + + int64_t getBatchSize() const { + if (value) return value->getHeight(); + if (ids) return ids->getSize(); + if (grad) return grad->getHeight(); + if (in) return in->getHeight(); + if (strs) return strs->size(); + return 0; + } + size_t getFrameHeight() const { return frameHeight; } + size_t getFrameWidth() const { return frameWidth; } + size_t getFrameDepth() const { return frameDepth; } + void setFrameHeight(size_t h) { frameHeight = h; } + void setFrameWidth(size_t w) { frameWidth = w; } + void setFrameDepth(size_t d) { frameDepth = d; } + + int64_t getNumSequences() const { + return sequenceStartPositions ? sequenceStartPositions->getSize() - 1 + : getBatchSize(); + } + + int64_t getNumSubSequences() const { + return subSequenceStartPositions ? subSequenceStartPositions->getSize() - 1 + : getBatchSize(); + } + + bool hasSeq() const { return sequenceStartPositions != nullptr; } + bool hasSubseq() const { return subSequenceStartPositions != nullptr; } + + const int* getCpuStartPositions() const { + return hasSubseq() ? subSequenceStartPositions->getData(false) + : sequenceStartPositions->getData(false); + } + + static inline real sum(const std::vector& arguments) { + real cost = 0; + for (auto& arg : arguments) { + if (arg.value) { + SetDevice device(arg.deviceId); + cost += arg.value->getSum(); + } + } + return cost; + } + + /** + * @brief (value, ids, grad, sequenceStartPositions) of output are subset of + * input. Note that, output share the same memory of input. + * + * @param input[in] input + * @param offset[in] offset in terms of rows + * @param height[in] height of output.value + * @param width[in] width of output.value + * @param useGpu[in] + * @param trans[in] whether input.value is transform + * @param seqFlag[in] whether input has sequenceStartPositions + * @param seqStart[in] offset of input.sequenceStartPositions + * @param seqSize[in] lenght of output.sequenceStartPositions + */ + void subArgFrom(const Argument& input, + size_t offset, + size_t height, + size_t width, + bool useGpu, + bool trans = false, + bool seqFlag = false, + size_t seqStart = 0, + size_t seqSize = 0); + /* + * for sequence input: + * startSeq: the sequence id of start + * copySize: how many sequences need to copy + * return value: how many samples are copied + * for non-sequence input: + * startSeq: the sample id of start + * copySize: how many samples need to copy + * return value: how many samples are copied + * Note that when specifying the stream explicitly in this case, + * synchronize should also be called somewhere after this function + */ + int32_t resizeAndCopyFrom(const Argument& src, + int32_t startSeq, + int32_t copySize, + bool useGpu, + hl_stream_t stream); + + /* + * same with the above function, except that the stream is + * HPPL_STREAM_DEFAULT and synchronize is automatically called + * inside it + */ + int32_t resizeAndCopyFrom(const Argument& src, + int32_t startSeq, + int32_t copySize, + bool useGpu = FLAGS_use_gpu); + + void resizeAndCopyFrom(const Argument& src, bool useGpu, hl_stream_t stream); + + /* + * same with the above function, except that the stream is + * HPPL_STREAM_DEFAULT and synchronize is automatically called + * inside it + */ + void resizeAndCopyFrom(const Argument& src, bool useGpu = FLAGS_use_gpu); + + /* + @brief Concatenate several arguments into one and put the result into it. + @param args : a vector of argument, each element of which is a frame in a + batch of sequences. + @param selectRows : select several row of args to concatenate + @param seqStartPos : sequence start positions in the final Argument + @param hl_stream_t : cuda stream + @param passTyoe : type of task, training or testing + */ + void concat(const std::vector& args, + const std::vector& selectRows, + const std::vector& seqStartPos, + const std::vector& copySize, + bool useGpu, + hl_stream_t stream, + PassType passType); + + /* + Concatenate several args into one and put the result into this. + */ + void concat(const std::vector& src, + bool useGpu = FLAGS_use_gpu, + hl_stream_t stream = HPPL_STREAM_DEFAULT, + PassType passType = PASS_TEST); + + /* + * split vector to several vectors according to dataId + */ + static void splitByDataId(const std::vector& argus, + std::vector>* arguGroups); + + struct SeqInfo { + // Equal to sequence length for sequence data + // Equal to number of subsequences for subsequence data + int topLevelLength; + + int seqStart; + int seqId; + + // Equal to topLevelLength for sequence data + // Equal to sum of the length of subsequences for subsequence data + int subLevelLength; + + // Only used for subsequence data, start position of this sequence + // is subSequenceStartPositions, i.e. + // subSequenceStartPositions[subSeqStart] == seqStart + int subSeqStart; + }; + /* + Get SeqInfo for each sequence of this argument + Elements in *seqInfo are sorted by topLevelLength in descending order + */ + void getSeqInfo(std::vector* segInfo) const; + + /* + Check Whether sequenceStartPositions is subset of + subSequenceStartPositions. + */ + void checkSubset() const; + + /* + sequence has sub-sequence degrades to a sequence. + */ + void degradeSequence(const Argument& input); + + /* + After pooling with stride n (n is smaller than sequence length), + a long sequence will be shorten. + This function is invalid for sequence having sub-sequence. + */ + void poolSequenceWithStride(const Argument& input, + size_t stride, + ICpuGpuVectorPtr* stridePositions, + bool reversed = false); + /** + * @brief getValueString will return the argument's output in string. There + * are several kinds of output. The keys of output dictionary are 'value', + * 'id', 'sequence pos', 'sub-sequence pos'. + * @param out [out]: the return values. + */ + void getValueString(std::unordered_map* out) const; + + /** + * @brief printValueString will print the argument's output in order of + * 'value', 'id', 'sequence pos', 'sub-sequence pos'. + * @param stream: Output stream + * @param prefix: line prefix for printing. + */ + void printValueString(std::ostream& stream, + const std::string& prefix = "") const; + + /** + * @brief reorganizeSeqInfo will reorganize sequenceStartPositions and + * subSequenceStartPositions into a 2 dimensional arrary: reorganizedSeqInfo. + * + * @param seqStartPos: sequenceStartPositions of an Argument. + * @param subSeqStartPos: subSequenceStartPositions of an Argument. + * @param the reorganized sequence start position information. + * + * Examples: + * seqStartPos: [0, 4, 15, 20, 28] + * subSeqStartPos: [0, 3, 4, 5, 7, 10, 15, 20, 22, 23, 25, 28] + * reorganizedSeqInfo: + * [ + * [0,3,4], + * [4,5,7,10,15], + * [15,20], + * [20,22,23,25,28] + * ] + */ + static void reorganizeSeqInfo( + const ICpuGpuVectorPtr seqStartPos, + const ICpuGpuVectorPtr subSeqStartPos, + std::vector>& reorganizedSeqInfo); +}; + +} // namespace paddle diff --git a/paddle/parameter/AverageOptimizer.cpp b/paddle/legacy/parameter/AverageOptimizer.cpp similarity index 100% rename from paddle/parameter/AverageOptimizer.cpp rename to paddle/legacy/parameter/AverageOptimizer.cpp diff --git a/paddle/parameter/AverageOptimizer.h b/paddle/legacy/parameter/AverageOptimizer.h similarity index 100% rename from paddle/parameter/AverageOptimizer.h rename to paddle/legacy/parameter/AverageOptimizer.h diff --git a/paddle/parameter/CMakeLists.txt b/paddle/legacy/parameter/CMakeLists.txt similarity index 100% rename from paddle/parameter/CMakeLists.txt rename to paddle/legacy/parameter/CMakeLists.txt diff --git a/paddle/legacy/parameter/FirstOrderOptimizer.cpp b/paddle/legacy/parameter/FirstOrderOptimizer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4f82a115f7bb467737b53b9891d88d3c4f501faf --- /dev/null +++ b/paddle/legacy/parameter/FirstOrderOptimizer.cpp @@ -0,0 +1,330 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "FirstOrderOptimizer.h" +#include "paddle/legacy/math/TrainingAlgorithmOp.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Util.h" + +#include + +DEFINE_bool(log_clipping, false, "enable log clipping or not"); + +namespace paddle { + +SparseMomentumParameterOptimizer::SparseMomentumParameterOptimizer( + const OptimizationConfig& optConfig) + : ParameterOptimizer(optConfig) { + addParameterType(PARAMETER_MOMENTUM); + addParameterType(PARAMETER_MOMENTUM_UT); + addParameterType(PARAMETER_MOMENTUM_VT); + alpha_ = 1; + beta_ = 1; + tau_ = -1; + threshold_ = 1e+06; +} + +void SparseMomentumParameterOptimizer::init(size_t numRows, + const ParameterConfig* config) { + isParameterSparse_ = numRows != 0; + t0Vec_.resize(numRows); + t0Vec_.assign(t0Vec_.size(), 0); + timer_ = 0; + momentum_ = config->momentum(); + decayRate_ = config->decay_rate(); + gamma_ = config->learning_rate(); +} + +void SparseMomentumParameterOptimizer::startBatch(int64_t numSamplesProcessed) { + learningRate_ = calcLearningRate(numSamplesProcessed, pass_); + if (isParameterSparse_) { + tau_ = tau_ + beta_ / alpha_; + alpha_ = alpha_ / momentum_; + beta_ = beta_ / (1 + decayRate_ * gamma_ * learningRate_); + } +} + +void SparseMomentumParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& paraConfig, + size_t sparseId) const { + if (sparseId != -1LU) { + CHECK_LT(sparseId, t0Vec_.size()); + if (t0Vec_[sparseId] == 0) { + vecs[PARAMETER_MOMENTUM_VT]->assign(*vecs[PARAMETER_VALUE]); + t0Vec_[sparseId] = 1; + } + vecs[PARAMETER_MOMENTUM_UT]->add(*vecs[PARAMETER_GRADIENT], + -alpha_ * gamma_ * learningRate_); + vecs[PARAMETER_MOMENTUM_VT]->add(*vecs[PARAMETER_GRADIENT], + tau_ * alpha_ * gamma_ * learningRate_); + vecs[PARAMETER_VALUE]->add(*vecs[PARAMETER_MOMENTUM_UT], + tau_ / beta_ + 1.0 / alpha_, + *vecs[PARAMETER_MOMENTUM_VT], + 1.0 / beta_); + + } else { + vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], + *vecs[PARAMETER_MOMENTUM], + learningRate_ * paraConfig.learning_rate(), + paraConfig.momentum(), + applyDecay_ ? paraConfig.decay_rate() : 0); + } +} + +ParameterOptimizer::TraverseCallback +SparseMomentumParameterOptimizer::needSpecialTraversal( + const ParameterConfig& config) const { + if (alpha_ > threshold_ && isParameterSparse_) { + // Restart to avoid large value multiplication + // 1. \alpha = 1, \beta = 1, \tau = 0 + // 2. Note that \tau * u_t + v_t = \beta \theta_t, therefore: + // u_t should be rescaled to u_t/alpha_ + // v_t should be reset to \theta_t + return [this](const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) { + vecs[PARAMETER_MOMENTUM_UT]->divScalar(alpha_); + vecs[PARAMETER_MOMENTUM_VT]->assign(*vecs[PARAMETER_VALUE]); + }; + } else { + return nullptr; + } +} + +void SparseMomentumParameterOptimizer::finishBatch() { + timer_++; + if (!isParameterSparse_) return; + if (alpha_ > threshold_) { + alpha_ = 1; + beta_ = 1; + tau_ = -1; + } +} + +void AdagradParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + BaseMatrix& value = *vecs[PARAMETER_VALUE]; + BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; + BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; + BaseMatrix& accum_buffer = *vecs[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& accum = *vecs[PARAMETER_GRADIENT_SQURESUM1]; + BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; + + real epsilon = optConfig_.ada_epsilon(); + real learningRate = learningRate_ * config.learning_rate(); + real momentum = config.momentum(); + real decayRate = applyDecay_ ? config.decay_rate() : 0; + + adagradApply(value, + grad, + mom, + accum_buffer, + accum, + lr, + epsilon, + learningRate, + momentum, + decayRate); +} + +ParameterOptimizer::TraverseCallback +AdagradParameterOptimizer::needSpecialTraversal( + const ParameterConfig& config) const { + if (numUpdates_ % kMaxNumAccumulates == 0) { + // Move the sum to a different buffer to avoid loss of precision + // due to too many sums. + return [](const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) { + vecs[PARAMETER_GRADIENT_SQURESUM]->add( + *vecs[PARAMETER_GRADIENT_SQURESUM1]); + vecs[PARAMETER_GRADIENT_SQURESUM1]->zeroMem(); + }; + } else { + return nullptr; + } +} + +void AdaDeltaParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + CHECK(sparseId == -1LU) << "Sparse update is not supported"; + + BaseMatrix& value = *vecs[PARAMETER_VALUE]; + BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; + BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; + BaseMatrix& accum = *vecs[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& accum_update = *vecs[PARAMETER_GRADIENT_SQURESUM1]; + BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; + + real learningRate = learningRate_ * config.learning_rate(); + real momentum = config.momentum(); + real decayRate = applyDecay_ ? config.decay_rate() : 0; + + adadeltaApply(value, + grad, + mom, + accum, + accum_update, + lr, + rou_, + epsilon_, + learningRate, + momentum, + decayRate); +} + +void RMSPropParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + BaseMatrix& value = *vecs[PARAMETER_VALUE]; + BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; + BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; + BaseMatrix& sum = *vecs[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& sum1 = *vecs[PARAMETER_GRADIENT_SQURESUM1]; + BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; + + real accumulatedRou = rou_; + bool firstTime = timer_ == 0; + if (sparseId != -1LU) { + CHECK_LT(sparseId, t0Vec_.size()); + accumulatedRou = std::pow(rou_, timer_ + 1 - t0Vec_[sparseId]); + firstTime = t0Vec_[sparseId] == 0; + t0Vec_[sparseId] = timer_ + 1; + } + + real epsilon = optConfig_.ada_epsilon(); + real learningRate = learningRate_ * config.learning_rate(); + real momentum = config.momentum(); + real decayRate = applyDecay_ ? config.decay_rate() : 0; + + rmspropApply(value, + grad, + mom, + sum, + sum1, + lr, + accumulatedRou, + rou_, + epsilon, + learningRate, + momentum, + decayRate, + firstTime); +} + +void DecayedAdagradParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + BaseMatrix& value = *vecs[PARAMETER_VALUE]; + BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; + BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; + BaseMatrix& sum = *vecs[PARAMETER_GRADIENT_SQURESUM]; + BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; + + real accumulatedRou = rou_; + bool firstTime = timer_ == 0; + if (sparseId != -1LU) { + CHECK_LT(sparseId, t0Vec_.size()); + accumulatedRou = std::pow(rou_, timer_ + 1 - t0Vec_[sparseId]); + firstTime = t0Vec_[sparseId] == 0; + t0Vec_[sparseId] = timer_ + 1; + } + + real epsilon = optConfig_.ada_epsilon(); + real learningRate = learningRate_ * config.learning_rate(); + real momentum = config.momentum(); + real decayRate = applyDecay_ ? config.decay_rate() : 0; + + decayedAdagradApply(value, + grad, + mom, + sum, + lr, + accumulatedRou, + rou_, + epsilon, + learningRate, + momentum, + decayRate, + firstTime); +} + +void AdamParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + CHECK(sparseId == -1UL) << "Sparse update is not supported"; + + real beta1_power = std::pow(beta1_, step_); + real beta2_power = std::pow(beta2_, step_); + real learningRate = config.learning_rate() * learningRate_; + + BaseMatrix& value = *vecs[PARAMETER_VALUE]; + BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; + BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; + BaseMatrix& v = *vecs[PARAMETER_SECOND_MOMENTUM]; + + adamApply(value, + grad, + mom, + v, + beta1_, + beta2_, + beta1_power, + beta2_power, + epsilon_, + learningRate); +} + +void AdamaxParameterOptimizer::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + CHECK(sparseId == -1UL) << "Sparse update is not supported"; + real learningRate = config.learning_rate() * learningRate_; + + BaseMatrix& value = *vecs[PARAMETER_VALUE]; + BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; + BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; + BaseMatrix& u = *vecs[PARAMETER_WEIGHTED_INFINITY_NORM]; + + adamaxApply(value, grad, mom, u, beta1_, beta2_, step_, learningRate); +} + +void OptimizerWithGradientClipping::update(const VectorPtr vecs[], + const ParameterConfig& config, + size_t sparseId) const { + real globalThreshold = optConfig_.gradient_clipping_threshold(); + real localThreshold = config.gradient_clipping_threshold(); + + // Use local gradient clipping threshold if it's enabled, + // otherwise using the global one. + real threshold = localThreshold > 0.0f ? localThreshold : globalThreshold; + std::string field = localThreshold > 0.0f ? "local" : "global"; + + real maxAbsGrad = vecs[PARAMETER_GRADIENT]->getAbsMax(); + if (maxAbsGrad > threshold) { + if (FLAGS_log_clipping) { + real avgAbsGrad = vecs[PARAMETER_GRADIENT]->getAbsSum() / + vecs[PARAMETER_GRADIENT]->getSize(); + LOG(INFO) << "parameter=" << config.name() << " need clipping by " + << field << " threshold=" << threshold + << ", max grad=" << maxAbsGrad << ", avg grad=" << avgAbsGrad; + } + vecs[PARAMETER_GRADIENT]->clip(-threshold, threshold); + } + optimizer_->update(vecs, config, sparseId); +} + +} // namespace paddle diff --git a/paddle/parameter/FirstOrderOptimizer.h b/paddle/legacy/parameter/FirstOrderOptimizer.h similarity index 100% rename from paddle/parameter/FirstOrderOptimizer.h rename to paddle/legacy/parameter/FirstOrderOptimizer.h diff --git a/paddle/legacy/parameter/LearningRateScheduler.cpp b/paddle/legacy/parameter/LearningRateScheduler.cpp new file mode 100644 index 0000000000000000000000000000000000000000..68c44a7ec49f64a1085609d906441c9ed4502888 --- /dev/null +++ b/paddle/legacy/parameter/LearningRateScheduler.cpp @@ -0,0 +1,173 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "LearningRateScheduler.h" +#include "paddle/legacy/utils/StringUtil.h" + +namespace paddle { + +ClassRegistrar + LearningRateScheduler::registrar_; + +LearningRateScheduler* LearningRateScheduler::create( + const OptimizationConfig& config) { + return registrar_.createByType(config.learning_rate_schedule(), config); +} + +// LRS stands for LearningRateScheduler + +class BaseLRS : public LearningRateScheduler { + public: + explicit BaseLRS(const OptimizationConfig& config) + : learningRate_(config.learning_rate()), + a_(config.learning_rate_decay_a()), + b_(config.learning_rate_decay_b()) {} + + protected: + real learningRate_; + real a_; + real b_; +}; + +class ConstLRS : public BaseLRS { + public: + explicit ConstLRS(const OptimizationConfig& config) : BaseLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + return learningRate_; + } +}; +REGISTER_LEARNING_RATE_SCHEDULER(constant, ConstLRS); + +class PolyLRS : public BaseLRS { + public: + explicit PolyLRS(const OptimizationConfig& config) : BaseLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + return learningRate_ * pow(1.0 + a_ * numSamplesProcessed, -b_); + } +}; +REGISTER_LEARNING_RATE_SCHEDULER(poly, PolyLRS); + +class CaffePolyLRS : public BaseLRS { + public: + explicit CaffePolyLRS(const OptimizationConfig& config) : BaseLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + if (numSamplesProcessed > a_) { + LOG_FIRST_N(WARNING, 1) + << "Using caffe_poly learning rate schedule, " + << "learning rate hits ZERO when " + << "numSamplesProcessed > config.learning_rate_decay_b(), " + << "training is over and you can stop it. " + << "See common/LearningRateScheduler.cpp for more info."; + return 0; + } else { + return learningRate_ * pow(1.0 - numSamplesProcessed / a_, b_); + } + } +}; +REGISTER_LEARNING_RATE_SCHEDULER(caffe_poly, CaffePolyLRS); + +class ExpLRS : public BaseLRS { + public: + explicit ExpLRS(const OptimizationConfig& config) : BaseLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + double decayRatio = (double)numSamplesProcessed / b_; + return learningRate_ * pow(a_, decayRatio); + } +}; +REGISTER_LEARNING_RATE_SCHEDULER(exp, ExpLRS); + +class DiscreteExpLRS : public BaseLRS { + public: + explicit DiscreteExpLRS(const OptimizationConfig& config) : BaseLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + int numDecays = floor(numSamplesProcessed / b_); + return learningRate_ * pow(a_, numDecays); + } +}; +REGISTER_LEARNING_RATE_SCHEDULER(discexp, DiscreteExpLRS); + +class LinearLRS : public BaseLRS { + public: + explicit LinearLRS(const OptimizationConfig& config) : BaseLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + return std::max(learningRate_ - a_ * numSamplesProcessed, b_); + } +}; +REGISTER_LEARNING_RATE_SCHEDULER(linear, LinearLRS); + +/* + specify learning rate through + learning_rate_args = 'seg0:rate0,seg1:rate1,...,segK:rateK' + if seg_{i-1} <= numSamples <= seg_i, + then learning_rate = learning_rate_base * rate_i +*/ +class ManualLRS : public BaseLRS { + public: + explicit ManualLRS(const OptimizationConfig& config) + : BaseLRS(config), currentSegment_(0), lastNum_(0) { + std::vector pieces; + str::split(config.learning_rate_args(), ',', &pieces); + rates_.reserve(pieces.size()); + std::string s1, s2; + + for (auto& piece : pieces) { + auto pos = piece.find(':'); + CHECK(pos != std::string::npos) << "Wrong format for learning_rate_args: " + << config.learning_rate_args(); + segments_.push_back(str::to(piece.substr(0, pos))); + rates_.push_back(str::to(piece.substr(pos + 1))); + } + } + + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + return calc(numSamplesProcessed); + } + + real calc(int64_t num) { + // We assume that num never decreases. + CHECK_LE(lastNum_, num); + lastNum_ = num; + while (currentSegment_ < rates_.size()) { + if (num <= segments_[currentSegment_]) { + return learningRate_ * rates_[currentSegment_]; + } + ++currentSegment_; + if (currentSegment_ < rates_.size()) { + LOG(INFO) << " learning_rate changes to " + << learningRate_ * rates_[currentSegment_]; + } + } + return learningRate_ * rates_.back(); + } + + protected: + std::vector rates_; + std::vector segments_; + size_t currentSegment_; + int64_t lastNum_; +}; + +REGISTER_LEARNING_RATE_SCHEDULER(manual, ManualLRS); + +class PassManualLRS : public ManualLRS { + public: + explicit PassManualLRS(const OptimizationConfig& config) + : ManualLRS(config) {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { + return calc(pass); + } +}; + +REGISTER_LEARNING_RATE_SCHEDULER(pass_manual, PassManualLRS); +} // namespace paddle diff --git a/paddle/legacy/parameter/LearningRateScheduler.h b/paddle/legacy/parameter/LearningRateScheduler.h new file mode 100644 index 0000000000000000000000000000000000000000..fc7e380a6af58577f4ba319d85522535b8f93a45 --- /dev/null +++ b/paddle/legacy/parameter/LearningRateScheduler.h @@ -0,0 +1,37 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "TrainerConfig.pb.h" +#include "paddle/legacy/utils/ClassRegistrar.h" + +namespace paddle { +// NOLINTNEXTLINES_4 +#define REGISTER_LEARNING_RATE_SCHEDULER(__type_name, __class_name) \ + static InitFunction __reg_type_##__type_name([]() { \ + LearningRateScheduler::registrar_.registerClass<__class_name>( \ + #__type_name); \ + }) + +class LearningRateScheduler { + public: + static LearningRateScheduler* create(const OptimizationConfig& config); + virtual ~LearningRateScheduler() {} + virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) = 0; + + static ClassRegistrar registrar_; +}; + +} // namespace paddle diff --git a/paddle/parameter/OptimizerFunctions.cpp b/paddle/legacy/parameter/OptimizerFunctions.cpp similarity index 100% rename from paddle/parameter/OptimizerFunctions.cpp rename to paddle/legacy/parameter/OptimizerFunctions.cpp diff --git a/paddle/parameter/OptimizerFunctions.h b/paddle/legacy/parameter/OptimizerFunctions.h similarity index 100% rename from paddle/parameter/OptimizerFunctions.h rename to paddle/legacy/parameter/OptimizerFunctions.h diff --git a/paddle/parameter/OptimizerWithRegularizer.cpp b/paddle/legacy/parameter/OptimizerWithRegularizer.cpp similarity index 100% rename from paddle/parameter/OptimizerWithRegularizer.cpp rename to paddle/legacy/parameter/OptimizerWithRegularizer.cpp diff --git a/paddle/parameter/OptimizerWithRegularizer.h b/paddle/legacy/parameter/OptimizerWithRegularizer.h similarity index 100% rename from paddle/parameter/OptimizerWithRegularizer.h rename to paddle/legacy/parameter/OptimizerWithRegularizer.h diff --git a/paddle/legacy/parameter/Parameter.cpp b/paddle/legacy/parameter/Parameter.cpp new file mode 100644 index 0000000000000000000000000000000000000000..666d808f0c13c5c828c51b2a36ee9d05f7f78c13 --- /dev/null +++ b/paddle/legacy/parameter/Parameter.cpp @@ -0,0 +1,425 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Parameter.h" +#include +#include +#include "AverageOptimizer.h" +#include "FirstOrderOptimizer.h" +#include "OptimizerFunctions.h" +#include "OptimizerWithRegularizer.h" +#include "ParameterUpdateFunctions.h" +#include "ThreadLocalBuffer.h" +#include "hl_gpu.h" +#include "paddle/legacy/math/CpuSparseMatrix.h" +#include "paddle/legacy/math/MathUtils.h" +#include "paddle/legacy/math/SparseRowMatrix.h" +#include "paddle/legacy/utils/Logging.h" + +DEFINE_int32(enable_grad_share, + (100 * 1024 * 1024), + "threshold for enable gradient parameter share for batch " + "multi-cpu training"); +DEFINE_int32( + grad_share_block_num, + 64, + "block number of gradient parameter share for batch multi-cpu training"); + +namespace paddle { + +const std::string Parameter::kMissParameterFail = "fail"; +const std::string Parameter::kMissParameterRand = "rand"; +const std::string Parameter::kMissParameterZero = "zero"; + +Parameter::Parameter(const ParameterConfig& config, bool useGpu, bool doInit) + : config_(config), + useGpu_(useGpu), + deviceId_(-1), + sharedCount_(0), + updateCounter_(0), + updated_(false), + headerFormat_(PARAM_FORMAT_ORIGINAL) { + setID(-1); /* capture uninitialized id */ + if (useGpu_ && FLAGS_parallel_nn) { + /* gpu environment is specified by device property */ + deviceId_ = config_.device(); + if (deviceId_ < 0) { + useGpu_ = false; + } + } + + if (doInit) { + initialize(); + } + + for (int i = 0; i < config.update_hooks_size(); ++i) { + this->updaterHooks_.push_back(IParameterUpdaterHook::create(config, i)); + } +} + +void Parameter::initialize() { + SetDevice device(deviceId_); + + bufs_[PARAMETER_VALUE] = + Vector::createParallelVector(config_.size(), useGpu_); + bufs_[PARAMETER_VALUE]->zeroMem(); + + if (config_.is_sparse()) { + enableSparseParameter(); + } + + if (!isStatic()) { + bufs_[PARAMETER_GRADIENT] = + Vector::createParallelVector(config_.size(), useGpu_); + bufs_[PARAMETER_MOMENTUM] = + Vector::createParallelVector(config_.size(), useGpu_); + + bufs_[PARAMETER_GRADIENT]->zeroMem(); + bufs_[PARAMETER_MOMENTUM]->zeroMem(); + } +} + +void Parameter::randomize(const VectorPtr& value, + const ParameterConfig& config) { + if (PARAMETER_INIT_UNIFORM == config.initial_strategy()) { + // initialize the parameter as uniform distribution + real initial_min = config.initial_mean() - config.initial_std(); + real initial_max = config.initial_mean() + config.initial_std(); + value->uniform(initial_min, initial_max); + VLOG(1) << config.name() << ": initial_min=" << initial_min + << ", initial_max=" << initial_max; + } else if (PARAMETER_INIT_NORMAL == config.initial_strategy()) { + /* Initialize the parameters randomly */ + value->randnorm(config.initial_mean(), config.initial_std()); + VLOG(1) << config.name() << ": initial_mean=" << config.initial_mean() + << ", initial_std=" << config.initial_std(); + } else { + LOG(FATAL) << "not supported initial_strategy: " + << config.initial_strategy(); + } +} + +void Parameter::randomize() { + if (!bufs_[PARAMETER_VALUE]) return; + SetDevice device(deviceId_); + Parameter::randomize(bufs_[PARAMETER_VALUE], config_); + + if (config_.is_sparse()) { + if (format_ == SPARSE_CSC) { + sparseRand(intBufs_[PARAMETER_COLS]->getData(), + intBufs_[PARAMETER_ROWS]->getData(), + config_.size(), + config_.dims(1) + 1, + config_.dims(0), + useGpu_); + } else { + sparseRand(intBufs_[PARAMETER_ROWS]->getData(), + intBufs_[PARAMETER_COLS]->getData(), + config_.size(), + config_.dims(0) + 1, + config_.dims(1), + useGpu_); + } + } + setValueUpdated(); +} + +void Parameter::zeroMem() { + if (!bufs_[PARAMETER_VALUE]) return; + bufs_[PARAMETER_VALUE]->zeroMem(); + setValueUpdated(); + LOG(INFO) << getName() << " set to 0"; +} + +bool Parameter::isGradShared(size_t* blockNum) { + if (!useGpu_ && !isStatic() && FLAGS_enable_grad_share > 0 && + !isGradSparseUpdate() && + this->getSize() > (size_t)FLAGS_enable_grad_share) { + if (blockNum) { + *blockNum = (size_t)FLAGS_grad_share_block_num; + } + return true; + } + return false; +} + +bool Parameter::isValueShared() { + return !useGpu_ && config_.is_shared() && FLAGS_trainer_count > 1; +} + +bool Parameter::isGradSparseUpdate() const { + return !useGpu_ && !isStatic() && + (config_.sparse_update() || config_.sparse_remote_update()); +} + +void Parameter::setMat(ParameterType pType, int matType) { + CHECK(!mats_[pType]); + + if (config_.dims_size() == 0 && matType == MAT_NORMAL) { + return; + } + + CHECK_EQ((size_t)config_.dims_size(), 2LU); + size_t height = config_.dims(0); + size_t width = config_.dims(1); + if (matType == MAT_NORMAL) { + if (!config_.is_sparse()) { + CHECK_EQ(height * width, bufs_[pType]->getSize()); + mats_[pType] = + Matrix::create(bufs_[pType]->getMemoryHandle(), height, width); + } else { + size_t size = bufs_[pType]->getSize(); + CHECK_GE(height * width, size); + if (format_ == SPARSE_CSR) { + CHECK_EQ(height + 1, intBufs_[PARAMETER_ROWS]->getSize()); + CHECK_EQ(size, intBufs_[PARAMETER_COLS]->getSize()); + } else { + CHECK_EQ(width + 1, intBufs_[PARAMETER_COLS]->getSize()); + CHECK_EQ(size, intBufs_[PARAMETER_ROWS]->getSize()); + } + mats_[pType] = + Matrix::createSparseMatrix(bufs_[pType]->getData(), + intBufs_[PARAMETER_ROWS]->getData(), + intBufs_[PARAMETER_COLS]->getData(), + height, + width, + bufs_[pType]->getSize(), + FLOAT_VALUE, + format_, + false, + useGpu_); + } + } +#ifndef PADDLE_MOBILE_INFERENCE + // NOLINTNEXTLINE + else if (matType == MAT_NORMAL_SHARED) { + CHECK_EQ(height * width, bufs_[pType]->getSize()); + size_t blockNum = 0; + CHECK(isGradShared(&blockNum)); + mats_[pType] = std::make_shared( + blockNum, + std::dynamic_pointer_cast( + bufs_[pType]->getMemoryHandle()), + height, + width); + } else if (matType == MAT_VALUE_SHARED) { + CHECK_EQ(height * width, bufs_[pType]->getSize()); + mats_[pType] = std::make_shared( + std::dynamic_pointer_cast( + bufs_[pType]->getMemoryHandle()), + height, + width); + } else if (matType == MAT_SPARSE_ROW_IDS) { + CHECK_EQ(height * width, bufs_[pType]->getSize()); + mats_[pType] = std::make_shared( + std::dynamic_pointer_cast( + bufs_[pType]->getMemoryHandle()), + height, + width); + } else if (matType == MAT_SPARSE_ROW) { + auto valueMat = + std::dynamic_pointer_cast(mats_[PARAMETER_VALUE]); + SparseRowCpuMatrix::IndexDictPtr indexDict(nullptr); + if (pType != PARAMETER_VALUE) { + CHECK(valueMat) << "The matrix for PARAMETER_VALUE must be set " + << " and its type must be MAT_SPARSE_ROW," + << " MAT_SPARSE_ROW_PREFETCH or MAT_CACHE_ROW"; + indexDict = valueMat->getIndexDictHandle(); + } + auto mat = + std::make_shared(nullptr, + height, + width, + // grad share index with value + indexDict); + mats_[pType] = mat; + } else if (matType == MAT_CACHE_ROW) { + CHECK(isGradSparseUpdate()); + auto mat = std::make_shared(height, width); + mats_[pType] = mat; + } else if (matType == MAT_SPARSE_ROW_PREFETCH_FULL_SIZE || + matType == MAT_SPARSE_ROW_PREFETCH) { + auto mat = std::make_shared( + bufs_[pType] ? std::dynamic_pointer_cast( + bufs_[pType]->getMemoryHandle()) + : nullptr, + height, + width, + nullptr, // indexDictHandle + getGlobalSyncThreadPool()); + mats_[pType] = mat; + } else if (matType == MAT_SPARSE_ROW_AUTO_GROW) { + CHECK(isGradSparseUpdate()); + mats_[pType] = std::make_shared(height, width); + } +#endif + // NOLINTNEXTLINE + else { + LOG(FATAL) << "Unsupported mat type" << matType; + } +} + +void Parameter::incUpdate(const UpdateCallback& callback) { + // Static parameter is fixed, and does not need to be updated + if (isStatic()) { + return; + } + + ++updateCounter_; + if (isUpdatable()) { + if (callback) callback(this); + clearUpdate(); + } +} + +bool Parameter::save(const std::string& filename) const { + std::ofstream fs(filename, std::ios_base::binary); + CHECK(fs) << "Fail to open " << filename; + return save(fs); +} + +bool Parameter::save(std::ostream& s) const { + CpuVector vec(*bufs_[PARAMETER_VALUE].get()); + Header header; + header.format = headerFormat_; + header.valueSize = sizeof(real); + header.size = getSize(); + + CHECK_EQ(header.size, vec.getSize()); + + CHECK(s.write(reinterpret_cast(&header), sizeof(header))) + << "Fail to write parameter " << getName(); + + CHECK(s.write(reinterpret_cast(vec.getData()), + header.size * sizeof(real))) + << "Fail to write parameter " << getName(); + if (config_.is_sparse()) { + CpuIVector rows(*intBufs_[PARAMETER_ROWS].get()); + CpuIVector cols(*intBufs_[PARAMETER_COLS].get()); + CHECK(s.write(reinterpret_cast(rows.getData()), + rows.getSize() * sizeof(int))) + << "Fail to write parameter " << getName(); + CHECK(s.write(reinterpret_cast(cols.getData()), + cols.getSize() * sizeof(int))) + << "Fail to write parameter " << getName(); + } + + return true; +} + +/** + * Load parameter value from a file + */ +bool Parameter::load(const std::string& filename) { + std::ifstream fs(filename, std::ios_base::binary); + if (!fs) { + LOG(INFO) << "missing parameters [" << filename << "] while loading model."; + if (kMissParameterFail == FLAGS_load_missing_parameter_strategy) { + LOG(FATAL) << getName() << " missing, not allowed."; + return false; + } + if (kMissParameterRand == FLAGS_load_missing_parameter_strategy) { + LOG(INFO) << getName() << " missing, set to random."; + randomize(); + return true; + } + if (kMissParameterZero == FLAGS_load_missing_parameter_strategy) { + LOG(INFO) << getName() << " missing, set to zero."; + zeroMem(); + return true; + } + LOG(FATAL) << "unsupported load_missing_parameter_strategy: " + << FLAGS_load_missing_parameter_strategy; + return false; + } + return load(fs); +} + +bool Parameter::load(std::istream& s) { + CpuVector vec(*bufs_[PARAMETER_VALUE].get()); + Header header; + CHECK(s.read(reinterpret_cast(&header), sizeof(header))) + << "Fail to read parameter " << getName(); + CHECK(isHeaderFormatSupported(header.format)) << "Incorrect format version: " + << header.format; + headerFormat_ = header.format; + CHECK_EQ(header.size, getSize()) + << "The size (" << header.size << ") in the file does not match the size " + << "(" << getSize() << ") of the parameter: " << getName(); + CHECK_EQ(header.valueSize, sizeof(real)) + << "Unsupported valueSize " << header.valueSize << " at: " << getName(); + CHECK(s.read(reinterpret_cast(vec.getData()), + header.size * sizeof(real))); + + auto& tmp = *bufs_[PARAMETER_VALUE].get(); + if (typeid(tmp) == typeid(GpuVector)) { + bufs_[PARAMETER_VALUE]->copyFrom(vec); + } + + if (config_.is_sparse() && config_.need_compact()) { + // load from dense parameter with many zero + CHECK_EQ(config_.dims_size(), 2); + auto height = config_.dims(0); + auto width = config_.dims(1); + auto mat = Matrix::create(vec.getData(), height, width); + CpuSparseMatrix sparseMat(height, + width, + 0, + FLOAT_VALUE, + format_, + /*trans*/ false); + sparseMat.copyFrom(*mat, HPPL_STREAM_DEFAULT); + auto nnz = sparseMat.getElementCnt(); + size_t rowSize = (format_ == SPARSE_CSR) ? height + 1 : nnz; + size_t colSize = (format_ == SPARSE_CSR) ? nnz : width + 1; + + intBufs_[PARAMETER_ROWS]->copyFrom(sparseMat.getRows(), rowSize); + intBufs_[PARAMETER_COLS]->copyFrom(sparseMat.getCols(), colSize); + bufs_[PARAMETER_VALUE]->resize(nnz); // for setMat check + bufs_[PARAMETER_VALUE]->copyFrom(sparseMat.getValue(), nnz); + config_.set_size(nnz); + LOG(INFO) << "compact nnz=" << (1. * nnz / (height * width)) + << " name=" << config_.name(); + } else if (config_.is_sparse()) { + CpuIVector rows(*intBufs_[PARAMETER_ROWS].get()); + CpuIVector cols(*intBufs_[PARAMETER_COLS].get()); + size_t rowSize, colSize; + CHECK_EQ(config_.dims_size(), 2); + if (format_ == SPARSE_CSR) { + rowSize = config_.dims(0) + 1; + colSize = config_.size(); + } else { + rowSize = config_.size(); + colSize = config_.dims(1) + 1; + } + CHECK( + s.read(reinterpret_cast(rows.getData()), rowSize * sizeof(int))); + CHECK( + s.read(reinterpret_cast(cols.getData()), colSize * sizeof(int))); + auto& paramRows = *intBufs_[PARAMETER_ROWS].get(); + if (typeid(paramRows) == typeid(GpuIVector)) { + intBufs_[PARAMETER_ROWS]->copyFrom(rows); + } + auto& paramCols = *intBufs_[PARAMETER_COLS].get(); + if (typeid(paramCols) == typeid(GpuIVector)) { + intBufs_[PARAMETER_COLS]->copyFrom(cols); + } + } + + setValueUpdated(); + + return true; +} + +} // namespace paddle diff --git a/paddle/legacy/parameter/Parameter.h b/paddle/legacy/parameter/Parameter.h new file mode 100644 index 0000000000000000000000000000000000000000..43b567dad045ad786b1b3f2d3614072f58310527 --- /dev/null +++ b/paddle/legacy/parameter/Parameter.h @@ -0,0 +1,380 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include +#include +#include + +#include "ParameterConfig.pb.h" +#include "TrainerConfig.pb.h" + +#include "ParameterUpdaterHook.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/ThreadLocal.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +typedef enum { + /// The paddle original basic format + PARAM_FORMAT_ORIGINAL = 0, + + /// See mkldnn_memory_format_t in + /// https://github.com/01org/mkl-dnn/blob/master/include/mkldnn_types.h + /// for a detailed description. + /// 2D weights tensor in the format (output channels, input channels). + PARAM_FORMAT_MKLDNN_OI, + + /// The total format items numbers + PARAM_FORMAT_ITEMS, +} PARAM_FORMAT; + +class SparsePrefetchRowCpuMatrix; + +class Parameter; +typedef std::function UpdateCallback; +typedef std::function ParamInitCallback; + +class Parameter; +typedef std::shared_ptr ParameterPtr; + +class Parameter { + public: + Parameter(const ParameterConfig& config, bool useGpu, bool doInit = true); + const std::string& getName() const { return config_.name(); } + + size_t getSize() const { return config_.size(); } + + bool isFullSize() const { + if (bufs_[PARAMETER_VALUE]) { + return this->getSize() == bufs_[PARAMETER_VALUE]->getSize(); + } + return false; + } + + inline bool useGpu() const { return useGpu_; } + + int getDeviceId() const { return deviceId_; } + + void setDevice(int deviceId) { deviceId_ = deviceId; } + + /// The id ranges from 0 to the_total_number_of_parameters - 1 + size_t getID() const { return config_.para_id(); } + + /// ID is a implict value created until neural network is built. + void setID(size_t id) { config_.set_para_id(id); } + + bool isStatic() const { return config_.is_static(); } + + enum MatType { + MAT_NORMAL, + /// both value and grad are shared + MAT_NORMAL_SHARED, + + /// Now used in BatchNorm in CPU mode + MAT_VALUE_SHARED, + + /// sparse matrix, which has full size parameter + MAT_SPARSE_ROW_IDS, + /// sparse matrix, parameter size scale by sparse rates. + MAT_SPARSE_ROW_AUTO_GROW, + MAT_CACHE_ROW, + MAT_SPARSE_ROW, + + /// sparse matrix for prefetching parameter from pserver + MAT_SPARSE_ROW_PREFETCH, + /// same as above, but parameter has full size for saving parameter in local + MAT_SPARSE_ROW_PREFETCH_FULL_SIZE, + }; + + void enableSparseParameter() { + if (config_.is_sparse()) { + if (config_.format() == "csr") { + size_t height = config_.dims(0); + size_t nnz = config_.size(); + enableIntType(PARAMETER_ROWS, height + 1); + enableIntType(PARAMETER_COLS, nnz); + format_ = SPARSE_CSR; + } else { + size_t width = config_.dims(1); + size_t nnz = config_.size(); + enableIntType(PARAMETER_COLS, width + 1); + enableIntType(PARAMETER_ROWS, nnz); + format_ = SPARSE_CSC; + } + } + } + + /// allocate buffer for the give type + void enableType(ParameterType type, MatType matType = MAT_NORMAL) { + if (bufs_[type] || mats_[type]) { + return; + } + SetDevice device(deviceId_); + if (config_.dims_size() == 2) { + if (matType == MAT_NORMAL || matType == MAT_NORMAL_SHARED || + matType == MAT_SPARSE_ROW_PREFETCH_FULL_SIZE || + matType == MAT_VALUE_SHARED || matType == MAT_SPARSE_ROW_IDS) { + bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_); + bufs_[type]->zeroMem(); + } else { + CHECK(isGradSparseUpdate()); + } + if (config_.is_sparse() && type == PARAMETER_VALUE) { + enableSparseParameter(); + } + setMat(type, matType); + } else { + bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_); + bufs_[type]->zeroMem(); + } + } + + void enableBufType(ParameterType type) { + if (bufs_[type]) return; + bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_); + bufs_[type]->zeroMem(); + } + + void enableIntType(ParameterType type, size_t intStoreSize = 0) { + if (!intBufs_[type]) { + SetDevice device(deviceId_); + size_t size = intStoreSize ? intStoreSize : config_.size(); + intBufs_[type] = IVector::create(size, useGpu_); + intBufs_[type]->zeroMem(); + } + } + + void enableSharedType(ParameterType type, + VectorPtr vec, + MatrixPtr mat = nullptr) { + if (!bufs_[type] && !mats_[type]) { + bufs_[type] = vec; + mats_[type] = mat; + } + } + + /// for batchGradientMachine: blockNum is number of partitions of the matrix. + bool isGradShared(size_t* blockNum = NULL); + + bool isValueShared(); + + // for AsgdSparseGradientMachine & SgdSparseGradientMachine: + // and MultiGradientMachine + bool isGradSparseUpdate() const; + + bool isSparseRemoteUpdate() const { + return config_.sparse_remote_update() && !useGpu(); + } + + const ParameterConfig& getConfig() const { return config_; } + + ParameterConfig& getConfig() { return config_; } + + bool hasType(ParameterType pType) const { + return bufs_[pType] || mats_[pType]; + } + + const VectorPtr& getBuf(ParameterType pType) const { + return this->bufs_[pType]; + } + + const VectorPtr* getBufs() const { return bufs_; } + + const MatrixPtr& getMat(ParameterType pType) const { return mats_[pType]; } + + void setValueUpdated() { updated_ = true; } + + void clearValueUpdated() { updated_ = false; } + + bool isValueUpdated() const { return updated_; } + + /** + * Save parameter value to a file + */ + bool save(const std::string& filename) const; + + /** + * Save parameter to ostream + */ + bool save(std::ostream& s) const; + + /** + * Load parameter value from a file + */ + bool load(const std::string& filename); + + /** + * Load parameter from istream + */ + bool load(std::istream& is); + + void incShared() { sharedCount_++; } + + /** + * After one of the parameter's gradient is merged + * You should call this function to do some additional processing, + */ + void incUpdate(const UpdateCallback& callbacks = NULL); + + void clearGradient() { + auto& mat = getMat(PARAMETER_GRADIENT); + if (mat) { + // zeroMem will also clear rows for SparseRowCpuMatrix + mat->zeroMem(); + } else { + auto& gradBuf = getBuf(PARAMETER_GRADIENT); + if (gradBuf) gradBuf->zeroMem(); + } + } + + void initialize(); + + /** + * Initialize the value according to config_: initial_mean, + * initial_std and initial_strategy. + */ + void randomize(); + static void randomize(const VectorPtr& value, const ParameterConfig& config); + + /// Initialize the value to 0 + void zeroMem(); + + /// file header structure + struct Header { + int32_t format; // = PARAM_FORMAT + uint32_t valueSize; // = sizeof(real) + uint64_t size; // = getSize() + }; + + /** + * @brief Is the header format supported. + */ + static bool isHeaderFormatSupported(int32_t fmt) { + return fmt < PARAM_FORMAT_ITEMS; + } + + /** + * @brief Get the format in header. + */ + int getHeaderFormat() { return headerFormat_; } + + /** + * @brief Set the format in header. + */ + void setHeaderFormat(int32_t fmt) { + CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: " + << fmt; + headerFormat_ = fmt; + } + + /** + * @brief Parameter Update Hook. + * + * The parameter's update hook before ParameterUpdater::updateImpl + * It could modify gradient/momentum/etc here. Such as drop some gradient, + * etc. + */ + void updateHook() { + for (auto& hook : updaterHooks_) { + hook->update(this); + } + } + + /** + * @brief Initialize all updater hook. + * + * This method should be invoked in ParameterUpdater::init() only. + */ + void initHook() { + for (auto& hook : updaterHooks_) { + hook->init(this); + } + } + + protected: + /** + * @brief create matrix to matType. + * + * used by gradient machine which needs specify matrix type, + * instead of creating in weights.cpp. + * + * @note pType should be enabled already. + */ + void setMat(ParameterType pType, int matType); + + bool isUpdatable() { return (updateCounter_ == sharedCount_); } + + void clearUpdate() { updateCounter_ = 0; } + + protected: + ParameterConfig config_; + + bool useGpu_; + + int deviceId_; + + /** + * @brief bufs_ stores parameter value and gradient. + * + * Layer should use bufs_[PARAMETER_VALUE] to form weight matrix for + * calculation and stores gradient to bufs_[PARAMETER_GRADIENT]. + */ + VectorPtr bufs_[NUM_PARAMETER_TYPES]; + + /** + * @brief Weight matrix for bufs_. + * + * It's helpfull when parameter shared by multi-layers. + * Caller should check, if mats exist, do not create it again. + */ + MatrixPtr mats_[NUM_PARAMETER_TYPES]; + + /// Int vectors, used in some User defined parameter types + IVectorPtr intBufs_[NUM_PARAMETER_TYPES]; + + int sharedCount_; + int updateCounter_; + + bool updated_; + SparseFormat format_; + + /// The header format for saving or loading param + int32_t headerFormat_; + + std::vector> updaterHooks_; + + public: + void setSharedCount(int cnt) { sharedCount_ = cnt; } + int getSharedCount() { return sharedCount_; } + + bool isSparse() { return config_.is_sparse(); } + SparseFormat getFormat() { return format_; } + + static const std::string kMissParameterFail; + static const std::string kMissParameterRand; + static const std::string kMissParameterZero; +}; + +typedef std::map ParameterMap; + +} // namespace paddle diff --git a/paddle/legacy/parameter/ParameterOptimizer.cpp b/paddle/legacy/parameter/ParameterOptimizer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b9dffa5afb4c99314869c7ed547ea9711d718b6e --- /dev/null +++ b/paddle/legacy/parameter/ParameterOptimizer.cpp @@ -0,0 +1,63 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Logging.h" + +#include + +#include "AverageOptimizer.h" +#include "FirstOrderOptimizer.h" +#include "OptimizerFunctions.h" +#include "OptimizerWithRegularizer.h" +#include "ParameterOptimizer.h" +#include "hl_gpu.h" + +namespace paddle { + +ParameterOptimizer* ParameterOptimizer::create( + const OptimizationConfig& optConfig, bool inPserver) { + if (inPserver && optConfig.num_batches_per_send_parameter() > 1) { + return new AddOptimizer(optConfig); + } + if (optConfig.learning_method() == "momentum") { + return new SgdOptimizer(optConfig); + } + if (optConfig.learning_method() == "torch_momentum") { + return new SgdOptimizer(optConfig); + } + if (optConfig.learning_method() == "adagrad") { + return new AdagradParameterOptimizer(optConfig); + } + if (optConfig.learning_method() == "adadelta") { + return new AdaDeltaParameterOptimizer(optConfig); + } + if (optConfig.learning_method() == "rmsprop") { + return new RMSPropParameterOptimizer(optConfig); + } + if (optConfig.learning_method() == "decayed_adagrad") { + return new DecayedAdagradParameterOptimizer(optConfig); + } + if (optConfig.learning_method() == "adam") { + return new AdamParameterOptimizer(optConfig); + } + if (optConfig.learning_method() == "adamax") { + return new AdamaxParameterOptimizer(optConfig); + } + if (optConfig.learning_method() == "sparse_momentum") { + return new SparseMomentumParameterOptimizer(optConfig); + } + return nullptr; +} + +} // namespace paddle diff --git a/paddle/parameter/ParameterOptimizer.h b/paddle/legacy/parameter/ParameterOptimizer.h similarity index 100% rename from paddle/parameter/ParameterOptimizer.h rename to paddle/legacy/parameter/ParameterOptimizer.h diff --git a/paddle/legacy/parameter/ParameterUpdateFunctions.cpp b/paddle/legacy/parameter/ParameterUpdateFunctions.cpp new file mode 100644 index 0000000000000000000000000000000000000000..72c9841acf6d3eb1d28d631e1599a1a403175013 --- /dev/null +++ b/paddle/legacy/parameter/ParameterUpdateFunctions.cpp @@ -0,0 +1,300 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Logging.h" +#ifdef __AVX__ +#include +#include +#endif + +#include "ParameterUpdateFunctions.h" + +namespace paddle { + +void sgdUpdateCpu(real learningRate, + real momentum, + real decayRate, + size_t size, + real* value, + const real* grad, + real* momentumVec) { + decayRate *= learningRate; +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for +#endif + for (size_t i = 0; i < size; ++i) { + momentumVec[i] = momentum * momentumVec[i] - learningRate * grad[i] - + decayRate * value[i]; + value[i] += momentumVec[i]; + } +} + +void sgdUpdate(real learningRate, + real momentum, + real decayRate, + Vector* value, + Vector* grad, + Vector* momentumVec) { + size_t size = value->getSize(); + real* val = value->getData(); + real* grd = grad->getData(); + real* mom = momentumVec->getData(); + if (typeid(*value) == typeid(CpuVector)) { + sgdUpdateCpu(learningRate, momentum, decayRate, size, val, grd, mom); + } else if (typeid(*value) == typeid(GpuVector)) { + value->sgdUpdate(*grad, *momentumVec, learningRate, momentum, decayRate); + } else { + LOG(FATAL) << "Wrong"; + } +} + +void sgdUpdateAvx(float learningRate, + float momentum, + float decayRate, + size_t size, + float* value, + const float* _grad, + float* momentumVec) { +#ifdef __AVX__ + float* grad = const_cast(_grad); // the gradient is not modified + // but when invoke simd functions + // need non-const pointer. + size_t gradientAlign = 0; + size_t gradientAlignHeader = (size_t)grad % sizeof(__m256); + CHECK_EQ(gradientAlignHeader, (size_t)momentumVec % sizeof(__m256)) + << "Gradent buffer didn't align with momentum buffer"; + CHECK_EQ(gradientAlignHeader, (size_t)value % sizeof(__m256)) + << "Gradent buffer didn't align with value buffer"; + if (0 != gradientAlignHeader) { + gradientAlignHeader = sizeof(__m256) - gradientAlignHeader; + gradientAlign = gradientAlignHeader / sizeof(real); + + // handle the unalign buffer + for (size_t i = 0; i < gradientAlign; i++) { + momentumVec[i] = momentum * momentumVec[i] - (learningRate * grad[i]) - + (decayRate * learningRate * value[i]); + value[i] += momentumVec[i]; + } + grad += gradientAlign; + momentumVec += gradientAlign; + value += gradientAlign; + } + + constexpr size_t kParallelNum = 8; + constexpr size_t nStepSize = (sizeof(__m256) / sizeof(real)) * kParallelNum; + size_t cntLoop = (size - gradientAlign) / nStepSize; + size_t cntRem = (size - gradientAlign) % nStepSize; + __m256 gradientTmp[kParallelNum]; + __m256 valueTmp[kParallelNum]; + __m256 lr, mom, dr; + std::function loopFun; + + learningRate *= -1; + lr = _mm256_set_ps(learningRate, + learningRate, + learningRate, + learningRate, + learningRate, + learningRate, + learningRate, + learningRate); + + if (0 != momentum) { + mom = _mm256_set_ps(momentum, + momentum, + momentum, + momentum, + momentum, + momentum, + momentum, + momentum); + } + + decayRate *= learningRate; + if (0 != decayRate) { + dr = _mm256_set_ps(decayRate, + decayRate, + decayRate, + decayRate, + decayRate, + decayRate, + decayRate, + decayRate); + } + + auto gradMulFun = [&](void) { + gradientTmp[0] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad), lr); + gradientTmp[1] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 8), lr); + gradientTmp[2] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 16), lr); + gradientTmp[3] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 24), lr); + gradientTmp[4] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 32), lr); + gradientTmp[5] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 40), lr); + gradientTmp[6] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 48), lr); + gradientTmp[7] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 56), lr); + }; + + auto valueMulFun = [&](void) { + valueTmp[0] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value), dr); + valueTmp[1] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 8), dr); + valueTmp[2] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 16), dr); + valueTmp[3] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 24), dr); + valueTmp[4] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 32), dr); + valueTmp[5] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 40), dr); + valueTmp[6] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 48), dr); + valueTmp[7] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 56), dr); + }; + + auto momentumMulFun = [&](void) { + *reinterpret_cast<__m256*>(momentumVec) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec), mom); + *reinterpret_cast<__m256*>(momentumVec + 8) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 8), mom); + *reinterpret_cast<__m256*>(momentumVec + 16) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 16), mom); + *reinterpret_cast<__m256*>(momentumVec + 24) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 24), mom); + *reinterpret_cast<__m256*>(momentumVec + 32) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 32), mom); + *reinterpret_cast<__m256*>(momentumVec + 40) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 40), mom); + *reinterpret_cast<__m256*>(momentumVec + 48) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 48), mom); + *reinterpret_cast<__m256*>(momentumVec + 56) = + _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 56), mom); + }; + + auto momentumAddGradFun = [&](void) { + *reinterpret_cast<__m256*>(momentumVec) = + _mm256_add_ps(*reinterpret_cast<__m256*>(momentumVec), gradientTmp[0]); + *reinterpret_cast<__m256*>(momentumVec + 8) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 8), gradientTmp[1]); + *reinterpret_cast<__m256*>(momentumVec + 16) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 16), gradientTmp[2]); + *reinterpret_cast<__m256*>(momentumVec + 24) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 24), gradientTmp[3]); + *reinterpret_cast<__m256*>(momentumVec + 32) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 32), gradientTmp[4]); + *reinterpret_cast<__m256*>(momentumVec + 40) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 40), gradientTmp[5]); + *reinterpret_cast<__m256*>(momentumVec + 48) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 48), gradientTmp[6]); + *reinterpret_cast<__m256*>(momentumVec + 56) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 56), gradientTmp[7]); + }; + + auto momentumZeroFun = [&](void) { + *reinterpret_cast<__m256*>(momentumVec) = gradientTmp[0]; + *reinterpret_cast<__m256*>(momentumVec + 8) = gradientTmp[1]; + *reinterpret_cast<__m256*>(momentumVec + 16) = gradientTmp[2]; + *reinterpret_cast<__m256*>(momentumVec + 24) = gradientTmp[3]; + *reinterpret_cast<__m256*>(momentumVec + 32) = gradientTmp[4]; + *reinterpret_cast<__m256*>(momentumVec + 40) = gradientTmp[5]; + *reinterpret_cast<__m256*>(momentumVec + 48) = gradientTmp[6]; + *reinterpret_cast<__m256*>(momentumVec + 56) = gradientTmp[7]; + }; + + auto momentumAddValueFun = [&](void) { + *reinterpret_cast<__m256*>(momentumVec) = + _mm256_add_ps(*reinterpret_cast<__m256*>(momentumVec), valueTmp[0]); + *reinterpret_cast<__m256*>(momentumVec + 8) = + _mm256_add_ps(*reinterpret_cast<__m256*>(momentumVec + 8), valueTmp[1]); + *reinterpret_cast<__m256*>(momentumVec + 16) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 16), valueTmp[2]); + *reinterpret_cast<__m256*>(momentumVec + 24) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 24), valueTmp[3]); + *reinterpret_cast<__m256*>(momentumVec + 32) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 32), valueTmp[4]); + *reinterpret_cast<__m256*>(momentumVec + 40) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 40), valueTmp[5]); + *reinterpret_cast<__m256*>(momentumVec + 48) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 48), valueTmp[6]); + *reinterpret_cast<__m256*>(momentumVec + 56) = _mm256_add_ps( + *reinterpret_cast<__m256*>(momentumVec + 56), valueTmp[7]); + }; + + auto valueAddMomentumFun = [&](void) { + *reinterpret_cast<__m256*>(value) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value), + *reinterpret_cast<__m256*>(momentumVec)); + *reinterpret_cast<__m256*>(value + 8) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 8), + *reinterpret_cast<__m256*>(momentumVec + 8)); + *reinterpret_cast<__m256*>(value + 16) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 16), + *reinterpret_cast<__m256*>(momentumVec + 16)); + *reinterpret_cast<__m256*>(value + 24) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 24), + *reinterpret_cast<__m256*>(momentumVec + 24)); + *reinterpret_cast<__m256*>(value + 32) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 32), + *reinterpret_cast<__m256*>(momentumVec + 32)); + *reinterpret_cast<__m256*>(value + 40) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 40), + *reinterpret_cast<__m256*>(momentumVec + 40)); + *reinterpret_cast<__m256*>(value + 48) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 48), + *reinterpret_cast<__m256*>(momentumVec + 48)); + *reinterpret_cast<__m256*>(value + 56) = + _mm256_add_ps(*reinterpret_cast<__m256*>(value + 56), + *reinterpret_cast<__m256*>(momentumVec + 56)); + }; + + if (0 == decayRate && 0 == momentum) { + loopFun = [&](void) { + gradMulFun(); + momentumZeroFun(); + valueAddMomentumFun(); + }; + } else if (0 == decayRate && 0 != momentum) { + loopFun = [&](void) { + gradMulFun(); + momentumMulFun(); + momentumAddGradFun(); + valueAddMomentumFun(); + }; + } else if (0 != decayRate && 0 == momentum) { + loopFun = [&](void) { + gradMulFun(); + valueMulFun(); + momentumZeroFun(); + momentumAddValueFun(); + valueAddMomentumFun(); + }; + } else if (0 != decayRate && 0 != momentum) { + loopFun = [&](void) { + gradMulFun(); + valueMulFun(); + momentumMulFun(); + momentumAddGradFun(); + momentumAddValueFun(); + valueAddMomentumFun(); + }; + } + + for (size_t i = 0; i < cntLoop; i++) { + loopFun(); + grad += nStepSize; + momentumVec += nStepSize; + value += nStepSize; + } + + for (size_t i = 0; i < cntRem; i++) { + momentumVec[i] = momentum * momentumVec[i] + (learningRate * grad[i]) + + (decayRate * value[i]); + value[i] += momentumVec[i]; + } +#endif +} + +} // namespace paddle diff --git a/paddle/legacy/parameter/ParameterUpdateFunctions.h b/paddle/legacy/parameter/ParameterUpdateFunctions.h new file mode 100644 index 0000000000000000000000000000000000000000..a7cc1c4c47b6c8723520221cb0efc2afb53a900c --- /dev/null +++ b/paddle/legacy/parameter/ParameterUpdateFunctions.h @@ -0,0 +1,56 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Common.h" + +namespace paddle { + +/** + * Performs the following operations. + * + * momentumVec = momentum * momentumVec + * - learningRate * grad + * - learningRate * decayRate * value + * + * value = value + momentumVec + * momentum = 0 or decayRate = 0 are specially handled to avoid unnecessary + * computation. + */ +void sgdUpdate(real learningRate, + real momentum, + real decayRate, + Vector* value, + Vector* grad, + Vector* momentumVec); + +void sgdUpdateCpu(real learningRate, + real momentum, + real decayRate, + size_t size, + real* value, + const real* grad, + real* momentumVec); + +void sgdUpdateAvx(float learningRate, + float momentum, + float decayRate, + size_t size, + float* value, + const float* grad, + float* momentumVec); + +} // namespace paddle diff --git a/paddle/legacy/parameter/ParameterUpdaterBase.cpp b/paddle/legacy/parameter/ParameterUpdaterBase.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7d9d3fad63160b76d6de0932f39596a8643d0a8e --- /dev/null +++ b/paddle/legacy/parameter/ParameterUpdaterBase.cpp @@ -0,0 +1,41 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ParameterUpdaterBase.h" +#include +#include "hl_gpu.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +void ParameterUpdater::init(const std::vector& parameters) { + parameters_ = parameters; + for (ParameterType type : getParameterTypes()) { + for (auto& para : parameters) { + para->enableType(type); + } + } + for (size_t pid = 0; pid < parameters_.size(); ++pid) { + nonStaticParaIDMap_.insert( + std::pair(parameters_[pid]->getID(), pid)); + } + + for (auto& para : parameters) { + if (!para->isStatic()) { + para->initHook(); + } + } +} + +} // namespace paddle diff --git a/paddle/parameter/ParameterUpdaterBase.h b/paddle/legacy/parameter/ParameterUpdaterBase.h similarity index 100% rename from paddle/parameter/ParameterUpdaterBase.h rename to paddle/legacy/parameter/ParameterUpdaterBase.h diff --git a/paddle/legacy/parameter/ParameterUpdaterHook.cpp b/paddle/legacy/parameter/ParameterUpdaterHook.cpp new file mode 100644 index 0000000000000000000000000000000000000000..bfb9769fb67fc71b6f96f09d44b2c108745eafa3 --- /dev/null +++ b/paddle/legacy/parameter/ParameterUpdaterHook.cpp @@ -0,0 +1,155 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ParameterUpdaterHook.h" + +#include +#include +#include +#include +#include +#include +#include + +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +/** + * The static pruning hook + * Static means user specify a sparsity_ratio before training started, and the + * network will prune the parameters based on the sparsity_ratio. More details + * can be found https://arxiv.org/pdf/1506.02626.pdf. + */ + +class StaticPruningHook : public IParameterUpdaterHook { + public: + explicit StaticPruningHook(const ParameterUpdaterHookConfig &hookConfig) + : initCount_(0) { + sparsityRatio_ = hookConfig.sparsity_ratio(); + } + + static bool sortPairAscend(const std::pair &pair1, + const std::pair &pair2) { + return pair1.first > pair2.first; + } + + void update(Parameter *para) { + updateThreadChecker_.check(); + auto &vec = para->getBuf(PARAMETER_GRADIENT); + if (vec) { + vec->dotMul(*maskVec_); + } + } + + void generateMask(Parameter *para) { + VectorPtr maskTemp = Vector::create(para->getSize(), false); + maskTemp->zeroMem(); + real *maskTempData = maskTemp->getData(); + size_t nonZeroNum = para->getSize() * (1 - sparsityRatio_); + + VectorPtr paraVec = para->getBuf(PARAMETER_VALUE); + VectorPtr paraCpuCopy = Vector::create(para->getSize(), false); + + paraCpuCopy->copyFrom(*paraVec); + std::vector> param; + + for (size_t i = 0; i < para->getSize(); i++) + param.push_back(std::make_pair(fabs(paraCpuCopy->getData()[i]), i)); + + std::partial_sort( + param.begin(), param.begin() + nonZeroNum, param.end(), sortPairAscend); + for (size_t i = 0; i < nonZeroNum; i++) maskTempData[param[i].second] = 1.0; + + // Currently just use a mask vector for hack. + if (para->useGpu()) { + maskVec_ = Vector::create(para->getSize(), para->useGpu()); + maskVec_->copyFrom(*maskTemp); + } else { + maskVec_ = maskTemp; + } + } + + void init(Parameter *para) { + generateMask(para); + size_t initCount = this->initCount_.fetch_add(1); + CHECK_EQ(initCount, 0UL) << "Currently the StaticPruningHook must invoke " + "in same ParamterUpdater"; + VLOG(3) << "Initialize Parameter " << para; + SetDevice device(para->getDeviceId()); + + auto ¶Vec = para->getBuf(PARAMETER_VALUE); + paraVec->dotMul(*maskVec_); + } + + private: + SameThreadChecker updateThreadChecker_; + std::atomic initCount_; + VectorPtr maskVec_; + real sparsityRatio_; +}; + +IParameterUpdaterHook::IParameterUpdaterHook() {} + +IParameterUpdaterHook::~IParameterUpdaterHook() {} + +/** + * A Hasher used by g_hooks. + * + * Use the independent hasher intendedly. There is a hasher in PServer for hash + * ParameterBlock. But not to use same hasher to reduce dependency. + * + * May be extracted to Util.h to unify the hasher. + */ +class StringIntPairHasher { + public: + size_t operator()(const std::pair &k) const { + return intHasher_(strHasher_(k.first) + k.second); + } + + private: + std::hash strHasher_; + std::hash intHasher_; +}; + +static WeakKVCache, + IParameterUpdaterHook, + StringIntPairHasher> + g_hookCache_; + +/** + * ParameterUpdaterHook actually factory method. + */ +static IParameterUpdaterHook *createImpl( + const ParameterUpdaterHookConfig &config) { + auto &type = config.type(); + if (type == "pruning") { + return new StaticPruningHook(config); + } + + LOG(FATAL) << "Unknown Hook type: " << type; + return nullptr; +} + +std::shared_ptr IParameterUpdaterHook::create( + const ParameterConfig ¶mConfig, int idx) { + std::pair key = {paramConfig.name(), idx}; + return g_hookCache_.get( + key, [&] { return createImpl(paramConfig.update_hooks(idx)); }); +} + +} // namespace paddle diff --git a/paddle/parameter/ParameterUpdaterHook.h b/paddle/legacy/parameter/ParameterUpdaterHook.h similarity index 100% rename from paddle/parameter/ParameterUpdaterHook.h rename to paddle/legacy/parameter/ParameterUpdaterHook.h diff --git a/paddle/legacy/parameter/Regularizer.cpp b/paddle/legacy/parameter/Regularizer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c1d5f4fa68403408bb44341e1e28f2ce3beb2e4c --- /dev/null +++ b/paddle/legacy/parameter/Regularizer.cpp @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Regularizer.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +Regularizer* Regularizer::get(const std::vector& types, + const ParameterConfig& paraConfig) { + bool useLearningRateVec = + std::find(types.begin(), types.end(), PARAMETER_LEARNING_RATE) != + types.end(); + if (paraConfig.decay_rate_l1() > 0.0f && + paraConfig.decay_rate() > 0.0f) { // use L1 and L2 + if (useLearningRateVec) { + static L1L2LrRegularizer regularizer_; + return ®ularizer_; + } + static L1L2Regularizer regularizer_; + return ®ularizer_; + } + if (paraConfig.decay_rate_l1() > 0.0f) { // use L1 only + if (useLearningRateVec) { + static L1LrRegularizer regularizer_; + return ®ularizer_; + } + static L1Regularizer regularizer_; + return ®ularizer_; + } + if (paraConfig.decay_rate() > 0.0f) { // use L2 only + if (useLearningRateVec) { + static L2LrRegularizer regularizer_; + return ®ularizer_; + } + static L2Regularizer regularizer_; + return ®ularizer_; + } + return nullptr; +} + +} // namespace paddle diff --git a/paddle/parameter/Regularizer.h b/paddle/legacy/parameter/Regularizer.h similarity index 100% rename from paddle/parameter/Regularizer.h rename to paddle/legacy/parameter/Regularizer.h diff --git a/paddle/parameter/ThreadLocalBuffer.cpp b/paddle/legacy/parameter/ThreadLocalBuffer.cpp similarity index 100% rename from paddle/parameter/ThreadLocalBuffer.cpp rename to paddle/legacy/parameter/ThreadLocalBuffer.cpp diff --git a/paddle/legacy/parameter/ThreadLocalBuffer.h b/paddle/legacy/parameter/ThreadLocalBuffer.h new file mode 100644 index 0000000000000000000000000000000000000000..d360feeed6c98ee60e3bdae924434054080576b0 --- /dev/null +++ b/paddle/legacy/parameter/ThreadLocalBuffer.h @@ -0,0 +1,22 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/legacy/math/Vector.h" + +namespace paddle { +namespace parameter { +extern VectorPtr* getThreadLocalBuffer(); +} // namespace parameter +} // namespace paddle diff --git a/paddle/legacy/parameter/Weight.cpp b/paddle/legacy/parameter/Weight.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9d94050a5cd8c3570c286e8e82c2a1470c40e6db --- /dev/null +++ b/paddle/legacy/parameter/Weight.cpp @@ -0,0 +1,84 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Weight.h" +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +Weight::Weight(size_t height, size_t width, ParameterPtr param) { + VectorPtr vPtr = param->getBuf(PARAMETER_VALUE); + VectorPtr gPtr = param->getBuf(PARAMETER_GRADIENT); + + // create a new weight + if (param->isSparse()) { + CHECK_LE(param->getSize(), width * height); + } else { + CHECK_EQ(param->getSize(), width * height); + } + + // weight_ + weight_ = param->getMat(PARAMETER_VALUE); + if (!weight_ && vPtr) { + weight_ = Matrix::create(vPtr->getMemoryHandle(), height, width); + } + if (weight_) { + CHECK_EQ(height, weight_->getHeight()); + CHECK_EQ(width, weight_->getWidth()); + } + + // weightGrad + weightGrad_ = param->getMat(PARAMETER_GRADIENT); + if (!weightGrad_ && gPtr) { + weightGrad_ = Matrix::create(gPtr->getMemoryHandle(), height, width); + } + if (weightGrad_) { + CHECK_EQ(height, weightGrad_->getHeight()); + CHECK_EQ(width, weightGrad_->getWidth()); + } + + parameter_ = param; +} + +Weight::Weight(size_t height, size_t width, ParameterPtr param, size_t offset) { + VectorPtr vPtr = param->getBuf(PARAMETER_VALUE); + VectorPtr gPtr = param->getBuf(PARAMETER_GRADIENT); + + // create a new weight + CHECK_LE(offset + width * height, param->getSize()); + + // weight_ + if (vPtr) { + weight_ = Matrix::create(vPtr->getData() + offset, + height, + width, + /* trans */ false, + param->useGpu()); + } + + // weightGrad + if (gPtr) { + weightGrad_ = Matrix::create(gPtr->getData() + offset, + height, + width, + /* trans */ false, + param->useGpu()); + } + + parameter_ = param; +} + +const ParameterPtr& Weight::getParameterPtr() { return parameter_; } +void Weight::setParameterPtr(ParameterPtr param) { parameter_ = param; } +} // namespace paddle diff --git a/paddle/legacy/parameter/Weight.h b/paddle/legacy/parameter/Weight.h new file mode 100644 index 0000000000000000000000000000000000000000..241c8d829cd0c7b57964324d3378bdfcf09e6a70 --- /dev/null +++ b/paddle/legacy/parameter/Weight.h @@ -0,0 +1,48 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include + +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/SparseRowMatrix.h" +#include "paddle/legacy/parameter/Parameter.h" + +namespace paddle { + +class Weight { + private: + MatrixPtr weight_; + MatrixPtr weightGrad_; + ParameterPtr parameter_; + + public: + Weight(size_t height, size_t width, ParameterPtr parameter); + Weight(size_t height, size_t width, ParameterPtr parameter, size_t offset); + + const MatrixPtr& getW() { return weight_; } + const MatrixPtr& getWGrad() { return weightGrad_; } + const ParameterPtr& getParameterPtr(); + + void incUpdate(const UpdateCallback& callback) { + getParameterPtr()->incUpdate(callback); + } + + void setParameterPtr(ParameterPtr param); +}; + +typedef std::vector> WeightList; + +} // namespace paddle diff --git a/paddle/parameter/tests/CMakeLists.txt b/paddle/legacy/parameter/tests/CMakeLists.txt similarity index 100% rename from paddle/parameter/tests/CMakeLists.txt rename to paddle/legacy/parameter/tests/CMakeLists.txt diff --git a/paddle/legacy/parameter/tests/test_argument.cpp b/paddle/legacy/parameter/tests/test_argument.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0c632e0cd10342431dfcada680a18d8f9eabeb9c --- /dev/null +++ b/paddle/legacy/parameter/tests/test_argument.cpp @@ -0,0 +1,57 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +using namespace paddle; // NOLINT + +TEST(Argument, poolSequenceWithStride) { + Argument input, output; + ICpuGpuVector::resizeOrCreate(input.sequenceStartPositions, 5, false); + int* inStart = input.sequenceStartPositions->getMutableData(false); + inStart[0] = 0; + inStart[1] = 9; + inStart[2] = 14; + inStart[3] = 17; + inStart[4] = 30; + + int strideResult[] = {0, 5, 9, 14, 17, 22, 27, 30}; + int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30}; + + for (auto reversed : {false, true}) { + ICpuGpuVectorPtr stridePositions; + output.poolSequenceWithStride( + input, 5 /* stride */, &stridePositions, reversed); + + const int* outStart = output.sequenceStartPositions->getData(false); + CHECK_EQ(outStart[0], 0); + CHECK_EQ(outStart[1], 2); + CHECK_EQ(outStart[2], 3); + CHECK_EQ(outStart[3], 4); + CHECK_EQ(outStart[4], 7); + + CHECK_EQ(stridePositions->getSize(), 8UL); + auto result = reversed ? strideResultReversed : strideResult; + for (int i = 0; i < 8; i++) { + CHECK_EQ(stridePositions->getData(false)[i], result[i]); + } + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/parameter/tests/test_common.cpp b/paddle/legacy/parameter/tests/test_common.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8de9d6da983553c0b9e574ac27ae8fca14bea5b7 --- /dev/null +++ b/paddle/legacy/parameter/tests/test_common.cpp @@ -0,0 +1,174 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include +#include +#include +#include +#include + +using namespace paddle; // NOLINT + +class CommonTest : public ::testing::Test { + protected: + CommonTest() : testStat_("test") {} + virtual ~CommonTest() {} + virtual void SetUp() { + const size_t buffSize[] = { + 100, 128, 500, 1024, 4096, 10240, 102400, 1000000}; + sizeVec_.resize(8); + memcpy(&sizeVec_[0], &buffSize[0], 8 * sizeof(size_t)); + valueUint_.resize(4); + valueUint_[0].first = 0.0; + valueUint_[0].second = 0.0; + valueUint_[1].first = 0.0; + valueUint_[1].second = 1.0; + valueUint_[2].first = 1.0; + valueUint_[2].second = 0.0; + valueUint_[3].first = 1.0; + valueUint_[3].second = 1.0; + learningRate_ = 1.0; + } + + void test_sgdUpadate(real* gradientBuffer, + real* valueBuffer, + real* momentumBuffer, + size_t size); + + virtual void TreaDown() { LOG(INFO) << "All Test Finished."; } + + protected: + std::vector> valueUint_; + std::vector sizeVec_; + real learningRate_; + StatSet testStat_; +}; + +void CommonTest::test_sgdUpadate(real* gradientBuffer, + real* valueBuffer, + real* momentumBuffer, + size_t size) { +// sgdUpdateAvx has no double version yet +#if defined(__AVX__) && !defined(PADDLE_TYPE_DOUBLE) + real valueSum1 = 0, valueSum2 = 0, momSum1 = 0, momSum2 = 0; + real* gradTmp = new real[size]; + real* valueTmp = new real[size]; + real* momentumTmp = new real[size]; + memcpy(gradTmp, gradientBuffer, size * sizeof(real)); + memcpy(valueTmp, valueBuffer, size * sizeof(real)); + memcpy(momentumTmp, momentumBuffer, size * sizeof(real)); + for (auto& arg : valueUint_) { + { + { + struct timeval t; + REGISTER_TIMER("gettimeofday", 0, testStat_); + gettimeofday(&t, NULL); + } + REGISTER_TIMER("avxTimer", 0); + sgdUpdateAvx(learningRate_, + arg.first, + arg.second, + size, + valueBuffer, + gradientBuffer, + momentumBuffer); + } + for (size_t i = 0; i < size; i++) { + valueSum1 += valueBuffer[i]; + momSum1 += momentumBuffer[i]; + // std::cout << "[" + // << valueBuffer[i] + // << "," << momentumBuffer[i] + // << "," << gradientBuffer[i] << "],"; + } + { + REGISTER_TIMER("cpuTimer", 0); + sgdUpdateCpu(learningRate_, + arg.first, + arg.second, + size, + valueTmp, + gradTmp, + momentumTmp); + } + for (size_t i = 0; i < size; i++) { + valueSum2 += valueTmp[i]; + momSum2 += momentumTmp[i]; + // std::cout << "[" + // << valueTmp[i] + // << "," << momentumTmp[i] + // << "," << gradTmp[i] << "],"; + } + + VLOG(3) << "valueSum1 = " << valueSum1 << " ; valueSum2 = " << valueSum2; + VLOG(3) << "momSum1 = " << momSum1 << " ; momSum2 = " << momSum2; + ASSERT_EQ(valueSum1, valueSum2); + ASSERT_EQ(momSum1, momSum2); + } + delete[] gradTmp; + delete[] valueTmp; + delete[] momentumTmp; +#endif +} + +TEST_F(CommonTest, sgdUpdate) { + const size_t alignHeader[] = {0, 2, 3, 5, 7, 8}; + for (auto& size : sizeVec_) { + real *gradientBuffer, *valueBuffer, *momentumBuffer; + CHECK_EQ(posix_memalign((void**)&gradientBuffer, 32, sizeof(real) * size), + 0); + CHECK_EQ(posix_memalign((void**)&valueBuffer, 32, sizeof(real) * size), 0); + CHECK_EQ(posix_memalign((void**)&momentumBuffer, 32, sizeof(real) * size), + 0); + + for (size_t i = 0; i < size; i++) { + gradientBuffer[i] = 1.0; + valueBuffer[i] = 2.0; + momentumBuffer[i] = 3.0; + } + for (int i = 0; i < 6; i++) { + LOG(INFO) << "----------------------" << size << ":" << alignHeader[i] + << "-------------------------"; + test_sgdUpadate(&gradientBuffer[alignHeader[i]], + &valueBuffer[alignHeader[i]], + &momentumBuffer[alignHeader[i]], + size - alignHeader[i]); + } + free(gradientBuffer); + free(valueBuffer); + free(momentumBuffer); + } + globalStat.printAllStatus(); + testStat_.printAllStatus(); +} + +TEST_F(CommonTest, syncThreadPool) { + SyncThreadPool pool(10); + + std::vector nums; + nums.resize(10); + + pool.exec([&](int tid, size_t numThreads) { nums[tid] = tid; }); + for (size_t i = 0; i < nums.size(); ++i) { + EXPECT_EQ((int)i, nums[i]); + } + + pool.exec([&](int tid, size_t numThreads) { nums[tid] -= tid; }); + for (size_t i = 0; i < nums.size(); ++i) { + EXPECT_EQ((int)0, nums[i]); + } +} diff --git a/paddle/legacy/pserver/BaseClient.cpp b/paddle/legacy/pserver/BaseClient.cpp new file mode 100644 index 0000000000000000000000000000000000000000..13bb8a1cc58580a8e0af31c23b420836c7422ad8 --- /dev/null +++ b/paddle/legacy/pserver/BaseClient.cpp @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "BaseClient.h" +#include +#include +#include +#include "paddle/legacy/utils/Stat.h" + +DECLARE_string(pservers); + +namespace paddle { + +BaseClient::BaseClient(bool separate, int numPorts) + : stopping_(false), numPorts_(numPorts), separateSendAndRecv_(separate) { + CHECK_GT(numPorts, 0); +} + +BaseClient::~BaseClient() {} + +void BaseClient::recvData() { recvSyncBarrier_->wait(); } + +void BaseClient::synchronize(SyncObject syncObjectId) { + SynchronizeRequest request; + request.set_sync_object_id(syncObjectId); + std::vector responses; + multiCall(__func__, request, &responses); +} + +void BaseClient::startThreads() { + if (!separateSendAndRecv_) { + return; + } + recvSyncBarrier_.reset(new ThreadBarrier(threadNum_ + 1)); + + sendThreads_.resize(threadNum_); + recvThreads_.resize(threadNum_); + sendJobQueue_.resize(threadNum_); + recvJobQueue_.resize(threadNum_); + + for (int i = 0; i < threadNum_; ++i) { + sendJobQueue_[i].reset(new SendQueue()); + recvJobQueue_[i].reset(new SendQueue()); + + sendThreads_[i].reset( + new std::thread([this](int id) { this->send(id); }, i)); + + recvThreads_[i].reset( + new std::thread([this](int id) { this->recv(id); }, i)); + } +} + +void BaseClient::finishThreads() { + if (!separateSendAndRecv_) { + return; + } + stopping_ = true; + for (int i = 0; i < threadNum_; i++) { + sendJobQueue_[i]->enqueue(nullptr); + } + for (auto& thread : sendThreads_) { + thread->join(); + } + for (auto& thread : recvThreads_) { + thread->join(); + } + stopping_ = false; +} +} // namespace paddle diff --git a/paddle/legacy/pserver/BaseClient.h b/paddle/legacy/pserver/BaseClient.h new file mode 100644 index 0000000000000000000000000000000000000000..66e8f39cd60998122bb8958b12b23ee7142be94d --- /dev/null +++ b/paddle/legacy/pserver/BaseClient.h @@ -0,0 +1,311 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ParameterService.pb.h" +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/pserver/ProtoServer.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Queue.h" + +namespace paddle { + +/** + * it manages all connections to pservers. + * it exists two modes to manage connections to all pservers. Firstly, one + * connection owns two threads that separately manage to send and receive + * data. Secondly, each thread uses one connection for all activation in it. + * the first solution arms with sendThreads_/recvThreads_ and sendJobQueue_/ + * recvJobQueue_. the second solution use some shared thread pool to manage + * connections. + */ +class BaseClient { + protected: + typedef std::unique_ptr ThreadPtr; + typedef std::vector> InputIovs; + typedef std::vector SendRequest; + typedef std::vector SendDataRequestVec; + + // TODO(yanfei): + // refine data structure to unify parameter and features communication + struct SendJob { + /// store parameters related blocks data + InputIovs parallelInputIovs; + /// store protobuf request + SendRequest parallelRequests; + /// store data, such as features for metric learning + SendDataRequestVec parallelDataRequests; + }; + + public: + explicit BaseClient(bool separate = false, int numPorts = FLAGS_ports_num); + + virtual ~BaseClient(); + + typedef std::shared_ptr SendJobPtr; + typedef Queue SendQueue; + + /// send data to server, support only synchronize + template + void putData(int clientId, + SendDataType type, + DataType* datas, + size_t size, + DataUpdateMode mode) { + synchronize(SYNC_DATA); + sendData(clientId, type, mode, datas, size); + recvData(); + synchronize(SYNC_DATA); + } + + template + void putOwnData(int clientId, + SendDataType type, + DataType* datas, + size_t size) { + putData(clientId, type, datas, size, DATA_UPDATE_MODE_SET_OWN); + } + + template + void getAllData(int clientId, + SendDataType type, + DataType* datas, + size_t size) { + sendData(clientId, + type, + DATA_UPDATE_MODE_GET_ALL, + reinterpret_cast(NULL), + 0); + recvData(); + size_t dataOffset = 0; + for (auto& recvMem : recvDataMems_) { + CHECK_LE(dataOffset, size); + size_t memSize = std::min(recvMem.get()->getSize(), + sizeof(DataType) * (size - dataOffset)); + CHECK_EQ(memSize % sizeof(DataType), size_t(0)); + memcpy(datas + dataOffset, recvMem.get()->getBuf(), memSize); + dataOffset += memSize / sizeof(DataType); + } + CHECK_EQ(dataOffset, size); + } + + /** + * Reduces values on all clients. + * This reduce just support SUM. + * The results are saved in recvBuf of rootId client + */ + template + void reduce(DataType* sendBuf, + DataType* recvBuf, + size_t size, + int clientId, + int rootId) { + putOwnData(clientId, DATA_REDUCE_SUM, sendBuf, size); + if (rootId == clientId) { + getAllData(clientId, DATA_REDUCE_SUM, recvBuf, size); + } + } + + /** + * return trans data type according to the input type + */ + virtual TransDataType getTransDtype(const std::type_info& info) { + TransDataType dataType; + if (typeid(int*) == info) { // NOLINT + dataType = TRANS_INT32; + } else if (typeid(uint32_t*) == info) { // NOLINT + dataType = TRANS_UINT32_T; + } else if (typeid(int64_t*) == info) { // NOLINT + dataType = TRANS_INT64_T; + } else if (typeid(uint64_t*) == info) { // NOLINT + dataType = TRANS_UINT64_T; + } else if (typeid(float*) == info) { // NOLINT + dataType = TRANS_FLOAT; + } else if (typeid(double*) == info) { // NOLINT + dataType = TRANS_DOUBLE; + } else { + LOG(FATAL) << "not supported"; + } + return dataType; + } + + protected: + /// for a > 0, b > 0: + /// return the smallest x s.t. b*x >= a + static int divup(int a, int b) { return (a + b - 1) / b; } + + int calcClientId(int i, int serviceNum) { + return (i + FLAGS_trainer_id * numPorts_) % serviceNum; + } + + /// start threads in sendThreads_ and recvThreads_ + void startThreads(); + + /// finish threads in sendThreads_ and recvThreads_ + void finishThreads(); + + template + void prepareData(int clientId, + SendDataType type, + DataUpdateMode updateMode, + DataType* datas, + size_t size, + SendJob* sendJob) { + sendJob->parallelDataRequests.resize(serviceNum_); + sendJob->parallelInputIovs.resize(serviceNum_); + for (int i = 0; i < serviceNum_; ++i) { + auto& request = sendJob->parallelDataRequests[i]; + request.set_update_mode(updateMode); + request.set_type(type); + request.set_client_id(clientId); + request.set_server_id(i); + } + + /// split datas which need send to Server into serviceNum_ pieces + if (!datas) { + CHECK(!size) << "ownSize should be zero since datas is nullptr"; + } + size_t baseSize = size / serviceNum_; + size_t dataOffset = 0; + for (int i = 0; i < serviceNum_; ++i) { + auto& request = sendJob->parallelDataRequests[i]; + DataBlock* block = request.add_blocks(); + size_t ownSize = size_t(i) < size % serviceNum_ ? baseSize + 1 : baseSize; + size_t realSize = datas ? std::max(ownSize, size_t(1)) : 0; + block->set_total_size(realSize * sizeof(DataType)); + block->set_data_size(sizeof(DataType)); + // TODO(yuyang18): The getTransDtype can be rewritten as template method + // to reduce runtime overhead. + block->set_data_type(getTransDtype(typeid(DataType*))); // NOLINT + if (datas) { + sendJob->parallelInputIovs[i].push_back( + {datas + dataOffset, realSize * sizeof(DataType)}); + } + dataOffset += ownSize; + } + CHECK_EQ(dataOffset, size); + } + + /** + * @brief send data to all data servers + * + * @note each trainer sends all its data to all data servers + * it's for broadcast data synchronization, such as features + * synchronization in metric learning. + */ + template + void sendData(int clientId, + SendDataType type, + DataUpdateMode updateMode, + DataType* datas, + size_t size) { + SendJobPtr sendJob = std::make_shared(); + prepareData(clientId, type, updateMode, datas, size, sendJob.get()); + for (int i = 0; i < threadNum_; ++i) { + sendJobQueue_[i]->enqueue(sendJob); + } + } + + /** + * @brief recv data from all data servers + * + * @note synchronize all recv threads + */ + void recvData(); + + /// send request, and recv responses + template + void multiCall(const char* funcName, + const ProtoIn& request, + std::vector* responses) { + responses->resize(clients_.size()); + size_t numClients = clients_.size(); + for (size_t i = 0; i < numClients; ++i) { + clients_[i].send(funcName, request); + } + for (size_t i = 0; i < numClients; ++i) { + clients_[i].recv(&(*responses)[i]); + } + } + + /** + * @brief synchronize all trainers and pservers + * + * @note used to ensure that data of all trainers have been received + */ + void synchronize(SyncObject syncObjectId = SYNC_DEFAULT); + + /** + * @brief use multithread to separately send data + * + * @note each thread should read its own JobQueue to handle requests + * each thread should calcClientId() to retrieve connections + * managed by himself. + * send and recv are implemented in child class. + */ + virtual void send(int threadId) = 0; + + /** + * @brief use multithread to separately receive data + * + * @note almost same as send() + */ + virtual void recv(int threadId) = 0; + + protected: + bool stopping_; + /// nodes * ports that means the number of real pservers + int serviceNum_; + /** + * threads num for managing all services. Normally the + * number of pservers are relatively less than several + * hundreds so that using thread-based parallelization + * can benifit traffic performance and pserver's sgd + * optimization performance. + */ + int threadNum_; + /// the connection manager at client end + std::vector clients_; + /// send threads for parallelization + std::vector sendThreads_; + /// recv threads for parallelization + std::vector recvThreads_; + std::unique_ptr recvSyncBarrier_; + + // TODO(yanfei): + // current pserver's will return value until all parameters' + // optimization are finished so that recv are not overlapped + // in reality. More robust implimentation should be to pipeline + // all send/recv action based on parameter unit level, and + // it will benifits deep and larger model training in future, + // especially local node compution power surpasses inter-connection + // such as GPU cluster, even with BOX GPU cluster. + // queue for buffering send request + /** + * send/recv queue cooperates with each other to accomplish + * overlapping communication with forwardBackward action. + */ + std::vector> sendJobQueue_; + /// queue for buffering recv request + std::vector> recvJobQueue_; + /// specific for dserver + SendJob sendJob_; + /// port num for each node + int numPorts_; + /// if set, overlapped optimization is disabled + bool separateSendAndRecv_; + std::vector recvDataMems_; +}; +} // namespace paddle diff --git a/paddle/pserver/CMakeLists.txt b/paddle/legacy/pserver/CMakeLists.txt similarity index 100% rename from paddle/pserver/CMakeLists.txt rename to paddle/legacy/pserver/CMakeLists.txt diff --git a/paddle/legacy/pserver/LightNetwork.cpp b/paddle/legacy/pserver/LightNetwork.cpp new file mode 100644 index 0000000000000000000000000000000000000000..469c95853ecdc02a6028417ca37b0020406eea09 --- /dev/null +++ b/paddle/legacy/pserver/LightNetwork.cpp @@ -0,0 +1,459 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +#include "LightNetwork.h" +#include "RDMANetwork.h" +#include "paddle/legacy/utils/StringUtil.h" +#include "paddle/legacy/utils/Util.h" + +/// quick ack can reduce the latency of small message +DEFINE_bool(small_messages, + false, + "if message size is small, recommend set it True to enable quick " + "ack and no delay"); + +/// reasonable sock_send_buf_size can control the traffic injected into switch +/// network. Injecting too many data into traffic could cause packets loss which +/// cause long latency and degrade the efficiency of communication. +DEFINE_int32(sock_send_buf_size, + 1024 * 1024 * 40, + "restrict sock send buff size, can reduce network congestion if " + "set carefully"); + +/// reasonable size can hold bursted packets and reduce packets loss +DEFINE_int32(sock_recv_buf_size, + 1024 * 1024 * 40, + "restrict sock recv buff size"); + +/// reasonable sock_listen_queue_size can control maximum pending connections. +DEFINE_int32(sock_listen_queue_size, + 1024, + "listen queue size when pserver listen a TCP port"); + +namespace paddle { + +/** + * @brief get ip address from interface name + * + * @param[in] device device interface name + */ +std::string getIpAddr(std::string &device) { + int sock; + struct sockaddr_in sin; + struct ifreq ifr; + + sock = socket(AF_INET, SOCK_DGRAM, 0); + CHECK(sock >= 0) << "Create socket error."; + + strncpy(ifr.ifr_name, device.c_str(), IFNAMSIZ); + ifr.ifr_name[IFNAMSIZ - 1] = 0; + + CHECK_GE(ioctl(sock, SIOCGIFADDR, &ifr), 0); + memcpy(&sin, &ifr.ifr_addr, sizeof(sin)); + close(sock); + return std::string(inet_ntoa(sin.sin_addr)); +} + +/** + * @brief set sock option + * + * @param[in] sockfd sock file descriptor + * + * @note adjust some default sock option for better performance + */ +void setOption(int sockfd) { +#if !defined(__APPLE__) && !defined(__OSX__) + int sendSize = FLAGS_sock_send_buf_size; + int recvSize = FLAGS_sock_recv_buf_size; + CHECK_GE( + setsockopt(sockfd, SOL_SOCKET, SO_RCVBUF, &recvSize, sizeof(recvSize)), + 0); + CHECK_GE( + setsockopt(sockfd, SOL_SOCKET, SO_SNDBUF, &sendSize, sizeof(sendSize)), + 0); +#endif + + if (FLAGS_small_messages) { + int optval = 1; + CHECK_GE( + setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, &optval, sizeof(optval)), + 0); +#ifdef TCP_QUICKACK + optval = 1; + CHECK_GE( + setsockopt(sockfd, IPPROTO_TCP, TCP_QUICKACK, &optval, sizeof(optval)), + 0); +#endif + } + int reuse = 1; + CHECK_GE(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &reuse, sizeof(reuse)), + 0); +} + +/** + * @brief class constructor for SocketServer + * @param[in] addr sock bind address + * @param[in] port sock bind port + * @param[in] rdmaCpu rdma sock bind cpu core + * + * @note start one socket server which hosts parameter server process. + * rdmaCpu is passed to rdma deamon for better performance, and + * start tcp socket instead of rdma socket if rdmaCpu is equal + * to -1. Each trainer process starts one connection to one socket + * server, and use --ports_num to build more connections to harness + * fat communication channel if necessary. + * each connection is controlled by single thread with blocking + * read and write. + */ +SocketServer::SocketServer(const std::string &addr, int port, int rdmaCpu) + : port_(port), addr_(addr), stopping_(false) { + if (rdmaCpu == -1) { + tcpRdma_ = F_TCP; + socket_ = 0; + maxPendingConnections_ = FLAGS_sock_listen_queue_size; + } else { + tcpRdma_ = F_RDMA; + rdmaCpu_ = rdmaCpu; + rdmaSocket_ = 0; + + std::stringstream ss; + ss << port; + rdmaUri_ = "rdma://" + addr + ":" + ss.str(); + } + + /// trigger to initialize RDMA lib + CHECK(RdmaClientDaemons::get()) << "initilizate RDMA failed\n"; +} + +SocketServer::~SocketServer() { + stopping_ = true; + /// trigger accept thread to stop + { + SocketClient trigger(addr_.empty() ? "127.0.0.1" : addr_, port_, tcpRdma_); + } + this->join(); +} + +/** + * @brief start one tcp server which hosts parameter server + * + * @note do tcp socket bind and listen. it will spawn one thread + * for each connection + */ +void SocketServer::tcpServer() { + int newsockfd; + socklen_t clilen; + struct sockaddr_in serv_addr, cli_addr; + struct hostent *server; + + /// First call to socket() function + socket_ = socket(AF_INET, SOCK_STREAM, 0); + CHECK(socket_ >= 0) << "ERROR opening socket"; + + /// Initialize socket structure + bzero((char *)&serv_addr, sizeof(serv_addr)); + serv_addr.sin_family = AF_INET; + serv_addr.sin_port = htons(port_); + if (!addr_.empty()) { + server = gethostbyname(addr_.c_str()); + CHECK(server) << "ERROR, no such host: " << addr_; + bcopy((char *)server->h_addr, + (char *)&serv_addr.sin_addr.s_addr, + server->h_length); + } else { + serv_addr.sin_addr.s_addr = INADDR_ANY; + } + + setOption(socket_); + + /// Now bind the host address using bind() call. + CHECK(bind(socket_, (struct sockaddr *)&serv_addr, sizeof(serv_addr)) >= 0) + << "ERROR on binding " << addr_; + + /// Now start listening for the clients, here process will + /// go in sleep mode and will wait for the incoming connection + listen(socket_, maxPendingConnections_); + clilen = sizeof(cli_addr); + + while (true) { + /// Accept actual connection from the client + newsockfd = accept(socket_, (struct sockaddr *)&cli_addr, &clilen); + if (stopping_) { + break; + } + CHECK(newsockfd >= 0) << "ERROR on accept"; + constexpr int kPeerNameLen = 128; + char peerName[kPeerNameLen]; + CHECK(inet_ntop(AF_INET, &cli_addr.sin_addr, peerName, kPeerNameLen)); + + SocketWorker *worker = + new SocketWorker(createChannel(newsockfd, std::string(peerName)), this); + worker->start(); + worker->detach(); + } + close(socket_); + LOG(INFO) << "pserver accept thread finish, addr=" << addr_ + << " port=" << port_; +} + +/** + * @brief start one rdma server which hosts parameter server + * + * @note do rdma bind and listen, which calling self-defined socket + * like rdma library. it will spawn one thread for each connection + */ +void SocketServer::rdmaServer() { + struct sxi_sock *newsock; + + /// First call to socket() function + rdmaSocket_ = rdma::ssocket(rdmaCpu_); + CHECK(rdmaSocket_) << "ERROR opening RDMA socket"; + + CHECK(rdma::bind(rdmaSocket_, rdmaUri_.c_str()) == 0) + << "ERROR bind RDMA socket"; + + /// Now start listening for the clients, here process will + /// go in sleep mode and will wait for the incoming connection + CHECK(rdma::listen(rdmaSocket_) == 0) << "ERROR listen RDMA socket"; + + while (true) { + /// Accept actual connection from the client + newsock = rdma::accept(rdmaSocket_); + if (stopping_) { + break; + } + CHECK(newsock) << "ERROR on accept"; + + constexpr int kPeerNameLen = 128; + char peerName[kPeerNameLen]; + + struct sockaddr_in *saddr = rdma::getSourceAddress(newsock); + CHECK(inet_ntop(AF_INET, &saddr->sin_addr, peerName, kPeerNameLen)); + + SocketWorker *worker = + new SocketWorker(createChannel(newsock, std::string(peerName)), this); + worker->start(); + worker->detach(); + } + rdma::close(rdmaSocket_); + LOG(INFO) << "pserver accept thread finish, rdma uri=" << rdmaUri_; +} + +/** + * @brief start a socket server + * + * @note framework for starting socket server + */ +void SocketServer::run() { + if (tcpRdma_ == F_TCP) { + LOG(INFO) << "tcp server start "; + tcpServer(); + } else if (tcpRdma_ == F_RDMA) { + LOG(INFO) << "rdma server start "; + rdmaServer(); + } +} + +/** + * @brief class constructor for rdma client deamons + * + * @note automatically start several client deamons for better performance + */ +std::unique_ptr RdmaClientDaemons::daemons_ = nullptr; +std::once_flag RdmaClientDaemons::initDataFlag_; + +RdmaClientDaemons::RdmaClientDaemons() { + if (FLAGS_rdma_tcp == "rdma") { + rdma::init(); + + struct sxi_socket *socket; + onlineCpus_ = rdma::numCpus(); + for (auto i = 0; i < onlineCpus_; i++) { + socket = rdma::csocket(i); + CHECK(socket) << "ERROR open client socket daemon"; + + rdmaClientSocket_.push_back(socket); + } + LOG(INFO) << "RDMA client daemons started, onlineCpus_:" << onlineCpus_; + /// round robin scheduler for new connection + curCpu_ = 0; + /// wait daemons to start completely. + sleep(2); + } +} + +RdmaClientDaemons::~RdmaClientDaemons() { + if (FLAGS_rdma_tcp == "rdma") { + for (auto i = 0; i < onlineCpus_; i++) { + rdma::close(rdmaClientSocket_[i]); + } + LOG(INFO) << "RDMA client daemons is destoryed, onlineCpus_ " + << onlineCpus_; + } +} + +/** + * @brief worker thread main context + * + * @note each connection from client(trainer) is controlled by single worker + * thread, which is for handling all parameter server requests + */ +void SocketWorker::run() { + LOG(INFO) << "worker started, peer = " << channel_->getPeerName(); + + std::vector inputIovs; + + while (true) { + std::unique_ptr msgReader = channel_->readMessage(); + if (!msgReader) { + break; + } + + auto callback = [this](const std::vector &outputIovs) { + channel_->writeMessage(outputIovs); + }; + + server_->handleRequest(std::move(msgReader), callback); + } + + LOG(INFO) << "worker begin to finish, peer = " << channel_->getPeerName(); + delete this; +} + +/** + * @brief start one tcp connection to tcp server + * @param[in] serverAddr tcp server ip + * @param[in] serverPort tcp server port + * + * @note each object contains one channel which accept byte stream + */ +void SocketClient::TcpClient(const std::string &serverAddr, int serverPort) { + struct sockaddr_in serv_addr; + struct hostent *server; + + int errRet; // temp for gethostbyname_r + + /// Create a socket point + int sockfd = socket(AF_INET, SOCK_STREAM, 0); + CHECK(sockfd >= 0) << "ERROR opening socket"; + +#if defined(__OSX__) || defined(__APPLE__) + server = getipnodebyname(serverAddr.c_str(), AF_INET, AI_DEFAULT, &errRet); + CHECK_NE(HOST_NOT_FOUND, errRet) << "ERROR, no such host: " << serverAddr + << " ret = " << errRet; + CHECK(server) << "getipnodebyname error!"; +#else + struct hostent hostinfo; + char buf[1024]; // temp for gethostbyname_r + CHECK_EQ( + 0, + gethostbyname_r( + serverAddr.c_str(), &hostinfo, buf, sizeof(buf), &server, &errRet)) + << "ERROR, no such host: " << serverAddr << " ret = " << errRet; + CHECK(server) << "gethostbyname_r error!"; +#endif + + bzero((char *)&serv_addr, sizeof(serv_addr)); + serv_addr.sin_family = AF_INET; + bcopy((char *)server->h_addr, + (char *)&serv_addr.sin_addr.s_addr, + server->h_length); + serv_addr.sin_port = htons(serverPort); + + setOption(sockfd); + + /// Now connect to the server + int retry_count = 0; + do { + if (connect(sockfd, (sockaddr *)&serv_addr, sizeof(serv_addr)) == 0) { + break; + } + + if (errno == ECONNREFUSED) { + LOG(WARNING) << "connection refused by pserver, try again!"; + if (retry_count++ >= 7) { + LOG(FATAL) << "connection refused by pserver, maybe pserver failed!"; + } + std::this_thread::sleep_for(std::chrono::seconds(1)); + } else { + CHECK(errno != 0) << "ERROR connecting to " << serverAddr << ":" + << serverPort << "errorno: " << errno; + } + } while (errno == ECONNREFUSED); + + channel_.reset(new SocketChannel(sockfd, serverAddr)); + tcpRdma_ = F_TCP; +} + +/** + * @brief start one RDMA connection to rdma server + * @param[in] serverAddr rdma server ip + * @param[in] serverPort rdma server port + * + * @note each object contains one channel which accept byte stream + * for rdma, low level sock also provide byte stream api. + */ +void SocketClient::RdmaClient(const std::string &serverAddr, int serverPort) { + struct sxi_sock *sock; + + std::stringstream ss; + ss << serverPort; + + std::string rdmaUri = "rdma://" + serverAddr + ":" + ss.str(); + + RdmaClientDaemons *daemons = RdmaClientDaemons::daemons_->get(); + socketDaemon_ = daemons->selectDaemon(); + + /// connect to server with socket daemon + sock = rdma::connect(socketDaemon_, rdmaUri.c_str()); + CHECK(sock) << "ERROR connect to server" << rdmaUri; + + std::vector seg; + str::split(rdmaUri, '/', &seg); + std::string server = seg.at(seg.size() - 1); + channel_.reset(new SocketChannel(sock, server)); + tcpRdma_ = F_RDMA; +} + +/** + * @brief class constructor + * @param[in] serverAddr pserver ip address + * @param[in] serverPort pserver port + * @param[in] ChannelType F_TCP or F_RDMA + * + * @note responsible for building one connection to specified pserver port + */ +SocketClient::SocketClient(const std::string &serverAddr, + int serverPort, + enum ChannelType channelType) { + if (channelType == F_RDMA) + RdmaClient(serverAddr, serverPort); + else + TcpClient(serverAddr, serverPort); +} + +} // namespace paddle diff --git a/paddle/legacy/pserver/LightNetwork.h b/paddle/legacy/pserver/LightNetwork.h new file mode 100644 index 0000000000000000000000000000000000000000..380f86832f5894fdf29588dde9a77068c624e066 --- /dev/null +++ b/paddle/legacy/pserver/LightNetwork.h @@ -0,0 +1,185 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "SocketChannel.h" + +#include +#include +#include +#include + +#include "paddle/legacy/utils/Thread.h" + +struct sxi_socket; + +namespace paddle { + +class SocketWorker; + +/** + * @brief class for holding all parameters processing for current port + * + * @note each parameter server inherits from one socket server, each + * server contains serveral woker threads which are to parallelize + * the processing of computation, but share some common datas stored + * in child class of socketserver. + */ +class SocketServer : public Thread { + // rdmaCpu controls the cpu affinity of RDMA server daemon, + // which could benifit performance. rdmaCpu = -1 means TCP + // is used instead of RDMA transport. + public: + SocketServer(const std::string& addr, int port, int rdmaCpu); + ~SocketServer(); + + virtual void run(); + + typedef std::function& outputIovs)> + ResponseCallback; + + protected: + // + // The derived class needs to implement this function + // to handle the request received by SocketWorker + // The request is encapsulated by MsgReader, which contains + // a set of blocks. + virtual void handleRequest(std::unique_ptr msgReader, + ResponseCallback callback) = 0; + + std::unique_ptr createChannel(int sock, + const std::string& peerName) { + return std::unique_ptr(new SocketChannel(sock, peerName)); + } + std::unique_ptr createChannel(struct sxi_sock* sock, + const std::string& peerName) { + return std::unique_ptr(new SocketChannel(sock, peerName)); + } + + friend class SocketWorker; + + private: + void rdmaServer(); + void tcpServer(); + + void detach() {} // detach accept thread is forbidden + + protected: + enum ChannelType tcpRdma_; + // for rdma + int rdmaCpu_; + std::string rdmaUri_; + sxi_socket* rdmaSocket_; + // for tcp + int port_; + std::string addr_; + int socket_; + int maxPendingConnections_; + bool stopping_; +}; + +/** + * @brief class for holding one connection from one trainer + * + * @note all parameter processing will run in the context of this worker + */ +class SocketWorker : public Thread { + public: + SocketWorker(std::unique_ptr&& channel, SocketServer* server) + : channel_(std::move(channel)), server_(server) {} + + virtual ~SocketWorker() {} + + virtual void run(); + + protected: + std::unique_ptr channel_; + SocketServer* server_; + enum ChannelType tcpRdma_; +}; + +/** + * @brief class for providing rdma client deamon thread + * + * @note the deamons are required by sock like rdam library. Here + * use singleton model for daemons. Each deamon hosts in + * single cpu core for better load balance performance + */ +class RdmaClientDaemons { + private: + RdmaClientDaemons(); + + static std::unique_ptr daemons_; + + public: + static RdmaClientDaemons* get() { + std::call_once(RdmaClientDaemons::initDataFlag_, + &RdmaClientDaemons::getInstance); + + return daemons_.get(); + } + + struct sxi_socket* selectDaemon() { + int cpu = curCpu_; + curCpu_ = (curCpu_ + 1) % onlineCpus_; + + LOG(INFO) << "select daemon " << cpu << "onlineCpus_ " << onlineCpus_; + return rdmaClientSocket_[cpu]; + } + + ~RdmaClientDaemons(); + + public: + friend class SocketClient; + + private: + static std::once_flag initDataFlag_; + static void getInstance() { + if (!daemons_.get()) daemons_.reset(new RdmaClientDaemons()); + } + + std::vector rdmaClientSocket_; + std::atomic curCpu_; + int onlineCpus_; +}; + +/** + * @brief management for client connection which are from trainers + * + * @note it contains one channel descriptor which used to write and + * read data + */ +class SocketClient { + public: + SocketClient(const std::string& serverAddr, + int serverPort, + enum ChannelType channelType); + + SocketChannel* getChannel() { return channel_.get(); } + + protected: + std::unique_ptr channel_; + struct sxi_socket* socketDaemon_; + enum ChannelType tcpRdma_; + + private: + void RdmaClient(const std::string& serverAddr, int serverPort); + void TcpClient(const std::string& serverAddr, int serverPort); +}; + +std::string getIpAddr(std::string& device); +void setOption(int sockfd); + +} // namespace paddle diff --git a/paddle/legacy/pserver/ParameterClient2.cpp b/paddle/legacy/pserver/ParameterClient2.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4c544ddc28517f50e7deb23d4fa7a82b34d42677 --- /dev/null +++ b/paddle/legacy/pserver/ParameterClient2.cpp @@ -0,0 +1,781 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "ParameterClient2.h" +#include "paddle/legacy/math/SparseRowMatrix.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/StringUtil.h" + +DEFINE_string(pservers, "127.0.0.1", "Comma separated addresses of pservers"); +DEFINE_int32(parallel_thread_num, 1, "Thread number for parameter send"); + +namespace paddle { + +template +void copyToRepeatedField(google::protobuf::RepeatedField* dest, + const T2* src, + size_t size) { + dest->Clear(); + dest->Reserve(size); + for (size_t i = 0; i < size; ++i) { + dest->AddAlreadyReserved(src[i]); + } +} + +ParameterClient2::ParameterClient2(bool separate, int port, int numPorts) + : BaseClient(separate, numPorts), port_(port) { +#ifndef PADDLE_DISABLE_TIMER + forwardbackwordTime_ = 0; +#endif +} + +int ParameterClient2::calcParameterBlockSize( + const std::vector& parameters, size_t serviceNum) { + size_t totalSize = 0; + for (auto& para : parameters) { + totalSize += para->getSize(); + } + size_t perServerSize = totalSize / serviceNum; + + int sizeBits = 64 - __builtin_clzl(perServerSize); + + /// 2^10 is min block size + /// 2^7 will be max number of blocks in one pserver + int blockSizeBits = std::max((sizeBits - 7), 10); + return 1 << blockSizeBits; +} + +void ParameterClient2::initThreads() { + threadNum_ = serviceNum_; + if (FLAGS_parallel_thread_num > 1) { + LOG(INFO) << "parallel_thread_num dosent need to set"; + } + syncThreadPool_.reset(new SyncThreadPool(threadNum_)); + startThreads(); +} + +bool ParameterClient2::init(const std::vector& parameters) { + destroy(); + + std::vector hosts; + str::split(FLAGS_pservers, ',', &hosts); + serviceNum_ = hosts.size() * numPorts_; + uint64_t denseBlockSize = calcParameterBlockSize(parameters, serviceNum_); + + /// setup prefetch matrix if exists + for (auto& para : parameters) { + /// set block size for each parameter + para->getConfig().set_parameter_block_size( + para->getConfig().sparse_remote_update() ? para->getConfig().dims(1) + : denseBlockSize); + } + + for (auto& para : parameters) { + CHECK_NE(-1UL, para->getID()) << "id in parameter is not initialized"; + parameterMap_[para->getID()] = para; + } + + allSegments_.reserve(parameters.size()); + + for (auto& para : parameters) { + ParameterSegments segments; + segments.name = para->getName(); + segments.id = para->getID(); + allSegments_.push_back(segments); + if (para->getConfig().sparse_remote_update()) { + CHECK_EQ(para->getConfig().parameter_block_size(), + para->getConfig().dims(1)) + << "For sparse remote update parameter," + << " block size is the width of each row."; + } + } + + /// init clients + clients_.reserve(serviceNum_); + recvDataMems_.resize(serviceNum_); + + for (size_t i = 0; i < hosts.size(); ++i) { + for (int j = 0; j < numPorts_; ++j) { + LOG(INFO) << "pserver " << i * numPorts_ + j << " " << hosts[i] << ":" + << port_ + j; + if (FLAGS_rdma_tcp == "rdma") { + clients_.emplace_back(hosts[i], port_ + j, F_RDMA); + } else { + clients_.emplace_back(hosts[i], port_ + j, F_TCP); + } + } + } + + sparseDistribution_.reset(new SparseParameterDistribution(serviceNum_)); + + sleep(2); + + initThreads(); + + return true; +} + +ParameterClient2::~ParameterClient2() { destroy(); } + +void ParameterClient2::destroy() { + if (clients_.empty()) { + /// this means not initialized. + return; + } + finishThreads(); + + parameterMap_.clear(); + allSegments_.clear(); + clients_.clear(); +} + +void ParameterClient2::sendParallel(int tid, + size_t numThreads, + ParameterType recvParameterType) { + int numMyClients = divup(serviceNum_ - tid, numThreads); + + for (int j = 0; j < numMyClients; ++j) { + REGISTER_TIMER("client_sendAndRecv_send"); + int i = numThreads * j + tid; + /// Try to make different clients to send data to different pservers + /// at the same time so that they will not flood data to the same + /// pserver. + i = calcClientId(i, serviceNum_); + clients_[i].send("sendParameter", + sendJob_.parallelRequests[i], + sendJob_.parallelInputIovs[i]); + + /// clear large structure + sendJob_.parallelRequests[i].Clear(); + sendJob_.parallelInputIovs[i].clear(); + } + + std::vector bufs; + SendParameterResponse response; + for (int j = 0; j < numMyClients; ++j) { + REGISTER_TIMER("client_sendAndRecv_recv"); + int i = numThreads * j + tid; + i = calcClientId(i, serviceNum_); + auto msgReader = clients_[i].recv(&response); + CHECK_EQ(msgReader->getNumBlocks(), (size_t)response.blocks_size()); + bufs.clear(); + bufs.reserve(response.blocks_size()); + for (auto& block : response.blocks()) { + auto it = parameterMap_.find(block.para_id()); + CHECK(it != parameterMap_.end()); + Parameter* parameter = it->second.get(); + real* buf = nullptr; + if (parameter->getBuf(recvParameterType)) { + buf = parameter->getBuf(recvParameterType)->getPoint(block.begin_pos()); + } else { + auto recvMat = dynamic_cast( + parameter->getMat(recvParameterType).get()); + CHECK(recvMat); + size_t width = parameter->getConfig().dims(1); + // TODO(wuyi): need add lock here? may also cause resize. + buf = recvMat->getLocalRow(block.begin_pos() / width); + } + /// sparse_id is not useful while receiving data since sparse data + /// storage is continuous, do commit recieved data as that of dense. + bufs.push_back(buf); + } + msgReader->readBlocks(bufs); + } +} + +void ParameterClient2::prepareSendData( + ParameterUpdateMode updateMode, + ParameterType parameterType, + const std::vector& parameterSegments, + int64_t numSamples, + real cost, + bool sendBackParameter, + ParameterType sendBackParameterType, + BatchStatus batchStatus, + SendJob* sendJob) { + sendJob->parallelRequests.resize(serviceNum_); + sendJob->parallelInputIovs.resize(serviceNum_); + + for (auto& request : sendJob->parallelRequests) { +#ifndef PADDLE_DISABLE_TIMER + if (updateMode == PSERVER_UPDATE_MODE_ADD_GRADIENT) { + request.set_forwardbackward_time(forwardbackwordTime_); + } +#endif + request.set_trainer_id(trainerId_); + request.set_update_mode(updateMode); + request.set_send_back_parameter(sendBackParameter); + request.set_send_back_parameter_type(sendBackParameterType); + request.set_num_samples(numSamples); + request.set_cost(cost); + request.set_batch_status(batchStatus); + CHECK_EQ(request.blocks_size(), 0); + VLOG(10) << "request: trainer_id: " << request.trainer_id() + << " update_mode" << request.update_mode() + << " send_back_parameter: " << request.send_back_parameter() + << " send_back_parameter_type: " + << request.send_back_parameter_type() + << " num_samples: " << request.num_samples() + << " cost: " << request.cost() + << " batch_status: " << request.batch_status(); + } + for (const auto& segments : parameterSegments) { + const auto it = parameterMap_.find(segments.id); + CHECK(it != parameterMap_.end()); + Parameter* parameter = it->second.get(); + CHECK(parameter != nullptr) << "parameter is nullptr"; + int64_t nameHash = std::hash()(segments.name); + bool sendingPara = !(updateMode == PSERVER_UPDATE_MODE_GET_PARAM || + updateMode == PSERVER_UPDATE_MODE_GET_PARAM_SPARSE || + updateMode == PSERVER_UPDATE_MODE_SET_PARAM_ZERO); + bool sparseUpdate = parameter->getConfig().sparse_remote_update() && + (updateMode == PSERVER_UPDATE_MODE_ADD_GRADIENT || + updateMode == PSERVER_UPDATE_MODE_ASYNC_SGD || + updateMode == PSERVER_UPDATE_MODE_GET_PARAM_SPARSE); + + const auto blockSize = parameter->getConfig().parameter_block_size(); + CHECK_GE(blockSize, 1LU) << "blockSize should > 0 " << blockSize; + const auto paraSize = parameter->getSize(); + if (sparseUpdate) { + auto prefetchMat = std::dynamic_pointer_cast( + parameter->getMat(PARAMETER_VALUE)); + CHECK(prefetchMat != nullptr) << "prefetchMat is nullptr"; + auto sendMat = dynamic_cast( + parameter->getMat(parameterType).get()); + CHECK(sendMat != nullptr) << "sendMat is nullptr"; + + syncThreadPool_->exec([&](int tid, size_t numThreads) { + std::lock_guard guard(sparseAutoGrowthMutex_); + const auto& localIndices = prefetchMat->getLocalIndices(); + /// num of sparse rows + size_t nLocalBlocks = localIndices.size(); + uint64_t beginDim = 0; + uint64_t endDim = 0; + + // HACK(typhoonzero): let it resize first + prefetchMat->getLocalRow(nLocalBlocks); + sendMat->getLocalRow(nLocalBlocks); + + for (size_t row = 0; row < nLocalBlocks; ++row) { + int64_t blockId = localIndices[row]; // local row -> sparse row + int serverId = std::abs((blockId + nameHash) % serviceNum_); + if (serverId % numThreads != (size_t)tid) { + continue; + } + + beginDim = blockId * blockSize; + endDim = std::min(beginDim + blockSize, paraSize); + + auto& request = sendJob->parallelRequests[serverId]; + ParameterBlock* block = request.add_blocks(); + block->set_para_id(segments.id); + /// global sparse row id + block->set_block_id(blockId); + /// local row offset + block->set_begin_pos(row * blockSize); + /// block len + block->set_block_size(endDim - beginDim); + if (sendingPara) { + sendJob->parallelInputIovs[serverId].push_back( + {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize}); + /// detect sparse parameter distribution + sparseDistribution_->probeDistribution(serverId, + sizeof(real) * blockSize); + } + } + }); + + } else { /// parameter set for dense and sparse + real* buf = + sendingPara ? parameter->getBuf(parameterType)->getPoint(0) : nullptr; + uint64_t endDim = 0; + for (uint64_t beginDim = 0; beginDim < paraSize; beginDim = endDim) { + endDim = std::min(beginDim + blockSize, paraSize); + int64_t blockId = beginDim / blockSize; + int serverId = std::abs((blockId + nameHash) % serviceNum_); + + auto& request = sendJob->parallelRequests[serverId]; + ParameterBlock* block = request.add_blocks(); + block->set_para_id(segments.id); + block->set_block_id(blockId); + block->set_begin_pos(beginDim); + block->set_block_size(endDim - beginDim); + if (buf) { + sendJob->parallelInputIovs[serverId].push_back( + {buf + beginDim, sizeof(real) * ((size_t)(endDim - beginDim))}); + } + } + } + } // parameterSegments + + sparseDistribution_->checkAndResetDistribution(); +} + +void ParameterClient2::sendAndReceiveParameter( + ParameterUpdateMode updateMode, + ParameterType parameterType, + const std::vector& parameterSegments, + int64_t numSamples, + real cost, + bool sendBackParameter, + ParameterType sendBackParameterType, + ParameterType recvParameterType) { + prepareSendData(updateMode, + parameterType, + parameterSegments, + numSamples, + cost, + sendBackParameter, + sendBackParameterType, + /*batchStatus = */ BATCH_START_AND_FINISH, + &sendJob_); + + syncThreadPool_->exec([&](int tid, size_t numThreads) { + this->sendParallel(tid, numThreads, recvParameterType); + }); +} + +void ParameterClient2::sendParameter( + ParameterUpdateMode updateMode, + ParameterType parameterType, + const std::vector& parameterSegments, + int64_t numSamples, + real cost, + bool sendBackParameter, + BatchStatus batchStatus) { + SendJobPtr sendJob = std::make_shared(); + prepareSendData(updateMode, + parameterType, + parameterSegments, + numSamples, + cost, + sendBackParameter, + PARAMETER_VALUE, + batchStatus, + sendJob.get()); + + for (int i = 0; i < threadNum_; i++) { + sendJobQueue_[i]->enqueue(sendJob); + } +} + +void ParameterClient2::recvParameter() { recvSyncBarrier_->wait(); } + +void ParameterClient2::send(int threadId) { + int index = threadId; + LOG(INFO) << "send thread " << threadId << " started"; + int numMyClients = divup(serviceNum_ - index, threadNum_); + while (true) { + SendJobPtr recvJob = sendJobQueue_[index]->dequeue(); + if (stopping_) { + recvJobQueue_[index]->enqueue(recvJob); + break; + } + for (int j = 0; j < numMyClients; ++j) { + REGISTER_TIMER("client_send"); + int i = threadNum_ * j + index; + /// Try to make different clients to send data to different pservers + /// at the same time so that they will not flood data to the same + /// pserver. + i = calcClientId(i, serviceNum_); + if (recvJob->parallelRequests.size()) { + clients_[i].send("sendParameter", + recvJob->parallelRequests[i], + recvJob->parallelInputIovs[i]); + } else { + clients_[i].send("sendData", + recvJob->parallelDataRequests[i], + recvJob->parallelInputIovs[i]); + } + } + recvJobQueue_[index]->enqueue(recvJob); + } +} + +void ParameterClient2::recv(int threadId) { + LOG(INFO) << "recv thread " << threadId << " started"; + int index = threadId; + int numMyClients = divup(serviceNum_ - index, threadNum_); + while (true) { + std::vector bufs; + SendParameterResponse response; + SendDataResponse dataResponse; + SendJobPtr recvJob = recvJobQueue_[index]->dequeue(); + if (stopping_) break; + for (int j = 0; j < numMyClients; ++j) { + REGISTER_TIMER("client_recv"); + int i = threadNum_ * j + index; + i = calcClientId(i, serviceNum_); + if (recvJob->parallelRequests.size()) { + auto msgReader = clients_[i].recv(&response); + CHECK_EQ(msgReader->getNumBlocks(), (size_t)response.blocks_size()); + bufs.clear(); + bufs.reserve(response.blocks_size()); + for (auto& block : response.blocks()) { + auto it = parameterMap_.find(block.para_id()); + CHECK(it != parameterMap_.end()); + Parameter* parameter = it->second.get(); + real* buf = + parameter->getBuf(PARAMETER_VALUE)->getPoint(block.begin_pos()); + CHECK_EQ(msgReader->getBlockLength(bufs.size()), + sizeof(real) * (block.block_size())); + bufs.push_back(buf); + } + msgReader->readBlocks(bufs); + } else { + auto msgReader = clients_[i].recv(&dataResponse); + CHECK_EQ(msgReader->getNumBlocks(), (size_t)dataResponse.blocks_size()); + size_t totalLen = msgReader->getTotalLength(); + if (0 == totalLen) { + continue; + } + auto& recvMem = recvDataMems_[dataResponse.server_id()]; + CHECK_EQ(dataResponse.blocks_size(), 1) + << "Only one block currently support now!"; + auto& block = dataResponse.blocks(0); + CHECK_EQ(totalLen % sizeof(block.data_size()), 0U); + recvMem = std::make_shared(totalLen); + msgReader->readNextBlock(recvMem.get()->getBuf()); + } + } + recvSyncBarrier_->wait(); + } +} + +void ParameterClient2::waitPassStart() { + WaitPassStartRequest request; + std::vector responses; + multiCall(__func__, request, &responses); +} + +void ParameterClient2::waitPassFinish() { + WaitPassFinishRequest request; + std::vector responses; + multiCall(__func__, request, &responses); +} + +void ParameterClient2::synchronize(SyncObject syncObjectId) { + SynchronizeRequest request; + request.set_sync_object_id(syncObjectId); + std::vector responses; + multiCall(__func__, request, &responses); +} + +void ParameterClient2::asyncFinishPass(SyncObject syncObjectId) { + SynchronizeRequest request; + request.set_sync_object_id(syncObjectId); + request.set_trainer_id(trainerId_); + std::vector responses; + multiCall(__func__, request, &responses); +} + +void ParameterClient2::setConfig(const OptimizationConfig& optConfig, + const std::string& saveDir, + bool isSparseServer) { + SetConfigRequest request; + std::vector responses; + + for (auto& nameAndPara : parameterMap_) { + *request.add_param_configs() = nameAndPara.second->getConfig(); + } + + *request.mutable_opt_config() = optConfig; + request.set_save_dir(saveDir); + request.set_is_sparse_server(isSparseServer); + + std::vector requests; + requests.resize(clients_.size()); + for (size_t i = 0; i < requests.size(); ++i) { + requests[i].CopyFrom(request); + requests[i].set_server_id(i); + } + + responses.resize(clients_.size()); + size_t numClients = clients_.size(); + for (size_t i = 0; i < numClients; ++i) { + clients_[i].send(__func__, requests[i]); + } + for (size_t i = 0; i < numClients; ++i) { + clients_[i].recv(&responses[i]); + } +} + +bool ParameterClient2::inStatus(PServerStatus status) { + GetStatusRequest request; + std::vector responses; + + bool ok = true; + multiCall("getStatus", request, &responses); + for (auto& response : responses) { + if (response.status() != status) { + ok = false; + } + } + + return ok; +} + +void ParameterClient2::setStatus(PServerStatus status) { + SetStatusRequest request; + request.set_status(status); + std::vector responses; + multiCall(__func__, request, &responses); +} + +void ParameterClient2::waitForStatus(PServerStatus status) { + while (!inStatus(status)) { + sleep(1); + } +} + +template +static void validateResponses(const std::vector& responses) { + for (auto& response : responses) { + CHECK(response.return_message().empty()) + << "client" << &response - &responses[0] + << " error:" << response.return_message(); + } +} + +PServerVector ParameterClient2::createVector() { + CreateVectorRequest request; + std::vector responses; + int64_t handle = -1; + + multiCall(__func__, request, &responses); + validateResponses(responses); + + for (auto& response : responses) { + if (handle == -1) { + handle = response.handle(); + } else { + CHECK_EQ(handle, response.handle()) << "Inconsistent handle from client" + << &response - &responses[0] << " " + << handle << " " << response.handle(); + } + } + return PServerVector{handle}; +} + +void ParameterClient2::releaseVector(PServerVector handle) { + ReleaseVectorRequest request; + std::vector responses; + + request.set_handle(handle.handle); + multiCall(__func__, request, &responses); + validateResponses(responses); +} + +PServerMatrix ParameterClient2::createMatrix(int32_t numCols) { + CreateMatrixRequest request; + std::vector responses; + int64_t handle = -1; + + request.set_num_cols(numCols); + multiCall(__func__, request, &responses); + validateResponses(responses); + + for (auto& response : responses) { + if (handle == -1) { + handle = response.handle(); + } else { + CHECK_EQ(handle, response.handle()) << "Inconsistent handle from client" + << &response - &responses[0] << " " + << handle << " " << response.handle(); + } + } + return PServerMatrix{handle}; +} + +void ParameterClient2::releaseMatrix(PServerMatrix handle) { + ReleaseMatrixRequest request; + std::vector responses; + + request.set_handle(handle.handle); + multiCall(__func__, request, &responses); + validateResponses(responses); +} + +void PreparedOperations::addOperationHelper(Operation* op, CpuVectorPtr vec) { + ProtoVector& pvec = *op->add_vectors(); + size_t dim = vec->getSize(); + pvec.set_dim(dim); + copyToRepeatedField(pvec.mutable_values(), vec->getData(), vec->getSize()); +} + +void PreparedOperations::addOperationHelper(Operation* op, CpuMatrixPtr mat) { + ProtoMatrix& pmat = *op->add_matrices(); + pmat.set_num_cols(mat->getWidth()); + pmat.set_num_rows(mat->getHeight()); + copyToRepeatedField( + pmat.mutable_values(), mat->getData(), pmat.num_cols() * pmat.num_rows()); +} + +static inline real addTwo(real a, double b) { return a + b; } + +void ParameterClient2::doOperation(PreparedOperations& ops, + bool waitForGradient, + bool sendBackGradient, + bool releasePass) { + std::vector responses; + ops.request_.set_wait_for_gradient(waitForGradient); + ops.request_.set_send_back_parameter(sendBackGradient); + ops.request_.set_release_pass(releasePass); + multiCall(__func__, ops.request_, &responses); + validateResponses(responses); + size_t numPassFinishServers = 0; + + size_t numOps = ops.request_.operations_size(); + for (auto& response : responses) { + numPassFinishServers += response.pass_finish(); + CHECK_EQ(numOps, (size_t)response.results_size()); + for (size_t opId = 0; opId < numOps; ++opId) { + const OperationResult& result = response.results(opId); + std::vector& resultScalars = ops.localResults_[opId].resultScalars; + std::vector& resultVectors = + ops.localResults_[opId].resultVectors; + std::vector& resultMatrices = + ops.localResults_[opId].resultMatrices; + + if (&response == &responses[0]) { + /// Initialize results to zero + + resultScalars.resize(result.scalars_size()); + for (auto p : resultScalars) { + if (!p) continue; + *p = 0; + } + size_t numVectors = result.vectors_size(); + resultVectors.resize(numVectors); + for (size_t i = 0; i < numVectors; ++i) { + if (!resultVectors[i]) continue; + resultVectors[i]->resize(result.vectors(i).dim()); + resultVectors[i]->zeroMem(); + } + size_t numMatrices = result.matrices_size(); + resultMatrices.resize(numMatrices); + for (size_t i = 0; i < numMatrices; ++i) { + if (!resultMatrices[i]) continue; + resultMatrices[i]->resize(result.matrices(i).num_rows(), + result.matrices(i).num_cols()); + resultMatrices[i]->zeroMem(); + } + } + + // aggregate results from each pserver to results + + CHECK_EQ(resultScalars.size(), (size_t)result.scalars_size()); + for (ssize_t i = 0; i < result.scalars_size(); ++i) { + real* rscalar = resultScalars[i]; + if (!rscalar) continue; + *rscalar += result.scalars(i); + } + + CHECK_EQ(resultVectors.size(), (size_t)result.vectors_size()); + for (auto& vec : result.vectors()) { + int i = &vec - &result.vectors(0); + CpuVectorPtr rvec = resultVectors[i]; + if (!rvec) continue; + CHECK_EQ(rvec->getSize(), (size_t)vec.dim()); + std::transform(rvec->getData(), + rvec->getData() + rvec->getSize(), + vec.values().data(), + rvec->getData(), + addTwo); + } + + CHECK_EQ(resultMatrices.size(), (size_t)result.matrices_size()); + for (auto& mat : result.matrices()) { + int i = &mat - &result.matrices(0); + CpuMatrixPtr rmat = resultMatrices[i]; + if (!rmat) continue; + CHECK_EQ(rmat->getHeight(), (size_t)mat.num_rows()); + CHECK_EQ(rmat->getWidth(), (size_t)mat.num_cols()); + + std::transform(rmat->getData(), + rmat->getData() + rmat->getElementCnt(), + mat.values().data(), + rmat->getData(), + addTwo); + } + } + } + passFinish_ = numPassFinishServers == clients_.size(); +} + +real ParameterClient2::vectorDotProduct(PServerVector u, PServerVector v) { + real result = 0.0; + PreparedOperations ops; + ops.addOperation(PSERVER_OP_utv, u, v)(&result); + doOperation(ops, false, false); + return result; +} + +void ParameterClient2::vectorScale(PServerVector u, real a) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_au, u, a); + doOperation(ops, false, false); +} + +void ParameterClient2::vectorCopy(PServerVector src, PServerVector dst) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_COPY, src, dst); + doOperation(ops, false, false); +} + +void ParameterClient2::vectorAddMult(PServerVector u, PServerVector v, real a) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_au_bv, v, u, a, (real)1); + doOperation(ops, false, false); +} + +void ParameterClient2::vectorAddMultInto(PServerVector u, + PServerVector v, + PServerVector w, + real a) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_au_bv_cw, v, w, u, (real)1, a, (real)0); + doOperation(ops, false, false); +} + +void ParameterClient2::vectorScaleInto(PServerVector u, + PServerVector v, + real a) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_au_bv, v, u, a, (real)0); + doOperation(ops, false, false); +} + +void ParameterClient2::loadValueVector(const std::string& dirName) { + LoadValueRequest request; + request.set_dir_name(dirName); + std::vector responses; + + multiCall(__func__, request, &responses); + validateResponses(responses); +} + +void ParameterClient2::saveValueVector(const std::string& dirName) { + SaveValueRequest request; + request.set_dir_name(dirName); + std::vector responses; + + multiCall(__func__, request, &responses); + validateResponses(responses); +} + +} // namespace paddle diff --git a/paddle/legacy/pserver/ParameterClient2.h b/paddle/legacy/pserver/ParameterClient2.h new file mode 100644 index 0000000000000000000000000000000000000000..9320e19c4df6c5439266f89e5599b9496f145172 --- /dev/null +++ b/paddle/legacy/pserver/ParameterClient2.h @@ -0,0 +1,602 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include + +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/pserver/BaseClient.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Queue.h" +#include "paddle/legacy/utils/Util.h" + +#include "ParameterService.pb.h" + +#include "ProtoServer.h" +#include "SparseParameterDistribution.h" + +DECLARE_int32(parallel_thread_num); + +namespace paddle { + +struct PServerMatrix { + int64_t handle; +}; + +struct PServerVector { + int64_t handle; +}; + +/** + * @brief A class to help to prepare server-side operations. + */ +class PreparedOperations { + protected: + class ResultsAdder; + struct LocalOperationResult; + + public: + /** + * Offers an easy way to prepare operations that will be performed on + * server-side. + * + * Usage: + * @code + * addOperation(optype, arguments...)(results...) + * @endcode + * + * Examples: + * 1. set pserver vector to 1: + * @code + * PServerVector u = parameterClient.createVector(); + * addOperation(PSERVER_OP_RESET, u, (real)1); + * @endcode + * + * 2. Compute inner product of to pserver vectors. + * @code + * PServerVector u = parameterClient.createVector(); + * PServerVector v = parameterClient.createVector(); + * real result; + * addOperation(PSERVER_OP_utv, u, v)(&result) + * @endcode + * + * @param[in] operation The operation that pserver will perform. + * @param[in] args Argument list of the operation + * @return A ResultsAdder object initialized with the last element of + * localResults_. + */ + template + ResultsAdder addOperation(MatrixVectorOperation operation, Args... args) { + Operation* op = request_.add_operations(); + op->set_operation(operation); + localResults_.emplace_back(); + addOperationHelper(op, args...); + return ResultsAdder(&localResults_.back()); + } + + protected: + void addOperationHelper(Operation* op) {} + + /** + * @brief Helper function to add an new operation that takes a PServerVector + * as an operand. + */ + void addOperationHelper(Operation* op, PServerVector arg) { + op->add_pvectors(arg.handle); + } + + /** + * @brief Helper function to add an new operation that takes a PServerMatrix + * as an operand. + */ + void addOperationHelper(Operation* op, PServerMatrix arg) { + op->add_pmatrices(arg.handle); + } + + /** + * @brief Helper function to add an new operation that takes a real valued + * scalar as an operand. + */ + void addOperationHelper(Operation* op, real arg) { op->add_scalars(arg); } + + /** + * @brief Helper function to add an new operation that takes a CpuVectorPtr + * as an operand. + * @note The array of CpuVectors that arg points to will be copied to + * op's vectors field. + */ + void addOperationHelper(Operation* op, CpuVectorPtr arg); + + /** + * @brief Helper function to add an new operation that takes a CpuMatrixPtr + * as an operand. + * @note The array of CpuMatrixs that arg points to will be copied to + * op's matrices field. + */ + void addOperationHelper(Operation* op, CpuMatrixPtr arg); + + /** + * @brief Helper function to add an new operation and prepare the operands. + * + * @tparam Arg An operand of the operation. + * @tparam Args A list of rest operands of the operation. + * @param op Pointer to an Operation object. + */ + template + void addOperationHelper(Operation* op, Arg arg, Args... args) { + addOperationHelper(op, arg); + addOperationHelper(op, args...); + } + + /** + * @brief ResultsAdder offers easy ways to quickly store operation results. + */ + class ResultsAdder { + public: + explicit ResultsAdder(LocalOperationResult* localResult) + : localResult_(localResult) {} + template + void operator()(Args... args) { + addResult(args...); + } + void addResult() {} + void addResult(real* arg) { localResult_->resultScalars.push_back(arg); } + void AddResult(CpuVectorPtr arg) { + localResult_->resultVectors.push_back(arg); + } + void AddResult(CpuMatrixPtr arg) { + localResult_->resultMatrices.push_back(arg); + } + template + void addResult(Arg arg, Args... args) { + addResult(arg); + addResult(args...); + } + + protected: + LocalOperationResult* localResult_; + }; + + protected: + DoOperationRequest request_; + std::vector inputIovs_; + struct LocalOperationResult { + std::vector resultScalars; + std::vector resultVectors; + std::vector resultMatrices; + }; + std::vector localResults_; + friend class ParameterClient2; +}; + +struct ParameterSegments { + std::string name; // name of the parameter + size_t id; // id of the parameter +}; + +/** + * The client interface for parameter server. ParameterClient2 supports 2 modes + * for managing connections to parameter servers, in the 1st mode one connection + * is shared by 2 threads that are separately responsible for sending and + * recieving activities, in the 2nd mode one connection is owned by only one + * thread, and all the sending and recieving activities run in that single + * thread. + * TODO(yanfei): + * Additional core idea to further optimizate pserver performance is + * to do sync-sgd based parameter level instead of pserver level. + * full-parallelization based parameter level for sync-sgd also can + * sense forwardbackward computation layer-by-layer for more deeper layer + * model. + * Firstly, pserver can do full-parallelization on all computation based + * parameter level instead of waiting for all gradients are finished and + * start to send back parameters value immediately if parameter is ready + * instead of waiting for all parameters value are ready + * Secondly, parameter client can write back parameters to GPU instead of + * waiting until all parameters are received to CPU host end. + */ +class ParameterClient2 : public BaseClient { + public: + /** Constructor. + * @param separate True if sending and recieving activities are separated + * into 2 threads, otherwise false. + * @param port Port number that parameter client runs on. + * @param numPorts Number of ports parameter clients occupies, + * numPorts * pserver number is the total number of + * connections the parameter client maintains. + */ + ParameterClient2(bool separate = false, + int port = FLAGS_port, + int numPorts = FLAGS_ports_num); + + ~ParameterClient2(); + + static int calcParameterBlockSize(const std::vector& parameters, + size_t serviceNum); + + public: + bool init(const std::vector& parameters); + + /// service functions + + /** + * @brief Sends the segments in parameter to parameter servers, then receives + * the response from the servers. + * @param[in] updateMode Indicates how parameters should be updated on the + * server side. + * @param[in] parameterType Type of parameter that will be sent. + * @param[in] segments Segments in the parameter that will be sent. + * @param[in] numSamples Number of samples this update is based on. + * @param[in] cost Cost of the batch, will be used to calculate global object + * value. + * @param[in] sendBackParameter True if the updated parameters should be sent + * back, otherwise false. + * @param[in] sendBackParameterType Send back parameter type on pserver, + * PARAMETER_VALUE by default + * @param[in] recvParameterType pserver[sendBackParameterType] will be copy to + * client[recvParameterType] + * @note Only parameterType will be sent. + */ + void sendAndReceiveParameter(ParameterUpdateMode updateMode, + ParameterType parameterType, + const std::vector& segments, + int64_t numSamples, + real cost, + bool sendBackParameter, + ParameterType sendBackParameterType, + ParameterType recvParameterType); + + /** + * @brief Sends all parameters to parameter servers, and receives the response + * from the servers. + */ + void sendAndReceiveParameter( + ParameterUpdateMode updateMode, + ParameterType parameterType, + int64_t numSamples, + real cost, + bool sendBackParameter, + ParameterType sendBackParameterType = PARAMETER_VALUE, + ParameterType recvParameterType = PARAMETER_VALUE) { + sendAndReceiveParameter(updateMode, + parameterType, + allSegments_, + numSamples, + cost, + sendBackParameter, + sendBackParameterType, + recvParameterType); + } + + /** + * @brief Sends the segments in parameter to parameter servers. Each + * sendParameter() must be paired with a recvParameter() in the future. + * Only parameterType will be sent. + * + * @param[in] updateMode Indicates how parameters should be updated on the + * server side. + * @param[in] parameterType Type of parameter that will be sent. + * @param[in] segments Segments in the parameter that will be sent. + * @param[in] numSamples Number of samples this update is based on. + * @param[in] cost Cost of the batch, will be used to calculate global object + * value. + * @param[in] sendBackParameter True if the updated parameters should be sent + * back, otherwise false. + * @param[in] batchStatus Status of the batch. + * @note This function is non-blocking. This means that parameter should + * not change between this call and recvParameter() + */ + void sendParameter(ParameterUpdateMode updateMode, + ParameterType parameterType, + const std::vector& segments, + int64_t numSamples, + real cost, + bool sendBackParameter, + BatchStatus batchStatus); + + void recvParameter(); + + /** + * Sends all parameters to parameter servers, recvParameter() have to be + * invoked + * afterwards. + * + * @note This function is non-blocking. This means that if parameter should + * not changes between this call and recvParameter() + */ + void sendParameter(ParameterUpdateMode updateMode, + ParameterType parameterType, + int64_t numSamples, + real cost, + bool sendBackParameter, + BatchStatus batchStatus) { + sendParameter(updateMode, + parameterType, + allSegments_, + numSamples, + cost, + sendBackParameter, + batchStatus); + } + + /// Get all parameters from parameter servers + void getParameter(ParameterType recvParameterType = PARAMETER_VALUE, + ParameterType sendBackParameterType = PARAMETER_VALUE) { + sendAndReceiveParameter(PSERVER_UPDATE_MODE_GET_PARAM, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true, // sendBackParameter = true + sendBackParameterType, + recvParameterType); + } + + /// Get parameters by sparse row ids from parameter servers + void getParameterSparse( + ParameterType recvParameterType = PARAMETER_VALUE, + ParameterType sendBackParameterType = PARAMETER_VALUE) { + sendAndReceiveParameter(PSERVER_UPDATE_MODE_GET_PARAM_SPARSE, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true, // sendBackParameter = true + sendBackParameterType, + recvParameterType); + } + + /// Set all parameters on parameter servers using the local parameters + void setParameter() { + sendAndReceiveParameter(PSERVER_UPDATE_MODE_SET_PARAM, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + false); // sendBackParameter = false + } + /** + * Set all parameters on parameter servers, values will be zero + * means do not sending local parameters + */ + void setParameterZero() { + sendAndReceiveParameter(PSERVER_UPDATE_MODE_SET_PARAM_ZERO, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + false); // sendBackParameter = false + } + + /** + * @brief Wait until all gradient servers start one pass. + * + * @note This is now only used by the gradient servers for "sgd" + * algorithm. Calling this function means that the calling gradient + * server is ready to start a new pass. + */ + void waitPassStart(); + + /** + * @brief Wait until all gradient servers finish one pass. + * + * @note This is now only used by the gradient servers for "sgd" algorithm. + * Calling this function means that the calling gradient server + * finishes one pass. + */ + void waitPassFinish(); + + /// Wait until all gradient servers call this function. + void synchronize(SyncObject syncObjectId = SYNC_DEFAULT); + + /// Called when async-sgd finish pass. + void asyncFinishPass(SyncObject syncObjectId = SYNC_DEFAULT); + + void asyncStartPass(SyncObject syncObjectId = SYNC_DEFAULT) { + return synchronize(syncObjectId); + } + + /** + * @brief Execute the prepared operations on pservers, fetch the results and + * aggregate results from different pservers. + * @param[in] ops Prepared operations that will be executed on pservers. + * @param[in] waitForGradient If true, wait for gradient to be ready before + * starting the operations. + * @param[in] sendBackParameter If true, send back the parameter to clients + * after the operations are finished. + * @param[in] If true, and if all clients call waitPassFinish, signal all + * clients finish the pass. + */ + void doOperation(PreparedOperations& ops, + bool waitForGradient, + bool sendBackParameter, + bool releasePass = true); + + /** + * Set the configuration of pserver, including parameter config and + * optimization config + */ + void setConfig(const OptimizationConfig& optConfig, + const std::string& saveDir = "", + bool isSparseServer = false); + + /// Return true if all pservers are in the given status + bool inStatus(PServerStatus status); + bool isPassFinish() { return passFinish_; } + + /// Set pserver status + void setStatus(PServerStatus status); + + /** + * @brief Wait until all pservers are at status + * @note This function is not suitable for frequent use, + * because it sleeps 1 second each time when condition is satisfied. + */ + void waitForStatus(PServerStatus status); + + /// Create a column vector. The size is the dimension of parameter. + PServerVector createVector(); + + /// Release the PServerVector given handle. + void releaseVector(PServerVector handle); + + /** + * Create a column major matrix. The number of rows is the dimension of + * parameter. The number of columns is specifed by numCols. + */ + PServerMatrix createMatrix(int32_t numCols); + + /// Release the PServerMatrix given handle. + void releaseMatrix(PServerMatrix handle); + + // Some basic algebra functions + /// Calculate the dot product of u and v + real vectorDotProduct(PServerVector u, PServerVector v); + + /// Scale u by a + void vectorScale(PServerVector u, real a); + + /// Copy from src to dest + void vectorCopy(PServerVector src, PServerVector dst); + + /// u += v * a + void vectorAddMult(PServerVector u, PServerVector v, real a); + + /// u = v + w * a + void vectorAddMultInto(PServerVector u, + PServerVector v, + PServerVector w, + real a); + /// u = v * a + void vectorScaleInto(PServerVector u, PServerVector v, real a); + + /// Return pserver parameter value. + PServerVector getPServerParameterValue() { + PServerVector vec; + vec.handle = PARAMETER_VALUE; + return vec; + } + + /// Return pserver parameter gradient. + PServerVector getPServerParameterGradient() { + PServerVector vec; + vec.handle = PARAMETER_GRADIENT; + return vec; + } + + /** + * Tell pservers to load value vector from file. + * + * @param[in] dirName The directory that contains the value vector file. + */ + void loadValueVector(const std::string& dirName); + + /// Tell pservers to save value vector to file. + void saveValueVector(const std::string& dirName); + + void setTrainerId(int trainerId) { trainerId_ = trainerId; } + +#ifndef PADDLE_DISABLE_TIMER + void setForwardbackwardTime(uint64_t delta) { forwardbackwordTime_ = delta; } +#endif + + protected: + template + void multiCall(const char* funcName, + const ProtoIn& request, + std::vector* responses) { + responses->resize(clients_.size()); + size_t numClients = clients_.size(); + for (size_t i = 0; i < numClients; ++i) { + clients_[i].send(funcName, request); + } + for (size_t i = 0; i < numClients; ++i) { + clients_[i].recv(&(*responses)[i]); + } + } + + private: + void destroy(); + + /** + * @brief management function for parallelizing send/recv all connections + * to all pservers. it is called under one SyncThreadPool. it + * supports to use N thread to control M connections. the receiving + * actions can be started until all sending action to all connections + * owned by current thread are finished. Different connections + * controlled + * by different threads can transfer data asynchronously. + */ + void sendParallel(int tid, + size_t numThreads, + ParameterType recvParameterType); + /// sending thread routine for asynchronously send data + void send(int threadId); + /// receiving thread routing for asynchronously receive data + void recv(int threadId); + + /** + * @brief main routine to build data for pserver + * + * @note it can prepare different kinds of parameter type data. it can + * be regarded as layer for bridging real parameters data and + * protobuf data for communication. + * TODO(yanfei): + * can abstract additional layer to encode and decode data to/from + * protobuf data. + */ + void prepareSendData( + ParameterUpdateMode updateMode, + ParameterType parameterType, // client send type + const std::vector& parameterSegments, + int64_t numSamples, + real cost, + bool sendBackParameter, + ParameterType sendBackParameterType, // send back type in pserver + BatchStatus batchStatus, + SendJob* sendJob); + + /// start necessary threads for threadPool + void initThreads(); + + protected: + /// start port number of pserver + /// it deduce all ports for dense and sparse with some rules + int port_; + /// identify the trainer id using this client + int trainerId_; + +#ifndef PADDLE_DISABLE_TIMER + uint64_t forwardbackwordTime_; +#endif + std::mutex sparseAutoGrowthMutex_; + + /// map id to parameter used for decoding protobuf data + std::unordered_map parameterMap_; + /// segments for all parameters that needed to sync + std::vector allSegments_; + + /// module for sensing sparse parameters distribution on all pservers + std::unique_ptr sparseDistribution_; + + /// thread pool for parallelizing all connections to pservers + std::unique_ptr syncThreadPool_; + + bool passFinish_; +}; + +} // namespace paddle diff --git a/paddle/legacy/pserver/ParameterServer2.cpp b/paddle/legacy/pserver/ParameterServer2.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8533a322d92d292ee613d44795cf60462082a11b --- /dev/null +++ b/paddle/legacy/pserver/ParameterServer2.cpp @@ -0,0 +1,1401 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ParameterServer2.h" + +#include +#include + +#include "paddle/legacy/math/SIMDFunctions.h" +#include "paddle/legacy/parameter/AverageOptimizer.h" +#include "paddle/legacy/parameter/FirstOrderOptimizer.h" +#include "paddle/legacy/parameter/OptimizerFunctions.h" +#include "paddle/legacy/parameter/OptimizerWithRegularizer.h" +#include "paddle/legacy/parameter/ParameterOptimizer.h" +#include "paddle/legacy/parameter/ParameterUpdateFunctions.h" +#include "paddle/legacy/parameter/Regularizer.h" +#include "paddle/legacy/parameter/ThreadLocalBuffer.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/StringUtil.h" + +DEFINE_int32(pserver_num_threads, 1, "number of threads for sync op exec"); +DEFINE_double(async_lagged_ratio_min, + 1.0, + "control config_.async_lagged_grad_discard_ratio() min value"); +DEFINE_double( + async_lagged_ratio_default, + 1.5, + "if async_lagged_grad_discard_ratio is not set in trainer_config.conf" + "use it as defalut value"); + +namespace paddle { + +const std::string ParameterServer2::kRetMsgInvalidMatrixHandle = + "Invalid matrix handle"; +const std::string ParameterServer2::kRetMsgInvalidVectorHandle = + "Invalid vector handle"; +const std::string ParameterServer2::kRetMsgUnknownOperation = + "Unknown operation"; + +ParameterServer2::ParameterServer2(const std::string& addr, + int port, + int rdmaCpu) + : ProtoServer(addr, port, rdmaCpu), + dataSize_(0), + size_(0), + gradientReadyBarrier_(FLAGS_num_gradient_servers + 1), + parameterReadyBarrier_(FLAGS_num_gradient_servers + 1), + passBarrier_(FLAGS_num_gradient_servers + 1), + numPassFinishClients_(0), + allClientPassFinish_(false), + serverId_(-1), + batchId_(-1) { + /** + * register function for remote client calling, these functions + * will be mapped to a data structure for quick looking up. each + * request from trainer can contains one function name to indicate + * remote action. this architecture looks like rpc style for pserver. + */ + REGISTER_SERVICE_FUNCTION_EX(ParameterServer2, sendParameter); + REGISTER_SERVICE_FUNCTION_EX(ParameterServer2, sendData); + REGISTER_SERVICE_FUNCTION(ParameterServer2, setConfig); + REGISTER_SERVICE_FUNCTION(ParameterServer2, setStatus); + REGISTER_SERVICE_FUNCTION(ParameterServer2, getStatus); + REGISTER_SERVICE_FUNCTION(ParameterServer2, doOperation); + REGISTER_SERVICE_FUNCTION(ParameterServer2, createVector); + REGISTER_SERVICE_FUNCTION(ParameterServer2, releaseVector); + REGISTER_SERVICE_FUNCTION(ParameterServer2, createMatrix); + REGISTER_SERVICE_FUNCTION(ParameterServer2, releaseMatrix); + REGISTER_SERVICE_FUNCTION(ParameterServer2, waitPassStart); + REGISTER_SERVICE_FUNCTION(ParameterServer2, waitPassFinish); + REGISTER_SERVICE_FUNCTION(ParameterServer2, synchronize); + REGISTER_SERVICE_FUNCTION(ParameterServer2, asyncFinishPass); + REGISTER_SERVICE_FUNCTION(ParameterServer2, loadValueVector); + REGISTER_SERVICE_FUNCTION(ParameterServer2, saveValueVector); + + /// thread pool for parallelizing some computations + if (FLAGS_pserver_num_threads > 1) { + syncThreadPool_.reset(new SyncThreadPool(FLAGS_pserver_num_threads, false)); + } +} + +bool ParameterServer2::init() { + vectors_.resize(NUM_PARAMETER_TYPES); + configMap_.clear(); + + numSamplesProcessed_ = 0; + cost_ = 0; + char* mpienv = getenv("OMPI_COMM_WORLD_SIZE"); + if (mpienv != NULL) { + mpiSize_ = atoi(mpienv); + } else { + mpiSize_ = 1; + } + status_ = PSERVER_STATUS_NOT_SET; + dataMems_.resize(FLAGS_num_gradient_servers); + synchronizeBarriers_.resize(SyncObject_ARRAYSIZE); + for (auto& barrier : synchronizeBarriers_) { + barrier.reset(new ThreadBarrier(FLAGS_num_gradient_servers)); + } + + // initialization for dicarding lagging gradient + asyncUpdateSteps_ = 0; + asyncTrainerSteps_.resize(FLAGS_num_gradient_servers); + asyncTrainerSteps_.assign(asyncTrainerSteps_.size(), 0); + asyncLaggedGradientsNum_ = 0; + asyncUpdateStat_.resize(static_cast(FLAGS_num_gradient_servers * + FLAGS_async_lagged_ratio_default)); + asyncUpdateStat_.assign(asyncUpdateStat_.size(), 0); + asyncTrainerDiscardStat_.resize(FLAGS_num_gradient_servers); + asyncTrainerDiscardStat_.assign(asyncTrainerDiscardStat_.size(), 0); + asyncTrainerCommitStat_.resize(FLAGS_num_gradient_servers); + asyncTrainerCommitStat_.assign(asyncTrainerCommitStat_.size(), 0); + + return true; +} + +void ParameterServer2::getStatus(const GetStatusRequest& request, + ProtoResponseCallback callback) { + (void)request; + GetStatusResponse response; + response.set_status(status_); + callback(response); +} + +void ParameterServer2::setStatus(const SetStatusRequest& request, + ProtoResponseCallback callback) { + status_ = request.status(); + SetStatusResponse response; + callback(response); +} + +void ParameterServer2::setConfig(const SetConfigRequest& request, + ProtoResponseCallback callback) { + { + std::lock_guard guard(parameterMutex_); + + serverId_ = request.server_id(); + isSparseServer_ = request.is_sparse_server(); + + if (!request.save_dir().empty()) { + mkDir(request.save_dir().c_str()); + } + + for (const auto& config : request.param_configs()) { + CHECK(!configMap_.count(config.para_id())) + << "Duplicated parameter name: " << config.name(); + configMap_[config.para_id()] = config; + CHECK_EQ(config.sparse_remote_update(), isSparseServer_); + } + + config_ = request.opt_config(); + if (config_.algorithm() == TrainAlgorithm::AsyncSGD) { + auto asyncLaggedRatio = config_.async_lagged_grad_discard_ratio(); + if (asyncLaggedRatio <= FLAGS_async_lagged_ratio_min) { + LOG(INFO) << "WARNING: async_lagged_grad_discard_ratio is too small" + << "reset to default, async_lagged_grad_discard_ratio = " + << FLAGS_async_lagged_ratio_default; + asyncLaggedRatio = FLAGS_async_lagged_ratio_default; + } + asyncLaggedThreshold_ = + static_cast(FLAGS_num_gradient_servers * asyncLaggedRatio); + LOG(INFO) << "discard lagged async gradient ratio: " << asyncLaggedRatio + << " asyncLaggedhreshold: " << asyncLaggedThreshold_; + } + if (isSparseServer_ && config_.num_batches_per_send_parameter() > 1) { + /// sparse server must NOT use local update mode + config_.set_num_batches_per_send_parameter(1); + } + + if (config_.num_batches_per_send_parameter() > 1 && + config_.center_parameter_update_method() == "average") { + /// scaling L1/L2 decay rate as large as L1/L2 apply in trainer + /// if parameter regularization in pserver + for (auto& pair : configMap_) { + ParameterConfig& config = pair.second; + if (config_.num_batches_per_send_parameter() == + config.num_batches_regularization()) { + real scale = + config_.delta_add_rate() * config.num_batches_regularization(); + if (config_.algorithm() == "sgd") { + scale *= FLAGS_num_gradient_servers; + } + config.set_decay_rate(config.decay_rate() * scale); + if (config.decay_rate() > 0.1f) { + LOG(FATAL) << "L2 decay=" << config.decay_rate() + << " for parameter:" << config.name() + << " is too large after scale in pserver!"; + } + config.set_decay_rate_l1(config.decay_rate_l1() * scale); + if (config.decay_rate_l1() > 0.1f) { + LOG(FATAL) << "L1 decay=" << config.decay_rate_l1() + << " for parameter:" << config.name() + << " is too large after scale in pserver!"; + } + + LOG(INFO) << "parameter:" << config.name() + << " decay apply in pserver," + << " L1 decay=" << config.decay_rate_l1() + << " L2 decay=" << config.decay_rate(); + } + } + } + } + + SetConfigResponse response; + callback(response); +} + +real bufferSum(const std::vector& buffers) { + real sum = 0; + for (const auto buffer : buffers) { + for (size_t i = 0; i < buffer.size; ++i) { + sum += buffer.base[i]; + } + } + return sum; +} + +void ParameterServer2::mergeSegments(BlockSegments* segments) { + if (segments->empty()) { + return; + } + std::sort(segments->begin(), segments->end()); + auto curr = segments->begin(); + for (auto it = segments->begin(); it != segments->end(); ++it) { + if (it->first <= curr->second) { + curr->second = std::max(curr->second, it->second); + } else { + ++curr; + *curr = *it; + } + } + ++curr; + segments->erase(curr, segments->end()); +} + +void ParameterServer2::setParameter(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + (void)response; + (void)outputBuffers; + LOG(INFO) << "pserver: setParameter"; + std::lock_guard guard(parameterMutex_); + + int64_t numBlocks = blockIdMap_.size(); + CHECK_EQ(blockIdMap_.size(), blockOffsetMap_.size()); + /// total bytes for all the added blocks + int64_t totalSize = size_; + std::vector offsets; + offsets.reserve(request.blocks_size()); + std::vector blockIds; + blockIds.reserve(request.blocks_size()); + int bufferIndex = 0; + + if (!request.blocks().size()) { + LOG(WARNING) + << "--ports_num or --ports_num_for_sparse might be too large, " + << "or total dense parameter size or sparse parameters size " + << "might be too small, this psever doesn't store any parameter."; + return; + } + + for (const auto& block : request.blocks()) { + /// block size for parameter(e.g. 128 for sparse row, 1K for dense) + uint64_t blockSize = getParameterConfig(block).parameter_block_size(); + BlockKey key(block.para_id(), block.block_id()); + if (inputBuffers.size()) { // if !=PSERVER_UPDATE_MODE_SET_PARAM_ZERO + Buffer buffer = inputBuffers[bufferIndex]; + ++bufferIndex; + CHECK_EQ(buffer.size, block.block_size()) + << "data size is too big:" + << " block_size=" << block.block_size() + << " data_size=" << buffer.size; + } + + /// add a new block + if (blockIdMap_.count(key) == 0) { + blockOffsetMap_[key] = totalSize; + blockIdMap_[key] = numBlocks; + ++numBlocks; + totalSize += blockSize; + } + offsets.push_back(blockOffsetMap_[key]); + blockIds.push_back(blockIdMap_[key]); + } + + size_ = totalSize; + LOG(INFO) << "pserver: new cpuvector: size=" << size_; + if (!vectors_[PARAMETER_VALUE]) { + /// vectors_ + const auto types = sgdOptimizerGetTypes(config_, true /*inPserver*/); + for (const auto type : types) { + vectors_[type].reset(new CpuVector(size_)); + vectors_[type]->zeroMem(); + } + + blockInfos_.resize(numBlocks); + for (auto& info : blockInfos_) { + info.lock.reset(new std::mutex()); + } + } else { + CHECK_EQ((size_t)size_, vectors_[PARAMETER_VALUE]->getSize()) + << "Currently adding new blocks is not supported. " + << "All blocks must be added in one setParameter call"; + } + + VectorPtr buf = vectors_[PARAMETER_VALUE]; + usedSegments_.reserve(offsets.size()); + /// if offsets is empty, means parameter_block_size is too big or too many + /// nodes. + if (offsets.empty()) { + LOG(WARNING) << "in setParameter: offsets is empty"; + } + for (size_t i = 0; i < offsets.size(); ++i) { + size_t blockId = blockIds[i]; + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(request.blocks(i)); + info.config = &config; + info.offset = offsets[i]; + info.optimizer.reset(sgdOptimizerCreate( + config_, config, config.sparse_remote_update(), true /*inPserver*/)); + if (config.sparse_remote_update()) { + size_t width = config.dims(1); + CHECK_EQ(config.parameter_block_size(), width) + << "block size: " << config.parameter_block_size() + << "width : " << width; + } + info.optimizer->init(1, info.config); + usedSegments_.push_back(std::make_pair( + offsets[i], offsets[i] + request.blocks(i).block_size())); + } + mergeSegments(&usedSegments_); + + if (request.update_mode() == PSERVER_UPDATE_MODE_SET_PARAM) { + /// copy param from trainer + for (size_t i = 0; i < offsets.size(); ++i) { + Buffer buffer = inputBuffers[i]; + real* start = buf->getPoint(offsets[i]); + CHECK_LE(offsets[i] + buffer.size, buf->getSize()); + memcpy(start, buffer.base, sizeof(real) * buffer.size); + } + } else { + CHECK(request.update_mode() == PSERVER_UPDATE_MODE_SET_PARAM_ZERO); + /// nothing to do, value vector zero mem already + } +} + +void ParameterServer2::addGradient(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + VLOG(1) << "pserver: addGradient"; + + { + ReadLockGuard guard(parameterMutex_); + int bufferIndex = 0; + for (const auto& block : request.blocks()) { + int64_t offset = getBlockOffset(block); + CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + + int64_t blockId = getBlockId(block); + CHECK_GE(blockId, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + + Buffer buffer = inputBuffers[bufferIndex]; + ++bufferIndex; + + const real* gradientBuffer = buffer.base; + real* gradientSumBuffer = vectors_[PARAMETER_GRADIENT]->getPoint(offset); + + size_t size = buffer.size; + + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(blockId); + if (config.sparse_remote_update()) { + CHECK_EQ(size, config.parameter_block_size()); + } else { // dense + CHECK_LE(size, config.parameter_block_size()); + } + std::lock_guard guard(*info.lock); + simd::addTo(gradientSumBuffer, gradientBuffer, size); + } + } + if (request.batch_status() == BATCH_FINISH || + request.batch_status() == BATCH_START_AND_FINISH) { + numSamplesProcessed_ += request.num_samples(); + cost_ += request.cost(); + VLOG(1) << "num samples: " << numSamplesProcessed_ + << ", new cost:" << cost_; + + /// notify doOperation gradient ready + gradientReadyBarrier_.wait(); + + /// wait doOperation finish + parameterReadyBarrier_.wait(); + VLOG(1) << "start send back"; + } +} + +bool ParameterServer2::asyncGrdientCommitCheckAndStat( + const SendParameterRequest& request) { + const auto trainerId = request.trainer_id(); + int64_t trainerSteps = asyncTrainerSteps_[trainerId]; + CHECK_GE(asyncUpdateSteps_, trainerSteps) + << " async update steps overflows " + << " trainer id: " << trainerId + << " async update steps in pserver: " << asyncUpdateSteps_ + << " async update steps in request: " << trainerSteps; + + asyncUpdateSteps_++; + bool commitGradient = true; + + int64_t delta = asyncUpdateSteps_ - trainerSteps; + if (delta >= asyncLaggedThreshold_) { + VLOG(1) << "discard Async Update: " + << " trainer id: " << trainerId + << " pserver steps: " << asyncUpdateSteps_ + << " request steps: " << trainerSteps; + asyncLaggedGradientsNum_++; + commitGradient = false; + } + /// stat on lagged steps, to get total discard distribution + if (static_cast(delta) < asyncUpdateStat_.size()) { + asyncUpdateStat_[delta]++; + } else { + asyncUpdateStat_[asyncUpdateStat_.size() - 1]++; + } + /// stat on trainerId and discard, to get trainer condition + if (commitGradient) { + asyncTrainerCommitStat_[trainerId]++; + } else { + asyncTrainerDiscardStat_[trainerId]++; + } + + return commitGradient; +} + +static ThreadLocal> localBlockBitset_; + +void ParameterServer2::asyncSGD(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + int64_t numBlocks = blockIdMap_.size(); + auto& localBlockBitset = *localBlockBitset_; + + if (isSparseServer_) { + if (localBlockBitset.empty()) { + localBlockBitset.resize(numBlocks); + } + localBlockBitset.assign(numBlocks, false); + } + + ReadLockGuard guard(parameterMutex_); + + if (request.send_back_parameter()) { + outputBuffers->reserve(request.blocks_size()); + } + + bool commitGradient = asyncGrdientCommitCheckAndStat(request); + + VectorPtr* vecs = parameter::getThreadLocalBuffer(); + size_t bufferIndex = 0; + for (const auto& block : request.blocks()) { + int64_t offset = getBlockOffset(block); + CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + int64_t blockId = getBlockId(block); + CHECK_GE(blockId, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + Buffer buffer = inputBuffers[bufferIndex]; + ++bufferIndex; + + size_t size = buffer.size; + + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(blockId); + + std::lock_guard guard(*info.lock); + /// gradients are too obsolete, will be discarded + if (commitGradient) { + info.optimizer->startBatch(numSamplesProcessed_); + + for (const auto type : info.optimizer->getParameterTypes()) { + vecs[type]->subVecFrom(*vectors_[type], offset, size); + } + vecs[PARAMETER_GRADIENT]->subVecFrom(buffer.base, 0, size); + info.optimizer->update(vecs, config, isSparseServer_ ? 0 : -1); + + if (auto callback = info.optimizer->needSpecialTraversal(config)) { + blockTraverse(info, config, offset, size, vecs, callback); + } + info.optimizer->finishBatch(); + } + + if (commitGradient && isSparseServer_) { + localBlockBitset[blockId] = true; + } + + if (!isSparseServer_ && request.send_back_parameter()) { // dense + int type = request.send_back_parameter_type(); + sendBackParameter(block, type, response, &buffer, outputBuffers); + } + } /// foreach block + + asyncTrainerSteps_[request.trainer_id()] = asyncUpdateSteps_; + + if (commitGradient && isSparseServer_) { + /// find blocks that trainer do not request update + for (int64_t blockId = 0; blockId < numBlocks; ++blockId) { + if (localBlockBitset[blockId]) { + continue; + } + + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = *info.config; + size_t size = config.parameter_block_size(); + + std::lock_guard guard(*info.lock); + info.optimizer->startBatch(numSamplesProcessed_); + if (auto callback = info.optimizer->needSpecialTraversal(config)) { + blockTraverse(info, config, info.offset, size, vecs, callback); + } + info.optimizer->finishBatch(); + } + } + + if (commitGradient && (request.batch_status() == BATCH_FINISH || + request.batch_status() == BATCH_START_AND_FINISH)) { + numSamplesProcessed_ += request.num_samples(); + } + + /// show some performance log if needed + if (request.trainer_id() == 0) { + /// batchId_ is approximately equal to "real batchId_" + batchId_++; + } +} + +void ParameterServer2::getParameter(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + (void)inputBuffers; + LOG(INFO) << "pserver: getParameter"; + ReadLockGuard guard(parameterMutex_); + for (const auto& block : request.blocks()) { + int type = request.send_back_parameter_type(); + sendBackParameter(block, type, response, outputBuffers); + } +} + +void ParameterServer2::getParameterSparse(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + (void)inputBuffers; + auto& buffer = *readWriteBuffer_; + size_t numReals = 0; + for (const auto& block : request.blocks()) { + numReals += getParameterConfig(block).dims(1); + } + buffer.resize(numReals); + + VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals; + + ReadLockGuard guard(parameterMutex_); + size_t offset = 0; + for (const auto& block : request.blocks()) { + size_t width = getParameterConfig(block).dims(1); + Buffer buf = {buffer.data() + offset, width}; + int type = request.send_back_parameter_type(); + sendBackParameterSparse(block, type, response, &buf, width, outputBuffers); + offset += width; + } +} + +void ParameterServer2::sendBackParameter(const ParameterBlock& block, + int parameterType, + SendParameterResponse* response, + std::vector* outputBuffers) { + ParameterBlock* returnBlock = response->add_blocks(); + returnBlock->set_para_id(block.para_id()); + returnBlock->set_block_id(block.block_id()); + returnBlock->set_begin_pos(block.begin_pos()); + returnBlock->set_block_size(block.block_size()); + + int64_t offset = getBlockOffset(block); + CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + + real* valueBuffer = vectors_[parameterType]->getPoint(offset); + outputBuffers->push_back({valueBuffer, (size_t)block.block_size()}); +} + +void ParameterServer2::sendBackParameter(const ParameterBlock& block, + int parameterType, + SendParameterResponse* response, + Buffer* buffer, + std::vector* outputBuffers) { + ParameterBlock* returnBlock = response->add_blocks(); + returnBlock->set_para_id(block.para_id()); + returnBlock->set_block_id(block.block_id()); + returnBlock->set_begin_pos(block.begin_pos()); + returnBlock->set_block_size(block.block_size()); + + int64_t offset = getBlockOffset(block); + CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + + size_t size = buffer->size; + real* valueBuffer = vectors_[parameterType]->getPoint(offset); + /// copy to second buffer to avoid to be polluted by other request + memcpy(buffer->base, valueBuffer, sizeof(real) * size); + outputBuffers->push_back({buffer->base, size}); +} + +void ParameterServer2::sendBackParameterSparse( + const ParameterBlock& block, + int parameterType, + SendParameterResponse* response, + Buffer* buffer, + size_t width, + std::vector* outputBuffers) { + ParameterBlock* returnBlock = response->add_blocks(); + returnBlock->set_para_id(block.para_id()); + returnBlock->set_block_id(block.block_id()); + returnBlock->set_begin_pos(block.begin_pos()); + returnBlock->set_block_size(block.block_size()); + int64_t offset = getBlockOffset(block); + CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " + << " id=" << block.para_id() + << " block id=" << block.block_id(); + + real* valueBuffer = vectors_[parameterType]->getPoint(offset); + CHECK_EQ(buffer->size, width); + memcpy(buffer->base, valueBuffer, width * sizeof(real)); + outputBuffers->push_back(*buffer); +} + +void ParameterServer2::readAllBlocks( + MsgReader* msgReader, std::vector* buffers) { + auto& buffer = *readWriteBuffer_; + size_t numBlocks = msgReader->getNumBlocks(); + buffer.resizeWithAlignHints(msgReader->getTotalLength() / sizeof(real), + numBlocks); + std::vector bufs(numBlocks); + buffers->clear(); + buffers->reserve(numBlocks); + buffer.resetAlignAlloc(); + for (size_t i = 0; i < numBlocks; ++i) { + size_t len = msgReader->getBlockLength(i); + CHECK_EQ(len % sizeof(real), (size_t)0); + size_t size = len / sizeof(real); + bufs[i] = buffer.nextBlock(size); + buffers->push_back({(real*)bufs[i], size}); + } + msgReader->readBlocks(bufs); +} + +void ParameterServer2::sendParameter(const SendParameterRequest& request, + std::unique_ptr msgReader, + ProtoResponseCallbackEx callback) { + SendParameterResponse response; + std::vector inputBuffers; + std::vector outputBuffers; + readAllBlocks(msgReader.get(), &inputBuffers); + msgReader.reset(); + + switch (request.update_mode()) { + case PSERVER_UPDATE_MODE_SET_PARAM: + case PSERVER_UPDATE_MODE_SET_PARAM_ZERO: + setParameter(request, inputBuffers, &response, &outputBuffers); + break; + case PSERVER_UPDATE_MODE_GET_PARAM: + getParameter(request, inputBuffers, &response, &outputBuffers); + break; + case PSERVER_UPDATE_MODE_GET_PARAM_SPARSE: + getParameterSparse(request, inputBuffers, &response, &outputBuffers); + break; + case PSERVER_UPDATE_MODE_ASYNC_SGD: + asyncSGD(request, inputBuffers, &response, &outputBuffers); + break; + case PSERVER_UPDATE_MODE_ADD_GRADIENT: + addGradient(request, inputBuffers, &response, &outputBuffers); + break; + case PSERVER_UPDATE_MODE_AVERAGE_PARAMETER: + break; + } + switch (request.update_mode()) { + case PSERVER_UPDATE_MODE_ADD_GRADIENT: + (*requestVec_).push_back(request); + (*callbackVec_).push_back(callback); + if (request.batch_status() == BATCH_FINISH || + request.batch_status() == BATCH_START_AND_FINISH) { + for (size_t i = 0; i < (*requestVec_).size(); i++) { + ReadLockGuard guard(parameterMutex_); + SendParameterRequest& request = (*requestVec_)[i]; + SendParameterResponse responseTemp; + + std::vector outputIovs; + if (request.send_back_parameter()) { + CHECK(!isSparseServer_); + std::vector outputBuffersTemp; + for (const auto& block : request.blocks()) { + int type = request.send_back_parameter_type(); + sendBackParameter(block, type, &responseTemp, &outputBuffersTemp); + } + outputIovs.reserve(outputBuffersTemp.size()); + for (auto buffer : outputBuffersTemp) { + outputIovs.push_back({buffer.base, buffer.size * sizeof(real)}); + } + } + + ProtoResponseCallbackEx& callbackTemp = (*callbackVec_)[i]; + callbackTemp(responseTemp, outputIovs); + } + (*requestVec_).clear(); + (*callbackVec_).clear(); + } + break; + case PSERVER_UPDATE_MODE_SET_PARAM: + case PSERVER_UPDATE_MODE_SET_PARAM_ZERO: + case PSERVER_UPDATE_MODE_GET_PARAM: + case PSERVER_UPDATE_MODE_GET_PARAM_SPARSE: + case PSERVER_UPDATE_MODE_ASYNC_SGD: + case PSERVER_UPDATE_MODE_AVERAGE_PARAMETER: + std::vector outputIovs; + outputIovs.reserve(outputBuffers.size()); + for (auto buffer : outputBuffers) { + outputIovs.push_back({buffer.base, buffer.size * sizeof(real)}); + } + callback(response, outputIovs); + break; + } +} + +template +void ParameterServer2::reduceAndSendData(const SendDataRequest& request, + std::unique_ptr& msgReader, + ProtoResponseCallbackEx& callback) { + SendDataResponse response; + response.set_type(request.type()); + response.set_server_id(serverId_); + + auto sendData = reinterpret_cast(dataMems_[0].get()->getBuf()); + size_t rawMemSize = dataMems_[0].get()->getSize(); + CHECK_EQ(rawMemSize % sizeof(Dtype), 0U); + size_t dataMemSize = rawMemSize / sizeof(Dtype); + for (size_t i = 1; i < dataMems_.size(); ++i) { + CHECK_EQ(dataMems_[i].get()->getSize(), rawMemSize); + auto data = reinterpret_cast(dataMems_[i].get()->getBuf()); + for (size_t j = 0; j < dataMemSize; ++j) { + sendData[j] += data[j]; + } + } + std::vector outputIovs; + auto block = response.add_blocks(); + outputIovs.push_back({sendData, rawMemSize}); + block->set_total_size(rawMemSize); + block->set_data_size(sizeof(Dtype)); + callback(response, outputIovs); +} + +void ParameterServer2::templateReduceSum(const SendDataRequest& request, + std::unique_ptr& msgReader, + ProtoResponseCallbackEx& callback) { + const auto& block = request.blocks(0); + switch (block.data_type()) { + case TRANS_FLOAT: + reduceAndSendData(request, msgReader, callback); + break; + case TRANS_DOUBLE: + reduceAndSendData(request, msgReader, callback); + break; + case TRANS_INT32: + reduceAndSendData(request, msgReader, callback); + break; + case TRANS_UINT32_T: + reduceAndSendData(request, msgReader, callback); + break; + case TRANS_INT64_T: + reduceAndSendData(request, msgReader, callback); + break; + case TRANS_UINT64_T: + reduceAndSendData(request, msgReader, callback); + break; + default: + LOG(FATAL) << "not supported"; + break; + } +} + +void ParameterServer2::sendData(const SendDataRequest& request, + std::unique_ptr msgReader, + ProtoResponseCallbackEx callback) { + SendDataResponse response; + response.set_type(request.type()); + response.set_server_id(serverId_); + + switch (request.update_mode()) { + case DATA_UPDATE_MODE_SET_OWN: { + CHECK_EQ(msgReader->getNumBlocks(), (size_t)(request.blocks_size())); + size_t totalLen = msgReader->getTotalLength(); + if (totalLen > 0) { + CHECK_EQ(msgReader->getNumBlocks(), 1U) + << "Only one block currently support now!"; + const auto& block = request.blocks(0); + if (0 == dataSize_) { + dataSize_ = block.data_size(); + } else { + CHECK_EQ(dataSize_, block.data_size()); + } + int64_t serverId = request.server_id(); + if (serverId_ < 0) { + serverId_ = serverId; + } else { + CHECK_EQ(serverId_, serverId); + } + int64_t clientId = request.client_id(); + dataMems_[clientId] = std::make_shared(totalLen); + CHECK_EQ(totalLen % sizeof(block.data_size()), 0U); + msgReader->readNextBlock(dataMems_[clientId].get()->getBuf()); + } + msgReader.reset(); + std::vector outputIovs; + callback(response, outputIovs); + break; + } + case DATA_UPDATE_MODE_GET_ALL: { + /// Currently only support DATA_REDUCE_SUM + /// And their Operations are just add + CHECK(DATA_REDUCE_SUM == request.type()); + templateReduceSum(request, msgReader, callback); + break; + } + default: { LOG(FATAL) << "not supported"; } + } +} + +void ParameterServer2::clearUnusedSegments(CpuVector* vec) { + real* data = vec->getData(); + if (usedSegments_.empty()) { + return; + } + memset(data, 0, sizeof(real) * usedSegments_[0].first); + memset(data + usedSegments_.back().second, + 0, + sizeof(real) * (size_ - usedSegments_.back().second)); + size_t n = size_ - usedSegments_.back().second; + + for (size_t i = 1; i < usedSegments_.size(); ++i) { + memset( + data + usedSegments_[i - 1].second, + 0, + sizeof(real) * (usedSegments_[i].first - usedSegments_[i - 1].second)); + n += usedSegments_[i].first - usedSegments_[i - 1].second; + } +} + +void ParameterServer2::parallelExecForEachBlock(ExecFunc func) { + SyncThreadPool::execHelper( + syncThreadPool_.get(), [&](int tid, size_t numThreads) { + int64_t numBlocks = blockIdMap_.size(); + VectorPtr* vecs = parameter::getThreadLocalBuffer(); + for (int64_t blockId = tid; blockId < numBlocks; + blockId += numThreads) { + func(blockId, vecs); + } + }); +} + +void ParameterServer2::blockTraverse( + BlockInfo& info, + const ParameterConfig& config, + int64_t offset, + size_t size, + const VectorPtr vecs[], + const ParameterOptimizer::TraverseCallback& callback) { + /// setup sub bufs + for (const auto type : info.optimizer->getParameterTypes()) { + vecs[type]->subVecFrom(*vectors_[type], offset, size); + } + callback(vecs, config, config.sparse_remote_update() ? 0 : -1LU); +} + +void ParameterServer2::op_SGD(const Operation& operation, + OperationResult* result) { + (void)operation; + (void)result; + + if (allClientPassFinish_) { + /// when all clients signal pass finished, the update + /// is empty. + return; + } + + { + parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(blockId); + int64_t offset = info.offset; + size_t size = config.parameter_block_size(); + + info.optimizer->startBatch(numSamplesProcessed_); + + for (const auto type : info.optimizer->getParameterTypes()) { + vecs[type]->subVecFrom(*vectors_[type], offset, size); + } + info.optimizer->update( + vecs, config, config.sparse_remote_update() ? 0 : -1LU); + vecs[PARAMETER_GRADIENT]->zeroMem(); + + if (auto callback = info.optimizer->needSpecialTraversal(config)) { + blockTraverse(info, config, offset, size, vecs, callback); + } + info.optimizer->finishBatch(); + }); + } + + batchId_++; +} + +void ParameterServer2::op_start_pass(const Operation& operation, + OperationResult* result) { + (void)operation; + (void)result; + + parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { + BlockInfo& info = blockInfos_[blockId]; + info.optimizer->startPass(); + }); +} + +void ParameterServer2::op_finish_pass(const Operation& operation, + OperationResult* result) { + (void)operation; + (void)result; + + parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(blockId); + size_t size = config.parameter_block_size(); + + /// catch up with + if (auto callback = info.optimizer->startCatchUpWith()) { + blockTraverse(info, config, info.offset, size, vecs, callback); + info.optimizer->finishCatchUpWith(); + } + + /// finish pass + info.optimizer->finishPass(); + }); + batchId_ = 0; +} + +void ParameterServer2::op_apply(const Operation& operation, + OperationResult* result) { + (void)operation; + (void)result; + + parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(blockId); + int64_t offset = info.offset; + size_t size = config.parameter_block_size(); + + // catch up with + if (auto callback = info.optimizer->startCatchUpWith()) { + blockTraverse(info, config, offset, size, vecs, callback); + info.optimizer->finishCatchUpWith(); + } + + // apply to PARAMETER_APPLY + if (auto callback = info.optimizer->apply()) { + blockTraverse(info, config, offset, size, vecs, callback); + } + }); +} + +void ParameterServer2::op_randomize(const Operation& operation, + OperationResult* result) { + LOG(INFO) << "ParameterServer2::op_randomize: serverId=" << serverId_; + + CpuVector& valueVec = *vectors_[PARAMETER_VALUE]; + + parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { + BlockInfo& info = blockInfos_[blockId]; + const ParameterConfig& config = getParameterConfig(blockId); + size_t size = config.parameter_block_size(); + + vecs[PARAMETER_VALUE]->subVecFrom(valueVec, info.offset, size); + Parameter::randomize(vecs[PARAMETER_VALUE], config); + }); +} + +void ParameterServer2::loadValueVector(const LoadValueRequest& request, + ProtoResponseCallback callback) { + LoadValueResponse response; + LOG(INFO) << "ParameterServer2::loadValueVector: serverId=" << serverId_; + + constexpr int kBufLen = 100; + char buf[kBufLen]; + snprintf(buf, kBufLen, "/pserver.%04d", static_cast(serverId_)); + std::string filename = request.dir_name() + buf; + + std::ifstream fs(filename, std::ios_base::binary); + CHECK(fs) << "Fail to open " << filename; + + CpuVector& vec = *vectors_[PARAMETER_VALUE]; + Parameter::Header header; + CHECK(fs.read(reinterpret_cast(&header), sizeof(header))) + << "Fail to read parameters in pserver"; + CHECK(Parameter::isHeaderFormatSupported(header.format)) + << "Incorrect format version: " << header.format; + CHECK_EQ(header.size, (size_t)size_) + << "The size (" << header.size << ") in the file does not match the size " + << "(" << size_ << ") of the pserver: " << serverId_; + CHECK_EQ(header.valueSize, sizeof(real)) << "Unsupported valueSize " + << header.valueSize; + CHECK(fs.read(reinterpret_cast(vec.getData()), + header.size * sizeof(real))); + + callback(response); +} + +void ParameterServer2::saveValueVector(const SaveValueRequest& request, + ProtoResponseCallback callback) { + SaveValueResponse response; + LOG(INFO) << "ParameterServer2::SaveValueVector: serverId=" << serverId_; + + mkDir(request.dir_name().c_str()); + + constexpr int kBufLen = 100; + char buf[kBufLen]; + snprintf(buf, kBufLen, "/pserver.%04d", static_cast(serverId_)); + std::string filename = request.dir_name() + buf; + + std::ofstream fs(filename, std::ios_base::binary); + CHECK(fs) << "Fail to open " << filename; + + CpuVector& vec = vectors_[PARAMETER_APPLY] ? *vectors_[PARAMETER_APPLY] + : *vectors_[PARAMETER_VALUE]; + Parameter::Header header; + // TODO(TJ): save param headerFormat_ + header.format = PARAM_FORMAT_ORIGINAL; + header.valueSize = sizeof(real); + header.size = size_; + + CHECK_EQ(header.size, vec.getSize()); + + CHECK(fs.write(reinterpret_cast(&header), sizeof(header))) + << "Fail to write parameter in pserver: " << serverId_; + + CHECK(fs.write(reinterpret_cast(vec.getData()), + header.size * sizeof(real))) + << "Fail to write parameter in pserver: " << serverId_; + + callback(response); +} + +void ParameterServer2::op_RESET(const Operation& operation, + OperationResult* result) { + (void)result; + CpuVector* u = vectors_[operation.pvectors(0)].get(); + u->reset(operation.scalars(0)); + clearUnusedSegments(u); +} + +void ParameterServer2::op_utv(const Operation& operation, + OperationResult* result) { + real* u = vectors_[operation.pvectors(0)]->getData(); + real* v = vectors_[operation.pvectors(1)]->getData(); + int64_t size = size_; + double sum = 0; + for (int64_t i = 0; i < size; ++i) { + sum += (double)u[i] * (double)v[i]; + } + result->add_scalars(sum); +} + +void ParameterServer2::op_au_bv(const Operation& operation, + OperationResult* result) { + (void)result; + real* u = vectors_[operation.pvectors(0)]->getData(); + real* v = vectors_[operation.pvectors(1)]->getData(); + int64_t size = size_; + real a = operation.scalars(0); + real b = operation.scalars(1); + for (int64_t i = 0; i < size; ++i) { + v[i] = a * u[i] + b * v[i]; + } +} + +void ParameterServer2::op_COPY(const Operation& operation, + OperationResult* result) { + (void)result; + real* u = vectors_[operation.pvectors(0)]->getData(); + real* v = vectors_[operation.pvectors(1)]->getData(); + int64_t size = size_; + for (int64_t i = 0; i < size; ++i) { + v[i] = u[i]; + } +} + +void ParameterServer2::op_au(const Operation& operation, + OperationResult* result) { + (void)result; + real* u = vectors_[operation.pvectors(0)]->getData(); + int64_t size = size_; + real a = operation.scalars(0); + for (int64_t i = 0; i < size; ++i) { + u[i] *= a; + } +} + +void ParameterServer2::op_au_bv_cw(const Operation& operation, + OperationResult* result) { + (void)result; + real* u = vectors_[operation.pvectors(0)]->getData(); + real* v = vectors_[operation.pvectors(1)]->getData(); + real* w = vectors_[operation.pvectors(2)]->getData(); + int64_t size = size_; + real a = operation.scalars(0); + real b = operation.scalars(1); + real c = operation.scalars(2); + for (int64_t i = 0; i < size; ++i) { + w[i] = a * u[i] + b * v[i] + c * w[i]; + } +} + +void ParameterServer2::op_make_steepest_desc_dir(const Operation& operation, + OperationResult* result) { + (void)result; + real* dir = vectors_[operation.pvectors(0)]->getData(); + real* grad = vectors_[operation.pvectors(1)]->getData(); + real* x = vectors_[operation.pvectors(2)]->getData(); + int64_t size = size_; + real l1weight = operation.scalars(0); + for (int64_t i = 0; i < size; ++i) { + if (x[i] < 0) { + dir[i] = -grad[i] + l1weight; + } else if (x[i] > 0) { + dir[i] = -grad[i] - l1weight; + } else { + if (grad[i] < -l1weight) { + dir[i] = -grad[i] - l1weight; + } else if (grad[i] > l1weight) { + dir[i] = -grad[i] + l1weight; + } else { + dir[i] = 0; + } + } + } +} + +void ParameterServer2::op_fix_dir_signs(const Operation& operation, + OperationResult* result) { + (void)result; + real* dir = vectors_[operation.pvectors(0)]->getData(); + real* steepestDescDir = vectors_[operation.pvectors(1)]->getData(); + int64_t size = size_; + for (int64_t i = 0; i < size; ++i) { + if (dir[i] * steepestDescDir[i] <= 0) { + dir[i] = 0; + } + } +} + +void ParameterServer2::op_fix_omega_signs(const Operation& operation, + OperationResult* result) { + (void)result; + real* x = vectors_[operation.pvectors(0)]->getData(); + real* newx = vectors_[operation.pvectors(1)]->getData(); + int64_t size = size_; + for (int64_t i = 0; i < size; ++i) { + if (x[i] * newx[i] < 0) { + newx[i] = 0; + } + } +} + +void ParameterServer2::op_dir_deriv(const Operation& operation, + OperationResult* result) { + real* dir = vectors_[operation.pvectors(0)]->getData(); + real* grad = vectors_[operation.pvectors(1)]->getData(); + real* x = vectors_[operation.pvectors(2)]->getData(); + int64_t size = size_; + real l1weight = operation.scalars(0); + double sum = 0; + for (int64_t i = 0; i < size; ++i) { + if (dir[i] != 0) { + if (x[i] < 0) { + sum += dir[i] * (grad[i] - l1weight); + } else if (x[i] > 0) { + sum += dir[i] * (grad[i] + l1weight); + } else if (dir[i] < 0) { + sum += dir[i] * (grad[i] - l1weight); + } else if (dir[i] > 0) { + sum += dir[i] * (grad[i] + l1weight); + } + } + } + result->add_scalars(sum); +} + +void ParameterServer2::op_cost(const Operation& operation, + OperationResult* result) { + real* x = vectors_[operation.pvectors(0)]->getData(); + real* newgrad = vectors_[operation.pvectors(1)]->getData(); + int64_t size = size_; + real l1weight = operation.scalars(0); + real l2weight = operation.scalars(1); + double cost_real = cost_ / mpiSize_; + double sum_weight_l1 = 0; + double sum_weight_l2 = 0; + for (int64_t i = 0; i < size; ++i) { + sum_weight_l1 += std::abs(x[i]); + sum_weight_l2 += x[i] * x[i]; + newgrad[i] += 2.0 * l2weight * x[i]; + } + cost_real += l1weight * sum_weight_l1 + l2weight * sum_weight_l2; + result->add_scalars(cost_real); +} + +ParameterServer2::OperatorFunction ParameterServer2::opFuncs[] = { + nullptr, // PSERVER_OP_utu = 0; + &ParameterServer2::op_utv, // PSERVER_OP_utv = 1; + &ParameterServer2::op_au, // PSERVER_OP_au = 2; + &ParameterServer2::op_au_bv, // PSERVER_OP_au_bv = 3; + nullptr, // PSERVER_OP_aAx_bu = 4; + &ParameterServer2::op_SGD, // PSERVER_OP_SGD = 5; + &ParameterServer2::op_RESET, // PSERVER_OP_RESET = 6; + &ParameterServer2::op_COPY, // PSERVER_OP_COPY = 7; + &ParameterServer2::op_au_bv_cw, // PSERVER_OP_au_bv_cw = 8; + &ParameterServer2::op_make_steepest_desc_dir, + /// PSERVER_OP_MAKE_STEEPEST_DESC_DIR = 9; + &ParameterServer2::op_fix_dir_signs, // PSERVER_OP_FIX_SIGNS = 10; + &ParameterServer2::op_dir_deriv, // PSERVER_OP_DIR_DERIV = 11; + &ParameterServer2::op_fix_omega_signs, // PSERVER_OP_FIX_OMEGA_SIGNS = 12; + &ParameterServer2::op_cost, // PSERVER_OP_COST = 13 + &ParameterServer2::op_start_pass, // PSERVER_OP_START_PASS = 14 + &ParameterServer2::op_finish_pass, // PSERVER_OP_FINISH_PASS = 15 + &ParameterServer2::op_randomize, // PSERVER_OP_RANDOMIZE = 16 + &ParameterServer2::op_apply, // PSERVER_OP_APPLY = 17 +}; + +void ParameterServer2::doOperation(const DoOperationRequest& request, + ProtoResponseCallback callback) { + if (request.wait_for_gradient()) { + /// wait gradient update + gradientReadyBarrier_.wait(); + allClientPassFinish_ = numPassFinishClients_ == FLAGS_num_gradient_servers; + } + + DoOperationResponse response; + response.set_pass_finish(allClientPassFinish_); + + for (const auto& op : request.operations()) { + OperationResult* opResult = response.add_results(); + if (op.operation() >= ARRAYSIZE(opFuncs)) { + LOG(ERROR) << "Unknown operation " << op.operation(); + response.set_return_message(kRetMsgUnknownOperation); + } + OperatorFunction opFunc = opFuncs[op.operation()]; + if (!opFunc) { + LOG(ERROR) << "Operation not implemented: " << op.operation(); + response.set_return_message(kRetMsgUnknownOperation); + } + (this->*opFunc)(op, opResult); + } + + if (request.send_back_parameter()) { + /// clean current cost + cost_ = 0; + + if (allClientPassFinish_ && request.release_pass()) { + /// This signals that all clients finish one pass, so waitPassFinish() + /// will stop waiting. + numPassFinishClients_ = 0; + } + + /// notify addGradient() to send back parameter + parameterReadyBarrier_.wait(); + } + callback(response); +} + +void ParameterServer2::waitPassStart(const WaitPassStartRequest& request, + ProtoResponseCallback callback) { + passBarrier_.wait(); + callback(WaitPassStartResponse()); +} + +void ParameterServer2::waitPassFinish(const WaitPassFinishRequest& request, + ProtoResponseCallback callback) { + numPassFinishClients_ += 1; + + while (numPassFinishClients_ != 0) { + /// notify doOperation gradient ready + gradientReadyBarrier_.wait(); + /// wait doOperation finish + parameterReadyBarrier_.wait(); + } + + callback(WaitPassFinishResponse()); +} + +void ParameterServer2::synchronize(const SynchronizeRequest& request, + ProtoResponseCallback callback) { + synchronizeBarriers_[request.sync_object_id()]->wait(); + dataSize_ = 0; + callback(SynchronizeResponse()); +} + +void ParameterServer2::asyncFinishPass(const SynchronizeRequest& request, + ProtoResponseCallback callback) { + synchronizeBarriers_[request.sync_object_id()]->wait(); + callback(SynchronizeResponse()); + + if (request.trainer_id() == 0) { + batchId_ = 0; + } +} + +void ParameterServer2::createVector(const CreateVectorRequest& request, + ProtoResponseCallback callback) { + (void)request; + CreateVectorResponse response; + LOG(INFO) << "ParameterServer2::createVector: size=" << size_; + CpuVectorPtr vec = std::make_shared(size_); + int64_t handle = -1; + { + std::lock_guard guard(parameterMutex_); + handle = vectors_.size(); + vectors_.push_back(vec); + } + response.set_handle(handle); + callback(response); +} + +void ParameterServer2::releaseVector(const ReleaseVectorRequest& request, + ProtoResponseCallback callback) { + ReleaseVectorResponse response; + CpuVectorPtr vec; + { + std::lock_guard guard(parameterMutex_); + vec.swap(vectors_[request.handle()]); + } + callback(response); +} + +void ParameterServer2::createMatrix(const CreateMatrixRequest& request, + ProtoResponseCallback callback) { + CreateMatrixResponse response; + /// We need to create column major matrix of size_ * num_cols + /// Matrix is row majoar. Need to tranpose when use it. + CpuMatrixPtr mat = std::make_shared(request.num_cols(), size_); + int64_t handle = -1; + { + std::lock_guard guard(parameterMutex_); + handle = matrices_.size(); + matrices_.push_back(mat); + } + response.set_handle(handle); + callback(response); +} + +void ParameterServer2::releaseMatrix(const ReleaseMatrixRequest& request, + ProtoResponseCallback callback) { + ReleaseMatrixResponse response; + CpuMatrixPtr mat; + { + std::lock_guard guard(parameterMutex_); + mat.swap(matrices_[request.handle()]); + } + callback(response); +} + +} // namespace paddle diff --git a/paddle/legacy/pserver/ParameterServer2.h b/paddle/legacy/pserver/ParameterServer2.h new file mode 100644 index 0000000000000000000000000000000000000000..069e730ea4ea4b253518d70142f0f242145cd326 --- /dev/null +++ b/paddle/legacy/pserver/ParameterServer2.h @@ -0,0 +1,696 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +#include +#include + +#include "paddle/legacy/math/Matrix.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/parameter/ParameterOptimizer.h" +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/ThreadLocal.h" + +#include "ParameterService.pb.h" + +#include "ProtoServer.h" + +DECLARE_int32(port); + +namespace paddle { + +// @TODO(yanfei): +// if armed with high density computation resource per node, pserver could also +// utilize GPU to reduce overhead. if this mechanism is used, it could pipeline +// network receiving and GPU computation to reduce the network overhead even +// further. the pipeline could help to accelerate BIG model training. +// @TODO:(yanfei) +// for cpu and less/low gpu machine, the time exhausted by forward and backward +// could be larger than optimization at pserver. However, if armed with lots of +// gpus per node and if the model size is so large enough that limited cpu +// computation causes big optmization latency, the GPU may be required by +// pserver. + +/** + * Client interface for the parameter server + * + * it implements several rpc API for remote parameter client usage. + * for sync-sgd, client needs one controller thread to build connections + * to all pservers, these controller connections do barriers + * synchronization with these connections used for transfering data. + * each data connection uses block based fine grained synchronization + * to gain better scalability. Merging gradients from different trainers + * are concurrently executed with block units, so that some network + * overhead will be hidden in merging gradient. + * for async-sgd, the difference is that pserver will do optimization + * immediately if the gradients are ready, so that pserver needs to + * prepare separate buffer to store value for sending back to trainer + * to prevent from being polluted. + */ +class ParameterServer2 : public ProtoServer { + protected: + /// parameter_ mutex. + RWLock parameterMutex_; + + typedef std::pair BlockKey; + struct BlockKeyHash { + size_t operator()(const BlockKey& key) const { + return std::hash()(key.first) + key.second; + } + }; + + // TODO(yanfei): + // if index data structure is based on parameters instead of blocks, the + // lookup performance could be better. In addition, the block memory + // access almost exhibits good locality, so index data structure and + // block data structure can be refined further, especially if gpu is used + // for pserver. + /** + * all parameters are stored in CpuVector with a blockMap_ data structure + * to index block data required by requests. + */ + typedef std::unordered_map BlockMap; + /// <(para, block), global offset(byte) in all parameters> + BlockMap blockOffsetMap_; + /// <(para, block), global idx [0, nBlocksInAllParameters]> + BlockMap blockIdMap_; + + std::vector vectors_; + std::vector matrices_; + std::vector dataMems_; + + // TODO(yanfei): + // if storing sparse_remote_update() flag in request instead of + // reading configMap_, and storing config within new block wise + // overview data structure, the config mapping, block mapping + // can be unified in single clean data structure. Use para_id + // to index parameters, use offset to index block within parameter + // and keep two index into single one. + /** + * mapping between parameter and config + * different parameter allows different config, such as decay_rate. + * for each request, it need to read config for adding gradient + * and optmization. + */ + std::unordered_map configMap_; + + /** + * to parallelize the multi-thread and multi-connnection + * computation at pserver, it use block unit to reduce + * the contention for computation, even further use block + * level optimizater control for each block for some special + * reason annotated below. + */ + struct BlockInfo { + const ParameterConfig* config; + std::unique_ptr lock; + /// global offset for all parameters + uint64_t offset; + /** + * + * Async sgd in pserver is very different from sync sgd. + * Each trainer follows startBatch, update*, finishBatch as in + * sync sgd, but all these actions are almost executed by + * multi-core and multi-thread simutaneously, so that async + * sgd optimization is based on block level in reality, then + * per block optimization is necessary indeed. In addition, + * per block optimization is also perfered for performance + * with multithreads. + */ + std::unique_ptr optimizer; + }; + std::vector blockInfos_; + + typedef std::vector> BlockSegments; + /// Because some blocks might not be fully used. We keep a + /// record of which segments are used. + BlockSegments usedSegments_; + + /// record pserver status, all status defined in ParameterService.pb + PServerStatus status_; + /// record all samples processed which could be used by optimizater + std::atomic numSamplesProcessed_; + double cost_; + int mpiSize_; + int dataSize_; + /// configuration for current parameter optimizer + OptimizationConfig config_; + + /** + * The ReadWriteBuffer is based on std::vector, but aligned for avx/sse + * compute. And add some helper method to allocate memory aligned blocks. + * + * @param T type of element. + * @param AlignBytes the memory aligned bytes for allocated blocks. + */ + template + class ReadWriteBuffer + : public std::vector> { + public: + static_assert(sizeof(T) % AlignBytes == 0 || AlignBytes % sizeof(T) == 0, + "Type T must be able to aligned."); + + /** + * @brief IsTLargerThanAlign compiled time calculated constant for is type + * T larger than alignments. + */ + constexpr static bool IsTLargerThanAlign = sizeof(T) >= AlignBytes; + + static_assert(std::is_pod::value, "T must be POD type."); + + /** + * @brief if AlignBytes > sizeof(T), then will calcuate how many elements + * can be stored in AlignBytes. + */ + constexpr static size_t AlignElementCount = AlignBytes / sizeof(T); + + static_assert(AlignElementCount == + (AlignElementCount & -AlignElementCount) || + AlignBytes > sizeof(T), + "AlignElementCount should be exp of 2"); + + /** + * @brief Resize Buffer, with block count that will be allocated. Each block + * will be memory aligned in AlignBytes. + * @param size The element count in all blocks. + * @param alignBlockCount The block count that will be allocated. + */ + void resizeWithAlignHints(size_t size, size_t alignBlockCount = 1) { + if (IsTLargerThanAlign) { //! So, each elements is memory aligned. + this->resize(size); + } else { + //! at most, we need such elements in buffer to make sure each block is + //! aligned. + this->resize(size + alignBlockCount * (AlignElementCount - 1)); + } + } + + /** + * @brief reset aligned allocate blocks. + */ + void resetAlignAlloc() { this->curOffset_ = 0; } + + /** + * @brief get next aligned block address. + * @param blockSize is the element count in each block. + * @return Aligned block address. + */ + T* nextBlock(size_t blockSize) { + T* r = &this->operator[](curOffset_); + curOffset_ += blockSize; + + if (!IsTLargerThanAlign) { + curOffset_ = + (curOffset_ + AlignElementCount - 1) & ~(AlignElementCount - 1); + } + return r; + } + + private: + size_t curOffset_; + }; + + /// to buffer the data from network for further processing to + /// reduce redundant memory allocation. + ThreadLocal> readWriteBuffer_; + + /// size of the parameter + int64_t size_; + + /// for synchronized training, check details in addGradient() + /// and doOperation() + ThreadBarrier gradientReadyBarrier_; + ThreadBarrier parameterReadyBarrier_; + ThreadBarrier passBarrier_; + ThreadLocal> requestVec_; + ThreadLocal> callbackVec_; + + std::atomic numPassFinishClients_; + bool allClientPassFinish_; + + std::vector> synchronizeBarriers_; + std::atomic serverId_; + + /** + * + * for lagged async gradient gradient commit control in Async Sgd. + * discard lagged gradients from too slow nodes, whose gradients + * exhibits bad quality. + * Algorithm: + * pserver: + * 1. initial asyncUpdaterSteps = 0, asyncTrainerSteps_[N] = 0. + * syncUpdaterSteps means + * the version of parameter value. + * 2. when pull arrives, record asyncUpdateSteps_ into + * syncTrainerSteps_[trainer_id] + * 3. when push arrives, compare asyncUpdateSteps_ with + * syncTrainerSteps_[trainer_id] + * if delta > threshold, discard current gradient, else commit + * gradient. + * 4. reset asyncUpdaterSteps_ and asyncTrainerSteps_[N] when pass + * finished + * Note: + * it can not discard all lag-gradient strictly in some special + * condition. part of gradients could be discarded if + * ConcurrentRemoteParameterUpdater is sed. + * this algorithm is implemented in asynSGD() + */ + int64_t asyncLaggedThreshold_; + std::atomic asyncUpdateSteps_; + std::vector asyncTrainerSteps_; + size_t asyncLaggedGradientsNum_; + /// stat all async update + std::vector asyncUpdateStat_; + /// stat per trainer_id + std::vector asyncTrainerDiscardStat_; + /// stat per trainer_id + std::vector asyncTrainerCommitStat_; + + /// only used by controller and other control cmd from trainer number 0 + std::unique_ptr syncThreadPool_; + + /// pserver for sparse remote update parameters + bool isSparseServer_; + + /// barrier performance tuning sync-sgd required + std::atomic batchId_; + + public: + struct Buffer { + real* base; + size_t size; + }; + + protected: + /// async gradient commit control + bool asyncGrdientCommitCheckAndStat(const SendParameterRequest& request); + + public: + /// disable default parameter for overloading + /// @rdmaCpu:the id of cpu core hosting RDMA server(0-N) + /// -1 means using TCP transport instead of RDMA + ParameterServer2(const std::string& addr, int port, int rdmaCpu = -1); + + ~ParameterServer2() {} + + static const std::string kRetMsgInvalidMatrixHandle; + static const std::string kRetMsgInvalidVectorHandle; + static const std::string kRetMsgUnknownOperation; + + /// service functions + template + void reduceAndSendData(const SendDataRequest& request, + std::unique_ptr& msgReader, + ProtoResponseCallbackEx& callback); + + void templateReduceSum(const SendDataRequest& request, + std::unique_ptr& msgReader, + ProtoResponseCallbackEx& callback); + + /** + * @brief framework for sending parameters + * + * @note different parameter data type can be sent to pserver. + * in most case, the api is used to send gradients from + * trainer to pserver. + * it also can be used to retrieve parameters from pserver + */ + void sendParameter(const SendParameterRequest& request, + std::unique_ptr msgReader, + ProtoResponseCallbackEx callback); + + void sendData(const SendDataRequest& request, + std::unique_ptr msgReader, + ProtoResponseCallbackEx callback); + + /** + * @brief send config to pserver + * + * @note it can help pserver to understand the configuration for + * optimization, + * logging control, duplicated initialization, etc. + */ + void setConfig(const SetConfigRequest& request, + ProtoResponseCallback callback); + + /** + * @brief get status for pserver + * + * @note used to check if parameters are ready at pserver + */ + void getStatus(const GetStatusRequest& request, + ProtoResponseCallback callback); + + /** + * @brief set status for pserver + * + * @note used to check if parameters are ready at pserver, since parameters + * at pserver are initialized by trainer + */ + void setStatus(const SetStatusRequest& request, + ProtoResponseCallback callback); + + /** + * @brief framework for doing some operation at pserver end + * + * @note if sync-sgd is used, controller will calling op_SGD action + * for gradient optimization. + * check avaiable operations in opFuncs[] + */ + void doOperation(const DoOperationRequest& request, + ProtoResponseCallback callback); + + /// Create a column vector. The size is the dimension of parameter + void createVector(const CreateVectorRequest& request, + ProtoResponseCallback callback); + + void releaseVector(const ReleaseVectorRequest& request, + ProtoResponseCallback callback); + + /// Create a column major matrix. The number of rows is the dimension of + /// parameter. The number of columns is specifed by num_cols. + void createMatrix(const CreateMatrixRequest& request, + ProtoResponseCallback callback); + + void releaseMatrix(const ReleaseMatrixRequest& request, + ProtoResponseCallback callback); + /** + * @brief stateful control for indicationg sync pass start + * + * @note it is valuable for logging and state control, + * especially for sync-sgd control + */ + void waitPassStart(const WaitPassStartRequest& request, + ProtoResponseCallback callback); + + /** + * @brief stateful control for indicationg sync pass end + * + * @note it is valuable for logging and state control, + * especially for sync-sgd control + */ + void waitPassFinish(const WaitPassFinishRequest& request, + ProtoResponseCallback callback); + + /** + * @brief synchronize all distributed trainers + * + * @note it's general api for synchronizing trainer and pserver + */ + void synchronize(const SynchronizeRequest& request, + ProtoResponseCallback callback); + + /** + * @brief stateful control for indicating async pass is finished + * + * @note it is valuable for logging control, state reset, etc. + */ + void asyncFinishPass(const SynchronizeRequest& request, + ProtoResponseCallback callback); + + void loadValueVector(const LoadValueRequest& request, + ProtoResponseCallback callback); + + void saveValueVector(const SaveValueRequest& request, + ProtoResponseCallback callback); + + public: + /** + * @brief initialize parameter server + */ + bool init(); + + /** + * @brief set parameters at pserver + * + * @note do parameter initialization if neccessy. + */ + void setParameter(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers); + + /** + * @brief receive gradients and do optimization for async-sgd + * + * @note this api asynchronizately receives all data from all + * trainers, and immediately do optimization and return + * optimizated value for trainer. + * this above routine are block based atomic updating, + * which means different block could based different stale + * gradient. + * it will discard some lagged gradients by default for + * better convergence. + */ + void asyncSGD(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers); + + /** + * @brief merge gradients from all trainer + * + * @note this api use block based parallelization as fine grained + * parallelization which benifits lock contention and latency + * hidden for communication, also can harness multi-core + * efficiently. + * it also implements the synchronization for sync-sgd + */ + void addGradient(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers); + + /** + * @brief get dense parameters from pserver + * + * @note for some specified condition, trainer will get parameters from + * pservers. + * e.g. + * if all parameters are stored at perver end for big model training + * trainer can use it to retrieve all parameters if necessary. + */ + void getParameter(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers); + + /** + * @brief get sparse value from parameter server + * + * @note with sparse enabled, pservers own all latest value + * while trainer only retrieve value that only are needed. + * e.g. + * trainer will do prefetch action to retrieve necessary latest + * value from pserver for sparse calculation. + */ + void getParameterSparse(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers); + + protected: + void mergeSegments(BlockSegments* segments); + + /// set the unused segments to zero + void clearUnusedSegments(CpuVector* vec); + + // TODO(yanfei): + // if read data and do optimization interleavely block by block, + // the performance could be better for gaining less network congestion. + /// read all data from connection and store it in static pre-allocated buffer + void readAllBlocks(MsgReader* msgReader, + std::vector* buffers); + + const ParameterConfig& getParameterConfig(const ParameterBlock& block) { + CHECK_LT(block.para_id(), -1UL) << "invalid parameter id:" + << block.para_id(); + const auto it = configMap_.find(block.para_id()); + CHECK(it != configMap_.end()) << "can not find parameter id: " + << block.para_id(); + return it->second; + } + + /// it implictly check blockOffsetMap_ while retrieving blockId + const ParameterConfig& getParameterConfig(int64_t blockId) const { + CHECK(blockId >= 0 && blockId < (int64_t)blockInfos_.size()) + << "block idx out of range, id: " << blockId + << " info size: " << blockInfos_.size(); + return *(blockInfos_[blockId].config); + } + + template + bool isValidVectorHandle(int64_t handle, Response* response) { + if (handle < 0 || (size_t)handle >= vectors_.size()) { + LOG(ERROR) << "Invalid vector handle " << handle; + response->set_return_message(kRetMsgInvalidVectorHandle); + return false; + } + return true; + } + + template + bool isValidMatrixHandle(int64_t handle, Response* response) { + if (handle < 0 || (size_t)handle >= matrices_.size()) { + LOG(ERROR) << "Invalid matrix handle " << handle; + response->set_return_message(kRetMsgInvalidMatrixHandle); + return false; + } + return true; + } + + /** + * @brief get block offset + * + * @note block.begin_dim is added to the block offset. + * return -1 if block cannot be found + */ + int64_t getBlockOffset(const ParameterBlock& block) const { + BlockKey key(block.para_id(), block.block_id()); + auto it = blockOffsetMap_.find(key); + if (it == blockOffsetMap_.end()) { + return -1; + } + return it->second; + } + + /// return -1 if block cannot be found + int64_t getBlockId(const ParameterBlock& block) const { + BlockKey key(block.para_id(), block.block_id()); + auto it = blockIdMap_.find(key); + if (it == blockIdMap_.end()) { + return -1; + } + return it->second; + } + + /** + * @brief prepare data for sending back + * + * @note modify reponse and outputBuffers for sending parameter + * back to client. The buffer for socket sending uses + * vectors_[parameterType] directly + * for dense with sync-sgd + */ + void sendBackParameter(const ParameterBlock& block, + int parameterType, + SendParameterResponse* response, + std::vector* outputBuffers); + + /** + * @brief prepare data for sending back + * + * @note modify response and outputBuffers for sending parameter + * back to client. The buffer for socket sending uses buffer->base + * The parameter values are copied from vectors_[parameterType] + * to buffer->base. + * for dense with async-sgd + */ + void sendBackParameter(const ParameterBlock& block, + int parameterType, + SendParameterResponse* response, + Buffer* buffer, + std::vector* outputBuffers); + /** + * @brief prepare data for sending back + * + * @note specified for sparse + */ + void sendBackParameterSparse(const ParameterBlock& block, + int parameterType, + SendParameterResponse* response, + Buffer* buffer, + size_t width, + std::vector* outputBuffers); + + /** + * framework routine for block parallelization + * e.g. + * for optimization on all blocks at pserver end, this routine can facilitize + * the parallelize of do optimization on all blocks with multithreads. + */ + typedef std::function ExecFunc; + void parallelExecForEachBlock(ExecFunc func); + void blockTraverse(BlockInfo& info, + const ParameterConfig& config, + int64_t offset, + size_t size, + const VectorPtr vecs[], + const ParameterOptimizer::TraverseCallback& callback); + + public: + typedef void (ParameterServer2::*OperatorFunction)(const Operation& operation, + OperationResult* result); + + /** + * doOperation will call following operations indirectly + * e.g. + * for sync-sgd control, the controller in remote updater will send op_SGD + * command to pserver, then send sendParameter request to pserver immediately. + * the two function at pserver end will do cooperation to achieve the sync-sgd + * gradient merge and optimization. + * the most following operations are specified for owlqn, all operations are + * under the context of doOperation function + */ + static OperatorFunction opFuncs[]; + + void op_SGD(const Operation& operation, OperationResult* result); + + void op_RESET(const Operation& operation, OperationResult* result); + + void op_utv(const Operation& operation, OperationResult* result); + + void op_au_bv(const Operation& operation, OperationResult* result); + + void op_COPY(const Operation& operation, OperationResult* result); + + void op_au(const Operation& operation, OperationResult* result); + + void op_au_bv_cw(const Operation& operation, OperationResult* result); + + void op_make_steepest_desc_dir(const Operation& operation, + OperationResult* result); + + void op_fix_dir_signs(const Operation& operation, OperationResult* result); + + void op_dir_deriv(const Operation& operation, OperationResult* result); + + void op_fix_omega_signs(const Operation& operation, OperationResult* result); + + void op_cost(const Operation& operation, OperationResult* result); + + void op_start_pass(const Operation& operation, OperationResult* result); + void op_finish_pass(const Operation& operation, OperationResult* result); + + void op_apply(const Operation& operation, OperationResult* result); + + void op_randomize(const Operation& operation, OperationResult* result); + + void op_load(const Operation& operation, OperationResult* result); + void op_save(const Operation& operation, OperationResult* result); +}; + +} // namespace paddle diff --git a/paddle/pserver/ParameterServer2Main.cpp b/paddle/legacy/pserver/ParameterServer2Main.cpp similarity index 100% rename from paddle/pserver/ParameterServer2Main.cpp rename to paddle/legacy/pserver/ParameterServer2Main.cpp diff --git a/paddle/pserver/ParameterServerController.cpp b/paddle/legacy/pserver/ParameterServerController.cpp similarity index 100% rename from paddle/pserver/ParameterServerController.cpp rename to paddle/legacy/pserver/ParameterServerController.cpp diff --git a/paddle/legacy/pserver/ParameterServerController.h b/paddle/legacy/pserver/ParameterServerController.h new file mode 100644 index 0000000000000000000000000000000000000000..b90d0cbceaa879b8cb281867b5326ff50c1e311a --- /dev/null +++ b/paddle/legacy/pserver/ParameterServerController.h @@ -0,0 +1,74 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "ParameterServer2.h" +#include "ParameterServerConfig.pb.h" +#include "RDMANetwork.h" +#include "paddle/legacy/utils/StringUtil.h" + +namespace paddle { + +/** + * @brief ParameterServerController is used for create, init and manage multi + * parameter server instances. The num of the instances is decided by port + * num(the ports number for parameter send) and network devices configured + * by gflags or proto. + */ +class ParameterServerController final { + public: + DISABLE_COPY(ParameterServerController); + + /** + * @brief Ctor, Create a ParameterServerController from ParameterServerConfig. + */ + explicit ParameterServerController(const ParameterServerConfig& config); + + /** + * @brief Dtor. + */ + ~ParameterServerController(); + + /** + * @brief create ParameterServerController from gflags, this is used for + * compatibility with the old usage of configuration by gflags. + */ + static ParameterServerController* createFromGflags(); + + /** + * @brief create ParameterServerController with ParameterServerConfig, remove + * gflags from ParameterServer. Init all ParameterServer2 instances according + * to + * the config. + */ + static ParameterServerController* create(const ParameterServerConfig& config); + + /** + * @brief start all ParameterServer2 instances in this + * ParameterServerController. + */ + void start(); + + /** + * @brief join and wait for all ParameterServer2 instances thread in this + * ParameterServerController. + */ + void wait(); + + private: + std::vector> parameterServers_; +}; + +} // namespace paddle diff --git a/paddle/pserver/ProtoServer.cpp b/paddle/legacy/pserver/ProtoServer.cpp similarity index 100% rename from paddle/pserver/ProtoServer.cpp rename to paddle/legacy/pserver/ProtoServer.cpp diff --git a/paddle/pserver/ProtoServer.h b/paddle/legacy/pserver/ProtoServer.h similarity index 100% rename from paddle/pserver/ProtoServer.h rename to paddle/legacy/pserver/ProtoServer.h diff --git a/paddle/legacy/pserver/RDMANetwork.h b/paddle/legacy/pserver/RDMANetwork.h new file mode 100644 index 0000000000000000000000000000000000000000..c87056f72c56647c827cdbd7bdd6a992b4bb1cf6 --- /dev/null +++ b/paddle/legacy/pserver/RDMANetwork.h @@ -0,0 +1,158 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#ifndef PADDLE_DISABLE_RDMA +#include "sxi_sock.h" +#else +#define PROMPT_ERR() LOG(FATAL) << "Paddle is not compiled with rdma" +#endif +#include "paddle/legacy/utils/Logging.h" + +#include +struct sxi_sock; +struct sxi_socket; + +#ifndef MAX_VEC_SIZE +// define default MAX_VEC_SIZE +#define MAX_VEC_SIZE (1UL << 16) +#endif + +namespace paddle { +/// Namespace rdma is adaptors for sxi_sock.h. Make paddle not depend on it +/// when disable rdma support +namespace rdma { +inline int numCpus() { +#ifndef PADDLE_DISABLE_RDMA + return sxi_num_configured_cpus(); +#else + return 0; +#endif +} + +inline sxi_socket* ssocket(int cpuId) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_ssocket(cpuId); +#else + PROMPT_ERR(); +#endif +} + +inline int listen(sxi_socket* s) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_listen(s); +#else + PROMPT_ERR(); +#endif +} + +inline int bind(sxi_socket* s, const char* str) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_bind(s, str); +#else + PROMPT_ERR(); +#endif +} + +inline sxi_sock* accept(sxi_socket* s) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_accept(s); +#else + PROMPT_ERR(); +#endif +} + +inline sockaddr_in* getSourceAddress(sxi_sock* sock) { +#ifndef PADDLE_DISABLE_RDMA + return reinterpret_cast(&sock->sa); +#else + PROMPT_ERR(); +#endif +} + +inline int close(sxi_socket* sock) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_socket_close(sock); +#else + PROMPT_ERR(); +#endif +} + +inline int close(sxi_sock* sock) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_sock_close(sock); +#else + PROMPT_ERR(); +#endif +} + +inline void init() { +#ifndef PADDLE_DISABLE_RDMA + sxi_module_init(); +#else + PROMPT_ERR(); +#endif +} + +inline sxi_socket* csocket(int cpuId) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_csocket(cpuId); +#else + PROMPT_ERR(); +#endif +} + +inline ssize_t read(sxi_sock* channel, void* data, size_t len) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_read(channel, data, len); +#else + PROMPT_ERR(); +#endif +} + +inline ssize_t write(sxi_sock* channel, void* data, size_t len) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_write(channel, data, len); +#else + PROMPT_ERR(); +#endif +} + +inline ssize_t readv(sxi_sock* channel, iovec* iov, int count) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_readv(channel, iov, count); +#else + PROMPT_ERR(); +#endif +} + +inline ssize_t writev(sxi_sock* channel, iovec* iov, int count) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_writev(channel, iov, count); +#else + PROMPT_ERR(); +#endif +} + +inline sxi_sock* connect(sxi_socket* socket, const char* url) { +#ifndef PADDLE_DISABLE_RDMA + return sxi_connect(socket, url); +#else + PROMPT_ERR(); +#endif +} + +} // namespace rdma +} // namespace paddle diff --git a/paddle/legacy/pserver/SocketChannel.cpp b/paddle/legacy/pserver/SocketChannel.cpp new file mode 100644 index 0000000000000000000000000000000000000000..79c763c62ba845067c7729eafb5b218fc7b91482 --- /dev/null +++ b/paddle/legacy/pserver/SocketChannel.cpp @@ -0,0 +1,235 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SocketChannel.h" + +#include +#include +#include +#include +#include +#include +#include "RDMANetwork.h" + +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +/** + * UIO_MAXIOV is documented in writev(2), but only + * declares it on osx/ios if defined(KERNEL) + */ +#ifndef UIO_MAXIOV +#define UIO_MAXIOV 512 +#endif + +SocketChannel::~SocketChannel() { + if (tcpRdma_ == F_TCP) + close(tcpSocket_); + else + rdma::close(rdmaSocket_); + LOG(INFO) << "destory connection in socket channel, peer = " << peerName_; +} + +size_t SocketChannel::read(void* buf, size_t size) { + size_t total = 0; + while (total < size) { + ssize_t len; + if (tcpRdma_ == F_TCP) + len = ::read(tcpSocket_, (char*)buf + total, size - total); + else + len = rdma::read(rdmaSocket_, (char*)buf + total, size - total); + + CHECK(len >= 0) << " peer=" << peerName_; + if (len <= 0) { + return total; + } + total += len; + } + return total; +} + +size_t SocketChannel::write(const void* buf, size_t size) { + size_t total = 0; + while (total < size) { + ssize_t len; + if (tcpRdma_ == F_TCP) + len = ::write(tcpSocket_, (const char*)buf + total, size - total); + else + len = rdma::write(rdmaSocket_, (char*)buf + total, size - total); + + CHECK(len >= 0) << " peer=" << peerName_; + if (len <= 0) { + return total; + } + total += len; + } + return total; +} + +template +static size_t readwritev(IOFunc iofunc, + SocketType socket, + iovec* iovs, + int iovcnt, + int maxiovs, + const std::string& peerName) { + int curIov = 0; + size_t total = 0; + + for (int i = 0; i < iovcnt; ++i) { + total += iovs[i].iov_len; + } + + size_t size = 0; + size_t curIovSizeDone = 0; + + while (size < total) { + ssize_t len = + iofunc(socket, &iovs[curIov], std::min(iovcnt - curIov, maxiovs)); + CHECK(len > 0) << " peer=" << peerName << " curIov=" << curIov + << " iovCnt=" << iovcnt + << " iovs[curIov].base=" << iovs[curIov].iov_base + << " iovs[curIov].iov_len=" << iovs[curIov].iov_len; + size += len; + + /// restore iovs[curIov] to the original value + iovs[curIov].iov_base = + (void*)((char*)iovs[curIov].iov_base - curIovSizeDone); + iovs[curIov].iov_len += curIovSizeDone; + + len += curIovSizeDone; + + while (curIov < iovcnt) { + if ((size_t)len < iovs[curIov].iov_len) break; + len -= iovs[curIov].iov_len; + ++curIov; + } + if (curIov < iovcnt) { + curIovSizeDone = len; + iovs[curIov].iov_base = (void*)((char*)iovs[curIov].iov_base + len); + iovs[curIov].iov_len -= len; + } + } + return size; +} + +/// rdma::readv and rdma::writev can take advantage of RDMA blocking offload +/// transfering +size_t SocketChannel::writev(const std::vector& iovs) { + if (tcpRdma_ == F_TCP) + return readwritev(::writev, + tcpSocket_, + const_cast(&iovs[0]), + iovs.size(), + UIO_MAXIOV, + peerName_); + else + return readwritev(rdma::writev, + rdmaSocket_, + const_cast(&iovs[0]), + iovs.size(), + MAX_VEC_SIZE, + peerName_); +} + +size_t SocketChannel::readv(std::vector* iovs) { + if (tcpRdma_ == F_TCP) + return readwritev(::readv, + tcpSocket_, + const_cast(&(*iovs)[0]), + iovs->size(), + UIO_MAXIOV, + peerName_); + else + return readwritev(rdma::readv, + rdmaSocket_, + const_cast(&(*iovs)[0]), + iovs->size(), + MAX_VEC_SIZE, + peerName_); +} + +void SocketChannel::writeMessage(const std::vector& userIovs) { + MessageHeader header; + header.numIovs = userIovs.size(); + + std::vector iovLengths; + iovLengths.reserve(userIovs.size()); + for (auto& iov : userIovs) { + iovLengths.push_back(iov.iov_len); + } + + std::vector iovs; + iovs.reserve(userIovs.size() + 2); + iovs.push_back({&header, sizeof(header)}); + iovs.push_back({&iovLengths[0], + static_cast(sizeof(iovLengths[0]) * header.numIovs)}); + iovs.insert(iovs.end(), userIovs.begin(), userIovs.end()); + + header.totalLength = 0; + for (auto& iov : iovs) { + header.totalLength += iov.iov_len; + } + + CHECK(writev(iovs) == (size_t)header.totalLength); +} + +std::unique_ptr SocketChannel::readMessage() { + MessageHeader header; + + size_t len = read(&header, sizeof(header)); + if (len == 0) { + return nullptr; + } + + CHECK(len == sizeof(header)); + + std::unique_ptr msgReader(new MsgReader(this, header.numIovs)); + + CHECK_EQ(msgReader->getTotalLength() + sizeof(header) + + msgReader->getNumBlocks() * sizeof(size_t), + (size_t)header.totalLength) + << " totalLength=" << msgReader->getTotalLength() + << " numBlocks=" << msgReader->getNumBlocks(); + return msgReader; +} + +MsgReader::MsgReader(SocketChannel* channel, size_t numBlocks) + : channel_(channel), blockLengths_(numBlocks), currentBlockIndex_(0) { + size_t size = numBlocks * sizeof(blockLengths_[0]); + CHECK(channel_->read(&blockLengths_[0], size) == size); +} + +void MsgReader::readBlocks(const std::vector& bufs) { + CHECK_LE(currentBlockIndex_ + bufs.size(), blockLengths_.size()); + std::vector iovs; + iovs.reserve(bufs.size()); + size_t totalLength = 0; + for (void* buf : bufs) { + iovs.push_back({buf, getNextBlockLength()}); + totalLength += getNextBlockLength(); + ++currentBlockIndex_; + } + + CHECK(channel_->readv(&iovs) == totalLength); +} + +void MsgReader::readNextBlock(void* buf) { + CHECK_LT(currentBlockIndex_, blockLengths_.size()); + CHECK(channel_->read(buf, getNextBlockLength()) == getNextBlockLength()); + ++currentBlockIndex_; +} + +} // namespace paddle diff --git a/paddle/legacy/pserver/SocketChannel.h b/paddle/legacy/pserver/SocketChannel.h new file mode 100644 index 0000000000000000000000000000000000000000..a7b3cd42f0aa32c3a74e14f87dbfe64d25473254 --- /dev/null +++ b/paddle/legacy/pserver/SocketChannel.h @@ -0,0 +1,153 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include + +#include +#include + +struct sxi_sock; + +namespace paddle { + +class SocketChannel; +enum ChannelType { + F_TCP = 1, + F_RDMA = 2, +}; + +/// reading a set of blocks of data from SocketChannel. +class MsgReader { + public: + MsgReader(SocketChannel* channel, size_t numIovs); + ~MsgReader() { + /// ensure all data blocks have been processed + CHECK_EQ(currentBlockIndex_, blockLengths_.size()); + } + /** + * @brief number of remaining parts + */ + size_t getNumBlocks() const { + return blockLengths_.size() - currentBlockIndex_; + } + + /** + * @brief lenght of next block + */ + size_t getNextBlockLength() const { return getBlockLength(0); } + + /** + * @brief get the total length of all the remaining blocks + */ + size_t getTotalLength() const { + size_t total = 0; + for (size_t i = currentBlockIndex_; i < blockLengths_.size(); ++i) { + total += blockLengths_[i]; + } + return total; + } + + /** + * @brief Get the length for block currentBlockIndex + i + */ + size_t getBlockLength(size_t i) const { + return blockLengths_[currentBlockIndex_ + i]; + } + + /** + * @brief read blocks data and store it to buf + */ + void readBlocks(const std::vector& bufs); + void readNextBlock(void* buf); + + protected: + SocketChannel* channel_; + std::vector blockLengths_; + size_t currentBlockIndex_; +}; + +/// APIs for reading and writing byte stream data or naive iov data +/// from the APIs both RDMA and TCP exhibits byte stream style +class SocketChannel { + public: + SocketChannel(int socket, const std::string& peerName) + : tcpSocket_(socket), peerName_(peerName) { + tcpRdma_ = F_TCP; + } + SocketChannel(struct sxi_sock* socket, const std::string& peerName) + : rdmaSocket_(socket), peerName_(peerName) { + tcpRdma_ = F_RDMA; + } + + ~SocketChannel(); + + const std::string& getPeerName() const { return peerName_; } + + /** + * @brief read size bytes. + * + * @note keep reading until getting size bytes or sock is closed + * is closed + */ + size_t read(void* buf, size_t size); + + /** + * @brief write size bytes. + * + * @note keep writing until writing size bytes or sock is closed + */ + size_t write(const void* buf, size_t size); + + /** + * @brief write a set of buffers. + * + * @note keep writing until all buffers are written or sock is closed + */ + size_t writev(const std::vector& iov); + + /** + * @brief read a set of buffers. + * + * @note keep reading until all buffers are full or sock is closed. + */ + size_t readv(std::vector* iov); + + /** + * @brief write a set of buffers. + * + * @note keep writing until all buffers are passed or sock is closed + */ + void writeMessage(const std::vector& iov); + + /// return null to indicate socket is closed + std::unique_ptr readMessage(); + + protected: + struct MessageHeader { + int64_t totalLength; /// include the header + int64_t numIovs; + int64_t iovLengths[0]; + }; + + int tcpSocket_; + struct sxi_sock* rdmaSocket_; + const std::string peerName_; + enum ChannelType tcpRdma_; +}; + +} // namespace paddle diff --git a/paddle/legacy/pserver/SparseParameterDistribution.cpp b/paddle/legacy/pserver/SparseParameterDistribution.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3f17b228f0e5fd33b7e7db2afe1fb9421acc69c5 --- /dev/null +++ b/paddle/legacy/pserver/SparseParameterDistribution.cpp @@ -0,0 +1,123 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/utils/Flags.h" + +#include "SparseParameterDistribution.h" + +DEFINE_bool(check_sparse_distribution_in_pserver, + false, + "check whether sparse parameter exhibts balanced distribution at " + "all pservers"); +DEFINE_bool(show_check_sparse_distribution_log, + false, + "show logs details for sparse parameter distribution in pserver"); +DEFINE_int32(check_sparse_distribution_batches, + 100, + "run sparse parameter distribution check for N batches"); +DEFINE_double( + check_sparse_distribution_ratio, + 0.6, + "if parameters dispatched to different pservers exhibit unbalanced " + " distribution for check_sparse_distribution_ratio * " + " check_sparse_distribution_batches times, crash program"); +DEFINE_double(check_sparse_distribution_unbalance_degree, + 2.0, + "the ratio of maximum data size and minimun data size for " + "different pserver"); + +namespace paddle { + +SparseParameterDistribution::SparseParameterDistribution(size_t serviceNum) { + totBytes_ = 0; + data_.resize(serviceNum); + + batchPassed_ = 0; + unbalanceCnt_ = 0; +} + +void SparseParameterDistribution::probeDistribution(int serverId, + size_t dataSize) { + if (!FLAGS_check_sparse_distribution_in_pserver || + batchPassed_ > FLAGS_check_sparse_distribution_batches) { + return; + } + + CHECK_LT((size_t)serverId, data_.size()) + << "invalid sparse parameter distribution probe"; + + data_[serverId] += dataSize; + totBytes_ += dataSize; +} + +void SparseParameterDistribution::checkAndResetDistribution() { + if (!FLAGS_check_sparse_distribution_in_pserver || + batchPassed_ >= FLAGS_check_sparse_distribution_batches) { + return; + } + + /// at runtime, prepareSendData is called by many contexts, + /// so need to check if data is avaiable. + if (!totBytes_) { + return; + } + + /// check if distribution is balanced + auto avgSize = totBytes_ / data_.size(); + auto unbalanceDegree = FLAGS_check_sparse_distribution_unbalance_degree; + for (auto& dataSize : data_) { + if (dataSize > unbalanceDegree * avgSize || + dataSize * unbalanceDegree < avgSize) { + unbalanceCnt_++; + break; + } + } + + auto printData = [&]() { + std::stringstream ss; + for (auto& dataSize : data_) { + ss << dataSize * 0.001 << "KB "; + } + ss << std::endl; + LOG(INFO) << ss.str(); + }; + + /// show all sparse data size for different pserver + if (FLAGS_show_check_sparse_distribution_log) { + LOG(INFO) << "sparse distribution:"; + printData(); + } + + totBytes_ = 0; + batchPassed_++; + + if (batchPassed_ == FLAGS_check_sparse_distribution_batches) { + LOG(INFO) << "show last parameter distribution sample:"; + printData(); + LOG(INFO) << "total unbalanced batches: " << unbalanceCnt_ + << " in passed batches: " << batchPassed_; + CHECK_LE((float)unbalanceCnt_ / (float)batchPassed_, + FLAGS_check_sparse_distribution_ratio) + << "unbalanced sparse parameter distribution for different pserver. " + << "it could be caused by unbalanced sparse ids distribution, try " + << "to shuffle dimensions in input samples"; + } + + std::fill(data_.begin(), data_.end(), 0); +} +} // namespace paddle diff --git a/paddle/legacy/pserver/SparseParameterDistribution.h b/paddle/legacy/pserver/SparseParameterDistribution.h new file mode 100644 index 0000000000000000000000000000000000000000..ee78029958f675d07ec0aba2d0c1ea92d664e8fd --- /dev/null +++ b/paddle/legacy/pserver/SparseParameterDistribution.h @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include + +#include +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +/* + * if sparse_remote_updater is used, different ParameterServer could + * be assigned with unbalanced gradients. the parameter value from + * ParameterServer also be not balanced. the distribution of different + * dimensions of sparse ids determines the unbalanced degree of data + * distributed among all ParameterServers. Even distribution will + * benifits cluster efficiency. + * do check the unbalanced degree of gradients at runtime, crash program + * if unbalanced distribution exhibts by default. + */ +class SparseParameterDistribution { + public: + /// serviceNum means the number of ParameterServers + explicit SparseParameterDistribution(size_t serviceNum); + ~SparseParameterDistribution() {} + /// collect data + void probeDistribution(int serverId, size_t data); + void checkAndResetDistribution(); + + private: + std::vector data_; + std::atomic totBytes_; + + /// after some batches, stop to check + int batchPassed_; + + /// stat on unbalanced distribution found + int unbalanceCnt_; +}; +} // namespace paddle diff --git a/paddle/pserver/test/.gitignore b/paddle/legacy/pserver/test/.gitignore similarity index 100% rename from paddle/pserver/test/.gitignore rename to paddle/legacy/pserver/test/.gitignore diff --git a/paddle/pserver/test/CMakeLists.txt b/paddle/legacy/pserver/test/CMakeLists.txt similarity index 100% rename from paddle/pserver/test/CMakeLists.txt rename to paddle/legacy/pserver/test/CMakeLists.txt diff --git a/paddle/legacy/pserver/test/SocketTest.cpp b/paddle/legacy/pserver/test/SocketTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3a781fcbf655b554e79fc753f3409d12f10f6646 --- /dev/null +++ b/paddle/legacy/pserver/test/SocketTest.cpp @@ -0,0 +1,256 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Util.h" + +#include +#include +#include +#include +#include + +#include + +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/utils/Logging.h" + +struct MessageHeader { + int64_t dataLength; +}; + +class Thread { + public: + void start(); + virtual void run() = 0; + virtual ~Thread() {} + + protected: + std::unique_ptr thread_; +}; + +void Thread::start() { + thread_.reset(new std::thread([this]() { this->run(); })); +} + +class SocketChannel { + public: + explicit SocketChannel(int socket) : socket_(socket) {} + int getSocketFd() const { return socket_; } + uint64_t readAll(void* buf, size_t size); + uint64_t writeAll(const void* buf, size_t size); + + protected: + int socket_; +}; + +uint64_t SocketChannel::readAll(void* buf, size_t size) { + uint64_t total = 0; + while (total < size) { + int64_t len = read(socket_, (char*)buf + total, size - total); + if (len <= 0) { + return total; + } + total += len; + } + return total; +} + +uint64_t SocketChannel::writeAll(const void* buf, size_t size) { + uint64_t total = 0; + while (total < size) { + int64_t len = write(socket_, (const char*)buf + total, size - total); + if (len <= 0) { + return total; + } + total += len; + } + return total; +} + +class SocketWorker : public Thread { + public: + explicit SocketWorker(int socket) : channel_(socket) {} + virtual void run(); + + // read n bytes. + int64_t readAll(char* buf, size_t n); + + // write n bytes + + protected: + SocketChannel channel_; + std::string buffer_; +}; + +class SocketServer : public Thread { + public: + explicit SocketServer(int port) + : port_(port), socket_(0), maxPendingConnections_(100) {} + + virtual void run(); + + protected: + int port_; + int socket_; + int maxPendingConnections_; +}; + +void SocketServer::run() { + int newsockfd; + socklen_t clilen; + struct sockaddr_in serv_addr, cli_addr; + + /* First call to socket() function */ + socket_ = socket(AF_INET, SOCK_STREAM, 0); + CHECK(socket_ >= 0) << "ERROR opening socket"; + + /* Initialize socket structure */ + bzero((char*)&serv_addr, sizeof(serv_addr)); + serv_addr.sin_family = AF_INET; + serv_addr.sin_addr.s_addr = INADDR_ANY; + serv_addr.sin_port = htons(port_); + + /* Now bind the host address using bind() call.*/ + CHECK(bind(socket_, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0) + << "ERROR on binding"; + + /* Now start listening for the clients, here process will + * go in sleep mode and will wait for the incoming connection + */ + listen(socket_, maxPendingConnections_); + clilen = sizeof(cli_addr); + + while (true) { + /* Accept actual connection from the client */ + newsockfd = accept(socket_, (struct sockaddr*)&cli_addr, &clilen); + CHECK(newsockfd >= 0) << "ERROR on accept"; + + SocketWorker* worker = new SocketWorker(newsockfd); + worker->start(); + } +} + +void SocketWorker::run() { + MessageHeader header; + + while (true) { + int64_t n = channel_.readAll(&header, sizeof(header)); + CHECK(n == sizeof(header)) << "ERROR reading from socket"; + + buffer_.resize(header.dataLength); + n = channel_.readAll(&buffer_[0], header.dataLength); + CHECK(n == header.dataLength) << "ERROR reading from socket"; + + /* Write a response to the client */ + n = channel_.writeAll(&header, sizeof(header)); + CHECK(n == sizeof(header)) << "ERROR reading from socket"; + n = channel_.writeAll(buffer_.data(), buffer_.size()); + CHECK(n == header.dataLength) << "ERROR writing to socket"; + } +} + +class SocketClient { + public: + SocketClient(const std::string& serverAddr, int serverPort); + SocketChannel* getChannel() const { return channel_.get(); } + + protected: + std::unique_ptr channel_; +}; + +SocketClient::SocketClient(const std::string& serverAddr, int serverPort) { + struct sockaddr_in serv_addr; + struct hostent* server; + + // char buffer[256]; + + /* Create a socket point */ + int sockfd = socket(AF_INET, SOCK_STREAM, 0); + CHECK(sockfd >= 0) << "ERROR opening socket"; + server = gethostbyname(serverAddr.c_str()); + CHECK(server) << "ERROR, no such host: " << serverAddr; + + bzero((char*)&serv_addr, sizeof(serv_addr)); + serv_addr.sin_family = AF_INET; + bcopy((char*)server->h_addr, + (char*)&serv_addr.sin_addr.s_addr, + server->h_length); + serv_addr.sin_port = htons(serverPort); + + /* Now connect to the server */ + CHECK(connect(sockfd, (sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0) + << "ERROR connecting"; + + channel_.reset(new SocketChannel(sockfd)); +} + +DEFINE_string(server_addr, "127.0.0.1", "Server address"); +DEFINE_int64(dim, 10000000, "Data size"); +DEFINE_int32(loop_time, 100000, "test loop time"); + +using namespace paddle; // NOLINT + +int main(int argc, char** argv) { + paddle::initMain(argc, argv); + SocketServer server(FLAGS_port); + server.start(); + sleep(1); + + SocketClient client(FLAGS_server_addr, FLAGS_port); + + SocketChannel* channel = client.getChannel(); + + MessageHeader header; + + uint64_t dataSize = FLAGS_dim * sizeof(real); + +#ifdef PADDLE_WITH_CUDA + GpuVector gpuParam(FLAGS_dim); + GpuVector gpuGrad(FLAGS_dim); +#else + CpuVector gpuParam(FLAGS_dim); + CpuVector gpuGrad(FLAGS_dim); +#endif + CpuVector cpuParam(FLAGS_dim); + CpuVector cpuGrad(FLAGS_dim); + + gpuParam.rand(); + gpuGrad.rand(); + cpuParam.rand(); + cpuGrad.rand(); + + for (int i = 0; i < FLAGS_loop_time; ++i) { + cpuGrad.copyFrom(gpuGrad); + + header.dataLength = dataSize; + CHECK(channel->writeAll(&header, sizeof(header)) == sizeof(header)) + << "Client write header error"; + + CHECK(channel->writeAll(cpuGrad.getData(), dataSize) == dataSize) + << "Client write data error"; + + /* Now read server response */ + CHECK(channel->readAll(&header, sizeof(header)) == sizeof(header)) + << "Client read header error"; + + CHECK_EQ((uint64_t)header.dataLength, dataSize); + CHECK(channel->readAll(cpuParam.getData(), dataSize) == dataSize) + << "Client read data error"; + + gpuParam.copyFrom(cpuParam); + + LOG_EVERY_N(INFO, 100) << "i=" << i; + } + exit(0); +} diff --git a/paddle/legacy/pserver/test/test_ParameterServer2.cpp b/paddle/legacy/pserver/test/test_ParameterServer2.cpp new file mode 100644 index 0000000000000000000000000000000000000000..542e80e046972be38d403bc3223f7e7fcd15e3f0 --- /dev/null +++ b/paddle/legacy/pserver/test/test_ParameterServer2.cpp @@ -0,0 +1,624 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_int32(num_gradient_servers); +DEFINE_string(server_addr, "127.0.0.1", "assign server address"); +DEFINE_int32(server_cpu, 0, "assign server cpu"); + +class ParameterServer2Tester : public ParameterServer2 { + public: + ParameterServer2Tester(std::string serverAddr, + int port, + int rdmaCpu = -1, + bool sepSendAndRecv = false) + : ParameterServer2(serverAddr, port, rdmaCpu), client_(sepSendAndRecv) {} + virtual ~ParameterServer2Tester() {} + void setup() { + CHECK(ParameterServer2::init()); + + parameters_.clear(); + clientConfigs_.clear(); + + clientConfigs_.resize(2); + { + ParameterConfig& config = clientConfigs_[0]; + config.set_name("para0"); + config.set_para_id(0); + config.set_size(10000); + config.set_device(-1); + config.set_learning_rate(1.0); + config.set_momentum(0.9); + } + + { + ParameterConfig& config = clientConfigs_[1]; + config.set_name("para1"); + config.set_para_id(1); + config.set_size(5000); + config.set_device(-1); + config.set_learning_rate(0.5); + config.set_momentum(0.4); + } + + for (auto& config : clientConfigs_) { + parameters_.emplace_back(new Parameter(config, /* useGpu= */ false)); + } + + size_t id = 0; + for (auto& para : parameters_) { + para->setID(id++); + } + + CHECK(client_.init(parameters_)); + OptimizationConfig optConfig; + optConfig.set_algorithm("async_sgd"); + optConfig.set_batch_size(100); + optConfig.set_learning_rate(0.1); + client_.setConfig(optConfig); + client_.setParameter(); + } + + void setConfigTest(); + void setStatusTest(); + void sendParameterTest(); + void sendDataTest(SendDataType type, size_t size); + void operationTest(); + void mergeBlockSegmentTest(); + void checkSegments(const BlockSegments& expected, const BlockSegments& segs); + void waitPassFinishTest(); + void synchronizeTest(); + + protected: + ParameterClient2 client_; + vector clientConfigs_; + vector parameters_; +}; + +std::unique_ptr g_server; + +void ParameterServer2Tester::setConfigTest() { + setup(); + + for (auto& config : clientConfigs_) { + auto it = configMap_.find(config.para_id()); + EXPECT_TRUE(it != configMap_.end()); + auto& serverConfig = it->second; + EXPECT_EQ(config.name(), serverConfig.name()); + EXPECT_EQ(config.size(), serverConfig.size()); + EXPECT_EQ(config.learning_rate(), serverConfig.learning_rate()); + EXPECT_EQ(config.momentum(), serverConfig.momentum()); + } +} + +void ParameterServer2Tester::setStatusTest() { + setup(); + EXPECT_TRUE(client_.inStatus(PSERVER_STATUS_NOT_SET)); + client_.setStatus(PSERVER_STATUS_PARAMETER_READY); + EXPECT_EQ(PSERVER_STATUS_PARAMETER_READY, status_); + EXPECT_TRUE(client_.inStatus(PSERVER_STATUS_PARAMETER_READY)); +} + +real sumVector(const CpuVector& vec) { + const real* data = vec.getData(); + size_t dim = vec.getSize(); + real sum = 0; + for (size_t i = 0; i < dim; ++i) { + sum += data[i]; + } + return sum; +} + +void ParameterServer2Tester::sendParameterTest() { + setup(); + + client_.sendAndReceiveParameter(PSERVER_UPDATE_MODE_SET_PARAM, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + false); // sendBackParameter = false + + vector parameterCopies; + + for (auto& parameter : parameters_) { + parameterCopies.emplace_back( + new Parameter(parameter->getConfig(), /* useGpu= */ false)); + parameterCopies.back() + ->getBuf(PARAMETER_VALUE) + ->copyFrom(*parameter->getBuf(PARAMETER_VALUE)); + } + + client_.sendAndReceiveParameter(PSERVER_UPDATE_MODE_GET_PARAM, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true); // sendBackParameter = true + + for (size_t i = 0; i != parameters_.size(); ++i) { + real* v1 = parameters_[i]->getBuf(PARAMETER_VALUE)->getData(); + real* v2 = parameterCopies[i]->getBuf(PARAMETER_VALUE)->getData(); + EXPECT_EQ(parameters_[i]->getSize(), parameterCopies[i]->getSize()); + size_t size = parameters_[i]->getSize(); + real sum1 = 0, sum2 = 0; + for (size_t j = 0; j < size; ++j) { + sum1 += v1[j]; + sum2 += v2[j]; + } + EXPECT_EQ(sum1, sum2); + } +} + +void ParameterServer2Tester::sendDataTest(SendDataType type, size_t size) { + ParameterClient2 client1(true); + client1.init(parameters_); + ParameterClient2 client2(true); + client2.init(parameters_); + ParameterClient2 client3(true); + client3.init(parameters_); + + ThreadWorker worker1; + ThreadWorker worker2; + ThreadWorker worker3; + + double* testData1 = new double[size]; + double* testData2 = new double[size]; + double* testData3 = new double[size]; + double* getDataExpect = new double[size]; + double* getDataReal = new double[size]; + for (size_t i = 0; i < size; ++i) { + testData1[i] = rand(); // NOLINT TODO(yuyang18): Use rand_r instead. + testData2[i] = rand(); // NOLINT + testData3[i] = rand(); // NOLINT + getDataExpect[i] = testData1[i] + testData2[i] + testData3[i]; + } + + auto put1 = [&]() { + LOG(INFO) << "putOwnData1 start"; + client1.putOwnData(0, type, testData1, size); + LOG(INFO) << "putOwnData1 finish"; + }; + + auto get1 = [&]() { + LOG(INFO) << "sendData1 get all start"; + client1.getAllData(0, type, getDataReal, size); + for (size_t i = 0; i < size; ++i) { + CHECK_EQ(getDataReal[i], getDataExpect[i]); + } + LOG(INFO) << "sendData1 get all finish"; + }; + + auto put2 = [&]() { + LOG(INFO) << "putOwnData2 start"; + client2.putOwnData(1, type, testData2, size); + LOG(INFO) << "putOwnData2 finish"; + }; + + auto put3 = [&]() { + LOG(INFO) << "putOwnData3 start"; + client3.putOwnData(2, type, testData3, size); + LOG(INFO) << "putOwnData3 finish"; + }; + + worker1.addJob(put1); + worker1.addJob(get1); + worker2.addJob(put2); + worker3.addJob(put3); + + worker1.addJob(put1); + worker2.addJob(put2); + worker3.addJob(put3); + worker1.addJob(get1); + + worker1.wait(); + worker2.wait(); + worker3.wait(); + free(testData1); + free(testData2); + free(testData3); + free(getDataExpect); + free(getDataReal); +} + +void ParameterServer2Tester::operationTest() { + PServerVector v1, v2; + v1 = client_.createVector(); + EXPECT_EQ(NUM_PARAMETER_TYPES, v1.handle); + + v2 = client_.createVector(); + EXPECT_EQ(NUM_PARAMETER_TYPES + 1, v2.handle); + + PreparedOperations ops; + ops.addOperation(PSERVER_OP_RESET, v1, (real)1); + ops.addOperation(PSERVER_OP_RESET, v2, (real)2); + + real res1, res2, res3; + ops.addOperation(PSERVER_OP_utv, v1, v2)(&res1); + + ops.addOperation(PSERVER_OP_au_bv, v1, v2, (real)-1, (real)1); + ops.addOperation(PSERVER_OP_utv, v1, v2)(&res2); + + ops.addOperation(PSERVER_OP_au_bv, v1, v2, (real)-1, (real)1); + ops.addOperation(PSERVER_OP_utv, v1, v2)(&res3); + client_.doOperation(ops, false, false); + + EXPECT_EQ(30000, res1); + EXPECT_EQ(15000, res2); + EXPECT_EQ(0, res3); + + PServerMatrix m1, m2; + m1 = client_.createMatrix(4); + EXPECT_EQ(0, m1.handle); + m2 = client_.createMatrix(8); + EXPECT_EQ(1, m2.handle); + + // TODO(yuyang18): add tests for other operations OP_COPY, OP_au + + client_.releaseVector(v1); + client_.releaseVector(v2); + client_.releaseMatrix(m1); + client_.releaseMatrix(m2); +} + +void ParameterServer2Tester::checkSegments(const BlockSegments& expected, + const BlockSegments& segs) { + EXPECT_EQ(expected.size(), segs.size()); + if (expected.size() != segs.size()) { + return; + } + for (size_t i = 0; i < expected.size(); ++i) { + EXPECT_EQ(expected[i], segs[i]); + } +} + +void ParameterServer2Tester::mergeBlockSegmentTest() { + { + BlockSegments segs{{10, 20}, {30, 45}, {50, 70}}; + mergeSegments(&segs); + checkSegments({{10, 20}, {30, 45}, {50, 70}}, segs); + } + { + BlockSegments segs{{30, 45}, {50, 70}, {10, 20}}; + mergeSegments(&segs); + checkSegments({{10, 20}, {30, 45}, {50, 70}}, segs); + } + { + BlockSegments segs{{30, 45}, {50, 70}, {10, 30}}; + mergeSegments(&segs); + checkSegments({{10, 45}, {50, 70}}, segs); + } + { + BlockSegments segs{{30, 45}, {10, 70}, {10, 30}}; + mergeSegments(&segs); + checkSegments({{10, 70}}, segs); + } + { + BlockSegments segs{{30, 45}, {50, 70}, {10, 35}}; + mergeSegments(&segs); + checkSegments({{10, 45}, {50, 70}}, segs); + } + { + BlockSegments segs{{30, 45}, {50, 70}, {10, 60}}; + mergeSegments(&segs); + checkSegments({{10, 70}}, segs); + } + { + BlockSegments segs{{30, 45}, {50, 70}, {30, 47}}; + mergeSegments(&segs); + checkSegments({{30, 47}, {50, 70}}, segs); + } +} + +void ParameterServer2Tester::waitPassFinishTest() { + ParameterClient2 client1; + ParameterClient2 client2; + ParameterClient2 client3; + + ThreadWorker worker1; + ThreadWorker worker2; + ThreadWorker worker3; + + auto init1 = [&]() { + LOG(INFO) << "init1 start"; + client1.init(parameters_); + LOG(INFO) << "init1 finish"; + }; + + auto init2 = [&]() { + LOG(INFO) << "init2 start"; + client2.init(parameters_); + LOG(INFO) << "init2 finish"; + }; + + auto init3 = [&]() { + LOG(INFO) << "init3 start"; + client3.init(parameters_); + LOG(INFO) << "init3 finish"; + }; + + auto update1 = [&]() { + LOG(INFO) << "update1 start"; + client1.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ADD_GRADIENT, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true); // sendBackParameter = false + LOG(INFO) << "update1 finish"; + }; + + auto wait1 = [&]() { + LOG(INFO) << "wait1 start"; + client1.waitPassFinish(); + LOG(INFO) << "wait1 finish"; + }; + + auto update2 = [&]() { + LOG(INFO) << "update2 start"; + client2.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ADD_GRADIENT, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true); // sendBackParameter = false + LOG(INFO) << "update2 finish"; + }; + + auto wait2 = [&]() { + LOG(INFO) << "wait2 start"; + client2.waitPassFinish(); + LOG(INFO) << "wait2 finish"; + }; + + auto op3 = [&]() { + LOG(INFO) << "op3 start"; + PreparedOperations ops; + ops.addOperation(PSERVER_OP_SGD); + client3.doOperation(ops, + /* waitForGradient= */ true, + /* sendBackarameter= */ true); + LOG(INFO) << "op3 finish"; + }; + + worker1.addJob(init1); + worker2.addJob(init2); + worker3.addJob(init3); + + worker1.addJob(update1); + worker2.addJob(update2); + worker3.addJob(op3); + + worker3.addJob(op3); + worker3.addJob(op3); + worker2.addJob(update2); + worker2.addJob(update2); + worker1.addJob(wait1); + + worker2.addJob(wait2); + worker3.addJob(op3); + + worker1.wait(); + worker2.wait(); + worker3.wait(); + + LOG(INFO) << "Pass 1 finished"; + + worker1.addJob(update1); + worker2.addJob(update2); + worker3.addJob(op3); + + worker1.wait(); + worker2.wait(); + worker3.wait(); + + worker3.addJob(op3); + worker3.addJob(op3); + worker1.addJob(update1); + worker1.addJob(wait1); + worker2.addJob(wait2); + + worker1.wait(); + worker2.wait(); + worker3.wait(); + + LOG(INFO) << "Pass 2 finished"; +} + +void ParameterServer2Tester::synchronizeTest() { + ParameterClient2 client1; + ParameterClient2 client2; + + ThreadWorker worker1; + ThreadWorker worker2; + + FLAGS_log_period_server = 2; + + auto init1 = [&]() { + LOG(INFO) << "init1 start"; + client1.init(parameters_); + client1.setTrainerId(0); + LOG(INFO) << "init1 finish"; + }; + + auto init2 = [&]() { + LOG(INFO) << "init2 start"; + client2.init(parameters_); + client2.setTrainerId(1); + LOG(INFO) << "init2 finish"; + }; + + auto update1 = [&]() { + LOG(INFO) << "update1 start"; + client1.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ASYNC_SGD, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true); // sendBackParameter = false + LOG(INFO) << "update1 finish"; + }; + + auto wait1 = [&]() { + LOG(INFO) << "wait1 start"; + client1.asyncFinishPass(); + LOG(INFO) << "wait1 finish"; + }; + + auto update2 = [&]() { + LOG(INFO) << "update2 start"; + client2.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ASYNC_SGD, + PARAMETER_VALUE, + 0, // numSamples = 0 + 0, // cost = 0 + true); // sendBackParameter = false + LOG(INFO) << "update2 finish"; + }; + + auto wait2 = [&]() { + LOG(INFO) << "wait2 start"; + client2.asyncFinishPass(); + LOG(INFO) << "wait2 finish"; + }; + + worker1.addJob(init1); + worker2.addJob(init2); + // call wait to reset some stats at pserver + worker1.addJob(wait1); + worker2.addJob(wait2); + + worker1.addJob(update1); + worker2.addJob(update2); + + worker2.addJob(update2); + worker2.addJob(update2); + worker1.addJob(wait1); + + worker2.addJob(wait2); + + worker1.wait(); + worker2.wait(); + LOG(INFO) << "Pass 1 finished"; + + worker1.addJob(update1); + worker2.addJob(update2); + + worker1.wait(); + worker2.wait(); + + worker1.addJob(update1); + worker2.addJob(update2); + worker1.addJob(update1); + worker1.addJob(update1); + worker1.addJob(update1); + worker1.addJob(update1); + worker1.addJob(update1); + worker1.addJob(update1); + worker1.addJob(wait1); + worker2.addJob(wait2); + + worker1.wait(); + worker2.wait(); + LOG(INFO) << "Pass 2 finished"; +} + +TEST(ParameterServer2, sendParameter) { g_server->sendParameterTest(); } + +TEST(ParameterServer2, setConfig) { g_server->setConfigTest(); } + +TEST(ParameterServer2, setStatus) { g_server->setStatusTest(); } + +TEST(ParameterServer2, operation) { g_server->operationTest(); } + +TEST(ParameterServer2, mergeBlockSegment) { g_server->mergeBlockSegmentTest(); } + +TEST(ParameterServer2, waitPassFinish) { g_server->waitPassFinishTest(); } + +TEST(ParameterServer2, synchronize) { g_server->synchronizeTest(); } + +TEST(ParameterServer2, sendData) { + // Set gserver and pserver all 3, so that the test is sufficient. + int oldFlagsPortsNUm = FLAGS_ports_num; + int oldFlagsNumGradientServers = FLAGS_num_gradient_servers; + int oldFlagsPort = FLAGS_port; + FLAGS_ports_num = 3; + FLAGS_num_gradient_servers = 3; + FLAGS_port = FLAGS_port + 1; + std::unique_ptr g_server1; + std::unique_ptr g_server2; + std::unique_ptr g_server3; + if (FLAGS_rdma_tcp == "rdma") { + g_server1.reset(new ParameterServer2Tester( + FLAGS_server_addr, FLAGS_port, FLAGS_server_cpu)); + g_server1->start(); + g_server2.reset(new ParameterServer2Tester( + FLAGS_server_addr, FLAGS_port + 1, FLAGS_server_cpu + 1)); + g_server2->start(); + g_server3.reset(new ParameterServer2Tester( + FLAGS_server_addr, FLAGS_port + 2, FLAGS_server_cpu + 2)); + g_server3->start(); + } else { // tcp + g_server1.reset(new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port)); + g_server1->start(); + g_server2.reset( + new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port + 1)); + g_server2->start(); + g_server3.reset( + new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port + 2)); + g_server3->start(); + } + + g_server2->init(); + g_server3->init(); + sleep(2); + g_server1->setup(); + g_server1->sendDataTest(DATA_REDUCE_SUM, 1 << 24); + sleep(2); + g_server1->sendDataTest(DATA_REDUCE_SUM, 2); + sleep(2); + g_server1.reset(); + g_server2.reset(); + g_server3.reset(); + + FLAGS_ports_num = oldFlagsPortsNUm; + FLAGS_num_gradient_servers = oldFlagsNumGradientServers; + FLAGS_port = oldFlagsPort; +} + +int main(int argc, char** argv) { + paddle::initMain(argc, argv); + testing::InitGoogleTest(&argc, argv); + + FLAGS_num_gradient_servers = 2; + + if (FLAGS_rdma_tcp == "rdma") { + g_server.reset(new ParameterServer2Tester( + FLAGS_server_addr, FLAGS_port, FLAGS_server_cpu)); + } else { + g_server.reset(new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port)); + } + + g_server->start(); + + sleep(2); + + int ret = RUN_ALL_TESTS(); + + g_server.reset(); + + exit(ret); +} diff --git a/paddle/legacy/pserver/test/test_ProtoServer.cpp b/paddle/legacy/pserver/test/test_ProtoServer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f7ab2e8af45f97a6537d41ca1afe51a4d3270b80 --- /dev/null +++ b/paddle/legacy/pserver/test/test_ProtoServer.cpp @@ -0,0 +1,169 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "ParameterService.pb.h" +#include "paddle/legacy/math/Vector.h" +#include "paddle/legacy/pserver/ProtoServer.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +DEFINE_string(server_addr, "127.0.0.1", "Server address"); +DEFINE_int64(dim, 50000000, "Data size"); +DEFINE_bool(test_proto_server, true, "whether to test ProtoServer"); +DEFINE_bool(benchmark, false, "Do benchmark. Skip some tests"); + +using namespace paddle; // NOLINT + +class MyServer : public ProtoServer { + public: + explicit MyServer(int port, int rdmaCpu = -1) + : ProtoServer(FLAGS_server_addr, port, rdmaCpu), + status_(PSERVER_STATUS_NOT_SET) { + REGISTER_SERVICE_FUNCTION(MyServer, getStatus); + REGISTER_SERVICE_FUNCTION(MyServer, setStatus); + REGISTER_SERVICE_FUNCTION_EX(MyServer, getStatusEx); + } + void getStatus(const GetStatusRequest& request, + ProtoResponseCallback callback) { + (void)request; + GetStatusResponse response; + response.set_status(status_); + callback(response); + } + + void getStatusEx(const GetStatusRequest& request, + std::unique_ptr msgReader, + ProtoResponseCallbackEx callback) { + (void)request; + GetStatusResponse response; + response.set_status(status_); + buffer_.resize(msgReader->getNextBlockLength()); + msgReader->readNextBlock(&buffer_[0]); + callback(response, {{&buffer_[0], buffer_.size()}}); + } + + void setStatus(const SetStatusRequest& request, + ProtoResponseCallback callback) { + SetStatusResponse response; + status_ = request.status(); + callback(response); + } + + protected: + PServerStatus status_; + std::string buffer_; +}; + +TEST(ProtoServer, regular) { + ProtoClient* client; + if (FLAGS_rdma_tcp == "rdma") + client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA); + else + client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_TCP); + { + GetStatusRequest request; + GetStatusResponse response; + auto msgReader = client->sendAndRecv("getStatus", request, &response); + EXPECT_EQ(response.status(), PSERVER_STATUS_NOT_SET); + EXPECT_EQ(msgReader->getNumBlocks(), (size_t)0); + } + + { + SetStatusRequest request; + SetStatusResponse response; + request.set_status(PSERVER_STATUS_PARAMETER_READY); + client->sendAndRecv("setStatus", request, &response); + } + + { + GetStatusRequest request; + GetStatusResponse response; + client->sendAndRecv("getStatus", request, &response); + EXPECT_EQ(response.status(), PSERVER_STATUS_PARAMETER_READY); + } + + delete client; +} + +TEST(ProtoServer, extended) { +#ifdef PADDLE_WITH_CUDA + ProtoClient* client; + if (FLAGS_rdma_tcp == "rdma") + client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA); + else + client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_TCP); + int64_t dataSize = FLAGS_dim * sizeof(real); + + GpuVector gpuParam(FLAGS_dim); + GpuVector gpuGrad(FLAGS_dim); + CpuVector cpuParam(FLAGS_dim); + CpuVector cpuGrad(FLAGS_dim); + + gpuParam.rand(); + gpuGrad.rand(); + cpuParam.rand(); + cpuGrad.rand(); + + for (int k = 0; k < 4; ++k) { + for (int i = 0; i < 10; ++i) { + cpuGrad.copyFrom(gpuGrad); + if (FLAGS_test_proto_server) { + GetStatusRequest request; + GetStatusResponse response; + { + REGISTER_TIMER("sendAndRecv"); + auto msgReader = + client->sendAndRecv("getStatusEx", + request, + {{cpuGrad.getData(), (size_t)dataSize}}, + &response); + + EXPECT_EQ(msgReader->getNumBlocks(), (size_t)1); + EXPECT_EQ(msgReader->getNextBlockLength(), (size_t)dataSize); + msgReader->readNextBlock(cpuParam.getData()); + } + if (!FLAGS_benchmark) { + real* v1 = cpuGrad.getData(); + real* v2 = cpuParam.getData(); + real sum1 = 0, sum2 = 0; + for (int j = 0; j < FLAGS_dim; ++j) { + sum1 += v1[j]; + sum2 += v2[j]; + } + EXPECT_EQ(sum1, sum2); + } + } + gpuParam.copyFrom(cpuParam); + + LOG_EVERY_N(INFO, 10) << "i=" << i; + } + globalStat.printAllStatus(); + globalStat.reset(); + } + + delete client; +#endif +} + +int main(int argc, char** argv) { + paddle::initMain(argc, argv); + testing::InitGoogleTest(&argc, argv); + MyServer server(FLAGS_port, FLAGS_rdma_tcp == "rdma" ? 0 : -1); + server.start(); + usleep(10000); + + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/pserver/test/test_ProtoServer.sh b/paddle/legacy/pserver/test/test_ProtoServer.sh new file mode 100755 index 0000000000000000000000000000000000000000..1439350847308cc5590329b0fe2a6d2c77d04409 --- /dev/null +++ b/paddle/legacy/pserver/test/test_ProtoServer.sh @@ -0,0 +1,33 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +set -x +for ((port=12340;port<=12360;port++)) +do + port_used_num=`netstat -a |grep $port|wc -l` + if [ $port_used_num -eq 0 ] + then + echo $port; + legacy/pserver/test/test_ProtoServer --port=$port + if [ $? -eq 0 ] + then + exit 0 + else + echo "test_ProtoServer run wrong" + exit 1 + fi +fi +done +echo "test_ProtoServer port not found" +exit 1 diff --git a/paddle/trainer/CMakeLists.txt b/paddle/legacy/trainer/CMakeLists.txt similarity index 100% rename from paddle/trainer/CMakeLists.txt rename to paddle/legacy/trainer/CMakeLists.txt diff --git a/paddle/legacy/trainer/MergeModel.cpp b/paddle/legacy/trainer/MergeModel.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8a3601f192224a43687191527374149d99285ae0 --- /dev/null +++ b/paddle/legacy/trainer/MergeModel.cpp @@ -0,0 +1,64 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "ParamUtil.h" +#include "Trainer.h" +#include "paddle/legacy/pserver/ParameterServer2.h" +#include "paddle/legacy/utils/PythonUtil.h" + +DEFINE_string(model_dir, "", "Directory for separated model files"); +DEFINE_string(config_file, "", "Config file for the model"); +DEFINE_string(model_file, "", "File for merged model file"); + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +int main(int argc, char** argv) { + initMain(argc, argv); + initPython(argc, argv); + + if (FLAGS_model_dir.empty() || FLAGS_config_file.empty() || + FLAGS_model_file.empty()) { + LOG(INFO) << "Usage: ./paddle_merge_model --model_dir=pass-00000 " + "--config_file=config.py --model_file=out.paddle"; + return 0; + } + + string confFile = FLAGS_config_file; +#ifndef PADDLE_WITH_CUDA + FLAGS_use_gpu = false; +#endif + auto config = std::make_shared(confFile); + unique_ptr gradientMachine(GradientMachine::create(*config)); + gradientMachine->loadParameters(FLAGS_model_dir); + + ofstream os(FLAGS_model_file); + + string buf; + config->getConfig().SerializeToString(&buf); + int64_t size = buf.size(); + os.write((char*)&size, sizeof(size)); + CHECK(os) << "Fail to write to " << FLAGS_model_file; + os.write(buf.data(), buf.size()); + vector& parameters = gradientMachine->getParameters(); + for (auto& para : parameters) { + para->save(os); + CHECK(os) << "Fail to write to " << FLAGS_model_file; + } + os.close(); + + return 0; +} diff --git a/paddle/legacy/trainer/NewRemoteParameterUpdater.cpp b/paddle/legacy/trainer/NewRemoteParameterUpdater.cpp new file mode 100644 index 0000000000000000000000000000000000000000..cdd832acd16e5c259a7f6463aac537e4e6537c97 --- /dev/null +++ b/paddle/legacy/trainer/NewRemoteParameterUpdater.cpp @@ -0,0 +1,150 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "NewRemoteParameterUpdater.h" +#include "Trainer.h" +#include "paddle/legacy/utils/Stat.h" + +DECLARE_int32(trainer_id); +DECLARE_string(save_dir); + +namespace paddle { +NewRemoteParameterUpdater::NewRemoteParameterUpdater( + const OptimizationConfig &config, const std::string pserverSpec) + : trainerConfig_(config), + parameterClient_(-1), + newParameters_(nullptr), + newGradients_(nullptr), + pserverSpec_(pserverSpec) {} + +NewRemoteParameterUpdater::NewRemoteParameterUpdater( + const OptimizationConfig &config, + const std::string pserverSpec, + const bool useEtcd) + : trainerConfig_(config), + parameterClient_(-1), + newParameters_(nullptr), + newGradients_(nullptr), + pserverSpec_(pserverSpec), + useEtcd_(useEtcd) {} + +void NewRemoteParameterUpdater::init( + const std::vector ¶meters) { + ParameterUpdater::init(parameters); + + // create parameter server client. + if (useEtcd_) { + parameterClient_ = + paddle_new_etcd_pserver_client((char *)pserverSpec_.c_str()); + } else { + parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(), + FLAGS_trainer_id == 0); + } + + // init new parameter and gradient. + newParameters_ = initNewParameter(PARAMETER_VALUE); + newGradients_ = initNewParameter(PARAMETER_GRADIENT); + + // init parameter, one trainer will get the opportunity to int parameter and + // send them to parameter server. Others will get the initialized parameter + // from parameter server + if (paddle_begin_init_params(parameterClient_)) { + LOG(INFO) << "paddle_begin_init_params start"; + // NOTE: convert V1 OptimizatioinConfig proto to V2 OptimizerConfig. + // This makes golang pserver compatible with handy V1 demos. + // TODO(wuyi): Refine or remove these ugly converting lines + OptimizerConfig optimizerConfigV2; + if (trainerConfig_.learning_method() == "momentum") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); + } else if (trainerConfig_.learning_method() == "adagrad") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad); + optimizerConfigV2.mutable_adagrad()->set_epsilon( + trainerConfig_.ada_epsilon()); + } else if (trainerConfig_.learning_method() == "adadelta") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad); + optimizerConfigV2.mutable_adadelta()->set_epsilon( + trainerConfig_.ada_epsilon()); + optimizerConfigV2.mutable_adadelta()->set_rho(trainerConfig_.ada_rou()); + } else if (trainerConfig_.learning_method() == "adam") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adam); + optimizerConfigV2.mutable_adam()->set_beta_1(trainerConfig_.adam_beta1()); + optimizerConfigV2.mutable_adam()->set_beta_2(trainerConfig_.adam_beta2()); + optimizerConfigV2.mutable_adam()->set_epsilon( + trainerConfig_.adam_epsilon()); + } else { + LOG(ERROR) << "got unsupported v1 optimizer config: " + << trainerConfig_.learning_method(); + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); + } + + if (trainerConfig_.learning_rate_schedule() == "constant") { + optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); + optimizerConfigV2.mutable_const_lr()->set_learning_rate( + trainerConfig_.learning_rate()); + } else if (trainerConfig_.learning_rate_schedule() == "linear") { + optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Linear); + optimizerConfigV2.mutable_linear_lr()->set_learning_rate( + trainerConfig_.learning_rate()); + optimizerConfigV2.mutable_linear_lr()->set_lr_decay_a( + trainerConfig_.learning_rate_decay_a()); + optimizerConfigV2.mutable_linear_lr()->set_lr_decay_b( + trainerConfig_.learning_rate_decay_b()); + } else { + LOG(ERROR) << "got unsupported v1 learning_rate_schedule config: " + << trainerConfig_.learning_rate_schedule() << ", set to const"; + optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); + optimizerConfigV2.mutable_const_lr()->set_learning_rate( + trainerConfig_.learning_rate()); + } + + // overwrite optimizerConfigV2 for per-parameter(layer) configs + for (int i = 0; i < parameterSize(); ++i) { + // FIXME(typhoonzero): paramConfig always have default values, + // how to check if it's default? + // TODO(typhoonzero): log output: optimizerConfigV2.DebugString(); + LOG(INFO) << "trainerConfig_: " << trainerConfig_.DebugString(); + // send param and config to pserver + std::string bytes = optimizerConfigV2.SerializeAsString(); + const char *array = bytes.data(); + int size = (int)bytes.size(); + paddle_init_param( + parameterClient_, *newParameters_[i], (void *)array, size); + } + paddle_finish_init_params(parameterClient_); + LOG(INFO) << "paddle_begin_init_params done"; + } else { + paddle_get_params(parameterClient_, newParameters_, parameterSize()); + } + + LOG(INFO) << "NewRemoteParameterUpdater initialized"; +} + +void NewRemoteParameterUpdater::updateImpl(Parameter *para) {} + +void NewRemoteParameterUpdater::finishBatch(real cost) { + // send gradient to parameter server. + paddle_send_grads(parameterClient_, newGradients_, parameterSize()); + // get the updated parameter from parameterClient. + paddle_get_params(parameterClient_, newParameters_, parameterSize()); + + // clear gradient after update parameter. + for (auto ¶ : parameters_) { + para->getBuf(PARAMETER_GRADIENT)->zeroMem(); + } +} + +void NewRemoteParameterUpdater::startPass() {} + +bool NewRemoteParameterUpdater::finishPass() { return true; } +} // namespace paddle diff --git a/paddle/legacy/trainer/NewRemoteParameterUpdater.h b/paddle/legacy/trainer/NewRemoteParameterUpdater.h new file mode 100644 index 0000000000000000000000000000000000000000..707e9ceb9b6a22d265f9bf7b02af7f3002930fd4 --- /dev/null +++ b/paddle/legacy/trainer/NewRemoteParameterUpdater.h @@ -0,0 +1,121 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "OptimizerConfig.pb.h" +#include "ParameterUpdater.h" +#include "libpaddle_pserver_cclient.h" +#include "paddle/legacy/pserver/ParameterClient2.h" +#include "paddle/legacy/utils/Queue.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +/** + * New remote parameter updater for dense parameters that use cclient of go. + */ +class NewRemoteParameterUpdater : public ParameterUpdater { + public: + NewRemoteParameterUpdater(const OptimizationConfig& config, + const std::string pserverSpec); + NewRemoteParameterUpdater(const OptimizationConfig& config, + const std::string pserverSpec, + const bool useEtcd); + ~NewRemoteParameterUpdater() { + releaseNewParameter(newParameters_); + releaseNewParameter(newGradients_); + if (parameterClient_ >= 0) paddle_pserver_client_release(parameterClient_); + } + + /** + * initialize the internal parameter client and itself. + */ + virtual void init(const std::vector& parameters); + /** + * @brief start batch + * + * @note one batch training exhibits stateful feature to help + * to do performance tuning, sgd optimization if necessary. + */ + virtual PassType startBatch(int64_t batchSize) { return PASS_TRAIN; } + + /** + * send parameters to pservers and get returned parameters + * from all pservers if necessary. + */ + virtual void finishBatch(real cost); + virtual void startPass(); + virtual bool finishPass(); + + protected: + /** + * work need to do after finishBatch + */ + virtual void updateImpl(Parameter* para); + + private: + int parameterSize() { return (int)parameters_.size(); } + + /** + * init parameter of go paddle pserver cclient. + * @param new_params + * @param type + */ + paddle_parameter** initNewParameter(ParameterType type) { + paddle_parameter** new_params = + (paddle_parameter**)malloc(sizeof(paddle_parameter*) * parameterSize()); + for (int i = 0; i < parameterSize(); ++i) { + new_params[i] = (paddle_parameter*)malloc(sizeof(paddle_parameter)); + memset(new_params[i], 0, sizeof(paddle_parameter)); + } + + for (int i = 0; i < parameterSize(); ++i) { + ParameterPtr param = parameters_[i]; + new_params[i]->element_type = PADDLE_ELEMENT_TYPE_FLOAT32; + new_params[i]->name = (char*)param->getName().c_str(); + new_params[i]->content = + (unsigned char*)(param->getBuf(type).get()->getData()); + new_params[i]->content_len = + (int)param->getBuf(type).get()->getSize() * sizeof(real); + } + return new_params; + } + + void releaseNewParameter(paddle_parameter** newParams) { + if (newParams != nullptr) { + for (int i = 0; i < parameterSize(); ++i) { + free(newParams[i]); + } + free(newParams); + } + } + + protected: + const OptimizationConfig& trainerConfig_; + /// internal parameter client object for exchanging data with pserver + paddle_pserver_client parameterClient_; + /// the parameters for new pserver client + paddle_parameter** newParameters_; + /// the gradinets for new pserver client + paddle_parameter** newGradients_; + /// the specification of parameter server "host1:port,host1:port" + std::string pserverSpec_; + /// true if pserverSpec_ is etcd endpoint, else pserverSpec_ is pserver addr + bool useEtcd_; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/ParamUtil.cpp b/paddle/legacy/trainer/ParamUtil.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b5aba32dee1d07015ae3fce1cc76242b8ae80fe5 --- /dev/null +++ b/paddle/legacy/trainer/ParamUtil.cpp @@ -0,0 +1,163 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ParamUtil.h" + +#include +#include + +#include +#include +#include +#include + +#include +#include + +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +#include "TesterConfig.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/gserver/layers/ValidationLayer.h" + +namespace paddle { + +ParameterUtil::ParameterUtil( + const std::shared_ptr &config, + std::unique_ptr &&intconfig, + const GradientMachinePtr &gradientMachine, + const std::shared_ptr ¶meterUpdater) { + config_ = config; + intConfig_ = std::move(intconfig); + gserver_ = gradientMachine; + pUpdater_ = parameterUpdater; +} + +bool ParameterUtil::loadParameters(int passId, bool local, bool remote) { + constexpr int kBufLen = 100; + char buf[kBufLen]; + snprintf(buf, kBufLen, "pass-%05d", passId); + std::string doneFile = path::join(config_->getSaveDir(), buf, "done"); + if (!fileExist(doneFile.c_str())) return false; + loadParametersWithPath(path::join(config_->getSaveDir(), buf), local, remote); + return true; +} + +void ParameterUtil::loadParametersWithPath(const std::string &dir, + bool local, + bool remote) { + if (local) { + gserver_->loadParameters(dir); + } + if (remote && pUpdater_) { + pUpdater_->loadParametersRemote(dir); + } +} + +void ParameterUtil::saveParametersOnePass(int passId, int passInnerId) { + pUpdater_->apply(); + saveParameters(passId, passInnerId); + if (intConfig_->save_only_one_ && passId >= intConfig_->saving_period_) { + deleteParameters(passId - intConfig_->saving_period_); + } + pUpdater_->restore(); +} + +void ParameterUtil::saveParameters(int passId, int passInnerId) { + constexpr int kBufLen = 100; + char buf[kBufLen]; + if (passInnerId > 0) { + snprintf(buf, kBufLen, "pass-%05d-%03d", passId, passInnerId); + } else { + snprintf(buf, kBufLen, "pass-%05d", passId); + } + + std::string basePath = config_->getSaveDir(); + if (basePath.find('/') == std::string::npos) { + basePath = "./" + basePath; + } + mkDirRecursively(basePath.c_str()); + + std::string saveDir = path::join(basePath, buf); + mkDir(saveDir.c_str()); + if (!intConfig_->load_save_param_pserver_) { + pUpdater_->getParametersRemote(true /*full parameter*/, + true /*after apply*/); + } + + gserver_->saveParameters(saveDir); + if (intConfig_->load_save_param_pserver_) { + pUpdater_->saveParametersRemote(saveDir); + } + std::string doneFile = path::join(saveDir, "done"); + touchFile(doneFile.c_str()); + std::ofstream out(doneFile); + version::printVersion(out); + out.close(); + VLOG(1) << "save dir " << saveDir; + saveConfigWithPath(saveDir); +} + +void ParameterUtil::deleteParameters(int passId, int passInnerId) { + constexpr int kBufLen = 100; + char buf[kBufLen]; + const std::string &saveDir = config_->getSaveDir(); + if (passInnerId > 0) { + snprintf(buf, + kBufLen, + "%s/pass-%05d-%03d", + saveDir.c_str(), + passId, + passInnerId); + } else { + snprintf(buf, kBufLen, "%s/pass-%05d", saveDir.c_str(), passId); + } + mkDir(saveDir.c_str()); + LOG(INFO) << "delete dir " << buf; + rmDir(buf); +} + +void ParameterUtil::saveConfigWithPath(const std::string &path) { + std::string src; + // save config in some path + if (!intConfig_->config_.empty()) { + src = intConfig_->config_; + } else { + bool ok; + src = config_->getConfigName(&ok); + if (!ok) { + return; + } + } + copyFileToPath(src, path); + + // save other import config file name to path.txt + std::string ss = path::join(path, "path.txt"); + std::ofstream os(ss); + std::string fileName = path::basename(src); + CHECK(os.write(fileName.c_str(), fileName.length())) + << "Fail to write config file name " << ss; + VLOG(1) << "fileName " << fileName; + os.close(); + + // copy other import config files + for (int i = 0; i < config_->getConfig().config_files_size(); ++i) { + copyFileToPath(config_->getConfig().config_files(i), path); + } +} + +} // namespace paddle diff --git a/paddle/legacy/trainer/ParamUtil.h b/paddle/legacy/trainer/ParamUtil.h new file mode 100644 index 0000000000000000000000000000000000000000..07786967762a7b9267d190de5275f0f94bbd21ef --- /dev/null +++ b/paddle/legacy/trainer/ParamUtil.h @@ -0,0 +1,125 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include + +#include "hl_gpu.h" +#include "paddle/legacy/gserver/dataproviders/DataProvider.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" + +#include +#include +#include "ParameterUpdater.h" +#include "TrainerConfig.pb.h" +#include "TrainerConfigHelper.h" + +namespace paddle { + +/** + * Configuration for parameter utils. + */ +struct ParameterUtilConfig { + DISABLE_COPY(ParameterUtilConfig); + + ParameterUtilConfig(bool save_only_one, + int saving_period, + bool load_save_parameters_in_pserver, + std::string config) + : save_only_one_(save_only_one), + saving_period_(saving_period), + load_save_param_pserver_(load_save_parameters_in_pserver), + config_(config) {} + + bool save_only_one_; + int saving_period_; + bool load_save_param_pserver_; + std::string config_; +}; + +/** + * ParameterUtil + * Utility class for loading and saving parameters + */ +class ParameterUtil { + public: + /** + * Ctor. + * + * @param config + * @param intconfig + * @param gradientMachine + * @param parameterUpdater + * @return + */ + ParameterUtil(const std::shared_ptr &config, + std::unique_ptr &&intconfig, + const GradientMachinePtr &gradientMachine, + const std::shared_ptr ¶meterUpdater); + + /// Load parameter from the saved parameter file as pass passId + /// if loadsave_parameters_in_pserver is set, some parameters MUST + /// load in pserver, which is "remote". + /// loadParameters can choose to load local/remote parameter, or both. + bool loadParameters(int passId, bool local = true, bool remote = false); + + /// load parameters given path info + void loadParametersWithPath(const std::string &dir, + bool local = true, + bool remote = false); + + /// Save parameter to dist for pass passId + /// passInnerId means saving times in one pass, some users want to + /// save parameters when have processed some batches in one pass + /// passInnerId = 0 means do not need to save in one inner pass + void saveParameters(int passId, int passInnerId = 0); + + /// save parameters for one pass, when passInnerId > 0 means saving + /// the passInnerId times in one pass + void saveParametersOnePass(int passId, int passInnerId = 0); + + /// delete parameter from disk via passId + void deleteParameters(int passId, int passInnerId = 0); + + /// save config given path info + void saveConfigWithPath(const std::string &path); + + /** + * Try to load parameter from config. + * @return true if can load from trainer config. + */ + inline bool tryLoadParametersFromConfig() { + auto &c = config_->getConfig(); + if (!c.init_model_path().empty()) { + loadParametersWithPath(c.init_model_path()); + return true; + } else if (c.start_pass() > 0) { + CHECK(loadParameters(c.start_pass() - 1)); + return true; + } else { + return false; + } + } + + private: + std::shared_ptr config_; + std::unique_ptr intConfig_; + GradientMachinePtr gserver_; + std::shared_ptr pUpdater_; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/ParameterUpdater.cpp b/paddle/legacy/trainer/ParameterUpdater.cpp new file mode 100644 index 0000000000000000000000000000000000000000..549fb0332da78053a261928b5558beb1ffbc79c5 --- /dev/null +++ b/paddle/legacy/trainer/ParameterUpdater.cpp @@ -0,0 +1,152 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ParameterUpdater.h" + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/utils/Thread.h" + +namespace paddle { + +static const hl_stream_t kDeviceToHostStream = HPPL_STREAM_1; +static const hl_stream_t kHostToDeviceStream = HPPL_STREAM_2; + +SgdUpdaterWithCpuAverager::SgdUpdaterWithCpuAverager( + const OptimizationConfig& optConfig) + : SgdLocalUpdater(optConfig, false /*with averager*/) { + CHECK(FLAGS_use_gpu && optConfig.do_average_in_cpu()); + averager_.reset(AverageOptimizer::create(optConfig, + new DummyOptimizer(optConfig), + false /*sparse*/, + true /*apply*/)); + updateWorker_.addJob([]() { hl_set_device(FLAGS_gpu_id); }); +} + +void SgdUpdaterWithCpuAverager::init( + const std::vector& parameters) { + SgdLocalUpdater::init(parameters); + averager_->init(parameters_.size(), nullptr); + copyEvents_.resize(parameters_.size()); + for (auto& parameter : parameters) { + SetDevice device(parameter->getDeviceId()); + cpuParameters_.emplace_back(new Parameter(parameter->getConfig(), + /* useGpu= */ false, + /* doInit= */ false)); + if (parameter->useGpu()) { + cpuParameters_.back()->enableType(PARAMETER_APPLY); + } else { + cpuParameters_.back()->enableSharedType( + PARAMETER_APPLY, parameter->getBuf(PARAMETER_VALUE)); + } + for (ParameterType type : averager_->getParameterTypes()) { + cpuParameters_.back()->enableType(type); + } + + hl_create_event(©Events_[nonStaticParaIDMap_[parameter->getID()]]); + } +} + +SgdUpdaterWithCpuAverager::~SgdUpdaterWithCpuAverager() { + for (auto& event : copyEvents_) { + hl_destroy_event(event); + } +} + +void SgdUpdaterWithCpuAverager::updateImpl(Parameter* para) { + SgdLocalUpdater::updateImpl(para); + + if (para->useGpu()) { + size_t pid = nonStaticParaIDMap_[para->getID()]; + Parameter* cpuPara = cpuParameters_[pid].get(); + cpuPara->getBuf(PARAMETER_VALUE) + ->copyFrom(*para->getBuf(PARAMETER_VALUE), kDeviceToHostStream); + hl_stream_record_event(kDeviceToHostStream, copyEvents_[pid]); + } + + updateWorker_.addJob( + std::bind(&SgdUpdaterWithCpuAverager::updateFunc, this, para)); +} + +void SgdUpdaterWithCpuAverager::updateFunc(Parameter* para) { + SetDevice setDevice(para->getDeviceId()); + size_t pid = nonStaticParaIDMap_[para->getID()]; + Parameter* cpuPara = cpuParameters_[pid].get(); + if (para->useGpu()) { + hl_event_synchronize(copyEvents_[pid]); + } + averager_->update(cpuPara->getBufs(), cpuPara->getConfig(), -1LU); +} + +void SgdUpdaterWithCpuAverager::finishBatch(real cost) { + SgdLocalUpdater::finishBatch(cost); + + updateWorker_.wait(); + for (auto para : cpuParameters_) { + if (auto callback = averager_->needSpecialTraversal(para->getConfig())) { + callback(para->getBufs(), para->getConfig(), -1LU); + } + } + averager_->finishBatch(); +} + +void SgdUpdaterWithCpuAverager::apply() { + // backup gpu value + for (auto& para : parameters_) { + SetDevice setDevice(para->getDeviceId()); + para->getBuf(PARAMETER_GRADIENT) + ->copyFrom(*para->getBuf(PARAMETER_VALUE), kHostToDeviceStream); + } + + // apply on cpu parameter + if (auto callback = averager_->apply()) { + for (auto para : cpuParameters_) { + callback(para->getBufs(), para->getConfig(), -1LU); + } + } + + // copy to gpu value + for (auto& para : parameters_) { + SetDevice setDevice(para->getDeviceId()); + size_t pid = nonStaticParaIDMap_[para->getID()]; + Parameter* cpuPara = cpuParameters_[pid].get(); + if (parameters_[pid]->useGpu()) { + para->getBuf(PARAMETER_VALUE) + ->copyFrom(*cpuPara->getBuf(PARAMETER_APPLY), kHostToDeviceStream); + } + } + hl_stream_synchronize(kHostToDeviceStream); + for (auto& para : parameters_) { + para->setValueUpdated(); + } +} + +void SgdUpdaterWithCpuAverager::restore() { + // restore on cpu parameter + if (auto callback = averager_->restore()) { + for (auto para : cpuParameters_) { + callback(para->getBufs(), para->getConfig(), -1LU); + } + } + + // restore gpu value + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(PARAMETER_VALUE)->copyFrom(*para->getBuf(PARAMETER_GRADIENT)); + para->getBuf(PARAMETER_GRADIENT)->zeroMem(); + para->setValueUpdated(); + } +} + +} // namespace paddle diff --git a/paddle/legacy/trainer/ParameterUpdater.h b/paddle/legacy/trainer/ParameterUpdater.h new file mode 100644 index 0000000000000000000000000000000000000000..acddc3702d78fdb198973f70a8642c5192af992b --- /dev/null +++ b/paddle/legacy/trainer/ParameterUpdater.h @@ -0,0 +1,265 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Thread.h" +#include "paddle/legacy/utils/Util.h" + +#include "paddle/legacy/parameter/AverageOptimizer.h" +#include "paddle/legacy/parameter/FirstOrderOptimizer.h" +#include "paddle/legacy/parameter/OptimizerFunctions.h" +#include "paddle/legacy/parameter/OptimizerWithRegularizer.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/parameter/ParameterUpdaterBase.h" + +#include "TrainerConfig.pb.h" +#include "paddle/legacy/gserver/layers/Layer.h" + +#include +#include + +namespace paddle { + +/** + * @brief Parameter Updater for SGD, and local(not cluster) run. + */ +class SgdLocalUpdater : public ParameterUpdater { + public: + /** + * @brief Ctor. Initialize optimizer locally by optConfig. + * @param optConfig optimization config. + * @param withAverager with average optimizer or not, default is true. + */ + explicit SgdLocalUpdater(const OptimizationConfig& optConfig, + bool withAverager = true) + : numSamplesProcessed_(0) { + auto baseOptimizer = ParameterOptimizer::create(optConfig); + optimizer_.reset(withAverager + ? AverageOptimizer::create(optConfig, baseOptimizer) + : baseOptimizer); + CHECK(optimizer_) << "fail to create optimizer: " + << optConfig.learning_method(); + auto types = optimizer_->getParameterTypes(); + for (auto type : types) { + addParameterType(type); + } + } + + /** + * @brief Initialize parameters and optimizer_. + * For example, + * If optimizer need hassien vector, then parameter's hassien will + * be initialized. + * @param parameters The parameter need to be initialized. + */ + virtual void init(const std::vector& parameters) { + ParameterUpdater::init(parameters); + optimizer_->init(parameters_.size(), nullptr); + // check no L1 decay in parameter configs + CHECK(std::find_if(parameters.begin(), + parameters.end(), + [](const ParameterPtr& para) { + return para->getConfig().decay_rate_l1() > 0.0f; + }) == parameters.end()) + << "SgdLocalUpdater cannot support L1 decay in parameter"; + } + + /** + * @brief Start a batch with current mini-batch size + * @param current mini-batch size. + * @return Always PASS_TRAIN. + */ + virtual PassType startBatch(int64_t batchSize) { + numSamplesProcessed_ += batchSize; + optimizer_->startBatch(numSamplesProcessed_); + return PASS_TRAIN; + } + + /** + * @brief finish a mini-batch. + */ + virtual void finishBatch(real cost) { optimizer_->finishBatch(); } + + /** + * @brief start a pass. + */ + virtual void startPass() { optimizer_->startPass(); } + + /** + * @brief finish a pass. + * @param cost sum cost during one pass. + * @return true if accept (used for owlqn). + */ + virtual bool finishPass() { + optimizer_->finishPass(); + return ParameterUpdater::finishPass(); + } + + /** + * @brief apply model average. + */ + virtual void apply() { + if (auto callback = optimizer_->apply()) { + for (auto para : parameters_) { + SetDevice device(para->getDeviceId()); + callback(para->getBufs(), para->getConfig(), -1UL); + } + } + } + + /** + * @brief restore parameter value before model average + */ + virtual void restore() { + if (auto callback = optimizer_->restore()) { + for (auto para : parameters_) { + SetDevice device(para->getDeviceId()); + callback(para->getBufs(), para->getConfig(), -1UL); + } + } + } + + protected: + /** + * @brief update method. Update value from gradient. + * @param para parameter that will be updated. + */ + virtual void updateImpl(Parameter* para) { + optimizer_->update(para->getBufs(), para->getConfig()); + if (auto callback = optimizer_->needSpecialTraversal(para->getConfig())) { + callback(para->getBufs(), para->getConfig(), -1UL); + } + + para->setValueUpdated(); + para->getBuf(PARAMETER_GRADIENT)->zeroMem(); + } + + std::unique_ptr optimizer_; + + /** + * @brief total number of samples processed. + */ + int64_t numSamplesProcessed_; +}; + +/** + * @brief SgdCpuUpdater is used only in recursive neural network + * @deprecated + */ +class SgdCpuUpdater : public SgdLocalUpdater, public Deprecated { + public: + explicit SgdCpuUpdater(const OptimizationConfig& optConfig) + : SgdLocalUpdater(optConfig), + Deprecated( + "SgdCpuUpdater is used only in recursive neural network, " + "and recursive neural network is deprecated in paddle. " + "Use it all by your own.") {} + + /** + * @brief update all parameter on finish batch. + * @param cost + */ + virtual void finishBatch(real cost) { + for (auto para : parameters_) { + SgdLocalUpdater::update(para.get()); + } + optimizer_->finishBatch(); + } + + protected: + /** + * @brief do nothing. + * @param para + */ + virtual void updateImpl(Parameter* para) {} +}; + +/** + * @brief Sgd Local Updater With average in cpu. + * + * It will do model average in cpu to reduce gpu memory comsuption. + */ +class SgdUpdaterWithCpuAverager : public SgdLocalUpdater { + public: + /** + * @brief Ctor. + * + * SgdUpdaterWithCpuAverager will do everything as a + * SgdLocalUpdater, then copy parameter from GPU to CPU, and do model + * average in cpu. + */ + explicit SgdUpdaterWithCpuAverager(const OptimizationConfig& optConfig); + ~SgdUpdaterWithCpuAverager(); + + /** + * @brief init. Initialize cpu parameters, model average optimizer. + * @param parameters + */ + virtual void init(const std::vector& parameters); + + virtual PassType startBatch(int64_t batchSize) { + averager_->startBatch(-1UL); + return SgdLocalUpdater::startBatch(batchSize); + } + virtual void finishBatch(real cost); + + virtual void startPass() { + averager_->startPass(); + SgdLocalUpdater::startPass(); + } + virtual bool finishPass() { + averager_->finishPass(); + return SgdLocalUpdater::finishPass(); + } + + /// apply the averaged parameter to PARAMETER_VALUE + /// use PARAETER_GRADIENT for backing up PARAMETER_VALUE + virtual void apply(); + + /** + * @brief Restore parameter before apply(). + */ + virtual void restore(); + + protected: + virtual void updateImpl(Parameter* para); + + void updateFunc(Parameter* para); + + protected: + std::unique_ptr averager_; + + /** + * @brief The thread worker which do model average. + * + * For each parameter, GPU->CPU parameter is async, and do model average in + * another thread. Because the training process don't need model average while + * training, and model average only used in evaluation stage and saving stage. + * So the model average is totally async. + */ + ThreadWorker updateWorker_; + + /** + * @brief The parameter mirror in cpu. + */ + std::vector cpuParameters_; + + /** + * @brief GPU -> CPU copy event. Model average will wait after copy done. + */ + std::vector copyEvents_; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/RemoteParameterUpdater.cpp b/paddle/legacy/trainer/RemoteParameterUpdater.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5de1cc7827aa8f219de60fe9da67fbb0595eb1d5 --- /dev/null +++ b/paddle/legacy/trainer/RemoteParameterUpdater.cpp @@ -0,0 +1,843 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "RemoteParameterUpdater.h" +#include "Trainer.h" +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/Stat.h" + +DECLARE_int32(trainer_id); +DECLARE_string(save_dir); + +namespace paddle { + +static const hl_stream_t kDeviceToHostStream = HPPL_STREAM_1; +static const hl_stream_t kHostToDeviceStream = HPPL_STREAM_2; +static const int kFinishBatchPid = -1; + +const std::string RemoteParameterUpdater::kAverage = "average"; +const std::string RemoteParameterUpdater::kElasticAverage = "elastic_average"; + +RemoteParameterUpdater::RemoteParameterUpdater( + const OptimizationConfig& config, + int expectedPassCount, + std::unique_ptr&& localUpdater) + : config_(config), + localUpdater_(std::move(localUpdater)), + numBatches_(0), + passCount_(0), + expectedPassCount_(expectedPassCount), + separateSendAndRecv_(false), + isFirstPass_(true), + useApplyInPserver_(false) { + addParameterType(PARAMETER_MOMENTUM); +} + +void RemoteParameterUpdater::init(const std::vector& parameters) { + ParameterUpdater::init(parameters); + + if (localUpdater_) { + localUpdater_->init(parameters); + + for (auto& parameter : parameters) { + parameter->enableType(PARAMETER_DELTA); + } + + CHECK(config_.center_parameter_update_method() == kAverage || + config_.center_parameter_update_method() == kElasticAverage) + << "unknown center_parameter_update_method"; + + // modify delta_add_rate + CHECK_GT(FLAGS_num_gradient_servers, 1) + << "FLAGS_num_gradient_servers should be set in trainer args."; + real delta_add_rate = config_.delta_add_rate() / FLAGS_num_gradient_servers; + config_.set_delta_add_rate(delta_add_rate); + LOG(INFO) << "center parameter in pserver," + << " modify delta_add_rate=" << delta_add_rate; + } + + if (!FLAGS_use_gpu) { + cpuParameters_ = parameters; + } else { + for (auto& parameter : parameters) { + cpuParameters_.emplace_back(new Parameter(parameter->getConfig(), + /* useGpu= */ false)); + cpuParameters_.back()->setID(parameter->getID()); + if (localUpdater_) { + cpuParameters_.back()->enableType(PARAMETER_DELTA); + } + } + } + + parameterClient_.reset(new ParameterClient2(separateSendAndRecv_)); + parameterClient_->init(cpuParameters_); + parameterClient_->setTrainerId(FLAGS_trainer_id); + + if (FLAGS_trainer_id == 0) { + parameterClient_->setConfig(config_); + copyParametersFromDevice(PARAMETER_VALUE); + parameterClient_->setParameter(); + parameterClient_->setStatus(PSERVER_STATUS_PARAMETER_READY); + } else { + parameterClient_->waitForStatus(PSERVER_STATUS_PARAMETER_READY); + parameterClient_->getParameter(); + copyParametersToDevice(PARAMETER_VALUE); + } + if (FLAGS_trainer_id == 0 && + (config_.algorithm() != TrainAlgorithm::AsyncSGD)) { + startController(); + useApplyInPserver_ = useApplyInPserver(config_); + } +} + +void RemoteParameterUpdater::startController() { + controllerThread_.reset(new std::thread([this]() { this->controller(); })); +} + +void RemoteParameterUpdater::controller() { + ParameterClient2 client(false); + client.init(cpuParameters_); + while (true) { + /*start pass*/ { + client.waitPassStart(); + + PreparedOperations ops; + ops.addOperation(PSERVER_OP_START_PASS); + client.doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false, + /* releasePass= */ false); + } + + while (true) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_SGD); + client.doOperation(ops, + /* waitForGradient= */ true, + /* sendBackarameter= */ true, + /* releasePass= */ false); + if (client.isPassFinish()) { + break; + } + } + + /*finish pass*/ { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_FINISH_PASS); + client.doOperation(ops, + /* waitForGradient= */ true, + /* sendBackarameter= */ true, + /* releasePass= */ true); + } + + passCount_++; + if (passCount_ == expectedPassCount_) { + break; + } + } +} + +void RemoteParameterUpdater::copyParametersToDevice( + ParameterType parameterType) { + if (!FLAGS_use_gpu) { + return; + } + int numParameters = cpuParameters_.size(); + for (int i = 0; i < numParameters; ++i) { + parameters_[i] + ->getBuf(parameterType) + ->copyFrom(*cpuParameters_[i]->getBuf(parameterType)); + if (parameterType == PARAMETER_VALUE) { + parameters_[i]->setValueUpdated(); + } + } +} + +void RemoteParameterUpdater::copyParametersFromDevice( + ParameterType parameterType) { + if (!FLAGS_use_gpu) { + return; + } + int numParameters = cpuParameters_.size(); + for (int i = 0; i < numParameters; ++i) { + cpuParameters_[i] + ->getBuf(parameterType) + ->copyFrom(*parameters_[i]->getBuf(parameterType)); + } +} + +void RemoteParameterUpdater::updateImpl(Parameter* para) { + REGISTER_TIMER("update"); + if (localUpdater_) { + localUpdater_->update(para); + } +} + +void RemoteParameterUpdater::finishBatch(real cost) { + if (localUpdater_) { + localUpdater_->finishBatch(cost); + } + + const std::string& algorithm = config_.algorithm(); + ParameterUpdateMode mode; + if (algorithm == TrainAlgorithm::AsyncSGD) { + mode = PSERVER_UPDATE_MODE_ASYNC_SGD; + } else if (algorithm == TrainAlgorithm::SGD) { + mode = PSERVER_UPDATE_MODE_ADD_GRADIENT; + } else { + LOG(FATAL) << "Unknown algorithm: " << algorithm; + } + + ParameterType sendType; + bool sendBackParameter = true; + if (localUpdater_) { + ++numBatches_; + if (numBatches_ % config_.num_batches_per_send_parameter() != 0) { + return; + } + + if (config_.center_parameter_update_method() == kElasticAverage) { + parameterClient_->getParameter(PARAMETER_DELTA); + copyParametersToDevice(PARAMETER_DELTA); + sendBackParameter = false; // no need send back after send + + // calc delta + for (auto& para : parameters_) { + // DELTA = LOCAL_VALUE - CENTER_VALUE/*store in DELTA*/ + para->getBuf(PARAMETER_DELTA) + ->add(*para->getBuf(PARAMETER_VALUE), -1.0f, 1.0f); + + // when delta send to pserver, pserver will do: + // CENTER_VALUE += alpha * (LOCAL_VALUE - CENTER_VALUE) + } + } else { + // calc delta + for (auto& para : parameters_) { + // DELTA = NEW_VALUE - OLD_VALUE/*store in DELTA*/ + para->getBuf(PARAMETER_DELTA) + ->add(*para->getBuf(PARAMETER_VALUE), -1.0f, 1.0f); + } + } + + sendType = PARAMETER_DELTA; + + } else { + // In this case, we perform SGD on pserver. + sendType = PARAMETER_GRADIENT; + } + + copyParametersFromDevice(sendType); + + { + REGISTER_TIMER("sendAndRecv_dense"); + parameterClient_->sendAndReceiveParameter(mode, + sendType, + batchSize_, + 0, // cost = 0 + sendBackParameter); + } + + if (sendBackParameter) { + copyParametersToDevice(PARAMETER_VALUE); + } + + if (localUpdater_) { + if (config_.center_parameter_update_method() == kElasticAverage) { + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + // LOCAL_VALUE += -alpha * (LOCAL_VALUE - CENTER_VALUE) + para->getBuf(PARAMETER_VALUE) + ->add(*para->getBuf(PARAMETER_DELTA), -config_.delta_add_rate()); + } + + } else { // average + // copy value to delta + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); + } + } + } else { + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(sendType)->zeroMem(); + } + } +} + +void RemoteParameterUpdater::startPass() { + if (config_.algorithm() == TrainAlgorithm::SGD) { + parameterClient_->waitPassStart(); + } else { + // sync could benifits reducing lagged trainer for async-sgd + // even if sync could not remove all lagged trainer for the + // sake of file loading, buffer etc. + parameterClient_->asyncStartPass(); + } + + if (localUpdater_) { + localUpdater_->startPass(); + numBatches_ = 0; + + if (config_.center_parameter_update_method() == kElasticAverage) { + if (!isFirstPass_) { + // restore local value from delta + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(PARAMETER_VALUE) + ->copyFrom(*para->getBuf(PARAMETER_DELTA)); + } + } + } else { // average + // copy value to delta + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); + } + } + } +} + +bool RemoteParameterUpdater::finishPass() { + if (localUpdater_) { + localUpdater_->finishPass(); + } + + if (config_.algorithm() == TrainAlgorithm::SGD) { + parameterClient_->waitPassFinish(); + } else { + parameterClient_->asyncFinishPass(); + } + if (localUpdater_) { + if (config_.center_parameter_update_method() == kElasticAverage) { + // backup local value to delta as we will get + // the remote parameter for saving/testing + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); + } + } + } + parameterClient_->getParameter(); + copyParametersToDevice(PARAMETER_VALUE); + + isFirstPass_ = false; + return true; +} + +void RemoteParameterUpdater::apply() { + if (useApplyInPserver_) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_APPLY); + parameterClient_->doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false); + parameterClient_->getParameter( + /* recvParameterType= */ PARAMETER_VALUE, + /* sendBackParameterType= */ PARAMETER_APPLY); + copyParametersToDevice(PARAMETER_VALUE); + } +} + +void RemoteParameterUpdater::restore() { + if (useApplyInPserver_) { + parameterClient_->getParameter(); + copyParametersToDevice(PARAMETER_VALUE); + } +} + +ConcurrentRemoteParameterUpdater::ConcurrentRemoteParameterUpdater( + OptimizationConfig config, + int passCount, + std::unique_ptr&& localUpdater) + : RemoteParameterUpdater(config, passCount, std::move(localUpdater)) { + sendThread_.reset(new std::thread([this]() { this->send(); })); + recvThread_.reset(new std::thread([this]() { this->recv(); })); + + stopping_ = false; + oneBatchFinished_ = false; + separateSendAndRecv_ = true; +} + +ConcurrentRemoteParameterUpdater::~ConcurrentRemoteParameterUpdater() { + stopping_ = true; + sendQueue_.enqueue(0); + sendThread_->join(); + recvQueue_.enqueue(0); + recvThread_->join(); +} + +void ConcurrentRemoteParameterUpdater::finishBatch(real cost) { + if (localUpdater_) { + localUpdater_->finishBatch(cost); + + if (!needToUpdateRemotely()) { + ++numBatches_; + return; + } + } + + sendQueue_.enqueue(kFinishBatchPid); + + finishBatchCond_.wait([this]() { return oneBatchFinished_; }); + oneBatchFinished_ = false; + { + REGISTER_TIMER("sync_hostToDeviceStream"); + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + hl_stream_synchronize(kHostToDeviceStream); + } + } + + if (localUpdater_) { + ++numBatches_; + } +} + +// Use para=NULL to signal the end of one batch +void ConcurrentRemoteParameterUpdater::send(Parameter* para) { + const std::string& algorithm = config_.algorithm(); + ParameterUpdateMode mode; + if (algorithm == TrainAlgorithm::AsyncSGD) { + mode = PSERVER_UPDATE_MODE_ASYNC_SGD; + } else if (algorithm == TrainAlgorithm::SGD) { + mode = PSERVER_UPDATE_MODE_ADD_GRADIENT; + } else { + LOG(FATAL) << "Unknown algorithm: " << algorithm; + } + ParameterType sendType; + if (localUpdater_) { + sendType = PARAMETER_DELTA; + } else { + // In this case, we perform SGD on pserver. + sendType = PARAMETER_GRADIENT; + } + std::vector paraSegment; + if (para == NULL) { + parameterClient_->sendParameter( + mode, + sendType, + paraSegment, + batchSize_, + 0, // cost=0 + true, // sendBackParameter = true + batchStatus_); // batchStatus_ = BATCH_FINISH + + } else { + ParameterSegments paraSegTemp; + paraSegment.reserve(1); + paraSegTemp.name = para->getName(); + paraSegTemp.id = para->getID(); + paraSegment.push_back(paraSegTemp); + { + SetDevice device(para->getDeviceId()); + REGISTER_TIMER("copySingleParaFromDevice"); + copySingleParaFromDevice(para, sendType); + hl_stream_synchronize(kDeviceToHostStream); + } + parameterClient_->sendParameter(mode, + sendType, + paraSegment, + batchSize_, + 0, // cost=0 + true, // sendBackParameter = true + batchStatus_); + if (batchStatus_ == BATCH_START) batchStatus_ = BATCH_ON; + } +} +void ConcurrentRemoteParameterUpdater::recv(Parameter* para) { + parameterClient_->recvParameter(); + if (para != NULL) { + REGISTER_TIMER("copySingleParaToDevice"); + SetDevice device(para->getDeviceId()); + copySingleParaToDevice(para, PARAMETER_VALUE); + + if (localUpdater_) { + para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); + } else { + // if cpu, parameter should not changes until recvParameter(). + // if gpu, zero mem when send finish + if (!FLAGS_use_gpu) { + para->getBuf(PARAMETER_GRADIENT)->zeroMem(); + } + } + } +} + +void ConcurrentRemoteParameterUpdater::recv() { + if (FLAGS_use_gpu) hl_set_device(FLAGS_gpu_id); + StatPtr stat = getStat("recv"); + FOR_TIMING(Timer timer); + while (true) { + int pid; + { + REGISTER_TIMER("recv_dequeue"); + pid = recvQueue_.dequeue(); + } + if (pid == kFinishBatchPid) { + Parameter* para = NULL; + FOR_TIMING(timer.start()); + recv(para); + FOR_TIMING(timer.stop()); + FOR_TIMING(stat->addSample(timer.get())); + FOR_TIMING(timer.reset()); + finishBatchCond_.notify_all([this] { oneBatchFinished_ = true; }); + } else { + if (stopping_) break; + Parameter* para = parameters_[pid].get(); + FOR_TIMING(timer.start()); + recv(para); + FOR_TIMING(timer.stop()); + oneBatchFinished_ = false; + } + } +} + +void ConcurrentRemoteParameterUpdater::send() { + if (FLAGS_use_gpu) hl_set_device(FLAGS_gpu_id); + StatPtr stat = getStat("send"); + FOR_TIMING(Timer timer); + while (true) { + int pid; + { + REGISTER_TIMER("send_dequeue"); + pid = sendQueue_.dequeue(); + } + if (pid == kFinishBatchPid) { + batchStatus_ = BATCH_FINISH; + if (!localUpdater_) { + // if cpu, parameter should not changes until recvParameter(). + // if gpu, zeroMem() at the end of batch so that it won't + // interfere with computation. + if (FLAGS_use_gpu) { + REGISTER_TIMER("para_zeroMem"); + for (auto& para : parameters_) { + SetDevice device(para->getDeviceId()); + para->getBuf(PARAMETER_GRADIENT)->zeroMem(); + } + } + } + Parameter* para = NULL; + FOR_TIMING(timer.start()); + send(para); + FOR_TIMING(timer.stop()); + FOR_TIMING(stat->addSample(timer.get())); + FOR_TIMING(timer.reset()); + recvQueue_.enqueue(pid); + } else { + if (stopping_) break; + Parameter* para = parameters_[pid].get(); + if (localUpdater_) { + // DELTA = NEW_VALUE - OLD_VALUE/*store in DELTA*/ + para->getBuf(PARAMETER_DELTA) + ->add(*para->getBuf(PARAMETER_VALUE), -1.0f, 1.0f); + } + FOR_TIMING(timer.start()); + send(para); + FOR_TIMING(timer.stop()); + recvQueue_.enqueue(nonStaticParaIDMap_[para->getID()]); + } + } +} + +void ConcurrentRemoteParameterUpdater::updateImpl(Parameter* para) { + REGISTER_TIMER("update"); + if (localUpdater_) { + localUpdater_->update(para); + if (!needToUpdateRemotely()) { + return; + } + } + sendQueue_.enqueue(nonStaticParaIDMap_[para->getID()]); +} + +void ConcurrentRemoteParameterUpdater::copySingleParaToDevice( + Parameter* para, ParameterType parameterType) { + if (!FLAGS_use_gpu) { + return; + } + int i = nonStaticParaIDMap_[para->getID()]; + para->getBuf(parameterType) + ->copyFrom(*cpuParameters_[i]->getBuf(parameterType), + kHostToDeviceStream); + if (parameterType == PARAMETER_VALUE) { + para->setValueUpdated(); + } +} + +void ConcurrentRemoteParameterUpdater::copySingleParaFromDevice( + Parameter* para, ParameterType parameterType) { + if (!FLAGS_use_gpu) { + return; + } + int i = nonStaticParaIDMap_[para->getID()]; + cpuParameters_[i] + ->getBuf(parameterType) + ->copyFrom(*para->getBuf(parameterType), kDeviceToHostStream); +} + +SparseRemoteParameterUpdater::SparseRemoteParameterUpdater( + const OptimizationConfig& config, int expectedPassCount, bool testing) + : config_(config), + passCount_(0), + expectedPassCount_(expectedPassCount), + testing_(testing), + useApplyInPserver_(false) {} + +void SparseRemoteParameterUpdater::init( + const std::vector& parameters) { + ParameterUpdater::init(parameters); + + parameterClient_.reset(new ParameterClient2( + false, FLAGS_port + FLAGS_ports_num, FLAGS_ports_num_for_sparse)); + parameterClient_->init(parameters_); + parameterClient_->setTrainerId(FLAGS_trainer_id); + + if (FLAGS_trainer_id == 0) { + parameterClient_->setConfig( + config_, FLAGS_save_dir, true /*is_sparse_server*/); + if (parameters[0]->isFullSize()) { + parameterClient_->setParameter(); + } else { // init in pserver + parameterClient_->setParameterZero(); + } + } + if (FLAGS_trainer_id == 0 && !testing_ && + config_.algorithm() == TrainAlgorithm::SGD) { + startController(); + useApplyInPserver_ = useApplyInPserver(config_); + } +} + +void SparseRemoteParameterUpdater::startController() { + controllerThread_.reset(new std::thread([this]() { this->controller(); })); +} + +void SparseRemoteParameterUpdater::controller() { + ParameterClient2 client( + false, FLAGS_port + FLAGS_ports_num, FLAGS_ports_num_for_sparse); + client.init(parameters_); + + while (true) { + /*start pass*/ { + client.waitPassStart(); + + PreparedOperations ops; + ops.addOperation(PSERVER_OP_START_PASS); + client.doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false, + /* releasePass= */ false); + } + + while (true) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_SGD); + client.doOperation(ops, + /* waitForGradient= */ true, + /* sendBackarameter= */ true, + /* releasePass= */ false); + if (client.isPassFinish()) { + break; + } + } + + /*finish pass*/ { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_FINISH_PASS); + client.doOperation(ops, + /* waitForGradient= */ true, + /* sendBackarameter= */ true, + /* releasePass= */ true); + } + + passCount_++; + if (passCount_ == expectedPassCount_) { + break; + } + } +} + +PassType SparseRemoteParameterUpdater::startBatch(int64_t batchSize) { + batchSize_ = batchSize; + return PASS_TRAIN; +} + +void SparseRemoteParameterUpdater::finishBatch(real cost) { + const std::string& algorithm = config_.algorithm(); + ParameterUpdateMode mode; + if (algorithm == TrainAlgorithm::AsyncSGD) { + mode = PSERVER_UPDATE_MODE_ASYNC_SGD; + } else if (algorithm == TrainAlgorithm::SGD) { + mode = PSERVER_UPDATE_MODE_ADD_GRADIENT; + } else { + LOG(FATAL) << "Unknown algorithm: " << algorithm; + } + + ParameterType sendType = PARAMETER_GRADIENT; + + REGISTER_TIMER("sendSparseParam"); + parameterClient_->sendAndReceiveParameter(mode, + sendType, + batchSize_, + 0, // cost = 0 + false); // sendBackParameter + + // grad zero move to sgd grad machine, before merge grad sparse remote +} + +void SparseRemoteParameterUpdater::startPass() { + if (config_.algorithm() == TrainAlgorithm::SGD) { + parameterClient_->waitPassStart(); + } else { + if (FLAGS_trainer_id == 0) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_START_PASS); + parameterClient_->doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false); + } + parameterClient_->asyncStartPass(); + } +} + +bool SparseRemoteParameterUpdater::finishPass() { + if (config_.algorithm() == TrainAlgorithm::SGD) { + parameterClient_->waitPassFinish(); + } else { + if (FLAGS_trainer_id == 0) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_FINISH_PASS); + parameterClient_->doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false); + } + parameterClient_->asyncFinishPass(); + } + + return true; +} + +// Trainer will call getParametersRemote at batch start or before save, +// so we do not get values in apply() and restore(). +void SparseRemoteParameterUpdater::apply() { + if (useApplyInPserver_) { + PreparedOperations ops; + ops.addOperation(PSERVER_OP_APPLY); + parameterClient_->doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false); + } +} + +void SparseRemoteParameterUpdater::restore() {} + +void SparseRemoteParameterUpdater::getParametersRemote(bool fullSize, + bool apply) { + ParameterType sendBackParameterType = + (useApplyInPserver_ && apply) ? PARAMETER_APPLY : PARAMETER_VALUE; + std::function getParams; + std::function applyL1; + if (fullSize) { + getParams = [&] { + parameterClient_->getParameter( + /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType); + }; + applyL1 = [](Parameter& para, real decayRate) { + para.getBuf(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate); + }; + } else { + getParams = [&] { + parameterClient_->getParameterSparse( + /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType); + }; + applyL1 = [](Parameter& para, real decayRate) { + para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate); + }; + } + { + REGISTER_TIMER("getParamDenseAndSparse"); + getParams(); + if (config_.shrink_parameter_value() > 0) { + for (auto& para : parameters_) { + if (para->getConfig().decay_rate_l1() > 0) { + applyL1(*para, config_.shrink_parameter_value()); + } + } + } + } +} + +void SparseRemoteParameterUpdater::randParametersRemote() { + CHECK_EQ(FLAGS_trainer_id, 0); + + PreparedOperations ops; + ops.addOperation(PSERVER_OP_RANDOMIZE); + parameterClient_->doOperation(ops, + /* waitForGradient= */ false, + /* sendBackarameter= */ false); +} + +void SparseRemoteParameterUpdater::loadParametersRemote( + const std::string& dirName) { + if (FLAGS_trainer_id == 0) { + parameterClient_->loadValueVector(dirName); + } + + if (testing_) { + // we do not use synchronize() here, + // because test mode may run only one tester + if (FLAGS_trainer_id == 0) { + parameterClient_->setStatus(PSERVER_STATUS_PARAMETER_READY); + } else { + parameterClient_->waitForStatus(PSERVER_STATUS_PARAMETER_READY); + } + } +} + +void SparseRemoteParameterUpdater::saveParametersRemote( + const std::string& dirName) { + if (FLAGS_trainer_id == 0) { + parameterClient_->saveValueVector(dirName); + } +} + +void SparseRemoteParameterUpdaterComposite::init( + const std::vector& parameters) { + parameters_ = parameters; + + std::vector parametersArray[NUMBER_UPDATERS]; + + for (auto& para : parameters_) { + if (para->isSparseRemoteUpdate()) { + parametersArray[UPDATER_SPARSE_REMOTE].push_back(para); + } else { + parametersArray[UPDATER_NORMAL].push_back(para); + } + } + CHECK(!parametersArray[UPDATER_SPARSE_REMOTE].empty()); + CHECK(!parametersArray[UPDATER_NORMAL].empty()); + + syncThreadPool_->execPlusOwner([&](int tid, size_t numThreads) { + updaters_[tid]->init(parametersArray[tid]); + }); + + parameterTypes_ = updaters_[UPDATER_NORMAL]->getParameterTypes(); +} + +std::vector> + ParameterUpdaterCreators::constructors_; + +} // namespace paddle diff --git a/paddle/legacy/trainer/RemoteParameterUpdater.h b/paddle/legacy/trainer/RemoteParameterUpdater.h new file mode 100644 index 0000000000000000000000000000000000000000..68468532981a49ef32f5f0da1170815d657d86c1 --- /dev/null +++ b/paddle/legacy/trainer/RemoteParameterUpdater.h @@ -0,0 +1,416 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "ParameterUpdater.h" +#include "paddle/legacy/pserver/ParameterClient2.h" +#include "paddle/legacy/utils/Queue.h" +#include "paddle/legacy/utils/Util.h" + +namespace paddle { + +// TODO(yanfei): +// I think that the biggest feature of rdma is packet lossless control +// feature instead of high bandwiths, zero copy and gpu-direct rdma in +// theroy. +// But zero-copy and gpu-direct rdma features can help to reduce latency +// caused by os system. +// So, for some specified cluster, such as high density gpu cluster, +// gpu-direct and zero copy could help to improve cluster communication +// performance. +// + +/** + * Normal remote parameter updater for dense parameters. + * + * It first packs all parameters for all pservers using ParameterClient + * module, then wait for merged parameters data from all pservers. + * The synchronization pattern specified by sync-sgd or async-sgd is + * achieved by all pservers with the help of the controller within this + * remote parameter updater. + * This module indeedly bridges the gradient machines and parameter servers. + * It helps to transfer the parameters from acceleration device to cpu end + * for network. It contains additional parameters copy buffers for + * acceleration devices at cpu end, such as gpu, otherwise it will + * directly use original parameters data to update pservers. + * + * This remote parameter updater does not use pipeline mechanism to hide + * copy latency from gpu to cpu buffer. In addition the overlapped between + * backward and communication is not supported. + */ +class RemoteParameterUpdater : public ParameterUpdater { + public: + RemoteParameterUpdater( + const OptimizationConfig& config, + int expectedPassCount, + std::unique_ptr&& localUpdater = nullptr); + ~RemoteParameterUpdater() { + if (controllerThread_) { + controllerThread_->join(); + } + } + + /** + * initialize the internal parameter client and itself. + */ + virtual void init(const std::vector& parameters); + /** + * @brief start batch + * + * @note one batch training exhibits stateful feature to help + * to do performance tuning, sgd optimization if necessary. + */ + virtual PassType startBatch(int64_t batchSize) { + if (localUpdater_) { + localUpdater_->startBatch(batchSize); + } + batchSize_ = batchSize; + batchStatus_ = BATCH_START; + return PASS_TRAIN; + } + + /** + * send parameters to pservers and get returned parameters + * from all pservers if necessary. it will implictly + * cooperate with controller thread for sync-sgd. + */ + virtual void finishBatch(real cost); + virtual void startPass(); + virtual bool finishPass(); + +#ifndef PADDLE_DISABLE_TIMER + virtual void setForwardbackwardTime(uint64_t delta) { + parameterClient_->setForwardbackwardTime(delta); + } +#endif + + virtual void apply(); + virtual void restore(); + + protected: + /** + * control all pservers with all trainers for sync-sgd + */ + virtual void controller(); + + /** + * work need to do after finishBatch + */ + virtual void updateImpl(Parameter* para); + + void startController(); + + /** + * @brief copy parameters from cpu host to device, such as gpu. + * + * @note return if all data are transfered. + */ + void copyParametersToDevice(ParameterType parameterType); + + /** + * @brief copy parameters from device to cpu host + * + * @note return if all data are transfered + */ + void copyParametersFromDevice(ParameterType parameterType); + + protected: + /// Optimization config used to guide initialization and finishBatch + OptimizationConfig config_; + /// internal parameter client object for exchanging data with pserver + std::unique_ptr parameterClient_; + /// internal shadow buffer at cpu host end, use original parameters_ + /// if no acceleration devices are used. + std::vector cpuParameters_; + /// local updater for aggregating multi-batches local delta + std::unique_ptr localUpdater_; + /// the size of mini-batch + int64_t batchSize_; + /// batches passed + int64_t numBatches_; + /// for stateful control + BatchStatus batchStatus_; + /// controller thread for sync-sgd + std::unique_ptr controllerThread_; + /// passed already finished + int64_t passCount_; + /// expected passes to finished + int64_t expectedPassCount_; + /// use normal synchronization communication if True + bool separateSendAndRecv_; + /// true if it's first pass + bool isFirstPass_; + bool useApplyInPserver_; + + static const std::string kAverage; + static const std::string kElasticAverage; +}; + +// TODO(yanfei): +// do parameters level synchronization Optimization at pserver end with +// ConcurrentRemoteParameterUpdater to get more parallelization, at last +// to really hide pserver latency in backward computation. +// +/** + * This updater add additional optimization for overlapping synchronization + * from pservers with backward computation. + * + * Parameter can be sent to pservers when related backward stage is finished. + * This concurrent udpater does data copy from acceleration device to host + * memory aynchronously. In addition internal parameter client reads data in + * host memory and send them to all pservers in next stage. So this class + * help to pipeline device-to-host copy and host-to-network to hide network + * latency in backward stage. + * It contains separate send and recv thread for pipeline usage. + */ +class ConcurrentRemoteParameterUpdater : public RemoteParameterUpdater { + public: + ConcurrentRemoteParameterUpdater( + OptimizationConfig config, + int expectedPassCount, + std::unique_ptr&& localUpdater); + ~ConcurrentRemoteParameterUpdater(); + + /** + * @brief send paraemeters to all pservers + * + * @note it just signal the end signal to internal parameter client + * to finished the aynchronous send action. In addition it also + * do synchronization for all asynchronous host-to-device copy. + */ + virtual void finishBatch(real cost); + + protected: + virtual void updateImpl(Parameter* para); + /// internal thread called in send thread + void send(Parameter* para); // para == NULL indicate end of a minibatch + /// internal function called in recv thread + void recv(Parameter* para); + /** + * @brief send thread for relaying data from gradient to parameter client + * + * @note just pipe data to internal parameter client for pipeline + */ + void send(); + /** + * @brief recv thread for relaying data from internal parameter client to + * host memory + * + * @note it contains the asynchronous data copy form host to device + */ + void recv(); + /// copy specified parameter from host to device + void copySingleParaToDevice(Parameter* para, ParameterType parameterType); + /// copy specified parameter from device to host + void copySingleParaFromDevice(Parameter* para, ParameterType parameterType); + bool needToUpdateRemotely() { + return (numBatches_ + 1) % config_.num_batches_per_send_parameter() == 0; + } + + private: + /// send thread used for overlapping + std::unique_ptr sendThread_; + /// recv thread used for overlapping + std::unique_ptr recvThread_; + /// buffer queue for overlapping + Queue sendQueue_; + /// buffer queue for overlapping + Queue recvQueue_; + /// flags indicating to stop + bool stopping_; + /// conditional variable for threads synchronization between the + /// thread calling finishBatch and internal recv thread + LockedCondition finishBatchCond_; + bool oneBatchFinished_; +}; + +// TODO(yanfei): +// merge sparse updater with dense updater, and could help to reduce +// the synchronization between sparse and dense udpater. it could also +// reduce the threads for managing all connections. +/** + * This class is specified for updating sparse parameters. + * + * It allows part of parameter to be exchanged with all pservers. + * If sparse input assigned, part gradients of first hidden layer + * could remained zero which can not need to be exchanged within + * all pservers. This is the key optimization point for this updater + * + * For updating sparse parameters, all latest parameters are stored + * in pservers instead of keeping full copy at train end, so need to + * prefetch parameters weight value which can be changed in next-batch + * before doing next forwardbackward. Also, with above fact that the + * parameters can be stored in pserver instead of trainer, we can + * fetch specified parmeters if necessary, and can support huge + * parameters which is larger enough than the RAM size in single + * node. + * + * Internally, this updater will direct internal parameter client + * to encapsulate sparse specified message for all pservers. + */ +class SparseRemoteParameterUpdater : public ParameterUpdater { + public: + SparseRemoteParameterUpdater(const OptimizationConfig& config, + int expectedPassCount, + bool testing); + ~SparseRemoteParameterUpdater() { + if (controllerThread_) { + controllerThread_->join(); + } + } + + /// initialization + virtual void init(const std::vector& parameters); + + /// stateful batch control + virtual PassType startBatch(int64_t batchSize); + /// send all sparse related parameters to all pservers + virtual void finishBatch(real cost); + virtual void startPass(); + virtual bool finishPass(); + + virtual void apply(); + virtual void restore(); + + /// load parameters from pservers + virtual void loadParametersRemote(const std::string& dirName); + /// save parameters to pservers + virtual void saveParametersRemote(const std::string& dirName); + /** + * @brief get latest sparse parameters value from all pservers + * + * @note call it before next mini-batch + */ + virtual void getParametersRemote(bool fullSize, bool apply); + virtual void randParametersRemote(); +#ifndef PADDLE_DISABLE_TIMER + virtual void setForwardbackwardTime(uint64_t delta) { + parameterClient_->setForwardbackwardTime(delta); + } +#endif + + protected: + /// update implimentation, not implemented + virtual void updateImpl(Parameter* para) {} + + /// internal controller routine for controller thread + virtual void controller(); + + /// start controller thread + void startController(); + + protected: + /// optimization config + OptimizationConfig config_; + /// internal parameter client + std::unique_ptr parameterClient_; + int64_t batchSize_; + std::unique_ptr controllerThread_; + int64_t passCount_; + int64_t expectedPassCount_; + bool testing_; + bool useApplyInPserver_; +}; + +/** + * Class for supporting normal updater and sparse updater + * + * Not all parts of one model are sparse, so it exists dense updater + * for normal layers while sparse updater is for sparse layers. + * + * it directly call internal dense and sparse udpater individually. + */ +class SparseRemoteParameterUpdaterComposite : public ParameterUpdaterComposite { + public: + enum { + UPDATER_SPARSE_REMOTE = 0, // execute in sync thread pool(tid:0) + UPDATER_NORMAL = 1, // execute in Owner thread(tid:1) + NUMBER_UPDATERS = 2, + }; + /** + * @brief create one dense updater and one sparse updater + * + * @note use syncThreadPool to synchronize these two updaters + */ + SparseRemoteParameterUpdaterComposite( + const OptimizationConfig& config, + int expectedPassCount, + bool testing, + std::unique_ptr&& normalUpdater) { + updaters_.resize(NUMBER_UPDATERS); + updaters_[UPDATER_SPARSE_REMOTE].reset( + new SparseRemoteParameterUpdater(config, expectedPassCount, testing)); + updaters_[UPDATER_NORMAL] = std::move(normalUpdater); + + syncThreadPool_.reset(new SyncThreadPool(NUMBER_UPDATERS - 1)); + } + + /// initialization of dense and sparse updaters + virtual void init(const std::vector& parameters); +}; + +class ParameterUpdaterCreators { + public: + /** + * @brief add a creator to create custom ParameterUpdater while training. + * The creator is a function with type (alogrithm, optConfig, isLocal, + * numPasses) -> ParameterUpdater*. Trainer will use this + * ParameterUpdater if creator can create a no nullptr + * ParameterUpdater. Return nullptr will use trainer's default + * updaters. + * + * @param creator method which can create ParameterUpdater. + */ + static void addCreator( + const std::function& creator) { // NOLINT explicit move closing ) in this line + // for readability + constructors_.push_back(creator); + } + + /** + * @brief Try to create an updater by given algo, optConfig, isLocal, + * numPasses. Return nullptr if cannot create anyone. + * @param algo algorithm string. + * @param optConfig optimization config. + * @param isLocal is in local mode or not. + * @param numPasses total passes that trainer will train. + * @return nullptr if fail, not nullptr if we can create an updater. + */ + static ParameterUpdater* tryCreateUpdater(const std::string& algo, + const OptimizationConfig& optConfig, + bool isLocal, + size_t numPasses) { + for (auto& c : constructors_) { + if (auto updater = c(algo, optConfig, isLocal, numPasses)) { + return updater; + } + } + return nullptr; + } + + private: + static std::vector> + constructors_; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/Tester.cpp b/paddle/legacy/trainer/Tester.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d977ca9657a7688c101ed060935c644e4876e6d1 --- /dev/null +++ b/paddle/legacy/trainer/Tester.cpp @@ -0,0 +1,380 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Tester.h" + +#include +#include + +#include +#include +#include +#include + +#include + +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +#include "TesterConfig.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachineMode.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/gserver/layers/ValidationLayer.h" + +namespace paddle { + +Tester::Tester(const std::shared_ptr& config, + std::unique_ptr&& intconfig, + const GradientMachinePtr& gradientMachine, + const std::shared_ptr& parameterUpdater, + std::shared_ptr testDataProvider) + : config_(config), + intconfig_(std::move(intconfig)), + gradientMachine_(gradientMachine), + parameterUpdater_(parameterUpdater), + testDataProvider_(testDataProvider) { + if (config_->getOptConfig().use_sparse_remote_updater()) { + LOG(FATAL) << "It's prohibited to set sparse_remote_update " + << "when doing train and test jobs in the same " + << "process. You could run paddle --job=test in " + << "a separate process."; + } + testEvaluator_.reset(gradientMachine_->makeEvaluator()); + if (intconfig_->distributeTest) { + testParameterClient_.reset(new ParameterClient2(true)); + } + + if (testParameterClient_) { + testParameterClient_->init(gradientMachine_->getParameters()); + } + + std::unique_ptr paramConfig( + new ParameterUtilConfig(intconfig_->saveOnlyOne, + intconfig_->savingPeriod, + intconfig_->loadsaveParametersInPserver, + intconfig_->config)); + + paramUtil_.reset(new ParameterUtil( + config_, std::move(paramConfig), gradientMachine_, parameterUpdater_)); +} + +void Tester::startTestPeriod() { + if (testDataProvider_) { + testDataProvider_->reset(); + } + testEvaluator_->start(); + testContext_.cost = 0; + testContext_.numSamples = 0; + + parameterUpdater_->apply(); + if (intconfig_->prevBatchState) { + gradientMachine_->getState(*intconfig_->trainState); + gradientMachine_->setState(*intconfig_->testState); + } +} + +void Tester::testOneDataBatch(const DataBatch& dataBatch, + std::vector* outArgs) { + testContext_.cost += + forwardOneBatch(dataBatch, testEvaluator_.get(), outArgs); + testContext_.numSamples += dataBatch.getSize(); +} + +void Tester::testOnePeriod() { + DataBatch dataBatch; + int64_t batchSize = config_->getOptConfig().batch_size(); + std::vector outArgs; + startTestPeriod(); + while (testDataProvider_->getNextBatch(batchSize, &dataBatch) != 0) { + testOneDataBatch(dataBatch, &outArgs); + } + finishTestPeriod(); +} + +void Tester::finishTestPeriod() { + if (intconfig_->prevBatchState) { + gradientMachine_->resetState(); + } + testEvaluator_->finish(); + CHECK_GT(testContext_.numSamples, 0) + << "There is no samples in your test batch. Possibly " + "wrong implementation of DataProvidor.reset()"; + LOG(INFO) << " Test samples=" << testContext_.numSamples + << " cost=" << testContext_.cost / testContext_.numSamples + << " Eval: " << *testEvaluator_; + parameterUpdater_->restore(); + if (intconfig_->prevBatchState) { + gradientMachine_->getState(*intconfig_->testState); + gradientMachine_->setState(*intconfig_->trainState); + } +} + +int64_t Tester::testOneBatchById(int64_t batchId) { + DataBatch dataBatch; + int32_t batchSize = config_->getOptConfig().batch_size(); + + testDataProvider_->getNextBatch(batchSize, &dataBatch); + + int64_t actualBatchSize = dataBatch.getSize(); + if (actualBatchSize == 0) { + return 0; + } + + std::vector outArgs; + + stats_ += std::pair{ + actualBatchSize, + forwardOneBatch(dataBatch, testEvaluator_.get(), &outArgs)}; + + if (((batchId + 1) % intconfig_->logPeriod) == 0) { + LOG(INFO) << " Batch=" << batchId + 1 << " " << stats_.getStats(false); + } + + return actualBatchSize; +} + +real Tester::forwardOneBatch(const DataBatch& dataBatch, + Evaluator* evaluator, + std::vector* pOutArgs) { + auto& outArgs = *pOutArgs; + const std::vector& inArgs = dataBatch.getStreams(); + if (intconfig_->loadsaveParametersInPserver) { + REGISTER_TIMER("prefetch"); + gradientMachine_->prefetch(inArgs); + parameterUpdater_->getParametersRemote(false /*full parameter*/, + true /*after apply*/); + } + + gradientMachine_->forward(inArgs, &outArgs, PASS_TEST); + + // write features if set this flag and outArgs is not empty + std::string featFile = intconfig_->featFile; + if (!featFile.empty() && outArgs.empty()) { + size_t numOutputs = outArgs.size(); + std::vector featMatrices; + featMatrices.resize(numOutputs); + for (size_t i = 0; i < numOutputs; ++i) { + featMatrices[i] = Matrix::create(outArgs[i].value->getHeight(), + outArgs[i].value->getWidth(), + false, + false); // CPU data buffer + featMatrices[i]->copyFrom(*(outArgs[i].value), HPPL_STREAM_DEFAULT); + } + hl_stream_synchronize(HPPL_STREAM_DEFAULT); + FILE* fp = fopen(featFile.c_str(), "ab+"); + CHECK(!ferror(fp)) << "Fail to open " << featFile; + + size_t sampleNum = featMatrices[0]->getHeight(); + for (size_t i = 0; i < sampleNum; ++i) { + for (size_t j = 0; j < numOutputs; ++j) { + size_t dim = featMatrices[j]->getWidth(); + fwrite(featMatrices[j]->getData() + i * dim, sizeof(real), dim, fp); + } + } + fclose(fp); + } + if (evaluator) { + gradientMachine_->eval(evaluator); + } + + // Save the output layers if predict_output_dir is not empty + std::string predictOutputDir = intconfig_->predictOutputDir; + if (!predictOutputDir.empty() && !outArgs.empty()) { + CHECK(intconfig_->testing) << "Only valid in test mode"; + if (!os_.is_open()) { + // TODO(yuyang18): Refactor these lines. + constexpr int kBufLen = 100; + char buf[kBufLen]; + snprintf(buf, kBufLen, "rank-%05d", intconfig_->trainerId); + mkDir(predictOutputDir.c_str()); + std::string filename = path::join(predictOutputDir, buf); + os_.open(filename, std::ofstream::trunc); + CHECK(os_.is_open()) << "Failed to open file " << filename; + } + printOutput(outArgs, os_); + return 0.0; // In this case, there is no meaning to calculate cost + } + + return Argument::sum(outArgs); +} + +void Tester::testOnePassBatch(int passId) { + stats_.reset(); + const std::vector inArgs; + gradientMachine_->forward(inArgs, nullptr, PASS_TEST); + int64_t num; + real cost; + gradientMachine_->getStats(cost, num); + stats_ += std::pair{num, cost}; + gradientMachine_->onPassEnd(); + + LOG(INFO) << " Pass=" << passId << " " << stats_.getStats(false); +} + +void Tester::testOnePass(int passId) { + stats_.reset(); + int64_t batchId = 0; + int num = 0; + if (intconfig_->prevBatchState) { + gradientMachine_->resetState(); + } + + testEvaluator_->start(); + + do { + num = testOneBatchById(batchId); + ++batchId; + } while (num > 0); + + gradientMachine_->onPassEnd(); + testEvaluator_->finish(); + + LOG(INFO) << " Pass=" << passId << " " << stats_.getStats(false) + << " Eval: " << *testEvaluator_; + + if (intconfig_->distributeTest) { + testEvaluator_->distributeEval(testParameterClient_.get()); + if (0 == intconfig_->trainerId) { + LOG(INFO) << "distribute eval: " << *testEvaluator_; + } + } +} + +void Tester::test() { + CHECK(testDataProvider_) << "TestData is not specified"; + testDataProvider_->setSkipShuffle(); + testDataProvider_->reset(); + gradientMachine_->start(); + + // For evaluation + std::vector modelList; + std::string modelListFromConfig = intconfig_->modelList; + std::string initModelPath = intconfig_->initModelPath; + if (!modelListFromConfig.empty()) { + loadFileList(modelListFromConfig, modelList); + intconfig_->testPass = 0; + intconfig_->numPasses = modelList.size(); + intconfig_->savingPeriod = 1; + CHECK_EQ(intconfig_->testWait, 0) << "--test_wait must be 0 for evaluation"; + } else if (!initModelPath.empty()) { + modelList.push_back(initModelPath); + intconfig_->testPass = 0; + intconfig_->numPasses = 1; + intconfig_->savingPeriod = 1; + CHECK_EQ(intconfig_->testWait, 0) << "--test_wait must be 0 for evaluation"; + } + + for (int i = intconfig_->testPass; i < intconfig_->numPasses; ++i) { + int passId = i; + if (passId % intconfig_->savingPeriod == 0) { + if (intconfig_->testWait) { + while (paramUtil_->loadParameters( + passId, true /*local*/, true /*remote*/) == false) { + LOG(INFO) << "Waiting for parameters of pass " << passId; + sleep(60); // sleep 60s + } + } else { + if (modelList.size() == 0) { + CHECK_EQ(paramUtil_->loadParameters( + passId, true /*local*/, true /*remote*/), + true); + } else { + paramUtil_->loadParametersWithPath( + modelList[i], true /*local*/, true /*remote*/); + } + } + if (IGradientMachineMode::trainWholeDataInOneBatch(intconfig_->mode)) { + testOnePassBatch(passId); + } else { + testOnePass(passId); + } + if (passId + intconfig_->savingPeriod < intconfig_->numPasses) { + // if there is at least 1 more pass to test, then call reset, + // otherwise not. + testDataProvider_->reset(); + } + } + } + + gradientMachine_->finish(); +} + +void Tester::printOutput(const std::vector& outArgs, + std::ostream& os) { + size_t numOutputs = outArgs.size(); + size_t numIns = outArgs[0].getBatchSize(); + if (cpuMat_.size() != numOutputs || cpuVec_.size() != numOutputs) { + cpuMat_.resize(numOutputs, nullptr); + cpuVec_.resize(numOutputs, nullptr); + } + + for (size_t i = 0; i < numOutputs; ++i) { + if (outArgs[i].value != nullptr) { + if (outArgs[i].value->useGpu()) { + if (dynamic_cast(outArgs[i].value.get())) { + size_t dim = outArgs[i].value->getWidth(); + Matrix::resizeOrCreate(cpuMat_[i], numIns, dim, false, false); + cpuMat_[i]->copyFrom(*outArgs[i].value); + } else if (dynamic_cast(outArgs[i].value.get())) { + auto sparseMat = + dynamic_cast(outArgs[i].value.get()); + cpuMat_[i] = Matrix::createSparseMatrix(sparseMat->getHeight(), + sparseMat->getWidth(), + sparseMat->getElementCnt(), + sparseMat->getValueType(), + sparseMat->format_, + false, /* trans */ + false); /* useGpu */ + hl_stream_t stream = HPPL_STREAM_DEFAULT; + cpuMat_[i]->copyFrom(*sparseMat, stream); + } else { + LOG(WARNING) << "Not supported gpu matrix type"; + } + } + } else if (outArgs[i].ids != nullptr) { + if (outArgs[i].ids->useGpu()) { + IVector::resizeOrCreate(cpuVec_[i], outArgs[i].ids->getSize(), false); + cpuVec_[i]->copyFrom(*outArgs[i].ids); + } + } else if (outArgs[i].strs != nullptr) { + continue; + } else { + LOG(WARNING) << "outArgs[" << i << "] has no data to print"; + } + } + + for (size_t i = 0; i < numIns; ++i) { + for (size_t j = 0; j < numOutputs; ++j) { + if (outArgs[j].value != nullptr) { + if (outArgs[j].value->useGpu()) { + cpuMat_[j]->printOneRow(os, i); + } else { + outArgs[j].value->printOneRow(os, i); + } + } else if (outArgs[j].ids != nullptr) { + if (outArgs[j].ids->useGpu()) { + cpuVec_[j]->printOneElement(os, i); + } else { + outArgs[j].ids->printOneElement(os, i); + } + } else if (outArgs[j].strs != nullptr) { + os << (*outArgs[j].strs)[i] << ";"; + } + } + os << std::endl; + } +} +} // namespace paddle diff --git a/paddle/legacy/trainer/Tester.h b/paddle/legacy/trainer/Tester.h new file mode 100644 index 0000000000000000000000000000000000000000..a298602d1d0894af90c098818908862a553cb3e7 --- /dev/null +++ b/paddle/legacy/trainer/Tester.h @@ -0,0 +1,149 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include + +#include "hl_gpu.h" +#include "paddle/legacy/gserver/dataproviders/DataProvider.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" + +#include "TrainerConfig.pb.h" + +#include +#include +#include "ParamUtil.h" +#include "ParameterUpdater.h" +#include "TesterConfig.h" +#include "TrainerInternalConfig.h" + +namespace paddle { + +/** + * Neural Network test logics code. + * It is a private class for Trainer. + */ +class Tester { + public: + /** + * Ctor + * @param config Trainer Config. + * @param intconfig Tester Config. + * @param gradientMachine Gradient machine(neuralnetwork) that will be tested. + * @param parameterUpdater Parameter Updater. Not for updating parameter, just + * for getting parameter from parameter-server. + * @param testDataProvider Test data provider. + */ + Tester(const std::shared_ptr& config, + std::unique_ptr&& intconfig, + const GradientMachinePtr& gradientMachine, + const std::shared_ptr& parameterUpdater, + std::shared_ptr testDataProvider); + + /** + * test one period. + * + * One period means 2 things. + * if test_period !=0 and not test_all_data_in_one_period, then + * will test test_period * batch_size data. + * else + * will test whole test data. + * + * It is convenience to test small set of data when test data set is large and + * is training at same time. + */ + void testOnePeriod(); + void startTestPeriod(); + void finishTestPeriod(); + void testOneDataBatch(const DataBatch& dataBatch, + std::vector* outArgs); + + /** + * Test for given data batch. + * @param dataBatch Data batch. + * @param evaluator Evaluator + * @return cost + */ + real forwardOneBatch(const DataBatch& dataBatch, + Evaluator* evaluator, + std::vector* outArgs); + + /** + * performance the full pass of test given test data provider + */ + void test(); + + protected: + std::shared_ptr testParameterClient_; + std::shared_ptr config_; + std::unique_ptr intconfig_; + GradientMachinePtr gradientMachine_; + std::shared_ptr parameterUpdater_; + std::unique_ptr testEvaluator_; + std::unique_ptr paramUtil_; + DataProviderPtr testDataProvider_; + TrainerStats stats_; + + // Used for saving the values of output layers + std::ofstream os_; + std::vector cpuMat_; + std::vector cpuVec_; + struct { + int64_t numSamples; + real cost; + } testContext_; + + private: + /** + * Test one batch by batchId. It is only used for testOnePass. + * + * Durning testOnePass, each log_period will print cost statistics. + * + * @param batchId current batch id (from 0) + * @return num of tested samples. Zero if end of pass. + */ + int64_t testOneBatchById(int64_t batchId); + + /** + * Test whole pass in one batch. + * + * + * @param passId current pass id (from 0) + */ + void testOnePassBatch(int passId); + + /** + * test for one pass in several mini-batches. + * + * Used for sgd method. + * + * @param passId current pass id (from 0) + */ + void testOnePass(int passId); + + /** + * print the outArgs to a stream + * + * used for save feature file + * + * @param [in] outArgs output arguments for network. + * @param [in,out] os output stream. + */ + void printOutput(const std::vector& outArgs, std::ostream& os); +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/TesterConfig.h b/paddle/legacy/trainer/TesterConfig.h new file mode 100644 index 0000000000000000000000000000000000000000..6c78f7cda347d5808d11e8af98672ef56898d643 --- /dev/null +++ b/paddle/legacy/trainer/TesterConfig.h @@ -0,0 +1,138 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include + +#include "hl_gpu.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" + +#include "TrainerConfig.pb.h" + +#include +#include +#include "ParameterUpdater.h" + +namespace paddle { + +/** + * TesterConfig + * general configs for training + */ +struct TesterConfig { + /** + * indicate test period + */ + int testPeriod; + + /** + * indicate whether to save previous batch state + */ + bool prevBatchState; + + /** + * log period + */ + int logPeriod; + + /** + * loadsave parameters in pserver + */ + bool loadsaveParametersInPserver; + + /** + * feat file + */ + std::string featFile; + + /** + * predict output dir + */ + std::string predictOutputDir; + + /** + * trianer id + */ + int trainerId; + + /** + * distribute test + */ + bool distributeTest; + + /** + * training state + */ + MachineState* trainState; + + /** + * test state + */ + MachineState* testState; + + /** + * model list + */ + std::string modelList; + + /** + * test passes + */ + int testPass; + + /** + * num passes + */ + int numPasses; + + /** + * saving period + */ + int savingPeriod; + + /** + * test wait + */ + int testWait; + + /** + * init model path + */ + std::string initModelPath; + + /** + * save only one + */ + bool saveOnlyOne; + + /** + * testing mode + */ + bool testing; + + /** + * mode + */ + int mode; + + /** + * config loc + */ + std::string config; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/ThreadParameterUpdater.cpp b/paddle/legacy/trainer/ThreadParameterUpdater.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0601bdf24e3150f5d182e2addde3a91609a967e4 --- /dev/null +++ b/paddle/legacy/trainer/ThreadParameterUpdater.cpp @@ -0,0 +1,309 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "ThreadParameterUpdater.h" + +#include "paddle/legacy/utils/Logging.h" + +#include "paddle/legacy/math/SparseRowMatrix.h" +#include "paddle/legacy/parameter/ThreadLocalBuffer.h" +#include "paddle/legacy/utils/Thread.h" + +DECLARE_int32(trainer_count); + +namespace paddle { + +SgdThreadUpdater::SgdThreadUpdater(const OptimizationConfig& optConfig) + : config_(optConfig), numSamplesProcessed_(0) { + // fill types + auto types = sgdOptimizerGetTypes(optConfig, false /*inPserver*/); + for (auto type : types) { + addParameterType(type); + } +} + +void SgdThreadUpdater::init(const std::vector& parameters) { + ParameterUpdater::init(parameters); + + // calc max parameter id + size_t maxId = 0; + for (auto& para : parameters_) { + maxId = std::max(maxId, para->getID()); + } + + optimizers_.resize(maxId + 1); + for (auto& para : parameters_) { + int pid = para->getID(); + optimizers_[pid].reset(sgdOptimizerCreate(config_, + para->getConfig(), + para->isGradSparseUpdate(), + false /*inPserver*/)); + size_t numRows = para->isGradSparseUpdate() ? para->getConfig().dims(0) : 0; + optimizers_[pid]->init(numRows, ¶->getConfig()); + if (para->isGradSparseUpdate() && FLAGS_trainer_count == 1) { + // For trainer_count=1, the gradient machine is NeuralNetwork, which does + // not create parameter buf for PARAMETER_GRADIENT for sparse update in + // Parameter::enableType(). But gradient parameter buf is still used + // in SgdThreadUpdater. We need to explicitly create it. + // + // The AverageOptimizer::restore/apply method will use PARAMETER_GRADIENT + // as a temp buffer. + para->enableBufType(PARAMETER_GRADIENT); + } + } +} + +void SgdThreadUpdater::startPass() { + for (auto& para : parameters_) { + int pid = para->getID(); + optimizers_[pid]->startPass(); + } +} + +bool SgdThreadUpdater::finishPass() { + catchUpWith(); + + for (auto& para : parameters_) { + int pid = para->getID(); + optimizers_[pid]->finishPass(); + } + return true; +} + +void SgdThreadUpdater::updateImpl(Parameter* para) { + if (!para->useGpu()) return; + SetDevice setDevice(para->getDeviceId()); + ParameterOptimizer* optimizer = optimizers_[para->getID()].get(); + optimizer->update(para->getBufs(), para->getConfig()); + if (auto callback = optimizer->needSpecialTraversal(para->getConfig())) { + callback(para->getBufs(), para->getConfig(), -1LU); + } + + para->setValueUpdated(); + para->clearGradient(); +} + +void SgdThreadUpdater::threadTraverse( + const ParameterOptimizer::TraverseCallback& callback, + int tid, + size_t numThreads, + Parameter* para) { + VectorPtr* vecs = parameter::getThreadLocalBuffer(); + if (para->isGradSparseUpdate()) { + size_t height = para->getConfig().dims(0); + size_t width = para->getConfig().dims(1); + for (size_t i = tid; i < height; i += numThreads) { + // setup sub bufs + for (auto type : parameterTypes_) { + vecs[type]->subVecFrom(*para->getBuf(type), i * width, width); + } + callback(vecs, para->getConfig(), i); + } + } else { // dense + // setup sub bufs + auto interval = calcSplitArrayInterval( + para->getSize(), (size_t)tid, numThreads, 8LU /*for avx*/); + for (auto type : parameterTypes_) { + vecs[type]->subVecFrom(*para->getBuf(type), interval); + } + + callback(vecs, para->getConfig(), -1LU); + } +} + +void SgdThreadUpdater::traverse(GetTraverseCallback getTraverseCallback) { + bool hasCpuPara = false; + bool hasGpuPara = false; + for (auto& para : parameters_) { + if (para->useGpu()) { + hasGpuPara = true; + } else { + hasCpuPara = true; + } + } + + auto cpuTraverse = [&](int tid, size_t numThreads) { + for (auto& para : parameters_) { + if (auto callback = getTraverseCallback(para.get())) { + threadTraverse(callback, tid, numThreads, para.get()); + } + } + }; + auto gpuTraverse = [&](int tid, size_t numThreads) { + for (auto& para : parameters_) { + if (para->useGpu()) { + if (auto callback = getTraverseCallback(para.get())) { + SetDevice setDevice(para->getDeviceId()); + callback(para->getBufs(), para->getConfig(), -1LU); + } + } + } + }; + + if (hasCpuPara && hasGpuPara) { + getGlobalSyncThreadPool()->exec(cpuTraverse, gpuTraverse); + } else if (hasCpuPara) { + getGlobalSyncThreadPool()->exec(cpuTraverse); + } else if (hasGpuPara) { + gpuTraverse(0, 0); + } +} + +void SgdThreadUpdater::catchUpWith() { + traverse([this](Parameter* para) { + return optimizers_[para->getID()]->startCatchUpWith(); + }); + + for (auto& para : parameters_) { + int pid = para->getID(); + optimizers_[pid]->finishCatchUpWith(); + } +} + +void SgdThreadUpdater::apply() { + catchUpWith(); + + traverse( + [this](Parameter* para) { return optimizers_[para->getID()]->apply(); }); +} + +void SgdThreadUpdater::restore() { + traverse([this](Parameter* para) { + return optimizers_[para->getID()]->restore(); + }); +} + +PassType SgdThreadUpdater::startBatch(int64_t batchSize) { + numSamplesProcessed_ += batchSize; + for (auto& para : parameters_) { + int pid = para->getID(); + optimizers_[pid]->startBatch(numSamplesProcessed_); + } + return PASS_TRAIN; +} + +void SgdThreadUpdater::finishBatch(real cost) { + getGlobalSyncThreadPool()->exec([&](int tid, size_t numThreads) { + for (auto& para : parameters_) { + if (para->isGradSparseUpdate()) { + threadUpdateSparse(tid, numThreads, para.get()); + } else if (!para->useGpu()) { + threadUpdateDense(tid, numThreads, para.get()); + } + } + }); + + for (auto& para : parameters_) { + int pid = para->getID(); + optimizers_[pid]->finishBatch(); + } +} + +void SgdThreadUpdater::threadUpdateSparse(int tid, + size_t numThreads, + Parameter* para) { + int pid = para->getID(); + ParameterOptimizer* optimizer = optimizers_[pid].get(); + VectorPtr* vecs = parameter::getThreadLocalBuffer(); + + size_t height = para->getConfig().dims(0); + size_t width = para->getConfig().dims(1); + + if (dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get())) { + // From MultiGradientMachine + SparseRowIdsCpuMatrix* mainMat = dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get()); + std::vector& sparseIds = mainMat->getIds(tid); + + for (auto id : sparseIds) { + // setup sub bufs + for (auto type : parameterTypes_) { + vecs[type]->subVecFrom(*para->getBuf(type), id * width, width); + } + optimizer->update(vecs, para->getConfig(), id); + vecs[PARAMETER_GRADIENT]->zeroMem(); + } + sparseIds.clear(); + } else if (dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get())) { + // From NeuralNetwork + SparseRowCpuMatrix* mainMat = dynamic_cast( + para->getMat(PARAMETER_GRADIENT).get()); + + std::vector& localIndices = + mainMat->getIndexDictHandle()->localIndices; + + auto interval = + calcSplitArrayInterval(localIndices.size(), tid, numThreads); + for (size_t i = interval.first; i < interval.second; ++i) { + auto id = localIndices[i]; + real* row = mainMat->getLocalRow(i); + // setup sub bufs + for (auto type : parameterTypes_) { + if (type == PARAMETER_GRADIENT) { + vecs[type]->subVecFrom(row, 0, width); + } else { + vecs[type]->subVecFrom(*para->getBuf(type), id * width, width); + } + } + optimizer->update(vecs, para->getConfig(), id); + vecs[PARAMETER_GRADIENT]->zeroMem(); + } + // For numThreads > 1, MultiGradientMachine is used, which goes + // to the above branch. + CHECK_EQ(numThreads, 1UL); + mainMat->clearIndices(); + } else { + auto& m = *para->getMat(PARAMETER_GRADIENT).get(); + LOG(FATAL) << "Internal error: " << para->getName() << " " + << typeid(m).name(); + } + + if (auto callback = optimizer->needSpecialTraversal(para->getConfig())) { + for (size_t i = tid; i < height; i += numThreads) { + // setup sub bufs + for (auto type : parameterTypes_) { + vecs[type]->subVecFrom(*para->getBuf(type), i * width, width); + } + callback(vecs, para->getConfig(), i); + } + } +} + +void SgdThreadUpdater::threadUpdateDense(int tid, + size_t numThreads, + Parameter* para) { + int pid = para->getID(); + ParameterOptimizer* optimizer = optimizers_[pid].get(); + VectorPtr* vecs = parameter::getThreadLocalBuffer(); + + auto interval = calcSplitArrayInterval( + para->getSize(), (size_t)tid, numThreads, 8LU /*for avx*/); + + // setup sub bufs + for (auto type : parameterTypes_) { + vecs[type]->subVecFrom(*para->getBuf(type), interval); + } + + // update + optimizer->update(vecs, para->getConfig()); + vecs[PARAMETER_GRADIENT]->zeroMem(); + + if (auto callback = optimizer->needSpecialTraversal(para->getConfig())) { + callback(vecs, para->getConfig(), -1LU); + } +} + +} // namespace paddle diff --git a/paddle/legacy/trainer/ThreadParameterUpdater.h b/paddle/legacy/trainer/ThreadParameterUpdater.h new file mode 100644 index 0000000000000000000000000000000000000000..172287d4eb56828c83e6670226b4c1f179fac6d8 --- /dev/null +++ b/paddle/legacy/trainer/ThreadParameterUpdater.h @@ -0,0 +1,85 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/parameter/AverageOptimizer.h" +#include "paddle/legacy/parameter/FirstOrderOptimizer.h" +#include "paddle/legacy/parameter/OptimizerFunctions.h" +#include "paddle/legacy/parameter/OptimizerWithRegularizer.h" +#include "paddle/legacy/parameter/Parameter.h" +#include "paddle/legacy/parameter/Regularizer.h" +#include "paddle/legacy/utils/Util.h" + +#include +#include + +namespace paddle { + +/** + * \brief A parameter updater that uses multiple threads to update parameters. + This parameter updater handles GPU and CPU updates differently, + because at the current moment, the merging on CPU is happening on the + main thread, and the its parameter size can be much larger than the one GPU. + Thus, for GPU, the parameter updates happens in updateImpl() function, which + is called by gradient machines as a callback function supplied to backward() + and forwardBackward(). + For CPU, the parameter updates happens in separate threads maintained by this + class. + */ +class SgdThreadUpdater : public ParameterUpdater { + public: + explicit SgdThreadUpdater(const OptimizationConfig& optConfig); + virtual ~SgdThreadUpdater() {} + + // Use the startPass() function of the base optimizer. + virtual void startPass(); + + // Use the finishPass() function of the base optimizer. + virtual bool finishPass(); + + virtual void init(const std::vector& parameters); + virtual PassType startBatch(int64_t batchSize); + // Call finishBatch for each optimizer. + virtual void finishBatch(real cost); + virtual void catchUpWith(); + virtual void apply(); + virtual void restore(); + + protected: + // This is the function that will be eventualy called by the GradientMachine. + // used only for GPU update. + virtual void updateImpl(Parameter* para); + OptimizationConfig config_; + int64_t numSamplesProcessed_; + + // One optimizers for each parameter. + std::vector> optimizers_; + + // The update function for CPU sparse parameters. + void threadUpdateSparse(int tid, size_t numThreads, Parameter* para); + + // The update function for CPU dense parameters. + void threadUpdateDense(int tid, size_t numThreads, Parameter* para); + // The update function for after update operations, such as averager. + void threadTraverse(const ParameterOptimizer::TraverseCallback& callback, + int tid, + size_t numThreads, + Parameter* para); + typedef std::function + GetTraverseCallback; + void traverse(GetTraverseCallback getTraverseCallback); +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/Trainer.cpp b/paddle/legacy/trainer/Trainer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2db754793cf19e0c29455f61ada5f1d15b3204af --- /dev/null +++ b/paddle/legacy/trainer/Trainer.cpp @@ -0,0 +1,653 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Trainer.h" + +#include + +#include +#include +#include +#include + +#include + +#include "paddle/legacy/utils/Common.h" +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +#include "RemoteParameterUpdater.h" +#include "TesterConfig.h" +#include "ThreadParameterUpdater.h" +#include "TrainerConfigHelper.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachineMode.h" +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/gserver/layers/ValidationLayer.h" + +DEFINE_string(config, "", "Trainer config file"); + +DEFINE_int32(test_period, + 0, + "if equal 0, do test on all test data at the end of " + "each pass. While if equal non-zero, do test on all test " + "data every test_period batches"); +DEFINE_bool(test_all_data_in_one_period, + false, + "This option was deprecated, since we will always do " + "test on all test set "); + +DEFINE_bool(local, true, "Train in local mode or not"); + +DEFINE_int32(average_test_period, + 0, + "Do test on average parameter every so" + " many batches. MUST be devided by FLAGS_log_period." + " Default 0 means do not test average parameter"); + +DEFINE_int32(saving_period, 1, "Save parameteres every so many passes"); +DEFINE_int64(saving_period_by_batches, + 0, + "Save parameters every so many batches in one pass"); +DEFINE_string(save_dir, "", "Directory for saving model parameter"); +DEFINE_int32(start_pass, + 0, + "Start training from this pass. " + "Will load parameter from the previous pass"); +DEFINE_int32(test_pass, -1, "Will load parameter start from this pass to test"); +DEFINE_int32(test_wait, 0, "Waiting for pass parameter if not exist"); +DEFINE_bool(with_cost, true, "enable cost layer or not"); +DEFINE_bool(distribute_test, false, "test in distribute mode"); + +DEFINE_int32(num_passes, 100, "train for so many passes"); + +DEFINE_string(config_args, + "", + "arguments passed to config file." + "Format: key1=value1,key2=value2"); + +DEFINE_bool(save_only_one, + false, + "Save only parameters in last pass, remove previous."); + +DEFINE_string(feat_file, "", "File name of extracted feature."); +DEFINE_string(predict_output_dir, + "", + "Directory that saves the predicted results of output layers"); +DEFINE_string(model_list, "", "File that saves the model list when evaluation"); + +namespace paddle { + +void Trainer::init(const std::shared_ptr& config, + bool testing, + const std::shared_ptr& gradientMachine, + const std::shared_ptr& dataProvider, + const std::shared_ptr& testDataProvider) { + this->stats_ = std::make_shared(); + + config_ = config; + + config_->updateConfigFromFlags(); + + testing_ = testing; + + // in testing, mode_ may GradientMachine::kTesting or + // GradientMachine::kSgdSparseCpuTraining + + if (FLAGS_local) { + CHECK(!FLAGS_loadsave_parameters_in_pserver) + << "local and loadsave_parameters_in_pserver can not both true"; + if (config_->getOptConfig().use_sparse_remote_updater()) { + config_->disableRemoteSparseUpdaterForEachParams(); + LOG(INFO) << "ignore sparse_remote_update=true due to --local=true"; + } + } + if (FLAGS_loadsave_parameters_in_pserver) { + CHECK(config_->getOptConfig().use_sparse_remote_updater()) + << "no parameter to load from pserver, please check network config"; + } + if (testing && !FLAGS_loadsave_parameters_in_pserver) { + if (config_->getOptConfig().use_sparse_remote_updater()) { + config_->disableRemoteSparseUpdater(); + LOG(INFO) << "because parameter is loaded local," + << "tester ignore sparse_remote_update flag"; + } + } + + CHECK(TrainAlgorithm::isValid(config_->getOptConfig().algorithm())) + << "invalid algorithm configuration: " + << config_->getOptConfig().algorithm(); + + bool useSparseUpdater = false; + for (auto& paraConfig : config_->getModelConfig().parameters()) { + if (paraConfig.sparse_update() || paraConfig.sparse_remote_update()) { + useSparseUpdater = true; + } + } + + if (FLAGS_use_mkldnn) { + CHECK_EQ(FLAGS_trainer_count, 1) << "MKLDNN only need 1 trainer"; + } + + if (testing) { + LOG(INFO) << "trainer: in testing mode"; + if (config_->getOptConfig().use_sparse_remote_updater() || + FLAGS_trainer_count > 1) { + mode_ = GradientMachine::kSgdSparseCpuTraining; + LOG(INFO) << "trainer mode: SgdSparseCpuTraining"; + } else { + mode_ = GradientMachine::kTesting; + LOG(INFO) << "trainer mode: Testing"; + } + } else if (IGradientMachineMode::tryGetMode( + (int*)&mode_, + config_->getOptConfig().algorithm(), + FLAGS_trainer_count, + FLAGS_local, + FLAGS_use_gpu)) { + LOG(INFO) << "Custom trainer mode."; + } else if ((config_->getOptConfig().algorithm() == TrainAlgorithm::SGD || + config_->getOptConfig().algorithm() == + TrainAlgorithm::AsyncSGD) && + useSparseUpdater) { + mode_ = GradientMachine::kSgdSparseCpuTraining; + LOG(INFO) << "trainer mode: SgdSparseCpuTraining"; + } else { + mode_ = GradientMachine::kNormal; + LOG(INFO) << "trainer mode: Normal"; + } + + // initialize trainer internal + trainerInternal_.init(config_, + gradientMachine, + TrainerInternalConfig::createFromMode(mode_), + stats_, + testing); + std::unique_ptr paramConfig( + new ParameterUtilConfig(FLAGS_save_only_one, + FLAGS_saving_period, + FLAGS_loadsave_parameters_in_pserver, + FLAGS_config)); + + paramUtil_.reset( + new paddle::ParameterUtil(config_, + std::move(paramConfig), + trainerInternal_.getGradientMachine(), + trainerInternal_.getParameterUpdater())); + + bool gpuData = + FLAGS_use_gpu && (!FLAGS_parallel_nn) && + (!IGradientMachineMode::dataMustInCpu(mode_, FLAGS_trainer_count)); + + dataProvider_ = dataProvider; + if (!dataProvider_ && config_->hasDataConfig() && !testing_) { + dataProvider_.reset(DataProvider::create(*config_, *config_, gpuData)); + } + if (!testDataProvider_) { + // No evaluator_ if there is testDataProvider but no dataProvider. + evaluator_.reset(trainerInternal_.getGradientMachine()->makeEvaluator()); + currentEvaluator_.reset( + trainerInternal_.getGradientMachine()->makeEvaluator()); + if (FLAGS_average_test_period > 0 && FLAGS_trainer_id == 0 && + config_->getOptConfig().average_window() > 0) { + CHECK_EQ(FLAGS_average_test_period % FLAGS_log_period, 0) + << "FLAGS_average_test_period must be divided by FALGS_log_period"; + averageEvaluator_.reset( + trainerInternal_.getGradientMachine()->makeEvaluator()); + } + } + + testDataProvider_ = testDataProvider; + if (!testDataProvider_ && config_->hasTestDataConfig()) { + testDataProvider_.reset( + DataProvider::create(config_->getTestDataConfig(), *config_, gpuData)); + } + if (testDataProvider_) { + createTester(); + } + + if (!testing && + (trainerInternal_.getGradientMachine()->hasStaticParameters())) { + CHECK(!FLAGS_loadsave_parameters_in_pserver) + << "is_static and loadsave_parameters_in_pserver can not both true"; + } + if (testing) { + // will load per pass for tester + } else if (paramUtil_->tryLoadParametersFromConfig()) { + // load from config already. + } else { + trainerInternal_.getGradientMachine()->randParameters(); + } + + // Only non static parameters need to be updated + std::vector& parameters = + trainerInternal_.getGradientMachine()->getNonStaticParameters(); + if (trainerInternal_.getParameterUpdater()) { + trainerInternal_.getParameterUpdater()->init(parameters); + + if (FLAGS_loadsave_parameters_in_pserver && FLAGS_trainer_id == 0) { + if (testing) { + // will load per pass for tester + } else if (!config_->getConfig().init_model_path().empty() && + (FLAGS_local || FLAGS_trainer_id == 0)) { + paramUtil_->loadParametersWithPath( + config_->getConfig().init_model_path(), + false /*local*/, + true /*remote*/); + } else if (config_->getConfig().start_pass() > 0 && + (FLAGS_local || FLAGS_trainer_id == 0)) { + CHECK(paramUtil_->loadParameters(config_->getConfig().start_pass() - 1, + false /*local*/, + true /*remote*/)); + } else { + trainerInternal_.getParameterUpdater()->randParametersRemote(); + } + } + } + + // set current evaluator and evalutor + trainerInternal_.setCurrentEvaluator(currentEvaluator_.get()); + trainerInternal_.setEvaluator(evaluator_.get()); +} + +void Trainer::train(size_t numPasses) { + startTrain(); + for (size_t i = 0; i < numPasses; ++i) { + if (IGradientMachineMode::trainWholeDataInOneBatch(mode_)) { + trainOnePassBatch(config_->getConfig().start_pass() + i); + } else { + trainOnePass(); + } + if (i < numPasses - 1) { + dataProvider_->reset(); + } + } + + finishTrain(); +} + +static double genPerturbation(real* d, real* grad, size_t dim) { + auto& reng = ThreadLocalRandomEngine::get(); + std::uniform_real_distribution dist(-1, 1); + double gradNorm = 0, dNorm = 0; + for (size_t i = 0; i < dim; ++i) { + d[i] = dist(reng); + dNorm += d[i] * d[i]; + gradNorm += grad[i] * grad[i]; + } + if (gradNorm > 0) { + real s = 0.5 * sqrt(gradNorm / dNorm); + for (size_t i = 0; i < dim; ++i) { + d[i] = s * d[i] + grad[i]; + } + } + double delta = 0; + for (size_t i = 0; i < dim; ++i) { + delta += grad[i] * d[i]; + } + return delta; +} + +real Trainer::checkGradient() { + trainerInternal_.getGradientMachine()->start(); + std::vector& parameters = + trainerInternal_.getGradientMachine()->getNonStaticParameters(); + DataBatch dataBatch; + int32_t batchSize = config_->getOptConfig().batch_size(); + + dataProvider_->getNextBatch(batchSize, &dataBatch); + + CHECK(dataBatch.getSize()) << "No data from data provider"; + std::vector& inArgs = dataBatch.getStreams(); + std::vector outArgs; + + trainerInternal_.getGradientMachine()->forward(inArgs, &outArgs, PASS_GC); + real cost = Argument::sum(outArgs); + LOG(INFO) << "original cost=" << cost; + trainerInternal_.getGradientMachine()->backward(); + + real maxDiff = 0; + char fill = ' '; + for (auto& parameter : parameters) { + CpuVector oldPara(parameter->getSize()); + CpuVector newPara(parameter->getSize()); + oldPara.copyFrom(*parameter->getBuf(PARAMETER_VALUE)); + real* newp = newPara.getData(); + real* oldp = oldPara.getData(); + CpuVector cpuGrad(*parameter->getBuf(PARAMETER_GRADIENT)); + real* grad = cpuGrad.getData(); + size_t dim = parameter->getSize(); + std::vector d(dim); + + double delta = genPerturbation(d.data(), grad, dim); + + // use a step such that delta / cost is FLAGS_checkgrad_eps + real step = + (delta != 0) ? cost / delta * FLAGS_checkgrad_eps : FLAGS_checkgrad_eps; + delta *= step; + for (size_t i = 0; i < dim; ++i) { + newp[i] = oldp[i] + step * d[i]; + } + + parameter->getBuf(PARAMETER_VALUE)->copyFrom(newPara); + parameter->setValueUpdated(); + trainerInternal_.getGradientMachine()->forward(inArgs, &outArgs, PASS_GC); + real newCost1 = Argument::sum(outArgs); + + for (size_t i = 0; i < dim; ++i) { + newp[i] = oldp[i] - step * d[i]; + } + + parameter->getBuf(PARAMETER_VALUE)->copyFrom(newPara); + parameter->setValueUpdated(); + trainerInternal_.getGradientMachine()->forward(inArgs, &outArgs, PASS_GC); + real newCost2 = Argument::sum(outArgs); + + real trueDelta = 0.5 * (newCost1 - newCost2); + real diff = (1e-20 + trueDelta) / (1e-20 + delta) - 1; + LOG(INFO) << std::setiosflags(std::ios::left) << std::setfill(fill) + << std::setw(20) << parameter->getName() + << "step=" << std::setw(15) << step << "cost1=" << std::setw(10) + << newCost1 << "cost2=" << std::setw(10) << newCost2 + << "true_delta=" << std::setw(15) << trueDelta + << "analytic_delta=" << std::setw(15) << delta << "diff=" << diff + << (std::abs(diff) > 0.01 ? " ***" : ""); + + maxDiff = std::max(maxDiff, std::abs(diff)); + + // restore parameter + parameter->getBuf(PARAMETER_VALUE)->copyFrom(oldPara); + parameter->setValueUpdated(); + + fill = (fill == ' ') ? '.' : ' '; + } + return maxDiff; +} + +void Trainer::startTrain() { + trainPassContext_.passId = config_->getConfig().start_pass(); + srand(config_->getConfig().start_pass() + 1); + if (dataProvider_) { + dataProvider_->reset(); + } + + trainerInternal_.getGradientMachine()->start(); +} + +void Trainer::finishTrain() { trainerInternal_.getGradientMachine()->finish(); } + +void Trainer::startTrainPass() { + stats_->reset(); + trainPassContext_.batchId = 0; + trainPassContext_.avgTestCost = 0; + trainPassContext_.numAvgTests = 0; + trainPassContext_.passInnerId = 1; + + trainerInternal_.getParameterUpdater()->startPass(); + evaluator_->start(); + if (FLAGS_prev_batch_state) { + trainerInternal_.getGradientMachine()->resetState(); + trainerInternal_.getGradientMachine()->getState(testState_); + } +} + +void Trainer::trainOneDataBatch(DataBatch& dataBatch) { + int num = dataBatch.getSize(); + if (averageEvaluator_) { + int64_t mod = trainPassContext_.batchId % FLAGS_average_test_period; + if (mod >= FLAGS_average_test_period - FLAGS_log_period) { + if (mod == FLAGS_average_test_period - FLAGS_log_period) { + averageEvaluator_->start(); + } + trainerInternal_.getParameterUpdater()->apply(); + if (FLAGS_prev_batch_state) { + trainerInternal_.getGradientMachine()->getState(trainState_); + } + trainPassContext_.avgTestCost += tester_->forwardOneBatch( + dataBatch, averageEvaluator_.get(), &forwardOutput_); + if (FLAGS_prev_batch_state) { + trainerInternal_.getGradientMachine()->setState(trainState_); + } + trainPassContext_.numAvgTests += num; + trainerInternal_.getParameterUpdater()->restore(); + } + } + { + REGISTER_TIMER("TrainBatch"); + trainerInternal_.trainOneBatch( + trainPassContext_.batchId, dataBatch, &forwardOutput_); + } + + if (averageEvaluator_ && + trainPassContext_.batchId % FLAGS_average_test_period == + FLAGS_average_test_period - 1) { + averageEvaluator_->finish(); + LOG(INFO) << " Averaged parameter:" + << " cost=" + << trainPassContext_.avgTestCost / trainPassContext_.numAvgTests + << " Eval: " << *averageEvaluator_; + trainPassContext_.numAvgTests = 0; + trainPassContext_.avgTestCost = 0; + } + + ++trainPassContext_.batchId; + + if (trainPassContext_.batchId % FLAGS_log_period == 0) { + FOR_TIMING(globalStat.setThreadInfo(true)); + FOR_TIMING(globalStat.printAllStatus()); + FOR_TIMING(globalStat.reset()); + } + + if (testDataProvider_ && FLAGS_test_period > 0 && + trainPassContext_.batchId % FLAGS_test_period == 0) { + tester_->testOnePeriod(); + } + + if (FLAGS_saving_period_by_batches > 0 && + trainPassContext_.batchId > + FLAGS_saving_period_by_batches * trainPassContext_.passInnerId && + 0 == FLAGS_trainer_id) { + trainerInternal_.getParameterUpdater()->catchUpWith(); + if (testDataProvider_) { + tester_->testOnePeriod(); + } + paramUtil_->saveParametersOnePass(trainPassContext_.passId, + trainPassContext_.passInnerId); + ++trainPassContext_.passInnerId; + } +} + +void Trainer::finishTrainPass() { + if (trainPassContext_.batchId == 0) { + // This means no more data from DataProvider + return; + } + + trainerInternal_.finishTrainPass(trainPassContext_.passId, + trainPassContext_.batchId); + + FOR_TIMING(globalStat.setThreadInfo(true)); + FOR_TIMING(globalStat.printAllStatus()); + FOR_TIMING(globalStat.reset()); + + if (testDataProvider_) { + tester_->testOnePeriod(); + } + + if (trainPassContext_.passId % FLAGS_saving_period == 0 && + FLAGS_trainer_id == 0) { + paramUtil_->saveParametersOnePass(trainPassContext_.passId); + } + ++trainPassContext_.passId; +} + +void Trainer::trainOnePass() { + startTrainPass(); + size_t batchSize = config_->getOptConfig().batch_size(); + while (true) { + DataBatch dataBatch; + + int num = 0; + { + REGISTER_TIMER("getTrainBatch"); + num = dataProvider_->getNextBatch(batchSize, &dataBatch); + } + if (num == 0) break; + CHECK_EQ(num, dataBatch.getSize()); + trainOneDataBatch(dataBatch); + } + + finishTrainPass(); +} + +void Trainer::trainOnePassBatch(int passId) { + this->stats_->reset(); + + trainerInternal_.getParameterUpdater()->startPass(); + const std::vector inArgs; + { + REGISTER_TIMER("onePass"); + trainerInternal_.getGradientMachine()->forwardBackward( + inArgs, nullptr, PASS_TRAIN, nullptr); + } + + real cost = .0; + int64_t num = 0; + trainerInternal_.getGradientMachine()->getStats(cost, num); + *stats_ += {num, cost}; + + trainerInternal_.getGradientMachine()->onPassEnd(); + + bool accepted = trainerInternal_.getParameterUpdater()->finishPass(); + + globalStat.setThreadInfo(true); + globalStat.printAllStatus(); + globalStat.reset(); + + LOG(INFO) << " Pass=" << passId + << " AcceptedPass=" << (accepted ? acceptedPassId_ : -1) + << stats_->getStats(false /*withCurrentCost*/); + + if (accepted) { + if (acceptedPassId_ % FLAGS_saving_period == 0 && FLAGS_trainer_id == 0) { + paramUtil_->saveParameters(acceptedPassId_); + } + acceptedPassId_++; + if (FLAGS_save_only_one && acceptedPassId_ >= FLAGS_saving_period) { + paramUtil_->deleteParameters(acceptedPassId_ - FLAGS_saving_period); + } + } +} + +real Trainer::calcGradient(const DataBatch& dataBatch, + const Vector& value, + Vector& gradient) { + CHECK_EQ(value.getSize(), gradient.getSize()); + std::vector& parameters = + trainerInternal_.getGradientMachine()->getParameters(); + + clearGradient(); + + size_t offset = 0; + size_t valueSize = value.getSize(); + + for (auto& para : parameters) { + CHECK_LE(offset + para->getSize(), valueSize); + VectorPtr val = + Vector::create(para->getSize(), value.getMemoryHandle(), offset); + para->getBuf(PARAMETER_VALUE)->copyFrom(*val); + para->setValueUpdated(); + offset += para->getSize(); + } + + CHECK_EQ(offset, valueSize); + + std::vector inArgs = dataBatch.getStreams(); + std::vector outArgs; + + trainerInternal_.getGradientMachine()->forwardBackward( + inArgs, &outArgs, PASS_TRAIN); + real cost = Argument::sum(outArgs); + + offset = 0; + for (auto& para : parameters) { + VectorPtr grad = + Vector::create(para->getSize(), gradient.getMemoryHandle(), offset); + if (para->getBuf(PARAMETER_GRADIENT)) { + grad->copyFrom(*para->getBuf(PARAMETER_GRADIENT)); + } + offset += para->getSize(); + } + + return cost; +} + +void Trainer::clearGradient() { + std::vector& parameters = + trainerInternal_.getGradientMachine()->getNonStaticParameters(); + for (auto& parameter : parameters) { + parameter->clearGradient(); + } +} + +int Trainer::getBatchSize() { return config_->getOptConfig().batch_size(); } + +void Trainer::createTester() { + tester_.reset(new paddle::Tester(config_, + createTesterConfig(), + trainerInternal_.getGradientMachine(), + trainerInternal_.getParameterUpdater(), + testDataProvider_)); +} + +void Trainer::test() { tester_->test(); } + +std::unique_ptr Trainer::createTesterConfig() { + TesterConfig* conf = new TesterConfig; + if (FLAGS_test_period) { + LOG(WARNING) << "The meaning of --test_period is changed: " + << "if equal 0, do test on all test data at the end of " + << "each pass. While if equal non-zero, do test on all test " + << "data every test_period batches "; + } + if (FLAGS_test_all_data_in_one_period) { + LOG(WARNING) << "--test_all_data_in_one_period was deprecated, since " + << "we will always do test on all test set "; + } + conf->testPeriod = FLAGS_test_period; + conf->prevBatchState = FLAGS_prev_batch_state; + conf->logPeriod = FLAGS_log_period; + conf->loadsaveParametersInPserver = FLAGS_loadsave_parameters_in_pserver; + conf->featFile = FLAGS_feat_file; + conf->predictOutputDir = FLAGS_predict_output_dir; + conf->trainerId = FLAGS_trainer_id; + conf->distributeTest = FLAGS_distribute_test; + conf->config = FLAGS_config; + conf->modelList = FLAGS_model_list; + conf->testPass = FLAGS_test_pass; + conf->numPasses = FLAGS_num_passes; + conf->savingPeriod = FLAGS_saving_period; + conf->testWait = FLAGS_test_wait; + conf->initModelPath = FLAGS_init_model_path; + conf->saveOnlyOne = FLAGS_save_only_one; + conf->testing = testing_; + conf->mode = mode_; + conf->trainState = &trainState_; + conf->testState = &testState_; + return std::unique_ptr(conf); +} + +ParameterUtil* Trainer::getParameterUtilPtr() { return paramUtil_.get(); } +} // namespace paddle diff --git a/paddle/legacy/trainer/Trainer.h b/paddle/legacy/trainer/Trainer.h new file mode 100644 index 0000000000000000000000000000000000000000..b467f9af0cf12a39dd3d119c59e6cafcb05474b4 --- /dev/null +++ b/paddle/legacy/trainer/Trainer.h @@ -0,0 +1,204 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include + +#include "hl_gpu.h" +#include "paddle/legacy/gserver/dataproviders/DataProvider.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" + +#include +#include +#include "ParamUtil.h" +#include "ParameterUpdater.h" +#include "Tester.h" +#include "TrainerConfigHelper.h" +#include "TrainerInternal.h" + +DECLARE_int32(num_passes); + +namespace paddle { + +/** + * Trainer Class + * + * Trainer combines GradientMachine, ParameterUpdater, DataProvider together to + * train/test a NeuralNetwork. + */ +class Trainer { + public: + /** + * Ctor. + * @return + */ + Trainer() : acceptedPassId_(0) {} + + virtual ~Trainer() {} + + /** + * initialize a new trainer using config + * + * @param config TrainerConfig. + * @param testing true if only for testing + * @param gradientMachine GradientMachine that will be trained. + * nullptr if create from config. + * @param dataProvider Train Data Provider. null if create from config. + * @param testDataProvider Test Data Provider. null if create from config. + */ + virtual void init( + const std::shared_ptr& config, + bool testing = false, + const std::shared_ptr& gradientMachine = nullptr, + const std::shared_ptr& dataProvider = nullptr, + const std::shared_ptr& testDataProvider = nullptr); + + /** + * Train until num_passes reached. + * One pass means neural network train through all training data. + * + * @param numPasses the number of traning pass. + * @note Durning neural network training, the num passes may set a very large + * value, and kill training process when result is good enough. + */ + void train(size_t numPasses = (size_t)FLAGS_num_passes); + + /** + * compare the gradient from bp with finite difference + * @return the maximal difference + */ + real checkGradient(); + + void startTrain(); + void finishTrain(); + void startTrainPass(); + void finishTrainPass(); + void trainOneDataBatch(DataBatch& dataBatch); + void time(); + + /** + * given a dataBatch and the current parameter value + * calculate its gradient and return the cost. + * + * TODO(yuyang18): I think this method is deprecated and buggy. Should it be + * removed? + */ + real calcGradient(const DataBatch& dataBatch, + const Vector& value, + Vector& gradient); + + /** + * Get Trainer Config. + */ + const TrainerConfig& getConfig() const { return config_->getConfig(); } + + /** + * Get Train Data Provider + */ + const DataProviderPtr& getDataProvider() { return dataProvider_; } + + /** + * Get Gradient Machine. + */ + const GradientMachinePtr& getGradientMachine() { + return trainerInternal_.getGradientMachine(); + } + + /** + * Get batch size in optimization config. + * @note This method didn't return the actual batch size. Just batch size + * set in the optimization config. The actual batch size in one trainer may + * less than batch size in config due to there are not enough data. + */ + int getBatchSize(); + + /** + * Do test job + */ + void test(); + + /** + * Get parameter util ptr + * + * TODO(yuyang18): Make it return a smart pointer. + */ + ParameterUtil* getParameterUtilPtr(); + + protected: + /** + * Train one pass of data. + * + * SGD Method. + */ + void trainOnePass(); + + /** + * Train one pass in one batch. + * + */ + void trainOnePassBatch(int passId); + + /** + * set parameter gradient to zero + */ + void clearGradient(); + + void createTester(); + + private: + std::unique_ptr createTesterConfig(); + + protected: + std::shared_ptr config_; + std::shared_ptr stats_; + + DataProviderPtr dataProvider_; + DataProviderPtr testDataProvider_; + MachineState trainState_; + MachineState testState_; + + struct TrainPassContext { + int64_t batchId; + real avgTestCost; + int64_t numAvgTests; + int passId; + int passInnerId; + }; + std::vector forwardOutput_; + + TrainPassContext trainPassContext_; + + std::unique_ptr evaluator_; + std::unique_ptr currentEvaluator_; + std::unique_ptr averageEvaluator_; + // training mode + // used to decide which GradientMachine and ParameterUpdater to create + GradientMachine::CreateMode mode_; + int testing_; + int acceptedPassId_; + + // trainer tester + std::unique_ptr tester_; + + // parameter util + std::unique_ptr paramUtil_; + + // trainer Internal + TrainerInternal trainerInternal_; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/TrainerBenchmark.cpp b/paddle/legacy/trainer/TrainerBenchmark.cpp new file mode 100644 index 0000000000000000000000000000000000000000..7f5bd2335481c417b466ac4ca9ca54798524045f --- /dev/null +++ b/paddle/legacy/trainer/TrainerBenchmark.cpp @@ -0,0 +1,71 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#undef PADDLE_DISABLE_TIMER + +#include "Trainer.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +DECLARE_int32(test_period); + +DEFINE_bool(feed_data, false, "Wether to read data from DataProvider."); + +namespace paddle { + +void Trainer::time() { + startTrain(); + + trainerInternal_.getParameterUpdater()->startPass(); + evaluator_->start(); + + DataBatch dataBatch; + int32_t batchSize = config_->getOptConfig().batch_size(); + int32_t num = dataProvider_->getNextBatch(batchSize, &dataBatch); + CHECK_EQ(num, batchSize) << "The sample number is less than batch size " + << num << " != " << batchSize; + + CHECK(dataBatch.getSize()) << "No data from data provider"; + + std::vector outputs; + // burning time + LOG(INFO) << "Burning time..."; + for (int n = 0; n < 10; ++n) { + trainerInternal_.trainOneBatch(n, dataBatch, &outputs); + } + LOG(INFO) << "Burning time end."; + + for (int n = 0; n < FLAGS_test_period; n++) { + if (FLAGS_feed_data) { + REGISTER_TIMER("GetData"); + num = dataProvider_->getNextBatch(batchSize, &dataBatch); + } + + if (num != batchSize) { + break; + } + + { + REGISTER_TIMER("FwdBwd"); + trainerInternal_.trainOneBatch(n, dataBatch, &outputs); + } + } + globalStat.setThreadInfo(true); + globalStat.printSegTimerStatus(); + globalStat.reset(); + + finishTrain(); +} + +} // namespace paddle diff --git a/paddle/legacy/trainer/TrainerConfigHelper.cpp b/paddle/legacy/trainer/TrainerConfigHelper.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4d31ba8d71d52ac51191affc612a79b6734dee74 --- /dev/null +++ b/paddle/legacy/trainer/TrainerConfigHelper.cpp @@ -0,0 +1,199 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "TrainerConfigHelper.h" +#include "ParamUtil.h" +#include "TrainerConfig.pb.h" +#include "paddle/legacy/utils/Flags.h" +#include "paddle/legacy/utils/PythonUtil.h" + +DECLARE_string(config); +DECLARE_string(init_model_path); +DECLARE_int32(start_pass); +DECLARE_string(save_dir); +DECLARE_int32(trainer_id); +DECLARE_bool(local); +DECLARE_bool(with_cost); +DECLARE_bool(with_gpu); +DECLARE_bool(parallel_nn); +DECLARE_string(config_args); +DECLARE_bool(use_mkldnn); +DECLARE_bool(use_mkl_packed); + +const char *kConfigParserModuleName = "paddle.trainer.config_parser"; +const char *kConfigParserFuncName = "parse_config_and_serialize"; + +namespace paddle { + +struct TrainerConfigHelperPrivate { + TrainerConfig conf; +}; + +TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath) + : m(new TrainerConfigHelperPrivate()) { + std::ostringstream configArgs; + configArgs << "trainer_id=" << FLAGS_trainer_id << ",local=" << FLAGS_local + << ",with_cost=" << FLAGS_with_cost << ",use_gpu=" << FLAGS_use_gpu + << ",parallel_nn=" << FLAGS_parallel_nn + << ",use_mkldnn=" << FLAGS_use_mkldnn + << ",use_mkl_packed=" << FLAGS_use_mkl_packed + << ",cudnn_version=" << hl_get_cudnn_lib_version(); + if (!FLAGS_config_args.empty()) { + configArgs << "," << FLAGS_config_args; + } + + VLOG(3) << "Parsing trainer config " << configFilePath; + std::string configProtoStr = + callPythonFunc(kConfigParserModuleName, + kConfigParserFuncName, + {configFilePath, configArgs.str()}); + CHECK(m->conf.ParseFromString(configProtoStr)); +} + +TrainerConfigHelper::TrainerConfigHelper(const TrainerConfig &config) + : m(new TrainerConfigHelperPrivate()) { + m->conf = config; +} + +TrainerConfigHelper::~TrainerConfigHelper() { delete m; } + +const TrainerConfig &TrainerConfigHelper::getConfig() const { return m->conf; } + +TrainerConfig &TrainerConfigHelper::getMutableConfig() { return m->conf; } + +const OptimizationConfig &TrainerConfigHelper::getOptConfig() const { + return m->conf.opt_config(); +} + +const ModelConfig &TrainerConfigHelper::getModelConfig() const { + return m->conf.model_config(); +} + +const DataConfig *TrainerConfigHelper::getDataConfigPtr() const { + if (m->conf.has_data_config()) { + return &m->conf.data_config(); + } else { + return nullptr; + } +} + +const DataConfig &TrainerConfigHelper::getTestDataConfig() const { + CHECK(m->conf.has_test_data_config()); + return m->conf.test_data_config(); +} + +bool TrainerConfigHelper::hasDataConfig() const { + return m->conf.has_data_config(); +} + +bool TrainerConfigHelper::hasTestDataConfig() const { + return m->conf.has_test_data_config(); +} + +void TrainerConfigHelper::updateConfigFromFlags() { + if (!FLAGS_save_dir.empty()) { + m->conf.set_save_dir(FLAGS_save_dir); + } + if (!FLAGS_init_model_path.empty()) { + m->conf.set_init_model_path(FLAGS_init_model_path); + } + if (FLAGS_start_pass != 0) { + m->conf.set_start_pass(FLAGS_start_pass); + } +} + +void TrainerConfigHelper::disableRemoteSparseUpdater() { + m->conf.mutable_opt_config()->set_use_sparse_remote_updater(false); +} + +void TrainerConfigHelper::disableRemoteSparseUpdaterForEachParams() { + this->disableRemoteSparseUpdater(); + for (int i = 0; i < m->conf.model_config().parameters_size(); ++i) { + m->conf.mutable_model_config() + ->mutable_parameters(i) + ->set_sparse_remote_update(false); + } +} + +OptimizationConfig &TrainerConfigHelper::getOptConfig() { + return *m->conf.mutable_opt_config(); +} + +void TrainerConfigHelper::setSaveDir(const std::string &saveDir) { + m->conf.set_save_dir(saveDir); +} + +const std::string &TrainerConfigHelper::getSaveDir() const { + return m->conf.save_dir(); +} + +std::string TrainerConfigHelper::getConfigNameFromPath( + const std::string &modelPath) { + std::ifstream s(path::join(modelPath, "path.txt")); + CHECK(s.is_open()) << " fail to open path.txt"; + std::string ss; + getline(s, ss); + VLOG(3) << "fileName " << path::join(modelPath, ss); + s.close(); + return path::join(modelPath, ss); +} + +std::string TrainerConfigHelper::getConfigNameFromPassId( + int passId, const std::string &modelPath) { + constexpr int kBufLen = 100; + char buf[kBufLen]; + snprintf(buf, kBufLen, "pass-%05d", passId); + return TrainerConfigHelper::getConfigNameFromPath(path::join(modelPath, buf)); +} + +std::string TrainerConfigHelper::getConfigName(bool *ok) const { + std::string retv = ""; + + if (!m->conf.config_file().empty()) { + retv = m->conf.config_file(); + } else if (!m->conf.init_model_path().empty()) { + retv = getConfigNameFromPath(m->conf.init_model_path()); + } else if (m->conf.start_pass() >= 1) { + retv = getConfigNameFromPassId(m->conf.start_pass(), m->conf.save_dir()); + } + + if (ok) { + *ok = !retv.empty(); + } + + return retv; +} + +std::shared_ptr TrainerConfigHelper::createFromFlags() { + std::string configPath; + if (!FLAGS_config.empty()) { + configPath = FLAGS_config; + } else if (!FLAGS_init_model_path.empty()) { + configPath = getConfigNameFromPath(FLAGS_init_model_path); + } else if (FLAGS_start_pass >= 1) { + configPath = + getConfigNameFromPassId(FLAGS_start_pass - 1, FLAGS_init_model_path); + } else { + return nullptr; + } + return std::make_shared(configPath); +} + +std::shared_ptr +TrainerConfigHelper::createFromFlagConfig() { + CHECK(!FLAGS_config.empty()); + return std::make_shared(FLAGS_config); +} + +} // namespace paddle diff --git a/paddle/legacy/trainer/TrainerConfigHelper.h b/paddle/legacy/trainer/TrainerConfigHelper.h new file mode 100644 index 0000000000000000000000000000000000000000..0e428bea2c4b44bf98772ccca8f8b10d315efbbd --- /dev/null +++ b/paddle/legacy/trainer/TrainerConfigHelper.h @@ -0,0 +1,205 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +namespace paddle { + +class TrainerConfig; +class OptimizationConfig; +struct TrainerConfigHelperPrivate; +class ModelConfig; +class DataConfig; + +/** + * @brief TrainerConfig Helper. A class wrap protobuf's TrainerConfig Object, + * simplize the usage for TrainerConfig. + * + * The all operation to TrainerConfig object should use this object. It remove + * many copy & paste code in trainer. + * + * @TODO(yuyang18): Make cmake check compiler support keyword 'final' or not. + * Define a macro to unify 'final' keyword + */ +class TrainerConfigHelper /*final*/ { + public: + DISABLE_COPY(TrainerConfigHelper); + + /** + * @brief Ctor, Create a TrainerConfig from config file + * @param configFilePath Config file path. + */ + explicit TrainerConfigHelper(const std::string& configFilePath); + explicit TrainerConfigHelper(const TrainerConfig& config); + + /** + * Dtor + * @warning this class is a final class. Should not be inherited. + */ + ~TrainerConfigHelper(); + + /** + * @brief Get Trainer Config itself. + */ + const TrainerConfig& getConfig() const; + + TrainerConfig& getMutableConfig(); + + /** + * @brief Get Optimizer Config. + */ + const OptimizationConfig& getOptConfig() const; + + /** + * @brief Get Model Config. + */ + const ModelConfig& getModelConfig() const; + + /** + * @brief Get Train Data Config Pointer. + * @return nullptr if there is no train data. Else will return pointer + */ + const DataConfig* getDataConfigPtr() const; + + /** + * @brief Get Tain Data Config. + * @warning Core when there is no train data. + */ + const DataConfig& getDataConfig() const { + CHECK(this->hasDataConfig()); + auto conf = this->getDataConfigPtr(); + return *conf; + } + + /** + * @brief Get test data config + * @warning Core when there is no test data. + */ + const DataConfig& getTestDataConfig() const; + + /** + * @brief Has train data config or not. + * @return true if has train data. + */ + bool hasDataConfig() const; + + /** + * @brief Has test data config or not. + * @return true if has test data. + */ + bool hasTestDataConfig() const; + + /** + * @brief Update trainer config from command line flags. + * Override config's (save_dir, init_model_path, start_pass) if command + * flags is existed. + */ + void updateConfigFromFlags(); + + /** + * @brief Disable optimization's sparse remote update. + */ + void disableRemoteSparseUpdater(); + + /** + * @brief Disable optimization and each parameter's sparse remote update. + */ + void disableRemoteSparseUpdaterForEachParams(); + + /** + * @brief implicit conversion. + */ + inline operator const TrainerConfig&() const { return this->getConfig(); } + + /** + * @brief implicit conversion. + */ + inline operator const OptimizationConfig&() const { + return this->getOptConfig(); + } + + /** + * @brief implicit conversion. + */ + inline operator const DataConfig&() const { return this->getDataConfig(); } + + /** + * @brief implicit conversion. + */ + inline operator const ModelConfig&() const { return this->getModelConfig(); } + + /** + * @brief Get mutable optimization config. + */ + OptimizationConfig& getOptConfig(); + + /** + * @brief set model save directory. + * @param saveDir Directory path. + */ + void setSaveDir(const std::string& saveDir); + + /** + * @brief get model save directory. + * @return save directory path. + */ + const std::string& getSaveDir() const; + + /** + * @brief Get config file name from model path. + * + * Paddle save model to a directory, and write a file 'path.txt' which save + * config filename. + * + * @param modelPath model saved directory. + * @return config file name. + */ + static std::string getConfigNameFromPath(const std::string& modelPath); + + /** + * @brief Get config file name from this config instance. + * @param[out] ok true if no error. + * @return config file name. + */ + std::string getConfigName(bool* ok = nullptr) const; + + /** + * @brief Try to create TrainerConfigHelper from all command line flags. + * Try to load from --config, --init_model_path, --start_pass one by + * one. Return nullptr if cannot load TrainerConfigHelper from all + * these place. + * @return nullptr if cannot load, otherwise return a TrainerConfigHelper. + */ + static std::shared_ptr createFromFlags(); + + /** + * @brief Try to create TrainerConfigHelper only from '--config' flag. + * @return nullptr if cannot load, otherwise return a TrainerConfigHelper. + */ + static std::shared_ptr createFromFlagConfig(); + + private: + static std::string getConfigNameFromPassId(int passId, + const std::string& modelPath); + + TrainerConfigHelperPrivate* m; +}; + +typedef std::shared_ptr TrainerConfigHelperPtr; + +} // namespace paddle diff --git a/paddle/legacy/trainer/TrainerInternal.cpp b/paddle/legacy/trainer/TrainerInternal.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ee3dea6340167ab16d2bfefe3d757b10f5d90bb5 --- /dev/null +++ b/paddle/legacy/trainer/TrainerInternal.cpp @@ -0,0 +1,303 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "TrainerInternal.h" + +#include +#include + +#include +#include +#include +#include + +#include + +#include "paddle/legacy/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/legacy/gserver/layers/ValidationLayer.h" +#include "paddle/legacy/utils/GlobalConstants.h" +#include "paddle/legacy/utils/PythonUtil.h" +#include "paddle/legacy/utils/Stat.h" +#include "paddle/legacy/utils/Util.h" + +#include "RemoteParameterUpdater.h" +#include "ThreadParameterUpdater.h" + +namespace paddle { + +void TrainerInternal::init(const std::shared_ptr& config, + const GradientMachinePtr& gradientMachine, + std::unique_ptr&& intconfig, + const std::shared_ptr& stats, + bool testing) { + config_ = config; + intconfig_ = std::move(intconfig); + stats_ = stats; + + //! in training will use parameter updater definitly. + //! But only use parameter in testing mode when some parameter in pserver. + if (!testing || (config_->getOptConfig().use_sparse_remote_updater() && + intconfig_->loadsave_parameters_in_pserver)) { + createParameterUpdater(testing); + } + + gradientMachine_ = gradientMachine; + if (!gradientMachine) { + CHECK(config_->getConfig().has_model_config()) + << "Missing model_config in trainer_config"; + gradientMachine_.reset( + GradientMachine::create(config_->getConfig().model_config(), + intconfig_->mode, + parameterUpdater_->getParameterTypes())); + } +} + +void TrainerInternal::trainOneBatch(int64_t batchId, + const DataBatch& dataBatch, + std::vector* outArgs) { + // true means updating parameter whenever gradient is ready during backward() + bool doPipelineUpdate = + (intconfig_->mode != GradientMachine::kSgdSparseCpuTraining) && + (intconfig_->local || intconfig_->use_gpu || + intconfig_->trainer_count <= 1); + + int64_t actualBatchSize = dataBatch.getSize(); + if (actualBatchSize == 0) { + return; + } + + bool showStats = intconfig_->show_param_stats_period > 0 && + (batchId + 1) % intconfig_->show_param_stats_period == 0 && + intconfig_->trainer_id == 0; + + std::vector paraStats; + if (showStats) { + paraStats.resize(gradientMachine_->getParameters().size()); + } + + const std::vector& inArgs = dataBatch.getStreams(); + + PassType passType = parameterUpdater_->startBatch(actualBatchSize); + + if (config_->getOptConfig().use_sparse_remote_updater()) { + REGISTER_TIMER("prefetch"); + gradientMachine_->prefetch(inArgs); + parameterUpdater_->getParametersRemote(); + } + + UpdateCallback updateCallback = [this, showStats, ¶Stats]( + Parameter* para) { + if (showStats) { + //! @TODO(yuyang18) Show stats is actually a ParameterHook, refactor + // it + //! to ParameterHook. + auto& grad = para->getBuf(PARAMETER_GRADIENT); + SetDevice device(para->getDeviceId()); + paraStats[para->getID()].avgAbsGrad = grad->getAbsSum() / para->getSize(); + paraStats[para->getID()].maxAbsGrad = grad->getAbsMax(); + } + parameterUpdater_->update(para); + }; + + { +#ifndef PADDLE_DISABLE_TIMER + Timer timer; + timer.start(); +#endif + REGISTER_TIMER("forwardBackward"); + forwardBackwardBatch( + inArgs, *outArgs, passType, updateCallback, doPipelineUpdate); +#ifndef PADDLE_DISABLE_TIMER + timer.stop(); + parameterUpdater_->setForwardbackwardTime(timer.get()); +#endif + } + + if (!doPipelineUpdate) { + auto& parameters = gradientMachine_->getNonStaticParameters(); + for (auto& para : parameters) { + updateCallback(para.get()); + } + } + + real cost = 0; + { + REGISTER_TIMER("sumCost"); + cost = Argument::sum(*outArgs); + } + + if (batchId % intconfig_->log_period == 0) { + currentEvaluator_->start(); + stats_->resetCurrentStat(); + } + { + REGISTER_TIMER("eval"); + gradientMachine_->eval(currentEvaluator_); + gradientMachine_->eval(evaluator_); + } + + *stats_ += {actualBatchSize, cost}; + { + REGISTER_TIMER("finishBatch"); + parameterUpdater_->finishBatch(cost); + } + + if (showStats) { + showParameterStats(paraStats); + } + if ((batchId + 1) % intconfig_->log_period == 0) { + currentEvaluator_->finish(); + + if (intconfig_->dot_period > 0) { + std::cerr << std::endl; + } + LOG(INFO) << " Batch=" << batchId + 1 << " " << *stats_ + << " Eval: " << *evaluator_ + << " CurrentEval: " << *currentEvaluator_; + } else if (intconfig_->dot_period > 0 && + (batchId + 1) % intconfig_->dot_period == 0) { + std::cerr << "."; + } +} + +/** + * finish train pass + */ +void TrainerInternal::finishTrainPass(int passId, int batchId) { + gradientMachine_->onPassEnd(); + parameterUpdater_->finishPass(); + evaluator_->finish(); + LOG(INFO) << " Pass=" << passId << " Batch=" << batchId << " " + << stats_->getStats(false /*without current cost*/) + << " Eval: " << *evaluator_; +} + +void TrainerInternal::showParameterStats( + const std::vector& paraStats) { + std::vector& parameters = gradientMachine_->getParameters(); + for (auto& parameter : parameters) { + SetDevice device(parameter->getDeviceId()); + real sum = parameter->getBuf(PARAMETER_VALUE)->getAbsSum(); + const auto& lr = parameter->getBuf(PARAMETER_LEARNING_RATE); + std::ostringstream osLrHistogram; + if (lr) { + if (VLOG_IS_ON(2)) { + osLrHistogram << " lr_histogram: "; + lr->histogram(osLrHistogram); + } else { + osLrHistogram << " max_lr=" << std::setw(11) << lr->getMax() + << " min_lr=" << std::setw(11) << lr->getMin() + << " avg_lr=" << std::setw(11) + << lr->getSum() / parameter->getSize(); + } + } + int pid = parameter->getID(); + LOG(INFO) << std::setiosflags(std::ios::left) << std::setfill(' ') + << std::setw(20) << parameter->getName() + << " avg_abs_val=" << std::setw(11) << sum / parameter->getSize() + << " max_val=" << std::setw(11) + << parameter->getBuf(PARAMETER_VALUE)->getAbsMax() + << " avg_abs_grad=" << std::setw(11) << paraStats[pid].avgAbsGrad + << " max_grad=" << std::setw(11) << paraStats[pid].maxAbsGrad + << osLrHistogram.str(); + } +} + +void TrainerInternal::createParameterUpdater(bool testing) { + const std::string& alg = config_->getOptConfig().algorithm(); + parameterUpdater_.reset(ParameterUpdaterCreators::tryCreateUpdater( + alg, config_->getOptConfig(), intconfig_->local, intconfig_->num_passes)); + if (parameterUpdater_) { + return; + } + + if (!intconfig_->local) { + if (testing && config_->getOptConfig().use_sparse_remote_updater()) { + std::unique_ptr localUpdater; + localUpdater.reset( + new SgdLocalUpdater(config_->getOptConfig())); // do nothing + parameterUpdater_.reset( + new SparseRemoteParameterUpdaterComposite(config_->getOptConfig(), + intconfig_->num_passes, + testing, + std::move(localUpdater))); + } else { + if (GradientMachine::kSgdSparseCpuTraining == intconfig_->mode && + !intconfig_->use_old_updater) { + intconfig_->use_old_updater = true; + LOG(INFO) << "Sgd sparse training can not work with" + << " ConcurrentRemoteParameterUpdater," + << " automatically reset --use_old_updater=true"; + } + + std::unique_ptr localUpdater; + if (config_->getOptConfig().num_batches_per_send_parameter() > 1) { + CHECK(alg == TrainAlgorithm::SGD || alg == TrainAlgorithm::AsyncSGD) + << "Unsupported algorithm in remote-local mode: " << alg; + if (GradientMachine::kSgdSparseCpuTraining == intconfig_->mode) { + localUpdater.reset(new SgdThreadUpdater(*config_)); + } else { + localUpdater.reset(new SgdLocalUpdater(*config_)); + } + } + + localUpdater.reset( + intconfig_->use_old_updater + ? new RemoteParameterUpdater( + *config_, intconfig_->num_passes, std::move(localUpdater)) + : new ConcurrentRemoteParameterUpdater( + *config_, intconfig_->num_passes, std::move(localUpdater))); + + if (config_->getOptConfig().use_sparse_remote_updater()) { + localUpdater.reset( + new SparseRemoteParameterUpdaterComposite(*config_, + intconfig_->num_passes, + testing, + std::move(localUpdater))); + } + + this->parameterUpdater_ = std::move(localUpdater); + } + } else { + CHECK_EQ(config_->getOptConfig().num_batches_per_send_parameter(), 1) + << "num_batches_per_send_parameter should be one in local mode!"; + + if (GradientMachine::kSgdSparseCpuTraining == intconfig_->mode) { + parameterUpdater_.reset(new SgdThreadUpdater(*config_)); + } else if (alg == TrainAlgorithm::SGD || alg == TrainAlgorithm::AsyncSGD) { + if (config_->getModelConfig().type() == "recursive_nn") { + parameterUpdater_.reset(new SgdCpuUpdater(*config_)); + } else if (intconfig_->use_gpu && + config_->getOptConfig().do_average_in_cpu() && + config_->getOptConfig().average_window() > 0) { + parameterUpdater_.reset(new SgdUpdaterWithCpuAverager(*config_)); + } else { + parameterUpdater_.reset(new SgdLocalUpdater(*config_)); + } + } else { + LOG(FATAL) << "Unsupported algorithm in local mode: " << alg; + } + } +} + +void TrainerInternal::forwardBackwardBatch(const std::vector& inArgs, + std::vector& outArgs, + PassType& passType, + UpdateCallback updateCallback, + bool doPipelineUpdate) { + gradientMachine_->forwardBackward( + inArgs, &outArgs, passType, doPipelineUpdate ? updateCallback : nullptr); +} + +} // namespace paddle diff --git a/paddle/legacy/trainer/TrainerInternal.h b/paddle/legacy/trainer/TrainerInternal.h new file mode 100644 index 0000000000000000000000000000000000000000..93919a68fca2930cdf106f45d112e2a459fe695a --- /dev/null +++ b/paddle/legacy/trainer/TrainerInternal.h @@ -0,0 +1,139 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include +#include +#include + +#include "ParameterUpdater.h" +#include "TrainerConfig.pb.h" +#include "TrainerConfigHelper.h" +#include "TrainerInternalConfig.h" +#include "hl_gpu.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" + +namespace paddle { + +/** + * TrainerInteral + * the core training class for driving training logic + */ +class TrainerInternal { + public: + struct ParaStat { + real maxAbsGrad; + real avgAbsGrad; + ParaStat() : maxAbsGrad(.0), avgAbsGrad(.0) {} + }; + + TrainerInternal() {} + + /** + * Intializes trainer internal class + * @param config network config + * @param machine gradient machine + * @param intconfig training config + * @param stats training stats + * @param testing if it is in testing phase + */ + void init(const std::shared_ptr& config, + const GradientMachinePtr& machine, + std::unique_ptr&& intconfig, + const std::shared_ptr& stats, + bool testing); + + virtual ~TrainerInternal() {} + + /** + * CreateParameterUpdater + * @param testing if it is in testing phase + */ + void createParameterUpdater(bool testing); + + /** + * FinishTrainPass + * @param passId current pass id + * @param batchId current batch id, starts from 0 + */ + void finishTrainPass(int passId, int batchId); + + /** + * trainOneBatch + * @param batchId current batch id + * @param dataBatch data for the batch + */ + void trainOneBatch(int64_t batchId, + const DataBatch& dataBatch, + std::vector* outArgs); + + /** + * showParameterStats + * @param paraStats training stats + */ + void showParameterStats(const std::vector& paraStats); + + /** + * getGradientMachine + */ + inline const GradientMachinePtr& getGradientMachine() const { + return gradientMachine_; + } + + /** + * getParameterUpdater + */ + inline const std::shared_ptr& getParameterUpdater() { + return parameterUpdater_; + } + + /** + * setCurrentEvaluator + * @param eval evaluator to set + */ + inline void setCurrentEvaluator(Evaluator* eval) { currentEvaluator_ = eval; } + + /** + * setEvaluator + * @param eval evaluator to set + */ + inline void setEvaluator(Evaluator* eval) { evaluator_ = eval; } + + /** + * forwardBackwardBatch + * @param inArgs input argument for data batch + * @param outArgs output argument from neural network + * @param updateCallback layerwise parameter gradient statistics + * @param doPipelineUpdate whether to do pipeline update + */ + virtual void forwardBackwardBatch(const std::vector& inArgs, + std::vector& outArgs, + PassType& passType, + UpdateCallback updateCallback, + bool doPipelineUpdate); + + protected: + std::shared_ptr parameterUpdater_; + GradientMachinePtr gradientMachine_; + std::shared_ptr config_; + std::unique_ptr intconfig_; + std::shared_ptr stats_; + Evaluator* currentEvaluator_; + Evaluator* evaluator_; +}; + +} // namespace paddle diff --git a/paddle/trainer/TrainerInternalConfig.cpp b/paddle/legacy/trainer/TrainerInternalConfig.cpp similarity index 100% rename from paddle/trainer/TrainerInternalConfig.cpp rename to paddle/legacy/trainer/TrainerInternalConfig.cpp diff --git a/paddle/legacy/trainer/TrainerInternalConfig.h b/paddle/legacy/trainer/TrainerInternalConfig.h new file mode 100644 index 0000000000000000000000000000000000000000..b91b53932381a8698b331a2989b5f16829c06a25 --- /dev/null +++ b/paddle/legacy/trainer/TrainerInternalConfig.h @@ -0,0 +1,233 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/legacy/utils/Util.h" + +#include + +#include "hl_gpu.h" +#include "paddle/legacy/gserver/gradientmachines/GradientMachine.h" + +#include "TrainerConfig.pb.h" + +#include +#include +#include +#include "ParameterUpdater.h" + +namespace paddle { +/** + * @brief TrainerStats object will statistics sample processed and total cost. + * + * There are two stats in it, the 'AvgCost' and 'CurrentAvgCost'. 'AvgCost' + * means cost through one pass(all mini-batches). 'CurrentAvgCost' means cost + * through one mini-batch. + */ +class TrainerStats { + public: + /** + * @brief reset all stats. + * + * often used before pass start. + */ + inline void reset() { + numProcessed_ = 0; + totalCost_ = .0; + this->resetCurrentStat(); + } + + /** + * @brief reset current stat. + * + * 'current' means the most recent --log_period mini-batches + */ + inline void resetCurrentStat() { + currentCost_ = .0; + currentSamples_ = 0; + } + + /** + * @brief add cost to stat. + * @param numProcessed current mini-batch size + * @param cost current mini-batch cost + */ + inline void addCost(int64_t numProcessed, real cost) { + this->numProcessed_ += numProcessed; + this->totalCost_ += cost; + this->currentSamples_ += numProcessed; + this->currentCost_ += cost; + } + + /** + * @brief get average cost through on pass(all processed mini-batches) + * @return pass average cost + */ + inline real getAvgCost() const { + CHECK_NE(this->numProcessed_, 0); + return this->totalCost_ / this->numProcessed_; + } + + /** + * @brief get current mini-batch's average cost. + * @return mini-batch average cost + */ + inline real getCurrentAvgCost() const { + CHECK_NE(this->currentSamples_, 0); + return this->currentCost_ / this->currentSamples_; + } + + /** + * @brief get all processed samples' number + * @return all processed samples' number + */ + inline int64_t getNumProcessed() const { return this->numProcessed_; } + + /** + * @brief same function as addCost. But it is simple to invoke. + * For example: + * + * @code{.cpp} + * TrainerStats stat; + * cost = neuralNetwork.forward(batchSize); + * stat += {batchSize, cost}; + * @endcode + * + * @param p a pair of parameter, first is numProcessed, second is cost. + * @return *this + */ + inline TrainerStats& operator+=(const std::pair& p) { + this->addCost(p.first, p.second); + return *this; + } + + /** + * @brief TrainerStats Constructor. + * + * reset stat when constructed. + */ + inline TrainerStats() { this->reset(); } + + /** + * @brief show stats to ostream. + * + * If there is no need to print current cost, set withCurrentCost to False. + * + * @param os output stream. + * @param withCurrentCost print current cost or not. + */ + void showStats(std::ostream& os, bool withCurrentCost = true) const { + os << "samples=" << this->getNumProcessed() + << " AvgCost=" << this->getAvgCost(); + if (withCurrentCost) { + os << " CurrentCost=" << this->getCurrentAvgCost(); + } + } + + /** + * @brief get stats to std::string + * @param withCurrentCost return current cost or not + * @return stats string + */ + std::string getStats(bool withCurrentCost = true) const { + std::ostringstream os; + this->showStats(os, withCurrentCost); + return os.str(); + } + + private: + int64_t numProcessed_; + real totalCost_; + real currentCost_; + int64_t currentSamples_; +}; + +inline std::ostream& operator<<(std::ostream& os, const TrainerStats& stats) { + stats.showStats(os); + return os; +} + +/** + * TrainerInternalConfig + * general configs for training + */ +struct TrainerInternalConfig { + /** + * @brief Create TrainerInternalConfig from GradientMachine::CreateMode and + * command line arguments. + * @param mode + * @return + */ + static std::unique_ptr createFromMode( + GradientMachine::CreateMode mode); + + /** + * indicate whether the training is local + * if local, no parameter server is used + */ + bool local; + + /** + * indicate whether training uses GPU + */ + bool use_gpu; + + /** + * indicate number of trainer + */ + int trainer_count; + + /** + * how frequently to show param stats + */ + int show_param_stats_period; + + /** + * current trainer id + */ + int trainer_id; + + /** + * frequency to dump log + */ + int log_period; + + /** + * dot period + */ + int dot_period; + + /** + * num passes for training + */ + int num_passes; + + /** + * use old updater + */ + bool use_old_updater; + + /** + * whether to load and save parameter in pserver + */ + bool loadsave_parameters_in_pserver; + + /** + * training mode + */ + GradientMachine::CreateMode mode; +}; + +} // namespace paddle diff --git a/paddle/legacy/trainer/TrainerMain.cpp b/paddle/legacy/trainer/TrainerMain.cpp new file mode 100644 index 0000000000000000000000000000000000000000..911aeba1928f7208aecb92910dac981f00fc6db5 --- /dev/null +++ b/paddle/legacy/trainer/TrainerMain.cpp @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/legacy/pserver/ParameterServerController.h" +#include "paddle/legacy/utils/PythonUtil.h" + +#include "ParamUtil.h" +#include "Trainer.h" + +DEFINE_bool(start_pserver, false, "Whether to start pserver"); +DECLARE_int32(gpu_id); +DEFINE_string(job, "train", "one of (train, test, checkgrad)"); +DECLARE_int32(start_pass); +DECLARE_string(config); +DECLARE_string(init_model_path); +DECLARE_string(rdma_tcp); + +using namespace paddle; // NOLINT + +int main(int argc, char** argv) { + // write logs instantly (never buffer log messages) + FLAGS_logbuflevel = -1; + + initMain(argc, argv); + initPython(argc, argv); + + std::unique_ptr parameterServerPtr(nullptr); + if (FLAGS_start_pserver) { + parameterServerPtr.reset( + paddle::ParameterServerController::createFromGflags()); + parameterServerPtr->start(); + } + Trainer trainer; + auto config = TrainerConfigHelper::createFromFlags(); + CHECK(config != nullptr) << "no valid config"; + + feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); + trainer.init(config, FLAGS_job == "test"); + + if (FLAGS_job == "train") { + trainer.train(); + } else if (FLAGS_job == "checkgrad") { + trainer.checkGradient(); + } else if (FLAGS_job == "test") { + trainer.test(); + } else if (FLAGS_job == "time") { + trainer.time(); + } else { + LOG(FATAL) << "Unknown job type: " << FLAGS_job; + } + + return 0; +} diff --git a/paddle/trainer/tests/.gitignore b/paddle/legacy/trainer/tests/.gitignore similarity index 100% rename from paddle/trainer/tests/.gitignore rename to paddle/legacy/trainer/tests/.gitignore diff --git a/paddle/legacy/trainer/tests/CMakeLists.txt b/paddle/legacy/trainer/tests/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..08548bea4c4a7fc4fa99d9305208abd4ee442572 --- /dev/null +++ b/paddle/legacy/trainer/tests/CMakeLists.txt @@ -0,0 +1,37 @@ +add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/sample_trainer_config.conf + COMMAND cp -r ${CMAKE_CURRENT_SOURCE_DIR}/* ${CMAKE_CURRENT_BINARY_DIR} +) +add_custom_target(copy_trainer_conf ALL DEPENDS sample_trainer_config.conf) + +set(PYTHON_PATH + ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d + ${PADDLE_BINARY_DIR}/python/:${PADDLE_BINARY_DIR}/paddle/legacy/trainer/tests) +function(trainer_test TARGET) + add_unittest_without_exec(${TARGET} ${TARGET}.cpp) + add_test(NAME ${TARGET} + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET} + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) +endfunction() + +trainer_test(test_Compare) +trainer_test(test_PyDataProviderWrapper) +trainer_test(test_recurrent_machine_generation) +trainer_test(test_Trainer) + +############### test_TrainerOnePass ########################## +if(WITH_PYTHON) + # only run test_TrainerOnePass when PYTHON is enabled, because train one pass + # is using PyDataProvider2. + add_unittest_without_exec(test_TrainerOnePass + test_TrainerOnePass.cpp) + add_test(NAME test_TrainerOnePass + COMMAND ${PYTHON_PATH} ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port + ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) +endif() + +#################### test_config_parser ######################### +add_test(NAME test_config_parser + COMMAND ${PYTHON_PATH} ${PYTHON_EXECUTABLE} + ${PADDLE_SOURCE_DIR}/paddle/legacy/trainer/tests/config_parser_test.py + WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) diff --git a/paddle/trainer/tests/__init__.py b/paddle/legacy/trainer/tests/__init__.py similarity index 100% rename from paddle/trainer/tests/__init__.py rename to paddle/legacy/trainer/tests/__init__.py diff --git a/paddle/legacy/trainer/tests/config_parser_test.py b/paddle/legacy/trainer/tests/config_parser_test.py new file mode 100644 index 0000000000000000000000000000000000000000..0d3d82cbdafcf85d42247e810fe7caa685a86e4d --- /dev/null +++ b/paddle/legacy/trainer/tests/config_parser_test.py @@ -0,0 +1,23 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer.config_parser import parse_config_and_serialize + +if __name__ == '__main__': + parse_config_and_serialize('legacy/trainer/tests/test_config.conf', '') + parse_config_and_serialize( + 'legacy/trainer/tests/sample_trainer_config.conf', + 'extension_module_name=paddle.trainer.config_parser_extension') + parse_config_and_serialize( + 'legacy/gserver/tests/pyDataProvider/trainer.conf', '') diff --git a/paddle/trainer/tests/fake_file_list.list b/paddle/legacy/trainer/tests/fake_file_list.list similarity index 100% rename from paddle/trainer/tests/fake_file_list.list rename to paddle/legacy/trainer/tests/fake_file_list.list diff --git a/paddle/trainer/tests/picojson.h b/paddle/legacy/trainer/tests/picojson.h similarity index 100% rename from paddle/trainer/tests/picojson.h rename to paddle/legacy/trainer/tests/picojson.h diff --git a/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.data b/paddle/legacy/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.data similarity index 100% rename from paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.data rename to paddle/legacy/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.data diff --git a/paddle/legacy/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.list b/paddle/legacy/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.list new file mode 100644 index 0000000000000000000000000000000000000000..11c1b1b38b9edacc4953fdf526906d28bcc2d720 --- /dev/null +++ b/paddle/legacy/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.list @@ -0,0 +1 @@ +legacy/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.data diff --git a/paddle/trainer/tests/rnn_gen_test_model_dir/r1.test.beam b/paddle/legacy/trainer/tests/rnn_gen_test_model_dir/r1.test.beam similarity index 100% rename from paddle/trainer/tests/rnn_gen_test_model_dir/r1.test.beam rename to paddle/legacy/trainer/tests/rnn_gen_test_model_dir/r1.test.beam diff --git a/paddle/trainer/tests/rnn_gen_test_model_dir/r1.test.nest b/paddle/legacy/trainer/tests/rnn_gen_test_model_dir/r1.test.nest similarity index 100% rename from paddle/trainer/tests/rnn_gen_test_model_dir/r1.test.nest rename to paddle/legacy/trainer/tests/rnn_gen_test_model_dir/r1.test.nest diff --git a/paddle/trainer/tests/rnn_gen_test_model_dir/r1.test.nobeam b/paddle/legacy/trainer/tests/rnn_gen_test_model_dir/r1.test.nobeam similarity index 100% rename from paddle/trainer/tests/rnn_gen_test_model_dir/r1.test.nobeam rename to paddle/legacy/trainer/tests/rnn_gen_test_model_dir/r1.test.nobeam diff --git a/paddle/trainer/tests/rnn_gen_test_model_dir/t1/transtable b/paddle/legacy/trainer/tests/rnn_gen_test_model_dir/t1/transtable similarity index 100% rename from paddle/trainer/tests/rnn_gen_test_model_dir/t1/transtable rename to paddle/legacy/trainer/tests/rnn_gen_test_model_dir/t1/transtable diff --git a/paddle/trainer/tests/rnn_gen_test_model_dir/t1/wordvec b/paddle/legacy/trainer/tests/rnn_gen_test_model_dir/t1/wordvec similarity index 100% rename from paddle/trainer/tests/rnn_gen_test_model_dir/t1/wordvec rename to paddle/legacy/trainer/tests/rnn_gen_test_model_dir/t1/wordvec diff --git a/paddle/trainer/tests/sample_data.txt b/paddle/legacy/trainer/tests/sample_data.txt similarity index 100% rename from paddle/trainer/tests/sample_data.txt rename to paddle/legacy/trainer/tests/sample_data.txt diff --git a/paddle/legacy/trainer/tests/sample_filelist.txt b/paddle/legacy/trainer/tests/sample_filelist.txt new file mode 100644 index 0000000000000000000000000000000000000000..8573f9e1795edd37cfa0d21f0effc08a80d38e29 --- /dev/null +++ b/paddle/legacy/trainer/tests/sample_filelist.txt @@ -0,0 +1 @@ +legacy/trainer/tests/sample_data.txt diff --git a/paddle/legacy/trainer/tests/sample_trainer_config.conf b/paddle/legacy/trainer/tests/sample_trainer_config.conf new file mode 100644 index 0000000000000000000000000000000000000000..5800b3625661efac80b84b19c2a5cedc34718488 --- /dev/null +++ b/paddle/legacy/trainer/tests/sample_trainer_config.conf @@ -0,0 +1,87 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +TrainData(SimpleData( + files = "legacy/trainer/tests/sample_filelist.txt", + feat_dim = 3, + context_len = 0, + buffer_capacity = 1000000)) + +TestData(SimpleData( + files = "legacy/trainer/tests/sample_filelist.txt", + feat_dim = 3, + context_len = 0, + buffer_capacity = 1000000)) + +settings(batch_size = 100) + +data = data_layer(name='input', size=3) + +fc1 = fc_layer(input=data, size=5, + bias_attr=False, + act=SigmoidActivation()) + +fc2 = fc_layer(input=data, size=9, + bias_attr=False, + act=LinearActivation()) + +fc3 = fc_layer(input=data, size=3, + bias_attr=False, + act=TanhActivation()) + +fc4 = fc_layer(input=data, size=5, + bias_attr=False, + act=LinearActivation(), + param_attr=ParamAttr(name='sharew')) + +fc5 = fc_layer(input=data, size=5, + bias_attr=False, + act=BReluActivation()) + +fc6 = fc_layer(input=data, size=5, + bias_attr=False, + act=SoftReluActivation()) + +fc7 = fc_layer(input=data, size=3, + bias_attr=False, + act=SquareActivation()) + +fc8 = fc_layer(input=data, size=5, + bias_attr=True, + act=SquareActivation()) + +with mixed_layer(size=3, act=SoftmaxActivation()) as layer9: + layer9 += full_matrix_projection(input=fc1) + layer9 += full_matrix_projection(input=fc2) + layer9 += full_matrix_projection(input=fc3) + layer9 += trans_full_matrix_projection(input=fc4, + param_attr=ParamAttr(name='sharew')) + layer9 += full_matrix_projection(input=fc5) + layer9 += full_matrix_projection(input=fc6) + layer9 += full_matrix_projection(input=fc7) + layer9 += full_matrix_projection(input=fc8) + +if get_config_arg('with_cost', bool, True): + # This is for training the neural network. + # We need to have another data layer for label + # and a layer for calculating cost + lbl = data_layer(name='label', size=1) + outputs(classification_cost(input=layer9, label=lbl)) +else: + # This is for prediction where we don't have label + # and don't need to calculate cost + outputs(layer9) diff --git a/paddle/legacy/trainer/tests/sample_trainer_config_hsigmoid.conf b/paddle/legacy/trainer/tests/sample_trainer_config_hsigmoid.conf new file mode 100644 index 0000000000000000000000000000000000000000..155c40b31f30c40e1ddeb65500f55162beb9a0ee --- /dev/null +++ b/paddle/legacy/trainer/tests/sample_trainer_config_hsigmoid.conf @@ -0,0 +1,53 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from paddle.trainer_config_helpers import * + +TrainData(SimpleData( + files = "legacy/trainer/tests/sample_filelist.txt", + feat_dim = 3, + context_len = 0, + buffer_capacity = 1000000, +)) + +settings(batch_size = 100) + +data = data_layer(name='input', size=3) + +fc1 = fc_layer(input=data, size=12, + bias_attr=False, + act=SigmoidActivation()) + +fc2 = fc_layer(input=data, size=19, + bias_attr=False, + act=LinearActivation()) + +fc3 = fc_layer(input=data, size=5, + bias_attr=False, + act=TanhActivation()) + +fc4 = fc_layer(input=data, size=5, + bias_attr=False, + act=LinearActivation()) + +# This is for training the neural network. +# We need to have another data layer for label +# and a layer for calculating cost +lbl = data_layer(name='label', size=1) + +outputs(hsigmoid(input=[fc1, fc2, fc3, fc4], + label=lbl, + num_classes=3)) diff --git a/paddle/legacy/trainer/tests/sample_trainer_config_parallel.conf b/paddle/legacy/trainer/tests/sample_trainer_config_parallel.conf new file mode 100644 index 0000000000000000000000000000000000000000..49cdde7fa2c55e6536a49633f959af6a888ec463 --- /dev/null +++ b/paddle/legacy/trainer/tests/sample_trainer_config_parallel.conf @@ -0,0 +1,86 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +TrainData(SimpleData( + files = "legacy/trainer/tests/sample_filelist.txt", + feat_dim = 3, + context_len = 0, + buffer_capacity = 1000000)) + +TestData(SimpleData( + files = "legacy/trainer/tests/sample_filelist.txt", + feat_dim = 3, + context_len = 0, + buffer_capacity = 1000000)) + +settings(batch_size = 100) + +# Output layer, label layer, cost layer, preferably set to the same environment. +output_device = 0 + +# Input Layer does not need to specify the device number. +data = data_layer(name='input', size=3) + +# Calculate in the CPU. +fc1 = fc_layer(input=data, size=5, + bias_attr=True, + layer_attr=ExtraAttr(device=-1), + act=SigmoidActivation()) + +# Calculate in the GPU 0. +fc2 = fc_layer(input=fc1, size=10, + bias_attr=True, + layer_attr=ExtraAttr(device=0), + act=SigmoidActivation()) + +# Calculate in the GPU 1. +fc3 = fc_layer(input=fc1, size=10, + bias_attr=True, + layer_attr=ExtraAttr(device=1), + act=SigmoidActivation()) + +# Calculate in the GPU 0. +fc4 = fc_layer(input=[fc2,fc3], size=10, + bias_attr=True, + layer_attr=ExtraAttr(device=0), + act=SigmoidActivation()) + +# Calculate in the GPU 1. +fc5 = fc_layer(input=[fc2,fc3], size=10, + bias_attr=True, + layer_attr=ExtraAttr(device=1), + act=SigmoidActivation()) + +output = fc_layer(input=[fc4,fc5], size=10, + bias_attr=True, + layer_attr=ExtraAttr(device=output_device), + act=SoftmaxActivation()) + +if get_config_arg('with_cost', bool, True): + # This is for training the neural network. + # We need to have another data layer for label + # and a layer for calculating cost + lbl = data_layer(name='label', size=1, + layer_attr=ExtraAttr(device=output_device)) + + outputs(classification_cost(input=output, + label=lbl, + layer_attr=ExtraAttr(device=output_device))) +else: + # This is for prediction where we don't have label + # and don't need to calculate cost + outputs(output) diff --git a/paddle/legacy/trainer/tests/sample_trainer_nest_rnn_gen.conf b/paddle/legacy/trainer/tests/sample_trainer_nest_rnn_gen.conf new file mode 100644 index 0000000000000000000000000000000000000000..51ef905a5a182464f69a1629e51bf8180eadb3fb --- /dev/null +++ b/paddle/legacy/trainer/tests/sample_trainer_nest_rnn_gen.conf @@ -0,0 +1,73 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from paddle.trainer_config_helpers import * + +settings(batch_size=15, learning_rate=0) + +num_words = 5 +beam_flag = get_config_arg('beam_search', bool, False) + +sent_id = data_layer(name="sent_id", size=1) + +# This layer has no actual use, but only to decide batch_size in generation. +# When generating, at least one Memory in RecurrentLayer MUST have a boot layer. +dummy_data = data_layer(name="dummy_data_input", size=2) + +def outer_step(dummy_data): + + gen_inputs = [StaticInput(input=dummy_data, size=2, is_seq=True), + GeneratedInput(size=num_words, + embedding_name="wordvec", + embedding_size=num_words)] + + def inner_step(dummy_memory, predict_word): + + # simplified RNN for testing + with mixed_layer(size=num_words) as layer: + layer += full_matrix_projection(input=predict_word, + param_attr=ParamAttr(name="transtable")) + + with mixed_layer(size=num_words, act=ExpActivation()) as out: + out += trans_full_matrix_projection(input=layer, + param_attr=ParamAttr(name="wordvec")) + + return out + + beam_gen = beam_search(name="rnn_gen", + step=inner_step, + input=gen_inputs, + bos_id=0, + eos_id=num_words-1, + beam_size=2 if beam_flag else 1, + num_results_per_sample=1, + max_length=10) + return beam_gen + +beam_gen_concat = recurrent_group(name="rnn_gen_concat", + step=outer_step, + input=[SubsequenceInput(dummy_data)]) + +seqtext_printer_evaluator(input=beam_gen_concat, + id_input=sent_id, + dict_file="./legacy/trainer/tests/test_gen_dict.txt", + result_file="./legacy/trainer/tests/dump_text.test") +#outputs(beam_gen_concat) +# In this config, as dummy_data_input doesn't work on beam_gen (we can find dummy_memory +# is read-only memory, and isn't used by other layers of step), we show the Inputs and Outputs +# as follows. Note that "__beam_search_predict__" is the default output name of beam_search. +Inputs("sent_id","dummy_data_input") +Outputs("__beam_search_predict__") diff --git a/paddle/legacy/trainer/tests/sample_trainer_rnn_gen.conf b/paddle/legacy/trainer/tests/sample_trainer_rnn_gen.conf new file mode 100644 index 0000000000000000000000000000000000000000..35c7f0fcd91f9b534a4f535387af720659d7f9b8 --- /dev/null +++ b/paddle/legacy/trainer/tests/sample_trainer_rnn_gen.conf @@ -0,0 +1,66 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from paddle.trainer_config_helpers import * + +settings(batch_size=15, learning_rate=0) + +num_words = 5 +beam_flag = get_config_arg('beam_search', bool, False) + +sent_id = data_layer(name="sent_id", size=1) + +# This layer has no actual use, but only to decide batch_size in generation. +# When generating, at least one Memory in RecurrentLayer MUST have a boot layer. +dummy_data = data_layer(name="dummy_data_input", size=2) + +gen_inputs = [StaticInput(input=dummy_data, size=2), + GeneratedInput(size=num_words, + embedding_name="wordvec", + embedding_size=num_words)] + +def step(dummy_memory, predict_word): + + # simplified RNN for testing + with mixed_layer(size=num_words) as layer: + layer += full_matrix_projection(input=predict_word, + param_attr=ParamAttr(name="transtable")) + + with mixed_layer(size=num_words, act=ExpActivation()) as out: + out += trans_full_matrix_projection(input=layer, + param_attr=ParamAttr(name="wordvec")) + + return out + +beam_gen = beam_search(name="rnn_gen", + step=step, + input=gen_inputs, + bos_id=0, + eos_id=num_words-1, + beam_size=2 if beam_flag else 1, + num_results_per_sample=2 if beam_flag else 1, + max_length=10) + +seqtext_printer_evaluator(input=beam_gen, + id_input=sent_id, + dict_file="./legacy/trainer/tests/test_gen_dict.txt", + result_file="./legacy/trainer/tests/dump_text.test") +#outputs(beam_gen) +# In this config, as dummy_data_input doesn't work on beam_gen (we can find dummy_memory +# is read-only memory, and isn't used by other layers of step), we show the Inputs and Outputs +# as follows. Note that "__beam_search_predict__" is the default output name of beam_search. +Inputs("sent_id","dummy_data_input") +Outputs("__beam_search_predict__") diff --git a/paddle/legacy/trainer/tests/simple_sparse_neural_network.py b/paddle/legacy/trainer/tests/simple_sparse_neural_network.py new file mode 100644 index 0000000000000000000000000000000000000000..9419f4d903b1de205a6c549c7dcd9bb85ed7396b --- /dev/null +++ b/paddle/legacy/trainer/tests/simple_sparse_neural_network.py @@ -0,0 +1,37 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +settings(batch_size=17, learning_method=AdaGradOptimizer(), learning_rate=1e-4) + +file_list = 'legacy/trainer/tests/fake_file_list.list' + +define_py_data_sources2( + train_list=file_list, + test_list=file_list, + module="simple_sparse_neural_network_dp", + obj="process") + +embedding = embedding_layer( + input=data_layer( + name="word_ids", size=8191), + size=128, + param_attr=ParamAttr(sparse_update=True)) +prediction = fc_layer(input=embedding, size=10, act=SoftmaxActivation()) + +outputs( + classification_cost( + input=prediction, label=data_layer( + name='label', size=10))) diff --git a/paddle/trainer/tests/simple_sparse_neural_network_dp.py b/paddle/legacy/trainer/tests/simple_sparse_neural_network_dp.py similarity index 100% rename from paddle/trainer/tests/simple_sparse_neural_network_dp.py rename to paddle/legacy/trainer/tests/simple_sparse_neural_network_dp.py diff --git a/paddle/trainer/tests/testPyDataWrapper.py b/paddle/legacy/trainer/tests/testPyDataWrapper.py similarity index 100% rename from paddle/trainer/tests/testPyDataWrapper.py rename to paddle/legacy/trainer/tests/testPyDataWrapper.py diff --git a/paddle/legacy/trainer/tests/test_Compare.cpp b/paddle/legacy/trainer/tests/test_Compare.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e37e546be8513b1cc7438810a01641859a4bad18 --- /dev/null +++ b/paddle/legacy/trainer/tests/test_Compare.cpp @@ -0,0 +1,158 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "paddle/legacy/trainer/Trainer.h" + +#include +#include + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static const string& configFile = + "legacy/trainer/tests/sample_trainer_config.conf"; + +DECLARE_int32(gpu_id); +DECLARE_bool(use_gpu); +DECLARE_string(config); +DECLARE_string(config_args); + +struct comData { + vector outArgs; + vector parameters; +}; + +void calcGradient(bool useGpu, comData& Data) { + FLAGS_use_gpu = useGpu; + FLAGS_config = configFile; + + *ThreadLocalRand::getSeed() = 0; + srand(0); + Trainer trainer; + trainer.init(TrainerConfigHelper::createFromFlagConfig()); + + Data.parameters = trainer.getGradientMachine()->getParameters(); + DataBatch dataBatch; + int32_t batchSize = trainer.getConfig().opt_config().batch_size(); + trainer.getDataProvider()->setSkipShuffle(); + trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch); + CHECK(dataBatch.getSize()) << "No data from data provider"; + vector& inArgs = dataBatch.getStreams(); + trainer.getGradientMachine()->start(); + for (int i = 0; i < 2; ++i) { + trainer.getGradientMachine()->forwardBackward( + inArgs, &Data.outArgs, PASS_TRAIN); + } + trainer.getGradientMachine()->finish(); +} + +void compareGradient(comData& comDataCpu, comData& comDataGpu); + +TEST(Trainer, create) { + int devCount = 0; + devCount = hl_get_device_count(); + FLAGS_config_args = "drop_rate=0"; + + comData comDataCpu; + calcGradient(false, comDataCpu); + LOG(INFO) << "Cpu is completed"; + + { + LOG(INFO) << "Test GPU"; + comData comData; + calcGradient(true, comData); + compareGradient(comDataCpu, comData); + LOG(INFO) << "Gpu is completed"; + } + + { + LOG(INFO) << "Test test multi gpu"; + comData comData; + FLAGS_trainer_count = devCount; + calcGradient(true, comData); + compareGradient(comDataCpu, comData); + LOG(INFO) << "Gpu4 is completed"; + } + + { + LOG(INFO) << "Test use_sparse_update=true"; + comData comData; + calcGradient(false, comData); + compareGradient(comDataCpu, comData); + LOG(INFO) << "Cpu4 is completed"; + } +} + +double checkBuffer(real* A, real* B, size_t len) { +#ifdef PADDLE_TYPE_DOUBLE + double precision = 1e-7; +#else + double precision = 2e-3; +#endif + int nNum = 0; + double maxE = 0; + for (size_t i = 0; i < len; ++i) { + double e = fabs(A[i] - B[i]); + maxE = std::max(e, maxE); + nNum += e > precision * fabs(A[i]); + } + EXPECT_EQ(0, nNum); + return maxE; +} + +void compareGradient(comData& comDataCpu, comData& comDataGpu) { + /*compare outArgs*/ + vector outArgs1 = comDataCpu.outArgs; + vector outArgs2 = comDataGpu.outArgs; + CpuMatrix out1(outArgs1[0].value->getHeight(), outArgs1[0].value->getWidth()); + CpuMatrix out2(outArgs2[0].value->getHeight(), outArgs2[0].value->getWidth()); + out1.copyFrom(*outArgs1[0].value); + out2.copyFrom(*outArgs2[0].value); + checkBuffer(out1.getData(), out2.getData(), out1.getElementCnt()); + + /*compare parameters*/ + vector& parameters1 = comDataCpu.parameters; + vector& parameters2 = comDataGpu.parameters; + for (size_t i = 0; i < parameters1.size(); ++i) { + ParameterPtr parameter1, parameter2; + parameter1 = parameters1[i]; + parameter2 = parameters2[i]; + /*compare parameters value*/ + CpuVector para1(parameter1->getSize()); + CpuVector para2(parameter2->getSize()); + para1.copyFrom(*parameter1->getBuf(PARAMETER_VALUE)); + para2.copyFrom(*parameter2->getBuf(PARAMETER_VALUE)); + checkBuffer(para1.getData(), para2.getData(), para1.getSize()); + + /*compare parameters grad*/ + CpuVector cpuGrad1(*parameter1->getBuf(PARAMETER_GRADIENT)); + CpuVector cpuGrad2(*parameter2->getBuf(PARAMETER_GRADIENT)); + double e = + checkBuffer(cpuGrad1.getData(), cpuGrad2.getData(), cpuGrad1.getSize()); + LOG(INFO) << parameter1->getName() << " max error=" << e; + } +} + +int main(int argc, char** argv) { +#ifndef PADDLE_WITH_CUDA + exit(0); +#endif + paddle::initMain(argc, argv); + testing::InitGoogleTest(&argc, argv); + initPython(argc, argv); + int ret = RUN_ALL_TESTS(); + exit(ret); +} diff --git a/paddle/legacy/trainer/tests/test_PyDataProviderWrapper.cpp b/paddle/legacy/trainer/tests/test_PyDataProviderWrapper.cpp new file mode 100644 index 0000000000000000000000000000000000000000..847adcfabada18e11203d3f18fb6dc355c670afb --- /dev/null +++ b/paddle/legacy/trainer/tests/test_PyDataProviderWrapper.cpp @@ -0,0 +1,220 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef PADDLE_NO_PYTHON +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "picojson.h" + +void checkValue(std::vector& arguments, picojson::array& arr); +const std::string kDir = "./legacy/trainer/tests/pydata_provider_wrapper_dir/"; + +TEST(PyDataProviderWrapper, SequenceData) { + paddle::DataConfig conf; + conf.set_type("py"); + conf.set_load_data_module("testPyDataWrapper"); + conf.set_load_data_object("processSeqAndGenerateData"); + conf.set_load_data_args(kDir + "test_pydata_provider_wrapper.json"); + conf.clear_files(); + conf.set_files(kDir + "test_pydata_provider_wrapper.list"); + paddle::DataProviderPtr provider(paddle::DataProvider::create(conf, false)); + provider->setSkipShuffle(); + provider->reset(); + paddle::DataBatch batchFromPy; + provider->getNextBatch(100, &batchFromPy); + + picojson::value val; + std::fstream fin; + fin.open(kDir + "test_pydata_provider_wrapper.json", std::ios_base::in); + EXPECT_TRUE(fin.is_open()); + if (fin.is_open()) { + std::string err = picojson::parse(val, fin); + EXPECT_TRUE(err.empty()); + EXPECT_TRUE(val.is()); + picojson::array& arr = val.get(); + std::vector& arguments = batchFromPy.getStreams(); + // CHECK Value + checkValue(arguments, arr); + // CHECK sequenceStartPositions + for (size_t i = 0; i < arr.size(); i++) { + int row_id = arr[i].get().size(); + EXPECT_EQ(0, arguments[i].sequenceStartPositions->getData(false)[0]); + EXPECT_EQ((int)row_id, + arguments[i].sequenceStartPositions->getData(false)[1]); + } + fin.close(); + } +} + +TEST(PyDataProviderWrapper, HasSubSequenceData) { + paddle::DataConfig conf; + conf.set_type("py"); + conf.set_load_data_module("testPyDataWrapper"); + conf.set_load_data_object("processSubSeqAndGenerateData"); + conf.set_load_data_args(kDir + "test_pydata_provider_wrapper.json"); + conf.clear_files(); + conf.set_files(kDir + "test_pydata_provider_wrapper.list"); + paddle::DataProviderPtr provider(paddle::DataProvider::create(conf, false)); + provider->setSkipShuffle(); + provider->reset(); + paddle::DataBatch batchFromPy; + provider->getNextBatch(1, &batchFromPy); + + picojson::value val; + std::fstream fin; + fin.open(kDir + "test_pydata_provider_wrapper.json", std::ios_base::in); + EXPECT_TRUE(fin.is_open()); + if (fin.is_open()) { + std::string err = picojson::parse(val, fin); + EXPECT_TRUE(err.empty()); + EXPECT_TRUE(val.is()); + picojson::array& arr = val.get(); + std::vector& arguments = batchFromPy.getStreams(); + // CHECK Value + checkValue(arguments, arr); + // CHECK sequenceStartPositions and subSequenceStartPositions + for (size_t i = 0; i < arr.size(); i++) { + int row_id = arr[i].get().size(); + EXPECT_EQ(0, arguments[i].sequenceStartPositions->getData(false)[0]); + EXPECT_EQ((int)row_id, + arguments[i].sequenceStartPositions->getData(false)[1]); + EXPECT_EQ(0, arguments[i].subSequenceStartPositions->getData(false)[0]); + EXPECT_EQ((int)row_id, + arguments[i].subSequenceStartPositions->getData(false)[1]); + } + fin.close(); + } +} + +int main(int argc, char** argv) { + paddle::initMain(argc, argv); + paddle::initPython(argc, argv); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} + +void checkValue(std::vector& arguments, + picojson::array& arr) { + // CHECK SLOT 0, Sparse Value. + paddle::Argument& sparse_values_seq = arguments[0]; + paddle::MatrixPtr& sparse_values_seq_rawmatrix = sparse_values_seq.value; + EXPECT_TRUE(sparse_values_seq_rawmatrix != nullptr); + paddle::CpuSparseMatrix* sparse_val_seq_sparse_mat = + dynamic_cast(sparse_values_seq_rawmatrix.get()); + EXPECT_TRUE(sparse_val_seq_sparse_mat != nullptr); + EXPECT_EQ(arr.size(), arguments.size()); + EXPECT_TRUE(arr[0].is()); + size_t row_id = 0; + for (picojson::value& sparse_val_seq : arr[0].get()) { + std::unordered_map cols; + for (picojson::value& kv : sparse_val_seq.get()) { + EXPECT_TRUE(kv.get(0).is()); + EXPECT_TRUE(kv.get(1).is()); + int col = (int)(kv.get(0).get()); + real val = (real)(kv.get(1).get()); + cols.insert({col, val}); + } + size_t colNum = sparse_val_seq_sparse_mat->getColNum(row_id); + EXPECT_EQ(cols.size(), colNum); + int* rowIds = sparse_val_seq_sparse_mat->getRowCols(row_id); + real* rowBuf = sparse_val_seq_sparse_mat->getRowValues(row_id); + for (size_t i = 0; i < colNum; ++i) { + int id = rowIds[i]; + auto it = cols.find(id); + EXPECT_NE(cols.end(), it); + real expect = it->second; + EXPECT_NEAR(expect, *rowBuf, 1e-5); + ++rowBuf; + } + ++row_id; + } + + // CHECK SLOT 1, Dense Value. + paddle::Argument& dense_arg = arguments[1]; + paddle::MatrixPtr& dense_mat = dense_arg.value; + EXPECT_NE(nullptr, dense_mat); + EXPECT_TRUE(arr[1].is()); + row_id = 0; + for (picojson::value& dense_seq : arr[1].get()) { + EXPECT_TRUE(dense_seq.is()); + picojson::array& row = dense_seq.get(); + EXPECT_EQ(row.size(), dense_mat->getWidth()); + real* rowBuf = dense_mat->getRowBuf(row_id++); + + for (picojson::value& val : row) { + EXPECT_TRUE(val.is()); + real expect = val.get(); + EXPECT_NEAR(expect, *rowBuf++, 1e-5); + } + } + + // CHECK SLOT 2, Sparse Non Value. + paddle::Argument& sparse_non_val_arg = arguments[2]; + paddle::MatrixPtr& sparse_non_val_rawm = sparse_non_val_arg.value; + EXPECT_NE(nullptr, sparse_non_val_rawm); + paddle::CpuSparseMatrix* sparse_non_val_m = + dynamic_cast(sparse_non_val_rawm.get()); + EXPECT_NE(nullptr, sparse_non_val_m); + row_id = 0; + for (picojson::value& row : arr[2].get()) { + EXPECT_TRUE(row.is()); + std::unordered_set ids; + for (picojson::value& id : row.get()) { + EXPECT_TRUE(id.is()); + ids.insert((int)(id.get())); + } + size_t colNum = sparse_non_val_m->getColNum(row_id); + EXPECT_EQ(ids.size(), colNum); + for (size_t i = 0; i < colNum; ++i) { + int col = sparse_non_val_m->getRowCols(row_id)[i]; + EXPECT_TRUE(ids.find(col) != ids.end()); + } + ++row_id; + } + + // CHECK SLOT 3, Index. + paddle::Argument& index_arg = arguments[3]; + paddle::IVectorPtr indices = index_arg.ids; + EXPECT_NE(nullptr, indices); + int* idPtr = indices->getData(); + for (picojson::value& id : arr[3].get()) { + EXPECT_TRUE(id.is()); + int _id = (int)(id.get()); + EXPECT_EQ(_id, *idPtr++); + } + + // CHECK SLOT 4, String. + paddle::Argument& strArg = arguments[4]; + std::vector* strPtr = strArg.strs.get(); + EXPECT_NE(nullptr, strPtr); + size_t vecIndex = 0; + for (picojson::value& str : arr[4].get()) { + EXPECT_TRUE(str.is()); + std::string _str = str.get(); + EXPECT_EQ(_str, (*strPtr)[vecIndex++]); + } +} + +#else +int main() { return 0; } + +#endif diff --git a/paddle/legacy/trainer/tests/test_Trainer.cpp b/paddle/legacy/trainer/tests/test_Trainer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..14ad0a265281a8df20a70b0da2873ea451338ddb --- /dev/null +++ b/paddle/legacy/trainer/tests/test_Trainer.cpp @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/legacy/trainer/Trainer.h" + +#include + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static const string& configFile1 = + "legacy/trainer/tests/sample_trainer_config.conf"; +static const string& configFile2 = + "legacy/trainer/tests/sample_trainer_config_hsigmoid.conf"; +static const string& configFile4 = + "legacy/trainer/tests/sample_trainer_config_parallel.conf"; + +DECLARE_bool(use_gpu); +DECLARE_string(config); +DECLARE_int32(gpu_id); +DECLARE_bool(allow_only_one_model_on_one_gpu); + +void checkGradientTest(const string& configFile, + bool useGpu, + bool parallel, + int trainerCount = 1) { + FLAGS_use_gpu = useGpu; + FLAGS_parallel_nn = parallel; + FLAGS_config = configFile; + FLAGS_trainer_count = trainerCount; + LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount + << " configFile=" << configFile; + + Trainer trainer; + trainer.init(TrainerConfigHelper::createFromFlagConfig()); + EXPECT_LE(fabs(trainer.checkGradient()), 0.02); +} + +TEST(checkGradient, cpu) { checkGradientTest(configFile1, false, false); } + +#ifdef PADDLE_WITH_CUDA +TEST(checkGradient, gpu) { checkGradientTest(configFile1, true, false); } + +TEST(checkGradient, multiGpu) { + int numGpu; + numGpu = hl_get_device_count(); + for (auto count : {2, 4}) { + if (count <= numGpu) { + checkGradientTest(configFile1, true, false, count); + } + } +} + +TEST(checkGradient, parallel) { + if (hl_get_device_count() >= 2) { + checkGradientTest(configFile4, true, true); + } +} + +TEST(checkGradient, multiParallel) { + FLAGS_allow_only_one_model_on_one_gpu = false; + checkGradientTest(configFile4, true, true, 2); + FLAGS_allow_only_one_model_on_one_gpu = true; +} + +#endif + +TEST(checkGradient, multi) { + int numGpu; + if (version::isWithGpu()) { + numGpu = hl_get_device_count(); + } else { + numGpu = 0; + } + for (bool useGpu : {false, true}) { + for (auto count : {2, 4}) { + if (useGpu && count > numGpu) continue; + checkGradientTest(configFile1, useGpu, false, count); + } + } +} + +TEST(checkGradient, hsigmoid) { checkGradientTest(configFile2, false, false); } + +TEST(checkGradient, non_parallel) { + checkGradientTest(configFile4, false, false); +} + +int main(int argc, char** argv) { + initMain(argc, argv); + initPython(argc, argv); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/trainer/tests/test_TrainerOnePass.cpp b/paddle/legacy/trainer/tests/test_TrainerOnePass.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3e5c5ea723f3fd80316ee826fe9c6566e7049b7b --- /dev/null +++ b/paddle/legacy/trainer/tests/test_TrainerOnePass.cpp @@ -0,0 +1,318 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include "paddle/legacy/trainer/Trainer.h" +#include "paddle/legacy/trainer/TrainerInternal.h" + +#include +#include + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static const string& configFile1 = + "legacy/trainer/tests/sample_trainer_config.conf"; +static const string& configFile2 = + "legacy/trainer/tests/sample_trainer_config_parallel.conf"; + +static const string& configFileSimpleSparse = + "legacy/trainer/tests/simple_sparse_neural_network.py"; + +DECLARE_bool(use_gpu); +DECLARE_string(config); +DECLARE_int32(gpu_id); +DECLARE_int32(seed); +DECLARE_int32(num_passes); +DECLARE_int32(saving_period); + +class TrainerForTest : public paddle::Trainer { + public: + inline const std::shared_ptr& getParameterUpdaterForTest() { + return this->trainerInternal_.getParameterUpdater(); + } +}; + +int gNumDevices = 0; + +void trainerOnePassTest(const string& configFile, + bool useGpu, + bool parallel, + int trainerCount = 1, + double averageWindow = 0.0f, + bool doAverageInCpu = false) { + FLAGS_use_gpu = useGpu; + FLAGS_parallel_nn = parallel; + FLAGS_config = configFile; + FLAGS_trainer_count = trainerCount; + LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount + << " configFile=" << configFile; + srand(FLAGS_seed); + + if (useGpu) { + if (gNumDevices < trainerCount) { + return; + } + } + + Trainer trainer; + auto config = TrainerConfigHelper::createFromFlagConfig(); + if (averageWindow > 0) { + config->getOptConfig().set_average_window(averageWindow); + config->getOptConfig().set_do_average_in_cpu(doAverageInCpu); + } + trainer.init(config); + trainer.train(); +} + +// 1. test trainer (cpu, gpu). +TEST(trainerOnePass, cpu) { trainerOnePassTest(configFile1, false, false); } + +#ifdef PADDLE_WITH_CUDA +TEST(trainerOnePass, gpu) { trainerOnePassTest(configFile1, true, false); } + +TEST(trainerOnePass, gpu2) { trainerOnePassTest(configFile1, true, false, 2); } + +TEST(trainerOnePass, gpu4) { trainerOnePassTest(configFile1, true, false, 4); } + +TEST(trainerOnePass, parallel) { + if (hl_get_device_count() >= 2) { + trainerOnePassTest(configFile2, true, true); + } +} +#endif + +// 2. test average_window. +#ifdef PADDLE_WITH_CUDA +TEST(average_window, gpu) { + trainerOnePassTest(configFile1, true, false, 4, 0.01); +} + +TEST(average_window, gpu2) { + FLAGS_num_passes = 20; + trainerOnePassTest(configFile1, true, false, 2, 0.01); + FLAGS_num_passes = 1; +} + +TEST(average_window, gpu4) { + FLAGS_num_passes = 20; + trainerOnePassTest(configFile1, true, false, 4, 0.01); + FLAGS_num_passes = 1; +} + +TEST(average_window_cpu, gpu2) { + FLAGS_num_passes = 20; + trainerOnePassTest(configFile1, true, false, 2, 0.01, true); + FLAGS_num_passes = 1; +} + +TEST(average_window_cpu, gpu4) { + FLAGS_num_passes = 20; + trainerOnePassTest(configFile1, true, false, 4, 0.01, true); + FLAGS_num_passes = 1; +} +#endif + +// 3. test trainer + pserver. +DECLARE_int32(num_gradient_servers); +DECLARE_int32(port); +DECLARE_bool(local); +DECLARE_bool(use_old_updater); + +double checkRemoteParameterUpdater(TrainerForTest& trainer) { + auto gradientMachine = trainer.getGradientMachine(); + auto parameterUpdater = trainer.getParameterUpdaterForTest(); + auto dataProvider = trainer.getDataProvider(); + auto& parameters = gradientMachine->getParameters(); + const TrainerConfig& config = trainer.getConfig(); + const string& alg = config.opt_config().algorithm(); + + vector parameterCheck; + for (auto& parameter : parameters) { + parameterCheck.emplace_back( + new Parameter(parameter->getConfig(), /* useGpu= */ false)); + parameterCheck.back() + ->getBuf(PARAMETER_VALUE) + ->copyFrom(*parameter->getBuf(PARAMETER_VALUE)); + parameterCheck.back() + ->getBuf(PARAMETER_GRADIENT) + ->copyFrom(*parameter->getBuf(PARAMETER_GRADIENT)); + } + + std::unique_ptr parameterUpdaterCheck; + if (alg == TrainAlgorithm::SGD) { + parameterUpdaterCheck.reset(new SgdLocalUpdater(config.opt_config())); + } else { + LOG(INFO) << "unsupported algorithm in remote parameter check: " << alg; + return -1.0; + } + parameterUpdaterCheck->init(parameterCheck); + + // gradientMachine->start(config, *dataProvider); + DataBatch dataBatch; + int32_t batchSize = config.opt_config().batch_size(); + dataProvider->getNextBatch(batchSize, &dataBatch); + CHECK(dataBatch.getSize()) << "No data from data provider"; + int64_t actualBatchSize = dataBatch.getSize(); + const vector& inArgs = dataBatch.getStreams(); + vector outArgs; + + UpdateCallback updateCallback = [parameterUpdater, + parameterCheck](Parameter* para) { + parameterCheck[para->getID()] + ->getBuf(PARAMETER_GRADIENT) + ->copyFrom(*para->getBuf(PARAMETER_GRADIENT)); + parameterUpdater->update(para); + }; + + parameterUpdater->startPass(); + parameterUpdaterCheck->startPass(); + + for (int i = 0; i < config.opt_config().num_batches_per_get_parameter() * 2; + ++i) { + PassType passType = parameterUpdater->startBatch(actualBatchSize); + gradientMachine->forwardBackward( + inArgs, &outArgs, passType, updateCallback); + parameterUpdater->finishBatch(0); + + parameterUpdaterCheck->startBatch(actualBatchSize); + for (auto& para : parameterCheck) { + parameterUpdaterCheck->update(para.get()); + } + parameterUpdaterCheck->finishBatch(0); + } + + double sum = 0.0f; + for (size_t i = 0; i != parameters.size(); ++i) { + real *v1, *v2; + CpuVector trainerPara(parameters[i]->getSize()); + trainerPara.copyFrom(*parameters[i]->getBuf(PARAMETER_VALUE)); + if (!FLAGS_use_gpu) { + v1 = parameters[i]->getBuf(PARAMETER_VALUE)->getData(); + } else { + v1 = trainerPara.getData(); + } + v2 = parameterCheck[i]->getBuf(PARAMETER_VALUE)->getData(); + + size_t size = parameters[i]->getSize(); + double diff = 0; + for (size_t j = 0; j < size; ++j) { + diff += fabs(v1[j] - v2[j]); + } + sum += diff; + LOG(INFO) << setiosflags(ios::left) << setfill(' ') << setw(20) + << parameters[i]->getName() << "diff=" << setw(15) << diff; + } + + parameterUpdater->finishPass(); + parameterUpdaterCheck->finishPass(); + gradientMachine->finish(); + return sum; +} + +void checkRemoteParameterUpdaterTest(const string& configFile, + bool useGpu, + bool parallel, + int trainerCount = 1, + bool useOldUpdater = false, + int num_batches_per_get_parameter = 1) { + FLAGS_use_gpu = useGpu; + FLAGS_parallel_nn = parallel; + FLAGS_config = configFile; + FLAGS_trainer_count = trainerCount; + FLAGS_use_old_updater = useOldUpdater; + LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount + << " configFile=" << configFile; + srand(FLAGS_seed); + + if (useGpu) { + if (gNumDevices < trainerCount) { + return; + } + } + + FLAGS_local = 0; + std::shared_ptr pserver; + pserver.reset(new ParameterServer2(std::string(), FLAGS_port)); + pserver->init(); + pserver->start(); + + TrainerForTest trainer; + auto config = TrainerConfigHelper::createFromFlagConfig(); + config->getOptConfig().set_num_batches_per_get_parameter( + num_batches_per_get_parameter); + trainer.init(config); + EXPECT_EQ(checkRemoteParameterUpdater(trainer), 0); + + FLAGS_local = 1; +} + +TEST(checkRemoteUpdater, cpuTrainer) { + checkRemoteParameterUpdaterTest(configFile1, false, false); +} + +TEST(checkRemoteUpdater, cpuTrainerOldUpdater) { + checkRemoteParameterUpdaterTest(configFile1, false, false, 1, true); +} + +#ifdef PADDLE_WITH_CUDA +TEST(checkRemoteUpdater, gpuTrainer) { + checkRemoteParameterUpdaterTest(configFile1, true, false); +} + +TEST(checkRemoteUpdater, gpu2Trainer) { + checkRemoteParameterUpdaterTest(configFile1, true, false, 2); +} + +TEST(checkRemoteUpdater, gpu4Trainer) { + checkRemoteParameterUpdaterTest(configFile1, true, false, 4); +} + +TEST(checkRemoteUpdater, gpuTrainerOldUpdater) { + checkRemoteParameterUpdaterTest(configFile1, true, false, 1, true); +} + +TEST(checkRemoteUpdater, gpu2TrainerOldUpdater) { + checkRemoteParameterUpdaterTest(configFile1, true, false, 2, true); +} + +TEST(checkRemoteUpdater, gpu4TrainerOldUpdater) { + checkRemoteParameterUpdaterTest(configFile1, true, false, 4, true); +} + +#endif + +TEST(checkRemoteUpdater, cpuDeltaTrainer) { + checkRemoteParameterUpdaterTest(configFile1, false, false, 1, false, 10); +} + +TEST(checkRemoteUpdater, cpuDeltaTrainerOldUpdater) { + checkRemoteParameterUpdaterTest(configFile1, false, false, 1, true, 10); +} + +TEST(SgdThreadUpdater, simpleSparseNN) { + trainerOnePassTest(configFileSimpleSparse, false, false, 1, 0.5, true); +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + initPython(argc, argv); + gNumDevices = hl_get_device_count(); + + FLAGS_num_passes = 1; // train one pass + FLAGS_saving_period = 100000; // do not save parameteres + return RUN_ALL_TESTS(); +} diff --git a/paddle/legacy/trainer/tests/test_config.conf b/paddle/legacy/trainer/tests/test_config.conf new file mode 100644 index 0000000000000000000000000000000000000000..bce687ad83686d465987d72defd37db2b50953a1 --- /dev/null +++ b/paddle/legacy/trainer/tests/test_config.conf @@ -0,0 +1,77 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from paddle.trainer_config_helpers import * + +TrainData(SimpleData( + files = "legacy/trainer/tests/sample_filelist.txt", + feat_dim = 3, + context_len = 0, + buffer_capacity = 1000000, + async_load_data = False)) + +settings(batch_size = 100) + +data = data_layer(name='input', size=3) + +wt = data_layer(name='weight', size=1) + +fc1 = fc_layer(input=data, size=5, + bias_attr=True, + act=SigmoidActivation()) + +fc2 = fc_layer(input=data, size=12, + bias_attr=True, + param_attr=ParamAttr(name='sharew'), + act=LinearActivation()) + +fc3 = fc_layer(input=data, size=3, + bias_attr=True, + act=TanhActivation()) + +fc4 = fc_layer(input=data, size=5, + bias_attr=True, + layer_attr=ExtraAttr(drop_rate=0.5), + act=SquareActivation()) + +pool = img_pool_layer(input=fc2, + pool_size=2, + pool_size_y=3, + num_channels=1, + padding=1, + padding_y=2, + stride=2, + stride_y=3, + pool_type=CudnnAvgPooling()) + +concat = concat_layer(input=[fc3, fc4]) + +with mixed_layer(size=3, act=SoftmaxActivation()) as output: + output += full_matrix_projection(input=fc1) + output += trans_full_matrix_projection(input=fc2, + param_attr=ParamAttr(name='sharew')) + output += full_matrix_projection(input=concat) + output += identity_projection(input=fc3) + +lbl = data_layer(name='label', size=1) + +cost = classification_cost(input=output, label=lbl, weight=wt, + layer_attr=ExtraAttr(device=-1)) + +nce = nce_layer(input=fc2, label=lbl, weight=wt, + num_classes=3, + neg_distribution=[0.1, 0.3, 0.6]) + +outputs(cost, nce) diff --git a/paddle/trainer/tests/test_gen_dict.txt b/paddle/legacy/trainer/tests/test_gen_dict.txt similarity index 100% rename from paddle/trainer/tests/test_gen_dict.txt rename to paddle/legacy/trainer/tests/test_gen_dict.txt diff --git a/paddle/legacy/trainer/tests/test_recurrent_machine_generation.cpp b/paddle/legacy/trainer/tests/test_recurrent_machine_generation.cpp new file mode 100644 index 0000000000000000000000000000000000000000..47b4e82cd32917fcf32dbb5ffabb47330dab93d9 --- /dev/null +++ b/paddle/legacy/trainer/tests/test_recurrent_machine_generation.cpp @@ -0,0 +1,157 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include +#include + +#include + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +static const string& CONFIG_FILE = + "legacy/trainer/tests/sample_trainer_rnn_gen.conf"; +static const string& NEST_CONFIG_FILE = + "legacy/trainer/tests/sample_trainer_nest_rnn_gen.conf"; +static const string& OUTPUT_DIR = "legacy/trainer/tests/dump_text.test"; +static string modelDir = + "legacy/trainer/tests/rnn_gen_test_model_dir/t1"; // NOLINT +static string expectFile = // NOLINT + "legacy/trainer/tests/rnn_gen_test_model_dir/r1.test"; // NOLINT + +DECLARE_string(config_args); + +vector readRetFile(const string& fname) { + ifstream inFile(fname); + float ret; + vector nums; + while (inFile >> ret) { + nums.push_back(ret); + } + return nums; +} + +void checkOutput(const string& expRetFile) { + vector rets = readRetFile(OUTPUT_DIR); + vector expRets = readRetFile(expRetFile); + EXPECT_EQ(rets.size(), expRets.size()); + for (size_t i = 0; i < rets.size(); i++) { + EXPECT_FLOAT_EQ(rets[i], expRets[i]); + } +} + +void prepareInArgs(vector& inArgs, + const size_t batchSize, + bool useGpu, + bool hasSubseq) { + inArgs.clear(); + // sentence id + Argument sentId; + sentId.value = nullptr; + if (hasSubseq) { + // as there is only one sequence, there is only one label. + IVector::resizeOrCreate(sentId.ids, 1, useGpu); + sentId.ids->setElement(0, 0); + } else { + // as there is batchSize word, there is batchSize label. + IVector::resizeOrCreate(sentId.ids, batchSize, useGpu); + for (size_t i = 0; i < batchSize; ++i) sentId.ids->setElement(i, i); + } + inArgs.emplace_back(sentId); + + // a dummy layer to decide batch size + Argument dummyInput; + dummyInput.value = Matrix::create(batchSize, 2, false, useGpu); + dummyInput.value->randomizeUniform(); + if (hasSubseq) { + // generate one sequence with batchSize subsequence, + // and each subsequence has only one word. + dummyInput.sequenceStartPositions = ICpuGpuVector::create(2, false); + int* buf = dummyInput.sequenceStartPositions->getMutableData(false); + dummyInput.subSequenceStartPositions = + ICpuGpuVector::create(batchSize + 1, false); + int* subBuf = dummyInput.subSequenceStartPositions->getMutableData(false); + buf[0] = 0; + buf[1] = batchSize; + for (size_t i = 0; i < batchSize + 1; i++) subBuf[i] = i; + } + inArgs.emplace_back(dummyInput); +} + +void testGeneration(const string& configFile, + bool useGpu, + bool hasSubseq, + const string& expRetFile) { + FLAGS_use_gpu = useGpu; + auto config = std::make_shared(configFile); + unique_ptr gradientMachine(GradientMachine::create(*config)); + gradientMachine->loadParameters(modelDir); + vector inArgs(2); + + const size_t batchSize = 15; + prepareInArgs(inArgs, batchSize, useGpu, hasSubseq); + vector outArgs; + unique_ptr testEvaluator(gradientMachine->makeEvaluator()); + testEvaluator->start(); + gradientMachine->forward(inArgs, &outArgs, PASS_TEST); + gradientMachine->eval(testEvaluator.get()); + testEvaluator->finish(); + checkOutput(expRetFile); +} + +#ifndef PADDLE_TYPE_DOUBLE + +TEST(RecurrentGradientMachine, test_generation) { +#ifndef PADDLE_WITH_CUDA + const auto useGpuConfs = {false}; +#else + const auto useGpuConfs = {true, false}; +#endif + auto testGen = [&](const string& configFile, + bool hasSubseq, + const string& expRetFile, + bool beam_search) { + FLAGS_config_args = beam_search ? "beam_search=1" : "beam_search=0"; + for (auto useGpu : useGpuConfs) { + LOG(INFO) << configFile << " useGpu=" << useGpu + << " beam_search=" << beam_search; + testGeneration(configFile, useGpu, hasSubseq, expRetFile); + } + }; + testGen(CONFIG_FILE, false, expectFile + ".nobeam", false); // no beam search + testGen(CONFIG_FILE, false, expectFile + ".beam", true); // beam search + // In hierarchical RNN, beam search and one way search are only in inner-RNN, + // outer-RNN will concat the generated inner-results (first for beam search) + // from inner-RNN. Thus, they have the same outer-results. + testGen(NEST_CONFIG_FILE, + true, + expectFile + ".nest", + false); // no beam search + testGen(NEST_CONFIG_FILE, true, expectFile + ".nest", true); // beam search +} +#endif + +int main(int argc, char** argv) { + initMain(argc, argv); + initPython(argc, argv); + CHECK(argc == 1 || argc == 3); + if (argc == 3) { + modelDir = argv[1]; + expectFile = argv[2]; + } + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/utils/.gitignore b/paddle/legacy/utils/.gitignore similarity index 100% rename from paddle/utils/.gitignore rename to paddle/legacy/utils/.gitignore diff --git a/paddle/utils/Any.h b/paddle/legacy/utils/Any.h similarity index 100% rename from paddle/utils/Any.h rename to paddle/legacy/utils/Any.h diff --git a/paddle/utils/CMakeLists.txt b/paddle/legacy/utils/CMakeLists.txt similarity index 100% rename from paddle/utils/CMakeLists.txt rename to paddle/legacy/utils/CMakeLists.txt diff --git a/paddle/utils/ClassRegistrar.h b/paddle/legacy/utils/ClassRegistrar.h similarity index 100% rename from paddle/utils/ClassRegistrar.h rename to paddle/legacy/utils/ClassRegistrar.h diff --git a/paddle/utils/Common.h b/paddle/legacy/utils/Common.h similarity index 100% rename from paddle/utils/Common.h rename to paddle/legacy/utils/Common.h diff --git a/paddle/legacy/utils/CpuId.cpp b/paddle/legacy/utils/CpuId.cpp new file mode 100644 index 0000000000000000000000000000000000000000..66e7c6606f070aef4fd954b8f4ada994b2f4fb96 --- /dev/null +++ b/paddle/legacy/utils/CpuId.cpp @@ -0,0 +1,66 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/CpuId.h" +#include "paddle/legacy/utils/Util.h" + +#ifdef _WIN32 + +#include + +/// for MSVC +#define CPUID(info, x) __cpuidex(info, x, 0) + +#else + +#if !defined(__arm__) && !defined(__aarch64__) +#include +/// for GCC/Clang +#define CPUID(info, x) __cpuid_count(x, 0, info[0], info[1], info[2], info[3]) +#endif + +#endif + +namespace paddle { + +SIMDFlags::SIMDFlags() { +#if defined(__arm__) || defined(__aarch64__) + simd_flags_ = SIMD_NEON; +#else + unsigned int cpuInfo[4]; + // CPUID: https://en.wikipedia.org/wiki/CPUID + // clang-format off + CPUID(cpuInfo, 0x00000001); + simd_flags_ |= cpuInfo[3] & (1 << 25) ? SIMD_SSE : SIMD_NONE; + simd_flags_ |= cpuInfo[3] & (1 << 26) ? SIMD_SSE2 : SIMD_NONE; + simd_flags_ |= cpuInfo[2] & (1 << 0) ? SIMD_SSE3 : SIMD_NONE; + simd_flags_ |= cpuInfo[2] & (1 << 9) ? SIMD_SSSE3 : SIMD_NONE; + simd_flags_ |= cpuInfo[2] & (1 << 19) ? SIMD_SSE41 : SIMD_NONE; + simd_flags_ |= cpuInfo[2] & (1 << 20) ? SIMD_SSE42 : SIMD_NONE; + simd_flags_ |= cpuInfo[2] & (1 << 12) ? SIMD_FMA3 : SIMD_NONE; + simd_flags_ |= cpuInfo[2] & (1 << 28) ? SIMD_AVX : SIMD_NONE; + + CPUID(cpuInfo, 0x00000007); + simd_flags_ |= cpuInfo[1] & (1 << 5) ? SIMD_AVX2 : SIMD_NONE; + simd_flags_ |= cpuInfo[1] & (1 << 16) ? SIMD_AVX512: SIMD_NONE; + + CPUID(cpuInfo, 0x80000001); + simd_flags_ |= cpuInfo[2] & (1 << 16) ? SIMD_FMA4 : SIMD_NONE; + // clang-fotmat on +#endif +} + +SIMDFlags const* SIMDFlags::instance() { + static SIMDFlags instance; + return &instance; +} + +} // namespace paddle diff --git a/paddle/utils/CpuId.h b/paddle/legacy/utils/CpuId.h similarity index 100% rename from paddle/utils/CpuId.h rename to paddle/legacy/utils/CpuId.h diff --git a/paddle/utils/CustomStackTrace.cpp b/paddle/legacy/utils/CustomStackTrace.cpp similarity index 100% rename from paddle/utils/CustomStackTrace.cpp rename to paddle/legacy/utils/CustomStackTrace.cpp diff --git a/paddle/utils/CustomStackTrace.h b/paddle/legacy/utils/CustomStackTrace.h similarity index 100% rename from paddle/utils/CustomStackTrace.h rename to paddle/legacy/utils/CustomStackTrace.h diff --git a/paddle/utils/DynamicLoader.cpp b/paddle/legacy/utils/DynamicLoader.cpp similarity index 100% rename from paddle/utils/DynamicLoader.cpp rename to paddle/legacy/utils/DynamicLoader.cpp diff --git a/paddle/utils/DynamicLoader.h b/paddle/legacy/utils/DynamicLoader.h similarity index 100% rename from paddle/utils/DynamicLoader.h rename to paddle/legacy/utils/DynamicLoader.h diff --git a/paddle/utils/Error.h b/paddle/legacy/utils/Error.h similarity index 100% rename from paddle/utils/Error.h rename to paddle/legacy/utils/Error.h diff --git a/paddle/utils/Excepts.h b/paddle/legacy/utils/Excepts.h similarity index 100% rename from paddle/utils/Excepts.h rename to paddle/legacy/utils/Excepts.h diff --git a/paddle/utils/Flags.cpp b/paddle/legacy/utils/Flags.cpp similarity index 100% rename from paddle/utils/Flags.cpp rename to paddle/legacy/utils/Flags.cpp diff --git a/paddle/utils/Flags.h b/paddle/legacy/utils/Flags.h similarity index 100% rename from paddle/utils/Flags.h rename to paddle/legacy/utils/Flags.h diff --git a/paddle/utils/GlobalConstants.cpp b/paddle/legacy/utils/GlobalConstants.cpp similarity index 100% rename from paddle/utils/GlobalConstants.cpp rename to paddle/legacy/utils/GlobalConstants.cpp diff --git a/paddle/utils/GlobalConstants.h b/paddle/legacy/utils/GlobalConstants.h similarity index 100% rename from paddle/utils/GlobalConstants.h rename to paddle/legacy/utils/GlobalConstants.h diff --git a/paddle/utils/Locks.h b/paddle/legacy/utils/Locks.h similarity index 100% rename from paddle/utils/Locks.h rename to paddle/legacy/utils/Locks.h diff --git a/paddle/utils/Logging.cpp b/paddle/legacy/utils/Logging.cpp similarity index 100% rename from paddle/utils/Logging.cpp rename to paddle/legacy/utils/Logging.cpp diff --git a/paddle/utils/Logging.h b/paddle/legacy/utils/Logging.h similarity index 100% rename from paddle/utils/Logging.h rename to paddle/legacy/utils/Logging.h diff --git a/paddle/legacy/utils/PythonUtil.cpp b/paddle/legacy/utils/PythonUtil.cpp new file mode 100644 index 0000000000000000000000000000000000000000..21ed049c4d2743d1fa914d6948d6c8c2862f0bfc --- /dev/null +++ b/paddle/legacy/utils/PythonUtil.cpp @@ -0,0 +1,215 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "PythonUtil.h" +#include +#include + +namespace paddle { + +#ifdef PADDLE_NO_PYTHON + +DEFINE_string(python_path, "", "python path"); +DEFINE_string(python_bin, "python2.7", "python bin"); + +constexpr int kExecuteCMDBufLength = 204800; + +int executeCMD(const char* cmd, char* result) { + char bufPs[kExecuteCMDBufLength]; + char ps[kExecuteCMDBufLength] = {0}; + FILE* ptr; + strncpy(ps, cmd, kExecuteCMDBufLength); + if ((ptr = popen(ps, "r")) != NULL) { + size_t count = fread(bufPs, 1, kExecuteCMDBufLength, ptr); + memcpy(result, + bufPs, + count - 1); // why count-1: remove the '\n' at the end + result[count] = 0; + pclose(ptr); + ptr = NULL; + return count - 1; + } else { + LOG(FATAL) << "popen failed"; + return -1; + } +} + +std::string callPythonFunc(const std::string& moduleName, + const std::string& funcName, + const std::vector& args) { + std::string pythonLibPath = ""; + std::string pythonBinPath = ""; + if (!FLAGS_python_path.empty()) { + pythonLibPath = FLAGS_python_path + "/lib:"; + pythonBinPath = FLAGS_python_path + "/bin/"; + } + std::string s = "LD_LIBRARY_PATH=" + pythonLibPath + "$LD_LIBRARY_PATH " + + pythonBinPath + std::string(FLAGS_python_bin) + + " -c 'import " + moduleName + "\n" + "print " + moduleName + + "." + funcName + "("; + for (auto& arg : args) { + s = s + "\"" + arg + "\", "; + } + s += ")'"; + char result[kExecuteCMDBufLength] = {0}; + LOG(INFO) << " cmd string: " << s; + int length = executeCMD(s.c_str(), result); + CHECK_NE(-1, length); + return std::string(result, length); +} + +#else + +static std::recursive_mutex g_pyMutex; + +PyGuard::PyGuard() : guard_(g_pyMutex) {} + +static void printPyErrorStack(std::ostream& os, + bool withEndl = false, + bool withPyPath = true) { + PyObject *ptype, *pvalue, *ptraceback; + PyErr_Fetch(&ptype, &pvalue, &ptraceback); + PyErr_NormalizeException(&ptype, &pvalue, &ptraceback); + PyErr_Clear(); + if (withPyPath) { + os << "Current PYTHONPATH: " << py::repr(PySys_GetObject(strdup("path"))); + if (withEndl) { + os << std::endl; + } + } + PyTracebackObject* obj = (PyTracebackObject*)ptraceback; + + os << "Python Error: " << PyString_AsString(PyObject_Str(ptype)) << " : " + << (pvalue == NULL ? "" : PyString_AsString(PyObject_Str(pvalue))); + if (withEndl) { + os << std::endl; + } + os << "Python Callstack: "; + if (withEndl) { + os << std::endl; + } + while (obj != NULL) { + int line = obj->tb_lineno; + const char* filename = + PyString_AsString(obj->tb_frame->f_code->co_filename); + os << " " << filename << " : " << line; + if (withEndl) { + os << std::endl; + } + obj = obj->tb_next; + } + + Py_XDECREF(ptype); + Py_XDECREF(pvalue); + Py_XDECREF(ptraceback); +} +PyObjectPtr callPythonFuncRetPyObj(const std::string& moduleName, + const std::string& funcName, + const std::vector& args) { + PyGuard guard; + PyObjectPtr pyModule = py::import(moduleName); + PyObjectPtr pyFunc(PyObject_GetAttrString(pyModule.get(), funcName.c_str())); + CHECK_PY(pyFunc) << "GetAttrString failed."; + PyObjectPtr pyArgs(PyTuple_New(args.size())); + for (size_t i = 0; i < args.size(); ++i) { + PyObjectPtr pyArg(PyString_FromString(args[i].c_str())); + CHECK_PY(pyArg) << "Import pyArg failed."; + PyTuple_SetItem(pyArgs.get(), i, pyArg.release()); // Maybe a problem + } + PyObjectPtr ret(PyObject_CallObject(pyFunc.get(), pyArgs.get())); + CHECK_PY(ret) << "Call Object failed."; + return ret; +} + +std::string callPythonFunc(const std::string& moduleName, + const std::string& funcName, + const std::vector& args) { + PyObjectPtr obj = callPythonFuncRetPyObj(moduleName, funcName, args); +#if PY_MAJOR_VERSION >= 3 + Py_ssize_t str_size = 0u; + const char* str = PyUnicode_AsUTF8AndSize(obj.get(), &str_size); + return std::string(str, (size_t)str_size); +#else + return std::string(PyString_AsString(obj.get()), PyString_Size(obj.get())); +#endif // PY_MAJOR_VERSION >= 3 +} + +PyObjectPtr createPythonClass( + const std::string& moduleName, + const std::string& className, + const std::vector& args, + const std::map& kwargs) { + PyGuard guard; + PyObjectPtr pyModule = py::import(moduleName); + LOG(INFO) << "createPythonClass moduleName.c_str:" << moduleName.c_str(); + CHECK_PY(pyModule) << "Import module " << moduleName << " failed."; + PyObjectPtr pyDict(PyModule_GetDict(pyModule.get())); + CHECK_PY(pyDict) << "Get Dict failed."; + PyObjectPtr pyClass(PyDict_GetItemString(pyDict.get(), className.c_str())); + LOG(INFO) << "createPythonClass className.c_str():" << className.c_str(); + CHECK_PY(pyClass) << "Import class " << className << " failed."; + PyObjectPtr argsObjectList(PyTuple_New(args.size())); + for (size_t i = 0; i < args.size(); ++i) { + PyObjectPtr pyArg(Py_BuildValue("s#", args[i].c_str(), args[i].length())); + PyTuple_SetItem(argsObjectList.get(), i, pyArg.release()); + } + + PyObjectPtr kwargsObjectList(PyDict_New()); + for (auto& x : kwargs) { + PyObjectPtr pyArg(Py_BuildValue("s#", x.second.c_str(), x.second.length())); + PyDict_SetItemString( + kwargsObjectList.get(), x.first.c_str(), pyArg.release()); + } + + PyObjectPtr pyInstance(PyInstance_New( + pyClass.get(), argsObjectList.release(), kwargsObjectList.release())); + CHECK_PY(pyInstance) << "Create class " << className << " failed."; + return pyInstance; +} + +namespace py { +char* repr(PyObject* obj) { return PyString_AsString(PyObject_Repr(obj)); } + +std::string getPyCallStack() { + std::ostringstream os; + printPyErrorStack(os, true); + return os.str(); +} + +PyObjectPtr import(const std::string& moduleName) { + auto module = PyImport_ImportModule(moduleName.c_str()); + CHECK_PY(module) << "Import " << moduleName << "Error"; + return PyObjectPtr(module); +} + +} // namespace py + +#endif +extern "C" { +extern const char enable_virtualenv_py[]; +} +void initPython(int argc, char** argv) { +#ifndef PADDLE_NO_PYTHON + Py_SetProgramName(argv[0]); + Py_Initialize(); + PySys_SetArgv(argc, argv); + // python blocks SIGINT. Need to enable it. + signal(SIGINT, SIG_DFL); + + // Manually activate virtualenv when user is using virtualenv + PyRun_SimpleString(enable_virtualenv_py); +#endif +} + +} // namespace paddle diff --git a/paddle/legacy/utils/PythonUtil.h b/paddle/legacy/utils/PythonUtil.h new file mode 100644 index 0000000000000000000000000000000000000000..d5b2dbddde21f5c2a0696aadeda2b057175fc5e9 --- /dev/null +++ b/paddle/legacy/utils/PythonUtil.h @@ -0,0 +1,381 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +// clang-format off +#include "paddle/legacy/utils/Util.h" + +#ifndef PADDLE_NO_PYTHON +// must include the following two blocks, otherwise, +// gcc compiler may produce warning +#ifdef __APPLE__ +#define _POSIX_SOURCE +#define _POSIX_C_SOURCE 200809L +#define _XOPEN_SOURCE 700 +#endif + +#ifdef _POSIX_C_SOURCE +#define __TEMP_POSIX_C_SOURCE _POSIX_C_SOURCE +#undef _POSIX_C_SOURCE +#endif +#ifdef _XOPEN_SOURCE +#define __TEMP_XOPEN_SOURCE _XOPEN_SOURCE +#undef _XOPEN_SOURCE +#endif +#include +#include +#endif + +#include +#include +#include +// clang-format on + +namespace paddle { + +std::string callPythonFunc(const std::string& moduleName, + const std::string& funcName, + const std::vector& args); + +#ifndef PADDLE_NO_PYTHON + +/** + * Global lock guard of python C-api invokes. + * NOTE: the lock of this guard is reentrant or recursive. + */ +class PyGuard { + public: + PyGuard(); + PyGuard(const PyGuard& other) = delete; + PyGuard& operator=(const PyGuard& other) = delete; + + private: + std::lock_guard guard_; +}; + +struct PyObjectDeleter { + void operator()(PyObject* obj) { + if (obj) { + Py_DECREF(obj); + } + } +}; + +typedef std::unique_ptr PyObjectPtr; + +PyObjectPtr callPythonFuncRetPyObj(const std::string& moduleName, + const std::string& funcName, + const std::vector& args); + +PyObjectPtr createPythonClass(const std::string& moduleName, + const std::string& className, + const std::vector& args, + const std::map& kwargs); + +#define CHECK_PY(x) CHECK((x) != nullptr) << ::paddle::py::getPyCallStack() + +namespace py { +PyObjectPtr import(const std::string& moduleName); + +#if PY_MAJOR_VERSION >= 3 +/** + * Cast a PyLong to int type T. + * @tparam T return type. + * @param [in] obj PyLong object. + * @param [out] ok status for casting. False if error occured. nullptr if user + * don't care is ok or not. + * @return The value of python object, or 0 if not ok. + */ +template +T castInt(PyObject* obj, bool* ok = nullptr) { + // Refer to https://www.python.org/dev/peps/pep-0237/, the int and long object + // were unified to long since python3 + if (PyLong_Check(obj)) { + if (ok) *ok = true; + return (T)PyLong_AsUnsignedLong(obj); + } else { + if (ok) *ok = false; + return (T)0; + } +} + +// Convert PyAPI from 2.x to 3.x +#define PyString_FromString PyUnicode_FromString +#define PyString_AsString PyUnicode_AsUTF8 + +#else +/** + * Cast a PyLong or PyInt to int type T. + * @tparam T return type. + * @param [in] obj PyLong or PyInt object. + * @param [out] ok status for casting. False if error occured. nullptr if user + * don't care is ok or not. + * @return The value of python object, or 0 if not ok. + */ +template +T castInt(PyObject* obj, bool* ok = nullptr) { + if (PyLong_Check(obj)) { + if (ok) *ok = true; + return (T)PyLong_AsUnsignedLong(obj); + } else if (PyInt_Check(obj)) { + if (ok) *ok = true; + return (T)PyInt_AsLong(obj); + } else { + if (ok) *ok = false; + return (T)0; + } +} +#endif // PY_MAJOR_VERSION >= 3 + +/** + * Invoke repr of python object. + * + * Just like toString method in java. + */ +char* repr(PyObject* obj); + +/** + * Invoke repr of python object. + */ +inline char* repr(const PyObjectPtr& obj) { return repr(obj.get()); } + +/** + * Get Python Error Stack String. + */ +std::string getPyCallStack(); + +/** + * Object Helper for PyObjectPtr. + * + * Implements getAttr method for object. + */ +class ObjectHelper { + public: + explicit ObjectHelper(const PyObjectPtr& obj) : obj_(obj) {} + + /** + * get attribute + */ + inline PyObject* getAttr(const std::string& field) const { + auto obj = PyObject_GetAttrString(obj_.get(), field.c_str()); + CHECK_PY(obj) << "Cannot get attribute on python object " << obj_.get(); + return obj; + } + + /** + * Get Int attribute + * @param [in] field attribute name. + * @param [out] ok true if this attribute is int. + * @tparam T int type. + * @return int value. + */ + template + T getIntAttr(const std::string& field, bool* ok = nullptr) const { + PyObjectPtr tmp(getAttr(field)); + return castInt(tmp.get(), ok); + } + + /** + * Get int attribute. Log(Fatal) when not ok + * @param field attribute name. + * @return int value. + */ + template + T getIntAttrWithError(const std::string& field) const { + bool ok; + T tmp = getIntAttr(field, &ok); + CHECK(ok) << "Cannot get integer attribute on object " << obj_.get(); + return tmp; + } + + /** + * Get bool attribute. + * @param field + * @param [out] isBoolType return true if attribute is bool type. If the + * attribute is not bool type, then an implicit + * conversion will happens, and will return the + * conversion result. + * + * Such as, if the attribute is 1, then the return + * value of function will be true, but the isBoolType + * will return false. + * @return + */ + bool getBoolAttr(const std::string& field, bool* isBoolType = nullptr) const { + PyObjectPtr tmp(getAttr(field)); + if (isBoolType) { + *isBoolType = PyBool_Check(tmp.get()); + } + return PyObject_IsTrue(tmp.get()); + } + + private: + const PyObjectPtr& obj_; +}; + +/** + * Python Sequence Helper + * + * The python sequence means list or tuple. + */ +class SequenceHelper { + public: + explicit SequenceHelper(const PyObjectPtr& seq) : seq_(seq.get()) { + CHECK(PySequence_Check(seq_)); + } + + explicit SequenceHelper(PyObject* seq) : seq_(seq) { + CHECK(PySequence_Check(seq_)); + } + + inline size_t size() const { return (size_t)PySequence_Size(seq_); } + + inline PyObject* operator[](size_t i) const { + return PySequence_Fast_GET_ITEM(seq_, i); + } + + inline double getDouble(size_t i) const { + auto* ptr = (*this)[i]; + return PyFloat_AsDouble(ptr); + } + + /** + * Set a sequence item o[i] = obj; + * @param i index + * @param obj setted item. + * @param steal if steal = true, sequence will move object in iteself, + * just like std::move. Otherwise, it will increase reference + * count. Default is false. + */ + inline void set(size_t i, const PyObjectPtr& obj, bool steal = false) { + this->set(i, obj.get(), steal); + } + + /** + * Set a sequence item o[i] = obj; + */ + inline void set(size_t i, PyObject* obj, bool steal = false) { + if (!steal) { + Py_XINCREF(obj); + } + if (PyTuple_Check(seq_)) { + CHECK_NE(PyTuple_SetItem(seq_, i, obj), -1) << getPyCallStack(); + } else { + CHECK_NE(PySequence_SetItem(seq_, i, obj), -1) << getPyCallStack(); + } + } + + private: + PyObject* seq_; +}; + +class DictHelper { + public: + explicit DictHelper(PyObject* d) : dict_(d) {} + + explicit DictHelper(const PyObjectPtr& d) : dict_(d.get()) {} + + void set(const std::string& key, PyObject* item) { + PyDict_SetItemString(dict_, key.c_str(), item); + } + + void setBool(const std::string& key, bool b) { + this->set(key, PyBool_FromLong(b)); + } + + void setStringList(const std::string& key, + const std::vector& items) { + auto* list = PyList_New(items.size()); + for (size_t i = 0; i < items.size(); ++i) { + PyList_SetItem(list, i, PyString_FromString(items[i].c_str())); + } + this->set(key, list); + } + + private: + inline void checkDict() { CHECK(PyDict_Check(this->dict_)); } + + PyObject* dict_; +}; + +inline static bool isCallable(const PyObjectPtr& obj) { + return PyCallable_Check(obj.get()); +} + +/** + * Wrap a callable object. + */ +class CallableHelper { + public: + explicit CallableHelper(const PyObjectPtr& obj) : obj_(obj) { + CHECK(py::isCallable(obj_)); + } + + ~CallableHelper() {} + + /** + * reset args, and create new tuple. + * @param sz args size. + */ + void setArgsSize(size_t sz) { args.reset(PyTuple_New(sz)); } + + /** + * Get args sequence. User can set/get by SequenceHelper. + */ + SequenceHelper getArgs() { return SequenceHelper(args); } + + /** + * Call python method, return an object. + */ + PyObject* operator()() { + PyGuard guard; + return PyObject_Call(obj_.get(), args.get(), kwargs.get()); + } + + private: + const PyObjectPtr& obj_; + PyObjectPtr args; + PyObjectPtr kwargs; +}; + +inline static PyObject* iterNext(const PyObjectPtr& context, bool* atEnd) { + PyGuard g; + PyObject* data = PyIter_Next(context.get()); + if (data == nullptr) { + if (PyErr_ExceptionMatches(PyExc_StopIteration)) { + PyErr_Clear(); + *atEnd = true; + return nullptr; + } else if (PyErr_Occurred()) { + CHECK_PY(data) << "Calling iterator next error"; + return nullptr; + } else { + *atEnd = false; + return data; // just return none in iterator. + } + } else { + *atEnd = false; + return data; + } +} +} // namespace py + +#endif + +/** + * Initialize python. + */ +void initPython(int argc, char** argv); + +} // namespace paddle diff --git a/paddle/utils/Queue.h b/paddle/legacy/utils/Queue.h similarity index 100% rename from paddle/utils/Queue.h rename to paddle/legacy/utils/Queue.h diff --git a/paddle/utils/Stat.cpp b/paddle/legacy/utils/Stat.cpp similarity index 100% rename from paddle/utils/Stat.cpp rename to paddle/legacy/utils/Stat.cpp diff --git a/paddle/utils/Stat.h b/paddle/legacy/utils/Stat.h similarity index 100% rename from paddle/utils/Stat.h rename to paddle/legacy/utils/Stat.h diff --git a/paddle/utils/StringUtil.cpp b/paddle/legacy/utils/StringUtil.cpp similarity index 100% rename from paddle/utils/StringUtil.cpp rename to paddle/legacy/utils/StringUtil.cpp diff --git a/paddle/utils/StringUtil.h b/paddle/legacy/utils/StringUtil.h similarity index 100% rename from paddle/utils/StringUtil.h rename to paddle/legacy/utils/StringUtil.h diff --git a/paddle/utils/Thread.h b/paddle/legacy/utils/Thread.h similarity index 100% rename from paddle/utils/Thread.h rename to paddle/legacy/utils/Thread.h diff --git a/paddle/utils/ThreadLocal.cpp b/paddle/legacy/utils/ThreadLocal.cpp similarity index 100% rename from paddle/utils/ThreadLocal.cpp rename to paddle/legacy/utils/ThreadLocal.cpp diff --git a/paddle/utils/ThreadLocal.h b/paddle/legacy/utils/ThreadLocal.h similarity index 100% rename from paddle/utils/ThreadLocal.h rename to paddle/legacy/utils/ThreadLocal.h diff --git a/paddle/utils/Util.cpp b/paddle/legacy/utils/Util.cpp similarity index 100% rename from paddle/utils/Util.cpp rename to paddle/legacy/utils/Util.cpp diff --git a/paddle/utils/Util.h b/paddle/legacy/utils/Util.h similarity index 100% rename from paddle/utils/Util.h rename to paddle/legacy/utils/Util.h diff --git a/paddle/utils/Version.cpp b/paddle/legacy/utils/Version.cpp similarity index 100% rename from paddle/utils/Version.cpp rename to paddle/legacy/utils/Version.cpp diff --git a/paddle/utils/Version.h b/paddle/legacy/utils/Version.h similarity index 100% rename from paddle/utils/Version.h rename to paddle/legacy/utils/Version.h diff --git a/paddle/legacy/utils/arch/linux/Locks.cpp b/paddle/legacy/utils/arch/linux/Locks.cpp new file mode 100644 index 0000000000000000000000000000000000000000..32d351e3328afd79007aea7a51e59cbfc941eeeb --- /dev/null +++ b/paddle/legacy/utils/arch/linux/Locks.cpp @@ -0,0 +1,149 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Locks.h" +#include +#include +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { +class SemaphorePrivate { + public: + sem_t sem; +}; + +Semaphore::Semaphore(int initValue) : m(new SemaphorePrivate()) { + sem_init(&m->sem, 0, initValue); +} + +Semaphore::~Semaphore() { + sem_destroy(&m->sem); + delete m; +} + +bool Semaphore::timeWait(struct timespec* ts) { + return (0 == sem_timedwait(&m->sem, ts)); +} + +void Semaphore::wait() { sem_wait(&m->sem); } + +void Semaphore::post() { sem_post(&m->sem); } + +/// SpinLockPrivate + +#ifdef PADDLE_USE_PTHREAD_SPINLOCK + +class SpinLockPrivate { + public: + inline SpinLockPrivate() { pthread_spin_init(&lock_, 0); } + inline ~SpinLockPrivate() { pthread_spin_destroy(&lock_); } + + inline void lock() { pthread_spin_lock(&lock_); } + inline void unlock() { pthread_spin_unlock(&lock_); } + + pthread_spinlock_t lock_; + char padding_[64 - sizeof(pthread_spinlock_t)]; +}; + +#else +// clang-format off +#include +#include +// clang-format on + +class SpinLockPrivate { + public: + inline void lock() { + while (lock_.test_and_set(std::memory_order_acquire)) { + } + } + inline void unlock() { lock_.clear(std::memory_order_release); } + + std::atomic_flag lock_ = ATOMIC_FLAG_INIT; + char padding_[64 - sizeof(lock_)]; // Padding to cache line size +}; + +#endif + +SpinLock::SpinLock() : m(new SpinLockPrivate()) {} +SpinLock::~SpinLock() { delete m; } +void SpinLock::lock() { m->lock(); } +void SpinLock::unlock() { m->unlock(); } + +/// ThreadBarrierPrivate + +#ifdef PADDLE_USE_PTHREAD_BARRIER + +class ThreadBarrierPrivate { + public: + pthread_barrier_t barrier_; + + inline explicit ThreadBarrierPrivate(int count) { + pthread_barrier_init(&barrier_, nullptr, count); + } + + inline ~ThreadBarrierPrivate() { pthread_barrier_destroy(&barrier_); } + + inline void wait() { pthread_barrier_wait(&barrier_); } +}; + +#else + +class ThreadBarrierPrivate { + public: + pthread_mutex_t mutex_; + pthread_cond_t cond_; + int count_; + int tripCount_; + + inline explicit ThreadBarrierPrivate(int cnt) : count_(0), tripCount_(cnt) { + CHECK_NE(cnt, 0); + CHECK_GE(pthread_mutex_init(&mutex_, 0), 0); + CHECK_GE(pthread_cond_init(&cond_, 0), 0); + } + + inline ~ThreadBarrierPrivate() { + pthread_cond_destroy(&cond_); + pthread_mutex_destroy(&mutex_); + } + + /** + * @brief wait + * @return true if the last wait + */ + inline bool wait() { + pthread_mutex_lock(&mutex_); + ++count_; + if (count_ >= tripCount_) { + count_ = 0; + pthread_cond_broadcast(&cond_); + pthread_mutex_unlock(&mutex_); + return true; + } else { + pthread_cond_wait(&cond_, &mutex_); + pthread_mutex_unlock(&mutex_); + return false; + } + } +}; + +#endif + +/// ThreadBarrier + +ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {} +ThreadBarrier::~ThreadBarrier() { delete m; } +void ThreadBarrier::wait() { m->wait(); } + +} // namespace paddle diff --git a/paddle/legacy/utils/arch/osx/Excepts.cpp b/paddle/legacy/utils/arch/osx/Excepts.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2b7d6dca8454417fd78f6da7f906785d24a6219b --- /dev/null +++ b/paddle/legacy/utils/arch/osx/Excepts.cpp @@ -0,0 +1,57 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Excepts.h" + +#if defined(__APPLE__) || defined(__OSX__) +#if defined(__arm__) || defined(__arm64__) +// TODO(liuyiqun): implement the arm version +int fegetexcept(void) { return -1; } +int feenableexcept(unsigned int excepts) { return -1; } +int fedisableexcept(unsigned int excepts) { return -1; } +#else +int fegetexcept(void) { + static fenv_t fenv; + return fegetenv(&fenv) ? -1 : (fenv.__control & FE_ALL_EXCEPT); +} + +int feenableexcept(unsigned int excepts) { + static fenv_t fenv; + unsigned int new_excepts = excepts & FE_ALL_EXCEPT, old_excepts; + + if (fegetenv(&fenv)) return -1; + old_excepts = fenv.__control & FE_ALL_EXCEPT; + + // unmask + fenv.__control &= ~new_excepts; + fenv.__mxcsr &= ~(new_excepts << 7); + + return (fesetenv(&fenv) ? -1 : old_excepts); +} + +int fedisableexcept(unsigned int excepts) { + static fenv_t fenv; + unsigned int new_excepts = excepts & FE_ALL_EXCEPT, old_excepts; + + if (fegetenv(&fenv)) return -1; + old_excepts = fenv.__control & FE_ALL_EXCEPT; + + // mask + fenv.__control |= new_excepts; + fenv.__mxcsr |= new_excepts << 7; + + return (fesetenv(&fenv) ? -1 : old_excepts); +} +#endif +#endif diff --git a/paddle/legacy/utils/arch/osx/Locks.cpp b/paddle/legacy/utils/arch/osx/Locks.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b68c48f0c31aa928a634e0369295ec084b9ccd8e --- /dev/null +++ b/paddle/legacy/utils/arch/osx/Locks.cpp @@ -0,0 +1,105 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Locks.h" +#include +#include +#include +#include "paddle/legacy/utils/Logging.h" + +namespace paddle { + +class SemaphorePrivate { + public: + ~SemaphorePrivate() { dispatch_release(sem); } + + dispatch_semaphore_t sem; +}; + +Semaphore::Semaphore(int initValue) : m(new SemaphorePrivate()) { + m->sem = dispatch_semaphore_create(initValue); +} + +Semaphore::~Semaphore() { delete m; } + +bool Semaphore::timeWait(timespec *ts) { + dispatch_time_t tm = dispatch_walltime(ts, 0); + return (0 == dispatch_semaphore_wait(m->sem, tm)); +} + +void Semaphore::wait() { + dispatch_semaphore_wait(m->sem, DISPATCH_TIME_FOREVER); +} + +void Semaphore::post() { dispatch_semaphore_signal(m->sem); } + +class SpinLockPrivate { + public: + std::atomic_flag lock_ = ATOMIC_FLAG_INIT; + char padding_[64 - sizeof(lock_)]; // Padding to cache line size +}; + +SpinLock::SpinLock() : m(new SpinLockPrivate()) {} +SpinLock::~SpinLock() { delete m; } + +void SpinLock::lock() { + while (m->lock_.test_and_set(std::memory_order_acquire)) { + } +} + +void SpinLock::unlock() { m->lock_.clear(std::memory_order_release); } + +class ThreadBarrierPrivate { + public: + pthread_mutex_t mutex_; + pthread_cond_t cond_; + int count_; + int tripCount_; + + inline explicit ThreadBarrierPrivate(int cnt) : count_(0), tripCount_(cnt) { + CHECK_NE(cnt, 0); + CHECK_GE(pthread_mutex_init(&mutex_, 0), 0); + CHECK_GE(pthread_cond_init(&cond_, 0), 0); + } + + inline ~ThreadBarrierPrivate() { + pthread_cond_destroy(&cond_); + pthread_mutex_destroy(&mutex_); + } + + /** + * @brief wait + * @return true if the last wait + */ + inline bool wait() { + pthread_mutex_lock(&mutex_); + ++count_; + if (count_ >= tripCount_) { + count_ = 0; + pthread_cond_broadcast(&cond_); + pthread_mutex_unlock(&mutex_); + return true; + } else { + pthread_cond_wait(&cond_, &mutex_); + pthread_mutex_unlock(&mutex_); + return false; + } + } +}; + +ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {} +ThreadBarrier::~ThreadBarrier() { delete m; } +void ThreadBarrier::wait() { m->wait(); } + +} // namespace paddle diff --git a/paddle/utils/enable_virtualenv.py b/paddle/legacy/utils/enable_virtualenv.py similarity index 100% rename from paddle/utils/enable_virtualenv.py rename to paddle/legacy/utils/enable_virtualenv.py diff --git a/paddle/legacy/utils/tests/CMakeLists.txt b/paddle/legacy/utils/tests/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..4af01db5c84cb497b756027cbb6ad06c081a8899 --- /dev/null +++ b/paddle/legacy/utils/tests/CMakeLists.txt @@ -0,0 +1,18 @@ +add_simple_unittest(test_Thread) +add_simple_unittest(test_StringUtils) +add_simple_unittest(test_CustomStackTrace) +add_simple_unittest(test_ThreadBarrier) +add_simple_unittest(test_SpinLock) +add_simple_unittest(test_SIMDFlags) +add_simple_unittest(test_Error) + +add_executable( + test_CustomStackTracePrint + test_CustomStackTracePrint.cpp +) +link_paddle_exe(test_CustomStackTracePrint) +if(NOT APPLE) + add_test(NAME test_CustomStackTracePrint + COMMAND ${PADDLE_SOURCE_DIR}/paddle/legacy/utils/tests/test_CustomStackTracePrint.sh + WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) +endif() diff --git a/paddle/legacy/utils/tests/test_CustomStackTrace.cpp b/paddle/legacy/utils/tests/test_CustomStackTrace.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2a418e3ae2277fc5dc6856d131dafa9daf0bad47 --- /dev/null +++ b/paddle/legacy/utils/tests/test_CustomStackTrace.cpp @@ -0,0 +1,92 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include // NOLINT +#include // NOLINT + +#include "paddle/legacy/utils/CustomStackTrace.h" +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/StringUtil.h" +#include "paddle/legacy/utils/Util.h" + +DEFINE_int32(test_thread_num, 10, "testing thread number"); + +void testNormalImpl( + const std::function&, + size_t, + size_t, + paddle::ThreadBarrier&, + paddle::ThreadBarrier&)>& callback) { + paddle::CustomStackTrace tracer; + paddle::ThreadBarrier doneBarrier(FLAGS_test_thread_num + 1); + paddle::ThreadBarrier startBarrier(FLAGS_test_thread_num + 1); + constexpr size_t countDown = 10; + constexpr size_t layerSize = 1000; + std::vector> threads; + threads.reserve(FLAGS_test_thread_num); + + for (int32_t i = 0; i < FLAGS_test_thread_num; ++i) { + threads.emplace_back( + new std::thread([&tracer, &startBarrier, &doneBarrier, &callback] { + callback(tracer, countDown, layerSize, startBarrier, doneBarrier); + })); + } + size_t cntDown = countDown; + while (cntDown-- > 0) { + startBarrier.wait(); + sleep(1); + doneBarrier.wait(); + ASSERT_TRUE(tracer.empty()); + } + + for (auto& thread : threads) { + thread->join(); + } +} + +TEST(CustomStackTrace, normalTrain) { + testNormalImpl([](paddle::CustomStackTrace& tracer, + size_t countDown, + size_t layerSize, + paddle::ThreadBarrier& start, + paddle::ThreadBarrier& finish) { + while (countDown-- > 0) { + start.wait(); + for (size_t i = 0; i < layerSize; ++i) { + tracer.push("layer_" + paddle::str::to_string(i)); + } + for (size_t i = 0; i < layerSize; ++i) { + tracer.pop("layer_" + paddle::str::to_string(layerSize - 1 - i)); + } + finish.wait(); + } + }); +} + +TEST(CustomStackTrace, normalTest) { + testNormalImpl([](paddle::CustomStackTrace& tracer, + size_t countDown, + size_t layerSize, + paddle::ThreadBarrier& start, + paddle::ThreadBarrier& finish) { + while (countDown-- > 0) { + start.wait(); + for (size_t i = 0; i < layerSize; ++i) { + tracer.push("layer_" + paddle::str::to_string(i)); + } + tracer.clear(); // in forward test, tracer will clear after forward. + finish.wait(); + } + }); +} diff --git a/paddle/legacy/utils/tests/test_CustomStackTracePrint.cpp b/paddle/legacy/utils/tests/test_CustomStackTracePrint.cpp new file mode 100644 index 0000000000000000000000000000000000000000..78886a3ed9f237a39079bbf604a376f98bd86b59 --- /dev/null +++ b/paddle/legacy/utils/tests/test_CustomStackTracePrint.cpp @@ -0,0 +1,30 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/CustomStackTrace.h" +#include "paddle/legacy/utils/StringUtil.h" +#include "paddle/legacy/utils/Util.h" + +int main(int argc, char** argv) { + paddle::initMain(argc, argv); + + for (size_t i = 0; i < 1000; ++i) { + paddle::gLayerStackTrace.push("layer_" + paddle::str::to_string(i)); + if (i == 998) { + throw "Unhandle exception"; + } + } + + return 0; +} diff --git a/paddle/utils/tests/test_CustomStackTracePrint.sh b/paddle/legacy/utils/tests/test_CustomStackTracePrint.sh similarity index 100% rename from paddle/utils/tests/test_CustomStackTracePrint.sh rename to paddle/legacy/utils/tests/test_CustomStackTracePrint.sh diff --git a/paddle/legacy/utils/tests/test_Error.cpp b/paddle/legacy/utils/tests/test_Error.cpp new file mode 100644 index 0000000000000000000000000000000000000000..250c4d58a64a0d284a15418e40264f1857d30050 --- /dev/null +++ b/paddle/legacy/utils/tests/test_Error.cpp @@ -0,0 +1,34 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/Error.h" + +#include + +TEST(Error, testAll) { + paddle::Error error; + ASSERT_TRUE(error.isOK()); + error = paddle::Error("I'm the error"); + ASSERT_FALSE(error.isOK()); + ASSERT_STREQ("I'm the error", error.msg()); + + error = paddle::Error("error2"); + ASSERT_FALSE(error.isOK()); + ASSERT_STREQ("error2", error.msg()); + + int i = 3; + auto error3 = paddle::Error("error%d", i); + ASSERT_FALSE(error3.isOK()); + ASSERT_STREQ("error3", error3.msg()); +} diff --git a/paddle/legacy/utils/tests/test_SIMDFlags.cpp b/paddle/legacy/utils/tests/test_SIMDFlags.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6362210acdaf26a26a2548ddaf8ed455b9c76618 --- /dev/null +++ b/paddle/legacy/utils/tests/test_SIMDFlags.cpp @@ -0,0 +1,48 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "paddle/legacy/utils/CpuId.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Util.h" + +using namespace paddle; // NOLINT + +TEST(SIMDFlags, gccTest) { +#if (defined(__GNUC__) || defined(__GNUG__)) && !(defined(__clang__)) && \ + !defined(__arm__) && !defined(__aarch64__) + // clang-format off + CHECK(!__builtin_cpu_supports("sse") != HAS_SSE); + CHECK(!__builtin_cpu_supports("sse2") != HAS_SSE2); + CHECK(!__builtin_cpu_supports("sse3") != HAS_SSE3); + CHECK(!__builtin_cpu_supports("ssse3") != HAS_SSSE3); + CHECK(!__builtin_cpu_supports("sse4.1") != HAS_SSE41); + CHECK(!__builtin_cpu_supports("sse4.2") != HAS_SSE42); + CHECK(!__builtin_cpu_supports("avx") != HAS_AVX); + CHECK(!__builtin_cpu_supports("avx2") != HAS_AVX2); +// clang-format on +#endif +} + +TEST(SIMDFlags, normalPrint) { + LOG(INFO) << "Has SSE: " << std::boolalpha << HAS_SSE; + LOG(INFO) << "Has SSE2: " << std::boolalpha << HAS_SSE2; + LOG(INFO) << "Has SSE3: " << std::boolalpha << HAS_SSE3; + LOG(INFO) << "Has SSSE3: " << std::boolalpha << HAS_SSSE3; + LOG(INFO) << "Has SSE4: " << std::boolalpha << HAS_SSE41 || HAS_SSE42; + LOG(INFO) << "Has FMA3: " << std::boolalpha << HAS_FMA3; + LOG(INFO) << "Has FMA4: " << std::boolalpha << HAS_FMA4; + LOG(INFO) << "Has AVX: " << std::boolalpha << HAS_AVX; + LOG(INFO) << "Has AVX2: " << std::boolalpha << HAS_AVX2; + LOG(INFO) << "Has AVX512: " << std::boolalpha << HAS_AVX512; + LOG(INFO) << "Has NEON: " << std::boolalpha << HAS_NEON; +} diff --git a/paddle/legacy/utils/tests/test_SpinLock.cpp b/paddle/legacy/utils/tests/test_SpinLock.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4cd7836d6af251b48925de95c2811361313d7b41 --- /dev/null +++ b/paddle/legacy/utils/tests/test_SpinLock.cpp @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include +#include + +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Util.h" + +DEFINE_int32(test_thread_num, 100, "testing thread number"); + +void testNormalImpl( + size_t thread_num, + const std::function& callback) { + paddle::SpinLock mutex; + std::vector threads; + threads.reserve(thread_num); + + size_t count = 0; + for (size_t i = 0; i < thread_num; ++i) { + threads.emplace_back([&thread_num, &count, &mutex, &callback] { + callback(thread_num, count, mutex); + }); + } + for (auto& thread : threads) { + thread.join(); + } + // Check whether all threads reach this point or not + CHECK_EQ(count, thread_num); +} + +TEST(ThreadSpinLock, normalTest) { + for (auto& thread_num : {10, 30, 50, 100, 300, 1000}) { + testNormalImpl( + thread_num, + [](size_t thread_num, size_t& count, paddle::SpinLock& mutex) { + std::lock_guard lock(mutex); + ++count; + }); + } +} diff --git a/paddle/legacy/utils/tests/test_StringUtils.cpp b/paddle/legacy/utils/tests/test_StringUtils.cpp new file mode 100644 index 0000000000000000000000000000000000000000..61d2815f097af7125bfefdc4909509564300d6aa --- /dev/null +++ b/paddle/legacy/utils/tests/test_StringUtils.cpp @@ -0,0 +1,23 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/legacy/utils/StringUtil.h" + +#include + +TEST(StringUtil, to) { + ASSERT_NEAR(paddle::str::to("12.45"), 12.45, 1e-5); + ASSERT_DEATH_IF_SUPPORTED(paddle::str::to("12.45x23"), ".*"); + ASSERT_DEATH_IF_SUPPORTED(paddle::str::to(""), ".*"); +} diff --git a/paddle/legacy/utils/tests/test_Thread.cpp b/paddle/legacy/utils/tests/test_Thread.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5e07da3236862c5f585671d9bb8e3fbbd1c5b5fc --- /dev/null +++ b/paddle/legacy/utils/tests/test_Thread.cpp @@ -0,0 +1,81 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include + +using paddle::AsyncThreadPool; // NOLINT + +TEST(AsyncThreadPool, addJob) { + AsyncThreadPool pool(8); + auto a = pool.addJob([] { return 1; }); + auto b = pool.addJob([] { return true; }); + auto c = pool.addJob([] { return false; }); + + ASSERT_EQ(a.get(), 1); + ASSERT_TRUE(b.get()); + ASSERT_FALSE(c.get()); +} + +TEST(AsyncThreadPool, addBatchJob) { + AsyncThreadPool pool(8); + std::atomic counter{0}; + + std::vector jobs; + + for (int i = 0; i < 10000; i++) { + jobs.emplace_back([&] { counter++; }); + } + + pool.addBatchJobs(jobs); + + ASSERT_EQ(counter, 10000); +} + +TEST(AsyncThreadPool, multiThreadAddBatchJob) { + AsyncThreadPool levelOnePool(200); + AsyncThreadPool levelTwoPool(200); + + std::shared_ptr mut = std::make_shared(); + int counter = 0; + const int numMonitors = 300; + const int numSlaves = 300; + std::vector moniterJobs(numMonitors, [&] { + std::vector slaveJobs(numSlaves, [mut, &counter] { + std::lock_guard lk(*mut); + counter++; + }); + levelTwoPool.addBatchJobs(slaveJobs); + }); + levelOnePool.addBatchJobs(moniterJobs); + ASSERT_EQ(counter, numMonitors * numSlaves); +} + +TEST(AsyncThreadPool, addBatchJobWithResults) { + AsyncThreadPool pool(100); + + std::vector> jobs; + const int numJobs = 100; + for (int i = 0; i < numJobs; i++) { + jobs.emplace_back([i] { return i; }); + } + + std::vector res; + pool.addBatchJobs(jobs, res); + + for (int i = 0; i < numJobs; i++) { + ASSERT_EQ(res[i], i); + } +} diff --git a/paddle/legacy/utils/tests/test_ThreadBarrier.cpp b/paddle/legacy/utils/tests/test_ThreadBarrier.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9c8851ae2112320c89aa3e7ed6e850d00cc14006 --- /dev/null +++ b/paddle/legacy/utils/tests/test_ThreadBarrier.cpp @@ -0,0 +1,66 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include +#include + +#include "paddle/legacy/utils/Locks.h" +#include "paddle/legacy/utils/Logging.h" +#include "paddle/legacy/utils/Util.h" + +DEFINE_int32(test_thread_num, 100, "testing thread number"); + +void testNormalImpl( + size_t thread_num, + const std::function&, + paddle::ThreadBarrier&)>& callback) { + std::mutex mutex; + std::set tids; + paddle::ThreadBarrier barrier(thread_num); + + std::vector threads; + threads.reserve(thread_num); + for (size_t i = 0; i < thread_num; ++i) { + threads.emplace_back([&thread_num, &mutex, &tids, &barrier, &callback] { + callback(thread_num, mutex, tids, barrier); + }); + } + + for (auto& thread : threads) { + thread.join(); + } +} + +TEST(ThreadBarrier, normalTest) { + for (auto& thread_num : {10, 30, 50, 100, 300, 1000}) { + testNormalImpl(thread_num, + [](size_t thread_num, + std::mutex& mutex, + std::set& tids, + paddle::ThreadBarrier& barrier) { + { + std::lock_guard guard(mutex); + tids.insert(std::this_thread::get_id()); + } + barrier.wait(); + // Check whether all threads reach this point or not + CHECK_EQ(tids.size(), thread_num); + }); + } +} diff --git a/paddle/math/Allocator.h b/paddle/math/Allocator.h deleted file mode 100644 index c43a83891eb6b7eae278169736149ad1d89e950e..0000000000000000000000000000000000000000 --- a/paddle/math/Allocator.h +++ /dev/null @@ -1,137 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "hl_gpu.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -/** - * @brief Allocator base class. - * - * This is the base class of all Allocator class. - */ -class Allocator { - public: - virtual ~Allocator() {} - virtual void* alloc(size_t size) = 0; - virtual void free(void* ptr) = 0; - virtual std::string getName() = 0; -}; - -/** - * @brief CPU allocator implementation. - */ -class CpuAllocator : public Allocator { - public: - ~CpuAllocator() {} - - /** - * @brief Aligned allocation on CPU. - * @param size Size to be allocated. - * @return Pointer to the allocated memory - */ - virtual void* alloc(size_t size) { - void* ptr; -#ifdef PADDLE_WITH_MKLDNN - // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp - // memory alignment - CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0); -#else - CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0); -#endif - CHECK(ptr) << "Fail to allocate CPU memory: size=" << size; - return ptr; - } - - /** - * @brief Free the memory space. - * @param ptr Pointer to be free. - */ - virtual void free(void* ptr) { - if (ptr) { - ::free(ptr); - } - } - - virtual std::string getName() { return "cpu_alloc"; } -}; - -/** - * @brief GPU allocator implementation. - */ -class GpuAllocator : public Allocator { - public: - ~GpuAllocator() {} - - /** - * @brief Allocate GPU memory. - * @param size Size to be allocated. - * @return Pointer to the allocated memory - */ - virtual void* alloc(size_t size) { - void* ptr = hl_malloc_device(size); - CHECK(ptr) << "Fail to allocate GPU memory " << size << " bytes"; - return ptr; - } - - /** - * @brief Free the GPU memory. - * @param ptr Pointer to be free. - */ - virtual void free(void* ptr) { - if (ptr) { - hl_free_mem_device(ptr); - } - } - - virtual std::string getName() { return "gpu_alloc"; } -}; - -/** - * @brief CPU pinned memory allocator implementation. - */ -class CudaHostAllocator : public Allocator { - public: - ~CudaHostAllocator() {} - - /** - * @brief Allocate pinned memory. - * @param size Size to be allocated. - * @return Pointer to the allocated memory - */ - virtual void* alloc(size_t size) { - void* ptr = hl_malloc_host(size); - CHECK(ptr) << "Fail to allocate pinned memory " << size << " bytes"; - return ptr; - } - - /** - * @brief Free the pinned memory. - * @param ptr Pointer to be free. - */ - virtual void free(void* ptr) { - if (ptr) { - hl_free_mem_host(ptr); - } - } - - virtual std::string getName() { return "cuda_host_alloc"; } -}; - -} // namespace paddle diff --git a/paddle/math/BaseMatrix.cu b/paddle/math/BaseMatrix.cu deleted file mode 100644 index 7b57419e5a510ba50aff0b47681d1294607e31f9..0000000000000000000000000000000000000000 --- a/paddle/math/BaseMatrix.cu +++ /dev/null @@ -1,1953 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include "BaseMatrix.h" -#include "MathFunctions.h" -#include "NEONFunctions.h" -#include "SIMDFunctions.h" -#include "hl_matrix_apply.cuh" -#include "hl_matrix_base.cuh" -#include "hl_matrix_ops.cuh" - -namespace paddle { - -const char* SPARSE_SUPPORT_ERROR = "Sparse Matrix/Vector is not supported."; - -template -template -int BaseMatrixT::applyUnary(Op op) { - MatrixOffset offset(0, 0); - applyUnary(op, height_, width_, offset); - return 0; -} - -template -template -int BaseMatrixT::applyUnary(Op op, - int numRows, - int numCols, - MatrixOffset& offset) { - CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; - int dimM = numRows; - int dimN = numCols; - int lda = stride_; - - T* A = data_; - CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - - CHECK_LE(dimM + offset.aRow_, this->height_); - CHECK_LE(dimN + offset.aCol_, this->width_); - if (true == useGpu_) { - hl_gpu_apply_unary_op(op, A, dimM, dimN, lda); - } else { - hl_cpu_apply_unary_op(op, A, dimM, dimN, lda); - } - return 0; -} - -template -template -int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b) { - CHECK(height_ == b.height_ && width_ == b.width_) - << "Matrix dimensions are not equal"; - - MatrixOffset offset(0, 0, 0, 0); - applyBinary(op, b, height_, width_, offset); - return 0; -} - -template -template -int BaseMatrixT::applyBinary( - Op op, BaseMatrixT& b, int numRows, int numCols, MatrixOffset& offset) { - applyBinary(op, b, numRows, numCols, offset, false_type(), false_type()); - return 0; -} - -template -template -int BaseMatrixT::applyBinary(Op op, - BaseMatrixT& b, - int numRows, - int numCols, - MatrixOffset& offset, - bAsRowVector, - bAsColVector) { - CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(useGpu_ == b.useGpu_) << "Matrix type mismatch"; - - int dimM = numRows; - int dimN = numCols; - int lda = stride_; - int ldb = b.stride_; - - T* A = data_; - T* B = b.data_; - CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS( - B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); - CHECK_LE(dimM + offset.aRow_, this->height_); - CHECK_LE(dimN + offset.aCol_, this->width_); - if (!bAsRowVector::value && !bAsColVector::value) { - CHECK_LE(dimM + offset.bRow_, b.height_); - CHECK_LE(dimN + offset.bCol_, b.width_); - } else if (bAsRowVector::value && !bAsColVector::value) { - CHECK_LE(dimN + offset.bCol_, b.width_); - } else if (!bAsRowVector::value && bAsColVector::value) { - CHECK_LE(dimM + offset.bRow_, b.height_); - } else { - } - if (true == useGpu_) { - hl_gpu_apply_binary_op( - op, A, B, dimM, dimN, lda, ldb); - } else { - hl_cpu_apply_binary_op( - op, A, B, dimM, dimN, lda, ldb); - } - - return 0; -} - -template -template -int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c) { - CHECK_EQ(height_, b.height_); - CHECK_EQ(width_, b.width_); - CHECK_EQ(height_, c.height_); - CHECK_EQ(width_, c.width_); - - MatrixOffset offset(0, 0, 0, 0, 0, 0); - applyTernary(op, b, c, height_, width_, offset); - - return 0; -} - -template -template -int BaseMatrixT::applyTernary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - int numRows, - int numCols, - MatrixOffset& offset) { - applyTernary(op, b, c, numRows, numCols, offset, false_type(), false_type()); - - return 0; -} - -template -template -int BaseMatrixT::applyTernary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - int numRows, - int numCols, - MatrixOffset& offset, - cAsRowVector, - cAsColVector) { - CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(!c.isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK_EQ(useGpu_, b.useGpu_); - CHECK_EQ(useGpu_, c.useGpu_); - - int dimM = numRows; - int dimN = numCols; - int lda = stride_; - int ldb = b.stride_; - int ldc = c.stride_; - - T* A = data_; - T* B = b.data_; - T* C = c.data_; - CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS( - B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); - CAL_MATRIX_START_ADDRESS( - C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); - - CHECK_LE(dimM + offset.aRow_, this->height_); - CHECK_LE(dimN + offset.aCol_, this->width_); - CHECK_LE(dimM + offset.bRow_, b.height_); - CHECK_LE(dimN + offset.bCol_, b.width_); - if (!cAsRowVector::value && !cAsColVector::value) { - CHECK_LE(dimM + offset.cRow_, c.height_); - CHECK_LE(dimN + offset.cCol_, c.width_); - } else if (cAsRowVector::value && !cAsColVector::value) { - CHECK_LE(dimN + offset.cCol_, c.width_); - } else if (!cAsRowVector::value && cAsColVector::value) { - CHECK_LE(dimM + offset.cRow_, c.height_); - } else { - } - - if (true == useGpu_) { - hl_gpu_apply_ternary_op( - op, A, B, C, dimM, dimN, lda, ldb, ldc); - } else { - hl_cpu_apply_ternary_op( - op, A, B, C, dimM, dimN, lda, ldb, ldc); - } - - return 0; -} - -template -template -int BaseMatrixT::applyQuaternary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - BaseMatrixT& d) { - CHECK_EQ(height_, b.height_); - CHECK_EQ(width_, b.width_); - CHECK_EQ(height_, c.height_); - CHECK_EQ(width_, c.width_); - CHECK_EQ(height_, d.height_); - CHECK_EQ(width_, d.width_); - - MatrixOffset offset(0, 0, 0, 0, 0, 0, 0, 0); - applyQuaternary(op, b, c, d, height_, width_, offset); - - return 0; -} - -template -template -int BaseMatrixT::applyQuaternary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - BaseMatrixT& d, - int numRows, - int numCols, - MatrixOffset& offset) { - CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(!c.isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK(!d.isSparse()) << SPARSE_SUPPORT_ERROR; - CHECK_EQ(useGpu_, b.useGpu_); - CHECK_EQ(useGpu_, c.useGpu_); - CHECK_EQ(useGpu_, d.useGpu_); - - int dimM = numRows; - int dimN = numCols; - int lda = stride_; - int ldb = b.stride_; - int ldc = c.stride_; - int ldd = d.stride_; - - T* A = data_; - T* B = b.data_; - T* C = c.data_; - T* D = d.data_; - CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS( - B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); - CAL_MATRIX_START_ADDRESS( - C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); - CAL_MATRIX_START_ADDRESS( - D, d.height_, d.width_, ldd, offset.dCol_, offset.dRow_); - - CHECK_LE(dimM + offset.aRow_, this->height_); - CHECK_LE(dimN + offset.aCol_, this->width_); - CHECK_LE(dimM + offset.bRow_, b.height_); - CHECK_LE(dimN + offset.bCol_, b.width_); - CHECK_LE(dimM + offset.cRow_, c.height_); - CHECK_LE(dimN + offset.cCol_, c.width_); - CHECK_LE(dimM + offset.dRow_, d.height_); - CHECK_LE(dimN + offset.dCol_, d.width_); - if (true == useGpu_) { - hl_gpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, ldc, ldd); - } else { - hl_cpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, ldc, ldd); - } - - return 0; -} - -template -template -int BaseMatrixT::aggregate(Agg agg, - Op op, - Saver sv, - BaseMatrixT& b, - int numRows, - int numCols, - MatrixOffset& offset, - aAsRowVector, - aAsColVector) { - CHECK_EQ(useGpu_, b.useGpu_); - - int ld = stride_; - int ldb = b.stride_; - - T* dst = data_; - T* B = b.data_; - CAL_MATRIX_START_ADDRESS( - dst, height_, width_, ld, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS( - B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); - - if (aAsRowVector::value && !aAsColVector::value) { - if (useGpu_) { - hl_gpu_matrix_column_op(agg, op, sv, numRows, numCols, dst, B, ldb); - } else { - hl_cpu_matrix_column_op(agg, op, sv, numRows, numCols, dst, B, ldb); - } - } else if (!aAsRowVector::value && aAsColVector::value) { - if (useGpu_) { - hl_gpu_matrix_row_op(agg, op, sv, numRows, numCols, dst, ld, B, ldb); - } else { - hl_cpu_matrix_row_op(agg, op, sv, numRows, numCols, dst, ld, B, ldb); - } - } else { - LOG(FATAL) << "not supported"; - } - - return 0; -} - -template -template -int BaseMatrixT::aggregate(Agg agg, - Op op, - Saver sv, - BaseMatrixT& b, - BaseMatrixT& c, - int numRows, - int numCols, - MatrixOffset& offset, - aAsRowVector, - aAsColVector) { - CHECK_EQ(useGpu_, b.useGpu_); - CHECK_EQ(useGpu_, c.useGpu_); - - int ld = stride_; - int ldb = b.stride_; - int ldc = c.stride_; - - T* dst = data_; - T* B = b.data_; - T* C = c.data_; - CAL_MATRIX_START_ADDRESS( - dst, height_, width_, ld, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS( - B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); - CAL_MATRIX_START_ADDRESS( - C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); - - if (aAsRowVector::value && !aAsColVector::value) { - if (useGpu_) { - hl_gpu_matrix_column_op( - agg, op, sv, numRows, numCols, dst, B, ldb, C, ldc); - } else { - hl_cpu_matrix_column_op( - agg, op, sv, numRows, numCols, dst, B, ldb, C, ldc); - } - } else if (!aAsRowVector::value && aAsColVector::value) { - if (useGpu_) { - hl_gpu_matrix_row_op( - agg, op, sv, numRows, numCols, dst, ld, B, ldb, C, ldc); - } else { - hl_cpu_matrix_row_op( - agg, op, sv, numRows, numCols, dst, ld, B, ldb, C, ldc); - } - } else { - LOG(FATAL) << "not supported"; - } - - return 0; -} - -/** - * @brief unary operator. - * - */ - -DEFINE_MATRIX_UNARY_OP(Neg, a = -a); -template -void BaseMatrixT::neg() { - applyUnary(unary::Neg()); -} - -DEFINE_MATRIX_UNARY_OP(Exp, a = exp(a)); -template <> -void BaseMatrixT::exp2() { - applyUnary(unary::Exp()); -} - -DEFINE_MATRIX_UNARY_OP(Log, a = log(a)); -template <> -void BaseMatrixT::log2() { - if (useGpu_) { - applyUnary(unary::Log()); - } else { - vLog(height_ * width_, data_, data_); - } -} - -DEFINE_MATRIX_UNARY_OP(Sqrt, a = sqrt(a)); -template <> -void BaseMatrixT::sqrt2() { - applyUnary(unary::Sqrt()); -} - -DEFINE_MATRIX_UNARY_OP(Square, a = a * a); -template -void BaseMatrixT::square2() { - applyUnary(unary::Square()); -} - -DEFINE_MATRIX_UNARY_OP(Reciprocal, a = 1.0f / a); -template -void BaseMatrixT::reciprocal2() { - applyUnary(unary::Reciprocal()); -} - -DEFINE_MATRIX_UNARY_OP(Abs, a = a > 0 ? a : -a); -template -void BaseMatrixT::abs2() { - applyUnary(unary::Abs()); -} - -DEFINE_MATRIX_UNARY_OP(Sign, a = (a > 0) - (a < 0)); -template -void BaseMatrixT::sign2() { - applyUnary(unary::Sign()); -} - -DEFINE_MATRIX_UNARY_OP(Zero, a = 0); -template -void BaseMatrixT::zero() { - applyUnary(unary::Zero()); -} - -template -void BaseMatrixT::zeroAtOffset(int64_t columnOffset, int64_t numColumns) { - int numRows = height_; - int numCols = numColumns; - MatrixOffset offset(columnOffset, 0); - applyUnary(unary::Zero(), numRows, numCols, offset); -} - -DEFINE_MATRIX_UNARY_OP(One, a = 1); -template -void BaseMatrixT::one() { - applyUnary(unary::One()); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(Pow, ONE_PARAMETER, a = pow(a, p)); -template <> -void BaseMatrixT::pow2(real p) { - if (useGpu_) { - applyUnary(unary::Pow(p)); - } else { - vPow(height_ * width_, data_, p, data_); - } -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(SubScalar, ONE_PARAMETER, a -= p); -template -void BaseMatrixT::subScalar(T p) { - applyUnary(unary::SubScalar(p)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(MulScalar, ONE_PARAMETER, a *= p); -template -void BaseMatrixT::mulScalar(T p) { - applyUnary(unary::MulScalar(p)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(DivScalar, ONE_PARAMETER, a /= p); -template -void BaseMatrixT::divScalar(T p) { - applyUnary(unary::DivScalar(p)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(Assign, ONE_PARAMETER, a = p); -template -void BaseMatrixT::assign(T p) { - applyUnary(unary::Assign(p)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(Add, ONE_PARAMETER, a += p); -template -void BaseMatrixT::add(T p) { - applyUnary(unary::Add(p)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(Add2, TWO_PARAMETER, a = a * p1 + p2); -template -void BaseMatrixT::add(T p1, T p2) { - applyUnary(unary::Add2(p1, p2)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, - TWO_PARAMETER, - a = a < p1 ? p1 : (a > p2 ? p2 : a)); -template -void BaseMatrixT::clip(T p1, T p2) { - applyUnary(unary::Clip(p1, p2)); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, - TWO_PARAMETER, - a = b < p1 ? 0 : (b > p2 ? 0 : 1)); -template -void BaseMatrixT::clipDerivative(BaseMatrixT& b, T p1, T p2) { - applyBinary(binary::ClipDerivative(p1, p2), b); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, - ONE_PARAMETER, - a = a > p ? 1.0f : 0.0f); -template -void BaseMatrixT::biggerThanScalar(T p) { - applyUnary(unary::BiggerThanScalar(p)); -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(DownClip, ONE_PARAMETER, a = a > p ? a : p); -template -void BaseMatrixT::downClip(T p) { - applyUnary(unary::DownClip(p)); -} - -/** - * @brief binary operator. - * - */ - -DEFINE_MATRIX_BINARY_OP(Add, a += b); -template -void BaseMatrixT::add(BaseMatrixT& b) { - applyBinary(binary::Add(), b); -} - -template <> -void BaseMatrixT::add(BaseMatrixT& b) { - if (useGpu_) { - applyBinary(binary::Add(), b); - } else { // cpu branch - CHECK_EQ(height_, b.height_); - CHECK_EQ(width_, b.width_); - vAdd(height_ * width_, data_, b.data_, data_); - } -} - -template -void BaseMatrixT::addAtOffset(BaseMatrixT& b, int64_t columnOffset) { - if (columnOffset + b.width_ <= width_) { - int numRows = height_; - int numCols = b.width_; - MatrixOffset offset(columnOffset, 0, 0, 0); - applyBinary(binary::Add(), b, numRows, numCols, offset); - } else if (columnOffset + width_ <= b.width_) { - int numRows = height_; - int numCols = width_; - MatrixOffset offset(0, 0, columnOffset, 0); - applyBinary(binary::Add(), b, numRows, numCols, offset); - } else { - LOG(FATAL) << "Wrong argument " - << " a.width=" << width_ << " b.width=" << b.width_ - << " columnOffset=" << columnOffset; - } -} - -template -void BaseMatrixT::addP2P(BaseMatrixT& b) { - T* A = data_; - T* B = b.data_; - int dimM = height_; - int dimN = width_; - - hl_gpu_apply_binary_op, 0, 0>( - binary::Add(), A, B, dimM, dimN, dimN, dimN); -} - -template -void BaseMatrixT::addColVector(BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::Add(), - b, - numRows, - numCols, - offset, - false_type(), - true_type() /* bAsColVector */); -} - -template -void BaseMatrixT::addRowVector(BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::Add(), - b, - numRows, - numCols, - offset, - true_type() /* bAsRowVector */, - false_type()); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(Add1, ONE_PARAMETER, a += b * p); -template -void BaseMatrixT::add(BaseMatrixT& b, T p) { - applyBinary(binary::Add1(p), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(Pow, ONE_PARAMETER, a = pow(b, p)); -template <> -void BaseMatrixT::pow2(BaseMatrixT& b, real p) { - if (useGpu_) { - applyBinary(binary::Pow(p), b); - } else { - vPow(height_ * width_, b.data_, p, data_); - } -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(Add2, TWO_PARAMETER, a = p1 * a + p2 * b); -template -void BaseMatrixT::add(BaseMatrixT& b, T p1, T p2) { - applyBinary(binary::Add2(p1, p2), b); -} - -template -void BaseMatrixT::addBias(BaseMatrixT& b, T scale) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::Add1(scale), - b, - numRows, - numCols, - offset, - true_type() /* bAsRowVector */, - false_type()); -} - -DEFINE_MATRIX_BINARY_OP(Sub, a -= b); -template -void BaseMatrixT::sub(BaseMatrixT& b) { - applyBinary(binary::Sub(), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(Sub1, ONE_PARAMETER, a -= b * p); -template -void BaseMatrixT::sub(BaseMatrixT& b, T p) { - applyBinary(binary::Sub1(p), b); -} - -DEFINE_MATRIX_BINARY_OP(Relu, b = a > 0.0f ? a : 0.0f); -template -void BaseMatrixT::relu(BaseMatrixT& b) { - applyBinary(binary::Relu(), b); -} - -#if defined(__ARM_NEON__) || defined(__ARM_NEON) -template <> -void BaseMatrixT::relu(BaseMatrixT& b) { - neon::relu(data_, b.data_, height_ * width_); -} -#endif - -DEFINE_MATRIX_BINARY_OP(ReluDerivative, a *= (b > 0.0f ? 1.0f : 0.0f)); -template -void BaseMatrixT::reluDerivative(BaseMatrixT& b) { - applyBinary(binary::ReluDerivative(), b); -} - -DEFINE_MATRIX_BINARY_OP(Softrelu, const T THRESHOLD = 40.0; - b = log(1.0 + exp((a > THRESHOLD) - ? THRESHOLD - : ((a < -THRESHOLD) ? (-THRESHOLD) - : a)))); -template <> -void BaseMatrixT::softrelu(BaseMatrixT& b) { - applyBinary(binary::Softrelu(), b); -} - -DEFINE_MATRIX_BINARY_OP( - SoftreluDerivative, const T THRESHOLD = 40.0; - a *= (1.0 - exp(-1.0 * ((b > THRESHOLD) - ? THRESHOLD - : ((b < -THRESHOLD) ? (-THRESHOLD) : b))))); -template <> -void BaseMatrixT::softreluDerivative(BaseMatrixT& b) { - applyBinary(binary::SoftreluDerivative(), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(Brelu, TWO_PARAMETER, b = a > p1 ? a : p1; - b = b < p2 ? b : p2); -template -void BaseMatrixT::brelu(BaseMatrixT& b) { - int p1 = 0, p2 = 24; //! TODO(yuyang18): Make p1,p2 configuable. - applyBinary(binary::Brelu(p1, p2), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(BreluDerivative, - TWO_PARAMETER, - a *= (b > p1 && b < p2) ? 1.0 : 0.0); -template -void BaseMatrixT::breluDerivative(BaseMatrixT& b) { - int p1 = 0, p2 = 24; - applyBinary(binary::BreluDerivative(p1, p2), b); -} - -DEFINE_MATRIX_BINARY_OP(Square, b = a * a); -template -void BaseMatrixT::square2(BaseMatrixT& b) { - applyBinary(binary::Square(), b); -} - -DEFINE_MATRIX_BINARY_OP(SquareDerivative, a *= 2.0 * b); -template -void BaseMatrixT::squareDerivative(BaseMatrixT& b) { - applyBinary(binary::SquareDerivative(), b); -} - -DEFINE_MATRIX_BINARY_OP(Tanh, T tmp = -2.0 * a; - tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp; - b = 2.0 / (1.0 + std::exp(tmp)) - 1.0); -template <> -void BaseMatrixT::tanh(BaseMatrixT& b) { - applyBinary(binary::Tanh(), b); -} - -DEFINE_MATRIX_BINARY_OP(TanhDerivative, a *= 1 - b * b); -template -void BaseMatrixT::tanhDerivative(BaseMatrixT& b) { - applyBinary(binary::TanhDerivative(), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP( - ScaledTanh, TWO_PARAMETER, b = p1 * (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0)); -template <> -void BaseMatrixT::scaledTanh(BaseMatrixT& b, real p1, real p2) { - applyBinary(binary::ScaledTanh(p1, p2), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(ScaledTanhDerivative, - TWO_PARAMETER, - a *= p2 * (p1 - b * b)); -template -void BaseMatrixT::scaledTanhDerivative(BaseMatrixT& b, T p1, T p2) { - applyBinary(binary::ScaledTanhDerivative(p1 * p1, p2 / p1), b); -} - -DEFINE_MATRIX_BINARY_OP(Reciprocal, b = 1.0f / a); -template -void BaseMatrixT::reciprocal2(BaseMatrixT& b) { - applyBinary(binary::Reciprocal(), b); -} - -DEFINE_MATRIX_BINARY_OP(ReciprocalDerivative, a *= -b * b); -template -void BaseMatrixT::reciprocalDerivative(BaseMatrixT& b) { - applyBinary(binary::ReciprocalDerivative(), b); -} - -DEFINE_MATRIX_BINARY_OP(Abs, b = a > 0.0f ? a : -a); -template -void BaseMatrixT::abs2(BaseMatrixT& b) { - applyBinary(binary::Abs(), b); -} - -DEFINE_MATRIX_BINARY_OP(AbsDerivative, a = (b > 0) ? a : (b < 0) ? -a : 0); -template -void BaseMatrixT::absDerivative(BaseMatrixT& b) { - applyBinary(binary::AbsDerivative(), b); -} - -DEFINE_MATRIX_BINARY_OP(Sigmoid, const T THRESHOLD_MIN = -40.0; - const T THRESHOLD_MAX = 13.0; - T tmp = (a < THRESHOLD_MIN) - ? THRESHOLD_MIN - : ((a > THRESHOLD_MAX) ? THRESHOLD_MAX : a); - b = 1.0f / (1.0f + exp(-tmp))); -template <> -void BaseMatrixT::sigmoid(BaseMatrixT& b) { - if (useGpu_) { - applyBinary(binary::Sigmoid(), b); - } else { // cpu versioni - size_t numSamples = this->height_; - size_t dim = this->width_; - CHECK_EQ(b.height_, numSamples); - CHECK_EQ(b.width_, dim); - const real* in = this->data_; - real* out = b.data_; - - // out = - in - const float THRESHOLD_MIN = -40.0; // make sure sigmoid(x) > 0 - const float THRESHOLD_MAX = 13.0; // make sure sigmoid(x) < 1 - for (size_t i = 0; i < numSamples * dim; ++i) { - real tmp = in[i]; - tmp = (tmp < THRESHOLD_MIN) - ? THRESHOLD_MIN - : ((tmp > THRESHOLD_MAX) ? THRESHOLD_MAX : tmp); - out[i] = -tmp; - } - - // out = exp(out) - vExp(numSamples * dim, out, out); - - // out = 1 / (1 + out) - for (size_t i = 0; i < numSamples * dim; ++i) { - out[i] = 1 / (1 + out[i]); - } - } -} - -DEFINE_MATRIX_BINARY_OP(SigmoidDerivative, a *= b * (1 - b)); -template -void BaseMatrixT::sigmoidDerivative(BaseMatrixT& b) { - applyBinary(binary::SigmoidDerivative(), b); -} - -DEFINE_MATRIX_BINARY_OP(ExpDerivative, a *= b); -template -void BaseMatrixT::expDerivative(BaseMatrixT& b) { - applyBinary(binary::ExpDerivative(), b); -} - -DEFINE_MATRIX_BINARY_OP(Sign, b = a > 0.0f ? 1.0f : -1.0f); -template -void BaseMatrixT::sign2(BaseMatrixT& b) { - applyBinary(binary::Sign(), b); -} - -DEFINE_MATRIX_BINARY_OP(Exp, a = exp(b)); -template <> -void BaseMatrixT::exp2(BaseMatrixT& b) { - applyBinary(binary::Exp(), b); -} - -DEFINE_MATRIX_BINARY_OP(Log, a = log(b)); -template <> -void BaseMatrixT::log2(BaseMatrixT& b) { - if (useGpu_) { - applyBinary(binary::Log(), b); - } else { - vLog(height_ * width_, b.data_, data_); - } -} - -DEFINE_MATRIX_BINARY_OP(Sqrt, a = sqrt(b)); -template <> -void BaseMatrixT::sqrt2(BaseMatrixT& b) { - applyBinary(binary::Sqrt(), b); -} - -DEFINE_MATRIX_BINARY_OP(InvSqrt, a = 1.0f / sqrt(b)); -template <> -void BaseMatrixT::invSqrt(BaseMatrixT& b) { - if (useGpu_) { - applyBinary(binary::InvSqrt(), b); - } else { // cpu branch - CHECK_EQ(height_, b.height_); - CHECK_EQ(width_, b.width_); - vInvSqrt(height_ * width_, b.data_, data_); - } -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(IsEqual, ONE_PARAMETER, a = (b == p)); -template -void BaseMatrixT::isEqualTo(BaseMatrixT& b, T value) { - applyBinary(binary::IsEqual(value), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(AddScalar, ONE_PARAMETER, a = b + p); -template -void BaseMatrixT::addScalar(BaseMatrixT& b, T p) { - applyBinary(binary::AddScalar(p), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(SubScalar, ONE_PARAMETER, a = b - p); -template -void BaseMatrixT::subScalar(BaseMatrixT& b, T p) { - applyBinary(binary::SubScalar(p), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(MulScalar, ONE_PARAMETER, a = b * p); -template -void BaseMatrixT::mulScalar(BaseMatrixT& b, T p) { - applyBinary(binary::MulScalar(p), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(DivScalar, ONE_PARAMETER, a = b / p); -template -void BaseMatrixT::divScalar(BaseMatrixT& b, T p) { - applyBinary(binary::DivScalar(p), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(ScalarDiv, ONE_PARAMETER, a = p / b); -template -void BaseMatrixT::scalarDiv(BaseMatrixT& b, T p) { - applyBinary(binary::ScalarDiv(p), b); -} - -/** - * @brief ternary operator. - * - */ - -DEFINE_MATRIX_TERNARY_OP(SoftCrossEntropy, - a = -c * log(b) - (1 - c) * log(1 - b)); -template <> -void BaseMatrixT::softCrossEntropy(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::SoftCrossEntropy(), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(SoftCrossEntropyBp, a += (b - c) / (b * (1 - b))); -template -void BaseMatrixT::softCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::SoftCrossEntropyBp(), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(BinaryCrossEntropy, - a = c > 0.5 ? -log(b) : -log(1.0 - b)); -template <> -void BaseMatrixT::binaryLabelCrossEntropy(BaseMatrixT& b, - BaseMatrixT& c) { - if (useGpu_) { - applyTernary(ternary::BinaryCrossEntropy(), b, c); - } else { - CHECK_EQ(height_, b.height_); - CHECK_EQ(height_, c.height_); - CHECK_EQ(width_, b.width_); - CHECK_EQ(width_, c.width_); - - size_t size = height_ * width_; - real* out = b.data_; - real* label = c.data_; - real* cost = data_; - - for (size_t i = 0; i < size; ++i) { - cost[i] = label[i] > 0.5 ? out[i] : 1.0 - out[i]; - } - vLog(size, cost, cost); - for (size_t i = 0; i < size; ++i) { - cost[i] *= -1.0; - } - } -} - -DEFINE_MATRIX_TERNARY_OP(BinaryCrossEntropyBp, - a += c > 0.5 ? -1.0 / b : 1.0 / (1.0 - b)); -template -void BaseMatrixT::binaryLabelCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::BinaryCrossEntropyBp(), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(Add, a = b + c); -template -void BaseMatrixT::add(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::Add(), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add1, TWO_PARAMETER, a = p1 * b + p2 * c); -template -void BaseMatrixT::add(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2) { - applyTernary(ternary::Add1(p1, p2), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(Sub, a = b - c); -template -void BaseMatrixT::sub(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::Sub(), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(Sub1, TWO_PARAMETER, a = p1 * b - p2 * c); -template -void BaseMatrixT::sub(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2) { - applyTernary(ternary::Sub1(p1, p2), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(Add2, a = a + b + c); -template -void BaseMatrixT::add2(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::Add2(), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add3, - THREE_PARAMETER, - a = p1 * a + p2 * b + p3 * c); -template -void BaseMatrixT::add2(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3) { - applyTernary(ternary::Add3(p1, p2, p3), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(SgdUpdate, - THREE_PARAMETER, - c = p2 * c - p1 * (b + p3 * a); - a = a + c); -template -void BaseMatrixT::sgdUpdate(BaseMatrixT& b, // grad - BaseMatrixT& c, // mom - T p1, // learningRate, - T p2, // momentum, - T p3) { // decayRate - applyTernary(ternary::SgdUpdate(p1, p2, p3), b, c); -} - -DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(SgdUpdate, - THREE_PARAMETER, - c = p2 * c - p1 * d * (b + p3 * a); - a += c); -template -void BaseMatrixT::sgdUpdate(BaseMatrixT& b, // grad, - BaseMatrixT& c, // mom, - BaseMatrixT& d, // lr, - T p1, // learningRate, - T p2, // momentum, - T p3) { // decayRate - applyQuaternary(quaternary::SgdUpdate(p1, p2, p3), b, c, d); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL1, ONE_PARAMETER, T lambda = p * b; - a = (a > lambda) - ? (a - lambda) - : (a < -lambda) ? (a + lambda) : 0); -template -void BaseMatrixT::applyL1(BaseMatrixT& lr, T learningRate, T decayRate) { - applyBinary(binary::ApplyL1(learningRate * decayRate), lr); -} - -template <> -void BaseMatrixT::applyL1(BaseMatrixT& lr, - real learningRate, - real decayRate) { - if (useGpu_) { - applyBinary(binary::ApplyL1(learningRate * decayRate), lr); - } else { - simd::decayL1(this->data_, - this->data_, - lr.data_, - learningRate * decayRate, - height_ * width_); - } -} - -DEFINE_MATRIX_UNARY_PARAMETER_OP(ApplyL1, ONE_PARAMETER, T lambda = p; - a = (a > lambda) - ? (a - lambda) - : (a < -lambda) ? (a + lambda) : 0); -template -void BaseMatrixT::applyL1(T learningRate, T decayRate) { - applyUnary(unary::ApplyL1(learningRate * decayRate)); -} - -template <> -void BaseMatrixT::applyL1(real learningRate, real decayRate) { - if (useGpu_) { - applyUnary(unary::ApplyL1(learningRate * decayRate)); - } else { - simd::decayL1( - this->data_, this->data_, learningRate * decayRate, height_ * width_); - } -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL2, - ONE_PARAMETER, - a *= (1.0f / (1.0f + p * b))); -template -void BaseMatrixT::applyL2(BaseMatrixT& lr, T learningRate, T decayRate) { - if (useGpu_) { - applyBinary(binary::ApplyL2(learningRate * decayRate), lr); - } else { - size_t size = this->height_ * this->width_; - T decay = learningRate * decayRate; - for (size_t j = 0; j < size; ++j) { - this->data_[j] *= 1.0f / (1.0f + decay * lr.data_[j]); - } - } -} - -template -void BaseMatrixT::applyL2(T learningRate, T decayRate) { - BaseMatrixT::mulScalar(1.0f / (1.0f + learningRate * decayRate)); -} - -DEFINE_MATRIX_BINARY_OP(DotMul, a *= b); -template -void BaseMatrixT::dotMul(BaseMatrixT& b) { - applyBinary(binary::DotMul(), b); -} - -DEFINE_MATRIX_TERNARY_OP(DotMul, a = b * c); -template -void BaseMatrixT::dotMul(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::DotMul(), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(DotDiv, a = (b == 0.0) ? 0.0 : b / c); -template -void BaseMatrixT::dotDiv(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::DotDiv(), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotDiv2P, - TWO_PARAMETER, - a = (b + p1) / (c + p2)); -template -void BaseMatrixT::dotDiv(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { - applyTernary(ternary::DotDiv2P(p1, p2), b, c); -} - -DEFINE_MATRIX_QUATERNARY_OP(RankLoss, const T THRESHOLD = 40.0; a = b - c; - a = (a > THRESHOLD) - ? THRESHOLD - : ((a < -THRESHOLD) ? (-THRESHOLD) : a); - a = log(1 + exp(a)) - a * d); -template <> -void BaseMatrixT::rankLoss(BaseMatrixT& b, - BaseMatrixT& c, - BaseMatrixT& d) { - applyQuaternary(quaternary::RankLoss(), b, c, d); -} - -DEFINE_MATRIX_QUATERNARY_OP(RankLossBp, const T THRESHOLD = 40.0; a = b - c; - a = (a > THRESHOLD) - ? THRESHOLD - : ((a < -THRESHOLD) ? (-THRESHOLD) : a); - a = exp(a); - a = (a / (1 + a) - d)); -template <> -void BaseMatrixT::rankLossBp(BaseMatrixT& b, - BaseMatrixT& c, - BaseMatrixT& d) { - applyQuaternary(quaternary::RankLossBp(), b, c, d); -} - -/* this = log(1 + exp(b)) - c * b */ -DEFINE_MATRIX_TERNARY_OP(LogisticRegressionLoss, const T THRESHOLD = 40.0; - T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) - ? -THRESHOLD - : b; - a = log(1 + exp(x)) - c * x); -template <> -void BaseMatrixT::logisticRegressionLoss(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::LogisticRegressionLoss(), b, c); -} - -/* this = exp(b)/(1+exp(b)) - c */ -DEFINE_MATRIX_TERNARY_OP(LogisticRegressionLossBp, const T THRESHOLD = 40.0; - T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) - ? -THRESHOLD - : b; - x = exp(x); - a = x / (1 + x) - c); -template <> -void BaseMatrixT::logisticRegressionLossBp(BaseMatrixT& b, - BaseMatrixT& c) { - applyTernary(ternary::LogisticRegressionLossBp(), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(BiggerThan, a = (b > c) ? 1.0f : 0.0f); -template -void BaseMatrixT::biggerThan(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::BiggerThan(), b, c); -} - -DEFINE_MATRIX_QUATERNARY_OP( - BiggerThan, a = ((b > c && d > 0.5f) || (b < c && d < 0.5f)) ? 1.0f : 0.0f); -template -void BaseMatrixT::biggerThan(BaseMatrixT& b, - BaseMatrixT& c, - BaseMatrixT& d) { - applyQuaternary(quaternary::BiggerThan(), b, c, d); -} - -DEFINE_MATRIX_TERNARY_OP(Max, a = (b > c) ? b : c); -template -void BaseMatrixT::max2(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::Max(), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(BinaryClassificationError, - ONE_PARAMETER, - c += ((a > p) == (b > p)) ? 0.0f : 1.0f); -template -void BaseMatrixT::binaryClassificationError2(size_t destCol, - BaseMatrixT& b, - BaseMatrixT& c, - T p) { - CHECK(!useGpu_) << "do not support gpu"; - MatrixOffset offset(0, 0, 0, 0, destCol, 0); - int numRows = b.height_; - int numCols = b.width_; - b.applyTernary(ternary::BinaryClassificationError(p), - c, - *this, - numRows, - numCols, - offset, - false_type(), - true_type() /*cAsColVector*/); -} - -template <> -void BaseMatrixT::binaryClassificationError(size_t destCol, - BaseMatrixT& b, - BaseMatrixT& c, - real p) { - MatrixOffset offset(destCol, 0, 0, 0, 0, 0); - int numRows = b.height_; - int numCols = b.width_; - aggregate(aggregate::sum(), - base::binary::classificationError(p), - base::binary::add(), - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*aAsColVector*/); -} - -DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(Add3, - THREE_PARAMETER, - a = p1 * b + p2 * c + p3 * d); -template -void BaseMatrixT::add3( - BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d, T p1, T p2, T p3) { - applyQuaternary(quaternary::Add3(p1, p2, p3), b, c, d); -} - -DEFINE_MATRIX_TERNARY_OP(DotMulSquare, a = b * c * c); -template -void BaseMatrixT::dotMulSquare(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::DotMulSquare(), b, c); -} - -DEFINE_MATRIX_TERNARY_OP(DotSquareSquare, a = b * b * c * c); -template -void BaseMatrixT::dotSquareSquare(BaseMatrixT& b, BaseMatrixT& c) { - applyTernary(ternary::DotSquareSquare(), b, c); -} - -DEFINE_MATRIX_BINARY_OP(DotMulSquare, a *= b * b); -template -void BaseMatrixT::dotMulSquare(BaseMatrixT& b) { - applyBinary(binary::DotMulSquare(), b); -} - -DEFINE_MATRIX_BINARY_OP(DotSquareMul, a = a * a * b); -template -void BaseMatrixT::dotSquareMul(BaseMatrixT& b) { - applyBinary(binary::DotSquareMul(), b); -} - -DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(AddSquareSum, - THREE_PARAMETER, - T tmp = p1 * b + p2 * c + p3 * d; - a += tmp * tmp); -template -void BaseMatrixT::addSquareSum( - BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT d, T p1, T p2, T p3) { - applyQuaternary(quaternary::AddSquareSum(p1, p2, p3), b, c, d); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(AddSquare, ONE_PARAMETER, a += p * b * b); -template -void BaseMatrixT::addSquare(BaseMatrixT& b, T p) { - applyBinary(binary::AddSquare(p), b); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(DecayAddSquare, - TWO_PARAMETER, - a = p1 * a + p2 * b * b); -template -void BaseMatrixT::decayAddSquare(BaseMatrixT& b, T p1, T p2) { - applyBinary(binary::DecayAddSquare(p1, p2), b); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DecayAddSquareMul, - TWO_PARAMETER, - a = p1 * a + p2 * b * b * c * c); -template -void BaseMatrixT::decayAddSquareMul(BaseMatrixT& b, - BaseMatrixT& c, - T p1, - T p2) { - applyTernary(ternary::DecayAddSquareMul(p1, p2), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(ReciprocalSum, - THREE_PARAMETER, - a = 1 / (p1 * b + p2 * c + p3)); -template -void BaseMatrixT::reciprocalSum( - BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3) { - applyTernary(ternary::ReciprocalSum(p1, p2, p3), b, c); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(Reciprocal2, - TWO_PARAMETER, - a = 1 / (p1 * b + p2)); -template -void BaseMatrixT::reciprocal2(BaseMatrixT& b, T p1, T p2) { - applyBinary(binary::Reciprocal2(p1, p2), b); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSquareSum, - TWO_PARAMETER, - T tmp = p1 * b + p2 * c; - a *= tmp * tmp); -template -void BaseMatrixT::dotMulSquareSum(BaseMatrixT& b, - BaseMatrixT& c, - T p1, - T p2) { - applyTernary(ternary::DotMulSquareSum(p1, p2), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotSquareSum, - TWO_PARAMETER, - T tmp = p1 * b + p2 * c; - a = tmp * tmp); -template -void BaseMatrixT::dotSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { - applyTernary(ternary::DotSquareSum(p1, p2), b, c); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSum, - TWO_PARAMETER, - a *= p1 * b + p2 * c); -template -void BaseMatrixT::dotMulSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { - applyTernary(ternary::DotMulSum(p1, p2), b, c); -} - -DEFINE_MATRIX_BINARY_OP(CopyAndClear, b = a; a = 0); -template -void BaseMatrixT::copyAndClear(BaseMatrixT& b) { - applyBinary(binary::CopyAndClear(), b); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(AddDotMul, - TWO_PARAMETER, - a = p1 * a + p2 * b * c); -template -void BaseMatrixT::addDotMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { - applyTernary(ternary::AddDotMul(p1, p2), b, c); -} - -DEFINE_MATRIX_BINARY_OP(Assign, a = b;); -template -void BaseMatrixT::assign(BaseMatrixT& b) { - if (useGpu_) { - applyBinary(binary::Assign(), b); - } else { // cpu version - CHECK_EQ(this->height_, b.height_); - CHECK_EQ(this->width_, b.width_); - memcpy(data_, b.data_, sizeof(T) * height_ * width_); - } -} - -template -void BaseMatrixT::assignAtOffset(BaseMatrixT& b, int64_t columnOffset) { - if (columnOffset + b.width_ <= width_) { - int numRows = height_; - int numCols = b.width_; - MatrixOffset offset(columnOffset, 0, 0, 0); - applyBinary(binary::Assign(), b, numRows, numCols, offset); - } else if (columnOffset + width_ <= b.width_) { - int numRows = height_; - int numCols = width_; - MatrixOffset offset(0, 0, columnOffset, 0); - applyBinary(binary::Assign(), b, numRows, numCols, offset); - } else { - LOG(FATAL) << "Wrong argument " - << " a.width=" << width_ << " b.width=" << b.width_ - << " columnOffset=" << columnOffset; - } -} - -DEFINE_MATRIX_BINARY_OP(DeepSwap, T tmp = a; a = b; b = tmp); -template -void BaseMatrixT::deepSwap(BaseMatrixT& b) { - applyBinary(binary::DeepSwap(), b); -} - -template <> -void BaseMatrixT::rowDotMul(size_t destCol, - BaseMatrixT& b, - BaseMatrixT& c) { - int numRows = b.height_; - int numCols = b.width_; - MatrixOffset offset(destCol, 0, 0, 0, 0, 0); - aggregate(aggregate::sum(), - base::binary::mul(), - base::binary::add(), - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*aAsColVector*/); -} - -template -void BaseMatrixT::rowDotMul2(size_t destCol, - BaseMatrixT& b, - BaseMatrixT& c) { - CHECK(!useGpu_) << "do not support gpu"; - - size_t height = this->height_; - CHECK_LT(destCol, this->width_); - CHECK_EQ(height, b.height_); - CHECK_EQ(height, c.height_); - CHECK_EQ(b.width_, c.width_); - size_t width = b.width_; - T* A = this->data_; - const T* B = b.data_; - const T* C = c.data_; - for (size_t i = 0; i < height; - ++i, A += this->width_, B += width, C += width) { - for (size_t j = 0; j < width; ++j) { - A[destCol] += B[j] * C[j]; - } - } -} - -template <> -void BaseMatrixT::addDotMulVMM(BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - int numRows = b.height_; - int numCols = b.width_; - aggregate(aggregate::sum(), - base::binary::mul(), - base::binary::add(), - b, - c, - numRows, - numCols, - offset, - true_type() /*aAsRowVector*/, - false_type()); -} - -template -void BaseMatrixT::addDotMulVMM2(BaseMatrixT& b, BaseMatrixT& c) { - CHECK(!useGpu_) << "do not support gpu"; - - CHECK_EQ(height_, 1LU); - CHECK_EQ(b.height_, c.height_); - CHECK_EQ(width_, b.width_); - CHECK_EQ(width_, c.width_); - size_t height = b.height_; - size_t width = b.width_; - T* A = this->data_; - const T* B = b.data_; - const T* C = c.data_; - for (size_t i = 0; i < height; ++i, B += width, C += width) { - for (size_t j = 0; j < width; ++j) { - A[j] += B[j] * C[j]; - } - } -} - -DEFINE_MATRIX_TERNARY_OP(addDotMulMMV, a += b * c); -template -void BaseMatrixT::addDotMulMMV(BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::addDotMulMMV(), - b, - c, - numRows, - numCols, - offset, - true_type() /*cAsRowVector*/, - false_type()); -} - -template -void BaseMatrixT::addDotMulMMV2(BaseMatrixT& b, BaseMatrixT& c) { - CHECK(!useGpu_) << "do not support gpu"; - - CHECK_EQ(c.height_, 1LU); - CHECK_EQ(height_, b.height_); - CHECK_EQ(width_, b.width_); - CHECK_EQ(width_, c.width_); - size_t height = height_; - size_t width = width_; - T* A = this->data_; - const T* B = b.data_; - const T* C = c.data_; - for (size_t i = 0; i < height; ++i, A += width, B += width) { - for (size_t j = 0; j < width; ++j) { - A[j] += B[j] * C[j]; - } - } -} - -template -void BaseMatrixT::rowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, cCol, 0); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::DotMul(), - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*cAsColVector*/); -} - -template -void BaseMatrixT::rowScale2(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { - CHECK(!useGpu_) << "do not support gpu"; - - size_t height = this->height_; - size_t width = this->width_; - CHECK_EQ(height, b.height_); - CHECK_EQ(width, b.width_); - CHECK_LT(cCol, c.width_); - CHECK_EQ(height, c.height_); - T* A = this->data_; - const T* B = b.data_; - const T* C = c.data_; - for (size_t i = 0; i < height; ++i, A += width, B += width, C += c.width_) { - for (size_t j = 0; j < width; ++j) { - A[j] = B[j] * C[cCol]; - } - } -} - -template -void BaseMatrixT::colScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, 0, cRow); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::DotMul(), - b, - c, - numRows, - numCols, - offset, - true_type() /* cAsRowVector */, - false_type() /* cAsColVector */); -} - -template -void BaseMatrixT::addColScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, 0, cRow); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::addDotMulMMV(), - b, - c, - numRows, - numCols, - offset, - true_type() /* cAsRowVector */, - false_type() /* cAsColVector */); -} - -template -void BaseMatrixT::addRowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, cCol, 0); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::addDotMulMMV(), - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*cAsColVector*/); -} - -DEFINE_MATRIX_TERNARY_PARAMETER_OP(RowAdd, ONE_PARAMETER, a = b + p * c); -template -void BaseMatrixT::rowAdd(size_t cCol, BaseMatrixT& b, BaseMatrixT& c, T p) { - MatrixOffset offset(0, 0, 0, 0, cCol, 0); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::RowAdd(p), - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*cAsColVector*/); -} - -DEFINE_MATRIX_TERNARY_OP(RowPow, a = pow(b, c)); -template <> -void BaseMatrixT::rowPow(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { - if (useGpu_) { - MatrixOffset offset(0, 0, 0, 0, cCol, 0); - int numRows = height_; - int numCols = width_; - applyTernary(ternary::RowPow(), - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*cAsColVector*/); - } else { - size_t height = this->height_; - size_t width = this->width_; - CHECK_EQ(height, b.height_); - CHECK_EQ(width, b.width_); - CHECK_LT(cCol, c.width_); - CHECK_EQ(height, c.height_); - real* A = this->data_; - const real* B = b.data_; - const real* C = c.data_; - for (size_t i = 0; i < height; ++i, A += width, B += width, C += c.width_) { - vPow(width, B, C[cCol], A); - } - } -} - -template -void BaseMatrixT::mulRowVector(BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::DotMul(), - b, - numRows, - numCols, - offset, - true_type() /* bAsRowVector */, - false_type()); -} - -DEFINE_MATRIX_BINARY_OP(DotDiv, a /= b); -template -void BaseMatrixT::divRowVector(BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::DotDiv(), - b, - numRows, - numCols, - offset, - true_type() /* bAsRowVector */, - false_type()); -} - -template -void BaseMatrixT::mulColVector(BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::DotMul(), - b, - numRows, - numCols, - offset, - false_type(), - true_type() /* bAsColVector */); -} - -template -void BaseMatrixT::divColVector(BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0); - int numRows = height_; - int numCols = width_; - applyBinary(binary::DotDiv(), - b, - numRows, - numCols, - offset, - false_type(), - true_type() /* bAsColVector */); -} - -template <> -template -int BaseMatrixT::applyRow(Agg agg, BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - size_t numRows = b.height_; - size_t numCols = b.width_; - CHECK_EQ(height_, numRows); - CHECK_EQ(width_, 1UL); - aggregate(agg, - base::unary::identity(), - base::binary::second(), - b, - numRows, - numCols, - offset, - false_type(), - true_type() /*aAsColVector*/); - - return 0; -} - -template <> -template -int BaseMatrixT::applyRow(Agg agg, Saver sv, BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - size_t numRows = b.height_; - size_t numCols = b.width_; - CHECK_EQ(height_, numRows); - CHECK_EQ(width_, 1UL); - aggregate(agg, - base::unary::identity(), - sv, - b, - numRows, - numCols, - offset, - false_type(), - true_type() /*aAsColVector*/); - - return 0; -} - -template <> -template -int BaseMatrixT::applyRow(Agg agg, - real scaleDest, - real scaleAgg, - BaseMatrixT& b) { - if (scaleDest != 0) { - applyRow(agg, base::binary::add2(scaleDest, scaleAgg), b); - } else { - applyRow(agg, base::binary::second(), b); - if (scaleAgg != 1) { - mulScalar(scaleAgg); - } - } - return 0; -} - -template <> -template -int BaseMatrixT::applyRow( - Agg agg, Op op, Saver sv, BaseMatrixT& b, BaseMatrixT& c) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - size_t numRows = b.height_; - size_t numCols = b.width_; - CHECK_EQ(height_, numRows); - CHECK_EQ(width_, 1UL); - CHECK_EQ(c.height_, numRows); - CHECK_EQ(c.width_, numCols); - aggregate(agg, - op, - sv, - b, - c, - numRows, - numCols, - offset, - false_type(), - true_type() /*aAsColVector*/); - return 0; -} - -template <> -template -int BaseMatrixT::applyRow(Agg agg, - Op op, - real scaleDest, - real scaleAgg, - BaseMatrixT& b, - BaseMatrixT& c) { - if (scaleDest != 0) { - applyRow(agg, op, base::binary::add2(scaleDest, scaleAgg), b, c); - } else { - applyRow(agg, op, base::binary::second(), b, c); - if (scaleAgg != 1) { - mulScalar(scaleAgg); - } - } - return 0; -} - -template <> -template -int BaseMatrixT::applyCol(Agg agg, BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - size_t numRows = b.height_; - size_t numCols = b.width_; - CHECK_EQ(width_, numCols); - CHECK_EQ(height_, 1UL); - aggregate(agg, - base::unary::identity(), - base::binary::second(), - b, - numRows, - numCols, - offset, - true_type() /*aAsRowVector*/, - false_type()); - - return 0; -} - -template <> -template -int BaseMatrixT::applyCol(Agg agg, Saver sv, BaseMatrixT& b) { - MatrixOffset offset(0, 0, 0, 0, 0, 0); - size_t numRows = b.height_; - size_t numCols = b.width_; - CHECK_EQ(width_, numCols); - CHECK_EQ(height_, 1UL); - aggregate(agg, - base::unary::identity(), - sv, - b, - numRows, - numCols, - offset, - true_type() /*aAsRowVector*/, - false_type()); - - return 0; -} - -template <> -template -int BaseMatrixT::applyCol(Agg agg, - real scaleDest, - real scaleAgg, - BaseMatrixT& b) { - if (scaleDest != 0) { - applyCol(agg, base::binary::add2(scaleDest, scaleAgg), b); - } else { - applyCol(agg, base::binary::second(), b); - if (scaleAgg != 1) { - mulScalar(scaleAgg); - } - } - return 0; -} - -template <> -void BaseMatrixT::sumRows(BaseMatrixT& b, real scaleSum, real scaleDest) { - applyRow(aggregate::sum(), scaleDest, scaleSum, b); -} - -template <> -void BaseMatrixT::maxRows(BaseMatrixT& b) { - applyRow(aggregate::max(), b); -} - -template <> -void BaseMatrixT::minRows(BaseMatrixT& b) { - applyRow(aggregate::min(), b); -} - -template <> -void BaseMatrixT::maxCols(BaseMatrixT& b) { - applyCol(aggregate::max(), b); -} - -template <> -void BaseMatrixT::minCols(BaseMatrixT& b) { - applyCol(aggregate::min(), b); -} - -template <> -void BaseMatrixT::sumCols(BaseMatrixT& b, real scaleSum, real scaleDest) { - applyCol(aggregate::sum(), scaleDest, scaleSum, b); -} - -template <> -void BaseMatrixT::sumOfSquaredDiffs(BaseMatrixT& b, - BaseMatrixT& c, - real scaleSum, - real scaleDest) { - applyRow( - aggregate::sum(), base::binary::squaredDiff(), scaleDest, scaleSum, b, c); -} - -template <> -void BaseMatrixT::sumOfProducts(BaseMatrixT& b, - BaseMatrixT& c, - real scaleSum, - real scaleDest) { - applyRow(aggregate::sum(), base::binary::mul(), scaleDest, scaleSum, b, c); -} - -template class BaseMatrixT; - -#ifndef PADDLE_MOBILE_INFERENCE - -template class BaseMatrixT; - -#else - -template <> -void BaseMatrixT::zero() { - applyUnary(unary::Zero()); -} - -template <> -void BaseMatrixT::assign(int p) { - applyUnary(unary::Assign(p)); -} - -template <> -void BaseMatrixT::isEqualTo(BaseMatrixT& b, int value) { - applyBinary(binary::IsEqual(value), b); -} - -template <> -void BaseMatrixT::neg() { - applyUnary(unary::Neg()); -} - -template <> -void BaseMatrixT::abs2() { - applyUnary(unary::Abs()); -} - -template <> -void BaseMatrixT::add(int p) { - applyUnary(unary::Add(p)); -} - -template <> -void BaseMatrixT::add(int p1, int p2) { - applyUnary(unary::Add2(p1, p2)); -} - -template <> -void BaseMatrixT::applyL1(int learningRate, int decayRate) { - applyUnary(unary::ApplyL1(learningRate * decayRate)); -} - -#endif -} // namespace paddle diff --git a/paddle/math/BaseMatrix.h b/paddle/math/BaseMatrix.h deleted file mode 100644 index 1958629aa0354fcc332b1e5677a64c29397e0d26..0000000000000000000000000000000000000000 --- a/paddle/math/BaseMatrix.h +++ /dev/null @@ -1,1095 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include "TensorExpression.h" -#include "paddle/utils/Common.h" - -namespace paddle { - -/* - * nvcc currently does not support C++11, - * so I realized false_type and true_type. - */ -template -struct bool_constant { - static const T value = v; -}; -typedef bool_constant false_type; -typedef bool_constant true_type; - -/** - * @brief Calculate matrix element address. - * - * For instance, address of A[i][j] = i * ld + j. - * - */ -#define CAL_MATRIX_START_ADDRESS(address, height, width, ld, col, row) \ - CHECK_LE(col, width); \ - CHECK_LE(row, height); \ - address += row * ld + col; - -class MatrixOffset { - public: - size_t aCol_; - size_t aRow_; - size_t bCol_; - size_t bRow_; - size_t cCol_; - size_t cRow_; - size_t dCol_; - size_t dRow_; - MatrixOffset(size_t aCol = 0, - size_t aRow = 0, - size_t bCol = 0, - size_t bRow = 0, - size_t cCol = 0, - size_t cRow = 0, - size_t dCol = 0, - size_t dRow = 0) - : aCol_(aCol), - aRow_(aRow), - bCol_(bCol), - bRow_(bRow), - cCol_(cCol), - cRow_(cRow), - dCol_(dCol), - dRow_(dRow) {} -}; - -template -class BaseMatrixT : public TensorExpression, T> { - public: - size_t height_, width_; - size_t stride_; - T* data_; - bool trans_; - bool useGpu_; - - public: - virtual ~BaseMatrixT() {} - BaseMatrixT(size_t height, size_t width, T* data, bool trans, bool useGpu) - : height_(height), - width_(width), - stride_(width), - data_(data), - trans_(trans), - useGpu_(useGpu) {} - - /** - * @note This constructor is for temporarily making a matrix with different - * useGpu flag as the original matrix so that mixed gpu/cpu operations - * can be performed successfully. - */ - BaseMatrixT(BaseMatrixT& mat, bool useGpu) - : height_(mat.height_), - width_(mat.width_), - stride_(mat.stride_), - data_(mat.data_), - trans_(mat.trans_), - useGpu_(useGpu) {} - - BaseMatrixT(size_t height, - size_t width, - size_t stride, - T* data, - bool trans, - bool use_gpu) - : height_(height), - width_(width), - stride_(stride), - data_(data), - trans_(trans), - useGpu_(use_gpu) { - /* CHECK_LE(width_, stride_); */ - } - - /// caller should make sure that the size of data is at least height*width - void setData(T* data) { data_ = data; } - - /** - * unary operator: element wise op(a). - * - * @code - * for 0 <= i < this->height_ & for 0 <= j < this->width_. - * @endcode - */ - template - int applyUnary(Op op); - - /** - * unary operator: element wise op(a). - * - * @code - * for 0 <= i < numRows & for 0 <= j < numCols. - * While matrix start address is: - * A = this->data_ + offset.aRow_*ld + offset.aCol_; - * @endcode - */ - template - int applyUnary(Op op, int numRows, int numCols, MatrixOffset& offset); - - /** - * binary operator: element wise op(a, b). - * - * @code - * for 0 <= i < this->height_ & for 0 <= j < this->width_. - * While this->height_ == b.height_ && this->width_ == b.width_. - * @endcode - */ - template - int applyBinary(Op op, BaseMatrixT& b); - - /** - * binary operator: element wise op(a, b) - * - * @code - * for 0 <= i < numRows & for 0 <= j < numCols. - * While matrix start address is: - * A = this->data_ + offset.aRow_*lda + offset.aCol_; - * B = b->data_ + offset.bRow_*ldb + offset.bCol_; - * - * if (bAsRowVector == false_type && bAsColVector == false_type) - * op(A[i * lda + j], B[i * ldb + j]) - * - * if (bAsRowVector == true_type && bAsColVector == false_type) - * op(A[i * lda + j], B[j]) - * - * if (bAsRowVector == false_type && bAsColVector == true_type) - * op(A[i * lda + j], B[i * ldb]) - * - * if (bAsRowVector == true_type && bAsColVector == true_type) - * op(A[i * lda + j], B[0]) - * @endcode - */ - template - int applyBinary(Op op, - BaseMatrixT& b, - int numRows, - int numCols, - MatrixOffset& offset, - bAsRowVector, - bAsColVector); - - template - int applyBinary( - Op op, BaseMatrixT& b, int numRows, int numCols, MatrixOffset& offset); - - /** - * ternary operator: element wise op(a, b, c). - * - * @code - * for 0 <= i < this->height_ & for 0 <= j < this->width_. - * - * While this->height_ == b.height_ && this->width_ == b.width_ - * && this->height_ == c.height_ && this->width_ == c.width_ - * @endcode - */ - template - int applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c); - - /** - * ternary operator: element wise op(a, b, c). - * - * @code - * for 0 <= i < numRows & for 0 <= j < numCols. - * While matrix start address is: - * - * A = this->data_ + offset.aRow_*lda + offset.aCol_; - * B = b->data_ + offset.bRow_*ldb + offset.bCol_; - * C = c->data_ + offset.cRow_*ldc + offset.cCol_; - * - * if (cAsRowVector == false_type && cAsColVector == false_type) - * op(A[i*lda + j], B[i*ldb + j], C[i*ldc + j]) - * - * if (cAsRowVector == true_type && cAsColVector == false_type) - * op(A[i*lda + j], B[i*ldb + j], C[j]) - * - * if (cAsRowVector == false_type && cAsColVector == true_type) - * op(A[i*lda + j], B[i*ldb + j], C[i*ldc]) - * - * if (cAsRowVector == 1 && cAsColVector == 1) - * op(A[i*lda + j], B[i*ldb + j], C[0]) - * @endcode - */ - template - int applyTernary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - int numRows, - int numCols, - MatrixOffset& offset, - cAsRowVector, - cAsColVector); - - template - int applyTernary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - int numRows, - int numCols, - MatrixOffset& offset); - - /** - * quaternary operator: element wise op(a, b, c, d). - * - * @code - * for 0 <= i < this->height_ & for 0 <= j < this->width_. - * - * While this->height_ == b.height_ && this->width_ == b.width_ - * && this->height_ == c.height_ && this->width_ == c.width_ - * && this->height_ == d.height_ && this->width_ == d.width_ - * @endcode - */ - template - int applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d); - - /** - * quaternary operator: element wise op(a, b, c, d). - * - * @code - * for 0 <= i < numRows & for 0 <= j < numCols. - * While matrix start address is: - * A = this->data_ + offset.aRow_*lda + offset.aCol_; - * B = b->data_ + offset.bRow_*ldb + offset.bCol_; - * C = c->data_ + offset.cRow_*ldc + offset.cCol_; - * D = d->data_ + offset.dRow_*ldd + offset.dCol_; - * @endcode - */ - template - int applyQuaternary(Op op, - BaseMatrixT& b, - BaseMatrixT& c, - BaseMatrixT& d, - int numRows, - int numCols, - MatrixOffset& offset); - - /** - * a aggregate expression that apply each row(or column) of matrix b. - * op and sv is element wise operator. - * - * @code - * if (aAsRowVector == true_type && aAsColVector == false_type) - * for each column j & 0 <= i < numRows, do: - * dst = agg(op(b[i*ldb + j])) - * a[j] = sv(a[j], dst) - * - * if (aAsRowVector == false_type && aAsColVector == true_type) - * for each row i & 0 <= j < numCols, do: - * dst = agg(op(b[i*ldb + j])) - * a[i] = sv(a[i], dst) - * @endcode - */ - template - int aggregate(Agg agg, - Op op, - Saver sv, - BaseMatrixT& b, - int numRows, - int numCols, - MatrixOffset& offset, - aAsRowVector, - aAsColVector); - - /** - * a aggregate expression that apply each row(or column) of matrix b and c. - * - * op and sv is element wise operator. - * - * @code - * if (aAsRowVector == true_type && aAsColVector == false_type) - * for each column j & 0 <= i < numRows, do: - * dst = agg(op(b[i*ldb + j], c[i*ldc + j])) - * a[j] = sv(a[j], dst) - * - * if (aAsRowVector == false_type && aAsColVector == true_type) - * for each row i & 0 <= j < numCols, do: - * dst = agg(op(b[i*ldb + j], c[i*ldc + j])) - * a[i] = sv(a[i], dst) - * @endcode - */ - template - int aggregate(Agg agg, - Op op, - Saver sv, - BaseMatrixT& b, - BaseMatrixT& c, - int numRows, - int numCols, - MatrixOffset& offset, - aAsRowVector, - aAsColVector); - - /** - * a aggregate expression that apply each row of matrix b. - * - * @code - * for each row i & 0 <= j < b.width_, do: - * this[i] = agg(b[i*ldb + j]) - * @endcode - */ - template - int applyRow(Agg agg, BaseMatrixT& b); - - /** - * a aggregate expression that apply each row of matrix b. - * - * @code - * for each row i & 0 <= j < b.width_, do: - * dst = agg(op(b[i*ldb + j], c[i*ldc + j]) - * this[i] = sv(this[i], dst) - * @endcode - */ - template - int applyRow(Agg agg, Op op, Saver sv, BaseMatrixT& b, BaseMatrixT& c); - - // Same as the above with the special handing of sv=add2(scaleDest, scaleAgg) - template - int applyRow(Agg agg, - Op op, - real scaleDest, - real scaleAgg, - BaseMatrixT& b, - BaseMatrixT& c); - - /** - * a aggregate expression that apply each row of matrix b. - * - * @code - * for each row i & 0 <= j < b.width_, do: - * dst = agg(b[i*ldb + j]) - * this[i] = sv(this[i], dst) - * @endcode - */ - template - int applyRow(Agg agg, Saver sv, BaseMatrixT& b); - - // Same as the above with the special handing of sv=add2(scaleDest, scaleAgg) - template - int applyRow(Agg agg, real scaleDest, real scaleAgg, BaseMatrixT& b); - - /** - * a aggregate expression that apply each column of matrix b. - * - * @code - * for each column j & 0 <= i < b.height_, do: - * this[j] = agg(b[i*ldb + j]) - * @endcode - */ - template - int applyCol(Agg agg, BaseMatrixT& b); - - /** - * a aggregate expression that apply each column of matrix b. - * - * @code - * for each column j & 0 <= i < b.height_, do: - * dst = agg(b[i*ldb + j]) - * this[j] = sv(this[j], dst) - * @endcode - */ - template - int applyCol(Agg agg, Saver sv, BaseMatrixT& b); - - // Same as the above with the special handing of sv=add2(scaleDest, scaleAgg) - template - int applyCol(Agg agg, real scaleDest, real scaleAgg, BaseMatrixT& b); - - bool useGpu() const { return useGpu_; } - - const T* rowBuf(size_t row) const { return data_ + width_ * row; } - - T* rowBuf(size_t row) { return data_ + width_ * row; } - - /** - * @brief unary operator. - * - */ - void neg(); - void exp2(); - void pow2(T p); - void log2(); - void sqrt2(); - void square2(); - void reciprocal2(); - void abs2(); - void sign2(); - void zero(); - - /** - * @code - * this(row, col + columnOffset) = 0 for 0 <= col < numColumns - * @endcode - */ - void zeroAtOffset(int64_t columnOffset, int64_t numColumns); - void one(); - void subScalar(T p); - void mulScalar(T p); - void divScalar(T p); - - /** - * @code - * this = p - * @endcode - */ - void assign(T p); - - /** - * @code - * swap(this, b) - * example: swap two Matrices - * MatrixPtr cpuA = std::make_shared(height, width); - * MatrixPtr cpuB = std::make_shared(height, width); - * cpuA->deepSwap(*cpuB); - * @endcode - */ - void deepSwap(BaseMatrixT& b); - - /** - * @code - * this = this + p - * @endcode - */ - void add(T p); - - /** - * @code - * this = this*p1 + p2 - * @endcode - */ - void add(T p1, T p2); - - /** - * this = this < low ? low : this - * - * this = this > high ? high : this - */ - void clip(T p1, T p2); - - /** - * this = b < low ? 0 : 1 - * - * this = b > high ? 0 : 1 - */ - void clipDerivative(BaseMatrixT& b, T p1, T p2); - - /** - * @code - * a = a > p ? 1.0f : 0.0f - * @endcode - */ - void biggerThanScalar(T p); - - /** - * @code - * a = a > p ? a : p - * @endcode - */ - void downClip(T p); - - /** - * @code - * this = b - * @endcode - */ - void assign(BaseMatrixT& b); - - /** - * @code - * If b.width + columOffset <= this.width - * this(row, col + columnOffset) = b(row, col) for 0 <= col < b.width - * - * If this.width + columnOffset <= b.width - * this(row, col) = b(row, col + columnOffset) for 0 <= col < this.width - * - * Otherwise, FATAL - * @endcode - */ - void assignAtOffset(BaseMatrixT& b, int64_t columnOffset); - - /// this = this + b - void add(BaseMatrixT& b); - - /** - * @code - * If b.width + columOffset <= this.width - * this(row, col + columnOffset) += b(row, col) for 0 <= col < b.width - * - * If this.width + columnOffset <= b.width - * this(row, col) += b(row, col + columnOffset) for 0 <= col < this.width - * - * Otherwise, FATAL - * @endcode - */ - void addAtOffset(BaseMatrixT& b, int64_t columnOffset); - - void addColVector(BaseMatrixT& b); - void addRowVector(BaseMatrixT& b); - void addBias(BaseMatrixT& b, T scale); - - void mulRowVector(BaseMatrixT& b); - void divRowVector(BaseMatrixT& b); - - void mulColVector(BaseMatrixT& b); - void divColVector(BaseMatrixT& b); - - void addP2P(BaseMatrixT& b); - - /** - * @code - * this = this + b*p - * @endcode - */ - void add(BaseMatrixT& b, T p); - - /** - * @code - * this = p1*this + p2*b - * @endcode - */ - void add(BaseMatrixT& b, T p1, T p2); - - /** - * @code - * this = this - b - * @endcode - */ - void sub(BaseMatrixT& b); - - /** - * @code - * this = this - b*p - * @endcode - */ - void sub(BaseMatrixT& b, T p); - - /** - * @code - * b = max(0, this) - * @endcode - */ - void relu(BaseMatrixT& b); - void reluDerivative(BaseMatrixT& b); - - /** - * @code - * b = log(1.0 + exp(this)) - * @endcode - */ - void softrelu(BaseMatrixT& b); - void softreluDerivative(BaseMatrixT& b); - - /** - * @code - * b = min(max(this, p1), p2) - * @endcode - */ - void brelu(BaseMatrixT& b); - void breluDerivative(BaseMatrixT& b); - - /** - * @code - * b = this * this - * @endcode - */ - void square2(BaseMatrixT& b); - void squareDerivative(BaseMatrixT& b); - - /** - * @code - * b = tanh(this) - * @endcode - */ - void tanh(BaseMatrixT& b); - void tanhDerivative(BaseMatrixT& b); - - /** - * @code - * b = p1 * tanh(p2 * this) - * @endcode - */ - void scaledTanh(BaseMatrixT& b, T p1, T p2); - void scaledTanhDerivative(BaseMatrixT& b, T p1, T p2); - - /** - * @code - * b = 1.0f / this - * @endcode - */ - void reciprocal2(BaseMatrixT& b); - void reciprocalDerivative(BaseMatrixT& b); - - /** - * @code - * b = this > 0.0f ? this : -this - * @endcode - */ - void abs2(BaseMatrixT& b); - void absDerivative(BaseMatrixT& b); - - /** - * @code - * b = 1.0f / (1.0f + exp(-this)) - * @endcode - */ - void sigmoid(BaseMatrixT& b); - void sigmoidDerivative(BaseMatrixT& b); - - /** - * @code - * b = a - * @endcode - */ - void expDerivative(BaseMatrixT& b); - - void sign2(BaseMatrixT& b); - - void exp2(BaseMatrixT& b); - void pow2(BaseMatrixT& b, T p); - void log2(BaseMatrixT& b); - void sqrt2(BaseMatrixT& b); - void addScalar(BaseMatrixT& b, T p); - void subScalar(BaseMatrixT& b, T p); - void mulScalar(BaseMatrixT& b, T p); - void divScalar(BaseMatrixT& b, T p); - void scalarDiv(BaseMatrixT& b, T p); - - /** - * @code - * this = 1.0f / sqrt(b) - * @endcode - */ - void invSqrt(BaseMatrixT& b); - - /// this = (b == value) - void isEqualTo(BaseMatrixT& b, T value); - - /** - * @brief ternary operator. - */ - void softCrossEntropy(BaseMatrixT& b, BaseMatrixT& c); - void softCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c); - void binaryLabelCrossEntropy(BaseMatrixT& b, BaseMatrixT& c); - void binaryLabelCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = b + c - * @endcode - */ - void add(BaseMatrixT& b, BaseMatrixT& c); - /** - * @code - * this = b*p1 + c*p2 - * @endcode - */ - void add(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2); - /** - * @code - * this = b - c - * @endcode - */ - void sub(BaseMatrixT& b, BaseMatrixT& c); - /** - * @code - * this = b*p1 - c*p2 - * @endcode - */ - void sub(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2); - - /** - * @code - * this = this + b + c - * @endcode - */ - void add2(BaseMatrixT& b, BaseMatrixT& c); - /** - * @code - * this = this*p1 + b*p2 + c*p3 - * @endcode - */ - void add2(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3); - - /** - * @code - * this = a*p1 + b*p2 + c*p3 - * @endcode - */ - void add3(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d, T p1, T p2, T p3); - - /** - * @code - * c = p2 * c - p1 * (b + p3 * this) - * this += mom - * @endcode - */ - void sgdUpdate(BaseMatrixT& b, // grad - BaseMatrixT& c, // mom - T p1, // learningRate, - T p2, // momentum, - T p3); // decayRate - - /** - * @code - * c = p2 * c - p1 * d * (b + p3 * this) - * this += mom - * @endcode - */ - void sgdUpdate(BaseMatrixT& b, // grad, - BaseMatrixT& c, // mom, - BaseMatrixT& d, // lr, - T p1, // learningRate, - T p2, // momentum, - T p3); // decayRate - - /// apply L1/L2 to *this* - virtual void applyL1(T learningRate, T decayRate); - void applyL1(BaseMatrixT& lr, T learningRate, T decayRate); - void applyL2(T learningRate, T decayRate); - void applyL2(BaseMatrixT& lr, T learningRate, T decayRate); - - /** - * @code - * this *= b - * @endcode - */ - void dotMul(BaseMatrixT& b); - - /** - * @code - * this = b * c - * @endcode - */ - void dotMul(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = b / c - * @endcode - */ - void dotDiv(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = (b + p1) / (c + p2) - * @endcode - */ - void dotDiv(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); - - /** - * @code - * this = log(1 + exp(b - c)) - d * (b - c) - * @endcode - */ - void rankLoss(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d); - void rankLossBp(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d); - - /** - * @code - * this = log(1 + exp(b)) - c * b - * @endcode - */ - void logisticRegressionLoss(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this += exp(b)/(1+exp(b)) - c - * @endcode - */ - void logisticRegressionLossBp(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = b > c ? 1.0 : 0.0 - * @endcode - */ - void biggerThan(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = ((b>c && d>0.5) || (bc ? b : c - * @endcode - */ - void max2(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this[destCol] += (b>p1 == c>p1) ? 0 : 1) - * @endcode - */ - void binaryClassificationError(size_t destCol, - BaseMatrixT& b, - BaseMatrixT& c, - T p); - void binaryClassificationError2(size_t destCol, - BaseMatrixT& b, - BaseMatrixT& c, - T p); - - /** - * @code - * this = this * b * b - * @endcode - */ - void dotMulSquare(BaseMatrixT& b); - - /** - * @code - * this = this * this * b - * @endcode - */ - void dotSquareMul(BaseMatrixT& b); - - /** - * @code - * this = b * c * c - * @endcode - */ - void dotMulSquare(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = b * b * c * c - * @endcode - */ - void dotSquareSquare(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = this * (p1*b + p2*c)^2 - * @endcode - */ - void dotMulSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); - - /** - * @code - * this = (p1*b + p2*c)^2 - * @endcode - */ - void dotSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); - - /** - * @code - * this= this * (p1*b + p2*c) - * @endcode - */ - void dotMulSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); - - /** - * @code - * this += sqr(p1*b + p2*c + p3*d) - * @endcode - */ - void addSquareSum( - BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT d, T p1, T p2, T p3); - - /** - * @code - * this += p * sqr(b) - * @endcode - */ - void addSquare(BaseMatrixT& b, T p); - - /** - * @code - * this = p1 * this + p2 * sqr(b) - * @endcode - */ - void decayAddSquare(BaseMatrixT& b, T p1, T p2); - - /** - * @code - * this = p1 * this + p2 * sqr(b * c) - * @endcode - */ - void decayAddSquareMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); - - /** - * @code - * this = 1 / (p1 * b + p2) - * @endcode - */ - void reciprocal2(BaseMatrixT& b, T p1, T p2); - - /** - * @code - * this = 1 / (p1 * b + p2 * c + p3) - * @endcode - */ - void reciprocalSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3); - - /** - * @code - * b = this; this = 0 - * @endcode - */ - void copyAndClear(BaseMatrixT& b); - - /** - * @code - * this_row[destCol] += dotprod(b_row, c_row) - * @endcode - */ - void rowDotMul(size_t destCol, BaseMatrixT& b, BaseMatrixT& c); - void rowDotMul2(size_t destCol, BaseMatrixT& b, BaseMatrixT& c); - - /** - * this is vector (one row matrix) - * - * @code - * for each row i, do: - * this_row += dotmul(b_row_i, c_row_i) - * @endcode - */ - void addDotMulVMM(BaseMatrixT& b, BaseMatrixT& c); - void addDotMulVMM2(BaseMatrixT& b, BaseMatrixT& c); - - /** - * c is vector (one row matrix) - * - * @code - * for each row i, do: - * this_row_i += dotmul(b_row_i, c_row) - * @endcode - */ - void addDotMulMMV(BaseMatrixT& b, BaseMatrixT& c); - void addDotMulMMV2(BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this = p1 * this + p2 * b * c - * @endcode - */ - void addDotMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2); - - /** - * @code - * this_row = b_row * c_row[cCol] - * @endcode - */ - void rowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); - void rowScale2(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this_col = b_col * c_col[cRow] - * @endcode - */ - void colScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this_col += b_col * c_col[cRow] - * @endcode - */ - void addColScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c); - - /** - * @code - * this_row += b_row * c_row[cCol] - * @endcode - */ - void addRowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); - - /// calculate the sum of each row of the matrix b. - /// this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} - void sumRows(BaseMatrixT& b, T scaleSum, T scaleDest); - - /// calculate the maximum value of each row of the matrix b. - void maxRows(BaseMatrixT& b); - /// calculate the minimum value of each row of the matrix b. - void minRows(BaseMatrixT& b); - - /// calculate the maximum value of each column of the matrix b. - void maxCols(BaseMatrixT& b); - /// calculate the minimum value of each column of the matrix b. - void minCols(BaseMatrixT& b); - - /// calculate the sum of each column of the matrix b. - /// this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji} - void sumCols(BaseMatrixT& b, T scaleSum, T scaleDest); - - /// this_i = scaleDest * this_i + scaleSum * \sum_j (b_{ij} - c_{ij})^2 - void sumOfSquaredDiffs(BaseMatrixT& b, - BaseMatrixT& c, - T scaleSum, - T scaleDest); - - /// this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij} - void sumOfProducts(BaseMatrixT& b, BaseMatrixT& c, T scaleSum, T scaleDest); - - /** - * @code - * this_row = b_row + p * ones * c_row[cCol] - * @endcode - */ - void rowAdd(size_t cCol, BaseMatrixT& b, BaseMatrixT& c, T p); - /** - * @code - * this_row = pow(b_row, c_row[cCol]) - * @endcode - */ - void rowPow(size_t cCol, BaseMatrixT& b, BaseMatrixT& c); - - virtual bool isSparse() const { return false; } - - template - void operator=(const ExpressionType& expr) { - if (useGpu_) { - TensorGpuApply(*this, expr); - } else { - TensorCpuApply(*this, expr); - } - } - - template - void operator+=(const ExpressionType& expr) { - (*this) = (*this) + expr; - } - template - void operator-=(const ExpressionType& expr) { - (*this) = (*this) - expr; - } - template - void operator*=(const ExpressionType& expr) { - (*this) = (*this) * expr; - } - template - void operator/=(const ExpressionType& expr) { - (*this) = (*this) / expr; - } -}; - -typedef BaseMatrixT BaseMatrix; -typedef BaseMatrixT IBaseMatrix; - -} // namespace paddle diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt deleted file mode 100644 index 3c897b5f3e09cd53ddd5b767333ce4759250da71..0000000000000000000000000000000000000000 --- a/paddle/math/CMakeLists.txt +++ /dev/null @@ -1,57 +0,0 @@ -# common package contains: -# * the utilities: -# * Thread Libs -# * Memory Manage libs -# * CommandLine Parser -# * Logging -# * Timer/Stats -# * the math libraries: -# * Matrix/Vector -# * the parameter optimizers. -# * the parameter updater functions. -# -# TODO(yuyang18): separate libs. -# -file(GLOB MATH_HEADERS . *.h) -file(GLOB MATH_SOURCES . *.cpp) - -if(NOT WITH_MKLDNN) - set(DNN_HEADER "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.h") - set(DNN_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.cpp") - list(REMOVE_ITEM MATH_HEADERS "${DNN_HEADER}") - list(REMOVE_ITEM MATH_SOURCES "${DNN_SOURCE}") - message(STATUS "Skip compiling with MKLDNNMatrix") -else() - message(STATUS "Compile with MKLDNNMatrix") -endif() - -if(MOBILE_INFERENCE) - # Remove sparse - list(REMOVE_ITEM MATH_HEADERS - ${CMAKE_CURRENT_SOURCE_DIR}/CpuSparseMatrix.h - ${CMAKE_CURRENT_SOURCE_DIR}/SparseMatrix.h - ${CMAKE_CURRENT_SOURCE_DIR}/SparseRowMatrix.h) - list(REMOVE_ITEM MATH_SOURCES - ${CMAKE_CURRENT_SOURCE_DIR}/CpuSparseMatrix.cpp - ${CMAKE_CURRENT_SOURCE_DIR}/SparseMatrix.cpp - ${CMAKE_CURRENT_SOURCE_DIR}/SparseRowMatrix.cpp) -endif() -set(MATH_SOURCES - "${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu" - "${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu" - ${MATH_SOURCES}) -if(NOT WITH_GPU) - # then compile BaseMatrix.cu as c++ file - compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu") - compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu") - add_library(paddle_math STATIC - ${MATH_SOURCES}) -else() - cuda_add_library(paddle_math ${MATH_SOURCES}) -endif() - - -add_dependencies(paddle_math paddle_proto ${external_project_dependencies}) # depends -if(WITH_TESTING) - add_subdirectory(tests) -endif() diff --git a/paddle/math/CpuSparseMatrix.cpp b/paddle/math/CpuSparseMatrix.cpp deleted file mode 100644 index 023450ffb794086399d7131ba5faa4dbefeaaf7d..0000000000000000000000000000000000000000 --- a/paddle/math/CpuSparseMatrix.cpp +++ /dev/null @@ -1,787 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "CpuSparseMatrix.h" -#include "SparseMatrix.h" -#include "float.h" -#include "hl_gpu.h" -#include "paddle/math/MathUtils.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -const size_t CpuSparseMatrix::DEFAULT_AVG_WIDTH; - -CpuSparseMatrix::CpuSparseMatrix(size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans) - : Matrix(NULL, height, width, trans, false) { - resize(height, width, nnz, valueType, format); -} - -CpuSparseMatrix::CpuSparseMatrix(CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans) - : Matrix(dataHandle, height, width, trans, false) { - resize(height, width, nnz, valueType, format); -} - -CpuSparseMatrix::CpuSparseMatrix(real* data, - int* rows, - int* cols, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans) - : Matrix(NULL, height, width, trans, false) { - cols_ = cols; - rows_ = rows; - value_ = data; - height_ = height; - width_ = width; - elementCnt_ = nnz; - valueType_ = valueType; - format_ = format; -} - -void CpuSparseMatrix::resize(size_t newHeight, - size_t newWidth, - size_t newNnz, - SparseValueType valueType, - SparseFormat format) { - CHECK_LE(newNnz, newHeight * newWidth); - size_t newSize = 0; - if (format == SPARSE_CSR) { - newSize = (newHeight + 1) * sizeof(int) + newNnz * sizeof(int); - } else { - newSize = (newWidth + 1) * sizeof(int) + newNnz * sizeof(int); - } - - if (NO_VALUE != valueType) { - newSize += newNnz * sizeof(real); - } - - if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) { - memoryHandle_ = std::make_shared(newSize); - } - - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newNnz; - valueType_ = valueType; - format_ = format; - sparseResize(); -} -void CpuSparseMatrix::sparseResize() { - if (format_ == SPARSE_CSR) { - rows_ = reinterpret_cast( - reinterpret_cast(memoryHandle_->getBuf())); - cols_ = reinterpret_cast( - reinterpret_cast(memoryHandle_->getBuf()) + - (height_ + 1) * sizeof(int)); - if (NO_VALUE != valueType_) { - value_ = reinterpret_cast( - reinterpret_cast(memoryHandle_->getBuf()) + - (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); - } else { - value_ = NULL; - } - } else { - cols_ = reinterpret_cast( - reinterpret_cast(memoryHandle_->getBuf())); - rows_ = reinterpret_cast( - reinterpret_cast(memoryHandle_->getBuf()) + - (width_ + 1) * sizeof(int)); - if (NO_VALUE != valueType_) { - value_ = reinterpret_cast( - reinterpret_cast(memoryHandle_->getBuf()) + - (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); - } else { - value_ = NULL; - } - } -} - -void CpuSparseMatrix::resize(size_t newHeight, size_t newWidth) { - resize(newHeight, - newWidth, - newHeight * std::min(DEFAULT_AVG_WIDTH, newWidth), - valueType_, - format_); -} - -MatrixPtr CpuSparseMatrix::getTranspose() { - if (!memoryHandle_ && !value_) { - MatrixPtr dest(new CpuSparseMatrix( - height_, width_, elementCnt_, valueType_, format_, true)); - return dest; - } else if (memoryHandle_) { - MatrixPtr dest(new CpuSparseMatrix( - std::dynamic_pointer_cast(memoryHandle_), - height_, - width_, - elementCnt_, - valueType_, - format_, - true)); - return dest; - } else if (value_) { - MatrixPtr dest(new CpuSparseMatrix(value_, - rows_, - cols_, - height_, - width_, - elementCnt_, - valueType_, - format_, - true)); - return dest; - } else { - return NULL; - } -} - -SparseValueType CpuSparseMatrix::getValueType() { return valueType_; } - -void CpuSparseMatrix::mul(const Matrix& a, - const Matrix& b, - real scaleAB, - real scaleT) { - CHECK(!isTransposed()) << "Not supported"; - const auto a_ptr = dynamic_cast(&a); - const auto b_ptr = dynamic_cast(&b); - - if (a_ptr && b_ptr) { - CpuMatrix::mul((CpuMatrix*)a_ptr, (CpuMatrix*)b_ptr, this, scaleAB, scaleT); - } else { - LOG(FATAL) << "not supported"; - } -} - -void CpuSparseMatrix::add3(CpuMatrix* b) { - CHECK(getFormat() != SPARSE_CSC) << "Not supported"; - CHECK(height_ == b->getHeight()); - CHECK(width_ == b->getWidth()); - real* A = getValue(); - real* B = b->getData(); - int* cols = getCols(); - for (size_t i = 0; i < height_; i++) { - size_t start = getRowStartIdx(i); - size_t end = getRowStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - A[j] = B[i * width_ + cols[j]]; - } - } -} - -void CpuSparseMatrix::add3(MatrixPtr b) { - if (dynamic_cast(b.get())) { - add3(dynamic_cast(b.get())); - } else { - LOG(FATAL) << "not supported"; - } -} - -void CpuSparseMatrix::addBias(Matrix& b, real scale) { - CHECK_EQ(b.getHeight(), (size_t)1); - CHECK_EQ(width_, b.getWidth()); - real* A = getValue(); - real* B = b.getData(); - int* cols = getCols(); - size_t nnz = getElementCnt(); - for (size_t i = 0; i < nnz; i++) { - A[i] += scale * B[cols[i]]; - } -} - -template -void printBuf(std::ostream& os, T* a, size_t len, const char* name) { - os << "\n: " << name << " ["; - for (size_t i = 0; i < len; i++) { - os << a[i] << " "; - } - os << "]\n"; -} - -void CpuSparseMatrix::print(std::ostream& os) const { - size_t rowSize = format_ == SPARSE_CSC ? elementCnt_ : height_ + 1; - size_t colSize = format_ == SPARSE_CSC ? width_ + 1 : elementCnt_; - printBuf(os, rows_, rowSize, "row"); - printBuf(os, cols_, colSize, "col"); - if (valueType_ == FLOAT_VALUE) { - printBuf(os, value_, elementCnt_, "value"); - } - return; -} - -void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const { - CHECK_LT(idx, height_); - if (format_ == SPARSE_CSC) { - LOG(FATAL) << "SPARSE_CSC not supported"; - return; - } - - const int* col = getRowCols(idx); - size_t num = getColNum(idx); - if (num > 0) { - if (valueType_ == FLOAT_VALUE) { - const real* data = getRowValues(idx); - os << col[0] << ":" << data[0]; - for (size_t i = 1; i < num; ++i) { - os << " " << col[i] << ":" << data[i]; - } - } else { - os << col[0]; - for (size_t i = 1; i < num; ++i) { - os << " " << col[i]; - } - } - } - os << ";"; -} - -void CpuSparseMatrix::rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c) { - CHECK(getFormat() != SPARSE_CSC) << "Not supported"; - CHECK_EQ(height_, b.getHeight()); - CHECK_EQ(width_, b.getWidth()); - real* A = getValue(); - real* B = b.getValue(); - if (b.getValueType() == FLOAT_VALUE) { - for (size_t i = 0; i < height_; i++) { - size_t start = getRowStartIdx(i); - size_t end = getRowStartIdx(i + 1); - CHECK_EQ(start, b.getRowStartIdx(i)); - CHECK_EQ(end, b.getRowStartIdx(i + 1)); - for (size_t j = start; j < end; j++) { - A[j] = B[j] * c.getElement(i, cCol); - } - } - } else if (b.getValueType() == NO_VALUE) { - for (size_t i = 0; i < height_; i++) { - size_t start = getRowStartIdx(i); - size_t end = getRowStartIdx(i + 1); - CHECK_EQ(start, b.getRowStartIdx(i)); - CHECK_EQ(end, b.getRowStartIdx(i + 1)); - for (size_t j = start; j < end; j++) { - A[j] = c.getElement(i, cCol); - } - } - } -} - -void CpuSparseMatrix::randomizeUniform() { - CHECK_LE(elementCnt_, height_ * width_); - if (valueType_ == FLOAT_VALUE) { - real* data = getValue(); - for (size_t i = 0; i < elementCnt_; ++i) { - *data++ = rand() / static_cast(RAND_MAX); // NOLINT - } - } - if (format_ == SPARSE_CSR) { - sparseRand(rows_, cols_, elementCnt_, height_ + 1, width_, false); - } else { - sparseRand(cols_, rows_, elementCnt_, width_ + 1, height_, false); - } -} - -void CpuSparseMatrix::copyFrom(std::vector& rows, - std::vector& cols, - std::vector& values) { - size_t size = format_ == SPARSE_CSR ? cols.size() : rows.size(); - resize(height_, width_, size, valueType_, format_); - if (valueType_ == FLOAT_VALUE) { - memcpy(&value_[0], &values[0], sizeof(real) * values.size()); - } - memcpy(&cols_[0], &cols[0], sizeof(int) * cols.size()); - memcpy(&rows_[0], &rows[0], sizeof(int) * rows.size()); -} - -// Copy from a CpuMatrix, only supported in sparse_float_value_t -// SparseMatrix. -void CpuSparseMatrix::copyFrom(const CpuMatrix& src) { - CHECK_EQ(getHeight(), src.getHeight()); - CHECK_EQ(getWidth(), src.getWidth()); - CHECK(!src.trans_ && !trans_); - if (format_ == SPARSE_CSR) { - std::vector rows(getHeight() + 1); - std::vector cols; - std::vector values; - rows[0] = 0; - for (size_t r = 0; r < getHeight(); ++r) { - for (size_t c = 0; c < getWidth(); ++c) { - real v = src.getElement(r, c); - if (fabs(v) > FLT_EPSILON) { - cols.push_back(c); - values.push_back(v); - } - } - rows[r + 1] = values.size(); - } - copyFrom(rows, cols, values); - } else { - std::vector cols(getWidth() + 1); - std::vector rows; - std::vector values; - cols[0] = 0; - for (size_t r = 0; r < getWidth(); ++r) { - for (size_t c = 0; c < getHeight(); ++c) { - real v = src.getElement(c, r); - if (fabs(v) > FLT_EPSILON) { - rows.push_back(c); - values.push_back(v); - } - } - cols[r + 1] = values.size(); - } - copyFrom(rows, cols, values); - } -} - -MatrixPtr CpuSparseMatrix::clone(size_t height, size_t width, bool useGpu) { - if (height == 0 && width == 0) { - height = height_; - width = width_; - } - CHECK(width && height); - if (!useGpu) { - return std::make_shared( - height, width, 0, valueType_, format_); - } else { - return std::make_shared( - height, width, elementCnt_, valueType_, format_); - } -} - -MatrixPtr CpuSparseMatrix::subMatrix(size_t startRow, size_t numRows) { - CHECK_LE(startRow + numRows, height_); - CHECK_EQ(format_, SPARSE_CSR); - if (valueType_ == NO_VALUE) { - return std::make_shared( - nullptr, - rows_ + startRow, - cols_, - numRows, - width_, - rows_[startRow + numRows] - rows_[startRow], - valueType_, - format_, - trans_); - } else { - return std::make_shared( - value_, - rows_ + startRow, - cols_, - numRows, - width_, - rows_[startRow + numRows] - rows_[startRow], - valueType_, - format_, - trans_); - } -} - -/* mem MUST be alloced outside (memAlloc=false) */ -void CpuSparseMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { - CHECK(!memAlloc); - CpuSparseMatrix* mat = dynamic_cast(matTrans.get()); - if (format_ == SPARSE_CSR) { - /*statistic element number in each col*/ - int* colCounters = mat->getRows() + 1; - memset(colCounters, 0, sizeof(int) * width_); - for (size_t i = 0; i < elementCnt_; ++i) { - int col = cols_[i]; - colCounters[col]++; - } - /*fill mat rows */ - mat->getRows()[0] = 0; - for (size_t i = 1; i < width_ + 1; i++) { - mat->getRows()[i] = mat->getRows()[i - 1] + mat->getRows()[i]; - } - /*fill mat values and cols*/ - std::vector colNumVec(width_, 0); - if (valueType_ == FLOAT_VALUE) { - for (size_t i = 0; i < height_; i++) { - for (int j = rows_[i]; j < rows_[i + 1]; j++) { - int colIdx = cols_[j]; - int index = mat->getRows()[colIdx] + colNumVec[colIdx]; - mat->getCols()[index] = i; - mat->getValue()[index] = value_[j]; - colNumVec[colIdx]++; - } - } - } else { - for (size_t i = 0; i < height_; i++) { - for (int j = rows_[i]; j < rows_[i + 1]; j++) { - int colIdx = cols_[j]; - int index = mat->getRows()[colIdx] + colNumVec[colIdx]; - mat->getCols()[index] = i; - colNumVec[colIdx]++; - } - } - } - } else { - /*statistic element number in each row*/ - int* rowCounters = mat->getCols() + 1; - memset(rowCounters, 0, sizeof(int) * height_); - for (size_t i = 0; i < elementCnt_; ++i) { - int row = rows_[i]; - rowCounters[row]++; - } - - /*fill mat cols */ - mat->getCols()[0] = 0; - for (size_t i = 1; i < height_ + 1; i++) { - mat->getCols()[i] = mat->getCols()[i - 1] + mat->getCols()[i]; - } - /*fill mat values and rows*/ - std::vector rowNumVec(height_, 0); - if (valueType_ == FLOAT_VALUE) { - for (size_t i = 0; i < width_; i++) { - for (int j = cols_[i]; j < cols_[i + 1]; j++) { - int rowIdx = rows_[j]; - int index = mat->getCols()[rowIdx] + rowNumVec[rowIdx]; - mat->getRows()[index] = i; - mat->getValue()[index] = value_[j]; - rowNumVec[rowIdx]++; - } - } - } else { - for (size_t i = 0; i < width_; i++) { - for (int j = cols_[i]; j < cols_[i + 1]; j++) { - int rowIdx = rows_[j]; - int index = mat->getCols()[rowIdx] + rowNumVec[rowIdx]; - mat->getRows()[index] = i; - rowNumVec[rowIdx]++; - } - } - } - } -} - -void CpuSparseMatrix::setRow(size_t row, - size_t colNum, - const unsigned int* cols, - const real* values) { - if (format_ == SPARSE_CSR) { - CHECK_LT(row, height_); - CHECK(NULL != cols); - if (0 == row) { - rows_[row] = 0; - } - rows_[row + 1] = rows_[row] + colNum; - for (size_t i = 0; i < colNum; ++i) { - cols_[rows_[row] + i] = cols[i]; - } - if (valueType_ == NO_VALUE) { - CHECK(!values); - } else { - for (size_t i = 0; i < colNum; ++i) { - value_[rows_[row] + i] = values[i]; - } - } - } else { - LOG(FATAL) << "not supported"; - } -} - -void CpuSparseMatrix::fillRowIndices(IVectorPtr& outVec) const { - if (format_ == SPARSE_CSR) { - auto nnz = getElementCnt(); - IVector::resizeOrCreate(outVec, nnz, false); - auto out = outVec->getData(); - int* rows = getRows(); - for (size_t i = 0; i < height_; i++) { - for (int j = rows[i]; j < rows[i + 1]; j++) { - out[j] = i; - } - } - } else { - LOG(FATAL) << "SPARSE_CSC not supported"; - } -} - -ThreadLocal> CpuSparseMatrix::cpuLocalMats_; - -CpuSparseMatrixPtr CpuSparseMatrix::getTmpSparseMatrix(size_t height, - size_t width) { - std::vector* localMats = cpuLocalMats_.get(); - auto it = localMats->begin(); - while (it != localMats->end()) { - if (it->unique()) { - (*it)->resize(height, width, elementCnt_, valueType_, format_); - return *it; - } - } - localMats->emplace_back(std::make_shared( - height, width, elementCnt_, valueType_, format_, false)); - return localMats->back(); -} - -void CpuSparseMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { - if (dynamic_cast(&src)) { - auto tmpSrc = dynamic_cast(&src); - copyFrom(*tmpSrc, stream); - } else if (dynamic_cast(&src)) { - auto tmpSrc = dynamic_cast(&src); - copyFrom(*tmpSrc); - } else if (dynamic_cast(&src)) { - auto tmpSrc = dynamic_cast(&src); - copyFrom(*tmpSrc); - } else { - LOG(FATAL) << "not implemented"; - } -} - -void CpuSparseMatrix::copyFrom(const Matrix& src) { - if (dynamic_cast(&src)) { - auto tmpSrc = dynamic_cast(&src); - copyFrom(*tmpSrc); - } else if (dynamic_cast(&src)) { - auto tmpSrc = dynamic_cast(&src); - copyFrom(*tmpSrc); - } else { - LOG(FATAL) << "not implemented"; - } -} - -void CpuSparseMatrix::copyFrom(const GpuSparseMatrix& src, hl_stream_t stream) { - CHECK_EQ(height_, src.getHeight()); - CHECK_EQ(width_, src.getWidth()); - CHECK_EQ(size_t(elementCnt_), src.getElementCnt()); - size_t valSize = valueType_ == NO_VALUE ? 0 : elementCnt_; - if (format_ == SPARSE_CSC) - hl_memcpy_from_csc_matrix(value_, - valSize, - rows_, - elementCnt_, - cols_, - width_ + 1, - src.sMatrix_.get(), - stream); - else - hl_memcpy_from_csr_matrix(value_, - valSize, - rows_, - height_ + 1, - cols_, - elementCnt_, - src.sMatrix_.get(), - stream); -} - -void CpuSparseMatrix::copyFrom(const CpuSparseMatrix& src) { - CHECK_EQ(height_, src.getHeight()); - CHECK_EQ(width_, src.getWidth()); - CHECK_EQ(format_, src.getFormat()); - int start = format_ == SPARSE_CSR ? src.getRows()[0] : src.getCols()[0]; - if (format_ == SPARSE_CSR) { - size_t totalColNum = 0; - for (size_t i = 0; i < height_; ++i) { - totalColNum += src.getColNum(i); - } - resize(height_, width_, totalColNum, valueType_, format_); - rows_[0] = 0; - for (size_t i = 0; i < height_; ++i) { - rows_[i + 1] = rows_[i] + src.getColNum(i); - } - memcpy(cols_, src.getCols() + start, totalColNum * sizeof(int)); - } else { - size_t totalColNum = 0; - for (size_t i = 0; i < width_; ++i) { - totalColNum += src.getRowNum(i); - } - resize(height_, width_, totalColNum, valueType_, format_); - cols_[0] = 0; - for (size_t i = 0; i < width_; ++i) { - cols_[i + 1] = cols_[i] + src.getRowNum(i); - } - memcpy(rows_, src.getRows() + start, totalColNum * sizeof(int)); - } - - // if have different value type, only copy rows and cols - if (valueType_ == FLOAT_VALUE && src.getValueType() == FLOAT_VALUE) { - memcpy(value_, src.getValue() + start, elementCnt_ * sizeof(real)); - } -} - -void CpuSparseMatrix::copyRow(int offsets, - size_t colNum, - const sparse_non_value_t* row) { - for (size_t j = 0; j < colNum; j++) { - cols_[offsets + j] = row[j].col; - } -} - -void CpuSparseMatrix::copyRow(int offsets, - size_t colNum, - const sparse_float_value_t* row) { - for (size_t j = 0; j < colNum; j++) { - cols_[offsets + j] = row[j].col; - value_[offsets + j] = row[j].value; - } -} - -template -void CpuSparseMatrix::copyFrom(int64_t* ids, int64_t* indices, T* data) { - size_t totalColNum = 0; - for (size_t i = 0; i < height_; ++i) { - int64_t id = ids[i]; - totalColNum += indices[id + 1] - indices[id]; - } - valueType_ = typeid(T) == typeid(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE; - - resize(height_, width_, totalColNum, valueType_, format_); - - rows_[0] = 0; - for (size_t i = 0; i < height_; ++i) { - int64_t id = ids[i]; - T* row = data + indices[id]; - size_t colNum = indices[id + 1] - indices[id]; - rows_[i + 1] = rows_[i] + colNum; - copyRow(rows_[i], colNum, row); - } -} - -template -void CpuSparseMatrix::copyFrom(int64_t* indices, T* data) { - CHECK(format_ == SPARSE_CSR); - size_t totalColNum = indices[height_] - indices[0]; - valueType_ = typeid(T) == typeid(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE; - resize(height_, width_, totalColNum, valueType_, format_); - - rows_[0] = 0; - for (size_t i = 0; i < height_; ++i) { - T* row = data + indices[i]; - size_t colNum = indices[i + 1] - indices[i]; - rows_[i + 1] = rows_[i] + colNum; - copyRow(rows_[i], colNum, row); - } -} - -void CpuSparseMatrix::trimFrom(const CpuSparseMatrix& src) { - CHECK_EQ(height_, src.getHeight()); - CHECK_LE(width_, src.getWidth()); - CHECK_EQ(format_, src.getFormat()); - CHECK_EQ(valueType_, src.getValueType()); - if (format_ == SPARSE_CSR) { - int* srcCols = src.getCols(); - size_t numLessWidth = - std::count_if(srcCols, srcCols + src.getElementCnt(), [this](size_t n) { - return n < this->width_; - }); - resize(height_, width_, numLessWidth, valueType_, format_); - rows_[0] = 0; - size_t index = 0; - for (size_t r = 0; r < height_; ++r) { - for (int i = src.getRows()[r]; i < src.getRows()[r + 1]; ++i) { - if (srcCols[i] < static_cast(width_)) { - cols_[index] = srcCols[i]; - if (valueType_ == FLOAT_VALUE) { - value_[index] = src.getValue()[i]; - } - ++index; - } - } - rows_[r + 1] = index; - } - CHECK_EQ(index, numLessWidth); - } else { - size_t numLessWidth = src.getCols()[width_] - src.getCols()[0]; - resize(height_, width_, numLessWidth, valueType_, format_); - cols_[0] = 0; - size_t index = 0; - // note: c < width_, not src.getWidth(); - for (size_t c = 0; c < width_; ++c) { - for (int i = src.getCols()[c]; i < src.getCols()[c + 1]; ++i) { - rows_[index] = src.getRows()[i]; - if (valueType_ == FLOAT_VALUE) { - value_[index] = src.getValue()[i]; - } - ++index; - } - cols_[c + 1] = index; - } - CHECK_EQ(index, numLessWidth); - } -} - -void CpuSparseMatrix::zeroMem() { - CHECK(valueType_ == FLOAT_VALUE); - memset(value_, 0, elementCnt_ * sizeof(real)); -} - -template void CpuSparseMatrix::copyFrom(int64_t* ids, - int64_t* indices, - sparse_non_value_t* data); - -template void CpuSparseMatrix::copyFrom(int64_t* ids, - int64_t* indices, - sparse_float_value_t* data); - -template void CpuSparseMatrix::copyFrom(int64_t* indices, - sparse_non_value_t* data); - -template void CpuSparseMatrix::copyFrom(int64_t* indices, - sparse_float_value_t* data); - -void CpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { - size_t numSamples = getHeight(); - size_t beam = maxVal.getWidth(); - CHECK_EQ(maxIds.getSize(), numSamples * beam); - CHECK_EQ(maxVal.getHeight(), numSamples); - maxVal.zeroMem(); - int* outids = maxIds.getData(); - real* outvalues = maxVal.getData(); - - typedef std::pair valuepair; - std::vector vec; - for (size_t i = 0; i < numSamples; i++) { - vec.clear(); - - auto num = getColNum(i); - auto ids = getRowCols(i); - auto values = getRowValues(i); - for (size_t j = 0; j < num; j++) { - vec.push_back(std::make_pair(values[j], ids[j])); - } - - size_t outsize = std::min(num, beam); - std::partial_sort(vec.begin(), - vec.begin() + outsize, - vec.end(), - [](const valuepair& a, const valuepair& b) { - return a.first > b.first; - }); - for (size_t j = 0; j < outsize; j++) { - outids[i * beam + j] = vec[j].second; - outvalues[i * beam + j] = vec[j].first; - } - if (outsize < beam) { - // if the number of values to sort are less than the output size, - // use -1 to indicate the end of valid sorted values. - outids[i * beam + outsize] = -1; - } - } -} - -} // namespace paddle diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h deleted file mode 100644 index d4a78f3e54b73add3c00e17f13d91359839d3d14..0000000000000000000000000000000000000000 --- a/paddle/math/MKLDNNMatrix.h +++ /dev/null @@ -1,256 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "Matrix.h" -#include "mkldnn.hpp" -#include "paddle/parameter/Parameter.h" - -namespace paddle { - -class MKLDNNMatrix; -typedef std::shared_ptr MKLDNNMatrixPtr; - -#define CHECK_PRIMITIVE_DESC_EQ(MAT, PD, ...) \ - CHECK(MAT) << " can not be empty."; \ - CHECK(MAT->getPrimitiveDesc() == PD) \ - << #MAT "->getPrimitiveDesc() and " #PD " should be equal.\n " \ - << "" __VA_ARGS__; - -/** - * @brief MKLDNN Matrix. - * - */ -class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory { - public: - MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd) - : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false), - mkldnn::memory(pd, m->getData()), - m_(m) {} - - ~MKLDNNMatrix() {} - - /** - * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc - */ - static MKLDNNMatrixPtr create(mkldnn::memory::primitive_desc pd, - MatrixPtr m = nullptr); - - /** - * Create MKLDNNMatrix from a MatrixPtr and memory details info - */ - static MKLDNNMatrixPtr create( - mkldnn::memory::dims dims, - mkldnn::memory::format fmt, - mkldnn::engine& eg, - MatrixPtr m = nullptr, - mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32); - - /** - * Create primitive descriptor. - * default with f32 dtype - */ - static mkldnn::memory::primitive_desc createPrimitiveDesc( - const mkldnn::memory::dims dims, - const mkldnn::memory::format& fmt, - const mkldnn::engine& eg, - const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) { - return mkldnn::memory::primitive_desc(memory::desc(dims, dtype, fmt), eg); - } - - /** - * Create Memory descriptor. - * default with any format and f32 dtype - */ - static mkldnn::memory::desc createMemoryDesc( - const mkldnn::memory::dims dims, - const mkldnn::memory::format& fmt = mkldnn::memory::format::any, - const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) { - return mkldnn::memory::desc(dims, dtype, fmt); - } - - /** - * Create reorder primitive. - * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst. - * checkData: whether to check the data handle of src and dst. - * if true, it will check the data and do not allow them equal; - * otherwise, it will not check them, then the reorder created - * may have inplace buffer. - * Do not set false, if you can not guarantee the inplace logical - * would work with your reorder. - */ - static std::shared_ptr createReorder( - const MKLDNNMatrixPtr& src, - const MKLDNNMatrixPtr& dst, - bool checkData = true); - - void copyFrom(const Matrix& src) { - // TODO(TJ): reorder data if this format is not nchw or x - m_->copyFrom(src); - } - - void copyTo(Matrix& dst) { - // TODO(TJ): reorder data if this format is not nchw or x - dst.copyFrom(*m_); - } - - public: - /** - * Reorder this MKLDNNMatrix from other format. - * Support inplace reorder. - * @note: this function would only reorder the data layout. - * will NOT change this original dim or format info - */ - void reorderDataFrom(const MKLDNNMatrixPtr& m, - memory::format srcFmt, - memory::dims targetDim); - - /** - * Reorder this MKLDNNMatrix to other format. - * Support inplace reorder. - * @note: this function would only reorder the data layout. - * will NOT change the dst dim or format info - */ - void reorderDataTo(const MKLDNNMatrixPtr& m, - memory::format dstFmt, - memory::dims targetDim); - - /** - * Dimensionality reduction. - * Change format "nchw --> nc" or "oihw --> oi" if the h and w are both 1 - */ - void downSpatial(); - - /** - * set the memory data handle. - * Caution: This will not check the buffer size of the data, - * it should be coverd by user. - */ - void setData(real* data) { - set_data_handle(data); - CpuMatrix::setData(data); - m_.reset(); - } - - /** - * override the CpuMatrix::resize - */ - void resize(size_t newHeight, size_t newWidth) override { - m_->resize(newHeight, newWidth); - if (data_ == m_->getData() && elementCnt_ == newHeight * newWidth) { - return; - } - CpuMatrix::setData(data_); - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newHeight * newWidth; - stride_ = width_; - auto pd = mkldnn::memory::primitive_desc( - mkldnn::memory::desc({(int)newHeight, (int)newWidth}, - getDtype(), - mkldnn::memory::format::nc), - getEngine()); - resetMKLDNNMemory(pd, data_); - } - - /** - * override Matrix::getData - * check data before return - */ - real* getData() override { - CHECK_EQ((void*)data_, get_data_handle()); - return data_; - } - - const real* getData() const override { - CHECK_EQ((void*)data_, get_data_handle()); - return data_; - } - - /** - * Get primitive descriptor. - */ - mkldnn::memory::primitive_desc getPrimitiveDesc() { - return this->get_primitive_desc(); - } - - /** - * Get memory descriptor. - */ - mkldnn::memory::desc getMemoryDesc() { return getPrimitiveDesc().desc(); } - - /** - * Get dimensions. - */ - mkldnn::memory::dims getDims() { - mkldnn::memory::desc md = getMemoryDesc(); - const int* src = md.data.dims; - int ndims = md.data.ndims; - mkldnn::memory::dims dst; - dst.resize(ndims); - for (int i = 0; i < ndims; ++i) { - dst[i] = src[i]; - } - return dst; - } - - /** - * Get format. - */ - mkldnn::memory::format getFormat() { - return (mkldnn::memory::format)(getMemoryDesc().data.format); - } - - /** - * Get memory data type. - */ - mkldnn::memory::data_type getDtype() { - return (mkldnn::memory::data_type)(getMemoryDesc().data.data_type); - } - - /** - * Get engine. - */ - mkldnn::engine getEngine() { return getPrimitiveDesc().get_engine(); } - - protected: - /** - * Do reorder once. - * Can support inplace. - */ - void reorderOnce(void* srcData, - void* dstData, - memory::format srcFmt, - memory::format dstFmt, - memory::dims dm); - /** - * reset this MKLDNN Memory from primitve desc - */ - void resetMKLDNNMemory(memory::primitive_desc pd, real* data) { - mkldnn_primitive_t result; - mkldnn::error::wrap_c_api( - mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr), - "could not create a memory primitive"); - reset(result); - set_data_handle(data); - } - - private: - // save the CpuMatrixPtr in case the buffer released outside - CpuMatrixPtr m_; -}; - -} // namespace paddle diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp deleted file mode 100644 index f48119aa511578b21602a225277f01b4c6a9e9a8..0000000000000000000000000000000000000000 --- a/paddle/math/MathFunctions.cpp +++ /dev/null @@ -1,348 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/math/MathFunctions.h" -#include "hl_matrix_apply.cuh" -#include "hl_matrix_ops.cuh" -#include "paddle/utils/DynamicLoader.h" - -namespace dynload { - -std::once_flag lapack_dso_flag; -void* lapack_dso_handle = nullptr; - -/** - * The following macro definition can generate structs - * (for each function) to dynamic load lapack routine - * via operator overloading. - * - * note: default dynamic linked libs - */ - -// The argument for stringizing operator is not macro-expanded first. -// We have to use two levels of macro to do the expansion. -// See https://gcc.gnu.org/onlinedocs/cpp/Stringizing.html -#define STR(x) #x - -// clang-format off -#ifndef LAPACK_FOUND -#define DYNAMIC_LOAD_LAPACK_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - auto operator()(Args... args) -> decltype(__name(args...)) { \ - using lapack_func = decltype(__name(args...)) (*)(Args...); \ - std::call_once(lapack_dso_flag, GetLapackDsoHandle, &lapack_dso_handle); \ - void* p_##__name = dlsym(lapack_dso_handle, STR(__name)); \ - CHECK(p_##__name) << "Cannot find symbol " << STR(__name) \ - << " in liblapack.so"; \ - return reinterpret_cast(p_##__name)(args...); \ - } \ - } __name; // struct DynLoad__##__name -#else -#define DYNAMIC_LOAD_LAPACK_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - auto operator()(Args... args) -> decltype(__name(args...)) { \ - return __name(args...); \ - } \ - } __name; // struct DynLoad__##__name -#endif - -#define PADDLE_SGETRF LAPACKE_sgetrf -#define PADDLE_DGETRF LAPACKE_dgetrf -#define PADDLE_SGETRI LAPACKE_sgetri -#define PADDLE_DGETRI LAPACKE_dgetri - -#define LAPACK_ROUTINE_EACH(__macro) \ - __macro(PADDLE_SGETRF) \ - __macro(PADDLE_DGETRF) \ - __macro(PADDLE_SGETRI) \ - __macro(PADDLE_DGETRI) -// clang-format on - -LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP) - -} // namespace dynload - -namespace paddle { - -#ifndef PADDLE_USE_EIGEN_FOR_BLAS -template <> -void gemm(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, - const int M, - const int N, - const int K, - const float alpha, - const float* A, - const int lda, - const float* B, - const int ldb, - const float beta, - float* C, - const int ldc) { - cblas_sgemm(CblasRowMajor, - transA, - transB, - M, - N, - K, - alpha, - A, - lda, - B, - ldb, - beta, - C, - ldc); -} - -template <> -void gemm(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, - const int M, - const int N, - const int K, - const double alpha, - const double* A, - const int lda, - const double* B, - const int ldb, - const double beta, - double* C, - const int ldc) { - cblas_dgemm(CblasRowMajor, - transA, - transB, - M, - N, - K, - alpha, - A, - lda, - B, - ldb, - beta, - C, - ldc); -} -#endif - -template <> -int getrf(const CBLAS_ORDER order, - const int M, - const int N, - float* A, - const int lda, - int* ipiv) { - return dynload::PADDLE_SGETRF(order, M, N, A, lda, ipiv); -} - -template <> -int getrf(const CBLAS_ORDER order, - const int M, - const int N, - double* A, - const int lda, - int* ipiv) { - return dynload::PADDLE_DGETRF(order, M, N, A, lda, ipiv); -} - -template <> -int getri(const CBLAS_ORDER order, - const int N, - float* A, - const int lda, - const int* ipiv) { - return dynload::PADDLE_SGETRI(order, N, A, lda, ipiv); -} - -template <> -int getri(const CBLAS_ORDER order, - const int N, - double* A, - const int lda, - const int* ipiv) { - return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv); -} - -#ifndef PADDLE_USE_EIGEN_FOR_BLAS -template <> -void axpy(const int n, const float alpha, const float* x, float* y) { - cblas_saxpy(n, alpha, x, 1, y, 1); -} - -template <> -void axpy(const int n, const double alpha, const double* x, double* y) { - cblas_daxpy(n, alpha, x, 1, y, 1); -} - -template <> -float dotProduct(const int n, const float* x, const float* y) { - return cblas_sdot(n, x, 1, y, 1); -} - -template <> -double dotProduct(const int n, const double* x, const double* y) { - return cblas_ddot(n, x, 1, y, 1); -} -#endif - -#if defined(PADDLE_WITH_MKLML) - -template <> -void vExp(const int n, const float* a, float* r) { - vsExp(n, a, r); -} - -template <> -void vExp(const int n, const double* a, double* r) { - vdExp(n, a, r); -} - -template <> -void vPow(const int n, const float* a, const float b, float* r) { - vsPowx(n, a, b, r); -} - -template <> -void vPow(const int n, const double* a, const double b, double* r) { - vdPowx(n, a, b, r); -} - -template <> -void vLog(const int n, const float* a, float* r) { - vsLn(n, a, r); -} - -template <> -void vLog(const int n, const double* a, double* r) { - vdLn(n, a, r); -} - -template <> -void vAdd(const int n, const float* a, const float* b, float* r) { - vsAdd(n, a, b, r); -} - -template <> -void vAdd(const int n, const double* a, const double* b, double* r) { - vdAdd(n, a, b, r); -} - -template <> -void vTanh(const int n, const float* a, float* r) { - vsTanh(n, a, r); -} - -template <> -void vTanh(const int n, const double* a, double* r) { - vdTanh(n, a, r); -} - -template <> -void vInvSqrt(const int n, const float* a, float* r) { - vsInvSqrt(n, a, r); -} - -template <> -void vInvSqrt(const int n, const double* a, double* r) { - vdInvSqrt(n, a, r); -} - -template <> -void vLog1p(const int n, const float* a, float* r) { - vsLog1p(n, a, r); -} - -template <> -void vLog1p(const int n, const double* a, double* r) { - vdLog1p(n, a, r); -} -#else - -DEFINE_MATRIX_BINARY_OP(vExp, b = std::exp(a)); -template -void vExp(const int n, const T* a, T* r) { - hl_cpu_apply_binary_op, 0, 0>( - binary::vExp(), const_cast(a), r, 1, n, n, n); -} - -DEFINE_MATRIX_BINARY_OP(vLog, b = std::log(a)); -template -void vLog(const int n, const T* a, T* r) { - hl_cpu_apply_binary_op, 0, 0>( - binary::vLog(), const_cast(a), r, 1, n, n, n); -} - -DEFINE_MATRIX_BINARY_PARAMETER_OP(vPow, ONE_PARAMETER, b = std::pow(a, p)); -template -void vPow(const int n, const T* a, const T b, T* r) { - hl_cpu_apply_binary_op, 0, 0>( - binary::vPow(b), const_cast(a), r, 1, n, n, n); -} - -DEFINE_MATRIX_TERNARY_OP(vAdd, c = a + b); -template -void vAdd(const int n, const T* a, const T* b, T* r) { - hl_cpu_apply_ternary_op, 0, 0>(ternary::vAdd(), - const_cast(a), - const_cast(b), - r, - 1, - n, - n, - n, - n); -} - -DEFINE_MATRIX_BINARY_OP(vInvSqrt, b = 1.0f / std::sqrt(a)); -template -void vInvSqrt(const int n, const T* a, T* r) { - hl_cpu_apply_binary_op, 0, 0>( - binary::vInvSqrt(), const_cast(a), r, 1, n, n, n); -} - -DEFINE_MATRIX_BINARY_OP(vLog1p, b = std::log(1.0f + a)); -template -void vLog1p(const int n, const T* a, T* r) { - hl_cpu_apply_binary_op, 0, 0>( - binary::vLog1p(), const_cast(a), r, 1, n, n, n); -} - -DEFINE_MATRIX_BINARY_OP(vTanh, T tmp = -2.0 * a; - tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp; - b = 2.0 / (1.0 + std::exp(tmp)) - 1.0); -template -void vTanh(const int n, const T* a, T* r) { - hl_cpu_apply_binary_op, 0, 0>( - binary::vTanh(), const_cast(a), r, 1, n, n, n); -} - -template void vExp(const int n, const float* a, float* r); -template void vExp(const int n, const double* a, double* r); -template void vLog(const int n, const float* a, float* r); -template void vLog(const int n, const double* a, double* r); -template void vPow(const int n, const float* a, const float b, float* r); -template void vPow(const int n, const double* a, const double b, double* r); -template void vAdd(const int n, const float* a, const float* b, float* r); -template void vAdd(const int n, const double* a, const double* b, double* r); -template void vInvSqrt(const int n, const double* a, double* r); -template void vInvSqrt(const int n, const float* a, float* r); -template void vLog1p(const int n, const float* a, float* r); -template void vLog1p(const int n, const double* a, double* r); -template void vTanh(const int n, const float* a, float* r); -template void vTanh(const int n, const double* a, double* r); -#endif -} // namespace paddle diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h deleted file mode 100644 index f3d8b1a39e849d5f5a9e79cf33252b60170ced81..0000000000000000000000000000000000000000 --- a/paddle/math/MathFunctions.h +++ /dev/null @@ -1,130 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifndef MATHFUNCTIONS_H_ -#define MATHFUNCTIONS_H_ - -#ifdef PADDLE_WITH_MKLML -#include -#include -#include -#endif - -#if defined(PADDLE_USE_VECLIB) -extern "C" { -#include -#include -} -#endif - -#ifdef PADDLE_USE_OPENBLAS -#include -#include -#endif - -#ifndef LAPACK_FOUND -extern "C" { -#ifndef PADDLE_USE_EIGEN_FOR_BLAS -#include -#else -typedef enum CBLAS_ORDER { - CblasRowMajor = 101, - CblasColMajor = 102 -} CBLAS_ORDER; -#endif -int LAPACKE_sgetrf( - int matrix_layout, int m, int n, float* a, int lda, int* ipiv); -int LAPACKE_dgetrf( - int matrix_layout, int m, int n, double* a, int lda, int* ipiv); -int LAPACKE_sgetri( - int matrix_layout, int n, float* a, int lda, const int* ipiv); -int LAPACKE_dgetri( - int matrix_layout, int n, double* a, int lda, const int* ipiv); -} -#endif - -#include - -namespace paddle { - -#ifndef PADDLE_USE_EIGEN_FOR_BLAS -template -void gemm(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc); -#endif - -template -int getrf(const CBLAS_ORDER Order, - const int M, - const int N, - T* A, - const int lda, - int* ipiv); - -template -int getri( - const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv); - -template -void axpy(const int n, const T alpha, const T* x, T* y) { - /// y = y + alpha * x - for (int i = 0; i < n; i++) { - y[i] = y[i] + alpha * x[i]; - } -} - -template -T dotProduct(const int n, const T* x, const T* y) { - T result = static_cast(0); - for (int i = 0; i < n; i++) { - result += x[i] * y[i]; - } - return result; -} - -template -void vExp(const int n, const T* a, T* r); - -template -void vPow(const int n, const T* a, const T b, T* r); - -template -void vLog(const int n, const T* a, T* r); - -template -void vAdd(const int n, const T* a, const T* b, T* r); - -template -void vInvSqrt(const int n, const T* a, T* r); - -template -void vLog1p(const int n, const T* a, T* r); - -template -void vTanh(const int n, const T* a, T* r); - -} // namespace paddle - -#endif // MATHFUNCTIONS_H_ diff --git a/paddle/math/MathUtils.cpp b/paddle/math/MathUtils.cpp deleted file mode 100644 index b2afdbcd51a3cf5d3e6f3e2bb14902bf78fe68c8..0000000000000000000000000000000000000000 --- a/paddle/math/MathUtils.cpp +++ /dev/null @@ -1,97 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "MathUtils.h" -#include -#include "Vector.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -/*if csc, major is cols and minor is rows, else - * major is rows and minor is cols, according to - * major value to initialize minor value" - */ -void sparseRand( - int* major, int* minor, int nnz, int majorLen, int minorMax, bool useGpu) { - CHECK(size_t(nnz) >= size_t(1)); - int* cpuMajor; - int* cpuMinor; - CpuIVector cpuMinorVec(nnz); - CpuIVector cpuMajorVec(majorLen); - if (useGpu) { - cpuMajor = cpuMajorVec.getData(); - cpuMinor = cpuMinorVec.getData(); - } else { - cpuMajor = major; - cpuMinor = minor; - } - - /*major value init*/ - for (int i = 0; i < majorLen - 1; i++) { - cpuMajor[i] = 1.0 * i * nnz / (majorLen - 1); - } - cpuMajor[majorLen - 1] = nnz; - - /*minor value init according to major value*/ - std::vector used(minorMax, 0); - for (int i = 0; i < majorLen - 1; i++) { - CHECK_LE(cpuMajor[i + 1] - cpuMajor[i], minorMax); - used.assign(minorMax, 0); - for (int j = cpuMajor[i]; j < cpuMajor[i + 1]; j++) { - int idx = ::rand() % minorMax; - while (used[idx]) { - idx = ::rand() % minorMax; - } - cpuMinor[j] = idx; - used[idx] = 1; - } - std::sort(cpuMinor + cpuMajor[i], - cpuMinor + cpuMajor[i + 1], - [](int a, int b) { return a < b; }); - } - /*memcpy result to gpu*/ - if (useGpu) { - hl_memcpy_host2device(major, cpuMajor, sizeof(int) * majorLen); - hl_memcpy_host2device(minor, cpuMinor, sizeof(int) * nnz); - } -} - -int outputSize( - int imageSize, int filterSize, int padding, int stride, bool caffeMode) { - int outputSize; - if (!caffeMode) { - outputSize = - (imageSize - filterSize + 2 * padding + stride - 1) / stride + 1; - } else { - outputSize = (imageSize - filterSize + 2 * padding) / stride + 1; - } - CHECK_GE(outputSize, 1); - return outputSize; -} - -int imageSize( - int outputSize, int filterSize, int padding, int stride, bool caffeMode) { - int imageSize; - if (!caffeMode) { - imageSize = - (outputSize - 1) * stride + filterSize - 2 * padding - stride + 1; - } else { - imageSize = (outputSize - 1) * stride + filterSize - 2 * padding; - } - CHECK_GE(imageSize, 1); - return imageSize; -} - -} // namespace paddle diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp deleted file mode 100644 index bcd6dfe1fda6b1243007b0c26a6e0087eedcc10c..0000000000000000000000000000000000000000 --- a/paddle/math/Matrix.cpp +++ /dev/null @@ -1,4787 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Matrix.h" -#include "MathFunctions.h" -#include "SparseMatrix.h" -#include "SparseRowMatrix.h" - -#include -#include -#include - -#include -#include "hl_cnn.h" -#include "hl_gpu.h" -#include "hl_table_apply.h" -#include "hl_top_k.h" -#include "paddle/utils/Logging.h" - -#include "NEONFunctions.h" -#include "paddle/function/GemmFunctor.h" -#include "paddle/utils/ThreadLocal.h" - -#include "SIMDFunctions.h" - -namespace paddle { - -inline real _pow(real a, real beta) { return std::pow(a, beta); } - -inline real _square(real a) { return a * a; } - -inline real _safelog(real a) { return a > 0.0f ? std::log(a) : -40.0f; } - -Matrix::Matrix(MemoryHandlePtr memHandle, - size_t height, - size_t width, - bool trans, - bool use_gpu) - : BaseMatrix( - height, - width, - memHandle ? (reinterpret_cast(memHandle->getBuf())) : nullptr, - trans, - use_gpu) { - elementCnt_ = width * height; - memoryHandle_ = memHandle; -} - -Matrix::Matrix( - real* data, size_t height, size_t width, bool trans, bool use_gpu) - : BaseMatrix(height, width, data, trans, use_gpu) { - elementCnt_ = width * height; -} - -Matrix::Matrix(real* data, - size_t height, - size_t width, - size_t stride, - bool trans, - bool use_gpu) - : BaseMatrix(height, width, stride, data, trans, use_gpu) { - elementCnt_ = width * height; -} - -MatrixPtr Matrix::createSparseMatrix(real* data, - int* row, - int* col, - size_t height, - size_t width, - size_t nnz, /* used to allocate space */ - SparseValueType valueType, /*value type*/ - SparseFormat format, - bool trans, - bool useGpu) { - if (useGpu) { - return std::make_shared( - data, row, col, height, width, nnz, valueType, format, trans); - } else { - return std::make_shared( - data, row, col, height, width, nnz, valueType, format, trans); - } -} - -MatrixPtr Matrix::createSparseMatrix(size_t height, - size_t width, - size_t nnz, /* used to allocate space */ - SparseValueType valueType, /*value type*/ - SparseFormat format, - bool trans, - bool useGpu) { - if (useGpu) { - return std::make_shared( - height, width, nnz, valueType, format, trans); - } else { - return std::make_shared( - height, width, nnz, valueType, format, trans); - } -} - -MatrixPtr Matrix::create(MemoryHandlePtr memHandle, - size_t height, - size_t width, - bool trans) { - if (auto gpuHandle = std::dynamic_pointer_cast(memHandle)) { - return std::make_shared(gpuHandle, height, width, trans); - } else if (auto cpuHandle = - std::dynamic_pointer_cast(memHandle)) { - return std::make_shared(cpuHandle, height, width, trans); - } else { - LOG(FATAL) << "Wrong"; - return nullptr; - } -} - -MatrixPtr Matrix::create(size_t height, size_t width, bool trans, bool useGpu) { - if (useGpu) { - return std::make_shared(height, width, trans); - } else { - return std::make_shared(height, width, trans); - } -} - -MatrixPtr Matrix::create( - real* data, size_t height, size_t width, bool trans, bool useGpu) { - if (useGpu) { - return std::make_shared(data, height, width, trans); - } else { - return std::make_shared(data, height, width, trans); - } -} - -MatrixPtr Matrix::create(real* data, - size_t height, - size_t width, - size_t stride, - bool trans, - bool useGpu) { - if (useGpu) { - return std::make_shared(data, height, width, stride, trans); - } else { - return std::make_shared(data, height, width, stride, trans); - } -} - -MatrixPtr Matrix::createSparseMatrix(size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - bool trans, - bool useGpu) { - if (useGpu) { - return std::make_shared( - height, width, nnz, valueType, SPARSE_CSR, trans); - } else { - return std::make_shared( - height, width, nnz, valueType, SPARSE_CSR, trans); - } -} - -void Matrix::resizeOrCreate( - MatrixPtr& matrix, size_t height, size_t width, bool trans, bool useGpu) { - if (!matrix) { - matrix = Matrix::create(height, width, trans, useGpu); - } else { - CHECK_EQ(matrix->useGpu(), useGpu); - matrix->resize(height, width); - } -} - -void Matrix::resizeOrCreateSparseMatrix(MatrixPtr& matrix, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans, - bool useGpu) { - if (!matrix) { - matrix = Matrix::createSparseMatrix( - height, width, nnz, valueType, format, trans, useGpu); - } else { - CHECK(dynamic_cast(matrix.get()) || - dynamic_cast(matrix.get())); - CHECK_EQ(matrix->useGpu(), useGpu); - matrix->resize(height, width, nnz, valueType, format); - } -} - -void Matrix::reshape(size_t height, size_t width) { - CHECK(isContiguous()); - CHECK(height_ * width_ == height * width); - height_ = height; - width_ = width; - stride_ = width_; -} - -MatrixPtr Matrix::subMatrix(size_t startRow, - size_t endRow, - size_t startCol, - size_t endCol) { - CHECK_LE(startRow, endRow); - CHECK_LE(endRow, getHeight()); - CHECK_LE(startCol, endCol); - CHECK_LE(endCol, getWidth()); - - return Matrix::create(getData() + startRow * getStride() + startCol, - endRow - startRow, - endCol - startCol, - getStride(), - trans_, - useGpu_); -} - -void Matrix::setDiag(real value) { - CHECK(data_ != NULL); - CHECK_EQ(height_, width_); - - zeroMem(); - BaseMatrix diag(height_, 1, stride_ + 1, data_, false, useGpu_); - diag.assign(value); -} - -GpuMatrix::GpuMatrix(size_t height, size_t width, bool trans) - : Matrix(std::make_shared(height * width * sizeof(real)), - height, - width, - trans, - true) {} - -GpuMatrix::~GpuMatrix() {} - -void GpuMatrix::zeroMem() { - CHECK(data_ != NULL); - zero(); -} - -void GpuMatrix::resetOne() { - CHECK(data_ != NULL); - one(); -} - -void GpuMatrix::resize(size_t newHeight, size_t newWidth) { - size_t newSize = newHeight * newWidth; - if (NULL == memoryHandle_.get() || - newSize * sizeof(real) > memoryHandle_->getAllocSize()) { - memoryHandle_ = std::make_shared(newSize * sizeof(real)); - data_ = reinterpret_cast(memoryHandle_->getBuf()); - } - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newSize; - stride_ = width_; -} - -real GpuMatrix::getElement(size_t x, size_t y) const { - real elem = 0; - hl_memcpy_device2host(&elem, &data_[x * stride_ + y], sizeof(real)); - return elem; -} - -real GpuMatrix::getSum() { - CHECK(isContiguous()); - real sum = 0.0f; - hl_vector_sum(data_, &sum, height_ * width_); - return sum; -} - -real GpuMatrix::getMin() { - CHECK(isContiguous()); - auto vec = GpuVector(height_ * width_, data_); - return vec.getMin(); -} - -real GpuMatrix::getMax() { - CHECK(isContiguous()); - auto vec = GpuVector(height_ * width_, data_); - return vec.getMax(); -} - -void GpuMatrix::accumulateColSum(Matrix& src) { - CHECK_EQ(getWidth(), src.getWidth()); - CHECK_EQ(getHeight(), (size_t)1); - sumCols(src, 1.0, 1.0); -} - -real GpuMatrix::getAbsSum() { - CHECK(isContiguous()); - real sum = 0.0f; - hl_vector_abs_sum(data_, &sum, height_ * width_); - return sum; -} - -void GpuMatrix::copyFrom(const Matrix& src) { - CHECK(isContiguous()); - CHECK(src.isContiguous()); - CHECK(elementCnt_ == src.getElementCnt()); - - if (typeid(src) == typeid(CpuMatrix)) { - hl_memcpy_host2device( - data_, const_cast(src.getData()), sizeof(real) * elementCnt_); - } else if (typeid(src) == typeid(GpuMatrix)) { - hl_memcpy_device2device( - data_, const_cast(src.getData()), sizeof(real) * elementCnt_); - } else { - LOG(FATAL) << "Wrong"; - } -} - -void GpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { - CHECK(isContiguous()); - CHECK(src.isContiguous()); - CHECK(elementCnt_ == src.getElementCnt()); - hl_memcpy_async(this->getData(), - const_cast(src.getData()), - sizeof(real) * elementCnt_, - stream); -} - -void GpuMatrix::copyFrom(const real* hostSrc, size_t size) { - CHECK(isContiguous()); - CHECK(size <= elementCnt_); - hl_memcpy_host2device(data_, const_cast(hostSrc), sizeof(real) * size); -} - -void GpuMatrix::copyFrom(const real* hostSrc, const int64_t* seq) { - LOG(FATAL) << "not implemented"; -} - -void GpuMatrix::copyFrom(const IVector& src) { - CHECK(isContiguous()); - CpuMatrix matrix(src.getSize(), 1, false); - matrix.copyFrom(src); - copyFrom(matrix); -} - -void GpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) { - size_t height = getHeight(); - size_t width = getWidth(); - CHECK_EQ(b.getWidth(), width); - real* dst = getData(); - real* src = b.getData(); - const int* index = rowIndex.getData(); - hl_sequence2batch_copy(dst, src, index, width, height, true); -} - -MatrixPtr GpuMatrix::clone(size_t height, size_t width, bool useGpu) { - CHECK(isContiguous()); - - if (height == 0 && width == 0) { - height = height_; - width = width_; - } - - CHECK(width && height); - - if (useGpu) { - return std::make_shared(height, width); - } else { - return std::make_shared(height, width); - } -} - -MatrixPtr GpuMatrix::getTranspose() { - if (memoryHandle_.get() != NULL) { - MatrixPtr copy_T( - new GpuMatrix(std::dynamic_pointer_cast(memoryHandle_), - height_, - width_, - true)); - return copy_T; - } else { - MatrixPtr copy_T(new GpuMatrix(data_, height_, width_, true)); - return copy_T; - } -} - -void GpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { - if (memAlloc) { - matTrans = std::make_shared(width_, height_); - } else { - CHECK(matTrans != NULL); - CHECK_EQ(matTrans->getHeight(), width_); - CHECK_EQ(matTrans->getWidth(), height_); - } - real* dataTrans = matTrans->getData(); - real* data = getData(); - int lda = getStride(); - int ldc = matTrans->getStride(); - - hl_matrix_transpose(data, dataTrans, height_, width_, lda, ldc); -} - -void GpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) { - if (memAlloc) { - matRot = std::make_shared(width_, height_); - } else { - CHECK(matRot != NULL); - CHECK_EQ(matRot->getHeight(), width_); - CHECK_EQ(matRot->getWidth(), height_); - } - - real* dataRot = matRot->getData(); - real* data = getData(); - hl_matrix_rotate(data, dataRot, height_, width_, clockWise); -} - -MatrixPtr GpuMatrix::getInverse() { - MatrixPtr matInv; - inverse(matInv, true); - return matInv; -} - -void GpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) { - CHECK_EQ(height_, width_); - - if (memAlloc) { - matInv = std::make_shared(height_, width_); - } else { - CHECK(matInv != NULL); - } - - real* data = getData(); - real* dataInv = matInv->getData(); - int lda = getStride(); - int ldc = matInv->getStride(); - - hl_matrix_inverse(data, dataInv, height_, lda, ldc); -} - -void GpuMatrix::addBias(Matrix& b, real scale) { - CHECK(b.getHeight() == 1) << "the Bias should be a vector"; - BaseMatrix::addBias(b, scale); -} - -void GpuMatrix::addSharedBias(Matrix& b, real scale) { - CHECK(b.getHeight() == 1) << "the Bias should be a vector"; - CHECK_LE(b.getWidth(), getWidth()); - CHECK_EQ(getWidth() % b.getWidth(), 0UL); - hl_matrix_add_shared_bias( - getData(), b.getData(), b.getWidth(), getHeight(), getWidth(), scale); -} - -void GpuMatrix::collectBias(Matrix& a, real scale) { -#ifdef PADDLE_WITH_CUDA - CHECK_EQ(getHeight(), (size_t)1); - CHECK_EQ(width_, a.getWidth()); - GpuSparseMatrix* sMatPtr = dynamic_cast(&a); - if (!sMatPtr) { - sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1); - } else { - real* data = getData(); - hl_sparse_matrix_s A_d = sMatPtr->sMatrix_.get(); - hl_sparse_matrix_column_sum(data, A_d, sMatPtr->getHeight(), width_, scale); - } -#endif -} - -void GpuMatrix::collectSharedBias(Matrix& a, real scale) { - CHECK_EQ(getHeight(), (size_t)1); - CHECK_EQ(a.getWidth() % getWidth(), 0UL); - hl_matrix_collect_shared_bias( - getData(), a.getData(), getWidth(), a.getHeight(), a.getWidth(), scale); -} - -void GpuMatrix::sequenceAvgForward(Matrix& a, - const IVector& startsPos, - int mode) { - size_t height = getHeight(); - size_t width = getWidth(); - CHECK_EQ(height, startsPos.getSize() - 1); - CHECK_EQ(width, a.getWidth()); - real* dst = getData(); - real* src = a.getData(); - const int* starts = startsPos.getData(); - - hl_sequence_avg_forward(dst, src, starts, height, width, mode); -} - -void GpuMatrix::sequenceAvgBackward(Matrix& a, - const IVector& startsPos, - int mode) { - size_t height = a.getHeight(); - size_t width = getWidth(); - CHECK_EQ(height, startsPos.getSize() - 1); - CHECK_EQ(width, a.getWidth()); - real* dst = getData(); - real* src = a.getData(); - const int* starts = startsPos.getData(); - - hl_sequence_avg_backward(dst, src, starts, height, width, mode); -} - -/* this = scaleAB*(a*b) + scaleT*this */ -void GpuMatrix::mul(const GpuMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT) { - CHECK(!isTransposed()) << "Not supported"; - - if (!a.isTransposed() && !b.isTransposed()) { - CHECK_EQ(width_, b.width_); - CHECK_EQ(height_, a.height_); - CHECK_EQ(a.width_, b.height_); - } else if (a.isTransposed() && !b.isTransposed()) { - CHECK_EQ(width_, b.width_); - CHECK_EQ(height_, a.width_); - CHECK_EQ(a.height_, b.height_); - } else if (!a.isTransposed() && b.isTransposed()) { - CHECK_EQ(width_, b.height_); - CHECK_EQ(height_, a.height_); - CHECK_EQ(a.width_, b.width_); - } else { - LOG(FATAL) << "Is not supported"; - } - - real* A_d = a.data_; - real* B_d = b.data_; - real* C_d = data_; - int dimM = getHeight(); - int dimN = getWidth(); - int dimK = !a.isTransposed() ? a.width_ : a.height_; - int lda = a.getStride(); - int ldb = b.getStride(); - int ldc = getStride(); - hl_trans_op_t transa = !a.isTransposed() ? HPPL_OP_N : HPPL_OP_T; - hl_trans_op_t transb = !b.isTransposed() ? HPPL_OP_N : HPPL_OP_T; - - hl_matrix_mul(A_d, - transa, - B_d, - transb, - C_d, - dimM, - dimN, - dimK, - scaleAB, - scaleT, - lda, - ldb, - ldc); -} - -void GpuMatrix::mul(const GpuSparseMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT) { -#ifdef PADDLE_WITH_CUDA - CHECK(isContiguous()); - CHECK(b.isContiguous()); - CHECK(b.useGpu_ == true) << "Matrix type are not equal"; - CHECK(!trans_ && !b.trans_) << "not supported"; - - if (!a.trans_) { - CHECK(width_ == b.width_ && height_ == a.height_ && a.width_ == b.height_) - << "Matrix dimensions are not equal"; - } else { - CHECK(width_ == b.width_ && height_ == a.width_ && a.height_ == b.height_) - << "Matrix dimensions are not equal"; - } - hl_trans_op_t transA = a.trans_ ? HPPL_OP_T : HPPL_OP_N; - hl_sparse_matrix_s A_d = a.sMatrix_.get(); - real* B_d = b.data_; - real* C_d = data_; - hl_matrix_csr_mul_dense(A_d, - transA, - B_d, - HPPL_OP_N, - C_d, - height_, - width_, - b.height_, - scaleAB, - scaleT); -#endif -} - -void GpuMatrix::mul(const GpuMatrix& a, - const GpuSparseMatrix& b, - real scaleAB, - real scaleT) { -#ifdef PADDLE_WITH_CUDA - CHECK(isContiguous()); - CHECK(a.isContiguous()); - CHECK(a.useGpu_ == true) << "Matrix type are not equal"; - - hl_sparse_matrix_s B_d = b.sMatrix_.get(); - real* A_d = a.data_; - real* C_d = data_; - hl_trans_op_t transB = b.trans_ ? HPPL_OP_T : HPPL_OP_N; - if (!b.trans_) { - CHECK(width_ == b.width_ && height_ == a.height_ && a.width_ == b.height_) - << "Matrix dimensions are not equal"; - } else { - CHECK(width_ == b.height_ && height_ == a.height_ && a.width_ == b.width_) - << "Matrix dimensions are not equal"; - } - if (b.format_ == SPARSE_CSC) { - hl_matrix_dense_mul_csc(A_d, - HPPL_OP_N, - B_d, - transB, - C_d, - height_, - width_, - a.width_, - scaleAB, - scaleT); - } else { - hl_matrix_dense_mul_csr(A_d, - HPPL_OP_N, - B_d, - transB, - C_d, - height_, - width_, - a.width_, - scaleAB, - scaleT); - } -#endif -} - -/* this = a*b */ -void GpuMatrix::mul(const Matrix& a, const Matrix& b) { mul(a, b, 1.0, 0.0); } - -void GpuMatrix::mul(const Matrix& a, - const Matrix& b, - real scaleAB, - real scaleT) { - const auto a_ptr = dynamic_cast(&a); - const auto b_ptr = dynamic_cast(&b); - const auto a_ptr_s = dynamic_cast(&a); - const auto b_ptr_s = dynamic_cast(&b); - - if (a_ptr && b_ptr) { - mul(*a_ptr, *b_ptr, scaleAB, scaleT); - } else if (a_ptr_s && b_ptr) { - mul(*a_ptr_s, *b_ptr, scaleAB, scaleT); - } else if (a_ptr && b_ptr_s) { - mul(*a_ptr, *b_ptr_s, scaleAB, scaleT); - } else { - LOG(FATAL) << "Not supported"; - } -} - -/* this = this* b */ -void GpuMatrix::rightMul(Matrix& b) { rightMul(b, 1.0, 0.0); } - -/* this = scaleAB*(this*b) + scaleT*this */ -void GpuMatrix::rightMul(Matrix& b, real scaleAB, real scaleT) { - CHECK(dynamic_cast(&b)); - CHECK(!isTransposed()) << "Not supported"; - CHECK(!b.isTransposed()) << "Not supported"; - mul(*this, *dynamic_cast(&b), scaleAB, scaleT); -} - -/* this = a*this */ -void GpuMatrix::leftMul(Matrix& a) { leftMul(a, 1.0, 0.0); } - -/* this = scaleAB*(a*this) + scaleT*this */ -void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) { - CHECK(dynamic_cast(&a)); - CHECK(!isTransposed()) << "Not supported"; - CHECK(!a.isTransposed()) << "Not supported"; - mul(*dynamic_cast(&a), *this, scaleAB, scaleT); -} - -void GpuMatrix::selectRows(Matrix& table, IVector& ids) { -#ifdef PADDLE_WITH_CUDA - CHECK(dynamic_cast(&table)); - CHECK(table.useGpu()); - CHECK(ids.useGpu()); - CHECK_EQ(getHeight(), ids.getSize()); - CHECK_EQ(getWidth(), table.getWidth()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - real* a = getData(); - size_t tableSize = table.getHeight(); - int* index = ids.getData(); - - hl_matrix_select_rows(a, - stride_, - table.getData(), - table.stride_, - index, - numSamples, - tableSize, - dim); -#endif -} - -void GpuMatrix::addToRows(Matrix& table, IVector& ids) { -#ifdef PADDLE_WITH_CUDA - CHECK(dynamic_cast(&table)); - CHECK(table.useGpu()); - CHECK(ids.useGpu()); - CHECK_EQ(getHeight(), ids.getSize()); - CHECK_EQ(getWidth(), table.getWidth()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - real* a = getData(); - size_t tableSize = table.getHeight(); - int* index = ids.getData(); - - hl_matrix_add_to_rows(table.getData(), - table.stride_, - a, - stride_, - index, - numSamples, - tableSize, - dim); -#endif -} - -void GpuMatrix::colMerge(Matrix& src) { - CHECK(src.height_ == height_); - if (!trans_ && !src.trans_) { - sumRows(src, /* scaleSum= */ 1, /* scaleDest= */ 0); - } else { - LOG(FATAL) << "Is not supported"; - } -} - -void GpuMatrix::rowSum(Matrix& sum) { - CHECK_EQ(sum.getHeight(), getHeight()); - CHECK_EQ(sum.getWidth(), (size_t)1); - - sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0); -} - -void GpuMatrix::rowMax(Matrix& max) { - CHECK_EQ(max.getHeight(), getHeight()); - CHECK_EQ(max.getWidth(), (size_t)1); - - max.maxRows(*this); -} - -void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { -#ifdef PADDLE_WITH_CUDA - CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; - size_t numSamples = getHeight(); - size_t beam = maxVal.getWidth(); - CHECK_EQ(maxIds.getSize(), numSamples * beam); - CHECK_EQ(maxVal.getHeight(), numSamples); - CHECK_EQ(maxVal.getWidth(), beam); - - hl_matrix_top_k(maxVal.getData(), - maxVal.getStride(), - maxIds.getData(), - this->getData(), - this->getStride(), - this->getWidth(), - beam, - numSamples); -#endif -} - -void GpuMatrix::colMax(Matrix& max) { - CHECK_EQ(max.getWidth(), getWidth()); - CHECK_EQ(max.getHeight(), (size_t)1); - - max.maxCols(*this); -} - -void GpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) { - LOG(FATAL) << "Is not supported"; -} - -void GpuMatrix::maxoutForward(Matrix& a, - IVector& id, - size_t channels, - size_t groups) { - CHECK(dynamic_cast(&a)); - CHECK(dynamic_cast(&id)); - CHECK_EQ(a.getHeight(), getHeight()); - - size_t size = getWidth(); - size_t batchSize = getHeight(); - const real* input = a.getData(); - real* output = getData(); - int* idForGpu = id.getData(); - - hl_maxout_forward( - input, output, idForGpu, batchSize, size, size / channels, groups); -} - -void GpuMatrix::maxoutBackward(Matrix& a, - IVector& id, - size_t channels, - size_t groups) { - CHECK(dynamic_cast(&a)); - CHECK(dynamic_cast(&id)); - CHECK_EQ(a.getHeight(), getHeight()); - - size_t size = a.getWidth(); - size_t batchSize = getHeight(); - real* input = getData(); - const real* output = a.getData(); - const int* idForGpu = id.getData(); - - hl_maxout_backward( - input, output, idForGpu, batchSize, size, size / channels, groups); -} - -/*calulate the error of classification */ -void GpuMatrix::classificationError(Matrix& output, - IVector& label, - size_t topkSize) { - auto gpuOutput = dynamic_cast(&output); - auto gpuLabel = dynamic_cast(&label); - size_t numSamples = this->getHeight(); - GpuMatrixPtr gpuTopVal = std::make_shared(numSamples, topkSize); - GpuIVectorPtr gpuTopIds = std::make_shared(numSamples * topkSize); - - CHECK(gpuOutput && gpuLabel) << "Invalid argument pointer"; - CHECK(gpuTopVal && gpuTopIds) << "Allocate GPU memory failed"; - CHECK(gpuLabel->getSize() == numSamples) << "Vector size is not equal"; - CHECK(numSamples == gpuOutput->getHeight() && this->getWidth() == 1) - << "Matrix dimensions are not equal"; - - size_t dim = gpuOutput->getWidth(); - hl_matrix_classification_error(gpuTopVal->getData(), - gpuTopVal->getStride(), - gpuTopIds->getData(), - gpuOutput->getData(), - gpuOutput->getStride(), - dim, - topkSize, - numSamples, - gpuLabel->getData(), - this->getData()); -} - -/* copy -log(output[i * width + label]) to this->data[i] */ -void GpuMatrix::oneHotCrossEntropy(Matrix& output, IVector& label) { - GpuMatrix* output_ptr = dynamic_cast(&output); - GpuIVector* label_ptr = dynamic_cast(&label); - - CHECK(output_ptr && label_ptr) << "Invalid argument pointer"; - - CHECK(height_ == label.getSize() && width_ == 1 && height_ == output.height_) - << "Matrix dimensions are not equal"; - - real* A_d = output_ptr->data_; - real* C_d = data_; - int* label_d = label_ptr->getData(); - - hl_matrix_cross_entropy(A_d, C_d, label_d, height_, output.width_); -} - -/* calculate the error of outputV according to label */ -void GpuMatrix::oneHotCrossEntropyBp(Matrix& outputV, IVector& label) { - GpuMatrix* output_ptr = dynamic_cast(&outputV); - GpuIVector* label_ptr = dynamic_cast(&label); - - CHECK(output_ptr && label_ptr) << "Invalid argument pointer"; - - CHECK(height_ == output_ptr->height_ && width_ == output_ptr->width_) - << "Matrix dimensions are not equal"; - - real* output_d = output_ptr->data_; - real* grad_d = data_; - int* label_d = label_ptr->getData(); - - hl_matrix_cross_entropy_bp(grad_d, output_d, label_d, height_, width_); -} - -void GpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output, - IVector& label, - real alpha) { - LOG(FATAL) << "Not implemented"; -} - -void GpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, - IVector& label, - real alpha) { - LOG(FATAL) << "Not implemented"; -} - -void GpuMatrix::softmax(Matrix& output) { - CHECK(output.useGpu()) << "Matrix type are not equal"; - - size_t height = getHeight(); - size_t width = getWidth(); - CHECK(height == output.getHeight() && width == output.getWidth()) - << "Matrix dimensions are not equal"; - - real* inputData = getData(); - real* outputData = output.getData(); - hl_matrix_softmax(inputData, outputData, height, width); -} - -void GpuMatrix::sequenceSoftmax(Matrix& output, const IVector& index) { - CHECK_EQ(getWidth(), 1UL); - CHECK_EQ(output.getWidth(), 1UL); - CHECK(isContiguous()); - - real* inputData = getData(); - real* outputData = output.getData(); - auto starts = index.getData(); - int numSequences = index.getSize() - 1; - hl_sequence_softmax_forward(inputData, outputData, starts, numSequences); -} - -void GpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) { - CHECK(output.useGpu_ == true && sftmaxSum.useGpu_ == true) - << "Matrix type are not equal"; - - CHECK(height_ == output.height_ && width_ == output.width_ && - height_ == sftmaxSum.height_) - << "Matrix dimensions are not equal"; - - real* output_d = output.data_; - real* sftmaxSum_d = sftmaxSum.data_; - real* grad_d = data_; - hl_matrix_softmax_derivative(grad_d, output_d, sftmaxSum_d, height_, width_); -} - -void GpuMatrix::softmaxBackward(Matrix& outputV) { - CHECK(outputV.useGpu()) << "Matrix type are not equal"; - - size_t height = getHeight(); - size_t width = getWidth(); - CHECK(height == outputV.getHeight() && width == outputV.getWidth()) - << "Matrix dimensions are not equal"; - - real* output_grad = getData(); - real* output_value = outputV.getData(); - hl_softmax_backward(output_value, output_grad, height, width); -} - -void GpuMatrix::sumOfSquares(Matrix& output, Matrix& label) { - CHECK_EQ(label.getHeight(), height_); - CHECK_EQ(output.getHeight(), height_); - CHECK_EQ(label.getWidth(), output.getWidth()); - CHECK_EQ((size_t)1, width_); - - auto labelptr = dynamic_cast(&label); - if (labelptr) { - LOG(FATAL) << "not supported: GpuSparseMatrix as label"; - } - - BaseMatrix::sumOfSquaredDiffs(output, - label, - /* scaleSum= */ 1, - /* scaleDest= */ 1); -} - -void GpuMatrix::sumOfSquaresBp(Matrix& outputV, Matrix& label) { - add2(outputV, label, 1, 2, -2); -} - -void GpuMatrix::tanh(Matrix& output) { BaseMatrix::tanh(output); } - -void GpuMatrix::tanhDerivative(Matrix& output) { - BaseMatrix::tanhDerivative(output); -} - -void GpuMatrix::softrelu(Matrix& output) { BaseMatrix::softrelu(output); } - -void GpuMatrix::softreluDerivative(Matrix& output) { - BaseMatrix::softreluDerivative(output); -} - -void GpuMatrix::scaledTanh(Matrix& output, real p1, real p2) { - BaseMatrix::scaledTanh(output, p1, p2); -} - -void GpuMatrix::randomizeUniform() { - CHECK(isContiguous()); - real* data = data_; - size_t size = height_ * width_; - - hl_rand(data, size); -} - -void GpuMatrix::print(std::ostream& os) const { - CHECK(isContiguous()); - CpuMatrix cpuMat(getHeight(), getWidth()); - cpuMat.copyFrom(*this); - cpuMat.print(os); -} - -void GpuMatrix::print(std::ostream& os, size_t height, size_t width) const { - CHECK(isContiguous()); - CpuMatrix cpuMat(getHeight(), getWidth()); - cpuMat.copyFrom(*this); - cpuMat.print(os, height, width); -} - -void GpuMatrix::check(std::ostream& os, Matrix& refMat, bool printDiff) { - CHECK(isContiguous()); - CHECK(height_ == refMat.getHeight()); - CHECK(width_ == refMat.getWidth()); - CpuMatrix cpuRef(height_, width_); - GpuMatrix gpuRef(height_, width_); - cpuRef.copyFrom(refMat); - gpuRef.copyFrom(*this); - size_t diffCnt = 0; - for (size_t i = 0; i < height_; ++i) { - for (size_t j = 0; j < width_; ++j) { - real a = gpuRef.getElement(i, j); - real b = cpuRef.getElement(i, j); - if (fabs(a - b) > 0.00001) { - ++diffCnt; - if (printDiff) { - os << "ref= " << a << " check= " << b << std::endl; - } - } - } - } - LOG(INFO) << "the diffCnt is " << diffCnt; -} - -void GpuMatrix::upsampleForward(Matrix& input, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW) { - CHECK(input.useGpu_ == true) << "Matrix type are not equal"; - CHECK(mask.useGpu_ == true) << "Matrix type are not equal"; - - real* inputData = input.getData(); - real* maskData = mask.getData(); - real* outData = data_; - - size_t batch = input.getHeight(); - - CHECK(imgSizeH * imgSizeW * channels == input.getWidth()); - CHECK(imgSizeH * imgSizeW * channels == mask.getWidth()); - CHECK_EQ(batch, this->getHeight()); - CHECK(width_ == outputH * outputW * channels); - hl_upsample_forward(inputData, - maskData, - batch, - imgSizeH, - imgSizeW, - channels, - outputH, - outputW, - outData); -} - -void GpuMatrix::upsampleBackward(Matrix& outputGrad, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW) { - CHECK(outputGrad.useGpu_ == true) << "Matrix type are not equal"; - CHECK(mask.useGpu_ == true) << "Matrix type are not equal"; - - real* outputGradData = outputGrad.getData(); - real* maskData = mask.getData(); - real* inputGradData = data_; - size_t batch = outputGrad.getHeight(); - - CHECK(imgSizeH * imgSizeW == this->getWidth() / channels); - CHECK_EQ(batch, this->getHeight()); - CHECK_EQ(channels * outputH * outputW, outputGrad.getWidth()); - hl_upsample_backward(outputGradData, - maskData, - batch, - imgSizeH, - imgSizeW, - channels, - outputH, - outputW, - inputGradData); -} - -void GpuMatrix::maxPoolForward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - MatrixPtr maskMatP) { - CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal"; - - real* inputData = inputMat.getData(); - real* maskData = NULL; - size_t frameNum = inputMat.getHeight(); - CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); - CHECK(height_ == inputMat.getHeight()); - CHECK(width_ == outputH * outputW * channels); - - if (maskMatP != NULL) { - CHECK(maskMatP->useGpu_ == true) << "Matrix type are not equal"; - CHECK(outputH * outputW * channels == maskMatP->getWidth()); - maskData = maskMatP->getData(); - } - - hl_maxpool_forward(frameNum, - inputData, - channels, - imgSizeH, - imgSizeW, - outputH, - outputW, - sizeX, - sizeY, - strideH, - strideW, - paddingH, - paddingW, - data_, - getStride(), - maskData); -} - -void GpuMatrix::maxPoolBackward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - Matrix& outGrad, - Matrix& outV, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW) { - CHECK(inputMat.useGpu_ == true && outGrad.useGpu_ == true && - outV.useGpu_ == true) - << "Matrix type are not equal"; - - real* inputData = inputMat.getData(); - real* outData = outV.getData(); - real* outDiff = outGrad.getData(); - size_t frameNum = inputMat.getHeight(); - size_t channels = outV.getWidth() / outputH / outputW; - CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); - CHECK(height_ == inputMat.getHeight()); - CHECK(outGrad.getHeight() == outV.getHeight() && - outGrad.getWidth() == outV.getWidth()); - - hl_maxpool_backward(frameNum, - inputData, - outData, - outDiff, - channels, - imgSizeH, - imgSizeW, - outputH, - outputW, - sizeX, - sizeY, - strideH, - strideW, - paddingH, - paddingW, - scaleTargets, - scaleOutput, - data_, - outGrad.getStride()); -} - -void GpuMatrix::avgPoolForward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - bool excludeMode) { - CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal"; - - real* inputData = inputMat.getData(); - size_t frameNum = inputMat.getHeight(); - CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); - CHECK(height_ == inputMat.getHeight()); - CHECK(width_ == outputH * outputW * channels); - - hl_avgpool_forward(frameNum, - inputData, - channels, - imgSizeH, - imgSizeW, - outputH, - outputW, - sizeX, - sizeY, - strideH, - strideW, - paddingH, - paddingW, - data_, - getStride(), - excludeMode); -} - -void GpuMatrix::avgPoolBackward(Matrix& outGrad, - size_t imgSizeH, - size_t imgSizeW, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW, - bool excludeMode) { - CHECK(outGrad.useGpu_ == true) << "Matrix type are not equal"; - - real* outDiff = outGrad.getData(); - size_t frameNum = outGrad.getHeight(); - size_t channels = outGrad.getWidth() / outputH / outputW; - CHECK(imgSizeH * imgSizeW * channels == width_); - CHECK(height_ == outGrad.getHeight()); - CHECK(outGrad.getWidth() == outputH * outputW * channels); - - hl_avgpool_backward(frameNum, - outDiff, - channels, - imgSizeH, - imgSizeW, - outputH, - outputW, - sizeX, - sizeY, - strideH, - strideW, - paddingH, - paddingW, - scaleTargets, - scaleOutput, - data_, - outGrad.getStride(), - excludeMode); -} - -void GpuMatrix::maxPool3DForward(Matrix& inputMat, - Matrix& maxPoolIdx, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW) { - CHECK(inputMat.useGpu_) << "Matrix type are not correct"; - - real* inputData = inputMat.getData(); - real* maxPoolIdxData = maxPoolIdx.getData(); - size_t num = inputMat.getHeight(); - CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth()); - CHECK(height_ == inputMat.getHeight()); - CHECK(width_ == outputD * outputH * outputW * channels); - - hl_maxpool3D_forward(num, - inputData, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outputD, - outputH, - outputW, - sizeZ, - sizeY, - sizeX, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - getData(), - maxPoolIdxData, - getStride()); -} - -void GpuMatrix::maxPool3DBackward(Matrix& outGrad, - Matrix& maxPoolIdx, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput) { - CHECK(outGrad.useGpu_ && maxPoolIdx.useGpu_) << "Matrix type are not equal"; - - real* outDiff = outGrad.getData(); - real* maxPoolIdxData = maxPoolIdx.getData(); - size_t frameNum = getHeight(); - size_t channels = outGrad.getWidth() / outputD / outputH / outputW; - CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth()); - CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() && - outGrad.getWidth() == maxPoolIdx.getWidth()); - - hl_maxpool3D_backward(frameNum, - outDiff, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outputD, - outputH, - outputW, - sizeZ, - sizeY, - sizeX, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - scaleTargets, - scaleOutput, - getData(), - maxPoolIdxData, - outGrad.getStride()); -} - -void GpuMatrix::avgPool3DForward(Matrix& inputMat, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW) { - CHECK(inputMat.useGpu_) << "Matrix type are not equal"; - - real* inputData = inputMat.getData(); - size_t frameNum = inputMat.getHeight(); - CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth()); - CHECK(height_ == inputMat.getHeight()); - CHECK(width_ == outputD * outputH * outputW * channels); - - hl_avgpool3D_forward(frameNum, - inputData, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outputD, - outputH, - outputW, - sizeZ, - sizeY, - sizeX, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - getData(), - getStride()); -} - -void GpuMatrix::avgPool3DBackward(Matrix& outGrad, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput) { - CHECK(outGrad.useGpu_) << "Matrix type are not equal"; - - real* outDiff = outGrad.getData(); - size_t frameNum = outGrad.getHeight(); - size_t channels = outGrad.getWidth() / outputD / outputH / outputW; - CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_); - CHECK(height_ == outGrad.getHeight()); - CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels); - - hl_avgpool3D_backward(frameNum, - outDiff, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outputD, - outputH, - outputW, - sizeZ, - sizeY, - sizeX, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - scaleTargets, - scaleOutput, - getData(), - outGrad.getStride()); -} - -void GpuMatrix::maxSequenceForward(Matrix& input, - const IVector& sequence, - IVector& index) { - CHECK(dynamic_cast(&input)); - CHECK(dynamic_cast(&sequence)); - CHECK(dynamic_cast(&index)); - - real* outData = getData(); - real* inputData = input.getData(); - const int* starts = sequence.getData(); - int* maxIndex = index.getData(); - size_t numSequences = getHeight(); - size_t dim = getWidth(); - - CHECK_EQ(dim, input.getWidth()); - CHECK_EQ(numSequences, sequence.getSize() - 1); - CHECK_EQ(numSequences * dim, index.getSize()); - - hl_max_sequence_forward( - inputData, starts, outData, maxIndex, numSequences, dim); -} - -void GpuMatrix::maxSequenceBackward(Matrix& outputGrad, - const IVector& sequence, - IVector& index) { - CHECK(dynamic_cast(&outputGrad)); - CHECK(dynamic_cast(&sequence)); - CHECK(dynamic_cast(&index)); - - real* inputGrad = getData(); - real* outGrad = outputGrad.getData(); - int* maxIndex = index.getData(); - size_t dim = getWidth(); - size_t numSequences = sequence.getSize() - 1; - - CHECK_EQ(dim, outputGrad.getWidth()); - CHECK_EQ(numSequences, outputGrad.getHeight()); - CHECK_EQ(numSequences * dim, index.getSize()); - - hl_max_sequence_backward(outGrad, maxIndex, inputGrad, numSequences, dim); -} - -void GpuMatrix::paramReluForward(Matrix& data, Matrix& W) { - CHECK(data.useGpu_ == true && W.useGpu_ == true) - << "Matrix type are not equal"; - real* input = data.getData(); - real* w = W.getData(); - size_t numElements = data.getWidth(); - size_t numSamples = data.getHeight(); - size_t paraSize = W.getHeight() * W.getWidth(); - CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init - size_t partial_sum = numElements / paraSize; - real* output = getData(); - hl_param_relu_forward(output, input, w, numElements, numSamples, partial_sum); -} - -void GpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) { - CHECK(oGrad.useGpu_ == true && data.useGpu_ == true) - << "Matrix type are not equal"; - real* ograd = oGrad.getData(); - real* input = data.getData(); - real* wgrad = data_; - size_t numElements = data.getWidth(); - size_t numSamples = data.getHeight(); - size_t paraSize = this->getHeight() * this->getWidth(); - CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init - size_t partial_sum = numElements / paraSize; - hl_param_relu_backward_w( - wgrad, ograd, input, numElements, numSamples, partial_sum); -} - -void GpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) { - real* diff = data_; - real* input = data.getData(); - real* ograd = oGrad.getData(); - real* w = W.getData(); - size_t numElements = data.getWidth(); - size_t numSamples = data.getHeight(); - size_t paraSize = W.getHeight() * W.getWidth(); - CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init - size_t partial_sum = numElements / paraSize; - hl_param_relu_backward_diff( - ograd, input, w, diff, numElements, numSamples, partial_sum); -} - -void GpuMatrix::addColumnVector(const Matrix& b) { - BaseMatrix::addColVector(const_cast(b)); -} - -void GpuMatrix::bilinearForward(const Matrix& in, - const size_t inImgH, - const size_t inImgW, - const size_t outImgH, - const size_t outImgW, - const size_t numChannels, - const real ratioH, - const real ratioW) { - CHECK(dynamic_cast(&in)); - - const size_t outputW = getWidth(); - const size_t outputH = getHeight(); - const size_t inputW = in.getWidth(); - const size_t inputH = in.getHeight(); - - real* outData = getData(); - const real* inData = in.getData(); - - if (inImgH == outImgW && inImgW == outImgW) { - this->copyFrom(in); - } else { - hl_bilinear_forward(inData, - inImgH, - inImgW, - inputH, - inputW, - outData, - outImgH, - outImgW, - outputH, - outputW, - numChannels, - ratioH, - ratioW); - } -} - -void GpuMatrix::bilinearBackward(const Matrix& out, - const size_t outImgH, - const size_t outImgW, - const size_t inImgH, - const size_t inImgW, - const size_t numChannels, - const real ratioH, - const real ratioW) { - CHECK(dynamic_cast(&out)); - - const size_t inputW = getWidth(); - const size_t inputH = getHeight(); - const size_t outputW = out.getWidth(); - const size_t outputH = out.getHeight(); - - real* inGrad = getData(); - const real* outGrad = out.getData(); - - if (outImgH == inImgH && outImgW == inImgW) { - this->add(const_cast(out)); - } else { - hl_bilinear_backward(inGrad, - inImgH, - inImgW, - inputH, - inputW, - outGrad, - outImgH, - outImgW, - outputH, - outputW, - numChannels, - ratioH, - ratioW); - } -} - -void GpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) { -#ifdef PADDLE_WITH_CUDA - GpuMatrix* outputPtr = dynamic_cast(&output); - auto labelPtr = dynamic_cast(&label); - - CHECK(outputPtr && labelPtr) << "Invalid argument pointer"; - CHECK(labelPtr->format_ == SPARSE_CSR) << "Matrix format not supported"; - CHECK(height_ == outputPtr->height_ && width_ == 1 && - outputPtr->width_ == labelPtr->getWidth() && - outputPtr->height_ == labelPtr->getHeight()) - << "Matrix dimensions are not equal"; - - real* output_d = outputPtr->data_; - real* entropy_d = data_; - hl_sparse_matrix_s mat_d = labelPtr->sMatrix_.get(); - hl_matrix_multi_binary_cross_entropy( - output_d, entropy_d, mat_d, height_, outputPtr->width_); -#endif -} - -void GpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { -#ifdef PADDLE_WITH_CUDA - GpuMatrix* outputPtr = dynamic_cast(&output); - auto labelPtr = dynamic_cast(&label); - - CHECK(outputPtr && labelPtr) << "Invalid argument pointer"; - CHECK(labelPtr->format_ == SPARSE_CSR) << "Matrix format not supported"; - CHECK(height_ == outputPtr->height_ && width_ == outputPtr->width_ && - outputPtr->width_ == labelPtr->getWidth() && - outputPtr->height_ == labelPtr->getHeight()) - << "Matrix dimensions are not equal"; - - real* output_d = outputPtr->data_; - real* grad_d = data_; - hl_sparse_matrix_s mat_d = labelPtr->sMatrix_.get(); - hl_matrix_multi_binary_cross_entropy_bp( - output_d, grad_d, mat_d, height_, width_); -#endif -} - -void GpuMatrix::vol2Col(real* dataSrc, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW) { - hl_matrix_vol2Col(dataSrc, - channels, - depth, - height, - width, - filterD, - filterH, - filterW, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - getData()); -} - -void GpuMatrix::col2Vol(real* dataDst, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - real alpha, - real beta) { - hl_matrix_col2Vol(dataDst, - channels, - depth, - height, - width, - filterD, - filterH, - filterW, - strideD, - strideH, - strideW, - paddingD, - paddingH, - paddingW, - getData(), - alpha, - beta); -} - -/** - * CpuMatrix - */ - -CpuMatrix::CpuMatrix(size_t height, size_t width, bool trans) - : Matrix(std::make_shared(height * width * sizeof(real)), - height, - width, - trans, - false) {} - -CpuMatrix::~CpuMatrix() {} - -void CpuMatrix::zeroMem() { - CHECK(data_ != NULL); - if (isContiguous()) { - memset(data_, 0, height_ * width_ * sizeof(real)); - } else { - BaseMatrix::zero(); - } -} -void CpuMatrix::resetOne() { - CHECK(data_ != NULL); - BaseMatrix::one(); -} - -void CpuMatrix::copyFrom(const Matrix& src) { - CHECK(isContiguous()); - if (typeid(src) == typeid(GpuMatrix)) { - CHECK(src.isContiguous()); - CHECK(elementCnt_ == src.getElementCnt()); - hl_memcpy_device2host( - data_, const_cast(src.getData()), sizeof(real) * elementCnt_); - } else if (typeid(src) == typeid(CpuMatrix) || - typeid(src) == typeid(SharedCpuMatrix)) { - CHECK(src.isContiguous()); - CHECK(elementCnt_ == src.getElementCnt()); - memcpy(data_, src.getData(), sizeof(real) * elementCnt_); - } else if (typeid(src) == typeid(CpuSparseMatrix)) { - CHECK_GE(elementCnt_, src.getElementCnt()); - copyFrom(dynamic_cast(const_cast(src))); - } else { - LOG(FATAL) << "Wrong"; - } -} - -void CpuMatrix::copyFrom(CpuSparseMatrix& src) { - CHECK(isContiguous()); - CHECK(height_ == src.getHeight()); - CHECK(width_ == src.getWidth()); - memset(data_, 0, sizeof(real) * height_ * width_); - if (src.getValueType() == FLOAT_VALUE) { - if (src.getFormat() == SPARSE_CSC) { - int* rows = src.getRows(); - real* vals = src.getValue(); - for (size_t i = 0; i < width_; i++) { - for (size_t j = src.getColStartIdx(i); j < src.getColStartIdx(i + 1); - j++) { - data_[rows[j] * width_ + i] = vals[j]; - } - } - } else { - int* cols = src.getCols(); - real* vals = src.getValue(); - for (size_t i = 0; i < height_; i++) { - for (size_t j = src.getRowStartIdx(i); j < src.getRowStartIdx(i + 1); - j++) { - data_[i * width_ + cols[j]] = vals[j]; - } - } - } - } else { - if (src.getFormat() == SPARSE_CSC) { - int* rows = src.getRows(); - for (size_t i = 0; i < width_; i++) { - for (size_t j = src.getColStartIdx(i); j < src.getColStartIdx(i + 1); - j++) { - data_[rows[j] * width_ + i] = 1.0; - } - } - } else { - int* cols = src.getCols(); - for (size_t i = 0; i < height_; i++) { - for (size_t j = src.getRowStartIdx(i); j < src.getRowStartIdx(i + 1); - j++) { - data_[i * width_ + cols[j]] = 1.0; - } - } - } - } -} - -void CpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { - CHECK(isContiguous()); - CHECK(src.isContiguous()); - CHECK(elementCnt_ == src.getElementCnt()); - if (typeid(src) == typeid(GpuMatrix)) { - hl_memcpy_async(this->getData(), - const_cast(src.getData()), - sizeof(real) * elementCnt_, - stream); - // There is a need to add synchronization to ensure that the data is copied. - hl_stream_synchronize(stream); - } else if (typeid(src) == typeid(CpuMatrix)) { - memcpy(data_, src.getData(), sizeof(real) * elementCnt_); - } else { - LOG(FATAL) << "Wrong"; - } -} - -void CpuMatrix::copyFrom(const real* cpuSrc, size_t size) { - CHECK(isContiguous()); - CHECK(size <= elementCnt_); - memcpy(data_, cpuSrc, sizeof(real) * size); -} - -void CpuMatrix::copyFrom(const real* cpuSrc, const int64_t* seq) { - CHECK(isContiguous()); - for (size_t i = 0; i < height_; i++) { - memcpy(data_ + i * width_, cpuSrc + seq[i] * width_, sizeof(real) * width_); - } -} - -void CpuMatrix::copyFrom(const IVector& src) { - CHECK(isContiguous()); - CHECK(elementCnt_ == src.getSize()) - << "the src and dst should have same size."; - const int* cpuSrc = NULL; - IVectorPtr tmp; - if (src.useGpu()) { - CpuIVector tmp(src.getSize()); - tmp.copyFrom(src); - cpuSrc = tmp.getData(); - } else { - cpuSrc = src.getData(); - } - for (size_t i = 0; i < elementCnt_; ++i) { - data_[i] = cpuSrc[i]; - } -} - -void CpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) { - size_t height = getHeight(); - size_t width = getWidth(); - CHECK_EQ(b.getWidth(), width); - const int* index = rowIndex.getData(); - for (size_t i = 0; i < height; i++) { - CHECK_LT(static_cast(index[i]), b.getHeight()); - real* src = b.getData() + index[i] * width; - real* dst = getData() + i * width; - memcpy(dst, src, sizeof(real) * width); - } -} - -MatrixPtr CpuMatrix::clone(size_t height, size_t width, bool useGpu) { - CHECK(isContiguous()); - - if (height == 0 && width == 0) { - height = height_; - width = width_; - } - - CHECK(width && height); - - if (useGpu) { - return std::make_shared(height, width); - } else { - return std::make_shared(height, width); - } -} - -void CpuMatrix::resize(size_t newHeight, size_t newWidth) { - size_t newSize = newHeight * newWidth; - if (NULL == memoryHandle_.get() || - newSize * sizeof(real) > memoryHandle_->getAllocSize()) { - memoryHandle_ = std::make_shared(newSize * sizeof(real)); - data_ = reinterpret_cast(memoryHandle_->getBuf()); - } - - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newSize; - stride_ = width_; -} - -real CpuMatrix::getElement(size_t x, size_t y) const { - return data_[x * stride_ + y]; -} - -real CpuMatrix::getSum() { - CHECK(isContiguous()); - double sum = 0; - for (size_t i = 0; i < height_; ++i) { - for (size_t j = 0; j < width_; ++j) { - sum += data_[i * width_ + j]; - } - } - return sum; -} - -void CpuMatrix::accumulateColSum(Matrix& src) { - CHECK_EQ(getWidth(), src.getWidth()); - CHECK_EQ(getHeight(), (size_t)1); - - sumCols(src, /* scaleSum= */ 1, /* scaleDest= */ 1); -} - -real CpuMatrix::getAbsSum() { - CHECK(isContiguous()); - double sum = 0; - for (size_t i = 0; i < height_; ++i) { - for (size_t j = 0; j < width_; ++j) { - sum += fabs(data_[i * width_ + j]); - } - } - return sum; -} - -MatrixPtr CpuMatrix::getTranspose() { - if (memoryHandle_.get() != NULL) { - return std::make_shared( - std::dynamic_pointer_cast(memoryHandle_), - height_, - width_, - true); - } else { - MatrixPtr copy_T(new CpuMatrix(data_, height_, width_, true)); - return copy_T; - } -} - -void CpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { - if (memAlloc) { - matTrans = std::make_shared(width_, height_); - } else { - CHECK(matTrans != NULL); - CHECK_EQ(matTrans->getHeight(), width_); - CHECK_EQ(matTrans->getWidth(), height_); - } - real* dataTrans = matTrans->getData(); - real* data = getData(); - int lda = getStride(); - int ldc = matTrans->getStride(); - - for (size_t i = 0; i < height_; i++) { - for (size_t j = 0; j < width_; j++) { - dataTrans[j * ldc + i] = data[i * lda + j]; - } - } -} - -void CpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) { - if (memAlloc) { - matRot = std::make_shared(width_, height_); - } else { - CHECK(matRot != NULL); - CHECK_EQ(matRot->getHeight(), width_); - CHECK_EQ(matRot->getWidth(), height_); - } - real* dataRot = matRot->getData(); - real* data = getData(); - - for (size_t i = 0; i < height_; i++) { - for (size_t j = 0; j < width_; j++) { - if (clockWise) { - dataRot[j * height_ + i] = data[(height_ - i - 1) * width_ + j]; - } else { - dataRot[j * height_ + i] = data[i * width_ + (width_ - j - 1)]; - } - } - } -} - -MatrixPtr CpuMatrix::getInverse() { - MatrixPtr matInv; - inverse(matInv, true); - return matInv; -} - -void CpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) { - CHECK_EQ(height_, width_); - - if (memAlloc) { - matInv = std::make_shared(height_, width_); - } else { - CHECK(matInv != NULL); - } - - CHECK_EQ(height_, matInv->getHeight()); - CHECK_EQ(width_, matInv->getWidth()); - matInv->copyFrom(*this); - - real* data = getData(); - real* dataInv = matInv->getData(); - int ldc = matInv->getStride(); - - if (height_ == 1) { - CHECK_NE(*data, 0); - *dataInv = 1.0 / (*data); - return; - } - - /* Compute the LU decomposition of the matrix */ - std::vector ipiv(height_); - CBLAS_ORDER order = (matInv->isTransposed() ? CblasColMajor : CblasRowMajor); - int info = getrf(order, height_, height_, dataInv, ldc, ipiv.data()); - CHECK_EQ(info, 0); - - /* Compute the inverse of the matrix given its LU decompsotion */ - info = getri(order, height_, dataInv, ldc, ipiv.data()); - CHECK_EQ(info, 0); -} - -void CpuMatrix::upsampleForward(Matrix& input, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW) { - real* inputData = input.getData(); - real* maskData = mask.getData(); - real* outData = data_; - size_t inLength = imgSizeH * imgSizeW; - size_t outLength = outputH * outputW; - size_t batch = input.getHeight(); - CHECK(inLength == input.getWidth() / channels); - CHECK_EQ(batch, this->getHeight()); - CHECK_EQ(channels * outLength, this->getWidth()); - - for (size_t k = 0; k < batch; k++) { - for (size_t c = 0; c < channels; c++) { - for (size_t i = 0; i < inLength; i++) { - size_t out_index = static_cast(maskData[i]); - if (out_index >= outLength) { - LOG(FATAL) << "upsample index " << out_index << " out of range."; - } - outData[out_index] = inputData[i]; - } - inputData += inLength; - maskData += inLength; - outData += outLength; - } - } -} - -void CpuMatrix::upsampleBackward(Matrix& outputGrad, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW) { - real* outputGradData = outputGrad.getData(); - real* maskData = mask.getData(); - real* inputGradData = data_; - size_t inLength = imgSizeH * imgSizeW; - size_t outLength = outputH * outputW; - size_t batch = outputGrad.getHeight(); - CHECK(inLength == this->getWidth() / channels); - CHECK_EQ(batch, this->getHeight()); - CHECK_EQ(channels * outLength, outputGrad.getWidth()); - - for (size_t k = 0; k < batch; k++) { - for (size_t c = 0; c < channels; c++) { - for (size_t i = 0; i < inLength; i++) { - size_t out_index = static_cast(maskData[i]); - if (out_index >= outLength) { - LOG(FATAL) << "upsample index " << out_index << " out of range."; - } - inputGradData[i] = outputGradData[out_index]; - } - inputGradData += inLength; - maskData += inLength; - outputGradData += outLength; - } - } -} - -void CpuMatrix::maxPoolForward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - MatrixPtr maskMatP) { - real* inputData = inputMat.getData(); - real* outData = data_; - real* maskData = NULL; - size_t num = inputMat.getHeight(); - size_t inLength = imgSizeH * imgSizeW; - size_t outLength = outputH * outputW; - CHECK(inLength == inputMat.getWidth() / channels); - CHECK_EQ(num, this->getHeight()); - CHECK_EQ(channels * outLength, this->getWidth()); - size_t outStride = getStride(); - - if (maskMatP != NULL) { - maskData = maskMatP->getData(); - CHECK_EQ(channels * outLength, maskMatP->getWidth()); - } - - /* pool max one by one */ - for (size_t n = 0; n < num; ++n) { // frame by frame - if (!isContiguous()) { - outData = data_ + n * outStride; - } - for (size_t c = 0; c < channels; ++c) { // channel by channel - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = hstart + sizeY; - hstart = hstart < 0 ? 0 : hstart; - hend = hend < (int)imgSizeH ? hend : (int)imgSizeH; - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = wstart + sizeX; - wstart = wstart < 0 ? 0 : wstart; - wend = wend < (int)imgSizeW ? wend : (int)imgSizeW; - - real maxval = -(real)FLT_MAX; - int max_index = -1; - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - if (maxval < inputData[h * imgSizeW + w]) { - maxval = inputData[h * imgSizeW + w]; - max_index = h * imgSizeW + w; - } - } - } - - outData[ph * outputW + pw] = maxval; - if (maskData != NULL) maskData[ph * outputW + pw] = max_index; - } - } - // compute offset - inputData += inLength; - outData += outLength; - - if (maskData != NULL) maskData += outLength; - } - } -} - -void CpuMatrix::maxPoolBackward(Matrix& image, - size_t imgSizeH, - size_t imgSizeW, - Matrix& outGrad, - Matrix& outV, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW) { - size_t num = image.getHeight(); - size_t inLength = imgSizeH * imgSizeW; - size_t outLength = outputH * outputW; - size_t channels = size_t(width_ / inLength); - CHECK(image.getWidth() == inLength * channels); - CHECK(image.getHeight() == height_ && image.getWidth() == width_); - CHECK(outV.getHeight() == outGrad.getHeight() && - outV.getWidth() == outGrad.getWidth()); - - real* tgtGrad = data_; - real* inData = image.getData(); - real* otData = outV.getData(); - real* otGrad = outGrad.getData(); - - size_t outStride = outV.getStride(); - real* origOutData = otData; - real* origOutGrad = otGrad; - - for (size_t n = 0; n < num; ++n) { - if (!outV.isContiguous()) { - otData = origOutData + n * outStride; - otGrad = origOutGrad + n * outStride; - } - for (size_t c = 0; c < channels; ++c) { - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = std::min(hstart + sizeY, imgSizeH); - hstart = std::max(hstart, 0); - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = std::min(wstart + sizeX, imgSizeW); - wstart = std::max(wstart, 0); - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - tgtGrad[h * imgSizeW + w] = - scaleTargets * tgtGrad[h * imgSizeW + w] + - scaleOutput * otGrad[ph * outputW + pw] * - (inData[h * imgSizeW + w] == otData[ph * outputW + pw]); - } - } - } - } - // offset - inData += inLength; - tgtGrad += inLength; - otData += outLength; - otGrad += outLength; - } - } -} - -void CpuMatrix::avgPoolForward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - bool excludeMode) { - // The main loop - size_t num = input.getHeight(); - size_t inLength = imgSizeH * imgSizeW; - size_t outLength = outputH * outputW; - CHECK(inLength * channels == input.getWidth()); - CHECK(outLength * channels * num == height_ * width_); - real* tgtData = data_; - real* inData = input.getData(); - - for (size_t n = 0; n < num; ++n) { - if (!isContiguous()) { - tgtData = data_ + n * getStride(); - } - for (size_t c = 0; c < channels; ++c) { - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = std::min(hstart + sizeY, imgSizeH); - hstart = std::max(hstart, 0); - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = std::min(wstart + sizeX, imgSizeW); - wstart = std::max(wstart, 0); - tgtData[ph * outputW + pw] = 0; // clear - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - tgtData[ph * outputW + pw] += inData[h * imgSizeW + w]; - } - } - int poolSize = - excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; - CHECK(poolSize); - tgtData[ph * outputW + pw] /= poolSize; - } - } - // compute offset - inData += inLength; - tgtData += outLength; - } - } -} - -void CpuMatrix::avgPoolBackward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW, - bool excludeMode) { - size_t num = input.getHeight(); - size_t channels = input.getWidth() / outputH / outputW; - size_t inLength = imgSizeH * imgSizeW; - size_t outLength = outputH * outputW; - CHECK(inLength * channels == getWidth()); - real* inData = input.getData(); - real* outData = getData(); - - for (size_t n = 0; n < num; ++n) { - if (!input.isContiguous()) { - inData = input.getData() + n * input.getStride(); - } - for (size_t c = 0; c < channels; ++c) { - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = std::min(hstart + sizeY, imgSizeH); - hstart = std::max(hstart, 0); - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = std::min(wstart + sizeX, imgSizeW); - wstart = std::max(wstart, 0); - int poolSize = - excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; - CHECK(poolSize); - - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - outData[h * imgSizeW + w] += inData[ph * outputW + pw] / poolSize; - } - } - } - } - // offset - outData += inLength; - inData += outLength; - } - } -} - -void CpuMatrix::maxPool3DForward(Matrix& inputMat, - Matrix& maxPoolIdx, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW) { - real* inputData = inputMat.getData(); - real* outData = getData(); - real* maxPoolIdxData = maxPoolIdx.getData(); - size_t num = inputMat.getHeight(); - size_t inLength = imgSizeH * imgSizeW * imgSizeD; - size_t outLength = outputH * outputW * outputD; - CHECK(inLength == inputMat.getWidth() / channels); - CHECK_EQ(num, this->getHeight()); - CHECK_EQ(channels * outLength, this->getWidth()); - size_t outStride = getStride(); - - /* initialize the data_ */ - for (size_t i = 0; i < height_; i++) { - for (size_t j = 0; j < width_; j++) { - outData[(i)*outStride + j] = -(real)FLT_MAX; - maxPoolIdxData[(i)*outStride + j] = -1; - } - } - - /* pool max one by one */ - for (size_t n = 0; n < num; ++n) { // frame by frame - if (!isContiguous()) { - outData = getData() + n * outStride; - maxPoolIdxData = maxPoolIdx.getData() + n * outStride; - } - for (size_t c = 0; c < channels; ++c) { // channel by channel - for (size_t pd = 0; pd < outputD; ++pd) { - int dstart = pd * strideD - paddingD; - int dend = std::min(dstart + sizeZ, imgSizeD); - dstart = std::max(dstart, 0); - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = std::min(hstart + sizeY, imgSizeH); - hstart = std::max(hstart, 0); - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = std::min(wstart + sizeX, imgSizeW); - wstart = std::max(wstart, 0); - int maxIdx = -1; - real maxOutData = outData[(pd * outputH + ph) * outputW + pw]; - for (int d = dstart; d < dend; ++d) { - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - if (maxOutData < - inputData[(d * imgSizeH + h) * imgSizeW + w]) { - maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w]; - maxIdx = (d * imgSizeH + h) * imgSizeW + w; - } - } - } - } - outData[(pd * outputH + ph) * outputW + pw] = maxOutData; - maxPoolIdxData[(pd * outputH + ph) * outputW + pw] = maxIdx; - } - } - } - // compute offset - inputData += inLength; - outData += outLength; - maxPoolIdxData += outLength; - } - } -} - -void CpuMatrix::maxPool3DBackward(Matrix& outGrad, - Matrix& maxPoolIdx, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput) { - size_t num = getHeight(); - size_t inLength = imgSizeH * imgSizeW * imgSizeD; - size_t outLength = outputH * outputW * outputD; - size_t channels = size_t(width_ / inLength); - CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() && - maxPoolIdx.getWidth() == outGrad.getWidth()); - - real* tgtGrad = getData(); - real* otGrad = outGrad.getData(); - real* maxPoolIdxData = maxPoolIdx.getData(); - size_t outStride = outGrad.getStride(); - - for (size_t n = 0; n < num; ++n) { - if (!outGrad.isContiguous()) { - otGrad = outGrad.getData() + n * outStride; - maxPoolIdxData = maxPoolIdx.getData() + n * outStride; - } - for (size_t c = 0; c < channels; ++c) { - for (size_t pd = 0; pd < outputD; ++pd) { - for (size_t ph = 0; ph < outputH; ++ph) { - for (size_t pw = 0; pw < outputW; ++pw) { - const size_t index = (pd * outputH + ph) * outputW + pw; - const size_t tgtIdx = static_cast(maxPoolIdxData[index]); - tgtGrad[tgtIdx] = - scaleTargets * tgtGrad[tgtIdx] + scaleOutput * otGrad[index]; - } - } - } - // offset - tgtGrad += inLength; - otGrad += outLength; - maxPoolIdxData += outLength; - } - } -} - -void CpuMatrix::avgPool3DForward(Matrix& input, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW) { - // The main loop - size_t num = input.getHeight(); - size_t inLength = imgSizeH * imgSizeW * imgSizeD; - size_t outLength = outputH * outputW * outputD; - CHECK(inLength * channels == input.getWidth()); - CHECK(outLength * channels * num == height_ * width_); - real* tgtData = getData(); - real* inData = input.getData(); - - for (size_t n = 0; n < num; ++n) { - if (!isContiguous()) { - tgtData = data_ + n * getStride(); - } - for (size_t c = 0; c < channels; ++c) { - for (size_t pd = 0; pd < outputD; ++pd) { - int dstart = pd * strideD - paddingD; - int dend = std::min(dstart + sizeZ, imgSizeD); - dstart = std::max(dstart, 0); - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = std::min(hstart + sizeY, imgSizeH); - hstart = std::max(hstart, 0); - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = std::min(wstart + sizeX, imgSizeW); - wstart = std::max(wstart, 0); - - tgtData[(pd * outputH + ph) * outputW + pw] = 0; // clear - for (int d = dstart; d < dend; ++d) { - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - tgtData[(pd * outputH + ph) * outputW + pw] += - inData[(d * imgSizeH + h) * imgSizeW + w]; - } - } - } - int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); - CHECK(poolSize); - tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize; - } - } - } - // compute offset - inData += inLength; - tgtData += outLength; - } - } -} - -void CpuMatrix::avgPool3DBackward(Matrix& input, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput) { - size_t num = input.getHeight(); - size_t inLength = imgSizeH * imgSizeW * imgSizeD; - size_t outLength = outputH * outputW * outputD; - size_t channels = input.getWidth() / outLength; - CHECK(inLength * channels == getWidth()); - real* inData = input.getData(); - real* outData = getData(); - - for (size_t n = 0; n < num; ++n) { - if (!input.isContiguous()) { - inData = input.getData() + n * input.getStride(); - } - for (size_t c = 0; c < channels; ++c) { - for (size_t pd = 0; pd < outputD; ++pd) { - int dstart = pd * strideD - paddingD; - int dend = std::min(dstart + sizeZ, imgSizeD); - dstart = std::max(dstart, 0); - for (size_t ph = 0; ph < outputH; ++ph) { - int hstart = ph * strideH - paddingH; - int hend = std::min(hstart + sizeY, imgSizeH); - hstart = std::max(hstart, 0); - for (size_t pw = 0; pw < outputW; ++pw) { - int wstart = pw * strideW - paddingW; - int wend = std::min(wstart + sizeX, imgSizeW); - wstart = std::max(wstart, 0); - int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); - CHECK(poolSize); - for (int d = dstart; d < dend; ++d) { - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - outData[(d * imgSizeH + h) * imgSizeW + w] += - inData[(pd * outputH + ph) * outputW + pw] / poolSize; - } - } - } - } - } - } - // offset - outData += inLength; - inData += outLength; - } - } -} - -/** - * Input: one or more sequences. Each sequence contains some instances. - * Output: output size is the number of input sequences (NOT input instances). - * output[i] is set to max_{for each instance in this sequence}{input[i]} - */ -void CpuMatrix::maxSequenceForward(Matrix& input, - const IVector& sequence, - IVector& index) { - CHECK(dynamic_cast(&input)); - CHECK(dynamic_cast(&sequence)); - CHECK(dynamic_cast(&index)); - - real* outData = getData(); - real* inputData = input.getData(); - const int* starts = sequence.getData(); - int* maxIndex = index.getData(); - size_t numSequences = getHeight(); - size_t dim = getWidth(); - - CHECK_EQ(dim, input.getWidth()); - CHECK_EQ(numSequences, sequence.getSize() - 1); - CHECK_EQ(starts[numSequences], (int)input.getHeight()); - CHECK_EQ(numSequences * dim, index.getSize()); - - for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) { - // current sequence, loop for each input instance - // (1) first instance: do not need compare, copy value to outV directly - for (size_t k = 0; k < dim; ++k) { - outData[sequenceId * dim + k] = inputData[starts[sequenceId] * dim + k]; - maxIndex[sequenceId * dim + k] = starts[sequenceId]; - } - // (2) other instance in same sequence - for (int insId = starts[sequenceId] + 1; insId < starts[sequenceId + 1]; - ++insId) { - // insId is the index on all instances - for (size_t k = 0; k < dim; ++k) { - // for each dim - if (inputData[insId * dim + k] > outData[sequenceId * dim + k]) { - // update max value and record index - outData[sequenceId * dim + k] = inputData[insId * dim + k]; - maxIndex[sequenceId * dim + k] = insId; - } - } - } - } -} - -void CpuMatrix::maxSequenceBackward(Matrix& outputGrad, - const IVector& sequence, - IVector& index) { - CHECK(dynamic_cast(&outputGrad)); - CHECK(dynamic_cast(&sequence)); - CHECK(dynamic_cast(&index)); - - real* inputGrad = getData(); - real* outGrad = outputGrad.getData(); - int* maxIndex = index.getData(); - size_t dim = getWidth(); - size_t numSequences = sequence.getSize() - 1; - - CHECK_EQ(dim, outputGrad.getWidth()); - CHECK_EQ(numSequences, outputGrad.getHeight()); - CHECK_EQ(numSequences * dim, index.getSize()); - - for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) { - // current sequence - for (size_t j = 0; j < dim; ++j) { - // each dim - int insId = maxIndex[sequenceId * dim + j]; - inputGrad[insId * dim + j] += outGrad[sequenceId * dim + j]; - } - } -} - -inline void vecAddTo(real* a, const real* b, size_t len) { - for (unsigned int i = 0; i < len; ++i) { - a[i] += b[i]; - } -} - -inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { - for (unsigned int i = 0; i < len; ++i) { - a[i] += scaleB * b[i]; - } -} - -inline void colVecAddTo( - real* a, const real* b, size_t len, size_t aWidth, size_t bWidth) { - for (unsigned int i = 0; i < len; ++i) { - a[i * aWidth] += b[i * bWidth]; - } -} - -inline void colVecAddTo( - real* a, real* b, real c, size_t len, size_t aWidth, size_t bWidth) { - for (unsigned int i = 0; i < len; ++i) { - a[i * aWidth] += b[i * bWidth] * c; - } -} - -void CpuMatrix::addBias(Matrix& b, real scale) { - CHECK(b.useGpu_ == false) << "Matrix type are not equal"; - - CHECK_EQ(b.getHeight(), (size_t)1); - CHECK_EQ(width_, b.getWidth()); - real* aData = getData(); - real* bData = b.getData(); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - - if (scale == 1 && getStride() % 32 == 0) { // use libaddto - // @TODO(yuyang18) Make input addr can be unaligned. - // So merge this if and else - CHECK_EQ((size_t)aData % 32, 0UL); - CHECK_EQ((size_t)bData % 32, 0UL); - for (size_t i = 0; i < numSamples; i++) { - simd::addTo(aData + i * getStride(), bData, dim); - } - } else { - for (size_t i = 0; i < numSamples; i++) { - for (size_t j = 0; j < dim; j++) { - aData[i * getStride() + j] += scale * bData[j]; - } - } - } -} - -void CpuMatrix::addSharedBias(Matrix& b, real scale) { - CHECK_EQ(b.getHeight(), (size_t)1); - real* aData = getData(); - real* bData = b.getData(); - size_t numSamples = getHeight(); - size_t channel = b.getWidth(); - CHECK_EQ(getWidth() % channel, 0UL); - size_t dim = getWidth() / channel; - - for (size_t i = 0; i < numSamples; i++) { - for (size_t c = 0; c < channel; c++) { - for (size_t j = 0; j < dim; j++) { - aData[i * getStride() + c * dim + j] += scale * bData[c]; - } - } - } -} - -void CpuMatrix::collectBias(Matrix& a, real scale) { - CHECK_EQ(getHeight(), (size_t)1); - CHECK_EQ(width_, a.getWidth()); - CpuSparseMatrix* aptr = dynamic_cast(&a); - if (!aptr) { - sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1); - } else { - size_t nnz = aptr->getElementCnt(); - int* cols = aptr->getCols(); - real* A = aptr->getValue(); - real* B = getData(); - for (size_t i = 0; i < nnz; i++) { - B[cols[i]] += scale * A[i]; - } - } -} - -void CpuMatrix::collectSharedBias(Matrix& a, real scale) { - CHECK_EQ(getHeight(), (size_t)1); - real* B = getData(); - real* A = a.getData(); - size_t numSamples = a.getHeight(); - size_t channel = getWidth(); - CHECK_EQ(a.getWidth() % channel, 0UL); - size_t dim = a.getWidth() / channel; - for (size_t i = 0; i < numSamples; i++) { - for (size_t c = 0; c < channel; c++) { - for (size_t j = 0; j < dim; j++) { - B[c] += scale * A[i * channel * dim + c * dim + j]; - } - } - } -} - -void CpuMatrix::sequenceAvgForward(Matrix& a, - const IVector& startsPos, - int mode) { - size_t height = getHeight(); - size_t width = getWidth(); - CHECK_EQ(height, startsPos.getSize() - 1); - CHECK_EQ(width, a.getWidth()); - real* dst = getData(); - real* src = a.getData(); - const int* starts = startsPos.getData(); - MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false); - MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false); - for (size_t i = 0; i < height; i++) { - int sequenceLength = starts[i + 1] - starts[i]; - if (0 == sequenceLength) { - // empty sequence - continue; - } - outMtx->setData(dst + i * width); - dataMtx->setData(src + starts[i] * width, sequenceLength, width); - if (mode == 0) { - // plain average - outMtx->sumCols(*dataMtx, - (real)1 / (real)sequenceLength, - /* scaleDest= */ 1); - } else if (mode == 1) { - // sum instead of average - outMtx->sumCols(*dataMtx, /* scaleSum= */ 1, /* scaleDest= */ 1); - } else if (mode == 2) { - // divide by square root of sequenceLength - outMtx->sumCols(*dataMtx, - (real)1 / std::sqrt(sequenceLength), - /* scaleDest= */ 1); - } else { - LOG(FATAL) << "should not reach here"; - } - } -} - -void CpuMatrix::sequenceAvgBackward(Matrix& a, - const IVector& startsPos, - int mode) { - size_t height = a.getHeight(); - size_t width = getWidth(); - CHECK_EQ(height, startsPos.getSize() - 1); - CHECK_EQ(width, a.getWidth()); - real* dst = getData(); - real* src = a.getData(); - const int* starts = startsPos.getData(); - MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false); - MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false); - for (size_t i = 0; i < height; ++i) { - int sequenceLength = starts[i + 1] - starts[i]; - if (0 == sequenceLength) { - // empty sequence - continue; - } - outMtx->setData(dst + starts[i] * width, sequenceLength, width); - dataMtx->setData(src + i * width); - if (mode == 0) { - // plain average - outMtx->addBias(*dataMtx, 1.0f / sequenceLength); - } else if (mode == 1) { - // sum instead of average - outMtx->addBias(*dataMtx, 1.0f); - } else if (mode == 2) { - // divide by square root of sequenceLength - outMtx->addBias(*dataMtx, 1.0f / std::sqrt(sequenceLength)); - } else { - LOG(FATAL) << "should not reach here"; - } - } -} - -/* this = scaleAB*(a*b) + scaleT*this*/ -void CpuMatrix::mul(const Matrix& a, - const Matrix& b, - real scaleAB, - real scaleT) { - CHECK(!isTransposed()) << "Not supported"; - const auto a_ptr = dynamic_cast(&a); - const auto b_ptr = dynamic_cast(&b); - const auto a_ptr_s = dynamic_cast(&a); - const auto b_ptr_s = dynamic_cast(&b); - - if (a_ptr && b_ptr) { - mul((CpuMatrix*)a_ptr, (CpuMatrix*)b_ptr, scaleAB, scaleT); - } else if (a_ptr_s && b_ptr) { - mul((CpuSparseMatrix*)a_ptr_s, (CpuMatrix*)b_ptr, scaleAB, scaleT); - } else if (a_ptr && b_ptr_s) { - mul((CpuMatrix*)a_ptr, (CpuSparseMatrix*)b_ptr_s, scaleAB, scaleT); - } else { - LOG(FATAL) << "Not supported"; - } -} - -void CpuMatrix::mul(CpuSparseMatrix* a, - CpuMatrix* b, - real scaleAB, - real scaleT) { - if (dynamic_cast(b)) { - return mul(a, dynamic_cast(b), this, scaleAB, scaleT); - } else if (dynamic_cast(b)) { - return mul(a, dynamic_cast(b), this, scaleAB, scaleT); - } else { - return mul(a, b, this, scaleAB, scaleT); - } -} - -void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { - CHECK(!isTransposed()) << "Not supported"; - - size_t a_col, b_col, a_row, b_row; - bool a_trans, b_trans; - if (!a->isTransposed()) { - a_col = a->getWidth(); - a_row = a->getHeight(); - a_trans = false; - } else { - a_col = a->getHeight(); - a_row = a->getWidth(); - a_trans = true; - } - if (!b->isTransposed()) { - b_col = b->getWidth(); - b_row = b->getHeight(); - b_trans = false; - } else { - b_col = b->getHeight(); - b_row = b->getWidth(); - b_trans = true; - } - - CHECK_EQ(a_col, b_row); - CHECK_EQ(a_row, getHeight()); - CHECK_EQ(b_col, getWidth()); - - real* A = a->getData(); - real* B = b->getData(); - real* C = getData(); - - int M = getHeight(); - int N = getWidth(); - int K = a_col; - int lda = a->getStride(); - int ldb = b->getStride(); - int ldc = getStride(); - BlasGemm::compute( - a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc); -} - -void CpuMatrix::mul( - CpuMatrix* a, CpuMatrix* b, CpuSparseMatrix* c, real scaleAB, real scaleT) { - CHECK(!c->isTransposed()) << "Not supported"; - CHECK_EQ(c->getValueType(), FLOAT_VALUE); - - real* A = a->getData(); - real* B = b->getData(); - real* C = c->getValue(); - int* rows = c->getRows(); - int* cols = c->getCols(); - size_t height = c->getHeight(); - size_t width = c->getWidth(); - if (scaleT == 0) { - c->zeroMem(); - } - - if (!a->isTransposed() && !b->isTransposed()) { - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), m); - CHECK_EQ(a->getHeight(), height); - CHECK_EQ(b->getWidth(), width); - if (c->getFormat() == SPARSE_CSC) { - for (size_t i = 0; i < width; i++) { - size_t start = c->getColStartIdx(i); - size_t end = c->getColStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - real sum = 0; - size_t rowIdx = rows[j]; - for (size_t k = 0; k < m; k++) { - sum += A[rowIdx * m + k] * B[k * width + i]; - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - } else { - for (size_t i = 0; i < height; i++) { - size_t start = c->getRowStartIdx(i); - size_t end = c->getRowStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - real sum = 0; - size_t colIdx = cols[j]; - for (size_t k = 0; k < m; k++) { - sum += A[i * m + k] * B[k * width + colIdx]; - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - } - } else if (a->isTransposed() && !b->isTransposed()) { - size_t m = a->getHeight(); - CHECK_EQ(m, b->getHeight()); - CHECK_EQ(b->getWidth(), width); - CHECK_EQ(a->getWidth(), height); - - if (c->getFormat() == SPARSE_CSC) { - for (size_t i = 0; i < width; i++) { - size_t start = c->getColStartIdx(i); - size_t end = c->getColStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - real sum = 0; - size_t rowIdx = rows[j]; - for (size_t k = 0; k < m; k++) { - sum += A[k * height + rowIdx] * B[k * width + i]; - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - } else { - for (size_t i = 0; i < height; i++) { - int start = c->getRowStartIdx(i); - int end = c->getRowStartIdx(i + 1); - for (int j = start; j < end; j++) { - real sum = 0; - size_t colIdx = cols[j]; - for (size_t k = 0; k < m; k++) { - sum += A[k * height + i] * B[k * width + colIdx]; - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - } - } else if (!a->isTransposed() && b->isTransposed()) { - size_t m = a->getWidth(); - CHECK_EQ(b->getWidth(), m); - CHECK_EQ(a->getHeight(), height); - CHECK_EQ(b->getHeight(), width); - if (c->getFormat() == SPARSE_CSR) { - for (size_t i = 0; i < height; i++) { - size_t start = c->getRowStartIdx(i); - size_t end = c->getRowStartIdx(i + 1); - for (size_t j = start; j < end; j++) { - real sum = 0; - size_t colIdx = cols[j]; - for (size_t k = 0; k < m; k++) { - sum += A[i * m + k] * B[colIdx * m + k]; - } - C[j] = scaleAB * sum + scaleT * C[j]; - } - } - } else { - LOG(FATAL) << "Not supported csc format " - "when a is not trans and b is trans"; - } - } else { - LOG(FATAL) << "Not supported"; - } -} - -void CpuMatrix::mul(CpuMatrix* a, - CpuSparseMatrix* b, - real scaleAB, - real scaleT) { - CHECK(!trans_) << "Not supported"; - CHECK(!a->isTransposed()) << "Not supported"; - CHECK(scaleT == 0 || scaleT == 1); - - // TODO(yuyang18): Maybe bug implementation here - CHECK_EQ(scaleAB, static_cast(1.0)); - - real* A = a->getData(); - real* B = b->getValue(); - real* C = getData(); - int* rows = b->getRows(); - int* cols = b->getCols(); - - if (scaleT == 0) { - zeroMem(); - } - if (b->getFormat() == SPARSE_CSC) { - if (!b->isTransposed()) { - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), m); - CHECK_EQ(a->getHeight(), height_); - CHECK_EQ(b->getWidth(), width_); - - if (b->getValueType() == NO_VALUE) { - for (size_t j = 0; j < b->getWidth(); ++j) { - int start = b->getColStartIdx(j); - int end = b->getColStartIdx(j + 1); - for (int i = start; i < end; ++i) { - colVecAddTo(C + j, A + rows[i], height_, width_, a->getWidth()); - } - } - } else if (b->getValueType() == FLOAT_VALUE) { - for (size_t j = 0; j < b->getWidth(); ++j) { - int start = b->getColStartIdx(j); - int end = b->getColStartIdx(j + 1); - for (int i = start; i < end; ++i) { - colVecAddTo( - C + j, A + rows[i], B[i], height_, width_, a->getWidth()); - } - } - } - } else /*if (b->isTransposed())*/ { - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), width_); - CHECK_EQ(a->getHeight(), height_); - CHECK_EQ(b->getWidth(), m); - if (b->getValueType() == NO_VALUE) { - for (size_t i = 0; i < b->getWidth(); ++i) { - int start = b->getColStartIdx(i); - int end = b->getColStartIdx(i + 1); - for (int j = start; j < end; ++j) { - colVecAddTo(C + rows[j], A + i, height_, width_, a->getWidth()); - } - } - } else if (b->getValueType() == FLOAT_VALUE) { - for (size_t i = 0; i < b->getWidth(); ++i) { - int start = b->getColStartIdx(i); - int end = b->getColStartIdx(i + 1); - for (int j = start; j < end; ++j) { - colVecAddTo( - C + rows[j], A + i, B[j], height_, width_, a->getWidth()); - } - } - } - } - } else { - if (!b->isTransposed()) { - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), m); - CHECK_EQ(a->getHeight(), height_); - CHECK_EQ(b->getWidth(), width_); - - if (b->getValueType() == NO_VALUE) { - for (size_t j = 0; j < b->getHeight(); ++j) { - int start = b->getRowStartIdx(j); - int end = b->getRowStartIdx(j + 1); - for (int i = start; i < end; ++i) { - colVecAddTo(C + cols[i], A + j, height_, width_, a->getWidth()); - } - } - } else if (b->getValueType() == FLOAT_VALUE) { - for (size_t j = 0; j < b->getHeight(); ++j) { - int start = b->getRowStartIdx(j); - int end = b->getRowStartIdx(j + 1); - for (int i = start; i < end; ++i) { - colVecAddTo( - C + cols[i], A + j, B[i], height_, width_, a->getWidth()); - } - } - } - } else /*if (b->isTransposed())*/ { - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), width_); - CHECK_EQ(a->getHeight(), height_); - CHECK_EQ(b->getWidth(), m); - if (b->getValueType() == NO_VALUE) { - for (size_t i = 0; i < b->getHeight(); ++i) { - int start = b->getRowStartIdx(i); - int end = b->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - colVecAddTo(C + i, A + cols[j], height_, width_, a->getWidth()); - } - } - } else if (b->getValueType() == FLOAT_VALUE) { - for (size_t i = 0; i < b->getHeight(); ++i) { - int start = b->getRowStartIdx(i); - int end = b->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - colVecAddTo( - C + i, A + cols[j], B[j], height_, width_, a->getWidth()); - } - } - } - } - } -} - -void CpuMatrix::selectRows(Matrix& table, IVector& ids) { - if (dynamic_cast(&table)) { - selectRowsImp(*dynamic_cast(&table), ids); - } else if (dynamic_cast(&table)) { - selectRowsImp(*dynamic_cast(&table), ids); - } else { - CHECK(table.isContiguous()); - selectRowsImp(*dynamic_cast(&table), ids); - } -} - -void CpuMatrix::selectElements(Matrix& table, IVector& ids) { - CHECK_EQ(table.getHeight(), ids.getSize()); - CHECK_EQ(getHeight(), ids.getSize()); - CHECK_EQ(getWidth(), 1U); - real* tableData = table.getData(); - int* idsData = ids.getData(); - for (size_t i = 0; i < table.getHeight(); i++) { - data_[i] += tableData[i * table.getWidth() + idsData[i]]; - } -} - -void CpuMatrix::addElements(Matrix& table, IVector& ids) { - CHECK_EQ(table.getHeight(), ids.getSize()); - CHECK_EQ(getHeight(), ids.getSize()); - CHECK_EQ(getWidth(), 1U); - real* tableData = table.getData(); - int* idsData = ids.getData(); - for (size_t i = 0; i < table.getHeight(); i++) { - tableData[i * table.getWidth() + idsData[i]] += data_[i]; - } -} - -// this.row[i] += table.row[ids[i]] -template -void CpuMatrix::selectRowsImp(TableMatType& table, IVector& ids) { - CHECK(!table.useGpu()); - CHECK(!ids.useGpu()); - CHECK_EQ(getHeight(), ids.getSize()); - CHECK_EQ(getWidth(), table.getWidth()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - real* a = getData(); - size_t tableSize = table.getHeight(); - int* index = ids.getData(); - - for (size_t i = 0; i < numSamples; ++i) { - if (index[i] == -1) continue; - CHECK_LT(index[i], (int)tableSize); - CHECK_GE(index[i], 0); - vecAddTo(a + i * stride_, table.getRow(index[i]), dim); - } -} - -void CpuMatrix::addToRows(Matrix& table, IVector& ids) { - if (dynamic_cast(&table)) { - addToRowsImp(*dynamic_cast(&table), ids); - } else if (dynamic_cast(&table)) { - addToRowsImp(*dynamic_cast(&table), ids); - } else if (dynamic_cast(&table)) { - addToRowsImp(*dynamic_cast(&table), ids); - } else { - CHECK(table.isContiguous()); - addToRowsImp(*dynamic_cast(&table), ids); - } -} - -// table.row[ids[i]] += this.row[i] -template -void CpuMatrix::addToRowsImp(TableMatType& table, IVector& ids) { - CHECK(!table.useGpu()); - CHECK(!ids.useGpu()); - CHECK_EQ(getHeight(), ids.getSize()); - CHECK_EQ(getWidth(), table.getWidth()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - real* a = getData(); - size_t tableSize = table.getHeight(); - int* index = ids.getData(); - - for (size_t i = 0; i < numSamples; ++i) { - if (index[i] == -1) continue; - CHECK_LT(index[i], (int)tableSize); - CHECK_GE(index[i], 0); - vecAddTo(table.getRow(index[i]), a + i * stride_, dim); - } -} - -static ThreadLocal> threadLocalColArray; - -template -void CpuMatrix::mul( - CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB, real scaleT) { - CHECK(!c->isTransposed()) << "Not supported"; - CHECK(!b->isTransposed()) << "Not supported"; - // TODO(yuyang18): Maybe bug implementation here. - CHECK(scaleAB == 1) << "Not supported"; - CHECK(scaleT == 0 || scaleT == 1) << "Not supported"; - CHECK_EQ(a->getFormat(), SPARSE_CSR) << "Not supported"; - - real* B = b->getData(); - real* C = c->getData(); - size_t height = c->getHeight(); - size_t width = c->getWidth(); - int* cols = a->getCols(); - real* values = a->getValue(); - - if (scaleT == 0) { - c->zeroMem(); - } - - if (!a->isTransposed()) { - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), m); - CHECK_EQ(a->getHeight(), height); - CHECK_EQ(b->getWidth(), width); - - if (a->getValueType() == NO_VALUE) { - if (width % 32 == 0) { // use libaddto - // @TODO(yuyang18) Make input addr can be unaligned. - // So merge this if and else - CHECK_EQ((size_t)B % 32, 0UL); - CHECK_EQ((size_t)C % 32, 0UL); - auto& colArray = *threadLocalColArray; - for (size_t i = 0; i < a->getHeight(); ++i) { - const int start = a->getRowStartIdx(i); - const int end = a->getRowStartIdx(i + 1); - size_t colNum = end - start; - colArray.resize(colNum); - for (int j = 0; j < end - start; ++j) { - colArray[j] = b->getRow(cols[j + start]); - } - simd::batchAddTo(c->getRow(i), &colArray[0], colNum, width); - } - - } else { - for (size_t i = 0; i < a->getHeight(); ++i) { - const int start = a->getRowStartIdx(i); - const int end = a->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - vecAddTo(c->getRow(i), b->getRow(cols[j]), width); - } - } - } - } else if (a->getValueType() == FLOAT_VALUE) { - for (size_t i = 0; i < a->getHeight(); ++i) { - const int start = a->getRowStartIdx(i); - const int end = a->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - vecAddTo(c->getRow(i), b->getRow(cols[j]), values[j], width); - } - } - } - } else /*if (a->isTransposed())*/ { - size_t m = a->getHeight(); - CHECK_EQ(b->getHeight(), m); - CHECK_EQ(a->getWidth(), height); - CHECK_EQ(b->getWidth(), width); - if (a->getValueType() == NO_VALUE) { - if (width % 32 == 0) { // use libaddto - // @TODO(yuyang18) Make input addr can be unaligned. - // So merge this if and else - CHECK_EQ((size_t)B % 32, 0UL); - CHECK_EQ((size_t)C % 32, 0UL); - for (size_t i = 0; i < a->getHeight(); ++i) { - const int start = a->getRowStartIdx(i); - const int end = a->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - simd::addTo(c->getRow(cols[j]), b->getRow(i), width); - } - } - - } else { - for (size_t i = 0; i < a->getHeight(); ++i) { - const int start = a->getRowStartIdx(i); - const int end = a->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - vecAddTo(c->getRow(cols[j]), b->getRow(i), width); - } - } - } - } else if (a->getValueType() == FLOAT_VALUE) { - for (size_t i = 0; i < a->getHeight(); ++i) { - const int start = a->getRowStartIdx(i); - const int end = a->getRowStartIdx(i + 1); - for (int j = start; j < end; ++j) { - vecAddTo(c->getRow(cols[j]), b->getRow(i), values[j], width); - } - } - } - } -} - -// instantiation mul() called in SparseRowMatrix.cpp -template void CpuMatrix::mul( - CpuSparseMatrix* a, - CpuMatrix* b, - SparseRowCpuMatrix* c, - real scaleAB, - real scaleT); -template void CpuMatrix::mul( - CpuSparseMatrix* a, - CpuMatrix* b, - SparseAutoGrowRowCpuMatrix* c, - real scaleAB, - real scaleT); -template void CpuMatrix::mul(CpuSparseMatrix* a, - CpuMatrix* b, - CacheRowCpuMatrix* c, - real scaleAB, - real scaleT); - -#ifndef PADDLE_MOBILE_INFERENCE -void SharedCpuMatrix::mul(CpuSparseMatrix* a, - CpuMatrix* b, - real scaleAB, - real scaleT) { - CHECK(!isTransposed()) << "Not supported"; - CHECK(!b->isTransposed()) << "Not supported"; - CHECK_EQ(scaleAB, 1) << "Not supported"; - CHECK_EQ(scaleT, 1) << "Not supported"; - CHECK_EQ(a->getFormat(), SPARSE_CSR) << "not supported"; - - real* B = b->getData(); - real* C = getData(); - size_t height = getHeight(); - size_t width = getWidth(); - - // get real trans - MatrixPtr aTrans; - if (a->isTransposed()) { - aTrans = a->getTmpSparseMatrix(a->getWidth(), a->getHeight()); - a->transpose(aTrans, false); - } - a = dynamic_cast(aTrans.get()); - - size_t m = a->getWidth(); - CHECK_EQ(b->getHeight(), m); - CHECK_EQ(a->getHeight(), height); - CHECK_EQ(b->getWidth(), width); - - size_t blockSize = (height / blockNum_) + 1; - CpuMatrixPtr localBuf = *localBuf_; - if (!localBuf) { - localBuf = std::make_shared(blockSize, width); - } else { - localBuf->resize(blockSize, width); - } - localBuf->zeroMem(); - real* localC = localBuf->getData(); - std::vector& blockSeq = *blockSeq_; - if (blockSeq.size() == 0) { - for (int k = 0; k < blockNum_; ++k) { - blockSeq.push_back(k); - } - std::shuffle( - blockSeq.begin(), blockSeq.end(), ThreadLocalRandomEngine::get()); - } - std::vector& localBufRows = *localBufRows_; - int* cols = a->getCols(); - real* value = a->getValue(); - - for (int k = 0; k < blockNum_; ++k) { - int blockId = blockSeq[k]; - size_t blockBegin = blockId * blockSize; - size_t blockEnd = (blockId + 1) * blockSize; - if (blockId == blockNum_ - 1) { - blockEnd = height; - } - if (a->getValueType() == NO_VALUE) { - for (size_t i = blockBegin; i < blockEnd; ++i) { - int start = a->getRowStartIdx(i); - int end = a->getRowStartIdx(i); - size_t colNum = a->getColNum(i); - if (colNum == 0) { - continue; - } // skip empty row - localBufRows.push_back(i); - size_t bufPos = localBufRows.size() - 1; - for (int j = start; j < end; ++j) { - vecAddTo(localC + bufPos * width, B + cols[j] * width, width); - } - } - } else if (a->getValueType() == FLOAT_VALUE) { - for (size_t i = blockBegin; i < blockEnd; ++i) { - int start = a->getRowStartIdx(i); - int end = a->getRowStartIdx(i); - size_t colNum = a->getColNum(i); - if (colNum == 0) { - continue; - } // skip empty row - localBufRows.push_back(i); - size_t bufPos = localBufRows.size() - 1; - for (int j = start; j < end; ++j) { - vecAddTo( - localC + bufPos * width, B + cols[j] * width, value[j], width); - } - } - } - - { - std::lock_guard guard(*blockLocks_[blockId]); - for (size_t i = 0; i < localBufRows.size(); ++i) { - vecAddTo(C + localBufRows[i] * width, localC + i * width, width); - } - } - memset(localC, 0, localBufRows.size() * width * sizeof(real)); - localBufRows.clear(); - } - - VLOG(2) << " B[0]=" << B[0] << " B[1]=" << B[1] << " C[0]=" << C[0] - << " C[1]=" << C[1]; -} - -void SharedCpuMatrix::add(Matrix& b, real p1, real p2) { - CHECK_EQ(blockNum_, 1); - std::lock_guard guard(*blockLocks_[0]); - CpuMatrix::add(b, p1, p2); -} - -void SharedCpuMatrix::add(real p1, real p2) { - CHECK_EQ(blockNum_, 1); - std::lock_guard guard(*blockLocks_[0]); - CpuMatrix::add(p1, p2); -} - -void SharedCpuMatrix::initShared(int blockNum) { - CHECK_GT(height_ * width_, 1UL * 1024 * 1024) - << "should not share small matrix"; - initBlock(blockNum); -} - -void SharedCpuMatrix::initBlock(int blockNum) { - CHECK_LE(blockNum, 200) << "should not use large block number"; - blockNum_ = blockNum; - blockLocks_.resize(blockNum); - for (auto& locker : blockLocks_) { - locker.reset(new std::mutex); - } -} - -#endif -/* Add a (column) vector b to matrix a, column by column */ -void CpuMatrix::addColumnVector(const Matrix& b) { - BaseMatrix::addColVector(const_cast(b)); -} - -/* this = a*b */ -void CpuMatrix::mul(const Matrix& a, const Matrix& b) { - return mul(a, b, 1.0, 0.0); -} - -/* this = scaleAB*(this*b) + scaleT*this */ -void CpuMatrix::rightMul(Matrix& b, real scaleAB, real scaleT) { - (void)b; - (void)scaleAB; - (void)scaleT; - LOG(FATAL) << "Not implemented"; -} - -/* this = this* b */ -void CpuMatrix::rightMul(Matrix& b) { return rightMul(b, 1.0, 0.0); } - -/* this = scaleAB*(a*this) + scaleT*this */ -void CpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) { - (void)a; - (void)scaleAB; - (void)scaleT; - LOG(FATAL) << "Not implemented"; -} - -/* this = a*this) */ -void CpuMatrix::leftMul(Matrix& a) { return leftMul(a, 1.0, 0.0); } - -void CpuMatrix::colMerge(Matrix& src) { src.rowSum(*this); } - -void CpuMatrix::rowSum(Matrix& sum) { - CHECK_EQ(sum.getHeight(), getHeight()); - CHECK_EQ(sum.getWidth(), (size_t)1); - - sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0); -} - -void CpuMatrix::rowMaxId(IVector& maxIds) { - CHECK(!maxIds.useGpu()) << "Matrix type are not equal"; - - size_t numSamples = getHeight(); - CHECK_EQ(maxIds.getSize(), numSamples); - - real* a = getData(); - int* s = maxIds.getData(); - size_t dim = getWidth(); - - for (size_t i = 0; i < numSamples; i++) { - real sm = a[i * dim]; - int maxId = 0; - for (size_t j = 1; j < dim; j++) { - if (a[i * dim + j] > sm) { - maxId = j; - sm = a[i * dim + j]; - } - } - s[i] = maxId; - } -} - -void CpuMatrix::rowMax(Matrix& max) { - CHECK_EQ(max.getHeight(), getHeight()); - CHECK_EQ(max.getWidth(), (size_t)1); - max.maxRows(*this); -} - -/* Get the top k elements of each row of this matrix */ -void CpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { - CHECK(isContiguous()); - CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal"; - size_t numSamples = getHeight(); - size_t beam = maxVal.getWidth(); - CHECK_EQ(maxIds.getSize(), numSamples * beam); - CHECK_EQ(maxVal.getHeight(), numSamples); - CHECK_EQ(maxVal.getWidth(), beam); - - real* a = getData(); - int* s = maxIds.getData(); - real* t = maxVal.getData(); - size_t dim = getWidth(); - for (size_t i = 0; i < numSamples; i++) { - std::vector> vec; - for (size_t j = 0; j < dim; j++) { - vec.push_back(std::pair(a[i * dim + j], j)); - } - - std::partial_sort( - vec.begin(), - vec.begin() + beam, - vec.end(), - [](const std::pair& l, const std::pair& r) { - return l.first > r.first; - }); - for (size_t j = 0; j < beam; j++) { - t[i * beam + j] = vec[j].first; - s[i * beam + j] = vec[j].second; - } - } -} - -void CpuMatrix::colMax(Matrix& max) { - CHECK_EQ(max.getWidth(), getWidth()); - CHECK_EQ(max.getHeight(), (size_t)1); - max.maxCols(*this); -} - -void CpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) { - CHECK(isContiguous()); - CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal"; - size_t numSamples = getWidth(); - size_t beam = maxVal.getHeight(); - CHECK_EQ(maxIds.getSize(), numSamples * beam); - CHECK_EQ(maxVal.getWidth(), numSamples); - - real* a = getData(); - int* s = maxIds.getData(); - real* t = maxVal.getData(); - size_t dim = getHeight(); - for (size_t i = 0; i < numSamples; i++) { - std::vector> vec; - for (size_t j = 0; j < dim; j++) { - vec.push_back(std::pair(a[i + j * numSamples], j)); - } - - std::partial_sort( - vec.begin(), - vec.begin() + beam, - vec.end(), - [](const std::pair& l, const std::pair& r) { - return l.first > r.first; - }); - for (size_t j = 0; j < beam; j++) { - t[i + j * numSamples] = vec[j].first; - s[i + j * numSamples] = vec[j].second; - } - } -} - -void CpuMatrix::maxoutForward(Matrix& a, - IVector& id, - size_t channels, - size_t groups) { - CHECK(dynamic_cast(&a)); - CHECK(dynamic_cast(&id)); - CHECK_EQ(a.getHeight(), getHeight()); - - size_t size = getWidth(); - size_t batchSize = getHeight(); - size_t featLen = size / channels; - const real* input = a.getData(); - int* idForCpu = id.getData(); - - MatrixPtr maxInMat, maxOutMat; - Matrix::resizeOrCreate(maxInMat, groups, size, false, false); - Matrix::resizeOrCreate(maxOutMat, 1, size, false, false); - - for (size_t batch_idx = 0; batch_idx < batchSize; ++batch_idx) { - size_t newIndex = batch_idx * size; - IVectorPtr tmpId = IVector::create(idForCpu + newIndex, size, false); - - for (size_t i = 0; i < channels; ++i) { - size_t newFeatLen = i * featLen; - for (size_t j = 0; j < groups; ++j) { - maxInMat->subMatrix(j, j + 1, newFeatLen, newFeatLen + featLen) - ->copyFrom(input + (newIndex + newFeatLen) * groups + j * featLen, - featLen); - } - } - maxInMat->colMax(*tmpId, *maxOutMat); - this->subRowMatrix(batch_idx, batch_idx + 1)->copyFrom(*maxOutMat); - } -} - -void CpuMatrix::maxoutBackward(Matrix& a, - IVector& id, - size_t channels, - size_t groups) { - CHECK(dynamic_cast(&a)); - CHECK(dynamic_cast(&id)); - CHECK_EQ(a.getHeight(), getHeight()); - - size_t size = a.getWidth(); - size_t batchSize = getHeight(); - size_t featLen = size / channels; - size_t newFeatLen = groups * featLen; - real* inputG = getData(); - const real* outG = a.getData(); - int* idForCpu = id.getData(); - - for (size_t batch_idx = 0; batch_idx < batchSize; ++batch_idx) { - size_t newIndex = batch_idx * size; - int* idData = idForCpu + newIndex; - - for (size_t i = 0; i < size; ++i) { - int gradIdx = - idData[i] * featLen + (i / featLen) * newFeatLen + i % featLen; - (inputG + newIndex * groups)[gradIdx] += (outG + newIndex)[i]; - } - } -} - -void CpuMatrix::rowNormalizeL1(Matrix& out) { - CHECK(!out.useGpu()); - - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(out.getHeight(), numSamples); - CHECK_EQ(out.getWidth(), dim); - real* a = getData(); - real* b = out.getData(); - for (size_t i = 0; i < numSamples; ++i) { - real s = 0; - for (size_t j = 0; j < dim; ++j) { - s += a[i * dim + j]; - } - // Right now, we just bet that sum won't be zero. If this really happens, - // we will figure out what should be done then. - CHECK_GT(s, 0); - s = 1 / s; - for (size_t j = 0; j < dim; ++j) { - b[i * dim + j] = s * a[i * dim + j]; - } - } -} - -/* calulate classification error */ -void CpuMatrix::classificationError(Matrix& output, - IVector& label, - size_t topkSize) { - size_t numSamples = this->getHeight(); - auto cpuOutput = dynamic_cast(&output); - auto cpuLabel = dynamic_cast(&label); - IVectorPtr cpuTopIds = std::make_shared(numSamples * topkSize); - MatrixPtr cpuTopVal = std::make_shared(numSamples, topkSize); - - CHECK(cpuOutput && cpuLabel) << "Invalid argument pointer"; - CHECK(cpuTopIds && cpuTopVal) << "Allocate cpu memory failed"; - CHECK(cpuLabel->getSize() == numSamples) << "Vector size is not equal"; - CHECK(cpuOutput->getHeight() == numSamples && this->getWidth() == 1) - << "Matrix dimensions are not equal"; - - // top k matrix classification - cpuOutput->rowMax(*cpuTopIds, *cpuTopVal); - - size_t dim = cpuOutput->getWidth(); - real* result = this->getData(); - int* ids = cpuTopIds->getData(); - int* lbl = cpuLabel->getData(); - for (size_t i = 0; i < numSamples; ++i) { - CHECK_GE(lbl[i], 0); - CHECK_LT((size_t)lbl[i], dim); - - for (size_t j = 0; j < topkSize; ++j) { - if (ids[j + i * topkSize] == lbl[i]) { - result[i] = 0; - break; - } - result[i] = 1.0f; - } - } -} - -/* copy -log(output[label]) to this->data[i] */ -void CpuMatrix::oneHotCrossEntropy(Matrix& output, IVector& label) { - CHECK(dynamic_cast(&output)); - CHECK(dynamic_cast(&label)); - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(label.getSize(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(getWidth(), (size_t)1); - - real* out = output.getData(); - real* cost = getData(); - int* lbl = label.getData(); - for (size_t i = 0; i < numSamples; ++i, out += dim) { - CHECK_GE(lbl[i], 0); - CHECK_LT((size_t)lbl[i], dim); - cost[i] = -std::log(out[lbl[i]]); - } -} - -/* calculate the error of outputV according to label */ -void CpuMatrix::oneHotCrossEntropyBp(Matrix& output, IVector& label) { - CHECK(dynamic_cast(&output)); - CHECK(dynamic_cast(&label)); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(output.getWidth(), dim); - real* out = output.getData(); - real* grad = getData(); - int* lbl = label.getData(); - for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) { - grad[lbl[i]] -= 1 / out[lbl[i]]; - } -} - -/* - We implement the matrix functionality in CostLayer.cpp, - but we define the scalar function here for sanity check - deletion of the function does not affect anything neverthelss -*/ -void CpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output, - IVector& label, - real alpha) { - CHECK(dynamic_cast(&output)); - CHECK(dynamic_cast(&label)); - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(label.getSize(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(getWidth(), (size_t)1); - - real* out = output.getData(); - real* cost = getData(); - int* lbl = label.getData(); - for (size_t i = 0; i < numSamples; ++i, out += dim) { - CHECK_GE(lbl[i], 0); - CHECK_LT((size_t)lbl[i], dim); - real sum = 0; - for (size_t j = 0; j < dim; ++j) { - sum += out[j]; - } - sum = _safelog(sum); - cost[i] = -_safelog(out[lbl[i]]) + sum + alpha * _square(sum); - } -} - -/* - We implement the matrix functionality in CostLayer.cpp, - but we define the scalar function here for sanity check - deletion of the function does not affect anything neverthelss -*/ -void CpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& output, - IVector& label, - real alpha) { - CHECK(dynamic_cast(&output)); - CHECK(dynamic_cast(&label)); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(output.getWidth(), dim); - real* out = output.getData(); - real* grad = getData(); - int* lbl = label.getData(); - - for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) { - grad[lbl[i]] -= 1 / out[lbl[i]]; - real sum = 0; - for (size_t j = 0; j < dim; ++j) { - sum += out[j]; - } - for (size_t j = 0; j < dim; ++j) { - if (j == (size_t)lbl[i]) { - grad[j] += -1 / out[j]; - } - grad[j] += 1 / sum + 2 * alpha * _safelog(sum) / sum; - } - } -} - -#define FORWARD_LOOP() \ - size_t numSamples = getHeight(); \ - size_t dim = getWidth(); \ - CHECK_EQ(output.getHeight(), numSamples); \ - CHECK_EQ(output.getWidth(), dim); \ - const real* in = getData(); \ - real* out = output.getData(); \ - for (size_t i = 0; i < numSamples; ++i, in += dim, out += dim) - -#define BACKWARD_LOOP() \ - size_t numSamples = getHeight(); \ - size_t dim = getWidth(); \ - CHECK_EQ(output.getHeight(), numSamples); \ - CHECK_EQ(output.getWidth(), dim); \ - real* grad = getData(); \ - real* out = output.getData(); \ - for (size_t i = 0; i < numSamples; ++i, grad += dim, out += dim) - -void CpuMatrix::softmax(Matrix& output) { - CHECK(!output.useGpu()); - - const float THRESHOLD = -64.0; - - FORWARD_LOOP() { - real max = -1.0e20; - for (size_t j = 0; j < dim; ++j) { - if (in[j] > max) { - max = in[j]; - } - } - for (size_t j = 0; j < dim; ++j) { - real a = in[j] - max; - if (a < THRESHOLD) { - a = THRESHOLD; - } - out[j] = a; - } - vExp(dim, out, out); - - real sum = 0; - for (size_t j = 0; j < dim; ++j) { - sum += out[j]; - } - sum = 1 / sum; - for (size_t j = 0; j < dim; ++j) { - out[j] *= sum; - } - } -} - -void CpuMatrix::sequenceSoftmax(Matrix& output, const IVector& index) { - CHECK_EQ(getWidth(), 1UL); - CHECK_EQ(output.getWidth(), 1UL); - CHECK(isContiguous()); - - MatrixPtr inTmp = Matrix::create(nullptr, - /* height= */ 1, - 1, - /* trans= */ false, - false); - MatrixPtr outTmp = Matrix::create(nullptr, - /* height= */ 1, - 1, - /* trans= */ false, - false); - size_t numSequences = index.getSize() - 1; - auto starts = index.getData(); - for (size_t i = 0; i < numSequences; ++i) { - size_t offset = starts[i]; - size_t size = starts[i + 1] - starts[i]; - inTmp->setData(getData() + offset, 1UL, size); - outTmp->setData(output.getData() + offset, 1UL, size); - inTmp->softmax(*outTmp); - } -} - -void CpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) { - CHECK(output.useGpu_ == false) << "Matrix type are not equal"; - CHECK_EQ(getHeight(), sftmaxSum.getHeight()); - - real* sums = sftmaxSum.getData(); - - BACKWARD_LOOP() { - real sum = sums[i]; - for (size_t j = 0; j < dim; ++j) { - grad[j] = out[j] * (grad[j] - sum); - } - } -} - -void CpuMatrix::sumOfSquares(Matrix& output, Matrix& label) { - CHECK(output.useGpu_ == false && label.useGpu_ == false) - << "Matrix type are not equal"; - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(label.getHeight(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(label.getWidth(), dim); - CHECK_EQ(getWidth(), (size_t)1); - real* out = output.getData(); - real* cost = getData(); - - auto labelptr = dynamic_cast(&label); - if (labelptr) { - // it is a CpuSparseMatrix - if (labelptr->getFormat() == SPARSE_CSR) { - // treat label as a SparseMatrix - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = 0; j < dim; ++j) { - cost[i] += _square(out[i * dim + j]); - } - } - if (labelptr->getValueType() == NO_VALUE) { - int* cols = labelptr->getCols(); - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = labelptr->getRowStartIdx(i); - j < labelptr->getRowStartIdx(i + 1); - ++j) { - cost[i] += 1.0 - 2.0 * out[i * dim + cols[j]]; - /* - * explanation of above line: original codes are follows: - * cost[i] -= _square(out[i * dim + feature.col]); - * cost[i] += _square(1.0 - out[i * dim + feature.col]); - */ - } - } - } else if (labelptr->getValueType() == FLOAT_VALUE) { - int* cols = labelptr->getCols(); - real* values = labelptr->getValue(); - for (size_t i = 0; i < numSamples; ++i) { - real sum1 = 0; - real sum2 = 0; - for (size_t j = labelptr->getRowStartIdx(i); - j < labelptr->getRowStartIdx(i + 1); - ++j) { - sum1 += values[j] * values[j]; - sum2 += values[j] * out[i * dim + cols[j]]; - /* - * explanation of above line: original codes are follows: - * cost[i] -= _square(out[i * dim + feature.col]); - * cost[i] += _square(value.col - out[i * dim + feature.col]); - */ - } - cost[i] += sum1 - 2.0 * sum2; - } - } else { - LOG(FATAL) << "unsupported sparse matrix value type in sumOfSquares"; - return; - } - return; - } else { - LOG(FATAL) << "unsupported sparse matrix format in sumOfSquares"; - return; - } - } - - BaseMatrix::sumOfSquaredDiffs(output, - label, - /* scaleSum= */ 1, - /* scaleDest= */ 1); -} - -/* calculate the error of outputV according to label */ -void CpuMatrix::sumOfSquaresBp(Matrix& output, Matrix& label) { - CHECK(output.useGpu_ == false && label.useGpu_ == false) - << "Matrix type are not equal"; - - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(output.getWidth(), dim); - CHECK_EQ(label.getWidth(), dim); - - real* out = output.getData(); - real* grad = getData(); - - auto labelptr = dynamic_cast(&label); - if (labelptr) { - // it is a CpuSparseMatrix - if (labelptr->getFormat() == SPARSE_CSR) { - // treat label as a SparseMatrix - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = 0; j < dim; ++j) { - grad[i * dim + j] += 2.0 * out[i * dim + j]; - } - } - if (labelptr->getValueType() == NO_VALUE) { - int* cols = labelptr->getCols(); - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = labelptr->getRowStartIdx(i); - j < labelptr->getRowStartIdx(i + 1); - ++j) { - grad[i * dim + cols[j]] -= 2.0; - /* - * explanation of above line: original codes are follows: - * grad[i * dim + feature.col] -= 2.0 * out[i * dim + feature.col]; - * grad[i * dim + feature.col] += 2.0 * (out[i * dim + feature.col] - * - 1); - */ - } - } - } else if (labelptr->getValueType() == FLOAT_VALUE) { - int* cols = labelptr->getCols(); - real* values = labelptr->getValue(); - for (size_t i = 0; i < numSamples; ++i) { - for (size_t j = labelptr->getRowStartIdx(i); - j < labelptr->getRowStartIdx(i + 1); - ++j) { - grad[i * dim + cols[j]] -= 2.0 * values[j]; - /* - * explanation of above line: original codes are follows: - * grad[i * dim + feature.col] -= 2.0 * out[i * dim + feature.col]; - * grad[i * dim + feature.col] += 2.0 * (out[i * dim + feature.col] - * - value.col); - */ - } - } - } else { - LOG(FATAL) << "unsupported sparse matrix value type in sumOfSquares"; - return; - } - return; - } else { - LOG(FATAL) << "unsupported sparse matrix format in sumOfSquares"; - return; - } - } - - real* lbl = label.getData(); - size_t ld = getStride(); - size_t outLd = output.getStride(); - size_t lblLd = label.getStride(); - CHECK(lbl); - for (size_t i = 0; i < numSamples; - ++i, out += outLd, lbl += lblLd, grad += ld) { - for (size_t j = 0; j < dim; ++j) { - grad[j] += 2.0 * (out[j] - lbl[j]); // positive gradient; - } - } -} - -void CpuMatrix::smoothL1(Matrix& output, Matrix& label, real destScale) { - CHECK(output.useGpu_ == false && label.useGpu_ == false) - << "Matrix type are not equal"; - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(label.getHeight(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(label.getWidth(), dim); - CHECK_EQ(getWidth(), (size_t)1); - - real* cost = getData(); - real* out = output.getData(); - real* lbl = label.getData(); - - for (size_t i = 0; i < numSamples; ++i, out += dim, lbl += dim) { - for (size_t j = 0; j < dim; ++j) { - real absVal = std::fabs(out[j] - lbl[j]); - cost[i] *= destScale; - if (absVal < 1.0) - cost[i] += 0.5 * absVal * absVal; - else - cost[i] += absVal - 0.5; - } - } -} - -void CpuMatrix::smoothL1Bp(Matrix& output, Matrix& label, real destScale) { - CHECK(output.useGpu_ == false && label.useGpu_ == false) - << "Matrix type are not equal"; - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(label.getHeight(), numSamples); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(label.getWidth(), dim); - CHECK_EQ(getWidth(), dim); - - real* out = output.getData(); - real* lbl = label.getData(); - real* grad = getData(); - - for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim, lbl += dim) { - for (size_t j = 0; j < dim; ++j) { - real val = out[j] - lbl[j]; - grad[j] *= destScale; - if (std::fabs(val) < 1) { - grad[j] += val; - } else { - grad[j] += (real(0) < val) - (val < real(0)); - } - } - } -} - -void CpuMatrix::tanh(Matrix& output) { - CHECK(isContiguous()); - CHECK(output.isContiguous()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(output.getWidth(), dim); - vTanh(numSamples * dim, getData(), output.getData()); -} - -void CpuMatrix::tanhDerivative(Matrix& output) { - BaseMatrix::tanhDerivative(output); -} - -void CpuMatrix::softrelu(Matrix& output) { - CHECK(isContiguous()); - CHECK(output.isContiguous()); - const real THRESHOLD = 40.0; - FORWARD_LOOP() { // TODO(yuyang18): SIMD it? - for (size_t j = 0; j < dim; ++j) { - real x = in[j]; - if (x > THRESHOLD) { - x = THRESHOLD; - } else if (x < -THRESHOLD) { - x = -THRESHOLD; - } - out[j] = x; - } - } - vExp(numSamples * dim, output.getData(), output.getData()); - vLog1p(numSamples * dim, output.getData(), output.getData()); -} - -void CpuMatrix::softreluDerivative(Matrix& output) { - CHECK(isContiguous()); - CHECK(output.isContiguous()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - size_t size = numSamples * dim; - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(output.getWidth(), dim); - real* grad = getData(); - MatrixPtr tmpMat = Matrix::create(numSamples, dim); - real* tmp = tmpMat->getData(); - - vExp(size, output.getData(), tmpMat->getData()); - - for (size_t i = 0; i < size; ++i) { - grad[i] *= (1.0 - 1.0 / tmp[i]); - } -} - -void CpuMatrix::scaledTanh(Matrix& output, real p1, real p2) { - CHECK(isContiguous()); - CHECK(output.isContiguous()); - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(output.getHeight(), numSamples); - CHECK_EQ(output.getWidth(), dim); - - const real* in = getData(); - real* out = output.getData(); - - // out = p2*in - for (size_t i = 0; i < numSamples * dim; ++i) { - out[i] = p2 * in[i]; - } - - vTanh(numSamples * dim, out, out); - - // out = p1 * out - for (size_t i = 0; i < numSamples * dim; ++i) { - out[i] = p1 * out[i]; - } -} - -/* uniform randomization, minimize precision = 1e-5 */ -void CpuMatrix::randomizeUniform() { - CHECK(isContiguous()); - real* data = getData(); - unsigned int* randSeed = ThreadLocalRand::getSeed(); - real recipRandMax = 1.0f / (real)RAND_MAX; - for (size_t i = 0; i < elementCnt_; ++i) { - *data++ = rand_r(randSeed) * recipRandMax; - } -} - -void CpuMatrix::print(std::ostream& os) const { - CHECK(isContiguous()); - for (size_t i = 0; i < height_; ++i) { - for (size_t j = 0; j < width_; ++j) { - os << data_[i * width_ + j] << " "; - } - os << std::endl; - } -} - -void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) { - real* input = data.getData(); - real* w = W.getData(); - real* output = data_; - size_t numElements = data.getWidth(); - size_t numSamples = data.getHeight(); - size_t paraSize = W.getHeight() * W.getWidth(); - CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init - - size_t partial_sum = numElements / paraSize; - if (paraSize == numElements) { - for (size_t n = 0; n < numSamples * numElements; ++n) { - output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements]; - } - return; - } - -#if defined(__ARM_NEON__) || defined(__ARM_NEON) - for (size_t n = 0; n < numSamples; ++n) { - for (size_t i = 0; i < paraSize; i++) { - neon::prelu( - input + i * partial_sum, w[i], output + i * partial_sum, partial_sum); - } - input = input + numElements; - output = output + numElements; - } -#else - for (size_t n = 0, k = 0; n < numSamples; ++n) { - for (size_t i = 0; i < numElements; ++i, ++k) { - output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum]; - } - } -#endif -} - -void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) { - real* ograd = oGrad.getData(); - real* input = data.getData(); - real* wgrad = data_; - size_t numElements = data.getWidth(); - size_t numSamples = data.getHeight(); - size_t paraSize = this->getHeight() * this->getWidth(); - CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init - size_t partial_sum = numElements / paraSize; - for (size_t n = 0, k = 0; n < numSamples; ++n) { - for (size_t i = 0; i < numElements; ++i, ++k) { - wgrad[i / partial_sum] += ograd[k] * (input[k] > 0 ? 0 : input[k]); - } - } -} - -void CpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) { - real* diff = data_; - real* input = data.getData(); - real* ograd = oGrad.getData(); - real* w = W.getData(); - size_t numElements = data.getWidth(); - size_t numSamples = data.getHeight(); - size_t paraSize = W.getHeight() * W.getWidth(); - CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init - size_t partial_sum = numElements / paraSize; - for (size_t n = 0, k = 0; n < numSamples; ++n) { - for (size_t i = 0; i < numElements; ++i, ++k) { - diff[k] += ograd[k] * (input[k] > 0 ? 1 : w[i / partial_sum]); - } - } -} - -void CpuMatrix::print(std::ostream& os, size_t height, size_t width) const { - CHECK(isContiguous()); - size_t h = height_ < height ? height_ : height; - size_t w = width_ < width ? width_ : width; - os.setf(std::ostream::scientific); - os << "["; - for (size_t i = 0; i < h; ++i) { - for (size_t j = 0; j < w; ++j) { - os << data_[i * width_ + j] << " "; - } - if (i == h - 1) { - os << "]"; - } - os << std::endl; - } -} - -void CpuMatrix::printOneRow(std::ostream& os, size_t idx) const { - CHECK_LT(idx, height_); - size_t offset = idx * stride_; - os << data_[offset]; - for (size_t i = 1; i < width_; ++i) { - os << " " << data_[offset + i]; - } - os << ";"; -} - -void CpuMatrix::check(std::ostream& os, Matrix& refMat, bool printDiff) { - CHECK(isContiguous()); - CHECK(height_ == refMat.getHeight()); - CHECK(width_ == refMat.getWidth()); - CpuMatrix cpuRef(height_, width_); - cpuRef.copyFrom(refMat); - size_t diffCnt = 0; - for (size_t i = 0; i < height_; ++i) { - for (size_t j = 0; j < width_; ++j) { - real a = getElement(i, j); - real b = cpuRef.getElement(i, j); - if (fabs(a - b) > 0.00001) { - ++diffCnt; - if (printDiff) { - os << "ref= " << a << " check= " << b << std::endl; - } - } - } - } - LOG(INFO) << "the diffCnt is " << diffCnt; -} - -real CpuMatrix::getMin() { - size_t size = getHeight() * getWidth(); - real* data = getData(); - real res = data[0]; - for (size_t i = 1; i < size; ++i) { - if (res > data[i]) { - res = data[i]; - } - } - return res; -} - -real CpuMatrix::getMax() { - size_t size = getHeight() * getWidth(); - real* data = getData(); - real res = data[0]; - for (size_t i = 1; i < size; ++i) { - if (res < data[i]) { - res = data[i]; - } - } - return res; -} - -void CpuMatrix::circularConv(Matrix& in0, Matrix& in1) { - size_t height = this->getHeight(); - size_t width0 = this->getWidth(); - size_t width1 = in1.getWidth(); - - CHECK_EQ(height, in0.getHeight()); - CHECK_EQ(width0, in0.getWidth()); - CHECK_EQ(height, in1.getHeight()); - - CHECK_EQ(width1 % 2, 1U); - - real* outV = this->getData(); - real* inV0 = in0.getData(); - real* inV1 = in1.getData(); - - int leftCtxLen = (width1 - 1) / 2; - for (size_t x = 0; x < height; - ++x, outV += width0, inV0 += width0, inV1 += width1) { - for (size_t i = 0; i < width0; ++i) { // each dimension of output - for (size_t j = 0; j < width1; ++j) { - // iterate over all dimentions of inV1 - int index = i + j - leftCtxLen; - index = (index + width0) % width0; - outV[i] += inV0[index] * inV1[j]; - } - } - } -} - -void CpuMatrix::circularConvDerivative( - Matrix& outG, Matrix& in0, Matrix& in1, Matrix& inG0, Matrix& inG1) { - size_t height = in0.getHeight(); - size_t width0 = in0.getWidth(); - size_t width1 = in1.getWidth(); - - CHECK_EQ(height, in1.getHeight()); - CHECK_EQ(height, inG0.getHeight()); - CHECK_EQ(width0, inG0.getWidth()); - CHECK_EQ(height, inG1.getHeight()); - CHECK_EQ(width1, inG1.getWidth()); - CHECK_EQ(height, outG.getHeight()); - CHECK_EQ(width0, outG.getWidth()); - - real* outGV = outG.getData(); - real* inV0 = in0.getData(); - real* inV1 = in1.getData(); - real* inGV0 = inG0.getData(); - real* inGV1 = inG1.getData(); - - int leftCtxLen = (width1 - 1) / 2; - for (size_t x = 0; x < height; ++x, - outGV += width0, - inV0 += width0, - inV1 += width1, - inGV0 += width0, - inGV1 += width1) { - for (size_t j = 0; j < width1; ++j) { // iterate over width1 - for (size_t i = 0; i < width0; ++i) { - // such over all dimensions of outG - int index = i + j - leftCtxLen; - index = (index + width0) % width0; - inGV0[index] += outGV[i] * inV1[j]; - inGV1[j] += outGV[i] * inV0[index]; - } - } - } -} - -void CpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) { - CHECK(dynamic_cast(&output)); - auto labelPtr = dynamic_cast(&label); - CHECK(labelPtr); - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(numSamples, output.getHeight()); - CHECK_EQ(numSamples, labelPtr->getHeight()); - CHECK_EQ(dim, labelPtr->getWidth()); - - real* out = output.getData(); - real* cost = getData(); - for (size_t i = 0; i < numSamples; ++i, out += dim) { - for (size_t j = 0; j < dim; ++j) { - CHECK(out[j] > 0 && out[j] < 1.0); - cost[i] -= std::log(1 - out[j]); - } - - const int* cols = labelPtr->getRowCols(i); - for (size_t j = 0; j < labelPtr->getColNum(i); ++j) { - CHECK_LT(size_t(cols[j]), dim); - cost[i] -= std::log(out[cols[j]] / (1 - out[cols[j]])); - } - } -} - -void CpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { - CHECK(dynamic_cast(&output)); - auto labelPtr = dynamic_cast(&label); - CHECK(labelPtr); - - size_t numSamples = getHeight(); - size_t dim = getWidth(); - CHECK_EQ(numSamples, output.getHeight()); - CHECK_EQ(numSamples, labelPtr->getHeight()); - CHECK_EQ(dim, output.getWidth()); - CHECK_EQ(dim, labelPtr->getWidth()); - - real* out = output.getData(); - real* grad = getData(); - for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) { - for (size_t j = 0; j < dim; ++j) { - CHECK(out[j] > 0 && out[j] < 1.0); - grad[j] += 1.0 / (1 - out[j]); - } - - const int* cols = labelPtr->getRowCols(i); - for (size_t j = 0; j < labelPtr->getColNum(i); ++j) { - CHECK_LT(size_t(cols[j]), dim); - grad[cols[j]] -= 1.0 / (out[cols[j]] * (1 - out[cols[j]])); - } - } -} - -/* calculate the classification error for multi binary label */ -void CpuMatrix::classificationErrorMulti(Matrix& output, - Matrix& label, - real threshold) { - CHECK(dynamic_cast(&output)); - auto labelPtr = dynamic_cast(&label); - CHECK(labelPtr); - - size_t numSamples = getHeight(); - size_t dim = output.getWidth(); - CHECK_EQ(numSamples, output.getHeight()); - CHECK_EQ(numSamples, labelPtr->getHeight()); - CHECK_EQ(dim, labelPtr->getWidth()); - - real* out = output.getData(); - real* result = getData(); - for (size_t i = 0; i < numSamples; ++i, out += dim) { - real sum = 0.0; - for (size_t j = 0; j < dim; ++j) { - if (out[j] >= threshold) { - sum += 1.0; - } - } - - const int* cols = labelPtr->getRowCols(i); - for (size_t j = 0; j < labelPtr->getColNum(i); ++j) { - CHECK_LT(size_t(cols[j]), dim); - if (out[cols[j]] < threshold) { - sum += 1.0; - } else { - sum -= 1.0; - } - } - result[i] = sum / dim; - } -} - -void CpuMatrix::bilinearForward(const Matrix& in, - const size_t inImgH, - const size_t inImgW, - const size_t outImgH, - const size_t outImgW, - const size_t numChannels, - const real ratioH, - const real ratioW) { - CHECK(dynamic_cast(&in)); - - size_t outputW = getWidth(); - size_t batchSize = getHeight(); - size_t inputW = in.getWidth(); - size_t inputH = in.getHeight(); - size_t inPosOffset = inImgH * inImgW; - size_t outPosOffset = outImgH * outImgW; - (void)(inputH); - - real* outData = getData(); - const real* inData = in.getData(); - - if (inImgH == outImgH && inImgW == outImgW) { - this->copyFrom(in); - } else { - for (size_t k = 0; k < batchSize; ++k) { // loop for batches - for (size_t i = 0; i < outImgH; ++i) { // loop for images - size_t h = ratioH * i; - size_t hid = (h < inImgH - 1) ? 1 : 0; - real h1lambda = ratioH * i - h; - real h2lambda = 1 - h1lambda; - - for (size_t j = 0; j < outImgW; ++j) { - size_t w = ratioW * j; - size_t wid = (w < inImgW - 1) ? 1 : 0; - real w1lambda = ratioW * j - w; - real w2lambda = 1 - w1lambda; - // calculate four position for bilinear interpolation - const real* inPos = &inData[k * inputW + h * inImgW + w]; - real* outPos = &outData[k * outputW + i * outImgW + j]; - for (size_t c = 0; c < numChannels; ++c) { // loop for channels - // bilinear interpolation - outPos[0] = - h2lambda * (w2lambda * inPos[0] + w1lambda * inPos[wid]) + - h1lambda * (w2lambda * inPos[hid * inImgW] + - w1lambda * inPos[hid * inImgW + wid]); - inPos += inPosOffset; - outPos += outPosOffset; - } - } - } - } - } -} - -void CpuMatrix::bilinearBackward(const Matrix& out, - const size_t outImgH, - const size_t outImgW, - const size_t inImgH, - const size_t inImgW, - const size_t numChannels, - const real ratioH, - const real ratioW) { - CHECK(dynamic_cast(&out)); - - size_t inputW = getWidth(); - size_t inputH = getHeight(); - size_t outputW = out.getWidth(); - size_t batchSize = out.getHeight(); - size_t inPosOffset = inImgH * inImgW; - size_t outPosOffset = outImgH * outImgW; - (void)(inputH); - - real* inGrad = getData(); - const real* outGrad = out.getData(); - - if (inImgH == outImgH && inImgW == outImgW) { - this->add(const_cast(out)); - } else { - for (size_t k = 0; k < batchSize; ++k) { // loop for batches - for (size_t i = 0; i < outImgH; ++i) { // loop for images - size_t h = ratioH * i; - size_t hid = (h < inImgH - 1) ? 1 : 0; - real h1lambda = ratioH * i - h; - real h2lambda = 1 - h1lambda; - for (size_t j = 0; j < outImgW; ++j) { - size_t w = ratioW * j; - size_t wid = (w < inImgW - 1) ? 1 : 0; - real w1lambda = ratioW * j - w; - real w2lambda = 1 - w1lambda; - - real* inPos = &inGrad[k * inputW + h * inImgW + w]; - const real* outPos = &outGrad[k * outputW + i * outImgW + j]; - for (size_t c = 0; c < numChannels; ++c) { // loop for channels - inPos[0] += h2lambda * w2lambda * outPos[0]; - inPos[wid] += h2lambda * w1lambda * outPos[0]; - inPos[hid * inImgW] += h1lambda * w2lambda * outPos[0]; - inPos[hid * inImgW + wid] += h1lambda * w1lambda * outPos[0]; - inPos += inPosOffset; - outPos += outPosOffset; - } - } - } - } - } -} - -void CpuMatrix::vol2Col(real* data, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW) { - real* outData = getData(); - int outHeight = (height + 2 * paddingH - filterH) / strideH + 1; - int outWidth = (width + 2 * paddingW - filterW) / strideW + 1; - int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1; - - int channelsCol = channels * filterD * filterH * filterW; - for (int c = 0; c < channelsCol; ++c) { - int wOffset = c % filterW; - int hOffset = (c / filterW) % filterH; - int dOffset = (c / filterW / filterH) % filterD; - int cIn = c / filterW / filterH / filterD; - for (int d = 0; d < outDepth; ++d) { - for (int h = 0; h < outHeight; ++h) { - for (int w = 0; w < outWidth; ++w) { - int dPad = d * strideD - paddingD + dOffset; - int hPad = h * strideH - paddingH + hOffset; - int wPad = w * strideW - paddingW + wOffset; - - if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width && - dPad >= 0 && dPad < depth) - outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = - data[((cIn * depth + dPad) * height + hPad) * width + wPad]; - else - outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = 0; - } - } - } - } -} - -void CpuMatrix::col2Vol(real* trg, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - real alpha, - real beta) { - real* src = getData(); - int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1; - int outHeight = (height + 2 * paddingH - filterH) / strideH + 1; - int outWidth = (width + 2 * paddingW - filterW) / strideW + 1; - int channelsCol = channels * filterD * filterH * filterW; - for (int c = 0; c < channelsCol; ++c) { - int wOffset = c % filterW; - int hOffset = (c / filterW) % filterH; - int dOffset = (c / filterW / filterH) % filterD; - int cIm = c / filterW / filterH / filterD; - for (int d = 0; d < outDepth; ++d) { - for (int h = 0; h < outHeight; ++h) { - for (int w = 0; w < outWidth; ++w) { - int dPad = d * strideD - paddingD + dOffset; - int hPad = h * strideH - paddingH + hOffset; - int wPad = w * strideW - paddingW + wOffset; - if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width && - dPad >= 0 && dPad < depth) - trg[((cIm * depth + dPad) * height + hPad) * width + wPad] = - alpha * - src[((c * outDepth + d) * outHeight + h) * outWidth + w] + - beta * - trg[((cIm * depth + dPad) * height + hPad) * width + wPad]; - } - } - } - } -} - -//////////////////////////////////////////////////////////////// -// functions executed via cpu // -//////////////////////////////////////////////////////////////// - -void GpuMatrix::selectElements(Matrix& table, IVector& ids) { - execViaCpu2(&CpuMatrix::selectElements, *this, table, ids); -} -} // namespace paddle diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h deleted file mode 100644 index 4c3b2c95361065372f5969a2da73bce0eb9d123f..0000000000000000000000000000000000000000 --- a/paddle/math/Matrix.h +++ /dev/null @@ -1,2189 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include - -#include "paddle/utils/Logging.h" -#include "paddle/utils/ThreadLocal.h" - -#include - -#include "BaseMatrix.h" -#include "MemoryHandle.h" -#include "Vector.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -/// TODO(tianbing), move to paddle/function/TensorType.h -enum SparseValueType { NO_VALUE = 0, FLOAT_VALUE = 1 }; - -/** - * @brief matrix sparse_format . - * - * nnz represents nonzero number in sparse matrix. - * - * SPARSE_CSR: row major matrix. length of row is height_ + 1, each element - * represents row start index in Matrix. length of col and value are nnz. - * - * SPARSE_CSC: col major matrix. length of col is width_ + 1, each element - * represents col start index in Matrix. length of col and value are nnz. - * - * @code - * for example: [0, 1, 0, 2, 0; - * 1, 0, 0, 0, 0; - * 0, 0, 0, 2, 5]; - * SPARSE_CSR row [0, 2, 3, 5]; - * col [1, 3, 0, 3, 4]; - * value [1, 2, 1, 2, 5] - * SPARSE_CSC col [0, 1, 2, 2, 4, 5]; - * row [1, 0, 0, 2, 2]; - * value [1, 1, 2, 2, 5] - * @endcode - */ -/// TODO(tianbing), move to paddle/function/TensorType.h -enum SparseFormat { SPARSE_CSR = 0, SPARSE_CSC = 1 }; - -class Matrix; -class GpuMatrix; -class CpuMatrix; -class CpuSparseMatrix; -class GpuSparseMatrix; -typedef std::shared_ptr MatrixPtr; -typedef std::shared_ptr GpuMatrixPtr; -typedef std::shared_ptr CpuMatrixPtr; -typedef std::shared_ptr GpuSparseMatrixPtr; -typedef std::shared_ptr CpuSparseMatrixPtr; - -/** - * Copy or assignemnt constructor will share the data as opposed to making a - * copy of the original data. To make a copy of the orinal data, use copyFrom() - * instead. - */ -class Matrix : public BaseMatrix { - protected: - Matrix(MemoryHandlePtr memHandle, - size_t height, - size_t width, - bool trans, - bool use_gpu); - - Matrix(real* data, size_t height, size_t width, bool trans, bool use_gpu); - - Matrix(real* data, - size_t height, - size_t width, - size_t stride, - bool trans, - bool use_gpu); - - static ThreadLocal tmpMat_; - - public: - size_t elementCnt_; // maximal number of elements which can be held in data_ - MemoryHandlePtr memoryHandle_; - - public: - virtual ~Matrix() {} - - static MatrixPtr create(MemoryHandlePtr memHandle, - size_t height, - size_t width, - bool trans = false); - static MatrixPtr create(size_t height, - size_t width, - bool trans = false, - bool useGpu = false); - static MatrixPtr create(real* data, - size_t height, - size_t width, - bool trans = false, - bool useGpu = false); - static MatrixPtr create(real* data, - size_t height, - size_t width, - size_t stride, - bool trans = false, - bool useGpu = false); - - static MatrixPtr createSparseMatrix(size_t height, - size_t width, - size_t nnz, - SparseValueType valueType = FLOAT_VALUE, - bool trans = false, - bool useGpu = false); - static MatrixPtr createSparseMatrix(size_t height, - size_t width, - size_t nnz, - SparseValueType valueType = FLOAT_VALUE, - SparseFormat foramt = SPARSE_CSR, - bool trans = false, - bool useGpu = false); - - static MatrixPtr createSparseMatrix(real* data, - int* row, - int* col, - size_t height, - size_t width, - size_t nnz, /* used to allocate space */ - SparseValueType valueType, /*value type*/ - SparseFormat format, - bool trans, - bool useGpu); - - static void resizeOrCreateSparseMatrix( - MatrixPtr& matrix, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType = FLOAT_VALUE, - SparseFormat foramt = SPARSE_CSR, - bool trans = false, - bool useGpu = false); - - static void resizeOrCreate(MatrixPtr& a, - size_t height, - size_t width, - bool trans = false, - bool useGpu = false); - - /** - * @brief set the data buffer used to hold the matrix data. - * - * caller should make sure that the size of data is at least - * sizeof(real)*height*width. - */ - void setData(real* data) { - BaseMatrix::setData(data); - memoryHandle_.reset(); - } - - /// the data should be contiguous - void setData(real* data, size_t newHeight, size_t newWidth) { - setData(data); - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newHeight * newWidth; - stride_ = width_; - } - - size_t getWidth() const { return width_; } - size_t getHeight() const { return height_; } - size_t getStride() const { return stride_; } - size_t getElementCnt() const { return elementCnt_; } - virtual real* getData() { return data_; } - virtual const real* getData() const { return data_; } - bool isTransposed() const { return trans_; } - bool isContiguous() const { return stride_ == width_ || height_ == 1; } - - // If sparse matrix, need to dynamic_cast to CpuSparseMatrix/GpuSparseMatrix - // befor call the following functions. - // Declare these functions in the base class just easy to call them. - // And these declarations should be moved to base class of sparse matrix - // if refactor sparse matrix - virtual int* getRows() const { - LOG(FATAL) << "Not implemented"; - return nullptr; //! suppress warning for no return value. - } - - virtual int* getCols() const { - LOG(FATAL) << "Not implemented"; - return nullptr; //! suppress warning for no return value. - } - - virtual SparseFormat getFormat() const { - LOG(FATAL) << "Not implemented"; - return SPARSE_CSR; //! suppress warning for no return value. - } - - virtual SparseValueType getValueType() const { - LOG(FATAL) << "Not implemented"; - return NO_VALUE; //! suppress warning for no return value. - } - - /** - * @brief matrix elment-wise add - * - * Named add3 just because add/add2 has been used in BaseMatrix.cu - * and they are not virtual function. - */ - virtual void add3(MatrixPtr b) { LOG(FATAL) << "Not implemented"; } - - MemoryHandlePtr getMemoryHandle() const { return memoryHandle_; } - - virtual void zeroMem() { LOG(FATAL) << "Not implemented"; } - - virtual void resetOne() { LOG(FATAL) << "Not implemented"; } - - void setDiag(real value); - - virtual void copyFrom(const Matrix& src) { LOG(FATAL) << "Not implemented"; } - - virtual void trimFrom(const CpuSparseMatrix& src) { - LOG(FATAL) << "Not implemented"; - } - - // For GpuMatrix this is an asynchronous copy interface - // For CpuMatrix this is an synchronous copy interface - virtual void copyFrom(const Matrix& src, hl_stream_t stream) { - LOG(FATAL) << "Not implemented"; - } - - MatrixPtr subMatrix(size_t startRow, - size_t endRow, - size_t startCol, - size_t endCol); - - MatrixPtr subRowMatrix(size_t startRow, size_t endRow) { - return subMatrix(startRow, endRow, 0, getWidth()); - } - - MatrixPtr subColMatrix(size_t startCol, size_t endCol) { - return subMatrix(0, getHeight(), startCol, endCol); - } - - virtual MatrixPtr subMatrix(size_t startRow, size_t numRows) { - CHECK_LE(startRow + numRows, getHeight()); - return Matrix::create(getData() + startRow * getWidth(), - numRows, - getWidth(), - trans_, - useGpu_); - } - virtual MatrixPtr subMatrix(size_t startRow, size_t numRows, MatrixPtr dest) { - CHECK_LE(startRow + numRows, getHeight()); - CHECK_EQ(useGpu_, dest->useGpu_); - dest->setData(this->rowBuf(startRow), numRows, getWidth()); - return dest; - } - - /** - * If this is GpuMatrix, src is assumed to be CPU memory - * - * If this is CpuMatrix, src is assumed to be CPU memory - */ - virtual void copyFrom(const real* src, size_t size) { - LOG(FATAL) << "Not implemented"; - } - - virtual void copyFrom(const real* src, const int64_t* seq) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @brief convert a int vector to a real matrix. - * - * (1) source and dest are both in CPU. - * - * (2) sizes are exactly match. - */ - virtual void copyFrom(const IVector& src) { - LOG(FATAL) << "copy data from int vector only available on CpuMatrix."; - } - - virtual void copyByRowIndex(Matrix& b, const IVector& rowIndex) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @brief Create a matrix with the same type (GpuMatrix, CpuMatrix, - * NonValueSparseMatrix, etc.) as this. - * - * If height and width is zero, the new matrix will have the same size - * as this, otherwise the new matrix will have the specified size. - * - */ - virtual MatrixPtr clone(size_t height = 0, - size_t width = 0, - bool useGpu = false) { - LOG(FATAL) << "Not implemented"; - return nullptr; - } - - virtual real* getRowBuf(size_t row) { - LOG(FATAL) << "Not implemented"; - return nullptr; - } - - virtual real getElement(size_t x, size_t y) const { - LOG(FATAL) << "Not implemented"; - return 0; - } - - virtual real getSum() { - LOG(FATAL) << "Not implemented"; - return 0; - } - - virtual void accumulateColSum(Matrix& src) { - LOG(FATAL) << "Not implemented"; - } - - virtual real getAbsSum() { - LOG(FATAL) << "Not implemented"; - return 0; - } - - /** - * @note Original data may not be preserved after resize(). - */ - virtual void resize(size_t newHeight, size_t newWidth) = 0; - - /** - * @note This should only be used for sparse matrix. - */ - virtual void resize(size_t newHeight, - size_t newWidth, - size_t newNnz, /* total item used to allocate space */ - SparseValueType valueType, - SparseFormat format) = 0; - - /** - * @brief This should only be used for sparse matrix. - * - * Currently must be called for each row in order. - * The matrix is not valid until setRow is called for the last row. - */ - virtual void setRow(size_t row, - size_t colNum, - const unsigned int* cols, - const real* values) = 0; - - virtual MatrixPtr getTranspose() = 0; - - /** - * @brief hard transpose. - * - * allocate matTrans' memory outside, then set memAlloc as false; - * else set as true. - */ - virtual void transpose(MatrixPtr& matTrans, bool memAlloc) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @brief rotate 90 degrees in clock-wise if clockWise=true; - * otherwise rotate in anti clock-wise - * clock-wise: - * \f[ - * y(j,i) = x(M-i-1,j) - * \f] - * anti clock-wise: - * \f[ - * y(j,i) = x(i, N-1-j) - * \f] - * where \f$x\f$ is (M x N) input, and \f$y\f$ is (N x M) output. - * - * allocate matRot' memory outside, then set memAlloc as false; - * else set as true. - */ - virtual void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) { - LOG(FATAL) << "Not implemented"; - } - - virtual MatrixPtr getInverse() { - LOG(FATAL) << "Not implemented"; - return nullptr; - } - - /** - * @brief inverse. - * - * if allocate matInv's memory outside, then set memAlloc as false; - * else set as true. - */ - virtual void inverse(MatrixPtr& matInv, bool memAlloc) { - LOG(FATAL) << "Not implemented"; - } - - public: - /// Only set all variables to 0 or NULL but not free them. - virtual void clear() { - height_ = 0; - width_ = 0; - data_ = NULL; - } - - void reshape(size_t height, size_t width); - - /// add b to each sample of this. - virtual void addBias(Matrix& b, real scale) { - LOG(FATAL) << "Not implemented"; - } - - virtual void addSharedBias(Matrix& b, real scale) { - LOG(FATAL) << "Not implemented"; - } - - void addBias(Matrix& b, real scale, bool sharedBias) { - if (!sharedBias) { - addBias(b, scale); - } else { - addSharedBias(b, scale); - } - } - - /// add each sample from a to this. - virtual void collectBias(Matrix& a, real scale) { - LOG(FATAL) << "Not implemented"; - } - - virtual void collectSharedBias(Matrix& a, real scale) { - LOG(FATAL) << "Not implemented"; - } - - void collectBias(Matrix& a, real scale, bool sharedBias) { - if (!sharedBias) { - collectBias(a, scale); - } else { - collectSharedBias(a, scale); - } - } - - virtual void sequenceAvgForward(Matrix& a, - const IVector& startsPos, - int mode) { - LOG(FATAL) << "Not implemented"; - } - - virtual void sequenceAvgBackward(Matrix& a, - const IVector& startsPos, - int mode) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * this = scaleAB*(a*b) + scaleT*this - * @endcode - */ - virtual void mul(const Matrix& a, - const Matrix& b, - real scaleAB, - real scaleT) { - LOG(FATAL) << "Not implemented"; - } - - /// Add a vector (column) b to matrix a, column by column. - virtual void addColumnVector(const Matrix& b) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * For j < codeLength: - * this(i, j) += vec(index(i, j), 0) - * where index(i, j) = ((codes(i) + numClasses) >> (j + 1)) - 1 - * @endcode - */ - virtual void addByBitCode(size_t numClasses, - const IVector& codes, - const Matrix& vec) { - (void)numClasses; - (void)codes; - (void)vec; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * For j < codeLength: - * vec(index(i, j), 0) += this(i, j) - * where index is same as the index for addByBitCode - * @endcode - */ - virtual void addByBitCodeBackward(size_t numClasses, - const IVector& codes, - Matrix& vec) { - (void)numClasses; - (void)codes; - (void)vec; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * For j < codeLength: - * this(i, j) += - * where index is same as the index for addByBitCode - * @endcode - */ - virtual void mulByBitCode(size_t numClasses, - const IVector& codes, - const Matrix& mat, - const Matrix& input) { - (void)numClasses; - (void)codes; - (void)mat; - (void)input; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * For j < codeLength: - * mat.row(index(i, j)) += this(i, j) * input.row(i) - * where index is same as the index for addByBitCode - * @endcode - */ - virtual void mulByBitCodeBackwardWeight(size_t numClasses, - const IVector& codes, - Matrix& mat, - const Matrix& input) { - (void)numClasses; - (void)codes; - (void)mat; - (void)input; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * For j < codeLength: - * input.row(i) += this(i, j) * mat.row(index(i, j)) - * where index is same as the index for addByBitCode - * @endcode - */ - virtual void mulByBitCodeBackwardError(size_t numClasses, - const IVector& codes, - const Matrix& mat, - Matrix& input) { - (void)numClasses; - (void)codes; - (void)mat; - (void)input; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * For j < codeLength - * sum(i, 0) = scaleSum * \sum_j bit(i, j) * this(i, j) - * where bit(i, j) = ((codes(i) + numClasses) & 2^j) ? 1 : 0 - * @endcode - */ - virtual void sumByBitCode(size_t numClasses, - IVector& codes, - Matrix& sum, - real scaleSum) { - (void)numClasses; - (void)codes; - (void)sum; - (void)scaleSum; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * For j < codeLength - * this(i, j) -= bit(i, j) - * where bit(i, j) is same as that for sumByBitCode - * @endcode - */ - virtual void subByBitCode(size_t numClasses_, IVector& codes) { - (void)numClasses_; - (void)codes; - LOG(FATAL) << "Not implemeted"; - } - - /** - * add the sum of each row of this to mat - */ - virtual void rowSum(Matrix& sum) { - (void)sum; - LOG(FATAL) << "Not implemeted"; - } - - /** - * set the max of each row of this to mat - */ - virtual void rowMax(Matrix& max) { - (void)max; - LOG(FATAL) << "Not implemeted"; - } - - /** - * set the max of each column of this to mat - */ - virtual void colMax(Matrix& max) { LOG(FATAL) << "not implemented"; } - - /** - * @brief Get the top k elements of each column of this matrix. - * - * The row ids and values of these elements are stored in - * maxIds and max respectively. where k is the size of maxIds. - * And note that the top k elements are not sorted. - */ - virtual void colMax(IVector& maxIds, Matrix& maxVal) { - LOG(FATAL) << "not implemented"; - } - - virtual void maxoutForward(Matrix& a, - IVector& id, - size_t channels, - size_t groups) { - LOG(FATAL) << "not implemented"; - } - - virtual void maxoutBackward(Matrix& a, - IVector& id, - size_t channels, - size_t groups) { - LOG(FATAL) << "not implemented"; - } - - virtual void rowMaxId(IVector& maxIds) { LOG(FATAL) << "Not implemented"; } - - /** - * @brief Get the top k elements of each row of this matrix. - * - * The column ids and values of these elements are stored in - * maxIds and max respectively. where k is the size of maxIds. - * And note that the top k elements are not sorted. - */ - virtual void rowMax(IVector& maxIds, Matrix& max) { - LOG(FATAL) << "Not implemented"; - } - - /// normalize each row so that the sum of each row is 1. - virtual void rowNormalizeL1(Matrix& out) { - (void)out; - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * this = a*b - * @endcode - */ - virtual void mul(const Matrix& a, const Matrix& b) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * this = scaleAB*(this*b) + scaleT*this - * @endcode - */ - virtual void rightMul(Matrix& b, real scaleAB, real scaleT) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * this = this* b - * @endcode - */ - virtual void rightMul(Matrix& b) { LOG(FATAL) << "Not implemented"; } - - /** - * @code - * this = scaleAB*(a*this) + scaleT*this - * @endcode - */ - virtual void leftMul(Matrix& a, real scaleAB, real scaleT) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * this = a*this) - * @endcode - */ - virtual void leftMul(Matrix& a) { LOG(FATAL) << "Not implemented"; } - - /// merge the element for each col. - virtual void colMerge(Matrix& src) { LOG(FATAL) << "Not implemented"; } - - /// copy -log(output[label]) to this->data[i]. - virtual void oneHotCrossEntropy(Matrix& output, IVector& label) { - LOG(FATAL) << "Not implemented"; - } - - /// calculate the error of outputV according to label. - virtual void oneHotCrossEntropyBp(Matrix& outputV, IVector& label) { - LOG(FATAL) << "Not implemented"; - } - - /// copy -log(output[label]) to this->data[i]. - virtual void oneHotCrossEntropyWithSelfNorm(Matrix& output, - IVector& label, - real alpha) { - LOG(FATAL) << "Not implemented"; - } - - /// calculate the error of outputV according to label. - virtual void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, - IVector& label, - real alpha) { - LOG(FATAL) << "Not implemented"; - } - - /** - * \f[ - * a[i] = \sum_{j=-(N-1)/2}^{(N-1)/2} b_{i+j} * c_{j} - * \f] - * - * b contains M elements, - * c contains N elements (N is odd), - * b's index arithmetic is computed modulo M, - * c's index arithmetic is computed modulo N. - */ - virtual void circularConv(Matrix& b, Matrix& c) { - LOG(FATAL) << "Not implemented"; - } - - virtual void circularConvDerivative(Matrix& output, - Matrix& prevOut1, - Matrix& prevOut2, - Matrix& prevGrad1, - Matrix& prevGrad2) { - LOG(FATAL) << "Not implemented"; - } - - /* output_ij = exp(this_{ij}) / (sum_j exp(this_ij)) */ - virtual void softmax(Matrix& output) { - (void)output; - LOG(FATAL) << "Not implemeted"; - } - virtual void sequenceSoftmax(Matrix& output, const IVector& index) { - (void)output; - LOG(FATAL) << "Not implemeted"; - } - - virtual void softmaxBackward(Matrix& outputV) { - (void)outputV; - LOG(FATAL) << "Not implemeted"; - } - - /* - sum_i = sum_j this_ij * output_ij - this_ij = output_ij* (this_ij - sum_i) - */ - virtual void softmaxDerivative(Matrix& output, Matrix& sftmaxSum) { - LOG(FATAL) << "Not implemented"; - } - - /// calculate the sum of squares diff cost. - virtual void sumOfSquares(Matrix& output, Matrix& label) { - LOG(FATAL) << "Not implemented"; - } - - /// gradient of sumOfSquares. - virtual void sumOfSquaresBp(Matrix& outputV, Matrix& label) { - LOG(FATAL) << "Not implemented"; - } - - virtual void smoothL1(Matrix& output, Matrix& label, real destScale) { - LOG(FATAL) << "Not implemented"; - } - - virtual void smoothL1Bp(Matrix& outputV, Matrix& label, real destScale) { - LOG(FATAL) << "Not implemented"; - } - - virtual void tanh(Matrix& output) { LOG(FATAL) << "Not implemented"; } - - virtual void tanhDerivative(Matrix& output) { - LOG(FATAL) << "Not implemented"; - } - - virtual void softrelu(Matrix& output) { LOG(FATAL) << "Not implemented"; } - - virtual void softreluDerivative(Matrix& output) { - LOG(FATAL) << "Not implemented"; - } - - virtual void scaledTanh(Matrix& output, real p1, real p2) { - LOG(FATAL) << "Not implemented"; - } - - /// print out the values of elements to os - virtual void print(std::ostream& os) const { - LOG(FATAL) << "Not implemented"; - } - - /** - * print a part of the matrix - * from the (top,left) value to the (height, width) value (not included) - */ - virtual void print(std::ostream& os, size_t height, size_t width) const { - LOG(FATAL) << "Not implemented"; - } - - /// print one row to os - virtual void printOneRow(std::ostream& os, size_t idx) const { - LOG(FATAL) << "Not implemented"; - } - - virtual void check(std::ostream& os, Matrix& refMat, bool printDiff = true) {} - - virtual real getMin() { - LOG(FATAL) << "Not implemented"; - return 0; - } - virtual real getMax() { - LOG(FATAL) << "Not implemented"; - return 0; - } - - virtual void randomizeUniform() { LOG(FATAL) << "Not implemented"; } - - /** - * @brief calulate the error of classification - * - * output[i] = 1 if row i is an error. - * - * output[i] = 0 if row i is correct. - * - */ - virtual void classificationError(Matrix& output, - IVector& label, - size_t topkSize = 1) { - LOG(FATAL) << "Not implemented"; - } - - virtual void upsampleForward(Matrix& input, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void upsampleBackward(Matrix& outputGrad, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW) { - LOG(FATAL) << "Not implemeted"; - } - - /** - * Pooling forward operation, pick out the largest element - * in the sizeX of value, if the maskMatP is not NULL, it will - * also caculate the location indices. - */ - virtual void maxPoolForward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - MatrixPtr maskMatP = NULL) { - LOG(FATAL) << "Not implemeted"; - } - - /// Pooling backward operation. - virtual void maxPoolBackward(Matrix& image, - size_t imgSizeH, - size_t imgSizeW, - Matrix& outGrad, - Matrix& outV, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW) { - LOG(FATAL) << "Not implemeted"; - } - - /// Pooling forward operation, caculate the average of sizeX elements. - virtual void avgPoolForward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - bool excludeMode = true) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void avgPoolBackward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW, - bool excludeMode = true) { - LOG(FATAL) << "Not implemeted"; - } - - /** - * Pooling 3D forward operation, pick out the largest element - * in the sizeX of value - */ - virtual void maxPool3DForward(Matrix& inputMat, - Matrix& maxPoolIdx, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void maxPool3DBackward(Matrix& outGrad, - Matrix& maxPoolIdx, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void avgPool3DForward(Matrix& input, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void avgPool3DBackward(Matrix& input, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput) { - LOG(FATAL) << "Not implemeted"; - } - - /** - * Input: one or more sequences. Each sequence contains some instances. - * - * Output: output size is the number of input sequences (NOT input - * instances). - * - * output[i] is set to max_input[i]. - */ - virtual void maxSequenceForward(Matrix& input, - const IVector& sequence, - IVector& index) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void maxSequenceBackward(Matrix& outputGrad, - const IVector& sequence, - IVector& index) { - LOG(FATAL) << "Not implemeted"; - } - - /** - * @code - * this.row[i] += table.row[ids[i]] - * if ids[i] == -1, it will be ignored - * @endcode - */ - virtual void selectRows(Matrix& table, IVector& ids) { - (void)table; - (void)ids; - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * this[i] = table[i, id[i]] - * @endcode - */ - virtual void selectElements(Matrix& table, IVector& ids) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * table.row[ids[i]] += this.row[i] - * if ids[i] == -1, it will be ignored - * @endcode - */ - virtual void addToRows(Matrix& table, IVector& ids) { - (void)table; - (void)ids; - LOG(FATAL) << "Not implemented"; - } - - /** - * @code - * table[i, id[i]] += this[i] - * @endcode - */ - virtual void addElements(Matrix& table, IVector& ids) { - LOG(FATAL) << "Not implemented"; - } - /** - * @brief cross entropy for multi binary labels - * - * @code - * this[i] = -sum(label[i][j]*log(output[i][j]) - * + (1-label[i][j])*log(1-output[i][j])) - * @endcode - */ - virtual void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @brief The gradient of cross entropy for multi binary labels on output - * - * @code - * this[i][j] = -label[i][j]/output[i][j] - * + (1-label[i][j])/(1-output[i][j]) - * @endcode - */ - virtual void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { - LOG(FATAL) << "Not implemented"; - } - - /** - * @brief Calculate the classification error for multi binary labels - * - * @code - * this[i] = sum((output[i][j] >= threshold && label[i][j] == 0) - * || (output[i][j] < threshold && label[i][j] == 1)) - * / output->getWidth() - * @endcode - */ - virtual void classificationErrorMulti(Matrix& output, - Matrix& label, - real threshold) { - LOG(FATAL) << "Not implemented"; - } - - virtual void paramReluForward(Matrix& data, Matrix& W) { - LOG(FATAL) << "Not implemented"; - } - virtual void paramReluBackwardW(Matrix& oGrad, Matrix& data) { - LOG(FATAL) << "Not implemented"; - } - virtual void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) { - LOG(FATAL) << "Not implemented"; - } - - virtual void vol2Col(real* data, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void col2Vol(real* trg, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - real alpha, - real beta) { - LOG(FATAL) << "Not implemeted"; - } - - virtual void bilinearForward(const Matrix& in, - const size_t inImgH, - const size_t inImgW, - const size_t outImgH, - const size_t outImgW, - const size_t numChannels, - const real ratioH, - const real ratioW) { - LOG(FATAL) << "Not implemented"; - } - virtual void bilinearBackward(const Matrix& out, - const size_t outImgH, - const size_t outImgW, - const size_t inImgH, - const size_t inImgW, - const size_t numChannels, - const real ratioH, - const real ratioW) { - LOG(FATAL) << "Not implemented"; - } - - template - void operator=(const ExpressionType& expr) { - if (useGpu_) { - TensorGpuApply(*this, expr); - } else { - TensorCpuApply(*this, expr); - } - } - - bool isEmpty() const { return data_ == nullptr; } - - explicit operator bool() const { return !isEmpty(); } -}; - -inline std::ostream& operator<<(std::ostream& os, const Matrix& mat) { - mat.print(os); - return os; -} - -class GpuMatrix : public Matrix { - public: - GpuMatrix(); - - GpuMatrix(size_t height, size_t width, bool trans = false); - GpuMatrix(real* data, size_t height, size_t width, bool trans = false) - : Matrix(data, height, width, trans, true) {} - GpuMatrix(real* data, - size_t height, - size_t width, - size_t stride, - bool trans = false) - : Matrix(data, height, width, stride, trans, true) {} - GpuMatrix(GpuMemHandlePtr dataHandle, - size_t height, - size_t width, - bool trans = false) - : Matrix(dataHandle, height, width, trans, true) {} - ~GpuMatrix(); - - void zeroMem(); - void resetOne(); - void setDiag(real value); - - void resize(size_t newHeight, size_t newWidth); - void resize(size_t newHeight, - size_t newWidth, - size_t newNnz, /* used to allocate space */ - SparseValueType valueType, - SparseFormat format) { - LOG(FATAL) << "Only Support Sparse Matrix"; - } - void setRow(size_t row, - size_t colNum, - const unsigned int* cols, - const real* values) { - LOG(FATAL) << "Only Support Sparse Matrix"; - } - - /** - * Copy the data from cpu_memory buffer - */ - void copyFrom(const real* hostSrc, size_t size); - - void copyFrom(const real* hostSrc, const int64_t* seq); - - void copyFrom(const Matrix& src, hl_stream_t stream); - - void copyFrom(const Matrix& src); - - void copyFrom(const IVector& src); - - void copyByRowIndex(Matrix& b, const IVector& rowIndex); - - MatrixPtr clone(size_t height, size_t width, bool useGpu = false); - - real getElement(size_t x, size_t y) const; - - real* getRow(size_t row) { return BaseMatrix::rowBuf(row); } - virtual real* getRowBuf(size_t row) { return getRow(row); } - - real getSum(); - void accumulateColSum(Matrix& src); - real getAbsSum(); - - real getMin(); - real getMax(); - - MatrixPtr getTranspose(); - void transpose(MatrixPtr& matTrans, bool memAlloc); - void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise); - - MatrixPtr getInverse(); - void inverse(MatrixPtr& matInv, bool memAlloc); - - /// add b to each sample of this. - void addBias(Matrix& b, real scale); - void addSharedBias(Matrix& b, real scale); - - /** - * @code - * add each sample from a to this. - * @endcode - */ - void collectBias(Matrix& a, real scale); - void collectSharedBias(Matrix& a, real scale); - - void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode); - void sequenceAvgBackward(Matrix& a, const IVector& startsPos, int mode); - - /** - * @code - * this.row[i] += table.row[ids[i]] - * @endcode - */ - virtual void selectRows(Matrix& table, IVector& ids); - - /** - * @code - * this[i] = table[i, id[i]] - * @endcode - */ - virtual void selectElements(Matrix& table, IVector& ids); - - /** - * @code - * table.row[ids[i]] += this.row[i] - * @endcode - */ - virtual void addToRows(Matrix& table, IVector& ids); - - void addColumnVector(const Matrix& b); - - /** - * @code - * this = scaleAB*(a*b) + scaleT*this - * @endcode - */ - void mul(const Matrix& a, const Matrix& b, real scaleAB, real scaleT); - - /** - * @code - * this = a*b - * @endcode - */ - void mul(const Matrix& a, const Matrix& b); - - void mul(const GpuMatrix& a, const GpuMatrix& b, real scaleAB, real scaleT); - - void mul(const GpuSparseMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT); - - void mul(const GpuMatrix& a, - const GpuSparseMatrix& b, - real scaleAB, - real scaleT); - - /** - * @code - * this = scaleAB*(this*b) + scaleT*this - * @endcode - */ - void rightMul(Matrix& b, real scaleAB, real scaleT); - - /** - * @code - * this = this* b - * @endcode - */ - void rightMul(Matrix& b); - - /** - * @code - * this = scaleAB*(a*this) + scaleT*this - * @endcode - */ - void leftMul(Matrix& a, real scaleAB, real scaleT); - - /** - * @code - * this = a*this - * @endcode - */ - void leftMul(Matrix& a); - - void colMerge(Matrix& src); - void rowSum(Matrix& sum); - void rowMax(Matrix& max); - void rowMax(IVector& maxIds, Matrix& max); - void colMax(Matrix& max); - void colMax(IVector& maxIds, Matrix& max); - void maxoutForward(Matrix& a, IVector& id, size_t channels, size_t groups); - void maxoutBackward(Matrix& a, IVector& id, size_t channels, size_t groups); - - void oneHotCrossEntropy(Matrix& output, IVector& label); - void oneHotCrossEntropyBp(Matrix& outputV, IVector& label); - void oneHotCrossEntropyWithSelfNorm(Matrix& output, - IVector& label, - real alpha); - void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, - IVector& label, - real alpha); - - void softmax(Matrix& output); - void sequenceSoftmax(Matrix& output, const IVector& index); - void softmaxBackward(Matrix& outputV); - void softmaxDerivative(Matrix& output, Matrix& sftmaxSum); - - /// calculate the sum of squares diff cost. - void sumOfSquares(Matrix& output, Matrix& label); - - /// gradient of sumOfSquares. - void sumOfSquaresBp(Matrix& outputV, Matrix& label); - void tanh(Matrix& output); - void tanhDerivative(Matrix& output); - void softrelu(Matrix& output); - void softreluDerivative(Matrix& output); - void scaledTanh(Matrix& output, real p1, real p2); - - virtual void print(std::ostream& os) const; - virtual void print(std::ostream& os, size_t height, size_t width) const; - - void paramReluForward(Matrix& data, Matrix& W); - void paramReluBackwardW(Matrix& oGrad, Matrix& data); - void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W); - - void check(std::ostream& os, Matrix& refMat, bool printDiff = true); - void randomizeUniform(); - - void classificationError(Matrix& output, IVector& label, size_t topkSize = 1); - - void upsampleForward(Matrix& input, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW); - - void upsampleBackward(Matrix& outputGrad, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW); - - void maxPoolForward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - MatrixPtr maskMatP); - - void maxPoolBackward(Matrix& image, - size_t imgSizeH, - size_t imgSizeW, - Matrix& outGrad, - Matrix& outV, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW); - - void avgPoolForward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - bool excludeMode = true); - - void avgPoolBackward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW, - bool excludeMode = true); - - void maxPool3DForward(Matrix& inputMat, - Matrix& maxPoolIdx, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW); - - void maxPool3DBackward(Matrix& outGrad, - Matrix& maxPoolIdx, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput); - - void avgPool3DForward(Matrix& input, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW); - - void avgPool3DBackward(Matrix& input, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput); - - void maxSequenceForward(Matrix& input, - const IVector& sequence, - IVector& index); - - void maxSequenceBackward(Matrix& outputGrad, - const IVector& sequence, - IVector& index); - - void bilinearForward(const Matrix& in, - const size_t inImgH, - const size_t inImgW, - const size_t outImgH, - const size_t outImgW, - const size_t numChannels, - const real ratioH, - const real ratioW); - - void bilinearBackward(const Matrix& out, - const size_t outImgH, - const size_t outImgW, - const size_t inImgH, - const size_t inImgW, - const size_t numChannels, - const real ratioH, - const real ratioW); - - void vol2Col(real* data, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW); - - void col2Vol(real* trg, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - real alpha, - real beta); - - void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label); - - void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label); - - template - void operator=(const ExpressionType& expr) { - TensorGpuApply(*this, expr); - } -}; - -class CpuMatrix : public Matrix { - private: - MatrixPtr sftmaxSum_; - MatrixPtr sftmaxDot_; - - public: - CpuMatrix(size_t height, size_t width, bool trans = false); - CpuMatrix(real* data, size_t height, size_t width, bool trans = false) - : Matrix(data, height, width, trans, false) {} - CpuMatrix(real* data, - size_t height, - size_t width, - size_t stride, - bool trans = false) - : Matrix(data, height, width, stride, trans, false) {} - - CpuMatrix(CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - bool trans = false) - : Matrix(dataHandle, height, width, trans, false) {} - - ~CpuMatrix(); - - void zeroMem(); - void resetOne(); - void setDiag(real value); - - void resize(size_t newHeight, size_t newWidth); - void resize(size_t newHeight, - size_t newWidth, - size_t newNnz, /* used to allocate space */ - SparseValueType valueType, - SparseFormat format) { - LOG(FATAL) << "Only Support Sparse Matrix"; - } - void setRow(size_t row, - size_t colNum, - const unsigned int* cols, - const real* values) { - LOG(FATAL) << "Only Support Sparse Matrix"; - } - - real getElement(size_t x, size_t y) const; - real getSum(); - void accumulateColSum(Matrix& src); - real getAbsSum(); - - MatrixPtr getTranspose(); - void transpose(MatrixPtr& matTrans, bool memAlloc); - void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise); - - MatrixPtr getInverse(); - void inverse(MatrixPtr& matInv, bool memAlloc); - - void copyFrom(const Matrix& src); - - void copyFrom(const Matrix& src, hl_stream_t stream); - - void copyFrom(const real* cpuSrc, size_t size); - - void copyFrom(const real* cpuSrc, const int64_t* seq); - - void copyFrom(const IVector& src); - - void copyFrom(CpuSparseMatrix& src); - - void copyByRowIndex(Matrix& b, const IVector& rowIndex); - - MatrixPtr clone(size_t height, size_t width, bool useGpu = false); - - void upsampleForward(Matrix& input, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW); - - void upsampleBackward(Matrix& outputGrad, - Matrix& mask, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t outputH, - size_t outputW); - - void maxPoolForward(Matrix& inputMat, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - MatrixPtr maskMatP); - - void maxPoolBackward(Matrix& image, - size_t imgSizeH, - size_t imgSizeW, - Matrix& outGrad, - Matrix& outV, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW); - - void avgPoolForward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t channels, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - size_t paddingH, - size_t paddingW, - bool excludeMode = true); - - void avgPoolBackward(Matrix& input, - size_t imgSizeH, - size_t imgSizeW, - size_t sizeX, - size_t sizeY, - size_t strideH, - size_t strideW, - size_t outputH, - size_t outputW, - real scaleTargets, - real scaleOutput, - size_t paddingH, - size_t paddingW, - bool excludeMode = true); - - void maxPool3DForward(Matrix& inputMat, - Matrix& maxPoolIdx, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW); - - void maxPool3DBackward(Matrix& outGrad, - Matrix& maxPoolIdx, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput); - - void avgPool3DForward(Matrix& input, - size_t channels, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW); - - void avgPool3DBackward(Matrix& input, - size_t imgSizeD, - size_t imgSizeH, - size_t imgSizeW, - size_t outputD, - size_t outputH, - size_t outputW, - size_t sizeZ, - size_t sizeY, - size_t sizeX, - size_t strideD, - size_t strideH, - size_t strideW, - size_t paddingD, - size_t paddingH, - size_t paddingW, - real scaleTargets, - real scaleOutput); - - void maxSequenceForward(Matrix& input, - const IVector& sequence, - IVector& index); - - void maxSequenceBackward(Matrix& outputGrad, - const IVector& sequence, - IVector& index); - - real* getRow(size_t row) { return BaseMatrix::rowBuf(row); } - virtual real* getRowBuf(size_t row) { return getRow(row); } - - public: - /// add b to each sample of this. - void addBias(Matrix& b, real scale); - void addSharedBias(Matrix& b, real scale); - - /// add each sample of a to this. - void collectBias(Matrix& a, real scale); - void collectSharedBias(Matrix& a, real scale); - - void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode); - void sequenceAvgBackward(Matrix& a, const IVector& startsPos, int mode); - - /** - * @code - * this.row[i] += table.row[ids[i]] - * @endcode - */ - virtual void selectRows(Matrix& table, IVector& ids); - - /** - * @code - * table.row[ids[i]] += this.row[i] - * @endcode - */ - virtual void addToRows(Matrix& table, IVector& ids); - - /** - * @code - * this[i] = table[i, id[i]] - * @endcode - */ - virtual void selectElements(Matrix& table, IVector& ids); - - /** - * @code - * table[i, id[i]] += this[i] - * @endcode - */ - virtual void addElements(Matrix& table, IVector& ids); - - /** - * use abstract getRow() to get row from table. - * - * Define table as template instead of virtual class for performance sake. - * internal used by above two virtual funcs. - */ - template - void selectRowsImp(TableMatType& table, IVector& ids); - template - void addToRowsImp(TableMatType& table, IVector& ids); - - void addColumnVector(const Matrix& b); - - void mul(const Matrix& a, const Matrix& b, real scaleAB, real scaleT); - void mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); - - void mul(CpuMatrix* a, CpuSparseMatrix* b, real scaleAB, real scaleT); - - static void mul(CpuMatrix* a, - CpuMatrix* b, - CpuSparseMatrix* c, - real scaleAB, - real scaleT); - - /** - * c = a * b - * - * use abstract getRow() to get row from B,C. - * Define B,C as template instead of virtual class for performance sake. - */ - template - static void mul( - CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB, real scaleT); - - virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); - - void mul(const Matrix& a, const Matrix& b); - - void rightMul(Matrix& b, real scaleAB, real scaleT); - void rightMul(Matrix& b); - - void leftMul(Matrix& a, real scaleAB, real scaleT); - void leftMul(Matrix& a); - void colMerge(Matrix& src); - void rowSum(Matrix& sum); - void rowMaxId(IVector& maxIds); - void rowMax(Matrix& max); - void rowMax(IVector& maxIds, Matrix& maxVal); - void colMax(Matrix& max); - void colMax(IVector& maxIds, Matrix& maxVal); - void maxoutForward(Matrix& a, IVector& id, size_t channels, size_t groups); - void maxoutBackward(Matrix& a, IVector& id, size_t channels, size_t groups); - void rowNormalizeL1(Matrix& out); - - void oneHotCrossEntropy(Matrix& output, IVector& label); - void oneHotCrossEntropyBp(Matrix& outputV, IVector& label); - void oneHotCrossEntropyWithSelfNorm(Matrix& output, - IVector& label, - real alpha); - void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, - IVector& label, - real alpha); - - void circularConv(Matrix& b, Matrix& c); - void circularConvDerivative(Matrix& output, - Matrix& prevOut1, - Matrix& prevOut2, - Matrix& prevGrad1, - Matrix& prevGrad2); - - void softmax(Matrix& output); - void sequenceSoftmax(Matrix& output, const IVector& index); - void softmaxDerivative(Matrix& output, Matrix& sftmaxSum); - - /// calculate the sum of squares diff cost. - void sumOfSquares(Matrix& output, Matrix& label); - - /// gradient of sumOfSquares. - void sumOfSquaresBp(Matrix& outputV, Matrix& label); - - void smoothL1(Matrix& output, Matrix& label, real destScale); - void smoothL1Bp(Matrix& output, Matrix& label, real destScale); - - void tanh(Matrix& output); - void tanhDerivative(Matrix& output); - - void softrelu(Matrix& output); - void softreluDerivative(Matrix& output); - void scaledTanh(Matrix& output, real p1, real p2); - - void print(std::ostream& os) const; - void print(std::ostream& os, size_t height, size_t width) const; - void printOneRow(std::ostream& os, size_t idx) const; - - void paramReluForward(Matrix& data, Matrix& W); - void paramReluBackwardW(Matrix& oGrad, Matrix& data); - void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W); - - void check(std::ostream& os, Matrix& refMat, bool printDiff = true); - - real getMin(); - real getMax(); - - void randomizeUniform(); - - void classificationError(Matrix& output, IVector& label, size_t topkSize = 1); - - void addByBitCode(size_t numClasses, const IVector& codes, const Matrix& vec); - - void addByBitCodeBackward(size_t numClasses, - const IVector& codes, - Matrix& vec); - - void mulByBitCode(size_t numClasses, - const IVector& codes, - const Matrix& mat, - const Matrix& input); - - void mulByBitCodeBackwardWeight(size_t numClasses, - const IVector& codes, - Matrix& mat, - const Matrix& input); - - void mulByBitCodeBackwardError(size_t numClasses, - const IVector& codes, - const Matrix& mat, - Matrix& input); - - void sumByBitCode(size_t numClasses, - IVector& codes, - Matrix& sum, - real scaleSum); - - void subByBitCode(size_t numClasses_, IVector& codes); - - void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label); - void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label); - void classificationErrorMulti(Matrix& output, Matrix& label, real threshold); - - void bilinearForward(const Matrix& in, - const size_t inImgH, - const size_t inImgW, - const size_t outImgH, - const size_t outImgW, - const size_t numChannels, - const real ratioH, - const real ratioW); - - void bilinearBackward(const Matrix& out, - const size_t outImgH, - const size_t outImgW, - const size_t inImgH, - const size_t inImgW, - const size_t numChannels, - const real ratioH, - const real ratioW); - - void vol2Col(real* data, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW); - - void col2Vol(real* trg, - int channels, - int depth, - int height, - int width, - int filterD, - int filterH, - int filterW, - int strideD, - int strideH, - int strideW, - int paddingD, - int paddingH, - int paddingW, - real alpha, - real beta); - - template - void operator=(const ExpressionType& expr) { - TensorCpuApply(*this, expr); - } -}; - -class SharedCpuMatrix : public CpuMatrix { - public: -#ifndef PADDLE_MOBILE_INFERENCE - /* blockNum is number of partitions of the matrix */ - SharedCpuMatrix(int blockNum, size_t height, size_t width, bool trans = false) - : CpuMatrix(height, width, trans) { - initShared(blockNum); - } - SharedCpuMatrix( - int blockNum, real* data, size_t height, size_t width, bool trans = false) - : CpuMatrix(data, height, width, trans) { - initShared(blockNum); - } - - SharedCpuMatrix(int blockNum, - CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - bool trans = false) - : CpuMatrix(dataHandle, height, width, trans) { - initShared(blockNum); - } - - SharedCpuMatrix(CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - bool trans = false) - : CpuMatrix(dataHandle, height, width, trans) { - initBlock(1); - } - - ~SharedCpuMatrix() {} - - public: - virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); - virtual void add(Matrix& b, real p1, real p2); - virtual void add(real p1, real p2); - - private: - using Matrix::mul; - void initShared(int blockNum); - void initBlock(int blockNum); - - int blockNum_; - std::vector> blockLocks_; - ThreadLocal localBuf_; - ThreadLocal> localBufRows_; - ThreadLocal> blockSeq_; -#endif -}; - -typedef struct { unsigned int col; } sparse_non_value_t; - -typedef struct { - unsigned int col; - float value; -} sparse_float_value_t; - -} // namespace paddle -#include "ExecViaCpu.h" diff --git a/paddle/math/MatrixBitCode.cpp b/paddle/math/MatrixBitCode.cpp deleted file mode 100644 index f7a949294b54a5a874e1239a13ca9dce3ba18e94..0000000000000000000000000000000000000000 --- a/paddle/math/MatrixBitCode.cpp +++ /dev/null @@ -1,291 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Matrix.h" -#include "hl_gpu.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -namespace { - -struct SimpleCode { - SimpleCode(size_t code, size_t numClasses) : c_(code + numClasses) {} - inline size_t calcIndex(int bit) const { return (c_ >> (bit + 1)) - 1; } - inline bool calcBit(int bit) const { return c_ & (1 << bit); } - inline int getLength() const { return findLastSet(c_) - 1; } - - private: - size_t c_; -}; - -struct SimpleCodeTable { - explicit SimpleCodeTable(size_t numClasses) : numClasses_(numClasses) {} - SimpleCode operator()(size_t code) const { - return SimpleCode(code, numClasses_); - } - size_t size() const { return numClasses_; } - int getMaxCodeLength() const { return findLastSet(numClasses_ - 1); } - - private: - size_t numClasses_; - int maxCodeLength_; -}; - -} // namespace - -/** - * CodeTable class should support 3 functions: - * - * size_t size() - * return the number of codes - * - * int getMaxCodeLength() - * return the maximal code length - * - * Code operator()(size_t i) - * return the i-th code. Code class is descriebed below. - * - * Code class should support 3 functions: - * - * int getLength() - * return the length of the code - * - * bool calcIndex(int bit) - * bit ranges from 0 to getLength() - 1 - * return the index for the (1+bit) level parent - * - * bool calcBit(int bit) - * return true if the bit level parent is the right child of (1+bit) level - * parent - * - */ - -/* - for i: - for j < codeLength: - op(tmat(i, j), vec(0, index(i, j))) -*/ -template -static void addByBitCodeT( - Op op, CodeTable codeTable, const IVector& codes, TMat& tmat, Mat& vec) { - CHECK(!vec.useGpu()); - - size_t numClasses = codeTable.size(); - size_t maxCodeLength = codeTable.getMaxCodeLength(); - size_t numSamples = tmat.getHeight(); - size_t oWidth = tmat.getWidth(); - CHECK_EQ(tmat.getWidth(), maxCodeLength); - CHECK_EQ(codes.getSize(), numSamples); - CHECK_EQ(vec.getHeight(), (size_t)1); - CHECK_EQ(vec.getWidth(), numClasses - 1); - - auto data = tmat.getData(); - auto v = vec.getData(); - const int* c = codes.getData(); - for (size_t i = 0; i < numSamples; ++i) { - auto code = codeTable(c[i]); - int codeLength = code.getLength(); - for (int j = 0; j < codeLength; ++j) { - size_t index = code.calcIndex(j); - op(data[i * oWidth + j], v[index]); - } - } -} - -/* For j < codeLength: - this(i, j) += vec(0, index(i, j)) -*/ -void CpuMatrix::addByBitCode(size_t numClasses, - const IVector& codes, - const Matrix& vec) { - auto op = [](real& t, real v) { t += v; }; - addByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, vec); -} - -/* For j < codeLength: - vec(0, index(i, j)) += this(i, j) -*/ -void CpuMatrix::addByBitCodeBackward(size_t numClasses, - const IVector& codes, - Matrix& vec) { - auto op = [](real t, real& v) { v += t; }; - addByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, vec); -} - -/* - for i: - for j < codeLength: - op(tmat(i, j), mat.row(index(i, j)), input.row(i)) -*/ -template -void mulByBitCodeT(Op op, - CodeTable codeTable, - IVec& codes, - TMat& tmat, - WMat& weight, - InMat& input) { - CHECK(!tmat.useGpu() && !weight.useGpu() && !input.useGpu()); - - size_t numClasses = codeTable.size(); - size_t maxCodeLength = codeTable.getMaxCodeLength(); - size_t numSamples = tmat.getHeight(); - size_t inputDim = input.getWidth(); - size_t oWidth = tmat.getWidth(); - CHECK_EQ(tmat.getWidth(), maxCodeLength); - CHECK_EQ(codes.getSize(), numSamples); - CHECK_EQ(input.getHeight(), numSamples); - CHECK_EQ(weight.getHeight(), numClasses - 1); - CHECK_EQ(weight.getWidth(), inputDim); - - real* data = tmat.getData(); - const int* c = codes.getData(); - for (size_t i = 0; i < numSamples; ++i) { - auto code = codeTable(c[i]); - int codeLength = code.getLength(); - for (int j = 0; j < codeLength; ++j) { - size_t index = code.calcIndex(j); - op(data[i * oWidth + j], weight.rowBuf(index), input.rowBuf(i), inputDim); - } - } -} - -/* For j < codeLength: - this(i, j) += -*/ -void CpuMatrix::mulByBitCode(size_t numClasses, - const IVector& codes, - const Matrix& weight, - const Matrix& input) { - auto op = []( - real& t, const real* weightRow, const real* inputRow, size_t inputDim) { - real sum = 0; - for (size_t k = 0; k < inputDim; ++k) { - sum += weightRow[k] * inputRow[k]; - } - t += sum; - }; - - mulByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, weight, input); -} - -/* For index(i, j) >= 0: - weight.row(index(i, j)) += this(i, j) * input.row(i) -*/ -void CpuMatrix::mulByBitCodeBackwardWeight(size_t numClasses, - const IVector& codes, - Matrix& weight, - const Matrix& input) { - auto op = []( - const real t, real* weightRow, const real* inputRow, size_t inputDim) { - for (size_t k = 0; k < inputDim; ++k) { - weightRow[k] += t * inputRow[k]; - } - }; - - mulByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, weight, input); -} - -/* For j < codeLength: - input.row(i) += this(i, j) * weight.row(index(i, j)) -*/ -void CpuMatrix::mulByBitCodeBackwardError(size_t numClasses, - const IVector& codes, - const Matrix& weight, - Matrix& input) { - auto op = []( - const real t, const real* weightRow, real* inputRow, size_t inputDim) { - for (size_t k = 0; k < inputDim; ++k) { - inputRow[k] += t * weightRow[k]; - } - }; - - mulByBitCodeT(op, SimpleCodeTable(numClasses), codes, *this, weight, input); -} - -template -void sumByBitCodeT(CodeTable codeTable, - IVector& codes, - const CpuMatrix& tmat, - Matrix& sum, - real scaleSum) { - size_t maxCodeLength = codeTable.getMaxCodeLength(); - size_t numSamples = tmat.getHeight(); - size_t oWidth = tmat.getWidth(); - CHECK_EQ(tmat.getWidth(), maxCodeLength); - CHECK_EQ(codes.getSize(), numSamples); - CHECK_EQ(sum.getHeight(), numSamples); - CHECK_EQ(sum.getWidth(), (size_t)1); - - const real* data = tmat.getData(); - real* s = sum.getData(); - int* c = codes.getData(); - for (size_t i = 0; i < numSamples; ++i) { - real sm = 0; - auto code = codeTable(c[i]); - int codeLength = code.getLength(); - for (int j = 0; j < codeLength; ++j) { - if (code.calcBit(j)) { - sm += data[i * oWidth + j]; - } - } - s[i] = scaleSum * sm; - } -} - -/* For j < codeLength: - sum(i, 0) = \sum_j bit(i, j) * this(i, j) -*/ -void CpuMatrix::sumByBitCode(size_t numClasses, - IVector& codes, - Matrix& sum, - real scaleSum) { - sumByBitCodeT(SimpleCodeTable(numClasses), codes, *this, sum, scaleSum); -} - -template -void subByBitCodeT(CodeTable codeTable, IVector& codes, CpuMatrix& tmat) { - size_t maxCodeLength = codeTable.getMaxCodeLength(); - size_t numSamples = tmat.getHeight(); - size_t oWidth = tmat.getWidth(); - CHECK_EQ(tmat.getWidth(), maxCodeLength); - CHECK_EQ(codes.getSize(), numSamples); - - real* data = tmat.getData(); - int* c = codes.getData(); - for (size_t i = 0; i < numSamples; ++i) { - auto code = codeTable(c[i]); - int codeLength = code.getLength(); - for (int j = 0; j < codeLength; ++j) { - if (code.calcBit(j)) { - data[i * oWidth + j] -= 1; - } - } - } -} - -/* For j < codeLength - this(i, j) -= bit(i, j) -*/ -void CpuMatrix::subByBitCode(size_t numClasses, IVector& codes) { - subByBitCodeT(SimpleCodeTable(numClasses), codes, *this); -} - -} // namespace paddle diff --git a/paddle/math/RowBuffer.h b/paddle/math/RowBuffer.h deleted file mode 100644 index 6950afaa21d60615b27c06a151b0afbb296653bf..0000000000000000000000000000000000000000 --- a/paddle/math/RowBuffer.h +++ /dev/null @@ -1,139 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include "MemoryHandle.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -/** - * @brief The RowBuffer class - * Represent the SparseRow Matrix Data. - * - * If not set memory handler, then the data could be auto growth. - */ -class RowBuffer { - public: - /** - * @brief RowBuffer create a auto-growth row buffer. The row length is width. - * @param width the length of each row, a.k.a matrix width. - */ - explicit RowBuffer(size_t width) : width_(width) {} - - /** - * @brief RowBuffer create a row buffer, which cannot be auto-growth. - * @param mem the pre-allocated memory. - * @param width the length of each row, a.k.a matrix width. - */ - RowBuffer(const CpuMemHandlePtr& mem, size_t width) - : preallocatedBuf_(mem), width_(width) {} - - /** - * @brief resize resize the buffer with rowCount - * @param rowCnt number of row. matrix height. - */ - inline void resize(int rowCnt) { - if (preallocatedBuf_) { - CHECK(preallocatedBuf_->getSize() >= rowCnt * width_ * sizeof(real)); - } else { - rowStore_.resize(rowCnt * width_); - } - } - - /** - * @brief get a row buffer with row index. - * @param row the index of row. - * @return row buffer. - */ - inline real* get(int row) const { - if (preallocatedBuf_) { - CHECK_LE((row)*width_ * sizeof(real), preallocatedBuf_->getSize()); - return reinterpret_cast(preallocatedBuf_->getBuf()) + row * width_; - } else { - CHECK_LE((row + 1) * width_, rowStore_.size()); - return const_cast(rowStore_.data() + row * width_); - } - } - - /** - * @brief get a row buffer with row index. If row index is larger than local - * buffer, the size of local buffer will grow. - * @param row the index of row. - * @return row buffer. - */ - inline real* getWithAutoGrowth(int row) { - if (preallocatedBuf_) { - return get(row); - } else { - if ((rowStore_.size() <= row * width_)) { - rowStore_.resize((row + 1) * width_); - } - return rowStore_.data() + row * width_; - } - } - - /** - * @return raw data buffer. - */ - inline real* data() { - if (preallocatedBuf_) { - return reinterpret_cast(preallocatedBuf_->getBuf()); - } else { - return rowStore_.data(); - } - } - - /** - * @brief clear local buffer. It only affect auto-growth buffer. - */ - inline void clear() { - // swap an empty vector to it to free the memory. - std::vector> empty; - rowStore_.swap(empty); - } - - /** - * @brief get current number of rows. - * @return number of rows. - */ - inline size_t getRowCount() const { - if (preallocatedBuf_) { - return preallocatedBuf_->getSize() / sizeof(real) / width_; - } else { - return rowStore_.size() / width_; - } - } - - /** - * @brief get is this buffer can automatically grow or not. - * @return ture if can automacitally grow. - */ - inline bool isAutoGrowth() const { return !preallocatedBuf_; } - - /** - * @brief return the width of matrix. a.k.a length of row. - * @return width of matrix - */ - inline size_t getWidth() const { return width_; } - - private: - //! TODO(yuyang18): Add resize method to CpuMemHandlePtr, then we can get rid - //! of std::vector here. - CpuMemHandlePtr preallocatedBuf_; - std::vector> rowStore_; - size_t width_; -}; -} // namespace paddle diff --git a/paddle/math/SparseMatrix.cpp b/paddle/math/SparseMatrix.cpp deleted file mode 100644 index 1faa343dbcef3d20b29b272a8da37f8e2bba654b..0000000000000000000000000000000000000000 --- a/paddle/math/SparseMatrix.cpp +++ /dev/null @@ -1,864 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SparseMatrix.h" -#include -#include -#include -#include "hl_gpu.h" -#include "hl_top_k.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -GpuSparseMatrix::GpuSparseMatrix(size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans) - : Matrix(NULL, height, width, trans, true) { - resize(height, width, nnz, valueType, format); -} - -GpuSparseMatrix::GpuSparseMatrix(GpuMemHandlePtr dataHandle, - hl_sparse_matrix_s_ptr sMatrix, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans, - MemoryHandlePtr sMemoryHandle) - : Matrix(dataHandle, height, width, trans, true) { - CHECK(dataHandle && sMatrix) << "Invalid argument pointer"; - - size_t size = 0; - if (format == SPARSE_CSR) { - size = (height + 1) * sizeof(int) + nnz * sizeof(int); - } else { - size = (width + 1) * sizeof(int) + nnz * sizeof(int); - } - - if (NO_VALUE != valueType) { - size += nnz * sizeof(real); - } - CHECK_LE(size, dataHandle->getSize()); - - sMatrix_ = sMatrix; - - if (sMemoryHandle == NULL) { - sMemoryHandle_ = std::make_shared(dataHandle->getSize()); - } else { - CHECK_EQ(sMemoryHandle->getSize(), dataHandle->getSize()); - sMemoryHandle_ = sMemoryHandle; - } - - elementCnt_ = nnz; - valueType_ = valueType; - format_ = format; - if (format_ == SPARSE_CSR) - sparseResizeCSR(); - else - sparseResizeCSC(); -} - -GpuSparseMatrix::GpuSparseMatrix(hl_sparse_matrix_s_ptr sMatrix, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans, - MemoryHandlePtr sMemoryHandle) - : Matrix(NULL, height, width, trans, true) { - CHECK(sMatrix) << "Invalid argument pointer"; - sMatrix_ = sMatrix; - sMemoryHandle_ = sMemoryHandle; - elementCnt_ = nnz; - format_ = format; - valueType_ = valueType; -} - -GpuSparseMatrix::GpuSparseMatrix(real* value, - int* rows, - int* cols, - size_t height, - size_t width, - size_t nnz, - SparseValueType valueType, - SparseFormat format, - bool trans) - : Matrix(NULL, height, width, trans, true) { - size_t size = 0; - if (format == SPARSE_CSR) { - size = (height + 1) * sizeof(int) + nnz * sizeof(int); - } else { - size = (width + 1) * sizeof(int) + nnz * sizeof(int); - } - - if (NO_VALUE != valueType) { - size += nnz * sizeof(real); - } - elementCnt_ = nnz; - valueType_ = valueType; - format_ = format; - - sMemoryHandle_ = std::make_shared(size); - if (format_ == SPARSE_CSR) { - rows_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf())); - cols_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf()) + - (height_ + 1) * sizeof(int)); - if (NO_VALUE != valueType_) { - value_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf()) + - (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); - } else { - value_ = NULL; - } - - if (sMatrix_ == NULL) { - /* construct hl_sparse_matrix_s */ - hl_sparse_matrix_s tmp; - hl_construct_sparse_matrix( - &tmp, - value, - rows, - cols, - HL_SPARSE_CSR, - valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, - height_, - width_, - elementCnt_); - hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); - sMatrix_ = tmp2; - } - - } else { - cols_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf())); - rows_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf()) + - (width_ + 1) * sizeof(int)); - if (NO_VALUE != valueType_) { - value_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf()) + - (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); - } else { - value_ = NULL; - } - - if (sMatrix_ == NULL) { - /* construct hl_sparse_matrix_s */ - hl_sparse_matrix_s tmp; - hl_construct_sparse_matrix( - &tmp, - value, - rows, - cols, - HL_SPARSE_CSC, - valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, - height_, - width_, - elementCnt_); - hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); - sMatrix_ = tmp2; - } - } -} - -void GpuSparseMatrix::sparseResizeCSR() { - rows_ = - reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf())); - cols_ = - reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf()) + - (height_ + 1) * sizeof(int)); - if (NO_VALUE != valueType_) { - value_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf()) + - (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); - } else { - value_ = NULL; - } - - if (sMatrix_ == NULL) { - /* construct hl_sparse_matrix_s */ - hl_sparse_matrix_s tmp; - hl_construct_sparse_matrix( - &tmp, - data_, - memoryHandle_->getSize(), - HL_SPARSE_CSR, - valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, - height_, - width_, - elementCnt_); - hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); - sMatrix_ = tmp2; - } -} - -void GpuSparseMatrix::sparseResizeCSC() { - cols_ = - reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf())); - rows_ = - reinterpret_cast(reinterpret_cast(sMemoryHandle_->getBuf()) + - (width_ + 1) * sizeof(int)); - if (NO_VALUE != valueType_) { - value_ = reinterpret_cast( - reinterpret_cast(sMemoryHandle_->getBuf()) + - (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int)); - } else { - value_ = NULL; - } - - if (sMatrix_ == NULL) { - /* construct hl_sparse_matrix_s */ - hl_sparse_matrix_s tmp; - hl_construct_sparse_matrix( - &tmp, - memoryHandle_->getBuf(), - memoryHandle_->getSize(), - HL_SPARSE_CSC, - valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE, - height_, - width_, - elementCnt_); - hl_sparse_matrix_s_ptr tmp2(tmp, hl_destruct_sparse_matrix); - sMatrix_ = tmp2; - } -} - -void GpuSparseMatrix::resize(size_t newHeight, - size_t newWidth, - size_t newNnz, - SparseValueType valueType, - SparseFormat format) { - if (format == SPARSE_CSR) { - resizeCSR(newHeight, newWidth, newNnz, valueType); - } else { - resizeCSC(newHeight, newWidth, newNnz, valueType); - } -} - -void GpuSparseMatrix::resizeCSR(size_t newHeight, - size_t newWidth, - size_t newNnz, - SparseValueType valueType) { - size_t newSize = (newHeight + 1) * sizeof(int) + newNnz * sizeof(int); - if (NO_VALUE != valueType) { - newSize += newNnz * sizeof(real); - } - - if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) { - memoryHandle_ = std::make_shared(newSize); - data_ = reinterpret_cast(memoryHandle_->getBuf()); - sMemoryHandle_ = std::make_shared(newSize); - end_ = reinterpret_cast(sMemoryHandle_->getBuf()) + - sMemoryHandle_->getSize(); - sMatrix_ = NULL; - } else if (valueType != valueType_) { - sMatrix_ = NULL; - } else { - /* - * newNnz > elementCnt_ is necessary for the following condition: - * Firstly, height_ is 9 elementCnt_ is 56 - * Secondly, height_ is 11 elementCnt_ is 44 - * ==> height_ is bigger, sMatrix_ will resize, and total item is 44 now - * Then, height_ is 10 elementCnt_ is 52 - * ==> Without newNnz > elementCnt_ condition, sMatrix_ will fail - */ - if ((ssize_t)((newHeight + 1) * sizeof(int)) > - ((char*)cols_ - (char*)rows_) || - newNnz > static_cast(sMatrix_->nnz)) { - sMatrix_ = NULL; - } else if (NO_VALUE == valueType) { - if ((ssize_t)(newNnz * sizeof(int)) > (end_ - (char*)cols_)) { - sMatrix_ = NULL; - } - } else { - if ((ssize_t)(newNnz * sizeof(int)) > ((char*)value_ - (char*)cols_) || - (ssize_t)(newNnz * sizeof(real)) > (end_ - (char*)value_)) { - sMatrix_ = NULL; - } - } - } - - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newNnz; - valueType_ = valueType; - format_ = SPARSE_CSR; - - if (sMatrix_ == NULL) { - sparseResizeCSR(); - } -} - -void GpuSparseMatrix::resizeCSC(size_t newHeight, - size_t newWidth, - size_t newNnz, - SparseValueType valueType) { - size_t newSize = (newWidth + 1) * sizeof(int) + newNnz * sizeof(int); - if (NO_VALUE != valueType) { - newSize += newNnz * sizeof(real); - } - - if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) { - memoryHandle_ = std::make_shared(newSize); - data_ = reinterpret_cast(memoryHandle_->getBuf()); - sMemoryHandle_ = std::make_shared(newSize); - end_ = reinterpret_cast(sMemoryHandle_->getBuf()) + - sMemoryHandle_->getSize(); - sMatrix_ = NULL; - } else if (valueType != valueType_) { - sMatrix_ = NULL; - } else { - /* - * newNnz > elementCnt_ is necessary for the following condition: - * Firstly, height_ is 9 elementCnt_ is 56 - * Secondly, height_ is 11 elementCnt_ is 44 - * ==> height_ is bigger, sMatrix_ will resize, - * and total item is 44 now - * Then, height_ is 10 elementCnt_ is 52 - * ==> Without newNnz > elementCnt_ condition, sMatrix_ will fail - */ - if ((ssize_t)((newWidth + 1) * sizeof(int)) > - ((char*)rows_ - (char*)cols_) || - newNnz > static_cast(sMatrix_->nnz)) { - sMatrix_ = NULL; - } else if (NO_VALUE == valueType) { - if ((ssize_t)(newNnz * sizeof(int)) > (end_ - (char*)rows_)) { - sMatrix_ = NULL; - } - } else { - if ((ssize_t)(newNnz * sizeof(int)) > ((char*)value_ - (char*)rows_) || - (ssize_t)(newNnz * sizeof(real)) > (end_ - (char*)value_)) { - sMatrix_ = NULL; - } - } - } - - height_ = newHeight; - width_ = newWidth; - elementCnt_ = newNnz; - valueType_ = valueType; - format_ = SPARSE_CSC; - - if (sMatrix_ == NULL) { - sparseResizeCSC(); - } -} - -void GpuSparseMatrix::resize(size_t newHeight, size_t newWidth) { - resize(newHeight, newWidth, elementCnt_, valueType_, format_); -} - -MatrixPtr GpuSparseMatrix::getTranspose() { - CHECK(memoryHandle_.get() || sMatrix_) << "not supported"; - if (memoryHandle_.get()) { - MatrixPtr copy_T(new GpuSparseMatrix( - std::dynamic_pointer_cast(memoryHandle_), - sMatrix_, - height_, - width_, - elementCnt_, - valueType_, - format_, - true, - sMemoryHandle_)); - return copy_T; - } else { - MatrixPtr copy_T(new GpuSparseMatrix(sMatrix_, - height_, - width_, - elementCnt_, - valueType_, - format_, - true, - sMemoryHandle_)); - return copy_T; - } -} - -void GpuSparseMatrix::copyRow(int offsets, - size_t colNum, - const sparse_non_value_t* row) { - memcpy(cols_ + offsets, row, sizeof(int) * colNum); -} - -void GpuSparseMatrix::copyRow(int offsets, - size_t colNum, - const sparse_float_value_t* row) { - for (size_t j = 0; j < colNum; j++) { - cols_[offsets + j] = row[j].col; - value_[offsets + j] = row[j].value; - } -} - -void GpuSparseMatrix::copyFrom(const Matrix& src, hl_stream_t stream) { - if (auto mat = dynamic_cast(&src)) { - copyFrom(*(const_cast(mat)), stream); - } else if (auto mat = dynamic_cast(&src)) { - copyFrom(*(const_cast(mat)), stream); - } else { - LOG(FATAL) << "Not implemented"; - } -} - -void GpuSparseMatrix::copyFrom(const Matrix& src) { - copyFrom(src, HPPL_STREAM_1); - hl_stream_synchronize(HPPL_STREAM_1); -} - -template -void GpuSparseMatrix::copyFrom(int64_t* ids, - int64_t* indices, - T* data, - hl_stream_t stream) { - CHECK_EQ(format_, SPARSE_CSR); - size_t nnz = 0; - for (size_t i = 0; i < height_; i++) { - int64_t id = ids[i]; - nnz += indices[id + 1] - indices[id]; - } - - resize(height_, - width_, - nnz, - sizeof(T) == sizeof(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE, - format_); - - rows_[0] = 0; - for (size_t i = 0; i < height_; i++) { - int64_t id = ids[i]; - size_t colNum = indices[id + 1] - indices[id]; - rows_[i + 1] = rows_[i] + colNum; - - T* row = data + indices[id]; - copyRow(rows_[i], colNum, row); - } - - sMatrix_->format = HL_SPARSE_CSR; - sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; - sMatrix_->rows = height_; - sMatrix_->cols = width_; - sMatrix_->nnz = nnz; - hl_memcpy_csr_matrix(sMatrix_.get(), value_, rows_, cols_, stream); -} - -void GpuSparseMatrix::setRow(size_t row, - size_t colNum, - const unsigned int* cols, - const real* values) { - CHECK_EQ(format_, SPARSE_CSR); - if (NO_VALUE == valueType_) { - CHECK_LT(row, height_); - CHECK(NULL != cols); - CHECK(NULL == values); - } else { - CHECK_LT(row, height_); - CHECK(NULL != cols); - CHECK(NULL != values); - } - if (0 == row) { - rows_[row] = 0; - } - rows_[row + 1] = rows_[row] + colNum; - - memcpy(cols_ + rows_[row], cols, sizeof(*cols) * colNum); - if (FLOAT_VALUE == valueType_) { - memcpy(value_ + rows_[row], values, sizeof(*values) * colNum); - } - - if (height_ - 1 == row) { - sMatrix_->format = HL_SPARSE_CSR; - sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; - sMatrix_->rows = height_; - sMatrix_->cols = width_; - sMatrix_->nnz = elementCnt_; - hl_memcpy_csr_matrix( - sMatrix_.get(), value_, rows_, cols_, HPPL_STREAM_DEFAULT); - } -} - -SparseValueType GpuSparseMatrix::getValueType() const { return valueType_; } - -void GpuSparseMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) { - CHECK_EQ(format_, SPARSE_CSC); - int nnz = sMatrix_->nnz; - if (memAlloc) { - matTrans = std::make_shared( - width_, height_, nnz, valueType_, format_, false); - } else { - CHECK(matTrans != nullptr); - } - - CpuIVector rows(nnz); - CpuIVector cols(width_ + 1); - CpuIVector cols_full(nnz); - CpuVector value(nnz); - hl_stream_t stream = HPPL_STREAM_1; - hl_memcpy_from_csc_matrix(value.getData(), - nnz, - rows.getData(), - nnz, - cols.getData(), - width_ + 1, - sMatrix_.get(), - stream); - - hl_stream_synchronize(stream); - - /*for every non zero number, get its column index*/ - std::vector dataVec; - for (size_t i = 0; i < width_; i++) { - for (int j = cols.getData()[i]; j < cols.getData()[i + 1]; j++) { - cols_full.getData()[j] = i; - } - } - - /*sort row index and column index by the ascending order*/ - for (int i = 0; i < nnz; i++) { - dataVec.emplace_back( - rows.getData()[i], cols_full.getData()[i], value.getData()[i]); - } - std::sort(dataVec.begin(), dataVec.end(), [](Element a, Element b) { - return a.row < b.row || (a.row == b.row && a.col < b.col); - }); - - /*get sorted data, row index, and col index, put them in the right place*/ - cols.resize(height_ + 1); - rows.resize(nnz); - value.resize(nnz); - - cols.getData()[0] = 0; - rows.getData()[0] = dataVec[0].col; - value.getData()[0] = dataVec[0].val; - for (int i = 1; i < nnz; i++) { - if (dataVec[i].row != dataVec[i - 1].row) { - for (int j = dataVec[i - 1].row + 1; j <= dataVec[i].row; j++) { - cols.getData()[j] = i; - } - } - rows.getData()[i] = dataVec[i].col; - value.getData()[i] = dataVec[i].val; - } - cols.getData()[height_] = nnz; - - /*copy back from cpu*/ - GpuSparseMatrixPtr dest = - std::dynamic_pointer_cast(matTrans); - hl_memcpy_csc_matrix((dest->sMatrix_).get(), - value.getData(), - rows.getData(), - cols.getData(), - stream); - hl_stream_synchronize(stream); -} - -void GpuSparseMatrix::mul(const GpuMatrix& a, - const GpuMatrix& b, - real scaleAB, - real scaleT) { - CHECK(a.useGpu_ && b.useGpu_) << "type not match"; - CHECK(!trans_) << "trans not supported"; - real* A_d = (real*)a.getData(); - real* B_d = (real*)b.getData(); - hl_sparse_matrix_s C_d = sMatrix_.get(); - hl_trans_op_t a_trans = a.trans_ ? HPPL_OP_T : HPPL_OP_N; - hl_trans_op_t b_trans = b.trans_ ? HPPL_OP_T : HPPL_OP_N; - - if (!a.trans_ && !b.trans_) { - CHECK(height_ == a.getHeight()); - CHECK(width_ == b.getWidth()); - CHECK(a.getWidth() == b.getHeight()); - } else if (a.trans_ && !b.trans_) { - CHECK(height_ == a.getWidth()); - CHECK(width_ == b.getWidth()); - CHECK(a.getHeight() == b.getHeight()); - } else if (!a.trans_ && b.trans_) { - CHECK(height_ == a.getHeight()); - CHECK(width_ == b.getHeight()); - CHECK(a.getWidth() == b.getWidth()); - } else { - LOG(INFO) << "Not support"; - } - int dimM = height_; - int dimN = width_; - int dimK = !b.trans_ ? b.getHeight() : b.getWidth(); - hl_sparse_matrix_mul( - A_d, a_trans, B_d, b_trans, C_d, dimM, dimN, dimK, scaleAB, scaleT); -} - -void GpuSparseMatrix::mul(const Matrix& a, - const Matrix& b, - real scaleAB, - real scaleT) { - const auto a_ptr = dynamic_cast(&a); - const auto b_ptr = dynamic_cast(&b); - if (a_ptr && b_ptr) { - mul(*a_ptr, *b_ptr, scaleAB, scaleT); - } else { - LOG(FATAL) << "not supported"; - } -} - -template -void printBuf(std::ostream& os, T* a, size_t len, const char* name) { - os << "\n: " << name << " ["; - for (size_t i = 0; i < len; i++) { - os << a[i] << " "; - } - os << "]\n"; -} - -void GpuSparseMatrix::print(std::ostream& os) const { - if (format_ == SPARSE_CSC) { - int nnz = sMatrix_->nnz; - IVectorPtr rows = IVector::create(nnz, false); - IVectorPtr cols = IVector::create(width_ + 1, false); - VectorPtr value = Vector::create(nnz, false); - hl_stream_t stream = HPPL_STREAM_DEFAULT; - hl_memcpy_from_csc_matrix(value->getData(), - value->getSize(), - rows->getData(), - rows->getSize(), - cols->getData(), - cols->getSize(), - sMatrix_.get(), - stream); - hl_stream_synchronize(stream); - - printBuf(os, cols->getData(), width_ + 1, "col idx"); - printBuf(os, rows->getData(), elementCnt_, "row idx"); - printBuf(os, value->getData(), elementCnt_, "value"); - } -} - -void GpuSparseMatrix::copyFromCSR(CpuSparseMatrix& src, hl_stream_t stream) { - trans_ = src.trans_; - size_t nnz = src.getElementCnt(); - - resize(src.getHeight(), src.getWidth(), nnz, valueType_, src.getFormat()); - // if have different value type, only copy rows and cols - SparseValueType vType = - valueType_ != src.getValueType() ? NO_VALUE : valueType_; - - sMatrix_->format = HL_SPARSE_CSR; - sMatrix_->type = vType == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; - sMatrix_->rows = height_; - sMatrix_->cols = width_; - sMatrix_->nnz = nnz; - - hl_memcpy_csr_matrix(sMatrix_.get(), - vType == NO_VALUE ? NULL : src.getValue(), - src.getRows(), - src.getCols(), - stream); - - // restore type of sMatrix_ - sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; -} - -void GpuSparseMatrix::copyFromCSC(CpuSparseMatrix& src, hl_stream_t stream) { - trans_ = src.trans_; - size_t nnz = src.getElementCnt(); - - resize(src.getHeight(), src.getWidth(), nnz, valueType_, src.getFormat()); - - // if have different value type, only copy rows and cols - SparseValueType vType = - valueType_ != src.getValueType() ? NO_VALUE : valueType_; - - sMatrix_->format = HL_SPARSE_CSC; - sMatrix_->type = vType == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; - sMatrix_->rows = height_; - sMatrix_->cols = width_; - sMatrix_->nnz = nnz; - - hl_memcpy_csc_matrix(sMatrix_.get(), - vType == NO_VALUE ? NULL : src.getValue(), - src.getRows(), - src.getCols(), - stream); - - // restore type of sMatrix_ - sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; -} - -void GpuSparseMatrix::copyFrom(GpuSparseMatrix& src, hl_stream_t stream) { - CHECK(trans_ == src.trans_); - CHECK(format_ == src.getFormat()); - resize(src.getHeight(), - src.getWidth(), - elementCnt_, - valueType_, - src.getFormat()); - - size_t rowSize = format_ == SPARSE_CSC ? elementCnt_ : height_ + 1; - size_t colSize = format_ == SPARSE_CSC ? width_ + 1 : elementCnt_; - - if (valueType_ == FLOAT_VALUE && src.getValueType() == FLOAT_VALUE) { - hl_memcpy_async( - getValue(), src.getValue(), sizeof(real) * elementCnt_, stream); - } - CHECK(getRows()); - CHECK(src.getRows()); - - hl_memcpy_async(getRows(), src.getRows(), sizeof(int) * rowSize, stream); - hl_memcpy_async(getCols(), src.getCols(), sizeof(int) * colSize, stream); -} - -void GpuSparseMatrix::copyFrom(CpuSparseMatrix& src, hl_stream_t stream) { - if (format_ == SPARSE_CSR) { - copyFromCSR(src, stream); - } else { - copyFromCSC(src, stream); - } -} - -void GpuSparseMatrix::trimFromCSR(const CpuSparseMatrix& src) { - trans_ = src.trans_; - int* srcCols = src.getCols(); - size_t nnz = std::count_if(srcCols, - srcCols + src.getElementCnt(), - [this](size_t n) { return n < this->width_; }); - resize(height_, width_, nnz, valueType_, format_); - - rows_[0] = 0; - size_t index = 0; - for (size_t r = 0; r < height_; ++r) { - for (int i = src.getRows()[r]; i < src.getRows()[r + 1]; ++i) { - if (srcCols[i] < (int)width_) { - cols_[index] = srcCols[i]; - if (valueType_ == FLOAT_VALUE) { - value_[index] = src.getValue()[i]; - } - ++index; - } - } - rows_[r + 1] = index; - } - CHECK_EQ(index, nnz); - - sMatrix_->format = HL_SPARSE_CSR; - sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; - sMatrix_->rows = height_; - sMatrix_->cols = width_; - sMatrix_->nnz = nnz; - - hl_memcpy_csr_matrix(sMatrix_.get(), - valueType_ == NO_VALUE ? NULL : value_, - rows_, - cols_, - /*default stream = */ HPPL_STREAM_DEFAULT); -} - -void GpuSparseMatrix::trimFromCSC(const CpuSparseMatrix& src) { - trans_ = src.trans_; - size_t nnz = src.getCols()[width_] - src.getCols()[0]; - resize(height_, width_, nnz, valueType_, format_); - - cols_[0] = 0; - for (size_t i = 0; i < width_; i++) { - cols_[i + 1] = cols_[i] + (int)(src.getRowNum(i)); - } - memcpy(rows_, src.getRows() + src.getCols()[0], sizeof(int) * nnz); - if (valueType_ == FLOAT_VALUE) { - memcpy(value_, src.getValue() + src.getCols()[0], sizeof(real) * nnz); - } - - sMatrix_->format = HL_SPARSE_CSC; - sMatrix_->type = valueType_ == NO_VALUE ? HL_NO_VALUE : HL_FLOAT_VALUE; - sMatrix_->rows = height_; - sMatrix_->cols = width_; - sMatrix_->nnz = nnz; - - hl_memcpy_csc_matrix(sMatrix_.get(), - valueType_ == NO_VALUE ? NULL : value_, - rows_, - cols_, - /*default stream = */ HPPL_STREAM_DEFAULT); -} - -void GpuSparseMatrix::trimFrom(const CpuSparseMatrix& src) { - if (format_ == SPARSE_CSR) { - trimFromCSR(src); - } else { - trimFromCSC(src); - } -} - -void GpuSparseMatrix::addBias(Matrix& b, real scale) { - CHECK(b.getHeight() == 1) << "the Bias should be a vector"; - hl_sparse_matrix_s A_d = sMatrix_.get(); - hl_sparse_matrix_add_bias(A_d, b.getData(), scale); -} - -void GpuSparseMatrix::add3(GpuMatrix* b) { - CHECK(getFormat() != SPARSE_CSC) << "Not supported"; - CHECK(height_ == b->getHeight()); - CHECK(width_ == b->getWidth()); - real* B_d = b->getData(); - hl_sparse_matrix_s A_d = sMatrix_.get(); - hl_sparse_matrix_add_dense(A_d, B_d, height_, width_, 1, 0); -} - -void GpuSparseMatrix::add3(MatrixPtr b) { - if (dynamic_cast(b.get())) { - add3(dynamic_cast(b.get())); - } else { - LOG(FATAL) << "not supported"; - } -} - -void GpuSparseMatrix::zeroMem() { - CHECK(valueType_ == FLOAT_VALUE); - real* value = getValue(); - if (value == NULL) { - LOG(FATAL) << "value is nullptr"; - } - hl_matrix_zero_mem(value, elementCnt_); -} - -void GpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { -#ifdef PADDLE_WITH_CUDA - CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; - size_t numSamples = getHeight(); - size_t beam = maxVal.getWidth(); - CHECK_EQ(maxIds.getSize(), numSamples * beam); - CHECK_EQ(maxVal.getHeight(), numSamples); - CHECK_EQ(format_, SPARSE_CSR) << "Only support SPARSE_CSR"; - - hl_sparse_matrix_top_k(maxVal.getData(), - maxVal.getStride(), - maxIds.getData(), - sMatrix_.get(), - beam, - numSamples); -#endif -} - -template void GpuSparseMatrix::copyFrom(int64_t* ids, - int64_t* indices, - sparse_non_value_t* data, - hl_stream_t stream); -template void GpuSparseMatrix::copyFrom(int64_t* ids, - int64_t* indices, - sparse_float_value_t* data, - hl_stream_t stream); -} // namespace paddle diff --git a/paddle/math/SparseRowMatrix.cpp b/paddle/math/SparseRowMatrix.cpp deleted file mode 100644 index 4254175aabc8c32edb243d4a82c2e34c81393f74..0000000000000000000000000000000000000000 --- a/paddle/math/SparseRowMatrix.cpp +++ /dev/null @@ -1,282 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SparseRowMatrix.h" -#include "CpuSparseMatrix.h" - -#include - -#include "paddle/utils/Logging.h" - -#include "SIMDFunctions.h" - -#include "paddle/utils/Thread.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -const unsigned int SparseRowCpuMatrix::kUnusedId_ = -1U; - -void SparseRowCpuMatrix::init(size_t height, size_t width) { - height_ = height; - if (!indexDictHandle_) { - indexDictHandle_.reset(new IndexDict); - indexDictHandle_->globalIndices.assign(height, kUnusedId_); - } - localIndices_ = &indexDictHandle_->localIndices; - globalIndices_ = indexDictHandle_->globalIndices.data(); -} - -void SparseRowCpuMatrix::mul(CpuSparseMatrix* a, - CpuMatrix* b, - real scaleAB, - real scaleT) { - CpuMatrix::mul(a, b, this, scaleAB, scaleT); -} - -void SparseRowCpuMatrix::copyFrom(const real* src, size_t size) { - LOG(FATAL) << "This should not be called"; -} - -void SparseRowCpuMatrix::zeroMem() { - apply([](real* buf, size_t len) { memset(buf, 0, sizeof(real) * len); }); - clearRows(); -} - -void SparseRowCpuMatrix::applyL1(real learningRate, real decayRate) { - apply([=](real* buf, size_t len) { - CpuVector value(0, nullptr); - value.subVecFrom(buf, 0, len); - value.applyL1(learningRate, decayRate); - }); -} - -void SparseRowCpuMatrix::sgdUpdate(BaseMatrix& value, - IVector& t0, - real learningRate, - int currentTime, - real decayRate, - bool useL1, - bool fini) { - std::vector& localIndices = indexDictHandle_->localIndices; - - // t0 and value are vectors - CHECK_EQ(t0.getSize(), this->height_); - CHECK_EQ(value.width_, this->height_ * this->width_); - - if (decayRate == 0.0f) { - if (fini) { - return; - } - - for (size_t i = 0; i < localIndices.size(); ++i) { - real* g = getLocalRow(i); - real* v = value.rowBuf(localIndices[i]); - for (size_t j = 0; j < this->width_; ++j) { - v[j] -= learningRate * g[j]; - } - } - return; - } // else - - if (useL1) { // L1 decay - if (fini) { - for (size_t i = 0; i < this->height_; ++i) { - real* v = value.rowBuf(i); - int* t = t0.getData() + i; - if (t[0] < currentTime) { - // W(t0) -> W(t+1) - int tDiff = currentTime - t[0]; - real delta = tDiff * learningRate * decayRate; - simd::decayL1(v, v, delta, this->width_); - } - } - return; - } // else - - for (size_t i = 0; i < localIndices.size(); ++i) { - real* g = getLocalRow(i); - real* v = value.rowBuf(localIndices[i]); - int* t = t0.getData() + localIndices[i]; - if (t[0] < currentTime) { - // W(t0) -> W(t) - int tDiff = currentTime - t[0]; - real delta = tDiff * learningRate * decayRate; - simd::decayL1(v, v, delta, this->width_); - } - - // W(t) -> W(t+1) - for (size_t j = 0; j < this->width_; ++j) { - v[j] -= learningRate * g[j]; - } - simd::decayL1(v, v, learningRate * decayRate, this->width_); - - // state update to t+1 - t[0] = currentTime + 1; - } - - } else { // L2 decay - if (fini) { - for (size_t i = 0; i < this->height_; ++i) { - real* v = value.rowBuf(i); - int* t = t0.getData() + i; - if (t[0] < currentTime) { - // W(t0) -> W(t+1) - int tDiff = currentTime - t[0]; - real recip = 1.0f / (1.0f + tDiff * learningRate * decayRate); - for (size_t j = 0; j < this->width_; ++j) { - v[j] *= recip; - } - } - } - return; - } // else - - real recipDecay = 1.0f / (1.0f + learningRate * decayRate); - - for (size_t i = 0; i < localIndices.size(); ++i) { - real* g = getLocalRow(i); - real* v = value.rowBuf(localIndices[i]); - int* t = t0.getData() + localIndices[i]; - if (t[0] < currentTime) { - // W(t0) -> W(t) - int tDiff = currentTime - t[0]; - real recip = 1.0f / (1.0f + tDiff * learningRate * decayRate); - for (size_t j = 0; j < this->width_; ++j) { - v[j] *= recip; - } - } - - // W(t) -> W(t+1) - for (size_t j = 0; j < this->width_; ++j) { - v[j] = recipDecay * (v[j] - learningRate * g[j]); - } - - // state update to t+1 - t[0] = currentTime + 1; - } - } -} - -void SparseRowCpuMatrix::addTo(BaseMatrix& dest, - std::vector& ids, - size_t tid, - size_t numThreads) { - CHECK(!dest.useGpu_); - CHECK_EQ(dest.height_ * dest.width_, this->height_ * this->width_); - - std::vector& localIndices = indexDictHandle_->localIndices; - for (size_t i = 0; i < localIndices.size(); ++i) { - uint32_t id = localIndices[i]; - if (id % numThreads == tid) { - simd::addTo(dest.rowBuf(id), getLocalRow(i), this->width_); - ids.push_back(id); - } - } -} - -void SparseRowCpuMatrix::addTo(SparseRowCpuMatrix& dest, - size_t tid, - size_t numThreads) { - CHECK(!dest.useGpu_); - CHECK_EQ(dest.height_ * dest.width_, this->height_ * this->width_); - - std::vector& localIndices = indexDictHandle_->localIndices; - for (size_t i = 0; i < localIndices.size(); ++i) { - uint32_t id = localIndices[i]; - if (id % numThreads == tid) { - dest.checkIndex(id); - simd::addTo(dest.getRow(id), getLocalRow(i), this->width_); - } - } -} - -void SparseRowCpuMatrix::zeroMemThread(size_t tid, size_t numThreads) { - std::vector& localIndices = indexDictHandle_->localIndices; - for (size_t i = 0; i < localIndices.size(); ++i) { - uint32_t id = localIndices[i]; - if (id % numThreads == tid) { - memset(this->getLocalRow(i), 0, this->width_ * sizeof(real)); - } - } -} - -void SparseAutoGrowRowCpuMatrix::mul(CpuSparseMatrix* a, - CpuMatrix* b, - real scaleAB, - real scaleT) { - CpuMatrix::mul( - a, b, this, scaleAB, scaleT); -} - -void CacheRowCpuMatrix::mul(CpuSparseMatrix* a, - CpuMatrix* b, - real scaleAB, - real scaleT) { - CpuMatrix::mul(a, b, this, scaleAB, scaleT); -} - -void SparsePrefetchRowCpuMatrix::addRows(const unsigned int* ids, size_t len) { - std::vector& localIndices = indexDictHandle_->localIndices; - for (size_t i = 0; i < len; i++) { - CHECK_LT(*(ids + i), this->getHeight()) - << "id:" << *(ids + i) << "Height:" << this->getHeight() - << "sparse id value exceeds the max input dimension, " - << "it could be caused invalid input data samples"; - } - localIndices.insert(localIndices.end(), ids, ids + len); -} - -void SparsePrefetchRowCpuMatrix::addRows(MatrixPtr input) { - CpuSparseMatrix* mat = dynamic_cast(input.get()); - CHECK(mat) << "only support sparse matrix"; - addRows(reinterpret_cast(mat->getCols()), - mat->getElementCnt()); -} - -void SparsePrefetchRowCpuMatrix::addRows(IVectorPtr ids) { - std::vector& localIndices = indexDictHandle_->localIndices; - size_t numSamples = ids->getSize(); - int* index = ids->getData(); - for (size_t i = 0; i < numSamples; ++i) { - if (index[i] == -1) continue; - - unsigned int id = (unsigned int)index[i]; - CHECK_LT(id, this->getHeight()) - << "id:" << id << "Height:" << this->getHeight() - << "sparse id value exceeds the max input dimension, " - << "it could be caused invalid input data samples"; - localIndices.push_back(id); - } -} - -void SparsePrefetchRowCpuMatrix::setupIndices() { - auto& localIndices = indexDictHandle_->localIndices; - uniqueIds(localIndices); - // for each sparse row - for (size_t id = 0; id < localIndices.size(); ++id) { - globalIndices_[localIndices[id]] = id; // sparse row -> local id - } - checkStoreSize(); -} - -void SparseRowCpuMatrix::checkIndices() { - std::vector& localIndices = indexDictHandle_->localIndices; - for (size_t i = 0; i < localIndices.size(); ++i) { - CHECK_EQ(globalIndices_[localIndices[i]], i); - } - checkStoreSize(); -} - -} // namespace paddle diff --git a/paddle/math/SparseRowMatrix.h b/paddle/math/SparseRowMatrix.h deleted file mode 100644 index cf6779e8b0b1d6b0c13b21a08ffff5af76e57ba6..0000000000000000000000000000000000000000 --- a/paddle/math/SparseRowMatrix.h +++ /dev/null @@ -1,341 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#ifndef PADDLE_MOBILE_INFERENCE - -#include -#include -#include -#include "Matrix.h" -#include "RowBuffer.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -/** - * Sparse Row - */ -class SparseRowCpuMatrix : public CpuMatrix { - public: - struct IndexDict { - // In the following, global id means the row id in the original matrix. - // Local id means the row id in the local storage which only contains - // the sparse rows. - std::vector localIndices; // local id -> global id - std::vector globalIndices; // global id -> local id - }; - typedef std::shared_ptr IndexDictPtr; - - /// heightStore is max number of rows of the sparse matrix. - SparseRowCpuMatrix(CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - IndexDictPtr indexDictHandle = nullptr, - bool trans = false) - : CpuMatrix(nullptr, height, width, trans), - indexDictHandle_(indexDictHandle) { - init(height, width); - buf_.reset(new RowBuffer(dataHandle, width)); - } - - virtual ~SparseRowCpuMatrix() {} - - public: - /** - * Get the row buf - * - * @param row row id in the original matrix - */ - real* getRow(size_t row) { - CHECK_NE(globalIndices_[row], kUnusedId_); - return getLocalRow(globalIndices_[row]); - } - - /** - * Get the row buf - * - * @param row row id in local storage - */ - real* getLocalRow(size_t row) { return buf_->getWithAutoGrowth(row); } - - /** - * reserve the storage for rows according to current size of - * indexDictHandle. - * - * This is only used when SparseRowCpuMatrix is constructed with - * indexDictHandle. - */ - void reserveStore() { buf_->resize(localIndices_->size()); } - - // row is the row id in the original matrix - virtual real* getRowBuf(size_t row) { return getRow(row); } - - virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); - - /** - * Fill data according to row indexs added, setup indices inside. - * - * *src* and *size* are data and size of normal dense CpuMatrix. - */ - virtual void copyFrom(const real* src, size_t size); - virtual void zeroMem(); - - /** - * apply L1 to all sparse rows, should be apply after indices ready. - */ - virtual void applyL1(real learningRate, real decayRate); - - void clearIndices() { clearRows(); } - void zeroMemThread(size_t tid, size_t numThreads); - - /** - * value -= grad * learningRate, this is gradient. - * - * If L1 decay set use L1, else if L2 set use L2, otherwise no decay atall. - * - * t0 is a int vector used by L1/L2 decay, size = height of parameter - * matrix, - * store the time that each weight row last updated. - * - * Time is batchId, currentTime is current batchId. - * - * While pass finished, caller should call this func one more time - * with (fini=true) to let weight decay catch up current time. - */ - void sgdUpdate(BaseMatrix& value, - IVector& t0, - real learningRate, - int currentTime, - real decayRate, - bool useL1, - bool fini = false); - - /** - * merge rows in *this* to *dest* for designated thread - * - * values add to *dest* matrix - * - * ids occured in *this* append to *ids* - * filtered by (id % numThreads == tid) - */ - void addTo(BaseMatrix& dest, - std::vector& ids, - size_t tid, - size_t numThreads); - - /** - * the second version addTo(), *dest* is a SparseRowCpuMatrix. - * - * The dest's indices should be setup already, addTo() will - * check src ids is exist in dest's indices. - */ - void addTo(SparseRowCpuMatrix& dest, size_t tid, size_t numThreads); - - const IndexDictPtr& getIndexDictHandle() const { return indexDictHandle_; } - - /** - * check all local and global indices consistency - */ - void checkIndices(); - /** - * check whether row *i* exist in indices - */ - void checkIndex(size_t i) { - size_t localId = globalIndices_[i]; - CHECK_LT(localId, localIndices_->size()); - CHECK_EQ((*localIndices_)[localId], i); - } - - std::vector& getLocalIndices() const { - return indexDictHandle_->localIndices; - } - - protected: - template - void apply(Func f) { - f(buf_->data(), localIndices_->size() * width_); - } - - void init(size_t height, size_t width); - - /// clear row indices. - void clearRows() { - for (auto id : *localIndices_) { - globalIndices_[id] = kUnusedId_; - } - localIndices_->clear(); - buf_->clear(); - } - - inline void checkStoreSize() { - if (buf_->isAutoGrowth()) { - if (buf_->getRowCount() > 0.5 * height_) { - LOG(WARNING) << "There are more than 0.5*height (" - << localIndices_->size() << ") rows are used for sparse " - << "update, which is not efficient. Considering not use " - << "sparse_update."; - } - } else { - CHECK_LE(localIndices_->size(), buf_->getRowCount()); - } - } - - std::unique_ptr buf_; - IndexDictPtr indexDictHandle_; - std::vector* localIndices_; // =&indexDictHandle_->localIndices - unsigned int* globalIndices_; // =indexDictHandle_->globalIndices.data(); - static const unsigned int kUnusedId_; -}; - -class SyncThreadPool; - -/// For prefetching parameters from remote Parameter server -class SparsePrefetchRowCpuMatrix : public SparseRowCpuMatrix { - public: - SparsePrefetchRowCpuMatrix(CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - IndexDictPtr indexDictHandle = nullptr, - SyncThreadPool* pool = nullptr, - bool trans = false) - : SparseRowCpuMatrix(dataHandle, height, width, indexDictHandle, trans), - pool_(pool) {} - - /** - * Extract feature ids from *input*, to fill row indexs. - * - * *input* must be sparse matrix. - * - * Can call many times before setup. - */ - void addRows(MatrixPtr input); - void addRows(IVectorPtr ids); - - /** - * setup global indices of SparseRowMatrix after finish add rows. - */ - void setupIndices(); - - protected: - void addRows(const unsigned int* ids, size_t len); - SyncThreadPool* pool_; -}; - -class SparseAutoGrowRowCpuMatrix : public SparseRowCpuMatrix { - public: - SparseAutoGrowRowCpuMatrix(size_t height, - size_t width, - IndexDictPtr indexDictHandle = nullptr, - bool trans = false) - : SparseRowCpuMatrix(nullptr, height, width, indexDictHandle, trans) {} - - real* getRow(size_t row) { - auto id = globalIndices_[row]; - if (id == kUnusedId_) { - id = globalIndices_[row] = localIndices_->size(); - localIndices_->push_back(row); - checkStoreSize(); - } - return getLocalRow(id); - } - - virtual real* getRowBuf(size_t row) { return getRow(row); } - - virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); -}; - -class CacheRowCpuMatrix : public SparseAutoGrowRowCpuMatrix { - public: - CacheRowCpuMatrix(size_t height, - size_t width, - IndexDictPtr indexDictHandle = nullptr, - bool trans = false) - : SparseAutoGrowRowCpuMatrix(height, width, indexDictHandle, trans), - sourceData_(nullptr) {} - - void setSourceData(CpuVectorPtr sourceVec) { - sourceDataVec_ = sourceVec; - sourceData_ = sourceVec->getData(); - } - - real* getRow(size_t row) { - auto id = globalIndices_[row]; - if (id == kUnusedId_) { - id = globalIndices_[row] = localIndices_->size(); - localIndices_->push_back(row); - checkStoreSize(); - memcpy( - getLocalRow(id), sourceData_ + width_ * row, sizeof(float) * width_); - } - return getLocalRow(id); - } - - virtual real* getRowBuf(size_t row) { return getRow(row); } - - virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); - - public: - CpuVectorPtr sourceDataVec_; - real* sourceData_; -}; - -/** - * Sparse Row Ids Matrix. - * - * mostly same as CpuMatrix, but maintain sparse row ids occured, - * ids are hashed by worker thread id. - */ -class SparseRowIdsCpuMatrix : public CpuMatrix { - public: - SparseRowIdsCpuMatrix(CpuMemHandlePtr dataHandle, - size_t height, - size_t width, - bool trans = false) - : CpuMatrix(dataHandle, height, width, trans) {} - - void setNumOfThreads(size_t numOfThreads) { idsArray_.resize(numOfThreads); } - - std::vector& getIds(size_t threadId) { return idsArray_[threadId]; } - - private: - std::vector> idsArray_; -}; - -} // namespace paddle - -#else -namespace paddle { - -class SparseRowCpuMatrix : public CpuMatrix { - public: - void reserveStore() {} - void clearIndices() {} -}; - -class SparsePrefetchRowCpuMatrix : public SparseRowCpuMatrix { - public: - void setupIndices() {} - void addRows(MatrixPtr input) {} - void addRows(IVectorPtr ids) {} -}; - -class SparseAutoGrowRowCpuMatrix : public SparseRowCpuMatrix {}; -class CacheRowCpuMatrix : public SparseAutoGrowRowCpuMatrix {}; -class SparseRowIdsCpuMatrix : public CpuMatrix {}; - -} // namespace paddle - -#endif diff --git a/paddle/math/Storage.cpp b/paddle/math/Storage.cpp deleted file mode 100644 index 5982bf2e5637ff4b4af6baae47e40b68e0c07c86..0000000000000000000000000000000000000000 --- a/paddle/math/Storage.cpp +++ /dev/null @@ -1,101 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Storage.h" -#include "Allocator.h" -#include "paddle/utils/StringUtil.h" -#include "paddle/utils/Util.h" - -#ifndef PADDLE_MOBILE_INFERENCE -DEFINE_int32(pool_limit_size, - 536870912, - "maximum memory size managed by a memory pool, default is 512M"); -#else -DEFINE_int32(pool_limit_size, 0, "default is 0"); -#endif - -namespace paddle { - -// Initialization StorageEngine singleton. -// Other modules may rely on storage management, -// so StorageEngine need to be initialized before other modules. -static InitFunction __init_storage_engine([]() { StorageEngine::singleton(); }, - std::numeric_limits::max()); - -StorageEngine::StorageEngine() : cpuAllocator_(nullptr) {} - -StorageEngine::~StorageEngine() { - delete cpuAllocator_; - for (auto it : gpuAllocator_) { - delete it; - } -} - -StorageEngine* StorageEngine::singleton() { - static StorageEngine storage; - return &storage; -} - -PoolAllocator* StorageEngine::getGpuAllocator(int deviceId) { - { - // if gpuAllocator_ has been constructed - ReadLockGuard guard(lock_); - if (deviceId < static_cast(gpuAllocator_.size()) && - (gpuAllocator_[deviceId] != nullptr)) { - return gpuAllocator_[deviceId]; - } - } - - { - // Construct gpuAllocator_ - std::lock_guard guard(lock_); - if (deviceId >= static_cast(gpuAllocator_.size())) { - gpuAllocator_.resize(deviceId + 1); - } - if (gpuAllocator_[deviceId] == nullptr) { - std::string name = - "gpu" + str::to_string(deviceId) + std::string("_pool"); - gpuAllocator_[deviceId] = - new PoolAllocator(new GpuAllocator(), FLAGS_pool_limit_size, name); - } - return gpuAllocator_[deviceId]; - } -} - -PoolAllocator* StorageEngine::getCpuAllocator() { - { - // if cpuAllocator_ has been constructed - ReadLockGuard guard(lock_); - if (cpuAllocator_ != nullptr) { - return cpuAllocator_; - } - } - - { - // Construct cpuAllocator_ - std::lock_guard guard(lock_); - if (cpuAllocator_ == nullptr) { - if (FLAGS_use_gpu) { - cpuAllocator_ = new PoolAllocator( - new CudaHostAllocator(), FLAGS_pool_limit_size, "cuda_host_pool"); - } else { - cpuAllocator_ = new PoolAllocator( - new CpuAllocator(), FLAGS_pool_limit_size, "cpu_pool"); - } - } - return cpuAllocator_; - } -} - -} // namespace paddle diff --git a/paddle/math/Storage.h b/paddle/math/Storage.h deleted file mode 100644 index 61a9aa2a07442d9e4ede80c961e17e079eb8b3ba..0000000000000000000000000000000000000000 --- a/paddle/math/Storage.h +++ /dev/null @@ -1,52 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "PoolAllocator.h" -#include "paddle/utils/Locks.h" - -namespace paddle { - -/** - * @brief Storage manager for multiple devices. - */ -class StorageEngine { - public: - /** - * @return Storage singleton - */ - static StorageEngine* singleton(); - - /** - * @return return one gpu allocator by deviceId - */ - PoolAllocator* getGpuAllocator(int deviceId); - - /** - * @return return cpu allocator - */ - PoolAllocator* getCpuAllocator(); - - protected: - StorageEngine(); - ~StorageEngine(); - RWLock lock_; - std::vector gpuAllocator_; - PoolAllocator* cpuAllocator_; -}; - -} // namespace paddle diff --git a/paddle/math/TensorAssign.h b/paddle/math/TensorAssign.h deleted file mode 100644 index 7d4726ddba43202970c37dd1a08f842104b24ada..0000000000000000000000000000000000000000 --- a/paddle/math/TensorAssign.h +++ /dev/null @@ -1,158 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "paddle/utils/Logging.h" - -namespace paddle { - -/** - * \brief Tensor Assign Expression(return by lazyAssign, - * and evaluated by AssignEvaluate) - */ -template -class TensorAssignOp { - public: - explicit TensorAssignOp(const LhsType& lhs, const RhsType& rhs) - : lhs_(lhs), rhs_(rhs) { -#ifndef __CUDA_ARCH__ - CHECK_EQ(lhs_.getWidth(), rhs_.getWidth()); - CHECK_EQ(lhs_.getHeight(), rhs_.getHeight()); - CHECK_EQ(lhs_.useGpu(), rhs_.useGpu()); -#endif - } - - INLINE void apply(const int i, const int j) { - lhs_.applyRef(i, j) = rhs_.apply(i, j); - } - INLINE void apply(const int index) { - lhs_.applyRef(index) = rhs_.apply(index); - } - - INLINE size_t getWidth() const { return lhs_.getWidth(); } - INLINE size_t getHeight() const { return rhs_.getHeight(); } - INLINE bool isContiguous() const { - return lhs_.isContiguous() && rhs_.isContiguous(); - } - INLINE bool useGpu() const { return lhs_.useGpu(); } - - private: - TensorApply lhs_; - TensorApply rhs_; -}; - -template -void AssignCpuEvaluate(int height, - int width, - bool isContiguous, - Assign&& assign, - AssignOp&&... args) { - if (isContiguous) { - int size = height * width; - for (int index = 0; index < size; index++) { - assign.apply(index); - __attribute__((unused)) int dummy[] = {(((args)).apply(index), 0)...}; - } - } else { - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - assign.apply(i, j); - __attribute__((unused)) int dummy[] = {(((args)).apply(i, j), 0)...}; - } - } - } -} - -#ifdef __NVCC__ -template -__global__ void AssignGpuEvaluate1(const int border, - Assign assign, - AssignOp... args) { - const int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < border) { - assign.apply(idx); - __attribute__((unused)) int dummy[] = {(((args)).apply(idx), 0)...}; - } -} - -template -__global__ void AssignGpuEvaluate2(const int height, - const int width, - Assign assign, - AssignOp... args) { - const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; - const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; - for (int i = rowIdx; i < height; i += gridDim.y * blockDim.y) { - for (int j = colIdx; j < width; j += gridDim.x * blockDim.x) { - assign.apply(i, j); - __attribute__((unused)) int dummy[] = {(((args)).apply(i, j), 0)...}; - } - } -} -#endif - -/** - * \brief Evaluate one or more TensorAssignOp objects. - * - * \note At least one assignment expression is required - */ -template -void AssignEvaluate(Assign&& assign, AssignOp&&... args) { - const bool useGpu_ = assign.useGpu(); - bool isContiguous_ = assign.isContiguous(); - const size_t height = assign.getHeight(); - const size_t width = assign.getWidth(); - - const int packSize = sizeof...(args); - const bool packUseGpu[] = {((args)).useGpu()...}; - const bool packIsContiguous[] = {((args)).isContiguous()...}; - const size_t packHeight[] = {((args)).getHeight()...}; - const size_t packWidth[] = {((args)).getWidth()...}; - - for (int i = 0; i < packSize; i++) { - CHECK_EQ(useGpu_, packUseGpu[i]); - CHECK_EQ(height, packHeight[i]); - CHECK_EQ(width, packWidth[i]); - isContiguous_ = isContiguous_ && packIsContiguous[i]; - } - - if (useGpu_) { -#ifdef __NVCC__ - if (isContiguous_) { - int size = height * width; - int blockSize = size <= 1024 ? size : 1024; - int gridSize = (size + 1024 - 1) / 1024; - AssignGpuEvaluate1<<>>( - size, assign, args...); - } else { - int blockSizeY = std::min(32, (int)height); - int blockSizeX = (32 / blockSizeY) * 32; - int gridSizeX = std::min(32, (int)(width + blockSizeX - 1) / blockSizeX); - int gridSizeY = std::min(32, (int)(height + blockSizeY - 1) / blockSizeY); - dim3 threads(blockSizeX, blockSizeY); - dim3 grid(gridSizeX, gridSizeY); - AssignGpuEvaluate2<<>>( - height, width, assign, args...); - } - - CHECK_SYNC("AssignEvaluate failed"); -#endif - } else { - AssignCpuEvaluate(height, width, isContiguous_, assign, args...); - } -} - -} // namespace paddle diff --git a/paddle/math/TensorEvaluate.h b/paddle/math/TensorEvaluate.h deleted file mode 100644 index 2a722016e777a131ef14636a6871d29d9b131044..0000000000000000000000000000000000000000 --- a/paddle/math/TensorEvaluate.h +++ /dev/null @@ -1,112 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include "hl_base.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -/** - * \brief The tensor cpu evaluate api. - */ -template -inline void TensorCpuApply(LeftType& lhs, const RightType& rhs) { - TensorApply lhs_(lhs); - TensorApply rhs_(rhs); - CHECK_EQ(lhs_.getWidth(), rhs_.getWidth()); - CHECK_EQ(lhs_.getHeight(), rhs_.getHeight()); - CHECK_EQ(lhs_.useGpu(), rhs_.useGpu()); - - int height = lhs_.getHeight(); - int width = lhs_.getWidth(); - if (lhs_.isContiguous() && rhs_.isContiguous()) { - int size = height * width; - for (int index = 0; index < size; index++) { - lhs_.applyRef(index) = rhs_.apply(index); - } - } else { - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - lhs_.applyRef(i, j) = rhs_.apply(i, j); - } - } - } -} - -#ifdef __NVCC__ -template -__global__ void TensorElementWiseOp(LeftType lhs, - RightType rhs, - const int border) { - const int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < border) { - lhs.applyRef(idx) = rhs.apply(idx); - } -} - -template -__global__ void TensorElementWiseOp(LeftType lhs, RightType rhs) { - const int colIdx = blockIdx.x * blockDim.x + threadIdx.x; - const int rowIdx = blockIdx.y * blockDim.y + threadIdx.y; - for (int i = rowIdx; i < lhs.getHeight(); i += gridDim.y * blockDim.y) { - for (int j = colIdx; j < lhs.getWidth(); j += gridDim.x * blockDim.x) { - lhs.applyRef(i, j) = rhs.apply(i, j); - } - } -} - -/** - * \brief The tensor gpu evaluate api. - */ -template -inline void TensorGpuApply(LeftType& lhs, const RightType& rhs) { - TensorApply lhs_(lhs); - TensorApply rhs_(rhs); - CHECK_EQ(lhs_.getWidth(), rhs_.getWidth()); - CHECK_EQ(lhs_.getHeight(), rhs_.getHeight()); - CHECK_EQ(lhs_.useGpu(), rhs_.useGpu()); - - int dimM = lhs_.getHeight(); - int dimN = lhs_.getWidth(); - - if (lhs_.isContiguous() && rhs_.isContiguous()) { - int size = dimM * dimN; - int blockSize = size <= 1024 ? size : 1024; - int gridSize = (size + 1024 - 1) / 1024; - TensorElementWiseOp<<>>( - lhs_, rhs_, size); - } else { - int blockSizeY = std::min(32, dimM); - int blockSizeX = (32 / blockSizeY) * 32; - int gridSizeX = std::min(32, (dimN + blockSizeX - 1) / blockSizeX); - int gridSizeY = std::min(32, (dimM + blockSizeY - 1) / blockSizeY); - dim3 threads(blockSizeX, blockSizeY); - dim3 grid(gridSizeX, gridSizeY); - TensorElementWiseOp<<>>(lhs_, rhs_); - } - - CHECK_SYNC("TensorGpuApply failed"); -} -#else -template -inline void TensorGpuApply(LeftType& lhs, RightType& rhs) { - LOG(FATAL) << "Since it is gcc compiled, " - "this calculation does not support GPU implementation."; -} -#endif - -} // namespace paddle diff --git a/paddle/math/TensorExpression.h b/paddle/math/TensorExpression.h deleted file mode 100644 index f6da9adfca50e49ca260e20313c8979a38e1b06b..0000000000000000000000000000000000000000 --- a/paddle/math/TensorExpression.h +++ /dev/null @@ -1,446 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include "hl_tensor_ops.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -template -class TensorConstant; -template -class TensorUnaryOp; -template -class TensorBinaryOp; -template -class TensorTernaryOp; - -template -class TensorAssignOp; - -/** - * \brief Tensor base class. - * - * This is the base class of all Tensor and Expression class. - */ -template -class TensorExpression { - public: - /** - * Element wise unary expression. - */ - template - const TensorUnaryOp unaryExpression( - const UnaryOp& op) const { - return TensorUnaryOp(op, derived()); - } - - const TensorUnaryOp, const Derived, T> operator+( - T p) const { - return unaryExpression(hppl::unary::add_scale(p)); - } - - const TensorUnaryOp, const Derived, T> operator-( - T p) const { - return unaryExpression(hppl::unary::sub_scale(p)); - } - - const TensorUnaryOp, const Derived, T> operator*( - T p) const { - return unaryExpression(hppl::unary::mul_scale(p)); - } - - const TensorUnaryOp, const Derived, T> operator/( - T p) const { - return unaryExpression(hppl::unary::div_scale(p)); - } - - const TensorUnaryOp, const Derived, T> operator-() const { - return unaryExpression(hppl::unary::neg()); - } - - const TensorUnaryOp, const Derived, T> exp() const { - return unaryExpression(hppl::unary::exp_op()); - } - - const TensorUnaryOp, const Derived, T> log() const { - return unaryExpression(hppl::unary::log_op()); - } - - const TensorUnaryOp, const Derived, T> sqrt() const { - return unaryExpression(hppl::unary::sqrt_op()); - } - - const TensorUnaryOp, const Derived, T> square() const { - return unaryExpression(hppl::unary::square()); - } - - const TensorUnaryOp, const Derived, T> reciprocal() - const { - return unaryExpression(hppl::unary::reciprocal()); - } - - const TensorUnaryOp, const Derived, T> abs() const { - return unaryExpression(hppl::unary::abs()); - } - - const TensorUnaryOp, const Derived, T> sign() const { - return unaryExpression(hppl::unary::sign()); - } - - const TensorUnaryOp, const Derived, T> pow(T p) const { - return unaryExpression(hppl::unary::pow_op(p)); - } - - const TensorUnaryOp, const Derived, T> min(T p) const { - return unaryExpression(hppl::unary::min(p)); - } - - const TensorUnaryOp, const Derived, T> max(T p) const { - return unaryExpression(hppl::unary::max(p)); - } - - const TensorUnaryOp, const Derived, T> operator==( - T p) const { - return unaryExpression(hppl::unary::cmp_eq(p)); - } - - const TensorUnaryOp, const Derived, T> operator!=( - T p) const { - return unaryExpression(hppl::unary::cmp_ne(p)); - } - - const TensorUnaryOp, const Derived, T> operator<=( - T p) const { - return unaryExpression(hppl::unary::cmp_le(p)); - } - - const TensorUnaryOp, const Derived, T> operator<( - T p) const { - return unaryExpression(hppl::unary::cmp_lt(p)); - } - - const TensorUnaryOp, const Derived, T> operator>=( - T p) const { - return unaryExpression(hppl::unary::cmp_ge(p)); - } - - const TensorUnaryOp, const Derived, T> operator>( - T p) const { - return unaryExpression(hppl::unary::cmp_gt(p)); - } - - const TensorUnaryOp, const Derived, T> operator&&( - T p) const { - return unaryExpression(hppl::unary::and_op(p)); - } - - const TensorUnaryOp, const Derived, T> operator||( - T p) const { - return unaryExpression(hppl::unary::or_op(p)); - } - - /** - * Element wise binary expression. - */ - template - const TensorBinaryOp - binaryExpression(const BinaryOp& op, const ExpressionType& expr) const { - return TensorBinaryOp( - op, derived(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator==(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::cmp_eq(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator!=(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::cmp_ne(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator<=(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::cmp_le(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator<(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::cmp_lt(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator>=(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::cmp_ge(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator>(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::cmp_gt(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator&&(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::and_op(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator||(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::or_op(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator+(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::add(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator-(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::sub(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator*(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::mul(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - operator/(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::div(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - min(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::min(), expr); - } - - template - const TensorBinaryOp, - const Derived, - const ExpressionType, - T> - max(const ExpressionType& expr) const { - return binaryExpression(hppl::binary::max(), expr); - } - - /** - * Element wise ternary expression. - * - * ternary conditional operator(?: operator). - * The conditional expression returns one of two values depending on - * the result of derived expression. - * If derived expression evaluates to true, then expression1 is evaluated. - * If derived expression evaluates to false, then expression2 is evaluated. - */ - template - const TensorTernaryOp - condition(const ExprType1& expr1, const ExprType2& expr2) const { - return TensorTernaryOp( - derived(), expr1, expr2); - } - - template - const TensorTernaryOp< - const Derived, - const TensorConstant, const Derived, T>, - const ExprType, - T> - condition(T p, const ExprType& expr) const { - return condition(constant(p), expr); - } - - template - const TensorTernaryOp< - const Derived, - const ExprType, - const TensorConstant, const Derived, T>, - T> - condition(const ExprType& expr, T p) const { - return condition(expr, constant(p)); - } - - const TensorTernaryOp< - const Derived, - const TensorConstant, const Derived, T>, - const TensorConstant, const Derived, T>, - T> - condition(T p1, T p2) const { - return condition(constant(p1), constant(p2)); - } - - /** - * return a TensorConstant. A TensorConstant object hold a constant value. - */ - const TensorConstant, const Derived, T> constant( - T p) const { - return TensorConstant, const Derived, T>( - hppl::unary::constant(p), derived()); - } - - /** - * return a TensorAssignOp, and use AssignEvaluate to evaluate one or more - * TensorAssignOp objects. - */ - template - TensorAssignOp lazyAssign( - const ExpressionType& expr) const { - return TensorAssignOp(derived(), expr); - } - - protected: - const Derived& derived() const { return *static_cast(this); } -}; - -/** - * \brief Unary Operator Expression - */ -template -class TensorUnaryOp - : public TensorExpression, T> { - public: - explicit TensorUnaryOp(const OP op, const ExprType& expr) - : op_(op), expr_(expr) {} - - const OP op_; - const ExprType expr_; -}; - -/** - * \brief Binary Operator Expression - */ -template -class TensorBinaryOp - : public TensorExpression, T> { - public: - explicit TensorBinaryOp(const OP op, const LhsType& lhs, const RhsType& rhs) - : op_(op), lhs_(lhs), rhs_(rhs) {} - - const OP op_; - const LhsType lhs_; - const RhsType rhs_; -}; - -/** - * \brief Ternary Operator Expression - */ -template -class TensorTernaryOp : public TensorExpression< - TensorTernaryOp, - T> { - public: - explicit TensorTernaryOp(const ExprType1& expr1, - const ExprType2& expr2, - const ExprType3& expr3) - : expr1_(expr1), expr2_(expr2), expr3_(expr3) {} - - const ExprType1 expr1_; - const ExprType2 expr2_; - const ExprType3 expr3_; -}; - -/** - * \brief Constant Expression - */ -template -class TensorConstant - : public TensorExpression, T> { - public: - explicit TensorConstant(const OP op, const ExprType& expr) - : op_(op), expr_(expr) {} - - const OP op_; - const ExprType expr_; -}; - -/** - * \brief operator+ overload - * \return a unary operator expression - */ -template -const TensorUnaryOp, const Derived, T> operator+( - T p, const TensorExpression& expr) { - return expr + p; -} - -/** - * \brief operator* overload - * \return a unary operator expression - */ -template -const TensorUnaryOp, const Derived, T> operator*( - T p, const TensorExpression& expr) { - return expr * p; -} - -} // namespace paddle - -#include "TensorApply.h" -#include "TensorEvaluate.h" diff --git a/paddle/math/TrainingAlgorithmOp.cu b/paddle/math/TrainingAlgorithmOp.cu deleted file mode 100644 index b844768d3b9fd05b5a0eada5e315b9e91588a4ee..0000000000000000000000000000000000000000 --- a/paddle/math/TrainingAlgorithmOp.cu +++ /dev/null @@ -1,356 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "BaseMatrix.h" -#include "TrainingAlgorithmOp.h" -#include "paddle/utils/Logging.h" - -#if __cplusplus > 199711L - -#include "TensorAssign.h" - -namespace paddle { - -void sparseMomentumApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& momU, - BaseMatrix& momV, - real alpha, - real beta, - real gamma, - real tau, - real learningRate) { - auto expr1 = momU.lazyAssign(momU - (alpha * gamma * learningRate) * grad); - auto expr2 = - momV.lazyAssign(momV + (tau * alpha * gamma * learningRate) * grad); - auto expr3 = value.lazyAssign((tau / beta + (real)1 / alpha) * momU + - ((real)1 / beta) * momV); - - AssignEvaluate(expr1, expr2, expr3); -} - -void adadeltaApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum, - BaseMatrix& accum_update, - BaseMatrix& lr, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate) { - auto expr1 = accum.lazyAssign(rou * accum + ((real)1 - rou) * grad.square()); - auto expr2 = - lr.lazyAssign(((accum_update + epsilon) / (accum + epsilon)).sqrt()); - auto expr3 = accum_update.lazyAssign(rou * accum_update + - ((real)1 - rou) * (grad * lr).square()); - auto expr4 = mom.lazyAssign(mom * momentum - - learningRate * lr * (grad + value * decayRate)); - auto expr5 = value.lazyAssign(value + mom); - - AssignEvaluate(expr1, expr2, expr3, expr4, expr5); -} - -void adagradApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum_buffer, - BaseMatrix& accum, - BaseMatrix& lr, - real epsilon, - real learningRate, - real momentum, - real decayRate) { - auto expr1 = accum.lazyAssign(accum + grad.square()); - auto expr2 = - lr.lazyAssign((accum_buffer + accum + epsilon).sqrt().reciprocal()); - auto expr3 = mom.lazyAssign(mom * momentum - - learningRate * lr * (grad + value * decayRate)); - auto expr4 = value.lazyAssign(value + mom); - - AssignEvaluate(expr1, expr2, expr3, expr4); -} - -void rmspropApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& g, - BaseMatrix& f, - BaseMatrix& lr, - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime) { - auto expr2 = f.lazyAssign(accumulatedRou * f + ((real)1 - rou) * grad); - auto expr3 = lr.lazyAssign((g - f.square() + epsilon).sqrt().reciprocal()); - auto expr4 = mom.lazyAssign(mom * momentum - - learningRate * lr * (grad + value * decayRate)); - auto expr5 = value.lazyAssign(value + mom); - - if (firstTime) { - auto expr1 = g.lazyAssign(accumulatedRou * g + grad.square()); - - AssignEvaluate(expr1, expr2, expr3, expr4, expr5); - } else { - auto expr1 = - g.lazyAssign(accumulatedRou * g + ((real)1 - rou) * grad.square()); - - AssignEvaluate(expr1, expr2, expr3, expr4, expr5); - } -} - -void decayedAdagradApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum, - BaseMatrix& lr, - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime) { - auto expr2 = lr.lazyAssign((accum + epsilon).sqrt().reciprocal()); - auto expr3 = mom.lazyAssign(mom * momentum - - learningRate * lr * (grad + value * decayRate)); - auto expr4 = value.lazyAssign(value + mom); - - if (firstTime) { - auto expr1 = accum.lazyAssign(accumulatedRou * accum + grad.square()); - - AssignEvaluate(expr1, expr2, expr3, expr4); - } else { - auto expr1 = accum.lazyAssign(accumulatedRou * accum + - ((real)1 - rou) * grad.square()); - - AssignEvaluate(expr1, expr2, expr3, expr4); - } -} - -void adamApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, // firse moment - BaseMatrix& v, // second moment - real beta1, - real beta2, - real beta1_power, - real beta2_power, - real epsilon, - real learningRate) { - real alpha = - learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); - - auto expr1 = mom.lazyAssign(beta1 * mom + ((real)1 - beta1) * grad); - auto expr2 = v.lazyAssign(beta2 * v + ((real)1 - beta2) * grad.square()); - auto expr3 = value.lazyAssign(value - (mom * alpha) / (v.sqrt() + epsilon)); - - AssignEvaluate(expr1, expr2, expr3); -} - -void adamaxApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, // firse moment - BaseMatrix& u, // weighted infinity norm - real beta1, - real beta2, - int64_t step, - real alpha) { - auto expr1 = mom.lazyAssign(beta1 * mom + ((real)1 - beta1) * grad); - auto expr2 = - u.lazyAssign((beta2 * u > grad.abs()).condition(beta2 * u, grad.abs())); - auto expr3 = value.lazyAssign( - value - (alpha / ((real)1 - (real)std::pow(beta1, step))) * (mom / u)); - - AssignEvaluate(expr1, expr2, expr3); -} - -} // namespace paddle - -#else - -namespace paddle { - -void sparseMomentumApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& momU, - BaseMatrix& momV, - real alpha, - real beta, - real gamma, - real tau, - real learningRate) { - /** - * \alpha_t = \alpha_{t-1} / k - * \beta_t = \beta_{t-1} / (1 + \lambda\gamma_t) - * u_t = u_{t-1} - \alpha_t \gamma_t g_t - * v_t = v_{t-1} + \tau_{t-1} \alpha_t \gamma_t g_t - * \tau_t = \tau_{t-1} + \beta_t / \alpha_t - */ - momU -= (alpha * gamma * learningRate) * grad; - momV += (tau * alpha * gamma * learningRate) * grad; - value = (tau / beta + (real)1 / alpha) * momU + ((real)1 / beta) * momV; -} - -void adadeltaApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum, - BaseMatrix& accum_update, - BaseMatrix& lr, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate) { - // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 - accum = rou * accum + ((real)1 - rou) * grad.square(); - - // learn_rate: sqrt(( E(dx_{t-1}^2) + epsilon ) / ( E(g_t^2) + epsilon )) - lr = ((accum_update + epsilon) / (accum + epsilon)).sqrt(); - - // E(dx_t^2) = \rou * E(dx_{t-1}^2) + (1-\rou) * (-g*learn_rate)^2 - accum_update = rou * accum_update + ((real)1 - rou) * (grad * lr).square(); - - mom = mom * momentum - learningRate * lr * (grad + value * decayRate); - value += mom; -} - -void adagradApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum_buffer, - BaseMatrix& accum, - BaseMatrix& lr, - real epsilon, - real learningRate, - real momentum, - real decayRate) { - accum += grad.square(); - lr = (accum_buffer + accum + epsilon).sqrt().reciprocal(); - mom = mom * momentum - learningRate * lr * (grad + value * decayRate); - value += mom; -} - -void rmspropApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& g, - BaseMatrix& f, - BaseMatrix& lr, - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime) { - // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 - // For the first time update, make the sum be the current square - // so that the initial estimation of E(g_t^2) will not be too small. - if (firstTime) { - g = accumulatedRou * g + grad.square(); - } else { - g = accumulatedRou * g + ((real)1 - rou) * grad.square(); - } - - // E(f_t) = \rou * E(f_{t-1}) + (1-\rou) * g - f = accumulatedRou * f + ((real)1 - rou) * grad; - - // learn_rate = 1/sqrt( ( E(g_t^2) - (E(f_t))^2 + epsilon ) - // Basiclly if the sign of the gradient changes more often, - // the learning rate will be decreased. - lr = (g - f.square() + epsilon).sqrt().reciprocal(); - - mom = mom * momentum - learningRate * lr * (grad + value * decayRate); - value += mom; -} - -void decayedAdagradApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum, - BaseMatrix& lr, - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime) { - // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 - // For the first time update, make the sum be the current square - // so that the initial estimation of E(g_t^2) will not be too small. - if (firstTime) { - accum = accumulatedRou * accum + grad.square(); - } else { - accum = accumulatedRou * accum + ((real)1 - rou) * grad.square(); - } - - // learn_rate = 1/sqrt( ( E(g_t^2) + epsilon ) - // Basiclly if the bigger the magnitude gradient is, - // the smaller the learning rate will be. - lr = (accum + epsilon).sqrt().reciprocal(); - - mom = mom * momentum - learningRate * lr * (grad + value * decayRate); - value += mom; -} - -void adamApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, // firse moment - BaseMatrix& v, // second moment - real beta1, - real beta2, - real beta1_power, - real beta2_power, - real epsilon, - real learningRate) { - real alpha = - learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); - - // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; - mom = beta1 * mom + ((real)1 - beta1) * grad; - - // v_t = \beta_2 * v_{t-1} + (1-\beta_2)* g_{t-1}^2 - v = beta2 * v + ((real)1 - beta2) * grad.square(); - - value -= (mom * alpha) / (v.sqrt() + epsilon); -} - -void adamaxApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, // firse moment - BaseMatrix& u, // weighted infinity norm - real beta1, - real beta2, - int64_t step, - real alpha) { - // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; - mom = beta1 * mom + ((real)1 - beta1) * grad; - - // u_t = max(\beta_2*u_{t-1}, abs(g_t)) - u = (beta2 * u > grad.abs()).condition(beta2 * u, grad.abs()); - - // \theta_t = \theta_{t-1} - (\alpha/(1-\beta_1^t))*m_t/u_t - value -= (alpha / ((real)1 - (real)std::pow(beta1, step))) * (mom / u); -} - -} // namespace paddle - -#endif diff --git a/paddle/math/TrainingAlgorithmOp.h b/paddle/math/TrainingAlgorithmOp.h deleted file mode 100644 index fe40fc2d36e796bd4be7b7fc1e12a6eafa5d4700..0000000000000000000000000000000000000000 --- a/paddle/math/TrainingAlgorithmOp.h +++ /dev/null @@ -1,122 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "BaseMatrix.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -/** - * \brief Sparse Momentum optimizer. - */ -extern void sparseMomentumApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& momU, - BaseMatrix& momV, - real alpha, - real beta, - real gamma, - real tau, - real learningRate); - -/** - * \brief AdaDelta optimizer. - */ -extern void adadeltaApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& sum, - BaseMatrix& sum1, - BaseMatrix& mom, - BaseMatrix& lr, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate); - -/** - * \brief AdaGrad optimizer. - */ -extern void adagradApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& sum, - BaseMatrix& sum1, - BaseMatrix& mom, - BaseMatrix& lr, - real epsilon, - real learningRate, - real momentum, - real decayRate); - -/** - * \brief RMSProp optimizer. - */ -extern void rmspropApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& g, - BaseMatrix& f, - BaseMatrix& mom, - BaseMatrix& lr, - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime); - -/** - * \brief Decayed AdaGrad optimizer. - */ -extern void decayedAdagradApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& accum, - BaseMatrix& lr, - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime); - -/** - * \brief Adam optimizer. - */ -extern void adamApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, - BaseMatrix& v, - real beta1, - real beta2, - real beta1_power, - real beta2_power, - real epsilon, - real learningRate); - -/** - * \brief AdaMax optimizer. - */ -extern void adamaxApply(BaseMatrix& value, - BaseMatrix& grad, - BaseMatrix& mom, // firse moment - BaseMatrix& u, // weighted infinity norm - real beta1, - real beta2, - int64_t step, - real alpha); -} // namespace paddle diff --git a/paddle/math/Vector.cpp b/paddle/math/Vector.cpp deleted file mode 100644 index 2a47ed7ef81a2e969757c244370cc346b13e1c03..0000000000000000000000000000000000000000 --- a/paddle/math/Vector.cpp +++ /dev/null @@ -1,1091 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Vector.h" -#include "paddle/utils/Util.h" - -#include -#include "Matrix.h" -#include "hl_gpu.h" -#include "hl_matrix.h" -#include "hl_table_apply.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Thread.h" -#include "paddle/utils/ThreadLocal.h" - -namespace paddle { - -template -std::shared_ptr> VectorT::create(size_t size, bool useGpu) { - if (useGpu) { - return std::make_shared>(size); - } else { - return std::make_shared>(size); - } -} - -template -std::shared_ptr> VectorT::createParallelVector( - size_t size, bool useGpu, SyncThreadPool* pool) { - if (!useGpu && FLAGS_trainer_count > 1 && FLAGS_enable_parallel_vector && - size >= (size_t)FLAGS_enable_parallel_vector) { - return std::make_shared>( - size, pool ? pool : getGlobalSyncThreadPool()); - } else { - return create(size, useGpu); - } -} - -template -std::shared_ptr> VectorT::create(T* data, - size_t size, - bool useGpu) { - if (useGpu) { - return std::make_shared>(size, data); - } else { - return std::make_shared>(size, data); - } -} - -template -std::shared_ptr> VectorT::create(size_t size, - MemoryHandlePtr memoryHandle, - size_t offset) { - if (auto cpuMemHandle = - std::dynamic_pointer_cast(memoryHandle)) { - return std::make_shared>(size, cpuMemHandle, offset); - } else if (auto gpuMemHandle = - std::dynamic_pointer_cast(memoryHandle)) { - return std::make_shared>(size, gpuMemHandle, offset); - } else { - LOG(FATAL) << "Wrong"; - return NULL; - } -} - -template <> -MatrixPtr VectorT::toOneHotSparseMatrix(size_t idRange, bool useGpu) { - LOG(FATAL) << "Wrong for real vector"; - return nullptr; -} - -template <> -MatrixPtr VectorT::toOneHotSparseMatrix(size_t idRange, bool useGpu) { - size_t height = getSize(); - size_t width = idRange; - MatrixPtr mat = Matrix::createSparseMatrix( - height, idRange, height, NO_VALUE, SPARSE_CSR, false, useGpu); - - CpuIVector cpuIds(height); - cpuIds.copyFrom(*this); - int* idData = cpuIds.getData(); - - for (decltype(height) i = 0; i < height; i++) { - const unsigned int id = idData[i]; - CHECK_LT(id, width); - mat->setRow(i, 1, &id, nullptr); - } - return mat; -} - -template <> -std::shared_ptr> VectorT::castToInt() { - std::shared_ptr> ret = IVector::create(this->getSize(), useGpu_); - if (useGpu_) { - hl_vector_cast2int(ret->getData(), this->getData(), this->getSize()); - } else { - for (size_t i = 0; i < getSize(); ++i) { - ret->getData()[i] = int(this->getData()[i]); - } - } - return ret; -} - -template -GpuVectorT::GpuVectorT(size_t size) - : VectorT(size, - std::make_shared(sizeof(T) * size), - 0, /* offset = 0 */ - true /* useGpu = true */) {} - -template -T GpuVectorT::getElement(size_t i) const { - T elem = 0; - hl_memcpy_device2host(&elem, const_cast(&this->getData()[i]), sizeof(T)); - return elem; -} -template -void GpuVectorT::setElement(size_t i, const T& value) { - hl_memcpy_host2device(&this->getData()[i], const_cast(&value), sizeof(T)); -} - -template -T* GpuVectorT::getPoint(const uint64_t beginPos) { - LOG(FATAL) << "Not implemented" << beginPos; - return NULL; -} - -template <> -int GpuVectorT::getAbsSum() { - LOG(FATAL) << "Not implemented"; - return 0; -} - -template <> -int GpuVectorT::getSum() { - LOG(FATAL) << "Not implemented"; - return 0; -} - -template <> -real GpuVectorT::getAbsSum() { - real* A = this->getData(); - real sum = 0; - hl_vector_abs_sum(A, &sum, this->getSize()); - return sum; -} - -template <> -real GpuVectorT::getSum() { - real* A = this->getData(); - real sum = 0; - hl_vector_sum(A, &sum, this->getSize()); - return sum; -} - -template <> -int GpuVectorT::getMax() { - CpuIVector cpuIVec = CpuIVector(this->getSize()); - copyTo(&cpuIVec); - return cpuIVec.getMax(); -} - -template <> -int GpuVectorT::getAbsMax() { - CpuIVector cpuIVec = CpuIVector(this->getSize()); - copyTo(&cpuIVec); - return cpuIVec.getAbsMax(); -} - -template -void GpuVectorT::isEqualTo(const VectorT& b, const T& value) { - BaseMatrixT::isEqualTo((BaseMatrixT&)b, value); -} - -template -void GpuVectorT::selectFrom(const VectorT& src, const VectorT& ids) { -#ifdef PADDLE_WITH_CUDA - hl_vector_select_from(this->getData(), - this->getSize(), - src.getData(), - src.getSize(), - ids.getData(), - ids.getSize()); -#endif -} - -template -real gpuRowFunc(Func f, GpuVector& v) { - static ThreadLocal>> local; - if (!*local) { - (*local).reset(new CpuVector(1)); - } - real* A = v.getData(); - f(A, (*local)->getData(), 1, v.getSize()); - return (*local)->getData()[0]; -} - -template <> -real GpuVectorT::getMax() { - return gpuRowFunc(hl_matrix_row_max, *this); -} - -template <> -real GpuVectorT::getAbsMax() { - return std::max(gpuRowFunc(hl_matrix_row_max, *this), - -gpuRowFunc(hl_matrix_row_min, *this)); -} - -template <> -int GpuVectorT::getMin() { - LOG(FATAL) << "Not implemented"; - return 0; -} - -template <> -real GpuVectorT::getMin() { - return gpuRowFunc(hl_matrix_row_min, *this); -} - -template -T GpuVectorT::get(size_t pos) { - T val = (T)0; - hl_memcpy_device2host((void*)&val, (void*)(this->getData() + pos), sizeof(T)); - return val; -} - -template -void GpuVectorT::histogram(std::ostream& os, int type) { - LOG(FATAL) << "Not implemented"; -} - -template -void GpuVectorT::zeroMem() { - BaseMatrixT::zero(); -} - -template -void GpuVectorT::reset(const T& value) { - BaseMatrixT::assign(value); -} - -template -void GpuVectorT::fillSequence() { - LOG(FATAL) << "not implemented"; -} - -template -void GpuVectorT::copyFrom(const VectorT& src) { - src.copyTo(this); -} - -template -void GpuVectorT::copyFrom(const VectorT& src, hl_stream_t stream) { - CHECK_EQ(src.getSize(), this->getSize()); - hl_memcpy_async((void*)this->getData(), - (void*)src.getData(), - sizeof(T) * this->getSize(), - stream); -} - -template -void GpuVectorT::copyFrom(const T* gpuSrc, size_t size) { - CHECK(gpuSrc != NULL); - CHECK_LE(size, this->size_); - - hl_memcpy((void*)this->getData(), (void*)gpuSrc, sizeof(T) * size); -} - -template -void GpuVectorT::copyFrom(const T* gpuSrc, size_t size, hl_stream_t stream) { - CHECK(gpuSrc != NULL); - CHECK_LE(size, this->size_); - - hl_memcpy_async( - (void*)this->getData(), (void*)gpuSrc, sizeof(T) * size, stream); -} - -template -void GpuVectorT::copyTo(CpuVectorT* dest) const { - CHECK_EQ(this->getSize(), dest->getSize()); - - hl_memcpy_device2host((void*)dest->getData(), - (void*)this->getData(), - sizeof(T) * this->getSize()); -} - -template -void GpuVectorT::copyTo(GpuVectorT* dest) const { - CHECK_EQ(this->getSize(), dest->getSize()); - - hl_memcpy_device2device((void*)dest->getData(), - (void*)this->getData(), - sizeof(T) * this->getSize()); -} - -template <> -void GpuVectorT::rand() { - LOG(FATAL) << "Not implemented"; -} - -template <> -void GpuVectorT::print(std::ostream& os, size_t num) const { - IVectorPtr dest = IVector::create(this->size_, false); - hl_memcpy_device2host((void*)dest->getData(), - (void*)this->getData(), - sizeof(int) * this->getSize()); - dest->print(os, num); -} - -template <> -void GpuVectorT::print(std::ostream& os, size_t num) const { - VectorPtr dest = Vector::create(this->size_, false); - hl_memcpy_device2host((void*)dest->getData(), - (void*)this->getData(), - sizeof(int) * this->getSize()); - dest->print(os, num); -} - -template <> -void GpuVectorT::printOneElement(std::ostream& os, size_t idx) const { - LOG(FATAL) << "Not implemented"; -} - -template <> -void GpuVectorT::printOneElement(std::ostream& os, size_t idx) const { - LOG(FATAL) << "Not implemented"; -} - -template <> -void CpuVectorT::rand() { - LOG(FATAL) << "Not implemented"; -} -template <> -void GpuVectorT::rand(size_t classNum) { - LOG(FATAL) << "Not implemented"; -} - -template <> -void CpuVectorT::rand(size_t classNum) { - LOG(FATAL) << "Not implemented"; -} - -template <> -void GpuVectorT::rand() { - VectorPtr cPtr = Vector::create(this->size_, false); - cPtr->rand(); - - hl_memcpy_host2device(data_, cPtr->getData(), this->size_ * sizeof(real)); -} - -template <> -void GpuVectorT::rand(size_t classNum) { - IVectorPtr cPtr = IVector::create(this->size_, false); - cPtr->rand(classNum); - - hl_memcpy_host2device(data_, cPtr->getData(), this->size_ * sizeof(int)); -} - -template <> -void CpuVectorT::rand(size_t classNum) { - size_t size = this->getSize(); - int* data = this->getData(); - for (size_t i = 0; i < size; i++) { - data[i] = - std::min(classNum - 1, - size_t(::rand() * (1. / ((double)RAND_MAX + 1)) * classNum)); - } -} - -template <> -void CpuVectorT::rand() { - size_t size = this->getSize(); - real* data = this->getData(); - for (size_t i = 0; i < size; i++) { - data[i] = ::rand() * (1. / (double)RAND_MAX); - // data[ii] = ((temp > RAND_MAX/2)? 1 : -1) * - // sqrt( abs((temp-RAND_MAX/2))/(double(RAND_MAX))/2048 ); - } -} - -template -void CpuVectorT::randnorm(real, real) { - LOG(FATAL) << "Not implemented"; -} - -template -void CpuVectorT::uniform(real, real) { - LOG(FATAL) << "Not implemented"; -} - -template -void GpuVectorT::randnorm(real, real) { - LOG(FATAL) << "Not implemented"; -} - -template -void GpuVectorT::uniform(real, real) { - LOG(FATAL) << "Not implemented"; -} - -template <> -void CpuVectorT::randnorm(real mean, real std) { - size_t size = this->getSize(); - real* data = this->getData(); - unsigned int* seed = ThreadLocalRand::getSeed(); - auto rand1 = [&]() { return (1. + ::rand_r(seed)) * (1. / (1. + RAND_MAX)); }; - for (size_t i = 0; i < size - 1; i += 2) { - real r1 = rand1(); - r1 = std::sqrt(-2 * std::log(r1)); - real r2 = rand1(); - data[i] = mean + std * r1 * cos(2 * M_PI * r2); - data[i + 1] = mean + std * r1 * sin(2 * M_PI * r2); - } - real r1 = rand1(); - r1 = std::sqrt(-2 * std::log(r1)); - real r2 = rand1(); - data[size - 1] = mean + std * r1 * cos(2 * M_PI * r2); -} - -template <> -void CpuVectorT::uniform(real left, real right) { - size_t size = this->getSize(); - real* data = this->getData(); - real range = right - left; - unsigned int* seed = ThreadLocalRand::getSeed(); - auto rand1 = [&]() { return ::rand_r(seed) * (1. / (1. + RAND_MAX)); }; - for (size_t i = 0; i < size; ++i) { - data[i] = rand1() * range + left; - } -} - -template <> -void GpuVectorT::randnorm(real mean, real std) { - CpuVector cpuVec = CpuVector(this->getSize()); - cpuVec.randnorm(mean, std); - - hl_memcpy_host2device( - data_, cpuVec.getData(), this->getSize() * sizeof(real)); -} - -template <> -void GpuVectorT::uniform(real left, real right) { - CpuVector cpuVec = CpuVector(this->getSize()); - cpuVec.uniform(left, right); - - hl_memcpy_host2device( - data_, cpuVec.getData(), this->getSize() * sizeof(real)); -} - -template -CpuVectorT::CpuVectorT(size_t size) - : VectorT(size, - std::make_shared(sizeof(T) * size), - 0, /* offset = 0 */ - false /* useGpu = false */) {} - -template -CpuVectorT::CpuVectorT(const VectorT& src) - : VectorT(src.getSize(), - src.getMemoryHandle(), - 0, /* offset = 0 */ - false /* useGpu = false */) { - if (typeid(*this->memoryHandle_.get()) != typeid(CpuMemoryHandle)) { - this->memoryHandle_ = - std::make_shared(sizeof(T) * this->getSize()); - this->data_ = reinterpret_cast(this->memoryHandle_->getBuf()); - } - src.copyTo(this); -} - -template -T CpuVectorT::getAbsSum() { - const T* A = this->getData(); - size_t size = this->getSize(); - T sum = 0; - for (size_t i = 0; i < size; i++) { - sum += (A[i] > 0) ? A[i] : -A[i]; - } - return sum; -} - -// cannot use above version, due to precision issue of float -template <> -real CpuVectorT::getAbsSum() { - const real* A = this->getData(); - size_t size = this->getSize(); - double sum = 0; - for (size_t i = 0; i < size; i++) { - sum += (A[i] > 0) ? A[i] : -A[i]; - } - return sum; -} - -template -T CpuVectorT::getSum() { - const T* A = this->getData(); - size_t size = this->getSize(); - T sum = 0; - for (size_t i = 0; i < size; i++) { - sum += A[i]; - } - return sum; -} - -template <> -real CpuVectorT::getSum() { - const real* A = this->getData(); - size_t size = this->getSize(); - double sum = 0; - for (size_t i = 0; i < size; i++) { - sum += A[i]; - } - return sum; -} - -template -T CpuVectorT::get(size_t pos) { - return this->getData()[pos]; -} - -template -T CpuVectorT::getMax() { - const T* A = this->getData(); - size_t size = this->getSize(); - T res = A[0]; - for (size_t i = 1; i < size; i++) { - if (res < A[i]) res = A[i]; - } - return res; -} - -template -T CpuVectorT::getAbsMax() { - const T* A = this->getData(); - size_t size = this->getSize(); - T res = std::abs(A[0]); - for (size_t i = 1; i < size; i++) { - if (res < std::abs(A[i])) res = std::abs(A[i]); - } - return res; -} - -template -T CpuVectorT::getMin() { - const T* A = this->getData(); - size_t size = this->getSize(); - T res = A[0]; - for (size_t i = 1; i < size; i++) { - if (res > A[i]) res = A[i]; - } - return res; -} - -template -void CpuVectorT::isEqualTo(const VectorT& b, const T& value) { - size_t size = this->getSize(); - CHECK_EQ(b.getSize(), size); - - const T* B = b.getData(); - T* A = this->getData(); - for (size_t i = 0; i < size; i++) { - A[i] = (B[i] == value); - } -} - -template -void CpuVectorT::selectFrom(const VectorT& src, const VectorT& ids) { - size_t size = this->getSize(); - CHECK_EQ(ids.getSize(), size); - - const int* indices = ids.getData(); - const T* B = src.getData(); - T* A = this->getData(); - for (size_t i = 0; i < size; i++) { - int index = indices[i]; - CHECK_LT(index, (int)src.getSize()); - A[i] = B[index]; - } -} - -static int getSignAndExponentOfFloat(float a) { - uint32_t* pa = reinterpret_cast(&a); - return *pa >> 23; -} - -template -void CpuVectorT::histogram(std::ostream& os, int type) { - LOG(FATAL) << "Not implemented"; -} - -template <> -void CpuVectorT::histogram(std::ostream& os, int type) { - int counters[512]; - memset(counters, 0, sizeof(counters)); - int counterZero = 0; - - const real* A = this->getData(); - size_t size = this->getSize(); - for (size_t i = 0; i < size; i++) { - if (A[i] == 0.0f) { - ++counterZero; - } else { - ++counters[getSignAndExponentOfFloat(A[i])]; - } - } - - int64_t sum = 0; - float sizeNonZero = size - counterZero; - os << "zero:" << counterZero; - for (int i = 0; i < 256; i++) { - int counter = counters[i]; - if (counter) { - os << " 2^" << i - 127 << ":" << counter / sizeNonZero * 100 << "%"; - sum += counter * (i - 127); - } - } - for (int i = 0; i < 256; i++) { - int counter = counters[i + 256]; - if (counter) { - os << " -2^" << i - 127 << ":" << counter / sizeNonZero * 100 << "%"; - sum += counter * (i - 127); - } - } - os << ", nonzero_exponent_avg=" << sum / sizeNonZero; -} - -template -void CpuVectorT::zeroMem() { - memset(this->getData(), 0, sizeof(T) * this->getSize()); -} - -template -void CpuVectorT::reset(const T& value) { - T* A = this->getData(); - size_t size = this->getSize(); - for (size_t i = 0; i < size; i++) { - A[i] = value; - } -} - -template -void CpuVectorT::fillSequence() { - T* A = this->getData(); - size_t size = this->getSize(); - for (size_t i = 0; i < size; i++) { - A[i] = i; - } -} - -template -void CpuVectorT::copyFrom(const VectorT& src) { - src.copyTo(this); -} - -template -void CpuVectorT::copyFrom(const VectorT& src, hl_stream_t stream) { - if (typeid(src) == typeid(GpuVectorT)) { - hl_memcpy_async((void*)this->getData(), - (void*)src.getData(), - sizeof(T) * this->getSize(), - stream); - // There is a need to add synchronization to ensure that the data is copied. - hl_stream_synchronize(stream); - } else { - src.copyTo(this); - } -} - -template -void CpuVectorT::copyFrom(const T* hostSrc, size_t size) { - CHECK(hostSrc != NULL); - CHECK_LE(size, this->size_); - memcpy(this->data_, hostSrc, sizeof(T) * size); -} - -template -void CpuVectorT::copyFrom(const T* hostSrc, - size_t size, - hl_stream_t stream) { - (void)stream; - - CHECK(hostSrc != NULL); - CHECK_LE(size, this->size_); - memcpy(this->data_, hostSrc, sizeof(T) * size); -} - -template -void CpuVectorT::copyTo(CpuVectorT* dest) const { - CHECK_EQ(this->getSize(), dest->getSize()); - memcpy(dest->getData(), this->getData(), sizeof(T) * this->getSize()); -} - -template -void CpuVectorT::copyTo(GpuVectorT* dest) const { - CHECK_EQ(this->getSize(), dest->getSize()); - hl_memcpy_host2device((void*)dest->getData(), - (void*)this->getData(), - sizeof(T) * this->getSize()); -} - -template <> -void CpuVectorT::print(std::ostream& os, size_t num) const { - size_t w = size_ < num ? size_ : num; - os << "["; - for (size_t i = 0; i < w; ++i) { - os << data_[i] << " "; - } - os << "]" << std::endl; -} - -template <> -void CpuVectorT::print(std::ostream& os, size_t num) const { - size_t w = size_ < num ? size_ : num; - os << "["; - for (size_t i = 0; i < w; ++i) { - os << (int)data_[i] << " "; - } - os << "]" << std::endl; -} - -template <> -void CpuVectorT::printOneElement(std::ostream& os, size_t idx) const { - CHECK_LT(idx, size_); - os << data_[idx] << ";"; -} - -template <> -void CpuVectorT::printOneElement(std::ostream& os, size_t idx) const { - CHECK_LT(idx, size_); - os << (int)data_[idx] << ";"; -} - -template -void ParallelCpuVectorT::parallelExec(ExecFunc func) { - LOG(FATAL) << "Not implemented"; -} - -template <> -void ParallelCpuVectorT::parallelExec(ExecFunc func) { - pool_->exec([this, func](int tid, size_t numThreads) { - auto interval = calcSplitArrayInterval( - this->getSize(), (size_t)tid, numThreads, 8LU /*for avx*/); - // setup sub bufs - CpuVector subVec(0, nullptr); - subVec.subVecFrom(*this, interval); - func(subVec); - }); -} - -template -void ParallelCpuVectorT::exec(SyncThreadPool::JobFunc func) { - LOG(FATAL) << "Not implemented"; -} - -template <> -void ParallelCpuVectorT::exec(SyncThreadPool::JobFunc func) { - pool_->exec(func); -} - -template -CpuGpuVectorT::CpuGpuVectorT(size_t size, bool useGpu) : sync_(nullptr) { - if (!useGpu) { - cpuVectorT_ = std::make_shared>(size); - } else { - gpuVectorT_ = std::make_shared>(size); - } - setSync(useGpu); -} - -template -CpuGpuVectorT::CpuGpuVectorT(const std::shared_ptr>& src) - : sync_(nullptr) { - bool useGpu = src->useGpu(); - if (useGpu) { - gpuVectorT_ = src; - } else { - cpuVectorT_ = src; - } - setSync(useGpu); -} - -template -CpuGpuVectorT::CpuGpuVectorT(size_t size, T* data, bool useGpu) - : sync_(nullptr) { - if (!useGpu) { - cpuVectorT_ = std::make_shared>(size, data); - setSync(DATA_AT_CPU); - } else { - gpuVectorT_ = std::make_shared>(size, data); - setSync(DATA_AT_GPU); - } -} - -template -std::shared_ptr> CpuGpuVectorT::create(size_t size, - bool useGpu) { - return std::make_shared>(size, useGpu); -} - -template -void CpuGpuVectorT::resize(size_t size, bool useGpu) { - if (useGpu) { - CHECK(gpuVectorT_) << "gpuVectorT_ is null"; - // If memoryHandle_ is nullptr, - // the data may be owned by the caller when it was constructed. - // It should not resize for this case. - if (gpuVectorT_->getMemoryHandle()) { - gpuVectorT_->resize(size); - } else { - CHECK_EQ(gpuVectorT_->getSize(), size); - } - } else { - CHECK(cpuVectorT_) << "cpuVectorT_ is null"; - // If memoryHandle_ is nullptr, - // the data may be owned by the caller when it was constructed. - // It should not resize for this case. - if (cpuVectorT_->getMemoryHandle()) { - cpuVectorT_->resize(size); - } else { - CHECK_EQ(cpuVectorT_->getSize(), size); - } - } - setSync(useGpu); -} - -template -void CpuGpuVectorT::resizeOrCreate(std::shared_ptr>& vec, - size_t size, - bool useGpu) { - if (vec) { - vec->resize(size, useGpu); - } else { - vec = create(size, useGpu); - } -} - -template -void CpuGpuVectorT::resizeOrCreate(size_t size, bool useGpu) { - if (useGpu && (!gpuVectorT_)) { - gpuVectorT_ = VectorT::create(size, true); - } else if ((!useGpu) && (!cpuVectorT_)) { - cpuVectorT_ = VectorT::create(size, false); - } else { - CHECK((useGpu && gpuVectorT_) || (!useGpu && cpuVectorT_)); - this->resize(size, useGpu); - } -} - -template -CpuGpuVectorT::CpuGpuVectorT(CpuGpuVectorT& src, - size_t offset, - size_t size) - : sync_(nullptr) { - CHECK_LE(offset + size, static_cast(src.getSize())); -#ifdef PADDLE_WITH_CUDA - SyncedFlag* flag = src.getSync(); - if (*flag == DATA_AT_CPU) { - src.copyToGpu(); // will set synchronous data between CPU and GPU - } else if (*flag == DATA_AT_GPU) { - src.copyToCpu(); // will set synchronous data between CPU and GPU - } -#endif - auto cMemHandle = (src.getVector(false))->getMemoryHandle(); - cpuVectorT_ = std::make_shared>( - size, std::dynamic_pointer_cast(cMemHandle), offset); -#ifdef PADDLE_WITH_CUDA - auto gMemHandle = (src.getVector(true))->getMemoryHandle(); - gpuVectorT_ = std::make_shared>( - size, std::dynamic_pointer_cast(gMemHandle), offset); - src.setSync(SYNCED); -#endif - setSync(src.getSync()); -} - -template -std::shared_ptr> CpuGpuVectorT::getVector( - bool useGpu) const { - auto* self = const_cast*>(this); - if (useGpu) { - self->copyToGpu(); - return std::const_pointer_cast>(gpuVectorT_); - } else { - self->copyToCpu(); - return std::const_pointer_cast>(cpuVectorT_); - } -} - -template -std::shared_ptr>& CpuGpuVectorT::getMutableVector(bool useGpu) { - setSync(useGpu); - if (useGpu) { - copyToGpu(); - return gpuVectorT_; - } else { - copyToCpu(); - return cpuVectorT_; - } -} - -template -const T* CpuGpuVectorT::getData(bool useGpu) const { - auto self = const_cast*>(this); - if (useGpu) { - self->copyToGpu(); - return gpuVectorT_->getData(); - } else { - self->copyToCpu(); - return cpuVectorT_->getData(); - } -} - -// Operation will change data and need to reset sync_ & syncFlag_. -#define MUTABLE_VECTOR_OP(OP, useGpu, args...) \ - do { \ - if (useGpu) { \ - copyToGpu(); \ - setSync(useGpu); \ - return gpuVectorT_->OP(args); \ - } else { \ - copyToCpu(); \ - setSync(useGpu); \ - return cpuVectorT_->OP(args); \ - } \ - } while (0) - -template -T* CpuGpuVectorT::getMutableData(bool useGpu) { - MUTABLE_VECTOR_OP(getData, useGpu); -} - -template -void CpuGpuVectorT::zeroMem(bool useGpu) { - MUTABLE_VECTOR_OP(zeroMem, useGpu); -} - -template -void CpuGpuVectorT::fillSequence(bool useGpu) { - MUTABLE_VECTOR_OP(fillSequence, useGpu); -} - -template -void CpuGpuVectorT::setElement(size_t i, const T& value, bool useGpu) { - MUTABLE_VECTOR_OP(setElement, useGpu, i, value); -} - -template -T CpuGpuVectorT::getElement(size_t i) const { - switch (*this->getSync()) { - case SYNCED: - case DATA_AT_CPU: - return cpuVectorT_->getElement(i); - break; - case DATA_AT_GPU: - return gpuVectorT_->getElement(i); - break; - default: - LOG(FATAL) << "Not support"; - break; - } -} - -template -void CpuGpuVectorT::copyFrom(const VectorT& src, hl_stream_t stream) { - auto cVec = dynamic_cast*>(&src); - auto gVec = dynamic_cast*>(&src); - if (cVec) { - copyToCpu(cVec->getData(), cVec->getSize(), stream); - } else if (gVec) { - copyToGpu(gVec->getData(), gVec->getSize(), stream); - } else { - LOG(FATAL) << "Invalid type of src"; - } -} - -template -void CpuGpuVectorT::copyFrom(const T* data, size_t size, bool useGpu) { - if (useGpu) { - copyToGpu(data, size); - } else { - copyToCpu(data, size); - } -} - -template -void CpuGpuVectorT::copyFrom(const T* data, - size_t size, - hl_stream_t stream, - bool useGpu) { - if (useGpu) { - copyToGpu(data, size, stream); - } else { - copyToCpu(data, size, stream); - } -} - -template -void CpuGpuVectorT::copyFrom(CpuGpuVectorT& src, - size_t offset, - size_t size, - bool useGpu, - hl_stream_t stream) { - if (useGpu) { - VectorT::resizeOrCreate(gpuVectorT_, size, true); - gpuVectorT_->copyFrom(src.getData(true) + offset, size, stream); - } else { - VectorT::resizeOrCreate(cpuVectorT_, size, false); - cpuVectorT_->copyFrom(src.getData(false) + offset, size, stream); - } - setSync(useGpu); -} - -template -void CpuGpuVectorT::copyFrom(CpuGpuVectorT& src, hl_stream_t stream) { - switch (*src.getSync()) { - case DATA_AT_CPU: - copyFrom(*(src.getVector(false)), stream); - break; - case DATA_AT_GPU: - copyFrom(*(src.getVector(true)), stream); - break; - case SYNCED: - copyFrom(*(src.getVector(false)), stream); - copyFrom(*(src.getVector(true)), stream); - setSync(SYNCED); - break; - default: - LOG(FATAL) << "Not support"; - break; - } -} - -template -void CpuGpuVectorT::copyToCpu() { - switch (*this->getSync()) { - case DATA_AT_GPU: - CHECK(gpuVectorT_); - this->resizeOrCreate(gpuVectorT_->getSize(), false); - cpuVectorT_->copyFrom(*gpuVectorT_); - setSync(SYNCED); - break; - case DATA_AT_CPU: - case SYNCED: - CHECK(cpuVectorT_); - break; - default: - LOG(FATAL) << "Not support"; - break; - } -} - -template -void CpuGpuVectorT::copyToGpu() { - switch (*this->getSync()) { - case DATA_AT_CPU: - CHECK(cpuVectorT_); - this->resizeOrCreate(cpuVectorT_->getSize(), true); - gpuVectorT_->copyFrom(*cpuVectorT_); - setSync(SYNCED); - break; - case DATA_AT_GPU: - case SYNCED: - CHECK(gpuVectorT_); - break; - default: - LOG(FATAL) << "Not support"; - break; - } -} - -template class VectorT; -template class VectorT; -template class CpuVectorT; -template class CpuVectorT; -template class GpuVectorT; -template class GpuVectorT; -template class CpuGpuVectorT; -template class CpuGpuVectorT; - -} // namespace paddle diff --git a/paddle/math/Vector.h b/paddle/math/Vector.h deleted file mode 100644 index 964b42cae52af9b487ab17103bc5e999514e4dd1..0000000000000000000000000000000000000000 --- a/paddle/math/Vector.h +++ /dev/null @@ -1,726 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include - -#include - -#include "BaseMatrix.h" -#include "MemoryHandle.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Thread.h" - -namespace paddle { - -template -class GpuVectorT; -template -class CpuVectorT; - -template -class BaseVector; - -class SyncThreadPool; - -class Matrix; - -template -class BaseVector : public BaseMatrixT { - public: - BaseVector(size_t size, T* data, bool useGpu) - : BaseMatrixT(1, size, data, false, useGpu), size_(this->width_) {} - - ~BaseVector() {} - - protected: - size_t& size_; -}; - -/** - * Copy or assignemnt constructor will share the data as opposed to making a - * copy of the original data. To make a copy of the orinal data, use copyFrom() - * instead. - */ -template -class VectorT : public BaseVector { - protected: - VectorT(size_t size, MemoryHandlePtr memoryHandle, size_t offset, bool useGpu) - : BaseVector(size, - reinterpret_cast(memoryHandle->getBuf()) + offset, - useGpu) { - memoryHandle_ = memoryHandle; - } - - // data is still owned by the caller. - // data should be valid during the life of this vector. - // Caller is responsible for release the memory. - VectorT(size_t size, T* data, bool useGpu) - : BaseVector(size, data, useGpu) {} - - public: - virtual ~VectorT() {} - - static std::shared_ptr> create(size_t size, bool useGpu); - - static std::shared_ptr> create(T* data, size_t size, bool useGpu); - - static std::shared_ptr> create(size_t size, - MemoryHandlePtr memoryHandle, - size_t offset = 0); - - // owner can set SyncThreadPool, - // if not set, will use globalSyncThreadPool, - // which can be used in main thread only. - static std::shared_ptr> createParallelVector( - size_t size, bool useGpu, SyncThreadPool* pool = nullptr); - - size_t getSize() const { return this->size_; } - const T* getData() const { return this->data_; } - T* getData() { return this->data_; } - - virtual void zeroMem() = 0; - // set all elements to value - virtual void reset(const T& value) = 0; - // fill data by 0, 1, 2, ... - virtual void fillSequence() = 0; - - MemoryHandlePtr getMemoryHandle() const { return memoryHandle_; } - - /** - * resizing to a big vector will not preserve old values. - */ - void resize(size_t newSize) { - if (!memoryHandle_ || newSize * sizeof(T) > memoryHandle_->getAllocSize()) { - memoryHandle_ = newMemory(newSize * sizeof(T)); - this->data_ = reinterpret_cast(memoryHandle_->getBuf()); - } - this->size_ = newSize; - } - - static void resizeOrCreate(std::shared_ptr>& vec, - size_t size, - bool useGpu) { - if (vec) { - vec->resize(size); - } else { - vec = create(size, useGpu); - } - } - - virtual MemoryHandlePtr newMemory(size_t size) = 0; - - /** - * form sub vector from *src*, shallow copy - */ - void subVecFrom(const VectorT& src, size_t start, size_t size) { - CHECK_EQ(BaseVector::useGpu_, src.useGpu_); - CHECK_LT(start, src.size_); - CHECK_LE(start + size, src.size_); - - BaseVector::size_ = size; - BaseVector::data_ = const_cast(src.data_) + start; - } - - std::shared_ptr> subVec(size_t start, size_t size) { - CHECK_LE(start + size, static_cast(getSize())); - return VectorT::create(getData() + start, size, BaseVector::useGpu_); - } - - /** - * form sub vector from *src*, shallow copy - */ - void subVecFrom(const T* src, size_t start, size_t size) { - BaseVector::size_ = size; - BaseVector::data_ = const_cast(src) + start; - } - - /** - * form sub vector from *src*, shallow copy - * in *interval* [interval.first, interval.second) - */ - void subVecFrom(const VectorT& src, std::pair interval) { - subVecFrom(src, interval.first, interval.second - interval.first); - } - - /** - * convert the vector to a sparse one_hot matrix of width idRange - * only applies to IVector - */ - std::shared_ptr toOneHotSparseMatrix(size_t idRange, bool useGpu); - - /** - * @brief cast vector of "real" elements to "int" elements. - * - * @note: float -> int must be casted, or you'll get wrong data. - */ - std::shared_ptr> castToInt(); - - /** - * This function will crash if the size of src and dest is different. - */ - virtual void copyFrom(const VectorT& src) = 0; - - /** - * If GpuVector, this function is an asynchronous interface, - * will push the copy-task to the specifed-stream and return immediately. - * - * If CpuVector, this function is an synchronous interface, - * same as the copyFrom(const VectorT& src). - */ - virtual void copyFrom(const VectorT& src, hl_stream_t stream) = 0; - - /** - * copy size elements from src - * - * If this is GpuVector, src can be cpu or gpu memory - * - * If this is CpuVector, src is assumed to be cpu memory - */ - virtual void copyFrom(const T* src, size_t size) = 0; - - /** - * copy size elements from src - * - * If this is GpuVector, src can be cpu or gpu memory - * - * If this is CpuVector, src is assumed to be cpu memory, - */ - virtual void copyFrom(const T* src, size_t size, hl_stream_t stream) = 0; - - /** - * exec a func in single/multi thread - */ - virtual void exec(SyncThreadPool::JobFunc func) { func(0, 1); } - - /// Get the buffer point with beginPos - virtual T* getPoint(const uint64_t beginPos) = 0; - - /// Get the value for the i'th element - virtual T getElement(size_t i) const = 0; - virtual void setElement(size_t i, const T& value) = 0; - - //---------- math operations ---------------- - - // sum of the absolute value of each elements - virtual T getAbsSum() = 0; - - virtual T getSum() = 0; - virtual T getMax() = 0; - virtual T getAbsMax() = 0; - virtual T getMin() = 0; - - /// element-wise calc: this = (b == value) - virtual void isEqualTo(const VectorT& b, const T& value) = 0; - - /// select elements indexed by *ids* from vector *src* - virtual void selectFrom(const VectorT& src, const VectorT& ids) = 0; - - enum HistogramType { - HISTOGRAM_EXPONENT = 0, - }; - - /** - * @brief print histogram of vector values - * - * @note only exponent histogram supported currently - */ - virtual void histogram(std::ostream& os, int type = HISTOGRAM_EXPONENT) = 0; - - /// generate uniform random value for each element - virtual void rand() = 0; - /** - * generate uniform random value for each element, - * data range is from 0 to (classes - 1). - */ - virtual void rand(size_t classes) = 0; - - /** - * Debug use only. Very inefficient for GPU vector. - * get the value at pos. - */ - virtual T get(size_t pos) = 0; - - /** - * generate univariate Gaussian distributed random numbers - * with given mean and standardDeviation. - */ - virtual void randnorm(real mean, real standardDeviation) = 0; - - /** - * generate uniform distributed random numbers - * with given range. - */ - virtual void uniform(real left, real right) = 0; - - /// print the first "num" elements of the Vector - virtual void print(std::ostream& os, size_t num) const = 0; - - /// print the "idx" element of the Vector - virtual void printOneElement(std::ostream& os, size_t idx) const = 0; - - template - void operator=(const ExpressionType& expr) { - if (BaseVector::useGpu_) { - TensorGpuApply(*this, expr); - } else { - TensorCpuApply(*this, expr); - } - } - - protected: - friend class GpuVectorT; - friend class CpuVectorT; - virtual void copyTo(CpuVectorT* dest) const = 0; - virtual void copyTo(GpuVectorT* dest) const = 0; - MemoryHandlePtr memoryHandle_; -}; - -template -std::ostream& operator<<(std::ostream& os, const VectorT& vec) { - vec.print(os, vec.getSize()); - return os; -} - -template -class GpuVectorT : public VectorT { - public: - explicit GpuVectorT(size_t size); - GpuVectorT(size_t size, GpuMemHandlePtr memHandle, size_t offset) - : VectorT(size, memHandle, offset, true) {} - - // data is still owned by the caller. - // data should be valid during the life of this vector. - // Caller is responsible for release the memory. - GpuVectorT(size_t size, T* data) : VectorT(size, data, true) {} - - virtual MemoryHandlePtr newMemory(size_t size) { - return std::make_shared(size); - } - virtual void zeroMem(); - virtual void reset(const T& value); - virtual void fillSequence(); - - virtual void copyFrom(const T* src, size_t size); - virtual void copyFrom(const T* src, size_t size, hl_stream_t stream); - virtual void copyFrom(const VectorT& src); - virtual void copyFrom(const VectorT& src, hl_stream_t stream); - virtual T getElement(size_t i) const; - virtual void setElement(size_t i, const T& value); - virtual T* getPoint(const uint64_t beginPos); - - virtual T getAbsSum(); - virtual T getSum(); - virtual T getMax(); - virtual T getAbsMax(); - virtual T getMin(); - virtual void isEqualTo(const VectorT& b, const T& value); - virtual void selectFrom(const VectorT& src, const VectorT& ids); - virtual void histogram(std::ostream& os, int type); - virtual void rand(); - virtual void rand(size_t classes); - virtual void randnorm(real mean, real standardDeviation); - virtual void uniform(real left, real right); - virtual T get(size_t pos); - virtual void print(std::ostream& os, size_t num) const; - virtual void printOneElement(std::ostream& os, size_t idx) const; - - template - void operator=(const ExpressionType& expr) { - TensorGpuApply(*this, expr); - } - - protected: - virtual void copyTo(CpuVectorT* dest) const; - virtual void copyTo(GpuVectorT* dest) const; -}; - -template -class CpuVectorT : public VectorT { - public: - explicit CpuVectorT(size_t size); - CpuVectorT(size_t size, MemoryHandlePtr memoryHandle, size_t offset) - : VectorT(size, memoryHandle, offset, false) {} - - // data is still owned by the caller. - // data should be valid during the life of this vector. - // Caller is responsible for release the memory. - CpuVectorT(size_t size, T* data) : VectorT(size, data, false) {} - - /** - * If src is a CpuVector, the new CpuVector will share the data with src - * - * If src is a GpuVector, the new CpuVector will copy data from src - */ - explicit CpuVectorT(const VectorT& src); - - virtual MemoryHandlePtr newMemory(size_t size) { - return std::make_shared(size); - } - - virtual void zeroMem(); - virtual void reset(const T& value); - virtual void fillSequence(); - virtual void copyFrom(const T* src, size_t size); - virtual void copyFrom(const T* src, size_t size, hl_stream_t stream); - virtual void copyFrom(const VectorT& src); - virtual void copyFrom(const VectorT& src, hl_stream_t stream); - virtual void copyTo(CpuVectorT* dest) const; - virtual void copyTo(GpuVectorT* dest) const; - - /// Get the buffer point with beginPos - virtual T* getPoint(const uint64_t beginPos) { - return this->getData() + beginPos; - } - - virtual T getElement(size_t i) const { return this->getData()[i]; } - virtual void setElement(size_t i, const T& value) { - this->getData()[i] = value; - } - - virtual T getAbsSum(); - virtual T getSum(); - virtual T getMax(); - virtual T getAbsMax(); - virtual T getMin(); - virtual void isEqualTo(const VectorT& b, const T& value); - virtual void selectFrom(const VectorT& src, const VectorT& ids); - virtual void histogram(std::ostream& os, int type); - virtual void rand(); - virtual void rand(size_t classes); - virtual void randnorm(real mean, real standardDeviation); - virtual void uniform(real left, real right); - virtual T get(size_t pos); - virtual void print(std::ostream& os, size_t num) const; - virtual void printOneElement(std::ostream& os, size_t idx) const; - - template - void operator=(const ExpressionType& expr) { - TensorCpuApply(*this, expr); - } -}; - -template -class ParallelCpuVectorT : public CpuVectorT { - public: - ParallelCpuVectorT(size_t size, SyncThreadPool* pool) - : CpuVectorT(size), pool_(pool) {} - - virtual void zeroMem() { - parallelExec([](CpuVectorT& vec) { vec.CpuVectorT::zeroMem(); }); - } - virtual void randnorm(real mean, real standardDeviation) { - parallelExec([=](CpuVectorT& vec) { - vec.CpuVectorT::randnorm(mean, standardDeviation); - }); - } - virtual void uniform(real left, real right) { - parallelExec( - [=](CpuVectorT& vec) { vec.CpuVectorT::uniform(left, right); }); - } - - virtual void exec(SyncThreadPool::JobFunc jobFunc); - - private: - typedef std::function& vec)> ExecFunc; - void parallelExec(ExecFunc func); - SyncThreadPool* pool_; -}; - -/** - * A class to do conversion between CpuVector and GpuVector automatically. - */ -template -class CpuGpuVectorT { - public: - /** - * @brief An enum type of SyncedFlag using to - * mark data memory is in CPU or GPU. - * - * DATA_AT_CPU: data is located in CPU. - * - * DATA_AT_GPU: data is located in GPU. - * - * SYNCED: data is located in CPU and GPU simultaneously. - */ - enum SyncedFlag { DATA_AT_CPU = 0, DATA_AT_GPU = 1, SYNCED = 2 }; - - /** - * @brief A constructor, create cpuVectorT_ or gpuVectorT_. - * - * @param[in] size data size. - * @param[in] useGpu use gpu or not. - */ - explicit CpuGpuVectorT(size_t size, bool useGpu); - - /** - * @brief A constructor, create CpuGpuVectorT by VectorT. - * - * If src is CpuVector, cpuVectorT_ is shared data with src. - * - * If src is GpuVector, gpuVectorT_ is shared data with src. - */ - explicit CpuGpuVectorT(const std::shared_ptr>& src); - - /** - * @brief A constructor. - * - * If useGpu is true, data should be located in device and - * create gpuVectorT_ with data. - * - * If useGpu is false, data should be located in host and - * create cpuVectorT_ with data. - * - * @note Data is owned by the caller and should be valid during - * the life of this vector. - * Caller is responsible for release the memory. - */ - CpuGpuVectorT(size_t size, T* data, bool useGpu); - - CpuGpuVectorT(CpuGpuVectorT& src, size_t offset, size_t size); - - virtual ~CpuGpuVectorT() {} - - static std::shared_ptr> create(size_t size, bool useGpu); - - /** - * @brief resize vector. - * - * If useGpu is true, resize gpuVectorT_ and set syncFlag_ to DATA_AT_GPU, - * - * otherwise resize cpuVectorT_ and set syncFlag_ to DATA_AT_CPU. - */ - void resize(size_t size, bool useGpu); - - /** - * @brief resize or create CpuGpuVectorT. - */ - static void resizeOrCreate(std::shared_ptr>& vec, - size_t size, - bool useGpu); - - /** - * @brief return a const cpuVectorT_ or gpuVectorT_. - * - * If useGpu is true, return gpuVectorT_. - * - * If useGpu is false, return cpuVectorT_. - * - * @note Caller should not change the data. - * If caller changes const attribute, - * should set syncFlag_. - */ - std::shared_ptr> getVector(bool useGpu) const; - - /** - * @brief return a const cpuVectorT_ or gpuVectorT_. - * - * @note: This interface will change syncFlag_, so if you will - * not change the data, you should call getVector. - */ - std::shared_ptr>& getMutableVector(bool useGpu); - - /** - * @brief return const T* data. - * - * If useGpu is true, return device data. - * - * If useGpu is false, return host data. - */ - const T* getData(bool useGpu) const; - - // TODO(yuyang18): Make getData more c++ style. - // inline T* getData(bool useGpu) { - // return getMutableData(useGpu); - // } - - T* getMutableData(bool useGpu); - - /** - * If useGpu is true, gpuVectorT_->Op(). - * - * If useGpu is false, cpuVectorT_->Op(). - * - * Op is zeroMem, fillSequence, ... - */ - void zeroMem(bool useGpu); - void fillSequence(bool useGpu); - void setElement(size_t i, const T& value, bool useGpu); - - /** - * @brief return i-th element. - */ - T getElement(size_t i) const; - - /** - * @brief return vector size. - */ - size_t getSize() const { - size_t size = 0; - switch (*sync_) { - case SYNCED: - case DATA_AT_CPU: - size = cpuVectorT_->getSize(); - break; - case DATA_AT_GPU: - size = gpuVectorT_->getSize(); - break; - default: - LOG(FATAL) << "Not support"; - break; - } - return size; - } - - /// copy data to cpuVectorT_. - inline void copyToCpu(const T* data, size_t size) { - this->resizeOrCreate(size, false); - cpuVectorT_->copyFrom(data, size); - setSync(DATA_AT_CPU); - } - /// copy data to cpuVectorT_ using specifed-stream. - inline void copyToCpu(const T* data, size_t size, hl_stream_t stream) { - this->resizeOrCreate(size, false); - cpuVectorT_->copyFrom(data, size, stream); - setSync(DATA_AT_CPU); - } - - /// copy data to gpuVectorT_. - inline void copyToGpu(const T* data, size_t size) { - this->resizeOrCreate(size, true); - gpuVectorT_->copyFrom(data, size); - setSync(DATA_AT_GPU); - } - /// copy data to gpuVectorT_ using specifed-stream. - inline void copyToGpu(const T* data, size_t size, hl_stream_t stream) { - this->resizeOrCreate(size, true); - gpuVectorT_->copyFrom(data, size, stream); - setSync(DATA_AT_GPU); - } - - /** - * @brief copy from src using specifed-stream. - * - * If src is CpuVectorT, copy to cpuVectorT_. - * - * If src is GpuVectorT, copy to gpuVectorT_. - */ - void copyFrom(const VectorT& src, hl_stream_t stream); - - /** - * @brief copy data. - * - * If useGpu is false, copy host data to cpuVectorT_. - * - * If useGpu is true, copy device data to gpuVectorT_. - * - * @note data address should consistent with useGpu. - */ - void copyFrom(const T* data, size_t size, bool useGpu); - void copyFrom(const T* data, size_t size, hl_stream_t stream, bool useGpu); - - /** - * @brief copy from (src + offset) using specifed-stream. - */ - void copyFrom(CpuGpuVectorT& src, - size_t offset, - size_t size, - bool useGpu, - hl_stream_t stream); - - /** - * @brief copy from src using specifed-stream. - */ - void copyFrom(CpuGpuVectorT& src, hl_stream_t stream); - - /** - * @brief return sync_. - */ - inline SyncedFlag* getSync() const { return sync_; } - - /** - * @brief set sync_. - */ - inline void setSync(SyncedFlag* sync) { sync_ = sync; } - - inline void setSync(SyncedFlag syncFlag) { - if (sync_) { - *sync_ = syncFlag; - } else { - syncFlag_ = syncFlag; - sync_ = &syncFlag_; - } - } - - inline void setSync(bool useGpu) { - SyncedFlag flag = useGpu ? DATA_AT_GPU : DATA_AT_CPU; - setSync(flag); - } - - protected: - void resizeOrCreate(size_t size, bool useGpu); - - /** - * @brief copy between cpuVectorT_ and gpuVectorT_. - * - * If syncFlag_ is DATA_AT_CPU and SYNCED, do nothing. - * - * If syncFlag_ is DATA_AT_GPU, copy gpuVectorT_ to cpuVectorT_ - * and set syncFlag_ to SYNCED. - */ - void copyToCpu(); - - /** - * @brief copy between cpuVectorT_ and gpuVectorT_. - * - * If syncFlag_ is DATA_AT_GPU and SYNCED, do nothing. - * - * If syncFlag_ is DATA_AT_CPU, copy cpuVectorT_ to gpuVectorT_ - * and set syncFlag_ to SYNCED. - */ - void copyToGpu(); - - /// host pointer. - std::shared_ptr> cpuVectorT_; - /// device pointer. - std::shared_ptr> gpuVectorT_; - /// specify current data address. - SyncedFlag syncFlag_; - SyncedFlag* sync_; -}; - -typedef VectorT Vector; -typedef CpuVectorT CpuVector; -typedef GpuVectorT GpuVector; - -typedef VectorT IVector; -typedef CpuVectorT CpuIVector; -typedef GpuVectorT GpuIVector; - -typedef std::shared_ptr VectorPtr; -typedef std::shared_ptr CpuVectorPtr; -typedef std::shared_ptr GpuVectorPtr; - -typedef std::shared_ptr IVectorPtr; -typedef std::shared_ptr CpuIVectorPtr; -typedef std::shared_ptr GpuIVectorPtr; - -typedef CpuGpuVectorT CpuGpuVector; -typedef CpuGpuVectorT ICpuGpuVector; -typedef std::shared_ptr CpuGpuVectorPtr; -typedef std::shared_ptr ICpuGpuVectorPtr; - -} // namespace paddle diff --git a/paddle/math/tests/OriginalOptimizerApi.h b/paddle/math/tests/OriginalOptimizerApi.h deleted file mode 100644 index e30d784b232dd7d477877d3f7c90cd185357328c..0000000000000000000000000000000000000000 --- a/paddle/math/tests/OriginalOptimizerApi.h +++ /dev/null @@ -1,201 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/math/Vector.h" -#include "paddle/utils/GlobalConstants.h" - -using namespace paddle; // NOLINT - -void SparseMomentumParameterOptimizer(const VectorPtr vecs[], - real alpha, - real beta, - real gamma, - real tau, - real learningRate) { - vecs[PARAMETER_MOMENTUM_UT]->add(*vecs[PARAMETER_GRADIENT], - -alpha * gamma * learningRate); - vecs[PARAMETER_MOMENTUM_VT]->add(*vecs[PARAMETER_GRADIENT], - tau * alpha * gamma * learningRate); - vecs[PARAMETER_VALUE]->add(*vecs[PARAMETER_MOMENTUM_UT], - tau / beta + 1.0 / alpha, - *vecs[PARAMETER_MOMENTUM_VT], - 1.0 / beta); -} - -void AdagradParameterOptimizer(const VectorPtr vecs[], - real epsilon, - real learningRate, - real momentum, - real decayRate) { - vecs[PARAMETER_GRADIENT_SQURESUM1]->addSquare(*vecs[PARAMETER_GRADIENT], - 1.0f); - vecs[PARAMETER_LEARNING_RATE]->add(*vecs[PARAMETER_GRADIENT_SQURESUM], - *vecs[PARAMETER_GRADIENT_SQURESUM1]); - vecs[PARAMETER_LEARNING_RATE]->add(epsilon); - vecs[PARAMETER_LEARNING_RATE]->invSqrt(*vecs[PARAMETER_LEARNING_RATE]); - - vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], - *vecs[PARAMETER_MOMENTUM], - *vecs[PARAMETER_LEARNING_RATE], - learningRate, - momentum, - decayRate); -} - -void AdaDeltaParameterOptimizer(const VectorPtr vecs[], - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate) { - // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 - vecs[PARAMETER_GRADIENT_SQURESUM]->decayAddSquare( - *vecs[PARAMETER_GRADIENT], rou, 1.0f - rou); - - // learn_rate = sqrt( ( E(dx_{t-1}^2) + epsilon ) / ( E(g_t^2) + epsilon ) ) - vecs[PARAMETER_LEARNING_RATE]->dotDiv(*vecs[PARAMETER_GRADIENT_SQURESUM1], - *vecs[PARAMETER_GRADIENT_SQURESUM], - epsilon, - epsilon); - vecs[PARAMETER_LEARNING_RATE]->sqrt2(); - - // E(dx_t^2) = \rou * E(dx_{t-1}^2) + (1-\rou) * (-g*learn_rate)^2 - vecs[PARAMETER_GRADIENT_SQURESUM1]->decayAddSquareMul( - *vecs[PARAMETER_GRADIENT], - *vecs[PARAMETER_LEARNING_RATE], - rou, - 1.0f - rou); - - vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], - *vecs[PARAMETER_MOMENTUM], - *vecs[PARAMETER_LEARNING_RATE], - learningRate, - momentum, - decayRate); -} - -void RMSPropParameterOptimizer(const VectorPtr vecs[], - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime) { - // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 - // For the first time update, make the sum be the current square - // so that the initial estimation of E(g_t^2) will not be too small. - vecs[PARAMETER_GRADIENT_SQURESUM]->decayAddSquare( - *vecs[PARAMETER_GRADIENT], accumulatedRou, firstTime ? 1.0f : 1.0f - rou); - - // E(g_t) = \rou * E(g_{t-1}) + (1-\rou) * g - vecs[PARAMETER_GRADIENT_SQURESUM1]->add( - *vecs[PARAMETER_GRADIENT], accumulatedRou, 1.0f - rou); - - // learn_rate = 1/sqrt( ( E(g_t^2) - (E(g_t))^2 + epsilon ) - // Basiclly if the sign of the gradient changes more often, - // the learning rate will be decreased. - vecs[PARAMETER_LEARNING_RATE]->assign(*vecs[PARAMETER_GRADIENT_SQURESUM]); - vecs[PARAMETER_LEARNING_RATE]->addSquare(*vecs[PARAMETER_GRADIENT_SQURESUM1], - -1.0f); - vecs[PARAMETER_LEARNING_RATE]->add(epsilon); - vecs[PARAMETER_LEARNING_RATE]->invSqrt(*vecs[PARAMETER_LEARNING_RATE]); - - vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], - *vecs[PARAMETER_MOMENTUM], - *vecs[PARAMETER_LEARNING_RATE], - learningRate, - momentum, - decayRate); -} - -void DecayedAdagradParameterOptimizer(const VectorPtr vecs[], - real accumulatedRou, - real rou, - real epsilon, - real learningRate, - real momentum, - real decayRate, - bool firstTime) { - // E(g_t^2) = \rou * E(g_{t-1}^2) + (1-\rou) * g^2 - // For the first time update, make the sum be the current square - // so that the initial estimation of E(g_t^2) will not be too small. - vecs[PARAMETER_GRADIENT_SQURESUM]->decayAddSquare( - *vecs[PARAMETER_GRADIENT], accumulatedRou, firstTime ? 1.0f : 1.0f - rou); - - // learn_rate = 1/sqrt( ( E(g_t^2) + epsilon ) - // Basiclly if the bigger the magnitude gradient is, - // the smaller the learning rate will be. - vecs[PARAMETER_LEARNING_RATE]->assign(epsilon); - vecs[PARAMETER_LEARNING_RATE]->add(*vecs[PARAMETER_GRADIENT_SQURESUM]); - vecs[PARAMETER_LEARNING_RATE]->invSqrt(*vecs[PARAMETER_LEARNING_RATE]); - - vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], - *vecs[PARAMETER_MOMENTUM], - *vecs[PARAMETER_LEARNING_RATE], - learningRate, - momentum, - decayRate); -} - -void AdamParameterOptimizer(const VectorPtr vecs[], - real beta1, - real beta2, - real beta1_power, - real beta2_power, - real epsilon, - real learningRate) { - Vector* m = vecs[PARAMETER_MOMENTUM].get(); - Vector* g = vecs[PARAMETER_GRADIENT].get(); - Vector* v = vecs[PARAMETER_SECOND_MOMENTUM].get(); - Vector* theta = vecs[PARAMETER_VALUE].get(); - - // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; - m->add(*g, beta1, 1 - beta1); - - // v_t = \beta_2 * v_{t-1} + (1-\beta_2)* g_{t-1}^2 - g->square2(); - v->add(*g, beta2, 1 - beta2); - - // tmp = m_t / ( \sqrt{v_t} + \epsilon ) - // \theta_t = \theta_{t-1} - \alpha * \sqrt(1-\beta_2^t) / (1-\beta_1^t) * tmp - g->sqrt2(*v); - g->dotDiv(*m, *g, 0., epsilon); - real alpha = - learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); - theta->add(*theta, 1.0, *g, -alpha); -} - -void AdamaxParameterOptimizer( - const VectorPtr vecs[], real beta1, real beta2, int64_t step, real alpha) { - Vector* m = vecs[PARAMETER_MOMENTUM].get(); - Vector* g = vecs[PARAMETER_GRADIENT].get(); - Vector* u = vecs[PARAMETER_WEIGHTED_INFINITY_NORM].get(); - Vector* theta = vecs[PARAMETER_VALUE].get(); - - // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; - m->add(*g, beta1, 1 - beta1); - - // u_t = max(\beta_2*u_{t-1}, abs(g_t)) - u->mulScalar(beta2); - g->abs2(); - u->max2(*u, *g); - - // \theta_t = \theta_{t-1} - (\alpha/(1-\beta_1^t))*m_t/u_t - g->dotDiv(*m, *u); - real learningRate = alpha / (1 - std::pow(beta1, step)); - theta->add(*theta, 1.0, *g, -learningRate); -} diff --git a/paddle/math/tests/PerfUtils.h b/paddle/math/tests/PerfUtils.h deleted file mode 100644 index bee2351e2fb80f9ccef670535c92485389f0c51a..0000000000000000000000000000000000000000 --- a/paddle/math/tests/PerfUtils.h +++ /dev/null @@ -1,46 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -// Performance Check -#ifdef PADDLE_DISABLE_TIMER - -#define EXPRESSION_PERFORMANCE(expression) expression; - -#else - -#include "paddle/utils/Stat.h" -using namespace paddle; // NOLINT - -#define EXPRESSION_PERFORMANCE(expression) \ - do { \ - char expr[30]; \ - strncpy(expr, #expression, 30); \ - if (expr[29] != '\0') { \ - expr[27] = '.'; \ - expr[28] = '.'; \ - expr[29] = '\0'; \ - } \ - expression; \ - for (int i = 0; i < 20; i++) { \ - REGISTER_TIMER(expr); \ - expression; \ - } \ - LOG(INFO) << std::setiosflags(std::ios::left) << std::setfill(' ') \ - << *globalStat.getStat(expr); \ - globalStat.reset(); \ - } while (0) - -#endif diff --git a/paddle/math/tests/TensorCheck.h b/paddle/math/tests/TensorCheck.h deleted file mode 100644 index 40ac04ef5d4baa0239bb03b04c3a6cce0fcac5a5..0000000000000000000000000000000000000000 --- a/paddle/math/tests/TensorCheck.h +++ /dev/null @@ -1,216 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -/** - * This file provides a TensorCheck template function, which can be used to - * compare CpuMatrix and GpuMatrix, CpuVector and GpuVector, and so on. - */ - -#include -#include "paddle/math/Matrix.h" - -namespace autotest { - -using paddle::Matrix; -using paddle::CpuMatrix; -using paddle::GpuMatrix; -using paddle::VectorT; -using paddle::CpuVectorT; -using paddle::GpuVectorT; - -class AssertEqual { - public: - AssertEqual(real err = 0) : err_(err) {} - - inline bool operator()(real a, real b) { - if (err_ == 0) { - if (a != b) { - return false; - } - } else { - if (std::fabs(a - b) > err_) { - if ((std::fabs(a - b) / std::fabs(a)) > (err_ / 10.0f)) { - return false; - } - } - } - - return true; - } - - private: - real err_; -}; - -template -class CopyToCpu; - -template <> -class CopyToCpu { - public: - explicit CopyToCpu(const CpuMatrix& arg) : arg_(arg) {} - const CpuMatrix& copiedArg() const { return arg_; } - - private: - const CpuMatrix& arg_; -}; - -template <> -class CopyToCpu { - public: - explicit CopyToCpu(const GpuMatrix& arg) - : arg_(arg.getHeight(), arg.getWidth()) { - arg_.copyFrom(arg); - } - CpuMatrix& copiedArg() { return arg_; } - - private: - CpuMatrix arg_; -}; - -template <> -class CopyToCpu { - public: - explicit CopyToCpu(const Matrix& arg) - : arg_(arg.getHeight(), arg.getWidth()) { - arg_.copyFrom(arg); - } - CpuMatrix& copiedArg() { return arg_; } - - private: - CpuMatrix arg_; -}; - -template -class CopyToCpu> { - public: - explicit CopyToCpu(const CpuVectorT& arg) : arg_(arg) {} - const CpuVectorT& copiedArg() const { return arg_; } - - private: - const CpuVectorT& arg_; -}; - -template -class CopyToCpu> { - public: - explicit CopyToCpu(const GpuVectorT& arg) : arg_(arg.getSize()) { - arg_.copyFrom(arg); - } - CpuVectorT& copiedArg() { return arg_; } - - private: - CpuVectorT arg_; -}; - -template -class CopyToCpu> { - public: - explicit CopyToCpu(const VectorT& arg) : arg_(arg.getSize()) { - arg_.copyFrom(arg); - } - CpuVectorT& copiedArg() { return arg_; } - - private: - CpuVectorT arg_; -}; - -template -void TensorCheck(AssertEq compare, - const CpuMatrix& matrix1, - const CpuMatrix& matrix2) { - CHECK(matrix1.getHeight() == matrix2.getHeight()); - CHECK(matrix1.getWidth() == matrix2.getWidth()); - - int height = matrix1.getHeight(); - int width = matrix1.getWidth(); - const real* data1 = matrix1.getData(); - const real* data2 = matrix2.getData(); - int count = 0; - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - real a = data1[i * width + j]; - real b = data2[i * width + j]; - if (!compare(a, b)) { - count++; - } - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; -} - -template -void TensorCheck(AssertEq compare, - const CpuVectorT& vector1, - const CpuVectorT& vector2) { - CHECK(vector1.getSize() == vector2.getSize()); - - const T* data1 = vector1.getData(); - const T* data2 = vector2.getData(); - size_t size = vector1.getSize(); - int count = 0; - for (size_t i = 0; i < size; i++) { - real a = data1[i]; - real b = data2[i]; - if (!compare(a, b)) { - count++; - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different elements."; -} - -template -void TensorCheck(AssertEq compare, - const Tensor1& tensor1, - const Tensor2& tensor2) { - TensorCheck(compare, - CopyToCpu(tensor1).copiedArg(), - CopyToCpu(tensor2).copiedArg()); -} - -template -void TensorCheck(AssertEq compare, real args1, real args2) { - EXPECT_EQ(compare(args1, args2), true) << "[Test error] args1 = " << args1 - << ", args2 = " << args2; -} - -template -void TensorCheck(AssertEq compare, size_t args1, size_t args2) { - EXPECT_EQ(args1, args2) << "[Test error] args1 = " << args1 - << ", args2 = " << args2; -} - -template -void TensorCheckEqual(const Tensor1& tensor1, const Tensor2& tensor2) { - AssertEqual compare(0); - TensorCheck(compare, - CopyToCpu(tensor1).copiedArg(), - CopyToCpu(tensor2).copiedArg()); -} - -template -void TensorCheckErr(const Tensor1& tensor1, const Tensor2& tensor2) { -#ifndef PADDLE_TYPE_DOUBLE - AssertEqual compare(1e-3); -#else - AssertEqual compare(1e-10); -#endif - TensorCheck(compare, - CopyToCpu(tensor1).copiedArg(), - CopyToCpu(tensor2).copiedArg()); -} - -} // namespace autotest diff --git a/paddle/math/tests/TestUtils.h b/paddle/math/tests/TestUtils.h deleted file mode 100644 index e1966ec8a74747960420ec80fdfbb957f7cf177f..0000000000000000000000000000000000000000 --- a/paddle/math/tests/TestUtils.h +++ /dev/null @@ -1,294 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -/** - * This file provides a AutoCompare calss to simplify the comparison - * of CPU and GPU member functions. - * - * This takes two steps - * 1. Construct an AutoCompare object. - * When constructing an AutoCompare object, you can set the err argument - * to specify the maximum error for CPU and GPU functions. - * - * 2. Use the template functions cmpWithArg or cmpWithoutArg. - * A. [cmpWithArg] Requires the caller construct the cpu arguments. - * - * AutoCompare test; - * Init Argument arg1,arg2... - * test.cmpWithArg(function, arg1, arg2....) - * - * B. [cmpWithoutArg] The caller do not need construct arguments. - * If matrix used in these functions arguments is the same size. - * Such as the element wise function and the aggregate function - * defined in the BaseMatrix.cpp. - * - * AutoCompare test; - * test.cmpWithoutArg(function, height, width) - */ - -#include -#include "TensorCheck.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" - -namespace autotest { - -using paddle::BaseMatrix; -using paddle::CpuMatrix; -using paddle::GpuMatrix; -using paddle::CpuIVector; -using paddle::GpuIVector; -using paddle::CpuSparseMatrix; -using paddle::GpuSparseMatrix; - -template -class ReplaceType { - public: - typedef T1 type; -}; - -template <> -class ReplaceType { - public: - typedef CpuMatrix type; -}; - -template <> -class ReplaceType { - public: - typedef GpuMatrix type; -}; - -template <> -class ReplaceType { - public: - typedef CpuMatrix type; -}; - -template <> -class ReplaceType { - public: - typedef GpuMatrix type; -}; - -// construct a argument -template -T construct(int height, int width); - -template <> -float construct(int height, int width) { - return 0.5; -} - -template <> -double construct(int height, int width) { - return 0.5; -} - -template <> -size_t construct(int height, int width) { - size_t offset = std::rand() % (height < width ? height : width); - return offset; -} - -template <> -CpuMatrix construct(int height, int width) { - CpuMatrix a(height, width); - return a; -} - -template <> -GpuMatrix construct(int height, int width) { - GpuMatrix a(height, width); - return a; -} - -// init a argument -template -void init(T& v) { - return; -} - -template <> -void init(CpuMatrix& v) { - v.randomizeUniform(); -} - -template <> -void init(GpuMatrix& v) { - v.randomizeUniform(); -} - -// init a tuple which contains a set of arguments. -template -inline typename std::enable_if::type initTuple( - std::tuple& t) {} - -template - inline typename std::enable_if < - I::type initTuple(std::tuple& t) { - init(std::get(t)); - initTuple(t); -} - -// copy a argument, copy src to dest -template -void copy(T1& dest, T2& src) { - dest = src; -} - -template <> -void copy(GpuMatrix& dest, CpuMatrix& src) { - dest.copyFrom(src); -} - -// copy a tuple, copy src to dest -template -inline typename std::enable_if::type copyTuple( - std::tuple& dest, std::tuple& src) {} - -template - inline typename std::enable_if < - I::type copyTuple(std::tuple& dest, - std::tuple& src) { - copy(std::get(dest), std::get(src)); - copyTuple(dest, src); -} - -// call member function -template -R call(C& obj, R (FC::*f)(FArgs...), Args&&... args) { - return (obj.*f)(args...); -} - -template -class ReturnType { - public: - typedef T type; -}; - -template <> -class ReturnType { - public: - typedef GpuMatrix type; -}; - -template <> -class ReturnType { - public: - typedef GpuIVector type; -}; - -template <> -class ReturnType { - public: - typedef GpuSparseMatrix type; -}; - -template -typename ReturnType::type autoArgs(T& v) { - return v; -} - -template <> -GpuMatrix autoArgs(CpuMatrix& v) { - GpuMatrix a(v.getHeight(), v.getWidth()); - a.copyFrom(v); - return a; -} - -template <> -GpuIVector autoArgs(CpuIVector& v) { - GpuIVector a(v.getSize()); - a.copyFrom(v); - return a; -} - -template <> -GpuSparseMatrix autoArgs(CpuSparseMatrix& v) { - GpuSparseMatrix a(v.getHeight(), - v.getWidth(), - v.getElementCnt(), - v.getValueType(), - v.getFormat()); - a.copyFrom(v, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - return a; -} - -class AutoCompare { - public: - /** - * err is the allowed calculation error. - * The smaller the value of err, - * the stricter the comparison is between CPU and GPU calculations. - */ - AutoCompare(size_t height, size_t width, real err = 1e-3) - : cpu(height, width), gpu(height, width), compare(err) { - init(cpu); - copy(gpu, cpu); - } - - template - void cmpWithArg(R (C::*f)(FArgs...), Args&&... args) { - static_assert(sizeof...(FArgs) == sizeof...(Args), - "size of parameter packs are not equal"); - call(cpu, f, args...); - call(gpu, f, autoArgs(args)...); - - TensorCheck(compare, cpu, gpu); - } - - template - void cmpWithoutArg(R (C::*f)(Args...), size_t height, size_t width) { - static_assert(sizeof...(I) == sizeof...(Args), - "size of parameter packs are not equal"); - (void)height; - (void)width; - auto tuple1 = std::make_tuple( - construct>::type>::type, - CpuMatrix>::type>(height, width)...); - - auto tuple2 = std::make_tuple( - construct>::type>::type, - GpuMatrix>::type>(height, width)...); - - initTuple(tuple1); - copyTuple(tuple2, tuple1); - - call(cpu, f, std::get(tuple1)...); - call(gpu, f, std::get(tuple2)...); - - TensorCheck(compare, cpu, gpu); - } - - protected: - CpuMatrix cpu; - GpuMatrix gpu; - AssertEqual compare; -}; - -} // namespace autotest diff --git a/paddle/math/tests/test_Allocator.cpp b/paddle/math/tests/test_Allocator.cpp deleted file mode 100644 index 84bc1c1d9e0a8368a69c1e53a63056eb45b9239f..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_Allocator.cpp +++ /dev/null @@ -1,122 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/utils/Logging.h" -#include "paddle/utils/Util.h" -#define private public -#include "paddle/math/Allocator.h" -#include "paddle/math/MemoryHandle.h" -#include "paddle/math/PoolAllocator.h" - -using namespace paddle; // NOLINT - -template -void testPoolAllocator() { - PoolAllocator* pool = - new PoolAllocator(new Allocator(), /* sizeLimit */ 1024); - - /* alloc from system memory */ - void* ptr1 = pool->alloc(10); - void* ptr2 = pool->alloc(200); - void* ptr3 = pool->alloc(200); - pool->free(ptr1, 10); - pool->free(ptr2, 200); - pool->free(ptr3, 200); - pool->printAll(); - EXPECT_EQ((size_t)2, pool->pool_.size()); - EXPECT_EQ((size_t)1, pool->pool_[10].size()); - EXPECT_EQ((size_t)2, pool->pool_[200].size()); - EXPECT_EQ(ptr1, pool->pool_[10][0]); - EXPECT_EQ(ptr2, pool->pool_[200][0]); - EXPECT_EQ(ptr3, pool->pool_[200][1]); - - /* alloc from pool */ - void* ptr4 = pool->alloc(10); - void* ptr5 = pool->alloc(200); - pool->printAll(); - EXPECT_EQ((size_t)0, pool->pool_[10].size()); - EXPECT_EQ((size_t)1, pool->pool_[200].size()); - EXPECT_EQ(ptr1, ptr4); - EXPECT_EQ(ptr3, ptr5); - pool->free(ptr4, 10); - pool->free(ptr5, 200); - - /* alloc size > sizeLimit */ - void* ptr6 = pool->alloc(1024); - pool->free(ptr6, 1024); - EXPECT_LE((size_t)1024, pool->poolMemorySize_); - - void* ptr7 = pool->alloc(1); - EXPECT_EQ((size_t)0, pool->poolMemorySize_); - EXPECT_EQ((size_t)0, pool->pool_.size()); - pool->free(ptr7, 1); - - delete pool; -} - -TEST(Allocator, Pool) { - testPoolAllocator(); -#ifdef PADDLE_WITH_CUDA - testPoolAllocator(); -#endif -} - -TEST(MemoryHandle, Cpu) { - for (auto size : {10, 30, 50, 100, 200, 512, 1000, 1023, 1024, 1025, 8193}) { - CpuMemoryHandle handle(size); - EXPECT_LE(handle.getSize(), handle.getAllocSize()); - } - - void* ptr1; - void* ptr2; - { - CpuMemoryHandle handle(256); - ptr1 = handle.getBuf(); - } - { - CpuMemoryHandle handle(256); - ptr2 = handle.getBuf(); - } - EXPECT_EQ(ptr1, ptr2); -} - -#ifdef PADDLE_WITH_CUDA -TEST(MemoryHandle, Gpu) { - int numGpu = hl_get_device_count(); - - /* alloc from system memory */ - void* ptr3[numGpu]; - void* ptr4[numGpu]; - for (int i = 0; i < numGpu; i++) { - SetDevice device(i); - GpuMemoryHandle handle1(30); - GpuMemoryHandle handle2(30); - GpuMemoryHandle handle3(4000); - GpuMemoryHandle handle4(500); - ptr3[i] = handle3.getBuf(); - ptr4[i] = handle4.getBuf(); - } - - /* alloc from pool */ - for (int i = 0; i < numGpu; i++) { - SetDevice device(i); - GpuMemoryHandle handle1(30); - GpuMemoryHandle handle3(4000); - GpuMemoryHandle handle4(500); - EXPECT_EQ(ptr3[i], handle3.getBuf()); - EXPECT_EQ(ptr4[i], handle4.getBuf()); - } -} -#endif diff --git a/paddle/math/tests/test_BaseMatrix.cpp b/paddle/math/tests/test_BaseMatrix.cpp deleted file mode 100644 index 6f7beb60c8f535d51b18c4984b89d1972f4c82bd..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_BaseMatrix.cpp +++ /dev/null @@ -1,247 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_CUDA -/** - * This test file use autotest::AutoCompare and cmpWithoutArg to compares the - * implementation of CPU and GPU member function in - * BaseMatrix.cpp and Matrix.cpp. - */ - -#include -#include "TestUtils.h" -#include "paddle/math/BaseMatrix.h" - -using paddle::BaseMatrix; -using paddle::Matrix; -using autotest::AutoCompare; - -// Test all void (BaseMatrix::*)() function -TEST(BaseMatrix, void) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - auto compare = [height, width](void (BaseMatrix::*f)()) { - AutoCompare test(height, width, 1e-5); - test.cmpWithoutArg(f, height, width); - }; - - compare(&BaseMatrix::neg); - compare(&BaseMatrix::exp2); - compare(&BaseMatrix::log2); - compare(&BaseMatrix::sqrt2); - compare(&BaseMatrix::square2); - compare(&BaseMatrix::reciprocal2); - compare(&BaseMatrix::abs2); - compare(&BaseMatrix::sign2); - compare(&BaseMatrix::zero); - compare(&BaseMatrix::one); - } - } -} - -// Test all void (BaseMatrix::*)(real) function -TEST(BaseMatrix, real) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - auto compare = [height, width](void (BaseMatrix::*f)(real)) { - AutoCompare test(height, width, 1e-5); - test.cmpWithoutArg<0>(f, height, width); - }; - - compare(&BaseMatrix::pow2); - compare(&BaseMatrix::subScalar); - compare(&BaseMatrix::mulScalar); - compare(&BaseMatrix::divScalar); - compare(&BaseMatrix::assign); - compare(&BaseMatrix::add); - compare(&BaseMatrix::biggerThanScalar); - compare(&BaseMatrix::downClip); - } - } -} - -// Test all void (BaseMatrix::*)(BaseMatrix&) function -TEST(BaseMatrix, BaseMatrix) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - auto compare = [height, width](void (BaseMatrix::*f)(BaseMatrix&)) { - AutoCompare test(height, width, 1e-5); - test.cmpWithoutArg<0>(f, height, width); - }; - - compare(&BaseMatrix::assign); - compare(&BaseMatrix::add); - compare(&BaseMatrix::relu); - compare(&BaseMatrix::reluDerivative); - compare(&BaseMatrix::softrelu); - compare(&BaseMatrix::softreluDerivative); - compare(&BaseMatrix::brelu); - compare(&BaseMatrix::breluDerivative); - compare(&BaseMatrix::square2); - compare(&BaseMatrix::squareDerivative); - compare(&BaseMatrix::tanh); - compare(&BaseMatrix::tanhDerivative); - compare(&BaseMatrix::reciprocal2); - compare(&BaseMatrix::reciprocalDerivative); - compare(&BaseMatrix::abs2); - compare(&BaseMatrix::absDerivative); - compare(&BaseMatrix::sigmoid); - compare(&BaseMatrix::sigmoidDerivative); - compare(&BaseMatrix::expDerivative); - compare(&BaseMatrix::sign2); - compare(&BaseMatrix::exp2); - compare(&BaseMatrix::log2); - compare(&BaseMatrix::sqrt2); - compare(&BaseMatrix::dotMul); - compare(&BaseMatrix::dotMulSquare); - compare(&BaseMatrix::dotSquareMul); - compare(&BaseMatrix::addColVector); - compare(&BaseMatrix::addRowVector); - compare(&BaseMatrix::mulRowVector); - compare(&BaseMatrix::divRowVector); - compare(&BaseMatrix::mulColVector); - compare(&BaseMatrix::divColVector); - compare(&BaseMatrix::addP2P); - compare(&BaseMatrix::invSqrt); - } - } -} - -// Test all void (BaseMatrix::*)(real, real) function -TEST(BaseMatrix, real_real) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - auto compare = [height, width](void (BaseMatrix::*f)(real, real)) { - AutoCompare test(height, width, 1e-5); - test.cmpWithoutArg<0, 1>(f, height, width); - }; - - compare(&BaseMatrix::add); - compare(&BaseMatrix::clip); - } - } -} - -// Test all void (BaseMatrix::*)(BaseMatrix&, real) function -TEST(BaseMatrix, BaseMatrix_real) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - auto compare = [height, width](void (BaseMatrix::*f)(BaseMatrix&, real)) { - AutoCompare test(height, width, 1e-5); - test.cmpWithoutArg<0, 1>(f, height, width); - }; - - compare(&BaseMatrix::addBias); - compare(&BaseMatrix::add); - compare(&BaseMatrix::sub); - compare(&BaseMatrix::pow2); - compare(&BaseMatrix::addScalar); - compare(&BaseMatrix::subScalar); - compare(&BaseMatrix::mulScalar); - compare(&BaseMatrix::divScalar); - compare(&BaseMatrix::scalarDiv); - compare(&BaseMatrix::addSquare); - compare(&BaseMatrix::isEqualTo); - } - } -} - -// Test all void (BaseMatrix::*)(BaseMatrix&, BaseMatrix&) function -TEST(BaseMatrix, BaseMatrix_BaseMatrix) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - auto compare = [height, - width](void (BaseMatrix::*f)(BaseMatrix&, BaseMatrix&)) { - AutoCompare test(height, width, 1e-5); - test.cmpWithoutArg<0, 1>(f, height, width); - }; - - compare(&BaseMatrix::softCrossEntropy); - compare(&BaseMatrix::softCrossEntropyBp); - compare(&BaseMatrix::binaryLabelCrossEntropy); - compare(&BaseMatrix::binaryLabelCrossEntropyBp); - compare(&BaseMatrix::sub); - compare(&BaseMatrix::add2); - compare(&BaseMatrix::dotMul); - compare(&BaseMatrix::dotDiv); - compare(&BaseMatrix::logisticRegressionLoss); - compare(&BaseMatrix::logisticRegressionLossBp); - compare(&BaseMatrix::biggerThan); - compare(&BaseMatrix::max2); - compare(&BaseMatrix::dotMulSquare); - compare(&BaseMatrix::dotSquareSquare); - } - } -} - -void TestEelementWise(size_t height, size_t width) { - AutoCompare rowScale(height, width); - rowScale.cmpWithoutArg<0, 1, 2>(&BaseMatrix::rowScale, height, width); - - AutoCompare rowDotMul(height, width); - rowDotMul.cmpWithoutArg<0, 1, 2>(&BaseMatrix::rowDotMul, height, width); - - AutoCompare binaryClassificationError(height, width); - binaryClassificationError.cmpWithoutArg<0, 1, 2, 3>( - &BaseMatrix::binaryClassificationError, height, width); - - AutoCompare sumOfSquaresBp(height, width); - sumOfSquaresBp.cmpWithoutArg<0, 1>(&Matrix::sumOfSquaresBp, height, width); -} - -void TestAggregateToRow(size_t height, size_t width) { - AutoCompare maxCols(1, width); - maxCols.cmpWithoutArg<0>(&BaseMatrix::maxCols, height, width); - - AutoCompare minCols(1, width); - minCols.cmpWithoutArg<0>(&BaseMatrix::minCols, height, width); - - AutoCompare addDotMulVMM(1, width); - addDotMulVMM.cmpWithoutArg<0, 1>(&BaseMatrix::addDotMulVMM, height, width); - - AutoCompare sumCols(1, width); - sumCols.cmpWithoutArg<0, 1, 2>(&BaseMatrix::sumCols, height, width); - - AutoCompare collectBias(1, width); - collectBias.cmpWithoutArg<0, 1>( - static_cast(&Matrix::collectBias), - height, - width); -} - -void TestAggregateToCol(size_t height, size_t width) { - AutoCompare maxRows(height, 1); - maxRows.cmpWithoutArg<0>(&BaseMatrix::maxRows, height, width); - - AutoCompare minRows(height, 1); - minRows.cmpWithoutArg<0>(&BaseMatrix::minRows, height, width); - - AutoCompare sumRows(height, 1); - sumRows.cmpWithoutArg<0, 1, 2>(&BaseMatrix::sumRows, height, width); - - AutoCompare sumOfSquares(height, 1); - sumOfSquares.cmpWithoutArg<0, 1>(&Matrix::sumOfSquares, height, width); -} - -TEST(BaseMatrix, Other) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - TestEelementWise(height, width); - TestAggregateToRow(height, width); - TestAggregateToCol(height, width); - } - } -} - -#endif diff --git a/paddle/math/tests/test_CpuGpuVector.cpp b/paddle/math/tests/test_CpuGpuVector.cpp deleted file mode 100644 index 395541a76ae5e5497fdaa8b4870e421cbf62608a..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_CpuGpuVector.cpp +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_CUDA - -#include -#include "paddle/math/Vector.h" -#include "paddle/utils/Util.h" -#include "test_matrixUtil.h" - -using namespace paddle; // NOLINT - -TEST(CpuGpuVector, getData) { - size_t size = 500; - hl_stream_t stream(HPPL_STREAM_DEFAULT); - CpuVectorPtr cpuVec = std::make_shared(size); - GpuVectorPtr gpuVec = std::make_shared(size); - cpuVec->uniform(0.0, 10.0); - gpuVec->copyFrom(*cpuVec, stream); - hl_stream_synchronize(stream); - - CpuGpuVectorPtr vec = std::make_shared(gpuVec); - auto a = vec->getData(false); - auto b = cpuVec->getData(); - hl_stream_synchronize(stream); - checkDataEqual(a, b, size); -} - -TEST(CpuGpuVector, subCreate) { - size_t size1 = 1024; - size_t offset = 100; - size_t size2 = 500; - hl_stream_t stream(HPPL_STREAM_DEFAULT); - CpuGpuVectorPtr v1 = std::make_shared(size1, /*useGpu*/ false); - auto vec = v1->getMutableVector(false); - vec->uniform(0.0, 10.0); - auto v2 = std::make_shared(*v1, offset, size2); - CHECK_EQ(*v1->getSync(), *v2->getSync()); - - // check subVec equal - checkDataEqual(v1->getData(false) + offset, v2->getData(false), size2); - - CpuVectorPtr v1Check = std::make_shared(size1); - CpuVectorPtr v2Check = std::make_shared(size2); - v1Check->copyFrom(*(v1->getVector(true)), stream); - v2Check->copyFrom(*(v2->getVector(true)), stream); - hl_stream_synchronize(stream); - - checkDataEqual(v2->getData(false), v2Check->getData(), size2); - checkDataEqual(v1Check->getData() + offset, v2Check->getData(), size2); - - CpuVectorPtr noise = std::make_shared(size2); - noise->uniform(0.0, 1.0); - auto v = v2->getMutableVector(false); // will change header - // add noise to subVec - v->add(*noise); - - // check v1_cpu_data == v2_cpu_data - checkDataEqual(v1->getData(false) + offset, v2->getData(false), size2); - - v1Check->copyFrom(*(v1->getVector(true)), stream); - v2Check->copyFrom(*(v2->getVector(true)), stream); - hl_stream_synchronize(stream); - - // check v1_gpu_data == v2_gpu_data - checkDataEqual(v1Check->getData() + offset, v2Check->getData(), size2); -} - -#endif diff --git a/paddle/math/tests/test_ExecViaCpu.cpp b/paddle/math/tests/test_ExecViaCpu.cpp deleted file mode 100644 index 72256cb9d4c93159418d27c7ca0d4f8b9a412a64..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_ExecViaCpu.cpp +++ /dev/null @@ -1,116 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include "paddle/math/SparseMatrix.h" - -using namespace paddle; // NOLINT - -const int height = 10; -const int width = 16; - -real f(Matrix& mat1, - const Matrix& mat2, - IVector& vec1, - const IVector& vec2, - real scalar) { - CHECK(!mat1.useGpu()); - CHECK(!mat2.useGpu()); - CHECK(!vec1.useGpu()); - CHECK(!vec2.useGpu()); - mat1.copyFrom(mat2); - vec1.copyFrom(vec2); - - return scalar; -} - -class Functor { - public: - real operator()(Matrix& mat1, - const Matrix& mat2, - IVector& vec1, - const IVector& vec2, - real scalar) { - a_ = f(mat1, mat2, vec1, vec2, scalar); - return a_; - } - - private: - real a_; -}; - -template -void testWrapper(F&& f) { - MatrixPtr cpumat1 = Matrix::create(height, width, false, /*useGpu=*/false); - MatrixPtr cpumat2 = Matrix::create(height, width, false, /*useGpu=*/false); - - IVectorPtr cpuvec1 = IVector::create(height, /*useGpu=*/false); - IVectorPtr cpuvec2 = IVector::create(height, /*useGpu=*/false); - - const real scalar = 1.23456; - - MatrixPtr gpumat1 = Matrix::create(height, width, false, /*useGpu=*/true); - MatrixPtr gpumat2 = Matrix::create(height, width, false, /*useGpu=*/true); - IVectorPtr gpuvec1 = IVector::create(height, /*useGpu=*/true); - IVectorPtr gpuvec2 = IVector::create(height, /*useGpu=*/true); - - cpumat2->randomizeUniform(); - cpuvec2->rand(width); - gpumat2->copyFrom(*cpumat2); - gpuvec2->copyFrom(*cpuvec2); - - real ret = execViaCpu(f, *gpumat1, *gpumat2, *gpuvec1, *gpuvec2, 1.23456); - EXPECT_EQ(ret, scalar); - cpumat1->copyFrom(*gpumat1); - cpuvec1->copyFrom(*gpuvec1); - - for (int i = 0; i < height; ++i) { - EXPECT_EQ(cpuvec1->getElement(i), cpuvec2->getElement(i)); - for (int j = 0; j < width; ++j) { - EXPECT_EQ(cpumat1->getElement(i, j), cpumat2->getElement(i, j)); - } - } - gpumat1->resize(height, 1); - execViaCpu2(&CpuMatrix::selectElements, *gpumat1, *gpumat2, *gpuvec1); - - cpumat1->resize(height, 1); - cpumat1->selectElements(*cpumat2, *cpuvec1); - for (int i = 0; i < height; ++i) { - EXPECT_EQ(cpumat1->getElement(i, 0), gpumat1->getElement(i, 0)); - } -} - -#ifdef PADDLE_WITH_CUDA -TEST(ExecViaCpu, test1) { - testWrapper(f); - testWrapper(&f); - - auto lambda = [](Matrix& mat1, - const Matrix& mat2, - IVector& vec1, - const IVector& vec2, - real scalar) -> real { - return f(mat1, mat2, vec1, vec2, scalar); - }; - LOG(INFO) << "lambda is_class=" << std::is_class::value - << " is_function=" << std::is_function::value; - testWrapper(lambda); - - Functor functor; - testWrapper(functor); -} -#endif diff --git a/paddle/math/tests/test_FPException.cpp b/paddle/math/tests/test_FPException.cpp deleted file mode 100644 index d87fdcda9edc8644301b7fe77f4c0c751d5a774a..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_FPException.cpp +++ /dev/null @@ -1,93 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -/** - * This test is about floating point calculation exception. - * Paddle catches FE_INVALID, FE DIVBYZERO and FE_OVERFLOW exceptions. - * - * Some exceptions occur in the middle of a set of formulas, - * that can be circumvented by some tricks. - * For example, - * calculate tanh - * b = 2.0 / (1.0 + exp(-2 * a)) - 1.0 - * - * If the result of (-2 * a) is too large, - * a FE_OVERFLOW exception occurs when calculating exp. - * But the result of tanh is no overflow problem, - * so we can add some tricks to prevent exp calculate an excessive value. - * - */ - -#include -#include "paddle/math/Matrix.h" -#include "paddle/utils/Common.h" - -using namespace paddle; // NOLINT - -void SetTensorValue(Matrix& matrix, real value) { - int height = matrix.getHeight(); - int width = matrix.getWidth(); - int stride = matrix.getStride(); - real* data = matrix.getData(); - for (int i = 0; i < height; i++) { - int j = rand() % width; // NOLINT - if (typeid(matrix) == typeid(CpuMatrix)) { - data[i * stride + j] = value; - } else if (typeid(matrix) == typeid(GpuMatrix)) { - hl_memcpy(&data[i * stride + j], &value, sizeof(real)); - } else { - LOG(FATAL) << "should not reach here"; - } - } -} - -template -void testTanh(real illegal) { - MatrixPtr A = std::make_shared(10, 10); - MatrixPtr B = std::make_shared(10, 10); - A->randomizeUniform(); - B->randomizeUniform(); - - SetTensorValue(*A, illegal); - - A->tanh(*B); -} - -template -void testSigmoid(real illegal) { - MatrixPtr A = std::make_shared(10, 10); - MatrixPtr B = std::make_shared(10, 10); - A->randomizeUniform(); - B->randomizeUniform(); - - SetTensorValue(*A, illegal); - - A->sigmoid(*B); -} - -TEST(fp, overflow) { - for (auto illegal : {-90.0, 90.0}) { - LOG(INFO) << " illegal=" << illegal; - testTanh(illegal); - testSigmoid(illegal); - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - - feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); - return RUN_ALL_TESTS(); -} diff --git a/paddle/math/tests/test_GpuProfiler.cpp b/paddle/math/tests/test_GpuProfiler.cpp deleted file mode 100644 index 828159660bae1ad1c0b56fd7202f0357549877ca..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_GpuProfiler.cpp +++ /dev/null @@ -1,165 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_CUDA - -#include -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" -#include "paddle/testing/TestUtil.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -void MatrixCheckErr(const Matrix& matrix1, const Matrix& matrix2) { - CHECK(matrix1.getHeight() == matrix2.getHeight()); - CHECK(matrix1.getWidth() == matrix2.getWidth()); -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - - int height = matrix1.getHeight(); - int width = matrix1.getWidth(); - const real* data1 = matrix1.getData(); - const real* data2 = matrix2.getData(); - int count = 0; - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - real a = data1[i * width + j]; - real b = data2[i * width + j]; - if (fabs(a - b) > err) { - if ((fabsf(a - b) / fabsf(a)) > (err / 10.0f)) { - count++; - } - } - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; -} - -void testBilinearFwdBwd(int numSamples, - int imgSizeH, - int imgSizeW, - int channels) { - int inWidth = imgSizeH * imgSizeW * channels; - int outWidth = 2 * imgSizeH * 2 * imgSizeW * channels; - real ratioH = 0.5; - real ratioW = 0.5; - - // forward - MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); - - MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); - MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); - - input->randomizeUniform(); - inputGpu->copyFrom(*input); - - { - // nvprof: GPU Proflier - REGISTER_GPU_PROFILER("testBilinearFwdBwd"); - target->bilinearForward(*input, - imgSizeH, - imgSizeW, - 2 * imgSizeH, - 2 * imgSizeW, - channels, - ratioH, - ratioW); - targetGpu->bilinearForward(*inputGpu, - imgSizeH, - imgSizeW, - 2 * imgSizeH, - 2 * imgSizeW, - channels, - ratioH, - ratioW); - } - - // check - targetCheck->copyFrom(*targetGpu); - MatrixCheckErr(*target, *targetCheck); - - // backward - MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); - - MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpuGrad = - GpuMatrix::create(numSamples, outWidth, false, true); - MatrixPtr targetCheckGrad = - CpuMatrix::create(numSamples, inWidth, false, false); - - inputGrad->randomizeUniform(); - targetGrad->randomizeUniform(); - inputGpuGrad->copyFrom(*inputGrad); - targetGpuGrad->copyFrom(*targetGrad); - - inputGrad->bilinearBackward(*targetGrad, - 2 * imgSizeH, - 2 * imgSizeW, - imgSizeH, - imgSizeW, - channels, - ratioH, - ratioW); - inputGpuGrad->bilinearBackward(*targetGpuGrad, - 2 * imgSizeH, - 2 * imgSizeW, - imgSizeH, - imgSizeW, - channels, - ratioH, - ratioW); - - // check - targetCheckGrad->copyFrom(*inputGpuGrad); - MatrixCheckErr(*inputGrad, *targetCheckGrad); -} - -TEST(Profiler, testBilinearFwdBwd) { - auto numSamples = 10; - auto channels = 16; - auto imgSize = 64; - { - // nvprof: GPU Proflier - REGISTER_GPU_PROFILER("testBilinearFwdBwd"); - // Paddle built-in timer - REGISTER_TIMER_INFO( - "testBilinearFwdBwd", - "numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"); - testBilinearFwdBwd(numSamples, imgSize, imgSize, channels); - } - globalStat.printAllStatus(); -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - - // nvprof: GPU Proflier - REGISTER_GPU_PROFILER( - "RecursiveProfilingTest", - "numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"); - - return RUN_ALL_TESTS(); -} - -#endif diff --git a/paddle/math/tests/test_RowBuffer.cpp b/paddle/math/tests/test_RowBuffer.cpp deleted file mode 100644 index e38de853e03874be3fd3582f7b39b1d490886d78..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_RowBuffer.cpp +++ /dev/null @@ -1,65 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/math/RowBuffer.h" - -TEST(RowBuffer, testAutoGrow) { - paddle::RowBuffer buf(128); - ASSERT_EQ(128UL, buf.getWidth()); - ASSERT_TRUE(buf.isAutoGrowth()); - buf.resize(2); - ASSERT_EQ(2UL, buf.getRowCount()); - for (size_t i = 0; i < buf.getWidth() * 2; ++i) { - buf.data()[i] = i; - } - for (size_t i = 0; i < buf.getRowCount(); ++i) { - for (size_t j = 0; j < buf.getWidth(); ++j) { - ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5); - } - } - - auto data = buf.getWithAutoGrowth(2); - for (size_t i = 0; i < buf.getWidth(); ++i) { - data[i] = i; - } - - ASSERT_EQ(3UL, buf.getRowCount()); - for (size_t i = 0; i < buf.getRowCount() - 1; ++i) { - for (size_t j = 0; j < buf.getWidth(); ++j) { - ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5); - } - } - for (size_t i = 0; i < buf.getWidth(); ++i) { - ASSERT_NEAR(i, buf.get(2)[i], 1e-5); - } -} - -TEST(RowBuffer, testWithMemBuf) { - paddle::CpuMemHandlePtr mem = - std::make_shared(128 * 2 * sizeof(real)); - paddle::RowBuffer buf(mem, 128); - ASSERT_TRUE(!buf.isAutoGrowth()); - ASSERT_EQ(2UL, buf.getRowCount()); - for (size_t i = 0; i < buf.getWidth() * 2; ++i) { - buf.data()[i] = i; - } - for (size_t i = 0; i < buf.getRowCount(); ++i) { - for (size_t j = 0; j < buf.getWidth(); ++j) { - ASSERT_NEAR(i * buf.getWidth() + j, buf.getWithAutoGrowth(i)[j], 1e-5); - } - } - - ASSERT_DEATH_IF_SUPPORTED(buf.getWithAutoGrowth(3), ".*"); -} diff --git a/paddle/math/tests/test_SIMDFunctions.cpp b/paddle/math/tests/test_SIMDFunctions.cpp deleted file mode 100644 index b692679436ee7bd3b8c4a675e969e15b065cc534..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_SIMDFunctions.cpp +++ /dev/null @@ -1,171 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/math/SIMDFunctions.h" -#include "paddle/utils/Util.h" - -#include - -#include -#include -#include -#include - -#include -#include - -static constexpr size_t VECTOR_LEN = 3072; -static constexpr size_t BATCH_SIZE = 64; -static constexpr size_t ALIGN = 32; -static_assert(VECTOR_LEN % ALIGN == 0, "VECTOR_LEN % ALIGN == 0"); -static_assert(BATCH_SIZE % ALIGN == 0, "BATCH_SIZE % ALIGN == 0"); -static constexpr float EPSILON = 1e-5; -static std::mt19937 RandomEngine(time(0)); - -inline static std::unique_ptr NewVector(size_t len = VECTOR_LEN, - size_t align = ALIGN) { - float* ptr; - CHECK_EQ(posix_memalign((void**)&ptr, align, len * sizeof(float)), 0); - return std::unique_ptr(ptr); -} - -inline static std::unique_ptr NewRandomVector(size_t len = VECTOR_LEN, - size_t align = ALIGN) { - std::uniform_real_distribution dist(-100.0f, 100.0f); - auto generator = std::bind(dist, RandomEngine); - auto retv = NewVector(len, align); - std::generate_n(retv.get(), len, generator); - return retv; -} - -TEST(SIMDFunction, addTo) { - typedef std::function AddToMethodType; - - AddToMethodType naive = paddle::simd::naive::addTo; - AddToMethodType simd = paddle::simd::addTo; - - auto A = NewRandomVector(); - auto B = NewRandomVector(); - - auto ACopy = NewVector(); - memcpy(ACopy.get(), A.get(), VECTOR_LEN * sizeof(float)); - - naive(A.get(), B.get(), VECTOR_LEN); - simd(ACopy.get(), B.get(), VECTOR_LEN); - - for (size_t i = 0; i < VECTOR_LEN; ++i) { - ASSERT_NEAR(A[i], ACopy[i], EPSILON); - } -} - -TEST(SIMDFunction, batchAddTo) { - auto A = NewRandomVector(); - auto ACopy = NewVector(); - memcpy(ACopy.get(), A.get(), sizeof(float) * VECTOR_LEN); - - std::vector> B; - for (size_t i = 0; i < BATCH_SIZE; ++i) { - B.emplace_back(NewRandomVector()); - } - std::unique_ptr BRaw(new float*[BATCH_SIZE]); - for (size_t i = 0; i < BATCH_SIZE; ++i) { - BRaw[i] = B[i].get(); - } - - typedef std::function - BatchAddToMethodType; - - BatchAddToMethodType naive = paddle::simd::naive::batchAddTo; - BatchAddToMethodType simd = paddle::simd::batchAddTo; - - naive(A.get(), (const float**)BRaw.get(), BATCH_SIZE, VECTOR_LEN); - simd(ACopy.get(), (const float**)BRaw.get(), BATCH_SIZE, VECTOR_LEN); - - for (size_t i = 0; i < VECTOR_LEN; ++i) { - ASSERT_NEAR(A[i], ACopy[i], EPSILON); - } -} - -TEST(SIMDFunction, colMax) { - auto A = NewRandomVector(VECTOR_LEN * BATCH_SIZE); - auto naiveResult = NewVector(BATCH_SIZE); - auto simdResult = NewVector(BATCH_SIZE); - - typedef std::function ColMaxMethodType; - ColMaxMethodType naive = paddle::simd::naive::colMax; - ColMaxMethodType simd = paddle::simd::colMax; - - naive(naiveResult.get(), A.get(), BATCH_SIZE, VECTOR_LEN); - simd(simdResult.get(), A.get(), BATCH_SIZE, VECTOR_LEN); - - for (size_t i = 0; i < BATCH_SIZE; ++i) { - ASSERT_NEAR(naiveResult[i], simdResult[i], EPSILON); - } -} - -TEST(SIMDFunction, decayL1_WithLR) { - auto dest = NewRandomVector(); - auto src = NewRandomVector(); - auto lr = NewRandomVector(); - auto lambda = 0.23f; - - auto simd_dest = NewVector(); - memcpy(simd_dest.get(), dest.get(), sizeof(float) * VECTOR_LEN); - - typedef std::function - DecayL1MethodType; - - DecayL1MethodType naive = []( - float* d, float* s, float* lr, float l, size_t len) { - paddle::simd::naive::decayL1(d, s, lr, l, len); - }; - - DecayL1MethodType simd = []( - float* d, float* s, float* lr, float l, size_t len) { - paddle::simd::decayL1(d, s, lr, l, len); - }; - - naive(dest.get(), src.get(), lr.get(), lambda, VECTOR_LEN); - simd(simd_dest.get(), src.get(), lr.get(), lambda, VECTOR_LEN); - - for (size_t i = 0; i < VECTOR_LEN; ++i) { - ASSERT_NEAR(dest[i], simd_dest[i], EPSILON); - } -} - -TEST(SIMDFunction, decayL1_WithoutLR) { - auto dest = NewRandomVector(); - auto src = NewRandomVector(); - auto lambda = 0.23; - - auto simd_dest = NewVector(); - memcpy(simd_dest.get(), dest.get(), sizeof(float) * VECTOR_LEN); - - typedef std::function DecayL1MethodType; - - DecayL1MethodType naive = [](float* d, float* s, float l, size_t len) { - paddle::simd::naive::decayL1(d, s, l, len); - }; - - DecayL1MethodType simd = [](float* d, float* s, float l, size_t len) { - paddle::simd::decayL1(d, s, l, len); - }; - - naive(dest.get(), src.get(), lambda, VECTOR_LEN); - simd(simd_dest.get(), src.get(), lambda, VECTOR_LEN); - - for (size_t i = 0; i < VECTOR_LEN; ++i) { - ASSERT_NEAR(dest[i], simd_dest[i], EPSILON); - } -} diff --git a/paddle/math/tests/test_SparseMatrix.cpp b/paddle/math/tests/test_SparseMatrix.cpp deleted file mode 100644 index dbcbeb8d506cf22c026bb7299bf7f71de488cb4a..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_SparseMatrix.cpp +++ /dev/null @@ -1,565 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "test_matrixUtil.h" - -using namespace paddle; // NOLINT - -TEST(Matrix, CopyCpuMatrixToSparseMatrix) { - const size_t HEIGHT = 20; - const size_t WIDTH = 10; - const size_t WIDTH_TEST = 15; - MatrixPtr testMatrix( - new CpuSparseMatrix(HEIGHT, WIDTH, HEIGHT * 5, FLOAT_VALUE, SPARSE_CSR)); - MatrixPtr testCpuMatrix(new CpuMatrix(HEIGHT, WIDTH)); - testCpuMatrix->randomizeUniform(); - testMatrix->copyFrom(*testCpuMatrix, HPPL_STREAM_DEFAULT); - MatrixPtr mulCpuMatrix(new CpuMatrix(WIDTH, WIDTH_TEST)); - mulCpuMatrix->randomizeUniform(); - MatrixPtr ret1(new CpuMatrix(HEIGHT, WIDTH_TEST)), - ret2(new CpuMatrix(HEIGHT, WIDTH_TEST)); - ret1->zeroMem(); - ret2->zeroMem(); - ret1->mul(*testMatrix, *mulCpuMatrix, 1.0, 1.0); - ret2->mul(*testCpuMatrix, *mulCpuMatrix, 1.0, 1.0); - checkMatrixEqual(ret1, ret2); -} - -struct MatrixPara { - size_t height; - size_t width; - bool trans; - bool sparse; - size_t nnz; - SparseFormat format; -}; - -#ifdef PADDLE_WITH_CUDA -void test_sparse_matrix_mul(MatrixPara paraA, - MatrixPara paraB, - MatrixPara paraC) { - // for cpu sparse matrix mul - MatrixPtr cpuMatrixA, cpuMatrixB, cpuMatrixC, gpuMatrixC_d2h; - // for gpu sparse matrix mul - MatrixPtr gpuMatrixA, gpuMatrixB, gpuMatrixC; - // for cpu dense matrix mul - MatrixPtr cpuDenseA, cpuDenseB, cpuDenseC; - - if (paraA.sparse) { - cpuMatrixA = Matrix::createSparseMatrix(paraA.height, - paraA.width, - paraA.nnz, - FLOAT_VALUE, - paraA.format, - paraA.trans, - false); - gpuMatrixA = Matrix::createSparseMatrix(paraA.height, - paraA.width, - paraA.nnz, - FLOAT_VALUE, - paraA.format, - paraA.trans, - true); - } else { - cpuMatrixA = Matrix::create(paraA.height, paraA.width, paraA.trans, false); - gpuMatrixA = Matrix::create(paraA.height, paraA.width, paraA.trans, true); - } - cpuDenseA = Matrix::create(paraA.height, paraA.width, paraA.trans, false); - - if (paraB.sparse) { - cpuMatrixB = Matrix::createSparseMatrix(paraB.height, - paraB.width, - paraB.nnz, - FLOAT_VALUE, - paraB.format, - paraB.trans, - false); - gpuMatrixB = Matrix::createSparseMatrix(paraB.height, - paraB.width, - paraB.nnz, - FLOAT_VALUE, - paraB.format, - paraB.trans, - true); - } else { - cpuMatrixB = Matrix::create(paraB.height, paraB.width, paraB.trans, false); - gpuMatrixB = Matrix::create(paraB.height, paraB.width, paraB.trans, true); - } - cpuDenseB = Matrix::create(paraB.height, paraB.width, paraB.trans, false); - - if (paraC.sparse) { - cpuMatrixC = Matrix::createSparseMatrix(paraC.height, - paraC.width, - paraC.nnz, - FLOAT_VALUE, - paraC.format, - paraC.trans, - false); - gpuMatrixC = Matrix::createSparseMatrix(paraC.height, - paraC.width, - paraC.nnz, - FLOAT_VALUE, - paraC.format, - paraC.trans, - true); - gpuMatrixC_d2h = Matrix::createSparseMatrix(paraC.height, - paraC.width, - paraC.nnz, - FLOAT_VALUE, - paraC.format, - paraC.trans, - false); - } else { - cpuMatrixC = Matrix::create(paraC.height, paraC.width, paraC.trans, false); - gpuMatrixC = Matrix::create(paraC.height, paraC.width, paraC.trans, true); - gpuMatrixC_d2h = - Matrix::create(paraC.height, paraC.width, paraC.trans, false); - } - cpuDenseC = Matrix::create(paraC.height, paraC.width, paraC.trans, false); - - /*matrix init*/ - hl_stream_t stream(HPPL_STREAM_1); - cpuMatrixA->randomizeUniform(); - cpuMatrixB->randomizeUniform(); - cpuMatrixC->randomizeUniform(); - - gpuMatrixA->copyFrom(*cpuMatrixA, stream); - gpuMatrixB->copyFrom(*cpuMatrixB, stream); - gpuMatrixC->copyFrom(*cpuMatrixC, stream); - - cpuDenseA->copyFrom(*cpuMatrixA); - cpuDenseB->copyFrom(*cpuMatrixB); - cpuDenseC->copyFrom(*cpuMatrixC); - - hl_stream_synchronize(stream); - - /*matrix mul*/ - cpuMatrixC->mul(*cpuMatrixA, *cpuMatrixB, 1.0, 1.0); - gpuMatrixC->mul(*gpuMatrixA, *gpuMatrixB, 1.0, 1.0); - cpuDenseC->mul(*cpuDenseA, *cpuDenseB, 1.0, 1.0); - - gpuMatrixC_d2h->copyFrom(*gpuMatrixC, stream); - hl_stream_synchronize(stream); - - /*check result*/ - if (paraC.sparse) { - checkSMatrixEqual( - std::dynamic_pointer_cast(cpuMatrixC), - std::dynamic_pointer_cast(gpuMatrixC_d2h)); - checkSMatrixEqual2Dense( - std::dynamic_pointer_cast(cpuMatrixC), - std::dynamic_pointer_cast(cpuDenseC)); - } else { - checkMatrixEqual(cpuMatrixC, gpuMatrixC_d2h); - checkMatrixEqual(cpuMatrixC, cpuDenseC); - } -} - -TEST(Matrix, SparseMatrixMul) { - const size_t DIM_M = 4; - const size_t DIM_N = 4; - const size_t DIM_K = 8; - const size_t NNZ = 5; - for (auto format : {SPARSE_CSC, SPARSE_CSR}) { - std::string str_format = format == SPARSE_CSC ? "CSC" : "CSR"; - LOG(INFO) << "test dense mul " << str_format; - test_sparse_matrix_mul( - {DIM_M, DIM_K, /*trans*/ false, /*sparse*/ false, NNZ, format}, - {DIM_K, DIM_N, /*trans*/ false, /*sparse*/ true, NNZ, format}, - {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}); - - LOG(INFO) << "test dense mul " << str_format << " trans"; - test_sparse_matrix_mul( - {DIM_M, DIM_K, /*trans*/ false, /*sparse*/ false, NNZ, format}, - {DIM_N, DIM_K, /*trans*/ true, /*sparse*/ true, NNZ, format}, - {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}); - - LOG(INFO) << "test dense mul dense 2 " << str_format; - test_sparse_matrix_mul( - {DIM_M, DIM_K, /*trans*/ false, /*sparse*/ false, NNZ, format}, - {DIM_K, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}, - {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ true, NNZ, format}); - - LOG(INFO) << "test denseT mul dense 2 " << str_format; - test_sparse_matrix_mul( - {DIM_K, DIM_M, /*trans*/ true, /*sparse*/ false, NNZ, format}, - {DIM_K, DIM_N, /*trans*/ false, /*sparse*/ false, NNZ, format}, - {DIM_M, DIM_N, /*trans*/ false, /*sparse*/ true, NNZ, format}); - } -} - -TEST(Matrix, CopySparseMatrixToGpuSparseMatrix) { - const size_t HEIGHT = 20; - const size_t WIDTH = 10; - const size_t WIDTH_TEST = 15; - MatrixPtr testMatrix( - new CpuSparseMatrix(HEIGHT, WIDTH, HEIGHT * 2, FLOAT_VALUE, SPARSE_CSR)); - MatrixPtr testCpuMatrix(new CpuMatrix(HEIGHT, WIDTH)); - testCpuMatrix->randomizeUniform(); - testMatrix->copyFrom(*testCpuMatrix, HPPL_STREAM_DEFAULT); - - MatrixPtr testGpuMatrix = testMatrix->clone(HEIGHT, WIDTH, true); - hl_stream_t gpuStream(HPPL_STREAM_3); - testGpuMatrix->copyFrom(*testMatrix, gpuStream); - hl_stream_synchronize(gpuStream); - - MatrixPtr mulCpuMatrix(new CpuMatrix(WIDTH, WIDTH_TEST)); - mulCpuMatrix->randomizeUniform(); - MatrixPtr mulGpuMatrix(new GpuMatrix(WIDTH, WIDTH_TEST)); - mulGpuMatrix->copyFrom(*mulCpuMatrix); - MatrixPtr ret1(new CpuMatrix(HEIGHT, WIDTH_TEST)); - MatrixPtr ret2(new GpuMatrix(HEIGHT, WIDTH_TEST)); - ret1->zeroMem(); - ret2->zeroMem(); - ret1->mul(*testMatrix, *mulCpuMatrix, 1.0, 1.0); - ret2->mul(*testGpuMatrix, *mulGpuMatrix, 1.0, 1.0); - checkMatrixEqual(ret1, ret2); -} - -#endif - -TEST(Matrix, SparseMatrixTranspose) { - for (auto height : {10, 50, 100}) { - for (auto width : {10, 50, 100}) { - auto nnz = height * width; - for (auto valueType : {FLOAT_VALUE, NO_VALUE}) { - for (auto format : {SPARSE_CSR, SPARSE_CSC}) { - for (auto sparseRate : {0.1, 0.2, 0.5}) { - MatrixPtr matA = Matrix::createSparseMatrix( - height, width, size_t(nnz * sparseRate), valueType, format); - MatrixPtr matB(new CpuSparseMatrix( - width, height, size_t(nnz * sparseRate), valueType, format)); - matA->randomizeUniform(); - matA->transpose(matB, false); - - /*dense matrix transpose*/ - CpuMatrixPtr matC(new CpuMatrix(height, width)); - matC->copyFrom(*matA); - MatrixPtr matD(new CpuMatrix(width, height)); - matC->transpose(matD, false); - - /*check result*/ - checkSMatrixEqual2Dense( - std::dynamic_pointer_cast(matB), - std::dynamic_pointer_cast(matD)); - } - } - } - } - } -} - -TEST(Matrix, CpuSparseMatrixSubMatrix) { - const size_t HEIGHT = 10; - const size_t WIDTH = 10; - const size_t NNZ = HEIGHT * WIDTH; - for (auto valueType : {FLOAT_VALUE, NO_VALUE}) { - size_t startRow = 3; - size_t rowNum = 2; - real sparseRate = 0.1; - /*sparse matrix init and get subMatrix*/ - CpuSparseMatrixPtr matA = std::make_shared( - HEIGHT, WIDTH, size_t(NNZ * sparseRate), valueType, SPARSE_CSR); - matA->randomizeUniform(); - CpuSparseMatrixPtr matB = std::dynamic_pointer_cast( - matA->subMatrix(startRow, rowNum)); - - int start = matA->getRows()[startRow]; - int end = matA->getRows()[startRow + rowNum]; - - /*compare two matrix*/ - ASSERT_EQ(matB->getElementCnt(), size_t(end - start)); - if (valueType == FLOAT_VALUE) { - for (size_t i = 0; i < matB->getElementCnt(); i++) { - ASSERT_FLOAT_EQ(matB->getValue()[start + i], - matA->getValue()[start + i]); - } - } - - for (size_t i = 0; i < matB->getElementCnt(); i++) { - ASSERT_EQ(matB->getCols()[start + i], matA->getCols()[start + i]); - } - for (size_t i = 0; i < rowNum; i++) { - ASSERT_EQ(matB->getRows()[i], matA->getRows()[startRow + i]); - } - } -} - -void sparseValid( - int* major, int* minor, size_t nnz, size_t majorLen, size_t minorLen) { - CHECK_EQ(nnz, size_t(major[majorLen - 1])); - CHECK_EQ(nnz, minorLen); - for (size_t i = 0; i < majorLen - 1; i++) { - EXPECT_LE(major[i], major[i + 1]); - for (int j = major[i]; j < major[i + 1] - 1; j++) { - EXPECT_LE(minor[j], minor[j + 1]); - } - } -} - -TEST(Matrix, CpuSparseMatrixRandUniform) { - const size_t HEIGHT = 5; - const size_t WIDTH = 10; - const size_t NNZ = HEIGHT * WIDTH; - int* major = nullptr; - int* minor = nullptr; - size_t majorLen = 0; - size_t minorLen = 0; - size_t nnz = 0; - for (auto valueType : {NO_VALUE, FLOAT_VALUE}) { - for (auto format : {SPARSE_CSR, SPARSE_CSC}) { - CpuSparseMatrixPtr matA = std::make_shared( - HEIGHT, WIDTH, size_t(NNZ * 0.1), valueType, format); - matA->randomizeUniform(); - nnz = matA->getElementCnt(); - if (format == SPARSE_CSR) { - majorLen = matA->getHeight() + 1; - minorLen = matA->getElementCnt(); - major = matA->getRows(); - minor = matA->getCols(); - } else { - majorLen = matA->getWidth() + 1; - minorLen = matA->getElementCnt(); - major = matA->getCols(); - minor = matA->getRows(); - } - sparseValid(major, minor, nnz, majorLen, minorLen); - } - } -} - -TEST(Matrix, CpuSparseMatrixCopyFrom) { - size_t height = 10; - size_t width = 8; - int64_t indices[11] = {0, 1, 5, 5, 9, 13, 15, 17, 19, 30, 32}; - sparse_non_value_t data[32]; - for (size_t i = 0; i < 32; i++) { - data[i].col = ::rand() % width; - } - CpuSparseMatrixPtr mat = std::make_shared( - height, width, 32, NO_VALUE, SPARSE_CSR, false); - mat->copyFrom(indices, data); - - /*compare indices*/ - size_t sum = 0; - CHECK_EQ(sum, size_t(mat->getRows()[0])); - for (size_t i = 1; i < height + 1; i++) { - sum += indices[i] - indices[i - 1]; - CHECK_EQ(sum, size_t(mat->getRows()[i])); - } - CHECK_EQ(mat->getElementCnt(), size_t(indices[height] - indices[0])); - for (size_t i = 0; i < mat->getElementCnt(); i++) { - CHECK_EQ(size_t(mat->getCols()[i]), size_t(data[i].col)); - } -} - -TEST(Matrix, SparseMatrixCSRFormatTrimFrom) { - size_t height = 10; - size_t width = 8; - int64_t indices[11] = {0, 1, 5, 5, 9, 13, 15, 17, 19, 27, 32}; - sparse_float_value_t data[32]; - int value[32] = { - 1, // row_0 : 1 - 5, 3, 1, 6, // row_1 : 4 - 0, 1, 2, 3, // row_3 : 4 - 4, 5, 6, 7, // row_4 : 4 - 2, 3, // row_5 : 2 - 3, 5, // row_6 : 2 - 0, 1, // row_7 : 2 - 0, 1, 2, 3, 4, 5, 6, 7, // row_8 : 8 - 2, 4, 7, 3, 1 // row_9 : 5 - }; - for (size_t i = 0; i < 32; i++) { - data[i].col = value[i]; - data[i].value = float(value[i]); - } - CpuSparseMatrixPtr mat = std::make_shared( - height, width, 32, FLOAT_VALUE, SPARSE_CSR, false); - mat->copyFrom(indices, data); - - /*compare indices*/ - size_t sum = 0; - CHECK_EQ(sum, size_t(mat->getRows()[0])); - for (size_t i = 1; i < height + 1; i++) { - sum += indices[i] - indices[i - 1]; - CHECK_EQ(sum, size_t(mat->getRows()[i])); - } - CHECK_EQ(mat->getElementCnt(), size_t(indices[height] - indices[0])); - for (size_t i = 0; i < mat->getElementCnt(); i++) { - CHECK_EQ(size_t(mat->getCols()[i]), size_t(data[i].col)); - } - - size_t trimedWidth = 4; - int64_t trimedIndices[11] = {0, 1, 3, 3, 7, 7, 9, 10, 12, 16, 19}; - sparse_float_value_t trimedData[19]; - int trimedValue[19] = { - 1, // row_0 : 1 - 3, - 1, // row_1 : 2 - 0, - 1, - 2, - 3, // row_3 : 4 - 2, - 3, // row_5 : 2 - 3, // row_6 : 1 - 0, - 1, // row_7 : 2 - 0, - 1, - 2, - 3, // row_8 : 4 - 2, - 3, - 1 // row_9 : 3 - }; - for (size_t i = 0; i < 19; i++) { - trimedData[i].col = trimedValue[i]; - trimedData[i].value = float(trimedValue[i]); - } - CpuSparseMatrixPtr matA = std::make_shared( - height, trimedWidth, 19, FLOAT_VALUE, SPARSE_CSR, false); - matA->copyFrom(trimedIndices, trimedData); - - /*compare indices*/ - sum = 0; - CHECK_EQ(sum, size_t(matA->getRows()[0])); - for (size_t i = 1; i < height + 1; i++) { - sum += trimedIndices[i] - trimedIndices[i - 1]; - CHECK_EQ(sum, size_t(matA->getRows()[i])); - } - CHECK_EQ(matA->getElementCnt(), - size_t(trimedIndices[height] - trimedIndices[0])); - for (size_t i = 0; i < matA->getElementCnt(); i++) { - CHECK_EQ(size_t(matA->getCols()[i]), size_t(trimedData[i].col)); - } - - CpuSparseMatrixPtr matB = std::make_shared( - height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, false); - matB->trimFrom(*mat); - checkSMatrixEqual2(matA, matB); - -#ifdef PADDLE_WITH_CUDA - GpuSparseMatrixPtr matC = std::make_shared( - height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, true); - matC->trimFrom(*mat); - - CpuSparseMatrixPtr matD = - std::make_shared(height, - trimedWidth, - matC->getElementCnt(), - FLOAT_VALUE, - SPARSE_CSR, - false); - matD->copyFrom(*matC, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - checkSMatrixEqual2(matA, matD); -#endif -} - -TEST(Matrix, SparseMatrixCSCFormatTrimFrom) { - size_t height = 8; - size_t width = 10; - int indices[11] = {0, 1, 5, 5, 9, 13, 15, 17, 19, 27, 32}; - int value[32] = { - 1, // col_0 : 1 - 5, 3, 1, 6, // col_1 : 4 - 0, 1, 2, 3, // col_3 : 4 - 4, 5, 6, 7, // col_4 : 4 - 2, 3, // col_5 : 2 - 3, 5, // col_6 : 2 - 0, 1, // col_7 : 2 - 0, 1, 2, 3, 4, 5, 6, 7, // col_8 : 8 - 2, 4, 7, 3, 1 // col_9 : 5 - }; - std::vector rows(value, value + 32); - std::vector cols(indices, indices + 11); - std::vector values(value, value + 32); - CpuSparseMatrixPtr mat = std::make_shared( - height, width, 32, FLOAT_VALUE, SPARSE_CSC, false); - mat->copyFrom(rows, cols, values); - - /*compare indices*/ - size_t sum = 0; - CHECK_EQ(sum, size_t(mat->getCols()[0])); - for (size_t i = 1; i < width + 1; i++) { - sum += indices[i] - indices[i - 1]; - CHECK_EQ(sum, size_t(mat->getCols()[i])); - } - CHECK_EQ(mat->getElementCnt(), size_t(indices[width] - indices[0])); - for (size_t i = 0; i < mat->getElementCnt(); i++) { - CHECK_EQ(size_t(mat->getRows()[i]), size_t(value[i])); - } - - size_t trimedWidth = 5; - int trimedIndices[6] = {0, 1, 5, 5, 9, 13}; - int trimedValue[13] = { - 1, // col_0 : 1 - 5, - 3, - 1, - 6, // col_1 : 4 - 0, - 1, - 2, - 3, // col_3 : 4 - 4, - 5, - 6, - 7 // col_4 : 4 - }; - std::vector rowsA(trimedValue, trimedValue + 13); - std::vector colsA(trimedIndices, trimedIndices + 6); - std::vector valuesA(trimedValue, trimedValue + 13); - CpuSparseMatrixPtr matA = std::make_shared( - height, trimedWidth, 13, FLOAT_VALUE, SPARSE_CSC, false); - matA->copyFrom(rowsA, colsA, valuesA); - - /*compare indices*/ - sum = 0; - CHECK_EQ(sum, size_t(matA->getCols()[0])); - for (size_t i = 1; i < trimedWidth + 1; i++) { - sum += trimedIndices[i] - trimedIndices[i - 1]; - CHECK_EQ(sum, size_t(matA->getCols()[i])); - } - CHECK_EQ(matA->getElementCnt(), - size_t(trimedIndices[trimedWidth] - trimedIndices[0])); - for (size_t i = 0; i < matA->getElementCnt(); i++) { - CHECK_EQ(size_t(matA->getRows()[i]), size_t(rowsA[i])); - } - - CpuSparseMatrixPtr matB = std::make_shared( - height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, false); - matB->trimFrom(*mat); - checkSMatrixEqual2(matA, matB); - -#ifdef PADDLE_WITH_CUDA - GpuSparseMatrixPtr matC = std::make_shared( - height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, true); - matC->trimFrom(*mat); - - CpuSparseMatrixPtr matD = - std::make_shared(height, - trimedWidth, - matC->getElementCnt(), - FLOAT_VALUE, - SPARSE_CSC, - false); - matD->copyFrom(*matC, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - checkSMatrixEqual2(matA, matD); -#endif -} diff --git a/paddle/math/tests/test_Tensor.cu b/paddle/math/tests/test_Tensor.cu deleted file mode 100644 index acb2da86d0f41d12fced97d1ddaf5be00959fb82..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_Tensor.cu +++ /dev/null @@ -1,1162 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "TensorCheck.h" -#include "paddle/math/Matrix.h" - -using paddle::Matrix; -using paddle::CpuMatrix; -using paddle::GpuMatrix; -using paddle::CpuVector; -using paddle::GpuVector; -using paddle::CpuIVector; -using paddle::GpuIVector; -using autotest::TensorCheckEqual; -using autotest::TensorCheckErr; - -#define INIT_UNARY(A1, A2) \ - Tensor A1(height, width); \ - Tensor A2(height, width); \ - A1.randomizeUniform(); \ - A2.copyFrom(A1) -#define INIT_BINARY(A1, A2, B) \ - INIT_UNARY(A1, A2); \ - Tensor B(height, width); \ - B.randomizeUniform() -#define INIT_TERNARY(A1, A2, B, C) \ - INIT_BINARY(A1, A2, B); \ - Tensor C(height, width); \ - C.randomizeUniform() -#define INIT_QUATERNARY(A1, A2, B, C, D) \ - INIT_TERNARY(A1, A2, B, C); \ - Tensor D(height, width); \ - D.randomizeUniform() - -template -struct TestUnaryMatrix { - typedef std::function UnaryFunc; - - explicit TestUnaryMatrix(UnaryFunc testUnaryFunc) { - for (auto height : {1, 11, 73, 128, 200, 330}) { - for (auto width : {1, 32, 100, 512, 1000, 3210}) { - LOG(INFO) << " height=" << height << " width=" << width; - INIT_UNARY(A1, A2); - testUnaryFunc(A1, A2); - } - } - } -}; - -template -struct TestBinaryMatrix { - typedef std::function BinaryFunc; - - explicit TestBinaryMatrix(BinaryFunc testBinaryFunc) { - for (auto height : {1, 11, 73, 128, 200, 330}) { - for (auto width : {1, 32, 100, 512, 1000, 3210}) { - LOG(INFO) << " height=" << height << " width=" << width; - INIT_BINARY(A1, A2, B); - testBinaryFunc(A1, A2, B); - } - } - } -}; - -template -struct TestTernaryMatrix { - typedef std::function - TernaryFunc; - - explicit TestTernaryMatrix(TernaryFunc testTernaryFunc) { - for (auto height : {1, 11, 73, 128, 200, 330}) { - for (auto width : {1, 32, 100, 512, 1000, 3210}) { - LOG(INFO) << " height=" << height << " width=" << width; - INIT_TERNARY(A1, A2, B, C); - testTernaryFunc(A1, A2, B, C); - } - } - } -}; - -template -struct TestQuaternaryMatrix { - typedef std::function - QuaternaryFunc; - - explicit TestQuaternaryMatrix(QuaternaryFunc testQuaternaryFunc) { - for (auto height : {1, 11, 73, 128, 200, 330}) { - for (auto width : {1, 32, 100, 512, 1000, 3210}) { - LOG(INFO) << " height=" << height << " width=" << width; - INIT_QUATERNARY(A1, A2, B, C, D); - testQuaternaryFunc(A1, A2, B, C, D); - } - } - } -}; - -template -struct TestUnaryVectorT { - typedef std::function UnaryFunc; - - explicit TestUnaryVectorT(UnaryFunc testUnaryFunc) { - for (auto size : {1, 11, 73, 128, 200, 330, 512, 1000, 4210}) { - LOG(INFO) << " size=" << size; - Tensor A1(size); - Tensor A2(size); - if (typeid(T) == typeid(real)) { - A1.rand(); - } else { - A1.rand(1000); - } - A2.copyFrom(A1); - testUnaryFunc(A1, A2); - } - } -}; - -void SetTensorValue(Matrix& matrix, real value) { - int height = matrix.getHeight(); - int width = matrix.getWidth(); - int stride = matrix.getStride(); - real* data = matrix.getData(); - for (int i = 0; i < height; i++) { - int j = rand() % width; // NOLINT - if (typeid(matrix) == typeid(CpuMatrix)) { - data[i * stride + j] = value; - } else if (typeid(matrix) == typeid(GpuMatrix)) { - hl_memcpy(&data[i * stride + j], &value, sizeof(real)); - } else { - } - } -} - -template -void testTensorAddScalar(Tensor& A1, Tensor& A2) { - real p1 = 2.5; - real p2 = 3.0; - A1.add(p1); // a += p - A2 += p1; - TensorCheckEqual(A1, A2); - - A1.add(p1, p2); // a = a * p1 + p2 - A2 = A2 * p1 + p2; - TensorCheckEqual(A1, A2); -} - -template -void testTensorSubScalar(Tensor& A1, Tensor& A2) { - real p = 2.5; - A1.subScalar(p); // a -= p - A2 -= p; - TensorCheckEqual(A1, A2); -} - -template -void testTensorMulScalar(Tensor& A1, Tensor& A2) { - real p = 2.5; - A1.mulScalar(p); // a *= p - A2 *= p; - TensorCheckEqual(A1, A2); - - real learningRate = 0.7f; - real decayRate = 1.2f; - A1.applyL2(learningRate, decayRate); - A2 = A2 * (1.0f / (1.0f + learningRate * decayRate)); - TensorCheckEqual(A1, A2); -} - -template -void testTensorDivScalar(Tensor& A1, Tensor& A2) { - real p = 2.5; - A1.divScalar(p); // a /= p - A2 /= p; - TensorCheckEqual(A1, A2); -} - -template -void testTensorNeg(Tensor& A1, Tensor& A2) { - A1.neg(); // a = -a - A2 = -A2; - TensorCheckEqual(A1, A2); -} - -template -void testTensorAbs(Tensor& A1, Tensor& A2) { - A1.abs2(); // a = a > 0 ? a : -a - A2 = A2.abs(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorSquare(Tensor& A1, Tensor& A2) { - A1.square2(); // a = a * a - A2 = A2.square(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorReciprocal(Tensor& A1, Tensor& A2) { - A1.reciprocal2(); // a = 1.0f / a - A2 = A2.reciprocal(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorSign(Tensor& A1, Tensor& A2) { - A1.sign2(); // a = (a > 0) - (a < 0) - A2 = A2.sign(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorAssign(Tensor& A1, Tensor& A2) { - A1.assign(1.5); // a = p - A2 = A2.constant(1.5); - TensorCheckEqual(A1, A2); - - A1.one(); // a = 1 - A2 = A2.constant(1.0); - TensorCheckEqual(A1, A2); - - A1.zero(); // a = 0 - A2 = A2.constant(0.0); - TensorCheckEqual(A1, A2); -} - -template -void testUnaryBaseOp(Tensor& A1, Tensor& A2) { - testTensorAddScalar(A1, A2); - testTensorSubScalar(A1, A2); - testTensorMulScalar(A1, A2); - testTensorDivScalar(A1, A2); - testTensorNeg(A1, A2); - testTensorAbs(A1, A2); - testTensorSquare(A1, A2); - testTensorReciprocal(A1, A2); - testTensorSign(A1, A2); - testTensorAssign(A1, A2); -} - -template -void testUnaryBaseOpInt(Tensor& A1, Tensor& A2) { - A1.add(2); // a += p - A2 += 2; - TensorCheckEqual(A1, A2); - - A1.add(3, 2); // a = a * p1 + p2 - A2 = A2 * 3 + 2; - TensorCheckEqual(A1, A2); - - testTensorNeg(A1, A2); - testTensorAbs(A1, A2); -} - -TEST(Unary, BaseOp) { - TestUnaryMatrix testCpuMatrix(testUnaryBaseOp); - TestUnaryVectorT testCpuVector(testUnaryBaseOp); - TestUnaryVectorT testCpuIVector( - testUnaryBaseOpInt); - -#ifdef PADDLE_WITH_GPU - TestUnaryMatrix testGpuMatrix(testUnaryBaseOp); - TestUnaryVectorT testGpuVector(testUnaryBaseOp); - TestUnaryVectorT testGpuIVector( - testUnaryBaseOpInt); -#endif -} - -template -void testTensorExp(Tensor& A1, Tensor& A2) { - A1.exp2(); // a = exp(a) - A2 = A2.exp(); - TensorCheckErr(A1, A2); -} - -template -void testTensorLog(Tensor& A1, Tensor& A2) { - A1.log2(); // a = log(a) - A2 = A2.log(); - TensorCheckErr(A1, A2); -} - -template -void testTensorSqrt(Tensor& A1, Tensor& A2) { - A1.sqrt2(); // a = sqrt(a) - A2 = A2.sqrt(); - TensorCheckErr(A1, A2); -} - -template -void testTensorPow(Tensor& A1, Tensor& A2) { - A1.pow2(3.2); // a = pow(a, p) - A2 = A2.pow(3.2); - TensorCheckErr(A1, A2); -} - -template -void testUnayrMathOp(Tensor& A1, Tensor& A2) { - testTensorExp(A1, A2); - testTensorLog(A1, A2); - testTensorSqrt(A1, A2); - testTensorPow(A1, A2); -} - -TEST(Unary, MathOp) { - TestUnaryMatrix testCpu(testUnayrMathOp); - -#ifdef PADDLE_WITH_GPU - TestUnaryMatrix testGpu(testUnayrMathOp); -#endif -} - -template -void testTensorClip(Tensor& A1, Tensor& A2) { - real p1 = 0.003f; - real p2 = 0.877f; - A1.clip(p1, p2); // a = a < p1 ? p1 : (a > p2 ? p2 : a) - // A2 = A2.min(0.877f).max(0.003f); - A2 = (A2 < p1).condition(p1, (A2 > p2).condition(p2, A2)); - TensorCheckEqual(A1, A2); -} - -template -void testTensorBiggerThanScalar(Tensor& A1, Tensor& A2) { - real p = 0.5f; - A1.biggerThanScalar(p); // a = a > p ? 1.0f : 0.0f - A2 = (A2 > p).condition((real)1.0, (real)0.0); - TensorCheckEqual(A1, A2); -} - -template -void testTensorapplyL1(Tensor& A1, Tensor& A2) { - /** - * T lambda = p; - * a = (a > lambda) ? (a - lambda) - * : (a < -lambda) ? (a + lambda) : 0 - * - * p = learningRate * decayRate; - */ - real learningRate = 0.7f; - real decayRate = 0.6f; - A1.applyL1(learningRate, decayRate); - A2 = (A2 > (learningRate * decayRate)) - .condition( - (A2 - (learningRate * decayRate)), - (A2 < -(learningRate * decayRate)) - .condition((A2 + (learningRate * decayRate)), (real)0.0)); - TensorCheckEqual(A1, A2); -} - -template -void testUnayrCompareOp(Tensor& A1, Tensor& A2) { - testTensorClip(A1, A2); - testTensorBiggerThanScalar(A1, A2); - - A1.randomizeUniform(); - A1.subScalar(0.5f); - A2.copyFrom(A1); - testTensorapplyL1(A1, A2); -} - -TEST(Unary, CompareOp) { - TestUnaryMatrix testCpu(testUnayrCompareOp); - -#ifdef PADDLE_WITH_GPU - TestUnaryMatrix testGpu(testUnayrCompareOp); -#endif -} - -template -void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B) { - real p1 = 2.5; - real p2 = 3.2; - A1.add(B); // a += b - A2 += B; - TensorCheckEqual(A1, A2); - - A1.add(B, p1); // a += b * p - A2 += B * p1; - TensorCheckEqual(A1, A2); - - A1.add(B, p1, p2); // a = p1 * a + p2 * b - A2 = A2 * p1 + B * p2; - TensorCheckEqual(A1, A2); - - A1.addScalar(B, p1); // a = b + p - A2 = B + p1; - TensorCheckEqual(A1, A2); - - A1.addSquare(B, p1); // a += p * b * b - A2 += B.constant(p1) * B * B; - TensorCheckEqual(A1, A2); - - A1.decayAddSquare(B, p1, p2); // a = p1 * a + p2 * b * b - A2 = A2 * p1 + B.constant(p2) * B * B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B) { - real p = 2.5; - A1.sub(B); // a -= b - A2 -= B; - TensorCheckEqual(A1, A2); - - A1.sub(B, p); // a -= b * p - A2 -= B * p; - TensorCheckEqual(A1, A2); - - A1.subScalar(B, p); // a = b - p - A2 = B - p; - TensorCheckEqual(A1, A2); -} - -template -void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B) { - real p = 2.5; - A1.mulScalar(B, p); // a = b * p - A2 = B * p; - TensorCheckEqual(A1, A2); - - A1.dotMulSquare(B); // a *= b * b - A2 *= B * B; - TensorCheckEqual(A1, A2); - - A1.dotSquareMul(B); // a = a * a * b - A2 = A2 * A2 * B; - TensorCheckEqual(A1, A2); - - A1.dotMul(B); // a *= b - A2 *= B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B) { - real p = 2.5; - A1.divScalar(B, p); // a = b / p - A2 = B / p; - TensorCheckEqual(A1, A2); - - A1.scalarDiv(B, p); // a = p / b - A2 = B.constant(p) / B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorAssign(Tensor& A1, Tensor& A2, Tensor& B) { - A1.assign(B); // a = b - A2 = B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorSquare(Tensor& A1, Tensor& A2, Tensor& B) { - B.square2(A1); // b = a * a - A2 = B.square(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorSquareDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.squareDerivative(B); // a *= 2.0 * b - A2 = A2 * (real)2.0 * B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B) { - B.reciprocal2(A1); // b = 1.0f / a - A2 = B.reciprocal(); - TensorCheckEqual(A1, A2); - - real p1 = 0.58; - real p2 = 0.32; - A1.reciprocal2(B, p1, p2); // a = 1 / (p1 * b + p2) - A2 = (B * p1 + p2).reciprocal(); - TensorCheckEqual(A1, A2); - - real learningRate = 0.7f; - real decayRate = 1.2f; - A1.applyL2(B, learningRate, decayRate); // a *= (1.0f / (1.0f + p * b)) - A2 *= (B.constant(1.0f) + B.constant(learningRate * decayRate) * B) - .reciprocal(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorReciprocalDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.reciprocalDerivative(B); // a *= -b * b - A2 *= (-B) * B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorSign(Tensor& A1, Tensor& A2, Tensor& B) { - B.sign2(A1); // b = a > 0.0f ? 1.0f : -1.0f - A2 = B.sign(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorAbs(Tensor& A1, Tensor& A2, Tensor& B) { - B.abs2(A1); // b = a > 0.0f ? a : -a - A2 = B.abs(); - TensorCheckEqual(A1, A2); -} - -template -void testBinaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B) { - testTensorAdd(A1, A2, B); - testTensorSub(A1, A2, B); - testTensorMul(A1, A2, B); - testTensorDiv(A1, A2, B); - testTensorSquare(A1, A2, B); - testTensorSquareDerivative(A1, A2, B); - testTensorReciprocal(A1, A2, B); - testTensorReciprocalDerivative(A1, A2, B); - testTensorAbs(A1, A2, B); - testTensorSign(A1, A2, B); - testTensorAssign(A1, A2, B); -} - -TEST(Binary, BaseOp) { - TestBinaryMatrix testCpu(testBinaryBaseOp); - -#ifdef PADDLE_WITH_GPU - TestBinaryMatrix testGpu(testBinaryBaseOp); -#endif -} - -template -void testTensorExp(Tensor& A1, Tensor& A2, Tensor& B) { - // a = exp(b) - A1.exp2(B); - A2 = B.exp(); - TensorCheckErr(A1, A2); -} - -template -void testTensorExpDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.expDerivative(B); // a *= b - A2 *= B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorLog(Tensor& A1, Tensor& A2, Tensor& B) { - // a = log(b) - A1.log2(B); - A2 = B.log(); - TensorCheckErr(A1, A2); -} - -template -void testTensorSqrt(Tensor& A1, Tensor& A2, Tensor& B) { - // a = sqrt(b) - A1.sqrt2(B); - A2 = B.sqrt(); - TensorCheckErr(A1, A2); -} - -template -void testTensorInvSqrt(Tensor& A1, Tensor& A2, Tensor& B) { - // a = 1.0f / sqrt(b) - A1.invSqrt(B); - A2 = B.sqrt().reciprocal(); - TensorCheckErr(A1, A2); -} - -template -void testTensorPow(Tensor& A1, Tensor& A2, Tensor& B) { - A1.pow2(B, 2.5f); // a = pow(b, p) - A2 = B.pow(2.5f); - TensorCheckErr(A1, A2); -} - -template -void testTensorSoftrelu(Tensor& A1, Tensor& A2, Tensor& B) { - /* - * const T THRESHOLD = 40.0; - * b = log(1.0 + - * exp((a > THRESHOLD) ? THRESHOLD - * : ((a < -THRESHOLD) ? (-THRESHOLD) : a))) - */ - B.softrelu(A1); - - real THRESHOLD = 40.0; - A2 = (B.constant(1.0f) + - (B > THRESHOLD) - .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)) - .exp()) - .log(); - TensorCheckErr(A1, A2); -} - -template -void testTensorSoftreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - /* - * const T THRESHOLD = 40.0; - * a *= (1.0 - exp(-1.0 * ((b > THRESHOLD) - * ? THRESHOLD - * : ((b < -THRESHOLD) ? (-THRESHOLD) : b))))); - */ - A1.softreluDerivative(B); - real THRESHOLD = 40.0; - A2 = A2 * - (B.constant(1.0f) - - (B.constant(-1.0f) * - (B > THRESHOLD) - .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B))) - .exp()); - TensorCheckErr(A1, A2); -} - -template -void testTensorSigmoid(Tensor& A1, Tensor& A2, Tensor& B) { - /* - const T THRESHOLD_MIN = -40.0; - const T THRESHOLD_MAX = 13.0; - T tmp = (a < THRESHOLD_MIN) ? THRESHOLD_MIN - : ((a > THRESHOLD_MAX) ? THRESHOLD_MAX : a); - b = 1.0f / (1.0f + exp(-tmp))) - */ - B.sigmoid(A1); - - const real THRESHOLD_MIN = -40.0; - const real THRESHOLD_MAX = 13.0; - auto tmp = (B < THRESHOLD_MIN) - .condition(THRESHOLD_MIN, - (B > THRESHOLD_MAX).condition(THRESHOLD_MAX, B)); - A2 = (B.constant(1.0f) + (-tmp).exp()).reciprocal(); - TensorCheckErr(A1, A2); -} - -template -void testTensorSigmoidDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.sigmoidDerivative(B); // a *= b * (1 - b) - A2 *= B * (B.constant(1.0f) - B); - TensorCheckEqual(A1, A2); -} - -template -void testTensorTanh(Tensor& A1, Tensor& A2, Tensor& B) { - B.tanh(A1); // b = 2.0 / (1.0 + exp(-2 * a)) - 1.0 - A2 = B.constant(2.0f) / ((B * ((real)-2.0f)).exp() + (real)1.0f) - (real)1.0f; - TensorCheckErr(A1, A2); -} - -template -void testTensorTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.tanhDerivative(B); // a *= 1 - b * b - A2 *= B.constant(1.0f) - B * B; - TensorCheckEqual(A1, A2); -} - -template -void testTensorScaledTanh(Tensor& A1, Tensor& A2, Tensor& B) { - real p1 = 2.5; - real p2 = 3.1; - // b = p1 * (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0) - B.scaledTanh(A1, p1, p2); - A2 = B.constant(p1) * - (B.constant(2.0f) / ((B.constant(-2.0f) * p2 * B).exp() + (real)1.0) - - (real)1.0); - TensorCheckErr(A1, A2); -} - -template -void testTensorScaledTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - real p1 = 2.5; - real p2 = 3.1; - // a *= (p2 / p1) * (p1 * p1 - b * b)); - A1.scaledTanhDerivative(B, p1, p2); - A2 = A2 * (B.constant(p2 / p1) * (B.constant(p1 * p1) - B * B)); - TensorCheckEqual(A1, A2); -} - -template -void testBinaryMathOp(Tensor& A1, Tensor& A2, Tensor& B) { - testTensorTanhDerivative(A1, A2, B); - testTensorScaledTanhDerivative(A1, A2, B); - testTensorSigmoidDerivative(A1, A2, B); - testTensorExpDerivative(A1, A2, B); - testTensorScaledTanh(A1, A2, B); - testTensorTanh(A1, A2, B); - testTensorExp(A1, A2, B); - testTensorLog(A1, A2, B); - testTensorSqrt(A1, A2, B); - testTensorInvSqrt(A1, A2, B); - testTensorPow(A1, A2, B); - - testTensorSoftrelu(A1, A2, B); - testTensorSoftreluDerivative(A1, A2, B); - testTensorSigmoid(A1, A2, B); -} - -TEST(Binary, MathOp) { - TestBinaryMatrix testCpu(testBinaryMathOp); - -#ifdef PADDLE_WITH_GPU - TestBinaryMatrix testGpu(testBinaryMathOp); -#endif -} - -template -void testTensorRelu(Tensor& A1, Tensor& A2, Tensor& B) { - B.relu(A1); // b = a > 0.0f ? a : 0.0f - A2 = (B > (real)0.0f).condition(B, (real)0.0f); - TensorCheckEqual(A1, A2); -} - -template -void testTensorReluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.reluDerivative(B); // a *= (b > 0.0f ? 1.0f : 0.0f) - A2 *= (B > (real)0.0).condition((real)1.0, (real)0.0); - TensorCheckEqual(A1, A2); -} - -template -void testTensorBrelu(Tensor& A1, Tensor& A2, Tensor& B) { - /* - * b = a > p1 ? a : p1 - * b = b < p2 ? b : p2 - * int p1 = 0, p2 = 24; - */ - SetTensorValue(B, 32.0f); - B.brelu(A1); - auto tmp = (B > (real)0.0f).condition(B, (real)0.0f); - A2 = (tmp < (real)24.0f).condition(tmp, (real)24.0f); - TensorCheckEqual(A1, A2); -} - -template -void testTensorBreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - SetTensorValue(B, 32.0f); - /* - * a *= (b > p1 && b < p2) ? 1.0 : 0.0 - * int p1 = 0, p2 = 24; - */ - A1.breluDerivative(B); - A2 *= (B > (real)0.0f && B < (real)24.0f).condition((real)1.0f, (real)0.0f); - TensorCheckEqual(A1, A2); -} - -template -void testTensorAbsDerivative(Tensor& A1, Tensor& A2, Tensor& B) { - A1.absDerivative(B); // a = (b > 0) ? a : (b < 0) ? -a : 0 - A2 = (B > (real)0.0f) - .condition(A2, (B < (real)0.0f).condition(-A2, (real)0.0f)); - TensorCheckEqual(A1, A2); -} - -template -void testTensorIsEqualTo(Tensor& A1, Tensor& A2, Tensor& B) { - real p = 0.613; - SetTensorValue(B, p); - A1.isEqualTo(B, p); // a = (b == p) - A2 = (B == p); - TensorCheckEqual(A1, A2); -} - -template -void testTensorapplyL1(Tensor& A1, Tensor& A2, Tensor& B) { - /** - * T lambda = p * b; - * a = (a > lambda) ? (a - lambda) - * : (a < -lambda) ? (a + lambda) : 0 - * - * p = learningRate * decayRate; - */ - real learningRate = 0.7f; - real decayRate = 0.6f; - A1.applyL1(B, learningRate, decayRate); - auto lambda = B.constant(learningRate * decayRate) * B; - A2 = (A2 > lambda) - .condition((A2 - lambda), - (A2 < -lambda).condition((A2 + lambda), (real)0.0f)); - TensorCheckEqual(A1, A2); -} - -template -void testBinaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B) { - B.subScalar(0.5f); - SetTensorValue(B, 0.0f); - testTensorReluDerivative(A1, A2, B); - - A1.randomizeUniform(); - A2.copyFrom(A1); - testTensorBreluDerivative(A1, A2, B); - - testTensorAbsDerivative(A1, A2, B); - testTensorRelu(A1, A2, B); - testTensorBrelu(A1, A2, B); - testTensorIsEqualTo(A1, A2, B); -} - -TEST(Binary, CompareOp) { - TestBinaryMatrix testCpu(testBinaryCompareOp); - -#ifdef PADDLE_WITH_GPU - TestBinaryMatrix testGpu(testBinaryCompareOp); -#endif -} - -template -void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.add(B, C); // a = b + c - A2 = B + C; - TensorCheckEqual(A1, A2); - - real p1 = 1.5; - real p2 = 2.5; - real p3 = 3.8; - A1.add(B, p1, C, p2); // a = p1 * b + p2 * c - A2 = B * p1 + C * p2; - TensorCheckEqual(A1, A2); - - A1.add2(B, C); // a = a + b + c - A2 = A2 + B + C; - TensorCheckEqual(A1, A2); - - A1.add2(B, C, p1, p2, p3); // a = p1 * a + p2 * b + p3 * c - A2 = A2 * p1 + B * p2 + C * p3; - TensorCheckEqual(A1, A2); - - A1.decayAddSquareMul(B, C, p1, p2); // a = p1 * a + p2 * b * b * c * c - A2 = A2 * p1 + B.constant(p2) * B * B * C * C; - TensorCheckEqual(A1, A2); -} - -template -void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.sub(B, C); // a = b - c - A2 = B - C; - TensorCheckEqual(A1, A2); - - real p1 = 1.5; - real p2 = 2.5; - A1.sub(B, p1, C, p2); // a = p1 * b - p2 * c - A2 = B * p1 - C * p2; - TensorCheckEqual(A1, A2); -} - -template -void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.dotMul(B, C); // a = b * c - A2 = B * C; - TensorCheckEqual(A1, A2); - - A1.dotMulSquare(B, C); // a = b * c * c - A2 = B * C * C; - TensorCheckEqual(A1, A2); - - A1.dotSquareSquare(B, C); // a = b * b * c * c - A2 = B * B * C * C; - TensorCheckEqual(A1, A2); - - real p1 = 1.5; - real p2 = 2.5; - - /* - * T tmp = p1 * b + p2 * c; - * a *= tmp * tmp - */ - A1.dotMulSquareSum(B, C, p1, p2); - auto tmp = B * p1 + C * p2; - A2 *= tmp * tmp; - TensorCheckEqual(A1, A2); - - /* - * T tmp = p1 * b + p2 * c; - * a = tmp * tmp - */ - A1.dotSquareSum(B, C, p1, p2); - auto tmp2 = B * p1 + C * p2; - A2 = tmp2 * tmp2; - TensorCheckEqual(A1, A2); - - // a *= p1 * b + p2 * c - A1.dotMulSum(B, C, p1, p2); - A2 *= B * p1 + C * p2; - TensorCheckEqual(A1, A2); - - // a = p1 * a + p2 * b * c - A1.addDotMul(B, C, p1, p2); - A2 = A2 * p1 + B.constant(p2) * B * C; - TensorCheckEqual(A1, A2); -} - -template -void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.dotDiv(B, C); // a = (b == 0.0) ? 0.0 : b / c - A2 = (B == (real)0.0).condition((real)0.0, B / C); - TensorCheckEqual(A1, A2); - - real p1 = 1.5; - real p2 = 2.5; - A1.dotDiv(B, C, p1, p2); // a = (b + p1) / (c + p2) - A2 = (B + p1) / (C + p2); - TensorCheckEqual(A1, A2); -} - -template -void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - real p1 = 1.5; - real p2 = 2.5; - real p3 = 3.5; - A1.reciprocalSum(B, C, p1, p2, p3); // a = 1 / (p1 * b + p2 * c + p3) - A2 = (B * p1 + C * p2 + p3).reciprocal(); - TensorCheckEqual(A1, A2); -} - -template -void testTensorSoftCrossEntropy(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.softCrossEntropy(B, C); // a = -c * log(b) - (1 - c) * log(1 - b) - A2 = -C * B.log() - (C.constant(1.0f) - C) * (B.constant(1.0f) - B).log(); - TensorCheckErr(A1, A2); -} - -template -void testTensorSoftCrossEntropyBp(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C) { - A1.softCrossEntropyBp(B, C); // a += (b - c) / (b * (1 - b)) - A2 += (B - C) / (B * (B.constant(1.0f) - B)); - TensorCheckEqual(A1, A2); -} - -template -void testTernaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - testTensorAdd(A1, A2, B, C); - testTensorSub(A1, A2, B, C); - testTensorMul(A1, A2, B, C); - testTensorDiv(A1, A2, B, C); - testTensorReciprocal(A1, A2, B, C); - testTensorSoftCrossEntropyBp(A1, A2, B, C); - - testTensorSoftCrossEntropy(A1, A2, B, C); -} - -TEST(Ternary, BaseOp) { - TestTernaryMatrix testCpu(testTernaryBaseOp); - -#ifdef PADDLE_WITH_GPU - TestTernaryMatrix testGpu(testTernaryBaseOp); -#endif -} - -template -void testTensorBinaryLabelCrossEntropy(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C) { - A1.binaryLabelCrossEntropy(B, C); // a = c > 0.5 ? -log(b) : -log(1.0 - b) - A2 = (C > (real)0.5).condition(-(B.log()), -((B.constant(1.0f) - B).log())); - TensorCheckErr(A1, A2); -} - -template -void testTensorBinaryLabelCrossEntropyBp(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C) { - // a += c > 0.5 ? -1.0 / b : 1.0 / (1.0 - b) - A1.binaryLabelCrossEntropyBp(B, C); - A2 += (C > (real)0.5) - .condition((B.constant(-1.0f) / B), - (B.constant(1.0f) - B).reciprocal()); - TensorCheckErr(A1, A2); -} - -template -void testTensorLogisticRegressionLoss(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C) { - SetTensorValue(B, 50.0f); - SetTensorValue(B, -50.0f); - /** - * const T THRESHOLD = 40.0; - * T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) - * ? -THRESHOLD - * : b; - * a = log(1 + exp(x)) - c * x - */ - A1.logisticRegressionLoss(B, C); - real THRESHOLD = 40.0; - auto tmp = - (B > THRESHOLD) - .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); - A2 = (C.constant(1.0f) + tmp.exp()).log() - C * tmp; - TensorCheckErr(A1, A2); -} - -template -void testTensorLogisticRegressionLossBp(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C) { - SetTensorValue(B, 50.0f); - SetTensorValue(B, -50.0f); - /** - * const T THRESHOLD = 40.0; - * T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) - * ? -THRESHOLD - * : b; - * x = exp(x); a = x / (1 + x) - c - */ - A1.logisticRegressionLossBp(B, C); - real THRESHOLD = 40.0; - auto tmp = - (B > THRESHOLD) - .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); - auto tmp2 = tmp.exp(); - A2 = tmp2 / (C.constant(1.0) + tmp2) - C; - TensorCheckErr(A1, A2); -} - -template -void testTensorBiggerThan(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.biggerThan(B, C); // a = (b > c) ? 1.0f : 0.0f - A2 = (B > C).condition((real)1.0f, (real)0.0f); - TensorCheckEqual(A1, A2); -} - -template -void testTensorMax(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - A1.max2(B, C); // a = (b > c) ? b : c - A2 = (B > C).condition(B, C); - TensorCheckEqual(A1, A2); -} - -template -void testTernaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { - testTensorBinaryLabelCrossEntropyBp(A1, A2, B, C); - testTensorBinaryLabelCrossEntropy(A1, A2, B, C); - testTensorBiggerThan(A1, A2, B, C); - testTensorMax(A1, A2, B, C); - - testTensorLogisticRegressionLoss(A1, A2, B, C); - testTensorLogisticRegressionLossBp(A1, A2, B, C); -} - -TEST(Ternary, CompareOp) { - TestTernaryMatrix testCpu(testTernaryCompareOp); - -#ifdef PADDLE_WITH_GPU - TestTernaryMatrix testGpu(testTernaryCompareOp); -#endif -} - -template -void testQuaternaryAdd( - Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { - // A1.add3(B, C, D, 1.5f, 2.5f, 3.5f); // a = p1 * b + p2 * c + p3 * d - // A2 = B * 1.5f + C * 2.5f + D * 3.5f; - // TensorCheckEqual(A1, A2); - - /* - * T tmp = p1 * b + p2 * c + p3 * d; - * a += tmp * tmp - */ - real p1 = 1.5f; - real p2 = 2.5f; - real p3 = 3.5f; - A1.addSquareSum(B, C, D, p1, p2, p3); - auto tmp = B * p1 + C * p2 + D * p3; - A2 += tmp * tmp; - TensorCheckEqual(A1, A2); -} - -TEST(Quaternary, BaseOp) { - TestQuaternaryMatrix testCpu(testQuaternaryAdd); - -#ifdef PADDLE_WITH_GPU - TestQuaternaryMatrix testGpu(testQuaternaryAdd); -#endif -} - -template -void testTensorBiggerThan( - Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { - // a = ((b > c && d > 0.5f) || (b < c && d < 0.5f)) ? 1.0f : 0.0f); - A1.biggerThan(B, C, D); - A2 = ((B > C && D > (real)0.5) || (B < C && D < (real)0.5)) - .condition((real)1.0, (real)0.0); - TensorCheckEqual(A1, A2); -} - -template -void testTensorRankLoss( - Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { - /** - * const T THRESHOLD = 40.0; a = b - c; - * a = (a > THRESHOLD) - * ? THRESHOLD - * : ((a < -THRESHOLD) ? (-THRESHOLD) : a); - * a = log(1 + exp(a)) - a * d - */ - A1.rankLoss(B, C, D); - - real THRESHOLD = 40.0; - auto tmp = B - C; - auto tmp2 = - (tmp > THRESHOLD) - .condition(THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); - A2 = (D.constant(1.0f) + tmp2.exp()).log() - tmp2 * D; - - TensorCheckErr(A1, A2); -} - -template -void testTensorRankLossBp( - Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { - /** - * const T THRESHOLD = 40.0; a = b - c; - * a = (a > THRESHOLD) - * ? THRESHOLD - * : ((a < -THRESHOLD) ? (-THRESHOLD) : a); - * a = exp(a); a = (a / (1 + a) - d) - */ - A1.rankLossBp(B, C, D); - real THRESHOLD = 40.0; - auto tmp = B - C; - auto tmp2 = - (tmp > THRESHOLD) - .condition(THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); - auto tmp3 = tmp2.exp(); - A2 = tmp3 / (D.constant(1.0f) + tmp3) - D; - - TensorCheckErr(A1, A2); -} - -template -void testQuaternaryCompareOp( - Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { - testTensorBiggerThan(A1, A2, B, C, D); - testTensorRankLoss(A1, A2, B, C, D); - testTensorRankLossBp(A1, A2, B, C, D); -} - -TEST(Quaternary, CompareOp) { - TestQuaternaryMatrix testCpu(testQuaternaryCompareOp); - -#ifdef PADDLE_WITH_GPU - TestQuaternaryMatrix testGpu(testQuaternaryCompareOp); -#endif -} diff --git a/paddle/math/tests/test_TrainingAlgorithm.cpp b/paddle/math/tests/test_TrainingAlgorithm.cpp deleted file mode 100644 index fb58d26734cab5d7d7bbbbe1cf8a920e4195b4bb..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_TrainingAlgorithm.cpp +++ /dev/null @@ -1,461 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "OriginalOptimizerApi.h" -#include "PerfUtils.h" -#include "TensorCheck.h" -#include "paddle/math/TrainingAlgorithmOp.h" -#include "paddle/utils/Util.h" - -using namespace paddle; // NOLINT - -#ifndef PADDLE_TYPE_DOUBLE -DEFINE_double(max_diff, 1e-5, "max diff allowed"); -#else -DEFINE_double(max_diff, 1e-13, "max diff allowed"); -#endif - -class SetMaxDiff { - public: - explicit SetMaxDiff(double max_diff) { - max_diff_ = FLAGS_max_diff; - FLAGS_max_diff = max_diff; - } - ~SetMaxDiff() { FLAGS_max_diff = max_diff_; } - - private: - double max_diff_; -}; - -#define COPY_VECTOR_TO_CPU(cpuVec, vector) \ - do { \ - if (vector->useGpu()) { \ - cpuVec = Vector::create(vector->getSize(), false); \ - cpuVec->copyFrom(*vector); \ - } else { \ - cpuVec = vector; \ - } \ - } while (0) - -int VectorCheckErr(const Vector& vector1, const Vector& vector2) { - CHECK(vector1.getSize() == vector2.getSize()); - - const real* data1 = vector1.getData(); - const real* data2 = vector2.getData(); - size_t size = vector1.getSize(); - int count = 0; - for (size_t i = 0; i < size; i++) { - real a = data1[i]; - real b = data2[i]; - if (fabs(a - b) > FLAGS_max_diff) { - if ((fabsf(a - b) / fabsf(a)) > (FLAGS_max_diff / 10.0f)) { - count++; - } - } - } - - return count; -} - -int VectorCheckErr(const VectorPtr& vector1, const VectorPtr& vector2) { - VectorPtr tmp1; - VectorPtr tmp2; - COPY_VECTOR_TO_CPU(tmp1, vector1); - COPY_VECTOR_TO_CPU(tmp2, vector2); - return VectorCheckErr(*tmp1, *tmp2); -} - -#ifdef PADDLE_DISABLE_TIMER - -#define CHECK_VECTORPTR(vector1, vector2) \ - EXPECT_EQ(VectorCheckErr(vector1, vector2), 0) - -#else - -#define CHECK_VECTORPTR(vector1, vector2) - -#endif - -typedef std::function testMatrixFunc; - -void testCase(testMatrixFunc matrixFunc) { -#ifdef PADDLE_WITH_CUDA - for (auto useGpu : {false, true}) { -#else - for (auto useGpu : {false}) { -#endif - for (auto size : {1, - 32, - 64, - 128, - 512, - 1024, - 4096, - 32768, - 65536, - 131072, - 262144, - 524288, - 1048576, - 2097152}) { - LOG(INFO) << " size=" << size << " useGpu=" << useGpu; - matrixFunc(size, useGpu); - } - } -} - -#define INIT_VECTOR(vec1, vec2, type, size, useGpu) \ - vec1[type] = Vector::create(size, useGpu); \ - vec2[type] = Vector::create(size, useGpu); \ - vec1[type]->rand(); \ - vec2[type]->copyFrom(*vec1[type]); - -void testAdagrad(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM1, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); - - real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT - real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT - real momentum = (real)rand() / (real)RAND_MAX; // NOLINT - real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT - - EXPRESSION_PERFORMANCE(AdagradParameterOptimizer( - bufs1, epsilon, learningRate, momentum, decayRate)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; - BaseMatrix& accum_buffer = *bufs2[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& accum = *bufs2[PARAMETER_GRADIENT_SQURESUM1]; - BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; - - EXPRESSION_PERFORMANCE(adagradApply(value, - grad, - mom, - accum_buffer, - accum, - lr, - epsilon, - learningRate, - momentum, - decayRate)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM1], - bufs2[PARAMETER_GRADIENT_SQURESUM1]); - CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], - bufs2[PARAMETER_LEARNING_RATE]); -} - -TEST(Training, Adagrad) { testCase(testAdagrad); } - -void testAdaDelta(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM1, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); - - real rou = (real)rand() / (real)RAND_MAX; // NOLINT - real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT - real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT - real momentum = (real)rand() / (real)RAND_MAX; // NOLINT - real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT - - EXPRESSION_PERFORMANCE(AdaDeltaParameterOptimizer( - bufs1, rou, epsilon, learningRate, momentum, decayRate)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; - BaseMatrix& accum = *bufs2[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& accum_update = *bufs2[PARAMETER_GRADIENT_SQURESUM1]; - BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; - - EXPRESSION_PERFORMANCE(adadeltaApply(value, - grad, - mom, - accum, - accum_update, - lr, - rou, - epsilon, - learningRate, - momentum, - decayRate)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM], - bufs2[PARAMETER_GRADIENT_SQURESUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM1], - bufs2[PARAMETER_GRADIENT_SQURESUM1]); - CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], - bufs2[PARAMETER_LEARNING_RATE]); -} - -TEST(Training, AdaDelta) { testCase(testAdaDelta); } - -template -void testRMSProp(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM1, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); - - /* make sure 'g - f.square()' greater than 0 */ - bufs1[PARAMETER_GRADIENT_SQURESUM]->add(1.0); - bufs2[PARAMETER_GRADIENT_SQURESUM]->copyFrom( - *bufs1[PARAMETER_GRADIENT_SQURESUM]); - - real rou = (real)rand() / (real)RAND_MAX; // NOLINT - real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT - real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT - real momentum = (real)rand() / (real)RAND_MAX; // NOLINT - real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT - real accumulatedRou = rou; - - EXPRESSION_PERFORMANCE(RMSPropParameterOptimizer(bufs1, - accumulatedRou, - rou, - epsilon, - learningRate, - momentum, - decayRate, - isFirstTime)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; - BaseMatrix& sum = *bufs2[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& sum1 = *bufs2[PARAMETER_GRADIENT_SQURESUM1]; - BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; - - EXPRESSION_PERFORMANCE(rmspropApply(value, - grad, - mom, - sum, - sum1, - lr, - accumulatedRou, - rou, - epsilon, - learningRate, - momentum, - decayRate, - isFirstTime)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM], - bufs2[PARAMETER_GRADIENT_SQURESUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM1], - bufs2[PARAMETER_GRADIENT_SQURESUM1]); - CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], - bufs2[PARAMETER_LEARNING_RATE]); -} - -TEST(Training, RMSProp) { - testCase(testRMSProp); - testCase(testRMSProp); -} - -template -void testDecayedAdagrad(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT_SQURESUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_LEARNING_RATE, size, useGpu); - - real rou = (real)rand() / (real)RAND_MAX; // NOLINT - real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT - real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT - real momentum = (real)rand() / (real)RAND_MAX; // NOLINT - real decayRate = (real)rand() / (real)RAND_MAX; // NOLINT - real accumulatedRou = rou; - - if (isFirstTime) { - bufs1[PARAMETER_GRADIENT_SQURESUM]->zeroMem(); - bufs2[PARAMETER_GRADIENT_SQURESUM]->zeroMem(); - } - - EXPRESSION_PERFORMANCE(DecayedAdagradParameterOptimizer(bufs1, - accumulatedRou, - rou, - epsilon, - learningRate, - momentum, - decayRate, - isFirstTime)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; - BaseMatrix& sum = *bufs2[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& lr = *bufs2[PARAMETER_LEARNING_RATE]; - - EXPRESSION_PERFORMANCE(decayedAdagradApply(value, - grad, - mom, - sum, - lr, - accumulatedRou, - rou, - epsilon, - learningRate, - momentum, - decayRate, - isFirstTime)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_GRADIENT_SQURESUM], - bufs2[PARAMETER_GRADIENT_SQURESUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_LEARNING_RATE], - bufs2[PARAMETER_LEARNING_RATE]); -} - -TEST(Training, DecayedAdagrad) { - testCase(testDecayedAdagrad); - testCase(testDecayedAdagrad); -} - -void testAdam(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_SECOND_MOMENTUM, size, useGpu); - - real beta1 = (real)rand() / (real)RAND_MAX; // NOLINT - real beta2 = (real)rand() / (real)RAND_MAX; // NOLINT - real beta1_power = (real)rand() / (real)RAND_MAX; // NOLINT - real beta2_power = (real)rand() / (real)RAND_MAX; // NOLINT - real epsilon = (real)rand() / (real)RAND_MAX; // NOLINT - real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT - - EXPRESSION_PERFORMANCE(AdamParameterOptimizer( - bufs1, beta1, beta2, beta1_power, beta2_power, epsilon, learningRate)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; - BaseMatrix& v = *bufs2[PARAMETER_SECOND_MOMENTUM]; - - EXPRESSION_PERFORMANCE(adamApply(value, - grad, - mom, - v, - beta1, - beta2, - beta1_power, - beta2_power, - epsilon, - learningRate)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_SECOND_MOMENTUM], - bufs2[PARAMETER_SECOND_MOMENTUM]); -} - -TEST(Training, Adam) { testCase(testAdam); } - -void testAdamax(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_WEIGHTED_INFINITY_NORM, size, useGpu); - - real beta1 = (real)rand() / (real)RAND_MAX; // NOLINT - real beta2 = (real)rand() / (real)RAND_MAX; // NOLINT - real alpha = (real)rand() / (real)RAND_MAX; // NOLINT - int64_t step = 2; - - EXPRESSION_PERFORMANCE( - AdamaxParameterOptimizer(bufs1, beta1, beta2, step, alpha)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& mom = *bufs2[PARAMETER_MOMENTUM]; - BaseMatrix& u = *bufs2[PARAMETER_WEIGHTED_INFINITY_NORM]; - - EXPRESSION_PERFORMANCE( - adamaxApply(value, grad, mom, u, beta1, beta2, step, alpha)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM], bufs2[PARAMETER_MOMENTUM]); - CHECK_VECTORPTR(bufs1[PARAMETER_WEIGHTED_INFINITY_NORM], - bufs2[PARAMETER_WEIGHTED_INFINITY_NORM]); -} - -TEST(Training, Adamax) { -#ifndef PADDLE_TYPE_DOUBLE - SetMaxDiff diff(1e-4); -#endif - testCase(testAdamax); -} - -void testSparseMomentum(size_t size, bool useGpu) { - VectorPtr bufs1[NUM_PARAMETER_TYPES]; - VectorPtr bufs2[NUM_PARAMETER_TYPES]; - INIT_VECTOR(bufs1, bufs2, PARAMETER_VALUE, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_GRADIENT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM_UT, size, useGpu); - INIT_VECTOR(bufs1, bufs2, PARAMETER_MOMENTUM_VT, size, useGpu); - - real alpha = (real)rand() / (real)RAND_MAX; // NOLINT - real beta = (real)rand() / (real)RAND_MAX; // NOLINT - real gamma = (real)rand() / (real)RAND_MAX; // NOLINT - real tau = (real)rand() / (real)RAND_MAX; // NOLINT - real learningRate = (real)rand() / (real)RAND_MAX; // NOLINT - - EXPRESSION_PERFORMANCE(SparseMomentumParameterOptimizer( - bufs1, alpha, beta, gamma, tau, learningRate)); - - BaseMatrix& value = *bufs2[PARAMETER_VALUE]; - BaseMatrix& grad = *bufs2[PARAMETER_GRADIENT]; - BaseMatrix& momU = *bufs2[PARAMETER_MOMENTUM_UT]; - BaseMatrix& momV = *bufs2[PARAMETER_MOMENTUM_VT]; - - EXPRESSION_PERFORMANCE(sparseMomentumApply( - value, grad, momU, momV, alpha, beta, gamma, tau, learningRate)); - - CHECK_VECTORPTR(bufs1[PARAMETER_VALUE], bufs2[PARAMETER_VALUE]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM_UT], bufs2[PARAMETER_MOMENTUM_UT]); - CHECK_VECTORPTR(bufs1[PARAMETER_MOMENTUM_VT], bufs2[PARAMETER_MOMENTUM_VT]); -} - -TEST(Training, SparseMomentum) { testCase(testSparseMomentum); } diff --git a/paddle/math/tests/test_lazyAssign.cu b/paddle/math/tests/test_lazyAssign.cu deleted file mode 100644 index cbd74bbfe33270f351632b58d7e89f8e60d15b83..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_lazyAssign.cu +++ /dev/null @@ -1,147 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "PerfUtils.h" -#include "TensorCheck.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/TensorAssign.h" - -using paddle::BaseMatrix; -using paddle::CpuMatrix; -using paddle::GpuMatrix; -using autotest::TensorCheckEqual; -using autotest::TensorCheckErr; - -typedef std::function testMatrixFunc; -void testMatrixCase(testMatrixFunc matrixFunc) { - for (auto height : {1}) { - for (auto width : {1, - 32, - 64, - 128, - 512, - 1024, - 4096, - 32768, - 65536, - 131072, - 262144, - 524288, - 1048576, - 2097152, - 4194304, - 8388608}) { - matrixFunc(height, width); - } - } -} - -template -void testLazyAssign(int height, int width) { - Tensor A1(height, width); - Tensor A2(height, width); - Tensor B(height, width); - Tensor C(height, width); - Tensor D(height, width); - A1.randomizeUniform(); - B.randomizeUniform(); - C.randomizeUniform(); - D.randomizeUniform(); - A2.copyFrom(A1); - - EXPRESSION_PERFORMANCE(A1 = B + C; A1 = A1 * D;); - - EXPRESSION_PERFORMANCE(auto expr1 = A2.lazyAssign(B + C); - auto expr2 = A2.lazyAssign(A2 * D); - AssignEvaluate(expr1, expr2);); - - TensorCheckErr(A1, A2); -} - -TEST(lazyAssign, CPU) { testMatrixCase(testLazyAssign); } - -#ifdef PADDLE_WITH_GPU -TEST(lazyAssign, GPU) { testMatrixCase(testLazyAssign); } -#endif - -template -void sgdUpdateTensor( - Tensor& A, Tensor& B, Tensor& C, Tensor& D, real p1, real p2, real p3) { - C = C * p2 - D * (B + A * p3) * p1; - A += C; -} - -void sgdUpdateLazyAssign(BaseMatrix& A, - BaseMatrix& B, - BaseMatrix& C, - BaseMatrix& D, - real p1, - real p2, - real p3) { - auto expr1 = C.lazyAssign(C * p2 - D * (B + A * p3) * p1); - auto expr2 = A.lazyAssign(A + C); - AssignEvaluate(expr1, expr2); -} - -template -void testSgdUpdate(int height, int width) { - Tensor A1(height, width); - Tensor A2(height, width); - Tensor A3(height, width); - A1.randomizeUniform(); - A2.copyFrom(A1); - A3.copyFrom(A1); - - Tensor B(height, width); - B.randomizeUniform(); - - Tensor C1(height, width); - Tensor C2(height, width); - Tensor C3(height, width); - C1.randomizeUniform(); - C2.copyFrom(C1); - C3.copyFrom(C1); - - Tensor D(height, width); - D.randomizeUniform(); - - real p1 = 0.2; - real p2 = 0.3; - real p3 = 0.5; - - /** - * c = p2 * c - p1 * (b + p3 * a); - * a = a + c; - */ - // BaseMatrix API - EXPRESSION_PERFORMANCE(A1.sgdUpdate(B, C1, D, p1, p2, p3);); - - // Tensor expression - EXPRESSION_PERFORMANCE(sgdUpdateTensor(A2, B, C2, D, p1, p2, p3)); - - // lazyAssign - EXPRESSION_PERFORMANCE(sgdUpdateLazyAssign(A3, B, C3, D, p1, p2, p3)); - - TensorCheckErr(A1, A2); - TensorCheckErr(A1, A3); - TensorCheckErr(C1, C2); - TensorCheckErr(C1, C3); -} - -TEST(sgdUpdate, CPU) { testMatrixCase(testSgdUpdate); } - -#ifdef PADDLE_WITH_GPU -TEST(sgdUpdate, GPU) { testMatrixCase(testSgdUpdate); } -#endif diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp deleted file mode 100644 index e45ddd433faf18dbcd647b305db3a36d38c90825..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_matrixCompare.cpp +++ /dev/null @@ -1,1698 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_CUDA -/// This unittest checks GpuMatrix/CpuMatrix get same result, so disable when -/// only cpu version. - -#include -#include "TensorCheck.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseMatrix.h" -#include "paddle/testing/TestUtil.h" -#include "paddle/utils/DynamicLoader.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT -using autotest::TensorCheckEqual; -using autotest::TensorCheckErr; - -void testMatrixMaxSequence(int batchSize, int inputDim) { - // forward - MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); - MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - - IVectorPtr cpuSequence; - generateSequenceStartPositions(batchSize, cpuSequence); - IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); - gpuSequence->copyFrom(*cpuSequence); - - int newBatchSize = cpuSequence->getSize() - 1; - MatrixPtr cpuOutput = std::make_shared(newBatchSize, inputDim); - MatrixPtr gpuOutput = std::make_shared(newBatchSize, inputDim); - cpuOutput->zero(); - gpuOutput->zero(); - - IVectorPtr cpuIndex = nullptr; - IVectorPtr gpuIndex = nullptr; - IVector::resizeOrCreate(cpuIndex, newBatchSize * inputDim, false); - IVector::resizeOrCreate(gpuIndex, newBatchSize * inputDim, true); - cpuIndex->zeroMem(); - gpuIndex->zeroMem(); - - cpuOutput->maxSequenceForward(*cpuInput, *cpuSequence, *cpuIndex); - gpuOutput->maxSequenceForward(*gpuInput, *gpuSequence, *gpuIndex); - - TensorCheckEqual(*cpuOutput, *gpuOutput); - TensorCheckEqual(*cpuIndex, *gpuIndex); - - // backward - MatrixPtr cpuOutputGrad = std::make_shared(newBatchSize, inputDim); - MatrixPtr gpuOutputGrad = std::make_shared(newBatchSize, inputDim); - cpuOutputGrad->randomizeUniform(); - gpuOutputGrad->copyFrom(*cpuOutputGrad); - - MatrixPtr cpuInputGrad = std::make_shared(batchSize, inputDim); - MatrixPtr gpuInputGrad = std::make_shared(batchSize, inputDim); - cpuInputGrad->randomizeUniform(); - gpuInputGrad->copyFrom(*cpuInputGrad); - - cpuInputGrad->maxSequenceBackward(*cpuOutputGrad, *cpuSequence, *cpuIndex); - gpuInputGrad->maxSequenceBackward(*gpuOutputGrad, *gpuSequence, *gpuIndex); - - TensorCheckEqual(*cpuInputGrad, *gpuInputGrad); -} - -TEST(Matrix, maxSequence) { - for (auto batchSize : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 - for (auto inputDim : {1, 7, 131}) { // prime numbers close to 1, 8, 128 - VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim; - testMatrixMaxSequence(batchSize, inputDim); - } - } -} - -void testMatrixGetSum(int height, int width) { - MatrixPtr cpuInput = std::make_shared(height, width); - MatrixPtr gpuInput = std::make_shared(height, width); - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - -#ifndef PADDLE_TYPE_DOUBLE - int x = log10(height * width); - real err = 1e-6 * pow(10, x); -#else - real err = 1e-8; -#endif - - real cpuSum = cpuInput->getSum(); - real gpuSum = gpuInput->getSum(); - - EXPECT_LE(fabs(cpuSum - gpuSum), err); -} - -void testMatrixGetMinMax(int height, int width) { - MatrixPtr cpuInput = std::make_shared(height, width); - MatrixPtr gpuInput = std::make_shared(height, width); - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - - real cpuMin = cpuInput->getMin(); - real gpuMin = gpuInput->getMin(); - real cpuMax = cpuInput->getMax(); - real gpuMax = gpuInput->getMax(); - - EXPECT_EQ(cpuMin, gpuMin); - EXPECT_EQ(cpuMax, gpuMax); -} - -void testMatrixZeroAtOffset(int height, int width) { - MatrixPtr cpuA = std::make_shared(height, width); - MatrixPtr gpuA = std::make_shared(height, width); - MatrixPtr cpuTest = std::make_shared(height, width); - - cpuA->randomizeUniform(); - gpuA->copyFrom(*cpuA); - cpuTest->copyFrom(*cpuA); - - int columnOffset = rand() % width; // NOLINT we just use rand() for test. - int numColumns = rand() % (width - columnOffset); // NOLINT - - if (numColumns == 0) return; - - cpuA->zeroAtOffset(columnOffset, numColumns); - gpuA->zeroAtOffset(columnOffset, numColumns); - - /* cpuTest */ - real* a = cpuTest->getData() + columnOffset; - for (int64_t i = 0; i < height; ++i) { - for (int64_t j = 0; j < numColumns; ++j) { - a[i * width + j] = 0; - } - } - - TensorCheckEqual(*cpuA, *gpuA); - TensorCheckEqual(*cpuA, *cpuTest); -} - -void testMatrixDeepSwap(int height, int width) { - MatrixPtr cpuA = std::make_shared(height, width); - MatrixPtr cpuB = std::make_shared(height, width); - MatrixPtr cpuCopyA = std::make_shared(height, width); - MatrixPtr cpuCopyB = std::make_shared(height, width); - - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - cpuCopyA->copyFrom(*cpuA); - cpuCopyB->copyFrom(*cpuB); - - // swap matrix cpuA and cpuB - cpuA->deepSwap(*cpuB); - - TensorCheckEqual(*cpuA, *cpuCopyB); - TensorCheckEqual(*cpuB, *cpuCopyA); -} - -void testMatrixTranspose(int height, int width) { - MatrixPtr cpu = std::make_shared(height, width); - MatrixPtr gpu = std::make_shared(height, width); - MatrixPtr cpuT = std::make_shared(width, height); - MatrixPtr gpuT = std::make_shared(width, height); - - cpu->randomizeUniform(); - gpu->copyFrom(*cpu); - cpu->transpose(cpuT, false); - gpu->transpose(gpuT, true); - - TensorCheckEqual(*cpuT, *gpuT); -} - -void testMatrixRotate(int height, int width) { - MatrixPtr cpu = std::make_shared(height, width); - MatrixPtr gpu = std::make_shared(height, width); - MatrixPtr cpuR = std::make_shared(width, height); - MatrixPtr gpuR = std::make_shared(width, height); - - cpu->randomizeUniform(); - gpu->copyFrom(*cpu); - - cpu->rotate(cpuR, false, true); - gpu->rotate(gpuR, true, true); - TensorCheckEqual(*cpuR, *gpuR); - - cpu->rotate(cpuR, true, false); - gpu->rotate(gpuR, false, false); - TensorCheckEqual(*cpuR, *gpuR); -} - -void testMatrixInverse(int height) { - MatrixPtr cpu = std::make_shared(height, height); - MatrixPtr gpu = std::make_shared(height, height); - MatrixPtr cpuI = std::make_shared(height, height); - MatrixPtr gpuI = std::make_shared(height, height); - - /* Make matrix well conditioned: cpu * cpuT + Identity */ - cpu->randomizeUniform(); - MatrixPtr cpuT = cpu->getTranspose(); - MatrixPtr outputCheck = std::make_shared(height, height); - outputCheck->mul(*cpu, *cpuT); - cpu->setDiag(1.0); - cpu->add(*outputCheck); - - gpu->copyFrom(*cpu); - cpu->inverse(cpuI, true); - gpu->inverse(gpuI, false); - - TensorCheckErr(*cpuI, *gpuI); - - outputCheck->mul(*cpu, *cpuI); - cpu->setDiag(1.0); - TensorCheckErr(*cpu, *outputCheck); -} - -TEST(Matrix, unary) { - for (auto height : {1, 3, 11, 73, 128, 200, 330}) { - for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) { - VLOG(3) << " height=" << height << " width=" << width; - - testMatrixDeepSwap(height, width); - testMatrixZeroAtOffset(height, width); - testMatrixGetSum(height, width); - testMatrixTranspose(height, width); - testMatrixRotate(height, width); - } -#ifdef LAPACK_FOUND - // inverse matrix - testMatrixInverse(height); -#else - LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK" - << "support so we cannot test matrix inverse. To test " - << "matrix inverse, please install LAPACKE " - << "and MKL/Openblas, and re-build PaddlePaddle."; -#endif - } -} - -void testMatrixSoftmax(int height, int width) { - MatrixPtr cpuInput = std::make_shared(height, width); - MatrixPtr cpuOutput = std::make_shared(height, width); - MatrixPtr gpuInput = std::make_shared(height, width); - MatrixPtr gpuOutput = std::make_shared(height, width); - - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - cpuOutput->zero(); - gpuOutput->zero(); - cpuInput->softmax(*cpuOutput); - gpuInput->softmax(*gpuOutput); - - TensorCheckErr(*cpuOutput, *gpuOutput); -} - -void testSequenceSoftmax(int batchSize) { - // forward - int inputDim = 1; - MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); - MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - - IVectorPtr cpuSequence; - generateSequenceStartPositions(batchSize, cpuSequence); - IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); - gpuSequence->copyFrom(*cpuSequence); - - cpuInput->sequenceSoftmax(*cpuInput, *cpuSequence); - gpuInput->sequenceSoftmax(*gpuInput, *gpuSequence); - - TensorCheckErr(*cpuInput, *gpuInput); -} - -void testMatrixSoftmaxThreshold(int height, int width) { - MatrixPtr cpuInput = std::make_shared(height, width); - MatrixPtr cpuOutput = std::make_shared(height, width); - MatrixPtr gpuInput = std::make_shared(height, width); - MatrixPtr gpuOutput = std::make_shared(height, width); - - cpuInput->randomizeUniform(); - cpuInput->getData()[0] = 100.0; - gpuInput->copyFrom(*cpuInput); - cpuOutput->zero(); - gpuOutput->zero(); - cpuInput->softmax(*cpuOutput); - gpuInput->softmax(*gpuOutput); - - MatrixPtr outputCheck = std::make_shared(height, width); - outputCheck->copyFrom(*gpuOutput); - // check output zero - int cpuCount = 0; - int gpuCount = 0; - auto zeroNum = [](MatrixPtr out, int& count) { - for (size_t i = 0; i < out->getHeight(); i++) { - for (size_t j = 0; j < out->getWidth(); j++) { - if (out->getElement(i, j) == 0) count++; - } - } - }; - zeroNum(cpuOutput, cpuCount); - zeroNum(outputCheck, gpuCount); - EXPECT_EQ(cpuCount, 0) << "Cpu softmax output value 0"; - EXPECT_EQ(gpuCount, 0) << "Gpu softmax output value 0"; -} - -void testMatrixSoftmaxBp(int height, int width) { - MatrixPtr cpuInput = std::make_shared(height, width); - MatrixPtr cpuOutput = std::make_shared(height, width); - MatrixPtr gpuInput = std::make_shared(height, width); - MatrixPtr gpuOutput = std::make_shared(height, width); - - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - cpuOutput->randomizeUniform(); - gpuOutput->copyFrom(*cpuOutput); - gpuOutput->softmaxBackward(*gpuInput); - - MatrixPtr sftMaxSum = std::make_shared(height, 1); - MatrixPtr sftMaxDot = std::make_shared(height, width); - sftMaxDot->dotMul(*cpuOutput, *cpuInput); - sftMaxSum->colMerge(*sftMaxDot); - cpuOutput->softmaxDerivative(*cpuInput, *sftMaxSum); - - TensorCheckErr(*cpuOutput, *gpuOutput); -} - -TEST(Matrix, softmax) { - for (auto height : {1, 3, 131}) { // prime numbers close to 1, 4, 127 - for (auto width : {1, 17, 251}) { // prime numbers close to 1, 16, 256 - VLOG(3) << " height=" << height << " width=" << width; - - testMatrixSoftmax(height, width); - testMatrixSoftmaxBp(height, width); - testMatrixSoftmaxThreshold(height, width); - } - testSequenceSoftmax(height); - } -} - -void testMatrixAddToRows(int numSamples, int tableSize, int inputDim) { - MatrixPtr cpuTable = std::make_shared(tableSize, inputDim); - MatrixPtr gpuTable = std::make_shared(tableSize, inputDim); - cpuTable->randomizeUniform(); - gpuTable->copyFrom(*cpuTable); - - IVectorPtr cpuIds; - IVectorPtr gpuIds; - cpuIds = VectorT::create(numSamples, false); - gpuIds = VectorT::create(numSamples, true); - cpuIds->rand(tableSize); - gpuIds->copyFrom(*cpuIds); - - MatrixPtr cpuOutput = std::make_shared(numSamples, inputDim); - MatrixPtr gpuOutput = std::make_shared(numSamples, inputDim); - cpuOutput->randomizeUniform(); - gpuOutput->copyFrom(*cpuOutput); - - cpuOutput->addToRows(*cpuTable, *cpuIds); - gpuOutput->addToRows(*gpuTable, *gpuIds); - - TensorCheckErr(*cpuTable, *gpuTable); -} - -TEST(Matrix, tableProjection) { - for (auto numSamples : {10, 100, 1000, 10000, 80000}) { - for (auto tableSize : {10, 100}) { - for (auto inputDim : {20, 50}) { - VLOG(3) << " numSamples=" << numSamples << " tableSize=" << tableSize - << " inputDim=" << inputDim; - testMatrixAddToRows(numSamples, tableSize, inputDim); - } - } - } -} - -void testMatrixMul(bool transa, bool transb, int dimM, int dimN, int dimK) { - int heightA = transa == false ? dimM : dimK; - int widthA = transa == false ? dimK : dimM; - int heightB = transb == false ? dimK : dimN; - int widthB = transb == false ? dimN : dimK; - int heightC = dimM; - int widthC = dimN; - - MatrixPtr cpuA = std::make_shared(heightA, widthA, transa); - MatrixPtr cpuB = std::make_shared(heightB, widthB, transb); - MatrixPtr cpuC = std::make_shared(heightC, widthC); - MatrixPtr gpuA = std::make_shared(heightA, widthA, transa); - MatrixPtr gpuB = std::make_shared(heightB, widthB, transb); - MatrixPtr gpuC = std::make_shared(heightC, widthC); - - real alpha = 1.5; - real beta = 2.0; - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - cpuC->randomizeUniform(); - gpuA->copyFrom(*cpuA); - gpuB->copyFrom(*cpuB); - gpuC->copyFrom(*cpuC); - - cpuC->mul(*cpuA, *cpuB, alpha, beta); - gpuC->mul(*gpuA, *gpuB, alpha, beta); - - TensorCheckErr(*cpuC, *gpuC); -} - -void testSubMatrixMul(bool transa, bool transb, int dimM, int dimN, int dimK) { - int heightA = transa == false ? dimM : dimK; - int widthA = transa == false ? dimK : dimM; - int heightB = transb == false ? dimK : dimN; - int widthB = transb == false ? dimN : dimK; - int heightC = dimM; - int widthC = dimN; - - MatrixPtr cpuA = std::make_shared(heightA, widthA, transa); - MatrixPtr cpuB = std::make_shared(heightB, widthB, transb); - MatrixPtr cpuC = std::make_shared(heightC, widthC); - MatrixPtr gpuA = std::make_shared(heightA, widthA, transa); - MatrixPtr gpuB = std::make_shared(heightB, widthB, transb); - MatrixPtr gpuC = std::make_shared(heightC, widthC); - - real alpha = 1.5; - real beta = 2.0; - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - cpuC->randomizeUniform(); - gpuA->copyFrom(*cpuA); - gpuB->copyFrom(*cpuB); - gpuC->copyFrom(*cpuC); - - auto subSize = [](int& start, int& end, int dim) { - if (dim == 1) { - start = 0; - end = dim; - } else { - int subDim = rand() % (dim - 1) + 1; // NOLINT - start = rand() % (dim - subDim); // NOLINT - end = start + subDim; - } - }; - - auto subMatrix = [](MatrixPtr& sub, - MatrixPtr matrix, - size_t startRow, - size_t endRow, - size_t startCol, - size_t endCol) { - if (!matrix->isTransposed()) { - sub = matrix->subMatrix(startRow, endRow, startCol, endCol); - } else { - sub = matrix->subMatrix(startCol, endCol, startRow, endRow); - } - }; - - int startM, endM; - int startN, endN; - int startK, endK; - subSize(startM, endM, dimM); - subSize(startN, endN, dimN); - subSize(startK, endK, dimK); - - MatrixPtr subCpuA; - MatrixPtr subCpuB; - MatrixPtr subGpuA; - MatrixPtr subGpuB; - subMatrix(subCpuA, cpuA, startM, endM, startK, endK); - subMatrix(subGpuA, gpuA, startM, endM, startK, endK); - subMatrix(subCpuB, cpuB, startK, endK, startN, endN); - subMatrix(subGpuB, gpuB, startK, endK, startN, endN); - MatrixPtr subCpuC = cpuC->subMatrix(startM, endM, startN, endN); - MatrixPtr subGpuC = gpuC->subMatrix(startM, endM, startN, endN); - - subCpuC->mul(*subCpuA, *subCpuB, alpha, beta); - subGpuC->mul(*subGpuA, *subGpuB, alpha, beta); - - TensorCheckErr(*cpuC, *gpuC); -} - -TEST(Matrix, mul) { - for (auto transa : {false, true}) { - for (auto transb : {false, true}) { - for (auto dimM : {1, 9, 53, 127, 345, 1023, 2135}) { - for (auto dimN : {1, 5, 37, 256, 1024}) { - for (auto dimK : {8, 45, 346, 784, 1025}) { - if (true == transa && true == transb) { - continue; - } - VLOG(3) << setiosflags(ios::left) << setfill(' ') - << " transa=" << transa << " transb=" << transb - << " dimM=" << setw(5) << dimM << " dimN=" << setw(5) - << dimN << " dimK=" << setw(5) << dimK; - - testMatrixMul(transa, transb, dimM, dimN, dimK); - testSubMatrixMul(transa, transb, dimM, dimN, dimK); - } - } - } - } - } -} - -void testVectorRowFunc(int size) { - CpuVectorPtr cpu = std::make_shared>(size); - GpuVectorPtr gpu = std::make_shared>(size); - - cpu->rand(); - gpu->copyFrom(*cpu); - - EXPECT_EQ(cpu->getMax(), gpu->getMax()); - EXPECT_EQ(cpu->getMin(), gpu->getMin()); - EXPECT_EQ(cpu->getAbsMax(), gpu->getAbsMax()); -} - -TEST(Vector, rowFunc) { - for (auto size : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 - VLOG(3) << " size=" << size; - testVectorRowFunc(size); - } -} - -template -void testVectorReset(int size) { - std::shared_ptr> cpu = std::make_shared>(size); - std::shared_ptr> gpu = std::make_shared>(size); - - T value = (T)((int)rand() % 100 + 1.0f / ((int)rand() % 100)); - cpu->reset(value); - gpu->reset(value); - - TensorCheckEqual(*cpu, *gpu); -} - -template -void testVecortSelectFrom(int size) { - std::shared_ptr> cpuDst = std::make_shared>(size); - std::shared_ptr> gpuDst = std::make_shared>(size); - std::shared_ptr> cpuSrc = - std::make_shared>(size * 2); - std::shared_ptr> gpuSrc = - std::make_shared>(size * 2); - CpuIVectorPtr cpuIds = std::make_shared>(size); - GpuIVectorPtr gpuIds = std::make_shared>(size); - - if (std::is_same::value) { - cpuSrc->rand(); - } else { - cpuSrc->rand(100000); - } - gpuSrc->copyFrom(*cpuSrc); - cpuIds->rand(size); - gpuIds->copyFrom(*cpuIds); - - cpuDst->selectFrom(*cpuSrc, *cpuIds); - gpuDst->selectFrom(*gpuSrc, *gpuIds); - - TensorCheckEqual(*cpuDst, *gpuDst); -} - -template -void testVecotrZeroMem(int size) { - std::shared_ptr> cpu = std::make_shared>(size); - std::shared_ptr> gpu = std::make_shared>(size); - - cpu->zeroMem(); - gpu->zeroMem(); - - TensorCheckEqual(*cpu, *gpu); -} - -template -void testVectorIsEqual(int size) { - std::shared_ptr> cpuA = std::make_shared>(size); - std::shared_ptr> cpuB = std::make_shared>(size); - std::shared_ptr> gpuA = std::make_shared>(size); - std::shared_ptr> gpuB = std::make_shared>(size); - - if (std::is_same::value) { - cpuB->rand(); - } else { - cpuB->rand(100000); - } - gpuB->copyFrom(*cpuB); - - T value = (T)((int)rand() % 100 + 1.0f / ((int)rand() % 100)); - cpuA->isEqualTo(*cpuB, value); - gpuA->isEqualTo(*gpuB, value); - - TensorCheckEqual(*cpuA, *gpuA); -} - -TEST(Vector, Equal) { - for (auto size : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 - VLOG(3) << " size=" << size; - testVectorReset(size); - testVectorReset(size); - testVecortSelectFrom(size); - testVecortSelectFrom(size); - testVecotrZeroMem(size); - testVecotrZeroMem(size); - testVectorIsEqual(size); - testVectorIsEqual(size); - } -} - -void testMatrixTopK(int samples, int dim, int beamSize) { - MatrixPtr cpuSrc = std::make_shared(samples, dim); - MatrixPtr gpuSrc = std::make_shared(samples, dim); - MatrixPtr cpuVal = std::make_shared(samples, beamSize); - MatrixPtr gpuVal = std::make_shared(samples, beamSize); - IVectorPtr cpuIds = std::make_shared(samples * beamSize); - IVectorPtr gpuIds = std::make_shared(samples * beamSize); - - cpuSrc->randomizeUniform(); - gpuSrc->copyFrom(*cpuSrc); - - cpuSrc->rowMax(*cpuIds, *cpuVal); - gpuSrc->rowMax(*gpuIds, *gpuVal); - - TensorCheckEqual(*cpuVal, *gpuVal); -} - -TEST(Matrix, topK) { - for (auto samples : {1, 17, 131}) { // prime numbers close to 1, 16, 127 - for (auto dim : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 - for (auto beamSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) { - if (beamSize > dim) continue; - VLOG(3) << " samples=" << samples << " beamSize=" << beamSize - << " dim=" << dim; - testMatrixTopK(samples, dim, beamSize); - } - } - } -} - -void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) { - int nnz = samples * dim * ratio; - if (nnz < 1) nnz = 1; // Because sparseRand in MathUtil.cpp requires this. - MatrixPtr cpuSrc = std::make_shared(samples, dim, nnz); - MatrixPtr gpuSrc = std::make_shared(samples, dim, nnz); - MatrixPtr cpuVal = std::make_shared(samples, beamSize); - MatrixPtr gpuVal = std::make_shared(samples, beamSize); - IVectorPtr cpuIds = std::make_shared(samples * beamSize); - IVectorPtr gpuIds = std::make_shared(samples * beamSize); - - cpuSrc->randomizeUniform(); - gpuSrc->copyFrom(*cpuSrc); - cpuVal->zero(); - cpuIds->zero(); - gpuVal->zero(); - gpuIds->zero(); - - cpuSrc->rowMax(*cpuIds, *cpuVal); - gpuSrc->rowMax(*gpuIds, *gpuVal); - - TensorCheckEqual(*cpuVal, *gpuVal); - - IVectorPtr outCheckIds = std::make_shared(samples * beamSize); - outCheckIds->copyFrom(*gpuIds); - - const int* data1 = cpuIds->getData(); - const int* data2 = outCheckIds->getData(); - size_t size = cpuIds->getSize(); - for (size_t i = 0; i < size; i++) { - if (data1[i] == -1 && data1[i] != data2[i]) { - EXPECT_EQ(data1[i], data2[i]); - } - } -} - -TEST(SMatrix, topK) { - for (auto samples : {1, 3, 61}) { - for (auto dim : {1, 3, 61}) { - for (auto beamSize : {1, 3, 61}) { - for (auto ratio : {0.01, 0.001}) { - if (beamSize > dim) continue; - VLOG(3) << " samples=" << samples << " beamSize=" << beamSize - << " dim=" << dim << " ratio=" << ratio; - testSMatrixTopK(samples, dim, beamSize, ratio); - } - } - } - } -} - -void testMatrixSequenceAvg(int batchSize, int inputDim, int mode) { - MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); - MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - - IVectorPtr cpuSequence; - generateSequenceStartPositions(batchSize, cpuSequence); - IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); - gpuSequence->copyFrom(*cpuSequence); - - int newBatchSize = cpuSequence->getSize() - 1; - MatrixPtr cpuOutput = std::make_shared(newBatchSize, inputDim); - MatrixPtr gpuOutput = std::make_shared(newBatchSize, inputDim); - cpuOutput->zero(); - gpuOutput->zero(); - - cpuOutput->sequenceAvgForward(*cpuInput, *cpuSequence, mode); - gpuOutput->sequenceAvgForward(*gpuInput, *gpuSequence, mode); - - TensorCheckErr(*cpuOutput, *gpuOutput); - - MatrixPtr cpuInGrad = std::make_shared(batchSize, inputDim); - MatrixPtr gpuInGrad = std::make_shared(batchSize, inputDim); - cpuInGrad->randomizeUniform(); - gpuInGrad->copyFrom(*cpuInGrad); - - cpuInGrad->sequenceAvgBackward(*cpuOutput, *cpuSequence, mode); - gpuInGrad->sequenceAvgBackward(*gpuOutput, *gpuSequence, mode); - - TensorCheckErr(*cpuInGrad, *gpuInGrad); -} - -TEST(Matrix, sequenceAvg) { - for (auto batchSize : {10, 128, 6000}) { - for (auto inputDim : {32, 100, 512}) { - for (auto mode : {0, 1, 2}) { - VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim - << " mode=" << mode; - testMatrixSequenceAvg(batchSize, inputDim, mode); - } - } - } -} - -void testParamReluBackwardDiff(int height, - int width, - int w_height, - int w_width) { - MatrixPtr oGrad = CpuMatrix::create(height, width, false, false); - MatrixPtr input = CpuMatrix::create(height, width, false, false); - MatrixPtr diff = CpuMatrix::create(height, width, false, false); - MatrixPtr w = CpuMatrix::create(w_height, w_width, false, false); - - oGrad->randomizeUniform(); - input->randomizeUniform(); - w->randomizeUniform(); - diff->randomizeUniform(); - input->add(-0.5); - - MatrixPtr oGradGpu = GpuMatrix::create(height, width, false, true); - MatrixPtr inputGpu = GpuMatrix::create(height, width, false, true); - MatrixPtr diffGpu = CpuMatrix::create(height, width, false, true); - MatrixPtr wGpu = GpuMatrix::create(w_height, w_width, false, true); - - oGradGpu->copyFrom(*oGrad); - inputGpu->copyFrom(*input); - wGpu->copyFrom(*w); - diffGpu->copyFrom(*diff); - - diff->paramReluBackwardDiff(*oGrad, *input, *w); - diffGpu->paramReluBackwardDiff(*oGradGpu, *inputGpu, *wGpu); - - TensorCheckErr(*diff, *diffGpu); -} - -TEST(Matrix, paramReluBackwardDiff) { - for (auto height : {10, 40, 100}) { - for (auto width : {10, 40, 100}) { - for (auto w_height : {1, 2}) { - for (auto w_width : {1, 2}) { - if (width % (w_height * w_width)) continue; - testParamReluBackwardDiff(height, width, w_height, w_width); - } - } - } - } -} - -void testClassificationError(int numSamples, int dim, int topkSize) { - MatrixPtr cpuError = std::make_shared(numSamples, 1); - MatrixPtr gpuError = std::make_shared(numSamples, 1); - MatrixPtr cpuOutput = std::make_shared(numSamples, dim); - MatrixPtr gpuOutput = std::make_shared(numSamples, dim); - IVectorPtr cpuLabel = std::make_shared(numSamples); - IVectorPtr gpuLabel = std::make_shared(numSamples); - - cpuOutput->randomizeUniform(); - cpuLabel->rand(dim); - gpuOutput->copyFrom(*cpuOutput); - gpuLabel->copyFrom(*cpuLabel); - - cpuError->classificationError(*cpuOutput, *cpuLabel, topkSize); - gpuError->classificationError(*gpuOutput, *gpuLabel, topkSize); - - TensorCheckEqual(*cpuError, *gpuError); -} - -TEST(Matrix, classificationError) { - for (auto numSamples : {1, 3, 31}) { - for (auto dim : {1, 3, 31}) { - for (auto topkSize : {1, 3, (int)rand() % dim + 1}) { - if (topkSize > dim) continue; - VLOG(3) << " sample= " << numSamples << " topkSize= " << topkSize - << " dim= " << dim; - testClassificationError(numSamples, dim, topkSize); - } - } - } -} - -void testMaxPoolFwdBwd(int numSamples, - int channels, - int imgSizeH, - int imgSizeW, - int ksizeH, - int ksizeW, - int strideH, - int strideW, - int padH, - int padW) { - int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); - int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); - - int inWidth = imgSizeH * imgSizeW * channels; - MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); - - int outWidth = channels * outH * outW; - MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); - - input->randomizeUniform(); - target->randomizeUniform(); - inputGpu->copyFrom(*input); - targetGpu->copyFrom(*target); - - target->maxPoolForward(*input, - imgSizeH, - imgSizeW, - channels, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - padH, - padW); - targetGpu->maxPoolForward(*inputGpu, - imgSizeH, - imgSizeW, - channels, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - padH, - padW); - MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); - targetCheck->copyFrom(*targetGpu); - checkMatrixEqual(target, targetCheck); - - MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); - MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpuGrad = - GpuMatrix::create(numSamples, outWidth, false, true); - - inputGrad->randomizeUniform(); - targetGrad->randomizeUniform(); - inputGpuGrad->copyFrom(*inputGrad); - targetGpuGrad->copyFrom(*targetGrad); - - inputGrad->maxPoolBackward(*input, - imgSizeH, - imgSizeW, - *targetGrad, - *target, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - 1.0, - 1.0, - padH, - padW); - inputGpuGrad->maxPoolBackward(*inputGpu, - imgSizeH, - imgSizeW, - *targetGpuGrad, - *targetGpu, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - 1.0, - 1.0, - padH, - padW); - MatrixPtr targetBwdCheck = - CpuMatrix::create(numSamples, inWidth, false, false); - targetBwdCheck->copyFrom(*inputGpuGrad); - checkMatrixEqual(inputGrad, targetBwdCheck); -} - -void testAvgPoolFwdBwd(int numSamples, - int channels, - int imgSizeH, - int imgSizeW, - int ksizeH, - int ksizeW, - int strideH, - int strideW, - int padH, - int padW) { - int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); - int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); - - int inWidth = imgSizeH * imgSizeW * channels; - MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); - - int outWidth = channels * outH * outW; - MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); - - input->randomizeUniform(); - target->randomizeUniform(); - inputGpu->copyFrom(*input); - targetGpu->copyFrom(*target); - - target->avgPoolForward(*input, - imgSizeH, - imgSizeW, - channels, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - padH, - padW); - targetGpu->avgPoolForward(*inputGpu, - imgSizeH, - imgSizeW, - channels, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - padH, - padW); - - TensorCheckErr(*target, *targetGpu); - - MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); - MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpuGrad = - GpuMatrix::create(numSamples, outWidth, false, true); - - inputGrad->randomizeUniform(); - targetGrad->randomizeUniform(); - inputGpuGrad->copyFrom(*inputGrad); - targetGpuGrad->copyFrom(*targetGrad); - - inputGrad->avgPoolBackward(*targetGrad, - imgSizeH, - imgSizeW, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - 1.0, - 1.0, - padH, - padW); - inputGpuGrad->avgPoolBackward(*targetGpuGrad, - imgSizeH, - imgSizeW, - ksizeW, - ksizeH, - strideH, - strideW, - outH, - outW, - 1.0, - 1.0, - padH, - padW); - - TensorCheckErr(*inputGrad, *inputGpuGrad); -} - -// TODO(yi): I noticed many such blindly combinatorial tests in this -// file. They are no help to locate defects at all. -TEST(Matrix, PoolFwdBwd) { - for (auto numSamples : {1, 3}) { - for (auto channels : {1, 3}) { - for (auto imgSizeH : {13, 17}) { - for (auto imgSizeW : {17, 19}) { - for (auto sizeX : {2, 3}) { - for (auto sizeY : {2, 3}) { - for (auto sH : {1, 2}) { - for (auto sW : {1, 2}) { - for (auto pH : {0, (sizeY - 1) / 2}) { - for (auto pW : {0, (sizeX - 1) / 2}) { - VLOG(3) << " numSamples=" << numSamples - << " channels=" << channels - << " imgSizeH=" << imgSizeH - << " imgSizeW=" << imgSizeW << " sizeX=" << sizeX - << " sizeY=" << sizeY << " strideH=" << sH - << " strideW=" << sW << " padingH=" << pH - << " padingW=" << pW; - testMaxPoolFwdBwd(numSamples, - channels, - imgSizeH, - imgSizeW, - sizeX, - sizeY, - sH, - sW, - pH, - pW); - testAvgPoolFwdBwd(numSamples, - channels, - imgSizeH, - imgSizeW, - sizeX, - sizeY, - sH, - sW, - pH, - pW); - } - } - } - } - } - } - } - } - } - } -} - -void testMaxOutFwdBwd( - int numSamples, int imgSizeH, int imgSizeW, int channels, int groups) { - int inWidth = imgSizeH * imgSizeW * channels; - int outChannels = channels / groups; - int outWidth = imgSizeH * imgSizeW * outChannels; - - // forward - MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); - - MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); - - IVectorPtr id = CpuIVector::create(numSamples * outWidth, false); - IVectorPtr idGpu = GpuIVector::create(numSamples * outWidth, true); - - input->randomizeUniform(); - inputGpu->copyFrom(*input); - - target->maxoutForward(*input, *id, outChannels, groups); - targetGpu->maxoutForward(*inputGpu, *idGpu, outChannels, groups); - - TensorCheckErr(*target, *targetGpu); - TensorCheckEqual(*id, *idGpu); - - // backward - MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); - - MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpuGrad = - GpuMatrix::create(numSamples, outWidth, false, true); - - inputGrad->randomizeUniform(); - targetGrad->randomizeUniform(); - inputGpuGrad->copyFrom(*inputGrad); - targetGpuGrad->copyFrom(*targetGrad); - - inputGrad->maxoutBackward(*targetGrad, *id, outChannels, groups); - inputGpuGrad->maxoutBackward(*targetGpuGrad, *idGpu, outChannels, groups); - - TensorCheckErr(*inputGrad, *inputGpuGrad); -} - -TEST(Matrix, MaxOutFwdBwd) { - for (auto numSamples : {5, 10}) { - for (auto channels : {8, 16}) { - for (auto imgSizeH : {14, 28}) { - for (auto imgSizeW : {16, 30}) { - for (auto groups : {2, 4}) { - VLOG(3) << " numSamples=" << numSamples << " channels=" << channels - << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW - << " groups=" << groups; - testMaxOutFwdBwd(numSamples, imgSizeH, imgSizeW, channels, groups); - } - } - } - } - } -} - -TEST(CpuMatrix, copyFrom) { - const size_t height = 31; - const size_t width = 53; - CpuMatrix cpu(height, width); - GpuMatrix gpu(height, width); - CpuMatrix copy(height, width); - - cpu.randomizeUniform(); - gpu.copyFrom(cpu); - copy.copyFrom(gpu, HPPL_STREAM_DEFAULT); - - TensorCheckEqual(cpu, copy); -} - -void testBatch2seqPadding(int batchSize, int inputDim) { - MatrixPtr cpuInput = std::make_shared(batchSize, inputDim); - MatrixPtr gpuInput = std::make_shared(batchSize, inputDim); - cpuInput->randomizeUniform(); - gpuInput->copyFrom(*cpuInput); - - IVectorPtr cpuSequence; - generateSequenceStartPositions(batchSize, cpuSequence); - for (int i = 0; i < int(cpuSequence->getSize()); ++i) { - (cpuSequence->getData())[i] += 1; // so no way that maxSeqLen is 0; - } - - IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); - gpuSequence->copyFrom(*cpuSequence); - - size_t numSeq = cpuSequence->getSize() - 1; - size_t maxSeqLen = *std::max_element(cpuSequence->getData(), - cpuSequence->getData() + numSeq); - - printf("numSeq = %ld, maxSeqLen = %ld\n", numSeq, maxSeqLen); - MatrixPtr cBatch = std::make_shared(numSeq * maxSeqLen, inputDim); - MatrixPtr gBatch = std::make_shared(numSeq * maxSeqLen, inputDim); - MatrixPtr cCheck = std::make_shared(numSeq * maxSeqLen, inputDim); - - // hl_sequence2batch_copy_padding(gBatch->getData(), - // gpuInput->getData(), - // cpuSequence->getData(), - // inputDim, - // maxSeqLen, - // numSeq, - // false, - // true); - // cCheck->copyFrom(*gBatch); - - // int* seqStart = cpuSequence->getData(); - // float* batchData = cBatch->getData(); - // float* seqData = cpuInput->getData(); - // for (size_t i = 0; i < maxSeqLen; i++) { - // for (size_t j = 0; j < numSeq; j++) { - // size_t sequenceStart = seqStart[j]; - // size_t sequenceLength = seqStart[j + 1] - seqStart[j]; - // if (i < sequenceLength) { - // memcpy(batchData + (i * numSeq + j) * inputDim, - // seqData + (sequenceStart + i) * inputDim, - // inputDim * sizeof(real)); - // } else { - // memset(batchData + (i * numSeq + j) * inputDim, - // 0, - // inputDim * sizeof(real)); - // } - // } - // } - - // TensorCheckErr(*cBatch, *cCheck); -} - -TEST(Matrix, warpCTC) { - for (auto batchSize : {1, 3, 17}) { - for (auto inputDim : {1, 3, 31}) { - VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim; - testBatch2seqPadding(batchSize, inputDim); - } - } -} - -void testMaxPool3DFwdBwd(int numSamples, - int channels, - int imgSizeD, - int imgSizeH, - int imgSizeW, - int ksizeD, - int ksizeH, - int ksizeW, - int strideD, - int strideH, - int strideW, - int padD, - int padH, - int padW) { - int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true); - int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); - int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); - - int inWidth = channels * imgSizeD * imgSizeH * imgSizeW; - MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); - - int outWidth = channels * outD * outH * outW; - MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); - MatrixPtr maxIdx = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr maxIdxGpu = GpuMatrix::create(numSamples, outWidth, false, true); - - input->randomizeUniform(); - target->randomizeUniform(); - inputGpu->copyFrom(*input); - targetGpu->copyFrom(*target); - - target->maxPool3DForward(*input, - *maxIdx, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW); - targetGpu->maxPool3DForward(*inputGpu, - *maxIdxGpu, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW); - MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); - targetCheck->copyFrom(*targetGpu); - checkMatrixEqual(target, targetCheck); - - MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); - MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpuGrad = - GpuMatrix::create(numSamples, outWidth, false, true); - - inputGrad->randomizeUniform(); - targetGrad->randomizeUniform(); - inputGpuGrad->copyFrom(*inputGrad); - targetGpuGrad->copyFrom(*targetGrad); - - inputGrad->maxPool3DBackward(*targetGrad, - *maxIdx, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW, - 1.0, - 1.0); - inputGpuGrad->maxPool3DBackward(*targetGpuGrad, - *maxIdxGpu, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW, - 1.0, - 1.0); - MatrixPtr targetBwdCheck = - CpuMatrix::create(numSamples, inWidth, false, false); - targetBwdCheck->copyFrom(*inputGpuGrad); - checkMatrixEqual(inputGrad, targetBwdCheck); -} - -void testAvgPool3DFwdBwd(int numSamples, - int channels, - int imgSizeD, - int imgSizeH, - int imgSizeW, - int ksizeD, - int ksizeH, - int ksizeW, - int strideD, - int strideH, - int strideW, - int padD, - int padH, - int padW) { - int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true); - int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); - int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); - - int inWidth = imgSizeD * imgSizeH * imgSizeW * channels; - MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); - - int outWidth = channels * outD * outH * outW; - MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); - - input->randomizeUniform(); - target->randomizeUniform(); - inputGpu->copyFrom(*input); - targetGpu->copyFrom(*target); - - target->avgPool3DForward(*input, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW); - - targetGpu->avgPool3DForward(*inputGpu, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW); - - TensorCheckErr(*target, *targetGpu); - - MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); - MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); - MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); - MatrixPtr targetGpuGrad = - GpuMatrix::create(numSamples, outWidth, false, true); - - inputGrad->randomizeUniform(); - targetGrad->randomizeUniform(); - inputGpuGrad->copyFrom(*inputGrad); - targetGpuGrad->copyFrom(*targetGrad); - - inputGrad->avgPool3DBackward(*targetGrad, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW, - 1.0, - 1.0); - - inputGpuGrad->avgPool3DBackward(*targetGpuGrad, - imgSizeD, - imgSizeH, - imgSizeW, - outD, - outH, - outW, - ksizeD, - ksizeH, - ksizeW, - strideD, - strideH, - strideW, - padD, - padH, - padW, - 1.0, - 1.0); - TensorCheckErr(*inputGrad, *inputGpuGrad); -} - -// TODO(yi): I noticed many such blindly combinatorial tests in this -// file. They are no help to locate defects at all. -TEST(Matrix, Pool3DFwdBwd) { - for (auto numSamples : {1, 3}) { - for (auto channels : {3}) { - for (auto imgSizeD : {9, 16}) { - for (auto imgSizeH : {9, 32}) { - for (auto imgSizeW : {9, 32}) { - for (auto sizeX : {3}) { - for (auto sizeY : {3}) { - for (auto sizeZ : {3}) { - for (auto sD : {2}) { - for (auto sH : {2}) { - for (auto sW : {2}) { - for (auto pD : {0, (sizeZ - 1) / 2}) { - for (auto pH : {0, (sizeY - 1) / 2}) { - for (auto pW : {0, (sizeX - 1) / 2}) { - VLOG(3) << " numSamples=" << numSamples - << " channels=" << channels - << " imgSizeD=" << imgSizeD - << " imgSizeH=" << imgSizeH - << " imgSizeW=" << imgSizeW - << " sizeX=" << sizeX - << " sizeY=" << sizeY - << " sizeZ=" << sizeZ << " strideD=" << sD - << " strideH=" << sH << " strideW=" << sW - << " padingD=" << pD << " padingH=" << pH - << " padingW=" << pW; - - testMaxPool3DFwdBwd(numSamples, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - sizeX, - sizeY, - sizeZ, - sD, - sH, - sW, - pD, - pH, - pW); - testAvgPool3DFwdBwd(numSamples, - channels, - imgSizeD, - imgSizeH, - imgSizeW, - sizeX, - sizeY, - sizeZ, - sD, - sH, - sW, - pD, - pH, - pW); - } - } - } - } - } - } - } - } - } - } - } - } - } - } - - // for (auto numSamples : {1, 3}) { - // for (auto channels : {1, 3}) { - // for (auto imgSizeD : {9,16}) { - // for (auto imgSizeH : {9, 32}) { - // for (auto imgSizeW : {9, 32}) { - // for (auto sizeX : {2, 3}) { - // for (auto sizeY : {2, 3}) { - // for (auto sizeZ : {2,3}){ - // for (auto sD : {1, 2}) { - // for (auto sH : {1, 2}) { - // for (auto sW : {1, 2}) { - // for (auto pD : {0, (sizeZ - 1) / 2}){ - // for (auto pH : {0, (sizeY - 1) / 2}) { - // for (auto pW : {0, (sizeX - 1) / 2}) { - // VLOG(3) << " numSamples=" << numSamples - // << " channels=" << channels - // << " imgSizeD=" << imgSizeD - // << " imgSizeH=" << imgSizeH - // << " imgSizeW=" << imgSizeW - // << " sizeX=" << sizeX - // << " sizeY=" << sizeY - // << " sizeZ=" << sizeZ - // << " strideD=" << sD - // << " strideH=" << sH - // << " strideW=" << sW - // << " padingD=" << pD - // << " padingH=" << pH - // << " padingW=" << pW; - // - // testMaxPool3DFwdBwd(numSamples, - // channels, - // imgSizeD, - // imgSizeH, - // imgSizeW, - // sizeX, - // sizeY, - // sizeZ, - // sD, - // sH, - // sW, - // pD, - // pH, - // pW); - // testAvgPool3DFwdBwd(numSamples, - // channels, - // imgSizeD, - // imgSizeH, - // imgSizeW, - // sizeX, - // sizeY, - // sizeZ, - // sD, - // sH, - // sW, - // pD, - // pH, - // pW); - // } - // } - // } - // } - // } - // } - // } - // } - // } - // } - // } - // } - // } - // } -} - -void testMatrixCol2Vol(int depth, int height, int width) { - int channel = 3; - int filterX = 3, filterY = 4, filterZ = 5; - int strideX = 2, strideY = 2, strideZ = 2; - int padX = 1, padY = 1, padZ = 1; - - MatrixPtr cpuImage = - std::make_shared(channel, depth * height * width); - MatrixPtr gpuImage = - std::make_shared(channel, depth * height * width); - cpuImage->randomizeUniform(); - gpuImage->copyFrom(*cpuImage); - - int outD = outputSize(depth, filterZ, padZ, strideZ, true); - int outH = outputSize(height, filterY, padY, strideY, true); - int outW = outputSize(width, filterX, padX, strideX, true); - - int colBufHeight = channel * filterZ * filterY * filterX; - int colBufWidth = outD * outH * outW; - MatrixPtr cpuColBuf = std::make_shared(colBufHeight, colBufWidth); - MatrixPtr gpuColBuf = std::make_shared(colBufHeight, colBufWidth); - cpuColBuf->vol2Col(cpuImage->getData(), - channel, - depth, - height, - width, - filterZ, - filterY, - filterX, - strideZ, - strideY, - strideX, - padZ, - padY, - padX); - gpuColBuf->vol2Col(gpuImage->getData(), - channel, - depth, - height, - width, - filterZ, - filterY, - filterX, - strideZ, - strideY, - strideX, - padZ, - padY, - padX); - TensorCheckEqual(*cpuColBuf, *gpuColBuf); - - cpuColBuf->randomizeUniform(); - gpuColBuf->copyFrom(*cpuColBuf); - cpuColBuf->col2Vol(cpuImage->getData(), - channel, - depth, - height, - width, - filterZ, - filterY, - filterX, - strideZ, - strideY, - strideX, - padZ, - padY, - padX, - 1.0, - 1.0); - gpuColBuf->col2Vol(gpuImage->getData(), - channel, - depth, - height, - width, - filterZ, - filterY, - filterX, - strideZ, - strideY, - strideX, - padZ, - padY, - padX, - 1.0, - 1.0); - TensorCheckErr(*cpuImage, *gpuImage); -} - -TEST(Matrix, col2Vol) { - for (auto depth : {9, 16, 64}) { - for (auto height : {9, 11, 128}) { - for (auto width : {9, 32, 128}) { - VLOG(3) << "depth=" << depth << " height=" << height - << " width=" << width; - testMatrixCol2Vol(depth, height, width); - } - } - } -} - -#endif diff --git a/paddle/math/tests/test_matrixUtil.h b/paddle/math/tests/test_matrixUtil.h deleted file mode 100644 index 86297547dcd83ca87d1c87a8489f7af2f3e9f492..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_matrixUtil.h +++ /dev/null @@ -1,233 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include "paddle/math/SparseMatrix.h" - -namespace paddle { - -void checkMatrixEqual(const MatrixPtr& a, const MatrixPtr& b) { - ASSERT_EQ(a->getWidth(), b->getWidth()); - ASSERT_EQ(a->getHeight(), b->getHeight()); - ASSERT_EQ(a->isTransposed(), b->isTransposed()); - for (size_t r = 0; r < a->getHeight(); ++r) { - for (size_t c = 0; c < a->getWidth(); ++c) { - ASSERT_FLOAT_EQ(a->getElement(r, c), b->getElement(r, c)); - } - } -} - -void checkSMatrixEqual(const CpuSparseMatrix& a, const CpuSparseMatrix& b) { - ASSERT_EQ(a.getWidth(), b.getWidth()); - ASSERT_EQ(a.getHeight(), b.getHeight()); - ASSERT_EQ(a.isTransposed(), b.isTransposed()); - ASSERT_EQ(a.getFormat(), b.getFormat()); - ASSERT_EQ(a.getElementCnt(), b.getElementCnt()); - for (size_t r = 0; r < a.getElementCnt(); ++r) { - ASSERT_FLOAT_EQ(a.getValue()[r], b.getValue()[r]); - } -} - -void checkSMatrixEqual(const CpuSparseMatrixPtr& a, - const CpuSparseMatrixPtr& b) { - ASSERT_EQ(a->getWidth(), b->getWidth()); - ASSERT_EQ(a->getHeight(), b->getHeight()); - ASSERT_EQ(a->isTransposed(), b->isTransposed()); - ASSERT_EQ(a->getFormat(), b->getFormat()); - ASSERT_EQ(a->getElementCnt(), b->getElementCnt()); - for (size_t r = 0; r < a->getElementCnt(); ++r) { - ASSERT_FLOAT_EQ(a->getValue()[r], b->getValue()[r]); - } -} - -void checkSMatrixEqual2(const CpuSparseMatrixPtr& a, - const CpuSparseMatrixPtr& b) { - ASSERT_EQ(a->getWidth(), b->getWidth()); - ASSERT_EQ(a->getHeight(), b->getHeight()); - ASSERT_EQ(a->isTransposed(), b->isTransposed()); - ASSERT_EQ(a->getFormat(), b->getFormat()); - ASSERT_EQ(a->getValueType(), b->getValueType()); - ASSERT_EQ(a->getElementCnt(), b->getElementCnt()); - if (a->getFormat() == SPARSE_CSR) { - for (size_t r = 0; r < a->getElementCnt(); ++r) { - ASSERT_EQ(a->getCols()[r], b->getCols()[r]); - if (a->getValueType() == FLOAT_VALUE) { - ASSERT_FLOAT_EQ(a->getValue()[r], b->getValue()[r]); - } - } - for (size_t r = 0; r <= a->getHeight(); r++) { - ASSERT_EQ(a->getRows()[r], b->getRows()[r]); - } - } else { - for (size_t r = 0; r < a->getElementCnt(); ++r) { - ASSERT_EQ(a->getRows()[r], b->getRows()[r]); - if (a->getValueType() == FLOAT_VALUE) { - ASSERT_FLOAT_EQ(a->getValue()[r], b->getValue()[r]); - } - } - for (size_t r = 0; r <= a->getWidth(); r++) { - ASSERT_EQ(a->getCols()[r], b->getCols()[r]); - } - } -} - -void checkSMatrixEqual2Dense(const CpuSparseMatrix& a, const CpuMatrix& b) { - ASSERT_EQ(a.getWidth(), b.getWidth()); - ASSERT_EQ(a.getHeight(), b.getHeight()); - ASSERT_EQ(a.isTransposed(), b.isTransposed()); - - if (a.getFormat() == SPARSE_CSC) { - int* rows = a.getRows(); - for (size_t i = 0; i < a.getWidth(); i++) { - for (size_t j = a.getColStartIdx(i); j < a.getColStartIdx(i + 1); j++) { - if (a.getValueType() == FLOAT_VALUE) { - ASSERT_FLOAT_EQ(a.getValue()[j], b.getElement(rows[j], i)); - } else { - ASSERT_FLOAT_EQ(1.0, b.getElement(rows[j], i)); - } - } - } - } else { - int* cols = a.getCols(); - for (size_t i = 0; i < a.getHeight(); i++) { - for (size_t j = a.getRowStartIdx(i); j < a.getRowStartIdx(i + 1); j++) { - if (a.getValueType() == FLOAT_VALUE) { - ASSERT_FLOAT_EQ(a.getValue()[j], b.getElement(i, cols[j])); - } else { - ASSERT_FLOAT_EQ(1.0, b.getElement(i, cols[j])); - } - } - } - } -} - -void checkSMatrixEqual2Dense(const CpuSparseMatrixPtr& a, - const CpuMatrixPtr& b) { - ASSERT_EQ(a->getWidth(), b->getWidth()); - ASSERT_EQ(a->getHeight(), b->getHeight()); - ASSERT_EQ(a->isTransposed(), b->isTransposed()); - - if (a->getFormat() == SPARSE_CSC) { - int* rows = a->getRows(); - for (size_t i = 0; i < a->getWidth(); i++) { - for (size_t j = a->getColStartIdx(i); j < a->getColStartIdx(i + 1); j++) { - if (a->getValueType() == FLOAT_VALUE) { - ASSERT_FLOAT_EQ(a->getValue()[j], b->getElement(rows[j], i)); - } else { - ASSERT_FLOAT_EQ(1.0, b->getElement(rows[j], i)); - } - } - } - } else { - int* cols = a->getCols(); - for (size_t i = 0; i < a->getHeight(); i++) { - for (size_t j = a->getRowStartIdx(i); j < a->getRowStartIdx(i + 1); j++) { - if (a->getValueType() == FLOAT_VALUE) { - ASSERT_FLOAT_EQ(a->getValue()[j], b->getElement(i, cols[j])); - } else { - ASSERT_FLOAT_EQ(1.0, b->getElement(i, cols[j])); - } - } - } - } -} - -void checkSMatrixErr(const CpuSparseMatrixPtr& a, const CpuSparseMatrixPtr& b) { -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - ASSERT_EQ(a->getWidth(), b->getWidth()); - ASSERT_EQ(a->getHeight(), b->getHeight()); - ASSERT_EQ(a->isTransposed(), b->isTransposed()); - ASSERT_EQ(a->getFormat(), b->getFormat()); - ASSERT_EQ(a->getValueType(), b->getValueType()); - ASSERT_EQ(a->getElementCnt(), b->getElementCnt()); - int count = 0; - if (a->getFormat() == SPARSE_CSR) { - for (size_t r = 0; r < a->getElementCnt(); ++r) { - ASSERT_EQ(a->getCols()[r], b->getCols()[r]); - if (a->getValueType() == FLOAT_VALUE) { - real aVal = a->getValue()[r]; - real bVal = b->getValue()[r]; - if (std::abs(aVal - bVal) > err) { - if ((std::abs(aVal - bVal) / std::abs(aVal)) > (err / 10.0f)) { - LOG(INFO) << "a=" << aVal << "\t" - << "b=" << bVal; - count++; - } - } - } - } - for (size_t r = 0; r <= a->getHeight(); r++) { - ASSERT_EQ(a->getRows()[r], b->getRows()[r]); - } - } else { - for (size_t r = 0; r < a->getElementCnt(); ++r) { - ASSERT_EQ(a->getRows()[r], b->getRows()[r]); - if (a->getValueType() == FLOAT_VALUE) { - real aVal = a->getValue()[r]; - real bVal = b->getValue()[r]; - if (std::abs(aVal - bVal) > err) { - if ((std::abs(aVal - bVal) / std::abs(aVal)) > (err / 10.0f)) { - count++; - } - } - } - } - for (size_t r = 0; r <= a->getWidth(); r++) { - ASSERT_EQ(a->getCols()[r], b->getCols()[r]); - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; -} - -void checkMatrixErr(const Matrix& matrix1, const Matrix& matrix2) { - CHECK(matrix1.getHeight() == matrix2.getHeight()); - CHECK(matrix1.getWidth() == matrix2.getWidth()); -#ifndef PADDLE_TYPE_DOUBLE - real err = 1e-3; -#else - real err = 1e-10; -#endif - - int height = matrix1.getHeight(); - int width = matrix1.getWidth(); - const real* data1 = matrix1.getData(); - const real* data2 = matrix2.getData(); - int count = 0; - for (int i = 0; i < height; i++) { - for (int j = 0; j < width; j++) { - real a = data1[i * width + j]; - real b = data2[i * width + j]; - if (std::abs(a - b) > err) { - if ((std::abs(a - b) / std::abs(a)) > (err / 10.0f)) { - count++; - } - } - } - } - EXPECT_EQ(count, 0) << "There are " << count << " different element."; -} - -void checkDataEqual(const real* a, const real* b, size_t size) { - for (size_t i = 0; i < size; ++i) { - ASSERT_FLOAT_EQ(a[i], b[i]); - } -} - -} // namespace paddle diff --git a/paddle/math/tests/test_sparseMatrixCompare.cpp b/paddle/math/tests/test_sparseMatrixCompare.cpp deleted file mode 100644 index 12647d21a29936e169b893ec8119b64fec9af580..0000000000000000000000000000000000000000 --- a/paddle/math/tests/test_sparseMatrixCompare.cpp +++ /dev/null @@ -1,174 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_CUDA -/// This unittest checks GpuSparseMatrix/CpuSparseMatrix get same result, -// so disable when -/// only cpu version. - -#include -#include "paddle/math/Matrix.h" -#include "paddle/utils/Util.h" -#include "test_matrixUtil.h" - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static inline int uniformRandom(int n) { return n == 0 ? 0 : rand() % n; } - -void testSpMatrixAddBias(int M, int N, real rate, real scale) { - int nnz = M * N * rate; - - MatrixPtr cpuA(new CpuSparseMatrix(M, N, nnz)); - MatrixPtr cpuB = std::make_shared(1, N); - - MatrixPtr gpuA(new GpuSparseMatrix(M, N, nnz)); - MatrixPtr gpuB = std::make_shared(1, N); - - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - - hl_stream_t stream(HPPL_STREAM_1); - gpuA->copyFrom(*cpuA, stream); - gpuB->copyFrom(*cpuB, stream); - hl_stream_synchronize(stream); - - cpuA->addBias(*cpuB, scale); - gpuA->addBias(*gpuB, scale); - - MatrixPtr outputCheck(new CpuSparseMatrix(M, N, nnz)); - outputCheck->copyFrom(*gpuA, stream); - hl_stream_synchronize(stream); - checkSMatrixEqual2(std::dynamic_pointer_cast(cpuA), - std::dynamic_pointer_cast(outputCheck)); -} - -void testSpMatrixAddDense(int M, int N, real rate) { // add3 - int nnz = M * N * rate; - - MatrixPtr cpuA(new CpuSparseMatrix(M, N, nnz)); - MatrixPtr cpuB = std::make_shared(M, N); - - MatrixPtr gpuA(new GpuSparseMatrix(M, N, nnz)); - MatrixPtr gpuB = std::make_shared(M, N); - - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - - hl_stream_t stream(HPPL_STREAM_3); - gpuA->copyFrom(*cpuA, stream); - gpuB->copyFrom(*cpuB, stream); - hl_stream_synchronize(stream); - - cpuA->add3(cpuB); - gpuA->add3(gpuB); - - MatrixPtr outputCheck(new CpuSparseMatrix(M, N, nnz)); - outputCheck->copyFrom(*gpuA, stream); - hl_stream_synchronize(stream); - checkSMatrixEqual2(std::dynamic_pointer_cast(cpuA), - std::dynamic_pointer_cast(outputCheck)); -} - -void testSpMatrixMul(int M, int N, int K, real rate) { - int nnz = M * N * rate; - - MatrixPtr cpuA = std::make_shared(M, K); - MatrixPtr cpuB = std::make_shared(N, K); - MatrixPtr cpuC(new CpuSparseMatrix(M, N, nnz)); - - MatrixPtr gpuA = std::make_shared(M, K); - MatrixPtr gpuB = std::make_shared(N, K); - MatrixPtr gpuC(new GpuSparseMatrix(M, N, nnz)); - - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - cpuC->randomizeUniform(); - - hl_stream_t stream(HPPL_STREAM_3); - gpuA->copyFrom(*cpuA, stream); - gpuB->copyFrom(*cpuB, stream); - gpuC->copyFrom(*cpuC, stream); - hl_stream_synchronize(stream); - - cpuC->mul(*cpuA, *cpuB->getTranspose(), 1, 1); - gpuC->mul(*gpuA, *gpuB->getTranspose(), 1, 1); - - MatrixPtr outputCheck(new CpuSparseMatrix(M, N, nnz)); - outputCheck->copyFrom(*gpuC, stream); - hl_stream_synchronize(stream); - checkSMatrixErr(std::dynamic_pointer_cast(cpuC), - std::dynamic_pointer_cast(outputCheck)); -} - -void testSpMatrixCollectBias(int M, int N, real rate) { - int nnz = M * N * rate; - LOG(INFO) << "nnz=" << nnz; - - MatrixPtr cpuA(new CpuSparseMatrix(M, N, nnz)); - MatrixPtr cpuB = std::make_shared(1, N); - - MatrixPtr gpuA(new GpuSparseMatrix(M, N, nnz)); - MatrixPtr gpuB = std::make_shared(1, N); - - cpuA->randomizeUniform(); - cpuB->randomizeUniform(); - - hl_stream_t stream(HPPL_STREAM_3); - gpuA->copyFrom(*cpuA, stream); - gpuB->copyFrom(*cpuB, stream); - hl_stream_synchronize(stream); - - cpuB->collectBias(*cpuA, 1); - gpuB->collectBias(*gpuA, 1); - - MatrixPtr outputCheck = std::make_shared(1, N); - outputCheck->copyFrom(*gpuB, stream); - hl_stream_synchronize(stream); - checkMatrixErr(*cpuB, *outputCheck); -} - -TEST(SMatrix, sMatrixOp) { - for (auto height : {1, 11, 200}) { - for (auto width : {200, 2048, 20480}) { - VLOG(3) << " height=" << height << " width=" << width; - for (auto rate : {0.02, 0.1}) { - testSpMatrixAddDense(height, width, rate); - testSpMatrixAddBias(height, width, rate, 1.0); - } - } - } -} - -TEST(SMatrix, sMatrixMul) { - for (auto M : {1, 40, 128, 200}) { - for (auto N : {100, 2000, 20480}) { - for (auto K : {100, 512, 1024}) { - VLOG(3) << " M=" << M << " N=" << N << " K=" << K; - testSpMatrixMul(M, N, K, 0.05); - } - } - } -} - -TEST(SMatrix, sMatrixCollectBias) { - for (auto height : {1, 128, 200}) { - for (auto width : {100, 2048, 20480}) { - VLOG(3) << " height=" << height << " width=" << width; - testSpMatrixCollectBias(height, width, 0.1); - } - } -} - -#endif diff --git a/paddle/optimizer/serialization.h b/paddle/optimizer/serialization.h deleted file mode 100644 index bf12eed15f0190b8e856163c68690f3f6eef9a12..0000000000000000000000000000000000000000 --- a/paddle/optimizer/serialization.h +++ /dev/null @@ -1,49 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include -#include -#include -#include "OptimizerConfig.pb.h" -#include "paddle/utils/Logging.h" -#include "tensor.h" - -namespace paddle { -namespace optimizer { - -static void TensorToProto(const Tensor& tensor, TensorProto* proto) { - proto->set_data_type(TensorProto::PADDLE_ELEMENT_TYPE_FLOAT32); - std::stringstream os; - for (size_t i = 0; i < tensor.size(); ++i) { - os << tensor[i]; - proto->add_content(os.str()); - os.str(std::string()); - } -} - -static void ProtoToTensor(const TensorProto& proto, Tensor* tensor) { - std::stringstream sin; - for (auto i = 0; i < proto.content_size(); ++i) { - sin << proto.content(i); - sin >> (*tensor)[i]; - sin.str(std::string()); - sin.clear(); - } -} - -} // namespace optimizer -} // namespace paddle diff --git a/paddle/optimizer/tensor.h b/paddle/optimizer/tensor.h deleted file mode 100644 index d2cef99074335be6f9852d60daa103b9b45a550d..0000000000000000000000000000000000000000 --- a/paddle/optimizer/tensor.h +++ /dev/null @@ -1,68 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. -#pragma once -/** - * @brief tensor used by optimizer - */ - -#include -#include -#include "paddle/utils/Common.h" -#include "paddle/utils/Logging.h" - -namespace paddle { -namespace optimizer { - -template -class TensorT { - public: - TensorT(size_t size) : height_(1), width_(size) { - // new T[size]() initializes all element to zero value. - data_ptr_ = std::shared_ptr(new T[size](), std::default_delete()); - data_ = data_ptr_.get(); - } - - TensorT(T* data, size_t size) - : height_(1), width_(size), data_ptr_(nullptr), data_(data) {} - - TensorT(T* data, size_t h, size_t w) - : height_(h), width_(w), data_ptr_(nullptr), data_(data) {} - - virtual ~TensorT() {} - - T* get_buffer() { return this->data_; } - - T& operator[](const size_t idx) { - CHECK(idx >= 0 && idx < this->width_) << "out of index range"; - return data_[idx]; - } - T& operator[](const size_t idx) const { - CHECK(idx >= 0 && idx < this->width_) << "out of index range"; - return data_[idx]; - } - // TODO: replace with tensorshape - size_t size() const { return this->width_ * this->height_; } - - protected: - size_t height_; - size_t width_; - std::shared_ptr data_ptr_; - T* data_; -}; - -// TODO(zhihong): design problem of dynamic datatype, need to fix it -typedef TensorT Tensor; - -} // namespace optimizer -} // namespace paddle diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp deleted file mode 100644 index 94522f718a0c19bfc704ca92eddef5c5a9cb6919..0000000000000000000000000000000000000000 --- a/paddle/parameter/Argument.cpp +++ /dev/null @@ -1,707 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Argument.h" -#include "paddle/math/SparseMatrix.h" - -#include - -namespace paddle { -static void resizeAndCopy(MatrixPtr& dest, - const MatrixPtr& src, - bool useGpu, - hl_stream_t stream) { - if (src) { - if (!dest) { - dest = src->clone(0, 0, useGpu); - } else { - CHECK_EQ(dest->useGpu(), useGpu); - dest->resize(src->getHeight(), src->getWidth()); - } - dest->copyFrom(*src, stream); - } else { - dest.reset(); - } -} - -static void resizeAndCopy(IVectorPtr& dest, - const IVectorPtr& src, - bool useGpu, - hl_stream_t stream) { - if (src) { - IVector::resizeOrCreate(dest, src->getSize(), useGpu); - dest->copyFrom(*src, stream); - } else { - dest.reset(); - } -} - -static void resizeAndCopy(ICpuGpuVectorPtr& dest, - const ICpuGpuVectorPtr& src, - bool useGpu, - hl_stream_t stream) { - if (src) { - ICpuGpuVector::resizeOrCreate(dest, src->getSize(), useGpu); - dest->copyFrom(*src, stream); - } else { - dest.reset(); - } -} - -static void resizeAndCopy(MatrixPtr& dest, - const MatrixPtr& src, - int32_t startRow, - int32_t copySize, - bool useGpu, - hl_stream_t stream = HPPL_STREAM_DEFAULT) { - if (src) { - CHECK_LE((size_t)startRow + copySize, src->getHeight()); - int height = copySize; - int width = src->getWidth(); - if (!dest) { - dest = src->clone(height, width, useGpu); - } else { - CHECK_EQ(dest->useGpu(), useGpu); - dest->resize(height, width); - } - MatrixPtr submat = src->subMatrix(startRow, copySize); - if (dynamic_cast(dest.get())) { - // copy a subMatrix of CpuSparseMatrix to GpuSparseMatrix. - // First copy it to CPU, and then copy it to the GPU. - MatrixPtr tmp = src->clone(height, width, false); - tmp->copyFrom(*submat, stream); - dest->copyFrom(*tmp, stream); - } else { - dest->copyFrom(*submat, stream); - } - } else { - dest.reset(); - } -} - -static void resizeAndCopy(IVectorPtr& dest, - const IVectorPtr& src, - int32_t startPos, - int32_t copySize, - bool useGpu, - hl_stream_t stream = HPPL_STREAM_DEFAULT) { - if (src) { - CHECK_LE((size_t)startPos + copySize, src->getSize()); - - int height = copySize; - IVector::resizeOrCreate(dest, height, useGpu); - dest->copyFrom(src->getData() + startPos, height, stream); - } else { - dest.reset(); - } -} - -static void resizeAndCopy(ICpuGpuVectorPtr& dest, - const ICpuGpuVectorPtr& src, - int32_t startPos, - int32_t copySize, - bool useGpu, - hl_stream_t stream = HPPL_STREAM_DEFAULT) { - if (src) { - CHECK_LE((size_t)startPos + copySize, src->getSize()); - - ICpuGpuVector::resizeOrCreate(dest, copySize, useGpu); - dest->copyFrom(*src, startPos, copySize, useGpu, stream); - } else { - dest.reset(); - } -} - -static void resizeAndCopy(SVectorPtr& dest, - const SVectorPtr& src, - bool useGpu, - hl_stream_t stream) { - if (src) { - size_t height = src->size(); - if (!dest) { - dest = std::make_shared>(height); - } else { - dest->resize(height); - } - std::copy_n(src->begin(), height, dest->begin()); - } else { - dest.reset(); - } -} - -static void resizeAndCopy(SVectorPtr& dest, - const SVectorPtr& src, - int32_t startPos, - int32_t copySize, - bool useGpu, - hl_stream_t stream = HPPL_STREAM_DEFAULT) { - if (src) { - CHECK_LE((size_t)startPos + copySize, src->size()); - size_t height = copySize; - if (!dest) { - dest = std::make_shared>(height); - } else { - dest->resize(height); - } - std::copy_n(src->begin() + startPos, height, dest->begin()); - } else { - dest.reset(); - } -} - -void Argument::resizeAndCopyFrom(const Argument& src, bool useGpu) { - resizeAndCopyFrom(src, useGpu, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); -} - -void Argument::resizeAndCopyFrom(const Argument& src, - bool useGpu, - hl_stream_t stream) { - dataId = src.dataId; - resizeAndCopy(value, src.value, useGpu, stream); - resizeAndCopy(grad, src.grad, useGpu, stream); - resizeAndCopy(in, src.in, useGpu, stream); - resizeAndCopy(ids, src.ids, useGpu, stream); - resizeAndCopy(sequenceStartPositions, - src.sequenceStartPositions, - false /* useGpu */, - stream); - if (src.hasSubseq()) { - resizeAndCopy(subSequenceStartPositions, - src.subSequenceStartPositions, - false /* useGpu */, - stream); - } - resizeAndCopy(strs, src.strs, useGpu, stream); - frameWidth = src.frameWidth; - frameHeight = src.frameHeight; - frameDepth = src.frameDepth; -} - -int32_t Argument::resizeAndCopyFrom(const Argument& src, - int32_t startSeq, - int32_t copySize, - bool useGpu) { - int32_t size = - resizeAndCopyFrom(src, startSeq, copySize, useGpu, HPPL_STREAM_DEFAULT); - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - return size; -} - -int32_t Argument::resizeAndCopyFrom(const Argument& src, - int32_t startSeq, - int32_t copySize, - bool useGpu, - hl_stream_t stream) { - dataId = src.dataId; - frameWidth = src.frameWidth; - frameHeight = src.frameHeight; - frameDepth = src.frameDepth; - - if (!src.sequenceStartPositions) { - // non-sequence input, copy samples directly - int32_t startRow = startSeq; - resizeAndCopy(in, src.in, startRow, copySize, useGpu, stream); - resizeAndCopy(value, src.value, startRow, copySize, useGpu, stream); - resizeAndCopy(grad, src.grad, startRow, copySize, useGpu, stream); - resizeAndCopy(ids, src.ids, startRow, copySize, useGpu, stream); - resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream); - return copySize; - } else { - // sequence input - const int* sequence = src.sequenceStartPositions->getData(false); - int32_t startRow = sequence[startSeq]; // sample start from here - int32_t endRow = sequence[startSeq + copySize]; // sample end - int32_t copyFeatureSize = endRow - startRow; // num of samples - resizeAndCopy(in, src.in, startRow, copyFeatureSize, useGpu, stream); - resizeAndCopy(value, src.value, startRow, copyFeatureSize, useGpu, stream); - resizeAndCopy(grad, src.grad, startRow, copyFeatureSize, useGpu, stream); - resizeAndCopy(ids, src.ids, startRow, copyFeatureSize, useGpu, stream); - resizeAndCopy(sequenceStartPositions, - src.sequenceStartPositions, - startSeq, - copySize + 1, - false, - stream); - // modify new sequenceStartPositions - int* destSequences = sequenceStartPositions->getMutableData(false); - for (int i = 0; i < copySize + 1; i++) { - destSequences[i] -= startRow; - } - CHECK_EQ(destSequences[0], 0); - CHECK_EQ(destSequences[copySize], copyFeatureSize); - if (src.hasSubseq()) { - // sequence has sub-sequence - int* subSequence = src.subSequenceStartPositions->getMutableData(false); - int32_t subStartSeq = 0; - int32_t subEndSeq = 0; - int numSubSequences = src.getNumSubSequences(); - for (int i = 0; i < numSubSequences + 1; i++) { - if (subSequence[i] == startRow) { - subStartSeq = i; - } else if (subSequence[i] == endRow) { - subEndSeq = i; - break; - } - } - int32_t copySubSize = subEndSeq - subStartSeq; - resizeAndCopy(subSequenceStartPositions, - src.subSequenceStartPositions, - subStartSeq, - copySubSize + 1, - false, - stream); - // modify new subSequenceStartPositions - int* destSubSequences = subSequenceStartPositions->getMutableData(false); - for (int i = 0; i < copySubSize + 1; i++) { - destSubSequences[i] -= startRow; - } - CHECK_EQ(destSubSequences[0], 0); - CHECK_EQ(destSubSequences[copySubSize], copyFeatureSize); - } - resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream); - return copyFeatureSize; - } -} - -void Argument::concat(const std::vector& args, - const std::vector& selectRows, - const std::vector& seqStartPos, - const std::vector& copySize, - bool useGpu, - hl_stream_t stream, - PassType passType) { - CHECK(!subSequenceStartPositions) - << "undefined behavior for subsequence positions"; - - size_t batchSize = 0; - for (size_t i = 0; i < copySize.size(); ++i) - batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]); - - auto copyArg = [batchSize, stream](MatrixPtr& dst, - MatrixPtr src, - int desStartRow, - int srcStartRow, - int size, - bool useGpu) { - if (!src) { - dst.reset(); - return; - } - size_t width = src->getWidth(); - if (!dst) { - dst = src->clone(batchSize, width, useGpu); - } else { - dst->resize(batchSize, width); - } - - MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size); - tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream); - }; - - auto copyIds = [batchSize, stream](IVectorPtr& dst, - const IVectorPtr& src, - int desStartRow, - int srcStartRow, - int size, - bool useGpu) { - if (!src) { - dst.reset(); - return; - } - IVector::resizeOrCreate(dst, batchSize, useGpu); - dst->subVec(desStartRow, size) - ->copyFrom(*src->subVec(srcStartRow, size), stream); - }; - - auto copyStrs = [batchSize](SVectorPtr& dst, - const SVectorPtr& src, - int desStartRow, - int srcStartRow, - int size, - bool useGpu) { - if (!src) { - dst.reset(); - return; - } - if (!dst) { - dst = std::make_shared>(batchSize); - } else { - dst->resize(batchSize); - } - std::copy(src->begin() + srcStartRow, - src->begin() + srcStartRow + size, - dst->begin() + desStartRow); - }; - - dataId = args[0].dataId; - CHECK_NE(seqStartPos.size(), 0UL); - int desStartRow = 0; - for (size_t i = 0; i < copySize.size(); ++i) { - int startPos = seqStartPos[i]; - int endPos = seqStartPos[i + 1]; - CHECK_GE(args.size(), static_cast(endPos - startPos)); - for (int j = startPos; j < endPos; ++j) { - const Argument& arg = args[j - startPos]; - CHECK_EQ(arg.dataId, dataId) << "Arguments to concatenate should have " - << "the same dataId."; - const int srcStartRow = selectRows[j]; - copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu); - copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu); - if (passType != PASS_TEST) { - copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu); - } - copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu); - copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu); - desStartRow += copySize[i]; - } - } - ICpuGpuVector::resizeOrCreate( - sequenceStartPositions, seqStartPos.size(), useGpu); - sequenceStartPositions->copyFrom( - seqStartPos.data(), seqStartPos.size(), useGpu); -} - -void Argument::concat(const std::vector& args, - bool useGpu, - hl_stream_t stream, - PassType passType) { - int32_t batchSize = 0; - int64_t numSequences = 0; - int64_t numSubSequences = 0; - for (auto& arg : args) { - batchSize += arg.getBatchSize(); - numSequences += arg.getNumSequences(); - numSubSequences += arg.getNumSubSequences(); - } - - auto copyArg = [batchSize, stream]( - MatrixPtr& dst, MatrixPtr src, int startRow, bool useGpu) { - if (!src) { - dst.reset(); - return; - } - size_t width = src->getWidth(); - if (!dst) { - dst = src->clone(batchSize, width, useGpu); - } else { - dst->resize(batchSize, width); - } - - MatrixPtr tmpMatrix = dst->subMatrix(startRow, src->getHeight()); - tmpMatrix->copyFrom(*src, stream); - }; - - auto copyIds = [batchSize, stream]( - IVectorPtr& dst, const IVectorPtr& src, int startRow, bool useGpu) { - if (!src) { - dst.reset(); - return; - } - IVector::resizeOrCreate(dst, batchSize, useGpu); - dst->subVec(startRow, src->getSize())->copyFrom(*src, stream); - }; - - auto copyStrs = [batchSize]( - SVectorPtr& dst, const SVectorPtr& src, int startRow, bool useGpu) { - if (!src) { - dst.reset(); - return; - } - if (!dst) { - dst = std::make_shared>(batchSize); - } else { - dst->resize(batchSize); - } - std::copy(src->begin(), src->end(), dst->begin() + startRow); - }; - - auto copySequencePos = [](ICpuGpuVectorPtr& dstSeq, - const ICpuGpuVectorPtr& srcSeq, - int dstNumSequences, - int srcNumSequences, - int& startSequences, - int startRow) { - if (srcSeq) { - ICpuGpuVector::resizeOrCreate(dstSeq, dstNumSequences + 1, false); - const int* src = srcSeq->getData(false); - int* dest = dstSeq->getMutableData(false); - for (int i = 0; i < srcNumSequences + 1; ++i) { - dest[i + startSequences] = src[i] + startRow; - } - startSequences += srcNumSequences; - } else { - dstSeq.reset(); - } - }; - - int startRow = 0; - int startSequences = 0; - int startSubSequences = 0; - dataId = args[0].dataId; - for (auto& arg : args) { - CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have" - << " same dataId"; - copyArg(in, arg.in, startRow, useGpu); - copyArg(value, arg.value, startRow, useGpu); - if (passType != PASS_TEST) copyArg(grad, arg.grad, startRow, useGpu); - copyIds(ids, arg.ids, startRow, useGpu); - copySequencePos(sequenceStartPositions, - arg.sequenceStartPositions, - numSequences, - arg.getNumSequences(), - startSequences, - startRow); - copySequencePos(subSequenceStartPositions, - arg.subSequenceStartPositions, - numSubSequences, - arg.getNumSubSequences(), - startSubSequences, - startRow); - copyStrs(strs, arg.strs, startRow, useGpu); - startRow += arg.getBatchSize(); - } -} - -void Argument::splitByDataId(const std::vector& argus, - std::vector>* arguGroups) { - arguGroups->clear(); - int lastDataId = -1; - for (const auto& argu : argus) { - if (argu.dataId == -1) { - // is -1, then create a new group - arguGroups->emplace_back(); - lastDataId = -1; - } else if (argu.dataId != lastDataId) { - // not -1, also not equal to last Argument, then create a new group - arguGroups->emplace_back(); - lastDataId = argu.dataId; - } else { - // not -1, and equal to last Argument, do nothing - } - arguGroups->back().push_back(argu); - } -} - -void Argument::getSeqInfo(std::vector* seqInfo) const { - const int* starts = sequenceStartPositions->getData(false); - const int* subStarts = - hasSubseq() ? subSequenceStartPositions->getData(false) : nullptr; - size_t numSequences = getNumSequences(); - seqInfo->reserve(numSequences); - int subSeqEnd = 0; - for (size_t i = 0; i < numSequences; ++i) { - SeqInfo info; - info.seqStart = starts[i]; - info.subLevelLength = starts[i + 1] - starts[i]; - info.seqId = i; - if (hasSubseq()) { - info.subSeqStart = subSeqEnd; - while (subStarts[subSeqEnd] < starts[i + 1]) { - ++subSeqEnd; - } - info.topLevelLength = subSeqEnd - info.subSeqStart; - } else { - info.topLevelLength = info.subLevelLength; - info.subSeqStart = 0; // not used - } - seqInfo->push_back(info); - } - std::sort( - seqInfo->begin(), seqInfo->end(), [](const SeqInfo& a, const SeqInfo& b) { - return a.topLevelLength > b.topLevelLength; - }); -} - -void Argument::checkSubset() const { - if (getNumSequences() > getNumSubSequences()) { - LOG(FATAL) << "numSubSequences is less than numSequences (" - << getNumSubSequences() << " vs. " << getNumSequences() << ")"; - } - const int* start = sequenceStartPositions->getData(false); - const int* subStart = subSequenceStartPositions->getData(false); - int seqId = 0; - int subSeqId = 0; - while (seqId < getNumSequences() && subSeqId < getNumSubSequences()) { - if (start[seqId] > subStart[subSeqId]) { - ++subSeqId; - } else if (start[seqId] == subStart[subSeqId]) { - ++subSeqId; - ++seqId; - } else { - LOG(FATAL) << "seqStartPositions is not subset of subSeqStartPositions"; - } - } - if (seqId < getNumSequences()) { - LOG(FATAL) << "seqStartPositions is not subset of subSeqStartPositions"; - } -} - -void Argument::degradeSequence(const Argument& input) { - CHECK_EQ(input.hasSubseq(), 1UL); - size_t numSequences = input.getNumSequences(); - size_t numSubSequences = input.getNumSubSequences(); - ICpuGpuVector::resizeOrCreate( - sequenceStartPositions, numSequences + 1, false); - int* tgtBuf = sequenceStartPositions->getMutableData(false); - const int* starts = input.sequenceStartPositions->getData(false); - const int* subStarts = input.subSequenceStartPositions->getData(false); - int seqId = 0; - for (size_t subSeqId = 0; subSeqId < numSubSequences; ++subSeqId) { - if (subStarts[subSeqId] == starts[seqId]) { - tgtBuf[seqId] = subSeqId; - seqId++; - } - } - tgtBuf[numSequences] = numSubSequences; -} - -void Argument::poolSequenceWithStride(const Argument& input, - size_t stride, - ICpuGpuVectorPtr* stridePostions, - bool reversed) { - // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5, - // then sequenceStartPositions = [0, 2, 3, 4, 7]. - // If reversed = false, stridePostions = [0, 5, 9, 14, 17, 22, 27, 30]; - // else reversed = true, stridePostions = [0, 4, 9, 14, 17, 20, 25, 30] - - CHECK(input.sequenceStartPositions); - CHECK_EQ(input.hasSubseq(), 0UL); - CHECK_GT(stride, 0UL) << "stride must larger than 0"; - size_t numSequences = input.getNumSequences(); - ICpuGpuVector::resizeOrCreate( - sequenceStartPositions, numSequences + 1, false); - const int* starts = input.sequenceStartPositions->getData(false); - int* tgtBuf = sequenceStartPositions->getMutableData(false); - // first index of target sequence and stride positions are both 0 - tgtBuf[0] = 0; - std::vector stridePos; - for (size_t seqId = 0; seqId < numSequences; ++seqId) { - size_t seqLength = starts[seqId + 1] - starts[seqId]; - stridePos.emplace_back(starts[seqId]); - if (seqLength == 0) { - // empty sequence - tgtBuf[seqId + 1] = tgtBuf[seqId]; - } else { - int size = ceil((float)seqLength / stride); - tgtBuf[seqId + 1] = tgtBuf[seqId] + size; - for (int i = 0; i < size - 1; ++i) { - int cur = reversed ? starts[seqId + 1] - (size - 1 - i) * stride - : stridePos.back() + stride; - stridePos.emplace_back(cur); - } - } - } - stridePos.emplace_back(starts[numSequences]); - int size = stridePos.size(); - CHECK_EQ(size - 1, tgtBuf[numSequences]); - ICpuGpuVector::resizeOrCreate(*stridePostions, size, false); - (*stridePostions)->getMutableVector(false)->copyFrom(stridePos.data(), size); -} - -void Argument::getValueString( - std::unordered_map* out) const { - if (value) { - std::ostringstream os; - value->print(os); - out->insert({"value", os.str()}); - } - if (ids) { - std::ostringstream os; - ids->print(os, ids->getSize()); - out->insert({"ids", os.str()}); - } - if (sequenceStartPositions) { - std::ostringstream os; - sequenceStartPositions->getVector(false)->print( - os, sequenceStartPositions->getSize()); - out->insert({"sequence pos", os.str()}); - } - if (subSequenceStartPositions) { - std::ostringstream os; - subSequenceStartPositions->getVector(false)->print( - os, subSequenceStartPositions->getSize()); - out->insert({"sub-sequence pos", os.str()}); - } -} - -void Argument::printValueString(std::ostream& stream, - const std::string& prefix) const { - std::unordered_map out; - getValueString(&out); - for (auto field : {"value", "ids", "sequence pos", "sub-sequence pos"}) { - auto it = out.find(field); - if (it != out.end()) { - stream << prefix << field << ":\n" << it->second; - } - } -} - -void Argument::subArgFrom(const Argument& input, - size_t offset, - size_t height, - size_t width, - bool useGpu, - bool trans, - bool seqFlag, - size_t seqStart, - size_t seqSize) { - if (input.value) { - value = Matrix::create( - input.value->getData() + offset * width, height, width, trans, useGpu); - } - if (input.ids) { - ids = IVector::create(input.ids->getData() + offset, height, useGpu); - } - if (input.grad) { - grad = Matrix::create( - input.grad->getData() + offset * width, height, width, trans, useGpu); - } - if (seqFlag) { - sequenceStartPositions = std::make_shared( - *(input.sequenceStartPositions), seqStart, seqSize); - } -} - -void Argument::reorganizeSeqInfo( - const ICpuGpuVectorPtr seqStartPos, - const ICpuGpuVectorPtr subSeqStartPos, - std::vector>& reorganizedSeqInfo) { - CHECK(seqStartPos); - reorganizedSeqInfo.clear(); - - int seqNum = seqStartPos->getSize() - 1; - int* seqStarts = seqStartPos->getMutableData(false); - - if (subSeqStartPos) { - int* subSeqStarts = subSeqStartPos->getMutableData(false); - reorganizedSeqInfo.resize(seqNum, std::vector()); - int seqIdx = 0; - for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) { - reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); - if (subSeqStarts[i] == seqStarts[seqIdx + 1]) { - seqIdx++; - if (seqIdx == seqNum) return; - reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); - } - } - } else { - reorganizedSeqInfo.resize(1, std::vector(seqNum + 1, 0)); - memcpy(reorganizedSeqInfo[0].data(), - seqStarts, - sizeof(int) * seqStartPos->getSize()); - } -} - -} // namespace paddle diff --git a/paddle/parameter/Argument.h b/paddle/parameter/Argument.h deleted file mode 100644 index e580d38216b699360fb30f135be8052ab56abf66..0000000000000000000000000000000000000000 --- a/paddle/parameter/Argument.h +++ /dev/null @@ -1,349 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "hl_gpu.h" - -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -typedef std::shared_ptr> SVectorPtr; - -struct Argument { - Argument() - : in(nullptr), - value(nullptr), - ids(nullptr), - grad(nullptr), - strs(nullptr), - frameHeight(0), - frameWidth(0), - frameDepth(0), - sequenceStartPositions(nullptr), - subSequenceStartPositions(nullptr), - cpuSequenceDims(nullptr), - deviceId(-1), - allCount(0), - valueCount(0), - gradCount(0), - dataId(0) {} - Argument(const Argument& argument) { - *this = argument; - valueCount = 0; - gradCount = 0; - dataId = argument.dataId; - } - ~Argument() {} - - void operator=(const Argument& argument) { - in = argument.in; - value = argument.value; - ids = argument.ids; - grad = argument.grad; - strs = argument.strs; - sequenceStartPositions = argument.sequenceStartPositions; - subSequenceStartPositions = argument.subSequenceStartPositions; - cpuSequenceDims = argument.cpuSequenceDims; - deviceId = argument.deviceId; - allCount = argument.allCount; - frameHeight = argument.frameHeight; - frameWidth = argument.frameWidth; - frameDepth = argument.frameDepth; - dataId = argument.dataId; - } - - MatrixPtr in; // used if needed - MatrixPtr value; - IVectorPtr ids; // a sequence of ids. Can be use for class id for costLayer - MatrixPtr grad; // If empty, gradient is not needed. - SVectorPtr strs; - - // A dataBatch includes batchSize frames, one frame maybe not only vector - size_t frameHeight; - size_t frameWidth; - size_t frameDepth; - - // If NULL, each position is treated independently. - // Otherwise, its size should be #NumberOfSequences + 1. - // The first position is always 0 and - // the last position should be equal to batchSize. - ICpuGpuVectorPtr sequenceStartPositions; - - // If NULL, each sequence has no subsequence. - // Otherwise, its size should be #NumberOfSubSequences + 1. - // The first position is always 0 and - // the last position should be equal to batchSize. - ICpuGpuVectorPtr subSequenceStartPositions; - - // dimension of sequence, stored only in CPU - IVectorPtr cpuSequenceDims; - - int deviceId; // the GPU device id which the argument in - int allCount; // the number of output layers using this argument - mutable int valueCount; // waiting this member when layer do forward - mutable int gradCount; // waiting this member when layer do backward - mutable LockedCondition valueReadyCond; - mutable LockedCondition gradReadyCond; - - int dataId; // dataProvider id - - /* Increase the reference count of the argument. */ - void countIncrement() { allCount++; } - - int getAllCount() const { return allCount; } - - void waitValueReady() const { - valueReadyCond.wait([this] { return (valueCount != 0); }); - - std::lock_guard guard(*valueReadyCond.mutex()); - valueCount--; - } - - void notifyValueReady() const { - valueReadyCond.notify_all([this] { valueCount = allCount; }); - } - - void waitGradReady() const { - gradReadyCond.wait([this] { return (gradCount == allCount); }); - gradCount = 0; - } - - void notifyGradReady() const { - gradReadyCond.notify_all([this] { gradCount++; }); - } - - int64_t getBatchSize() const { - if (value) return value->getHeight(); - if (ids) return ids->getSize(); - if (grad) return grad->getHeight(); - if (in) return in->getHeight(); - if (strs) return strs->size(); - return 0; - } - size_t getFrameHeight() const { return frameHeight; } - size_t getFrameWidth() const { return frameWidth; } - size_t getFrameDepth() const { return frameDepth; } - void setFrameHeight(size_t h) { frameHeight = h; } - void setFrameWidth(size_t w) { frameWidth = w; } - void setFrameDepth(size_t d) { frameDepth = d; } - - int64_t getNumSequences() const { - return sequenceStartPositions ? sequenceStartPositions->getSize() - 1 - : getBatchSize(); - } - - int64_t getNumSubSequences() const { - return subSequenceStartPositions ? subSequenceStartPositions->getSize() - 1 - : getBatchSize(); - } - - bool hasSeq() const { return sequenceStartPositions != nullptr; } - bool hasSubseq() const { return subSequenceStartPositions != nullptr; } - - const int* getCpuStartPositions() const { - return hasSubseq() ? subSequenceStartPositions->getData(false) - : sequenceStartPositions->getData(false); - } - - static inline real sum(const std::vector& arguments) { - real cost = 0; - for (auto& arg : arguments) { - if (arg.value) { - SetDevice device(arg.deviceId); - cost += arg.value->getSum(); - } - } - return cost; - } - - /** - * @brief (value, ids, grad, sequenceStartPositions) of output are subset of - * input. Note that, output share the same memory of input. - * - * @param input[in] input - * @param offset[in] offset in terms of rows - * @param height[in] height of output.value - * @param width[in] width of output.value - * @param useGpu[in] - * @param trans[in] whether input.value is transform - * @param seqFlag[in] whether input has sequenceStartPositions - * @param seqStart[in] offset of input.sequenceStartPositions - * @param seqSize[in] lenght of output.sequenceStartPositions - */ - void subArgFrom(const Argument& input, - size_t offset, - size_t height, - size_t width, - bool useGpu, - bool trans = false, - bool seqFlag = false, - size_t seqStart = 0, - size_t seqSize = 0); - /* - * for sequence input: - * startSeq: the sequence id of start - * copySize: how many sequences need to copy - * return value: how many samples are copied - * for non-sequence input: - * startSeq: the sample id of start - * copySize: how many samples need to copy - * return value: how many samples are copied - * Note that when specifying the stream explicitly in this case, - * synchronize should also be called somewhere after this function - */ - int32_t resizeAndCopyFrom(const Argument& src, - int32_t startSeq, - int32_t copySize, - bool useGpu, - hl_stream_t stream); - - /* - * same with the above function, except that the stream is - * HPPL_STREAM_DEFAULT and synchronize is automatically called - * inside it - */ - int32_t resizeAndCopyFrom(const Argument& src, - int32_t startSeq, - int32_t copySize, - bool useGpu = FLAGS_use_gpu); - - void resizeAndCopyFrom(const Argument& src, bool useGpu, hl_stream_t stream); - - /* - * same with the above function, except that the stream is - * HPPL_STREAM_DEFAULT and synchronize is automatically called - * inside it - */ - void resizeAndCopyFrom(const Argument& src, bool useGpu = FLAGS_use_gpu); - - /* - @brief Concatenate several arguments into one and put the result into it. - @param args : a vector of argument, each element of which is a frame in a - batch of sequences. - @param selectRows : select several row of args to concatenate - @param seqStartPos : sequence start positions in the final Argument - @param hl_stream_t : cuda stream - @param passTyoe : type of task, training or testing - */ - void concat(const std::vector& args, - const std::vector& selectRows, - const std::vector& seqStartPos, - const std::vector& copySize, - bool useGpu, - hl_stream_t stream, - PassType passType); - - /* - Concatenate several args into one and put the result into this. - */ - void concat(const std::vector& src, - bool useGpu = FLAGS_use_gpu, - hl_stream_t stream = HPPL_STREAM_DEFAULT, - PassType passType = PASS_TEST); - - /* - * split vector to several vectors according to dataId - */ - static void splitByDataId(const std::vector& argus, - std::vector>* arguGroups); - - struct SeqInfo { - // Equal to sequence length for sequence data - // Equal to number of subsequences for subsequence data - int topLevelLength; - - int seqStart; - int seqId; - - // Equal to topLevelLength for sequence data - // Equal to sum of the length of subsequences for subsequence data - int subLevelLength; - - // Only used for subsequence data, start position of this sequence - // is subSequenceStartPositions, i.e. - // subSequenceStartPositions[subSeqStart] == seqStart - int subSeqStart; - }; - /* - Get SeqInfo for each sequence of this argument - Elements in *seqInfo are sorted by topLevelLength in descending order - */ - void getSeqInfo(std::vector* segInfo) const; - - /* - Check Whether sequenceStartPositions is subset of - subSequenceStartPositions. - */ - void checkSubset() const; - - /* - sequence has sub-sequence degrades to a sequence. - */ - void degradeSequence(const Argument& input); - - /* - After pooling with stride n (n is smaller than sequence length), - a long sequence will be shorten. - This function is invalid for sequence having sub-sequence. - */ - void poolSequenceWithStride(const Argument& input, - size_t stride, - ICpuGpuVectorPtr* stridePositions, - bool reversed = false); - /** - * @brief getValueString will return the argument's output in string. There - * are several kinds of output. The keys of output dictionary are 'value', - * 'id', 'sequence pos', 'sub-sequence pos'. - * @param out [out]: the return values. - */ - void getValueString(std::unordered_map* out) const; - - /** - * @brief printValueString will print the argument's output in order of - * 'value', 'id', 'sequence pos', 'sub-sequence pos'. - * @param stream: Output stream - * @param prefix: line prefix for printing. - */ - void printValueString(std::ostream& stream, - const std::string& prefix = "") const; - - /** - * @brief reorganizeSeqInfo will reorganize sequenceStartPositions and - * subSequenceStartPositions into a 2 dimensional arrary: reorganizedSeqInfo. - * - * @param seqStartPos: sequenceStartPositions of an Argument. - * @param subSeqStartPos: subSequenceStartPositions of an Argument. - * @param the reorganized sequence start position information. - * - * Examples: - * seqStartPos: [0, 4, 15, 20, 28] - * subSeqStartPos: [0, 3, 4, 5, 7, 10, 15, 20, 22, 23, 25, 28] - * reorganizedSeqInfo: - * [ - * [0,3,4], - * [4,5,7,10,15], - * [15,20], - * [20,22,23,25,28] - * ] - */ - static void reorganizeSeqInfo( - const ICpuGpuVectorPtr seqStartPos, - const ICpuGpuVectorPtr subSeqStartPos, - std::vector>& reorganizedSeqInfo); -}; - -} // namespace paddle diff --git a/paddle/parameter/FirstOrderOptimizer.cpp b/paddle/parameter/FirstOrderOptimizer.cpp deleted file mode 100644 index 182e833405e8f8bc3a4c9ffddbf628040f9cceaa..0000000000000000000000000000000000000000 --- a/paddle/parameter/FirstOrderOptimizer.cpp +++ /dev/null @@ -1,330 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "FirstOrderOptimizer.h" -#include "paddle/math/TrainingAlgorithmOp.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Util.h" - -#include - -DEFINE_bool(log_clipping, false, "enable log clipping or not"); - -namespace paddle { - -SparseMomentumParameterOptimizer::SparseMomentumParameterOptimizer( - const OptimizationConfig& optConfig) - : ParameterOptimizer(optConfig) { - addParameterType(PARAMETER_MOMENTUM); - addParameterType(PARAMETER_MOMENTUM_UT); - addParameterType(PARAMETER_MOMENTUM_VT); - alpha_ = 1; - beta_ = 1; - tau_ = -1; - threshold_ = 1e+06; -} - -void SparseMomentumParameterOptimizer::init(size_t numRows, - const ParameterConfig* config) { - isParameterSparse_ = numRows != 0; - t0Vec_.resize(numRows); - t0Vec_.assign(t0Vec_.size(), 0); - timer_ = 0; - momentum_ = config->momentum(); - decayRate_ = config->decay_rate(); - gamma_ = config->learning_rate(); -} - -void SparseMomentumParameterOptimizer::startBatch(int64_t numSamplesProcessed) { - learningRate_ = calcLearningRate(numSamplesProcessed, pass_); - if (isParameterSparse_) { - tau_ = tau_ + beta_ / alpha_; - alpha_ = alpha_ / momentum_; - beta_ = beta_ / (1 + decayRate_ * gamma_ * learningRate_); - } -} - -void SparseMomentumParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& paraConfig, - size_t sparseId) const { - if (sparseId != -1LU) { - CHECK_LT(sparseId, t0Vec_.size()); - if (t0Vec_[sparseId] == 0) { - vecs[PARAMETER_MOMENTUM_VT]->assign(*vecs[PARAMETER_VALUE]); - t0Vec_[sparseId] = 1; - } - vecs[PARAMETER_MOMENTUM_UT]->add(*vecs[PARAMETER_GRADIENT], - -alpha_ * gamma_ * learningRate_); - vecs[PARAMETER_MOMENTUM_VT]->add(*vecs[PARAMETER_GRADIENT], - tau_ * alpha_ * gamma_ * learningRate_); - vecs[PARAMETER_VALUE]->add(*vecs[PARAMETER_MOMENTUM_UT], - tau_ / beta_ + 1.0 / alpha_, - *vecs[PARAMETER_MOMENTUM_VT], - 1.0 / beta_); - - } else { - vecs[PARAMETER_VALUE]->sgdUpdate(*vecs[PARAMETER_GRADIENT], - *vecs[PARAMETER_MOMENTUM], - learningRate_ * paraConfig.learning_rate(), - paraConfig.momentum(), - applyDecay_ ? paraConfig.decay_rate() : 0); - } -} - -ParameterOptimizer::TraverseCallback -SparseMomentumParameterOptimizer::needSpecialTraversal( - const ParameterConfig& config) const { - if (alpha_ > threshold_ && isParameterSparse_) { - // Restart to avoid large value multiplication - // 1. \alpha = 1, \beta = 1, \tau = 0 - // 2. Note that \tau * u_t + v_t = \beta \theta_t, therefore: - // u_t should be rescaled to u_t/alpha_ - // v_t should be reset to \theta_t - return [this](const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) { - vecs[PARAMETER_MOMENTUM_UT]->divScalar(alpha_); - vecs[PARAMETER_MOMENTUM_VT]->assign(*vecs[PARAMETER_VALUE]); - }; - } else { - return nullptr; - } -} - -void SparseMomentumParameterOptimizer::finishBatch() { - timer_++; - if (!isParameterSparse_) return; - if (alpha_ > threshold_) { - alpha_ = 1; - beta_ = 1; - tau_ = -1; - } -} - -void AdagradParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - BaseMatrix& value = *vecs[PARAMETER_VALUE]; - BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; - BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; - BaseMatrix& accum_buffer = *vecs[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& accum = *vecs[PARAMETER_GRADIENT_SQURESUM1]; - BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; - - real epsilon = optConfig_.ada_epsilon(); - real learningRate = learningRate_ * config.learning_rate(); - real momentum = config.momentum(); - real decayRate = applyDecay_ ? config.decay_rate() : 0; - - adagradApply(value, - grad, - mom, - accum_buffer, - accum, - lr, - epsilon, - learningRate, - momentum, - decayRate); -} - -ParameterOptimizer::TraverseCallback -AdagradParameterOptimizer::needSpecialTraversal( - const ParameterConfig& config) const { - if (numUpdates_ % kMaxNumAccumulates == 0) { - // Move the sum to a different buffer to avoid loss of precision - // due to too many sums. - return [](const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) { - vecs[PARAMETER_GRADIENT_SQURESUM]->add( - *vecs[PARAMETER_GRADIENT_SQURESUM1]); - vecs[PARAMETER_GRADIENT_SQURESUM1]->zeroMem(); - }; - } else { - return nullptr; - } -} - -void AdaDeltaParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - CHECK(sparseId == -1LU) << "Sparse update is not supported"; - - BaseMatrix& value = *vecs[PARAMETER_VALUE]; - BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; - BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; - BaseMatrix& accum = *vecs[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& accum_update = *vecs[PARAMETER_GRADIENT_SQURESUM1]; - BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; - - real learningRate = learningRate_ * config.learning_rate(); - real momentum = config.momentum(); - real decayRate = applyDecay_ ? config.decay_rate() : 0; - - adadeltaApply(value, - grad, - mom, - accum, - accum_update, - lr, - rou_, - epsilon_, - learningRate, - momentum, - decayRate); -} - -void RMSPropParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - BaseMatrix& value = *vecs[PARAMETER_VALUE]; - BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; - BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; - BaseMatrix& sum = *vecs[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& sum1 = *vecs[PARAMETER_GRADIENT_SQURESUM1]; - BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; - - real accumulatedRou = rou_; - bool firstTime = timer_ == 0; - if (sparseId != -1LU) { - CHECK_LT(sparseId, t0Vec_.size()); - accumulatedRou = std::pow(rou_, timer_ + 1 - t0Vec_[sparseId]); - firstTime = t0Vec_[sparseId] == 0; - t0Vec_[sparseId] = timer_ + 1; - } - - real epsilon = optConfig_.ada_epsilon(); - real learningRate = learningRate_ * config.learning_rate(); - real momentum = config.momentum(); - real decayRate = applyDecay_ ? config.decay_rate() : 0; - - rmspropApply(value, - grad, - mom, - sum, - sum1, - lr, - accumulatedRou, - rou_, - epsilon, - learningRate, - momentum, - decayRate, - firstTime); -} - -void DecayedAdagradParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - BaseMatrix& value = *vecs[PARAMETER_VALUE]; - BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; - BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; - BaseMatrix& sum = *vecs[PARAMETER_GRADIENT_SQURESUM]; - BaseMatrix& lr = *vecs[PARAMETER_LEARNING_RATE]; - - real accumulatedRou = rou_; - bool firstTime = timer_ == 0; - if (sparseId != -1LU) { - CHECK_LT(sparseId, t0Vec_.size()); - accumulatedRou = std::pow(rou_, timer_ + 1 - t0Vec_[sparseId]); - firstTime = t0Vec_[sparseId] == 0; - t0Vec_[sparseId] = timer_ + 1; - } - - real epsilon = optConfig_.ada_epsilon(); - real learningRate = learningRate_ * config.learning_rate(); - real momentum = config.momentum(); - real decayRate = applyDecay_ ? config.decay_rate() : 0; - - decayedAdagradApply(value, - grad, - mom, - sum, - lr, - accumulatedRou, - rou_, - epsilon, - learningRate, - momentum, - decayRate, - firstTime); -} - -void AdamParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - CHECK(sparseId == -1UL) << "Sparse update is not supported"; - - real beta1_power = std::pow(beta1_, step_); - real beta2_power = std::pow(beta2_, step_); - real learningRate = config.learning_rate() * learningRate_; - - BaseMatrix& value = *vecs[PARAMETER_VALUE]; - BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; - BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; - BaseMatrix& v = *vecs[PARAMETER_SECOND_MOMENTUM]; - - adamApply(value, - grad, - mom, - v, - beta1_, - beta2_, - beta1_power, - beta2_power, - epsilon_, - learningRate); -} - -void AdamaxParameterOptimizer::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - CHECK(sparseId == -1UL) << "Sparse update is not supported"; - real learningRate = config.learning_rate() * learningRate_; - - BaseMatrix& value = *vecs[PARAMETER_VALUE]; - BaseMatrix& grad = *vecs[PARAMETER_GRADIENT]; - BaseMatrix& mom = *vecs[PARAMETER_MOMENTUM]; - BaseMatrix& u = *vecs[PARAMETER_WEIGHTED_INFINITY_NORM]; - - adamaxApply(value, grad, mom, u, beta1_, beta2_, step_, learningRate); -} - -void OptimizerWithGradientClipping::update(const VectorPtr vecs[], - const ParameterConfig& config, - size_t sparseId) const { - real globalThreshold = optConfig_.gradient_clipping_threshold(); - real localThreshold = config.gradient_clipping_threshold(); - - // Use local gradient clipping threshold if it's enabled, - // otherwise using the global one. - real threshold = localThreshold > 0.0f ? localThreshold : globalThreshold; - std::string field = localThreshold > 0.0f ? "local" : "global"; - - real maxAbsGrad = vecs[PARAMETER_GRADIENT]->getAbsMax(); - if (maxAbsGrad > threshold) { - if (FLAGS_log_clipping) { - real avgAbsGrad = vecs[PARAMETER_GRADIENT]->getAbsSum() / - vecs[PARAMETER_GRADIENT]->getSize(); - LOG(INFO) << "parameter=" << config.name() << " need clipping by " - << field << " threshold=" << threshold - << ", max grad=" << maxAbsGrad << ", avg grad=" << avgAbsGrad; - } - vecs[PARAMETER_GRADIENT]->clip(-threshold, threshold); - } - optimizer_->update(vecs, config, sparseId); -} - -} // namespace paddle diff --git a/paddle/parameter/LearningRateScheduler.cpp b/paddle/parameter/LearningRateScheduler.cpp deleted file mode 100644 index d57d2189a45dc8cbcea7a8a5f25c5ec7ac71cca3..0000000000000000000000000000000000000000 --- a/paddle/parameter/LearningRateScheduler.cpp +++ /dev/null @@ -1,173 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "LearningRateScheduler.h" -#include "paddle/utils/StringUtil.h" - -namespace paddle { - -ClassRegistrar - LearningRateScheduler::registrar_; - -LearningRateScheduler* LearningRateScheduler::create( - const OptimizationConfig& config) { - return registrar_.createByType(config.learning_rate_schedule(), config); -} - -// LRS stands for LearningRateScheduler - -class BaseLRS : public LearningRateScheduler { - public: - explicit BaseLRS(const OptimizationConfig& config) - : learningRate_(config.learning_rate()), - a_(config.learning_rate_decay_a()), - b_(config.learning_rate_decay_b()) {} - - protected: - real learningRate_; - real a_; - real b_; -}; - -class ConstLRS : public BaseLRS { - public: - explicit ConstLRS(const OptimizationConfig& config) : BaseLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - return learningRate_; - } -}; -REGISTER_LEARNING_RATE_SCHEDULER(constant, ConstLRS); - -class PolyLRS : public BaseLRS { - public: - explicit PolyLRS(const OptimizationConfig& config) : BaseLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - return learningRate_ * pow(1.0 + a_ * numSamplesProcessed, -b_); - } -}; -REGISTER_LEARNING_RATE_SCHEDULER(poly, PolyLRS); - -class CaffePolyLRS : public BaseLRS { - public: - explicit CaffePolyLRS(const OptimizationConfig& config) : BaseLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - if (numSamplesProcessed > a_) { - LOG_FIRST_N(WARNING, 1) - << "Using caffe_poly learning rate schedule, " - << "learning rate hits ZERO when " - << "numSamplesProcessed > config.learning_rate_decay_b(), " - << "training is over and you can stop it. " - << "See common/LearningRateScheduler.cpp for more info."; - return 0; - } else { - return learningRate_ * pow(1.0 - numSamplesProcessed / a_, b_); - } - } -}; -REGISTER_LEARNING_RATE_SCHEDULER(caffe_poly, CaffePolyLRS); - -class ExpLRS : public BaseLRS { - public: - explicit ExpLRS(const OptimizationConfig& config) : BaseLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - double decayRatio = (double)numSamplesProcessed / b_; - return learningRate_ * pow(a_, decayRatio); - } -}; -REGISTER_LEARNING_RATE_SCHEDULER(exp, ExpLRS); - -class DiscreteExpLRS : public BaseLRS { - public: - explicit DiscreteExpLRS(const OptimizationConfig& config) : BaseLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - int numDecays = floor(numSamplesProcessed / b_); - return learningRate_ * pow(a_, numDecays); - } -}; -REGISTER_LEARNING_RATE_SCHEDULER(discexp, DiscreteExpLRS); - -class LinearLRS : public BaseLRS { - public: - explicit LinearLRS(const OptimizationConfig& config) : BaseLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - return std::max(learningRate_ - a_ * numSamplesProcessed, b_); - } -}; -REGISTER_LEARNING_RATE_SCHEDULER(linear, LinearLRS); - -/* - specify learning rate through - learning_rate_args = 'seg0:rate0,seg1:rate1,...,segK:rateK' - if seg_{i-1} <= numSamples <= seg_i, - then learning_rate = learning_rate_base * rate_i -*/ -class ManualLRS : public BaseLRS { - public: - explicit ManualLRS(const OptimizationConfig& config) - : BaseLRS(config), currentSegment_(0), lastNum_(0) { - std::vector pieces; - str::split(config.learning_rate_args(), ',', &pieces); - rates_.reserve(pieces.size()); - std::string s1, s2; - - for (auto& piece : pieces) { - auto pos = piece.find(':'); - CHECK(pos != std::string::npos) << "Wrong format for learning_rate_args: " - << config.learning_rate_args(); - segments_.push_back(str::to(piece.substr(0, pos))); - rates_.push_back(str::to(piece.substr(pos + 1))); - } - } - - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - return calc(numSamplesProcessed); - } - - real calc(int64_t num) { - // We assume that num never decreases. - CHECK_LE(lastNum_, num); - lastNum_ = num; - while (currentSegment_ < rates_.size()) { - if (num <= segments_[currentSegment_]) { - return learningRate_ * rates_[currentSegment_]; - } - ++currentSegment_; - if (currentSegment_ < rates_.size()) { - LOG(INFO) << " learning_rate changes to " - << learningRate_ * rates_[currentSegment_]; - } - } - return learningRate_ * rates_.back(); - } - - protected: - std::vector rates_; - std::vector segments_; - size_t currentSegment_; - int64_t lastNum_; -}; - -REGISTER_LEARNING_RATE_SCHEDULER(manual, ManualLRS); - -class PassManualLRS : public ManualLRS { - public: - explicit PassManualLRS(const OptimizationConfig& config) - : ManualLRS(config) {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { - return calc(pass); - } -}; - -REGISTER_LEARNING_RATE_SCHEDULER(pass_manual, PassManualLRS); -} // namespace paddle diff --git a/paddle/parameter/LearningRateScheduler.h b/paddle/parameter/LearningRateScheduler.h deleted file mode 100644 index 3fad97040248dcf8a22988c38153df31f267ed37..0000000000000000000000000000000000000000 --- a/paddle/parameter/LearningRateScheduler.h +++ /dev/null @@ -1,37 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "TrainerConfig.pb.h" -#include "paddle/utils/ClassRegistrar.h" - -namespace paddle { -// NOLINTNEXTLINES_4 -#define REGISTER_LEARNING_RATE_SCHEDULER(__type_name, __class_name) \ - static InitFunction __reg_type_##__type_name([]() { \ - LearningRateScheduler::registrar_.registerClass<__class_name>( \ - #__type_name); \ - }) - -class LearningRateScheduler { - public: - static LearningRateScheduler* create(const OptimizationConfig& config); - virtual ~LearningRateScheduler() {} - virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) = 0; - - static ClassRegistrar registrar_; -}; - -} // namespace paddle diff --git a/paddle/parameter/Parameter.cpp b/paddle/parameter/Parameter.cpp deleted file mode 100644 index 0e6ea90f3d582e843c62bda000313eb71289d5b4..0000000000000000000000000000000000000000 --- a/paddle/parameter/Parameter.cpp +++ /dev/null @@ -1,425 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Parameter.h" -#include -#include -#include "AverageOptimizer.h" -#include "FirstOrderOptimizer.h" -#include "OptimizerFunctions.h" -#include "OptimizerWithRegularizer.h" -#include "ParameterUpdateFunctions.h" -#include "ThreadLocalBuffer.h" -#include "hl_gpu.h" -#include "paddle/math/CpuSparseMatrix.h" -#include "paddle/math/MathUtils.h" -#include "paddle/math/SparseRowMatrix.h" -#include "paddle/utils/Logging.h" - -DEFINE_int32(enable_grad_share, - (100 * 1024 * 1024), - "threshold for enable gradient parameter share for batch " - "multi-cpu training"); -DEFINE_int32( - grad_share_block_num, - 64, - "block number of gradient parameter share for batch multi-cpu training"); - -namespace paddle { - -const std::string Parameter::kMissParameterFail = "fail"; -const std::string Parameter::kMissParameterRand = "rand"; -const std::string Parameter::kMissParameterZero = "zero"; - -Parameter::Parameter(const ParameterConfig& config, bool useGpu, bool doInit) - : config_(config), - useGpu_(useGpu), - deviceId_(-1), - sharedCount_(0), - updateCounter_(0), - updated_(false), - headerFormat_(PARAM_FORMAT_ORIGINAL) { - setID(-1); /* capture uninitialized id */ - if (useGpu_ && FLAGS_parallel_nn) { - /* gpu environment is specified by device property */ - deviceId_ = config_.device(); - if (deviceId_ < 0) { - useGpu_ = false; - } - } - - if (doInit) { - initialize(); - } - - for (int i = 0; i < config.update_hooks_size(); ++i) { - this->updaterHooks_.push_back(IParameterUpdaterHook::create(config, i)); - } -} - -void Parameter::initialize() { - SetDevice device(deviceId_); - - bufs_[PARAMETER_VALUE] = - Vector::createParallelVector(config_.size(), useGpu_); - bufs_[PARAMETER_VALUE]->zeroMem(); - - if (config_.is_sparse()) { - enableSparseParameter(); - } - - if (!isStatic()) { - bufs_[PARAMETER_GRADIENT] = - Vector::createParallelVector(config_.size(), useGpu_); - bufs_[PARAMETER_MOMENTUM] = - Vector::createParallelVector(config_.size(), useGpu_); - - bufs_[PARAMETER_GRADIENT]->zeroMem(); - bufs_[PARAMETER_MOMENTUM]->zeroMem(); - } -} - -void Parameter::randomize(const VectorPtr& value, - const ParameterConfig& config) { - if (PARAMETER_INIT_UNIFORM == config.initial_strategy()) { - // initialize the parameter as uniform distribution - real initial_min = config.initial_mean() - config.initial_std(); - real initial_max = config.initial_mean() + config.initial_std(); - value->uniform(initial_min, initial_max); - VLOG(1) << config.name() << ": initial_min=" << initial_min - << ", initial_max=" << initial_max; - } else if (PARAMETER_INIT_NORMAL == config.initial_strategy()) { - /* Initialize the parameters randomly */ - value->randnorm(config.initial_mean(), config.initial_std()); - VLOG(1) << config.name() << ": initial_mean=" << config.initial_mean() - << ", initial_std=" << config.initial_std(); - } else { - LOG(FATAL) << "not supported initial_strategy: " - << config.initial_strategy(); - } -} - -void Parameter::randomize() { - if (!bufs_[PARAMETER_VALUE]) return; - SetDevice device(deviceId_); - Parameter::randomize(bufs_[PARAMETER_VALUE], config_); - - if (config_.is_sparse()) { - if (format_ == SPARSE_CSC) { - sparseRand(intBufs_[PARAMETER_COLS]->getData(), - intBufs_[PARAMETER_ROWS]->getData(), - config_.size(), - config_.dims(1) + 1, - config_.dims(0), - useGpu_); - } else { - sparseRand(intBufs_[PARAMETER_ROWS]->getData(), - intBufs_[PARAMETER_COLS]->getData(), - config_.size(), - config_.dims(0) + 1, - config_.dims(1), - useGpu_); - } - } - setValueUpdated(); -} - -void Parameter::zeroMem() { - if (!bufs_[PARAMETER_VALUE]) return; - bufs_[PARAMETER_VALUE]->zeroMem(); - setValueUpdated(); - LOG(INFO) << getName() << " set to 0"; -} - -bool Parameter::isGradShared(size_t* blockNum) { - if (!useGpu_ && !isStatic() && FLAGS_enable_grad_share > 0 && - !isGradSparseUpdate() && - this->getSize() > (size_t)FLAGS_enable_grad_share) { - if (blockNum) { - *blockNum = (size_t)FLAGS_grad_share_block_num; - } - return true; - } - return false; -} - -bool Parameter::isValueShared() { - return !useGpu_ && config_.is_shared() && FLAGS_trainer_count > 1; -} - -bool Parameter::isGradSparseUpdate() const { - return !useGpu_ && !isStatic() && - (config_.sparse_update() || config_.sparse_remote_update()); -} - -void Parameter::setMat(ParameterType pType, int matType) { - CHECK(!mats_[pType]); - - if (config_.dims_size() == 0 && matType == MAT_NORMAL) { - return; - } - - CHECK_EQ((size_t)config_.dims_size(), 2LU); - size_t height = config_.dims(0); - size_t width = config_.dims(1); - if (matType == MAT_NORMAL) { - if (!config_.is_sparse()) { - CHECK_EQ(height * width, bufs_[pType]->getSize()); - mats_[pType] = - Matrix::create(bufs_[pType]->getMemoryHandle(), height, width); - } else { - size_t size = bufs_[pType]->getSize(); - CHECK_GE(height * width, size); - if (format_ == SPARSE_CSR) { - CHECK_EQ(height + 1, intBufs_[PARAMETER_ROWS]->getSize()); - CHECK_EQ(size, intBufs_[PARAMETER_COLS]->getSize()); - } else { - CHECK_EQ(width + 1, intBufs_[PARAMETER_COLS]->getSize()); - CHECK_EQ(size, intBufs_[PARAMETER_ROWS]->getSize()); - } - mats_[pType] = - Matrix::createSparseMatrix(bufs_[pType]->getData(), - intBufs_[PARAMETER_ROWS]->getData(), - intBufs_[PARAMETER_COLS]->getData(), - height, - width, - bufs_[pType]->getSize(), - FLOAT_VALUE, - format_, - false, - useGpu_); - } - } -#ifndef PADDLE_MOBILE_INFERENCE - // NOLINTNEXTLINE - else if (matType == MAT_NORMAL_SHARED) { - CHECK_EQ(height * width, bufs_[pType]->getSize()); - size_t blockNum = 0; - CHECK(isGradShared(&blockNum)); - mats_[pType] = std::make_shared( - blockNum, - std::dynamic_pointer_cast( - bufs_[pType]->getMemoryHandle()), - height, - width); - } else if (matType == MAT_VALUE_SHARED) { - CHECK_EQ(height * width, bufs_[pType]->getSize()); - mats_[pType] = std::make_shared( - std::dynamic_pointer_cast( - bufs_[pType]->getMemoryHandle()), - height, - width); - } else if (matType == MAT_SPARSE_ROW_IDS) { - CHECK_EQ(height * width, bufs_[pType]->getSize()); - mats_[pType] = std::make_shared( - std::dynamic_pointer_cast( - bufs_[pType]->getMemoryHandle()), - height, - width); - } else if (matType == MAT_SPARSE_ROW) { - auto valueMat = - std::dynamic_pointer_cast(mats_[PARAMETER_VALUE]); - SparseRowCpuMatrix::IndexDictPtr indexDict(nullptr); - if (pType != PARAMETER_VALUE) { - CHECK(valueMat) << "The matrix for PARAMETER_VALUE must be set " - << " and its type must be MAT_SPARSE_ROW," - << " MAT_SPARSE_ROW_PREFETCH or MAT_CACHE_ROW"; - indexDict = valueMat->getIndexDictHandle(); - } - auto mat = - std::make_shared(nullptr, - height, - width, - // grad share index with value - indexDict); - mats_[pType] = mat; - } else if (matType == MAT_CACHE_ROW) { - CHECK(isGradSparseUpdate()); - auto mat = std::make_shared(height, width); - mats_[pType] = mat; - } else if (matType == MAT_SPARSE_ROW_PREFETCH_FULL_SIZE || - matType == MAT_SPARSE_ROW_PREFETCH) { - auto mat = std::make_shared( - bufs_[pType] ? std::dynamic_pointer_cast( - bufs_[pType]->getMemoryHandle()) - : nullptr, - height, - width, - nullptr, // indexDictHandle - getGlobalSyncThreadPool()); - mats_[pType] = mat; - } else if (matType == MAT_SPARSE_ROW_AUTO_GROW) { - CHECK(isGradSparseUpdate()); - mats_[pType] = std::make_shared(height, width); - } -#endif - // NOLINTNEXTLINE - else { - LOG(FATAL) << "Unsupported mat type" << matType; - } -} - -void Parameter::incUpdate(const UpdateCallback& callback) { - // Static parameter is fixed, and does not need to be updated - if (isStatic()) { - return; - } - - ++updateCounter_; - if (isUpdatable()) { - if (callback) callback(this); - clearUpdate(); - } -} - -bool Parameter::save(const std::string& filename) const { - std::ofstream fs(filename, std::ios_base::binary); - CHECK(fs) << "Fail to open " << filename; - return save(fs); -} - -bool Parameter::save(std::ostream& s) const { - CpuVector vec(*bufs_[PARAMETER_VALUE].get()); - Header header; - header.format = headerFormat_; - header.valueSize = sizeof(real); - header.size = getSize(); - - CHECK_EQ(header.size, vec.getSize()); - - CHECK(s.write(reinterpret_cast(&header), sizeof(header))) - << "Fail to write parameter " << getName(); - - CHECK(s.write(reinterpret_cast(vec.getData()), - header.size * sizeof(real))) - << "Fail to write parameter " << getName(); - if (config_.is_sparse()) { - CpuIVector rows(*intBufs_[PARAMETER_ROWS].get()); - CpuIVector cols(*intBufs_[PARAMETER_COLS].get()); - CHECK(s.write(reinterpret_cast(rows.getData()), - rows.getSize() * sizeof(int))) - << "Fail to write parameter " << getName(); - CHECK(s.write(reinterpret_cast(cols.getData()), - cols.getSize() * sizeof(int))) - << "Fail to write parameter " << getName(); - } - - return true; -} - -/** - * Load parameter value from a file - */ -bool Parameter::load(const std::string& filename) { - std::ifstream fs(filename, std::ios_base::binary); - if (!fs) { - LOG(INFO) << "missing parameters [" << filename << "] while loading model."; - if (kMissParameterFail == FLAGS_load_missing_parameter_strategy) { - LOG(FATAL) << getName() << " missing, not allowed."; - return false; - } - if (kMissParameterRand == FLAGS_load_missing_parameter_strategy) { - LOG(INFO) << getName() << " missing, set to random."; - randomize(); - return true; - } - if (kMissParameterZero == FLAGS_load_missing_parameter_strategy) { - LOG(INFO) << getName() << " missing, set to zero."; - zeroMem(); - return true; - } - LOG(FATAL) << "unsupported load_missing_parameter_strategy: " - << FLAGS_load_missing_parameter_strategy; - return false; - } - return load(fs); -} - -bool Parameter::load(std::istream& s) { - CpuVector vec(*bufs_[PARAMETER_VALUE].get()); - Header header; - CHECK(s.read(reinterpret_cast(&header), sizeof(header))) - << "Fail to read parameter " << getName(); - CHECK(isHeaderFormatSupported(header.format)) << "Incorrect format version: " - << header.format; - headerFormat_ = header.format; - CHECK_EQ(header.size, getSize()) - << "The size (" << header.size << ") in the file does not match the size " - << "(" << getSize() << ") of the parameter: " << getName(); - CHECK_EQ(header.valueSize, sizeof(real)) - << "Unsupported valueSize " << header.valueSize << " at: " << getName(); - CHECK(s.read(reinterpret_cast(vec.getData()), - header.size * sizeof(real))); - - auto& tmp = *bufs_[PARAMETER_VALUE].get(); - if (typeid(tmp) == typeid(GpuVector)) { - bufs_[PARAMETER_VALUE]->copyFrom(vec); - } - - if (config_.is_sparse() && config_.need_compact()) { - // load from dense parameter with many zero - CHECK_EQ(config_.dims_size(), 2); - auto height = config_.dims(0); - auto width = config_.dims(1); - auto mat = Matrix::create(vec.getData(), height, width); - CpuSparseMatrix sparseMat(height, - width, - 0, - FLOAT_VALUE, - format_, - /*trans*/ false); - sparseMat.copyFrom(*mat, HPPL_STREAM_DEFAULT); - auto nnz = sparseMat.getElementCnt(); - size_t rowSize = (format_ == SPARSE_CSR) ? height + 1 : nnz; - size_t colSize = (format_ == SPARSE_CSR) ? nnz : width + 1; - - intBufs_[PARAMETER_ROWS]->copyFrom(sparseMat.getRows(), rowSize); - intBufs_[PARAMETER_COLS]->copyFrom(sparseMat.getCols(), colSize); - bufs_[PARAMETER_VALUE]->resize(nnz); // for setMat check - bufs_[PARAMETER_VALUE]->copyFrom(sparseMat.getValue(), nnz); - config_.set_size(nnz); - LOG(INFO) << "compact nnz=" << (1. * nnz / (height * width)) - << " name=" << config_.name(); - } else if (config_.is_sparse()) { - CpuIVector rows(*intBufs_[PARAMETER_ROWS].get()); - CpuIVector cols(*intBufs_[PARAMETER_COLS].get()); - size_t rowSize, colSize; - CHECK_EQ(config_.dims_size(), 2); - if (format_ == SPARSE_CSR) { - rowSize = config_.dims(0) + 1; - colSize = config_.size(); - } else { - rowSize = config_.size(); - colSize = config_.dims(1) + 1; - } - CHECK( - s.read(reinterpret_cast(rows.getData()), rowSize * sizeof(int))); - CHECK( - s.read(reinterpret_cast(cols.getData()), colSize * sizeof(int))); - auto& paramRows = *intBufs_[PARAMETER_ROWS].get(); - if (typeid(paramRows) == typeid(GpuIVector)) { - intBufs_[PARAMETER_ROWS]->copyFrom(rows); - } - auto& paramCols = *intBufs_[PARAMETER_COLS].get(); - if (typeid(paramCols) == typeid(GpuIVector)) { - intBufs_[PARAMETER_COLS]->copyFrom(cols); - } - } - - setValueUpdated(); - - return true; -} - -} // namespace paddle diff --git a/paddle/parameter/Parameter.h b/paddle/parameter/Parameter.h deleted file mode 100644 index ef519bf35a4f051b4477eb04b5eb2c5f0b5e29e8..0000000000000000000000000000000000000000 --- a/paddle/parameter/Parameter.h +++ /dev/null @@ -1,380 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include - -#include -#include -#include - -#include "ParameterConfig.pb.h" -#include "TrainerConfig.pb.h" - -#include "ParameterUpdaterHook.h" -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/ThreadLocal.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -typedef enum { - /// The paddle original basic format - PARAM_FORMAT_ORIGINAL = 0, - - /// See mkldnn_memory_format_t in - /// https://github.com/01org/mkl-dnn/blob/master/include/mkldnn_types.h - /// for a detailed description. - /// 2D weights tensor in the format (output channels, input channels). - PARAM_FORMAT_MKLDNN_OI, - - /// The total format items numbers - PARAM_FORMAT_ITEMS, -} PARAM_FORMAT; - -class SparsePrefetchRowCpuMatrix; - -class Parameter; -typedef std::function UpdateCallback; -typedef std::function ParamInitCallback; - -class Parameter; -typedef std::shared_ptr ParameterPtr; - -class Parameter { - public: - Parameter(const ParameterConfig& config, bool useGpu, bool doInit = true); - const std::string& getName() const { return config_.name(); } - - size_t getSize() const { return config_.size(); } - - bool isFullSize() const { - if (bufs_[PARAMETER_VALUE]) { - return this->getSize() == bufs_[PARAMETER_VALUE]->getSize(); - } - return false; - } - - inline bool useGpu() const { return useGpu_; } - - int getDeviceId() const { return deviceId_; } - - void setDevice(int deviceId) { deviceId_ = deviceId; } - - /// The id ranges from 0 to the_total_number_of_parameters - 1 - size_t getID() const { return config_.para_id(); } - - /// ID is a implict value created until neural network is built. - void setID(size_t id) { config_.set_para_id(id); } - - bool isStatic() const { return config_.is_static(); } - - enum MatType { - MAT_NORMAL, - /// both value and grad are shared - MAT_NORMAL_SHARED, - - /// Now used in BatchNorm in CPU mode - MAT_VALUE_SHARED, - - /// sparse matrix, which has full size parameter - MAT_SPARSE_ROW_IDS, - /// sparse matrix, parameter size scale by sparse rates. - MAT_SPARSE_ROW_AUTO_GROW, - MAT_CACHE_ROW, - MAT_SPARSE_ROW, - - /// sparse matrix for prefetching parameter from pserver - MAT_SPARSE_ROW_PREFETCH, - /// same as above, but parameter has full size for saving parameter in local - MAT_SPARSE_ROW_PREFETCH_FULL_SIZE, - }; - - void enableSparseParameter() { - if (config_.is_sparse()) { - if (config_.format() == "csr") { - size_t height = config_.dims(0); - size_t nnz = config_.size(); - enableIntType(PARAMETER_ROWS, height + 1); - enableIntType(PARAMETER_COLS, nnz); - format_ = SPARSE_CSR; - } else { - size_t width = config_.dims(1); - size_t nnz = config_.size(); - enableIntType(PARAMETER_COLS, width + 1); - enableIntType(PARAMETER_ROWS, nnz); - format_ = SPARSE_CSC; - } - } - } - - /// allocate buffer for the give type - void enableType(ParameterType type, MatType matType = MAT_NORMAL) { - if (bufs_[type] || mats_[type]) { - return; - } - SetDevice device(deviceId_); - if (config_.dims_size() == 2) { - if (matType == MAT_NORMAL || matType == MAT_NORMAL_SHARED || - matType == MAT_SPARSE_ROW_PREFETCH_FULL_SIZE || - matType == MAT_VALUE_SHARED || matType == MAT_SPARSE_ROW_IDS) { - bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_); - bufs_[type]->zeroMem(); - } else { - CHECK(isGradSparseUpdate()); - } - if (config_.is_sparse() && type == PARAMETER_VALUE) { - enableSparseParameter(); - } - setMat(type, matType); - } else { - bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_); - bufs_[type]->zeroMem(); - } - } - - void enableBufType(ParameterType type) { - if (bufs_[type]) return; - bufs_[type] = Vector::createParallelVector(config_.size(), useGpu_); - bufs_[type]->zeroMem(); - } - - void enableIntType(ParameterType type, size_t intStoreSize = 0) { - if (!intBufs_[type]) { - SetDevice device(deviceId_); - size_t size = intStoreSize ? intStoreSize : config_.size(); - intBufs_[type] = IVector::create(size, useGpu_); - intBufs_[type]->zeroMem(); - } - } - - void enableSharedType(ParameterType type, - VectorPtr vec, - MatrixPtr mat = nullptr) { - if (!bufs_[type] && !mats_[type]) { - bufs_[type] = vec; - mats_[type] = mat; - } - } - - /// for batchGradientMachine: blockNum is number of partitions of the matrix. - bool isGradShared(size_t* blockNum = NULL); - - bool isValueShared(); - - // for AsgdSparseGradientMachine & SgdSparseGradientMachine: - // and MultiGradientMachine - bool isGradSparseUpdate() const; - - bool isSparseRemoteUpdate() const { - return config_.sparse_remote_update() && !useGpu(); - } - - const ParameterConfig& getConfig() const { return config_; } - - ParameterConfig& getConfig() { return config_; } - - bool hasType(ParameterType pType) const { - return bufs_[pType] || mats_[pType]; - } - - const VectorPtr& getBuf(ParameterType pType) const { - return this->bufs_[pType]; - } - - const VectorPtr* getBufs() const { return bufs_; } - - const MatrixPtr& getMat(ParameterType pType) const { return mats_[pType]; } - - void setValueUpdated() { updated_ = true; } - - void clearValueUpdated() { updated_ = false; } - - bool isValueUpdated() const { return updated_; } - - /** - * Save parameter value to a file - */ - bool save(const std::string& filename) const; - - /** - * Save parameter to ostream - */ - bool save(std::ostream& s) const; - - /** - * Load parameter value from a file - */ - bool load(const std::string& filename); - - /** - * Load parameter from istream - */ - bool load(std::istream& is); - - void incShared() { sharedCount_++; } - - /** - * After one of the parameter's gradient is merged - * You should call this function to do some additional processing, - */ - void incUpdate(const UpdateCallback& callbacks = NULL); - - void clearGradient() { - auto& mat = getMat(PARAMETER_GRADIENT); - if (mat) { - // zeroMem will also clear rows for SparseRowCpuMatrix - mat->zeroMem(); - } else { - auto& gradBuf = getBuf(PARAMETER_GRADIENT); - if (gradBuf) gradBuf->zeroMem(); - } - } - - void initialize(); - - /** - * Initialize the value according to config_: initial_mean, - * initial_std and initial_strategy. - */ - void randomize(); - static void randomize(const VectorPtr& value, const ParameterConfig& config); - - /// Initialize the value to 0 - void zeroMem(); - - /// file header structure - struct Header { - int32_t format; // = PARAM_FORMAT - uint32_t valueSize; // = sizeof(real) - uint64_t size; // = getSize() - }; - - /** - * @brief Is the header format supported. - */ - static bool isHeaderFormatSupported(int32_t fmt) { - return fmt < PARAM_FORMAT_ITEMS; - } - - /** - * @brief Get the format in header. - */ - int getHeaderFormat() { return headerFormat_; } - - /** - * @brief Set the format in header. - */ - void setHeaderFormat(int32_t fmt) { - CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: " - << fmt; - headerFormat_ = fmt; - } - - /** - * @brief Parameter Update Hook. - * - * The parameter's update hook before ParameterUpdater::updateImpl - * It could modify gradient/momentum/etc here. Such as drop some gradient, - * etc. - */ - void updateHook() { - for (auto& hook : updaterHooks_) { - hook->update(this); - } - } - - /** - * @brief Initialize all updater hook. - * - * This method should be invoked in ParameterUpdater::init() only. - */ - void initHook() { - for (auto& hook : updaterHooks_) { - hook->init(this); - } - } - - protected: - /** - * @brief create matrix to matType. - * - * used by gradient machine which needs specify matrix type, - * instead of creating in weights.cpp. - * - * @note pType should be enabled already. - */ - void setMat(ParameterType pType, int matType); - - bool isUpdatable() { return (updateCounter_ == sharedCount_); } - - void clearUpdate() { updateCounter_ = 0; } - - protected: - ParameterConfig config_; - - bool useGpu_; - - int deviceId_; - - /** - * @brief bufs_ stores parameter value and gradient. - * - * Layer should use bufs_[PARAMETER_VALUE] to form weight matrix for - * calculation and stores gradient to bufs_[PARAMETER_GRADIENT]. - */ - VectorPtr bufs_[NUM_PARAMETER_TYPES]; - - /** - * @brief Weight matrix for bufs_. - * - * It's helpfull when parameter shared by multi-layers. - * Caller should check, if mats exist, do not create it again. - */ - MatrixPtr mats_[NUM_PARAMETER_TYPES]; - - /// Int vectors, used in some User defined parameter types - IVectorPtr intBufs_[NUM_PARAMETER_TYPES]; - - int sharedCount_; - int updateCounter_; - - bool updated_; - SparseFormat format_; - - /// The header format for saving or loading param - int32_t headerFormat_; - - std::vector> updaterHooks_; - - public: - void setSharedCount(int cnt) { sharedCount_ = cnt; } - int getSharedCount() { return sharedCount_; } - - bool isSparse() { return config_.is_sparse(); } - SparseFormat getFormat() { return format_; } - - static const std::string kMissParameterFail; - static const std::string kMissParameterRand; - static const std::string kMissParameterZero; -}; - -typedef std::map ParameterMap; - -} // namespace paddle diff --git a/paddle/parameter/ParameterOptimizer.cpp b/paddle/parameter/ParameterOptimizer.cpp deleted file mode 100644 index 638daa58f1e5f3f416d7f90ad2662a523eaf6741..0000000000000000000000000000000000000000 --- a/paddle/parameter/ParameterOptimizer.cpp +++ /dev/null @@ -1,63 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Logging.h" - -#include - -#include "AverageOptimizer.h" -#include "FirstOrderOptimizer.h" -#include "OptimizerFunctions.h" -#include "OptimizerWithRegularizer.h" -#include "ParameterOptimizer.h" -#include "hl_gpu.h" - -namespace paddle { - -ParameterOptimizer* ParameterOptimizer::create( - const OptimizationConfig& optConfig, bool inPserver) { - if (inPserver && optConfig.num_batches_per_send_parameter() > 1) { - return new AddOptimizer(optConfig); - } - if (optConfig.learning_method() == "momentum") { - return new SgdOptimizer(optConfig); - } - if (optConfig.learning_method() == "torch_momentum") { - return new SgdOptimizer(optConfig); - } - if (optConfig.learning_method() == "adagrad") { - return new AdagradParameterOptimizer(optConfig); - } - if (optConfig.learning_method() == "adadelta") { - return new AdaDeltaParameterOptimizer(optConfig); - } - if (optConfig.learning_method() == "rmsprop") { - return new RMSPropParameterOptimizer(optConfig); - } - if (optConfig.learning_method() == "decayed_adagrad") { - return new DecayedAdagradParameterOptimizer(optConfig); - } - if (optConfig.learning_method() == "adam") { - return new AdamParameterOptimizer(optConfig); - } - if (optConfig.learning_method() == "adamax") { - return new AdamaxParameterOptimizer(optConfig); - } - if (optConfig.learning_method() == "sparse_momentum") { - return new SparseMomentumParameterOptimizer(optConfig); - } - return nullptr; -} - -} // namespace paddle diff --git a/paddle/parameter/ParameterUpdateFunctions.cpp b/paddle/parameter/ParameterUpdateFunctions.cpp deleted file mode 100644 index db1153c2d6430e453d776b92b63152c311771668..0000000000000000000000000000000000000000 --- a/paddle/parameter/ParameterUpdateFunctions.cpp +++ /dev/null @@ -1,300 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Logging.h" -#ifdef __AVX__ -#include -#include -#endif - -#include "ParameterUpdateFunctions.h" - -namespace paddle { - -void sgdUpdateCpu(real learningRate, - real momentum, - real decayRate, - size_t size, - real* value, - const real* grad, - real* momentumVec) { - decayRate *= learningRate; -#ifdef PADDLE_WITH_MKLML -#pragma omp parallel for -#endif - for (size_t i = 0; i < size; ++i) { - momentumVec[i] = momentum * momentumVec[i] - learningRate * grad[i] - - decayRate * value[i]; - value[i] += momentumVec[i]; - } -} - -void sgdUpdate(real learningRate, - real momentum, - real decayRate, - Vector* value, - Vector* grad, - Vector* momentumVec) { - size_t size = value->getSize(); - real* val = value->getData(); - real* grd = grad->getData(); - real* mom = momentumVec->getData(); - if (typeid(*value) == typeid(CpuVector)) { - sgdUpdateCpu(learningRate, momentum, decayRate, size, val, grd, mom); - } else if (typeid(*value) == typeid(GpuVector)) { - value->sgdUpdate(*grad, *momentumVec, learningRate, momentum, decayRate); - } else { - LOG(FATAL) << "Wrong"; - } -} - -void sgdUpdateAvx(float learningRate, - float momentum, - float decayRate, - size_t size, - float* value, - const float* _grad, - float* momentumVec) { -#ifdef __AVX__ - float* grad = const_cast(_grad); // the gradient is not modified - // but when invoke simd functions - // need non-const pointer. - size_t gradientAlign = 0; - size_t gradientAlignHeader = (size_t)grad % sizeof(__m256); - CHECK_EQ(gradientAlignHeader, (size_t)momentumVec % sizeof(__m256)) - << "Gradent buffer didn't align with momentum buffer"; - CHECK_EQ(gradientAlignHeader, (size_t)value % sizeof(__m256)) - << "Gradent buffer didn't align with value buffer"; - if (0 != gradientAlignHeader) { - gradientAlignHeader = sizeof(__m256) - gradientAlignHeader; - gradientAlign = gradientAlignHeader / sizeof(real); - - // handle the unalign buffer - for (size_t i = 0; i < gradientAlign; i++) { - momentumVec[i] = momentum * momentumVec[i] - (learningRate * grad[i]) - - (decayRate * learningRate * value[i]); - value[i] += momentumVec[i]; - } - grad += gradientAlign; - momentumVec += gradientAlign; - value += gradientAlign; - } - - constexpr size_t kParallelNum = 8; - constexpr size_t nStepSize = (sizeof(__m256) / sizeof(real)) * kParallelNum; - size_t cntLoop = (size - gradientAlign) / nStepSize; - size_t cntRem = (size - gradientAlign) % nStepSize; - __m256 gradientTmp[kParallelNum]; - __m256 valueTmp[kParallelNum]; - __m256 lr, mom, dr; - std::function loopFun; - - learningRate *= -1; - lr = _mm256_set_ps(learningRate, - learningRate, - learningRate, - learningRate, - learningRate, - learningRate, - learningRate, - learningRate); - - if (0 != momentum) { - mom = _mm256_set_ps(momentum, - momentum, - momentum, - momentum, - momentum, - momentum, - momentum, - momentum); - } - - decayRate *= learningRate; - if (0 != decayRate) { - dr = _mm256_set_ps(decayRate, - decayRate, - decayRate, - decayRate, - decayRate, - decayRate, - decayRate, - decayRate); - } - - auto gradMulFun = [&](void) { - gradientTmp[0] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad), lr); - gradientTmp[1] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 8), lr); - gradientTmp[2] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 16), lr); - gradientTmp[3] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 24), lr); - gradientTmp[4] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 32), lr); - gradientTmp[5] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 40), lr); - gradientTmp[6] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 48), lr); - gradientTmp[7] = _mm256_mul_ps(*reinterpret_cast<__m256*>(grad + 56), lr); - }; - - auto valueMulFun = [&](void) { - valueTmp[0] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value), dr); - valueTmp[1] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 8), dr); - valueTmp[2] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 16), dr); - valueTmp[3] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 24), dr); - valueTmp[4] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 32), dr); - valueTmp[5] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 40), dr); - valueTmp[6] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 48), dr); - valueTmp[7] = _mm256_mul_ps(*reinterpret_cast<__m256*>(value + 56), dr); - }; - - auto momentumMulFun = [&](void) { - *reinterpret_cast<__m256*>(momentumVec) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec), mom); - *reinterpret_cast<__m256*>(momentumVec + 8) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 8), mom); - *reinterpret_cast<__m256*>(momentumVec + 16) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 16), mom); - *reinterpret_cast<__m256*>(momentumVec + 24) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 24), mom); - *reinterpret_cast<__m256*>(momentumVec + 32) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 32), mom); - *reinterpret_cast<__m256*>(momentumVec + 40) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 40), mom); - *reinterpret_cast<__m256*>(momentumVec + 48) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 48), mom); - *reinterpret_cast<__m256*>(momentumVec + 56) = - _mm256_mul_ps(*reinterpret_cast<__m256*>(momentumVec + 56), mom); - }; - - auto momentumAddGradFun = [&](void) { - *reinterpret_cast<__m256*>(momentumVec) = - _mm256_add_ps(*reinterpret_cast<__m256*>(momentumVec), gradientTmp[0]); - *reinterpret_cast<__m256*>(momentumVec + 8) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 8), gradientTmp[1]); - *reinterpret_cast<__m256*>(momentumVec + 16) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 16), gradientTmp[2]); - *reinterpret_cast<__m256*>(momentumVec + 24) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 24), gradientTmp[3]); - *reinterpret_cast<__m256*>(momentumVec + 32) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 32), gradientTmp[4]); - *reinterpret_cast<__m256*>(momentumVec + 40) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 40), gradientTmp[5]); - *reinterpret_cast<__m256*>(momentumVec + 48) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 48), gradientTmp[6]); - *reinterpret_cast<__m256*>(momentumVec + 56) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 56), gradientTmp[7]); - }; - - auto momentumZeroFun = [&](void) { - *reinterpret_cast<__m256*>(momentumVec) = gradientTmp[0]; - *reinterpret_cast<__m256*>(momentumVec + 8) = gradientTmp[1]; - *reinterpret_cast<__m256*>(momentumVec + 16) = gradientTmp[2]; - *reinterpret_cast<__m256*>(momentumVec + 24) = gradientTmp[3]; - *reinterpret_cast<__m256*>(momentumVec + 32) = gradientTmp[4]; - *reinterpret_cast<__m256*>(momentumVec + 40) = gradientTmp[5]; - *reinterpret_cast<__m256*>(momentumVec + 48) = gradientTmp[6]; - *reinterpret_cast<__m256*>(momentumVec + 56) = gradientTmp[7]; - }; - - auto momentumAddValueFun = [&](void) { - *reinterpret_cast<__m256*>(momentumVec) = - _mm256_add_ps(*reinterpret_cast<__m256*>(momentumVec), valueTmp[0]); - *reinterpret_cast<__m256*>(momentumVec + 8) = - _mm256_add_ps(*reinterpret_cast<__m256*>(momentumVec + 8), valueTmp[1]); - *reinterpret_cast<__m256*>(momentumVec + 16) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 16), valueTmp[2]); - *reinterpret_cast<__m256*>(momentumVec + 24) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 24), valueTmp[3]); - *reinterpret_cast<__m256*>(momentumVec + 32) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 32), valueTmp[4]); - *reinterpret_cast<__m256*>(momentumVec + 40) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 40), valueTmp[5]); - *reinterpret_cast<__m256*>(momentumVec + 48) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 48), valueTmp[6]); - *reinterpret_cast<__m256*>(momentumVec + 56) = _mm256_add_ps( - *reinterpret_cast<__m256*>(momentumVec + 56), valueTmp[7]); - }; - - auto valueAddMomentumFun = [&](void) { - *reinterpret_cast<__m256*>(value) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value), - *reinterpret_cast<__m256*>(momentumVec)); - *reinterpret_cast<__m256*>(value + 8) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 8), - *reinterpret_cast<__m256*>(momentumVec + 8)); - *reinterpret_cast<__m256*>(value + 16) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 16), - *reinterpret_cast<__m256*>(momentumVec + 16)); - *reinterpret_cast<__m256*>(value + 24) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 24), - *reinterpret_cast<__m256*>(momentumVec + 24)); - *reinterpret_cast<__m256*>(value + 32) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 32), - *reinterpret_cast<__m256*>(momentumVec + 32)); - *reinterpret_cast<__m256*>(value + 40) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 40), - *reinterpret_cast<__m256*>(momentumVec + 40)); - *reinterpret_cast<__m256*>(value + 48) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 48), - *reinterpret_cast<__m256*>(momentumVec + 48)); - *reinterpret_cast<__m256*>(value + 56) = - _mm256_add_ps(*reinterpret_cast<__m256*>(value + 56), - *reinterpret_cast<__m256*>(momentumVec + 56)); - }; - - if (0 == decayRate && 0 == momentum) { - loopFun = [&](void) { - gradMulFun(); - momentumZeroFun(); - valueAddMomentumFun(); - }; - } else if (0 == decayRate && 0 != momentum) { - loopFun = [&](void) { - gradMulFun(); - momentumMulFun(); - momentumAddGradFun(); - valueAddMomentumFun(); - }; - } else if (0 != decayRate && 0 == momentum) { - loopFun = [&](void) { - gradMulFun(); - valueMulFun(); - momentumZeroFun(); - momentumAddValueFun(); - valueAddMomentumFun(); - }; - } else if (0 != decayRate && 0 != momentum) { - loopFun = [&](void) { - gradMulFun(); - valueMulFun(); - momentumMulFun(); - momentumAddGradFun(); - momentumAddValueFun(); - valueAddMomentumFun(); - }; - } - - for (size_t i = 0; i < cntLoop; i++) { - loopFun(); - grad += nStepSize; - momentumVec += nStepSize; - value += nStepSize; - } - - for (size_t i = 0; i < cntRem; i++) { - momentumVec[i] = momentum * momentumVec[i] + (learningRate * grad[i]) + - (decayRate * value[i]); - value[i] += momentumVec[i]; - } -#endif -} - -} // namespace paddle diff --git a/paddle/parameter/ParameterUpdateFunctions.h b/paddle/parameter/ParameterUpdateFunctions.h deleted file mode 100644 index 7434baa2d3d6297cc6d8d99b46cff516d6ac49f9..0000000000000000000000000000000000000000 --- a/paddle/parameter/ParameterUpdateFunctions.h +++ /dev/null @@ -1,56 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/math/Vector.h" -#include "paddle/utils/Common.h" - -namespace paddle { - -/** - * Performs the following operations. - * - * momentumVec = momentum * momentumVec - * - learningRate * grad - * - learningRate * decayRate * value - * - * value = value + momentumVec - * momentum = 0 or decayRate = 0 are specially handled to avoid unnecessary - * computation. - */ -void sgdUpdate(real learningRate, - real momentum, - real decayRate, - Vector* value, - Vector* grad, - Vector* momentumVec); - -void sgdUpdateCpu(real learningRate, - real momentum, - real decayRate, - size_t size, - real* value, - const real* grad, - real* momentumVec); - -void sgdUpdateAvx(float learningRate, - float momentum, - float decayRate, - size_t size, - float* value, - const float* grad, - float* momentumVec); - -} // namespace paddle diff --git a/paddle/parameter/ParameterUpdaterBase.cpp b/paddle/parameter/ParameterUpdaterBase.cpp deleted file mode 100644 index 7815856b45d93406597b332469de1c57a7781da5..0000000000000000000000000000000000000000 --- a/paddle/parameter/ParameterUpdaterBase.cpp +++ /dev/null @@ -1,41 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ParameterUpdaterBase.h" -#include -#include "hl_gpu.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -void ParameterUpdater::init(const std::vector& parameters) { - parameters_ = parameters; - for (ParameterType type : getParameterTypes()) { - for (auto& para : parameters) { - para->enableType(type); - } - } - for (size_t pid = 0; pid < parameters_.size(); ++pid) { - nonStaticParaIDMap_.insert( - std::pair(parameters_[pid]->getID(), pid)); - } - - for (auto& para : parameters) { - if (!para->isStatic()) { - para->initHook(); - } - } -} - -} // namespace paddle diff --git a/paddle/parameter/ParameterUpdaterHook.cpp b/paddle/parameter/ParameterUpdaterHook.cpp deleted file mode 100644 index 989185b66a5b7785bb0572fba59a72adeef9797b..0000000000000000000000000000000000000000 --- a/paddle/parameter/ParameterUpdaterHook.cpp +++ /dev/null @@ -1,155 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ParameterUpdaterHook.h" - -#include -#include -#include -#include -#include -#include -#include - -#include "paddle/math/Vector.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -/** - * The static pruning hook - * Static means user specify a sparsity_ratio before training started, and the - * network will prune the parameters based on the sparsity_ratio. More details - * can be found https://arxiv.org/pdf/1506.02626.pdf. - */ - -class StaticPruningHook : public IParameterUpdaterHook { - public: - explicit StaticPruningHook(const ParameterUpdaterHookConfig &hookConfig) - : initCount_(0) { - sparsityRatio_ = hookConfig.sparsity_ratio(); - } - - static bool sortPairAscend(const std::pair &pair1, - const std::pair &pair2) { - return pair1.first > pair2.first; - } - - void update(Parameter *para) { - updateThreadChecker_.check(); - auto &vec = para->getBuf(PARAMETER_GRADIENT); - if (vec) { - vec->dotMul(*maskVec_); - } - } - - void generateMask(Parameter *para) { - VectorPtr maskTemp = Vector::create(para->getSize(), false); - maskTemp->zeroMem(); - real *maskTempData = maskTemp->getData(); - size_t nonZeroNum = para->getSize() * (1 - sparsityRatio_); - - VectorPtr paraVec = para->getBuf(PARAMETER_VALUE); - VectorPtr paraCpuCopy = Vector::create(para->getSize(), false); - - paraCpuCopy->copyFrom(*paraVec); - std::vector> param; - - for (size_t i = 0; i < para->getSize(); i++) - param.push_back(std::make_pair(fabs(paraCpuCopy->getData()[i]), i)); - - std::partial_sort( - param.begin(), param.begin() + nonZeroNum, param.end(), sortPairAscend); - for (size_t i = 0; i < nonZeroNum; i++) maskTempData[param[i].second] = 1.0; - - // Currently just use a mask vector for hack. - if (para->useGpu()) { - maskVec_ = Vector::create(para->getSize(), para->useGpu()); - maskVec_->copyFrom(*maskTemp); - } else { - maskVec_ = maskTemp; - } - } - - void init(Parameter *para) { - generateMask(para); - size_t initCount = this->initCount_.fetch_add(1); - CHECK_EQ(initCount, 0UL) << "Currently the StaticPruningHook must invoke " - "in same ParamterUpdater"; - VLOG(3) << "Initialize Parameter " << para; - SetDevice device(para->getDeviceId()); - - auto ¶Vec = para->getBuf(PARAMETER_VALUE); - paraVec->dotMul(*maskVec_); - } - - private: - SameThreadChecker updateThreadChecker_; - std::atomic initCount_; - VectorPtr maskVec_; - real sparsityRatio_; -}; - -IParameterUpdaterHook::IParameterUpdaterHook() {} - -IParameterUpdaterHook::~IParameterUpdaterHook() {} - -/** - * A Hasher used by g_hooks. - * - * Use the independent hasher intendedly. There is a hasher in PServer for hash - * ParameterBlock. But not to use same hasher to reduce dependency. - * - * May be extracted to Util.h to unify the hasher. - */ -class StringIntPairHasher { - public: - size_t operator()(const std::pair &k) const { - return intHasher_(strHasher_(k.first) + k.second); - } - - private: - std::hash strHasher_; - std::hash intHasher_; -}; - -static WeakKVCache, - IParameterUpdaterHook, - StringIntPairHasher> - g_hookCache_; - -/** - * ParameterUpdaterHook actually factory method. - */ -static IParameterUpdaterHook *createImpl( - const ParameterUpdaterHookConfig &config) { - auto &type = config.type(); - if (type == "pruning") { - return new StaticPruningHook(config); - } - - LOG(FATAL) << "Unknown Hook type: " << type; - return nullptr; -} - -std::shared_ptr IParameterUpdaterHook::create( - const ParameterConfig ¶mConfig, int idx) { - std::pair key = {paramConfig.name(), idx}; - return g_hookCache_.get( - key, [&] { return createImpl(paramConfig.update_hooks(idx)); }); -} - -} // namespace paddle diff --git a/paddle/parameter/Regularizer.cpp b/paddle/parameter/Regularizer.cpp deleted file mode 100644 index d223fd2df679af1e983e84f48a4d3b0715ce1569..0000000000000000000000000000000000000000 --- a/paddle/parameter/Regularizer.cpp +++ /dev/null @@ -1,54 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Regularizer.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -Regularizer* Regularizer::get(const std::vector& types, - const ParameterConfig& paraConfig) { - bool useLearningRateVec = - std::find(types.begin(), types.end(), PARAMETER_LEARNING_RATE) != - types.end(); - if (paraConfig.decay_rate_l1() > 0.0f && - paraConfig.decay_rate() > 0.0f) { // use L1 and L2 - if (useLearningRateVec) { - static L1L2LrRegularizer regularizer_; - return ®ularizer_; - } - static L1L2Regularizer regularizer_; - return ®ularizer_; - } - if (paraConfig.decay_rate_l1() > 0.0f) { // use L1 only - if (useLearningRateVec) { - static L1LrRegularizer regularizer_; - return ®ularizer_; - } - static L1Regularizer regularizer_; - return ®ularizer_; - } - if (paraConfig.decay_rate() > 0.0f) { // use L2 only - if (useLearningRateVec) { - static L2LrRegularizer regularizer_; - return ®ularizer_; - } - static L2Regularizer regularizer_; - return ®ularizer_; - } - return nullptr; -} - -} // namespace paddle diff --git a/paddle/parameter/ThreadLocalBuffer.h b/paddle/parameter/ThreadLocalBuffer.h deleted file mode 100644 index 07c96e59d0bc0a58ce9956a54e7de359896e5618..0000000000000000000000000000000000000000 --- a/paddle/parameter/ThreadLocalBuffer.h +++ /dev/null @@ -1,22 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "paddle/math/Vector.h" - -namespace paddle { -namespace parameter { -extern VectorPtr* getThreadLocalBuffer(); -} // namespace parameter -} // namespace paddle diff --git a/paddle/parameter/Weight.cpp b/paddle/parameter/Weight.cpp deleted file mode 100644 index ba4ddce69fb9c2ad0fa937efca5ba470247978e4..0000000000000000000000000000000000000000 --- a/paddle/parameter/Weight.cpp +++ /dev/null @@ -1,84 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Weight.h" -#include "paddle/utils/Logging.h" - -namespace paddle { - -Weight::Weight(size_t height, size_t width, ParameterPtr param) { - VectorPtr vPtr = param->getBuf(PARAMETER_VALUE); - VectorPtr gPtr = param->getBuf(PARAMETER_GRADIENT); - - // create a new weight - if (param->isSparse()) { - CHECK_LE(param->getSize(), width * height); - } else { - CHECK_EQ(param->getSize(), width * height); - } - - // weight_ - weight_ = param->getMat(PARAMETER_VALUE); - if (!weight_ && vPtr) { - weight_ = Matrix::create(vPtr->getMemoryHandle(), height, width); - } - if (weight_) { - CHECK_EQ(height, weight_->getHeight()); - CHECK_EQ(width, weight_->getWidth()); - } - - // weightGrad - weightGrad_ = param->getMat(PARAMETER_GRADIENT); - if (!weightGrad_ && gPtr) { - weightGrad_ = Matrix::create(gPtr->getMemoryHandle(), height, width); - } - if (weightGrad_) { - CHECK_EQ(height, weightGrad_->getHeight()); - CHECK_EQ(width, weightGrad_->getWidth()); - } - - parameter_ = param; -} - -Weight::Weight(size_t height, size_t width, ParameterPtr param, size_t offset) { - VectorPtr vPtr = param->getBuf(PARAMETER_VALUE); - VectorPtr gPtr = param->getBuf(PARAMETER_GRADIENT); - - // create a new weight - CHECK_LE(offset + width * height, param->getSize()); - - // weight_ - if (vPtr) { - weight_ = Matrix::create(vPtr->getData() + offset, - height, - width, - /* trans */ false, - param->useGpu()); - } - - // weightGrad - if (gPtr) { - weightGrad_ = Matrix::create(gPtr->getData() + offset, - height, - width, - /* trans */ false, - param->useGpu()); - } - - parameter_ = param; -} - -const ParameterPtr& Weight::getParameterPtr() { return parameter_; } -void Weight::setParameterPtr(ParameterPtr param) { parameter_ = param; } -} // namespace paddle diff --git a/paddle/parameter/Weight.h b/paddle/parameter/Weight.h deleted file mode 100644 index 113dd6530c82fe1e831ad4a35e9cbcb9880b9243..0000000000000000000000000000000000000000 --- a/paddle/parameter/Weight.h +++ /dev/null @@ -1,48 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include - -#include "paddle/math/Matrix.h" -#include "paddle/math/SparseRowMatrix.h" -#include "paddle/parameter/Parameter.h" - -namespace paddle { - -class Weight { - private: - MatrixPtr weight_; - MatrixPtr weightGrad_; - ParameterPtr parameter_; - - public: - Weight(size_t height, size_t width, ParameterPtr parameter); - Weight(size_t height, size_t width, ParameterPtr parameter, size_t offset); - - const MatrixPtr& getW() { return weight_; } - const MatrixPtr& getWGrad() { return weightGrad_; } - const ParameterPtr& getParameterPtr(); - - void incUpdate(const UpdateCallback& callback) { - getParameterPtr()->incUpdate(callback); - } - - void setParameterPtr(ParameterPtr param); -}; - -typedef std::vector> WeightList; - -} // namespace paddle diff --git a/paddle/parameter/tests/test_argument.cpp b/paddle/parameter/tests/test_argument.cpp deleted file mode 100644 index 54ceb3e08714e37abb5d491c8973bee631b993be..0000000000000000000000000000000000000000 --- a/paddle/parameter/tests/test_argument.cpp +++ /dev/null @@ -1,57 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -using namespace paddle; // NOLINT - -TEST(Argument, poolSequenceWithStride) { - Argument input, output; - ICpuGpuVector::resizeOrCreate(input.sequenceStartPositions, 5, false); - int* inStart = input.sequenceStartPositions->getMutableData(false); - inStart[0] = 0; - inStart[1] = 9; - inStart[2] = 14; - inStart[3] = 17; - inStart[4] = 30; - - int strideResult[] = {0, 5, 9, 14, 17, 22, 27, 30}; - int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30}; - - for (auto reversed : {false, true}) { - ICpuGpuVectorPtr stridePositions; - output.poolSequenceWithStride( - input, 5 /* stride */, &stridePositions, reversed); - - const int* outStart = output.sequenceStartPositions->getData(false); - CHECK_EQ(outStart[0], 0); - CHECK_EQ(outStart[1], 2); - CHECK_EQ(outStart[2], 3); - CHECK_EQ(outStart[3], 4); - CHECK_EQ(outStart[4], 7); - - CHECK_EQ(stridePositions->getSize(), 8UL); - auto result = reversed ? strideResultReversed : strideResult; - for (int i = 0; i < 8; i++) { - CHECK_EQ(stridePositions->getData(false)[i], result[i]); - } - } -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/parameter/tests/test_common.cpp b/paddle/parameter/tests/test_common.cpp deleted file mode 100644 index 89dcc6c751eb2ec07bfe8297c93d56c824086211..0000000000000000000000000000000000000000 --- a/paddle/parameter/tests/test_common.cpp +++ /dev/null @@ -1,174 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include -#include -#include -#include -#include - -using namespace paddle; // NOLINT - -class CommonTest : public ::testing::Test { - protected: - CommonTest() : testStat_("test") {} - virtual ~CommonTest() {} - virtual void SetUp() { - const size_t buffSize[] = { - 100, 128, 500, 1024, 4096, 10240, 102400, 1000000}; - sizeVec_.resize(8); - memcpy(&sizeVec_[0], &buffSize[0], 8 * sizeof(size_t)); - valueUint_.resize(4); - valueUint_[0].first = 0.0; - valueUint_[0].second = 0.0; - valueUint_[1].first = 0.0; - valueUint_[1].second = 1.0; - valueUint_[2].first = 1.0; - valueUint_[2].second = 0.0; - valueUint_[3].first = 1.0; - valueUint_[3].second = 1.0; - learningRate_ = 1.0; - } - - void test_sgdUpadate(real* gradientBuffer, - real* valueBuffer, - real* momentumBuffer, - size_t size); - - virtual void TreaDown() { LOG(INFO) << "All Test Finished."; } - - protected: - std::vector> valueUint_; - std::vector sizeVec_; - real learningRate_; - StatSet testStat_; -}; - -void CommonTest::test_sgdUpadate(real* gradientBuffer, - real* valueBuffer, - real* momentumBuffer, - size_t size) { -// sgdUpdateAvx has no double version yet -#if defined(__AVX__) && !defined(PADDLE_TYPE_DOUBLE) - real valueSum1 = 0, valueSum2 = 0, momSum1 = 0, momSum2 = 0; - real* gradTmp = new real[size]; - real* valueTmp = new real[size]; - real* momentumTmp = new real[size]; - memcpy(gradTmp, gradientBuffer, size * sizeof(real)); - memcpy(valueTmp, valueBuffer, size * sizeof(real)); - memcpy(momentumTmp, momentumBuffer, size * sizeof(real)); - for (auto& arg : valueUint_) { - { - { - struct timeval t; - REGISTER_TIMER("gettimeofday", 0, testStat_); - gettimeofday(&t, NULL); - } - REGISTER_TIMER("avxTimer", 0); - sgdUpdateAvx(learningRate_, - arg.first, - arg.second, - size, - valueBuffer, - gradientBuffer, - momentumBuffer); - } - for (size_t i = 0; i < size; i++) { - valueSum1 += valueBuffer[i]; - momSum1 += momentumBuffer[i]; - // std::cout << "[" - // << valueBuffer[i] - // << "," << momentumBuffer[i] - // << "," << gradientBuffer[i] << "],"; - } - { - REGISTER_TIMER("cpuTimer", 0); - sgdUpdateCpu(learningRate_, - arg.first, - arg.second, - size, - valueTmp, - gradTmp, - momentumTmp); - } - for (size_t i = 0; i < size; i++) { - valueSum2 += valueTmp[i]; - momSum2 += momentumTmp[i]; - // std::cout << "[" - // << valueTmp[i] - // << "," << momentumTmp[i] - // << "," << gradTmp[i] << "],"; - } - - VLOG(3) << "valueSum1 = " << valueSum1 << " ; valueSum2 = " << valueSum2; - VLOG(3) << "momSum1 = " << momSum1 << " ; momSum2 = " << momSum2; - ASSERT_EQ(valueSum1, valueSum2); - ASSERT_EQ(momSum1, momSum2); - } - delete[] gradTmp; - delete[] valueTmp; - delete[] momentumTmp; -#endif -} - -TEST_F(CommonTest, sgdUpdate) { - const size_t alignHeader[] = {0, 2, 3, 5, 7, 8}; - for (auto& size : sizeVec_) { - real *gradientBuffer, *valueBuffer, *momentumBuffer; - CHECK_EQ(posix_memalign((void**)&gradientBuffer, 32, sizeof(real) * size), - 0); - CHECK_EQ(posix_memalign((void**)&valueBuffer, 32, sizeof(real) * size), 0); - CHECK_EQ(posix_memalign((void**)&momentumBuffer, 32, sizeof(real) * size), - 0); - - for (size_t i = 0; i < size; i++) { - gradientBuffer[i] = 1.0; - valueBuffer[i] = 2.0; - momentumBuffer[i] = 3.0; - } - for (int i = 0; i < 6; i++) { - LOG(INFO) << "----------------------" << size << ":" << alignHeader[i] - << "-------------------------"; - test_sgdUpadate(&gradientBuffer[alignHeader[i]], - &valueBuffer[alignHeader[i]], - &momentumBuffer[alignHeader[i]], - size - alignHeader[i]); - } - free(gradientBuffer); - free(valueBuffer); - free(momentumBuffer); - } - globalStat.printAllStatus(); - testStat_.printAllStatus(); -} - -TEST_F(CommonTest, syncThreadPool) { - SyncThreadPool pool(10); - - std::vector nums; - nums.resize(10); - - pool.exec([&](int tid, size_t numThreads) { nums[tid] = tid; }); - for (size_t i = 0; i < nums.size(); ++i) { - EXPECT_EQ((int)i, nums[i]); - } - - pool.exec([&](int tid, size_t numThreads) { nums[tid] -= tid; }); - for (size_t i = 0; i < nums.size(); ++i) { - EXPECT_EQ((int)0, nums[i]); - } -} diff --git a/paddle/pserver/BaseClient.cpp b/paddle/pserver/BaseClient.cpp deleted file mode 100644 index a6204ef47ea553246ddadbb2eae6cc714cafe594..0000000000000000000000000000000000000000 --- a/paddle/pserver/BaseClient.cpp +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "BaseClient.h" -#include -#include -#include -#include "paddle/utils/Stat.h" - -DECLARE_string(pservers); - -namespace paddle { - -BaseClient::BaseClient(bool separate, int numPorts) - : stopping_(false), numPorts_(numPorts), separateSendAndRecv_(separate) { - CHECK_GT(numPorts, 0); -} - -BaseClient::~BaseClient() {} - -void BaseClient::recvData() { recvSyncBarrier_->wait(); } - -void BaseClient::synchronize(SyncObject syncObjectId) { - SynchronizeRequest request; - request.set_sync_object_id(syncObjectId); - std::vector responses; - multiCall(__func__, request, &responses); -} - -void BaseClient::startThreads() { - if (!separateSendAndRecv_) { - return; - } - recvSyncBarrier_.reset(new ThreadBarrier(threadNum_ + 1)); - - sendThreads_.resize(threadNum_); - recvThreads_.resize(threadNum_); - sendJobQueue_.resize(threadNum_); - recvJobQueue_.resize(threadNum_); - - for (int i = 0; i < threadNum_; ++i) { - sendJobQueue_[i].reset(new SendQueue()); - recvJobQueue_[i].reset(new SendQueue()); - - sendThreads_[i].reset( - new std::thread([this](int id) { this->send(id); }, i)); - - recvThreads_[i].reset( - new std::thread([this](int id) { this->recv(id); }, i)); - } -} - -void BaseClient::finishThreads() { - if (!separateSendAndRecv_) { - return; - } - stopping_ = true; - for (int i = 0; i < threadNum_; i++) { - sendJobQueue_[i]->enqueue(nullptr); - } - for (auto& thread : sendThreads_) { - thread->join(); - } - for (auto& thread : recvThreads_) { - thread->join(); - } - stopping_ = false; -} -} // namespace paddle diff --git a/paddle/pserver/BaseClient.h b/paddle/pserver/BaseClient.h deleted file mode 100644 index d50230e73a3a7d128cbfd1d70517fddd228fb1bb..0000000000000000000000000000000000000000 --- a/paddle/pserver/BaseClient.h +++ /dev/null @@ -1,311 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ParameterService.pb.h" -#include "paddle/math/Matrix.h" -#include "paddle/pserver/ProtoServer.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Queue.h" - -namespace paddle { - -/** - * it manages all connections to pservers. - * it exists two modes to manage connections to all pservers. Firstly, one - * connection owns two threads that separately manage to send and receive - * data. Secondly, each thread uses one connection for all activation in it. - * the first solution arms with sendThreads_/recvThreads_ and sendJobQueue_/ - * recvJobQueue_. the second solution use some shared thread pool to manage - * connections. - */ -class BaseClient { - protected: - typedef std::unique_ptr ThreadPtr; - typedef std::vector> InputIovs; - typedef std::vector SendRequest; - typedef std::vector SendDataRequestVec; - - // TODO(yanfei): - // refine data structure to unify parameter and features communication - struct SendJob { - /// store parameters related blocks data - InputIovs parallelInputIovs; - /// store protobuf request - SendRequest parallelRequests; - /// store data, such as features for metric learning - SendDataRequestVec parallelDataRequests; - }; - - public: - explicit BaseClient(bool separate = false, int numPorts = FLAGS_ports_num); - - virtual ~BaseClient(); - - typedef std::shared_ptr SendJobPtr; - typedef Queue SendQueue; - - /// send data to server, support only synchronize - template - void putData(int clientId, - SendDataType type, - DataType* datas, - size_t size, - DataUpdateMode mode) { - synchronize(SYNC_DATA); - sendData(clientId, type, mode, datas, size); - recvData(); - synchronize(SYNC_DATA); - } - - template - void putOwnData(int clientId, - SendDataType type, - DataType* datas, - size_t size) { - putData(clientId, type, datas, size, DATA_UPDATE_MODE_SET_OWN); - } - - template - void getAllData(int clientId, - SendDataType type, - DataType* datas, - size_t size) { - sendData(clientId, - type, - DATA_UPDATE_MODE_GET_ALL, - reinterpret_cast(NULL), - 0); - recvData(); - size_t dataOffset = 0; - for (auto& recvMem : recvDataMems_) { - CHECK_LE(dataOffset, size); - size_t memSize = std::min(recvMem.get()->getSize(), - sizeof(DataType) * (size - dataOffset)); - CHECK_EQ(memSize % sizeof(DataType), size_t(0)); - memcpy(datas + dataOffset, recvMem.get()->getBuf(), memSize); - dataOffset += memSize / sizeof(DataType); - } - CHECK_EQ(dataOffset, size); - } - - /** - * Reduces values on all clients. - * This reduce just support SUM. - * The results are saved in recvBuf of rootId client - */ - template - void reduce(DataType* sendBuf, - DataType* recvBuf, - size_t size, - int clientId, - int rootId) { - putOwnData(clientId, DATA_REDUCE_SUM, sendBuf, size); - if (rootId == clientId) { - getAllData(clientId, DATA_REDUCE_SUM, recvBuf, size); - } - } - - /** - * return trans data type according to the input type - */ - virtual TransDataType getTransDtype(const std::type_info& info) { - TransDataType dataType; - if (typeid(int*) == info) { // NOLINT - dataType = TRANS_INT32; - } else if (typeid(uint32_t*) == info) { // NOLINT - dataType = TRANS_UINT32_T; - } else if (typeid(int64_t*) == info) { // NOLINT - dataType = TRANS_INT64_T; - } else if (typeid(uint64_t*) == info) { // NOLINT - dataType = TRANS_UINT64_T; - } else if (typeid(float*) == info) { // NOLINT - dataType = TRANS_FLOAT; - } else if (typeid(double*) == info) { // NOLINT - dataType = TRANS_DOUBLE; - } else { - LOG(FATAL) << "not supported"; - } - return dataType; - } - - protected: - /// for a > 0, b > 0: - /// return the smallest x s.t. b*x >= a - static int divup(int a, int b) { return (a + b - 1) / b; } - - int calcClientId(int i, int serviceNum) { - return (i + FLAGS_trainer_id * numPorts_) % serviceNum; - } - - /// start threads in sendThreads_ and recvThreads_ - void startThreads(); - - /// finish threads in sendThreads_ and recvThreads_ - void finishThreads(); - - template - void prepareData(int clientId, - SendDataType type, - DataUpdateMode updateMode, - DataType* datas, - size_t size, - SendJob* sendJob) { - sendJob->parallelDataRequests.resize(serviceNum_); - sendJob->parallelInputIovs.resize(serviceNum_); - for (int i = 0; i < serviceNum_; ++i) { - auto& request = sendJob->parallelDataRequests[i]; - request.set_update_mode(updateMode); - request.set_type(type); - request.set_client_id(clientId); - request.set_server_id(i); - } - - /// split datas which need send to Server into serviceNum_ pieces - if (!datas) { - CHECK(!size) << "ownSize should be zero since datas is nullptr"; - } - size_t baseSize = size / serviceNum_; - size_t dataOffset = 0; - for (int i = 0; i < serviceNum_; ++i) { - auto& request = sendJob->parallelDataRequests[i]; - DataBlock* block = request.add_blocks(); - size_t ownSize = size_t(i) < size % serviceNum_ ? baseSize + 1 : baseSize; - size_t realSize = datas ? std::max(ownSize, size_t(1)) : 0; - block->set_total_size(realSize * sizeof(DataType)); - block->set_data_size(sizeof(DataType)); - // TODO(yuyang18): The getTransDtype can be rewritten as template method - // to reduce runtime overhead. - block->set_data_type(getTransDtype(typeid(DataType*))); // NOLINT - if (datas) { - sendJob->parallelInputIovs[i].push_back( - {datas + dataOffset, realSize * sizeof(DataType)}); - } - dataOffset += ownSize; - } - CHECK_EQ(dataOffset, size); - } - - /** - * @brief send data to all data servers - * - * @note each trainer sends all its data to all data servers - * it's for broadcast data synchronization, such as features - * synchronization in metric learning. - */ - template - void sendData(int clientId, - SendDataType type, - DataUpdateMode updateMode, - DataType* datas, - size_t size) { - SendJobPtr sendJob = std::make_shared(); - prepareData(clientId, type, updateMode, datas, size, sendJob.get()); - for (int i = 0; i < threadNum_; ++i) { - sendJobQueue_[i]->enqueue(sendJob); - } - } - - /** - * @brief recv data from all data servers - * - * @note synchronize all recv threads - */ - void recvData(); - - /// send request, and recv responses - template - void multiCall(const char* funcName, - const ProtoIn& request, - std::vector* responses) { - responses->resize(clients_.size()); - size_t numClients = clients_.size(); - for (size_t i = 0; i < numClients; ++i) { - clients_[i].send(funcName, request); - } - for (size_t i = 0; i < numClients; ++i) { - clients_[i].recv(&(*responses)[i]); - } - } - - /** - * @brief synchronize all trainers and pservers - * - * @note used to ensure that data of all trainers have been received - */ - void synchronize(SyncObject syncObjectId = SYNC_DEFAULT); - - /** - * @brief use multithread to separately send data - * - * @note each thread should read its own JobQueue to handle requests - * each thread should calcClientId() to retrieve connections - * managed by himself. - * send and recv are implemented in child class. - */ - virtual void send(int threadId) = 0; - - /** - * @brief use multithread to separately receive data - * - * @note almost same as send() - */ - virtual void recv(int threadId) = 0; - - protected: - bool stopping_; - /// nodes * ports that means the number of real pservers - int serviceNum_; - /** - * threads num for managing all services. Normally the - * number of pservers are relatively less than several - * hundreds so that using thread-based parallelization - * can benifit traffic performance and pserver's sgd - * optimization performance. - */ - int threadNum_; - /// the connection manager at client end - std::vector clients_; - /// send threads for parallelization - std::vector sendThreads_; - /// recv threads for parallelization - std::vector recvThreads_; - std::unique_ptr recvSyncBarrier_; - - // TODO(yanfei): - // current pserver's will return value until all parameters' - // optimization are finished so that recv are not overlapped - // in reality. More robust implimentation should be to pipeline - // all send/recv action based on parameter unit level, and - // it will benifits deep and larger model training in future, - // especially local node compution power surpasses inter-connection - // such as GPU cluster, even with BOX GPU cluster. - // queue for buffering send request - /** - * send/recv queue cooperates with each other to accomplish - * overlapping communication with forwardBackward action. - */ - std::vector> sendJobQueue_; - /// queue for buffering recv request - std::vector> recvJobQueue_; - /// specific for dserver - SendJob sendJob_; - /// port num for each node - int numPorts_; - /// if set, overlapped optimization is disabled - bool separateSendAndRecv_; - std::vector recvDataMems_; -}; -} // namespace paddle diff --git a/paddle/pserver/LightNetwork.cpp b/paddle/pserver/LightNetwork.cpp deleted file mode 100644 index 4c0da2217e880b7509ea5f42da5ac7ffe93a53ec..0000000000000000000000000000000000000000 --- a/paddle/pserver/LightNetwork.cpp +++ /dev/null @@ -1,459 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include -#include -#include - -#include -#include -#include -#include - -#include "LightNetwork.h" -#include "RDMANetwork.h" -#include "paddle/utils/StringUtil.h" -#include "paddle/utils/Util.h" - -/// quick ack can reduce the latency of small message -DEFINE_bool(small_messages, - false, - "if message size is small, recommend set it True to enable quick " - "ack and no delay"); - -/// reasonable sock_send_buf_size can control the traffic injected into switch -/// network. Injecting too many data into traffic could cause packets loss which -/// cause long latency and degrade the efficiency of communication. -DEFINE_int32(sock_send_buf_size, - 1024 * 1024 * 40, - "restrict sock send buff size, can reduce network congestion if " - "set carefully"); - -/// reasonable size can hold bursted packets and reduce packets loss -DEFINE_int32(sock_recv_buf_size, - 1024 * 1024 * 40, - "restrict sock recv buff size"); - -/// reasonable sock_listen_queue_size can control maximum pending connections. -DEFINE_int32(sock_listen_queue_size, - 1024, - "listen queue size when pserver listen a TCP port"); - -namespace paddle { - -/** - * @brief get ip address from interface name - * - * @param[in] device device interface name - */ -std::string getIpAddr(std::string &device) { - int sock; - struct sockaddr_in sin; - struct ifreq ifr; - - sock = socket(AF_INET, SOCK_DGRAM, 0); - CHECK(sock >= 0) << "Create socket error."; - - strncpy(ifr.ifr_name, device.c_str(), IFNAMSIZ); - ifr.ifr_name[IFNAMSIZ - 1] = 0; - - CHECK_GE(ioctl(sock, SIOCGIFADDR, &ifr), 0); - memcpy(&sin, &ifr.ifr_addr, sizeof(sin)); - close(sock); - return std::string(inet_ntoa(sin.sin_addr)); -} - -/** - * @brief set sock option - * - * @param[in] sockfd sock file descriptor - * - * @note adjust some default sock option for better performance - */ -void setOption(int sockfd) { -#if !defined(__APPLE__) && !defined(__OSX__) - int sendSize = FLAGS_sock_send_buf_size; - int recvSize = FLAGS_sock_recv_buf_size; - CHECK_GE( - setsockopt(sockfd, SOL_SOCKET, SO_RCVBUF, &recvSize, sizeof(recvSize)), - 0); - CHECK_GE( - setsockopt(sockfd, SOL_SOCKET, SO_SNDBUF, &sendSize, sizeof(sendSize)), - 0); -#endif - - if (FLAGS_small_messages) { - int optval = 1; - CHECK_GE( - setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, &optval, sizeof(optval)), - 0); -#ifdef TCP_QUICKACK - optval = 1; - CHECK_GE( - setsockopt(sockfd, IPPROTO_TCP, TCP_QUICKACK, &optval, sizeof(optval)), - 0); -#endif - } - int reuse = 1; - CHECK_GE(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &reuse, sizeof(reuse)), - 0); -} - -/** - * @brief class constructor for SocketServer - * @param[in] addr sock bind address - * @param[in] port sock bind port - * @param[in] rdmaCpu rdma sock bind cpu core - * - * @note start one socket server which hosts parameter server process. - * rdmaCpu is passed to rdma deamon for better performance, and - * start tcp socket instead of rdma socket if rdmaCpu is equal - * to -1. Each trainer process starts one connection to one socket - * server, and use --ports_num to build more connections to harness - * fat communication channel if necessary. - * each connection is controlled by single thread with blocking - * read and write. - */ -SocketServer::SocketServer(const std::string &addr, int port, int rdmaCpu) - : port_(port), addr_(addr), stopping_(false) { - if (rdmaCpu == -1) { - tcpRdma_ = F_TCP; - socket_ = 0; - maxPendingConnections_ = FLAGS_sock_listen_queue_size; - } else { - tcpRdma_ = F_RDMA; - rdmaCpu_ = rdmaCpu; - rdmaSocket_ = 0; - - std::stringstream ss; - ss << port; - rdmaUri_ = "rdma://" + addr + ":" + ss.str(); - } - - /// trigger to initialize RDMA lib - CHECK(RdmaClientDaemons::get()) << "initilizate RDMA failed\n"; -} - -SocketServer::~SocketServer() { - stopping_ = true; - /// trigger accept thread to stop - { - SocketClient trigger(addr_.empty() ? "127.0.0.1" : addr_, port_, tcpRdma_); - } - this->join(); -} - -/** - * @brief start one tcp server which hosts parameter server - * - * @note do tcp socket bind and listen. it will spawn one thread - * for each connection - */ -void SocketServer::tcpServer() { - int newsockfd; - socklen_t clilen; - struct sockaddr_in serv_addr, cli_addr; - struct hostent *server; - - /// First call to socket() function - socket_ = socket(AF_INET, SOCK_STREAM, 0); - CHECK(socket_ >= 0) << "ERROR opening socket"; - - /// Initialize socket structure - bzero((char *)&serv_addr, sizeof(serv_addr)); - serv_addr.sin_family = AF_INET; - serv_addr.sin_port = htons(port_); - if (!addr_.empty()) { - server = gethostbyname(addr_.c_str()); - CHECK(server) << "ERROR, no such host: " << addr_; - bcopy((char *)server->h_addr, - (char *)&serv_addr.sin_addr.s_addr, - server->h_length); - } else { - serv_addr.sin_addr.s_addr = INADDR_ANY; - } - - setOption(socket_); - - /// Now bind the host address using bind() call. - CHECK(bind(socket_, (struct sockaddr *)&serv_addr, sizeof(serv_addr)) >= 0) - << "ERROR on binding " << addr_; - - /// Now start listening for the clients, here process will - /// go in sleep mode and will wait for the incoming connection - listen(socket_, maxPendingConnections_); - clilen = sizeof(cli_addr); - - while (true) { - /// Accept actual connection from the client - newsockfd = accept(socket_, (struct sockaddr *)&cli_addr, &clilen); - if (stopping_) { - break; - } - CHECK(newsockfd >= 0) << "ERROR on accept"; - constexpr int kPeerNameLen = 128; - char peerName[kPeerNameLen]; - CHECK(inet_ntop(AF_INET, &cli_addr.sin_addr, peerName, kPeerNameLen)); - - SocketWorker *worker = - new SocketWorker(createChannel(newsockfd, std::string(peerName)), this); - worker->start(); - worker->detach(); - } - close(socket_); - LOG(INFO) << "pserver accept thread finish, addr=" << addr_ - << " port=" << port_; -} - -/** - * @brief start one rdma server which hosts parameter server - * - * @note do rdma bind and listen, which calling self-defined socket - * like rdma library. it will spawn one thread for each connection - */ -void SocketServer::rdmaServer() { - struct sxi_sock *newsock; - - /// First call to socket() function - rdmaSocket_ = rdma::ssocket(rdmaCpu_); - CHECK(rdmaSocket_) << "ERROR opening RDMA socket"; - - CHECK(rdma::bind(rdmaSocket_, rdmaUri_.c_str()) == 0) - << "ERROR bind RDMA socket"; - - /// Now start listening for the clients, here process will - /// go in sleep mode and will wait for the incoming connection - CHECK(rdma::listen(rdmaSocket_) == 0) << "ERROR listen RDMA socket"; - - while (true) { - /// Accept actual connection from the client - newsock = rdma::accept(rdmaSocket_); - if (stopping_) { - break; - } - CHECK(newsock) << "ERROR on accept"; - - constexpr int kPeerNameLen = 128; - char peerName[kPeerNameLen]; - - struct sockaddr_in *saddr = rdma::getSourceAddress(newsock); - CHECK(inet_ntop(AF_INET, &saddr->sin_addr, peerName, kPeerNameLen)); - - SocketWorker *worker = - new SocketWorker(createChannel(newsock, std::string(peerName)), this); - worker->start(); - worker->detach(); - } - rdma::close(rdmaSocket_); - LOG(INFO) << "pserver accept thread finish, rdma uri=" << rdmaUri_; -} - -/** - * @brief start a socket server - * - * @note framework for starting socket server - */ -void SocketServer::run() { - if (tcpRdma_ == F_TCP) { - LOG(INFO) << "tcp server start "; - tcpServer(); - } else if (tcpRdma_ == F_RDMA) { - LOG(INFO) << "rdma server start "; - rdmaServer(); - } -} - -/** - * @brief class constructor for rdma client deamons - * - * @note automatically start several client deamons for better performance - */ -std::unique_ptr RdmaClientDaemons::daemons_ = nullptr; -std::once_flag RdmaClientDaemons::initDataFlag_; - -RdmaClientDaemons::RdmaClientDaemons() { - if (FLAGS_rdma_tcp == "rdma") { - rdma::init(); - - struct sxi_socket *socket; - onlineCpus_ = rdma::numCpus(); - for (auto i = 0; i < onlineCpus_; i++) { - socket = rdma::csocket(i); - CHECK(socket) << "ERROR open client socket daemon"; - - rdmaClientSocket_.push_back(socket); - } - LOG(INFO) << "RDMA client daemons started, onlineCpus_:" << onlineCpus_; - /// round robin scheduler for new connection - curCpu_ = 0; - /// wait daemons to start completely. - sleep(2); - } -} - -RdmaClientDaemons::~RdmaClientDaemons() { - if (FLAGS_rdma_tcp == "rdma") { - for (auto i = 0; i < onlineCpus_; i++) { - rdma::close(rdmaClientSocket_[i]); - } - LOG(INFO) << "RDMA client daemons is destoryed, onlineCpus_ " - << onlineCpus_; - } -} - -/** - * @brief worker thread main context - * - * @note each connection from client(trainer) is controlled by single worker - * thread, which is for handling all parameter server requests - */ -void SocketWorker::run() { - LOG(INFO) << "worker started, peer = " << channel_->getPeerName(); - - std::vector inputIovs; - - while (true) { - std::unique_ptr msgReader = channel_->readMessage(); - if (!msgReader) { - break; - } - - auto callback = [this](const std::vector &outputIovs) { - channel_->writeMessage(outputIovs); - }; - - server_->handleRequest(std::move(msgReader), callback); - } - - LOG(INFO) << "worker begin to finish, peer = " << channel_->getPeerName(); - delete this; -} - -/** - * @brief start one tcp connection to tcp server - * @param[in] serverAddr tcp server ip - * @param[in] serverPort tcp server port - * - * @note each object contains one channel which accept byte stream - */ -void SocketClient::TcpClient(const std::string &serverAddr, int serverPort) { - struct sockaddr_in serv_addr; - struct hostent *server; - - int errRet; // temp for gethostbyname_r - - /// Create a socket point - int sockfd = socket(AF_INET, SOCK_STREAM, 0); - CHECK(sockfd >= 0) << "ERROR opening socket"; - -#if defined(__OSX__) || defined(__APPLE__) - server = getipnodebyname(serverAddr.c_str(), AF_INET, AI_DEFAULT, &errRet); - CHECK_NE(HOST_NOT_FOUND, errRet) << "ERROR, no such host: " << serverAddr - << " ret = " << errRet; - CHECK(server) << "getipnodebyname error!"; -#else - struct hostent hostinfo; - char buf[1024]; // temp for gethostbyname_r - CHECK_EQ( - 0, - gethostbyname_r( - serverAddr.c_str(), &hostinfo, buf, sizeof(buf), &server, &errRet)) - << "ERROR, no such host: " << serverAddr << " ret = " << errRet; - CHECK(server) << "gethostbyname_r error!"; -#endif - - bzero((char *)&serv_addr, sizeof(serv_addr)); - serv_addr.sin_family = AF_INET; - bcopy((char *)server->h_addr, - (char *)&serv_addr.sin_addr.s_addr, - server->h_length); - serv_addr.sin_port = htons(serverPort); - - setOption(sockfd); - - /// Now connect to the server - int retry_count = 0; - do { - if (connect(sockfd, (sockaddr *)&serv_addr, sizeof(serv_addr)) == 0) { - break; - } - - if (errno == ECONNREFUSED) { - LOG(WARNING) << "connection refused by pserver, try again!"; - if (retry_count++ >= 7) { - LOG(FATAL) << "connection refused by pserver, maybe pserver failed!"; - } - std::this_thread::sleep_for(std::chrono::seconds(1)); - } else { - CHECK(errno != 0) << "ERROR connecting to " << serverAddr << ":" - << serverPort << "errorno: " << errno; - } - } while (errno == ECONNREFUSED); - - channel_.reset(new SocketChannel(sockfd, serverAddr)); - tcpRdma_ = F_TCP; -} - -/** - * @brief start one RDMA connection to rdma server - * @param[in] serverAddr rdma server ip - * @param[in] serverPort rdma server port - * - * @note each object contains one channel which accept byte stream - * for rdma, low level sock also provide byte stream api. - */ -void SocketClient::RdmaClient(const std::string &serverAddr, int serverPort) { - struct sxi_sock *sock; - - std::stringstream ss; - ss << serverPort; - - std::string rdmaUri = "rdma://" + serverAddr + ":" + ss.str(); - - RdmaClientDaemons *daemons = RdmaClientDaemons::daemons_->get(); - socketDaemon_ = daemons->selectDaemon(); - - /// connect to server with socket daemon - sock = rdma::connect(socketDaemon_, rdmaUri.c_str()); - CHECK(sock) << "ERROR connect to server" << rdmaUri; - - std::vector seg; - str::split(rdmaUri, '/', &seg); - std::string server = seg.at(seg.size() - 1); - channel_.reset(new SocketChannel(sock, server)); - tcpRdma_ = F_RDMA; -} - -/** - * @brief class constructor - * @param[in] serverAddr pserver ip address - * @param[in] serverPort pserver port - * @param[in] ChannelType F_TCP or F_RDMA - * - * @note responsible for building one connection to specified pserver port - */ -SocketClient::SocketClient(const std::string &serverAddr, - int serverPort, - enum ChannelType channelType) { - if (channelType == F_RDMA) - RdmaClient(serverAddr, serverPort); - else - TcpClient(serverAddr, serverPort); -} - -} // namespace paddle diff --git a/paddle/pserver/LightNetwork.h b/paddle/pserver/LightNetwork.h deleted file mode 100644 index bcfc9655e989e80e08e9dce9b8734c0643cbf661..0000000000000000000000000000000000000000 --- a/paddle/pserver/LightNetwork.h +++ /dev/null @@ -1,185 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "SocketChannel.h" - -#include -#include -#include -#include - -#include "paddle/utils/Thread.h" - -struct sxi_socket; - -namespace paddle { - -class SocketWorker; - -/** - * @brief class for holding all parameters processing for current port - * - * @note each parameter server inherits from one socket server, each - * server contains serveral woker threads which are to parallelize - * the processing of computation, but share some common datas stored - * in child class of socketserver. - */ -class SocketServer : public Thread { - // rdmaCpu controls the cpu affinity of RDMA server daemon, - // which could benifit performance. rdmaCpu = -1 means TCP - // is used instead of RDMA transport. - public: - SocketServer(const std::string& addr, int port, int rdmaCpu); - ~SocketServer(); - - virtual void run(); - - typedef std::function& outputIovs)> - ResponseCallback; - - protected: - // - // The derived class needs to implement this function - // to handle the request received by SocketWorker - // The request is encapsulated by MsgReader, which contains - // a set of blocks. - virtual void handleRequest(std::unique_ptr msgReader, - ResponseCallback callback) = 0; - - std::unique_ptr createChannel(int sock, - const std::string& peerName) { - return std::unique_ptr(new SocketChannel(sock, peerName)); - } - std::unique_ptr createChannel(struct sxi_sock* sock, - const std::string& peerName) { - return std::unique_ptr(new SocketChannel(sock, peerName)); - } - - friend class SocketWorker; - - private: - void rdmaServer(); - void tcpServer(); - - void detach() {} // detach accept thread is forbidden - - protected: - enum ChannelType tcpRdma_; - // for rdma - int rdmaCpu_; - std::string rdmaUri_; - sxi_socket* rdmaSocket_; - // for tcp - int port_; - std::string addr_; - int socket_; - int maxPendingConnections_; - bool stopping_; -}; - -/** - * @brief class for holding one connection from one trainer - * - * @note all parameter processing will run in the context of this worker - */ -class SocketWorker : public Thread { - public: - SocketWorker(std::unique_ptr&& channel, SocketServer* server) - : channel_(std::move(channel)), server_(server) {} - - virtual ~SocketWorker() {} - - virtual void run(); - - protected: - std::unique_ptr channel_; - SocketServer* server_; - enum ChannelType tcpRdma_; -}; - -/** - * @brief class for providing rdma client deamon thread - * - * @note the deamons are required by sock like rdam library. Here - * use singleton model for daemons. Each deamon hosts in - * single cpu core for better load balance performance - */ -class RdmaClientDaemons { - private: - RdmaClientDaemons(); - - static std::unique_ptr daemons_; - - public: - static RdmaClientDaemons* get() { - std::call_once(RdmaClientDaemons::initDataFlag_, - &RdmaClientDaemons::getInstance); - - return daemons_.get(); - } - - struct sxi_socket* selectDaemon() { - int cpu = curCpu_; - curCpu_ = (curCpu_ + 1) % onlineCpus_; - - LOG(INFO) << "select daemon " << cpu << "onlineCpus_ " << onlineCpus_; - return rdmaClientSocket_[cpu]; - } - - ~RdmaClientDaemons(); - - public: - friend class SocketClient; - - private: - static std::once_flag initDataFlag_; - static void getInstance() { - if (!daemons_.get()) daemons_.reset(new RdmaClientDaemons()); - } - - std::vector rdmaClientSocket_; - std::atomic curCpu_; - int onlineCpus_; -}; - -/** - * @brief management for client connection which are from trainers - * - * @note it contains one channel descriptor which used to write and - * read data - */ -class SocketClient { - public: - SocketClient(const std::string& serverAddr, - int serverPort, - enum ChannelType channelType); - - SocketChannel* getChannel() { return channel_.get(); } - - protected: - std::unique_ptr channel_; - struct sxi_socket* socketDaemon_; - enum ChannelType tcpRdma_; - - private: - void RdmaClient(const std::string& serverAddr, int serverPort); - void TcpClient(const std::string& serverAddr, int serverPort); -}; - -std::string getIpAddr(std::string& device); -void setOption(int sockfd); - -} // namespace paddle diff --git a/paddle/pserver/ParameterClient2.cpp b/paddle/pserver/ParameterClient2.cpp deleted file mode 100644 index 43e4902b0f0f73840624041f19ba7f4eb9a45844..0000000000000000000000000000000000000000 --- a/paddle/pserver/ParameterClient2.cpp +++ /dev/null @@ -1,781 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "ParameterClient2.h" -#include "paddle/math/SparseRowMatrix.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/StringUtil.h" - -DEFINE_string(pservers, "127.0.0.1", "Comma separated addresses of pservers"); -DEFINE_int32(parallel_thread_num, 1, "Thread number for parameter send"); - -namespace paddle { - -template -void copyToRepeatedField(google::protobuf::RepeatedField* dest, - const T2* src, - size_t size) { - dest->Clear(); - dest->Reserve(size); - for (size_t i = 0; i < size; ++i) { - dest->AddAlreadyReserved(src[i]); - } -} - -ParameterClient2::ParameterClient2(bool separate, int port, int numPorts) - : BaseClient(separate, numPorts), port_(port) { -#ifndef PADDLE_DISABLE_TIMER - forwardbackwordTime_ = 0; -#endif -} - -int ParameterClient2::calcParameterBlockSize( - const std::vector& parameters, size_t serviceNum) { - size_t totalSize = 0; - for (auto& para : parameters) { - totalSize += para->getSize(); - } - size_t perServerSize = totalSize / serviceNum; - - int sizeBits = 64 - __builtin_clzl(perServerSize); - - /// 2^10 is min block size - /// 2^7 will be max number of blocks in one pserver - int blockSizeBits = std::max((sizeBits - 7), 10); - return 1 << blockSizeBits; -} - -void ParameterClient2::initThreads() { - threadNum_ = serviceNum_; - if (FLAGS_parallel_thread_num > 1) { - LOG(INFO) << "parallel_thread_num dosent need to set"; - } - syncThreadPool_.reset(new SyncThreadPool(threadNum_)); - startThreads(); -} - -bool ParameterClient2::init(const std::vector& parameters) { - destroy(); - - std::vector hosts; - str::split(FLAGS_pservers, ',', &hosts); - serviceNum_ = hosts.size() * numPorts_; - uint64_t denseBlockSize = calcParameterBlockSize(parameters, serviceNum_); - - /// setup prefetch matrix if exists - for (auto& para : parameters) { - /// set block size for each parameter - para->getConfig().set_parameter_block_size( - para->getConfig().sparse_remote_update() ? para->getConfig().dims(1) - : denseBlockSize); - } - - for (auto& para : parameters) { - CHECK_NE(-1UL, para->getID()) << "id in parameter is not initialized"; - parameterMap_[para->getID()] = para; - } - - allSegments_.reserve(parameters.size()); - - for (auto& para : parameters) { - ParameterSegments segments; - segments.name = para->getName(); - segments.id = para->getID(); - allSegments_.push_back(segments); - if (para->getConfig().sparse_remote_update()) { - CHECK_EQ(para->getConfig().parameter_block_size(), - para->getConfig().dims(1)) - << "For sparse remote update parameter," - << " block size is the width of each row."; - } - } - - /// init clients - clients_.reserve(serviceNum_); - recvDataMems_.resize(serviceNum_); - - for (size_t i = 0; i < hosts.size(); ++i) { - for (int j = 0; j < numPorts_; ++j) { - LOG(INFO) << "pserver " << i * numPorts_ + j << " " << hosts[i] << ":" - << port_ + j; - if (FLAGS_rdma_tcp == "rdma") { - clients_.emplace_back(hosts[i], port_ + j, F_RDMA); - } else { - clients_.emplace_back(hosts[i], port_ + j, F_TCP); - } - } - } - - sparseDistribution_.reset(new SparseParameterDistribution(serviceNum_)); - - sleep(2); - - initThreads(); - - return true; -} - -ParameterClient2::~ParameterClient2() { destroy(); } - -void ParameterClient2::destroy() { - if (clients_.empty()) { - /// this means not initialized. - return; - } - finishThreads(); - - parameterMap_.clear(); - allSegments_.clear(); - clients_.clear(); -} - -void ParameterClient2::sendParallel(int tid, - size_t numThreads, - ParameterType recvParameterType) { - int numMyClients = divup(serviceNum_ - tid, numThreads); - - for (int j = 0; j < numMyClients; ++j) { - REGISTER_TIMER("client_sendAndRecv_send"); - int i = numThreads * j + tid; - /// Try to make different clients to send data to different pservers - /// at the same time so that they will not flood data to the same - /// pserver. - i = calcClientId(i, serviceNum_); - clients_[i].send("sendParameter", - sendJob_.parallelRequests[i], - sendJob_.parallelInputIovs[i]); - - /// clear large structure - sendJob_.parallelRequests[i].Clear(); - sendJob_.parallelInputIovs[i].clear(); - } - - std::vector bufs; - SendParameterResponse response; - for (int j = 0; j < numMyClients; ++j) { - REGISTER_TIMER("client_sendAndRecv_recv"); - int i = numThreads * j + tid; - i = calcClientId(i, serviceNum_); - auto msgReader = clients_[i].recv(&response); - CHECK_EQ(msgReader->getNumBlocks(), (size_t)response.blocks_size()); - bufs.clear(); - bufs.reserve(response.blocks_size()); - for (auto& block : response.blocks()) { - auto it = parameterMap_.find(block.para_id()); - CHECK(it != parameterMap_.end()); - Parameter* parameter = it->second.get(); - real* buf = nullptr; - if (parameter->getBuf(recvParameterType)) { - buf = parameter->getBuf(recvParameterType)->getPoint(block.begin_pos()); - } else { - auto recvMat = dynamic_cast( - parameter->getMat(recvParameterType).get()); - CHECK(recvMat); - size_t width = parameter->getConfig().dims(1); - // TODO(wuyi): need add lock here? may also cause resize. - buf = recvMat->getLocalRow(block.begin_pos() / width); - } - /// sparse_id is not useful while receiving data since sparse data - /// storage is continuous, do commit recieved data as that of dense. - bufs.push_back(buf); - } - msgReader->readBlocks(bufs); - } -} - -void ParameterClient2::prepareSendData( - ParameterUpdateMode updateMode, - ParameterType parameterType, - const std::vector& parameterSegments, - int64_t numSamples, - real cost, - bool sendBackParameter, - ParameterType sendBackParameterType, - BatchStatus batchStatus, - SendJob* sendJob) { - sendJob->parallelRequests.resize(serviceNum_); - sendJob->parallelInputIovs.resize(serviceNum_); - - for (auto& request : sendJob->parallelRequests) { -#ifndef PADDLE_DISABLE_TIMER - if (updateMode == PSERVER_UPDATE_MODE_ADD_GRADIENT) { - request.set_forwardbackward_time(forwardbackwordTime_); - } -#endif - request.set_trainer_id(trainerId_); - request.set_update_mode(updateMode); - request.set_send_back_parameter(sendBackParameter); - request.set_send_back_parameter_type(sendBackParameterType); - request.set_num_samples(numSamples); - request.set_cost(cost); - request.set_batch_status(batchStatus); - CHECK_EQ(request.blocks_size(), 0); - VLOG(10) << "request: trainer_id: " << request.trainer_id() - << " update_mode" << request.update_mode() - << " send_back_parameter: " << request.send_back_parameter() - << " send_back_parameter_type: " - << request.send_back_parameter_type() - << " num_samples: " << request.num_samples() - << " cost: " << request.cost() - << " batch_status: " << request.batch_status(); - } - for (const auto& segments : parameterSegments) { - const auto it = parameterMap_.find(segments.id); - CHECK(it != parameterMap_.end()); - Parameter* parameter = it->second.get(); - CHECK(parameter != nullptr) << "parameter is nullptr"; - int64_t nameHash = std::hash()(segments.name); - bool sendingPara = !(updateMode == PSERVER_UPDATE_MODE_GET_PARAM || - updateMode == PSERVER_UPDATE_MODE_GET_PARAM_SPARSE || - updateMode == PSERVER_UPDATE_MODE_SET_PARAM_ZERO); - bool sparseUpdate = parameter->getConfig().sparse_remote_update() && - (updateMode == PSERVER_UPDATE_MODE_ADD_GRADIENT || - updateMode == PSERVER_UPDATE_MODE_ASYNC_SGD || - updateMode == PSERVER_UPDATE_MODE_GET_PARAM_SPARSE); - - const auto blockSize = parameter->getConfig().parameter_block_size(); - CHECK_GE(blockSize, 1LU) << "blockSize should > 0 " << blockSize; - const auto paraSize = parameter->getSize(); - if (sparseUpdate) { - auto prefetchMat = std::dynamic_pointer_cast( - parameter->getMat(PARAMETER_VALUE)); - CHECK(prefetchMat != nullptr) << "prefetchMat is nullptr"; - auto sendMat = dynamic_cast( - parameter->getMat(parameterType).get()); - CHECK(sendMat != nullptr) << "sendMat is nullptr"; - - syncThreadPool_->exec([&](int tid, size_t numThreads) { - std::lock_guard guard(sparseAutoGrowthMutex_); - const auto& localIndices = prefetchMat->getLocalIndices(); - /// num of sparse rows - size_t nLocalBlocks = localIndices.size(); - uint64_t beginDim = 0; - uint64_t endDim = 0; - - // HACK(typhoonzero): let it resize first - prefetchMat->getLocalRow(nLocalBlocks); - sendMat->getLocalRow(nLocalBlocks); - - for (size_t row = 0; row < nLocalBlocks; ++row) { - int64_t blockId = localIndices[row]; // local row -> sparse row - int serverId = std::abs((blockId + nameHash) % serviceNum_); - if (serverId % numThreads != (size_t)tid) { - continue; - } - - beginDim = blockId * blockSize; - endDim = std::min(beginDim + blockSize, paraSize); - - auto& request = sendJob->parallelRequests[serverId]; - ParameterBlock* block = request.add_blocks(); - block->set_para_id(segments.id); - /// global sparse row id - block->set_block_id(blockId); - /// local row offset - block->set_begin_pos(row * blockSize); - /// block len - block->set_block_size(endDim - beginDim); - if (sendingPara) { - sendJob->parallelInputIovs[serverId].push_back( - {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize}); - /// detect sparse parameter distribution - sparseDistribution_->probeDistribution(serverId, - sizeof(real) * blockSize); - } - } - }); - - } else { /// parameter set for dense and sparse - real* buf = - sendingPara ? parameter->getBuf(parameterType)->getPoint(0) : nullptr; - uint64_t endDim = 0; - for (uint64_t beginDim = 0; beginDim < paraSize; beginDim = endDim) { - endDim = std::min(beginDim + blockSize, paraSize); - int64_t blockId = beginDim / blockSize; - int serverId = std::abs((blockId + nameHash) % serviceNum_); - - auto& request = sendJob->parallelRequests[serverId]; - ParameterBlock* block = request.add_blocks(); - block->set_para_id(segments.id); - block->set_block_id(blockId); - block->set_begin_pos(beginDim); - block->set_block_size(endDim - beginDim); - if (buf) { - sendJob->parallelInputIovs[serverId].push_back( - {buf + beginDim, sizeof(real) * ((size_t)(endDim - beginDim))}); - } - } - } - } // parameterSegments - - sparseDistribution_->checkAndResetDistribution(); -} - -void ParameterClient2::sendAndReceiveParameter( - ParameterUpdateMode updateMode, - ParameterType parameterType, - const std::vector& parameterSegments, - int64_t numSamples, - real cost, - bool sendBackParameter, - ParameterType sendBackParameterType, - ParameterType recvParameterType) { - prepareSendData(updateMode, - parameterType, - parameterSegments, - numSamples, - cost, - sendBackParameter, - sendBackParameterType, - /*batchStatus = */ BATCH_START_AND_FINISH, - &sendJob_); - - syncThreadPool_->exec([&](int tid, size_t numThreads) { - this->sendParallel(tid, numThreads, recvParameterType); - }); -} - -void ParameterClient2::sendParameter( - ParameterUpdateMode updateMode, - ParameterType parameterType, - const std::vector& parameterSegments, - int64_t numSamples, - real cost, - bool sendBackParameter, - BatchStatus batchStatus) { - SendJobPtr sendJob = std::make_shared(); - prepareSendData(updateMode, - parameterType, - parameterSegments, - numSamples, - cost, - sendBackParameter, - PARAMETER_VALUE, - batchStatus, - sendJob.get()); - - for (int i = 0; i < threadNum_; i++) { - sendJobQueue_[i]->enqueue(sendJob); - } -} - -void ParameterClient2::recvParameter() { recvSyncBarrier_->wait(); } - -void ParameterClient2::send(int threadId) { - int index = threadId; - LOG(INFO) << "send thread " << threadId << " started"; - int numMyClients = divup(serviceNum_ - index, threadNum_); - while (true) { - SendJobPtr recvJob = sendJobQueue_[index]->dequeue(); - if (stopping_) { - recvJobQueue_[index]->enqueue(recvJob); - break; - } - for (int j = 0; j < numMyClients; ++j) { - REGISTER_TIMER("client_send"); - int i = threadNum_ * j + index; - /// Try to make different clients to send data to different pservers - /// at the same time so that they will not flood data to the same - /// pserver. - i = calcClientId(i, serviceNum_); - if (recvJob->parallelRequests.size()) { - clients_[i].send("sendParameter", - recvJob->parallelRequests[i], - recvJob->parallelInputIovs[i]); - } else { - clients_[i].send("sendData", - recvJob->parallelDataRequests[i], - recvJob->parallelInputIovs[i]); - } - } - recvJobQueue_[index]->enqueue(recvJob); - } -} - -void ParameterClient2::recv(int threadId) { - LOG(INFO) << "recv thread " << threadId << " started"; - int index = threadId; - int numMyClients = divup(serviceNum_ - index, threadNum_); - while (true) { - std::vector bufs; - SendParameterResponse response; - SendDataResponse dataResponse; - SendJobPtr recvJob = recvJobQueue_[index]->dequeue(); - if (stopping_) break; - for (int j = 0; j < numMyClients; ++j) { - REGISTER_TIMER("client_recv"); - int i = threadNum_ * j + index; - i = calcClientId(i, serviceNum_); - if (recvJob->parallelRequests.size()) { - auto msgReader = clients_[i].recv(&response); - CHECK_EQ(msgReader->getNumBlocks(), (size_t)response.blocks_size()); - bufs.clear(); - bufs.reserve(response.blocks_size()); - for (auto& block : response.blocks()) { - auto it = parameterMap_.find(block.para_id()); - CHECK(it != parameterMap_.end()); - Parameter* parameter = it->second.get(); - real* buf = - parameter->getBuf(PARAMETER_VALUE)->getPoint(block.begin_pos()); - CHECK_EQ(msgReader->getBlockLength(bufs.size()), - sizeof(real) * (block.block_size())); - bufs.push_back(buf); - } - msgReader->readBlocks(bufs); - } else { - auto msgReader = clients_[i].recv(&dataResponse); - CHECK_EQ(msgReader->getNumBlocks(), (size_t)dataResponse.blocks_size()); - size_t totalLen = msgReader->getTotalLength(); - if (0 == totalLen) { - continue; - } - auto& recvMem = recvDataMems_[dataResponse.server_id()]; - CHECK_EQ(dataResponse.blocks_size(), 1) - << "Only one block currently support now!"; - auto& block = dataResponse.blocks(0); - CHECK_EQ(totalLen % sizeof(block.data_size()), 0U); - recvMem = std::make_shared(totalLen); - msgReader->readNextBlock(recvMem.get()->getBuf()); - } - } - recvSyncBarrier_->wait(); - } -} - -void ParameterClient2::waitPassStart() { - WaitPassStartRequest request; - std::vector responses; - multiCall(__func__, request, &responses); -} - -void ParameterClient2::waitPassFinish() { - WaitPassFinishRequest request; - std::vector responses; - multiCall(__func__, request, &responses); -} - -void ParameterClient2::synchronize(SyncObject syncObjectId) { - SynchronizeRequest request; - request.set_sync_object_id(syncObjectId); - std::vector responses; - multiCall(__func__, request, &responses); -} - -void ParameterClient2::asyncFinishPass(SyncObject syncObjectId) { - SynchronizeRequest request; - request.set_sync_object_id(syncObjectId); - request.set_trainer_id(trainerId_); - std::vector responses; - multiCall(__func__, request, &responses); -} - -void ParameterClient2::setConfig(const OptimizationConfig& optConfig, - const std::string& saveDir, - bool isSparseServer) { - SetConfigRequest request; - std::vector responses; - - for (auto& nameAndPara : parameterMap_) { - *request.add_param_configs() = nameAndPara.second->getConfig(); - } - - *request.mutable_opt_config() = optConfig; - request.set_save_dir(saveDir); - request.set_is_sparse_server(isSparseServer); - - std::vector requests; - requests.resize(clients_.size()); - for (size_t i = 0; i < requests.size(); ++i) { - requests[i].CopyFrom(request); - requests[i].set_server_id(i); - } - - responses.resize(clients_.size()); - size_t numClients = clients_.size(); - for (size_t i = 0; i < numClients; ++i) { - clients_[i].send(__func__, requests[i]); - } - for (size_t i = 0; i < numClients; ++i) { - clients_[i].recv(&responses[i]); - } -} - -bool ParameterClient2::inStatus(PServerStatus status) { - GetStatusRequest request; - std::vector responses; - - bool ok = true; - multiCall("getStatus", request, &responses); - for (auto& response : responses) { - if (response.status() != status) { - ok = false; - } - } - - return ok; -} - -void ParameterClient2::setStatus(PServerStatus status) { - SetStatusRequest request; - request.set_status(status); - std::vector responses; - multiCall(__func__, request, &responses); -} - -void ParameterClient2::waitForStatus(PServerStatus status) { - while (!inStatus(status)) { - sleep(1); - } -} - -template -static void validateResponses(const std::vector& responses) { - for (auto& response : responses) { - CHECK(response.return_message().empty()) - << "client" << &response - &responses[0] - << " error:" << response.return_message(); - } -} - -PServerVector ParameterClient2::createVector() { - CreateVectorRequest request; - std::vector responses; - int64_t handle = -1; - - multiCall(__func__, request, &responses); - validateResponses(responses); - - for (auto& response : responses) { - if (handle == -1) { - handle = response.handle(); - } else { - CHECK_EQ(handle, response.handle()) << "Inconsistent handle from client" - << &response - &responses[0] << " " - << handle << " " << response.handle(); - } - } - return PServerVector{handle}; -} - -void ParameterClient2::releaseVector(PServerVector handle) { - ReleaseVectorRequest request; - std::vector responses; - - request.set_handle(handle.handle); - multiCall(__func__, request, &responses); - validateResponses(responses); -} - -PServerMatrix ParameterClient2::createMatrix(int32_t numCols) { - CreateMatrixRequest request; - std::vector responses; - int64_t handle = -1; - - request.set_num_cols(numCols); - multiCall(__func__, request, &responses); - validateResponses(responses); - - for (auto& response : responses) { - if (handle == -1) { - handle = response.handle(); - } else { - CHECK_EQ(handle, response.handle()) << "Inconsistent handle from client" - << &response - &responses[0] << " " - << handle << " " << response.handle(); - } - } - return PServerMatrix{handle}; -} - -void ParameterClient2::releaseMatrix(PServerMatrix handle) { - ReleaseMatrixRequest request; - std::vector responses; - - request.set_handle(handle.handle); - multiCall(__func__, request, &responses); - validateResponses(responses); -} - -void PreparedOperations::addOperationHelper(Operation* op, CpuVectorPtr vec) { - ProtoVector& pvec = *op->add_vectors(); - size_t dim = vec->getSize(); - pvec.set_dim(dim); - copyToRepeatedField(pvec.mutable_values(), vec->getData(), vec->getSize()); -} - -void PreparedOperations::addOperationHelper(Operation* op, CpuMatrixPtr mat) { - ProtoMatrix& pmat = *op->add_matrices(); - pmat.set_num_cols(mat->getWidth()); - pmat.set_num_rows(mat->getHeight()); - copyToRepeatedField( - pmat.mutable_values(), mat->getData(), pmat.num_cols() * pmat.num_rows()); -} - -static inline real addTwo(real a, double b) { return a + b; } - -void ParameterClient2::doOperation(PreparedOperations& ops, - bool waitForGradient, - bool sendBackGradient, - bool releasePass) { - std::vector responses; - ops.request_.set_wait_for_gradient(waitForGradient); - ops.request_.set_send_back_parameter(sendBackGradient); - ops.request_.set_release_pass(releasePass); - multiCall(__func__, ops.request_, &responses); - validateResponses(responses); - size_t numPassFinishServers = 0; - - size_t numOps = ops.request_.operations_size(); - for (auto& response : responses) { - numPassFinishServers += response.pass_finish(); - CHECK_EQ(numOps, (size_t)response.results_size()); - for (size_t opId = 0; opId < numOps; ++opId) { - const OperationResult& result = response.results(opId); - std::vector& resultScalars = ops.localResults_[opId].resultScalars; - std::vector& resultVectors = - ops.localResults_[opId].resultVectors; - std::vector& resultMatrices = - ops.localResults_[opId].resultMatrices; - - if (&response == &responses[0]) { - /// Initialize results to zero - - resultScalars.resize(result.scalars_size()); - for (auto p : resultScalars) { - if (!p) continue; - *p = 0; - } - size_t numVectors = result.vectors_size(); - resultVectors.resize(numVectors); - for (size_t i = 0; i < numVectors; ++i) { - if (!resultVectors[i]) continue; - resultVectors[i]->resize(result.vectors(i).dim()); - resultVectors[i]->zeroMem(); - } - size_t numMatrices = result.matrices_size(); - resultMatrices.resize(numMatrices); - for (size_t i = 0; i < numMatrices; ++i) { - if (!resultMatrices[i]) continue; - resultMatrices[i]->resize(result.matrices(i).num_rows(), - result.matrices(i).num_cols()); - resultMatrices[i]->zeroMem(); - } - } - - // aggregate results from each pserver to results - - CHECK_EQ(resultScalars.size(), (size_t)result.scalars_size()); - for (ssize_t i = 0; i < result.scalars_size(); ++i) { - real* rscalar = resultScalars[i]; - if (!rscalar) continue; - *rscalar += result.scalars(i); - } - - CHECK_EQ(resultVectors.size(), (size_t)result.vectors_size()); - for (auto& vec : result.vectors()) { - int i = &vec - &result.vectors(0); - CpuVectorPtr rvec = resultVectors[i]; - if (!rvec) continue; - CHECK_EQ(rvec->getSize(), (size_t)vec.dim()); - std::transform(rvec->getData(), - rvec->getData() + rvec->getSize(), - vec.values().data(), - rvec->getData(), - addTwo); - } - - CHECK_EQ(resultMatrices.size(), (size_t)result.matrices_size()); - for (auto& mat : result.matrices()) { - int i = &mat - &result.matrices(0); - CpuMatrixPtr rmat = resultMatrices[i]; - if (!rmat) continue; - CHECK_EQ(rmat->getHeight(), (size_t)mat.num_rows()); - CHECK_EQ(rmat->getWidth(), (size_t)mat.num_cols()); - - std::transform(rmat->getData(), - rmat->getData() + rmat->getElementCnt(), - mat.values().data(), - rmat->getData(), - addTwo); - } - } - } - passFinish_ = numPassFinishServers == clients_.size(); -} - -real ParameterClient2::vectorDotProduct(PServerVector u, PServerVector v) { - real result = 0.0; - PreparedOperations ops; - ops.addOperation(PSERVER_OP_utv, u, v)(&result); - doOperation(ops, false, false); - return result; -} - -void ParameterClient2::vectorScale(PServerVector u, real a) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_au, u, a); - doOperation(ops, false, false); -} - -void ParameterClient2::vectorCopy(PServerVector src, PServerVector dst) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_COPY, src, dst); - doOperation(ops, false, false); -} - -void ParameterClient2::vectorAddMult(PServerVector u, PServerVector v, real a) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_au_bv, v, u, a, (real)1); - doOperation(ops, false, false); -} - -void ParameterClient2::vectorAddMultInto(PServerVector u, - PServerVector v, - PServerVector w, - real a) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_au_bv_cw, v, w, u, (real)1, a, (real)0); - doOperation(ops, false, false); -} - -void ParameterClient2::vectorScaleInto(PServerVector u, - PServerVector v, - real a) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_au_bv, v, u, a, (real)0); - doOperation(ops, false, false); -} - -void ParameterClient2::loadValueVector(const std::string& dirName) { - LoadValueRequest request; - request.set_dir_name(dirName); - std::vector responses; - - multiCall(__func__, request, &responses); - validateResponses(responses); -} - -void ParameterClient2::saveValueVector(const std::string& dirName) { - SaveValueRequest request; - request.set_dir_name(dirName); - std::vector responses; - - multiCall(__func__, request, &responses); - validateResponses(responses); -} - -} // namespace paddle diff --git a/paddle/pserver/ParameterClient2.h b/paddle/pserver/ParameterClient2.h deleted file mode 100644 index c96bb787151a525556c8217629109de201762cff..0000000000000000000000000000000000000000 --- a/paddle/pserver/ParameterClient2.h +++ /dev/null @@ -1,602 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include -#include - -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/pserver/BaseClient.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/Queue.h" -#include "paddle/utils/Util.h" - -#include "ParameterService.pb.h" - -#include "ProtoServer.h" -#include "SparseParameterDistribution.h" - -DECLARE_int32(parallel_thread_num); - -namespace paddle { - -struct PServerMatrix { - int64_t handle; -}; - -struct PServerVector { - int64_t handle; -}; - -/** - * @brief A class to help to prepare server-side operations. - */ -class PreparedOperations { - protected: - class ResultsAdder; - struct LocalOperationResult; - - public: - /** - * Offers an easy way to prepare operations that will be performed on - * server-side. - * - * Usage: - * @code - * addOperation(optype, arguments...)(results...) - * @endcode - * - * Examples: - * 1. set pserver vector to 1: - * @code - * PServerVector u = parameterClient.createVector(); - * addOperation(PSERVER_OP_RESET, u, (real)1); - * @endcode - * - * 2. Compute inner product of to pserver vectors. - * @code - * PServerVector u = parameterClient.createVector(); - * PServerVector v = parameterClient.createVector(); - * real result; - * addOperation(PSERVER_OP_utv, u, v)(&result) - * @endcode - * - * @param[in] operation The operation that pserver will perform. - * @param[in] args Argument list of the operation - * @return A ResultsAdder object initialized with the last element of - * localResults_. - */ - template - ResultsAdder addOperation(MatrixVectorOperation operation, Args... args) { - Operation* op = request_.add_operations(); - op->set_operation(operation); - localResults_.emplace_back(); - addOperationHelper(op, args...); - return ResultsAdder(&localResults_.back()); - } - - protected: - void addOperationHelper(Operation* op) {} - - /** - * @brief Helper function to add an new operation that takes a PServerVector - * as an operand. - */ - void addOperationHelper(Operation* op, PServerVector arg) { - op->add_pvectors(arg.handle); - } - - /** - * @brief Helper function to add an new operation that takes a PServerMatrix - * as an operand. - */ - void addOperationHelper(Operation* op, PServerMatrix arg) { - op->add_pmatrices(arg.handle); - } - - /** - * @brief Helper function to add an new operation that takes a real valued - * scalar as an operand. - */ - void addOperationHelper(Operation* op, real arg) { op->add_scalars(arg); } - - /** - * @brief Helper function to add an new operation that takes a CpuVectorPtr - * as an operand. - * @note The array of CpuVectors that arg points to will be copied to - * op's vectors field. - */ - void addOperationHelper(Operation* op, CpuVectorPtr arg); - - /** - * @brief Helper function to add an new operation that takes a CpuMatrixPtr - * as an operand. - * @note The array of CpuMatrixs that arg points to will be copied to - * op's matrices field. - */ - void addOperationHelper(Operation* op, CpuMatrixPtr arg); - - /** - * @brief Helper function to add an new operation and prepare the operands. - * - * @tparam Arg An operand of the operation. - * @tparam Args A list of rest operands of the operation. - * @param op Pointer to an Operation object. - */ - template - void addOperationHelper(Operation* op, Arg arg, Args... args) { - addOperationHelper(op, arg); - addOperationHelper(op, args...); - } - - /** - * @brief ResultsAdder offers easy ways to quickly store operation results. - */ - class ResultsAdder { - public: - explicit ResultsAdder(LocalOperationResult* localResult) - : localResult_(localResult) {} - template - void operator()(Args... args) { - addResult(args...); - } - void addResult() {} - void addResult(real* arg) { localResult_->resultScalars.push_back(arg); } - void AddResult(CpuVectorPtr arg) { - localResult_->resultVectors.push_back(arg); - } - void AddResult(CpuMatrixPtr arg) { - localResult_->resultMatrices.push_back(arg); - } - template - void addResult(Arg arg, Args... args) { - addResult(arg); - addResult(args...); - } - - protected: - LocalOperationResult* localResult_; - }; - - protected: - DoOperationRequest request_; - std::vector inputIovs_; - struct LocalOperationResult { - std::vector resultScalars; - std::vector resultVectors; - std::vector resultMatrices; - }; - std::vector localResults_; - friend class ParameterClient2; -}; - -struct ParameterSegments { - std::string name; // name of the parameter - size_t id; // id of the parameter -}; - -/** - * The client interface for parameter server. ParameterClient2 supports 2 modes - * for managing connections to parameter servers, in the 1st mode one connection - * is shared by 2 threads that are separately responsible for sending and - * recieving activities, in the 2nd mode one connection is owned by only one - * thread, and all the sending and recieving activities run in that single - * thread. - * TODO(yanfei): - * Additional core idea to further optimizate pserver performance is - * to do sync-sgd based parameter level instead of pserver level. - * full-parallelization based parameter level for sync-sgd also can - * sense forwardbackward computation layer-by-layer for more deeper layer - * model. - * Firstly, pserver can do full-parallelization on all computation based - * parameter level instead of waiting for all gradients are finished and - * start to send back parameters value immediately if parameter is ready - * instead of waiting for all parameters value are ready - * Secondly, parameter client can write back parameters to GPU instead of - * waiting until all parameters are received to CPU host end. - */ -class ParameterClient2 : public BaseClient { - public: - /** Constructor. - * @param separate True if sending and recieving activities are separated - * into 2 threads, otherwise false. - * @param port Port number that parameter client runs on. - * @param numPorts Number of ports parameter clients occupies, - * numPorts * pserver number is the total number of - * connections the parameter client maintains. - */ - ParameterClient2(bool separate = false, - int port = FLAGS_port, - int numPorts = FLAGS_ports_num); - - ~ParameterClient2(); - - static int calcParameterBlockSize(const std::vector& parameters, - size_t serviceNum); - - public: - bool init(const std::vector& parameters); - - /// service functions - - /** - * @brief Sends the segments in parameter to parameter servers, then receives - * the response from the servers. - * @param[in] updateMode Indicates how parameters should be updated on the - * server side. - * @param[in] parameterType Type of parameter that will be sent. - * @param[in] segments Segments in the parameter that will be sent. - * @param[in] numSamples Number of samples this update is based on. - * @param[in] cost Cost of the batch, will be used to calculate global object - * value. - * @param[in] sendBackParameter True if the updated parameters should be sent - * back, otherwise false. - * @param[in] sendBackParameterType Send back parameter type on pserver, - * PARAMETER_VALUE by default - * @param[in] recvParameterType pserver[sendBackParameterType] will be copy to - * client[recvParameterType] - * @note Only parameterType will be sent. - */ - void sendAndReceiveParameter(ParameterUpdateMode updateMode, - ParameterType parameterType, - const std::vector& segments, - int64_t numSamples, - real cost, - bool sendBackParameter, - ParameterType sendBackParameterType, - ParameterType recvParameterType); - - /** - * @brief Sends all parameters to parameter servers, and receives the response - * from the servers. - */ - void sendAndReceiveParameter( - ParameterUpdateMode updateMode, - ParameterType parameterType, - int64_t numSamples, - real cost, - bool sendBackParameter, - ParameterType sendBackParameterType = PARAMETER_VALUE, - ParameterType recvParameterType = PARAMETER_VALUE) { - sendAndReceiveParameter(updateMode, - parameterType, - allSegments_, - numSamples, - cost, - sendBackParameter, - sendBackParameterType, - recvParameterType); - } - - /** - * @brief Sends the segments in parameter to parameter servers. Each - * sendParameter() must be paired with a recvParameter() in the future. - * Only parameterType will be sent. - * - * @param[in] updateMode Indicates how parameters should be updated on the - * server side. - * @param[in] parameterType Type of parameter that will be sent. - * @param[in] segments Segments in the parameter that will be sent. - * @param[in] numSamples Number of samples this update is based on. - * @param[in] cost Cost of the batch, will be used to calculate global object - * value. - * @param[in] sendBackParameter True if the updated parameters should be sent - * back, otherwise false. - * @param[in] batchStatus Status of the batch. - * @note This function is non-blocking. This means that parameter should - * not change between this call and recvParameter() - */ - void sendParameter(ParameterUpdateMode updateMode, - ParameterType parameterType, - const std::vector& segments, - int64_t numSamples, - real cost, - bool sendBackParameter, - BatchStatus batchStatus); - - void recvParameter(); - - /** - * Sends all parameters to parameter servers, recvParameter() have to be - * invoked - * afterwards. - * - * @note This function is non-blocking. This means that if parameter should - * not changes between this call and recvParameter() - */ - void sendParameter(ParameterUpdateMode updateMode, - ParameterType parameterType, - int64_t numSamples, - real cost, - bool sendBackParameter, - BatchStatus batchStatus) { - sendParameter(updateMode, - parameterType, - allSegments_, - numSamples, - cost, - sendBackParameter, - batchStatus); - } - - /// Get all parameters from parameter servers - void getParameter(ParameterType recvParameterType = PARAMETER_VALUE, - ParameterType sendBackParameterType = PARAMETER_VALUE) { - sendAndReceiveParameter(PSERVER_UPDATE_MODE_GET_PARAM, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true, // sendBackParameter = true - sendBackParameterType, - recvParameterType); - } - - /// Get parameters by sparse row ids from parameter servers - void getParameterSparse( - ParameterType recvParameterType = PARAMETER_VALUE, - ParameterType sendBackParameterType = PARAMETER_VALUE) { - sendAndReceiveParameter(PSERVER_UPDATE_MODE_GET_PARAM_SPARSE, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true, // sendBackParameter = true - sendBackParameterType, - recvParameterType); - } - - /// Set all parameters on parameter servers using the local parameters - void setParameter() { - sendAndReceiveParameter(PSERVER_UPDATE_MODE_SET_PARAM, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - false); // sendBackParameter = false - } - /** - * Set all parameters on parameter servers, values will be zero - * means do not sending local parameters - */ - void setParameterZero() { - sendAndReceiveParameter(PSERVER_UPDATE_MODE_SET_PARAM_ZERO, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - false); // sendBackParameter = false - } - - /** - * @brief Wait until all gradient servers start one pass. - * - * @note This is now only used by the gradient servers for "sgd" - * algorithm. Calling this function means that the calling gradient - * server is ready to start a new pass. - */ - void waitPassStart(); - - /** - * @brief Wait until all gradient servers finish one pass. - * - * @note This is now only used by the gradient servers for "sgd" algorithm. - * Calling this function means that the calling gradient server - * finishes one pass. - */ - void waitPassFinish(); - - /// Wait until all gradient servers call this function. - void synchronize(SyncObject syncObjectId = SYNC_DEFAULT); - - /// Called when async-sgd finish pass. - void asyncFinishPass(SyncObject syncObjectId = SYNC_DEFAULT); - - void asyncStartPass(SyncObject syncObjectId = SYNC_DEFAULT) { - return synchronize(syncObjectId); - } - - /** - * @brief Execute the prepared operations on pservers, fetch the results and - * aggregate results from different pservers. - * @param[in] ops Prepared operations that will be executed on pservers. - * @param[in] waitForGradient If true, wait for gradient to be ready before - * starting the operations. - * @param[in] sendBackParameter If true, send back the parameter to clients - * after the operations are finished. - * @param[in] If true, and if all clients call waitPassFinish, signal all - * clients finish the pass. - */ - void doOperation(PreparedOperations& ops, - bool waitForGradient, - bool sendBackParameter, - bool releasePass = true); - - /** - * Set the configuration of pserver, including parameter config and - * optimization config - */ - void setConfig(const OptimizationConfig& optConfig, - const std::string& saveDir = "", - bool isSparseServer = false); - - /// Return true if all pservers are in the given status - bool inStatus(PServerStatus status); - bool isPassFinish() { return passFinish_; } - - /// Set pserver status - void setStatus(PServerStatus status); - - /** - * @brief Wait until all pservers are at status - * @note This function is not suitable for frequent use, - * because it sleeps 1 second each time when condition is satisfied. - */ - void waitForStatus(PServerStatus status); - - /// Create a column vector. The size is the dimension of parameter. - PServerVector createVector(); - - /// Release the PServerVector given handle. - void releaseVector(PServerVector handle); - - /** - * Create a column major matrix. The number of rows is the dimension of - * parameter. The number of columns is specifed by numCols. - */ - PServerMatrix createMatrix(int32_t numCols); - - /// Release the PServerMatrix given handle. - void releaseMatrix(PServerMatrix handle); - - // Some basic algebra functions - /// Calculate the dot product of u and v - real vectorDotProduct(PServerVector u, PServerVector v); - - /// Scale u by a - void vectorScale(PServerVector u, real a); - - /// Copy from src to dest - void vectorCopy(PServerVector src, PServerVector dst); - - /// u += v * a - void vectorAddMult(PServerVector u, PServerVector v, real a); - - /// u = v + w * a - void vectorAddMultInto(PServerVector u, - PServerVector v, - PServerVector w, - real a); - /// u = v * a - void vectorScaleInto(PServerVector u, PServerVector v, real a); - - /// Return pserver parameter value. - PServerVector getPServerParameterValue() { - PServerVector vec; - vec.handle = PARAMETER_VALUE; - return vec; - } - - /// Return pserver parameter gradient. - PServerVector getPServerParameterGradient() { - PServerVector vec; - vec.handle = PARAMETER_GRADIENT; - return vec; - } - - /** - * Tell pservers to load value vector from file. - * - * @param[in] dirName The directory that contains the value vector file. - */ - void loadValueVector(const std::string& dirName); - - /// Tell pservers to save value vector to file. - void saveValueVector(const std::string& dirName); - - void setTrainerId(int trainerId) { trainerId_ = trainerId; } - -#ifndef PADDLE_DISABLE_TIMER - void setForwardbackwardTime(uint64_t delta) { forwardbackwordTime_ = delta; } -#endif - - protected: - template - void multiCall(const char* funcName, - const ProtoIn& request, - std::vector* responses) { - responses->resize(clients_.size()); - size_t numClients = clients_.size(); - for (size_t i = 0; i < numClients; ++i) { - clients_[i].send(funcName, request); - } - for (size_t i = 0; i < numClients; ++i) { - clients_[i].recv(&(*responses)[i]); - } - } - - private: - void destroy(); - - /** - * @brief management function for parallelizing send/recv all connections - * to all pservers. it is called under one SyncThreadPool. it - * supports to use N thread to control M connections. the receiving - * actions can be started until all sending action to all connections - * owned by current thread are finished. Different connections - * controlled - * by different threads can transfer data asynchronously. - */ - void sendParallel(int tid, - size_t numThreads, - ParameterType recvParameterType); - /// sending thread routine for asynchronously send data - void send(int threadId); - /// receiving thread routing for asynchronously receive data - void recv(int threadId); - - /** - * @brief main routine to build data for pserver - * - * @note it can prepare different kinds of parameter type data. it can - * be regarded as layer for bridging real parameters data and - * protobuf data for communication. - * TODO(yanfei): - * can abstract additional layer to encode and decode data to/from - * protobuf data. - */ - void prepareSendData( - ParameterUpdateMode updateMode, - ParameterType parameterType, // client send type - const std::vector& parameterSegments, - int64_t numSamples, - real cost, - bool sendBackParameter, - ParameterType sendBackParameterType, // send back type in pserver - BatchStatus batchStatus, - SendJob* sendJob); - - /// start necessary threads for threadPool - void initThreads(); - - protected: - /// start port number of pserver - /// it deduce all ports for dense and sparse with some rules - int port_; - /// identify the trainer id using this client - int trainerId_; - -#ifndef PADDLE_DISABLE_TIMER - uint64_t forwardbackwordTime_; -#endif - std::mutex sparseAutoGrowthMutex_; - - /// map id to parameter used for decoding protobuf data - std::unordered_map parameterMap_; - /// segments for all parameters that needed to sync - std::vector allSegments_; - - /// module for sensing sparse parameters distribution on all pservers - std::unique_ptr sparseDistribution_; - - /// thread pool for parallelizing all connections to pservers - std::unique_ptr syncThreadPool_; - - bool passFinish_; -}; - -} // namespace paddle diff --git a/paddle/pserver/ParameterServer2.cpp b/paddle/pserver/ParameterServer2.cpp deleted file mode 100644 index f8814714c29a9776adde6a979a84241f733f65bd..0000000000000000000000000000000000000000 --- a/paddle/pserver/ParameterServer2.cpp +++ /dev/null @@ -1,1401 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ParameterServer2.h" - -#include -#include - -#include "paddle/math/SIMDFunctions.h" -#include "paddle/parameter/AverageOptimizer.h" -#include "paddle/parameter/FirstOrderOptimizer.h" -#include "paddle/parameter/OptimizerFunctions.h" -#include "paddle/parameter/OptimizerWithRegularizer.h" -#include "paddle/parameter/ParameterOptimizer.h" -#include "paddle/parameter/ParameterUpdateFunctions.h" -#include "paddle/parameter/Regularizer.h" -#include "paddle/parameter/ThreadLocalBuffer.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/StringUtil.h" - -DEFINE_int32(pserver_num_threads, 1, "number of threads for sync op exec"); -DEFINE_double(async_lagged_ratio_min, - 1.0, - "control config_.async_lagged_grad_discard_ratio() min value"); -DEFINE_double( - async_lagged_ratio_default, - 1.5, - "if async_lagged_grad_discard_ratio is not set in trainer_config.conf" - "use it as defalut value"); - -namespace paddle { - -const std::string ParameterServer2::kRetMsgInvalidMatrixHandle = - "Invalid matrix handle"; -const std::string ParameterServer2::kRetMsgInvalidVectorHandle = - "Invalid vector handle"; -const std::string ParameterServer2::kRetMsgUnknownOperation = - "Unknown operation"; - -ParameterServer2::ParameterServer2(const std::string& addr, - int port, - int rdmaCpu) - : ProtoServer(addr, port, rdmaCpu), - dataSize_(0), - size_(0), - gradientReadyBarrier_(FLAGS_num_gradient_servers + 1), - parameterReadyBarrier_(FLAGS_num_gradient_servers + 1), - passBarrier_(FLAGS_num_gradient_servers + 1), - numPassFinishClients_(0), - allClientPassFinish_(false), - serverId_(-1), - batchId_(-1) { - /** - * register function for remote client calling, these functions - * will be mapped to a data structure for quick looking up. each - * request from trainer can contains one function name to indicate - * remote action. this architecture looks like rpc style for pserver. - */ - REGISTER_SERVICE_FUNCTION_EX(ParameterServer2, sendParameter); - REGISTER_SERVICE_FUNCTION_EX(ParameterServer2, sendData); - REGISTER_SERVICE_FUNCTION(ParameterServer2, setConfig); - REGISTER_SERVICE_FUNCTION(ParameterServer2, setStatus); - REGISTER_SERVICE_FUNCTION(ParameterServer2, getStatus); - REGISTER_SERVICE_FUNCTION(ParameterServer2, doOperation); - REGISTER_SERVICE_FUNCTION(ParameterServer2, createVector); - REGISTER_SERVICE_FUNCTION(ParameterServer2, releaseVector); - REGISTER_SERVICE_FUNCTION(ParameterServer2, createMatrix); - REGISTER_SERVICE_FUNCTION(ParameterServer2, releaseMatrix); - REGISTER_SERVICE_FUNCTION(ParameterServer2, waitPassStart); - REGISTER_SERVICE_FUNCTION(ParameterServer2, waitPassFinish); - REGISTER_SERVICE_FUNCTION(ParameterServer2, synchronize); - REGISTER_SERVICE_FUNCTION(ParameterServer2, asyncFinishPass); - REGISTER_SERVICE_FUNCTION(ParameterServer2, loadValueVector); - REGISTER_SERVICE_FUNCTION(ParameterServer2, saveValueVector); - - /// thread pool for parallelizing some computations - if (FLAGS_pserver_num_threads > 1) { - syncThreadPool_.reset(new SyncThreadPool(FLAGS_pserver_num_threads, false)); - } -} - -bool ParameterServer2::init() { - vectors_.resize(NUM_PARAMETER_TYPES); - configMap_.clear(); - - numSamplesProcessed_ = 0; - cost_ = 0; - char* mpienv = getenv("OMPI_COMM_WORLD_SIZE"); - if (mpienv != NULL) { - mpiSize_ = atoi(mpienv); - } else { - mpiSize_ = 1; - } - status_ = PSERVER_STATUS_NOT_SET; - dataMems_.resize(FLAGS_num_gradient_servers); - synchronizeBarriers_.resize(SyncObject_ARRAYSIZE); - for (auto& barrier : synchronizeBarriers_) { - barrier.reset(new ThreadBarrier(FLAGS_num_gradient_servers)); - } - - // initialization for dicarding lagging gradient - asyncUpdateSteps_ = 0; - asyncTrainerSteps_.resize(FLAGS_num_gradient_servers); - asyncTrainerSteps_.assign(asyncTrainerSteps_.size(), 0); - asyncLaggedGradientsNum_ = 0; - asyncUpdateStat_.resize(static_cast(FLAGS_num_gradient_servers * - FLAGS_async_lagged_ratio_default)); - asyncUpdateStat_.assign(asyncUpdateStat_.size(), 0); - asyncTrainerDiscardStat_.resize(FLAGS_num_gradient_servers); - asyncTrainerDiscardStat_.assign(asyncTrainerDiscardStat_.size(), 0); - asyncTrainerCommitStat_.resize(FLAGS_num_gradient_servers); - asyncTrainerCommitStat_.assign(asyncTrainerCommitStat_.size(), 0); - - return true; -} - -void ParameterServer2::getStatus(const GetStatusRequest& request, - ProtoResponseCallback callback) { - (void)request; - GetStatusResponse response; - response.set_status(status_); - callback(response); -} - -void ParameterServer2::setStatus(const SetStatusRequest& request, - ProtoResponseCallback callback) { - status_ = request.status(); - SetStatusResponse response; - callback(response); -} - -void ParameterServer2::setConfig(const SetConfigRequest& request, - ProtoResponseCallback callback) { - { - std::lock_guard guard(parameterMutex_); - - serverId_ = request.server_id(); - isSparseServer_ = request.is_sparse_server(); - - if (!request.save_dir().empty()) { - mkDir(request.save_dir().c_str()); - } - - for (const auto& config : request.param_configs()) { - CHECK(!configMap_.count(config.para_id())) - << "Duplicated parameter name: " << config.name(); - configMap_[config.para_id()] = config; - CHECK_EQ(config.sparse_remote_update(), isSparseServer_); - } - - config_ = request.opt_config(); - if (config_.algorithm() == TrainAlgorithm::AsyncSGD) { - auto asyncLaggedRatio = config_.async_lagged_grad_discard_ratio(); - if (asyncLaggedRatio <= FLAGS_async_lagged_ratio_min) { - LOG(INFO) << "WARNING: async_lagged_grad_discard_ratio is too small" - << "reset to default, async_lagged_grad_discard_ratio = " - << FLAGS_async_lagged_ratio_default; - asyncLaggedRatio = FLAGS_async_lagged_ratio_default; - } - asyncLaggedThreshold_ = - static_cast(FLAGS_num_gradient_servers * asyncLaggedRatio); - LOG(INFO) << "discard lagged async gradient ratio: " << asyncLaggedRatio - << " asyncLaggedhreshold: " << asyncLaggedThreshold_; - } - if (isSparseServer_ && config_.num_batches_per_send_parameter() > 1) { - /// sparse server must NOT use local update mode - config_.set_num_batches_per_send_parameter(1); - } - - if (config_.num_batches_per_send_parameter() > 1 && - config_.center_parameter_update_method() == "average") { - /// scaling L1/L2 decay rate as large as L1/L2 apply in trainer - /// if parameter regularization in pserver - for (auto& pair : configMap_) { - ParameterConfig& config = pair.second; - if (config_.num_batches_per_send_parameter() == - config.num_batches_regularization()) { - real scale = - config_.delta_add_rate() * config.num_batches_regularization(); - if (config_.algorithm() == "sgd") { - scale *= FLAGS_num_gradient_servers; - } - config.set_decay_rate(config.decay_rate() * scale); - if (config.decay_rate() > 0.1f) { - LOG(FATAL) << "L2 decay=" << config.decay_rate() - << " for parameter:" << config.name() - << " is too large after scale in pserver!"; - } - config.set_decay_rate_l1(config.decay_rate_l1() * scale); - if (config.decay_rate_l1() > 0.1f) { - LOG(FATAL) << "L1 decay=" << config.decay_rate_l1() - << " for parameter:" << config.name() - << " is too large after scale in pserver!"; - } - - LOG(INFO) << "parameter:" << config.name() - << " decay apply in pserver," - << " L1 decay=" << config.decay_rate_l1() - << " L2 decay=" << config.decay_rate(); - } - } - } - } - - SetConfigResponse response; - callback(response); -} - -real bufferSum(const std::vector& buffers) { - real sum = 0; - for (const auto buffer : buffers) { - for (size_t i = 0; i < buffer.size; ++i) { - sum += buffer.base[i]; - } - } - return sum; -} - -void ParameterServer2::mergeSegments(BlockSegments* segments) { - if (segments->empty()) { - return; - } - std::sort(segments->begin(), segments->end()); - auto curr = segments->begin(); - for (auto it = segments->begin(); it != segments->end(); ++it) { - if (it->first <= curr->second) { - curr->second = std::max(curr->second, it->second); - } else { - ++curr; - *curr = *it; - } - } - ++curr; - segments->erase(curr, segments->end()); -} - -void ParameterServer2::setParameter(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers) { - (void)response; - (void)outputBuffers; - LOG(INFO) << "pserver: setParameter"; - std::lock_guard guard(parameterMutex_); - - int64_t numBlocks = blockIdMap_.size(); - CHECK_EQ(blockIdMap_.size(), blockOffsetMap_.size()); - /// total bytes for all the added blocks - int64_t totalSize = size_; - std::vector offsets; - offsets.reserve(request.blocks_size()); - std::vector blockIds; - blockIds.reserve(request.blocks_size()); - int bufferIndex = 0; - - if (!request.blocks().size()) { - LOG(WARNING) - << "--ports_num or --ports_num_for_sparse might be too large, " - << "or total dense parameter size or sparse parameters size " - << "might be too small, this psever doesn't store any parameter."; - return; - } - - for (const auto& block : request.blocks()) { - /// block size for parameter(e.g. 128 for sparse row, 1K for dense) - uint64_t blockSize = getParameterConfig(block).parameter_block_size(); - BlockKey key(block.para_id(), block.block_id()); - if (inputBuffers.size()) { // if !=PSERVER_UPDATE_MODE_SET_PARAM_ZERO - Buffer buffer = inputBuffers[bufferIndex]; - ++bufferIndex; - CHECK_EQ(buffer.size, block.block_size()) - << "data size is too big:" - << " block_size=" << block.block_size() - << " data_size=" << buffer.size; - } - - /// add a new block - if (blockIdMap_.count(key) == 0) { - blockOffsetMap_[key] = totalSize; - blockIdMap_[key] = numBlocks; - ++numBlocks; - totalSize += blockSize; - } - offsets.push_back(blockOffsetMap_[key]); - blockIds.push_back(blockIdMap_[key]); - } - - size_ = totalSize; - LOG(INFO) << "pserver: new cpuvector: size=" << size_; - if (!vectors_[PARAMETER_VALUE]) { - /// vectors_ - const auto types = sgdOptimizerGetTypes(config_, true /*inPserver*/); - for (const auto type : types) { - vectors_[type].reset(new CpuVector(size_)); - vectors_[type]->zeroMem(); - } - - blockInfos_.resize(numBlocks); - for (auto& info : blockInfos_) { - info.lock.reset(new std::mutex()); - } - } else { - CHECK_EQ((size_t)size_, vectors_[PARAMETER_VALUE]->getSize()) - << "Currently adding new blocks is not supported. " - << "All blocks must be added in one setParameter call"; - } - - VectorPtr buf = vectors_[PARAMETER_VALUE]; - usedSegments_.reserve(offsets.size()); - /// if offsets is empty, means parameter_block_size is too big or too many - /// nodes. - if (offsets.empty()) { - LOG(WARNING) << "in setParameter: offsets is empty"; - } - for (size_t i = 0; i < offsets.size(); ++i) { - size_t blockId = blockIds[i]; - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(request.blocks(i)); - info.config = &config; - info.offset = offsets[i]; - info.optimizer.reset(sgdOptimizerCreate( - config_, config, config.sparse_remote_update(), true /*inPserver*/)); - if (config.sparse_remote_update()) { - size_t width = config.dims(1); - CHECK_EQ(config.parameter_block_size(), width) - << "block size: " << config.parameter_block_size() - << "width : " << width; - } - info.optimizer->init(1, info.config); - usedSegments_.push_back(std::make_pair( - offsets[i], offsets[i] + request.blocks(i).block_size())); - } - mergeSegments(&usedSegments_); - - if (request.update_mode() == PSERVER_UPDATE_MODE_SET_PARAM) { - /// copy param from trainer - for (size_t i = 0; i < offsets.size(); ++i) { - Buffer buffer = inputBuffers[i]; - real* start = buf->getPoint(offsets[i]); - CHECK_LE(offsets[i] + buffer.size, buf->getSize()); - memcpy(start, buffer.base, sizeof(real) * buffer.size); - } - } else { - CHECK(request.update_mode() == PSERVER_UPDATE_MODE_SET_PARAM_ZERO); - /// nothing to do, value vector zero mem already - } -} - -void ParameterServer2::addGradient(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers) { - VLOG(1) << "pserver: addGradient"; - - { - ReadLockGuard guard(parameterMutex_); - int bufferIndex = 0; - for (const auto& block : request.blocks()) { - int64_t offset = getBlockOffset(block); - CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - - int64_t blockId = getBlockId(block); - CHECK_GE(blockId, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - - Buffer buffer = inputBuffers[bufferIndex]; - ++bufferIndex; - - const real* gradientBuffer = buffer.base; - real* gradientSumBuffer = vectors_[PARAMETER_GRADIENT]->getPoint(offset); - - size_t size = buffer.size; - - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(blockId); - if (config.sparse_remote_update()) { - CHECK_EQ(size, config.parameter_block_size()); - } else { // dense - CHECK_LE(size, config.parameter_block_size()); - } - std::lock_guard guard(*info.lock); - simd::addTo(gradientSumBuffer, gradientBuffer, size); - } - } - if (request.batch_status() == BATCH_FINISH || - request.batch_status() == BATCH_START_AND_FINISH) { - numSamplesProcessed_ += request.num_samples(); - cost_ += request.cost(); - VLOG(1) << "num samples: " << numSamplesProcessed_ - << ", new cost:" << cost_; - - /// notify doOperation gradient ready - gradientReadyBarrier_.wait(); - - /// wait doOperation finish - parameterReadyBarrier_.wait(); - VLOG(1) << "start send back"; - } -} - -bool ParameterServer2::asyncGrdientCommitCheckAndStat( - const SendParameterRequest& request) { - const auto trainerId = request.trainer_id(); - int64_t trainerSteps = asyncTrainerSteps_[trainerId]; - CHECK_GE(asyncUpdateSteps_, trainerSteps) - << " async update steps overflows " - << " trainer id: " << trainerId - << " async update steps in pserver: " << asyncUpdateSteps_ - << " async update steps in request: " << trainerSteps; - - asyncUpdateSteps_++; - bool commitGradient = true; - - int64_t delta = asyncUpdateSteps_ - trainerSteps; - if (delta >= asyncLaggedThreshold_) { - VLOG(1) << "discard Async Update: " - << " trainer id: " << trainerId - << " pserver steps: " << asyncUpdateSteps_ - << " request steps: " << trainerSteps; - asyncLaggedGradientsNum_++; - commitGradient = false; - } - /// stat on lagged steps, to get total discard distribution - if (static_cast(delta) < asyncUpdateStat_.size()) { - asyncUpdateStat_[delta]++; - } else { - asyncUpdateStat_[asyncUpdateStat_.size() - 1]++; - } - /// stat on trainerId and discard, to get trainer condition - if (commitGradient) { - asyncTrainerCommitStat_[trainerId]++; - } else { - asyncTrainerDiscardStat_[trainerId]++; - } - - return commitGradient; -} - -static ThreadLocal> localBlockBitset_; - -void ParameterServer2::asyncSGD(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers) { - int64_t numBlocks = blockIdMap_.size(); - auto& localBlockBitset = *localBlockBitset_; - - if (isSparseServer_) { - if (localBlockBitset.empty()) { - localBlockBitset.resize(numBlocks); - } - localBlockBitset.assign(numBlocks, false); - } - - ReadLockGuard guard(parameterMutex_); - - if (request.send_back_parameter()) { - outputBuffers->reserve(request.blocks_size()); - } - - bool commitGradient = asyncGrdientCommitCheckAndStat(request); - - VectorPtr* vecs = parameter::getThreadLocalBuffer(); - size_t bufferIndex = 0; - for (const auto& block : request.blocks()) { - int64_t offset = getBlockOffset(block); - CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - int64_t blockId = getBlockId(block); - CHECK_GE(blockId, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - Buffer buffer = inputBuffers[bufferIndex]; - ++bufferIndex; - - size_t size = buffer.size; - - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(blockId); - - std::lock_guard guard(*info.lock); - /// gradients are too obsolete, will be discarded - if (commitGradient) { - info.optimizer->startBatch(numSamplesProcessed_); - - for (const auto type : info.optimizer->getParameterTypes()) { - vecs[type]->subVecFrom(*vectors_[type], offset, size); - } - vecs[PARAMETER_GRADIENT]->subVecFrom(buffer.base, 0, size); - info.optimizer->update(vecs, config, isSparseServer_ ? 0 : -1); - - if (auto callback = info.optimizer->needSpecialTraversal(config)) { - blockTraverse(info, config, offset, size, vecs, callback); - } - info.optimizer->finishBatch(); - } - - if (commitGradient && isSparseServer_) { - localBlockBitset[blockId] = true; - } - - if (!isSparseServer_ && request.send_back_parameter()) { // dense - int type = request.send_back_parameter_type(); - sendBackParameter(block, type, response, &buffer, outputBuffers); - } - } /// foreach block - - asyncTrainerSteps_[request.trainer_id()] = asyncUpdateSteps_; - - if (commitGradient && isSparseServer_) { - /// find blocks that trainer do not request update - for (int64_t blockId = 0; blockId < numBlocks; ++blockId) { - if (localBlockBitset[blockId]) { - continue; - } - - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = *info.config; - size_t size = config.parameter_block_size(); - - std::lock_guard guard(*info.lock); - info.optimizer->startBatch(numSamplesProcessed_); - if (auto callback = info.optimizer->needSpecialTraversal(config)) { - blockTraverse(info, config, info.offset, size, vecs, callback); - } - info.optimizer->finishBatch(); - } - } - - if (commitGradient && (request.batch_status() == BATCH_FINISH || - request.batch_status() == BATCH_START_AND_FINISH)) { - numSamplesProcessed_ += request.num_samples(); - } - - /// show some performance log if needed - if (request.trainer_id() == 0) { - /// batchId_ is approximately equal to "real batchId_" - batchId_++; - } -} - -void ParameterServer2::getParameter(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers) { - (void)inputBuffers; - LOG(INFO) << "pserver: getParameter"; - ReadLockGuard guard(parameterMutex_); - for (const auto& block : request.blocks()) { - int type = request.send_back_parameter_type(); - sendBackParameter(block, type, response, outputBuffers); - } -} - -void ParameterServer2::getParameterSparse(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers) { - (void)inputBuffers; - auto& buffer = *readWriteBuffer_; - size_t numReals = 0; - for (const auto& block : request.blocks()) { - numReals += getParameterConfig(block).dims(1); - } - buffer.resize(numReals); - - VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals; - - ReadLockGuard guard(parameterMutex_); - size_t offset = 0; - for (const auto& block : request.blocks()) { - size_t width = getParameterConfig(block).dims(1); - Buffer buf = {buffer.data() + offset, width}; - int type = request.send_back_parameter_type(); - sendBackParameterSparse(block, type, response, &buf, width, outputBuffers); - offset += width; - } -} - -void ParameterServer2::sendBackParameter(const ParameterBlock& block, - int parameterType, - SendParameterResponse* response, - std::vector* outputBuffers) { - ParameterBlock* returnBlock = response->add_blocks(); - returnBlock->set_para_id(block.para_id()); - returnBlock->set_block_id(block.block_id()); - returnBlock->set_begin_pos(block.begin_pos()); - returnBlock->set_block_size(block.block_size()); - - int64_t offset = getBlockOffset(block); - CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - - real* valueBuffer = vectors_[parameterType]->getPoint(offset); - outputBuffers->push_back({valueBuffer, (size_t)block.block_size()}); -} - -void ParameterServer2::sendBackParameter(const ParameterBlock& block, - int parameterType, - SendParameterResponse* response, - Buffer* buffer, - std::vector* outputBuffers) { - ParameterBlock* returnBlock = response->add_blocks(); - returnBlock->set_para_id(block.para_id()); - returnBlock->set_block_id(block.block_id()); - returnBlock->set_begin_pos(block.begin_pos()); - returnBlock->set_block_size(block.block_size()); - - int64_t offset = getBlockOffset(block); - CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - - size_t size = buffer->size; - real* valueBuffer = vectors_[parameterType]->getPoint(offset); - /// copy to second buffer to avoid to be polluted by other request - memcpy(buffer->base, valueBuffer, sizeof(real) * size); - outputBuffers->push_back({buffer->base, size}); -} - -void ParameterServer2::sendBackParameterSparse( - const ParameterBlock& block, - int parameterType, - SendParameterResponse* response, - Buffer* buffer, - size_t width, - std::vector* outputBuffers) { - ParameterBlock* returnBlock = response->add_blocks(); - returnBlock->set_para_id(block.para_id()); - returnBlock->set_block_id(block.block_id()); - returnBlock->set_begin_pos(block.begin_pos()); - returnBlock->set_block_size(block.block_size()); - int64_t offset = getBlockOffset(block); - CHECK_GE(offset, 0) << "Only existing parameter block is allowed: " - << " id=" << block.para_id() - << " block id=" << block.block_id(); - - real* valueBuffer = vectors_[parameterType]->getPoint(offset); - CHECK_EQ(buffer->size, width); - memcpy(buffer->base, valueBuffer, width * sizeof(real)); - outputBuffers->push_back(*buffer); -} - -void ParameterServer2::readAllBlocks( - MsgReader* msgReader, std::vector* buffers) { - auto& buffer = *readWriteBuffer_; - size_t numBlocks = msgReader->getNumBlocks(); - buffer.resizeWithAlignHints(msgReader->getTotalLength() / sizeof(real), - numBlocks); - std::vector bufs(numBlocks); - buffers->clear(); - buffers->reserve(numBlocks); - buffer.resetAlignAlloc(); - for (size_t i = 0; i < numBlocks; ++i) { - size_t len = msgReader->getBlockLength(i); - CHECK_EQ(len % sizeof(real), (size_t)0); - size_t size = len / sizeof(real); - bufs[i] = buffer.nextBlock(size); - buffers->push_back({(real*)bufs[i], size}); - } - msgReader->readBlocks(bufs); -} - -void ParameterServer2::sendParameter(const SendParameterRequest& request, - std::unique_ptr msgReader, - ProtoResponseCallbackEx callback) { - SendParameterResponse response; - std::vector inputBuffers; - std::vector outputBuffers; - readAllBlocks(msgReader.get(), &inputBuffers); - msgReader.reset(); - - switch (request.update_mode()) { - case PSERVER_UPDATE_MODE_SET_PARAM: - case PSERVER_UPDATE_MODE_SET_PARAM_ZERO: - setParameter(request, inputBuffers, &response, &outputBuffers); - break; - case PSERVER_UPDATE_MODE_GET_PARAM: - getParameter(request, inputBuffers, &response, &outputBuffers); - break; - case PSERVER_UPDATE_MODE_GET_PARAM_SPARSE: - getParameterSparse(request, inputBuffers, &response, &outputBuffers); - break; - case PSERVER_UPDATE_MODE_ASYNC_SGD: - asyncSGD(request, inputBuffers, &response, &outputBuffers); - break; - case PSERVER_UPDATE_MODE_ADD_GRADIENT: - addGradient(request, inputBuffers, &response, &outputBuffers); - break; - case PSERVER_UPDATE_MODE_AVERAGE_PARAMETER: - break; - } - switch (request.update_mode()) { - case PSERVER_UPDATE_MODE_ADD_GRADIENT: - (*requestVec_).push_back(request); - (*callbackVec_).push_back(callback); - if (request.batch_status() == BATCH_FINISH || - request.batch_status() == BATCH_START_AND_FINISH) { - for (size_t i = 0; i < (*requestVec_).size(); i++) { - ReadLockGuard guard(parameterMutex_); - SendParameterRequest& request = (*requestVec_)[i]; - SendParameterResponse responseTemp; - - std::vector outputIovs; - if (request.send_back_parameter()) { - CHECK(!isSparseServer_); - std::vector outputBuffersTemp; - for (const auto& block : request.blocks()) { - int type = request.send_back_parameter_type(); - sendBackParameter(block, type, &responseTemp, &outputBuffersTemp); - } - outputIovs.reserve(outputBuffersTemp.size()); - for (auto buffer : outputBuffersTemp) { - outputIovs.push_back({buffer.base, buffer.size * sizeof(real)}); - } - } - - ProtoResponseCallbackEx& callbackTemp = (*callbackVec_)[i]; - callbackTemp(responseTemp, outputIovs); - } - (*requestVec_).clear(); - (*callbackVec_).clear(); - } - break; - case PSERVER_UPDATE_MODE_SET_PARAM: - case PSERVER_UPDATE_MODE_SET_PARAM_ZERO: - case PSERVER_UPDATE_MODE_GET_PARAM: - case PSERVER_UPDATE_MODE_GET_PARAM_SPARSE: - case PSERVER_UPDATE_MODE_ASYNC_SGD: - case PSERVER_UPDATE_MODE_AVERAGE_PARAMETER: - std::vector outputIovs; - outputIovs.reserve(outputBuffers.size()); - for (auto buffer : outputBuffers) { - outputIovs.push_back({buffer.base, buffer.size * sizeof(real)}); - } - callback(response, outputIovs); - break; - } -} - -template -void ParameterServer2::reduceAndSendData(const SendDataRequest& request, - std::unique_ptr& msgReader, - ProtoResponseCallbackEx& callback) { - SendDataResponse response; - response.set_type(request.type()); - response.set_server_id(serverId_); - - auto sendData = reinterpret_cast(dataMems_[0].get()->getBuf()); - size_t rawMemSize = dataMems_[0].get()->getSize(); - CHECK_EQ(rawMemSize % sizeof(Dtype), 0U); - size_t dataMemSize = rawMemSize / sizeof(Dtype); - for (size_t i = 1; i < dataMems_.size(); ++i) { - CHECK_EQ(dataMems_[i].get()->getSize(), rawMemSize); - auto data = reinterpret_cast(dataMems_[i].get()->getBuf()); - for (size_t j = 0; j < dataMemSize; ++j) { - sendData[j] += data[j]; - } - } - std::vector outputIovs; - auto block = response.add_blocks(); - outputIovs.push_back({sendData, rawMemSize}); - block->set_total_size(rawMemSize); - block->set_data_size(sizeof(Dtype)); - callback(response, outputIovs); -} - -void ParameterServer2::templateReduceSum(const SendDataRequest& request, - std::unique_ptr& msgReader, - ProtoResponseCallbackEx& callback) { - const auto& block = request.blocks(0); - switch (block.data_type()) { - case TRANS_FLOAT: - reduceAndSendData(request, msgReader, callback); - break; - case TRANS_DOUBLE: - reduceAndSendData(request, msgReader, callback); - break; - case TRANS_INT32: - reduceAndSendData(request, msgReader, callback); - break; - case TRANS_UINT32_T: - reduceAndSendData(request, msgReader, callback); - break; - case TRANS_INT64_T: - reduceAndSendData(request, msgReader, callback); - break; - case TRANS_UINT64_T: - reduceAndSendData(request, msgReader, callback); - break; - default: - LOG(FATAL) << "not supported"; - break; - } -} - -void ParameterServer2::sendData(const SendDataRequest& request, - std::unique_ptr msgReader, - ProtoResponseCallbackEx callback) { - SendDataResponse response; - response.set_type(request.type()); - response.set_server_id(serverId_); - - switch (request.update_mode()) { - case DATA_UPDATE_MODE_SET_OWN: { - CHECK_EQ(msgReader->getNumBlocks(), (size_t)(request.blocks_size())); - size_t totalLen = msgReader->getTotalLength(); - if (totalLen > 0) { - CHECK_EQ(msgReader->getNumBlocks(), 1U) - << "Only one block currently support now!"; - const auto& block = request.blocks(0); - if (0 == dataSize_) { - dataSize_ = block.data_size(); - } else { - CHECK_EQ(dataSize_, block.data_size()); - } - int64_t serverId = request.server_id(); - if (serverId_ < 0) { - serverId_ = serverId; - } else { - CHECK_EQ(serverId_, serverId); - } - int64_t clientId = request.client_id(); - dataMems_[clientId] = std::make_shared(totalLen); - CHECK_EQ(totalLen % sizeof(block.data_size()), 0U); - msgReader->readNextBlock(dataMems_[clientId].get()->getBuf()); - } - msgReader.reset(); - std::vector outputIovs; - callback(response, outputIovs); - break; - } - case DATA_UPDATE_MODE_GET_ALL: { - /// Currently only support DATA_REDUCE_SUM - /// And their Operations are just add - CHECK(DATA_REDUCE_SUM == request.type()); - templateReduceSum(request, msgReader, callback); - break; - } - default: { LOG(FATAL) << "not supported"; } - } -} - -void ParameterServer2::clearUnusedSegments(CpuVector* vec) { - real* data = vec->getData(); - if (usedSegments_.empty()) { - return; - } - memset(data, 0, sizeof(real) * usedSegments_[0].first); - memset(data + usedSegments_.back().second, - 0, - sizeof(real) * (size_ - usedSegments_.back().second)); - size_t n = size_ - usedSegments_.back().second; - - for (size_t i = 1; i < usedSegments_.size(); ++i) { - memset( - data + usedSegments_[i - 1].second, - 0, - sizeof(real) * (usedSegments_[i].first - usedSegments_[i - 1].second)); - n += usedSegments_[i].first - usedSegments_[i - 1].second; - } -} - -void ParameterServer2::parallelExecForEachBlock(ExecFunc func) { - SyncThreadPool::execHelper( - syncThreadPool_.get(), [&](int tid, size_t numThreads) { - int64_t numBlocks = blockIdMap_.size(); - VectorPtr* vecs = parameter::getThreadLocalBuffer(); - for (int64_t blockId = tid; blockId < numBlocks; - blockId += numThreads) { - func(blockId, vecs); - } - }); -} - -void ParameterServer2::blockTraverse( - BlockInfo& info, - const ParameterConfig& config, - int64_t offset, - size_t size, - const VectorPtr vecs[], - const ParameterOptimizer::TraverseCallback& callback) { - /// setup sub bufs - for (const auto type : info.optimizer->getParameterTypes()) { - vecs[type]->subVecFrom(*vectors_[type], offset, size); - } - callback(vecs, config, config.sparse_remote_update() ? 0 : -1LU); -} - -void ParameterServer2::op_SGD(const Operation& operation, - OperationResult* result) { - (void)operation; - (void)result; - - if (allClientPassFinish_) { - /// when all clients signal pass finished, the update - /// is empty. - return; - } - - { - parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(blockId); - int64_t offset = info.offset; - size_t size = config.parameter_block_size(); - - info.optimizer->startBatch(numSamplesProcessed_); - - for (const auto type : info.optimizer->getParameterTypes()) { - vecs[type]->subVecFrom(*vectors_[type], offset, size); - } - info.optimizer->update( - vecs, config, config.sparse_remote_update() ? 0 : -1LU); - vecs[PARAMETER_GRADIENT]->zeroMem(); - - if (auto callback = info.optimizer->needSpecialTraversal(config)) { - blockTraverse(info, config, offset, size, vecs, callback); - } - info.optimizer->finishBatch(); - }); - } - - batchId_++; -} - -void ParameterServer2::op_start_pass(const Operation& operation, - OperationResult* result) { - (void)operation; - (void)result; - - parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { - BlockInfo& info = blockInfos_[blockId]; - info.optimizer->startPass(); - }); -} - -void ParameterServer2::op_finish_pass(const Operation& operation, - OperationResult* result) { - (void)operation; - (void)result; - - parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(blockId); - size_t size = config.parameter_block_size(); - - /// catch up with - if (auto callback = info.optimizer->startCatchUpWith()) { - blockTraverse(info, config, info.offset, size, vecs, callback); - info.optimizer->finishCatchUpWith(); - } - - /// finish pass - info.optimizer->finishPass(); - }); - batchId_ = 0; -} - -void ParameterServer2::op_apply(const Operation& operation, - OperationResult* result) { - (void)operation; - (void)result; - - parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(blockId); - int64_t offset = info.offset; - size_t size = config.parameter_block_size(); - - // catch up with - if (auto callback = info.optimizer->startCatchUpWith()) { - blockTraverse(info, config, offset, size, vecs, callback); - info.optimizer->finishCatchUpWith(); - } - - // apply to PARAMETER_APPLY - if (auto callback = info.optimizer->apply()) { - blockTraverse(info, config, offset, size, vecs, callback); - } - }); -} - -void ParameterServer2::op_randomize(const Operation& operation, - OperationResult* result) { - LOG(INFO) << "ParameterServer2::op_randomize: serverId=" << serverId_; - - CpuVector& valueVec = *vectors_[PARAMETER_VALUE]; - - parallelExecForEachBlock([&](int64_t blockId, const VectorPtr vecs[]) { - BlockInfo& info = blockInfos_[blockId]; - const ParameterConfig& config = getParameterConfig(blockId); - size_t size = config.parameter_block_size(); - - vecs[PARAMETER_VALUE]->subVecFrom(valueVec, info.offset, size); - Parameter::randomize(vecs[PARAMETER_VALUE], config); - }); -} - -void ParameterServer2::loadValueVector(const LoadValueRequest& request, - ProtoResponseCallback callback) { - LoadValueResponse response; - LOG(INFO) << "ParameterServer2::loadValueVector: serverId=" << serverId_; - - constexpr int kBufLen = 100; - char buf[kBufLen]; - snprintf(buf, kBufLen, "/pserver.%04d", static_cast(serverId_)); - std::string filename = request.dir_name() + buf; - - std::ifstream fs(filename, std::ios_base::binary); - CHECK(fs) << "Fail to open " << filename; - - CpuVector& vec = *vectors_[PARAMETER_VALUE]; - Parameter::Header header; - CHECK(fs.read(reinterpret_cast(&header), sizeof(header))) - << "Fail to read parameters in pserver"; - CHECK(Parameter::isHeaderFormatSupported(header.format)) - << "Incorrect format version: " << header.format; - CHECK_EQ(header.size, (size_t)size_) - << "The size (" << header.size << ") in the file does not match the size " - << "(" << size_ << ") of the pserver: " << serverId_; - CHECK_EQ(header.valueSize, sizeof(real)) << "Unsupported valueSize " - << header.valueSize; - CHECK(fs.read(reinterpret_cast(vec.getData()), - header.size * sizeof(real))); - - callback(response); -} - -void ParameterServer2::saveValueVector(const SaveValueRequest& request, - ProtoResponseCallback callback) { - SaveValueResponse response; - LOG(INFO) << "ParameterServer2::SaveValueVector: serverId=" << serverId_; - - mkDir(request.dir_name().c_str()); - - constexpr int kBufLen = 100; - char buf[kBufLen]; - snprintf(buf, kBufLen, "/pserver.%04d", static_cast(serverId_)); - std::string filename = request.dir_name() + buf; - - std::ofstream fs(filename, std::ios_base::binary); - CHECK(fs) << "Fail to open " << filename; - - CpuVector& vec = vectors_[PARAMETER_APPLY] ? *vectors_[PARAMETER_APPLY] - : *vectors_[PARAMETER_VALUE]; - Parameter::Header header; - // TODO(TJ): save param headerFormat_ - header.format = PARAM_FORMAT_ORIGINAL; - header.valueSize = sizeof(real); - header.size = size_; - - CHECK_EQ(header.size, vec.getSize()); - - CHECK(fs.write(reinterpret_cast(&header), sizeof(header))) - << "Fail to write parameter in pserver: " << serverId_; - - CHECK(fs.write(reinterpret_cast(vec.getData()), - header.size * sizeof(real))) - << "Fail to write parameter in pserver: " << serverId_; - - callback(response); -} - -void ParameterServer2::op_RESET(const Operation& operation, - OperationResult* result) { - (void)result; - CpuVector* u = vectors_[operation.pvectors(0)].get(); - u->reset(operation.scalars(0)); - clearUnusedSegments(u); -} - -void ParameterServer2::op_utv(const Operation& operation, - OperationResult* result) { - real* u = vectors_[operation.pvectors(0)]->getData(); - real* v = vectors_[operation.pvectors(1)]->getData(); - int64_t size = size_; - double sum = 0; - for (int64_t i = 0; i < size; ++i) { - sum += (double)u[i] * (double)v[i]; - } - result->add_scalars(sum); -} - -void ParameterServer2::op_au_bv(const Operation& operation, - OperationResult* result) { - (void)result; - real* u = vectors_[operation.pvectors(0)]->getData(); - real* v = vectors_[operation.pvectors(1)]->getData(); - int64_t size = size_; - real a = operation.scalars(0); - real b = operation.scalars(1); - for (int64_t i = 0; i < size; ++i) { - v[i] = a * u[i] + b * v[i]; - } -} - -void ParameterServer2::op_COPY(const Operation& operation, - OperationResult* result) { - (void)result; - real* u = vectors_[operation.pvectors(0)]->getData(); - real* v = vectors_[operation.pvectors(1)]->getData(); - int64_t size = size_; - for (int64_t i = 0; i < size; ++i) { - v[i] = u[i]; - } -} - -void ParameterServer2::op_au(const Operation& operation, - OperationResult* result) { - (void)result; - real* u = vectors_[operation.pvectors(0)]->getData(); - int64_t size = size_; - real a = operation.scalars(0); - for (int64_t i = 0; i < size; ++i) { - u[i] *= a; - } -} - -void ParameterServer2::op_au_bv_cw(const Operation& operation, - OperationResult* result) { - (void)result; - real* u = vectors_[operation.pvectors(0)]->getData(); - real* v = vectors_[operation.pvectors(1)]->getData(); - real* w = vectors_[operation.pvectors(2)]->getData(); - int64_t size = size_; - real a = operation.scalars(0); - real b = operation.scalars(1); - real c = operation.scalars(2); - for (int64_t i = 0; i < size; ++i) { - w[i] = a * u[i] + b * v[i] + c * w[i]; - } -} - -void ParameterServer2::op_make_steepest_desc_dir(const Operation& operation, - OperationResult* result) { - (void)result; - real* dir = vectors_[operation.pvectors(0)]->getData(); - real* grad = vectors_[operation.pvectors(1)]->getData(); - real* x = vectors_[operation.pvectors(2)]->getData(); - int64_t size = size_; - real l1weight = operation.scalars(0); - for (int64_t i = 0; i < size; ++i) { - if (x[i] < 0) { - dir[i] = -grad[i] + l1weight; - } else if (x[i] > 0) { - dir[i] = -grad[i] - l1weight; - } else { - if (grad[i] < -l1weight) { - dir[i] = -grad[i] - l1weight; - } else if (grad[i] > l1weight) { - dir[i] = -grad[i] + l1weight; - } else { - dir[i] = 0; - } - } - } -} - -void ParameterServer2::op_fix_dir_signs(const Operation& operation, - OperationResult* result) { - (void)result; - real* dir = vectors_[operation.pvectors(0)]->getData(); - real* steepestDescDir = vectors_[operation.pvectors(1)]->getData(); - int64_t size = size_; - for (int64_t i = 0; i < size; ++i) { - if (dir[i] * steepestDescDir[i] <= 0) { - dir[i] = 0; - } - } -} - -void ParameterServer2::op_fix_omega_signs(const Operation& operation, - OperationResult* result) { - (void)result; - real* x = vectors_[operation.pvectors(0)]->getData(); - real* newx = vectors_[operation.pvectors(1)]->getData(); - int64_t size = size_; - for (int64_t i = 0; i < size; ++i) { - if (x[i] * newx[i] < 0) { - newx[i] = 0; - } - } -} - -void ParameterServer2::op_dir_deriv(const Operation& operation, - OperationResult* result) { - real* dir = vectors_[operation.pvectors(0)]->getData(); - real* grad = vectors_[operation.pvectors(1)]->getData(); - real* x = vectors_[operation.pvectors(2)]->getData(); - int64_t size = size_; - real l1weight = operation.scalars(0); - double sum = 0; - for (int64_t i = 0; i < size; ++i) { - if (dir[i] != 0) { - if (x[i] < 0) { - sum += dir[i] * (grad[i] - l1weight); - } else if (x[i] > 0) { - sum += dir[i] * (grad[i] + l1weight); - } else if (dir[i] < 0) { - sum += dir[i] * (grad[i] - l1weight); - } else if (dir[i] > 0) { - sum += dir[i] * (grad[i] + l1weight); - } - } - } - result->add_scalars(sum); -} - -void ParameterServer2::op_cost(const Operation& operation, - OperationResult* result) { - real* x = vectors_[operation.pvectors(0)]->getData(); - real* newgrad = vectors_[operation.pvectors(1)]->getData(); - int64_t size = size_; - real l1weight = operation.scalars(0); - real l2weight = operation.scalars(1); - double cost_real = cost_ / mpiSize_; - double sum_weight_l1 = 0; - double sum_weight_l2 = 0; - for (int64_t i = 0; i < size; ++i) { - sum_weight_l1 += std::abs(x[i]); - sum_weight_l2 += x[i] * x[i]; - newgrad[i] += 2.0 * l2weight * x[i]; - } - cost_real += l1weight * sum_weight_l1 + l2weight * sum_weight_l2; - result->add_scalars(cost_real); -} - -ParameterServer2::OperatorFunction ParameterServer2::opFuncs[] = { - nullptr, // PSERVER_OP_utu = 0; - &ParameterServer2::op_utv, // PSERVER_OP_utv = 1; - &ParameterServer2::op_au, // PSERVER_OP_au = 2; - &ParameterServer2::op_au_bv, // PSERVER_OP_au_bv = 3; - nullptr, // PSERVER_OP_aAx_bu = 4; - &ParameterServer2::op_SGD, // PSERVER_OP_SGD = 5; - &ParameterServer2::op_RESET, // PSERVER_OP_RESET = 6; - &ParameterServer2::op_COPY, // PSERVER_OP_COPY = 7; - &ParameterServer2::op_au_bv_cw, // PSERVER_OP_au_bv_cw = 8; - &ParameterServer2::op_make_steepest_desc_dir, - /// PSERVER_OP_MAKE_STEEPEST_DESC_DIR = 9; - &ParameterServer2::op_fix_dir_signs, // PSERVER_OP_FIX_SIGNS = 10; - &ParameterServer2::op_dir_deriv, // PSERVER_OP_DIR_DERIV = 11; - &ParameterServer2::op_fix_omega_signs, // PSERVER_OP_FIX_OMEGA_SIGNS = 12; - &ParameterServer2::op_cost, // PSERVER_OP_COST = 13 - &ParameterServer2::op_start_pass, // PSERVER_OP_START_PASS = 14 - &ParameterServer2::op_finish_pass, // PSERVER_OP_FINISH_PASS = 15 - &ParameterServer2::op_randomize, // PSERVER_OP_RANDOMIZE = 16 - &ParameterServer2::op_apply, // PSERVER_OP_APPLY = 17 -}; - -void ParameterServer2::doOperation(const DoOperationRequest& request, - ProtoResponseCallback callback) { - if (request.wait_for_gradient()) { - /// wait gradient update - gradientReadyBarrier_.wait(); - allClientPassFinish_ = numPassFinishClients_ == FLAGS_num_gradient_servers; - } - - DoOperationResponse response; - response.set_pass_finish(allClientPassFinish_); - - for (const auto& op : request.operations()) { - OperationResult* opResult = response.add_results(); - if (op.operation() >= ARRAYSIZE(opFuncs)) { - LOG(ERROR) << "Unknown operation " << op.operation(); - response.set_return_message(kRetMsgUnknownOperation); - } - OperatorFunction opFunc = opFuncs[op.operation()]; - if (!opFunc) { - LOG(ERROR) << "Operation not implemented: " << op.operation(); - response.set_return_message(kRetMsgUnknownOperation); - } - (this->*opFunc)(op, opResult); - } - - if (request.send_back_parameter()) { - /// clean current cost - cost_ = 0; - - if (allClientPassFinish_ && request.release_pass()) { - /// This signals that all clients finish one pass, so waitPassFinish() - /// will stop waiting. - numPassFinishClients_ = 0; - } - - /// notify addGradient() to send back parameter - parameterReadyBarrier_.wait(); - } - callback(response); -} - -void ParameterServer2::waitPassStart(const WaitPassStartRequest& request, - ProtoResponseCallback callback) { - passBarrier_.wait(); - callback(WaitPassStartResponse()); -} - -void ParameterServer2::waitPassFinish(const WaitPassFinishRequest& request, - ProtoResponseCallback callback) { - numPassFinishClients_ += 1; - - while (numPassFinishClients_ != 0) { - /// notify doOperation gradient ready - gradientReadyBarrier_.wait(); - /// wait doOperation finish - parameterReadyBarrier_.wait(); - } - - callback(WaitPassFinishResponse()); -} - -void ParameterServer2::synchronize(const SynchronizeRequest& request, - ProtoResponseCallback callback) { - synchronizeBarriers_[request.sync_object_id()]->wait(); - dataSize_ = 0; - callback(SynchronizeResponse()); -} - -void ParameterServer2::asyncFinishPass(const SynchronizeRequest& request, - ProtoResponseCallback callback) { - synchronizeBarriers_[request.sync_object_id()]->wait(); - callback(SynchronizeResponse()); - - if (request.trainer_id() == 0) { - batchId_ = 0; - } -} - -void ParameterServer2::createVector(const CreateVectorRequest& request, - ProtoResponseCallback callback) { - (void)request; - CreateVectorResponse response; - LOG(INFO) << "ParameterServer2::createVector: size=" << size_; - CpuVectorPtr vec = std::make_shared(size_); - int64_t handle = -1; - { - std::lock_guard guard(parameterMutex_); - handle = vectors_.size(); - vectors_.push_back(vec); - } - response.set_handle(handle); - callback(response); -} - -void ParameterServer2::releaseVector(const ReleaseVectorRequest& request, - ProtoResponseCallback callback) { - ReleaseVectorResponse response; - CpuVectorPtr vec; - { - std::lock_guard guard(parameterMutex_); - vec.swap(vectors_[request.handle()]); - } - callback(response); -} - -void ParameterServer2::createMatrix(const CreateMatrixRequest& request, - ProtoResponseCallback callback) { - CreateMatrixResponse response; - /// We need to create column major matrix of size_ * num_cols - /// Matrix is row majoar. Need to tranpose when use it. - CpuMatrixPtr mat = std::make_shared(request.num_cols(), size_); - int64_t handle = -1; - { - std::lock_guard guard(parameterMutex_); - handle = matrices_.size(); - matrices_.push_back(mat); - } - response.set_handle(handle); - callback(response); -} - -void ParameterServer2::releaseMatrix(const ReleaseMatrixRequest& request, - ProtoResponseCallback callback) { - ReleaseMatrixResponse response; - CpuMatrixPtr mat; - { - std::lock_guard guard(parameterMutex_); - mat.swap(matrices_[request.handle()]); - } - callback(response); -} - -} // namespace paddle diff --git a/paddle/pserver/ParameterServer2.h b/paddle/pserver/ParameterServer2.h deleted file mode 100644 index 0b8ef5c170c01ec8a5d53f01db9888f82ca68eec..0000000000000000000000000000000000000000 --- a/paddle/pserver/ParameterServer2.h +++ /dev/null @@ -1,696 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include -#include -#include -#include -#include - -#include -#include - -#include "paddle/math/Matrix.h" -#include "paddle/math/Vector.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/parameter/ParameterOptimizer.h" -#include "paddle/utils/Common.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/ThreadLocal.h" - -#include "ParameterService.pb.h" - -#include "ProtoServer.h" - -DECLARE_int32(port); - -namespace paddle { - -// @TODO(yanfei): -// if armed with high density computation resource per node, pserver could also -// utilize GPU to reduce overhead. if this mechanism is used, it could pipeline -// network receiving and GPU computation to reduce the network overhead even -// further. the pipeline could help to accelerate BIG model training. -// @TODO:(yanfei) -// for cpu and less/low gpu machine, the time exhausted by forward and backward -// could be larger than optimization at pserver. However, if armed with lots of -// gpus per node and if the model size is so large enough that limited cpu -// computation causes big optmization latency, the GPU may be required by -// pserver. - -/** - * Client interface for the parameter server - * - * it implements several rpc API for remote parameter client usage. - * for sync-sgd, client needs one controller thread to build connections - * to all pservers, these controller connections do barriers - * synchronization with these connections used for transfering data. - * each data connection uses block based fine grained synchronization - * to gain better scalability. Merging gradients from different trainers - * are concurrently executed with block units, so that some network - * overhead will be hidden in merging gradient. - * for async-sgd, the difference is that pserver will do optimization - * immediately if the gradients are ready, so that pserver needs to - * prepare separate buffer to store value for sending back to trainer - * to prevent from being polluted. - */ -class ParameterServer2 : public ProtoServer { - protected: - /// parameter_ mutex. - RWLock parameterMutex_; - - typedef std::pair BlockKey; - struct BlockKeyHash { - size_t operator()(const BlockKey& key) const { - return std::hash()(key.first) + key.second; - } - }; - - // TODO(yanfei): - // if index data structure is based on parameters instead of blocks, the - // lookup performance could be better. In addition, the block memory - // access almost exhibits good locality, so index data structure and - // block data structure can be refined further, especially if gpu is used - // for pserver. - /** - * all parameters are stored in CpuVector with a blockMap_ data structure - * to index block data required by requests. - */ - typedef std::unordered_map BlockMap; - /// <(para, block), global offset(byte) in all parameters> - BlockMap blockOffsetMap_; - /// <(para, block), global idx [0, nBlocksInAllParameters]> - BlockMap blockIdMap_; - - std::vector vectors_; - std::vector matrices_; - std::vector dataMems_; - - // TODO(yanfei): - // if storing sparse_remote_update() flag in request instead of - // reading configMap_, and storing config within new block wise - // overview data structure, the config mapping, block mapping - // can be unified in single clean data structure. Use para_id - // to index parameters, use offset to index block within parameter - // and keep two index into single one. - /** - * mapping between parameter and config - * different parameter allows different config, such as decay_rate. - * for each request, it need to read config for adding gradient - * and optmization. - */ - std::unordered_map configMap_; - - /** - * to parallelize the multi-thread and multi-connnection - * computation at pserver, it use block unit to reduce - * the contention for computation, even further use block - * level optimizater control for each block for some special - * reason annotated below. - */ - struct BlockInfo { - const ParameterConfig* config; - std::unique_ptr lock; - /// global offset for all parameters - uint64_t offset; - /** - * - * Async sgd in pserver is very different from sync sgd. - * Each trainer follows startBatch, update*, finishBatch as in - * sync sgd, but all these actions are almost executed by - * multi-core and multi-thread simutaneously, so that async - * sgd optimization is based on block level in reality, then - * per block optimization is necessary indeed. In addition, - * per block optimization is also perfered for performance - * with multithreads. - */ - std::unique_ptr optimizer; - }; - std::vector blockInfos_; - - typedef std::vector> BlockSegments; - /// Because some blocks might not be fully used. We keep a - /// record of which segments are used. - BlockSegments usedSegments_; - - /// record pserver status, all status defined in ParameterService.pb - PServerStatus status_; - /// record all samples processed which could be used by optimizater - std::atomic numSamplesProcessed_; - double cost_; - int mpiSize_; - int dataSize_; - /// configuration for current parameter optimizer - OptimizationConfig config_; - - /** - * The ReadWriteBuffer is based on std::vector, but aligned for avx/sse - * compute. And add some helper method to allocate memory aligned blocks. - * - * @param T type of element. - * @param AlignBytes the memory aligned bytes for allocated blocks. - */ - template - class ReadWriteBuffer - : public std::vector> { - public: - static_assert(sizeof(T) % AlignBytes == 0 || AlignBytes % sizeof(T) == 0, - "Type T must be able to aligned."); - - /** - * @brief IsTLargerThanAlign compiled time calculated constant for is type - * T larger than alignments. - */ - constexpr static bool IsTLargerThanAlign = sizeof(T) >= AlignBytes; - - static_assert(std::is_pod::value, "T must be POD type."); - - /** - * @brief if AlignBytes > sizeof(T), then will calcuate how many elements - * can be stored in AlignBytes. - */ - constexpr static size_t AlignElementCount = AlignBytes / sizeof(T); - - static_assert(AlignElementCount == - (AlignElementCount & -AlignElementCount) || - AlignBytes > sizeof(T), - "AlignElementCount should be exp of 2"); - - /** - * @brief Resize Buffer, with block count that will be allocated. Each block - * will be memory aligned in AlignBytes. - * @param size The element count in all blocks. - * @param alignBlockCount The block count that will be allocated. - */ - void resizeWithAlignHints(size_t size, size_t alignBlockCount = 1) { - if (IsTLargerThanAlign) { //! So, each elements is memory aligned. - this->resize(size); - } else { - //! at most, we need such elements in buffer to make sure each block is - //! aligned. - this->resize(size + alignBlockCount * (AlignElementCount - 1)); - } - } - - /** - * @brief reset aligned allocate blocks. - */ - void resetAlignAlloc() { this->curOffset_ = 0; } - - /** - * @brief get next aligned block address. - * @param blockSize is the element count in each block. - * @return Aligned block address. - */ - T* nextBlock(size_t blockSize) { - T* r = &this->operator[](curOffset_); - curOffset_ += blockSize; - - if (!IsTLargerThanAlign) { - curOffset_ = - (curOffset_ + AlignElementCount - 1) & ~(AlignElementCount - 1); - } - return r; - } - - private: - size_t curOffset_; - }; - - /// to buffer the data from network for further processing to - /// reduce redundant memory allocation. - ThreadLocal> readWriteBuffer_; - - /// size of the parameter - int64_t size_; - - /// for synchronized training, check details in addGradient() - /// and doOperation() - ThreadBarrier gradientReadyBarrier_; - ThreadBarrier parameterReadyBarrier_; - ThreadBarrier passBarrier_; - ThreadLocal> requestVec_; - ThreadLocal> callbackVec_; - - std::atomic numPassFinishClients_; - bool allClientPassFinish_; - - std::vector> synchronizeBarriers_; - std::atomic serverId_; - - /** - * - * for lagged async gradient gradient commit control in Async Sgd. - * discard lagged gradients from too slow nodes, whose gradients - * exhibits bad quality. - * Algorithm: - * pserver: - * 1. initial asyncUpdaterSteps = 0, asyncTrainerSteps_[N] = 0. - * syncUpdaterSteps means - * the version of parameter value. - * 2. when pull arrives, record asyncUpdateSteps_ into - * syncTrainerSteps_[trainer_id] - * 3. when push arrives, compare asyncUpdateSteps_ with - * syncTrainerSteps_[trainer_id] - * if delta > threshold, discard current gradient, else commit - * gradient. - * 4. reset asyncUpdaterSteps_ and asyncTrainerSteps_[N] when pass - * finished - * Note: - * it can not discard all lag-gradient strictly in some special - * condition. part of gradients could be discarded if - * ConcurrentRemoteParameterUpdater is sed. - * this algorithm is implemented in asynSGD() - */ - int64_t asyncLaggedThreshold_; - std::atomic asyncUpdateSteps_; - std::vector asyncTrainerSteps_; - size_t asyncLaggedGradientsNum_; - /// stat all async update - std::vector asyncUpdateStat_; - /// stat per trainer_id - std::vector asyncTrainerDiscardStat_; - /// stat per trainer_id - std::vector asyncTrainerCommitStat_; - - /// only used by controller and other control cmd from trainer number 0 - std::unique_ptr syncThreadPool_; - - /// pserver for sparse remote update parameters - bool isSparseServer_; - - /// barrier performance tuning sync-sgd required - std::atomic batchId_; - - public: - struct Buffer { - real* base; - size_t size; - }; - - protected: - /// async gradient commit control - bool asyncGrdientCommitCheckAndStat(const SendParameterRequest& request); - - public: - /// disable default parameter for overloading - /// @rdmaCpu:the id of cpu core hosting RDMA server(0-N) - /// -1 means using TCP transport instead of RDMA - ParameterServer2(const std::string& addr, int port, int rdmaCpu = -1); - - ~ParameterServer2() {} - - static const std::string kRetMsgInvalidMatrixHandle; - static const std::string kRetMsgInvalidVectorHandle; - static const std::string kRetMsgUnknownOperation; - - /// service functions - template - void reduceAndSendData(const SendDataRequest& request, - std::unique_ptr& msgReader, - ProtoResponseCallbackEx& callback); - - void templateReduceSum(const SendDataRequest& request, - std::unique_ptr& msgReader, - ProtoResponseCallbackEx& callback); - - /** - * @brief framework for sending parameters - * - * @note different parameter data type can be sent to pserver. - * in most case, the api is used to send gradients from - * trainer to pserver. - * it also can be used to retrieve parameters from pserver - */ - void sendParameter(const SendParameterRequest& request, - std::unique_ptr msgReader, - ProtoResponseCallbackEx callback); - - void sendData(const SendDataRequest& request, - std::unique_ptr msgReader, - ProtoResponseCallbackEx callback); - - /** - * @brief send config to pserver - * - * @note it can help pserver to understand the configuration for - * optimization, - * logging control, duplicated initialization, etc. - */ - void setConfig(const SetConfigRequest& request, - ProtoResponseCallback callback); - - /** - * @brief get status for pserver - * - * @note used to check if parameters are ready at pserver - */ - void getStatus(const GetStatusRequest& request, - ProtoResponseCallback callback); - - /** - * @brief set status for pserver - * - * @note used to check if parameters are ready at pserver, since parameters - * at pserver are initialized by trainer - */ - void setStatus(const SetStatusRequest& request, - ProtoResponseCallback callback); - - /** - * @brief framework for doing some operation at pserver end - * - * @note if sync-sgd is used, controller will calling op_SGD action - * for gradient optimization. - * check avaiable operations in opFuncs[] - */ - void doOperation(const DoOperationRequest& request, - ProtoResponseCallback callback); - - /// Create a column vector. The size is the dimension of parameter - void createVector(const CreateVectorRequest& request, - ProtoResponseCallback callback); - - void releaseVector(const ReleaseVectorRequest& request, - ProtoResponseCallback callback); - - /// Create a column major matrix. The number of rows is the dimension of - /// parameter. The number of columns is specifed by num_cols. - void createMatrix(const CreateMatrixRequest& request, - ProtoResponseCallback callback); - - void releaseMatrix(const ReleaseMatrixRequest& request, - ProtoResponseCallback callback); - /** - * @brief stateful control for indicationg sync pass start - * - * @note it is valuable for logging and state control, - * especially for sync-sgd control - */ - void waitPassStart(const WaitPassStartRequest& request, - ProtoResponseCallback callback); - - /** - * @brief stateful control for indicationg sync pass end - * - * @note it is valuable for logging and state control, - * especially for sync-sgd control - */ - void waitPassFinish(const WaitPassFinishRequest& request, - ProtoResponseCallback callback); - - /** - * @brief synchronize all distributed trainers - * - * @note it's general api for synchronizing trainer and pserver - */ - void synchronize(const SynchronizeRequest& request, - ProtoResponseCallback callback); - - /** - * @brief stateful control for indicating async pass is finished - * - * @note it is valuable for logging control, state reset, etc. - */ - void asyncFinishPass(const SynchronizeRequest& request, - ProtoResponseCallback callback); - - void loadValueVector(const LoadValueRequest& request, - ProtoResponseCallback callback); - - void saveValueVector(const SaveValueRequest& request, - ProtoResponseCallback callback); - - public: - /** - * @brief initialize parameter server - */ - bool init(); - - /** - * @brief set parameters at pserver - * - * @note do parameter initialization if neccessy. - */ - void setParameter(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers); - - /** - * @brief receive gradients and do optimization for async-sgd - * - * @note this api asynchronizately receives all data from all - * trainers, and immediately do optimization and return - * optimizated value for trainer. - * this above routine are block based atomic updating, - * which means different block could based different stale - * gradient. - * it will discard some lagged gradients by default for - * better convergence. - */ - void asyncSGD(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers); - - /** - * @brief merge gradients from all trainer - * - * @note this api use block based parallelization as fine grained - * parallelization which benifits lock contention and latency - * hidden for communication, also can harness multi-core - * efficiently. - * it also implements the synchronization for sync-sgd - */ - void addGradient(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers); - - /** - * @brief get dense parameters from pserver - * - * @note for some specified condition, trainer will get parameters from - * pservers. - * e.g. - * if all parameters are stored at perver end for big model training - * trainer can use it to retrieve all parameters if necessary. - */ - void getParameter(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers); - - /** - * @brief get sparse value from parameter server - * - * @note with sparse enabled, pservers own all latest value - * while trainer only retrieve value that only are needed. - * e.g. - * trainer will do prefetch action to retrieve necessary latest - * value from pserver for sparse calculation. - */ - void getParameterSparse(const SendParameterRequest& request, - std::vector& inputBuffers, - SendParameterResponse* response, - std::vector* outputBuffers); - - protected: - void mergeSegments(BlockSegments* segments); - - /// set the unused segments to zero - void clearUnusedSegments(CpuVector* vec); - - // TODO(yanfei): - // if read data and do optimization interleavely block by block, - // the performance could be better for gaining less network congestion. - /// read all data from connection and store it in static pre-allocated buffer - void readAllBlocks(MsgReader* msgReader, - std::vector* buffers); - - const ParameterConfig& getParameterConfig(const ParameterBlock& block) { - CHECK_LT(block.para_id(), -1UL) << "invalid parameter id:" - << block.para_id(); - const auto it = configMap_.find(block.para_id()); - CHECK(it != configMap_.end()) << "can not find parameter id: " - << block.para_id(); - return it->second; - } - - /// it implictly check blockOffsetMap_ while retrieving blockId - const ParameterConfig& getParameterConfig(int64_t blockId) const { - CHECK(blockId >= 0 && blockId < (int64_t)blockInfos_.size()) - << "block idx out of range, id: " << blockId - << " info size: " << blockInfos_.size(); - return *(blockInfos_[blockId].config); - } - - template - bool isValidVectorHandle(int64_t handle, Response* response) { - if (handle < 0 || (size_t)handle >= vectors_.size()) { - LOG(ERROR) << "Invalid vector handle " << handle; - response->set_return_message(kRetMsgInvalidVectorHandle); - return false; - } - return true; - } - - template - bool isValidMatrixHandle(int64_t handle, Response* response) { - if (handle < 0 || (size_t)handle >= matrices_.size()) { - LOG(ERROR) << "Invalid matrix handle " << handle; - response->set_return_message(kRetMsgInvalidMatrixHandle); - return false; - } - return true; - } - - /** - * @brief get block offset - * - * @note block.begin_dim is added to the block offset. - * return -1 if block cannot be found - */ - int64_t getBlockOffset(const ParameterBlock& block) const { - BlockKey key(block.para_id(), block.block_id()); - auto it = blockOffsetMap_.find(key); - if (it == blockOffsetMap_.end()) { - return -1; - } - return it->second; - } - - /// return -1 if block cannot be found - int64_t getBlockId(const ParameterBlock& block) const { - BlockKey key(block.para_id(), block.block_id()); - auto it = blockIdMap_.find(key); - if (it == blockIdMap_.end()) { - return -1; - } - return it->second; - } - - /** - * @brief prepare data for sending back - * - * @note modify reponse and outputBuffers for sending parameter - * back to client. The buffer for socket sending uses - * vectors_[parameterType] directly - * for dense with sync-sgd - */ - void sendBackParameter(const ParameterBlock& block, - int parameterType, - SendParameterResponse* response, - std::vector* outputBuffers); - - /** - * @brief prepare data for sending back - * - * @note modify response and outputBuffers for sending parameter - * back to client. The buffer for socket sending uses buffer->base - * The parameter values are copied from vectors_[parameterType] - * to buffer->base. - * for dense with async-sgd - */ - void sendBackParameter(const ParameterBlock& block, - int parameterType, - SendParameterResponse* response, - Buffer* buffer, - std::vector* outputBuffers); - /** - * @brief prepare data for sending back - * - * @note specified for sparse - */ - void sendBackParameterSparse(const ParameterBlock& block, - int parameterType, - SendParameterResponse* response, - Buffer* buffer, - size_t width, - std::vector* outputBuffers); - - /** - * framework routine for block parallelization - * e.g. - * for optimization on all blocks at pserver end, this routine can facilitize - * the parallelize of do optimization on all blocks with multithreads. - */ - typedef std::function ExecFunc; - void parallelExecForEachBlock(ExecFunc func); - void blockTraverse(BlockInfo& info, - const ParameterConfig& config, - int64_t offset, - size_t size, - const VectorPtr vecs[], - const ParameterOptimizer::TraverseCallback& callback); - - public: - typedef void (ParameterServer2::*OperatorFunction)(const Operation& operation, - OperationResult* result); - - /** - * doOperation will call following operations indirectly - * e.g. - * for sync-sgd control, the controller in remote updater will send op_SGD - * command to pserver, then send sendParameter request to pserver immediately. - * the two function at pserver end will do cooperation to achieve the sync-sgd - * gradient merge and optimization. - * the most following operations are specified for owlqn, all operations are - * under the context of doOperation function - */ - static OperatorFunction opFuncs[]; - - void op_SGD(const Operation& operation, OperationResult* result); - - void op_RESET(const Operation& operation, OperationResult* result); - - void op_utv(const Operation& operation, OperationResult* result); - - void op_au_bv(const Operation& operation, OperationResult* result); - - void op_COPY(const Operation& operation, OperationResult* result); - - void op_au(const Operation& operation, OperationResult* result); - - void op_au_bv_cw(const Operation& operation, OperationResult* result); - - void op_make_steepest_desc_dir(const Operation& operation, - OperationResult* result); - - void op_fix_dir_signs(const Operation& operation, OperationResult* result); - - void op_dir_deriv(const Operation& operation, OperationResult* result); - - void op_fix_omega_signs(const Operation& operation, OperationResult* result); - - void op_cost(const Operation& operation, OperationResult* result); - - void op_start_pass(const Operation& operation, OperationResult* result); - void op_finish_pass(const Operation& operation, OperationResult* result); - - void op_apply(const Operation& operation, OperationResult* result); - - void op_randomize(const Operation& operation, OperationResult* result); - - void op_load(const Operation& operation, OperationResult* result); - void op_save(const Operation& operation, OperationResult* result); -}; - -} // namespace paddle diff --git a/paddle/pserver/ParameterServerController.h b/paddle/pserver/ParameterServerController.h deleted file mode 100644 index 1308d62fb1787f19123fe37d49f8e14039c5a39a..0000000000000000000000000000000000000000 --- a/paddle/pserver/ParameterServerController.h +++ /dev/null @@ -1,74 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "ParameterServer2.h" -#include "ParameterServerConfig.pb.h" -#include "RDMANetwork.h" -#include "paddle/utils/StringUtil.h" - -namespace paddle { - -/** - * @brief ParameterServerController is used for create, init and manage multi - * parameter server instances. The num of the instances is decided by port - * num(the ports number for parameter send) and network devices configured - * by gflags or proto. - */ -class ParameterServerController final { - public: - DISABLE_COPY(ParameterServerController); - - /** - * @brief Ctor, Create a ParameterServerController from ParameterServerConfig. - */ - explicit ParameterServerController(const ParameterServerConfig& config); - - /** - * @brief Dtor. - */ - ~ParameterServerController(); - - /** - * @brief create ParameterServerController from gflags, this is used for - * compatibility with the old usage of configuration by gflags. - */ - static ParameterServerController* createFromGflags(); - - /** - * @brief create ParameterServerController with ParameterServerConfig, remove - * gflags from ParameterServer. Init all ParameterServer2 instances according - * to - * the config. - */ - static ParameterServerController* create(const ParameterServerConfig& config); - - /** - * @brief start all ParameterServer2 instances in this - * ParameterServerController. - */ - void start(); - - /** - * @brief join and wait for all ParameterServer2 instances thread in this - * ParameterServerController. - */ - void wait(); - - private: - std::vector> parameterServers_; -}; - -} // namespace paddle diff --git a/paddle/pserver/RDMANetwork.h b/paddle/pserver/RDMANetwork.h deleted file mode 100644 index 83db6b9df71274c3a8eb3403457877b68f2b6dea..0000000000000000000000000000000000000000 --- a/paddle/pserver/RDMANetwork.h +++ /dev/null @@ -1,158 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#ifndef PADDLE_DISABLE_RDMA -#include "sxi_sock.h" -#else -#define PROMPT_ERR() LOG(FATAL) << "Paddle is not compiled with rdma" -#endif -#include "paddle/utils/Logging.h" - -#include -struct sxi_sock; -struct sxi_socket; - -#ifndef MAX_VEC_SIZE -// define default MAX_VEC_SIZE -#define MAX_VEC_SIZE (1UL << 16) -#endif - -namespace paddle { -/// Namespace rdma is adaptors for sxi_sock.h. Make paddle not depend on it -/// when disable rdma support -namespace rdma { -inline int numCpus() { -#ifndef PADDLE_DISABLE_RDMA - return sxi_num_configured_cpus(); -#else - return 0; -#endif -} - -inline sxi_socket* ssocket(int cpuId) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_ssocket(cpuId); -#else - PROMPT_ERR(); -#endif -} - -inline int listen(sxi_socket* s) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_listen(s); -#else - PROMPT_ERR(); -#endif -} - -inline int bind(sxi_socket* s, const char* str) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_bind(s, str); -#else - PROMPT_ERR(); -#endif -} - -inline sxi_sock* accept(sxi_socket* s) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_accept(s); -#else - PROMPT_ERR(); -#endif -} - -inline sockaddr_in* getSourceAddress(sxi_sock* sock) { -#ifndef PADDLE_DISABLE_RDMA - return reinterpret_cast(&sock->sa); -#else - PROMPT_ERR(); -#endif -} - -inline int close(sxi_socket* sock) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_socket_close(sock); -#else - PROMPT_ERR(); -#endif -} - -inline int close(sxi_sock* sock) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_sock_close(sock); -#else - PROMPT_ERR(); -#endif -} - -inline void init() { -#ifndef PADDLE_DISABLE_RDMA - sxi_module_init(); -#else - PROMPT_ERR(); -#endif -} - -inline sxi_socket* csocket(int cpuId) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_csocket(cpuId); -#else - PROMPT_ERR(); -#endif -} - -inline ssize_t read(sxi_sock* channel, void* data, size_t len) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_read(channel, data, len); -#else - PROMPT_ERR(); -#endif -} - -inline ssize_t write(sxi_sock* channel, void* data, size_t len) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_write(channel, data, len); -#else - PROMPT_ERR(); -#endif -} - -inline ssize_t readv(sxi_sock* channel, iovec* iov, int count) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_readv(channel, iov, count); -#else - PROMPT_ERR(); -#endif -} - -inline ssize_t writev(sxi_sock* channel, iovec* iov, int count) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_writev(channel, iov, count); -#else - PROMPT_ERR(); -#endif -} - -inline sxi_sock* connect(sxi_socket* socket, const char* url) { -#ifndef PADDLE_DISABLE_RDMA - return sxi_connect(socket, url); -#else - PROMPT_ERR(); -#endif -} - -} // namespace rdma -} // namespace paddle diff --git a/paddle/pserver/SocketChannel.cpp b/paddle/pserver/SocketChannel.cpp deleted file mode 100644 index 72e6943408a1856db214262ff0b0698a2eb89a91..0000000000000000000000000000000000000000 --- a/paddle/pserver/SocketChannel.cpp +++ /dev/null @@ -1,235 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "SocketChannel.h" - -#include -#include -#include -#include -#include -#include -#include "RDMANetwork.h" - -#include "paddle/utils/Util.h" - -namespace paddle { - -/** - * UIO_MAXIOV is documented in writev(2), but only - * declares it on osx/ios if defined(KERNEL) - */ -#ifndef UIO_MAXIOV -#define UIO_MAXIOV 512 -#endif - -SocketChannel::~SocketChannel() { - if (tcpRdma_ == F_TCP) - close(tcpSocket_); - else - rdma::close(rdmaSocket_); - LOG(INFO) << "destory connection in socket channel, peer = " << peerName_; -} - -size_t SocketChannel::read(void* buf, size_t size) { - size_t total = 0; - while (total < size) { - ssize_t len; - if (tcpRdma_ == F_TCP) - len = ::read(tcpSocket_, (char*)buf + total, size - total); - else - len = rdma::read(rdmaSocket_, (char*)buf + total, size - total); - - CHECK(len >= 0) << " peer=" << peerName_; - if (len <= 0) { - return total; - } - total += len; - } - return total; -} - -size_t SocketChannel::write(const void* buf, size_t size) { - size_t total = 0; - while (total < size) { - ssize_t len; - if (tcpRdma_ == F_TCP) - len = ::write(tcpSocket_, (const char*)buf + total, size - total); - else - len = rdma::write(rdmaSocket_, (char*)buf + total, size - total); - - CHECK(len >= 0) << " peer=" << peerName_; - if (len <= 0) { - return total; - } - total += len; - } - return total; -} - -template -static size_t readwritev(IOFunc iofunc, - SocketType socket, - iovec* iovs, - int iovcnt, - int maxiovs, - const std::string& peerName) { - int curIov = 0; - size_t total = 0; - - for (int i = 0; i < iovcnt; ++i) { - total += iovs[i].iov_len; - } - - size_t size = 0; - size_t curIovSizeDone = 0; - - while (size < total) { - ssize_t len = - iofunc(socket, &iovs[curIov], std::min(iovcnt - curIov, maxiovs)); - CHECK(len > 0) << " peer=" << peerName << " curIov=" << curIov - << " iovCnt=" << iovcnt - << " iovs[curIov].base=" << iovs[curIov].iov_base - << " iovs[curIov].iov_len=" << iovs[curIov].iov_len; - size += len; - - /// restore iovs[curIov] to the original value - iovs[curIov].iov_base = - (void*)((char*)iovs[curIov].iov_base - curIovSizeDone); - iovs[curIov].iov_len += curIovSizeDone; - - len += curIovSizeDone; - - while (curIov < iovcnt) { - if ((size_t)len < iovs[curIov].iov_len) break; - len -= iovs[curIov].iov_len; - ++curIov; - } - if (curIov < iovcnt) { - curIovSizeDone = len; - iovs[curIov].iov_base = (void*)((char*)iovs[curIov].iov_base + len); - iovs[curIov].iov_len -= len; - } - } - return size; -} - -/// rdma::readv and rdma::writev can take advantage of RDMA blocking offload -/// transfering -size_t SocketChannel::writev(const std::vector& iovs) { - if (tcpRdma_ == F_TCP) - return readwritev(::writev, - tcpSocket_, - const_cast(&iovs[0]), - iovs.size(), - UIO_MAXIOV, - peerName_); - else - return readwritev(rdma::writev, - rdmaSocket_, - const_cast(&iovs[0]), - iovs.size(), - MAX_VEC_SIZE, - peerName_); -} - -size_t SocketChannel::readv(std::vector* iovs) { - if (tcpRdma_ == F_TCP) - return readwritev(::readv, - tcpSocket_, - const_cast(&(*iovs)[0]), - iovs->size(), - UIO_MAXIOV, - peerName_); - else - return readwritev(rdma::readv, - rdmaSocket_, - const_cast(&(*iovs)[0]), - iovs->size(), - MAX_VEC_SIZE, - peerName_); -} - -void SocketChannel::writeMessage(const std::vector& userIovs) { - MessageHeader header; - header.numIovs = userIovs.size(); - - std::vector iovLengths; - iovLengths.reserve(userIovs.size()); - for (auto& iov : userIovs) { - iovLengths.push_back(iov.iov_len); - } - - std::vector iovs; - iovs.reserve(userIovs.size() + 2); - iovs.push_back({&header, sizeof(header)}); - iovs.push_back({&iovLengths[0], - static_cast(sizeof(iovLengths[0]) * header.numIovs)}); - iovs.insert(iovs.end(), userIovs.begin(), userIovs.end()); - - header.totalLength = 0; - for (auto& iov : iovs) { - header.totalLength += iov.iov_len; - } - - CHECK(writev(iovs) == (size_t)header.totalLength); -} - -std::unique_ptr SocketChannel::readMessage() { - MessageHeader header; - - size_t len = read(&header, sizeof(header)); - if (len == 0) { - return nullptr; - } - - CHECK(len == sizeof(header)); - - std::unique_ptr msgReader(new MsgReader(this, header.numIovs)); - - CHECK_EQ(msgReader->getTotalLength() + sizeof(header) + - msgReader->getNumBlocks() * sizeof(size_t), - (size_t)header.totalLength) - << " totalLength=" << msgReader->getTotalLength() - << " numBlocks=" << msgReader->getNumBlocks(); - return msgReader; -} - -MsgReader::MsgReader(SocketChannel* channel, size_t numBlocks) - : channel_(channel), blockLengths_(numBlocks), currentBlockIndex_(0) { - size_t size = numBlocks * sizeof(blockLengths_[0]); - CHECK(channel_->read(&blockLengths_[0], size) == size); -} - -void MsgReader::readBlocks(const std::vector& bufs) { - CHECK_LE(currentBlockIndex_ + bufs.size(), blockLengths_.size()); - std::vector iovs; - iovs.reserve(bufs.size()); - size_t totalLength = 0; - for (void* buf : bufs) { - iovs.push_back({buf, getNextBlockLength()}); - totalLength += getNextBlockLength(); - ++currentBlockIndex_; - } - - CHECK(channel_->readv(&iovs) == totalLength); -} - -void MsgReader::readNextBlock(void* buf) { - CHECK_LT(currentBlockIndex_, blockLengths_.size()); - CHECK(channel_->read(buf, getNextBlockLength()) == getNextBlockLength()); - ++currentBlockIndex_; -} - -} // namespace paddle diff --git a/paddle/pserver/SocketChannel.h b/paddle/pserver/SocketChannel.h deleted file mode 100644 index 8b45ac56090ef82e77514566e7df6b366958655e..0000000000000000000000000000000000000000 --- a/paddle/pserver/SocketChannel.h +++ /dev/null @@ -1,153 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include - -#include -#include - -struct sxi_sock; - -namespace paddle { - -class SocketChannel; -enum ChannelType { - F_TCP = 1, - F_RDMA = 2, -}; - -/// reading a set of blocks of data from SocketChannel. -class MsgReader { - public: - MsgReader(SocketChannel* channel, size_t numIovs); - ~MsgReader() { - /// ensure all data blocks have been processed - CHECK_EQ(currentBlockIndex_, blockLengths_.size()); - } - /** - * @brief number of remaining parts - */ - size_t getNumBlocks() const { - return blockLengths_.size() - currentBlockIndex_; - } - - /** - * @brief lenght of next block - */ - size_t getNextBlockLength() const { return getBlockLength(0); } - - /** - * @brief get the total length of all the remaining blocks - */ - size_t getTotalLength() const { - size_t total = 0; - for (size_t i = currentBlockIndex_; i < blockLengths_.size(); ++i) { - total += blockLengths_[i]; - } - return total; - } - - /** - * @brief Get the length for block currentBlockIndex + i - */ - size_t getBlockLength(size_t i) const { - return blockLengths_[currentBlockIndex_ + i]; - } - - /** - * @brief read blocks data and store it to buf - */ - void readBlocks(const std::vector& bufs); - void readNextBlock(void* buf); - - protected: - SocketChannel* channel_; - std::vector blockLengths_; - size_t currentBlockIndex_; -}; - -/// APIs for reading and writing byte stream data or naive iov data -/// from the APIs both RDMA and TCP exhibits byte stream style -class SocketChannel { - public: - SocketChannel(int socket, const std::string& peerName) - : tcpSocket_(socket), peerName_(peerName) { - tcpRdma_ = F_TCP; - } - SocketChannel(struct sxi_sock* socket, const std::string& peerName) - : rdmaSocket_(socket), peerName_(peerName) { - tcpRdma_ = F_RDMA; - } - - ~SocketChannel(); - - const std::string& getPeerName() const { return peerName_; } - - /** - * @brief read size bytes. - * - * @note keep reading until getting size bytes or sock is closed - * is closed - */ - size_t read(void* buf, size_t size); - - /** - * @brief write size bytes. - * - * @note keep writing until writing size bytes or sock is closed - */ - size_t write(const void* buf, size_t size); - - /** - * @brief write a set of buffers. - * - * @note keep writing until all buffers are written or sock is closed - */ - size_t writev(const std::vector& iov); - - /** - * @brief read a set of buffers. - * - * @note keep reading until all buffers are full or sock is closed. - */ - size_t readv(std::vector* iov); - - /** - * @brief write a set of buffers. - * - * @note keep writing until all buffers are passed or sock is closed - */ - void writeMessage(const std::vector& iov); - - /// return null to indicate socket is closed - std::unique_ptr readMessage(); - - protected: - struct MessageHeader { - int64_t totalLength; /// include the header - int64_t numIovs; - int64_t iovLengths[0]; - }; - - int tcpSocket_; - struct sxi_sock* rdmaSocket_; - const std::string peerName_; - enum ChannelType tcpRdma_; -}; - -} // namespace paddle diff --git a/paddle/pserver/SparseParameterDistribution.cpp b/paddle/pserver/SparseParameterDistribution.cpp deleted file mode 100644 index bb247f389cc26b32ff79d36bdf5c81ba8591dc58..0000000000000000000000000000000000000000 --- a/paddle/pserver/SparseParameterDistribution.cpp +++ /dev/null @@ -1,123 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "paddle/utils/Logging.h" - -#include "paddle/utils/Flags.h" - -#include "SparseParameterDistribution.h" - -DEFINE_bool(check_sparse_distribution_in_pserver, - false, - "check whether sparse parameter exhibts balanced distribution at " - "all pservers"); -DEFINE_bool(show_check_sparse_distribution_log, - false, - "show logs details for sparse parameter distribution in pserver"); -DEFINE_int32(check_sparse_distribution_batches, - 100, - "run sparse parameter distribution check for N batches"); -DEFINE_double( - check_sparse_distribution_ratio, - 0.6, - "if parameters dispatched to different pservers exhibit unbalanced " - " distribution for check_sparse_distribution_ratio * " - " check_sparse_distribution_batches times, crash program"); -DEFINE_double(check_sparse_distribution_unbalance_degree, - 2.0, - "the ratio of maximum data size and minimun data size for " - "different pserver"); - -namespace paddle { - -SparseParameterDistribution::SparseParameterDistribution(size_t serviceNum) { - totBytes_ = 0; - data_.resize(serviceNum); - - batchPassed_ = 0; - unbalanceCnt_ = 0; -} - -void SparseParameterDistribution::probeDistribution(int serverId, - size_t dataSize) { - if (!FLAGS_check_sparse_distribution_in_pserver || - batchPassed_ > FLAGS_check_sparse_distribution_batches) { - return; - } - - CHECK_LT((size_t)serverId, data_.size()) - << "invalid sparse parameter distribution probe"; - - data_[serverId] += dataSize; - totBytes_ += dataSize; -} - -void SparseParameterDistribution::checkAndResetDistribution() { - if (!FLAGS_check_sparse_distribution_in_pserver || - batchPassed_ >= FLAGS_check_sparse_distribution_batches) { - return; - } - - /// at runtime, prepareSendData is called by many contexts, - /// so need to check if data is avaiable. - if (!totBytes_) { - return; - } - - /// check if distribution is balanced - auto avgSize = totBytes_ / data_.size(); - auto unbalanceDegree = FLAGS_check_sparse_distribution_unbalance_degree; - for (auto& dataSize : data_) { - if (dataSize > unbalanceDegree * avgSize || - dataSize * unbalanceDegree < avgSize) { - unbalanceCnt_++; - break; - } - } - - auto printData = [&]() { - std::stringstream ss; - for (auto& dataSize : data_) { - ss << dataSize * 0.001 << "KB "; - } - ss << std::endl; - LOG(INFO) << ss.str(); - }; - - /// show all sparse data size for different pserver - if (FLAGS_show_check_sparse_distribution_log) { - LOG(INFO) << "sparse distribution:"; - printData(); - } - - totBytes_ = 0; - batchPassed_++; - - if (batchPassed_ == FLAGS_check_sparse_distribution_batches) { - LOG(INFO) << "show last parameter distribution sample:"; - printData(); - LOG(INFO) << "total unbalanced batches: " << unbalanceCnt_ - << " in passed batches: " << batchPassed_; - CHECK_LE((float)unbalanceCnt_ / (float)batchPassed_, - FLAGS_check_sparse_distribution_ratio) - << "unbalanced sparse parameter distribution for different pserver. " - << "it could be caused by unbalanced sparse ids distribution, try " - << "to shuffle dimensions in input samples"; - } - - std::fill(data_.begin(), data_.end(), 0); -} -} // namespace paddle diff --git a/paddle/pserver/SparseParameterDistribution.h b/paddle/pserver/SparseParameterDistribution.h deleted file mode 100644 index e168f36c75e9452fff547f139a67a553cc6b796a..0000000000000000000000000000000000000000 --- a/paddle/pserver/SparseParameterDistribution.h +++ /dev/null @@ -1,52 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include - -#include -#include "paddle/utils/Logging.h" - -namespace paddle { - -/* - * if sparse_remote_updater is used, different ParameterServer could - * be assigned with unbalanced gradients. the parameter value from - * ParameterServer also be not balanced. the distribution of different - * dimensions of sparse ids determines the unbalanced degree of data - * distributed among all ParameterServers. Even distribution will - * benifits cluster efficiency. - * do check the unbalanced degree of gradients at runtime, crash program - * if unbalanced distribution exhibts by default. - */ -class SparseParameterDistribution { - public: - /// serviceNum means the number of ParameterServers - explicit SparseParameterDistribution(size_t serviceNum); - ~SparseParameterDistribution() {} - /// collect data - void probeDistribution(int serverId, size_t data); - void checkAndResetDistribution(); - - private: - std::vector data_; - std::atomic totBytes_; - - /// after some batches, stop to check - int batchPassed_; - - /// stat on unbalanced distribution found - int unbalanceCnt_; -}; -} // namespace paddle diff --git a/paddle/pserver/test/SocketTest.cpp b/paddle/pserver/test/SocketTest.cpp deleted file mode 100644 index 206cd17c379f529579c103893cfb492524bc6f8d..0000000000000000000000000000000000000000 --- a/paddle/pserver/test/SocketTest.cpp +++ /dev/null @@ -1,256 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Util.h" - -#include -#include -#include -#include -#include - -#include - -#include "paddle/math/Vector.h" -#include "paddle/utils/Logging.h" - -struct MessageHeader { - int64_t dataLength; -}; - -class Thread { - public: - void start(); - virtual void run() = 0; - virtual ~Thread() {} - - protected: - std::unique_ptr thread_; -}; - -void Thread::start() { - thread_.reset(new std::thread([this]() { this->run(); })); -} - -class SocketChannel { - public: - explicit SocketChannel(int socket) : socket_(socket) {} - int getSocketFd() const { return socket_; } - uint64_t readAll(void* buf, size_t size); - uint64_t writeAll(const void* buf, size_t size); - - protected: - int socket_; -}; - -uint64_t SocketChannel::readAll(void* buf, size_t size) { - uint64_t total = 0; - while (total < size) { - int64_t len = read(socket_, (char*)buf + total, size - total); - if (len <= 0) { - return total; - } - total += len; - } - return total; -} - -uint64_t SocketChannel::writeAll(const void* buf, size_t size) { - uint64_t total = 0; - while (total < size) { - int64_t len = write(socket_, (const char*)buf + total, size - total); - if (len <= 0) { - return total; - } - total += len; - } - return total; -} - -class SocketWorker : public Thread { - public: - explicit SocketWorker(int socket) : channel_(socket) {} - virtual void run(); - - // read n bytes. - int64_t readAll(char* buf, size_t n); - - // write n bytes - - protected: - SocketChannel channel_; - std::string buffer_; -}; - -class SocketServer : public Thread { - public: - explicit SocketServer(int port) - : port_(port), socket_(0), maxPendingConnections_(100) {} - - virtual void run(); - - protected: - int port_; - int socket_; - int maxPendingConnections_; -}; - -void SocketServer::run() { - int newsockfd; - socklen_t clilen; - struct sockaddr_in serv_addr, cli_addr; - - /* First call to socket() function */ - socket_ = socket(AF_INET, SOCK_STREAM, 0); - CHECK(socket_ >= 0) << "ERROR opening socket"; - - /* Initialize socket structure */ - bzero((char*)&serv_addr, sizeof(serv_addr)); - serv_addr.sin_family = AF_INET; - serv_addr.sin_addr.s_addr = INADDR_ANY; - serv_addr.sin_port = htons(port_); - - /* Now bind the host address using bind() call.*/ - CHECK(bind(socket_, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0) - << "ERROR on binding"; - - /* Now start listening for the clients, here process will - * go in sleep mode and will wait for the incoming connection - */ - listen(socket_, maxPendingConnections_); - clilen = sizeof(cli_addr); - - while (true) { - /* Accept actual connection from the client */ - newsockfd = accept(socket_, (struct sockaddr*)&cli_addr, &clilen); - CHECK(newsockfd >= 0) << "ERROR on accept"; - - SocketWorker* worker = new SocketWorker(newsockfd); - worker->start(); - } -} - -void SocketWorker::run() { - MessageHeader header; - - while (true) { - int64_t n = channel_.readAll(&header, sizeof(header)); - CHECK(n == sizeof(header)) << "ERROR reading from socket"; - - buffer_.resize(header.dataLength); - n = channel_.readAll(&buffer_[0], header.dataLength); - CHECK(n == header.dataLength) << "ERROR reading from socket"; - - /* Write a response to the client */ - n = channel_.writeAll(&header, sizeof(header)); - CHECK(n == sizeof(header)) << "ERROR reading from socket"; - n = channel_.writeAll(buffer_.data(), buffer_.size()); - CHECK(n == header.dataLength) << "ERROR writing to socket"; - } -} - -class SocketClient { - public: - SocketClient(const std::string& serverAddr, int serverPort); - SocketChannel* getChannel() const { return channel_.get(); } - - protected: - std::unique_ptr channel_; -}; - -SocketClient::SocketClient(const std::string& serverAddr, int serverPort) { - struct sockaddr_in serv_addr; - struct hostent* server; - - // char buffer[256]; - - /* Create a socket point */ - int sockfd = socket(AF_INET, SOCK_STREAM, 0); - CHECK(sockfd >= 0) << "ERROR opening socket"; - server = gethostbyname(serverAddr.c_str()); - CHECK(server) << "ERROR, no such host: " << serverAddr; - - bzero((char*)&serv_addr, sizeof(serv_addr)); - serv_addr.sin_family = AF_INET; - bcopy((char*)server->h_addr, - (char*)&serv_addr.sin_addr.s_addr, - server->h_length); - serv_addr.sin_port = htons(serverPort); - - /* Now connect to the server */ - CHECK(connect(sockfd, (sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0) - << "ERROR connecting"; - - channel_.reset(new SocketChannel(sockfd)); -} - -DEFINE_string(server_addr, "127.0.0.1", "Server address"); -DEFINE_int64(dim, 10000000, "Data size"); -DEFINE_int32(loop_time, 100000, "test loop time"); - -using namespace paddle; // NOLINT - -int main(int argc, char** argv) { - paddle::initMain(argc, argv); - SocketServer server(FLAGS_port); - server.start(); - sleep(1); - - SocketClient client(FLAGS_server_addr, FLAGS_port); - - SocketChannel* channel = client.getChannel(); - - MessageHeader header; - - uint64_t dataSize = FLAGS_dim * sizeof(real); - -#ifdef PADDLE_WITH_CUDA - GpuVector gpuParam(FLAGS_dim); - GpuVector gpuGrad(FLAGS_dim); -#else - CpuVector gpuParam(FLAGS_dim); - CpuVector gpuGrad(FLAGS_dim); -#endif - CpuVector cpuParam(FLAGS_dim); - CpuVector cpuGrad(FLAGS_dim); - - gpuParam.rand(); - gpuGrad.rand(); - cpuParam.rand(); - cpuGrad.rand(); - - for (int i = 0; i < FLAGS_loop_time; ++i) { - cpuGrad.copyFrom(gpuGrad); - - header.dataLength = dataSize; - CHECK(channel->writeAll(&header, sizeof(header)) == sizeof(header)) - << "Client write header error"; - - CHECK(channel->writeAll(cpuGrad.getData(), dataSize) == dataSize) - << "Client write data error"; - - /* Now read server response */ - CHECK(channel->readAll(&header, sizeof(header)) == sizeof(header)) - << "Client read header error"; - - CHECK_EQ((uint64_t)header.dataLength, dataSize); - CHECK(channel->readAll(cpuParam.getData(), dataSize) == dataSize) - << "Client read data error"; - - gpuParam.copyFrom(cpuParam); - - LOG_EVERY_N(INFO, 100) << "i=" << i; - } - exit(0); -} diff --git a/paddle/pserver/test/test_ParameterServer2.cpp b/paddle/pserver/test/test_ParameterServer2.cpp deleted file mode 100644 index 01d179258dffaf996a57022801ee3bd60a268f77..0000000000000000000000000000000000000000 --- a/paddle/pserver/test/test_ParameterServer2.cpp +++ /dev/null @@ -1,624 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include -#include -#include - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -DECLARE_int32(num_gradient_servers); -DEFINE_string(server_addr, "127.0.0.1", "assign server address"); -DEFINE_int32(server_cpu, 0, "assign server cpu"); - -class ParameterServer2Tester : public ParameterServer2 { - public: - ParameterServer2Tester(std::string serverAddr, - int port, - int rdmaCpu = -1, - bool sepSendAndRecv = false) - : ParameterServer2(serverAddr, port, rdmaCpu), client_(sepSendAndRecv) {} - virtual ~ParameterServer2Tester() {} - void setup() { - CHECK(ParameterServer2::init()); - - parameters_.clear(); - clientConfigs_.clear(); - - clientConfigs_.resize(2); - { - ParameterConfig& config = clientConfigs_[0]; - config.set_name("para0"); - config.set_para_id(0); - config.set_size(10000); - config.set_device(-1); - config.set_learning_rate(1.0); - config.set_momentum(0.9); - } - - { - ParameterConfig& config = clientConfigs_[1]; - config.set_name("para1"); - config.set_para_id(1); - config.set_size(5000); - config.set_device(-1); - config.set_learning_rate(0.5); - config.set_momentum(0.4); - } - - for (auto& config : clientConfigs_) { - parameters_.emplace_back(new Parameter(config, /* useGpu= */ false)); - } - - size_t id = 0; - for (auto& para : parameters_) { - para->setID(id++); - } - - CHECK(client_.init(parameters_)); - OptimizationConfig optConfig; - optConfig.set_algorithm("async_sgd"); - optConfig.set_batch_size(100); - optConfig.set_learning_rate(0.1); - client_.setConfig(optConfig); - client_.setParameter(); - } - - void setConfigTest(); - void setStatusTest(); - void sendParameterTest(); - void sendDataTest(SendDataType type, size_t size); - void operationTest(); - void mergeBlockSegmentTest(); - void checkSegments(const BlockSegments& expected, const BlockSegments& segs); - void waitPassFinishTest(); - void synchronizeTest(); - - protected: - ParameterClient2 client_; - vector clientConfigs_; - vector parameters_; -}; - -std::unique_ptr g_server; - -void ParameterServer2Tester::setConfigTest() { - setup(); - - for (auto& config : clientConfigs_) { - auto it = configMap_.find(config.para_id()); - EXPECT_TRUE(it != configMap_.end()); - auto& serverConfig = it->second; - EXPECT_EQ(config.name(), serverConfig.name()); - EXPECT_EQ(config.size(), serverConfig.size()); - EXPECT_EQ(config.learning_rate(), serverConfig.learning_rate()); - EXPECT_EQ(config.momentum(), serverConfig.momentum()); - } -} - -void ParameterServer2Tester::setStatusTest() { - setup(); - EXPECT_TRUE(client_.inStatus(PSERVER_STATUS_NOT_SET)); - client_.setStatus(PSERVER_STATUS_PARAMETER_READY); - EXPECT_EQ(PSERVER_STATUS_PARAMETER_READY, status_); - EXPECT_TRUE(client_.inStatus(PSERVER_STATUS_PARAMETER_READY)); -} - -real sumVector(const CpuVector& vec) { - const real* data = vec.getData(); - size_t dim = vec.getSize(); - real sum = 0; - for (size_t i = 0; i < dim; ++i) { - sum += data[i]; - } - return sum; -} - -void ParameterServer2Tester::sendParameterTest() { - setup(); - - client_.sendAndReceiveParameter(PSERVER_UPDATE_MODE_SET_PARAM, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - false); // sendBackParameter = false - - vector parameterCopies; - - for (auto& parameter : parameters_) { - parameterCopies.emplace_back( - new Parameter(parameter->getConfig(), /* useGpu= */ false)); - parameterCopies.back() - ->getBuf(PARAMETER_VALUE) - ->copyFrom(*parameter->getBuf(PARAMETER_VALUE)); - } - - client_.sendAndReceiveParameter(PSERVER_UPDATE_MODE_GET_PARAM, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true); // sendBackParameter = true - - for (size_t i = 0; i != parameters_.size(); ++i) { - real* v1 = parameters_[i]->getBuf(PARAMETER_VALUE)->getData(); - real* v2 = parameterCopies[i]->getBuf(PARAMETER_VALUE)->getData(); - EXPECT_EQ(parameters_[i]->getSize(), parameterCopies[i]->getSize()); - size_t size = parameters_[i]->getSize(); - real sum1 = 0, sum2 = 0; - for (size_t j = 0; j < size; ++j) { - sum1 += v1[j]; - sum2 += v2[j]; - } - EXPECT_EQ(sum1, sum2); - } -} - -void ParameterServer2Tester::sendDataTest(SendDataType type, size_t size) { - ParameterClient2 client1(true); - client1.init(parameters_); - ParameterClient2 client2(true); - client2.init(parameters_); - ParameterClient2 client3(true); - client3.init(parameters_); - - ThreadWorker worker1; - ThreadWorker worker2; - ThreadWorker worker3; - - double* testData1 = new double[size]; - double* testData2 = new double[size]; - double* testData3 = new double[size]; - double* getDataExpect = new double[size]; - double* getDataReal = new double[size]; - for (size_t i = 0; i < size; ++i) { - testData1[i] = rand(); // NOLINT TODO(yuyang18): Use rand_r instead. - testData2[i] = rand(); // NOLINT - testData3[i] = rand(); // NOLINT - getDataExpect[i] = testData1[i] + testData2[i] + testData3[i]; - } - - auto put1 = [&]() { - LOG(INFO) << "putOwnData1 start"; - client1.putOwnData(0, type, testData1, size); - LOG(INFO) << "putOwnData1 finish"; - }; - - auto get1 = [&]() { - LOG(INFO) << "sendData1 get all start"; - client1.getAllData(0, type, getDataReal, size); - for (size_t i = 0; i < size; ++i) { - CHECK_EQ(getDataReal[i], getDataExpect[i]); - } - LOG(INFO) << "sendData1 get all finish"; - }; - - auto put2 = [&]() { - LOG(INFO) << "putOwnData2 start"; - client2.putOwnData(1, type, testData2, size); - LOG(INFO) << "putOwnData2 finish"; - }; - - auto put3 = [&]() { - LOG(INFO) << "putOwnData3 start"; - client3.putOwnData(2, type, testData3, size); - LOG(INFO) << "putOwnData3 finish"; - }; - - worker1.addJob(put1); - worker1.addJob(get1); - worker2.addJob(put2); - worker3.addJob(put3); - - worker1.addJob(put1); - worker2.addJob(put2); - worker3.addJob(put3); - worker1.addJob(get1); - - worker1.wait(); - worker2.wait(); - worker3.wait(); - free(testData1); - free(testData2); - free(testData3); - free(getDataExpect); - free(getDataReal); -} - -void ParameterServer2Tester::operationTest() { - PServerVector v1, v2; - v1 = client_.createVector(); - EXPECT_EQ(NUM_PARAMETER_TYPES, v1.handle); - - v2 = client_.createVector(); - EXPECT_EQ(NUM_PARAMETER_TYPES + 1, v2.handle); - - PreparedOperations ops; - ops.addOperation(PSERVER_OP_RESET, v1, (real)1); - ops.addOperation(PSERVER_OP_RESET, v2, (real)2); - - real res1, res2, res3; - ops.addOperation(PSERVER_OP_utv, v1, v2)(&res1); - - ops.addOperation(PSERVER_OP_au_bv, v1, v2, (real)-1, (real)1); - ops.addOperation(PSERVER_OP_utv, v1, v2)(&res2); - - ops.addOperation(PSERVER_OP_au_bv, v1, v2, (real)-1, (real)1); - ops.addOperation(PSERVER_OP_utv, v1, v2)(&res3); - client_.doOperation(ops, false, false); - - EXPECT_EQ(30000, res1); - EXPECT_EQ(15000, res2); - EXPECT_EQ(0, res3); - - PServerMatrix m1, m2; - m1 = client_.createMatrix(4); - EXPECT_EQ(0, m1.handle); - m2 = client_.createMatrix(8); - EXPECT_EQ(1, m2.handle); - - // TODO(yuyang18): add tests for other operations OP_COPY, OP_au - - client_.releaseVector(v1); - client_.releaseVector(v2); - client_.releaseMatrix(m1); - client_.releaseMatrix(m2); -} - -void ParameterServer2Tester::checkSegments(const BlockSegments& expected, - const BlockSegments& segs) { - EXPECT_EQ(expected.size(), segs.size()); - if (expected.size() != segs.size()) { - return; - } - for (size_t i = 0; i < expected.size(); ++i) { - EXPECT_EQ(expected[i], segs[i]); - } -} - -void ParameterServer2Tester::mergeBlockSegmentTest() { - { - BlockSegments segs{{10, 20}, {30, 45}, {50, 70}}; - mergeSegments(&segs); - checkSegments({{10, 20}, {30, 45}, {50, 70}}, segs); - } - { - BlockSegments segs{{30, 45}, {50, 70}, {10, 20}}; - mergeSegments(&segs); - checkSegments({{10, 20}, {30, 45}, {50, 70}}, segs); - } - { - BlockSegments segs{{30, 45}, {50, 70}, {10, 30}}; - mergeSegments(&segs); - checkSegments({{10, 45}, {50, 70}}, segs); - } - { - BlockSegments segs{{30, 45}, {10, 70}, {10, 30}}; - mergeSegments(&segs); - checkSegments({{10, 70}}, segs); - } - { - BlockSegments segs{{30, 45}, {50, 70}, {10, 35}}; - mergeSegments(&segs); - checkSegments({{10, 45}, {50, 70}}, segs); - } - { - BlockSegments segs{{30, 45}, {50, 70}, {10, 60}}; - mergeSegments(&segs); - checkSegments({{10, 70}}, segs); - } - { - BlockSegments segs{{30, 45}, {50, 70}, {30, 47}}; - mergeSegments(&segs); - checkSegments({{30, 47}, {50, 70}}, segs); - } -} - -void ParameterServer2Tester::waitPassFinishTest() { - ParameterClient2 client1; - ParameterClient2 client2; - ParameterClient2 client3; - - ThreadWorker worker1; - ThreadWorker worker2; - ThreadWorker worker3; - - auto init1 = [&]() { - LOG(INFO) << "init1 start"; - client1.init(parameters_); - LOG(INFO) << "init1 finish"; - }; - - auto init2 = [&]() { - LOG(INFO) << "init2 start"; - client2.init(parameters_); - LOG(INFO) << "init2 finish"; - }; - - auto init3 = [&]() { - LOG(INFO) << "init3 start"; - client3.init(parameters_); - LOG(INFO) << "init3 finish"; - }; - - auto update1 = [&]() { - LOG(INFO) << "update1 start"; - client1.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ADD_GRADIENT, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true); // sendBackParameter = false - LOG(INFO) << "update1 finish"; - }; - - auto wait1 = [&]() { - LOG(INFO) << "wait1 start"; - client1.waitPassFinish(); - LOG(INFO) << "wait1 finish"; - }; - - auto update2 = [&]() { - LOG(INFO) << "update2 start"; - client2.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ADD_GRADIENT, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true); // sendBackParameter = false - LOG(INFO) << "update2 finish"; - }; - - auto wait2 = [&]() { - LOG(INFO) << "wait2 start"; - client2.waitPassFinish(); - LOG(INFO) << "wait2 finish"; - }; - - auto op3 = [&]() { - LOG(INFO) << "op3 start"; - PreparedOperations ops; - ops.addOperation(PSERVER_OP_SGD); - client3.doOperation(ops, - /* waitForGradient= */ true, - /* sendBackarameter= */ true); - LOG(INFO) << "op3 finish"; - }; - - worker1.addJob(init1); - worker2.addJob(init2); - worker3.addJob(init3); - - worker1.addJob(update1); - worker2.addJob(update2); - worker3.addJob(op3); - - worker3.addJob(op3); - worker3.addJob(op3); - worker2.addJob(update2); - worker2.addJob(update2); - worker1.addJob(wait1); - - worker2.addJob(wait2); - worker3.addJob(op3); - - worker1.wait(); - worker2.wait(); - worker3.wait(); - - LOG(INFO) << "Pass 1 finished"; - - worker1.addJob(update1); - worker2.addJob(update2); - worker3.addJob(op3); - - worker1.wait(); - worker2.wait(); - worker3.wait(); - - worker3.addJob(op3); - worker3.addJob(op3); - worker1.addJob(update1); - worker1.addJob(wait1); - worker2.addJob(wait2); - - worker1.wait(); - worker2.wait(); - worker3.wait(); - - LOG(INFO) << "Pass 2 finished"; -} - -void ParameterServer2Tester::synchronizeTest() { - ParameterClient2 client1; - ParameterClient2 client2; - - ThreadWorker worker1; - ThreadWorker worker2; - - FLAGS_log_period_server = 2; - - auto init1 = [&]() { - LOG(INFO) << "init1 start"; - client1.init(parameters_); - client1.setTrainerId(0); - LOG(INFO) << "init1 finish"; - }; - - auto init2 = [&]() { - LOG(INFO) << "init2 start"; - client2.init(parameters_); - client2.setTrainerId(1); - LOG(INFO) << "init2 finish"; - }; - - auto update1 = [&]() { - LOG(INFO) << "update1 start"; - client1.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ASYNC_SGD, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true); // sendBackParameter = false - LOG(INFO) << "update1 finish"; - }; - - auto wait1 = [&]() { - LOG(INFO) << "wait1 start"; - client1.asyncFinishPass(); - LOG(INFO) << "wait1 finish"; - }; - - auto update2 = [&]() { - LOG(INFO) << "update2 start"; - client2.sendAndReceiveParameter(PSERVER_UPDATE_MODE_ASYNC_SGD, - PARAMETER_VALUE, - 0, // numSamples = 0 - 0, // cost = 0 - true); // sendBackParameter = false - LOG(INFO) << "update2 finish"; - }; - - auto wait2 = [&]() { - LOG(INFO) << "wait2 start"; - client2.asyncFinishPass(); - LOG(INFO) << "wait2 finish"; - }; - - worker1.addJob(init1); - worker2.addJob(init2); - // call wait to reset some stats at pserver - worker1.addJob(wait1); - worker2.addJob(wait2); - - worker1.addJob(update1); - worker2.addJob(update2); - - worker2.addJob(update2); - worker2.addJob(update2); - worker1.addJob(wait1); - - worker2.addJob(wait2); - - worker1.wait(); - worker2.wait(); - LOG(INFO) << "Pass 1 finished"; - - worker1.addJob(update1); - worker2.addJob(update2); - - worker1.wait(); - worker2.wait(); - - worker1.addJob(update1); - worker2.addJob(update2); - worker1.addJob(update1); - worker1.addJob(update1); - worker1.addJob(update1); - worker1.addJob(update1); - worker1.addJob(update1); - worker1.addJob(update1); - worker1.addJob(wait1); - worker2.addJob(wait2); - - worker1.wait(); - worker2.wait(); - LOG(INFO) << "Pass 2 finished"; -} - -TEST(ParameterServer2, sendParameter) { g_server->sendParameterTest(); } - -TEST(ParameterServer2, setConfig) { g_server->setConfigTest(); } - -TEST(ParameterServer2, setStatus) { g_server->setStatusTest(); } - -TEST(ParameterServer2, operation) { g_server->operationTest(); } - -TEST(ParameterServer2, mergeBlockSegment) { g_server->mergeBlockSegmentTest(); } - -TEST(ParameterServer2, waitPassFinish) { g_server->waitPassFinishTest(); } - -TEST(ParameterServer2, synchronize) { g_server->synchronizeTest(); } - -TEST(ParameterServer2, sendData) { - // Set gserver and pserver all 3, so that the test is sufficient. - int oldFlagsPortsNUm = FLAGS_ports_num; - int oldFlagsNumGradientServers = FLAGS_num_gradient_servers; - int oldFlagsPort = FLAGS_port; - FLAGS_ports_num = 3; - FLAGS_num_gradient_servers = 3; - FLAGS_port = FLAGS_port + 1; - std::unique_ptr g_server1; - std::unique_ptr g_server2; - std::unique_ptr g_server3; - if (FLAGS_rdma_tcp == "rdma") { - g_server1.reset(new ParameterServer2Tester( - FLAGS_server_addr, FLAGS_port, FLAGS_server_cpu)); - g_server1->start(); - g_server2.reset(new ParameterServer2Tester( - FLAGS_server_addr, FLAGS_port + 1, FLAGS_server_cpu + 1)); - g_server2->start(); - g_server3.reset(new ParameterServer2Tester( - FLAGS_server_addr, FLAGS_port + 2, FLAGS_server_cpu + 2)); - g_server3->start(); - } else { // tcp - g_server1.reset(new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port)); - g_server1->start(); - g_server2.reset( - new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port + 1)); - g_server2->start(); - g_server3.reset( - new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port + 2)); - g_server3->start(); - } - - g_server2->init(); - g_server3->init(); - sleep(2); - g_server1->setup(); - g_server1->sendDataTest(DATA_REDUCE_SUM, 1 << 24); - sleep(2); - g_server1->sendDataTest(DATA_REDUCE_SUM, 2); - sleep(2); - g_server1.reset(); - g_server2.reset(); - g_server3.reset(); - - FLAGS_ports_num = oldFlagsPortsNUm; - FLAGS_num_gradient_servers = oldFlagsNumGradientServers; - FLAGS_port = oldFlagsPort; -} - -int main(int argc, char** argv) { - paddle::initMain(argc, argv); - testing::InitGoogleTest(&argc, argv); - - FLAGS_num_gradient_servers = 2; - - if (FLAGS_rdma_tcp == "rdma") { - g_server.reset(new ParameterServer2Tester( - FLAGS_server_addr, FLAGS_port, FLAGS_server_cpu)); - } else { - g_server.reset(new ParameterServer2Tester(FLAGS_server_addr, FLAGS_port)); - } - - g_server->start(); - - sleep(2); - - int ret = RUN_ALL_TESTS(); - - g_server.reset(); - - exit(ret); -} diff --git a/paddle/pserver/test/test_ProtoServer.cpp b/paddle/pserver/test/test_ProtoServer.cpp deleted file mode 100644 index a66b14a1cc58d11988e4936a9c35d98b8bf5edc1..0000000000000000000000000000000000000000 --- a/paddle/pserver/test/test_ProtoServer.cpp +++ /dev/null @@ -1,169 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "ParameterService.pb.h" -#include "paddle/math/Vector.h" -#include "paddle/pserver/ProtoServer.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -DEFINE_string(server_addr, "127.0.0.1", "Server address"); -DEFINE_int64(dim, 50000000, "Data size"); -DEFINE_bool(test_proto_server, true, "whether to test ProtoServer"); -DEFINE_bool(benchmark, false, "Do benchmark. Skip some tests"); - -using namespace paddle; // NOLINT - -class MyServer : public ProtoServer { - public: - explicit MyServer(int port, int rdmaCpu = -1) - : ProtoServer(FLAGS_server_addr, port, rdmaCpu), - status_(PSERVER_STATUS_NOT_SET) { - REGISTER_SERVICE_FUNCTION(MyServer, getStatus); - REGISTER_SERVICE_FUNCTION(MyServer, setStatus); - REGISTER_SERVICE_FUNCTION_EX(MyServer, getStatusEx); - } - void getStatus(const GetStatusRequest& request, - ProtoResponseCallback callback) { - (void)request; - GetStatusResponse response; - response.set_status(status_); - callback(response); - } - - void getStatusEx(const GetStatusRequest& request, - std::unique_ptr msgReader, - ProtoResponseCallbackEx callback) { - (void)request; - GetStatusResponse response; - response.set_status(status_); - buffer_.resize(msgReader->getNextBlockLength()); - msgReader->readNextBlock(&buffer_[0]); - callback(response, {{&buffer_[0], buffer_.size()}}); - } - - void setStatus(const SetStatusRequest& request, - ProtoResponseCallback callback) { - SetStatusResponse response; - status_ = request.status(); - callback(response); - } - - protected: - PServerStatus status_; - std::string buffer_; -}; - -TEST(ProtoServer, regular) { - ProtoClient* client; - if (FLAGS_rdma_tcp == "rdma") - client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA); - else - client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_TCP); - { - GetStatusRequest request; - GetStatusResponse response; - auto msgReader = client->sendAndRecv("getStatus", request, &response); - EXPECT_EQ(response.status(), PSERVER_STATUS_NOT_SET); - EXPECT_EQ(msgReader->getNumBlocks(), (size_t)0); - } - - { - SetStatusRequest request; - SetStatusResponse response; - request.set_status(PSERVER_STATUS_PARAMETER_READY); - client->sendAndRecv("setStatus", request, &response); - } - - { - GetStatusRequest request; - GetStatusResponse response; - client->sendAndRecv("getStatus", request, &response); - EXPECT_EQ(response.status(), PSERVER_STATUS_PARAMETER_READY); - } - - delete client; -} - -TEST(ProtoServer, extended) { -#ifdef PADDLE_WITH_CUDA - ProtoClient* client; - if (FLAGS_rdma_tcp == "rdma") - client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA); - else - client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_TCP); - int64_t dataSize = FLAGS_dim * sizeof(real); - - GpuVector gpuParam(FLAGS_dim); - GpuVector gpuGrad(FLAGS_dim); - CpuVector cpuParam(FLAGS_dim); - CpuVector cpuGrad(FLAGS_dim); - - gpuParam.rand(); - gpuGrad.rand(); - cpuParam.rand(); - cpuGrad.rand(); - - for (int k = 0; k < 4; ++k) { - for (int i = 0; i < 10; ++i) { - cpuGrad.copyFrom(gpuGrad); - if (FLAGS_test_proto_server) { - GetStatusRequest request; - GetStatusResponse response; - { - REGISTER_TIMER("sendAndRecv"); - auto msgReader = - client->sendAndRecv("getStatusEx", - request, - {{cpuGrad.getData(), (size_t)dataSize}}, - &response); - - EXPECT_EQ(msgReader->getNumBlocks(), (size_t)1); - EXPECT_EQ(msgReader->getNextBlockLength(), (size_t)dataSize); - msgReader->readNextBlock(cpuParam.getData()); - } - if (!FLAGS_benchmark) { - real* v1 = cpuGrad.getData(); - real* v2 = cpuParam.getData(); - real sum1 = 0, sum2 = 0; - for (int j = 0; j < FLAGS_dim; ++j) { - sum1 += v1[j]; - sum2 += v2[j]; - } - EXPECT_EQ(sum1, sum2); - } - } - gpuParam.copyFrom(cpuParam); - - LOG_EVERY_N(INFO, 10) << "i=" << i; - } - globalStat.printAllStatus(); - globalStat.reset(); - } - - delete client; -#endif -} - -int main(int argc, char** argv) { - paddle::initMain(argc, argv); - testing::InitGoogleTest(&argc, argv); - MyServer server(FLAGS_port, FLAGS_rdma_tcp == "rdma" ? 0 : -1); - server.start(); - usleep(10000); - - return RUN_ALL_TESTS(); -} diff --git a/paddle/pserver/test/test_ProtoServer.sh b/paddle/pserver/test/test_ProtoServer.sh deleted file mode 100755 index 970c90b494c2a256cf22f3de7b7ea7964fed58ab..0000000000000000000000000000000000000000 --- a/paddle/pserver/test/test_ProtoServer.sh +++ /dev/null @@ -1,33 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -set -x -for ((port=12340;port<=12360;port++)) -do - port_used_num=`netstat -a |grep $port|wc -l` - if [ $port_used_num -eq 0 ] - then - echo $port; - pserver/test/test_ProtoServer --port=$port - if [ $? -eq 0 ] - then - exit 0 - else - echo "test_ProtoServer run wrong" - exit 1 - fi -fi -done -echo "test_ProtoServer port not found" -exit 1 diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index 8eeea1805d8610f6f27f422337f3526688b73de3..8460f93b841fe136db138e0dc7576f3aacdbeb5f 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -19,10 +19,12 @@ # Utils #================================================= +set -ex + function print_usage() { echo -e "\n${RED}Usage${NONE}: ${BOLD}${SCRIPT_NAME}${NONE} [OPTION]" - + echo -e "\n${RED}Options${NONE}: ${BLUE}build${NONE}: run build for x86 platform ${BLUE}build_android${NONE}: run build for android platform @@ -37,6 +39,7 @@ function print_usage() { ${BLUE}fluid_inference_lib${NONE}: deploy fluid inference library ${BLUE}check_style${NONE}: run code style check ${BLUE}cicheck${NONE}: run CI tasks + ${BLUE}assert_api_not_changed${NONE}: check api compability " } @@ -78,6 +81,12 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" + elif [ "$1" == "cp35-cp35m" ]; then + export LD_LIBRARY_PATH=/opt/_internal/cpython-3.5.1/lib/:${LD_LIBRARY_PATH} + export PATH=/opt/_internal/cpython-3.5.1/bin/:${PATH} + export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.5.1/bin/python3 + -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.5.1/include/python3.5m + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.5.1/lib/libpython3.so" fi fi @@ -106,6 +115,9 @@ function cmake_gen() { -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DWITH_CONTRIB=${WITH_CONTRIB:-ON} + -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} + -DWITH_INFERENCE_DEMO=${WITH_INFERENCE_DEMO:-ON} + -DPY_VERSION=${PY_VERSION:-2.7} ======================================== EOF # Disable UNITTEST_USE_VIRTUALENV in docker because @@ -132,7 +144,10 @@ EOF -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ - -DWITH_CONTRIB=${WITH_CONTRIB:-ON} + -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \ + -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} \ + -DWITH_INFERENCE_DEMO=${WITH_INFERENCE_DEMO:-ON} \ + -DPY_VERSION=${PY_VERSION:-2.7} } function abort(){ @@ -145,19 +160,17 @@ function check_style() { trap 'abort' 0 set -e - # install glide - curl https://glide.sh/get | bash - eval "$(GIMME_GO_VERSION=1.8.3 gimme)" + if [ -x "$(command -v gimme)" ]; then + eval "$(GIMME_GO_VERSION=1.8.3 gimme)" + fi # set up go environment for running gometalinter mkdir -p $GOPATH/src/github.com/PaddlePaddle/ ln -sf ${PADDLE_ROOT} $GOPATH/src/github.com/PaddlePaddle/Paddle - cd $GOPATH/src/github.com/PaddlePaddle/Paddle/go; glide install; cd - - - go get github.com/alecthomas/gometalinter - gometalinter --install + mkdir -p ./build/go + cp go/glide.* build/go + cd build/go; glide install; cd - - cd ${PADDLE_ROOT} export PATH=/usr/bin:$PATH pre-commit install clang-format --version @@ -183,6 +196,7 @@ function build() { ============================================ EOF make clean + make -j `nproc` make install -j `nproc` } @@ -198,7 +212,7 @@ function build_android() { fi ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API - + cat < new.spec + python ${PADDLE_ROOT}/tools/diff_api.py ${PADDLE_ROOT}/paddle/fluid/API.spec new.spec + deactivate + + API_CHANGE=`git diff --name-only upstream/develop | grep "paddle/fluid/API.spec" || true` + echo "checking API.spec change, PR: ${GIT_PR_ID}, changes: ${API_CHANGE}" + if [ ${API_CHANGE} ] && [ "${GIT_PR_ID}" != "" ]; then + # TODO: curl -H 'Authorization: token ${TOKEN}' + APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews | \ + python ${PADDLE_ROOT}/tools/check_pr_approval.py 2 7845005 2887803 728699 13348433` + echo "current pr ${GIT_PR_ID} got approvals: ${APPROVALS}" + if [ "${APPROVALS}" == "FALSE" ]; then + echo "You must have at least 2 approvals for the api change!" + exit 1 + fi + fi +} + + function single_test() { TEST_NAME=$1 if [ -z "${TEST_NAME}" ]; then @@ -331,14 +370,14 @@ EOF function bind_test() { # the number of process to run tests NUM_PROC=6 - + # calculate and set the memory usage for each process MEM_USAGE=$(printf "%.2f" `echo "scale=5; 1.0 / $NUM_PROC" | bc`) export FLAGS_fraction_of_gpu_memory_to_use=$MEM_USAGE - + # get the CUDA device count CUDA_DEVICE_COUNT=$(nvidia-smi -L | wc -l) - + for (( i = 0; i < $NUM_PROC; i++ )); do cuda_list=() for (( j = 0; j < $CUDA_DEVICE_COUNT; j++ )); do @@ -380,6 +419,25 @@ EOF linkchecker doc/v2/en/html/index.html linkchecker doc/v2/cn/html/index.html linkchecker doc/v2/api/en/html/index.html + + if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi; + + # Deploy to the the content server if its a "develop" or "release/version" branch + # The "develop_doc" branch is reserved to test full deploy process without impacting the real content. + if [ "$TRAVIS_BRANCH" == "develop_doc" ]; then + PPO_SCRIPT_BRANCH=develop + elif [[ "$TRAVIS_BRANCH" == "develop" || "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then + PPO_SCRIPT_BRANCH=master + else + # Early exit, this branch doesn't require documentation build + return 0; + fi + # Fetch the paddlepaddle.org deploy_docs.sh from the appopriate branch + export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/$PPO_SCRIPT_BRANCH/scripts/deploy/deploy_docs.sh + export PYTHONPATH=$PYTHONPATH:${PADDLE_ROOT}/build/python:/paddle/build/python + cd .. + curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH ${PADDLE_ROOT} ${PADDLE_ROOT}/build/doc/ ${PPO_SCRIPT_BRANCH} + cd - } function gen_html() { @@ -449,7 +507,7 @@ EOF # run paddle version to install python packages first RUN apt-get update &&\ ${NCCL_DEPS}\ - apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \ + apt-get install -y wget python-pip python-opencv libgtk2.0-dev dmidecode python-tk && easy_install -U pip && \ pip install /*.whl; apt-get install -f -y && \ apt-get clean -y && \ rm -f /*.whl && \ @@ -491,15 +549,28 @@ function gen_fluid_inference_lib() { Deploying fluid inference library ... ======================================== EOF + cmake .. -DWITH_DISTRIBUTE=OFF make -j `nproc` inference_lib_dist cd ${PADDLE_ROOT}/build - mv fluid_install_dir fluid - tar -cf fluid.tgz fluid + cp -r fluid_install_dir fluid + tar -czf fluid.tgz fluid + fi +} + +function test_fluid_inference_lib() { + if [ ${WITH_C_API:-OFF} == "OFF" ] ; then + cat < -#include "paddle/utils/Util.h" +#include "paddle/legacy/utils/Util.h" int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); diff --git a/paddle/testing/TestUtil.cpp b/paddle/testing/TestUtil.cpp index cfb8c713d96008a74287fb1248657c30f3b81164..fa8efc20f59addb4526d2cbeaf34f161307c588a 100644 --- a/paddle/testing/TestUtil.cpp +++ b/paddle/testing/TestUtil.cpp @@ -14,7 +14,7 @@ limitations under the License. */ #include "TestUtil.h" #include -#include "paddle/math/SparseMatrix.h" +#include "paddle/legacy/math/SparseMatrix.h" DEFINE_int32(fixed_seq_length, 0, "Produce some sequence of fixed length"); diff --git a/paddle/testing/TestUtil.h b/paddle/testing/TestUtil.h index ec86469aebbafbf5406a21e6825eda6c105a6b9d..98b864e3c56f1938075bd039ba13a49ec457de50 100644 --- a/paddle/testing/TestUtil.h +++ b/paddle/testing/TestUtil.h @@ -15,7 +15,7 @@ limitations under the License. */ #pragma once #include -#include "paddle/math/Matrix.h" +#include "paddle/legacy/math/Matrix.h" namespace paddle { diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc index 586ec48477f085a14d2f15b265a95d596705694f..cfea2059c3ce20fb44732d990e9708ad6f8d81a1 100644 --- a/paddle/testing/paddle_gtest_main.cc +++ b/paddle/testing/paddle_gtest_main.cc @@ -16,8 +16,8 @@ limitations under the License. */ #include "gflags/gflags.h" #include "gtest/gtest.h" -#include "paddle/fluid/framework/init.h" #include "paddle/fluid/memory/memory.h" +#include "paddle/fluid/platform/init.h" int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); @@ -30,7 +30,9 @@ int main(int argc, char** argv) { new_argv.push_back( strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory")); #else - new_argv.push_back(strdup("--tryfromenv=use_pinned_memory")); + new_argv.push_back(strdup( + "--tryfromenv=use_pinned_memory,use_mkldnn,initial_cpu_memory_in_mb")); + new_argv.push_back(strdup("--undefok=use_mkldnn,initial_cpu_memory_in_mb")); #endif int new_argc = static_cast(new_argv.size()); char** new_argv_address = new_argv.data(); diff --git a/paddle/trainer/MergeModel.cpp b/paddle/trainer/MergeModel.cpp deleted file mode 100644 index 56c38015fb2398f8b39fac6b5a5d4af1c2fd56aa..0000000000000000000000000000000000000000 --- a/paddle/trainer/MergeModel.cpp +++ /dev/null @@ -1,64 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "ParamUtil.h" -#include "Trainer.h" -#include "paddle/pserver/ParameterServer2.h" -#include "paddle/utils/PythonUtil.h" - -DEFINE_string(model_dir, "", "Directory for separated model files"); -DEFINE_string(config_file, "", "Config file for the model"); -DEFINE_string(model_file, "", "File for merged model file"); - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -int main(int argc, char** argv) { - initMain(argc, argv); - initPython(argc, argv); - - if (FLAGS_model_dir.empty() || FLAGS_config_file.empty() || - FLAGS_model_file.empty()) { - LOG(INFO) << "Usage: ./paddle_merge_model --model_dir=pass-00000 " - "--config_file=config.py --model_file=out.paddle"; - return 0; - } - - string confFile = FLAGS_config_file; -#ifndef PADDLE_WITH_CUDA - FLAGS_use_gpu = false; -#endif - auto config = std::make_shared(confFile); - unique_ptr gradientMachine(GradientMachine::create(*config)); - gradientMachine->loadParameters(FLAGS_model_dir); - - ofstream os(FLAGS_model_file); - - string buf; - config->getConfig().SerializeToString(&buf); - int64_t size = buf.size(); - os.write((char*)&size, sizeof(size)); - CHECK(os) << "Fail to write to " << FLAGS_model_file; - os.write(buf.data(), buf.size()); - vector& parameters = gradientMachine->getParameters(); - for (auto& para : parameters) { - para->save(os); - CHECK(os) << "Fail to write to " << FLAGS_model_file; - } - os.close(); - - return 0; -} diff --git a/paddle/trainer/NewRemoteParameterUpdater.cpp b/paddle/trainer/NewRemoteParameterUpdater.cpp deleted file mode 100644 index 410ac6d95c4d65ce6fb25c05351bb8ddb24473f4..0000000000000000000000000000000000000000 --- a/paddle/trainer/NewRemoteParameterUpdater.cpp +++ /dev/null @@ -1,150 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "NewRemoteParameterUpdater.h" -#include "Trainer.h" -#include "paddle/utils/Stat.h" - -DECLARE_int32(trainer_id); -DECLARE_string(save_dir); - -namespace paddle { -NewRemoteParameterUpdater::NewRemoteParameterUpdater( - const OptimizationConfig &config, const std::string pserverSpec) - : trainerConfig_(config), - parameterClient_(-1), - newParameters_(nullptr), - newGradients_(nullptr), - pserverSpec_(pserverSpec) {} - -NewRemoteParameterUpdater::NewRemoteParameterUpdater( - const OptimizationConfig &config, - const std::string pserverSpec, - const bool useEtcd) - : trainerConfig_(config), - parameterClient_(-1), - newParameters_(nullptr), - newGradients_(nullptr), - pserverSpec_(pserverSpec), - useEtcd_(useEtcd) {} - -void NewRemoteParameterUpdater::init( - const std::vector ¶meters) { - ParameterUpdater::init(parameters); - - // create parameter server client. - if (useEtcd_) { - parameterClient_ = - paddle_new_etcd_pserver_client((char *)pserverSpec_.c_str()); - } else { - parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(), - FLAGS_trainer_id == 0); - } - - // init new parameter and gradient. - newParameters_ = initNewParameter(PARAMETER_VALUE); - newGradients_ = initNewParameter(PARAMETER_GRADIENT); - - // init parameter, one trainer will get the opportunity to int parameter and - // send them to parameter server. Others will get the initialized parameter - // from parameter server - if (paddle_begin_init_params(parameterClient_)) { - LOG(INFO) << "paddle_begin_init_params start"; - // NOTE: convert V1 OptimizatioinConfig proto to V2 OptimizerConfig. - // This makes golang pserver compatible with handy V1 demos. - // TODO(wuyi): Refine or remove these ugly converting lines - OptimizerConfig optimizerConfigV2; - if (trainerConfig_.learning_method() == "momentum") { - optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); - } else if (trainerConfig_.learning_method() == "adagrad") { - optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad); - optimizerConfigV2.mutable_adagrad()->set_epsilon( - trainerConfig_.ada_epsilon()); - } else if (trainerConfig_.learning_method() == "adadelta") { - optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad); - optimizerConfigV2.mutable_adadelta()->set_epsilon( - trainerConfig_.ada_epsilon()); - optimizerConfigV2.mutable_adadelta()->set_rho(trainerConfig_.ada_rou()); - } else if (trainerConfig_.learning_method() == "adam") { - optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adam); - optimizerConfigV2.mutable_adam()->set_beta_1(trainerConfig_.adam_beta1()); - optimizerConfigV2.mutable_adam()->set_beta_2(trainerConfig_.adam_beta2()); - optimizerConfigV2.mutable_adam()->set_epsilon( - trainerConfig_.adam_epsilon()); - } else { - LOG(ERROR) << "got unsupported v1 optimizer config: " - << trainerConfig_.learning_method(); - optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); - } - - if (trainerConfig_.learning_rate_schedule() == "constant") { - optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); - optimizerConfigV2.mutable_const_lr()->set_learning_rate( - trainerConfig_.learning_rate()); - } else if (trainerConfig_.learning_rate_schedule() == "linear") { - optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Linear); - optimizerConfigV2.mutable_linear_lr()->set_learning_rate( - trainerConfig_.learning_rate()); - optimizerConfigV2.mutable_linear_lr()->set_lr_decay_a( - trainerConfig_.learning_rate_decay_a()); - optimizerConfigV2.mutable_linear_lr()->set_lr_decay_b( - trainerConfig_.learning_rate_decay_b()); - } else { - LOG(ERROR) << "got unsupported v1 learning_rate_schedule config: " - << trainerConfig_.learning_rate_schedule() << ", set to const"; - optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); - optimizerConfigV2.mutable_const_lr()->set_learning_rate( - trainerConfig_.learning_rate()); - } - - // overwrite optimizerConfigV2 for per-parameter(layer) configs - for (int i = 0; i < parameterSize(); ++i) { - // FIXME(typhoonzero): paramConfig always have default values, - // how to check if it's default? - // TODO(typhoonzero): log output: optimizerConfigV2.DebugString(); - LOG(INFO) << "trainerConfig_: " << trainerConfig_.DebugString(); - // send param and config to pserver - std::string bytes = optimizerConfigV2.SerializeAsString(); - const char *array = bytes.data(); - int size = (int)bytes.size(); - paddle_init_param( - parameterClient_, *newParameters_[i], (void *)array, size); - } - paddle_finish_init_params(parameterClient_); - LOG(INFO) << "paddle_begin_init_params done"; - } else { - paddle_get_params(parameterClient_, newParameters_, parameterSize()); - } - - LOG(INFO) << "NewRemoteParameterUpdater initialized"; -} - -void NewRemoteParameterUpdater::updateImpl(Parameter *para) {} - -void NewRemoteParameterUpdater::finishBatch(real cost) { - // send gradient to parameter server. - paddle_send_grads(parameterClient_, newGradients_, parameterSize()); - // get the updated parameter from parameterClient. - paddle_get_params(parameterClient_, newParameters_, parameterSize()); - - // clear gradient after update parameter. - for (auto ¶ : parameters_) { - para->getBuf(PARAMETER_GRADIENT)->zeroMem(); - } -} - -void NewRemoteParameterUpdater::startPass() {} - -bool NewRemoteParameterUpdater::finishPass() { return true; } -} // namespace paddle diff --git a/paddle/trainer/NewRemoteParameterUpdater.h b/paddle/trainer/NewRemoteParameterUpdater.h deleted file mode 100644 index 02693c675e6f5cb574e52e9681963a5904676028..0000000000000000000000000000000000000000 --- a/paddle/trainer/NewRemoteParameterUpdater.h +++ /dev/null @@ -1,121 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "OptimizerConfig.pb.h" -#include "ParameterUpdater.h" -#include "libpaddle_pserver_cclient.h" -#include "paddle/pserver/ParameterClient2.h" -#include "paddle/utils/Queue.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -/** - * New remote parameter updater for dense parameters that use cclient of go. - */ -class NewRemoteParameterUpdater : public ParameterUpdater { - public: - NewRemoteParameterUpdater(const OptimizationConfig& config, - const std::string pserverSpec); - NewRemoteParameterUpdater(const OptimizationConfig& config, - const std::string pserverSpec, - const bool useEtcd); - ~NewRemoteParameterUpdater() { - releaseNewParameter(newParameters_); - releaseNewParameter(newGradients_); - if (parameterClient_ >= 0) paddle_pserver_client_release(parameterClient_); - } - - /** - * initialize the internal parameter client and itself. - */ - virtual void init(const std::vector& parameters); - /** - * @brief start batch - * - * @note one batch training exhibits stateful feature to help - * to do performance tuning, sgd optimization if necessary. - */ - virtual PassType startBatch(int64_t batchSize) { return PASS_TRAIN; } - - /** - * send parameters to pservers and get returned parameters - * from all pservers if necessary. - */ - virtual void finishBatch(real cost); - virtual void startPass(); - virtual bool finishPass(); - - protected: - /** - * work need to do after finishBatch - */ - virtual void updateImpl(Parameter* para); - - private: - int parameterSize() { return (int)parameters_.size(); } - - /** - * init parameter of go paddle pserver cclient. - * @param new_params - * @param type - */ - paddle_parameter** initNewParameter(ParameterType type) { - paddle_parameter** new_params = - (paddle_parameter**)malloc(sizeof(paddle_parameter*) * parameterSize()); - for (int i = 0; i < parameterSize(); ++i) { - new_params[i] = (paddle_parameter*)malloc(sizeof(paddle_parameter)); - memset(new_params[i], 0, sizeof(paddle_parameter)); - } - - for (int i = 0; i < parameterSize(); ++i) { - ParameterPtr param = parameters_[i]; - new_params[i]->element_type = PADDLE_ELEMENT_TYPE_FLOAT32; - new_params[i]->name = (char*)param->getName().c_str(); - new_params[i]->content = - (unsigned char*)(param->getBuf(type).get()->getData()); - new_params[i]->content_len = - (int)param->getBuf(type).get()->getSize() * sizeof(real); - } - return new_params; - } - - void releaseNewParameter(paddle_parameter** newParams) { - if (newParams != nullptr) { - for (int i = 0; i < parameterSize(); ++i) { - free(newParams[i]); - } - free(newParams); - } - } - - protected: - const OptimizationConfig& trainerConfig_; - /// internal parameter client object for exchanging data with pserver - paddle_pserver_client parameterClient_; - /// the parameters for new pserver client - paddle_parameter** newParameters_; - /// the gradinets for new pserver client - paddle_parameter** newGradients_; - /// the specification of parameter server "host1:port,host1:port" - std::string pserverSpec_; - /// true if pserverSpec_ is etcd endpoint, else pserverSpec_ is pserver addr - bool useEtcd_; -}; - -} // namespace paddle diff --git a/paddle/trainer/ParamUtil.cpp b/paddle/trainer/ParamUtil.cpp deleted file mode 100644 index ffbca42e106591ddeb2cefcfafbeb408c544371b..0000000000000000000000000000000000000000 --- a/paddle/trainer/ParamUtil.cpp +++ /dev/null @@ -1,163 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ParamUtil.h" - -#include -#include - -#include -#include -#include -#include - -#include -#include - -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -#include "TesterConfig.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/gserver/layers/ValidationLayer.h" - -namespace paddle { - -ParameterUtil::ParameterUtil( - const std::shared_ptr &config, - std::unique_ptr &&intconfig, - const GradientMachinePtr &gradientMachine, - const std::shared_ptr ¶meterUpdater) { - config_ = config; - intConfig_ = std::move(intconfig); - gserver_ = gradientMachine; - pUpdater_ = parameterUpdater; -} - -bool ParameterUtil::loadParameters(int passId, bool local, bool remote) { - constexpr int kBufLen = 100; - char buf[kBufLen]; - snprintf(buf, kBufLen, "pass-%05d", passId); - std::string doneFile = path::join(config_->getSaveDir(), buf, "done"); - if (!fileExist(doneFile.c_str())) return false; - loadParametersWithPath(path::join(config_->getSaveDir(), buf), local, remote); - return true; -} - -void ParameterUtil::loadParametersWithPath(const std::string &dir, - bool local, - bool remote) { - if (local) { - gserver_->loadParameters(dir); - } - if (remote && pUpdater_) { - pUpdater_->loadParametersRemote(dir); - } -} - -void ParameterUtil::saveParametersOnePass(int passId, int passInnerId) { - pUpdater_->apply(); - saveParameters(passId, passInnerId); - if (intConfig_->save_only_one_ && passId >= intConfig_->saving_period_) { - deleteParameters(passId - intConfig_->saving_period_); - } - pUpdater_->restore(); -} - -void ParameterUtil::saveParameters(int passId, int passInnerId) { - constexpr int kBufLen = 100; - char buf[kBufLen]; - if (passInnerId > 0) { - snprintf(buf, kBufLen, "pass-%05d-%03d", passId, passInnerId); - } else { - snprintf(buf, kBufLen, "pass-%05d", passId); - } - - std::string basePath = config_->getSaveDir(); - if (basePath.find('/') == std::string::npos) { - basePath = "./" + basePath; - } - mkDirRecursively(basePath.c_str()); - - std::string saveDir = path::join(basePath, buf); - mkDir(saveDir.c_str()); - if (!intConfig_->load_save_param_pserver_) { - pUpdater_->getParametersRemote(true /*full parameter*/, - true /*after apply*/); - } - - gserver_->saveParameters(saveDir); - if (intConfig_->load_save_param_pserver_) { - pUpdater_->saveParametersRemote(saveDir); - } - std::string doneFile = path::join(saveDir, "done"); - touchFile(doneFile.c_str()); - std::ofstream out(doneFile); - version::printVersion(out); - out.close(); - VLOG(1) << "save dir " << saveDir; - saveConfigWithPath(saveDir); -} - -void ParameterUtil::deleteParameters(int passId, int passInnerId) { - constexpr int kBufLen = 100; - char buf[kBufLen]; - const std::string &saveDir = config_->getSaveDir(); - if (passInnerId > 0) { - snprintf(buf, - kBufLen, - "%s/pass-%05d-%03d", - saveDir.c_str(), - passId, - passInnerId); - } else { - snprintf(buf, kBufLen, "%s/pass-%05d", saveDir.c_str(), passId); - } - mkDir(saveDir.c_str()); - LOG(INFO) << "delete dir " << buf; - rmDir(buf); -} - -void ParameterUtil::saveConfigWithPath(const std::string &path) { - std::string src; - // save config in some path - if (!intConfig_->config_.empty()) { - src = intConfig_->config_; - } else { - bool ok; - src = config_->getConfigName(&ok); - if (!ok) { - return; - } - } - copyFileToPath(src, path); - - // save other import config file name to path.txt - std::string ss = path::join(path, "path.txt"); - std::ofstream os(ss); - std::string fileName = path::basename(src); - CHECK(os.write(fileName.c_str(), fileName.length())) - << "Fail to write config file name " << ss; - VLOG(1) << "fileName " << fileName; - os.close(); - - // copy other import config files - for (int i = 0; i < config_->getConfig().config_files_size(); ++i) { - copyFileToPath(config_->getConfig().config_files(i), path); - } -} - -} // namespace paddle diff --git a/paddle/trainer/ParamUtil.h b/paddle/trainer/ParamUtil.h deleted file mode 100644 index 10746b4d58e3a82c081987a6aaad9e0b42272a03..0000000000000000000000000000000000000000 --- a/paddle/trainer/ParamUtil.h +++ /dev/null @@ -1,125 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include - -#include "hl_gpu.h" -#include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" - -#include -#include -#include "ParameterUpdater.h" -#include "TrainerConfig.pb.h" -#include "TrainerConfigHelper.h" - -namespace paddle { - -/** - * Configuration for parameter utils. - */ -struct ParameterUtilConfig { - DISABLE_COPY(ParameterUtilConfig); - - ParameterUtilConfig(bool save_only_one, - int saving_period, - bool load_save_parameters_in_pserver, - std::string config) - : save_only_one_(save_only_one), - saving_period_(saving_period), - load_save_param_pserver_(load_save_parameters_in_pserver), - config_(config) {} - - bool save_only_one_; - int saving_period_; - bool load_save_param_pserver_; - std::string config_; -}; - -/** - * ParameterUtil - * Utility class for loading and saving parameters - */ -class ParameterUtil { - public: - /** - * Ctor. - * - * @param config - * @param intconfig - * @param gradientMachine - * @param parameterUpdater - * @return - */ - ParameterUtil(const std::shared_ptr &config, - std::unique_ptr &&intconfig, - const GradientMachinePtr &gradientMachine, - const std::shared_ptr ¶meterUpdater); - - /// Load parameter from the saved parameter file as pass passId - /// if loadsave_parameters_in_pserver is set, some parameters MUST - /// load in pserver, which is "remote". - /// loadParameters can choose to load local/remote parameter, or both. - bool loadParameters(int passId, bool local = true, bool remote = false); - - /// load parameters given path info - void loadParametersWithPath(const std::string &dir, - bool local = true, - bool remote = false); - - /// Save parameter to dist for pass passId - /// passInnerId means saving times in one pass, some users want to - /// save parameters when have processed some batches in one pass - /// passInnerId = 0 means do not need to save in one inner pass - void saveParameters(int passId, int passInnerId = 0); - - /// save parameters for one pass, when passInnerId > 0 means saving - /// the passInnerId times in one pass - void saveParametersOnePass(int passId, int passInnerId = 0); - - /// delete parameter from disk via passId - void deleteParameters(int passId, int passInnerId = 0); - - /// save config given path info - void saveConfigWithPath(const std::string &path); - - /** - * Try to load parameter from config. - * @return true if can load from trainer config. - */ - inline bool tryLoadParametersFromConfig() { - auto &c = config_->getConfig(); - if (!c.init_model_path().empty()) { - loadParametersWithPath(c.init_model_path()); - return true; - } else if (c.start_pass() > 0) { - CHECK(loadParameters(c.start_pass() - 1)); - return true; - } else { - return false; - } - } - - private: - std::shared_ptr config_; - std::unique_ptr intConfig_; - GradientMachinePtr gserver_; - std::shared_ptr pUpdater_; -}; - -} // namespace paddle diff --git a/paddle/trainer/ParameterUpdater.cpp b/paddle/trainer/ParameterUpdater.cpp deleted file mode 100644 index 4e9e890c85945aedd7e604f52a06902191c95d4c..0000000000000000000000000000000000000000 --- a/paddle/trainer/ParameterUpdater.cpp +++ /dev/null @@ -1,152 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ParameterUpdater.h" - -#include "paddle/utils/Logging.h" - -#include "paddle/utils/Thread.h" - -namespace paddle { - -static const hl_stream_t kDeviceToHostStream = HPPL_STREAM_1; -static const hl_stream_t kHostToDeviceStream = HPPL_STREAM_2; - -SgdUpdaterWithCpuAverager::SgdUpdaterWithCpuAverager( - const OptimizationConfig& optConfig) - : SgdLocalUpdater(optConfig, false /*with averager*/) { - CHECK(FLAGS_use_gpu && optConfig.do_average_in_cpu()); - averager_.reset(AverageOptimizer::create(optConfig, - new DummyOptimizer(optConfig), - false /*sparse*/, - true /*apply*/)); - updateWorker_.addJob([]() { hl_set_device(FLAGS_gpu_id); }); -} - -void SgdUpdaterWithCpuAverager::init( - const std::vector& parameters) { - SgdLocalUpdater::init(parameters); - averager_->init(parameters_.size(), nullptr); - copyEvents_.resize(parameters_.size()); - for (auto& parameter : parameters) { - SetDevice device(parameter->getDeviceId()); - cpuParameters_.emplace_back(new Parameter(parameter->getConfig(), - /* useGpu= */ false, - /* doInit= */ false)); - if (parameter->useGpu()) { - cpuParameters_.back()->enableType(PARAMETER_APPLY); - } else { - cpuParameters_.back()->enableSharedType( - PARAMETER_APPLY, parameter->getBuf(PARAMETER_VALUE)); - } - for (ParameterType type : averager_->getParameterTypes()) { - cpuParameters_.back()->enableType(type); - } - - hl_create_event(©Events_[nonStaticParaIDMap_[parameter->getID()]]); - } -} - -SgdUpdaterWithCpuAverager::~SgdUpdaterWithCpuAverager() { - for (auto& event : copyEvents_) { - hl_destroy_event(event); - } -} - -void SgdUpdaterWithCpuAverager::updateImpl(Parameter* para) { - SgdLocalUpdater::updateImpl(para); - - if (para->useGpu()) { - size_t pid = nonStaticParaIDMap_[para->getID()]; - Parameter* cpuPara = cpuParameters_[pid].get(); - cpuPara->getBuf(PARAMETER_VALUE) - ->copyFrom(*para->getBuf(PARAMETER_VALUE), kDeviceToHostStream); - hl_stream_record_event(kDeviceToHostStream, copyEvents_[pid]); - } - - updateWorker_.addJob( - std::bind(&SgdUpdaterWithCpuAverager::updateFunc, this, para)); -} - -void SgdUpdaterWithCpuAverager::updateFunc(Parameter* para) { - SetDevice setDevice(para->getDeviceId()); - size_t pid = nonStaticParaIDMap_[para->getID()]; - Parameter* cpuPara = cpuParameters_[pid].get(); - if (para->useGpu()) { - hl_event_synchronize(copyEvents_[pid]); - } - averager_->update(cpuPara->getBufs(), cpuPara->getConfig(), -1LU); -} - -void SgdUpdaterWithCpuAverager::finishBatch(real cost) { - SgdLocalUpdater::finishBatch(cost); - - updateWorker_.wait(); - for (auto para : cpuParameters_) { - if (auto callback = averager_->needSpecialTraversal(para->getConfig())) { - callback(para->getBufs(), para->getConfig(), -1LU); - } - } - averager_->finishBatch(); -} - -void SgdUpdaterWithCpuAverager::apply() { - // backup gpu value - for (auto& para : parameters_) { - SetDevice setDevice(para->getDeviceId()); - para->getBuf(PARAMETER_GRADIENT) - ->copyFrom(*para->getBuf(PARAMETER_VALUE), kHostToDeviceStream); - } - - // apply on cpu parameter - if (auto callback = averager_->apply()) { - for (auto para : cpuParameters_) { - callback(para->getBufs(), para->getConfig(), -1LU); - } - } - - // copy to gpu value - for (auto& para : parameters_) { - SetDevice setDevice(para->getDeviceId()); - size_t pid = nonStaticParaIDMap_[para->getID()]; - Parameter* cpuPara = cpuParameters_[pid].get(); - if (parameters_[pid]->useGpu()) { - para->getBuf(PARAMETER_VALUE) - ->copyFrom(*cpuPara->getBuf(PARAMETER_APPLY), kHostToDeviceStream); - } - } - hl_stream_synchronize(kHostToDeviceStream); - for (auto& para : parameters_) { - para->setValueUpdated(); - } -} - -void SgdUpdaterWithCpuAverager::restore() { - // restore on cpu parameter - if (auto callback = averager_->restore()) { - for (auto para : cpuParameters_) { - callback(para->getBufs(), para->getConfig(), -1LU); - } - } - - // restore gpu value - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(PARAMETER_VALUE)->copyFrom(*para->getBuf(PARAMETER_GRADIENT)); - para->getBuf(PARAMETER_GRADIENT)->zeroMem(); - para->setValueUpdated(); - } -} - -} // namespace paddle diff --git a/paddle/trainer/ParameterUpdater.h b/paddle/trainer/ParameterUpdater.h deleted file mode 100644 index ef7ab92eca77bab2a8481561713f8034d2b8505d..0000000000000000000000000000000000000000 --- a/paddle/trainer/ParameterUpdater.h +++ /dev/null @@ -1,265 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Thread.h" -#include "paddle/utils/Util.h" - -#include "paddle/parameter/AverageOptimizer.h" -#include "paddle/parameter/FirstOrderOptimizer.h" -#include "paddle/parameter/OptimizerFunctions.h" -#include "paddle/parameter/OptimizerWithRegularizer.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/parameter/ParameterUpdaterBase.h" - -#include "TrainerConfig.pb.h" -#include "paddle/gserver/layers/Layer.h" - -#include -#include - -namespace paddle { - -/** - * @brief Parameter Updater for SGD, and local(not cluster) run. - */ -class SgdLocalUpdater : public ParameterUpdater { - public: - /** - * @brief Ctor. Initialize optimizer locally by optConfig. - * @param optConfig optimization config. - * @param withAverager with average optimizer or not, default is true. - */ - explicit SgdLocalUpdater(const OptimizationConfig& optConfig, - bool withAverager = true) - : numSamplesProcessed_(0) { - auto baseOptimizer = ParameterOptimizer::create(optConfig); - optimizer_.reset(withAverager - ? AverageOptimizer::create(optConfig, baseOptimizer) - : baseOptimizer); - CHECK(optimizer_) << "fail to create optimizer: " - << optConfig.learning_method(); - auto types = optimizer_->getParameterTypes(); - for (auto type : types) { - addParameterType(type); - } - } - - /** - * @brief Initialize parameters and optimizer_. - * For example, - * If optimizer need hassien vector, then parameter's hassien will - * be initialized. - * @param parameters The parameter need to be initialized. - */ - virtual void init(const std::vector& parameters) { - ParameterUpdater::init(parameters); - optimizer_->init(parameters_.size(), nullptr); - // check no L1 decay in parameter configs - CHECK(std::find_if(parameters.begin(), - parameters.end(), - [](const ParameterPtr& para) { - return para->getConfig().decay_rate_l1() > 0.0f; - }) == parameters.end()) - << "SgdLocalUpdater cannot support L1 decay in parameter"; - } - - /** - * @brief Start a batch with current mini-batch size - * @param current mini-batch size. - * @return Always PASS_TRAIN. - */ - virtual PassType startBatch(int64_t batchSize) { - numSamplesProcessed_ += batchSize; - optimizer_->startBatch(numSamplesProcessed_); - return PASS_TRAIN; - } - - /** - * @brief finish a mini-batch. - */ - virtual void finishBatch(real cost) { optimizer_->finishBatch(); } - - /** - * @brief start a pass. - */ - virtual void startPass() { optimizer_->startPass(); } - - /** - * @brief finish a pass. - * @param cost sum cost during one pass. - * @return true if accept (used for owlqn). - */ - virtual bool finishPass() { - optimizer_->finishPass(); - return ParameterUpdater::finishPass(); - } - - /** - * @brief apply model average. - */ - virtual void apply() { - if (auto callback = optimizer_->apply()) { - for (auto para : parameters_) { - SetDevice device(para->getDeviceId()); - callback(para->getBufs(), para->getConfig(), -1UL); - } - } - } - - /** - * @brief restore parameter value before model average - */ - virtual void restore() { - if (auto callback = optimizer_->restore()) { - for (auto para : parameters_) { - SetDevice device(para->getDeviceId()); - callback(para->getBufs(), para->getConfig(), -1UL); - } - } - } - - protected: - /** - * @brief update method. Update value from gradient. - * @param para parameter that will be updated. - */ - virtual void updateImpl(Parameter* para) { - optimizer_->update(para->getBufs(), para->getConfig()); - if (auto callback = optimizer_->needSpecialTraversal(para->getConfig())) { - callback(para->getBufs(), para->getConfig(), -1UL); - } - - para->setValueUpdated(); - para->getBuf(PARAMETER_GRADIENT)->zeroMem(); - } - - std::unique_ptr optimizer_; - - /** - * @brief total number of samples processed. - */ - int64_t numSamplesProcessed_; -}; - -/** - * @brief SgdCpuUpdater is used only in recursive neural network - * @deprecated - */ -class SgdCpuUpdater : public SgdLocalUpdater, public Deprecated { - public: - explicit SgdCpuUpdater(const OptimizationConfig& optConfig) - : SgdLocalUpdater(optConfig), - Deprecated( - "SgdCpuUpdater is used only in recursive neural network, " - "and recursive neural network is deprecated in paddle. " - "Use it all by your own.") {} - - /** - * @brief update all parameter on finish batch. - * @param cost - */ - virtual void finishBatch(real cost) { - for (auto para : parameters_) { - SgdLocalUpdater::update(para.get()); - } - optimizer_->finishBatch(); - } - - protected: - /** - * @brief do nothing. - * @param para - */ - virtual void updateImpl(Parameter* para) {} -}; - -/** - * @brief Sgd Local Updater With average in cpu. - * - * It will do model average in cpu to reduce gpu memory comsuption. - */ -class SgdUpdaterWithCpuAverager : public SgdLocalUpdater { - public: - /** - * @brief Ctor. - * - * SgdUpdaterWithCpuAverager will do everything as a - * SgdLocalUpdater, then copy parameter from GPU to CPU, and do model - * average in cpu. - */ - explicit SgdUpdaterWithCpuAverager(const OptimizationConfig& optConfig); - ~SgdUpdaterWithCpuAverager(); - - /** - * @brief init. Initialize cpu parameters, model average optimizer. - * @param parameters - */ - virtual void init(const std::vector& parameters); - - virtual PassType startBatch(int64_t batchSize) { - averager_->startBatch(-1UL); - return SgdLocalUpdater::startBatch(batchSize); - } - virtual void finishBatch(real cost); - - virtual void startPass() { - averager_->startPass(); - SgdLocalUpdater::startPass(); - } - virtual bool finishPass() { - averager_->finishPass(); - return SgdLocalUpdater::finishPass(); - } - - /// apply the averaged parameter to PARAMETER_VALUE - /// use PARAETER_GRADIENT for backing up PARAMETER_VALUE - virtual void apply(); - - /** - * @brief Restore parameter before apply(). - */ - virtual void restore(); - - protected: - virtual void updateImpl(Parameter* para); - - void updateFunc(Parameter* para); - - protected: - std::unique_ptr averager_; - - /** - * @brief The thread worker which do model average. - * - * For each parameter, GPU->CPU parameter is async, and do model average in - * another thread. Because the training process don't need model average while - * training, and model average only used in evaluation stage and saving stage. - * So the model average is totally async. - */ - ThreadWorker updateWorker_; - - /** - * @brief The parameter mirror in cpu. - */ - std::vector cpuParameters_; - - /** - * @brief GPU -> CPU copy event. Model average will wait after copy done. - */ - std::vector copyEvents_; -}; - -} // namespace paddle diff --git a/paddle/trainer/RemoteParameterUpdater.cpp b/paddle/trainer/RemoteParameterUpdater.cpp deleted file mode 100644 index 7314266cb24da9b9e9f0f1cbe61ed363247f51fe..0000000000000000000000000000000000000000 --- a/paddle/trainer/RemoteParameterUpdater.cpp +++ /dev/null @@ -1,843 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "RemoteParameterUpdater.h" -#include "Trainer.h" -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/Stat.h" - -DECLARE_int32(trainer_id); -DECLARE_string(save_dir); - -namespace paddle { - -static const hl_stream_t kDeviceToHostStream = HPPL_STREAM_1; -static const hl_stream_t kHostToDeviceStream = HPPL_STREAM_2; -static const int kFinishBatchPid = -1; - -const std::string RemoteParameterUpdater::kAverage = "average"; -const std::string RemoteParameterUpdater::kElasticAverage = "elastic_average"; - -RemoteParameterUpdater::RemoteParameterUpdater( - const OptimizationConfig& config, - int expectedPassCount, - std::unique_ptr&& localUpdater) - : config_(config), - localUpdater_(std::move(localUpdater)), - numBatches_(0), - passCount_(0), - expectedPassCount_(expectedPassCount), - separateSendAndRecv_(false), - isFirstPass_(true), - useApplyInPserver_(false) { - addParameterType(PARAMETER_MOMENTUM); -} - -void RemoteParameterUpdater::init(const std::vector& parameters) { - ParameterUpdater::init(parameters); - - if (localUpdater_) { - localUpdater_->init(parameters); - - for (auto& parameter : parameters) { - parameter->enableType(PARAMETER_DELTA); - } - - CHECK(config_.center_parameter_update_method() == kAverage || - config_.center_parameter_update_method() == kElasticAverage) - << "unknown center_parameter_update_method"; - - // modify delta_add_rate - CHECK_GT(FLAGS_num_gradient_servers, 1) - << "FLAGS_num_gradient_servers should be set in trainer args."; - real delta_add_rate = config_.delta_add_rate() / FLAGS_num_gradient_servers; - config_.set_delta_add_rate(delta_add_rate); - LOG(INFO) << "center parameter in pserver," - << " modify delta_add_rate=" << delta_add_rate; - } - - if (!FLAGS_use_gpu) { - cpuParameters_ = parameters; - } else { - for (auto& parameter : parameters) { - cpuParameters_.emplace_back(new Parameter(parameter->getConfig(), - /* useGpu= */ false)); - cpuParameters_.back()->setID(parameter->getID()); - if (localUpdater_) { - cpuParameters_.back()->enableType(PARAMETER_DELTA); - } - } - } - - parameterClient_.reset(new ParameterClient2(separateSendAndRecv_)); - parameterClient_->init(cpuParameters_); - parameterClient_->setTrainerId(FLAGS_trainer_id); - - if (FLAGS_trainer_id == 0) { - parameterClient_->setConfig(config_); - copyParametersFromDevice(PARAMETER_VALUE); - parameterClient_->setParameter(); - parameterClient_->setStatus(PSERVER_STATUS_PARAMETER_READY); - } else { - parameterClient_->waitForStatus(PSERVER_STATUS_PARAMETER_READY); - parameterClient_->getParameter(); - copyParametersToDevice(PARAMETER_VALUE); - } - if (FLAGS_trainer_id == 0 && - (config_.algorithm() != TrainAlgorithm::AsyncSGD)) { - startController(); - useApplyInPserver_ = useApplyInPserver(config_); - } -} - -void RemoteParameterUpdater::startController() { - controllerThread_.reset(new std::thread([this]() { this->controller(); })); -} - -void RemoteParameterUpdater::controller() { - ParameterClient2 client(false); - client.init(cpuParameters_); - while (true) { - /*start pass*/ { - client.waitPassStart(); - - PreparedOperations ops; - ops.addOperation(PSERVER_OP_START_PASS); - client.doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false, - /* releasePass= */ false); - } - - while (true) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_SGD); - client.doOperation(ops, - /* waitForGradient= */ true, - /* sendBackarameter= */ true, - /* releasePass= */ false); - if (client.isPassFinish()) { - break; - } - } - - /*finish pass*/ { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_FINISH_PASS); - client.doOperation(ops, - /* waitForGradient= */ true, - /* sendBackarameter= */ true, - /* releasePass= */ true); - } - - passCount_++; - if (passCount_ == expectedPassCount_) { - break; - } - } -} - -void RemoteParameterUpdater::copyParametersToDevice( - ParameterType parameterType) { - if (!FLAGS_use_gpu) { - return; - } - int numParameters = cpuParameters_.size(); - for (int i = 0; i < numParameters; ++i) { - parameters_[i] - ->getBuf(parameterType) - ->copyFrom(*cpuParameters_[i]->getBuf(parameterType)); - if (parameterType == PARAMETER_VALUE) { - parameters_[i]->setValueUpdated(); - } - } -} - -void RemoteParameterUpdater::copyParametersFromDevice( - ParameterType parameterType) { - if (!FLAGS_use_gpu) { - return; - } - int numParameters = cpuParameters_.size(); - for (int i = 0; i < numParameters; ++i) { - cpuParameters_[i] - ->getBuf(parameterType) - ->copyFrom(*parameters_[i]->getBuf(parameterType)); - } -} - -void RemoteParameterUpdater::updateImpl(Parameter* para) { - REGISTER_TIMER("update"); - if (localUpdater_) { - localUpdater_->update(para); - } -} - -void RemoteParameterUpdater::finishBatch(real cost) { - if (localUpdater_) { - localUpdater_->finishBatch(cost); - } - - const std::string& algorithm = config_.algorithm(); - ParameterUpdateMode mode; - if (algorithm == TrainAlgorithm::AsyncSGD) { - mode = PSERVER_UPDATE_MODE_ASYNC_SGD; - } else if (algorithm == TrainAlgorithm::SGD) { - mode = PSERVER_UPDATE_MODE_ADD_GRADIENT; - } else { - LOG(FATAL) << "Unknown algorithm: " << algorithm; - } - - ParameterType sendType; - bool sendBackParameter = true; - if (localUpdater_) { - ++numBatches_; - if (numBatches_ % config_.num_batches_per_send_parameter() != 0) { - return; - } - - if (config_.center_parameter_update_method() == kElasticAverage) { - parameterClient_->getParameter(PARAMETER_DELTA); - copyParametersToDevice(PARAMETER_DELTA); - sendBackParameter = false; // no need send back after send - - // calc delta - for (auto& para : parameters_) { - // DELTA = LOCAL_VALUE - CENTER_VALUE/*store in DELTA*/ - para->getBuf(PARAMETER_DELTA) - ->add(*para->getBuf(PARAMETER_VALUE), -1.0f, 1.0f); - - // when delta send to pserver, pserver will do: - // CENTER_VALUE += alpha * (LOCAL_VALUE - CENTER_VALUE) - } - } else { - // calc delta - for (auto& para : parameters_) { - // DELTA = NEW_VALUE - OLD_VALUE/*store in DELTA*/ - para->getBuf(PARAMETER_DELTA) - ->add(*para->getBuf(PARAMETER_VALUE), -1.0f, 1.0f); - } - } - - sendType = PARAMETER_DELTA; - - } else { - // In this case, we perform SGD on pserver. - sendType = PARAMETER_GRADIENT; - } - - copyParametersFromDevice(sendType); - - { - REGISTER_TIMER("sendAndRecv_dense"); - parameterClient_->sendAndReceiveParameter(mode, - sendType, - batchSize_, - 0, // cost = 0 - sendBackParameter); - } - - if (sendBackParameter) { - copyParametersToDevice(PARAMETER_VALUE); - } - - if (localUpdater_) { - if (config_.center_parameter_update_method() == kElasticAverage) { - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - // LOCAL_VALUE += -alpha * (LOCAL_VALUE - CENTER_VALUE) - para->getBuf(PARAMETER_VALUE) - ->add(*para->getBuf(PARAMETER_DELTA), -config_.delta_add_rate()); - } - - } else { // average - // copy value to delta - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); - } - } - } else { - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(sendType)->zeroMem(); - } - } -} - -void RemoteParameterUpdater::startPass() { - if (config_.algorithm() == TrainAlgorithm::SGD) { - parameterClient_->waitPassStart(); - } else { - // sync could benifits reducing lagged trainer for async-sgd - // even if sync could not remove all lagged trainer for the - // sake of file loading, buffer etc. - parameterClient_->asyncStartPass(); - } - - if (localUpdater_) { - localUpdater_->startPass(); - numBatches_ = 0; - - if (config_.center_parameter_update_method() == kElasticAverage) { - if (!isFirstPass_) { - // restore local value from delta - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(PARAMETER_VALUE) - ->copyFrom(*para->getBuf(PARAMETER_DELTA)); - } - } - } else { // average - // copy value to delta - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); - } - } - } -} - -bool RemoteParameterUpdater::finishPass() { - if (localUpdater_) { - localUpdater_->finishPass(); - } - - if (config_.algorithm() == TrainAlgorithm::SGD) { - parameterClient_->waitPassFinish(); - } else { - parameterClient_->asyncFinishPass(); - } - if (localUpdater_) { - if (config_.center_parameter_update_method() == kElasticAverage) { - // backup local value to delta as we will get - // the remote parameter for saving/testing - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); - } - } - } - parameterClient_->getParameter(); - copyParametersToDevice(PARAMETER_VALUE); - - isFirstPass_ = false; - return true; -} - -void RemoteParameterUpdater::apply() { - if (useApplyInPserver_) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_APPLY); - parameterClient_->doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false); - parameterClient_->getParameter( - /* recvParameterType= */ PARAMETER_VALUE, - /* sendBackParameterType= */ PARAMETER_APPLY); - copyParametersToDevice(PARAMETER_VALUE); - } -} - -void RemoteParameterUpdater::restore() { - if (useApplyInPserver_) { - parameterClient_->getParameter(); - copyParametersToDevice(PARAMETER_VALUE); - } -} - -ConcurrentRemoteParameterUpdater::ConcurrentRemoteParameterUpdater( - OptimizationConfig config, - int passCount, - std::unique_ptr&& localUpdater) - : RemoteParameterUpdater(config, passCount, std::move(localUpdater)) { - sendThread_.reset(new std::thread([this]() { this->send(); })); - recvThread_.reset(new std::thread([this]() { this->recv(); })); - - stopping_ = false; - oneBatchFinished_ = false; - separateSendAndRecv_ = true; -} - -ConcurrentRemoteParameterUpdater::~ConcurrentRemoteParameterUpdater() { - stopping_ = true; - sendQueue_.enqueue(0); - sendThread_->join(); - recvQueue_.enqueue(0); - recvThread_->join(); -} - -void ConcurrentRemoteParameterUpdater::finishBatch(real cost) { - if (localUpdater_) { - localUpdater_->finishBatch(cost); - - if (!needToUpdateRemotely()) { - ++numBatches_; - return; - } - } - - sendQueue_.enqueue(kFinishBatchPid); - - finishBatchCond_.wait([this]() { return oneBatchFinished_; }); - oneBatchFinished_ = false; - { - REGISTER_TIMER("sync_hostToDeviceStream"); - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - hl_stream_synchronize(kHostToDeviceStream); - } - } - - if (localUpdater_) { - ++numBatches_; - } -} - -// Use para=NULL to signal the end of one batch -void ConcurrentRemoteParameterUpdater::send(Parameter* para) { - const std::string& algorithm = config_.algorithm(); - ParameterUpdateMode mode; - if (algorithm == TrainAlgorithm::AsyncSGD) { - mode = PSERVER_UPDATE_MODE_ASYNC_SGD; - } else if (algorithm == TrainAlgorithm::SGD) { - mode = PSERVER_UPDATE_MODE_ADD_GRADIENT; - } else { - LOG(FATAL) << "Unknown algorithm: " << algorithm; - } - ParameterType sendType; - if (localUpdater_) { - sendType = PARAMETER_DELTA; - } else { - // In this case, we perform SGD on pserver. - sendType = PARAMETER_GRADIENT; - } - std::vector paraSegment; - if (para == NULL) { - parameterClient_->sendParameter( - mode, - sendType, - paraSegment, - batchSize_, - 0, // cost=0 - true, // sendBackParameter = true - batchStatus_); // batchStatus_ = BATCH_FINISH - - } else { - ParameterSegments paraSegTemp; - paraSegment.reserve(1); - paraSegTemp.name = para->getName(); - paraSegTemp.id = para->getID(); - paraSegment.push_back(paraSegTemp); - { - SetDevice device(para->getDeviceId()); - REGISTER_TIMER("copySingleParaFromDevice"); - copySingleParaFromDevice(para, sendType); - hl_stream_synchronize(kDeviceToHostStream); - } - parameterClient_->sendParameter(mode, - sendType, - paraSegment, - batchSize_, - 0, // cost=0 - true, // sendBackParameter = true - batchStatus_); - if (batchStatus_ == BATCH_START) batchStatus_ = BATCH_ON; - } -} -void ConcurrentRemoteParameterUpdater::recv(Parameter* para) { - parameterClient_->recvParameter(); - if (para != NULL) { - REGISTER_TIMER("copySingleParaToDevice"); - SetDevice device(para->getDeviceId()); - copySingleParaToDevice(para, PARAMETER_VALUE); - - if (localUpdater_) { - para->getBuf(PARAMETER_DELTA)->copyFrom(*para->getBuf(PARAMETER_VALUE)); - } else { - // if cpu, parameter should not changes until recvParameter(). - // if gpu, zero mem when send finish - if (!FLAGS_use_gpu) { - para->getBuf(PARAMETER_GRADIENT)->zeroMem(); - } - } - } -} - -void ConcurrentRemoteParameterUpdater::recv() { - if (FLAGS_use_gpu) hl_set_device(FLAGS_gpu_id); - StatPtr stat = getStat("recv"); - FOR_TIMING(Timer timer); - while (true) { - int pid; - { - REGISTER_TIMER("recv_dequeue"); - pid = recvQueue_.dequeue(); - } - if (pid == kFinishBatchPid) { - Parameter* para = NULL; - FOR_TIMING(timer.start()); - recv(para); - FOR_TIMING(timer.stop()); - FOR_TIMING(stat->addSample(timer.get())); - FOR_TIMING(timer.reset()); - finishBatchCond_.notify_all([this] { oneBatchFinished_ = true; }); - } else { - if (stopping_) break; - Parameter* para = parameters_[pid].get(); - FOR_TIMING(timer.start()); - recv(para); - FOR_TIMING(timer.stop()); - oneBatchFinished_ = false; - } - } -} - -void ConcurrentRemoteParameterUpdater::send() { - if (FLAGS_use_gpu) hl_set_device(FLAGS_gpu_id); - StatPtr stat = getStat("send"); - FOR_TIMING(Timer timer); - while (true) { - int pid; - { - REGISTER_TIMER("send_dequeue"); - pid = sendQueue_.dequeue(); - } - if (pid == kFinishBatchPid) { - batchStatus_ = BATCH_FINISH; - if (!localUpdater_) { - // if cpu, parameter should not changes until recvParameter(). - // if gpu, zeroMem() at the end of batch so that it won't - // interfere with computation. - if (FLAGS_use_gpu) { - REGISTER_TIMER("para_zeroMem"); - for (auto& para : parameters_) { - SetDevice device(para->getDeviceId()); - para->getBuf(PARAMETER_GRADIENT)->zeroMem(); - } - } - } - Parameter* para = NULL; - FOR_TIMING(timer.start()); - send(para); - FOR_TIMING(timer.stop()); - FOR_TIMING(stat->addSample(timer.get())); - FOR_TIMING(timer.reset()); - recvQueue_.enqueue(pid); - } else { - if (stopping_) break; - Parameter* para = parameters_[pid].get(); - if (localUpdater_) { - // DELTA = NEW_VALUE - OLD_VALUE/*store in DELTA*/ - para->getBuf(PARAMETER_DELTA) - ->add(*para->getBuf(PARAMETER_VALUE), -1.0f, 1.0f); - } - FOR_TIMING(timer.start()); - send(para); - FOR_TIMING(timer.stop()); - recvQueue_.enqueue(nonStaticParaIDMap_[para->getID()]); - } - } -} - -void ConcurrentRemoteParameterUpdater::updateImpl(Parameter* para) { - REGISTER_TIMER("update"); - if (localUpdater_) { - localUpdater_->update(para); - if (!needToUpdateRemotely()) { - return; - } - } - sendQueue_.enqueue(nonStaticParaIDMap_[para->getID()]); -} - -void ConcurrentRemoteParameterUpdater::copySingleParaToDevice( - Parameter* para, ParameterType parameterType) { - if (!FLAGS_use_gpu) { - return; - } - int i = nonStaticParaIDMap_[para->getID()]; - para->getBuf(parameterType) - ->copyFrom(*cpuParameters_[i]->getBuf(parameterType), - kHostToDeviceStream); - if (parameterType == PARAMETER_VALUE) { - para->setValueUpdated(); - } -} - -void ConcurrentRemoteParameterUpdater::copySingleParaFromDevice( - Parameter* para, ParameterType parameterType) { - if (!FLAGS_use_gpu) { - return; - } - int i = nonStaticParaIDMap_[para->getID()]; - cpuParameters_[i] - ->getBuf(parameterType) - ->copyFrom(*para->getBuf(parameterType), kDeviceToHostStream); -} - -SparseRemoteParameterUpdater::SparseRemoteParameterUpdater( - const OptimizationConfig& config, int expectedPassCount, bool testing) - : config_(config), - passCount_(0), - expectedPassCount_(expectedPassCount), - testing_(testing), - useApplyInPserver_(false) {} - -void SparseRemoteParameterUpdater::init( - const std::vector& parameters) { - ParameterUpdater::init(parameters); - - parameterClient_.reset(new ParameterClient2( - false, FLAGS_port + FLAGS_ports_num, FLAGS_ports_num_for_sparse)); - parameterClient_->init(parameters_); - parameterClient_->setTrainerId(FLAGS_trainer_id); - - if (FLAGS_trainer_id == 0) { - parameterClient_->setConfig( - config_, FLAGS_save_dir, true /*is_sparse_server*/); - if (parameters[0]->isFullSize()) { - parameterClient_->setParameter(); - } else { // init in pserver - parameterClient_->setParameterZero(); - } - } - if (FLAGS_trainer_id == 0 && !testing_ && - config_.algorithm() == TrainAlgorithm::SGD) { - startController(); - useApplyInPserver_ = useApplyInPserver(config_); - } -} - -void SparseRemoteParameterUpdater::startController() { - controllerThread_.reset(new std::thread([this]() { this->controller(); })); -} - -void SparseRemoteParameterUpdater::controller() { - ParameterClient2 client( - false, FLAGS_port + FLAGS_ports_num, FLAGS_ports_num_for_sparse); - client.init(parameters_); - - while (true) { - /*start pass*/ { - client.waitPassStart(); - - PreparedOperations ops; - ops.addOperation(PSERVER_OP_START_PASS); - client.doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false, - /* releasePass= */ false); - } - - while (true) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_SGD); - client.doOperation(ops, - /* waitForGradient= */ true, - /* sendBackarameter= */ true, - /* releasePass= */ false); - if (client.isPassFinish()) { - break; - } - } - - /*finish pass*/ { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_FINISH_PASS); - client.doOperation(ops, - /* waitForGradient= */ true, - /* sendBackarameter= */ true, - /* releasePass= */ true); - } - - passCount_++; - if (passCount_ == expectedPassCount_) { - break; - } - } -} - -PassType SparseRemoteParameterUpdater::startBatch(int64_t batchSize) { - batchSize_ = batchSize; - return PASS_TRAIN; -} - -void SparseRemoteParameterUpdater::finishBatch(real cost) { - const std::string& algorithm = config_.algorithm(); - ParameterUpdateMode mode; - if (algorithm == TrainAlgorithm::AsyncSGD) { - mode = PSERVER_UPDATE_MODE_ASYNC_SGD; - } else if (algorithm == TrainAlgorithm::SGD) { - mode = PSERVER_UPDATE_MODE_ADD_GRADIENT; - } else { - LOG(FATAL) << "Unknown algorithm: " << algorithm; - } - - ParameterType sendType = PARAMETER_GRADIENT; - - REGISTER_TIMER("sendSparseParam"); - parameterClient_->sendAndReceiveParameter(mode, - sendType, - batchSize_, - 0, // cost = 0 - false); // sendBackParameter - - // grad zero move to sgd grad machine, before merge grad sparse remote -} - -void SparseRemoteParameterUpdater::startPass() { - if (config_.algorithm() == TrainAlgorithm::SGD) { - parameterClient_->waitPassStart(); - } else { - if (FLAGS_trainer_id == 0) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_START_PASS); - parameterClient_->doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false); - } - parameterClient_->asyncStartPass(); - } -} - -bool SparseRemoteParameterUpdater::finishPass() { - if (config_.algorithm() == TrainAlgorithm::SGD) { - parameterClient_->waitPassFinish(); - } else { - if (FLAGS_trainer_id == 0) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_FINISH_PASS); - parameterClient_->doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false); - } - parameterClient_->asyncFinishPass(); - } - - return true; -} - -// Trainer will call getParametersRemote at batch start or before save, -// so we do not get values in apply() and restore(). -void SparseRemoteParameterUpdater::apply() { - if (useApplyInPserver_) { - PreparedOperations ops; - ops.addOperation(PSERVER_OP_APPLY); - parameterClient_->doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false); - } -} - -void SparseRemoteParameterUpdater::restore() {} - -void SparseRemoteParameterUpdater::getParametersRemote(bool fullSize, - bool apply) { - ParameterType sendBackParameterType = - (useApplyInPserver_ && apply) ? PARAMETER_APPLY : PARAMETER_VALUE; - std::function getParams; - std::function applyL1; - if (fullSize) { - getParams = [&] { - parameterClient_->getParameter( - /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType); - }; - applyL1 = [](Parameter& para, real decayRate) { - para.getBuf(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate); - }; - } else { - getParams = [&] { - parameterClient_->getParameterSparse( - /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType); - }; - applyL1 = [](Parameter& para, real decayRate) { - para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate); - }; - } - { - REGISTER_TIMER("getParamDenseAndSparse"); - getParams(); - if (config_.shrink_parameter_value() > 0) { - for (auto& para : parameters_) { - if (para->getConfig().decay_rate_l1() > 0) { - applyL1(*para, config_.shrink_parameter_value()); - } - } - } - } -} - -void SparseRemoteParameterUpdater::randParametersRemote() { - CHECK_EQ(FLAGS_trainer_id, 0); - - PreparedOperations ops; - ops.addOperation(PSERVER_OP_RANDOMIZE); - parameterClient_->doOperation(ops, - /* waitForGradient= */ false, - /* sendBackarameter= */ false); -} - -void SparseRemoteParameterUpdater::loadParametersRemote( - const std::string& dirName) { - if (FLAGS_trainer_id == 0) { - parameterClient_->loadValueVector(dirName); - } - - if (testing_) { - // we do not use synchronize() here, - // because test mode may run only one tester - if (FLAGS_trainer_id == 0) { - parameterClient_->setStatus(PSERVER_STATUS_PARAMETER_READY); - } else { - parameterClient_->waitForStatus(PSERVER_STATUS_PARAMETER_READY); - } - } -} - -void SparseRemoteParameterUpdater::saveParametersRemote( - const std::string& dirName) { - if (FLAGS_trainer_id == 0) { - parameterClient_->saveValueVector(dirName); - } -} - -void SparseRemoteParameterUpdaterComposite::init( - const std::vector& parameters) { - parameters_ = parameters; - - std::vector parametersArray[NUMBER_UPDATERS]; - - for (auto& para : parameters_) { - if (para->isSparseRemoteUpdate()) { - parametersArray[UPDATER_SPARSE_REMOTE].push_back(para); - } else { - parametersArray[UPDATER_NORMAL].push_back(para); - } - } - CHECK(!parametersArray[UPDATER_SPARSE_REMOTE].empty()); - CHECK(!parametersArray[UPDATER_NORMAL].empty()); - - syncThreadPool_->execPlusOwner([&](int tid, size_t numThreads) { - updaters_[tid]->init(parametersArray[tid]); - }); - - parameterTypes_ = updaters_[UPDATER_NORMAL]->getParameterTypes(); -} - -std::vector> - ParameterUpdaterCreators::constructors_; - -} // namespace paddle diff --git a/paddle/trainer/RemoteParameterUpdater.h b/paddle/trainer/RemoteParameterUpdater.h deleted file mode 100644 index 3a40a46354efd6b92278884c8f5b72504a3ff283..0000000000000000000000000000000000000000 --- a/paddle/trainer/RemoteParameterUpdater.h +++ /dev/null @@ -1,416 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include "ParameterUpdater.h" -#include "paddle/pserver/ParameterClient2.h" -#include "paddle/utils/Queue.h" -#include "paddle/utils/Util.h" - -namespace paddle { - -// TODO(yanfei): -// I think that the biggest feature of rdma is packet lossless control -// feature instead of high bandwiths, zero copy and gpu-direct rdma in -// theroy. -// But zero-copy and gpu-direct rdma features can help to reduce latency -// caused by os system. -// So, for some specified cluster, such as high density gpu cluster, -// gpu-direct and zero copy could help to improve cluster communication -// performance. -// - -/** - * Normal remote parameter updater for dense parameters. - * - * It first packs all parameters for all pservers using ParameterClient - * module, then wait for merged parameters data from all pservers. - * The synchronization pattern specified by sync-sgd or async-sgd is - * achieved by all pservers with the help of the controller within this - * remote parameter updater. - * This module indeedly bridges the gradient machines and parameter servers. - * It helps to transfer the parameters from acceleration device to cpu end - * for network. It contains additional parameters copy buffers for - * acceleration devices at cpu end, such as gpu, otherwise it will - * directly use original parameters data to update pservers. - * - * This remote parameter updater does not use pipeline mechanism to hide - * copy latency from gpu to cpu buffer. In addition the overlapped between - * backward and communication is not supported. - */ -class RemoteParameterUpdater : public ParameterUpdater { - public: - RemoteParameterUpdater( - const OptimizationConfig& config, - int expectedPassCount, - std::unique_ptr&& localUpdater = nullptr); - ~RemoteParameterUpdater() { - if (controllerThread_) { - controllerThread_->join(); - } - } - - /** - * initialize the internal parameter client and itself. - */ - virtual void init(const std::vector& parameters); - /** - * @brief start batch - * - * @note one batch training exhibits stateful feature to help - * to do performance tuning, sgd optimization if necessary. - */ - virtual PassType startBatch(int64_t batchSize) { - if (localUpdater_) { - localUpdater_->startBatch(batchSize); - } - batchSize_ = batchSize; - batchStatus_ = BATCH_START; - return PASS_TRAIN; - } - - /** - * send parameters to pservers and get returned parameters - * from all pservers if necessary. it will implictly - * cooperate with controller thread for sync-sgd. - */ - virtual void finishBatch(real cost); - virtual void startPass(); - virtual bool finishPass(); - -#ifndef PADDLE_DISABLE_TIMER - virtual void setForwardbackwardTime(uint64_t delta) { - parameterClient_->setForwardbackwardTime(delta); - } -#endif - - virtual void apply(); - virtual void restore(); - - protected: - /** - * control all pservers with all trainers for sync-sgd - */ - virtual void controller(); - - /** - * work need to do after finishBatch - */ - virtual void updateImpl(Parameter* para); - - void startController(); - - /** - * @brief copy parameters from cpu host to device, such as gpu. - * - * @note return if all data are transfered. - */ - void copyParametersToDevice(ParameterType parameterType); - - /** - * @brief copy parameters from device to cpu host - * - * @note return if all data are transfered - */ - void copyParametersFromDevice(ParameterType parameterType); - - protected: - /// Optimization config used to guide initialization and finishBatch - OptimizationConfig config_; - /// internal parameter client object for exchanging data with pserver - std::unique_ptr parameterClient_; - /// internal shadow buffer at cpu host end, use original parameters_ - /// if no acceleration devices are used. - std::vector cpuParameters_; - /// local updater for aggregating multi-batches local delta - std::unique_ptr localUpdater_; - /// the size of mini-batch - int64_t batchSize_; - /// batches passed - int64_t numBatches_; - /// for stateful control - BatchStatus batchStatus_; - /// controller thread for sync-sgd - std::unique_ptr controllerThread_; - /// passed already finished - int64_t passCount_; - /// expected passes to finished - int64_t expectedPassCount_; - /// use normal synchronization communication if True - bool separateSendAndRecv_; - /// true if it's first pass - bool isFirstPass_; - bool useApplyInPserver_; - - static const std::string kAverage; - static const std::string kElasticAverage; -}; - -// TODO(yanfei): -// do parameters level synchronization Optimization at pserver end with -// ConcurrentRemoteParameterUpdater to get more parallelization, at last -// to really hide pserver latency in backward computation. -// -/** - * This updater add additional optimization for overlapping synchronization - * from pservers with backward computation. - * - * Parameter can be sent to pservers when related backward stage is finished. - * This concurrent udpater does data copy from acceleration device to host - * memory aynchronously. In addition internal parameter client reads data in - * host memory and send them to all pservers in next stage. So this class - * help to pipeline device-to-host copy and host-to-network to hide network - * latency in backward stage. - * It contains separate send and recv thread for pipeline usage. - */ -class ConcurrentRemoteParameterUpdater : public RemoteParameterUpdater { - public: - ConcurrentRemoteParameterUpdater( - OptimizationConfig config, - int expectedPassCount, - std::unique_ptr&& localUpdater); - ~ConcurrentRemoteParameterUpdater(); - - /** - * @brief send paraemeters to all pservers - * - * @note it just signal the end signal to internal parameter client - * to finished the aynchronous send action. In addition it also - * do synchronization for all asynchronous host-to-device copy. - */ - virtual void finishBatch(real cost); - - protected: - virtual void updateImpl(Parameter* para); - /// internal thread called in send thread - void send(Parameter* para); // para == NULL indicate end of a minibatch - /// internal function called in recv thread - void recv(Parameter* para); - /** - * @brief send thread for relaying data from gradient to parameter client - * - * @note just pipe data to internal parameter client for pipeline - */ - void send(); - /** - * @brief recv thread for relaying data from internal parameter client to - * host memory - * - * @note it contains the asynchronous data copy form host to device - */ - void recv(); - /// copy specified parameter from host to device - void copySingleParaToDevice(Parameter* para, ParameterType parameterType); - /// copy specified parameter from device to host - void copySingleParaFromDevice(Parameter* para, ParameterType parameterType); - bool needToUpdateRemotely() { - return (numBatches_ + 1) % config_.num_batches_per_send_parameter() == 0; - } - - private: - /// send thread used for overlapping - std::unique_ptr sendThread_; - /// recv thread used for overlapping - std::unique_ptr recvThread_; - /// buffer queue for overlapping - Queue sendQueue_; - /// buffer queue for overlapping - Queue recvQueue_; - /// flags indicating to stop - bool stopping_; - /// conditional variable for threads synchronization between the - /// thread calling finishBatch and internal recv thread - LockedCondition finishBatchCond_; - bool oneBatchFinished_; -}; - -// TODO(yanfei): -// merge sparse updater with dense updater, and could help to reduce -// the synchronization between sparse and dense udpater. it could also -// reduce the threads for managing all connections. -/** - * This class is specified for updating sparse parameters. - * - * It allows part of parameter to be exchanged with all pservers. - * If sparse input assigned, part gradients of first hidden layer - * could remained zero which can not need to be exchanged within - * all pservers. This is the key optimization point for this updater - * - * For updating sparse parameters, all latest parameters are stored - * in pservers instead of keeping full copy at train end, so need to - * prefetch parameters weight value which can be changed in next-batch - * before doing next forwardbackward. Also, with above fact that the - * parameters can be stored in pserver instead of trainer, we can - * fetch specified parmeters if necessary, and can support huge - * parameters which is larger enough than the RAM size in single - * node. - * - * Internally, this updater will direct internal parameter client - * to encapsulate sparse specified message for all pservers. - */ -class SparseRemoteParameterUpdater : public ParameterUpdater { - public: - SparseRemoteParameterUpdater(const OptimizationConfig& config, - int expectedPassCount, - bool testing); - ~SparseRemoteParameterUpdater() { - if (controllerThread_) { - controllerThread_->join(); - } - } - - /// initialization - virtual void init(const std::vector& parameters); - - /// stateful batch control - virtual PassType startBatch(int64_t batchSize); - /// send all sparse related parameters to all pservers - virtual void finishBatch(real cost); - virtual void startPass(); - virtual bool finishPass(); - - virtual void apply(); - virtual void restore(); - - /// load parameters from pservers - virtual void loadParametersRemote(const std::string& dirName); - /// save parameters to pservers - virtual void saveParametersRemote(const std::string& dirName); - /** - * @brief get latest sparse parameters value from all pservers - * - * @note call it before next mini-batch - */ - virtual void getParametersRemote(bool fullSize, bool apply); - virtual void randParametersRemote(); -#ifndef PADDLE_DISABLE_TIMER - virtual void setForwardbackwardTime(uint64_t delta) { - parameterClient_->setForwardbackwardTime(delta); - } -#endif - - protected: - /// update implimentation, not implemented - virtual void updateImpl(Parameter* para) {} - - /// internal controller routine for controller thread - virtual void controller(); - - /// start controller thread - void startController(); - - protected: - /// optimization config - OptimizationConfig config_; - /// internal parameter client - std::unique_ptr parameterClient_; - int64_t batchSize_; - std::unique_ptr controllerThread_; - int64_t passCount_; - int64_t expectedPassCount_; - bool testing_; - bool useApplyInPserver_; -}; - -/** - * Class for supporting normal updater and sparse updater - * - * Not all parts of one model are sparse, so it exists dense updater - * for normal layers while sparse updater is for sparse layers. - * - * it directly call internal dense and sparse udpater individually. - */ -class SparseRemoteParameterUpdaterComposite : public ParameterUpdaterComposite { - public: - enum { - UPDATER_SPARSE_REMOTE = 0, // execute in sync thread pool(tid:0) - UPDATER_NORMAL = 1, // execute in Owner thread(tid:1) - NUMBER_UPDATERS = 2, - }; - /** - * @brief create one dense updater and one sparse updater - * - * @note use syncThreadPool to synchronize these two updaters - */ - SparseRemoteParameterUpdaterComposite( - const OptimizationConfig& config, - int expectedPassCount, - bool testing, - std::unique_ptr&& normalUpdater) { - updaters_.resize(NUMBER_UPDATERS); - updaters_[UPDATER_SPARSE_REMOTE].reset( - new SparseRemoteParameterUpdater(config, expectedPassCount, testing)); - updaters_[UPDATER_NORMAL] = std::move(normalUpdater); - - syncThreadPool_.reset(new SyncThreadPool(NUMBER_UPDATERS - 1)); - } - - /// initialization of dense and sparse updaters - virtual void init(const std::vector& parameters); -}; - -class ParameterUpdaterCreators { - public: - /** - * @brief add a creator to create custom ParameterUpdater while training. - * The creator is a function with type (alogrithm, optConfig, isLocal, - * numPasses) -> ParameterUpdater*. Trainer will use this - * ParameterUpdater if creator can create a no nullptr - * ParameterUpdater. Return nullptr will use trainer's default - * updaters. - * - * @param creator method which can create ParameterUpdater. - */ - static void addCreator( - const std::function& creator) { // NOLINT explicit move closing ) in this line - // for readability - constructors_.push_back(creator); - } - - /** - * @brief Try to create an updater by given algo, optConfig, isLocal, - * numPasses. Return nullptr if cannot create anyone. - * @param algo algorithm string. - * @param optConfig optimization config. - * @param isLocal is in local mode or not. - * @param numPasses total passes that trainer will train. - * @return nullptr if fail, not nullptr if we can create an updater. - */ - static ParameterUpdater* tryCreateUpdater(const std::string& algo, - const OptimizationConfig& optConfig, - bool isLocal, - size_t numPasses) { - for (auto& c : constructors_) { - if (auto updater = c(algo, optConfig, isLocal, numPasses)) { - return updater; - } - } - return nullptr; - } - - private: - static std::vector> - constructors_; -}; - -} // namespace paddle diff --git a/paddle/trainer/Tester.cpp b/paddle/trainer/Tester.cpp deleted file mode 100644 index 16e676d60248dfe6d443c50fbf34970e63c1f412..0000000000000000000000000000000000000000 --- a/paddle/trainer/Tester.cpp +++ /dev/null @@ -1,380 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Tester.h" - -#include -#include - -#include -#include -#include -#include - -#include - -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -#include "TesterConfig.h" -#include "paddle/gserver/gradientmachines/GradientMachineMode.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/gserver/layers/ValidationLayer.h" - -namespace paddle { - -Tester::Tester(const std::shared_ptr& config, - std::unique_ptr&& intconfig, - const GradientMachinePtr& gradientMachine, - const std::shared_ptr& parameterUpdater, - std::shared_ptr testDataProvider) - : config_(config), - intconfig_(std::move(intconfig)), - gradientMachine_(gradientMachine), - parameterUpdater_(parameterUpdater), - testDataProvider_(testDataProvider) { - if (config_->getOptConfig().use_sparse_remote_updater()) { - LOG(FATAL) << "It's prohibited to set sparse_remote_update " - << "when doing train and test jobs in the same " - << "process. You could run paddle --job=test in " - << "a separate process."; - } - testEvaluator_.reset(gradientMachine_->makeEvaluator()); - if (intconfig_->distributeTest) { - testParameterClient_.reset(new ParameterClient2(true)); - } - - if (testParameterClient_) { - testParameterClient_->init(gradientMachine_->getParameters()); - } - - std::unique_ptr paramConfig( - new ParameterUtilConfig(intconfig_->saveOnlyOne, - intconfig_->savingPeriod, - intconfig_->loadsaveParametersInPserver, - intconfig_->config)); - - paramUtil_.reset(new ParameterUtil( - config_, std::move(paramConfig), gradientMachine_, parameterUpdater_)); -} - -void Tester::startTestPeriod() { - if (testDataProvider_) { - testDataProvider_->reset(); - } - testEvaluator_->start(); - testContext_.cost = 0; - testContext_.numSamples = 0; - - parameterUpdater_->apply(); - if (intconfig_->prevBatchState) { - gradientMachine_->getState(*intconfig_->trainState); - gradientMachine_->setState(*intconfig_->testState); - } -} - -void Tester::testOneDataBatch(const DataBatch& dataBatch, - std::vector* outArgs) { - testContext_.cost += - forwardOneBatch(dataBatch, testEvaluator_.get(), outArgs); - testContext_.numSamples += dataBatch.getSize(); -} - -void Tester::testOnePeriod() { - DataBatch dataBatch; - int64_t batchSize = config_->getOptConfig().batch_size(); - std::vector outArgs; - startTestPeriod(); - while (testDataProvider_->getNextBatch(batchSize, &dataBatch) != 0) { - testOneDataBatch(dataBatch, &outArgs); - } - finishTestPeriod(); -} - -void Tester::finishTestPeriod() { - if (intconfig_->prevBatchState) { - gradientMachine_->resetState(); - } - testEvaluator_->finish(); - CHECK_GT(testContext_.numSamples, 0) - << "There is no samples in your test batch. Possibly " - "wrong implementation of DataProvidor.reset()"; - LOG(INFO) << " Test samples=" << testContext_.numSamples - << " cost=" << testContext_.cost / testContext_.numSamples - << " Eval: " << *testEvaluator_; - parameterUpdater_->restore(); - if (intconfig_->prevBatchState) { - gradientMachine_->getState(*intconfig_->testState); - gradientMachine_->setState(*intconfig_->trainState); - } -} - -int64_t Tester::testOneBatchById(int64_t batchId) { - DataBatch dataBatch; - int32_t batchSize = config_->getOptConfig().batch_size(); - - testDataProvider_->getNextBatch(batchSize, &dataBatch); - - int64_t actualBatchSize = dataBatch.getSize(); - if (actualBatchSize == 0) { - return 0; - } - - std::vector outArgs; - - stats_ += std::pair{ - actualBatchSize, - forwardOneBatch(dataBatch, testEvaluator_.get(), &outArgs)}; - - if (((batchId + 1) % intconfig_->logPeriod) == 0) { - LOG(INFO) << " Batch=" << batchId + 1 << " " << stats_.getStats(false); - } - - return actualBatchSize; -} - -real Tester::forwardOneBatch(const DataBatch& dataBatch, - Evaluator* evaluator, - std::vector* pOutArgs) { - auto& outArgs = *pOutArgs; - const std::vector& inArgs = dataBatch.getStreams(); - if (intconfig_->loadsaveParametersInPserver) { - REGISTER_TIMER("prefetch"); - gradientMachine_->prefetch(inArgs); - parameterUpdater_->getParametersRemote(false /*full parameter*/, - true /*after apply*/); - } - - gradientMachine_->forward(inArgs, &outArgs, PASS_TEST); - - // write features if set this flag and outArgs is not empty - std::string featFile = intconfig_->featFile; - if (!featFile.empty() && outArgs.empty()) { - size_t numOutputs = outArgs.size(); - std::vector featMatrices; - featMatrices.resize(numOutputs); - for (size_t i = 0; i < numOutputs; ++i) { - featMatrices[i] = Matrix::create(outArgs[i].value->getHeight(), - outArgs[i].value->getWidth(), - false, - false); // CPU data buffer - featMatrices[i]->copyFrom(*(outArgs[i].value), HPPL_STREAM_DEFAULT); - } - hl_stream_synchronize(HPPL_STREAM_DEFAULT); - FILE* fp = fopen(featFile.c_str(), "ab+"); - CHECK(!ferror(fp)) << "Fail to open " << featFile; - - size_t sampleNum = featMatrices[0]->getHeight(); - for (size_t i = 0; i < sampleNum; ++i) { - for (size_t j = 0; j < numOutputs; ++j) { - size_t dim = featMatrices[j]->getWidth(); - fwrite(featMatrices[j]->getData() + i * dim, sizeof(real), dim, fp); - } - } - fclose(fp); - } - if (evaluator) { - gradientMachine_->eval(evaluator); - } - - // Save the output layers if predict_output_dir is not empty - std::string predictOutputDir = intconfig_->predictOutputDir; - if (!predictOutputDir.empty() && !outArgs.empty()) { - CHECK(intconfig_->testing) << "Only valid in test mode"; - if (!os_.is_open()) { - // TODO(yuyang18): Refactor these lines. - constexpr int kBufLen = 100; - char buf[kBufLen]; - snprintf(buf, kBufLen, "rank-%05d", intconfig_->trainerId); - mkDir(predictOutputDir.c_str()); - std::string filename = path::join(predictOutputDir, buf); - os_.open(filename, std::ofstream::trunc); - CHECK(os_.is_open()) << "Failed to open file " << filename; - } - printOutput(outArgs, os_); - return 0.0; // In this case, there is no meaning to calculate cost - } - - return Argument::sum(outArgs); -} - -void Tester::testOnePassBatch(int passId) { - stats_.reset(); - const std::vector inArgs; - gradientMachine_->forward(inArgs, nullptr, PASS_TEST); - int64_t num; - real cost; - gradientMachine_->getStats(cost, num); - stats_ += std::pair{num, cost}; - gradientMachine_->onPassEnd(); - - LOG(INFO) << " Pass=" << passId << " " << stats_.getStats(false); -} - -void Tester::testOnePass(int passId) { - stats_.reset(); - int64_t batchId = 0; - int num = 0; - if (intconfig_->prevBatchState) { - gradientMachine_->resetState(); - } - - testEvaluator_->start(); - - do { - num = testOneBatchById(batchId); - ++batchId; - } while (num > 0); - - gradientMachine_->onPassEnd(); - testEvaluator_->finish(); - - LOG(INFO) << " Pass=" << passId << " " << stats_.getStats(false) - << " Eval: " << *testEvaluator_; - - if (intconfig_->distributeTest) { - testEvaluator_->distributeEval(testParameterClient_.get()); - if (0 == intconfig_->trainerId) { - LOG(INFO) << "distribute eval: " << *testEvaluator_; - } - } -} - -void Tester::test() { - CHECK(testDataProvider_) << "TestData is not specified"; - testDataProvider_->setSkipShuffle(); - testDataProvider_->reset(); - gradientMachine_->start(); - - // For evaluation - std::vector modelList; - std::string modelListFromConfig = intconfig_->modelList; - std::string initModelPath = intconfig_->initModelPath; - if (!modelListFromConfig.empty()) { - loadFileList(modelListFromConfig, modelList); - intconfig_->testPass = 0; - intconfig_->numPasses = modelList.size(); - intconfig_->savingPeriod = 1; - CHECK_EQ(intconfig_->testWait, 0) << "--test_wait must be 0 for evaluation"; - } else if (!initModelPath.empty()) { - modelList.push_back(initModelPath); - intconfig_->testPass = 0; - intconfig_->numPasses = 1; - intconfig_->savingPeriod = 1; - CHECK_EQ(intconfig_->testWait, 0) << "--test_wait must be 0 for evaluation"; - } - - for (int i = intconfig_->testPass; i < intconfig_->numPasses; ++i) { - int passId = i; - if (passId % intconfig_->savingPeriod == 0) { - if (intconfig_->testWait) { - while (paramUtil_->loadParameters( - passId, true /*local*/, true /*remote*/) == false) { - LOG(INFO) << "Waiting for parameters of pass " << passId; - sleep(60); // sleep 60s - } - } else { - if (modelList.size() == 0) { - CHECK_EQ(paramUtil_->loadParameters( - passId, true /*local*/, true /*remote*/), - true); - } else { - paramUtil_->loadParametersWithPath( - modelList[i], true /*local*/, true /*remote*/); - } - } - if (IGradientMachineMode::trainWholeDataInOneBatch(intconfig_->mode)) { - testOnePassBatch(passId); - } else { - testOnePass(passId); - } - if (passId + intconfig_->savingPeriod < intconfig_->numPasses) { - // if there is at least 1 more pass to test, then call reset, - // otherwise not. - testDataProvider_->reset(); - } - } - } - - gradientMachine_->finish(); -} - -void Tester::printOutput(const std::vector& outArgs, - std::ostream& os) { - size_t numOutputs = outArgs.size(); - size_t numIns = outArgs[0].getBatchSize(); - if (cpuMat_.size() != numOutputs || cpuVec_.size() != numOutputs) { - cpuMat_.resize(numOutputs, nullptr); - cpuVec_.resize(numOutputs, nullptr); - } - - for (size_t i = 0; i < numOutputs; ++i) { - if (outArgs[i].value != nullptr) { - if (outArgs[i].value->useGpu()) { - if (dynamic_cast(outArgs[i].value.get())) { - size_t dim = outArgs[i].value->getWidth(); - Matrix::resizeOrCreate(cpuMat_[i], numIns, dim, false, false); - cpuMat_[i]->copyFrom(*outArgs[i].value); - } else if (dynamic_cast(outArgs[i].value.get())) { - auto sparseMat = - dynamic_cast(outArgs[i].value.get()); - cpuMat_[i] = Matrix::createSparseMatrix(sparseMat->getHeight(), - sparseMat->getWidth(), - sparseMat->getElementCnt(), - sparseMat->getValueType(), - sparseMat->format_, - false, /* trans */ - false); /* useGpu */ - hl_stream_t stream = HPPL_STREAM_DEFAULT; - cpuMat_[i]->copyFrom(*sparseMat, stream); - } else { - LOG(WARNING) << "Not supported gpu matrix type"; - } - } - } else if (outArgs[i].ids != nullptr) { - if (outArgs[i].ids->useGpu()) { - IVector::resizeOrCreate(cpuVec_[i], outArgs[i].ids->getSize(), false); - cpuVec_[i]->copyFrom(*outArgs[i].ids); - } - } else if (outArgs[i].strs != nullptr) { - continue; - } else { - LOG(WARNING) << "outArgs[" << i << "] has no data to print"; - } - } - - for (size_t i = 0; i < numIns; ++i) { - for (size_t j = 0; j < numOutputs; ++j) { - if (outArgs[j].value != nullptr) { - if (outArgs[j].value->useGpu()) { - cpuMat_[j]->printOneRow(os, i); - } else { - outArgs[j].value->printOneRow(os, i); - } - } else if (outArgs[j].ids != nullptr) { - if (outArgs[j].ids->useGpu()) { - cpuVec_[j]->printOneElement(os, i); - } else { - outArgs[j].ids->printOneElement(os, i); - } - } else if (outArgs[j].strs != nullptr) { - os << (*outArgs[j].strs)[i] << ";"; - } - } - os << std::endl; - } -} -} // namespace paddle diff --git a/paddle/trainer/Tester.h b/paddle/trainer/Tester.h deleted file mode 100644 index 801c77e3116369732bf4b03107adce6a71dc2184..0000000000000000000000000000000000000000 --- a/paddle/trainer/Tester.h +++ /dev/null @@ -1,149 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include - -#include "hl_gpu.h" -#include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" - -#include "TrainerConfig.pb.h" - -#include -#include -#include "ParamUtil.h" -#include "ParameterUpdater.h" -#include "TesterConfig.h" -#include "TrainerInternalConfig.h" - -namespace paddle { - -/** - * Neural Network test logics code. - * It is a private class for Trainer. - */ -class Tester { - public: - /** - * Ctor - * @param config Trainer Config. - * @param intconfig Tester Config. - * @param gradientMachine Gradient machine(neuralnetwork) that will be tested. - * @param parameterUpdater Parameter Updater. Not for updating parameter, just - * for getting parameter from parameter-server. - * @param testDataProvider Test data provider. - */ - Tester(const std::shared_ptr& config, - std::unique_ptr&& intconfig, - const GradientMachinePtr& gradientMachine, - const std::shared_ptr& parameterUpdater, - std::shared_ptr testDataProvider); - - /** - * test one period. - * - * One period means 2 things. - * if test_period !=0 and not test_all_data_in_one_period, then - * will test test_period * batch_size data. - * else - * will test whole test data. - * - * It is convenience to test small set of data when test data set is large and - * is training at same time. - */ - void testOnePeriod(); - void startTestPeriod(); - void finishTestPeriod(); - void testOneDataBatch(const DataBatch& dataBatch, - std::vector* outArgs); - - /** - * Test for given data batch. - * @param dataBatch Data batch. - * @param evaluator Evaluator - * @return cost - */ - real forwardOneBatch(const DataBatch& dataBatch, - Evaluator* evaluator, - std::vector* outArgs); - - /** - * performance the full pass of test given test data provider - */ - void test(); - - protected: - std::shared_ptr testParameterClient_; - std::shared_ptr config_; - std::unique_ptr intconfig_; - GradientMachinePtr gradientMachine_; - std::shared_ptr parameterUpdater_; - std::unique_ptr testEvaluator_; - std::unique_ptr paramUtil_; - DataProviderPtr testDataProvider_; - TrainerStats stats_; - - // Used for saving the values of output layers - std::ofstream os_; - std::vector cpuMat_; - std::vector cpuVec_; - struct { - int64_t numSamples; - real cost; - } testContext_; - - private: - /** - * Test one batch by batchId. It is only used for testOnePass. - * - * Durning testOnePass, each log_period will print cost statistics. - * - * @param batchId current batch id (from 0) - * @return num of tested samples. Zero if end of pass. - */ - int64_t testOneBatchById(int64_t batchId); - - /** - * Test whole pass in one batch. - * - * - * @param passId current pass id (from 0) - */ - void testOnePassBatch(int passId); - - /** - * test for one pass in several mini-batches. - * - * Used for sgd method. - * - * @param passId current pass id (from 0) - */ - void testOnePass(int passId); - - /** - * print the outArgs to a stream - * - * used for save feature file - * - * @param [in] outArgs output arguments for network. - * @param [in,out] os output stream. - */ - void printOutput(const std::vector& outArgs, std::ostream& os); -}; - -} // namespace paddle diff --git a/paddle/trainer/TesterConfig.h b/paddle/trainer/TesterConfig.h deleted file mode 100644 index 68d4c931ff2df8e24acaa9fe6b35bfd613197c72..0000000000000000000000000000000000000000 --- a/paddle/trainer/TesterConfig.h +++ /dev/null @@ -1,138 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include - -#include "hl_gpu.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" - -#include "TrainerConfig.pb.h" - -#include -#include -#include "ParameterUpdater.h" - -namespace paddle { - -/** - * TesterConfig - * general configs for training - */ -struct TesterConfig { - /** - * indicate test period - */ - int testPeriod; - - /** - * indicate whether to save previous batch state - */ - bool prevBatchState; - - /** - * log period - */ - int logPeriod; - - /** - * loadsave parameters in pserver - */ - bool loadsaveParametersInPserver; - - /** - * feat file - */ - std::string featFile; - - /** - * predict output dir - */ - std::string predictOutputDir; - - /** - * trianer id - */ - int trainerId; - - /** - * distribute test - */ - bool distributeTest; - - /** - * training state - */ - MachineState* trainState; - - /** - * test state - */ - MachineState* testState; - - /** - * model list - */ - std::string modelList; - - /** - * test passes - */ - int testPass; - - /** - * num passes - */ - int numPasses; - - /** - * saving period - */ - int savingPeriod; - - /** - * test wait - */ - int testWait; - - /** - * init model path - */ - std::string initModelPath; - - /** - * save only one - */ - bool saveOnlyOne; - - /** - * testing mode - */ - bool testing; - - /** - * mode - */ - int mode; - - /** - * config loc - */ - std::string config; -}; - -} // namespace paddle diff --git a/paddle/trainer/ThreadParameterUpdater.cpp b/paddle/trainer/ThreadParameterUpdater.cpp deleted file mode 100644 index 3c85c3aaac68fc29da90c24d1208887a17009d5f..0000000000000000000000000000000000000000 --- a/paddle/trainer/ThreadParameterUpdater.cpp +++ /dev/null @@ -1,309 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "ThreadParameterUpdater.h" - -#include "paddle/utils/Logging.h" - -#include "paddle/math/SparseRowMatrix.h" -#include "paddle/parameter/ThreadLocalBuffer.h" -#include "paddle/utils/Thread.h" - -DECLARE_int32(trainer_count); - -namespace paddle { - -SgdThreadUpdater::SgdThreadUpdater(const OptimizationConfig& optConfig) - : config_(optConfig), numSamplesProcessed_(0) { - // fill types - auto types = sgdOptimizerGetTypes(optConfig, false /*inPserver*/); - for (auto type : types) { - addParameterType(type); - } -} - -void SgdThreadUpdater::init(const std::vector& parameters) { - ParameterUpdater::init(parameters); - - // calc max parameter id - size_t maxId = 0; - for (auto& para : parameters_) { - maxId = std::max(maxId, para->getID()); - } - - optimizers_.resize(maxId + 1); - for (auto& para : parameters_) { - int pid = para->getID(); - optimizers_[pid].reset(sgdOptimizerCreate(config_, - para->getConfig(), - para->isGradSparseUpdate(), - false /*inPserver*/)); - size_t numRows = para->isGradSparseUpdate() ? para->getConfig().dims(0) : 0; - optimizers_[pid]->init(numRows, ¶->getConfig()); - if (para->isGradSparseUpdate() && FLAGS_trainer_count == 1) { - // For trainer_count=1, the gradient machine is NeuralNetwork, which does - // not create parameter buf for PARAMETER_GRADIENT for sparse update in - // Parameter::enableType(). But gradient parameter buf is still used - // in SgdThreadUpdater. We need to explicitly create it. - // - // The AverageOptimizer::restore/apply method will use PARAMETER_GRADIENT - // as a temp buffer. - para->enableBufType(PARAMETER_GRADIENT); - } - } -} - -void SgdThreadUpdater::startPass() { - for (auto& para : parameters_) { - int pid = para->getID(); - optimizers_[pid]->startPass(); - } -} - -bool SgdThreadUpdater::finishPass() { - catchUpWith(); - - for (auto& para : parameters_) { - int pid = para->getID(); - optimizers_[pid]->finishPass(); - } - return true; -} - -void SgdThreadUpdater::updateImpl(Parameter* para) { - if (!para->useGpu()) return; - SetDevice setDevice(para->getDeviceId()); - ParameterOptimizer* optimizer = optimizers_[para->getID()].get(); - optimizer->update(para->getBufs(), para->getConfig()); - if (auto callback = optimizer->needSpecialTraversal(para->getConfig())) { - callback(para->getBufs(), para->getConfig(), -1LU); - } - - para->setValueUpdated(); - para->clearGradient(); -} - -void SgdThreadUpdater::threadTraverse( - const ParameterOptimizer::TraverseCallback& callback, - int tid, - size_t numThreads, - Parameter* para) { - VectorPtr* vecs = parameter::getThreadLocalBuffer(); - if (para->isGradSparseUpdate()) { - size_t height = para->getConfig().dims(0); - size_t width = para->getConfig().dims(1); - for (size_t i = tid; i < height; i += numThreads) { - // setup sub bufs - for (auto type : parameterTypes_) { - vecs[type]->subVecFrom(*para->getBuf(type), i * width, width); - } - callback(vecs, para->getConfig(), i); - } - } else { // dense - // setup sub bufs - auto interval = calcSplitArrayInterval( - para->getSize(), (size_t)tid, numThreads, 8LU /*for avx*/); - for (auto type : parameterTypes_) { - vecs[type]->subVecFrom(*para->getBuf(type), interval); - } - - callback(vecs, para->getConfig(), -1LU); - } -} - -void SgdThreadUpdater::traverse(GetTraverseCallback getTraverseCallback) { - bool hasCpuPara = false; - bool hasGpuPara = false; - for (auto& para : parameters_) { - if (para->useGpu()) { - hasGpuPara = true; - } else { - hasCpuPara = true; - } - } - - auto cpuTraverse = [&](int tid, size_t numThreads) { - for (auto& para : parameters_) { - if (auto callback = getTraverseCallback(para.get())) { - threadTraverse(callback, tid, numThreads, para.get()); - } - } - }; - auto gpuTraverse = [&](int tid, size_t numThreads) { - for (auto& para : parameters_) { - if (para->useGpu()) { - if (auto callback = getTraverseCallback(para.get())) { - SetDevice setDevice(para->getDeviceId()); - callback(para->getBufs(), para->getConfig(), -1LU); - } - } - } - }; - - if (hasCpuPara && hasGpuPara) { - getGlobalSyncThreadPool()->exec(cpuTraverse, gpuTraverse); - } else if (hasCpuPara) { - getGlobalSyncThreadPool()->exec(cpuTraverse); - } else if (hasGpuPara) { - gpuTraverse(0, 0); - } -} - -void SgdThreadUpdater::catchUpWith() { - traverse([this](Parameter* para) { - return optimizers_[para->getID()]->startCatchUpWith(); - }); - - for (auto& para : parameters_) { - int pid = para->getID(); - optimizers_[pid]->finishCatchUpWith(); - } -} - -void SgdThreadUpdater::apply() { - catchUpWith(); - - traverse( - [this](Parameter* para) { return optimizers_[para->getID()]->apply(); }); -} - -void SgdThreadUpdater::restore() { - traverse([this](Parameter* para) { - return optimizers_[para->getID()]->restore(); - }); -} - -PassType SgdThreadUpdater::startBatch(int64_t batchSize) { - numSamplesProcessed_ += batchSize; - for (auto& para : parameters_) { - int pid = para->getID(); - optimizers_[pid]->startBatch(numSamplesProcessed_); - } - return PASS_TRAIN; -} - -void SgdThreadUpdater::finishBatch(real cost) { - getGlobalSyncThreadPool()->exec([&](int tid, size_t numThreads) { - for (auto& para : parameters_) { - if (para->isGradSparseUpdate()) { - threadUpdateSparse(tid, numThreads, para.get()); - } else if (!para->useGpu()) { - threadUpdateDense(tid, numThreads, para.get()); - } - } - }); - - for (auto& para : parameters_) { - int pid = para->getID(); - optimizers_[pid]->finishBatch(); - } -} - -void SgdThreadUpdater::threadUpdateSparse(int tid, - size_t numThreads, - Parameter* para) { - int pid = para->getID(); - ParameterOptimizer* optimizer = optimizers_[pid].get(); - VectorPtr* vecs = parameter::getThreadLocalBuffer(); - - size_t height = para->getConfig().dims(0); - size_t width = para->getConfig().dims(1); - - if (dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get())) { - // From MultiGradientMachine - SparseRowIdsCpuMatrix* mainMat = dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get()); - std::vector& sparseIds = mainMat->getIds(tid); - - for (auto id : sparseIds) { - // setup sub bufs - for (auto type : parameterTypes_) { - vecs[type]->subVecFrom(*para->getBuf(type), id * width, width); - } - optimizer->update(vecs, para->getConfig(), id); - vecs[PARAMETER_GRADIENT]->zeroMem(); - } - sparseIds.clear(); - } else if (dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get())) { - // From NeuralNetwork - SparseRowCpuMatrix* mainMat = dynamic_cast( - para->getMat(PARAMETER_GRADIENT).get()); - - std::vector& localIndices = - mainMat->getIndexDictHandle()->localIndices; - - auto interval = - calcSplitArrayInterval(localIndices.size(), tid, numThreads); - for (size_t i = interval.first; i < interval.second; ++i) { - auto id = localIndices[i]; - real* row = mainMat->getLocalRow(i); - // setup sub bufs - for (auto type : parameterTypes_) { - if (type == PARAMETER_GRADIENT) { - vecs[type]->subVecFrom(row, 0, width); - } else { - vecs[type]->subVecFrom(*para->getBuf(type), id * width, width); - } - } - optimizer->update(vecs, para->getConfig(), id); - vecs[PARAMETER_GRADIENT]->zeroMem(); - } - // For numThreads > 1, MultiGradientMachine is used, which goes - // to the above branch. - CHECK_EQ(numThreads, 1UL); - mainMat->clearIndices(); - } else { - auto& m = *para->getMat(PARAMETER_GRADIENT).get(); - LOG(FATAL) << "Internal error: " << para->getName() << " " - << typeid(m).name(); - } - - if (auto callback = optimizer->needSpecialTraversal(para->getConfig())) { - for (size_t i = tid; i < height; i += numThreads) { - // setup sub bufs - for (auto type : parameterTypes_) { - vecs[type]->subVecFrom(*para->getBuf(type), i * width, width); - } - callback(vecs, para->getConfig(), i); - } - } -} - -void SgdThreadUpdater::threadUpdateDense(int tid, - size_t numThreads, - Parameter* para) { - int pid = para->getID(); - ParameterOptimizer* optimizer = optimizers_[pid].get(); - VectorPtr* vecs = parameter::getThreadLocalBuffer(); - - auto interval = calcSplitArrayInterval( - para->getSize(), (size_t)tid, numThreads, 8LU /*for avx*/); - - // setup sub bufs - for (auto type : parameterTypes_) { - vecs[type]->subVecFrom(*para->getBuf(type), interval); - } - - // update - optimizer->update(vecs, para->getConfig()); - vecs[PARAMETER_GRADIENT]->zeroMem(); - - if (auto callback = optimizer->needSpecialTraversal(para->getConfig())) { - callback(vecs, para->getConfig(), -1LU); - } -} - -} // namespace paddle diff --git a/paddle/trainer/ThreadParameterUpdater.h b/paddle/trainer/ThreadParameterUpdater.h deleted file mode 100644 index b5e6a7ce3c8457364b10c921bca3386fbb6f6cbf..0000000000000000000000000000000000000000 --- a/paddle/trainer/ThreadParameterUpdater.h +++ /dev/null @@ -1,85 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/parameter/AverageOptimizer.h" -#include "paddle/parameter/FirstOrderOptimizer.h" -#include "paddle/parameter/OptimizerFunctions.h" -#include "paddle/parameter/OptimizerWithRegularizer.h" -#include "paddle/parameter/Parameter.h" -#include "paddle/parameter/Regularizer.h" -#include "paddle/utils/Util.h" - -#include -#include - -namespace paddle { - -/** - * \brief A parameter updater that uses multiple threads to update parameters. - This parameter updater handles GPU and CPU updates differently, - because at the current moment, the merging on CPU is happening on the - main thread, and the its parameter size can be much larger than the one GPU. - Thus, for GPU, the parameter updates happens in updateImpl() function, which - is called by gradient machines as a callback function supplied to backward() - and forwardBackward(). - For CPU, the parameter updates happens in separate threads maintained by this - class. - */ -class SgdThreadUpdater : public ParameterUpdater { - public: - explicit SgdThreadUpdater(const OptimizationConfig& optConfig); - virtual ~SgdThreadUpdater() {} - - // Use the startPass() function of the base optimizer. - virtual void startPass(); - - // Use the finishPass() function of the base optimizer. - virtual bool finishPass(); - - virtual void init(const std::vector& parameters); - virtual PassType startBatch(int64_t batchSize); - // Call finishBatch for each optimizer. - virtual void finishBatch(real cost); - virtual void catchUpWith(); - virtual void apply(); - virtual void restore(); - - protected: - // This is the function that will be eventualy called by the GradientMachine. - // used only for GPU update. - virtual void updateImpl(Parameter* para); - OptimizationConfig config_; - int64_t numSamplesProcessed_; - - // One optimizers for each parameter. - std::vector> optimizers_; - - // The update function for CPU sparse parameters. - void threadUpdateSparse(int tid, size_t numThreads, Parameter* para); - - // The update function for CPU dense parameters. - void threadUpdateDense(int tid, size_t numThreads, Parameter* para); - // The update function for after update operations, such as averager. - void threadTraverse(const ParameterOptimizer::TraverseCallback& callback, - int tid, - size_t numThreads, - Parameter* para); - typedef std::function - GetTraverseCallback; - void traverse(GetTraverseCallback getTraverseCallback); -}; - -} // namespace paddle diff --git a/paddle/trainer/Trainer.cpp b/paddle/trainer/Trainer.cpp deleted file mode 100644 index 3e4a2b5fa8a3981f6362edc1dc61ae1616e257ef..0000000000000000000000000000000000000000 --- a/paddle/trainer/Trainer.cpp +++ /dev/null @@ -1,653 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "Trainer.h" - -#include - -#include -#include -#include -#include - -#include - -#include "paddle/utils/Common.h" -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -#include "RemoteParameterUpdater.h" -#include "TesterConfig.h" -#include "ThreadParameterUpdater.h" -#include "TrainerConfigHelper.h" -#include "paddle/gserver/gradientmachines/GradientMachineMode.h" -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/gserver/layers/ValidationLayer.h" - -DEFINE_string(config, "", "Trainer config file"); - -DEFINE_int32(test_period, - 0, - "if equal 0, do test on all test data at the end of " - "each pass. While if equal non-zero, do test on all test " - "data every test_period batches"); -DEFINE_bool(test_all_data_in_one_period, - false, - "This option was deprecated, since we will always do " - "test on all test set "); - -DEFINE_bool(local, true, "Train in local mode or not"); - -DEFINE_int32(average_test_period, - 0, - "Do test on average parameter every so" - " many batches. MUST be devided by FLAGS_log_period." - " Default 0 means do not test average parameter"); - -DEFINE_int32(saving_period, 1, "Save parameteres every so many passes"); -DEFINE_int64(saving_period_by_batches, - 0, - "Save parameters every so many batches in one pass"); -DEFINE_string(save_dir, "", "Directory for saving model parameter"); -DEFINE_int32(start_pass, - 0, - "Start training from this pass. " - "Will load parameter from the previous pass"); -DEFINE_int32(test_pass, -1, "Will load parameter start from this pass to test"); -DEFINE_int32(test_wait, 0, "Waiting for pass parameter if not exist"); -DEFINE_bool(with_cost, true, "enable cost layer or not"); -DEFINE_bool(distribute_test, false, "test in distribute mode"); - -DEFINE_int32(num_passes, 100, "train for so many passes"); - -DEFINE_string(config_args, - "", - "arguments passed to config file." - "Format: key1=value1,key2=value2"); - -DEFINE_bool(save_only_one, - false, - "Save only parameters in last pass, remove previous."); - -DEFINE_string(feat_file, "", "File name of extracted feature."); -DEFINE_string(predict_output_dir, - "", - "Directory that saves the predicted results of output layers"); -DEFINE_string(model_list, "", "File that saves the model list when evaluation"); - -namespace paddle { - -void Trainer::init(const std::shared_ptr& config, - bool testing, - const std::shared_ptr& gradientMachine, - const std::shared_ptr& dataProvider, - const std::shared_ptr& testDataProvider) { - this->stats_ = std::make_shared(); - - config_ = config; - - config_->updateConfigFromFlags(); - - testing_ = testing; - - // in testing, mode_ may GradientMachine::kTesting or - // GradientMachine::kSgdSparseCpuTraining - - if (FLAGS_local) { - CHECK(!FLAGS_loadsave_parameters_in_pserver) - << "local and loadsave_parameters_in_pserver can not both true"; - if (config_->getOptConfig().use_sparse_remote_updater()) { - config_->disableRemoteSparseUpdaterForEachParams(); - LOG(INFO) << "ignore sparse_remote_update=true due to --local=true"; - } - } - if (FLAGS_loadsave_parameters_in_pserver) { - CHECK(config_->getOptConfig().use_sparse_remote_updater()) - << "no parameter to load from pserver, please check network config"; - } - if (testing && !FLAGS_loadsave_parameters_in_pserver) { - if (config_->getOptConfig().use_sparse_remote_updater()) { - config_->disableRemoteSparseUpdater(); - LOG(INFO) << "because parameter is loaded local," - << "tester ignore sparse_remote_update flag"; - } - } - - CHECK(TrainAlgorithm::isValid(config_->getOptConfig().algorithm())) - << "invalid algorithm configuration: " - << config_->getOptConfig().algorithm(); - - bool useSparseUpdater = false; - for (auto& paraConfig : config_->getModelConfig().parameters()) { - if (paraConfig.sparse_update() || paraConfig.sparse_remote_update()) { - useSparseUpdater = true; - } - } - - if (FLAGS_use_mkldnn) { - CHECK_EQ(FLAGS_trainer_count, 1) << "MKLDNN only need 1 trainer"; - } - - if (testing) { - LOG(INFO) << "trainer: in testing mode"; - if (config_->getOptConfig().use_sparse_remote_updater() || - FLAGS_trainer_count > 1) { - mode_ = GradientMachine::kSgdSparseCpuTraining; - LOG(INFO) << "trainer mode: SgdSparseCpuTraining"; - } else { - mode_ = GradientMachine::kTesting; - LOG(INFO) << "trainer mode: Testing"; - } - } else if (IGradientMachineMode::tryGetMode( - (int*)&mode_, - config_->getOptConfig().algorithm(), - FLAGS_trainer_count, - FLAGS_local, - FLAGS_use_gpu)) { - LOG(INFO) << "Custom trainer mode."; - } else if ((config_->getOptConfig().algorithm() == TrainAlgorithm::SGD || - config_->getOptConfig().algorithm() == - TrainAlgorithm::AsyncSGD) && - useSparseUpdater) { - mode_ = GradientMachine::kSgdSparseCpuTraining; - LOG(INFO) << "trainer mode: SgdSparseCpuTraining"; - } else { - mode_ = GradientMachine::kNormal; - LOG(INFO) << "trainer mode: Normal"; - } - - // initialize trainer internal - trainerInternal_.init(config_, - gradientMachine, - TrainerInternalConfig::createFromMode(mode_), - stats_, - testing); - std::unique_ptr paramConfig( - new ParameterUtilConfig(FLAGS_save_only_one, - FLAGS_saving_period, - FLAGS_loadsave_parameters_in_pserver, - FLAGS_config)); - - paramUtil_.reset( - new paddle::ParameterUtil(config_, - std::move(paramConfig), - trainerInternal_.getGradientMachine(), - trainerInternal_.getParameterUpdater())); - - bool gpuData = - FLAGS_use_gpu && (!FLAGS_parallel_nn) && - (!IGradientMachineMode::dataMustInCpu(mode_, FLAGS_trainer_count)); - - dataProvider_ = dataProvider; - if (!dataProvider_ && config_->hasDataConfig() && !testing_) { - dataProvider_.reset(DataProvider::create(*config_, *config_, gpuData)); - } - if (!testDataProvider_) { - // No evaluator_ if there is testDataProvider but no dataProvider. - evaluator_.reset(trainerInternal_.getGradientMachine()->makeEvaluator()); - currentEvaluator_.reset( - trainerInternal_.getGradientMachine()->makeEvaluator()); - if (FLAGS_average_test_period > 0 && FLAGS_trainer_id == 0 && - config_->getOptConfig().average_window() > 0) { - CHECK_EQ(FLAGS_average_test_period % FLAGS_log_period, 0) - << "FLAGS_average_test_period must be divided by FALGS_log_period"; - averageEvaluator_.reset( - trainerInternal_.getGradientMachine()->makeEvaluator()); - } - } - - testDataProvider_ = testDataProvider; - if (!testDataProvider_ && config_->hasTestDataConfig()) { - testDataProvider_.reset( - DataProvider::create(config_->getTestDataConfig(), *config_, gpuData)); - } - if (testDataProvider_) { - createTester(); - } - - if (!testing && - (trainerInternal_.getGradientMachine()->hasStaticParameters())) { - CHECK(!FLAGS_loadsave_parameters_in_pserver) - << "is_static and loadsave_parameters_in_pserver can not both true"; - } - if (testing) { - // will load per pass for tester - } else if (paramUtil_->tryLoadParametersFromConfig()) { - // load from config already. - } else { - trainerInternal_.getGradientMachine()->randParameters(); - } - - // Only non static parameters need to be updated - std::vector& parameters = - trainerInternal_.getGradientMachine()->getNonStaticParameters(); - if (trainerInternal_.getParameterUpdater()) { - trainerInternal_.getParameterUpdater()->init(parameters); - - if (FLAGS_loadsave_parameters_in_pserver && FLAGS_trainer_id == 0) { - if (testing) { - // will load per pass for tester - } else if (!config_->getConfig().init_model_path().empty() && - (FLAGS_local || FLAGS_trainer_id == 0)) { - paramUtil_->loadParametersWithPath( - config_->getConfig().init_model_path(), - false /*local*/, - true /*remote*/); - } else if (config_->getConfig().start_pass() > 0 && - (FLAGS_local || FLAGS_trainer_id == 0)) { - CHECK(paramUtil_->loadParameters(config_->getConfig().start_pass() - 1, - false /*local*/, - true /*remote*/)); - } else { - trainerInternal_.getParameterUpdater()->randParametersRemote(); - } - } - } - - // set current evaluator and evalutor - trainerInternal_.setCurrentEvaluator(currentEvaluator_.get()); - trainerInternal_.setEvaluator(evaluator_.get()); -} - -void Trainer::train(size_t numPasses) { - startTrain(); - for (size_t i = 0; i < numPasses; ++i) { - if (IGradientMachineMode::trainWholeDataInOneBatch(mode_)) { - trainOnePassBatch(config_->getConfig().start_pass() + i); - } else { - trainOnePass(); - } - if (i < numPasses - 1) { - dataProvider_->reset(); - } - } - - finishTrain(); -} - -static double genPerturbation(real* d, real* grad, size_t dim) { - auto& reng = ThreadLocalRandomEngine::get(); - std::uniform_real_distribution dist(-1, 1); - double gradNorm = 0, dNorm = 0; - for (size_t i = 0; i < dim; ++i) { - d[i] = dist(reng); - dNorm += d[i] * d[i]; - gradNorm += grad[i] * grad[i]; - } - if (gradNorm > 0) { - real s = 0.5 * sqrt(gradNorm / dNorm); - for (size_t i = 0; i < dim; ++i) { - d[i] = s * d[i] + grad[i]; - } - } - double delta = 0; - for (size_t i = 0; i < dim; ++i) { - delta += grad[i] * d[i]; - } - return delta; -} - -real Trainer::checkGradient() { - trainerInternal_.getGradientMachine()->start(); - std::vector& parameters = - trainerInternal_.getGradientMachine()->getNonStaticParameters(); - DataBatch dataBatch; - int32_t batchSize = config_->getOptConfig().batch_size(); - - dataProvider_->getNextBatch(batchSize, &dataBatch); - - CHECK(dataBatch.getSize()) << "No data from data provider"; - std::vector& inArgs = dataBatch.getStreams(); - std::vector outArgs; - - trainerInternal_.getGradientMachine()->forward(inArgs, &outArgs, PASS_GC); - real cost = Argument::sum(outArgs); - LOG(INFO) << "original cost=" << cost; - trainerInternal_.getGradientMachine()->backward(); - - real maxDiff = 0; - char fill = ' '; - for (auto& parameter : parameters) { - CpuVector oldPara(parameter->getSize()); - CpuVector newPara(parameter->getSize()); - oldPara.copyFrom(*parameter->getBuf(PARAMETER_VALUE)); - real* newp = newPara.getData(); - real* oldp = oldPara.getData(); - CpuVector cpuGrad(*parameter->getBuf(PARAMETER_GRADIENT)); - real* grad = cpuGrad.getData(); - size_t dim = parameter->getSize(); - std::vector d(dim); - - double delta = genPerturbation(d.data(), grad, dim); - - // use a step such that delta / cost is FLAGS_checkgrad_eps - real step = - (delta != 0) ? cost / delta * FLAGS_checkgrad_eps : FLAGS_checkgrad_eps; - delta *= step; - for (size_t i = 0; i < dim; ++i) { - newp[i] = oldp[i] + step * d[i]; - } - - parameter->getBuf(PARAMETER_VALUE)->copyFrom(newPara); - parameter->setValueUpdated(); - trainerInternal_.getGradientMachine()->forward(inArgs, &outArgs, PASS_GC); - real newCost1 = Argument::sum(outArgs); - - for (size_t i = 0; i < dim; ++i) { - newp[i] = oldp[i] - step * d[i]; - } - - parameter->getBuf(PARAMETER_VALUE)->copyFrom(newPara); - parameter->setValueUpdated(); - trainerInternal_.getGradientMachine()->forward(inArgs, &outArgs, PASS_GC); - real newCost2 = Argument::sum(outArgs); - - real trueDelta = 0.5 * (newCost1 - newCost2); - real diff = (1e-20 + trueDelta) / (1e-20 + delta) - 1; - LOG(INFO) << std::setiosflags(std::ios::left) << std::setfill(fill) - << std::setw(20) << parameter->getName() - << "step=" << std::setw(15) << step << "cost1=" << std::setw(10) - << newCost1 << "cost2=" << std::setw(10) << newCost2 - << "true_delta=" << std::setw(15) << trueDelta - << "analytic_delta=" << std::setw(15) << delta << "diff=" << diff - << (std::abs(diff) > 0.01 ? " ***" : ""); - - maxDiff = std::max(maxDiff, std::abs(diff)); - - // restore parameter - parameter->getBuf(PARAMETER_VALUE)->copyFrom(oldPara); - parameter->setValueUpdated(); - - fill = (fill == ' ') ? '.' : ' '; - } - return maxDiff; -} - -void Trainer::startTrain() { - trainPassContext_.passId = config_->getConfig().start_pass(); - srand(config_->getConfig().start_pass() + 1); - if (dataProvider_) { - dataProvider_->reset(); - } - - trainerInternal_.getGradientMachine()->start(); -} - -void Trainer::finishTrain() { trainerInternal_.getGradientMachine()->finish(); } - -void Trainer::startTrainPass() { - stats_->reset(); - trainPassContext_.batchId = 0; - trainPassContext_.avgTestCost = 0; - trainPassContext_.numAvgTests = 0; - trainPassContext_.passInnerId = 1; - - trainerInternal_.getParameterUpdater()->startPass(); - evaluator_->start(); - if (FLAGS_prev_batch_state) { - trainerInternal_.getGradientMachine()->resetState(); - trainerInternal_.getGradientMachine()->getState(testState_); - } -} - -void Trainer::trainOneDataBatch(DataBatch& dataBatch) { - int num = dataBatch.getSize(); - if (averageEvaluator_) { - int64_t mod = trainPassContext_.batchId % FLAGS_average_test_period; - if (mod >= FLAGS_average_test_period - FLAGS_log_period) { - if (mod == FLAGS_average_test_period - FLAGS_log_period) { - averageEvaluator_->start(); - } - trainerInternal_.getParameterUpdater()->apply(); - if (FLAGS_prev_batch_state) { - trainerInternal_.getGradientMachine()->getState(trainState_); - } - trainPassContext_.avgTestCost += tester_->forwardOneBatch( - dataBatch, averageEvaluator_.get(), &forwardOutput_); - if (FLAGS_prev_batch_state) { - trainerInternal_.getGradientMachine()->setState(trainState_); - } - trainPassContext_.numAvgTests += num; - trainerInternal_.getParameterUpdater()->restore(); - } - } - { - REGISTER_TIMER("TrainBatch"); - trainerInternal_.trainOneBatch( - trainPassContext_.batchId, dataBatch, &forwardOutput_); - } - - if (averageEvaluator_ && - trainPassContext_.batchId % FLAGS_average_test_period == - FLAGS_average_test_period - 1) { - averageEvaluator_->finish(); - LOG(INFO) << " Averaged parameter:" - << " cost=" - << trainPassContext_.avgTestCost / trainPassContext_.numAvgTests - << " Eval: " << *averageEvaluator_; - trainPassContext_.numAvgTests = 0; - trainPassContext_.avgTestCost = 0; - } - - ++trainPassContext_.batchId; - - if (trainPassContext_.batchId % FLAGS_log_period == 0) { - FOR_TIMING(globalStat.setThreadInfo(true)); - FOR_TIMING(globalStat.printAllStatus()); - FOR_TIMING(globalStat.reset()); - } - - if (testDataProvider_ && FLAGS_test_period > 0 && - trainPassContext_.batchId % FLAGS_test_period == 0) { - tester_->testOnePeriod(); - } - - if (FLAGS_saving_period_by_batches > 0 && - trainPassContext_.batchId > - FLAGS_saving_period_by_batches * trainPassContext_.passInnerId && - 0 == FLAGS_trainer_id) { - trainerInternal_.getParameterUpdater()->catchUpWith(); - if (testDataProvider_) { - tester_->testOnePeriod(); - } - paramUtil_->saveParametersOnePass(trainPassContext_.passId, - trainPassContext_.passInnerId); - ++trainPassContext_.passInnerId; - } -} - -void Trainer::finishTrainPass() { - if (trainPassContext_.batchId == 0) { - // This means no more data from DataProvider - return; - } - - trainerInternal_.finishTrainPass(trainPassContext_.passId, - trainPassContext_.batchId); - - FOR_TIMING(globalStat.setThreadInfo(true)); - FOR_TIMING(globalStat.printAllStatus()); - FOR_TIMING(globalStat.reset()); - - if (testDataProvider_) { - tester_->testOnePeriod(); - } - - if (trainPassContext_.passId % FLAGS_saving_period == 0 && - FLAGS_trainer_id == 0) { - paramUtil_->saveParametersOnePass(trainPassContext_.passId); - } - ++trainPassContext_.passId; -} - -void Trainer::trainOnePass() { - startTrainPass(); - size_t batchSize = config_->getOptConfig().batch_size(); - while (true) { - DataBatch dataBatch; - - int num = 0; - { - REGISTER_TIMER("getTrainBatch"); - num = dataProvider_->getNextBatch(batchSize, &dataBatch); - } - if (num == 0) break; - CHECK_EQ(num, dataBatch.getSize()); - trainOneDataBatch(dataBatch); - } - - finishTrainPass(); -} - -void Trainer::trainOnePassBatch(int passId) { - this->stats_->reset(); - - trainerInternal_.getParameterUpdater()->startPass(); - const std::vector inArgs; - { - REGISTER_TIMER("onePass"); - trainerInternal_.getGradientMachine()->forwardBackward( - inArgs, nullptr, PASS_TRAIN, nullptr); - } - - real cost = .0; - int64_t num = 0; - trainerInternal_.getGradientMachine()->getStats(cost, num); - *stats_ += {num, cost}; - - trainerInternal_.getGradientMachine()->onPassEnd(); - - bool accepted = trainerInternal_.getParameterUpdater()->finishPass(); - - globalStat.setThreadInfo(true); - globalStat.printAllStatus(); - globalStat.reset(); - - LOG(INFO) << " Pass=" << passId - << " AcceptedPass=" << (accepted ? acceptedPassId_ : -1) - << stats_->getStats(false /*withCurrentCost*/); - - if (accepted) { - if (acceptedPassId_ % FLAGS_saving_period == 0 && FLAGS_trainer_id == 0) { - paramUtil_->saveParameters(acceptedPassId_); - } - acceptedPassId_++; - if (FLAGS_save_only_one && acceptedPassId_ >= FLAGS_saving_period) { - paramUtil_->deleteParameters(acceptedPassId_ - FLAGS_saving_period); - } - } -} - -real Trainer::calcGradient(const DataBatch& dataBatch, - const Vector& value, - Vector& gradient) { - CHECK_EQ(value.getSize(), gradient.getSize()); - std::vector& parameters = - trainerInternal_.getGradientMachine()->getParameters(); - - clearGradient(); - - size_t offset = 0; - size_t valueSize = value.getSize(); - - for (auto& para : parameters) { - CHECK_LE(offset + para->getSize(), valueSize); - VectorPtr val = - Vector::create(para->getSize(), value.getMemoryHandle(), offset); - para->getBuf(PARAMETER_VALUE)->copyFrom(*val); - para->setValueUpdated(); - offset += para->getSize(); - } - - CHECK_EQ(offset, valueSize); - - std::vector inArgs = dataBatch.getStreams(); - std::vector outArgs; - - trainerInternal_.getGradientMachine()->forwardBackward( - inArgs, &outArgs, PASS_TRAIN); - real cost = Argument::sum(outArgs); - - offset = 0; - for (auto& para : parameters) { - VectorPtr grad = - Vector::create(para->getSize(), gradient.getMemoryHandle(), offset); - if (para->getBuf(PARAMETER_GRADIENT)) { - grad->copyFrom(*para->getBuf(PARAMETER_GRADIENT)); - } - offset += para->getSize(); - } - - return cost; -} - -void Trainer::clearGradient() { - std::vector& parameters = - trainerInternal_.getGradientMachine()->getNonStaticParameters(); - for (auto& parameter : parameters) { - parameter->clearGradient(); - } -} - -int Trainer::getBatchSize() { return config_->getOptConfig().batch_size(); } - -void Trainer::createTester() { - tester_.reset(new paddle::Tester(config_, - createTesterConfig(), - trainerInternal_.getGradientMachine(), - trainerInternal_.getParameterUpdater(), - testDataProvider_)); -} - -void Trainer::test() { tester_->test(); } - -std::unique_ptr Trainer::createTesterConfig() { - TesterConfig* conf = new TesterConfig; - if (FLAGS_test_period) { - LOG(WARNING) << "The meaning of --test_period is changed: " - << "if equal 0, do test on all test data at the end of " - << "each pass. While if equal non-zero, do test on all test " - << "data every test_period batches "; - } - if (FLAGS_test_all_data_in_one_period) { - LOG(WARNING) << "--test_all_data_in_one_period was deprecated, since " - << "we will always do test on all test set "; - } - conf->testPeriod = FLAGS_test_period; - conf->prevBatchState = FLAGS_prev_batch_state; - conf->logPeriod = FLAGS_log_period; - conf->loadsaveParametersInPserver = FLAGS_loadsave_parameters_in_pserver; - conf->featFile = FLAGS_feat_file; - conf->predictOutputDir = FLAGS_predict_output_dir; - conf->trainerId = FLAGS_trainer_id; - conf->distributeTest = FLAGS_distribute_test; - conf->config = FLAGS_config; - conf->modelList = FLAGS_model_list; - conf->testPass = FLAGS_test_pass; - conf->numPasses = FLAGS_num_passes; - conf->savingPeriod = FLAGS_saving_period; - conf->testWait = FLAGS_test_wait; - conf->initModelPath = FLAGS_init_model_path; - conf->saveOnlyOne = FLAGS_save_only_one; - conf->testing = testing_; - conf->mode = mode_; - conf->trainState = &trainState_; - conf->testState = &testState_; - return std::unique_ptr(conf); -} - -ParameterUtil* Trainer::getParameterUtilPtr() { return paramUtil_.get(); } -} // namespace paddle diff --git a/paddle/trainer/Trainer.h b/paddle/trainer/Trainer.h deleted file mode 100644 index 78127b7be5cef34f51a4b540852c139625b571dd..0000000000000000000000000000000000000000 --- a/paddle/trainer/Trainer.h +++ /dev/null @@ -1,204 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include - -#include "hl_gpu.h" -#include "paddle/gserver/dataproviders/DataProvider.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" - -#include -#include -#include "ParamUtil.h" -#include "ParameterUpdater.h" -#include "Tester.h" -#include "TrainerConfigHelper.h" -#include "TrainerInternal.h" - -DECLARE_int32(num_passes); - -namespace paddle { - -/** - * Trainer Class - * - * Trainer combines GradientMachine, ParameterUpdater, DataProvider together to - * train/test a NeuralNetwork. - */ -class Trainer { - public: - /** - * Ctor. - * @return - */ - Trainer() : acceptedPassId_(0) {} - - virtual ~Trainer() {} - - /** - * initialize a new trainer using config - * - * @param config TrainerConfig. - * @param testing true if only for testing - * @param gradientMachine GradientMachine that will be trained. - * nullptr if create from config. - * @param dataProvider Train Data Provider. null if create from config. - * @param testDataProvider Test Data Provider. null if create from config. - */ - virtual void init( - const std::shared_ptr& config, - bool testing = false, - const std::shared_ptr& gradientMachine = nullptr, - const std::shared_ptr& dataProvider = nullptr, - const std::shared_ptr& testDataProvider = nullptr); - - /** - * Train until num_passes reached. - * One pass means neural network train through all training data. - * - * @param numPasses the number of traning pass. - * @note Durning neural network training, the num passes may set a very large - * value, and kill training process when result is good enough. - */ - void train(size_t numPasses = (size_t)FLAGS_num_passes); - - /** - * compare the gradient from bp with finite difference - * @return the maximal difference - */ - real checkGradient(); - - void startTrain(); - void finishTrain(); - void startTrainPass(); - void finishTrainPass(); - void trainOneDataBatch(DataBatch& dataBatch); - void time(); - - /** - * given a dataBatch and the current parameter value - * calculate its gradient and return the cost. - * - * TODO(yuyang18): I think this method is deprecated and buggy. Should it be - * removed? - */ - real calcGradient(const DataBatch& dataBatch, - const Vector& value, - Vector& gradient); - - /** - * Get Trainer Config. - */ - const TrainerConfig& getConfig() const { return config_->getConfig(); } - - /** - * Get Train Data Provider - */ - const DataProviderPtr& getDataProvider() { return dataProvider_; } - - /** - * Get Gradient Machine. - */ - const GradientMachinePtr& getGradientMachine() { - return trainerInternal_.getGradientMachine(); - } - - /** - * Get batch size in optimization config. - * @note This method didn't return the actual batch size. Just batch size - * set in the optimization config. The actual batch size in one trainer may - * less than batch size in config due to there are not enough data. - */ - int getBatchSize(); - - /** - * Do test job - */ - void test(); - - /** - * Get parameter util ptr - * - * TODO(yuyang18): Make it return a smart pointer. - */ - ParameterUtil* getParameterUtilPtr(); - - protected: - /** - * Train one pass of data. - * - * SGD Method. - */ - void trainOnePass(); - - /** - * Train one pass in one batch. - * - */ - void trainOnePassBatch(int passId); - - /** - * set parameter gradient to zero - */ - void clearGradient(); - - void createTester(); - - private: - std::unique_ptr createTesterConfig(); - - protected: - std::shared_ptr config_; - std::shared_ptr stats_; - - DataProviderPtr dataProvider_; - DataProviderPtr testDataProvider_; - MachineState trainState_; - MachineState testState_; - - struct TrainPassContext { - int64_t batchId; - real avgTestCost; - int64_t numAvgTests; - int passId; - int passInnerId; - }; - std::vector forwardOutput_; - - TrainPassContext trainPassContext_; - - std::unique_ptr evaluator_; - std::unique_ptr currentEvaluator_; - std::unique_ptr averageEvaluator_; - // training mode - // used to decide which GradientMachine and ParameterUpdater to create - GradientMachine::CreateMode mode_; - int testing_; - int acceptedPassId_; - - // trainer tester - std::unique_ptr tester_; - - // parameter util - std::unique_ptr paramUtil_; - - // trainer Internal - TrainerInternal trainerInternal_; -}; - -} // namespace paddle diff --git a/paddle/trainer/TrainerBenchmark.cpp b/paddle/trainer/TrainerBenchmark.cpp deleted file mode 100644 index 173653c81688fe4606731c68ea1854268b3f4590..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerBenchmark.cpp +++ /dev/null @@ -1,71 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#undef PADDLE_DISABLE_TIMER - -#include "Trainer.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -DECLARE_int32(test_period); - -DEFINE_bool(feed_data, false, "Wether to read data from DataProvider."); - -namespace paddle { - -void Trainer::time() { - startTrain(); - - trainerInternal_.getParameterUpdater()->startPass(); - evaluator_->start(); - - DataBatch dataBatch; - int32_t batchSize = config_->getOptConfig().batch_size(); - int32_t num = dataProvider_->getNextBatch(batchSize, &dataBatch); - CHECK_EQ(num, batchSize) << "The sample number is less than batch size " - << num << " != " << batchSize; - - CHECK(dataBatch.getSize()) << "No data from data provider"; - - std::vector outputs; - // burning time - LOG(INFO) << "Burning time..."; - for (int n = 0; n < 10; ++n) { - trainerInternal_.trainOneBatch(n, dataBatch, &outputs); - } - LOG(INFO) << "Burning time end."; - - for (int n = 0; n < FLAGS_test_period; n++) { - if (FLAGS_feed_data) { - REGISTER_TIMER("GetData"); - num = dataProvider_->getNextBatch(batchSize, &dataBatch); - } - - if (num != batchSize) { - break; - } - - { - REGISTER_TIMER("FwdBwd"); - trainerInternal_.trainOneBatch(n, dataBatch, &outputs); - } - } - globalStat.setThreadInfo(true); - globalStat.printSegTimerStatus(); - globalStat.reset(); - - finishTrain(); -} - -} // namespace paddle diff --git a/paddle/trainer/TrainerConfigHelper.cpp b/paddle/trainer/TrainerConfigHelper.cpp deleted file mode 100644 index 2b68d89e48a3efd5de205ce33643b7e6320a4303..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerConfigHelper.cpp +++ /dev/null @@ -1,199 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "TrainerConfigHelper.h" -#include "ParamUtil.h" -#include "TrainerConfig.pb.h" -#include "paddle/utils/Flags.h" -#include "paddle/utils/PythonUtil.h" - -DECLARE_string(config); -DECLARE_string(init_model_path); -DECLARE_int32(start_pass); -DECLARE_string(save_dir); -DECLARE_int32(trainer_id); -DECLARE_bool(local); -DECLARE_bool(with_cost); -DECLARE_bool(with_gpu); -DECLARE_bool(parallel_nn); -DECLARE_string(config_args); -DECLARE_bool(use_mkldnn); -DECLARE_bool(use_mkl_packed); - -const char *kConfigParserModuleName = "paddle.trainer.config_parser"; -const char *kConfigParserFuncName = "parse_config_and_serialize"; - -namespace paddle { - -struct TrainerConfigHelperPrivate { - TrainerConfig conf; -}; - -TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath) - : m(new TrainerConfigHelperPrivate()) { - std::ostringstream configArgs; - configArgs << "trainer_id=" << FLAGS_trainer_id << ",local=" << FLAGS_local - << ",with_cost=" << FLAGS_with_cost << ",use_gpu=" << FLAGS_use_gpu - << ",parallel_nn=" << FLAGS_parallel_nn - << ",use_mkldnn=" << FLAGS_use_mkldnn - << ",use_mkl_packed=" << FLAGS_use_mkl_packed - << ",cudnn_version=" << hl_get_cudnn_lib_version(); - if (!FLAGS_config_args.empty()) { - configArgs << "," << FLAGS_config_args; - } - - VLOG(3) << "Parsing trainer config " << configFilePath; - std::string configProtoStr = - callPythonFunc(kConfigParserModuleName, - kConfigParserFuncName, - {configFilePath, configArgs.str()}); - CHECK(m->conf.ParseFromString(configProtoStr)); -} - -TrainerConfigHelper::TrainerConfigHelper(const TrainerConfig &config) - : m(new TrainerConfigHelperPrivate()) { - m->conf = config; -} - -TrainerConfigHelper::~TrainerConfigHelper() { delete m; } - -const TrainerConfig &TrainerConfigHelper::getConfig() const { return m->conf; } - -TrainerConfig &TrainerConfigHelper::getMutableConfig() { return m->conf; } - -const OptimizationConfig &TrainerConfigHelper::getOptConfig() const { - return m->conf.opt_config(); -} - -const ModelConfig &TrainerConfigHelper::getModelConfig() const { - return m->conf.model_config(); -} - -const DataConfig *TrainerConfigHelper::getDataConfigPtr() const { - if (m->conf.has_data_config()) { - return &m->conf.data_config(); - } else { - return nullptr; - } -} - -const DataConfig &TrainerConfigHelper::getTestDataConfig() const { - CHECK(m->conf.has_test_data_config()); - return m->conf.test_data_config(); -} - -bool TrainerConfigHelper::hasDataConfig() const { - return m->conf.has_data_config(); -} - -bool TrainerConfigHelper::hasTestDataConfig() const { - return m->conf.has_test_data_config(); -} - -void TrainerConfigHelper::updateConfigFromFlags() { - if (!FLAGS_save_dir.empty()) { - m->conf.set_save_dir(FLAGS_save_dir); - } - if (!FLAGS_init_model_path.empty()) { - m->conf.set_init_model_path(FLAGS_init_model_path); - } - if (FLAGS_start_pass != 0) { - m->conf.set_start_pass(FLAGS_start_pass); - } -} - -void TrainerConfigHelper::disableRemoteSparseUpdater() { - m->conf.mutable_opt_config()->set_use_sparse_remote_updater(false); -} - -void TrainerConfigHelper::disableRemoteSparseUpdaterForEachParams() { - this->disableRemoteSparseUpdater(); - for (int i = 0; i < m->conf.model_config().parameters_size(); ++i) { - m->conf.mutable_model_config() - ->mutable_parameters(i) - ->set_sparse_remote_update(false); - } -} - -OptimizationConfig &TrainerConfigHelper::getOptConfig() { - return *m->conf.mutable_opt_config(); -} - -void TrainerConfigHelper::setSaveDir(const std::string &saveDir) { - m->conf.set_save_dir(saveDir); -} - -const std::string &TrainerConfigHelper::getSaveDir() const { - return m->conf.save_dir(); -} - -std::string TrainerConfigHelper::getConfigNameFromPath( - const std::string &modelPath) { - std::ifstream s(path::join(modelPath, "path.txt")); - CHECK(s.is_open()) << " fail to open path.txt"; - std::string ss; - getline(s, ss); - VLOG(3) << "fileName " << path::join(modelPath, ss); - s.close(); - return path::join(modelPath, ss); -} - -std::string TrainerConfigHelper::getConfigNameFromPassId( - int passId, const std::string &modelPath) { - constexpr int kBufLen = 100; - char buf[kBufLen]; - snprintf(buf, kBufLen, "pass-%05d", passId); - return TrainerConfigHelper::getConfigNameFromPath(path::join(modelPath, buf)); -} - -std::string TrainerConfigHelper::getConfigName(bool *ok) const { - std::string retv = ""; - - if (!m->conf.config_file().empty()) { - retv = m->conf.config_file(); - } else if (!m->conf.init_model_path().empty()) { - retv = getConfigNameFromPath(m->conf.init_model_path()); - } else if (m->conf.start_pass() >= 1) { - retv = getConfigNameFromPassId(m->conf.start_pass(), m->conf.save_dir()); - } - - if (ok) { - *ok = !retv.empty(); - } - - return retv; -} - -std::shared_ptr TrainerConfigHelper::createFromFlags() { - std::string configPath; - if (!FLAGS_config.empty()) { - configPath = FLAGS_config; - } else if (!FLAGS_init_model_path.empty()) { - configPath = getConfigNameFromPath(FLAGS_init_model_path); - } else if (FLAGS_start_pass >= 1) { - configPath = - getConfigNameFromPassId(FLAGS_start_pass - 1, FLAGS_init_model_path); - } else { - return nullptr; - } - return std::make_shared(configPath); -} - -std::shared_ptr -TrainerConfigHelper::createFromFlagConfig() { - CHECK(!FLAGS_config.empty()); - return std::make_shared(FLAGS_config); -} - -} // namespace paddle diff --git a/paddle/trainer/TrainerConfigHelper.h b/paddle/trainer/TrainerConfigHelper.h deleted file mode 100644 index b21dda964e70fce6e5e9672cc131595ad5af3bbc..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerConfigHelper.h +++ /dev/null @@ -1,205 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include -#include -#include - -namespace paddle { - -class TrainerConfig; -class OptimizationConfig; -struct TrainerConfigHelperPrivate; -class ModelConfig; -class DataConfig; - -/** - * @brief TrainerConfig Helper. A class wrap protobuf's TrainerConfig Object, - * simplize the usage for TrainerConfig. - * - * The all operation to TrainerConfig object should use this object. It remove - * many copy & paste code in trainer. - * - * @TODO(yuyang18): Make cmake check compiler support keyword 'final' or not. - * Define a macro to unify 'final' keyword - */ -class TrainerConfigHelper /*final*/ { - public: - DISABLE_COPY(TrainerConfigHelper); - - /** - * @brief Ctor, Create a TrainerConfig from config file - * @param configFilePath Config file path. - */ - explicit TrainerConfigHelper(const std::string& configFilePath); - explicit TrainerConfigHelper(const TrainerConfig& config); - - /** - * Dtor - * @warning this class is a final class. Should not be inherited. - */ - ~TrainerConfigHelper(); - - /** - * @brief Get Trainer Config itself. - */ - const TrainerConfig& getConfig() const; - - TrainerConfig& getMutableConfig(); - - /** - * @brief Get Optimizer Config. - */ - const OptimizationConfig& getOptConfig() const; - - /** - * @brief Get Model Config. - */ - const ModelConfig& getModelConfig() const; - - /** - * @brief Get Train Data Config Pointer. - * @return nullptr if there is no train data. Else will return pointer - */ - const DataConfig* getDataConfigPtr() const; - - /** - * @brief Get Tain Data Config. - * @warning Core when there is no train data. - */ - const DataConfig& getDataConfig() const { - CHECK(this->hasDataConfig()); - auto conf = this->getDataConfigPtr(); - return *conf; - } - - /** - * @brief Get test data config - * @warning Core when there is no test data. - */ - const DataConfig& getTestDataConfig() const; - - /** - * @brief Has train data config or not. - * @return true if has train data. - */ - bool hasDataConfig() const; - - /** - * @brief Has test data config or not. - * @return true if has test data. - */ - bool hasTestDataConfig() const; - - /** - * @brief Update trainer config from command line flags. - * Override config's (save_dir, init_model_path, start_pass) if command - * flags is existed. - */ - void updateConfigFromFlags(); - - /** - * @brief Disable optimization's sparse remote update. - */ - void disableRemoteSparseUpdater(); - - /** - * @brief Disable optimization and each parameter's sparse remote update. - */ - void disableRemoteSparseUpdaterForEachParams(); - - /** - * @brief implicit conversion. - */ - inline operator const TrainerConfig&() const { return this->getConfig(); } - - /** - * @brief implicit conversion. - */ - inline operator const OptimizationConfig&() const { - return this->getOptConfig(); - } - - /** - * @brief implicit conversion. - */ - inline operator const DataConfig&() const { return this->getDataConfig(); } - - /** - * @brief implicit conversion. - */ - inline operator const ModelConfig&() const { return this->getModelConfig(); } - - /** - * @brief Get mutable optimization config. - */ - OptimizationConfig& getOptConfig(); - - /** - * @brief set model save directory. - * @param saveDir Directory path. - */ - void setSaveDir(const std::string& saveDir); - - /** - * @brief get model save directory. - * @return save directory path. - */ - const std::string& getSaveDir() const; - - /** - * @brief Get config file name from model path. - * - * Paddle save model to a directory, and write a file 'path.txt' which save - * config filename. - * - * @param modelPath model saved directory. - * @return config file name. - */ - static std::string getConfigNameFromPath(const std::string& modelPath); - - /** - * @brief Get config file name from this config instance. - * @param[out] ok true if no error. - * @return config file name. - */ - std::string getConfigName(bool* ok = nullptr) const; - - /** - * @brief Try to create TrainerConfigHelper from all command line flags. - * Try to load from --config, --init_model_path, --start_pass one by - * one. Return nullptr if cannot load TrainerConfigHelper from all - * these place. - * @return nullptr if cannot load, otherwise return a TrainerConfigHelper. - */ - static std::shared_ptr createFromFlags(); - - /** - * @brief Try to create TrainerConfigHelper only from '--config' flag. - * @return nullptr if cannot load, otherwise return a TrainerConfigHelper. - */ - static std::shared_ptr createFromFlagConfig(); - - private: - static std::string getConfigNameFromPassId(int passId, - const std::string& modelPath); - - TrainerConfigHelperPrivate* m; -}; - -typedef std::shared_ptr TrainerConfigHelperPtr; - -} // namespace paddle diff --git a/paddle/trainer/TrainerInternal.cpp b/paddle/trainer/TrainerInternal.cpp deleted file mode 100644 index 4c5d4a0913aaf3a9932b3d67806378ece4245304..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerInternal.cpp +++ /dev/null @@ -1,303 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "TrainerInternal.h" - -#include -#include - -#include -#include -#include -#include - -#include - -#include "paddle/gserver/gradientmachines/NeuralNetwork.h" -#include "paddle/gserver/layers/ValidationLayer.h" -#include "paddle/utils/GlobalConstants.h" -#include "paddle/utils/PythonUtil.h" -#include "paddle/utils/Stat.h" -#include "paddle/utils/Util.h" - -#include "RemoteParameterUpdater.h" -#include "ThreadParameterUpdater.h" - -namespace paddle { - -void TrainerInternal::init(const std::shared_ptr& config, - const GradientMachinePtr& gradientMachine, - std::unique_ptr&& intconfig, - const std::shared_ptr& stats, - bool testing) { - config_ = config; - intconfig_ = std::move(intconfig); - stats_ = stats; - - //! in training will use parameter updater definitly. - //! But only use parameter in testing mode when some parameter in pserver. - if (!testing || (config_->getOptConfig().use_sparse_remote_updater() && - intconfig_->loadsave_parameters_in_pserver)) { - createParameterUpdater(testing); - } - - gradientMachine_ = gradientMachine; - if (!gradientMachine) { - CHECK(config_->getConfig().has_model_config()) - << "Missing model_config in trainer_config"; - gradientMachine_.reset( - GradientMachine::create(config_->getConfig().model_config(), - intconfig_->mode, - parameterUpdater_->getParameterTypes())); - } -} - -void TrainerInternal::trainOneBatch(int64_t batchId, - const DataBatch& dataBatch, - std::vector* outArgs) { - // true means updating parameter whenever gradient is ready during backward() - bool doPipelineUpdate = - (intconfig_->mode != GradientMachine::kSgdSparseCpuTraining) && - (intconfig_->local || intconfig_->use_gpu || - intconfig_->trainer_count <= 1); - - int64_t actualBatchSize = dataBatch.getSize(); - if (actualBatchSize == 0) { - return; - } - - bool showStats = intconfig_->show_param_stats_period > 0 && - (batchId + 1) % intconfig_->show_param_stats_period == 0 && - intconfig_->trainer_id == 0; - - std::vector paraStats; - if (showStats) { - paraStats.resize(gradientMachine_->getParameters().size()); - } - - const std::vector& inArgs = dataBatch.getStreams(); - - PassType passType = parameterUpdater_->startBatch(actualBatchSize); - - if (config_->getOptConfig().use_sparse_remote_updater()) { - REGISTER_TIMER("prefetch"); - gradientMachine_->prefetch(inArgs); - parameterUpdater_->getParametersRemote(); - } - - UpdateCallback updateCallback = [this, showStats, ¶Stats]( - Parameter* para) { - if (showStats) { - //! @TODO(yuyang18) Show stats is actually a ParameterHook, refactor - // it - //! to ParameterHook. - auto& grad = para->getBuf(PARAMETER_GRADIENT); - SetDevice device(para->getDeviceId()); - paraStats[para->getID()].avgAbsGrad = grad->getAbsSum() / para->getSize(); - paraStats[para->getID()].maxAbsGrad = grad->getAbsMax(); - } - parameterUpdater_->update(para); - }; - - { -#ifndef PADDLE_DISABLE_TIMER - Timer timer; - timer.start(); -#endif - REGISTER_TIMER("forwardBackward"); - forwardBackwardBatch( - inArgs, *outArgs, passType, updateCallback, doPipelineUpdate); -#ifndef PADDLE_DISABLE_TIMER - timer.stop(); - parameterUpdater_->setForwardbackwardTime(timer.get()); -#endif - } - - if (!doPipelineUpdate) { - auto& parameters = gradientMachine_->getNonStaticParameters(); - for (auto& para : parameters) { - updateCallback(para.get()); - } - } - - real cost = 0; - { - REGISTER_TIMER("sumCost"); - cost = Argument::sum(*outArgs); - } - - if (batchId % intconfig_->log_period == 0) { - currentEvaluator_->start(); - stats_->resetCurrentStat(); - } - { - REGISTER_TIMER("eval"); - gradientMachine_->eval(currentEvaluator_); - gradientMachine_->eval(evaluator_); - } - - *stats_ += {actualBatchSize, cost}; - { - REGISTER_TIMER("finishBatch"); - parameterUpdater_->finishBatch(cost); - } - - if (showStats) { - showParameterStats(paraStats); - } - if ((batchId + 1) % intconfig_->log_period == 0) { - currentEvaluator_->finish(); - - if (intconfig_->dot_period > 0) { - std::cerr << std::endl; - } - LOG(INFO) << " Batch=" << batchId + 1 << " " << *stats_ - << " Eval: " << *evaluator_ - << " CurrentEval: " << *currentEvaluator_; - } else if (intconfig_->dot_period > 0 && - (batchId + 1) % intconfig_->dot_period == 0) { - std::cerr << "."; - } -} - -/** - * finish train pass - */ -void TrainerInternal::finishTrainPass(int passId, int batchId) { - gradientMachine_->onPassEnd(); - parameterUpdater_->finishPass(); - evaluator_->finish(); - LOG(INFO) << " Pass=" << passId << " Batch=" << batchId << " " - << stats_->getStats(false /*without current cost*/) - << " Eval: " << *evaluator_; -} - -void TrainerInternal::showParameterStats( - const std::vector& paraStats) { - std::vector& parameters = gradientMachine_->getParameters(); - for (auto& parameter : parameters) { - SetDevice device(parameter->getDeviceId()); - real sum = parameter->getBuf(PARAMETER_VALUE)->getAbsSum(); - const auto& lr = parameter->getBuf(PARAMETER_LEARNING_RATE); - std::ostringstream osLrHistogram; - if (lr) { - if (VLOG_IS_ON(2)) { - osLrHistogram << " lr_histogram: "; - lr->histogram(osLrHistogram); - } else { - osLrHistogram << " max_lr=" << std::setw(11) << lr->getMax() - << " min_lr=" << std::setw(11) << lr->getMin() - << " avg_lr=" << std::setw(11) - << lr->getSum() / parameter->getSize(); - } - } - int pid = parameter->getID(); - LOG(INFO) << std::setiosflags(std::ios::left) << std::setfill(' ') - << std::setw(20) << parameter->getName() - << " avg_abs_val=" << std::setw(11) << sum / parameter->getSize() - << " max_val=" << std::setw(11) - << parameter->getBuf(PARAMETER_VALUE)->getAbsMax() - << " avg_abs_grad=" << std::setw(11) << paraStats[pid].avgAbsGrad - << " max_grad=" << std::setw(11) << paraStats[pid].maxAbsGrad - << osLrHistogram.str(); - } -} - -void TrainerInternal::createParameterUpdater(bool testing) { - const std::string& alg = config_->getOptConfig().algorithm(); - parameterUpdater_.reset(ParameterUpdaterCreators::tryCreateUpdater( - alg, config_->getOptConfig(), intconfig_->local, intconfig_->num_passes)); - if (parameterUpdater_) { - return; - } - - if (!intconfig_->local) { - if (testing && config_->getOptConfig().use_sparse_remote_updater()) { - std::unique_ptr localUpdater; - localUpdater.reset( - new SgdLocalUpdater(config_->getOptConfig())); // do nothing - parameterUpdater_.reset( - new SparseRemoteParameterUpdaterComposite(config_->getOptConfig(), - intconfig_->num_passes, - testing, - std::move(localUpdater))); - } else { - if (GradientMachine::kSgdSparseCpuTraining == intconfig_->mode && - !intconfig_->use_old_updater) { - intconfig_->use_old_updater = true; - LOG(INFO) << "Sgd sparse training can not work with" - << " ConcurrentRemoteParameterUpdater," - << " automatically reset --use_old_updater=true"; - } - - std::unique_ptr localUpdater; - if (config_->getOptConfig().num_batches_per_send_parameter() > 1) { - CHECK(alg == TrainAlgorithm::SGD || alg == TrainAlgorithm::AsyncSGD) - << "Unsupported algorithm in remote-local mode: " << alg; - if (GradientMachine::kSgdSparseCpuTraining == intconfig_->mode) { - localUpdater.reset(new SgdThreadUpdater(*config_)); - } else { - localUpdater.reset(new SgdLocalUpdater(*config_)); - } - } - - localUpdater.reset( - intconfig_->use_old_updater - ? new RemoteParameterUpdater( - *config_, intconfig_->num_passes, std::move(localUpdater)) - : new ConcurrentRemoteParameterUpdater( - *config_, intconfig_->num_passes, std::move(localUpdater))); - - if (config_->getOptConfig().use_sparse_remote_updater()) { - localUpdater.reset( - new SparseRemoteParameterUpdaterComposite(*config_, - intconfig_->num_passes, - testing, - std::move(localUpdater))); - } - - this->parameterUpdater_ = std::move(localUpdater); - } - } else { - CHECK_EQ(config_->getOptConfig().num_batches_per_send_parameter(), 1) - << "num_batches_per_send_parameter should be one in local mode!"; - - if (GradientMachine::kSgdSparseCpuTraining == intconfig_->mode) { - parameterUpdater_.reset(new SgdThreadUpdater(*config_)); - } else if (alg == TrainAlgorithm::SGD || alg == TrainAlgorithm::AsyncSGD) { - if (config_->getModelConfig().type() == "recursive_nn") { - parameterUpdater_.reset(new SgdCpuUpdater(*config_)); - } else if (intconfig_->use_gpu && - config_->getOptConfig().do_average_in_cpu() && - config_->getOptConfig().average_window() > 0) { - parameterUpdater_.reset(new SgdUpdaterWithCpuAverager(*config_)); - } else { - parameterUpdater_.reset(new SgdLocalUpdater(*config_)); - } - } else { - LOG(FATAL) << "Unsupported algorithm in local mode: " << alg; - } - } -} - -void TrainerInternal::forwardBackwardBatch(const std::vector& inArgs, - std::vector& outArgs, - PassType& passType, - UpdateCallback updateCallback, - bool doPipelineUpdate) { - gradientMachine_->forwardBackward( - inArgs, &outArgs, passType, doPipelineUpdate ? updateCallback : nullptr); -} - -} // namespace paddle diff --git a/paddle/trainer/TrainerInternal.h b/paddle/trainer/TrainerInternal.h deleted file mode 100644 index 48ee53a5e60f950bfc3cc299c754b0e72601c818..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerInternal.h +++ /dev/null @@ -1,139 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include -#include -#include - -#include "ParameterUpdater.h" -#include "TrainerConfig.pb.h" -#include "TrainerConfigHelper.h" -#include "TrainerInternalConfig.h" -#include "hl_gpu.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" - -namespace paddle { - -/** - * TrainerInteral - * the core training class for driving training logic - */ -class TrainerInternal { - public: - struct ParaStat { - real maxAbsGrad; - real avgAbsGrad; - ParaStat() : maxAbsGrad(.0), avgAbsGrad(.0) {} - }; - - TrainerInternal() {} - - /** - * Intializes trainer internal class - * @param config network config - * @param machine gradient machine - * @param intconfig training config - * @param stats training stats - * @param testing if it is in testing phase - */ - void init(const std::shared_ptr& config, - const GradientMachinePtr& machine, - std::unique_ptr&& intconfig, - const std::shared_ptr& stats, - bool testing); - - virtual ~TrainerInternal() {} - - /** - * CreateParameterUpdater - * @param testing if it is in testing phase - */ - void createParameterUpdater(bool testing); - - /** - * FinishTrainPass - * @param passId current pass id - * @param batchId current batch id, starts from 0 - */ - void finishTrainPass(int passId, int batchId); - - /** - * trainOneBatch - * @param batchId current batch id - * @param dataBatch data for the batch - */ - void trainOneBatch(int64_t batchId, - const DataBatch& dataBatch, - std::vector* outArgs); - - /** - * showParameterStats - * @param paraStats training stats - */ - void showParameterStats(const std::vector& paraStats); - - /** - * getGradientMachine - */ - inline const GradientMachinePtr& getGradientMachine() const { - return gradientMachine_; - } - - /** - * getParameterUpdater - */ - inline const std::shared_ptr& getParameterUpdater() { - return parameterUpdater_; - } - - /** - * setCurrentEvaluator - * @param eval evaluator to set - */ - inline void setCurrentEvaluator(Evaluator* eval) { currentEvaluator_ = eval; } - - /** - * setEvaluator - * @param eval evaluator to set - */ - inline void setEvaluator(Evaluator* eval) { evaluator_ = eval; } - - /** - * forwardBackwardBatch - * @param inArgs input argument for data batch - * @param outArgs output argument from neural network - * @param updateCallback layerwise parameter gradient statistics - * @param doPipelineUpdate whether to do pipeline update - */ - virtual void forwardBackwardBatch(const std::vector& inArgs, - std::vector& outArgs, - PassType& passType, - UpdateCallback updateCallback, - bool doPipelineUpdate); - - protected: - std::shared_ptr parameterUpdater_; - GradientMachinePtr gradientMachine_; - std::shared_ptr config_; - std::unique_ptr intconfig_; - std::shared_ptr stats_; - Evaluator* currentEvaluator_; - Evaluator* evaluator_; -}; - -} // namespace paddle diff --git a/paddle/trainer/TrainerInternalConfig.h b/paddle/trainer/TrainerInternalConfig.h deleted file mode 100644 index 43aae381029784278ad58c9398f64af24dffa1df..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerInternalConfig.h +++ /dev/null @@ -1,233 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once - -#include "paddle/utils/Util.h" - -#include - -#include "hl_gpu.h" -#include "paddle/gserver/gradientmachines/GradientMachine.h" - -#include "TrainerConfig.pb.h" - -#include -#include -#include -#include "ParameterUpdater.h" - -namespace paddle { -/** - * @brief TrainerStats object will statistics sample processed and total cost. - * - * There are two stats in it, the 'AvgCost' and 'CurrentAvgCost'. 'AvgCost' - * means cost through one pass(all mini-batches). 'CurrentAvgCost' means cost - * through one mini-batch. - */ -class TrainerStats { - public: - /** - * @brief reset all stats. - * - * often used before pass start. - */ - inline void reset() { - numProcessed_ = 0; - totalCost_ = .0; - this->resetCurrentStat(); - } - - /** - * @brief reset current stat. - * - * 'current' means the most recent --log_period mini-batches - */ - inline void resetCurrentStat() { - currentCost_ = .0; - currentSamples_ = 0; - } - - /** - * @brief add cost to stat. - * @param numProcessed current mini-batch size - * @param cost current mini-batch cost - */ - inline void addCost(int64_t numProcessed, real cost) { - this->numProcessed_ += numProcessed; - this->totalCost_ += cost; - this->currentSamples_ += numProcessed; - this->currentCost_ += cost; - } - - /** - * @brief get average cost through on pass(all processed mini-batches) - * @return pass average cost - */ - inline real getAvgCost() const { - CHECK_NE(this->numProcessed_, 0); - return this->totalCost_ / this->numProcessed_; - } - - /** - * @brief get current mini-batch's average cost. - * @return mini-batch average cost - */ - inline real getCurrentAvgCost() const { - CHECK_NE(this->currentSamples_, 0); - return this->currentCost_ / this->currentSamples_; - } - - /** - * @brief get all processed samples' number - * @return all processed samples' number - */ - inline int64_t getNumProcessed() const { return this->numProcessed_; } - - /** - * @brief same function as addCost. But it is simple to invoke. - * For example: - * - * @code{.cpp} - * TrainerStats stat; - * cost = neuralNetwork.forward(batchSize); - * stat += {batchSize, cost}; - * @endcode - * - * @param p a pair of parameter, first is numProcessed, second is cost. - * @return *this - */ - inline TrainerStats& operator+=(const std::pair& p) { - this->addCost(p.first, p.second); - return *this; - } - - /** - * @brief TrainerStats Constructor. - * - * reset stat when constructed. - */ - inline TrainerStats() { this->reset(); } - - /** - * @brief show stats to ostream. - * - * If there is no need to print current cost, set withCurrentCost to False. - * - * @param os output stream. - * @param withCurrentCost print current cost or not. - */ - void showStats(std::ostream& os, bool withCurrentCost = true) const { - os << "samples=" << this->getNumProcessed() - << " AvgCost=" << this->getAvgCost(); - if (withCurrentCost) { - os << " CurrentCost=" << this->getCurrentAvgCost(); - } - } - - /** - * @brief get stats to std::string - * @param withCurrentCost return current cost or not - * @return stats string - */ - std::string getStats(bool withCurrentCost = true) const { - std::ostringstream os; - this->showStats(os, withCurrentCost); - return os.str(); - } - - private: - int64_t numProcessed_; - real totalCost_; - real currentCost_; - int64_t currentSamples_; -}; - -inline std::ostream& operator<<(std::ostream& os, const TrainerStats& stats) { - stats.showStats(os); - return os; -} - -/** - * TrainerInternalConfig - * general configs for training - */ -struct TrainerInternalConfig { - /** - * @brief Create TrainerInternalConfig from GradientMachine::CreateMode and - * command line arguments. - * @param mode - * @return - */ - static std::unique_ptr createFromMode( - GradientMachine::CreateMode mode); - - /** - * indicate whether the training is local - * if local, no parameter server is used - */ - bool local; - - /** - * indicate whether training uses GPU - */ - bool use_gpu; - - /** - * indicate number of trainer - */ - int trainer_count; - - /** - * how frequently to show param stats - */ - int show_param_stats_period; - - /** - * current trainer id - */ - int trainer_id; - - /** - * frequency to dump log - */ - int log_period; - - /** - * dot period - */ - int dot_period; - - /** - * num passes for training - */ - int num_passes; - - /** - * use old updater - */ - bool use_old_updater; - - /** - * whether to load and save parameter in pserver - */ - bool loadsave_parameters_in_pserver; - - /** - * training mode - */ - GradientMachine::CreateMode mode; -}; - -} // namespace paddle diff --git a/paddle/trainer/TrainerMain.cpp b/paddle/trainer/TrainerMain.cpp deleted file mode 100644 index c5c1d484e5f85c774fd4b8f1d4a8d46abfa2f547..0000000000000000000000000000000000000000 --- a/paddle/trainer/TrainerMain.cpp +++ /dev/null @@ -1,65 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include "paddle/pserver/ParameterServerController.h" -#include "paddle/utils/PythonUtil.h" - -#include "ParamUtil.h" -#include "Trainer.h" - -DEFINE_bool(start_pserver, false, "Whether to start pserver"); -DECLARE_int32(gpu_id); -DEFINE_string(job, "train", "one of (train, test, checkgrad)"); -DECLARE_int32(start_pass); -DECLARE_string(config); -DECLARE_string(init_model_path); -DECLARE_string(rdma_tcp); - -using namespace paddle; // NOLINT - -int main(int argc, char** argv) { - // write logs instantly (never buffer log messages) - FLAGS_logbuflevel = -1; - - initMain(argc, argv); - initPython(argc, argv); - - std::unique_ptr parameterServerPtr(nullptr); - if (FLAGS_start_pserver) { - parameterServerPtr.reset( - paddle::ParameterServerController::createFromGflags()); - parameterServerPtr->start(); - } - Trainer trainer; - auto config = TrainerConfigHelper::createFromFlags(); - CHECK(config != nullptr) << "no valid config"; - - feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW); - trainer.init(config, FLAGS_job == "test"); - - if (FLAGS_job == "train") { - trainer.train(); - } else if (FLAGS_job == "checkgrad") { - trainer.checkGradient(); - } else if (FLAGS_job == "test") { - trainer.test(); - } else if (FLAGS_job == "time") { - trainer.time(); - } else { - LOG(FATAL) << "Unknown job type: " << FLAGS_job; - } - - return 0; -} diff --git a/paddle/trainer/tests/CMakeLists.txt b/paddle/trainer/tests/CMakeLists.txt deleted file mode 100644 index 12c9ea8cef79a6bdbd6e26c35612d0abbe00257b..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/CMakeLists.txt +++ /dev/null @@ -1,37 +0,0 @@ -add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/sample_trainer_config.conf - COMMAND cp -r ${CMAKE_CURRENT_SOURCE_DIR}/* ${CMAKE_CURRENT_BINARY_DIR} -) -add_custom_target(copy_trainer_conf ALL DEPENDS sample_trainer_config.conf) - -set(PYTHON_PATH - ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d - ${PADDLE_BINARY_DIR}/python/:${PADDLE_BINARY_DIR}/paddle/trainer/tests) -function(trainer_test TARGET) - add_unittest_without_exec(${TARGET} ${TARGET}.cpp) - add_test(NAME ${TARGET} - COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET} - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) -endfunction() - -trainer_test(test_Compare) -trainer_test(test_PyDataProviderWrapper) -trainer_test(test_recurrent_machine_generation) -trainer_test(test_Trainer) - -############### test_TrainerOnePass ########################## -if(WITH_PYTHON) - # only run test_TrainerOnePass when PYTHON is enabled, because train one pass - # is using PyDataProvider2. - add_unittest_without_exec(test_TrainerOnePass - test_TrainerOnePass.cpp) - add_test(NAME test_TrainerOnePass - COMMAND ${PYTHON_PATH} ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port - ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) -endif() - -#################### test_config_parser ######################### -add_test(NAME test_config_parser - COMMAND ${PYTHON_PATH} ${PYTHON_EXECUTABLE} - ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/config_parser_test.py - WORKING_DIRECTORY ${PADDLE_BINARY_DIR}/paddle/) diff --git a/paddle/trainer/tests/config_parser_test.py b/paddle/trainer/tests/config_parser_test.py deleted file mode 100644 index db66ebb5b7c13fe53df14a07918aad62ba895ffa..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/config_parser_test.py +++ /dev/null @@ -1,22 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer.config_parser import parse_config_and_serialize - -if __name__ == '__main__': - parse_config_and_serialize('trainer/tests/test_config.conf', '') - parse_config_and_serialize( - 'trainer/tests/sample_trainer_config.conf', - 'extension_module_name=paddle.trainer.config_parser_extension') - parse_config_and_serialize('gserver/tests/pyDataProvider/trainer.conf', '') diff --git a/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.list b/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.list deleted file mode 100644 index 0db50f34dd24b5e6fbc33a1e8dd3c16cb59eb56e..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.list +++ /dev/null @@ -1 +0,0 @@ -trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.data diff --git a/paddle/trainer/tests/sample_filelist.txt b/paddle/trainer/tests/sample_filelist.txt deleted file mode 100644 index 7db4c735359a380dc150e24368653d2a6a55a453..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/sample_filelist.txt +++ /dev/null @@ -1 +0,0 @@ -trainer/tests/sample_data.txt diff --git a/paddle/trainer/tests/sample_trainer_config.conf b/paddle/trainer/tests/sample_trainer_config.conf deleted file mode 100644 index 2697832840f35a33c07f1664ef18a229d656d784..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/sample_trainer_config.conf +++ /dev/null @@ -1,87 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -TrainData(SimpleData( - files = "trainer/tests/sample_filelist.txt", - feat_dim = 3, - context_len = 0, - buffer_capacity = 1000000)) - -TestData(SimpleData( - files = "trainer/tests/sample_filelist.txt", - feat_dim = 3, - context_len = 0, - buffer_capacity = 1000000)) - -settings(batch_size = 100) - -data = data_layer(name='input', size=3) - -fc1 = fc_layer(input=data, size=5, - bias_attr=False, - act=SigmoidActivation()) - -fc2 = fc_layer(input=data, size=9, - bias_attr=False, - act=LinearActivation()) - -fc3 = fc_layer(input=data, size=3, - bias_attr=False, - act=TanhActivation()) - -fc4 = fc_layer(input=data, size=5, - bias_attr=False, - act=LinearActivation(), - param_attr=ParamAttr(name='sharew')) - -fc5 = fc_layer(input=data, size=5, - bias_attr=False, - act=BReluActivation()) - -fc6 = fc_layer(input=data, size=5, - bias_attr=False, - act=SoftReluActivation()) - -fc7 = fc_layer(input=data, size=3, - bias_attr=False, - act=SquareActivation()) - -fc8 = fc_layer(input=data, size=5, - bias_attr=True, - act=SquareActivation()) - -with mixed_layer(size=3, act=SoftmaxActivation()) as layer9: - layer9 += full_matrix_projection(input=fc1) - layer9 += full_matrix_projection(input=fc2) - layer9 += full_matrix_projection(input=fc3) - layer9 += trans_full_matrix_projection(input=fc4, - param_attr=ParamAttr(name='sharew')) - layer9 += full_matrix_projection(input=fc5) - layer9 += full_matrix_projection(input=fc6) - layer9 += full_matrix_projection(input=fc7) - layer9 += full_matrix_projection(input=fc8) - -if get_config_arg('with_cost', bool, True): - # This is for training the neural network. - # We need to have another data layer for label - # and a layer for calculating cost - lbl = data_layer(name='label', size=1) - outputs(classification_cost(input=layer9, label=lbl)) -else: - # This is for prediction where we don't have label - # and don't need to calculate cost - outputs(layer9) diff --git a/paddle/trainer/tests/sample_trainer_config_hsigmoid.conf b/paddle/trainer/tests/sample_trainer_config_hsigmoid.conf deleted file mode 100644 index e4abe31d480b69bc2ff4741649b336714818515b..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/sample_trainer_config_hsigmoid.conf +++ /dev/null @@ -1,53 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -from paddle.trainer_config_helpers import * - -TrainData(SimpleData( - files = "trainer/tests/sample_filelist.txt", - feat_dim = 3, - context_len = 0, - buffer_capacity = 1000000, -)) - -settings(batch_size = 100) - -data = data_layer(name='input', size=3) - -fc1 = fc_layer(input=data, size=12, - bias_attr=False, - act=SigmoidActivation()) - -fc2 = fc_layer(input=data, size=19, - bias_attr=False, - act=LinearActivation()) - -fc3 = fc_layer(input=data, size=5, - bias_attr=False, - act=TanhActivation()) - -fc4 = fc_layer(input=data, size=5, - bias_attr=False, - act=LinearActivation()) - -# This is for training the neural network. -# We need to have another data layer for label -# and a layer for calculating cost -lbl = data_layer(name='label', size=1) - -outputs(hsigmoid(input=[fc1, fc2, fc3, fc4], - label=lbl, - num_classes=3)) diff --git a/paddle/trainer/tests/sample_trainer_config_parallel.conf b/paddle/trainer/tests/sample_trainer_config_parallel.conf deleted file mode 100644 index e2b8b3ecdab83b4614dbe468c3a295c05867f7f9..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/sample_trainer_config_parallel.conf +++ /dev/null @@ -1,86 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -TrainData(SimpleData( - files = "trainer/tests/sample_filelist.txt", - feat_dim = 3, - context_len = 0, - buffer_capacity = 1000000)) - -TestData(SimpleData( - files = "trainer/tests/sample_filelist.txt", - feat_dim = 3, - context_len = 0, - buffer_capacity = 1000000)) - -settings(batch_size = 100) - -# Output layer, label layer, cost layer, preferably set to the same environment. -output_device = 0 - -# Input Layer does not need to specify the device number. -data = data_layer(name='input', size=3) - -# Calculate in the CPU. -fc1 = fc_layer(input=data, size=5, - bias_attr=True, - layer_attr=ExtraAttr(device=-1), - act=SigmoidActivation()) - -# Calculate in the GPU 0. -fc2 = fc_layer(input=fc1, size=10, - bias_attr=True, - layer_attr=ExtraAttr(device=0), - act=SigmoidActivation()) - -# Calculate in the GPU 1. -fc3 = fc_layer(input=fc1, size=10, - bias_attr=True, - layer_attr=ExtraAttr(device=1), - act=SigmoidActivation()) - -# Calculate in the GPU 0. -fc4 = fc_layer(input=[fc2,fc3], size=10, - bias_attr=True, - layer_attr=ExtraAttr(device=0), - act=SigmoidActivation()) - -# Calculate in the GPU 1. -fc5 = fc_layer(input=[fc2,fc3], size=10, - bias_attr=True, - layer_attr=ExtraAttr(device=1), - act=SigmoidActivation()) - -output = fc_layer(input=[fc4,fc5], size=10, - bias_attr=True, - layer_attr=ExtraAttr(device=output_device), - act=SoftmaxActivation()) - -if get_config_arg('with_cost', bool, True): - # This is for training the neural network. - # We need to have another data layer for label - # and a layer for calculating cost - lbl = data_layer(name='label', size=1, - layer_attr=ExtraAttr(device=output_device)) - - outputs(classification_cost(input=output, - label=lbl, - layer_attr=ExtraAttr(device=output_device))) -else: - # This is for prediction where we don't have label - # and don't need to calculate cost - outputs(output) diff --git a/paddle/trainer/tests/sample_trainer_nest_rnn_gen.conf b/paddle/trainer/tests/sample_trainer_nest_rnn_gen.conf deleted file mode 100644 index 741a0aa71df7866c180ab2513f28638117d0f1ca..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/sample_trainer_nest_rnn_gen.conf +++ /dev/null @@ -1,73 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -from paddle.trainer_config_helpers import * - -settings(batch_size=15, learning_rate=0) - -num_words = 5 -beam_flag = get_config_arg('beam_search', bool, False) - -sent_id = data_layer(name="sent_id", size=1) - -# This layer has no actual use, but only to decide batch_size in generation. -# When generating, at least one Memory in RecurrentLayer MUST have a boot layer. -dummy_data = data_layer(name="dummy_data_input", size=2) - -def outer_step(dummy_data): - - gen_inputs = [StaticInput(input=dummy_data, size=2, is_seq=True), - GeneratedInput(size=num_words, - embedding_name="wordvec", - embedding_size=num_words)] - - def inner_step(dummy_memory, predict_word): - - # simplified RNN for testing - with mixed_layer(size=num_words) as layer: - layer += full_matrix_projection(input=predict_word, - param_attr=ParamAttr(name="transtable")) - - with mixed_layer(size=num_words, act=ExpActivation()) as out: - out += trans_full_matrix_projection(input=layer, - param_attr=ParamAttr(name="wordvec")) - - return out - - beam_gen = beam_search(name="rnn_gen", - step=inner_step, - input=gen_inputs, - bos_id=0, - eos_id=num_words-1, - beam_size=2 if beam_flag else 1, - num_results_per_sample=1, - max_length=10) - return beam_gen - -beam_gen_concat = recurrent_group(name="rnn_gen_concat", - step=outer_step, - input=[SubsequenceInput(dummy_data)]) - -seqtext_printer_evaluator(input=beam_gen_concat, - id_input=sent_id, - dict_file="./trainer/tests/test_gen_dict.txt", - result_file="./trainer/tests/dump_text.test") -#outputs(beam_gen_concat) -# In this config, as dummy_data_input doesn't work on beam_gen (we can find dummy_memory -# is read-only memory, and isn't used by other layers of step), we show the Inputs and Outputs -# as follows. Note that "__beam_search_predict__" is the default output name of beam_search. -Inputs("sent_id","dummy_data_input") -Outputs("__beam_search_predict__") diff --git a/paddle/trainer/tests/sample_trainer_rnn_gen.conf b/paddle/trainer/tests/sample_trainer_rnn_gen.conf deleted file mode 100644 index 58d27f15ae1c0a38885ee105a7963b6e7bd55906..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/sample_trainer_rnn_gen.conf +++ /dev/null @@ -1,66 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -from paddle.trainer_config_helpers import * - -settings(batch_size=15, learning_rate=0) - -num_words = 5 -beam_flag = get_config_arg('beam_search', bool, False) - -sent_id = data_layer(name="sent_id", size=1) - -# This layer has no actual use, but only to decide batch_size in generation. -# When generating, at least one Memory in RecurrentLayer MUST have a boot layer. -dummy_data = data_layer(name="dummy_data_input", size=2) - -gen_inputs = [StaticInput(input=dummy_data, size=2), - GeneratedInput(size=num_words, - embedding_name="wordvec", - embedding_size=num_words)] - -def step(dummy_memory, predict_word): - - # simplified RNN for testing - with mixed_layer(size=num_words) as layer: - layer += full_matrix_projection(input=predict_word, - param_attr=ParamAttr(name="transtable")) - - with mixed_layer(size=num_words, act=ExpActivation()) as out: - out += trans_full_matrix_projection(input=layer, - param_attr=ParamAttr(name="wordvec")) - - return out - -beam_gen = beam_search(name="rnn_gen", - step=step, - input=gen_inputs, - bos_id=0, - eos_id=num_words-1, - beam_size=2 if beam_flag else 1, - num_results_per_sample=2 if beam_flag else 1, - max_length=10) - -seqtext_printer_evaluator(input=beam_gen, - id_input=sent_id, - dict_file="./trainer/tests/test_gen_dict.txt", - result_file="./trainer/tests/dump_text.test") -#outputs(beam_gen) -# In this config, as dummy_data_input doesn't work on beam_gen (we can find dummy_memory -# is read-only memory, and isn't used by other layers of step), we show the Inputs and Outputs -# as follows. Note that "__beam_search_predict__" is the default output name of beam_search. -Inputs("sent_id","dummy_data_input") -Outputs("__beam_search_predict__") diff --git a/paddle/trainer/tests/simple_sparse_neural_network.py b/paddle/trainer/tests/simple_sparse_neural_network.py deleted file mode 100644 index 970fb466dc5061713fe7815d5247cbbde93be821..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/simple_sparse_neural_network.py +++ /dev/null @@ -1,37 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -settings(batch_size=17, learning_method=AdaGradOptimizer(), learning_rate=1e-4) - -file_list = 'trainer/tests/fake_file_list.list' - -define_py_data_sources2( - train_list=file_list, - test_list=file_list, - module="simple_sparse_neural_network_dp", - obj="process") - -embedding = embedding_layer( - input=data_layer( - name="word_ids", size=8191), - size=128, - param_attr=ParamAttr(sparse_update=True)) -prediction = fc_layer(input=embedding, size=10, act=SoftmaxActivation()) - -outputs( - classification_cost( - input=prediction, label=data_layer( - name='label', size=10))) diff --git a/paddle/trainer/tests/test_Compare.cpp b/paddle/trainer/tests/test_Compare.cpp deleted file mode 100644 index f3a964acb69be059a43470f7b68910a3b6cecaab..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_Compare.cpp +++ /dev/null @@ -1,157 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "paddle/trainer/Trainer.h" - -#include -#include - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static const string& configFile = "trainer/tests/sample_trainer_config.conf"; - -DECLARE_int32(gpu_id); -DECLARE_bool(use_gpu); -DECLARE_string(config); -DECLARE_string(config_args); - -struct comData { - vector outArgs; - vector parameters; -}; - -void calcGradient(bool useGpu, comData& Data) { - FLAGS_use_gpu = useGpu; - FLAGS_config = configFile; - - *ThreadLocalRand::getSeed() = 0; - srand(0); - Trainer trainer; - trainer.init(TrainerConfigHelper::createFromFlagConfig()); - - Data.parameters = trainer.getGradientMachine()->getParameters(); - DataBatch dataBatch; - int32_t batchSize = trainer.getConfig().opt_config().batch_size(); - trainer.getDataProvider()->setSkipShuffle(); - trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch); - CHECK(dataBatch.getSize()) << "No data from data provider"; - vector& inArgs = dataBatch.getStreams(); - trainer.getGradientMachine()->start(); - for (int i = 0; i < 2; ++i) { - trainer.getGradientMachine()->forwardBackward( - inArgs, &Data.outArgs, PASS_TRAIN); - } - trainer.getGradientMachine()->finish(); -} - -void compareGradient(comData& comDataCpu, comData& comDataGpu); - -TEST(Trainer, create) { - int devCount = 0; - devCount = hl_get_device_count(); - FLAGS_config_args = "drop_rate=0"; - - comData comDataCpu; - calcGradient(false, comDataCpu); - LOG(INFO) << "Cpu is completed"; - - { - LOG(INFO) << "Test GPU"; - comData comData; - calcGradient(true, comData); - compareGradient(comDataCpu, comData); - LOG(INFO) << "Gpu is completed"; - } - - { - LOG(INFO) << "Test test multi gpu"; - comData comData; - FLAGS_trainer_count = devCount; - calcGradient(true, comData); - compareGradient(comDataCpu, comData); - LOG(INFO) << "Gpu4 is completed"; - } - - { - LOG(INFO) << "Test use_sparse_update=true"; - comData comData; - calcGradient(false, comData); - compareGradient(comDataCpu, comData); - LOG(INFO) << "Cpu4 is completed"; - } -} - -double checkBuffer(real* A, real* B, size_t len) { -#ifdef PADDLE_TYPE_DOUBLE - double precision = 1e-7; -#else - double precision = 2e-3; -#endif - int nNum = 0; - double maxE = 0; - for (size_t i = 0; i < len; ++i) { - double e = fabs(A[i] - B[i]); - maxE = std::max(e, maxE); - nNum += e > precision * fabs(A[i]); - } - EXPECT_EQ(0, nNum); - return maxE; -} - -void compareGradient(comData& comDataCpu, comData& comDataGpu) { - /*compare outArgs*/ - vector outArgs1 = comDataCpu.outArgs; - vector outArgs2 = comDataGpu.outArgs; - CpuMatrix out1(outArgs1[0].value->getHeight(), outArgs1[0].value->getWidth()); - CpuMatrix out2(outArgs2[0].value->getHeight(), outArgs2[0].value->getWidth()); - out1.copyFrom(*outArgs1[0].value); - out2.copyFrom(*outArgs2[0].value); - checkBuffer(out1.getData(), out2.getData(), out1.getElementCnt()); - - /*compare parameters*/ - vector& parameters1 = comDataCpu.parameters; - vector& parameters2 = comDataGpu.parameters; - for (size_t i = 0; i < parameters1.size(); ++i) { - ParameterPtr parameter1, parameter2; - parameter1 = parameters1[i]; - parameter2 = parameters2[i]; - /*compare parameters value*/ - CpuVector para1(parameter1->getSize()); - CpuVector para2(parameter2->getSize()); - para1.copyFrom(*parameter1->getBuf(PARAMETER_VALUE)); - para2.copyFrom(*parameter2->getBuf(PARAMETER_VALUE)); - checkBuffer(para1.getData(), para2.getData(), para1.getSize()); - - /*compare parameters grad*/ - CpuVector cpuGrad1(*parameter1->getBuf(PARAMETER_GRADIENT)); - CpuVector cpuGrad2(*parameter2->getBuf(PARAMETER_GRADIENT)); - double e = - checkBuffer(cpuGrad1.getData(), cpuGrad2.getData(), cpuGrad1.getSize()); - LOG(INFO) << parameter1->getName() << " max error=" << e; - } -} - -int main(int argc, char** argv) { -#ifndef PADDLE_WITH_CUDA - exit(0); -#endif - paddle::initMain(argc, argv); - testing::InitGoogleTest(&argc, argv); - initPython(argc, argv); - int ret = RUN_ALL_TESTS(); - exit(ret); -} diff --git a/paddle/trainer/tests/test_PyDataProviderWrapper.cpp b/paddle/trainer/tests/test_PyDataProviderWrapper.cpp deleted file mode 100644 index 92dc8aa9ec5ce281d1950d84260c1b9555e686a7..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_PyDataProviderWrapper.cpp +++ /dev/null @@ -1,220 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifndef PADDLE_NO_PYTHON -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include "picojson.h" - -void checkValue(std::vector& arguments, picojson::array& arr); -const std::string kDir = "./trainer/tests/pydata_provider_wrapper_dir/"; - -TEST(PyDataProviderWrapper, SequenceData) { - paddle::DataConfig conf; - conf.set_type("py"); - conf.set_load_data_module("testPyDataWrapper"); - conf.set_load_data_object("processSeqAndGenerateData"); - conf.set_load_data_args(kDir + "test_pydata_provider_wrapper.json"); - conf.clear_files(); - conf.set_files(kDir + "test_pydata_provider_wrapper.list"); - paddle::DataProviderPtr provider(paddle::DataProvider::create(conf, false)); - provider->setSkipShuffle(); - provider->reset(); - paddle::DataBatch batchFromPy; - provider->getNextBatch(100, &batchFromPy); - - picojson::value val; - std::fstream fin; - fin.open(kDir + "test_pydata_provider_wrapper.json", std::ios_base::in); - EXPECT_TRUE(fin.is_open()); - if (fin.is_open()) { - std::string err = picojson::parse(val, fin); - EXPECT_TRUE(err.empty()); - EXPECT_TRUE(val.is()); - picojson::array& arr = val.get(); - std::vector& arguments = batchFromPy.getStreams(); - // CHECK Value - checkValue(arguments, arr); - // CHECK sequenceStartPositions - for (size_t i = 0; i < arr.size(); i++) { - int row_id = arr[i].get().size(); - EXPECT_EQ(0, arguments[i].sequenceStartPositions->getData(false)[0]); - EXPECT_EQ((int)row_id, - arguments[i].sequenceStartPositions->getData(false)[1]); - } - fin.close(); - } -} - -TEST(PyDataProviderWrapper, HasSubSequenceData) { - paddle::DataConfig conf; - conf.set_type("py"); - conf.set_load_data_module("testPyDataWrapper"); - conf.set_load_data_object("processSubSeqAndGenerateData"); - conf.set_load_data_args(kDir + "test_pydata_provider_wrapper.json"); - conf.clear_files(); - conf.set_files(kDir + "test_pydata_provider_wrapper.list"); - paddle::DataProviderPtr provider(paddle::DataProvider::create(conf, false)); - provider->setSkipShuffle(); - provider->reset(); - paddle::DataBatch batchFromPy; - provider->getNextBatch(1, &batchFromPy); - - picojson::value val; - std::fstream fin; - fin.open(kDir + "test_pydata_provider_wrapper.json", std::ios_base::in); - EXPECT_TRUE(fin.is_open()); - if (fin.is_open()) { - std::string err = picojson::parse(val, fin); - EXPECT_TRUE(err.empty()); - EXPECT_TRUE(val.is()); - picojson::array& arr = val.get(); - std::vector& arguments = batchFromPy.getStreams(); - // CHECK Value - checkValue(arguments, arr); - // CHECK sequenceStartPositions and subSequenceStartPositions - for (size_t i = 0; i < arr.size(); i++) { - int row_id = arr[i].get().size(); - EXPECT_EQ(0, arguments[i].sequenceStartPositions->getData(false)[0]); - EXPECT_EQ((int)row_id, - arguments[i].sequenceStartPositions->getData(false)[1]); - EXPECT_EQ(0, arguments[i].subSequenceStartPositions->getData(false)[0]); - EXPECT_EQ((int)row_id, - arguments[i].subSequenceStartPositions->getData(false)[1]); - } - fin.close(); - } -} - -int main(int argc, char** argv) { - paddle::initMain(argc, argv); - paddle::initPython(argc, argv); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} - -void checkValue(std::vector& arguments, - picojson::array& arr) { - // CHECK SLOT 0, Sparse Value. - paddle::Argument& sparse_values_seq = arguments[0]; - paddle::MatrixPtr& sparse_values_seq_rawmatrix = sparse_values_seq.value; - EXPECT_TRUE(sparse_values_seq_rawmatrix != nullptr); - paddle::CpuSparseMatrix* sparse_val_seq_sparse_mat = - dynamic_cast(sparse_values_seq_rawmatrix.get()); - EXPECT_TRUE(sparse_val_seq_sparse_mat != nullptr); - EXPECT_EQ(arr.size(), arguments.size()); - EXPECT_TRUE(arr[0].is()); - size_t row_id = 0; - for (picojson::value& sparse_val_seq : arr[0].get()) { - std::unordered_map cols; - for (picojson::value& kv : sparse_val_seq.get()) { - EXPECT_TRUE(kv.get(0).is()); - EXPECT_TRUE(kv.get(1).is()); - int col = (int)(kv.get(0).get()); - real val = (real)(kv.get(1).get()); - cols.insert({col, val}); - } - size_t colNum = sparse_val_seq_sparse_mat->getColNum(row_id); - EXPECT_EQ(cols.size(), colNum); - int* rowIds = sparse_val_seq_sparse_mat->getRowCols(row_id); - real* rowBuf = sparse_val_seq_sparse_mat->getRowValues(row_id); - for (size_t i = 0; i < colNum; ++i) { - int id = rowIds[i]; - auto it = cols.find(id); - EXPECT_NE(cols.end(), it); - real expect = it->second; - EXPECT_NEAR(expect, *rowBuf, 1e-5); - ++rowBuf; - } - ++row_id; - } - - // CHECK SLOT 1, Dense Value. - paddle::Argument& dense_arg = arguments[1]; - paddle::MatrixPtr& dense_mat = dense_arg.value; - EXPECT_NE(nullptr, dense_mat); - EXPECT_TRUE(arr[1].is()); - row_id = 0; - for (picojson::value& dense_seq : arr[1].get()) { - EXPECT_TRUE(dense_seq.is()); - picojson::array& row = dense_seq.get(); - EXPECT_EQ(row.size(), dense_mat->getWidth()); - real* rowBuf = dense_mat->getRowBuf(row_id++); - - for (picojson::value& val : row) { - EXPECT_TRUE(val.is()); - real expect = val.get(); - EXPECT_NEAR(expect, *rowBuf++, 1e-5); - } - } - - // CHECK SLOT 2, Sparse Non Value. - paddle::Argument& sparse_non_val_arg = arguments[2]; - paddle::MatrixPtr& sparse_non_val_rawm = sparse_non_val_arg.value; - EXPECT_NE(nullptr, sparse_non_val_rawm); - paddle::CpuSparseMatrix* sparse_non_val_m = - dynamic_cast(sparse_non_val_rawm.get()); - EXPECT_NE(nullptr, sparse_non_val_m); - row_id = 0; - for (picojson::value& row : arr[2].get()) { - EXPECT_TRUE(row.is()); - std::unordered_set ids; - for (picojson::value& id : row.get()) { - EXPECT_TRUE(id.is()); - ids.insert((int)(id.get())); - } - size_t colNum = sparse_non_val_m->getColNum(row_id); - EXPECT_EQ(ids.size(), colNum); - for (size_t i = 0; i < colNum; ++i) { - int col = sparse_non_val_m->getRowCols(row_id)[i]; - EXPECT_TRUE(ids.find(col) != ids.end()); - } - ++row_id; - } - - // CHECK SLOT 3, Index. - paddle::Argument& index_arg = arguments[3]; - paddle::IVectorPtr indices = index_arg.ids; - EXPECT_NE(nullptr, indices); - int* idPtr = indices->getData(); - for (picojson::value& id : arr[3].get()) { - EXPECT_TRUE(id.is()); - int _id = (int)(id.get()); - EXPECT_EQ(_id, *idPtr++); - } - - // CHECK SLOT 4, String. - paddle::Argument& strArg = arguments[4]; - std::vector* strPtr = strArg.strs.get(); - EXPECT_NE(nullptr, strPtr); - size_t vecIndex = 0; - for (picojson::value& str : arr[4].get()) { - EXPECT_TRUE(str.is()); - std::string _str = str.get(); - EXPECT_EQ(_str, (*strPtr)[vecIndex++]); - } -} - -#else -int main() { return 0; } - -#endif diff --git a/paddle/trainer/tests/test_Trainer.cpp b/paddle/trainer/tests/test_Trainer.cpp deleted file mode 100644 index 394038cf730f13cb957fbbc5ae0e5719b8fe9db6..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_Trainer.cpp +++ /dev/null @@ -1,106 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "paddle/trainer/Trainer.h" - -#include - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static const string& configFile1 = "trainer/tests/sample_trainer_config.conf"; -static const string& configFile2 = - "trainer/tests/sample_trainer_config_hsigmoid.conf"; -static const string& configFile4 = - "trainer/tests/sample_trainer_config_parallel.conf"; - -DECLARE_bool(use_gpu); -DECLARE_string(config); -DECLARE_int32(gpu_id); -DECLARE_bool(allow_only_one_model_on_one_gpu); - -void checkGradientTest(const string& configFile, - bool useGpu, - bool parallel, - int trainerCount = 1) { - FLAGS_use_gpu = useGpu; - FLAGS_parallel_nn = parallel; - FLAGS_config = configFile; - FLAGS_trainer_count = trainerCount; - LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount - << " configFile=" << configFile; - - Trainer trainer; - trainer.init(TrainerConfigHelper::createFromFlagConfig()); - EXPECT_LE(fabs(trainer.checkGradient()), 0.02); -} - -TEST(checkGradient, cpu) { checkGradientTest(configFile1, false, false); } - -#ifdef PADDLE_WITH_CUDA -TEST(checkGradient, gpu) { checkGradientTest(configFile1, true, false); } - -TEST(checkGradient, multiGpu) { - int numGpu; - numGpu = hl_get_device_count(); - for (auto count : {2, 4}) { - if (count <= numGpu) { - checkGradientTest(configFile1, true, false, count); - } - } -} - -TEST(checkGradient, parallel) { - if (hl_get_device_count() >= 2) { - checkGradientTest(configFile4, true, true); - } -} - -TEST(checkGradient, multiParallel) { - FLAGS_allow_only_one_model_on_one_gpu = false; - checkGradientTest(configFile4, true, true, 2); - FLAGS_allow_only_one_model_on_one_gpu = true; -} - -#endif - -TEST(checkGradient, multi) { - int numGpu; - if (version::isWithGpu()) { - numGpu = hl_get_device_count(); - } else { - numGpu = 0; - } - for (bool useGpu : {false, true}) { - for (auto count : {2, 4}) { - if (useGpu && count > numGpu) continue; - checkGradientTest(configFile1, useGpu, false, count); - } - } -} - -TEST(checkGradient, hsigmoid) { checkGradientTest(configFile2, false, false); } - -TEST(checkGradient, non_parallel) { - checkGradientTest(configFile4, false, false); -} - -int main(int argc, char** argv) { - initMain(argc, argv); - initPython(argc, argv); - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/trainer/tests/test_TrainerOnePass.cpp b/paddle/trainer/tests/test_TrainerOnePass.cpp deleted file mode 100644 index de12c4d649c6041f497c0eeac0904ebfc0d5bf97..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_TrainerOnePass.cpp +++ /dev/null @@ -1,317 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include "paddle/trainer/Trainer.h" -#include "paddle/trainer/TrainerInternal.h" - -#include -#include - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static const string& configFile1 = "trainer/tests/sample_trainer_config.conf"; -static const string& configFile2 = - "trainer/tests/sample_trainer_config_parallel.conf"; - -static const string& configFileSimpleSparse = - "trainer/tests/simple_sparse_neural_network.py"; - -DECLARE_bool(use_gpu); -DECLARE_string(config); -DECLARE_int32(gpu_id); -DECLARE_int32(seed); -DECLARE_int32(num_passes); -DECLARE_int32(saving_period); - -class TrainerForTest : public paddle::Trainer { - public: - inline const std::shared_ptr& getParameterUpdaterForTest() { - return this->trainerInternal_.getParameterUpdater(); - } -}; - -int gNumDevices = 0; - -void trainerOnePassTest(const string& configFile, - bool useGpu, - bool parallel, - int trainerCount = 1, - double averageWindow = 0.0f, - bool doAverageInCpu = false) { - FLAGS_use_gpu = useGpu; - FLAGS_parallel_nn = parallel; - FLAGS_config = configFile; - FLAGS_trainer_count = trainerCount; - LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount - << " configFile=" << configFile; - srand(FLAGS_seed); - - if (useGpu) { - if (gNumDevices < trainerCount) { - return; - } - } - - Trainer trainer; - auto config = TrainerConfigHelper::createFromFlagConfig(); - if (averageWindow > 0) { - config->getOptConfig().set_average_window(averageWindow); - config->getOptConfig().set_do_average_in_cpu(doAverageInCpu); - } - trainer.init(config); - trainer.train(); -} - -// 1. test trainer (cpu, gpu). -TEST(trainerOnePass, cpu) { trainerOnePassTest(configFile1, false, false); } - -#ifdef PADDLE_WITH_CUDA -TEST(trainerOnePass, gpu) { trainerOnePassTest(configFile1, true, false); } - -TEST(trainerOnePass, gpu2) { trainerOnePassTest(configFile1, true, false, 2); } - -TEST(trainerOnePass, gpu4) { trainerOnePassTest(configFile1, true, false, 4); } - -TEST(trainerOnePass, parallel) { - if (hl_get_device_count() >= 2) { - trainerOnePassTest(configFile2, true, true); - } -} -#endif - -// 2. test average_window. -#ifdef PADDLE_WITH_CUDA -TEST(average_window, gpu) { - trainerOnePassTest(configFile1, true, false, 4, 0.01); -} - -TEST(average_window, gpu2) { - FLAGS_num_passes = 20; - trainerOnePassTest(configFile1, true, false, 2, 0.01); - FLAGS_num_passes = 1; -} - -TEST(average_window, gpu4) { - FLAGS_num_passes = 20; - trainerOnePassTest(configFile1, true, false, 4, 0.01); - FLAGS_num_passes = 1; -} - -TEST(average_window_cpu, gpu2) { - FLAGS_num_passes = 20; - trainerOnePassTest(configFile1, true, false, 2, 0.01, true); - FLAGS_num_passes = 1; -} - -TEST(average_window_cpu, gpu4) { - FLAGS_num_passes = 20; - trainerOnePassTest(configFile1, true, false, 4, 0.01, true); - FLAGS_num_passes = 1; -} -#endif - -// 3. test trainer + pserver. -DECLARE_int32(num_gradient_servers); -DECLARE_int32(port); -DECLARE_bool(local); -DECLARE_bool(use_old_updater); - -double checkRemoteParameterUpdater(TrainerForTest& trainer) { - auto gradientMachine = trainer.getGradientMachine(); - auto parameterUpdater = trainer.getParameterUpdaterForTest(); - auto dataProvider = trainer.getDataProvider(); - auto& parameters = gradientMachine->getParameters(); - const TrainerConfig& config = trainer.getConfig(); - const string& alg = config.opt_config().algorithm(); - - vector parameterCheck; - for (auto& parameter : parameters) { - parameterCheck.emplace_back( - new Parameter(parameter->getConfig(), /* useGpu= */ false)); - parameterCheck.back() - ->getBuf(PARAMETER_VALUE) - ->copyFrom(*parameter->getBuf(PARAMETER_VALUE)); - parameterCheck.back() - ->getBuf(PARAMETER_GRADIENT) - ->copyFrom(*parameter->getBuf(PARAMETER_GRADIENT)); - } - - std::unique_ptr parameterUpdaterCheck; - if (alg == TrainAlgorithm::SGD) { - parameterUpdaterCheck.reset(new SgdLocalUpdater(config.opt_config())); - } else { - LOG(INFO) << "unsupported algorithm in remote parameter check: " << alg; - return -1.0; - } - parameterUpdaterCheck->init(parameterCheck); - - // gradientMachine->start(config, *dataProvider); - DataBatch dataBatch; - int32_t batchSize = config.opt_config().batch_size(); - dataProvider->getNextBatch(batchSize, &dataBatch); - CHECK(dataBatch.getSize()) << "No data from data provider"; - int64_t actualBatchSize = dataBatch.getSize(); - const vector& inArgs = dataBatch.getStreams(); - vector outArgs; - - UpdateCallback updateCallback = [parameterUpdater, - parameterCheck](Parameter* para) { - parameterCheck[para->getID()] - ->getBuf(PARAMETER_GRADIENT) - ->copyFrom(*para->getBuf(PARAMETER_GRADIENT)); - parameterUpdater->update(para); - }; - - parameterUpdater->startPass(); - parameterUpdaterCheck->startPass(); - - for (int i = 0; i < config.opt_config().num_batches_per_get_parameter() * 2; - ++i) { - PassType passType = parameterUpdater->startBatch(actualBatchSize); - gradientMachine->forwardBackward( - inArgs, &outArgs, passType, updateCallback); - parameterUpdater->finishBatch(0); - - parameterUpdaterCheck->startBatch(actualBatchSize); - for (auto& para : parameterCheck) { - parameterUpdaterCheck->update(para.get()); - } - parameterUpdaterCheck->finishBatch(0); - } - - double sum = 0.0f; - for (size_t i = 0; i != parameters.size(); ++i) { - real *v1, *v2; - CpuVector trainerPara(parameters[i]->getSize()); - trainerPara.copyFrom(*parameters[i]->getBuf(PARAMETER_VALUE)); - if (!FLAGS_use_gpu) { - v1 = parameters[i]->getBuf(PARAMETER_VALUE)->getData(); - } else { - v1 = trainerPara.getData(); - } - v2 = parameterCheck[i]->getBuf(PARAMETER_VALUE)->getData(); - - size_t size = parameters[i]->getSize(); - double diff = 0; - for (size_t j = 0; j < size; ++j) { - diff += fabs(v1[j] - v2[j]); - } - sum += diff; - LOG(INFO) << setiosflags(ios::left) << setfill(' ') << setw(20) - << parameters[i]->getName() << "diff=" << setw(15) << diff; - } - - parameterUpdater->finishPass(); - parameterUpdaterCheck->finishPass(); - gradientMachine->finish(); - return sum; -} - -void checkRemoteParameterUpdaterTest(const string& configFile, - bool useGpu, - bool parallel, - int trainerCount = 1, - bool useOldUpdater = false, - int num_batches_per_get_parameter = 1) { - FLAGS_use_gpu = useGpu; - FLAGS_parallel_nn = parallel; - FLAGS_config = configFile; - FLAGS_trainer_count = trainerCount; - FLAGS_use_old_updater = useOldUpdater; - LOG(INFO) << " useGpu=" << useGpu << " trainerCount=" << trainerCount - << " configFile=" << configFile; - srand(FLAGS_seed); - - if (useGpu) { - if (gNumDevices < trainerCount) { - return; - } - } - - FLAGS_local = 0; - std::shared_ptr pserver; - pserver.reset(new ParameterServer2(std::string(), FLAGS_port)); - pserver->init(); - pserver->start(); - - TrainerForTest trainer; - auto config = TrainerConfigHelper::createFromFlagConfig(); - config->getOptConfig().set_num_batches_per_get_parameter( - num_batches_per_get_parameter); - trainer.init(config); - EXPECT_EQ(checkRemoteParameterUpdater(trainer), 0); - - FLAGS_local = 1; -} - -TEST(checkRemoteUpdater, cpuTrainer) { - checkRemoteParameterUpdaterTest(configFile1, false, false); -} - -TEST(checkRemoteUpdater, cpuTrainerOldUpdater) { - checkRemoteParameterUpdaterTest(configFile1, false, false, 1, true); -} - -#ifdef PADDLE_WITH_CUDA -TEST(checkRemoteUpdater, gpuTrainer) { - checkRemoteParameterUpdaterTest(configFile1, true, false); -} - -TEST(checkRemoteUpdater, gpu2Trainer) { - checkRemoteParameterUpdaterTest(configFile1, true, false, 2); -} - -TEST(checkRemoteUpdater, gpu4Trainer) { - checkRemoteParameterUpdaterTest(configFile1, true, false, 4); -} - -TEST(checkRemoteUpdater, gpuTrainerOldUpdater) { - checkRemoteParameterUpdaterTest(configFile1, true, false, 1, true); -} - -TEST(checkRemoteUpdater, gpu2TrainerOldUpdater) { - checkRemoteParameterUpdaterTest(configFile1, true, false, 2, true); -} - -TEST(checkRemoteUpdater, gpu4TrainerOldUpdater) { - checkRemoteParameterUpdaterTest(configFile1, true, false, 4, true); -} - -#endif - -TEST(checkRemoteUpdater, cpuDeltaTrainer) { - checkRemoteParameterUpdaterTest(configFile1, false, false, 1, false, 10); -} - -TEST(checkRemoteUpdater, cpuDeltaTrainerOldUpdater) { - checkRemoteParameterUpdaterTest(configFile1, false, false, 1, true, 10); -} - -TEST(SgdThreadUpdater, simpleSparseNN) { - trainerOnePassTest(configFileSimpleSparse, false, false, 1, 0.5, true); -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - initMain(argc, argv); - initPython(argc, argv); - gNumDevices = hl_get_device_count(); - - FLAGS_num_passes = 1; // train one pass - FLAGS_saving_period = 100000; // do not save parameteres - return RUN_ALL_TESTS(); -} diff --git a/paddle/trainer/tests/test_config.conf b/paddle/trainer/tests/test_config.conf deleted file mode 100644 index 2f86aaa75316fa2a5a28edfef31c01e15a44b3d0..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_config.conf +++ /dev/null @@ -1,77 +0,0 @@ -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -TrainData(SimpleData( - files = "trainer/tests/sample_filelist.txt", - feat_dim = 3, - context_len = 0, - buffer_capacity = 1000000, - async_load_data = False)) - -settings(batch_size = 100) - -data = data_layer(name='input', size=3) - -wt = data_layer(name='weight', size=1) - -fc1 = fc_layer(input=data, size=5, - bias_attr=True, - act=SigmoidActivation()) - -fc2 = fc_layer(input=data, size=12, - bias_attr=True, - param_attr=ParamAttr(name='sharew'), - act=LinearActivation()) - -fc3 = fc_layer(input=data, size=3, - bias_attr=True, - act=TanhActivation()) - -fc4 = fc_layer(input=data, size=5, - bias_attr=True, - layer_attr=ExtraAttr(drop_rate=0.5), - act=SquareActivation()) - -pool = img_pool_layer(input=fc2, - pool_size=2, - pool_size_y=3, - num_channels=1, - padding=1, - padding_y=2, - stride=2, - stride_y=3, - pool_type=CudnnAvgPooling()) - -concat = concat_layer(input=[fc3, fc4]) - -with mixed_layer(size=3, act=SoftmaxActivation()) as output: - output += full_matrix_projection(input=fc1) - output += trans_full_matrix_projection(input=fc2, - param_attr=ParamAttr(name='sharew')) - output += full_matrix_projection(input=concat) - output += identity_projection(input=fc3) - -lbl = data_layer(name='label', size=1) - -cost = classification_cost(input=output, label=lbl, weight=wt, - layer_attr=ExtraAttr(device=-1)) - -nce = nce_layer(input=fc2, label=lbl, weight=wt, - num_classes=3, - neg_distribution=[0.1, 0.3, 0.6]) - -outputs(cost, nce) diff --git a/paddle/trainer/tests/test_recurrent_machine_generation.cpp b/paddle/trainer/tests/test_recurrent_machine_generation.cpp deleted file mode 100644 index a8fbe31c2b1e228107dfc19483444409bfcbf788..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_recurrent_machine_generation.cpp +++ /dev/null @@ -1,155 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include -#include - -#include - -using namespace paddle; // NOLINT -using namespace std; // NOLINT - -static const string& CONFIG_FILE = "trainer/tests/sample_trainer_rnn_gen.conf"; -static const string& NEST_CONFIG_FILE = - "trainer/tests/sample_trainer_nest_rnn_gen.conf"; -static const string& OUTPUT_DIR = "trainer/tests/dump_text.test"; -static string modelDir = "trainer/tests/rnn_gen_test_model_dir/t1"; // NOLINT -static string expectFile = // NOLINT - "trainer/tests/rnn_gen_test_model_dir/r1.test"; // NOLINT - -DECLARE_string(config_args); - -vector readRetFile(const string& fname) { - ifstream inFile(fname); - float ret; - vector nums; - while (inFile >> ret) { - nums.push_back(ret); - } - return nums; -} - -void checkOutput(const string& expRetFile) { - vector rets = readRetFile(OUTPUT_DIR); - vector expRets = readRetFile(expRetFile); - EXPECT_EQ(rets.size(), expRets.size()); - for (size_t i = 0; i < rets.size(); i++) { - EXPECT_FLOAT_EQ(rets[i], expRets[i]); - } -} - -void prepareInArgs(vector& inArgs, - const size_t batchSize, - bool useGpu, - bool hasSubseq) { - inArgs.clear(); - // sentence id - Argument sentId; - sentId.value = nullptr; - if (hasSubseq) { - // as there is only one sequence, there is only one label. - IVector::resizeOrCreate(sentId.ids, 1, useGpu); - sentId.ids->setElement(0, 0); - } else { - // as there is batchSize word, there is batchSize label. - IVector::resizeOrCreate(sentId.ids, batchSize, useGpu); - for (size_t i = 0; i < batchSize; ++i) sentId.ids->setElement(i, i); - } - inArgs.emplace_back(sentId); - - // a dummy layer to decide batch size - Argument dummyInput; - dummyInput.value = Matrix::create(batchSize, 2, false, useGpu); - dummyInput.value->randomizeUniform(); - if (hasSubseq) { - // generate one sequence with batchSize subsequence, - // and each subsequence has only one word. - dummyInput.sequenceStartPositions = ICpuGpuVector::create(2, false); - int* buf = dummyInput.sequenceStartPositions->getMutableData(false); - dummyInput.subSequenceStartPositions = - ICpuGpuVector::create(batchSize + 1, false); - int* subBuf = dummyInput.subSequenceStartPositions->getMutableData(false); - buf[0] = 0; - buf[1] = batchSize; - for (size_t i = 0; i < batchSize + 1; i++) subBuf[i] = i; - } - inArgs.emplace_back(dummyInput); -} - -void testGeneration(const string& configFile, - bool useGpu, - bool hasSubseq, - const string& expRetFile) { - FLAGS_use_gpu = useGpu; - auto config = std::make_shared(configFile); - unique_ptr gradientMachine(GradientMachine::create(*config)); - gradientMachine->loadParameters(modelDir); - vector inArgs(2); - - const size_t batchSize = 15; - prepareInArgs(inArgs, batchSize, useGpu, hasSubseq); - vector outArgs; - unique_ptr testEvaluator(gradientMachine->makeEvaluator()); - testEvaluator->start(); - gradientMachine->forward(inArgs, &outArgs, PASS_TEST); - gradientMachine->eval(testEvaluator.get()); - testEvaluator->finish(); - checkOutput(expRetFile); -} - -#ifndef PADDLE_TYPE_DOUBLE - -TEST(RecurrentGradientMachine, test_generation) { -#ifndef PADDLE_WITH_CUDA - const auto useGpuConfs = {false}; -#else - const auto useGpuConfs = {true, false}; -#endif - auto testGen = [&](const string& configFile, - bool hasSubseq, - const string& expRetFile, - bool beam_search) { - FLAGS_config_args = beam_search ? "beam_search=1" : "beam_search=0"; - for (auto useGpu : useGpuConfs) { - LOG(INFO) << configFile << " useGpu=" << useGpu - << " beam_search=" << beam_search; - testGeneration(configFile, useGpu, hasSubseq, expRetFile); - } - }; - testGen(CONFIG_FILE, false, expectFile + ".nobeam", false); // no beam search - testGen(CONFIG_FILE, false, expectFile + ".beam", true); // beam search - // In hierarchical RNN, beam search and one way search are only in inner-RNN, - // outer-RNN will concat the generated inner-results (first for beam search) - // from inner-RNN. Thus, they have the same outer-results. - testGen(NEST_CONFIG_FILE, - true, - expectFile + ".nest", - false); // no beam search - testGen(NEST_CONFIG_FILE, true, expectFile + ".nest", true); // beam search -} -#endif - -int main(int argc, char** argv) { - initMain(argc, argv); - initPython(argc, argv); - CHECK(argc == 1 || argc == 3); - if (argc == 3) { - modelDir = argv[1]; - expectFile = argv[2]; - } - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/utils/CpuId.cpp b/paddle/utils/CpuId.cpp deleted file mode 100644 index 7186feef041eb3b1be459a506294f83f9a00ad94..0000000000000000000000000000000000000000 --- a/paddle/utils/CpuId.cpp +++ /dev/null @@ -1,66 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/CpuId.h" -#include "paddle/utils/Util.h" - -#ifdef _WIN32 - -#include - -/// for MSVC -#define CPUID(info, x) __cpuidex(info, x, 0) - -#else - -#if !defined(__arm__) && !defined(__aarch64__) -#include -/// for GCC/Clang -#define CPUID(info, x) __cpuid_count(x, 0, info[0], info[1], info[2], info[3]) -#endif - -#endif - -namespace paddle { - -SIMDFlags::SIMDFlags() { -#if defined(__arm__) || defined(__aarch64__) - simd_flags_ = SIMD_NEON; -#else - unsigned int cpuInfo[4]; - // CPUID: https://en.wikipedia.org/wiki/CPUID - // clang-format off - CPUID(cpuInfo, 0x00000001); - simd_flags_ |= cpuInfo[3] & (1 << 25) ? SIMD_SSE : SIMD_NONE; - simd_flags_ |= cpuInfo[3] & (1 << 26) ? SIMD_SSE2 : SIMD_NONE; - simd_flags_ |= cpuInfo[2] & (1 << 0) ? SIMD_SSE3 : SIMD_NONE; - simd_flags_ |= cpuInfo[2] & (1 << 9) ? SIMD_SSSE3 : SIMD_NONE; - simd_flags_ |= cpuInfo[2] & (1 << 19) ? SIMD_SSE41 : SIMD_NONE; - simd_flags_ |= cpuInfo[2] & (1 << 20) ? SIMD_SSE42 : SIMD_NONE; - simd_flags_ |= cpuInfo[2] & (1 << 12) ? SIMD_FMA3 : SIMD_NONE; - simd_flags_ |= cpuInfo[2] & (1 << 28) ? SIMD_AVX : SIMD_NONE; - - CPUID(cpuInfo, 0x00000007); - simd_flags_ |= cpuInfo[1] & (1 << 5) ? SIMD_AVX2 : SIMD_NONE; - simd_flags_ |= cpuInfo[1] & (1 << 16) ? SIMD_AVX512: SIMD_NONE; - - CPUID(cpuInfo, 0x80000001); - simd_flags_ |= cpuInfo[2] & (1 << 16) ? SIMD_FMA4 : SIMD_NONE; - // clang-fotmat on -#endif -} - -SIMDFlags const* SIMDFlags::instance() { - static SIMDFlags instance; - return &instance; -} - -} // namespace paddle diff --git a/paddle/utils/PythonUtil.cpp b/paddle/utils/PythonUtil.cpp deleted file mode 100644 index 7faeff55c28b9065179ad27b3b604a9f411249e5..0000000000000000000000000000000000000000 --- a/paddle/utils/PythonUtil.cpp +++ /dev/null @@ -1,209 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "PythonUtil.h" -#include -#include - -namespace paddle { - -#ifdef PADDLE_NO_PYTHON - -DEFINE_string(python_path, "", "python path"); -DEFINE_string(python_bin, "python2.7", "python bin"); - -constexpr int kExecuteCMDBufLength = 204800; - -int executeCMD(const char* cmd, char* result) { - char bufPs[kExecuteCMDBufLength]; - char ps[kExecuteCMDBufLength] = {0}; - FILE* ptr; - strncpy(ps, cmd, kExecuteCMDBufLength); - if ((ptr = popen(ps, "r")) != NULL) { - size_t count = fread(bufPs, 1, kExecuteCMDBufLength, ptr); - memcpy(result, - bufPs, - count - 1); // why count-1: remove the '\n' at the end - result[count] = 0; - pclose(ptr); - ptr = NULL; - return count - 1; - } else { - LOG(FATAL) << "popen failed"; - return -1; - } -} - -std::string callPythonFunc(const std::string& moduleName, - const std::string& funcName, - const std::vector& args) { - std::string pythonLibPath = ""; - std::string pythonBinPath = ""; - if (!FLAGS_python_path.empty()) { - pythonLibPath = FLAGS_python_path + "/lib:"; - pythonBinPath = FLAGS_python_path + "/bin/"; - } - std::string s = "LD_LIBRARY_PATH=" + pythonLibPath + "$LD_LIBRARY_PATH " + - pythonBinPath + std::string(FLAGS_python_bin) + - " -c 'import " + moduleName + "\n" + "print " + moduleName + - "." + funcName + "("; - for (auto& arg : args) { - s = s + "\"" + arg + "\", "; - } - s += ")'"; - char result[kExecuteCMDBufLength] = {0}; - LOG(INFO) << " cmd string: " << s; - int length = executeCMD(s.c_str(), result); - CHECK_NE(-1, length); - return std::string(result, length); -} - -#else - -static std::recursive_mutex g_pyMutex; - -PyGuard::PyGuard() : guard_(g_pyMutex) {} - -static void printPyErrorStack(std::ostream& os, - bool withEndl = false, - bool withPyPath = true) { - PyObject *ptype, *pvalue, *ptraceback; - PyErr_Fetch(&ptype, &pvalue, &ptraceback); - PyErr_NormalizeException(&ptype, &pvalue, &ptraceback); - PyErr_Clear(); - if (withPyPath) { - os << "Current PYTHONPATH: " << py::repr(PySys_GetObject(strdup("path"))); - if (withEndl) { - os << std::endl; - } - } - PyTracebackObject* obj = (PyTracebackObject*)ptraceback; - - os << "Python Error: " << PyString_AsString(PyObject_Str(ptype)) << " : " - << (pvalue == NULL ? "" : PyString_AsString(PyObject_Str(pvalue))); - if (withEndl) { - os << std::endl; - } - os << "Python Callstack: "; - if (withEndl) { - os << std::endl; - } - while (obj != NULL) { - int line = obj->tb_lineno; - const char* filename = - PyString_AsString(obj->tb_frame->f_code->co_filename); - os << " " << filename << " : " << line; - if (withEndl) { - os << std::endl; - } - obj = obj->tb_next; - } - - Py_XDECREF(ptype); - Py_XDECREF(pvalue); - Py_XDECREF(ptraceback); -} -PyObjectPtr callPythonFuncRetPyObj(const std::string& moduleName, - const std::string& funcName, - const std::vector& args) { - PyGuard guard; - PyObjectPtr pyModule = py::import(moduleName); - PyObjectPtr pyFunc(PyObject_GetAttrString(pyModule.get(), funcName.c_str())); - CHECK_PY(pyFunc) << "GetAttrString failed."; - PyObjectPtr pyArgs(PyTuple_New(args.size())); - for (size_t i = 0; i < args.size(); ++i) { - PyObjectPtr pyArg(PyString_FromString(args[i].c_str())); - CHECK_PY(pyArg) << "Import pyArg failed."; - PyTuple_SetItem(pyArgs.get(), i, pyArg.release()); // Maybe a problem - } - PyObjectPtr ret(PyObject_CallObject(pyFunc.get(), pyArgs.get())); - CHECK_PY(ret) << "Call Object failed."; - return ret; -} - -std::string callPythonFunc(const std::string& moduleName, - const std::string& funcName, - const std::vector& args) { - PyObjectPtr obj = callPythonFuncRetPyObj(moduleName, funcName, args); - return std::string(PyString_AsString(obj.get()), PyString_Size(obj.get())); -} - -PyObjectPtr createPythonClass( - const std::string& moduleName, - const std::string& className, - const std::vector& args, - const std::map& kwargs) { - PyGuard guard; - PyObjectPtr pyModule = py::import(moduleName); - LOG(INFO) << "createPythonClass moduleName.c_str:" << moduleName.c_str(); - CHECK_PY(pyModule) << "Import module " << moduleName << " failed."; - PyObjectPtr pyDict(PyModule_GetDict(pyModule.get())); - CHECK_PY(pyDict) << "Get Dict failed."; - PyObjectPtr pyClass(PyDict_GetItemString(pyDict.get(), className.c_str())); - LOG(INFO) << "createPythonClass className.c_str():" << className.c_str(); - CHECK_PY(pyClass) << "Import class " << className << " failed."; - PyObjectPtr argsObjectList(PyTuple_New(args.size())); - for (size_t i = 0; i < args.size(); ++i) { - PyObjectPtr pyArg(Py_BuildValue("s#", args[i].c_str(), args[i].length())); - PyTuple_SetItem(argsObjectList.get(), i, pyArg.release()); - } - - PyObjectPtr kwargsObjectList(PyDict_New()); - for (auto& x : kwargs) { - PyObjectPtr pyArg(Py_BuildValue("s#", x.second.c_str(), x.second.length())); - PyDict_SetItemString( - kwargsObjectList.get(), x.first.c_str(), pyArg.release()); - } - - PyObjectPtr pyInstance(PyInstance_New( - pyClass.get(), argsObjectList.release(), kwargsObjectList.release())); - CHECK_PY(pyInstance) << "Create class " << className << " failed."; - return pyInstance; -} - -namespace py { -char* repr(PyObject* obj) { return PyString_AsString(PyObject_Repr(obj)); } - -std::string getPyCallStack() { - std::ostringstream os; - printPyErrorStack(os, true); - return os.str(); -} - -PyObjectPtr import(const std::string& moduleName) { - auto module = PyImport_ImportModule(moduleName.c_str()); - CHECK_PY(module) << "Import " << moduleName << "Error"; - return PyObjectPtr(module); -} - -} // namespace py - -#endif -extern "C" { -extern const char enable_virtualenv_py[]; -} -void initPython(int argc, char** argv) { -#ifndef PADDLE_NO_PYTHON - Py_SetProgramName(argv[0]); - Py_Initialize(); - PySys_SetArgv(argc, argv); - // python blocks SIGINT. Need to enable it. - signal(SIGINT, SIG_DFL); - - // Manually activate virtualenv when user is using virtualenv - PyRun_SimpleString(enable_virtualenv_py); -#endif -} - -} // namespace paddle diff --git a/paddle/utils/PythonUtil.h b/paddle/utils/PythonUtil.h deleted file mode 100644 index 6f8d7e09309503e47aca7ae2d20774c748703b21..0000000000000000000000000000000000000000 --- a/paddle/utils/PythonUtil.h +++ /dev/null @@ -1,353 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -// clang-format off -#include "paddle/utils/Util.h" - -#ifndef PADDLE_NO_PYTHON -// must include the following two blocks, otherwise, -// gcc compiler may produce warning -#ifdef __APPLE__ -#define _POSIX_SOURCE -#define _POSIX_C_SOURCE 200809L -#define _XOPEN_SOURCE 700 -#endif - -#ifdef _POSIX_C_SOURCE -#define __TEMP_POSIX_C_SOURCE _POSIX_C_SOURCE -#undef _POSIX_C_SOURCE -#endif -#ifdef _XOPEN_SOURCE -#define __TEMP_XOPEN_SOURCE _XOPEN_SOURCE -#undef _XOPEN_SOURCE -#endif -#include -#include -#endif - -#include -#include -#include -// clang-format on - -namespace paddle { - -std::string callPythonFunc(const std::string& moduleName, - const std::string& funcName, - const std::vector& args); - -#ifndef PADDLE_NO_PYTHON - -/** - * Global lock guard of python C-api invokes. - * NOTE: the lock of this guard is reentrant or recursive. - */ -class PyGuard { - public: - PyGuard(); - PyGuard(const PyGuard& other) = delete; - PyGuard& operator=(const PyGuard& other) = delete; - - private: - std::lock_guard guard_; -}; - -struct PyObjectDeleter { - void operator()(PyObject* obj) { - if (obj) { - Py_DECREF(obj); - } - } -}; - -typedef std::unique_ptr PyObjectPtr; - -PyObjectPtr callPythonFuncRetPyObj(const std::string& moduleName, - const std::string& funcName, - const std::vector& args); - -PyObjectPtr createPythonClass(const std::string& moduleName, - const std::string& className, - const std::vector& args, - const std::map& kwargs); - -#define CHECK_PY(x) CHECK((x) != nullptr) << ::paddle::py::getPyCallStack() - -namespace py { -PyObjectPtr import(const std::string& moduleName); - -/** - * Cast a PyLong or PyInt to int type T. - * @tparam T return type. - * @param [in] obj PyLong or PyInt object. - * @param [out] ok status for casting. False if error occured. nullptr if user - * don't care is ok or not. - * @return The value of python object, or 0 if not ok. - */ -template -T castInt(PyObject* obj, bool* ok = nullptr) { - if (PyLong_Check(obj)) { - if (ok) *ok = true; - return (T)PyLong_AsUnsignedLong(obj); - } else if (PyInt_Check(obj)) { - if (ok) *ok = true; - return (T)PyInt_AsLong(obj); - } else { - if (ok) *ok = false; - return (T)0; - } -} - -/** - * Invoke repr of python object. - * - * Just like toString method in java. - */ -char* repr(PyObject* obj); - -/** - * Invoke repr of python object. - */ -inline char* repr(const PyObjectPtr& obj) { return repr(obj.get()); } - -/** - * Get Python Error Stack String. - */ -std::string getPyCallStack(); - -/** - * Object Helper for PyObjectPtr. - * - * Implements getAttr method for object. - */ -class ObjectHelper { - public: - explicit ObjectHelper(const PyObjectPtr& obj) : obj_(obj) {} - - /** - * get attribute - */ - inline PyObject* getAttr(const std::string& field) const { - auto obj = PyObject_GetAttrString(obj_.get(), field.c_str()); - CHECK_PY(obj) << "Cannot get attribute on python object " << obj_.get(); - return obj; - } - - /** - * Get Int attribute - * @param [in] field attribute name. - * @param [out] ok true if this attribute is int. - * @tparam T int type. - * @return int value. - */ - template - T getIntAttr(const std::string& field, bool* ok = nullptr) const { - PyObjectPtr tmp(getAttr(field)); - return castInt(tmp.get(), ok); - } - - /** - * Get int attribute. Log(Fatal) when not ok - * @param field attribute name. - * @return int value. - */ - template - T getIntAttrWithError(const std::string& field) const { - bool ok; - T tmp = getIntAttr(field, &ok); - CHECK(ok) << "Cannot get integer attribute on object " << obj_.get(); - return tmp; - } - - /** - * Get bool attribute. - * @param field - * @param [out] isBoolType return true if attribute is bool type. If the - * attribute is not bool type, then an implicit - * conversion will happens, and will return the - * conversion result. - * - * Such as, if the attribute is 1, then the return - * value of function will be true, but the isBoolType - * will return false. - * @return - */ - bool getBoolAttr(const std::string& field, bool* isBoolType = nullptr) const { - PyObjectPtr tmp(getAttr(field)); - if (isBoolType) { - *isBoolType = PyBool_Check(tmp.get()); - } - return PyObject_IsTrue(tmp.get()); - } - - private: - const PyObjectPtr& obj_; -}; - -/** - * Python Sequence Helper - * - * The python sequence means list or tuple. - */ -class SequenceHelper { - public: - explicit SequenceHelper(const PyObjectPtr& seq) : seq_(seq.get()) { - CHECK(PySequence_Check(seq_)); - } - - explicit SequenceHelper(PyObject* seq) : seq_(seq) { - CHECK(PySequence_Check(seq_)); - } - - inline size_t size() const { return (size_t)PySequence_Size(seq_); } - - inline PyObject* operator[](size_t i) const { - return PySequence_Fast_GET_ITEM(seq_, i); - } - - inline double getDouble(size_t i) const { - auto* ptr = (*this)[i]; - return PyFloat_AsDouble(ptr); - } - - /** - * Set a sequence item o[i] = obj; - * @param i index - * @param obj setted item. - * @param steal if steal = true, sequence will move object in iteself, - * just like std::move. Otherwise, it will increase reference - * count. Default is false. - */ - inline void set(size_t i, const PyObjectPtr& obj, bool steal = false) { - this->set(i, obj.get(), steal); - } - - /** - * Set a sequence item o[i] = obj; - */ - inline void set(size_t i, PyObject* obj, bool steal = false) { - if (!steal) { - Py_XINCREF(obj); - } - if (PyTuple_Check(seq_)) { - CHECK_NE(PyTuple_SetItem(seq_, i, obj), -1) << getPyCallStack(); - } else { - CHECK_NE(PySequence_SetItem(seq_, i, obj), -1) << getPyCallStack(); - } - } - - private: - PyObject* seq_; -}; - -class DictHelper { - public: - explicit DictHelper(PyObject* d) : dict_(d) {} - - explicit DictHelper(const PyObjectPtr& d) : dict_(d.get()) {} - - void set(const std::string& key, PyObject* item) { - PyDict_SetItemString(dict_, key.c_str(), item); - } - - void setBool(const std::string& key, bool b) { - this->set(key, PyBool_FromLong(b)); - } - - void setStringList(const std::string& key, - const std::vector& items) { - auto* list = PyList_New(items.size()); - for (size_t i = 0; i < items.size(); ++i) { - PyList_SetItem(list, i, PyString_FromString(items[i].c_str())); - } - this->set(key, list); - } - - private: - inline void checkDict() { CHECK(PyDict_Check(this->dict_)); } - - PyObject* dict_; -}; - -inline static bool isCallable(const PyObjectPtr& obj) { - return PyCallable_Check(obj.get()); -} - -/** - * Wrap a callable object. - */ -class CallableHelper { - public: - explicit CallableHelper(const PyObjectPtr& obj) : obj_(obj) { - CHECK(py::isCallable(obj_)); - } - - ~CallableHelper() {} - - /** - * reset args, and create new tuple. - * @param sz args size. - */ - void setArgsSize(size_t sz) { args.reset(PyTuple_New(sz)); } - - /** - * Get args sequence. User can set/get by SequenceHelper. - */ - SequenceHelper getArgs() { return SequenceHelper(args); } - - /** - * Call python method, return an object. - */ - PyObject* operator()() { - PyGuard guard; - return PyObject_Call(obj_.get(), args.get(), kwargs.get()); - } - - private: - const PyObjectPtr& obj_; - PyObjectPtr args; - PyObjectPtr kwargs; -}; - -inline static PyObject* iterNext(const PyObjectPtr& context, bool* atEnd) { - PyGuard g; - PyObject* data = PyIter_Next(context.get()); - if (data == nullptr) { - if (PyErr_ExceptionMatches(PyExc_StopIteration)) { - PyErr_Clear(); - *atEnd = true; - return nullptr; - } else if (PyErr_Occurred()) { - CHECK_PY(data) << "Calling iterator next error"; - return nullptr; - } else { - *atEnd = false; - return data; // just return none in iterator. - } - } else { - *atEnd = false; - return data; - } -} -} // namespace py - -#endif - -/** - * Initialize python. - */ -void initPython(int argc, char** argv); - -} // namespace paddle diff --git a/paddle/utils/arch/linux/Locks.cpp b/paddle/utils/arch/linux/Locks.cpp deleted file mode 100644 index 409af8bce3621c51bfd7a69c6b4ec1f9cc6be8e4..0000000000000000000000000000000000000000 --- a/paddle/utils/arch/linux/Locks.cpp +++ /dev/null @@ -1,149 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Locks.h" -#include -#include -#include "paddle/utils/Logging.h" - -namespace paddle { -class SemaphorePrivate { - public: - sem_t sem; -}; - -Semaphore::Semaphore(int initValue) : m(new SemaphorePrivate()) { - sem_init(&m->sem, 0, initValue); -} - -Semaphore::~Semaphore() { - sem_destroy(&m->sem); - delete m; -} - -bool Semaphore::timeWait(struct timespec* ts) { - return (0 == sem_timedwait(&m->sem, ts)); -} - -void Semaphore::wait() { sem_wait(&m->sem); } - -void Semaphore::post() { sem_post(&m->sem); } - -/// SpinLockPrivate - -#ifdef PADDLE_USE_PTHREAD_SPINLOCK - -class SpinLockPrivate { - public: - inline SpinLockPrivate() { pthread_spin_init(&lock_, 0); } - inline ~SpinLockPrivate() { pthread_spin_destroy(&lock_); } - - inline void lock() { pthread_spin_lock(&lock_); } - inline void unlock() { pthread_spin_unlock(&lock_); } - - pthread_spinlock_t lock_; - char padding_[64 - sizeof(pthread_spinlock_t)]; -}; - -#else -// clang-format off -#include -#include -// clang-format on - -class SpinLockPrivate { - public: - inline void lock() { - while (lock_.test_and_set(std::memory_order_acquire)) { - } - } - inline void unlock() { lock_.clear(std::memory_order_release); } - - std::atomic_flag lock_ = ATOMIC_FLAG_INIT; - char padding_[64 - sizeof(lock_)]; // Padding to cache line size -}; - -#endif - -SpinLock::SpinLock() : m(new SpinLockPrivate()) {} -SpinLock::~SpinLock() { delete m; } -void SpinLock::lock() { m->lock(); } -void SpinLock::unlock() { m->unlock(); } - -/// ThreadBarrierPrivate - -#ifdef PADDLE_USE_PTHREAD_BARRIER - -class ThreadBarrierPrivate { - public: - pthread_barrier_t barrier_; - - inline explicit ThreadBarrierPrivate(int count) { - pthread_barrier_init(&barrier_, nullptr, count); - } - - inline ~ThreadBarrierPrivate() { pthread_barrier_destroy(&barrier_); } - - inline void wait() { pthread_barrier_wait(&barrier_); } -}; - -#else - -class ThreadBarrierPrivate { - public: - pthread_mutex_t mutex_; - pthread_cond_t cond_; - int count_; - int tripCount_; - - inline explicit ThreadBarrierPrivate(int cnt) : count_(0), tripCount_(cnt) { - CHECK_NE(cnt, 0); - CHECK_GE(pthread_mutex_init(&mutex_, 0), 0); - CHECK_GE(pthread_cond_init(&cond_, 0), 0); - } - - inline ~ThreadBarrierPrivate() { - pthread_cond_destroy(&cond_); - pthread_mutex_destroy(&mutex_); - } - - /** - * @brief wait - * @return true if the last wait - */ - inline bool wait() { - pthread_mutex_lock(&mutex_); - ++count_; - if (count_ >= tripCount_) { - count_ = 0; - pthread_cond_broadcast(&cond_); - pthread_mutex_unlock(&mutex_); - return true; - } else { - pthread_cond_wait(&cond_, &mutex_); - pthread_mutex_unlock(&mutex_); - return false; - } - } -}; - -#endif - -/// ThreadBarrier - -ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {} -ThreadBarrier::~ThreadBarrier() { delete m; } -void ThreadBarrier::wait() { m->wait(); } - -} // namespace paddle diff --git a/paddle/utils/arch/osx/Excepts.cpp b/paddle/utils/arch/osx/Excepts.cpp deleted file mode 100644 index ac444615786fa9f89f96504a31b2289eae7bb643..0000000000000000000000000000000000000000 --- a/paddle/utils/arch/osx/Excepts.cpp +++ /dev/null @@ -1,57 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Excepts.h" - -#if defined(__APPLE__) || defined(__OSX__) -#if defined(__arm__) || defined(__arm64__) -// TODO(liuyiqun): implement the arm version -int fegetexcept(void) { return -1; } -int feenableexcept(unsigned int excepts) { return -1; } -int fedisableexcept(unsigned int excepts) { return -1; } -#else -int fegetexcept(void) { - static fenv_t fenv; - return fegetenv(&fenv) ? -1 : (fenv.__control & FE_ALL_EXCEPT); -} - -int feenableexcept(unsigned int excepts) { - static fenv_t fenv; - unsigned int new_excepts = excepts & FE_ALL_EXCEPT, old_excepts; - - if (fegetenv(&fenv)) return -1; - old_excepts = fenv.__control & FE_ALL_EXCEPT; - - // unmask - fenv.__control &= ~new_excepts; - fenv.__mxcsr &= ~(new_excepts << 7); - - return (fesetenv(&fenv) ? -1 : old_excepts); -} - -int fedisableexcept(unsigned int excepts) { - static fenv_t fenv; - unsigned int new_excepts = excepts & FE_ALL_EXCEPT, old_excepts; - - if (fegetenv(&fenv)) return -1; - old_excepts = fenv.__control & FE_ALL_EXCEPT; - - // mask - fenv.__control |= new_excepts; - fenv.__mxcsr |= new_excepts << 7; - - return (fesetenv(&fenv) ? -1 : old_excepts); -} -#endif -#endif diff --git a/paddle/utils/arch/osx/Locks.cpp b/paddle/utils/arch/osx/Locks.cpp deleted file mode 100644 index f3905091bd024ab02c3f5d39cfed6dbc38fabbbc..0000000000000000000000000000000000000000 --- a/paddle/utils/arch/osx/Locks.cpp +++ /dev/null @@ -1,105 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Locks.h" -#include -#include -#include -#include "paddle/utils/Logging.h" - -namespace paddle { - -class SemaphorePrivate { - public: - ~SemaphorePrivate() { dispatch_release(sem); } - - dispatch_semaphore_t sem; -}; - -Semaphore::Semaphore(int initValue) : m(new SemaphorePrivate()) { - m->sem = dispatch_semaphore_create(initValue); -} - -Semaphore::~Semaphore() { delete m; } - -bool Semaphore::timeWait(timespec *ts) { - dispatch_time_t tm = dispatch_walltime(ts, 0); - return (0 == dispatch_semaphore_wait(m->sem, tm)); -} - -void Semaphore::wait() { - dispatch_semaphore_wait(m->sem, DISPATCH_TIME_FOREVER); -} - -void Semaphore::post() { dispatch_semaphore_signal(m->sem); } - -class SpinLockPrivate { - public: - std::atomic_flag lock_ = ATOMIC_FLAG_INIT; - char padding_[64 - sizeof(lock_)]; // Padding to cache line size -}; - -SpinLock::SpinLock() : m(new SpinLockPrivate()) {} -SpinLock::~SpinLock() { delete m; } - -void SpinLock::lock() { - while (m->lock_.test_and_set(std::memory_order_acquire)) { - } -} - -void SpinLock::unlock() { m->lock_.clear(std::memory_order_release); } - -class ThreadBarrierPrivate { - public: - pthread_mutex_t mutex_; - pthread_cond_t cond_; - int count_; - int tripCount_; - - inline explicit ThreadBarrierPrivate(int cnt) : count_(0), tripCount_(cnt) { - CHECK_NE(cnt, 0); - CHECK_GE(pthread_mutex_init(&mutex_, 0), 0); - CHECK_GE(pthread_cond_init(&cond_, 0), 0); - } - - inline ~ThreadBarrierPrivate() { - pthread_cond_destroy(&cond_); - pthread_mutex_destroy(&mutex_); - } - - /** - * @brief wait - * @return true if the last wait - */ - inline bool wait() { - pthread_mutex_lock(&mutex_); - ++count_; - if (count_ >= tripCount_) { - count_ = 0; - pthread_cond_broadcast(&cond_); - pthread_mutex_unlock(&mutex_); - return true; - } else { - pthread_cond_wait(&cond_, &mutex_); - pthread_mutex_unlock(&mutex_); - return false; - } - } -}; - -ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {} -ThreadBarrier::~ThreadBarrier() { delete m; } -void ThreadBarrier::wait() { m->wait(); } - -} // namespace paddle diff --git a/paddle/utils/tests/CMakeLists.txt b/paddle/utils/tests/CMakeLists.txt deleted file mode 100644 index c770ce169878d9998e559b1d417fc1acc88cde97..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/CMakeLists.txt +++ /dev/null @@ -1,18 +0,0 @@ -add_simple_unittest(test_Thread) -add_simple_unittest(test_StringUtils) -add_simple_unittest(test_CustomStackTrace) -add_simple_unittest(test_ThreadBarrier) -add_simple_unittest(test_SpinLock) -add_simple_unittest(test_SIMDFlags) -add_simple_unittest(test_Error) - -add_executable( - test_CustomStackTracePrint - test_CustomStackTracePrint.cpp -) -link_paddle_exe(test_CustomStackTracePrint) -if(NOT APPLE) - add_test(NAME test_CustomStackTracePrint - COMMAND ${PADDLE_SOURCE_DIR}/paddle/utils/tests/test_CustomStackTracePrint.sh - WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) -endif() diff --git a/paddle/utils/tests/test_CustomStackTrace.cpp b/paddle/utils/tests/test_CustomStackTrace.cpp deleted file mode 100644 index 4d5540b24cb9d52482cfa5a77dfa956b8bf4ef38..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_CustomStackTrace.cpp +++ /dev/null @@ -1,92 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include // NOLINT -#include // NOLINT - -#include "paddle/utils/CustomStackTrace.h" -#include "paddle/utils/Locks.h" -#include "paddle/utils/StringUtil.h" -#include "paddle/utils/Util.h" - -DEFINE_int32(test_thread_num, 10, "testing thread number"); - -void testNormalImpl( - const std::function&, - size_t, - size_t, - paddle::ThreadBarrier&, - paddle::ThreadBarrier&)>& callback) { - paddle::CustomStackTrace tracer; - paddle::ThreadBarrier doneBarrier(FLAGS_test_thread_num + 1); - paddle::ThreadBarrier startBarrier(FLAGS_test_thread_num + 1); - constexpr size_t countDown = 10; - constexpr size_t layerSize = 1000; - std::vector> threads; - threads.reserve(FLAGS_test_thread_num); - - for (int32_t i = 0; i < FLAGS_test_thread_num; ++i) { - threads.emplace_back( - new std::thread([&tracer, &startBarrier, &doneBarrier, &callback] { - callback(tracer, countDown, layerSize, startBarrier, doneBarrier); - })); - } - size_t cntDown = countDown; - while (cntDown-- > 0) { - startBarrier.wait(); - sleep(1); - doneBarrier.wait(); - ASSERT_TRUE(tracer.empty()); - } - - for (auto& thread : threads) { - thread->join(); - } -} - -TEST(CustomStackTrace, normalTrain) { - testNormalImpl([](paddle::CustomStackTrace& tracer, - size_t countDown, - size_t layerSize, - paddle::ThreadBarrier& start, - paddle::ThreadBarrier& finish) { - while (countDown-- > 0) { - start.wait(); - for (size_t i = 0; i < layerSize; ++i) { - tracer.push("layer_" + paddle::str::to_string(i)); - } - for (size_t i = 0; i < layerSize; ++i) { - tracer.pop("layer_" + paddle::str::to_string(layerSize - 1 - i)); - } - finish.wait(); - } - }); -} - -TEST(CustomStackTrace, normalTest) { - testNormalImpl([](paddle::CustomStackTrace& tracer, - size_t countDown, - size_t layerSize, - paddle::ThreadBarrier& start, - paddle::ThreadBarrier& finish) { - while (countDown-- > 0) { - start.wait(); - for (size_t i = 0; i < layerSize; ++i) { - tracer.push("layer_" + paddle::str::to_string(i)); - } - tracer.clear(); // in forward test, tracer will clear after forward. - finish.wait(); - } - }); -} diff --git a/paddle/utils/tests/test_CustomStackTracePrint.cpp b/paddle/utils/tests/test_CustomStackTracePrint.cpp deleted file mode 100644 index 360c61c88a757da708b01d2bb54068b948b235cc..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_CustomStackTracePrint.cpp +++ /dev/null @@ -1,30 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/CustomStackTrace.h" -#include "paddle/utils/StringUtil.h" -#include "paddle/utils/Util.h" - -int main(int argc, char** argv) { - paddle::initMain(argc, argv); - - for (size_t i = 0; i < 1000; ++i) { - paddle::gLayerStackTrace.push("layer_" + paddle::str::to_string(i)); - if (i == 998) { - throw "Unhandle exception"; - } - } - - return 0; -} diff --git a/paddle/utils/tests/test_Error.cpp b/paddle/utils/tests/test_Error.cpp deleted file mode 100644 index 6f311fa6b80191de1e11ce1f63c31b64fe2eeb80..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_Error.cpp +++ /dev/null @@ -1,34 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/Error.h" - -#include - -TEST(Error, testAll) { - paddle::Error error; - ASSERT_TRUE(error.isOK()); - error = paddle::Error("I'm the error"); - ASSERT_FALSE(error.isOK()); - ASSERT_STREQ("I'm the error", error.msg()); - - error = paddle::Error("error2"); - ASSERT_FALSE(error.isOK()); - ASSERT_STREQ("error2", error.msg()); - - int i = 3; - auto error3 = paddle::Error("error%d", i); - ASSERT_FALSE(error3.isOK()); - ASSERT_STREQ("error3", error3.msg()); -} diff --git a/paddle/utils/tests/test_SIMDFlags.cpp b/paddle/utils/tests/test_SIMDFlags.cpp deleted file mode 100644 index a808d456a69866f72502bcf1ae244cec14738e22..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_SIMDFlags.cpp +++ /dev/null @@ -1,48 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include "paddle/utils/CpuId.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Util.h" - -using namespace paddle; // NOLINT - -TEST(SIMDFlags, gccTest) { -#if (defined(__GNUC__) || defined(__GNUG__)) && !(defined(__clang__)) && \ - !defined(__arm__) && !defined(__aarch64__) - // clang-format off - CHECK(!__builtin_cpu_supports("sse") != HAS_SSE); - CHECK(!__builtin_cpu_supports("sse2") != HAS_SSE2); - CHECK(!__builtin_cpu_supports("sse3") != HAS_SSE3); - CHECK(!__builtin_cpu_supports("ssse3") != HAS_SSSE3); - CHECK(!__builtin_cpu_supports("sse4.1") != HAS_SSE41); - CHECK(!__builtin_cpu_supports("sse4.2") != HAS_SSE42); - CHECK(!__builtin_cpu_supports("avx") != HAS_AVX); - CHECK(!__builtin_cpu_supports("avx2") != HAS_AVX2); -// clang-format on -#endif -} - -TEST(SIMDFlags, normalPrint) { - LOG(INFO) << "Has SSE: " << std::boolalpha << HAS_SSE; - LOG(INFO) << "Has SSE2: " << std::boolalpha << HAS_SSE2; - LOG(INFO) << "Has SSE3: " << std::boolalpha << HAS_SSE3; - LOG(INFO) << "Has SSSE3: " << std::boolalpha << HAS_SSSE3; - LOG(INFO) << "Has SSE4: " << std::boolalpha << HAS_SSE41 || HAS_SSE42; - LOG(INFO) << "Has FMA3: " << std::boolalpha << HAS_FMA3; - LOG(INFO) << "Has FMA4: " << std::boolalpha << HAS_FMA4; - LOG(INFO) << "Has AVX: " << std::boolalpha << HAS_AVX; - LOG(INFO) << "Has AVX2: " << std::boolalpha << HAS_AVX2; - LOG(INFO) << "Has AVX512: " << std::boolalpha << HAS_AVX512; - LOG(INFO) << "Has NEON: " << std::boolalpha << HAS_NEON; -} diff --git a/paddle/utils/tests/test_SpinLock.cpp b/paddle/utils/tests/test_SpinLock.cpp deleted file mode 100644 index cc34eb1f868003d3db9221578c0c20c44be285eb..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_SpinLock.cpp +++ /dev/null @@ -1,55 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include -#include - -#include "paddle/utils/Locks.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Util.h" - -DEFINE_int32(test_thread_num, 100, "testing thread number"); - -void testNormalImpl( - size_t thread_num, - const std::function& callback) { - paddle::SpinLock mutex; - std::vector threads; - threads.reserve(thread_num); - - size_t count = 0; - for (size_t i = 0; i < thread_num; ++i) { - threads.emplace_back([&thread_num, &count, &mutex, &callback] { - callback(thread_num, count, mutex); - }); - } - for (auto& thread : threads) { - thread.join(); - } - // Check whether all threads reach this point or not - CHECK_EQ(count, thread_num); -} - -TEST(ThreadSpinLock, normalTest) { - for (auto& thread_num : {10, 30, 50, 100, 300, 1000}) { - testNormalImpl( - thread_num, - [](size_t thread_num, size_t& count, paddle::SpinLock& mutex) { - std::lock_guard lock(mutex); - ++count; - }); - } -} diff --git a/paddle/utils/tests/test_StringUtils.cpp b/paddle/utils/tests/test_StringUtils.cpp deleted file mode 100644 index 248f58a7f26e26e82b55110930964cee04fb558b..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_StringUtils.cpp +++ /dev/null @@ -1,23 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/utils/StringUtil.h" - -#include - -TEST(StringUtil, to) { - ASSERT_NEAR(paddle::str::to("12.45"), 12.45, 1e-5); - ASSERT_DEATH_IF_SUPPORTED(paddle::str::to("12.45x23"), ".*"); - ASSERT_DEATH_IF_SUPPORTED(paddle::str::to(""), ".*"); -} diff --git a/paddle/utils/tests/test_Thread.cpp b/paddle/utils/tests/test_Thread.cpp deleted file mode 100644 index 6e2580c4913f0adc7ba1e63c9cebce308775aac6..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_Thread.cpp +++ /dev/null @@ -1,81 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -#include - -using paddle::AsyncThreadPool; // NOLINT - -TEST(AsyncThreadPool, addJob) { - AsyncThreadPool pool(8); - auto a = pool.addJob([] { return 1; }); - auto b = pool.addJob([] { return true; }); - auto c = pool.addJob([] { return false; }); - - ASSERT_EQ(a.get(), 1); - ASSERT_TRUE(b.get()); - ASSERT_FALSE(c.get()); -} - -TEST(AsyncThreadPool, addBatchJob) { - AsyncThreadPool pool(8); - std::atomic counter{0}; - - std::vector jobs; - - for (int i = 0; i < 10000; i++) { - jobs.emplace_back([&] { counter++; }); - } - - pool.addBatchJobs(jobs); - - ASSERT_EQ(counter, 10000); -} - -TEST(AsyncThreadPool, multiThreadAddBatchJob) { - AsyncThreadPool levelOnePool(200); - AsyncThreadPool levelTwoPool(200); - - std::shared_ptr mut = std::make_shared(); - int counter = 0; - const int numMonitors = 300; - const int numSlaves = 300; - std::vector moniterJobs(numMonitors, [&] { - std::vector slaveJobs(numSlaves, [mut, &counter] { - std::lock_guard lk(*mut); - counter++; - }); - levelTwoPool.addBatchJobs(slaveJobs); - }); - levelOnePool.addBatchJobs(moniterJobs); - ASSERT_EQ(counter, numMonitors * numSlaves); -} - -TEST(AsyncThreadPool, addBatchJobWithResults) { - AsyncThreadPool pool(100); - - std::vector> jobs; - const int numJobs = 100; - for (int i = 0; i < numJobs; i++) { - jobs.emplace_back([i] { return i; }); - } - - std::vector res; - pool.addBatchJobs(jobs, res); - - for (int i = 0; i < numJobs; i++) { - ASSERT_EQ(res[i], i); - } -} diff --git a/paddle/utils/tests/test_ThreadBarrier.cpp b/paddle/utils/tests/test_ThreadBarrier.cpp deleted file mode 100644 index 554b1c1d4adce7a0196b304281dcf878a0b6426e..0000000000000000000000000000000000000000 --- a/paddle/utils/tests/test_ThreadBarrier.cpp +++ /dev/null @@ -1,66 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -#include -#include - -#include "paddle/utils/Locks.h" -#include "paddle/utils/Logging.h" -#include "paddle/utils/Util.h" - -DEFINE_int32(test_thread_num, 100, "testing thread number"); - -void testNormalImpl( - size_t thread_num, - const std::function&, - paddle::ThreadBarrier&)>& callback) { - std::mutex mutex; - std::set tids; - paddle::ThreadBarrier barrier(thread_num); - - std::vector threads; - threads.reserve(thread_num); - for (size_t i = 0; i < thread_num; ++i) { - threads.emplace_back([&thread_num, &mutex, &tids, &barrier, &callback] { - callback(thread_num, mutex, tids, barrier); - }); - } - - for (auto& thread : threads) { - thread.join(); - } -} - -TEST(ThreadBarrier, normalTest) { - for (auto& thread_num : {10, 30, 50, 100, 300, 1000}) { - testNormalImpl(thread_num, - [](size_t thread_num, - std::mutex& mutex, - std::set& tids, - paddle::ThreadBarrier& barrier) { - { - std::lock_guard guard(mutex); - tids.insert(std::this_thread::get_id()); - } - barrier.wait(); - // Check whether all threads reach this point or not - CHECK_EQ(tids.size(), thread_num); - }); - } -} diff --git a/patches/grpc/completion_queue.h b/patches/grpc/completion_queue.h new file mode 100644 index 0000000000000000000000000000000000000000..6e92c60ea2db00cc6e227830228888f9a06735c4 --- /dev/null +++ b/patches/grpc/completion_queue.h @@ -0,0 +1,386 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/// A completion queue implements a concurrent producer-consumer queue, with +/// two main API-exposed methods: \a Next and \a AsyncNext. These +/// methods are the essential component of the gRPC C++ asynchronous API. +/// There is also a \a Shutdown method to indicate that a given completion queue +/// will no longer have regular events. This must be called before the +/// completion queue is destroyed. +/// All completion queue APIs are thread-safe and may be used concurrently with +/// any other completion queue API invocation; it is acceptable to have +/// multiple threads calling \a Next or \a AsyncNext on the same or different +/// completion queues, or to call these methods concurrently with a \a Shutdown +/// elsewhere. +/// \remark{All other API calls on completion queue should be completed before +/// a completion queue destructor is called.} +#ifndef GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H +#define GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H + +#include + +#include +#include +#include +#include +#include +#include + +struct grpc_completion_queue; + +namespace grpc { + +template +class ClientReader; +template +class ClientWriter; +template +class ClientReaderWriter; +template +class ServerReader; +template +class ServerWriter; +namespace internal { +template +class ServerReaderWriterBody; +} // namespace internal + +class Channel; +class ChannelInterface; +class ClientContext; +class CompletionQueue; +class Server; +class ServerBuilder; +class ServerContext; +class ServerInterface; + +namespace internal { +class CompletionQueueTag; +class RpcMethod; +template +class RpcMethodHandler; +template +class ClientStreamingHandler; +template +class ServerStreamingHandler; +template +class BidiStreamingHandler; +class UnknownMethodHandler; +template +class TemplatedBidiStreamingHandler; +template +class BlockingUnaryCallImpl; +} // namespace internal + +extern CoreCodegenInterface* g_core_codegen_interface; + +/// A thin wrapper around \ref grpc_completion_queue (see \ref +/// src/core/lib/surface/completion_queue.h). +/// See \ref doc/cpp/perf_notes.md for notes on best practices for high +/// performance servers. +class CompletionQueue : private GrpcLibraryCodegen { + public: + /// Default constructor. Implicitly creates a \a grpc_completion_queue + /// instance. + CompletionQueue() + : CompletionQueue(grpc_completion_queue_attributes{ + GRPC_CQ_CURRENT_VERSION, GRPC_CQ_NEXT, GRPC_CQ_DEFAULT_POLLING}) {} + + /// Wrap \a take, taking ownership of the instance. + /// + /// \param take The completion queue instance to wrap. Ownership is taken. + explicit CompletionQueue(grpc_completion_queue* take); + + /// Destructor. Destroys the owned wrapped completion queue / instance. + ~CompletionQueue() { + if (typeid(*g_core_codegen_interface).hash_code() != + typeid(CoreCodegenInterface).hash_code()) { + g_core_codegen_interface->grpc_completion_queue_destroy(cq_); + } + } + + /// Tri-state return for AsyncNext: SHUTDOWN, GOT_EVENT, TIMEOUT. + enum NextStatus { + SHUTDOWN, ///< The completion queue has been shutdown and fully-drained + GOT_EVENT, ///< Got a new event; \a tag will be filled in with its + ///< associated value; \a ok indicating its success. + TIMEOUT ///< deadline was reached. + }; + + /// Read from the queue, blocking until an event is available or the queue is + /// shutting down. + /// + /// \param tag[out] Updated to point to the read event's tag. + /// \param ok[out] true if read a successful event, false otherwise. + /// + /// Note that each tag sent to the completion queue (through RPC operations + /// or alarms) will be delivered out of the completion queue by a call to + /// Next (or a related method), regardless of whether the operation succeeded + /// or not. Success here means that this operation completed in the normal + /// valid manner. + /// + /// Server-side RPC request: \a ok indicates that the RPC has indeed + /// been started. If it is false, the server has been Shutdown + /// before this particular call got matched to an incoming RPC. + /// + /// Client-side StartCall/RPC invocation: \a ok indicates that the RPC is + /// going to go to the wire. If it is false, it not going to the wire. This + /// would happen if the channel is either permanently broken or + /// transiently broken but with the fail-fast option. (Note that async unary + /// RPCs don't post a CQ tag at this point, nor do client-streaming + /// or bidi-streaming RPCs that have the initial metadata corked option set.) + /// + /// Client-side Write, Client-side WritesDone, Server-side Write, + /// Server-side Finish, Server-side SendInitialMetadata (which is + /// typically included in Write or Finish when not done explicitly): + /// \a ok means that the data/metadata/status/etc is going to go to the + /// wire. If it is false, it not going to the wire because the call + /// is already dead (i.e., canceled, deadline expired, other side + /// dropped the channel, etc). + /// + /// Client-side Read, Server-side Read, Client-side + /// RecvInitialMetadata (which is typically included in Read if not + /// done explicitly): \a ok indicates whether there is a valid message + /// that got read. If not, you know that there are certainly no more + /// messages that can ever be read from this stream. For the client-side + /// operations, this only happens because the call is dead. For the + /// server-sider operation, though, this could happen because the client + /// has done a WritesDone already. + /// + /// Client-side Finish: \a ok should always be true + /// + /// Server-side AsyncNotifyWhenDone: \a ok should always be true + /// + /// Alarm: \a ok is true if it expired, false if it was canceled + /// + /// \return true if got an event, false if the queue is fully drained and + /// shut down. + bool Next(void** tag, bool* ok) { + return (AsyncNextInternal(tag, + ok, + g_core_codegen_interface->gpr_inf_future( + GPR_CLOCK_REALTIME)) != SHUTDOWN); + } + + /// Read from the queue, blocking up to \a deadline (or the queue's shutdown). + /// Both \a tag and \a ok are updated upon success (if an event is available + /// within the \a deadline). A \a tag points to an arbitrary location usually + /// employed to uniquely identify an event. + /// + /// \param tag[out] Upon sucess, updated to point to the event's tag. + /// \param ok[out] Upon sucess, true if a successful event, false otherwise + /// See documentation for CompletionQueue::Next for explanation of ok + /// \param deadline[in] How long to block in wait for an event. + /// + /// \return The type of event read. + template + NextStatus AsyncNext(void** tag, bool* ok, const T& deadline) { + TimePoint deadline_tp(deadline); + return AsyncNextInternal(tag, ok, deadline_tp.raw_time()); + } + + /// EXPERIMENTAL + /// First executes \a F, then reads from the queue, blocking up to + /// \a deadline (or the queue's shutdown). + /// Both \a tag and \a ok are updated upon success (if an event is available + /// within the \a deadline). A \a tag points to an arbitrary location usually + /// employed to uniquely identify an event. + /// + /// \param F[in] Function to execute before calling AsyncNext on this queue. + /// \param tag[out] Upon sucess, updated to point to the event's tag. + /// \param ok[out] Upon sucess, true if read a regular event, false otherwise. + /// \param deadline[in] How long to block in wait for an event. + /// + /// \return The type of event read. + template + NextStatus DoThenAsyncNext(F&& f, void** tag, bool* ok, const T& deadline) { + CompletionQueueTLSCache cache = CompletionQueueTLSCache(this); + f(); + if (cache.Flush(tag, ok)) { + return GOT_EVENT; + } else { + return AsyncNext(tag, ok, deadline); + } + } + + /// Request the shutdown of the queue. + /// + /// \warning This method must be called at some point if this completion queue + /// is accessed with Next or AsyncNext. \a Next will not return false + /// until this method has been called and all pending tags have been drained. + /// (Likewise for \a AsyncNext returning \a NextStatus::SHUTDOWN .) + /// Only once either one of these methods does that (that is, once the queue + /// has been \em drained) can an instance of this class be destroyed. + /// Also note that applications must ensure that no work is enqueued on this + /// completion queue after this method is called. + void Shutdown(); + + /// Returns a \em raw pointer to the underlying \a grpc_completion_queue + /// instance. + /// + /// \warning Remember that the returned instance is owned. No transfer of + /// owership is performed. + grpc_completion_queue* cq() { return cq_; } + + protected: + /// Private constructor of CompletionQueue only visible to friend classes + CompletionQueue(const grpc_completion_queue_attributes& attributes) { + cq_ = g_core_codegen_interface->grpc_completion_queue_create( + g_core_codegen_interface->grpc_completion_queue_factory_lookup( + &attributes), + &attributes, + NULL); + InitialAvalanching(); // reserve this for the future shutdown + } + + private: + // Friend synchronous wrappers so that they can access Pluck(), which is + // a semi-private API geared towards the synchronous implementation. + template + friend class ::grpc::ClientReader; + template + friend class ::grpc::ClientWriter; + template + friend class ::grpc::ClientReaderWriter; + template + friend class ::grpc::ServerReader; + template + friend class ::grpc::ServerWriter; + template + friend class ::grpc::internal::ServerReaderWriterBody; + template + friend class ::grpc::internal::RpcMethodHandler; + template + friend class ::grpc::internal::ClientStreamingHandler; + template + friend class ::grpc::internal::ServerStreamingHandler; + template + friend class ::grpc::internal::TemplatedBidiStreamingHandler; + friend class ::grpc::internal::UnknownMethodHandler; + friend class ::grpc::Server; + friend class ::grpc::ServerContext; + friend class ::grpc::ServerInterface; + template + friend class ::grpc::internal::BlockingUnaryCallImpl; + + /// EXPERIMENTAL + /// Creates a Thread Local cache to store the first event + /// On this completion queue queued from this thread. Once + /// initialized, it must be flushed on the same thread. + class CompletionQueueTLSCache { + public: + CompletionQueueTLSCache(CompletionQueue* cq); + ~CompletionQueueTLSCache(); + bool Flush(void** tag, bool* ok); + + private: + CompletionQueue* cq_; + bool flushed_; + }; + + NextStatus AsyncNextInternal(void** tag, bool* ok, gpr_timespec deadline); + + /// Wraps \a grpc_completion_queue_pluck. + /// \warning Must not be mixed with calls to \a Next. + bool Pluck(internal::CompletionQueueTag* tag) { + auto deadline = + g_core_codegen_interface->gpr_inf_future(GPR_CLOCK_REALTIME); + auto ev = g_core_codegen_interface->grpc_completion_queue_pluck( + cq_, tag, deadline, nullptr); + bool ok = ev.success != 0; + void* ignored = tag; + GPR_CODEGEN_ASSERT(tag->FinalizeResult(&ignored, &ok)); + GPR_CODEGEN_ASSERT(ignored == tag); + // Ignore mutations by FinalizeResult: Pluck returns the C API status + return ev.success != 0; + } + + /// Performs a single polling pluck on \a tag. + /// \warning Must not be mixed with calls to \a Next. + /// + /// TODO: sreek - This calls tag->FinalizeResult() even if the cq_ is already + /// shutdown. This is most likely a bug and if it is a bug, then change this + /// implementation to simple call the other TryPluck function with a zero + /// timeout. i.e: + /// TryPluck(tag, gpr_time_0(GPR_CLOCK_REALTIME)) + void TryPluck(internal::CompletionQueueTag* tag) { + auto deadline = g_core_codegen_interface->gpr_time_0(GPR_CLOCK_REALTIME); + auto ev = g_core_codegen_interface->grpc_completion_queue_pluck( + cq_, tag, deadline, nullptr); + if (ev.type == GRPC_QUEUE_TIMEOUT) return; + bool ok = ev.success != 0; + void* ignored = tag; + // the tag must be swallowed if using TryPluck + GPR_CODEGEN_ASSERT(!tag->FinalizeResult(&ignored, &ok)); + } + + /// Performs a single polling pluck on \a tag. Calls tag->FinalizeResult if + /// the pluck() was successful and returned the tag. + /// + /// This exects tag->FinalizeResult (if called) to return 'false' i.e expects + /// that the tag is internal not something that is returned to the user. + void TryPluck(internal::CompletionQueueTag* tag, gpr_timespec deadline) { + auto ev = g_core_codegen_interface->grpc_completion_queue_pluck( + cq_, tag, deadline, nullptr); + if (ev.type == GRPC_QUEUE_TIMEOUT || ev.type == GRPC_QUEUE_SHUTDOWN) { + return; + } + + bool ok = ev.success != 0; + void* ignored = tag; + GPR_CODEGEN_ASSERT(!tag->FinalizeResult(&ignored, &ok)); + } + + /// Manage state of avalanching operations : completion queue tags that + /// trigger other completion queue operations. The underlying core completion + /// queue should not really shutdown until all avalanching operations have + /// been finalized. Note that we maintain the requirement that an avalanche + /// registration must take place before CQ shutdown (which must be maintained + /// elsehwere) + void InitialAvalanching() { + gpr_atm_rel_store(&avalanches_in_flight_, static_cast(1)); + } + void RegisterAvalanching() { + gpr_atm_no_barrier_fetch_add(&avalanches_in_flight_, + static_cast(1)); + } + void CompleteAvalanching(); + + grpc_completion_queue* cq_; // owned + + gpr_atm avalanches_in_flight_; +}; + +/// A specific type of completion queue used by the processing of notifications +/// by servers. Instantiated by \a ServerBuilder. +class ServerCompletionQueue : public CompletionQueue { + public: + bool IsFrequentlyPolled() { return polling_type_ != GRPC_CQ_NON_LISTENING; } + + private: + grpc_cq_polling_type polling_type_; + friend class ServerBuilder; + /// \param is_frequently_polled Informs the GRPC library about whether the + /// server completion queue would be actively polled (by calling Next() or + /// AsyncNext()). By default all server completion queues are assumed to be + /// frequently polled. + ServerCompletionQueue(grpc_cq_polling_type polling_type) + : CompletionQueue(grpc_completion_queue_attributes{ + GRPC_CQ_CURRENT_VERSION, GRPC_CQ_NEXT, polling_type}), + polling_type_(polling_type) {} +}; + +} // namespace grpc + +#endif // GRPCPP_IMPL_CODEGEN_COMPLETION_QUEUE_H diff --git a/patches/grpc/grpc_library.h b/patches/grpc/grpc_library.h new file mode 100644 index 0000000000000000000000000000000000000000..4870a1cda4b2a6489bc379fe53cf3e9659fffc47 --- /dev/null +++ b/patches/grpc/grpc_library.h @@ -0,0 +1,64 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifndef GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H +#define GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H + +#include + +#include + +namespace grpc { + +class GrpcLibraryInterface { + public: + virtual ~GrpcLibraryInterface() = default; + virtual void init() = 0; + virtual void shutdown() = 0; +}; + +/// Initialized by \a grpc::GrpcLibraryInitializer from +/// +extern GrpcLibraryInterface* g_glip; + +/// Classes that require gRPC to be initialized should inherit from this class. +class GrpcLibraryCodegen { + public: + GrpcLibraryCodegen(bool call_grpc_init = true) : grpc_init_called_(false) { + if (call_grpc_init) { + GPR_CODEGEN_ASSERT(g_glip && + "gRPC library not initialized. See " + "grpc::internal::GrpcLibraryInitializer."); + g_glip->init(); + grpc_init_called_ = true; + } + } + virtual ~GrpcLibraryCodegen() { + if (grpc_init_called_ && + typeid(*g_glip).hash_code() != + typeid(GrpcLibraryInterface).hash_code()) { + GPR_CODEGEN_ASSERT(g_glip && + "gRPC library not initialized. See " + "grpc::internal::GrpcLibraryInitializer."); + g_glip->shutdown(); + } + } + + private: + bool grpc_init_called_; +}; + +} // namespace grpc + +#endif // GRPCPP_IMPL_CODEGEN_GRPC_LIBRARY_H diff --git a/proto/README.md b/proto/README.md new file mode 100644 index 0000000000000000000000000000000000000000..dda7ed7b3c8ea4b541eaafbd0fd239eea789b40e --- /dev/null +++ b/proto/README.md @@ -0,0 +1,3 @@ +## protos in this folder are legacy v2 protos. + +## Please refer to paddle/fluid for latest version. diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index ea25f3ab351ca1feb085a8fbbfe53d8cee397bbf..25900811509aee8b37fdaf09cf902ea2ae3eee57 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -1,4 +1,4 @@ -file(GLOB UTILS_PY_FILES . ./paddle/utils/*.py) +file(GLOB UTILS_PY_FILES . ./paddle/legacy/utils/*.py) file(GLOB_RECURSE FLUID_PY_FILES ./paddle/fluid/*.py) set(PY_FILES paddle/__init__.py ${UTILS_PY_FILES} @@ -91,3 +91,16 @@ endif() install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR} DESTINATION opt/paddle/share/wheels ) + +if(APPLE) + find_program(INSTALL_NAME_TOOL_EXECUTABLE install_name_tool) + if(NOT INSTALL_NAME_TOOL_EXECUTABLE) + message(FATAL_ERROR "install_name_tool not found, please check.\n") + endif() +else(APPLE) + find_program(PATCHELF_EXECUTABLE patchelf) + if(NOT PATCHELF_EXECUTABLE) + message(FATAL_ERROR "patchelf not found, please install it.\n" + "For Ubuntu, the command is: apt-get install -y patchelf.") + endif() +endif(APPLE) diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py index d1cf04161ae4444ebc7da7fbc20e37dafe6c0fb1..241a07a35297e85763781a42696fd727733459a3 100644 --- a/python/paddle/__init__.py +++ b/python/paddle/__init__.py @@ -12,16 +12,16 @@ # See the License for the specific language governing permissions and # limitations under the License. try: - from version import full_version as __version__ - from version import commit as __git_commit__ + from paddle.version import full_version as __version__ + from paddle.version import commit as __git_commit__ except ImportError: import sys - sys.stderr.write('''Warning with import paddle: you should not + sys.stderr.write('''Warning with import paddle: you should not import paddle from the source directory; please install paddlepaddle*.whl firstly.''' ) -import reader -import dataset -import batch +import paddle.reader +import paddle.dataset +import paddle.batch batch = batch.batch diff --git a/python/paddle/batch.py b/python/paddle/batch.py index 317cf037c69f8639e3760fbfce20565127794fcb..008509660739d61245526278735064472b8b06dd 100644 --- a/python/paddle/batch.py +++ b/python/paddle/batch.py @@ -15,7 +15,7 @@ __all__ = ['batch'] -def batch(reader, batch_size): +def batch(reader, batch_size, drop_last=False): """ Create a batched reader. @@ -23,6 +23,8 @@ def batch(reader, batch_size): :type reader: callable :param batch_size: size of each mini-batch :type batch_size: int + :param drop_last: drop the last batch, if the size of last batch is not equal to batch_size. + :type drop_last: bool :return: the batched reader. :rtype: callable """ @@ -35,7 +37,13 @@ def batch(reader, batch_size): if len(b) == batch_size: yield b b = [] - if b: + if drop_last == False and len(b) != 0: yield b + # Batch size check + batch_size = int(batch_size) + if batch_size <= 0: + raise ValueError("batch_size should be a positive integeral value, " + "but got batch_size={}".format(batch_size)) + return batch_reader diff --git a/python/paddle/dataset/__init__.py b/python/paddle/dataset/__init__.py index 3315e826e82a33dfeb9c5223ce196cffb1ae7234..54aa3edc51d3734633ce077a59bd86cec8d09032 100644 --- a/python/paddle/dataset/__init__.py +++ b/python/paddle/dataset/__init__.py @@ -15,20 +15,20 @@ Dataset package. """ -import mnist -import imikolov -import imdb -import cifar -import movielens -import conll05 -import uci_housing -import sentiment -import wmt14 -import wmt16 -import mq2007 -import flowers -import voc2012 -import image +import paddle.dataset.mnist +import paddle.dataset.imikolov +import paddle.dataset.imdb +import paddle.dataset.cifar +import paddle.dataset.movielens +import paddle.dataset.conll05 +import paddle.dataset.uci_housing +import paddle.dataset.sentiment +import paddle.dataset.wmt14 +import paddle.dataset.wmt16 +import paddle.dataset.mq2007 +import paddle.dataset.flowers +import paddle.dataset.voc2012 +import paddle.dataset.image __all__ = [ 'mnist', diff --git a/python/paddle/dataset/cifar.py b/python/paddle/dataset/cifar.py index 07f4dcbdab2fecf84a0a7042a48a8c8a9e5f880d..f6b4ff8fbd0f83b1d652d37c1b2d04efd3c73cbb 100644 --- a/python/paddle/dataset/cifar.py +++ b/python/paddle/dataset/cifar.py @@ -28,11 +28,12 @@ images per class. """ -import cPickle import itertools import numpy import paddle.dataset.common import tarfile +from six.moves import zip +from six.moves import cPickle as pickle __all__ = ['train100', 'test100', 'train10', 'test10', 'convert'] @@ -43,12 +44,12 @@ CIFAR100_URL = URL_PREFIX + 'cifar-100-python.tar.gz' CIFAR100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85' -def reader_creator(filename, sub_name): +def reader_creator(filename, sub_name, cycle=False): def read_batch(batch): data = batch['data'] labels = batch.get('labels', batch.get('fine_labels', None)) assert labels is not None - for sample, label in itertools.izip(data, labels): + for sample, label in zip(data, labels): yield (sample / 255.0).astype(numpy.float32), int(label) def reader(): @@ -56,10 +57,13 @@ def reader_creator(filename, sub_name): names = (each_item.name for each_item in f if sub_name in each_item.name) - for name in names: - batch = cPickle.load(f.extractfile(name)) - for item in read_batch(batch): - yield item + while True: + for name in names: + batch = pickle.load(f.extractfile(name)) + for item in read_batch(batch): + yield item + if not cycle: + break return reader @@ -94,34 +98,40 @@ def test100(): 'test') -def train10(): +def train10(cycle=False): """ CIFAR-10 training set creator. It returns a reader creator, each sample in the reader is image pixels in [0, 1] and label in [0, 9]. + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: Training reader creator :rtype: callable """ return reader_creator( paddle.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), - 'data_batch') + 'data_batch', + cycle=cycle) -def test10(): +def test10(cycle=False): """ CIFAR-10 test set creator. It returns a reader creator, each sample in the reader is image pixels in [0, 1] and label in [0, 9]. + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: Test reader creator. :rtype: callable """ return reader_creator( paddle.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), - 'test_batch') + 'test_batch', + cycle=cycle) def fetch(): diff --git a/python/paddle/dataset/common.py b/python/paddle/dataset/common.py index 68660601c161d2332b17b448fae089506238ba78..6195cc50df338e83bea1f4ad416529464636a33e 100644 --- a/python/paddle/dataset/common.py +++ b/python/paddle/dataset/common.py @@ -20,9 +20,8 @@ import shutil import sys import importlib import paddle.dataset -import cPickle +import six.moves.cPickle as pickle import glob -import cPickle as pickle __all__ = [ 'DATA_HOME', @@ -75,13 +74,13 @@ def download(url, module_name, md5sum, save_name=None): retry_limit = 3 while not (os.path.exists(filename) and md5file(filename) == md5sum): if os.path.exists(filename): - print "file md5", md5file(filename), md5sum + print("file md5", md5file(filename), md5sum) if retry < retry_limit: retry += 1 else: raise RuntimeError("Cannot download {0} within retry limit {1}". format(url, retry_limit)) - print "Cache file %s not found, downloading %s" % (filename, url) + print("Cache file %s not found, downloading %s" % (filename, url)) r = requests.get(url, stream=True) total_length = r.headers.get('content-length') @@ -104,8 +103,9 @@ def download(url, module_name, md5sum, save_name=None): def fetch_all(): - for module_name in filter(lambda x: not x.startswith("__"), - dir(paddle.dataset)): + for module_name in [ + x for x in dir(paddle.dataset) if not x.startswith("__") + ]: if "fetch" in dir( importlib.import_module("paddle.dataset.%s" % module_name)): getattr( @@ -114,8 +114,9 @@ def fetch_all(): def fetch_all_recordio(path): - for module_name in filter(lambda x: not x.startswith("__"), - dir(paddle.dataset)): + for module_name in [ + x for x in dir(paddle.dataset) if not x.startswith("__") + ]: if "convert" in dir( importlib.import_module("paddle.dataset.%s" % module_name)) and \ not module_name == "common": @@ -126,7 +127,7 @@ def fetch_all_recordio(path): "convert")(ds_path) -def split(reader, line_count, suffix="%05d.pickle", dumper=cPickle.dump): +def split(reader, line_count, suffix="%05d.pickle", dumper=pickle.dump): """ you can call the function as: @@ -167,7 +168,7 @@ def split(reader, line_count, suffix="%05d.pickle", dumper=cPickle.dump): def cluster_files_reader(files_pattern, trainer_count, trainer_id, - loader=cPickle.load): + loader=pickle.load): """ Create a reader that yield element from the given files, select a file set according trainer count and trainer_id @@ -188,7 +189,7 @@ def cluster_files_reader(files_pattern, my_file_list = [] for idx, fn in enumerate(file_list): if idx % trainer_count == trainer_id: - print "append file: %s" % fn + print("append file: %s" % fn) my_file_list.append(fn) for fn in my_file_list: with open(fn, "r") as f: @@ -221,7 +222,7 @@ def convert(output_path, reader, line_count, name_prefix): for l in lines: # FIXME(Yancey1989): # dumps with protocol: pickle.HIGHEST_PROTOCOL - writer.write(cPickle.dumps(l)) + writer.write(pickle.dumps(l)) writer.close() lines = [] diff --git a/python/paddle/dataset/conll05.py b/python/paddle/dataset/conll05.py index 4e94ce89892f8e6822c15fdc510805e75dfca988..a97c95d067b876a87f0aa19b2ddd0702a848bd4a 100644 --- a/python/paddle/dataset/conll05.py +++ b/python/paddle/dataset/conll05.py @@ -24,18 +24,19 @@ import tarfile import gzip import itertools import paddle.dataset.common +from six.moves import zip __all__ = ['test, get_dict', 'get_embedding', 'convert'] DATA_URL = 'http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz' DATA_MD5 = '387719152ae52d60422c016e92a742fc' -WORDDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/wordDict.txt' +WORDDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FwordDict.txt' WORDDICT_MD5 = 'ea7fb7d4c75cc6254716f0177a506baa' -VERBDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/verbDict.txt' +VERBDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FverbDict.txt' VERBDICT_MD5 = '0d2977293bbb6cbefab5b0f97db1e77c' -TRGDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/targetDict.txt' +TRGDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FtargetDict.txt' TRGDICT_MD5 = 'd8c7f03ceb5fc2e5a0fa7503a4353751' -EMB_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/emb' +EMB_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2Femb' EMB_MD5 = 'bf436eb0faa1f6f9103017f8be57cdb7' UNK_IDX = 0 @@ -87,12 +88,12 @@ def corpus_reader(data_path, words_name, props_name): sentences = [] labels = [] one_seg = [] - for word, label in itertools.izip(words_file, props_file): + for word, label in zip(words_file, props_file): word = word.strip() label = label.strip().split() if len(label) == 0: # end of sentence - for i in xrange(len(one_seg[0])): + for i in range(len(one_seg[0])): a_kind_lable = [x[i] for x in one_seg] labels.append(a_kind_lable) diff --git a/python/paddle/dataset/flowers.py b/python/paddle/dataset/flowers.py index f082e33be3357fbe405ab1a1ef5e0e601108a363..914dae348bc94d061072543aa14aba2219f4b52d 100644 --- a/python/paddle/dataset/flowers.py +++ b/python/paddle/dataset/flowers.py @@ -28,10 +28,9 @@ Graphics and Image Processing (2008) http://www.robots.ox.ac.uk/~vgg/publications/papers/nilsback08.{pdf,ps.gz}. """ -import cPickle import itertools import functools -from common import download +from .common import download import tarfile import scipy.io as scio from paddle.dataset.image import * @@ -39,6 +38,8 @@ from paddle.reader import * import os import numpy as np from multiprocessing import cpu_count +from six.moves import cPickle as pickle +from six.moves import zip __all__ = ['train', 'test', 'valid'] DATA_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/102flowers.tgz' @@ -76,7 +77,8 @@ def reader_creator(data_file, dataset_name, mapper, buffered_size=1024, - use_xmap=True): + use_xmap=True, + cycle=False): ''' 1. read images from tar file and merge images into batch files in 102flowers.tgz_batch/ @@ -96,6 +98,8 @@ def reader_creator(data_file, :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: data reader :rtype: callable ''' @@ -108,23 +112,27 @@ def reader_creator(data_file, file_list = batch_images_from_tar(data_file, dataset_name, img2label) def reader(): - for file in open(file_list): - file = file.strip() - batch = None - with open(file, 'r') as f: - batch = cPickle.load(f) - data = batch['data'] - labels = batch['label'] - for sample, label in itertools.izip(data, batch['label']): - yield sample, int(label) - 1 + while True: + for file in open(file_list): + file = file.strip() + batch = None + with open(file, 'r') as f: + batch = pickle.load(f) + data = batch['data'] + labels = batch['label'] + for sample, label in zip(data, batch['label']): + yield sample, int(label) - 1 + if not cycle: + break if use_xmap: - return xmap_readers(mapper, reader, cpu_count(), buffered_size) + cpu_num = int(os.environ.get('CPU_NUM', cpu_count())) + return xmap_readers(mapper, reader, cpu_num, buffered_size) else: return map_readers(mapper, reader) -def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): +def train(mapper=train_mapper, buffered_size=1024, use_xmap=True, cycle=False): ''' Create flowers training set reader. It returns a reader, each sample in the reader is @@ -137,17 +145,23 @@ def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: train data reader :rtype: callable ''' return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), TRAIN_FLAG, mapper, - buffered_size, use_xmap) + download(SETID_URL, 'flowers', SETID_MD5), + TRAIN_FLAG, + mapper, + buffered_size, + use_xmap, + cycle=cycle) -def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): +def test(mapper=test_mapper, buffered_size=1024, use_xmap=True, cycle=False): ''' Create flowers test set reader. It returns a reader, each sample in the reader is @@ -160,14 +174,20 @@ def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: test data reader :rtype: callable ''' return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), TEST_FLAG, mapper, - buffered_size, use_xmap) + download(SETID_URL, 'flowers', SETID_MD5), + TEST_FLAG, + mapper, + buffered_size, + use_xmap, + cycle=cycle) def valid(mapper=test_mapper, buffered_size=1024, use_xmap=True): diff --git a/python/paddle/dataset/image.py b/python/paddle/dataset/image.py index 9235c41e9eb95b25a0dc53a494a203e7a4525981..3b3d89c93c48d611dccf6f14958c310a6cac1a7b 100644 --- a/python/paddle/dataset/image.py +++ b/python/paddle/dataset/image.py @@ -36,7 +36,7 @@ except ImportError: cv2 = None import os import tarfile -import cPickle +import six.moves.cPickle as pickle __all__ = [ "load_image_bytes", "load_image", "resize_short", "to_chw", "center_crop", @@ -86,10 +86,10 @@ def batch_images_from_tar(data_file, output = {} output['label'] = labels output['data'] = data - cPickle.dump( + pickle.dump( output, open('%s/batch_%d' % (out_path, file_id), 'w'), - protocol=cPickle.HIGHEST_PROTOCOL) + protocol=pickle.HIGHEST_PROTOCOL) file_id += 1 data = [] labels = [] @@ -97,10 +97,10 @@ def batch_images_from_tar(data_file, output = {} output['label'] = labels output['data'] = data - cPickle.dump( + pickle.dump( output, open('%s/batch_%d' % (out_path, file_id), 'w'), - protocol=cPickle.HIGHEST_PROTOCOL) + protocol=pickle.HIGHEST_PROTOCOL) with open(meta_file, 'a') as meta: for file in os.listdir(out_path): diff --git a/python/paddle/dataset/imdb.py b/python/paddle/dataset/imdb.py index 5ff05b1e9b7f4c42909370a21beb140ecdcd6868..e7fe4e0b7e5832c2bc7ca1307725936a70292c39 100644 --- a/python/paddle/dataset/imdb.py +++ b/python/paddle/dataset/imdb.py @@ -42,13 +42,13 @@ def tokenize(pattern): # sequential access of member files, other than # tarfile.extractfile, which does random access and might # destroy hard disks. - tf = tarf.next() + tf = next(tarf) while tf != None: if bool(pattern.match(tf.name)): # newline and punctuations removal and ad-hoc tokenization. yield tarf.extractfile(tf).read().rstrip("\n\r").translate( None, string.punctuation).lower().split() - tf = tarf.next() + tf = next(tarf) def build_dict(pattern, cutoff): @@ -62,11 +62,11 @@ def build_dict(pattern, cutoff): word_freq[word] += 1 # Not sure if we should prune less-frequent words here. - word_freq = filter(lambda x: x[1] > cutoff, word_freq.items()) + word_freq = [x for x in list(word_freq.items()) if x[1] > cutoff] dictionary = sorted(word_freq, key=lambda x: (-x[1], x[0])) words, _ = list(zip(*dictionary)) - word_idx = dict(zip(words, xrange(len(words)))) + word_idx = dict(list(zip(words, list(range(len(words)))))) word_idx[''] = len(words) return word_idx diff --git a/python/paddle/dataset/imikolov.py b/python/paddle/dataset/imikolov.py index c6c0a0f54373dd068b2c493f6fbc9c8578593aef..bc007c9d3c8e2f1e4ff091f7c2c93eacbbe8d0e0 100644 --- a/python/paddle/dataset/imikolov.py +++ b/python/paddle/dataset/imikolov.py @@ -64,11 +64,11 @@ def build_dict(min_word_freq=50): # remove for now, since we will set it as last index del word_freq[''] - word_freq = filter(lambda x: x[1] > min_word_freq, word_freq.items()) + word_freq = [x for x in list(word_freq.items()) if x[1] > min_word_freq] word_freq_sorted = sorted(word_freq, key=lambda x: (-x[1], x[0])) words, _ = list(zip(*word_freq_sorted)) - word_idx = dict(zip(words, xrange(len(words)))) + word_idx = dict(list(zip(words, list(range(len(words)))))) word_idx[''] = len(words) return word_idx diff --git a/python/paddle/dataset/mnist.py b/python/paddle/dataset/mnist.py index 6a1b8b5fac223c0d134cae69a61a0c2c00bc1feb..ffa9008c80129b80b3807dbab37bc198e59cf5a2 100644 --- a/python/paddle/dataset/mnist.py +++ b/python/paddle/dataset/mnist.py @@ -65,11 +65,17 @@ def reader_creator(image_filename, label_filename, buffer_size): images = images / 255.0 * 2.0 - 1.0 - for i in xrange(buffer_size): + for i in range(buffer_size): yield images[i, :], int(labels[i]) finally: - m.terminate() - l.terminate() + try: + m.terminate() + except: + pass + try: + l.terminate() + except: + pass return reader @@ -111,7 +117,7 @@ def fetch(): paddle.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', TRAIN_IMAGE_MD5) paddle.dataset.common.download(TRAIN_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) paddle.dataset.common.download(TEST_IMAGE_URL, 'mnist', TEST_IMAGE_MD5) - paddle.dataset.common.download(TEST_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) + paddle.dataset.common.download(TEST_LABEL_URL, 'mnist', TEST_LABEL_MD5) def convert(path): diff --git a/python/paddle/dataset/movielens.py b/python/paddle/dataset/movielens.py index ab11716202a8298c182e23b661eb1d2ac74bf4da..056ec2178607329dd6daa1764820c2312bbaed59 100644 --- a/python/paddle/dataset/movielens.py +++ b/python/paddle/dataset/movielens.py @@ -16,7 +16,7 @@ Movielens 1-M dataset. Movielens 1-M dataset contains 1 million ratings from 6000 users on 4000 movies, which was collected by GroupLens Research. This module will download -Movielens 1-M dataset from +Movielens 1-M dataset from http://files.grouplens.org/datasets/movielens/ml-1m.zip and parse training set and test set into paddle reader creators. @@ -187,7 +187,7 @@ def max_movie_id(): Get the maximum value of movie id. """ __initialize_meta_info__() - return reduce(__max_index_info__, MOVIE_INFO.viewvalues()).index + return reduce(__max_index_info__, list(MOVIE_INFO.values())).index def max_user_id(): @@ -195,7 +195,7 @@ def max_user_id(): Get the maximum value of user id. """ __initialize_meta_info__() - return reduce(__max_index_info__, USER_INFO.viewvalues()).index + return reduce(__max_index_info__, list(USER_INFO.values())).index def __max_job_id_impl__(a, b): @@ -210,7 +210,7 @@ def max_job_id(): Get the maximum value of job id. """ __initialize_meta_info__() - return reduce(__max_job_id_impl__, USER_INFO.viewvalues()).job_id + return reduce(__max_job_id_impl__, list(USER_INFO.values())).job_id def movie_categories(): @@ -243,7 +243,7 @@ def unittest(): for test_count, _ in enumerate(test()()): pass - print train_count, test_count + print(train_count, test_count) def fetch(): diff --git a/python/paddle/dataset/mq2007.py b/python/paddle/dataset/mq2007.py index d3b3dd524c34be660c5f2d4fc5ce2fa0420efbc1..cc4d088316dfd490dc9d6b247c66c2495cedf2c3 100644 --- a/python/paddle/dataset/mq2007.py +++ b/python/paddle/dataset/mq2007.py @@ -26,7 +26,7 @@ http://research.microsoft.com/en-us/um/beijing/projects/letor/LETOR4.0/Data/MQ20 import os import functools import rarfile -from common import download +from .common import download import numpy as np # URL = "http://research.microsoft.com/en-us/um/beijing/projects/letor/LETOR4.0/Data/MQ2007.rar" @@ -53,7 +53,7 @@ class Query(object): ---------- query_id : int query_id in dataset, mapping from query to relevance documents - relevance_score : int + relevance_score : int relevance score of query and document pair feature_vector : array, dense feature feature in vector format @@ -92,7 +92,7 @@ class Query(object): sys.stdout.write("expect 48 space split parts, get %d" % (len(parts))) return None - # format : 0 qid:10 1:0.000272 2:0.000000 .... + # format : 0 qid:10 1:0.000272 2:0.000000 .... self.relevance_score = int(parts[0]) self.query_id = int(parts[1].split(':')[1]) for p in parts[2:]: @@ -295,7 +295,7 @@ def __reader__(filepath, format="pairwise", shuffle=False, fill_missing=-1): -------- filename : string fill_missing : fill the missing value. default in MQ2007 is -1 - + Returns ------ yield @@ -330,4 +330,4 @@ if __name__ == "__main__": mytest = functools.partial( __reader__, filepath="MQ2007/MQ2007/Fold1/sample", format="listwise") for label, query in mytest(): - print label, query + print(label, query) diff --git a/python/paddle/dataset/sentiment.py b/python/paddle/dataset/sentiment.py index f5461164fe6b816356e42fc7b7dcf388eccfadfb..953ada057bc114ebbfe39011d2fd3b5b7a2b0d37 100644 --- a/python/paddle/dataset/sentiment.py +++ b/python/paddle/dataset/sentiment.py @@ -43,11 +43,11 @@ def download_data_if_not_yet(): nltk.data.path.append(paddle.dataset.common.DATA_HOME) movie_reviews.categories() except LookupError: - print "Downloading movie_reviews data set, please wait....." + print("Downloading movie_reviews data set, please wait.....") nltk.download( 'movie_reviews', download_dir=paddle.dataset.common.DATA_HOME) - print "Download data set success....." - print "Path is " + nltk.data.find('corpora/movie_reviews').path + print("Download data set success.....") + print("Path is " + nltk.data.find('corpora/movie_reviews').path) def get_word_dict(): @@ -64,7 +64,7 @@ def get_word_dict(): for field in movie_reviews.fileids(category): for words in movie_reviews.words(field): word_freq_dict[words] += 1 - words_sort_list = word_freq_dict.items() + words_sort_list = list(word_freq_dict.items()) words_sort_list.sort(cmp=lambda a, b: b[1] - a[1]) for index, word in enumerate(words_sort_list): words_freq_sorted.append((word[0], index)) @@ -80,7 +80,8 @@ def sort_files(): files_list = list() neg_file_list = movie_reviews.fileids('neg') pos_file_list = movie_reviews.fileids('pos') - files_list = list(chain.from_iterable(zip(neg_file_list, pos_file_list))) + files_list = list( + chain.from_iterable(list(zip(neg_file_list, pos_file_list)))) return files_list diff --git a/python/paddle/dataset/tests/common_test.py b/python/paddle/dataset/tests/common_test.py index e7cc02aa83061599ffefa18de6cb02ac0fc9e9b7..777cd06a19726f8ad73774c958c8cb512808a3aa 100644 --- a/python/paddle/dataset/tests/common_test.py +++ b/python/paddle/dataset/tests/common_test.py @@ -36,7 +36,7 @@ class TestCommon(unittest.TestCase): def test_split(self): def test_reader(): def reader(): - for x in xrange(10): + for x in range(10): yield x return reader @@ -49,7 +49,7 @@ class TestCommon(unittest.TestCase): def test_cluster_file_reader(self): _, temp_path = tempfile.mkstemp() - for x in xrange(5): + for x in range(5): with open(temp_path + '/%05d.test' % x) as f: f.write('%d\n' % x) reader = paddle.dataset.common.cluster_files_reader( @@ -63,7 +63,7 @@ class TestCommon(unittest.TestCase): def test_reader(): def reader(): - for x in xrange(record_num): + for x in range(record_num): yield x return reader diff --git a/python/paddle/dataset/tests/imikolov_test.py b/python/paddle/dataset/tests/imikolov_test.py index 233fd9fc8cea4cd0b5cd052580030fc8c993693c..50f50d947d221686d6308a6ed44cbcff3b10c6f5 100644 --- a/python/paddle/dataset/tests/imikolov_test.py +++ b/python/paddle/dataset/tests/imikolov_test.py @@ -59,7 +59,7 @@ class TestMikolov(unittest.TestCase): self.assertEqual(first_line, read_line) def test_total(self): - _, idx = zip(*WORD_DICT.items()) + _, idx = list(zip(*list(WORD_DICT.items()))) self.assertEqual(sorted(idx)[-1], len(WORD_DICT) - 1) diff --git a/python/paddle/dataset/tests/test_sentiment.py b/python/paddle/dataset/tests/test_sentiment.py index 543f4b7378b583ea3857bf785cf330c43e535c2a..37326517f7b39fb74c694684eb8a547d5f021946 100644 --- a/python/paddle/dataset/tests/test_sentiment.py +++ b/python/paddle/dataset/tests/test_sentiment.py @@ -24,9 +24,8 @@ from nltk.corpus import movie_reviews class TestSentimentMethods(unittest.TestCase): def test_get_word_dict(self): word_dict = st.get_word_dict()[0:10] - test_word_list = [(u',', 0), (u'the', 1), (u'.', 2), (u'a', 3), - (u'and', 4), (u'of', 5), (u'to', 6), (u"'", 7), - (u'is', 8), (u'in', 9)] + test_word_list = [(',', 0), ('the', 1), ('.', 2), ('a', 3), ('and', 4), + ('of', 5), ('to', 6), ("'", 7), ('is', 8), ('in', 9)] for idx, each in enumerate(word_dict): self.assertEqual(each, test_word_list[idx]) self.assertTrue("/root/.cache/paddle/dataset" in nltk.data.path) diff --git a/python/paddle/dataset/uci_housing.py b/python/paddle/dataset/uci_housing.py index fbfa477d055eb5f484989eacce38cee8d617d729..410ca7af0d6d1dc26acbf92fce5e49fce7d3a3bb 100644 --- a/python/paddle/dataset/uci_housing.py +++ b/python/paddle/dataset/uci_housing.py @@ -49,9 +49,12 @@ def feature_range(maximums, minimums): import matplotlib.pyplot as plt fig, ax = plt.subplots() feature_num = len(maximums) - ax.bar(range(feature_num), maximums - minimums, color='r', align='center') + ax.bar(list(range(feature_num)), + maximums - minimums, + color='r', + align='center') ax.set_title('feature scale') - plt.xticks(range(feature_num), feature_names) + plt.xticks(list(range(feature_num)), feature_names) plt.xlim([-1, feature_num]) fig.set_figheight(6) fig.set_figwidth(10) @@ -71,7 +74,7 @@ def load_data(filename, feature_num=14, ratio=0.8): maximums, minimums, avgs = data.max(axis=0), data.min(axis=0), data.sum( axis=0) / data.shape[0] feature_range(maximums[:-1], minimums[:-1]) - for i in xrange(feature_num - 1): + for i in range(feature_num - 1): data[:, i] = (data[:, i] - avgs[i]) / (maximums[i] - minimums[i]) offset = int(data.shape[0] * ratio) UCI_TRAIN_DATA = data[:offset] diff --git a/python/paddle/dataset/wmt14.py b/python/paddle/dataset/wmt14.py index f0908c737874fa7335cca5b5f0cba83190c9f90f..7504474591fa486428d0310f10387818c4cf0300 100644 --- a/python/paddle/dataset/wmt14.py +++ b/python/paddle/dataset/wmt14.py @@ -36,11 +36,10 @@ URL_DEV_TEST = ('http://www-lium.univ-lemans.fr/~schwenk/' MD5_DEV_TEST = '7d7897317ddd8ba0ae5c5fa7248d3ff5' # this is a small set of data for test. The original data is too large and # will be add later. -URL_TRAIN = ('http://paddlepaddle.cdn.bcebos.com/demo/' - 'wmt_shrinked_data/wmt14.tgz') +URL_TRAIN = ('http://paddlemodels.bj.bcebos.com/wmt/wmt14.tgz') MD5_TRAIN = '0791583d57d5beb693b9414c5b36798c' # BLEU of this trained model is 26.92 -URL_MODEL = 'http://paddlepaddle.bj.bcebos.com/demo/wmt_14/wmt14_model.tar.gz' +URL_MODEL = 'http://paddlemodels.bj.bcebos.com/wmt%2Fwmt14.tgz' MD5_MODEL = '0cb4a5366189b6acba876491c8724fa3' START = "" @@ -154,8 +153,8 @@ def get_dict(dict_size, reverse=True): tar_file = paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN) src_dict, trg_dict = __read_to_dict(tar_file, dict_size) if reverse: - src_dict = {v: k for k, v in src_dict.items()} - trg_dict = {v: k for k, v in trg_dict.items()} + src_dict = {v: k for k, v in list(src_dict.items())} + trg_dict = {v: k for k, v in list(trg_dict.items())} return src_dict, trg_dict diff --git a/python/paddle/dataset/wmt16.py b/python/paddle/dataset/wmt16.py index 540d43b692e0f65460f558dd74a52715ff4db68d..4e3c466c38e402cc574e93ef3a5935edf8f9dd3b 100644 --- a/python/paddle/dataset/wmt16.py +++ b/python/paddle/dataset/wmt16.py @@ -70,7 +70,9 @@ def __build_dict(tar_file, dict_size, save_path, lang): fout.write("%s\n%s\n%s\n" % (START_MARK, END_MARK, UNK_MARK)) for idx, word in enumerate( sorted( - word_dict.iteritems(), key=lambda x: x[1], reverse=True)): + iter(list(word_dict.items())), + key=lambda x: x[1], + reverse=True)): if idx + 3 == dict_size: break fout.write("%s\n" % (word[0])) diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index 859605d005328c030980a49a349742772de1cb6d..9aac3c7fc16ae1ded2700662764895385b043130 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -14,75 +14,79 @@ from __future__ import print_function # import all class inside framework into fluid module -import framework -from framework import * +from . import framework +from .framework import * # import all class inside executor into fluid module -import executor -from executor import * - -import trainer -from trainer import Trainer -from trainer import BeginEpochEvent -from trainer import EndEpochEvent -from trainer import BeginStepEvent -from trainer import EndStepEvent - -import inferencer -from inferencer import Inferencer - -import io -import evaluator -import initializer -import layers -import nets -import optimizer -import backward -import regularizer -import average -import metrics -import transpiler -from param_attr import ParamAttr, WeightNormParamAttr -from data_feeder import DataFeeder -from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace -from transpiler import DistributeTranspiler, SimpleDistributeTranspiler, \ - InferenceTranspiler, memory_optimize, release_memory -from concurrency import (Go, make_channel, channel_send, channel_recv, - channel_close, Select) -from lod_tensor import create_lod_tensor, create_random_int_lodtensor -import clip -import profiler -import unique_name -import recordio_writer -import parallel_executor -from parallel_executor import * +from . import executor +from .executor import * + +from . import trainer +from .trainer import Trainer +from .trainer import BeginEpochEvent +from .trainer import EndEpochEvent +from .trainer import BeginStepEvent +from .trainer import EndStepEvent +from .trainer import CheckpointConfig + +from . import inferencer +from .inferencer import Inferencer + +from . import io +from . import evaluator +from . import initializer +from . import layers +from . import contrib +from . import nets +from . import optimizer +from . import backward +from . import regularizer +from . import average +from . import metrics +from . import transpiler +from .param_attr import ParamAttr, WeightNormParamAttr +from .data_feeder import DataFeeder +from .core import LoDTensor, LoDTensorArray, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope +from .transpiler import DistributeTranspiler, InferenceTranspiler, \ + memory_optimize, release_memory, DistributeTranspilerConfig +from .lod_tensor import create_lod_tensor, create_random_int_lodtensor +from . import clip +from . import profiler +from . import unique_name +from . import recordio_writer +from . import parallel_executor +from .parallel_executor import * +from paddle.fluid.layers.math_op_patch import monkey_patch_variable Tensor = LoDTensor -__all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \ - trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \ - parallel_executor.__all__ + lod_tensor.__all__ + [ - 'io', - 'initializer', - 'layers', - 'transpiler' - 'nets', - 'optimizer', - 'learning_rate_decay', - 'backward', - 'regularizer', - 'LoDTensor', - 'CPUPlace', - 'CUDAPlace', - 'CUDAPinnedPlace', - 'Tensor', - 'ParamAttr', - 'WeightNormParamAttr', - 'DataFeeder', - 'clip', - 'profiler', - 'unique_name', - 'recordio_writer', - ] +__all__ = framework.__all__ + executor.__all__ + \ + trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \ + parallel_executor.__all__ + lod_tensor.__all__ + [ + 'io', + 'initializer', + 'layers', + 'contrib', + 'transpiler', + 'nets', + 'optimizer', + 'learning_rate_decay', + 'backward', + 'regularizer', + 'LoDTensor', + 'LoDTensorArray', + 'CPUPlace', + 'CUDAPlace', + 'CUDAPinnedPlace', + 'Tensor', + 'ParamAttr', + 'WeightNormParamAttr', + 'DataFeeder', + 'clip', + 'profiler', + 'unique_name', + 'recordio_writer', + 'Scope', + ] def __bootstrap__(): @@ -93,8 +97,8 @@ def __bootstrap__(): None """ import sys - import core import os + from . import core in_test = 'unittest' in sys.modules @@ -116,11 +120,18 @@ def __bootstrap__(): read_env_flags = [ 'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir', - 'eager_delete_scope' + 'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb', + 'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads', + "dist_threadpool_size", 'cpu_deterministic' ] + if core.is_compiled_with_dist(): + read_env_flags.append('rpc_deadline') + read_env_flags.append('rpc_server_profile_period') + read_env_flags.append('rpc_server_profile_path') + if core.is_compiled_with_cuda(): read_env_flags += [ - 'fraction_of_gpu_memory_to_use', 'cudnn_algo_use_autotune' + 'fraction_of_gpu_memory_to_use', 'cudnn_deterministic' ] core.init_gflags([sys.argv[0]] + ["--tryfromenv=" + ",".join(read_env_flags)]) @@ -131,5 +142,5 @@ def __bootstrap__(): # TODO(panyx0718): Avoid doing complex initialization logic in __init__.py. # Consider paddle.init(args) or paddle.main(args) -layers.monkey_patch_variable() +monkey_patch_variable() __bootstrap__() diff --git a/python/paddle/fluid/annotations.py b/python/paddle/fluid/annotations.py new file mode 100644 index 0000000000000000000000000000000000000000..15e7976354f2a22065f1723bfa696d056181dac2 --- /dev/null +++ b/python/paddle/fluid/annotations.py @@ -0,0 +1,39 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function +import functools +import sys + +__all__ = ['deprecated'] + + +def deprecated(since, instead, extra_message=""): + def decorator(func): + err_msg = "API {0} is deprecated since {1}. Please use {2} instead.".format( + func.__name__, since, instead) + if len(extra_message) != 0: + err_msg += "\n" + err_msg += extra_message + + @functools.wraps(func) + def wrapper(*args, **kwargs): + print(err_msg, file=sys.stderr) + return func(*args, **kwargs) + + wrapper.__doc__ += "\n " + wrapper.__doc__ += err_msg + return wrapper + + return decorator diff --git a/python/paddle/fluid/average.py b/python/paddle/fluid/average.py index 6abe8233b07c484494848c566e9898600a7d8f5c..358e24df31bb517604481bb48b9180e579f8460d 100644 --- a/python/paddle/fluid/average.py +++ b/python/paddle/fluid/average.py @@ -36,6 +36,25 @@ def _is_number_or_matrix_(var): class WeightedAverage(object): + """ + Calculate weighted average. + + The average calculating is accomplished via Python totally. + They do not change Paddle's Program, nor do anything to + modify NN model's configuration. They are completely + wrappers of Python functions. + + Examples: + .. code-block:: python + avg = fluid.average.WeightedAverage() + avg.add(value=2.0, weight=1) + avg.add(value=4.0, weight=2) + avg.eval() + + # The result is 3.333333333. + # For (2.0 * 1 + 4.0 * 2) / (1 + 2) = 3.333333333 + """ + def __init__(self): warnings.warn( "The %s is deprecated, please use fluid.metrics.Accuracy instead." % diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py index 4f9622d04dc98f41b503ceb780802d2a4e4c58a0..fd6a76dd0cfa347328d87093884e5cd324395497 100644 --- a/python/paddle/fluid/backward.py +++ b/python/paddle/fluid/backward.py @@ -16,12 +16,10 @@ from paddle.fluid import framework as framework from . import core import collections import copy -import unique_name +import six +from . import unique_name -__all__ = [ - 'append_backward', - 'calc_gradient', -] +__all__ = ['append_backward'] def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None): @@ -47,17 +45,25 @@ def _create_op_desc_(op_type, inputs, outputs, attrs): """ op_desc = core.OpDesc() op_desc.set_type(op_type) - for para, args in inputs.iteritems(): - op_desc.set_input(para, args) - for para, args in outputs.iteritems(): - op_desc.set_output(para, args) + for para, args in list(inputs.items()): + op_desc.set_input( + para, + list( + map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg, + args))) + for para, args in list(outputs.items()): + op_desc.set_output( + para, + list( + map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg, + args))) op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName() if op_role_attr_name not in attrs: attrs[ op_role_attr_name] = core.op_proto_and_checker_maker.OpRole.Backward - for name, val in attrs.iteritems(): + for name, val in list(attrs.items()): if isinstance(val, framework.Block): op_desc.set_block_attr(name, val.desc) else: @@ -108,7 +114,9 @@ def _strip_grad_suffix_(name): e.g. x@GRAD ==> x y@GRAD@RENAME@1 ==> y """ - pos = name.find(core.grad_var_suffix()) + if isinstance(name, six.text_type): + name = name.encode() + pos = name.find(six.b(core.grad_var_suffix())) return name[:pos] if pos != -1 else name @@ -117,13 +125,16 @@ def _append_grad_suffix_(name): Append grad suffix to the given variable name e.g. x ==> x@GRAD """ - return name + core.grad_var_suffix() + if isinstance(name, six.text_type): + name = name.encode() + return name + six.b(core.grad_var_suffix()) def _addup_repetitive_outputs_(op_descs): """ In backward part, an variable may be the output of more than one ops. - In this case, the variable should be the accumulation of all the outputs. + And one op may yield its multiple outputs to the same variable. + In these cases, the variable should be the accumulation of all the outputs. `sum_op`s are added to implement the accumulate. """ pending_sum_ops = [] @@ -132,37 +143,55 @@ def _addup_repetitive_outputs_(op_descs): for idx, op_desc in enumerate(op_descs): for var_name in op_desc.input_arg_names(): if len(renamed_vars[var_name]) > 1: - pending_sum_ops.append( - (_create_op_desc_("sum", {"X": renamed_vars[var_name]}, - {"Out": [var_name]}, {}), idx)) + pending_sum_ops.append((_create_op_desc_( + "sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]}, + {"use_mkldnn": False}), idx)) renamed_vars[var_name] = [var_name] - for var_name in op_desc.output_arg_names(): - if var_name == core.empty_var_name( - ) or var_name in op_desc.input_arg_names(): - # empty variable or inplace op - continue - if len(renamed_vars[var_name]) == 0: - # it's the first time we get the variable - renamed_vars[var_name] = [var_name] - else: - if len(renamed_vars[var_name]) == 1: + for param_idx, param_name in enumerate(op_desc.output_names()): + arg_names = op_desc.output(param_name) + for arg_idx, var_name in enumerate(arg_names): + if var_name == core.empty_var_name( + ) or var_name in op_desc.input_arg_names(): + # empty variable or inplace op + continue + if len(renamed_vars[var_name]) == 0: + # it's the first time we get the variable + renamed_vars[var_name] = [var_name] + else: + if len(renamed_vars[var_name]) == 1: + new_name = var_name + "@RENAME@" + \ + str(var_rename_count[var_name]) + var_rename_count[var_name] += 1 + # rename original var_name + renamed_vars[var_name][0] = new_name + _rename_arg_(op_descs, var_name, new_name, 0, idx) + _rename_arg_(pending_sum_ops, var_name, new_name) + + for p in op_desc.output_names()[:param_idx]: + p_arg_names = op_desc.output(p) + if var_name in p_arg_names: + op_desc.set_output(p, [ + new_name if x == var_name else x + for x in p_arg_names + ]) + + arg_names = [ + new_name if x == var_name else x + for x in arg_names[:arg_idx] + ] + arg_names[arg_idx:] + new_name = var_name + "@RENAME@" + \ - str(var_rename_count[var_name]) + str(var_rename_count[var_name]) var_rename_count[var_name] += 1 - # rename original var_name - renamed_vars[var_name][0] = new_name - _rename_arg_(op_descs, var_name, new_name, 0, idx) - _rename_arg_(pending_sum_ops, var_name, new_name) - - new_name = var_name + "@RENAME@" + \ - str(var_rename_count[var_name]) - var_rename_count[var_name] += 1 - op_desc.rename_output(var_name, new_name) - renamed_vars[var_name].append(new_name) - for var_name, inputs in renamed_vars.iteritems(): + arg_names[arg_idx] = new_name + op_desc.set_output(param_name, arg_names) + renamed_vars[var_name].append(new_name) + + for var_name, inputs in list(renamed_vars.items()): if len(inputs) > 1: - pending_sum_ops.append((_create_op_desc_( - "sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs))) + pending_sum_ops.append( + (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]}, + {"use_mkldnn": False}), len(op_descs))) # sum_op descs are sorted according to their insert position for p in reversed(pending_sum_ops): op_descs.insert(p[1], p[0]) @@ -182,16 +211,19 @@ def _remove_no_grad_branch_(op_descs, no_grad_set): out_arg_names = op_desc.output_arg_names() if len(out_arg_names) == 0 or _all_in_set_(out_arg_names, no_grad_set): return True - if _all_in_set_( - filter(lambda name: name.find(core.grad_var_suffix()) != -1, - op_desc.input_arg_names()), no_grad_set): + if _all_in_set_([ + name for name in op_desc.input_arg_names() + if name.find(core.grad_var_suffix()) != -1 + ], no_grad_set): no_grad_set.update(out_arg_names) return True return False # Remove ops whose outputs are all in no_grad_dict - op_descs = filter( - lambda op_desc: not _op_can_be_removed_(op_desc, no_grad_set), op_descs) + op_descs = [ + op_desc for op_desc in op_descs + if not _op_can_be_removed_(op_desc, no_grad_set) + ] # Insert fill_zeros_like_op to_insert = [] for idx, op_desc in enumerate(op_descs): @@ -201,12 +233,12 @@ def _remove_no_grad_branch_(op_descs, no_grad_set): "X": [_strip_grad_suffix_(arg)] }, {"Out": [arg]}, {}), idx)) - map(lambda p: op_descs.insert(p[1], p[0]), reversed(to_insert)) + list([op_descs.insert(p[1], p[0]) for p in reversed(to_insert)]) return op_descs -import proto.framework_pb2 as framework_pb2 +from .proto import framework_pb2 def serialize_op_decs(op_desc): @@ -228,8 +260,10 @@ def _callback_lookup_(op): if op.type == 'parallel_do' and op.attr('use_nccl'): all_vars = op.block.vars param_names = set(op.input('parameters')) - param_names = filter(lambda name: all_vars[name].stop_gradient is False, - param_names) + param_names = [ + name for name in param_names + if all_vars[name].stop_gradient is False + ] param_grad_names = [n + "@GRAD" for n in param_names] class ParallelDoCallBack(object): @@ -310,9 +344,9 @@ def _append_backward_ops_(block, grad_sub_block_list = [] # If the op has its own sub-block, deal with the sub-block first if op.has_attr("sub_block"): - sub_block = program.block(op.block_attr("sub_block")) + sub_block = program.block(op.block_attr_id("sub_block")) grad_sub_block = program.create_block() - grad_sub_block.set_forward_block_idx(sub_block.idx) + grad_sub_block._set_forward_block_idx(sub_block.idx) cb = _callback_lookup_(op) if cb is not None: if callbacks is None: @@ -372,7 +406,7 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map): for op_idx in range(start_op_idx, block.desc.op_size()): op_desc = block.desc.op(op_idx) if op_desc.has_attr("sub_block"): - sub_block = block.program.block(op_desc.block_attr("sub_block")) + sub_block = block.program.block(op_desc.block_attr_id("sub_block")) _append_backward_vars_(sub_block, 0, grad_to_var, grad_info_map) new_vars = set() # create new gradient variables @@ -383,7 +417,7 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map): continue block.desc.var(grad_var_name) new_vars.add(grad_var_name) - if not grad_to_var.has_key(grad_var_name): + if grad_var_name not in grad_to_var: continue grad_info_map[grad_to_var[grad_var_name]] = (grad_var_name, block) # infer_shape and infer_type @@ -411,7 +445,7 @@ def _rename_grad_(block, start_op_idx, grad_to_var, target_grad_map): op_desc.rename_output(name, new_name) var_map[name] = new_name - for g, ng in var_map.iteritems(): + for g, ng in list(var_map.items()): if g in grad_to_var: grad_to_var[ng] = grad_to_var[g] grad_to_var.pop(g) @@ -423,7 +457,7 @@ def _get_stop_gradients_(program): for block in program.blocks: assert isinstance(block, framework.Block) block_no_grad_set = set() - for var in block.vars.itervalues(): + for var in list(block.vars.values()): assert isinstance(var, framework.Variable) if var.stop_gradient: block_no_grad_set.add(_append_grad_suffix_(var.name)) @@ -434,18 +468,65 @@ def _get_stop_gradients_(program): def append_backward(loss, parameter_list=None, no_grad_set=None, callbacks=None): """ - Append backward part to main_program + Append backward part to main_program. - Args: - loss(Variable): The variable generated by cost function. - parameter_list(list[string]): Parameters that need to be updated by - optimizer. If None, it means all parameters need to be updated. - no_grad_set(set): Variables that have no gradients in Block 0. - All variables with `step_gradient=True` from all blocks will be - automatically added. + A complete neural network training is made up of forward and backward + propagation. However, when we configure a network, we only need to + specify its forwrd part. The backward part is generated automatically + according to the forward part by this function. - Return: - (list[(Variable,Variable)]): list of (parameter, gradient) pair. + In most cases, users do not need to invoke this function manually. It + will be automatically invoked by the optimizer's `minimize` function. + + Args: + loss(Variable): The loss variable of the network. + parameter_list(list[string]|None): Names of parameters that need + to be updated by optimizers. + If it is None, all parameters + will be updated. + Default: None + no_grad_set(set|None): Variables in the Block 0 whose gradients + should be ignored. All variables with + `step_gradient=True` from all blocks will + be automatically added into this set. + Default: None + callbacks(list[callable object]|None): The callbacks are used for + doing some custom jobs during + backward part building. All + callable objects in it will + be invoked once each time a + new gradient operator is added + into the program. The callable + object must has two input + parameters: 'block' and 'context'. + The 'block' is the block which + the new gradient operator will + be added to. The 'context' is a + map, whose keys are gradient + variable names and values are + corresponding original variables. + In addition to this, the 'context' + has another special key-value pair: + the key is string '__current_op_desc__' + and the value is the op_desc of the + gradient operator who has just + triggered the callable object. + + Returns: + list[(Variable,Variable)]: Pairs of parameter and its + corresponding gradients. The key is the parameter and the + value is gradient variable. + + Raises: + AssertionError: If `loss` is not an instance of Variable. + + Examples: + .. code-block:: python + + # network configuration code + # ... + avg_loss = fluid.layers.mean(loss) + param_grad_list = fluid.backward.append_backward(loss=avg_loss) """ assert isinstance(loss, framework.Variable) @@ -472,7 +553,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, no_grad_set = set() no_grad_set = copy.copy(no_grad_set) no_grad_dict = _get_stop_gradients_(program) - no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set)) + no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set))) grad_info_map = dict() root_block = program.block(0) @@ -495,7 +576,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0])) op_path = _find_op_path_(root_block, [loss], [], block_no_grad_set) - no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set)) + no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set))) _append_backward_ops_(root_block, op_path, root_block, no_grad_dict, grad_to_var, callbacks) @@ -508,9 +589,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map) program.current_block_idx = current_block_idx - program.sync_with_cpp() - # FIXME(zcd): prevent loss.grad optimized by mem_opt. - loss.block.var(_append_grad_suffix_(loss.name)).persistable = True + program._sync_with_cpp() if parameter_list is not None: parameters = parameter_list @@ -636,7 +715,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None): no_grad_set = set() no_grad_set = copy.copy(no_grad_set) no_grad_dict = _get_stop_gradients_(prog) - no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set)) + no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set))) fwd_op_num = block.desc.op_size() @@ -670,7 +749,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None): block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0])) op_path = _find_op_path_(block, targets, inputs, block_no_grad_set) - no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set)) + no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set))) grad_to_var = dict() grad_info_map = dict() _append_backward_ops_(block, op_path, block, no_grad_dict, grad_to_var) @@ -681,7 +760,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None): _rename_grad_(block, fwd_op_num, grad_to_var, target_grad_map) _append_backward_vars_(block, fwd_op_num, grad_to_var, grad_info_map) - prog.sync_with_cpp() + prog._sync_with_cpp() grad_vars = [] for input_var in inputs: diff --git a/python/paddle/fluid/clip.py b/python/paddle/fluid/clip.py index 66c3fc6b66d61bc9578f84594409ad0f24c99910..4b0a792f784fffcce3f911d3e7448b472d39f8e1 100644 --- a/python/paddle/fluid/clip.py +++ b/python/paddle/fluid/clip.py @@ -13,10 +13,11 @@ # limitations under the License. import copy +import six import functools -import layers -import framework +from . import layers +from . import framework from . import core __all__ = [ @@ -24,8 +25,6 @@ __all__ = [ 'GradientClipByValue', 'GradientClipByNorm', 'GradientClipByGlobalNorm', - 'append_gradient_clip_ops', - 'error_clip_callback', ] @@ -33,11 +32,30 @@ class BaseErrorClipAttr(object): def __str__(self): raise NotImplementedError() - def append_clip_op(self, block, grad_name): + def _append_clip_op(self, block, grad_name): raise NotImplementedError() class ErrorClipByValue(BaseErrorClipAttr): + """ + Clips tensor values to the range [min, max]. + + Given a tensor t, this operation clips its value to min and max inplace. + + - Any values less than min are set to min. + - Any values greater than max are set to max. + + Args: + max (float): The maximum value to clip by. + min (float, optional): The minimum value to clip by. if not set by user, \ + will be set to -max by framework. + + Examples: + .. code-block:: python + + var = fluid.framework.Variable(..., error_clip=ErrorClipByValue(max=5.0), ...) + """ + def __init__(self, max, min=None): max = float(max) if min is None: @@ -50,7 +68,7 @@ class ErrorClipByValue(BaseErrorClipAttr): def __str__(self): return "ByValue, min=%f, max=%f" % (self.min, self.max) - def append_clip_op(self, block, grad_name): + def _append_clip_op(self, block, grad_name): clip_op_desc = block.desc.append_op() clip_op_desc.set_type("clip") clip_op_desc.set_input("X", [grad_name]) @@ -63,9 +81,8 @@ def error_clip_callback(block, context): # the context is a grad_to_var map grad_to_var = context op_desc = block.desc.op(block.desc.op_size() - 1) - for grad_n in filter(lambda n: grad_to_var.has_key(n), - op_desc.output_arg_names()): - fwd_var = block.var_recursive(grad_to_var[grad_n]) + for grad_n in [n for n in op_desc.output_arg_names() if n in grad_to_var]: + fwd_var = block._var_recursive(grad_to_var[grad_n]) error_clip = getattr(fwd_var, "error_clip", None) if not (error_clip is None or isinstance(error_clip, BaseErrorClipAttr)): @@ -73,17 +90,17 @@ def error_clip_callback(block, context): "Variable's error_clip should be an instance of BaseErrorClipAttr or None." ) if error_clip is not None: - error_clip.append_clip_op(block, grad_n) + error_clip._append_clip_op(block, grad_n) class BaseGradientClipAttr(object): def __str__(self): raise NotImplementedError() - def process_context(self, context, param, grad): + def _process_context(self, context, param, grad): raise NotImplementedError() - def create_operators(self, param, grad): + def _create_operators(self, param, grad): raise NotImplementedError() @@ -91,14 +108,39 @@ class NullGradientClipAttr(BaseGradientClipAttr): def __str__(self): return "Null" - def process_context(self, context, param, grad): + def _process_context(self, context, param, grad): pass - def create_operators(self, param, grad): + def _create_operators(self, param, grad): return param, grad class GradientClipByValue(BaseGradientClipAttr): + """ + Clips gradient values to the range [min, max]. + + Given a tensor t, this operation clips its value to min and max inplace. + + - Any values less than min are set to min. + - Any values greater than max are set to max. + + Args: + max (float): The maximum value to clip by. + min (float, optional): The minimum value to clip by. if not set by user, \ + will be set to -max by framework. + + Examples: + .. code-block:: python + + w_param_attrs = ParamAttr(name=None, + initializer=UniformInitializer(low=-1.0, high=1.0, seed=0), + learning_rate=1.0, + regularizer=L1Decay(1.0), + trainable=True, + clip=GradientClipByValue(-1.0, 1.0)) + y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs) + """ + def __init__(self, max, min=None): max = float(max) if min is None: @@ -111,33 +153,102 @@ class GradientClipByValue(BaseGradientClipAttr): def __str__(self): return "ByValue, min=%f, max=%f" % (self.min, self.max) - def process_context(self, context, param, grad): + def _process_context(self, context, param, grad): pass - def create_operators(self, param, grad): + def _create_operators(self, param, grad): new_grad = layers.clip(x=grad, min=self.min, max=self.max) return param, new_grad class GradientClipByNorm(BaseGradientClipAttr): + """ + Clips tensor values to a maximum L2-norm. + + This operator limits the L2 norm of the input :math:`X` within :math:`max\_norm`. + If the L2 norm of :math:`X` is less than or equal to :math:`max\_norm`, :math:`Out` + will be the same as :math:`X`. If the L2 norm of :math:`X` is greater than + :math:`max\_norm`, :math:`X` will be linearly scaled to make the L2 norm of + :math:`Out` equal to :math:`max\_norm`, as shown in the following formula: + + .. math:: + + Out = \\frac{max\_norm * X}{norm(X)}, + + where :math:`norm(X)` represents the L2 norm of :math:`X`. + + Args: + clip_norm (float): The maximum norm value + + Examples: + .. code-block:: python + + w_param_attrs = ParamAttr(name=None, + initializer=UniformInitializer(low=-1.0, high=1.0, seed=0), + learning_rate=1.0, + regularizer=L1Decay(1.0), + trainable=True, + clip=GradientClipByNorm(clip_norm=2.0)) + y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs) + + """ + def __init__(self, clip_norm): self.clip_norm = clip_norm def __str__(self): return "ByNorm, clip_norm=%f" % self.clip_norm - def process_context(self, context, param, grad): + def _process_context(self, context, param, grad): pass - def create_operators(self, param, grad): + def _create_operators(self, param, grad): new_grad = layers.clip_by_norm(x=grad, max_norm=self.clip_norm) return param, new_grad class GradientClipByGlobalNorm(BaseGradientClipAttr): + """ + Clips values of multiple tensors by the ratio of the sum of their norms. + + Given a list of tensors t_list, and a clipping ratio clip_norm, this + operation returns a list of clipped tensors list_clipped and the global + norm (global_norm) of all tensors in t_list. + + To perform the clipping, the values :math:`t\_list[i]` are set to: + + .. math:: + + t\_list[i] = t\_list[i] * \\frac{clip\_norm}{\max(global\_norm, clip\_norm)} + + where: + + .. math:: + + global\_norm = \sqrt{\sum_{i=0}^{N-1}(l2norm(t\_list[i]))^2} + + If :math:`clip\_norm > global\_norm` then the entries in t_list remain as they are, + otherwise they're all shrunk by the global ratio. + + Args: + clip_norm (float): The maximum norm value + group_name (str, optional): The group name for this clip. + + Examples: + .. code-block:: python + + p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip) + + with fluid.program_guard(main_program=prog_clip): + fluid.clip.set_gradient_clip( + fluid.clip.GradientClipByGlobalNorm(clip_norm=2.0)) + p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip) + + """ + def __init__(self, clip_norm, group_name="default_group"): - if not isinstance(group_name, basestring): - raise TypeError("'group_name' must be a basestring.") + if not isinstance(group_name, six.string_types): + raise TypeError("'group_name' must be a %s." % (six.string_types)) self.clip_norm = clip_norm self.group_name = group_name @@ -146,7 +257,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr): return "ByGlobalNorm, group_name=%s, clip_norm=%f" % (self.group_name, self.clip_norm) - def process_context(self, context, param, grad): + def _process_context(self, context, param, grad): if self.group_name not in context: context[self.group_name] = [] context[self.group_name + "_clip_value"] = self.clip_norm @@ -163,7 +274,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr): self.context = context - def create_operators(self, param, grad): + def _create_operators(self, param, grad): group_scale_name = self.group_name + "_scale" if group_scale_name not in self.context: group_norm_var = layers.sums(input=self.context[self.group_name]) @@ -173,7 +284,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr): x=clip_var, y=layers.elementwise_max( x=clip_var, y=group_norm_var)) - assert group_scale_var.shape == (1L, ) + assert group_scale_var.shape == (1, ) self.context[group_scale_name] = group_scale_var new_grad = layers.elementwise_mul( @@ -183,15 +294,16 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr): def set_gradient_clip(clip, param_list=None, program=None): """ - To specify parameters that require gradient clip. - Args: - clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr, - which describes the type and detailed attributes of required gradient clip. - param_list(list, None by default): Parameters that require gradient clip. - It can be a list of parameter or a list of parameter's name. - When it's None, all parameters in the program will be included. - program(Program, None by default): The program where parameters are. - Will be the default main program when assigned with None. + To specify parameters that require gradient clip. + + Args: + clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr, + which describes the type and detailed attributes of required gradient clip. + param_list(list(Variable)): Parameters that require gradient clip. + It can be a list of parameter or a list of parameter's name. + When it's None, all parameters in the program will be included. + program(Program): The program where parameters are. + Will be the default main program when assigned with None. """ if not isinstance(clip, BaseGradientClipAttr): raise TypeError( @@ -201,7 +313,7 @@ def set_gradient_clip(clip, param_list=None, program=None): program = framework.default_main_program() if param_list is None: param_list = program.block(0).all_parameters() - if all(isinstance(elem, basestring) for elem in param_list): + if all(isinstance(elem, six.string_types) for elem in param_list): param_list = [program.block(0).var(elem) for elem in param_list] if not all(isinstance(elem, framework.Parameter) for elem in param_list): raise TypeError( @@ -212,10 +324,12 @@ def set_gradient_clip(clip, param_list=None, program=None): param.gradient_clip_attr = copy.deepcopy(clip) -def append_gradient_clip_ops(param_grad): +def append_gradient_clip_ops(param_grads): context = dict() - for p, g in param_grad: - with p.block.program.optimized_guard(p): + for p, g in param_grads: + if g is None: + continue + with p.block.program.optimized_guard([p, g]): clip_attr = getattr(p, 'gradient_clip_attr', NullGradientClipAttr()) if clip_attr is None: clip_attr = NullGradientClipAttr() @@ -224,12 +338,14 @@ def append_gradient_clip_ops(param_grad): "clip attribute should be an instance of BaseGradientClipAttr" ) - clip_attr.process_context(context=context, param=p, grad=g) + clip_attr._process_context(context=context, param=p, grad=g) res = [] - for p, g in param_grad: - with p.block.program.optimized_guard(p): - res.append(clip_attr.create_operators(param=p, grad=g)) + for p, g in param_grads: + if g is None: + continue + with p.block.program.optimized_guard([p, g]): + res.append(clip_attr._create_operators(param=p, grad=g)) return res diff --git a/python/paddle/fluid/concurrency.py b/python/paddle/fluid/concurrency.py index 470dd0df524936a773f6e740c8079f0efa8ef7b4..676a52a917dd1f9700ec38de32932938ec339be5 100644 --- a/python/paddle/fluid/concurrency.py +++ b/python/paddle/fluid/concurrency.py @@ -12,15 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. -from layers.control_flow import BlockGuard, equal +from .layers.control_flow import BlockGuard, equal from .framework import Operator -from layer_helper import LayerHelper, unique_name -from layers import fill_constant -import core +from .layer_helper import LayerHelper, unique_name +from .layers import fill_constant +from . import core __all__ = [ - 'Go', 'make_channel', 'channel_send', 'channel_recv', 'channel_close', - 'Select' + 'make_channel', 'channel_send', 'channel_recv', 'channel_close', 'Select' ] @@ -35,10 +34,10 @@ class Go(BlockGuard): def __exit__(self, exc_type, exc_val, exc_tb): if exc_type is not None: return False - self.construct_go_op() + self._construct_go_op() return super(Go, self).__exit__(exc_type, exc_val, exc_tb) - def construct_go_op(self): + def _construct_go_op(self): main_program = self.helper.main_program go_block = main_program.current_block() parent_block = main_program.block(main_program.current_block() @@ -69,8 +68,10 @@ class Go(BlockGuard): parent_block.append_op( type='go', inputs={ - 'X': - [parent_block.var_recursive(x_name) for x_name in x_name_list] + 'X': [ + parent_block._var_recursive(x_name) + for x_name in x_name_list + ] }, outputs={}, attrs={'sub_block': go_block}) @@ -259,7 +260,7 @@ class Select(BlockGuard): if var_name in intermediate ] - X = [select_block.var_recursive(x_name) for x_name in params] + X = [select_block._var_recursive(x_name) for x_name in params] # Needs to be used by `equal` inside the cases block. X.append(self.case_to_execute) diff --git a/python/paddle/fluid/contrib/__init__.py b/python/paddle/fluid/contrib/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..58f2da1c3ba2f84602e7a18c7b1c78d1f0d2ede1 --- /dev/null +++ b/python/paddle/fluid/contrib/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from . import decoder +from .decoder import * +from . import memory_usage_calc +from .memory_usage_calc import * + +__all__ = decoder.__all__ + memory_usage_calc.__all__ diff --git a/python/paddle/fluid/contrib/decoder/__init__.py b/python/paddle/fluid/contrib/decoder/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6343c1543d206f82e605c5c986fa91d70c467113 --- /dev/null +++ b/python/paddle/fluid/contrib/decoder/__init__.py @@ -0,0 +1,18 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from . import beam_search_decoder +from .beam_search_decoder import * + +__all__ = beam_search_decoder.__all__ diff --git a/python/paddle/fluid/contrib/decoder/beam_search_decoder.py b/python/paddle/fluid/contrib/decoder/beam_search_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d268a948f7a2cf038a419c95521b81088ed8215f --- /dev/null +++ b/python/paddle/fluid/contrib/decoder/beam_search_decoder.py @@ -0,0 +1,840 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This module provides a general beam search decoder API for RNN based decoders. +The purpose of this API is to allow users to highly customize the behavior +within their RNN decoder(vanilla RNN, LSTM, attention + LSTM, future etc.), +without using the low level API such as while ops. + +This API is still under active development and may change drastically. +""" + +import contextlib +import numpy as np +import six + +from ... import layers +from ...framework import Variable +from ... import core +from ... import framework, unique_name +from ...layer_helper import LayerHelper + +__all__ = ['InitState', 'StateCell', 'TrainingDecoder', 'BeamSearchDecoder'] + + +class _DecoderType: + TRAINING = 1 + BEAM_SEARCH = 2 + + +class InitState(object): + """ + The initial hidden state object. The state objects holds a variable, and may + use it to initialize the hidden state cell of RNN. Usually used as input to + `StateCell` class. + + Args: + init (Variable): The initial variable of the hidden state. If set None, + the variable will be created as a tensor with constant value based + on `shape` and `value` param. + shape (tuple|list): If `init` is None, new Variable's shape. Default + None. + value (float): If `init` is None, new Variable's value. Default None. + init_boot (Variable): If provided, the initial variable will be created + with the same shape as this variable. + need_reorder (bool): If set true, the init will be sorted by its lod + rank within its batches. This should be used if `batch_size > 1`. + dtype (np.dtype|core.VarDesc.VarType|str): Data type of the initial + variable. + + Returns: + An initialized state object. + + Examples: + See `StateCell`. + """ + + def __init__(self, + init=None, + shape=None, + value=0.0, + init_boot=None, + need_reorder=False, + dtype='float32'): + if init is not None: + self._init = init + elif init_boot is None: + raise ValueError( + 'init_boot must be provided to infer the shape of InitState .\n') + else: + self._init = layers.fill_constant_batch_size_like( + input=init_boot, value=value, shape=shape, dtype=dtype) + + self._shape = shape + self._value = value + self._need_reorder = need_reorder + self._dtype = dtype + + @property + def value(self): + return self._init + + @property + def need_reorder(self): + return self._need_reorder + + +class _MemoryState(object): + def __init__(self, state_name, rnn_obj, init_state): + self._state_name = state_name # each is a rnn.memory + self._rnn_obj = rnn_obj + self._state_mem = self._rnn_obj.memory( + init=init_state.value, need_reorder=init_state.need_reorder) + + def get_state(self): + return self._state_mem + + def update_state(self, state): + self._rnn_obj.update_memory(self._state_mem, state) + + +class _ArrayState(object): + def __init__(self, state_name, block, init_state): + self._state_name = state_name + self._block = block + + self._state_array = self._block.create_var( + name=unique_name.generate('array_state_array'), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=init_state.value.dtype) + + self._counter = self._block.create_var( + name=unique_name.generate('array_state_counter'), + type=core.VarDesc.VarType.LOD_TENSOR, + dtype='int64') + + # initialize counter + self._block.append_op( + type='fill_constant', + inputs={}, + outputs={'Out': [self._counter]}, + attrs={ + 'shape': [1], + 'dtype': self._counter.dtype, + 'value': float(0.0), + 'force_cpu': True + }) + + self._counter.stop_gradient = True + + # write initial state + block.append_op( + type='write_to_array', + inputs={'X': init_state.value, + 'I': self._counter}, + outputs={'Out': self._state_array}) + + def get_state(self): + state = layers.array_read(array=self._state_array, i=self._counter) + return state + + def update_state(self, state): + layers.increment(x=self._counter, value=1, in_place=True) + layers.array_write(state, array=self._state_array, i=self._counter) + + +class StateCell(object): + """ + The state cell class stores the hidden state of the RNN cell. A typical RNN + cell has one or more hidden states, and one or more step inputs. This class + allows you to defines the name of hidden states as well as step inputs, and + their associated variables. + + Args: + inputs (dict): A feeding dict of {name(str) : Variable}. It specifies + the names of step inputs for RNN cell, and the associated variables. + The variable could initially be None and set manually during each + RNN step. + states (dict): A feeding dict of {name(str) : InitState object}. It + specifies the names of hidden states and their initialized state. + out_state (str): A string that specifies the name of hidden state that + will be used to compute the score in beam search process. + name (str): The name of the RNN cell. Default None. + + Raises: + `ValueError`: If the initial state is not an instance of InitState, or + the out_state is not in the dict of states. + + Returns: + StateCell: The initialized StateCell object. + + Examples: + .. code-block:: python + hidden_state = InitState(init=encoder_out, need_reorder=True) + state_cell = StateCell( + inputs={'current_word': None}, + states={'h': hidden_state}, + out_state='h') + """ + + def __init__(self, inputs, states, out_state, name=None): + self._helper = LayerHelper('state_cell', name=name) + self._cur_states = {} + self._state_names = [] + for state_name, state in six.iteritems(states): + if not isinstance(state, InitState): + raise ValueError('state must be an InitState object.') + self._cur_states[state_name] = state + self._state_names.append(state_name) + self._inputs = inputs # inputs is place holder here + self._cur_decoder_obj = None + self._in_decoder = False + self._states_holder = {} + self._switched_decoder = False + self._state_updater = None + self._out_state = out_state + if self._out_state not in self._cur_states: + raise ValueError('out_state must be one state in states') + + def _enter_decoder(self, decoder_obj): + if self._in_decoder == True or self._cur_decoder_obj is not None: + raise ValueError('StateCell has already entered a decoder.') + self._in_decoder = True + self._cur_decoder_obj = decoder_obj + self._switched_decoder = False + + def _leave_decoder(self, decoder_obj): + if not self._in_decoder: + raise ValueError('StateCell not in decoder, ' + 'invalid leaving operation.') + + if self._cur_decoder_obj != decoder_obj: + raise ValueError('Inconsistent decoder object in StateCell.') + + self._in_decoder = False + self._cur_decoder_obj = None + self._switched_decoder = False + + def _switch_decoder(self): # lazy switch + if not self._in_decoder: + raise ValueError('StateCell must be enter a decoder.') + + if self._switched_decoder: + raise ValueError('StateCell already done switching.') + + for state_name in self._state_names: + if state_name not in self._states_holder: + state = self._cur_states[state_name] + + if not isinstance(state, InitState): + raise ValueError('Current type of state is %s, should be ' + 'an InitState object.' % type(state)) + + self._states_holder[state_name] = {} + + if self._cur_decoder_obj.type == _DecoderType.TRAINING: + self._states_holder[state_name][id(self._cur_decoder_obj)] \ + = _MemoryState(state_name, + self._cur_decoder_obj.dynamic_rnn, + state) + elif self._cur_decoder_obj.type == _DecoderType.BEAM_SEARCH: + self._states_holder[state_name][id(self._cur_decoder_obj)] \ + = _ArrayState(state_name, + self._cur_decoder_obj._parent_block(), + state) + else: + raise ValueError('Unknown decoder type, only support ' + '[TRAINING, BEAM_SEARCH]') + + # Read back, since current state should be LoDTensor + self._cur_states[state_name] = \ + self._states_holder[state_name][ + id(self._cur_decoder_obj)].get_state() + + self._switched_decoder = True + + def get_state(self, state_name): + """ + The getter of state object. Find the state variable by its name. + + Args: + state_name (str): A string of the state's name. + + Returns: + The associated state object. + """ + if self._in_decoder and not self._switched_decoder: + self._switch_decoder() + + if state_name not in self._cur_states: + raise ValueError( + 'Unknown state %s. Please make sure _switch_decoder() ' + 'invoked.' % state_name) + + return self._cur_states[state_name] + + def get_input(self, input_name): + """ + The getter of input variable. Find the input variable by its name. + + Args: + input_name (str): The string of the input's name. + + Returns: + The associated input variable. + """ + if input_name not in self._inputs or self._inputs[input_name] is None: + raise ValueError('Invalid input %s.' % input_name) + return self._inputs[input_name] + + def set_state(self, state_name, state_value): + """ + The setter of the state variable. Change the variable of the given + `state_name`. + + Args: + state_name (str): The name of the state to change. + state_value (Var): The variable of the new state. + """ + self._cur_states[state_name] = state_value + + def state_updater(self, updater): + """ + Set up the updater to update the hidden state every RNN step. The + behavior of updater could be customized by users. The updater should be + a function that takes a `StateCell` object as input and update the + hidden state within it. The hidden state could be accessed through + `get_state` method. + + Args: + updater (func): the updater to update the state cell. + """ + self._state_updater = updater + + def _decorator(state_cell): + if state_cell == self: + raise TypeError('Updater should only accept a StateCell object ' + 'as argument.') + updater(state_cell) + + return _decorator + + def compute_state(self, inputs): + """ + Provide the step input of RNN cell, and compute the new hidden state + with updater and give step input. + + Args: + inputs (dict): A feed dict, {name(str): Variable}. name should be + the names of step inputs for this RNN cell, and Variable should be + the associated variables. + + Examples: + .. code-block:: python + state_cell.compute_state(inputs={'x': current_word}) + """ + if self._in_decoder and not self._switched_decoder: + self._switch_decoder() + + for input_name, input_value in six.iteritems(inputs): + if input_name not in self._inputs: + raise ValueError('Unknown input %s. ' + 'Please make sure %s in input ' + 'place holder.' % (input_name, input_name)) + self._inputs[input_name] = input_value + self._state_updater(self) + + def update_states(self): + """ + Update and record state information after each RNN step. + """ + if self._in_decoder and not self._switched_decoder: + self._switched_decoder() + + for state_name, decoder_state in six.iteritems(self._states_holder): + if id(self._cur_decoder_obj) not in decoder_state: + raise ValueError('Unknown decoder object, please make sure ' + 'switch_decoder been invoked.') + decoder_state[id(self._cur_decoder_obj)].update_state( + self._cur_states[state_name]) + + def out_state(self): + """ + Get the output state variable. This must be called after update_states. + + Returns: + The output variable of the RNN cell. + """ + return self._cur_states[self._out_state] + + +class TrainingDecoder(object): + """ + A decoder that can only be used for training. The decoder could be + initialized with a `StateCell` object. The computation within the RNN cell + could be defined with decoder's block. + + Args: + state_cell (StateCell): A StateCell object that handles the input and + state variables. + name (str): The name of this decoder. Default None. + + Returns: + TrainingDecoder: The initialized TrainingDecoder object. + + Examples: + .. code-block:: python + decoder = TrainingDecoder(state_cell) + with decoder.block(): + current_word = decoder.step_input(trg_embedding) + decoder.state_cell.compute_state(inputs={'x': current_word}) + current_score = layers.fc(input=decoder.state_cell.get_state('h'), + size=32, + act='softmax') + decoder.state_cell.update_states() + decoder.output(current_score) + """ + BEFORE_DECODER = 0 + IN_DECODER = 1 + AFTER_DECODER = 2 + + def __init__(self, state_cell, name=None): + self._helper = LayerHelper('training_decoder', name=name) + self._status = TrainingDecoder.BEFORE_DECODER + self._dynamic_rnn = layers.DynamicRNN() + self._type = _DecoderType.TRAINING + self._state_cell = state_cell + self._state_cell._enter_decoder(self) + + @contextlib.contextmanager + def block(self): + """ + Define the behavior of the decoder for each RNN time step. + """ + if self._status != TrainingDecoder.BEFORE_DECODER: + raise ValueError('decoder.block() can only be invoked once') + self._status = TrainingDecoder.IN_DECODER + with self._dynamic_rnn.block(): + yield + self._status = TrainingDecoder.AFTER_DECODER + self._state_cell._leave_decoder(self) + + @property + def state_cell(self): + self._assert_in_decoder_block('state_cell') + return self._state_cell + + @property + def dynamic_rnn(self): + return self._dynamic_rnn + + @property + def type(self): + return self._type + + def step_input(self, x): + """ + Set the input variable as a step input to the RNN cell. For example, + in machine translation, each time step we read one word from the target + sentences, then the target sentence is a step input to the RNN cell. + + Args: + x (Variable): the variable to be used as step input. + + Returns: + Variable: The variable as input of current step. + + Examples: + .. code-block:: python + current_word = decoder.step_input(trg_embedding) + """ + self._assert_in_decoder_block('step_input') + return self._dynamic_rnn.step_input(x) + + def static_input(self, x): + """ + Set the input variable as a static input of RNN cell. In contrast to + step input, this variable will be used as a whole within the RNN decode + loop and will not be scattered into time steps. + + Args: + x (Variable): the variable to be used as static input. + + Returns: + Variable: The variable as input of current step. + + Examples: + .. code-block:: python + encoder_vec = decoder.static_input(encoded_vector) + """ + self._assert_in_decoder_block('static_input') + return self._dynamic_rnn.static_input(x) + + def __call__(self, *args, **kwargs): + """ + Get the output of RNN. This API should only be invoked after RNN.block() + + Returns: + Variable: The specified output of the RNN cell. + """ + if self._status != TrainingDecoder.AFTER_DECODER: + raise ValueError('Output of training decoder can only be visited ' + 'outside the block.') + return self._dynamic_rnn(*args, **kwargs) + + def output(self, *outputs): + """ + Set the output variable of the RNN cell. + + Args: + *outputs (Variables): a series of variables that treated as output + of the RNN cell. + + Examples: + .. code-block:: python + out = fluid.layers.fc(input=h, + size=32, + bias_attr=True, + act='softmax') + decoder.output(out) + """ + self._assert_in_decoder_block('output') + self._dynamic_rnn.output(*outputs) + + def _assert_in_decoder_block(self, method): + if self._status != TrainingDecoder.IN_DECODER: + raise ValueError('%s should be invoked inside block of ' + 'TrainingDecoder object.' % method) + + +class BeamSearchDecoder(object): + """ + A beam search decoder that can be used for inference. The decoder should be + initialized with a `StateCell` object. The decode process can be defined + within its block. + + Args: + state_cell (StateCell): A StateCell object that handles the input and + state variables. + init_ids (Variable): The init beam search token ids. + init_scores (Variable): The associated score of each id. + target_dict_dim (int): Size of dictionary. + word_dim (int): Word embedding dimension. + input_var_dict (dict): A feeding dict to feed the required input + variables to the state cell. It will be used by state_cell 's + compute method. Default empty. + topk_size (int): The topk size used for beam search. Default 50. + max_len (int): The maximum allowed length of the generated sentence. + Default 100. + beam_size (int): The beam width of beam search decode. Default 1. + end_id (int): The id of end token within beam search. + name (str): The name of this decoder. Default None. + + Returns: + BeamSearchDecoder: A initialized BeamSearchDecoder object. + + Examples: + .. code-block:: python + decoder = BeamSearchDecoder( + state_cell=state_cell, + init_ids=init_ids, + init_scores=init_scores, + target_dict_dim=target_dict_dim, + word_dim=word_dim, + init_var_dict={}, + topk_size=topk_size, + sparse_emb=IS_SPARSE, + max_len=max_length, + beam_size=beam_size, + end_id=1, + name=None + ) + decoder.decode() + translation_ids, translation_scores = decoder() + """ + BEFORE_BEAM_SEARCH_DECODER = 0 + IN_BEAM_SEARCH_DECODER = 1 + AFTER_BEAM_SEARCH_DECODER = 2 + + def __init__(self, + state_cell, + init_ids, + init_scores, + target_dict_dim, + word_dim, + input_var_dict={}, + topk_size=50, + sparse_emb=True, + max_len=100, + beam_size=1, + end_id=1, + name=None): + self._helper = LayerHelper('beam_search_decoder', name=name) + self._counter = layers.zeros(shape=[1], dtype='int64') + self._counter.stop_gradient = True + self._type = _DecoderType.BEAM_SEARCH + self._max_len = layers.fill_constant( + shape=[1], dtype='int64', value=max_len) + self._cond = layers.less_than( + x=self._counter, + y=layers.fill_constant( + shape=[1], dtype='int64', value=max_len)) + self._while_op = layers.While(self._cond) + self._state_cell = state_cell + self._state_cell._enter_decoder(self) + self._status = BeamSearchDecoder.BEFORE_BEAM_SEARCH_DECODER + self._zero_idx = layers.fill_constant( + shape=[1], value=0, dtype='int64', force_cpu=True) + self._array_dict = {} + self._array_link = [] + self._ids_array = None + self._scores_array = None + self._beam_size = beam_size + self._end_id = end_id + + self._init_ids = init_ids + self._init_scores = init_scores + self._target_dict_dim = target_dict_dim + self._topk_size = topk_size + self._sparse_emb = sparse_emb + self._word_dim = word_dim + self._input_var_dict = input_var_dict + + @contextlib.contextmanager + def block(self): + """ + Define the behavior of the decoder for each RNN time step. + """ + if self._status != BeamSearchDecoder.BEFORE_BEAM_SEARCH_DECODER: + raise ValueError('block() can only be invoke once.') + + self._status = BeamSearchDecoder.IN_BEAM_SEARCH_DECODER + + with self._while_op.block(): + yield + with layers.Switch() as switch: + with switch.case(self._cond): + layers.increment(x=self._counter, value=1.0, in_place=True) + + for value, array in self._array_link: + layers.array_write( + x=value, i=self._counter, array=array) + + layers.less_than( + x=self._counter, y=self._max_len, cond=self._cond) + + self._status = BeamSearchDecoder.AFTER_BEAM_SEARCH_DECODER + self._state_cell._leave_decoder(self) + + @property + def type(self): + return self._type + + def early_stop(self): + """ + Stop the generation process in advance. Could be used as "break". + """ + layers.fill_constant( + shape=[1], value=0, dtype='bool', force_cpu=True, out=self._cond) + + def decode(self): + """ + Set up the computation within the decoder. Then you could call the + decoder to get the result of beam search decode. If you want to define + a more specific decoder, you could override this function. + + Examples: + .. code-block:: python + decoder.decode() + translation_ids, translation_scores = decoder() + """ + with self.block(): + prev_ids = self.read_array(init=self._init_ids, is_ids=True) + prev_scores = self.read_array( + init=self._init_scores, is_scores=True) + prev_ids_embedding = layers.embedding( + input=prev_ids, + size=[self._target_dict_dim, self._word_dim], + dtype='float32', + is_sparse=self._sparse_emb) + + feed_dict = {} + update_dict = {} + + for init_var_name, init_var in six.iteritems(self._input_var_dict): + if init_var_name not in self.state_cell._inputs: + raise ValueError('Variable ' + init_var_name + + ' not found in StateCell!\n') + + read_var = self.read_array(init=init_var) + update_dict[init_var_name] = read_var + feed_var_expanded = layers.sequence_expand(read_var, + prev_scores) + feed_dict[init_var_name] = feed_var_expanded + + for state_str in self._state_cell._state_names: + prev_state = self.state_cell.get_state(state_str) + prev_state_expanded = layers.sequence_expand(prev_state, + prev_scores) + self.state_cell.set_state(state_str, prev_state_expanded) + + for i, input_name in enumerate(self._state_cell._inputs): + if input_name not in feed_dict: + feed_dict[input_name] = prev_ids_embedding + + self.state_cell.compute_state(inputs=feed_dict) + current_state = self.state_cell.out_state() + current_state_with_lod = layers.lod_reset( + x=current_state, y=prev_scores) + scores = layers.fc(input=current_state_with_lod, + size=self._target_dict_dim, + act='softmax') + topk_scores, topk_indices = layers.topk(scores, k=self._topk_size) + accu_scores = layers.elementwise_add( + x=layers.log(x=topk_scores), + y=layers.reshape( + prev_scores, shape=[-1]), + axis=0) + selected_ids, selected_scores = layers.beam_search( + prev_ids, + prev_scores, + topk_indices, + accu_scores, + self._beam_size, + end_id=1, + level=0) + + with layers.Switch() as switch: + with switch.case(layers.is_empty(selected_ids)): + self.early_stop() + with switch.default(): + self.state_cell.update_states() + self.update_array(prev_ids, selected_ids) + self.update_array(prev_scores, selected_scores) + for update_name, var_to_update in six.iteritems( + update_dict): + self.update_array(var_to_update, feed_dict[update_name]) + + def read_array(self, init, is_ids=False, is_scores=False): + """ + Read an array to get the decoded ids and scores generated by previous + RNN step. At the first step of RNN, the init variable mut be used to + initialize the array. + + Args: + init (Variable): The initial variable for first step usage. init + must be provided. + is_ids (bool): Specify whether the variable is an id. + is_scores (bool): Specify whether the variable is a score. + + Returns: + The associated variable generated during previous RNN steps. + + Examples: + .. code-block:: python + prev_ids = decoder.read_array(init=init_ids, is_ids=True) + prev_scores = decoder.read_array(init=init_scores, is_scores=True) + """ + self._assert_in_decoder_block('read_array') + + if is_ids and is_scores: + raise ValueError('Shouldn\'t mark current array be ids array and' + 'scores array at the same time.') + + if not isinstance(init, Variable): + raise TypeError('The input argument `init` must be a Variable.') + + parent_block = self._parent_block() + array = parent_block.create_var( + name=unique_name.generate('beam_search_decoder_array'), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=init.dtype) + parent_block.append_op( + type='write_to_array', + inputs={'X': init, + 'I': self._zero_idx}, + outputs={'Out': array}) + + if is_ids: + self._ids_array = array + elif is_scores: + self._scores_array = array + + read_value = layers.array_read(array=array, i=self._counter) + self._array_dict[read_value.name] = array + return read_value + + def update_array(self, array, value): + """ + Store the value generated in current step in an array for each RNN step. + This array could be accessed by read_array method. + + Args: + array (Variable): The array to append the new variable to. + value (Variable): The newly generated value to be stored. + """ + self._assert_in_decoder_block('update_array') + + if not isinstance(array, Variable): + raise TypeError( + 'The input argument `array` of must be a Variable.') + if not isinstance(value, Variable): + raise TypeError('The input argument `value` of must be a Variable.') + + array = self._array_dict.get(array.name, None) + if array is None: + raise ValueError('Please invoke read_array before update_array.') + self._array_link.append((value, array)) + + def __call__(self): + """ + Run the decode process and return the final decode result. + + Returns: + A tuple of decoded (id, score) pairs. id is a Variable that holds + the generated tokens, and score is a Variable with the same shape + as id, holds the score for each generated token. + """ + if self._status != BeamSearchDecoder.AFTER_BEAM_SEARCH_DECODER: + raise ValueError('Output of BeamSearchDecoder object can ' + 'only be visited outside the block.') + return layers.beam_search_decode( + ids=self._ids_array, + scores=self._scores_array, + beam_size=self._beam_size, + end_id=self._end_id) + + @property + def state_cell(self): + self._assert_in_decoder_block('state_cell') + return self._state_cell + + def _parent_block(self): + """ + Getter of parent block. + + Returns: + The parent block of decoder. + """ + program = self._helper.main_program + parent_block_idx = program.current_block().parent_idx + if parent_block_idx < 0: + raise ValueError('Invalid block with index %d.' % parent_block_idx) + parent_block = program.block(parent_block_idx) + return parent_block + + def _assert_in_decoder_block(self, method): + if self._status != BeamSearchDecoder.IN_BEAM_SEARCH_DECODER: + raise ValueError('%s should be invoked inside block of ' + 'BeamSearchDecoder object.' % method) diff --git a/python/paddle/fluid/contrib/memory_usage_calc.py b/python/paddle/fluid/contrib/memory_usage_calc.py new file mode 100644 index 0000000000000000000000000000000000000000..5da846edb63c28efd791fdfac4046cfa56c24181 --- /dev/null +++ b/python/paddle/fluid/contrib/memory_usage_calc.py @@ -0,0 +1,102 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This module privides a memory usage calculate function for user. +The purpose of this API is to allow users to estimate memory usage of +a program under a special batch size, then user can set appropriate +batch size to fully utilize a GPU. + +This API is still under active development and may change drastically. +""" + +from .. import core +from ..framework import Program, Variable + +__all__ = ['memory_usage'] + +dtype_to_size = { + core.VarDesc.VarType.FP16: 2, + core.VarDesc.VarType.FP32: 4, + core.VarDesc.VarType.FP64: 8, + core.VarDesc.VarType.INT16: 2, + core.VarDesc.VarType.INT32: 4, + core.VarDesc.VarType.INT64: 8, + core.VarDesc.VarType.BOOL: 1, + core.VarDesc.VarType.UINT8: 1, +} + +DEBUG = False + + +def memory_usage(program, batch_size): + """ + Get the estimate memory usage of program with input batch size. + + Args: + program(Program): The current Program. + batch_size(int): The current input data batch_size. + + Returns: + min_total_memory(float): the estimate memory usage lower bound. + max_total_memory(float): the estimate memory usage upper bound. + unit_str(string): the unit of estimate usage result. + + Examples: + + >>> import paddle.fluid as fluid + >>> lower_usage, upper_usage, unit = fluid.contrib.memory_usage( + fluid.default_main_program(), batch_size=10) + >>> print "memory usage is about %.3f - %.3f %s" % \ + (lower_usage, upper_usage, unit) + + """ + + # Parameters check + if not isinstance(program, Program): + raise TypeError( + "Calculating Memory Usage requires Program as its Parameter." + "But you passed in %s" % (type(prgram))) + if batch_size <= 0: + raise ValueError("The batch size need to be positive.") + + # Get the var_name list of first block and calculate + total_memory = 0.0 + for var in program.global_block().vars.itervalues(): + data_count = 1 + for x in var.shape: + if x == -1: + data_count *= batch_size + else: + data_count *= x + var_memory = data_count * dtype_to_size[var.dtype] + if DEBUG: + print "%s memory usage: %d" % (var.name, var_memory) + total_memory += var_memory + if DEBUG: + print "total memory usage: %.2f" % (total_memory) + + # Convert appropriate unit + unit_str = "B" + if total_memory > 1024: + total_memory /= 1024 + unit_str = "KB" + if total_memory > 1024: + total_memory /= 1024 + unit_str = "MB" + + # Append extra memory consumption (5% - 10%) + min_total_memory = total_memory * 1.05 + max_total_memory = total_memory * 1.1 + + return min_total_memory, max_total_memory, unit_str diff --git a/python/paddle/fluid/data_feeder.py b/python/paddle/fluid/data_feeder.py index 7940dabcfb03cc9eb46f678365685a6e99bcceec..9452cf0e2a3a2eddb761149466bfc1ee3d23dfd9 100644 --- a/python/paddle/fluid/data_feeder.py +++ b/python/paddle/fluid/data_feeder.py @@ -12,13 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function -import core +from . import core import numpy -import six.moves as six +import os +import six +from six.moves import zip, range, xrange import multiprocessing -from framework import Variable, default_main_program +from .framework import Variable, default_main_program __all__ = ['DataFeeder'] @@ -28,6 +29,13 @@ class DataToLoDTensorConverter(object): self.place = place self.lod_level = lod_level self.shape = shape + negtive_count = 0 + for s in self.shape: + if s < 0: + negtive_count += 1 + if negtive_count > 1: + self.shape = None + break if dtype == core.VarDesc.VarType.FP32: self.dtype = 'float32' elif dtype == core.VarDesc.VarType.INT64: @@ -45,8 +53,8 @@ class DataToLoDTensorConverter(object): self.data = [] self.lod = [] - for i in six.range(lod_level): - self.lod.append([0]) + for i in six.moves.range(lod_level): + self.lod.append([]) def feed(self, data): self._feed_impl_(data, self.lod, self.lod_level) @@ -55,21 +63,77 @@ class DataToLoDTensorConverter(object): if lod_level == 0: self.data.append(data) else: - cur_lod_len = len(data) - lod[0].append(lod[0][-1] + cur_lod_len) + lod[0].append(len(data)) for each_data in data: self._feed_impl_(each_data, lod[1:], lod_level - 1) def done(self): - arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape) + arr = numpy.array(self.data, dtype=self.dtype) + if self.shape: + arr = arr.reshape(self.shape) t = core.LoDTensor() t.set(arr, self.place) if self.lod_level > 0: - t.set_lod(self.lod) + t.set_recursive_sequence_lengths(self.lod) return t class DataFeeder(object): + """ + DataFeeder converts the data that returned by a reader into a data + structure that can feed into Executor and ParallelExecutor. The reader + usually returns a list of mini-batch data entries. Each data entry in + the list is one sample. Each sample is a list or a tuple with one + feature or multiple features. + + The simple usage shows below: + + .. code-block:: python + + place = fluid.CPUPlace() + img = fluid.layers.data(name='image', shape=[1, 28, 28]) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([img, label], fluid.CPUPlace()) + result = feeder.feed([([0] * 784, [9]), ([1] * 784, [1])]) + + + If you want to feed data into GPU side separately in advance when you + use multi-GPU to train a model, you can use `decorate_reader` function. + + .. code-block:: python + + place=fluid.CUDAPlace(0) + feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) + reader = feeder.decorate_reader( + paddle.batch(flowers.train(), batch_size=16)) + + Args: + feed_list(list): The Variables or Variables'name that will + feed into model. + place(Place): place indicates feed data into CPU or GPU, if you want to + feed data into GPU, please using `fluid.CUDAPlace(i)` (`i` represents + the GPU id), or if you want to feed data into CPU, please using + `fluid.CPUPlace()`. + program(Program): The Program that will feed data into, if program + is None, it will use default_main_program(). Default None. + + Raises: + ValueError: If some Variable is not in this Program. + + Examples: + .. code-block:: python + + # ... + place = fluid.CPUPlace() + feed_list = [ + main_program.global_block().var(var_name) for var_name in feed_vars_name + ] # feed_vars_name is a list of variables' name. + feeder = fluid.DataFeeder(feed_list, place) + for data in reader(): + outs = exe.run(program=main_program, + feed=feeder.feed(data)) + """ + def __init__(self, feed_list, place, program=None): self.feed_dtypes = [] self.feed_names = [] @@ -78,7 +142,7 @@ class DataFeeder(object): if program is None: program = default_main_program() for each_var in feed_list: - if isinstance(each_var, basestring): + if isinstance(each_var, six.string_types): each_var = program.block(0).var(each_var) if not isinstance(each_var, Variable): raise TypeError("Feed list should contain a list of variable") @@ -99,8 +163,18 @@ class DataFeeder(object): self.place = place def feed(self, iterable): + """ + According to feed_list and iterable, converters the input into + a data structure that can feed into Executor and ParallelExecutor. + + Args: + iterable(list|tuple): the input data. + + Returns: + dict: the result of conversion. + """ converter = [] - for lod_level, shape, dtype in six.zip( + for lod_level, shape, dtype in six.moves.zip( self.feed_lod_level, self.feed_shapes, self.feed_dtypes): converter.append( DataToLoDTensorConverter( @@ -113,23 +187,41 @@ class DataFeeder(object): assert len(each_sample) == len(converter), ( "The number of fields in data (%s) does not match " + "len(feed_list) (%s)") % (len(each_sample), len(converter)) - for each_converter, each_slot in six.zip(converter, each_sample): + for each_converter, each_slot in six.moves.zip(converter, + each_sample): each_converter.feed(each_slot) ret_dict = {} - for each_name, each_converter in six.zip(self.feed_names, converter): + for each_name, each_converter in six.moves.zip(self.feed_names, + converter): ret_dict[each_name] = each_converter.done() return ret_dict def feed_parallel(self, iterable, num_places=None): + """ + Takes multiple mini-batches. Each mini-batch will be feed on each + device in advance. + + Args: + iterable(list|tuple): the input data. + num_places(int): the number of devices. Default None. + + Returns: + dict: the result of conversion. + + Notes: + The number of devices and number of mini-batches must be same. + """ if isinstance(self.place, core.CUDAPlace): places = [ core.CUDAPlace(i) - for i in six.xrange(self._get_number_of_places_(num_places)) + for i in six.moves.xrange( + self._get_number_of_places_(num_places)) ] else: places = [ core.CPUPlace() - for _ in six.xrange(self._get_number_of_places_(num_places)) + for _ in six.moves.xrange( + self._get_number_of_places_(num_places)) ] if len(iterable) != len(places): @@ -139,7 +231,7 @@ class DataFeeder(object): "must be same.") place = self.place - for p, batch in six.zip(places, iterable): + for p, batch in six.moves.zip(places, iterable): self.place = p yield self.feed(batch) self.place = place @@ -150,13 +242,33 @@ class DataFeeder(object): elif isinstance(self.place, core.CUDAPlace): return core.get_cuda_device_count() else: - return multiprocessing.cpu_count() + cpu_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + return cpu_num def decorate_reader(self, reader, multi_devices, num_places=None, drop_last=True): + """ + Converter the input data into a data that returned by reader into + multiple mini-batches. Each mini-batch will be feed on each device. + + Args: + reader(fun): the input data. + multi_devices(bool): the number of places. Default None. + num_places(int): the number of places. Default None. + drop_last(bool): the number of places. Default None. + + Returns: + dict: the result of conversion. + + Raises: + ValueError: If drop_last is False and the data batch which cannot + fit for devices. + """ + def __reader_creator__(): if not multi_devices: for item in reader(): diff --git a/python/paddle/fluid/debugger.py b/python/paddle/fluid/debugger.py index 1c56064a1e8bdc5d975837cb5a75a40d557765ad..b7a92cf044900acdd41ede378dd68aa2d9c6b2dc 100644 --- a/python/paddle/fluid/debugger.py +++ b/python/paddle/fluid/debugger.py @@ -14,8 +14,8 @@ import sys import re -from graphviz import GraphPreviewGenerator -import proto.framework_pb2 as framework_pb2 +from .graphviz import GraphPreviewGenerator +from .proto import framework_pb2 from google.protobuf import text_format _vartype2str_ = [ diff --git a/python/paddle/fluid/evaluator.py b/python/paddle/fluid/evaluator.py index 7c6ad6f27dcfd7040f79c72c01413c8cc84a28ba..c0671cce9a1f169f02ba03a839c45b6e4df2c47a 100644 --- a/python/paddle/fluid/evaluator.py +++ b/python/paddle/fluid/evaluator.py @@ -15,11 +15,11 @@ import warnings import numpy as np -import layers -from framework import Program, Variable, program_guard -import unique_name -from layer_helper import LayerHelper -from initializer import Constant +from . import layers +from .framework import Program, Variable, program_guard +from . import unique_name +from .layer_helper import LayerHelper +from .initializer import Constant __all__ = [ 'ChunkEvaluator', @@ -41,7 +41,12 @@ def _clone_var_(block, var): class Evaluator(object): """ - Base Class for all evaluators + Warning: better to use the fluid.metrics.* things, more + flexible support via pure Python and Operator, and decoupled + with executor. Short doc are intended to urge new user + start from Metrics. + + Base Class for all evaluators. Args: name(str): The name of evaluator. such as, "accuracy". Used for generate @@ -69,6 +74,10 @@ class Evaluator(object): def reset(self, executor, reset_program=None): """ reset metric states at the begin of each pass/user specified batch + + Args: + executor(Executor|ParallelExecutor): a executor for executing the reset_program + reset_program(Program): a single Program for reset process """ if reset_program is None: reset_program = Program() @@ -85,15 +94,16 @@ class Evaluator(object): def eval(self, executor, eval_program=None): """ Evaluate the statistics merged by multiple mini-batches. + Args: + executor(Executor|ParallelExecutor): a executor for executing the eval_program + eval_program(Program): a single Program for eval process """ raise NotImplementedError() - def create_state(self, suffix, dtype, shape): + def _create_state(self, suffix, dtype, shape): """ Create state variable. - NOTE: It is not a public API. - Args: suffix(str): the state suffix. dtype(str|core.VarDesc.VarType): the state data type @@ -113,9 +123,35 @@ class Evaluator(object): class ChunkEvaluator(Evaluator): """ + Warning: This would be deprecated in the future. Please use fluid.metrics.ChunkEvaluator + instead. + Accumulate counter numbers output by chunk_eval from mini-batches and compute the precision recall and F1-score using the accumulated counter numbers. + For some basics of chunking, please refer to + 'Chunking with Support Vector Machines '. + + Args: + input (Variable): prediction output of the network. + label (Variable): label of the test data set. + chunk_scheme (str): can be IOB/IOE/IOBES and IO. See the chunk_eval op for details. + num_chunk_types (int): the number of chunk type. + excluded_chunk_types (list): A list including chunk type ids, indicating chunk types that are not counted. + + Returns: + tuple: tuple containing: precision, recall, f1_score + + Examples: + .. code-block:: python + + exe = fluid.executor(place) + evaluator = fluid.Evaluator.ChunkEvaluator(input, label) + for epoch in PASS_NUM: + evaluator.reset(exe) + for data in batches: + loss = exe.run(fetch_list=[cost]) + distance, instance_error = distance_evaluator.eval(exe) """ def __init__( @@ -130,11 +166,11 @@ class ChunkEvaluator(Evaluator): if main_program.current_block().idx != 0: raise ValueError("You can only invoke Evaluator in root block") - self.num_infer_chunks = self.create_state( + self.num_infer_chunks = self._create_state( dtype='int64', shape=[1], suffix='num_infer_chunks') - self.num_label_chunks = self.create_state( + self.num_label_chunks = self._create_state( dtype='int64', shape=[1], suffix='num_label_chunks') - self.num_correct_chunks = self.create_state( + self.num_correct_chunks = self._create_state( dtype='int64', shape=[1], suffix='num_correct_chunks') precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks = layers.chunk_eval( input=input, @@ -178,6 +214,8 @@ class ChunkEvaluator(Evaluator): class EditDistance(Evaluator): """ + Warning: This would be deprecated in the future. Please use fluid.metrics.EditDistance + instead. Accumulate edit distance sum and sequence number from mini-batches and compute the average edit_distance and instance error of all batches. @@ -188,15 +226,16 @@ class EditDistance(Evaluator): ignored_tokens(list of int): Tokens that should be removed before calculating edit distance. - Example: + Examples: + .. code-block:: python - exe = fluid.executor(place) - distance_evaluator = fluid.Evaluator.EditDistance(input, label) - for epoch in PASS_NUM: - distance_evaluator.reset(exe) - for data in batches: - loss = exe.run(fetch_list=[cost]) - distance, instance_error = distance_evaluator.eval(exe) + exe = fluid.executor(place) + distance_evaluator = fluid.Evaluator.EditDistance(input, label) + for epoch in PASS_NUM: + distance_evaluator.reset(exe) + for data in batches: + loss = exe.run(fetch_list=[cost]) + distance, instance_error = distance_evaluator.eval(exe) In the above example: 'distance' is the average of the edit distance in a pass. @@ -210,11 +249,11 @@ class EditDistance(Evaluator): if main_program.current_block().idx != 0: raise ValueError("You can only invoke Evaluator in root block") - self.total_distance = self.create_state( + self.total_distance = self._create_state( dtype='float32', shape=[1], suffix='total_distance') - self.seq_num = self.create_state( + self.seq_num = self._create_state( dtype='int64', shape=[1], suffix='seq_num') - self.instance_error = self.create_state( + self.instance_error = self._create_state( dtype='int64', shape=[1], suffix='instance_error') distances, seq_num = layers.edit_distance( input=input, label=label, ignored_tokens=ignored_tokens) @@ -256,9 +295,10 @@ class EditDistance(Evaluator): class DetectionMAP(Evaluator): """ + Warning: This would be deprecated in the future. Please use fluid.metrics.DetectionMAP + instead. Calculate the detection mean average precision (mAP). - TODO (Dang Qingqing): update the following doc. The general steps are as follows: 1. calculate the true positive and false positive according to the input of detection and labels. @@ -293,17 +333,18 @@ class DetectionMAP(Evaluator): - 11point: the 11-point interpolated average precision. - integral: the natural integral of the precision-recall curve. - Example: + Examples: + .. code-block:: python - exe = fluid.executor(place) - map_evaluator = fluid.Evaluator.DetectionMAP(input, - gt_label, gt_box, gt_difficult) - cur_map, accum_map = map_evaluator.get_map_var() - fetch = [cost, cur_map, accum_map] - for epoch in PASS_NUM: - map_evaluator.reset(exe) - for data in batches: - loss, cur_map_v, accum_map_v = exe.run(fetch_list=fetch) + exe = fluid.executor(place) + map_evaluator = fluid.Evaluator.DetectionMAP(input, + gt_label, gt_box, gt_difficult) + cur_map, accum_map = map_evaluator.get_map_var() + fetch = [cost, cur_map, accum_map] + for epoch in PASS_NUM: + map_evaluator.reset(exe) + for data in batches: + loss, cur_map_v, accum_map_v = exe.run(fetch_list=fetch) In the above example: @@ -340,9 +381,10 @@ class DetectionMAP(Evaluator): evaluate_difficult=evaluate_difficult, ap_version=ap_version) - self.create_state(dtype='int32', shape=None, suffix='accum_pos_count') - self.create_state(dtype='float32', shape=None, suffix='accum_true_pos') - self.create_state(dtype='float32', shape=None, suffix='accum_false_pos') + self._create_state(dtype='int32', shape=None, suffix='accum_pos_count') + self._create_state(dtype='float32', shape=None, suffix='accum_true_pos') + self._create_state( + dtype='float32', shape=None, suffix='accum_false_pos') self.has_state = None var = self.helper.create_variable( diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 93aa5f908ec929a33089a62caa2186ba9e57fffe..e24b9faae24084ccc743a5b5126db9667089e128 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -14,21 +14,27 @@ import numpy as np import contextlib -from framework import Program, default_main_program, Variable +import six +from .framework import Program, default_main_program, Variable from . import core -__all__ = [ - 'Executor', 'global_scope', 'scope_guard', 'switch_scope', 'fetch_var' -] +__all__ = ['Executor', 'global_scope', 'scope_guard', '_switch_scope'] g_scope = core.Scope() def global_scope(): + """ + Get the global/default scope instance. There are a lot of APIs use + :code:`global_scope` as its default value, e.g., :code:`Executor.run` + + Returns: + Scope: The global/default scope instance. + """ return g_scope -def switch_scope(scope): +def _switch_scope(scope): global g_scope ex = g_scope g_scope = scope @@ -37,12 +43,42 @@ def switch_scope(scope): @contextlib.contextmanager def scope_guard(scope): - ex = switch_scope(scope) + """ + Change the global/default scope instance by Python `with` statement. All + variable in runtime will assigned to the new scope. + + Examples: + >>> import paddle.fluid as fluid + >>> new_scope = fluid.Scope() + >>> with fluid.scope_guard(new_scope): + >>> ... + + Args: + scope: The new global/default scope. + """ + ex = _switch_scope(scope) yield - switch_scope(ex) + _switch_scope(ex) def as_numpy(tensor): + """ + Convert a Tensor to a numpy.ndarray, its only support Tensor without LoD information. + For higher dimensional sequence data, please use LoDTensor directly. + Examples: + >>> import paddle.fluid as fluid + >>> outs = executor.run(...) + >>> np_outs = map(lambda x: as_numpy(x), outs) + >>> ... + + Args: + tensor(Variable): a instance of Tensor + + Returns: + numpy.ndarray + """ + if isinstance(tensor, core.LoDTensorArray): + return [as_numpy(t) for t in tensor] if isinstance(tensor, list): return [as_numpy(t) for t in tensor] assert isinstance(tensor, core.LoDTensor) @@ -133,16 +169,20 @@ def has_fetch_operators(block, fetch_targets, fetch_holder_name): return fetch_count > 0 -def fetch_var(name, scope=None, return_numpy=True): +def _fetch_var(name, scope=None, return_numpy=True): """ - Fetch the value of the variable with the given name from the given scope + Fetch the value of the variable with the given name from the + given scope. + Args: name(str): name of the variable. Typically, only persistable variables can be found in the scope used for running the program. scope(core.Scope|None): scope object. It should be the scope where you pass to Executor.run() when running your program. - If None, global_scope() will be used. - return_numpy(bool): whether convert the tensor to numpy.ndarray + If None, global_scope() will be used. Default None. + return_numpy(bool): whether convert the tensor to numpy.ndarray. + Default True. + Returns: LodTensor|numpy.ndarray """ @@ -162,41 +202,82 @@ def fetch_var(name, scope=None, return_numpy=True): return tensor -def get_program_cache_key(feed, fetch_list): - feed_var_names = feed.keys() +def _get_program_cache_key(feed, fetch_list): + feed_var_names = list(feed.keys()) def to_name_str(var): if isinstance(var, Variable): return var.desc.name() elif isinstance(var, str): return var + elif isinstance(var, six.string_types): + return str(var) else: raise TypeError(str(var) + " should be Variable or str") - fetch_var_names = map(to_name_str, fetch_list) + fetch_var_names = list(map(to_name_str, fetch_list)) return str(feed_var_names + fetch_var_names) +def _as_lodtensor(data, place): + """ + Convert numpy.ndarray to Tensor, its only support Tensor without LoD information. + For higher dimensional sequence data, please use LoDTensor directly. + + Examples: + >>> import paddle.fluid as fluid + >>> place = fluid.CPUPlace() + >>> exe = fluid.executor(place) + >>> data = np.array(size=(100, 200, 300)) + >>> np_outs = map(lambda x: fluid.executor._as_lodtensor(x, place), data) + >>> ... + + Args: + data(numpy.ndarray): a instance of array + + Returns: + LoDTensor + """ + if isinstance(data, list): + raise RuntimeError("Some of your feed data hold LoD information. \ + They can not be completely cast from a list of Python \ + ndarray to LoDTensor. Please convert data to LoDTensor \ + directly before feeding the data.\ + ") + # single tensor case + tensor = core.LoDTensor() + tensor.set(data, place) + return tensor + + class Executor(object): + """ + An Executor in Python, only support the single-GPU running. For multi-cards, please refer to + ParallelExecutor. + Python executor takes a program, add feed operators and fetch operators to this program according + to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides + the variables(or names) that user want to get after program run. Note: the executor will run all + operators in the program but not only the operators dependent by the fetch_list. + It store the global variables into the global scope, and create a local scope for the temporary + variables. The local scope contents will be discarded after every minibatch forward/backward finished. + But the global scope variables will be persistent through different runs. + All of ops in program will be running in sequence. + + Args: + place(core.CPUPlace|core.CUDAPlace(n)): indicate the executor run on which device + + Note: For debugging complicated network in parallel-GPUs, you can test it on the executor. + They has the exactly same arguments, and expected the same results. + """ + def __init__(self, place): self.place = place p = core.Place() p.set_place(place) self.executor = core.Executor(p) self.program_caches = dict() - - def as_lodtensor(self, data): - if isinstance(data, list): - raise RuntimeError("Some of your feed data hold LoD information. \ - They can not be completely cast from a list of Python \ - ndarray to LoDTensor. Please convert data to LoDTensor \ - directly before feeding the data.\ - ") - # single tensor case - tensor = core.LoDTensor() - tensor.set(data, self.place) - return tensor + self._closed = False def _get_program_cache(self, program_cache_key): return self.program_caches.get(program_cache_key, None) @@ -230,7 +311,7 @@ class Executor(object): if not has_feed_operators(global_block, feed, feed_var_name): for i, name in enumerate(feed): out = global_block.var(name) - global_block.prepend_op( + global_block._prepend_op( type='feed', inputs={'X': [feed_var]}, outputs={'Out': [out]}, @@ -256,7 +337,7 @@ class Executor(object): feed_target_name = op.desc.output('Out')[0] cur_feed = feed[feed_target_name] if not isinstance(cur_feed, core.LoDTensor): - cur_feed = self.as_lodtensor(cur_feed) + cur_feed = _as_lodtensor(cur_feed, self.place) idx = op.desc.attr('col') core.set_feed_variable(scope, cur_feed, feed_var_name, idx) else: @@ -265,10 +346,28 @@ class Executor(object): def _fetch_data(self, fetch_list, fetch_var_name, scope): outs = [ core.get_fetch_variable(scope, fetch_var_name, i) - for i in xrange(len(fetch_list)) + for i in range(len(fetch_list)) ] return outs + def close(self): + """ + Close this executor. + + You can no long use this executor after calling this method. + For the distributed training, this method would free the resource on PServers related to + the current Trainer. + + Example: + >>> cpu = core.CPUPlace() + >>> exe = Executor(cpu) + >>> ... + >>> exe.close() + """ + if not self._closed: + self.executor.close() + self._closed = True + def run(self, program=None, feed=None, @@ -278,24 +377,52 @@ class Executor(object): scope=None, return_numpy=True, use_program_cache=False): - """ Run program by this Executor. Feed data by feed map, fetch result by fetch_list. - + """ + Run program by this Executor. Feed data by feed map, fetch result by fetch_list. Python executor takes a program, add feed operators and fetch operators to this program according to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides - the variables(or names) that user want to get after program run. Note: the executor will run all + the variables(or names) that user want to get after program run. + + Note: the executor will run all operators in the program but not only the operators dependent by the fetch_list - :param program: the program that need to run, if not provied, then default_main_program will be used. - :param feed: feed variable map, e.g. {"image": ImageData, "label": LableData} - :param fetch_list: a list of variable or variable names that user want to get, run will return them according - to this list. - :param feed_var_name: the name for the input variable of feed Operator. - :param fetch_var_name: the name for the output variable of feed Operator. - :param scope: the scope used to run this program, you can switch it to different scope. default is global_scope - :param return_numpy: if convert the fetched tensor to numpy - :param use_program_cache: set use_program_cache to true if program not changed compare to the last step. - :return: result according to fetch_list. + Args: + program(Program): the program that need to run, if not provied, then default_main_program will be used. + feed(dict): feed variable map, e.g. {"image": ImageData, "label": LableData} + fetch_list(list): a list of variable or variable names that user want to get, run will return them according to this list. + feed_var_name(str): the name for the input variable of feed Operator. + fetch_var_name(str): the name for the output variable of fetch Operator. + scope(Scope): the scope used to run this program, you can switch it to different scope. default is global_scope + return_numpy(bool): if convert the fetched tensor to numpy + use_program_cache(bool): set use_program_cache to true if program not changed compare to the last step. + + Returns: + + list(numpy.array): fetch result according to fetch_list. + + + Examples: + + >>> data = layers.data(name='X', shape=[1], dtype='float32') + >>> hidden = layers.fc(input=data, size=10) + >>> layers.assign(hidden, out) + >>> loss = layers.mean(out) + >>> adam = fluid.optimizer.Adam() + >>> adam.minimize(loss) + + >>> cpu = core.CPUPlace() + >>> exe = Executor(cpu) + >>> exe.run(default_startup_program()) + + >>> x = numpy.random.random(size=(10, 1)).astype('float32') + >>> outs = exe.run( + >>> feed={'X': x}, + >>> fetch_list=[loss.name]) """ + + if self._closed: + raise RuntimeError("Attempted to use a closed Executor") + if feed is None: feed = {} if not isinstance(feed, dict): @@ -315,7 +442,7 @@ class Executor(object): if scope is None: scope = global_scope() - cache_key = get_program_cache_key(feed, fetch_list) + cache_key = _get_program_cache_key(feed, fetch_list) if use_program_cache: cached_program = self._get_program_cache(cache_key) if cached_program is None: diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 33b5caa0eab0ec192eb4a3b63cf82a672c58d2cb..45b3abb88c9431f52705bb62df2c32779dd0cf9d 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -15,23 +15,30 @@ import collections import contextlib import re +import six import numpy as np -import proto.framework_pb2 as framework_pb2 -from . import core -import unique_name +from .proto import framework_pb2 +try: + from . import core +except ImportError as e: + raise ImportError( + """NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\" + if you encounters \"libmkldnn.so not found\" errors. If you have python + installed in other directory, replace \"/usr/local/lib\" with your own + directory. The original error is: \n""" + e.message) +except Exception as e: + raise e +from . import unique_name __all__ = [ - 'Block', - 'Variable', 'Program', 'Operator', + 'Parameter', 'default_startup_program', 'default_main_program', 'program_guard', - 'switch_startup_program', - 'switch_main_program', 'get_var', ] @@ -43,7 +50,8 @@ ZERO_VAR_SUFFIX = core.kZeroVarSuffix() def grad_var_name(var_name): """ - return gradient name for a certain var name + Returns: + str: gradient name for a certain var name """ return var_name + GRAD_VAR_SUFFIX @@ -51,10 +59,12 @@ def grad_var_name(var_name): def convert_np_dtype_to_dtype_(np_dtype): """ Convert the data type in numpy to the data type in Paddle + Args: - np_dtype(np.dtype): the data type in numpy + np_dtype(np.dtype): the data type in numpy. - Returns(core.VarDesc.VarType): the data type in Paddle + Returns: + core.VarDesc.VarType: the data type in Paddle. """ dtype = np.dtype(np_dtype) @@ -72,10 +82,12 @@ def convert_np_dtype_to_dtype_(np_dtype): return core.VarDesc.VarType.INT64 elif dtype == np.bool: return core.VarDesc.VarType.BOOL + elif dtype == np.uint16: + return core.VarDesc.VarType.INT16 elif dtype == np.uint8: return core.VarDesc.VarType.UINT8 else: - raise ValueError("Not supported numpy dtype " + str(dtype)) + raise ValueError("Not supported numpy dtype " + six.binary_type(dtype)) def dtype_is_floating(dtype): @@ -118,37 +130,53 @@ def _debug_string_(proto, throw_on_error=True): class Variable(object): """ - Python variable. Every input and output of an operator is a variable. Every - variable belongs to a block. The variable has a name and two variables in - different blocks could have the same name. - - There are many kinds of variables. Please reference the framework.proto for - details. + In Fluid, every input and output of an operator is a variable. In most + cases, variables are used for holding different kinds of data or training + labels. A variable belongs to a block. All variable has its own name and + two variables in different blocks could have the same name. - Notes: The constructor of Variable should not be invoked directly. Please - use `Block.create_var` to create a variable. + There are many kinds of variables. Each kind of them has its own attributes + and usages. Please reference the framework.proto for details. - >>> cur_program = Program() - >>> cur_block = cur_program.current_block() - >>> new_variable = cur_block.create_var( - >>> name="X", shape=[-1, 23, 48], dtype='float32') + Most of a Variable's member variables can be setted to be None. It mean + it is not available or will be specified later. Args: - block(Block): The associated block. It will be passed by - `Block.create_var` automatically. + block(Block): The block that the variable belongs to. type(core.VarDesc.VarType): Variable type. Please reference the framework.proto for details. - shape(tuple|list|None): The shape of variable. -1 means the batch size. + name(str|None): The name of the variable. If setted None, it will be + generated automatically. Default: None + shape(tuple|list|None): The shape of the variable. -1 means the batch size. Some kinds of variable do not contain shape, just set it to None. - dtype(np.dtype|core.VarDesc.VarType|str): The data type of variable. - lod_level(int): The level of lod tensor. 0 means it is not a time + Default: None + dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable. + Default: None + lod_level (int|None): The level of lod tensor. 0 means it is not a time series data. - capacity(int): The capacity of Channel variable. Ignored - for other types. - persistable(bool): True if the variable should be saved as check point. - Defaults to False. - stop_gradient(bool): True if the variable will stop to calculate - gradients when backward. Defaults to False. + Default: None + capacity (int|None): The capacity of Channel variable. Ignored for other + types. Default: None + persistable (bool|None): True if the variable is persistable. A persistable + variable will not be deleted after an iteration ending. Defaults: None. + error_clip (BaseErrorClipAttr|None): The error clip attributes of the + corresponding gradient variable. Default: None + stop_gradient (bool): True if the variable will stop to calculate its + gradients when backward. Default: False. + is_data (bool): True if the variable is an input data. Default: False + + Notes: + The constructor of Variable should not be invoked directly. Please + use `Block.create_var` to create a variable. + + Examples: + .. code-block:: python + + cur_program = Program() + cur_block = cur_program.current_block() + new_variable = cur_block.create_var(name="X", + shape=[-1, 23, 48], + dtype='float32') """ def __init__(self, @@ -170,6 +198,7 @@ class Variable(object): if name is None: name = unique_name.generate('_generated_var') is_new_var = False + name = name if isinstance(name, six.binary_type) else name.encode() self.desc = self.block.desc.find_var(name) if self.desc is None: @@ -251,29 +280,39 @@ class Variable(object): Get debug string. Args: - throw_on_error(bool): True if raise an exception when self is not - intialized. + throw_on_error(bool): True if raise an exception when self is + not initialized. with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True - - Returns(str): The debug string. + (e.g. trainable, optimize_attr, ...) will be printed when + with_details is True. Default False; + Returns: + str: The debug string. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, bool) protostr = self.desc.serialize_to_string() - proto = framework_pb2.VarDesc.FromString(str(protostr)) + proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr)) res_str = _debug_string_(proto, throw_on_error) if with_details: additional_attr = ("error_clip", "stop_gradient") for attr_name in additional_attr: - res_str += "%s: %s\n" % (attr_name, - str(getattr(self, attr_name))) + res_str += "%s: %s\n" % ( + attr_name, six.binary_type(getattr(self, attr_name))) return res_str __repr__ = __str__ - def set_desc(self, input): + def _set_desc(self, input): + """ + Set the variable description. + + Args: + input(core.VarDesc): The new VarDesc. + + Returns: + None + """ self.desc = input @property @@ -309,7 +348,16 @@ class Variable(object): def type(self): return self.desc.type() - def set_error_clip(self, error_clip): + def _set_error_clip(self, error_clip): + """ + Set the error_clip. + + Args: + error_clip(BaseErrorClipAttr) : The new error_clip. + + Returns: + None + """ self.error_clip = error_clip @@ -317,13 +365,13 @@ def get_all_op_protos(): """ Get all registered op proto from PaddlePaddle C++ end. - Returns(list): list of OpProto - + Returns: + list: list of OpProto. """ protostrs = core.get_all_op_protos() ret_values = [] for pbstr in protostrs: - op_proto = framework_pb2.OpProto.FromString(str(pbstr)) + op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr)) ret_values.append(op_proto) return ret_values @@ -361,13 +409,63 @@ class OpProtoHolder(object): raise ValueError("Operator \"%s\" has not been registered." % type) return self.op_proto_map[type] + @staticmethod + def generated_op_attr_names(): + return { + core.op_proto_and_checker_maker.kOpRoleAttrName(), + core.op_proto_and_checker_maker.kOpRoleVarAttrName() + } + class Operator(object): """ - Python Operator class. The operator represents the build in instructions in a - Block. Users can use the build in instructions to describe their neural - network. + In Fluid, all the operation are represented by Operator, and Operator + is regarded as a build in an instruction of a Block. Users can use the + build in instructions to describe their neural network. + + Args: + block(Block): The block has the current operator. + desc(core.OpDesc): The protobuf description of Operator. + type(str): The type of operator. Default None. + inputs(dict): The input of this Operator. it is a dictionary, for every + element, key is the input parameter name, and value is a list of + variables. Default None. + outputs(dict): The output of this Operator. it is a dictionary, for + every element, key is the input parameter name, and value is a list + of variables. Default None. + attrs(dict): The attributes of this Operator. it is a dictionary, for + every element, key is attribute name, and value is the attribute value. + The attribute type should be as same as the type registered in C++ side. + Default None. + + Returns: + Operator: The initialized Operator. + + Raises: + ValueError: If the passed input, output and attrs doesn't match the + initializing Operator's that registered in C++ side. + + Notes: + The constructor of operator should not be invoked directly. Use + Block.append_op or Block._prepend_op instead. + + Examples: + .. code-block:: python + + cur_program = Program() + cur_block = cur_program.current_block() + # var1 += var2 + var3 + cur_block.append_op(type="sum", + inputs={"X": [var1, var2, var3]}, + outputs={"Out": [var1]}) """ + OP_WITHOUT_KERNEL_SET = { + 'feed', 'fetch', 'save', 'load', 'recurrent', 'go', + 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv', + 'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine', + 'ncclInit', 'channel_create', 'channel_close', 'channel_send', + 'channel_recv', 'select', 'checkpoint_notify', 'gen_nccl_id' + } def __init__(self, block, @@ -376,50 +474,27 @@ class Operator(object): inputs=None, outputs=None, attrs=None): - """ - Constructor. - - Notes: The constructor of operator should not be invoked directly. Use - Block.append_op or Block.prepend_op instead. - - >>> cur_program = Program() - >>> cur_block = cur_program.current_block() - >>> # var1 += var2 + var3 - >>> cur_block.append_op(type="sum", - >>> inputs={"X": [var1, var2, var3]}, - >>> outputs={"Out": [var1]}) - - Args: - block(Block): The block has the current operator. - desc(core.OpDesc): The protobuf description. - type(str): The type of operator. - inputs(dict): The input dictionary. Key is the input parameter name. - Value is a list of variables. - outputs(dict): The output dictionary which has the same format with - inputs. - attrs(dict): The attributes dictionary. Key is attribute name. Value - is the attribute value. The attribute type should be as same as - the type registered in C++ - """ self.block = block self.desc = desc - self.attrs = attrs - if self.attrs is None: - self.attrs = dict() + # note: not add self.attrs here: + # https://github.com/PaddlePaddle/Paddle/pull/12583#pullrequestreview-145093173 + op_attrs = attrs + if op_attrs is None: + op_attrs = dict() del attrs op_maker = core.op_proto_and_checker_maker - if op_maker.kOpRoleAttrName() not in self.attrs: - self.attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role + if op_maker.kOpRoleAttrName() not in op_attrs: + op_attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role role_var_name = op_maker.kOpRoleVarAttrName() if len(self.block.program. - op_role_var) != 0 and role_var_name not in self.attrs: - self.attrs[role_var_name] = self.block.program.op_role_var + op_role_var) != 0 and role_var_name not in op_attrs: + op_attrs[role_var_name] = self.block.program.op_role_var - if role_var_name in self.attrs and len(self.attrs[role_var_name]) == 0: - del self.attrs[role_var_name] + if role_var_name in op_attrs and len(op_attrs[role_var_name]) == 0: + del op_attrs[role_var_name] if len(self.desc.type()) != 0: return @@ -451,10 +526,19 @@ class Operator(object): % (in_proto.name, len(in_args))) in_arg_names = [] for arg in in_args: - if isinstance(arg, basestring): + if isinstance(arg, six.string_types): in_arg_names.append(arg) + elif isinstance(arg, six.binary_type): + in_arg_names.append(arg.decode()) else: - in_arg_names.append(arg.name) + if isinstance(arg.name, six.string_types): + in_arg_names.append(arg.name) + elif isinstance(arg.name, six.binary_type): + in_arg_names.append(arg.name.decode()) + else: + raise TypeError( + "arguments require unicode, str or bytes, but get %s instead." + % (type(arg.name))) self.desc.set_input(in_proto.name, in_arg_names) else: self.desc.set_input(in_proto.name, []) @@ -469,8 +553,9 @@ class Operator(object): if not given == need: raise ValueError(("Incorrect setting for output(s) of " "operator \"%s\". Need: [%s] Given: [%s]") % - (type, ", ".join(str(e) for e in need), - ", ".join(str(e) for e in given))) + (type, + ", ".join(six.binary_type(e) for e in need), + ", ".join(six.binary_type(e) for e in given))) for out_proto in proto.outputs: out_args = outputs[out_proto.name] @@ -482,51 +567,49 @@ class Operator(object): (out_proto.name, len(out_args))) out_arg_names = [] for arg in out_args: - out_arg_names.append(arg.name) + if isinstance(arg.name, six.string_types): + out_arg_names.append(arg.name) + elif isinstance(arg.name, six.binary_type): + out_arg_names.append(arg.name.decode()) + else: + raise TypeError( + "arguments require unicode, str or bytes, but get %s instead." + % (type(arg.name))) arg.op = self self.desc.set_output(out_proto.name, out_arg_names) - if self.attrs is not None: - if not isinstance(self.attrs, dict): + if op_attrs is not None: + if not isinstance(op_attrs, dict): raise TypeError("'attrs' should be a dict.") for attr in proto.attrs: attr_name = attr.name - if (attr_name not in self.attrs) or ( - self.attrs[attr_name] is None): + if (attr_name not in op_attrs) or (op_attrs[attr_name] is None): continue - if isinstance(self.attrs[attr_name], Block): - self.desc.set_block_attr(attr_name, - self.attrs[attr_name].desc) - elif isinstance(self.attrs[attr_name], core.BlockDesc) or \ - isinstance(self.attrs[attr_name], core.ProgramDesc): - self.desc.set_serialized_attr( - attr_name, self.attrs[attr_name].serialize_to_string()) - else: - self.desc.set_attr(attr_name, self.attrs[attr_name]) + attr_val = op_attrs[attr_name] + self._update_desc_attr(attr_name, attr_val) + self.desc.check_attrs() - no_kernel_op_set = { - 'feed', 'fetch', 'save', 'load', 'recurrent', 'go', - 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', - 'recv', 'listen_and_serv', 'parallel_do', 'save_combine', - 'load_combine', 'ncclInit', 'channel_create', 'channel_close', - 'channel_send', 'channel_recv', 'select', 'gen_nccl_id' - } - if type not in no_kernel_op_set: + if self.has_kernel(type): self.desc.infer_var_type(self.block.desc) self.desc.infer_shape(self.block.desc) + def has_kernel(self, op_type): + return op_type not in self.OP_WITHOUT_KERNEL_SET + def to_string(self, throw_on_error): """ - To debug string. + Get debug string. + Args: - throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True + throw_on_error(bool): Whether to raise exception if self is not + initialized. - Returns(str): The debug string. + Returns: + str: The debug string. """ protostr = self.desc.serialize_to_string() - proto = framework_pb2.OpDesc.FromString(str(protostr)) + proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr)) return _debug_string_(proto, throw_on_error) def __str__(self): @@ -540,29 +623,45 @@ class Operator(object): def input(self, name): """ - Get input arguments by the input parameter name - Args: - name(str): The input parameter name + Get the input arguments according to the input parameter name. - Returns(list): return the list of argument names associated with the - specific parameter name. + Args: + name(str): The input parameter name. + Returns: + list: return the list of argument names that associated with \ + the specific parameter name. """ return self.desc.input(name) def rename_input(self, old_name, new_name): + """ + Rename the `old_name` to `new_name`. + + Args: + old_name(str): The old name of the Operator's input. + new_name(str): The new name of the Operator's input. + + Returns: + None + """ self.desc.rename_input(old_name, new_name) def rename_output(self, old_name, new_name): + """ + Rename the `old_name` to `new_name`. + + Args: + old_name(str): The old name of the Operator's output. + new_name(str): The new name of the Operator's output. + + Returns: + None + """ self.desc.rename_output(old_name, new_name) @property def input_names(self): - """ - Get all input parameter names - Returns(list): return a list of input parameter names - - """ return self.desc.input_names() @property @@ -575,33 +674,23 @@ class Operator(object): def output(self, name): """ - Get output arguments by the output parameter name - Args: - name(str): The output parameter name + Get output arguments by the output parameter name. - Returns(list): return the list of argument names associated with the - specific parameter name. + Args: + name(str): The output parameter name. + Returns: + list: return the list of argument names associated with \ + the specific parameter name. """ return self.desc.output(name) @property def output_names(self): - """ - Get all output parameter names - Returns(list): return a list of output parameter names - - """ return self.desc.output_names() @property def idx(self): - """ - Return the array index of current operator. - Returns(int): The array index in block.ops array - Raises: - ValueError: when the operator is not found. - """ for i, op in enumerate(self.block.ops): if op == self: return i @@ -610,78 +699,192 @@ class Operator(object): def has_attr(self, name): """ - operator has the attribute with name or not. + Whether this Operator has the attribute with name or not. + Args: - name(str): the attribute name + name(str): the attribute name. - Returns(bool): True if has this attribute. + Returns: + bool: True if has this attribute. """ return self.desc.has_attr(name) def attr_type(self, name): """ - Get the type of attribute by attribute name - Args: - name(str): the attribute name + Get the type of attribute by attribute's name. - Returns(core.AttrType): the attribute type + Args: + name(str): the attribute name. + Returns: + core.AttrType: the attribute type. """ return self.desc.attr_type(name) def set_attr(self, name, val): - self.attrs[name] = val - self.desc.set_attr(name, val) + """ + Set the value of attribute by attribute's name. - @property - def attr_names(self): + Args: + name(str): the attribute name. + val(bool|int|str|float|list): the value of the attribute. + + Raises: + ValueError: If the type of value doesn't match with desc.attr_type(name). """ - Get all attribute names - Returns(list): The list of attribute name + self._update_desc_attr(name, val) + + def _update_desc_attr(self, name, val): + """ + Update the value of desc's attribute by attribute's name. + + Args: + name(str): the attribute name. + val(bool|int|str|float|list): the value of the attribute. + Raises: + ValueError: If the type of value doesn't match with desc.attr_type(name). """ + if isinstance(val, Block): + self.desc.set_block_attr(name, val.desc) + elif isinstance(val, list) and val and all( + isinstance(v, Block) for v in val): + self.desc.set_blocks_attr(name, [v.desc for v in val]) + elif isinstance(val, core.BlockDesc) or \ + isinstance(val, core.ProgramDesc): + self.desc.set_serialized_attr(name, val.serialize_to_string()) + else: + self.desc.set_attr(name, val) + + @property + def attr_names(self): return self.desc.attr_names() def attr(self, name): """ - Get attribute by name + Get the attribute by name. + Args: - name(str): the attribute name + name(str): the attribute name. - Returns(bool|int|str|float|list): The attribute value. The return value + Returns: + bool|int|str|float|list: The attribute value. The return value can be any valid attribute type. - """ return self.desc.attr(name) + def block_attr_id(self, name): + """ + Get the block attribute's id by name. + + Args: + name(str): the attribute name. + + Returns: + int: the block index. + """ + return self.desc.block_attr_id(name) + def block_attr(self, name): """ - Get the block attribute by name + Get the block attribute by name. + + Args: + name(str): the attribute name. + + Returns: + block: the block attribute. + """ + + id = self.block_attr_id(name) + assert (id >= 0 and id < len(self.block.program.blocks)) + return self.block.program.blocks[id] + + def blocks_attr(self, name): + """ + Get the blocks attribute by name. + Args: - name(str): the attribute name + name(str): the attribute name. + + Returns: + list: list of the blocks attribute. + """ + attrs = [] + for i in self.blocks_attr_ids(name): + assert (i >= 0 and i < len(self.block.program.blocks)) + attrs.append(self.block.program.blocks[i]) - Returns(int): the block index + return attrs + def blocks_attr_ids(self, name): + """ + Get the blocks attribute's ids by name. + + Args: + name(str): the attribute name. + + Returns: + list: list of the blocks ids. """ - return self.desc.block_attr(name) + + return self.desc.blocks_attr_ids(name) def all_attrs(self): """ - Get the attribute dict - Returns(dict): The Operator's attribute dict + Get the attribute dict. + + Returns: + dict: The Operator's attribute dict, name->attr. """ attr_names = self.attr_names attr_map = {} for n in attr_names: - if n == 'sub_block': + attr_type = self.desc.attr_type(n) + if attr_type == core.AttrType.BLOCK: attr_map[n] = self.block_attr(n) - else: - attr_map[n] = self.attr(n) + continue + + if attr_type == core.AttrType.BLOCKS: + attr_map[n] = self.blocks_attr(n) + continue + + attr_map[n] = self.attr(n) + return attr_map class Block(object): + """ + In Fluid, a Program is consistence of multi-Block, and Block stores + VarDesc and OpDesc. In a specific Block, a VarDesc have a unique name. + One block could have some child blocks, and child block's name scopes + should inherit the parent's so that OpDesc in child block can reference + a VarDesc that is stored in the parent block. + Please reference the framework.proto for details. + + Args: + program(Program): The Program that the Block belongs to. + idx(int): The block's id in the Program. + + Notes: + The constructor of Block should not be invoked directly. Please + use `Program.create_block()` to create a block. + + Examples: + .. code-block:: python + + cur_program = Program() + cur_block = cur_program.current_block() + var = cur_block.create_var(name="X", + shape=[-1, 23, 48], + dtype='float32') + cur_block.append_op(type="abs", + inputs={"X": [var]}, + outputs={"Out": [var]}) + """ + def __init__(self, program, idx): self.desc = program.desc.block(idx) self.vars = collections.OrderedDict() # var_name --> var @@ -694,15 +897,17 @@ class Block(object): def to_string(self, throw_on_error, with_details=False): """ - To debug string. + Get debug string. + Args: throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True + when throw_on_error is True. with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True - - Returns(str): The debug string. + (e.g. trainable, optimize_attr, ...) will be printed when + with_details is True. Default False. + Returns: + str: The debug string. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, bool) @@ -710,7 +915,7 @@ class Block(object): re_add_indent = re.compile(r"\n(.)") res_str = "blocks {\n idx: %d\n parent_idx: %d" % ( self.idx, self.parent_idx) - for var in self.vars.itervalues(): + for var in list(self.vars.values()): res_str += "\n vars {\n %s }" % re_add_indent.sub( r"\n \1", var.to_string(throw_on_error, with_details)) for op in self.ops: @@ -719,7 +924,8 @@ class Block(object): res_str += "\n}" else: protostr = self.desc.serialize_to_string() - proto = framework_pb2.BlockDesc.FromString(str(protostr)) + proto = framework_pb2.BlockDesc.FromString( + six.binary_type(protostr)) res_str = _debug_string_(proto, throw_on_error) return res_str @@ -733,22 +939,60 @@ class Block(object): def forward_block_idx(self): return self.desc.get_forward_block_idx() - def set_forward_block_idx(self, idx): - self.desc.set_forward_block_idx(idx) + def _set_forward_block_idx(self, idx): + """ + Set the forward block Idx. + + Args: + idx(int): the block index. + + Returns: + None + """ + self.desc._set_forward_block_idx(idx) @property def idx(self): return self.desc.id def var(self, name): - if not isinstance(name, basestring): - raise TypeError() + """ + Get a Variable by name from this block. + + Args: + name(str): the Variable's name. + + Raises: + ValueError: The If input's type is not str, or this block + doesn't have a Variable with the giving name. + + Returns: + Variable: the Variable with the giving name. + """ + if not isinstance(name, six.string_types): + if not isinstance(name, six.binary_type): + raise TypeError( + "var require string as parameter, but get %s instead." % + (type(name))) v = self.vars.get(name, None) if v is None: raise ValueError("var %s not in this block" % name) return v - def var_recursive(self, name): + def _var_recursive(self, name): + """ + Get a Variable by name from this block recursively. + + Args: + name(str): the Variable's name. + + Raises: + ValueError: this block and this parent block doesn't + have a Variable with the giving name. + + Returns: + Variable: the Variable with the giving name. + """ frontier = list() visited = set() @@ -780,7 +1024,7 @@ class Block(object): return list(self.iter_parameters()) def iter_parameters(self): - return (item[1] for item in self.vars.iteritems() + return (item[1] for item in list(self.vars.items()) if isinstance(item[1], Parameter)) def create_var(self, *args, **kwargs): @@ -792,9 +1036,21 @@ class Block(object): def has_var(self, name): return name in self.vars - def rename_var(self, name, new_name): + def _rename_var(self, name, new_name): """ Rename variable in vars and ops' inputs and outputs + + Args: + name(str): the name that need to be renamed. + new_name(str): the name that need to rename to. + + Raises: + ValueError: If this block doesn't have this the giving name, + or the type of the var with the giving name is not Parameter + or Variable. + + Returns: + Variable: the Variable with the giving name. """ if not self.has_var(name): raise ValueError("var %s is not in current block" % name) @@ -814,8 +1070,8 @@ class Block(object): else: raise ValueError("unsupported var type: %s", type(v)) orig_var_type = v.type - self.desc.rename_var(name, new_name) - # NOTE: v is destroyed by C++ after calling rename_var. + self.desc._rename_var(name, new_name) + # NOTE: v is destroyed by C++ after calling _rename_var. d = self.desc.find_var(new_name) if var_type == "Parameter": var = Parameter( @@ -838,57 +1094,109 @@ class Block(object): error_clip=error_clip, stop_gradient=stop_gradient) - # rename the python side, sync_with_cpp will only add + # rename the python side, _sync_with_cpp will only add # new vars/ops to python side. self.vars[new_name] = var del self.vars[name] - self.sync_with_cpp() + self._sync_with_cpp() return var - def remove_var(self, name): - self.sync_with_cpp() - self.desc.remove_var(name) + def _remove_var(self, name): + self._sync_with_cpp() + self.desc._remove_var(name) del self.vars[name] def create_parameter(self, *args, **kwargs): global_block = self.program.global_block() param = Parameter(global_block, *args, **kwargs) if 'initializer' in kwargs: - kwargs['initializer'](param, self) + + def _is_inited_by(block, var): + init_ops = [] + for op in block.ops: + if var.name in op.output_arg_names: + init_ops.append(op) + return init_ops + + initializer = kwargs['initializer'] + init_ops = _is_inited_by(global_block, param) + init_ops_len = len(init_ops) + if init_ops_len > 1: + raise RuntimeError("param " + param.name + + " is inited by multiple init ops " + str( + init_ops)) + elif init_ops_len == 1: + #TODO already inited, do nothing, should log a warning + pass + else: + initializer(param, self) return param def append_op(self, *args, **kwargs): + """ + Appends a new Operator according to the giving arguments. + + Returns: + Operator: the append Operator. + """ op_desc = self.desc.append_op() op = Operator(block=self, desc=op_desc, *args, **kwargs) self.ops.append(op) return op - def insert_op(self, index, *args, **kwargs): - self.sync_with_cpp() - op_desc = self.desc.insert_op(index) + def _insert_op(self, index, *args, **kwargs): + """ + Insert a Operator according to the giving arguments. + + Args: + index(int): the place that the operator to insert. + + Returns: + Operator: the insert Operator. + """ + self._sync_with_cpp() + op_desc = self.desc._insert_op(index) op = Operator(block=self, desc=op_desc, *args, **kwargs) self.ops.insert(index, op) return op - def remove_op(self, index): - self.sync_with_cpp() - self.desc.remove_op(index, index + 1) + def _remove_op(self, index): + """ + Remove the specific position operator. + + Args: + index(int): the position that the operator to insert. + + Returns: + None + """ + self._sync_with_cpp() + self.desc._remove_op(index, index + 1) del self.ops[index] - def slice_ops(self, start, end): + def _slice_ops(self, start, end): + """ + Return the Operator between start and end. + + Args: + start(int): the start position. + end(int): the end position. + + Returns: + list: the Operators between start and end. + """ return self.ops[start:end] - def prepend_op(self, *args, **kwargs): - op_desc = self.desc.prepend_op() + def _prepend_op(self, *args, **kwargs): + op_desc = self.desc._prepend_op() op = Operator(self, op_desc, *args, **kwargs) self.ops.insert(0, op) return op - def sync_with_cpp(self): + def _sync_with_cpp(self): """ - Sync from the desc on the c++ end. - - This method is used to synchronize the c++ desc instance generated by backward. + Sync from the desc on the c++ end. This method is used to synchronize + the c++ desc instance generated by backward. """ # sync variables from cpp for var in self.desc.all_vars(): @@ -896,7 +1204,7 @@ class Block(object): self.create_var(name=var.name(), desc=var, type=var.type()) # sync variables removed from c++ end - for var in self.vars.keys(): + for var in list(self.vars.keys()): if not self.desc.find_var(var): self.vars.pop(var) @@ -951,22 +1259,28 @@ class Block(object): for index in range(len(self.ops)): assert self.ops[index].desc == ops_in_cpp[index] - def copy_param_info_from(self, other): + def _copy_param_info_from(self, other): """ - Copy the information of parameters from the other block + Copy the information of parameters from the other block. + Args: - other(Block): the other block + other(Block): the other block. + + Raises: + ValueError: If type of input is not Block, or the `other` and this + block is not in the same topology. Returns: None """ if not isinstance(other, Block): - raise TypeError("copy_param_info_from should be invoked with Block") + raise TypeError( + "_copy_param_info_from should be invoked with Block") for p in other.iter_parameters(): assert isinstance(p, Parameter) v = self.vars.get(p.name, None) if v is None: - raise ValueError("copy_param_info_from should be invoked with " + raise ValueError("_copy_param_info_from should be invoked with " "same topology") assert isinstance(v, Variable) new_p = Parameter( @@ -984,14 +1298,15 @@ class Block(object): name=v.name) self.vars[new_p.name] = new_p - def clone_variable(self, var): + def _clone_variable(self, var): """ Clone a variable into current block. + Args: var: the variable to be cloned. Returns: - The new variable cloned from 'var' in current block. + Variable: the new variable cloned from 'var' in current block. """ assert isinstance(var, Variable) ret_var = None @@ -999,6 +1314,9 @@ class Block(object): if var.type == core.VarDesc.VarType.STEP_SCOPES: ret_var = self.create_var( name=var.name, persistable=var.persistable, type=var.type) + elif var.type == core.VarDesc.VarType.RAW: + ret_var = self.create_var( + name=var.name, persistable=var.persistable, type=var.type) elif var.type == core.VarDesc.VarType.SELECTED_ROWS: ret_var = self.create_var( name=var.name, @@ -1020,6 +1338,32 @@ class Block(object): class Program(object): + """ + Python Program. Beneath it is a ProgramDesc, which is used for + create c++ Program. A program is a self-contained programing + language like container. It has at least one Block, when the + control flow op like conditional_block, while_op is included, + it will contains nested block. + Please reference the framework.proto for details. + + Notes: we have default_startup_program and default_main_program + by default, a pair of them will shared the parameters. + The default_startup_program only run once to initialize parameters, + default_main_program run in every mini batch and adjust the weights. + + Returns: + A empty program. + + Examples: + >>> main_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(main_program=main_program, startup_program=startup_program): + >>> fluid.layers.data(name="x", shape=[-1, 784], dtype='float32') + >>> fluid.layers.data(name="y", shape=[-1, 1], dtype='int32') + >>> fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu") + + """ + def __init__(self): self.desc = core.ProgramDesc() self.blocks = [Block(self, 0)] @@ -1030,6 +1374,19 @@ class Program(object): @property def op_role(self): + """ + The operator role. In a enum {Forward, Backward, Optimize}. + + Notes: this is a low level API. It is used only for ParallelExecutor to + duplicate or schedule operator to devices. + + For example, the forward operator should be executed on every device. + The backward operator should be executed on every device and the + parameter gradient of backward (use :code:`op_role_var` to get this + variable) operator should be merged to one device. The optimization + operators should be executed on only one device and broadcast the + optimization result, i.e., the new parameter, to every other device. + """ return self._current_role @op_role.setter @@ -1038,6 +1395,13 @@ class Program(object): @property def op_role_var(self): + """ + The auxiliary variables for :code:`op_role` property. + + See Also: :code:`Program.op_role`'s documentation for details. + + Notes: This is a very low-level API. Users should not use it directly. + """ return self._op_role_var @op_role_var.setter @@ -1045,27 +1409,62 @@ class Program(object): self._op_role_var = [var_name] @contextlib.contextmanager - def optimized_guard(self, var): + def optimized_guard(self, param_and_grads): + """ + A with guard to set :code:`Optimization` :code:`OpRole` and + :code:`OpRoleVar` automatically. + + Notes: This is a very low level API. Users should not use it directly. + + Args: + param_and_grads(list): The variables (names) to be optimized. + + Examples: + + >>> p, g = backward(...) + >>> with program.optimized_guard([p,g]): + >>> p = p - 0.001 * g + """ OpRole = core.op_proto_and_checker_maker.OpRole self._current_role = OpRole.Optimize - self._op_role_var = [var.name if isinstance(var, Variable) else var] + self._op_role_var = [ + var.name if isinstance(var, Variable) else var + for var in param_and_grads + ] yield self._op_role_var = [] self._current_role = OpRole.Forward def __str__(self): + """ + Get the protobuf debug string of this Program. + + Returns: + (str): The protobuf debug string. + + Raises: + ValueError: If any of required fields is not set. + """ return self.to_string(True) def to_string(self, throw_on_error, with_details=False): """ To debug string. + Args: - throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True - with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True + throw_on_error(bool): raise Value error when any of required fields + is not set. - Returns(str): The debug string. + with_details(bool): True if more details about variables and + parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need + to print. + + Returns + (str): The debug string. + + Raises: + ValueError: If any of required fields is not set and throw_on_error is + True. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, @@ -1076,41 +1475,134 @@ class Program(object): res_str += block.to_string(throw_on_error, with_details) else: protostr = self.desc.serialize_to_string() - proto = framework_pb2.ProgramDesc.FromString(str(protostr)) + proto = framework_pb2.ProgramDesc.FromString( + six.binary_type(protostr)) res_str = _debug_string_(proto, throw_on_error) return res_str def get_desc(self): + """ + Get the C++ side of `ProgramDesc` object pointer. The C++ object is + exposed by :code:`pybind`. + + Notes: This is a very low level API. Users should not use this API + directly. + """ return self.desc def clone(self, for_test=False): - """Clone the Program object + """ + Create a new, duplicated program. + + + Some operators, e.g., :code:`batch_norm`, behave differently between + training and testing. They have an attribute, :code:`is_test`, to + control this behaviour. This method will change the :code:`is_test` + attribute of them to :code:`True` when :code:`for_test=True`. - Set for_test to False when we want to clone the program for training. - Set for_test to True when we want to clone the program for testing. + * Set for_test to False when we want to clone the program for training. + * Set for_test to True when we want to clone the program for testing. + + Notes: This API DOES NOT prune any operator. Use + :code:`clone(for_test=True)` before backward and optimization please. e.g. + + >>> test_program = fluid.default_main_program().clone(for_test=True) + >>> optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9) + >>> optimizer.minimize() Args: - for_test(bool): Some operators, such as batch_norm and drop_out ops, - behave differently in training and testing. If for_test is True, - the is_test attributes in these operators will be set to True for - testing purposes, otherwise, they remain unchanged. + for_test(bool): True if change the :code:`is_test` attribute of + operators to :code:`True`. - Returns(Program): - The cloned Program object. + Returns: + Program: The new, duplicated Program object. + + Examples: + + 1. To clone a test program, the sample code is: + + >>> import paddle.fluid as fluid + >>> train_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(train_program, startup_program): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> hidden = fluid.layers.fc(input=img, size=200, act='relu') + >>> hidden = fluid.layers.dropout(hidden, dropout_prob=0.5) + >>> loss = fluid.layers.cross_entropy( + >>> input=fluid.layers.fc(hidden, size=10, act='softmax'), + >>> label=fluid.layers.data(name='label', shape=[1], dtype='int64')) + >>> + >>> test_program = train_program.clone(for_test=True) + >>> + >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3) + >>> with fluid.program_guard(train_program, startup_program): + >>> sgd.minimize(loss) + + 2. The :code:`clone` method can be avoid if you create program for + training and program for testing individually. + + >>> import paddle.fluid as fluid + >>> + >>> def network(is_test): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> hidden = fluid.layers.fc(input=img, size=200, act='relu') + >>> hidden = fluid.layers.dropout(hidden, dropout_prob=0.5, is_test=is_test) + >>> loss = fluid.layers.cross_entropy( + >>> input=fluid.layers.fc(hidden, size=10, act='softmax'), + >>> label=fluid.layers.data(name='label', shape=[1], dtype='int64')) + >>> return loss + >>> + >>> train_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> test_program = fluid.Program() + >>> + >>> with fluid.program_guard(train_program, startup_program): + >>> with fluid.unique_name.guard(): + >>> loss = network(is_test=False) + >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3) + >>> sgd.minimize(loss) + >>> + >>> # the test startup program is not used. + >>> with fluid.program_guard(test_program, fluid.Program()): + >>> with fluid.unique_name.guard(): + >>> loss = network(is_test=True) + + The two code snippets above will generate same programs. """ if for_test: - p = self.inference_optimize() + p = self.inference_optimize(export_for_deployment=False) else: p = Program() + p.current_block_idx = self.current_block_idx + p._seed = self._seed p.desc = core.ProgramDesc(self.desc) p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] - p.sync_with_cpp() - p.copy_param_info_from(self) + p._current_role = self._current_role + p._op_role_var = self._op_role_var + + p._sync_with_cpp() + + p._copy_param_info_from(self) p.copy_data_info_from(self) return p def prune(self, targets): + """ + Prune operators and variables which are not needed to generate + :code:`targets`. + + Notes: This is a very low level API. Users should not use this API + directly. This API is in flux and not stable. + + Args: + targets(list|Variable|Operator): A list of variables or operators + need to be pruned + + Returns: + Program: A new, pruned program. + + """ if not isinstance(targets, list): targets = [targets] targets_idx = [] @@ -1140,39 +1632,97 @@ class Program(object): targets_idx.append([t.block.idx, t.idx]) res = Program() res.desc = core.prune(self.desc, targets_idx) - res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())] - res.sync_with_cpp() + res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())] + res._sync_with_cpp() return res - def inference_optimize(self): + def inference_optimize(self, export_for_deployment=True): + """ + This method will create a new program and do following adjustments on it: + 1. Remove all reader variables and their creator ops if exist. + + 2. Remove the :code:`read_op` if exists. + + 3. change the :code:`is_test` + attribute of operators to :code:`True`. All the :code:`Parameter` + information will be lost. + + Args: + export_for_deployment(bool): remove the read ops that are added by py_reader + for cpp inference library + + Notes: This API is a very low level API. Use + :code:`Program.clone(for_test=True)` instead. + + Returns: + Program: The new program. + """ # this is an alternative implement before # core.inference_optimize being fixed. res = Program() res.desc = core.ProgramDesc(self.desc) - for i in xrange(res.desc.num_blocks()): + + # remove all readers and the read_op if exist + read_op_idx = 0 + root_block = res.desc.block(0) + if export_for_deployment: + while True: + if read_op_idx >= root_block.op_size() or root_block.op( + read_op_idx).type() == 'read': + break + read_op_idx += 1 + if read_op_idx < root_block.op_size(): + root_block._remove_op(0, read_op_idx + 1) + for var in root_block.all_vars(): + if var.type() == core.VarDesc.VarType.READER: + root_block._remove_var(var.name()) + + # change all `is_test` attributes to True + for i in range(res.desc.num_blocks()): block = res.desc.block(i) - for j in xrange(block.op_size()): + for j in range(block.op_size()): op = block.op(j) if op.has_attr('is_test'): op.set_attr('is_test', True) - res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())] - res.sync_with_cpp() + res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())] + res._sync_with_cpp() return res @staticmethod def parse_from_string(binary_str): + """ + Deserialize a program desc from protobuf binary string. + + Notes: All information about parameters will be lost after serialization + and deserialization. + + Args: + binary_str_type(str): The binary prootbuf string. + + Returns: + Program: A deserialized program desc. + """ p = Program() p.desc = core.ProgramDesc(binary_str) - p.blocks = [Block(p, i) for i in xrange(p.desc.num_blocks())] - p.sync_with_cpp() + p.blocks = [Block(p, i) for i in range(p.desc.num_blocks())] + p._sync_with_cpp() return p @property def random_seed(self): + """ + The default random seed for random operators in Program. Zero means get + the random seed from random device. + + Notes: It must be set before the operators have been added. + """ return self._seed @property def num_blocks(self): + """ + The number of blocks in this program. + """ return self.desc.num_blocks() @random_seed.setter @@ -1182,18 +1732,43 @@ class Program(object): self._seed = seed def __repr__(self): - return str(self) + return self.__str__() def global_block(self): + """ + Get the first block of this program. + """ return self.blocks[0] def block(self, index): + """ + Get the :code:`index` block of this program + Args: + index(int): The index of block to get + + Returns: + Block: The :code:`index` block + """ return self.blocks[index] def current_block(self): + """ + Get the current block. The :code:`current` block is the block to append + operators. + """ return self.blocks[self.current_block_idx] def create_block(self, parent_idx=None): + """ + Create a new block with the :code:`parent_idx` and change the current block + to new block. + + Args: + parent_idx(int): The parent block index. + + Returns: + Block: The new block. + """ new_block_idx = len(self.blocks) parent = self.current_block() if parent_idx is None else self.block( parent_idx) @@ -1203,17 +1778,36 @@ class Program(object): return self.current_block() def rollback(self): + """ + Exit a code block, i.e., roll back to the parent block. + Returns: + None + """ self.current_block_idx = self.current_block().parent_idx - def sync_with_cpp(self): + def _sync_with_cpp(self): + """ + Synchronize Python instance to its binding C++ object instance. + If the program is modified in C++ space, this method should be invoked. + + Notes: This is a very low level API. Users should not invoke it + directly. + + Returns: + None + """ for block_idx in range(len(self.blocks), self.desc.num_blocks()): self.blocks.append(Block(self, block_idx)) for block in self.blocks: - block.sync_with_cpp() + block._sync_with_cpp() - def copy_param_info_from(self, other): + def _copy_param_info_from(self, other): """ Copy the information of parameters from other program. + + Notes: This is a very low level API. Users should not invoke it + directly. + Args: other(Program): Other program @@ -1221,17 +1815,21 @@ class Program(object): None """ if not isinstance(other, Program): - raise TypeError("copy_param_info_from should be invoked with " + raise TypeError("_copy_param_info_from should be invoked with " "Program") if len(self.blocks) != len(other.blocks): - raise ValueError("copy_param_info_from should be invoked with two " + raise ValueError("_copy_param_info_from should be invoked with two " "program, with represent the same topology") - self.global_block().copy_param_info_from(other.global_block()) + self.global_block()._copy_param_info_from(other.global_block()) def copy_data_info_from(self, other): """ Copy the information of data variables from other program. + + Notes: This is a very low level API. Users should not invoke it + directly. + Args: other(Program): Other program @@ -1239,23 +1837,52 @@ class Program(object): None """ if not isinstance(other, Program): - raise TypeError("copy_param_info_from should be invoked with " + raise TypeError("_copy_param_info_from should be invoked with " "Program") if len(self.blocks) != len(other.blocks): - raise ValueError("copy_param_info_from should be invoked with two " + raise ValueError("_copy_param_info_from should be invoked with two " "program, with represent the same topology") - for var in other.global_block().vars.itervalues(): + for var in list(other.global_block().vars.values()): if var.is_data: self.global_block().var(var.name).is_data = True def list_vars(self): + """ + Get all variables from this Program. A iterable object is returned. + + Returns: + iterable: The generator will yield every variable in this program. + """ for each_block in self.blocks: - for each_var in each_block.vars.itervalues(): + for each_var in list(each_block.vars.values()): yield each_var class Parameter(Variable): + """ + Parameter is derived from Variable. A parameter is a persistable + Variable, and will be updated by optimizers after each iteration. + The training of a neural network is essentially the updating of + its parameters. + + Relative to a general Variable, a Parameter has several its own + member variables: + + Args: + trainable(bool): True if the parameter need to be updated after + iterations. + optimize_attr(map): Parameter attributes related with optimizing. + Currently, it only contains 'learning_rate'. + Default: {'learning_rate': 1.0} + regularizer(WeightDecayRegularizer): The Regularizer which will + be applied on the parameter. Default: None + gradient_clip_attr(BaseGradientClipAttr): The gradint clip strategy + which will be applied on the parameter. Default: None + do_model_average(bool): True if the model average strategy will + be applied on this parameter. + """ + def __init__(self, block, shape, dtype, **kwargs): if shape is None or dtype is None: raise ValueError("Parameter must set shape and dtype") @@ -1285,6 +1912,7 @@ class Parameter(Variable): def to_string(self, throw_on_error, with_details=False): """ To debug string. + Args: throw_on_error(bool): raise exception when self is not initialized when throw_on_error is True @@ -1301,8 +1929,8 @@ class Parameter(Variable): additional_attr = ("trainable", "optimize_attr", "regularizer", "gradient_clip_attr", "do_model_average") for attr_name in additional_attr: - res_str += "%s: %s\n" % (attr_name, - str(getattr(self, attr_name))) + res_str += "%s: %s\n" % ( + attr_name, six.binary_type(getattr(self, attr_name))) else: res_str = Variable.to_string(self, throw_on_error, False) return res_str @@ -1317,8 +1945,15 @@ _startup_program_ = Program() def default_startup_program(): """ - Get default startup program. In startup program, Paddle will initialize - parameters, initialize nccl handle, etc. + Get default/global startup program. + + The layer function in :code:`fluid.layers` will create parameters, readers, + NCCL handles as global variables. The :code:`startup_program` will + initialize them by the operators in startup program. The layer function will + append these initialization operators into startup program. + + This method will return the :code:`default` or the :code:`current` startup + program. Users can use :code:`fluid.program_guard` to switch program. Returns: Program: startup program @@ -1328,7 +1963,15 @@ def default_startup_program(): def default_main_program(): """ - Get default main program. The main program is used for training or testing. + Get default/global main program. The main program is used for training or + testing. + + All layer function in :code:`fluid.layers` will append operators and + variables to the :code:`default_main_program`. + + The :code:`default_main_program` is the default program in a lot of APIs. + For example, the :code:`Executor.run()` will execute the + :code:`default_main_program` when the program is not specified. Returns: Program: main program @@ -1370,20 +2013,34 @@ def switch_startup_program(program): @contextlib.contextmanager def program_guard(main_program, startup_program=None): """ - Switch program with `with` statement + Change the global main program and startup program with `with` statement. + Layer functions in the Python `with` block will append operators and + variables to the new main programs. + + Examples: + + >>> import paddle.fluid as fluid + >>> main_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(main_program, startup_program): + >>> data = fluid.layers.data(...) + >>> hidden = fluid.layers.fc(...) + + Notes: The temporary :code:`Program` can be used if the user does not need + to construct either of startup program or main program. Examples: - >>> with program_guard(Program()): - >>> data = fluid.layers.data(...) - >>> hidden = fluid.layers.fc(...) + + >>> import paddle.fluid as fluid + >>> main_program = fluid.Program() + >>> # does not care about startup program. Just pass a temporary value. + >>> with fluid.program_guard(main_program, fluid.Program()): + >>> data = ... Args: - main_program(Program): New main program inside `with` statement + main_program(Program): New main program inside `with` statement. startup_program(Program): New startup program inside `with` statement. None means do not change startup program. - - Returns: - None """ if not isinstance(main_program, Program): raise TypeError("main_program should be Program") @@ -1400,11 +2057,12 @@ def program_guard(main_program, startup_program=None): def get_var(name, program=None): """ - Get a variable by name from the global block of a program + Get a variable by name from the global block of a program. + Args: name(str): name of the variable program(Program|None): program object. - If None, default_global_program() will be used. + If None, default_global_program() will be used. Returns: Variable diff --git a/python/paddle/fluid/graphviz.py b/python/paddle/fluid/graphviz.py index 125b4efa9d476e561bd78d0365cd92bbf7e66605..ba67bf5ae6fe44ea23414d444a270c436c195326 100644 --- a/python/paddle/fluid/graphviz.py +++ b/python/paddle/fluid/graphviz.py @@ -14,12 +14,13 @@ import os import random +import six import subprocess import logging def crepr(v): - if type(v) is str or type(v) is unicode: + if isinstance(v, six.string_types): return '"%s"' % v return str(v) @@ -104,7 +105,7 @@ class Graph(object): def _rank_repr(self): ranks = sorted( - self.rank_groups.items(), + list(self.rank_groups.items()), cmp=lambda a, b: a[1].priority > b[1].priority) repr = [] for x in ranks: @@ -148,7 +149,7 @@ class Node(object): name=self.name, label=self.label, extra=',' + ','.join("%s=%s" % (key, crepr(value)) - for key, value in self.attrs.items()) + for key, value in list(self.attrs.items())) if self.attrs else "") return reprs @@ -172,7 +173,7 @@ class Edge(object): target=self.target.name, extra="" if not self.attrs else "[" + ','.join("{}={}".format(attr[0], crepr(attr[1])) - for attr in self.attrs.items()) + "]") + for attr in list(self.attrs.items())) + "]") return repr diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py index 9f242cf29a56573349f192307a68e135a409a4be..ff382d8b832b4b2bc6779dbb28d3fd95c8a0984e 100644 --- a/python/paddle/fluid/inferencer.py +++ b/python/paddle/fluid/inferencer.py @@ -14,26 +14,43 @@ import contextlib -import core +from . import core -import executor -import framework -import io -import parallel_executor -import unique_name -from trainer import check_and_get_place +from . import executor +from . import framework +from . import io +from . import parallel_executor +from . import unique_name +from .trainer import check_and_get_place __all__ = ['Inferencer', ] class Inferencer(object): + """ + Inferencer High Level API. + + Args: + infer_func (Python func): Infer function that will return predict Variable + param_path (str): The path where the inference model is saved by fluid.io.save_params + place (Place): place to do the inference + parallel (bool): use parallel_executor to run the inference, it will use multi CPU/GPU. + + Examples: + .. code-block:: python + + def inference_program(): + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + return y_predict + + place = fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path="/tmp/model", place=place) + + """ + def __init__(self, infer_func, param_path, place=None, parallel=False): - """ - :param infer_func: a function that will return predict Variable - :param param_path: the path where the inference model is saved by fluid.io.save_params - :param place: place to do the inference - :param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU. - """ self.param_path = param_path self.scope = core.Scope() self.parallel = parallel @@ -56,11 +73,24 @@ class Inferencer(object): else: self.exe = executor.Executor(self.place) + self.inference_program = self.inference_program.clone(for_test=True) + def infer(self, inputs, return_numpy=True): """ - :param inputs: a map of {"input_name": input_var} that will be feed into the inference program - to get the predict value - :return: the predict value of the inference model + Do Inference for Inputs + + Args: + inputs (map): a map of {"input_name": input_var} that will be feed into the inference program + return_numpy (bool): transform return value into numpy or not + + Returns: + Tensor or Numpy: the predict value of the inference model for the inputs + + Examples: + .. code-block:: python + + tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32") + results = inferencer.infer({'x': tensor_x}) """ if not isinstance(inputs, dict): raise ValueError( diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py index 4e132ed26183eaa5e572128e679cdbffd42e5a42..6dedbae7a6586f862328c7f23d0aea6ba5022614 100644 --- a/python/paddle/fluid/initializer.py +++ b/python/paddle/fluid/initializer.py @@ -12,31 +12,45 @@ # See the License for the specific language governing permissions and # limitations under the License. -import framework +from . import framework import numpy as np import contextlib +from .core import VarDesc __all__ = [ - 'Constant', 'Uniform', 'Normal', 'Xavier', 'force_init_on_cpu', - 'init_on_cpu', 'ConstantInitializer', 'UniformInitializer', - 'NormalInitializer', 'XavierInitializer' + 'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'MSRA', + 'force_init_on_cpu', 'init_on_cpu', 'ConstantInitializer', + 'UniformInitializer', 'NormalInitializer', 'XavierInitializer', + 'BilinearInitializer', 'MSRAInitializer' ] _force_init_on_cpu_ = False def force_init_on_cpu(): + """ + The flag of whether force to init variables on CPU. + + Examples: + .. code-block:: python + + if force_init_on_cpu(): + pass + + """ return _force_init_on_cpu_ @contextlib.contextmanager def init_on_cpu(): """ - Switch program with `with` statement + Force the variable to be inited on CPU. Examples: - >>> with init_on_cpu(): - >>> step = layers.create_global_var() + .. code-block:: python + + with init_on_cpu(): + step = layers.create_global_var() """ global _force_init_on_cpu_ @@ -102,14 +116,18 @@ class Initializer(object): class ConstantInitializer(Initializer): """Implements the constant initializer + + Args: + value (float): constant value to initialize the variable + + Examples: + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10, + param_attr=fluid.initializer.Constant(value=2.0)) """ def __init__(self, value=0.0, force_cpu=False): - """Constructor for ConstantInitializer - - Args: - value: constant value to initialize the variable - """ assert value is not None super(ConstantInitializer, self).__init__() self._value = value @@ -129,7 +147,7 @@ class ConstantInitializer(Initializer): assert isinstance(var, framework.Variable) assert isinstance(block, framework.Block) # Initialization Ops should be prepended and not appended - op = block.prepend_op( + op = block._prepend_op( type="fill_constant", outputs={"Out": var}, attrs={ @@ -144,16 +162,20 @@ class ConstantInitializer(Initializer): class UniformInitializer(Initializer): """Implements the random uniform distribution initializer + + Args: + low (float): lower boundary of the uniform distribution + high (float): upper boundary of the uniform distribution + seed (int): random seed + + Examples: + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10, + param_attr=fluid.initializer.Uniform(low=-0.5, high=0.5)) """ def __init__(self, low=-1.0, high=1.0, seed=0): - """Constructor for UniformInitializer - - Args: - low: lower boundary of the uniform distribution - high: upper boundary of the uniform distribution - seed: random seed - """ assert low is not None assert high is not None assert high >= low @@ -179,7 +201,7 @@ class UniformInitializer(Initializer): # Initialization Ops should be prepended and not appended if self._seed == 0: self._seed = block.program.random_seed - op = block.prepend_op( + op = block._prepend_op( type="uniform_random", outputs={"Out": var}, attrs={ @@ -194,17 +216,21 @@ class UniformInitializer(Initializer): class NormalInitializer(Initializer): - """Implements the random Normal(Gaussian) distribution initializer + """Implements the Random Normal(Gaussian) distribution initializer + + Args: + loc (float): mean of the normal distribution + scale (float): standard deviation of the normal distribution + seed (int): random seed + + Examples: + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10, + param_attr=fluid.initializer.Normal(loc=0.0, scale=2.0)) """ def __init__(self, loc=0.0, scale=1.0, seed=0): - """Constructor for NormalInitializer - - Args: - loc: mean of the normal distribution - scale: standard deviation of the normal distribution - seed: random seed - """ assert loc is not None assert scale is not None assert seed is not None @@ -229,7 +255,7 @@ class NormalInitializer(Initializer): # Initialization Ops should be prepended and not appended if self._seed == 0: self._seed = block.program.random_seed - op = block.prepend_op( + op = block._prepend_op( type="gaussian_random", outputs={"Out": var}, attrs={ @@ -237,46 +263,57 @@ class NormalInitializer(Initializer): "dtype": int(var.dtype), "mean": self._mean, "std": self._std_dev, - "seed": self._seed + "seed": self._seed, + "use_mkldnn": False }) var.op = op return op class XavierInitializer(Initializer): - """Implements the Xavier initializer - + """ This class implements the Xavier weight initializer from the paper - Understanding the difficulty of training deep feedforward neural - networks[1] by Xavier Glorot and Yoshua Bengio. + `Understanding the difficulty of training deep feedforward neural + networks `_ + by Xavier Glorot and Yoshua Bengio. This initializer is designed to keep the scale of the gradients approximately same in all the layers. In case of Uniform distribution, - the range is [-x, x], where x = sqrt(6 / (fan_in + fan_out)). + the range is [-x, x], where + + .. math:: + + x = \sqrt{\\frac{6.0}{fan\_in + fan\_out}} + In case of Normal distribution, the mean is 0 and the standard deviation - is sqrt(2/ (fan_in + fan_out)). + is + + .. math:: + + \sqrt{\\frac{2.0}{fan\_in + fan\_out}} + + + Args: + uniform (bool): whether to use uniform or normal distribution + fan_in (float): fan_in for Xavier initialization. If None, it is + inferred from the variable. + fan_out (float): fan_out for Xavier initialization. If None, it is + inferred from the variable. + seed (int): random seed + + Note: + It is recommended to set fan_in and fan_out to None for most cases. + + Examples: + .. code-block:: python + + fc = fluid.layers.fc( + input=queries, size=10, + param_attr=fluid.initializer.Xavier(uniform=False)) - References: - [1] Understanding the difficulty of training deep feedforward neural - networks. International conference on artificial intelligence and - statistics. - (http://proceedings.mlr.press/v9/glorot10a.html) """ def __init__(self, uniform=True, fan_in=None, fan_out=None, seed=0): - """Constructor for XavierInitializer - - Args: - uniform: whether to use uniform or normal distribution - fan_in: fan_in for Xavier initialization. If None, it is - inferred from the variable. - fan_out: fan_out for Xavier initialization. If None, it is - inferred from the variable. - seed: random seed - - Note: It is recommended to set fan_in and fan_out to None for - most cases. - """ assert uniform is not None assert seed is not None super(XavierInitializer, self).__init__() @@ -309,7 +346,7 @@ class XavierInitializer(Initializer): if self._uniform: limit = np.sqrt(6.0 / float(fan_in + fan_out)) - op = block.prepend_op( + op = block._prepend_op( type="uniform_random", outputs={"Out": var}, attrs={ @@ -322,7 +359,7 @@ class XavierInitializer(Initializer): else: std = np.sqrt(2.0 / float(fan_in + fan_out)) - op = block.prepend_op( + op = block._prepend_op( type="gaussian_random", outputs={"Out": var}, attrs={ @@ -340,30 +377,42 @@ class MSRAInitializer(Initializer): """Implements the MSRA initializer a.k.a. Kaiming Initializer This class implements the weight initialization from the paper - Delving Deep into Rectifiers: Surpassing Human-Level Performance on - ImageNet Classification[1] by Kaiming He, Xiangyu Zhang, Shaoqing Ren - and Jian Sun. This is a robust initialization method that particularly - considers the rectifier nonlinearities. In case of Uniform distribution, - the range is [-x, x], where x = sqrt(6 / fan_in). In case of Normal - distribution, the mean is 0 and the standard deviation - is sqrt(2/ fan_in). - - References: - [1] Delving Deep into Rectifiers: Surpassing Human-Level Performance - on ImageNet Classification - (https://arxiv.org/abs/1502.01852) + `Delving Deep into Rectifiers: Surpassing Human-Level Performance on + ImageNet Classification `_ + by Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. This is a + robust initialization method that particularly considers the rectifier + nonlinearities. In case of Uniform distribution, the range is [-x, x], where + + .. math:: + + x = \sqrt{\\frac{6.0}{fan\_in}} + + In case of Normal distribution, the mean is 0 and the standard deviation + is + + .. math:: + + \sqrt{\\frac{2.0}{fan\_in}} + + Args: + uniform (bool): whether to use uniform or normal distribution + fan_in (float): fan_in for MSRAInitializer. If None, it is\ + inferred from the variable. + seed (int): random seed + + Note: + It is recommended to set fan_in to None for most cases. + + Examples: + .. code-block:: python + + fc = fluid.layers.fc( + input=queries, size=10, + param_attr=fluid.initializer.MSRA(uniform=False)) """ def __init__(self, uniform=True, fan_in=None, seed=0): """Constructor for MSRAInitializer - - Args: - uniform: whether to use uniform or normal distribution - fan_in: fan_in for MSRAInitializer. If None, it is - inferred from the variable. - seed: random seed - - Note: It is recommended to set fan_in to None for most cases. """ assert uniform is not None assert seed is not None @@ -395,7 +444,7 @@ class MSRAInitializer(Initializer): if self._uniform: limit = np.sqrt(6.0 / float(fan_in)) - op = block.prepend_op( + op = block._prepend_op( type="uniform_random", outputs={"Out": var}, attrs={ @@ -408,7 +457,7 @@ class MSRAInitializer(Initializer): else: std = np.sqrt(2.0 / float(fan_in)) - op = block.prepend_op( + op = block._prepend_op( type="gaussian_random", outputs={"Out": var}, attrs={ @@ -422,6 +471,104 @@ class MSRAInitializer(Initializer): return op +class BilinearInitializer(Initializer): + """ + This initializer can be used in transposed convolution operator to + act as upsampling. Users can upsample a feature map with shape of + (B, C, H, W) by any integer factor. The usage is: + + Examples: + + .. code-block:: python + + factor = 2 + w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.), + initializer=Bilinear()) + conv_up = fluid.layers.conv2d_transpose( + input, + num_filters=C, + output_size=None, + filter_size=2 * factor - factor % 2, + padding=ceil((factor - 1) / 2.), + stride=factor, + groups=C, + param_attr=w_attr, + bias_attr=False) + + Where, `num_filters=C` and `groups=C` means this is channel-wise transposed + convolution. The filter shape will be (C, 1, K, K) where K is `filer_size`, + This initializer will set a (K, K) interpolation kernel for every channel + of the filter identically. The resulting shape of the output feature map + will be (B, C, factor * H, factor * W). Note that the learning rate and the + weight decay are set to 0 in order to keep coefficient values of bilinear + interpolation unchanged during training. + + """ + + def __init__(self): + """Constructor for BilinearInitializer. + """ + super(BilinearInitializer, self).__init__() + + def __call__(self, var, block): + """Add biliear initialization ops for a variable + + Args: + var (Variable): Variable that needs to be initialized. + block (Block): The block in which initialization ops should + be added. + + Returns: + Operator: the initialization op + + Raises: + ValueError: If type of `var` and `block` is not right. + If the shape of `var` size is not 4 and + var.shape[2] != var.shape[3]. + """ + if not isinstance(var, framework.Variable): + raise ValueError("var must be framework.Variable.") + + if not isinstance(block, framework.Block): + raise ValueError("block must be framework.Block.") + + shape = var.shape + if len(shape) != 4: + raise ValueError("the length of shape must be 4.") + if shape[2] != shape[3]: + raise ValueError("shape[2] must be equal to shape[3].") + + weight = np.zeros(np.prod(var.shape), dtype='float32') + size = shape[3] + # factor + f = np.ceil(size / 2.) + # center + c = (2 * f - 1 - f % 2) / (2. * f) + for i in range(np.prod(shape)): + x = i % size + y = (i / size) % size + weight[i] = (1 - abs(x / f - c)) * (1 - abs(y / f - c)) + weight = np.reshape(weight, shape) + + if var.dtype == VarDesc.VarType.FP32: + value_name = "fp32_values" + values = [float(v) for v in weight.flat] + else: + raise ValueError("Unsupported dtype %s", input.dtype) + if np.prod(shape) > 1024 * 1024: + raise ValueError("The size of input is too big. ") + op = block.append_op( + type='assign_value', + outputs={'Out': [var]}, + attrs={ + 'dtype': var.dtype, + 'shape': list(shape), + value_name: values + }) + var.op = op + return op + + # We short the class name, since users will use the initializer with the package # name. The sample code: # @@ -436,3 +583,4 @@ Uniform = UniformInitializer Normal = NormalInitializer Xavier = XavierInitializer MSRA = MSRAInitializer +Bilinear = BilinearInitializer diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index 8e58e5eb794e1bb507ab05394a1f7b57a1d2ed42..af734210323913a36f861380dc38a98253aca0a1 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -13,38 +13,62 @@ # limitations under the License. import os +import errno import time import shutil +import six from paddle.fluid.evaluator import Evaluator -from paddle.fluid.framework import Program, Parameter, default_main_program, Variable +from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable from . import core __all__ = [ 'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params', 'load_persistables', 'save_inference_model', 'load_inference_model', - 'get_inference_program', 'save_checkpoint', 'load_checkpoint', - 'clean_checkpoint' + 'get_inference_program' ] def is_parameter(var): - """Check whether the variable is a Parameter. - - This function checks whether the input variable is a Parameter. + """ + Check whether the given variable is an instance of Parameter. Args: - var : The input variable. + var(Variable): The variable to be checked. Returns: - boolean result whether the variable is a Parameter. + bool: True if the given `var` is an instance of Parameter, + False if not. + + Examples: + .. code-block:: python + + param = fluid.default_main_program().global_block().var('fc.w') + res = fluid.io.is_parameter(param) """ return isinstance(var, Parameter) def is_persistable(var): + """ + Check whether the given variable is persistable. + + Args: + var(Variable): The variable to be checked. + + Returns: + bool: True if the given `var` is persistable + False if not. + + Examples: + .. code-block:: python + + param = fluid.default_main_program().global_block().var('fc.w') + res = fluid.io.is_persistable(param) + """ if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ - var.desc.type() == core.VarDesc.VarType.FETCH_LIST: + var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ + var.desc.type() == core.VarDesc.VarType.READER: return False return var.persistable @@ -67,20 +91,69 @@ def save_vars(executor, predicate=None, filename=None): """ - Save variables to directory by executor. - - :param executor: executor that save variable - :param dirname: directory path - :param main_program: program. If vars is None, then filter all variables in this - program which fit `predicate`. Default default_main_program. - :param predicate: The Predicate describes a callable that returns a variable - as a bool. If it returns true, the corresponding input variable will be saved. - :param vars: variables need to be saved. If vars is specified, program & predicate - will be ignored - :param filename: The name of a single file that all vars are saved to. - If it is None, save variables to separate files. - - :return: None + Save variables to the given directory by executor. + + There are two ways to specify variables to be saved: The first way, list + variables in a list and assign it to the `vars`. The second way, assign the + `main_program` with an existing program, then all variables in the program + will be saved. The first way has a higher priority. In other words, if `vars` + are assigned, the `main_program` and the `predicate` will be ignored. + + The `dirname` are used to specify the folder where to save variables. + If you prefer to save variables in separate files in the folder `dirname`, + set `filename` None; if you prefer to save all variables in a single file, + use `filename` to specify it. + + Args: + executor(Executor): The executor to run for saving variables. + dirname(str): The directory path. + main_program(Program|None): The program whose variables will be saved. + If it is None, the default main program will + be used automatically. + Default: None + vars(list[Variable]|None): The list that contains all variables to save. + It has a higher priority than the `main_program`. + Default: None + predicate(function|None): If it is not None, only variables in the + `main_program` that makes predicate(variable)==True + will be saved. It only works when we are using the + `main_program` to specify variables (In other words + `vars` is None). + Default: None + filename(str|None): The file which to save all variables. If you prefer to save + variables separately, set it to None. + Default: None + + Returns: + None + + Raises: + TypeError: If `main_program` is not an instance of Program nor None. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + + # The first usage: using `main_program` to specify variables + def name_has_fc(var): + res = "fc" in var.name + return res + + prog = fluid.default_main_program() + fluid.io.save_vars(executor=exe, dirname=path, main_program=prog, + vars=None) + # All variables in `main_program` whose name includes "fc" will be saved. + # And variables are going to be saved separately. + + + # The second usage: using `vars` to specify variables + var_list = [var_a, var_b, var_c] + fluid.io.save_vars(executor=exe, dirname=path, vars=var_list, + filename="vars_file") + # var_a, var_b and var_c will be saved. And they are going to be + # saved in the same file named 'var_file' in the path "./my_paddle_model". """ if vars is None: if main_program is None: @@ -91,7 +164,7 @@ def save_vars(executor, save_vars( executor, dirname=dirname, - vars=filter(predicate, main_program.list_vars()), + vars=list(filter(predicate, main_program.list_vars())), filename=filename) else: save_program = Program() @@ -128,7 +201,42 @@ def save_vars(executor, def save_params(executor, dirname, main_program=None, filename=None): """ - Save all parameters to directory with executor. + This function filters out all parameters from the give `main_program` + and then save them to the folder `dirname` or the file `filename`. + + Use the `dirname` to specify the saving folder. If you would like to + save parameters in separate files, set `filename` None; if you would + like to save all parameters in a single file, use `filename` to specify + the file name. + + NOTICE: Some variables are not Parameter while they are necessary for + training. So you can NOT save and continue your training just by + `save_params()` and `load_params()`. Please use `save_persistables()` + and `load_persistables()` instead. + + Args: + executor(Executor): The executor to run for saving parameters. + dirname(str): The saving directory path. + main_program(Program|None): The program whose parameters will be + saved. If it is None, the default + main program will be used automatically. + Default: None + filename(str|None): The file to save all parameters. If you prefer + to save parameters in differnet files, set it + to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.save_params(executor=exe, dirname=param_path, + main_program=None) """ save_vars( executor, @@ -141,7 +249,37 @@ def save_params(executor, dirname, main_program=None, filename=None): def save_persistables(executor, dirname, main_program=None, filename=None): """ - Save all persistables to directory with executor. + This function filters out all variables with `persistable==True` from the + give `main_program` and then saves these variables to the folder `dirname` + or file `filename`. + + The `dirname` is used to specify the folder where persistable variables + are going to be saved. If you would like to save variables in separate + files, set `filename` None; if you would like to save all variables in a + single file, use `filename` to specify the file name. + + Args: + executor(Executor): The executor to run for saving persistable variables. + dirname(str): The directory path. + main_program(Program|None): The program whose persistbale variables will + be saved. If it is None, the default main + program will be used automatically. + Default: None + filename(str|None): The file to saved all variables. If you prefer to + save variables in differnet files, set it to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.save_persistables(executor=exe, dirname=param_path, + main_program=None) """ save_vars( executor, @@ -159,20 +297,69 @@ def load_vars(executor, predicate=None, filename=None): """ - Load variables from directory by executor. - - :param executor: executor that load variable - :param dirname: directory path - :param main_program: program. If vars is None, then filter all variables in this - program which fit `predicate`. Default default_main_program(). - :param predicate: The Predicate describes a callable that returns a variable - as a bool. If it returns true, the corresponding input variable will be loaded. - :param vars: variables need to be loaded. If vars is specified, program & - predicate will be ignored - :param filename: The name of the single file that all vars are loaded from. - If it is None, load variables from separate files. - - :return: None + Load variables from the given directory by executor. + + There are two ways to specify variables to be loaded: The first way, list + variables in a list and assign it to the `vars`. The second way, assign the + `main_program` with an existing program, then all variables in the program + will be loaded. The first way has a higher priority. In other words if `vars` + are assigned, the `main_program` and the `predicate` will be ignored. + + The `dirname` are used to specify the folder where to load variables. + If variables were saved in separate files in the folder `dirname`, + set `filename` None; if all variables were saved in a single file, + use `filename` to specify it. + + Args: + executor(Executor): The executor to run for loading variables. + dirname(str): The directory path. + main_program(Program|None): The program whose variables will be loaded. + If it is None, the default main program will + be used automatically. + Default: None + vars(list[Variable]|None): The list that contains all variables to load. + It has a higher priority than the `main_program`. + Default: None + predicate(function|None): If it is not None, only variables in the + `main_program` that makes predicate(variable)==True + will be loaded. It only works when we are using the + `main_program` to specify variables (In other words + `vars` is None). + Default: None + filename(str|None): The file which saved all required variables. If variables + were saved in differnet files, set it to None. + Default: None + + Returns: + None + + Raises: + TypeError: If `main_program` is not an instance of Program nor None. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + + # The first usage: using `main_program` to specify variables + def name_has_fc(var): + res = "fc" in var.name + return res + + prog = fluid.default_main_program() + fluid.io.load_vars(executor=exe, dirname=path, main_program=prog, + vars=None) + # All variables in `main_program` whose name includes "fc" will be loaded. + # And all the variables are supposed to have been saved in differnet files. + + + # The second usage: using `vars` to specify variables + var_list = [var_a, var_b, var_c] + fluid.io.load_vars(executor=exe, dirname=path, vars=var_list, + filename="vars_file") + # var_a, var_b and var_c will be loaded. And they are supposed to haven + # been saved in the same file named 'var_file' in the path "./my_paddle_model". """ if vars is None: if main_program is None: @@ -183,7 +370,7 @@ def load_vars(executor, load_vars( executor, dirname=dirname, - vars=filter(predicate, main_program.list_vars()), + vars=list(filter(predicate, main_program.list_vars())), filename=filename) else: load_prog = Program() @@ -220,7 +407,42 @@ def load_vars(executor, def load_params(executor, dirname, main_program=None, filename=None): """ - load all parameters from directory by executor. + This function filters out all parameters from the give `main_program` + and then trys to load these parameters from the folder `dirname` or + the file `filename`. + + Use the `dirname` to specify the folder where parameters were saved. If + parameters were saved in separate files in the folder `dirname`, set + `filename` None; if all parameters were saved in a single file, use + `filename` to specify the file name. + + NOTICE: Some variables are not Parameter while they are necessary for + training. So you can NOT save and continue your training just by + `save_params()` and `load_params()`. Please use `save_persistables()` + and `load_persistables()` instead. + + Args: + executor(Executor): The executor to run for loading parameters. + dirname(str): The directory path. + main_program(Program|None): The program whose parameters will be + loaded. If it is None, the default + main program will be used automatically. + Default: None + filename(str|None): The file which saved all parameters. If parameters + were saved in differnet files, set it to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.load_params(executor=exe, dirname=param_path, + main_program=None) """ load_vars( executor, @@ -232,7 +454,37 @@ def load_params(executor, dirname, main_program=None, filename=None): def load_persistables(executor, dirname, main_program=None, filename=None): """ - load all persistables from directory by executor. + This function filters out all variables with `persistable==True` from the + give `main_program` and then trys to load these variables from the folder + `dirname` or the file `filename`. + + Use the `dirname` to specify the folder where persistable variables were + saved. If variables were saved in separate files, set `filename` None; + if all variables were saved in a single file, use `filename` to specify + the file name. + + Args: + executor(Executor): The executor to run for loading persistable variables. + dirname(str): The directory path. + main_program(Program|None): The program whose persistbale variables will + be loaded. If it is None, the default main + program will be used automatically. + Default: None + filename(str|None): The file which saved all variables. If variables were + saved in differnet files, set it to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.load_persistables(executor=exe, dirname=param_path, + main_program=None) """ load_vars( executor, @@ -273,7 +525,7 @@ def prepend_feed_ops(inference_program, for i, name in enumerate(feed_target_names): out = global_block.var(name) - global_block.prepend_op( + global_block._prepend_op( type='feed', inputs={'X': [feed_var]}, outputs={'Out': [out]}, @@ -303,32 +555,73 @@ def save_inference_model(dirname, executor, main_program=None, model_filename=None, - params_filename=None): + params_filename=None, + export_for_deployment=True): """ - Build a model especially for inference, - and save it to directory by the executor. - - :param dirname: directory path - :param feeded_var_names: Names of variables that need to be feeded data during inference - :param target_vars: Variables from which we can get inference results. - :param executor: executor that save inference model - :param main_program: original program, which will be pruned to build the inference model. - Default default_main_program(). - :param model_filename: The name of file to save inference program. - If not specified, default filename `__model__` will be used. - :param params_filename: The name of file to save parameters. - It is used for the case that all parameters are saved in a single binary file. - If not specified, parameters are considered saved in separate files. - - :return: None + Prune the given `main_program` to build a new program especially for inference, + and then save it and all related parameters to given `dirname` by the `executor`. + + Args: + dirname(str): The directory path to save the inference model. + feeded_var_names(list[str]): Names of variables that need to be feeded data + during inference. + target_vars(list[Variable]): Variables from which we can get inference + results. + executor(Executor): The executor that saves the inference model. + main_program(Program|None): The original program, which will be pruned to + build the inference model. If is setted None, + the default main program will be used. + Default: None. + model_filename(str|None): The name of file to save the inference program + itself. If is setted None, a default filename + `__model__` will be used. + params_filename(str|None): The name of file to save all related parameters. + If it is setted None, parameters will be saved + in separate files . + export_for_deployment(bool): remove the read ops that are added by py_reader + for cpp inference lib. Default True + + Returns: + None + + Raises: + ValueError: If `feed_var_names` is not a list of basestring. + ValueError: If `target_vars` is not a list of Variable. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./infer_model" + fluid.io.save_inference_model(dirname=path, feeded_var_names=['img'], + target_vars=[predict_var], executor=exe) + + # In this exsample, the function will prune the default main program + # to make it suitable for infering the `predict_var`. The pruned + # inference program is going to be saved in the "./infer_model/__model__" + # and parameters are going to be saved in separate files under folder + # "./infer_model". + """ - if isinstance(feeded_var_names, basestring): + if isinstance(feeded_var_names, six.binary_type): feeded_var_names = [feeded_var_names] + elif isinstance(feeded_var_names, six.text_type): + feeded_var_names = [feeded_var_names.encode()] else: if len(feeded_var_names) > 0: + # TODO(paddle-dev): polish these code blocks if not (bool(feeded_var_names) and all( - isinstance(name, basestring) for name in feeded_var_names)): - raise ValueError("'feed_var_names' should be a list of str.") + isinstance(name, six.binary_type) + for name in feeded_var_names)): + if not (all( + isinstance(name, six.text_type) + for name in feeded_var_names)): + raise ValueError( + "'feed_var_names' should be a list of str.") + else: + feeded_var_names = [ + name.encode() for name in feeded_var_names + ] if isinstance(target_vars, Variable): target_vars = [target_vars] @@ -349,11 +642,12 @@ def save_inference_model(dirname, for i, op in enumerate(global_block.ops): op.desc.set_is_target(False) if op.type == "feed" or op.type == "fetch": - global_block.remove_op(i) + global_block._remove_op(i) copy_program.desc.flush() pruned_program = copy_program.prune(targets=target_vars) - inference_program = pruned_program.inference_optimize() + inference_program = pruned_program.inference_optimize( + export_for_deployment=export_for_deployment) fetch_var_names = [v.name for v in target_vars] prepend_feed_ops(inference_program, feeded_var_names) @@ -381,18 +675,49 @@ def load_inference_model(dirname, """ Load inference model from a directory - :param dirname: directory path - :param executor: executor that load inference model - :param model_filename: The name of file to load inference program. - If not specified, default filename `__model__` will be used. - :param params_filename: The name of file to load parameters. - It is used for the case that all parameters are saved in a single binary file. - If not specified, parameters are considered saved in separate files. - - :return: [program, feed_target_names, fetch_targets] - program: program especially for inference. - feed_target_names: Names of variables that need to feed data - fetch_targets: Variables from which we can get inference results. + Args: + dirname(str): The directory path + executor(Executor): The executor to run for loading inference model. + model_filename(str|None): The name of file to load inference program. + If it is None, the default filename + '__model__' will be used. + Default: None + params_filename(str|None): The name of file to load all parameters. + It is only used for the case that all + parameters were saved in a single binary + file. If parameters were saved in separate + files, set it as 'None'. + + Returns: + tuple: The return of this function is a tuple with three elements: + (program, feed_target_names, fetch_targets). The `program` is a + Program, it's the program for inference. The `feed_target_names` is + a list of str, it contains Names of variables that need to feed + data in the inference program. The `fetch_targets` is a list of + Variable. It contains variables from which we can get inference + results. + + Raises: + ValueError: If `dirname` is not a existing directory. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./infer_model" + [inference_program, feed_target_names, fetch_targets] = + fluid.io.load_inference_model(dirname=path, executor=exe) + results = exe.run(inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + + # In this exsample, the inference program was saved in the + # "./infer_model/__model__" and parameters were saved in + # separate files in ""./infer_model". + # After getting inference program, feed target names and + # fetch targets, we can use an Executor to run the inference + # program to get the inference result. + """ if not os.path.isdir(dirname): raise ValueError("There is no directory named '%s'", dirname) @@ -423,12 +748,25 @@ def load_inference_model(dirname, def get_parameter_value(para, executor): """ - Get the LoDTensor for the parameter + Get the LoDTensor value of the given parameter. + + Args: + para(Parameter): The parameter to get value from. + executor(Executor): The executor to run for retrieving the value. - :param executor: executor for retrieving the value - :param para: the given parameter + Returns: + numpy.array: The given parameter's values. + + Raises: + AssertionError: If the `para` is not an instance of Parameter. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param = fluid.default_main_program().global_block().var('fc.w') + p = fluid.io.get_parameter_value(param, exe) - :return: the LoDTensor for the parameter """ assert is_parameter(para) @@ -440,205 +778,32 @@ def get_parameter_value(para, executor): def get_parameter_value_by_name(name, executor, program=None): """ - Get the LoDTensor for paramter with the given name + Get the LoDTensor value of a certain parameter by its name. + + Args: + name(str): The parameter's name. + executor(Executor): The executor to run for retrieving the value. + program(Program | None): The program where to find the parameter. + If it's set to be None, the function will + try to find the parameter in the default + main program. + + Returns: + numpy.array: The parameter's values. - :param executor: executor for retrieving the value - :param name: the name of the parameter - :param program: the program where the variable is found - Default default_main_program(). + Raises: + TypeError: If given `name` is not an instance of basestring. + TypeError: If the parameter with the given name doesn't exist. + AssertionError: If there is a varibale named `name` in the + given program but it is not a Parameter. - :return: the LoDTensor for the variable + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + p = fluid.io.get_parameter_value('fc.w', exe) """ if program is None: program = default_main_program() var = program.global_block().var(name) return get_parameter_value(var, executor) - - -SUCCESS_MARK_FILENAME = "_SUCCESS" -CHECKPOINT_PREFIX = "checkpoint" -CHECKPOINT_SEPARATOR = "_" - - -def save_checkpoint(executor, - checkpoint_dir=None, - max_num_checkpoints=3, - save_interval_secs=600, - main_program=None): - """ - Save Checkpoint will save persistable LodTensor variables from main_program in checkpoint directory, - the directory named by serial number from 0 to (n -1), save_checkpoint use LRU strategy - to keep numbers of checkpoint directory, the numbers of checkpoint directory are max_num_checkpoints at most, - The interval between two saved checkpoints must greater than save_interval_secs. - - :param executor - :param checkpoint_dir - :param max_num_checkpoints - :param save_interval_secs - :param main_program - """ - if checkpoint_dir is None: - checkpoint_dir = os.getcwd() - - if not os.path.isdir(checkpoint_dir): - os.makedirs(checkpoint_dir) - - serial = _get_lastest_checkpoint_dir(checkpoint_dir) - if serial >= 0 and not _interval_secs_exceed( - _get_serial_dir(serial, checkpoint_dir), save_interval_secs): - return - - serial += 1 - cur_dir = _get_serial_dir(serial, checkpoint_dir) - - save_vars( - executor, - dirname=cur_dir, - main_program=main_program, - vars=None, - predicate=_is_checkpoint_var, - filename=None) - _write_success(cur_dir) - _lru_delete(checkpoint_dir, max_num_checkpoints) - - -def load_checkpoint(executor, checkpoint_dir=None, main_program=None): - """ - Load checkpoint from a directory by executor, - it will find the most recent saved checkpoint file and load it auto. - - :param executor - :param checkpoint_dir - :param main_program - """ - - if checkpoint_dir is None: - checkpoint_dir = os.getcwd() - - serial = _get_lastest_checkpoint_dir(checkpoint_dir) - - if serial < 0: - return - - cur_dir = _get_serial_dir(serial, checkpoint_dir) - - load_vars( - executor, - dirname=cur_dir, - main_program=main_program, - predicate=_is_checkpoint_var, - filename=None) - - -def clean_checkpoint(checkpoint_dir, delete_dir=False): - """ - clean the checkpoint dir, when the train exits normally, the trainer will call clean_checkpoint to delete checkpoint directory saved before. - delete_dir only works when the directory is empty, otherwise, OSError is raised. - """ - if checkpoint_dir is None: - checkpoint_dir = os.getcwd() - _lru_delete(checkpoint_dir, max_num_checkpoints=0) - - if delete_dir and not os.listdir(checkpoint_dir): - os.rmdir(checkpoint_dir) - - -def _get_serial_dir(serial, checkpoint_dir): - serial_folder = CHECKPOINT_PREFIX + CHECKPOINT_SEPARATOR + str(serial) - return os.path.join(checkpoint_dir, serial_folder) - - -def _is_checkpoint_var(var): - """ - the checkpoint will not save or load all the variables. - var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded. - - :param var - """ - if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ - var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ - var.desc.type() == core.VarDesc.VarType.RAW: - return False - - if var.name.endswith("@GRAD"): - return False - - return var.persistable - - -def _interval_secs_exceed(dirname, save_interval_secs): - dir_time = os.path.getmtime(dirname) - if save_interval_secs > (time.time() - dir_time): - return False - return True - - -def _lru_delete(dirname, max_num_checkpoints=3): - dirs = os.listdir(dirname) - serials = [] - for serial in dirs: - try: - serials.append(int(serial)) - except ValueError: - continue - - if len(serials) <= max_num_checkpoints: - return - - serials.sort(reverse=True) - serials = serials[max_num_checkpoints:] - for serial in serials: - cur_dir = os.path.join(dirname, str(serial)) - shutil.rmtree(cur_dir) - - -def _write_success(dirname): - """ - write an empty file named "_SUCCESS" in checkpoint dir, indicate this checkpoint is correct. - - :param dirname - """ - success_file = os.path.join(dirname, SUCCESS_MARK_FILENAME) - with open(success_file, 'a') as f: - now = time.ctime() - f.write(now) - - -def _get_lastest_checkpoint_dir(checkpoint_dir): - """ - get the latest file in checkpoint directory, the _SUCCESS file must exist in the directory - - :param checkpoint_dir - """ - if not checkpoint_dir.strip(): - return -1 - - def has_success(checkpoint_dir, cur_dir): - """ - is _SUCCESS in this dir - """ - _, serial = cur_dir.split(CHECKPOINT_SEPARATOR) - - try: - int(serial) - except ValueError: - return -1 - - if not os.path.isdir(os.path.join(checkpoint_dir, cur_dir)): - return -1 - - success_path = os.path.join( - _get_serial_dir(serial, checkpoint_dir), SUCCESS_MARK_FILENAME) - if os.path.isfile(success_path): - return int(serial) - - if not os.path.isdir(checkpoint_dir): - return -1 - - current_dir = -1 - dirs = os.listdir(checkpoint_dir) - for cur_dir in dirs: - success_num = has_success(checkpoint_dir, cur_dir) - if success_num > current_dir: - current_dir = success_num - return current_dir diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index 86efd1ff51cf29485ee28b4d60ffb1439af1aad9..0c2b1eb795860373220eb254612161f7dc816ffd 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -14,12 +14,14 @@ import copy import itertools +import six -from framework import Variable, Parameter, default_main_program, default_startup_program, dtype_is_floating -import unique_name +from .framework import Variable, Parameter, default_main_program, default_startup_program, dtype_is_floating +from . import unique_name from paddle.fluid.initializer import Constant, Xavier -from param_attr import ParamAttr, WeightNormParamAttr -import core +from .param_attr import ParamAttr, WeightNormParamAttr +from . import core +from six.moves import zip class LayerHelper(object): @@ -68,11 +70,11 @@ class LayerHelper(object): @property def param_attr(self): - return ParamAttr.to_attr(self.kwargs.get('param_attr', None)) + return ParamAttr._to_attr(self.kwargs.get('param_attr', None)) @property def bias_attr(self): - return ParamAttr.to_attr(self.kwargs.get('bias_attr', None)) + return ParamAttr._to_attr(self.kwargs.get('bias_attr', None)) def multiple_param_attr(self, length): param_attr = self.param_attr @@ -83,7 +85,7 @@ class LayerHelper(object): raise ValueError("parameter number mismatch") elif len(param_attr) == 1 and length != 1: tmp = [None] * length - for i in xrange(length): + for i in range(length): tmp[i] = copy.deepcopy(param_attr[0]) param_attr = tmp return param_attr @@ -91,7 +93,7 @@ class LayerHelper(object): def iter_inputs_and_params(self, input_param_name='input'): inputs = self.multiple_input(input_param_name) param_attrs = self.multiple_param_attr(len(inputs)) - for ipt, param_attr in itertools.izip(inputs, param_attrs): + for ipt, param_attr in zip(inputs, param_attrs): yield ipt, param_attr def input_dtype(self, input_param_name='input'): @@ -218,7 +220,7 @@ class LayerHelper(object): norm = __norm_op(reshape, dim=0, block=block) __reshape_op(norm, out=out, shape=out_shape, block=block) else: - perm = range(len(x.shape)) + perm = list(range(len(x.shape))) perm[0], perm[dim] = dim, 0 transpose = __transpose_op(x, perm, block=block) norm = __norm_op(transpose, dim=0, block=block) @@ -262,11 +264,11 @@ class LayerHelper(object): g_param = self.startup_program.global_block().create_parameter( dtype=dtype, shape=g_param_shape, - **g_param_attr.to_kwargs(with_initializer=False)) + **g_param_attr._to_kwargs(with_initializer=False)) v_param = self.startup_program.global_block().create_parameter( dtype=dtype, shape=v_param_shape, - **v_param_attr.to_kwargs(with_initializer=True)) + **v_param_attr._to_kwargs(with_initializer=True)) __norm_except_dim( x=v_param, out=g_param, @@ -275,9 +277,9 @@ class LayerHelper(object): # Add weight normalization to main_program g_param = self.main_program.global_block().create_parameter( - dtype=dtype, shape=g_param_shape, **g_param_attr.to_kwargs()) + dtype=dtype, shape=g_param_shape, **g_param_attr._to_kwargs()) v_param = self.main_program.global_block().create_parameter( - dtype=dtype, shape=v_param_shape, **v_param_attr.to_kwargs()) + dtype=dtype, shape=v_param_shape, **v_param_attr._to_kwargs()) w_param = __weight_normalize(g_param, v_param, dim=attr.dim) return w_param @@ -296,11 +298,11 @@ class LayerHelper(object): if default_initializer is None and attr.initializer is None: if is_bias: - attr.set_default_bias_initializer() + attr._set_default_bias_initializer() else: - attr.set_default_param_initializer() + attr._set_default_param_initializer() else: - attr.set_default_initializer(default_initializer) + attr._set_default_initializer(default_initializer) # If weight normalization is set, insert extra parameters and ops. # Refer to https://arxiv.org/pdf/1602.07868.pdf @@ -310,9 +312,9 @@ class LayerHelper(object): return param self.startup_program.global_block().create_parameter( - dtype=dtype, shape=shape, **attr.to_kwargs(with_initializer=True)) + dtype=dtype, shape=shape, **attr._to_kwargs(with_initializer=True)) return self.main_program.global_block().create_parameter( - dtype=dtype, shape=shape, **attr.to_kwargs()) + dtype=dtype, shape=shape, **attr._to_kwargs()) def get_parameter(self, name): param = self.main_program.global_block().var(name) @@ -397,8 +399,10 @@ class LayerHelper(object): act = self.kwargs.get('act', None) if act is None: return input_var - if isinstance(act, basestring): + if isinstance(act, six.string_types): act = {'type': act} + else: + raise TypeError(str(act) + " should be unicode or str") if 'use_cudnn' in self.kwargs and self.kwargs.get('use_cudnn'): act['use_cudnn'] = self.kwargs.get('use_cudnn') diff --git a/python/paddle/fluid/layers/__init__.py b/python/paddle/fluid/layers/__init__.py index a568f61dcb2da976baa7847ae26281a34d6f88dd..a48e360463456ab7e00534dc0684aa153c8205cd 100644 --- a/python/paddle/fluid/layers/__init__.py +++ b/python/paddle/fluid/layers/__init__.py @@ -12,28 +12,27 @@ # See the License for the specific language governing permissions and # limitations under the License. -import ops -from ops import * -import nn -from nn import * -import io -from io import * -import tensor -from tensor import * -import control_flow -from control_flow import * -import device -from device import * -import math_op_patch -from math_op_patch import * -import detection -from detection import * -import metric -from metric import * -from learning_rate_scheduler import * +from . import ops +from .ops import * +from . import nn +from .nn import * +from . import io +from .io import * +from . import tensor +from .tensor import * +from . import control_flow +from .control_flow import * +from . import device +from .device import * +from . import math_op_patch +from .math_op_patch import * +from . import detection +from .detection import * +from . import metric_op +from .metric_op import * +from .learning_rate_scheduler import * __all__ = [] -__all__ += math_op_patch.__all__ __all__ += nn.__all__ __all__ += io.__all__ __all__ += tensor.__all__ @@ -41,5 +40,5 @@ __all__ += control_flow.__all__ __all__ += ops.__all__ __all__ += device.__all__ __all__ += detection.__all__ -__all__ += metric.__all__ +__all__ += metric_op.__all__ __all__ += learning_rate_scheduler.__all__ diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index d1ea9f148566d20988a43f4c9d421c4452697ef1..9fb7b4d0cad67db2d2d4b56e43d8837b8160cdb0 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -13,38 +13,29 @@ # limitations under the License. import contextlib -from layer_function_generator import autodoc -from tensor import assign, fill_constant +from .layer_function_generator import autodoc, templatedoc +from .tensor import assign, fill_constant from .. import core from ..framework import Program, Variable, Operator from ..layer_helper import LayerHelper, unique_name from ..initializer import force_init_on_cpu -from ops import logical_and, logical_not, logical_or +from .ops import logical_and, logical_not, logical_or +import numpy +import warnings +from functools import reduce __all__ = [ - 'split_lod_tensor', - 'merge_lod_tensor', - 'BlockGuard', - 'BlockGuardWithCompletion', - 'StaticRNNMemoryLink', - 'WhileGuard', 'While', 'Switch', - 'lod_rank_table', - 'max_sequence_len', - 'lod_tensor_to_array', - 'array_to_lod_tensor', 'increment', 'array_write', 'create_array', 'less_than', 'equal', 'array_read', - 'shrink_memory', 'array_length', 'IfElse', 'DynamicRNN', - 'ConditionalBlock', 'StaticRNN', 'reorder_lod_tensor_by_rank', 'ParallelDo', @@ -55,34 +46,36 @@ __all__ = [ def split_lod_tensor(input, mask, level=0): """ - **split_lod_tensor** - This function takes in an input that contains the complete lod information, and takes in a mask which is used to mask certain parts of the input. The output is the true branch and the false branch with the mask applied to - the input at a certain level in the tensor. + the input at a certain level in the tensor. Mainly used in IfElse to split + data into two parts. Args: input(tuple|list|None): The input tensor that contains complete lod information needed to construct the output. mask(list): A bool column vector which masks the input. - level(int): The specific lod level to rank. + level(int): The specific lod level to split. Returns: - Variable: The true branch of tensor as per the mask applied to input. - Variable: The false branch of tensor as per the mask applied to input. + tuple(Variable, Variable): + The true branch of tensor as per the mask applied to input. + + The false branch of tensor as per the mask applied to input. Examples: .. code-block:: python - x = layers.data(name='x', shape=[1]) + x = fluid.layers.data(name='x', shape=[1]) x.persistable = True - y = layers.data(name='y', shape=[1]) + y = fluid.layers.data(name='y', shape=[1]) y.persistable = True - out_true, out_false = layers.split_lod_tensor( + out_true, out_false = fluid.layers.split_lod_tensor( input=x, mask=y, level=level) + """ helper = LayerHelper('split_lod_tensor', **locals()) out_true = helper.create_tmp_variable(dtype=input.dtype) @@ -105,8 +98,9 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0): This function takes in an input :math:`x`, the True branch, the False branch and a binary :math:`mask`. Using this information, this function - merges the True and False branches of the tensor into a single Output - at a certain lod level indiacted by :math:`level`. + merges the True and False branches of the tensor into a single tensor as + output at a certain lod level indicated by :math:`level`. Used in IfElse + to merge the output if True block and False Block. Args: in_true(tuple|list|None): The True branch to be merged. @@ -114,7 +108,7 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0): x(tuple|list|None): The input tensor that contains complete lod information needed to construct the output. mask(list): A bool column vector which masks the input. - level(int): The specific lod level to rank. + level(int): The specific lod level to merge. Returns: Variable: The merged output tensor. @@ -182,12 +176,14 @@ def Print(input, Returns: Variable: Output tensor, same data with input tensor. + Examples: + .. code-block:: python - value = some_layer(...) - Print(value, summarize=10, - message="The content of some_layer: ") + value = some_layer(...) + Print(value, summarize=10, + message="The content of some_layer: ") ''' helper = LayerHelper('print', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) @@ -233,12 +229,62 @@ class BlockGuard(object): class ParallelDo(object): """ - ParallelDo class. + ParallelDo is used to represent multi-thread data parallel processing. + + Its vanilla implementation can be shown as the following (:math:`|` means + single thread and :math:`||||` means multiple threads) + + .. code-block:: text + + In the forward pass + | Split input onto different devices + | Copy parameter onto different devices + |||| Compute forward pass in parallel + | Merge output from different devices + + In the backward pass + | Split output@grad onto different devices + |||| Compute backward pass in parallel + | accumulate param@grad from different devices to the first device + | Merge input@grad from different devices + | Copy param@grad to the place of parallel_do_op - ParallelDo class is used to create a ParallelDo. + Examples: + + .. code-block:: python + + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # ParallelDo version & Single-thread version + if thread_num > 1: + places = fluid.layers.get_places(thread_num) + pd = fluid.layers.ParallelDo(places) + with pd.do(): + images = pd.read_input(images) + label = pd.read_input(label) + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + + avg_cost = fluid.layers.mean(x=cost) + pd.write_output(avg_cost) + + avg_cost = pd() + avg_cost = fluid.layers.mean(avg_cost) + else: + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + .. warning:: + + It will be soon deprecated, please use ParallelExecutor instead. """ def __init__(self, places, use_nccl=False, name=None): + warnings.warn( + "API ParallelDo is deprecated since 0.15.0. Please use ParallelExecutor instead.", + Warning) self.helper = LayerHelper("parallel_do", name=name) self.inputs = [] self.places = places @@ -297,7 +343,7 @@ class ParallelDo(object): return [parent_block.var(name) for name in params] - def complete_op(self): + def _complete_op(self): main_program = self.helper.main_program current_block = main_program.current_block() parent_block = self.parent_block() @@ -353,7 +399,7 @@ class BlockGuardWithCompletion(BlockGuard): if exc_type is not None: return False self.rnn.status = StaticRNN.AFTER_RNN_BLOCK - self.rnn.complete_op() + self.rnn._complete_op() return super(BlockGuardWithCompletion, self).__exit__(exc_type, exc_val, exc_tb) @@ -362,16 +408,17 @@ class StaticRNNMemoryLink(object): """ StaticRNNMemoryLink class. - Args: - init: the initial variable for Memory - init: Variable - pre_mem: the memory variable in previous time step - pre_mem: Variable - mem: the memory variable in current time step - mem: Variable - StaticRNNMemoryLink class is used to create a link between two memory cells of a StaticRNN. + + + NOTE: This is a internal data structure of a very low-level API. + Please use StaticRNN instead. + + Args: + init(Variable): the initial variable for Memory. + pre_mem(Variable): the memory variable in previous time step. + mem(Variable): the memory variable in current time step. """ def __init__(self, init, pre_mem, mem=None): @@ -428,7 +475,7 @@ class StaticRNN(object): if shape is None or batch_ref is None: raise ValueError( "if init is None, memory at least need shape and batch_ref") - parent_block = self.parent_block() + parent_block = self._parent_block() var_name = unique_name.generate("@".join( [self.helper.name, "memory_boot"])) boot_var = parent_block.create_var( @@ -485,7 +532,7 @@ class StaticRNN(object): outputs={'Out': tmp_o}, attrs={'dtype': o.dtype}) - out_var = self.parent_block().create_var( + out_var = self._parent_block().create_var( name=tmp_o.name, shape=[self.seq_len] + list(tmp_o.shape), dtype=tmp_o.dtype) @@ -501,7 +548,7 @@ class StaticRNN(object): raise TypeError("update memory should take variables") self.memories[mem.name].mem = var - def parent_block(self): + def _parent_block(self): prog = self.helper.main_program parent_idx = prog.current_block().parent_idx assert parent_idx >= 0 @@ -518,10 +565,10 @@ class StaticRNN(object): else: return self.outputs - def complete_op(self): + def _complete_op(self): main_program = self.helper.main_program rnn_block = main_program.current_block() - parent_block = self.parent_block() + parent_block = self._parent_block() local_inputs = set() @@ -555,7 +602,7 @@ class StaticRNN(object): boot_memories = [] pre_memories = [] memories = [] - for _, mem in self.memories.iteritems(): + for _, mem in list(self.memories.items()): boot_memories.append(mem.init) pre_memories.append(mem.pre_mem.name) mem_var = rnn_block.var(mem.mem.name) @@ -601,11 +648,34 @@ class WhileGuard(BlockGuard): if exc_type is not None: return False self.while_op.status = While.AFTER_WHILE_BLOCK - self.while_op.complete() + self.while_op._complete() return super(WhileGuard, self).__exit__(exc_type, exc_val, exc_tb) class While(object): + """ + while loop control flow. + + Args: + cond (Variable): condition used to compare. + name (str): The name of this layer. + + Examples: + .. code-block:: python + + d0 = layers.data("d0", shape=[10], dtype='float32') + data_array = layers.array_write(x=d0, i=i) + array_len = layers.fill_constant(shape=[1],dtype='int64', value=3) + + cond = layers.less_than(x=i, y=array_len) + while_op = layers.While(cond=cond) + with while_op.block(): + d = layers.array_read(array=data_array, i=i) + i = layers.increment(x=i, in_place=True) + layers.array_write(result, i=i, array=d) + layers.less_than(x=i, y=array_len, cond=cond) + """ + BEFORE_WHILE_BLOCK = 0 IN_WHILE_BLOCK = 1 AFTER_WHILE_BLOCK = 2 @@ -625,7 +695,7 @@ class While(object): def block(self): return WhileGuard(self) - def complete(self): + def _complete(self): main_program = self.helper.main_program while_block = main_program.current_block() parent_block = main_program.block(main_program.current_block() @@ -654,8 +724,10 @@ class While(object): parent_block.append_op( type='while', inputs={ - 'X': - [parent_block.var_recursive(x_name) for x_name in x_name_list], + 'X': [ + parent_block._var_recursive(x_name) + for x_name in x_name_list + ], 'Condition': [self.cond_var] }, outputs={'Out': out_vars, @@ -675,8 +747,8 @@ def lod_rank_table(x, level=0): .. code-block:: text x is a LoDTensor: - x.lod = [[0, 2, 3], - [0, 5, 6, 7]] + x.lod = [[2, 1], + [5, 1, 1]] x.data = [a, b, c, d, e, f, g] 1. set level to 0: @@ -706,7 +778,7 @@ def lod_rank_table(x, level=0): .. code-block:: python x = fluid.layers.data(name='x', shape=[10], - dtype='float32', lod_level=1) + dtype='float32', lod_level=1) out = layers.lod_rank_table(x=x, level=0) """ helper = LayerHelper("lod_rank_table", **locals()) @@ -721,26 +793,22 @@ def lod_rank_table(x, level=0): return table +@templatedoc() def max_sequence_len(rank_table): - """Max Sequence Len Operator. Given a LoDRankTable object, this layer - returns the max length of a batch of sequences. In fact, a LoDRankTable - object contains a list of tuples() and - the list is already sorted by sequence length in descending order, so the - operator just returns the sequence length of the first tuple element. + """ + ${comment} + + >>> import paddle.fluid as fluid + >>> x = fluid.layers.data(name='x', shape=[10], dtype='float32', + >>> lod_level=1) + >>> rank_table = layers.lod_rank_table(x=x, level=0) + >>> max_seq_len = layers.max_sequence_len(rank_table) Args: - rank_table (Variable): Input variable which is a LoDRankTable object. + rank_table(${rank_table_type}): ${rank_table_comment}. Returns: - Variable: The max length of sequence. - - Examples: - .. code-block:: python - - x = fluid.layers.data(name='x', shape=[10], - dtype='float32', lod_level=1) - rank_table = layers.lod_rank_table(x=x, level=0) - max_seq_len = layers.max_sequence_len(rank_table) + ${out_comment}. """ helper = LayerHelper("max_seqence_len", **locals()) res = helper.create_tmp_variable(dtype="int64") @@ -752,17 +820,25 @@ def max_sequence_len(rank_table): def lod_tensor_to_array(x, table): - """ Convert a LOD_TENSOR to an LOD_TENSOR_ARRAY. + """ + Convert a LoDTensor to a LoDTensorArray. + + This function split a LoDTesnor to a LoDTensorArray according to its LoD + information. LoDTensorArray is an alias of C++ std::vector in + PaddlePaddle. The generated LoDTensorArray of this function can be further read + or written by `read_from_array()` and `write_to_array()` operators. However, + this function is generally an internal component of PaddlePaddle `DynamicRNN`. + Users should not use it directly. Args: - x (Variable|list): The LOD tensor to be converted to a LOD tensor array. + x (Variable|list): The LoDTensor to be converted to a LoDTensorArray. table (ParamAttr|list): The variable that stores the level of lod which is ordered by sequence length in - descending order. + descending order. It is generally generated + by `layers.lod_rank_table()` API. Returns: - Variable: The variable of type array that has been converted from a - tensor. + Variable: The LoDTensorArray that has been converted from the input tensor. Examples: .. code-block:: python @@ -827,8 +903,7 @@ def increment(x, value=1.0, in_place=True): in_place (bool): If the increment should be performed in-place. Returns: - Variable: The tensor variable storing the transformation of - element-wise increment of each value in the input. + Variable: The elementwise-incremented object. Examples: .. code-block:: python @@ -870,7 +945,7 @@ def array_write(x, i, array=None): Variable: The output LOD_TENSOR_ARRAY where the input tensor is written. Examples: - .. code-block::python + .. code-block:: python tmp = fluid.layers.zeros(shape=[10], dtype='int32') i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) @@ -891,14 +966,17 @@ def array_write(x, i, array=None): def create_array(dtype): - """This function creates an array of type :math:`LOD_TENSOR_ARRAY` using the - LayerHelper. + """ + **Create LoDTensorArray** + + This function creates an array of LOD_TENSOR_ARRAY . It is mainly used to + implement RNN with array_write, array_read and While. Args: - dtype (int|float): The data type of the elements in the array. + dtype (int|float): The data type of the elements in the lod_tensor_array. Returns: - Variable: The tensor variable storing the elements of data type. + Variable: The lod_tensor_array variable storing the elements of data type. Examples: .. code-block:: python @@ -913,37 +991,40 @@ def create_array(dtype): dtype=dtype) -def less_than(x, y, force_cpu=True, cond=None, **ignored): +@templatedoc() +def less_than(x, y, force_cpu=None, cond=None, **ignored): """ - **Less than** + ${comment} - This layer returns the truth value of :math:`x < y` elementwise. + >>> import paddle.fluid as fluid + >>> less = fluid.layers.less_than(x=label, y=limit) Args: - x(Variable): First operand of *less_than* - y(Variable): Second operand of *less_than* - force_cpu(Bool|True): The output data will be on CPU if set true. + x(${x_type}): ${x_comment}. + y(${y_type}): ${y_comment}. + force_cpu(${force_cpu_type}): ${force_cpu_comment}. cond(Variable|None): Optional output variable to store the result of *less_than* Returns: - Variable: The tensor variable storing the output of *less_than*. - - Examples: - .. code-block:: python - - less = fluid.layers.less_than(x=label, y=limit) + ${out_comment}. """ helper = LayerHelper("less_than", **locals()) if cond is None: cond = helper.create_tmp_variable(dtype='bool') cond.stop_gradient = True + attrs = dict() + if force_cpu is not None: + attrs['force_cpu'] = force_cpu + elif force_init_on_cpu(): + attrs['force_cpu'] = force_init_on_cpu() + helper.append_op( type='less_than', inputs={'X': [x], 'Y': [y]}, outputs={'Out': [cond]}, - attrs={'force_cpu': force_cpu or force_init_on_cpu()}) + attrs=attrs) return cond @@ -978,16 +1059,34 @@ def equal(x, y, cond=None, **ignored): def array_read(array, i): - """This function performs the operation to read the data in as an + """ + This function performs the operation to read the data in as an LOD_TENSOR_ARRAY. + + .. code-block:: text + + Given: + + array = [0.6, 0.1, 0.3, 0.1] + + And: + + i = 2 + + Then: + + output = 0.3 + Args: - array (Variable|list): The input tensor that will be written to an array. - i (Variable|list): The subscript index in tensor array, that points the - place where data will be written to. + array (Variable|list): The input tensor that store data to be read. + i (Variable|list): The index of the data to be read from input array. + Returns: Variable: The tensor type variable that has the data written to it. + Examples: - .. code-block::python + .. code-block:: python + tmp = fluid.layers.zeros(shape=[10], dtype='int32') i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) arr = layers.array_read(tmp, i=i) @@ -1008,8 +1107,28 @@ def array_read(array, i): def shrink_memory(x, i, table): """ - This function creates an operator to shrink_rnn_memory using the RankTable + This function creates an operator to shrink rnn memory using the RankTable as mentioned in the input parameter. + + NOTE: This API is very low-level API. It is used by DynamicRNN only. + + Since the Dynamic RNN uses no-padding way to implement RNN. The sequence + will be sorted by order, and the length of valid memory will be shrink after + each time step. + + Args: + x(Variable): The memory object in the previous time step. + i(Variable): The step count variable. A int scalar as LoDTensor. + table(Variable): The RNNRankTable object. + + Returns: + the memory variable after shrink. + + Examples: + + Since this API is very low level API. The example is not provided. + Please reference the implementation of class DynamicRNN for detail + usage. """ helper = LayerHelper('shrink_memory', **locals()) out = helper.create_tmp_variable(dtype=x.dtype) @@ -1024,9 +1143,14 @@ def shrink_memory(x, i, table): def array_length(array): - """This function performs the operation to find the length of the input + """ + **Get the Length of Input LoDTensorArray** + + This function performs the operation to find the length of the input LOD_TENSOR_ARRAY. + Related API: array_read, array_write, While. + Args: array (LOD_TENSOR_ARRAY): The input array that will be used to compute the length. @@ -1035,12 +1159,13 @@ def array_length(array): Variable: The length of the input LoDTensorArray. Examples: - .. code-block::python + .. code-block:: python tmp = fluid.layers.zeros(shape=[10], dtype='int32') i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) arr = fluid.layers.array_write(tmp, i=i) arr_len = fluid.layers.array_length(arr) + """ helper = LayerHelper('array_length', **locals()) tmp = helper.create_tmp_variable(dtype='int64') @@ -1051,6 +1176,13 @@ def array_length(array): class ConditionalBlockGuard(BlockGuard): + """ + ConditionalBlockGuard is derived from BlockGuard. It is dedicated for + holding a ConditionalBlock, and helping users entering and exiting the + ConditionalBlock via Python's 'with' keyword. However, ConditionalBlockGuard + is generally an internal component of IfElse, users should not use it directly. + """ + def __init__(self, block): if not isinstance(block, ConditionalBlock): raise TypeError("block should be conditional block") @@ -1067,6 +1199,31 @@ class ConditionalBlockGuard(BlockGuard): class ConditionalBlock(object): + ''' + **ConditionalBlock** + + ConditionalBlock is an operator that bind a block to a specific condition, + if the condition matches, the corresponding block will be executed. + + Args: + inputs (Variable): bool conditions. + is_scalar_condition (bool): whether the branch is controled by a scalar. + name(str): name of this ConditionalBlock. + + Examples: + .. code-block:: python + + cond = layers.less_than(x=label, y=limit) + true_image, false_image = layers.split_lod_tensor( + input=image, mask=cond) + true_cond = layers.ConditionalBlock([true_image]) + + with true_cond.block(): + ... + with false_cond.block(): + ... + ''' + def __init__(self, inputs, is_scalar_condition=False, name=None): for each_input in inputs: if not isinstance(each_input, Variable): @@ -1098,7 +1255,7 @@ class ConditionalBlock(object): input_set = set([ipt.name for ipt in self.inputs]) param_list = [ - parent_block.var_recursive(each_name) for each_name in params + parent_block._var_recursive(each_name) for each_name in params if each_name not in input_set ] @@ -1124,6 +1281,42 @@ class ConditionalBlock(object): class Switch(object): + """ + Switch class works just like a `if-elif-else`. Can be used in learning rate scheduler + to modify learning rate + + The Semantics: + + 1. A `switch` control-flow checks cases one-by-one. + + 2. The condition of each case is a boolean value, which is a scalar Variable. + + 3. It runs the first matched case, or the default case if there is one. + + 4. Once it matches a case, it runs the corresponding branch and only that branch. + + Examples: + .. code-block:: python + + lr = fluid.layers.tensor.create_global_var( + shape=[1], + value=0.0, + dtype='float32', + persistable=True, + name="learning_rate") + one_var = tensor.fill_constant( + shape=[1], dtype='float32', value=1.0) + two_var = tensor.fill_constant( + shape=[1], dtype='float32', value=2.0) + + with fluid.layers.control_flow.Switch() as switch: + with switch.case(global_step == zero_var): + fluid.layers.tensor.assign(input=one_var, output=lr) + with switch.default(): + fluid.layers.tensor.assign(input=two_var, output=lr) + + """ + def __init__(self, name=None): self.helper = LayerHelper('switch', name=name) self.inside_scope = False @@ -1153,7 +1346,8 @@ class Switch(object): return ConditionalBlockGuard(cond_block) def default(self): - """create a default case for this switch + """ + create a default case for this switch """ pre_cond_num = len(self.pre_not_conditions) if pre_cond_num == 0: @@ -1213,6 +1407,34 @@ class IfElseBlockGuard(object): class IfElse(object): + """ + if-else control flow. + + Args: + cond (Variable): condition used to compare. + name (str, default None): The name of this layer. + + Examples: + .. code-block:: python + + limit = fluid.layers.fill_constant_batch_size_like( + input=label, dtype='int64', shape=[1], value=5.0) + cond = fluid.layers.less_than(x=label, y=limit) + ie = fluid.layers.IfElse(cond) + with ie.true_block(): + true_image = ie.input(image) + hidden = fluid.layers.fc(input=true_image, size=100, act='tanh') + prob = fluid.layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + + with ie.false_block(): + false_image = ie.input(image) + hidden = fluid.layers.fc( + input=false_image, size=200, act='tanh') + prob = fluid.layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + prob = ie() + """ OUT_IF_ELSE_BLOCKS = 0 IN_IF_ELSE_TRUE_BLOCKS = 1 IN_IF_ELSE_FALSE_BLOCKS = 2 @@ -1232,7 +1454,7 @@ class IfElse(object): if self.status == IfElse.OUT_IF_ELSE_BLOCKS: raise ValueError("input must in true/false blocks") if id(x) not in self.input_table: - parent_block = self.parent_block() + parent_block = self._parent_block() out_true = parent_block.create_var( name=unique_name.generate('ifelse_input' + self.helper.name), dtype=x.dtype) @@ -1258,7 +1480,7 @@ class IfElse(object): else: return out_false - def parent_block(self): + def _parent_block(self): current_block = self.helper.main_program.current_block() return self.helper.main_program.block(current_block.parent_idx) @@ -1274,7 +1496,7 @@ class IfElse(object): out_table = self.output_table[1 if self.status == self.IN_IF_ELSE_TRUE_BLOCKS else 0] - parent_block = self.parent_block() + parent_block = self._parent_block() for each_out in outs: if not isinstance(each_out, Variable): raise TypeError("Each output should be a variable") @@ -1291,7 +1513,7 @@ class IfElse(object): def __call__(self): if self.status != self.OUT_IF_ELSE_BLOCKS: raise ValueError("IfElse::__call__ must be out of sub-block") - false_len, true_len = map(len, self.output_table) + false_len, true_len = list(map(len, self.output_table)) if false_len == 0 and true_len == 0: raise ValueError("Must invoke true_block/false_block before " "__call__") @@ -1315,6 +1537,38 @@ class IfElse(object): class DynamicRNN(object): + """ + The dynamic RNN can process a batch of sequence data. The length of each + sample sequence can be different. This API automatically process them in + batch. + + The input lod must be set. Please reference `lod_tensor` + + >>> import paddle.fluid as fluid + >>> data = fluid.layers.data(name='sentence', dtype='int64', lod_level=1) + >>> embedding = fluid.layers.embedding(input=data, size=[65535, 32], + >>> is_sparse=True) + >>> + >>> drnn = fluid.layers.DynamicRNN() + >>> with drnn.block(): + >>> word = drnn.step_input(embedding) + >>> prev = drnn.memory(shape=[200]) + >>> hidden = fluid.layers.fc(input=[word, prev], size=200, act='relu') + >>> drnn.update_memory(prev, hidden) # set prev to hidden + >>> drnn.output(hidden) + >>> + >>> # last is the last time step of rnn. It is the encoding result. + >>> last = fluid.layers.sequence_last_step(drnn()) + + The dynamic RNN will unfold sequence into timesteps. Users need to define + how to process each time step during the :code:`with` block. + + The `memory` is used staging data cross time step. The initial value of + memory can be zero or another variable. + + The dynamic RNN can mark multiple variables as its output. Use `drnn()` to + get the output sequence. + """ BEFORE_RNN = 0 IN_RNN = 1 AFTER_RNN = 2 @@ -1337,6 +1591,15 @@ class DynamicRNN(object): self.mem_link = [] def step_input(self, x): + """ + Mark a sequence as a dynamic RNN input. + Args: + x(Variable): The input sequence. + + Returns: + The current timestep in the input sequence. + + """ self._assert_in_rnn_block_("step_input") if not isinstance(x, Variable): raise TypeError( @@ -1380,6 +1643,15 @@ class DynamicRNN(object): return array_read(array=input_array, i=self.step_idx) def static_input(self, x): + """ + Mark a variable as a RNN input. The input will not be scattered into + time steps. + Args: + x(Variable): The input variable. + + Returns: + The input variable that can access in RNN. + """ self._assert_in_rnn_block_("static_input") if not isinstance(x, Variable): raise TypeError( @@ -1401,6 +1673,10 @@ class DynamicRNN(object): @contextlib.contextmanager def block(self): + """ + The block for user to define operators in RNN. See the class docstring + for more details. + """ if self.status != DynamicRNN.BEFORE_RNN: raise ValueError("rnn.block() can only be invoke once") self.step_idx = fill_constant( @@ -1427,6 +1703,9 @@ class DynamicRNN(object): x=each_array, table=self.lod_rank_table)) def __call__(self, *args, **kwargs): + """ + Get the output of RNN. This API should only be invoked after RNN.block() + """ if self.status != DynamicRNN.AFTER_RNN: raise ValueError(("Output of the dynamic RNN can only be visited " "outside the rnn block.")) @@ -1441,6 +1720,70 @@ class DynamicRNN(object): value=0.0, need_reorder=False, dtype='float32'): + """ + Create a memory variable for dynamic rnn. + + If the :code:`init` is not None, :code:`memory` will be initialized by + this variable. The :code:`need_reorder` is used to reorder the memory as + the input variable. It should be set to true when the initialized memory + depends on the input sample. + + For example, + + >>> import paddle.fluid as fluid + >>> sentence = fluid.layers.data( + >>> name='sentence', dtype='float32', shape=[32]) + >>> boot_memory = fluid.layers.data( + >>> name='boot', dtype='float32', shape=[10]) + >>> + >>> drnn = fluid.layers.DynamicRNN() + >>> with drnn.block(): + >>> word = drnn.step_input(sentence) + >>> memory = drnn.memory(init=boot_memory, need_reorder=True) + >>> hidden = fluid.layers.fc( + >>> input=[word, memory], size=10, act='tanh') + >>> drnn.update_memory(ex_mem=memory, new_mem=hidden) + >>> drnn.output(hidden) + >>> rnn_output = drnn() + + + Otherwise, if :code:`shape`, :code:`value`, :code:`dtype` are set, the + :code:`memory` will be initialized by this :code:`value`. + + For example, + + >>> import paddle.fluid as fluid + >>> sentence = fluid.layers.data( + >>> name='sentence', dtype='float32', shape=[32]) + >>> + >>> drnn = fluid.layers.DynamicRNN() + >>> with drnn.block(): + >>> word = drnn.step_input(sentence) + >>> memory = drnn.memory(shape=[10], dtype='float32', value=0) + >>> hidden = fluid.layers.fc( + >>> input=[word, memory], size=10, act='tanh') + >>> drnn.update_memory(ex_mem=memory, new_mem=hidden) + >>> drnn.output(hidden) + >>> rnn_output = drnn() + + + Args: + init(Variable|None): The initialized variable. + + shape(list|tuple): The memory shape. NOTE the shape does not contain + batch_size. + + value(float): the initalized value. + + need_reorder(bool): True if the initialized memory depends on the + input sample. + + dtype(str|numpy.dtype): The data type of the initialized memory. + + Returns: + the memory variable. + + """ self._assert_in_rnn_block_('memory') if init is not None: if not isinstance(init, Variable): @@ -1508,6 +1851,16 @@ class DynamicRNN(object): return self.memory(init=init) def update_memory(self, ex_mem, new_mem): + """ + Update the memory from ex_mem to new_mem. NOTE that the shape and data + type of :code:`ex_mem` and :code:`new_mem` must be same. + Args: + ex_mem(Variable): the memory variable. + new_mem(Variable): the plain variable generated in RNN block. + + Returns: + None + """ self._assert_in_rnn_block_('update_memory') if not isinstance(ex_mem, Variable): raise TypeError("The input arg `ex_mem` of update_memory() must " @@ -1525,6 +1878,15 @@ class DynamicRNN(object): self.mem_link.append((new_mem, mem_array)) def output(self, *outputs): + """ + mark the RNN output variables. + + Args: + outputs: The output variables. + + Returns: + None + """ self._assert_in_rnn_block_('output') parent_block = self._parent_block_() for each in outputs: @@ -1567,26 +1929,26 @@ def reorder_lod_tensor_by_rank(x, rank_table): def is_empty(x, cond=None, **ignored): """ - **Is Empty** - - This layer returns the truth value of whether the variable is empty. + Test whether a Variable is empty. Args: - x(Variable): Operand of *is_empty* - cond(Variable|None): Optional output variable to store the result - of *is_empty* + x (Variable): The Variable to be tested. + cond (Variable|None): Output parameter. Returns the test result + of given 'x'. Default: None Returns: - Variable: The tensor variable storing the output of *is_empty*. + Variable: A bool scalar. True if 'x' is an empty Variable. Raises: TypeError: If input cond is not a variable, or cond's dtype is - not bool + not bool. Examples: .. code-block:: python - less = fluid.layers.is_empty(x=input) + res = fluid.layers.is_empty(x=input) + # or: + fluid.layers.is_empty(x=input, cond=res) """ helper = LayerHelper("is_empty", **locals()) if cond is None: diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index 3a83db12fd13651578deeac6b562bac2f1e4e4b6..b996c8368862184f9bc8b177f3b6e43aebdfb007 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -15,12 +15,15 @@ All layers just related to the detection neural network. """ -from layer_function_generator import generate_layer_fn -from layer_function_generator import autodoc +from .layer_function_generator import generate_layer_fn +from .layer_function_generator import autodoc, templatedoc from ..layer_helper import LayerHelper -import tensor -import nn +from . import tensor +from . import nn +from . import ops import math +import numpy +from functools import reduce __all__ = [ 'prior_box', @@ -30,11 +33,14 @@ __all__ = [ 'detection_output', 'ssd_loss', 'detection_map', + 'rpn_target_assign', + 'anchor_generator', ] __auto__ = [ 'iou_similarity', 'box_coder', + 'polygon_box_transform', ] __all__ += __auto__ @@ -43,6 +49,135 @@ for _OP in set(__auto__): globals()[_OP] = generate_layer_fn(_OP) +def rpn_target_assign(loc, + scores, + anchor_box, + gt_box, + rpn_batch_size_per_im=256, + fg_fraction=0.25, + rpn_positive_overlap=0.7, + rpn_negative_overlap=0.3): + """ + ** Target Assign Layer for region proposal network (RPN) in Faster-RCNN detection. ** + + This layer can be, for given the Intersection-over-Union (IoU) overlap + between anchors and ground truth boxes, to assign classification and + regression targets to each each anchor, these target labels are used for + train RPN. The classification targets is a binary class label (of being + an object or not). Following the paper of Faster-RCNN, the positive labels + are two kinds of anchors: (i) the anchor/anchors with the highest IoU + overlap with a ground-truth box, or (ii) an anchor that has an IoU overlap + higher than rpn_positive_overlap(0.7) with any ground-truth box. Note + that a single ground-truth box may assign positive labels to multiple + anchors. A non-positive anchor is when its IoU ratio is lower than + rpn_negative_overlap (0.3) for all ground-truth boxes. Anchors that are + neither positive nor negative do not contribute to the training objective. + The regression targets are the encoded ground-truth boxes associated with + the positive anchors. + + Args: + loc(Variable): A 3-D Tensor with shape [N, M, 4] represents the + predicted locations of M bounding bboxes. N is the batch size, + and each bounding box has four coordinate values and the layout + is [xmin, ymin, xmax, ymax]. + scores(Variable): A 3-D Tensor with shape [N, M, C] represents the + predicted confidence predictions. N is the batch size, C is the + class number, M is number of bounding boxes. For each category + there are total M scores which corresponding M bounding boxes. + anchor_box(Variable): A 2-D Tensor with shape [M, 4] holds M boxes, + each box is represented as [xmin, ymin, xmax, ymax], + [xmin, ymin] is the left top coordinate of the anchor box, + if the input is image feature map, they are close to the origin + of the coordinate system. [xmax, ymax] is the right bottom + coordinate of the anchor box. + gt_box (Variable): The ground-truth boudding boxes (bboxes) are a 2D + LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth + bboxes of mini-batch input. + rpn_batch_size_per_im(int): Total number of RPN examples per image. + fg_fraction(float): Target fraction of RoI minibatch that is labeled + foreground (i.e. class > 0), 0-th class is background. + rpn_positive_overlap(float): Minimum overlap required between an anchor + and ground-truth box for the (anchor, gt box) pair to be a positive + example. + rpn_negative_overlap(float): Maximum overlap allowed between an anchor + and ground-truth box for the (anchor, gt box) pair to be a negative + examples. + + Returns: + tuple: + A tuple(predicted_scores, predicted_location, target_label, + target_bbox) is returned. The predicted_scores and + predicted_location is the predicted result of the RPN. + The target_label and target_bbox is the ground truth, + respectively. The predicted_location is a 2D Tensor with shape + [F, 4], and the shape of target_bbox is same as the shape of + the predicted_location, F is the number of the foreground + anchors. The predicted_scores is a 2D Tensor with shape + [F + B, 1], and the shape of target_label is same as the shape + of the predicted_scores, B is the number of the background + anchors, the F and B is depends on the input of this operator. + + Examples: + .. code-block:: python + + loc = layers.data(name='location', shape=[2, 80], + append_batch_size=False, dtype='float32') + scores = layers.data(name='scores', shape=[2, 40], + append_batch_size=False, dtype='float32') + anchor_box = layers.data(name='anchor_box', shape=[20, 4], + append_batch_size=False, dtype='float32') + gt_box = layers.data(name='gt_box', shape=[10, 4], + append_batch_size=False, dtype='float32') + loc_pred, score_pred, loc_target, score_target = + fluid.layers.detection_output(loc=location, + scores=scores, + anchor_box=anchor_box, + gt_box=gt_box) + """ + + helper = LayerHelper('rpn_target_assign', **locals()) + # 1. Compute the regression target bboxes + target_bbox = box_coder( + prior_box=anchor_box, + target_box=gt_box, + code_type='encode_center_size', + box_normalized=False) + + # 2. Compute overlaps between the prior boxes and the gt boxes overlaps + iou = iou_similarity(x=gt_box, y=anchor_box) + + # 3. Assign target label to anchors + loc_index = helper.create_tmp_variable(dtype=anchor_box.dtype) + score_index = helper.create_tmp_variable(dtype=anchor_box.dtype) + target_label = helper.create_tmp_variable(dtype=anchor_box.dtype) + helper.append_op( + type="rpn_target_assign", + inputs={'Overlap': iou, }, + outputs={ + 'LocationIndex': loc_index, + 'ScoreIndex': score_index, + 'TargetLabel': target_label, + }, + attrs={ + 'rpn_batch_size_per_im': rpn_batch_size_per_im, + 'rpn_positive_overlap': rpn_positive_overlap, + 'rpn_negative_overlap': rpn_negative_overlap, + 'fg_fraction': fg_fraction, + }) + + # 4. Reshape and gather the target entry + scores = nn.reshape(x=scores, shape=(-1, 2)) + loc = nn.reshape(x=loc, shape=(-1, 4)) + target_label = nn.reshape(x=target_label, shape=(-1, 1)) + target_bbox = nn.reshape(x=target_bbox, shape=(-1, 4)) + + predicted_scores = nn.gather(scores, score_index) + predicted_location = nn.gather(loc, loc_index) + target_label = nn.gather(target_label, score_index) + target_bbox = nn.gather(target_bbox, loc_index) + return predicted_scores, predicted_loc, target_label, target_bbox + + def detection_output(loc, scores, prior_box, @@ -97,7 +232,9 @@ def detection_output(loc, nms_eta(float): The parameter for adaptive NMS. Returns: - Variable: The detection outputs is a LoDTensor with shape [No, 6]. + Variable: + + The detection outputs is a LoDTensor with shape [No, 6]. Each row has six values: [label, confidence, xmin, ymin, xmax, ymax]. `No` is the total number of detections in this mini-batch. For each instance, the offsets in first dimension are called LoD, the offset @@ -110,15 +247,15 @@ def detection_output(loc, Examples: .. code-block:: python - pb = layers.data(name='prior_box', shape=[10, 4], + pb = layers.data(name='prior_box', shape=[10, 4], append_batch_size=False, dtype='float32') - pbv = layers.data(name='prior_box_var', shape=[10, 4], + pbv = layers.data(name='prior_box_var', shape=[10, 4], append_batch_size=False, dtype='float32') - loc = layers.data(name='target_box', shape=[2, 21, 4], + loc = layers.data(name='target_box', shape=[2, 21, 4], append_batch_size=False, dtype='float32') - scores = layers.data(name='scores', shape=[2, 21, 10], + scores = layers.data(name='scores', shape=[2, 21, 10], append_batch_size=False, dtype='float32') - nmsed_outs = fluid.layers.detection_output(scores=scores, + nmsed_outs = fluid.layers.detection_output(scores=scores, loc=loc, prior_box=pb, prior_box_var=pbv) @@ -129,10 +266,11 @@ def detection_output(loc, prior_box_var=prior_box_var, target_box=loc, code_type='decode_center_size') - old_shape = scores.shape - scores = nn.reshape(x=scores, shape=(-1, old_shape[-1])) + compile_shape = scores.shape + run_shape = ops.shape(scores) + scores = nn.flatten(x=scores, axis=2) scores = nn.softmax(input=scores) - scores = nn.reshape(x=scores, shape=old_shape) + scores = nn.reshape(x=scores, shape=compile_shape, actual_shape=run_shape) scores = nn.transpose(scores, perm=[0, 2, 1]) scores.stop_gradient = True nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype) @@ -153,7 +291,7 @@ def detection_output(loc, return nmsed_outs -@autodoc() +@templatedoc() def detection_map(detect_res, label, class_num, @@ -164,6 +302,47 @@ def detection_map(detect_res, input_states=None, out_states=None, ap_version='integral'): + """ + ${comment} + + Args: + detect_res: ${detect_res_comment} + label: ${label_comment} + class_num: ${class_num_comment} + background_label: ${background_label_comment} + overlap_threshold: ${overlap_threshold_comment} + evaluate_difficult: ${evaluate_difficult_comment} + has_state: ${has_state_comment} + input_states: If not None, It contains 3 elements: + 1. pos_count ${pos_count_comment}. + 2. true_pos ${true_pos_comment}. + 3. false_pos ${false_pos_comment}. + out_states: If not None, it contains 3 elements. + 1. accum_pos_count ${accum_pos_count_comment}. + 2. accum_true_pos ${accum_true_pos_comment}. + 3. accum_false_pos ${accum_false_pos_comment}. + ap_version: ${ap_type_comment} + + Returns: + ${map_comment} + + + Examples: + .. code-block:: python + + detect_res = fluid.layers.data( + name='detect_res', + shape=[10, 6], + append_batch_size=False, + dtype='float32') + label = fluid.layers.data( + name='label', + shape=[10, 6], + append_batch_size=False, + dtype='float32') + + map_out = fluid.layers.detection_map(detect_res, label, 21) + """ helper = LayerHelper("detection_map", **locals()) def __create_var(type): @@ -210,53 +389,68 @@ def bipartite_match(dist_matrix, dist_threshold=None, name=None): """ - **Bipartite matchint operator** - - This operator is a greedy bipartite matching algorithm, which is used to - obtain the matching with the maximum distance based on the input + This operator implements a greedy bipartite matching algorithm, which is + used to obtain the matching with the maximum distance based on the input distance matrix. For input 2D matrix, the bipartite matching algorithm can - find the matched column for each row, also can find the matched row for - each column. And this operator only calculate matched indices from column - to row. For each instance, the number of matched indices is the number of - of columns of the input ditance matrix. - - There are two outputs to save matched indices and distance. - A simple description, this algothrim matched the best (maximum distance) + find the matched column for each row (matched means the largest distance), + also can find the matched row for each column. And this operator only + calculate matched indices from column to row. For each instance, + the number of matched indices is the column number of the input distance + matrix. + + There are two outputs, matched indices and distance. + A simple description, this algorithm matched the best (maximum distance) row entity to the column entity and the matched indices are not duplicated in each row of ColToRowMatchIndices. If the column entity is not matched any row entity, set -1 in ColToRowMatchIndices. - Please note that the input DistMat can be LoDTensor (with LoD) or Tensor. + NOTE: the input DistMat can be LoDTensor (with LoD) or Tensor. If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size. If Tensor, the height of ColToRowMatchIndices is 1. + NOTE: This API is a very low level API. It is used by :code:`ssd_loss` + layer. Please consider to use :code:`ssd_loss` instead. + Args: dist_matrix(Variable): This input is a 2-D LoDTensor with shape [K, M]. It is pair-wise distance matrix between the entities represented by each row and each column. For example, assumed one entity is A with shape [K], another entity is B with shape [M]. The - dist_matirx[i][j] is the distance between A[i] and B[j]. The bigger - the distance is, the better macthing the pairs are. Please note, - This tensor can contain LoD information to represent a batch of - inputs. One instance of this batch can contain different numbers of - entities. + dist_matrix[i][j] is the distance between A[i] and B[j]. The bigger + the distance is, the better matching the pairs are. + + NOTE: This tensor can contain LoD information to represent a batch + of inputs. One instance of this batch can contain different numbers + of entities. match_type(string|None): The type of matching method, should be - 'bipartite' or 'per_prediction', 'bipartite' by defalut. + 'bipartite' or 'per_prediction'. [default 'bipartite']. dist_threshold(float|None): If `match_type` is 'per_prediction', this threshold is to determine the extra matching bboxes based - on the maximum distance, 0.5 by defalut. + on the maximum distance, 0.5 by default. Returns: - match_indices(Variable): A 2-D Tensor with shape [N, M] in int type. - N is the batch size. If match_indices[i][j] is -1, it - means B[j] does not match any entity in i-th instance. - Otherwise, it means B[j] is matched to row - match_indices[i][j] in i-th instance. The row number of - i-th instance is saved in match_indices[i][j]. - match_distance(Variable): A 2-D Tensor with shape [N, M] in float type. - N is batch size. If match_indices[i][j] is -1, - match_distance[i][j] is also -1.0. Otherwise, assumed - match_distance[i][j] = d, and the row offsets of each instance - are called LoD. Then match_distance[i][j] = dist_matrix[d+LoD[i]][j]. + tuple: a tuple with two elements is returned. The first is + matched_indices, the second is matched_distance. + + The matched_indices is a 2-D Tensor with shape [N, M] in int type. + N is the batch size. If match_indices[i][j] is -1, it + means B[j] does not match any entity in i-th instance. + Otherwise, it means B[j] is matched to row + match_indices[i][j] in i-th instance. The row number of + i-th instance is saved in match_indices[i][j]. + + The matched_distance is a 2-D Tensor with shape [N, M] in float type + . N is batch size. If match_indices[i][j] is -1, + match_distance[i][j] is also -1.0. Otherwise, assumed + match_distance[i][j] = d, and the row offsets of each instance + are called LoD. Then match_distance[i][j] = + dist_matrix[d+LoD[i]][j]. + + Examples: + + >>> x = fluid.layers.data(name='x', shape=[4], dtype='float32') + >>> y = fluid.layers.data(name='y', shape=[4], dtype='float32') + >>> iou = fluid.layers.iou_similarity(x=x, y=y) + >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou) """ helper = LayerHelper('bipartite_match', **locals()) match_indices = helper.create_tmp_variable(dtype='int32') @@ -281,8 +475,6 @@ def target_assign(input, mismatch_value=None, name=None): """ - **Target assigner operator** - This operator can be, for given the target bounding boxes or labels, to assign classification and regression targets to each prediction as well as weights to prediction. The weights is used to specify which prediction would @@ -296,20 +488,24 @@ def target_assign(input, 1. Assigning all outpts based on `match_indices`: - If id = match_indices[i][j] > 0, + .. code-block:: text + + If id = match_indices[i][j] > 0, - out[i][j][0 : K] = X[lod[i] + id][j % P][0 : K] - out_weight[i][j] = 1. + out[i][j][0 : K] = X[lod[i] + id][j % P][0 : K] + out_weight[i][j] = 1. - Otherwise, + Otherwise, - out[j][j][0 : K] = {mismatch_value, mismatch_value, ...} - out_weight[i][j] = 0. + out[j][j][0 : K] = {mismatch_value, mismatch_value, ...} + out_weight[i][j] = 0. 2. Assigning out_weight based on `neg_indices` if `neg_indices` is provided: Assumed that the row offset for each instance in `neg_indices` is called neg_lod, for i-th instance and each `id` of neg_indices in this instance: + + .. code-block:: text out[i][id][0 : K] = {mismatch_value, mismatch_value, ...} out_weight[i][id] = 1.0 @@ -326,10 +522,22 @@ def target_assign(input, mismatch_value (float32): Fill this value to the mismatched location. Returns: - out (Variable): The output is a 3D Tensor with shape [N, P, K], - N and P is the same as they are in `neg_indices`, K is the - same as it in input of X. If `match_indices[i][j]`. - out_weight (Variable): The weight for output with the shape of [N, P, 1]. + tuple: + A tuple(out, out_weight) is returned. out is a 3D Tensor with + shape [N, P, K], N and P is the same as they are in + `neg_indices`, K is the same as it in input of X. If + `match_indices[i][j]`. out_weight is the weight for output with + the shape of [N, P, 1]. + + Examples: + + .. code-block:: python + + matched_indices, matched_dist = fluid.layers.bipartite_match(iou) + gt = layers.data( + name='gt', shape=[1, 1], dtype='int32', lod_level=1) + trg, trg_weight = layers.target_assign( + gt, matched_indices, mismatch_value=0) """ helper = LayerHelper('target_assign', **locals()) out = helper.create_tmp_variable(dtype=input.dtype) @@ -364,7 +572,7 @@ def ssd_loss(location, normalize=True, sample_size=None): """ - **Multi-box loss layer for object dection algorithm of SSD** + **Multi-box loss layer for object detection algorithm of SSD** This layer is to compute dection loss for SSD given the location offset predictions, confidence predictions, prior boxes and ground-truth boudding @@ -372,21 +580,35 @@ def ssd_loss(location, is a weighted sum of the localization loss (or regression loss) and confidence loss (or classification loss) by performing the following steps: - 1. Find matched boundding box by bipartite matching algorithm. + 1. Find matched bounding box by bipartite matching algorithm. + 1.1 Compute IOU similarity between ground-truth boxes and prior boxes. + 1.2 Compute matched boundding box by bipartite matching algorithm. + 2. Compute confidence for mining hard examples + 2.1. Get the target label based on matched indices. + 2.2. Compute confidence loss. + 3. Apply hard example mining to get the negative example indices and update the matched indices. + 4. Assign classification and regression targets + 4.1. Encoded bbox according to the prior boxes. + 4.2. Assign regression targets. + 4.3. Assign classification targets. + 5. Compute the overall objective loss. + 5.1 Compute confidence loss. + 5.1 Compute localization loss. + 5.3 Compute the overall weighted loss. Args: @@ -421,39 +643,36 @@ def ssd_loss(location, mining_type (str): The hard example mining type, should be 'hard_example' or 'max_negative', now only support `max_negative`. normalize (bool): Whether to normalize the SSD loss by the total number - of output locations, True by defalut. + of output locations, True by default. sample_size (int): The max sample size of negative box, used only when mining_type is 'hard_example'. Returns: - Variable: The weighted sum of the localization loss and confidence loss, - with shape [N * Np, 1], N and Np are the same as they are - in `location`. + The weighted sum of the localization loss and confidence loss, with \ + shape [N * Np, 1], N and Np are the same as they are in `location`. Raises: - ValueError: If mining_type is 'hard_example', now only support - mining type of `max_negative`. + ValueError: If mining_type is 'hard_example', now only support mining \ + type of `max_negative`. Examples: - .. code-block:: python - - pb = layers.data( - name='prior_box', - shape=[10, 4], - append_batch_size=False, - dtype='float32') - pbv = layers.data( - name='prior_box_var', - shape=[10, 4], - append_batch_size=False, - dtype='float32') - loc = layers.data(name='target_box', shape=[10, 4], dtype='float32') - scores = layers.data(name='scores', shape=[10, 21], dtype='float32') - gt_box = layers.data( - name='gt_box', shape=[4], lod_level=1, dtype='float32') - gt_label = layers.data( - name='gt_label', shape=[1], lod_level=1, dtype='float32') - loss = layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv) + >>> pb = fluid.layers.data( + >>> name='prior_box', + >>> shape=[10, 4], + >>> append_batch_size=False, + >>> dtype='float32') + >>> pbv = fluid.layers.data( + >>> name='prior_box_var', + >>> shape=[10, 4], + >>> append_batch_size=False, + >>> dtype='float32') + >>> loc = fluid.layers.data(name='target_box', shape=[10, 4], dtype='float32') + >>> scores = fluid.layers.data(name='scores', shape=[10, 21], dtype='float32') + >>> gt_box = fluid.layers.data( + >>> name='gt_box', shape=[4], lod_level=1, dtype='float32') + >>> gt_label = fluid.layers.data( + >>> name='gt_label', shape=[1], lod_level=1, dtype='float32') + >>> loss = fluid.layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv) """ helper = LayerHelper('ssd_loss', **locals()) @@ -461,9 +680,10 @@ def ssd_loss(location, raise ValueError("Only support mining_type == max_negative now.") num, num_prior, num_class = confidence.shape + conf_shape = ops.shape(confidence) def __reshape_to_2d(var): - return nn.reshape(x=var, shape=[-1, var.shape[-1]]) + return nn.flatten(x=var, axis=2) # 1. Find matched boundding box by prior box. # 1.1 Compute IOU similarity between ground-truth boxes and prior boxes. @@ -474,7 +694,8 @@ def ssd_loss(location, # 2. Compute confidence for mining hard examples # 2.1. Get the target label based on matched indices - gt_label = nn.reshape(x=gt_label, shape=gt_label.shape + (1, )) + gt_label = nn.reshape( + x=gt_label, shape=(len(gt_label.shape) - 1) * (0, ) + (-1, 1)) gt_label.stop_gradient = True target_label, _ = target_assign( gt_label, matched_indices, mismatch_value=background_label) @@ -485,9 +706,12 @@ def ssd_loss(location, target_label = __reshape_to_2d(target_label) target_label.stop_gradient = True conf_loss = nn.softmax_with_cross_entropy(confidence, target_label) - # 3. Mining hard examples - conf_loss = nn.reshape(x=conf_loss, shape=(num, num_prior)) + conf_loss = nn.reshape( + x=conf_loss, + shape=(num, num_prior), + actual_shape=ops.slice( + conf_shape, axes=[0], starts=[0], ends=[2])) conf_loss.stop_gradient = True neg_indices = helper.create_tmp_variable(dtype='int32') dtype = matched_indices.dtype @@ -506,7 +730,7 @@ def ssd_loss(location, }, attrs={ 'neg_pos_ratio': neg_pos_ratio, - 'neg_dist_threshold': neg_pos_ratio, + 'neg_dist_threshold': neg_overlap, 'mining_type': mining_type, 'sample_size': sample_size, }) @@ -556,7 +780,11 @@ def ssd_loss(location, # 5.3 Compute overall weighted loss. loss = conf_loss_weight * conf_loss + loc_loss_weight * loc_loss # reshape to [N, Np], N is the batch size and Np is the prior box number. - loss = nn.reshape(x=loss, shape=[-1, num_prior]) + loss = nn.reshape( + x=loss, + shape=(num, num_prior), + actual_shape=ops.slice( + conf_shape, axes=[0], starts=[0], ends=[2])) loss = nn.reduce_sum(loss, dim=1, keep_dim=True) if normalize: normalizer = nn.reduce_sum(target_loc_weight) @@ -575,9 +803,10 @@ def prior_box(input, clip=False, steps=[0.0, 0.0], offset=0.5, - name=None): + name=None, + min_max_aspect_ratios_order=False): """ - **Prior box operator** + **Prior Box Operator** Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm. Each position of the input produce N prior boxes, N is determined by @@ -604,28 +833,37 @@ def prior_box(input, Default: [0., 0.] offset(float): Prior boxes center offset. Default: 0.5 name(str): Name of the prior box op. Default: None. + min_max_aspect_ratios_order(bool): If set True, the output prior box is + in order of [min, max, aspect_ratios], which is consistent with + Caffe. Please note, this order affects the weights order of + convolution layer followed by and does not affect the final + detection results. Default: False. Returns: - boxes(Variable): the output prior boxes of PriorBox. - The layout is [H, W, num_priors, 4]. - H is the height of input, W is the width of input, - num_priors is the total - box count of each position of input. - Variances(Variable): the expanded variances of PriorBox. - The layout is [H, W, num_priors, 4]. - H is the height of input, W is the width of input - num_priors is the total - box count of each position of input + tuple: A tuple with two Variable (boxes, variances) + + boxes: the output prior boxes of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input, + num_priors is the total + box count of each position of input. + + variances: the expanded variances of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input + num_priors is the total + box count of each position of input Examples: .. code-block:: python - box, var = prior_box( - input=conv1, - image=images, - min_sizes=[100.], - flip=True, - clip=True) + + box, var = fluid.layers.prior_box( + input=conv1, + image=images, + min_sizes=[100.], + flip=True, + clip=True) """ helper = LayerHelper("prior_box", **locals()) dtype = helper.input_dtype() @@ -653,7 +891,8 @@ def prior_box(input, 'clip': clip, 'step_w': steps[0], 'step_h': steps[1], - 'offset': offset + 'offset': offset, + 'min_max_aspect_ratios_order': min_max_aspect_ratios_order } if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0: if not _is_list_or_tuple_(max_sizes): @@ -693,13 +932,12 @@ def multi_box_head(inputs, kernel_size=1, pad=0, stride=1, - name=None): + name=None, + min_max_aspect_ratios_order=False): """ - **Prior_boxes** - Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm. The details of this algorithm, please refer the - section 2.2 of SSD paper (SSD: Single Shot MultiBox Detector) + section 2.2 of SSD paper `SSD: Single Shot MultiBox Detector `_ . Args: @@ -738,26 +976,34 @@ def multi_box_head(inputs, pad(int|list|tuple): The padding of conv2d. Default:0. stride(int|list|tuple): The stride of conv2d. Default:1, name(str): Name of the prior box layer. Default: None. + min_max_aspect_ratios_order(bool): If set True, the output prior box is + in order of [min, max, aspect_ratios], which is consistent with + Caffe. Please note, this order affects the weights order of + convolution layer followed by and does not affect the fininal + detection results. Default: False. Returns: - mbox_loc(Variable): The predicted boxes' location of the inputs. - The layout is [N, H*W*Priors, 4]. where Priors - is the number of predicted boxes each position of each input. - mbox_conf(Variable): The predicted boxes' confidence of the inputs. - The layout is [N, H*W*Priors, C]. where Priors - is the number of predicted boxes each position of each input - and C is the number of Classes. - boxes(Variable): the output prior boxes of PriorBox. - The layout is [num_priors, 4]. num_priors is the total - box count of each position of inputs. - Variances(Variable): the expanded variances of PriorBox. - The layout is [num_priors, 4]. num_priors is the total - box count of each position of inputs + tuple: A tuple with four Variables. (mbox_loc, mbox_conf, boxes, variances) + + mbox_loc: The predicted boxes' location of the inputs. The layout + is [N, H*W*Priors, 4]. where Priors is the number of predicted + boxes each position of each input. + + mbox_conf: The predicted boxes' confidence of the inputs. The layout + is [N, H*W*Priors, C]. where Priors is the number of predicted boxes + each position of each input and C is the number of Classes. + + boxes: the output prior boxes of PriorBox. The layout is [num_priors, 4]. + num_priors is the total box count of each position of inputs. + + variances: the expanded variances of PriorBox. The layout is + [num_priors, 4]. num_priors is the total box count of each position of inputs Examples: .. code-block:: python - mbox_locs, mbox_confs, box, var = layers.multi_box_head( + + mbox_locs, mbox_confs, box, var = fluid.layers.multi_box_head( inputs=[conv1, conv2, conv3, conv4, conv5, conv5], image=images, num_classes=21, @@ -771,13 +1017,7 @@ def multi_box_head(inputs, """ def _reshape_with_axis_(input, axis=1): - if not (axis > 0 and axis < len(input.shape)): - raise ValueError("The axis should be smaller than " - "the arity of input and bigger than 0.") - new_shape = [ - -1, reduce(lambda x, y: x * y, input.shape[axis:len(input.shape)]) - ] - out = nn.reshape(x=input, shape=new_shape) + out = nn.flatten(x=input, axis=axis) return out def _is_list_or_tuple_(data): @@ -799,7 +1039,7 @@ def multi_box_head(inputs, min_sizes = [] max_sizes = [] step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2))) - for ratio in xrange(min_ratio, max_ratio + 1, step): + for ratio in range(min_ratio, max_ratio + 1, step): min_sizes.append(base_size * ratio / 100.) max_sizes.append(base_size * (ratio + step) / 100.) min_sizes = [base_size * .10] + min_sizes @@ -849,7 +1089,8 @@ def multi_box_head(inputs, step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0] box, var = prior_box(input, image, min_size, max_size, aspect_ratio, - variance, flip, clip, step, offset) + variance, flip, clip, step, offset, None, + min_max_aspect_ratios_order) box_results.append(box) var_results.append(var) @@ -866,11 +1107,13 @@ def multi_box_head(inputs, stride=stride) mbox_loc = nn.transpose(mbox_loc, perm=[0, 2, 3, 1]) - new_shape = [ + compile_shape = [ mbox_loc.shape[0], mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3] / 4, 4 ] - mbox_loc_flatten = nn.reshape(mbox_loc, shape=new_shape) + run_shape = tensor.assign(numpy.array([0, -1, 4]).astype("int32")) + mbox_loc_flatten = nn.reshape( + mbox_loc, shape=compile_shape, actual_shape=run_shape) mbox_locs.append(mbox_loc_flatten) # get conf @@ -882,11 +1125,15 @@ def multi_box_head(inputs, padding=pad, stride=stride) conf_loc = nn.transpose(conf_loc, perm=[0, 2, 3, 1]) - new_shape = [ + new_shape = [0, -1, num_classes] + compile_shape = [ conf_loc.shape[0], conf_loc.shape[1] * conf_loc.shape[2] * conf_loc.shape[3] / num_classes, num_classes ] - conf_loc_flatten = nn.reshape(conf_loc, shape=new_shape) + run_shape = tensor.assign( + numpy.array([0, -1, num_classes]).astype("int32")) + conf_loc_flatten = nn.reshape( + conf_loc, shape=compile_shape, actual_shape=run_shape) mbox_confs.append(conf_loc_flatten) if len(box_results) == 1: @@ -909,3 +1156,95 @@ def multi_box_head(inputs, box.stop_gradient = True var.stop_gradient = True return mbox_locs_concat, mbox_confs_concat, box, var + + +def anchor_generator(input, + anchor_sizes=None, + aspect_ratios=None, + variance=[0.1, 0.1, 0.2, 0.2], + stride=None, + offset=0.5, + name=None): + """ + **Anchor generator operator** + + Generate anchors for Faster RCNN algorithm. + Each position of the input produce N anchors, N = + size(anchor_sizes) * size(aspect_ratios). The order of generated anchors + is firstly aspect_ratios loop then anchor_sizes loop. + + Args: + input(Variable): The input feature map, the format is NCHW. + anchor_sizes(list|tuple|float): The anchor sizes of generated anchors, + given in absolute pixels e.g. [64., 128., 256., 512.]. + For instance, the anchor size of 64 means the area of this anchor equals to 64**2. + aspect_ratios(list|tuple|float): The height / width ratios of generated + anchors, e.g. [0.5, 1.0, 2.0]. + variance(list|tuple): The variances to be used in box regression deltas. + Default:[0.1, 0.1, 0.2, 0.2]. + stride(list|turple): The anchors stride across width and height, + e.g. [16.0, 16.0] + offset(float): Prior boxes center offset. Default: 0.5 + name(str): Name of the prior box op. Default: None. + + Returns: + Anchors(Variable): The output anchors with a layout of [H, W, num_anchors, 4]. + H is the height of input, W is the width of input, + num_anchors is the box count of each position. + Each anchor is in (xmin, ymin, xmax, ymax) format an unnormalized. + Variances(Variable): The expanded variances of anchors + with a layout of [H, W, num_priors, 4]. + H is the height of input, W is the width of input + num_anchors is the box count of each position. + Each variance is in (xcenter, ycenter, w, h) format. + + + Examples: + + .. code-block:: python + + anchor, var = anchor_generator( + input=conv1, + anchor_sizes=[64, 128, 256, 512], + aspect_ratios=[0.5, 1.0, 2.0], + variance=[0.1, 0.1, 0.2, 0.2], + stride=[16.0, 16.0], + offset=0.5) + """ + helper = LayerHelper("anchor_generator", **locals()) + dtype = helper.input_dtype() + + def _is_list_or_tuple_(data): + return (isinstance(data, list) or isinstance(data, tuple)) + + if not _is_list_or_tuple_(anchor_sizes): + anchor_sizes = [anchor_sizes] + if not _is_list_or_tuple_(aspect_ratios): + aspect_ratios = [aspect_ratios] + if not (_is_list_or_tuple_(stride) and len(stride) == 2): + raise ValueError('stride should be a list or tuple ', + 'with length 2, (stride_width, stride_height).') + + anchor_sizes = list(map(float, anchor_sizes)) + aspect_ratios = list(map(float, aspect_ratios)) + stride = list(map(float, stride)) + + attrs = { + 'anchor_sizes': anchor_sizes, + 'aspect_ratios': aspect_ratios, + 'variances': variance, + 'stride': stride, + 'offset': offset + } + + anchor = helper.create_tmp_variable(dtype) + var = helper.create_tmp_variable(dtype) + helper.append_op( + type="anchor_generator", + inputs={"Input": input}, + outputs={"Anchors": anchor, + "Variances": var}, + attrs=attrs, ) + anchor.stop_gradient = True + var.stop_gradient = True + return anchor, var diff --git a/python/paddle/fluid/layers/device.py b/python/paddle/fluid/layers/device.py index e0c1aab230aeed7fb858e91e7da7eae58032ee16..bb1fb7fd571a56acf367e663af0cf9431211bcea 100644 --- a/python/paddle/fluid/layers/device.py +++ b/python/paddle/fluid/layers/device.py @@ -15,13 +15,15 @@ All util layers. """ -from layer_function_generator import autodoc +from .layer_function_generator import autodoc from ..framework import unique_name from ..layer_helper import LayerHelper +from ..annotations import deprecated -__all__ = ['get_places'] +__all__ = [] +@deprecated(since='0.15.0', instead="ParallelExecutor") @autodoc() def get_places(device_count=None, device_type=None): helper = LayerHelper('get_places', **locals()) diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 8758ac9f94ab91b5be5fc70917c64db38997d1c1..327ae309816344a0bcebfe70ffb59a00eab1d86f 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -12,18 +12,23 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib +import multiprocessing +import threading +from ..data_feeder import DataFeeder +from .control_flow import BlockGuard +from .layer_function_generator import templatedoc from .. import core -from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program, Program -from ..unique_name import generate as unique_name -from control_flow import BlockGuard -from ..layer_helper import LayerHelper from ..executor import global_scope +from ..framework import convert_np_dtype_to_dtype_, default_main_program, \ + default_startup_program, program_guard, Program +from ..layer_helper import LayerHelper +from ..unique_name import generate as unique_name __all__ = [ - 'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file', - 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer', - 'random_data_generator', 'Preprocessor' + 'data', 'open_recordio_file', 'open_files', 'read_file', 'shuffle', 'batch', + 'double_buffer', 'random_data_generator', 'py_reader', 'Preprocessor', + 'load' ] @@ -64,7 +69,7 @@ def data(name, """ helper = LayerHelper('data', **locals()) shape = list(shape) - for i in xrange(len(shape)): + for i in range(len(shape)): if shape[i] is None: shape[i] = -1 append_batch_size = False @@ -108,10 +113,35 @@ class BlockGuardServ(BlockGuard): class ListenAndServ(object): """ - ListenAndServ class. + **ListenAndServ Layer** + + ListenAndServ is used to create a rpc server bind and listen + on specific TCP port, this server will run the sub-block when + received variables from clients. + + Args: + endpoint(string): IP:port string which the server will listen on. + inputs(list): a list of variables that the server will get from clients. + fan_in(int): how many client are expected to report to this server, default: 1. + optimizer_mode(bool): whether to run the server as a parameter server, default: True. + + Examples: + .. code-block:: python - ListenAndServ class is used to wrap listen_and_serv op to create a server - which can receive variables from clients and run a block. + with fluid.program_guard(main): + serv = layers.ListenAndServ( + "127.0.0.1:6170", ["X"], optimizer_mode=False) + with serv.do(): + x = layers.data( + shape=[32, 32], + dtype='float32', + name="X", + append_batch_size=False) + fluid.initializer.Constant(value=1.0)(x, main.global_block()) + layers.scale(x=x, scale=10.0, out=out_var) + + exe = fluid.Executor(place) + exe.run(main) """ def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True): @@ -160,7 +190,6 @@ class ListenAndServ(object): main_program = self.helper.main_program current_block = main_program.current_block() parent_block = self.parent_block() - empty_block = Program().global_block() parent_block.append_op( type='listen_and_serv', @@ -169,25 +198,25 @@ class ListenAndServ(object): attrs={ 'endpoint': self.endpoint, 'Fanin': self.fan_in, - 'OptimizeBlock': current_block, - 'PrefetchBlock': empty_block, + 'optimize_blocks': [ + current_block + ], # did not support multiple optimize blocks in layers 'sync_mode': True, # did not support async now in layers 'grad_to_block_id': [""] }) -def Send(endpoints, send_vars, get_vars=None): +def Send(endpoints, send_vars, sync=True): """ - Send layer + Send variables to the server side, and get vars from server + side when server have finished running server side program. Args: - endpoints: comma seperated IP:PORT pairs in the order + endpoints (str): comma seperated IP:PORT pairs in the order of send_vars to send - send_vars: vars to send - get_vars: vars to get from server after send completes. + send_vars (list): variables to send to server + sync (bool): whether to wait the request finish - Send variables to the server side, and get vars from server - side when server have finished running server side program. """ assert (type(send_vars) == list) @@ -195,40 +224,33 @@ def Send(endpoints, send_vars, get_vars=None): endpoints = list(set(epmap)) helper = LayerHelper("Send", **locals()) - if not get_vars: - get_vars = [] - for s in send_vars: - v = helper.create_tmp_variable(dtype=s.dtype, stop_gradient=True) - get_vars.append(v) rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName() helper.append_op( type="send", inputs={"X": send_vars}, - outputs={"Out": get_vars}, attrs={ "endpoints": endpoints, "epmap": epmap, rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC }) - - return get_vars + if sync: + helper.append_op(type="send_barrier", attrs={"endpoints": endpoints}) -def Recv(endpoints, get_vars): +def Recv(endpoints, get_vars, sync=True): """ - Recv layer + Receive variables from server side Args: - endpoints: comma seperated IP:PORT pairs in the order + endpoints (str): comma seperated IP:PORT pairs in the order of send_vars to send - send_vars: vars to send - get_vars: vars to get from server after send completes. + get_vars (list): vars to get from server after send completes. + sync (bool): whether to wait the request finish - Send variables to the server side, and get vars from server - side when server have finished running server side program. + Returns: + list: list of received variables """ - assert (type(send_vars) == list) assert (type(get_vars) == list) epmap = endpoints.split(",") @@ -241,6 +263,9 @@ def Recv(endpoints, get_vars): outputs={"Out": get_vars}, attrs={"endpoints": endpoints, "epmap": epmap}) + if sync: + helper.append_op(type="fetch_barrier", attrs={"endpoints": endpoints}) + return get_vars def monkey_patch_reader_methods(reader): @@ -291,6 +316,7 @@ def _copy_reader_create_op_(block, op): return new_op +@templatedoc(op_type='create_recordio_file_reader') def open_recordio_file(filename, shapes, lod_levels, @@ -298,34 +324,30 @@ def open_recordio_file(filename, pass_num=1, for_parallel=True): """ - Open a RecordIO file - - This layer takes a RecordIO file to read from and returns a Reader Variable. - Via the Reader Variable, we can get data from the given RecordIO file. + ${comment} Args: - filename(str): The RecordIO file's name. + filename(${filename_type}): ${filename_comment}. shapes(list): List of tuples which declaring data shapes. - lod_levels(list): List of ints which declaring data lod_level. + lod_levels(${lod_levels_type}): ${lod_levels_comment}. dtypes(list): List of strs which declaring data type. pass_num(int): Number of passes to run. for_parallel(Bool): Set it as True if you are going to run subsequent operators in parallel. Returns: - Variable: A Reader Variable via which we can get RecordIO file data. + ${out_comment}. Examples: - .. code-block:: python - reader = fluid.layers.io.open_recordio_file( - filename='./data.recordio', - shapes=[(3,224,224), (1)], - lod_levels=[0, 0], - dtypes=['float32', 'int64']) - - # Via the reader, we can use 'read_file' layer to get data: - image, label = fluid.layers.io.read_file(reader) + >>> import paddle.fluid as fluid + >>> reader = fluid.layers.io.open_recordio_file( + >>> filename='./data.recordio', + >>> shapes=[(3,224,224), (1)], + >>> lod_levels=[0, 0], + >>> dtypes=['float32', 'int64']) + >>> # Via the reader, we can use 'read_file' layer to get data: + >>> image, label = fluid.layers.io.read_file(reader) """ dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] shape_concat = [] @@ -357,9 +379,6 @@ def open_recordio_file(filename, if pass_num > 1: main_prog_var = multi_pass(reader=main_prog_var, pass_num=pass_num) - if for_parallel: - main_prog_var = parallel(reader=main_prog_var) - return monkey_patch_reader_methods(main_prog_var) @@ -368,9 +387,9 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): Create a uniform random data generator This layer returns a Reader Variable. - Instead of opening a file and reading data from it, this - Reader Variable generates float uniform random data by itself. - It can be used as a dummy reader to test a network without + Instead of opening a file and reading data from it, this + Reader Variable generates float uniform random data by itself. + It can be used as a dummy reader to test a network without opening a real file. Args: @@ -385,16 +404,16 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): Variable: A Reader Variable from which we can get random data. Examples: - .. code-block:: python - reader = fluid.layers.io.random_data_generator( - low=0.0, - high=1.0, - shapes=[(3,224,224), (1)], - lod_levels=[0, 0]) + .. code-block:: python - # Via the reader, we can use 'read_file' layer to get data: - image, label = fluid.layers.io.read_file(reader) + reader = fluid.layers.random_data_generator( + low=0.0, + high=1.0, + shapes=[[3,224,224], [1]], + lod_levels=[0, 0]) + # Via the reader, we can use 'read_file' layer to get data: + image, label = fluid.layers.read_file(reader) """ dtypes = [core.VarDesc.VarType.FP32] * len(shapes) shape_concat = [] @@ -424,37 +443,287 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): main_prog_var = _copy_reader_var_(default_main_program().current_block(), startup_var) - if for_parallel: - main_prog_var = parallel(reader=main_prog_var) - return monkey_patch_reader_methods(main_prog_var) +def py_reader(capacity, + shapes, + dtypes, + lod_levels=None, + name=None, + use_double_buffer=True): + """ + Create a Python reader for data feeding in Python + + This layer returns a Reader Variable. + The Reader provides :code:`decorate_paddle_reader()` and + :code:`decorate_tensor_provider()` to set a Python generator as the data + source in Python side. When :code:`Executor::Run()` is invoked in C++ + side, the data from the generator would be read automatically. Unlike + :code:`DataFeeder.feed()`, the data reading process and + :code:`Executor::Run()` process can run in parallel using + :code:`py_reader`. The :code:`start()` method of the Reader should be + called when each pass begins, while the :code:`reset()` method should be + called when the pass ends and :code:`fluid.core.EOFException` raises. + Note that :code:`Program.clone()` method cannot clone :code:`py_reader`. + + Args: + capacity(int): The buffer capacity maintained by :code:`py_reader`. + shapes(list|tuple): List of tuples which declaring data shapes. + dtypes(list|tuple): List of strs which declaring data type. + lod_levels(list|tuple): List of ints which declaring data lod_level. + name(basestring): The prefix Python queue name and Reader name. None will + be generated automatically. + use_double_buffer(bool): Whether use double buffer or not. + + Returns: + Variable: A Reader from which we can get feeding data. + + Examples: + + 1. The basic usage of :code:`py_reader` is as follows: + + >>> import paddle.v2 + >>> import paddle.fluid as fluid + >>> import paddle.dataset.mnist as mnist + >>> + >>> reader = fluid.layers.py_reader(capacity=64, + >>> shapes=[(-1,3,224,224), (-1,1)], + >>> dtypes=['float32', 'int64']) + >>> reader.decorate_paddle_reader( + >>> paddle.v2.reader.shuffle(paddle.batch(mnist.train()) + >>> + >>> img, label = fluid.layers.read_file(reader) + >>> loss = network(img, label) # some network definition + >>> + >>> fluid.Executor(fluid.CUDAPlace(0)).run(fluid.default_startup_program()) + >>> + >>> exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name) + >>> for epoch_id in range(10): + >>> reader.start() + >>> try: + >>> while True: + >>> exe.run(fetch_list=[loss.name]) + >>> except fluid.core.EOFException: + >>> reader.reset() + + 2. When training and testing are both performed, two different + :code:`py_reader` should be created with different names, e.g.: + + >>> import paddle.v2 + >>> import paddle.fluid as fluid + >>> import paddle.dataset.mnist as mnist + >>> + >>> def network(reader): + >>> img, label = fluid.layers.read_file(reader) + >>> # Here, we omitted the network definition + >>> return loss + >>> + >>> train_reader = fluid.layers.py_reader(capacity=64, + >>> shapes=[(-1,3,224,224), (-1,1)], + >>> dtypes=['float32', 'int64'], + >>> name='train_reader') + >>> train_reader.decorate_paddle_reader( + >>> paddle.v2.reader.shuffle(paddle.batch(mnist.train()) + >>> + >>> test_reader = fluid.layers.py_reader(capacity=32, + >>> shapes=[(-1,3,224,224), (-1,1)], + >>> dtypes=['float32', 'int64'], + >>> name='test_reader') + >>> test_reader.decorate_paddle_reader(paddle.batch(mnist.test(), 512)) + >>> + >>> # Create train_main_prog and train_startup_prog + >>> train_main_prog = fluid.Program() + >>> train_startup_prog = fluid.Program() + >>> with fluid.program_guard(train_main_prog, train_startup_prog): + >>> # Use fluid.unique_name.guard() to share parameters with test program + >>> with fluid.unique_name.guard(): + >>> train_loss = network(train_reader) # some network definition + >>> adam = fluid.optimizer.Adam(learning_rate=0.01) + >>> adam.minimize(loss) + >>> + >>> # Create test_main_prog and test_startup_prog + >>> test_main_prog = fluid.Program() + >>> test_startup_prog = fluid.Program() + >>> with fluid.program_guard(test_main_prog, test_startup_prog): + >>> # Use fluid.unique_name.guard() to share parameters with train program + >>> with fluid.unique_name.guard(): + >>> test_loss = network(test_reader) + >>> + >>> fluid.Executor(fluid.CUDAPlace(0)).run(train_startup_prog) + >>> fluid.Executor(fluid.CUDAPlace(0)).run(test_startup_prog) + >>> + >>> train_exe = fluid.ParallelExecutor(use_cuda=True, + >>> loss_name=train_loss.name, main_program=train_main_prog) + >>> test_exe = fluid.ParallelExecutor(use_cuda=True, + >>> loss_name=test_loss.name, main_program=test_main_prog) + >>> for epoch_id in range(10): + >>> train_reader.start() + >>> try: + >>> while True: + >>> train_exe.run(fetch_list=[train_loss.name]) + >>> except fluid.core.EOFException: + >>> train_reader.reset() + >>> + >>> test_reader.start() + >>> try: + >>> while True: + >>> test_exe.run(fetch_list=[test_loss.name]) + >>> except fluid.core.EOFException: + >>> test_reader.reset() + """ + dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] + shape_concat = [] + ranks = [] + + for shape in shapes: + shape_concat.extend(shape) + ranks.append(len(shape)) + + if lod_levels is None: + lod_levels = [0] * len(shapes) + + if name is None: + queue_name = unique_name('lod_tensor_blocking_queue') + reader_name = unique_name('create_py_reader') + double_buffer_name = unique_name('double_buffer') + else: + queue_name = "_".join([name, "queue"]) + reader_name = "_".join([name, "reader"]) + double_buffer_name = "_".join([name, "double_buffer"]) + + var = global_scope().var(queue_name) + feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) + + startup_blk = default_startup_program().current_block() + startup_var = startup_blk.create_var(name=reader_name) + startup_blk.append_op( + type='create_py_reader', + inputs={'blocking_queue': [queue_name]}, + outputs={'Out': [startup_var]}, + attrs={ + 'shape_concat': shape_concat, + 'lod_levels': lod_levels, + 'ranks': ranks + }) + + startup_var.desc.set_dtypes(dtypes) + startup_var.persistable = True + + main_prog_var = _copy_reader_var_(default_main_program().current_block(), + startup_var) + + reader = monkey_patch_reader_methods(main_prog_var) + if use_double_buffer: + double_buffer_reader = double_buffer(reader, name=double_buffer_name) + # we return a double buffer reader. However, the reset method comes from + # py_reader. + double_buffer_reader.reset = reader.reset + reader = double_buffer_reader + + # monkey patch py_reader special methods + reader.queue = feed_queue + current_reset_method = reader.reset + reader.thread = None + reader.tensor_provider = None + reader.exited = False + + def start_provide_thread(func): + def __provider_thread__(): + for tensors in func(): + array = core.LoDTensorArray() + for item in tensors: + if not isinstance(item, core.LoDTensor): + tmp = core.LoDTensor() + tmp.set(item, core.CPUPlace()) + item = tmp + + array.append(item) + + if reader.exited: + break + feed_queue.push(array) + if reader.exited: + break + feed_queue.close() + + reader.thread = threading.Thread(target=__provider_thread__) + reader.thread.daemon = True + reader.thread.start() + + def __set_tensor_provider__(func): + reader.tensor_provider = func + + def __set_paddle_reader__(paddle_reader): + with program_guard(Program(), Program()): + feed_list = [] + counter = 0 + for dtype, shape, lod_level in zip(dtypes, shapes, lod_levels): + name = str(counter) + feed_list.append( + data( + name=name, + dtype=dtype, + shape=shape, + lod_level=lod_level)) + counter += 1 + + feeder = DataFeeder(feed_list=feed_list, place=core.CPUPlace()) + paddle_reader = feeder.decorate_reader( + paddle_reader, multi_devices=False) + + def __tensor_provider__(): + for slots in paddle_reader(): + yield [slots[str(idx)] for idx in xrange(counter)] + + __set_tensor_provider__(__tensor_provider__) + + def __reset__(): + current_reset_method() + if reader.thread is not None and reader.tensor_provider is not None: + reader.exited = True + reader.thread.join() + reader.exited = False + + def __start__(): + start_provide_thread(reader.tensor_provider) + + reader.reset = __reset__ + reader.decorate_tensor_provider = __set_tensor_provider__ + reader.decorate_paddle_reader = __set_paddle_reader__ + reader.start = __start__ + + return reader + + def open_files(filenames, shapes, lod_levels, dtypes, - thread_num, + thread_num=None, buffer_size=None, pass_num=1, - for_parallel=True): + is_test=None): """ Open files - This layer takes a list of files to read from and returns a Reader Variable. - Via the Reader Variable, we can get data from given files. All files must - have name suffixs to indicate their formats, e.g., '*.recordio'. + This layer takes a list of files to read from and returns a Reader Variable. + Via the Reader Variable, we can get data from given files. All files must + have name suffixs to indicate their formats, e.g., '*.recordio'. Args: filenames(list): The list of file names. shapes(list): List of tuples which declaring data shapes. lod_levels(list): List of ints which declaring data lod_level. dtypes(list): List of strs which declaring data type. - thread_num(int): The maximal concurrent prefetch thread number. - buffer_size(int): The size of prefetch buffer. + thread_num(None): The number of thread to read files. + Default: min(len(filenames), cpu_number). + buffer_size(None): The buffer size of reader. Default: 3 * thread_num pass_num(int): Number of passes to run. - for_parallel(Bool): Set it as True if you are going to run - subsequent operators in parallel. + is_test(bool|None): Whether `open_files` used for testing or not. If it + is used for testing, the order of data generated is same as the file + order. Otherwise, it is not guaranteed the order of data is same + between every epoch. [Default: False]. Returns: Variable: A Reader Variable via which we can get file data. @@ -466,15 +735,21 @@ def open_files(filenames, './data2.recordio'], shapes=[(3,224,224), (1)], lod_levels=[0, 0], - dtypes=['float32', 'int64'], - thread_num=2, - buffer_size=2) + dtypes=['float32', 'int64']) # Via the reader, we can use 'read_file' layer to get data: image, label = fluid.layers.io.read_file(reader) """ + if thread_num is None: + thread_num = min(len(filenames), multiprocessing.cpu_count()) + else: + thread_num = int(thread_num) + if buffer_size is None: - buffer_size = thread_num + buffer_size = 3 * thread_num + else: + buffer_size = int(buffer_size) + if isinstance(filenames, basestring): filenames = [filenames] dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] @@ -488,17 +763,18 @@ def open_files(filenames, multi_file_reader_name = unique_name('multi_file_reader') startup_blk = default_startup_program().current_block() startup_reader = startup_blk.create_var(name=multi_file_reader_name) + attrs = { + 'shape_concat': shape_concat, + 'lod_levels': lod_levels, + 'ranks': ranks, + 'file_names': filenames, + 'thread_num': thread_num, + 'buffer_size': buffer_size + } + if is_test is not None: + attrs['is_test'] = is_test startup_blk.append_op( - type='open_files', - outputs={'Out': [startup_reader]}, - attrs={ - 'shape_concat': shape_concat, - 'lod_levels': lod_levels, - 'ranks': ranks, - 'file_names': filenames, - 'thread_num': thread_num, - 'buffer_size': buffer_size - }) + type='open_files', outputs={'Out': [startup_reader]}, attrs=attrs) startup_reader.desc.set_dtypes(dtypes) startup_reader.persistable = True @@ -508,9 +784,6 @@ def open_files(filenames, main_prog_reader = multi_pass( reader=main_prog_reader, pass_num=pass_num) - if for_parallel: - main_prog_reader = parallel(reader=main_prog_reader) - return monkey_patch_reader_methods(main_prog_reader) @@ -543,16 +816,77 @@ def __create_unshared_decorated_reader__(op_type, reader, attrs, name=None): def shuffle(reader, buffer_size): + """ + Shuffle the reader. + """ return __create_unshared_decorated_reader__( 'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)}) def batch(reader, batch_size): + """ + This layer is a reader decorator. It takes a reader and adds + 'batching' decoration on it. When reading with the result + decorated reader, output data will be automatically organized + to the form of batches. + + Args: + reader(Variable): The reader to be decorated with 'batching'. + batch_size(int): The batch size. + + Returns: + Variable: The reader which has been decorated with 'batching'. + + Examples: + .. code-block:: python + + raw_reader = fluid.layers.io.open_files(filenames=['./data1.recordio', + './data2.recordio'], + shapes=[(3,224,224), (1)], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + thread_num=2, + buffer_size=2) + batch_reader = fluid.layers.batch(reader=raw_reader, batch_size=5) + + # If we read data with the raw_reader: + # data = fluid.layers.read_file(raw_reader) + # We can only get data instance by instance. + # + # However, if we read data with the batch_reader: + # data = fluid.layers.read_file(batch_reader) + # Each 5 adjacent instances will be automatically combined together + # to become a batch. So what we get('data') is a batch data instead + # of an instance. + """ return __create_unshared_decorated_reader__( 'create_batch_reader', reader, {'batch_size': int(batch_size)}) def double_buffer(reader, place=None, name=None): + """ + Wrap a double buffer reader. The data will copy to target place with a + double buffer queue. If the target place is None, the place that executor + perform on will be used. + + Args: + reader(Variable): the reader variable need to be wrapped. + place(Place): the place of target data. Default is the sample place of + executor perform. + + name(str): Variable name. None if the user does not care. + + Returns: + wrapped reader with double buffer. + + Examples: + + >>> reader = fluid.layers.open_files(filenames=['somefile'], + >>> shapes=[[-1, 784], [-1, 1]], + >>> dtypes=['float32', 'int64']) + >>> reader = fluid.layers.double_buffer(reader) + >>> img, label = fluid.layers.read_file(reader) + """ attrs = dict() if place is not None: attrs['place'] = str(place).upper() @@ -565,20 +899,41 @@ def multi_pass(reader, pass_num): 'create_multi_pass_reader', reader, {'pass_num': int(pass_num)}) -def parallel(reader): - return __create_shared_decorated_reader__('create_threaded_reader', reader, - {}) +def read_file(reader): + """ + Execute the given reader and get data via it. + A reader is also a Variable. It can be a raw reader generated by + `fluid.layers.open_files()` or a decorated one generated by + `fluid.layers.double_buffer()` and so on. -def read_file(file_obj): + Args: + + reader(Variable): The reader to execute. + + Returns: + Tuple[Variable]: Data read via the given reader. + + Examples: + .. code-block:: python + + data_file = fluid.layers.open_files( + filenames=['mnist.recordio'], + shapes=[(-1, 748), (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"]) + data_file = fluid.layers.double_buffer( + fluid.layers.batch(data_file, batch_size=64)) + input, label = fluid.layers.read_file(data_file) + """ helper = LayerHelper('read_file') out = [ helper.create_tmp_variable( stop_gradient=True, dtype='float32') - for _ in range(len(file_obj.desc.shapes())) + for _ in range(len(reader.desc.shapes())) ] helper.append_op( - type='read', inputs={'Reader': [file_obj]}, outputs={'Out': out}) + type='read', inputs={'Reader': [reader]}, outputs={'Out': out}) if len(out) == 1: return out[0] else: @@ -586,6 +941,26 @@ def read_file(file_obj): class Preprocessor(object): + """ + A block for data pre-processing in reader. + + Args: + reader (Variable): A reader variable. + name (str, default None): The name of the reader. + + Examples: + .. code-block:: python + + preprocessor = fluid.layers.io.Preprocessor(reader=reader) + with preprocessor.block(): + img, lbl = preprocessor.inputs() + img_out = img / 2 + lbl_out = lbl + 1 + preprocessor.outputs(img_out, lbl_out) + + data_file = fluid.layers.io.double_buffer(preprocessor()) + + """ BEFORE_SUB_BLOCK = 0 IN_SUB_BLOCK = 1 AFTER_SUB_BLOCK = 2 @@ -602,7 +977,7 @@ class Preprocessor(object): self.sink_var_names = None self.status = Preprocessor.BEFORE_SUB_BLOCK - def is_completed(self): + def _is_completed(self): return self.sub_block and self.source_var_names and self.sink_var_names @contextlib.contextmanager @@ -612,7 +987,7 @@ class Preprocessor(object): yield self.main_prog.rollback() self.status = Preprocessor.AFTER_SUB_BLOCK - if not self.is_completed(): + if not self._is_completed(): raise RuntimeError( "The definition of preprocessor is incompleted! " "Please make sure that you have set input and output " @@ -630,7 +1005,7 @@ class Preprocessor(object): source_lod_levels = self.underlying_reader.desc.lod_levels() self.source_var_names = [ unique_name("preprocessor_source") - for _ in xrange(len(source_shapes)) + for _ in range(len(source_shapes)) ] source_vars = [] for var_name, shape, dtype, lod_level in zip( @@ -662,3 +1037,29 @@ class Preprocessor(object): "sink_var_names": self.sink_var_names }) return monkey_patch_reader_methods(self.reader) + + +@templatedoc() +def load(out, file_path, load_as_fp16=None): + """ + ${comment} + + >>> import paddle.fluid as fluid + >>> tmp_tensor = fluid.layers.create_tensor(dtype='float32') + >>> fluid.layers.load(tmp_tensor, "./tmp_tensor.bin") + + Args: + out(${out_type}): ${out_comment}. + + file_path(${file_path_type}): ${file_path_comment}. + + load_as_fp16(${load_as_fp16_type}): ${load_as_fp16_comment}. + + Returns: + None + """ + helper = LayerHelper("load", **locals()) + attrs = {"file_path": file_path} + if load_as_fp16 is not None: + attrs['load_as_fp16'] = load_as_fp16 + helper.append_op(type="load", inputs={}, output={"Out": out}, args=attrs) diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 295d1b7190ec39bcc6efdf72aebede14a99807aa..c0d72620b1ddb183f43ebce766688518b5a737ac 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -12,19 +12,16 @@ # See the License for the specific language governing permissions and # limitations under the License. import re -import cStringIO import functools import warnings +import string +from six.moves import cStringIO from ..proto import framework_pb2 from ..framework import OpProtoHolder, Variable from ..layer_helper import LayerHelper -__all__ = [ - 'deprecated', - 'generate_layer_fn', - 'autodoc', -] +__all__ = ['deprecated', 'generate_layer_fn', 'autodoc', 'templatedoc'] def _convert_(name): @@ -43,6 +40,22 @@ def _convert_(name): return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower() +def _type_to_str_(tp): + return framework_pb2.AttrType.Name(tp) + + +_two_dollar_pattern_ = re.compile(r"\$\$([^\$]+)\$\$") +_single_dollar_pattern_ = re.compile(r"\$([^\$]+)\$") +_two_bang_pattern_ = re.compile(r"!!([^!]+)!!") + + +def escape_math(text): + return _two_bang_pattern_.sub( + r'$$\1$$', + _single_dollar_pattern_.sub(r':math:`\1`', + _two_dollar_pattern_.sub(r"!!\1!!", text))) + + def _generate_doc_string_(op_proto): """ Generate docstring by OpProto @@ -54,34 +67,33 @@ def _generate_doc_string_(op_proto): str: the document string """ - def _type_to_str_(tp): - return framework_pb2.AttrType.Name(tp) - if not isinstance(op_proto, framework_pb2.OpProto): raise TypeError("OpProto should be `framework_pb2.OpProto`") - buf = cStringIO.StringIO() - buf.write(op_proto.comment) + buf = cStringIO() + buf.write(escape_math(op_proto.comment)) buf.write('\nArgs:\n') for each_input in op_proto.inputs: line_begin = ' {0}: '.format(_convert_(each_input.name)) buf.write(line_begin) - buf.write(each_input.comment) - buf.write('\n') - buf.write(' ' * len(line_begin)) - buf.write('Duplicable: ') - buf.write(str(each_input.duplicable)) - buf.write(' Optional: ') - buf.write(str(each_input.dispensable)) + buf.write(escape_math(each_input.comment)) + if each_input.duplicable: + buf.write(" Duplicatable.") + if each_input.dispensable: + buf.write(" Optional.") buf.write('\n') + skip_attrs = OpProtoHolder.generated_op_attr_names() + for each_attr in op_proto.attrs: + if each_attr.name in skip_attrs: + continue buf.write(' ') buf.write(each_attr.name) buf.write(' (') buf.write(_type_to_str_(each_attr.type)) buf.write('): ') - buf.write(each_attr.comment) + buf.write(escape_math(each_attr.comment)) buf.write('\n') if len(op_proto.outputs) != 0: @@ -90,7 +102,7 @@ def _generate_doc_string_(op_proto): for each_opt in op_proto.outputs: if not each_opt.intermediate: break - buf.write(each_opt.comment) + buf.write(escape_math(each_opt.comment)) return buf.getvalue() @@ -107,9 +119,9 @@ def generate_layer_fn(op_type): """ op_proto = OpProtoHolder.instance().get_op_proto(op_type) not_intermediate_outputs = \ - filter(lambda output: not output.intermediate, op_proto.outputs) + [output for output in op_proto.outputs if not output.intermediate] intermediate_outputs = \ - filter(lambda output: output.intermediate, op_proto.outputs) + [output for output in op_proto.outputs if output.intermediate] if len(not_intermediate_outputs) != 1: raise ValueError("Only one non intermediate output operator can be", @@ -220,3 +232,61 @@ def autodoc(comment=""): return func return __impl__ + + +def templatedoc(op_type=None): + """ + Decorator of layer function. It will use the docstring from the layer + function as the template. The template arguments are: + + * ${comment}: The operator comment written in CPP. + * ${{name}_comment}: The comment of ${name} written with AddAttr, AddOutput, + and AddInput. The ${name} is Python snake style. i.e., xxx_xxx. + * ${{name}_type}: The type of ${name}. + + Returns: + Decorated function. + """ + + def trim_ending_dot(msg): + return msg.rstrip('.') + + def __impl__(func): + if op_type is None: + op_type_name = func.__name__ + else: + op_type_name = op_type + op_proto = OpProtoHolder.instance().get_op_proto(op_type_name) + tmpl = string.Template(func.__doc__) + + comment_lines = op_proto.comment.split("\n") + comment = "" + for line in comment_lines: + line = line.strip() + if len(line) != 0: + comment += escape_math(line) + comment += " " + elif len(comment) != 0: + comment += "\n \n " + + args = {"comment": trim_ending_dot(comment)} + for each_input in op_proto.inputs: + input_name = _convert_(each_input.name) + args["{0}_comment".format(input_name)] = trim_ending_dot( + each_input.comment) + args["{0}_type".format(input_name)] = "Variable" + for each_attr in op_proto.attrs: + input_name = _convert_(each_attr.name) + args["{0}_comment".format(input_name)] = trim_ending_dot( + each_attr.comment) + args["{0}_type".format(input_name)] = _type_to_str_(each_attr.type) + + for each_opt in op_proto.outputs: + output_name = _convert_(each_opt.name) + args["{0}_comment".format(output_name)] = trim_ending_dot( + each_opt.comment) + args["{0}_type".format(output_name)] = "Variable" + func.__doc__ = tmpl.substitute(args) + return func + + return __impl__ diff --git a/python/paddle/fluid/layers/learning_rate_scheduler.py b/python/paddle/fluid/layers/learning_rate_scheduler.py index d13c54daa5a985e2e1bf9357630fe29d24a17bb4..daf91a40f7ad7935d355a287819ad1dbcdd84eb8 100644 --- a/python/paddle/fluid/layers/learning_rate_scheduler.py +++ b/python/paddle/fluid/layers/learning_rate_scheduler.py @@ -11,17 +11,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - -import control_flow -import nn -import ops -import tensor -from ..initializer import init_on_cpu - -__all__ = [ - 'exponential_decay', 'natural_exp_decay', 'inverse_time_decay', - 'polynomial_decay', 'piecewise_decay', 'noam_decay' -] """ When training a model, it's often useful to decay the learning rate during training process, this is called @@ -31,6 +20,18 @@ User can also implement their own learning_rate_decay strategy according to this module. """ +from . import control_flow +from . import nn +from . import ops +from . import tensor +from ..initializer import init_on_cpu +from ..framework import default_main_program, Parameter + +__all__ = [ + 'exponential_decay', 'natural_exp_decay', 'inverse_time_decay', + 'polynomial_decay', 'piecewise_decay', 'noam_decay', 'append_LARS' +] + def _decay_step_counter(begin=0): # the first global step is zero in learning rate decay @@ -41,57 +42,76 @@ def _decay_step_counter(begin=0): def noam_decay(d_model, warmup_steps): - """Apply decay to learning rate. - ```python - lr_value = np.power(d_model, -0.5) * np.min([ - np.power(current_steps, -0.5), - np.power(warmup_steps, -1.5) * current_steps - ]) - ``` + """ + Noam decay method. The numpy implementation of noam decay as follows. + + >>> import numpy as np + >>> lr_value = np.power(d_model, -0.5) * np.min([ + >>> np.power(current_steps, -0.5), + >>> np.power(warmup_steps, -1.5) * current_steps]) + + Please reference `attention is all you need + `_. Args: d_model(Variable): The dimensionality of input and output of model. - Reference: attention is all you need - https://arxiv.org/pdf/1706.03762.pdf + warmup_steps(Variable): A super parameter. Returns: The decayed learning rate. """ global_step = _decay_step_counter(1) - with init_on_cpu(): - a = global_step**-0.5 - b = (warmup_steps**-1.5) * global_step - lr_value = (d_model**-0.5) * ops.elementwise_min(a, b) + + a = global_step**-0.5 + b = (warmup_steps**-1.5) * global_step + lr_value = (d_model**-0.5) * ops.elementwise_min(a, b) return lr_value def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False): - """Applies exponential decay to the learning rate. + """ + Applies exponential decay to the learning rate. + + When training a model, it is often recommended to lower the learning rate as the + training progresses. By using this function, the learning rate will be decayed by + 'decay_rate' every 'decay_steps' steps. + + >>> if staircase == True: + >>> decayed_learning_rate = learning_rate * decay_rate ^ floor(global_step / decay_steps) + >>> else: + >>> decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps) - ```python - decayed_learning_rate = learning_rate * - decay_rate ^ (global_step / decay_steps) - ``` Args: - learning_rate: A scalar float32 value or a Variable. This - will be the initial learning rate during training - decay_steps: A Python `int32` number. - decay_rate: A Python `float` number. - staircase: Boolean. If set true, decay the learning rate every decay_steps. + learning_rate(Variable|float): The initial learning rate. + decay_steps(int): See the decay computation above. + decay_rate(float): The decay rate. See the decay computation above. + staircase(Boolean): If True, decay the learning rate at discrete intervals. + Default: False Returns: - The decayed learning rate + Variable: The decayed learning rate + + Examples: + .. code-block:: python + + base_lr = 0.1 + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.exponential_decay( + learning_rate=base_lr, + decay_steps=10000, + decay_rate=0.5, + staircase=True)) + sgd_optimizer.minimize(avg_cost) + """ global_step = _decay_step_counter() - with init_on_cpu(): - # update learning_rate - div_res = global_step / decay_steps - if staircase: - div_res = ops.floor(div_res) - decayed_lr = learning_rate * (decay_rate**div_res) + div_res = global_step / decay_steps + if staircase: + div_res = ops.floor(div_res) + decayed_lr = learning_rate * (decay_rate**div_res) return decayed_lr @@ -116,41 +136,56 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False): """ global_step = _decay_step_counter() - with init_on_cpu(): - div_res = global_step / decay_steps - if staircase: - div_res = ops.floor(div_res) - decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res) + div_res = global_step / decay_steps + if staircase: + div_res = ops.floor(div_res) + decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res) return decayed_lr def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False): - """Applies inverse time decay to the initial learning rate. + """ + Applies inverse time decay to the initial learning rate. - >>> if staircase: + When training a model, it is often recommended to lower the learning rate as the + training progresses. By using this function, an inverse decay function will be + applied to the initial learning rate. + + >>> if staircase == True: >>> decayed_learning_rate = learning_rate / (1 + decay_rate * floor(global_step / decay_step)) >>> else: >>> decayed_learning_rate = learning_rate / (1 + decay_rate * global_step / decay_step) Args: - learning_rate: A scalar float32 value or a Variable. This - will be the initial learning rate during training. - decay_steps: A Python `int32` number. - decay_rate: A Python `float` number. - staircase: Boolean. If set true, decay the learning rate every decay_steps. + learning_rate(Variable|float): The initial learning rate. + decay_steps(int): See the decay computation above. + decay_rate(float): The decay rate. See the decay computation above. + staircase(Boolean): If True, decay the learning rate at discrete intervals. + Default: False Returns: - The decayed learning rate + Variable: The decayed learning rate + + Examples: + .. code-block:: python + + base_lr = 0.1 + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.inverse_time_decay( + learning_rate=base_lr, + decay_steps=10000, + decay_rate=0.5, + staircase=True)) + sgd_optimizer.minimize(avg_cost) """ global_step = _decay_step_counter() - with init_on_cpu(): - div_res = global_step / decay_steps - if staircase: - div_res = ops.floor(div_res) + div_res = global_step / decay_steps + if staircase: + div_res = ops.floor(div_res) - decayed_lr = learning_rate / (1 + decay_rate * div_res) + decayed_lr = learning_rate / (1 + decay_rate * div_res) return decayed_lr @@ -160,62 +195,74 @@ def polynomial_decay(learning_rate, end_learning_rate=0.0001, power=1.0, cycle=False): - """Applies polynomial decay to the initial learning rate. + """ + Applies polynomial decay to the initial learning rate. + + .. code-block:: python + + if cycle: + decay_steps = decay_steps * ceil(global_step / decay_steps) + else: + global_step = min(global_step, decay_steps) + decayed_learning_rate = (learning_rate - end_learning_rate) * + (1 - global_step / decay_steps) ^ power + end_learning_rate - >>> if cycle: - >>> decay_steps = decay_steps * ceil(global_step / decay_steps) - >>> else: - >>> global_step = min(global_step, decay_steps) - >>> decayed_learning_rate = (learning_rate - end_learning_rate) * - >>> (1 - global_step / decay_steps) ^ power + - >>> end_learning_rate Args: - learning_rate: A scalar float32 value or a Variable. This - will be the initial learning rate during training - decay_steps: A Python `int32` number. - end_learning_rate: A Python `float` number. - power: A Python `float` number - cycle: Boolean. If set true, decay the learning rate every decay_steps. + learning_rate(Variable|float32): A scalar float32 value or a Variable. This + will be the initial learning rate during training. + decay_steps(int32): A Python `int32` number. + end_learning_rate(float): A Python `float` number. + power(float): A Python `float` number. + cycle(bool): If set true, decay the learning rate every decay_steps. Returns: - The decayed learning rate + Variable: The decayed learning rate """ global_step = _decay_step_counter() - with init_on_cpu(): - if cycle: - div_res = ops.ceil(global_step / decay_steps) - zero_var = tensor.fill_constant( - shape=[1], dtype='float32', value=0.0) - one_var = tensor.fill_constant( - shape=[1], dtype='float32', value=1.0) - - with control_flow.Switch() as switch: - with switch.case(global_step == zero_var): - tensor.assign(input=one_var, output=div_res) - decay_steps = decay_steps * div_res - else: - decay_steps_var = tensor.fill_constant( - shape=[1], dtype='float32', value=float(decay_steps)) - global_step = ops.elementwise_min(x=global_step, y=decay_steps_var) + if cycle: + div_res = ops.ceil(global_step / decay_steps) + zero_var = tensor.fill_constant(shape=[1], dtype='float32', value=0.0) + one_var = tensor.fill_constant(shape=[1], dtype='float32', value=1.0) - decayed_lr = (learning_rate - end_learning_rate) * \ - ((1 - global_step / decay_steps) ** power) + end_learning_rate + with control_flow.Switch() as switch: + with switch.case(global_step == zero_var): + tensor.assign(input=one_var, output=div_res) + decay_steps = decay_steps * div_res + else: + decay_steps_var = tensor.fill_constant( + shape=[1], dtype='float32', value=float(decay_steps)) + global_step = ops.elementwise_min(x=global_step, y=decay_steps_var) + + decayed_lr = (learning_rate - end_learning_rate) * \ + ((1 - global_step / decay_steps) ** power) + end_learning_rate return decayed_lr def piecewise_decay(boundaries, values): """Applies piecewise decay to the initial learning rate. - >>> boundaries = [10000, 20000] - >>> values = [1.0, 0.5, 0.1] - >>> - >>> if step < 10000: - >>> learning_rate = 1.0 - >>> elif 10000 <= step < 20000: - >>> learning_rate = 0.5 - >>> else: - >>> learning_rate = 0.1 + The algorithm can be described as the code below. + + .. code-block:: python + + boundaries = [10000, 20000] + values = [1.0, 0.5, 0.1] + if step < 10000: + learning_rate = 1.0 + elif 10000 <= step < 20000: + learning_rate = 0.5 + else: + learning_rate = 0.1 + Args: + boundaries: A list of steps numbers. + values: A list of learning rate values that will be picked during + different step boundaries. + + Returns: + The decayed learning rate. + + """ if len(values) - len(boundaries) != 1: @@ -223,27 +270,65 @@ def piecewise_decay(boundaries, values): global_step = _decay_step_counter() - with init_on_cpu(): - lr = tensor.create_global_var( - shape=[1], - value=0.0, - dtype='float32', - persistable=True, - name="learning_rate") + lr = tensor.create_global_var( + shape=[1], + value=0.0, + dtype='float32', + persistable=True, + name="learning_rate") - with control_flow.Switch() as switch: - for i in range(len(boundaries)): - boundary_val = tensor.fill_constant( - shape=[1], dtype='float32', value=float(boundaries[i])) - value_var = tensor.fill_constant( - shape=[1], dtype='float32', value=float(values[i])) - with switch.case(global_step < boundary_val): - tensor.assign(value_var, lr) - last_value_var = tensor.fill_constant( + with control_flow.Switch() as switch: + for i in range(len(boundaries)): + boundary_val = tensor.fill_constant( shape=[1], dtype='float32', - value=float(values[len(values) - 1])) - with switch.default(): - tensor.assign(last_value_var, lr) + value=float(boundaries[i]), + force_cpu=True) + value_var = tensor.fill_constant( + shape=[1], dtype='float32', value=float(values[i])) + with switch.case(global_step < boundary_val): + tensor.assign(value_var, lr) + last_value_var = tensor.fill_constant( + shape=[1], dtype='float32', value=float(values[len(values) - 1])) + with switch.default(): + tensor.assign(last_value_var, lr) return lr + + +def append_LARS(params_grads, learning_rate, weight_decay): + """Applies LARS (LAYER-WISE ADAPTIVE RATE SCALING) to learning rate for + each layer. + + ```python + learning_rate *= local_gw_ratio * sqrt(sumsq(param)) + / (sqrt(sumsq(gradient))+ weight_decay * sqrt(sumsq(param))) + ``` + + Args: + learning_rate: A learning rate Variable. This + is the global learning rate for LARS. + weight_decay: A Python `float` number. + + Returns: + The decayed learning rate + """ + + def _balanced_weight(param_norm, grad_norm): + if weight_decay == 1.0: + return grad_norm + param_norm + else: + return grad_norm + weight_decay * param_norm + + for param, grad in params_grads: + param_lr = param.optimize_attr['learning_rate'] + param_norm = ops.sqrt(nn.reduce_sum(input=ops.square(param))) + grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad))) + if type(param_lr) == float and param_lr == 1.0: + decayed_lr = learning_rate * param_norm \ + / _balanced_weight(param_norm, grad_norm) + else: + decayed_lr = learning_rate * param_lr * param_norm \ + / _balanced_weight(param_norm, grad_norm) + # set back param local learning rate + param.optimize_attr['learning_rate'] = decayed_lr diff --git a/python/paddle/fluid/layers/math_op_patch.py b/python/paddle/fluid/layers/math_op_patch.py index 1754061c4ba6f5b97bced3548bc412dfb1b7932c..0e10a91d25877984396f9bcf9aae6438707eeab1 100644 --- a/python/paddle/fluid/layers/math_op_patch.py +++ b/python/paddle/fluid/layers/math_op_patch.py @@ -13,11 +13,9 @@ # limitations under the License. from ..framework import Variable, unique_name -from layer_function_generator import OpProtoHolder +from .layer_function_generator import OpProtoHolder from ..initializer import force_init_on_cpu -__all__ = ['monkey_patch_variable'] - def monkey_patch_variable(): def unique_tmp_name(): diff --git a/python/paddle/fluid/layers/metric.py b/python/paddle/fluid/layers/metric.py deleted file mode 100644 index cab2eb55510542bdd4dd7eca7667601697759181..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/layers/metric.py +++ /dev/null @@ -1,81 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -All layers just related to metric. -""" - -import warnings -from ..layer_helper import LayerHelper -from ..initializer import Normal, Constant -from ..framework import Variable -from ..param_attr import ParamAttr -import nn - -__all__ = ['accuracy', 'auc'] - - -def accuracy(input, label, k=1, correct=None, total=None): - """ - This function computes the accuracy using the input and label. - The output is the top k inputs and their indices. - """ - helper = LayerHelper("accuracy", **locals()) - topk_out, topk_indices = nn.topk(input, k=k) - acc_out = helper.create_tmp_variable(dtype="float32") - if correct is None: - correct = helper.create_tmp_variable(dtype="int64") - if total is None: - total = helper.create_tmp_variable(dtype="int64") - helper.append_op( - type="accuracy", - inputs={ - "Out": [topk_out], - "Indices": [topk_indices], - "Label": [label] - }, - outputs={ - "Accuracy": [acc_out], - "Correct": [correct], - "Total": [total], - }) - return acc_out - - -def auc(input, label, curve='ROC', num_thresholds=200): - warnings.warn( - "This interface not recommended, fluid.layers.auc compute the auc at every minibatch, \ - but can not aggregate them and get the pass AUC, because pass \ - auc can not be averaged with weighted from the minibatch auc value. \ - Please use fluid.metrics.Auc, it can compute the auc value via Python natively, \ - which can get every minibatch and every pass auc value.", Warning) - helper = LayerHelper("auc", **locals()) - topk_out = helper.create_tmp_variable(dtype=input.dtype) - topk_indices = helper.create_tmp_variable(dtype="int64") - topk_out, topk_indices = nn.topk(input, k=k) - auc_out = helper.create_tmp_variable(dtype="float32") - if correct is None: - correct = helper.create_tmp_variable(dtype="int64") - if total is None: - total = helper.create_tmp_variable(dtype="int64") - helper.append_op( - type="accuracy", - inputs={ - "Out": [topk_out], - "Indices": [topk_indices], - "Label": [label] - }, - attrs={"curve": curve, - "num_thresholds": num_thresholds}, - outputs={"AUC": [auc_out], }) - return auc_out diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py new file mode 100644 index 0000000000000000000000000000000000000000..49bae1e8af768d93294120e1d13ef0242313aa3c --- /dev/null +++ b/python/paddle/fluid/layers/metric_op.py @@ -0,0 +1,148 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +All layers just related to metric. +""" + +import warnings +from ..layer_helper import LayerHelper +from ..initializer import Normal, Constant +from ..framework import Variable +from ..param_attr import ParamAttr +from . import nn + +__all__ = ['accuracy', 'auc'] + + +def accuracy(input, label, k=1, correct=None, total=None): + """ + accuracy layer. + Refer to the https://en.wikipedia.org/wiki/Precision_and_recall + + This function computes the accuracy using the input and label. + If the correct label occurs in top k predictions, then correct will increment by one. + Note: the dtype of accuracy is determined by input. the input and label dtype can be different. + + Args: + input(Variable): The input of accuracy layer, which is the predictions of network. + Carry LoD information is supported. + label(Variable): The label of dataset. + k(int): The top k predictions for each class will be checked. + correct(Variable): The correct predictions count. + total(Variable): The total entries count. + + Returns: + Variable: The correct rate. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[-1, 32, 32], dtype="float32") + label = fluid.layers.data(name="data", shape=[-1,1], dtype="int32") + predict = fluid.layers.fc(input=data, size=10) + acc = fluid.layers.accuracy(input=predict, label=label, k=5) + + """ + helper = LayerHelper("accuracy", **locals()) + topk_out, topk_indices = nn.topk(input, k=k) + acc_out = helper.create_tmp_variable(dtype="float32") + if correct is None: + correct = helper.create_tmp_variable(dtype="int64") + if total is None: + total = helper.create_tmp_variable(dtype="int64") + helper.append_op( + type="accuracy", + inputs={ + "Out": [topk_out], + "Indices": [topk_indices], + "Label": [label] + }, + outputs={ + "Accuracy": [acc_out], + "Correct": [correct], + "Total": [total], + }) + return acc_out + + +def auc(input, label, curve='ROC', num_thresholds=200, topk=1): + """ + **Area Under the Curve (AUC) Layer** + + This implementation computes the AUC according to forward output and label. + It is used very widely in binary classification evaluation. + + Note: If input label contains values other than 0 and 1, it will be cast + to `bool`. Find the relevant definitions `here `_. + + There are two types of possible curves: + + 1. ROC: Receiver operating characteristic; + 2. PR: Precision Recall + + Args: + input(Variable): A floating-point 2D Variable, values are in the range + [0, 1]. Each row is sorted in descending order. This + input should be the output of topk. Typically, this + Variable indicates the probability of each label. + label(Variable): A 2D int Variable indicating the label of the training + data. The height is batch size and width is always 1. + curve(str): Curve type, can be 'ROC' or 'PR'. Default 'ROC'. + num_thresholds(int): The number of thresholds to use when discretizing + the roc curve. Default 200. + topk(int): only topk number of prediction output will be used for auc. + + Returns: + Variable: A scalar representing the current AUC. + + Examples: + .. code-block:: python + + # network is a binary classification model and label the ground truth + prediction = network(image, is_infer=True) + auc_out=fluid.layers.auc(input=prediction, label=label) + """ + helper = LayerHelper("auc", **locals()) + auc_out = helper.create_tmp_variable(dtype="float64") + # make tp, tn, fp, fn persistable, so that can accumulate all batches. + tp = helper.create_global_variable(persistable=True, dtype='int64') + tn = helper.create_global_variable(persistable=True, dtype='int64') + fp = helper.create_global_variable(persistable=True, dtype='int64') + fn = helper.create_global_variable(persistable=True, dtype='int64') + for var in [tp, tn, fp, fn]: + helper.set_variable_initializer( + var, Constant( + value=0.0, force_cpu=True)) + + helper.append_op( + type="auc", + inputs={ + "Predict": [input], + "Label": [label], + "TP": [tp], + "TN": [tn], + "FP": [fp], + "FN": [fn] + }, + attrs={"curve": curve, + "num_thresholds": num_thresholds}, + outputs={ + "AUC": [auc_out], + "TPOut": [tp], + "TNOut": [tn], + "FPOut": [fp], + "FNOut": [fn] + }) + return auc_out, [tp, tn, fp, fn] diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 63ec83151477770ea64070cae4f5e4fcc497f7af..c75e7eeb4384f28ca0dd95e3b79b7de5a3031351 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -1,4 +1,18 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# Copyright (c ) 2018 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -19,9 +33,12 @@ from ..layer_helper import LayerHelper from ..initializer import Normal, Constant from ..framework import Variable from ..param_attr import ParamAttr -from layer_function_generator import autodoc -from tensor import concat -import utils +from .layer_function_generator import autodoc, templatedoc +from .tensor import concat +from . import utils +import random +from .. import unique_name +from functools import reduce __all__ = [ 'fc', @@ -38,13 +55,16 @@ __all__ = [ 'chunk_eval', 'sequence_conv', 'conv2d', + 'conv3d', 'sequence_pool', 'sequence_softmax', 'softmax', 'pool2d', + 'pool3d', 'batch_norm', 'beam_search_decode', 'conv2d_transpose', + 'conv3d_transpose', 'sequence_expand', 'lstm_unit', 'reduce_sum', @@ -66,6 +86,7 @@ __all__ = [ 'transpose', 'im2sequence', 'nce', + 'hsigmoid', 'beam_search', 'row_conv', 'multiplex', @@ -81,8 +102,17 @@ __all__ = [ 'label_smooth', 'roi_pool', 'dice_loss', - 'upsampling_bilinear2d', + 'image_resize', + 'image_resize_short', + 'resize_bilinear', + 'gather', 'random_crop', + 'mean_iou', + 'relu', + 'log', + 'crop', + 'rank_loss', + 'flatten', ] @@ -91,7 +121,6 @@ def fc(input, num_flatten_dims=1, param_attr=None, bias_attr=None, - use_cudnn=False, use_mkldnn=False, act=None, is_test=False, @@ -99,14 +128,15 @@ def fc(input, """ **Fully Connected Layer** - The fully connected layer can take multiple tensors as its inputs. It - creates a variable called weights for each input tensor, which represents - a fully connected weight matrix from each input unit to each output unit. - The fully connected layer multiplies each input tensor with its coresponding - weight to produce an output Tensor. If multiple input tensors are given, - the results of multiple multiplications will be sumed up. If bias_attr is - not None, a bias variable will be created and added to the output. Finally, - if activation is not None, it will be applied to the output as well. + This function creates a fully connected layer in the network. It can take + multiple tensors as its inputs. It creates a variable called weights for + each input tensor, which represents a fully connected weight matrix from + each input unit to each output unit. The fully connected layer multiplies + each input tensor with its coresponding weight to produce an output Tensor. + If multiple input tensors are given, the results of multiple multiplications + will be sumed up. If bias_attr is not None, a bias variable will be created + and added to the output. Finally, if activation is not None, it will be applied + to the output as well. This process can be formulated as follows: @@ -139,7 +169,8 @@ def fc(input, param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable parameters/weights of this layer. bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias - of this layer. If it is set to None, no bias will be added to the output units. + of this layer. If it is set to False, no bias will be added to the output units. + If it is set to None, the bias is initialized zero. Default: None. act (str, default None): Activation to be applied to the output of this layer. is_test(bool): A flag indicating whether execution is in test phase. use_mkldnn(bool): Use mkldnn kernel or not, it is valid only when the mkldnn @@ -147,7 +178,7 @@ def fc(input, name (str, default None): The name of this layer. Returns: - A tensor variable storing the transformation result. + Variable: The transformation result. Raises: ValueError: If rank of the input tensor is less than 2. @@ -155,8 +186,7 @@ def fc(input, Examples: .. code-block:: python - data = fluid.layers.data( - name="data", shape=[32, 32], dtype="float32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") fc = fluid.layers.fc(input=data, size=1000, act="tanh") """ @@ -188,7 +218,10 @@ def fc(input, else: pre_bias = helper.create_tmp_variable(dtype) helper.append_op( - type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias}) + type="sum", + inputs={"X": mul_results}, + outputs={"Out": pre_bias}, + attrs={"use_mkldnn": use_mkldnn}) # add bias pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims) # add activation @@ -218,10 +251,11 @@ def embedding(input, have two elements which indicate the size of the dictionary of embeddings and the size of each embedding vector respectively. is_sparse(bool): The flag indicating whether to use sparse update. + is_distributed(bool): Whether to run lookup table from remote parameter server. padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup. Otherwise the given :attr:`padding_idx` indicates padding the output with zeros whenever lookup encounters it in :attr:`input`. If - :math:`padding_idx < 0`, the padding_idx to use in lookup is + :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is :math:`size[0] + dim`. param_attr(ParamAttr): Parameters for this layer dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc @@ -257,9 +291,11 @@ def embedding(input, return tmp -# TODO(qijun): expose H0 and C0 +@templatedoc(op_type="lstm") def dynamic_lstm(input, size, + h_0=None, + c_0=None, param_attr=None, bias_attr=None, use_peepholes=True, @@ -270,56 +306,18 @@ def dynamic_lstm(input, dtype='float32', name=None): """ - **Dynamic LSTM Layer** - - The defalut implementation is diagonal/peephole connection - (https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows: - - .. math:: - - i_t & = \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) - - f_t & = \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) - - \\tilde{c_t} & = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) - - o_t & = \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) - - c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t} - - h_t & = o_t \odot act_h(c_t) - - where the :math:`W` terms denote weight matrices (e.g. :math:`W_{xi}` is - the matrix of weights from the input gate to the input), :math:`W_{ic}, \ - W_{fc}, W_{oc}` are diagonal weight matrices for peephole connections. In - our implementation, we use vectors to reprenset these diagonal weight - matrices. The :math:`b` terms denote bias vectors (:math:`b_i` is the input - gate bias vector), :math:`\sigma` is the non-linear activations, such as - logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input - gate, forget gate, output gate, and cell activation vectors, respectively, - all of which have the same size as the cell output activation vector :math:`h`. - - The :math:`\odot` is the element-wise product of the vectors. :math:`act_g` - and :math:`act_h` are the cell input and cell output activation functions - and `tanh` is usually used for them. :math:`\\tilde{c_t}` is also called - candidate hidden state, which is computed based on the current input and - the previous hidden state. - - Set `use_peepholes` to `False` to disable peephole connection. The formula - is omitted here, please refer to the paper - http://www.bioinf.jku.at/publications/older/2604.pdf for details. - - Note that these :math:`W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}` - operations on the input :math:`x_{t}` are NOT included in this operator. - Users can choose to use fully-connect layer before LSTM layer. + ${comment} Args: - input(Variable): The input of dynamic_lstm layer, which supports - variable-time length input sequence. The underlying - tensor in this Variable is a matrix with shape - (T X 4D), where T is the total time steps in this - mini-batch, D is the hidden size. - size(int): 4 * hidden size. + input (Variable): ${input_comment} + size (int): 4 * hidden size. + h_0(Variable): The initial hidden state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size and D is the hidden size. + c_0(Variable): The initial cell state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size. `h_0` and `c_0` can be NULL but only at the same time. + param_attr(ParamAttr|None): The parameter attribute for the learnable hidden-hidden weights. @@ -327,33 +325,26 @@ def dynamic_lstm(input, W_{fh}, W_{oh}`} - The shape is (D x 4D), where D is the hidden size. - bias_attr(ParamAttr|None): The bias attribute for the learnable bias + bias_attr (ParamAttr|None): The bias attribute for the learnable bias weights, which contains two parts, input-hidden bias weights and peephole connections weights if setting `use_peepholes` to `True`. 1. `use_peepholes = False` - - Biases = {:math:`b_c, b_i, b_f, b_o`}. - - The shape is (1 x 4D). + - Biases = {:math:`b_c, b_i, b_f, b_o`}. + - The shape is (1 x 4D). 2. `use_peepholes = True` - - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ + - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ W_{fc}, W_{oc}`}. - - The shape is (1 x 7D). - use_peepholes(bool): Whether to enable diagonal/peephole connections, - default `True`. - is_reverse(bool): Whether to compute reversed LSTM, default `False`. - gate_activation(str): The activation for input gate, forget gate and - output gate. Choices = ["sigmoid", "tanh", "relu", - "identity"], default "sigmoid". - cell_activation(str): The activation for cell output. Choices = ["sigmoid", - "tanh", "relu", "identity"], default "tanh". - candidate_activation(str): The activation for candidate hidden state. - Choices = ["sigmoid", "tanh", - "relu", "identity"], - default "tanh". - dtype(str): Data type. Choices = ["float32", "float64"], default "float32". - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + - The shape is (1 x 7D). + use_peepholes (bool): ${use_peepholes_comment} + is_reverse (bool): ${is_reverse_comment} + gate_activation (str): ${gate_activation_comment} + cell_activation (str): ${cell_activation_comment} + candidate_activation (str): ${candidate_activation_comment} + dtype (str): Data type. Choices = ["float32", "float64"], default "float32". + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: tuple: The hidden state, and cell state of LSTM. The shape of both \ @@ -383,12 +374,20 @@ def dynamic_lstm(input, cell = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype) batch_cell_pre_act = helper.create_tmp_variable(dtype) + inputs = {'Input': input, 'Weight': weight, 'Bias': bias} + batch_size = input.shape[0] + if h_0: + assert h_0.shape == (batch_size, size), \ + 'The shape of h0 should be (batch_size, %d)' % size + inputs['H0'] = h_0 + if c_0: + assert c_0.shape == (batch_size, size), \ + 'The shape of c0 should be (batch_size, %d)' % size + inputs['C0'] = c_0 helper.append_op( type='lstm', - inputs={'Input': input, - 'Weight': weight, - 'Bias': bias}, + inputs=inputs, outputs={ 'Hidden': hidden, 'Cell': cell, @@ -516,27 +515,31 @@ def dynamic_lstmp(input, cell_activation(str): The activation for cell output. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". candidate_activation(str): The activation for candidate hidden state. - Choices = ["sigmoid", "tanh", - "relu", "identity"], + Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". proj_activation(str): The activation for projection output. - Choices = ["sigmoid", "tanh", - "relu", "identity"], + Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". dtype(str): Data type. Choices = ["float32", "float64"], default "float32". name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. Returns: - tuple: The projection of hidden state, and cell state of LSTMP. The \ - shape of projection is (T x P), for the cell state which is \ - (T x D), and both LoD is the same with the `input`. + tuple: A tuple of two output variable: the projection of hidden state, \ + and cell state of LSTMP. The shape of projection is (T x P), \ + for the cell state which is (T x D), and both LoD is the same \ + with the `input`. Examples: + .. code-block:: python + dict_dim, emb_dim = 128, 64 + data = fluid.layers.data(name='sequence', shape=[1], + dtype='int32', lod_level=1) + emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim]) hidden_dim, proj_dim = 512, 256 - fc_out = fluid.layers.fc(input=input_seq, size=hidden_dim * 4, + fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4, act=None, bias_attr=None) proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out, size=hidden_dim * 4, @@ -602,10 +605,10 @@ def dynamic_gru(input, candidate_activation='tanh', h_0=None): """ - **Dynamic GRU Layer** + **Gated Recurrent Unit (GRU) Layer** Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on - Sequence Modeling `_ + Sequence Modeling `_ . The formula is as follows: @@ -650,18 +653,27 @@ def dynamic_gru(input, :attr:`False`. gate_activation(str): The activation for update gate and reset gate. Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid". - activation(str): The activation for candidate hidden state. + candidate_activation(str): The activation for candidate hidden state. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". + h_0 (Variable): This is initial hidden state. If not set, default is + zero. This is a tensor with shape (N x D), where N is the number of + total time steps of input mini-batch feature and D is the hidden + size. Returns: Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \ - and lod is the same with the input. + and sequence length is the same with the input. Examples: + .. code-block:: python + dict_dim, emb_dim = 128, 64 + data = fluid.layers.data(name='sequence', shape=[1], + dtype='int32', lod_level=1) + emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim]) hidden_dim = 512 - x = fluid.layers.fc(input=data, size=hidden_dim * 3) + x = fluid.layers.fc(input=emb, size=hidden_dim * 3) hidden = fluid.layers.dynamic_gru(input=x, dim=hidden_dim) """ @@ -672,11 +684,13 @@ def dynamic_gru(input, attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) bias = helper.create_parameter( attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True) + batch_size = input.shape[0] inputs = {'Input': input, 'Weight': weight, 'Bias': bias} if h_0 != None: assert h_0.shape == ( - size, size), 'The shape of h0 should be(%d, %d)' % (size, size) - inputs['h0'] = h_0 + batch_size, size + ), 'The shape of h0 should be(batch_size, %d)' % size + inputs['H0'] = h_0 hidden = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype) @@ -798,7 +812,25 @@ def gru_unit(input, return updated_hidden, reset_hidden_pre, gate +@templatedoc() def linear_chain_crf(input, label, param_attr=None): + """ + Linear Chain CRF. + + ${comment} + + Args: + input(${emission_type}): ${emission_comment} + input(${transition_type}): ${transition_comment} + label(${label_type}): ${label_comment} + param_attr(ParamAttr): The attribute of the learnable parameter. + + Returns: + output(${emission_exps_type}): ${emission_exps_comment} \n + output(${transition_exps_type}): ${transition_exps_comment} \n + output(${log_likelihood_type}): ${log_likelihood_comment} + + """ helper = LayerHelper('linear_chain_crf', **locals()) size = input.shape[1] transition = helper.create_parameter( @@ -824,7 +856,27 @@ def linear_chain_crf(input, label, param_attr=None): return log_likelihood +@templatedoc() def crf_decoding(input, param_attr, label=None): + """ + ${comment} + + Args: + input(${emission_type}): ${emission_comment} + + param_attr(ParamAttr): The parameter attribute for training. + + label(${label_type}): ${label_comment} + + Returns: + Variable: ${viterbi_path_comment} + + Examples: + .. code-block:: python + + crf_decode = layers.crf_decoding( + input=hidden, param_attr=ParamAttr(name="crfw")) + """ helper = LayerHelper('crf_decoding', **locals()) transition = helper.get_parameter(param_attr.name) viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype()) @@ -838,10 +890,17 @@ def crf_decoding(input, param_attr, label=None): return viterbi_path +@templatedoc() def cos_sim(X, Y): """ - This function performs the cosine similarity between two tensors - X and Y and returns that as the output. + ${comment} + + Args: + X (Variable): ${x_comment}. + Y (Variable): ${y_comment}. + + Returns: + Variable: the output of cosine(X, Y). """ helper = LayerHelper('cos_sim', **locals()) out = helper.create_tmp_variable(dtype=X.dtype) @@ -863,34 +922,39 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): Drop or keep each element of `x` independently. Dropout is a regularization technique for reducing overfitting by preventing neuron co-adaption during - training. The dropout operator randomly set (according to the given dropout + training. The dropout operator randomly sets (according to the given dropout probability) the outputs of some units to zero, while others are remain unchanged. Args: - x(variable): The input tensor. - dropout_prob(float): Probability of setting units to zero. - is_test(bool): A flag indicating whether it is in test phrase or not. - seed(int): A Python integer used to create random seeds. If this - parameter is set to None, a random seed is used. - NOTE: If an integer seed is given, always the same output - units will be dropped. DO NOT use a fixed seed in training. - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + x (Variable): The input tensor variable. + dropout_prob (float): Probability of setting units to zero. + is_test (bool): A flag indicating whether it is in test phrase or not. + seed (int): A Python integer used to create random seeds. If this + parameter is set to None, a random seed is used. + NOTE: If an integer seed is given, always the same output + units will be dropped. DO NOT use a fixed seed in training. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: - Variable: A tensor variable. + Variable: A tensor variable is the shape with `x`. Examples: + .. code-block:: python - x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") - droped = fluid.layers.dropout(input=x, dropout_rate=0.5) + x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + droped = fluid.layers.dropout(x, dropout_prob=0.5) """ helper = LayerHelper('dropout', **locals()) out = helper.create_tmp_variable(dtype=x.dtype) mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True) + + if (seed is None or seed == 0) and helper.main_program.random_seed != 0: + seed = helper.main_program.random_seed + helper.append_op( type='dropout', inputs={'X': [x]}, @@ -998,8 +1062,8 @@ def square_error_cost(input, label): * :math:`Out`: Output value, same shape with :math:`X`. Args: - input(Variable): Input tensor, has predictions. - label(Variable): Label tensor, has target labels. + input (Variable): Input tensor, has predictions. + label (Variable): Label tensor, has target labels. Returns: Variable: The tensor variable storing the element-wise squared error \ @@ -1028,14 +1092,101 @@ def square_error_cost(input, label): return square_out +@templatedoc() def chunk_eval(input, label, chunk_scheme, num_chunk_types, excluded_chunk_types=None): """ + **Chunk Evaluator** + This function computes and outputs the precision, recall and F1-score of chunk detection. + + For some basics of chunking, please refer to + 'Chunking with Support Vector Machines '. + + ChunkEvalOp computes the precision, recall, and F1-score of chunk detection, + and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes. + Here is a NER example of labeling for these tagging schemes: + + .. code-block:: python + + ====== ====== ====== ===== == ============ ===== ===== ===== == ========= + Li Ming works at Agricultural Bank of China in Beijing. + ====== ====== ====== ===== == ============ ===== ===== ===== == ========= + IO I-PER I-PER O O I-ORG I-ORG I-ORG I-ORG O I-LOC + IOB B-PER I-PER O O B-ORG I-ORG I-ORG I-ORG O B-LOC + IOE I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC + IOBES B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC + ====== ====== ====== ===== == ============ ===== ===== ===== == ========= + + There are three chunk types(named entity types) including PER(person), ORG(organization) + and LOC(LOCATION), and we can see that the labels have the form -. + + Since the calculations actually use label ids rather than labels, extra attention + should be paid when mapping labels to ids to make CheckEvalOp work. The key point + is that the listed equations are satisfied by ids. + + .. code-block:: python + + tag_type = label % num_tag_type + chunk_type = label / num_tag_type + + where `num_tag_type` is the num of tag types in the tagging scheme, `num_chunk_type` + is the num of chunk types, and `tag_type` get its value from the following table. + + .. code-block:: python + + Scheme Begin Inside End Single + plain 0 - - - + IOB 0 1 - - + IOE - 0 1 - + IOBES 0 1 2 3 + + Still use NER as example, assuming the tagging scheme is IOB while chunk types are ORG, + PER and LOC. To satisfy the above equations, the label map can be like this: + + .. code-block:: python + + B-ORG 0 + I-ORG 1 + B-PER 2 + I-PER 3 + B-LOC 4 + I-LOC 5 + O 6 + + It's not hard to verify the equations noting that the num of chunk types + is 3 and the num of tag types in IOB scheme is 2. For example, the label + id of I-LOC is 5, the tag type id of I-LOC is 1, and the chunk type id of + I-LOC is 2, which consistent with the results from the equations. + + Args: + input (Variable): prediction output of the network. + label (Variable): label of the test data set. + chunk_scheme (str): ${chunk_scheme_comment} + num_chunk_types (int): ${num_chunk_types_comment} + excluded_chunk_types (list): ${excluded_chunk_types_comment} + + Returns: + tuple: tuple containing: precision, recall, f1_score, + num_infer_chunks, num_label_chunks, + num_correct_chunks + + Examples: + .. code-block:: python + + crf = fluid.layers.linear_chain_crf( + input=hidden, label=label, param_attr=ParamAttr(name="crfw")) + crf_decode = fluid.layers.crf_decoding( + input=hidden, param_attr=ParamAttr(name="crfw")) + fluid.layers.chunk_eval( + input=crf_decode, + label=label, + chunk_scheme="IOB", + num_chunk_types=(label_dict_len - 1) / 2) """ helper = LayerHelper("chunk_eval", **locals()) @@ -1068,6 +1219,7 @@ def chunk_eval(input, num_correct_chunks) +@templatedoc() def sequence_conv(input, num_filters, filter_size=3, @@ -1080,11 +1232,20 @@ def sequence_conv(input, This function creates the op for sequence_conv, using the inputs and other convolutional configurations for the filters and stride as given in the input parameters to the function. - """ - # FIXME(dzh) : want to unify the argument of python layer - # function. So we ignore some unecessary attributes. - # such as, padding_trainable, context_start. + Args: + input (Variable): ${x_comment} + num_filters (int): number of filters. + filter_size (int): the filter size (H and W). + filter_stride (int): stride of the filter. + padding (bool): if True, add paddings. + bias_attr (ParamAttr|None): attributes for bias + param_attr (ParamAttr|None): attributes for parameter + act (str): the activation type + + Returns: + Variable: output of sequence_conv + """ helper = LayerHelper('sequence_conv', **locals()) dtype = helper.input_dtype() @@ -1110,6 +1271,41 @@ def sequence_conv(input, def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): + """ + This function computes the softmax activation among all time-steps for each + sequence. The dimension of each time-step should be 1. Thus, the shape of + input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N` + is the sum of the length of all sequences. + + For i-th sequence in a mini-batch: + + .. math:: + + Out(X[lod[i]:lod[i+1]], :) = \\frac{\exp(X[lod[i]:lod[i+1], :])}{\sum(\exp(X[lod[i]:lod[i+1], :]))} + + For example, for a mini-batch of 3 sequences with variable-length, + each containing 2, 3, 2 time-steps, the lod of which is [0, 2, 5, 7], + then softmax will be computed among :math:`X[0:2, :]`, :math:`X[2:5, :]`, + :math:`X[5:7, :]`, and :math:`N` turns out to be 7. + + Args: + input (Variable): The input variable which is a LoDTensor. + bias_attr (ParamAttr|None): attributes for bias + param_attr (ParamAttr|None): attributes for parameter + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ + library is installed. Default: True + + Returns: + Variable: output of sequence_softmax + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[7, 1], + dtype='float32', lod_level=1) + x_sequence_softmax = fluid.layers.sequence_softmax(input=x) + """ helper = LayerHelper('sequence_softmax', **locals()) dtype = helper.input_dtype() softmax_out = helper.create_tmp_variable(dtype) @@ -1122,6 +1318,48 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): + """ + The input of the softmax operator is a tensor of any rank. The output tensor + has the same shape as the input. + + The input tensor will first be logically flattened to a 2-D matrix. The matrix's + second dimension(row length) is as same as the last dimension of the input + tensor, and the first dimension(column length) is the product of all other + dimensions of the input tensor. For each row of the matrix, the softmax operator + squashes the K-dimensional(K is the width of the matrix, which is also the size + of the input tensor's last dimension) vector of arbitrary real values to a + K-dimensional vector of real values in the range [0, 1] that add up to 1. + + It computes the exponential of the given dimension and the sum of exponential + values of all the other dimensions in the K-dimensional vector input. + Then the ratio of the exponential of the given dimension and the sum of + exponential values of all the other dimensions is the output of the softmax + operator. + + For each row :math:`i` and each column :math:`j` in the matrix, we have: + + .. math:: + + Out[i, j] = \\frac{\exp(X[i, j])}{\sum_j(exp(X[i, j])} + + Args: + input (Variable): The input variable. + bias_attr (ParamAttr): attributes for bias + param_attr (ParamAttr): attributes for parameter + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ + library is installed. + + Returns: + Variable: output of softmax + + Examples: + + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10) + softmax = fluid.layers.softmax(input=fc) + + """ helper = LayerHelper('softmax', **locals()) dtype = helper.input_dtype() softmax_out = helper.create_tmp_variable(dtype) @@ -1147,14 +1385,17 @@ def conv2d(input, act=None, name=None): """ - **Convlution2D Layer** - The convolution2D layer calculates the output based on the input, filter - and strides, paddings, dilations, groups parameters. Input(Input) and - Output(Output) are in NCHW format. Where N is batch size, C is the number of + and strides, paddings, dilations, groups parameters. Input and + Output are in NCHW format, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature. - The details of convolution layer, please refer UFLDL's `convolution, - `_ . + Filter is in MCHW format, where M is the number of output image channels, + C is the number of input image channels, H is the height of the filter, + and W is the width of the filter. If the groups is greater than 1, + C will equal the number of input image channels divided by the groups. + Please refer to UFLDL's `convolution + `_ + for more detials. If bias attribution and activation type are provided, bias is added to the output of the convolution, and the corresponding activation function is applied to the final result. @@ -1165,62 +1406,64 @@ def conv2d(input, Out = \sigma (W \\ast X + b) - In the above equation: + Where: * :math:`X`: Input value, a tensor with NCHW format. * :math:`W`: Filter value, a tensor with MCHW format. * :math:`\\ast`: Convolution operation. * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. * :math:`\\sigma`: Activation function. - * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be - different. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. Example: - Input: - Input shape: $(N, C_{in}, H_{in}, W_{in})$ + Input shape: :math:`(N, C_{in}, H_{in}, W_{in})` - Filter shape: $(C_{out}, C_{in}, H_f, W_f)$ + Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)` - Output: - Output shape: $(N, C_{out}, H_{out}, W_{out})$ + + Output shape: :math:`(N, C_{out}, H_{out}, W_{out})` Where .. math:: - H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\ - W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1 - - Args: - input(Variable): The input image with [N, C, H, W] format. - num_filters(int): The number of filter. It is as same as the output - image channel. - filter_size(int|tuple|None): The filter size. If filter_size is a tuple, - it must contain two integers, (filter_size_H, filter_size_W). - Otherwise, the filter will be a square. - stride(int|tuple): The stride size. If stride is a tuple, it must - contain two integers, (stride_H, stride_W). Otherwise, the - stride_H = stride_W = stride. Default: stride = 1. - padding(int|tuple): The padding size. If padding is a tuple, it must - contain two integers, (padding_H, padding_W). Otherwise, the - padding_H = padding_W = padding. Default: padding = 0. - dilation(int|tuple): The dilation size. If dilation is a tuple, it must - contain two integers, (dilation_H, dilation_W). Otherwise, the - dilation_H = dilation_W = dilation. Default: dilation = 1. - groups(int): The groups number of the Conv2d Layer. According to grouped - convolution in Alex Krizhevsky's Deep CNN paper: when group=2, - the first half of the filters is only connected to the first half - of the input channels, while the second half of the filters is only - connected to the second half of the input channels. Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None - bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None - use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn - library is installed. Default: True - act(str): Activation type. Default: None - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\ + W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1 + + Args: + input (Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of filter. It is as same as the output + image channel. + filter_size (int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. + stride (int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. Default: stride = 1. + padding (int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. Default: padding = 0. + dilation (int|tuple): The dilation size. If dilation is a tuple, it must + contain two integers, (dilation_H, dilation_W). Otherwise, the + dilation_H = dilation_W = dilation. Default: dilation = 1. + groups (int): The groups number of the Conv2d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: when group=2, + the first half of the filters is only connected to the first half + of the input channels, while the second half of the filters is only + connected to the second half of the input channels. Default: groups=1 + param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None + bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled + with mkldnn library. Default: False + act (str): Activation type. Default: None + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The tensor variable storing the convolution and \ @@ -1233,13 +1476,9 @@ def conv2d(input, Examples: .. code-block:: python - data = fluid.layers.data( - name='data', shape=[3, 32, 32], dtype='float32') - conv2d = fluid.layers.conv2d( - input=data, num_filters=2, filter_size=3, act="relu") + data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32') + conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu") """ - if stride is None: - stride = [1, 1] num_channels = input.shape[1] @@ -1302,68 +1541,230 @@ def conv2d(input, return helper.append_activation(pre_act) -def sequence_pool(input, pool_type): +def conv3d(input, + num_filters, + filter_size, + stride=1, + padding=0, + dilation=1, + groups=None, + param_attr=None, + bias_attr=None, + use_cudnn=True, + use_mkldnn=False, + act=None, + name=None): """ - This function add the operator for sequence pooling. - It pools features of all time-steps of each instance, and is applied - on top of the input using pool_type mentioned in the parameters. + **Convlution3D Layer** - It supports four pool_type: + The convolution3D layer calculates the output based on the input, filter + and strides, paddings, dilations, groups parameters. Input(Input) and + Output(Output) are in NCDHW format. Where N is batch size C is the number of + channels, D is the depth of the feature, H is the height of the feature, + and W is the width of the feature. Convlution3D is similar with Convlution2D + but adds one dimension(depth). If bias attribution and activation type are + provided, bias is added to the output of the convolution, and the + corresponding activation function is applied to the final result. - - average: :math:`Out[i] = \\frac{\sum_i X_i}{N}` - - sum: :math:`Out[i] = \sum_jX_{ij}` - - sqrt: :math:`Out[i] = \\frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}` - - max: :math:`Out[i] = max(X_i)` + For each input :math:`X`, the equation is: - .. code-block:: text + .. math:: - x is a 1-level LoDTensor: - x.lod = [[0, 2, 5, 7]] - x.data = [1, 3, 2, 4, 6, 5, 1] - x.dims = [7, 1] + Out = \sigma (W \\ast X + b) - then output is a Tensor: - out.dim = [3, 1] - with condition len(x.lod[-1]) - 1 == out.dims[0] + In the above equation: - for different pool_type: - average: out.data = [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2 - sum : out.data = [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1 - sqrt : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), - 6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2) - max : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1) - last : out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) - first : out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1) + * :math:`X`: Input value, a tensor with NCDHW format. + * :math:`W`: Filter value, a tensor with MCDHW format. + * :math:`\\ast`: Convolution operation. + * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. + * :math:`\\sigma`: Activation function. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. + + Example: + + - Input: + + Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})` + + Filter shape: :math:`(C_{out}, C_{in}, D_f, H_f, W_f)` + + - Output: + Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})` + + Where + + .. math:: + + D_{out}&= \\frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{strides[0]} + 1 \\\\ + H_{out}&= \\frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{strides[1]} + 1 \\\\ + W_{out}&= \\frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{strides[2]} + 1 Args: - input(variable): The input variable which is a LoDTensor. - pool_type (string): The pooling type of sequence_pool. - It supports average, sum, sqrt and max. + input (Variable): The input image with [N, C, D, H, W] format. + num_filters(int): The number of filter. It is as same as the output + image channel. + filter_size (int|tuple|None): The filter size. If filter_size is a tuple, + it must contain three integers, (filter_size_D, filter_size_H, filter_size_W). + Otherwise, the filter will be a square. + stride (int|tuple): The stride size. If stride is a tuple, it must + contain three integers, (stride_D, stride_H, stride_W). Otherwise, the + stride_D = stride_H = stride_W = stride. Default: stride = 1. + padding (int|tuple): The padding size. If padding is a tuple, it must + contain three integers, (padding_D, padding_H, padding_W). Otherwise, the + padding_D = padding_H = padding_W = padding. Default: padding = 0. + dilation (int|tuple): The dilation size. If dilation is a tuple, it must + contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the + dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1. + groups (int): The groups number of the Conv3d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: when group=2, + the first half of the filters is only connected to the first half + of the input channels, while the second half of the filters is only + connected to the second half of the input channels. Default: groups=1 + param_attr (ParamAttr): The parameters to the Conv3d Layer. Default: None + bias_attr (ParamAttr): Bias parameter for the Conv3d layer. Default: None + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not. + act (str): Activation type. Default: None + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: - The sequence pooling variable which is a Tensor. + Variable: The tensor variable storing the convolution and \ + non-linearity activation result. - Examples: + Raises: + ValueError: If the shapes of input, filter_size, stride, padding and + groups mismatch. + Examples: .. code-block:: python - x = fluid.layers.data(name='x', shape=[7, 1], - dtype='float32', lod_level=1) - avg_x = fluid.layers.sequence_pool(input=x, pool_type='average') - sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum') - sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt') - max_x = fluid.layers.sequence_pool(input=x, pool_type='max') - last_x = fluid.layers.sequence_pool(input=x, pool_type='last') - first_x = fluid.layers.sequence_pool(input=x, pool_type='first') + data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32') + conv3d = fluid.layers.conv3d(input=data, num_filters=2, filter_size=3, act="relu") """ - helper = LayerHelper('sequence_pool', **locals()) + + l_type = 'conv3d' + + helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) - max_index = helper.create_tmp_variable(dtype) - helper.append_op( - type="sequence_pool", - inputs={"X": input}, + num_channels = input.shape[1] + + if groups is None: + num_filter_channels = num_channels + else: + if num_channels % groups != 0: + raise ValueError("num_channels must be divisible by groups.") + num_filter_channels = num_channels / groups + + filter_size = utils.convert_to_list(filter_size, 3, 'filter_size') + stride = utils.convert_to_list(stride, 3, 'stride') + padding = utils.convert_to_list(padding, 3, 'padding') + dilation = utils.convert_to_list(dilation, 3, 'dilation') + + if not isinstance(use_cudnn, bool): + raise ValueError("use_cudnn should be True or False") + + input_shape = input.shape + filter_shape = [num_filters, num_filter_channels] + filter_size + + def _get_default_param_initializer(): + std = (2.0 / (filter_size[0]**3 * num_channels))**0.5 + return Normal(0.0, std, 0) + + filter_param = helper.create_parameter( + attr=helper.param_attr, + shape=filter_shape, + dtype=dtype, + default_initializer=_get_default_param_initializer()) + + pre_bias = helper.create_tmp_variable(dtype) + + helper.append_op( + type=l_type, + inputs={ + 'Input': input, + 'Filter': filter_param, + }, + outputs={"Output": pre_bias}, + attrs={ + 'strides': stride, + 'paddings': padding, + 'dilations': dilation, + 'groups': groups, + 'use_cudnn': use_cudnn, + 'use_mkldnn': use_mkldnn + }) + + pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) + + return helper.append_activation(pre_act) + + +def sequence_pool(input, pool_type): + """ + This function add the operator for sequence pooling. + It pools features of all time-steps of each instance, and is applied + on top of the input using pool_type mentioned in the parameters. + + It supports four pool_type: + + - average: :math:`Out[i] = \\frac{\sum_i X_i}{N}` + - sum: :math:`Out[i] = \sum_jX_{ij}` + - sqrt: :math:`Out[i] = \\frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}` + - max: :math:`Out[i] = max(X_i)` + + .. code-block:: text + + x is a 1-level LoDTensor: + x.lod = [[2, 3, 2]] + x.data = [1, 3, 2, 4, 6, 5, 1] + x.dims = [7, 1] + + then output is a Tensor: + out.dim = [3, 1] + with condition len(x.lod[-1]) == out.dims[0] + + for different pool_type: + average: out.data = [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2 + sum : out.data = [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1 + sqrt : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), + 6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2) + max : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1) + last : out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) + first : out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1) + + Args: + input(variable): The input variable which is a LoDTensor. + pool_type (string): The pooling type of sequence_pool. + It supports average, sum, sqrt and max. + + Returns: + The sequence pooling variable which is a Tensor. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[7, 1], + dtype='float32', lod_level=1) + avg_x = fluid.layers.sequence_pool(input=x, pool_type='average') + sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum') + sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt') + max_x = fluid.layers.sequence_pool(input=x, pool_type='max') + last_x = fluid.layers.sequence_pool(input=x, pool_type='last') + first_x = fluid.layers.sequence_pool(input=x, pool_type='first') + """ + helper = LayerHelper('sequence_pool', **locals()) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + max_index = helper.create_tmp_variable(dtype) + + helper.append_op( + type="sequence_pool", + inputs={"X": input}, outputs={"Out": pool_out, "MaxIndex": max_index}, attrs={"pooltype": pool_type.upper()}) @@ -1378,18 +1779,18 @@ def sequence_pool(input, pool_type): def sequence_first_step(input): """ - This funciton get the first step of sequence. + This function gets the first step of sequence. .. code-block:: text x is a 1-level LoDTensor: - x.lod = [[0, 2, 5, 7]] + x.lod = [[2, 3, 2]] x.data = [1, 3, 2, 4, 6, 5, 1] x.dims = [7, 1] then output is a Tensor: out.dim = [3, 1] - with condition len(x.lod[-1]) - 1 == out.dims[0] + with condition len(x.lod[-1]) == out.dims[0] out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1) Args: @@ -1411,18 +1812,18 @@ def sequence_first_step(input): def sequence_last_step(input): """ - This funciton get the last step of sequence. + This function gets the last step of sequence. .. code-block:: text x is a 1-level LoDTensor: - x.lod = [[0, 2, 5, 7]] + x.lod = [[2, 3, 2]] x.data = [1, 3, 2, 4, 6, 5, 1] x.dims = [7, 1] then output is a Tensor: out.dim = [3, 1] - with condition len(x.lod[-1]) - 1 == out.dims[0] + with condition len(x.lod[-1]) == out.dims[0] out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) Args: @@ -1442,6 +1843,7 @@ def sequence_last_step(input): return sequence_pool(input=input, pool_type="last") +@templatedoc() def pool2d(input, pool_size=-1, pool_type="max", @@ -1453,8 +1855,45 @@ def pool2d(input, use_mkldnn=False, name=None): """ - This function adds the operator for pooling in 2 dimensions, using the - pooling configurations mentioned in input parameters. + ${comment} + + Args: + input (Variable): The input tensor of pooling operator. The format of + input tensor is NCHW, where N is batch size, C is + the number of channels, H is the height of the + feature, and W is the width of the feature. + pool_size (int): The side length of pooling windows. All pooling + windows are squares with pool_size on a side. + pool_type: ${pooling_type_comment} + pool_stride (int): stride of the pooling layer. + pool_padding (int): padding size. + global_pooling: ${global_pooling_comment} + use_cudnn: ${use_cudnn_comment} + ceil_mode: ${ceil_mode_comment} + use_mkldnn: ${use_mkldnn_comment} + name (str|None): A name for this layer(optional). If set None, the + layer will be named automatically. + + Returns: + Variable: The pooling result. + + Raises: + ValueError: If 'pool_type' is not "max" nor "avg" + ValueError: If 'global_pooling' is False and 'pool_size' is -1 + ValueError: If 'use_cudnn' is not a bool value. + + Examples: + + .. code-block:: python + + data = fluid.layers.data( + name='data', shape=[3, 32, 32], dtype='float32') + conv2d = fluid.layers.pool2d( + input=data, + pool_size=2, + pool_type='max', + pool_stride=1, + global_pooling=False) """ if pool_type not in ["max", "avg"]: raise ValueError( @@ -1473,12 +1912,84 @@ def pool2d(input, if not isinstance(use_cudnn, bool): raise ValueError("use_cudnn should be True or False") - helper = LayerHelper('pool2d', **locals()) + l_type = 'pool2d' + + helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() pool_out = helper.create_tmp_variable(dtype) helper.append_op( - type="pool2d", + type=l_type, + inputs={"X": input}, + outputs={"Out": pool_out}, + attrs={ + "pooling_type": pool_type, + "ksize": pool_size, + "global_pooling": global_pooling, + "strides": pool_stride, + "paddings": pool_padding, + "use_cudnn": use_cudnn, + "ceil_mode": ceil_mode, + "use_mkldnn": use_mkldnn + }) + + return pool_out + + +def pool3d(input, + pool_size=-1, + pool_type="max", + pool_stride=1, + pool_padding=0, + global_pooling=False, + use_cudnn=True, + ceil_mode=False, + use_mkldnn=False, + name=None): + """ + This function adds the operator for pooling in 3-dimensions, using the + pooling configurations mentioned in input parameters. + + Args: + input (Variable): ${input_comment} + pool_size (int): ${ksize_comment} + pool_type (str): ${pooling_type_comment} + pool_stride (int): stride of the pooling layer. + pool_padding (int): padding size. + global_pooling (bool): ${global_pooling_comment} + use_cudnn (bool): ${use_cudnn_comment} + ceil_mode (bool): ${ceil_mode_comment} + use_mkldnn (bool): ${use_mkldnn_comment} + name (str): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: output of pool3d layer. + """ + if pool_type not in ["max", "avg"]: + raise ValueError( + "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.", + str(pool_type)) + + if global_pooling is False and pool_size == -1: + raise ValueError( + "When the global_pooling is False, pool_size must be passed " + "and be a valid value. Received pool_size: " + str(pool_size)) + + pool_size = utils.convert_to_list(pool_size, 3, 'pool_size') + pool_padding = utils.convert_to_list(pool_padding, 3, 'pool_padding') + pool_stride = utils.convert_to_list(pool_stride, 3, 'pool_stride') + + if not isinstance(use_cudnn, bool): + raise ValueError("use_cudnn should be True or False") + + l_type = "pool3d" + helper = LayerHelper(l_type, **locals()) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + + helper.append_op( + type=l_type, inputs={"X": input}, outputs={"Out": pool_out}, attrs={ @@ -1508,10 +2019,61 @@ def batch_norm(input, name=None, moving_mean_name=None, moving_variance_name=None, - do_model_average_for_mean_and_var=False): + do_model_average_for_mean_and_var=False, + fuse_with_relu=False): """ - This function helps create an operator to implement - the BatchNorm layer using the configurations from the input parameters. + **Batch Normalization Layer** + + Can be used as a normalizer function for conv2d and fully_connected operations. + The required data format for this layer is one of the following: + + 1. NHWC `[batch, in_height, in_width, in_channels]` + + 2. NCHW `[batch, in_channels, in_height, in_width]` + + Refer to `Batch Normalization: Accelerating Deep Network Training by Reducing + Internal Covariate Shift `_ + for more details. + + :math:`input` is the input features over a mini-batch. + + .. math:: + + \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\ + \ mini-batch\ mean \\\\ + \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\ + \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\ + \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\ + \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\ + y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift + + Args: + input(variable): The input variable which is a LoDTensor. + act(string, Default None): Activation type, linear|relu|prelu|... + is_test(bool, Default False): Used for training or training. + momentum(float, Default 0.9): + epsilon(float, Default 1e-05): + param_attr(ParamAttr): The parameter attribute for Parameter `scale`. + bias_attr(ParamAttr): The parameter attribute for Parameter `bias`. + data_layout(string, default NCHW): NCHW|NHWC + in_place(bool, Default False): Make the input and output of batch norm reuse memory. + use_mkldnn(bool, Default false): ${use_mkldnn_comment} + name(string, Default None): A name for this layer(optional). If set None, the layer + will be named automatically. + moving_mean_name(string, Default None): The name of moving_mean which store the global Mean. + moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance. + do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not. + fuse_with_relu (bool): if True, this OP performs relu after batch norm. + + Returns: + Variable: A tensor variable which is the result after applying batch normalization on the input. + + Examples: + + .. code-block:: python + + hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w') + hidden2 = fluid.layers.batch_norm(input=hidden1) """ helper = LayerHelper('batch_norm', **locals()) dtype = helper.input_dtype() @@ -1587,12 +2149,14 @@ def batch_norm(input, "momentum": momentum, "epsilon": epsilon, "is_test": is_test, - "use_mkldnn": use_mkldnn + "use_mkldnn": use_mkldnn, + "fuse_with_relu": fuse_with_relu }) return helper.append_activation(batch_norm_out) +@templatedoc() def layer_norm(input, scale=True, shift=True, @@ -1603,20 +2167,11 @@ def layer_norm(input, act=None, name=None): """ - **Layer Normalization** - - Assume feature vectors exist on dimensions - :attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics - along these dimensions for each feature vector :math:`a` with size - :math:`H`, then normalize each feature vector using the corresponding - statistics. After that, apply learnable gain and bias on the normalized - tensor to scale and shift if :attr:`scale` and :attr:`shift` are set. - - Refer to `Layer Normalization `_ + ${comment} The formula is as follows: - .. math:: + .. math:: \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} a_i @@ -1624,6 +2179,15 @@ def layer_norm(input, h & = f(\\frac{g}{\\sigma}(a - \\mu) + b) + * :math:`a`: the vector representation of the summed inputs to the neurons + in that layer. + + * :math:`H`: the number of hidden units in a layers + + * :math:`g`: the trainable scale parameter. + + * :math:`b`: the trainable bias parameter. + Args: input(Variable): The input tensor variable. scale(bool): Whether to learn the adaptive gain :math:`g` after @@ -1639,16 +2203,16 @@ def layer_norm(input, bias_attr(ParamAttr|None): The parameter attribute for the learnable bias :math:`b`. act(str): Activation to be applied to the output of layer normalizaiton. + name (str): The name of this layer. It is optional. Returns: - Variable: A tensor variable with the same shape as the input. + ${y_comment} Examples: - .. code-block:: python - data = fluid.layers.data( - name='data', shape=[3, 32, 32], dtype='float32') - x = fluid.layers.layer_norm(input=data, begin_norm_axis=1) + >>> data = fluid.layers.data(name='data', shape=[3, 32, 32], + >>> dtype='float32') + >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1) """ helper = LayerHelper('layer_norm', **locals()) dtype = helper.input_dtype() @@ -1689,23 +2253,6 @@ def layer_norm(input, return helper.append_activation(layer_norm_out) -def beam_search_decode(ids, scores, name=None): - helper = LayerHelper('beam_search_decode', **locals()) - sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) - sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) - - helper.append_op( - type="beam_search_decode", - inputs={"Ids": ids, - "Scores": scores}, - outputs={ - "SentenceIds": sentence_ids, - "SentenceScores": sentence_scores - }) - - return sentence_ids, sentence_scores - - def conv2d_transpose(input, num_filters, output_size=None, @@ -1730,32 +2277,36 @@ def conv2d_transpose(input, represent height and width, respectively. The details of convolution transpose layer, please refer to the following explanation and references `therein `_. + If bias attribution and activation type are provided, bias is added to + the output of the convolution, and the corresponding activation function + is applied to the final result. For each input :math:`X`, the equation is: .. math:: - Out = W \\ast X + Out = \sigma (W \\ast X + b) - In the above equation: + Where: * :math:`X`: Input value, a tensor with NCHW format. * :math:`W`: Filter value, a tensor with MCHW format. - * :math:`\\ast` : Convolution transpose operation. - * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be - different. + * :math:`\\ast`: Convolution operation. + * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. + * :math:`\\sigma`: Activation function. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. Example: - Input: - Input shape: $(N, C_{in}, H_{in}, W_{in})$ + Input shape: :math:`(N, C_{in}, H_{in}, W_{in})` - Filter shape: $(C_{in}, C_{out}, H_f, W_f)$ + Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)` - Output: - Output shape: $(N, C_{out}, H_{out}, W_{out})$ + Output shape: :math:`(N, C_{out}, H_{out}, W_{out})` Where @@ -1765,59 +2316,64 @@ def conv2d_transpose(input, W_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1 Args: - input(Variable): The input image with [N, C, H, W] format. - num_filters(int): The number of the filter. It is as same as the output - image channel. - output_size(int|tuple|None): The output image size. If output size is a - tuple, it must contain two integers, (image_H, image_W). This - parameter only works when filter_size is None. - filter_size(int|tuple|None): The filter size. If filter_size is a tuple, - it must contain two integers, (filter_size_H, filter_size_W). - Otherwise, the filter will be a square. None if use output size to - calculate filter_size. - padding(int|tuple): The padding size. If padding is a tuple, it must - contain two integers, (padding_H, padding_W). Otherwise, the - padding_H = padding_W = padding. Default: padding = 0. - stride(int|tuple): The stride size. If stride is a tuple, it must - contain two integers, (stride_H, stride_W). Otherwise, the - stride_H = stride_W = stride. Default: stride = 1. - dilation(int|tuple): The dilation size. If dilation is a tuple, it must - contain two integers, (dilation_H, dilation_W). Otherwise, the - dilation_H = dilation_W = dilation. Default: dilation = 1. - groups(int): The groups number of the Conv2d transpose layer. Inspired by - grouped convolution in Alex Krizhevsky's Deep CNN paper, in which - when group=2, the first half of the filters is only connected to the - first half of the input channels, while the second half of the - filters is only connected to the second half of the input channels. - Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. - Default: None - bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None - use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn - library is installed. Default: True - act(str): Activation type. Default: None - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. - - Returns: - Variable: The tensor variable storing the convolution transpose result. + input(Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of the filter. It is as same as the output + image channel. + output_size(int|tuple|None): The output image size. If output size is a + tuple, it must contain two integers, (image_H, image_W). This + parameter only works when filter_size is None. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. None if use output size to + calculate filter_size. + padding(int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. Default: padding = 0. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. Default: stride = 1. + dilation(int|tuple): The dilation size. If dilation is a tuple, it must + contain two integers, (dilation_H, dilation_W). Otherwise, the + dilation_H = dilation_W = dilation. Default: dilation = 1. + groups(int): The groups number of the Conv2d transpose layer. Inspired by + grouped convolution in Alex Krizhevsky's Deep CNN paper, in which + when group=2, the first half of the filters is only connected to the + first half of the input channels, while the second half of the + filters is only connected to the second half of the input channels. + Default: groups=1 + param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. + Default: None + bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None + use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + act(str): Activation type. Default: None + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The tensor variable storing the convolution transpose result. Raises: - ValueError: If the shapes of input, filter_size, stride, padding and - groups mismatch. + ValueError: If the shapes of input, filter_size, stride, padding and + groups mismatch. Examples: .. code-block:: python - data = fluid.layers.data( - name='data', shape=[3, 32, 32], dtype='float32') - conv2d_transpose = fluid.layers.conv2d_transpose( - input=data, num_filters=2, filter_size=3) + data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32') + conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3) """ - helper = LayerHelper("conv2d_transpose", **locals()) + + input_channel = input.shape[1] + + op_type = 'conv2d_transpose' + if (input_channel == groups and num_filters == input_channel and + not use_cudnn): + op_type = 'depthwise_conv2d_transpose' + + helper = LayerHelper(op_type, **locals()) if not isinstance(input, Variable): raise TypeError("Input of conv2d_transpose must be Variable") - input_channel = input.shape[1] padding = utils.convert_to_list(padding, 2, 'padding') stride = utils.convert_to_list(stride, 2, 'stride') @@ -1851,7 +2407,176 @@ def conv2d_transpose(input, pre_bias = helper.create_tmp_variable(dtype=input.dtype) helper.append_op( - type='conv2d_transpose', + type=op_type, + inputs={'Input': [input], + 'Filter': [img_filter]}, + outputs={'Output': pre_bias}, + attrs={ + 'strides': stride, + 'paddings': padding, + 'dilations': dilation, + 'groups': groups, + 'use_cudnn': use_cudnn + }) + + pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) + out = helper.append_activation(pre_act) + return out + + +def conv3d_transpose(input, + num_filters, + output_size=None, + filter_size=None, + padding=0, + stride=1, + dilation=1, + groups=None, + param_attr=None, + bias_attr=None, + use_cudnn=True, + act=None, + name=None): + """ + **Convlution3D transpose layer** + + The convolution3D transpose layer calculates the output based on the input, + filter, and dilations, strides, paddings. Input(Input) and output(Output) + are in NCDHW format. Where N is batch size, C is the number of channels, + D is the depth of the feature, H is the height of the feature, and W + is the width of the feature. Parameters(dilations, strides, paddings) are + two elements. These two elements represent height and width, respectively. + The details of convolution transpose layer, please refer to the following + explanation and references `therein `_. + If bias attribution and activation type are provided, bias is added to + the output of the convolution, and the corresponding activation function + is applied to the final result. + + For each input :math:`X`, the equation is: + + .. math:: + + Out = \sigma (W \\ast X + b) + + In the above equation: + + * :math:`X`: Input value, a tensor with NCDHW format. + * :math:`W`: Filter value, a tensor with MCDHW format. + * :math:`\\ast`: Convolution operation. + * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. + * :math:`\\sigma`: Activation function. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. + + Example: + + - Input: + + Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})` + + Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)` + + - Output: + + Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})` + + Where + + .. math:: + + D_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\ + H_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\ + W_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1 + + Args: + input(Variable): The input image with [N, C, D, H, W] format. + num_filters(int): The number of the filter. It is as same as the output + image channel. + output_size(int|tuple|None): The output image size. If output size is a + tuple, it must contain three integers, (image_D, image_H, image_W). This + parameter only works when filter_size is None. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain three integers, (filter_size_D, filter_size_H, filter_size_W). + Otherwise, the filter will be a square. None if use output size to + calculate filter_size. + padding(int|tuple): The padding size. If padding is a tuple, it must + contain three integers, (padding_D, padding_H, padding_W). Otherwise, the + padding_D = padding_H = padding_W = padding. Default: padding = 0. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain three integers, (stride_D, stride_H, stride_W). Otherwise, the + stride_D = stride_H = stride_W = stride. Default: stride = 1. + dilation(int|tuple): The dilation size. If dilation is a tuple, it must + contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the + dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1. + groups(int): The groups number of the Conv3d transpose layer. Inspired by + grouped convolution in Alex Krizhevsky's Deep CNN paper, in which + when group=2, the first half of the filters is only connected to the + first half of the input channels, while the second half of the + filters is only connected to the second half of the input channels. + Default: groups=1 + param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer. + Default: None + bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None + use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + act(str): Activation type. Default: None + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The tensor variable storing the convolution transpose result. + + Raises: + ValueError: If the shapes of input, filter_size, stride, padding and + groups mismatch. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32') + conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3) + """ + l_type = "conv3d_transpose" + helper = LayerHelper(l_type, **locals()) + if not isinstance(input, Variable): + raise TypeError("Input of conv3d_transpose must be Variable") + input_channel = input.shape[1] + + padding = utils.convert_to_list(padding, 3, 'padding') + stride = utils.convert_to_list(stride, 3, 'stride') + dilation = utils.convert_to_list(dilation, 3, 'dilation') + + if not isinstance(use_cudnn, bool): + raise ValueError("use_cudnn should be True or False") + + if filter_size is None: + if output_size is None: + raise ValueError("output_size must be set when filter_size is None") + if isinstance(output_size, int): + output_size = [output_size, output_size] + + d_in = input.shape[2] + h_in = input.shape[3] + w_in = input.shape[4] + + filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 * + padding[0] - 1) / dilation[0] + 1 + filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 * + padding[1] - 1) / dilation[1] + 1 + filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 * + padding[2] - 1) / dilation[2] + 1 + filter_size = [filter_size_d, filter_size_h, filter_size_w] + else: + filter_size = utils.convert_to_list(filter_size, 3, + 'conv3d_transpose.filter_size') + + groups = 1 if groups is None else groups + filter_shape = [input_channel, num_filters / groups] + filter_size + img_filter = helper.create_parameter( + dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) + + pre_bias = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type=l_type, inputs={'Input': [input], 'Filter': [img_filter]}, outputs={'Output': pre_bias}, @@ -1879,18 +2604,18 @@ def sequence_expand(x, y, ref_level=-1, name=None): * Case 1 x is a LoDTensor: - x.lod = [[0, 2, 4]] + x.lod = [[2, 2]] x.data = [[a], [b], [c], [d]] x.dims = [4, 1] y is a LoDTensor: - y.lod = [[0, 2, 4], - [0, 3, 6, 7, 8]] + y.lod = [[2, 2], + [3, 3, 1, 1]] ref_level: 0 then output is a 1-level LoDTensor: - out.lod = [[0, 2, 4, 6, 8]] + out.lod = [[2, 2, 2, 2]] out.data = [[a], [b], [a], [b], [c], [d], [c], [d]] out.dims = [8, 1] @@ -1900,7 +2625,7 @@ def sequence_expand(x, y, ref_level=-1, name=None): x.dims = [3, 1] y is a LoDTensor: - y.lod = [[0, 2, 2, 5]] + y.lod = [[2, 0, 3]] ref_level: -1 @@ -1938,10 +2663,89 @@ def sequence_expand(x, y, ref_level=-1, name=None): return tmp -def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0): - ''' - This function implements the beam search algorithm. - ''' +def beam_search(pre_ids, + pre_scores, + ids, + scores, + beam_size, + end_id, + level=0, + name=None): + """ + Beam search is a classical algorithm for selecting candidate words in a + machine translation task. + + Refer to `Beam search `_ + for more details. + + This layer does the search in beams for one time step. Specifically, it + selects the top-K candidate word ids of current step from :attr:`ids` + according to their :attr:`scores` for all source sentences, where K is + :attr:`beam_size` and :attr:`ids, scores` are predicted results from the + computation cell. Additionally, :attr:`pre_ids` and :attr:`pre_scores` are + the output of beam_search at previous step, they are needed for special use + to handle ended candidate translations. + + Note that the :attr:`scores` passed in should be accumulated scores, and + length penalty should be done with extra operators before calculating the + accumulated scores if needed, also suggest finding top-K before it and + using the top-K candidates following. + + Please see the following demo for a fully beam search usage example: + + fluid/tests/book/test_machine_translation.py + + Args: + pre_ids(Variable): The LodTensor variable which is the output of + beam_search at previous step. It should be a LodTensor with shape + :math:`(batch_size, 1)` and lod + :math:`[[0, 1, ... , batch_size], [0, 1, ..., batch_size]]` at the + first step. + pre_scores(Variable): The LodTensor variable which is the output of + beam_search at previous step. + ids(Variable): The LodTensor variable containing the candidates ids. + Its shape should be :math:`(batch_size \\times beam_size, K)`, + where :math:`K` supposed to be :attr:`beam_size`. + scores(Variable): The LodTensor variable containing the accumulated + scores corresponding to :attr:`ids` and its shape is the same as + the shape of :attr:`ids`. + beam_size(int): The beam width used in beam search. + end_id(int): The id of end token. + level(int, default 0): It can be ignored and mustn't change currently. + It means the source level of lod, which is explained as following. + The lod level of :attr:`ids` should be 2. The first level is source + level which describes how many prefixes (branchs) for each source + sentece (beam), and the second level is sentence level which + describes how these candidates belong to the prefix. The paths + linking prefixes and selected candidates are organized and reserved + in lod. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The LodTensor pair containing the selected ids and the \ + corresponding scores. + + Examples: + .. code-block:: python + + # Suppose `probs` contains predicted results from the computation + # cell and `pre_ids` and `pre_scores` is the output of beam_search + # at previous step. + topk_scores, topk_indices = layers.topk(probs, k=beam_size) + accu_scores = layers.elementwise_add( + x=layers.log(x=topk_scores)), + y=layers.reshape( + pre_scores, shape=[-1]), + axis=0) + selected_ids, selected_scores = layers.beam_search( + pre_ids=pre_ids, + pre_scores=pre_scores, + ids=topk_indices, + scores=accu_scores, + beam_size=beam_size, + end_id=end_id) + """ helper = LayerHelper('beam_search', **locals()) score_type = scores.dtype id_type = ids.dtype @@ -1953,6 +2757,7 @@ def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0): type='beam_search', inputs={ 'pre_ids': pre_ids, + 'pre_scores': pre_scores, 'ids': ids, 'scores': scores, }, @@ -1970,6 +2775,56 @@ def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0): return selected_ids, selected_scores +def beam_search_decode(ids, scores, beam_size, end_id, name=None): + """ + Beam Search Decode Layer. This layer constructs the full hypotheses for + each source sentence by walking back along the LoDTensorArray :attr:`ids` + whose lods can be used to restore the path in the beam search tree. + Please see the following demo for a fully beam search usage example: + fluid/tests/book/test_machine_translation.py + + Args: + ids(Variable): The LodTensorArray variable containing the selected ids + of all steps. + scores(Variable): The LodTensorArray variable containing the selected + scores of all steps. + beam_size(int): The beam width used in beam search. + end_id(int): The id of end token. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The LodTensor pair containing the generated id sequences \ + and the corresponding scores. The shapes and lods of the two \ + LodTensor are same. The lod level is 2 and the two levels \ + separately indicate how many hypotheses each source sentence has \ + and how many ids each hypothesis has. + + Examples: + .. code-block:: python + # Suppose `ids` and `scores` are LodTensorArray variables reserving + # the selected ids and scores of all steps + finished_ids, finished_scores = layers.beam_search_decode( + ids, scores, beam_size=5, end_id=0) + """ + helper = LayerHelper('beam_search_decode', **locals()) + sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) + sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) + + helper.append_op( + type="beam_search_decode", + inputs={"Ids": ids, + "Scores": scores}, + outputs={ + "SentenceIds": sentence_ids, + "SentenceScores": sentence_scores + }, + attrs={"beam_size": beam_size, + "end_id": end_id}) + + return sentence_ids, sentence_scores + + def lstm_unit(x_t, hidden_t_prev, cell_t_prev, @@ -2116,7 +2971,7 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None): # x is a Tensor variable with following elements: # [[0.2, 0.3, 0.5, 0.9] # [0.1, 0.2, 0.6, 0.7]] - # Each example is followed by the correspending output tensor. + # Each example is followed by the corresponding output tensor. fluid.layers.reduce_sum(x) # [3.5] fluid.layers.reduce_sum(x, dim=0) # [0.3, 0.5, 1.1, 1.6] fluid.layers.reduce_sum(x, dim=-1) # [1.9, 1.6] @@ -2125,7 +2980,7 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None): # x is a Tensor variable with shape [2, 2, 2] and elements as below: # [[[1, 2], [3, 4]], # [[5, 6], [7, 8]]] - # Each example is followed by the correspending output tensor. + # Each example is followed by the corresponding output tensor. fluid.layers.reduce_sum(x, dim=[1, 2]) # [10, 26] fluid.layers.reduce_sum(x, dim=[0, 1]) # [16, 20] @@ -2148,23 +3003,24 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None): def reduce_mean(input, dim=None, keep_dim=False, name=None): """ - Computes the mean of tensor elements over the given dimension. + Computes the mean of the input tensor's elements along the given dimension. Args: input (Variable): The input variable which is a Tensor or LoDTensor. - dim (list|int|None): The dimensions along which the mean is computed. If - :attr:`None`, compute the mean over all elements of :attr:`input` - and return a Tensor variable with a single element, otherwise + dim (list|int|None): The dimension along which the mean is computed. If + `None`, compute the mean over all elements of :attr:`input` + and return a variable with a single element, otherwise it must be in the range :math:`[-rank(input), rank(input))`. If - :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. + :math:`dim[i] < 0`, the dimension to reduce is + :math:`rank(input) + dim[i]`. keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. - name(str|None): A name for this layer(optional). If set None, the layer + name(str|None): A name for this layer(optional). If set `None`, the layer will be named automatically. Returns: - Variable: The reduced Tensor variable. + Variable: The reduced mean Variable. Examples: .. code-block:: python @@ -2386,7 +3242,7 @@ def split(input, num_or_sections, dim=-1, name=None): will be named automatically. Returns: - List: The list of segmented tensor variables. + list(Variable): The list of segmented tensor variables. Examples: .. code-block:: python @@ -2436,77 +3292,51 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): The l2 normalize layer normalizes `x` along dimension `axis` using an L2 norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes - output = x / sqrt(max(sum(x**2), epsilon)) + .. math:: + + y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }} For `x` with more dimensions, this layer independently normalizes each 1-D slice along dimension `axis`. Args: - x(Variable|list): The input tensor to l2_normalize layer. - axis(int): Dimension along which to normalize the input. - epsilon(float): A lower bound value for `x`'s l2 norm. sqrt(epsilon) will - be used as the divisor if the l2 norm of `x` is less than - sqrt(epsilon). - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. - + x(Variable|list): The input tensor to l2_normalize layer. + axis(int): The axis on which to apply normalization. If `axis < 0`, \ + the dimension to normalization is rank(X) + axis. -1 is the + last dimension. + epsilon(float): The epsilon value is used to avoid division by zero, \ + the defalut value is 1e-10. + name(str|None): A name for this layer(optional). If set None, the layer \ + will be named automatically. Returns: - Variable: The output tensor variable. + Variable: The output tensor variable is the same shape with `x`. Examples: + .. code-block:: python - data = fluid.layers.data(name="data", - shape=(3, 17, 13), - dtype="float32") - normed = fluid.layers.l2_normalize(x=data, axis=1) + data = fluid.layers.data(name="data", + shape=(3, 17, 13), + dtype="float32") + normed = fluid.layers.l2_normalize(x=data, axis=1) """ if len(x.shape) == 1: axis = 0 helper = LayerHelper("l2_normalize", **locals()) - square = helper.create_tmp_variable(dtype=x.dtype) - helper.append_op(type="square", inputs={"X": x}, outputs={"Out": square}) - - reduced_sum = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_tmp_variable(dtype=x.dtype) + norm = helper.create_tmp_variable(dtype=x.dtype) helper.append_op( - type="reduce_sum", - inputs={"X": square}, - outputs={"Out": reduced_sum}, + type="norm", + inputs={"X": x}, + outputs={"Out": out, + "Norm": norm}, attrs={ - "dim": [1] if axis is None else [axis], - "keep_dim": True, - "reduce_all": False + "axis": 1 if axis is None else axis, + "epsilon": epsilon, }) - - # TODO(caoying) A lower bound value epsilon for the norm is needed to - # imporve the numeric stability of reciprocal. This requires a maximum_op. - rsquare = helper.create_tmp_variable(dtype=x.dtype) - helper.append_op( - type="reciprocal", inputs={"X": reduced_sum}, outputs={"Out": rsquare}) - - # TODO(caoying) the current elementwise_mul operator does not support a - # general broadcast rule which broadcasts input(Y) to have the same - # dimension with Input(X) starting from a specified dimension. So this - # exanpsion is requred. Once a general broadcast rule is spported, this - # expanding canbe removed. - rsquare_expanded = helper.create_tmp_variable(dtype=x.dtype) - expand_times = [1] * len(x.shape) - expand_times[axis] = int(x.shape[axis]) - helper.append_op( - type="expand", - inputs={"X": rsquare}, - outputs={"Out": rsquare_expanded}, - attrs={"expand_times": expand_times}) - - out = helper.create_tmp_variable(dtype=x.dtype) - helper.append_op( - type="elementwise_mul", - inputs={"X": x, - "Y": rsquare_expanded}, - outputs={"Out": out}) return out @@ -2621,25 +3451,51 @@ def topk(input, k, name=None): This operator is used to find values and indices of the k largest entries for the last dimension. - If the input is a vector (rank=1), finds the k largest entries in the vector + If the input is a vector (1-D Tensor), finds the k largest entries in the vector and outputs their values and indices as vectors. Thus values[j] is the j-th largest entry in input, and its index is indices[j]. If the input is a Tensor with higher rank, this operator computes the top k entries along the last dimension. + For example: + + .. code-block:: text + + If: + input = [[5, 4, 2, 3], + [9, 7, 10, 25], + [6, 2, 10, 1]] + k = 2 + + Then: + The first output: + values = [[5, 4], + [10, 25], + [6, 10]] + + The second output: + indices = [[0, 1], + [2, 3], + [0, 2]] + Args: input(Variable): The input variable which can be a vector or Tensor with higher rank. - k(int): An integer value to specify the top k largest elements. + k(int): The number of top elements to look for along the last dimension + of input. name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. + Default: None Returns: - values(Variable): The k largest elements along each last dimensional - slice. - indices(Variable): The indices of values within the last dimension of - input. + Tuple[Variable]: A tuple with two elements. Each element is a Variable. + The first one is k largest elements along each last + dimensional slice. The second one is indices of values + within the last dimension of input. + + Raises: + ValueError: If k < 1 or k is not less than the last dimension of input Examples: .. code-block:: python @@ -2647,7 +3503,7 @@ def topk(input, k, name=None): top5_values, top5_indices = layers.topk(input, k=5) """ shape = input.shape - if k < 1 and k >= shape[-1]: + if k < 1 or k >= shape[-1]: raise ValueError("k must be greater than 0 and less than %d." % (shape[-1])) @@ -2665,8 +3521,7 @@ def topk(input, k, name=None): return values, indices -def edit_distance(input, label, normalized=True, ignored_tokens=None, - name=None): +def edit_distance(input, label, normalized=True, ignored_tokens=None): """ EditDistance operator computes the edit distances between a batch of hypothesis strings and their references. Edit distance, also called @@ -2680,26 +3535,23 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None, "kitten" -> "sitten" -> "sittin" -> "sitting" - Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with + The input is a LoDTensor consisting of all the hypothesis strings with the total number denoted by `batch_size`, and the separation is specified by the LoD information. And the `batch_size` reference strings are arranged - in order in the same way in the LoDTensor Input(Refs). + in order in the same way in the input LoDTensor. - Output(Out) contains the `batch_size` results and each stands for the edit + The output contains the `batch_size` results and each stands for the edit distance for a pair of strings respectively. If Attr(normalized) is true, the edit distance will be divided by the length of reference string. Args: - input(Variable): The indices for hypothesis strings. - label(Variable): The indices for reference strings. - - normalized(bool): Indicated whether to normalize the edit distance by + normalized(bool, default True): Indicated whether to normalize the edit distance by the length of reference string. - - ignored_tokens(list of int): Tokens that should be removed before + ignored_tokens(list, default None): Tokens that should be removed before calculating edit distance. + name (str): The name of this layer. It is optional. Returns: Variable: sequence-to-sequence edit distance in shape [batch_size, 1]. @@ -2709,7 +3561,6 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None, x = fluid.layers.data(name='x', shape=[8], dtype='float32') y = fluid.layers.data(name='y', shape=[7], dtype='float32') - cost = fluid.layers.edit_distance(input=x,label=y) """ helper = LayerHelper("edit_distance", **locals()) @@ -2750,6 +3601,7 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None, def ctc_greedy_decoder(input, blank, name=None): """ This op is used to decode sequences by greedy policy by below steps: + 1. Get the indexes of max value for each row in input. a.k.a. numpy.argmax(input, axis=0). 2. For each sequence in result of step1, merge repeated tokens between two @@ -2771,7 +3623,7 @@ def ctc_greedy_decoder(input, blank, name=None): [0.2, 0.2, 0.1, 0.5], [0.5, 0.1, 0.3, 0.1]] - input.lod = [[0, 4, 8]] + input.lod = [[4, 4]] Then: @@ -2779,7 +3631,7 @@ def ctc_greedy_decoder(input, blank, name=None): [1], [3]] - output.lod = [[0, 2, 3]] + output.lod = [[2, 1]] Args: @@ -2789,14 +3641,14 @@ def ctc_greedy_decoder(input, blank, name=None): where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label). - blank(int): the blank label index of Connectionist Temporal Classification (CTC) loss, which is in thehalf-opened interval [0, num_classes + 1). + name (str): The name of this layer. It is optional. Returns: Variable: CTC greedy decode result. If all the sequences in result were - empty, the result LoDTensor will be [-1] with LoD [[0]] and dims [1, 1]. + empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1]. Examples: .. code-block:: python @@ -2829,35 +3681,33 @@ def warpctc(input, label, blank=0, norm_by_times=False): input tensor. Args: - input(Variable): (LodTensor, default: LoDTensor), - the unscaled probabilities of variable-length sequences, + input (Variable): The unscaled probabilities of variable-length sequences, which is a 2-D Tensor with LoD information. It's shape is [Lp, num_classes + 1], where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label). - label(Variable): (LodTensor, default: LoDTensor), the ground truth - of variable-length sequence, which is a 2-D Tensor with LoD - information. It is of the shape [Lg, 1], where Lg is th sum of - all labels' length. - blank: (int, default: 0), the blank label index of Connectionist + label (Variable): The ground truth of variable-length sequence, + which is a 2-D Tensor with LoD information. It is of the shape [Lg, 1], + where Lg is th sum of all labels' length. + blank (int, default 0): The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). - norm_by_times: (bool, default: false), whether to normalize - the gradients by the number of time-step, which is also the - sequence's length. There is no need to normalize the gradients - if warpctc layer was follewed by a mean_op. + norm_by_times(bool, default false): Whether to normalize the gradients + by the number of time-step, which is also the sequence's length. + There is no need to normalize the gradients if warpctc layer was + follewed by a mean_op. Returns: Variable: The Connectionist Temporal Classification (CTC) loss, which is a 2-D Tensor of the shape [batch_size, 1]. Examples: + .. code-block:: python - y = layers.data( - name='y', shape=[11, 8], dtype='float32', lod_level=1) - y_predict = layers.data( - name='y_predict', shape=[11, 1], dtype='float32') - cost = layers.warpctc(input=y_predict, label=y) + + label = fluid.layers.data(shape=[11, 8], dtype='float32', lod_level=1) + predict = fluid.layers.data(shape=[11, 1], dtype='float32') + cost = fluid.layers.warpctc(input=predict, label=label) """ helper = LayerHelper('warpctc', **locals()) @@ -2887,16 +3737,20 @@ def sequence_reshape(input, new_dim): x is a LoDTensor: x.lod = [[0, 2, 6]] - x.data = [[1, 2], [3, 4], - [5, 6], [7, 8], [9, 10], [11, 12]] + x.data = [[1, 2], [3, 4], + [5, 6], [7, 8], + [9, 10], [11, 12]] x.dims = [6, 2] set new_dim = 4 then out is a LoDTensor: + out.lod = [[0, 1, 3]] - out.data = [[1, 2, 3, 4], - [5, 6, 7, 8], [9, 10, 11, 12]] + + out.data = [[1, 2, 3, 4], + [5, 6, 7, 8], + [9, 10, 11, 12]] out.dims = [3, 4] Currently, only 1-level LoDTensor is supported and please make sure @@ -2904,19 +3758,19 @@ def sequence_reshape(input, new_dim): no remainder for each sequence. Args: - input (Variable): (LodTensor, default: LoDTensor), a 2-D LoDTensor - with shape being [N, M] where M for dimension. - new_dim (int): New dimension which the input LoDTensor is reshaped to. + + input (Variable): A 2-D LoDTensor with shape being [N, M] where M for dimension. + new_dim (int): New dimension that the input LoDTensor is reshaped to. Returns: + Variable: Reshaped LoDTensor according to new dimension. Examples: .. code-block:: python - x = fluid.layers.data(name='x', shape=[5, 20], - dtype='float32', lod_level=1) - x_reshaped = layers.sequence_reshape(input=x, new_dim=10) + x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1) + x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10) """ helper = LayerHelper('sequence_reshape', **locals()) out = helper.create_tmp_variable(helper.input_dtype()) @@ -2928,7 +3782,10 @@ def sequence_reshape(input, new_dim): return out -@autodoc() +# FIXME(wuyi): let docstring_checker.py understand @autodoc. +# For now, the comments in c++ use types like Tensor, but in python side +# the type is often "Variable", and arguments may vary. +@templatedoc(op_type="nce") def nce(input, label, num_total_classes, @@ -2936,6 +3793,49 @@ def nce(input, param_attr=None, bias_attr=None, num_neg_samples=None): + """ + ${comment} + + Args: + input (Variable): input variable. + label (Variable): label. + num_total_classes (int):${num_total_classes_comment} + sample_weight (Variable|None): A Variable of shape [batch_size, 1] + storing a weight for each sample. The default weight for each + sample is 1.0. + param_attr (ParamAttr|None): attributes for parameter + bias_attr (ParamAttr|None): attributes for bias + num_neg_samples (int): ${num_neg_samples_comment} + + Returns: + Variable: The output nce loss. + + Examples: + .. code-block:: python + + window_size = 5 + words = [] + for i in xrange(window_size): + words.append(layers.data( + name='word_{0}'.format(i), shape=[1], dtype='int64')) + + dict_size = 10000 + label_word = int(window_size / 2) + 1 + + embs = [] + for i in xrange(window_size): + if i == label_word: + continue + + emb = layers.embedding(input=words[i], size=[dict_size, 32], + param_attr='emb.w', is_sparse=True) + embs.append(emb) + + embs = layers.concat(input=embs, axis=1) + loss = layers.nce(input=embs, label=words[label_word], + num_total_classes=dict_size, param_attr='nce.w', + bias_attr='nce.b') + """ helper = LayerHelper('nce', **locals()) assert isinstance(input, Variable) dim = input.shape[1] @@ -2983,18 +3883,85 @@ def nce(input, return cost / (num_neg_samples + 1) -def transpose(x, perm, name=None): +def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): """ - **transpose Layer** + The hierarchical sigmoid operator is used to accelerate the training + process of language model. This operator organizes the classes into a + complete binary tree, each leaf node represents a class(a word) and each + internal node acts as a binary classifier. For each word there's a unique + path from root to it's leaf node, hsigmoid calculate the cost for each + internal node on the path, and sum them to get a total cost. hsigmoid can + achive a acceleration from :math:`O(N)` to :math:`O(logN)`, where :math:`N` + represents the size of word dict. + + Refer to `Hierarchical Probabilistic Neural Network Language Model + `_ + + Args: + input (Variable): The input tensor variable with shape + :math:`[N \\times D]`, where :math:`N` is the size of mini-batch, + and :math:`D` is the feature size. + label (Variable): The tensor variable contains labels of training data. + It's a tensor with shape is :math:`[N \\times 1]`. + num_classes: (int), The number of classes, must not be less than 2. + param_attr (ParamAttr|list of ParamAttr, default None): The parameter + attribute for learnable parameters/weights of this layer. + bias_attr (ParamAttr|list of ParamAttr, default None): The parameter + attribute for the bias of this layer. If it is set to False, no + bias will be applied. + + Returns: + Out: (Tensor) The cost of hierarchical sigmoid operator. the shape is [N, 1] + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[2], dtype='float32') + y = fluid.layers.data(name='y', shape=[1], dtype='int64') + out = fluid.layers.hsigmoid(input=x, label=y, num_classes=6) + """ + + helper = LayerHelper('hierarchical_sigmoid', **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + pre_out = helper.create_tmp_variable(dtype) + dim = input.shape[1] + if num_classes < 2: + raise ValueError("num_classes must not be less than 2.") + weights = helper.create_parameter( + attr=helper.param_attr, + shape=[num_classes - 1, dim], + is_bias=False, + dtype=input.dtype) + inputs = {"X": input, "W": weights, "Label": label} + if helper.bias_attr: + bias = helper.create_parameter( + attr=helper.bias_attr, + shape=[1, num_classes - 1], + is_bias=True, + dtype=input.dtype) + inputs['Bias'] = bias + helper.append_op( + type="hierarchical_sigmoid", + inputs=inputs, + outputs={"Out": out, + "PreOut": pre_out}, + attrs={"num_classes": num_classes}) + return out + + +def transpose(x, perm, name=None): + """ Permute the dimensions of `input` according to `perm`. The `i`-th dimension of the returned tensor will correspond to the perm[i]-th dimension of `input`. Args: - input (Variable): (Tensor), A Tensor. - perm (list): A permutation of the dimensions of `input`. + x (Variable): The input Tensor. + perm (list): A permutation of the dimensions of `input`. + name (str): The name of this layer. It is optional. Returns: Variable: A transposed Tensor. @@ -3027,7 +3994,13 @@ def transpose(x, perm, name=None): return out -def im2sequence(input, filter_size=1, stride=1, padding=0, name=None): +def im2sequence(input, + filter_size=1, + stride=1, + padding=0, + input_image_size=None, + out_stride=1, + name=None): """ Extracts image patches from the input tensor to form a tensor of shape {input.batch_size * output_height * output_width, filter_size_H * @@ -3064,6 +4037,15 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None): padding_up = padding_down = padding_left = padding_right = padding Default: padding = 0. + input_image_size(Variable): the input contains image real size.It's dim + is [batchsize, 2]. It is dispensable.It is just for batch inference. + + out_stride(int|tuple): The scaling of image through CNN. It is + dispensable. It is valid only when input_image_size is not null. + If out_stride is tuple, it must contain two intergers, + (out_stride_H, out_stride_W). Otherwise, + the out_stride_H = out_stride_W = out_stride. + name (int): The name of this layer. It is optional. Returns: @@ -3075,8 +4057,6 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None): Examples: - As an example: - .. code-block:: text Given: @@ -3116,11 +4096,11 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None): [ 5. 7. 2. 4. 1. 3. 9. 0.] [ 7. 9. 4. 8. 3. 5. 0. 8.]] - output.dims = {8, 9} + output.dims = {8, 8} - output.lod = [[0, 4, 8]] + output.lod = [[4, 4]] - The simple usage is: + Examples: .. code-block:: python @@ -3138,44 +4118,27 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None): if len(padding) == 2: padding.append(padding[0]) padding.append(padding[1]) - + inputs = {"X": input} + attrs = {"kernels": filter_size, "strides": stride, "padding": padding} + if input_image_size: + if isinstance(out_stride, int): + out_stride = [out_stride, out_stride] + inputs["Y"] = input_image_size + attrs["out_stride"] = out_stride helper = LayerHelper('im2sequence', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) helper.append_op( - type='im2sequence', - inputs={'X': input}, - outputs={'Out': out}, - attrs={ - 'kernels': filter_size, - 'strides': stride, - 'paddings': padding, - }) + type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs) return out +@templatedoc() def row_conv(input, future_context_size, param_attr=None, act=None): - """Row Conv Operator. This layer will apply lookahead convolution to - **input**. The input variable should be a 2D LoDTensor with shape [T, D]. - Parameters with shape [future_context_size + 1, D] will be created. The math - equation of row convolution is as follows: - - .. math:: - Out_{i} = \sum_{j = i} ^ {i + \\tau} X_{j} \odot W_{i - j} - - In the above equation: - - * :math:`Out_{i}`: The i-th row of output variable with shape [1, D]. - * :math:`\\tau`: Future context size. - * :math:`X_{j}`: The j-th row of input variable with shape [1, D]. - * :math:`W_{i-j}`: The (i-j)-th row of parameters with shape [1, D]. - - More details about row_conv please refer to the paper \ - (http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf) and - the design document \ - (https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645). + """ + ${comment} Args: - input (Variable): Input variable, a 2D LoDTensor with shape [T, D]. + input (${x_type}): ${x_comment}. future_context_size (int): Future context size. Please note, the shape of convolution kernel is [future_context_size + 1, D]. param_attr (ParamAttr): Attributes of parameters, including @@ -3183,14 +4146,13 @@ def row_conv(input, future_context_size, param_attr=None, act=None): act (str): Non-linear activation to be applied to output variable. Returns: - Variable: The output tensor with same shape as input tensor. + ${out_comment}. Examples: - .. code-block:: python - - x = fluid.layers.data(name='x', shape=[16], - dtype='float32', lod_level=1) - out = fluid.layers.row_conv(input=x, future_context_size=2) + >>> import paddle.fluid as fluid + >>> x = fluid.layers.data(name='x', shape=[16], + >>> dtype='float32', lod_level=1) + >>> out = fluid.layers.row_conv(input=x, future_context_size=2) """ helper = LayerHelper('row_conv', **locals()) dtype = helper.input_dtype() @@ -3206,42 +4168,23 @@ def row_conv(input, future_context_size, param_attr=None, act=None): return helper.append_activation(out) +@templatedoc() def multiplex(inputs, index): """ - **Multiplex Layer** + ${comment} - Referring to the given index variable, this layer selects rows from the - input variables to construct a multiplex variable. Assuming that there are - :math:`m` input variables and :math:`I_i` represents the i-th input - variable and :math:`i` is in [0, :math:`m`). All input variables are - tensors with same shape [:math:`d_0`, :math:`d_1`, ..., :math:`d_R`]. - Please note that rank of the input tensor should be at least 2. Each input - variable will be treated as a 2-D matrix with shape [:math:`M`, :math:`N`] - where :math:`M` for :math:`d_0` and :math:`N` for :math:`d_1` * :math:`d_2` - * ... * :math:`d_R`. Let :math:`I_i[j]` be the j-th row of the i-th input - variable. The given index variable should be a 2-D tensor with shape - [:math:`M`, 1]. Let `ID[i]` be the i-th index value of the index variable. - Then the output variable will be a tensor with shape [:math:`d_0`, - :math:`d_1`, ..., :math:`d_R`]. If we treat the output tensor as a 2-D - matrix with shape [:math:`M`, :math:`N`] and let :math:`O[i]` be the i-th - row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`. + >>> import paddle.fluid as fluid + >>> x1 = fluid.layers.data(name='x1', shape=[4], dtype='float32') + >>> x2 = fluid.layers.data(name='x2', shape=[4], dtype='float32') + >>> index = fluid.layers.data(name='index', shape=[1], dtype='int32') + >>> out = fluid.layers.multiplex(inputs=[x1, x2], index=index) Args: - inputs (list): A list of variables to gather from. All variables have the - same shape and the rank is at least 2. - index (Variable): Tensor, index variable which is a 2-D tensor - with shape [M, 1] where M is the batch size. + inputs (list): ${x_comment}. + index (${ids_type}): ${ids_comment}. Returns: - Variable: Multiplex variable gathered from input variables. - - Examples: - .. code-block:: python - - x1 = fluid.layers.data(name='x1', shape=[4], dtype='float32') - x2 = fluid.layers.data(name='x2', shape=[4], dtype='float32') - index = fluid.layers.data(name='index', shape=[1], dtype='int32') - out = fluid.layers.multiplex(inputs=[x1, x2], index=index) + ${out_comment}. """ helper = LayerHelper('multiplex', **locals()) @@ -3327,31 +4270,30 @@ def softmax_with_cross_entropy(logits, label, soft_label=False): def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): """ - **Smooth L1 Loss Operator. ** - - This operator computes the smooth L1 loss for X and Y. - The operator takes the first dimension of X and Y as batch size. + This layer computes the smooth L1 loss for Variable :attr:`x` and :attr:`y`. + It takes the first dimension of :attr:`x` and :attr:`y` as batch size. For each instance, it computes the smooth L1 loss element by element first - and then sums all the losses. So the shape of Out is [batch_size, 1]. + and then sums all the losses. So the shape of ouput Variable is + [batch_size, 1]. Args: x (Variable): A tensor with rank at least 2. The input value of smooth L1 loss op with shape [batch_size, dim1, ..., dimN]. y (Variable): A tensor with rank at least 2. The target value of smooth - L1 loss op with same shape as x. + L1 loss op with same shape as :attr:`x`. inside_weight (Variable|None): A tensor with rank at least 2. This - input is optional and should have same shape with x. If provided, - the result of (x - y) will be multiplied by this tensor element by - element. + input is optional and should have same shape with :attr:`x`. If + provided, the result of (:attr:`x` - :attr:`y`) will be multiplied + by this tensor element by element. outside_weight (Variable|None): A tensor with rank at least 2. This - input is optional and should have same shape with x. If provided, - the out smooth L1 loss will be multiplied by this tensor element - by element. - sigma (float|None): Hyper parameter of smooth L1 loss op. A float scalar - with default value 1.0. + input is optional and should have same shape with :attr:`x`. If + provided, the out smooth L1 loss will be multiplied by this tensor + element by element. + sigma (float|None): Hyper parameter of smooth L1 loss layer. A float + scalar with default value 1.0. + Returns: - Variable: A tensor with rank be 2. The output smooth L1 loss with - shape [batch_size, 1]. + Variable: The output smooth L1 loss with shape [batch_size, 1]. Examples: .. code-block:: python @@ -3362,6 +4304,7 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): fc = fluid.layers.fc(input=data, size=100) out = fluid.layers.smooth_l1(x=fc, y=label) """ + helper = LayerHelper('smooth_l1_loss', **locals()) diff = helper.create_tmp_variable(dtype=x.dtype) loss = helper.create_tmp_variable(dtype=x.dtype) @@ -3381,32 +4324,20 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): def one_hot(input, depth): """ - One Hot Operator. This operator creates the one-hot representations for input - index values. The following example will help to explain the function of this - operator. + This layer creates the one-hot representations for input indices. Args: - input(variable): A Tensor/LodTensor of indices, last dimension must be 1. - depth(scalar): an interger defining the depth of the one hot dimension. + input(Variable): Input indices, last dimension must be 1. + depth(scalar): An interger defining the depth of the one-hot dimension. Returns: - The one-hot tensor or LodTensor, same as input. + Variable: The one-hot representations of input. Examples: .. code-block:: python - X is a LoDTensor: - X.lod = [[0, 1, 4]] - X.shape = [4, 1] - X.data = [[1], [1], [3], [0]] - set depth = 4 - Out is a LoDTensor: - Out.lod = [[0, 1, 4]] - Out.shape = [4, 4] - Out.data = [[0., 1., 0., 0.], - [0., 1., 0., 0.], - [0., 0., 0., 1.], - [1., 0., 0., 0.]] + label = layers.data(name="label", shape=[1], dtype="float32") + one_hot_label = layers.one_hot(input=label, depth=10) """ helper = LayerHelper("one_hot", **locals()) one_hot_out = helper.create_tmp_variable(dtype='float32') @@ -3420,15 +4351,23 @@ def one_hot(input, depth): def autoincreased_step_counter(counter_name=None, begin=1, step=1): """ - NOTE: The counter will be automatically increased by 1 every mini-batch - Return the run counter of the main program, which is started with 1. + Create an auto-increase variable + which will be automatically increased by 1 every mini-batch + Return the run counter of the main program, default is started from 1. Args: counter_name(str): The counter name, default is '@STEP_COUNTER@'. begin(int): The first value of this counter. step(int): The increment step between each execution. - Returns(Variable): The global run counter. + Returns: + Variable: The global run counter. + + Examples: + .. code-block:: python + + global_step = fluid.layers.autoincreased_step_counter( + counter_name='@LR_DECAY_COUNTER@', begin=begin, step=1) """ helper = LayerHelper('global_step_counter') if counter_name is None: @@ -3439,7 +4378,7 @@ def autoincreased_step_counter(counter_name=None, begin=1, step=1): helper.set_variable_initializer( counter, initializer=Constant( value=begin - 1, force_cpu=True)) - helper.main_program.global_block().prepend_op( + helper.main_program.global_block()._prepend_op( type='increment', inputs={'X': [counter]}, outputs={'Out': [counter]}, @@ -3489,7 +4428,7 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): the corresponding dimension of x. Args: - input(variable): The input tensor. + x(variable): The input tensor. shape(list): The new shape. At most one dimension of the new shape can be -1. actual_shape(variable): An optional input. If provided, reshape @@ -3498,11 +4437,17 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): say :attr:`actual_shape` has a higher priority than :attr:`shape`. act (str): The non-linear activation to be applied to output variable. - inplace(bool): If this flag is set true, a new output tensor is created - whose data is copied from input x, otherwise the output - shares data with input without copying. + inplace(bool): If this flag is set true, the output + shares data with input without copying, otherwise + a new output tensor is created + whose data is copied from input x. + name (str): The name of this layer. It is optional. + + Returns: + Variable: The output tensor. - Returns(variable): The output tensor. + Raises: + TypeError: if actual_shape is neither Variable nor None. Examples: .. code-block:: python @@ -3515,6 +4460,11 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): if not (isinstance(shape, list) or isinstance(shape, tuple)): raise ValueError("Input shape must be a python lsit or tuple.") + inputs = {"X": x} + if isinstance(actual_shape, Variable): + inputs["Shape"] = actual_shape + elif actual_shape is not None: + raise TypeError("actual_shape should either be Variable or None") # Validate the shape unk_dim_idx = -1 @@ -3532,88 +4482,86 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): "except one unknown dimension.") helper = LayerHelper("reshape", **locals()) - reshaped = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_tmp_variable(dtype=x.dtype) helper.append_op( type="reshape", - inputs={"X": x, - "Shape": actual_shape} - if isinstance(actual_shape, Variable) else {"X": x}, - attrs={"shape": shape, - "inplace": inplace}, - outputs={"Out": reshaped}) + inputs=inputs, + attrs={"shape": shape}, + outputs={"Out": out}) - return helper.append_activation(reshaped) + return helper.append_activation(out) def lod_reset(x, y=None, target_lod=None): """ - LoD Reset Operator. Set LoD of **x** to a new one specified by **y** or - **target_lod**. When **y** provided, **y.lod** would be considered as target - LoD first, otherwise **y.data** would be considered as target LoD. If **y** - is not provided, target LoD should be specified by **target_lod**. - If target LoD is specified by **Y.data** or **target_lod**, only one level - LoD is supported. + Set LoD of :attr:`x` to a new one specified by :attr:`y` or + :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be + considered as target LoD first, otherwise :attr:`y.data` would be + considered as target LoD. If :attr:`y` is not provided, target LoD should + be specified by :attr:`target_lod`. If target LoD is specified by + :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported. .. code-block:: text * Example 1: Given a 1-level LoDTensor x: - x.lod = [[ 0, 2, 5 6 ]] + x.lod = [[ 2, 3, 1 ]] x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] x.dims = [6, 1] - target_lod: [0, 4, 6] + target_lod: [4, 2] then we get a 1-level LoDTensor: - out.lod = [[ 0, 4, 6 ]] + out.lod = [[4, 2]] out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] out.dims = [6, 1] * Example 2: Given a 1-level LoDTensor x: - x.lod = [[ 0, 2, 5 6 ]] + x.lod = [[2, 3, 1]] x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] x.dims = [6, 1] y is a Tensor: - y.data = [[0, 2, 6]] + y.data = [[2, 4]] y.dims = [1, 3] then we get a 1-level LoDTensor: - out.lod = [[ 0, 2, 6 ]] + out.lod = [[2, 4]] out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] out.dims = [6, 1] * Example 3: Given a 1-level LoDTensor x: - x.lod = [[ 0, 2, 5 6 ]] + x.lod = [[2, 3, 1]] x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] x.dims = [6, 1] y is a 2-level LoDTensor: - y.lod = [[0, 2, 4], [0, 2, 5, 6]] + y.lod = [[2, 2], [2, 2, 1, 1]] y.data = [[1.1], [2.1], [3.1], [4.1], [5.1], [6.1]] y.dims = [6, 1] then we get a 2-level LoDTensor: - out.lod = [[0, 2, 4], [0, 2, 5, 6]] + out.lod = [[2, 2], [2, 2, 1, 1]] out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] out.dims = [6, 1] Args: x (Variable): Input variable which could be a Tensor or LodTensor. - y (Variable|None): If provided, output's LoD would be derived from y. + y (Variable|None): If provided, output's LoD would be derived + from :attr:`y`. target_lod (list|tuple|None): One level LoD which should be considered - as target LoD when y not provided. + as target LoD when :attr:`y` not provided. Returns: - Variable: Output variable with LoD specified by this operator. + Variable: Output variable with LoD specified by this layer. Raises: - ValueError: If y and target_lod are both None. + ValueError: If :attr:`y` and :attr:`target_lod` are both None. Examples: .. code-block:: python @@ -3649,9 +4597,7 @@ def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None): .. math:: - Output(i, x, y) = Input(i, x, y) / \left( - k + \alpha \sum\limits^{\min(C, c + n/2)}_{j = \max(0, c - n/2)} - (Input(j, x, y))^2 \right)^{\beta} + Output(i, x, y) = Input(i, x, y) / \\left(k + \\alpha \\sum\\limits^{\\min(C, c + n/2)}_{j = \\max(0, c - n/2)}(Input(j, x, y))^2\\right)^{\\beta} In the above equation: @@ -3835,34 +4781,20 @@ def label_smooth(label, return smooth_label +@templatedoc() def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): """ - Region of interest pooling (also known as RoI pooling) is to perform - is to perform max pooling on inputs of nonuniform sizes to obtain - fixed-size feature maps (e.g. 7*7). - The operator has three steps: - 1. Dividing each region proposal into equal-sized sections with - the pooled_width and pooled_height - 2. Finding the largest value in each section - 3. Copying these max values to the output buffer + ${comment} Args: - input (Variable): The input for ROI pooling. - rois (Variable): ROIs (Regions of Interest) to pool over. It should - be a 2-D one level LoTensor of shape [num_rois, 4]. - The layout is [x1, y1, x2, y2], where (x1, y1) - is the top left coordinates, and (x2, y2) is the - bottom right coordinates. The num_rois is the - total number of ROIs in this batch data. - pooled_height (integer): The pooled output height. Default: 1 - pooled_width (integer): The pooled output width. Default: 1 - spatial_scale (float): Multiplicative spatial scale factor. To - translate ROI coords from their input scale - to the scale used when pooling. Default: 1.0 + input (Variable): ${x_comment} + rois (Variable): ROIs (Regions of Interest) to pool over. + pooled_height (integer): ${pooled_height_comment} Default: 1 + pooled_width (integer): ${pooled_width_comment} Default: 1 + spatial_scale (float): ${spatial_scale_comment} Default: 1.0 Returns: - pool_out (Variable): The output is a 4-D tensor of the shape - (num_rois, channels, pooled_h, pooled_w). + Variable: ${out_comment}. Examples: .. code-block:: python @@ -3889,7 +4821,6 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): def dice_loss(input, label, epsilon=0.00001): """ - **Dice loss Layer** Dice loss for comparing the similarity of two batch of data, usually is used for binary image segmentation i.e. labels are binary. The dice loss can be defined as below equation: @@ -3920,7 +4851,7 @@ def dice_loss(input, label, epsilon=0.00001): loss = fluid.layers.dice_loss(input=predictions, label=label, 2) """ label = one_hot(label, depth=input.shape[-1]) - reduce_dim = range(1, len(input.shape)) + reduce_dim = list(range(1, len(input.shape))) inse = reduce_sum(input * label, dim=reduce_dim) dice_denominator = reduce_sum( input, dim=reduce_dim) + reduce_sum( @@ -3929,40 +4860,50 @@ def dice_loss(input, label, epsilon=0.00001): return reduce_mean(dice_score) -def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): +def image_resize(input, + out_shape=None, + scale=None, + name=None, + resample='BILINEAR'): """ - The mathematical meaning of upsampling_bilinear2d is also called - Bilinear interpolation. - Bilinear interpolation is an extension of linear interpolation for - interpolating functions of two variables (e.g. H-direction and - W-direction in this layer) on a rectilinear 2D grid. + **Resize a Batch of Images** + + The input must be a tensor of the shape (num_batches, channels, in_h, in_w), + and the resizing only applies on the last two dimensions(hight and width). + + Supporting resample methods: - For details, please refer to Wikipedia: - https://en.wikipedia.org/wiki/Bilinear_interpolation + 'BILINEAR' : Bilinear interpolation Args: - input (Variable): The input tensor of bilinear interpolation, + input (Variable): The input tensor of image resize layer, This is a 4-D tensor of the shape (num_batches, channels, in_h, in_w). - out_shape(list|tuple|None): Output shape of bilinear interpolation + out_shape(list|tuple|Variable|None): Output shape of image resize layer, the shape is (out_h, out_w). Default: None - scale(int|None): The multiplier for the input height or width. + scale(float|None): The multiplier for the input height or width. At least one of out_shape or scale must be set. And out_shape has a higher priority than scale. Default: None name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. + resample(str): The resample method. It can only be 'BILINEAR' currently. + Default: 'BILINEAR' Returns: - out (Variable): The output is a 4-D tensor of the shape - (num_batches, channls, out_h, out_w). + Variable: The output is a 4-D tensor of the shape + (num_batches, channls, out_h, out_w). Examples: .. code-block:: python - out = fluid.layers.bilinear_interp(input, out_shape=[12, 12]) + out = fluid.layers.image_resize(input, out_shape=[12, 12]) """ + resample_methods = {'BILINEAR': 'bilinear_interp'} + if resample not in resample_methods: + raise ValueError( + "The 'resample' of image_resize can only be 'BILINEAR' currently.") if out_shape is None and scale is None: raise ValueError("One of out_shape and scale must not be None") helper = LayerHelper('bilinear_interp', **locals()) @@ -3971,52 +4912,520 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): def _is_list_or_turple_(data): return (isinstance(data, list) or isinstance(data, tuple)) + out_h = 0 + out_w = 0 + inputs = {"X": input} if out_shape is not None: - if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2): - raise ValueError('out_shape should be a list or tuple ', - 'with length 2, (out_h, out_w).') - out_shape = list(map(int, out_shape)) - out_h = out_shape[0] - out_w = out_shape[1] + if not (_is_list_or_turple_(out_shape) and + len(out_shape) == 2) and not isinstance(out_shape, Variable): + raise ValueError('out_shape should be a list or tuple or variable') + if _is_list_or_turple_(out_shape): + out_shape = list(map(int, out_shape)) + out_h = out_shape[0] + out_w = out_shape[1] + else: + inputs['OutSize'] = out_shape else: out_h = int(input.shape[2] * scale) out_w = int(input.shape[3] * scale) out = helper.create_tmp_variable(dtype) helper.append_op( - type="bilinear_interp", - inputs={"X": input}, + type=resample_methods[resample], + inputs=inputs, outputs={"Out": out}, attrs={"out_h": out_h, "out_w": out_w}) return out -def random_crop(input, shape, seed=1): - helper = LayerHelper("random_crop", **locals()) +@templatedoc(op_type="bilinear_interp") +def resize_bilinear(input, out_shape=None, scale=None, name=None): + """ + ${comment} + + Args: + input(${x_type}): ${x_comment}. + + out_shape(${out_size_type}): ${out_size_comment}. + + scale(float|None): The multiplier for the input height or width. At + least one of out_shape or scale must be set. And out_shape has + a higher priority than scale. Default: None. + + name(str|None): The output variable name. + + Returns: + ${out_comment}. + """ + + return image_resize(input, out_shape, scale, name, 'BILINEAR') + + +def image_resize_short(input, out_short_len, resample='BILINEAR'): + """ + Resize a batch of images. The short edge of input images will be + resized to the given 'out_short_len'. The long edge of input images + will be resized proportionately to make images' length-width ratio + constant. + + Args: + input (Variable): The input tensor of image resize layer, + This is a 4-D tensor of the shape + (num_batches, channels, in_h, in_w). + out_short_len(int): The length of output images' short edge. + resample (str): resample method, default: BILINEAR. + + Returns: + Variable: The output is a 4-D tensor of the shape + (num_batches, channls, out_h, out_w). + """ + in_shape = input.shape + if len(in_shape) != 4: + raise ValueError( + "The rank of input must be 4 (num_batches, channels, in_h, in_w).") + hw = in_shape[2:4] + short_idx = hw.index(min(hw)) + long_idx = 1 - short_idx + out_shape = list(hw) + out_shape[short_idx] = out_short_len + out_shape[long_idx] = int( + float(out_shape[long_idx]) * (float(out_short_len) / float(hw[ + short_idx])) + 0.5) + return image_resize(input=input, out_shape=out_shape, resample=resample) + + +def gather(input, index): + """ + **Gather Layer** + + Output is obtained by gathering entries of the outer-most dimension + of X indexed by `index` and concatenate them together. + + .. math:: + + Out = X[Index] + + + .. code-block:: text + + + Given: + + X = [[1, 2], + [3, 4], + [5, 6]] + + Index = [1, 2] + + Then: + + Out = [[3, 4], + [5, 6]] + + Args: + input (Variable): The source input with rank>=1. + index (Variable): The index input with rank=1. + + Returns: + output (Variable): The output is a tensor with the same rank as input. + + Examples: + + .. code-block:: python + + output = fluid.layers.gather(x, index) + """ + helper = LayerHelper('gather', **locals()) dtype = helper.input_dtype() out = helper.create_tmp_variable(dtype) + helper.append_op( + type="gather", + inputs={"X": input, + "Index": index}, + outputs={"Out": out}) + return out + + +@templatedoc() +def random_crop(x, shape, seed=None): + """ + ${comment} + + Args: + x(${x_type}): ${x_comment} + shape(${shape_type}): ${shape_comment} + seed(int|${seed_type}|None): ${seed_comment} By default, the seed will + get from `random.randint(-65536, 65535)`. + + Returns: + ${out_comment} + + Examples: + >>> img = fluid.layers.data("img", [3, 256, 256]) + >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224]) + """ + helper = LayerHelper("random_crop", **locals()) + dtype = x.dtype + out = helper.create_tmp_variable(dtype) + if seed is None: + seed = random.randint(-65536, 65535) + op_attrs = {"shape": shape} if isinstance(seed, int): - seed_value = seed - seed = helper.create_tmp_variable(dtype="int64") - helper.append_op( - type="fill_constant", - inputs={}, - outputs={"Out": seed}, - attrs={ - "dtype": seed.dtype, - "shape": [1], - "value": float(seed_value), - "force_cpu": True - }) + op_attrs["startup_seed"] = seed + seed = helper.create_variable( + name=unique_name.generate("random_crop_seed"), + dtype="int64", + persistable=True) elif not isinstance(seed, Variable): raise ValueError("'seed' must be a Variable or an int.") - seed_out = helper.create_tmp_variable(dtype="int64") helper.append_op( type="random_crop", - inputs={"X": input, + inputs={"X": x, "Seed": seed}, outputs={"Out": out, - "SeedOut": seed_out}, - attrs={"shape": shape}) + "SeedOut": seed}, + attrs=op_attrs) + return out + + +def log(x): + """ + Calculates the natural log of the given input tensor, element-wise. + + .. math:: + + Out = \\ln(x) + + Args: + x (Variable): Input tensor. + + Returns: + Variable: The natural log of the input tensor computed element-wise. + + Examples: + + .. code-block:: python + + output = fluid.layers.log(x) + """ + helper = LayerHelper('log', **locals()) + dtype = helper.input_dtype(input_param_name='x') + out = helper.create_tmp_variable(dtype) + helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out}) + return out + + +def relu(x): + """ + Relu takes one input data (Tensor) and produces one output data (Tensor) + where the rectified linear function, y = max(0, x), is applied to + the tensor elementwise. + + .. math:: + + Out = \\max(0, x) + + Args: + x (Variable): The input tensor. + + Returns: + Variable: The output tensor with the same shape as input. + + Examples: + + .. code-block:: python + + output = fluid.layers.relu(x) + """ + helper = LayerHelper('relu', **locals()) + dtype = helper.input_dtype(input_param_name='x') + out = helper.create_tmp_variable(dtype) + helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out}) + return out + + +def mean_iou(input, label, num_classes): + """ + Mean Intersection-Over-Union is a common evaluation metric for + semantic image segmentation, which first computes the IOU for each + semantic class and then computes the average over classes. + IOU is defined as follows: + + .. math:: + + IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}. + + The predictions are accumulated in a confusion matrix and mean-IOU + is then calculated from it. + + + Args: + input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64. + label (Variable): A Tensor of ground truth labels with type int32 or int64. + Its shape should be the same as input. + num_classes (int): The possible number of labels. + + Returns: + mean_iou (Variable): A Tensor representing the mean intersection-over-union with shape [1]. + out_wrong(Variable): A Tensor with shape [num_classes]. The wrong numbers of each class. + out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class. + + Examples: + + .. code-block:: python + + iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes) + """ + helper = LayerHelper('mean_iou', **locals()) + dtype = helper.input_dtype() + out_mean_iou = helper.create_tmp_variable(dtype='float32') + out_wrong = helper.create_tmp_variable(dtype='int32') + out_correct = helper.create_tmp_variable(dtype='int32') + helper.append_op( + type="mean_iou", + inputs={"Predictions": input, + "Labels": label}, + outputs={ + "OutMeanIou": out_mean_iou, + "OutWrong": out_wrong, + "OutCorrect": out_correct + }, + attrs={"num_classes": num_classes}) + return out_mean_iou, out_wrong, out_correct + + +def crop(x, shape=None, offsets=None, name=None): + """ + Crop input into output, as specified by offsets and shape. + + .. code-block:: text + + * Case 1: + Given + X = [[0, 1, 2, 0, 0] + [0, 3, 4, 0, 0] + [0, 0, 0, 0, 0]], + and + shape = [2, 2], + offsets = [0, 1], + output is: + Out = [[1, 2], + [3, 4]]. + * Case 2: + Given + X = [[0, 1, 2, 5, 0] + [0, 3, 4, 6, 0] + [0, 0, 0, 0, 0]], + and shape is tensor + shape = [[0, 0, 0] + [0, 0, 0]] + and + offsets = [0, 1], + + output is: + Out = [[1, 2, 5], + [3, 4, 6]]. + + Args: + x (Variable): The input tensor variable. + shape (Variable|list/tuple of integer): The output shape is specified + by `shape`, which can a Variable or a list/tupe of integer. + If a tensor Variable, it's rank must be the same as `x`. This way + is suitable for the case that the output shape may be changed each + iteration. If a list/tupe of integer, it's length must be the same + as the rank of `x` + offsets (Variable|list/tuple of integer|None): Specifies the copping + offsets at each dimension. It can be a Variable or or a list/tupe + of integer. If a tensor Variable, it's rank must be the same as `x`. + This way is suitable for the case that the offsets may be changed + each iteration. If a list/tupe of integer, it's length must be the + same as the rank of `x`. If None, the offsets are 0 at each + dimension. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The cropped tensor variable. + + Raises: + ValueError: If shape is not a list, tuple or Variable. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name="x", shape=[3, 5], dtype="float32") + y = fluid.layers.data(name="y", shape=[2, 3], dtype="float32") + crop = fluid.layers.crop(x, shape=y) + + # or + z = fluid.layers.data(name="z", shape=[3, 5], dtype="float32") + crop = fluid.layers.crop(z, shape=[2, 3]) + + """ + helper = LayerHelper('crop', **locals()) + + if not (isinstance(shape, list) or isinstance(shape, tuple) or \ + isinstance(shape, Variable)): + raise ValueError("The shape should be a list, tuple or Variable.") + + if offsets is None: + offsets = [0] * len(x.shape) + + out = helper.create_tmp_variable(x.dtype) + ipts = {'X': x} + attrs = {} + if isinstance(shape, Variable): + ipts['Y'] = shape + else: + attrs['shape'] = shape + if isinstance(offsets, Variable): + ipts['Offsets'] = offsets + else: + attrs['offsets'] = offsets + + helper.append_op( + type='crop', + inputs=ipts, + outputs={'Out': out}, + attrs=None if len(attrs) == 0 else attrs) + return out + + +def rank_loss(label, left, right, name=None): + """ + **Rank loss layer for RankNet** + + RankNet(http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf) + is a pairwise ranking model with a training sample consisting of a pair + of documents, A and B. Label P indicates whether A is ranked higher than B + or not: + + P = {0, 1} or {0, 0.5, 1}, where 0.5 means that there is no information + about the rank of the input pair. + + Rank loss layer takes three inputs: left (o_i), right (o_j) and + label (P_{i,j}). The inputs respectively represent RankNet's output scores + for documents A and B and the value of label P. The following equation + computes rank loss C_{i,j} from the inputs: + + $$ + C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\ + o_{i,j} = o_i - o_j \\ + \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \} + $$ + + Rank loss layer takes batch inputs with size batch_size (batch_size >= 1). + + Args: + label (Variable): Indicats whether A ranked higher than B or not. + left (Variable): RankNet's output score for doc A. + right (Variable): RankNet's output score for doc B. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + list: The value of rank loss. + + Raises: + ValueError: Any of label, left, and right is not a variable. + + Examples: + + .. code-block:: python + + label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32") + left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32") + right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32") + out = fluid.layers.rank_loss(label, left, right) + + + """ + helper = LayerHelper('rank_loss', **locals()) + + if not (isinstance(label, Variable)): + raise ValueError("The label should be a Variable") + + if not (isinstance(left, Variable)): + raise ValueError("The left should be a Variable") + + if not (isinstance(right, Variable)): + raise ValueError("The right should be a Variable") + + out = helper.create_tmp_variable("float32") + + helper.append_op( + type='rank_loss', + inputs={"Label": label, + "Left": left, + "Right": right}, + outputs={'Out': out}) + return out + + +def flatten(x, axis=1, name=None): + """ + **Flatten layer** + Flattens the input tensor into a 2D matrix. + + Examples: + Case 1: + Given + X.shape = (3, 100, 100, 4) + and + axis = 2 + We get: + Out.shape = (3 * 100, 4 * 100) + + Case 2: + Given + X.shape = (3, 100, 100, 4) + and + axis = 0 + We get: + Out.shape = (1, 3 * 100 * 100 * 4) + + Args: + x (Variable): A tensor of rank >= axis. + axis (int): Indicate up to which input dimensions (exclusive) should + be flattened to the outer dimension of the output. + The value for axis must be in the range [0, R], where R + is the rank of the input tensor. When axis = 0, the shape + of the output tensor is (1, (d_0 X d_1 ... d_n), where the + shape of the input tensor is (d_0, d_1, ... d_n). + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: A 2D tensor with the contents of the input tensor, with input + dimensions up to axis flattened to the outer dimension of + the output and remaining input dimensions flattened into the + inner dimension of the output. + + Raises: + ValueError: If x is not a variable. + ValueError: If axis is not in range [0, rank(x)]. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name="x", shape=[4, 4, 3], dtype="float32") + out = fluid.layers.flatten(x=x, axis=2) + """ + helper = LayerHelper('flatten', **locals()) + + if not (isinstance(x, Variable)): + raise ValueError("The input x should be a Variable") + + if not (isinstance(axis, int)) or axis > len(x.shape) or axis < 0: + raise ValueError("The axis should be a int, and in range [0, rank(x)]") + + out = helper.create_tmp_variable(x.dtype) + helper.append_op( + type='flatten', + inputs={"X": x}, + outputs={'Out': out}, + attrs={"axis": axis}) return out diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index a9fe25744cc0b385479c9366af1b731ec221dd5a..f70c7f2258ce588444cf46d6c8affc4c9555203e 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -11,13 +11,12 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from layer_function_generator import generate_layer_fn +from .layer_function_generator import generate_layer_fn __activations__ = [ 'sigmoid', 'logsigmoid', 'exp', - 'relu', 'tanh', 'tanh_shrink', 'softshrink', @@ -29,7 +28,6 @@ __activations__ = [ 'sin', 'round', 'reciprocal', - 'log', 'square', 'softplus', 'softsign', @@ -40,8 +38,6 @@ __activations__ = [ 'relu6', 'pow', 'stanh', - 'hard_shrink', - 'thresholded_relu', 'hard_sigmoid', 'swish', ] @@ -64,14 +60,100 @@ __all__ = [ 'logical_or', 'logical_xor', 'logical_not', - 'uniform_random', 'uniform_random_batch_size_like', 'gaussian_random', 'gaussian_random_batch_size_like', - 'cumsum', 'scatter', 'sum', + 'slice', + 'shape', + 'maxout', ] + __activations__ for _OP in set(__all__): globals()[_OP] = generate_layer_fn(_OP) + +__all__ += ["uniform_random"] + +_uniform_random_ = generate_layer_fn('uniform_random') + + +def uniform_random(shape, dtype=None, min=None, max=None, seed=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + return _uniform_random_(**kwargs) + + +uniform_random.__doc__ = _uniform_random_.__doc__ + """ +Examples: + + >>> result = fluid.layers.uniform_random(shape=[32, 784]) +""" + +__all__ += ['hard_shrink'] + +_hard_shrink_ = generate_layer_fn('hard_shrink') + + +def hard_shrink(x, threshold=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + return _hard_shrink_(**kwargs) + + +hard_shrink.__doc__ = _hard_shrink_.__doc__ + """ +Examples: + + >>> data = fluid.layers.data(name="input", shape=[784]) + >>> result = fluid.layers.hard_shrink(x=data, threshold=0.3) +""" + +__all__ += ['cumsum'] + +_cum_sum_ = generate_layer_fn('cumsum') + + +def cumsum(x, axis=None, exclusive=None, reverse=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + + return _cum_sum_(**kwargs) + + +cumsum.__doc__ = _cum_sum_.__doc__ + """ +Examples: + + >>> data = fluid.layers.data(name="input", shape=[32, 784]) + >>> result = fluid.layers.cumsum(data, axis=0) +""" + +__all__ += ['thresholded_relu'] + +_thresholded_relu_ = generate_layer_fn('thresholded_relu') + + +def thresholded_relu(x, threshold=None): + kwargs = dict() + for name in locals(): + val = locals()[name] + if val is not None: + kwargs[name] = val + + _thresholded_relu_(**kwargs) + + +thresholded_relu.__doc__ = _thresholded_relu_.__doc__ + """ +Examples: + + >>> data = fluid.layers.data(name="input", shape=[1]) + >>> result = fluid.layers.thresholded_relu(data, threshold=0.4) +""" diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index be34cc81a5d5ca0e781e5984b6c3eeaa4e25eb90..b93d721c12cb6ead044dc790f2f2af8a61a63b60 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -6,7 +6,7 @@ # # http://www.apache.org/licenses/LICENSE-2.0 # -# Unless required by applicable law or agreed to in writing, software +# Unlessf required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and @@ -18,6 +18,7 @@ from ..framework import convert_np_dtype_to_dtype_ from ..framework import Variable from ..initializer import Constant, force_init_on_cpu from ..core import VarDesc +from .layer_function_generator import templatedoc import numpy __all__ = [ @@ -30,12 +31,34 @@ __all__ = [ 'assign', 'fill_constant_batch_size_like', 'fill_constant', + 'argmin', + 'argmax', + 'argsort', 'ones', 'zeros', + 'reverse', ] def create_tensor(dtype, name=None, persistable=False): + """ + Create an variable, which will hold a LoDTensor with data type dtype. + + Args: + dtype(string): 'float32'|'int32'|..., the data type of the + created tensor. + name(string): The name of the created tensor, if not set, + the name will be a random unique one. + persistable(bool): Set the persistable flag of the create tensor. + + Returns: + Variable: The tensor variable storing the created tensor. + + Examples: + .. code-block:: python + + tensor = fluid.layers.create_tensor(dtype='float32') + """ helper = LayerHelper("create_tensor", **locals()) return helper.create_variable( name=helper.name, dtype=dtype, persistable=persistable) @@ -48,7 +71,12 @@ def create_parameter(shape, is_bias=False, default_initializer=None): """ - Create a parameter + Create a parameter. The parameter is a learnable variable, which can have + gradient, and can be optimized. + + NOTE: this is a very low-level API. This API is useful when you create + operator by your self. instead of using layers. + Args: shape(list[int]): shape of the parameter dtype(string): element type of the parameter @@ -60,7 +88,12 @@ def create_parameter(shape, default_initializer(Initializer): initializer for the parameter Returns: - Parameter: the created parameter + the created parameter. + + Examples: + >>> W = fluid.layers.create_parameter(shape=[784, 200], dtype='float32') + >>> data = fluid.layers.data(name="img", shape=[64, 784], append_batch_size=False) + >>> hidden = fluid.layers.matmul(x=data, y=W) """ helper = LayerHelper("create_parameter", **locals()) if attr is None: @@ -76,16 +109,29 @@ def create_global_var(shape, force_cpu=False, name=None): """ - Create a global variable. such as global_step + Create a new variable in the global block(block 0). + Args: shape(list[int]): shape of the variable - value(float): the value of the variable - dtype(string): element type of the parameter - persistable(bool): if this variable is persistable - force_cpu(bool): force this variable to be on CPU + value(float): the value of the variable. The new created + variable will be filled with it. + dtype(string): data type of the variable + persistable(bool): if this variable is persistable. + Default: False + force_cpu(bool): force this variable to be on CPU. + Default: False + name(str|None): The name of the variable. If set to None the variable + name will be generated automatically. + Default: None Returns: Variable: the created Variable + + Examples: + .. code-block:: python + + var = fluid.create_global_var(shape=[2,3], value=1.0, dtype='float32', + persistable=True, force_cpu=True, name='new_var') """ helper = LayerHelper("global_var", **locals()) var = helper.create_global_variable( @@ -98,8 +144,21 @@ def create_global_var(shape, def cast(x, dtype): """ - This function takes in the input with input_dtype - and casts it to the output_dtype as the output. + This layer takes in the Variable :attr:`x` with :attr:`x.dtype` and casts + it to the output with :attr:`dtype`. + + Args: + x (Variable): The input Variable for casting. + dtype(np.dtype|core.VarDesc.VarType|str): Data type of the output Variable. + + Returns: + Variable: The output Variable after casting. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name='x', shape=[13], dtype='float32') + result = fluid.layers.cast(x=data, dtype='float64') """ helper = LayerHelper('cast', **locals()) out = helper.create_tmp_variable(dtype=dtype) @@ -130,7 +189,8 @@ def concat(input, axis=0, name=None): Examples: .. code-block:: python - out = fluid.layers.concat(input=[Efirst, Esecond, Ethird, Efourth]) + + out = fluid.layers.concat(input=[Efirst, Esecond, Ethird, Efourth]) """ helper = LayerHelper('concat', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) @@ -143,19 +203,21 @@ def concat(input, axis=0, name=None): def sums(input, out=None): - """This function performs the sum operation on the input and returns the + """ + This function performs the sum operation on the input and returns the result as the output. Args: input (Variable|list): The input tensor that has the elements that need to be summed up. + out (Variable|None): Output parameter. The sum result. + Default: None Returns: - Variable: The tensor type variable that has the sum of input - written to it. + Variable: the sum of input. The same as the argument 'out' Examples: - .. code-block::python + .. code-block:: python tmp = fluid.layers.zeros(shape=[10], dtype='int32') i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) @@ -169,11 +231,15 @@ def sums(input, out=None): helper = LayerHelper('sum', **locals()) if out is None: out = helper.create_tmp_variable(dtype=helper.input_dtype()) - helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out}) + helper.append_op( + type='sum', + inputs={'X': input}, + outputs={'Out': out}, + attrs={'use_mkldnn': False}) return out -def assign(input, output): +def assign(input, output=None): """ **Assign** @@ -181,18 +247,21 @@ def assign(input, output): Args: input(Variable|numpy.ndarray): The source variable - output(Variable): The destination variable + output(Variable|None): The destination variable Returns: Variable: The destination variable that was supplied as the *output*. Examples: .. code-block:: python + out = fluid.layers.create_tensor(dtype='float32') hidden = fluid.layers.fc(input=data, size=10) fluid.layers.assign(hidden, out) """ helper = LayerHelper('assign', **locals()) + if output is None: + output = helper.create_tmp_variable(dtype=input.dtype) if isinstance(input, Variable): helper.append_op( type='assign', inputs={'X': [input]}, outputs={'Out': [output]}) @@ -266,6 +335,7 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None): return out +@templatedoc() def fill_constant_batch_size_like(input, shape, dtype, @@ -273,30 +343,28 @@ def fill_constant_batch_size_like(input, input_dim_idx=0, output_dim_idx=0): """ - **fill_constant_batch_size_like** - - This function creates a tensor of specified *shape*, *dtype* and batch size, - and initializes this with a constant supplied in *value*. The batch size is - obtained from the `input` tensor. + ${comment} It also sets *stop_gradient* to True. + >>> data = fluid.layers.fill_constant_batch_size_like( + >>> input=like, shape=[1], value=0, dtype='int64') + Args: - input(Variable): Tensor whose dimensions will be used to get batch size - shape(tuple|list|None): Shape of output tensor - dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor - value(float): Constant value to initialize the output tensor - input_dim_idx(int): Index of input's batch size dimension - output_dim_idx(int): Index of output's batch size dimension + input(${input_type}): ${input_comment}. - Returns: - Variable: The tensor variable storing the output + shape(${shape_type}): ${shape_comment}. - Examples: - .. code-block:: python + dtype(${dtype_type}): ${dtype_comment}. + + value(${value_type}): ${value_comment}. + + input_dim_idx(${input_dim_idx_type}): ${input_dim_idx_comment}. + + output_dim_idx(${output_dim_idx_type}): ${output_dim_idx_comment}. - data = fluid.layers.fill_constant_batch_size_like( - input=like, shape=[1], value=0, dtype='int64') + Returns: + ${out_comment}. """ helper = LayerHelper("fill_constant_batch_size_like", **locals()) out = helper.create_tmp_variable(dtype=dtype) @@ -315,6 +383,120 @@ def fill_constant_batch_size_like(input, return out +def argmin(x, axis=0): + """ + **argmin** + + This function computes the indices of the min elements + of the input tensor's element along the provided axis. + + Args: + x(Variable): The input to compute the indices of + the min elements. + axis(int): Axis to compute indices along. + + Returns: + Variable: The tensor variable storing the output + + Examples: + .. code-block:: python + + out = fluid.layers.argmin(x=in, axis=0) + out = fluid.layers.argmin(x=in, axis=-1) + """ + helper = LayerHelper("arg_min", **locals()) + out = helper.create_tmp_variable(VarDesc.VarType.INT64) + helper.append_op( + type='arg_min', + inputs={'X': x}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + +def argmax(x, axis=0): + """ + **argmax** + + This function computes the indices of the max elements + of the input tensor's element along the provided axis. + + Args: + x(Variable): The input to compute the indices of + the max elements. + axis(int): Axis to compute indices along. + + Returns: + Variable: The tensor variable storing the output + + Examples: + .. code-block:: python + + out = fluid.layers.argmax(x=in, axis=0) + out = fluid.layers.argmax(x=in, axis=-1) + """ + helper = LayerHelper("arg_max", **locals()) + out = helper.create_tmp_variable(VarDesc.VarType.INT64) + helper.append_op( + type='arg_max', + inputs={'X': x}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + +def argsort(input, axis=-1, name=None): + """ + Performs sorting on the input Variable along the given axis, and outputs + sorted data Varibale and its corresponding index Variable with the same + shape as :attr:`input`. + + .. code-block:: text + + For example, the given axis is -1 and the input Variable + + input = [[0.15849551, 0.45865775, 0.8563702 ], + [0.12070083, 0.28766365, 0.18776911]], + + after argsort, the sorted Vairable becomes + + out = [[0.15849551, 0.45865775, 0.8563702 ], + [0.12070083, 0.18776911, 0.28766365]], + + and the sorted indices along the given axis turn outs to be + + indices = [[0, 1, 2], + [0, 2, 1]] + + Args: + input(Variable): The input Variable for sorting. + axis(int): The axis along which to sort the input Variable. When + :attr:`axis` < 0, the actual axis will be :attr:`axis` + + rank(:attr:`input`). Default -1, the last dimension. + name(str|None): (optional) A name for this layer. If set None, the + layer will be named automatically. + + Returns: + tuple: A tuple of sorted data Variable and the sorted indices. + + Examples: + .. code-block:: python + + input = fluid.layers.data(data=[2, 3]) + out, indices = fluid.layers.argsort(input, axis=0) + """ + helper = LayerHelper("argsort", **locals()) + out = helper.create_tmp_variable(dtype=input.dtype, stop_gradient=True) + ids = helper.create_tmp_variable(VarDesc.VarType.INT64, stop_gradient=True) + helper.append_op( + type='argsort', + inputs={'X': input}, + outputs={'Out': out, + 'Indices': ids}, + attrs={'axis': axis}) + return out, ids + + def ones(shape, dtype, force_cpu=False): """ **ones** @@ -349,11 +531,12 @@ def zeros(shape, dtype, force_cpu=False): It also sets *stop_gradient* to True. Args: - shape(tuple|list|None): Shape of output tensor - dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor + shape(tuple|list|None): Shape of output tensor. + dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor. + force_cpu(bool, default False): Whether to make output stay on CPU. Returns: - Variable: The tensor variable storing the output + Variable: The tensor variable storing the output. Examples: .. code-block:: python @@ -363,6 +546,40 @@ def zeros(shape, dtype, force_cpu=False): return fill_constant(value=0.0, **locals()) +def reverse(x, axis): + """ + **reverse** + + This function reverse the input 'x' along given axises. + + Args: + x(Vairbale): the input to be reversed. + axis(int|tuple|list): Axis that along which order of elements + is reversed. If it is a tuple or a list, reversing + will be apply on each axis in the tuple or list. + + Returns: + Variable: The reversed tensor. + + Examples: + .. code-block:: python + + out = fluid.layers.reverse(x=in, axis=0) + # or: + out = fluid.layers.reverse(x=in, axis=[0,1]) + """ + if isinstance(axis, int): + axis = [axis] + helper = LayerHelper("reverse", **locals()) + out = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type='reverse', + inputs={'Input': x}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + def save(x, file_path, overwrite=True): """ Saves a variable as a file. @@ -370,9 +587,9 @@ def save(x, file_path, overwrite=True): Args: x(variable): The Tensor/LoDTensor to be saved. file_path(str): The file path where the variable will be saved. - overwrite(bool): Whether or not cover the given file when it has already - existed. If it's set 'False' and the file is existed, a runtime - error will be thrown. + overwrite(bool): Whether or not cover the given file when it has already + existed. If it's set 'False' and the file is existed, a runtime + error will be thrown. """ helper = LayerHelper("save", **locals()) helper.append_op( @@ -388,11 +605,27 @@ def save_combine(x, file_path, overwrite=True): Saves a list of variables into a single file. Args: - x(list): A list of Tensor/LoDTensor to be saved together in a single file. + x(list): A list of Tensor/LoDTensor variables to be saved together in + a single file. file_path(str): The file path where variables will be saved. - overwrite(bool): Whether or not cover the given file when it has already - existed. If it's set 'False' and the file is existed, a runtime - error will be thrown. + overwrite(bool): Whether or not cover the given file when it has already + existed. If it's set 'False' and the file is existed, a runtime + error will be thrown. + + Returns: + There is no return value. + + Examples: + + .. code-block:: python + + v1 = fluid.layers.data(name="data", + shape=(4, 6), + dtype="float32") + v2 = fluid.layers.data(name="data", + shape=(6, 8, 4), + dtype="float32") + normed = fluid.layers.save_combine([v1, v2], file_path="output") """ helper = LayerHelper("save_combine", **locals()) helper.append_op( @@ -403,22 +636,6 @@ def save_combine(x, file_path, overwrite=True): "overwrite": overwrite}) -def load(out, file_path): - """ - Loads a variable from a given file. - - Args: - out(variable): The variable to be read from the disk file. - file_path(str): The path of the disk file. - """ - helper = LayerHelper("load", **locals()) - helper.append_op( - type="load", - inputs={}, - output={"Out": out}, - args={"file_path": file_path}) - - def load_combine(out, file_path): """ Loads a list of vairables from a single file. diff --git a/python/paddle/fluid/lod_tensor.py b/python/paddle/fluid/lod_tensor.py index 9946d0a4ff33b2f5040f6d2e31aa20fcf9c609a7..53c33616f55be5f5ef7068a6e94418e17d739e3c 100644 --- a/python/paddle/fluid/lod_tensor.py +++ b/python/paddle/fluid/lod_tensor.py @@ -12,178 +12,123 @@ # See the License for the specific language governing permissions and # limitations under the License. -import core +from . import core import numpy as np __all__ = ['create_lod_tensor', 'create_random_int_lodtensor'] -def _validate_lod(lod, tensor_height=-1): - """Check whether the input length-based lod info is valid. - - There are several things to check: - 1. lod should be a list of lists. Empty list is fine. - 2. The length of each sublist (a lod level) should be at least one. - 3. Each element in each lod level should be an integer greater than 0. - 4. The sum of one lod level should be equal to the length of the next lod level. - 5. The sum of the last lod level should be equal to the tensor height. - Bypass this check if user does not provide tensor_height as input. - - Args: - lod: the length-based lod info, e.g., [[2, 3], [2, 1, 2, 3, 4]]. - tensor_height: the outermost dimension of the tensor with which the input - lod is associated with. - - Returns: - A boolean indicating whether the input lod is valid or not. +def create_lod_tensor(data, recursive_seq_lens, place): """ - assert isinstance(lod, list), "lod should be a list" - # Empty lod is fine - if len(lod) == 0: - return True - - lod_sum = [] - for level in lod: - assert isinstance(level, list), "each item in lod should be a list" - # Each level of lod should have at least one length info - if len(level) < 1: - return False - level_sum = 0 - for lod_len in level: - # Each length in a level should be > 0 - if lod_len <= 0: - return False - level_sum += lod_len - lod_sum.append(level_sum) - - for idx, val in enumerate(lod_sum[:-1]): - # Each level's sum should be equal to - # the number of items in the next level - if val != len(lod[idx + 1]): - return False - - if tensor_height == -1: - return True - else: - # Last level's sum should be equal to the tensor height - return lod_sum[-1] == tensor_height + Create a lod tensor from a numpy array, a list, or an existing lod tensor. + Create a lod tensor by doing the following: -def _convert_lod(lod): - """Convert a length-based lod to a offset-based lod. + 1. Check that the length-based level of detail (LoD) also known as + recursive_sequence_lengths of the input is valid. - If the length-based lod is [[2, 3], [2, 1, 2, 3, 4]], - then the offset-based lod is [[0, 2, 5], [0, 2, 3, 5, 8, 12]]. + 2. Convert recursive_sequence_lengths to a offset-based LoD. - Args: - lod: a length-based lod info. - - Returns: - A list of lists as the offset-based lod converted to from the input lod. - """ - new_lod = [] - for level in lod: - cur_len = 0 - new_level = [cur_len] - for lod_len in level: - cur_len += lod_len - new_level.append(cur_len) - new_lod.append(new_level) - return new_lod + 3. Copy the data from a numpy array, a list or a existing lod tensor to + CPU or GPU device (based on input place). + 4. Set the level of detail (LoD) using the offset-based LoD. -def create_lod_tensor(data, lod, place): - """Create a lod tensor from a numpy array, a list, or an existing lod tensor. + Examples: - Create a lod tensor by doing the following: - 1. Check that the length-based input lod is valid. - 2. Convert the length-based lod to a offset-based LoD. - 3. Copy the data from a numpy array, a list or a existing lod tensor to - CPU or GPU device (based on input place). - 4. Set the level of detail (LoD) using the offset-based LoD. - - Use example: - Suppose we want LoDTensor to hold data for sequences of word, where each word is - represented by an integer. If we want to create a LoDTensor to represent two - sentences, one of 2 words, and one of 3 words. + Suppose we want LoDTensor to hold data for sequences of word, where each + word is represented by an integer. If we want to create a LoDTensor to + represent two sentences, one of 2 words, and one of 3 words. - Then 'data' can be a numpy array of integers with shape (5, 1). - 'lod' will be [[2, 3]], indicating the length(# of words) in each sentence. - This length-based input lod [[2, 3]] will be converted to offset-based lod [[0, 2, 5]] - inside the function call. + Then :code:`data` can be a numpy array of integers with shape (5, 1). + :code:`recursive_seq_lens` will be [[2, 3]], indicating the length(# of words) in each + sentence. This length-based :code:`recursive_seq_lens` [[2, 3]] will be converted to + offset-based LoD [[0, 2, 5]] inside the function call. - Please refer to - github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md - for more details regarding LoD. + Please reference :ref:`api_guide_low_level_lod_tensor` for more details + regarding LoD. Args: - data: a numpy array or a LoDTensor or a list holding the data to be copied. - lod: a list of lists indicating the length-based LoD info specified by the user. - place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. + data(numpy.ndarray|list|LoDTensor): a numpy array or a LoDTensor or a + list holding the data to be copied. + recursive_seq_lens(list): a list of lists indicating the length-based level of detail + info specified by the user. + place(Place): CPU or GPU place indicating where the data in the new + LoDTensor will be stored. Returns: - A fluid LoDTensor object with tensor data and lod info. + A fluid LoDTensor object with tensor data and recursive_seq_lens info. """ if isinstance(data, core.LoDTensor): - return create_lod_tensor(np.array(data), lod, place) + return create_lod_tensor(np.array(data), recursive_seq_lens, place) elif isinstance(data, list): - # When input data is a list, it only deal with the case where the base element - # is an index of shape [1] and dtype int64 (e.g., word id). Hence, the generated - # LoDTensor will be of shape [n, 1] and dtype int64, where `n` is the total number - # of words or other indexes in the sequence. - new_lod = [] + # When input data is a list, it only deal with the case where the base element + # is an index of shape [1] and dtype int64 (e.g., word id). Hence, the generated + # LoDTensor will be of shape [n, 1] and dtype int64, where `n` is the total number + # of words or other indexes in the sequence. + new_recursive_seq_lens = [] for seq in data: - new_lod.append(len(seq)) - assert [new_lod] == lod, "data and lod do not match" + new_recursive_seq_lens.append(len(seq)) + assert [ + new_recursive_seq_lens + ] == recursive_seq_lens, "data and recursive_seq_lens do not match" flattened_data = np.concatenate(data, axis=0).astype("int64") flattened_data = flattened_data.reshape([len(flattened_data), 1]) - return create_lod_tensor(flattened_data, lod, place) + return create_lod_tensor(flattened_data, recursive_seq_lens, place) elif isinstance(data, np.ndarray): - assert _validate_lod(lod, - data.shape[0]), "the provided lod info is invalid" tensor = core.LoDTensor() tensor.set(data, place) - tensor.set_lod(_convert_lod(lod)) + tensor.set_recursive_sequence_lengths(recursive_seq_lens) + assert tensor.has_valid_recursive_sequence_lengths( + ), "the provided lod info is invalid" return tensor else: raise TypeError( "data should be either a LoDTensor, a Numpy array or a list") -def create_random_int_lodtensor(lod, base_shape, place, low, high): - """Create a LoDTensor containing random integers. +def create_random_int_lodtensor(recursive_seq_lens, base_shape, place, low, + high): + """ + Create a LoDTensor containing random integers. - This function is frequently used in the book examples. So we revised it based on - the new create_lod_tensor API and put it here in the lod_tensor module to simplify - the code. + This function is frequently used in the book examples. So we revised it + based on the new create_lod_tensor API and put it here in the lod_tensor + module to simplify the code. The function does the following: - 1. Calculate the overall shape of the LoDTensor based on the length-based 'lod' input - and the shape of the basic element in 'base_shape'. + + 1. Calculate the overall shape of the LoDTensor based on the length-based + :code:`recursive_seq_lens` input and the shape of the basic element in + :code:`base_shape`. + 2. Create a numpy array of this shape. + 3. Create the LoDTensor using create_lod_tensor API. - Suppose we want LoDTensor to hold data for sequences of word, where each word is - represented by an integer. If we want to create a LoDTensor to represent two - sentences, one of 2 words, and one of 3 words. Then 'base_shape' is [1], input - length-based 'lod' is [[2, 3]]. Then the overall shape of the LoDTensor would be - [5, 1], holding 5 words for two sentences. + Suppose we want LoDTensor to hold data for sequences of word, where each + word is represented by an integer. If we want to create a LoDTensor to + represent two sentences, one of 2 words, and one of 3 words. Then + 'base_shape' is [1], input length-based 'recursive_seq_lens' is [[2, 3]]. + Then the overall shape of the LoDTensor would be [5, 1], holding 5 words + for two sentences. Args: - data: a numpy array or a LoDTensor holding the data to be copied. - lod: a list of lists indicating the length-based LoD info specified by the user. - base_shape: the shape of the basic element to be held by the LoDTensor. - place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. - low: the lower bound of the random integers. - high: the upper bound of the random integers. + recursive_seq_lens(list): a list of lists indicating the length-based + level of detail info specified by the user. + base_shape(list): the shape of the basic element to be held by the + LoDTensor. + place(Place): CPU or GPU place indicating where the data in the new + LoDTensor will be stored. + low(int): the lower bound of the random integers. + high(int): the upper bound of the random integers. Returns: - A fluid LoDTensor object with tensor data and lod info. + A fluid LoDTensor object with tensor data and recursive_seq_lens info. """ assert isinstance(base_shape, list), "base_shape should be a list" - converted_lod = _convert_lod(lod) # append the total number of basic elements to the front of its shape - overall_shape = [converted_lod[-1][-1]] + base_shape - # the range of integer data elements is [low, high] + overall_shape = [sum(recursive_seq_lens[-1])] + base_shape + # the range of integer data elements is [low, high] data = np.random.random_integers(low, high, overall_shape).astype("int64") - return create_lod_tensor(data, lod, place) + return create_lod_tensor(data, recursive_seq_lens, place) diff --git a/python/paddle/fluid/metrics.py b/python/paddle/fluid/metrics.py index bb9c6fdc60089fc2b43573a6421a6f9781d2d4a8..cd8934522755691217a99a2cca271badda55368e 100644 --- a/python/paddle/fluid/metrics.py +++ b/python/paddle/fluid/metrics.py @@ -23,6 +23,8 @@ import warnings __all__ = [ 'MetricBase', 'CompositeMetric', + 'Precision', + 'Recall', 'Accuracy', 'ChunkEvaluator', 'EditDistance', @@ -46,40 +48,41 @@ def _is_number_or_matrix_(var): class MetricBase(object): """ - Base Class for all evaluators + Base Class for all Metrics. + MetricBase define a group of interfaces for the + model evaluation methods. Metrics accumulate metric states between + consecutive minibatches, at every minibatch, use update + interface to add current minibatch value to global states. + Use eval to compute accumative metric value from last reset() + or from scratch on. + If you need to custom a new metric, please inherit from MetricBase and + custom implementation. Args: - name(str): The name of evaluator. such as, "accuracy". Used for generate - temporary variable name. - Interface: - Note(*) : the states is the attributes who not has _ prefix. - - get_config(): print current states and configuration - reset(): clear the states. If the Metrics states type is not (int, float, np.ndarray), - Please override this method. - update(): update states at every minibatch - eval(): get metric evaluation in numpy type. + name(str): The name of metric instance. such as, "accuracy". + It needed if you want to distinct different metrics in a model. + """ - def __init__(self, name, **kwargs): + def __init__(self, name): self._name = str(name) if name != None else self.__class__.__name__ - self._kwargs = kwargs if kwargs != None else dict() - self.reset() def __str__(self): return self._name def reset(self): """ - states is the attributes who not has _ prefix. - reset the states of metrics. + reset clear the states of metrics. By default, the states + are the members who do not has _ prefix, reset set them to inital states. + If you violate the implicit name rule, please also custom the reset + interface. """ states = { attr: value - for attr, value in self.__dict__.iteritems() + for attr, value in list(self.__dict__.items()) if not attr.startswith("_") } - for attr, value in states.iteritems(): + for attr, value in list(states.items()): if isinstance(value, int): setattr(self, attr, 0) elif isinstance(value, float): @@ -90,61 +93,231 @@ class MetricBase(object): setattr(self, attr, None) def get_config(self): + """ + Get the metric and current states. + The states are the members who do not has "_" prefix. + + Args: + None + + Returns: + dict: a dict of metric and states + """ states = { attr: value - for attr, value in self.__dict__.iteritems() + for attr, value in list(self.__dict__.items()) if not attr.startswith("_") } - config = copy.deepcopy(self._kwargs) + config = {} config.update({"name": self._name, "states": copy.deepcopy(states)}) return config - def update(self): - raise NotImplementedError() + def update(self, preds, labels): + """ + Updates the metric states at every minibatch. + One user can compute the minibatch metric via pure Python, or + via a c++ operator. + + Args: + preds(numpy.array): the predictions of current minibatch + labels(numpy.array): the labels of current minibatch, if the label is one-hot + or soft-label, should custom the corresponding update rule. + """ + raise NotImplementedError( + "Should not use it directly, please extend it.") def eval(self): - raise NotImplementedError() + """ + Evalute the current metrics based the accumulated states. + + Returns: + float|list(float)|numpy.array: the metrics via Python. + """ + raise NotImplementedError( + "Should not use it directly, please extend it.") class CompositeMetric(MetricBase): """ - Compute multiple metrics in each minibatch. + Composite multiple metrics in one instance. for example, merge F1, accuracy, recall into one Metric. + + Examples: + .. code-block:: python + + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + comp = fluid.metrics.CompositeMetric() + acc = fluid.metrics.Precision() + recall = fluid.metrics.Recall() + comp.add_metric(acc) + comp.add_metric(recall) + for pass in range(PASSES): + comp.reset() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + comp.update(preds=preds, labels=labels) + numpy_acc, numpy_recall = comp.eval() """ - def __init__(self, name=None, **kwargs): - super(CompositeMetric, self).__init__(name, kwargs) + def __init__(self, name=None): + super(CompositeMetric, self).__init__(name) self._metrics = [] def add_metric(self, metric): + """ + add one metric instance to CompositeMetric. + + Args: + metric: a instance of MetricBase. + """ if not isinstance(metric, MetricBase): raise ValueError("SubMetric should be inherit from MetricBase.") self._metrics.append(metric) + def update(self, preds, labels): + """ + Update every metrics in sequence. + + Args: + preds(numpy.array): the predictions of current minibatch + labels(numpy.array): the labels of current minibatch, if the label is one-hot + or soft-label, should custom the corresponding update rule. + """ + for m in self._metrics: + ans.append(m.update(preds, labels)) + def eval(self): + """ + Evaluate every metrics in sequence. + + Returns: + list(float|numpy.array): a list of metrics value in Python. + """ ans = [] for m in self._metrics: ans.append(m.eval()) return ans +class Precision(MetricBase): + """ + Precision (also called positive predictive value) is the fraction of + relevant instances among the retrieved instances. + https://en.wikipedia.org/wiki/Evaluation_of_binary_classifiers + + Note Precision is different with Accuracy in binary classifiers. + accuracy = true positive / total instances + precision = true positive / all positive instance + + Examples: + .. code-block:: python + + metric = fluid.metrics.Precision() + for pass in range(PASSES): + metric.reset() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(preds=preds, labels=labels) + numpy_precision = metric.eval() + """ + + def __init__(self, name=None): + super(Precision, self).__init__(name) + self.tp = 0 # true positive + self.fp = 0 # false positive + + def update(self, preds, labels): + if not _is_numpy_(preds): + raise ValueError("The 'preds' must be a numpy ndarray.") + if not _is_numpy_(labels): + raise ValueError("The 'labels' must be a numpy ndarray.") + sample_num = labels[0] + for i in range(sample_num): + pred = preds[i].astype("int32") + label = labels[i] + if label == 1: + if pred == label: + self.tp += 1 + else: + self.fp += 1 + + def eval(self): + ap = self.tp + self.fp + return float(self.tp) / ap if ap != 0 else .0 + + +class Recall(MetricBase): + """ + Recall (also known as sensitivity) is the fraction of + relevant instances that have been retrieved over the + total amount of relevant instances + + https://en.wikipedia.org/wiki/Precision_and_recall + + Examples: + .. code-block:: python + + metric = fluid.metrics.Recall() + for pass in range(PASSES): + metric.reset() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(preds=preds, labels=labels) + numpy_recall = metric.eval() + """ + + def __init__(self, name=None): + super(Recall, self).__init__(name) + self.tp = 0 # true positive + self.fn = 0 # false negtive + + def update(self, preds, labels): + if not _is_numpy_(preds): + raise ValueError("The 'preds' must be a numpy ndarray.") + if not _is_numpy_(labels): + raise ValueError("The 'labels' must be a numpy ndarray.") + sample_num = labels[0] + for i in range(sample_num): + pred = preds[i].astype("int32") + label = labels[i] + if label == 1: + if pred == label: + self.tp += 1 + else: + if pred != label: + self.fn += 1 + + def eval(self): + recall = self.tp + self.fn + return float(self.tp) / recall if recall != 0 else .0 + + class Accuracy(MetricBase): """ Accumulate the accuracy from minibatches and compute the average accuracy for every pass. + https://en.wikipedia.org/wiki/Accuracy_and_precision Args: name: the metrics name - Example: - minibatch_accuracy = fluid.layers.accuracy(pred, label) - accuracy_evaluator = fluid.metrics.Accuracy() - for epoch in PASS_NUM: - accuracy_evaluator.reset() - for data in batches: - loss = exe.run(fetch_list=[cost, minibatch_accuracy]) - accuracy_evaluator.update(value=minibatch_accuracy, weight=batches) - accuracy = accuracy_evaluator.eval() + Examples: + .. code-block:: python + + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + minibatch_accuracy = fluid.layers.accuracy(pred, label) + accuracy_evaluator = fluid.metrics.Accuracy() + for pass in range(PASSES): + accuracy_evaluator.reset() + for data in train_reader(): + batch_size = data[0] + loss = exe.run(fetch_list=[cost, minibatch_accuracy]) + accuracy_evaluator.update(value=minibatch_accuracy, weight=batch_size) + numpy_acc = accuracy_evaluator.eval() """ def __init__(self, name=None): @@ -153,6 +326,13 @@ class Accuracy(MetricBase): self.weight = .0 def update(self, value, weight): + """ + Update minibatch states. + + Args: + value(float|numpy.array): accuracy of one minibatch. + weight(int|float): batch size. + """ if not _is_number_or_matrix_(value): raise ValueError( "The 'value' must be a number(int, float) or a numpy ndarray.") @@ -163,9 +343,8 @@ class Accuracy(MetricBase): def eval(self): if self.weight == 0: - raise ValueError( - "There is no data in Accuracy Metrics. Please check layers.accuracy output has added to Accuracy." - ) + raise ValueError("There is no data in Accuracy Metrics. \ + Please check layers.accuracy output has added to Accuracy.") return self.value / self.weight @@ -174,6 +353,25 @@ class ChunkEvaluator(MetricBase): Accumulate counter numbers output by chunk_eval from mini-batches and compute the precision recall and F1-score using the accumulated counter numbers. + For some basics of chunking, please refer to + 'Chunking with Support Vector Machines '. + ChunkEvalEvaluator computes the precision, recall, and F1-score of chunk detection, + and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes. + + Examples: + .. code-block:: python + + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks = layers.chunk_eval( + input=pred, + label=label) + metric = fluid.metrics.ChunkEvaluator() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(num_infer_chunks, num_label_chunks, num_correct_chunks) + numpy_precision, numpy_recall, numpy_f1 = metric.eval() """ def __init__(self, name=None): @@ -183,9 +381,17 @@ class ChunkEvaluator(MetricBase): self.num_correct_chunks = 0 def update(self, num_infer_chunks, num_label_chunks, num_correct_chunks): + """ + Update the states based on the layers.chunk_eval() ouputs. + Args: + num_infer_chunks(int|numpy.array): The number of chunks in Inference on the given minibatch. + num_label_chunks(int|numpy.array): The number of chunks in Label on the given mini-batch. + num_correct_chunks(int|float|numpy.array): The number of chunks both in Inference and Label on the + given mini-batch. + """ if not _is_number_or_matrix_(num_infer_chunks): raise ValueError( - "The 'num_infer_chunks' must be a number(int, float) or a numpy ndarray." + "The 'num_infer_chunks' must be a number(int) or a numpy ndarray." ) if not _is_number_or_matrix_(num_label_chunks): raise ValueError( @@ -212,21 +418,28 @@ class ChunkEvaluator(MetricBase): class EditDistance(MetricBase): """ + Edit distance is a way of quantifying how dissimilar two strings + (e.g., words) are to one another by counting the minimum number + of operations required to transform one string into the other. + Refer to https://en.wikipedia.org/wiki/Edit_distance + Accumulate edit distance sum and sequence number from mini-batches and compute the average edit_distance and instance error of all batches. Args: name: the metrics name - Example: - edit_distance_metrics = fluid.layers.edit_distance(input, label) - distance_evaluator = fluid.metrics.EditDistance() - for epoch in PASS_NUM: - distance_evaluator.reset() - for data in batches: - loss = exe.run(fetch_list=[cost] + list(edit_distance_metrics)) - distance_evaluator.update(*edit_distance_metrics) - distance, instance_error = distance_evaluator.eval() + Examples: + .. code-block:: python + + distances, seq_num = fluid.layers.edit_distance(input, label) + distance_evaluator = fluid.metrics.EditDistance() + for epoch in PASS_NUM: + distance_evaluator.reset() + for data in batches: + loss = exe.run(fetch_list=[cost] + list(edit_distance_metrics)) + distance_evaluator.update(distances, seq_num) + distance, instance_error = distance_evaluator.eval() In the above example: 'distance' is the average of the edit distance in a pass. @@ -264,16 +477,38 @@ class EditDistance(MetricBase): class DetectionMAP(MetricBase): """ Calculate the detection mean average precision (mAP). - - TODO (Dang Qingqing): update the following doc. - The general steps are as follows: - 1. calculate the true positive and false positive according to the input - of detection and labels. - 2. calculate mAP value, support two versions: '11 point' and 'integral'. - + mAP is the metric to measure the accuracy of object detectors + like Faster R-CNN, SSD, etc. + It is the average of the maximum precisions at different recall values. Please get more information from the following articles: https://sanchom.wordpress.com/tag/average-precision/ + https://arxiv.org/abs/1512.02325 + + The general steps are as follows: + + 1. calculate the true positive and false positive according to the input + of detection and labels. + 2. calculate mAP value, support two versions: '11 point' and 'integral'. + + Examples: + .. code-block:: python + + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + batch_map = layers.detection_map( + input, + label, + class_num, + background_label, + overlap_threshold=overlap_threshold, + evaluate_difficult=evaluate_difficult, + ap_version=ap_version) + metric = fluid.metrics.DetectionMAP() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, batch_map]) + batch_size = data[0] + metric.update(value=batch_map, weight=batch_size) + numpy_map = metric.eval() """ def __init__(self, name=None): @@ -302,17 +537,18 @@ class DetectionMAP(MetricBase): class Auc(MetricBase): """ - Auc Metrics which adapts to binary classification. - Need to note that auc metrics compute the value via Python natively. + Auc metric adapts to the binary classification. + Refer to https://en.wikipedia.org/wiki/Receiver_operating_characteristic#Area_under_the_curve + Need to note that auc metric compute the value via Python natively. If you concern the speed, please use the fluid.layers.auc instead. The `auc` function creates four local variables, `true_positives`, - `true_negatives`, `false_positives` and `false_negatives` that are used to - compute the AUC. To discretize the AUC curve, a linearly spaced set of - thresholds is used to compute pairs of recall and precision values. The area - under the ROC-curve is therefore computed using the height of the recall - values by the false positive rate, while the area under the PR-curve is the - computed using the height of the precision values by the recall. + `true_negatives`, `false_positives` and `false_negatives` that are used to + compute the AUC. To discretize the AUC curve, a linearly spaced set of + thresholds is used to compute pairs of recall and precision values. The area + under the ROC-curve is therefore computed using the height of the recall + values by the false positive rate, while the area under the PR-curve is the + computed using the height of the precision values by the recall. Args: name: metric name @@ -322,22 +558,32 @@ class Auc(MetricBase): curve. "NOTE: only implement the ROC curve type via Python now." + + Examples: + .. code-block:: python + + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + metric = fluid.metrics.Auc() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(preds, labels) + numpy_auc = metric.eval() """ def __init__(self, name, curve='ROC', num_thresholds=200): - super(MetricBase, self).__init__(name, curve, num_thresholds) + super(Auc, self).__init__(name=name) self._curve = curve self._num_thresholds = num_thresholds self._epsilon = 1e-6 - self.tp_list = np.ndarray((num_thresholds, )) - self.fn_list = np.ndarray((num_thresholds, )) - self.tn_list = np.ndarray((num_thresholds, )) - self.fp_list = np.ndarray((num_thresholds, )) + self.tp_list = np.zeros((num_thresholds, )) + self.fn_list = np.zeros((num_thresholds, )) + self.tn_list = np.zeros((num_thresholds, )) + self.fp_list = np.zeros((num_thresholds, )) - def update(self, labels, predictions, axis=1): + def update(self, preds, labels): if not _is_numpy_(labels): raise ValueError("The 'labels' must be a numpy ndarray.") - if not _is_numpy_(predictions): + if not _is_numpy_(preds): raise ValueError("The 'predictions' must be a numpy ndarray.") kepsilon = 1e-7 # to account for floating point imprecisions @@ -345,17 +591,17 @@ class Auc(MetricBase): for i in range(self._num_thresholds - 2)] thresholds = [0.0 - kepsilon] + thresholds + [1.0 + kepsilon] - # caculate TP, FN, TN, FP count + # calculate TP, FN, TN, FP count for idx_thresh, thresh in enumerate(thresholds): tp, fn, tn, fp = 0, 0, 0, 0 for i, lbl in enumerate(labels): if lbl: - if predictions[i, 0] >= thresh: + if preds[i, 1] >= thresh: tp += 1 else: fn += 1 else: - if predictions[i, 0] >= thresh: + if preds[i, 1] >= thresh: fp += 1 else: tn += 1 diff --git a/python/paddle/fluid/net_drawer.py b/python/paddle/fluid/net_drawer.py index 73946a0721dc4a6d03074a4708cf574951412e66..623a7d3fd05567a26bb6923550f597a0e1e27e32 100644 --- a/python/paddle/fluid/net_drawer.py +++ b/python/paddle/fluid/net_drawer.py @@ -24,7 +24,7 @@ logger = logging.getLogger(__name__) logger.setLevel(logging.INFO) try: - from graphviz import Digraph + from .graphviz import Digraph except ImportError: logger.info( 'Cannot import graphviz, which is required for drawing a network. This ' @@ -77,7 +77,7 @@ def parse_graph(program, graph, var_dict, **kwargs): # fill the known variables for block in program.blocks: for var in block.vars: - if not var_dict.has_key(var): + if var not in var_dict: var_dict[var] = "Feed" temp_id = 0 @@ -93,17 +93,17 @@ def parse_graph(program, graph, var_dict, **kwargs): var_dict[arg] = op.type for e in op.inputs: for arg in e.arguments: - if var_dict.has_key(arg): + if arg in var_dict: graph.edge(**draw_edge(var_dict, op, e, arg)) break # only plot the first block def draw_graph(startup_program, main_program, **kwargs): - if kwargs.has_key("graph_attr"): + if "graph_attr" in kwargs: GRAPH_STYLE.update(kwargs[graph_attr]) - if kwargs.has_key("node_attr"): + if "node_attr" in kwargs: OP_STYLE.update(kwargs[node_attr]) - if kwargs.has_key("edge_attr"): + if "edge_attr" in kwargs: VAR_STYLE.update(kwargs[edge_attr]) graph_id = unique_id() diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index bbedf6fde0872fd32d81c103bf5fe61449b7f57b..08480671d8a5c50bbec97930c451cbcdc241e1fe 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -11,7 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -import layers +from . import layers __all__ = [ "simple_img_conv_pool", @@ -26,16 +26,87 @@ def simple_img_conv_pool(input, filter_size, pool_size, pool_stride, - act, - param_attr=None, + pool_padding=0, pool_type='max', + global_pooling=False, + conv_stride=1, + conv_padding=0, + conv_dilation=1, + conv_groups=1, + param_attr=None, + bias_attr=None, + act=None, use_cudnn=True, use_mkldnn=False): + """ + The simple_img_conv_pool is composed with one Convolution2d and one Pool2d. + + Args: + input (Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of filter. It is as same as the output + feature channel. + filter_size (int|list|tuple): The filter size. If filter_size is a list or + tuple, it must contain two integers, (filter_size_H, filter_size_W). Otherwise, + the filter_size_H = filter_size_W = filter_size. + pool_size (int|list|tuple): The pooling size of Pool2d layer. If pool_size + is a list or tuple, it must contain two integers, (pool_size_H, pool_size_W). + Otherwise, the pool_size_H = pool_size_W = pool_size. + pool_stride (int|list|tuple): The pooling stride of Pool2d layer. If pool_stride + is a list or tuple, it must contain two integers, (pooling_stride_H, pooling_stride_W). + Otherwise, the pooling_stride_H = pooling_stride_W = pool_stride. + pool_padding (int|list|tuple): The padding of Pool2d layer. If pool_padding is a list or + tuple, it must contain two integers, (pool_padding_H, pool_padding_W). + Otherwise, the pool_padding_H = pool_padding_W = pool_padding. Default 0. + pool_type (str): Pooling type can be :math:`max` for max-pooling and :math:`avg` for + average-pooling. Default :math:`max`. + global_pooling (bool): Whether to use the global pooling. If global_pooling = true, + pool_size and pool_padding while be ignored. Default False + conv_stride (int|list|tuple): The stride size of the Conv2d Layer. If stride is a + list or tuple, it must contain two integers, (conv_stride_H, conv_stride_W). Otherwise, + the conv_stride_H = conv_stride_W = conv_stride. Default: conv_stride = 1. + conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is + a list or tuple, it must contain two integers, (conv_padding_H, conv_padding_W). + Otherwise, the conv_padding_H = conv_padding_W = conv_padding. Default: conv_padding = 0. + conv_dilation (int|list|tuple): The dilation size of the Conv2d Layer. If dilation is + a list or tuple, it must contain two integers, (conv_dilation_H, conv_dilation_W). + Otherwise, the conv_dilation_H = conv_dilation_W = conv_dilation. Default: conv_dilation = 1. + conv_groups (int): The groups number of the Conv2d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: when group=2, + the first half of the filters is only connected to the first half + of the input channels, while the second half of the filters is only + connected to the second half of the input channels. Default: groups=1 + param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None + bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None + act (str): Activation type for Conv2d. Default: None + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled + with mkldnn library. Default: False + + Return: + Variable: The result of input after Convolution2d and Pool2d. + + Examples: + .. code-block:: python + + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + conv_pool = fluid.nets.simple_img_conv_pool(input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + """ conv_out = layers.conv2d( input=input, num_filters=num_filters, filter_size=filter_size, + stride=conv_stride, + padding=conv_padding, + dilation=conv_dilation, + groups=conv_groups, param_attr=param_attr, + bias_attr=bias_attr, act=act, use_cudnn=use_cudnn, use_mkldnn=use_mkldnn) @@ -45,6 +116,8 @@ def simple_img_conv_pool(input, pool_size=pool_size, pool_type=pool_type, pool_stride=pool_stride, + pool_padding=pool_padding, + global_pooling=global_pooling, use_cudnn=use_cudnn, use_mkldnn=use_mkldnn) return pool_out @@ -60,11 +133,65 @@ def img_conv_group(input, conv_with_batchnorm=False, conv_batchnorm_drop_rate=0.0, pool_stride=1, - pool_type=None, + pool_type="max", use_cudnn=True, use_mkldnn=False): """ - Image Convolution Group, Used for vgg net. + The Image Convolution Group is composed of Convolution2d, BatchNorm, DropOut, + and Pool2d. According to the input arguments, img_conv_group will do serials of + computation for Input using Convolution2d, BatchNorm, DropOut, and pass the last + result to Pool2d. + + Args: + input (Variable): The input image with [N, C, H, W] format. + conv_num_filter(list|tuple): Indicates the numbers of filter of this group. + pool_size (int|list|tuple): The pooling size of Pool2d Layer. If pool_size + is a list or tuple, it must contain two integers, (pool_size_H, pool_size_W). + Otherwise, the pool_size_H = pool_size_W = pool_size. + conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is + a list or tuple, its length must be equal to the length of conv_num_filter. + Otherwise the conv_padding of all Conv2d Layers are the same. Default 1. + conv_filter_size (int|list|tuple): The filter size. If filter_size is a list or + tuple, its length must be equal to the length of conv_num_filter. + Otherwise the conv_filter_size of all Conv2d Layers are the same. Default 3. + conv_act (str): Activation type for Conv2d Layer that is not followed by BatchNorm. + Default: None. + param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None + conv_with_batchnorm (bool|list): Indicates whether to use BatchNorm after Conv2d Layer. + If conv_with_batchnorm is a list, its length must be equal to the length of + conv_num_filter. Otherwise, conv_with_batchnorm indicates whether all the + Conv2d Layer follows a BatchNorm. Default False. + conv_batchnorm_drop_rate (float|list): Indicates the drop_rate of Dropout Layer + after BatchNorm. If conv_batchnorm_drop_rate is a list, its length must be + equal to the length of conv_num_filter. Otherwise, drop_rate of all Dropout + Layers is conv_batchnorm_drop_rate. Default 0.0. + pool_stride (int|list|tuple): The pooling stride of Pool2d layer. If pool_stride + is a list or tuple, it must contain two integers, (pooling_stride_H, + pooling_stride_W). Otherwise, the pooling_stride_H = pooling_stride_W = pool_stride. + Default 1. + pool_type (str): Pooling type can be :math:`max` for max-pooling and :math:`avg` for + average-pooling. Default :math:`max`. + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled + with mkldnn library. Default: False + + Return: + Variable: The final result after serial computation using Convolution2d, + BatchNorm, DropOut, and Pool2d. + + Examples: + .. code-block:: python + + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + conv_pool = fluid.nets.img_conv_group(input=img, + num_channels=3, + conv_padding=1, + conv_num_filter=[3, 3], + conv_filter_size=3, + conv_act="relu", + pool_size=2, + pool_stride=2) """ tmp = input assert isinstance(conv_num_filter, list) or \ @@ -74,6 +201,7 @@ def img_conv_group(input, if not hasattr(obj, '__len__'): return [obj] * len(conv_num_filter) else: + assert len(obj) == len(conv_num_filter) return obj conv_padding = __extend_list__(conv_padding) @@ -82,7 +210,7 @@ def img_conv_group(input, conv_with_batchnorm = __extend_list__(conv_with_batchnorm) conv_batchnorm_drop_rate = __extend_list__(conv_batchnorm_drop_rate) - for i in xrange(len(conv_num_filter)): + for i in range(len(conv_num_filter)): local_conv_act = conv_act if conv_with_batchnorm[i]: local_conv_act = None @@ -119,6 +247,39 @@ def sequence_conv_pool(input, param_attr=None, act="sigmoid", pool_type="max"): + """ + The sequence_conv_pool is composed with Sequence Convolution and Pooling. + + Args: + input (Variable): The input of sequence_conv, which supports variable-time + length input sequence. The underlying of input is a matrix with shape + (T, N), where T is the total time steps in this mini-batch and N is + the input_hidden_size + num_filters(int): The number of filter. + filter_size (int): The filter size. + param_attr (ParamAttr): The parameters to the Sequence_conv Layer. Default: None. + act (str): Activation type for Sequence_conv Layer. Default: "sigmoid". + pool_type (str): Pooling type can be :math:`max` for max-pooling, :math:`average` for + average-pooling, :math:`sum` for sum-pooling, :math:`sqrt` for sqrt-pooling. + Default :math:`max`. + + Return: + Variable: The final result after Sequence Convolution and Pooling. + + Examples: + .. code-block:: python + + input_dim = len(word_dict) + emb_dim = 128 + hid_dim = 512 + data = fluid.layers.data( ame="words", shape=[1], dtype="int64", lod_level=1) + emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim], is_sparse=True) + seq_conv = fluid.nets.sequence_conv_pool(input=emb, + num_filters=hid_dim, + filter_size=3, + act="tanh", + pool_type="sqrt") + """ conv_out = layers.sequence_conv( input=input, num_filters=num_filters, @@ -132,9 +293,9 @@ def sequence_conv_pool(input, def glu(input, dim=-1): """ - The gated linear unit composed by split, sigmoid activation and elementwise - multiplication. Specifically, Split the input into two equal sized parts - :math:`a` and :math:`b` along the given dimension and then compute as + The Gated Linear Units(GLU) composed by split, sigmoid activation and element-wise + multiplication. Specifically, Split the input into two equal sized parts, + :math:`a` and :math:`b`, along the given dimension and then compute as following: .. math:: @@ -147,16 +308,16 @@ def glu(input, dim=-1): Args: input (Variable): The input variable which is a Tensor or LoDTensor. dim (int): The dimension along which to split. If :math:`dim < 0`, the - dimension to split along is :math:`rank(input) + dim`. + dimension to split along is :math:`rank(input) + dim`. Default -1. Returns: - Variable: The Tensor variable with half the size of input. + Variable: Variable with half the size of input. Examples: .. code-block:: python - # x is a Tensor variable with shape [3, 6, 9] - fluid.nets.glu(input=x, dim=1) # shape of output: [3, 3, 9] + data = fluid.layers.data(name="words", shape=[3, 6, 9], dtype="float32") + output = fluid.nets.glu(input=data, dim=1) # shape of output: [3, 3, 9] """ a, b = layers.split(input, num_or_sections=2, dim=dim) @@ -189,40 +350,48 @@ def scaled_dot_product_attention(queries, `_. Args: - queries (Variable): The input variable which should be a 3-D Tensor. keys (Variable): The input variable which should be a 3-D Tensor. values (Variable): The input variable which should be a 3-D Tensor. num_heads (int): Head number to compute the scaled dot product - attention. Default value is 1. + attention. Default: 1. dropout_rate (float): The dropout rate to drop the attention weight. - Default value is 0. + Default: 0.0. Returns: - - Variable: A 3-D Tensor computed by multi-head scaled dot product \ - attention. + Variable: A 3-D Tensor computed by multi-head scaled dot product\ + attention. Raises: - ValueError: If input queries, keys, values are not 3-D Tensors. - NOTE: + NOTES: 1. When num_heads > 1, three linear projections are learned respectively - to map input queries, keys and values into queries', keys' and values'. - queries', keys' and values' have the same shapes with queries, keys - and values. - - 1. When num_heads == 1, scaled_dot_product_attention has no learnable - parameters. + to map input queries, keys and values into queries', keys' and values'. + queries', keys' and values' have the same shapes with queries, keys + and values. + 2. When num_heads == 1, scaled_dot_product_attention has no learnable + parameters. Examples: .. code-block:: python - # Suppose q, k, v are Tensors with the following shape: - # q: [3, 5, 9], k: [3, 6, 9], v: [3, 6, 10] - - contexts = fluid.nets.scaled_dot_product_attention(q, k, v) + queries = fluid.layers.data(name="queries", + shape=[3, 5, 9], + dtype="float32", + append_batch_size=False) + queries.stop_gradient = False + keys = fluid.layers.data(name="keys", + shape=[3, 6, 9], + dtype="float32", + append_batch_size=False) + keys.stop_gradient = False + values = fluid.layers.data(name="values", + shape=[3, 6, 10], + dtype="float32", + append_batch_size=False) + values.stop_gradient = False + contexts = fluid.nets.scaled_dot_product_attention(queries, keys, values) contexts.shape # [3, 5, 10] """ if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3): @@ -319,10 +488,11 @@ def scaled_dot_product_attention(queries, trans_x = layers.transpose(x, perm=[0, 2, 1, 3]) return layers.reshape( x=trans_x, - shape=map(int, [ - trans_x.shape[0], trans_x.shape[1], - trans_x.shape[2] * trans_x.shape[3] - ])) + shape=list( + map(int, [ + trans_x.shape[0], trans_x.shape[1], trans_x.shape[2] * + trans_x.shape[3] + ]))) q, k, v = __compute_qkv(queries, keys, values, num_heads) diff --git a/python/paddle/fluid/op.py b/python/paddle/fluid/op.py index 0b76e94157e378b40baff641c466968e239d8a83..93f021a360ac61f64e769d057df188d79f6f2bb6 100644 --- a/python/paddle/fluid/op.py +++ b/python/paddle/fluid/op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +import six + import paddle.fluid.core as core import paddle.fluid.proto.framework_pb2 as framework_pb2 @@ -24,13 +26,13 @@ def get_all_op_protos(): protostrs = core.get_all_op_protos() ret_values = [] for pbstr in protostrs: - op_proto = framework_pb2.OpProto.FromString(str(pbstr)) + op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr)) ret_values.append(op_proto) return ret_values def is_str(s): - return isinstance(s, str) or isinstance(s, unicode) + return isinstance(s, six.string_types) class OpDescCreationMethod(object): @@ -189,7 +191,7 @@ class OperatorFactory(object): return self.get_op_info(t).method(**kwargs) def types(self): - return self.op_methods.keys() + return list(self.op_methods.keys()) def get_op_info(self, t): if t not in self.op_methods: @@ -197,13 +199,13 @@ class OperatorFactory(object): return self.op_methods.get(t) def get_op_input_names(self, type): - return map(lambda x: x[0], self.get_op_info(type).inputs) + return [x[0] for x in self.get_op_info(type).inputs] def get_op_inputs(self, type): return self.get_op_info(type).inputs def get_op_output_names(self, type): - return map(lambda x: x[0], self.get_op_info(type).outputs) + return [x[0] for x in self.get_op_info(type).outputs] def get_op_outputs(self, type): return self.get_op_info(type).outputs diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 115362c6bf33018342699a442c688e7356f3c206..a07325f46a2892222c2d1dcd74aa7cb01f6760a1 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -13,23 +13,23 @@ # limitations under the License. import re from collections import defaultdict -from paddle.fluid.framework import Program -import framework -import layers -from backward import append_backward -from framework import program_guard -import unique_name -from initializer import Constant -from layer_helper import LayerHelper -from regularizer import append_regularization_ops -from clip import append_gradient_clip_ops, error_clip_callback +from paddle.fluid.framework import Program, Variable +from . import framework +from . import layers +from .backward import append_backward +from .framework import program_guard +from . import unique_name +from .initializer import Constant +from .layer_helper import LayerHelper +from .regularizer import append_regularization_ops +from .clip import append_gradient_clip_ops, error_clip_callback from contextlib import contextmanager __all__ = [ - 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', + 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer', 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer', - 'Adadelta', 'ModelAverage', 'Optimizer' + 'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'RMSPropOptimizer' ] @@ -41,7 +41,10 @@ class Optimizer(object): but need to use one of it's implementation. """ - def __init__(self, learning_rate, regularization=None): + def __init__(self, + learning_rate, + regularization=None, + LARS_weight_decay=0.0): if not isinstance(learning_rate, float) and \ not isinstance(learning_rate, framework.Variable): raise TypeError("learning rate should be float or Variable") @@ -61,9 +64,10 @@ class Optimizer(object): # {accum_name : { paramter_name : accumulator_for_parameter, ...}, ...} self._accumulators = defaultdict(lambda: dict()) self.helper = None + self._LARS_weight_decay = LARS_weight_decay def _create_global_learning_rate(self): - lr = self.global_learning_rate() + lr = self._global_learning_rate() if isinstance(lr, framework.Variable): return @@ -82,7 +86,7 @@ class Optimizer(object): dtype='float32' if self._dtype == None else self._dtype, persistable=True) - def global_learning_rate(self, program=None): + def _global_learning_rate(self, program=None): """ get global decayed learning rate :return: @@ -100,10 +104,15 @@ class Optimizer(object): # create learning rate variable for every parameter param = param_and_grad[0] param_lr = param.optimize_attr['learning_rate'] - if param_lr == 1.0: - return self.global_learning_rate() + if type(param_lr) == Variable: + # param learning rate has been updated (LARS) + print("returns updated param lr ", param_lr) + return param_lr else: - return self.global_learning_rate() * param_lr + if param_lr == 1.0: + return self._global_learning_rate() + else: + return self._global_learning_rate() * param_lr def _create_accumulators(self, block, parameters): """Create all accumulators needed by the parameters @@ -114,7 +123,7 @@ class Optimizer(object): """ pass - def _finish_update(self, block): + def _finish_update(self, block, parameters_and_grads): """Finish any custom updates needed before completing an optimization step @@ -123,7 +132,7 @@ class Optimizer(object): parameters: list of parameter variables for the optimizer Returns: - list of finish ops or None + None """ pass @@ -176,22 +185,22 @@ class Optimizer(object): format(name, param.name)) return self._accumulators[name][param.name] - def create_optimization_pass(self, - parameters_and_grads, - loss, - startup_program=None): + def _create_optimization_pass(self, + parameters_and_grads, + loss, + startup_program=None): """Add optimization operators to update gradients to variables. Args: - loss: the target that this optimization is for. - parameters_and_grads: a list of (variable, gradient) pair to update. + loss(Variable): the target that this optimization is for. + parameters_and_grads(list(tuple(Variable, Variable))): + a list of (variable, gradient) pair to update. Returns: return_op_list: a list of operators that will complete one step of optimization. This will include parameter update ops, global step update ops and any other custom ops required by subclasses to manage their internal state. - :param startup_program: """ # This is a default implementation of create_optimization_pass that # can be shared by most optimizers. This implementation assumes that @@ -210,23 +219,28 @@ class Optimizer(object): self._create_accumulators(loss.block, [p[0] for p in parameters_and_grads]) self._create_global_learning_rate() + if self._LARS_weight_decay > 0.0: + layers.append_LARS(parameters_and_grads, + self._global_learning_rate(), + self._LARS_weight_decay) optimize_ops = [] for param_and_grad in parameters_and_grads: + if param_and_grad[1] is None: + continue with param_and_grad[0].block.program.optimized_guard( - param_and_grad[0]): - if param_and_grad[0].trainable is True and param_and_grad[ - 1] is not None: + param_and_grad): + if param_and_grad[0].trainable is True: optimize_op = self._append_optimize_op(loss.block, param_and_grad) optimize_ops.append(optimize_op) # Get custom finish ops for subclasses # FIXME: Need to fix this once we figure out how to handle dependencies - self._finish_update(loss.block) + self._finish_update(loss.block, parameters_and_grads) end = len(global_block.ops) - return global_block.slice_ops(start, end) + return global_block._slice_ops(start, end) def minimize(self, loss, @@ -249,13 +263,28 @@ class Optimizer(object): params_grads = append_regularization_ops(params_grads, self.regularization) - optimize_ops = self.create_optimization_pass(params_grads, loss, - startup_program) + optimize_ops = self._create_optimization_pass(params_grads, loss, + startup_program) return optimize_ops, params_grads class SGDOptimizer(Optimizer): - """ Simple SGD optimizer without any state. + """ + Optimizer of the stochastic gradient descent algorithm. + + .. math:: + + param\_out = param - learning\_rate * grad + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + + Examples: + .. code-block:: python + + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.2) + sgd_optimizer.minimize(cost) """ def __init__(self, learning_rate, **kwargs): @@ -281,7 +310,37 @@ class SGDOptimizer(Optimizer): class MomentumOptimizer(Optimizer): - """Simple Momentum optimizer with velocity state + """ + + Simple Momentum optimizer with velocity state + + This optimizer has a flag for Nestrov Momentum. + + The update equations are as follows: + + .. math:: + + & velocity = mu * velocity + gradient + + & if (use\_nesterov): + + &\quad param = param - (gradient + mu * velocity) * learning\_rate + + & else: + + &\quad param = param - learning\_rate * velocity + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + momentum (float): momentum factor + use_nesterov (bool): enables Nesterov momentum + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1) + optimizer.minimize(cost) """ _velocity_acc_str = "velocity" @@ -325,7 +384,32 @@ class MomentumOptimizer(Optimizer): class AdagradOptimizer(Optimizer): - """Simple Adagrad optimizer with moment state + """ + **Adaptive Gradient Algorithm (Adagrad)** + + The update is done as follows: + + .. math:: + + moment\_out &= moment + grad * grad + + param\_out &= param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon} + + The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + does not have the epsilon attribute. It is added here in our implementation + as also proposed here: http://cs231n.github.io/neural-networks-3/#ada + for numerical stability to avoid the division by zero error. + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adagrad(learning_rate=0.2) + optimizer.minimize(cost) """ _moment_acc_str = "moment" @@ -366,10 +450,45 @@ class AdagradOptimizer(Optimizer): class AdamOptimizer(Optimizer): - """Implements the Adam Optimizer + """ + This implements the Adam optimizer from Section 2 of the Adam + paper : https://arxiv.org/abs/1412.6980. + Adam is a first-order gradient-based optimization method based on + adaptive estimates of lower-order moments. + + Adam updates: + + .. math:: + + t & = t + 1 + + moment\_1\_out & = {\\beta}_1 * moment\_1 + (1 - {\\beta}_1) * grad + + moment\_2\_out & = {\\beta}_2 * moment\_2 + (1 - {\\beta}_2) * grad * grad + + learning\_rate & = learning\_rate * \\ + \\frac{\sqrt{1 - {\\beta}_2^t}}{1 - {\\beta}_1^t} + + param\_out & = param - learning\_rate * \\frac{moment\_1}{\sqrt{moment\_2} + \epsilon} + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + beta1 (float): The exponential decay rate for the 1st moment estimates. + beta2 (float): The exponential decay rate for the 2nd moment estimates. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adam(learning_rate=0.2) + optimizer.minimize(cost) + """ _moment1_acc_str = "moment1" _moment2_acc_str = "moment2" + _beta1_pow_acc_str = "beta1_pow_acc" + _beta2_pow_acc_str = "beta2_pow_acc" def __init__(self, learning_rate=0.001, @@ -391,32 +510,22 @@ class AdamOptimizer(Optimizer): def _create_accumulators(self, block, parameters): assert isinstance(block, framework.Block) - main_block = block.program.global_block() - # Create beta1 and beta2 power tensors - beta_shape = [1] - self._beta1_pow_acc = self.helper.create_global_variable( - name=unique_name.generate('beta1_pow_acc'), - dtype='float32' if self._dtype == None else self._dtype, - shape=beta_shape, - lod_level=0, - persistable=True) - self.helper.set_variable_initializer( - self._beta1_pow_acc, initializer=Constant(self._beta1)) - - self._beta2_pow_acc = self.helper.create_global_variable( - name=unique_name.generate('beta2_pow_acc'), - dtype='float32' if self._dtype == None else self._dtype, - shape=beta_shape, - lod_level=0, - persistable=True) - - self.helper.set_variable_initializer( - self._beta2_pow_acc, initializer=Constant(self._beta2)) - # Create accumulator tensors for first and second moments for p in parameters: self._add_accumulator(self._moment1_acc_str, p) self._add_accumulator(self._moment2_acc_str, p) + self._add_accumulator( + name=self._beta1_pow_acc_str, + param=p, + dtype='float32', + fill_value=self._beta1, + shape=[1]) + self._add_accumulator( + name=self._beta2_pow_acc_str, + param=p, + dtype='float32', + fill_value=self._beta2, + shape=[1]) def _append_optimize_op(self, block, param_and_grad): assert isinstance(block, framework.Block) @@ -425,6 +534,11 @@ class AdamOptimizer(Optimizer): param_and_grad[0]) moment2 = self._get_accumulator(self._moment2_acc_str, param_and_grad[0]) + beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str, + param_and_grad[0]) + beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str, + param_and_grad[0]) + # create the adam optimize op adam_op = block.append_op( type=self.type, @@ -434,8 +548,8 @@ class AdamOptimizer(Optimizer): "LearningRate": self._create_param_lr(param_and_grad), "Moment1": moment1, "Moment2": moment2, - "Beta1Pow": self._beta1_pow_acc, - "Beta2Pow": self._beta2_pow_acc + "Beta1Pow": beta1_pow_acc, + "Beta2Pow": beta2_pow_acc }, outputs={ "ParamOut": param_and_grad[0], @@ -450,31 +564,73 @@ class AdamOptimizer(Optimizer): return adam_op - def _finish_update(self, block): + def _finish_update(self, block, param_and_grads): """Update Beta1 and Beta2 Power accumulators """ assert isinstance(block, framework.Block) main_block = block.program.global_block() - scale_beta1 = main_block.append_op( - type="scale", - inputs={"X": self._beta1_pow_acc}, - outputs={"Out": self._beta1_pow_acc}, - attrs={"scale": self._beta1}) + for param, grad in param_and_grads: + if grad is None: + continue + with param.block.program.optimized_guard([param, grad]): + beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str, + param) + beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str, + param) + main_block.append_op( + type="scale", + inputs={"X": beta1_pow_acc}, + outputs={"Out": beta1_pow_acc}, + attrs={"scale": self._beta1}) + + main_block.append_op( + type="scale", + inputs={"X": beta2_pow_acc}, + outputs={"Out": beta2_pow_acc}, + attrs={"scale": self._beta2}) - scale_beta2 = main_block.append_op( - type="scale", - inputs={"X": self._beta2_pow_acc}, - outputs={"Out": self._beta2_pow_acc}, - attrs={"scale": self._beta2}) - return [scale_beta1, scale_beta2] +class AdamaxOptimizer(Optimizer): + """ + We implement the Adamax optimizer from Section 7 of the Adam + paper: https://arxiv.org/abs/1412.6980. Adamax is a variant of the + Adam algorithm based on the infinity norm. + Adamax updates: -class AdamaxOptimizer(Optimizer): - """Implements the Adamax Optimizer + .. math:: + + t & = t + 1 + + moment\_out & = {\\beta}_1 * moment + (1 - {\\beta}_1) * grad + + inf\_norm\_out & = max({\\beta}_2 * inf\_norm + \epsilon, |grad|) + + learning\_rate & = \\frac{learning\_rate}{1 - {\\beta}_1^t} + + param\_out & = param - learning\_rate * \\frac{moment\_out}{inf\_norm\_out} + + + The original paper does not have an epsilon attribute. + However, it is added here for numerical stability to prevent the + division by 0 error. + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + beta1 (float): The exponential decay rate for the 1st moment estimates. + beta2 (float): The exponential decay rate for the 2nd moment estimates. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adamax(learning_rate=0.2) + optimizer.minimize(cost) """ _moment_acc_str = "moment" _inf_norm_acc_str = "inf_norm" + _beta1_pow_acc_str = "beta1_pow_acc" def __init__(self, learning_rate=0.001, @@ -494,21 +650,16 @@ class AdamaxOptimizer(Optimizer): self._epsilon = epsilon def _create_accumulators(self, block, parameters): - # Create beta1 power accumulator tensor - beta_shape = [1] - self._beta1_pow_acc = self.helper.create_global_variable( - name=unique_name.generate('beta1_pow_acc'), - dtype='float32' if self._dtype == None else self._dtype, - shape=beta_shape, - lod_level=0, - persistable=True) - self.helper.set_variable_initializer( - self._beta1_pow_acc, initializer=Constant(self._beta1)) - # Create accumulator tensors for first moment and infinity norm for p in parameters: self._add_accumulator(self._moment_acc_str, p) self._add_accumulator(self._inf_norm_acc_str, p) + self._add_accumulator( + name=self._beta1_pow_acc_str, + param=p, + dtype='float32', + fill_value=self._beta1, + shape=[1]) def _append_optimize_op(self, block, param_and_grad): assert isinstance(block, framework.Block) @@ -516,6 +667,8 @@ class AdamaxOptimizer(Optimizer): moment = self._get_accumulator(self._moment_acc_str, param_and_grad[0]) inf_norm = self._get_accumulator(self._inf_norm_acc_str, param_and_grad[0]) + beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str, + param_and_grad[0]) # create the adamax optimize op adamax_op = block.append_op( type=self.type, @@ -525,7 +678,7 @@ class AdamaxOptimizer(Optimizer): "LearningRate": self._create_param_lr(param_and_grad), "Moment": moment, "InfNorm": inf_norm, - "Beta1Pow": self._beta1_pow_acc + "Beta1Pow": beta1_pow_acc }, outputs={ "ParamOut": param_and_grad[0], @@ -540,22 +693,53 @@ class AdamaxOptimizer(Optimizer): return adamax_op - def _finish_update(self, block): + def _finish_update(self, block, parameters_and_grads): """Update Beta1 Power accumulator """ assert isinstance(block, framework.Block) main_block = block.program.global_block() - scale_beta1 = main_block.append_op( - type="scale", - inputs={"X": self._beta1_pow_acc}, - outputs={"Out": self._beta1_pow_acc}, - attrs={"scale": self._beta1}) - - return [scale_beta1] + for param, grad in parameters_and_grads: + if grad is None: + continue + with param.block.program.optimized_guard([param, grad]): + beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str, + param) + main_block.append_op( + type="scale", + inputs={"X": beta1_pow_acc}, + outputs={"Out": beta1_pow_acc}, + attrs={"scale": self._beta1}) class DecayedAdagradOptimizer(Optimizer): - """Simple Decayed Adagrad optimizer with moment state + """ + **Decayed Adagrad Optimizer** + + The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + + The update is done as follows: + + .. math:: + + moment\_out & = decay * moment + (1 - decay) * grad * grad + + param\_out & = param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon} + + The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + does not have an epsilon attribute. It is added here for numerical + stability to avoid the division by zero error. + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + decay (float): decay rate. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2) + optimizer.minimize(cost) """ _moment_acc_str = "moment" @@ -601,6 +785,7 @@ class DecayedAdagradOptimizer(Optimizer): class AdadeltaOptimizer(Optimizer): """ **Adadelta Optimizer** + Simple Adadelta optimizer with average squared grad state and average squared update state. The details of adadelta please refer to this @@ -615,7 +800,7 @@ class AdadeltaOptimizer(Optimizer): E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2 Args: - learning_rate(float): global leraning rate + learning_rate(float): global learning rate rho(float): rho in equation epsilon(float): epsilon in equation @@ -690,37 +875,37 @@ class RMSPropOptimizer(Optimizer): .. math:: - r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\ + r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 w & = w - \\frac{\\eta} {\\sqrt{r(w,t) + \\epsilon}} \\nabla Q_{i}(w) The first equation calculates moving average of the squared gradient for - each weight. Then dividing the gradient by :math: `sqrt{v(w,t)}`. + each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`. In some cases, adding a momentum term :math: `\\beta` is beneficial. In our implementation, Nesterov momentum is used: .. math:: - r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\ + r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{v(w,t) + \\epsilon}} \\nabla Q_{i}(w) w & = w - v(w, t) - where, :math: `\\rho` is a hyperparameter and typical values are 0.9, 0.95 + where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95 and so on. :math: `beta` is the momentum term. :math: `\\epsilon` is a smoothing term to avoid division by zero, usually set somewhere in range from 1e-4 to 1e-8. Args: - learning_rate(float): global leraning rate. + learning_rate(float): global learning rate. rho(float): rho is :math: `\\rho` in equation, set 0.95 by default. epsilon(float): :math: `\\epsilon` in equation is smoothing term to avoid division by zero, set 1e-6 by default. - momentum(float): :math: `\\beta` in equation is the momentum term, + momentum(float): :math:`\\beta` in equation is the momentum term, set 0.0 by default. Raises: @@ -797,6 +982,113 @@ class RMSPropOptimizer(Optimizer): return rmsprop_op +class FtrlOptimizer(Optimizer): + """ + FTRL (Follow The Regularized Leader) Optimizer. + + The paper that proposed Follow The Regularized Leader (FTRL): + (https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf) + + .. math:: + + &new\_accum = squared\_accum + grad^2 + + &if (lr\_power == -0.5): + + &\quad linear\_accum += grad - \\frac{\\sqrt{new\_accum} - \\sqrt{squared\_accum}}{learning\_rate * param} + + &else: + + &\quad linear\_accum += grad - \\frac{new\_accum^{-lr\_power} - accum^{-lr\_power}}{learning\_rate * param} + + + &x = l1 * sign(linear\_accum) - linear\_accum + + &if (lr\_power == -0.5): + + &\quad y = \\frac{\\sqrt{new\_accum}}{learning\_rate} + (2 * l2) + + &\quad pre\_shrink = \\frac{x}{y} + + &\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0) + + &else: + + &\quad y = \\frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2) + + &\quad pre\_shrink = \\frac{x}{y} + + &\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0) + + &squared\_accum += grad^2 + + Args: + learning_rate (float|Variable): global learning rate. + l1 (float): + l2 (float): + lr_power (float): + + Raises: + ValueError: If learning_rate, rho, epsilon, momentum are None. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Ftrl(0.0001) + _, params_grads = optimizer.minimize(cost) + """ + + _squared_acc_str = "squared" + _linear_acc_str = "linear" + + def __init__(self, learning_rate, l1=0.0, l2=0.0, lr_power=-0.5, **kwargs): + super(FtrlOptimizer, self).__init__( + learning_rate=learning_rate, **kwargs) + if learning_rate is None: + raise ValueError("learning_rate is not set.") + + self.type = "ftrl" + self._l1 = l1 + self._l2 = l2 + self._lr_power = lr_power + + def _create_accumulators(self, block, parameters): + if not isinstance(block, framework.Block): + raise TypeError("block is not instance of framework.Block.") + + for p in parameters: + self._add_accumulator(self._squared_acc_str, p) + self._add_accumulator(self._linear_acc_str, p) + + def _append_optimize_op(self, block, param_and_grad): + if not isinstance(block, framework.Block): + raise TypeError("block is not instance of framework.Block.") + + squared_acc = self._get_accumulator(self._squared_acc_str, + param_and_grad[0]) + linear_acc = self._get_accumulator(self._linear_acc_str, + param_and_grad[0]) + ftrl_op = block.append_op( + type=self.type, + inputs={ + "Param": param_and_grad[0], + "Grad": param_and_grad[1], + "SquaredAccumulator": squared_acc, + "LinearAccumulator": linear_acc, + "LearningRate": self._create_param_lr(param_and_grad), + }, + outputs={ + "ParamOut": param_and_grad[0], + "SquaredAccumOut": squared_acc, + "LinearAccumOut": linear_acc + }, + attrs={"l1": self._l1, + "l2": self._l1, + "lr_power": self._lr_power}) + + return ftrl_op + + # We short the class name, since users will use the optimizer with the package # name. The sample code: # @@ -813,6 +1105,7 @@ Adamax = AdamaxOptimizer DecayedAdagrad = DecayedAdagradOptimizer Adadelta = AdadeltaOptimizer RMSProp = RMSPropOptimizer +Ftrl = FtrlOptimizer class ModelAverage(Optimizer): @@ -826,15 +1119,16 @@ class ModelAverage(Optimizer): Args: average_window_rate: The rate of average window. - params_grads: A list of parameter-grad variable pairs. min_average_window: The minimum size of average window. max_average_window: The maximum size of average window. Examples: - ... + + .. code-block:: python + optimizer = fluid.optimizer.Momentum() - _, params_grads = optimizer.minimize(cost) - model_average = fluid.optimizer.ModelAverage(params_grads, 0.15, + optimizer.minimize(cost) + model_average = fluid.optimizer.ModelAverage(0.15, min_average_window=10000, max_average_window=20000) for pass_id in range(args.pass_num): @@ -848,7 +1142,6 @@ class ModelAverage(Optimizer): def __init__(self, average_window_rate, - params_grads=None, min_average_window=10000, max_average_window=10000, **kwargs): @@ -857,24 +1150,22 @@ class ModelAverage(Optimizer): self.min_average_window = min_average_window self.max_average_window = max_average_window - self.params_grads = [] if params_grads is None else params_grads - params = {} - for param, grad in self.params_grads: - if param.do_model_average != False: - params[param.name] = (param, grad) + self.params_grads = [] for param in framework.default_main_program().global_block( ).all_parameters(): - if param.name not in params and param.do_model_average != False: + if param.do_model_average != False: grad = param.block.create_var( name=unique_name.generate(".".join([param.name, 'tmp'])), dtype=param.dtype, persistable=False, stop_gradient=True) - params[param.name] = (param, grad) - self.params_grads = params.values() + self.params_grads.append((param, grad)) for param, grad in self.params_grads: - self._append_average_accumulate_op(param) + if grad is None: + continue + with param.block.program.optimized_guard([param, grad]): + self._append_average_accumulate_op(param) self.apply_program = Program() block = self.apply_program.global_block() @@ -889,16 +1180,16 @@ class ModelAverage(Optimizer): self._add_average_restore_op(block, param_grad) def _add_average_apply_op(self, block, param_grad): - param = block.clone_variable(param_grad[0]) - grad = block.clone_variable(param_grad[1]) - sum_1 = block.clone_variable(self._get_accumulator('sum_1', param)) - sum_2 = block.clone_variable(self._get_accumulator('sum_2', param)) - sum_3 = block.clone_variable(self._get_accumulator('sum_3', param)) - num_accumulates = block.clone_variable( + param = block._clone_variable(param_grad[0]) + grad = block._clone_variable(param_grad[1]) + sum_1 = block._clone_variable(self._get_accumulator('sum_1', param)) + sum_2 = block._clone_variable(self._get_accumulator('sum_2', param)) + sum_3 = block._clone_variable(self._get_accumulator('sum_3', param)) + num_accumulates = block._clone_variable( self._get_accumulator('num_accumulates', param)) - old_num_accumulates = block.clone_variable( + old_num_accumulates = block._clone_variable( self._get_accumulator('old_num_accumulates', param)) - num_updates = block.clone_variable( + num_updates = block._clone_variable( self._get_accumulator('num_updates', param)) # backup param value to grad layers.assign(input=param, output=grad) @@ -912,8 +1203,8 @@ class ModelAverage(Optimizer): layers.elementwise_div(x=sum, y=tmp, out=param) def _add_average_restore_op(self, block, param_grad): - param = block.clone_variable(param_grad[0]) - grad = block.clone_variable(param_grad[1]) + param = block._clone_variable(param_grad[0]) + grad = block._clone_variable(param_grad[1]) layers.assign(input=grad, output=param) def _append_average_accumulate_op(self, param): diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 3117dfe00c7a3df1035c439dc31b81e67781d0cc..2a3555ebdde4d54f63bb420218896560c1b40ffd 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -12,12 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. -import core +from __future__ import print_function import multiprocessing -import framework -import executor +from . import core +from . import framework +from . import executor import warnings import sys +import os __all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy'] @@ -26,6 +28,40 @@ BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): + """ + ParallelExecutor can run program in parallel. + + Args: + use_cuda (bool): Whether to use CUDA or not. + loss_name (str): The loss name must set in training. Default None. + main_program (Program): The program that need to run, if not provided, + then default_main_program will be used. Default None. + share_vars_from(ParallelExecutor): If provied, it will share variables + from the specified ParallelExecutor. Default None. + num_trainers(int): If greater than 1, NCCL will be initialized with + multiple rank of nodes, each node should have same number of GPUs. + Distributed training will be enabled then. Default 1. + trainer_id(int: Must use together with num_trainers. trainer_id is the + "rank" of current node starts from 0. Default 0. + + Returns: + ParallelExecutor: The initialized ParallelExecutor object. + + Raises: + TypeError: If share_vars_from is provided, but not ParallelExecutor object. + + Examples: + .. code-block:: python + + train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name) + test_exe = fluid.ParallelExecutor(use_cuda=True, + main_program=test_program, + share_vars_from=train_exe) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) + test_loss, = test_exe.run([loss.name], feed=feed_dict) + """ + def __init__(self, use_cuda, loss_name=None, @@ -36,42 +72,6 @@ class ParallelExecutor(object): num_trainers=1, trainer_id=0, **kwargs): - """ - ParallelExecutor can run program in parallel. - - Args: - use_cuda(bool): Whether to use CUDA or not. - loss_name(str, default None): The loss name must set in training. - main_program(Program, default None): The program that need to run, - if not provided, then default_main_program will be used. - share_vars_from(ParallelExecutor, default None): If provied, - it will share variables from the specified ParallelExecutor. - num_trainers(int, default 1): If greater than 1, NCCL will be - initialized with multpile rank of nodes, each node should have - same number of GPUs. Distributed training will be enabled then. - trainer_id(int, default 0): Must use together with num_trainers. - trainer_id is the "rank" of current node starts from 0. - - Returns: - A ParallelExecutor object. - - Raises: - TypeError: If share_vars_from is provided, but not ParallelExecutor - object. - - Examples: - .. code-block:: python - - train_exe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name) - test_exe = fluid.ParallelExecutor( - use_cuda=True, - main_program=test_program, - share_vars_from=train_exe) - - train_loss, = train_exe.run([loss.name], feed=feed_dict) - test_loss, = test_exe.run([loss.name], feed=feed_dict) - """ if len(kwargs) != 0: err_msg = "" for key in kwargs: @@ -95,13 +95,15 @@ class ParallelExecutor(object): self._places = [] self._act_places = [] if use_cuda: - for i in xrange(core.get_cuda_device_count()): + for i in range(core.get_cuda_device_count()): p = core.Place() self._act_places.append(core.CUDAPlace(i)) p.set_place(self._act_places[-1]) self._places.append(p) else: - for i in xrange(multiprocessing.cpu_count()): + cpu_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + for i in range(cpu_num): p = core.Place() self._act_places.append(core.CPUPlace()) p.set_place(self._act_places[-1]) @@ -110,19 +112,17 @@ class ParallelExecutor(object): if exec_strategy is None: exec_strategy = ExecutionStrategy() - if use_cuda: - exec_strategy.use_event = True - else: - exec_strategy.use_event = False + exec_strategy.use_cuda = use_cuda if exec_strategy.num_threads == 0: if use_cuda: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. - exec_strategy.num_threads = len(self._places) * 2 + exec_strategy.num_threads = len(self._places) * 4 else: - exec_strategy.num_threads = min( - len(self._places) * 2, multiprocessing.cpu_count()) + cpu_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + exec_strategy.num_threads = cpu_num * 2 if build_strategy is None: build_strategy = BuildStrategy() @@ -130,18 +130,24 @@ class ParallelExecutor(object): main = main_program main = main if main else framework.default_main_program() scope = executor.global_scope() + # FIXME(Yancey1989): it's a temporary approach to determinate the distribute + # train program, call self.bcast_param() at the end of each mini-batch. + self.is_dist = True if "recv" in [ + op.type for op in main.global_block().ops + ] else False if share_vars_from and not isinstance(share_vars_from, ParallelExecutor): raise TypeError("share_vars_from must be ParallelExecutor.") + local_scopes = share_vars_from.executor.local_scopes( ) if share_vars_from else [] self.persistable_vars = [ - v.name - for v in filter( - lambda var: var.persistable and var.type != core.VarDesc.VarType.RAW, - main.list_vars()) + v.name for v in [ + var for var in main.list_vars() + if var.persistable and var.type != core.VarDesc.VarType.RAW + ] ] self.executor = core.ParallelExecutor( @@ -155,7 +161,7 @@ class ParallelExecutor(object): build_strategy, num_trainers, trainer_id) self.scope = scope - def run(self, fetch_list, feed=None, feed_dict=None): + def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=True): """ Run a parallel executor with fetch_list. @@ -165,12 +171,14 @@ class ParallelExecutor(object): element in the list will be copied to each device directly. For example, if the feed is a dict: + >>> exe = ParallelExecutor() >>> # the image will be splitted into devices. If there is two devices >>> # each device will process an image with shape (24, 1, 28, 28) >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))}) For example, if the feed is a list: + >>> exe = ParallelExecutor() >>> # each device will process each element in the list. >>> # the 1st device will process an image with shape (48, 1, 28, 28) @@ -181,22 +189,48 @@ class ParallelExecutor(object): >>> {"image": numpy.random.random(size=(32, 1, 28, 28))}, >>> ]) - Args: fetch_list(list): The fetched variable names feed(list|dict|None): The feed variables. If the feed is a dict, tensors in that dict will be splitted into each devices. If the feed is a list, each element of the list will be copied - to each device. + to each device. Default None. feed_dict: Alias for feed parameter, for backward compatibility. - This parameter is deprecated. + This parameter has been deprecated. Default None. + return_numpy(bool): Whether converts the fetched tensor to numpy. + Default: True. - Returns: fetched result list. + Returns: + List: The fetched result list. + + Raises: + ValueError: If the feed is a list, but its length is not equal the + length of active places, or its element's is not dict. + + NOTES: + 1. If the feed's type is dict, the number of data that feeds to + ParallelExecutor must be bigger than active places. Otherwise, + it will throw exception from C++ side. Special attention should be + paid to check whether the last batch of the dataset is bigger + than active places. + 2. If active places are more than one, the fetch results for each + variable is a list, and each element of this list is the variable of + respective active place. + Examples: + .. code-block:: python + + pe = fluid.ParallelExecutor(use_cuda=use_cuda, + loss_name=avg_cost.name, + main_program=fluid.default_main_program()) + loss = pe.run(feed=feeder.feed(cur_batch), + fetch_list=[avg_cost.name])) """ if feed is None and feed_dict is not None: feed = feed_dict - print >> sys.stderr, "`feed_dict` is deprecated. Please use `feed=`" + print( + "`feed_dict` is deprecated. Please use `feed=`", + file=sys.stderr) if isinstance(feed, dict): feed_tensor_dict = dict() @@ -237,10 +271,21 @@ class ParallelExecutor(object): fetch_var_name = '@FETCHED_VAR_NAME@' self.executor.run(fetch_list, fetch_var_name) arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array() + + if self.is_dist: + self._bcast_params() + + if return_numpy: + return executor.as_numpy(arr) + return [arr[i] for i in range(len(arr))] - def bcast_params(self): - self.executor.bcast_params(set(self.persistable_vars)) + def _bcast_params(self): + """ + Broadcast the parameters to other devices. It is used during + distributed training. + """ + self.executor._bcast_params(set(self.persistable_vars)) @property def device_count(self): diff --git a/python/paddle/fluid/param_attr.py b/python/paddle/fluid/param_attr.py index 1c6970441bccdc1c1221503256c30c83502bd123..afae577656c8970338f3b02208fcb4c738628ab6 100644 --- a/python/paddle/fluid/param_attr.py +++ b/python/paddle/fluid/param_attr.py @@ -12,8 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. -from initializer import Initializer, Xavier, Constant -from regularizer import WeightDecayRegularizer +import six + +from .initializer import Initializer, Xavier, Constant +from .regularizer import WeightDecayRegularizer __all__ = [ 'ParamAttr', @@ -22,6 +24,35 @@ __all__ = [ class ParamAttr(object): + """ + Parameter attributes object. To fine-tuning network training process, user + can set parameter's attributes to control training details. Such as learning rate, + regularization, trainable, do_model_average and the method to initialize param. + + + Args: + name(str): The parameter's name. Default None. + initializer(Initializer): The method to initial this parameter. Default None. + learning_rate(float): The parameter's learning rate. The learning rate when + optimize is :math:`global\_lr * parameter\_lr * scheduler\_factor`. + Default 1.0. + regularizer(WeightDecayRegularizer): Regularization factor. Default None. + trainable(bool): Whether this parameter is trainable. Default True. + gradient_clip(BaseGradientClipAttr): The method to clip this parameter's + gradient. Default None. + do_model_average(bool): Whether this parameter should do model average. + Default False. + + Examples: + .. code-block:: python + + w_param_attrs = fluid.ParamAttr(name="fc_weight", + learning_rate=0.5, + regularizer=fluid.L2Decay(1.0), + trainable=True) + y_predict = fluid.layers.fc(input=x, size=10, param_attr=w_param_attrs) + """ + def __init__(self, name=None, initializer=None, @@ -29,7 +60,7 @@ class ParamAttr(object): regularizer=None, trainable=True, gradient_clip=None, - do_model_average=None): + do_model_average=False): self.name = name self.initializer = initializer self.learning_rate = learning_rate @@ -38,7 +69,17 @@ class ParamAttr(object): self.gradient_clip = gradient_clip self.model_average = do_model_average - def set_default_initializer(self, initializer): + def _set_default_initializer(self, initializer): + """ + Set the default initializer, the initializer should be Constant, + Uniform, Normal, Xavier, MSRA. + + Args: + initializer(Initializer): the initializer to set. + + Returns: + None + """ if initializer is None: if self.initializer is None: raise ValueError("ParamAttr.initializer is not set") @@ -49,32 +90,73 @@ class ParamAttr(object): self.initializer = initializer - def set_default_param_initializer(self): - self.set_default_initializer(Xavier()) + def _set_default_param_initializer(self): + """ + Set the default initializer for the parameter with Xavier. + + Args: + None. + + Returns: + None. + """ + self._set_default_initializer(Xavier()) - def set_default_bias_initializer(self): - self.set_default_initializer(Constant(0.0)) + def _set_default_bias_initializer(self): + """ + Set the default initializer for the bias with Constant(0.0). + + Args: + None. + + Returns: + None. + """ + self._set_default_initializer(Constant(0.0)) @staticmethod - def to_attr(arg): + def _to_attr(arg): + """ + Create ParamAttr[s]. + + Args: + arg: Arguments to initialize ParamAttr[s]. arg's type can be + str, Initializer, float, WeightDecayRegularizer, BaseGradientClipAttr, + bool, ParamAttr, or a list of above type. + + Returns: + ParamAttr[s]: ParamAttr[s] initialized with arg. + + Raises: + arg can not initialize a ParamAttr. + """ if arg is None: return ParamAttr() elif isinstance(arg, list) or isinstance(arg, tuple): - return [ParamAttr.to_attr(a) for a in arg] + return [ParamAttr._to_attr(a) for a in arg] elif isinstance(arg, ParamAttr): return arg - elif isinstance(arg, str) or isinstance(arg, unicode): + elif isinstance(arg, six.string_types): return ParamAttr(name=arg) elif isinstance(arg, Initializer): return ParamAttr(initializer=arg) elif isinstance(arg, WeightDecayRegularizer): return ParamAttr(regularizer=arg) elif isinstance(arg, bool): - return ParamAttr.to_attr(None) if arg else False + return ParamAttr._to_attr(None) if arg else False else: raise TypeError("{0} cast to ParamAttr".format(type(arg))) - def to_kwargs(self, with_initializer=False): + def _to_kwargs(self, with_initializer=False): + """ + Returns the attributes of this parameter. + + Args: + with_initializer(bool): Whether to add initializer attr. + + Returns: + Parameter attributes(map): The attributes of this parameter. + """ kwargs = { 'name': self.name, 'optimize_attr': { @@ -92,9 +174,27 @@ class ParamAttr(object): class WeightNormParamAttr(ParamAttr): """ - Used for weight normalization. Any field in ParamAttr can also be set here. - Besides, an extra field dim can be set to indicate the dimension except - which to normalize. + Used for weight Norm. Weight Norm is a reparameterization of the weight vectors + in a neural network that decouples the length of those weight vectors from + their direction. Weight Norm has been implemented as discussed in this + paper: `Weight Normalization: A Simple Reparameterization to Accelerate + Training of Deep Neural Networks + `_. + + Args: + dim(list): The parameter's name. Default None. + kwargs: Any field in ParamAttr. Default None. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[3, 32, 32], dtype="float32") + fc = fluid.layers.fc(input=data, + size=1000, + param_attr=WeightNormParamAttr( + dim=None, + name='weight_norm_param')) + """ # List to record the parameters reparameterized by weight normalization. # If these parameters are treated as Variable rather than Parameter, diff --git a/python/paddle/fluid/profiler.py b/python/paddle/fluid/profiler.py index e2bd1d4c9a1ea5ddc0dfd19c769dcb40bfd6d04c..01983a830351b018770e6358f604781ffaae5800 100644 --- a/python/paddle/fluid/profiler.py +++ b/python/paddle/fluid/profiler.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import core +from . import core from contextlib import contextmanager import os @@ -42,6 +42,9 @@ def cuda_profiler(output_file, output_mode=None, config=None): counters/options for profiling by `config` argument. The default config is ['gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d', 'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace']. + Then users can use NVIDIA Visual Profiler + (https://developer.nvidia.com/nvidia-visual-profiler) tools to load this + this output file to visualize results. Args: output_file (string) : The output file name, the result will be @@ -50,6 +53,33 @@ def cuda_profiler(output_file, output_mode=None, config=None): Comma separated values format. It should be 'kvp' or 'csv'. config (list of string) : The profiler options and counters can refer to "Compute Command Line Profiler User Guide". + + Raises: + ValueError: If `output_mode` is not in ['kvp', 'csv']. + + Examples: + + .. code-block:: python + + import paddle.fluid as fluid + import paddle.fluid.profiler as profiler + + epoc = 8 + dshape = [4, 3, 28, 28] + data = fluid.layers.data(name='data', shape=[3, 28, 28], dtype='float32') + conv = fluid.layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1]) + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + + output_file = 'cuda_profiler.txt' + with profiler.cuda_profiler(output_file, 'csv') as nvprof: + for i in range(epoc): + input = np.random.random(dshape).astype('float32') + exe.run(fluid.default_main_program(), feed={'data': input}) + # then use NVIDIA Visual Profiler (nvvp) to load this output file + # to visualize results. """ if output_mode is None: output_mode = 'csv' @@ -69,19 +99,52 @@ def cuda_profiler(output_file, output_mode=None, config=None): def reset_profiler(): - """The profiler clear interface. - reset_profiler will clear the previous time record. + """ + Clear the previous time record. This interface does not work for + `fluid.profiler.cuda_profiler`, it only works for + `fluid.profiler.start_profiler`, `fluid.profiler.stop_profiler`, + and `fluid.profiler.profiler`. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + with profiler.profiler(state, 'total', '/tmp/profile'): + for iter in range(10): + if iter == 2: + profiler.reset_profiler() + # ... """ core.reset_profiler() def start_profiler(state): - """Enable the profiler. + """ + Enable the profiler. Uers can use `fluid.profiler.start_profiler` and + `fluid.profiler.stop_profiler` to insert the code, except the usage of + `fluid.profiler.profiler` interface. Args: state (string) : The profiling state, which should be 'CPU', 'GPU' or 'All'. 'CPU' means only profile CPU. 'GPU' means profiling GPU as well. 'All' also generates timeline. + + Raises: + ValueError: If `state` is not in ['CPU', 'GPU', 'All']. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + + profiler.start_profiler('GPU') + for iter in range(10): + if iter == 2: + profiler.reset_profiler() + # except each iteration + profiler.stop_profiler('total', '/tmp/profile') """ if core.is_profiler_enabled(): return @@ -97,7 +160,10 @@ def start_profiler(state): def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): - """Stop the profiler. + """ + Stop the profiler. Uers can use `fluid.profiler.start_profiler` and + `fluid.profiler.stop_profiler` to insert the code, except the usage of + `fluid.profiler.profiler` interface. Args: sorted_key (string) : If None, the profiling results will be printed @@ -111,6 +177,23 @@ def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): The `ave` means sorting by the average execution time. profile_path (string) : If state == 'All', it will write a profile proto output file. + + Raises: + ValueError: If `sorted_key` is not in + ['calls', 'total', 'max', 'min', 'ave']. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + + profiler.start_profiler('GPU') + for iter in range(10): + if iter == 2: + profiler.reset_profiler() + # except each iteration + profiler.stop_profiler('total', '/tmp/profile') """ if not core.is_profiler_enabled(): return @@ -135,15 +218,20 @@ def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): def profiler(state, sorted_key=None, profile_path='/tmp/profile'): """The profiler interface. Different from cuda_profiler, this profiler can be used to profile both CPU - and GPU program. By defalut, it records the CPU and GPU operator kernels, + and GPU program. By default, it records the CPU and GPU operator kernels, if you want to profile other program, you can refer the profiling tutorial - to add more records. + to add more records in C++ code. + + If the state == 'All', a profile proto file will be written to + `profile_path`. This file records timeline information during the execution. + Then users can visualize this file to see the timeline, please refer + https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/timeline.md Args: state (string) : The profiling state, which should be 'CPU' or 'GPU', telling the profiler to use CPU timer or GPU timer for profiling. Although users may have already specified the execution place - (CPUPlace/CUDAPlace) in the begining, for flexibility the profiler + (CPUPlace/CUDAPlace) in the beginning, for flexibility the profiler would not inherit this place. sorted_key (string) : If None, the profiling results will be printed in the order of first end time of events. Otherwise, the profiling @@ -156,6 +244,25 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'): The `ave` means sorting by the average execution time. profile_path (string) : If state == 'All', it will write a profile proto output file. + + Raises: + ValueError: If `state` is not in ['CPU', 'GPU', 'All']. If `sorted_key` is + not in ['calls', 'total', 'max', 'min', 'ave']. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + + with profiler.profiler('All', 'total', '/tmp/profile') as prof: + for pass_id in range(pass_num): + for batch_id, data in enumerate(train_reader()): + exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[], + use_program_cache=True) + # ... """ start_profiler(state) yield diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index 5accaacd5361165d30b92c71ae4fd62e23e44e07..93b38ad3fa37bd4bff04c529cd5518a8138e55ea 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -12,10 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. -import core +import os import contextlib - -__all__ = ['convert_reader_to_recordio_file'] +from . import core +__all__ = [ + 'convert_reader_to_recordio_file', 'convert_reader_to_recordio_files' +] @contextlib.contextmanager @@ -34,6 +36,45 @@ def convert_reader_to_recordio_file( compressor=core.RecordIOWriter.Compressor.Snappy, max_num_records=1000, feed_order=None): + """ + Convert a Python Reader to a recordio file. + + Please see :ref:`api_guide_python_reader` and :ref:`api_guide_reader_op` for + details. + + Examples: + + >>> import paddle.fluid as fluid + >>> import paddle.dataset.mnist as mnist + >>> import paddle + >>> + >>> tmp_program = fluid.Program() + >>> with fluid.program_guard(tmp_program): + >>> img = fluid.layers.data(name='img', shape=[784]) + >>> label = fluid.layers.data(name='label', shape=[1], dtype='int64') + >>> feeder = fluid.DataFeeder(feed_list=[img, label], place=fluid.CPUPlace()) + >>> # mnist.recordio will be generated in current directory + >>> fluid.recordio_writer.convert_reader_to_recordio_file( + >>> filename="mnist.recordio", + >>> reader_creator=paddle.batch(mnist.train(), batch_size=32), + >>> feeder=feeder) + + Args: + filename(str): The recordio filename. + reader_creator(callable): The Python Reader Creator. See + :ref:`api_guide_python_reader`. + feeder(DataFeeder): The DataFeeder instance. Used to convert + :code:`reader_creator` to :code: `lod_tensor` + compressor: Must in fluid.core.RecordIOWriter.Compressor.Snappy or + fluid.core.RecordIOWriter.Compressor.NoCompress. Use :code:`Snappy` + by default. + max_num_records(int): Maximum number of records in one chuck. Each record + is each return value from reader function + feed_order(list): The order of variable names that the reader returns + + Returns: + int: the number of record that saved. + """ if feed_order is None: feed_order = feeder.feed_names counter = 0 @@ -46,3 +87,47 @@ def convert_reader_to_recordio_file( writer.complete_append_tensor() counter += 1 return counter + + +def convert_reader_to_recordio_files( + filename, + batch_per_file, + reader_creator, + feeder, + compressor=core.RecordIOWriter.Compressor.Snappy, + max_num_records=1000, + feed_order=None): + """ + convert a python reader to many recordio files. + + This API is basically same as :code:`convert_reader_to_recordio_file`, + instead of it will create many recordio files. Each file contains at + most :code:`batch_per_file` records. + + Please reference + :ref:`api_fluid_recordio_writer_convert_reader_to_recordio_file` for more + details. + """ + if feed_order is None: + feed_order = feeder.feed_names + f_name, f_ext = os.path.splitext(filename) + assert (f_ext == ".recordio") + + lines = [] + f_idx = 0 + counter = 0 + for idx, batch in enumerate(reader_creator()): + lines.append(batch) + if idx >= batch_per_file and idx % batch_per_file == 0: + filename = "%s-%05d%s" % (f_name, f_idx, f_ext) + with create_recordio_writer(filename, compressor, + max_num_records) as writer: + for l in lines: + res = feeder.feed(l) + for each in feed_order: + writer.append_tensor(res[each]) + writer.complete_append_tensor() + counter += 1 + lines = [] + f_idx += 1 + return counter diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index c4d6829599616cb3ea7791a189e7070974de6ae3..6eaac4432d4df1288f37607a01484434542f1138 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -12,13 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. -import framework +from . import framework from . import core -__all__ = [ - 'append_regularization_ops', 'WeightDecayRegularizer', 'L1Decay', 'L2Decay', - 'L1DecayRegularizer', 'L2DecayRegularizer' -] +__all__ = ['L1Decay', 'L2Decay', 'L1DecayRegularizer', 'L2DecayRegularizer'] def append_regularization_ops(parameters_and_grads, regularization=None): @@ -36,19 +33,19 @@ def append_regularization_ops(parameters_and_grads, regularization=None): set. It will be applied with regularizer. Returns: - list of (parameters, gradients) pair with the regularized gradient + list[(Variable, Variable)]: list of (parameters, gradients) \ + pair with the regularized gradient Raises: Exception: Unknown regularization type """ params_and_grads = [] for param, grad in parameters_and_grads: - with param.block.program.optimized_guard(param): - # If no gradient then we don't need to do anything - if grad is None: - params_and_grads.append((param, grad)) - continue - + # If no gradient then we don't need to do anything + if grad is None: + params_and_grads.append((param, grad)) + continue + with param.block.program.optimized_guard([param, grad]): regularization_term = None if param.regularizer is not None: # Add variable for regularization term in grad block @@ -100,6 +97,24 @@ class WeightDecayRegularizer(object): class L2DecayRegularizer(WeightDecayRegularizer): """Implements the L2 Weight Decay Regularization + + Small values of L2 can help prevent over fitting the training data. + + .. math:: + + L2WeightDecay = reg\_coeff * parameter + + Args: + regularization_coeff(float): regularization coeff + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)) + optimizer.minimize(avg_cost) """ def __init__(self, regularization_coeff=0.0): @@ -127,14 +142,20 @@ class L2DecayRegularizer(WeightDecayRegularizer): dtype="float32", shape=param.shape, lod_level=param.lod_level) if grad.type == core.VarDesc.VarType.SELECTED_ROWS: + idx = block.create_var( + dtype="int64", + shape=param.shape, + type=core.VarDesc.VarType.LOD_TENSOR) decay = block.create_var( dtype="float32", shape=param.shape, type=core.VarDesc.VarType.SELECTED_ROWS) + block.append_op( + type='extract_rows', inputs={'X': grad}, outputs={'Out': idx}) block.append_op( type='lookup_table', inputs={'W': param, - 'Ids': grad}, + 'Ids': idx}, outputs={'Out': decay}, attrs={'is_sparse': True}) param = decay @@ -154,6 +175,27 @@ class L2DecayRegularizer(WeightDecayRegularizer): class L1DecayRegularizer(WeightDecayRegularizer): """Implements the L1 Weight Decay Regularization + + L1 regularization encourages sparsity. + + .. math:: + + L1WeightDecay = reg\_coeff * sign(parameter) + + Args: + regularization_coeff(float): regularization coeff + + Examples: + .. code-block:: python + + program = fluid.framework.Program() + block = program.global_block() + mul_x = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + regularizer=fluid.regularizer.L1DecayRegularizer(0.5)) """ def __init__(self, regularization_coeff=0.0): @@ -180,14 +222,20 @@ class L1DecayRegularizer(WeightDecayRegularizer): dtype="float32", shape=param.shape, lod_level=param.lod_level) if grad.type == core.VarDesc.VarType.SELECTED_ROWS: + idx = block.create_var( + dtype="int64", + shape=param.shape, + type=core.VarDesc.VarType.LOD_TENSOR) decay = block.create_var( dtype="float32", shape=param.shape, type=core.VarDesc.VarType.SELECTED_ROWS) + block.append_op( + type='extract_rows', inputs={'X': grad}, outputs={'Out': idx}) block.append_op( type='lookup_table', inputs={'W': param, - 'Ids': grad}, + 'Ids': idx}, outputs={'Out': decay}, attrs={'is_sparse': True}) diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py index de3906fc6a005181b0ab04a846eb2e7ce14004c2..36a1a223cfd7c69aff3e8648da990d23e4e75202 100644 --- a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py @@ -38,7 +38,7 @@ def inference_program(): return y_predict -def linear(): +def train_program(): y = fluid.layers.data(name='y', shape=[1], dtype='float32') y_predict = inference_program() @@ -48,20 +48,22 @@ def linear(): return avg_loss +def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, - place=place, - optimizer=fluid.optimizer.SGD(learning_rate=0.001)) + train_func=train_program, place=place, optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndStepEvent): if event.step == 10: test_metrics = trainer.test( reader=test_reader, feed_order=['x', 'y']) - print test_metrics + print(test_metrics) ''' ... ['25.768919467926025'] @@ -102,7 +104,7 @@ def main(use_cuda): # Directory for saving the trained model params_dirname = "fit_a_line.inference.model" - train(use_cuda, linear, params_dirname) + train(use_cuda, train_program, params_dirname) infer(use_cuda, inference_program, params_dirname) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py index 7fed6d914f75b690e34411aa154359c93b6ca989..9e4c384d92943227c2d68da829e6019e649a35fb 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py @@ -28,11 +28,12 @@ images per class. """ -import cPickle import itertools import numpy import paddle.v2.dataset.common import tarfile +from six.moves import cPickle as pickle +from six.moves import zip __all__ = ['train10'] @@ -46,7 +47,7 @@ def reader_creator(filename, sub_name, batch_size=None): data = batch['data'] labels = batch.get('labels', batch.get('fine_labels', None)) assert labels is not None - for sample, label in itertools.izip(data, labels): + for sample, label in zip(data, labels): yield (sample / 255.0).astype(numpy.float32), int(label) def reader(): @@ -56,7 +57,7 @@ def reader_creator(filename, sub_name, batch_size=None): batch_count = 0 for name in names: - batch = cPickle.load(f.extractfile(name)) + batch = pickle.load(f.extractfile(name)) for item in read_batch(batch): if isinstance(batch_size, int) and batch_count > batch_size: break diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py index 63dc1b6ce30974ede22a3f7772b76bf207bbae39..a1f62db093904b617f0e37dc20d586ccea7eacd2 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid import numpy @@ -85,6 +83,10 @@ def train_network(): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): BATCH_SIZE = 128 EPOCH_NUM = 1 @@ -92,10 +94,11 @@ def train(use_cuda, train_program, params_dirname): train_reader = paddle.batch( paddle.reader.shuffle( cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), - batch_size=BATCH_SIZE) + batch_size=BATCH_SIZE, + drop_last=False) test_reader = paddle.batch( - paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False) def event_handler(event): if isinstance(event, fluid.EndStepEvent): @@ -111,9 +114,7 @@ def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, - optimizer=fluid.optimizer.Adam(learning_rate=0.001), - place=place) + train_func=train_program, optimizer_func=optimizer_func, place=place) trainer.train( reader=train_reader, diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py index 0bf8f265a1c1b11364ecfa11061af183ce20d51e..8429551765740e7db0eda82ce0b17cff129359b0 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid import numpy @@ -64,15 +62,20 @@ def train_network(): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): BATCH_SIZE = 128 train_reader = paddle.batch( paddle.reader.shuffle( cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), - batch_size=BATCH_SIZE) + batch_size=BATCH_SIZE, + drop_last=False) test_reader = paddle.batch( - paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False) def event_handler(event): if isinstance(event, fluid.EndStepEvent): @@ -88,9 +91,7 @@ def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, - place=place, - optimizer=fluid.optimizer.Adam(learning_rate=0.001)) + train_func=train_program, place=place, optimizer_func=optimizer_func) trainer.train( reader=train_reader, diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py index 8cce398ff33695dc15ae6fb01a887194596af001..e3602e2d5643c233b2575d1adb7f181127f60287 100755 --- a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py +++ b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid import numpy as np @@ -141,12 +139,16 @@ def train_program(): return [avg_cost] +def optimize_func(): + return fluid.optimizer.SGD(learning_rate=fluid.layers.exponential_decay( + learning_rate=0.01, decay_steps=100000, decay_rate=0.5, staircase=True)) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.SGD(learning_rate=0.01) trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer=optimizer) + train_func=train_program, place=place, optimizer_func=optimize_func) feed_order = [ 'word_data', 'ctx_n2_data', 'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data', @@ -174,14 +176,15 @@ def train(use_cuda, train_program, params_dirname): if float(avg_cost) < 100.0: # Large value to increase CI speed trainer.save_params(params_dirname) else: - print('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, - float(avg_cost))) + print( + ('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, + float(avg_cost)))) if math.isnan(float(avg_cost)): sys.exit("got NaN loss, training failed.") elif isinstance(event, fluid.EndStepEvent): print("Step {0}, Epoch {1} Metrics {2}".format( - event.step, event.epoch, map(np.array, event.metrics))) + event.step, event.epoch, list(map(np.array, event.metrics)))) if event.step == 1: # Run 2 iterations to speed CI trainer.save_params(params_dirname) trainer.stop() @@ -202,35 +205,35 @@ def infer(use_cuda, inference_program, params_dirname): inferencer = fluid.Inferencer( inference_program, param_path=params_dirname, place=place) - # Setup inputs by creating LoDTensors to represent sequences of words. - # Here each word is the basic element of these LoDTensors and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], - # which has only one lod level. Then the created LoDTensors will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that lod info should be a list of lists. - lod = [[3, 4, 2]] + # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[3, 4, 2]] base_shape = [1] # The range of random integers is [low, high] word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=WORD_DICT_LEN - 1) ctx_n2 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=WORD_DICT_LEN - 1) ctx_n1 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=WORD_DICT_LEN - 1) ctx_0 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=WORD_DICT_LEN - 1) ctx_p1 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=WORD_DICT_LEN - 1) ctx_p2 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=WORD_DICT_LEN - 1) pred = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=PRED_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=PRED_DICT_LEN - 1) mark = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=MARK_DICT_LEN - 1) + recursive_seq_lens, base_shape, place, low=0, high=MARK_DICT_LEN - 1) results = inferencer.infer( { @@ -245,7 +248,7 @@ def infer(use_cuda, inference_program, params_dirname): }, return_numpy=False) - print("infer results: ", np.array(results[0])) + print("infer results: ", np.array(results[0]).shape) def main(use_cuda): diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py index d4b723d3e6b619709ab3dc76a32ae87f1cdec274..6fb0c85a8be2b4560ea1fdb32f01146a9206ee78 100644 --- a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py @@ -127,9 +127,19 @@ def decode(context, is_sparse): current_score = pd.fc(input=current_state_with_lod, size=target_dict_dim, act='softmax') - topk_scores, topk_indices = pd.topk(current_score, k=topk_size) + topk_scores, topk_indices = pd.topk(current_score, k=beam_size) + # calculate accumulated scores after topk to reduce computation cost + accu_scores = pd.elementwise_add( + x=pd.log(topk_scores), y=pd.reshape( + pre_score, shape=[-1]), axis=0) selected_ids, selected_scores = pd.beam_search( - pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) + pre_ids, + pre_score, + topk_indices, + accu_scores, + beam_size, + end_id=10, + level=0) pd.increment(x=counter, value=1, in_place=True) @@ -138,10 +148,14 @@ def decode(context, is_sparse): pd.array_write(selected_ids, array=ids_array, i=counter) pd.array_write(selected_scores, array=scores_array, i=counter) - pd.less_than(x=counter, y=array_len, cond=cond) + # update the break condition: up to the max length or all candidates of + # source sentences have ended. + length_cond = pd.less_than(x=counter, y=array_len) + finish_cond = pd.logical_not(pd.is_empty(x=selected_ids)) + pd.logical_and(x=length_cond, y=finish_cond, out=cond) translation_ids, translation_scores = pd.beam_search_decode( - ids=ids_array, scores=scores_array) + ids=ids_array, scores=scores_array, beam_size=beam_size, end_id=10) # return init_ids, init_scores @@ -158,6 +172,13 @@ def train_program(is_sparse): return avg_cost +def optimizer_func(): + return fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)) + + def train(use_cuda, is_sparse, is_local=True): EPOCH_NUM = 1 @@ -182,11 +203,8 @@ def train(use_cuda, is_sparse, is_local=True): trainer = fluid.Trainer( train_func=partial(train_program, is_sparse), - optimizer=fluid.optimizer.Adagrad( - learning_rate=1e-4, - regularization=fluid.regularizer.L2DecayRegularizer( - regularization_coeff=0.1)), - place=place) + place=place, + optimizer_func=optimizer_func) trainer.train( reader=train_reader, @@ -211,11 +229,13 @@ def decode_main(use_cuda, is_sparse): [1. for _ in range(batch_size)], dtype='float32') init_ids_data = init_ids_data.reshape((batch_size, 1)) init_scores_data = init_scores_data.reshape((batch_size, 1)) - init_lod = [1] * batch_size - init_lod = [init_lod, init_lod] + init_recursive_seq_lens = [1] * batch_size + init_recursive_seq_lens = [init_recursive_seq_lens, init_recursive_seq_lens] - init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) - init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) + init_ids = fluid.create_lod_tensor(init_ids_data, init_recursive_seq_lens, + place) + init_scores = fluid.create_lod_tensor(init_scores_data, + init_recursive_seq_lens, place) train_data = paddle.batch( paddle.reader.shuffle( @@ -230,7 +250,7 @@ def decode_main(use_cuda, is_sparse): feeder = fluid.DataFeeder(feed_list, place) for data in train_data(): - feed_dict = feeder.feed(map(lambda x: [x[0]], data)) + feed_dict = feeder.feed([[x[0]] for x in data]) feed_dict['init_ids'] = init_ids feed_dict['init_scores'] = init_scores @@ -239,7 +259,7 @@ def decode_main(use_cuda, is_sparse): feed=feed_dict, fetch_list=[translation_ids, translation_scores], return_numpy=False) - print result_ids.lod() + print(result_ids.recursive_sequence_lengths()) break diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py index 03439cbd37671b4727879bf3d0793f016f55247a..898807db6f343cbefcc877e0f03ed6c5b82dd669 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py @@ -11,9 +11,10 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function + import argparse import paddle.fluid as fluid +import paddle.fluid.core as core import paddle import sys import numpy @@ -57,14 +58,17 @@ def train_program(): return [avg_cost, acc] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adam(learning_rate=0.001) trainer = fluid.Trainer( train_func=train_program, place=place, - optimizer=optimizer, + optimizer_func=optimizer_func, parallel=True) def event_handler(event): @@ -85,8 +89,10 @@ def train(use_cuda, train_program, params_dirname): if math.isnan(avg_cost): sys.exit("got NaN loss, training failed.") elif isinstance(event, fluid.EndStepEvent): - print("Step {0}, Epoch {1} Metrics {2}".format( - event.step, event.epoch, map(numpy.array, event.metrics))) + print( + ("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, + list(map(numpy.array, event.metrics))))) train_reader = paddle.batch( paddle.reader.shuffle( @@ -131,4 +137,4 @@ def main(use_cuda): if __name__ == '__main__': # for use_cuda in (False, True): - main(use_cuda=True) + main(use_cuda=core.is_compiled_with_cuda()) diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py index 89bbd21bea7d64a8dd6fc32829b6addb680da62e..6dd64be315159f1835244fa027e578434e6cb038 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py @@ -11,7 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function + import argparse import paddle.fluid as fluid import paddle @@ -44,12 +44,15 @@ def train_program(): return [avg_cost, acc] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adam(learning_rate=0.001) trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer=optimizer) + train_func=train_program, place=place, optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py index dfc7325acf23176c05fe42761b9997b98d23372a..60f3d8e105209938360487d963b0328d95e7b1f0 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py +++ b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py @@ -155,12 +155,15 @@ def train_program(): return [avg_cost, scale_infer] +def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.2) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.SGD(learning_rate=0.2) trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer=optimizer) + train_func=train_program, place=place, optimizer_func=optimizer_func) feed_order = [ 'user_id', 'gender_id', 'age_id', 'job_id', 'movie_id', 'category_id', @@ -183,8 +186,9 @@ def train(use_cuda, train_program, params_dirname): trainer.save_params(params_dirname) trainer.stop() else: - print('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, - float(avg_cost))) + print( + ('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, + float(avg_cost)))) if math.isnan(float(avg_cost)): sys.exit("got NaN loss, training failed.") @@ -206,13 +210,15 @@ def infer(use_cuda, inference_program, params_dirname): inference_program, param_path=params_dirname, place=place) # Use the first data from paddle.dataset.movielens.test() as input. - # Use create_lod_tensor(data, lod, place) API to generate LoD Tensor, - # where `data` is a list of sequences of index numbers, `lod` is - # the level of detail (lod) info associated with `data`. + # Use create_lod_tensor(data, recursive_sequence_lengths, place) API + # to generate LoD Tensor where `data` is a list of sequences of index + # numbers, `recursive_sequence_lengths` is the length-based level of detail + # (lod) info associated with `data`. # For example, data = [[10, 2, 3], [2, 3]] means that it contains # two sequences of indexes, of length 3 and 2, respectively. - # Correspondingly, lod = [[3, 2]] contains one level of detail info, - # indicating that `data` consists of two sequences of length 3 and 2. + # Correspondingly, recursive_sequence_lengths = [[3, 2]] contains one + # level of detail info, indicating that `data` consists of two sequences + # of length 3 and 2, respectively. user_id = fluid.create_lod_tensor([[1]], [[1]], place) gender_id = fluid.create_lod_tensor([[1]], [[1]], place) age_id = fluid.create_lod_tensor([[0]], [[1]], place) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py index 11e9fd1bec1450f6753dbe38c7014090d6e136b6..24e65d1bd54cff7ad64453a3a61f50351d32ef08 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid from functools import partial @@ -64,15 +62,18 @@ def train_program(word_dict): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adagrad(learning_rate=0.002) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() trainer = fluid.Trainer( train_func=partial(train_program, word_dict), place=place, - optimizer=optimizer) + optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): @@ -95,7 +96,7 @@ def train(use_cuda, train_program, params_dirname): sys.exit("got NaN loss, training failed.") elif isinstance(event, fluid.EndStepEvent): print("Step {0}, Epoch {1} Metrics {2}".format( - event.step, event.epoch, map(np.array, event.metrics))) + event.step, event.epoch, list(map(np.array, event.metrics)))) if event.step == 1: # Run 2 iterations to speed CI trainer.save_params(params_dirname) trainer.stop() @@ -122,20 +123,20 @@ def infer(use_cuda, inference_program, params_dirname=None): place=place) # Setup input by creating LoDTensor to represent sequence of words. - # Here each word is the basic element of the LoDTensor and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], - # which has only one lod level. Then the created LoDTensor will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that lod info should be a list of lists. - lod = [[3, 4, 2]] + # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[3, 4, 2]] base_shape = [1] # The range of random integers is [low, high] tensor_words = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=len(word_dict) - 1) + recursive_seq_lens, base_shape, place, low=0, high=len(word_dict) - 1) results = inferencer.infer({'words': tensor_words}) print("infer results: ", results) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py index 90757d54f99715163518ce5a094e6ba3a67efed3..b3b1505a0fad07144f3f53c22abd5553054d8c51 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid from functools import partial @@ -79,15 +77,18 @@ def train_program(word_dict): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adagrad(learning_rate=0.002) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() trainer = fluid.Trainer( train_func=partial(train_program, word_dict), place=place, - optimizer=optimizer) + optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): @@ -110,7 +111,7 @@ def train(use_cuda, train_program, params_dirname): sys.exit("got NaN loss, training failed.") elif isinstance(event, fluid.EndStepEvent): print("Step {0}, Epoch {1} Metrics {2}".format( - event.step, event.epoch, map(np.array, event.metrics))) + event.step, event.epoch, list(map(np.array, event.metrics)))) if event.step == 1: # Run 2 iterations to speed CI trainer.save_params(params_dirname) trainer.stop() @@ -137,20 +138,20 @@ def infer(use_cuda, inference_program, params_dirname=None): place=place) # Setup input by creating LoDTensor to represent sequence of words. - # Here each word is the basic element of the LoDTensor and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], - # which has only one lod level. Then the created LoDTensor will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that lod info should be a list of lists. - lod = [[3, 4, 2]] + # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[3, 4, 2]] base_shape = [1] # The range of random integers is [low, high] tensor_words = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=len(word_dict) - 1) + recursive_seq_lens, base_shape, place, low=0, high=len(word_dict) - 1) results = inferencer.infer({'words': tensor_words}) print("infer results: ", results) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py index 52b7d4a83779d01936afb3d9d1e4864b05d55b5a..25f99ff0fd2d1050bb62338a6bf87aa29f913fb6 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid from functools import partial @@ -71,20 +69,25 @@ def train_program(word_dict): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adagrad(learning_rate=0.002) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() trainer = fluid.Trainer( train_func=partial(train_program, word_dict), place=place, - optimizer=optimizer) + optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): test_reader = paddle.batch( - paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + paddle.dataset.imdb.test(word_dict), + batch_size=BATCH_SIZE, + drop_last=False) avg_cost, acc = trainer.test( reader=test_reader, feed_order=['words', 'label']) @@ -102,7 +105,7 @@ def train(use_cuda, train_program, params_dirname): sys.exit("got NaN loss, training failed.") elif isinstance(event, fluid.EndStepEvent): print("Step {0}, Epoch {1} Metrics {2}".format( - event.step, event.epoch, map(np.array, event.metrics))) + event.step, event.epoch, list(map(np.array, event.metrics)))) if event.step == 1: # Run 2 iterations to speed CI trainer.save_params(params_dirname) trainer.stop() @@ -110,7 +113,8 @@ def train(use_cuda, train_program, params_dirname): train_reader = paddle.batch( paddle.reader.shuffle( paddle.dataset.imdb.train(word_dict), buf_size=25000), - batch_size=BATCH_SIZE) + batch_size=BATCH_SIZE, + drop_last=False) trainer.train( num_epochs=1, @@ -129,20 +133,20 @@ def infer(use_cuda, inference_program, params_dirname=None): place=place) # Setup input by creating LoDTensor to represent sequence of words. - # Here each word is the basic element of the LoDTensor and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], - # which has only one lod level. Then the created LoDTensor will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that lod info should be a list of lists. - lod = [[3, 4, 2]] + # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[3, 4, 2]] base_shape = [1] # The range of random integers is [low, high] tensor_words = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=len(word_dict) - 1) + recursive_seq_lens, base_shape, place, low=0, high=len(word_dict) - 1) results = inferencer.infer({'words': tensor_words}) print("infer results: ", results) diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py index eeb8e67087334ea96aab9cdb6272e34e2eb99939..02e65cf56c4d1bd262831320befd2edc735c0d1c 100644 --- a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py +++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py @@ -80,6 +80,10 @@ def train_program(is_sparse): return avg_cost +def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): train_reader = paddle.batch( paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) @@ -104,9 +108,7 @@ def train(use_cuda, train_program, params_dirname): sys.exit("got NaN loss, training failed.") trainer = fluid.Trainer( - train_func=train_program, - optimizer=fluid.optimizer.SGD(learning_rate=0.001), - place=place) + train_func=train_program, optimizer_func=optimizer_func, place=place) trainer.train( reader=train_reader, @@ -122,21 +124,22 @@ def infer(use_cuda, inference_program, params_dirname=None): # Setup inputs by creating 4 LoDTensors representing 4 words. Here each word # is simply an index to look up for the corresponding word vector and hence - # the shape of word (base_shape) should be [1]. The length-based level of - # detail (lod) info of each LoDtensor should be [[1]] meaning there is only - # one lod_level and there is only one sequence of one word on this level. - # Note that lod info should be a list of lists. - lod = [[1]] + # the shape of word (base_shape) should be [1]. The recursive_sequence_lengths, + # which is length-based level of detail (lod) of each LoDTensor, should be [[1]] + # meaning there is only one level of detail and there is only one sequence of + # one word on this level. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[1]] base_shape = [1] # The range of random integers is [low, high] first_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) second_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) third_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) fourth_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) result = inferencer.infer( { diff --git a/python/paddle/fluid/tests/book/notest_understand_sentiment.py b/python/paddle/fluid/tests/book/notest_understand_sentiment.py index c6687e8ad7fcc45c82d6dcb2256e9055a81cc61c..ce6342c2dad0b33e57d0ea90fc6ef1660ae4e68b 100644 --- a/python/paddle/fluid/tests/book/notest_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/notest_understand_sentiment.py @@ -11,8 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function +from paddle.fluid.layers.device import get_places import unittest import paddle.fluid as fluid import paddle @@ -144,7 +144,7 @@ def train(word_dict, cost, acc_out, prediction = net_method( data, label, input_dim=dict_dim, class_dim=class_dim) else: - places = fluid.layers.get_places() + places = get_places() pd = fluid.layers.ParallelDo(places) with pd.do(): cost, acc, _ = net_method( @@ -175,7 +175,7 @@ def train(word_dict, def train_loop(main_program): exe.run(fluid.default_startup_program()) - for pass_id in xrange(PASS_NUM): + for pass_id in range(PASS_NUM): for data in train_data(): cost_val, acc_val = exe.run(main_program, feed=feeder.feed(data), @@ -194,16 +194,16 @@ def train(word_dict, if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -235,20 +235,24 @@ def infer(word_dict, use_cuda, save_dirname=None): word_dict_len = len(word_dict) # Setup input by creating LoDTensor to represent sequence of words. - # Here each word is the basic element of the LoDTensor and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], - # which has only one lod level. Then the created LoDTensor will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that lod info should be a list of lists. - lod = [[3, 4, 2]] + # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[3, 4, 2]] base_shape = [1] # The range of random integers is [low, high] tensor_words = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. @@ -257,7 +261,7 @@ def infer(word_dict, use_cuda, save_dirname=None): feed={feed_target_names[0]: tensor_words}, fetch_list=fetch_targets, return_numpy=False) - print(results[0].lod()) + print(results[0].recursive_sequence_lengths()) np_data = np.array(results[0]) print("Inference Shape: ", np_data.shape) print("Inference results: ", np_data) diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py index b1a6b524d33cae97c8982ffb8f780b1b07761a09..37b64fa94a9aad7042e153e414ed29de3142db5a 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -69,16 +69,16 @@ def train(use_cuda, save_dirname, is_local): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -110,14 +110,23 @@ def infer(use_cuda, save_dirname=None): # The input's dimension should be 2-D and the second dim is 13 # The input data should be >= 0 batch_size = 10 - tensor_x = numpy.random.uniform(0, 10, - [batch_size, 13]).astype("float32") + + test_reader = paddle.batch( + paddle.dataset.uci_housing.test(), batch_size=batch_size) + + test_data = next(test_reader()) + test_feat = numpy.array( + [data[0] for data in test_data]).astype("float32") + test_label = numpy.array( + [data[1] for data in test_data]).astype("float32") + assert feed_target_names[0] == 'x' results = exe.run(inference_program, - feed={feed_target_names[0]: tensor_x}, + feed={feed_target_names[0]: numpy.array(test_feat)}, fetch_list=fetch_targets) print("infer shape: ", results[0].shape) print("infer results: ", results[0]) + print("ground truth: ", test_label) def main(use_cuda, is_local=True): diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index 0f3a4c9242a81a3c1fb90268245715a8e59a207a..de6fe5f140a86545e3291db165af824739a814ef 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import paddle import paddle.fluid as fluid import contextlib @@ -121,7 +119,7 @@ def train(net_type, use_cuda, save_dirname, is_local): avg_cost = fluid.layers.mean(cost) acc = fluid.layers.accuracy(input=predict, label=label) - # Test program + # Test program test_program = fluid.default_main_program().clone(for_test=True) optimizer = fluid.optimizer.Adam(learning_rate=0.001) @@ -178,16 +176,16 @@ def train(net_type, use_cuda, save_dirname, is_local): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index bc8a1aafc82d62501cecfa71be0cc3851c75eae2..b7ac911cafdc751f38c7f66bc48263a17a84dc08 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -76,8 +76,7 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, emb_layers.append(mark_embedding) hidden_0_layers = [ - fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') - for emb in emb_layers + fluid.layers.fc(input=emb, size=hidden_dim) for emb in emb_layers ] hidden_0 = fluid.layers.sums(input=hidden_0_layers) @@ -94,8 +93,8 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, for i in range(1, depth): mix_hidden = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') + fluid.layers.fc(input=input_tmp[0], size=hidden_dim), + fluid.layers.fc(input=input_tmp[1], size=hidden_dim) ]) lstm = fluid.layers.dynamic_lstm( @@ -182,7 +181,7 @@ def train(use_cuda, save_dirname=None, is_local=True): start_time = time.time() batch_id = 0 - for pass_id in xrange(PASS_NUM): + for pass_id in range(PASS_NUM): for data in train_data(): cost = exe.run(main_program, feed=feeder.feed(data), @@ -210,16 +209,16 @@ def train(use_cuda, save_dirname=None, is_local=True): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -248,35 +247,67 @@ def infer(use_cuda, save_dirname=None): [inference_program, feed_target_names, fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) - # Setup inputs by creating LoDTensors to represent sequences of words. - # Here each word is the basic element of these LoDTensors and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], - # which has only one lod level. Then the created LoDTensors will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that lod info should be a list of lists. - lod = [[3, 4, 2]] + # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[3, 4, 2]] base_shape = [1] # The range of random integers is [low, high] word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) pred = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=pred_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=pred_dict_len - 1) ctx_n2 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) ctx_n1 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) ctx_0 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) ctx_p1 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) ctx_p2 = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=word_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=word_dict_len - 1) mark = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=mark_dict_len - 1) + recursive_seq_lens, + base_shape, + place, + low=0, + high=mark_dict_len - 1) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. @@ -302,7 +333,7 @@ def infer(use_cuda, save_dirname=None): }, fetch_list=fetch_targets, return_numpy=False) - print(results[0].lod()) + print(results[0].recursive_sequence_lengths()) np_data = np.array(results[0]) print("Inference Shape: ", np_data.shape) diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 23e5900f127a7a3253c551f8f7fbceba08382209..462faad3e1cb7108f3bd6934017efe25fb9a4276 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -108,7 +108,7 @@ def decoder_decode(context, is_sparse): pre_state = pd.array_read(array=state_array, i=counter) pre_score = pd.array_read(array=scores_array, i=counter) - # expand the lod of pre_state to be the same with pre_score + # expand the recursive_sequence_lengths of pre_state to be the same with pre_score pre_state_expanded = pd.sequence_expand(pre_state, pre_score) pre_ids_emb = pd.embedding( @@ -126,9 +126,19 @@ def decoder_decode(context, is_sparse): current_score = pd.fc(input=current_state_with_lod, size=target_dict_dim, act='softmax') - topk_scores, topk_indices = pd.topk(current_score, k=50) + topk_scores, topk_indices = pd.topk(current_score, k=beam_size) + # calculate accumulated scores after topk to reduce computation cost + accu_scores = pd.elementwise_add( + x=pd.log(topk_scores), y=pd.reshape( + pre_score, shape=[-1]), axis=0) selected_ids, selected_scores = pd.beam_search( - pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) + pre_ids, + pre_score, + topk_indices, + accu_scores, + beam_size, + end_id=10, + level=0) pd.increment(x=counter, value=1, in_place=True) @@ -137,10 +147,14 @@ def decoder_decode(context, is_sparse): pd.array_write(selected_ids, array=ids_array, i=counter) pd.array_write(selected_scores, array=scores_array, i=counter) - pd.less_than(x=counter, y=array_len, cond=cond) + # update the break condition: up to the max length or all candidates of + # source sentences have ended. + length_cond = pd.less_than(x=counter, y=array_len) + finish_cond = pd.logical_not(pd.is_empty(x=selected_ids)) + pd.logical_and(x=length_cond, y=finish_cond, out=cond) translation_ids, translation_scores = pd.beam_search_decode( - ids=ids_array, scores=scores_array) + ids=ids_array, scores=scores_array, beam_size=beam_size, end_id=10) # return init_ids, init_scores @@ -185,7 +199,7 @@ def train_main(use_cuda, is_sparse, is_local=True): feeder = fluid.DataFeeder(feed_list, place) batch_id = 0 - for pass_id in xrange(1): + for pass_id in range(1): for data in train_data(): outs = exe.run(main_program, feed=feeder.feed(data), @@ -200,16 +214,16 @@ def train_main(use_cuda, is_sparse, is_local=True): if is_local: train_loop(framework.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -238,11 +252,13 @@ def decode_main(use_cuda, is_sparse): [1. for _ in range(batch_size)], dtype='float32') init_ids_data = init_ids_data.reshape((batch_size, 1)) init_scores_data = init_scores_data.reshape((batch_size, 1)) - init_lod = [1] * batch_size - init_lod = [init_lod, init_lod] + init_recursive_seq_lens = [1] * batch_size + init_recursive_seq_lens = [init_recursive_seq_lens, init_recursive_seq_lens] - init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) - init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) + init_ids = fluid.create_lod_tensor(init_ids_data, init_recursive_seq_lens, + place) + init_scores = fluid.create_lod_tensor(init_scores_data, + init_recursive_seq_lens, place) train_data = paddle.batch( paddle.reader.shuffle( @@ -257,7 +273,7 @@ def decode_main(use_cuda, is_sparse): feeder = fluid.DataFeeder(feed_list, place) for data in train_data(): - feed_dict = feeder.feed(map(lambda x: [x[0]], data)) + feed_dict = feeder.feed([[x[0]] for x in data]) feed_dict['init_ids'] = init_ids feed_dict['init_scores'] = init_scores @@ -266,7 +282,7 @@ def decode_main(use_cuda, is_sparse): feed=feed_dict, fetch_list=[translation_ids, translation_scores], return_numpy=False) - print result_ids.lod() + print(result_ids.recursive_sequence_lengths()) break diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index 578b1162fbd7e3a1b1c0cc934406818f2e07e019..3e5f76d12d41d016c995e5c85feda3c1847e356f 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -11,16 +11,18 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function -import argparse -import paddle.fluid as fluid -import paddle -import sys -import numpy -import unittest + +import paddle.fluid.core as core import math -import sys import os +import sys +import unittest + +import numpy + +import paddle +import paddle.fluid as fluid +from paddle.fluid.layers.device import get_places BATCH_SIZE = 64 @@ -76,7 +78,7 @@ def train(nn_type, net_conf = conv_net if parallel: - places = fluid.layers.get_places() + places = get_places() pd = fluid.layers.ParallelDo(places) with pd.do(): img_ = pd.read_input(img) @@ -94,7 +96,7 @@ def train(nn_type, test_program = fluid.default_main_program().clone(for_test=True) - optimizer = fluid.optimizer.Adam(learning_rate=0.001) + optimizer = fluid.optimizer.Adam(learning_rate=0.001, LARS_weight_decay=0.3) optimizer.minimize(avg_loss) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() @@ -151,16 +153,16 @@ def train(nn_type, if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -255,6 +257,8 @@ def inject_test_method(use_cuda, parallel, nn_type, combine): def inject_all_tests(): for use_cuda in (False, True): + if use_cuda and not core.is_compiled_with_cuda(): + continue for parallel in (False, True): for nn_type in ('mlp', 'conv'): inject_test_method(use_cuda, parallel, nn_type, True) diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index 65d6552acc9b3d31a97a45290e4613a633fffa3c..b30c8771fcf267260c4c5aa7076bedc89e3b7e8b 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -220,16 +220,16 @@ def train(use_cuda, save_dirname, is_local=True): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -260,13 +260,15 @@ def infer(use_cuda, save_dirname=None): # Use the first data from paddle.dataset.movielens.test() as input assert feed_target_names[0] == "user_id" - # Use create_lod_tensor(data, lod, place) API to generate LoD Tensor - # where `data` is a list of sequences of index numbers, `lod` is - # the level of detail (lod) info associated with `data`. + # Use create_lod_tensor(data, recursive_sequence_lengths, place) API + # to generate LoD Tensor where `data` is a list of sequences of index + # numbers, `recursive_sequence_lengths` is the length-based level of detail + # (lod) info associated with `data`. # For example, data = [[10, 2, 3], [2, 3]] means that it contains # two sequences of indexes, of length 3 and 2, respectively. - # Correspondingly, lod = [[3, 2]] contains one level of detail info, - # indicating that `data` consists of two sequences of length 3 and 2. + # Correspondingly, recursive_sequence_lengths = [[3, 2]] contains one + # level of detail info, indicating that `data` consists of two sequences + # of length 3 and 2, respectively. user_id = fluid.create_lod_tensor([[1]], [[1]], place) assert feed_target_names[1] == "gender_id" diff --git a/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py b/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py index 7ada57def6bfedb113ea1a56f9677116b80488ea..2e79be2bd0fc7a368df86e188b7fa616055bb3e7 100644 --- a/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py +++ b/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py @@ -175,7 +175,7 @@ def train(use_cuda, save_dirname=None): feeder = fluid.DataFeeder(feed_list, place) batch_id = 0 - for pass_id in xrange(2): + for pass_id in range(2): for data in train_data(): outs = exe.run(framework.default_main_program(), feed=feeder.feed(data), @@ -213,22 +213,22 @@ def infer(use_cuda, save_dirname=None): fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) # Setup input by creating LoDTensor to represent sequence of words. - # Here each word is the basic element of the LoDTensor and the shape of - # each word (base_shape) should be [1] since it is simply an index to + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to # look up for the corresponding word vector. - # Suppose the length_based level of detail (lod) info is set to [[4, 6]], - # which has only one lod level. Then the created LoDTensor will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for two sentences of - # length 4 and 6, respectively. - # Note that lod info should be a list of lists. - lod = [[4, 6]] + # Suppose the recursive_sequence_lengths info is set to [[4, 6]], + # which has only one level of detail. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for two sentences of + # length 4 and 6, respectively. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[4, 6]] base_shape = [1] # The range of random integers is [low, high] word_data = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=1) + recursive_seq_lens, base_shape, place, low=0, high=1) trg_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=1) + recursive_seq_lens, base_shape, place, low=0, high=1) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. @@ -241,7 +241,7 @@ def infer(use_cuda, save_dirname=None): }, fetch_list=fetch_targets, return_numpy=False) - print(results[0].lod()) + print(results[0].recursive_sequence_lengths()) np_data = np.array(results[0]) print("Inference shape: ", np_data.shape) print("Inference results: ", np_data) diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index 3118d88701e5f64ae50f7ee774ea8174aa7758eb..e761e05795313da23a9d984263ac2e202939b1e7 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -14,6 +14,7 @@ import paddle import paddle.fluid as fluid +from paddle.fluid.layers.device import get_places import unittest import os import numpy as np @@ -80,13 +81,15 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): avg_cost, predict_word = __network__( [first_word, second_word, third_word, forth_word, next_word]) else: - places = fluid.layers.get_places() + places = get_places() pd = fluid.layers.ParallelDo(places) with pd.do(): avg_cost, predict_word = __network__( - map(pd.read_input, [ - first_word, second_word, third_word, forth_word, next_word - ])) + list( + map(pd.read_input, [ + first_word, second_word, third_word, forth_word, + next_word + ]))) pd.write_output(avg_cost) avg_cost = fluid.layers.mean(pd()) @@ -125,16 +128,16 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": @@ -166,23 +169,24 @@ def infer(use_cuda, save_dirname=None): word_dict = paddle.dataset.imikolov.build_dict() dict_size = len(word_dict) - # Setup inputs by creating 4 LoDTensors representing 4 words. Here each word - # is simply an index to look up for the corresponding word vector and hence - # the shape of word (base_shape) should be [1]. The length-based level of - # detail (lod) info of each LoDtensor should be [[1]] meaning there is only - # one lod_level and there is only one sequence of one word on this level. - # Note that lod info should be a list of lists. - lod = [[1]] + # Setup inputs by creating 4 LoDTensors representing 4 words. Here each word + # is simply an index to look up for the corresponding word vector and hence + # the shape of word (base_shape) should be [1]. The recursive_sequence_lengths, + # which is length-based level of detail (lod) of each LoDTensor, should be [[1]] + # meaning there is only one level of detail and there is only one sequence of + # one word on this level. + # Note that recursive_sequence_lengths should be a list of lists. + recursive_seq_lens = [[1]] base_shape = [1] # The range of random integers is [low, high] first_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) second_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) third_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) fourth_word = fluid.create_random_int_lodtensor( - lod, base_shape, place, low=0, high=dict_size - 1) + recursive_seq_lens, base_shape, place, low=0, high=dict_size - 1) assert feed_target_names[0] == 'firstw' assert feed_target_names[1] == 'secondw' @@ -200,7 +204,7 @@ def infer(use_cuda, save_dirname=None): }, fetch_list=fetch_targets, return_numpy=False) - print(results[0].lod()) + print(results[0].recursive_sequence_lengths()) np_data = np.array(results[0]) print("Inference Shape: ", np_data.shape) @@ -243,7 +247,7 @@ def inject_test_method(use_cuda, is_sparse, is_parallel): is_sparse=is_sparse, is_parallel=is_parallel) - if use_cuda and is_sparse: + if (not fluid.core.is_compiled_with_cuda() or use_cuda) and is_sparse: fn = __impl__ else: # skip the other test when on CI server diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py index 8818cf96fa8f08036f9e23aae786f67b5614b2b9..ccc62b442f62fa9fa175de031b0732febe38ee9a 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py @@ -12,12 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. -import numpy as np -import paddle -import paddle.fluid as fluid import math import sys +import paddle +import paddle.fluid as fluid +from paddle.fluid.layers.device import get_places + # need to fix random seed and training data to compare the loss # value accurately calculated by the default and the memory optimization # version. @@ -34,7 +35,7 @@ if fluid.core.is_compiled_with_cuda(): use_nccl = False place = fluid.CUDAPlace(0) -places = fluid.layers.get_places(device_count=0, device_type=device_type) +places = get_places(device_count=0, device_type=device_type) pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl) with pd.do(): x_ = pd.read_input(x) @@ -56,7 +57,7 @@ BATCH_SIZE = 200 # fix the order of training data train_reader = paddle.batch( - paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE) + paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE, drop_last=False) # train_reader = paddle.batch( # paddle.reader.shuffle( @@ -77,7 +78,7 @@ for pass_id in range(PASS_NUM): if avg_loss_value[0] < 10.0: exit(0) # if avg cost less than 10.0, we think our code is good. - print avg_loss_value[0] + print(avg_loss_value[0]) if math.isnan(float(avg_loss_value)): sys.exit("got NaN loss, training failed.") exit(1) diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py index dfebb9a06ea4f290f128c486dcaccaeccdcef8c4..b2a59d27da9b3348b581d51a68d769bbf3b90d35 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py @@ -12,8 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function - import sys import paddle diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py index a1ca6d981fafb401985d03e9f2d63d1cb41b21b5..323ddfb6911fdd57b32344933373189370005126 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py @@ -80,21 +80,6 @@ def encoder_decoder(): return rnn() -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - res = core.LoDTensor() - res.set(flattened_data, place) - res.set_lod([lod]) - return res - - def main(): rnn_out = encoder_decoder() label = layers.data( @@ -122,18 +107,21 @@ def main(): exe.run(framework.default_startup_program()) + feed_order = [ + 'src_word_id', 'target_language_word', 'target_language_next_word' + ] + + feed_list = [ + fluid.default_main_program().global_block().var(var_name) + for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + batch_id = 0 - for pass_id in xrange(10): + for pass_id in range(10): for data in train_data(): - word_data = to_lodtensor(map(lambda x: x[0], data), place) - trg_word = to_lodtensor(map(lambda x: x[1], data), place) - trg_word_next = to_lodtensor(map(lambda x: x[2], data), place) outs = exe.run(fluid.default_main_program(), - feed={ - 'src_word_id': word_data, - 'target_language_word': trg_word, - 'target_language_next_word': trg_word_next - }, + feed=feeder.feed(data), fetch_list=[avg_cost]) avg_cost_val = np.array(outs[0]) print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) + diff --git a/python/paddle/fluid/tests/demo/fc_gan.py b/python/paddle/fluid/tests/demo/fc_gan.py index 8ea1b2b15cc0c0eb5bca67a9c5a6ac6c6774e7e2..3d92f50f0adeca79adefc291cdfba6a012fc2118 100644 --- a/python/paddle/fluid/tests/demo/fc_gan.py +++ b/python/paddle/fluid/tests/demo/fc_gan.py @@ -137,7 +137,7 @@ def main(): generated_img = exe.run(g_program, feed={'noise': n}, fetch_list={g_img})[0] - real_data = numpy.array(map(lambda x: x[0], data)).astype('float32') + real_data = numpy.array([x[0] for x in data]).astype('float32') real_data = real_data.reshape(num_true, 784) total_data = numpy.concatenate([real_data, generated_img]) total_label = numpy.concatenate([ @@ -150,7 +150,7 @@ def main(): feed={'img': total_data, 'label': total_label}, fetch_list={d_loss})[0] - for _ in xrange(NUM_TRAIN_TIMES_OF_DG): + for _ in range(NUM_TRAIN_TIMES_OF_DG): n = numpy.random.uniform( low=-1.0, high=1.0, size=[2 * num_true * NOISE_SIZE]).astype('float32').reshape( diff --git a/python/paddle/fluid/tests/demo/text_classification/.gitignore b/python/paddle/fluid/tests/demo/file_reader/.gitignore similarity index 100% rename from python/paddle/fluid/tests/demo/text_classification/.gitignore rename to python/paddle/fluid/tests/demo/file_reader/.gitignore diff --git a/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py b/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py new file mode 100644 index 0000000000000000000000000000000000000000..a00325d79be2eba4d7f770b5316c5857952fe272 --- /dev/null +++ b/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py @@ -0,0 +1,62 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import sys +import paddle.fluid as fluid +import paddle.v2 as paddle + + +def load_vocab(filename): + """ + load vocabulary + """ + vocab = {} + with open(filename) as f: + wid = 0 + for line in f: + vocab[line.strip()] = wid + wid += 1 + return vocab + + +# load word dict with paddle inner function +if len(sys.argv) == 1: + word_dict = paddle.dataset.imdb.word_dict() +else: + word_dict = load_vocab(sys.argv[1]) + word_dict[""] = len(word_dict) +print("Dict dim = ", len(word_dict)) + +# input text data +data = fluid.layers.data(name="words", shape=[1], dtype="int64", lod_level=1) + +# label data +label = fluid.layers.data(name="label", shape=[1], dtype="int64") +# like placeholder +feeder = fluid.DataFeeder(feed_list=[data, label], place=fluid.CPUPlace()) + +# train data set +BATCH_SIZE = 128 +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=25000), + batch_size=BATCH_SIZE) + +test_reader = paddle.batch( + paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + +fluid.recordio_writer.convert_reader_to_recordio_file( + "train.recordio", feeder=feeder, reader_creator=train_reader) +fluid.recordio_writer.convert_reader_to_recordio_file( + "test.recordio", feeder=feeder, reader_creator=test_reader) diff --git a/python/paddle/fluid/tests/demo/file_reader/train.py b/python/paddle/fluid/tests/demo/file_reader/train.py new file mode 100644 index 0000000000000000000000000000000000000000..bc3a6dc81d24afec66ed1489aead1cff79a59bca --- /dev/null +++ b/python/paddle/fluid/tests/demo/file_reader/train.py @@ -0,0 +1,138 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import numpy +import sys + +TRAIN_FILES = ['train.recordio'] +TEST_FILES = ['test.recordio'] + +DICT_DIM = 5147 + +# embedding dim +emb_dim = 128 + +# hidden dim +hid_dim = 128 + +# class num +class_dim = 2 + +# epoch num +epoch_num = 10 + + +def build_program(is_train): + file_obj_handle = fluid.layers.io.open_files( + filenames=TRAIN_FILES if is_train else TEST_FILES, + shapes=[[-1, 1], [-1, 1]], + lod_levels=[1, 0], + dtypes=['int64', 'int64']) + + file_obj = fluid.layers.io.double_buffer(file_obj_handle) + + with fluid.unique_name.guard(): + + data, label = fluid.layers.read_file(file_obj) + + emb = fluid.layers.embedding(input=data, size=[DICT_DIM, emb_dim]) + + conv_3 = fluid.nets.sequence_conv_pool( + input=emb, + num_filters=hid_dim, + filter_size=3, + act="tanh", + pool_type="sqrt") + + conv_4 = fluid.nets.sequence_conv_pool( + input=emb, + num_filters=hid_dim, + filter_size=4, + act="tanh", + pool_type="sqrt") + + prediction = fluid.layers.fc(input=[conv_3, conv_4], + size=class_dim, + act="softmax") + + # cross entropy loss + cost = fluid.layers.cross_entropy(input=prediction, label=label) + + # mean loss + avg_cost = fluid.layers.mean(x=cost) + acc = fluid.layers.accuracy(input=prediction, label=label) + + if is_train: + # SGD optimizer + sgd_optimizer = fluid.optimizer.Adagrad(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + return {'loss': avg_cost, 'log': [avg_cost, acc], 'file': file_obj_handle} + + +def main(): + train = fluid.Program() + startup = fluid.Program() + test = fluid.Program() + + with fluid.program_guard(train, startup): + train_args = build_program(is_train=True) + + with fluid.program_guard(test, startup): + test_args = build_program(is_train=False) + + use_cuda = fluid.core.is_compiled_with_cuda() + # startup + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place=place) + exe.run(startup) + + train_exe = fluid.ParallelExecutor( + use_cuda=use_cuda, + loss_name=train_args['loss'].name, + main_program=train) + test_exe = fluid.ParallelExecutor( + use_cuda=use_cuda, main_program=test, share_vars_from=train_exe) + + fetch_var_list = [var.name for var in train_args['log']] + for epoch_id in range(epoch_num): + # train + try: + batch_id = 0 + while True: + loss, acc = map(numpy.array, + train_exe.run(fetch_list=fetch_var_list)) + print 'Train epoch', epoch_id, 'batch', batch_id, 'loss:', loss, 'acc:', acc + batch_id += 1 + except fluid.core.EOFException: + print 'End of epoch', epoch_id + train_args['file'].reset() + + # test + loss = [] + acc = [] + try: + while True: + loss_np, acc_np = map(numpy.array, + test_exe.run(fetch_list=fetch_var_list)) + loss.append(loss_np[0]) + acc.append(acc_np[0]) + except: + test_args['file'].reset() + print 'Test loss:', numpy.mean(loss), 'acc:', numpy.mean(acc) + + +if __name__ == '__main__': + main() diff --git a/python/paddle/fluid/tests/demo/pyreader.py b/python/paddle/fluid/tests/demo/pyreader.py new file mode 100644 index 0000000000000000000000000000000000000000..82065401935036ca346fa395c033f0f57100f01b --- /dev/null +++ b/python/paddle/fluid/tests/demo/pyreader.py @@ -0,0 +1,98 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy + +import paddle +import paddle.dataset.mnist as mnist +import paddle.fluid as fluid +import paddle.v2 + + +def network(is_train): + reader = fluid.layers.py_reader( + capacity=10, + shapes=((-1, 784), (-1, 1)), + dtypes=('float32', 'int64'), + name="train_reader" if is_train else "test_reader", + use_double_buffer=True) + img, label = fluid.layers.read_file(reader) + + hidden = img + + for i in xrange(2): + hidden = fluid.layers.fc(input=hidden, size=100, act='tanh') + hidden = fluid.layers.dropout( + hidden, dropout_prob=0.5, is_test=not is_train) + + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + return fluid.layers.mean(loss), reader + + +def main(): + train_prog = fluid.Program() + startup_prog = fluid.Program() + + with fluid.program_guard(train_prog, startup_prog): + with fluid.unique_name.guard(): + loss, train_reader = network(True) + adam = fluid.optimizer.Adam(learning_rate=0.01) + adam.minimize(loss) + + test_prog = fluid.Program() + test_startup = fluid.Program() + with fluid.program_guard(test_prog, test_startup): + with fluid.unique_name.guard(): + test_loss, test_reader = network(False) + + use_cuda = fluid.core.is_compiled_with_cuda() + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + fluid.Executor(place).run(startup_prog) + fluid.Executor(place).run(test_startup) + + trainer = fluid.ParallelExecutor( + use_cuda=use_cuda, loss_name=loss.name, main_program=train_prog) + + tester = fluid.ParallelExecutor( + use_cuda=use_cuda, share_vars_from=trainer, main_program=test_prog) + + train_reader.decorate_paddle_reader( + paddle.v2.reader.shuffle( + paddle.batch(mnist.train(), 512), buf_size=8192)) + + test_reader.decorate_paddle_reader(paddle.batch(mnist.test(), 512)) + + for epoch_id in xrange(10): + train_reader.start() + try: + while True: + print 'train_loss', numpy.array( + trainer.run(fetch_list=[loss.name])) + except fluid.core.EOFException: + print 'End of epoch', epoch_id + train_reader.reset() + + test_reader.start() + try: + while True: + print 'test loss', numpy.array( + tester.run(fetch_list=[test_loss.name])) + except fluid.core.EOFException: + print 'End of testing' + test_reader.reset() + + +if __name__ == '__main__': + main() diff --git a/python/paddle/fluid/tests/demo/text_classification/convert_data_to_recordio.py b/python/paddle/fluid/tests/demo/text_classification/convert_data_to_recordio.py deleted file mode 100644 index 9425d472a48056e71da5da364f659971ef6c2520..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/demo/text_classification/convert_data_to_recordio.py +++ /dev/null @@ -1,59 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import sys -import paddle.fluid as fluid -import paddle.v2 as paddle - - -def load_vocab(filename): - """ - load vocabulary - """ - vocab = {} - with open(filename) as f: - wid = 0 - for line in f: - vocab[line.strip()] = wid - wid += 1 - return vocab - - -# load word dict with paddle inner function -word_dict = load_vocab(sys.argv[1]) -word_dict[""] = len(word_dict) -print "Dict dim = ", len(word_dict) - -# input text data -data = fluid.layers.data(name="words", shape=[1], dtype="int64", lod_level=1) - -# label data -label = fluid.layers.data(name="label", shape=[1], dtype="int64") -# like placeholder -feeder = fluid.DataFeeder(feed_list=[data, label], place=fluid.CPUPlace()) - -# train data set -BATCH_SIZE = 128 -train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.imdb.train(word_dict), buf_size=10000), - batch_size=BATCH_SIZE) - -test_reader = paddle.batch( - paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) - -fluid.recordio_writer.convert_reader_to_recordio_file( - "train.recordio", feeder=feeder, reader_creator=train_reader) -fluid.recordio_writer.convert_reader_to_recordio_file( - "test.recordio", feeder=feeder, reader_creator=test_reader) diff --git a/python/paddle/fluid/tests/demo/text_classification/train.py b/python/paddle/fluid/tests/demo/text_classification/train.py deleted file mode 100644 index e408684c6e0941a1b317ffeac66f071c1382836d..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/demo/text_classification/train.py +++ /dev/null @@ -1,148 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import paddle.fluid as fluid -import numpy -import sys - -TRAIN_FILES = ['train.recordio'] -TEST_FILES = ['test.recordio'] - -DICT_DIM = 89528 - -# embedding dim -emb_dim = 128 - -# hidden dim -hid_dim = 128 - -# hidden dim2 -hid_dim2 = 96 - -# class num -class_dim = 2 - - -def network_cfg(is_train, pass_num=100): - with fluid.unique_name.guard(): - train_file_obj = fluid.layers.open_files( - filenames=TRAIN_FILES, - pass_num=pass_num, - shapes=[[-1, 1], [-1, 1]], - lod_levels=[1, 0], - dtypes=['int64', 'int64'], - thread_num=1) - - test_file_obj = fluid.layers.open_files( - filenames=TEST_FILES, - pass_num=1, - shapes=[[-1, 1], [-1, 1]], - lod_levels=[1, 0], - dtypes=['int64', 'int64'], - thread_num=1) - - if is_train: - file_obj = fluid.layers.shuffle(train_file_obj, buffer_size=1000) - else: - file_obj = test_file_obj - - file_obj = fluid.layers.double_buffer( - file_obj, - name="train_double_buffer" if is_train else 'test_double_buffer') - - data, label = fluid.layers.read_file(file_obj) - - emb = fluid.layers.embedding(input=data, size=[DICT_DIM, emb_dim]) - - # sequence conv with window size = 3 - win_size = 3 - conv_3 = fluid.nets.sequence_conv_pool( - input=emb, - num_filters=hid_dim, - filter_size=win_size, - act="tanh", - pool_type="max") - - # fc layer after conv - fc_1 = fluid.layers.fc(input=[conv_3], size=hid_dim2) - - # probability of each class - prediction = fluid.layers.fc(input=[fc_1], - size=class_dim, - act="softmax") - # cross entropy loss - cost = fluid.layers.cross_entropy(input=prediction, label=label) - - # mean loss - avg_cost = fluid.layers.mean(x=cost) - acc = fluid.layers.accuracy(input=prediction, label=label) - - if is_train: - # SGD optimizer - sgd_optimizer = fluid.optimizer.Adagrad(learning_rate=0.01) - sgd_optimizer.minimize(avg_cost) - - return { - 'loss': avg_cost, - 'log': [avg_cost, acc], - 'file': train_file_obj if is_train else test_file_obj - } - - -def main(): - train = fluid.Program() - startup = fluid.Program() - - with fluid.program_guard(train, startup): - train_args = network_cfg(is_train=True) - - test = fluid.Program() - - with fluid.program_guard(test, fluid.Program()): - test_args = network_cfg(is_train=False) - - # startup - place = fluid.CUDAPlace(0) - exe = fluid.Executor(place=place) - exe.run(startup) - - train_exe = fluid.ParallelExecutor( - use_cuda=True, loss_name=train_args['loss'].name, main_program=train) - - fetch_var_list = [var.name for var in train_args['log']] - for i in xrange(sys.maxint): - result = map(numpy.array, - train_exe.run(fetch_list=fetch_var_list - if i % 1000 == 0 else [])) - if len(result) != 0: - print 'Train: ', result - - if i % 1000 == 0: - test_exe = fluid.ParallelExecutor( - use_cuda=True, main_program=test, share_vars_from=train_exe) - loss = [] - acc = [] - try: - while True: - loss_np, acc_np = map( - numpy.array, test_exe.run(fetch_list=fetch_var_list)) - loss.append(loss_np[0]) - acc.append(acc_np[0]) - except: - test_args['file'].reset() - print 'TEST: ', numpy.mean(loss), numpy.mean(acc) - - -if __name__ == '__main__': - main() diff --git a/python/paddle/fluid/tests/no_test_concurrency.py b/python/paddle/fluid/tests/no_test_concurrency.py new file mode 100644 index 0000000000000000000000000000000000000000..3bc0c9808e2345b610dea79abc56cfb0065ea46f --- /dev/null +++ b/python/paddle/fluid/tests/no_test_concurrency.py @@ -0,0 +1,258 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid import framework, unique_name, layer_helper +from paddle.fluid.executor import Executor +from paddle.fluid.layers import fill_constant, assign, While, elementwise_add, Print + + +class TestRoutineOp(unittest.TestCase): + def test_simple_routine(self): + ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + + # Create LOD_TENSOR and put it into the scope. This placeholder + # variable will be filled in and returned by fluid.channel_recv + result = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT64) + + with fluid.Go(): + input_value = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.FP64, value=1234) + fluid.channel_send(ch, input_value) + + result, status = fluid.channel_recv(ch, result) + fluid.channel_close(ch) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + outs = exe.run(fetch_list=[result]) + self.assertEqual(outs[0], 1234) + + def test_daisy_chain(self): + ''' + Mimics classic Daisy-chain test: https://talks.golang.org/2012/concurrency.slide#39 + ''' + n = 100 + + leftmost = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + left = leftmost + + # TODO(thuan): Use fluid.While() after scope capture is implemented. + # https://github.com/PaddlePaddle/Paddle/issues/8502 + for i in range(n): + right = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + with fluid.Go(): + one_tensor = self._create_one_dim_tensor(1) + result = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT64) + + result, status = fluid.channel_recv(right, result) + one_added = fluid.layers.elementwise_add(x=one_tensor, y=result) + fluid.channel_send(left, one_added) + left = right + + # Trigger the channel propagation by sending a "1" to rightmost channel + with fluid.Go(): + one_tensor = self._create_one_dim_tensor(1) + fluid.channel_send(right, one_tensor) + + leftmost_result = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT64) + leftmost_result, status = fluid.channel_recv(leftmost, leftmost_result) + + cpu = core.CPUPlace() + exe = Executor(cpu) + leftmost_data = exe.run(fetch_list=[leftmost_result]) + + # The leftmost_data should be equal to the number of channels + 1 + self.assertEqual(leftmost_data[0][0], n + 1) + + def _create_one_dim_tensor(self, value): + one_dim_tensor = fill_constant(shape=[1], dtype='int', value=value) + one_dim_tensor.stop_gradient = True + return one_dim_tensor + + def _create_tensor(self, name, type, dtype): + return framework.default_main_program().current_block().create_var( + name=unique_name.generate(name), type=type, dtype=dtype) + + def _create_persistable_tensor(self, name, type, dtype): + return framework.default_main_program().current_block().create_var( + name=unique_name.generate(name), + type=type, + dtype=dtype, + persistable=True) + + def test_select(self): + with framework.program_guard(framework.Program()): + ch1 = fluid.make_channel( + dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) + + result1 = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + input_value = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.FP64, value=10) + + with fluid.Select() as select: + with select.case(fluid.channel_send, ch1, input_value): + # Execute something. + pass + + with select.default(): + pass + + # This should not block because we are using a buffered channel. + result1, status = fluid.channel_recv(ch1, result1) + fluid.channel_close(ch1) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + result = exe.run(fetch_list=[result1]) + self.assertEqual(result[0][0], 10) + + def test_fibonacci(self): + """ + Mimics Fibonacci Go example: https://tour.golang.org/concurrency/5 + """ + with framework.program_guard(framework.Program()): + quit_ch_input_var = self._create_persistable_tensor( + 'quit_ch_input', core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT32) + quit_ch_input = fill_constant( + shape=[1], + dtype=core.VarDesc.VarType.INT32, + value=0, + out=quit_ch_input_var) + + result = self._create_persistable_tensor( + 'result', core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT32) + fill_constant( + shape=[1], + dtype=core.VarDesc.VarType.INT32, + value=0, + out=result) + + x = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + y = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=1) + + while_cond = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.BOOL, value=True) + + while_false = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.BOOL, value=False) + + x_tmp = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + + def fibonacci(channel, quit_channel): + while_op = While(cond=while_cond) + with while_op.block(): + result2 = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + + with fluid.Select() as select: + with select.case( + fluid.channel_send, channel, x, is_copy=True): + assign(input=x, output=x_tmp) + assign(input=y, output=x) + assign(elementwise_add(x=x_tmp, y=y), output=y) + + with select.case(fluid.channel_recv, quit_channel, + result2): + # Quit + helper = layer_helper.LayerHelper('assign') + helper.append_op( + type='assign', + inputs={'X': [while_false]}, + outputs={'Out': [while_cond]}) + + ch1 = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + quit_ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + + with fluid.Go(): + for i in range(10): + fluid.channel_recv(ch1, result) + Print(result) + + fluid.channel_send(quit_ch, quit_ch_input) + + fibonacci(ch1, quit_ch) + + fluid.channel_close(ch1) + fluid.channel_close(quit_ch) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + exe_result = exe.run(fetch_list=[result]) + self.assertEqual(exe_result[0][0], 34) + + def test_ping_pong(self): + """ + Mimics Ping Pong example: https://gobyexample.com/channel-directions + """ + with framework.program_guard(framework.Program()): + result = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + ping_result = self._create_tensor('ping_return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + def ping(ch, message): + fluid.channel_send(ch, message, is_copy=True) + + def pong(ch1, ch2): + fluid.channel_recv(ch1, ping_result) + fluid.channel_send(ch2, ping_result, is_copy=True) + + pings = fluid.make_channel( + dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) + pongs = fluid.make_channel( + dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) + + msg = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.FP64, value=9) + + ping(pings, msg) + pong(pings, pongs) + + fluid.channel_recv(pongs, result) + + fluid.channel_close(pings) + fluid.channel_close(pongs) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + exe_result = exe.run(fetch_list=[result]) + self.assertEqual(exe_result[0][0], 9) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/test_beam_search_decoder.py b/python/paddle/fluid/tests/test_beam_search_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..8bf750940d570cdad5e110168afc5f632202e869 --- /dev/null +++ b/python/paddle/fluid/tests/test_beam_search_decoder.py @@ -0,0 +1,265 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +A simple machine translation demo using beam search decoder. +""" + +import contextlib +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.framework as framework +import paddle.fluid.layers as layers +from paddle.fluid.executor import Executor +from paddle.fluid.contrib.decoder.beam_search_decoder import * +import unittest +import os + +dict_size = 30000 +source_dict_dim = target_dict_dim = dict_size +src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size) +hidden_dim = 32 +word_dim = 32 +decoder_size = hidden_dim +IS_SPARSE = True +batch_size = 2 +max_length = 8 +topk_size = 50 +trg_dic_size = 10000 +beam_size = 2 + + +def encoder(): + # encoder + src_word = layers.data( + name="src_word", shape=[1], dtype='int64', lod_level=1) + src_embedding = layers.embedding( + input=src_word, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=IS_SPARSE) + + fc1 = layers.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') + lstm_hidden0, lstm_0 = layers.dynamic_lstm(input=fc1, size=hidden_dim * 4) + encoder_out = layers.sequence_last_step(input=lstm_hidden0) + return encoder_out + + +def decoder_state_cell(context): + h = InitState(init=context, need_reorder=True) + state_cell = StateCell(inputs={'x': None}, states={'h': h}, out_state='h') + + @state_cell.state_updater + def updater(state_cell): + current_word = state_cell.get_input('x') + prev_h = state_cell.get_state('h') + # make sure lod of h heritted from prev_h + h = layers.fc(input=[prev_h, current_word], + size=decoder_size, + act='tanh') + state_cell.set_state('h', h) + + return state_cell + + +def decoder_train(state_cell): + # decoder + trg_language_word = layers.data( + name="target_word", shape=[1], dtype='int64', lod_level=1) + trg_embedding = layers.embedding( + input=trg_language_word, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=IS_SPARSE) + + decoder = TrainingDecoder(state_cell) + + with decoder.block(): + current_word = decoder.step_input(trg_embedding) + decoder.state_cell.compute_state(inputs={'x': current_word}) + current_score = layers.fc(input=decoder.state_cell.get_state('h'), + size=target_dict_dim, + act='softmax') + decoder.state_cell.update_states() + decoder.output(current_score) + + return decoder() + + +def decoder_decode(state_cell): + init_ids = layers.data( + name="init_ids", shape=[1], dtype="int64", lod_level=2) + init_scores = layers.data( + name="init_scores", shape=[1], dtype="float32", lod_level=2) + + decoder = BeamSearchDecoder( + state_cell=state_cell, + init_ids=init_ids, + init_scores=init_scores, + target_dict_dim=target_dict_dim, + word_dim=word_dim, + input_var_dict={}, + topk_size=topk_size, + sparse_emb=IS_SPARSE, + max_len=max_length, + beam_size=beam_size, + end_id=1, + name=None) + decoder.decode() + translation_ids, translation_scores = decoder() + + return translation_ids, translation_scores + + +def train_main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + context = encoder() + state_cell = decoder_state_cell(context) + rnn_out = decoder_train(state_cell) + label = layers.data( + name="target_next_word", shape=[1], dtype='int64', lod_level=1) + cost = layers.cross_entropy(input=rnn_out, label=label) + avg_cost = layers.mean(x=cost) + + optimizer = fluid.optimizer.Adagrad(learning_rate=1e-3) + optimizer.minimize(avg_cost) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.wmt14.train(dict_size), buf_size=1000), + batch_size=batch_size) + feed_order = ['src_word', 'target_word', 'target_next_word'] + + exe = Executor(place) + + def train_loop(main_program): + exe.run(framework.default_startup_program()) + + feed_list = [ + main_program.global_block().var(var_name) for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + + for pass_id in range(1): + for batch_id, data in enumerate(train_reader()): + outs = exe.run(main_program, + feed=feeder.feed(data), + fetch_list=[avg_cost]) + avg_cost_val = np.array(outs[0]) + print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) + + " avg_cost=" + str(avg_cost_val)) + if batch_id > 3: + break + + train_loop(framework.default_main_program()) + + +def decode_main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + context = encoder() + state_cell = decoder_state_cell(context) + translation_ids, translation_scores = decoder_decode(state_cell) + + exe = Executor(place) + exe.run(framework.default_startup_program()) + + init_ids_data = np.array([0 for _ in range(batch_size)], dtype='int64') + init_scores_data = np.array( + [1. for _ in range(batch_size)], dtype='float32') + init_ids_data = init_ids_data.reshape((batch_size, 1)) + init_scores_data = init_scores_data.reshape((batch_size, 1)) + init_lod = [1] * batch_size + init_lod = [init_lod, init_lod] + + init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) + init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.wmt14.train(dict_size), buf_size=1000), + batch_size=batch_size) + + feed_order = ['src_word'] + feed_list = [ + framework.default_main_program().global_block().var(var_name) + for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + + data = next(train_reader()) + feed_dict = feeder.feed([[x[0]] for x in data]) + feed_dict['init_ids'] = init_ids + feed_dict['init_scores'] = init_scores + + result_ids, result_scores = exe.run( + framework.default_main_program(), + feed=feed_dict, + fetch_list=[translation_ids, translation_scores], + return_numpy=False) + print(result_ids.lod()) + + +class TestBeamSearchDecoder(unittest.TestCase): + pass + + +@contextlib.contextmanager +def scope_prog_guard(): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +def inject_test_train(use_cuda): + f_name = 'test_{0}_train'.format('cuda' if use_cuda else 'cpu') + + def f(*args): + with scope_prog_guard(): + train_main(use_cuda) + + setattr(TestBeamSearchDecoder, f_name, f) + + +def inject_test_decode(use_cuda, decorator=None): + f_name = 'test_{0}_decode'.format('cuda' if use_cuda else 'cpu', 'sparse') + + def f(*args): + with scope_prog_guard(): + decode_main(use_cuda) + + if decorator is not None: + f = decorator(f) + + setattr(TestBeamSearchDecoder, f_name, f) + + +for _use_cuda_ in (False, True): + inject_test_train(_use_cuda_) + +for _use_cuda_ in (False, True): + _decorator_ = None + inject_test_decode(use_cuda=_use_cuda_, decorator=_decorator_) + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/test_concurrency.py b/python/paddle/fluid/tests/test_concurrency.py deleted file mode 100644 index e8f6cfb4a907b2c01e9662e7e9bf2cb0fbd6cb1b..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/test_concurrency.py +++ /dev/null @@ -1,258 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import unittest -import paddle.fluid as fluid -import paddle.fluid.core as core -from paddle.fluid import framework, unique_name, layer_helper -from paddle.fluid.executor import Executor -from paddle.fluid.layers import fill_constant, assign, While, elementwise_add, Print - - -class TestRoutineOp(unittest.TestCase): - def test_simple_routine(self): - ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) - - # Create LOD_TENSOR and put it into the scope. This placeholder - # variable will be filled in and returned by fluid.channel_recv - result = self._create_tensor('return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.INT64) - - with fluid.Go(): - input_value = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.FP64, value=1234) - fluid.channel_send(ch, input_value) - - result, status = fluid.channel_recv(ch, result) - fluid.channel_close(ch) - - cpu = core.CPUPlace() - exe = Executor(cpu) - - outs = exe.run(fetch_list=[result]) - self.assertEqual(outs[0], 1234) - - def test_daisy_chain(self): - ''' - Mimics classic Daisy-chain test: https://talks.golang.org/2012/concurrency.slide#39 - ''' - n = 100 - - leftmost = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) - left = leftmost - - # TODO(thuan): Use fluid.While() after scope capture is implemented. - # https://github.com/PaddlePaddle/Paddle/issues/8502 - for i in range(n): - right = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) - with fluid.Go(): - one_tensor = self._create_one_dim_tensor(1) - result = self._create_tensor('return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.INT64) - - result, status = fluid.channel_recv(right, result) - one_added = fluid.layers.elementwise_add(x=one_tensor, y=result) - fluid.channel_send(left, one_added) - left = right - - # Trigger the channel propagation by sending a "1" to rightmost channel - with fluid.Go(): - one_tensor = self._create_one_dim_tensor(1) - fluid.channel_send(right, one_tensor) - - leftmost_result = self._create_tensor('return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.INT64) - leftmost_result, status = fluid.channel_recv(leftmost, leftmost_result) - - cpu = core.CPUPlace() - exe = Executor(cpu) - leftmost_data = exe.run(fetch_list=[leftmost_result]) - - # The leftmost_data should be equal to the number of channels + 1 - self.assertEqual(leftmost_data[0][0], n + 1) - - def _create_one_dim_tensor(self, value): - one_dim_tensor = fill_constant(shape=[1], dtype='int', value=value) - one_dim_tensor.stop_gradient = True - return one_dim_tensor - - def _create_tensor(self, name, type, dtype): - return framework.default_main_program().current_block().create_var( - name=unique_name.generate(name), type=type, dtype=dtype) - - def _create_persistable_tensor(self, name, type, dtype): - return framework.default_main_program().current_block().create_var( - name=unique_name.generate(name), - type=type, - dtype=dtype, - persistable=True) - - def test_select(self): - with framework.program_guard(framework.Program()): - ch1 = fluid.make_channel( - dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) - - result1 = self._create_tensor('return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.FP64) - - input_value = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.FP64, value=10) - - with fluid.Select() as select: - with select.case(fluid.channel_send, ch1, input_value): - # Execute something. - pass - - with select.default(): - pass - - # This should not block because we are using a buffered channel. - result1, status = fluid.channel_recv(ch1, result1) - fluid.channel_close(ch1) - - cpu = core.CPUPlace() - exe = Executor(cpu) - - result = exe.run(fetch_list=[result1]) - self.assertEqual(result[0][0], 10) - - def test_fibonacci(self): - """ - Mimics Fibonacci Go example: https://tour.golang.org/concurrency/5 - """ - with framework.program_guard(framework.Program()): - quit_ch_input_var = self._create_persistable_tensor( - 'quit_ch_input', core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.INT32) - quit_ch_input = fill_constant( - shape=[1], - dtype=core.VarDesc.VarType.INT32, - value=0, - out=quit_ch_input_var) - - result = self._create_persistable_tensor( - 'result', core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.INT32) - fill_constant( - shape=[1], - dtype=core.VarDesc.VarType.INT32, - value=0, - out=result) - - x = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) - y = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=1) - - while_cond = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.BOOL, value=True) - - while_false = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.BOOL, value=False) - - x_tmp = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) - - def fibonacci(channel, quit_channel): - while_op = While(cond=while_cond) - with while_op.block(): - result2 = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) - - with fluid.Select() as select: - with select.case( - fluid.channel_send, channel, x, is_copy=True): - assign(input=x, output=x_tmp) - assign(input=y, output=x) - assign(elementwise_add(x=x_tmp, y=y), output=y) - - with select.case(fluid.channel_recv, quit_channel, - result2): - # Quit - helper = layer_helper.LayerHelper('assign') - helper.append_op( - type='assign', - inputs={'X': [while_false]}, - outputs={'Out': [while_cond]}) - - ch1 = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) - quit_ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) - - with fluid.Go(): - for i in xrange(10): - fluid.channel_recv(ch1, result) - Print(result) - - fluid.channel_send(quit_ch, quit_ch_input) - - fibonacci(ch1, quit_ch) - - fluid.channel_close(ch1) - fluid.channel_close(quit_ch) - - cpu = core.CPUPlace() - exe = Executor(cpu) - - exe_result = exe.run(fetch_list=[result]) - self.assertEqual(exe_result[0][0], 34) - - def test_ping_pong(self): - """ - Mimics Ping Pong example: https://gobyexample.com/channel-directions - """ - with framework.program_guard(framework.Program()): - result = self._create_tensor('return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.FP64) - - ping_result = self._create_tensor('ping_return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.FP64) - - def ping(ch, message): - fluid.channel_send(ch, message, is_copy=True) - - def pong(ch1, ch2): - fluid.channel_recv(ch1, ping_result) - fluid.channel_send(ch2, ping_result, is_copy=True) - - pings = fluid.make_channel( - dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) - pongs = fluid.make_channel( - dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) - - msg = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.FP64, value=9) - - ping(pings, msg) - pong(pings, pongs) - - fluid.channel_recv(pongs, result) - - fluid.channel_close(pings) - fluid.channel_close(pongs) - - cpu = core.CPUPlace() - exe = Executor(cpu) - - exe_result = exe.run(fetch_list=[result]) - self.assertEqual(exe_result[0][0], 9) - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/fluid/tests/test_data_feeder.py b/python/paddle/fluid/tests/test_data_feeder.py index ce3ba3ebc50d7b015f379b5e80b179463a7b231a..30b7a634a2b978df85d6432854ef12285460be44 100644 --- a/python/paddle/fluid/tests/test_data_feeder.py +++ b/python/paddle/fluid/tests/test_data_feeder.py @@ -22,12 +22,11 @@ class TestDataFeeder(unittest.TestCase): label = fluid.layers.data(name='label', shape=[1], dtype='int64') feeder = fluid.DataFeeder([img, label], fluid.CPUPlace()) result = feeder.feed([([0] * 784, [9]), ([1] * 784, [1])]) - print(result) self.assertEqual(result['image'].shape(), [2, 1, 28, 28]) self.assertEqual(result['label'].shape(), [2, 1]) - self.assertEqual(result['image'].lod(), []) - self.assertEqual(result['label'].lod(), []) + self.assertEqual(result['image'].recursive_sequence_lengths(), []) + self.assertEqual(result['label'].recursive_sequence_lengths(), []) def test_lod_level_1_converter(self): # lod_level = 1 @@ -42,12 +41,12 @@ class TestDataFeeder(unittest.TestCase): # label = [1] * len(data) result = feeder.feed( [([1, 2, 3], [1]), ([4, 5], [1]), ([6, 7, 8, 9], [1])]) - print(result) self.assertEqual(result['sentences'].shape(), [9, 1]) self.assertEqual(result['label'].shape(), [3, 1]) - self.assertEqual(result['sentences'].lod(), [[0, 3, 5, 9]]) - self.assertEqual(result['label'].lod(), []) + self.assertEqual(result['sentences'].recursive_sequence_lengths(), + [[3, 2, 4]]) + self.assertEqual(result['label'].recursive_sequence_lengths(), []) def test_lod_level_2_converter(self): # lod_level = 2 @@ -62,12 +61,12 @@ class TestDataFeeder(unittest.TestCase): # label = [1] * len(data) result = feeder.feed( [([[1, 2, 3], [4, 5]], [1]), ([[6, 7, 8, 9]], [1])]) - print(result) self.assertEqual(result['paragraphs'].shape(), [9, 1]) self.assertEqual(result['label'].shape(), [2, 1]) - self.assertEqual(result['paragraphs'].lod(), [[0, 2, 3], [0, 3, 5, 9]]) - self.assertEqual(result['label'].lod(), []) + self.assertEqual(result['paragraphs'].recursive_sequence_lengths(), + [[2, 1], [3, 2, 4]]) + self.assertEqual(result['label'].recursive_sequence_lengths(), []) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py index 8569d838bdd414eb84c6c87674990a25a2fdcdf9..fd45abd0a77cb54a3ca8e60cf80a1efe9f9d2060 100644 --- a/python/paddle/fluid/tests/test_detection.py +++ b/python/paddle/fluid/tests/test_detection.py @@ -12,7 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function import paddle.fluid as fluid import paddle.fluid.layers as layers from paddle.fluid.framework import Program, program_guard @@ -127,6 +126,24 @@ class TestPriorBox(unittest.TestCase): assert box.shape[3] == 4 +class TestAnchorGenerator(unittest.TestCase): + def test_anchor_generator(self): + data_shape = [3, 224, 224] + images = fluid.layers.data( + name='pixel', shape=data_shape, dtype='float32') + conv1 = fluid.layers.conv2d(images, 3, 3, 2) + anchor, var = fluid.layers.anchor_generator( + input=conv1, + anchor_sizes=[64, 128, 256, 512], + aspect_ratios=[0.5, 1.0, 2.0], + variance=[0.1, 0.1, 0.2, 0.2], + stride=[16.0, 16.0], + offset=0.5) + assert len(anchor.shape) == 4 + assert anchor.shape == var.shape + assert anchor.shape[3] == 4 + + class TestMultiBoxHead(unittest.TestCase): def test_multi_box_head(self): data_shape = [3, 224, 224] diff --git a/python/paddle/fluid/tests/test_error_clip.py b/python/paddle/fluid/tests/test_error_clip.py index 89f4c64975802dc1827ec17ed3626b91e36d6971..e8edd7fbbb31b1a6ecbf2a25a7d39e7b3f66363a 100644 --- a/python/paddle/fluid/tests/test_error_clip.py +++ b/python/paddle/fluid/tests/test_error_clip.py @@ -12,7 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. -from __future__ import print_function import numpy as np import paddle import paddle.fluid as fluid @@ -36,7 +35,7 @@ with fluid.program_guard(main_program=prog): avg_cost = fluid.layers.mean(cost) prog_clip = prog.clone() -prog_clip.block(0).var(hidden1.name).set_error_clip( +prog_clip.block(0).var(hidden1.name)._set_error_clip( fluid.clip.ErrorClipByValue( max=CLIP_MAX, min=CLIP_MIN)) diff --git a/python/paddle/fluid/tests/test_if_else_op.py b/python/paddle/fluid/tests/test_if_else_op.py new file mode 100644 index 0000000000000000000000000000000000000000..082f64c146f65eee4be0757d07495c33764fa841 --- /dev/null +++ b/python/paddle/fluid/tests/test_if_else_op.py @@ -0,0 +1,211 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle +import paddle.fluid.layers as layers +from paddle.fluid.framework import Program, program_guard +from paddle.fluid.executor import Executor +from paddle.fluid.optimizer import MomentumOptimizer +import paddle.fluid.core as core +import paddle.fluid as fluid +from paddle.fluid.layers.control_flow import split_lod_tensor +from paddle.fluid.layers.control_flow import merge_lod_tensor +from paddle.fluid.layers.control_flow import ConditionalBlock + +import unittest +import numpy as np + + +class TestMNISTIfElseOp(unittest.TestCase): + def test_raw_api(self): + prog = Program() + startup_prog = Program() + with program_guard(prog, startup_prog): + image = layers.data(name='x', shape=[784], dtype='float32') + + label = layers.data(name='y', shape=[1], dtype='int64') + + limit = layers.fill_constant(shape=[1], dtype='int64', value=5) + cond = layers.less_than(x=label, y=limit) + true_image, false_image = split_lod_tensor(input=image, mask=cond) + + true_out = layers.create_tensor(dtype='float32') + true_cond = ConditionalBlock([cond]) + + with true_cond.block(): + hidden = layers.fc(input=true_image, size=100, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + layers.assign(input=prob, output=true_out) + + false_out = layers.create_tensor(dtype='float32') + false_cond = ConditionalBlock([cond]) + + with false_cond.block(): + hidden = layers.fc(input=false_image, size=200, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + layers.assign(input=prob, output=false_out) + + prob = merge_lod_tensor( + in_true=true_out, in_false=false_out, mask=cond, x=image) + loss = layers.cross_entropy(input=prob, label=label) + avg_loss = layers.mean(loss) + + optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) + optimizer.minimize(avg_loss, startup_prog) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=10) + + place = core.CPUPlace() + exe = Executor(place) + + exe.run(startup_prog) + PASS_NUM = 100 + for pass_id in range(PASS_NUM): + for data in train_reader(): + x_data = np.array([x[0] for x in data]).astype("float32") + y_data = np.array([x[1] for x in data]).astype("int64") + y_data = np.expand_dims(y_data, axis=1) + + outs = exe.run(prog, + feed={'x': x_data, + 'y': y_data}, + fetch_list=[avg_loss]) + print(outs[0]) + if outs[0] < 1.0: + return + self.assertFalse(True) + + def test_ifelse(self): + prog = Program() + startup_prog = Program() + with program_guard(prog, startup_prog): + image = layers.data(name='x', shape=[784], dtype='float32') + + label = layers.data(name='y', shape=[1], dtype='int64') + + limit = layers.fill_constant(shape=[1], dtype='int64', value=5) + cond = layers.less_than(x=label, y=limit) + ie = layers.IfElse(cond) + + with ie.true_block(): + true_image = ie.input(image) + hidden = layers.fc(input=true_image, size=100, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + + with ie.false_block(): + false_image = ie.input(image) + hidden = layers.fc(input=false_image, size=200, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + + prob = ie() + loss = layers.cross_entropy(input=prob[0], label=label) + avg_loss = layers.mean(loss) + + optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) + optimizer.minimize(avg_loss, startup_prog) + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=200) + + place = core.CPUPlace() + exe = Executor(place) + + exe.run(startup_prog) + PASS_NUM = 100 + for pass_id in range(PASS_NUM): + for data in train_reader(): + x_data = np.array([x[0] for x in data]).astype("float32") + y_data = np.array([x[1] for x in data]).astype("int64") + y_data = y_data.reshape((y_data.shape[0], 1)) + + outs = exe.run(prog, + feed={'x': x_data, + 'y': y_data}, + fetch_list=[avg_loss]) + print(outs[0]) + if outs[0] < 1.0: + return + self.assertFalse(True) + + +class TestIfElse(unittest.TestCase): + def set_test_case(self): + # condiction is: self.data < self.cond_value + self.cond_value = 0.5 + self.data = np.random.rand(25, 1).astype(np.float32) + + def compare_ifelse_op_and_numpy(self, place): + self.set_test_case() + + prog = Program() + startup_prog = Program() + with program_guard(prog, startup_prog): + src = layers.data(name='data', shape=[1], dtype='float32') + cond = layers.fill_constant( + [1], dtype='float32', value=self.cond_value) + ifcond = layers.less_than(x=src, y=cond) + ie = layers.IfElse(ifcond) + with ie.true_block(): + true_target = ie.input(src) + ie.output(true_target) + + with ie.false_block(): + false_target = ie.input(src) + ie.output(false_target) + if_out = ie() + out = layers.reduce_sum(if_out) + + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + fetch_list = [out] + o1, = exe.run(fluid.default_main_program(), + feed={'data': self.data}, + fetch_list=[out]) + o2 = np.sum(self.data) + self.assertTrue( + np.allclose( + o1, o2, atol=1e-8), + "IfElse result : " + str(o1) + "\n Numpy result :" + str(o2)) + + def test_cpu(self): + self.compare_ifelse_op_and_numpy(fluid.CPUPlace()) + + def test_cuda(self): + if not core.is_compiled_with_cuda(): + return + self.compare_ifelse_op_and_numpy(fluid.CUDAPlace(0)) + + +class TestIfElseTrueBranch(TestIfElse): + def set_test_case(self): + # condiction is: self.data < self.cond_value + self.cond_value = 10. + self.data = np.random.rand(25, 1).astype(np.float32) + + +class TestIfElseFalseBranch(TestIfElse): + def set_test_case(self): + # condiction is: self.data < self.cond_value + self.cond_value = -10. + self.data = np.random.rand(25, 1).astype(np.float32) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/test_lod_tensor.py b/python/paddle/fluid/tests/test_lod_tensor.py index 013d72f418cf7ac11eb31fd221052039e896e203..f7a9dd4129027417a06a6c25ff9a801fff259c5e 100644 --- a/python/paddle/fluid/tests/test_lod_tensor.py +++ b/python/paddle/fluid/tests/test_lod_tensor.py @@ -13,77 +13,85 @@ # limitations under the License. import paddle.fluid as fluid -from paddle.fluid.lod_tensor import create_lod_tensor, create_random_int_lodtensor, _validate_lod, _convert_lod -import numpy +from paddle.fluid.lod_tensor import create_lod_tensor, create_random_int_lodtensor +import numpy as np import unittest class TestLoDTensor(unittest.TestCase): - def test_validate_lod(self): - lod = (1, 2, 1) - self.assertRaises(AssertionError, _validate_lod, lod, -1) - lod = [[1, 2], (2, 3)] - self.assertRaises(AssertionError, _validate_lod, lod, -1) - lod = [1, 2, 3] - self.assertRaises(AssertionError, _validate_lod, lod, -1) + def test_pybind_recursive_seq_lens(self): + tensor = fluid.LoDTensor() + recursive_seq_lens = [] + tensor.set_recursive_sequence_lengths(recursive_seq_lens) + recursive_seq_lens = [[], [1], [3]] + self.assertRaises(Exception, tensor.set_recursive_sequence_lengths, + recursive_seq_lens) + recursive_seq_lens = [[0], [2], [3]] + self.assertRaises(Exception, tensor.set_recursive_sequence_lengths, + recursive_seq_lens) - lod = [] - self.assertTrue(_validate_lod(lod, -1)) - lod = [[], [1], [3]] - self.assertFalse(_validate_lod(lod, -1)) - lod = [[0], [-1], [3]] - self.assertFalse(_validate_lod(lod, -1)) + recursive_seq_lens = [[1, 2, 3]] + tensor.set_recursive_sequence_lengths(recursive_seq_lens) + self.assertEqual(tensor.recursive_sequence_lengths(), + recursive_seq_lens) + tensor.set(np.random.random([6, 1]), fluid.CPUPlace()) + self.assertTrue(tensor.has_valid_recursive_sequence_lengths()) + tensor.set(np.random.random([9, 1]), fluid.CPUPlace()) + self.assertFalse(tensor.has_valid_recursive_sequence_lengths()) # Each level's sum should be equal to the number of items in the next level # Moreover, last level's sum should be equal to the tensor height - lod = [[2, 3], [1, 3, 1, 2, 1]] - self.assertTrue(_validate_lod(lod, tensor_height=8)) - lod = [[1, 3], [2, 1, 3]] - self.assertFalse(_validate_lod(lod, tensor_height=6)) - lod = [[1, 3], [2, 1, 3, 4]] - self.assertFalse(_validate_lod(lod, tensor_height=5)) - - def test_convert_lod(self): - lod = [[1, 2, 3]] - converted_lod = [[0, 1, 3, 6]] - self.assertEqual(_convert_lod(lod), converted_lod) - - lod = [[2, 3], [1, 3, 1, 2, 1]] - converted_lod = [[0, 2, 5], [0, 1, 4, 5, 7, 8]] - self.assertEqual(_convert_lod(lod), converted_lod) + recursive_seq_lens = [[2, 3], [1, 3, 1, 2, 2]] + tensor.set_recursive_sequence_lengths(recursive_seq_lens) + self.assertEqual(tensor.recursive_sequence_lengths(), + recursive_seq_lens) + tensor.set(np.random.random([8, 1]), fluid.CPUPlace()) + self.assertFalse(tensor.has_valid_recursive_sequence_lengths()) + recursive_seq_lens = [[2, 3], [1, 3, 1, 2, 1]] + tensor.set_recursive_sequence_lengths(recursive_seq_lens) + self.assertTrue(tensor.has_valid_recursive_sequence_lengths()) + tensor.set(np.random.random([9, 1]), fluid.CPUPlace()) + self.assertFalse(tensor.has_valid_recursive_sequence_lengths()) def test_create_lod_tensor(self): # Create LoDTensor from a list data = [[1, 2, 3], [3, 4]] - wrong_lod = [[2, 2]] - correct_lod = [[3, 2]] - self.assertRaises(AssertionError, create_lod_tensor, data, wrong_lod, - fluid.CPUPlace()) - tensor = create_lod_tensor(data, correct_lod, fluid.CPUPlace()) - self.assertEqual(tensor.lod(), [[0, 3, 5]]) + wrong_recursive_seq_lens = [[2, 2]] + correct_recursive_seq_lens = [[3, 2]] + self.assertRaises(AssertionError, create_lod_tensor, data, + wrong_recursive_seq_lens, fluid.CPUPlace()) + tensor = create_lod_tensor(data, correct_recursive_seq_lens, + fluid.CPUPlace()) + self.assertEqual(tensor.recursive_sequence_lengths(), + correct_recursive_seq_lens) # Create LoDTensor from numpy array - data = numpy.random.random([10, 1]) - lod = [[2, 1], [3, 3, 4]] - tensor = create_lod_tensor(data, lod, fluid.CPUPlace()) - self.assertEqual(tensor.lod(), [[0, 2, 3], [0, 3, 6, 10]]) + data = np.random.random([10, 1]) + recursive_seq_lens = [[2, 1], [3, 3, 4]] + tensor = create_lod_tensor(data, recursive_seq_lens, fluid.CPUPlace()) + self.assertEqual(tensor.recursive_sequence_lengths(), + recursive_seq_lens) # Create LoDTensor from another LoDTensor, they are differnt instances - new_lod = [[2, 2, 1], [1, 2, 2, 3, 2]] - new_tensor = create_lod_tensor(tensor, new_lod, fluid.CPUPlace()) - self.assertEqual(tensor.lod(), [[0, 2, 3], [0, 3, 6, 10]]) - self.assertEqual(new_tensor.lod(), [[0, 2, 4, 5], [0, 1, 3, 5, 8, 10]]) + new_recursive_seq_lens = [[2, 2, 1], [1, 2, 2, 3, 2]] + new_tensor = create_lod_tensor(tensor, new_recursive_seq_lens, + fluid.CPUPlace()) + self.assertEqual(tensor.recursive_sequence_lengths(), + recursive_seq_lens) + self.assertEqual(new_tensor.recursive_sequence_lengths(), + new_recursive_seq_lens) def test_create_random_int_lodtensor(self): # The shape of a word, commonly used in speech and NLP problem, is [1] shape = [1] - lod = [[2, 3, 5]] + recursive_seq_lens = [[2, 3, 5]] dict_size = 10000 low = 0 high = dict_size - 1 - tensor = create_random_int_lodtensor(lod, shape, + tensor = create_random_int_lodtensor(recursive_seq_lens, shape, fluid.CPUPlace(), low, high) - self.assertEqual(tensor.lod(), [[0, 2, 5, 10]]) + self.assertEqual(tensor.recursive_sequence_lengths(), + recursive_seq_lens) self.assertEqual(tensor.shape(), [10, 1]) diff --git a/python/paddle/fluid/tests/test_mnist_if_else_op.py b/python/paddle/fluid/tests/test_mnist_if_else_op.py deleted file mode 100644 index d34f52db5ffc889f17513d034ad2c99f696b0cdf..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/test_mnist_if_else_op.py +++ /dev/null @@ -1,148 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import paddle -import paddle.fluid.layers as layers -from paddle.fluid.framework import Program, program_guard, default_main_program, default_startup_program -from paddle.fluid.executor import Executor -from paddle.fluid.optimizer import MomentumOptimizer -import paddle.fluid.core as core -import unittest -import numpy as np - - -class TestMNISTIfElseOp(unittest.TestCase): - def test_raw_api(self): - prog = Program() - startup_prog = Program() - with program_guard(prog, startup_prog): - image = layers.data(name='x', shape=[784], dtype='float32') - - label = layers.data(name='y', shape=[1], dtype='int64') - - limit = layers.fill_constant_batch_size_like( - input=label, dtype='int64', shape=[1], value=5.0) - cond = layers.less_than(x=label, y=limit) - true_image, false_image = layers.split_lod_tensor( - input=image, mask=cond) - - true_out = layers.create_tensor(dtype='float32') - true_cond = layers.ConditionalBlock([true_image]) - - with true_cond.block(): - hidden = layers.fc(input=true_image, size=100, act='tanh') - prob = layers.fc(input=hidden, size=10, act='softmax') - layers.assign(input=prob, output=true_out) - - false_out = layers.create_tensor(dtype='float32') - false_cond = layers.ConditionalBlock([false_image]) - - with false_cond.block(): - hidden = layers.fc(input=false_image, size=200, act='tanh') - prob = layers.fc(input=hidden, size=10, act='softmax') - layers.assign(input=prob, output=false_out) - - prob = layers.merge_lod_tensor( - in_true=true_out, in_false=false_out, mask=cond, x=image) - loss = layers.cross_entropy(input=prob, label=label) - avg_loss = layers.mean(loss) - - optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) - optimizer.minimize(avg_loss, startup_prog) - - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.mnist.train(), buf_size=8192), - batch_size=200) - - place = core.CPUPlace() - exe = Executor(place) - - exe.run(startup_prog) - PASS_NUM = 100 - for pass_id in range(PASS_NUM): - for data in train_reader(): - x_data = np.array(map(lambda x: x[0], data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = np.expand_dims(y_data, axis=1) - - outs = exe.run(prog, - feed={'x': x_data, - 'y': y_data}, - fetch_list=[avg_loss]) - print outs[0] - if outs[0] < 1.0: - return - self.assertFalse(True) - - def test_ifelse(self): - prog = Program() - startup_prog = Program() - with program_guard(prog, startup_prog): - image = layers.data(name='x', shape=[784], dtype='float32') - - label = layers.data(name='y', shape=[1], dtype='int64') - - limit = layers.fill_constant_batch_size_like( - input=label, dtype='int64', shape=[1], value=5.0) - cond = layers.less_than(x=label, y=limit) - ie = layers.IfElse(cond) - - with ie.true_block(): - true_image = ie.input(image) - hidden = layers.fc(input=true_image, size=100, act='tanh') - prob = layers.fc(input=hidden, size=10, act='softmax') - ie.output(prob) - - with ie.false_block(): - false_image = ie.input(image) - hidden = layers.fc(input=false_image, size=200, act='tanh') - prob = layers.fc(input=hidden, size=10, act='softmax') - ie.output(prob) - - prob = ie() - loss = layers.cross_entropy(input=prob[0], label=label) - avg_loss = layers.mean(loss) - - optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) - optimizer.minimize(avg_loss, startup_prog) - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.mnist.train(), buf_size=8192), - batch_size=200) - - place = core.CPUPlace() - exe = Executor(place) - - exe.run(kwargs['startup_program']) - PASS_NUM = 100 - for pass_id in range(PASS_NUM): - for data in train_reader(): - x_data = np.array(map(lambda x: x[0], data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape((y_data.shape[0], 1)) - - outs = exe.run(kwargs['main_program'], - feed={'x': x_data, - 'y': y_data}, - fetch_list=[avg_loss]) - print outs[0] - if outs[0] < 1.0: - return - self.assertFalse(True) - - -if __name__ == '__main__': - # temp disable if else unittest since it could be buggy. - exit(0) diff --git a/python/paddle/fluid/tests/unittests/.gitignore b/python/paddle/fluid/tests/unittests/.gitignore index 3538a9c2009bb133609153427981fb66974377fa..b1e8fda03aa42f5f7528eafb46c16d55b868bae5 100644 --- a/python/paddle/fluid/tests/unittests/.gitignore +++ b/python/paddle/fluid/tests/unittests/.gitignore @@ -4,3 +4,5 @@ mnist_1.recordio mnist_2.recordio flowers.recordio wmt16.recordio +data_balance_test.recordio +data_balance_with_lod_test.recordio diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index fead95ffdab25c7ea96b7ef223efc0abf7eea3e3..e7dd85ef5c3641be04261dc5d4166fa8452b4200 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -12,10 +12,15 @@ endif(NOT WITH_MKLDNN) if(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_recv_op) + list(REMOVE_ITEM TEST_OPS test_dist_transpiler) + list(REMOVE_ITEM TEST_OPS test_simple_dist_transpiler) + list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op) + LIST(REMOVE_ITEM TEST_OPS test_dist_mnist) + LIST(REMOVE_ITEM TEST_OPS test_dist_word2vec) endif(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290 -list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184 +list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184 list(REMOVE_ITEM TEST_OPS test_lstm_unit_op) # # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5185 list(REMOVE_ITEM TEST_OPS test_nce) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778 list(REMOVE_ITEM TEST_OPS test_recurrent_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152 @@ -35,7 +40,7 @@ function(py_test_modules TARGET_NAME) ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/tools/test_runner.py ${py_test_modules_MODULES} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) if (py_test_modules_SERIAL) - set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1) + set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1) endif() endif() endfunction() @@ -43,10 +48,23 @@ list(REMOVE_ITEM TEST_OPS test_warpctc_op) list(REMOVE_ITEM TEST_OPS test_dist_train) list(REMOVE_ITEM TEST_OPS test_parallel_executor_crf) list(REMOVE_ITEM TEST_OPS test_parallel_executor_fetch_feed) +list(REMOVE_ITEM TEST_OPS test_dist_se_resnext) +list(REMOVE_ITEM TEST_OPS test_dist_transformer) +list(REMOVE_ITEM TEST_OPS test_parallel_executor_transformer) +list(REMOVE_ITEM TEST_OPS test_image_classification_resnet) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR} SERIAL) -py_test_modules(test_dist_train MODULES test_dist_train SERIAL) -# tests that need to be done in fixed timeout -set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20) +if(WITH_DISTRIBUTE) + py_test_modules(test_dist_train MODULES test_dist_train SERIAL) + set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20) + set_tests_properties(test_dist_mnist PROPERTIES TIMEOUT 200) + set_tests_properties(test_dist_word2vec PROPERTIES TIMEOUT 200) +endif() +py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL) +py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) +py_test_modules(test_dist_transformer MODULES test_dist_transformer SERIAL) +py_test_modules(test_dist_se_resnext MODULES test_dist_se_resnext SERIAL) +py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL) +py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) diff --git a/python/paddle/fluid/tests/unittests/benchmark.py b/python/paddle/fluid/tests/unittests/benchmark.py new file mode 100644 index 0000000000000000000000000000000000000000..b98a92dcbe5626c6cca93b3f5894302399793bf9 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/benchmark.py @@ -0,0 +1,114 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import unittest +import time +import itertools +import six + +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.op import Operator +from op_test import OpTest + + +class BenchmarkSuite(OpTest): + def timeit_function(self, callback, iters, *args, **kwargs): + assert iters != 0, "Iters should >= 1" + start = time.time() + for i in range(iters): + callback(*args, **kwargs) + elapse = time.time() - start + return elapse / iters + + def _assert_cpu_gpu_same(self, cpu_outs, gpu_outs, fetch_list, atol): + for item_cpu_out, item_gpu_out, variable in zip(cpu_outs, gpu_outs, + fetch_list): + # the cpu version is baseline, expect gpu version keep same with cpu version. + expect = item_cpu_out + expect_t = np.array(item_cpu_out) + actual = item_gpu_out + actual_t = np.array(item_gpu_out) + var_name = variable if isinstance( + variable, six.string_types) else variable.name + self.assertTrue( + np.allclose( + actual_t, expect_t, atol=atol), + "Output (" + var_name + ") has diff" + str(actual_t) + "\n" + + str(expect_t)) + self.assertListEqual(actual.lod(), + expect.lod(), + "Output (" + var_name + ") has different lod") + + def _get_input_names(self): + inputs = [] + for name, value in list(self.inputs.items()): + if isinstance(value, list): + inputs.extend([sub_name for sub_name, _ in value]) + inputs.append(name) + return inputs + + def _get_output_names(self): + outputs = [] + for var_name, var in list(self.outputs.items()): + if isinstance(var, list): + for sub_var_name, sub_var in var: + outputs.append(sub_var_name) + else: + outputs.append(var_name) + if len(outputs) == 0: + for out_name, out_dup in Operator.get_op_outputs(self.op_type): + outputs.append(str(out_name)) + return outputs + + def check_output_stability(self, atol=1e-8): + places = self._get_places() + if len(places) < 2: + return + cpu_outs, fetch_list = self._calc_output(places[0]) + gpu_outs, _ = self._calc_output(places[1]) + self._assert_cpu_gpu_same(cpu_outs, gpu_outs, fetch_list, atol) + + def timeit_output_with_place(self, place, iters): + return self.timeit_function(self.calc_output, iters, place) + + def timeit_output(self, iters=100): + places = self._get_places() + elapses = [] + for place in places: + elapses.append(self.timeit_output_with_place(place, iters)) + for place, elapse in zip(places, elapses): + print("One pass of ({2}_op) at {0} cost {1}".format( + str(place), elapse, self.op_type)) + + def timeit_grad_with_place(self, place, iters=100): + inputs_to_check = self._get_input_names() + output_names = self._get_output_names() + return self.timeit_function( + self._get_gradient, + iters, + inputs_to_check, + place, + output_names, + no_grad_set=None) + + def timeit_grad(self, iters=100): + places = self._get_places() + elapses = [] + for place in places: + elapses.append(self.timeit_grad_with_place(place, iters)) + for place, elapse in zip(places, elapses): + print("One pass of ({2}_grad_op) at {0} cost {1}".format( + str(place), elapse, self.op_type)) diff --git a/python/paddle/fluid/tests/unittests/benchmark_sum_op.py b/python/paddle/fluid/tests/unittests/benchmark_sum_op.py new file mode 100644 index 0000000000000000000000000000000000000000..91a5f1bca4441d80489a02eb9283928e38321826 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/benchmark_sum_op.py @@ -0,0 +1,82 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +import paddle.fluid as fluid +from benchmark import BenchmarkSuite +from op_test import OpTest + +# This is a demo op test case for operator benchmarking and high resolution number stability alignment. + + +class TestSumOp(BenchmarkSuite): + def setUp(self): + self.op_type = "sum" + self.customize_testcase() + self.customize_fetch_list() + + def customize_fetch_list(self): + """ + customize fetch list, configure the wanted variables. + >>> self.fetch_list = ["Out"] + """ + self.fetch_list = ["Out"] + # pass + + def customize_testcase(self): + # a test case + x0 = np.random.random((300, 400)).astype('float32') + x1 = np.random.random((300, 400)).astype('float32') + x2 = np.random.random((300, 400)).astype('float32') + + # NOTE: if the output is empty, then it will autofilled by benchmarkSuite. + # only the output dtype is used, the shape, lod and data is computed from input. + self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} + self.outputs = {"Out": x0 + x1 + x2} + + def test_check_output(self): + """ + compare the output with customized output. In this case, + you should set the correct output by hands. + >>> self.outputs = {"Out": x0 + x1 + x2} + """ + self.check_output(atol=1e-8) + + def test_output_stability(self): + # compare the cpu gpu output in high resolution. + self.check_output_stability() + + def test_timeit_output(self): + """ + perf the op, time cost will be averged in iters. + output example + >>> One pass of (sum_op) at CPUPlace cost 0.000461330413818 + >>> One pass of (sum_op) at CUDAPlace(0) cost 0.000556070804596 + """ + self.timeit_output(iters=100) + + def test_timeit_grad(self): + """ + perf the op gradient, time cost will be averged in iters. + output example + >>> One pass of (sum_grad_op) at CPUPlace cost 0.00279935121536 + >>> One pass of (sum_grad_op) at CUDAPlace(0) cost 0.00500632047653 + """ + self.timeit_grad(iters=100) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/dist_mnist.py b/python/paddle/fluid/tests/unittests/dist_mnist.py new file mode 100644 index 0000000000000000000000000000000000000000..8f5ba33f7cbf5286edc4503c219fd3cdff60c517 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_mnist.py @@ -0,0 +1,103 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import argparse +import time +import math + +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler +from paddle.fluid import core +import unittest +from multiprocessing import Process +import os +import signal +from functools import reduce +from test_dist_base import TestDistRunnerBase, runtime_main + +DTYPE = "float32" +paddle.dataset.mnist.fetch() + +# Fix seed for test +fluid.default_startup_program().random_seed = 1 +fluid.default_main_program().random_seed = 1 + + +def cnn_model(data): + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=data, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu", + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant())) + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu", + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant())) + + SIZE = 10 + input_shape = conv_pool_2.shape + param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE] + scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5 + + predict = fluid.layers.fc( + input=conv_pool_2, + size=SIZE, + act="softmax", + param_attr=fluid.param_attr.ParamAttr( + initializer=fluid.initializer.NormalInitializer( + loc=0.0, scale=scale, seed=1))) + return predict + + +class TestDistMnist2x2(TestDistRunnerBase): + def get_model(self, batch_size=2): + # Input data + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + inference_program = fluid.default_main_program().clone() + # Optimization + opt = fluid.optimizer.AdamOptimizer( + learning_rate=0.001, beta1=0.9, beta2=0.999) + + # Reader + train_reader = paddle.batch( + paddle.dataset.mnist.train(), batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + opt.minimize(avg_cost) + return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict + + +if __name__ == "__main__": + runtime_main(TestDistMnist2x2) diff --git a/python/paddle/fluid/tests/unittests/dist_se_resnext.py b/python/paddle/fluid/tests/unittests/dist_se_resnext.py new file mode 100644 index 0000000000000000000000000000000000000000..d576a173ce2546119ede49128ef69d240c7cf482 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_se_resnext.py @@ -0,0 +1,248 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import argparse +import six +import time +import math + +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler +from paddle.fluid import core +import unittest +from multiprocessing import Process +import os +import sys +import signal +from test_dist_base import TestDistRunnerBase, runtime_main + +# Fix seed for test +fluid.default_startup_program().random_seed = 1 +fluid.default_main_program().random_seed = 1 + +train_parameters = { + "input_size": [3, 224, 224], + "input_mean": [0.485, 0.456, 0.406], + "input_std": [0.229, 0.224, 0.225], + "learning_strategy": { + "name": "piecewise_decay", + "epochs": [30, 60, 90], + "steps": [0.1, 0.01, 0.001, 0.0001] + } +} + + +class SE_ResNeXt(): + def __init__(self, layers=50): + self.params = train_parameters + self.layers = layers + + def net(self, input, class_dim=1000): + layers = self.layers + supported_layers = [50, 101, 152] + assert layers in supported_layers, \ + "supported layers are {} but input layer is {}".format(supported_layers, layers) + if layers == 50: + cardinality = 32 + reduction_ratio = 16 + depth = [3, 4, 6, 3] + num_filters = [128, 256, 512, 1024] + + conv = self.conv_bn_layer( + input=input, + num_filters=64, + filter_size=7, + stride=2, + act='relu') + conv = fluid.layers.pool2d( + input=conv, + pool_size=3, + pool_stride=2, + pool_padding=1, + pool_type='max') + elif layers == 101: + cardinality = 32 + reduction_ratio = 16 + depth = [3, 4, 23, 3] + num_filters = [128, 256, 512, 1024] + + conv = self.conv_bn_layer( + input=input, + num_filters=64, + filter_size=7, + stride=2, + act='relu') + conv = fluid.layers.pool2d( + input=conv, + pool_size=3, + pool_stride=2, + pool_padding=1, + pool_type='max') + elif layers == 152: + cardinality = 64 + reduction_ratio = 16 + depth = [3, 8, 36, 3] + num_filters = [128, 256, 512, 1024] + + conv = self.conv_bn_layer( + input=input, + num_filters=64, + filter_size=3, + stride=2, + act='relu') + conv = self.conv_bn_layer( + input=conv, num_filters=64, filter_size=3, stride=1, act='relu') + conv = self.conv_bn_layer( + input=conv, + num_filters=128, + filter_size=3, + stride=1, + act='relu') + conv = fluid.layers.pool2d( + input=conv, pool_size=3, pool_stride=2, pool_padding=1, \ + pool_type='max') + + for block in range(len(depth)): + for i in range(depth[block]): + conv = self.bottleneck_block( + input=conv, + num_filters=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + cardinality=cardinality, + reduction_ratio=reduction_ratio) + + pool = fluid.layers.pool2d( + input=conv, pool_size=7, pool_type='avg', global_pooling=True) + drop = fluid.layers.dropout(x=pool, dropout_prob=0.2) + stdv = 1.0 / math.sqrt(drop.shape[1] * 1.0) + out = fluid.layers.fc(input=drop, size=class_dim, act='softmax') + return out + + def shortcut(self, input, ch_out, stride): + ch_in = input.shape[1] + if ch_in != ch_out or stride != 1: + filter_size = 1 + return self.conv_bn_layer(input, ch_out, filter_size, stride) + else: + return input + + def bottleneck_block(self, input, num_filters, stride, cardinality, + reduction_ratio): + conv0 = self.conv_bn_layer( + input=input, num_filters=num_filters, filter_size=1, act='relu') + conv1 = self.conv_bn_layer( + input=conv0, + num_filters=num_filters, + filter_size=3, + stride=stride, + groups=cardinality, + act='relu') + conv2 = self.conv_bn_layer( + input=conv1, num_filters=num_filters * 2, filter_size=1, act=None) + scale = self.squeeze_excitation( + input=conv2, + num_channels=num_filters * 2, + reduction_ratio=reduction_ratio) + + short = self.shortcut(input, num_filters * 2, stride) + + return fluid.layers.elementwise_add(x=short, y=scale, act='relu') + + def conv_bn_layer(self, + input, + num_filters, + filter_size, + stride=1, + groups=1, + act=None): + conv = fluid.layers.conv2d( + input=input, + num_filters=num_filters, + filter_size=filter_size, + stride=stride, + padding=(filter_size - 1) / 2, + groups=groups, + act=None, + # avoid pserver CPU init differs from GPU + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant()), + bias_attr=False) + return fluid.layers.batch_norm(input=conv, act=act) + + def squeeze_excitation(self, input, num_channels, reduction_ratio): + pool = fluid.layers.pool2d( + input=input, pool_size=0, pool_type='avg', global_pooling=True) + stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0) + squeeze = fluid.layers.fc(input=pool, + size=num_channels / reduction_ratio, + act='relu') + stdv = 1.0 / math.sqrt(squeeze.shape[1] * 1.0) + excitation = fluid.layers.fc(input=squeeze, + size=num_channels, + act='sigmoid') + scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0) + return scale + + +class DistSeResneXt2x2(TestDistRunnerBase): + def get_model(self, batch_size=2): + # Input data + image = fluid.layers.data( + name="data", shape=[3, 224, 224], dtype='float32') + label = fluid.layers.data(name="int64", shape=[1], dtype='int64') + + # Train program + model = SE_ResNeXt(layers=50) + out = model.net(input=image, class_dim=102) + cost = fluid.layers.cross_entropy(input=out, label=label) + + avg_cost = fluid.layers.mean(x=cost) + acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1) + acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5) + + # Evaluator + test_program = fluid.default_main_program().clone(for_test=True) + + # Optimization + total_images = 6149 # flowers + epochs = [30, 60, 90] + step = int(total_images / batch_size + 1) + + bd = [step * e for e in epochs] + base_lr = 0.1 + lr = [] + lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)] + + optimizer = fluid.optimizer.Momentum( + # FIXME(typhoonzero): add back LR decay once ParallelExecutor fixed. + #learning_rate=fluid.layers.piecewise_decay( + # boundaries=bd, values=lr), + learning_rate=base_lr, + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + optimizer.minimize(avg_cost) + + # Reader + train_reader = paddle.batch( + paddle.dataset.flowers.train(), batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.flowers.test(use_xmap=False), batch_size=batch_size) + + return test_program, avg_cost, train_reader, test_reader, acc_top1, out + + +if __name__ == "__main__": + runtime_main(DistSeResneXt2x2) diff --git a/python/paddle/fluid/tests/unittests/dist_transformer.py b/python/paddle/fluid/tests/unittests/dist_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..ee8020a73546cb9037e9dc4be589c62bb1b6b937 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_transformer.py @@ -0,0 +1,280 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import argparse +import time +import math + +import paddle +import paddle.fluid as fluid +from paddle.fluid import core +import os +import sys +import transformer_model +import paddle.dataset.wmt16 as wmt16 + +# Fix seed for test +fluid.default_startup_program().random_seed = 1 +fluid.default_main_program().random_seed = 1 + +WMT16_RECORDIO_FILE = "/tmp/wmt16.recordio" + + +class ModelHyperParams(object): + # Dictionary size for source and target language. This model directly uses + # paddle.dataset.wmt16 in which , and token has + # alreay been added, but the token is not added. Transformer requires + # sequences in a mini-batch are padded to have the same length. A token is + # added into the original dictionary in paddle.dateset.wmt16. + + # size of source word dictionary. + src_vocab_size = 10000 + # index for token in source language. + src_pad_idx = src_vocab_size + + # size of target word dictionay + trg_vocab_size = 10000 + # index for token in target language. + trg_pad_idx = trg_vocab_size + + # position value corresponding to the token. + pos_pad_idx = 0 + + # max length of sequences. It should plus 1 to include position + # padding token for position encoding. + max_length = 50 + + # the dimension for word embeddings, which is also the last dimension of + # the input and output of multi-head attention, position-wise feed-forward + # networks, encoder and decoder. + + d_model = 512 + # size of the hidden layer in position-wise feed-forward networks. + d_inner_hid = 1024 + # the dimension that keys are projected to for dot-product attention. + d_key = 64 + # the dimension that values are projected to for dot-product attention. + d_value = 64 + # number of head used in multi-head attention. + n_head = 8 + # number of sub-layers to be stacked in the encoder and decoder. + n_layer = 6 + # dropout rate used by all dropout layers. + dropout = 0.1 + + +def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): + """ + Pad the instances to the max sequence length in batch, and generate the + corresponding position data and attention bias. Then, convert the numpy + data to tensors and return a dict mapping names to tensors. + """ + + def __pad_batch_data(insts, + pad_idx, + is_target=False, + return_pos=True, + return_attn_bias=True, + return_max_len=True): + """ + Pad the instances to the max sequence length in batch, and generate the + corresponding position data and attention bias. + """ + return_list = [] + max_len = max(len(inst) for inst in insts) + inst_data = np.array( + [inst + [pad_idx] * (max_len - len(inst)) for inst in insts]) + return_list += [inst_data.astype("int64").reshape([-1, 1])] + if return_pos: + inst_pos = np.array([[ + pos_i + 1 if w_i != pad_idx else 0 + for pos_i, w_i in enumerate(inst) + ] for inst in inst_data]) + + return_list += [inst_pos.astype("int64").reshape([-1, 1])] + if return_attn_bias: + if is_target: + # This is used to avoid attention on paddings and subsequent + # words. + slf_attn_bias_data = np.ones((inst_data.shape[0], max_len, + max_len)) + slf_attn_bias_data = np.triu(slf_attn_bias_data, 1).reshape( + [-1, 1, max_len, max_len]) + slf_attn_bias_data = np.tile(slf_attn_bias_data, + [1, n_head, 1, 1]) * [-1e9] + else: + # This is used to avoid attention on paddings. + slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] * + (max_len - len(inst)) + for inst in insts]) + slf_attn_bias_data = np.tile( + slf_attn_bias_data.reshape([-1, 1, 1, max_len]), + [1, n_head, max_len, 1]) + return_list += [slf_attn_bias_data.astype("float32")] + if return_max_len: + return_list += [max_len] + return return_list if len(return_list) > 1 else return_list[0] + + src_word, src_pos, src_slf_attn_bias, src_max_len = __pad_batch_data( + [inst[0] for inst in insts], src_pad_idx, is_target=False) + trg_word, trg_pos, trg_slf_attn_bias, trg_max_len = __pad_batch_data( + [inst[1] for inst in insts], trg_pad_idx, is_target=True) + trg_src_attn_bias = np.tile(src_slf_attn_bias[:, :, ::src_max_len, :], + [1, 1, trg_max_len, 1]).astype("float32") + lbl_word = __pad_batch_data([inst[2] for inst in insts], trg_pad_idx, False, + False, False, False) + lbl_weight = (lbl_word != trg_pad_idx).astype("float32").reshape([-1, 1]) + + return [ + src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, + trg_slf_attn_bias, trg_src_attn_bias, lbl_word, lbl_weight + ] + + +def transformer(use_feed): + assert not use_feed, "transfomer doesn't support feed yet" + return transformer_model.transformer( + ModelHyperParams.src_vocab_size + 1, + ModelHyperParams.trg_vocab_size + 1, ModelHyperParams.max_length + 1, + ModelHyperParams.n_layer, ModelHyperParams.n_head, + ModelHyperParams.d_key, ModelHyperParams.d_value, + ModelHyperParams.d_model, ModelHyperParams.d_inner_hid, + ModelHyperParams.dropout, ModelHyperParams.src_pad_idx, + ModelHyperParams.trg_pad_idx, ModelHyperParams.pos_pad_idx) + + +def get_model(): + avg_cost = transformer(use_feed=False) + optimizer = fluid.optimizer.Adam() + optimizer.minimize(avg_cost) + return avg_cost + + +def get_transpiler(trainer_id, main_program, pserver_endpoints, trainers): + t = fluid.DistributeTranspiler() + t.transpile( + trainer_id=trainer_id, + program=main_program, + pservers=pserver_endpoints, + trainers=trainers) + return t + + +class DistTransformer2x2(object): + def run_pserver(self, pserver_endpoints, trainers, current_endpoint, + trainer_id): + get_model() + t = get_transpiler(trainer_id, + fluid.default_main_program(), pserver_endpoints, + trainers) + pserver_prog = t.get_pserver_program(current_endpoint) + startup_prog = t.get_startup_program(current_endpoint, pserver_prog) + + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + exe.run(pserver_prog) + + def _wait_ps_ready(self, pid): + retry_times = 20 + while True: + assert retry_times >= 0, "wait ps ready failed" + time.sleep(3) + print("waiting ps ready: ", pid) + try: + # the listen_and_serv_op would touch a file which contains the listen port + # on the /tmp directory until it was ready to process all the RPC call. + os.stat("/tmp/paddle.%d.port" % pid) + return + except os.error: + retry_times -= 1 + + def run_trainer(self, place, endpoints, trainer_id, trainers, is_dist=True): + avg_cost = get_model() + if is_dist: + t = get_transpiler(trainer_id, + fluid.default_main_program(), endpoints, + trainers) + trainer_prog = t.get_trainer_program() + else: + trainer_prog = fluid.default_main_program() + + startup_exe = fluid.Executor(place) + startup_exe.run(fluid.default_startup_program()) + + strategy = fluid.ExecutionStrategy() + strategy.num_threads = 1 + strategy.allow_op_delay = False + exe = fluid.ParallelExecutor( + True, loss_name=avg_cost.name, exec_strategy=strategy) + + first_loss, = exe.run(fetch_list=[avg_cost.name]) + print(first_loss) + for i in xrange(5): + _ = exe.run(fetch_list=[avg_cost.name]) + last_loss, = exe.run(fetch_list=[avg_cost.name]) + print(last_loss) + + +def main(role="pserver", + endpoints="127.0.0.1:9123", + trainer_id=0, + current_endpoint="127.0.0.1:9123", + trainers=1, + is_dist=True): + + reader = paddle.batch( + wmt16.train(ModelHyperParams.src_vocab_size, + ModelHyperParams.trg_vocab_size), + batch_size=transformer_model.batch_size) + + with fluid.recordio_writer.create_recordio_writer( + WMT16_RECORDIO_FILE) as writer: + for batch in reader(): + for tensor in prepare_batch_input( + batch, ModelHyperParams.src_pad_idx, + ModelHyperParams.trg_pad_idx, ModelHyperParams.n_head): + t = fluid.LoDTensor() + t.set(tensor, fluid.CPUPlace()) + writer.append_tensor(t) + writer.complete_append_tensor() + + model = DistTransformer2x2() + if role == "pserver": + model.run_pserver(endpoints, trainers, current_endpoint, trainer_id) + else: + p = fluid.CUDAPlace(0) if core.is_compiled_with_cuda( + ) else fluid.CPUPlace() + model.run_trainer(p, endpoints, trainer_id, trainers, is_dist) + + +if __name__ == "__main__": + if len(sys.argv) != 7: + print( + "Usage: python dist_transformer.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist]" + ) + role = sys.argv[1] + endpoints = sys.argv[2] + trainer_id = int(sys.argv[3]) + current_endpoint = sys.argv[4] + trainers = int(sys.argv[5]) + is_dist = True if sys.argv[6] == "TRUE" else False + main( + role=role, + endpoints=endpoints, + trainer_id=trainer_id, + current_endpoint=current_endpoint, + trainers=trainers, + is_dist=is_dist) diff --git a/python/paddle/fluid/tests/unittests/dist_word2vec.py b/python/paddle/fluid/tests/unittests/dist_word2vec.py new file mode 100644 index 0000000000000000000000000000000000000000..54a70f4adb4a9bb24e3c618a7fe71f42a376609b --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_word2vec.py @@ -0,0 +1,119 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import argparse +import time +import math +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler +from paddle.fluid import core +import unittest +from multiprocessing import Process +import os +import signal +from test_dist_base import TestDistRunnerBase, runtime_main + +IS_SPARSE = True +EMBED_SIZE = 32 +HIDDEN_SIZE = 256 +N = 5 + +# Fix seed for test +fluid.default_startup_program().random_seed = 1 +fluid.default_main_program().random_seed = 1 + + +class TestDistWord2vec2x2(TestDistRunnerBase): + def get_model(self, batch_size=2): + BATCH_SIZE = batch_size + + def __network__(words): + embed_first = fluid.layers.embedding( + input=words[0], + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr=fluid.ParamAttr( + name='shared_w', initializer=fluid.initializer.Constant())) + embed_second = fluid.layers.embedding( + input=words[1], + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr=fluid.ParamAttr( + name='shared_w', initializer=fluid.initializer.Constant())) + embed_third = fluid.layers.embedding( + input=words[2], + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr=fluid.ParamAttr( + name='shared_w', initializer=fluid.initializer.Constant())) + embed_forth = fluid.layers.embedding( + input=words[3], + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr=fluid.ParamAttr( + name='shared_w', initializer=fluid.initializer.Constant())) + + concat_embed = fluid.layers.concat( + input=[embed_first, embed_second, embed_third, embed_forth], + axis=1) + hidden1 = fluid.layers.fc( + input=concat_embed, + size=HIDDEN_SIZE, + act='sigmoid', + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant())) + predict_word = fluid.layers.fc( + input=hidden1, + size=dict_size, + act='softmax', + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant())) + cost = fluid.layers.cross_entropy( + input=predict_word, label=words[4]) + avg_cost = fluid.layers.mean(cost) + return avg_cost, predict_word + + word_dict = paddle.dataset.imikolov.build_dict() + dict_size = len(word_dict) + + first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64') + second_word = fluid.layers.data( + name='secondw', shape=[1], dtype='int64') + third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64') + forth_word = fluid.layers.data(name='forthw', shape=[1], dtype='int64') + next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') + avg_cost, predict_word = __network__( + [first_word, second_word, third_word, forth_word, next_word]) + + inference_program = paddle.fluid.default_main_program().clone() + + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + train_reader = paddle.batch( + paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) + test_reader = paddle.batch( + paddle.dataset.imikolov.test(word_dict, N), BATCH_SIZE) + + return inference_program, avg_cost, train_reader, test_reader, None, predict_word + + +if __name__ == "__main__": + runtime_main(TestDistWord2vec2x2) diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index b611470fa1ff326df960c349b71006f52d586d8e..b27d773f09d9a6daad5a10b65e683f4e11881de1 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -15,91 +15,30 @@ import unittest import numpy as np import random +import time import itertools -import paddle.fluid.core as core import collections + +import paddle.fluid as fluid +import paddle.fluid.core as core from paddle.fluid.backward import append_backward from paddle.fluid.op import Operator from paddle.fluid.executor import Executor -from paddle.fluid.framework import Program, OpProtoHolder +from paddle.fluid.framework import Program, OpProtoHolder, Variable +from testsuite import create_op, set_input, append_input_output, append_loss_ops +from functools import reduce +from six.moves import zip def randomize_probability(batch_size, class_num, dtype='float32'): prob = np.random.uniform( 0.1, 1.0, size=(batch_size, class_num)).astype(dtype) prob_sum = prob.sum(axis=1) - for i in xrange(len(prob)): + for i in range(len(prob)): prob[i] /= prob_sum[i] return prob -def create_op(scope, op_type, inputs, outputs, attrs): - kwargs = dict() - - op_maker = core.op_proto_and_checker_maker - op_role_attr_name = op_maker.kOpRoleAttrName() - - if op_role_attr_name not in attrs: - attrs[op_role_attr_name] = int(op_maker.OpRole.Forward) - - def __create_var__(name, var_name): - scope.var(var_name).get_tensor() - kwargs[name].append(var_name) - - for in_name, in_dup in Operator.get_op_inputs(op_type): - if in_name in inputs: - kwargs[in_name] = [] - if in_dup: - sub_in = inputs[in_name] - for item in sub_in: - sub_in_name, _ = item[0], item[1] - __create_var__(in_name, sub_in_name) - else: - __create_var__(in_name, in_name) - - for out_name, out_dup in Operator.get_op_outputs(op_type): - if out_name in outputs: - kwargs[out_name] = [] - if out_dup: - sub_out = outputs[out_name] - for item in sub_out: - sub_out_name, _ = item[0], item[1] - __create_var__(out_name, sub_out_name) - else: - __create_var__(out_name, out_name) - - for attr_name in Operator.get_op_attr_names(op_type): - if attr_name in attrs: - kwargs[attr_name] = attrs[attr_name] - - return Operator(op_type, **kwargs) - - -def set_input(scope, op, inputs, place): - def __set_input__(var_name, var): - if isinstance(var, tuple) or isinstance(var, np.ndarray): - tensor = scope.find_var(var_name).get_tensor() - if isinstance(var, tuple): - tensor.set_lod(var[1]) - var = var[0] - tensor.set_dims(var.shape) - tensor.set(var, place) - elif isinstance(var, float): - scope.find_var(var_name).set_float(var) - elif isinstance(var, int): - scope.find_var(var_name).set_int(var) - - for in_name, in_dup in Operator.get_op_inputs(op.type()): - if in_name in inputs: - if in_dup: - sub_in = inputs[in_name] - for item in sub_in: - sub_in_name, sub_in_val = item[0], item[1] - __set_input__(sub_in_name, sub_in_val) - else: - __set_input__(in_name, inputs[in_name]) - - def get_numeric_gradient(place, scope, op, @@ -123,12 +62,16 @@ def get_numeric_gradient(place, return np.array(sum).mean() tensor_to_check = scope.find_var(input_to_check).get_tensor() - tensor_size = product(tensor_to_check.get_dims()) - tensor_to_check_dtype = tensor_to_check.dtype() + tensor_size = product(tensor_to_check.shape()) + tensor_to_check_dtype = tensor_to_check._dtype() if tensor_to_check_dtype == core.VarDesc.VarType.FP32: tensor_to_check_dtype = np.float32 elif tensor_to_check_dtype == core.VarDesc.VarType.FP64: tensor_to_check_dtype = np.float64 + elif tensor_to_check_dtype == core.VarDesc.VarType.FP16: + tensor_to_check_dtype = np.float16 + # set delta as np.float16, will automatic convert to float32, float64 + delta = np.array(delta).astype(np.float16) else: raise ValueError("Not supported data type " + str( tensor_to_check_dtype)) @@ -136,20 +79,31 @@ def get_numeric_gradient(place, gradient_flat = np.zeros(shape=(tensor_size, ), dtype=tensor_to_check_dtype) def __get_elem__(tensor, i): - if tensor_to_check_dtype == np.float32: - return tensor.get_float_element(i) + if tensor_to_check_dtype == np.float16: + numpy_tensor = np.array(tensor).astype(np.float16) + numpy_tensor = numpy_tensor.flatten() + return numpy_tensor[i] + elif tensor_to_check_dtype == np.float32: + return tensor._get_float_element(i) else: - return tensor.get_double_element(i) + return tensor._get_double_element(i) def __set_elem__(tensor, i, e): - if tensor_to_check_dtype == np.float32: - tensor.set_float_element(i, e) + if tensor_to_check_dtype == np.float16: + numpy_tensor = np.array(tensor).astype(np.float16) + shape = numpy_tensor.shape + numpy_tensor = numpy_tensor.flatten() + numpy_tensor[i] = e + numpy_tensor = numpy_tensor.reshape(shape).view(np.uint16) + tensor.set(numpy_tensor, place) + elif tensor_to_check_dtype == np.float32: + tensor._set_float_element(i, e) else: - tensor.set_double_element(i, e) + tensor._set_double_element(i, e) # we only compute gradient of one element each time. # we use a for loop to compute the gradient of every element. - for i in xrange(tensor_size): + for i in range(tensor_size): if in_place: set_input(scope, op, inputs, place) @@ -170,49 +124,7 @@ def get_numeric_gradient(place, __set_elem__(tensor_to_check, i, origin) gradient_flat[i] = (y_pos - y_neg) / delta / 2 - return gradient_flat.reshape(tensor_to_check.get_dims()) - - -def append_input_output(block, op_proto, np_list, is_input): - '''Insert VarDesc and generate Python variable instance''' - proto_list = op_proto.inputs if is_input else op_proto.outputs - - def create_var(block, name, np_list, var_proto): - if name not in np_list: - assert var_proto.intermediate, "{} not found".format(name) - shape = None - lod_level = None - else: - np_value = np_list[name] - if isinstance(np_value, tuple): - shape = list(np_value[0].shape) - lod_level = len(np_value[1]) - else: - shape = list(np_value.shape) - lod_level = 0 - return block.create_var( - dtype="float32", shape=shape, lod_level=lod_level, name=name) - - var_dict = {} - for var_proto in proto_list: - var_name = str(var_proto.name) - if is_input: - if (var_name not in np_list) and var_proto.dispensable: - continue - assert (var_name in np_list) or (var_proto.dispensable), \ - "Missing {} as input".format(var_name) - if var_proto.duplicable: - assert isinstance(np_list[var_name], list), \ - "Duplicable {} should be set as list".format(var_name) - var_list = [] - for (name, np_value) in np_list[var_name]: - var_list.append( - create_var(block, name, {name: np_value}, var_proto)) - var_dict[var_name] = var_list - else: - var_dict[var_name] = create_var(block, var_name, np_list, var_proto) - - return var_dict + return gradient_flat.reshape(tensor_to_check.shape()) class OpTest(unittest.TestCase): @@ -221,16 +133,49 @@ class OpTest(unittest.TestCase): '''Fix random seeds to remove randomness from tests''' cls._np_rand_state = np.random.get_state() cls._py_rand_state = random.getstate() + cls.call_once = False + cls.dtype = "float32" + cls.outputs = {} np.random.seed(123) random.seed(124) @classmethod def tearDownClass(cls): - '''Restore random seeds''' + """Restore random seeds""" np.random.set_state(cls._np_rand_state) random.setstate(cls._py_rand_state) + def try_call_once(self, data_type): + if not self.call_once: + self.call_once = True + self.dtype = data_type + # See the comment of np_dtype_to_fluid_dtype + # If the input type is uint16, we assume use float16 + # for lodtensor dtype. + if self.dtype == np.uint16: + self.dtype == np.float16 + + def infer_dtype_from_inputs_outputs(self, inputs, outputs): + def infer_dtype(numpy_dict): + assert isinstance( + numpy_dict, + dict), "self.inputs, self.outputs must be numpy_dict" + for var_name, var_value in numpy_dict.items(): + if isinstance(var_value, (np.ndarray, np.generic)): + self.try_call_once(var_value.dtype) + elif isinstance(var_value, (list, tuple)): + # the case of self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} + if len(var_value) > 1 and isinstance(var_value[1], ( + np.ndarray, np.generic)): + instance = var_value[1] + self.try_call_once(instance[1].dtype) + else: + self.try_call_once("float32") + + infer_dtype(inputs) + infer_dtype(outputs) + def feed_var(self, input_vars, place): feed_map = {} for var_name in input_vars: @@ -238,34 +183,37 @@ class OpTest(unittest.TestCase): for name, np_value in self.inputs[var_name]: tensor = core.LoDTensor() if isinstance(np_value, tuple): - tensor.set(np_value[0], place) - tensor.set_lod(np_value[1]) + tensor.set( + OpTest.np_value_to_fluid_value(np_value[0]), place) + tensor.set_recursive_sequence_lengths(np_value[1]) else: - tensor.set(np_value, place) + tensor.set( + OpTest.np_value_to_fluid_value(np_value), place) feed_map[name] = tensor else: tensor = core.LoDTensor() if isinstance(self.inputs[var_name], tuple): - tensor.set(self.inputs[var_name][0], place) - tensor.set_lod(self.inputs[var_name][1]) + tensor.set( + OpTest.np_value_to_fluid_value(self.inputs[var_name][ + 0]), place) + tensor.set_recursive_sequence_lengths(self.inputs[var_name][ + 1]) else: - tensor.set(self.inputs[var_name], place) + tensor.set( + OpTest.np_value_to_fluid_value(self.inputs[var_name]), + place) feed_map[var_name] = tensor return feed_map - def calc_output(self, place): - outs, _ = self._calc_output(place) - return outs - - def _calc_output(self, place): + def _append_ops(self, block): op_proto = OpProtoHolder.instance().get_op_proto(self.op_type) - - program = Program() - block = program.global_block() - - inputs = append_input_output(block, op_proto, self.inputs, True) - outputs = append_input_output(block, op_proto, self.outputs, False) + "infer datatype from inputs and outputs for this test case" + self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs) + inputs = append_input_output(block, op_proto, self.inputs, True, + self.dtype) + outputs = append_input_output(block, op_proto, self.outputs, False, + self.dtype) op = block.append_op( type=self.op_type, inputs=inputs, @@ -275,22 +223,68 @@ class OpTest(unittest.TestCase): op.desc.infer_var_type(block.desc) op.desc.infer_shape(block.desc) - fetch_list = [] - for var_name, var in outputs.iteritems(): - if var_name in self.outputs: + def _get_io_vars(self, block, numpy_inputs): + inputs = {} + for name, value in numpy_inputs.items(): + if isinstance(value, list): + var_list = [ + block.var(sub_name) for sub_name, sub_value in value + ] + inputs[name] = var_list + else: + inputs[name] = block.var(name) + return inputs + + def _get_inputs(self, block): + return self._get_io_vars(block, self.inputs) + + def _get_outputs(self, block): + return self._get_io_vars(block, self.outputs) + + def calc_output(self, place): + outs, _ = self._calc_output(place) + return outs + + def _calc_output(self, place, parallel=False): + + program = Program() + block = program.global_block() + self._append_ops(block) + + inputs = self._get_inputs(block) + outputs = self._get_outputs(block) + feed_map = self.feed_var(inputs, place) + + if parallel: + use_cuda = False + if isinstance(place, fluid.CUDAPlace(0)): + use_cuda = True + executor = fluid.ParallelExecutor( + use_cuda=use_cuda, loss_name=loss.name, main_program=program) + else: + executor = Executor(place) + + fetch_list = getattr(self, "fetch_list", []) + # if the fetch_list is customized by user, we use it directly. + # if not, fill the fetch_list by the user configured outputs in test. + if len(fetch_list) == 0: + for var_name, var in outputs.items(): if isinstance(var, list): for v in var: fetch_list.append(v) else: fetch_list.append(var) - - feed_map = self.feed_var(inputs, place) - - exe = Executor(place) - outs = exe.run(program, - feed=feed_map, - fetch_list=fetch_list, - return_numpy=False) + # if the fetch_list still empty, fill the fetch_list by the operator output. + if len(fetch_list) == 0: + for out_name, out_dup in Operator.get_op_outputs(self.op_type): + fetch_list.append(str(out_name)) + # fetch_list = map(block.var, fetch_list) + if not isinstance(fetch_list[0], fluid.framework.Variable): + fetch_list = list(map(block.var, fetch_list)) + outs = executor.run(program, + feed=feed_map, + fetch_list=fetch_list, + return_numpy=False) return outs, fetch_list def check_output_with_place(self, place, atol): @@ -328,7 +322,8 @@ class OpTest(unittest.TestCase): str(place)) if isinstance(expect, tuple): self.assertListEqual( - actual.lod(), expect[1], "Output (" + sub_out_name + + actual.recursive_sequence_lengths(), expect[1], + "Output (" + sub_out_name + ") has different lod at " + str(place)) else: idx = find_actual(out_name, fetch_list) @@ -340,23 +335,34 @@ class OpTest(unittest.TestCase): np.allclose( actual_t, expect_t, atol=atol), "Output (" + out_name + ") has diff at " + str(place) + - str(actual_t) + "\n" + str(expect_t)) + "\nExpect " + str(expect_t) + "\n" + "But Got" + + str(actual_t)) if isinstance(expect, tuple): - self.assertListEqual(actual.lod(), expect[1], - "Output (" + out_name + + self.assertListEqual(actual.recursive_sequence_lengths(), + expect[1], "Output (" + out_name + ") has different lod at " + str(place)) - def check_output(self, atol=1e-5): - places = [core.CPUPlace()] + def _get_places(self): + if self.dtype == np.float16: + if core.is_compiled_with_cuda() and core.op_support_gpu( + self.op_type): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + return [place] + else: + return [] + places = [fluid.CPUPlace()] if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): places.append(core.CUDAPlace(0)) + return places + + def check_output(self, atol=1e-5): + places = self._get_places() for place in places: self.check_output_with_place(place, atol) def check_output_customized(self, checker): - places = [core.CPUPlace()] - if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): - places.append(core.CUDAPlace(0)) + places = self._get_places() for place in places: outs = self.calc_output(place) outs = [np.array(out) for out in outs] @@ -365,7 +371,7 @@ class OpTest(unittest.TestCase): def __assert_is_close(self, numeric_grads, analytic_grads, names, max_relative_error, msg_prefix): - for a, b, name in itertools.izip(numeric_grads, analytic_grads, names): + for a, b, name in zip(numeric_grads, analytic_grads, names): abs_a = np.abs(a) abs_a[abs_a < 1e-3] = 1 @@ -375,9 +381,9 @@ class OpTest(unittest.TestCase): def err_msg(): offset = np.argmax(diff_mat > max_relative_error) return ("%s Variable %s max gradient diff %f over limit %f, " - "the first error element is %d, %f, %f") % ( - msg_prefix, name, max_diff, max_relative_error, - offset, a.flatten()[offset], b.flatten()[offset]) + "the first error element is %d, expected %f, but got %f" + ) % (msg_prefix, name, max_diff, max_relative_error, + offset, a.flatten()[offset], b.flatten()[offset]) self.assertLessEqual(max_diff, max_relative_error, err_msg()) @@ -389,9 +395,7 @@ class OpTest(unittest.TestCase): in_place=False, max_relative_error=0.005, user_defined_grads=None): - places = [core.CPUPlace()] - if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): - places.append(core.CUDAPlace(0)) + places = self._get_places() for place in places: self.check_grad_with_place(place, inputs_to_check, output_names, no_grad_set, numeric_grad_delta, @@ -438,41 +442,12 @@ class OpTest(unittest.TestCase): max_relative_error, "Gradient Check On %s" % str(place)) - @staticmethod - def _create_var_descs_(block, var_dict): - # FIXME: Try unify with `append_input_output` - for param_name in var_dict: - var = var_dict[param_name] - if not isinstance(var, list) and not isinstance(var, tuple): - var = [(param_name, var, None)] - if not isinstance(var[0], list) and not isinstance(var[0], tuple): - var = [(param_name, var[0], var[1])] - - for i, item in enumerate(var): - if not isinstance(item[0], basestring): - item = [[param_name] + list(item)] - if len(item) == 2: - if isinstance(item[1], tuple): - var[i] = [item[0], item[1][0], item[1][1]] - else: - # only set var name and value, set lod to None - var[i] = list(item) + [None] - var_descs = [(block.create_var( - name=name, shape=each.shape, dtype=each.dtype), each, lod) - for name, each, lod in var] - - yield param_name, var_descs - - @staticmethod - def _merge_list(iterable): - return reduce(lambda a, b: list(a) + list(b), iterable, []) - @staticmethod def _numpy_to_lod_tensor(np_value, lod, place): tensor = core.LoDTensor() tensor.set(np_value, place) if lod is not None: - tensor.set_lod(lod) + tensor.set_recursive_sequence_lengths(lod) return tensor @staticmethod @@ -497,83 +472,46 @@ class OpTest(unittest.TestCase): input.dtype = np.uint16 return input - def _get_gradient(self, input_to_check, place, output_names, no_grad_set): - prog = Program() - block = prog.global_block() - inputs_with_np = { - key: value - for (key, value) in OpTest._create_var_descs_( - block, getattr(self, 'inputs', {})) - } - outputs_with_np = { - key: val - for (key, val) in OpTest._create_var_descs_( - block, getattr(self, 'outputs', {})) - } - inputs = { - k: [item[0] for item in inputs_with_np[k]] - for k in inputs_with_np - } - outputs = { - k: [item[0] for item in outputs_with_np[k]] - for k in outputs_with_np - } - - op = block.append_op( - type=self.op_type, - inputs=inputs, - outputs=outputs, - attrs=getattr(self, 'attrs', {})) - - # infer variable type and infer shape in compile-time - op.desc.infer_var_type(block.desc) - op.desc.infer_shape(block.desc) - - mean_inputs = map(block.var, output_names) + @staticmethod + def fluid_dtype_to_np_dtype(self, dtype): + """ + See above, convert the dtype to normal type. + """ + if dtype == np.uint16: + dtype = np.float16 + return dtype - if len(mean_inputs) == 1: - loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1]) - op = block.append_op( - inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean') - op.desc.infer_var_type(block.desc) - op.desc.infer_shape(block.desc) - else: - avg_sum = [] - for cur_loss in mean_inputs: - cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1]) - op = block.append_op( - inputs={"X": [cur_loss]}, - outputs={"Out": [cur_avg_loss]}, - type="mean") - op.desc.infer_var_type(block.desc) - op.desc.infer_shape(block.desc) - avg_sum.append(cur_avg_loss) - - loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1]) - op_sum = block.append_op( - inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum') - op_sum.desc.infer_var_type(block.desc) - op_sum.desc.infer_shape(block.desc) - - loss = block.create_var(dtype=loss_sum.dtype, shape=[1]) - op_loss = block.append_op( - inputs={"X": loss_sum}, - outputs={"Out": loss}, - type='scale', - attrs={'scale': 1.0 / float(len(avg_sum))}) - op_loss.desc.infer_var_type(block.desc) - op_loss.desc.infer_shape(block.desc) + @staticmethod + def np_value_to_fluid_value(input): + if input.dtype == np.float16: + input = input.view(np.uint16) + return input + def _get_gradient(self, + input_to_check, + place, + output_names, + no_grad_set, + parallel=False): + prog = Program() + block = prog.global_block() + self._append_ops(block) + loss = append_loss_ops(block, output_names) param_grad_list = append_backward( loss=loss, parameter_list=input_to_check, no_grad_set=no_grad_set) - feed_dict = { - item[0].name: OpTest._numpy_to_lod_tensor(item[1], item[2], place) - for p_name in inputs_with_np for item in inputs_with_np[p_name] - } + inputs = self._get_inputs(block) + feed_dict = self.feed_var(inputs, place) fetch_list = [g for p, g in param_grad_list] - executor = Executor(place) - return map(np.array, - executor.run(prog, feed_dict, fetch_list, - return_numpy=False)) + if parallel: + use_cuda = False + if isinstance(place, fluid.CUDAPlace(0)): + use_cuda = True + executor = fluid.ParallelExecutor( + use_cuda=use_cuda, loss_name=loss.name, main_program=prog) + else: + executor = Executor(place) + return list( + map(np.array, + executor.run(prog, feed_dict, fetch_list, return_numpy=False))) diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index c9c3c648717814c28c39a401487925824e885946..67c35e9de7e83699bf30ca946856bb907152cbdd 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -12,10 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. +import multiprocessing +import os import unittest import paddle.fluid as fluid import time import numpy as np +import math +import sys __all__ = ['TestParallelExecutorBase'] @@ -23,6 +27,7 @@ __all__ = ['TestParallelExecutorBase'] class TestParallelExecutorBase(unittest.TestCase): def check_network_convergence(self, method, + use_cuda=True, memory_opt=True, iter=50, batch_size=None, @@ -30,7 +35,8 @@ class TestParallelExecutorBase(unittest.TestCase): feed_dict=None, seed=None, use_parallel_executor=True, - balance_parameter_opt_between_cards=False): + use_reduce=False, + optimizer=fluid.optimizer.Adam): def run_executor(exe, feed, fetch_list, program=None): if isinstance(exe, fluid.ParallelExecutor): res = exe.run(fetch_list=fetch_list, feed=feed) @@ -45,26 +51,32 @@ class TestParallelExecutorBase(unittest.TestCase): main = fluid.Program() startup = fluid.Program() startup.random_seed = 1 # Fix random seed + main.random_seed = 1 with fluid.program_guard(main, startup): if seed is not None: startup.random_seed = seed + main.random_seed = seed + loss = method(use_feed=feed_dict is not None) - adam = fluid.optimizer.Adam() - adam.minimize(loss) + + optimizer().minimize(loss) + if memory_opt: fluid.memory_optimize(main) - place = fluid.CUDAPlace(0) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() startup_exe = fluid.Executor(place) startup_exe.run(startup) exec_strategy = fluid.ExecutionStrategy() exec_strategy.allow_op_delay = allow_op_delay build_strategy = fluid.BuildStrategy() - build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce if balance_parameter_opt_between_cards else fluid.BuildStrategy.ReduceStrategy.AllReduce + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \ + if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce if use_parallel_executor: exe = fluid.ParallelExecutor( - True, + use_cuda, loss_name=loss.name, exec_strategy=exec_strategy, build_strategy=build_strategy) @@ -72,13 +84,14 @@ class TestParallelExecutorBase(unittest.TestCase): exe = fluid.Executor(place=place) if batch_size is not None: - batch_size *= fluid.core.get_cuda_device_count() + batch_size *= fluid.core.get_cuda_device_count( + ) if use_cuda else int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) begin = time.time() first_loss, = run_executor( exe=exe, feed=feed_dict, fetch_list=[loss.name]) - first_loss = np.array(first_loss) - for i in xrange(iter): + for i in range(iter): run_executor(exe=exe, feed=feed_dict, fetch_list=[]) last_loss, = run_executor( @@ -86,11 +99,15 @@ class TestParallelExecutorBase(unittest.TestCase): end = time.time() if batch_size is not None: - print "%.4f Instance per second" % ( - (batch_size * iter + 2) / (end - begin)) + print("%.4f Instance per second" % ( + (batch_size * iter + 2) / (end - begin))) - last_loss = np.array(last_loss) + avg_last_loss_val = np.array(last_loss).mean() + avg_first_loss_val = np.array(first_loss).mean() + if math.isnan(float(avg_last_loss_val)) or math.isnan( + float(avg_first_loss_val)): + sys.exit("got NaN loss, training failed.") - print first_loss, last_loss + print(first_loss, last_loss) # self.assertGreater(first_loss[0], last_loss[0]) return first_loss, last_loss diff --git a/python/paddle/fluid/tests/unittests/test_accuracy_op.py b/python/paddle/fluid/tests/unittests/test_accuracy_op.py index 212a87e529da83c40ba8852e81bdf43d4611897b..db1861fd10e371ebe631a16380af591875886769 100644 --- a/python/paddle/fluid/tests/unittests/test_accuracy_op.py +++ b/python/paddle/fluid/tests/unittests/test_accuracy_op.py @@ -26,7 +26,7 @@ class TestAccuracyOp(OpTest): label = np.random.randint(0, 2, (n, 1)) self.inputs = {'Out': infer, 'Indices': indices, "Label": label} num_correct = 0 - for rowid in xrange(n): + for rowid in range(n): for ele in indices[rowid]: if ele == label[rowid]: num_correct += 1 diff --git a/python/paddle/fluid/tests/unittests/test_activation_op.py b/python/paddle/fluid/tests/unittests/test_activation_op.py index 5ed387fb1247f1a91147cb6981f1adc7c2eeb8a2..34f9cf0620fd1351111e93e16ed5f7e765d7078b 100644 --- a/python/paddle/fluid/tests/unittests/test_activation_op.py +++ b/python/paddle/fluid/tests/unittests/test_activation_op.py @@ -313,9 +313,9 @@ class TestAbs(OpTest): self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) - # Because we set delta = 0.005 in caculating numeric gradient, + # Because we set delta = 0.005 in calculating numeric gradient, # if x is too small, such as 0.002, x_neg will be -0.003 - # x_pos will be 0.007, so the numeric gradient is unaccurate. + # x_pos will be 0.007, so the numeric gradient is inaccurate. # we should avoid this x[np.abs(x) < 0.005] = 0.02 out = np.abs(x) diff --git a/python/paddle/fluid/tests/unittests/test_adam_op.py b/python/paddle/fluid/tests/unittests/test_adam_op.py index 3c65f3d44adcebdca92f78f7834d4878a9fa3dfe..fa4b39879c0ede569b6802502b2c71a93b163373 100644 --- a/python/paddle/fluid/tests/unittests/test_adam_op.py +++ b/python/paddle/fluid/tests/unittests/test_adam_op.py @@ -273,7 +273,7 @@ class TestSparseAdamOp(unittest.TestCase): self.setup(scope, place) op_args = dict() - for key, np_array in self.dense_inputs.iteritems(): + for key, np_array in self.dense_inputs.items(): var = scope.var(key).get_tensor() var.set(np_array, place) op_args[key] = key @@ -290,7 +290,7 @@ class TestSparseAdamOp(unittest.TestCase): adam_op = Operator("adam", **op_args) adam_op.run(scope, place) - for key, np_array in self.outputs.iteritems(): + for key, np_array in self.outputs.items(): out_var = scope.var(key).get_tensor() actual = np.array(out_var) actual = actual.reshape([actual.size]) diff --git a/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py b/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py new file mode 100644 index 0000000000000000000000000000000000000000..9c7d5d41f0c512a9fb609dce304c1eed929d28b5 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py @@ -0,0 +1,110 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://w_idxw.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import sys +import math +from op_test import OpTest + + +def anchor_generator_in_python(input_feat, anchor_sizes, aspect_ratios, + variances, stride, offset): + num_anchors = len(aspect_ratios) * len(anchor_sizes) + layer_h = input_feat.shape[2] + layer_w = input_feat.shape[3] + out_dim = (layer_h, layer_w, num_anchors, 4) + out_anchors = np.zeros(out_dim).astype('float32') + + for h_idx in range(layer_h): + for w_idx in range(layer_w): + x_ctr = (w_idx * stride[0]) + offset * (stride[0] - 1) + y_ctr = (h_idx * stride[1]) + offset * (stride[1] - 1) + idx = 0 + for r in range(len(aspect_ratios)): + ar = aspect_ratios[r] + for s in range(len(anchor_sizes)): + anchor_size = anchor_sizes[s] + area = stride[0] * stride[1] + area_ratios = area / ar + base_w = np.round(np.sqrt(area_ratios)) + base_h = np.round(base_w * ar) + scale_w = anchor_size / stride[0] + scale_h = anchor_size / stride[1] + w = scale_w * base_w + h = scale_h * base_h + out_anchors[h_idx, w_idx, idx, :] = [ + (x_ctr - 0.5 * (w - 1)), (y_ctr - 0.5 * (h - 1)), + (x_ctr + 0.5 * (w - 1)), (y_ctr + 0.5 * (h - 1)) + ] + idx += 1 + + # set the variance. + out_var = np.tile(variances, (layer_h, layer_w, num_anchors, 1)) + out_anchors = out_anchors.astype('float32') + out_var = out_var.astype('float32') + return out_anchors, out_var + + +class TestAnchorGeneratorOp(OpTest): + def set_data(self): + self.init_test_params() + self.init_test_input() + self.init_test_output() + self.inputs = {'Input': self.input} + + self.attrs = { + 'anchor_sizes': self.anchor_sizes, + 'aspect_ratios': self.aspect_ratios, + 'stride': self.stride, + 'offset': self.offset, + 'variances': self.variances, + } + + self.outputs = {'Anchors': self.out_anchors, 'Variances': self.out_var} + + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "anchor_generator" + self.set_data() + + def init_test_params(self): + self.batch_size = 1 + self.input_channels = 2 + self.layer_h = 2 + self.layer_w = 2 + + self.anchor_sizes = [64., 128., 256., 512.] + self.aspect_ratios = [0.5, 1., 2.] + self.stride = [16., 16.] + + self.offset = 0.5 + + self.variances = [0.1, 0.1, 0.2, 0.2] + + def init_test_input(self): + self.input = np.random.random( + (self.batch_size, self.input_channels, self.layer_h, + self.layer_w)).astype('float32') + + def init_test_output(self): + self.out_anchors, self.out_var = anchor_generator_in_python( + self.input, self.anchor_sizes, self.aspect_ratios, self.variances, + self.stride, self.offset) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py b/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py new file mode 100644 index 0000000000000000000000000000000000000000..e04412f809cdd75d07d28a60f0c2f19041a684f6 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py @@ -0,0 +1,82 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class BaseTestCase(OpTest): + def initTestCase(self): + self.op_type = 'arg_min' + self.dims = (3, 4, 5) + self.dtype = 'float32' + self.axis = 0 + + def setUp(self): + self.initTestCase() + self.x = (1000 * np.random.random(self.dims)).astype(self.dtype) + self.inputs = {'X': self.x} + self.attrs = {'axis': self.axis} + if self.op_type == "arg_min": + self.outputs = {'Out': np.argmin(self.x, axis=self.axis)} + else: + self.outputs = {'Out': np.argmax(self.x, axis=self.axis)} + + def test_check_output(self): + self.check_output() + + +class TestCase0(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_max' + self.dims = (3, 4, 5) + self.dtype = 'float32' + self.axis = 0 + + +class TestCase1(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_min' + self.dims = (3, 4) + self.dtype = 'float64' + self.axis = 1 + + +class TestCase2(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_max' + self.dims = (3, 4) + self.dtype = 'int64' + self.axis = 0 + + +class TestCase3(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_max' + self.dims = (3, ) + self.dtype = 'int64' + self.axis = 0 + + +class TestCase4(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_min' + self.dims = (1, ) + self.dtype = 'int32' + self.axis = 0 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_argsort_op.py b/python/paddle/fluid/tests/unittests/test_argsort_op.py new file mode 100644 index 0000000000000000000000000000000000000000..b29a102a3880406156481fdac54ca7043d3415db --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_argsort_op.py @@ -0,0 +1,56 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestArgsortOp(OpTest): + def setUp(self): + self.init_axis() + x = np.random.random((2, 3, 4, 5, 10)).astype("float32") + if self.axis < 0: + self.axis = self.axis + len(x.shape) + self.indices = np.argsort(x, kind='quicksort', axis=self.axis) + self.out = np.sort(x, kind='quicksort', axis=self.axis) + self.op_type = "argsort" + self.inputs = {'X': x} + self.attrs = {'axis': self.axis} + self.outputs = {'Indices': self.indices, 'Out': self.out} + + def init_axis(self): + self.axis = -1 + + def test_check_output(self): + self.check_output() + + +class TestArgsortOpAxis0(TestArgsortOp): + def init_axis(self): + self.axis = 0 + + +class TestArgsortOpAxis1(TestArgsortOp): + def init_axis(self): + self.axis = 1 + + +class TestArgsortOpAxisNeg2(TestArgsortOp): + def init_axis(self): + self.axis = -2 + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_array_read_write_op.py b/python/paddle/fluid/tests/unittests/test_array_read_write_op.py index a49e9035a43e04fc1d1b2328d7562c053320b24b..0000fb0958a129e9e1098de1fad888c503cfbdc5 100644 --- a/python/paddle/fluid/tests/unittests/test_array_read_write_op.py +++ b/python/paddle/fluid/tests/unittests/test_array_read_write_op.py @@ -80,8 +80,9 @@ class TestArrayReadWrite(unittest.TestCase): append_backward(total_sum_scaled) - g_vars = map(default_main_program().global_block().var, - [each_x.name + "@GRAD" for each_x in x]) + g_vars = list( + map(default_main_program().global_block().var, + [each_x.name + "@GRAD" for each_x in x])) g_out = [ item.sum() for item in exe.run( diff --git a/python/paddle/fluid/tests/unittests/test_auc_op.py b/python/paddle/fluid/tests/unittests/test_auc_op.py index 948836039be48ad74d5556100f06231bb89f26d3..6580c70ca68c4ba24919f03d071f6f88fb68953c 100644 --- a/python/paddle/fluid/tests/unittests/test_auc_op.py +++ b/python/paddle/fluid/tests/unittests/test_auc_op.py @@ -15,63 +15,42 @@ import unittest import numpy as np from op_test import OpTest +from paddle.fluid import metrics class TestAucOp(OpTest): def setUp(self): self.op_type = "auc" pred = np.random.random((128, 2)).astype("float32") - indices = np.random.randint(0, 2, (128, 2)) labels = np.random.randint(0, 2, (128, 1)) num_thresholds = 200 - self.inputs = {'Out': pred, 'Indices': indices, 'Label': labels} + tp = np.zeros((num_thresholds, )).astype("int64") + tn = np.zeros((num_thresholds, )).astype("int64") + fp = np.zeros((num_thresholds, )).astype("int64") + fn = np.zeros((num_thresholds, )).astype("int64") + + self.inputs = { + 'Predict': pred, + 'Label': labels, + 'TP': tp, + 'TN': tn, + 'FP': fp, + 'FN': fn + } self.attrs = {'curve': 'ROC', 'num_thresholds': num_thresholds} - # NOTE: sklearn use a different way to generate thresholds - # which will cause the result differs slightly: - # from sklearn.metrics import roc_curve, auc - # fpr, tpr, thresholds = roc_curve(labels, pred) - # auc_value = auc(fpr, tpr) - # we caculate AUC again using numpy for testing - kepsilon = 1e-7 # to account for floating point imprecisions - thresholds = [(i + 1) * 1.0 / (num_thresholds - 1) - for i in range(num_thresholds - 2)] - thresholds = [0.0 - kepsilon] + thresholds + [1.0 + kepsilon] - # caculate TP, FN, TN, FP count - tp_list = np.ndarray((num_thresholds, )) - fn_list = np.ndarray((num_thresholds, )) - tn_list = np.ndarray((num_thresholds, )) - fp_list = np.ndarray((num_thresholds, )) - for idx_thresh, thresh in enumerate(thresholds): - tp, fn, tn, fp = 0, 0, 0, 0 - for i, lbl in enumerate(labels): - if lbl: - if pred[i, 0] >= thresh: - tp += 1 - else: - fn += 1 - else: - if pred[i, 0] >= thresh: - fp += 1 - else: - tn += 1 - tp_list[idx_thresh] = tp - fn_list[idx_thresh] = fn - tn_list[idx_thresh] = tn - fp_list[idx_thresh] = fp - - epsilon = 1e-6 - tpr = (tp_list.astype("float32") + epsilon) / ( - tp_list + fn_list + epsilon) - fpr = fp_list.astype("float32") / (fp_list + tn_list + epsilon) - rec = (tp_list.astype("float32") + epsilon) / ( - tp_list + fp_list + epsilon) - - x = fpr[:num_thresholds - 1] - fpr[1:] - y = (tpr[:num_thresholds - 1] + tpr[1:]) / 2.0 - auc_value = np.sum(x * y) - - self.outputs = {'AUC': auc_value} + python_auc = metrics.Auc(name="auc", + curve='ROC', + num_thresholds=num_thresholds) + python_auc.update(pred, labels) + + self.outputs = { + 'AUC': python_auc.eval(), + 'TPOut': python_auc.tp_list, + 'FNOut': python_auc.fn_list, + 'TNOut': python_auc.tn_list, + 'FPOut': python_auc.fp_list + } def test_check_output(self): self.check_output() diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py index f6097d4b846e8da1c4ee3cc49b31f9873660056d..18fa5461590134d2032a29e40699109c12092c6d 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py @@ -52,5 +52,17 @@ class TestMKLDNNBatchNormOpInference(TestBatchNormOpInference): self.check_with_place(place, data_format, self.dtype, [2, 3, 4, 5]) +class TestMKLDNNBatchNormOpWithReluInference(TestBatchNormOpInference): + def init_kernel_type(self): + self.use_mkldnn = True + self.fuse_with_relu = True + + def test_check_output(self): + place = core.CPUPlace() + data_format = "NCHW" + + self.check_with_place(place, data_format, self.dtype, [2, 3, 4, 5]) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py index 4216d83653b27ec7f18034e576fbedbecc3f1cfe..f805fdc35f624bf6e9d94d66839dcb2a0143a29b 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py @@ -128,8 +128,7 @@ def create_or_get_tensor(scope, var_name, var, place): tensor = scope.var(var_name).get_tensor() if var is not None: assert isinstance(var, np.ndarray) - tensor.set_lod([[]]) - tensor.set_dims(var.shape) + tensor.set_recursive_sequence_lengths([]) tensor.set(var, place) return tensor @@ -159,6 +158,7 @@ class TestBatchNormOpInference(unittest.TestCase): def setUp(self): self.dtype = np.float32 self.use_mkldnn = False + self.fuse_with_relu = False self.init_kernel_type() def __assert_close(self, tensor, np_array, msg, atol=1e-4): @@ -180,6 +180,8 @@ class TestBatchNormOpInference(unittest.TestCase): scale_shape = [c] x_val = np.random.random_sample(x_shape).astype(dtype) + # generate some negative values to test case with relu fused + x_val = x_val - 0.5 scale_val = np.random.random_sample(scale_shape).astype(np.float32) bias_val = np.random.random_sample(scale_shape).astype(np.float32) @@ -188,6 +190,8 @@ class TestBatchNormOpInference(unittest.TestCase): y_out = _reference_testing(x_val, scale_val, bias_val, mean, variance, epsilon, data_layout).astype(dtype) + if self.fuse_with_relu: + y_out = np.maximum(y_out, 0) scope = core.Scope() @@ -233,6 +237,7 @@ class TestBatchNormOpInference(unittest.TestCase): is_test=True, data_layout=data_layout, use_mkldnn=self.use_mkldnn, + fuse_with_relu=self.fuse_with_relu, epsilon=epsilon) batch_norm_op.run(scope, place) @@ -265,6 +270,7 @@ class TestFP16BatchNormOpInference(TestBatchNormOpInference): def setUp(self): self.dtype = np.float16 self.use_mkldnn = False + self.fuse_with_relu = False self.init_kernel_type() def test_check_output(self): @@ -284,6 +290,7 @@ class TestFP16BatchNormOpInference(TestBatchNormOpInference): class TestBatchNormOpTraining(unittest.TestCase): def setUp(self): self.use_mkldnn = False + self.fuse_with_relu = False self.data_formats = ["NCHW", "NHWC"] self.init_kernel_type() @@ -367,7 +374,8 @@ class TestBatchNormOpTraining(unittest.TestCase): "epsilon": epsilon, "is_test": False, "data_layout": data_layout, - "use_mkldnn": self.use_mkldnn + "use_mkldnn": self.use_mkldnn, + "fuse_with_relu": self.fuse_with_relu }) block.create_var(name='y@GRAD', dtype='float32', shape=y.shape) @@ -407,7 +415,7 @@ class TestBatchNormOpTraining(unittest.TestCase): self.__assert_close(scale_grad, out[6], "scale_grad") self.__assert_close(bias_grad, out[7], "bias_grad") - print "op test forward passed: ", str(place), data_layout + print("op test forward passed: ", str(place), data_layout) places = [core.CPUPlace()] diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py index 7976dd7c3f14390fb00bc8ab39121b6a686e3039..4a3ac2a31e072eb1a15af31f558cf9f626a7ac53 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py @@ -20,6 +20,8 @@ from paddle.fluid.op import Operator class TestBeamSearchDecodeOp(unittest.TestCase): + """unittest of beam_search_decode_op""" + def setUp(self): self.scope = core.Scope() self.place = core.CPUPlace() @@ -32,32 +34,44 @@ class TestBeamSearchDecodeOp(unittest.TestCase): def test_get_set(self): ids = self.scope.var("ids").get_lod_tensor_array() - self.append_lod_tensor( - ids, [[0, 3, 6], [0, 1, 2, 3, 4, 5, 6]], - np.array( - [1, 2, 3, 4, 5, 6], dtype="int64")) - self.append_lod_tensor( - ids, [[0, 3, 6], [0, 1, 1, 3, 5, 5, 6]], - np.array( - [0, 1, 2, 3, 4, 5], dtype="int64")) - self.append_lod_tensor( - ids, [[0, 3, 6], [0, 0, 1, 2, 3, 4, 5]], - np.array( - [0, 1, 2, 3, 4], dtype="int64")) - scores = self.scope.var("scores").get_lod_tensor_array() - self.append_lod_tensor( - scores, [[0, 3, 6], [0, 1, 2, 3, 4, 5, 6]], - np.array( - [1, 2, 3, 4, 5, 6], dtype="float64")) - self.append_lod_tensor( - scores, [[0, 3, 6], [0, 1, 1, 3, 5, 5, 6]], - np.array( - [0, 1, 2, 3, 4, 5], dtype="float64")) - self.append_lod_tensor( - scores, [[0, 3, 6], [0, 0, 1, 2, 3, 4, 5]], - np.array( - [0, 1, 2, 3, 4], dtype="float64")) + # Construct sample data with 5 steps and 2 source sentences + # beam_size = 2, end_id = 1 + # start with start_id + [ + self.append_lod_tensor( + array, [[0, 1, 2], [0, 1, 2]], np.array( + [0, 0], dtype=dtype)) + for array, dtype in ((ids, "int64"), (scores, "float32")) + ] + [ + self.append_lod_tensor( + array, [[0, 1, 2], [0, 2, 4]], + np.array( + [2, 3, 4, 5], dtype=dtype)) + for array, dtype in ((ids, "int64"), (scores, "float32")) + ] + [ + self.append_lod_tensor( + array, [[0, 2, 4], [0, 2, 2, 4, 4]], + np.array( + [3, 1, 5, 4], dtype=dtype)) + for array, dtype in ((ids, "int64"), (scores, "float32")) + ] + [ + self.append_lod_tensor( + array, [[0, 2, 4], [0, 1, 2, 3, 4]], + np.array( + [1, 1, 3, 5], dtype=dtype)) + for array, dtype in ((ids, "int64"), (scores, "float32")) + ] + [ + self.append_lod_tensor( + array, [[0, 2, 4], [0, 0, 0, 2, 2]], + np.array( + [5, 1], dtype=dtype)) + for array, dtype in ((ids, "int64"), (scores, "float32")) + ] sentence_ids = self.scope.var("sentence_ids").get_tensor() sentence_scores = self.scope.var("sentence_scores").get_tensor() @@ -69,21 +83,25 @@ class TestBeamSearchDecodeOp(unittest.TestCase): Scores="scores", # outputs SentenceIds="sentence_ids", - SentenceScores="sentence_scores") + SentenceScores="sentence_scores", + beam_size=2, + end_id=1, ) beam_search_decode_op.run(self.scope, self.place) - expected_lod = [[0, 4, 8], [0, 1, 3, 6, 9, 10, 13, 16, 19]] + expected_lod = [[0, 2, 4], [0, 4, 7, 12, 17]] self.assertEqual(sentence_ids.lod(), expected_lod) self.assertEqual(sentence_scores.lod(), expected_lod) expected_data = np.array( - [2, 1, 0, 3, 1, 0, 3, 2, 1, 5, 4, 3, 2, 4, 4, 3, 6, 5, 4], "int64") + [0, 2, 3, 1, 0, 2, 1, 0, 4, 5, 3, 5, 0, 4, 5, 3, 1], "int64") self.assertTrue(np.array_equal(np.array(sentence_ids), expected_data)) self.assertTrue( np.array_equal(np.array(sentence_scores), expected_data)) +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestBeamSearchDecodeOpGPU(TestBeamSearchDecodeOp): def setUp(self): self.scope = core.Scope() diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_op.py index bc708f3aff54f54d290684d68afa503a50a32dac..e8283fc9422d93af5735aaec1a165b46ac1ef78e 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_op.py @@ -26,9 +26,12 @@ def create_tensor(scope, name, np_data): class BeamSearchOpTester(unittest.TestCase): + """unittest of beam_search_op""" + def setUp(self): self.scope = core.Scope() self._create_ids() + self._create_pre_scores() self._create_scores() self._create_pre_ids() self.scope.var('selected_ids') @@ -37,7 +40,8 @@ class BeamSearchOpTester(unittest.TestCase): def test_run(self): op = Operator( 'beam_search', - pre_ids="pre_ids", + pre_ids='pre_ids', + pre_scores='pre_scores', ids='ids', scores='scores', selected_ids='selected_ids', @@ -47,15 +51,26 @@ class BeamSearchOpTester(unittest.TestCase): end_id=0, ) op.run(self.scope, core.CPUPlace()) selected_ids = self.scope.find_var("selected_ids").get_tensor() - print 'selected_ids', np.array(selected_ids) - print 'lod', selected_ids.lod() + selected_scores = self.scope.find_var("selected_scores").get_tensor() + self.assertTrue( + np.allclose( + np.array(selected_ids), np.array([4, 2, 3, 8])[:, np.newaxis])) + self.assertTrue( + np.allclose( + np.array(selected_scores), + np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis])) + self.assertEqual(selected_ids.lod(), [[0, 2, 4], [0, 1, 2, 3, 4]]) def _create_pre_ids(self): np_data = np.array([[1, 2, 3, 4]], dtype='int64') - tensor = create_tensor(self.scope, "pre_ids", np_data) + tensor = create_tensor(self.scope, 'pre_ids', np_data) + + def _create_pre_scores(self): + np_data = np.array([[0.1, 0.2, 0.3, 0.4]], dtype='float32') + tensor = create_tensor(self.scope, 'pre_scores', np_data) def _create_ids(self): - self.lod = [[0, 1, 4], [0, 1, 2, 3, 4]] + self.lod = [[0, 2, 4], [0, 1, 2, 3, 4]] np_data = np.array( [[4, 2, 5], [2, 1, 3], [3, 5, 2], [8, 2, 1]], dtype='int64') tensor = create_tensor(self.scope, "ids", np_data) diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py index bffb4f3b666a7ddcc133b7c30fab132b49aa1d0e..b04f25ef874cc6204211a4f5f5991a0ec8c473dd 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -15,9 +15,13 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core -def bilinear_interp_np(input, out_h, out_w): +def bilinear_interp_np(input, out_h, out_w, out_size): + if out_size is not None: + out_h = out_size[0] + out_w = out_size[1] batch_size, channel, in_h, in_w = input.shape if out_h > 1: ratio_h = (in_h - 1.0) / (out_h - 1.0) @@ -42,19 +46,22 @@ def bilinear_interp_np(input, out_h, out_w): out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] + w1lambda*input[:, :, h, w+wid]) + \ - h1lambda*(w2lambda*input[:, :, h+hid, w] + - w1lambda*input[:, :, h+hid, w+wid]) - return out.astype("float32") + h1lambda*(w2lambda*input[:, :, h+hid, w] + + w1lambda*input[:, :, h+hid, w+wid]) + return out.astype(input.dtype) class TestBilinearInterpOp(OpTest): def setUp(self): + self.out_size = None self.init_test_case() self.op_type = "bilinear_interp" input_np = np.random.random(self.input_shape).astype("float32") - output_np = bilinear_interp_np(input_np, self.out_h, self.out_w) - + output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, + self.out_size) self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} self.outputs = {'Out': output_np} @@ -68,6 +75,7 @@ class TestBilinearInterpOp(OpTest): self.input_shape = [2, 3, 4, 4] self.out_h = 2 self.out_w = 2 + self.out_size = np.array([3, 3]).astype("int32") class TestCase1(TestBilinearInterpOp): @@ -91,5 +99,68 @@ class TestCase3(TestBilinearInterpOp): self.out_w = 128 +class TestCase4(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + self.out_size = np.array([2, 2]).astype("int32") + + +class TestCase5(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + self.out_size = np.array([11, 11]).astype("int32") + + +class TestCase6(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + self.out_size = np.array([65, 129]).astype("int32") + + +class TestBilinearInterpOpUint8(OpTest): + def setUp(self): + self.out_size = None + self.init_test_case() + self.op_type = "bilinear_interp" + input_np = np.random.randint( + low=0, high=256, size=self.input_shape).astype("uint8") + output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, + self.out_size) + self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size + self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} + self.outputs = {'Out': output_np} + + def test_check_output(self): + self.check_output_with_place(place=core.CPUPlace(), atol=1) + + def init_test_case(self): + self.input_shape = [1, 3, 9, 6] + self.out_h = 10 + self.out_w = 9 + + +class TestCase1Uint8(TestBilinearInterpOpUint8): + def init_test_case(self): + self.input_shape = [2, 3, 128, 64] + self.out_h = 120 + self.out_w = 50 + + +class TestCase2Uint8(TestBilinearInterpOpUint8): + def init_test_case(self): + self.input_shape = [4, 1, 7, 8] + self.out_h = 5 + self.out_w = 13 + self.out_size = np.array([6, 15]).astype("int32") + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py b/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py index f7461ee6dab699064153332116449c8e20a0bac0..ceeca25b74d85ed2874d672e402e3186c4ce7d47 100644 --- a/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py +++ b/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py @@ -48,7 +48,7 @@ def bipartite_match(distance, match_indices, match_dist): def argmax_match(distance, match_indices, match_dist, threshold): r, c = distance.shape - for j in xrange(c): + for j in range(c): if match_indices[j] != -1: continue col_dist = distance[:, j] @@ -65,23 +65,25 @@ def batch_bipartite_match(distance, lod, match_type=None, dist_threshold=None): distance (numpy.array) : The distance of two entries with shape [M, N]. lod (list of int): The offsets of each input in this batch. """ - n = len(lod) - 1 + n = len(lod) m = distance.shape[1] match_indices = -1 * np.ones((n, m), dtype=np.int) match_dist = np.zeros((n, m), dtype=np.float32) - for i in range(len(lod) - 1): - bipartite_match(distance[lod[i]:lod[i + 1], :], match_indices[i, :], - match_dist[i, :]) + cur_offset = 0 + for i in range(n): + bipartite_match(distance[cur_offset:(cur_offset + lod[i]), :], + match_indices[i, :], match_dist[i, :]) if match_type == 'per_prediction': - argmax_match(distance[lod[i]:lod[i + 1], :], match_indices[i, :], - match_dist[i, :], dist_threshold) + argmax_match(distance[cur_offset:(cur_offset + lod[i]), :], + match_indices[i, :], match_dist[i, :], dist_threshold) + cur_offset += lod[i] return match_indices, match_dist class TestBipartiteMatchOpWithLoD(OpTest): def setUp(self): self.op_type = 'bipartite_match' - lod = [[0, 5, 11, 23]] + lod = [[5, 6, 12]] dist = np.random.random((23, 217)).astype('float32') match_indices, match_dist = batch_bipartite_match(dist, lod[0]) @@ -98,7 +100,7 @@ class TestBipartiteMatchOpWithLoD(OpTest): class TestBipartiteMatchOpWithoutLoD(OpTest): def setUp(self): self.op_type = 'bipartite_match' - lod = [[0, 8]] + lod = [[8]] dist = np.random.random((8, 17)).astype('float32') match_indices, match_dist = batch_bipartite_match(dist, lod[0]) @@ -112,10 +114,27 @@ class TestBipartiteMatchOpWithoutLoD(OpTest): self.check_output() +class TestBipartiteMatchOpWithoutLoDLargeScaleInput(OpTest): + def setUp(self): + self.op_type = 'bipartite_match' + lod = [[300]] + dist = np.random.random((300, 17)).astype('float32') + match_indices, match_dist = batch_bipartite_match(dist, lod[0]) + + self.inputs = {'DistMat': dist} + self.outputs = { + 'ColToRowMatchIndices': match_indices, + 'ColToRowMatchDist': match_dist, + } + + def test_check_output(self): + self.check_output() + + class TestBipartiteMatchOpWithPerPredictionType(OpTest): def setUp(self): self.op_type = 'bipartite_match' - lod = [[0, 5, 11, 23]] + lod = [[5, 6, 12]] dist = np.random.random((23, 237)).astype('float32') match_indices, match_dist = batch_bipartite_match(dist, lod[0], 'per_prediction', 0.5) diff --git a/python/paddle/fluid/tests/unittests/test_box_coder_op.py b/python/paddle/fluid/tests/unittests/test_box_coder_op.py index 56f5af91d8e58086c12fde6948229675569aa271..4ce9a4783e2332b6882164a70e1462c6a6d31bef 100644 --- a/python/paddle/fluid/tests/unittests/test_box_coder_op.py +++ b/python/paddle/fluid/tests/unittests/test_box_coder_op.py @@ -19,7 +19,8 @@ import math from op_test import OpTest -def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): +def box_coder(target_box, prior_box, prior_box_var, output_box, code_type, + box_normalized): prior_box_x = ( (prior_box[:, 2] + prior_box[:, 0]) / 2).reshape(1, prior_box.shape[0]) prior_box_y = ( @@ -30,6 +31,9 @@ def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): (prior_box[:, 3] - prior_box[:, 1])).reshape(1, prior_box.shape[0]) prior_box_var = prior_box_var.reshape(1, prior_box_var.shape[0], prior_box_var.shape[1]) + if not box_normalized: + prior_box_height = prior_box_height + 1 + prior_box_width = prior_box_width + 1 if (code_type == "EncodeCenterSize"): target_box_x = ((target_box[:, 2] + target_box[:, 0]) / 2).reshape( @@ -40,6 +44,9 @@ def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): target_box.shape[0], 1) target_box_height = ((target_box[:, 3] - target_box[:, 1])).reshape( target_box.shape[0], 1) + if not box_normalized: + target_box_height = target_box_height + 1 + target_box_width = target_box_width + 1 output_box[:,:,0] = (target_box_x - prior_box_x) / prior_box_width / \ prior_box_var[:,:,0] @@ -64,21 +71,29 @@ def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): output_box[:, :, 1] = target_box_y - target_box_height / 2 output_box[:, :, 2] = target_box_x + target_box_width / 2 output_box[:, :, 3] = target_box_y + target_box_height / 2 + if not box_normalized: + output_box[:, :, 2] = output_box[:, :, 2] - 1 + output_box[:, :, 3] = output_box[:, :, 3] - 1 -def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type): +def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type, + box_normalized): n = target_box.shape[0] m = prior_box.shape[0] output_box = np.zeros((n, m, 4), dtype=np.float32) - for i in range(len(lod) - 1): + cur_offset = 0 + for i in range(len(lod)): if (code_type == "EncodeCenterSize"): - box_coder(target_box[lod[i]:lod[i + 1], :], prior_box, - prior_box_var, output_box[lod[i]:lod[i + 1], :, :], - code_type) + box_coder(target_box[cur_offset:(cur_offset + lod[i]), :], + prior_box, prior_box_var, + output_box[cur_offset:(cur_offset + lod[i]), :, :], + code_type, box_normalized) elif (code_type == "DecodeCenterSize"): - box_coder(target_box[lod[i]:lod[i + 1], :, :], prior_box, - prior_box_var, output_box[lod[i]:lod[i + 1], :, :], - code_type) + box_coder(target_box[cur_offset:(cur_offset + lod[i]), :, :], + prior_box, prior_box_var, + output_box[cur_offset:(cur_offset + lod[i]), :, :], + code_type, box_normalized) + cur_offset += lod[i] return output_box @@ -88,20 +103,50 @@ class TestBoxCoderOp(OpTest): def setUp(self): self.op_type = "box_coder" - lod = [[0, 1, 2, 3, 4, 5]] + lod = [[1, 1, 1, 1, 1]] prior_box = np.random.random((10, 4)).astype('float32') prior_box_var = np.random.random((10, 4)).astype('float32') target_box = np.random.random((5, 10, 4)).astype('float32') code_type = "DecodeCenterSize" + box_normalized = False output_box = batch_box_coder(prior_box, prior_box_var, target_box, - lod[0], code_type) + lod[0], code_type, box_normalized) self.inputs = { 'PriorBox': prior_box, 'PriorBoxVar': prior_box_var, 'TargetBox': target_box, } - self.attrs = {'code_type': 'decode_center_size'} + self.attrs = { + 'code_type': 'decode_center_size', + 'box_normalized': False + } + self.outputs = {'OutputBox': output_box} + + +class TestBoxCoderOpWithoutBoxVar(OpTest): + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "box_coder" + lod = [[0, 1, 2, 3, 4, 5]] + prior_box = np.random.random((10, 4)).astype('float32') + prior_box_var = np.ones((10, 4)).astype('float32') + target_box = np.random.random((5, 10, 4)).astype('float32') + code_type = "DecodeCenterSize" + box_normalized = False + output_box = batch_box_coder(prior_box, prior_box_var, target_box, + lod[0], code_type, box_normalized) + + self.inputs = { + 'PriorBox': prior_box, + 'TargetBox': target_box, + } + self.attrs = { + 'code_type': 'decode_center_size', + 'box_normalized': False + } self.outputs = {'OutputBox': output_box} @@ -111,20 +156,21 @@ class TestBoxCoderOpWithLoD(OpTest): def setUp(self): self.op_type = "box_coder" - lod = [[0, 4, 12, 20]] + lod = [[4, 8, 8]] prior_box = np.random.random((10, 4)).astype('float32') prior_box_var = np.random.random((10, 4)).astype('float32') target_box = np.random.random((20, 4)).astype('float32') code_type = "EncodeCenterSize" + box_normalized = True output_box = batch_box_coder(prior_box, prior_box_var, target_box, - lod[0], code_type) + lod[0], code_type, box_normalized) self.inputs = { 'PriorBox': prior_box, 'PriorBoxVar': prior_box_var, 'TargetBox': (target_box, lod), } - self.attrs = {'code_type': 'encode_center_size'} + self.attrs = {'code_type': 'encode_center_size', 'box_normalized': True} self.outputs = {'OutputBox': output_box} diff --git a/python/paddle/fluid/tests/unittests/test_calc_gradient.py b/python/paddle/fluid/tests/unittests/test_calc_gradient.py index 06e676cd83e77549afd679e730426c590cc046bf..7f2a9e6971ed933463216e38498d48ab132a1a37 100644 --- a/python/paddle/fluid/tests/unittests/test_calc_gradient.py +++ b/python/paddle/fluid/tests/unittests/test_calc_gradient.py @@ -16,8 +16,6 @@ import unittest import paddle.fluid as fluid import paddle.fluid.layers as layers -import paddle.fluid.framework as framework -import paddle.fluid.optimizer as optimizer from paddle.fluid.backward import calc_gradient diff --git a/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py b/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py index 050df2801c98e8f4167cdd1b4dde858c9f9f07dd..354110f1f96f6b4aad1a4866c8d1337dec3acd16 100644 --- a/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py +++ b/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py @@ -63,7 +63,7 @@ class TestChunkEvalOp(OpTest): # generate chunk beginnings chunk_begins = sorted( np.random.choice( - range(starts[-1]), num_chunks, replace=False)) + list(range(starts[-1])), num_chunks, replace=False)) seq_chunk_begins = [] begin_idx = 0 # divide chunks into sequences @@ -93,7 +93,7 @@ class TestChunkEvalOp(OpTest): self.num_infer_chunks + self.num_label_chunks - self.num_correct_chunks) correct_chunks = np.random.choice( - range(len(chunks)), self.num_correct_chunks, replace=False) + list(range(len(chunks))), self.num_correct_chunks, replace=False) infer_chunks = np.random.choice( [x for x in range(len(chunks)) if x not in correct_chunks], self.num_infer_chunks - self.num_correct_chunks, @@ -138,16 +138,17 @@ class TestChunkEvalOp(OpTest): infer.fill(self.num_chunk_types * self.num_tag_types) label = np.copy(infer) starts = np.random.choice( - range(1, self.batch_size), self.num_sequences - 1, + list(range(1, self.batch_size)), + self.num_sequences - 1, replace=False).tolist() starts.extend([0, self.batch_size]) starts = sorted(starts) self.num_correct_chunks, self.num_infer_chunks, self.num_label_chunks = self.gen_chunks( infer, label, starts) - self.inputs = { - 'Inference': (infer, [starts]), - 'Label': (label, [starts]) - } + lod = [] + for i in range(len(starts) - 1): + lod.append(starts[i + 1] - starts[i]) + self.inputs = {'Inference': (infer, [lod]), 'Label': (label, [lod])} precision = float( self.num_correct_chunks ) / self.num_infer_chunks if self.num_infer_chunks else 0 diff --git a/python/paddle/fluid/tests/unittests/test_concat_op.py b/python/paddle/fluid/tests/unittests/test_concat_op.py index 1e00d67d5480bfa77a60e1aed52cafac6e8242ca..e9f3c45dc40b3333fe7304f8e4313d156bd5374c 100644 --- a/python/paddle/fluid/tests/unittests/test_concat_op.py +++ b/python/paddle/fluid/tests/unittests/test_concat_op.py @@ -43,7 +43,7 @@ class TestConcatOp(OpTest): self.axis = 1 -class TestConcatOp2(OpTest): +class TestConcatOp2(TestConcatOp): def init_test_data(self): self.x0 = np.random.random((2, 3, 4, 5)).astype('float32') self.x1 = np.random.random((2, 3, 4, 5)).astype('float32') @@ -51,5 +51,16 @@ class TestConcatOp2(OpTest): self.axis = 1 +class TestConcatOp3(TestConcatOp): + def init_test_data(self): + self.x0 = np.random.random((1, 256, 170, 256)).astype('float32') + self.x1 = np.random.random((1, 128, 170, 256)).astype('float32') + self.x2 = np.random.random((1, 128, 170, 256)).astype('float32') + self.axis = 1 + + def test_check_grad(self): + pass + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_conditional_block.py b/python/paddle/fluid/tests/unittests/test_conditional_block.py index 084b8d37386fac0366c190f5f30dd39467072498..77869a1242e08d348bfb1031b8f5b1ab5c81d868 100644 --- a/python/paddle/fluid/tests/unittests/test_conditional_block.py +++ b/python/paddle/fluid/tests/unittests/test_conditional_block.py @@ -18,14 +18,15 @@ import paddle.fluid.core as core from paddle.fluid.framework import default_startup_program, default_main_program from paddle.fluid.executor import Executor from paddle.fluid.backward import append_backward +from paddle.fluid.layers.control_flow import ConditionalBlock import numpy -class ConditionalBlock(unittest.TestCase): +class ConditionalBlockTest(unittest.TestCase): def test_forward(self): data = layers.data(name='X', shape=[1], dtype='float32') data.stop_gradient = False - cond = layers.ConditionalBlock(inputs=[data]) + cond = ConditionalBlock(inputs=[data]) out = layers.create_tensor(dtype='float32') with cond.block(): hidden = layers.fc(input=data, size=10) @@ -38,7 +39,7 @@ class ConditionalBlock(unittest.TestCase): x = numpy.random.random(size=(10, 1)).astype('float32') outs = exe.run(feed={'X': x}, fetch_list=[out])[0] - print outs + print(outs) loss = layers.mean(out) append_backward(loss=loss) outs = exe.run( @@ -46,7 +47,7 @@ class ConditionalBlock(unittest.TestCase): fetch_list=[ default_main_program().block(0).var(data.name + "@GRAD") ])[0] - print outs + print(outs) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_const_value.py b/python/paddle/fluid/tests/unittests/test_const_value.py index d1075d514e9b2b692f271f10a005815a66b421fb..58ac6fa0a9a30a08a831111513777cca59062724 100644 --- a/python/paddle/fluid/tests/unittests/test_const_value.py +++ b/python/paddle/fluid/tests/unittests/test_const_value.py @@ -16,7 +16,7 @@ import unittest import paddle.fluid.framework as framework -class ConditionalBlock(unittest.TestCase): +class ConstantTest(unittest.TestCase): def test_const_value(self): self.assertEqual(framework.GRAD_VAR_SUFFIX, "@GRAD") self.assertEqual(framework.TEMP_VAR_NAME, "@TEMP@") diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py index db6be21baaa54d33af9f5c44d1815e4b389eb884..d0de7ad52c8a851c16cbbbf544d479f696dee136 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py @@ -20,16 +20,19 @@ from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride class TestMKLDNN(TestConv2dOp): def init_kernel_type(self): self.use_mkldnn = True + self.data_format = "NCHW" class TestMKLDNNWithPad(TestWithPad): def init_kernel_type(self): self.use_mkldnn = True + self.data_format = "NCHW" class TestMKLDNNWithStride(TestWithStride): def init_kernel_type(self): self.use_mkldnn = True + self.data_format = "NCHW" if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_op.py index a478649541ba9828e55c4239090d5aee554223ac..bb1cd87d615fa341b7244e9f3e113b9fb4765ac2 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_op.py @@ -66,6 +66,7 @@ class TestConv2dOp(OpTest): self.op_type = "conv2d" self.use_cudnn = False self.use_mkldnn = False + self.data_format = "AnyLayout" self.dtype = np.float32 self.init_kernel_type() self.init_group() @@ -93,7 +94,8 @@ class TestConv2dOp(OpTest): 'groups': self.groups, 'dilations': self.dilations, 'use_cudnn': self.use_cudnn, - 'use_mkldnn': self.use_mkldnn + 'use_mkldnn': self.use_mkldnn, + 'data_format': self.data_format } self.outputs = {'Output': output} @@ -101,59 +103,35 @@ class TestConv2dOp(OpTest): return core.is_compiled_with_cuda() and self.use_cudnn def test_check_output(self): - if self.testcudnn(): - place = core.CUDAPlace(0) - self.check_output_with_place(place, atol=1e-5) - else: - self.check_output() + place = core.CUDAPlace(0) if self.testcudnn() else core.CPUPlace() + self.check_output_with_place(place, atol=1e-5) def test_check_grad(self): if self.dtype == np.float16: return - if self.testcudnn(): - place = core.CUDAPlace(0) - self.check_grad_with_place( - place, - set(['Input', 'Filter']), - 'Output', - max_relative_error=0.02) - else: - self.check_grad( - set(['Input', 'Filter']), 'Output', max_relative_error=0.02) + place = core.CUDAPlace(0) if self.testcudnn() else core.CPUPlace() + self.check_grad_with_place( + place, set(['Input', 'Filter']), 'Output', max_relative_error=0.02) def test_check_grad_no_filter(self): if self.dtype == np.float16: return - if self.testcudnn(): - place = core.CUDAPlace(0) - self.check_grad_with_place( - place, ['Input'], - 'Output', - max_relative_error=0.02, - no_grad_set=set(['Filter'])) - else: - self.check_grad( - ['Input'], - 'Output', - max_relative_error=0.02, - no_grad_set=set(['Filter'])) + place = core.CUDAPlace(0) if self.testcudnn() else core.CPUPlace() + self.check_grad_with_place( + place, ['Input'], + 'Output', + max_relative_error=0.02, + no_grad_set=set(['Filter'])) def test_check_grad_no_input(self): if self.dtype == np.float16: return - if self.testcudnn(): - place = core.CUDAPlace(0) - self.check_grad_with_place( - place, ['Filter'], - 'Output', - max_relative_error=0.02, - no_grad_set=set(['Input'])) - else: - self.check_grad( - ['Filter'], - 'Output', - max_relative_error=0.02, - no_grad_set=set(['Input'])) + place = core.CUDAPlace(0) if self.testcudnn() else core.CPUPlace() + self.check_grad_with_place( + place, ['Filter'], + 'Output', + max_relative_error=0.02, + no_grad_set=set(['Input'])) def init_test_case(self): self.pad = [0, 0] diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py index ded2f130288a4a959a1c859b2cc8ccf0912efb12..af6cd99b0d7e6b0a2dfd4fc1d33e8390017a5906 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py @@ -191,12 +191,16 @@ class TestWithDilation(TestConv2dTransposeOp): # ------------ test_cudnn ------------ +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNN(TestConv2dTransposeOp): def init_op_type(self): self.use_cudnn = True self.op_type = "conv2d_transpose" +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNNWithPad(TestWithPad): def init_test_case(self): self.pad = [1, 1] @@ -212,6 +216,8 @@ class TestCUDNNWithPad(TestWithPad): self.op_type = "conv2d_transpose" +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNNWithStride(TestWithStride): def init_test_case(self): self.pad = [1, 1] @@ -227,6 +233,8 @@ class TestCUDNNWithStride(TestWithStride): self.op_type = "conv2d_transpose" +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNNWithGroups(TestWithGroups): def init_test_case(self): self.pad = [1, 1] @@ -242,6 +250,19 @@ class TestCUDNNWithGroups(TestWithGroups): self.op_type = "conv2d_transpose" +class TestDepthwiseConvTranspose(TestConv2dTransposeOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [2, 2] + self.dilations = [1, 1] + self.input_size = [2, 8, 16, 16] # NCHW + self.groups = 8 + assert np.mod(self.input_size[1], self.groups) == 0 + f_c = self.input_size[1] / self.groups + self.filter_size = [self.input_size[1], f_c, 4, 4] + self.op_type = "depthwise_conv2d_transpose" + + # Please Don't remove the following code. # Currently, CI use cudnn V5.0 which not support dilation conv. # class TestCUDNNWithDilation(TestWithDilation): diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py b/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py index c9f26d10df8ff39d6bd77b1597336600f676d362..300fa5e8bde001e0f66c5f924a81c30add99aead 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py @@ -197,12 +197,16 @@ class TestWithDilation(TestConv3dTransposeOp): # ------------ test_cudnn ------------ +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNN(TestConv3dTransposeOp): def init_op_type(self): self.use_cudnn = True self.op_type = "conv3d_transpose" +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNNWithPad(TestWithPad): def init_test_case(self): self.pad = [1, 1, 1] @@ -218,6 +222,8 @@ class TestCUDNNWithPad(TestWithPad): self.op_type = "conv3d_transpose" +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNNWithStride(TestWithStride): def init_test_case(self): self.pad = [1, 1, 1] @@ -233,6 +239,8 @@ class TestCUDNNWithStride(TestWithStride): self.op_type = "conv3d_transpose" +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestCUDNNWithGroups(TestWithGroups): def init_test_case(self): self.pad = [1, 1, 1] diff --git a/python/paddle/fluid/tests/unittests/test_conv_shift_op.py b/python/paddle/fluid/tests/unittests/test_conv_shift_op.py index 5d4d244f439a671d895f9237b793e6c6bbf2895b..9fdb7baa90d2184c3c439e76b6bb5f0668f5f9ee 100644 --- a/python/paddle/fluid/tests/unittests/test_conv_shift_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv_shift_op.py @@ -22,8 +22,8 @@ def conv_shift_forward(x, y): M = x.shape[1] N = y.shape[1] y_half_width = (N - 1) / 2 - for i in xrange(M): - for j in xrange(N): + for i in range(M): + for j in range(N): out[:, i] += x[:, (i + j + M - y_half_width) % M] * y[:, j] return out diff --git a/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py b/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py index 5e6f9a20a93e467980f5a4f23fbcb6118317fe44..07c89eefc32fab37ce093e91d96fbe4471ecddc6 100644 --- a/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py +++ b/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py @@ -18,7 +18,7 @@ import paddle.fluid.layers as layers class TestDocString(unittest.TestCase): def test_layer_doc_string(self): - print layers.dropout.__doc__ + print(layers.dropout.__doc__) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py b/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py index f397f542bb07519886d75618e2a915c2dbf61fce..122b076c2d3e3a69f52a2c335e2bc89707b4fa9b 100644 --- a/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py +++ b/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py @@ -22,9 +22,9 @@ from op_test import OpTest class CRFDecoding(object): def __init__(self, emission_weights, transition_weights, seq_start_positions): - assert (emission_weights.shape[0] == seq_start_positions[-1]) + assert (emission_weights.shape[0] == sum(seq_start_positions)) self.tag_num = emission_weights.shape[1] - self.seq_num = len(seq_start_positions) - 1 + self.seq_num = len(seq_start_positions) self.seq_start_positions = seq_start_positions self.x = emission_weights @@ -34,9 +34,9 @@ class CRFDecoding(object): self.w = transition_weights[2:, :] self.track = np.zeros( - (seq_start_positions[-1], self.tag_num), dtype="int64") + (sum(seq_start_positions), self.tag_num), dtype="int64") self.decoded_path = np.zeros( - (seq_start_positions[-1], 1), dtype="int64") + (sum(seq_start_positions), 1), dtype="int64") def _decode_one_sequence(self, decoded_path, x): seq_len, tag_num = x.shape @@ -71,9 +71,11 @@ class CRFDecoding(object): decoded_path[i - 1] = max_idx = track[i, max_idx] def decode(self): + cur_pos = 0 for i in range(self.seq_num): - start = self.seq_start_positions[i] - end = self.seq_start_positions[i + 1] + start = cur_pos + cur_pos += self.seq_start_positions[i] + end = cur_pos self._decode_one_sequence(self.decoded_path[start:end, :], self.x[start:end, :]) return self.decoded_path @@ -90,11 +92,13 @@ class TestCRFDecodingOp1(OpTest): TAG_NUM = 17 MAX_SEQ_LEN = 10 - lod = [[0]] + lod = [[]] + total_len = 0 for i in range(SEQ_NUM): - lod[-1].append(lod[-1][-1] + random.randint(1, MAX_SEQ_LEN)) + lod[-1].append(random.randint(1, MAX_SEQ_LEN)) + total_len += lod[-1][-1] emission = np.random.uniform(-1, 1, - [lod[-1][-1], TAG_NUM]).astype("float64") + [total_len, TAG_NUM]).astype("float64") transition = np.random.uniform(-0.5, 0.5, [TAG_NUM + 2, TAG_NUM]).astype("float64") @@ -126,7 +130,8 @@ class TestCRFDecodingOp2(OpTest): self.op_type = "crf_decoding" TAG_NUM = 5 - lod = [[0, 1, 3, 6, 10]] + lod = [[1, 2, 3, 4]] + total_len = sum(lod[-1]) transition = np.repeat( np.arange( TAG_NUM, dtype="float64").reshape(1, TAG_NUM), @@ -135,13 +140,13 @@ class TestCRFDecodingOp2(OpTest): emission = np.repeat( np.arange( TAG_NUM, dtype="float64").reshape(1, TAG_NUM), - lod[-1][-1], + total_len, axis=0) labels = np.random.randint( - low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int64") + low=0, high=TAG_NUM, size=(total_len, 1), dtype="int64") predicted_labels = np.ones( - (lod[-1][-1], 1), dtype="int64") * (TAG_NUM - 1) + (total_len, 1), dtype="int64") * (TAG_NUM - 1) expected_output = (labels == predicted_labels).astype("int64") self.inputs = { diff --git a/python/paddle/fluid/tests/unittests/test_crop_op.py b/python/paddle/fluid/tests/unittests/test_crop_op.py index 20cc3a643f1adfc04faad15e1b7baad3e22d9d29..4016089c01644f0389855ab114360f90c50a1bbe 100644 --- a/python/paddle/fluid/tests/unittests/test_crop_op.py +++ b/python/paddle/fluid/tests/unittests/test_crop_op.py @@ -42,9 +42,9 @@ class TestCropOp(OpTest): def setUp(self): self.op_type = "crop" self.crop_by_input = False + self.offset_by_input = False self.attrs = {} self.initTestCase() - self.attrs['offsets'] = self.offsets if self.crop_by_input: self.inputs = { 'X': np.random.random(self.x_shape).astype("float32"), @@ -55,6 +55,10 @@ class TestCropOp(OpTest): self.inputs = { 'X': np.random.random(self.x_shape).astype("float32"), } + if self.offset_by_input: + self.inputs['Offsets'] = np.array(self.offsets).astype('int32') + else: + self.attrs['offsets'] = self.offsets self.outputs = { 'Out': crop(self.inputs['X'], self.offsets, self.crop_shape) } @@ -101,5 +105,22 @@ class TestCase4(TestCropOp): self.crop_by_input = True +class TestCase5(TestCropOp): + def initTestCase(self): + self.x_shape = (3, 4, 5) + self.crop_shape = [2, 2, 3] + self.offsets = [1, 0, 2] + self.offset_by_input = True + + +class TestCase6(TestCropOp): + def initTestCase(self): + self.x_shape = (10, 9, 14) + self.crop_shape = [3, 3, 5] + self.offsets = [3, 5, 4] + self.crop_by_input = True + self.offset_by_input = True + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py index c5b9e92d69133e593a2ce223e83006eda590daa5..86ac159323a5f9f6149ce5ed4437402eb885c6bc 100644 --- a/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py +++ b/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py @@ -105,5 +105,107 @@ class TestCrossEntropyOp3(OpTest): ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) +class TestCrossEntropyOp4(OpTest): + """Test high rank tensor cross-entropy with discrete one-hot labels. + """ + + def setUp(self): + self.op_type = "cross_entropy" + shape = [10, 2, 4] + ins_num = np.prod(np.array(shape)) + class_num = 10 + + X_2d = randomize_probability(ins_num, class_num, dtype='float64') + + label_2d = np.random.randint(0, class_num, (ins_num, 1), dtype="int64") + cross_entropy_2d = np.asmatrix( + [[-np.log(X_2d[i][label_2d[i][0]])] for i in range(X_2d.shape[0])], + dtype="float64") + + X = X_2d.reshape(shape + [class_num]) + label = label_2d.reshape(shape + [1]) + cross_entropy = np.array(cross_entropy_2d).reshape(shape + [1]) + + self.inputs = {"X": X, "Label": label} + self.outputs = {"Y": cross_entropy} + self.attrs = {"soft_label": False} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Y", numeric_grad_delta=0.001) + + +class TestCrossEntropyOp5(OpTest): + """Test high rank tensor cross-entropy with vectorized soft labels. + """ + + def setUp(self): + self.op_type = "cross_entropy" + shape = [4, 3] + ins_num = np.prod(np.array(shape)) + class_num = 37 + + X_2d = randomize_probability(ins_num, class_num) + label_2d = np.random.uniform(0.1, 1.0, + [ins_num, class_num]).astype("float32") + label_2d /= label_2d.sum(axis=1, keepdims=True) + cross_entropy_2d = (-label_2d * np.log(X_2d)).sum( + axis=1, keepdims=True).astype("float32") + + X = X_2d.reshape(shape + [class_num]) + label = label_2d.reshape(shape + [class_num]) + cross_entropy = np.array(cross_entropy_2d).reshape(shape + [1]) + + self.inputs = {"X": X, "Label": label} + self.outputs = {"Y": cross_entropy} + self.attrs = {"soft_label": True} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad( + ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) + + +class TestCrossEntropyOp6(OpTest): + """Test high rank tensor cross-entropy with vectorized one-hot representation of labels. + """ + + def setUp(self): + self.op_type = "cross_entropy" + shape = [4, 3, 2] + ins_num = np.prod(np.array(shape)) + class_num = 17 + + X_2d = randomize_probability(ins_num, class_num) + label_index_2d = np.random.randint( + 0, class_num, (ins_num), dtype="int32") + label_2d = np.zeros(X_2d.shape) + label_2d[np.arange(ins_num), label_index_2d] = 1 + + cross_entropy_2d = np.asmatrix( + [[-np.log(X_2d[i][label_index_2d[i]])] + for i in range(X_2d.shape[0])], + dtype="float32") + + X = X_2d.reshape(shape + [class_num]) + label = label_2d.reshape(shape + [class_num]) + cross_entropy = np.array(cross_entropy_2d).reshape(shape + [1]) + + self.inputs = {"X": X, "Label": label.astype(np.float32)} + self.outputs = {"Y": cross_entropy} + self.attrs = {"soft_label": True} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad( + ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_ctc_align.py b/python/paddle/fluid/tests/unittests/test_ctc_align.py index f166031a1cbbaa5e312f5c7919b39648d0dad013..131b4076f45ae25b45bb3f64da07a5c3aacc43d5 100644 --- a/python/paddle/fluid/tests/unittests/test_ctc_align.py +++ b/python/paddle/fluid/tests/unittests/test_ctc_align.py @@ -22,14 +22,16 @@ from test_softmax_op import stable_softmax def CTCAlign(input, lod, blank, merge_repeated): lod0 = lod[0] result = [] - for i in range(len(lod0) - 1): + cur_offset = 0 + for i in range(len(lod0)): prev_token = -1 - for j in range(lod0[i], lod0[i + 1]): + for j in range(cur_offset, cur_offset + lod0[i]): token = input[j][0] if (token != blank) and not (merge_repeated and token == prev_token): result.append(token) prev_token = token + cur_offset += lod0[i] result = np.array(result).reshape([len(result), 1]).astype("int32") if len(result) == 0: result = np.array([-1]) @@ -39,7 +41,7 @@ def CTCAlign(input, lod, blank, merge_repeated): class TestCTCAlignOp(OpTest): def config(self): self.op_type = "ctc_align" - self.input_lod = [[0, 11, 18]] + self.input_lod = [[11, 7]] self.blank = 0 self.merge_repeated = False self.input = np.array( @@ -66,7 +68,7 @@ class TestCTCAlignOp(OpTest): class TestCTCAlignOpCase1(TestCTCAlignOp): def config(self): self.op_type = "ctc_align" - self.input_lod = [[0, 11, 19]] + self.input_lod = [[11, 8]] self.blank = 0 self.merge_repeated = True self.input = np.array( @@ -77,7 +79,7 @@ class TestCTCAlignOpCase1(TestCTCAlignOp): class TestCTCAlignOpCase2(TestCTCAlignOp): def config(self): self.op_type = "ctc_align" - self.input_lod = [[0, 4]] + self.input_lod = [[4]] self.blank = 0 self.merge_repeated = True self.input = np.array([0, 0, 0, 0]).reshape([4, 1]).astype("int32") diff --git a/python/paddle/fluid/tests/unittests/test_data_balance.py b/python/paddle/fluid/tests/unittests/test_data_balance.py new file mode 100644 index 0000000000000000000000000000000000000000..951282e8bab5018204c0d31caa10f8f84a8f3d6c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_data_balance.py @@ -0,0 +1,195 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import paddle.fluid as fluid +import paddle.v2 as paddle +import numpy as np + + +class TestDataBalance(unittest.TestCase): + def prepare_data(self): + def fake_data_generator(): + for n in range(self.total_ins_num): + yield np.ones((3, 4)) * n, n + + # Prepare data + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch( + fake_data_generator, batch_size=self.batch_size) + feeder = fluid.DataFeeder( + feed_list=[ + fluid.layers.data( + name='image', shape=[3, 4], dtype='float32'), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + self.num_batches = fluid.recordio_writer.convert_reader_to_recordio_file( + self.data_file_name, reader, feeder) + + def prepare_lod_data(self): + def fake_data_generator(): + for n in range(1, self.total_ins_num + 1): + d1 = (np.ones((n, 3)) * n).astype('float32') + d2 = (np.array(n).reshape((1, 1))).astype('int32') + yield d1, d2 + + # Prepare lod data + with fluid.program_guard(fluid.Program(), fluid.Program()): + with fluid.recordio_writer.create_recordio_writer( + filename=self.lod_data_file_name) as writer: + eof = False + generator = fake_data_generator() + while (not eof): + data_batch = [ + np.array([]).reshape((0, 3)), np.array([]).reshape( + (0, 1)) + ] + lod = [0] + for _ in range(self.batch_size): + try: + ins = next(generator) + except StopIteration: + eof = True + break + for i, d in enumerate(ins): + data_batch[i] = np.concatenate( + (data_batch[i], d), axis=0) + lod.append(lod[-1] + ins[0].shape[0]) + if data_batch[0].shape[0] > 0: + for i, d in enumerate(data_batch): + t = fluid.LoDTensor() + t.set(data_batch[i], fluid.CPUPlace()) + if i == 0: + t.set_lod([lod]) + writer.append_tensor(t) + writer.complete_append_tensor() + + def setUp(self): + self.use_cuda = fluid.core.is_compiled_with_cuda() + self.data_file_name = './data_balance_test.recordio' + self.lod_data_file_name = './data_balance_with_lod_test.recordio' + self.total_ins_num = 50 + self.batch_size = 10 + self.prepare_data() + self.prepare_lod_data() + + def main(self): + main_prog = fluid.Program() + startup_prog = fluid.Program() + with fluid.program_guard(main_prog, startup_prog): + data_reader = fluid.layers.io.open_files( + filenames=[self.data_file_name], + shapes=[[-1, 3, 4], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + if self.use_cuda: + data_reader = fluid.layers.double_buffer(data_reader) + image, label = fluid.layers.read_file(data_reader) + + place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + + build_strategy = fluid.BuildStrategy() + build_strategy.enable_data_balance = True + parallel_exe = fluid.ParallelExecutor( + use_cuda=self.use_cuda, + main_program=main_prog, + build_strategy=build_strategy) + + if (parallel_exe.device_count > self.batch_size): + print("WARNING: Unittest TestDataBalance skipped. \ + For the result is not correct when device count \ + is larger than batch size.") + exit(0) + fetch_list = [image.name, label.name] + + data_appeared = [False] * self.total_ins_num + while (True): + try: + image_val, label_val = parallel_exe.run(fetch_list, + return_numpy=True) + except fluid.core.EOFException: + break + ins_num = image_val.shape[0] + broadcasted_label = np.ones( + (ins_num, 3, 4)) * label_val.reshape((ins_num, 1, 1)) + self.assertEqual(image_val.all(), broadcasted_label.all()) + for l in label_val: + self.assertFalse(data_appeared[l[0]]) + data_appeared[l[0]] = True + for i in data_appeared: + self.assertTrue(i) + + def main_lod(self): + main_prog = fluid.Program() + startup_prog = fluid.Program() + with fluid.program_guard(main_prog, startup_prog): + data_reader = fluid.layers.io.open_files( + filenames=[self.lod_data_file_name], + shapes=[[-1, 3], [-1, 1]], + lod_levels=[1, 0], + dtypes=['float32', 'int32']) + ins, label = fluid.layers.read_file(data_reader) + + place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + build_strategy = fluid.BuildStrategy() + build_strategy.enable_data_balance = True + parallel_exe = fluid.ParallelExecutor( + use_cuda=self.use_cuda, + main_program=main_prog, + build_strategy=build_strategy) + + if parallel_exe.device_count > self.batch_size: + print("WARNING: Unittest TestDataBalance skipped. \ + For the result is not correct when device count \ + is larger than batch size.") + exit(0) + fetch_list = [ins.name, label.name] + + data_appeared = [False] * self.total_ins_num + while (True): + try: + ins_tensor, label_tensor = parallel_exe.run( + fetch_list, return_numpy=False) + except fluid.core.EOFException: + break + + ins_val = np.array(ins_tensor) + label_val = np.array(label_tensor) + ins_lod = ins_tensor.lod()[0] + self.assertEqual(ins_val.shape[1], 3) + self.assertEqual(label_val.shape[1], 1) + self.assertEqual(len(ins_lod) - 1, label_val.shape[0]) + for i in range(0, len(ins_lod) - 1): + ins_elem = ins_val[ins_lod[i]:ins_lod[i + 1]][:] + label_elem = label_val[i][0] + self.assertEqual(ins_elem.all(), label_elem.all()) + self.assertFalse(data_appeared[int(label_elem - 1)]) + data_appeared[int(label_elem - 1)] = True + + for i in data_appeared: + self.assertTrue(i) + + def test_all(self): + self.main() + self.main_lod() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py b/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py index a3bf7b544b91c70ffe3894219c118ec9887aba81..868bcca881a65dad7d0ecabb1e388818cdd0997e 100644 --- a/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py +++ b/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py @@ -39,7 +39,7 @@ class TestDefaultScopeFuncs(unittest.TestCase): self.assertTrue(i.is_int()) self.assertEqual(10, i.get_int()) - for _ in xrange(10): + for _ in range(10): scoped_function(__new_scope__) diff --git a/python/paddle/fluid/tests/unittests/test_desc_clone.py b/python/paddle/fluid/tests/unittests/test_desc_clone.py new file mode 100644 index 0000000000000000000000000000000000000000..8603d3a5b3b5d368fe87b8dcf9dc7363f95caf86 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_desc_clone.py @@ -0,0 +1,196 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import argparse +import time +import math + +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler +from paddle.fluid import core +import unittest +from multiprocessing import Process +import os +import signal +import collections + +SEED = 1 +DTYPE = "float32" +paddle.dataset.mnist.fetch() + + +# random seed must set before configuring the network. +# fluid.default_startup_program().random_seed = SEED +def cnn_model(data): + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=data, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + + # TODO(dzhwinter) : refine the initializer and random seed settting + SIZE = 10 + input_shape = conv_pool_2.shape + param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE] + scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5 + + predict = fluid.layers.fc( + input=conv_pool_2, + size=SIZE, + act="softmax", + param_attr=fluid.param_attr.ParamAttr( + initializer=fluid.initializer.NormalInitializer( + loc=0.0, scale=scale))) + return predict + + +def get_model(batch_size): + # Input data + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + inference_program = fluid.default_main_program().clone() + # Optimization + opt = fluid.optimizer.AdamOptimizer( + learning_rate=0.001, beta1=0.9, beta2=0.999) + + # Reader + train_reader = paddle.batch( + paddle.dataset.mnist.train(), batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + opt.minimize(avg_cost) + return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict + + +def get_transpiler(trainer_id, main_program, pserver_endpoints, trainers): + t = fluid.DistributeTranspiler() + t.transpile( + trainer_id=trainer_id, + program=main_program, + pservers=pserver_endpoints, + trainers=trainers) + return t + + +def operator_equal(a, b): + for k, v in a.__dict__.iteritems(): + if isinstance(v, fluid.framework.Program) or \ + isinstance(v, fluid.framework.Block): + continue + + elif isinstance(v, core.OpDesc): + if v.serialize_to_string() != b.__dict__[k].serialize_to_string(): + raise ValueError("In operator_equal not equal:{0}\n".format(k)) + + elif isinstance(v, collections.OrderedDict): + v0 = sorted(v.iteritems(), key=lambda x: x[0]) + v1 = sorted(b.__dict__[k].iteritems(), key=lambda x: x[0]) + + if v0 != v1: + raise ValueError("In operator_equal not equal:{0}\n".format(k)) + + elif (v != b.__dict__[k]): + raise ValueError("In operator_equal not equal:{0}\n".format(k)) + + return True + + +def block_equal(a, b): + for k, v in a.__dict__.iteritems(): + if isinstance(v, core.ProgramDesc) or isinstance( + v, fluid.framework.Program) or isinstance(v, core.BlockDesc): + continue + + elif k == "ops": + for i in range(0, len(a.ops)): + if not operator_equal(a.ops[i], b.ops[i]): + raise ValueError("In block_equal not equal:{0}\n".format(k)) + assert (len(a.ops) == len(b.ops)) + + elif isinstance(v, collections.OrderedDict): + v0 = sorted(v.iteritems(), key=lambda x: x[0]) + v1 = sorted(b.__dict__[k].iteritems(), key=lambda x: x[0]) + + if v0 != v1: + raise ValueError("In block_equal not equal:{0}\n".format(k)) + + elif (v != b.__dict__[k]): + raise ValueError("In block_equal not equal:{0}\n".format(k)) + + return True + + +def program_equal(a, b): + for k, v in a.__dict__.iteritems(): + if isinstance(v, core.ProgramDesc): + continue + + elif k == 'blocks': + for i in range(0, len(a.blocks)): + if not block_equal(a.blocks[i], b.blocks[i]): + raise ValueError("In operator_equal not equal:{0}\n".format( + k)) + return False + assert (len(a.blocks) == len(b.blocks)) + + elif (v != b.__dict__[k]): + raise ValueError("In program_equal not equal:{0}\n".format(k)) + + return True + + +class TestDistMnist(unittest.TestCase): + def test_desc_clone(self): + get_model(batch_size=20) + + pserver_endpoints = "127.0.0.1:9123" + trainers = 1 + current_endpoint = "127.0.0.1:9123" + t = get_transpiler(0, + fluid.default_main_program(), pserver_endpoints, + trainers) + + pserver_prog = t.get_pserver_program(current_endpoint) + startup_prog = t.get_startup_program(current_endpoint, pserver_prog) + main = pserver_prog.clone() + startup = startup_prog.clone() + + self.assertTrue(program_equal(main, pserver_prog)) + self.assertTrue(program_equal(startup, startup_prog)) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_detection_map_op.py b/python/paddle/fluid/tests/unittests/test_detection_map_op.py index f545ad155ccd28c2d34e424d307eed49b37f20fb..8b66d1b270980a18fd1bbd068917e982a450ad6f 100644 --- a/python/paddle/fluid/tests/unittests/test_detection_map_op.py +++ b/python/paddle/fluid/tests/unittests/test_detection_map_op.py @@ -74,13 +74,13 @@ class TestDetectionMAPOp(OpTest): self.evaluate_difficult = True self.ap_type = "integral" - self.label_lod = [[0, 2, 4]] + self.label_lod = [[2, 2]] # label difficult xmin ymin xmax ymax self.label = [[1, 0, 0.1, 0.1, 0.3, 0.3], [1, 1, 0.6, 0.6, 0.8, 0.8], [2, 0, 0.3, 0.3, 0.6, 0.5], [1, 0, 0.7, 0.1, 0.9, 0.3]] # label score xmin ymin xmax ymax difficult - self.detect_lod = [[0, 3, 7]] + self.detect_lod = [[3, 4]] self.detect = [ [1, 0.3, 0.1, 0.0, 0.4, 0.3], [1, 0.7, 0.0, 0.1, 0.2, 0.3], [1, 0.9, 0.7, 0.6, 0.8, 0.8], [2, 0.8, 0.2, 0.1, 0.4, 0.4], @@ -89,7 +89,7 @@ class TestDetectionMAPOp(OpTest): ] # label score true_pos false_pos - self.tf_pos_lod = [[0, 3, 7]] + self.tf_pos_lod = [[3, 4]] self.tf_pos = [[1, 0.9, 1, 0], [1, 0.7, 1, 0], [1, 0.3, 0, 1], [1, 0.2, 1, 0], [2, 0.8, 0, 1], [2, 0.1, 1, 0], [3, 0.2, 0, 1]] @@ -112,15 +112,19 @@ class TestDetectionMAPOp(OpTest): for i, count in enumerate(class_pos_count): class_pos_count_dict[i] = count - for i in range(len(true_pos_lod[0]) - 1): - start = true_pos_lod[0][i] - end = true_pos_lod[0][i + 1] + cur_pos = 0 + for i in range(len(true_pos_lod[0])): + start = cur_pos + cur_pos += true_pos_lod[0][i] + end = cur_pos for j in range(start, end): true_pos_dict[i].append(true_pos[j]) - for i in range(len(false_pos_lod[0]) - 1): - start = false_pos_lod[0][i] - end = false_pos_lod[0][i + 1] + cur_pos = 0 + for i in range(len(false_pos_lod[0])): + start = cur_pos + cur_pos += false_pos_lod[0][i] + end = cur_pos for j in range(start, end): false_pos_dict[i].append(false_pos[j]) @@ -130,19 +134,19 @@ class TestDetectionMAPOp(OpTest): label_number = self.class_num out_class_pos_count = [] - out_true_pos_lod = [0] + out_true_pos_lod = [] out_true_pos = [] - out_false_pos_lod = [0] + out_false_pos_lod = [] out_false_pos = [] for i in range(label_number): out_class_pos_count.append([label_count[i]]) true_pos_list = true_pos[i] out_true_pos += true_pos_list - out_true_pos_lod.append(len(out_true_pos)) + out_true_pos_lod.append(len(true_pos_list)) false_pos_list = false_pos[i] out_false_pos += false_pos_list - out_false_pos_lod.append(len(out_false_pos)) + out_false_pos_lod.append(len(false_pos_list)) return out_class_pos_count, out_true_pos, [ out_true_pos_lod @@ -172,7 +176,7 @@ class TestDetectionMAPOp(OpTest): true_pos[label].append([score, tp]) false_pos[label].append([score, fp]) - for (label, label_pos_num) in label_count.items(): + for (label, label_pos_num) in list(label_count.items()): if label_pos_num == 0 or label not in true_pos: continue label_true_pos = true_pos[label] label_false_pos = false_pos[label] @@ -241,7 +245,7 @@ class TestDetectionMAPOpSkipDiff(TestDetectionMAPOp): self.evaluate_difficult = False - self.tf_pos_lod = [[0, 2, 6]] + self.tf_pos_lod = [[2, 4]] # label score true_pos false_pos self.tf_pos = [[1, 0.7, 1, 0], [1, 0.3, 0, 1], [1, 0.2, 1, 0], [2, 0.8, 0, 1], [2, 0.1, 1, 0], [3, 0.2, 0, 1]] @@ -267,9 +271,9 @@ class TestDetectionMAPOpMultiBatch(TestDetectionMAPOp): def init_test_case(self): super(TestDetectionMAPOpMultiBatch, self).init_test_case() self.class_pos_count = [0, 2, 1] - self.true_pos_lod = [[0, 0, 3, 5]] + self.true_pos_lod = [[0, 3, 2]] self.true_pos = [[0.7, 1.], [0.3, 0.], [0.2, 1.], [0.8, 0.], [0.1, 1.]] - self.false_pos_lod = [[0, 0, 3, 5]] + self.false_pos_lod = [[0, 3, 2]] self.false_pos = [[0.7, 0.], [0.3, 1.], [0.2, 0.], [0.8, 1.], [0.1, 0.]] diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py new file mode 100644 index 0000000000000000000000000000000000000000..4379463aca4443eb7a886ce78446440cc59f3b30 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_base.py @@ -0,0 +1,282 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import time + +import unittest +import os +import sys +import signal +import subprocess +import six + + +class TestDistRunnerBase(object): + def get_model(self, batch_size=2): + raise NotImplementedError( + "get_model should be implemented by child classes.") + + def get_transpiler(self, trainer_id, main_program, pserver_endpoints, + trainers): + # NOTE: import fluid until runtime, or else forking processes will cause error. + import paddle + import paddle.fluid as fluid + t = fluid.DistributeTranspiler() + t.transpile( + trainer_id=trainer_id, + program=main_program, + pservers=pserver_endpoints, + trainers=trainers) + return t + + def run_pserver(self, pserver_endpoints, trainers, current_endpoint, + trainer_id): + import paddle + import paddle.fluid as fluid + self.get_model(batch_size=2) + t = self.get_transpiler(trainer_id, + fluid.default_main_program(), pserver_endpoints, + trainers) + pserver_prog = t.get_pserver_program(current_endpoint) + startup_prog = t.get_startup_program(current_endpoint, pserver_prog) + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + exe.run(pserver_prog) + + def run_trainer(self, place, endpoints, trainer_id, trainers, is_dist=True): + import paddle + import paddle.fluid as fluid + test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \ + self.get_model(batch_size=2) + if is_dist: + t = self.get_transpiler(trainer_id, + fluid.default_main_program(), endpoints, + trainers) + trainer_prog = t.get_trainer_program() + else: + trainer_prog = fluid.default_main_program() + + startup_exe = fluid.Executor(place) + startup_exe.run(fluid.default_startup_program()) + + strategy = fluid.ExecutionStrategy() + strategy.num_threads = 1 + strategy.allow_op_delay = False + exe = fluid.ParallelExecutor( + True, loss_name=avg_cost.name, exec_strategy=strategy) + + feed_var_list = [ + var for var in trainer_prog.global_block().vars.values() + if var.is_data + ] + + feeder = fluid.DataFeeder(feed_var_list, place) + reader_generator = test_reader() + + data = next(reader_generator) + first_loss, = exe.run(fetch_list=[avg_cost.name], + feed=feeder.feed(data)) + print(first_loss) + + for i in six.moves.xrange(5): + data = next(reader_generator) + loss, = exe.run(fetch_list=[avg_cost.name], feed=feeder.feed(data)) + + data = next(reader_generator) + last_loss, = exe.run(fetch_list=[avg_cost.name], feed=feeder.feed(data)) + print(last_loss) + + +def runtime_main(test_class): + import paddle + import paddle.fluid as fluid + import paddle.fluid.core as core + + if len(sys.argv) != 7: + print( + "Usage: python dist_se_resnext.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist]" + ) + role = sys.argv[1] + endpoints = sys.argv[2] + trainer_id = int(sys.argv[3]) + current_endpoint = sys.argv[4] + trainers = int(sys.argv[5]) + is_dist = True if sys.argv[6] == "TRUE" else False + + model = test_class() + if role == "pserver": + model.run_pserver(endpoints, trainers, current_endpoint, trainer_id) + else: + p = fluid.CUDAPlace(0) if core.is_compiled_with_cuda( + ) else fluid.CPUPlace() + model.run_trainer(p, endpoints, trainer_id, trainers, is_dist) + + +class TestDistBase(unittest.TestCase): + def setUp(self): + self._trainers = 2 + self._pservers = 2 + self._ps_endpoints = "127.0.0.1:9123,127.0.0.1:9124" + self._python_interp = "python" + + def start_pserver(self, model_file, check_error_log): + ps0_ep, ps1_ep = self._ps_endpoints.split(",") + ps0_cmd = "%s %s pserver %s 0 %s %d TRUE" % \ + (self._python_interp, model_file, self._ps_endpoints, ps0_ep, + self._trainers) + ps1_cmd = "%s %s pserver %s 0 %s %d TRUE" % \ + (self._python_interp, model_file, self._ps_endpoints, ps1_ep, + self._trainers) + + ps0_pipe = subprocess.PIPE + ps1_pipe = subprocess.PIPE + if check_error_log: + print("ps0_cmd:", ps0_cmd) + print("ps1_cmd:", ps1_cmd) + ps0_pipe = open("/tmp/ps0_err.log", "wb") + ps1_pipe = open("/tmp/ps1_err.log", "wb") + + ps0_proc = subprocess.Popen( + ps0_cmd.split(" "), stdout=subprocess.PIPE, stderr=ps0_pipe) + ps1_proc = subprocess.Popen( + ps1_cmd.split(" "), stdout=subprocess.PIPE, stderr=ps1_pipe) + + if not check_error_log: + return ps0_proc, ps1_proc, None, None + else: + return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe + + def _wait_ps_ready(self, pid): + retry_times = 50 + while True: + assert retry_times >= 0, "wait ps ready failed" + time.sleep(3) + try: + # the listen_and_serv_op would touch a file which contains the listen port + # on the /tmp directory until it was ready to process all the RPC call. + os.stat("/tmp/paddle.%d.port" % pid) + return + except os.error as e: + sys.stderr.write('waiting for pserver: %s, left retry %d\n' % + (e, retry_times)) + retry_times -= 1 + + def check_with_place(self, model_file, delta=1e-3, check_error_log=False): + # *ATTENTION* THIS TEST NEEDS AT LEAST 2GPUS TO RUN + required_envs = { + "PATH": os.getenv("PATH"), + "PYTHONPATH": os.getenv("PYTHONPATH"), + "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH"), + "FLAGS_fraction_of_gpu_memory_to_use": "0.15", + "FLAGS_cudnn_deterministic": "1" + } + + if check_error_log: + required_envs["GLOG_v"] = "7" + required_envs["GLOG_logtostderr"] = "1" + + # Run local to get a base line + env_local = {"CUDA_VISIBLE_DEVICES": "0"} + env_local.update(required_envs) + local_cmd = "%s %s trainer %s 0 %s %d FLASE" % \ + (self._python_interp, model_file, + "127.0.0.1:1234", "127.0.0.1:1234", 1) + if not check_error_log: + local_proc = subprocess.Popen( + local_cmd.split(" "), + stdout=subprocess.PIPE, + stderr=subprocess.PIPE, + env=env_local) + else: + print("trainer cmd:", local_cmd) + err_log = open("/tmp/trainer.err.log", "wb") + local_proc = subprocess.Popen( + local_cmd.split(" "), + stdout=subprocess.PIPE, + stderr=err_log, + env=env_local) + + local_proc.wait() + out, err = local_proc.communicate() + local_ret = out + sys.stderr.write('local_loss: %s\n' % local_ret) + sys.stderr.write('local_stderr: %s\n' % err) + + # Run dist train to compare with local results + ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(model_file, + check_error_log) + self._wait_ps_ready(ps0.pid) + self._wait_ps_ready(ps1.pid) + + ps0_ep, ps1_ep = self._ps_endpoints.split(",") + tr0_cmd = "%s %s trainer %s 0 %s %d TRUE" % \ + (self._python_interp, model_file, self._ps_endpoints, ps0_ep, + self._trainers) + tr1_cmd = "%s %s trainer %s 1 %s %d TRUE" % \ + (self._python_interp, model_file, self._ps_endpoints, ps1_ep, + self._trainers) + + env0 = {"CUDA_VISIBLE_DEVICES": "0"} + env1 = {"CUDA_VISIBLE_DEVICES": "1"} + env0.update(required_envs) + env1.update(required_envs) + FNULL = open(os.devnull, 'w') + + tr0_pipe = subprocess.PIPE + tr1_pipe = subprocess.PIPE + if check_error_log: + print("tr0_cmd:", tr0_cmd) + print("tr1_cmd:", tr1_cmd) + tr0_pipe = open("/tmp/tr0_err.log", "wb") + tr1_pipe = open("/tmp/tr1_err.log", "wb") + + tr0_proc = subprocess.Popen( + tr0_cmd.split(" "), + stdout=subprocess.PIPE, + stderr=tr0_pipe, + env=env0) + tr1_proc = subprocess.Popen( + tr1_cmd.split(" "), + stdout=subprocess.PIPE, + stderr=tr1_pipe, + env=env1) + + tr0_proc.wait() + tr1_proc.wait() + out, err = tr0_proc.communicate() + sys.stderr.write('dist_stderr: %s\n' % err) + loss_data0 = out + sys.stderr.write('dist_loss: %s\n' % loss_data0) + lines = loss_data0.split("\n") + dist_first_loss = eval(lines[0].replace(" ", ","))[0] + dist_last_loss = eval(lines[1].replace(" ", ","))[0] + + local_lines = local_ret.split("\n") + local_first_loss = eval(local_lines[0])[0] + local_last_loss = eval(local_lines[1])[0] + + # close trainer file + if check_error_log: + tr0_pipe.close() + tr1_pipe.close() + + ps0_pipe.close() + ps1_pipe.close() + # FIXME: use terminate() instead of sigkill. + os.kill(ps0.pid, signal.SIGKILL) + os.kill(ps1.pid, signal.SIGKILL) + FNULL.close() + + self.assertAlmostEqual(local_first_loss, dist_first_loss, delta=delta) + self.assertAlmostEqual(local_last_loss, dist_last_loss, delta=delta) diff --git a/python/paddle/fluid/tests/unittests/test_dist_mnist.py b/python/paddle/fluid/tests/unittests/test_dist_mnist.py new file mode 100644 index 0000000000000000000000000000000000000000..b3ccec9a7d65de57778a1f013465d41a5a267676 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_mnist.py @@ -0,0 +1,24 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import unittest +from test_dist_base import TestDistBase + + +class TestDistSeResneXt2x2(TestDistBase): + def test_se_resnext(self): + self.check_with_place("dist_mnist.py", delta=1e-7) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py new file mode 100644 index 0000000000000000000000000000000000000000..a33a338fc11e4301a8ec0eb565686d62b547b7f7 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py @@ -0,0 +1,24 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import unittest +from test_dist_base import TestDistBase + + +class TestDistSeResneXt2x2(TestDistBase): + def test_se_resnext(self): + self.check_with_place("dist_se_resnext.py", delta=1e-7) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index 2314bb2ed8a4eeb34752fd5d040f8a8476798aa6..55aa923f5ab229bc8e9a0b891e0ac9c2ec49d31b 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -16,17 +16,24 @@ import os import time import unittest from multiprocessing import Process +import signal import numpy import paddle.fluid as fluid import paddle.fluid.layers as layers +from paddle.fluid.layers.io import ListenAndServ +from paddle.fluid.layers.io import Recv +from paddle.fluid.layers.io import Send + +from paddle.fluid import core + +RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName( +) +RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC class TestSendOp(unittest.TestCase): - @unittest.skip( - "This test is buggy. We cannot use time.sleep to sync processes, the connection may fail in unittest." - ) def test_send(self): # Run init_serv in a thread place = fluid.CPUPlace() @@ -35,7 +42,9 @@ class TestSendOp(unittest.TestCase): p.daemon = True p.start() - time.sleep(10) + self.ps_timeout = 5 + self._wait_ps_ready(p.pid) + with open("/tmp/paddle.%d.port" % p.pid, "r") as fn: selected_port = int(fn.readlines()[0]) self.init_client(place, selected_port) @@ -44,15 +53,28 @@ class TestSendOp(unittest.TestCase): self.assertTrue(numpy.allclose(self.local_out, self.dist_out)) # FIXME(typhoonzero): find a way to gracefully shutdown the server. - os.system("kill -9 %d" % p.pid) + os.kill(p.pid, signal.SIGKILL) p.join() + def _wait_ps_ready(self, pid): + start_left_time = self.ps_timeout + sleep_time = 0.5 + while True: + assert start_left_time >= 0, "wait ps ready failed" + time.sleep(sleep_time) + try: + # the listen_and_serv_op would touch a file which contains the listen port + # on the /tmp directory until it was ready to process all the RPC call. + os.stat("/tmp/paddle.%d.port" % pid) + return + except os.error: + start_left_time -= sleep_time + def init_serv(self, place): main = fluid.Program() with fluid.program_guard(main): - serv = layers.ListenAndServ( - "127.0.0.1:0", ["X"], optimizer_mode=False) + serv = ListenAndServ("127.0.0.1:0", ["X"], optimizer_mode=False) with serv.do(): out_var = main.global_block().create_var( name="scale_0.tmp_0", @@ -73,18 +95,32 @@ class TestSendOp(unittest.TestCase): def init_client(self, place, port): main = fluid.Program() with fluid.program_guard(main): + main.global_block().append_op( + type="fetch_barrier", + inputs={}, + outputs={}, + attrs={ + "endpoints": ["127.0.0.1:{0}".format(port)], + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + x = layers.data( shape=[32, 32], dtype='float32', name='X', append_batch_size=False) fluid.initializer.Constant(value=2.3)(x, main.global_block()) + get_var = main.global_block().create_var( name="scale_0.tmp_0", # server side var dtype="float32", persistable=False, shape=[32, 32]) - o = layers.Send("127.0.0.1:%d" % port, [x], [get_var]) + fluid.initializer.Constant(value=2.3)(get_var, main.global_block()) + + Send("127.0.0.1:%d" % port, [x]) + o = Recv("127.0.0.1:%d" % port, [get_var]) + exe = fluid.Executor(place) self.dist_out = exe.run(main, fetch_list=o) # o is a list diff --git a/python/paddle/fluid/tests/unittests/test_dist_transformer.py b/python/paddle/fluid/tests/unittests/test_dist_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..68cd35d751dbce7eef9919dc8678fc0dd117757b --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_transformer.py @@ -0,0 +1,27 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +from test_dist_base import TestDistBase + + +class TestDistTransformer2x2(TestDistBase): + def test_transformer(self): + # TODO(paddle-dev): check if the delta is OK. + # Usually start around ~8000 and converge to ~5000 + self.check_with_place("dist_transformer.py", delta=400) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index fa49bd41a5876847d046682dce5c3d3868a18500..124abf4ccde98d565b3286c72793c91fd26bb71c 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -12,104 +12,601 @@ # See the License for the specific language governing permissions and # limitations under the License. -import unittest +import math +import unittest import paddle.fluid as fluid -import paddle.fluid.core as core -import paddle.fluid.layers as layers from paddle.fluid.transpiler.distribute_transpiler import delete_ops -import numpy +import traceback +import collections -class TestDistTranspiler(unittest.TestCase): +class TranspilerTest(unittest.TestCase): def setUp(self): self.trainer_id = 0 self.trainers = 2 self.pservers = 2 + # NOTE: we do not actually bind this port self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175" - self.current_pserver_ep = "127.0.0.1:6174" + self.pserver1_ep = "127.0.0.1:6174" + self.pserver2_ep = "127.0.0.1:6175" + self.sync_mode = True + self.transpiler = None def net_conf(self): x = fluid.layers.data(name='x', shape=[1000], dtype='float32') - y_predict = fluid.layers.fc(input=x, size=1000, act=None, - param_attr=fluid.ParamAttr(name='fc_w')) - + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) y = fluid.layers.data(name='y', shape=[1], dtype='float32') - cost = fluid.layers.square_error_cost(input=y_predict, label=y) avg_cost = fluid.layers.mean(cost) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) + sgd_optimizer.minimize(avg_cost) + return + + def get_main_program(self): + main = fluid.Program() + with fluid.program_guard(main): + self.net_conf() + self.origin_prog = main.clone() + return main + + def get_trainer(self, config=None): + src = fluid.default_startup_program().clone() + + t = self._transpiler_instance(config) + + trainer_main = t.get_trainer_program() + trainer_startup = fluid.default_startup_program() + + assert (src.num_blocks == 1) + assert (trainer_startup.num_blocks == src.num_blocks) + + return trainer_main, trainer_startup + + def get_pserver(self, ep, config=None, sync_mode=True): + t = self._transpiler_instance(config, sync_mode) + pserver = t.get_pserver_program(ep) + startup = t.get_startup_program(ep, pserver) + return pserver, startup + + def _transpiler_instance(self, config=None, sync_mode=True): + if not self.transpiler: + main = self.get_main_program() + self.transpiler = fluid.DistributeTranspiler(config=config) + self.transpiler.transpile( + self.trainer_id, + program=main, + pservers=self.pserver_eps, + trainers=self.trainers, + sync_mode=sync_mode) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) - return optimize_ops, params_grads + return self.transpiler + + def transpiler_test_impl(self): + pass def test_transpiler(self): - trainer = self.get_trainer() - pserver, startup = self.get_pserver(self.current_pserver_ep) - self.assertEqual([op.type for op in trainer.global_block().ops], - self.get_expect_trainer_ops()) + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + self.transpiler_test_impl() + + +class TestBasicModel(TranspilerTest): + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + pserver2, startup2 = self.get_pserver(self.pserver2_ep) + + trainer, trainer_startup = self.get_trainer() + + # splited var blocks should be in startup program + self.assertTrue("fc_w.block0" in trainer_startup.global_block().vars) + self.assertTrue("fc_w.block1" in trainer_startup.global_block().vars) + self.assertTrue("fc_w" in trainer_startup.global_block().vars) + self.assertTrue("fc_b" in trainer_startup.global_block().vars) + self.assertTrue("fc_w@GRAD" not in trainer_startup.global_block().vars) + self.assertTrue("fc_b@GRAD" not in trainer_startup.global_block().vars) + + src = [op.type for op in trainer_startup.global_block().ops] + dst = ['fill_constant', 'fill_constant', 'uniform_random', 'recv', 'recv', \ + 'fetch_barrier', 'concat'] + + self.assertEqual(src, dst) + + self.assertEqual([op.type for op in trainer.global_block().ops], [ + 'mul', 'elementwise_add', 'elementwise_sub', 'square', 'mean', + 'fill_constant', 'mean_grad', 'square_grad', 'elementwise_sub_grad', + 'elementwise_add_grad', 'send', 'mul_grad', 'split_byref', 'send', + 'send_barrier', 'recv', 'recv', 'fetch_barrier', 'concat' + ]) self.assertEqual(len(pserver.blocks), 3) # block0: listen_and_serv self.assertEqual([op.type for op in pserver.blocks[0].ops], ["listen_and_serv"]) - # block2: optimize pass + # block1~2: optimize pass self.assertEqual([op.type for op in pserver.blocks[1].ops], ["sum", "scale", "sgd"]) - # confirm startup program + self.assertEqual([op.type for op in startup.global_block().ops], + ["fill_constant", "fill_constant", "uniform_random"]) + # the variable #fc_w will be split into two blocks + fc_w_var = startup.global_block().var("fc_w.block1") + self.assertEqual(fc_w_var.shape, (500, 1000)) + # all parameters should be optimized on pserver + + pserver_params = [] + for prog in [pserver, pserver2]: + for blk in prog.blocks: + for op in blk.ops: + if "Param" in op.input_names: + param_name = op.input("Param")[0] + is_block_idx = param_name.find(".block") + if is_block_idx != -1: + origin_param_name = param_name[:is_block_idx] + else: + origin_param_name = param_name + pserver_params.append(origin_param_name) + trainer_params = [] + for op in self.origin_prog.global_block().ops: + if "Param" in op.input_names: + trainer_params.append(op.input("Param")[0]) + self.assertEqual(set(pserver_params), set(trainer_params)) - self.assertEqual([op.type for op in startup.global_block().ops], [ - "fill_constant", "fill_constant", "uniform_random", "uniform_random" + +class TestBasicModelWithLargeBlockSize(TranspilerTest): + def transpiler_test_impl(self): + config = fluid.DistributeTranspilerConfig() + config.min_block_size = 1048576 + + pserver, startup = self.get_pserver(self.pserver1_ep, config) + pserver2, startup2 = self.get_pserver(self.pserver2_ep, config) + + trainer, _ = self.get_trainer(config) + + self.assertEqual([op.type for op in trainer.global_block().ops], [ + 'mul', 'elementwise_add', 'elementwise_sub', 'square', 'mean', + 'fill_constant', 'mean_grad', 'square_grad', 'elementwise_sub_grad', + 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'send_barrier', + 'recv', 'recv', 'fetch_barrier' ]) + self.assertEqual(len(pserver.blocks), 2) + # block0: listen_and_serv + self.assertEqual([op.type for op in pserver.blocks[0].ops], + ["listen_and_serv"]) + # block1~2: optimize pass + self.assertEqual([op.type for op in pserver.blocks[1].ops], + ["sum", "scale", "sgd"]) + # confirm startup program + self.assertEqual([op.type for op in startup.global_block().ops], + ["fill_constant", "fill_constant"]) # the variable #fc_w will be split into two blocks - fc_w_var = startup.global_block().var("fc_w.block1") - self.assertEqual(fc_w_var.shape, (500, 1000)) + fc_w_var = startup2.global_block().var("fc_w") + self.assertEqual(fc_w_var.shape, (1000, 1000)) + # all parameters should be optimized on pserver - def get_main_program(self): - main = fluid.Program() + pserver_params = [] + for prog in [pserver, pserver2]: + for blk in prog.blocks: + for op in blk.ops: + if "Param" in op.input_names: + param_name = op.input("Param")[0] + is_block_idx = param_name.find(".block") + if is_block_idx != -1: + origin_param_name = param_name[:is_block_idx] + else: + origin_param_name = param_name + pserver_params.append(origin_param_name) + trainer_params = [] + for op in self.origin_prog.global_block().ops: + if "Param" in op.input_names: + trainer_params.append(op.input("Param")[0]) + self.assertEqual(set(pserver_params), set(trainer_params)) - with fluid.program_guard(main): - self.net_conf() - return main +class TestNoSliceVar(TranspilerTest): + def setUp(self): + super(TestNoSliceVar, self).setUp() + + def transpiler_test_impl(self): + config = fluid.DistributeTranspilerConfig() + config.slice_var_up = False + + _, startup = self.get_pserver(self.pserver1_ep, config) + _, startup2 = self.get_pserver(self.pserver2_ep, config) + + if "fc_w" in startup.global_block().vars: + fc_w_var = startup.global_block().vars["fc_w"] + elif "fc_w" in startup2.global_block().vars: + fc_w_var = startup2.global_block().vars["fc_w"] - def get_expect_trainer_ops(self): - trainer = fluid.Program() + self.assertEqual(fc_w_var.shape, (1000, 1000)) - with fluid.program_guard(trainer): - optimize_ops, params_grads = self.net_conf() - delete_ops(trainer.global_block(), optimize_ops) - ops = [op.type for op in trainer.global_block().ops] + [ - "split_byref", "send_vars", "send_barrier", "recv", "recv", - "fetch_barrier", "concat" +class TestLRDecay(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.exponential_decay( + learning_rate=1.0, + decay_steps=2100, + decay_rate=0.1, + staircase=True)) + sgd_optimizer.minimize(avg_cost) + return + + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + trainer, _ = self.get_trainer() + + self.assertEqual(len(pserver.blocks), 4) + lr_decay_ops = [op.type for op in pserver.blocks[1].ops] + self.assertEqual(lr_decay_ops, [ + "increment", "cast", "fill_constant", "elementwise_div", "floor", + "fill_constant", "elementwise_pow", "fill_constant", + "elementwise_mul" + ]) + + +class TestLRDecayConditional(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.piecewise_decay([10000, 20000], + [1.0, 0.5, 1.0])) + sgd_optimizer.minimize(avg_cost) + return + + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + trainer, _ = self.get_trainer() + + serv_op = pserver.blocks[0].ops[0] + sub_blocks = [] + optimize_blocks = [] + for b in serv_op.all_attrs()["optimize_blocks"]: + optimize_blocks.append(b.idx) + for b in pserver.blocks: + if b.idx not in optimize_blocks: + sub_blocks.append(b.idx) + + self.assertEqual(len(pserver.blocks), 7) + lr_decay_ops = [op.type for op in pserver.blocks[1].ops] + self.assertEqual(lr_decay_ops, [ + "increment", "cast", "fill_constant", "fill_constant", "less_than", + "logical_not", "conditional_block", "fill_constant", + "fill_constant", "less_than", "logical_not", "logical_and", + "logical_and", "conditional_block", "fill_constant", + "conditional_block" + ]) + # test the condition blocks + for b in sub_blocks: + if b == 0: + continue + block = pserver.blocks[b] + self.assertEqual([op.type for op in block.ops], ["assign"]) + + +class TestL2Decay(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc( + input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr( + name='fc_w', + regularizer=fluid.regularizer.L2Decay(), + gradient_clip=fluid.clip.GradientClipByValue(0.1)), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) + sgd_optimizer.minimize(avg_cost) + return + + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + trainer, _ = self.get_trainer() + + self.assertEqual(len(pserver.blocks), 3) + self.assertEqual([op.type for op in pserver.blocks[1].ops], + ["sum", "scale", "clip", "sgd"]) + self.assertEqual( + [op.type for op in pserver.blocks[2].ops], + ["sum", "scale", "clip", "scale", "elementwise_add", "sgd"]) + # TODO(typhoonzero): test clipping and L2Decay ops are removed from trainer + + +class TestL2DecayWithPiecewise(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + base_lr = 1.0 + bd = [1, 10, 20, 30] + lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)] + sgd_optimizer = fluid.optimizer.Momentum( + learning_rate=fluid.layers.piecewise_decay( + boundaries=bd, values=lr), + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + sgd_optimizer.minimize(avg_cost) + return + + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + trainer, _ = self.get_trainer() + + self.assertEqual(len(pserver.blocks), 9) + self.assertEqual([op.type for op in pserver.blocks[1].ops], [ + "increment", "cast", "fill_constant", "fill_constant", "less_than", + "logical_not", "conditional_block", "fill_constant", + "fill_constant", "less_than", "logical_not", "logical_and", + "logical_and", "conditional_block", "fill_constant", + "fill_constant", "less_than", "logical_not", "logical_and", + "logical_and", "conditional_block", "fill_constant", + "fill_constant", "less_than", "logical_not", "logical_and", + "logical_and", "conditional_block", "fill_constant", + "conditional_block" + ]) + self.assertEqual( + [op.type for op in pserver.blocks[7].ops], + ["sum", "scale", "scale", "elementwise_add", "momentum"]) + self.assertEqual( + [op.type for op in pserver.blocks[8].ops], + ["sum", "scale", "scale", "elementwise_add", "momentum"]) + + +class TestDistLookupTableBase(TranspilerTest): + def network_with_table(self, is_sparse, is_distributed): + self.table_size = 1000 + self.emb_size = 64 + + def emb_pool(ids): + emb = fluid.layers.embedding( + input=ids, + size=[self.table_size, self.emb_size], + dtype='float32', + param_attr='shared_w', # share parameter + is_sparse=is_sparse, + is_distributed=is_distributed) + pool = fluid.layers.sequence_pool(input=emb, pool_type='average') + return pool + + title_ids = fluid.layers.data( + name='title_ids', shape=[1], dtype='int64', lod_level=1) + brand_ids = fluid.layers.data( + name='brand_ids', shape=[1], dtype='int64', lod_level=1) + title_emb = emb_pool(title_ids) + brand_emb = emb_pool(brand_ids) + fc0 = fluid.layers.concat(input=[title_emb, brand_emb], axis=1) + predict = fluid.layers.fc(input=fc0, + size=2, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + optimizer = fluid.optimizer.Adam(learning_rate=0.003) + optimizer.minimize(avg_cost) + + +class TestLocalLookupTable(TestDistLookupTableBase): + def net_conf(self): + self.network_with_table(is_sparse=True, is_distributed=False) + + def transpiler_test_impl(self): + pserver1, startup1 = self.get_pserver(self.pserver1_ep) + + self.assertEqual(len(pserver1.blocks), 3) + # 0 listen_and_serv + # 1 optimize for fc_w or fc_b adam + self.assertEqual([op.type for op in pserver1.blocks[1].ops], + ["sum", "scale", "adam", "scale", "scale"]) + # 2 optimize for table adam + # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num + self.assertEqual([op.type for op in pserver1.blocks[2].ops], + ["sum", "adam", "scale", "scale"]) + + trainer, _ = self.get_trainer() + self.assertEqual(len(trainer.blocks), 1) + ops = [ + 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', + 'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', + 'fill_constant', 'mean_grad', 'cross_entropy_grad', + 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', + 'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', + 'lookup_table_grad', 'sum', 'split_selected_rows', 'send', + 'send_barrier', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat' ] - ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars") - return ops + self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) - def get_trainer(self): - return self._transpiler_instance().get_trainer_program() - def get_pserver(self, ep): - t = self._transpiler_instance() - pserver = t.get_pserver_program(ep) - startup = t.get_startup_program(ep, pserver) - return pserver, startup +class TestDistLookupTable(TestDistLookupTableBase): + def net_conf(self): + self.network_with_table(is_sparse=True, is_distributed=True) - def _transpiler_instance(self): - main = self.get_main_program() - t = fluid.DistributeTranspiler() - t.transpile( - self.trainer_id, - program=main, - pservers=self.pserver_eps, - trainers=self.trainers) - return t + def transpiler_test_impl(self): + pserver1, startup1 = self.get_pserver(self.pserver1_ep) + + self.assertEqual(len(pserver1.blocks), 6) + # 0 listen_and_serv + # 1 optimize for fc_w or fc_b adam + self.assertEqual([op.type for op in pserver1.blocks[1].ops], + ["sum", "scale", "adam", "scale", "scale"]) + # 2 optimize for table sgd + self.assertEqual([op.type for op in pserver1.blocks[2].ops], + ["sum", "sgd"]) + # 3 prefetch -> lookup_sparse_table for data0 + self.assertEqual([op.type for op in pserver1.blocks[3].ops], + ["lookup_sparse_table"]) + # 4 prefetch -> lookup_sparse_table for data1 + self.assertEqual([op.type for op in pserver1.blocks[4].ops], + ["lookup_sparse_table"]) + # 5 save table + self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"]) + + trainer, _ = self.get_trainer() + self.assertEqual(len(trainer.blocks), 1) + ops = [ + 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', 'split_ids', + 'prefetch', 'merge_ids', 'sequence_pool', 'concat', 'mul', + 'elementwise_add', 'cross_entropy', 'mean', 'fill_constant', + 'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send', + 'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad', + 'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad', + 'sum', 'split_ids', 'send', 'send_barrier', 'recv', 'recv', + 'fetch_barrier' + ] + self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) + + +class TestAsyncLocalLookupTable(TestDistLookupTableBase): + def net_conf(self): + self.network_with_table(is_sparse=True, is_distributed=False) + + def transpiler_test_impl(self): + config = fluid.DistributeTranspilerConfig() + pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False) + + self.assertEqual(len(pserver1.blocks), 3) + # 0 listen_and_serv + # 1 optimize for fc_w or fc_b adam + self.assertEqual([op.type for op in pserver1.blocks[1].ops], + ["adam", "scale", "scale"]) + # 2 optimize for table adam + # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num + self.assertEqual([op.type for op in pserver1.blocks[2].ops], + ["adam", "scale", "scale"]) + + trainer, _ = self.get_trainer(config) + self.assertEqual(len(trainer.blocks), 1) + ops = [ + 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', + 'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', + 'fill_constant', 'mean_grad', 'cross_entropy_grad', + 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', + 'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', + 'lookup_table_grad', 'sum', 'split_selected_rows', 'send', 'recv', + 'recv', 'recv', 'concat' + ] + self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) + + +class TestAsyncDistLookupTable(TestDistLookupTableBase): + def net_conf(self): + self.network_with_table(is_sparse=True, is_distributed=True) + + def transpiler_test_impl(self): + config = fluid.DistributeTranspilerConfig() + + pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False) + + self.assertEqual(len(pserver1.blocks), 6) + # 0 listen_and_serv + # 1 optimize for fc_w or fc_b adam + self.assertEqual([op.type for op in pserver1.blocks[1].ops], + ["adam", "scale", "scale"]) + # 2 optimize for table sgd + self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["sgd"]) + # 3 prefetch -> lookup_sparse_table for data0 + self.assertEqual([op.type for op in pserver1.blocks[3].ops], + ["lookup_sparse_table"]) + # 4 prefetch -> lookup_sparse_table for data1 + self.assertEqual([op.type for op in pserver1.blocks[4].ops], + ["lookup_sparse_table"]) + # 5 save table + self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"]) + + trainer, _ = self.get_trainer(config) + self.assertEqual(len(trainer.blocks), 1) + ops = [ + 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', 'split_ids', + 'prefetch', 'merge_ids', 'sequence_pool', 'concat', 'mul', + 'elementwise_add', 'cross_entropy', 'mean', 'fill_constant', + 'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send', + 'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad', + 'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad', + 'sum', 'split_ids', 'send', 'recv', 'recv' + ] + self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) + + +class TestDistLookupTableSliceSize(TestDistLookupTableBase): + def net_conf(self): + self.network_with_table(is_sparse=True, is_distributed=True) + + def transpiler_test_impl(self): + config = fluid.DistributeTranspilerConfig() + pserver1, startup1 = self.get_pserver(self.pserver1_ep, config) + + self.assertTrue(self.transpiler.has_distributed_lookup_table) + lookup_table_var = pserver1.global_block().vars[ + self.transpiler.table_name] + row_size = lookup_table_var.shape[0] + calc_row_size = int(math.ceil(self.table_size / self.pservers)) + self.assertEqual(row_size, calc_row_size) + + +class TestRMSPropOptimizer(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + optimizer = fluid.optimizer.RMSProp(learning_rate=0.1) + optimizer.minimize(avg_cost) + return + + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + pserver2, startup2 = self.get_pserver(self.pserver2_ep) + + self.assertEqual(len(pserver.blocks), 3) + # block1~2: optimize pass + self.assertEqual([op.type for op in pserver.blocks[1].ops], + ["sum", "scale", "rmsprop"]) + # the variable #fc_w will be split into two blocks + fc_w_var = startup.global_block().var("fc_w.block1") + self.assertEqual(fc_w_var.shape, (500, 1000)) + moment_var = startup.global_block().var("momentum_1") + self.assertEqual(moment_var.shape, (500, 1000)) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_dist_word2vec.py b/python/paddle/fluid/tests/unittests/test_dist_word2vec.py new file mode 100644 index 0000000000000000000000000000000000000000..543d0f9dc2c9b8cdcfaaaa14a7a4f197d210d951 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_word2vec.py @@ -0,0 +1,24 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import unittest +from test_dist_base import TestDistBase + + +class TestDistSeResneXt2x2(TestDistBase): + def test_se_resnext(self): + self.check_with_place("dist_word2vec.py", delta=1e-7) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py index 0faed94deb4808783027d776e0f4c61da0db457a..fdc6adc93bc2488d4faffed61fde5d54bbbbfd57 100644 --- a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py +++ b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py @@ -17,6 +17,12 @@ import paddle import unittest import numpy +from paddle.fluid.layers.control_flow import lod_rank_table +from paddle.fluid.layers.control_flow import max_sequence_len +from paddle.fluid.layers.control_flow import lod_tensor_to_array +from paddle.fluid.layers.control_flow import array_to_lod_tensor +from paddle.fluid.layers.control_flow import shrink_memory + class TestDynRNN(unittest.TestCase): def setUp(self): @@ -38,12 +44,11 @@ class TestDynRNN(unittest.TestCase): label = fluid.layers.data(name='label', shape=[1], dtype='float32') - rank_table = fluid.layers.lod_rank_table(x=sent_emb) + rank_table = lod_rank_table(x=sent_emb) - sent_emb_array = fluid.layers.lod_tensor_to_array( - x=sent_emb, table=rank_table) + sent_emb_array = lod_tensor_to_array(x=sent_emb, table=rank_table) - seq_len = fluid.layers.max_sequence_len(rank_table=rank_table) + seq_len = max_sequence_len(rank_table=rank_table) i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=0) i.stop_gradient = False @@ -66,7 +71,7 @@ class TestDynRNN(unittest.TestCase): mem = fluid.layers.array_read(array=mem_array, i=i) ipt = fluid.layers.array_read(array=sent_emb_array, i=i) - mem = fluid.layers.shrink_memory(x=mem, i=i, table=rank_table) + mem = shrink_memory(x=mem, i=i, table=rank_table) hidden = fluid.layers.fc(input=[mem, ipt], size=100, act='tanh') @@ -75,8 +80,7 @@ class TestDynRNN(unittest.TestCase): fluid.layers.array_write(x=hidden, i=i, array=mem_array) fluid.layers.less_than(x=i, y=seq_len, cond=cond) - all_timesteps = fluid.layers.array_to_lod_tensor( - x=out, table=rank_table) + all_timesteps = array_to_lod_tensor(x=out, table=rank_table) last = fluid.layers.sequence_last_step(input=all_timesteps) logits = fluid.layers.fc(input=last, size=1, act=None) loss = fluid.layers.sigmoid_cross_entropy_with_logits( @@ -131,7 +135,7 @@ class TestDynRNN(unittest.TestCase): loss_0 = exe.run(main_program, feed=feeder.feed(data), fetch_list=[loss])[0] - for _ in xrange(100): + for _ in range(100): val = exe.run(main_program, feed=feeder.feed(data), fetch_list=[loss])[0] diff --git a/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py b/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py index 22329390754d8d010dced0d1aca35617140cd097..7756885166c88eadb77c2c6d56aab767015abc51 100644 --- a/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py +++ b/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py @@ -30,9 +30,6 @@ class Memory(object): assert val.dtype == self.ex.dtype self.cur = val - def ex(self): - return self.ex - def next(self): self.ex = self.cur self.cur = None @@ -64,13 +61,13 @@ class BaseRNN(object): self.num_seq = num_seq self.inputs = collections.defaultdict(list) - for _ in xrange(num_seq): + for _ in range(num_seq): seq_len = random.randint(1, max_seq_len - 1) for iname in ins: ishape = ins[iname].get('shape', None) idtype = ins[iname].get('dtype', 'float32') lst = [] - for _ in xrange(seq_len): + for _ in range(seq_len): lst.append(numpy.random.random(size=ishape).astype(idtype)) self.inputs[iname].append(lst) @@ -99,16 +96,16 @@ class BaseRNN(object): for out in self.outputs: retv[out] = [] - for seq_id in xrange(self.num_seq): + for seq_id in range(self.num_seq): for mname in self.mems: self.mems[mname].reset() for out in self.outputs: self.outputs[out].next_sequence() - iname0 = self.inputs.keys()[0] + iname0 = list(self.inputs.keys())[0] seq_len = len(self.inputs[iname0][seq_id]) - for step_id in xrange(seq_len): + for step_id in range(seq_len): xargs = dict() for iname in self.inputs: @@ -139,16 +136,16 @@ class BaseRNN(object): feed_dict = dict() for iname in self.inputs: - lod = [0] + lod = [] np_flatten = [] - for seq_id in xrange(len(self.inputs[iname])): + for seq_id in range(len(self.inputs[iname])): seq_len = len(self.inputs[iname][seq_id]) - lod.append(lod[-1] + seq_len) + lod.append(seq_len) np_flatten.extend(self.inputs[iname][seq_id]) t = fluid.Tensor() t.set(numpy.array(np_flatten), place) - t.set_lod([lod]) + t.set_recursive_sequence_lengths([lod]) feed_dict[iname] = t for pname in self.params: @@ -162,8 +159,8 @@ class BaseRNN(object): " which is not matrix") g = numpy.zeros(shape=p.shape, dtype=p.dtype) - for i in xrange(p.shape[0]): - for j in xrange(p.shape[1]): + for i in range(p.shape[0]): + for j in range(p.shape[1]): o = p[i][j] p[i][j] += delta pos = self._exe_mean_out_() @@ -187,7 +184,7 @@ class BaseRNN(object): if len(item.shape) != 1: raise ValueError("Not support") - for i in xrange(len(item)): + for i in range(len(item)): o = item[i] item[i] += delta pos = self._exe_mean_out_() @@ -201,14 +198,14 @@ class BaseRNN(object): if not return_one_tensor: return grad - for i in xrange(len(grad)): + for i in range(len(grad)): grad[i] = numpy.concatenate(grad[i]) grad = numpy.concatenate(grad) return grad def _exe_mean_out_(self): outs = self.exe() - return numpy.array([o.mean() for o in outs.itervalues()]).mean() + return numpy.array([o.mean() for o in outs.values()]).mean() class SeedFixedTestCase(unittest.TestCase): @@ -277,13 +274,14 @@ class TestSimpleMul(SeedFixedTestCase): cpu = fluid.CPUPlace() exe = fluid.Executor(cpu) - out, w_g, i_g = map(numpy.array, - exe.run(feed=py_rnn.to_feed(cpu), - fetch_list=[ - out, self.PARAM_NAME + "@GRAD", - self.DATA_NAME + "@GRAD" - ], - return_numpy=False)) + out, w_g, i_g = list( + map(numpy.array, + exe.run(feed=py_rnn.to_feed(cpu), + fetch_list=[ + out, self.PARAM_NAME + "@GRAD", self.DATA_NAME + + "@GRAD" + ], + return_numpy=False))) out_by_python = py_rnn.exe()[self.OUT_NAME] self.assertTrue(numpy.allclose(out, out_by_python)) w_g_num = py_rnn.get_numeric_gradient_of_param(self.PARAM_NAME) @@ -354,14 +352,15 @@ class TestSimpleMulWithMemory(SeedFixedTestCase): cpu = fluid.CPUPlace() exe = fluid.Executor(cpu) feed = py_rnn.to_feed(cpu) - last_np, w_g, i_g = map(numpy.array, - exe.run(feed=feed, - fetch_list=[ - last, self.PARAM_NAME + "@GRAD", - self.DATA_NAME + "@GRAD" - ], - return_numpy=False)) - last_by_py, = py_rnn.exe().values() + last_np, w_g, i_g = list( + map(numpy.array, + exe.run(feed=feed, + fetch_list=[ + last, self.PARAM_NAME + "@GRAD", self.DATA_NAME + + "@GRAD" + ], + return_numpy=False))) + last_by_py, = list(py_rnn.exe().values()) w_g_num = py_rnn.get_numeric_gradient_of_param(self.PARAM_NAME) self.assertTrue(numpy.allclose(last_np, last_by_py)) diff --git a/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py b/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py index d3f63ee2c414a71309be8f0af6d3e5912078ecdb..d182889a970fb178dec4976aebbd79d05dc3e91e 100644 --- a/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py +++ b/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py @@ -39,20 +39,20 @@ class TestDyRnnStaticInput(unittest.TestCase): def prepare_x_tensor(self): self.x_tensor_dim = 10 - lod = [[0, 2, 3, 6]] - shape = [lod[0][-1], self.x_tensor_dim] + lod = [[2, 1, 3]] + shape = [sum(lod[0]), self.x_tensor_dim] self.x_tensor_data = np.random.random(shape).astype('float32') self.x_tensor = core.LoDTensor() - self.x_tensor.set_lod(lod) + self.x_tensor.set_recursive_sequence_lengths(lod) self.x_tensor.set(self.x_tensor_data, self.place) def prepare_static_input_tensor(self): self.static_input_tensor_dim = 4 - lod = [[0, 1, 3, 6]] - shape = [lod[0][-1], self.static_input_tensor_dim] + lod = [[1, 2, 3]] + shape = [sum(lod[0]), self.static_input_tensor_dim] self.static_input_data = np.random.random(shape).astype('float32') self.static_input_tensor = core.LoDTensor() - self.static_input_tensor.set_lod(lod) + self.static_input_tensor.set_recursive_sequence_lengths(lod) self.static_input_tensor.set(self.static_input_data, self.place) def fetch_value(self, var): @@ -65,11 +65,11 @@ class TestDyRnnStaticInput(unittest.TestCase): return self._lodtensor_to_ndarray(fetch_outs[0]) def _lodtensor_to_ndarray(self, lod_tensor): - dims = lod_tensor.get_dims() + dims = lod_tensor.shape() ndarray = np.zeros(shape=dims).astype('float32') - for i in xrange(np.product(dims)): - ndarray.ravel()[i] = lod_tensor.get_float_element(i) - return ndarray, lod_tensor.lod() + for i in range(np.product(dims)): + ndarray.ravel()[i] = lod_tensor._get_float_element(i) + return ndarray, lod_tensor.recursive_sequence_lengths() def build_graph(self, only_forward=False): x_tensor = fluid.layers.data( @@ -114,7 +114,7 @@ class TestDyRnnStaticInput(unittest.TestCase): shape=[1], dtype='int64', value=0) step_idx.stop_gradient = True - for i in xrange(self._max_sequence_len): + for i in range(self._max_sequence_len): step_out = fluid.layers.array_read(static_input_out_array, step_idx) step_out.stop_gradient = True @@ -131,39 +131,40 @@ class TestDyRnnStaticInput(unittest.TestCase): framework.grad_var_name('static_input_tensor')) return static_input_grad, loss - def get_seq_len_from_lod(self, lod): - return [lod[0][i + 1] - lod[0][i] for i in xrange(len(lod[0]) - 1)] - def get_expected_static_step_outs(self): - x_lod = self.x_tensor.lod() - x_seq_len = self.get_seq_len_from_lod(x_lod) + x_lod = self.x_tensor.recursive_sequence_lengths() + x_seq_len = x_lod[0] x_seq_len_sorted = sorted(x_seq_len) x_sorted_indices = np.argsort(x_seq_len)[::-1] - static_lod = self.static_input_tensor.lod() - static_sliced = [ - self.static_input_data[static_lod[0][i]:static_lod[0][i + 1]] - for i in xrange(len(static_lod[0]) - 1) - ] - static_seq_len = self.get_seq_len_from_lod(static_lod) + static_lod = self.static_input_tensor.recursive_sequence_lengths() + static_sliced = [] + cur_offset = 0 + for i in range(len(static_lod[0])): + static_sliced.append(self.static_input_data[cur_offset:( + cur_offset + static_lod[0][i])]) + cur_offset += static_lod[0][i] + static_seq_len = static_lod[0] static_reordered = [] - for i in xrange(len(x_sorted_indices)): + for i in range(len(x_sorted_indices)): static_reordered.extend(static_sliced[x_sorted_indices[i]].tolist()) static_seq_len_reordered = [ static_seq_len[x_sorted_indices[i]] - for i in xrange(len(x_sorted_indices)) + for i in range(len(x_sorted_indices)) ] static_step_outs = [] static_step_lods = [] - for i in xrange(self._max_sequence_len): + for i in range(self._max_sequence_len): end = len(x_seq_len) - bisect.bisect_left(x_seq_len_sorted, i + 1) - lod = [0] - for i in xrange(end): - lod.append(static_seq_len_reordered[i] + lod[-1]) + lod = [] + total_len = 0 + for i in range(end): + lod.append(static_seq_len_reordered[i]) + total_len += lod[-1] static_step_lods.append([lod]) - end = lod[-1] + end = total_len static_step_outs.append( np.array(static_reordered[:end]).astype('float32')) @@ -173,7 +174,7 @@ class TestDyRnnStaticInput(unittest.TestCase): static_step_outs = self.build_graph(only_forward=True) self.exe.run(framework.default_startup_program()) expected_outs, expected_lods = self.get_expected_static_step_outs() - for i in xrange(self._max_sequence_len): + for i in range(self._max_sequence_len): step_out, lod = self.fetch_value(static_step_outs[i]) self.assertTrue(np.allclose(step_out, expected_outs[i])) self.assertTrue(np.allclose(lod, expected_lods[i])) @@ -184,22 +185,24 @@ class TestDyRnnStaticInput(unittest.TestCase): actual_gradients, actual_lod = self.fetch_value(static_input_grad) - static_input_shape = self.static_input_tensor.get_dims() + static_input_shape = self.static_input_tensor.shape() numeric_gradients = np.zeros(shape=static_input_shape).astype('float32') # calculate numeric gradients tensor_size = np.product(static_input_shape) - for i in xrange(tensor_size): - origin = self.static_input_tensor.get_float_element(i) + for i in range(tensor_size): + origin = self.static_input_tensor._get_float_element(i) x_pos = origin + self._delta - self.static_input_tensor.set_float_element(i, x_pos) + self.static_input_tensor._set_float_element(i, x_pos) y_pos = self.fetch_value(loss)[0][0] x_neg = origin - self._delta - self.static_input_tensor.set_float_element(i, x_neg) + self.static_input_tensor._set_float_element(i, x_neg) y_neg = self.fetch_value(loss)[0][0] - self.static_input_tensor.set_float_element(i, origin) + self.static_input_tensor._set_float_element(i, origin) numeric_gradients.ravel()[i] = (y_pos - y_neg) / self._delta / 2 self.assertTrue(np.allclose(actual_gradients, numeric_gradients, 0.001)) - self.assertTrue(np.allclose(actual_lod, self.static_input_tensor.lod())) + self.assertTrue( + np.allclose(actual_lod, + self.static_input_tensor.recursive_sequence_lengths())) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_edit_distance_op.py b/python/paddle/fluid/tests/unittests/test_edit_distance_op.py index 2957fb50586c8bce74bbf8066e0e9bf24d79cb7d..816562621b4fc749f3c6b0eca8ee3c5850ef1ba9 100644 --- a/python/paddle/fluid/tests/unittests/test_edit_distance_op.py +++ b/python/paddle/fluid/tests/unittests/test_edit_distance_op.py @@ -52,23 +52,29 @@ class TestEditDistanceOp(OpTest): def setUp(self): self.op_type = "edit_distance" normalized = False - x1 = np.array([[0, 12, 3, 5, 8, 2]]).astype("int64") - x2 = np.array([[0, 12, 4, 7, 8]]).astype("int64") + x1 = np.array([[12, 3, 5, 8, 2]]).astype("int64") + x2 = np.array([[12, 4, 7, 8]]).astype("int64") x1 = np.transpose(x1) x2 = np.transpose(x2) - x1_lod = [0, 1, 5] - x2_lod = [0, 3, 4] + x1_lod = [1, 4] + x2_lod = [3, 1] - num_strs = len(x1_lod) - 1 + num_strs = len(x1_lod) distance = np.zeros((num_strs, 1)).astype("float32") sequence_num = np.array(2).astype("int64") + + x1_offset = 0 + x2_offset = 0 for i in range(0, num_strs): distance[i] = Levenshtein( - hyp=x1[x1_lod[i]:x1_lod[i + 1]], - ref=x2[x2_lod[i]:x2_lod[i + 1]]) + hyp=x1[x1_offset:(x1_offset + x1_lod[i])], + ref=x2[x2_offset:(x2_offset + x2_lod[i])]) + x1_offset += x1_lod[i] + x2_offset += x2_lod[i] if normalized is True: - len_ref = x2_lod[i + 1] - x2_lod[i] + len_ref = x2_lod[i] distance[i] = distance[i] / len_ref + self.attrs = {'normalized': normalized} self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])} self.outputs = {'Out': distance, 'SequenceNum': sequence_num} @@ -81,23 +87,29 @@ class TestEditDistanceOpNormalized(OpTest): def setUp(self): self.op_type = "edit_distance" normalized = True - x1 = np.array([[0, 10, 3, 6, 5, 8, 2]]).astype("int64") - x2 = np.array([[0, 10, 4, 6, 7, 8]]).astype("int64") + x1 = np.array([[10, 3, 6, 5, 8, 2]]).astype("int64") + x2 = np.array([[10, 4, 6, 7, 8]]).astype("int64") x1 = np.transpose(x1) x2 = np.transpose(x2) - x1_lod = [0, 1, 3, 6] - x2_lod = [0, 2, 3, 5] + x1_lod = [1, 2, 3] + x2_lod = [2, 1, 2] - num_strs = len(x1_lod) - 1 + num_strs = len(x1_lod) distance = np.zeros((num_strs, 1)).astype("float32") sequence_num = np.array(3).astype("int64") + + x1_offset = 0 + x2_offset = 0 for i in range(0, num_strs): distance[i] = Levenshtein( - hyp=x1[x1_lod[i]:x1_lod[i + 1]], - ref=x2[x2_lod[i]:x2_lod[i + 1]]) + hyp=x1[x1_offset:(x1_offset + x1_lod[i])], + ref=x2[x2_offset:(x2_offset + x2_lod[i])]) + x1_offset += x1_lod[i] + x2_offset += x2_lod[i] if normalized is True: - len_ref = x2_lod[i + 1] - x2_lod[i] + len_ref = x2_lod[i] distance[i] = distance[i] / len_ref + self.attrs = {'normalized': normalized} self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])} self.outputs = {'Out': distance, 'SequenceNum': sequence_num} diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py new file mode 100644 index 0000000000000000000000000000000000000000..bcdbfc8e527d0dc9a95eddaf040f8035207b6c20 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py @@ -0,0 +1,130 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import unittest +import numpy as np +import paddle.fluid.core as core +from op_test import OpTest +from test_elementwise_add_op import * +''' +Some tests differ from the tests defined in test_elementwise_add_op.py +because MKLDNN does not support tensors of number of dimensions 3. +Such dimensions cause exceptions in MKLDNN reorder primitive. +''' + + +class TestMKLDNNElementwiseAddOp(TestElementwiseAddOp): + def init_input_output(self): + self.x = np.random.uniform(0.1, 1, [2, 3, 4, 5]).astype(self.dtype) + self.y = np.random.uniform(0.1, 1, [2, 3, 4, 5]).astype(self.dtype) + self.out = np.add(self.x, self.y) + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_scalar(TestElementwiseAddOp_scalar): + def init_input_output(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(1).astype(self.dtype) + self.out = self.x + self.y + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_scalar2(TestElementwiseAddOp_scalar2): + def init_input_output(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(1, 1).astype(self.dtype) + self.out = self.x + self.y + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_Vector(TestElementwiseAddOp_Vector): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TesMKLDNNtElementwiseAddOp_broadcast_0(TestElementwiseAddOp_broadcast_0): + def init_input_output(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(2).astype(self.dtype) + self.out = self.x + self.y.reshape(2, 1, 1, 1) + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_broadcast_1(TestElementwiseAddOp_broadcast_1): + def init_input_output(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(3).astype(self.dtype) + self.out = self.x + self.y.reshape(1, 3, 1, 1) + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_broadcast_2(TestElementwiseAddOp_broadcast_2): + def init_input_output(self): + self.x = np.random.rand(2, 2, 3, 4).astype(self.dtype) + self.y = np.random.rand(4).astype(self.dtype) + self.out = self.x + self.y.reshape(1, 1, 1, 4) + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_broadcast_3(TestElementwiseAddOp_broadcast_3): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_broadcast_4(TestElementwiseAddOp_broadcast_4): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_rowwise_add_0( + TestElementwiseAddOp_rowwise_add_0): + def init_input_output(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(3, 4).astype(self.dtype) + self.out = self.x + self.y.reshape(1, 3, 4, 1) + + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_rowwise_add_1( + TestElementwiseAddOp_rowwise_add_1): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNElementwiseAddOp_channelwise_add( + TestElementwiseAddOp_channelwise_add): + def init_input_output(self): + self.x = np.random.rand(3, 5, 20, 20).astype(self.dtype) + self.y = np.random.rand(3, 1, 1, 1).astype(self.dtype) + self.out = self.x + self.y + + def init_kernel_type(self): + self.use_mkldnn = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py index 1f52bd90d0d49bda6c180019e90ebd923c91439c..fb9a496126f0b6efcad73590c78efe5a47b88cd6 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py @@ -18,19 +18,23 @@ from op_test import OpTest class TestElementwiseAddOp(OpTest): + def init_kernel_type(self): + self.use_mkldnn = False + def setUp(self): self.op_type = "elementwise_add" self.dtype = np.float32 self.axis = -1 self.init_dtype() self.init_input_output() + self.init_kernel_type() self.init_axis() self.inputs = { 'X': OpTest.np_dtype_to_fluid_dtype(self.x), 'Y': OpTest.np_dtype_to_fluid_dtype(self.y) } - self.attrs = {'axis': self.axis} + self.attrs = {'axis': self.axis, 'use_mkldnn': self.use_mkldnn} self.outputs = {'Out': self.out} def test_check_output(self): @@ -252,5 +256,25 @@ class TestFP16ElementwiseAddOp_rowwise_add_1(TestFP16ElementwiseAddOp): self.axis = 1 +class TestElementwiseAddOp_channelwise_add(TestElementwiseAddOp): + def init_input_output(self): + self.x = np.random.rand(3, 20, 20).astype(self.dtype) + self.y = np.random.rand(3, 1, 1).astype(self.dtype) + self.out = self.x + self.y + + def init_axis(self): + self.axis = -1 + + +class TestFP16ElementwiseAddOp_channelwise_add(TestFP16ElementwiseAddOp): + def init_input_output(self): + self.x = np.random.rand(3, 10, 20).astype(self.dtype) + self.y = np.random.rand(3, 1, 1).astype(self.dtype) + self.out = self.x + self.y + + def init_axis(self): + self.axis = -1 + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py index c6f45381af8ac64d117eb27325f25763fbf6cae7..6f350044892a4ba2a985b5bc2328ab1fc20c5504 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py @@ -26,7 +26,7 @@ class TestElementWiseAddOp(unittest.TestCase): def test_with_place(place): out_grad = np.random.random_sample(self.x.shape).astype(np.float32) x_grad = out_grad - sum_axis = range(0, len(self.x.shape)) + sum_axis = list(range(0, len(self.x.shape))) del sum_axis[self.axis] y_grad = np.sum(out_grad, axis=tuple(sum_axis)) diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py index acf652d3fb9743d69b7f7e248ff7a3ee83fc4c50..1854232194963bcbe302010320a30d85747eea96 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py @@ -20,8 +20,8 @@ class TestElementwiseOp(OpTest): def setUp(self): self.op_type = "elementwise_sub" self.inputs = { - 'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"), - 'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32") + 'X': np.random.uniform(0.1, 1, [2, 3]).astype("float32"), + 'Y': np.random.uniform(0.1, 1, [2, 3]).astype("float32") } self.outputs = {'Out': self.inputs['X'] - self.inputs['Y']} diff --git a/python/paddle/fluid/tests/unittests/test_extract_rows_op.py b/python/paddle/fluid/tests/unittests/test_extract_rows_op.py new file mode 100644 index 0000000000000000000000000000000000000000..6a41c44fe655b18626bdb727745dae032babe8ad --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_extract_rows_op.py @@ -0,0 +1,58 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import paddle.fluid.core as core +from paddle.fluid.op import Operator +from op_test import OpTest + + +class TestExtractRows(OpTest): + def check_with_place(self, place): + scope = core.Scope() + + # create and initialize Variable + feature_len = 12 + rows = [0, 4, 4, 7] + np_array = np.ones((len(rows), feature_len)).astype("float32") + + in_x = scope.var('X').get_selected_rows() + in_x.set_height(len(rows)) + in_x.set_rows(rows) + in_x_tensor = in_x.get_tensor() + in_x_tensor.set(np_array, place) + + # create Out Variable + out_tensor = scope.var('Out').get_tensor() + + # create and run lookup_table operator + extract_rows_op = Operator("extract_rows", X='X', Out='Out') + extract_rows_op.run(scope, place) + + # get result from Out + result_array = np.array(out_tensor) + result_array = [ele[0] for ele in result_array] + assert result_array == rows + + def test_concat_rows(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py b/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py index 281068e945e76a42635868d19573498f79fde1f3..026ac2112b2d78644b3315b9cab8019ca27e9714 100644 --- a/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py +++ b/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py @@ -40,7 +40,6 @@ class TestFakeDequantizeMaxAbsOp(OpTest): self.op_type = "fake_dequantize_max_abs" x = np.random.randn(31, 65).astype("float32") yq, scale = quantize_max_abs(x, self.num_bits) - print 'scale ', scale ydq = dequantize_max_abs(yq, self.num_bits, scale) self.inputs = {'X': yq} diff --git a/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py b/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py new file mode 100644 index 0000000000000000000000000000000000000000..6c6aa9d3bb656740c528c728efafc6a47e8bff91 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py @@ -0,0 +1,51 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestFakeQuantizeOp(OpTest): + def setUp(self): + self.op_type = "fake_quantize" + self.attrs = { + 'bit_length': 8, + 'quantize_type': 'abs_max', + 'window_size': 10000 + } + self.inputs = { + 'X': np.random.random((10, 10)).astype("float32"), + 'InScales': np.zeros(self.attrs['window_size']).astype("float32"), + 'InCurrentIter': np.zeros(1).astype("float32"), + 'InMovingScale': np.zeros(1).astype("float32") + } + self.scale = { + 'abs_max': np.max(np.abs(self.inputs['X'])).astype("float32") + } + self.outputs = { + 'Out': np.round(self.inputs['X'] / self.scale['abs_max'] * ( + (1 << (self.attrs['bit_length'] - 1)) - 1)), + 'OutScales': np.zeros(self.attrs['window_size']).astype("float32"), + 'OutMovingScale': + np.array([self.scale['abs_max']]).astype("float32"), + 'OutCurrentIter': np.zeros(1).astype("float32") + } + + def test_check_output(self): + self.check_output() + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py index 3f547f3c484bf034a87823a75d946ef130a5cb70..099e6e60642e9637f8f3648696e844c667e1c406 100644 --- a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py @@ -22,6 +22,7 @@ def fully_connected_naive(input, weights, bias_data=None): w_h, w_c = weights.shape x_data = np.reshape(input, [in_n, in_c * in_h * in_w]) + # this transpose should be implemented at C code w_data = np.transpose(np.reshape(weights, (w_c, in_c * in_h * in_w))) result = None @@ -43,15 +44,11 @@ class TestFCMKLDNNOp(OpTest): def setUp(self): self.op_type = "fc" self.use_mkldnn = True - self.with_bias = True self.matrix = MatrixGenerate(1, 10, 15, 3, 3) self.inputs = {'Input': self.matrix.input, 'W': self.matrix.weights} - self.attrs = { - 'use_mkldnn': self.use_mkldnn, - 'with_bias': self.with_bias - } + self.attrs = {'use_mkldnn': self.use_mkldnn, } self.outputs = { 'Out': fully_connected_naive(self.matrix.input, self.matrix.weights) @@ -85,13 +82,11 @@ class TestFCMKLDNNOp3(TestFCMKLDNNOp): class TestFCMKLDNNOp4(TestFCMKLDNNOp): def init_op_type(self): - self.with_bias = False self.matrix = MatrixGenerate(2, 32, 48, 2, 2) class TestFCMKLDNNOp4(TestFCMKLDNNOp): def init_op_type(self): - self.with_bias = False self.matrix = MatrixGenerate(2, 32, 1000, 6, 6) diff --git a/python/paddle/fluid/tests/unittests/test_fc_op.py b/python/paddle/fluid/tests/unittests/test_fc_op.py new file mode 100644 index 0000000000000000000000000000000000000000..2bb920710a9b10f3a8159bad3b33dd15ffbada19 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fc_op.py @@ -0,0 +1,90 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +def fc_refer(matrix, with_bias): + in_n, in_c, in_h, in_w = matrix.input.shape + w_i, w_o = matrix.weights.shape + + x_data = np.reshape(matrix.input, [in_n, in_c * in_h * in_w]) + w_data = np.reshape(matrix.weights, [w_i, w_o]) + b_data = np.reshape(matrix.bias, [1, w_o]) + result = None + + if with_bias: + result = np.dot(x_data, w_data) + b_data + else: + result = np.dot(x_data, w_data) + + return result + + +class MatrixGenerate: + def __init__(self, mb, ic, oc, h, w): + self.input = np.random.random((mb, ic, h, w)).astype("float32") + self.weights = np.random.random((ic * h * w, oc)).astype("float32") + self.bias = np.random.random((1, oc)).astype("float32") + + +class TestFCOp(OpTest): + def setUp(self): + self.op_type = "fc" + self.matrix = MatrixGenerate(1, 10, 15, 3, 3) + + self.with_bias = True + if self.with_bias: + self.inputs = { + 'Input': self.matrix.input, + 'W': self.matrix.weights, + 'Bias': self.matrix.bias + } + else: + self.inputs = {'Input': self.matrix.input, 'W': self.matrix.weights} + + self.attrs = {'use_mkldnn': False} + + self.outputs = {'Out': fc_refer(self.matrix, self.with_bias)} + + def test_check_output(self): + self.check_output() + + +class TestFCOpBiasBoth(TestFCOp): + def init_shapes(self, mb, ic, oc, h, w): + for with_bias in {True, False}: + self.with_bias = with_bias + self.matrix = MatrixGenerate(mb, ic, oc, h, w) + + +class TestFCOp1(TestFCOpBiasBoth): + def init_op_type(self): + self.init_shapes(2, 8, 10, 1, 1) + + +class TestFCOp2(TestFCOpBiasBoth): + def init_op_type(self): + self.init_shapes(4, 5, 6, 2, 2) + + +class TestFCOp4(TestFCOpBiasBoth): + def init_op_type(self): + self.init_shapes(1, 32, 64, 3, 3) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py b/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py index 9d724a6479f061996359b1efcc5f61f0564331c7..8b9da843115409c65055927d317867d1290c8f0e 100644 --- a/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py +++ b/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py @@ -24,17 +24,16 @@ class TestFeedFetch(unittest.TestCase): input_array = np.ones((4, 4, 6)).astype("float32") input_array[0, 0, 0] = 3 input_array[3, 3, 5] = 10 - input_tensor = core.LoDTensor([[0, 2, 4]]) + input_tensor = core.LoDTensor([[2, 2]]) input_tensor.set(input_array, place) core.set_feed_variable(scope, input_tensor, "feed", 0) output_tensor = core.get_fetch_variable(scope, "feed", 0) - output_lod = output_tensor.lod() - self.assertEqual(0, output_lod[0][0]) + output_lod = output_tensor.recursive_sequence_lengths() + self.assertEqual(2, output_lod[0][0]) self.assertEqual(2, output_lod[0][1]) - self.assertEqual(4, output_lod[0][2]) output_array = np.array(output_tensor) self.assertEqual(3, output_array[0, 0, 0]) diff --git a/python/paddle/fluid/tests/unittests/test_fetch_var.py b/python/paddle/fluid/tests/unittests/test_fetch_var.py index 46c3bbb6712c6276e48dd9328d7741a447f28b91..e6f37f0b4ca781e4ec83a00f8f2605ef02716bd7 100644 --- a/python/paddle/fluid/tests/unittests/test_fetch_var.py +++ b/python/paddle/fluid/tests/unittests/test_fetch_var.py @@ -26,7 +26,7 @@ class TestFetchVar(op_test.OpTest): layers.assign(input=val, output=x) exe = fluid.Executor(fluid.CPUPlace()) exe.run(fluid.default_main_program(), feed={}, fetch_list=[]) - fetched_x = fluid.fetch_var("x") + fetched_x = fluid.executor._fetch_var("x") self.assertTrue( numpy.array_equal(fetched_x, val), "fetch_x=%s val=%s" % (fetched_x, val)) diff --git a/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py b/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py index 533d8ccfac82a2e298af16181ab16bf7aa3db282..0c75cf33f5f208d11081a6802910c25553b8c4ec 100644 --- a/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py +++ b/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py @@ -55,7 +55,7 @@ class TestFillConstantBatchSizeLikeWithLoDTensor(OpTest): self.op_type = "fill_constant_batch_size_like" self.inputs = { 'Input': (np.random.random((31, 28)).astype("float32"), - [[0, 9, 23, 31]]) + [[9, 14, 8]]) } self.attrs = { 'value': 3.5, diff --git a/python/paddle/fluid/tests/unittests/test_flatten_op.py b/python/paddle/fluid/tests/unittests/test_flatten_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f8692ce2ea66ef61c63bc41e77df050398ac63fd --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_flatten_op.py @@ -0,0 +1,68 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +from op_test import OpTest + + +class TestFlattenOp(OpTest): + def setUp(self): + self.op_type = "flatten" + self.init_test_case() + self.inputs = {"X": np.random.random(self.in_shape).astype("float32")} + self.init_attrs() + self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + def init_test_case(self): + self.in_shape = (3, 2, 2, 5) + self.axis = 1 + self.new_shape = (3, 20) + + def init_attrs(self): + self.attrs = {"axis": self.axis} + + +class TestFlattenOp(TestFlattenOp): + def init_test_case(self): + self.in_shape = (3, 2, 2, 3) + self.axis = 0 + self.new_shape = (1, 36) + + +class TestFlattenOpWithDefaultAxis(TestFlattenOp): + def init_test_case(self): + self.in_shape = (3, 2, 2, 3) + self.new_shape = (3, 12) + + def init_attrs(self): + self.attrs = {} + + +class TestFlattenOpSixDims(TestFlattenOp): + def init_test_case(self): + self.in_shape = (3, 2, 3, 2, 4, 4) + self.axis = 4 + self.new_shape = (36, 16) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py b/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py new file mode 100644 index 0000000000000000000000000000000000000000..ec0a939e9ec21952a6657ea849bb9844bb69cc8d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py @@ -0,0 +1,818 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import unittest +import numpy as np +import paddle.fluid.core as core +from op_test import OpTest + +# scale + add +# TestElementwiseAddOp +# TestFusedOperatorsOp_scalar +# TestFusedOperatorsOp_scalar2 +# TestFusedOperatorsOp_Vector +# TestFusedOperatorsOp_broadcast_0 +# TestFusedOperatorsOp_broadcast_1 +# TestFusedOperatorsOp_broadcast_2 +# TestFusedOperatorsOp_broadcast_3 +# TestFusedOperatorsOp_broadcast_4 +# TestFusedOperatorsOp_rowwise_add_0 +# TestFusedOperatorsOp_rowwise_add_1 +# TestFusedOperatorsOp_channelwise_add + + +class TestElementwiseAddOp(OpTest): + def setUp(self): + self.op_type = "fused_elemwise_activation" + self.dtype = np.float32 + self.axis = -1 + + self.init_axis() + self.init_dtype() + self.init_input() + self.init_output() + self.init_attr() + + self.inputs = { + 'X': OpTest.np_dtype_to_fluid_dtype(self.x), + 'Y': OpTest.np_dtype_to_fluid_dtype(self.y) + } + self.outputs = {'Out': self.out} + + def init_input(self): + self.x = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype) + self.y = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype) + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["scale", "elementwise_add"] + } + + def init_dtype(self): + pass + + def init_axis(self): + pass + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) + + +class TestFusedOperatorsOp_scalar(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4).astype(self.dtype) + self.y = np.random.rand(1).astype(self.dtype) + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y) * self.scale + + +class TestFusedOperatorsOp_scalar2(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4).astype(self.dtype) + self.y = np.random.rand(1, 1).astype(self.dtype) + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y) * self.scale + + +class TestFusedOperatorsOp_Vector(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.random((32, )).astype(self.dtype) + self.y = np.random.random((32, )).astype(self.dtype) + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y) * self.scale + + +class TestFusedOperatorsOp_broadcast_0(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4).astype(self.dtype) + self.y = np.random.rand(2).astype(self.dtype) + + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(2, 1, 1)) * self.scale + + +class TestFusedOperatorsOp_broadcast_1(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4).astype(self.dtype) + self.y = np.random.rand(3).astype(self.dtype) + + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(1, 3, 1)) * self.scale + + +class TestFusedOperatorsOp_broadcast_2(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4).astype(self.dtype) + self.y = np.random.rand(4).astype(self.dtype) + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(1, 1, 4)) * self.scale + + +class TestFusedOperatorsOp_broadcast_3(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(3, 4).astype(self.dtype) + + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(1, 3, 4, 1)) * self.scale + + +class TestFusedOperatorsOp_broadcast_4(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype) + self.y = np.random.rand(2, 1).astype(self.dtype) + + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(2, 1, 1, 1)) * self.scale + + +class TestFusedOperatorsOp_rowwise_add_0(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 3, 4).astype(self.dtype) + self.y = np.random.rand(3, 4).astype(self.dtype) + + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(1, 3, 4)) * self.scale + + +class TestFusedOperatorsOp_rowwise_add_1(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(2, 1).astype(self.dtype) + self.y = np.random.rand(1).astype(self.dtype) + + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y.reshape(1, 1)) * self.scale + + +class TestFusedOperatorsOp_channelwise_add(TestElementwiseAddOp): + def init_input(self): + self.x = np.random.rand(3, 20, 20).astype(self.dtype) + self.y = np.random.rand(3, 1, 1).astype(self.dtype) + + def init_axis(self): + self.axis = -1 + + def init_output(self): + self.scale = 0.1 + self.out = (self.x + self.y) * self.scale + + +# add + scale +# TestElementwiseAddOp_f_add_scale +# TestFusedOperatorsOp_scalar_f_add_scale +# TestFusedOperatorsOp_scalar2_f_add_scale +# TestFusedOperatorsOp_Vector_f_add_scale +# TestFusedOperatorsOp_broadcast_0_f_add_scale +# TestFusedOperatorsOp_broadcast_1_f_add_scale +# TestFusedOperatorsOp_broadcast_2_f_add_scale +# TestFusedOperatorsOp_broadcast_3_f_add_scale +# TestFusedOperatorsOp_broadcast_4_f_add_scale +# TestFusedOperatorsOp_rowwise_add_0_f_add_scale +# TestFusedOperatorsOp_rowwise_add_1_f_add_scale +# TestFusedOperatorsOp_channelwise_add_f_add_scale + + +class TestFusedOperatorsOp_f_add_scale(TestElementwiseAddOp): + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_scalar_f_add_scale(TestFusedOperatorsOp_scalar): + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_scalar2_f_add_scale(TestFusedOperatorsOp_scalar2): + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_Vector_f_add_scale(TestFusedOperatorsOp_Vector): + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_broadcast_0_f_add_scale( + TestFusedOperatorsOp_broadcast_0): + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y.reshape(2, 1, 1) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_broadcast_1_f_add_scale( + TestFusedOperatorsOp_broadcast_1): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y.reshape(1, 3, 1) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_broadcast_2_f_add_scale( + TestFusedOperatorsOp_broadcast_2): + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y.reshape(1, 1, 4) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_broadcast_3_f_add_scale( + TestFusedOperatorsOp_broadcast_3): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y.reshape(1, 3, 4, 1) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_broadcast_4_f_add_scale( + TestFusedOperatorsOp_broadcast_4): + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.scale = 0.2 + self.out = self.x + self.y.reshape(2, 1, 1, 1) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_rowwise_add_0_f_add_scale( + TestFusedOperatorsOp_rowwise_add_0): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.1 + self.out = self.x + self.y.reshape(1, 3, 4) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_rowwise_add_1_f_add_scale( + TestFusedOperatorsOp_rowwise_add_1): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.scale = 0.2 + self.out = self.x + self.y.reshape(1, 1) * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +class TestFusedOperatorsOp_channelwise_add_f_add_scale( + TestFusedOperatorsOp_channelwise_add): + def init_axis(self): + self.axis = -1 + + def init_output(self): + self.scale = 0.2 + self.out = self.x + self.y * self.scale + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'scale': self.scale, + 'functor_list': ["elementwise_add", "scale"] + } + + +# add + relu +# TestElementwiseAddOp_f_add_relu +# TestFusedOperatorsOp_scalar_f_add_relu +# TestFusedOperatorsOp_scalar2_f_add_relu +# TestFusedOperatorsOp_Vector_f_add_relu +# TestFusedOperatorsOp_broadcast_0_f_add_relu +# TestFusedOperatorsOp_broadcast_1_f_add_relu +# TestFusedOperatorsOp_broadcast_2_f_add_relu +# TestFusedOperatorsOp_broadcast_3_f_add_relu +# TestFusedOperatorsOp_broadcast_4_f_add_relu +# TestFusedOperatorsOp_rowwise_add_0_f_add_relu +# TestFusedOperatorsOp_rowwise_add_1_f_add_relu +# TestFusedOperatorsOp_channelwise_add_f_add_relu + + +class TestFusedOperatorsOp_f_add_relu(TestElementwiseAddOp): + def init_output(self): + # Copy from test_activation_op.py + # Because we set delta = 0.005 in calculating numeric gradient, + # if x is too small, such as 0.002, x_neg will be -0.003 + # x_pos will be 0.007, so the numeric gradient is inaccurate. + # we should avoid this + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y, 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_scalar_f_add_relu(TestFusedOperatorsOp_scalar): + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y, 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_scalar2_f_add_relu(TestFusedOperatorsOp_scalar2): + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y, 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_Vector_f_add_relu(TestFusedOperatorsOp_Vector): + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y, 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_broadcast_0_f_add_relu( + TestFusedOperatorsOp_broadcast_0): + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(2, 1, 1), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_broadcast_1_f_add_relu( + TestFusedOperatorsOp_broadcast_1): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(1, 3, 1), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_broadcast_2_f_add_relu( + TestFusedOperatorsOp_broadcast_2): + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(1, 1, 4), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_broadcast_3_f_add_relu( + TestFusedOperatorsOp_broadcast_3): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(1, 3, 4, 1), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_broadcast_4_f_add_relu( + TestFusedOperatorsOp_broadcast_4): + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(2, 1, 1, 1), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_rowwise_add_0_f_add_relu( + TestFusedOperatorsOp_rowwise_add_0): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(1, 3, 4), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_rowwise_add_1_f_add_relu( + TestFusedOperatorsOp_rowwise_add_1): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y.reshape(1, 1), 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +class TestFusedOperatorsOp_channelwise_add_f_add_relu( + TestFusedOperatorsOp_channelwise_add): + def init_axis(self): + self.axis = -1 + + def init_output(self): + self.y[np.abs(self.y) < 0.005] = 0.02 + self.out = self.x + np.maximum(self.y, 0) + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["elementwise_add", "relu"] + } + + +# relu + add +# TestElementwiseAddOp_f_relu_add +# TestFusedOperatorsOp_scalar_f_relu_add +# TestFusedOperatorsOp_scalar2_f_relu_add +# TestFusedOperatorsOp_Vector_f_relu_add +# TestFusedOperatorsOp_broadcast_0_f_relu_add +# TestFusedOperatorsOp_broadcast_1_f_relu_add +# TestFusedOperatorsOp_broadcast_2_f_relu_add +# TestFusedOperatorsOp_broadcast_3_f_relu_add +# TestFusedOperatorsOp_broadcast_4_f_relu_add +# TestFusedOperatorsOp_rowwise_add_0_f_relu_add +# TestFusedOperatorsOp_rowwise_add_1_f_relu_add +# TestFusedOperatorsOp_channelwise_add_f_relu_add + + +class TestFusedOperatorsOp_f_relu_add(TestElementwiseAddOp): + def init_output(self): + # Copy from test_activation_op.py + # Because we set delta = 0.005 in calculating numeric gradient, + # if x is too small, such as 0.002, x_neg will be -0.003 + # x_pos will be 0.007, so the numeric gradient is inaccurate. + # we should avoid this + self.out = self.x + self.y + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_scalar_f_relu_add(TestFusedOperatorsOp_scalar): + def init_output(self): + self.out = self.x + self.y + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_scalar2_f_relu_add(TestFusedOperatorsOp_scalar2): + def init_output(self): + self.out = self.x + self.y + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_Vector_f_relu_add(TestFusedOperatorsOp_Vector): + def init_output(self): + self.out = self.x + self.y + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_broadcast_0_f_relu_add( + TestFusedOperatorsOp_broadcast_0): + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.out = self.x + self.y.reshape(2, 1, 1) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_broadcast_1_f_relu_add( + TestFusedOperatorsOp_broadcast_1): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.out = self.x + self.y.reshape(1, 3, 1) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_broadcast_2_f_relu_add( + TestFusedOperatorsOp_broadcast_2): + def init_output(self): + self.out = self.x + self.y.reshape(1, 1, 4) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_broadcast_3_f_relu_add( + TestFusedOperatorsOp_broadcast_3): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.out = self.x + self.y.reshape(1, 3, 4, 1) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_broadcast_4_f_relu_add( + TestFusedOperatorsOp_broadcast_4): + def init_axis(self): + self.axis = 0 + + def init_output(self): + self.out = self.x + self.y.reshape(2, 1, 1, 1) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_rowwise_add_0_f_relu_add( + TestFusedOperatorsOp_rowwise_add_0): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.out = self.x + self.y.reshape(1, 3, 4) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_rowwise_add_1_f_relu_add( + TestFusedOperatorsOp_rowwise_add_1): + def init_axis(self): + self.axis = 1 + + def init_output(self): + self.out = self.x + self.y.reshape(1, 1) + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +class TestFusedOperatorsOp_channelwise_add_f_relu_add( + TestFusedOperatorsOp_channelwise_add): + def init_axis(self): + self.axis = -1 + + def init_output(self): + self.out = self.x + self.y + self.out = np.maximum(self.out, 0) + self.out[np.abs(self.out) < 0.005] = 0.02 + + def init_attr(self): + self.attrs = { + 'axis': self.axis, + 'functor_list': ["relu", "elementwise_add"] + } + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gather_op.py b/python/paddle/fluid/tests/unittests/test_gather_op.py index 6fd043c27e27db53c95be3630b6c08216e8e35f4..4ae90864806204197c52bbbdc5516f141afd4613 100644 --- a/python/paddle/fluid/tests/unittests/test_gather_op.py +++ b/python/paddle/fluid/tests/unittests/test_gather_op.py @@ -20,8 +20,9 @@ from op_test import OpTest class TestGatherOp(OpTest): def setUp(self): self.op_type = "gather" - xnp = np.random.random((10, 20)).astype("float32") - self.inputs = {'X': xnp, 'Index': np.array([1, 3, 5]).astype("int32")} + self.config() + xnp = np.random.random(self.x_shape).astype("float32") + self.inputs = {'X': xnp, 'Index': np.array(self.index).astype("int32")} self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]} def test_check_output(self): @@ -30,6 +31,16 @@ class TestGatherOp(OpTest): def test_check_grad(self): self.check_grad(['X'], 'Out') + def config(self): + self.x_shape = (10, 20) + self.index = [1, 3, 5] + + +class TestCase1(TestGatherOp): + def config(self): + self.x_shape = (10) + self.index = [1, 3, 5] + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py new file mode 100644 index 0000000000000000000000000000000000000000..3ae877a60818744f852d3af9a02ffebf5e2affc8 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py @@ -0,0 +1,26 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +from test_gaussian_random_op import TestGaussianRandomOp + + +class TestMKLDNN(TestGaussianRandomOp): + def init_kernel_type(self): + self.use_mkldnn = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py index 272caceaf38699438ccae41691bf26b2eb4d2a22..8481500fd78f0ccf34f09c66bec27e195b9aada3 100644 --- a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py @@ -25,7 +25,15 @@ class TestGaussianRandomOp(unittest.TestCase): def setUp(self): self.op_type = "gaussian_random" self.inputs = {} - self.attrs = {"shape": [1000, 784], "mean": .0, "std": 1., "seed": 10} + self.use_mkldnn = False + self.init_kernel_type() + self.attrs = { + "shape": [1000, 784], + "mean": .0, + "std": 1., + "seed": 10, + "use_mkldnn": self.use_mkldnn + } self.outputs = ["Out"] @@ -58,6 +66,9 @@ class TestGaussianRandomOp(unittest.TestCase): self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1) self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1) + def init_kernel_type(self): + pass + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_get_places_op.py b/python/paddle/fluid/tests/unittests/test_get_places_op.py index 6dab1e22f0c50ab011d6b8e8944097600cf3fecc..964423e2d2638224244b4ca774d8eee08f3ec989 100644 --- a/python/paddle/fluid/tests/unittests/test_get_places_op.py +++ b/python/paddle/fluid/tests/unittests/test_get_places_op.py @@ -13,6 +13,7 @@ # limitations under the License. import paddle.fluid as fluid +from paddle.fluid.layers.device import get_places import decorators import unittest @@ -20,7 +21,7 @@ import unittest class TestGetPlaces(unittest.TestCase): @decorators.prog_scope() def test_get_places(self): - places = fluid.layers.get_places() + places = get_places() cpu = fluid.CPUPlace() exe = fluid.Executor(cpu) exe.run(fluid.default_main_program()) diff --git a/python/paddle/fluid/tests/unittests/test_gru_op.py b/python/paddle/fluid/tests/unittests/test_gru_op.py index 3a13eb872a8646cede126b667864dfc3784ebd0b..86a2c674d01f45b2b141572c8191d2fba7fa312f 100644 --- a/python/paddle/fluid/tests/unittests/test_gru_op.py +++ b/python/paddle/fluid/tests/unittests/test_gru_op.py @@ -20,8 +20,8 @@ from test_lstm_op import identity, sigmoid, tanh, relu class TestGRUOp(OpTest): - lod = [[0, 2, 6, 9]] - batch_size = lod[0][-1] + lod = [[2, 4, 3]] + batch_size = sum(lod[0]) frame_size = 5 activate = { 'identity': identity, @@ -33,12 +33,12 @@ class TestGRUOp(OpTest): @staticmethod def seq_to_batch(lod, is_reverse): idx_in_seq_list = [] - seq_starts = lod[0] - seq_lens = [] - for i in range(len(seq_starts) - 1): - seq_lens.append(seq_starts[i + 1] - seq_starts[i]) + seq_lens = lod[0] + seq_starts = [0] + for i in range(len(seq_lens)): + seq_starts.append(seq_starts[-1] + seq_lens[i]) sorted_seqs = sorted( - range(len(seq_lens)), lambda x, y: seq_lens[y] - seq_lens[x]) + list(range(len(seq_lens))), lambda x, y: seq_lens[y] - seq_lens[x]) num_batch = seq_lens[sorted_seqs[0]] for batch_idx in range(num_batch): idx_in_seq = [] @@ -74,15 +74,16 @@ class TestGRUOp(OpTest): def gru(self): input, lod = self.inputs['Input'] w = self.inputs['Weight'] - b = self.inputs['Bias'] if self.inputs.has_key('Bias') else np.zeros( + b = self.inputs['Bias'] if 'Bias' in self.inputs else np.zeros( (1, self.frame_size * 3)) batch_gate = self.outputs['BatchGate'] batch_reset_hidden_prev = self.outputs['BatchResetHiddenPrev'] batch_hidden = self.outputs['BatchHidden'] hidden = self.outputs['Hidden'] idx_in_seq_list = self.idx_in_seq_list - h_p = self.inputs['H0'][self.sorted_seqs] if self.inputs.has_key( - 'H0') else np.zeros((len(idx_in_seq_list[0]), self.frame_size)) + h_p = self.inputs['H0'][ + self.sorted_seqs] if 'H0' in self.inputs else np.zeros( + (len(idx_in_seq_list[0]), self.frame_size)) num_batch = len(idx_in_seq_list) end_idx = 0 for batch_idx in range(num_batch): diff --git a/python/paddle/fluid/tests/unittests/test_gru_unit_op.py b/python/paddle/fluid/tests/unittests/test_gru_unit_op.py index c56b1eefd3a3dfe1478bd0526fa32077edcac9ba..87a9eba4d97459082cdf1499efeddf24ed51e1b1 100644 --- a/python/paddle/fluid/tests/unittests/test_gru_unit_op.py +++ b/python/paddle/fluid/tests/unittests/test_gru_unit_op.py @@ -76,7 +76,7 @@ class TestGRUUnitOp(OpTest): x = self.inputs['Input'] h_p = self.inputs['HiddenPrev'] w = self.inputs['Weight'] - b = self.inputs['Bias'] if self.inputs.has_key('Bias') else np.zeros( + b = self.inputs['Bias'] if 'Bias' in self.inputs else np.zeros( (1, frame_size * 3)) g = x + np.tile(b, (batch_size, 1)) w_u_r = w.flatten()[:frame_size * frame_size * 2].reshape( diff --git a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py new file mode 100644 index 0000000000000000000000000000000000000000..daa5da8d95129af0305b326832a557daeb4c5c9c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py @@ -0,0 +1,101 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import math +from op_test import OpTest + +np.random.seed(100) + + +def find_latest_set(num): + return 1 + int(math.floor(math.log(num, 2))) + + +class CodeTable(object): + def __init__(self, num_classes, code): + self.c = num_classes + code + + def cal_index(self, bit): + return (self.c >> (bit + 1)) - 1 + + def get_length(self): + return find_latest_set(self.c) - 1 + + def cal_bit(self, bit): + return self.c & (1 << bit) + + +def hsigmoid(x, w, label, bias, num_classes): + batch_size = x.shape[0] + code_length = find_latest_set(num_classes - 1) + code_table = [0 for _ in range(code_length)] + pre_output = np.zeros((batch_size, code_length)) + pre_sum = np.zeros((batch_size, 1)) + out = np.zeros((batch_size, 1)).astype("float32") + for i in range(batch_size): + code_table = CodeTable(num_classes, label[i]) + length = code_table.get_length() + for j in range(length): + idx = code_table.cal_index(j) + pre_output[i][j] += bias[0][idx] + for i in range(batch_size): + code_table = CodeTable(num_classes, label[i]) + length = code_table.get_length() + for j in range(length): + idx = code_table.cal_index(j) + pre_output[i][j] += np.dot(w[idx], x[i]) + # clip[-40.0, 40.0] + pre_output = np.clip(pre_output, -40.0, 40.0) + # out(i, 0) = \sum_j bit(i, j) * preout(i, j) + for i in range(batch_size): + code_table = CodeTable(num_classes, label[i]) + length = code_table.get_length() + sum = 0.0 + for j in range(length): + if code_table.cal_bit(j): + sum += pre_output[i][j] + out[i] = -1.0 * sum + # soft relu + pre_output = np.log(1 + np.exp(pre_output)) + pre_sum = pre_output.sum(1).reshape((batch_size, 1)) + out += pre_sum + return pre_output, out + + +class TestHSigmoidOp(OpTest): + def setUp(self): + self.op_type = "hierarchical_sigmoid" + num_classes = 6 + feature_size = 8 + batch_size = 4 + x = np.random.random((batch_size, feature_size)).astype("float32") + w = np.random.random((num_classes - 1, feature_size)).astype("float32") + label = np.random.randint(0, num_classes, (batch_size, 1)) + bias = np.random.random((1, num_classes - 1)).astype("float32") + self.attrs = {'num_classes': num_classes} + self.inputs = {'X': x, 'W': w, 'Label': label, 'Bias': bias} + pre_output, out = hsigmoid(x, w, label, bias, num_classes) + self.outputs = {'PreOut': pre_output, 'Out': out} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['Bias', 'X', 'W'], ['Out'], no_grad_set=set('Label')) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_im2sequence_op.py b/python/paddle/fluid/tests/unittests/test_im2sequence_op.py index 4946475f11a4fc0ccaffeec6821d3976ea7c6560..13bc5768740ece00bbe285a0b47d82bb8a42d2c7 100644 --- a/python/paddle/fluid/tests/unittests/test_im2sequence_op.py +++ b/python/paddle/fluid/tests/unittests/test_im2sequence_op.py @@ -16,23 +16,48 @@ import numpy as np from op_test import OpTest -def get_output_shape(attrs, in_shape): +def get_output_shape(attrs, in_shape, img_real_size): + batchsize = in_shape[0] img_height = in_shape[2] img_width = in_shape[3] + paddings = np.array(attrs['paddings']).astype("int32") + kernels = np.array(attrs['kernels']).astype("int32") + strides = np.array(attrs['strides']).astype("int32") + output_height = np.zeros((1, batchsize)).astype("int32") + output_width = np.zeros((1, batchsize)).astype("int32") + if len(img_real_size): + out_stride = np.array(attrs['out_stride']).astype("int32") + imgreal_h = 0 + imgreal_w = 0 + for index in range(batchsize): + if img_real_size[index, 0] % out_stride[0] == 0: + imgreal_h = img_real_size[index, 0] / out_stride[0] + else: + imgreal_h = img_real_size[index, 0] / out_stride[0] + 1 + if img_real_size[index, 0] % out_stride[1] == 0: + imgreal_w = img_real_size[index, 1] / out_stride[1] + else: + imgreal_w = img_real_size[index, 0] / out_stride[1] + 1 + output_height[0,index] = \ + 1 + \ + (imgreal_h + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \ + strides[0] - paddings = attrs['paddings'] - kernels = attrs['kernels'] - strides = attrs['strides'] + output_width[0,index] = \ + 1 + \ + (imgreal_w + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \ + strides[1] + else: + for index in range(batchsize): + output_height[0,index] = \ + 1 + \ + (img_height + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \ + strides[0] - output_height = \ - 1 + \ - (img_height + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \ - strides[0] - - output_width = \ - 1 + \ - (img_width + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \ - strides[1] + output_width[0,index] = \ + 1 + \ + (img_width + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \ + strides[1] return output_height, output_width @@ -75,22 +100,25 @@ def im2col(attrs, im, col): im_row_offset][im_col_offset] -def Im2Sequence(inputs, attrs): - output_height, output_width = get_output_shape(attrs, inputs.shape) +def Im2Sequence(inputs, img_real_size, attrs): + output_height, output_width = get_output_shape(attrs, inputs.shape, + img_real_size) img_channels = inputs.shape[1] batch_size = inputs.shape[0] - out = np.zeros([ - batch_size, output_height, output_width, img_channels, - attrs['kernels'][0], attrs['kernels'][1] - ]).astype("float32") - - for i in range(len(inputs)): - im2col(attrs, inputs[i], out[i]) - - out = out.reshape([ - batch_size * output_height * output_width, - img_channels * attrs['kernels'][0] * attrs['kernels'][1] - ]) + out = [] + for index in range(batch_size): + tmp = np.zeros([ + output_height[0, index], output_width[0, index], img_channels, + attrs['kernels'][0], attrs['kernels'][1] + ]).astype("float32") + out.append(tmp) + for index in range(len(inputs)): + im2col(attrs, inputs[index], out[index]) + out[index] = out[index].reshape([ + output_height[0, index] * output_width[0, index], + img_channels * attrs['kernels'][0] * attrs['kernels'][1] + ]) + out = np.concatenate(out, axis=0) return out @@ -103,7 +131,7 @@ class TestBlockExpandOp(OpTest): self.attrs = { 'kernels': [2, 2], 'strides': [1, 1], - 'paddings': [1, 1, 1, 1] + 'paddings': [1, 1, 1, 1], } def setUp(self): @@ -113,7 +141,8 @@ class TestBlockExpandOp(OpTest): self.batch_size, self.img_channels, self.img_height, self.img_width ]).astype("float32") - out = Im2Sequence(x, self.attrs) + real_size = np.array([]).astype("float32") + out = Im2Sequence(x, real_size, self.attrs) self.inputs = {'X': x} self.outputs = {'Out': out} @@ -133,20 +162,20 @@ class TestBlockExpandOpCase2(TestBlockExpandOp): self.attrs = { 'kernels': [2, 1], 'strides': [2, 1], - 'paddings': [2, 1, 2, 1] + 'paddings': [2, 1, 2, 1], } class TestBlockExpandOpCase3(TestBlockExpandOp): def config(self): - self.batch_size = 3 + self.batch_size = 2 self.img_channels = 1 self.img_height = 4 self.img_width = 5 self.attrs = { 'kernels': [2, 1], 'strides': [2, 1], - 'paddings': [2, 0, 2, 0] + 'paddings': [2, 0, 2, 0], } @@ -159,9 +188,94 @@ class TestBlockExpandOpCase4(TestBlockExpandOp): self.attrs = { 'kernels': [2, 2], 'strides': [1, 1], - 'paddings': [0, 0, 0, 0] + 'paddings': [0, 0, 0, 0], + } + + +class TestBlockExpandOpCase5(OpTest): + def config(self): + self.batch_size = 1 + self.img_channels = 3 + self.img_height = 4 + self.img_width = 5 + self.attrs = { + 'kernels': [2, 1], + 'strides': [2, 1], + 'paddings': [2, 1, 2, 1], + 'out_stride': [2, 2], + } + + def setUp(self): + self.config() + self.op_type = "im2sequence" + x = np.random.uniform(0.1, 1, [ + self.batch_size, self.img_channels, self.img_height, self.img_width + ]).astype("float32") + real_size = np.array([[8, 10], [5, 8]]).astype("float32") + out = np.array(Im2Sequence(x, real_size, self.attrs)) + self.inputs = {'X': x, 'Y': real_size} #l ?? + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +class TestBlockExpandOpCase6(OpTest): + def config(self): + self.batch_size = 3 + self.img_channels = 1 + self.img_height = 4 + self.img_width = 5 + self.attrs = { + 'kernels': [2, 1], + 'strides': [1, 1], + 'paddings': [0, 0, 0, 0], + 'out_stride': [1, 1], + } + + def setUp(self): + self.config() + self.op_type = "im2sequence" + x = np.random.uniform(0.1, 1, [ + self.batch_size, self.img_channels, self.img_height, self.img_width + ]).astype("float32") + real_size = np.array([[8, 10], [5, 8], [5, 8]]).astype("float32") + out = np.array(Im2Sequence(x, real_size, self.attrs)) + self.inputs = {'X': x, 'Y': real_size} #l ?? + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +class TestBlockExpandOpCase7(OpTest): + def config(self): + self.batch_size = 2 + self.img_channels = 2 + self.img_height = 3 + self.img_width = 3 + self.attrs = { + 'kernels': [2, 2], + 'strides': [1, 1], + 'paddings': [1, 0, 1, 0], + 'out_stride': [2, 2], } + def setUp(self): + self.config() + self.op_type = "im2sequence" + x = np.random.uniform(0.1, 1, [ + self.batch_size, self.img_channels, self.img_height, self.img_width + ]).astype("float32") + real_size = np.array([[6, 6], [4, 4]]).astype("float32") + out = np.array(Im2Sequence(x, real_size, self.attrs)) + self.inputs = {'X': x, 'Y': real_size} + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + if __name__ == '__main__': unittest.main() +#set shiftwidth=4 set expandtab set tabstop=4 diff --git a/python/paddle/fluid/tests/unittests/test_image_classification_layer.py b/python/paddle/fluid/tests/unittests/test_image_classification_layer.py index 6ecfa9ea213fe0cf57e18fa83bbb85c223727d71..23b1ed957ad15bb631cd5160eb48328c76302987 100644 --- a/python/paddle/fluid/tests/unittests/test_image_classification_layer.py +++ b/python/paddle/fluid/tests/unittests/test_image_classification_layer.py @@ -43,7 +43,7 @@ class TestLayer(unittest.TestCase): hidden2 = fluid.layers.fc(input=hidden1, size=128, act='relu') fluid.layers.batch_norm(input=hidden2) - print str(main_program) + print(str(main_program)) def test_dropout_layer(self): main_program = Program() @@ -53,7 +53,7 @@ class TestLayer(unittest.TestCase): name='pixel', shape=[3, 48, 48], dtype='float32') fluid.layers.dropout(x=images, dropout_prob=0.5) - print str(main_program) + print(str(main_program)) def test_img_conv_group(self): main_program = Program() @@ -65,7 +65,7 @@ class TestLayer(unittest.TestCase): conv1 = conv_block(images, 64, 2, [0.3, 0]) conv_block(conv1, 256, 3, [0.4, 0.4, 0]) - print str(main_program) + print(str(main_program)) def test_elementwise_add_with_act(self): main_program = Program() diff --git a/python/paddle/fluid/tests/unittests/test_inference_model_io.py b/python/paddle/fluid/tests/unittests/test_inference_model_io.py index 51460cbb1370f6794e13d18fe099865b4713691f..4cd203155f446df07d2fe6c1d56e0d20f1113679 100644 --- a/python/paddle/fluid/tests/unittests/test_inference_model_io.py +++ b/python/paddle/fluid/tests/unittests/test_inference_model_io.py @@ -48,7 +48,7 @@ class TestBook(unittest.TestCase): exe.run(init_program, feed={}, fetch_list=[]) - for i in xrange(100): + for i in range(100): tensor_x = np.array( [[1, 1], [1, 2], [3, 4], [5, 2]]).astype("float32") tensor_y = np.array([[-2], [-3], [-7], [-7]]).astype("float32") diff --git a/python/paddle/fluid/tests/unittests/test_initializer.py b/python/paddle/fluid/tests/unittests/test_initializer.py index 587e2025e1045f63a5825f884d4dcad8b4685e62..b215e379864e919af03591ab2566c08dddbb5743 100644 --- a/python/paddle/fluid/tests/unittests/test_initializer.py +++ b/python/paddle/fluid/tests/unittests/test_initializer.py @@ -27,12 +27,13 @@ class TestConstantInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.ConstantInitializer()) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.ConstantInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'fill_constant') @@ -43,12 +44,13 @@ class TestConstantInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.ConstantInitializer(2.3)) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.ConstantInitializer(2.3)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'fill_constant') @@ -61,12 +63,13 @@ class TestUniformInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.UniformInitializer()) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.UniformInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -80,18 +83,19 @@ class TestUniformInitializer(unittest.TestCase): program = framework.Program() program.random_seed = 123 block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.UniformInitializer()) - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.UniformInitializer(seed=456)) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param1", + initializer=initializer.UniformInitializer()) + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param2", + initializer=initializer.UniformInitializer(seed=456)) init_op = block.ops[1] self.assertEqual(init_op.attr("seed"), 123) init_op1 = block.ops[0] @@ -102,12 +106,13 @@ class TestUniformInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.UniformInitializer(-4.2, 3.1, 123)) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.UniformInitializer(-4.2, 3.1, 123)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -115,6 +120,25 @@ class TestUniformInitializer(unittest.TestCase): self.assertAlmostEqual(init_op.attr('max'), 3.1, delta=DELTA) self.assertEqual(init_op.attr('seed'), 123) + def test_uniform_initializer_two_op(self): + """Test uniform initializer with supplied attributes + """ + program = framework.Program() + block = program.global_block() + for i in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.UniformInitializer(-4.2, float(i), 123)) + self.assertEqual(len(block.ops), 1) + init_op0 = block.ops[0] + self.assertEqual(init_op0.type, 'uniform_random') + self.assertAlmostEqual(init_op0.attr('min'), -4.2, delta=DELTA) + self.assertAlmostEqual(init_op0.attr('max'), 0.0, delta=DELTA) + self.assertEqual(init_op0.attr('seed'), 123) + class TestNormalInitializer(unittest.TestCase): def test_normal_initializer_default_value(self): @@ -122,12 +146,13 @@ class TestNormalInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.NormalInitializer()) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.NormalInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'gaussian_random') @@ -140,12 +165,13 @@ class TestNormalInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.NormalInitializer(2.3, 1.9, 123)) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.NormalInitializer(2.3, 1.9, 123)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'gaussian_random') @@ -161,12 +187,13 @@ class TestXavierInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.XavierInitializer()) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.XavierInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -181,12 +208,13 @@ class TestXavierInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10, 15, 20], - lod_level=0, - name="param", - initializer=initializer.XavierInitializer()) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10, 15, 20], + lod_level=0, + name="param", + initializer=initializer.XavierInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -203,12 +231,13 @@ class TestXavierInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.XavierInitializer(uniform=False)) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.XavierInitializer(uniform=False)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'gaussian_random') @@ -223,12 +252,13 @@ class TestXavierInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10, 15, 20], - lod_level=0, - name="param", - initializer=initializer.XavierInitializer(uniform=False)) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10, 15, 20], + lod_level=0, + name="param", + initializer=initializer.XavierInitializer(uniform=False)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'gaussian_random') @@ -244,13 +274,14 @@ class TestXavierInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.XavierInitializer( - fan_in=12, fan_out=23, seed=134)) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.XavierInitializer( + fan_in=12, fan_out=23, seed=134)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -267,12 +298,13 @@ class TestMSRAInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.MSRAInitializer()) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -287,12 +319,13 @@ class TestMSRAInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10, 15, 20], - lod_level=0, - name="param", - initializer=initializer.MSRAInitializer()) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10, 15, 20], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer()) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -308,12 +341,13 @@ class TestMSRAInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.MSRAInitializer(uniform=False)) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer(uniform=False)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'gaussian_random') @@ -328,12 +362,13 @@ class TestMSRAInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - param = block.create_parameter( - dtype="float32", - shape=[5, 10, 15, 20], - lod_level=0, - name="param", - initializer=initializer.MSRAInitializer(uniform=False)) + for _ in range(2): + param = block.create_parameter( + dtype="float32", + shape=[5, 10, 15, 20], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer(uniform=False)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'gaussian_random') @@ -348,13 +383,14 @@ class TestMSRAInitializer(unittest.TestCase): """ program = framework.Program() block = program.global_block() - block.create_parameter( - dtype="float32", - shape=[5, 10], - lod_level=0, - name="param", - initializer=initializer.MSRAInitializer( - fan_in=12, seed=134)) + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer( + fan_in=12, seed=134)) self.assertEqual(len(block.ops), 1) init_op = block.ops[0] self.assertEqual(init_op.type, 'uniform_random') @@ -364,5 +400,23 @@ class TestMSRAInitializer(unittest.TestCase): self.assertEqual(init_op.attr('seed'), 134) +class TestMSRAInitializer(unittest.TestCase): + def test_bilinear_initializer(self): + """Test the bilinear initializer with supplied arguments + """ + program = framework.Program() + block = program.global_block() + for _ in range(2): + block.create_parameter( + dtype="float32", + shape=[8, 1, 3, 3], + lod_level=0, + name="param", + initializer=initializer.BilinearInitializer()) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'assign_value') + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py b/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py index 8f62ac20a5c13257a1519128292e2abc4962bf84..eff4212d91e609a7ef531280bbd3cf3671a59830 100644 --- a/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py +++ b/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py @@ -58,8 +58,8 @@ class TestIOUSimilarityOpWithLoD(TestIOUSimilarityOp): def setUp(self): super(TestIOUSimilarityOpWithLoD, self).setUp() - self.boxes1_lod = [[0, 1, 2]] - self.output_lod = [[0, 1, 2]] + self.boxes1_lod = [[1, 1]] + self.output_lod = [[1, 1]] self.inputs = {'X': (self.boxes1, self.boxes1_lod), 'Y': self.boxes2} self.outputs = {'Out': (self.output, self.output_lod)} diff --git a/python/paddle/fluid/tests/unittests/test_layer_norm_op.py b/python/paddle/fluid/tests/unittests/test_layer_norm_op.py index 69365db4d104a1b69916a605534eff83e242289f..295887ccd171a3101329eb1255da146914fa9264 100644 --- a/python/paddle/fluid/tests/unittests/test_layer_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_layer_norm_op.py @@ -17,6 +17,7 @@ import numpy as np from operator import mul import paddle.fluid.core as core import paddle.fluid as fluid +from functools import reduce np.random.random(123) diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 60dc1f83fc32e2551eb2a04ef35f1c8a0ffec769..38a138a8faee7746d9e7630d39206066634956f8 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -16,6 +16,7 @@ from __future__ import print_function import unittest import paddle.fluid.layers as layers +from paddle.fluid.layers.device import get_places import paddle.fluid.nets as nets from paddle.fluid.framework import Program, program_guard, default_main_program from paddle.fluid.param_attr import ParamAttr @@ -173,6 +174,16 @@ class TestBook(unittest.TestCase): x=dat, label=lbl)) print(str(program)) + def test_hsigmoid(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[2], dtype='float32') + y = layers.data(name='y', shape=[2], dtype='int64') + self.assertIsNotNone( + layers.hsigmoid( + input=x, label=y, num_classes=2)) + print(str(program)) + def test_sequence_expand(self): program = Program() with program_guard(program): @@ -238,7 +249,7 @@ class TestBook(unittest.TestCase): def test_get_places(self): program = Program() with program_guard(program): - x = layers.get_places(device_count=4) + x = get_places(device_count=4) self.assertIsNotNone(x) print(str(program)) @@ -251,12 +262,16 @@ class TestBook(unittest.TestCase): print(str(program)) def test_im2sequence(self): - print("test_im2sequence") program = Program() with program_guard(program): x = layers.data(name='x', shape=[3, 128, 128], dtype='float32') + y = layers.data(name='y', shape=[], dtype='float32') output = layers.im2sequence( - input=x, stride=[1, 1], filter_size=[2, 2]) + input=x, + input_image_size=y, + stride=[1, 1], + filter_size=[2, 2], + out_stride=[1, 1]) self.assertIsNotNone(output) print(str(program)) @@ -264,7 +279,7 @@ class TestBook(unittest.TestCase): def test_nce(self): window_size = 5 words = [] - for i in xrange(window_size): + for i in range(window_size): words.append( layers.data( name='word_{0}'.format(i), shape=[1], dtype='int64')) @@ -273,7 +288,7 @@ class TestBook(unittest.TestCase): label_word = int(window_size / 2) + 1 embs = [] - for i in xrange(window_size): + for i in range(window_size): if i == label_word: continue @@ -369,16 +384,107 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) - def test_upsampling_bilinear2d(self): + def test_resize_bilinear(self): program = Program() with program_guard(program): x = layers.data(name='x', shape=[3, 9, 6], dtype="float32") - output = layers.upsampling_bilinear2d(x, out_shape=[12, 12]) + output = layers.resize_bilinear(x, out_shape=[12, 12]) + self.assertIsNotNone(output) + output = layers.resize_bilinear(x, scale=3) + self.assertIsNotNone(output) + print(str(program)) + + def test_polygon_box_transform(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[8, 4, 4], dtype="float32") + output = layers.polygon_box_transform(input=x) + self.assertIsNotNone(output) + print(str(program)) + + def test_l2_normalize(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[8, 7, 10], dtype="float32") + output = layers.l2_normalize(x, axis=1) + + def test_maxout(self): + program = Program() + with program_guard(program): + data = layers.data(name='x', shape=[8, 6, 6], dtype="float32") + output = layers.maxout(x=data, groups=2) self.assertIsNotNone(output) - output = layers.upsampling_bilinear2d(x, scale=3) + print(str(program)) + + def test_crop(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[3, 5], dtype="float32") + y = layers.data(name='y', shape=[2, 3], dtype="float32") + output = layers.crop(x, shape=y) self.assertIsNotNone(output) print(str(program)) + def test_mean_iou(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[16], dtype='float32') + y = layers.data(name='label', shape=[1], dtype='int64') + iou = layers.mean_iou(x, y, 2) + self.assertIsNotNone(iou) + print(str(program)) + + def test_argsort(self): + program = Program() + with program_guard(program): + data = layers.data(name='x', shape=[2, 3, 3], dtype="float32") + out, ids = layers.argsort(input=data, axis=1) + self.assertIsNotNone(out) + self.assertIsNotNone(ids) + print(str(program)) + + def test_rank_loss(self): + program = Program() + with program_guard(program): + label = layers.data( + name='label', + append_batch_size=False, + shape=[16, 1], + dtype="float32") + left = layers.data( + name='left', + append_batch_size=False, + shape=[16, 1], + dtype="float32") + right = layers.data( + name='right', + append_batch_size=False, + shape=[16, 1], + dtype="float32") + out = layers.rank_loss(label, left, right, name="rank_loss") + self.assertIsNotNone(out) + print(str(program)) + + def test_flatten(self): + program = Program() + with program_guard(program): + x = layers.data( + name='x', + append_batch_size=False, + shape=[4, 4, 3], + dtype="float32") + out = layers.flatten(x, axis=1, name="flatten") + self.assertIsNotNone(out) + + def test_shape(self): + program = Program() + with program_guard(program): + input = layers.data( + name="input", shape=[3, 100, 100], dtype="float32") + out = layers.shape(input, name="shape") + self.assertIsNotNone(out) + print(str(program)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py index 6382e290eb30c621da64d5c600be6d8a7c6254f1..e628195e7265ec564bd64a212c4a35fdff495063 100644 --- a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py +++ b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py @@ -91,20 +91,21 @@ class TestLearningRateDecay(unittest.TestCase): def check_decay_with_place(self, place, python_decay_fn, fluid_decay_fn, kwargs): + main_prog = fluid.Program() + startup_prog = fluid.Program() - decayed_lr = fluid_decay_fn(**kwargs) + with fluid.program_guard(main_prog, startup_prog): + decayed_lr = fluid_decay_fn(**kwargs) place = fluid.CPUPlace() exe = fluid.Executor(place) - exe.run(fluid.default_startup_program()) + exe.run(startup_prog) - fluid.memory_optimize(fluid.default_main_program()) + fluid.memory_optimize(main_prog) for step in range(10): - lr_val, = exe.run(fluid.default_main_program(), - feed={}, - fetch_list=[decayed_lr]) + lr_val, = exe.run(main_prog, feed={}, fetch_list=[decayed_lr]) python_decayed_lr = python_decay_fn( global_step=float(step), **kwargs) self.assertAlmostEqual( diff --git a/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py b/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py index f49f7635f76c9feb5b5593438cb445df9488c69b..696d0ab4fa81a409a2bf0d6f6f23779ec26eb6d2 100644 --- a/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py +++ b/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py @@ -105,11 +105,13 @@ class TestLinearChainCrfOp(OpTest): MAX_SEQ_LEN = 5 # the linear_chain_crf operator only supports sequence (LoD level = 1) - lod = [[0]] + lod = [[]] + seq_start_pos = [0] for i in range(SEQ_NUM): - lod[-1].append(lod[-1][-1] + random.randint(1, MAX_SEQ_LEN)) - emission = np.random.uniform(-1, 1, - [lod[-1][-1], TAG_NUM]).astype("float64") + lod[-1].append(random.randint(1, MAX_SEQ_LEN)) + seq_start_pos.append(seq_start_pos[-1] + lod[-1][-1]) + emission = np.random.uniform( + -1, 1, [seq_start_pos[-1], TAG_NUM]).astype("float64") emission_row_max = np.amax(emission, axis=1, keepdims=True) emission_exps = np.exp(emission - emission_row_max) @@ -118,14 +120,14 @@ class TestLinearChainCrfOp(OpTest): transition_exps = np.exp(transition) labels = np.random.randint( - low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int64") + low=0, high=TAG_NUM, size=(seq_start_pos[-1], 1), dtype="int64") self.inputs = { "Emission": (emission, lod), "Transition": transition, "Label": (labels, lod) } - crf = LinearChainCrfForward(lod[0], emission, emission_row_max, + crf = LinearChainCrfForward(seq_start_pos, emission, emission_row_max, emission_exps, transition, transition_exps, labels) alpha, log_likelihood = crf.crf_forward_compute() diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py index cf89f9d0ebf6200933e539ef7fa8cbdc8f6db058..1cdc69501043d120b9e3cc8ccda3a1212d205886 100644 --- a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -23,7 +23,7 @@ from multiprocessing import Process from op_test import OpTest -def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): +def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id): x = fluid.layers.data(name='x', shape=[1], dtype='float32') y_predict = fluid.layers.fc(input=x, size=1, act=None) y = fluid.layers.data(name='y', shape=[1], dtype='float32') @@ -39,15 +39,8 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) - port = os.getenv("PADDLE_INIT_PORT", port) - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS", ip) # ip,ip... - eplist = [] - for ip in pserver_ips.split(","): - eplist.append(':'.join([ip, port])) - pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS", trainer_count)) - current_endpoint = os.getenv("POD_IP", ip) + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID", trainer_id)) + pserver_endpoints = ip + ":" + port + current_endpoint = ip + ":" + port t = fluid.DistributeTranspiler() t.transpile( trainer_id, @@ -62,47 +55,55 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): class TestListenAndServOp(OpTest): def setUp(self): - self.sleep_time = 5 + self.ps_timeout = 5 self.ip = "127.0.0.1" - self.port = "6173" - self.trainer_count = 1 - self.trainer_id = 1 - - def _raise_signal(self, parent_pid, raised_signal): - time.sleep(self.sleep_time) - ps_command = subprocess.Popen( - "ps -o pid --ppid %d --noheaders" % parent_pid, - shell=True, - stdout=subprocess.PIPE) - ps_output = ps_command.stdout.read() - retcode = ps_command.wait() - assert retcode == 0, "ps command returned %d" % retcode - - for pid_str in ps_output.split("\n")[:-1]: - try: - os.kill(int(pid_str), raised_signal) - except Exception: - continue + self.port = "0" + self.trainers = 1 + self.trainer_id = 0 def _start_pserver(self, use_cuda, sync_mode): p = Process( target=run_pserver, - args=(use_cuda, sync_mode, self.ip, self.port, self.trainer_count, + args=(use_cuda, sync_mode, self.ip, self.port, self.trainers, self.trainer_id)) + p.daemon = True p.start() + return p + + def _wait_ps_ready(self, pid): + start_left_time = self.ps_timeout + sleep_time = 0.5 + while True: + assert start_left_time >= 0, "wait ps ready failed" + time.sleep(sleep_time) + try: + # the listen_and_serv_op would touch a file which contains the listen port + # on the /tmp directory until it was ready to process all the RPC call. + os.stat("/tmp/paddle.%d.port" % pid) + return + except os.error: + start_left_time -= sleep_time + + def test_rpc_interfaces(self): + # TODO(Yancey1989): need to make sure the rpc interface correctly. + pass def test_handle_signal_in_serv_op(self): # run pserver on CPU in sync mode - self._start_pserver(False, True) + p1 = self._start_pserver(False, True) + self._wait_ps_ready(p1.pid) - # raise SIGINT to pserver - self._raise_signal(os.getpid(), signal.SIGINT) + # raise SIGTERM to pserver + os.kill(p1.pid, signal.SIGINT) + p1.join() # run pserver on CPU in async mode - self._start_pserver(False, False) + p2 = self._start_pserver(False, False) + self._wait_ps_ready(p2.pid) # raise SIGTERM to pserver - self._raise_signal(os.getpid(), signal.SIGTERM) + os.kill(p2.pid, signal.SIGTERM) + p2.join() if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_lod_rank_table.py b/python/paddle/fluid/tests/unittests/test_lod_rank_table.py index 093eecb8370b8ae7e4c43ce7ca6f50f5d302bd60..d53ead381d301e797d5a19784aed49a5d6f99319 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_rank_table.py +++ b/python/paddle/fluid/tests/unittests/test_lod_rank_table.py @@ -12,7 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. -from paddle.fluid.layers import lod_rank_table, data +from paddle.fluid.layers import data +from paddle.fluid.layers.control_flow import lod_rank_table from paddle.fluid.executor import Executor import paddle.fluid.core as core import numpy @@ -30,11 +31,12 @@ class TestLoDRankTable(unittest.TestCase): tensor = core.LoDTensor() tensor.set(numpy.random.random(size=(17, 100)), cpu) - tensor.set_lod([[0, 1, 3], [0, 5, 6, 7], [0, 3, 4, 9, 10, 13, 16, 17]]) + tensor.set_recursive_sequence_lengths( + [[1, 2], [5, 1, 1], [3, 1, 5, 1, 3, 3, 1]]) exe.run(scope=scope, feed={'x': tensor}) var = scope.find_var(rank_table.name) table = var.get_lod_rank_table() - self.assertEqual([(0, 5), (1, 1), (2, 1)], table.items()) + self.assertEqual([(0, 5), (1, 1), (2, 1)], list(table.items())) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py index 6b6d4c824aeae319dacf224408ce96a0d9c5bb35..77905c4b96499c855fd5c5e704b8051ccdb7a323 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py +++ b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py @@ -21,11 +21,15 @@ class TestLodResetOpByAttr(OpTest): def setUp(self): self.op_type = "lod_reset" x = np.random.random((10, 20)).astype("float32") - lod = [[0, 3, 5, 10]] - target_lod_0 = [0, 7, 10] + lod = [[3, 2, 5]] + # target_offset_lod and target_lod are the same lod info represented + # in offset-based format and length-based format, respectively. + target_offset_lod = [0, 7, 10] + target_lod = [7, 3] self.inputs = {'X': (x, lod)} - self.attrs = {'target_lod': target_lod_0} - self.outputs = {'Out': (x, [target_lod_0])} + # The `target_lod` attribute is still based on offset + self.attrs = {'target_lod': target_offset_lod} + self.outputs = {'Out': (x, [target_lod])} def test_check_output(self): self.check_output() @@ -38,13 +42,16 @@ class TestLodResetOpByInput(OpTest): def setUp(self): self.op_type = "lod_reset" x = np.random.random((10, 20)).astype("float32") - lod = [[0, 3, 5, 10]] - target_lod_0 = [0, 4, 7, 10] + lod = [[3, 2, 5]] + # target_offset_lod and target_lod are the same lod info represented + # in offset-based format and length-based format, respectively. + target_offset_lod = [0, 4, 7, 10] + target_lod = [4, 3, 3] self.inputs = { 'X': (x, lod), - 'Y': np.array([target_lod_0]).astype('int32') + 'Y': np.array([target_offset_lod]).astype('int32') } - self.outputs = {'Out': (x, [target_lod_0])} + self.outputs = {'Out': (x, [target_lod])} def test_check_output(self): self.check_output() @@ -57,15 +64,16 @@ class TestLodResetOpBoth(OpTest): def setUp(self): self.op_type = "lod_reset" x = np.random.random((10, 20)).astype("float32") - lod = [[0, 3, 5, 10]] - target_lod_0_attr = [0, 7, 10] - target_lod_0_in = [0, 4, 7, 10] + lod = [[3, 2, 5]] + target_offset_lod_attr = [0, 7, 10] + target_offset_lod_in = [0, 4, 7, 10] + target_lod_in = [4, 3, 3] self.inputs = { 'X': (x, lod), - 'Y': np.array(target_lod_0_in).astype('int32') + 'Y': np.array(target_offset_lod_in).astype('int32') } - self.attrs = {'target_lod': target_lod_0_attr} - self.outputs = {'Out': (x, [target_lod_0_in])} + self.attrs = {'target_lod': target_offset_lod_attr} + self.outputs = {'Out': (x, [target_lod_in])} def test_check_output(self): self.check_output() @@ -78,11 +86,11 @@ class TestLodResetOpYIsLoDTensor(OpTest): def setUp(self): self.op_type = "lod_reset" x = np.random.random((10, 20)).astype("float32") - lod = [[0, 3, 5, 10]] + lod = [[3, 2, 5]] y = np.random.random((10, 10)).astype("float32") - target_lod_0 = [[0, 4, 7, 10]] - self.inputs = {'X': (x, lod), 'Y': (y, target_lod_0)} - self.outputs = {'Out': (x, target_lod_0)} + target_lod = [[4, 3, 3]] + self.inputs = {'X': (x, lod), 'Y': (y, target_lod)} + self.outputs = {'Out': (x, target_lod)} def test_check_output(self): self.check_output() diff --git a/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py b/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py index 63b17a5ccd62ed79b3d611e039c2b2705a133272..0ac6d9b81df0ecbe9c6560cdb0ab0507c3c2ed18 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py +++ b/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py @@ -24,28 +24,28 @@ class TestLoDTensorArray(unittest.TestCase): tensor_array = arr.get_lod_tensor_array() self.assertEqual(0, len(tensor_array)) cpu = core.CPUPlace() - for i in xrange(10): + for i in range(10): t = core.LoDTensor() t.set(numpy.array([i], dtype='float32'), cpu) - t.set_lod([[0, 1]]) + t.set_recursive_sequence_lengths([[1]]) tensor_array.append(t) self.assertEqual(10, len(tensor_array)) - for i in xrange(10): + for i in range(10): t = tensor_array[i] self.assertEqual(numpy.array(t), numpy.array([i], dtype='float32')) - self.assertEqual([[0, 1]], t.lod()) + self.assertEqual([[1]], t.recursive_sequence_lengths()) t = core.LoDTensor() t.set(numpy.array([i + 10], dtype='float32'), cpu) - t.set_lod([[0, 2]]) + t.set_recursive_sequence_lengths([[1]]) tensor_array[i] = t t = tensor_array[i] self.assertEqual( numpy.array(t), numpy.array( [i + 10], dtype='float32')) - self.assertEqual([[0, 2]], t.lod()) + self.assertEqual([[1]], t.recursive_sequence_lengths()) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py b/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py index 66a03640c148d769787593f41a44cd4d1aaa10b1..9789ff4af648b41a1b53844be89249bd260de61b 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py +++ b/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py @@ -20,6 +20,11 @@ from paddle.fluid.framework import Program, program_guard from paddle.fluid.executor import Executor from paddle.fluid.backward import append_backward +from paddle.fluid.layers.control_flow import lod_rank_table +from paddle.fluid.layers.control_flow import max_sequence_len +from paddle.fluid.layers.control_flow import lod_tensor_to_array +from paddle.fluid.layers.control_flow import array_to_lod_tensor + class TestCPULoDTensorArrayOps(unittest.TestCase): def place(self): @@ -29,9 +34,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor = core.LoDTensor() tensor.set( numpy.arange(10).reshape(10, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 3, 9, 10]]) - expect = map(lambda x: numpy.array(x).astype('int32'), - [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]]) + tensor.set_recursive_sequence_lengths([[3, 6, 1]]) + expect = [ + numpy.array(x).astype('int32') + for x in [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]] + ] self.main( tensor=tensor, expect_array=expect, @@ -42,9 +49,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor = core.LoDTensor() tensor.set( numpy.arange(10).reshape(10, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 3, 9, 9, 10]]) - expect = map(lambda x: numpy.array(x).astype('int32'), - [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]]) + tensor.set_recursive_sequence_lengths([[3, 6, 0, 1]]) + expect = [ + numpy.array(x).astype('int32') + for x in [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]] + ] self.main( tensor=tensor, expect_array=expect, @@ -55,7 +64,7 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor = core.LoDTensor() tensor.set( numpy.arange(20).reshape(20, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 2, 5], [0, 3, 9, 11, 17, 20]]) + tensor.set_recursive_sequence_lengths([[2, 3], [3, 6, 2, 6, 3]]) expect = [ numpy.array( @@ -65,7 +74,7 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): [17, 18, 19], dtype='int32') ] - lod = [[[0, 2, 5]], [[0, 6, 12]], [[0, 3]]] + lod = [[[2, 3]], [[6, 6]], [[3]]] self.main( tensor=tensor, expect_array=expect, @@ -77,8 +86,8 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor.set( numpy.arange(31).reshape(31, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 3, 5, 9, 11], - [0, 3, 7, 11, 11, 12, 17, 19, 21, 23, 30, 31]]) + tensor.set_recursive_sequence_lengths( + [[3, 2, 4, 2], [3, 4, 4, 0, 1, 5, 2, 2, 2, 7, 1]]) expect = [ numpy.array( @@ -88,7 +97,7 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): ], [17, 18, 3, 4, 5, 6, 11, 30], [19, 20, 7, 8, 9, 10], [21, 22]] ] - lod = [[[0, 5, 8, 8, 15]], [[0, 2, 6, 7, 8]], [[0, 2, 6]], [[0, 2]]] + lod = [[[5, 3, 0, 7]], [[2, 4, 1, 1]], [[2, 4]], [[2]]] self.main( tensor=tensor, expect_array=expect, @@ -99,17 +108,18 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor = core.LoDTensor() tensor.set( numpy.arange(50).reshape(50, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 2, 5, 6], [0, 2, 5, 6, 10, 12, 13], - [0, 3, 7, 11, 17, 21, 22, 23, 27, 31, 39, 45, 46, 50]]) + tensor.set_recursive_sequence_lengths( + [[2, 3, 1], [2, 3, 1, 4, 2, 1], + [3, 4, 4, 6, 4, 1, 1, 4, 4, 8, 6, 1, 4]]) expect = [ numpy.array( item, dtype='int32') - for item in [[21, 0, 1, 2, 3, 4, 5, 6, 46, 47, 48, 49], range( - 22, 39) + range(7, 21), range(39, 46)] + for item in [[21, 0, 1, 2, 3, 4, 5, 6, 46, 47, 48, 49], list( + range(22, 39)) + list(range(7, 21)), list(range(39, 46))] ] - lod = [[[0, 1, 3, 4], [0, 1, 4, 8, 12]], - [[0, 4, 7], [0, 1, 5, 9, 17, 21, 27, 31]], [[0, 2], [0, 6, 7]]] + lod = [[[1, 2, 1], [1, 3, 4, 4]], [[4, 3], [1, 4, 4, 8, 4, 6, 4]], + [[2], [6, 1]]] self.main( tensor=tensor, expect_array=expect, @@ -120,8 +130,9 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor = core.LoDTensor() tensor.set( numpy.arange(50).reshape(50, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 2, 5, 6], [0, 2, 5, 6, 10, 12, 13], - [0, 3, 7, 11, 17, 21, 22, 23, 27, 31, 39, 45, 46, 50]]) + tensor.set_recursive_sequence_lengths( + [[2, 3, 1], [2, 3, 1, 4, 2, 1], + [3, 4, 4, 6, 4, 1, 1, 4, 4, 8, 6, 1, 4]]) self.main( tensor=tensor, expect_array=None, @@ -135,13 +146,13 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): with program_guard(program): x = layers.data(name='x', shape=[10]) x.persistable = True - table = layers.lod_rank_table(x, level=level) - max_len = layers.max_sequence_len(table) + table = lod_rank_table(x, level=level) + max_len = max_sequence_len(table) max_len.persistable = True - array = layers.lod_tensor_to_array(x, table) + array = lod_tensor_to_array(x, table) array.persistable = True - result = layers.array_to_lod_tensor(array, table) + result = array_to_lod_tensor(array, table) result.persistable = True exe = Executor(place) scope = core.Scope() @@ -162,12 +173,13 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): exp_tensor, exp_lod = exp exp_tensor = numpy.expand_dims(exp_tensor, axis=1) self.assertTrue(numpy.allclose(exp_tensor, numpy.array(array[i]))) - self.assertEqual(exp_lod, array[i].lod()) + self.assertEqual(exp_lod, array[i].recursive_sequence_lengths()) def check_tensor_same(self, actual, expect): self.assertTrue( numpy.allclose(numpy.array(actual), numpy.array(expect))) - self.assertEqual(actual.lod(), expect.lod()) + self.assertEqual(actual.recursive_sequence_lengths(), + expect.recursive_sequence_lengths()) class TestCPULoDTensorArrayOpGrad(unittest.TestCase): @@ -178,9 +190,9 @@ class TestCPULoDTensorArrayOpGrad(unittest.TestCase): with program_guard(program): x = layers.data( name='x', shape=[1], dtype='float32', stop_gradient=False) - table = layers.lod_rank_table(x, level=0) - array = layers.lod_tensor_to_array(x, table) - result = layers.array_to_lod_tensor(array, table) + table = lod_rank_table(x, level=0) + array = lod_tensor_to_array(x, table) + result = array_to_lod_tensor(array, table) mean = layers.mean(result) @@ -188,7 +200,7 @@ class TestCPULoDTensorArrayOpGrad(unittest.TestCase): tensor = core.LoDTensor() tensor.set(numpy.arange(10).reshape(10, 1).astype('float32'), place) - tensor.set_lod([[0, 3, 9, 10]]) + tensor.set_recursive_sequence_lengths([[3, 6, 1]]) g_vars = program.global_block().var(x.name + "@GRAD") diff --git a/python/paddle/fluid/tests/unittests/test_lookup_table_op.py b/python/paddle/fluid/tests/unittests/test_lookup_table_op.py index f8d5785fbfe64843f4aa3b96b24809df60980c74..ac25f432dffd544d4b336983ec868f2431a5b91a 100644 --- a/python/paddle/fluid/tests/unittests/test_lookup_table_op.py +++ b/python/paddle/fluid/tests/unittests/test_lookup_table_op.py @@ -35,77 +35,59 @@ class TestLookupTableOp(OpTest): self.check_grad(['W'], 'Out', no_grad_set=set('Ids')) +class TestLookupTableOpWithTensorIds(OpTest): + def setUp(self): + self.op_type = "lookup_table" + table = np.random.random((17, 31)).astype("float32") + ids = np.random.randint( + low=0, high=17, size=(2, 4, 5, 1)).astype("int64") + self.inputs = {'W': table, 'Ids': ids} + self.outputs = {'Out': table[ids.flatten()].reshape((2, 4, 5, 31))} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['W'], 'Out', no_grad_set=set('Ids')) + + class TestLookupTableOpWithPadding(TestLookupTableOp): def test_check_output(self): ids = np.squeeze(self.inputs['Ids']) padding_idx = np.random.choice(ids, 1)[0] self.outputs['Out'][ids == padding_idx] = np.zeros(31) - self.attrs = {'padding_idx': long(padding_idx)} + self.attrs = {'padding_idx': int(padding_idx)} self.check_output() def test_check_grad(self): - # Since paddings are not trainable and fixed in forward, the gradient of + # Since paddings are not trainable and fixed in forward, the gradient of # paddings makes no sense and we don't test the gradient here. pass -class TestLookupTableIdsIsSelectedRows(OpTest): - def check_with_place(self, place): - scope = core.Scope() - - # create and initialize Variable - height = 10 - rows = [0, 4, 4, 7] - row_numel = 12 - - # create and initialize W Variable - W = scope.var('W').get_tensor() - W_array = np.full((height, row_numel), 1.0).astype("float32") - for i in range(height): - W_array[i] *= i - W.set(W_array, place) - - # create and initialize Ids Variable - ids_selected_rows = scope.var('Ids').get_selected_rows() - ids_selected_rows.set_height(len(rows)) - ids_selected_rows.set_rows(rows) - np_array = np.ones((len(rows), row_numel)).astype("float32") - ids_tensor = ids_selected_rows.get_tensor() - ids_tensor.set(np_array, place) - - # create Out Variable - Out = scope.var('Out').get_selected_rows() - - # create and run lookup_table operator - concat_rows_op = Operator("lookup_table", W='W', Ids='Ids', Out='Out') - concat_rows_op.run(scope, place) - - # get result from Out - Out_tensor = Out.get_tensor() - result_array = np.array(Out_tensor) - - # all(): return True if all elements of the iterable are true (or if the iterable is empty) - for idx, row in enumerate(rows): - assert (row == result_array[idx]).all() +class TestLookupTableOpWithTensorIdsAndPadding(TestLookupTableOpWithTensorIds): + def test_check_output(self): + ids = self.inputs['Ids'] + flatten_idx = ids.flatten() + padding_idx = np.random.choice(flatten_idx, 1)[0] + self.outputs['Out'][np.squeeze(ids == padding_idx)] = np.zeros(31) + self.attrs = {'padding_idx': long(padding_idx)} + self.check_output() - def test_concat_rows(self): - places = [core.CPUPlace()] - if core.is_compiled_with_cuda(): - places.append(core.CUDAPlace(0)) - for place in places: - self.check_with_place(place) + def test_check_grad(self): + # Since paddings are not trainable and fixed in forward, the gradient of + # paddings makes no sense and we don't test the gradient here. + pass class TestLookupTableWIsSelectedRows(OpTest): - def check_with_place(self, place): - scope = core.Scope() - - # create and initialize Id Variable + def prepare_ids(self, scope, place): ids_tensor = scope.var('Ids').get_tensor() ids_array = np.array([[0], [4], [3], [5]]).astype("int64") ids_tensor.set(ids_array, place) + return ids_array - # create and initialize W Variable + def prepare_w(self, scope, place): rows = [0, 1, 2, 3, 4, 5, 6] row_numel = 12 @@ -118,8 +100,22 @@ class TestLookupTableWIsSelectedRows(OpTest): w_tensor = w_selected_rows.get_tensor() w_tensor.set(w_array, place) - # create Out Variable - out_tensor = scope.var('Out').get_tensor() + def create_out_tensor(self, scope, place): + return scope.var('Out').get_tensor() + + def check_result(self, ids_array, result_array): + # all(): return True if all elements of the iterable are true (or if the iterable is empty) + for idx, row in enumerate(ids_array): + assert (row[0] == result_array[idx]).all() + + def check_with_place(self, place): + scope = core.Scope() + + ids_array = self.prepare_ids(scope, place) + + self.prepare_w(scope, place) + + out_tensor = self.create_out_tensor(scope, place) # create and run lookup_table operator lookup_table = Operator("lookup_table", W='W', Ids='Ids', Out='Out') @@ -127,9 +123,8 @@ class TestLookupTableWIsSelectedRows(OpTest): # get result from Out result_array = np.array(out_tensor) - # all(): return True if all elements of the iterable are true (or if the iterable is empty) - for idx, row in enumerate(ids_array): - assert (row[0] == result_array[idx]).all() + + self.check_result(ids_array, result_array) def test_w_is_selected_rows(self): places = [core.CPUPlace()] @@ -138,5 +133,19 @@ class TestLookupTableWIsSelectedRows(OpTest): self.check_with_place(place) +class TestLookupTableWithTensorIdsWIsSelectedRows( + TestLookupTableWIsSelectedRows): + def prepare_ids(self, scope, place): + ids_tensor = scope.var('Ids').get_tensor() + ids_array = np.random.randint( + low=0, high=6, size=(2, 4, 3, 1)).astype("int64") + ids_tensor.set(ids_array, place) + return ids_array + + def check_result(self, ids_array, result_array): + for idx, row in np.ndenumerate(ids_array): + assert (row == result_array[idx]).all() + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_lstm_op.py b/python/paddle/fluid/tests/unittests/test_lstm_op.py index f8ff5a3361af66612f08b2aa4eaffa363f04c594..705a24bd8f39a55e0a352944d961f8d33aaf96ff 100644 --- a/python/paddle/fluid/tests/unittests/test_lstm_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstm_op.py @@ -84,15 +84,17 @@ def lstm( h = g_o * act_cell(c) return h, c - def _reverse(x, lod): + def _reverse(x, offset): y = np.zeros_like(x) - for i in range(len(lod) - 1): - b, e = lod[i], lod[i + 1] + for i in range(len(offset) - 1): + b, e = offset[i], offset[i + 1] y[b:e, :] = np.flip(x[b:e, :], 0) return y - offset = lod[0] - batch_size = len(offset) - 1 + offset = [0] + for l in lod[0]: + offset.append(offset[-1] + l) + batch_size = len(lod[0]) hidden = [] cell = [] input = _reverse(input, offset) if is_reverse else input @@ -100,7 +102,7 @@ def lstm( input = input + np.tile(w_b, (offset[-1], 1)) for i in range(batch_size): # compute one sequence - seq_len = offset[i + 1] - offset[i] + seq_len = lod[0][i] x = input[offset[i]:offset[i + 1], :] h_pre = h0[i] # 1 x D c_pre = c0[i] # 1 x D @@ -124,7 +126,7 @@ def lstm( class TestLstmOp(OpTest): def set_argument(self): - self.lod = [[0, 2, 5, 7]] + self.lod = [[2, 3, 2]] self.D = 16 self.act_gate = 'sigmoid' @@ -139,8 +141,8 @@ class TestLstmOp(OpTest): self.set_argument() self.op_type = 'lstm' - T = self.lod[0][-1] - N = len(self.lod[0]) - 1 + T = sum(self.lod[0]) + N = len(self.lod[0]) x = np.random.normal(size=(T, 4 * self.D)).astype('float64') if self.has_initial_state: @@ -186,7 +188,7 @@ class TestLstmOp(OpTest): def test_check_grad(self): # TODO(qingqing) remove folowing lines after the check_grad is refined. - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( (N, self.D)).astype('float64') @@ -194,107 +196,104 @@ class TestLstmOp(OpTest): ['Input', 'Weight', 'Bias'], ['Hidden'], max_relative_error=5e-4) -class TestLstmOpHasInitial(TestLstmOp): - def set_argument(self): - self.lod = [[0, 2, 5, 7]] - self.D = 16 - - self.act_gate = 'sigmoid' - self.act_cell = 'tanh' - self.act_cand = 'tanh' - - self.has_initial_state = True - self.is_reverse = True - self.use_peepholes = True - - def test_check_grad(self): - # TODO(qingqing) remove folowing lines after the check_grad is refined. - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'], - max_relative_error=5e-4) - - def test_check_grad_ingore_bias(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('Bias')) - - def test_check_grad_ingore_weight(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Bias'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('Weight')) - - def test_check_grad_ingore_input(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Weight', 'Bias'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('Input')) - - def test_check_grad_ingore_h0(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight', 'Bias', 'C0'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('H0')) - - def test_check_grad_ingore_c0(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight', 'Bias', 'H0'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('C0')) - - -class TestLstmOpRerverse(TestLstmOp): - def set_argument(self): - self.lod = [[0, 2, 5, 7]] - self.D = 16 - - self.act_gate = 'sigmoid' - self.act_cell = 'tanh' - self.act_cand = 'tanh' - - self.has_initial_state = False - self.is_reverse = True - self.use_peepholes = True - - -class TestLstmOpNotUsePeepholes(TestLstmOp): - def set_argument(self): - self.lod = [[0, 2, 5, 7]] - self.D = 16 - - self.act_gate = 'sigmoid' - self.act_cell = 'tanh' - self.act_cand = 'tanh' - - self.has_initial_state = False - self.is_reverse = True - self.use_peepholes = False - +# class TestLstmOpHasInitial(TestLstmOp): +# def set_argument(self): +# self.lod = [[2, 3, 2]] +# self.D = 16 + +# self.act_gate = 'sigmoid' +# self.act_cell = 'tanh' +# self.act_cand = 'tanh' + +# self.has_initial_state = True +# self.is_reverse = True +# self.use_peepholes = True + +# def test_check_grad(self): +# # TODO(qingqing) remove folowing lines after the check_grad is refined. +# N = len(self.lod[0]) +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'], +# max_relative_error=5e-4) + +# def test_check_grad_ingore_bias(self): +# N = len(self.lod[0]) +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('Bias')) + +# def test_check_grad_ingore_weight(self): +# N = len(self.lod[0]) +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Bias'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('Weight')) + +# def test_check_grad_ingore_input(self): +# N = len(self.lod[0]) +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Weight', 'Bias'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('Input')) + +# def test_check_grad_ingore_h0(self): +# N = len(self.lod[0]) +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight', 'Bias', 'C0'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('H0')) + +# def test_check_grad_ingore_c0(self): +# N = len(self.lod[0]) +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight', 'Bias', 'H0'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('C0')) + +# class TestLstmOpRerverse(TestLstmOp): +# def set_argument(self): +# self.lod = [[2, 3, 2]] +# self.D = 16 + +# self.act_gate = 'sigmoid' +# self.act_cell = 'tanh' +# self.act_cand = 'tanh' + +# self.has_initial_state = False +# self.is_reverse = True +# self.use_peepholes = True + +# class TestLstmOpNotUsePeepholes(TestLstmOp): +# def set_argument(self): +# self.lod = [[2, 3, 2]] +# self.D = 16 + +# self.act_gate = 'sigmoid' +# self.act_cell = 'tanh' +# self.act_cand = 'tanh' + +# self.has_initial_state = False +# self.is_reverse = True +# self.use_peepholes = False if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_lstmp_op.py b/python/paddle/fluid/tests/unittests/test_lstmp_op.py index afff133f6c6cfe45d1aca4014dc8b92e6562e6b8..ed2262da4bc727657c2e65d69cb1922891e17b09 100644 --- a/python/paddle/fluid/tests/unittests/test_lstmp_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstmp_op.py @@ -64,15 +64,17 @@ def lstmp( r = act_proj(r) return r, c - def _reverse(x, lod): + def _reverse(x, offset): y = np.zeros_like(x) - for i in range(len(lod) - 1): - b, e = lod[i], lod[i + 1] + for i in range(len(offset) - 1): + b, e = offset[i], offset[i + 1] y[b:e, :] = np.flip(x[b:e, :], 0) return y - offset = lod[0] - batch_size = len(offset) - 1 + offset = [0] + for l in lod[0]: + offset.append(offset[-1] + l) + batch_size = len(lod[0]) # recurrent projection state projection = [] cell = [] @@ -81,7 +83,7 @@ def lstmp( input = input + np.tile(w_b, (offset[-1], 1)) for i in range(batch_size): # compute one sequence - seq_len = offset[i + 1] - offset[i] + seq_len = lod[0][i] x = input[offset[i]:offset[i + 1], :] r_pre = np.dot(h0[i], w_rh) # 1 x P r_pre = act_proj(r_pre) @@ -117,8 +119,8 @@ class TestLstmpOp(LstmTest.TestLstmOp): self.reset_argument() self.op_type = 'lstmp' - T = self.lod[0][-1] - N = len(self.lod[0]) - 1 + T = sum(self.lod[0]) + N = len(self.lod[0]) x = np.random.normal(size=(T, 4 * self.D)).astype('float64') if self.has_initial_state: @@ -166,7 +168,7 @@ class TestLstmpOp(LstmTest.TestLstmOp): def test_check_grad(self): # TODO(qingqing) remove folowing lines after the check_grad is refined. - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -183,7 +185,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): def test_check_grad(self): # TODO(qingqing) remove folowing lines after the check_grad is refined. - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -195,7 +197,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): max_relative_error=1e-2) def test_check_grad_ingore_bias(self): - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -207,7 +209,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): no_grad_set=set('Bias')) def test_check_grad_ingore_weight(self): - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -219,7 +221,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): no_grad_set=set('Weight')) def test_check_grad_ingore_proj_weight(self): - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -231,7 +233,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): no_grad_set=set('ProjWeight')) def test_check_grad_ingore_input(self): - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -243,7 +245,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): no_grad_set=set('Input')) def test_check_grad_ingore_h0(self): - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') @@ -255,7 +257,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): no_grad_set=set('H0')) def test_check_grad_ingore_c0(self): - N = len(self.lod[0]) - 1 + N = len(self.lod[0]) self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') diff --git a/python/paddle/fluid/tests/unittests/test_mean_iou.py b/python/paddle/fluid/tests/unittests/test_mean_iou.py new file mode 100644 index 0000000000000000000000000000000000000000..32b4ee184787cd4cda0fd889f67a609141a3cb27 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_mean_iou.py @@ -0,0 +1,114 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import division +import unittest +import numpy as np +from op_test import OpTest + + +def compute_mean_iou(predictions, labels, num_classes, in_wrongs, in_corrects, + in_mean_ious): + assert predictions.shape == labels.shape + predictions = predictions.flatten() + labels = labels.flatten() + + out_wrong = np.zeros([num_classes]).astype("int32") + for _, wrong in in_wrongs: + out_wrong += wrong + out_correct = np.zeros([num_classes]).astype("int32") + for _, correct in in_corrects: + out_correct += correct + + for pred, label in zip(predictions, labels): + if pred == label: + out_correct[pred] += 1 + else: + out_wrong[pred] += 1 + out_wrong[label] += 1 + + denominator = out_wrong + out_correct + valid_count = (denominator != 0).sum() + denominator = np.where(denominator > 0, denominator, + np.ones(denominator.shape)) + mean_iou = (out_correct / denominator).sum() / valid_count + + for _, in_mean_iou in in_mean_ious: + mean_iou += in_mean_iou + return mean_iou, out_wrong, out_correct + + +class TestMeanIOUOp(OpTest): + def setUp(self): + self.config() + self.op_type = "mean_iou" + predictions = np.random.randint(0, self.num_classes, + self.image_size).astype("int32") + labels = np.random.randint(0, self.num_classes, + self.image_size).astype("int32") + + in_wrongs = [] + for i in range(self.in_wrong_num): + in_wrongs.append(("in_wrong_%d" % i, np.random.randint( + 0, 10, [self.num_classes]).astype("int32"))) + + in_corrects = [] + for i in range(self.in_correct_num): + in_corrects.append(("in_correct_%d" % i, np.random.randint( + 0, 10, [self.num_classes]).astype("int32"))) + + in_mean_ious = [] + for i in range(self.in_mean_iou_num): + in_mean_ious.append(("in_mean_iou_%d" % i, np.random.uniform( + 0, 1, [1]).astype("float32"))) + + self.inputs = { + 'Predictions': predictions, + 'Labels': labels, + 'InWrongs': in_wrongs, + 'InCorrects': in_corrects, + 'InMeanIou': in_mean_ious + } + self.attrs = {'num_classes': int(self.num_classes)} + mean_iou, out_wrong, out_correct = compute_mean_iou( + predictions, labels, self.num_classes, in_wrongs, in_corrects, + in_mean_ious) + self.outputs = { + 'OutMeanIou': mean_iou, + 'OutWrong': out_wrong, + 'OutCorrect': out_correct + } + + def config(self): + self.num_classes = 10 + self.image_size = [128, 128] + self.in_wrong_num = 0 + self.in_correct_num = 0 + self.in_mean_iou_num = 0 + + def test_check_output(self): + self.check_output() + + +class TestCase1(TestMeanIOUOp): + def config(self): + self.num_classes = 5 + self.image_size = [100, 128] + self.in_wrong_num = 2 + self.in_correct_num = 2 + self.in_mean_iou_num = 2 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py b/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py index cfd6e63e12258a92447e68b4afbc7ead91b68cc1..67733807f8f8582f68dcfa3f361e13a631a29597 100644 --- a/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py @@ -43,5 +43,29 @@ class TestControlFlowGraph(unittest.TestCase): print(str(result_program)) +class TestMemoryTranspiler2(unittest.TestCase): + def setUp(self): + program = Program() + with program_guard(program, startup_program=Program()): + x = layers.data(name='x', shape=[13], dtype='float32') + fc = layers.fc(input=x, size=10, act=None) + reshape = layers.reshape(x=fc, shape=[-1, 2, 5]) + fc = layers.reshape(x=reshape, shape=[-1, 5, 2]) + y_predict = layers.fc(input=fc, size=1, act=None) + y = layers.data(name='y', shape=[1], dtype='float32') + cost = layers.square_error_cost(input=y_predict, label=y) + avg_cost = layers.mean(cost) + opt = optimizer.SGD(learning_rate=0.001) + opt.minimize(avg_cost) + self.program = program + + def test_inplace_ops(self): + print("before optimization") + print(str(self.program)) + result_program = memory_optimize(self.program) + print("after optimization") + print(str(result_program)) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_memory_usage.py b/python/paddle/fluid/tests/unittests/test_memory_usage.py new file mode 100644 index 0000000000000000000000000000000000000000..f9daf83652e18faab0ab31402b9f5889a0beceaf --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_memory_usage.py @@ -0,0 +1,69 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function +import paddle +import paddle.fluid as fluid +import contextlib +import unittest + + +def train_simulator(test_batch_size=10): + if test_batch_size <= 0: + raise ValueError("batch_size should be a positive integeral value, " + "but got batch_size={}".format(test_batch_size)) + + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + # Calculate memory usage in current network config + lower_usage, upper_usage, unit = fluid.contrib.memory_usage( + fluid.default_main_program(), batch_size=test_batch_size) + + print("memory usage is about %.3f - %.3f %s" % + (lower_usage, upper_usage, unit)) + + +class TestMemoryUsage(unittest.TestCase): + def test_with_unit_B(self): + with self.program_scope_guard(): + train_simulator() + + def test_with_unit_KB(self): + with self.program_scope_guard(): + train_simulator(test_batch_size=1000) + + def test_with_unit_MB(self): + with self.program_scope_guard(): + train_simulator(test_batch_size=100000) + + @contextlib.contextmanager + def program_scope_guard(self): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_merge_ids_op.py b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f209bdf30faffc0b2c7932b7b10f384d6d61a831 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py @@ -0,0 +1,38 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestMergeIdsOp(OpTest): + def setUp(self): + self.op_type = "merge_ids" + ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64') + x0 = np.array([[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]]).astype('float32') + x1 = np.array([]).astype('float32') + x2 = np.array([[0.4, 0.5], [0.4, 0.5], [0.5, 0.6], + [0.5, 0.6]]).astype('float32') + out = np.array([[0.1, 0.2], [0.4, 0.5], [0.4, 0.5], [0.2, 0.3], + [0.5, 0.6], [0.5, 0.6], [0.3, 0.4]]).astype('float32') + self.inputs = {'Ids': ids, "X": [('x0', x0), ('x1', x1), ('x2', x2)]} + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py b/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py index c27573c3d69037bc48e0b6a90636b3f027f15a41..54ee85c1a7a539fe9517f32adb35ab99b5ae2a07 100644 --- a/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py +++ b/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py @@ -70,7 +70,7 @@ class TestMineHardExamplesOp(OpTest): self.updated_match_indices = self.match_indices - self.neg_indices_lod = [[0, 1, 2]] + self.neg_indices_lod = [[1, 1]] self.neg_indices = np.array([[1], [0]]).astype('int32') @@ -92,7 +92,7 @@ class TestMineHardExamplesOpHardExample(TestMineHardExamplesOp): self.updated_match_indices = np.array([[0, -1, -1], [-1, -1, -1]]).astype('int32') - self.neg_indices_lod = [[0, 1, 3]] + self.neg_indices_lod = [[1, 2]] self.neg_indices = np.array([[2], [0], [2]]).astype('int32') diff --git a/python/paddle/fluid/tests/unittests/test_momentum_op.py b/python/paddle/fluid/tests/unittests/test_momentum_op.py index aaea9c1809213c5707e8540eebbdd6f269836fdc..c75d3bd276a5b494090c1aa1fea0bb4f2c067173 100644 --- a/python/paddle/fluid/tests/unittests/test_momentum_op.py +++ b/python/paddle/fluid/tests/unittests/test_momentum_op.py @@ -39,7 +39,7 @@ class TestMomentumOp1(OpTest): velocity_out = mu * velocity + grad if use_nesterov: - param_out = param - grad * learning_rate + \ + param_out = param - grad * learning_rate - \ velocity_out * mu * learning_rate else: param_out = param - learning_rate * velocity_out @@ -75,7 +75,7 @@ class TestMomentumOp2(OpTest): velocity_out = mu * velocity + grad if use_nesterov: - param_out = param - grad * learning_rate + \ + param_out = param - grad * learning_rate - \ velocity_out * mu * learning_rate else: param_out = param - learning_rate * velocity_out diff --git a/python/paddle/fluid/tests/unittests/test_mul_op.py b/python/paddle/fluid/tests/unittests/test_mul_op.py index 862b7f8cb93620da4dd4673028776cfe565eeb0b..bbc782c1bce302df68ab30013f3a7667e51ed479 100644 --- a/python/paddle/fluid/tests/unittests/test_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_mul_op.py @@ -22,8 +22,8 @@ class TestMulOp(OpTest): def setUp(self): self.op_type = "mul" self.inputs = { - 'X': np.random.random((32, 84)).astype("float32"), - 'Y': np.random.random((84, 100)).astype("float32") + 'X': np.random.random((2, 5)).astype("float32"), + 'Y': np.random.random((5, 3)).astype("float32") } self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} @@ -46,13 +46,16 @@ class TestMulOp2(OpTest): def setUp(self): self.op_type = "mul" self.inputs = { - 'X': np.random.random((15, 4, 12, 10)).astype("float32"), - 'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32") + 'X': np.random.random((3, 4, 4, 3)).astype("float32"), + 'Y': np.random.random((2, 6, 1, 2, 3)).astype("float32") } - self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} - result = np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10), - self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9)) - result = result.reshape(15, 4, 8, 2, 9) + self.attrs = { + 'x_num_col_dims': 2, + 'y_num_col_dims': 2, + } + result = np.dot(self.inputs['X'].reshape(3 * 4, 4 * 3), + self.inputs['Y'].reshape(2 * 6, 1 * 2 * 3)) + result = result.reshape(3, 4, 1, 2, 3) self.outputs = {'Out': result} def test_check_output(self): @@ -73,9 +76,9 @@ class TestMulOp2(OpTest): class TestFP16MulOp1(OpTest): def setUp(self): self.op_type = "mul" - x = np.random.random((32, 84)).astype("float16") - y = np.random.random((84, 100)).astype("float16") - self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} + x = np.random.random((3, 5)).astype("float16") + y = np.random.random((5, 4)).astype("float16") + self.inputs = {'X': x.view(np.float16), 'Y': y.view(np.float16)} self.outputs = {'Out': np.dot(x, y)} def test_check_output(self): @@ -88,13 +91,15 @@ class TestFP16MulOp1(OpTest): class TestFP16MulOp2(OpTest): def setUp(self): self.op_type = "mul" - x = np.random.random((15, 4, 12, 10)).astype("float16") - y = np.random.random((4, 30, 8, 2, 9)).astype("float16") - self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} - self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} - result = np.dot( - x.reshape(15 * 4, 12 * 10), y.reshape(4 * 30, 8 * 2 * 9)) - result = result.reshape(15, 4, 8, 2, 9) + x = np.random.random((3, 4, 4, 3)).astype("float16") + y = np.random.random((2, 6, 1, 2, 3)).astype("float16") + self.inputs = {'X': x.view(np.float16), 'Y': y.view(np.float16)} + self.attrs = { + 'x_num_col_dims': 2, + 'y_num_col_dims': 2, + } + result = np.dot(x.reshape(3 * 4, 4 * 3), y.reshape(2 * 6, 1 * 2 * 3)) + result = result.reshape(3, 4, 1, 2, 3) self.outputs = {'Out': result} def test_check_output(self): diff --git a/python/paddle/fluid/tests/unittests/test_multi_file_reader.py b/python/paddle/fluid/tests/unittests/test_multi_file_reader.py index 3f940203b9393d266d75b50c9cbf62e89c36cbdf..cb0ea96ff69ce32b0bb1b49f0318c353aa08d388 100644 --- a/python/paddle/fluid/tests/unittests/test_multi_file_reader.py +++ b/python/paddle/fluid/tests/unittests/test_multi_file_reader.py @@ -39,17 +39,17 @@ class TestMultipleReader(unittest.TestCase): copyfile('./mnist_0.recordio', './mnist_1.recordio') copyfile('./mnist_0.recordio', './mnist_2.recordio') - def main(self, thread_num): + def main(self, is_test=False): file_list = [ './mnist_0.recordio', './mnist_1.recordio', './mnist_2.recordio' ] with fluid.program_guard(fluid.Program(), fluid.Program()): data_files = fluid.layers.open_files( filenames=file_list, - thread_num=thread_num, shapes=[(-1, 784), (-1, 1)], lod_levels=[0, 0], - dtypes=['float32', 'int64']) + dtypes=['float32', 'int64'], + is_test=is_test) img, label = fluid.layers.read_file(data_files) if fluid.core.is_compiled_with_cuda(): @@ -64,14 +64,16 @@ class TestMultipleReader(unittest.TestCase): while True: try: img_val, = exe.run(fetch_list=[img]) - except fluid.core.EnforceNotMet as ex: - self.assertIn("There is no next data.", ex.message) + except fluid.core.EOFException: break batch_count += 1 self.assertLessEqual(img_val.shape[0], self.batch_size) self.assertEqual(batch_count, self.num_batch * 3) def test_main(self): - self.main(thread_num=3) # thread number equals to file number - self.main(thread_num=10) # thread number is larger than file number - self.main(thread_num=2) # thread number is less than file number + self.main(is_test=False) + self.main(is_test=True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py b/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py index 52e7cc1ffbba40a63ce3cec645c7c0a7a499c1bf..7fc9f550440d3d0e1a8182a69f5692b3df0aa258 100644 --- a/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py +++ b/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py @@ -59,8 +59,7 @@ class TestMultipleReader(unittest.TestCase): while True: try: img_val, = exe.run(fetch_list=[img]) - except fluid.core.EnforceNotMet as ex: - self.assertIn("There is no next data.", ex.message) + except fluid.core.EOFException: break batch_count += 1 self.assertLessEqual(img_val.shape[0], self.batch_size) diff --git a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py index 6459913c0162374e17d0249627e7107a195babf8..10cb78a08db0471699bcc0b7323d5346e3af64c7 100644 --- a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py +++ b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py @@ -112,7 +112,7 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold, if keep_top_k > -1 and num_det > keep_top_k: score_index = [] - for c, indices in selected_indices.iteritems(): + for c, indices in selected_indices.items(): for idx in indices: score_index.append((scores[c][idx], c, idx)) @@ -135,15 +135,15 @@ def batched_multiclass_nms(boxes, scores, background, score_threshold, batch_size = scores.shape[0] det_outs = [] - lod = [0] + lod = [] for n in range(batch_size): nmsed_outs, nmsed_num = multiclass_nms(boxes[n], scores[n], background, score_threshold, nms_threshold, nms_top_k, keep_top_k) - lod.append(lod[-1] + nmsed_num) + lod.append(nmsed_num) if nmsed_num == 0: continue - for c, indices in nmsed_outs.iteritems(): + for c, indices in nmsed_outs.items(): for idx in indices: xmin, ymin, xmax, ymax = boxes[n][idx][:] det_outs.append([c, scores[n][c][idx], xmin, ymin, xmax, ymax]) diff --git a/python/paddle/fluid/tests/unittests/test_nce.py b/python/paddle/fluid/tests/unittests/test_nce.py index 76ecc8ba08ba31798040080a0ae99fe515c28cec..7431a142c53a64e58872390776904ce8f781d6a9 100644 --- a/python/paddle/fluid/tests/unittests/test_nce.py +++ b/python/paddle/fluid/tests/unittests/test_nce.py @@ -66,7 +66,7 @@ class TestNCE(OpTest): self.attrs = { 'num_total_classes': num_classes, 'num_neg_samples': num_neg_samples, - 'custom_neg_classes': range(num_neg_samples) + 'custom_neg_classes': list(range(num_neg_samples)) } self.inputs = { 'Input': input, diff --git a/python/paddle/fluid/tests/unittests/test_norm_op.py b/python/paddle/fluid/tests/unittests/test_norm_op.py index 6feda175fb537db894ac7f19e22297f6062a4d61..108a665f37f5cd652ec83f784a56ca52e6b49fe8 100644 --- a/python/paddle/fluid/tests/unittests/test_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_norm_op.py @@ -17,44 +17,23 @@ import numpy as np from op_test import OpTest -def norm(input, scale, epsilon): - s0, s1, s2, s3 = input.shape - x_square = input * input - for i in xrange(s0): - input_batch = input[i:i + 1, :, :, :] - input_batch = input_batch.reshape(s1, s2 * s3) - x_square_batch = x_square[i:i + 1, :, :, :] - x_square_batch = x_square_batch.reshape(s1, s2 * s3) - square_colsum = x_square_batch.sum(axis=0) + epsilon - tmp = pow(square_colsum, 0.5) - tmp = np.reciprocal(tmp) - tmp_tile = np.tile(tmp, s1) - tmp_tile = tmp_tile.reshape(s1, s2 * s3) - scale_tile = np.tile(scale, (1, s2 * s3)) - scale_tile = scale_tile.reshape(s1, s2 * s3) - out_batch = input_batch * tmp_tile * scale_tile - out_batch = out_batch.reshape(1, s1, s2, s3) - if i == 0: - out = out_batch - else: - out = np.concatenate((out, out_batch), 0) - out.reshape(s0, s1, s2, s3) - return out +def l2_norm(x, axis, epsilon): + x2 = x**2 + s = np.sum(x2, axis=axis, keepdims=True) + r = np.sqrt(s + epsilon) + y = x / np.broadcast_to(r, x.shape) + return y, r class TestNormOp(OpTest): def setUp(self): self.op_type = "norm" self.init_test_case() - input = np.random.random(self.shape).astype("float32") - scale = np.array([10, 10, 10]) - self.inputs = { - 'X': input.astype('float32'), - 'Scale': scale.astype('float32') - } - self.attrs = {'epsilon': self.epsilon} - output = norm(input, scale, self.epsilon) - self.outputs = {'Out': output.astype('float32')} + x = np.random.random(self.shape).astype("float64") + y, norm = l2_norm(x, self.axis, self.epsilon) + self.inputs = {'X': x} + self.attrs = {'epsilon': self.epsilon, 'axis': self.axis} + self.outputs = {'Out': y, 'Norm': norm} def test_check_output(self): self.check_output() @@ -63,8 +42,23 @@ class TestNormOp(OpTest): self.check_grad(['X'], 'Out') def init_test_case(self): - self.shape = [2, 3, 2, 2] - self.epsilon = 1e-6 + self.shape = [2, 3, 4, 4] + self.axis = 1 + self.epsilon = 1e-8 + + +class TestNormOp2(TestNormOp): + def init_test_case(self): + self.shape = [5, 3, 9, 7] + self.axis = 0 + self.epsilon = 1e-8 + + +class TestNormOp3(TestNormOp): + def init_test_case(self): + self.shape = [5, 3, 2, 7] + self.axis = -1 + self.epsilon = 1e-8 if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py b/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py index ef34893943d8f6bf91b1eb14378e463c178de84d..198c68866d399023c51c2a43b588aa8ec49c3c9a 100644 --- a/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py +++ b/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py @@ -70,8 +70,9 @@ class TestNormalization(unittest.TestCase): def l2_normalize(self, data, axis, epsilon): """ Compute the groundtruth. """ - output = data * np.reciprocal( - np.sum(np.square(data), axis=axis, keepdims=True)) + output = data / np.broadcast_to( + np.sqrt(np.sum(np.square(data), axis=axis, keepdims=True)), + data.shape) return output def test_l2_normalize(self): diff --git a/python/paddle/fluid/tests/unittests/test_one_hot_op.py b/python/paddle/fluid/tests/unittests/test_one_hot_op.py index cd78cce8729ab2b5a0bb4817cf3022e53932283a..06fccd39ac65ab62ee5618ac19d1a0535b481d06 100644 --- a/python/paddle/fluid/tests/unittests/test_one_hot_op.py +++ b/python/paddle/fluid/tests/unittests/test_one_hot_op.py @@ -27,14 +27,14 @@ class TestOneHotOp(OpTest): self.op_type = 'one_hot' depth = 10 dimension = 12 - x_lod = [[0, 4, 5, 8, 11]] - x = [np.random.randint(0, depth - 1) for i in xrange(x_lod[0][-1])] - x = np.array(x).astype('int').reshape([x_lod[0][-1], 1]) + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int').reshape([sum(x_lod[0]), 1]) out = np.zeros(shape=(np.product(x.shape[:-1]), depth)).astype('float32') - for i in xrange(np.product(x.shape)): + for i in range(np.product(x.shape)): out[i, x[i]] = 1.0 self.inputs = {'X': (x, x_lod)} @@ -50,14 +50,14 @@ class TestOneHotOp_default_dtype(OpTest): self.op_type = 'one_hot' depth = 10 dimension = 12 - x_lod = [[0, 4, 5, 8, 11]] - x = [np.random.randint(0, depth - 1) for i in xrange(x_lod[0][-1])] - x = np.array(x).astype('int').reshape([x_lod[0][-1], 1]) + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int').reshape([sum(x_lod[0]), 1]) out = np.zeros(shape=(np.product(x.shape[:-1]), depth)).astype('float32') - for i in xrange(np.product(x.shape)): + for i in range(np.product(x.shape)): out[i, x[i]] = 1.0 self.inputs = {'X': (x, x_lod)} @@ -75,11 +75,11 @@ class TestOneHotOp_exception(OpTest): self.place = core.CPUPlace() self.dimension = 12 self.x = core.LoDTensor() - x_lod = [[0, 4, 5, 8, 11]] - data = [np.random.randint(11, 20) for i in xrange(x_lod[0][-1])] - data = np.array(data).astype('int').reshape([x_lod[0][-1], 1]) + x_lod = [[4, 1, 3, 3]] + data = [np.random.randint(11, 20) for i in range(sum(x_lod[0]))] + data = np.array(data).astype('int').reshape([sum(x_lod[0]), 1]) self.x.set(data, self.place) - self.x.set_lod(x_lod) + self.x.set_recursive_sequence_lengths(x_lod) def test_check_output(self): program = Program() diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py index e775db1d10f4561b6fb90631757a25c9f74cb777..18921d727f94a85b69259c07273f09c3e19390c6 100644 --- a/python/paddle/fluid/tests/unittests/test_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_optimizer.py @@ -97,7 +97,7 @@ class TestMomentumOptimizer(unittest.TestCase): params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(momentum_optimizer.get_accumulators()), 0) - opts = momentum_optimizer.create_optimization_pass( + opts = momentum_optimizer._create_optimization_pass( params_grads, mul_out, init_program) self.assertEqual(len(opts), 3) sgd_op = opts[-1] @@ -151,7 +151,7 @@ class TestMomentumOptimizer(unittest.TestCase): params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(momentum_optimizer.get_accumulators()), 0) - opts = momentum_optimizer.create_optimization_pass( + opts = momentum_optimizer._create_optimization_pass( params_grads, mul_out, init_program) self.assertEqual(len(opts), 3) sgd_op = opts[-1] @@ -214,8 +214,8 @@ class TestAdagradOptimizer(unittest.TestCase): params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0) - opts = adagrad_optimizer.create_optimization_pass(params_grads, mul_out, - init_program) + opts = adagrad_optimizer._create_optimization_pass( + params_grads, mul_out, init_program) self.assertEqual(len(opts), 3) self.assertEqual([op.type for op in opts], ["fill_constant", "elementwise_mul", "adagrad"]) @@ -278,8 +278,8 @@ class TestAdamOptimizer(unittest.TestCase): params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(adam_optimizer.get_accumulators()), 0) - opts = adam_optimizer.create_optimization_pass(params_grads, mul_out, - init_program) + opts = adam_optimizer._create_optimization_pass(params_grads, mul_out, + init_program) self.assertEqual(len(opts), 5) self.assertEqual( [op.type for op in opts], @@ -287,7 +287,7 @@ class TestAdamOptimizer(unittest.TestCase): # Check accumulators accumulators = adam_optimizer.get_accumulators() - self.assertEqual(len(accumulators), 2) + self.assertEqual(len(accumulators), 4) self.assertTrue(adam_optimizer.get_moment1_str() in accumulators) self.assertTrue(adam_optimizer.get_moment2_str() in accumulators) moment1_acc = accumulators[adam_optimizer.get_moment1_str()] @@ -345,8 +345,8 @@ class TestAdamaxOptimizer(unittest.TestCase): params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(adamax_optimizer.get_accumulators()), 0) - opts = adamax_optimizer.create_optimization_pass(params_grads, mul_out, - init_program) + opts = adamax_optimizer._create_optimization_pass(params_grads, mul_out, + init_program) self.assertEqual(len(opts), 4) self.assertEqual( [op.type for op in opts], @@ -354,7 +354,7 @@ class TestAdamaxOptimizer(unittest.TestCase): # Check accumulators accumulators = adamax_optimizer.get_accumulators() - self.assertEqual(len(accumulators), 2) + self.assertEqual(len(accumulators), 3) self.assertTrue(adamax_optimizer.get_moment_str() in accumulators) self.assertTrue(adamax_optimizer.get_inf_norm_str() in accumulators) moment_acc = accumulators[adamax_optimizer.get_moment_str()] @@ -409,7 +409,7 @@ class TestDecayedAdagradOptimizer(unittest.TestCase): params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0) - opts = decayed_adagrad_optimizer.create_optimization_pass( + opts = decayed_adagrad_optimizer._create_optimization_pass( params_grads, mul_out, init_program) self.assertEqual(len(opts), 3) self.assertEqual( @@ -434,5 +434,71 @@ class TestDecayedAdagradOptimizer(unittest.TestCase): self.assertAlmostEqual(init_ops[1].attr('value'), 0.0) +class TestFtrlOptimizer(unittest.TestCase): + class MockFtrl(optimizer.FtrlOptimizer): + def get_accumulators(self): + return self._accumulators + + def get_squared_str(self): + return self._squared_acc_str + + def get_linear_str(self): + return self._linear_acc_str + + def test_ftrl_optimizer(self): + init_program = framework.Program() + program = framework.Program() + block = program.global_block() + mul_x = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) + mul_y = block.create_var( + dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") + mul_out = block.create_var( + dtype="float32", shape=[5, 8], lod_level=0, name="mul.out") + block.append_op( + type="mul", + inputs={"X": mul_x, + "Y": mul_y}, + outputs={"Out": mul_out}, + attrs={"x_num_col_dims": 1}) + mean_out = block.create_var( + dtype="float32", shape=[1], lod_level=0, name="mean.out") + block.append_op( + type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) + learning_rate = 0.01 + ftrl_optimizer = self.MockFtrl( + learning_rate=learning_rate, l1=0.0, l2=0.0, lr_power=-0.5) + params_grads = append_backward(mean_out) + self.assertEqual(len(params_grads), 1) + self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0) + opts = ftrl_optimizer._create_optimization_pass(params_grads, mul_out, + init_program) + self.assertEqual(len(opts), 3) + self.assertEqual([op.type for op in opts], + ["fill_constant", "elementwise_mul", "ftrl"]) + + # Check accumulators + accumulators = ftrl_optimizer.get_accumulators() + self.assertEqual(len(accumulators), 2) + self.assertTrue(ftrl_optimizer.get_squared_str() in accumulators) + self.assertTrue(ftrl_optimizer.get_linear_str() in accumulators) + squared_acc = accumulators[ftrl_optimizer.get_squared_str()] + linear_acc = accumulators[ftrl_optimizer.get_linear_str()] + self.assertEqual(len(squared_acc), 1) + self.assertEqual(len(linear_acc), 1) + self.assertTrue(mul_x.name in squared_acc) + self.assertTrue(mul_x.name in linear_acc) + + # Check init_program + init_ops = init_program.global_block().ops + self.assertEqual(len(init_ops), 3) + self.assertEqual(init_ops[0].type, "fill_constant") + self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py index 66e138b03f3b170aca4fb2207438eb9af1783c33..d17e493c36a2ffcba632f5f85c7a1d2e5066dd1c 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py @@ -17,6 +17,7 @@ import paddle.fluid as fluid import unittest import paddle import numpy as np +import os word_dict, verb_dict, label_dict = conll05.get_dict() word_dict_len = len(word_dict) @@ -101,7 +102,11 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, class TestCRFModel(unittest.TestCase): - def check_network_convergence(self, is_sparse, build_strategy=None): + def check_network_convergence(self, + is_sparse, + build_strategy=None, + use_cuda=True): + os.environ['CPU_NUM'] = str(4) main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): @@ -145,12 +150,12 @@ class TestCRFModel(unittest.TestCase): paddle.dataset.conll05.test(), buf_size=8192), batch_size=16) - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup) pe = fluid.ParallelExecutor( - use_cuda=True, + use_cuda=use_cuda, loss_name=avg_cost.name, build_strategy=build_strategy) @@ -162,35 +167,46 @@ class TestCRFModel(unittest.TestCase): place=fluid.CPUPlace()) data = train_data() - for i in xrange(10): + for i in range(10): cur_batch = next(data) - print map(np.array, - pe.run(feed=feeder.feed(cur_batch), - fetch_list=[avg_cost.name]))[0] + print(pe.run(feed=feeder.feed(cur_batch), + fetch_list=[avg_cost.name])[0]) + @unittest.skip(reason="CI hangs") def test_update_sparse_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce self.check_network_convergence( - is_sparse=True, build_strategy=build_strategy) + is_sparse=True, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy, use_cuda=False) + @unittest.skip(reason="CI hangs") def test_update_dense_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce self.check_network_convergence( - is_sparse=False, build_strategy=build_strategy) + is_sparse=False, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy, use_cuda=False) + @unittest.skip(reason="CI hangs") def test_update_sparse_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce self.check_network_convergence( - is_sparse=True, build_strategy=build_strategy) + is_sparse=True, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy, use_cuda=False) + @unittest.skip(reason="CI hangs") def test_update_dense_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce self.check_network_convergence( - is_sparse=False, build_strategy=build_strategy) + is_sparse=False, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy, use_cuda=False) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py index 24f8d28c0304a77a99213374b25d0db728eca265..a43f2e7c49c02ce779344da44e640cabbf27986c 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py @@ -15,9 +15,11 @@ import paddle.dataset.flowers as flowers import math import paddle.fluid as fluid +import paddle.fluid.core as core import unittest import numpy as np import paddle +import os def Lenet(data, class_dim): @@ -35,7 +37,7 @@ def Lenet(data, class_dim): class TestFetchOp(unittest.TestCase): - def parallel_exe(self, train_inputs, seed): + def parallel_exe(self, train_inputs, seed, use_cuda): main = fluid.Program() startup = fluid.Program() startup.random_seed = seed @@ -59,22 +61,24 @@ class TestFetchOp(unittest.TestCase): # conv2d_1.b_0@GRAD. Those variables should not be pruned. # fluid.memory_optimize(main) - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup) feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) pe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name, main_program=main) + use_cuda=use_cuda, loss_name=loss.name, main_program=main) fetch_list = [] all_vars = main.global_block().vars - for k, v in all_vars.iteritems(): + for k, v in all_vars.items(): if 'tmp' not in k and k[0] is not '_' or v.persistable: fetch_list.append(k) for data in train_inputs: - ret = pe.run(fetch_list, feed=feeder.feed(data)) + ret = pe.run(fetch_list, + feed=feeder.feed(data), + return_numpy=True) for i in range(len(fetch_list)): assert not math.isnan(np.sum(ret[i])) and \ not math.isinf(np.sum(ret[i])) @@ -86,16 +90,19 @@ class TestFetchOp(unittest.TestCase): iters = 3 train_inputs = [] for i in range(iters): - train_inputs.append(tst_reader_iter.next()) + train_inputs.append(next(tst_reader_iter)) - self.parallel_exe(train_inputs, seed=1) + os.environ['CPU_NUM'] = str(4) + if core.is_compiled_with_cuda(): + self.parallel_exe(train_inputs, seed=1, use_cuda=True) + self.parallel_exe(train_inputs, seed=1, use_cuda=False) class TestFeedParallel(unittest.TestCase): - def test_main(self): + def parallel_exe(self, use_cuda, seed): main = fluid.Program() startup = fluid.Program() - startup.random_seed = 1 + startup.random_seed = seed with fluid.scope_guard(fluid.core.Scope()): with fluid.program_guard(main, startup): data = fluid.layers.data( @@ -111,22 +118,31 @@ class TestFeedParallel(unittest.TestCase): regularization=fluid.regularizer.L2Decay(1e-4)) opt.minimize(loss) - place = fluid.CUDAPlace(0) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) reader = feeder.decorate_reader( paddle.batch( flowers.train(), batch_size=16), multi_devices=True) + exe = fluid.Executor(place) exe.run(startup) + pe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name, main_program=main) + use_cuda=use_cuda, loss_name=loss.name, main_program=main) for batch_id, data in enumerate(reader()): - loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0]) - print batch_id, loss_np + loss_np = pe.run(feed=data, fetch_list=[loss.name])[0] + print(batch_id, loss_np) if batch_id == 2: break + def test_feed_op(self): + os.environ['CPU_NUM'] = str(4) + if core.is_compiled_with_cuda(): + self.parallel_exe(use_cuda=True, seed=1) + self.parallel_exe(use_cuda=False, seed=1) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py index 015703c3e25f4e11e64ab6a7de99da12bee608f6..9448d89cd58f4e5cff4bac49821fbc44c5a46246 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py @@ -14,10 +14,12 @@ from parallel_executor_test_base import TestParallelExecutorBase import paddle.fluid as fluid +import paddle.fluid.core as core import numpy as np import paddle import paddle.dataset.mnist as mnist import unittest +import os MNIST_RECORDIO_FILE = "./mnist_test_pe.recordio" @@ -31,13 +33,11 @@ def simple_fc_net(use_feed): filenames=[MNIST_RECORDIO_FILE], shapes=[[-1, 784], [-1, 1]], lod_levels=[0, 0], - dtypes=['float32', 'int64'], - thread_num=1, - for_parallel=True) + dtypes=['float32', 'int64']) reader = fluid.layers.io.double_buffer(reader) img, label = fluid.layers.read_file(reader) hidden = img - for _ in xrange(4): + for _ in range(4): hidden = fluid.layers.fc( hidden, size=200, @@ -59,14 +59,12 @@ def fc_with_batchnorm(use_feed): filenames=[MNIST_RECORDIO_FILE], shapes=[[-1, 784], [-1, 1]], lod_levels=[0, 0], - dtypes=['float32', 'int64'], - thread_num=1, - for_parallel=True) + dtypes=['float32', 'int64']) reader = fluid.layers.io.double_buffer(reader) img, label = fluid.layers.read_file(reader) hidden = img - for _ in xrange(1): + for _ in range(1): hidden = fluid.layers.fc( hidden, size=200, @@ -85,6 +83,7 @@ def fc_with_batchnorm(use_feed): class TestMNIST(TestParallelExecutorBase): @classmethod def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) # Convert mnist to recordio file with fluid.program_guard(fluid.Program(), fluid.Program()): reader = paddle.batch(mnist.train(), batch_size=4) @@ -99,72 +98,119 @@ class TestMNIST(TestParallelExecutorBase): fluid.recordio_writer.convert_reader_to_recordio_file( MNIST_RECORDIO_FILE, reader, feeder) - def check_simple_fc_convergence(self, balance_parameter_opt_between_cards): - self.check_network_convergence(simple_fc_net) - self.check_network_convergence(simple_fc_net, allow_op_delay=True) - - img = np.zeros(shape=[32, 784], dtype='float32') + def _init_data(self): + np.random.seed(5) + img = np.random.random(size=[32, 784]).astype(np.float32) label = np.ones(shape=[32, 1], dtype='int64') + return img, label + + def _compare_reduce_and_allreduce(self, model, use_cuda): + if use_cuda and not core.is_compiled_with_cuda(): + return + self.check_network_convergence( + model, use_cuda=use_cuda, use_reduce=True) + self.check_network_convergence( + model, use_cuda=use_cuda, allow_op_delay=True, use_reduce=True) + + img, label = self._init_data() + + all_reduce_first_loss, all_reduce_last_loss = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_reduce=False) + reduce_first_loss, reduce_last_loss = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_reduce=True) + + for loss in zip(all_reduce_first_loss, reduce_first_loss): + self.assertAlmostEqual(loss[0], loss[1], delta=1e-6) + for loss in zip(all_reduce_last_loss, reduce_last_loss): + self.assertAlmostEqual(loss[0], loss[1], delta=1e-4) + + # simple_fc + def check_simple_fc_convergence(self, use_cuda, use_reduce=False): + if use_cuda and not core.is_compiled_with_cuda(): + return + self.check_network_convergence(simple_fc_net, use_cuda=use_cuda) + self.check_network_convergence( + simple_fc_net, use_cuda=use_cuda, allow_op_delay=True) + + img, label = self._init_data() + self.check_network_convergence( simple_fc_net, feed_dict={"image": img, "label": label}, - balance_parameter_opt_between_cards=balance_parameter_opt_between_cards - ) + use_cuda=use_cuda, + use_reduce=use_reduce) def test_simple_fc(self): + # use_cuda + self.check_simple_fc_convergence(True) self.check_simple_fc_convergence(False) def test_simple_fc_with_new_strategy(self): - self.check_simple_fc_convergence(True) + # use_cuda, use_reduce + self._compare_reduce_and_allreduce(simple_fc_net, True) + self._compare_reduce_and_allreduce(simple_fc_net, False) + + def check_simple_fc_parallel_accuracy(self, use_cuda): + if use_cuda and not core.is_compiled_with_cuda(): + return + + img, label = self._init_data() - def check_simple_fc_parallel_accuracy(self, - balance_parameter_opt_between_cards): - img = np.zeros(shape=[32, 784], dtype='float32') - label = np.ones(shape=[32, 1], dtype='int64') single_first_loss, single_last_loss = self.check_network_convergence( method=simple_fc_net, - seed=1000, + seed=1, feed_dict={"image": img, "label": label}, + use_cuda=use_cuda, use_parallel_executor=False) parallel_first_loss, parallel_last_loss = self.check_network_convergence( method=simple_fc_net, - seed=1000, + seed=1, feed_dict={"image": img, "label": label}, - use_parallel_executor=True, - balance_parameter_opt_between_cards=balance_parameter_opt_between_cards - ) + use_cuda=use_cuda, + use_parallel_executor=True) - for p_f in parallel_first_loss: - self.assertAlmostEquals(p_f, single_first_loss[0], delta=1e-6) - for p_l in parallel_last_loss: - self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6) + self.assertAlmostEquals( + np.mean(parallel_first_loss), single_first_loss, delta=1e-6) + self.assertAlmostEquals( + np.mean(parallel_last_loss), single_last_loss, delta=1e-6) def test_simple_fc_parallel_accuracy(self): + self.check_simple_fc_parallel_accuracy(True) self.check_simple_fc_parallel_accuracy(False) - def test_simple_fc_parallel_accuracy_with_new_strategy(self): - self.check_simple_fc_parallel_accuracy(True) + def check_batchnorm_fc_convergence(self, use_cuda): + if use_cuda and not core.is_compiled_with_cuda(): + return + + self.check_network_convergence(fc_with_batchnorm, use_cuda=use_cuda) + + img, label = self._init_data() - def check_batchnorm_fc_convergence(self, - balance_parameter_opt_between_cards): - self.check_network_convergence(fc_with_batchnorm) - img = np.zeros(shape=[32, 784], dtype='float32') - label = np.ones(shape=[32, 1], dtype='int64') self.check_network_convergence( fc_with_batchnorm, feed_dict={"image": img, "label": label}, - balance_parameter_opt_between_cards=balance_parameter_opt_between_cards - ) + use_cuda=use_cuda) def test_batchnorm_fc(self): + self.check_batchnorm_fc_convergence(True) self.check_batchnorm_fc_convergence(False) def test_batchnorm_fc_with_new_strategy(self): - self.check_batchnorm_fc_convergence(True) + # FIXME(zcd): close this test temporally. + # self._compare_reduce_and_allreduce(fc_with_batchnorm, True) + self._compare_reduce_and_allreduce(fc_with_batchnorm, False) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py index a3fa140cbb7994a36d2cbee26d598165f1f771d2..a28428d8dee201ba105e18684c15d4b4582d989f 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py @@ -13,8 +13,27 @@ # limitations under the License. import paddle.fluid as fluid +import paddle.fluid.layers.ops as ops +from paddle.fluid.initializer import init_on_cpu +from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter +import paddle.fluid.core as core from parallel_executor_test_base import TestParallelExecutorBase import unittest +import math +import os +import numpy as np + +# FIXME(zcd): If the neural net has dropout_op, the output of ParallelExecutor +# and Executor is different. Because, for ParallelExecutor, the dropout_op of +# the neural net will be copied N copies(N is the number of device). This will +# lead to the random numbers generated by ParallelExecutor and Executor are different. +# So, if we compare the loss of ParallelExecutor and Executor, we should remove the +# dropout_op. +remove_dropout = False + +# FIXME(zcd): If the neural net has batch_norm, the output of ParallelExecutor +# and Executor is different. +remove_bn = False def squeeze_excitation(input, num_channels, reduction_ratio): @@ -47,7 +66,8 @@ def conv_bn_layer(input, num_filters, filter_size, stride=1, groups=1, groups=groups, act=None, bias_attr=False) - return fluid.layers.batch_norm(input=conv, act=act, momentum=0.1) + return conv if remove_bn else fluid.layers.batch_norm( + input=conv, act=act, momentum=0.1) def shortcut(input, ch_out, stride): @@ -86,13 +106,14 @@ def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio): return fluid.layers.elementwise_add(x=short, y=scale, act='relu') -def SE_ResNeXt50Small(batch_size=2, use_feed=False): - assert not use_feed, "SE_ResNeXt doesn't support feed yet" +batch_size = 12 +img_shape = [3, 224, 224] + - img = fluid.layers.fill_constant( - shape=[batch_size, 3, 224, 224], dtype='float32', value=0.0) - label = fluid.layers.fill_constant( - shape=[batch_size, 1], dtype='int64', value=0.0) +def SE_ResNeXt50Small(use_feed): + + img = fluid.layers.data(name='image', shape=img_shape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') conv = conv_bn_layer( input=img, num_filters=16, filter_size=3, stride=2, act='relu') @@ -121,7 +142,8 @@ def SE_ResNeXt50Small(batch_size=2, use_feed=False): reshape = fluid.layers.reshape( x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) pool = fluid.layers.reduce_mean(input=reshape, dim=2) - dropout = fluid.layers.dropout(x=pool, dropout_prob=0.2) + dropout = pool if remove_dropout else fluid.layers.dropout( + x=pool, dropout_prob=0.2, seed=1) # Classifier layer: prediction = fluid.layers.fc(input=dropout, size=1000, act='softmax') loss = fluid.layers.cross_entropy(input=prediction, label=label) @@ -129,23 +151,135 @@ def SE_ResNeXt50Small(batch_size=2, use_feed=False): return loss +def cosine_decay(learning_rate, step_each_epoch, epochs=120): + """ + Applies cosine decay to the learning rate. + lr = 0.05 * (math.cos(epoch * (math.pi / 120)) + 1) + """ + global_step = _decay_step_counter() + + with init_on_cpu(): + epoch = ops.floor(global_step / step_each_epoch) + decayed_lr = learning_rate * \ + (ops.cos(epoch * (math.pi / epochs)) + 1)/2 + return decayed_lr + + +def optimizer(learning_rate=0.01): + optimizer = fluid.optimizer.Momentum( + learning_rate=cosine_decay( + learning_rate=learning_rate, step_each_epoch=2, epochs=1), + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + return optimizer + + class TestResnet(TestParallelExecutorBase): - def check_resnet_convergence(self, balance_parameter_opt_between_cards): - import functools - batch_size = 2 - self.check_network_convergence( - functools.partial( - SE_ResNeXt50Small, batch_size=batch_size), - iter=20, + @classmethod + def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) + global remove_dropout + global remove_bn + remove_dropout = False + remove_bn = False + + def _init_data(self, batch_size=2, random=True): + np.random.seed(5) + if random: + img = np.random.random( + size=[batch_size] + img_shape).astype(np.float32) + else: + img = np.ones(shape=[batch_size] + img_shape, dtype='float32') + label = [np.random.randint(0, 999) for _ in range(batch_size)] + label = np.array(label).astype(np.int64).reshape(-1, 1) + return img, label + + def _compare_reduce_and_allreduce(self, + model, + use_cuda, + iter=20, + delta2=1e-6): + if use_cuda and not core.is_compiled_with_cuda(): + return + + global remove_bn + remove_bn = True + + img, label = self._init_data(batch_size=batch_size) + all_reduce_first_loss, all_reduce_last_loss = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + iter=iter, batch_size=batch_size, - balance_parameter_opt_between_cards=balance_parameter_opt_between_cards - ) + use_cuda=use_cuda, + use_reduce=False, + optimizer=optimizer) + reduce_first_loss, reduce_last_loss = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + iter=iter, + batch_size=batch_size, + use_cuda=use_cuda, + use_reduce=True, + optimizer=optimizer) + + for loss in zip(all_reduce_first_loss, reduce_first_loss): + self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + for loss in zip(all_reduce_last_loss, reduce_last_loss): + self.assertAlmostEquals(loss[0], loss[1], delta=delta2) + + def _check_resnet_convergence(self, + model, + use_cuda=True, + use_reduce=False, + iter=20, + delta2=1e-6): + if use_cuda and not core.is_compiled_with_cuda(): + return + + global remove_dropout + global remove_bn + remove_dropout = True + remove_bn = True + + img, label = self._init_data(batch_size=batch_size) + single_first_loss, single_last_loss = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + iter=iter, + batch_size=batch_size, + use_cuda=use_cuda, + use_reduce=use_reduce, + optimizer=optimizer, + use_parallel_executor=False) + parallel_first_loss, parallel_last_loss = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + iter=iter, + batch_size=batch_size, + use_cuda=use_cuda, + use_reduce=use_reduce, + optimizer=optimizer) + + self.assertAlmostEquals( + np.mean(parallel_first_loss), single_first_loss[0], delta=1e-6) + self.assertAlmostEquals( + np.mean(parallel_last_loss), single_last_loss[0], delta=delta2) - def test_resnet(self): - self.check_resnet_convergence(False) + def test_seresnext_with_learning_rate_decay(self): + self._check_resnet_convergence(model=SE_ResNeXt50Small, use_cuda=True) + self._check_resnet_convergence( + model=SE_ResNeXt50Small, use_cuda=False, iter=2, delta2=1e-3) - def test_resnet_with_new_strategy(self): - self.check_resnet_convergence(True) + def test_seresnext_with_new_strategy(self): + self._compare_reduce_and_allreduce( + model=SE_ResNeXt50Small, use_cuda=True, delta2=1e-2) + self._compare_reduce_and_allreduce( + model=SE_ResNeXt50Small, use_cuda=False, iter=5) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py index 93a5f767867d68110cf7b8f441cc740ecd843cf9..fcb5947ff05efd1c48ab9ec129ac9d17255d7020 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py @@ -13,15 +13,19 @@ # limitations under the License. import paddle.fluid as fluid +import paddle.fluid.core as core import numpy as np import unittest +import os +import sys +import math def simple_fc_net(): img = fluid.layers.data(name='image', shape=[784], dtype='float32') label = fluid.layers.data(name='label', shape=[1], dtype='int64') hidden = img - for _ in xrange(4): + for _ in range(4): hidden = fluid.layers.fc( hidden, size=200, @@ -35,7 +39,8 @@ def simple_fc_net(): class ParallelExecutorTestingDuringTraining(unittest.TestCase): - def check_network_convergence(self, build_strategy=None): + def check_network_convergence(self, use_cuda, build_strategy=None): + os.environ['CPU_NUM'] = str(4) main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): @@ -49,29 +54,36 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): image = np.random.normal(size=(batch_size, 784)).astype('float32') label = np.random.randint(0, 10, (batch_size, 1), dtype="int64") - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup) feed_dict = {'image': image, 'label': label} train_exe = fluid.ParallelExecutor( - use_cuda=True, + use_cuda=use_cuda, loss_name=loss.name, main_program=main, build_strategy=build_strategy) test_exe = fluid.ParallelExecutor( - use_cuda=True, + use_cuda=use_cuda, main_program=test_program, share_vars_from=train_exe, build_strategy=build_strategy) - for i in xrange(5): + for i in range(5): test_loss, = test_exe.run([loss.name], feed=feed_dict) - test_loss = np.array(test_loss) train_loss, = train_exe.run([loss.name], feed=feed_dict) - train_loss = np.array(train_loss) + + avg_test_loss_val = np.array(test_loss).mean() + if math.isnan(float(avg_test_loss_val)): + sys.exit("got NaN loss, testing failed.") + + avg_train_loss_val = np.array(train_loss).mean() + if math.isnan(float(avg_train_loss_val)): + sys.exit("got NaN loss, training failed.") + self.assertTrue( np.allclose( train_loss, test_loss, atol=1e-8), @@ -81,12 +93,20 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): def test_parallel_testing(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce - self.check_network_convergence(build_strategy) + if core.is_compiled_with_cuda(): + self.check_network_convergence( + use_cuda=True, build_strategy=build_strategy) + self.check_network_convergence( + use_cuda=False, build_strategy=build_strategy) def test_parallel_testing_with_new_strategy(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce - self.check_network_convergence(build_strategy) + if core.is_compiled_with_cuda(): + self.check_network_convergence( + use_cuda=True, build_strategy=build_strategy) + self.check_network_convergence( + use_cuda=False, build_strategy=build_strategy) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py index c81df66d987f3d3856af0e19fc935df7de2edacc..8203d5d1fce0950130ab71db40fb306f73c41bd4 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py @@ -19,8 +19,9 @@ from parallel_executor_test_base import TestParallelExecutorBase import unittest import paddle import paddle.dataset.wmt16 as wmt16 +import os -WMT16_RECORDIO_FILE = "./wmt16_test_pe.recordio" +WMT16_RECORDIO_FILE = "/tmp/wmt16.recordio" class ModelHyperParams(object): @@ -149,6 +150,7 @@ def transformer(use_feed): class TestTransformer(TestParallelExecutorBase): @classmethod def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) reader = paddle.batch( wmt16.train(ModelHyperParams.src_vocab_size, ModelHyperParams.trg_vocab_size), @@ -165,9 +167,9 @@ class TestTransformer(TestParallelExecutorBase): writer.append_tensor(t) writer.complete_append_tensor() - @unittest.skip("transformer is buggy in multi gpu") def test_main(self): - self.check_network_convergence(transformer) + self.check_network_convergence(transformer, use_cuda=True) + self.check_network_convergence(transformer, use_cuda=False, iter=5) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_op.py b/python/paddle/fluid/tests/unittests/test_parallel_op.py index 79bea148f9398152a02d70946cdc5fff1f47ba6b..c9617e36778740ce9620c3ad495c64c17277fde1 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_op.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_op.py @@ -15,8 +15,10 @@ import unittest import paddle.fluid as fluid +from paddle.fluid.layers.device import get_places import paddle.fluid.profiler as profiler import numpy +import six class BaseParallelForTest(unittest.TestCase): @@ -24,20 +26,20 @@ class BaseParallelForTest(unittest.TestCase): """ Run the unittest for parallel.for Args: - callback(callable): A callable function returns a generator. There - are two yields in the generator function. The first yield - returns the data layers, and the second yield returns the loss. - The modified data variables will be sent back during the first + callback(callable): A callable function returns a generator. There + are two yields in the generator function. The first yield + returns the data layers, and the second yield returns the loss. + The modified data variables will be sent back during the first yield. feed(dict): The executor feeding dictionary. - fetch(list|basestr): The fetch name lists. + fetch(list|basestr): The fetch name lists. Returns: None Raises: - AssertionError when the computation of cpu, parallel.for in cpu, + AssertionError when the computation of cpu, parallel.for in cpu, gpu, parallel.for in gpu are different. """ @@ -94,14 +96,14 @@ class BaseParallelForTest(unittest.TestCase): """ Run a single test, returns the fetch values Args: - place(Place): the computation place. - use_parallel(bool): Whether use parallel.for or not. + place(Place): the computation place. + use_parallel(bool): Whether use parallel.for or not. Returns: Fetched numpy arrays. """ - if isinstance(fetch, basestring): + if isinstance(fetch, six.string_types): fetch = [fetch] main = fluid.Program() startup = fluid.Program() @@ -113,15 +115,17 @@ class BaseParallelForTest(unittest.TestCase): generator = callback() # Automatically insert parallel do if use_parallel = True if use_parallel: - places = fluid.layers.get_places() + thread_num = fluid.core.get_cuda_device_count( + ) if use_gpu else 8 + places = get_places(thread_num) pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl) data = next(generator) - if isinstance(data, fluid.Variable): + if isinstance(data, fluid.framework.Variable): data = [data] with pd.do(): - ins = map(pd.read_input, data) + ins = list(map(pd.read_input, data)) if len(ins) == 1: ins = ins[0] loss = generator.send(ins) # patch input @@ -153,7 +157,7 @@ class BaseParallelForTest(unittest.TestCase): Returns: None - + Raises: AssertionError diff --git a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py index 2105d320665367e3ec1bfd7b3a353a144c91244f..8aff4e87f67bc61a162f09e982cf0a7a61639257 100644 --- a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py +++ b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py @@ -23,9 +23,9 @@ def PolygonBoxRestore(input): geo_channels = shape[1] h = shape[2] w = shape[3] - h_indexes = np.array(range(h) * w).reshape( + h_indexes = np.array(list(range(h)) * w).reshape( [w, h]).transpose()[np.newaxis, :] # [1, h, w] - w_indexes = np.array(range(w) * h).reshape( + w_indexes = np.array(list(range(w)) * h).reshape( [h, w])[np.newaxis, :] # [1, h, w] indexes = np.concatenate( (w_indexes, h_indexes))[np.newaxis, :] # [1, 2, h, w] diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_op.py b/python/paddle/fluid/tests/unittests/test_pool2d_op.py index f7e1e8573290766cde0c35816d687e7ba6fa4220..1cf70311b40bc7648b7462e93f201aa33c77b137 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool2d_op.py @@ -35,8 +35,8 @@ def max_pool2D_forward_naive(x, ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) - for i in xrange(H_out): - for j in xrange(W_out): + for i in range(H_out): + for j in range(W_out): r_start = np.max((i * strides[0] - paddings[0], 0)) r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) c_start = np.max((j * strides[1] - paddings[1], 0)) @@ -63,8 +63,8 @@ def avg_pool2D_forward_naive(x, ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) - for i in xrange(H_out): - for j in xrange(W_out): + for i in range(H_out): + for j in range(W_out): r_start = np.max((i * strides[0] - paddings[0], 0)) r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) c_start = np.max((j * strides[1] - paddings[1], 0)) diff --git a/python/paddle/fluid/tests/unittests/test_pool3d_op.py b/python/paddle/fluid/tests/unittests/test_pool3d_op.py index 142165f29beeaedfaa660f04424147e06710d192..92c64b37921eafd4c90e247a235f2dacea8fea1e 100644 --- a/python/paddle/fluid/tests/unittests/test_pool3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool3d_op.py @@ -38,13 +38,13 @@ def max_pool3D_forward_naive(x, ) / strides[2] + 1 if ceil_mode else (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1 out = np.zeros((N, C, D_out, H_out, W_out)) - for k in xrange(D_out): + for k in range(D_out): d_start = np.max((k * strides[0] - paddings[0], 0)) d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D)) - for i in xrange(H_out): + for i in range(H_out): h_start = np.max((i * strides[0] - paddings[0], 0)) h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) - for j in xrange(W_out): + for j in range(W_out): w_start = np.max((j * strides[1] - paddings[1], 0)) w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] @@ -72,13 +72,13 @@ def avg_pool3D_forward_naive(x, ) / strides[2] + 1 if ceil_mode else (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1 out = np.zeros((N, C, D_out, H_out, W_out)) - for k in xrange(D_out): + for k in range(D_out): d_start = np.max((k * strides[0] - paddings[0], 0)) d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D)) - for i in xrange(H_out): + for i in range(H_out): h_start = np.max((i * strides[0] - paddings[0], 0)) h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) - for j in xrange(W_out): + for j in range(W_out): w_start = np.max((j * strides[1] - paddings[1], 0)) w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] diff --git a/python/paddle/fluid/tests/unittests/test_pool_max_op.py b/python/paddle/fluid/tests/unittests/test_pool_max_op.py index cf9b7639224ef3804b946f729bb6a9cead4aae23..e6a9f6f08cf1445c14494506641b0c3502591c37 100644 --- a/python/paddle/fluid/tests/unittests/test_pool_max_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool_max_op.py @@ -29,21 +29,21 @@ def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=False): W_out = (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1 out = np.zeros((N, C, D_out, H_out, W_out)) mask = np.zeros((N, C, D_out, H_out, W_out)) - for k in xrange(D_out): + for k in range(D_out): d_start = np.max((k * strides[0] - paddings[0], 0)) d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D)) - for i in xrange(H_out): + for i in range(H_out): h_start = np.max((i * strides[0] - paddings[0], 0)) h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) - for j in xrange(W_out): + for j in range(W_out): w_start = np.max((j * strides[1] - paddings[1], 0)) w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] out[:, :, k, i, j] = np.max(x_masked, axis=(2, 3, 4)) - for n in xrange(N): - for c in xrange(C): + for n in range(N): + for c in range(C): arr = x_masked[n, c, :, :, :] index = np.where(arr == np.max(arr)) sub_deep = index[0][0] @@ -67,8 +67,8 @@ def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=False): W_out = (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) mask = np.zeros((N, C, H_out, W_out)) - for i in xrange(H_out): - for j in xrange(W_out): + for i in range(H_out): + for j in range(W_out): r_start = np.max((i * strides[0] - paddings[0], 0)) r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H)) c_start = np.max((j * strides[1] - paddings[1], 0)) @@ -77,8 +77,8 @@ def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=False): out[:, :, i, j] = np.max(x_masked, axis=(2, 3)) - for n in xrange(N): - for c in xrange(C): + for n in range(N): + for c in range(C): arr = x_masked[n, c, :, :] index = np.where(arr == np.max(arr)) sub_row = index[0][0] diff --git a/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py b/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py index 091cfc9c72769fefc9c792bfeaa872cb357736b7..8c76393bdaccc0b701b409efebf08fac95aa5f1a 100644 --- a/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py +++ b/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py @@ -32,7 +32,7 @@ def py_pnpair_op(score, label, query, column=-1, weight=None): # accumulate statistics pos, neg, neu = 0, 0, 0 - for _, ranks in predictions.items(): + for _, ranks in list(predictions.items()): for e1, e2 in itertools.combinations(ranks, 2): s1, s2, l1, l2, w1, w2 = e1[0], e2[0], e1[1], e2[1], e1[2], e2[2] w = (w1 + w2) * 0.5 diff --git a/python/paddle/fluid/tests/unittests/test_precision_recall_op.py b/python/paddle/fluid/tests/unittests/test_precision_recall_op.py index 7830ba29583d369c4b9f6f3077dda1dda1fd1c46..5ae425fee18b9b1baa0b945782268b79d6bb6625 100644 --- a/python/paddle/fluid/tests/unittests/test_precision_recall_op.py +++ b/python/paddle/fluid/tests/unittests/test_precision_recall_op.py @@ -39,19 +39,19 @@ def get_states(idxs, labels, cls_num, weights=None): ins_num = idxs.shape[0] # TP FP TN FN states = np.zeros((cls_num, 4)).astype('float32') - for i in xrange(ins_num): + for i in range(ins_num): w = weights[i] if weights is not None else 1.0 idx = idxs[i][0] label = labels[i][0] if idx == label: states[idx][0] += w - for j in xrange(cls_num): + for j in range(cls_num): states[j][2] += w states[idx][2] -= w else: states[label][3] += w states[idx][1] += w - for j in xrange(cls_num): + for j in range(cls_num): states[j][2] += w states[label][2] -= w states[idx][2] -= w @@ -64,7 +64,7 @@ def compute_metrics(states, cls_num): total_fn_count = 0.0 macro_avg_precision = 0.0 macro_avg_recall = 0.0 - for i in xrange(cls_num): + for i in range(cls_num): total_tp_count += states[i][0] total_fp_count += states[i][1] total_fn_count += states[i][3] @@ -90,9 +90,9 @@ class TestPrecisionRecallOp_0(OpTest): ins_num = 64 cls_num = 10 max_probs = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32') - idxs = np.random.choice(xrange(cls_num), ins_num).reshape( + idxs = np.random.choice(range(cls_num), ins_num).reshape( (ins_num, 1)).astype('int32') - labels = np.random.choice(xrange(cls_num), ins_num).reshape( + labels = np.random.choice(range(cls_num), ins_num).reshape( (ins_num, 1)).astype('int32') states = get_states(idxs, labels, cls_num) metrics = compute_metrics(states, cls_num) @@ -117,10 +117,10 @@ class TestPrecisionRecallOp_1(OpTest): ins_num = 64 cls_num = 10 max_probs = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32') - idxs = np.random.choice(xrange(cls_num), ins_num).reshape( + idxs = np.random.choice(range(cls_num), ins_num).reshape( (ins_num, 1)).astype('int32') weights = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32') - labels = np.random.choice(xrange(cls_num), ins_num).reshape( + labels = np.random.choice(range(cls_num), ins_num).reshape( (ins_num, 1)).astype('int32') states = get_states(idxs, labels, cls_num, weights) @@ -151,10 +151,10 @@ class TestPrecisionRecallOp_2(OpTest): ins_num = 64 cls_num = 10 max_probs = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32') - idxs = np.random.choice(xrange(cls_num), ins_num).reshape( + idxs = np.random.choice(range(cls_num), ins_num).reshape( (ins_num, 1)).astype('int32') weights = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32') - labels = np.random.choice(xrange(cls_num), ins_num).reshape( + labels = np.random.choice(range(cls_num), ins_num).reshape( (ins_num, 1)).astype('int32') states = np.random.randint(0, 30, (cls_num, 4)).astype('float32') diff --git a/python/paddle/fluid/tests/unittests/test_print_op.py b/python/paddle/fluid/tests/unittests/test_print_op.py index c75080fbb96d472810e5d6a1d02a77c456006f66..b461c5c9401d74ef8dcf4afc84dc0ea6920a2419 100644 --- a/python/paddle/fluid/tests/unittests/test_print_op.py +++ b/python/paddle/fluid/tests/unittests/test_print_op.py @@ -28,7 +28,7 @@ class TestPrintOpCPU(unittest.TestCase): self.x_tensor = core.LoDTensor() tensor_np = np.random.random(size=(2, 3)).astype('float32') self.x_tensor.set(tensor_np, self.place) - self.x_tensor.set_lod([[0, 1, 1]]) + self.x_tensor.set_recursive_sequence_lengths([[1, 1]]) def build_network(self, only_forward, **kargs): x = layers.data('x', shape=[3], dtype='float32', lod_level=1) @@ -56,13 +56,15 @@ class TestPrintOpCPU(unittest.TestCase): return_numpy=False) +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestPrintOpGPU(TestPrintOpCPU): def setUp(self): self.place = core.CUDAPlace(0) self.x_tensor = core.LoDTensor() tensor_np = np.random.random(size=(2, 3)).astype('float32') self.x_tensor.set(tensor_np, self.place) - self.x_tensor.set_lod([[0, 1, 1]]) + self.x_tensor.set_recursive_sequence_lengths([[1, 1]]) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_prior_box_op.py b/python/paddle/fluid/tests/unittests/test_prior_box_op.py index bcbc02a2baa46b9ab583ecf3006bd3262e6038fd..e15554737b9f3fa36382dde15ded928271679538 100644 --- a/python/paddle/fluid/tests/unittests/test_prior_box_op.py +++ b/python/paddle/fluid/tests/unittests/test_prior_box_op.py @@ -32,6 +32,7 @@ class TestPriorBoxOp(OpTest): 'variances': self.variances, 'flip': self.flip, 'clip': self.clip, + 'min_max_aspect_ratios_order': self.min_max_aspect_ratios_order, 'step_w': self.step_w, 'step_h': self.step_h, 'offset': self.offset @@ -52,6 +53,9 @@ class TestPriorBoxOp(OpTest): max_sizes = [5, 10] self.max_sizes = np.array(max_sizes).astype('float32').tolist() + def set_min_max_aspect_ratios_order(self): + self.min_max_aspect_ratios_order = False + def init_test_params(self): self.layer_w = 32 self.layer_h = 32 @@ -71,6 +75,7 @@ class TestPriorBoxOp(OpTest): self.set_max_sizes() self.aspect_ratios = [2.0, 3.0] self.flip = True + self.set_min_max_aspect_ratios_order() self.real_aspect_ratios = [1, 2.0, 1.0 / 2.0, 3.0, 1.0 / 3.0] self.aspect_ratios = np.array( self.aspect_ratios, dtype=np.float).flatten() @@ -78,7 +83,6 @@ class TestPriorBoxOp(OpTest): self.variances = np.array(self.variances, dtype=np.float).flatten() self.clip = True - self.num_priors = len(self.real_aspect_ratios) * len(self.min_sizes) if len(self.max_sizes) > 0: self.num_priors += len(self.max_sizes) @@ -106,26 +110,60 @@ class TestPriorBoxOp(OpTest): idx = 0 for s in range(len(self.min_sizes)): min_size = self.min_sizes[s] - # rest of priors - for r in range(len(self.real_aspect_ratios)): - ar = self.real_aspect_ratios[r] - c_w = min_size * math.sqrt(ar) / 2 - c_h = (min_size / math.sqrt(ar)) / 2 - out_boxes[h, w, idx, :] = [(c_x - c_w) / self.image_w, - (c_y - c_h) / self.image_h, - (c_x + c_w) / self.image_w, - (c_y + c_h) / self.image_h] - idx += 1 - - if len(self.max_sizes) > 0: - max_size = self.max_sizes[s] - # second prior: aspect_ratio = 1, - c_w = c_h = math.sqrt(min_size * max_size) / 2 + if not self.min_max_aspect_ratios_order: + # rest of priors + for r in range(len(self.real_aspect_ratios)): + ar = self.real_aspect_ratios[r] + c_w = min_size * math.sqrt(ar) / 2 + c_h = (min_size / math.sqrt(ar)) / 2 + out_boxes[h, w, idx, :] = [ + (c_x - c_w) / self.image_w, (c_y - c_h) / + self.image_h, (c_x + c_w) / self.image_w, + (c_y + c_h) / self.image_h + ] + idx += 1 + + if len(self.max_sizes) > 0: + max_size = self.max_sizes[s] + # second prior: aspect_ratio = 1, + c_w = c_h = math.sqrt(min_size * max_size) / 2 + out_boxes[h, w, idx, :] = [ + (c_x - c_w) / self.image_w, (c_y - c_h) / + self.image_h, (c_x + c_w) / self.image_w, + (c_y + c_h) / self.image_h + ] + idx += 1 + else: + c_w = c_h = min_size / 2. out_boxes[h, w, idx, :] = [(c_x - c_w) / self.image_w, (c_y - c_h) / self.image_h, (c_x + c_w) / self.image_w, (c_y + c_h) / self.image_h] idx += 1 + if len(self.max_sizes) > 0: + max_size = self.max_sizes[s] + # second prior: aspect_ratio = 1, + c_w = c_h = math.sqrt(min_size * max_size) / 2 + out_boxes[h, w, idx, :] = [ + (c_x - c_w) / self.image_w, (c_y - c_h) / + self.image_h, (c_x + c_w) / self.image_w, + (c_y + c_h) / self.image_h + ] + idx += 1 + + # rest of priors + for r in range(len(self.real_aspect_ratios)): + ar = self.real_aspect_ratios[r] + if abs(ar - 1.) < 1e-6: + continue + c_w = min_size * math.sqrt(ar) / 2 + c_h = (min_size / math.sqrt(ar)) / 2 + out_boxes[h, w, idx, :] = [ + (c_x - c_w) / self.image_w, (c_y - c_h) / + self.image_h, (c_x + c_w) / self.image_w, + (c_y + c_h) / self.image_h + ] + idx += 1 # clip the prior's coordidate such that it is within[0, 1] if self.clip: @@ -137,10 +175,15 @@ class TestPriorBoxOp(OpTest): self.out_var = out_var.astype('float32') -class TestPriorBoxOpWithMaxSize(TestPriorBoxOp): +class TestPriorBoxOpWithoutMaxSize(TestPriorBoxOp): def set_max_sizes(self): self.max_sizes = [] +class TestPriorBoxOpWithSpecifiedOutOrder(TestPriorBoxOp): + def set_min_max_aspect_ratios_order(self): + self.min_max_aspect_ratios_order = True + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_profiler.py b/python/paddle/fluid/tests/unittests/test_profiler.py index cf6fe14a86aa1ab6ea3f60ad15f33d708e9b803a..9f8d33f9bbfc78b6f1a0c089b34b2f41d561c640 100644 --- a/python/paddle/fluid/tests/unittests/test_profiler.py +++ b/python/paddle/fluid/tests/unittests/test_profiler.py @@ -79,12 +79,18 @@ class TestProfiler(unittest.TestCase): pass_acc_calculator.add(value=acc, weight=b_size) pass_acc = pass_acc_calculator.eval() + @unittest.skipIf(not core.is_compiled_with_cuda(), + "profiler is enabled only with GPU") def test_cpu_profiler(self): self.net_profiler('CPU') + @unittest.skipIf(not core.is_compiled_with_cuda(), + "profiler is enabled only with GPU") def test_cuda_profiler(self): self.net_profiler('GPU') + @unittest.skipIf(not core.is_compiled_with_cuda(), + "profiler is enabled only with GPU") def test_all_profiler(self): self.net_profiler('All', '/tmp/profile_out') with open('/tmp/profile_out', 'r') as f: diff --git a/python/paddle/fluid/tests/unittests/test_program.py b/python/paddle/fluid/tests/unittests/test_program.py index c51a48239330621d8e008415f81361616467cabf..0997afc97a97333c914a3027103ec48733b410dc 100644 --- a/python/paddle/fluid/tests/unittests/test_program.py +++ b/python/paddle/fluid/tests/unittests/test_program.py @@ -17,6 +17,7 @@ import unittest from paddle.fluid.framework import Program, default_main_program, program_guard, grad_var_name import paddle.fluid.layers as layers +import paddle.fluid as fluid main_program = default_main_program() @@ -98,6 +99,39 @@ class TestProgram(unittest.TestCase): new_program = main_program.clone() self.assertNotEqual(0, len(new_program.blocks[0].all_parameters())) + def test_program_inference_optimize(self): + def net(): + reader = fluid.layers.py_reader( + capacity=10, + shapes=[[-1, 10], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + use_double_buffer=True) + in_data, label = fluid.layers.read_file(reader) + predict_label = fluid.layers.fc(in_data, size=2, act='softmax') + loss = fluid.layers.mean( + fluid.layers.cross_entropy( + input=predict_label, label=label)) + + optimizer = fluid.optimizer.Adam() + optimizer.minimize(loss) + + startup_program = fluid.Program() + main_program = fluid.Program() + with fluid.program_guard(main_program, startup_program): + net() + no_read_program = main_program.inference_optimize() + keep_read_program = main_program.inference_optimize( + export_for_deployment=False) + no_read_ops = no_read_program.global_block().ops + keep_read_ops = keep_read_program.global_block().ops + self.assertEqual(len(keep_read_ops) - len(no_read_ops), 2) + self.assertEqual(keep_read_ops[0].type, 'create_double_buffer_reader') + self.assertEqual(keep_read_ops[1].type, 'read') + + for i in range(len(no_read_ops)): + self.assertEqual(no_read_ops[i].type, keep_read_ops[i + 2].type) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py index 3f9059fb5b31cd009c068ccddc9a8938adae5772..9853fb4e9a89944bfdf2954e3d3d86fef92ac93c 100644 --- a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py +++ b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py @@ -68,7 +68,7 @@ class TestOpDesc(unittest.TestCase): self.assertEqual(8, len(op.attr_names())) op.set_block_attr("block_attr", program_desc.block(0)) - self.assertEqual(0, op.block_attr("block_attr")) + self.assertEqual(0, op.block_attr_id("block_attr")) mul_op = block.append_op() mul_op.set_type("mul") @@ -181,13 +181,13 @@ class TestBlockDesc(unittest.TestCase): self.assertIsNotNone(block) op1 = block.append_op() op2 = block.append_op() - op0 = block.prepend_op() + op0 = block._prepend_op() all_ops = [] - for idx in xrange(0, block.op_size()): + for idx in range(0, block.op_size()): all_ops.append(block.op(idx)) self.assertEqual(all_ops, [op0, op1, op2]) - def test_remove_op(self): + def test__remove_op(self): program = Program() program_desc = program.desc self.assertIsNotNone(program_desc) @@ -201,11 +201,11 @@ class TestBlockDesc(unittest.TestCase): op1.set_type("test") op2.set_type("test") - block.remove_op(1, 2) - program.sync_with_cpp() + block._remove_op(1, 2) + program._sync_with_cpp() all_ops = [] - for idx in xrange(0, block.op_size()): + for idx in range(0, block.op_size()): all_ops.append(block.op(idx)) self.assertEqual(all_ops, [op0, op2]) diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py b/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py new file mode 100644 index 0000000000000000000000000000000000000000..f9bda5e4701f693f41fe7041ba0f5ec80b6fc31c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py @@ -0,0 +1,100 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import paddle.fluid as fluid +import numpy as np +from threading import Thread + + +def feed_data(feed_queue, inputs): + for in_data in inputs: + feed_queue.push(in_data) + + +class TestPyReader(unittest.TestCase): + def setUp(self): + self.capacity = 10 + self.batch_size_min = 10 + self.batch_size_max = 20 + self.shapes = [(-1, 3, 2, 1), (-1, 1)] + self.lod_levels = [0, 0] + self.dtypes = ['float32', 'int64'] + self.iterations = 20 + + def test_single_thread_main(self): + self.main(use_thread=False) + + def test_multiple_thread_main(self): + self.main(use_thread=True) + + def main(self, use_thread=False): + with fluid.program_guard(fluid.Program(), fluid.Program()): + place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda( + ) else fluid.CPUPlace() + executor = fluid.Executor(place) + + data_file = fluid.layers.py_reader( + capacity=self.capacity, + dtypes=self.dtypes, + lod_levels=self.lod_levels, + shapes=self.shapes) + feed_queue = data_file.queue + read_out_data = fluid.layers.read_file(data_file) + self.inputs = [] + + for i in range(self.iterations): + in_data = fluid.LoDTensorArray() + batch_size = np.random.random_integers(self.batch_size_min, + self.batch_size_max) + for shape, dtype in zip(self.shapes, self.dtypes): + next_data = np.random.uniform( + low=0, high=1000, + size=(batch_size, ) + shape[1:]).astype(dtype) + in_data.append( + fluid.executor._as_lodtensor(next_data, place)) + + self.inputs.append(in_data) + + executor.run(fluid.default_startup_program()) + self.outputs = [] + if use_thread: + thread = Thread( + target=feed_data, args=(feed_queue, self.inputs)) + thread.start() + for in_data in self.inputs: + self.outputs.append( + executor.run(fetch_list=list(read_out_data))) + else: + for in_data in self.inputs: + feed_queue.push(in_data) + self.outputs.append( + executor.run(fetch_list=list(read_out_data))) + + feed_queue.close() + self.validate() + + def validate(self): + self.assertEqual(len(self.inputs), len(self.outputs)) + for in_data_list, out_data_list in zip(self.inputs, self.outputs): + self.assertEqual(len(in_data_list), len(out_data_list)) + in_data_list_np = [ + np.array(in_lod_tensor) for in_lod_tensor in in_data_list + ] + for in_data, out_data in zip(in_data_list_np, out_data_list): + self.assertTrue((in_data == out_data).all()) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py new file mode 100644 index 0000000000000000000000000000000000000000..9a379bdbaa7e278879117a8cdc2dddb335a10ca1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py @@ -0,0 +1,226 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import paddle.fluid as fluid +import paddle.fluid.core as core +import numpy as np +import threading +import multiprocessing +import os + + +def as_tensor(np_array_or_tensor, place=None): + if isinstance(np_array_or_tensor, fluid.LoDTensor): + return np_array_or_tensor + + if place is None: + place = fluid.CPUPlace() + + tensor = fluid.LoDTensor() + tensor.set(np_array_or_tensor, place) + return tensor + + +def as_numpy(tensor_or_numpy): + return tensor_or_numpy if isinstance( + tensor_or_numpy, np.ndarray) else np.array(tensor_or_numpy) + + +def feed_data(feed_queue, reader): + data_generator = reader() + while True: + data = next(data_generator, None) + if data is None or not feed_queue.push(data): + break + + +def simple_fc_net(in_size, + class_num, + hidden_sizes, + batch_size, + queue_capacity, + use_double_buffer=False): + reader = fluid.layers.py_reader( + capacity=queue_capacity, + shapes=[[-1, in_size], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + use_double_buffer=False) + feed_queue = reader.queue + reader = fluid.layers.batch(reader, batch_size=batch_size) + if use_double_buffer: + reader = fluid.layers.double_buffer(reader) + + in_data, label = fluid.layers.read_file(reader) + + hidden = in_data + for hidden_size in hidden_sizes: + hidden = fluid.layers.fc( + hidden, + size=hidden_size, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + + predict_label = fluid.layers.fc(hidden, size=class_num, act='softmax') + loss = fluid.layers.mean( + fluid.layers.cross_entropy( + input=predict_label, label=label)) + + optimizer = fluid.optimizer.Adam() + optimizer.minimize(loss) + return in_data, label, loss, optimizer, feed_queue + + +class TestPyReaderUsingExecutor(unittest.TestCase): + def setUp(self): + self.in_size = 1000 + self.hidden_sizes = [50, 30, 20] + self.class_num = 10 + self.batch_size = 32 + self.iterations = 10 + self.queue_capacity = 50 + + def test(self): + for use_cuda in [False, True]: + for use_parallel_executor in [False, True]: + for use_double_buffer in [False, True]: + print('Test Parameters:'), + print({ + 'use_cuda': use_cuda, + 'use_parallel_executor': use_parallel_executor, + 'use_double_buffer': use_double_buffer + }) + self.main(use_cuda, use_parallel_executor, + use_double_buffer) + + def random_reader(self): + def reader(): + self.inputs = [] + cnt = 0 + while True: + tensors = fluid.LoDTensorArray() + in_data = np.random.uniform( + low=0, high=1, size=(1, self.in_size)).astype('float32') + tensors.append(as_tensor(in_data)) + label = np.random.random_integers( + low=0, high=self.class_num - 1, size=(1, 1)).astype('int64') + tensors.append(as_tensor(label)) + + if cnt < self.iterations * self.batch_size * self.batch_size_times: + if cnt % (self.batch_size * self.batch_size_times) == 0: + self.inputs.append([in_data, label]) + else: + self.inputs[-1][0] = np.concatenate( + (self.inputs[-1][0], in_data), axis=0) + self.inputs[-1][1] = np.concatenate( + (self.inputs[-1][1], label), axis=0) + elif not self.use_double_buffer: + break + + yield tensors + cnt += 1 + + yield None + + return reader + + def main(self, + use_cuda=True, + use_parallel_executor=False, + use_double_buffer=False): + assert not use_cuda or use_cuda and core.is_compiled_with_cuda() + + self.use_cuda = use_cuda + self.use_parallel_executor = use_parallel_executor + self.use_double_buffer = use_double_buffer + + startup_program = fluid.Program() + main_program = fluid.Program() + + with fluid.program_guard(main_program, startup_program): + in_data, label, loss, optimizer, feed_queue = simple_fc_net( + in_size=self.in_size, + class_num=self.class_num, + hidden_sizes=self.hidden_sizes, + batch_size=self.batch_size, + queue_capacity=self.queue_capacity, + use_double_buffer=self.use_double_buffer) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + startup_exe = fluid.Executor(place) + startup_exe.run(startup_program) + + if use_parallel_executor: + main_exe = fluid.ParallelExecutor(use_cuda, loss_name=loss.name) + if use_cuda: + self.batch_size_times = core.get_cuda_device_count() + else: + self.batch_size_times = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + else: + main_exe = startup_exe + self.batch_size_times = 1 + + reader = self.random_reader() + thread = threading.Thread( + target=feed_data, args=(feed_queue, reader)) + thread.start() + + self.outputs = [] + for _ in range(self.iterations): + fetches = main_exe.run(fetch_list=[in_data.name, label.name]) + fetches = [as_numpy(fetch) for fetch in fetches] + self.outputs.append(fetches) + + feed_queue.close() + self.validate() + + def validate(self): + self.assertEqual(len(self.inputs), len(self.outputs)) + for batch_in, batch_out in zip(self.inputs, self.outputs): + self.assertEqual(len(batch_in), len(batch_out)) + if self.use_parallel_executor and not self.use_double_buffer: + self.validate_unordered_batch(batch_in, batch_out) + else: + for in_data, out_data in zip(batch_in, batch_out): + self.assertEqual(in_data.shape, out_data.shape) + if not self.use_parallel_executor: + self.assertTrue((in_data == out_data).all()) + + def validate_unordered_batch(self, batch_in, batch_out): + out_index_left_set = set(range(self.batch_size * self.batch_size_times)) + mapping_num = 0 + for i in range(self.batch_size * self.batch_size_times): + for j in out_index_left_set: + flag = True + for k in range(len(batch_in)): + in_data = batch_in[k][i] + out_data = batch_out[k][j] + if (in_data != out_data).any(): + flag = False + break + + if flag: + out_index_left_set.remove(j) + mapping_num += 1 + break + + self.assertEqual(mapping_num, self.batch_size * self.batch_size_times) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_reader_reset.py b/python/paddle/fluid/tests/unittests/test_reader_reset.py new file mode 100644 index 0000000000000000000000000000000000000000..3ad85d57485956e0cadb197dadd172516fa15c39 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_reader_reset.py @@ -0,0 +1,116 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import paddle.v2 as paddle +import numpy as np +import unittest + + +class TestReaderReset(unittest.TestCase): + def prepare_data(self): + def fake_data_generator(): + for n in range(self.total_ins_num): + yield np.ones(self.ins_shape) * n, n + + # Prepare data + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(fake_data_generator, batch_size=1) + feeder = fluid.DataFeeder( + feed_list=[ + fluid.layers.data( + name='data', shape=[3], dtype='float32'), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + fluid.recordio_writer.convert_reader_to_recordio_file( + self.data_file_name, reader, feeder) + + def setUp(self): + self.use_cuda = fluid.core.is_compiled_with_cuda() + self.data_file_name = './reader_reset_test.recordio' + self.ins_shape = [3] + self.batch_size = 5 + self.total_ins_num = self.batch_size * 20 + self.test_pass_num = 100 + self.prepare_data() + + def main(self, with_double_buffer): + main_prog = fluid.Program() + startup_prog = fluid.Program() + + with fluid.program_guard(main_prog, startup_prog): + data_reader_handle = fluid.layers.io.open_files( + filenames=[self.data_file_name], + shapes=[[-1] + self.ins_shape, [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + thread_num=1, + pass_num=1) + data_reader = fluid.layers.io.batch(data_reader_handle, + self.batch_size) + if with_double_buffer: + data_reader = fluid.layers.double_buffer(data_reader) + image, label = fluid.layers.read_file(data_reader) + fetch_list = [image.name, label.name] + + place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + + build_strategy = fluid.BuildStrategy() + if with_double_buffer: + build_strategy.enable_data_balance = True + exec_strategy = fluid.ExecutionStrategy() + parallel_exe = fluid.ParallelExecutor( + use_cuda=self.use_cuda, + main_program=main_prog, + build_strategy=build_strategy, + exec_strategy=exec_strategy) + + data_appeared = [False] * self.total_ins_num + pass_count = 0 + while (True): + try: + data_val, label_val = parallel_exe.run(fetch_list, + return_numpy=True) + ins_num = data_val.shape[0] + broadcasted_label = np.ones((ins_num, ) + tuple( + self.ins_shape)) * label_val.reshape((ins_num, 1)) + self.assertEqual(data_val.all(), broadcasted_label.all()) + for l in label_val: + self.assertFalse(data_appeared[l[0]]) + data_appeared[l[0]] = True + + except fluid.core.EOFException: + pass_count += 1 + if with_double_buffer: + data_appeared = data_appeared[:-parallel_exe.device_count * + self.batch_size] + for i in data_appeared: + self.assertTrue(i) + if pass_count < self.test_pass_num: + data_appeared = [False] * self.total_ins_num + data_reader_handle.reset() + else: + break + + def test_all(self): + self.main(with_double_buffer=False) + self.main(with_double_buffer=True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_recordio_reader.py b/python/paddle/fluid/tests/unittests/test_recordio_reader.py index f32050014d7ace5aee4aca75a47bfc6a75ff91c2..69a522e273db017ac55b408276b4a28f5f907c42 100644 --- a/python/paddle/fluid/tests/unittests/test_recordio_reader.py +++ b/python/paddle/fluid/tests/unittests/test_recordio_reader.py @@ -68,8 +68,7 @@ class TestRecordIO(unittest.TestCase): while True: try: tmp, = exe.run(fetch_list=[avg_loss]) - except fluid.core.EnforceNotMet as ex: - self.assertIn("There is no next data.", ex.message) + except fluid.core.EOFException: break avg_loss_np.append(tmp) diff --git a/python/paddle/fluid/tests/unittests/test_recurrent_op.py b/python/paddle/fluid/tests/unittests/test_recurrent_op.py index d6ff18430e319e236f03d5661381e923cc956590..2e22df2beba9d74e28788fb72f6f7f7f2bef534e 100644 --- a/python/paddle/fluid/tests/unittests/test_recurrent_op.py +++ b/python/paddle/fluid/tests/unittests/test_recurrent_op.py @@ -203,12 +203,12 @@ class RecurrentOpTest1(unittest.TestCase): num_grad[idx], ana_grad[idx], rtol=0.1).all()) def check_forward(self): - print 'test recurrent op forward' + print('test recurrent op forward') pd_output = self.forward() py_output = self.py_rnn.forward() - print 'pd_output', pd_output + print('pd_output', pd_output) print - print 'py_output', py_output + print('py_output', py_output) self.assertEqual(pd_output.shape, py_output.shape) self.assertTrue(np.isclose(pd_output, py_output, rtol=0.1).all()) @@ -445,7 +445,7 @@ class RecurrentOpNoMemBootTest(RecurrentOpTest1): self.py_rnn = RecurrentOpNoMemBootTest.PySimpleRNN4(self.input_shape, self.output_shape) self.output = layers.mean(self.create_rnn_op(), **self.p_info) - print self.main_program + print(self.main_program) def create_rnn_op(self): x = layers.data( diff --git a/python/paddle/fluid/tests/unittests/test_reduce_op.py b/python/paddle/fluid/tests/unittests/test_reduce_op.py index 865c2b7df085aa6a6cb0d6eb461c342ce08695cd..06d116601bf733986ccf9c725340456ab1258be2 100644 --- a/python/paddle/fluid/tests/unittests/test_reduce_op.py +++ b/python/paddle/fluid/tests/unittests/test_reduce_op.py @@ -89,15 +89,11 @@ class TestProdOp(OpTest): self.check_grad(['X'], 'Out') -class TestKeepDimReduce(OpTest): +class Test1DReduce(OpTest): def setUp(self): self.op_type = "reduce_sum" - self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} - self.attrs = {'dim': [-2], 'keep_dim': True} - self.outputs = { - 'Out': - self.inputs['X'].sum(axis=tuple(self.attrs['dim']), keepdims=True) - } + self.inputs = {'X': np.random.random(20).astype("float64")} + self.outputs = {'Out': self.inputs['X'].sum(axis=0)} def test_check_output(self): self.check_output() @@ -106,32 +102,82 @@ class TestKeepDimReduce(OpTest): self.check_grad(['X'], 'Out') -class Test1DReduce(OpTest): +class Test2DReduce0(Test1DReduce): def setUp(self): self.op_type = "reduce_sum" - self.inputs = {'X': np.random.random(20).astype("float64")} + self.attrs = {'dim': [0]} + self.inputs = {'X': np.random.random((20, 10)).astype("float64")} self.outputs = {'Out': self.inputs['X'].sum(axis=0)} - def test_check_output(self): - self.check_output() - def test_check_grad(self): - self.check_grad(['X'], 'Out') +class Test2DReduce1(Test1DReduce): + def setUp(self): + self.op_type = "reduce_sum" + self.attrs = {'dim': [1]} + self.inputs = {'X': np.random.random((20, 10)).astype("float64")} + self.outputs = { + 'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim'])) + } -class TestReduceAll(OpTest): +class Test3DReduce0(Test1DReduce): + def setUp(self): + self.op_type = "reduce_sum" + self.attrs = {'dim': [1]} + self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")} + self.outputs = { + 'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim'])) + } + + +class Test3DReduce1(Test1DReduce): + def setUp(self): + self.op_type = "reduce_sum" + self.attrs = {'dim': [2]} + self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")} + self.outputs = { + 'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim'])) + } + + +class Test3DReduce2(Test1DReduce): + def setUp(self): + self.op_type = "reduce_sum" + self.attrs = {'dim': [-2]} + self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")} + self.outputs = { + 'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim'])) + } + + +class Test3DReduce3(Test1DReduce): + def setUp(self): + self.op_type = "reduce_sum" + self.attrs = {'dim': [1, 2]} + self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")} + self.outputs = { + 'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim'])) + } + + +class TestKeepDimReduce(Test1DReduce): + def setUp(self): + self.op_type = "reduce_sum" + self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} + self.attrs = {'dim': [1], 'keep_dim': True} + self.outputs = { + 'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']), + keepdims=self.attrs['keep_dim']) + } + + +class TestReduceAll(Test1DReduce): def setUp(self): self.op_type = "reduce_sum" self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")} self.attrs = {'reduce_all': True} self.outputs = {'Out': self.inputs['X'].sum()} - def test_check_output(self): - self.check_output() - - def test_check_grad(self): - self.check_grad(['X'], 'Out') - ## reduction in multi dims class TestReduceMeanOpMultiAxises(OpTest): diff --git a/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py b/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py index 76d0d2f2fe80e409dc1b7fa858d43fbc6ad960ef..6e1cd56b3e309fc014dc981a1e3aa841159fca15 100644 --- a/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py +++ b/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py @@ -15,6 +15,7 @@ import unittest import paddle.fluid as fluid import paddle.fluid.core as core +from paddle.fluid.layers.control_flow import lod_rank_table import numpy @@ -34,7 +35,7 @@ class TestReorderLoDTensor(unittest.TestCase): dat.stop_gradient = False rank_dat = fluid.layers.data( name=cls.data_desc[1][0], shape=cls.data_desc[1][1]) - table = fluid.layers.lod_rank_table(rank_dat) + table = lod_rank_table(rank_dat) new_dat = fluid.layers.reorder_lod_tensor_by_rank( x=dat, rank_table=table) loss = fluid.layers.reduce_sum(new_dat) @@ -70,11 +71,10 @@ class TestReorderLoDTensor(unittest.TestCase): lod_level_i = numpy.random.randint( low=1, high=5, - size=self.num_seq if i == 0 else lod_level_i[-1]) - lod_level_i = [0] + numpy.cumsum(lod_level_i).tolist() + size=self.num_seq if i == 0 else sum(lod_level_i)).tolist() data_lod.append(lod_level_i) data_value = numpy.random.random( - size=[data_lod[-1][-1] if data_lod else self.num_seq + size=[sum(data_lod[-1]) if data_lod else self.num_seq ] + data_shape).astype('float32') self.data[data_name] = (data_value, data_lod) @@ -84,29 +84,36 @@ class TestReorderLoDTensor(unittest.TestCase): tensor = fluid.Tensor() tensor.set(self.data[desc[0]][0], place) if self.data[desc[0]][1]: - tensor.set_lod(self.data[desc[0]][1]) + tensor.set_recursive_sequence_lengths(self.data[desc[0]][1]) self.inputs[desc[0]] = tensor def reorder(self): - level = 0 + def convert_to_offset(lod): + offset_lod = [[0] for i in lod] + for i, level in enumerate(lod): + for seq_len in level: + offset_lod[i].append(offset_lod[i][-1] + seq_len) + return offset_lod + level = 0 # compute the rank_table according to ref_lod ref_lod = self.data[self.data_desc[1][0]][1][level] rank_table = [] # list of (index, length) - for i in range(len(ref_lod) - 1): - rank_table.append((i, ref_lod[i + 1] - ref_lod[i])) + for i in range(len(ref_lod)): + rank_table.append((i, ref_lod[i])) rank_table = sorted(rank_table, lambda x, y: y[1] - x[1]) # compute the input sequence info according to input_lod input_value, input_lod = self.data[self.data_desc[0][0]] + offset_lod = convert_to_offset(input_lod) input_table = [] # list of (offset, length, sub_lod) - if input_lod: - for i in range(len(input_lod[level]) - 1): + if offset_lod: + for i in range(len(offset_lod[level]) - 1): start_idx = i end_idx = i + 1 sub_lod = [] - for lod_level_i in input_lod[level:]: + for lod_level_i in offset_lod[level:]: sub_lod_i = [] for idx in range(start_idx, end_idx): sub_lod_i.append(lod_level_i[idx + 1] - lod_level_i[ @@ -132,10 +139,9 @@ class TestReorderLoDTensor(unittest.TestCase): input_seq_sub_lod = input_table[index][2] if len(output_lod) == 0: - output_lod = [[0] for i in input_seq_sub_lod] - for i, sub_lod_i in enumerate(input_seq_sub_lod): - for idx_sub in sub_lod_i: - output_lod[i].append(output_lod[i][-1] + idx_sub) + output_lod = [[] for i in input_seq_sub_lod] + for i, level in enumerate(input_seq_sub_lod): + output_lod[i].extend(level) return output_value, output_lod def test_reorder_lod_tensor(self): @@ -148,7 +154,8 @@ class TestReorderLoDTensor(unittest.TestCase): self.assertTrue( numpy.allclose( numpy.array(actual_output), expect_output, atol=0.001)) - self.assertEqual(expect_output_lod, actual_output.lod()) + self.assertEqual(expect_output_lod, + actual_output.recursive_sequence_lengths()) # check gradient expect_grad = numpy.ones_like(self.data[self.data_desc[0][0]][0]) expect_grad_lod = self.data[self.data_desc[0][0]][1] @@ -156,7 +163,8 @@ class TestReorderLoDTensor(unittest.TestCase): self.assertTrue( numpy.allclose( numpy.array(actual_grad), expect_grad, atol=0.001)) - self.assertEqual(expect_grad_lod, actual_grad.lod()) + self.assertEqual(expect_grad_lod, + actual_grad.recursive_sequence_lengths()) def test_reorder_tensor(self): self.data_desc[0][-1] = 0 # input is tensor @@ -168,7 +176,8 @@ class TestReorderLoDTensor(unittest.TestCase): self.assertTrue( numpy.allclose( numpy.array(actual_output), expect_output, atol=0.001)) - self.assertEqual(expect_output_lod, actual_output.lod()) + self.assertEqual(expect_output_lod, + actual_output.recursive_sequence_lengths()) # check gradient expect_grad = numpy.ones_like(self.data[self.data_desc[0][0]][0]) expect_grad_lod = self.data[self.data_desc[0][0]][1] @@ -176,14 +185,14 @@ class TestReorderLoDTensor(unittest.TestCase): self.assertTrue( numpy.allclose( numpy.array(actual_grad), expect_grad, atol=0.001)) - self.assertEqual(expect_grad_lod, actual_grad.lod()) + self.assertEqual(expect_grad_lod, + actual_grad.recursive_sequence_lengths()) # compare outputs between LodTensors with explicit and implicit lod # use the same data but set the input lod explicitly - input_lod = [[ - i for i in range(len(self.data[self.data_desc[0][0]][0]) + 1) - ]] - self.inputs[self.data_desc[0][0]].set_lod(input_lod) + input_lod = [[1] * len(self.data[self.data_desc[0][0]][0])] + self.inputs[self.data_desc[0][0]].set_recursive_sequence_lengths( + input_lod) # preserve the output of LodTensor with implicit lod to compare expect_output = [ numpy.array(actual_output) for actual_output in self.actual_outputs diff --git a/python/paddle/fluid/tests/unittests/test_reshape_op.py b/python/paddle/fluid/tests/unittests/test_reshape_op.py index f51b5a7e9907294a5b91c920a363830d8b9a7137..2f5558578ac2a002a83c2a7e027ec5a96d8b4414 100644 --- a/python/paddle/fluid/tests/unittests/test_reshape_op.py +++ b/python/paddle/fluid/tests/unittests/test_reshape_op.py @@ -25,7 +25,7 @@ class TestReshapeOp(OpTest): self.op_type = "reshape" self.inputs = {"X": np.random.random(ori_shape).astype("float32")} - self.attrs = {"shape": new_shape, "inplace": False} + self.attrs = {"shape": new_shape} self.outputs = {"Out": self.inputs["X"].reshape(new_shape)} def test_check_output(self): @@ -42,7 +42,7 @@ class TestReshapeOpDimInfer1(OpTest): self.op_type = "reshape" self.inputs = {"X": np.random.random(ori_shape).astype("float32")} - self.attrs = {"shape": new_shape, "inplace": False} + self.attrs = {"shape": new_shape} self.outputs = {"Out": self.inputs["X"].reshape(self.attrs["shape"])} def test_check_output(self): @@ -60,7 +60,7 @@ class TestReshapeOpDimInfer2(OpTest): self.op_type = "reshape" self.inputs = {"X": np.random.random(ori_shape).astype("float32")} - self.attrs = {"shape": new_shape, "inplace": False} + self.attrs = {"shape": new_shape} self.outputs = {"Out": self.inputs["X"].reshape(infered_shape)} def test_check_output(self): diff --git a/python/paddle/fluid/tests/unittests/test_reverse_op.py b/python/paddle/fluid/tests/unittests/test_reverse_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f845575a02869f08299d76b5600074598ca27f6c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_reverse_op.py @@ -0,0 +1,67 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestReverseOp(OpTest): + def initTestCase(self): + self.x = np.random.random((3, 4)).astype('float32') + self.axis = [0] + + def setUp(self): + self.initTestCase() + self.op_type = "reverse" + self.inputs = {"X": self.x} + self.attrs = {'axis': self.axis} + out = self.x + for a in self.axis: + out = np.flip(out, axis=a) + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + +class TestCase0(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4)).astype('float32') + self.axis = [1] + + +class TestCase1(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4)).astype('float32') + self.axis = [0, 1] + + +class TestCase2(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4, 5)).astype('float32') + self.axis = [0, 2] + + +class TestCase3(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4, 5)).astype('float32') + self.axis = [1, 2] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py index 3d754aff3a73e7168e2123483b26e5e3a3585a4e..df5684ab173a4889dd7b693f9246bafd12e0345f 100644 --- a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py +++ b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py @@ -107,7 +107,7 @@ class TestROIPoolOp(OpTest): rois = [] self.rois_lod = [[]] for bno in range(self.batch_size): - self.rois_lod[0].append(len(rois)) + self.rois_lod[0].append(bno + 1) for i in range(bno + 1): x1 = np.random.random_integers( 0, self.width / self.spatial_scale - self.pooled_width) @@ -121,7 +121,6 @@ class TestROIPoolOp(OpTest): roi = [bno, x1, y1, x2, y2] rois.append(roi) - self.rois_lod[0].append(len(rois)) self.rois_num = len(rois) self.rois = np.array(rois).astype("int64") diff --git a/python/paddle/fluid/tests/unittests/test_row_conv_op.py b/python/paddle/fluid/tests/unittests/test_row_conv_op.py index 30f1efbcbcb11332c85c9d5489f22c17b06c2b36..07dcd108689ae6069e30fe22029258d192215549 100644 --- a/python/paddle/fluid/tests/unittests/test_row_conv_op.py +++ b/python/paddle/fluid/tests/unittests/test_row_conv_op.py @@ -19,8 +19,10 @@ from op_test import OpTest def row_conv_forward(x, lod, wt): out = np.zeros_like(x) - seq_info = lod[0] - num_sequences = len(seq_info) - 1 + num_sequences = len(lod[0]) + seq_info = [0] + for seq_len in lod[0]: + seq_info.append(seq_info[-1] + seq_len) context_length = wt.shape[0] for i in range(num_sequences): # loop over number of sequences @@ -32,7 +34,6 @@ def row_conv_forward(x, lod, wt): cur_timesteps = end - start for j in range(cur_timesteps): # loop over different timesteps for k in range(context_length): - if j + k >= cur_timesteps: continue curoutput[j, :] += curinput[j + k, :] * wt[k, :] @@ -44,8 +45,8 @@ class TestRowConvOp1(OpTest): def setUp(self): self.op_type = "row_conv" - lod = [[0, 2, 5, 7]] - T = lod[0][-1] + lod = [[2, 3, 2]] + T = sum(lod[0]) D = 16 context_length = 2 @@ -75,8 +76,8 @@ class TestRowConvOp2(OpTest): def setUp(self): self.op_type = "row_conv" - lod = [[0, 20, 50, 100]] - T = lod[0][-1] + lod = [[20, 30, 50]] + T = sum(lod[0]) D = 35 context_length = 35 diff --git a/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py b/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py new file mode 100644 index 0000000000000000000000000000000000000000..df6e0faaca6fd007b39a8f358d964055e149a025 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py @@ -0,0 +1,103 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import paddle.fluid.core as core +from op_test import OpTest + + +def rpn_target_assign(iou, rpn_batch_size_per_im, rpn_positive_overlap, + rpn_negative_overlap, fg_fraction): + iou = np.transpose(iou) + anchor_to_gt_max = iou.max(axis=1) + gt_to_anchor_argmax = iou.argmax(axis=0) + gt_to_anchor_max = iou[gt_to_anchor_argmax, np.arange(iou.shape[1])] + anchors_with_max_overlap = np.where(iou == gt_to_anchor_max)[0] + + tgt_lbl = np.ones((iou.shape[0], ), dtype=np.int32) * -1 + tgt_lbl[anchors_with_max_overlap] = 1 + tgt_lbl[anchor_to_gt_max >= rpn_positive_overlap] = 1 + + num_fg = int(fg_fraction * rpn_batch_size_per_im) + fg_inds = np.where(tgt_lbl == 1)[0] + if len(fg_inds) > num_fg: + disable_inds = np.random.choice( + fg_inds, size=(len(fg_inds) - num_fg), replace=False) + tgt_lbl[disable_inds] = -1 + fg_inds = np.where(tgt_lbl == 1)[0] + + num_bg = rpn_batch_size_per_im - np.sum(tgt_lbl == 1) + bg_inds = np.where(anchor_to_gt_max < rpn_negative_overlap)[0] + if len(bg_inds) > num_bg: + enable_inds = bg_inds[np.random.randint(len(bg_inds), size=num_bg)] + tgt_lbl[enable_inds] = 0 + bg_inds = np.where(tgt_lbl == 0)[0] + + loc_index = fg_inds + score_index = np.hstack((fg_inds, bg_inds)) + tgt_lbl = np.expand_dims(tgt_lbl, axis=1) + return loc_index, score_index, tgt_lbl + + +class TestRpnTargetAssignOp(OpTest): + def setUp(self): + iou = np.random.random((10, 8)).astype("float32") + self.op_type = "rpn_target_assign" + self.inputs = {'DistMat': iou} + self.attrs = { + 'rpn_batch_size_per_im': 256, + 'rpn_positive_overlap': 0.95, + 'rpn_negative_overlap': 0.3, + 'fg_fraction': 0.25, + 'fix_seed': True + } + loc_index, score_index, tgt_lbl = rpn_target_assign(iou, 256, 0.95, 0.3, + 0.25) + self.outputs = { + 'LocationIndex': loc_index, + 'ScoreIndex': score_index, + 'TargetLabel': tgt_lbl, + } + + def test_check_output(self): + self.check_output() + + +class TestRpnTargetAssignOp2(OpTest): + def setUp(self): + iou = np.random.random((10, 20)).astype("float32") + self.op_type = "rpn_target_assign" + self.inputs = {'DistMat': iou} + self.attrs = { + 'rpn_batch_size_per_im': 128, + 'rpn_positive_overlap': 0.5, + 'rpn_negative_overlap': 0.5, + 'fg_fraction': 0.5, + 'fix_seed': True + } + loc_index, score_index, tgt_lbl = rpn_target_assign(iou, 128, 0.5, 0.5, + 0.5) + self.outputs = { + 'LocationIndex': loc_index, + 'ScoreIndex': score_index, + 'TargetLabel': tgt_lbl, + } + + def test_check_output(self): + self.check_output() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_selected_rows.py b/python/paddle/fluid/tests/unittests/test_selected_rows.py index 3d7b86787fbf0a855bcd86b8a873c9134cb1d5cc..f504a06ffff8cb636498652554fca05e22bb905d 100644 --- a/python/paddle/fluid/tests/unittests/test_selected_rows.py +++ b/python/paddle/fluid/tests/unittests/test_selected_rows.py @@ -40,12 +40,12 @@ class TestSelectedRows(unittest.TestCase): # compare tensor self.assertAlmostEqual(2.0, - selected_rows.get_tensor().get_float_element(0)) + selected_rows.get_tensor()._get_float_element(0)) self.assertAlmostEqual(1.0, - selected_rows.get_tensor().get_float_element(1)) + selected_rows.get_tensor()._get_float_element(1)) self.assertAlmostEqual( 4.0, - selected_rows.get_tensor().get_float_element(2 * row_numel + 8)) + selected_rows.get_tensor()._get_float_element(2 * row_numel + 8)) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_seq_concat_op.py b/python/paddle/fluid/tests/unittests/test_seq_concat_op.py index 10592d127fafdf202c65fcfa91b5c464cc60e96c..11ffa761a690eb1f9f6dc50c45128a99301741db 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_concat_op.py +++ b/python/paddle/fluid/tests/unittests/test_seq_concat_op.py @@ -18,14 +18,19 @@ import sys from op_test import OpTest -def to_abs_lod(lod): - if len(lod) == 0 or len(lod) == 1: - return lod +def to_abs_offset_lod(lod): + offset_lod = [[0] for i in lod] + for i, level in enumerate(lod): + for seq_len in level: + offset_lod[i].append(offset_lod[i][-1] + seq_len) + + if len(offset_lod) == 0 or len(offset_lod) == 1: + return offset_lod import copy - new_lod = copy.deepcopy(lod) - for idx, val in enumerate(lod[0]): - new_lod[0][idx] = lod[1][val] - return new_lod + new_offset_lod = copy.deepcopy(offset_lod) + for idx, val in enumerate(offset_lod[0]): + new_offset_lod[0][idx] = offset_lod[1][val] + return new_offset_lod def seq_concat(inputs, level): @@ -35,11 +40,11 @@ def seq_concat(inputs, level): x1 = inputs['X'][1][1][0] level_idx = len(lod0) - level - 1 outs = [] - for i in range(len(lod0[level_idx]) - 1): - sub_x0 = x0[to_abs_lod(lod0)[level_idx][i]:to_abs_lod(lod0)[level_idx][ - i + 1], :] - sub_x1 = x1[to_abs_lod(lod1)[level_idx][i]:to_abs_lod(lod1)[level_idx][ - i + 1], :] + for i in range(len(lod0[level_idx])): + sub_x0 = x0[to_abs_offset_lod(lod0)[level_idx][i]:to_abs_offset_lod( + lod0)[level_idx][i + 1], :] + sub_x1 = x1[to_abs_offset_lod(lod1)[level_idx][i]:to_abs_offset_lod( + lod1)[level_idx][i + 1], :] outs.append(np.concatenate((sub_x0, sub_x1), axis=0)) return np.concatenate(outs, axis=0) @@ -48,9 +53,9 @@ class TestSeqConcatOp(OpTest): def set_data(self): # two level, batch size is 3 x0 = np.random.random((4, 6, 3)).astype('float32') - lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]] + lod0 = [[2, 2], [1, 1, 1, 1]] x1 = np.random.random((4, 8, 3)).astype('float32') - lod1 = [[0, 2, 4], [0, 1, 2, 3, 4]] + lod1 = [[2, 2], [1, 1, 1, 1]] axis = 1 level = 1 self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]} @@ -72,14 +77,14 @@ class TestSeqConcatOpLevelZeroNestedSequence(TestSeqConcatOp): def set_data(self): # two level, batch size is 3 x0 = np.random.random((4, 6, 3)).astype('float32') - lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]] + lod0 = [[2, 2], [1, 1, 1, 1]] x1 = np.random.random((7, 6, 3)).astype('float32') - lod1 = [[0, 2, 4], [0, 1, 3, 5, 7]] + lod1 = [[2, 2], [1, 2, 2, 2]] axis = 0 level = 0 self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]} self.attrs = {'axis': axis, 'level': level} - out_lod = [[0, 2, 4], [0, 2, 5, 8, 11]] + out_lod = [[2, 2], [2, 3, 3, 3]] self.outputs = {'Out': (seq_concat(self.inputs, level), out_lod)} @@ -87,14 +92,14 @@ class TestSeqConcatOplevelOneNestedSequence(TestSeqConcatOp): def set_data(self): # two level, batch size is 3 x0 = np.random.random((4, 6, 3)).astype('float32') - lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]] + lod0 = [[2, 2], [1, 1, 1, 1]] x1 = np.random.random((7, 6, 3)).astype('float32') - lod1 = [[0, 3, 4], [0, 1, 3, 5, 7]] + lod1 = [[3, 1], [1, 2, 2, 2]] axis = 0 level = 1 self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]} self.attrs = {'axis': axis, 'level': level} - out_lod = [[0, 5, 8], [0, 1, 2, 3, 5, 7, 8, 9, 11]] + out_lod = [[5, 3], [1, 1, 1, 2, 2, 1, 1, 2]] self.outputs = {'Out': (seq_concat(self.inputs, level), out_lod)} @@ -102,14 +107,14 @@ class TestSeqConcatOpLevelZeroSequence(TestSeqConcatOp): def set_data(self): # two level, batch size is 3 x0 = np.random.random((4, 3, 4)).astype('float32') - lod0 = [[0, 1, 2, 3, 4]] + lod0 = [[1, 1, 1, 1]] x1 = np.random.random((7, 3, 4)).astype('float32') - lod1 = [[0, 1, 3, 5, 7]] + lod1 = [[1, 2, 2, 2]] axis = 0 level = 0 self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]} self.attrs = {'axis': axis, 'level': level} - out_lod = [[0, 2, 5, 8, 11]] + out_lod = [[2, 3, 3, 3]] self.outputs = {'Out': (seq_concat(self.inputs, level), out_lod)} diff --git a/python/paddle/fluid/tests/unittests/test_seq_conv.py b/python/paddle/fluid/tests/unittests/test_seq_conv.py index 51dbf1f61834ff0093d76ed546be27a585697d40..1a6e1aad799e77b8e746353bee93680691939d24 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_conv.py +++ b/python/paddle/fluid/tests/unittests/test_seq_conv.py @@ -26,9 +26,9 @@ class TestSeqProject(OpTest): if self.context_length == 1 \ and self.context_start == 0 \ and self.padding_trainable: - print "If context_start is 0 " \ + print("If context_start is 0 " \ "and context_length is 1," \ - " padding_trainable should be false." + " padding_trainable should be false.") return # one level, batch size @@ -75,35 +75,38 @@ class TestSeqProject(OpTest): pading_data = self.pad_data out = np.zeros((self.input_size[0], self.context_length * self.input_size[1])).astype('float32') - lod = lod[0] + offset = [0] + for seq_len in lod[0]: + offset.append(offset[-1] + seq_len) begin_pad = np.max([0, -self.context_start]) - for i in range(len(lod) - 1): + for i in range(len(offset) - 1): for j in range(self.context_length): - in_begin = lod[i] + self.context_start + j - in_end = lod[i + 1] + self.context_start + j - out_begin = lod[i] - out_end = lod[i + 1] - if in_begin < lod[i]: - pad_size = np.min([lod[i] - in_begin, lod[i + 1] - lod[i]]) + in_begin = offset[i] + self.context_start + j + in_end = offset[i + 1] + self.context_start + j + out_begin = offset[i] + out_end = offset[i + 1] + if in_begin < offset[i]: + pad_size = np.min( + [offset[i] - in_begin, offset[i + 1] - offset[i]]) if self.padding_trainable: sub_w = pading_data[j:j + pad_size, :] - out[lod[i]:lod[i] + pad_size, j * self.input_size[1]:( - j + 1) * self.input_size[1]] = sub_w - out_begin = lod[i] + pad_size - in_begin = lod[i] + out[offset[i]:offset[i] + pad_size, j * self.input_size[ + 1]:(j + 1) * self.input_size[1]] = sub_w + out_begin = offset[i] + pad_size + in_begin = offset[i] - if in_end > lod[i + 1]: + if in_end > offset[i + 1]: pad_size = np.min( - [in_end - lod[i + 1], lod[i + 1] - lod[i]]) + [in_end - offset[i + 1], offset[i + 1] - offset[i]]) if self.padding_trainable: sub_w = pading_data[begin_pad + self.context_start + j - pad_size:begin_pad + self.context_start + j, :] - out[lod[i + 1] - pad_size:lod[i + 1], j * self. + out[offset[i + 1] - pad_size:offset[i + 1], j * self. input_size[1]:(j + 1) * self.input_size[1]] = sub_w - in_end = lod[i + 1] - out_end = lod[i + 1] - pad_size + in_end = offset[i + 1] + out_end = offset[i + 1] - pad_size if in_end <= in_begin: continue @@ -175,7 +178,11 @@ class TestSeqProject(OpTest): self.context_stride = 1 self.input_size = [self.input_row, 23] - self.lod = [[0, 4, 5, 8, self.input_row]] + offset_lod = [[0, 4, 5, 8, self.input_row]] + self.lod = [[]] + # convert from offset-based lod to length-based lod + for i in range(len(offset_lod[0]) - 1): + self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) self.output_represention = 8 # output feature size @@ -188,7 +195,11 @@ class TestSeqProjectCase1(TestSeqProject): self.context_stride = 1 self.input_size = [self.input_row, 23] - self.lod = [[0, 4, 5, 8, self.input_row]] + offset_lod = [[0, 4, 5, 8, self.input_row]] + self.lod = [[]] + # convert from offset-based lod to length-based lod + for i in range(len(offset_lod[0]) - 1): + self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) self.output_represention = 8 # output feature size @@ -201,10 +212,14 @@ class TestSeqProjectCase2(TestSeqProject): self.context_stride = 1 self.input_size = [self.input_row, 23] - idx = range(self.input_size[0]) + idx = list(range(self.input_size[0])) del idx[0] - self.lod = [[0] + np.sort(random.sample(idx, 8)).tolist() + - [self.input_size[0]]] + offset_lod = [[0] + np.sort(random.sample(idx, 8)).tolist() + + [self.input_size[0]]] + self.lod = [[]] + # convert from offset-based lod to length-based lod + for i in range(len(offset_lod[0]) - 1): + self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) self.output_represention = 8 # output feature size diff --git a/python/paddle/fluid/tests/unittests/test_seq_pool.py b/python/paddle/fluid/tests/unittests/test_seq_pool.py index 2e48ef0e880839f6d5b4e515a174f427a35e7e6f..0b3659d7a67956f7546d368346bd102eeedf1d97 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_pool.py +++ b/python/paddle/fluid/tests/unittests/test_seq_pool.py @@ -18,26 +18,34 @@ from op_test import OpTest class TestSeqAvgPool(OpTest): + def convert_to_offset(self, lod): + offset = [[0] for i in lod] + for i, level in enumerate(lod): + for seq_len in level: + offset[i].append(offset[i][-1] + seq_len) + return offset + def set_data(self): self.op_type = 'sequence_pool' # one level, batch size is 4 x = np.random.uniform(0.1, 1, [11, 23]).astype('float32') - lod = [[0, 4, 5, 8, 11]] + lod = [[4, 1, 3, 3]] self.inputs = {'X': (x, lod)} + offset = self.convert_to_offset(lod) out = np.zeros((4, 23)).astype('float32') self.outputs = {'Out': out} - return x, lod, out + return x, offset, out - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "AVERAGE"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] + for i in range(len(offset[0]) - 1): + sub_x = x[offset[0][i]:offset[0][i + 1], :] out[i] = sub_x.mean(axis=0) def setUp(self): - x, lod, out = self.set_data() - self.compute(x, lod, out) + x, offset, out = self.set_data() + self.compute(x, offset, out) def test_check_output(self): self.check_output() @@ -50,10 +58,10 @@ class TestSeqAvgPool(OpTest): class TestSeqSumPool(TestSeqAvgPool): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "SUM"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] + for i in range(len(offset[0]) - 1): + sub_x = x[offset[0][i]:offset[0][i + 1], :] out[i] = sub_x.sum(axis=0) @@ -61,46 +69,47 @@ class TestSeqMaxPool(TestSeqAvgPool): def set_data(self): self.op_type = 'sequence_pool' x = np.random.uniform(0.1, 1, [13, 23]).astype('float32') - lod = [[0, 4, 5, 8, 13]] - for i in range(4): - l = lod[0][i + 1] - lod[0][i] - x[lod[0][i] + np.random.randint(l), :] += 2.0 + lod = [[4, 1, 3, 5]] + offset = self.convert_to_offset(lod) + for i in range(len(offset[0]) - 1): + l = offset[0][i + 1] - offset[0][i] + x[offset[0][i] + np.random.randint(l), :] += 2.0 self.inputs = {'X': (x, lod)} out = np.zeros((4, 23)).astype('float32') self.outputs = {'Out': out} - return x, lod, out + return x, offset, out - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "MAX"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] + for i in range(len(offset[0]) - 1): + sub_x = x[offset[0][i]:offset[0][i + 1], :] out[i] = np.amax(sub_x, axis=0) class TestSeqSqrtPool(TestSeqAvgPool): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "SQRT"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - len = lod[0][i + 1] - lod[0][i] - out[i] = sub_x.sum(axis=0) / np.sqrt(len) + for i in range(len(offset[0]) - 1): + sub_x = x[offset[0][i]:offset[0][i + 1], :] + seq_len = offset[0][i + 1] - offset[0][i] + out[i] = sub_x.sum(axis=0) / np.sqrt(seq_len) class TestSeqLastPool(TestSeqAvgPool): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "LAST"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] + for i in range(len(offset[0]) - 1): + sub_x = x[offset[0][i]:offset[0][i + 1], :] out[i] = sub_x[-1, :] class TestSeqFirstPool(TestSeqAvgPool): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "FIRST"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] + for i in range(len(offset[0]) - 1): + sub_x = x[offset[0][i]:offset[0][i + 1], :] out[i] = sub_x[0, :] @@ -109,35 +118,39 @@ class TestSeqAvgPool2D(TestSeqAvgPool): self.op_type = 'sequence_pool' # one level, batch size is 4 x = np.random.uniform(0.1, 1, [13, 3, 17]).astype('float32') - lod = [[0, 4, 5, 8, 13]] + lod = [[4, 1, 3, 5]] self.inputs = {'X': (x, lod)} + offset = self.convert_to_offset(lod) out = np.zeros((4, 3, 17)).astype('float32') self.outputs = {'Out': out} - return x, lod, out + return x, offset, out - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "AVERAGE"} - for i in range(4): - sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17)) + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 17)) out[i] = np.reshape(sub_x.mean(axis=0), (3, 17)) class TestSeqSumPool2D(TestSeqAvgPool2D): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "SUM"} - for i in range(4): - sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17)) + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 17)) out[i] = np.reshape(sub_x.sum(axis=0), (3, 17)) class TestSeqSqrtPool2D(TestSeqAvgPool2D): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "SQRT"} - for i in range(4): - sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17)) - len = lod[0][i + 1] - lod[0][i] - out[i] = np.reshape(sub_x.sum(axis=0) / np.sqrt(len), (3, 17)) + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 17)) + seq_len = offset[0][i + 1] - offset[0][i] + out[i] = np.reshape(sub_x.sum(axis=0) / np.sqrt(seq_len), (3, 17)) def test_check_grad(self): # Remove MaxIndex after check_grad is refined. @@ -150,36 +163,40 @@ class TestSeqMaxPool2D(TestSeqAvgPool2D): def set_data(self): self.op_type = 'sequence_pool' x = np.random.uniform(0.1, 1, [13, 3, 11]).astype('float32') - lod = [[0, 4, 5, 8, 13]] + lod = [[4, 1, 3, 5]] self.inputs = {'X': (x, lod)} - for i in range(4): - l = lod[0][i + 1] - lod[0][i] - x[lod[0][i] + np.random.randint(l), :] += 1.0 + offset = self.convert_to_offset(lod) + for i in range(len(offset[0]) - 1): + l = offset[0][i + 1] - offset[0][i] + x[offset[0][i] + np.random.randint(l), :] += 1.0 out = np.zeros((4, 3, 11)).astype('float32') self.outputs = {'Out': out} - return x, lod, out + return x, offset, out - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "MAX"} - for i in range(4): - sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 11)) + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 11)) out[i] = np.reshape(np.amax(sub_x, axis=0), (3, 11)) class TestSeqLastPool2D(TestSeqAvgPool2D): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "LAST"} - for i in range(4): - sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17)) + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 17)) out[i] = np.reshape(sub_x[-1, :], (3, 17)) class TestSeqFirstPool2D(TestSeqAvgPool2D): - def compute(self, x, lod, out): + def compute(self, x, offset, out): self.attrs = {'pooltype': "FIRST"} - for i in range(4): - sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17)) + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 17)) out[i] = np.reshape(sub_x[0, :], (3, 17)) diff --git a/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py b/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py index ebab77e8041d5ff1bd845fb121e5901116fd0254..8f0765277ae85af2b17ad96d4fd0c1148c393ff0 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py @@ -18,15 +18,17 @@ from op_test import OpTest def sequence_erase(in_seq, lod0, tokens): - new_lod0 = [0] + new_lod0 = [] out_seq = [] - for i in range(0, len(lod0) - 1): + offset = 0 + for i in range(0, len(lod0)): num_out = 0 - for dat in in_seq[lod0[i]:lod0[i + 1]]: + for dat in in_seq[offset:(offset + lod0[i])]: if dat not in tokens: out_seq.append(dat) num_out += 1 - new_lod0.append(new_lod0[-1] + num_out) + offset += lod0[i] + new_lod0.append(num_out) return np.array(out_seq).astype("int32"), new_lod0 @@ -34,7 +36,7 @@ class TestSequenceEraseOpInt32(OpTest): def setUp(self): self.op_type = "sequence_erase" in_seq = np.random.randint(0, 10, (30, 1)).astype("int32") - lod = [[0, 9, 13, 24, 30]] + lod = [[9, 4, 11, 6]] tokens = [2, 3, 5] out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens) self.attrs = {'tokens': tokens} @@ -49,7 +51,7 @@ class TestSequenceEraseOpInt64(OpTest): def setUp(self): self.op_type = "sequence_erase" in_seq = np.random.randint(0, 10, (30, 1)).astype("int64") - lod = [[0, 9, 13, 24, 30]] + lod = [[9, 4, 11, 6]] tokens = [2, 3, 5] out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens) self.attrs = {'tokens': tokens} @@ -64,7 +66,7 @@ class TestSequenceEraseOpEmpty(OpTest): def setUp(self): self.op_type = "sequence_erase" in_seq = np.random.randint(0, 10, (30, 1)).astype("int32") - lod = [[0, 9, 13, 24, 30]] + lod = [[9, 4, 11, 6]] tokens = [] out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens) self.attrs = {'tokens': tokens} diff --git a/python/paddle/fluid/tests/unittests/test_sequence_expand.py b/python/paddle/fluid/tests/unittests/test_sequence_expand.py index 4c8ec1426c6e103498af544ea5928ec630707d46..5ff0dab23e516ae8114b8264492fb2a9d5c0b3f8 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_expand.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_expand.py @@ -21,7 +21,7 @@ class TestSequenceExpand(OpTest): def set_data(self): x_data = np.random.uniform(0.1, 1, [3, 1]).astype('float32') y_data = np.random.uniform(0.1, 1, [8, 1]).astype('float32') - y_lod = [[0, 1, 4, 8]] + y_lod = [[1, 3, 4]] self.inputs = {'X': x_data, 'Y': (y_data, y_lod)} def compute(self): @@ -37,23 +37,27 @@ class TestSequenceExpand(OpTest): out = np.zeros(shape=((0, ) + x_data.shape[1:]), dtype=x_data.dtype) if x_lod is None: - x_idx = [i for i in xrange(x_data.shape[0] + 1)] + # x_idx = [i for i in xrange(x_data.shape[0] + 1)] + x_idx = [1] * x_data.shape[0] else: x_idx = x_lod[0] - out_lod = [[0]] + out_lod = [[]] + + offset = 0 + for i in range(len(y_lod[ref_level])): + repeat_num = y_lod[ref_level][i] + x_len = x_idx[i] - for i in xrange(1, len(y_lod[ref_level])): - repeat_num = y_lod[ref_level][i] - y_lod[ref_level][i - 1] - x_len = x_idx[i] - x_idx[i - 1] if repeat_num > 0: - x_sub = x_data[x_idx[i - 1]:x_idx[i], :] + x_sub = x_data[offset:(offset + x_len), :] stacked_x_sub = x_sub for r in range(repeat_num - 1): stacked_x_sub = np.vstack((stacked_x_sub, x_sub)) out = np.vstack((out, stacked_x_sub)) if x_lod is not None: - for j in xrange(repeat_num): - out_lod[0].append(out_lod[0][-1] + x_len) + for j in range(repeat_num): + out_lod[0].append(x_len) + offset += x_len if x_lod is None: self.outputs = {'Out': out} @@ -75,9 +79,9 @@ class TestSequenceExpand(OpTest): class TestSequenceExpandCase1(TestSequenceExpand): def set_data(self): x_data = np.random.uniform(0.1, 1, [5, 1]).astype('float32') - x_lod = [[0, 2, 5]] + x_lod = [[2, 3]] y_data = np.random.uniform(0.1, 1, [13, 1]).astype('float32') - y_lod = [[0, 2, 5], [0, 2, 4, 7, 10, 13]] + y_lod = [[2, 3], [2, 2, 3, 3, 3]] self.inputs = {'X': x_data, 'Y': (y_data, y_lod)} self.attrs = {'ref_level': 0} @@ -85,9 +89,9 @@ class TestSequenceExpandCase1(TestSequenceExpand): class TestSequenceExpandCase2(TestSequenceExpand): def set_data(self): x_data = np.random.uniform(0.1, 1, [1, 2, 2]).astype('float32') - x_lod = [[0, 1]] + x_lod = [[1]] y_data = np.random.uniform(0.1, 1, [2, 2, 2]).astype('float32') - y_lod = [[0, 2], [0, 2]] + y_lod = [[2], [1, 1]] self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} self.attrs = {'ref_level': 0} @@ -95,9 +99,9 @@ class TestSequenceExpandCase2(TestSequenceExpand): class TestSequenceExpandCase3(TestSequenceExpand): def set_data(self): x_data = np.random.uniform(0.1, 1, [4, 1]).astype('float32') - x_lod = [[0, 1, 2, 3, 4]] - y_data = np.random.uniform(0.1, 1, [6, 1]).astype('float32') - y_lod = [[0, 2, 4, 4, 6]] + x_lod = [[1, 1, 1, 1]] + y_data = np.random.uniform(0.1, 1, [8, 1]).astype('float32') + y_lod = [[2, 2, 2, 2]] self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} @@ -105,9 +109,9 @@ class TestSequenceExpandCase4(TestSequenceExpand): def set_data(self): data = np.random.uniform(0.1, 1, [5 * 2, 1]) x_data = np.array(data).reshape([5, 2]).astype('float32') - x_lod = [[0, 2, 5]] - y_data = np.random.uniform(0.1, 1, [3, 1]).astype('float32') - y_lod = [[0, 1, 3], [0, 1, 3]] + x_lod = [[2, 3]] + y_data = np.random.uniform(0.1, 1, [5, 1]).astype('float32') + y_lod = [[2], [2, 3]] self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} diff --git a/python/paddle/fluid/tests/unittests/test_sequence_reshape.py b/python/paddle/fluid/tests/unittests/test_sequence_reshape.py index efeab560392d8c03b1bb5db83f59c12d4fef64b0..39b02ecf6ddb40737c4c1737d652c1a1b744d923 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_reshape.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_reshape.py @@ -22,7 +22,7 @@ class TestSequenceReshape(OpTest): def setUp(self): self.op_type = 'sequence_reshape' dimension = 12 - x_lod = [[0, 4, 5, 8, 11]] + x_lod = [[4, 1, 3, 3]] x = np.random.uniform(0.1, 1, [11, 24]).astype('float32') self.inputs = {'X': (x, x_lod)} @@ -34,13 +34,13 @@ class TestSequenceReshape(OpTest): def compute_output(self, x, x_lod, dimension): x_width = x.shape[1] - out_lod = [[0]] - for i in xrange(len(x_lod[0]) - 1): - seq_len = x_lod[0][i + 1] - x_lod[0][i] + out_lod = [[]] + for i in range(len(x_lod[0])): + seq_len = x_lod[0][i] offset = (seq_len * x_width) / dimension assert int(offset) * dimension == seq_len * x_width - out_lod[0].append(out_lod[0][-1] + int(offset)) - out = np.zeros(shape=(out_lod[0][-1], dimension)).astype('float32') + out_lod[0].append(int(offset)) + out = np.zeros(shape=(sum(out_lod[0]), dimension)).astype('float32') out.ravel()[:] = x.ravel()[:] return out, out_lod @@ -55,7 +55,7 @@ class TestSequenceReshape_reduce(TestSequenceReshape): def setUp(self): self.op_type = 'sequence_reshape' dimension = 24 - x_lod = [[0, 4, 6, 8, 12]] + x_lod = [[4, 2, 2, 4]] x = np.random.uniform(0.1, 1, [12, 12]).astype('float32') self.inputs = {'X': (x, x_lod)} @@ -70,7 +70,7 @@ class TestSequenceReshape_same(TestSequenceReshape): def setUp(self): self.op_type = 'sequence_reshape' dimension = 12 - x_lod = [[0, 4, 6, 8, 12]] + x_lod = [[4, 2, 2, 4]] x = np.random.uniform(0.1, 1, [12, 12]).astype('float32') self.inputs = {'X': (x, x_lod)} diff --git a/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py b/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py index 660b4a171d09ddfc0e78b650a467db6b576c7ee3..313e485d1e3080f2c59c68256cbc5c81aa6558cd 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py @@ -29,20 +29,20 @@ class TestSequenceSliceOp(OpTest): self.inputs = {'X': (x, lod), 'Offset': offset, 'Length': length} outs = [] #np.zeros((100, 3, 2)).astype('float32') - out_lod = [[0]] - out_lod_offset = 0 + out_lod = [[]] + lod_offset = 0 for i in range(len(offset)): - sub_x = x[lod[0][i] + offset[i, 0]:lod[0][i] + offset[i, 0] + + sub_x = x[lod_offset + offset[i, 0]:lod_offset + offset[i, 0] + length[i, 0], :] - out_lod_offset = out_lod_offset + len(sub_x) outs.append(sub_x) - out_lod[0].append(out_lod_offset) + out_lod[0].append(len(sub_x)) + lod_offset += lod[0][i] outs = np.concatenate(outs, axis=0) self.outputs = {'Out': (outs, out_lod)} def init_test_case(self): self.x_dim = (100, 3, 2) - self.x_lod = [[0, 20, 40, 60, 80, 100]] + self.x_lod = [[20, 20, 20, 20, 20]] self.offset = [[1], [2], [3], [4], [5]] self.length = [[10], [8], [6], [4], [2]] diff --git a/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py b/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py index d6dc99bb3106feee33daa52bffb386f07cc16de5..c4fc8b74cf80c3596b0af9f7f0434864591195bd 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py @@ -26,15 +26,16 @@ class TestSequenceSoftmaxOp(OpTest): self.init_op_type() x = np.random.uniform(0.1, 1, (11, 1)).astype("float32") - lod = [[0, 4, 5, 8, 11]] + lod = [[4, 1, 3, 3]] out = np.zeros((11, 1)).astype("float32") - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - sub_x = sub_x.reshape(1, lod[0][i + 1] - lod[0][i]) + offset = 0 + for i in range(len(lod[0])): + sub_x = x[offset:offset + lod[0][i], :] + sub_x = sub_x.reshape(1, lod[0][i]) sub_out = stable_softmax(sub_x) - out[lod[0][i]:lod[0][i + 1], :] = sub_out.reshape( - lod[0][i + 1] - lod[0][i], 1) + out[offset:offset + lod[0][i], :] = sub_out.reshape(lod[0][i], 1) + offset += lod[0][i] self.inputs = {"X": (x, lod)} self.outputs = {"Out": out} @@ -60,6 +61,8 @@ class TestSequenceSoftmaxOp(OpTest): # ----------------cudnn Sequencesoftmax---------------- +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestSequenceSoftmaxCUDNNOp(TestSequenceSoftmaxOp): def init_op_type(self): self.use_cudnn = True diff --git a/python/paddle/fluid/tests/unittests/test_shape_op.py b/python/paddle/fluid/tests/unittests/test_shape_op.py new file mode 100644 index 0000000000000000000000000000000000000000..a62ee050075cb8c9f8817c142825a89c24bdfedf --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_shape_op.py @@ -0,0 +1,47 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestShapeOp(OpTest): + def setUp(self): + self.op_type = "shape" + self.config() + self.shape = [2, 3] + input = np.zeros(self.shape) + self.inputs = {'Input': input} + self.outputs = {'Out': np.array(self.shape)} + + def config(self): + self.shape = [2, 3] + + def test_check_output(self): + self.check_output() + + +class case1(TestShapeOp): + def config(self): + self.shape = [2] + + +class case2(TestShapeOp): + def config(self): + self.shape = [1, 2, 3] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py b/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py index 1d93230e7b74c5b6c00bbe125e3ae2d3a649b4b9..a994bf181a74ca71a970da0105fe767f82750a6c 100644 --- a/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py +++ b/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py @@ -21,6 +21,9 @@ from paddle.fluid.framework import default_main_program, switch_main_program from paddle.fluid.framework import Program import numpy as np +from paddle.fluid.layers.control_flow import shrink_memory +from paddle.fluid.layers.control_flow import lod_rank_table + class TestShrinkRNNMemoryBase(unittest.TestCase): def setUp(self): @@ -30,23 +33,23 @@ class TestShrinkRNNMemoryBase(unittest.TestCase): x.stop_gradient = False rank_table_tensor = layers.data( 'rank_table_tensor', shape=[1], dtype='float32', lod_level=1) - table = layers.lod_rank_table(x=rank_table_tensor) + table = lod_rank_table(x=rank_table_tensor) i = layers.zeros(dtype='int64', shape=[1]) - self.mem1 = layers.shrink_memory(x=x, i=i, table=table) + self.mem1 = shrink_memory(x=x, i=i, table=table) i = layers.increment(x=i) i.stop_gradient = True - self.mem2 = layers.shrink_memory(x=self.mem1, i=i, table=table) + self.mem2 = shrink_memory(x=self.mem1, i=i, table=table) i = layers.increment(x=i) i.stop_gradient = True - self.mem3 = layers.shrink_memory(x=self.mem2, i=i, table=table) + self.mem3 = shrink_memory(x=self.mem2, i=i, table=table) mem3_mean = layers.mean(self.mem3) append_backward(loss=mem3_mean) self.x_grad = self.main_program.global_block().var('x@GRAD') def sum_lodtensor(self, tensor): sum_res = 0.0 - for i in xrange(np.product(tensor.get_dims())): - sum_res += tensor.get_float_element(i) + for i in range(np.product(tensor.shape())): + sum_res += tensor._get_float_element(i) return sum_res @@ -54,12 +57,12 @@ class TestShrinkRNNMemoryReferLoD(TestShrinkRNNMemoryBase): def test_refer_lod(self): cpu = core.CPUPlace() x_tensor = core.LoDTensor() - x_tensor.set_lod([[0, 2, 5, 6]]) + x_tensor.set_recursive_sequence_lengths([[2, 3, 1]]) tensor_np = np.random.random(size=(6, 100)).astype('float32') x_tensor.set(tensor_np, cpu) rank_table_tensor = core.LoDTensor() - rank_table_tensor.set_lod([[0, 1, 3, 6]]) + rank_table_tensor.set_recursive_sequence_lengths([[1, 2, 3]]) rank_table_tensor.set(np.random.random(size=(6, 1)).astype('float32'), cpu) @@ -83,7 +86,7 @@ class TestShrinkRNNMemoryNoLoD(TestShrinkRNNMemoryBase): x_tensor.set(tensor_np, cpu) rank_table_tensor = core.LoDTensor() - rank_table_tensor.set_lod([[0, 1, 3, 6]]) + rank_table_tensor.set_recursive_sequence_lengths([[1, 2, 3]]) rank_table_tensor.set(np.random.random(size=(6, 1)).astype('float32'), cpu) diff --git a/python/paddle/fluid/tests/unittests/test_slice_op.py b/python/paddle/fluid/tests/unittests/test_slice_op.py new file mode 100644 index 0000000000000000000000000000000000000000..1a48bce3bb7c74551a365fd471f6869b128babac --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_slice_op.py @@ -0,0 +1,62 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestSliceOp(OpTest): + def setUp(self): + self.op_type = "slice" + self.config() + self.inputs = {'Input': self.input} + self.outputs = {'Out': self.out} + self.attrs = { + 'axes': self.axes, + 'starts': self.starts, + 'ends': self.ends + } + + def config(self): + self.input = np.random.random([3, 4, 5, 6]).astype("float32") + self.starts = [1, 0, 2] + self.ends = [3, 3, 4] + self.axes = [0, 1, 2] + self.out = self.input[1:3, 0:3, 2:4, :] + + def test_check_output(self): + self.check_output() + + +class TestCase1(TestSliceOp): + def config(self): + self.input = np.random.random([3, 4, 5, 6]).astype("float32") + self.starts = [-3, 0, 2] + self.ends = [3, 100, -1] + self.axes = [0, 1, 2] + self.out = self.input[-3:3, 0:100, 2:-1, :] + + +class TestCase2(TestSliceOp): + def config(self): + self.input = np.random.random([3, 4, 5, 6]).astype("float32") + self.starts = [-3, 0, 2] + self.ends = [3, 100, -1] + self.axes = [0, 1, 3] + self.out = self.input[-3:3, 0:100, :, 2:-1] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_slice_var.py b/python/paddle/fluid/tests/unittests/test_slice_var.py new file mode 100644 index 0000000000000000000000000000000000000000..82305b23a1a1e2cee8cef6b291d848581fe5b509 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_slice_var.py @@ -0,0 +1,64 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +import unittest +from paddle.fluid.transpiler.distribute_transpiler import slice_variable +import paddle.fluid as fluid +import paddle.fluid.core as core +import random + + +class TestSliceVar(unittest.TestCase): + def check_slice_output(self, shapes, expected_sizes, min_size): + var_list = [] + program = fluid.Program() + for shape in shapes: + var = program.global_block().create_var( + name=str(random.randint(10000, 99999)), + persistable=True, + # dtype=core.VarDesc.VarType.LOD_TENSOR, + shape=shape) + var_list.append(var) + blocks = slice_variable(var_list, 10, min_size) + all_sizes = [] + for s in expected_sizes: + for s2 in s: + all_sizes.append(s2) + for i, block_str in enumerate(blocks): + varname, block_id, size = block_str.split(":") + self.assertEqual(int(size), all_sizes[i]) + + def test_1k(self): + shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]] + expected_sizes = [ + [15], [1024], + [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784], + [2040, 2040, 2040, 2040], + [1150, 1150, 1150, 1150, 1150, 1150, 1100] + ] + + self.check_slice_output(shapes, expected_sizes, 1024) + + def test_check_output_8k(self): + shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10], + [6, 33, 33, 33]] + expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000], + [35937, 35937, 35937, 35937, 35937, 35937]] + + self.check_slice_output(shapes, expected_sizes, 8192) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_softmax_op.py b/python/paddle/fluid/tests/unittests/test_softmax_op.py index 279f3073f73d1c36f54bb901d92441a7403ac23f..70ad05597c4a160cf6a25aeb3c379320cef69c63 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_op.py @@ -26,15 +26,22 @@ def stable_softmax(x): class TestSoftmaxOp(OpTest): + def get_x_shape(self): + return [10, 10] + def setUp(self): self.op_type = "softmax" self.use_cudnn = False self.use_mkldnn = False self.dtype = np.float32 self.init_kernel_type() + self.shape = self.get_x_shape() + + x = np.random.uniform(0.1, 1, self.shape).astype(self.dtype) + out = np.apply_along_axis(stable_softmax, 1, + x.reshape([-1, self.shape[-1]])) + out = out.reshape(self.shape) - x = np.random.uniform(0.1, 1, [10, 10]).astype(self.dtype) - out = np.apply_along_axis(stable_softmax, 1, x) self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} self.attrs = { @@ -63,11 +70,27 @@ class TestSoftmaxOp(OpTest): self.check_grad(["X"], "Out", max_relative_error=0.01) +class TestSoftmaxOp2(TestSoftmaxOp): + def get_x_shape(self): + return [2, 3, 4, 5] + + +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestSoftmaxCUDNNOp(TestSoftmaxOp): def init_kernel_type(self): self.use_cudnn = True +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestSoftmaxCUDNNOp2(TestSoftmaxCUDNNOp): + def get_x_shape(self): + return [2, 3, 4, 5] + + +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestSoftmaxFP16Op(TestSoftmaxOp): def init_kernel_type(self): self.dtype = np.float16 @@ -79,6 +102,15 @@ class TestSoftmaxFP16Op(TestSoftmaxOp): self.check_output_with_place(place, atol=1e-3) +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestSoftmaxFP16Op2(TestSoftmaxFP16Op): + def get_x_shape(self): + return [2, 3, 4, 5] + + +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") class TestSoftmaxFP16CUDNNOp(TestSoftmaxOp): def init_kernel_type(self): self.use_cudnn = True @@ -91,10 +123,22 @@ class TestSoftmaxFP16CUDNNOp(TestSoftmaxOp): self.check_output_with_place(place, atol=1e-3) +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestSoftmaxFP16CUDNNOp2(TestSoftmaxFP16CUDNNOp): + def get_x_shape(self): + return [2, 3, 4, 5] + + class TestSoftmaxMKLDNNOp(TestSoftmaxOp): def init_kernel_type(self): self.use_mkldnn = True +class TestSoftmaxMKLDNNOp2(TestSoftmaxMKLDNNOp): + def get_x_shape(self): + return [2, 3, 4, 5] + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py b/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py index 02cc7da84918041c33bf5c8def46025bc87a2b9e..ea1146166d34a31efbd859318b411cea50895fe1 100644 --- a/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py @@ -19,6 +19,8 @@ import paddle.fluid.layers as layers from paddle.fluid.framework import Program, program_guard from paddle.fluid.executor import Executor from paddle.fluid.backward import append_backward +from paddle.fluid.layers.control_flow import split_lod_tensor +from paddle.fluid.layers.control_flow import merge_lod_tensor class TestCPULoDTensorArrayOps(unittest.TestCase): @@ -56,7 +58,7 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): def test_split_and_merge_lod_tensor_level_0(self): tensor = core.LoDTensor() tensor.set(np.arange(10).reshape(10, 1).astype('int32'), self.place()) - tensor.set_lod([[0, 3, 9, 10]]) + tensor.set_recursive_sequence_lengths([[3, 6, 1]]) mask_np = np.array([0, 1, 0]).astype('bool') mask_np = np.expand_dims(mask_np, axis=1) @@ -68,15 +70,15 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): expect_true_tensor = np.expand_dims(expect_true_tensor, axis=1) expect_true = core.LoDTensor() expect_true.set(expect_true_tensor, self.place()) - expect_true.set_lod([[0, 6]]) + expect_true.set_recursive_sequence_lengths([[6]]) expect_false_tensor = np.array([0, 1, 2, 9]).astype('int32') expect_false_tensor = np.expand_dims(expect_false_tensor, axis=1) - expect_false_lod = [[0, 3, 4]] + expect_false_lod = [[3, 1]] expect_false = core.LoDTensor() expect_false.set(expect_false_tensor, self.place()) - expect_false.set_lod(expect_false_lod) + expect_false.set_recursive_sequence_lengths(expect_false_lod) self.main( tensor=tensor, @@ -96,12 +98,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): y = layers.data(name='y', shape=[1]) y.persistable = True - out_true, out_false = layers.split_lod_tensor( - input=x, mask=y, level=level) + out_true, out_false = split_lod_tensor(input=x, mask=y, level=level) out_true.persistable = True out_false.persistable = True - out = layers.merge_lod_tensor( + out = merge_lod_tensor( in_true=out_true, in_false=out_false, mask=y, x=x, level=level) out.persistable = True @@ -126,7 +127,8 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): def check_tensor_same(self, actual, expect): self.assertTrue(np.allclose(np.array(actual), np.array(expect))) - self.assertEqual(actual.lod(), expect.lod()) + self.assertEqual(actual.recursive_sequence_lengths(), + expect.recursive_sequence_lengths()) class TestCPUSplitMergeLoDTensorGrad(unittest.TestCase): @@ -141,9 +143,8 @@ class TestCPUSplitMergeLoDTensorGrad(unittest.TestCase): level = 0 - out_true, out_false = layers.split_lod_tensor( - input=x, mask=y, level=level) - out = layers.merge_lod_tensor( + out_true, out_false = split_lod_tensor(input=x, mask=y, level=level) + out = merge_lod_tensor( in_true=out_true, in_false=out_false, mask=y, x=x, level=level) mean = layers.mean(out) @@ -151,7 +152,7 @@ class TestCPUSplitMergeLoDTensorGrad(unittest.TestCase): tensor = core.LoDTensor() tensor.set(np.arange(10).reshape(10, 1).astype('float32'), place) - tensor.set_lod([[0, 3, 9, 10]]) + tensor.set_recursive_sequence_lengths([[3, 6, 1]]) mask_np = np.array([0, 1, 0]).astype('bool') mask_np = np.expand_dims(mask_np, axis=1) diff --git a/python/paddle/fluid/tests/unittests/test_split_ids_op.py b/python/paddle/fluid/tests/unittests/test_split_ids_op.py index e9f0a06a56b42952800411d548bb3fc1732e031e..ca7861309839d183e18c168403881a0b1b5bf309 100644 --- a/python/paddle/fluid/tests/unittests/test_split_ids_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_ids_op.py @@ -15,6 +15,8 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core +from paddle.fluid.op import Operator class TestSplitIdsOp(OpTest): @@ -31,5 +33,55 @@ class TestSplitIdsOp(OpTest): self.check_output() +class TestSpliteIds(unittest.TestCase): + def get_places(self): + places = [core.CPUPlace()] + return places + + def test_check_output(self): + for place in self.get_places(): + self.check_with_place(place) + + def check_with_place(self, place): + scope = core.Scope() + rows = [0, 5, 7, 4, 9] + height = 20 + row_numel = 2 + + # initialize input variable X + x = scope.var('X').get_selected_rows() + x.set_rows(rows) + x.set_height(height) + np_array = np.ones((len(rows), row_numel)).astype("float32") + for i in range(len(rows)): + for j in range(row_numel): + np_array[i, j] = rows[i] + j + x_tensor = x.get_tensor() + x_tensor.set(np_array, place) + + outs_name = ["out%d" % i for i in xrange(3)] + outs = [ + scope.var(var_name).get_selected_rows() for var_name in outs_name + ] + + # expected output selected rows + expected_out_rows = [[0, 9], [7, 4], [5]] + + op = Operator("split_ids", Ids="X", Out=outs_name) + + for _ in range(3): + op.run(scope, place) + + for i in range(len(outs)): + expected_rows = expected_out_rows[i] + self.assertEqual(outs[i].rows(), expected_rows) + for j in range(len(expected_rows)): + row = expected_rows[j] + self.assertAlmostEqual( + float(row), np.array(outs[i].get_tensor())[j, 0]) + self.assertAlmostEqual( + float(row + 1), np.array(outs[i].get_tensor())[j, 1]) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_split_op.py b/python/paddle/fluid/tests/unittests/test_split_op.py index eb49a53e54f4bdb6bcd6cb1991423970f29997bb..6b67a52e81b978ed78c72629f9177759f8e2c4e1 100644 --- a/python/paddle/fluid/tests/unittests/test_split_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_op.py @@ -26,7 +26,7 @@ class TestSplitOp(OpTest): self.inputs = {'X': x} self.attrs = {'axis': axis, 'sections': [2, 1, 2]} self.outputs = {'Out': [('out%d' % i, out[i]) \ - for i in xrange(len(out))]} + for i in range(len(out))]} def _set_op_type(self): self.op_type = "split" diff --git a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py index 61040a39ced6dc57d05a10bf0605c80011db45c3..2b261820e04b08234477fc0a9adde95262f99bba 100644 --- a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py @@ -53,7 +53,7 @@ class TestSpliteSelectedRows(unittest.TestCase): height_sections = [5, 5, 5, 5, 3] # initialize output variables [out0, out1] - outs_name = ["out%d" % i for i in xrange(len(height_sections))] + outs_name = ["out%d" % i for i in range(len(height_sections))] outs = [ scope.var(var_name).get_selected_rows() for var_name in outs_name ] diff --git a/python/paddle/fluid/tests/unittests/test_split_var.py b/python/paddle/fluid/tests/unittests/test_split_var.py deleted file mode 100644 index 157def9b56e44092a86023035d1ab444c938aa07..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/unittests/test_split_var.py +++ /dev/null @@ -1,64 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import math -import unittest -from paddle.fluid.transpiler.distribute_transpiler import split_variable -import paddle.fluid as fluid -import paddle.fluid.core as core -import random - - -class TestSplitVar(unittest.TestCase): - def check_split_output(self, shapes, expected_sizes, min_size): - var_list = [] - program = fluid.Program() - for shape in shapes: - var = program.global_block().create_var( - name=str(random.randint(10000, 99999)), - persistable=True, - # dtype=core.VarDesc.VarType.LOD_TENSOR, - shape=shape) - var_list.append(var) - blocks = split_variable(var_list, 10, min_size) - all_sizes = [] - for s in expected_sizes: - for s2 in s: - all_sizes.append(s2) - for i, block_str in enumerate(blocks): - varname, block_id, size = block_str.split(":") - self.assertEqual(int(size), all_sizes[i]) - - def test_1k(self): - shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]] - expected_sizes = [ - [15], [1024], - [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784], - [2040, 2040, 2040, 2040], - [1150, 1150, 1150, 1150, 1150, 1150, 1100] - ] - - self.check_split_output(shapes, expected_sizes, 1024) - - def test_check_output_8k(self): - shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10], - [6, 33, 33, 33]] - expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000], - [35937, 35937, 35937, 35937, 35937, 35937]] - - self.check_split_output(shapes, expected_sizes, 8192) - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_spp_op.py b/python/paddle/fluid/tests/unittests/test_spp_op.py index f0ab5909df62835b252154709e5ff75ca38235c8..3cbfc2a703f1c4a24674d468cd1152bfa6eb8ad2 100644 --- a/python/paddle/fluid/tests/unittests/test_spp_op.py +++ b/python/paddle/fluid/tests/unittests/test_spp_op.py @@ -26,7 +26,7 @@ class TestSppOp(OpTest): input = np.random.random(self.shape).astype("float32") nsize, csize, hsize, wsize = input.shape out_level_flatten = [] - for i in xrange(self.pyramid_height): + for i in range(self.pyramid_height): bins = np.power(2, i) kernel_size = [0, 0] padding = [0, 0] diff --git a/python/paddle/fluid/tests/unittests/test_squeeze_op.py b/python/paddle/fluid/tests/unittests/test_squeeze_op.py new file mode 100644 index 0000000000000000000000000000000000000000..bca6af2fd5dfadbc48cf1a76cfa6ffd4f1fdfdef --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_squeeze_op.py @@ -0,0 +1,114 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +from op_test import OpTest + + +# Correct: General. +class TestSqueezeOp(OpTest): + def setUp(self): + self.op_type = "squeeze" + self.init_test_case() + self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")} + self.init_attrs() + self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + def init_test_case(self): + self.ori_shape = (1, 3, 1, 5) + self.axes = (0, 2) + self.new_shape = (3, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": False} + + +# Correct: There is mins axis. +class TestSqueezeOp1(TestSqueezeOp): + def init_test_case(self): + self.ori_shape = (1, 3, 1, 5) + self.axes = (0, -2) + self.new_shape = (3, 5) + + +# Correct: No axes input. +class TestSqueezeOp2(TestSqueezeOp): + def init_test_case(self): + self.ori_shape = (1, 3, 1, 5) + self.axes = () + self.new_shape = (3, 5) + + +# Correct: Just part of axes be squeezed. +class TestSqueezeOp3(TestSqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 1, 5, 1, 4, 1) + self.axes = (1, -1) + self.new_shape = (3, 5, 1, 4) + + +# Correct: Inplace. +class TestSqueezeOpInplace1(TestSqueezeOp): + def init_test_case(self): + self.ori_shape = (1, 3, 1, 5) + self.axes = (0, 2) + self.new_shape = (3, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +# Correct: Inplace. There is mins axis. +class TestSqueezeOpInplace2(TestSqueezeOp): + def inti_test_case(self): + self.ori_shape = (1, 3, 1, 5) + self.axes = (0, -2) + self.new_shape = (3, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +# Correct: Inplace. No axes input. +class TestSqueezeOpInplace3(TestSqueezeOp): + def init_test_case(self): + self.ori_shape = (1, 3, 1, 5) + self.axes = () + self.new_shape = (3, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +# Correct: Inpalce. Just part of axes be squeezed. +class TestSqueezeOpInplace4(TestSqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 1, 5, 1, 4, 1) + self.axes = (1, -1) + self.new_shape = (3, 5, 1, 4) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py new file mode 100644 index 0000000000000000000000000000000000000000..7956897d68a3fb49d62ba696d0b6400b4f909989 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py @@ -0,0 +1,26 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +from test_sum_op import TestSumOp + + +class TestMKLDNN(TestSumOp): + def init_kernel_type(self): + self.use_mkldnn = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sum_op.py b/python/paddle/fluid/tests/unittests/test_sum_op.py index 2faf5b10647a1fa1d44e4847f017db177ee8808a..1d90414e137a70e6265042e24e106fe565802778 100644 --- a/python/paddle/fluid/tests/unittests/test_sum_op.py +++ b/python/paddle/fluid/tests/unittests/test_sum_op.py @@ -20,12 +20,15 @@ from op_test import OpTest class TestSumOp(OpTest): def setUp(self): self.op_type = "sum" + self.use_mkldnn = False + self.init_kernel_type() x0 = np.random.random((3, 4)).astype('float32') x1 = np.random.random((3, 4)).astype('float32') x2 = np.random.random((3, 4)).astype('float32') self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} y = x0 + x1 + x2 self.outputs = {'Out': y} + self.attrs = {'use_mkldnn': self.use_mkldnn} def test_check_output(self): self.check_output() @@ -33,6 +36,9 @@ class TestSumOp(OpTest): def test_check_grad(self): self.check_grad(['x0'], 'Out') + def init_kernel_type(self): + pass + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_target_assign_op.py b/python/paddle/fluid/tests/unittests/test_target_assign_op.py index ccb41e56c5555b8c79674449c9139ada0bc47aac..bd208897520122b6a5dcf71da325b1b9dba632f6 100644 --- a/python/paddle/fluid/tests/unittests/test_target_assign_op.py +++ b/python/paddle/fluid/tests/unittests/test_target_assign_op.py @@ -22,22 +22,23 @@ def gen_match_and_neg_indices(num_prior, gt_lod, neg_lod): if len(gt_lod) != len(neg_lod): raise AssertionError("The input arguments are illegal.") - batch_size = len(gt_lod) - 1 + batch_size = len(gt_lod) match_indices = -1 * np.ones((batch_size, num_prior)).astype('int32') - neg_indices = np.zeros((neg_lod[-1], 1)).astype('int32') + neg_indices = np.zeros((sum(neg_lod), 1)).astype('int32') + offset = 0 for n in range(batch_size): - gt_num = gt_lod[n + 1] - gt_lod[n] + gt_num = gt_lod[n] ids = random.sample([i for i in range(num_prior)], gt_num) match_indices[n, ids] = [i for i in range(gt_num)] ret_ids = set([i for i in range(num_prior)]) - set(ids) - s = neg_lod[n] - e = neg_lod[n + 1] - l = e - s + l = neg_lod[n] neg_ids = random.sample(ret_ids, l) - neg_indices[s:e, :] = np.array(neg_ids).astype('int32').reshape(l, 1) + neg_indices[offset:offset + neg_lod[n], :] = np.array(neg_ids).astype( + 'int32').reshape(l, 1) + offset += neg_lod[n] return match_indices, neg_indices @@ -56,24 +57,28 @@ def target_assign(encoded_box, gt_label, match_indices, neg_indices, gt_lod, # init weight for target label trg_label_wt = np.zeros((batch_size, num_prior, 1)).astype('float32') + gt_offset = 0 + neg_offset = 0 for i in range(batch_size): cur_indices = match_indices[i] col_ids = np.where(cur_indices > -1) col_val = cur_indices[col_ids] - gt_start = gt_lod[i] # target bbox - for v, c in zip(col_val + gt_start, col_ids[0].tolist()): + for v, c in zip(col_val + gt_offset, col_ids[0].tolist()): trg_box[i][c][:] = encoded_box[v][c][:] # weight for target bbox trg_box_wt[i][col_ids] = 1.0 - trg_label[i][col_ids] = gt_label[col_val + gt_start] + trg_label[i][col_ids] = gt_label[col_val + gt_offset] trg_label_wt[i][col_ids] = 1.0 # set target label weight to 1.0 for the negative samples if neg_indices is not None: - neg_ids = neg_indices[neg_lod[i]:neg_lod[i + 1]] + neg_ids = neg_indices[neg_offset:neg_offset + neg_lod[i]] trg_label_wt[i][neg_ids] = 1.0 + # update offset + gt_offset += gt_lod[i] + neg_offset += neg_lod[i] return trg_box, trg_box_wt, trg_label, trg_label_wt @@ -83,11 +88,11 @@ class TestTargetAssginFloatType(OpTest): self.op_type = "target_assign" num_prior = 120 num_class = 21 - gt_lod = [0, 5, 11, 23] - neg_lod = [0, 4, 7, 13] + gt_lod = [5, 6, 12] + neg_lod = [4, 3, 6] mismatch_value = 0 - batch_size = len(gt_lod) - 1 - num_gt = gt_lod[-1] + batch_size = len(gt_lod) + num_gt = sum(gt_lod) encoded_box = np.random.random((num_gt, num_prior, 4)).astype('float32') gt_label = np.random.randint( @@ -121,11 +126,11 @@ class TestTargetAssginIntType(OpTest): self.op_type = "target_assign" num_prior = 120 num_class = 21 - gt_lod = [0, 5, 11, 23] - neg_lod = [0, 4, 7, 13] + gt_lod = [5, 6, 12] + neg_lod = [4, 3, 6] mismatch_value = 0 - batch_size = len(gt_lod) - 1 - num_gt = gt_lod[-1] + batch_size = len(gt_lod) + num_gt = sum(gt_lod) encoded_box = np.random.random((num_gt, num_prior, 4)).astype('float32') gt_label = np.random.randint( diff --git a/python/paddle/fluid/tests/unittests/test_tensor.py b/python/paddle/fluid/tests/unittests/test_tensor.py index 379081c3287ce81dbf2bd7307cb5eac2620b13db..5ccc876ae8e6e20f76c77c1892f4de59d72bffc8 100644 --- a/python/paddle/fluid/tests/unittests/test_tensor.py +++ b/python/paddle/fluid/tests/unittests/test_tensor.py @@ -25,8 +25,8 @@ class TestTensor(unittest.TestCase): tensor = var.get_tensor() - tensor.set_dims([1000, 784]) - tensor.alloc_int(place) + tensor._set_dims([1000, 784]) + tensor._alloc_int(place) tensor_array = numpy.array(tensor) self.assertEqual((1000, 784), tensor_array.shape) tensor_array[3, 9] = 1 @@ -44,8 +44,8 @@ class TestTensor(unittest.TestCase): tensor = var.get_tensor() - tensor.set_dims([1000, 784]) - tensor.alloc_float(place) + tensor._set_dims([1000, 784]) + tensor._alloc_float(place) tensor_array = numpy.array(tensor) self.assertEqual((1000, 784), tensor_array.shape) @@ -63,21 +63,20 @@ class TestTensor(unittest.TestCase): var_lod = scope.var("test_lod_tensor") lod_tensor = var_lod.get_tensor() - lod_tensor.set_dims([4, 4, 6]) - lod_tensor.alloc_int(place) + lod_tensor._set_dims([4, 4, 6]) + lod_tensor._alloc_int(place) array = numpy.array(lod_tensor) array[0, 0, 0] = 3 array[3, 3, 5] = 10 lod_tensor.set(array, place) - lod_tensor.set_lod([[0, 2, 4]]) + lod_tensor.set_recursive_sequence_lengths([[2, 2]]) lod_v = numpy.array(lod_tensor) self.assertTrue(numpy.alltrue(array == lod_v)) - lod = lod_tensor.lod() - self.assertEqual(0, lod[0][0]) + lod = lod_tensor.recursive_sequence_lengths() + self.assertEqual(2, lod[0][0]) self.assertEqual(2, lod[0][1]) - self.assertEqual(4, lod[0][2]) def test_float_lod_tensor(self): place = core.CPUPlace() @@ -85,8 +84,8 @@ class TestTensor(unittest.TestCase): var_lod = scope.var("test_lod_tensor") lod_tensor = var_lod.get_tensor() - lod_tensor.set_dims([5, 2, 3, 4]) - lod_tensor.alloc_float(place) + lod_tensor._set_dims([5, 2, 3, 4]) + lod_tensor._alloc_float(place) tensor_array = numpy.array(lod_tensor) self.assertEqual((5, 2, 3, 4), tensor_array.shape) @@ -97,22 +96,21 @@ class TestTensor(unittest.TestCase): lod_v = numpy.array(lod_tensor) self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0]) self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1]) - self.assertEqual(len(lod_tensor.lod()), 0) + self.assertEqual(len(lod_tensor.recursive_sequence_lengths()), 0) - lod_py = [[0, 2, 5], [0, 2, 4, 5]] - lod_tensor.set_lod(lod_py) - lod = lod_tensor.lod() + lod_py = [[2, 1], [1, 2, 2]] + lod_tensor.set_recursive_sequence_lengths(lod_py) + lod = lod_tensor.recursive_sequence_lengths() self.assertListEqual(lod_py, lod) def test_lod_tensor_init(self): - scope = core.Scope() place = core.CPUPlace() - lod_py = [[0, 2, 5], [0, 2, 4, 5]] + lod_py = [[2, 1], [1, 2, 2]] lod_tensor = core.LoDTensor() - lod_tensor.set_dims([5, 2, 3, 4]) - lod_tensor.set_lod(lod_py) - lod_tensor.alloc_float(place) + lod_tensor._set_dims([5, 2, 3, 4]) + lod_tensor.set_recursive_sequence_lengths(lod_py) + lod_tensor._alloc_float(place) tensor_array = numpy.array(lod_tensor) tensor_array[0, 0, 0, 0] = 1.0 tensor_array[0, 0, 0, 1] = 2.0 @@ -121,18 +119,18 @@ class TestTensor(unittest.TestCase): lod_v = numpy.array(lod_tensor) self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0]) self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1]) - self.assertListEqual(lod_py, lod_tensor.lod()) + self.assertListEqual(lod_py, lod_tensor.recursive_sequence_lengths()) def test_lod_tensor_gpu_init(self): if not core.is_compiled_with_cuda(): return place = core.CUDAPlace(0) - lod_py = [[0, 2, 5], [0, 2, 4, 5]] + lod_py = [[2, 1], [1, 2, 2]] lod_tensor = core.LoDTensor() - lod_tensor.set_dims([5, 2, 3, 4]) - lod_tensor.set_lod(lod_py) - lod_tensor.alloc_float(place) + lod_tensor._set_dims([5, 2, 3, 4]) + lod_tensor.set_recursive_sequence_lengths(lod_py) + lod_tensor._alloc_float(place) tensor_array = numpy.array(lod_tensor) tensor_array[0, 0, 0, 0] = 1.0 tensor_array[0, 0, 0, 1] = 2.0 @@ -141,7 +139,7 @@ class TestTensor(unittest.TestCase): lod_v = numpy.array(lod_tensor) self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0]) self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1]) - self.assertListEqual(lod_py, lod_tensor.lod()) + self.assertListEqual(lod_py, lod_tensor.recursive_sequence_lengths()) def test_empty_tensor(self): place = core.CPUPlace() @@ -150,15 +148,15 @@ class TestTensor(unittest.TestCase): tensor = var.get_tensor() - tensor.set_dims([0, 1]) - tensor.alloc_float(place) + tensor._set_dims([0, 1]) + tensor._alloc_float(place) tensor_array = numpy.array(tensor) self.assertEqual((0, 1), tensor_array.shape) if core.is_compiled_with_cuda(): gpu_place = core.CUDAPlace(0) - tensor.alloc_float(gpu_place) + tensor._alloc_float(gpu_place) tensor_array = numpy.array(tensor) self.assertEqual((0, 1), tensor_array.shape) diff --git a/python/paddle/fluid/tests/unittests/test_top_k_op.py b/python/paddle/fluid/tests/unittests/test_top_k_op.py index cc2fcc5ec0a076679c7dd85a7e8f8da6a170172b..cbc3da550306b9febe8a8fd22e7f71efa572a3d0 100644 --- a/python/paddle/fluid/tests/unittests/test_top_k_op.py +++ b/python/paddle/fluid/tests/unittests/test_top_k_op.py @@ -28,7 +28,7 @@ class TestTopkOp(OpTest): self.inputs = {'X': input} self.attrs = {'k': k} - for rowid in xrange(32): + for rowid in range(32): row = input[rowid] output[rowid] = np.sort(row)[-k:] indices[rowid] = row.argsort()[-k:] @@ -52,7 +52,7 @@ class TestTopkOp3d(OpTest): self.inputs = {'X': input_flat_2d} self.attrs = {'k': k} - for rowid in xrange(64): + for rowid in range(64): row = input_flat_2d[rowid] output[rowid] = np.sort(row)[-k:] indices[rowid] = row.argsort()[-k:] diff --git a/python/paddle/fluid/tests/unittests/test_unpool_op.py b/python/paddle/fluid/tests/unittests/test_unpool_op.py index a97d6dfdda9b79eed3be6302fb2b1c3810f189dc..ecce4cdde2d648fe7d65427e34c77f5f9ad61417 100644 --- a/python/paddle/fluid/tests/unittests/test_unpool_op.py +++ b/python/paddle/fluid/tests/unittests/test_unpool_op.py @@ -22,10 +22,10 @@ def unpool2dmax_forward_naive(input, indices, ksize, strides, paddings): out_hsize = (s2 - 1) * strides[0] - 2 * paddings[0] + ksize[0] out_wsize = (s2 - 1) * strides[1] - 2 * paddings[1] + ksize[1] out = np.zeros((s0, s1, out_hsize, out_wsize)) - for nidx in xrange(s0): - for cidx in xrange(s1): - for h in xrange(s2): - for w in xrange(s3): + for nidx in range(s0): + for cidx in range(s1): + for h in range(s2): + for w in range(s3): index = indices[nidx, cidx, h, w] hidx = (index - index % out_wsize) / out_wsize widx = index % out_wsize @@ -47,16 +47,16 @@ class TestUnpoolOp(OpTest): self.strides[1] + 1 input = np.zeros((nsize, csize, hsize_out, wsize_out)) indices = np.zeros((nsize, csize, hsize_out, wsize_out)) - for i in xrange(hsize_out): - for j in xrange(wsize_out): + for i in range(hsize_out): + for j in range(wsize_out): r_start = np.max((i * self.strides[0] - self.paddings[0], 0)) r_end = np.min((i * self.strides[0] + self.ksize[0] - \ self.paddings[0], hsize)) c_start = np.max((j * self.strides[1] - self.paddings[1], 0)) c_end = np.min((j * self.strides[1] + self.ksize[1] - \ self.paddings[1], wsize)) - for nidx in xrange(nsize): - for cidx in xrange(csize): + for nidx in range(nsize): + for cidx in range(csize): x_masked = pre_input[nidx, cidx, r_start:r_end, \ c_start:c_end] input[nidx, cidx, i, j] = x_masked.max() diff --git a/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py b/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py new file mode 100644 index 0000000000000000000000000000000000000000..7a4aa0a40b5eb494f6027e800ca6b466bbe1c302 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py @@ -0,0 +1,111 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +from op_test import OpTest + + +# Correct: General. +class TestUnsqueezeOp(OpTest): + def setUp(self): + self.init_test_case() + self.op_type = "unsqueeze" + self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")} + self.init_attrs() + self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + def init_test_case(self): + self.ori_shape = (3, 5) + self.axes = (1, 2) + self.new_shape = (3, 1, 1, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": False} + + +# Correct: Single input index. +class TestUnsqueezeOp1(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 5) + self.axes = (-1, ) + self.new_shape = (3, 5, 1) + + +# Correct: Mixed input axis. +class TestUnsqueezeOp2(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 5) + self.axes = (0, -1) + self.new_shape = (1, 3, 5, 1) + + +# Correct: There is duplicated axis. +class TestUnsqueezeOp3(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 2, 5) + self.axes = (0, 3, 3) + self.new_shape = (1, 3, 2, 1, 1, 5) + + +# Correct: Reversed axes. +class TestUnsqueezeOp4(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 2, 5) + self.axes = (3, 1, 1) + self.new_shape = (3, 1, 1, 2, 5, 1) + + +# Correct: Inplace. +class TestUnsqueezeOpInplace1(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 5) + self.axes = (0, 2) + self.new_shape = (1, 3, 1, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +# Correct: Inplace. There is mins index. +class TestUnsqueezeOpInplace2(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 5) + self.axes = (0, -2) + self.new_shape = (1, 3, 1, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +# Correct: Inplace. There is duplicated axis. +class TestUnsqueezeOpInplace3(TestUnsqueezeOp): + def init_test_case(self): + self.ori_shape = (3, 2, 5) + self.axes = (0, 3, 3) + self.new_shape = (1, 3, 2, 1, 1, 5) + + def init_attrs(self): + self.attrs = {"axes": self.axes, "inplace": True} + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_version.py b/python/paddle/fluid/tests/unittests/test_version.py new file mode 100644 index 0000000000000000000000000000000000000000..a09c8a759b9461edcf7d5ddbd62d74408d5f292e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_version.py @@ -0,0 +1,48 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import re + +import paddle.version as fluid_version + + +class VersionTest(unittest.TestCase): + def setUp(self): + self._major_regex = "[0-9]+" + self._minor_regex = "[0-9]+" + self._patch_regex = "[0-9]+(\\.(a|b|rc)\\.[0-9]+)?" + self._rc_regex = "[0-9]+" + self._version_regex = "[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?" + self._commit_regex = "[0-9a-f]{5,49}" + + def test_check_output(self): + # check commit format + self.assertTrue(re.match(self._commit_regex, fluid_version.commit)) + self.assertTrue(isinstance(fluid_version.istaged, bool)) + + # check version format + if fluid_version.istaged: + self.assertEqual(fluid_version.major, 0) + self.assertEqual(fluid_version.minor, 0) + self.assertEqual(fluid_version.patch, "0") + self.assertEqual(fluid_version.rc, 0) + self.assertEqual(fluid_version.full_version, "0.0.0") + else: + self.assertTrue(re.match(self._major_regex, fluid_version.major)) + self.assertTrue(re.match(self._minor_regex, fluid_version.minor)) + self.assertTrue(re.match(self._patch_regex, fluid_version.patch)) + self.assertTrue(re.match(self._rc_regex, fluid_version.rc)) + self.assertTrue( + re.match(self._version_regex, fluid_version.full_version)) diff --git a/python/paddle/fluid/tests/unittests/test_warpctc_op.py b/python/paddle/fluid/tests/unittests/test_warpctc_op.py index ac638f7836f8205f80e31cfd5eb8892b2c7aee08..9f1aaee472f918da7deb8816a0a4654dafe74a30 100644 --- a/python/paddle/fluid/tests/unittests/test_warpctc_op.py +++ b/python/paddle/fluid/tests/unittests/test_warpctc_op.py @@ -34,8 +34,8 @@ class CTCForward(object): self.level = 0 self.num_classes = softmax.shape[1] - self.batch_size = len(softmax_lod[self.level]) - 1 - assert self.batch_size == len(labels_lod[self.level]) - 1 + self.batch_size = len(softmax_lod[self.level]) + assert self.batch_size == len(labels_lod[self.level]) self.loss = np.zeros([self.batch_size, 1], dtype="float32") self.gradient = np.zeros(self.softmax.shape, dtype="float32") @@ -156,16 +156,20 @@ class CTCForward(object): return -log_prob def forward(self): + softmax_offset = 0 + labels_offset = 0 for i in range(self.batch_size): - softmax_start_i = self.softmax_lod[self.level][i] - softmax_end_i = self.softmax_lod[self.level][i + 1] - labels_start_i = self.labels_lod[self.level][i] - labels_end_i = self.labels_lod[self.level][i + 1] + softmax_start_i = softmax_offset + softmax_end_i = softmax_offset + self.softmax_lod[self.level][i] + labels_start_i = labels_offset + labels_end_i = labels_offset + self.labels_lod[self.level][i] softmax_a_sequence = self.softmax[softmax_start_i:softmax_end_i, :] labels_a_sequence = self.labels[labels_start_i:labels_end_i, :] self.loss[i] = self.forward_a_sequence(softmax_a_sequence, labels_a_sequence) + softmax_offset += self.softmax_lod[self.level][i] + labels_offset += self.labels_lod[self.level][i] return self.loss @@ -173,8 +177,8 @@ class TestWarpCTCOp(OpTest): def config(self): self.batch_size = 4 self.num_classes = 8 - self.logits_lod = [[0, 4, 5, 8, 11]] - self.labels_lod = [[0, 3, 4, 8, 12]] + self.logits_lod = [[4, 1, 3, 3]] + self.labels_lod = [[3, 1, 4, 4]] self.blank = self.num_classes - 1 self.norm_by_times = False @@ -184,11 +188,13 @@ class TestWarpCTCOp(OpTest): logits = np.random.uniform( 0.1, 1.0, - [self.logits_lod[0][-1], self.num_classes]).astype("float32") + [sum(self.logits_lod[0]), self.num_classes]).astype("float32") softmax = np.apply_along_axis(stable_softmax, 1, logits) # labels should not be blank labels = np.random.randint( - 0, self.num_classes - 1, [self.labels_lod[0][-1], 1], dtype="int32") + 0, + self.num_classes - 1, [sum(self.labels_lod[0]), 1], + dtype="int32") ctc = CTCForward(softmax, self.logits_lod, labels, self.labels_lod, self.blank, self.norm_by_times) @@ -196,9 +202,8 @@ class TestWarpCTCOp(OpTest): max_sequence_length = 0 for i in range(self.batch_size): - max_sequence_length = max( - max_sequence_length, - self.logits_lod[0][i + 1] - self.logits_lod[0][i]) + max_sequence_length = max(max_sequence_length, + self.logits_lod[0][i]) self.gradient = np.zeros( [max_sequence_length, self.batch_size, self.num_classes], dtype="float32") @@ -222,8 +227,8 @@ class TestWarpCTCOpCase1(TestWarpCTCOp): def config(self): self.batch_size = 4 self.num_classes = CUDA_BLOCK_SIZE + 2 - self.logits_lod = [[0, 4, 5, 8, 11]] - self.labels_lod = [[0, 3, 4, 8, 12]] + self.logits_lod = [[4, 1, 3, 3]] + self.labels_lod = [[3, 1, 4, 4]] self.blank = 0 self.norm_by_times = False diff --git a/python/paddle/fluid/tests/unittests/test_weight_normalization.py b/python/paddle/fluid/tests/unittests/test_weight_normalization.py index 2adf917bc5d3bb35842a817c57a983627b759f22..436f9b9f86fb86270e47c8e30c5c0701787ca0f1 100644 --- a/python/paddle/fluid/tests/unittests/test_weight_normalization.py +++ b/python/paddle/fluid/tests/unittests/test_weight_normalization.py @@ -76,11 +76,11 @@ class TestWeightNormalization(unittest.TestCase): lod_level_i = numpy.random.randint( low=1, high=5, - size=self.batch_size if i == 0 else lod_level_i[-1]) - lod_level_i = [0] + numpy.cumsum(lod_level_i).tolist() + size=self.batch_size + if i == 0 else sum(lod_level_i)).tolist() data_lod.append(lod_level_i) data_value = numpy.random.random( - size=[data_lod[-1][-1] if data_lod else self.batch_size + size=[sum(data_lod[-1]) if data_lod else self.batch_size ] + data_shape).astype('float32') self.data[data_name] = (data_value, data_lod) @@ -90,7 +90,7 @@ class TestWeightNormalization(unittest.TestCase): tensor = fluid.Tensor() tensor.set(self.data[desc[0]][0], place) if self.data[desc[0]][1]: - tensor.set_lod(self.data[desc[0]][1]) + tensor.set_recursive_sequence_lengths(self.data[desc[0]][1]) self.inputs[desc[0]] = tensor def weight_normalize(self): diff --git a/python/paddle/fluid/tests/unittests/test_while_op.py b/python/paddle/fluid/tests/unittests/test_while_op.py index fe8808bc044684c96fb3382836be32dac1d241f3..790e6afe5f02236b00d9c67b7b25a881e07abace 100644 --- a/python/paddle/fluid/tests/unittests/test_while_op.py +++ b/python/paddle/fluid/tests/unittests/test_while_op.py @@ -66,7 +66,7 @@ class TestWhileOp(unittest.TestCase): exe = Executor(cpu) d = [] - for i in xrange(3): + for i in range(3): d.append(numpy.random.random(size=[10]).astype('float32')) outs = exe.run(feed={'d0': d[0], diff --git a/python/paddle/fluid/tests/unittests/testsuite.py b/python/paddle/fluid/tests/unittests/testsuite.py new file mode 100644 index 0000000000000000000000000000000000000000..c6e176ca31c57e623addd9594be81c0abdce489b --- /dev/null +++ b/python/paddle/fluid/tests/unittests/testsuite.py @@ -0,0 +1,190 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np + +import paddle.fluid.core as core +from paddle.fluid.op import Operator + + +def create_op(scope, op_type, inputs, outputs, attrs): + kwargs = dict() + + op_maker = core.op_proto_and_checker_maker + op_role_attr_name = op_maker.kOpRoleAttrName() + + if op_role_attr_name not in attrs: + attrs[op_role_attr_name] = int(op_maker.OpRole.Forward) + + def __create_var__(name, var_name): + scope.var(var_name).get_tensor() + kwargs[name].append(var_name) + + for in_name, in_dup in Operator.get_op_inputs(op_type): + if in_name in inputs: + kwargs[in_name] = [] + if in_dup: + sub_in = inputs[in_name] + for item in sub_in: + sub_in_name, _ = item[0], item[1] + __create_var__(in_name, sub_in_name) + else: + __create_var__(in_name, in_name) + + for out_name, out_dup in Operator.get_op_outputs(op_type): + if out_name in outputs: + kwargs[out_name] = [] + if out_dup: + sub_out = outputs[out_name] + for item in sub_out: + sub_out_name, _ = item[0], item[1] + __create_var__(out_name, sub_out_name) + else: + __create_var__(out_name, out_name) + + for attr_name in Operator.get_op_attr_names(op_type): + if attr_name in attrs: + kwargs[attr_name] = attrs[attr_name] + + return Operator(op_type, **kwargs) + + +def set_input(scope, op, inputs, place): + def np_value_to_fluid_value(input): + if input.dtype == np.float16: + input = input.view(np.uint16) + return input + + def __set_input__(var_name, var): + if isinstance(var, tuple) or isinstance(var, np.ndarray): + tensor = scope.find_var(var_name).get_tensor() + if isinstance(var, tuple): + tensor.set_recursive_sequence_lengths(var[1]) + var = var[0] + tensor._set_dims(var.shape) + tensor.set(np_value_to_fluid_value(var), place) + elif isinstance(var, float): + scope.find_var(var_name).set_float(var) + elif isinstance(var, int): + scope.find_var(var_name).set_int(var) + + for in_name, in_dup in Operator.get_op_inputs(op.type()): + if in_name in inputs: + if in_dup: + sub_in = inputs[in_name] + for item in sub_in: + sub_in_name, sub_in_val = item[0], item[1] + __set_input__(sub_in_name, sub_in_val) + else: + __set_input__(in_name, inputs[in_name]) + + +def append_input_output(block, op_proto, np_list, is_input, dtype): + '''Insert VarDesc and generate Python variable instance''' + proto_list = op_proto.inputs if is_input else op_proto.outputs + + def create_var(block, name, np_list, var_proto): + dtype = None + shape = None + lod_level = None + if name not in np_list: + assert var_proto.intermediate, "{} not found".format(name) + else: + # inferece the dtype from numpy value. + np_value = np_list[name] + if isinstance(np_value, tuple): + dtype = np_value[0].dtype + # output shape, lod should be infered from input. + if is_input: + shape = list(np_value[0].shape) + lod_level = len(np_value[1]) + else: + dtype = np_value.dtype + if is_input: + shape = list(np_value.shape) + lod_level = 0 + # NOTE(dzhwinter): type hacking + # numpy float16 is binded to paddle::platform::float16 + # in tensor_py.h via the help of uint16 datatype. Because + # the internal memory representation of float16 is + # actually uint16_t in paddle. So we use np.uint16 in numpy for + # raw memory, it can pass through the pybind. So in the testcase, + # we feed data use data.view(uint16), but the dtype is float16 in fact. + # The data.view(uint16) means do not cast the data type, but process data as the uint16 + if dtype == np.uint16: + dtype = np.float16 + return block.create_var( + dtype=dtype, shape=shape, lod_level=lod_level, name=name) + + var_dict = {} + for var_proto in proto_list: + var_name = str(var_proto.name) + if is_input: + if (var_name not in np_list) and var_proto.dispensable: + continue + assert (var_name in np_list) or (var_proto.dispensable), \ + "Missing {} as input".format(var_name) + if var_proto.duplicable: + assert isinstance(np_list[var_name], list), \ + "Duplicable {} should be set as list".format(var_name) + var_list = [] + for (name, np_value) in np_list[var_name]: + var_list.append( + create_var(block, name, {name: np_value}, var_proto)) + var_dict[var_name] = var_list + else: + var_dict[var_name] = create_var(block, var_name, np_list, var_proto) + + return var_dict + + +def append_loss_ops(block, output_names): + mean_inputs = list(map(block.var, output_names)) + # for item in mean_inputs: + # print(item) + # print("Item", item.dtype) + + if len(mean_inputs) == 1: + loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1]) + op = block.append_op( + inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean') + op.desc.infer_var_type(block.desc) + op.desc.infer_shape(block.desc) + else: + avg_sum = [] + for cur_loss in mean_inputs: + cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1]) + op = block.append_op( + inputs={"X": [cur_loss]}, + outputs={"Out": [cur_avg_loss]}, + type="mean") + op.desc.infer_var_type(block.desc) + op.desc.infer_shape(block.desc) + avg_sum.append(cur_avg_loss) + + loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1]) + op_sum = block.append_op( + inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum') + op_sum.desc.infer_var_type(block.desc) + op_sum.desc.infer_shape(block.desc) + + loss = block.create_var(dtype=loss_sum.dtype, shape=[1]) + op_loss = block.append_op( + inputs={"X": loss_sum}, + outputs={"Out": loss}, + type='scale', + attrs={'scale': 1.0 / float(len(avg_sum))}) + op_loss.desc.infer_var_type(block.desc) + op_loss.desc.infer_shape(block.desc) + return loss diff --git a/python/paddle/fluid/tests/unittests/transformer_model.py b/python/paddle/fluid/tests/unittests/transformer_model.py index c62792face3c353db1f2e3c77eaf4bd32fbded69..868a0248be6833d0e8fed8a26549352562c279c1 100644 --- a/python/paddle/fluid/tests/unittests/transformer_model.py +++ b/python/paddle/fluid/tests/unittests/transformer_model.py @@ -22,7 +22,7 @@ pos_enc_param_names = ( "src_pos_enc_table", "trg_pos_enc_table", ) -batch_size = 64 +batch_size = 2 def position_encoding_init(n_position, d_pos_vec): @@ -118,8 +118,9 @@ def multi_head_attention(queries, # FIXME(guosheng): Decouple the program desc with batch_size. return layers.reshape( x=trans_x, - shape=map(int, - [batch_size, -1, trans_x.shape[2] * trans_x.shape[3]])) + shape=list( + map(int, [batch_size, -1, trans_x.shape[2] * trans_x.shape[3] + ]))) def scaled_dot_product_attention(q, k, v, attn_bias, d_model, dropout_rate): """ @@ -403,7 +404,7 @@ def transformer( trg_pad_idx, pos_pad_idx, ): file_obj = fluid.layers.open_recordio_file( - filename='./wmt16.recordio', + filename='/tmp/wmt16.recordio', shapes=[ [batch_size * max_length, 1], [batch_size * max_length, 1], diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index 7da123dd92ed9d111d68cd70efb8ce1493452609..eed9b49ef40b591d5b6481846dab714423f57990 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -14,51 +14,129 @@ import contextlib import os +import errno +import shutil +import time -import core - -import data_feeder -import executor -import framework -import io +from . import core +from . import data_feeder +from . import executor +from . import framework +from . import io # optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module -import optimizer as opt_module -import parallel_executor -from transpiler import distribute_transpiler +from . import optimizer as opt_module +from . import parallel_executor +from .transpiler import distribute_transpiler __all__ = [ - 'Trainer', - 'BeginEpochEvent', - 'EndEpochEvent', - 'BeginStepEvent', - 'EndStepEvent', + 'Trainer', 'BeginEpochEvent', 'EndEpochEvent', 'BeginStepEvent', + 'EndStepEvent', 'CheckpointConfig' ] class BeginEpochEvent(object): + """ + The begin of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + """ + def __init__(self, epoch_id): self.epoch = epoch_id class EndEpochEvent(object): + """ + The end of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + """ + def __init__(self, epoch_id): self.epoch = epoch_id class BeginStepEvent(object): + """ + The begin of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + step_id(int): The current step ID. + """ + def __init__(self, epoch_id, step_id): self.epoch = epoch_id self.step = step_id self.fetch_metrics = True + """ + If fetch_metrics is true, the metrics will be fetched at the + EndStepEvent. Default is True. + """ class EndStepEvent(object): + """ + The end of a training step. + + Args: + epoch_id(int): The current epoch ID. + step_id(int): The current step ID. + metrics(list): A list of fetched tensor. The order of this list is same + as the :code:`train_func` returns. + """ + def __init__(self, epoch_id, step_id, metrics): self.epoch = epoch_id self.step = step_id self.metrics = metrics +class CheckpointConfig(object): + """ + Parameter object for :code:`save_checkpoint` and + :code:`fluid.Trainer`. Used to configuration how to save checkpoint. + + Args: + checkpoint_dir(str): Directory path to save check point. Default is the + current directory. + + max_num_checkpoints(int): The max number of local check points. + epoch_interval(int): Every number of epoch to save check point. + step_interval(int): Every number of step to save check point. + + Examples: + >>> config = fluid.CheckpointConfig("./checkpoints") + >>> trainer = fluid.Trainer(train_func=train_program, + >>> place=place, + >>> optimizer_func=optimizer_func, + >>> checkpoint_config=config) + >>> trainer.train(...) + """ + + def __init__(self, + checkpoint_dir=None, + max_num_checkpoints=3, + epoch_interval=1, + step_interval=10): + + assert epoch_interval >= 1 + assert step_interval >= 1 + + self.checkpoint_dir = checkpoint_dir \ + if checkpoint_dir is not None else os.getcwd() + self.max_num_checkpoints = max_num_checkpoints + self.epoch_interval = epoch_interval + self.step_interval = step_interval + self.epoch_id = 0 + self.step_id = 0 + self.load_serial = None + self.pserver_id = None + self.lookup_table_name = None + + def check_and_get_place(place): """ Check the type of place or get the default place @@ -87,28 +165,89 @@ def check_and_get_place(place): class Trainer(object): """ + A trainer wraps MultiGPU/MultiNode training loops and can be used to train a + simple neural network easily. + + This API takes a :code:`train_func`. A :code:`train_func` is a function that + return loss as it first return value. The reset value can be fetched by + EndStepEvent.metrics + + This API also takes a :code:`optimizer_func` that will return an optimizer + instance. + + For example, to train a MLP for MNIST dataset, the sample program is + + >>> import paddle.fluid as fluid + >>> + >>> def mlp(image, layer_sizes=[200, 100], activation="relu", num_classes=10): + >>> hidden = image + >>> for layer_size in layer_sizes: + >>> hidden = fluid.layers.fc(input=hidden, size=layer_size, act=activation) + >>> return fluid.layers.fc(input=hidden, size=num_classes, act="softmax") + >>> + >>> def train_mnist_mlp(): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> label = fluid.layers.data(name='label', shape=[1], dtype='int64') + >>> prediction = mlp(img) + >>> return fluid.layers.mean(fluid.layers.cross_entropy(prediction, label)) + >>> + >>> def optimizer(): + >>> return fluid.optimizer.Adam() + >>> + >>> trainer = Trainer(train_func=train_mnist_mlp, + >>> optimizer_func=optimizer, + >>> place=fluid.CUDAPlace(0), + >>> parallel=True) + >>> + >>> def train_callback(event): + >>> if isinstance(event, fluid.EndStepEvent): + >>> print "Epoch ID", event.epoch, "Step ID",\ + >>> event.step, "AvgLoss", event.metrics[0] + >>> elif isinstance(event, fluid.EndEpochEvent): + >>> trainer.save_params("./model_{0}".format(event.epoch)) + >>> + >>> trainer.train(num_epochs=100, event_handler=train_callback) + + For more example, please see :ref:`api_guide_high_level_api`. + Args: - train_func(callable): A function which will return loss. The loss must be a scalar. - optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer - place: The device place of this trainer. + train_func(callable): A function which will return loss. The loss must be + a scalar tensor. + optimizer_func(callable): A function that returns an Optimizer object. + place(CUDAPlace|CPUPlace): The device place of this trainer. If + :code:`parallel=True,` all CUDA Places will be used if :code:`place` + is a :code:`CUDAPlace`. + parallel(bool): True if use multiple devices. + checkpoint_config(CheckpointConfig): Configuration about how to save + checkpoints. """ def __init__(self, train_func, - optimizer, + optimizer_func, param_path=None, place=None, - parallel=False): + parallel=False, + checkpoint_config=None): self.__stop = False self.parallel = parallel + + # config for checkpoint + # only chief worker will save variables + self.trainer_id = 0 + self.checkpoint_cfg = checkpoint_config + if self.checkpoint_cfg: + assert isinstance(self.checkpoint_cfg, CheckpointConfig) + serial = _get_latest_checkpoint_serial( + self.checkpoint_cfg.checkpoint_dir) + self.checkpoint_cfg.load_serial = serial if serial >= 0 else None + + self.scope = core.Scope() + # 1. we need to generate a framework.Program by calling # program_func. Reference: fluid.program_guard in # test_word2vec.py - if not isinstance(optimizer, opt_module.Optimizer): - raise TypeError("The optimizer should be an instance of Optimizer") - - self.scope = core.Scope() self.startup_program = framework.Program() self.train_program = framework.Program() @@ -117,12 +256,15 @@ class Trainer(object): program_func_outs = train_func() self.train_func_outputs = program_func_outs if isinstance( program_func_outs, list) else [program_func_outs] - self.test_program = self.train_program.clone() + self.test_program = self.train_program.clone(for_test=True) + + # The first element of program_func_outs is loss. + loss = self.train_func_outputs[0] + + optimizer = optimizer_func() if not isinstance(optimizer, opt_module.Optimizer): raise TypeError( "The optimizer should be an instance of Optimizer") - # The fisrt element of program_func_outs is loss. - loss = self.train_func_outputs[0] optimize_ops, params_grads = optimizer.minimize(loss) self.place = check_and_get_place(place) @@ -136,9 +278,15 @@ class Trainer(object): exe = executor.Executor(place) exe.run(self.startup_program) - if param_path: + if self.checkpoint_cfg and self.checkpoint_cfg.load_serial is not None: + self._load_checkpoint() + + if param_path and os.path.isdir(param_path): # load params from param_path into scope - io.load_persistables(exe, dirname=param_path) + io.load_persistables( + executor=exe, + dirname=param_path, + main_program=self.startup_program) def _transpile_nccl2_dist(self): # PADDLE_TRAINER_IPS @@ -152,7 +300,7 @@ class Trainer(object): for ip in worker_ips.split(","): worker_endpoints.append(':'.join([ip, port])) self.num_trainers = len(worker_endpoints) - current_endpoint = os.getenv("POD_IP") + ":" + port + current_endpoint = os.getenv("PADDLE_CURRENT_IP") + ":" + port worker_endpoints.remove(current_endpoint) # TODO(wuyi): use self.nccl_id_var, self.num_trainers and self.trainer_id # in ParallelExecutor to start @@ -193,14 +341,21 @@ class Trainer(object): current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port # the unique trainer id, starting from 0, needed by trainer # only - trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + self.trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + # the role, should be either PSERVER or TRAINER training_role = os.getenv("PADDLE_TRAINING_ROLE") with self._prog_and_scope_guard(): t = distribute_transpiler.DistributeTranspiler() t.transpile( - trainer_id, pservers=pserver_endpoints, trainers=trainers) + self.trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": + if self.checkpoint_cfg: + pserver_id = eplist.index(current_endpoint) + self.checkpoint_cfg.pserver_id = pserver_id + if t.has_distributed_lookup_table: + self.checkpoint_cfg.lookup_table_name = t.table_name + self.train_program = t.get_pserver_program(current_endpoint) self.startup_program = t.get_startup_program(current_endpoint, self.train_program) @@ -219,17 +374,18 @@ class Trainer(object): def train(self, num_epochs, event_handler, reader=None, feed_order=None): """ - Train the model. + Start the train loop to train the model. Args: - num_epochs: The number of epoch. An epoch will process all data in reader - event_handler: The event handler. A function with type (ev:Event)->void - reader: - feed_order: Feeding order of reader. None will following the defining + num_epochs(int): The number of epoch. An epoch will process all data in reader + event_handler(callable): The event handler. A function with type (ev:Event)->void + reader(callable): A reader creator object. See also + :ref:`api_guide_python_reader` . + feed_order(list): Feeding order of reader. None will following the defining order in program Returns: - + None """ training_role = os.getenv("PADDLE_TRAINING_ROLE", "") if training_role == "PSERVER": @@ -249,16 +405,24 @@ class Trainer(object): Test the model on given test data Args: - reader: The reader that yields test data. - feed_order: Feeding order of reader. None will following the defining - order in program + reader(callable): The reader that yields test data. + feed_order(list): Feeding order of reader. None will following the + defining order in program """ return self._test_by_executor(reader, feed_order, self.train_func_outputs) def save_params(self, param_path): - # reference: save_persistables in io.py + """ + Save all parameters into :code:`param_path`. + + Args: + param_path(str): The path to save parameters. + + Returns: + None + """ with self._prog_and_scope_guard(): exe = executor.Executor(self.place) io.save_persistables(exe, dirname=param_path) @@ -293,11 +457,26 @@ class Trainer(object): self._train_by_any_executor(event_handler, exe, num_epochs, reader) def _train_by_any_executor(self, event_handler, exe, num_epochs, reader): - for epoch_id in range(num_epochs): + if self.checkpoint_cfg: + epochs = [ + epoch_id for epoch_id in range(num_epochs) + if epoch_id >= self.checkpoint_cfg.epoch_id + ] + else: + epochs = [epoch_id for epoch_id in range(num_epochs)] + + for epoch_id in epochs: event_handler(BeginEpochEvent(epoch_id)) for step_id, data in enumerate(reader()): if self.__stop: + if self.checkpoint_cfg: + self._clean_checkpoint() return + + if self.checkpoint_cfg and self.checkpoint_cfg.load_serial \ + and self.checkpoint_cfg.step_id >= step_id and self.checkpoint_cfg.epoch_id == epoch_id: + continue + begin_event = BeginStepEvent(epoch_id, step_id) event_handler(begin_event) if begin_event.fetch_metrics: @@ -308,8 +487,13 @@ class Trainer(object): ]) else: metrics = exe.run(feed=data, fetch_list=[]) + + if self.checkpoint_cfg: + self._save_checkpoint(epoch_id, step_id) event_handler(EndStepEvent(epoch_id, step_id, metrics)) event_handler(EndEpochEvent(epoch_id)) + if self.checkpoint_cfg: + self._clean_checkpoint() def _test_by_executor(self, reader, feed_order, fetch_list): with executor.scope_guard(self.scope): @@ -348,6 +532,74 @@ class Trainer(object): loss_name=self.train_func_outputs[0].name) return self._get_parallel_executor() + def _clean_checkpoint(self): + assert self.checkpoint_cfg + clean_checkpoint(checkpoint_dir=self.checkpoint_cfg.checkpoint_dir) + + def _get_checkpoint_load_args(self): + """ + epoch_id and step_id are runtime arguments, they are not variables, will load them independently. + """ + return ["epoch_id", "step_id"] + + def _get_checkpoint_save_args(self, epoch_id, step_id): + """ + epoch_id and step_id are runtime arguments, they are not variables, will save them independently. + """ + trainer_args = {} + trainer_args["epoch_id"] = epoch_id + trainer_args["step_id"] = step_id + return trainer_args + + def _save_checkpoint(self, epoch_id, step_id): + assert self.checkpoint_cfg + + if epoch_id % self.checkpoint_cfg.epoch_interval == 0 \ + and step_id % self.checkpoint_cfg.step_interval == 0: + exe = executor.Executor(self.place) + save_checkpoint( + executor=exe, + checkpoint_dir=self.checkpoint_cfg.checkpoint_dir, + trainer_id=self.trainer_id, + trainer_args=self._get_checkpoint_save_args(epoch_id, step_id), + main_program=self.train_program, + max_num_checkpoints=self.checkpoint_cfg.max_num_checkpoints) + + def _load_checkpoint(self): + with self._prog_and_scope_guard(): + exe = executor.Executor(self.place) + load_checkpoint( + executor=exe, + checkpoint_dir=self.checkpoint_cfg.checkpoint_dir, + main_program=self.startup_program) + + if not self.checkpoint_cfg.pserver_id: + load_trainer_args = self._get_checkpoint_load_args() + trainer_args = load_checkpoint( + executor=exe, + checkpoint_dir=self.checkpoint_cfg.checkpoint_dir, + main_program=self.startup_program, + role_id=self.trainer_id, + is_trainer=True, + load_trainer_args=load_trainer_args) + + if len(trainer_args) != 2: + raise ValueError( + "the return trainer_args length do not equal _get_checkpoint_load_args" + ) + self.checkpoint_cfg.epoch_id = int(trainer_args[0]) + self.checkpoint_cfg.step_id = int(trainer_args[1]) + else: + if self.checkpoint_cfg.lookup_table_name: + load_checkpoint( + executor=exe, + checkpoint_dir=self.checkpoint_cfg.checkpoint_dir, + main_program=self.startup_program, + role_id=self.checkpoint_cfg.pserver_id, + is_trainer=False, + load_trainer_args=None, + load_lookup_table=self.checkpoint_cfg.lookup_table_name) + def build_feed_var_list(program, feed_order): if not isinstance(program, framework.Program): @@ -361,12 +613,620 @@ def build_feed_var_list(program, feed_order): if not isinstance(feed_order, dict): raise TypeError( "The 'feed_order' should be either None, list or dict.") - if not sorted(feed_order.values()) == range(len(feed_order)): + if not sorted(feed_order.values()) == list(range(len(feed_order))): raise ValueError( "The values of 'feed_order' should be a permutation of [0, len(feed_order))" ) - sorted_pair_list = sorted(feed_order.items(), key=lambda item: item[1]) + sorted_pair_list = sorted( + list(feed_order.items()), key=lambda item: item[1]) feed_var_list = [ program.global_block().var(pair[0]) for pair in sorted_pair_list ] return feed_var_list + + +# move Checkpoint APIs from io.py to trainer.py, make all of them are private. +SUCCESS_MARK_FILENAME = "_SUCCESS" +CHECKPOINT_PREFIX = "checkpoint" +MODEL_DIR = "__model__" +LOOKUP_TABLE_DIR = "__lookup_table__" +TRAINER_PREFIX = "trainer" +CHECKPOINT_SEPARATOR = "_" + + +def save_checkpoint(executor, + checkpoint_dir, + trainer_id, + main_program, + trainer_args=None, + max_num_checkpoints=3, + lookup_table=None, + pserver_endpoints=None): + """ + This function filters out all checkpoint variables from the give + main_program and then saves these variables to the `checkpoint_dir` + directory. + + In the training precess, we generally save a checkpoint in each + iteration. So there might be a lot of checkpoints in the + `checkpoint_dir`. To avoid them taking too much disk space, the + `max_num_checkpoints` are introduced to limit the total number of + checkpoints. If the number of existing checkpints is greater than + the `max_num_checkpoints`, oldest ones will be scroll deleted. + + A variable is a checkpoint variable and will be saved if it meets + all following conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". + + Args: + executor(Executor): The executor to run for save checkpoint. + checkpoint_dir(str): The folder where to save checkpoints. + trainer_id(int): currect trainer id, if id is equal to 0, the trainer + is chief. + trainer_args(dict|None): Current training arguments. Such as 'epoch_id' + and 'step_id'. + Defaut: None + main_program(Program): The program whose checkpoint variables will + be saved. + max_num_checkpoints(int): The max number of total number of existing + checkpoints. + Default: 3 + lookup_table(string|None): the lookup table name, when use distribute + lookup table, we can get lookup table name by DistributeTranspiler. + table_name + pserver_endpoints(list|None): the parameter server ip:port list. + when use distribute lookup table, we can get pserver_endpoints by + distribute arguments. + + Returns: + None + + Raises: + ValueError: If `checkpoint_dir` is None. + AssertionError: If `trainer_args` is not a dict. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./checkpoints" + prog = fluid.default_main_program() + trainer_args = {"epoch_id": 200, + "step_id": 20} # just an example + table_name = "share_w" + ps_endpoints = ["127.0.0.1:6000","127.0.0.1:6001"] + + save_checkpoint(executor=exe, + checkpoint_dir=path, + trainer_id=0, + trainer_args=trainer_args, + main_program=prog, + max_num_checkpoints=3, + lookup_table=table_name, + pserver_endpoints = ps_endpoints) + """ + if checkpoint_dir is None: + raise ValueError("'checkpoint_dir' should not be None") + + if main_program is None: + raise ValueError('main_program should not be None.') + + if trainer_args: + assert isinstance(trainer_args, dict) + + is_chief = trainer_id == 0 + + _make_chekcpoint_dirs(checkpoint_dir) + serial = _get_latest_checkpoint_serial(checkpoint_dir) + 1 + cur_dir = _get_serial_dir(checkpoint_dir, serial) + + _save_trainer_args(cur_dir, trainer_id, trainer_args) + + if is_chief: + _save_persist_vars_without_grad(executor, cur_dir, main_program) + + if is_chief and lookup_table and pserver_endpoints: + _save_pserver_vars_by_notify(executor, cur_dir, lookup_table, + pserver_endpoints) + + _scroll_delete(checkpoint_dir, max_num_checkpoints) + + +def load_checkpoint(executor, + checkpoint_dir, + main_program, + role_id=0, + is_trainer=True, + load_trainer_args=None, + load_lookup_table=None): + """ + This function filters out all checkpoint variables from the give + main_program and then try to load these variables from the + `checkpoint_dir` directory. + + In the training precess, we generally save a checkpoint in each + iteration. So there are more than one checkpoint in the + `checkpoint_dir` (each checkpoint has its own sub folder), use + `serial` to specify which serial of checkpoint you would like to + load. + + A variable is a checkpoint variable and will be loaded if it meets + all following conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". + + Args: + executor(Executor): The executor to run for loading checkpoint. + checkpoint_dir(str): The folder where all checkpoints are. + serial(int): The serial of checkpoint you would like to load. + main_program(Program): The program whose checkpoint variables will + be loaded. + role_id(int): the trainer id or the parameter server id. + is_trainer(bool): trainer is True and parameter server is False. + load_trainer_args(list|None): list about load trainer args. + load_lookup_table(str|None): the lookup table name + + Returns: + None + + Raises: + ValueError: If `checkpoint_dir` is None. + ValueError: If `main_program` is None. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./checkpoints" + prog = fluid.default_main_program() + load_checkpoint(executor=exe, checkpoint_dir=path, + serial=9, main_program=prog) + + # In this example, `load_checkpoint` function + # will first filters out all checkpoint variables in the default + # main program, and then try to load these variables form the + # folder "./checkpoints/checkpoint_9/__model__". + """ + + if checkpoint_dir is None: + raise ValueError("'checkpoint_dir' should not be None") + + serial = _get_latest_checkpoint_serial(checkpoint_dir) + + # there are nothing need to be loaded + if serial is None or serial < 0: + return + + if main_program is None: + raise ValueError('main_program should not be None.') + + if is_trainer and load_trainer_args is None: + cur_dir = _get_serial_dir(checkpoint_dir, serial) + _load_persist_vars_without_grad(executor, cur_dir, main_program, True) + return + + if is_trainer and load_trainer_args: + return _load_trainer_args(checkpoint_dir, serial, role_id, + load_trainer_args) + + if not is_trainer and load_lookup_table: + _load_lookup_table_vars(executor, checkpoint_dir, main_program, role_id, + load_lookup_table) + + +def clean_checkpoint(checkpoint_dir, delete_dir=False): + """ + clean the checkpoint dir, when the train exits normally, + the trainer will call clean_checkpoint to delete checkpoint directory saved before. + delete_dir only works when the directory is empty, otherwise, OSError is raised. + + : param checkpoint_dir + : param delete_dir + """ + + if checkpoint_dir is None: + raise ValueError("'checkpoint_dir' should not be None") + _scroll_delete(checkpoint_dir, max_num_checkpoints=0) + + if delete_dir and not os.listdir(checkpoint_dir): + os.rmdir(checkpoint_dir) + + +def _load_persist_vars_without_grad(executor, + dirname, + program, + has_model_dir=False): + """ + This function filters out all checkpoint variables from the give + program and then trys to load these variables from the given directory. + + A variable is a checkpoint variable if it meets all following + conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". + + Args: + executor(Executor): The executor to run for loading variables. + dirname(str): The directory path. + program(Program): The program whose checkpoint variables will + be loaded. + has_model_dir(bool): if True, the function loads variables + from a sub directory named '__model__'. + Default: False + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + _load_persist_vars_without_grad(executor=exe, + dirname=param_path, program=prog, has_model_dir=True) + + # In this example, `_load_persist_vars_without_grad` function + # will first filters out all checkpoint variables in the default + # main program, and then trys to load these variables form the + # folder "./my_paddle_model/__model__". + """ + + if has_model_dir: + dirname = _get_model_dir(dirname) + + io.load_vars( + executor, + dirname=dirname, + main_program=program, + predicate=_is_checkpoint_var, + filename=None) + + +def _load_lookup_table_vars(executor, dirname, program, pserver_id, table_name): + """ + The parameter server will load lookup table's local file in + selectedrows variable. + + Args: + executor(Executor): The executor to run for loading persistable variables + dirname(str): The directory path + main_program(Program): Find the variable named table_name in main_program + pserver_id(int): the serial number in pserver_endpoints list + table_name(str): lookup table name + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + dirname = "./checkpoints/checkpoint_9/" + prog = fluid.default_main_program() + pserver_id = 1 + table_name = "share_w" + _load_lookup_table_vars(executor=exe, + dirname=dirname, program=prog, pserver_id=pserver_id, + table_name=table_name) + """ + + for var in program.list_vars(): + if var.name == table_name: + lookup_table_var = var + break + + assert lookup_table_var is not None + + lookup_table_dir = os.path.join(dirname, LOOKUP_TABLE_DIR) + table_file = table_name + CHECKPOINT_SEPARATOR + str(pserver_id) + + load_prog = framework.Program() + load_block = load_prog.global_block() + + load_block.append_op( + type='load', + inputs={}, + outputs={'Out': [lookup_table_var]}, + attrs={'file_path': os.path.join(lookup_table_dir, table_file)}) + + executor.run(load_prog) + + +def _save_persist_vars_without_grad(executor, dirname, program): + """ + This function filters out all checkpoint variables from the give + program and then save these variables to a sub-folder '__model__' of + the given directory. + + A variable is a checkpoint variable if it meets all following + conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". + + Args: + executor(Executor): The executor to run for saving variables. + dirname(str): The directory path. + program(Program): The program whose checkpoint variables will + be saved. + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + _save_persist_vars_without_grad(executor=exe, + dirname=param_path, program=prog) + + # In this example, `_save_persist_vars_without_grad` function + # will first filters out all checkpoint variables in the default + # main program, and then saves these variables to the folder + # "./my_paddle_model/__model__". + """ + cur_dir = _get_model_dir(dirname) + io.save_vars( + executor, + dirname=cur_dir, + main_program=program, + vars=None, + predicate=_is_checkpoint_var, + filename=None) + _write_success(cur_dir) + + +def _save_pserver_vars_by_notify(executor, dirname, lookup_table, + ps_endpoint_list): + """ + This function will send checkpoint notify message from Trainer 0 + to all the pservers. + The checkpoint notify message contains lookup table name, + the absolute path on pserver to save lookup_table. + + Args: + executor(Executor): The executor to run for send checkpoint notify. + dirname(str): The folder where to save checkpoints. + lookup_table(string): the lookup table name, when use distribute + lookup table, we can get lookup table name by DistributeTranspiler. + table_name + ps_endpoint_list(list): the parameter server ip:port list. + when use distribute lookup table, we can get ps_endpoint_list by + distribute arguments. + Return: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + table_name = "share_w" + ps_endpoints = ["127.0.0.1:6000","127.0.0.1:6001"] + + _save_pserver_vars_by_notify(executor=exe, + dirname=param_path, lookup_table=table_name, + ps_endpoint_list=ps_endpoints) + """ + cur_dir = _get_lookuptable_dir(dirname) + + checkpoint_notify_program = framework.Program() + checkpoint_notify_block = checkpoint_notify_program.global_block() + + attrs = {} + attrs['epmap'] = ps_endpoint_list + attrs['dir'] = cur_dir + attrs['lookup_table'] = lookup_table + + checkpoint_notify_block.append_op( + type='checkpoint_notify', inputs={}, outputs={}, attrs=attrs) + executor.run(checkpoint_notify_program) + + +def _save_trainer_args(dirname, trainer_id, trainer_args): + assert isinstance(trainer_args, dict) + + cur_dir = _get_trainer_dir(dirname, trainer_id) + + for name, value in list(trainer_args.items()): + args_file = os.path.join(cur_dir, name) + with open(args_file, 'w') as f: + f.write(str(value)) + _write_success(cur_dir) + + +def _load_trainer_args(checkpoint_dir, serial, trainer_id, trainer_args): + """ + trainer will load some args from it's independent directory, + such as epoch_id and step_id. + + Args: + checkpoint_dir(str): The folder where all checkpoints are. + serial(int): The serial of checkpoint you would like to load. + trainer_id(int): current trainer id. + trainer_args(list): list about load trainer args + Return: + None + + Examples: + .. code-block:: python + + param_path = "./checkpoint/" + serial = 7 + trainer_id = 2 + trainer_args = ["epoch_id", "step_id"] + + _load_trainer_args(checkpoint_dir=param_path, serial=serial, + trainer_id=trainer_id, trainer_args=trainer_args) + """ + assert isinstance(trainer_args, list) + + cur_dir = _get_serial_dir(checkpoint_dir, serial) + cur_dir = _get_trainer_dir(cur_dir, trainer_id) + + ret_values = [] + + for arg in trainer_args: + cur_file = os.path.join(cur_dir, arg) + with open(cur_file, 'r') as f: + contents = f.read() + ret_values.append(contents.strip()) + return ret_values + + +def _is_checkpoint_var(var): + """ + the checkpoint will not save or load all the variables. + var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded. + + : param var(Variable) + """ + if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ + var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ + var.desc.type() == core.VarDesc.VarType.RAW: + return False + # @GRAD are named for gradient variables, checkpoint will not save it. + if "@GRAD" in var.name: + return False + # .trainer_ are named for distribute train variables, checkpoint will not save it. + if ".trainer_" in var.name: + return False + + # .block is named for distribute train variables, checkpoint will not save it. + if ".block" in var.name: + return False + + return var.persistable + + +def _make_chekcpoint_dirs(dirs): + """ + _make_chekcpoint_dirs will makdir local directory directly, when the directory is exist, it will igore it. + """ + assert dirs is not None + + if os.path.isfile(dirs): + raise OSError(errno.ENOTDIR, "dirs path shoule be a Directory.", dirs) + + if not os.path.isdir(dirs): + try: + os.makedirs(dirs) + except OSError as err: + if err.errno != errno.EEXIST: + raise err + + +def _get_dir_serial(dirname): + _, serial = dirname.split(CHECKPOINT_SEPARATOR) + + try: + serial_num = int(serial) + except ValueError: + serial_num = -1 + return serial_num + + +def _get_serial_dir(dirname, serial): + serial_folder = CHECKPOINT_PREFIX + CHECKPOINT_SEPARATOR + str(serial) + serial_dir = os.path.join(dirname, serial_folder) + _make_chekcpoint_dirs(serial_dir) + + return serial_dir + + +def _get_model_dir(dirname): + model_dir = os.path.join(dirname, MODEL_DIR) + _make_chekcpoint_dirs(model_dir) + return model_dir + + +def _get_lookuptable_dir(dirname): + lookuptable_dir = os.path.join(dirname, LOOKUP_TABLE_DIR) + _make_chekcpoint_dirs(lookuptable_dir) + return lookuptable_dir + + +def _get_trainer_dir(dirname, trainer_id): + trainer_folder = TRAINER_PREFIX + CHECKPOINT_SEPARATOR + str(trainer_id) + trainer_dir = os.path.join(dirname, trainer_folder) + _make_chekcpoint_dirs(trainer_dir) + return trainer_dir + + +def _scroll_delete(dirname, max_num_checkpoints=3): + dirs = os.listdir(dirname) + serial_map = {} + for serial in dirs: + serial_num = _get_dir_serial(serial) + serial_map[serial_num] = serial + + if len(list(serial_map.keys())) <= max_num_checkpoints: + return + + serials = list(serial_map.keys()) + serials.sort(reverse=True) + serials = serials[max_num_checkpoints:] + for serial in serials: + cur_dir = _get_serial_dir(dirname, serial) + try: + shutil.rmtree(cur_dir) + except OSError as err: + if err.errno != errno.ENOENT: + raise err + + +def _write_success(dirname): + """ + write an empty file named "_SUCCESS" in checkpoint dir, indicate this checkpoint is correct. + + : param dirname + """ + success_file = os.path.join(dirname, SUCCESS_MARK_FILENAME) + with open(success_file, 'a') as f: + now = time.ctime() + f.write(now) + + +def _get_latest_checkpoint_serial(checkpoint_dir): + """ + get the latest file in checkpoint directory, the _SUCCESS file must exist in the directory + + : param checkpoint_dir + """ + if not checkpoint_dir: + return -1 + + def has_success(checkpoint_dir, cur_dir): + """ + is _SUCCESS in this dir + """ + + serial = _get_dir_serial(cur_dir) + if serial == -1 or not os.path.isdir( + os.path.join(checkpoint_dir, cur_dir)): + return -1 + + success_path = os.path.join( + _get_serial_dir(checkpoint_dir, serial), MODEL_DIR, + SUCCESS_MARK_FILENAME) + if os.path.isfile(success_path): + return serial + + if not os.path.isdir(checkpoint_dir): + return -1 + + current_dir = -1 + dirs = os.listdir(checkpoint_dir) + for cur_dir in dirs: + success_num = has_success(checkpoint_dir, cur_dir) + if success_num > current_dir: + current_dir = success_num + return current_dir diff --git a/python/paddle/fluid/transpiler/__init__.py b/python/paddle/fluid/transpiler/__init__.py index 045ca537b2e84c02298d6375a7ef5bdbb5517380..a8622ad54433fff40f68520955f0294e2955577e 100644 --- a/python/paddle/fluid/transpiler/__init__.py +++ b/python/paddle/fluid/transpiler/__init__.py @@ -12,13 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. -from distribute_transpiler import DistributeTranspiler -from inference_transpiler import InferenceTranspiler -from memory_optimization_transpiler import memory_optimize, release_memory -from distribute_transpiler_simple import SimpleDistributeTranspiler -from ps_dispatcher import HashName, RoundRobin +from .distribute_transpiler import DistributeTranspiler, DistributeTranspilerConfig +from .inference_transpiler import InferenceTranspiler +from .memory_optimization_transpiler import memory_optimize, release_memory +from .ps_dispatcher import HashName, RoundRobin __all__ = [ - "DistributeTranspiler", "InferenceTranspiler", "SimpleDistributeTranspiler", - "memory_optimize", "release_memory", "HashName", "RoundRobin" + "DistributeTranspiler", "InferenceTranspiler", "memory_optimize", + "release_memory", "HashName", "RoundRobin", "DistributeTranspilerConfig" ] diff --git a/python/paddle/fluid/transpiler/details/__init__.py b/python/paddle/fluid/transpiler/details/__init__.py index dc597c33849dc06cc975b245099672f64c3539d3..1bfab1f219f8a2f08a0fb5c0042d87a3ad707dd5 100644 --- a/python/paddle/fluid/transpiler/details/__init__.py +++ b/python/paddle/fluid/transpiler/details/__init__.py @@ -12,5 +12,5 @@ # See the License for the specific language governing permissions and # limitations under the License. -from program_utils import * -from ufind import * +from .program_utils import * +from .ufind import * diff --git a/python/paddle/fluid/transpiler/details/program_utils.py b/python/paddle/fluid/transpiler/details/program_utils.py index f10b496306a002ee131d01798a0698b807d379ca..76d10777f5f9ed6d27d55a640108bd036d8d8bac 100644 --- a/python/paddle/fluid/transpiler/details/program_utils.py +++ b/python/paddle/fluid/transpiler/details/program_utils.py @@ -17,10 +17,10 @@ def delete_ops(block, ops): try: start = list(block.ops).index(ops[0]) end = list(block.ops).index(ops[-1]) - [block.remove_op(start) for _ in xrange(end - start + 1)] - except Exception, e: + [block._remove_op(start) for _ in range(end - start + 1)] + except Exception as e: raise e - block.program.sync_with_cpp() + block.program._sync_with_cpp() def find_op_by_input_arg(block, arg_name): diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index 06b0a1375ce6568cca864cd8a2dd69ee46b223a7..ce4709f23b752cc061f3b767a262f82378b86707 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -12,21 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. """ -Transpile the program to distributed data-parallelism programs. -The main_program will be transformed to use a remote parameter server -to do parameter optimization. And the optimization graph will be put -into a parameter server program. - -Use different methods to split trainable variables to different -parameter servers. - Steps to transpile trainer: 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". 3. modify trainer program add split_op to each grad variable. -4. append send_op to send splited variables to server and fetch - params(splited blocks or origin param) from server. -5. append concat_op to merge splited blocks to update local weights. +4. append send_op to send splited variables to server and +5. add recv_op to fetch params(splited blocks or origin param) from server. +6. append concat_op to merge splited blocks to update local weights. Steps to transpile pserver: 1. create new program for parameter server. @@ -36,16 +28,17 @@ Steps to transpile pserver: 5. add listen_and_serv op """ -from __future__ import print_function - import math +import random +import numpy as np -from ps_dispatcher import RoundRobin, HashName, PSDispatcher +from .ps_dispatcher import RoundRobin, HashName, PSDispatcher from .. import core, framework from ..framework import Program, default_main_program, \ - default_startup_program, \ - Variable, Parameter, grad_var_name -from details import * + default_startup_program, Block, \ + Parameter, grad_var_name +from .details import * +from functools import reduce LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" @@ -70,7 +63,7 @@ def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") -def split_variable(var_list, service_count, min_block_size=8192): +def slice_variable(var_list, slice_count, min_block_size): """ We may need to split dense tensor to one or more blocks and put them equally onto parameter server. One block is a sub-tensor @@ -78,25 +71,25 @@ def split_variable(var_list, service_count, min_block_size=8192): We need to have a minimal block size so that the calculations in the parameter server side can gain better performance. By default - minimum block size 8K elements (maybe 16bit or 32bit or 64bit). + minimum block size 8K elements (maybe 16bit or 32bit or 64bit). Args: var_list (list): List of variables. - service_count (int): Numel of pserver services. A pserver may have two - or more listening ports. + slice_count (int): Numel of count that variables will be sliced, which + could be the pserver services' count. min_block_size (int): Minimum splitted block size. Returns: - blocks (list[(varname, block_id, current_block_size)]): A list + blocks (list[(varname, block_id, current_block_size)]): A list of VarBlocks. Each VarBlock specifies a shard of the var. """ blocks = [] for var in var_list: - split_count = service_count + split_count = slice_count var_numel = reduce(lambda x, y: x * y, var.shape) max_pserver_count = int(math.floor(var_numel / float(min_block_size))) if max_pserver_count == 0: max_pserver_count = 1 - if max_pserver_count < service_count: + if max_pserver_count < slice_count: split_count = max_pserver_count block_size = int(math.ceil(var_numel / float(split_count))) @@ -108,7 +101,7 @@ def split_variable(var_list, service_count, min_block_size=8192): block_size += dim1 - remains # update split_count after aligning split_count = int(math.ceil(var_numel / float(block_size))) - for block_id in xrange(split_count): + for block_id in range(split_count): curr_block_size = min(block_size, var_numel - ( (block_id) * block_size)) block = VarBlock(var.name, block_id, curr_block_size) @@ -116,141 +109,95 @@ def split_variable(var_list, service_count, min_block_size=8192): return blocks -class DistributeTranspiler: - def _has_distributed_lookup_table(self): - # process lookup_table_op - # 1. check all lookup_table_op is distributed - # 2. check all lookup_table_op share the same table. - distributed_lookup_table_ops = [] - # support only one distributed_lookup_table now - self.table_name = None - for op in self.origin_program.global_block().ops: - if op.type == LOOKUP_TABLE_TYPE: - if op.attrs['is_distributed'] is True: - if self.table_name is None: - self.table_name = op.input("W")[0] - if self.table_name != op.input("W")[0]: - raise RuntimeError("all distributed lookup_table_ops" - " should have only one table") - distributed_lookup_table_ops.append(op) - else: - if self.table_name is not None: - assert op.input("W")[0] != self.table_name +class DistributeTranspilerConfig(object): + """ + slice_var_up (bool): Do Tensor slice for pservers, default is True. + split_method (PSDispatcher): RoundRobin or HashName can be used + try to choose the best method to balance loads for pservers. + min_block_size (int): Minimum splitted element number in block. + According:https://github.com/PaddlePaddle/Paddle/issues/8638#issuecomment-369912156 + We can use bandwidth effiently when data size is larger than 2MB.If you + want to change it, please be sure you see the slice_variable function. + """ - return len(distributed_lookup_table_ops) > 0 + slice_var_up = True + split_method = None + min_block_size = 8192 - def _update_dist_lookup_table_vars(self, param_list, grad_list, - params_grads): - # TODO(wuyi): put find a way to put dist lookup table stuff all together. - # update self.table_param_grad and self.trainer_side_table_grad_list - program = self.origin_program - if self.has_distributed_lookup_table: - param_list = [ - param for param in param_list if param.name != self.table_name - ] - grad_list = [ - grad for grad in grad_list - if grad.name != grad_var_name(self.table_name) - ] - self.table_param_grad = [ - param_grad for param_grad in params_grads - if param_grad[0].name == self.table_name - ][0] - table_grad_var = self.table_param_grad[1] - if self.sync_mode: - self.trainer_side_table_grad_list = [ - program.global_block().create_var( - name="%s.trainer_%d.pserver_%d" % - (table_grad_var.name, self.trainer_id, index), - type=table_grad_var.type, - shape=table_grad_var.shape, - dtype=table_grad_var.dtype) - for index in range(len(self.pserver_endpoints)) - ] - else: - self.trainer_side_table_grad_list = [ - program.global_block().create_var( - name="%s.pserver_%d" % (table_grad_var.name, index), - type=table_grad_var.type, - shape=table_grad_var.shape, - dtype=table_grad_var.dtype) - for index in range(len(self.pserver_endpoints)) - ] - def _init_splited_vars(self, split_method): - # update these mappings for further transpile: - # 1. param_var_mapping: param var name -> [splited params vars] - # 2. grad_var_mapping: grad var name -> [splited grads vars] - # 3. grad_param_mapping: grad.blockx -> param.blockx - # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} +class DistributeTranspiler(object): + """ + **DistributeTranspiler** + + Convert the fluid program to distributed data-parallelism programs. + + The main_program will be transformed to use a remote parameter server + to do parameter optimization. And the optimization graph will be put + into a parameter server program. + + Examples: + .. code-block:: python + + # Define your model before these codes. + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "") + eplist = [] + for ip in pserver_ips.split(","): + eplist.append(':'.join([ip, port])) + pserver_endpoints = ",".join(eplist) + trainers = int(os.getenv("PADDLE_TRAINERS")) + current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port + trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + role = os.getenv("PADDLE_TRAINING_ROLE") + + t = distribute_transpiler.DistributeTranspiler() + t.transpile( + trainer_id, pservers=pserver_endpoints, trainers=trainers) + if role == "PSERVER": + pserver_program = t.get_pserver_program(current_endpoint) + pserver_startup_program = t.get_startup_program(current_endpoint, + pserver_program) + elif role == "TRAINER": + trainer_program = t.get_trainer_program() + """ - param_list = [] - grad_list = [] - for p, g in self.params_grads: - # skip parameter marked not trainable - if type(p) == Parameter and p.trainable == False: - continue - param_list.append(p) - grad_list.append(g) + def __init__(self, config=None): + if config is not None: + self.config = config + else: + self.config = DistributeTranspilerConfig() - self._update_dist_lookup_table_vars(param_list, grad_list, - self.params_grads) + if self.config.split_method is None: + self.config.split_method = RoundRobin - grad_blocks = split_variable(grad_list, len(self.pserver_endpoints)) - param_blocks = split_variable(param_list, len(self.pserver_endpoints)) - assert (len(grad_blocks) == len(param_blocks)) - # origin_varname -> [splited_var] - self.param_var_mapping = self._create_vars_from_blocklist( - self.origin_program, param_blocks) - self.grad_var_mapping = self._create_vars_from_blocklist( - self.origin_program, - grad_blocks, - add_trainer_suffix=self.trainer_num > 1) - self.grad_param_mapping = dict() - for g, p in zip(grad_blocks, param_blocks): - g_name, g_bid, _ = g.split(":") - p_name, p_bid, _ = p.split(":") - self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ - self.param_var_mapping[p_name][int(p_bid)] - - # create mapping of endpoint -> split var to create pserver side program - self.param_grad_ep_mapping = dict() - [ - self.param_grad_ep_mapping.update({ - ep: { - "params": [], - "grads": [] - } - }) for ep in self.pserver_endpoints - ] + assert (self.config.min_block_size >= 8192) + assert (self.config.split_method.__bases__[0] == PSDispatcher) def transpile(self, trainer_id, program=None, pservers="127.0.0.1:6174", trainers=1, - split_method=RoundRobin, sync_mode=True): """ - :param trainer_id: one unique id for each trainer in a job. - :type trainer_id: int - :param program: program to transpile, default is default_main_program - :type program: Program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - :param trainers: total number of workers/trainers in the job - :type trainers: int - :param split_method: A function to determin how to split variables - to different servers equally. - :type split_method: function - :type sync_mode: boolean default True - :param sync_mode: if sync_mode is set True, it means that dist transpiler - will transpile the program into sync_mode pserver and trainer program. + Run the transpiler. + + Args: + trainer_id (int): id for current trainer worker, if you have + n workers, the id may range from 0 ~ n-1 + program (Program|None): program to transpile, + default is fluid.default_main_program(). + pservers (str): comma separated ip:port string for the pserver + list. + trainers (int): number of trainers in the distributed job. + sync_mode (bool): Do sync training or not, default is True. """ - assert (split_method.__bases__[0] == PSDispatcher) if program is None: program = default_main_program() self.origin_program = program + self.origin_startup_program = default_startup_program().clone() + + self.startup_program = default_startup_program() self.trainer_num = trainers self.sync_mode = sync_mode self.trainer_id = trainer_id @@ -258,17 +205,32 @@ class DistributeTranspiler: self.pserver_endpoints = pserver_endpoints self.optimize_ops, self.params_grads = self._get_optimize_pass() - ps_dispatcher = split_method(self.pserver_endpoints) + ps_dispatcher = self.config.split_method(self.pserver_endpoints) self.has_distributed_lookup_table = self._has_distributed_lookup_table() - # split and create vars, then put splited vars in dicts for later use. - self._init_splited_vars(split_method) + # step 1: split and create vars, then put splited vars in dicts for later use. + self._init_splited_vars() - # step 3.1: insert send op to send gradient vars to parameter servers + # step 2: insert send op to send gradient vars to parameter servers ps_dispatcher.reset() send_vars = [] - for orig_varname, splited_vars in self.grad_var_mapping.items(): + + # in general cases, the number of pservers is times of 2, and this + # will lead to uneven distribution among weights and bias: + # fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1 + # fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2 + # shuffle the map will avoid the uneven distribution above + grad_var_mapping_items = list(self.grad_var_mapping.items()) + if not self.config.slice_var_up: + random.seed(self.trainer_num) + random.shuffle(grad_var_mapping_items) + + for orig_varname, splited_vars in grad_var_mapping_items: eplist = ps_dispatcher.dispatch(splited_vars) + + if not self.config.slice_var_up: + assert (len(splited_vars) == 1) + if len(splited_vars) == 1: orig_varname = splited_vars[0].name index = find_op_by_output_arg(program.global_block(), @@ -283,9 +245,9 @@ class DistributeTranspiler: AssertionError("Can not insert the send op by original " "variable name :", orig_varname) - program.global_block().insert_op( + program.global_block()._insert_op( index=index + 1, - type="send_vars", + type="send", inputs={"X": splited_vars}, outputs={}, attrs={ @@ -306,7 +268,7 @@ class DistributeTranspiler: RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - # step 3.2: insert recv op to receive parameters from parameter server + # step 3: insert recv op to receive parameters from parameter server recv_vars = [] for _, var in enumerate(send_vars): recv_vars.append(self.grad_param_mapping[var]) @@ -316,8 +278,9 @@ class DistributeTranspiler: for i, ep in enumerate(eplist): self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i]) self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) + # step4: Concat the parameters splits together after recv. - for varname, splited_var in self.param_var_mapping.iteritems(): + for varname, splited_var in list(self.param_var_mapping.items()): eps = [] for var in splited_var: index = [v.name for v in recv_vars].index(var.name) @@ -332,16 +295,17 @@ class DistributeTranspiler: RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - program.global_block().append_op( - type="fetch_barrier", - inputs={}, - outputs={}, - attrs={ - "endpoints": pserver_endpoints, - RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE - }) + if self.sync_mode: + program.global_block().append_op( + type="fetch_barrier", + inputs={}, + outputs={}, + attrs={ + "endpoints": pserver_endpoints, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) - for varname, splited_var in self.param_var_mapping.iteritems(): + for varname, splited_var in list(self.param_var_mapping.items()): if len(splited_var) <= 1: continue orig_param = program.global_block().vars[varname] @@ -351,28 +315,114 @@ class DistributeTranspiler: outputs={"Out": [orig_param]}, attrs={"axis": 0}) + self._get_trainer_startup_program(recv_vars=recv_vars, eplist=eplist) + if self.has_distributed_lookup_table: self._replace_lookup_table_op_with_prefetch(program, pserver_endpoints) self._split_table_grad_and_add_send_vars(program, pserver_endpoints) def get_trainer_program(self): + """ + Get transpiled trainer side program. + + Returns: + Program: trainer side program. + """ # remove optimize ops and add a send op to main_program + # FIXME(typhoonzero): Also ops like clip_gradient, lrn_decay? delete_ops(self.origin_program.global_block(), self.optimize_ops) - # FIXME(typhoonzero): serialize once will fix error occurs when clone. self.origin_program.__str__() + return self.origin_program + def _get_trainer_startup_program(self, + recv_vars, + eplist, + startup_program=None): + """ + Get transpiled trainer side startup program. + + Args: + startup_program(Program): Startup program. + + Returns: + Program: trainer side startup program. + """ + if startup_program is None: + startup_program = self.startup_program + + # FIXME(gongwb): delete not need ops. + # note that: some parameter is not trainable and those ops can't be deleted. + + for varname, splited_var in self.param_var_mapping.iteritems(): + # Get the eplist of recv vars + eps = [] + for var in splited_var: + index = [v.name for v in recv_vars].index(var.name) + eps.append(eplist[index]) + + for var in splited_var: + if startup_program.global_block().has_var(var.name): + continue + + startup_program.global_block().create_var( + name=var.name, + persistable=False, + type=var.type, + dtype=var.dtype, + shape=var.shape, + lod_level=var.lod_level) + + op = startup_program.global_block().append_op( + type="recv", + inputs={}, + outputs={"Out": splited_var}, + attrs={ + "epmap": eps, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + + startup_program.global_block().append_op( + type="fetch_barrier", + inputs={}, + outputs={}, + attrs={ + "endpoints": self.pserver_endpoints, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + + for varname, splited_var in self.param_var_mapping.iteritems(): + #add concat ops to merge splited parameters received from parameter servers. + if len(splited_var) <= 1: + continue + orig_param = startup_program.global_block().vars[varname] + startup_program.global_block().append_op( + type="concat", + inputs={"X": splited_var}, + outputs={"Out": [orig_param]}, + attrs={"axis": 0}) + + return startup_program + def get_pserver_program(self, endpoint): """ - Get pserver side program using the endpoint. - TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers. - NOTE: assume blocks of the same variable is not distributed - on the same pserver, only change param/grad varnames for - trainers to fetch. + Get parameter server side program. + + Args: + endpoint (str): current parameter server endpoint. + + Returns: + Program: the program for current parameter server to run. """ + # TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers. + # NOTE: assume blocks of the same variable is not distributed + # on the same pserver, only change param/grad varnames for + # trainers to fetch. + # step1 pserver_program = Program() + pserver_program.random_seed = self.origin_program.random_seed # step2: Create vars to receive vars at parameter servers. recv_inputs = [] for v in self.param_grad_ep_mapping[endpoint]["params"]: @@ -397,7 +447,7 @@ class DistributeTranspiler: dtype=v.dtype, shape=v.shape) if self.sync_mode and self.trainer_num > 1: - for trainer_id in xrange(self.trainer_num): + for trainer_id in range(self.trainer_num): var = pserver_program.global_block().create_var( name="%s.trainer_%d" % (orig_var_name, trainer_id), persistable=False, @@ -427,18 +477,14 @@ class DistributeTranspiler: # append it into the sub program. global_ops = [] - # HACK: optimization global ops only used to scale beta1 and beta2 - # replace it with dependency engine. - for op in self.optimize_ops: - if self._is_adam_connected_op(op): - global_ops.append(op) - def __append_optimize_op__(op, block, grad_to_block_id, merged_var): + def __append_optimize_op__(op, block, grad_to_block_id, merged_var, + lr_ops): if self._is_optimizer_op(op): self._append_pserver_ops(block, op, endpoint, grad_to_block_id, self.origin_program, merged_var) - else: - self._append_pserver_non_opt_ops(block, op, endpoint) + elif op not in lr_ops: + self._append_pserver_non_opt_ops(block, op) def __op_have_grad_input__(op): for varname in op.input_arg_names: @@ -446,20 +492,53 @@ class DistributeTranspiler: return varname return "" + def __clone_lr_op_sub_block__(op, program, lr_block): + if not op.has_attr('sub_block'): + return + + origin_block_desc = op.attr('sub_block') + origin_block = self.origin_program.block(origin_block_desc.id) + assert isinstance(origin_block, Block) + # we put the new sub block to new block to follow the block + # hierarchy of the original blocks + new_sub_block = program.create_block(lr_block.idx) + + # clone vars + for var in origin_block.vars: + new_sub_block._clone_variable(var) + + # clone ops + for origin_op in origin_block.ops: + cloned_op = self._clone_lr_op(program, new_sub_block, origin_op) + # clone sub_block of op + __clone_lr_op_sub_block__(cloned_op, program, new_sub_block) + + # reset the block of op + op.set_attr('sub_block', new_sub_block) + # append lr decay ops to the child block if exists lr_ops = self._get_lr_ops() + # record optimize blocks and we can run them on pserver parallel + optimize_blocks = [] if len(lr_ops) > 0: lr_decay_block = pserver_program.create_block( pserver_program.num_blocks - 1) + optimize_blocks.append(lr_decay_block) for _, op in enumerate(lr_ops): - self._append_pserver_non_opt_ops(lr_decay_block, op, endpoint) + cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op) + # append sub blocks to pserver_program in lr_decay_op + __clone_lr_op_sub_block__(cloned_op, pserver_program, + lr_decay_block) # append op to the current block grad_to_block_id = [] pre_block_idx = pserver_program.num_blocks - 1 for idx, opt_op in enumerate(opt_op_on_pserver): per_opt_block = pserver_program.create_block(pre_block_idx) + optimize_blocks.append(per_opt_block) # append grad merging ops before clip and weight decay + # cases may like: + # L2Decay op -> clip op -> optimize for _, op in enumerate(self.optimize_ops): # find the origin @GRAD var before clipping grad_varname_for_block = __op_have_grad_input__(op) @@ -467,62 +546,90 @@ class DistributeTranspiler: merged_var = self._append_pserver_grad_merge_ops( per_opt_block, grad_varname_for_block, endpoint, grad_to_block_id, self.origin_program) + break # append optimize op once then append other ops. for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and op not in global_ops: __append_optimize_op__(op, per_opt_block, grad_to_block_id, - merged_var) + merged_var, lr_ops) + # dedup grad to ids list + grad_to_block_id = list(set(grad_to_block_id)) # append global ops if global_ops: opt_state_block = pserver_program.create_block( pserver_program.num_blocks - 1) + optimize_blocks.append(opt_state_block) for glb_op in global_ops: __append_optimize_op__(glb_op, opt_state_block, - grad_to_block_id, None) + grad_to_block_id, None, lr_ops) # process distributed lookup_table - prefetch_block = None + prefetch_var_name_to_block_id = [] if self.has_distributed_lookup_table: pserver_index = self.pserver_endpoints.index(endpoint) table_opt_block = self._create_table_optimize_block( pserver_index, pserver_program, pre_block_idx, grad_to_block_id) - prefetch_block = self._create_prefetch_block( + optimize_blocks.append(table_opt_block) + prefetch_var_name_to_block_id = self._create_prefetch_block( pserver_index, pserver_program, table_opt_block) + checkpoint_block_id = self._create_checkpoint_save_block( + pserver_program, table_opt_block.idx) # NOTE: if has_distributed_lookup_table is False, then prefetch_block will # not be executed, so it's safe to use optimize_block to hold the place if self.has_distributed_lookup_table: - assert prefetch_block is not None + assert len(prefetch_var_name_to_block_id) > 0 else: - assert prefetch_block is None - prefetch_block = pserver_program.global_block() + assert len(prefetch_var_name_to_block_id) == 0 + + attrs = { + "optimize_blocks": optimize_blocks, + "endpoint": endpoint, + "Fanin": self.trainer_num, + "sync_mode": self.sync_mode, + "grad_to_block_id": grad_to_block_id, + } + if len(prefetch_var_name_to_block_id) > 0: + attrs['prefetch_var_name_to_block_id'] \ + = prefetch_var_name_to_block_id + attrs['checkpint_block_id'] = checkpoint_block_id # step5 append the listen_and_serv op pserver_program.global_block().append_op( type="listen_and_serv", inputs={'X': recv_inputs}, outputs={}, - attrs={ - "OptimizeBlock": pserver_program.block(1), - "endpoint": endpoint, - "Fanin": self.trainer_num, - "PrefetchBlock": prefetch_block, - "sync_mode": self.sync_mode, - "grad_to_block_id": grad_to_block_id - }) + attrs=attrs) - pserver_program.sync_with_cpp() + pserver_program._sync_with_cpp() return pserver_program - def get_startup_program(self, endpoint, pserver_program): + def get_startup_program(self, + endpoint, + pserver_program, + startup_program=None): """ Get startup program for current parameter server. Modify operator input variables if there are variables that were split to several blocks. + + Args: + endpoint (str): current pserver endpoint. + pserver_program (Program): call get_pserver_program first and + pass the result here. + startup_program (Program): if pass None, will use + default_startup_program + + Returns: + Program: parameter server side startup program. """ s_prog = Program() - orig_s_prog = default_startup_program() + if not startup_program: + orig_s_prog = default_startup_program() + else: + orig_s_prog = startup_program + s_prog.random_seed = orig_s_prog.random_seed params = self.param_grad_ep_mapping[endpoint]["params"] def _get_splited_name_and_shape(varname): @@ -535,48 +642,184 @@ class DistributeTranspiler: # 1. create vars in pserver program to startup program pserver_vars = pserver_program.global_block().vars created_var_map = dict() - for _, var in pserver_vars.iteritems(): - tmpvar = s_prog.global_block().clone_variable(var) + for _, var in list(pserver_vars.items()): + tmpvar = s_prog.global_block()._clone_variable(var) created_var_map[var.name] = tmpvar # 2. rename op outputs for op in orig_s_prog.global_block().ops: - new_inputs = dict() new_outputs = dict() # do not append startup op if var is not on this pserver op_on_pserver = False - for key in op.output_names: - newname, _ = _get_splited_name_and_shape(op.output(key)[0]) - if newname: - op_on_pserver = True - new_outputs[key] = created_var_map[newname] - elif op.output(key)[0] in pserver_vars: - op_on_pserver = True - new_outputs[key] = pserver_vars[op.output(key)[0]] - - # most startup program ops have no inputs - new_inputs = self._get_input_map_from_op(pserver_vars, op) + # TODO(gongwb): remove this line. + if op.type not in ["recv", "fetch_barrier", "concat"]: + for key in op.output_names: + newname, _ = _get_splited_name_and_shape(op.output(key)[0]) + if newname: + op_on_pserver = True + new_outputs[key] = created_var_map[newname] + elif op.output(key)[0] in pserver_vars: + op_on_pserver = True + new_outputs[key] = pserver_vars[op.output(key)[0]] if op_on_pserver: + # most startup program ops have no inputs + new_inputs = self._get_input_map_from_op(pserver_vars, op) + if op.type in [ "gaussian_random", "fill_constant", "uniform_random" ]: - op.attrs["shape"] = new_outputs["Out"].shape + op.set_attr("shape", list(new_outputs["Out"].shape)) s_prog.global_block().append_op( type=op.type, inputs=new_inputs, outputs=new_outputs, - attrs=op.attrs) + attrs=op.all_attrs()) return s_prog # ====================== private transpiler functions ===================== + def _has_distributed_lookup_table(self): + # process lookup_table_op + # 1. check all lookup_table_op is distributed + # 2. check all lookup_table_op share the same table. + distributed_lookup_table_ops = [] + # support only one distributed_lookup_table now + self.table_name = None + for op in self.origin_program.global_block().ops: + if op.type == LOOKUP_TABLE_TYPE: + if op.attr('is_distributed') is True: + if self.table_name is None: + self.table_name = op.input("W")[0] + if self.table_name != op.input("W")[0]: + raise RuntimeError("all distributed lookup_table_ops" + " should have only one table") + distributed_lookup_table_ops.append(op) + else: + if self.table_name is not None: + assert op.input("W")[0] != self.table_name + + return len(distributed_lookup_table_ops) > 0 + + def _update_dist_lookup_table_vars(self, param_list, grad_list, + params_grads): + # TODO(wuyi): put find a way to put dist lookup table stuff all together. + # update self.table_param_grad and self.trainer_side_table_grad_list + program = self.origin_program + if self.has_distributed_lookup_table: + param_list = [ + param for param in param_list if param.name != self.table_name + ] + grad_list = [ + grad for grad in grad_list + if grad.name != grad_var_name(self.table_name) + ] + self.table_param_grad = [ + param_grad for param_grad in params_grads + if param_grad[0].name == self.table_name + ][0] + table_grad_var = self.table_param_grad[1] + if self.sync_mode: + self.trainer_side_table_grad_list = [ + program.global_block().create_var( + name="%s.trainer_%d.pserver_%d" % + (table_grad_var.name, self.trainer_id, index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(len(self.pserver_endpoints)) + ] + else: + self.trainer_side_table_grad_list = [ + program.global_block().create_var( + name="%s.pserver_%d" % (table_grad_var.name, index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(len(self.pserver_endpoints)) + ] + return param_list, grad_list + + def _init_splited_vars(self): + # update these mappings for further transpile: + # 1. param_var_mapping: param var name -> [splited params vars] + # 2. grad_var_mapping: grad var name -> [splited grads vars] + # 3. grad_param_mapping: grad.blockx -> param.blockx + # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} + + param_list = [] + grad_list = [] + param_grad_set = set() + for p, g in self.params_grads: + # skip parameter marked not trainable + if type(p) == Parameter and p.trainable == False: + continue + if p.name not in param_grad_set: + param_list.append(p) + param_grad_set.add(p.name) + if g.name not in param_grad_set: + grad_list.append(g) + param_grad_set.add(g.name) + + param_list, grad_list = self._update_dist_lookup_table_vars( + param_list, grad_list, self.params_grads) + + if self.config.slice_var_up: + # when we slice var up into blocks, we will slice the var according to + # pserver services' count. A pserver may have two or more listening ports. + grad_blocks = slice_variable(grad_list, + len(self.pserver_endpoints), + self.config.min_block_size) + param_blocks = slice_variable(param_list, + len(self.pserver_endpoints), + self.config.min_block_size) + else: + # when we do NOT slice var up into blocks, we will always slice params + # grads into one block. + grad_blocks = slice_variable(grad_list, 1, + self.config.min_block_size) + param_blocks = slice_variable(param_list, 1, + self.config.min_block_size) + assert (len(grad_blocks) == len(param_blocks)) + + # origin_varname -> [splited_var] + self.param_var_mapping = self._create_vars_from_blocklist( + self.origin_program, param_blocks) + self.grad_var_mapping = self._create_vars_from_blocklist( + self.origin_program, + grad_blocks, + add_trainer_suffix=self.trainer_num > 1) + self.grad_param_mapping = dict() + for g, p in zip(grad_blocks, param_blocks): + g_name, g_bid, _ = g.split(":") + p_name, p_bid, _ = p.split(":") + self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ + self.param_var_mapping[p_name][int(p_bid)] + + # create mapping of endpoint -> split var to create pserver side program + self.param_grad_ep_mapping = dict() + [ + self.param_grad_ep_mapping.update({ + ep: { + "params": [], + "grads": [] + } + }) for ep in self.pserver_endpoints + ] + # transpiler function for dis lookup_table def _replace_lookup_table_op_with_prefetch(self, program, pserver_endpoints): # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op - self.prefetch_input_vars = None - self.prefetch_output_vars = None + # self.all_prefetch_input_vars = + # [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1] + # [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]] + self.all_prefetch_input_vars = [] + + # self.all_prefetch_input_vars = + # [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1] + # [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]] + self.all_prefetch_output_vars = [] continue_search_lookup_table_op = True while continue_search_lookup_table_op: @@ -586,26 +829,27 @@ class DistributeTranspiler: if op.type == LOOKUP_TABLE_TYPE: continue_search_lookup_table_op = True - op_index = list(all_ops).index(op) + lookup_table_op_index = list(all_ops).index(op) ids_name = op.input("Ids") out_name = op.output("Out") - if self.prefetch_input_vars is None: - ids_var = program.global_block().vars[ids_name[0]] - self.prefetch_input_vars = self.create_splited_vars( - source_var=ids_var, - block=program.global_block(), - tag="_prefetch_in_") - if self.prefetch_output_vars is None: - out_var = program.global_block().vars[out_name[0]] - self.prefetch_output_vars = self.create_splited_vars( - source_var=out_var, - block=program.global_block(), - tag="_prefetch_out_") + ids_var = program.global_block().vars[ids_name[0]] + prefetch_input_vars = self._create_splited_vars( + source_var=ids_var, + block=program.global_block(), + tag="_prefetch_in_") + self.all_prefetch_input_vars.append(prefetch_input_vars) + + out_var = program.global_block().vars[out_name[0]] + prefetch_output_vars = self._create_splited_vars( + source_var=out_var, + block=program.global_block(), + tag="_prefetch_out_") + self.all_prefetch_output_vars.append(prefetch_output_vars) # insert split_ids_op - program.global_block().insert_op( - index=op_index, + program.global_block()._insert_op( + index=lookup_table_op_index, type="split_ids", inputs={ 'Ids': [ @@ -613,31 +857,38 @@ class DistributeTranspiler: for varname in ids_name ] }, - outputs={"Out": self.prefetch_input_vars}) + outputs={"Out": prefetch_input_vars}) # insert prefetch_op - program.global_block().insert_op( - index=op_index + 1, + program.global_block()._insert_op( + index=lookup_table_op_index + 1, type="prefetch", - inputs={'X': self.prefetch_input_vars}, - outputs={"Out": self.prefetch_output_vars}, + inputs={'X': prefetch_input_vars}, + outputs={"Out": prefetch_output_vars}, attrs={ "epmap": pserver_endpoints, - RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + # FIXME(qiao) temporarily disable this config because prefetch + # is not act as other rpc op, it's more like a forward op + # RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) # insert concat_op - program.global_block().insert_op( - index=op_index + 2, - type="concat", - inputs={'X': self.prefetch_output_vars}, + program.global_block()._insert_op( + index=lookup_table_op_index + 2, + type="merge_ids", + inputs={ + 'Ids': [ + program.global_block().vars[varname] + for varname in ids_name + ], + 'X': prefetch_output_vars + }, outputs={ "Out": [ program.global_block().vars[varname] for varname in out_name ] - }, - attrs={"axis": 0}) + }) # delete lookup_table_op delete_ops(program.global_block(), [op]) @@ -645,7 +896,7 @@ class DistributeTranspiler: break def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints): - # 2. add split_ids_op and send_vars_op to send gradient to pservers + # 2. add split_ids_op and send_op to send gradient to pservers # there should only be one table_name all_ops = program.global_block().ops table_grad_name = grad_var_name(self.table_name) @@ -653,20 +904,20 @@ class DistributeTranspiler: if table_grad_name in op.output_arg_names: op_index = list(all_ops).index(op) # insert split_ids_op - program.global_block().insert_op( + program.global_block()._insert_op( index=op_index + 1, type="split_ids", inputs={ 'Ids': [program.global_block().vars[table_grad_name]] }, outputs={"Out": self.trainer_side_table_grad_list}) - program.global_block().insert_op( + program.global_block()._insert_op( index=op_index + 2, - type="send_vars", + type="send", inputs={'X': self.trainer_side_table_grad_list}, outputs={}, attrs={ - "sync_send": True, + "sync_mode": True, "epmap": pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) @@ -676,30 +927,34 @@ class DistributeTranspiler: optimize_block): # STEP: create prefetch block table_var = pserver_program.global_block().vars[self.table_name] - prefetch_block = pserver_program.create_block(optimize_block.idx) - trainer_ids = self.prefetch_input_vars[pserver_index] - pserver_ids = pserver_program.global_block().create_var( - name=trainer_ids.name, - type=trainer_ids.type, - shape=trainer_ids.shape, - dtype=trainer_ids.dtype) - trainer_out = self.prefetch_output_vars[pserver_index] - pserver_out = pserver_program.global_block().create_var( - name=trainer_out.name, - type=trainer_out.type, - shape=trainer_out.shape, - dtype=trainer_out.dtype) - prefetch_block.append_op( - type="lookup_sparse_table", - inputs={'Ids': pserver_ids, - "W": table_var}, - outputs={"Out": pserver_out}, - attrs={ - "is_sparse": True, # has no effect on lookup_table op - "is_distributed": True, - "padding_idx": -1 - }) - return prefetch_block + prefetch_var_name_to_block_id = [] + for index in range(len(self.all_prefetch_input_vars)): + prefetch_block = pserver_program.create_block(optimize_block.idx) + trainer_ids = self.all_prefetch_input_vars[index][pserver_index] + pserver_ids = pserver_program.global_block().create_var( + name=trainer_ids.name, + type=trainer_ids.type, + shape=trainer_ids.shape, + dtype=trainer_ids.dtype) + trainer_out = self.all_prefetch_output_vars[index][pserver_index] + pserver_out = pserver_program.global_block().create_var( + name=trainer_out.name, + type=trainer_out.type, + shape=trainer_out.shape, + dtype=trainer_out.dtype) + prefetch_block.append_op( + type="lookup_sparse_table", + inputs={'Ids': pserver_ids, + "W": table_var}, + outputs={"Out": pserver_out}, + attrs={ + "is_sparse": True, # has no effect on lookup_table op + "is_distributed": True, + "padding_idx": -1 + }) + prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str( + prefetch_block.idx)) + return prefetch_var_name_to_block_id def _create_table_optimize_block(self, pserver_index, pserver_program, pre_block_idx, grad_to_block_id): @@ -707,26 +962,31 @@ class DistributeTranspiler: # create table param and grad var in pserver program origin_param_var = self.origin_program.global_block().vars[ self.table_name] + + zero_dim = int( + math.ceil(origin_param_var.shape[0] / len(self.pserver_endpoints))) + table_shape = list(origin_param_var.shape) + table_shape[0] = zero_dim + param_var = pserver_program.global_block().create_var( name=origin_param_var.name, - shape=origin_param_var.shape, + shape=table_shape, dtype=origin_param_var.dtype, type=core.VarDesc.VarType.SELECTED_ROWS, persistable=True) # parameter must be selected rows param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS) - grad_var = pserver_program.global_block().clone_variable( + grad_var = pserver_program.global_block()._clone_variable( self.origin_program.global_block().vars[grad_var_name( self.table_name)]) # create table optimize block in pserver program table_opt_op = [ op for op in self.optimize_ops - if op.input("Param")[0] == self.table_name + if 'Param' in op.input_names and op.input("Param")[0] == + self.table_name ][0] table_opt_block = pserver_program.create_block(pre_block_idx) - # only support sgd now - assert table_opt_op.type == "sgd" if self.sync_mode: # create grad vars in pserver program @@ -745,7 +1005,8 @@ class DistributeTranspiler: table_opt_block.append_op( type="sum", inputs={"X": pserver_side_table_grad_list}, - outputs={"Out": [grad_var]}) + outputs={"Out": [grad_var]}, + attrs={"use_mkldnn": False}) else: # in async_mode, for table gradient, it also need to be splited to each parameter server origin_grad_name = grad_var.name @@ -754,7 +1015,7 @@ class DistributeTranspiler: if not splited_grad_name.startswith(origin_grad_name): raise ValueError("origin_grad_var: " + splited_grad_name + " grad_var:" + grad_var.name) - grad_var = pserver_program.global_block().rename_var( + grad_var = pserver_program.global_block()._rename_var( origin_grad_name, splited_grad_name) lr_var = pserver_program.global_block().vars[table_opt_op.input( @@ -765,17 +1026,39 @@ class DistributeTranspiler: "LearningRate": [lr_var] } outputs = {"ParamOut": [param_var]} - table_opt_block.append_op( - type=table_opt_op.type, - inputs=inputs, - outputs=outputs, - attrs=table_opt_op.attrs) + # only support sgd now + import logging + logging.warn( + "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of " + + table_opt_op.type) + table_opt_block.append_op(type="sgd", inputs=inputs, outputs=outputs) # add table parameter gradient and it's block id to grad_to_block_id grad_to_block_id.append(grad_var.name + ":" + str(table_opt_block.idx)) return table_opt_block + def _create_checkpoint_save_block(self, pserver_program, pre_block_idx): + """ + create a new block to handle save checkpoint. + """ + import os + + pserver_program.global_block().create_var( + name="kLookupTablePath", + persistable=True, + type=core.VarDesc.VarType.RAW) + + checkpoint_save_block = pserver_program.create_block(pre_block_idx) + # this 'file_path' do not be used in save lookup table variable + checkpoint_save_block.append_op( + type='save', + inputs={'X': [self.table_name]}, + outputs={}, + attrs={'file_path': "none"}) + + return checkpoint_save_block.idx + def _create_vars_from_blocklist(self, program, block_list, @@ -788,8 +1071,8 @@ class DistributeTranspiler: program (ProgramDesc): ProgramDesc which gradients blong. block_list (list[(varname, block_id, block_size)]): List of gradient blocks. add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True. - Returns: - var_mapping (dict(varname->[new_varname_variable])):A dict mapping + Returns: + var_mapping (dict(varname->[new_varname_variable])):A dict mapping from original var name to each var split. """ @@ -799,23 +1082,23 @@ class DistributeTranspiler: var_mapping = dict() for block_str in block_list: varname, offset, size = block_str.split(":") - if not block_map.has_key(varname): + if varname not in block_map: block_map[varname] = [] - block_map[varname].append((long(offset), long(size))) - for varname, splited in block_map.iteritems(): + block_map[varname].append((int(offset), int(size))) + + for varname, splited in list(block_map.items()): orig_var = program.global_block().var(varname) if len(splited) == 1: if self.sync_mode and add_trainer_suffix: new_var_name = "%s.trainer_%d" % \ (orig_var.name, self.trainer_id) - program.global_block().rename_var(varname, new_var_name) + program.global_block()._rename_var(varname, new_var_name) var_mapping[varname] = \ [program.global_block().var(new_var_name)] else: var_mapping[varname] = \ [program.global_block().var(orig_var.name)] continue - var_mapping[varname] = [] orig_shape = orig_var.shape orig_dim1_flatten = 1 @@ -842,10 +1125,10 @@ class DistributeTranspiler: type=orig_var.type, shape=splited_shape) # flattend splited var var_mapping[varname].append(var) - program.global_block().sync_with_cpp() + program.global_block()._sync_with_cpp() return var_mapping - def create_splited_vars(self, source_var, block, tag): + def _create_splited_vars(self, source_var, block, tag): return [ block.create_var( name=str(source_var.name + tag + str(index)), @@ -856,7 +1139,6 @@ class DistributeTranspiler: ] def _clone_var(self, block, var, persistable=True): - assert isinstance(var, Variable) return block.create_var( name=var.name, shape=var.shape, @@ -870,7 +1152,7 @@ class DistributeTranspiler: height_sections = [] for v in splited_vars: height_sections.append(v.shape[0]) - program.global_block().insert_op( + program.global_block()._insert_op( index=index + 1, type="split_selected_rows", inputs={"X": orig_var}, @@ -880,7 +1162,7 @@ class DistributeTranspiler: sections = [] for v in splited_vars: sections.append(v.shape[0]) - program.global_block().insert_op( + program.global_block()._insert_op( index=index + 1, type="split_byref", inputs={"X": orig_var}, @@ -966,7 +1248,7 @@ class DistributeTranspiler: grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx)) if self.sync_mode and self.trainer_num > 1: vars2merge = [] - for i in xrange(self.trainer_num): + for i in range(self.trainer_num): per_trainer_name = "%s.trainer_%d" % \ (merged_var_name, i) vars2merge.append(pserver_block.vars[per_trainer_name]) @@ -974,7 +1256,8 @@ class DistributeTranspiler: optimize_block.append_op( type="sum", inputs={"X": vars2merge}, - outputs={"Out": merged_var}) + outputs={"Out": merged_var}, + attrs={"use_mkldnn": False}) # TODO(panyx0718): What if it's SELECTED_ROWS. if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: optimize_block.append_op( @@ -989,18 +1272,39 @@ class DistributeTranspiler: program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() + # update param/grad shape first, then other inputs like # moment can use the updated shape + def _get_param_block(opt_op): + # param is already created on global program + param_block = None + for p in self.param_grad_ep_mapping[endpoint]["params"]: + if same_or_split_var(p.name, opt_op.input("Param")[0]): + param_block = p + break + return param_block + for key in opt_op.input_names: if key == "Grad": new_inputs[key] = merged_var + # For RMSProp optimizer + elif key == "Moment" or key == "MeanSquare": + param_block = _get_param_block(opt_op) + if not param_block: + return + moment_var = origin_program.global_block().vars[opt_op.input( + key)[0]] + tmpvar = pserver_block.create_var( + name=moment_var.name, + persistable=moment_var.persistable, + dtype=moment_var.dtype, + # change to use same shape as param + # TODO(typhoonzero): didn't append .block in the var name, + # may affect checkpoint saving? Need to verify. + shape=param_block.shape) + new_inputs[key] = tmpvar elif key == "Param": - # param is already created on global program - param_block = None - for p in self.param_grad_ep_mapping[endpoint]["params"]: - if same_or_split_var(p.name, opt_op.input(key)[0]): - param_block = p - break + param_block = _get_param_block(opt_op) if not param_block: return tmpvar = pserver_block.create_var( @@ -1013,7 +1317,7 @@ class DistributeTranspiler: # learning rate variable has already be created by non-optimize op, # don't create it once again. lr_varname = opt_op.input(key)[0] - if pserver_block.vars.has_key(lr_varname): + if lr_varname in pserver_block.vars: new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]] else: origin_var = origin_program.global_block().vars[lr_varname] @@ -1026,7 +1330,7 @@ class DistributeTranspiler: for key in opt_op.input_names: new_shape = None - if key in ["Param", "Grad", "LearningRate"]: + if key in ["Param", "Grad", "LearningRate", "Moment", "MeanSquare"]: continue var = self.origin_program.global_block().vars[opt_op.input(key)[0]] # update accumulator variable shape @@ -1049,23 +1353,47 @@ class DistributeTranspiler: type=opt_op.type, inputs=new_inputs, outputs=outputs, - attrs=opt_op.attrs) + attrs=opt_op.all_attrs()) def _is_splited_grad_var(self, var, var_dict): grad_block = None - for _, g in var_dict.iteritems(): + # TODO(minqiyang): replace these items() with six.iteritems() to + # improve memory + for _, g in list(var_dict.items()): if self._orig_varname(g.name) == self._orig_varname(var.name): if g.name.find(".trainer_") == -1: grad_block = g break return grad_block - def _append_pserver_non_opt_ops(self, optimize_block, opt_op, endpoint): + def _clone_lr_op(self, program, block, op): + inputs = self._get_input_map_from_op( + self.origin_program.global_block().vars, op) + for key, varlist in list(inputs.items()): + if not isinstance(varlist, list): + varlist = [varlist] + for var in varlist: + if var not in program.global_block().vars: + block._clone_variable(var) + + outputs = self._get_output_map_from_op( + self.origin_program.global_block().vars, op) + for key, varlist in list(outputs.items()): + if not isinstance(varlist, list): + varlist = [varlist] + for var in varlist: + if var not in program.global_block().vars: + block._clone_variable(var) + + return block.append_op( + type=op.type, inputs=inputs, outputs=outputs, attrs=op.all_attrs()) + + def _append_pserver_non_opt_ops(self, optimize_block, opt_op): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, opt_op) - for key, varlist in inputs.iteritems(): + for key, varlist in list(inputs.items()): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: @@ -1075,7 +1403,7 @@ class DistributeTranspiler: var, program.global_block().vars) if grad_block: inputs[key] = grad_block - elif not program.global_block().vars.has_key(var.name): + elif var.name not in program.global_block().vars: program.global_block().create_var( name=var.name, persistable=var.persistable, @@ -1084,7 +1412,7 @@ class DistributeTranspiler: outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) - for key, varlist in outputs.iteritems(): + for key, varlist in list(outputs.items()): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: @@ -1092,63 +1420,35 @@ class DistributeTranspiler: var, program.global_block().vars) if grad_block: outputs[key] = grad_block - elif not program.global_block().vars.has_key(var.name): - program.global_block().clone_variable(var) + elif var.name not in program.global_block().vars: + program.global_block()._clone_variable(var) - optimize_block.append_op( + return optimize_block.append_op( type=opt_op.type, inputs=inputs, outputs=outputs, - attrs=opt_op.attrs) + attrs=opt_op.all_attrs()) def _is_op_connected(self, op1, op2): # If one op's input is another op's output or # one op's output is another op's input, we say # the two operator is connected. - def _append_inname_remove_beta(varname_list): - op_input_names = [] - for in_name in varname_list: - # HACK: remove beta1 and beta2 to avoid let all - # ops connected. - if in_name.startswith("beta2_pow_acc") or \ - in_name.startswith("beta1_pow_acc"): - continue - else: - op_input_names.append(in_name) - return op_input_names - - op1_input_names = _append_inname_remove_beta(op1.desc.input_arg_names()) - op1_output_names = op1.desc.output_arg_names() - - op2_input_names = _append_inname_remove_beta(op2.desc.input_arg_names()) - op2_output_names = op2.desc.output_arg_names() - - if set(op1_output_names) & set(op2_input_names) or \ - set(op1_input_names) & set(op2_output_names): + if set(op1.desc.output_arg_names()) & set(op2.desc.input_arg_names()) or \ + set(op1.desc.input_arg_names()) & set(op2.desc.output_arg_names()): return True return False def _create_ufind(self, optimize_ops): # Create a unit find data struct by optimize ops ufind = UnionFind(optimize_ops) - for i in xrange(len(optimize_ops)): - for j in xrange(i, len(optimize_ops)): + for i in range(len(optimize_ops)): + for j in range(i, len(optimize_ops)): op1 = optimize_ops[i] op2 = optimize_ops[j] if self._is_op_connected(op1, op2): ufind.union(op1, op2) return ufind - def _is_opt_role_op(self, op): - # NOTE: depend on oprole to find out whether this op is for - # optimize - op_maker = core.op_proto_and_checker_maker - optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize - if op_maker.kOpRoleAttrName() in op.attrs and \ - int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role): - return True - return False - def _is_optimizer_op(self, op): if "Param" in op.input_names and \ "LearningRate" in op.input_names: @@ -1227,9 +1527,19 @@ class DistributeTranspiler: break return lr_ops + def _is_opt_role_op(self, op): + # NOTE: depend on oprole to find out whether this op is for + # optimize + op_maker = core.op_proto_and_checker_maker + optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize + if op_maker.kOpRoleAttrName() in op.attr_names and \ + int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role): + return True + return False + def _get_optimize_pass(self): """ - Get optimizer operators, paramters and gradients from origin_program + Get optimizer operators, parameters and gradients from origin_program Returns: opt_ops (list): optimize operators. params_grads (dict): paramter->gradient. @@ -1246,26 +1556,12 @@ class DistributeTranspiler: # and op_role_var to get the pair. for input_name in op.input_arg_names: if input_name.find("@GRAD") != -1 and \ - op.attrs[RPC_OP_ROLE_ATTR_NAME]: - param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0] + op.attr(RPC_OP_ROLE_ATTR_NAME): + param_name = op.attr(OP_ROLE_VAR_ATTR_NAME)[0] params_grads.append([ origin_var_dict[param_name], origin_var_dict[input_name] ]) - elif self._is_adam_connected_op(op): - opt_ops.append(op) else: pass return opt_ops, params_grads - - def _is_adam_connected_op(self, op): - """ - A hack function to determinate whether the input operator - is connected to optimize operator. - """ - if op.type == "scale": - for in_name in op.input_arg_names: - if in_name.startswith("beta1_pow_acc") or \ - in_name.startswith("beta2_pow_acc"): - return True - return False diff --git a/python/paddle/fluid/transpiler/distribute_transpiler_simple.py b/python/paddle/fluid/transpiler/distribute_transpiler_simple.py deleted file mode 100644 index ea8c27cdca885dbbf90349b35df9691951264061..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/transpiler/distribute_transpiler_simple.py +++ /dev/null @@ -1,254 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from ..framework import Program, default_main_program, Parameter, Variable -from ..layer_helper import LayerHelper - - -def hash_name_to_server(params_grads, pserver_endpoints): - """ - :param param_grads: - :return: a map of pserver endpoint -> - params -> [param list] - grads -> [grad list] - """ - - def _hash_param(param_name, total): - return hash(param_name) % total - - param_grad_map = dict() - for param, grad in params_grads: - if param.trainable is True and grad is not None: - server_id = _hash_param(param.name, len(pserver_endpoints)) - server_for_param = pserver_endpoints[server_id] - if not param_grad_map.has_key(server_for_param): - param_grad_map[server_for_param] = {"params": [], "grads": []} - param_grad_map[server_for_param]["params"].append(param) - param_grad_map[server_for_param]["grads"].append(grad) - - return param_grad_map - - -def round_robin(params_grads, pserver_endpoints): - assert (len(params_grads) > len(pserver_endpoints)) - - param_grad_map = dict() - pserver_idx = 0 - for param, grad in params_grads: - if param.trainable is True: - server_for_param = pserver_endpoints[pserver_idx] - if not param_grad_map.has_key(server_for_param): - param_grad_map[server_for_param] = {"params": [], "grads": []} - - param_grad_map[server_for_param]["params"].append(param) - param_grad_map[server_for_param]["grads"].append(grad) - - pserver_idx += 1 - if pserver_idx >= len(pserver_endpoints): - pserver_idx = 0 - return param_grad_map - - -class SimpleDistributeTranspiler: - def transpile(self, - optimize_ops, - params_grads, - program=None, - pservers="127.0.0.1:6174", - trainers=1, - split_method=round_robin): - """ - Transpile the program to a distributed data-parallelism programs. - - The main_program will be transform to use a remote parameter server - to do parameter optimization. And the optimization graph will be put - in to a parameter server program. - - Use different methods to split trainable varialbles to different - parameter servers. - - Example to run: - - exe = fluid.Executor(place) - t = fluid.DistributeTranspiler() - t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1) - - pserver_endpoint = os.getenv("PSERVER") - if pserver_endpoint: - pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops) - exe.run(fluid.default_startup_program()) - exe.run(pserver_prog) - else: - feeder = fluid.DataFeeder(feed_list=[images, label], place=place) - exe.run(fluid.default_startup_program()) - - for pass_id in range(PASS_NUM): - ... - - :param optimize_ops: op list of optimization, should be the - return value of Optimizer.minimize - :type optimize_ops: list - :param program: program to optimize, default default_main_program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - - :return: return a list of programs - """ - if program is None: - program = default_main_program() - self.program = program - self.trainers = trainers - self.optimize_ops = optimize_ops - self._optimize_distributed( - optimize_ops, - program, - params_grads, - pservers=pservers, - trainers=trainers, - split_method=split_method) - - def _clone_param(self, block, v): - assert isinstance(v, Parameter) - new_p = Parameter( - block=block, - shape=v.shape, - dtype=v.dtype, - type=v.type, - lod_level=v.lod_level, - stop_gradient=v.stop_gradient, - trainable=v.trainable, - optimize_attr=v.optimize_attr, - regularizer=v.regularizer, - name=v.name) - block.vars[new_p.name] = new_p - - def _clone_var(self, block, var): - assert isinstance(var, Variable) - return block.create_var( - name=var.name, - shape=var.shape, - dtype=var.dtype, - type=var.type, - lod_level=var.lod_level, - persistable=var.persistable) - - def _optimize_distributed(self, optimize_ops, program, params_and_grads, - **kwargs): - if kwargs.has_key("split_method"): - split_method = kwargs["split_method"] - else: - split_method = round_robin - - assert (callable(split_method)) - pserver_endpoints = kwargs["pservers"].split(",") - self.param_grad_map = split_method(params_and_grads, pserver_endpoints) - - send_op_ordered_inputs = [] - send_op_ordered_outputs = [] - epmap = [] - for ep, v in self.param_grad_map.iteritems(): - send_op_ordered_inputs.extend(v["grads"]) - send_op_ordered_outputs.extend(v["params"]) - for i in v["grads"]: - epmap.append(ep) - send_op = program.global_block().append_op( - type="send", - inputs={"X": send_op_ordered_inputs - }, # inputs is a list of tensors to be send - outputs={"Out": send_op_ordered_outputs}, - attrs={"endpoints": pserver_endpoints, - "epmap": epmap}) - - def get_trainer_program(self): - # remove optimize ops and add a send op to main_program - self.program.global_block().delete_ops(self.optimize_ops) - return self.program - - def _create_var_for_trainers(self, block, var, trainers): - var_list = [] - for i in xrange(trainers): - var_each = block.create_var( - name="%s.trainer_%d" % (var.name, i), - psersistable=var.persistable, - dtype=var.dtype, - shape=var.shape) - var_list.append(var_each) - return var_list - - def get_pserver_program(self, endpoint, optimize_ops): - pserver_program = Program() - for v in self.param_grad_map[endpoint]["params"]: - self._clone_param(pserver_program.global_block(), v) - - optimize_sub_program = Program() - grad_var_names = [ - var.name for var in self.param_grad_map[endpoint]["grads"] - ] - for opt_op in optimize_ops: - for _, var in opt_op.inputs.iteritems(): - # NOTE: append operators to merge gradients from multiple - # trainers. If trainers == 1, this is not needed. - if self.trainers > 1 and var.name in grad_var_names: - vars2merge = self._create_var_for_trainers( - optimize_sub_program.global_block(), var, self.trainers) - merged_var = optimize_sub_program.global_block().create_var( - name=var.name, - persistable=var.persistable, - dtype=var.dtype, - shape=var.shape) - optimize_sub_program.global_block().append_op( - type="sum", - inputs={"X": vars2merge}, - outputs={"Out": merged_var}) - optimize_sub_program.global_block().append_op( - type="scale", - inputs={"X": merged_var}, - outputs={"Out": merged_var}, - attrs={"scale": 1.0 / float(self.trainers)}) - else: - optimize_sub_program.global_block().create_var( - name=var.name, - persistable=var.persistable, - dtype=var.dtype, - shape=var.shape) - - if opt_op.inputs.has_key("Grad"): - if opt_op.inputs["Grad"].name in grad_var_names: - optimize_sub_program.global_block().append_op( - type=opt_op.type, - inputs=opt_op.inputs, - outputs=opt_op.outputs, - attrs=opt_op.attrs) - else: - optimize_sub_program.global_block().append_op( - type=opt_op.type, - inputs=opt_op.inputs, - outputs=opt_op.outputs, - attrs=opt_op.attrs) - pserver_program.global_block().append_op( - type="recv", - inputs={"RX": - self.param_grad_map[endpoint]["grads"]}, # grads to recv - outputs={}, - attrs={ - "OptimizeBlock": optimize_sub_program.global_block(), - "endpoint": endpoint, - "ParamList": - [p.name for p in self.param_grad_map[endpoint]["params"]], - "GradList": - [p.name for p in self.param_grad_map[endpoint]["grads"]], - "Trainers": self.trainers - }) - pserver_program.sync_with_cpp() - return pserver_program diff --git a/python/paddle/fluid/transpiler/inference_transpiler.py b/python/paddle/fluid/transpiler/inference_transpiler.py index 202aa76084432b4b2378470919b2e924301f2130..87f20bbccf3138585841952efacef5b0a3cbbace 100644 --- a/python/paddle/fluid/transpiler/inference_transpiler.py +++ b/python/paddle/fluid/transpiler/inference_transpiler.py @@ -12,23 +12,41 @@ # See the License for the specific language governing permissions and # limitations under the License. +import os import numpy as np from .. import core from ..framework import Program from ..executor import global_scope -class InferenceTranspiler: +class InferenceTranspiler(object): + ''' + Convert the fluid program to optimized inference program. + + There are several optimizations: + + - fuse convolution and batch normalization + - fuse batch normalization and relu (MKLDNN only) + + Examples: + + .. code-block:: python + + # As InferenceTranspiler will modify the original program, + # please clone before use it. + inference_transpiler_program = program.clone() + t = fluid.InferenceTranspiler() + t.transpile(inference_transpiler_program, place) + ''' + def transpile(self, program, place, scope=None): ''' - Transpile the program. Support only fuse batch normalization now. + Run the transpiler. - :param program: program to transpile - :type program: Program - :param place: inference place - :type place: Place - :param scope: inference scope - :type scope: Scope or None + Args: + program (Program): program to transpile + place (Place): inference place + scope (Scope|None): inference Scope ''' if not isinstance(program, Program): raise TypeError("program should be as Program type") @@ -39,59 +57,111 @@ class InferenceTranspiler: scope = global_scope() if not isinstance(scope, core.Scope): raise TypeError("scope should be as Scope type or None") - self.fuse_batch_norm(program, place, scope) + self._fuse_batch_norm(program, place, scope) + self._fuse_relu_mkldnn(program) + + def _fuse_relu_mkldnn(self, program): + ''' + Transpile the program by fused relu activation for MKLDNN program. + + Relu activation following batch norm OP can be fused by adding + :math:`fuse_with_relu` attribute to batch norm OP. + + The result of fuse is: + + - before: + + - batch_norm->relu->any_other_op + + - after: + + - batch_norm->any_other_op + + :param program: program to transpile + :type program: Program + ''' + use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False)) + if not use_mkldnn: + return + + self.block = program.block(0) + + i = 0 + while i < len(self.block.ops) - 1: + current_op = self.block.ops[i] + if current_op.type in ['batch_norm']: + next_op = self.block.ops[i + 1] + if next_op.type == 'relu': + # modify bnorm OP to include relu + current_op.set_attr("fuse_with_relu", True) + # remove relu OP + self.block._remove_op(i + 1) + i = i + 1 - def fuse_batch_norm(self, program, place, scope): + self._remove_unused_var() + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. + # And a better solution will be considered later. + program = program.clone() + + def _fuse_batch_norm(self, program, place, scope): ''' Transpile the program by fused batch normalization. - - The batch normalization followed the convolution or fully connected layer - can be integrated with them. Doing so will give us a forward acceleration, + + The batch normalization followed the convolution or fully connected layer + can be integrated with them. Doing so will give us a forward acceleration, especially in environments like mobile or embedded. - - For input X: - - Conv process: X = input * W + bias - - Batch norm process: X' = (X - mean) / std - - Scale Process: Y = a * X' + b + + For input :math:`X`: + + - Conv process: :math:`X = input * W + bias` + - Batch norm process: :math:`X' = (X - mean) / std` + - Scale Process: :math:`Y = a * X' + b` After fuse into one operation: - Y = (input * W + bias - mean) / std * a + b - = input * a * W / std + ((bias - mean) / std * a + b) + .. math:: + + Y &= (input * W + bias - mean) / std * a + b \\\\ + &= input * a * W / std + ((bias - mean) / std * a + b) + + The operator transformation is: - The operator transformation is: - before: + - conv->batch_norm->any_other_op (bias == 0) - conv->elementwise_add->batch_norm->any_other_op (bias != 0) - - after: + + - after: + - conv->elementwise_add->any_other_op - + The transpile stages are: + 1. insert elementwise_add op when bias == 0. 2. fuse the batch_norm's parameters to conv and elementwise_add operators. 3. remove batch_norm ops which are not used in any other ops. 4. adjust the input of any_other_op to be the output of elementwise_add operator. 5. remove unused variables. - :param program: program to transpile - :type program: Program - :param place: inference place - :type place: Place - :param scope: inference scope - :type scope: Scope + Args: + program (Program): program to transpile + place (Place): inference place + scope (Scope): inference Scope + ''' self.scope = scope self.place = place self.block = program.block(0) - self.input_map = {} # store the input names should be adjusted + self.input_map = {} # store the input names should be adjusted i = 0 - while i < len(self.block.ops): + while i < len(self.block.ops) - 2: current_op = self.block.ops[i] # TODO(luotao1): consider only conv2d now. fc would be delt later. if current_op.type in ['conv2d']: - # TODO(luotao1): consider single chain network now. - # For branch network, we counldn't use block.ops[i + 1] as + # TODO(luotao1): consider single chain network now. + # For branch network, we counldn't use block.ops[i + 1] as # the judgment condition. next_op = self.block.ops[i + 1] # conv2d without bias @@ -101,7 +171,7 @@ class InferenceTranspiler: # fuse batch_norm self._fuse_param(current_op, next_op, bias_op, 0) # remove batch_norm_op - self.block.remove_op(i + 2) + self.block._remove_op(i + 2) i = i + 1 # conv2d with bias, the next_op.type is elementwise_add elif (next_op.type == 'elementwise_add'): @@ -110,23 +180,23 @@ class InferenceTranspiler: # fuse batch_norm self._fuse_param(current_op, next_next_op, next_op, 1) # remove batch_norm_op - self.block.remove_op(i + 2) + self.block._remove_op(i + 2) i = i + 1 i = i + 1 self._adjust_input() self._remove_unused_var() - # TODO(luotao): use clone() method to flush the program.desc in force, - # since some large program.desc will not be flushed immediately. + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. # And a better solution will be considered later. program = program.clone() # ====================== private transpiler functions ===================== def _insert_bias_op(self, index, current_op, bn_op): ''' - Construct elementwise_add operator for adding bias + Construct elementwise_add operator for adding bias and insert it into program. - + :param index: insert location of bias_op :type index: Int :param current_op: current operator (conv or fc) @@ -142,7 +212,7 @@ class InferenceTranspiler: y_var = self.block.var(bn_op.input("Bias")[0]) out_var = self.block.var(bn_op.output("Y")[0]) - bias_op = self.block.insert_op( + bias_op = self.block._insert_op( index, type="elementwise_add", inputs={"X": x_var, @@ -154,14 +224,14 @@ class InferenceTranspiler: def _fuse_param(self, current_op, bn_op, bias_op, with_bias): ''' fuse the batch_norm_op' parameters to current_op (conv or fc) - + :param current_op: current operator (conv or fc) :type current_op: Operator :param bn_op: batch norm operator :type bn_op: Operator :param bias_op: elementwise_add operator for adding bias :type bias_op: Operator - :param with_bias: If current operator has bias, with_bias = 1; otherwise 0. + :param with_bias: If current operator has bias, with_bias = 1; otherwise 0. :type with_bias: Int ''' @@ -235,6 +305,6 @@ class InferenceTranspiler: args += current_op.output_arg_names args = list(set(args)) # unique the input and output arguments - for var in self.block.vars.keys(): + for var in list(self.block.vars.keys()): if var not in args: - self.block.remove_var(var) + self.block._remove_var(var) diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 9ff0ae6fca27d4681891b2033e2f8f95bd825942..20ba7ed2b0b9df0d0432727ee1f69f61533c402e 100644 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -14,8 +14,10 @@ from collections import defaultdict from .. import core -from ..framework import Program, default_main_program, Parameter, Variable +from ..framework import Program, default_main_program, Parameter from ..backward import _rename_arg_ +from functools import reduce +from six.moves import range dtype_to_size = { core.VarDesc.VarType.FP16: 2, @@ -107,7 +109,7 @@ class ControlFlowGraph(object): # Repeatedly apply liveness updates until the algorithm stablize # on a complete set live input vars and live output vars. while True: - for i in reversed(range(self.op_size)): + for i in reversed(list(range(self.op_size))): live_in[i] = set(self._live_in[i]) live_out[i] = set(self._live_out[i]) for s in self._successors[i]: @@ -157,9 +159,11 @@ class ControlFlowGraph(object): if op.type() == "fill_constant" and op.attr("force_cpu") == True: self._skip_opt.update(op.output_arg_names()) - def release_memory(self): + def release_memory(self, skip_opt_set=None): self._dataflow_analyze() self._update_skip_opt_set() + if skip_opt_set: + self._skip_opt.update(skip_opt_set) fwd_id = 0 bwd_id = 0 for i in range(self.op_size): @@ -170,12 +174,13 @@ class ControlFlowGraph(object): is_forward = i < self._forward_num in_diff, out_diff = self._get_diff(self._live_in[i], self._live_out[i]) - can_optimize = filter( - lambda x: self._check_var_validity(block_desc, x, is_forward), - in_diff) + can_optimize = [ + x for x in in_diff + if self._check_var_validity(block_desc, x, is_forward) + ] if can_optimize: index = i + fwd_id + 1 if is_forward else i - self._forward_num + bwd_id + 1 - delete_op = block_desc.insert_op(index) + delete_op = block_desc._insert_op(index) delete_op.set_type("delete_var") delete_op.set_input("X", can_optimize) if is_forward: @@ -183,7 +188,7 @@ class ControlFlowGraph(object): else: bwd_id += 1 - def memory_optimize(self, level=0): + def memory_optimize(self, skip_opt_set=None, level=0): def compare_shape(x_shape, cache_shape, opt_level): if opt_level == 0: return x_shape == cache_shape @@ -200,6 +205,9 @@ class ControlFlowGraph(object): self._dataflow_analyze() self._update_skip_opt_set() + # update skip set to meet users' demand + if skip_opt_set: + self._skip_opt.update(skip_opt_set) self.pool = [] for i in range(self.op_size): op = self._ops[i] @@ -208,9 +216,10 @@ class ControlFlowGraph(object): block_desc = op.block() is_forward = i < self._forward_num if self.pool: - defs_can_optimize = filter( - lambda x: self._check_var_validity(block_desc, x, is_forward), - self._defs[i]) + defs_can_optimize = [ + x for x in self._defs[i] + if self._check_var_validity(block_desc, x, is_forward) + ] out_pair = [ (x, self._find_var(block_desc, x, is_forward).shape()) for x in defs_can_optimize @@ -256,9 +265,10 @@ class ControlFlowGraph(object): break in_diff, _ = self._get_diff(self._live_in[i], self._live_out[i]) - can_optimize = filter( - lambda x: self._check_var_validity(block_desc, x, is_forward), - in_diff) + can_optimize = [ + x for x in in_diff + if self._check_var_validity(block_desc, x, is_forward) + ] if can_optimize: for var_name in can_optimize: self.pool.append((var_name, self._find_var( @@ -319,6 +329,8 @@ def _process_sub_block_pair(pdesc, sub_block_pair): sub_op_output = set() sub_op_output.update(sub_op_dict[fwd_id].output_arg_names()) sub_op_output.update(sub_op_dict[grad_id].output_arg_names()) + sub_op_output.update(sub_op_dict[fwd_id].input_arg_names()) + sub_op_output.update(sub_op_dict[grad_id].input_arg_names()) ops_list.append((sub_block_ops, block_op_size, sub_op_output)) # Process rest fwd_op block ops @@ -330,6 +342,7 @@ def _process_sub_block_pair(pdesc, sub_block_pair): sub_block_ops.append(sub_block.op(i)) sub_op_output = set() sub_op_output.update(sub_op_dict[fwd_id].output_arg_names()) + sub_op_output.update(sub_op_dict[fwd_id].input_arg_names()) ops_list.append((sub_block_ops, sub_block_op_size, sub_op_output)) return ops_list @@ -344,13 +357,17 @@ def _get_cfgs(input_program): pdesc = input_program.get_desc() block_desc = pdesc.block(0) op_size = block_desc.op_size() - # Get global block ops - ops_list.append( - ([block_desc.op(i) for i in range(op_size)], op_size, set())) # Only process one level of nested subblock. ops_list.extend(_process_sub_block_pair(pdesc, SUB_BLOCK_PAIR)) + skip_opt_set = set() + for _, _, skip_opt in ops_list: + skip_opt_set.update(skip_opt) + + # Get global block ops + ops_list.insert( + 0, ([block_desc.op(i) for i in range(op_size)], op_size, skip_opt_set)) cfgs = [ ControlFlowGraph(input_program, ops, forward_num, skip_opt) for ops, forward_num, skip_opt in ops_list @@ -358,7 +375,7 @@ def _get_cfgs(input_program): return cfgs -def memory_optimize(input_program, print_log=False, level=0): +def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0): """Optimize memory by reusing var memory. Note: it doesn't not support subblock nested in subblock. @@ -374,10 +391,20 @@ def memory_optimize(input_program, print_log=False, level=0): PRINT_LOG = print_log cfgs = _get_cfgs(input_program) for cfg in cfgs: - cfg.memory_optimize(level) + cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level) -def release_memory(input_program): +def release_memory(input_program, skip_opt_set=None): + """ + Modify the input program and insert :code:`delete_op` to early drop not used + variables. The modification will be performed inplace. + + Notes: This is an experimental API and could be removed in next few + releases. Users should not use this API. + + Args: + input_program(Program): The program will be inserted :code:`delete_op`. + """ cfgs = _get_cfgs(input_program) for cfg in cfgs: - cfg.release_memory() + cfg.release_memory(skip_opt_set=skip_opt_set) diff --git a/python/paddle/fluid/transpiler/ps_dispatcher.py b/python/paddle/fluid/transpiler/ps_dispatcher.py index d6a68677527deb09ace0e3a23cbc093d6d7b4349..dcffadd531719431f27feb464ed58a65c04770ee 100644 --- a/python/paddle/fluid/transpiler/ps_dispatcher.py +++ b/python/paddle/fluid/transpiler/ps_dispatcher.py @@ -33,15 +33,21 @@ class PSDispatcher(object): def dispatch(self, varlist): """ - :param varlist: a list of Variables - :return: a map of pserver endpoint -> varname + Args: + varlist(list): a list of Variables + Returns: + a map of pserver endpoint -> varname """ AssertionError("Interface has not been implemented.") class HashName(PSDispatcher): """ - Hash variable names to several endpoints + Hash variable names to several endpoints using python + "hash()" function. + + Args: + pserver_endpoints (list): list of endpoint(ip:port). """ def __init__(self, pserver_endpoints): @@ -61,7 +67,11 @@ class HashName(PSDispatcher): class RoundRobin(PSDispatcher): """ - Distribute variables to serveral endpoints. + Distribute variables to serveral endpoints using + RondRobin method. + + Args: + pserver_endpoints (list): list of endpoint(ip:port). """ def __init__(self, pserver_endpoints): diff --git a/python/paddle/fluid/unique_name.py b/python/paddle/fluid/unique_name.py index 33c53113ae7e8ed9aeada31f2aed6990b6fea110..b125eba4f83c588fa2fa81a357604a7d8592ea80 100644 --- a/python/paddle/fluid/unique_name.py +++ b/python/paddle/fluid/unique_name.py @@ -14,9 +14,10 @@ import collections import contextlib +import six import sys -__all__ = ['generate', 'switch', 'guard', 'UniqueNameGenerator'] +__all__ = ['generate', 'switch', 'guard'] class UniqueNameGenerator(object): @@ -67,8 +68,10 @@ def switch(new_generator=None): @contextlib.contextmanager def guard(new_generator=None): - if isinstance(new_generator, basestring): + if isinstance(new_generator, six.string_types): new_generator = UniqueNameGenerator(new_generator) + elif isinstance(new_generator, six.binary_type): + new_generator = UniqueNameGenerator(new_generator.decode()) old = switch(new_generator) yield switch(old) diff --git a/python/paddle/libs/__init__.py b/python/paddle/libs/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..34d4f4d07ed0d452c1965c5f1f198230571931aa --- /dev/null +++ b/python/paddle/libs/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# used for setup.py.in to store the thirdparty shared libraries diff --git a/python/paddle/reader/__init__.py b/python/paddle/reader/__init__.py index 3b059735a924d58714cd88a761eb83143f1192d6..678026cf95970e8ff58c1bad20246059ffb464c1 100644 --- a/python/paddle/reader/__init__.py +++ b/python/paddle/reader/__init__.py @@ -66,9 +66,9 @@ An example implementation for multiple item data reader creator: TODO(yuyang18): Should we add whole design doc here? """ -import decorator -from decorator import * +import paddle.reader.decorator +from paddle.reader.decorator import * -import creator +import paddle.reader.creator __all__ = decorator.__all__ + ['creator'] diff --git a/python/paddle/reader/creator.py b/python/paddle/reader/creator.py index 4c905d959fad4e8c1a8826ce8dc60c5fa834514d..c861020225fb6fe0a29653363c2151b20dc8f578 100644 --- a/python/paddle/reader/creator.py +++ b/python/paddle/reader/creator.py @@ -67,11 +67,14 @@ def recordio(paths, buf_size=100): import recordio as rec import paddle.reader.decorator as dec - import cPickle as pickle + import six + import six.moves.cPickle as pickle def reader(): - if isinstance(paths, basestring): + if isinstance(paths, six.string_types): path = paths + elif isinstance(paths, six.binary_type): + path = paths.decode() else: path = ",".join(paths) f = rec.reader(path) diff --git a/python/paddle/reader/decorator.py b/python/paddle/reader/decorator.py index 44a6e344630bb35d28ee29078bf8727053a24bef..ce410e61b92e7d3f32fa5bfeb415e4b6c5fa9df6 100644 --- a/python/paddle/reader/decorator.py +++ b/python/paddle/reader/decorator.py @@ -20,7 +20,10 @@ __all__ = [ from threading import Thread import subprocess -from Queue import Queue +from six.moves.queue import Queue +from six.moves import zip_longest +from six.moves import map +from six.moves import zip import itertools import random import zlib @@ -42,7 +45,7 @@ def map_readers(func, *readers): rs = [] for r in readers: rs.append(r()) - for e in itertools.imap(func, *rs): + for e in map(func, *rs): yield e return reader @@ -148,16 +151,16 @@ def compose(*readers, **kwargs): for r in readers: rs.append(r()) if not check_alignment: - for outputs in itertools.izip(*rs): - yield sum(map(make_tuple, outputs), ()) + for outputs in zip(*rs): + yield sum(list(map(make_tuple, outputs)), ()) else: - for outputs in itertools.izip_longest(*rs): + for outputs in zip_longest(*rs): for o in outputs: if o is None: # None will be not be present if compose is aligned raise ComposeNotAligned( "outputs of readers are not aligned.") - yield sum(map(make_tuple, outputs), ()) + yield sum(list(map(make_tuple, outputs)), ()) return reader @@ -306,7 +309,7 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False): args = (in_queue, out_queue, mapper, out_order) if order else ( in_queue, out_queue, mapper) workers = [] - for i in xrange(process_num): + for i in range(process_num): worker = Thread(target=target, args=args) worker.daemon = True workers.append(worker) @@ -336,7 +339,7 @@ def _buf2lines(buf, line_break="\n"): class PipeReader: """ - PipeReader read data by stream from a command, take it's + PipeReader read data by stream from a command, take it's stdout into a pipe buffer and redirect it to the parser to parse, then yield data as your desired format. @@ -352,7 +355,7 @@ class PipeReader: An example: .. code-block:: python - + def example_reader(): for f in myfiles: pr = PipeReader("cat %s"%f) diff --git a/python/paddle/reader/tests/decorator_test.py b/python/paddle/reader/tests/decorator_test.py index bee24d3b6579db5e99ec66931df201fdf9e1af07..537df489b9738864933b3a7922d178701db3d19f 100644 --- a/python/paddle/reader/tests/decorator_test.py +++ b/python/paddle/reader/tests/decorator_test.py @@ -136,7 +136,7 @@ class TestXmap(unittest.TestCase): reader = paddle.reader.xmap_readers(mapper, reader_creator_10(0), tNum, size, order) - for n in xrange(3): + for n in range(3): result = [] for i in reader(): result.append(i) @@ -156,7 +156,7 @@ class TestPipeReader(unittest.TestCase): import tempfile - records = [str(i) for i in xrange(5)] + records = [str(i) for i in range(5)] temp = tempfile.NamedTemporaryFile() try: with open(temp.name, 'w') as f: diff --git a/python/paddle/trainer/PyDataProviderWrapper.py b/python/paddle/trainer/PyDataProviderWrapper.py index 6af250772859811b3c48434ab005e50b435dd320..374976db9f17ad9b1fd33c5d4adf77155336d100 100644 --- a/python/paddle/trainer/PyDataProviderWrapper.py +++ b/python/paddle/trainer/PyDataProviderWrapper.py @@ -42,7 +42,7 @@ except ImportError: try: import cPickle as pickle except ImportError: - import pickle + import six.moves.cPickle as pickle import io diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index 460eb3b3491a0626eb6ecbf89132e24177a2adaa..5b90facd49d655f56c037e087d86e41372cbfdb9 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -67,7 +67,7 @@ extension_module_name=[MODULE_NAME], then config_parser will call MODULE_NAME.get_config_funcs(g_config) MODULE_NAME.get_config_funcs() should return a dictionary of name to functions, those functions will be available in the config file. -See trainer/tests/config_parser_test.py for example +See legacy/trainer/tests/config_parser_test.py for example To use this from paddle_trainer, paddle_trainer should be called with --config_args=extension_module_name=[MODULE_NAME] diff --git a/python/paddle/trainer_config_helpers/attrs.py b/python/paddle/trainer_config_helpers/attrs.py index e6f87ce61b1d16d4f98f111626776aa52c2ec35b..4e3beaf639bad9fed2862a5477095b66ef4b9aee 100644 --- a/python/paddle/trainer_config_helpers/attrs.py +++ b/python/paddle/trainer_config_helpers/attrs.py @@ -240,14 +240,15 @@ class ExtraLayerAttribute(object): :type error_clipping_threshold: float :param drop_rate: Dropout rate. Dropout will create a mask on layer output. The dropout rate is the zero rate of this mask. The - details of what dropout is please refer to `here - `_. + details of what dropout is please refer to `JMLRdropout + `_. :type drop_rate: float :param device: device ID of layer. device=-1, use CPU. device>=0, use GPU. - The details allocation in parallel_nn please refer to `here - `_. + The details allocation in parallel_nn please refer to `use_case + `_. :type device: int """ diff --git a/python/paddle/trainer_config_helpers/data_sources.py b/python/paddle/trainer_config_helpers/data_sources.py index ab9a2562dcccb394c0b24741ceeb10061e40cb9a..a2a32d848cbc4200397e6a12a3662419102da0a9 100644 --- a/python/paddle/trainer_config_helpers/data_sources.py +++ b/python/paddle/trainer_config_helpers/data_sources.py @@ -20,7 +20,7 @@ from .utils import deprecated try: import cPickle as pickle except ImportError: - import pickle + import six.moves.cPickle as pickle __all__ = ['define_py_data_sources2'] diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index ebc31b23e0f5504b4bebccabe996b054c7fbce3b..ee34c157334b533b9c330b8103424964d7df510b 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -28,7 +28,7 @@ from .default_decorators import * try: import cPickle as pickle except ImportError: - import pickle + import six.moves.cPickle as pickle import copy __all__ = [ @@ -2556,7 +2556,7 @@ def img_conv_layer(input, the output will be obtained by concatenating the two results. The details of grouped convolution, please refer to: - `ImageNet Classification with Deep Convolutional Neural Networks + `ImageNet Classification With Deep Convolutional Neural Networks `_ The example usage is: @@ -4182,9 +4182,9 @@ def recurrent_group(step, input, reverse=False, name=None, targetInlink=None): You can see following configs for further usages: - - time steps: lstmemory_group, paddle/gserver/tests/sequence_layer_group.conf, \ + - time steps: lstmemory_group, paddle/legacy/gserver/tests/sequence_layer_group.conf, \ demo/seqToseq/seqToseq_net.py - - sequence steps: paddle/gserver/tests/sequence_nest_layer_group.conf + - sequence steps: paddle/legacy/gserver/tests/sequence_nest_layer_group.conf :param step: A step function which takes the input of recurrent_group as its own input and returns values as recurrent_group's output every time step. @@ -5678,8 +5678,8 @@ def warp_ctc_layer(input, `_ library, which is used in `Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin `_, to compute Connectionist Temporal - Classification (CTC) loss. Besides, another `warp-ctc - `_ repository, which is forked from + Classification (CTC) loss. Besides, another `warp-ctc repository + `_ , which is forked from the official one, is maintained to enable more compiling options. During the building process, PaddlePaddle will clone the source codes, build and install it to :code:`third_party/install/warpctc` directory. diff --git a/python/paddle/v2/dataset/cifar.py b/python/paddle/v2/dataset/cifar.py index 0a2a1ced11ee5cb2fb407b229ce810d553c2fa46..662655c836dbc54bd6187dcd3dac7354d6c8ecd1 100644 --- a/python/paddle/v2/dataset/cifar.py +++ b/python/paddle/v2/dataset/cifar.py @@ -43,7 +43,7 @@ CIFAR100_URL = URL_PREFIX + 'cifar-100-python.tar.gz' CIFAR100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85' -def reader_creator(filename, sub_name): +def reader_creator(filename, sub_name, cycle=False): def read_batch(batch): data = batch['data'] labels = batch.get('labels', batch.get('fine_labels', None)) @@ -56,10 +56,13 @@ def reader_creator(filename, sub_name): names = (each_item.name for each_item in f if sub_name in each_item.name) - for name in names: - batch = cPickle.load(f.extractfile(name)) - for item in read_batch(batch): - yield item + while True: + for name in names: + batch = cPickle.load(f.extractfile(name)) + for item in read_batch(batch): + yield item + if not cycle: + break return reader @@ -94,34 +97,40 @@ def test100(): 'test') -def train10(): +def train10(cycle=False): """ CIFAR-10 training set creator. It returns a reader creator, each sample in the reader is image pixels in [0, 1] and label in [0, 9]. + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: Training reader creator :rtype: callable """ return reader_creator( paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), - 'data_batch') + 'data_batch', + cycle=cycle) -def test10(): +def test10(cycle=False): """ CIFAR-10 test set creator. It returns a reader creator, each sample in the reader is image pixels in [0, 1] and label in [0, 9]. + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: Test reader creator. :rtype: callable """ return reader_creator( paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), - 'test_batch') + 'test_batch', + cycle=cycle) def fetch(): diff --git a/python/paddle/v2/dataset/conll05.py b/python/paddle/v2/dataset/conll05.py index 0d544efac9cd20157f87b5cd3b68f97ab5ed2dbc..8312900dc43fdd64cc1a205ab846b6f1deaecf5d 100644 --- a/python/paddle/v2/dataset/conll05.py +++ b/python/paddle/v2/dataset/conll05.py @@ -29,13 +29,13 @@ __all__ = ['test, get_dict', 'get_embedding', 'convert'] DATA_URL = 'http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz' DATA_MD5 = '387719152ae52d60422c016e92a742fc' -WORDDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/wordDict.txt' +WORDDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FwordDict.txt' WORDDICT_MD5 = 'ea7fb7d4c75cc6254716f0177a506baa' -VERBDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/verbDict.txt' +VERBDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FverbDict.txt' VERBDICT_MD5 = '0d2977293bbb6cbefab5b0f97db1e77c' -TRGDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/targetDict.txt' +TRGDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FtargetDict.txt' TRGDICT_MD5 = 'd8c7f03ceb5fc2e5a0fa7503a4353751' -EMB_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/emb' +EMB_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2Femb' EMB_MD5 = 'bf436eb0faa1f6f9103017f8be57cdb7' UNK_IDX = 0 diff --git a/python/paddle/v2/dataset/flowers.py b/python/paddle/v2/dataset/flowers.py index 7bdddeaabec733ef26b3f766c6437f5c53d65044..db12076d54064781bd1060947497622b14783768 100644 --- a/python/paddle/v2/dataset/flowers.py +++ b/python/paddle/v2/dataset/flowers.py @@ -76,7 +76,8 @@ def reader_creator(data_file, dataset_name, mapper, buffered_size=1024, - use_xmap=True): + use_xmap=True, + cycle=False): ''' 1. read images from tar file and merge images into batch files in 102flowers.tgz_batch/ @@ -96,6 +97,8 @@ def reader_creator(data_file, :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: data reader :rtype: callable ''' @@ -108,23 +111,27 @@ def reader_creator(data_file, file_list = batch_images_from_tar(data_file, dataset_name, img2label) def reader(): - for file in open(file_list): - file = file.strip() - batch = None - with open(file, 'r') as f: - batch = cPickle.load(f) - data = batch['data'] - labels = batch['label'] - for sample, label in itertools.izip(data, batch['label']): - yield sample, int(label) - 1 + while True: + for file in open(file_list): + file = file.strip() + batch = None + with open(file, 'r') as f: + batch = cPickle.load(f) + data = batch['data'] + labels = batch['label'] + for sample, label in itertools.izip(data, batch['label']): + yield sample, int(label) - 1 + if not cycle: + break if use_xmap: - return xmap_readers(mapper, reader, cpu_count(), buffered_size) + cpu_num = int(os.environ.get('CPU_NUM', cpu_count())) + return xmap_readers(mapper, reader, cpu_num, buffered_size) else: return map_readers(mapper, reader) -def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): +def train(mapper=train_mapper, buffered_size=1024, use_xmap=True, cycle=False): ''' Create flowers training set reader. It returns a reader, each sample in the reader is @@ -137,17 +144,23 @@ def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: train data reader :rtype: callable ''' return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), TRAIN_FLAG, mapper, - buffered_size, use_xmap) + download(SETID_URL, 'flowers', SETID_MD5), + TRAIN_FLAG, + mapper, + buffered_size, + use_xmap, + cycle=cycle) -def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): +def test(mapper=test_mapper, buffered_size=1024, use_xmap=True, cycle=False): ''' Create flowers test set reader. It returns a reader, each sample in the reader is @@ -160,14 +173,20 @@ def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: test data reader :rtype: callable ''' return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), TEST_FLAG, mapper, - buffered_size, use_xmap) + download(SETID_URL, 'flowers', SETID_MD5), + TEST_FLAG, + mapper, + buffered_size, + use_xmap, + cycle=cycle) def valid(mapper=test_mapper, buffered_size=1024, use_xmap=True): diff --git a/python/paddle/v2/dataset/mnist.py b/python/paddle/v2/dataset/mnist.py index 9f675bed895223e054cd3bb6e504fe1607f19858..026cf501cfb35ab3fe35d24f52d3c271565482ef 100644 --- a/python/paddle/v2/dataset/mnist.py +++ b/python/paddle/v2/dataset/mnist.py @@ -68,8 +68,14 @@ def reader_creator(image_filename, label_filename, buffer_size): for i in xrange(buffer_size): yield images[i, :], int(labels[i]) finally: - m.terminate() - l.terminate() + try: + m.terminate() + except: + pass + try: + l.terminate() + except: + pass return reader @@ -112,7 +118,7 @@ def fetch(): paddle.v2.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', TRAIN_IMAGE_MD5) paddle.v2.dataset.common.download(TRAIN_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) paddle.v2.dataset.common.download(TEST_IMAGE_URL, 'mnist', TEST_IMAGE_MD5) - paddle.v2.dataset.common.download(TEST_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) + paddle.v2.dataset.common.download(TEST_LABEL_URL, 'mnist', TEST_LABEL_MD5) def convert(path): diff --git a/python/paddle/v2/dataset/wmt14.py b/python/paddle/v2/dataset/wmt14.py index 5104e29051e4480f3a7eb18421f1b519841b009b..b9e602f324ad9bf43416b420c6d5697050a5c802 100644 --- a/python/paddle/v2/dataset/wmt14.py +++ b/python/paddle/v2/dataset/wmt14.py @@ -15,7 +15,7 @@ WMT14 dataset. The original WMT14 dataset is too large and a small set of data for set is provided. This module will download dataset from -http://paddlepaddle.cdn.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz and +http://paddlemodels.bj.bcebos.com/wmt/wmt14.tgz and parse training set and test set into paddle reader creators. """ @@ -37,11 +37,10 @@ URL_DEV_TEST = ('http://www-lium.univ-lemans.fr/~schwenk/' MD5_DEV_TEST = '7d7897317ddd8ba0ae5c5fa7248d3ff5' # this is a small set of data for test. The original data is too large and # will be add later. -URL_TRAIN = ('http://paddlepaddle.cdn.bcebos.com/demo/' - 'wmt_shrinked_data/wmt14.tgz') +URL_TRAIN = ('http://paddlemodels.bj.bcebos.com/wmt/wmt14.tgz') MD5_TRAIN = '0791583d57d5beb693b9414c5b36798c' # BLEU of this trained model is 26.92 -URL_MODEL = 'http://paddlepaddle.bj.bcebos.com/demo/wmt_14/wmt14_model.tar.gz' +URL_MODEL = 'http://paddlemodels.bj.bcebos.com/wmt%2Fwmt14.tgz' MD5_MODEL = '0cb4a5366189b6acba876491c8724fa3' START = "" diff --git a/python/paddle/v2/image.py b/python/paddle/v2/image.py index 9235c41e9eb95b25a0dc53a494a203e7a4525981..08d8bd68f9b7eb703c15f7cb5ad1300969db5713 100644 --- a/python/paddle/v2/image.py +++ b/python/paddle/v2/image.py @@ -182,7 +182,7 @@ def resize_short(im, size): h_new = size * h / w else: w_new = size * w / h - im = cv2.resize(im, (h_new, w_new), interpolation=cv2.INTER_CUBIC) + im = cv2.resize(im, (w_new, h_new), interpolation=cv2.INTER_CUBIC) return im @@ -324,7 +324,6 @@ def simple_transform(im, if np.random.randint(2) == 0: im = left_right_flip(im, is_color) else: - im = center_crop(im, crop_size, is_color) im = center_crop(im, crop_size, is_color=is_color) if len(im.shape) == 3: im = to_chw(im) diff --git a/python/paddle/v2/inference.py b/python/paddle/v2/inference.py index 14b64742fd09bf6c197c5d1aa2354271293df239..28ee042282a08be32c13d91312fd97b211277522 100644 --- a/python/paddle/v2/inference.py +++ b/python/paddle/v2/inference.py @@ -63,7 +63,7 @@ class Inference(object): assert isinstance(val, api.Vector) val.copyFromNumpyArray(parameters.get(name).flatten()) # the setValueUpdated function is called in randomize, zeroMem, - # load function in paddle/parameter/Parameter.cpp. But in the + # load function in paddle/legacy/parameter/Parameter.cpp. But in the # inference mode, the setValueUpdated is never called, it will # cause the parameter will not be dispatched # in MultiGradientMachine for multi-GPU. So setValueUpdated is diff --git a/python/paddle/v2/minibatch.py b/python/paddle/v2/minibatch.py index 317cf037c69f8639e3760fbfce20565127794fcb..3c6a53db3c2287e8ef5931a06ca5dad455665ee0 100644 --- a/python/paddle/v2/minibatch.py +++ b/python/paddle/v2/minibatch.py @@ -15,7 +15,7 @@ __all__ = ['batch'] -def batch(reader, batch_size): +def batch(reader, batch_size, drop_last=True): """ Create a batched reader. @@ -23,6 +23,8 @@ def batch(reader, batch_size): :type reader: callable :param batch_size: size of each mini-batch :type batch_size: int + :param drop_last: drop the last batch, if the size of last batch is not equal to batch_size. + :type drop_last: bool :return: the batched reader. :rtype: callable """ @@ -35,7 +37,7 @@ def batch(reader, batch_size): if len(b) == batch_size: yield b b = [] - if b: + if drop_last == False and len(b) != 0: yield b return batch_reader diff --git a/python/requirements.txt b/python/requirements.txt index ea827e9d5a0dcf8eb2ede1f6eaa88c777a138816..f8298a63612cb217ce0e711e78fffdf86b73313d 100644 --- a/python/requirements.txt +++ b/python/requirements.txt @@ -1,5 +1,5 @@ requests==2.9.2 -numpy>=1.12 +numpy>=1.12,<=1.14 #TODO:change to ">=1.12" when numpy fix bug in 1.15 and higher version protobuf==3.1 recordio>=0.1.0 matplotlib @@ -8,4 +8,4 @@ scipy>=0.19.0 Pillow nltk>=3.2.2 graphviz -LinkChecker +six diff --git a/python/setup.py.in b/python/setup.py.in index 8257f1d5e212a84188a4c51bc2d0f4d4c7af91fb..4a6cddbbea4903f5a65123aa19b7e978b335f32b 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -1,14 +1,13 @@ from setuptools import setup, Distribution, Extension import subprocess +import os +import re +import shutil class BinaryDistribution(Distribution): def has_ext_modules(foo): return True -MAJOR = 0 -MINOR = 11 -PATCH = 0 RC = 0 -ISTAGED = False @@ -18,16 +17,51 @@ def git_commit(): git_commit = subprocess.Popen(cmd, stdout = subprocess.PIPE).communicate()[0].strip() except: git_commit = 'Unknown' - return git_commit + git_commit = git_commit.decode() + return str(git_commit) + +def _get_version_detail(idx): + assert idx < 3, "vesion info consists of %(major)d.%(minor)d.%(patch)d, \ + so detail index must less than 3" + + if re.match('@TAG_VERSION_REGEX@', '@PADDLE_VERSION@'): + version_details = '@PADDLE_VERSION@'.split('.') + + if len(version_details) == 3: + return version_details[idx] + + return 0 + +def get_major(): + return int(_get_version_detail(0)) + +def get_minor(): + return int(_get_version_detail(1)) + +def get_patch(): + return str(_get_version_detail(2)) + +def is_taged(): + try: + cmd = ['git', 'describe', '--exact-match', '--tags', 'HEAD', '2>/dev/null'] + git_tag = subprocess.Popen(cmd, stdout = subprocess.PIPE).communicate()[0].strip() + git_tag = git_tag.decode() + except: + return False + + if str(git_tag).replace('v', '') == '@PADDLE_VERSION@': + return True + else: + return False def write_version_py(filename='paddle/version.py'): cnt = ''' # THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY # -full_version = '%(major)d.%(minor)d.%(patch)d' +full_version = '%(major)d.%(minor)d.%(patch)s' major = '%(major)d' minor = '%(minor)d' -patch = '%(patch)d' +patch = '%(patch)s' rc = '%(rc)d' istaged = %(istaged)s commit = '%(commit)s' @@ -35,13 +69,13 @@ with_mkl = '%(with_mkl)s' def show(): if istaged: - print 'full_version:', full_version - print 'major:', major - print 'minor:', minor - print 'patch:', patch - print 'rc:', rc + print('full_version:', full_version) + print('major:', major) + print('minor:', minor) + print('patch:', patch) + print('rc:', rc) else: - print 'commit:', commit + print('commit:', commit) def mkl(): return with_mkl @@ -49,19 +83,20 @@ def mkl(): commit = git_commit() with open(filename, 'w') as f: f.write(cnt % { - 'major': MAJOR, - 'minor': MINOR, - 'patch': PATCH, + 'major': get_major(), + 'minor': get_minor(), + 'patch': get_patch(), 'rc': RC, 'version': '${PADDLE_VERSION}', 'commit': commit, - 'istaged': ISTAGED, + 'istaged': is_taged(), 'with_mkl': '@WITH_MKL@'}) write_version_py(filename='@PADDLE_BINARY_DIR@/python/paddle/version.py') packages=['paddle', + 'paddle.libs', 'paddle.utils', 'paddle.dataset', 'paddle.reader', @@ -69,6 +104,8 @@ packages=['paddle', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', 'paddle.fluid.layers', + 'paddle.fluid.contrib', + 'paddle.fluid.contrib.decoder', 'paddle.fluid.transpiler', 'paddle.fluid.transpiler.details'] @@ -93,9 +130,9 @@ if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: paddle_bins = '' if '${WITH_FLUID_ONLY}'== 'OFF': paddle_bin_dir = 'opt/paddle/bin' - paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', - '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', - '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main', + paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/legacy/trainer/paddle_trainer', + '${PADDLE_BINARY_DIR}/paddle/legacy/trainer/paddle_merge_model', + '${PADDLE_BINARY_DIR}/paddle/legacy/pserver/paddle_pserver_main', '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] package_data={'paddle.fluid': ['core.so']} @@ -113,12 +150,49 @@ package_dir={ } if '${WITH_FLUID_ONLY}'== 'OFF': package_dir['py_paddle']='${PADDLE_BINARY_DIR}/python/py_paddle' - -paddle_rt_lib_dir = 'lib' -paddle_rt_libs = ['${WARPCTC_LIBRARIES}'] -if '${MKL_SHARED_LIBS}'!= '': - paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';') +# put all thirdparty libraries in paddle.libs +package_data['paddle.libs']=['libwarpctc.so'] +libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs' +shutil.copy('${WARPCTC_LIBRARIES}', libs_path) +if '${WITH_MKL}' == 'ON': + shutil.copy('${MKLML_LIB}', libs_path) + shutil.copy('${MKLML_IOMP_LIB}', libs_path) + package_data['paddle.libs']+=['libmklml_intel.so','libiomp5.so'] +if '${WITH_MKLDNN}' == 'ON': + # TODO(typhoonzero): use install_name_tool to patch mkl libs once + # we can support mkl on mac. + # + # change rpath of libmkldnn.so.0, add $ORIGIN/ to it. + # The reason is that all thirdparty libraries in the same directory, + # thus, libmkldnn.so.0 will find libmklml_intel.so and libiomp5.so. + command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}" + if os.system(command) != 0: + raise Exception("patch libmkldnn.so failed, command: %s" % command) + package_data['paddle.libs']+=['libmkldnn.so.0'] + shutil.copy('${MKLDNN_SHARED_LIB}', libs_path) +# remove unused paddle/libs/__init__.py +os.remove(libs_path+'/__init__.py') +package_dir['paddle.libs']=libs_path + +# change rpath of core.so, add $ORIGIN/../libs/ to it. +# The reason is that libwarpctc.so, libiomp5.so etc are in paddle.libs, and +# core.so is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries. +# This operation will fix https://github.com/PaddlePaddle/Paddle/issues/3213 +if "@APPLE@" == "1": + command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so" +else: + command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so" +if os.system(command) != 0: + raise Exception("patch core.so failed, command: %s" % command) +if '${WITH_FLUID_ONLY}'== 'OFF': + # change rpath of _swig_paddle.so. + if "@APPLE@" == "1": + command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so" + else: + command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so" + if os.system(command) != 0: + raise Exception("patch _swig_paddle.so failed, command: %s" % command) setup(name='${PACKAGE_NAME}', version='${PADDLE_VERSION}', @@ -128,6 +202,5 @@ setup(name='${PACKAGE_NAME}', ext_modules=[Extension('_foo', ['stub.cc'])], package_data=package_data, package_dir=package_dir, - scripts=paddle_bins, - data_files=[(paddle_rt_lib_dir, paddle_rt_libs)] + scripts=paddle_bins ) diff --git a/tools/check_ctest_hung.py b/tools/check_ctest_hung.py new file mode 100644 index 0000000000000000000000000000000000000000..7de76c381b29a1ff8dcf2167f0e861dc261aa47b --- /dev/null +++ b/tools/check_ctest_hung.py @@ -0,0 +1,53 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import sys +import re + + +def escape(input): + o = input.replace("\n", "") + o = o.replace("\r", "") + return o + + +def main(): + usage = """Usage: +1. Download the Paddle_PR_CI_*.log from TeamCity +2. run: python check_ctest_hung.py Paddle_PR_CI_*.log +3. If there is hung ctest, the result likes: +Diff: set(['test_parallel_executor_crf']) + """ + if len(sys.argv) < 2: + print(usage) + exit(0) + + logfile = sys.argv[1] + started = set() + passed = set() + with open(logfile, "r") as fn: + for l in fn.readlines(): + if l.find("Test ") != -1 and \ + l.find("Passed") != -1: + m = re.search("Test\s+#[0-9]*\:\s([a-z0-9_]+)", escape(l)) + passed.add(m.group(1)) + if l.find("Start ") != -1: + start_parts = escape(l).split(" ") + m = re.search("Start\s+[0-9]+\:\s([a-z0-9_]+)", escape(l)) + started.add(m.group(1)) + print "Diff: ", started - passed + + +if __name__ == "__main__": + main() diff --git a/tools/check_pr_approval.py b/tools/check_pr_approval.py new file mode 100644 index 0000000000000000000000000000000000000000..937b0be7562fab93157c16b942631f0a580dfc68 --- /dev/null +++ b/tools/check_pr_approval.py @@ -0,0 +1,49 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function +import sys +import json + + +def check_approval(count, required_reviewers): + json_buff = "" + for line in sys.stdin: + json_buff = "".join([json_buff, line]) + json_resp = json.loads(json_buff) + approves = 0 + approved_user_ids = [] + for review in json_resp: + if review["state"] == "APPROVED": + approves += 1 + approved_user_ids.append(review["user"]["id"]) + + # convert to int + required_reviewers_int = set() + for rr in required_reviewers: + required_reviewers_int.add(int(rr)) + + if len(set(approved_user_ids) & required_reviewers_int) >= count: + print("TRUE") + else: + print("FALSE") + + +if __name__ == "__main__": + if len(sys.argv) > 1 and sys.argv[1].isdigit(): + check_approval(int(sys.argv[1]), sys.argv[2:]) + else: + print( + "Usage: python check_pr_approval.py [count] [required reviewer id] ..." + ) diff --git a/tools/codestyle/.gitignore b/tools/codestyle/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..0d20b6487c61e7d1bde93acf4a14b7a89083a16d --- /dev/null +++ b/tools/codestyle/.gitignore @@ -0,0 +1 @@ +*.pyc diff --git a/.clang_format.hook b/tools/codestyle/clang_format.hook similarity index 100% rename from .clang_format.hook rename to tools/codestyle/clang_format.hook diff --git a/.copyright.hook b/tools/codestyle/copyright.hook similarity index 100% rename from .copyright.hook rename to tools/codestyle/copyright.hook diff --git a/tools/codestyle/cpplint_pre_commit.hook b/tools/codestyle/cpplint_pre_commit.hook index b194af76dc529fd52b0aedfab9c41d625fe64c0d..aa14d3a2a12208eda11e82d88bc582eb3d2f5893 100755 --- a/tools/codestyle/cpplint_pre_commit.hook +++ b/tools/codestyle/cpplint_pre_commit.hook @@ -4,10 +4,10 @@ TOTAL_ERRORS=0 # The trick to remove deleted files: https://stackoverflow.com/a/2413151 for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do - if [[ $file =~ ^(paddle/api/.*|paddle/capi/.*|paddle/contrib/.*|paddle/cuda/.*|paddle/function/.*|paddle/gserver/.*|paddle/math/.*|paddle/optimizer/.*|paddle/parameter/.*|paddle/pserver/.*|paddle/trainer/.*|paddle/utils/.*) ]]; then + if [[ $file =~ ^(paddle/legacy/api/.*|paddle/legacy/capi/.*|paddle/contrib/.*|paddle/legacy/cuda/.*|paddle/legacy/function/.*|paddle/legacy/gserver/.*|paddle/legacy/math/.*|paddle/legacy/optimizer/.*|paddle/legacy/parameter/.*|paddle/legacy/pserver/.*|paddle/legacy/trainer/.*|paddle/legacy/utils/.*|paddle/testing/TestUtil.*|patches/grpc/.*) ]]; then continue; else - cpplint $file; + cpplint --filter=-readability/fn_size $file; TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); fi done diff --git a/tools/codestyle/docstring_checker.py b/tools/codestyle/docstring_checker.py index 48100e5bf989520043b5ca372b02883faea8a9fd..8d4b24a0cf6b743b72dca58fd885f927560964bf 100644 --- a/tools/codestyle/docstring_checker.py +++ b/tools/codestyle/docstring_checker.py @@ -126,9 +126,10 @@ class DocstringChecker(BaseChecker): 'W9002': ('Doc string does not end with "." period', symbol + "-end-with", 'Used when a doc string does not end with a period'), - 'W9003': ('All args with their types must be mentioned in doc string', - symbol + "-with-all-args", - 'Used when not all arguments are in the doc string '), + 'W9003': + ('All args with their types must be mentioned in doc string %s', + symbol + "-with-all-args", + 'Used when not all arguments are in the doc string '), 'W9005': ('Missing docstring or docstring is too short', symbol + "-missing", 'Add docstring longer >=10'), 'W9006': ('Docstring indent error, use 4 space for indent', @@ -178,6 +179,8 @@ class DocstringChecker(BaseChecker): self.indent_style(node) def missing_doc_string(self, node): + if node.name.startswith("__") or node.name.startswith("_"): + return True if node.tolineno - node.fromlineno <= 10: return True @@ -199,12 +202,16 @@ class DocstringChecker(BaseChecker): doc = node.doc lines = doc.splitlines() + line_num = 0 for l in lines: + if line_num == 0: + continue cur_indent = len(l) - len(l.lstrip()) if cur_indent % indent != 0: self.add_message('W9006', node=node, line=node.fromlineno) return False + line_num += 1 return True @@ -284,6 +291,8 @@ class DocstringChecker(BaseChecker): True if successful otherwise False. """ + if node.name.startswith("__") or node.name.startswith("_"): + return True find = False for t in node.body: if not isinstance(t, astroid.Return): @@ -309,6 +318,8 @@ class DocstringChecker(BaseChecker): Returns: True if successful otherwise False. """ + if node.name.startswith("__") or node.name.startswith("_"): + return True args = [] for arg in node.args.get_children(): if (not isinstance(arg, astroid.AssignName)) \ @@ -320,15 +331,19 @@ class DocstringChecker(BaseChecker): return True parsed_args = doc.args + args_not_documented = set(args) - set(parsed_args) if len(args) > 0 and len(parsed_args) <= 0: - print "debug:parsed args: ", parsed_args - self.add_message('W9003', node=node, line=node.fromlineno) + self.add_message( + 'W9003', + node=node, + line=node.fromlineno, + args=list(args_not_documented)) return False for t in args: if t not in parsed_args: - print t, " with (type) not in ", parsed_args - self.add_message('W9003', node=node, line=node.fromlineno) + self.add_message( + 'W9003', node=node, line=node.fromlineno, args=[t, ]) return False return True diff --git a/tools/codestyle/pylint_pre_commit.hook b/tools/codestyle/pylint_pre_commit.hook index e7c92ba671e0eb778b2ab5447bea7c4b14fe761b..150a3f5666bd39d30b7e6518e58a14fb5fe2f14b 100755 --- a/tools/codestyle/pylint_pre_commit.hook +++ b/tools/codestyle/pylint_pre_commit.hook @@ -7,13 +7,13 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" export PYTHONPATH=$DIR:$PYTHONPATH # The trick to remove deleted files: https://stackoverflow.com/a/2413151 -for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do +for file in $(git diff --name-status | awk '$1 != "D" {print $2}'); do pylint --disable=all --load-plugins=docstring_checker \ --enable=doc-string-one-line,doc-string-end-with,doc-string-with-all-args,doc-string-triple-quotes,doc-string-missing,doc-string-indent-error,doc-string-with-returns,doc-string-with-raises $file; TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); done -#exit $TOTAL_ERRORS +exit $TOTAL_ERRORS #For now, just warning: -exit 0 +#exit 0 diff --git a/tools/diff_api.py b/tools/diff_api.py new file mode 100644 index 0000000000000000000000000000000000000000..97c739ed2a5627ad9fd326f206976a4579dc26a3 --- /dev/null +++ b/tools/diff_api.py @@ -0,0 +1,29 @@ +#!/usr/bin/env python +from __future__ import print_function +import difflib +import sys + +with open(sys.argv[1], 'r') as f: + origin = f.read() + origin = origin.splitlines() + +with open(sys.argv[2], 'r') as f: + new = f.read() + new = new.splitlines() + +differ = difflib.Differ() +result = differ.compare(origin, new) + +error = False +print('API Difference is: ') +for each_diff in result: + if each_diff[0] in ['-', '?']: # delete or change API is not allowed + error = True + elif each_diff[0] == '+': + error = True + + if each_diff[0] != ' ': + print(each_diff) + +if error: + sys.exit(1) diff --git a/tools/manylinux1/Dockerfile.x64 b/tools/manylinux1/Dockerfile.x64 index 34c54303bd0b30fc57f1b72bbbecef45ea398e53..987fc01cfc13bcefbb9d792a6780fdfff158755d 100644 --- a/tools/manylinux1/Dockerfile.x64 +++ b/tools/manylinux1/Dockerfile.x64 @@ -13,7 +13,7 @@ ENV PATH /opt/rh/devtoolset-2/root/usr/bin:$PATH ENV LD_LIBRARY_PATH /opt/rh/devtoolset-2/root/usr/lib64:/opt/rh/devtoolset-2/root/usr/lib:/usr/local/lib64:/usr/local/lib:${LD_LIBRARY_PATH} ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig -RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool xz freetype-devel libpng-devel +RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool xz freetype-devel libpng-devel graphviz COPY build_scripts /build_scripts RUN bash build_scripts/build.sh && \ bash build_scripts/install_nccl2.sh && rm -r build_scripts @@ -40,11 +40,13 @@ RUN wget -O /root/requirements.txt https://raw.githubusercontent.com/PaddlePaddl RUN LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27mu/bin/pip install -r /root/requirements.txt && \ LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27m/bin/pip install -r /root/requirements.txt && \ + LD_LIBRARY_PATH=/opt/_internal/cpython-3.5.1/lib/:${LD_LIBRARY_PATH} /opt/_internal/cpython-3.5.1/bin/pip3 install -r /root/requirements.txt && \ go get github.com/Masterminds/glide && \ rm -rf /root/requirements.txt RUN LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27mu/bin/pip install pre-commit 'ipython==5.3.0' opencv-python && \ - LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27m/bin/pip install pre-commit 'ipython==5.3.0' opencv-python + LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27m/bin/pip install pre-commit 'ipython==5.3.0' opencv-python && \ + LD_LIBRARY_PATH=/opt/_internal/cpython-3.5.1/lib/:${LD_LIBRARY_PATH} /opt/_internal/cpython-3.5.1/bin/pip3 install pre-commit 'ipython==5.3.0' opencv-python RUN wget -O /opt/swig-2.0.12.tar.gz https://cytranet.dl.sourceforge.net/project/swig/swig/swig-2.0.12/swig-2.0.12.tar.gz && \ cd /opt && tar xzf swig-2.0.12.tar.gz && cd /opt/swig-2.0.12 && ./configure && make && make install && cd /opt && rm swig-2.0.12.tar.gz diff --git a/tools/print_signatures.py b/tools/print_signatures.py new file mode 100644 index 0000000000000000000000000000000000000000..5e7ffd44c7b0ba2270069bc4467dc377a58b2417 --- /dev/null +++ b/tools/print_signatures.py @@ -0,0 +1,67 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Print all signature of a python module in alphabet order. + +Usage: + ./print_signature "paddle.fluid" > signature.txt +""" +import importlib +import inspect +import collections +import sys +import pydoc + +member_dict = collections.OrderedDict() + + +def visit_member(parent_name, member): + cur_name = ".".join([parent_name, member.__name__]) + if inspect.isclass(member): + for name, value in inspect.getmembers(member): + if hasattr(value, '__name__') and (not name.startswith("_") or + name == "__init__"): + visit_member(cur_name, value) + elif callable(member): + try: + member_dict[cur_name] = inspect.getargspec(member) + except TypeError: # special for PyBind method + member_dict[cur_name] = " ".join([ + line.strip() for line in pydoc.render_doc(member).split('\n') + if "->" in line + ]) + + else: + raise RuntimeError("Unsupported generate signature of member, type {0}". + format(str(type(member)))) + + +def visit_all_module(mod): + for member_name in ( + name + for name in (mod.__all__ if hasattr(mod, "__all__") else dir(mod)) + if not name.startswith("_")): + instance = getattr(mod, member_name, None) + if instance is None: + continue + if inspect.ismodule(instance): + visit_all_module(instance) + else: + visit_member(mod.__name__, instance) + + +visit_all_module(importlib.import_module(sys.argv[1])) + +for name in member_dict: + print name, member_dict[name] diff --git a/tools/test_runner.py b/tools/test_runner.py index 9dc750b89058cd73355a2f7984d577252c03526d..2d6a3cf8a97a3bbaa69b66f5343c54b750624329 100644 --- a/tools/test_runner.py +++ b/tools/test_runner.py @@ -12,19 +12,20 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function import unittest import os import sys import paddle.fluid as fluid import importlib -import cStringIO +from six.moves import cStringIO def main(): sys.path.append(os.getcwd()) some_test_failed = False for module_name in sys.argv[1:]: - buffer = cStringIO.StringIO() + buffer = cStringIO() main = fluid.Program() startup = fluid.Program() scope = fluid.core.Scope() @@ -37,8 +38,11 @@ def main(): res = unittest.TextTestRunner(stream=buffer).run(tests) if not res.wasSuccessful(): some_test_failed = True - print >> sys.stderr, module_name, 'failed\n', buffer.getvalue( - ) + print( + module_name, + 'failed\n', + buffer.getvalue(), + file=sys.stderr) if some_test_failed: exit(1)